1 /*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 *
21 * $FreeBSD$
22 */
23
24 /*
25 * Copyright (c) 2003, 2010, Oracle and/or its affiliates. All rights reserved.
26 * Copyright (c) 2016, Joyent, Inc. All rights reserved.
27 * Copyright (c) 2012, 2014 by Delphix. All rights reserved.
28 */
29
30 /*
31 * DTrace - Dynamic Tracing for Solaris
32 *
33 * This is the implementation of the Solaris Dynamic Tracing framework
34 * (DTrace). The user-visible interface to DTrace is described at length in
35 * the "Solaris Dynamic Tracing Guide". The interfaces between the libdtrace
36 * library, the in-kernel DTrace framework, and the DTrace providers are
37 * described in the block comments in the <sys/dtrace.h> header file. The
38 * internal architecture of DTrace is described in the block comments in the
39 * <sys/dtrace_impl.h> header file. The comments contained within the DTrace
40 * implementation very much assume mastery of all of these sources; if one has
41 * an unanswered question about the implementation, one should consult them
42 * first.
43 *
44 * The functions here are ordered roughly as follows:
45 *
46 * - Probe context functions
47 * - Probe hashing functions
48 * - Non-probe context utility functions
49 * - Matching functions
50 * - Provider-to-Framework API functions
51 * - Probe management functions
52 * - DIF object functions
53 * - Format functions
54 * - Predicate functions
55 * - ECB functions
56 * - Buffer functions
57 * - Enabling functions
58 * - DOF functions
59 * - Anonymous enabling functions
60 * - Consumer state functions
61 * - Helper functions
62 * - Hook functions
63 * - Driver cookbook functions
64 *
65 * Each group of functions begins with a block comment labelled the "DTrace
66 * [Group] Functions", allowing one to find each block by searching forward
67 * on capital-f functions.
68 */
69 #include <sys/errno.h>
70 #include <sys/param.h>
71 #include <sys/types.h>
72 #ifndef illumos
73 #include <sys/time.h>
74 #endif
75 #include <sys/stat.h>
76 #include <sys/conf.h>
77 #include <sys/systm.h>
78 #include <sys/endian.h>
79 #ifdef illumos
80 #include <sys/ddi.h>
81 #include <sys/sunddi.h>
82 #endif
83 #include <sys/cpuvar.h>
84 #include <sys/kmem.h>
85 #ifdef illumos
86 #include <sys/strsubr.h>
87 #endif
88 #include <sys/sysmacros.h>
89 #include <sys/dtrace_impl.h>
90 #include <sys/atomic.h>
91 #include <sys/cmn_err.h>
92 #ifdef illumos
93 #include <sys/mutex_impl.h>
94 #include <sys/rwlock_impl.h>
95 #endif
96 #include <sys/ctf_api.h>
97 #ifdef illumos
98 #include <sys/panic.h>
99 #include <sys/priv_impl.h>
100 #endif
101 #ifdef illumos
102 #include <sys/cred_impl.h>
103 #include <sys/procfs_isa.h>
104 #endif
105 #include <sys/taskq.h>
106 #ifdef illumos
107 #include <sys/mkdev.h>
108 #include <sys/kdi.h>
109 #endif
110 #include <sys/zone.h>
111 #include <sys/socket.h>
112 #include <netinet/in.h>
113 #include "strtolctype.h"
114
115 /* FreeBSD includes: */
116 #ifndef illumos
117 #include <sys/callout.h>
118 #include <sys/ctype.h>
119 #include <sys/eventhandler.h>
120 #include <sys/limits.h>
121 #include <sys/linker.h>
122 #include <sys/kdb.h>
123 #include <sys/jail.h>
124 #include <sys/kernel.h>
125 #include <sys/malloc.h>
126 #include <sys/lock.h>
127 #include <sys/mutex.h>
128 #include <sys/ptrace.h>
129 #include <sys/random.h>
130 #include <sys/rwlock.h>
131 #include <sys/sx.h>
132 #include <sys/sysctl.h>
133
134
135 #include <sys/mount.h>
136 #undef AT_UID
137 #undef AT_GID
138 #include <sys/vnode.h>
139 #include <sys/cred.h>
140
141 #include <sys/dtrace_bsd.h>
142
143 #include <netinet/in.h>
144
145 #include "dtrace_cddl.h"
146 #include "dtrace_debug.c"
147 #endif
148
149 #include "dtrace_xoroshiro128_plus.h"
150
151 /*
152 * DTrace Tunable Variables
153 *
154 * The following variables may be tuned by adding a line to /etc/system that
155 * includes both the name of the DTrace module ("dtrace") and the name of the
156 * variable. For example:
157 *
158 * set dtrace:dtrace_destructive_disallow = 1
159 *
160 * In general, the only variables that one should be tuning this way are those
161 * that affect system-wide DTrace behavior, and for which the default behavior
162 * is undesirable. Most of these variables are tunable on a per-consumer
163 * basis using DTrace options, and need not be tuned on a system-wide basis.
164 * When tuning these variables, avoid pathological values; while some attempt
165 * is made to verify the integrity of these variables, they are not considered
166 * part of the supported interface to DTrace, and they are therefore not
167 * checked comprehensively. Further, these variables should not be tuned
168 * dynamically via "mdb -kw" or other means; they should only be tuned via
169 * /etc/system.
170 */
171 int dtrace_destructive_disallow = 0;
172 #ifndef illumos
173 /* Positive logic version of dtrace_destructive_disallow for loader tunable */
174 int dtrace_allow_destructive = 1;
175 #endif
176 dtrace_optval_t dtrace_nonroot_maxsize = (16 * 1024 * 1024);
177 size_t dtrace_difo_maxsize = (256 * 1024);
178 dtrace_optval_t dtrace_dof_maxsize = (8 * 1024 * 1024);
179 size_t dtrace_statvar_maxsize = (16 * 1024);
180 size_t dtrace_actions_max = (16 * 1024);
181 size_t dtrace_retain_max = 1024;
182 dtrace_optval_t dtrace_helper_actions_max = 128;
183 dtrace_optval_t dtrace_helper_providers_max = 32;
184 dtrace_optval_t dtrace_dstate_defsize = (1 * 1024 * 1024);
185 size_t dtrace_strsize_default = 256;
186 dtrace_optval_t dtrace_cleanrate_default = 9900990; /* 101 hz */
187 dtrace_optval_t dtrace_cleanrate_min = 200000; /* 5000 hz */
188 dtrace_optval_t dtrace_cleanrate_max = (uint64_t)60 * NANOSEC; /* 1/minute */
189 dtrace_optval_t dtrace_aggrate_default = NANOSEC; /* 1 hz */
190 dtrace_optval_t dtrace_statusrate_default = NANOSEC; /* 1 hz */
191 dtrace_optval_t dtrace_statusrate_max = (hrtime_t)10 * NANOSEC; /* 6/minute */
192 dtrace_optval_t dtrace_switchrate_default = NANOSEC; /* 1 hz */
193 dtrace_optval_t dtrace_nspec_default = 1;
194 dtrace_optval_t dtrace_specsize_default = 32 * 1024;
195 dtrace_optval_t dtrace_stackframes_default = 20;
196 dtrace_optval_t dtrace_ustackframes_default = 20;
197 dtrace_optval_t dtrace_jstackframes_default = 50;
198 dtrace_optval_t dtrace_jstackstrsize_default = 512;
199 int dtrace_msgdsize_max = 128;
200 hrtime_t dtrace_chill_max = MSEC2NSEC(500); /* 500 ms */
201 hrtime_t dtrace_chill_interval = NANOSEC; /* 1000 ms */
202 int dtrace_devdepth_max = 32;
203 int dtrace_err_verbose;
204 hrtime_t dtrace_deadman_interval = NANOSEC;
205 hrtime_t dtrace_deadman_timeout = (hrtime_t)10 * NANOSEC;
206 hrtime_t dtrace_deadman_user = (hrtime_t)30 * NANOSEC;
207 hrtime_t dtrace_unregister_defunct_reap = (hrtime_t)60 * NANOSEC;
208 #ifndef illumos
209 int dtrace_memstr_max = 4096;
210 int dtrace_bufsize_max_frac = 128;
211 #endif
212
213 /*
214 * DTrace External Variables
215 *
216 * As dtrace(7D) is a kernel module, any DTrace variables are obviously
217 * available to DTrace consumers via the backtick (`) syntax. One of these,
218 * dtrace_zero, is made deliberately so: it is provided as a source of
219 * well-known, zero-filled memory. While this variable is not documented,
220 * it is used by some translators as an implementation detail.
221 */
222 const char dtrace_zero[256] = { 0 }; /* zero-filled memory */
223
224 /*
225 * DTrace Internal Variables
226 */
227 #ifdef illumos
228 static dev_info_t *dtrace_devi; /* device info */
229 #endif
230 #ifdef illumos
231 static vmem_t *dtrace_arena; /* probe ID arena */
232 static vmem_t *dtrace_minor; /* minor number arena */
233 #else
234 static taskq_t *dtrace_taskq; /* task queue */
235 static struct unrhdr *dtrace_arena; /* Probe ID number. */
236 #endif
237 static dtrace_probe_t **dtrace_probes; /* array of all probes */
238 static int dtrace_nprobes; /* number of probes */
239 static dtrace_provider_t *dtrace_provider; /* provider list */
240 static dtrace_meta_t *dtrace_meta_pid; /* user-land meta provider */
241 static int dtrace_opens; /* number of opens */
242 static int dtrace_helpers; /* number of helpers */
243 static int dtrace_getf; /* number of unpriv getf()s */
244 #ifdef illumos
245 static void *dtrace_softstate; /* softstate pointer */
246 #endif
247 static dtrace_hash_t *dtrace_bymod; /* probes hashed by module */
248 static dtrace_hash_t *dtrace_byfunc; /* probes hashed by function */
249 static dtrace_hash_t *dtrace_byname; /* probes hashed by name */
250 static dtrace_toxrange_t *dtrace_toxrange; /* toxic range array */
251 static int dtrace_toxranges; /* number of toxic ranges */
252 static int dtrace_toxranges_max; /* size of toxic range array */
253 static dtrace_anon_t dtrace_anon; /* anonymous enabling */
254 static kmem_cache_t *dtrace_state_cache; /* cache for dynamic state */
255 static uint64_t dtrace_vtime_references; /* number of vtimestamp refs */
256 static kthread_t *dtrace_panicked; /* panicking thread */
257 static dtrace_ecb_t *dtrace_ecb_create_cache; /* cached created ECB */
258 static dtrace_genid_t dtrace_probegen; /* current probe generation */
259 static dtrace_helpers_t *dtrace_deferred_pid; /* deferred helper list */
260 static dtrace_enabling_t *dtrace_retained; /* list of retained enablings */
261 static dtrace_genid_t dtrace_retained_gen; /* current retained enab gen */
262 static dtrace_dynvar_t dtrace_dynhash_sink; /* end of dynamic hash chains */
263 static int dtrace_dynvar_failclean; /* dynvars failed to clean */
264 #ifndef illumos
265 static struct mtx dtrace_unr_mtx;
266 MTX_SYSINIT(dtrace_unr_mtx, &dtrace_unr_mtx, "Unique resource identifier", MTX_DEF);
267 static eventhandler_tag dtrace_kld_load_tag;
268 static eventhandler_tag dtrace_kld_unload_try_tag;
269 #endif
270
271 /*
272 * DTrace Locking
273 * DTrace is protected by three (relatively coarse-grained) locks:
274 *
275 * (1) dtrace_lock is required to manipulate essentially any DTrace state,
276 * including enabling state, probes, ECBs, consumer state, helper state,
277 * etc. Importantly, dtrace_lock is _not_ required when in probe context;
278 * probe context is lock-free -- synchronization is handled via the
279 * dtrace_sync() cross call mechanism.
280 *
281 * (2) dtrace_provider_lock is required when manipulating provider state, or
282 * when provider state must be held constant.
283 *
284 * (3) dtrace_meta_lock is required when manipulating meta provider state, or
285 * when meta provider state must be held constant.
286 *
287 * The lock ordering between these three locks is dtrace_meta_lock before
288 * dtrace_provider_lock before dtrace_lock. (In particular, there are
289 * several places where dtrace_provider_lock is held by the framework as it
290 * calls into the providers -- which then call back into the framework,
291 * grabbing dtrace_lock.)
292 *
293 * There are two other locks in the mix: mod_lock and cpu_lock. With respect
294 * to dtrace_provider_lock and dtrace_lock, cpu_lock continues its historical
295 * role as a coarse-grained lock; it is acquired before both of these locks.
296 * With respect to dtrace_meta_lock, its behavior is stranger: cpu_lock must
297 * be acquired _between_ dtrace_meta_lock and any other DTrace locks.
298 * mod_lock is similar with respect to dtrace_provider_lock in that it must be
299 * acquired _between_ dtrace_provider_lock and dtrace_lock.
300 */
301 static kmutex_t dtrace_lock; /* probe state lock */
302 static kmutex_t dtrace_provider_lock; /* provider state lock */
303 static kmutex_t dtrace_meta_lock; /* meta-provider state lock */
304
305 #ifndef illumos
306 /* XXX FreeBSD hacks. */
307 #define cr_suid cr_svuid
308 #define cr_sgid cr_svgid
309 #define ipaddr_t in_addr_t
310 #define mod_modname pathname
311 #define vuprintf vprintf
312 #ifndef crgetzoneid
313 #define crgetzoneid(_a) 0
314 #endif
315 #define ttoproc(_a) ((_a)->td_proc)
316 #define SNOCD 0
317 #define CPU_ON_INTR(_a) 0
318
319 #define PRIV_EFFECTIVE (1 << 0)
320 #define PRIV_DTRACE_KERNEL (1 << 1)
321 #define PRIV_DTRACE_PROC (1 << 2)
322 #define PRIV_DTRACE_USER (1 << 3)
323 #define PRIV_PROC_OWNER (1 << 4)
324 #define PRIV_PROC_ZONE (1 << 5)
325 #define PRIV_ALL ~0
326
327 SYSCTL_DECL(_debug_dtrace);
328 SYSCTL_DECL(_kern_dtrace);
329 #endif
330
331 #ifdef illumos
332 #define curcpu CPU->cpu_id
333 #endif
334
335
336 /*
337 * DTrace Provider Variables
338 *
339 * These are the variables relating to DTrace as a provider (that is, the
340 * provider of the BEGIN, END, and ERROR probes).
341 */
342 static dtrace_pattr_t dtrace_provider_attr = {
343 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
344 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
345 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
346 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
347 { DTRACE_STABILITY_STABLE, DTRACE_STABILITY_STABLE, DTRACE_CLASS_COMMON },
348 };
349
350 static void
351 dtrace_nullop(void)
352 {}
353
354 static dtrace_pops_t dtrace_provider_ops = {
355 .dtps_provide = (void (*)(void *, dtrace_probedesc_t *))dtrace_nullop,
356 .dtps_provide_module = (void (*)(void *, modctl_t *))dtrace_nullop,
357 .dtps_enable = (void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
358 .dtps_disable = (void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
359 .dtps_suspend = (void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
360 .dtps_resume = (void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
361 .dtps_getargdesc = NULL,
362 .dtps_getargval = NULL,
363 .dtps_usermode = NULL,
364 .dtps_destroy = (void (*)(void *, dtrace_id_t, void *))dtrace_nullop,
365 };
366
367 static dtrace_id_t dtrace_probeid_begin; /* special BEGIN probe */
368 static dtrace_id_t dtrace_probeid_end; /* special END probe */
369 dtrace_id_t dtrace_probeid_error; /* special ERROR probe */
370
371 /*
372 * DTrace Helper Tracing Variables
373 *
374 * These variables should be set dynamically to enable helper tracing. The
375 * only variables that should be set are dtrace_helptrace_enable (which should
376 * be set to a non-zero value to allocate helper tracing buffers on the next
377 * open of /dev/dtrace) and dtrace_helptrace_disable (which should be set to a
378 * non-zero value to deallocate helper tracing buffers on the next close of
379 * /dev/dtrace). When (and only when) helper tracing is disabled, the
380 * buffer size may also be set via dtrace_helptrace_bufsize.
381 */
382 int dtrace_helptrace_enable = 0;
383 int dtrace_helptrace_disable = 0;
384 int dtrace_helptrace_bufsize = 16 * 1024 * 1024;
385 uint32_t dtrace_helptrace_nlocals;
386 static dtrace_helptrace_t *dtrace_helptrace_buffer;
387 static uint32_t dtrace_helptrace_next = 0;
388 static int dtrace_helptrace_wrapped = 0;
389
390 /*
391 * DTrace Error Hashing
392 *
393 * On DEBUG kernels, DTrace will track the errors that has seen in a hash
394 * table. This is very useful for checking coverage of tests that are
395 * expected to induce DIF or DOF processing errors, and may be useful for
396 * debugging problems in the DIF code generator or in DOF generation . The
397 * error hash may be examined with the ::dtrace_errhash MDB dcmd.
398 */
399 #ifdef DEBUG
400 static dtrace_errhash_t dtrace_errhash[DTRACE_ERRHASHSZ];
401 static const char *dtrace_errlast;
402 static kthread_t *dtrace_errthread;
403 static kmutex_t dtrace_errlock;
404 #endif
405
406 /*
407 * DTrace Macros and Constants
408 *
409 * These are various macros that are useful in various spots in the
410 * implementation, along with a few random constants that have no meaning
411 * outside of the implementation. There is no real structure to this cpp
412 * mishmash -- but is there ever?
413 */
414 #define DTRACE_HASHSTR(hash, probe) \
415 dtrace_hash_str(*((char **)((uintptr_t)(probe) + (hash)->dth_stroffs)))
416
417 #define DTRACE_HASHNEXT(hash, probe) \
418 (dtrace_probe_t **)((uintptr_t)(probe) + (hash)->dth_nextoffs)
419
420 #define DTRACE_HASHPREV(hash, probe) \
421 (dtrace_probe_t **)((uintptr_t)(probe) + (hash)->dth_prevoffs)
422
423 #define DTRACE_HASHEQ(hash, lhs, rhs) \
424 (strcmp(*((char **)((uintptr_t)(lhs) + (hash)->dth_stroffs)), \
425 *((char **)((uintptr_t)(rhs) + (hash)->dth_stroffs))) == 0)
426
427 #define DTRACE_AGGHASHSIZE_SLEW 17
428
429 #define DTRACE_V4MAPPED_OFFSET (sizeof (uint32_t) * 3)
430
431 /*
432 * The key for a thread-local variable consists of the lower 61 bits of the
433 * t_did, plus the 3 bits of the highest active interrupt above LOCK_LEVEL.
434 * We add DIF_VARIABLE_MAX to t_did to assure that the thread key is never
435 * equal to a variable identifier. This is necessary (but not sufficient) to
436 * assure that global associative arrays never collide with thread-local
437 * variables. To guarantee that they cannot collide, we must also define the
438 * order for keying dynamic variables. That order is:
439 *
440 * [ key0 ] ... [ keyn ] [ variable-key ] [ tls-key ]
441 *
442 * Because the variable-key and the tls-key are in orthogonal spaces, there is
443 * no way for a global variable key signature to match a thread-local key
444 * signature.
445 */
446 #ifdef illumos
447 #define DTRACE_TLS_THRKEY(where) { \
448 uint_t intr = 0; \
449 uint_t actv = CPU->cpu_intr_actv >> (LOCK_LEVEL + 1); \
450 for (; actv; actv >>= 1) \
451 intr++; \
452 ASSERT(intr < (1 << 3)); \
453 (where) = ((curthread->t_did + DIF_VARIABLE_MAX) & \
454 (((uint64_t)1 << 61) - 1)) | ((uint64_t)intr << 61); \
455 }
456 #else
457 #define DTRACE_TLS_THRKEY(where) { \
458 solaris_cpu_t *_c = &solaris_cpu[curcpu]; \
459 uint_t intr = 0; \
460 uint_t actv = _c->cpu_intr_actv; \
461 for (; actv; actv >>= 1) \
462 intr++; \
463 ASSERT(intr < (1 << 3)); \
464 (where) = ((curthread->td_tid + DIF_VARIABLE_MAX) & \
465 (((uint64_t)1 << 61) - 1)) | ((uint64_t)intr << 61); \
466 }
467 #endif
468
469 #define DT_BSWAP_8(x) ((x) & 0xff)
470 #define DT_BSWAP_16(x) ((DT_BSWAP_8(x) << 8) | DT_BSWAP_8((x) >> 8))
471 #define DT_BSWAP_32(x) ((DT_BSWAP_16(x) << 16) | DT_BSWAP_16((x) >> 16))
472 #define DT_BSWAP_64(x) ((DT_BSWAP_32(x) << 32) | DT_BSWAP_32((x) >> 32))
473
474 #define DT_MASK_LO 0x00000000FFFFFFFFULL
475
476 #define DTRACE_STORE(type, tomax, offset, what) \
477 *((type *)((uintptr_t)(tomax) + (uintptr_t)offset)) = (type)(what);
478
479 #if !defined(__x86) && !defined(__aarch64__)
480 #define DTRACE_ALIGNCHECK(addr, size, flags) \
481 if (addr & (size - 1)) { \
482 *flags |= CPU_DTRACE_BADALIGN; \
483 cpu_core[curcpu].cpuc_dtrace_illval = addr; \
484 return (0); \
485 }
486 #else
487 #define DTRACE_ALIGNCHECK(addr, size, flags)
488 #endif
489
490 /*
491 * Test whether a range of memory starting at testaddr of size testsz falls
492 * within the range of memory described by addr, sz. We take care to avoid
493 * problems with overflow and underflow of the unsigned quantities, and
494 * disallow all negative sizes. Ranges of size 0 are allowed.
495 */
496 #define DTRACE_INRANGE(testaddr, testsz, baseaddr, basesz) \
497 ((testaddr) - (uintptr_t)(baseaddr) < (basesz) && \
498 (testaddr) + (testsz) - (uintptr_t)(baseaddr) <= (basesz) && \
499 (testaddr) + (testsz) >= (testaddr))
500
501 #define DTRACE_RANGE_REMAIN(remp, addr, baseaddr, basesz) \
502 do { \
503 if ((remp) != NULL) { \
504 *(remp) = (uintptr_t)(baseaddr) + (basesz) - (addr); \
505 } \
506 } while (0)
507
508
509 /*
510 * Test whether alloc_sz bytes will fit in the scratch region. We isolate
511 * alloc_sz on the righthand side of the comparison in order to avoid overflow
512 * or underflow in the comparison with it. This is simpler than the INRANGE
513 * check above, because we know that the dtms_scratch_ptr is valid in the
514 * range. Allocations of size zero are allowed.
515 */
516 #define DTRACE_INSCRATCH(mstate, alloc_sz) \
517 ((mstate)->dtms_scratch_base + (mstate)->dtms_scratch_size - \
518 (mstate)->dtms_scratch_ptr >= (alloc_sz))
519
520 #define DTRACE_LOADFUNC(bits) \
521 /*CSTYLED*/ \
522 uint##bits##_t \
523 dtrace_load##bits(uintptr_t addr) \
524 { \
525 size_t size = bits / NBBY; \
526 /*CSTYLED*/ \
527 uint##bits##_t rval; \
528 int i; \
529 volatile uint16_t *flags = (volatile uint16_t *) \
530 &cpu_core[curcpu].cpuc_dtrace_flags; \
531 \
532 DTRACE_ALIGNCHECK(addr, size, flags); \
533 \
534 for (i = 0; i < dtrace_toxranges; i++) { \
535 if (addr >= dtrace_toxrange[i].dtt_limit) \
536 continue; \
537 \
538 if (addr + size <= dtrace_toxrange[i].dtt_base) \
539 continue; \
540 \
541 /* \
542 * This address falls within a toxic region; return 0. \
543 */ \
544 *flags |= CPU_DTRACE_BADADDR; \
545 cpu_core[curcpu].cpuc_dtrace_illval = addr; \
546 return (0); \
547 } \
548 \
549 *flags |= CPU_DTRACE_NOFAULT; \
550 /*CSTYLED*/ \
551 rval = *((volatile uint##bits##_t *)addr); \
552 *flags &= ~CPU_DTRACE_NOFAULT; \
553 \
554 return (!(*flags & CPU_DTRACE_FAULT) ? rval : 0); \
555 }
556
557 #ifdef _LP64
558 #define dtrace_loadptr dtrace_load64
559 #else
560 #define dtrace_loadptr dtrace_load32
561 #endif
562
563 #define DTRACE_DYNHASH_FREE 0
564 #define DTRACE_DYNHASH_SINK 1
565 #define DTRACE_DYNHASH_VALID 2
566
567 #define DTRACE_MATCH_NEXT 0
568 #define DTRACE_MATCH_DONE 1
569 #define DTRACE_ANCHORED(probe) ((probe)->dtpr_func[0] != '\0')
570 #define DTRACE_STATE_ALIGN 64
571
572 #define DTRACE_FLAGS2FLT(flags) \
573 (((flags) & CPU_DTRACE_BADADDR) ? DTRACEFLT_BADADDR : \
574 ((flags) & CPU_DTRACE_ILLOP) ? DTRACEFLT_ILLOP : \
575 ((flags) & CPU_DTRACE_DIVZERO) ? DTRACEFLT_DIVZERO : \
576 ((flags) & CPU_DTRACE_KPRIV) ? DTRACEFLT_KPRIV : \
577 ((flags) & CPU_DTRACE_UPRIV) ? DTRACEFLT_UPRIV : \
578 ((flags) & CPU_DTRACE_TUPOFLOW) ? DTRACEFLT_TUPOFLOW : \
579 ((flags) & CPU_DTRACE_BADALIGN) ? DTRACEFLT_BADALIGN : \
580 ((flags) & CPU_DTRACE_NOSCRATCH) ? DTRACEFLT_NOSCRATCH : \
581 ((flags) & CPU_DTRACE_BADSTACK) ? DTRACEFLT_BADSTACK : \
582 DTRACEFLT_UNKNOWN)
583
584 #define DTRACEACT_ISSTRING(act) \
585 ((act)->dta_kind == DTRACEACT_DIFEXPR && \
586 (act)->dta_difo->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING)
587
588 /* Function prototype definitions: */
589 static size_t dtrace_strlen(const char *, size_t);
590 static dtrace_probe_t *dtrace_probe_lookup_id(dtrace_id_t id);
591 static void dtrace_enabling_provide(dtrace_provider_t *);
592 static int dtrace_enabling_match(dtrace_enabling_t *, int *);
593 static void dtrace_enabling_matchall(void);
594 static void dtrace_enabling_reap(void);
595 static dtrace_state_t *dtrace_anon_grab(void);
596 static uint64_t dtrace_helper(int, dtrace_mstate_t *,
597 dtrace_state_t *, uint64_t, uint64_t);
598 static dtrace_helpers_t *dtrace_helpers_create(proc_t *);
599 static void dtrace_buffer_drop(dtrace_buffer_t *);
600 static int dtrace_buffer_consumed(dtrace_buffer_t *, hrtime_t when);
601 static intptr_t dtrace_buffer_reserve(dtrace_buffer_t *, size_t, size_t,
602 dtrace_state_t *, dtrace_mstate_t *);
603 static int dtrace_state_option(dtrace_state_t *, dtrace_optid_t,
604 dtrace_optval_t);
605 static int dtrace_ecb_create_enable(dtrace_probe_t *, void *);
606 static void dtrace_helper_provider_destroy(dtrace_helper_provider_t *);
607 uint16_t dtrace_load16(uintptr_t);
608 uint32_t dtrace_load32(uintptr_t);
609 uint64_t dtrace_load64(uintptr_t);
610 uint8_t dtrace_load8(uintptr_t);
611 void dtrace_dynvar_clean(dtrace_dstate_t *);
612 dtrace_dynvar_t *dtrace_dynvar(dtrace_dstate_t *, uint_t, dtrace_key_t *,
613 size_t, dtrace_dynvar_op_t, dtrace_mstate_t *, dtrace_vstate_t *);
614 uintptr_t dtrace_dif_varstr(uintptr_t, dtrace_state_t *, dtrace_mstate_t *);
615 static int dtrace_priv_proc(dtrace_state_t *);
616 static void dtrace_getf_barrier(void);
617 static int dtrace_canload_remains(uint64_t, size_t, size_t *,
618 dtrace_mstate_t *, dtrace_vstate_t *);
619 static int dtrace_canstore_remains(uint64_t, size_t, size_t *,
620 dtrace_mstate_t *, dtrace_vstate_t *);
621
622 /*
623 * DTrace Probe Context Functions
624 *
625 * These functions are called from probe context. Because probe context is
626 * any context in which C may be called, arbitrarily locks may be held,
627 * interrupts may be disabled, we may be in arbitrary dispatched state, etc.
628 * As a result, functions called from probe context may only call other DTrace
629 * support functions -- they may not interact at all with the system at large.
630 * (Note that the ASSERT macro is made probe-context safe by redefining it in
631 * terms of dtrace_assfail(), a probe-context safe function.) If arbitrary
632 * loads are to be performed from probe context, they _must_ be in terms of
633 * the safe dtrace_load*() variants.
634 *
635 * Some functions in this block are not actually called from probe context;
636 * for these functions, there will be a comment above the function reading
637 * "Note: not called from probe context."
638 */
639 void
640 dtrace_panic(const char *format, ...)
641 {
642 va_list alist;
643
644 va_start(alist, format);
645 #ifdef __FreeBSD__
646 vpanic(format, alist);
647 #else
648 dtrace_vpanic(format, alist);
649 #endif
650 va_end(alist);
651 }
652
653 int
654 dtrace_assfail(const char *a, const char *f, int l)
655 {
656 dtrace_panic("assertion failed: %s, file: %s, line: %d", a, f, l);
657
658 /*
659 * We just need something here that even the most clever compiler
660 * cannot optimize away.
661 */
662 return (a[(uintptr_t)f]);
663 }
664
665 /*
666 * Atomically increment a specified error counter from probe context.
667 */
668 static void
669 dtrace_error(uint32_t *counter)
670 {
671 /*
672 * Most counters stored to in probe context are per-CPU counters.
673 * However, there are some error conditions that are sufficiently
674 * arcane that they don't merit per-CPU storage. If these counters
675 * are incremented concurrently on different CPUs, scalability will be
676 * adversely affected -- but we don't expect them to be white-hot in a
677 * correctly constructed enabling...
678 */
679 uint32_t oval, nval;
680
681 do {
682 oval = *counter;
683
684 if ((nval = oval + 1) == 0) {
685 /*
686 * If the counter would wrap, set it to 1 -- assuring
687 * that the counter is never zero when we have seen
688 * errors. (The counter must be 32-bits because we
689 * aren't guaranteed a 64-bit compare&swap operation.)
690 * To save this code both the infamy of being fingered
691 * by a priggish news story and the indignity of being
692 * the target of a neo-puritan witch trial, we're
693 * carefully avoiding any colorful description of the
694 * likelihood of this condition -- but suffice it to
695 * say that it is only slightly more likely than the
696 * overflow of predicate cache IDs, as discussed in
697 * dtrace_predicate_create().
698 */
699 nval = 1;
700 }
701 } while (dtrace_cas32(counter, oval, nval) != oval);
702 }
703
704 /*
705 * Use the DTRACE_LOADFUNC macro to define functions for each of loading a
706 * uint8_t, a uint16_t, a uint32_t and a uint64_t.
707 */
708 /* BEGIN CSTYLED */
709 DTRACE_LOADFUNC(8)
710 DTRACE_LOADFUNC(16)
711 DTRACE_LOADFUNC(32)
712 DTRACE_LOADFUNC(64)
713 /* END CSTYLED */
714
715 static int
716 dtrace_inscratch(uintptr_t dest, size_t size, dtrace_mstate_t *mstate)
717 {
718 if (dest < mstate->dtms_scratch_base)
719 return (0);
720
721 if (dest + size < dest)
722 return (0);
723
724 if (dest + size > mstate->dtms_scratch_ptr)
725 return (0);
726
727 return (1);
728 }
729
730 static int
731 dtrace_canstore_statvar(uint64_t addr, size_t sz, size_t *remain,
732 dtrace_statvar_t **svars, int nsvars)
733 {
734 int i;
735 size_t maxglobalsize, maxlocalsize;
736
737 if (nsvars == 0)
738 return (0);
739
740 maxglobalsize = dtrace_statvar_maxsize + sizeof (uint64_t);
741 maxlocalsize = maxglobalsize * NCPU;
742
743 for (i = 0; i < nsvars; i++) {
744 dtrace_statvar_t *svar = svars[i];
745 uint8_t scope;
746 size_t size;
747
748 if (svar == NULL || (size = svar->dtsv_size) == 0)
749 continue;
750
751 scope = svar->dtsv_var.dtdv_scope;
752
753 /*
754 * We verify that our size is valid in the spirit of providing
755 * defense in depth: we want to prevent attackers from using
756 * DTrace to escalate an orthogonal kernel heap corruption bug
757 * into the ability to store to arbitrary locations in memory.
758 */
759 VERIFY((scope == DIFV_SCOPE_GLOBAL && size <= maxglobalsize) ||
760 (scope == DIFV_SCOPE_LOCAL && size <= maxlocalsize));
761
762 if (DTRACE_INRANGE(addr, sz, svar->dtsv_data,
763 svar->dtsv_size)) {
764 DTRACE_RANGE_REMAIN(remain, addr, svar->dtsv_data,
765 svar->dtsv_size);
766 return (1);
767 }
768 }
769
770 return (0);
771 }
772
773 /*
774 * Check to see if the address is within a memory region to which a store may
775 * be issued. This includes the DTrace scratch areas, and any DTrace variable
776 * region. The caller of dtrace_canstore() is responsible for performing any
777 * alignment checks that are needed before stores are actually executed.
778 */
779 static int
780 dtrace_canstore(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
781 dtrace_vstate_t *vstate)
782 {
783 return (dtrace_canstore_remains(addr, sz, NULL, mstate, vstate));
784 }
785
786 /*
787 * Implementation of dtrace_canstore which communicates the upper bound of the
788 * allowed memory region.
789 */
790 static int
791 dtrace_canstore_remains(uint64_t addr, size_t sz, size_t *remain,
792 dtrace_mstate_t *mstate, dtrace_vstate_t *vstate)
793 {
794 /*
795 * First, check to see if the address is in scratch space...
796 */
797 if (DTRACE_INRANGE(addr, sz, mstate->dtms_scratch_base,
798 mstate->dtms_scratch_size)) {
799 DTRACE_RANGE_REMAIN(remain, addr, mstate->dtms_scratch_base,
800 mstate->dtms_scratch_size);
801 return (1);
802 }
803
804 /*
805 * Now check to see if it's a dynamic variable. This check will pick
806 * up both thread-local variables and any global dynamically-allocated
807 * variables.
808 */
809 if (DTRACE_INRANGE(addr, sz, vstate->dtvs_dynvars.dtds_base,
810 vstate->dtvs_dynvars.dtds_size)) {
811 dtrace_dstate_t *dstate = &vstate->dtvs_dynvars;
812 uintptr_t base = (uintptr_t)dstate->dtds_base +
813 (dstate->dtds_hashsize * sizeof (dtrace_dynhash_t));
814 uintptr_t chunkoffs;
815 dtrace_dynvar_t *dvar;
816
817 /*
818 * Before we assume that we can store here, we need to make
819 * sure that it isn't in our metadata -- storing to our
820 * dynamic variable metadata would corrupt our state. For
821 * the range to not include any dynamic variable metadata,
822 * it must:
823 *
824 * (1) Start above the hash table that is at the base of
825 * the dynamic variable space
826 *
827 * (2) Have a starting chunk offset that is beyond the
828 * dtrace_dynvar_t that is at the base of every chunk
829 *
830 * (3) Not span a chunk boundary
831 *
832 * (4) Not be in the tuple space of a dynamic variable
833 *
834 */
835 if (addr < base)
836 return (0);
837
838 chunkoffs = (addr - base) % dstate->dtds_chunksize;
839
840 if (chunkoffs < sizeof (dtrace_dynvar_t))
841 return (0);
842
843 if (chunkoffs + sz > dstate->dtds_chunksize)
844 return (0);
845
846 dvar = (dtrace_dynvar_t *)((uintptr_t)addr - chunkoffs);
847
848 if (dvar->dtdv_hashval == DTRACE_DYNHASH_FREE)
849 return (0);
850
851 if (chunkoffs < sizeof (dtrace_dynvar_t) +
852 ((dvar->dtdv_tuple.dtt_nkeys - 1) * sizeof (dtrace_key_t)))
853 return (0);
854
855 DTRACE_RANGE_REMAIN(remain, addr, dvar, dstate->dtds_chunksize);
856 return (1);
857 }
858
859 /*
860 * Finally, check the static local and global variables. These checks
861 * take the longest, so we perform them last.
862 */
863 if (dtrace_canstore_statvar(addr, sz, remain,
864 vstate->dtvs_locals, vstate->dtvs_nlocals))
865 return (1);
866
867 if (dtrace_canstore_statvar(addr, sz, remain,
868 vstate->dtvs_globals, vstate->dtvs_nglobals))
869 return (1);
870
871 return (0);
872 }
873
874
875 /*
876 * Convenience routine to check to see if the address is within a memory
877 * region in which a load may be issued given the user's privilege level;
878 * if not, it sets the appropriate error flags and loads 'addr' into the
879 * illegal value slot.
880 *
881 * DTrace subroutines (DIF_SUBR_*) should use this helper to implement
882 * appropriate memory access protection.
883 */
884 static int
885 dtrace_canload(uint64_t addr, size_t sz, dtrace_mstate_t *mstate,
886 dtrace_vstate_t *vstate)
887 {
888 return (dtrace_canload_remains(addr, sz, NULL, mstate, vstate));
889 }
890
891 /*
892 * Implementation of dtrace_canload which communicates the uppoer bound of the
893 * allowed memory region.
894 */
895 static int
896 dtrace_canload_remains(uint64_t addr, size_t sz, size_t *remain,
897 dtrace_mstate_t *mstate, dtrace_vstate_t *vstate)
898 {
899 volatile uintptr_t *illval = &cpu_core[curcpu].cpuc_dtrace_illval;
900 file_t *fp;
901
902 /*
903 * If we hold the privilege to read from kernel memory, then
904 * everything is readable.
905 */
906 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0) {
907 DTRACE_RANGE_REMAIN(remain, addr, addr, sz);
908 return (1);
909 }
910
911 /*
912 * You can obviously read that which you can store.
913 */
914 if (dtrace_canstore_remains(addr, sz, remain, mstate, vstate))
915 return (1);
916
917 /*
918 * We're allowed to read from our own string table.
919 */
920 if (DTRACE_INRANGE(addr, sz, mstate->dtms_difo->dtdo_strtab,
921 mstate->dtms_difo->dtdo_strlen)) {
922 DTRACE_RANGE_REMAIN(remain, addr,
923 mstate->dtms_difo->dtdo_strtab,
924 mstate->dtms_difo->dtdo_strlen);
925 return (1);
926 }
927
928 if (vstate->dtvs_state != NULL &&
929 dtrace_priv_proc(vstate->dtvs_state)) {
930 proc_t *p;
931
932 /*
933 * When we have privileges to the current process, there are
934 * several context-related kernel structures that are safe to
935 * read, even absent the privilege to read from kernel memory.
936 * These reads are safe because these structures contain only
937 * state that (1) we're permitted to read, (2) is harmless or
938 * (3) contains pointers to additional kernel state that we're
939 * not permitted to read (and as such, do not present an
940 * opportunity for privilege escalation). Finally (and
941 * critically), because of the nature of their relation with
942 * the current thread context, the memory associated with these
943 * structures cannot change over the duration of probe context,
944 * and it is therefore impossible for this memory to be
945 * deallocated and reallocated as something else while it's
946 * being operated upon.
947 */
948 if (DTRACE_INRANGE(addr, sz, curthread, sizeof (kthread_t))) {
949 DTRACE_RANGE_REMAIN(remain, addr, curthread,
950 sizeof (kthread_t));
951 return (1);
952 }
953
954 if ((p = curthread->t_procp) != NULL && DTRACE_INRANGE(addr,
955 sz, curthread->t_procp, sizeof (proc_t))) {
956 DTRACE_RANGE_REMAIN(remain, addr, curthread->t_procp,
957 sizeof (proc_t));
958 return (1);
959 }
960
961 if (curthread->t_cred != NULL && DTRACE_INRANGE(addr, sz,
962 curthread->t_cred, sizeof (cred_t))) {
963 DTRACE_RANGE_REMAIN(remain, addr, curthread->t_cred,
964 sizeof (cred_t));
965 return (1);
966 }
967
968 #ifdef illumos
969 if (p != NULL && p->p_pidp != NULL && DTRACE_INRANGE(addr, sz,
970 &(p->p_pidp->pid_id), sizeof (pid_t))) {
971 DTRACE_RANGE_REMAIN(remain, addr, &(p->p_pidp->pid_id),
972 sizeof (pid_t));
973 return (1);
974 }
975
976 if (curthread->t_cpu != NULL && DTRACE_INRANGE(addr, sz,
977 curthread->t_cpu, offsetof(cpu_t, cpu_pause_thread))) {
978 DTRACE_RANGE_REMAIN(remain, addr, curthread->t_cpu,
979 offsetof(cpu_t, cpu_pause_thread));
980 return (1);
981 }
982 #endif
983 }
984
985 if ((fp = mstate->dtms_getf) != NULL) {
986 uintptr_t psz = sizeof (void *);
987 vnode_t *vp;
988 vnodeops_t *op;
989
990 /*
991 * When getf() returns a file_t, the enabling is implicitly
992 * granted the (transient) right to read the returned file_t
993 * as well as the v_path and v_op->vnop_name of the underlying
994 * vnode. These accesses are allowed after a successful
995 * getf() because the members that they refer to cannot change
996 * once set -- and the barrier logic in the kernel's closef()
997 * path assures that the file_t and its referenced vode_t
998 * cannot themselves be stale (that is, it impossible for
999 * either dtms_getf itself or its f_vnode member to reference
1000 * freed memory).
1001 */
1002 if (DTRACE_INRANGE(addr, sz, fp, sizeof (file_t))) {
1003 DTRACE_RANGE_REMAIN(remain, addr, fp, sizeof (file_t));
1004 return (1);
1005 }
1006
1007 if ((vp = fp->f_vnode) != NULL) {
1008 size_t slen;
1009 #ifdef illumos
1010 if (DTRACE_INRANGE(addr, sz, &vp->v_path, psz)) {
1011 DTRACE_RANGE_REMAIN(remain, addr, &vp->v_path,
1012 psz);
1013 return (1);
1014 }
1015 slen = strlen(vp->v_path) + 1;
1016 if (DTRACE_INRANGE(addr, sz, vp->v_path, slen)) {
1017 DTRACE_RANGE_REMAIN(remain, addr, vp->v_path,
1018 slen);
1019 return (1);
1020 }
1021 #endif
1022
1023 if (DTRACE_INRANGE(addr, sz, &vp->v_op, psz)) {
1024 DTRACE_RANGE_REMAIN(remain, addr, &vp->v_op,
1025 psz);
1026 return (1);
1027 }
1028
1029 #ifdef illumos
1030 if ((op = vp->v_op) != NULL &&
1031 DTRACE_INRANGE(addr, sz, &op->vnop_name, psz)) {
1032 DTRACE_RANGE_REMAIN(remain, addr,
1033 &op->vnop_name, psz);
1034 return (1);
1035 }
1036
1037 if (op != NULL && op->vnop_name != NULL &&
1038 DTRACE_INRANGE(addr, sz, op->vnop_name,
1039 (slen = strlen(op->vnop_name) + 1))) {
1040 DTRACE_RANGE_REMAIN(remain, addr,
1041 op->vnop_name, slen);
1042 return (1);
1043 }
1044 #endif
1045 }
1046 }
1047
1048 DTRACE_CPUFLAG_SET(CPU_DTRACE_KPRIV);
1049 *illval = addr;
1050 return (0);
1051 }
1052
1053 /*
1054 * Convenience routine to check to see if a given string is within a memory
1055 * region in which a load may be issued given the user's privilege level;
1056 * this exists so that we don't need to issue unnecessary dtrace_strlen()
1057 * calls in the event that the user has all privileges.
1058 */
1059 static int
1060 dtrace_strcanload(uint64_t addr, size_t sz, size_t *remain,
1061 dtrace_mstate_t *mstate, dtrace_vstate_t *vstate)
1062 {
1063 size_t rsize;
1064
1065 /*
1066 * If we hold the privilege to read from kernel memory, then
1067 * everything is readable.
1068 */
1069 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0) {
1070 DTRACE_RANGE_REMAIN(remain, addr, addr, sz);
1071 return (1);
1072 }
1073
1074 /*
1075 * Even if the caller is uninterested in querying the remaining valid
1076 * range, it is required to ensure that the access is allowed.
1077 */
1078 if (remain == NULL) {
1079 remain = &rsize;
1080 }
1081 if (dtrace_canload_remains(addr, 0, remain, mstate, vstate)) {
1082 size_t strsz;
1083 /*
1084 * Perform the strlen after determining the length of the
1085 * memory region which is accessible. This prevents timing
1086 * information from being used to find NULs in memory which is
1087 * not accessible to the caller.
1088 */
1089 strsz = 1 + dtrace_strlen((char *)(uintptr_t)addr,
1090 MIN(sz, *remain));
1091 if (strsz <= *remain) {
1092 return (1);
1093 }
1094 }
1095
1096 return (0);
1097 }
1098
1099 /*
1100 * Convenience routine to check to see if a given variable is within a memory
1101 * region in which a load may be issued given the user's privilege level.
1102 */
1103 static int
1104 dtrace_vcanload(void *src, dtrace_diftype_t *type, size_t *remain,
1105 dtrace_mstate_t *mstate, dtrace_vstate_t *vstate)
1106 {
1107 size_t sz;
1108 ASSERT(type->dtdt_flags & DIF_TF_BYREF);
1109
1110 /*
1111 * Calculate the max size before performing any checks since even
1112 * DTRACE_ACCESS_KERNEL-credentialed callers expect that this function
1113 * return the max length via 'remain'.
1114 */
1115 if (type->dtdt_kind == DIF_TYPE_STRING) {
1116 dtrace_state_t *state = vstate->dtvs_state;
1117
1118 if (state != NULL) {
1119 sz = state->dts_options[DTRACEOPT_STRSIZE];
1120 } else {
1121 /*
1122 * In helper context, we have a NULL state; fall back
1123 * to using the system-wide default for the string size
1124 * in this case.
1125 */
1126 sz = dtrace_strsize_default;
1127 }
1128 } else {
1129 sz = type->dtdt_size;
1130 }
1131
1132 /*
1133 * If we hold the privilege to read from kernel memory, then
1134 * everything is readable.
1135 */
1136 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0) {
1137 DTRACE_RANGE_REMAIN(remain, (uintptr_t)src, src, sz);
1138 return (1);
1139 }
1140
1141 if (type->dtdt_kind == DIF_TYPE_STRING) {
1142 return (dtrace_strcanload((uintptr_t)src, sz, remain, mstate,
1143 vstate));
1144 }
1145 return (dtrace_canload_remains((uintptr_t)src, sz, remain, mstate,
1146 vstate));
1147 }
1148
1149 /*
1150 * Convert a string to a signed integer using safe loads.
1151 *
1152 * NOTE: This function uses various macros from strtolctype.h to manipulate
1153 * digit values, etc -- these have all been checked to ensure they make
1154 * no additional function calls.
1155 */
1156 static int64_t
1157 dtrace_strtoll(char *input, int base, size_t limit)
1158 {
1159 uintptr_t pos = (uintptr_t)input;
1160 int64_t val = 0;
1161 int x;
1162 boolean_t neg = B_FALSE;
1163 char c, cc, ccc;
1164 uintptr_t end = pos + limit;
1165
1166 /*
1167 * Consume any whitespace preceding digits.
1168 */
1169 while ((c = dtrace_load8(pos)) == ' ' || c == '\t')
1170 pos++;
1171
1172 /*
1173 * Handle an explicit sign if one is present.
1174 */
1175 if (c == '-' || c == '+') {
1176 if (c == '-')
1177 neg = B_TRUE;
1178 c = dtrace_load8(++pos);
1179 }
1180
1181 /*
1182 * Check for an explicit hexadecimal prefix ("0x" or "0X") and skip it
1183 * if present.
1184 */
1185 if (base == 16 && c == '' && ((cc = dtrace_load8(pos + 1)) == 'x' ||
1186 cc == 'X') && isxdigit(ccc = dtrace_load8(pos + 2))) {
1187 pos += 2;
1188 c = ccc;
1189 }
1190
1191 /*
1192 * Read in contiguous digits until the first non-digit character.
1193 */
1194 for (; pos < end && c != '\0' && lisalnum(c) && (x = DIGIT(c)) < base;
1195 c = dtrace_load8(++pos))
1196 val = val * base + x;
1197
1198 return (neg ? -val : val);
1199 }
1200
1201 /*
1202 * Compare two strings using safe loads.
1203 */
1204 static int
1205 dtrace_strncmp(char *s1, char *s2, size_t limit)
1206 {
1207 uint8_t c1, c2;
1208 volatile uint16_t *flags;
1209
1210 if (s1 == s2 || limit == 0)
1211 return (0);
1212
1213 flags = (volatile uint16_t *)&cpu_core[curcpu].cpuc_dtrace_flags;
1214
1215 do {
1216 if (s1 == NULL) {
1217 c1 = '\0';
1218 } else {
1219 c1 = dtrace_load8((uintptr_t)s1++);
1220 }
1221
1222 if (s2 == NULL) {
1223 c2 = '\0';
1224 } else {
1225 c2 = dtrace_load8((uintptr_t)s2++);
1226 }
1227
1228 if (c1 != c2)
1229 return (c1 - c2);
1230 } while (--limit && c1 != '\0' && !(*flags & CPU_DTRACE_FAULT));
1231
1232 return (0);
1233 }
1234
1235 /*
1236 * Compute strlen(s) for a string using safe memory accesses. The additional
1237 * len parameter is used to specify a maximum length to ensure completion.
1238 */
1239 static size_t
1240 dtrace_strlen(const char *s, size_t lim)
1241 {
1242 uint_t len;
1243
1244 for (len = 0; len != lim; len++) {
1245 if (dtrace_load8((uintptr_t)s++) == '\0')
1246 break;
1247 }
1248
1249 return (len);
1250 }
1251
1252 /*
1253 * Check if an address falls within a toxic region.
1254 */
1255 static int
1256 dtrace_istoxic(uintptr_t kaddr, size_t size)
1257 {
1258 uintptr_t taddr, tsize;
1259 int i;
1260
1261 for (i = 0; i < dtrace_toxranges; i++) {
1262 taddr = dtrace_toxrange[i].dtt_base;
1263 tsize = dtrace_toxrange[i].dtt_limit - taddr;
1264
1265 if (kaddr - taddr < tsize) {
1266 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
1267 cpu_core[curcpu].cpuc_dtrace_illval = kaddr;
1268 return (1);
1269 }
1270
1271 if (taddr - kaddr < size) {
1272 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
1273 cpu_core[curcpu].cpuc_dtrace_illval = taddr;
1274 return (1);
1275 }
1276 }
1277
1278 return (0);
1279 }
1280
1281 /*
1282 * Copy src to dst using safe memory accesses. The src is assumed to be unsafe
1283 * memory specified by the DIF program. The dst is assumed to be safe memory
1284 * that we can store to directly because it is managed by DTrace. As with
1285 * standard bcopy, overlapping copies are handled properly.
1286 */
1287 static void
1288 dtrace_bcopy(const void *src, void *dst, size_t len)
1289 {
1290 if (len != 0) {
1291 uint8_t *s1 = dst;
1292 const uint8_t *s2 = src;
1293
1294 if (s1 <= s2) {
1295 do {
1296 *s1++ = dtrace_load8((uintptr_t)s2++);
1297 } while (--len != 0);
1298 } else {
1299 s2 += len;
1300 s1 += len;
1301
1302 do {
1303 *--s1 = dtrace_load8((uintptr_t)--s2);
1304 } while (--len != 0);
1305 }
1306 }
1307 }
1308
1309 /*
1310 * Copy src to dst using safe memory accesses, up to either the specified
1311 * length, or the point that a nul byte is encountered. The src is assumed to
1312 * be unsafe memory specified by the DIF program. The dst is assumed to be
1313 * safe memory that we can store to directly because it is managed by DTrace.
1314 * Unlike dtrace_bcopy(), overlapping regions are not handled.
1315 */
1316 static void
1317 dtrace_strcpy(const void *src, void *dst, size_t len)
1318 {
1319 if (len != 0) {
1320 uint8_t *s1 = dst, c;
1321 const uint8_t *s2 = src;
1322
1323 do {
1324 *s1++ = c = dtrace_load8((uintptr_t)s2++);
1325 } while (--len != 0 && c != '\0');
1326 }
1327 }
1328
1329 /*
1330 * Copy src to dst, deriving the size and type from the specified (BYREF)
1331 * variable type. The src is assumed to be unsafe memory specified by the DIF
1332 * program. The dst is assumed to be DTrace variable memory that is of the
1333 * specified type; we assume that we can store to directly.
1334 */
1335 static void
1336 dtrace_vcopy(void *src, void *dst, dtrace_diftype_t *type, size_t limit)
1337 {
1338 ASSERT(type->dtdt_flags & DIF_TF_BYREF);
1339
1340 if (type->dtdt_kind == DIF_TYPE_STRING) {
1341 dtrace_strcpy(src, dst, MIN(type->dtdt_size, limit));
1342 } else {
1343 dtrace_bcopy(src, dst, MIN(type->dtdt_size, limit));
1344 }
1345 }
1346
1347 /*
1348 * Compare s1 to s2 using safe memory accesses. The s1 data is assumed to be
1349 * unsafe memory specified by the DIF program. The s2 data is assumed to be
1350 * safe memory that we can access directly because it is managed by DTrace.
1351 */
1352 static int
1353 dtrace_bcmp(const void *s1, const void *s2, size_t len)
1354 {
1355 volatile uint16_t *flags;
1356
1357 flags = (volatile uint16_t *)&cpu_core[curcpu].cpuc_dtrace_flags;
1358
1359 if (s1 == s2)
1360 return (0);
1361
1362 if (s1 == NULL || s2 == NULL)
1363 return (1);
1364
1365 if (s1 != s2 && len != 0) {
1366 const uint8_t *ps1 = s1;
1367 const uint8_t *ps2 = s2;
1368
1369 do {
1370 if (dtrace_load8((uintptr_t)ps1++) != *ps2++)
1371 return (1);
1372 } while (--len != 0 && !(*flags & CPU_DTRACE_FAULT));
1373 }
1374 return (0);
1375 }
1376
1377 /*
1378 * Zero the specified region using a simple byte-by-byte loop. Note that this
1379 * is for safe DTrace-managed memory only.
1380 */
1381 static void
1382 dtrace_bzero(void *dst, size_t len)
1383 {
1384 uchar_t *cp;
1385
1386 for (cp = dst; len != 0; len--)
1387 *cp++ = 0;
1388 }
1389
1390 static void
1391 dtrace_add_128(uint64_t *addend1, uint64_t *addend2, uint64_t *sum)
1392 {
1393 uint64_t result[2];
1394
1395 result[0] = addend1[0] + addend2[0];
1396 result[1] = addend1[1] + addend2[1] +
1397 (result[0] < addend1[0] || result[0] < addend2[0] ? 1 : 0);
1398
1399 sum[0] = result[0];
1400 sum[1] = result[1];
1401 }
1402
1403 /*
1404 * Shift the 128-bit value in a by b. If b is positive, shift left.
1405 * If b is negative, shift right.
1406 */
1407 static void
1408 dtrace_shift_128(uint64_t *a, int b)
1409 {
1410 uint64_t mask;
1411
1412 if (b == 0)
1413 return;
1414
1415 if (b < 0) {
1416 b = -b;
1417 if (b >= 64) {
1418 a[0] = a[1] >> (b - 64);
1419 a[1] = 0;
1420 } else {
1421 a[0] >>= b;
1422 mask = 1LL << (64 - b);
1423 mask -= 1;
1424 a[0] |= ((a[1] & mask) << (64 - b));
1425 a[1] >>= b;
1426 }
1427 } else {
1428 if (b >= 64) {
1429 a[1] = a[0] << (b - 64);
1430 a[0] = 0;
1431 } else {
1432 a[1] <<= b;
1433 mask = a[0] >> (64 - b);
1434 a[1] |= mask;
1435 a[0] <<= b;
1436 }
1437 }
1438 }
1439
1440 /*
1441 * The basic idea is to break the 2 64-bit values into 4 32-bit values,
1442 * use native multiplication on those, and then re-combine into the
1443 * resulting 128-bit value.
1444 *
1445 * (hi1 << 32 + lo1) * (hi2 << 32 + lo2) =
1446 * hi1 * hi2 << 64 +
1447 * hi1 * lo2 << 32 +
1448 * hi2 * lo1 << 32 +
1449 * lo1 * lo2
1450 */
1451 static void
1452 dtrace_multiply_128(uint64_t factor1, uint64_t factor2, uint64_t *product)
1453 {
1454 uint64_t hi1, hi2, lo1, lo2;
1455 uint64_t tmp[2];
1456
1457 hi1 = factor1 >> 32;
1458 hi2 = factor2 >> 32;
1459
1460 lo1 = factor1 & DT_MASK_LO;
1461 lo2 = factor2 & DT_MASK_LO;
1462
1463 product[0] = lo1 * lo2;
1464 product[1] = hi1 * hi2;
1465
1466 tmp[0] = hi1 * lo2;
1467 tmp[1] = 0;
1468 dtrace_shift_128(tmp, 32);
1469 dtrace_add_128(product, tmp, product);
1470
1471 tmp[0] = hi2 * lo1;
1472 tmp[1] = 0;
1473 dtrace_shift_128(tmp, 32);
1474 dtrace_add_128(product, tmp, product);
1475 }
1476
1477 /*
1478 * This privilege check should be used by actions and subroutines to
1479 * verify that the user credentials of the process that enabled the
1480 * invoking ECB match the target credentials
1481 */
1482 static int
1483 dtrace_priv_proc_common_user(dtrace_state_t *state)
1484 {
1485 cred_t *cr, *s_cr = state->dts_cred.dcr_cred;
1486
1487 /*
1488 * We should always have a non-NULL state cred here, since if cred
1489 * is null (anonymous tracing), we fast-path bypass this routine.
1490 */
1491 ASSERT(s_cr != NULL);
1492
1493 if ((cr = CRED()) != NULL &&
1494 s_cr->cr_uid == cr->cr_uid &&
1495 s_cr->cr_uid == cr->cr_ruid &&
1496 s_cr->cr_uid == cr->cr_suid &&
1497 s_cr->cr_gid == cr->cr_gid &&
1498 s_cr->cr_gid == cr->cr_rgid &&
1499 s_cr->cr_gid == cr->cr_sgid)
1500 return (1);
1501
1502 return (0);
1503 }
1504
1505 /*
1506 * This privilege check should be used by actions and subroutines to
1507 * verify that the zone of the process that enabled the invoking ECB
1508 * matches the target credentials
1509 */
1510 static int
1511 dtrace_priv_proc_common_zone(dtrace_state_t *state)
1512 {
1513 #ifdef illumos
1514 cred_t *cr, *s_cr = state->dts_cred.dcr_cred;
1515
1516 /*
1517 * We should always have a non-NULL state cred here, since if cred
1518 * is null (anonymous tracing), we fast-path bypass this routine.
1519 */
1520 ASSERT(s_cr != NULL);
1521
1522 if ((cr = CRED()) != NULL && s_cr->cr_zone == cr->cr_zone)
1523 return (1);
1524
1525 return (0);
1526 #else
1527 return (1);
1528 #endif
1529 }
1530
1531 /*
1532 * This privilege check should be used by actions and subroutines to
1533 * verify that the process has not setuid or changed credentials.
1534 */
1535 static int
1536 dtrace_priv_proc_common_nocd(void)
1537 {
1538 proc_t *proc;
1539
1540 if ((proc = ttoproc(curthread)) != NULL &&
1541 !(proc->p_flag & SNOCD))
1542 return (1);
1543
1544 return (0);
1545 }
1546
1547 static int
1548 dtrace_priv_proc_destructive(dtrace_state_t *state)
1549 {
1550 int action = state->dts_cred.dcr_action;
1551
1552 if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE) == 0) &&
1553 dtrace_priv_proc_common_zone(state) == 0)
1554 goto bad;
1555
1556 if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER) == 0) &&
1557 dtrace_priv_proc_common_user(state) == 0)
1558 goto bad;
1559
1560 if (((action & DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG) == 0) &&
1561 dtrace_priv_proc_common_nocd() == 0)
1562 goto bad;
1563
1564 return (1);
1565
1566 bad:
1567 cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1568
1569 return (0);
1570 }
1571
1572 static int
1573 dtrace_priv_proc_control(dtrace_state_t *state)
1574 {
1575 if (state->dts_cred.dcr_action & DTRACE_CRA_PROC_CONTROL)
1576 return (1);
1577
1578 if (dtrace_priv_proc_common_zone(state) &&
1579 dtrace_priv_proc_common_user(state) &&
1580 dtrace_priv_proc_common_nocd())
1581 return (1);
1582
1583 cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1584
1585 return (0);
1586 }
1587
1588 static int
1589 dtrace_priv_proc(dtrace_state_t *state)
1590 {
1591 if (state->dts_cred.dcr_action & DTRACE_CRA_PROC)
1592 return (1);
1593
1594 cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_UPRIV;
1595
1596 return (0);
1597 }
1598
1599 static int
1600 dtrace_priv_kernel(dtrace_state_t *state)
1601 {
1602 if (state->dts_cred.dcr_action & DTRACE_CRA_KERNEL)
1603 return (1);
1604
1605 cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_KPRIV;
1606
1607 return (0);
1608 }
1609
1610 static int
1611 dtrace_priv_kernel_destructive(dtrace_state_t *state)
1612 {
1613 if (state->dts_cred.dcr_action & DTRACE_CRA_KERNEL_DESTRUCTIVE)
1614 return (1);
1615
1616 cpu_core[curcpu].cpuc_dtrace_flags |= CPU_DTRACE_KPRIV;
1617
1618 return (0);
1619 }
1620
1621 /*
1622 * Determine if the dte_cond of the specified ECB allows for processing of
1623 * the current probe to continue. Note that this routine may allow continued
1624 * processing, but with access(es) stripped from the mstate's dtms_access
1625 * field.
1626 */
1627 static int
1628 dtrace_priv_probe(dtrace_state_t *state, dtrace_mstate_t *mstate,
1629 dtrace_ecb_t *ecb)
1630 {
1631 dtrace_probe_t *probe = ecb->dte_probe;
1632 dtrace_provider_t *prov = probe->dtpr_provider;
1633 dtrace_pops_t *pops = &prov->dtpv_pops;
1634 int mode = DTRACE_MODE_NOPRIV_DROP;
1635
1636 ASSERT(ecb->dte_cond);
1637
1638 #ifdef illumos
1639 if (pops->dtps_mode != NULL) {
1640 mode = pops->dtps_mode(prov->dtpv_arg,
1641 probe->dtpr_id, probe->dtpr_arg);
1642
1643 ASSERT((mode & DTRACE_MODE_USER) ||
1644 (mode & DTRACE_MODE_KERNEL));
1645 ASSERT((mode & DTRACE_MODE_NOPRIV_RESTRICT) ||
1646 (mode & DTRACE_MODE_NOPRIV_DROP));
1647 }
1648
1649 /*
1650 * If the dte_cond bits indicate that this consumer is only allowed to
1651 * see user-mode firings of this probe, call the provider's dtps_mode()
1652 * entry point to check that the probe was fired while in a user
1653 * context. If that's not the case, use the policy specified by the
1654 * provider to determine if we drop the probe or merely restrict
1655 * operation.
1656 */
1657 if (ecb->dte_cond & DTRACE_COND_USERMODE) {
1658 ASSERT(mode != DTRACE_MODE_NOPRIV_DROP);
1659
1660 if (!(mode & DTRACE_MODE_USER)) {
1661 if (mode & DTRACE_MODE_NOPRIV_DROP)
1662 return (0);
1663
1664 mstate->dtms_access &= ~DTRACE_ACCESS_ARGS;
1665 }
1666 }
1667 #endif
1668
1669 /*
1670 * This is more subtle than it looks. We have to be absolutely certain
1671 * that CRED() isn't going to change out from under us so it's only
1672 * legit to examine that structure if we're in constrained situations.
1673 * Currently, the only times we'll this check is if a non-super-user
1674 * has enabled the profile or syscall providers -- providers that
1675 * allow visibility of all processes. For the profile case, the check
1676 * above will ensure that we're examining a user context.
1677 */
1678 if (ecb->dte_cond & DTRACE_COND_OWNER) {
1679 cred_t *cr;
1680 cred_t *s_cr = state->dts_cred.dcr_cred;
1681 proc_t *proc;
1682
1683 ASSERT(s_cr != NULL);
1684
1685 if ((cr = CRED()) == NULL ||
1686 s_cr->cr_uid != cr->cr_uid ||
1687 s_cr->cr_uid != cr->cr_ruid ||
1688 s_cr->cr_uid != cr->cr_suid ||
1689 s_cr->cr_gid != cr->cr_gid ||
1690 s_cr->cr_gid != cr->cr_rgid ||
1691 s_cr->cr_gid != cr->cr_sgid ||
1692 (proc = ttoproc(curthread)) == NULL ||
1693 (proc->p_flag & SNOCD)) {
1694 if (mode & DTRACE_MODE_NOPRIV_DROP)
1695 return (0);
1696
1697 #ifdef illumos
1698 mstate->dtms_access &= ~DTRACE_ACCESS_PROC;
1699 #endif
1700 }
1701 }
1702
1703 #ifdef illumos
1704 /*
1705 * If our dte_cond is set to DTRACE_COND_ZONEOWNER and we are not
1706 * in our zone, check to see if our mode policy is to restrict rather
1707 * than to drop; if to restrict, strip away both DTRACE_ACCESS_PROC
1708 * and DTRACE_ACCESS_ARGS
1709 */
1710 if (ecb->dte_cond & DTRACE_COND_ZONEOWNER) {
1711 cred_t *cr;
1712 cred_t *s_cr = state->dts_cred.dcr_cred;
1713
1714 ASSERT(s_cr != NULL);
1715
1716 if ((cr = CRED()) == NULL ||
1717 s_cr->cr_zone->zone_id != cr->cr_zone->zone_id) {
1718 if (mode & DTRACE_MODE_NOPRIV_DROP)
1719 return (0);
1720
1721 mstate->dtms_access &=
1722 ~(DTRACE_ACCESS_PROC | DTRACE_ACCESS_ARGS);
1723 }
1724 }
1725 #endif
1726
1727 return (1);
1728 }
1729
1730 /*
1731 * Note: not called from probe context. This function is called
1732 * asynchronously (and at a regular interval) from outside of probe context to
1733 * clean the dirty dynamic variable lists on all CPUs. Dynamic variable
1734 * cleaning is explained in detail in <sys/dtrace_impl.h>.
1735 */
1736 void
1737 dtrace_dynvar_clean(dtrace_dstate_t *dstate)
1738 {
1739 dtrace_dynvar_t *dirty;
1740 dtrace_dstate_percpu_t *dcpu;
1741 dtrace_dynvar_t **rinsep;
1742 int i, j, work = 0;
1743
1744 for (i = 0; i < NCPU; i++) {
1745 dcpu = &dstate->dtds_percpu[i];
1746 rinsep = &dcpu->dtdsc_rinsing;
1747
1748 /*
1749 * If the dirty list is NULL, there is no dirty work to do.
1750 */
1751 if (dcpu->dtdsc_dirty == NULL)
1752 continue;
1753
1754 if (dcpu->dtdsc_rinsing != NULL) {
1755 /*
1756 * If the rinsing list is non-NULL, then it is because
1757 * this CPU was selected to accept another CPU's
1758 * dirty list -- and since that time, dirty buffers
1759 * have accumulated. This is a highly unlikely
1760 * condition, but we choose to ignore the dirty
1761 * buffers -- they'll be picked up a future cleanse.
1762 */
1763 continue;
1764 }
1765
1766 if (dcpu->dtdsc_clean != NULL) {
1767 /*
1768 * If the clean list is non-NULL, then we're in a
1769 * situation where a CPU has done deallocations (we
1770 * have a non-NULL dirty list) but no allocations (we
1771 * also have a non-NULL clean list). We can't simply
1772 * move the dirty list into the clean list on this
1773 * CPU, yet we also don't want to allow this condition
1774 * to persist, lest a short clean list prevent a
1775 * massive dirty list from being cleaned (which in
1776 * turn could lead to otherwise avoidable dynamic
1777 * drops). To deal with this, we look for some CPU
1778 * with a NULL clean list, NULL dirty list, and NULL
1779 * rinsing list -- and then we borrow this CPU to
1780 * rinse our dirty list.
1781 */
1782 for (j = 0; j < NCPU; j++) {
1783 dtrace_dstate_percpu_t *rinser;
1784
1785 rinser = &dstate->dtds_percpu[j];
1786
1787 if (rinser->dtdsc_rinsing != NULL)
1788 continue;
1789
1790 if (rinser->dtdsc_dirty != NULL)
1791 continue;
1792
1793 if (rinser->dtdsc_clean != NULL)
1794 continue;
1795
1796 rinsep = &rinser->dtdsc_rinsing;
1797 break;
1798 }
1799
1800 if (j == NCPU) {
1801 /*
1802 * We were unable to find another CPU that
1803 * could accept this dirty list -- we are
1804 * therefore unable to clean it now.
1805 */
1806 dtrace_dynvar_failclean++;
1807 continue;
1808 }
1809 }
1810
1811 work = 1;
1812
1813 /*
1814 * Atomically move the dirty list aside.
1815 */
1816 do {
1817 dirty = dcpu->dtdsc_dirty;
1818
1819 /*
1820 * Before we zap the dirty list, set the rinsing list.
1821 * (This allows for a potential assertion in
1822 * dtrace_dynvar(): if a free dynamic variable appears
1823 * on a hash chain, either the dirty list or the
1824 * rinsing list for some CPU must be non-NULL.)
1825 */
1826 *rinsep = dirty;
1827 dtrace_membar_producer();
1828 } while (dtrace_casptr(&dcpu->dtdsc_dirty,
1829 dirty, NULL) != dirty);
1830 }
1831
1832 if (!work) {
1833 /*
1834 * We have no work to do; we can simply return.
1835 */
1836 return;
1837 }
1838
1839 dtrace_sync();
1840
1841 for (i = 0; i < NCPU; i++) {
1842 dcpu = &dstate->dtds_percpu[i];
1843
1844 if (dcpu->dtdsc_rinsing == NULL)
1845 continue;
1846
1847 /*
1848 * We are now guaranteed that no hash chain contains a pointer
1849 * into this dirty list; we can make it clean.
1850 */
1851 ASSERT(dcpu->dtdsc_clean == NULL);
1852 dcpu->dtdsc_clean = dcpu->dtdsc_rinsing;
1853 dcpu->dtdsc_rinsing = NULL;
1854 }
1855
1856 /*
1857 * Before we actually set the state to be DTRACE_DSTATE_CLEAN, make
1858 * sure that all CPUs have seen all of the dtdsc_clean pointers.
1859 * This prevents a race whereby a CPU incorrectly decides that
1860 * the state should be something other than DTRACE_DSTATE_CLEAN
1861 * after dtrace_dynvar_clean() has completed.
1862 */
1863 dtrace_sync();
1864
1865 dstate->dtds_state = DTRACE_DSTATE_CLEAN;
1866 }
1867
1868 /*
1869 * Depending on the value of the op parameter, this function looks-up,
1870 * allocates or deallocates an arbitrarily-keyed dynamic variable. If an
1871 * allocation is requested, this function will return a pointer to a
1872 * dtrace_dynvar_t corresponding to the allocated variable -- or NULL if no
1873 * variable can be allocated. If NULL is returned, the appropriate counter
1874 * will be incremented.
1875 */
1876 dtrace_dynvar_t *
1877 dtrace_dynvar(dtrace_dstate_t *dstate, uint_t nkeys,
1878 dtrace_key_t *key, size_t dsize, dtrace_dynvar_op_t op,
1879 dtrace_mstate_t *mstate, dtrace_vstate_t *vstate)
1880 {
1881 uint64_t hashval = DTRACE_DYNHASH_VALID;
1882 dtrace_dynhash_t *hash = dstate->dtds_hash;
1883 dtrace_dynvar_t *free, *new_free, *next, *dvar, *start, *prev = NULL;
1884 processorid_t me = curcpu, cpu = me;
1885 dtrace_dstate_percpu_t *dcpu = &dstate->dtds_percpu[me];
1886 size_t bucket, ksize;
1887 size_t chunksize = dstate->dtds_chunksize;
1888 uintptr_t kdata, lock, nstate;
1889 uint_t i;
1890
1891 ASSERT(nkeys != 0);
1892
1893 /*
1894 * Hash the key. As with aggregations, we use Jenkins' "One-at-a-time"
1895 * algorithm. For the by-value portions, we perform the algorithm in
1896 * 16-bit chunks (as opposed to 8-bit chunks). This speeds things up a
1897 * bit, and seems to have only a minute effect on distribution. For
1898 * the by-reference data, we perform "One-at-a-time" iterating (safely)
1899 * over each referenced byte. It's painful to do this, but it's much
1900 * better than pathological hash distribution. The efficacy of the
1901 * hashing algorithm (and a comparison with other algorithms) may be
1902 * found by running the ::dtrace_dynstat MDB dcmd.
1903 */
1904 for (i = 0; i < nkeys; i++) {
1905 if (key[i].dttk_size == 0) {
1906 uint64_t val = key[i].dttk_value;
1907
1908 hashval += (val >> 48) & 0xffff;
1909 hashval += (hashval << 10);
1910 hashval ^= (hashval >> 6);
1911
1912 hashval += (val >> 32) & 0xffff;
1913 hashval += (hashval << 10);
1914 hashval ^= (hashval >> 6);
1915
1916 hashval += (val >> 16) & 0xffff;
1917 hashval += (hashval << 10);
1918 hashval ^= (hashval >> 6);
1919
1920 hashval += val & 0xffff;
1921 hashval += (hashval << 10);
1922 hashval ^= (hashval >> 6);
1923 } else {
1924 /*
1925 * This is incredibly painful, but it beats the hell
1926 * out of the alternative.
1927 */
1928 uint64_t j, size = key[i].dttk_size;
1929 uintptr_t base = (uintptr_t)key[i].dttk_value;
1930
1931 if (!dtrace_canload(base, size, mstate, vstate))
1932 break;
1933
1934 for (j = 0; j < size; j++) {
1935 hashval += dtrace_load8(base + j);
1936 hashval += (hashval << 10);
1937 hashval ^= (hashval >> 6);
1938 }
1939 }
1940 }
1941
1942 if (DTRACE_CPUFLAG_ISSET(CPU_DTRACE_FAULT))
1943 return (NULL);
1944
1945 hashval += (hashval << 3);
1946 hashval ^= (hashval >> 11);
1947 hashval += (hashval << 15);
1948
1949 /*
1950 * There is a remote chance (ideally, 1 in 2^31) that our hashval
1951 * comes out to be one of our two sentinel hash values. If this
1952 * actually happens, we set the hashval to be a value known to be a
1953 * non-sentinel value.
1954 */
1955 if (hashval == DTRACE_DYNHASH_FREE || hashval == DTRACE_DYNHASH_SINK)
1956 hashval = DTRACE_DYNHASH_VALID;
1957
1958 /*
1959 * Yes, it's painful to do a divide here. If the cycle count becomes
1960 * important here, tricks can be pulled to reduce it. (However, it's
1961 * critical that hash collisions be kept to an absolute minimum;
1962 * they're much more painful than a divide.) It's better to have a
1963 * solution that generates few collisions and still keeps things
1964 * relatively simple.
1965 */
1966 bucket = hashval % dstate->dtds_hashsize;
1967
1968 if (op == DTRACE_DYNVAR_DEALLOC) {
1969 volatile uintptr_t *lockp = &hash[bucket].dtdh_lock;
1970
1971 for (;;) {
1972 while ((lock = *lockp) & 1)
1973 continue;
1974
1975 if (dtrace_casptr((volatile void *)lockp,
1976 (volatile void *)lock, (volatile void *)(lock + 1)) == (void *)lock)
1977 break;
1978 }
1979
1980 dtrace_membar_producer();
1981 }
1982
1983 top:
1984 prev = NULL;
1985 lock = hash[bucket].dtdh_lock;
1986
1987 dtrace_membar_consumer();
1988
1989 start = hash[bucket].dtdh_chain;
1990 ASSERT(start != NULL && (start->dtdv_hashval == DTRACE_DYNHASH_SINK ||
1991 start->dtdv_hashval != DTRACE_DYNHASH_FREE ||
1992 op != DTRACE_DYNVAR_DEALLOC));
1993
1994 for (dvar = start; dvar != NULL; dvar = dvar->dtdv_next) {
1995 dtrace_tuple_t *dtuple = &dvar->dtdv_tuple;
1996 dtrace_key_t *dkey = &dtuple->dtt_key[0];
1997
1998 if (dvar->dtdv_hashval != hashval) {
1999 if (dvar->dtdv_hashval == DTRACE_DYNHASH_SINK) {
2000 /*
2001 * We've reached the sink, and therefore the
2002 * end of the hash chain; we can kick out of
2003 * the loop knowing that we have seen a valid
2004 * snapshot of state.
2005 */
2006 ASSERT(dvar->dtdv_next == NULL);
2007 ASSERT(dvar == &dtrace_dynhash_sink);
2008 break;
2009 }
2010
2011 if (dvar->dtdv_hashval == DTRACE_DYNHASH_FREE) {
2012 /*
2013 * We've gone off the rails: somewhere along
2014 * the line, one of the members of this hash
2015 * chain was deleted. Note that we could also
2016 * detect this by simply letting this loop run
2017 * to completion, as we would eventually hit
2018 * the end of the dirty list. However, we
2019 * want to avoid running the length of the
2020 * dirty list unnecessarily (it might be quite
2021 * long), so we catch this as early as
2022 * possible by detecting the hash marker. In
2023 * this case, we simply set dvar to NULL and
2024 * break; the conditional after the loop will
2025 * send us back to top.
2026 */
2027 dvar = NULL;
2028 break;
2029 }
2030
2031 goto next;
2032 }
2033
2034 if (dtuple->dtt_nkeys != nkeys)
2035 goto next;
2036
2037 for (i = 0; i < nkeys; i++, dkey++) {
2038 if (dkey->dttk_size != key[i].dttk_size)
2039 goto next; /* size or type mismatch */
2040
2041 if (dkey->dttk_size != 0) {
2042 if (dtrace_bcmp(
2043 (void *)(uintptr_t)key[i].dttk_value,
2044 (void *)(uintptr_t)dkey->dttk_value,
2045 dkey->dttk_size))
2046 goto next;
2047 } else {
2048 if (dkey->dttk_value != key[i].dttk_value)
2049 goto next;
2050 }
2051 }
2052
2053 if (op != DTRACE_DYNVAR_DEALLOC)
2054 return (dvar);
2055
2056 ASSERT(dvar->dtdv_next == NULL ||
2057 dvar->dtdv_next->dtdv_hashval != DTRACE_DYNHASH_FREE);
2058
2059 if (prev != NULL) {
2060 ASSERT(hash[bucket].dtdh_chain != dvar);
2061 ASSERT(start != dvar);
2062 ASSERT(prev->dtdv_next == dvar);
2063 prev->dtdv_next = dvar->dtdv_next;
2064 } else {
2065 if (dtrace_casptr(&hash[bucket].dtdh_chain,
2066 start, dvar->dtdv_next) != start) {
2067 /*
2068 * We have failed to atomically swing the
2069 * hash table head pointer, presumably because
2070 * of a conflicting allocation on another CPU.
2071 * We need to reread the hash chain and try
2072 * again.
2073 */
2074 goto top;
2075 }
2076 }
2077
2078 dtrace_membar_producer();
2079
2080 /*
2081 * Now set the hash value to indicate that it's free.
2082 */
2083 ASSERT(hash[bucket].dtdh_chain != dvar);
2084 dvar->dtdv_hashval = DTRACE_DYNHASH_FREE;
2085
2086 dtrace_membar_producer();
2087
2088 /*
2089 * Set the next pointer to point at the dirty list, and
2090 * atomically swing the dirty pointer to the newly freed dvar.
2091 */
2092 do {
2093 next = dcpu->dtdsc_dirty;
2094 dvar->dtdv_next = next;
2095 } while (dtrace_casptr(&dcpu->dtdsc_dirty, next, dvar) != next);
2096
2097 /*
2098 * Finally, unlock this hash bucket.
2099 */
2100 ASSERT(hash[bucket].dtdh_lock == lock);
2101 ASSERT(lock & 1);
2102 hash[bucket].dtdh_lock++;
2103
2104 return (NULL);
2105 next:
2106 prev = dvar;
2107 continue;
2108 }
2109
2110 if (dvar == NULL) {
2111 /*
2112 * If dvar is NULL, it is because we went off the rails:
2113 * one of the elements that we traversed in the hash chain
2114 * was deleted while we were traversing it. In this case,
2115 * we assert that we aren't doing a dealloc (deallocs lock
2116 * the hash bucket to prevent themselves from racing with
2117 * one another), and retry the hash chain traversal.
2118 */
2119 ASSERT(op != DTRACE_DYNVAR_DEALLOC);
2120 goto top;
2121 }
2122
2123 if (op != DTRACE_DYNVAR_ALLOC) {
2124 /*
2125 * If we are not to allocate a new variable, we want to
2126 * return NULL now. Before we return, check that the value
2127 * of the lock word hasn't changed. If it has, we may have
2128 * seen an inconsistent snapshot.
2129 */
2130 if (op == DTRACE_DYNVAR_NOALLOC) {
2131 if (hash[bucket].dtdh_lock != lock)
2132 goto top;
2133 } else {
2134 ASSERT(op == DTRACE_DYNVAR_DEALLOC);
2135 ASSERT(hash[bucket].dtdh_lock == lock);
2136 ASSERT(lock & 1);
2137 hash[bucket].dtdh_lock++;
2138 }
2139
2140 return (NULL);
2141 }
2142
2143 /*
2144 * We need to allocate a new dynamic variable. The size we need is the
2145 * size of dtrace_dynvar plus the size of nkeys dtrace_key_t's plus the
2146 * size of any auxiliary key data (rounded up to 8-byte alignment) plus
2147 * the size of any referred-to data (dsize). We then round the final
2148 * size up to the chunksize for allocation.
2149 */
2150 for (ksize = 0, i = 0; i < nkeys; i++)
2151 ksize += P2ROUNDUP(key[i].dttk_size, sizeof (uint64_t));
2152
2153 /*
2154 * This should be pretty much impossible, but could happen if, say,
2155 * strange DIF specified the tuple. Ideally, this should be an
2156 * assertion and not an error condition -- but that requires that the
2157 * chunksize calculation in dtrace_difo_chunksize() be absolutely
2158 * bullet-proof. (That is, it must not be able to be fooled by
2159 * malicious DIF.) Given the lack of backwards branches in DIF,
2160 * solving this would presumably not amount to solving the Halting
2161 * Problem -- but it still seems awfully hard.
2162 */
2163 if (sizeof (dtrace_dynvar_t) + sizeof (dtrace_key_t) * (nkeys - 1) +
2164 ksize + dsize > chunksize) {
2165 dcpu->dtdsc_drops++;
2166 return (NULL);
2167 }
2168
2169 nstate = DTRACE_DSTATE_EMPTY;
2170
2171 do {
2172 retry:
2173 free = dcpu->dtdsc_free;
2174
2175 if (free == NULL) {
2176 dtrace_dynvar_t *clean = dcpu->dtdsc_clean;
2177 void *rval;
2178
2179 if (clean == NULL) {
2180 /*
2181 * We're out of dynamic variable space on
2182 * this CPU. Unless we have tried all CPUs,
2183 * we'll try to allocate from a different
2184 * CPU.
2185 */
2186 switch (dstate->dtds_state) {
2187 case DTRACE_DSTATE_CLEAN: {
2188 void *sp = &dstate->dtds_state;
2189
2190 if (++cpu >= NCPU)
2191 cpu = 0;
2192
2193 if (dcpu->dtdsc_dirty != NULL &&
2194 nstate == DTRACE_DSTATE_EMPTY)
2195 nstate = DTRACE_DSTATE_DIRTY;
2196
2197 if (dcpu->dtdsc_rinsing != NULL)
2198 nstate = DTRACE_DSTATE_RINSING;
2199
2200 dcpu = &dstate->dtds_percpu[cpu];
2201
2202 if (cpu != me)
2203 goto retry;
2204
2205 (void) dtrace_cas32(sp,
2206 DTRACE_DSTATE_CLEAN, nstate);
2207
2208 /*
2209 * To increment the correct bean
2210 * counter, take another lap.
2211 */
2212 goto retry;
2213 }
2214
2215 case DTRACE_DSTATE_DIRTY:
2216 dcpu->dtdsc_dirty_drops++;
2217 break;
2218
2219 case DTRACE_DSTATE_RINSING:
2220 dcpu->dtdsc_rinsing_drops++;
2221 break;
2222
2223 case DTRACE_DSTATE_EMPTY:
2224 dcpu->dtdsc_drops++;
2225 break;
2226 }
2227
2228 DTRACE_CPUFLAG_SET(CPU_DTRACE_DROP);
2229 return (NULL);
2230 }
2231
2232 /*
2233 * The clean list appears to be non-empty. We want to
2234 * move the clean list to the free list; we start by
2235 * moving the clean pointer aside.
2236 */
2237 if (dtrace_casptr(&dcpu->dtdsc_clean,
2238 clean, NULL) != clean) {
2239 /*
2240 * We are in one of two situations:
2241 *
2242 * (a) The clean list was switched to the
2243 * free list by another CPU.
2244 *
2245 * (b) The clean list was added to by the
2246 * cleansing cyclic.
2247 *
2248 * In either of these situations, we can
2249 * just reattempt the free list allocation.
2250 */
2251 goto retry;
2252 }
2253
2254 ASSERT(clean->dtdv_hashval == DTRACE_DYNHASH_FREE);
2255
2256 /*
2257 * Now we'll move the clean list to our free list.
2258 * It's impossible for this to fail: the only way
2259 * the free list can be updated is through this
2260 * code path, and only one CPU can own the clean list.
2261 * Thus, it would only be possible for this to fail if
2262 * this code were racing with dtrace_dynvar_clean().
2263 * (That is, if dtrace_dynvar_clean() updated the clean
2264 * list, and we ended up racing to update the free
2265 * list.) This race is prevented by the dtrace_sync()
2266 * in dtrace_dynvar_clean() -- which flushes the
2267 * owners of the clean lists out before resetting
2268 * the clean lists.
2269 */
2270 dcpu = &dstate->dtds_percpu[me];
2271 rval = dtrace_casptr(&dcpu->dtdsc_free, NULL, clean);
2272 ASSERT(rval == NULL);
2273 goto retry;
2274 }
2275
2276 dvar = free;
2277 new_free = dvar->dtdv_next;
2278 } while (dtrace_casptr(&dcpu->dtdsc_free, free, new_free) != free);
2279
2280 /*
2281 * We have now allocated a new chunk. We copy the tuple keys into the
2282 * tuple array and copy any referenced key data into the data space
2283 * following the tuple array. As we do this, we relocate dttk_value
2284 * in the final tuple to point to the key data address in the chunk.
2285 */
2286 kdata = (uintptr_t)&dvar->dtdv_tuple.dtt_key[nkeys];
2287 dvar->dtdv_data = (void *)(kdata + ksize);
2288 dvar->dtdv_tuple.dtt_nkeys = nkeys;
2289
2290 for (i = 0; i < nkeys; i++) {
2291 dtrace_key_t *dkey = &dvar->dtdv_tuple.dtt_key[i];
2292 size_t kesize = key[i].dttk_size;
2293
2294 if (kesize != 0) {
2295 dtrace_bcopy(
2296 (const void *)(uintptr_t)key[i].dttk_value,
2297 (void *)kdata, kesize);
2298 dkey->dttk_value = kdata;
2299 kdata += P2ROUNDUP(kesize, sizeof (uint64_t));
2300 } else {
2301 dkey->dttk_value = key[i].dttk_value;
2302 }
2303
2304 dkey->dttk_size = kesize;
2305 }
2306
2307 ASSERT(dvar->dtdv_hashval == DTRACE_DYNHASH_FREE);
2308 dvar->dtdv_hashval = hashval;
2309 dvar->dtdv_next = start;
2310
2311 if (dtrace_casptr(&hash[bucket].dtdh_chain, start, dvar) == start)
2312 return (dvar);
2313
2314 /*
2315 * The cas has failed. Either another CPU is adding an element to
2316 * this hash chain, or another CPU is deleting an element from this
2317 * hash chain. The simplest way to deal with both of these cases
2318 * (though not necessarily the most efficient) is to free our
2319 * allocated block and re-attempt it all. Note that the free is
2320 * to the dirty list and _not_ to the free list. This is to prevent
2321 * races with allocators, above.
2322 */
2323 dvar->dtdv_hashval = DTRACE_DYNHASH_FREE;
2324
2325 dtrace_membar_producer();
2326
2327 do {
2328 free = dcpu->dtdsc_dirty;
2329 dvar->dtdv_next = free;
2330 } while (dtrace_casptr(&dcpu->dtdsc_dirty, free, dvar) != free);
2331
2332 goto top;
2333 }
2334
2335 /*ARGSUSED*/
2336 static void
2337 dtrace_aggregate_min(uint64_t *oval, uint64_t nval, uint64_t arg)
2338 {
2339 if ((int64_t)nval < (int64_t)*oval)
2340 *oval = nval;
2341 }
2342
2343 /*ARGSUSED*/
2344 static void
2345 dtrace_aggregate_max(uint64_t *oval, uint64_t nval, uint64_t arg)
2346 {
2347 if ((int64_t)nval > (int64_t)*oval)
2348 *oval = nval;
2349 }
2350
2351 static void
2352 dtrace_aggregate_quantize(uint64_t *quanta, uint64_t nval, uint64_t incr)
2353 {
2354 int i, zero = DTRACE_QUANTIZE_ZEROBUCKET;
2355 int64_t val = (int64_t)nval;
2356
2357 if (val < 0) {
2358 for (i = 0; i < zero; i++) {
2359 if (val <= DTRACE_QUANTIZE_BUCKETVAL(i)) {
2360 quanta[i] += incr;
2361 return;
2362 }
2363 }
2364 } else {
2365 for (i = zero + 1; i < DTRACE_QUANTIZE_NBUCKETS; i++) {
2366 if (val < DTRACE_QUANTIZE_BUCKETVAL(i)) {
2367 quanta[i - 1] += incr;
2368 return;
2369 }
2370 }
2371
2372 quanta[DTRACE_QUANTIZE_NBUCKETS - 1] += incr;
2373 return;
2374 }
2375
2376 ASSERT(0);
2377 }
2378
2379 static void
2380 dtrace_aggregate_lquantize(uint64_t *lquanta, uint64_t nval, uint64_t incr)
2381 {
2382 uint64_t arg = *lquanta++;
2383 int32_t base = DTRACE_LQUANTIZE_BASE(arg);
2384 uint16_t step = DTRACE_LQUANTIZE_STEP(arg);
2385 uint16_t levels = DTRACE_LQUANTIZE_LEVELS(arg);
2386 int32_t val = (int32_t)nval, level;
2387
2388 ASSERT(step != 0);
2389 ASSERT(levels != 0);
2390
2391 if (val < base) {
2392 /*
2393 * This is an underflow.
2394 */
2395 lquanta[0] += incr;
2396 return;
2397 }
2398
2399 level = (val - base) / step;
2400
2401 if (level < levels) {
2402 lquanta[level + 1] += incr;
2403 return;
2404 }
2405
2406 /*
2407 * This is an overflow.
2408 */
2409 lquanta[levels + 1] += incr;
2410 }
2411
2412 static int
2413 dtrace_aggregate_llquantize_bucket(uint16_t factor, uint16_t low,
2414 uint16_t high, uint16_t nsteps, int64_t value)
2415 {
2416 int64_t this = 1, last, next;
2417 int base = 1, order;
2418
2419 ASSERT(factor <= nsteps);
2420 ASSERT(nsteps % factor == 0);
2421
2422 for (order = 0; order < low; order++)
2423 this *= factor;
2424
2425 /*
2426 * If our value is less than our factor taken to the power of the
2427 * low order of magnitude, it goes into the zeroth bucket.
2428 */
2429 if (value < (last = this))
2430 return (0);
2431
2432 for (this *= factor; order <= high; order++) {
2433 int nbuckets = this > nsteps ? nsteps : this;
2434
2435 if ((next = this * factor) < this) {
2436 /*
2437 * We should not generally get log/linear quantizations
2438 * with a high magnitude that allows 64-bits to
2439 * overflow, but we nonetheless protect against this
2440 * by explicitly checking for overflow, and clamping
2441 * our value accordingly.
2442 */
2443 value = this - 1;
2444 }
2445
2446 if (value < this) {
2447 /*
2448 * If our value lies within this order of magnitude,
2449 * determine its position by taking the offset within
2450 * the order of magnitude, dividing by the bucket
2451 * width, and adding to our (accumulated) base.
2452 */
2453 return (base + (value - last) / (this / nbuckets));
2454 }
2455
2456 base += nbuckets - (nbuckets / factor);
2457 last = this;
2458 this = next;
2459 }
2460
2461 /*
2462 * Our value is greater than or equal to our factor taken to the
2463 * power of one plus the high magnitude -- return the top bucket.
2464 */
2465 return (base);
2466 }
2467
2468 static void
2469 dtrace_aggregate_llquantize(uint64_t *llquanta, uint64_t nval, uint64_t incr)
2470 {
2471 uint64_t arg = *llquanta++;
2472 uint16_t factor = DTRACE_LLQUANTIZE_FACTOR(arg);
2473 uint16_t low = DTRACE_LLQUANTIZE_LOW(arg);
2474 uint16_t high = DTRACE_LLQUANTIZE_HIGH(arg);
2475 uint16_t nsteps = DTRACE_LLQUANTIZE_NSTEP(arg);
2476
2477 llquanta[dtrace_aggregate_llquantize_bucket(factor,
2478 low, high, nsteps, nval)] += incr;
2479 }
2480
2481 /*ARGSUSED*/
2482 static void
2483 dtrace_aggregate_avg(uint64_t *data, uint64_t nval, uint64_t arg)
2484 {
2485 data[0]++;
2486 data[1] += nval;
2487 }
2488
2489 /*ARGSUSED*/
2490 static void
2491 dtrace_aggregate_stddev(uint64_t *data, uint64_t nval, uint64_t arg)
2492 {
2493 int64_t snval = (int64_t)nval;
2494 uint64_t tmp[2];
2495
2496 data[0]++;
2497 data[1] += nval;
2498
2499 /*
2500 * What we want to say here is:
2501 *
2502 * data[2] += nval * nval;
2503 *
2504 * But given that nval is 64-bit, we could easily overflow, so
2505 * we do this as 128-bit arithmetic.
2506 */
2507 if (snval < 0)
2508 snval = -snval;
2509
2510 dtrace_multiply_128((uint64_t)snval, (uint64_t)snval, tmp);
2511 dtrace_add_128(data + 2, tmp, data + 2);
2512 }
2513
2514 /*ARGSUSED*/
2515 static void
2516 dtrace_aggregate_count(uint64_t *oval, uint64_t nval, uint64_t arg)
2517 {
2518 *oval = *oval + 1;
2519 }
2520
2521 /*ARGSUSED*/
2522 static void
2523 dtrace_aggregate_sum(uint64_t *oval, uint64_t nval, uint64_t arg)
2524 {
2525 *oval += nval;
2526 }
2527
2528 /*
2529 * Aggregate given the tuple in the principal data buffer, and the aggregating
2530 * action denoted by the specified dtrace_aggregation_t. The aggregation
2531 * buffer is specified as the buf parameter. This routine does not return
2532 * failure; if there is no space in the aggregation buffer, the data will be
2533 * dropped, and a corresponding counter incremented.
2534 */
2535 static void
2536 dtrace_aggregate(dtrace_aggregation_t *agg, dtrace_buffer_t *dbuf,
2537 intptr_t offset, dtrace_buffer_t *buf, uint64_t expr, uint64_t arg)
2538 {
2539 dtrace_recdesc_t *rec = &agg->dtag_action.dta_rec;
2540 uint32_t i, ndx, size, fsize;
2541 uint32_t align = sizeof (uint64_t) - 1;
2542 dtrace_aggbuffer_t *agb;
2543 dtrace_aggkey_t *key;
2544 uint32_t hashval = 0, limit, isstr;
2545 caddr_t tomax, data, kdata;
2546 dtrace_actkind_t action;
2547 dtrace_action_t *act;
2548 uintptr_t offs;
2549
2550 if (buf == NULL)
2551 return;
2552
2553 if (!agg->dtag_hasarg) {
2554 /*
2555 * Currently, only quantize() and lquantize() take additional
2556 * arguments, and they have the same semantics: an increment
2557 * value that defaults to 1 when not present. If additional
2558 * aggregating actions take arguments, the setting of the
2559 * default argument value will presumably have to become more
2560 * sophisticated...
2561 */
2562 arg = 1;
2563 }
2564
2565 action = agg->dtag_action.dta_kind - DTRACEACT_AGGREGATION;
2566 size = rec->dtrd_offset - agg->dtag_base;
2567 fsize = size + rec->dtrd_size;
2568
2569 ASSERT(dbuf->dtb_tomax != NULL);
2570 data = dbuf->dtb_tomax + offset + agg->dtag_base;
2571
2572 if ((tomax = buf->dtb_tomax) == NULL) {
2573 dtrace_buffer_drop(buf);
2574 return;
2575 }
2576
2577 /*
2578 * The metastructure is always at the bottom of the buffer.
2579 */
2580 agb = (dtrace_aggbuffer_t *)(tomax + buf->dtb_size -
2581 sizeof (dtrace_aggbuffer_t));
2582
2583 if (buf->dtb_offset == 0) {
2584 /*
2585 * We just kludge up approximately 1/8th of the size to be
2586 * buckets. If this guess ends up being routinely
2587 * off-the-mark, we may need to dynamically readjust this
2588 * based on past performance.
2589 */
2590 uintptr_t hashsize = (buf->dtb_size >> 3) / sizeof (uintptr_t);
2591
2592 if ((uintptr_t)agb - hashsize * sizeof (dtrace_aggkey_t *) <
2593 (uintptr_t)tomax || hashsize == 0) {
2594 /*
2595 * We've been given a ludicrously small buffer;
2596 * increment our drop count and leave.
2597 */
2598 dtrace_buffer_drop(buf);
2599 return;
2600 }
2601
2602 /*
2603 * And now, a pathetic attempt to try to get a an odd (or
2604 * perchance, a prime) hash size for better hash distribution.
2605 */
2606 if (hashsize > (DTRACE_AGGHASHSIZE_SLEW << 3))
2607 hashsize -= DTRACE_AGGHASHSIZE_SLEW;
2608
2609 agb->dtagb_hashsize = hashsize;
2610 agb->dtagb_hash = (dtrace_aggkey_t **)((uintptr_t)agb -
2611 agb->dtagb_hashsize * sizeof (dtrace_aggkey_t *));
2612 agb->dtagb_free = (uintptr_t)agb->dtagb_hash;
2613
2614 for (i = 0; i < agb->dtagb_hashsize; i++)
2615 agb->dtagb_hash[i] = NULL;
2616 }
2617
2618 ASSERT(agg->dtag_first != NULL);
2619 ASSERT(agg->dtag_first->dta_intuple);
2620
2621 /*
2622 * Calculate the hash value based on the key. Note that we _don't_
2623 * include the aggid in the hashing (but we will store it as part of
2624 * the key). The hashing algorithm is Bob Jenkins' "One-at-a-time"
2625 * algorithm: a simple, quick algorithm that has no known funnels, and
2626 * gets good distribution in practice. The efficacy of the hashing
2627 * algorithm (and a comparison with other algorithms) may be found by
2628 * running the ::dtrace_aggstat MDB dcmd.
2629 */
2630 for (act = agg->dtag_first; act->dta_intuple; act = act->dta_next) {
2631 i = act->dta_rec.dtrd_offset - agg->dtag_base;
2632 limit = i + act->dta_rec.dtrd_size;
2633 ASSERT(limit <= size);
2634 isstr = DTRACEACT_ISSTRING(act);
2635
2636 for (; i < limit; i++) {
2637 hashval += data[i];
2638 hashval += (hashval << 10);
2639 hashval ^= (hashval >> 6);
2640
2641 if (isstr && data[i] == '\0')
2642 break;
2643 }
2644 }
2645
2646 hashval += (hashval << 3);
2647 hashval ^= (hashval >> 11);
2648 hashval += (hashval << 15);
2649
2650 /*
2651 * Yes, the divide here is expensive -- but it's generally the least
2652 * of the performance issues given the amount of data that we iterate
2653 * over to compute hash values, compare data, etc.
2654 */
2655 ndx = hashval % agb->dtagb_hashsize;
2656
2657 for (key = agb->dtagb_hash[ndx]; key != NULL; key = key->dtak_next) {
2658 ASSERT((caddr_t)key >= tomax);
2659 ASSERT((caddr_t)key < tomax + buf->dtb_size);
2660
2661 if (hashval != key->dtak_hashval || key->dtak_size != size)
2662 continue;
2663
2664 kdata = key->dtak_data;
2665 ASSERT(kdata >= tomax && kdata < tomax + buf->dtb_size);
2666
2667 for (act = agg->dtag_first; act->dta_intuple;
2668 act = act->dta_next) {
2669 i = act->dta_rec.dtrd_offset - agg->dtag_base;
2670 limit = i + act->dta_rec.dtrd_size;
2671 ASSERT(limit <= size);
2672 isstr = DTRACEACT_ISSTRING(act);
2673
2674 for (; i < limit; i++) {
2675 if (kdata[i] != data[i])
2676 goto next;
2677
2678 if (isstr && data[i] == '\0')
2679 break;
2680 }
2681 }
2682
2683 if (action != key->dtak_action) {
2684 /*
2685 * We are aggregating on the same value in the same
2686 * aggregation with two different aggregating actions.
2687 * (This should have been picked up in the compiler,
2688 * so we may be dealing with errant or devious DIF.)
2689 * This is an error condition; we indicate as much,
2690 * and return.
2691 */
2692 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
2693 return;
2694 }
2695
2696 /*
2697 * This is a hit: we need to apply the aggregator to
2698 * the value at this key.
2699 */
2700 agg->dtag_aggregate((uint64_t *)(kdata + size), expr, arg);
2701 return;
2702 next:
2703 continue;
2704 }
2705
2706 /*
2707 * We didn't find it. We need to allocate some zero-filled space,
2708 * link it into the hash table appropriately, and apply the aggregator
2709 * to the (zero-filled) value.
2710 */
2711 offs = buf->dtb_offset;
2712 while (offs & (align - 1))
2713 offs += sizeof (uint32_t);
2714
2715 /*
2716 * If we don't have enough room to both allocate a new key _and_
2717 * its associated data, increment the drop count and return.
2718 */
2719 if ((uintptr_t)tomax + offs + fsize >
2720 agb->dtagb_free - sizeof (dtrace_aggkey_t)) {
2721 dtrace_buffer_drop(buf);
2722 return;
2723 }
2724
2725 /*CONSTCOND*/
2726 ASSERT(!(sizeof (dtrace_aggkey_t) & (sizeof (uintptr_t) - 1)));
2727 key = (dtrace_aggkey_t *)(agb->dtagb_free - sizeof (dtrace_aggkey_t));
2728 agb->dtagb_free -= sizeof (dtrace_aggkey_t);
2729
2730 key->dtak_data = kdata = tomax + offs;
2731 buf->dtb_offset = offs + fsize;
2732
2733 /*
2734 * Now copy the data across.
2735 */
2736 *((dtrace_aggid_t *)kdata) = agg->dtag_id;
2737
2738 for (i = sizeof (dtrace_aggid_t); i < size; i++)
2739 kdata[i] = data[i];
2740
2741 /*
2742 * Because strings are not zeroed out by default, we need to iterate
2743 * looking for actions that store strings, and we need to explicitly
2744 * pad these strings out with zeroes.
2745 */
2746 for (act = agg->dtag_first; act->dta_intuple; act = act->dta_next) {
2747 int nul;
2748
2749 if (!DTRACEACT_ISSTRING(act))
2750 continue;
2751
2752 i = act->dta_rec.dtrd_offset - agg->dtag_base;
2753 limit = i + act->dta_rec.dtrd_size;
2754 ASSERT(limit <= size);
2755
2756 for (nul = 0; i < limit; i++) {
2757 if (nul) {
2758 kdata[i] = '\0';
2759 continue;
2760 }
2761
2762 if (data[i] != '\0')
2763 continue;
2764
2765 nul = 1;
2766 }
2767 }
2768
2769 for (i = size; i < fsize; i++)
2770 kdata[i] = 0;
2771
2772 key->dtak_hashval = hashval;
2773 key->dtak_size = size;
2774 key->dtak_action = action;
2775 key->dtak_next = agb->dtagb_hash[ndx];
2776 agb->dtagb_hash[ndx] = key;
2777
2778 /*
2779 * Finally, apply the aggregator.
2780 */
2781 *((uint64_t *)(key->dtak_data + size)) = agg->dtag_initial;
2782 agg->dtag_aggregate((uint64_t *)(key->dtak_data + size), expr, arg);
2783 }
2784
2785 /*
2786 * Given consumer state, this routine finds a speculation in the INACTIVE
2787 * state and transitions it into the ACTIVE state. If there is no speculation
2788 * in the INACTIVE state, 0 is returned. In this case, no error counter is
2789 * incremented -- it is up to the caller to take appropriate action.
2790 */
2791 static int
2792 dtrace_speculation(dtrace_state_t *state)
2793 {
2794 int i = 0;
2795 dtrace_speculation_state_t curstate;
2796 uint32_t *stat = &state->dts_speculations_unavail, count;
2797
2798 while (i < state->dts_nspeculations) {
2799 dtrace_speculation_t *spec = &state->dts_speculations[i];
2800
2801 curstate = spec->dtsp_state;
2802
2803 if (curstate != DTRACESPEC_INACTIVE) {
2804 if (curstate == DTRACESPEC_COMMITTINGMANY ||
2805 curstate == DTRACESPEC_COMMITTING ||
2806 curstate == DTRACESPEC_DISCARDING)
2807 stat = &state->dts_speculations_busy;
2808 i++;
2809 continue;
2810 }
2811
2812 if (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2813 curstate, DTRACESPEC_ACTIVE) == curstate)
2814 return (i + 1);
2815 }
2816
2817 /*
2818 * We couldn't find a speculation. If we found as much as a single
2819 * busy speculation buffer, we'll attribute this failure as "busy"
2820 * instead of "unavail".
2821 */
2822 do {
2823 count = *stat;
2824 } while (dtrace_cas32(stat, count, count + 1) != count);
2825
2826 return (0);
2827 }
2828
2829 /*
2830 * This routine commits an active speculation. If the specified speculation
2831 * is not in a valid state to perform a commit(), this routine will silently do
2832 * nothing. The state of the specified speculation is transitioned according
2833 * to the state transition diagram outlined in <sys/dtrace_impl.h>
2834 */
2835 static void
2836 dtrace_speculation_commit(dtrace_state_t *state, processorid_t cpu,
2837 dtrace_specid_t which)
2838 {
2839 dtrace_speculation_t *spec;
2840 dtrace_buffer_t *src, *dest;
2841 uintptr_t daddr, saddr, dlimit, slimit;
2842 dtrace_speculation_state_t curstate, new = 0;
2843 intptr_t offs;
2844 uint64_t timestamp;
2845
2846 if (which == 0)
2847 return;
2848
2849 if (which > state->dts_nspeculations) {
2850 cpu_core[cpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
2851 return;
2852 }
2853
2854 spec = &state->dts_speculations[which - 1];
2855 src = &spec->dtsp_buffer[cpu];
2856 dest = &state->dts_buffer[cpu];
2857
2858 do {
2859 curstate = spec->dtsp_state;
2860
2861 if (curstate == DTRACESPEC_COMMITTINGMANY)
2862 break;
2863
2864 switch (curstate) {
2865 case DTRACESPEC_INACTIVE:
2866 case DTRACESPEC_DISCARDING:
2867 return;
2868
2869 case DTRACESPEC_COMMITTING:
2870 /*
2871 * This is only possible if we are (a) commit()'ing
2872 * without having done a prior speculate() on this CPU
2873 * and (b) racing with another commit() on a different
2874 * CPU. There's nothing to do -- we just assert that
2875 * our offset is 0.
2876 */
2877 ASSERT(src->dtb_offset == 0);
2878 return;
2879
2880 case DTRACESPEC_ACTIVE:
2881 new = DTRACESPEC_COMMITTING;
2882 break;
2883
2884 case DTRACESPEC_ACTIVEONE:
2885 /*
2886 * This speculation is active on one CPU. If our
2887 * buffer offset is non-zero, we know that the one CPU
2888 * must be us. Otherwise, we are committing on a
2889 * different CPU from the speculate(), and we must
2890 * rely on being asynchronously cleaned.
2891 */
2892 if (src->dtb_offset != 0) {
2893 new = DTRACESPEC_COMMITTING;
2894 break;
2895 }
2896 /*FALLTHROUGH*/
2897
2898 case DTRACESPEC_ACTIVEMANY:
2899 new = DTRACESPEC_COMMITTINGMANY;
2900 break;
2901
2902 default:
2903 ASSERT(0);
2904 }
2905 } while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
2906 curstate, new) != curstate);
2907
2908 /*
2909 * We have set the state to indicate that we are committing this
2910 * speculation. Now reserve the necessary space in the destination
2911 * buffer.
2912 */
2913 if ((offs = dtrace_buffer_reserve(dest, src->dtb_offset,
2914 sizeof (uint64_t), state, NULL)) < 0) {
2915 dtrace_buffer_drop(dest);
2916 goto out;
2917 }
2918
2919 /*
2920 * We have sufficient space to copy the speculative buffer into the
2921 * primary buffer. First, modify the speculative buffer, filling
2922 * in the timestamp of all entries with the curstate time. The data
2923 * must have the commit() time rather than the time it was traced,
2924 * so that all entries in the primary buffer are in timestamp order.
2925 */
2926 timestamp = dtrace_gethrtime();
2927 saddr = (uintptr_t)src->dtb_tomax;
2928 slimit = saddr + src->dtb_offset;
2929 while (saddr < slimit) {
2930 size_t size;
2931 dtrace_rechdr_t *dtrh = (dtrace_rechdr_t *)saddr;
2932
2933 if (dtrh->dtrh_epid == DTRACE_EPIDNONE) {
2934 saddr += sizeof (dtrace_epid_t);
2935 continue;
2936 }
2937 ASSERT3U(dtrh->dtrh_epid, <=, state->dts_necbs);
2938 size = state->dts_ecbs[dtrh->dtrh_epid - 1]->dte_size;
2939
2940 ASSERT3U(saddr + size, <=, slimit);
2941 ASSERT3U(size, >=, sizeof (dtrace_rechdr_t));
2942 ASSERT3U(DTRACE_RECORD_LOAD_TIMESTAMP(dtrh), ==, UINT64_MAX);
2943
2944 DTRACE_RECORD_STORE_TIMESTAMP(dtrh, timestamp);
2945
2946 saddr += size;
2947 }
2948
2949 /*
2950 * Copy the buffer across. (Note that this is a
2951 * highly subobtimal bcopy(); in the unlikely event that this becomes
2952 * a serious performance issue, a high-performance DTrace-specific
2953 * bcopy() should obviously be invented.)
2954 */
2955 daddr = (uintptr_t)dest->dtb_tomax + offs;
2956 dlimit = daddr + src->dtb_offset;
2957 saddr = (uintptr_t)src->dtb_tomax;
2958
2959 /*
2960 * First, the aligned portion.
2961 */
2962 while (dlimit - daddr >= sizeof (uint64_t)) {
2963 *((uint64_t *)daddr) = *((uint64_t *)saddr);
2964
2965 daddr += sizeof (uint64_t);
2966 saddr += sizeof (uint64_t);
2967 }
2968
2969 /*
2970 * Now any left-over bit...
2971 */
2972 while (dlimit - daddr)
2973 *((uint8_t *)daddr++) = *((uint8_t *)saddr++);
2974
2975 /*
2976 * Finally, commit the reserved space in the destination buffer.
2977 */
2978 dest->dtb_offset = offs + src->dtb_offset;
2979
2980 out:
2981 /*
2982 * If we're lucky enough to be the only active CPU on this speculation
2983 * buffer, we can just set the state back to DTRACESPEC_INACTIVE.
2984 */
2985 if (curstate == DTRACESPEC_ACTIVE ||
2986 (curstate == DTRACESPEC_ACTIVEONE && new == DTRACESPEC_COMMITTING)) {
2987 uint32_t rval = dtrace_cas32((uint32_t *)&spec->dtsp_state,
2988 DTRACESPEC_COMMITTING, DTRACESPEC_INACTIVE);
2989
2990 ASSERT(rval == DTRACESPEC_COMMITTING);
2991 }
2992
2993 src->dtb_offset = 0;
2994 src->dtb_xamot_drops += src->dtb_drops;
2995 src->dtb_drops = 0;
2996 }
2997
2998 /*
2999 * This routine discards an active speculation. If the specified speculation
3000 * is not in a valid state to perform a discard(), this routine will silently
3001 * do nothing. The state of the specified speculation is transitioned
3002 * according to the state transition diagram outlined in <sys/dtrace_impl.h>
3003 */
3004 static void
3005 dtrace_speculation_discard(dtrace_state_t *state, processorid_t cpu,
3006 dtrace_specid_t which)
3007 {
3008 dtrace_speculation_t *spec;
3009 dtrace_speculation_state_t curstate, new = 0;
3010 dtrace_buffer_t *buf;
3011
3012 if (which == 0)
3013 return;
3014
3015 if (which > state->dts_nspeculations) {
3016 cpu_core[cpu].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
3017 return;
3018 }
3019
3020 spec = &state->dts_speculations[which - 1];
3021 buf = &spec->dtsp_buffer[cpu];
3022
3023 do {
3024 curstate = spec->dtsp_state;
3025
3026 switch (curstate) {
3027 case DTRACESPEC_INACTIVE:
3028 case DTRACESPEC_COMMITTINGMANY:
3029 case DTRACESPEC_COMMITTING:
3030 case DTRACESPEC_DISCARDING:
3031 return;
3032
3033 case DTRACESPEC_ACTIVE:
3034 case DTRACESPEC_ACTIVEMANY:
3035 new = DTRACESPEC_DISCARDING;
3036 break;
3037
3038 case DTRACESPEC_ACTIVEONE:
3039 if (buf->dtb_offset != 0) {
3040 new = DTRACESPEC_INACTIVE;
3041 } else {
3042 new = DTRACESPEC_DISCARDING;
3043 }
3044 break;
3045
3046 default:
3047 ASSERT(0);
3048 }
3049 } while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
3050 curstate, new) != curstate);
3051
3052 buf->dtb_offset = 0;
3053 buf->dtb_drops = 0;
3054 }
3055
3056 /*
3057 * Note: not called from probe context. This function is called
3058 * asynchronously from cross call context to clean any speculations that are
3059 * in the COMMITTINGMANY or DISCARDING states. These speculations may not be
3060 * transitioned back to the INACTIVE state until all CPUs have cleaned the
3061 * speculation.
3062 */
3063 static void
3064 dtrace_speculation_clean_here(dtrace_state_t *state)
3065 {
3066 dtrace_icookie_t cookie;
3067 processorid_t cpu = curcpu;
3068 dtrace_buffer_t *dest = &state->dts_buffer[cpu];
3069 dtrace_specid_t i;
3070
3071 cookie = dtrace_interrupt_disable();
3072
3073 if (dest->dtb_tomax == NULL) {
3074 dtrace_interrupt_enable(cookie);
3075 return;
3076 }
3077
3078 for (i = 0; i < state->dts_nspeculations; i++) {
3079 dtrace_speculation_t *spec = &state->dts_speculations[i];
3080 dtrace_buffer_t *src = &spec->dtsp_buffer[cpu];
3081
3082 if (src->dtb_tomax == NULL)
3083 continue;
3084
3085 if (spec->dtsp_state == DTRACESPEC_DISCARDING) {
3086 src->dtb_offset = 0;
3087 continue;
3088 }
3089
3090 if (spec->dtsp_state != DTRACESPEC_COMMITTINGMANY)
3091 continue;
3092
3093 if (src->dtb_offset == 0)
3094 continue;
3095
3096 dtrace_speculation_commit(state, cpu, i + 1);
3097 }
3098
3099 dtrace_interrupt_enable(cookie);
3100 }
3101
3102 /*
3103 * Note: not called from probe context. This function is called
3104 * asynchronously (and at a regular interval) to clean any speculations that
3105 * are in the COMMITTINGMANY or DISCARDING states. If it discovers that there
3106 * is work to be done, it cross calls all CPUs to perform that work;
3107 * COMMITMANY and DISCARDING speculations may not be transitioned back to the
3108 * INACTIVE state until they have been cleaned by all CPUs.
3109 */
3110 static void
3111 dtrace_speculation_clean(dtrace_state_t *state)
3112 {
3113 int work = 0, rv;
3114 dtrace_specid_t i;
3115
3116 for (i = 0; i < state->dts_nspeculations; i++) {
3117 dtrace_speculation_t *spec = &state->dts_speculations[i];
3118
3119 ASSERT(!spec->dtsp_cleaning);
3120
3121 if (spec->dtsp_state != DTRACESPEC_DISCARDING &&
3122 spec->dtsp_state != DTRACESPEC_COMMITTINGMANY)
3123 continue;
3124
3125 work++;
3126 spec->dtsp_cleaning = 1;
3127 }
3128
3129 if (!work)
3130 return;
3131
3132 dtrace_xcall(DTRACE_CPUALL,
3133 (dtrace_xcall_t)dtrace_speculation_clean_here, state);
3134
3135 /*
3136 * We now know that all CPUs have committed or discarded their
3137 * speculation buffers, as appropriate. We can now set the state
3138 * to inactive.
3139 */
3140 for (i = 0; i < state->dts_nspeculations; i++) {
3141 dtrace_speculation_t *spec = &state->dts_speculations[i];
3142 dtrace_speculation_state_t curstate, new;
3143
3144 if (!spec->dtsp_cleaning)
3145 continue;
3146
3147 curstate = spec->dtsp_state;
3148 ASSERT(curstate == DTRACESPEC_DISCARDING ||
3149 curstate == DTRACESPEC_COMMITTINGMANY);
3150
3151 new = DTRACESPEC_INACTIVE;
3152
3153 rv = dtrace_cas32((uint32_t *)&spec->dtsp_state, curstate, new);
3154 ASSERT(rv == curstate);
3155 spec->dtsp_cleaning = 0;
3156 }
3157 }
3158
3159 /*
3160 * Called as part of a speculate() to get the speculative buffer associated
3161 * with a given speculation. Returns NULL if the specified speculation is not
3162 * in an ACTIVE state. If the speculation is in the ACTIVEONE state -- and
3163 * the active CPU is not the specified CPU -- the speculation will be
3164 * atomically transitioned into the ACTIVEMANY state.
3165 */
3166 static dtrace_buffer_t *
3167 dtrace_speculation_buffer(dtrace_state_t *state, processorid_t cpuid,
3168 dtrace_specid_t which)
3169 {
3170 dtrace_speculation_t *spec;
3171 dtrace_speculation_state_t curstate, new = 0;
3172 dtrace_buffer_t *buf;
3173
3174 if (which == 0)
3175 return (NULL);
3176
3177 if (which > state->dts_nspeculations) {
3178 cpu_core[cpuid].cpuc_dtrace_flags |= CPU_DTRACE_ILLOP;
3179 return (NULL);
3180 }
3181
3182 spec = &state->dts_speculations[which - 1];
3183 buf = &spec->dtsp_buffer[cpuid];
3184
3185 do {
3186 curstate = spec->dtsp_state;
3187
3188 switch (curstate) {
3189 case DTRACESPEC_INACTIVE:
3190 case DTRACESPEC_COMMITTINGMANY:
3191 case DTRACESPEC_DISCARDING:
3192 return (NULL);
3193
3194 case DTRACESPEC_COMMITTING:
3195 ASSERT(buf->dtb_offset == 0);
3196 return (NULL);
3197
3198 case DTRACESPEC_ACTIVEONE:
3199 /*
3200 * This speculation is currently active on one CPU.
3201 * Check the offset in the buffer; if it's non-zero,
3202 * that CPU must be us (and we leave the state alone).
3203 * If it's zero, assume that we're starting on a new
3204 * CPU -- and change the state to indicate that the
3205 * speculation is active on more than one CPU.
3206 */
3207 if (buf->dtb_offset != 0)
3208 return (buf);
3209
3210 new = DTRACESPEC_ACTIVEMANY;
3211 break;
3212
3213 case DTRACESPEC_ACTIVEMANY:
3214 return (buf);
3215
3216 case DTRACESPEC_ACTIVE:
3217 new = DTRACESPEC_ACTIVEONE;
3218 break;
3219
3220 default:
3221 ASSERT(0);
3222 }
3223 } while (dtrace_cas32((uint32_t *)&spec->dtsp_state,
3224 curstate, new) != curstate);
3225
3226 ASSERT(new == DTRACESPEC_ACTIVEONE || new == DTRACESPEC_ACTIVEMANY);
3227 return (buf);
3228 }
3229
3230 /*
3231 * Return a string. In the event that the user lacks the privilege to access
3232 * arbitrary kernel memory, we copy the string out to scratch memory so that we
3233 * don't fail access checking.
3234 *
3235 * dtrace_dif_variable() uses this routine as a helper for various
3236 * builtin values such as 'execname' and 'probefunc.'
3237 */
3238 uintptr_t
3239 dtrace_dif_varstr(uintptr_t addr, dtrace_state_t *state,
3240 dtrace_mstate_t *mstate)
3241 {
3242 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
3243 uintptr_t ret;
3244 size_t strsz;
3245
3246 /*
3247 * The easy case: this probe is allowed to read all of memory, so
3248 * we can just return this as a vanilla pointer.
3249 */
3250 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) != 0)
3251 return (addr);
3252
3253 /*
3254 * This is the tougher case: we copy the string in question from
3255 * kernel memory into scratch memory and return it that way: this
3256 * ensures that we won't trip up when access checking tests the
3257 * BYREF return value.
3258 */
3259 strsz = dtrace_strlen((char *)addr, size) + 1;
3260
3261 if (mstate->dtms_scratch_ptr + strsz >
3262 mstate->dtms_scratch_base + mstate->dtms_scratch_size) {
3263 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3264 return (0);
3265 }
3266
3267 dtrace_strcpy((const void *)addr, (void *)mstate->dtms_scratch_ptr,
3268 strsz);
3269 ret = mstate->dtms_scratch_ptr;
3270 mstate->dtms_scratch_ptr += strsz;
3271 return (ret);
3272 }
3273
3274 /*
3275 * Return a string from a memoy address which is known to have one or
3276 * more concatenated, individually zero terminated, sub-strings.
3277 * In the event that the user lacks the privilege to access
3278 * arbitrary kernel memory, we copy the string out to scratch memory so that we
3279 * don't fail access checking.
3280 *
3281 * dtrace_dif_variable() uses this routine as a helper for various
3282 * builtin values such as 'execargs'.
3283 */
3284 static uintptr_t
3285 dtrace_dif_varstrz(uintptr_t addr, size_t strsz, dtrace_state_t *state,
3286 dtrace_mstate_t *mstate)
3287 {
3288 char *p;
3289 size_t i;
3290 uintptr_t ret;
3291
3292 if (mstate->dtms_scratch_ptr + strsz >
3293 mstate->dtms_scratch_base + mstate->dtms_scratch_size) {
3294 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
3295 return (0);
3296 }
3297
3298 dtrace_bcopy((const void *)addr, (void *)mstate->dtms_scratch_ptr,
3299 strsz);
3300
3301 /* Replace sub-string termination characters with a space. */
3302 for (p = (char *) mstate->dtms_scratch_ptr, i = 0; i < strsz - 1;
3303 p++, i++)
3304 if (*p == '\0')
3305 *p = ' ';
3306
3307 ret = mstate->dtms_scratch_ptr;
3308 mstate->dtms_scratch_ptr += strsz;
3309 return (ret);
3310 }
3311
3312 /*
3313 * This function implements the DIF emulator's variable lookups. The emulator
3314 * passes a reserved variable identifier and optional built-in array index.
3315 */
3316 static uint64_t
3317 dtrace_dif_variable(dtrace_mstate_t *mstate, dtrace_state_t *state, uint64_t v,
3318 uint64_t ndx)
3319 {
3320 /*
3321 * If we're accessing one of the uncached arguments, we'll turn this
3322 * into a reference in the args array.
3323 */
3324 if (v >= DIF_VAR_ARG0 && v <= DIF_VAR_ARG9) {
3325 ndx = v - DIF_VAR_ARG0;
3326 v = DIF_VAR_ARGS;
3327 }
3328
3329 switch (v) {
3330 case DIF_VAR_ARGS:
3331 ASSERT(mstate->dtms_present & DTRACE_MSTATE_ARGS);
3332 if (ndx >= sizeof (mstate->dtms_arg) /
3333 sizeof (mstate->dtms_arg[0])) {
3334 int aframes = mstate->dtms_probe->dtpr_aframes + 2;
3335 dtrace_provider_t *pv;
3336 uint64_t val;
3337
3338 pv = mstate->dtms_probe->dtpr_provider;
3339 if (pv->dtpv_pops.dtps_getargval != NULL)
3340 val = pv->dtpv_pops.dtps_getargval(pv->dtpv_arg,
3341 mstate->dtms_probe->dtpr_id,
3342 mstate->dtms_probe->dtpr_arg, ndx, aframes);
3343 else
3344 val = dtrace_getarg(ndx, aframes);
3345
3346 /*
3347 * This is regrettably required to keep the compiler
3348 * from tail-optimizing the call to dtrace_getarg().
3349 * The condition always evaluates to true, but the
3350 * compiler has no way of figuring that out a priori.
3351 * (None of this would be necessary if the compiler
3352 * could be relied upon to _always_ tail-optimize
3353 * the call to dtrace_getarg() -- but it can't.)
3354 */
3355 if (mstate->dtms_probe != NULL)
3356 return (val);
3357
3358 ASSERT(0);
3359 }
3360
3361 return (mstate->dtms_arg[ndx]);
3362
3363 case DIF_VAR_REGS:
3364 case DIF_VAR_UREGS: {
3365 struct trapframe *tframe;
3366
3367 if (!dtrace_priv_proc(state))
3368 return (0);
3369
3370 if (v == DIF_VAR_REGS)
3371 tframe = curthread->t_dtrace_trapframe;
3372 else
3373 tframe = curthread->td_frame;
3374
3375 if (tframe == NULL) {
3376 DTRACE_CPUFLAG_SET(CPU_DTRACE_BADADDR);
3377 cpu_core[curcpu].cpuc_dtrace_illval = 0;
3378 return (0);
3379 }
3380
3381 return (dtrace_getreg(tframe, ndx));
3382 }
3383
3384 case DIF_VAR_CURTHREAD:
3385 if (!dtrace_priv_proc(state))
3386 return (0);
3387 return ((uint64_t)(uintptr_t)curthread);
3388
3389 case DIF_VAR_TIMESTAMP:
3390 if (!(mstate->dtms_present & DTRACE_MSTATE_TIMESTAMP)) {
3391 mstate->dtms_timestamp = dtrace_gethrtime();
3392 mstate->dtms_present |= DTRACE_MSTATE_TIMESTAMP;
3393 }
3394 return (mstate->dtms_timestamp);
3395
3396 case DIF_VAR_VTIMESTAMP:
3397 ASSERT(dtrace_vtime_references != 0);
3398 return (curthread->t_dtrace_vtime);
3399
3400 case DIF_VAR_WALLTIMESTAMP:
3401 if (!(mstate->dtms_present & DTRACE_MSTATE_WALLTIMESTAMP)) {
3402 mstate->dtms_walltimestamp = dtrace_gethrestime();
3403 mstate->dtms_present |= DTRACE_MSTATE_WALLTIMESTAMP;
3404 }
3405 return (mstate->dtms_walltimestamp);
3406
3407 #ifdef illumos
3408 case DIF_VAR_IPL:
3409 if (!dtrace_priv_kernel(state))
3410 return (0);
3411 if (!(mstate->dtms_present & DTRACE_MSTATE_IPL)) {
3412 mstate->dtms_ipl = dtrace_getipl();
3413 mstate->dtms_present |= DTRACE_MSTATE_IPL;
3414 }
3415 return (mstate->dtms_ipl);
3416 #endif
3417
3418 case DIF_VAR_EPID:
3419 ASSERT(mstate->dtms_present & DTRACE_MSTATE_EPID);
3420 return (mstate->dtms_epid);
3421
3422 case DIF_VAR_ID:
3423 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3424 return (mstate->dtms_probe->dtpr_id);
3425
3426 case DIF_VAR_STACKDEPTH:
3427 if (!dtrace_priv_kernel(state))
3428 return (0);
3429 if (!(mstate->dtms_present & DTRACE_MSTATE_STACKDEPTH)) {
3430 int aframes = mstate->dtms_probe->dtpr_aframes + 2;
3431
3432 mstate->dtms_stackdepth = dtrace_getstackdepth(aframes);
3433 mstate->dtms_present |= DTRACE_MSTATE_STACKDEPTH;
3434 }
3435 return (mstate->dtms_stackdepth);
3436
3437 case DIF_VAR_USTACKDEPTH:
3438 if (!dtrace_priv_proc(state))
3439 return (0);
3440 if (!(mstate->dtms_present & DTRACE_MSTATE_USTACKDEPTH)) {
3441 /*
3442 * See comment in DIF_VAR_PID.
3443 */
3444 if (DTRACE_ANCHORED(mstate->dtms_probe) &&
3445 CPU_ON_INTR(CPU)) {
3446 mstate->dtms_ustackdepth = 0;
3447 } else {
3448 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3449 mstate->dtms_ustackdepth =
3450 dtrace_getustackdepth();
3451 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3452 }
3453 mstate->dtms_present |= DTRACE_MSTATE_USTACKDEPTH;
3454 }
3455 return (mstate->dtms_ustackdepth);
3456
3457 case DIF_VAR_CALLER:
3458 if (!dtrace_priv_kernel(state))
3459 return (0);
3460 if (!(mstate->dtms_present & DTRACE_MSTATE_CALLER)) {
3461 int aframes = mstate->dtms_probe->dtpr_aframes + 2;
3462
3463 if (!DTRACE_ANCHORED(mstate->dtms_probe)) {
3464 /*
3465 * If this is an unanchored probe, we are
3466 * required to go through the slow path:
3467 * dtrace_caller() only guarantees correct
3468 * results for anchored probes.
3469 */
3470 pc_t caller[2] = {0, 0};
3471
3472 dtrace_getpcstack(caller, 2, aframes,
3473 (uint32_t *)(uintptr_t)mstate->dtms_arg[0]);
3474 mstate->dtms_caller = caller[1];
3475 } else if ((mstate->dtms_caller =
3476 dtrace_caller(aframes)) == -1) {
3477 /*
3478 * We have failed to do this the quick way;
3479 * we must resort to the slower approach of
3480 * calling dtrace_getpcstack().
3481 */
3482 pc_t caller = 0;
3483
3484 dtrace_getpcstack(&caller, 1, aframes, NULL);
3485 mstate->dtms_caller = caller;
3486 }
3487
3488 mstate->dtms_present |= DTRACE_MSTATE_CALLER;
3489 }
3490 return (mstate->dtms_caller);
3491
3492 case DIF_VAR_UCALLER:
3493 if (!dtrace_priv_proc(state))
3494 return (0);
3495
3496 if (!(mstate->dtms_present & DTRACE_MSTATE_UCALLER)) {
3497 uint64_t ustack[3];
3498
3499 /*
3500 * dtrace_getupcstack() fills in the first uint64_t
3501 * with the current PID. The second uint64_t will
3502 * be the program counter at user-level. The third
3503 * uint64_t will contain the caller, which is what
3504 * we're after.
3505 */
3506 ustack[2] = 0;
3507 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
3508 dtrace_getupcstack(ustack, 3);
3509 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
3510 mstate->dtms_ucaller = ustack[2];
3511 mstate->dtms_present |= DTRACE_MSTATE_UCALLER;
3512 }
3513
3514 return (mstate->dtms_ucaller);
3515
3516 case DIF_VAR_PROBEPROV:
3517 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3518 return (dtrace_dif_varstr(
3519 (uintptr_t)mstate->dtms_probe->dtpr_provider->dtpv_name,
3520 state, mstate));
3521
3522 case DIF_VAR_PROBEMOD:
3523 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3524 return (dtrace_dif_varstr(
3525 (uintptr_t)mstate->dtms_probe->dtpr_mod,
3526 state, mstate));
3527
3528 case DIF_VAR_PROBEFUNC:
3529 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3530 return (dtrace_dif_varstr(
3531 (uintptr_t)mstate->dtms_probe->dtpr_func,
3532 state, mstate));
3533
3534 case DIF_VAR_PROBENAME:
3535 ASSERT(mstate->dtms_present & DTRACE_MSTATE_PROBE);
3536 return (dtrace_dif_varstr(
3537 (uintptr_t)mstate->dtms_probe->dtpr_name,
3538 state, mstate));
3539
3540 case DIF_VAR_PID:
3541 if (!dtrace_priv_proc(state))
3542 return (0);
3543
3544 #ifdef illumos
3545 /*
3546 * Note that we are assuming that an unanchored probe is
3547 * always due to a high-level interrupt. (And we're assuming
3548 * that there is only a single high level interrupt.)
3549 */
3550 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3551 return (pid0.pid_id);
3552
3553 /*
3554 * It is always safe to dereference one's own t_procp pointer:
3555 * it always points to a valid, allocated proc structure.
3556 * Further, it is always safe to dereference the p_pidp member
3557 * of one's own proc structure. (These are truisms becuase
3558 * threads and processes don't clean up their own state --
3559 * they leave that task to whomever reaps them.)
3560 */
3561 return ((uint64_t)curthread->t_procp->p_pidp->pid_id);
3562 #else
3563 return ((uint64_t)curproc->p_pid);
3564 #endif
3565
3566 case DIF_VAR_PPID:
3567 if (!dtrace_priv_proc(state))
3568 return (0);
3569
3570 #ifdef illumos
3571 /*
3572 * See comment in DIF_VAR_PID.
3573 */
3574 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3575 return (pid0.pid_id);
3576
3577 /*
3578 * It is always safe to dereference one's own t_procp pointer:
3579 * it always points to a valid, allocated proc structure.
3580 * (This is true because threads don't clean up their own
3581 * state -- they leave that task to whomever reaps them.)
3582 */
3583 return ((uint64_t)curthread->t_procp->p_ppid);
3584 #else
3585 if (curproc->p_pid == proc0.p_pid)
3586 return (curproc->p_pid);
3587 else
3588 return (curproc->p_pptr->p_pid);
3589 #endif
3590
3591 case DIF_VAR_TID:
3592 #ifdef illumos
3593 /*
3594 * See comment in DIF_VAR_PID.
3595 */
3596 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3597 return (0);
3598 #endif
3599
3600 return ((uint64_t)curthread->t_tid);
3601
3602 case DIF_VAR_EXECARGS: {
3603 struct pargs *p_args = curthread->td_proc->p_args;
3604
3605 if (p_args == NULL)
3606 return(0);
3607
3608 return (dtrace_dif_varstrz(
3609 (uintptr_t) p_args->ar_args, p_args->ar_length, state, mstate));
3610 }
3611
3612 case DIF_VAR_EXECNAME:
3613 #ifdef illumos
3614 if (!dtrace_priv_proc(state))
3615 return (0);
3616
3617 /*
3618 * See comment in DIF_VAR_PID.
3619 */
3620 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3621 return ((uint64_t)(uintptr_t)p0.p_user.u_comm);
3622
3623 /*
3624 * It is always safe to dereference one's own t_procp pointer:
3625 * it always points to a valid, allocated proc structure.
3626 * (This is true because threads don't clean up their own
3627 * state -- they leave that task to whomever reaps them.)
3628 */
3629 return (dtrace_dif_varstr(
3630 (uintptr_t)curthread->t_procp->p_user.u_comm,
3631 state, mstate));
3632 #else
3633 return (dtrace_dif_varstr(
3634 (uintptr_t) curthread->td_proc->p_comm, state, mstate));
3635 #endif
3636
3637 case DIF_VAR_ZONENAME:
3638 #ifdef illumos
3639 if (!dtrace_priv_proc(state))
3640 return (0);
3641
3642 /*
3643 * See comment in DIF_VAR_PID.
3644 */
3645 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3646 return ((uint64_t)(uintptr_t)p0.p_zone->zone_name);
3647
3648 /*
3649 * It is always safe to dereference one's own t_procp pointer:
3650 * it always points to a valid, allocated proc structure.
3651 * (This is true because threads don't clean up their own
3652 * state -- they leave that task to whomever reaps them.)
3653 */
3654 return (dtrace_dif_varstr(
3655 (uintptr_t)curthread->t_procp->p_zone->zone_name,
3656 state, mstate));
3657 #elif defined(__FreeBSD__)
3658 /*
3659 * On FreeBSD, we introduce compatibility to zonename by falling through
3660 * into jailname.
3661 */
3662 case DIF_VAR_JAILNAME:
3663 if (!dtrace_priv_kernel(state))
3664 return (0);
3665
3666 return (dtrace_dif_varstr(
3667 (uintptr_t)curthread->td_ucred->cr_prison->pr_name,
3668 state, mstate));
3669
3670 case DIF_VAR_JID:
3671 if (!dtrace_priv_kernel(state))
3672 return (0);
3673
3674 return ((uint64_t)curthread->td_ucred->cr_prison->pr_id);
3675 #else
3676 return (0);
3677 #endif
3678
3679 case DIF_VAR_UID:
3680 if (!dtrace_priv_proc(state))
3681 return (0);
3682
3683 #ifdef illumos
3684 /*
3685 * See comment in DIF_VAR_PID.
3686 */
3687 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3688 return ((uint64_t)p0.p_cred->cr_uid);
3689
3690 /*
3691 * It is always safe to dereference one's own t_procp pointer:
3692 * it always points to a valid, allocated proc structure.
3693 * (This is true because threads don't clean up their own
3694 * state -- they leave that task to whomever reaps them.)
3695 *
3696 * Additionally, it is safe to dereference one's own process
3697 * credential, since this is never NULL after process birth.
3698 */
3699 return ((uint64_t)curthread->t_procp->p_cred->cr_uid);
3700 #else
3701 return ((uint64_t)curthread->td_ucred->cr_uid);
3702 #endif
3703
3704 case DIF_VAR_GID:
3705 if (!dtrace_priv_proc(state))
3706 return (0);
3707
3708 #ifdef illumos
3709 /*
3710 * See comment in DIF_VAR_PID.
3711 */
3712 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3713 return ((uint64_t)p0.p_cred->cr_gid);
3714
3715 /*
3716 * It is always safe to dereference one's own t_procp pointer:
3717 * it always points to a valid, allocated proc structure.
3718 * (This is true because threads don't clean up their own
3719 * state -- they leave that task to whomever reaps them.)
3720 *
3721 * Additionally, it is safe to dereference one's own process
3722 * credential, since this is never NULL after process birth.
3723 */
3724 return ((uint64_t)curthread->t_procp->p_cred->cr_gid);
3725 #else
3726 return ((uint64_t)curthread->td_ucred->cr_gid);
3727 #endif
3728
3729 case DIF_VAR_ERRNO: {
3730 #ifdef illumos
3731 klwp_t *lwp;
3732 if (!dtrace_priv_proc(state))
3733 return (0);
3734
3735 /*
3736 * See comment in DIF_VAR_PID.
3737 */
3738 if (DTRACE_ANCHORED(mstate->dtms_probe) && CPU_ON_INTR(CPU))
3739 return (0);
3740
3741 /*
3742 * It is always safe to dereference one's own t_lwp pointer in
3743 * the event that this pointer is non-NULL. (This is true
3744 * because threads and lwps don't clean up their own state --
3745 * they leave that task to whomever reaps them.)
3746 */
3747 if ((lwp = curthread->t_lwp) == NULL)
3748 return (0);
3749
3750 return ((uint64_t)lwp->lwp_errno);
3751 #else
3752 return (curthread->td_errno);
3753 #endif
3754 }
3755 #ifndef illumos
3756 case DIF_VAR_CPU: {
3757 return curcpu;
3758 }
3759 #endif
3760 default:
3761 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
3762 return (0);
3763 }
3764 }
3765
3766
3767 typedef enum dtrace_json_state {
3768 DTRACE_JSON_REST = 1,
3769 DTRACE_JSON_OBJECT,
3770 DTRACE_JSON_STRING,
3771 DTRACE_JSON_STRING_ESCAPE,
3772 DTRACE_JSON_STRING_ESCAPE_UNICODE,
3773 DTRACE_JSON_COLON,
3774 DTRACE_JSON_COMMA,
3775 DTRACE_JSON_VALUE,
3776 DTRACE_JSON_IDENTIFIER,
3777 DTRACE_JSON_NUMBER,
3778 DTRACE_JSON_NUMBER_FRAC,
3779 DTRACE_JSON_NUMBER_EXP,
3780 DTRACE_JSON_COLLECT_OBJECT
3781 } dtrace_json_state_t;
3782
3783 /*
3784 * This function possesses just enough knowledge about JSON to extract a single
3785 * value from a JSON string and store it in the scratch buffer. It is able
3786 * to extract nested object values, and members of arrays by index.
3787 *
3788 * elemlist is a list of JSON keys, stored as packed NUL-terminated strings, to
3789 * be looked up as we descend into the object tree. e.g.
3790 *
3791 * foo[0].bar.baz[32] --> "foo" NUL "" NUL "bar" NUL "baz" NUL "32" NUL
3792 * with nelems = 5.
3793 *
3794 * The run time of this function must be bounded above by strsize to limit the
3795 * amount of work done in probe context. As such, it is implemented as a
3796 * simple state machine, reading one character at a time using safe loads
3797 * until we find the requested element, hit a parsing error or run off the
3798 * end of the object or string.
3799 *
3800 * As there is no way for a subroutine to return an error without interrupting
3801 * clause execution, we simply return NULL in the event of a missing key or any
3802 * other error condition. Each NULL return in this function is commented with
3803 * the error condition it represents -- parsing or otherwise.
3804 *
3805 * The set of states for the state machine closely matches the JSON
3806 * specification (http://json.org/). Briefly:
3807 *
3808 * DTRACE_JSON_REST:
3809 * Skip whitespace until we find either a top-level Object, moving
3810 * to DTRACE_JSON_OBJECT; or an Array, moving to DTRACE_JSON_VALUE.
3811 *
3812 * DTRACE_JSON_OBJECT:
3813 * Locate the next key String in an Object. Sets a flag to denote
3814 * the next String as a key string and moves to DTRACE_JSON_STRING.
3815 *
3816 * DTRACE_JSON_COLON:
3817 * Skip whitespace until we find the colon that separates key Strings
3818 * from their values. Once found, move to DTRACE_JSON_VALUE.
3819 *
3820 * DTRACE_JSON_VALUE:
3821 * Detects the type of the next value (String, Number, Identifier, Object
3822 * or Array) and routes to the states that process that type. Here we also
3823 * deal with the element selector list if we are requested to traverse down
3824 * into the object tree.
3825 *
3826 * DTRACE_JSON_COMMA:
3827 * Skip whitespace until we find the comma that separates key-value pairs
3828 * in Objects (returning to DTRACE_JSON_OBJECT) or values in Arrays
3829 * (similarly DTRACE_JSON_VALUE). All following literal value processing
3830 * states return to this state at the end of their value, unless otherwise
3831 * noted.
3832 *
3833 * DTRACE_JSON_NUMBER, DTRACE_JSON_NUMBER_FRAC, DTRACE_JSON_NUMBER_EXP:
3834 * Processes a Number literal from the JSON, including any exponent
3835 * component that may be present. Numbers are returned as strings, which
3836 * may be passed to strtoll() if an integer is required.
3837 *
3838 * DTRACE_JSON_IDENTIFIER:
3839 * Processes a "true", "false" or "null" literal in the JSON.
3840 *
3841 * DTRACE_JSON_STRING, DTRACE_JSON_STRING_ESCAPE,
3842 * DTRACE_JSON_STRING_ESCAPE_UNICODE:
3843 * Processes a String literal from the JSON, whether the String denotes
3844 * a key, a value or part of a larger Object. Handles all escape sequences
3845 * present in the specification, including four-digit unicode characters,
3846 * but merely includes the escape sequence without converting it to the
3847 * actual escaped character. If the String is flagged as a key, we
3848 * move to DTRACE_JSON_COLON rather than DTRACE_JSON_COMMA.
3849 *
3850 * DTRACE_JSON_COLLECT_OBJECT:
3851 * This state collects an entire Object (or Array), correctly handling
3852 * embedded strings. If the full element selector list matches this nested
3853 * object, we return the Object in full as a string. If not, we use this
3854 * state to skip to the next value at this level and continue processing.
3855 *
3856 * NOTE: This function uses various macros from strtolctype.h to manipulate
3857 * digit values, etc -- these have all been checked to ensure they make
3858 * no additional function calls.
3859 */
3860 static char *
3861 dtrace_json(uint64_t size, uintptr_t json, char *elemlist, int nelems,
3862 char *dest)
3863 {
3864 dtrace_json_state_t state = DTRACE_JSON_REST;
3865 int64_t array_elem = INT64_MIN;
3866 int64_t array_pos = 0;
3867 uint8_t escape_unicount = 0;
3868 boolean_t string_is_key = B_FALSE;
3869 boolean_t collect_object = B_FALSE;
3870 boolean_t found_key = B_FALSE;
3871 boolean_t in_array = B_FALSE;
3872 uint32_t braces = 0, brackets = 0;
3873 char *elem = elemlist;
3874 char *dd = dest;
3875 uintptr_t cur;
3876
3877 for (cur = json; cur < json + size; cur++) {
3878 char cc = dtrace_load8(cur);
3879 if (cc == '\0')
3880 return (NULL);
3881
3882 switch (state) {
3883 case DTRACE_JSON_REST:
3884 if (isspace(cc))
3885 break;
3886
3887 if (cc == '{') {
3888 state = DTRACE_JSON_OBJECT;
3889 break;
3890 }
3891
3892 if (cc == '[') {
3893 in_array = B_TRUE;
3894 array_pos = 0;
3895 array_elem = dtrace_strtoll(elem, 10, size);
3896 found_key = array_elem == 0 ? B_TRUE : B_FALSE;
3897 state = DTRACE_JSON_VALUE;
3898 break;
3899 }
3900
3901 /*
3902 * ERROR: expected to find a top-level object or array.
3903 */
3904 return (NULL);
3905 case DTRACE_JSON_OBJECT:
3906 if (isspace(cc))
3907 break;
3908
3909 if (cc == '"') {
3910 state = DTRACE_JSON_STRING;
3911 string_is_key = B_TRUE;
3912 break;
3913 }
3914
3915 /*
3916 * ERROR: either the object did not start with a key
3917 * string, or we've run off the end of the object
3918 * without finding the requested key.
3919 */
3920 return (NULL);
3921 case DTRACE_JSON_STRING:
3922 if (cc == '\\') {
3923 *dd++ = '\\';
3924 state = DTRACE_JSON_STRING_ESCAPE;
3925 break;
3926 }
3927
3928 if (cc == '"') {
3929 if (collect_object) {
3930 /*
3931 * We don't reset the dest here, as
3932 * the string is part of a larger
3933 * object being collected.
3934 */
3935 *dd++ = cc;
3936 collect_object = B_FALSE;
3937 state = DTRACE_JSON_COLLECT_OBJECT;
3938 break;
3939 }
3940 *dd = '\0';
3941 dd = dest; /* reset string buffer */
3942 if (string_is_key) {
3943 if (dtrace_strncmp(dest, elem,
3944 size) == 0)
3945 found_key = B_TRUE;
3946 } else if (found_key) {
3947 if (nelems > 1) {
3948 /*
3949 * We expected an object, not
3950 * this string.
3951 */
3952 return (NULL);
3953 }
3954 return (dest);
3955 }
3956 state = string_is_key ? DTRACE_JSON_COLON :
3957 DTRACE_JSON_COMMA;
3958 string_is_key = B_FALSE;
3959 break;
3960 }
3961
3962 *dd++ = cc;
3963 break;
3964 case DTRACE_JSON_STRING_ESCAPE:
3965 *dd++ = cc;
3966 if (cc == 'u') {
3967 escape_unicount = 0;
3968 state = DTRACE_JSON_STRING_ESCAPE_UNICODE;
3969 } else {
3970 state = DTRACE_JSON_STRING;
3971 }
3972 break;
3973 case DTRACE_JSON_STRING_ESCAPE_UNICODE:
3974 if (!isxdigit(cc)) {
3975 /*
3976 * ERROR: invalid unicode escape, expected
3977 * four valid hexidecimal digits.
3978 */
3979 return (NULL);
3980 }
3981
3982 *dd++ = cc;
3983 if (++escape_unicount == 4)
3984 state = DTRACE_JSON_STRING;
3985 break;
3986 case DTRACE_JSON_COLON:
3987 if (isspace(cc))
3988 break;
3989
3990 if (cc == ':') {
3991 state = DTRACE_JSON_VALUE;
3992 break;
3993 }
3994
3995 /*
3996 * ERROR: expected a colon.
3997 */
3998 return (NULL);
3999 case DTRACE_JSON_COMMA:
4000 if (isspace(cc))
4001 break;
4002
4003 if (cc == ',') {
4004 if (in_array) {
4005 state = DTRACE_JSON_VALUE;
4006 if (++array_pos == array_elem)
4007 found_key = B_TRUE;
4008 } else {
4009 state = DTRACE_JSON_OBJECT;
4010 }
4011 break;
4012 }
4013
4014 /*
4015 * ERROR: either we hit an unexpected character, or
4016 * we reached the end of the object or array without
4017 * finding the requested key.
4018 */
4019 return (NULL);
4020 case DTRACE_JSON_IDENTIFIER:
4021 if (islower(cc)) {
4022 *dd++ = cc;
4023 break;
4024 }
4025
4026 *dd = '\0';
4027 dd = dest; /* reset string buffer */
4028
4029 if (dtrace_strncmp(dest, "true", 5) == 0 ||
4030 dtrace_strncmp(dest, "false", 6) == 0 ||
4031 dtrace_strncmp(dest, "null", 5) == 0) {
4032 if (found_key) {
4033 if (nelems > 1) {
4034 /*
4035 * ERROR: We expected an object,
4036 * not this identifier.
4037 */
4038 return (NULL);
4039 }
4040 return (dest);
4041 } else {
4042 cur--;
4043 state = DTRACE_JSON_COMMA;
4044 break;
4045 }
4046 }
4047
4048 /*
4049 * ERROR: we did not recognise the identifier as one
4050 * of those in the JSON specification.
4051 */
4052 return (NULL);
4053 case DTRACE_JSON_NUMBER:
4054 if (cc == '.') {
4055 *dd++ = cc;
4056 state = DTRACE_JSON_NUMBER_FRAC;
4057 break;
4058 }
4059
4060 if (cc == 'x' || cc == 'X') {
4061 /*
4062 * ERROR: specification explicitly excludes
4063 * hexidecimal or octal numbers.
4064 */
4065 return (NULL);
4066 }
4067
4068 /* FALLTHRU */
4069 case DTRACE_JSON_NUMBER_FRAC:
4070 if (cc == 'e' || cc == 'E') {
4071 *dd++ = cc;
4072 state = DTRACE_JSON_NUMBER_EXP;
4073 break;
4074 }
4075
4076 if (cc == '+' || cc == '-') {
4077 /*
4078 * ERROR: expect sign as part of exponent only.
4079 */
4080 return (NULL);
4081 }
4082 /* FALLTHRU */
4083 case DTRACE_JSON_NUMBER_EXP:
4084 if (isdigit(cc) || cc == '+' || cc == '-') {
4085 *dd++ = cc;
4086 break;
4087 }
4088
4089 *dd = '\0';
4090 dd = dest; /* reset string buffer */
4091 if (found_key) {
4092 if (nelems > 1) {
4093 /*
4094 * ERROR: We expected an object, not
4095 * this number.
4096 */
4097 return (NULL);
4098 }
4099 return (dest);
4100 }
4101
4102 cur--;
4103 state = DTRACE_JSON_COMMA;
4104 break;
4105 case DTRACE_JSON_VALUE:
4106 if (isspace(cc))
4107 break;
4108
4109 if (cc == '{' || cc == '[') {
4110 if (nelems > 1 && found_key) {
4111 in_array = cc == '[' ? B_TRUE : B_FALSE;
4112 /*
4113 * If our element selector directs us
4114 * to descend into this nested object,
4115 * then move to the next selector
4116 * element in the list and restart the
4117 * state machine.
4118 */
4119 while (*elem != '\0')
4120 elem++;
4121 elem++; /* skip the inter-element NUL */
4122 nelems--;
4123 dd = dest;
4124 if (in_array) {
4125 state = DTRACE_JSON_VALUE;
4126 array_pos = 0;
4127 array_elem = dtrace_strtoll(
4128 elem, 10, size);
4129 found_key = array_elem == 0 ?
4130 B_TRUE : B_FALSE;
4131 } else {
4132 found_key = B_FALSE;
4133 state = DTRACE_JSON_OBJECT;
4134 }
4135 break;
4136 }
4137
4138 /*
4139 * Otherwise, we wish to either skip this
4140 * nested object or return it in full.
4141 */
4142 if (cc == '[')
4143 brackets = 1;
4144 else
4145 braces = 1;
4146 *dd++ = cc;
4147 state = DTRACE_JSON_COLLECT_OBJECT;
4148 break;
4149 }
4150
4151 if (cc == '"') {
4152 state = DTRACE_JSON_STRING;
4153 break;
4154 }
4155
4156 if (islower(cc)) {
4157 /*
4158 * Here we deal with true, false and null.
4159 */
4160 *dd++ = cc;
4161 state = DTRACE_JSON_IDENTIFIER;
4162 break;
4163 }
4164
4165 if (cc == '-' || isdigit(cc)) {
4166 *dd++ = cc;
4167 state = DTRACE_JSON_NUMBER;
4168 break;
4169 }
4170
4171 /*
4172 * ERROR: unexpected character at start of value.
4173 */
4174 return (NULL);
4175 case DTRACE_JSON_COLLECT_OBJECT:
4176 if (cc == '\0')
4177 /*
4178 * ERROR: unexpected end of input.
4179 */
4180 return (NULL);
4181
4182 *dd++ = cc;
4183 if (cc == '"') {
4184 collect_object = B_TRUE;
4185 state = DTRACE_JSON_STRING;
4186 break;
4187 }
4188
4189 if (cc == ']') {
4190 if (brackets-- == 0) {
4191 /*
4192 * ERROR: unbalanced brackets.
4193 */
4194 return (NULL);
4195 }
4196 } else if (cc == '}') {
4197 if (braces-- == 0) {
4198 /*
4199 * ERROR: unbalanced braces.
4200 */
4201 return (NULL);
4202 }
4203 } else if (cc == '{') {
4204 braces++;
4205 } else if (cc == '[') {
4206 brackets++;
4207 }
4208
4209 if (brackets == 0 && braces == 0) {
4210 if (found_key) {
4211 *dd = '\0';
4212 return (dest);
4213 }
4214 dd = dest; /* reset string buffer */
4215 state = DTRACE_JSON_COMMA;
4216 }
4217 break;
4218 }
4219 }
4220 return (NULL);
4221 }
4222
4223 /*
4224 * Emulate the execution of DTrace ID subroutines invoked by the call opcode.
4225 * Notice that we don't bother validating the proper number of arguments or
4226 * their types in the tuple stack. This isn't needed because all argument
4227 * interpretation is safe because of our load safety -- the worst that can
4228 * happen is that a bogus program can obtain bogus results.
4229 */
4230 static void
4231 dtrace_dif_subr(uint_t subr, uint_t rd, uint64_t *regs,
4232 dtrace_key_t *tupregs, int nargs,
4233 dtrace_mstate_t *mstate, dtrace_state_t *state)
4234 {
4235 volatile uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags;
4236 volatile uintptr_t *illval = &cpu_core[curcpu].cpuc_dtrace_illval;
4237 dtrace_vstate_t *vstate = &state->dts_vstate;
4238
4239 #ifdef illumos
4240 union {
4241 mutex_impl_t mi;
4242 uint64_t mx;
4243 } m;
4244
4245 union {
4246 krwlock_t ri;
4247 uintptr_t rw;
4248 } r;
4249 #else
4250 struct thread *lowner;
4251 union {
4252 struct lock_object *li;
4253 uintptr_t lx;
4254 } l;
4255 #endif
4256
4257 switch (subr) {
4258 case DIF_SUBR_RAND:
4259 regs[rd] = dtrace_xoroshiro128_plus_next(
4260 state->dts_rstate[curcpu]);
4261 break;
4262
4263 #ifdef illumos
4264 case DIF_SUBR_MUTEX_OWNED:
4265 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
4266 mstate, vstate)) {
4267 regs[rd] = 0;
4268 break;
4269 }
4270
4271 m.mx = dtrace_load64(tupregs[0].dttk_value);
4272 if (MUTEX_TYPE_ADAPTIVE(&m.mi))
4273 regs[rd] = MUTEX_OWNER(&m.mi) != MUTEX_NO_OWNER;
4274 else
4275 regs[rd] = LOCK_HELD(&m.mi.m_spin.m_spinlock);
4276 break;
4277
4278 case DIF_SUBR_MUTEX_OWNER:
4279 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
4280 mstate, vstate)) {
4281 regs[rd] = 0;
4282 break;
4283 }
4284
4285 m.mx = dtrace_load64(tupregs[0].dttk_value);
4286 if (MUTEX_TYPE_ADAPTIVE(&m.mi) &&
4287 MUTEX_OWNER(&m.mi) != MUTEX_NO_OWNER)
4288 regs[rd] = (uintptr_t)MUTEX_OWNER(&m.mi);
4289 else
4290 regs[rd] = 0;
4291 break;
4292
4293 case DIF_SUBR_MUTEX_TYPE_ADAPTIVE:
4294 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
4295 mstate, vstate)) {
4296 regs[rd] = 0;
4297 break;
4298 }
4299
4300 m.mx = dtrace_load64(tupregs[0].dttk_value);
4301 regs[rd] = MUTEX_TYPE_ADAPTIVE(&m.mi);
4302 break;
4303
4304 case DIF_SUBR_MUTEX_TYPE_SPIN:
4305 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (kmutex_t),
4306 mstate, vstate)) {
4307 regs[rd] = 0;
4308 break;
4309 }
4310
4311 m.mx = dtrace_load64(tupregs[0].dttk_value);
4312 regs[rd] = MUTEX_TYPE_SPIN(&m.mi);
4313 break;
4314
4315 case DIF_SUBR_RW_READ_HELD: {
4316 uintptr_t tmp;
4317
4318 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
4319 mstate, vstate)) {
4320 regs[rd] = 0;
4321 break;
4322 }
4323
4324 r.rw = dtrace_loadptr(tupregs[0].dttk_value);
4325 regs[rd] = _RW_READ_HELD(&r.ri, tmp);
4326 break;
4327 }
4328
4329 case DIF_SUBR_RW_WRITE_HELD:
4330 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (krwlock_t),
4331 mstate, vstate)) {
4332 regs[rd] = 0;
4333 break;
4334 }
4335
4336 r.rw = dtrace_loadptr(tupregs[0].dttk_value);
4337 regs[rd] = _RW_WRITE_HELD(&r.ri);
4338 break;
4339
4340 case DIF_SUBR_RW_ISWRITER:
4341 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (krwlock_t),
4342 mstate, vstate)) {
4343 regs[rd] = 0;
4344 break;
4345 }
4346
4347 r.rw = dtrace_loadptr(tupregs[0].dttk_value);
4348 regs[rd] = _RW_ISWRITER(&r.ri);
4349 break;
4350
4351 #else /* !illumos */
4352 case DIF_SUBR_MUTEX_OWNED:
4353 if (!dtrace_canload(tupregs[0].dttk_value,
4354 sizeof (struct lock_object), mstate, vstate)) {
4355 regs[rd] = 0;
4356 break;
4357 }
4358 l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
4359 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4360 regs[rd] = LOCK_CLASS(l.li)->lc_owner(l.li, &lowner);
4361 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4362 break;
4363
4364 case DIF_SUBR_MUTEX_OWNER:
4365 if (!dtrace_canload(tupregs[0].dttk_value,
4366 sizeof (struct lock_object), mstate, vstate)) {
4367 regs[rd] = 0;
4368 break;
4369 }
4370 l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
4371 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4372 LOCK_CLASS(l.li)->lc_owner(l.li, &lowner);
4373 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4374 regs[rd] = (uintptr_t)lowner;
4375 break;
4376
4377 case DIF_SUBR_MUTEX_TYPE_ADAPTIVE:
4378 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (struct mtx),
4379 mstate, vstate)) {
4380 regs[rd] = 0;
4381 break;
4382 }
4383 l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
4384 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4385 regs[rd] = (LOCK_CLASS(l.li)->lc_flags & LC_SLEEPLOCK) != 0;
4386 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4387 break;
4388
4389 case DIF_SUBR_MUTEX_TYPE_SPIN:
4390 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (struct mtx),
4391 mstate, vstate)) {
4392 regs[rd] = 0;
4393 break;
4394 }
4395 l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
4396 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4397 regs[rd] = (LOCK_CLASS(l.li)->lc_flags & LC_SPINLOCK) != 0;
4398 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4399 break;
4400
4401 case DIF_SUBR_RW_READ_HELD:
4402 case DIF_SUBR_SX_SHARED_HELD:
4403 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
4404 mstate, vstate)) {
4405 regs[rd] = 0;
4406 break;
4407 }
4408 l.lx = dtrace_loadptr((uintptr_t)&tupregs[0].dttk_value);
4409 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4410 regs[rd] = LOCK_CLASS(l.li)->lc_owner(l.li, &lowner) &&
4411 lowner == NULL;
4412 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4413 break;
4414
4415 case DIF_SUBR_RW_WRITE_HELD:
4416 case DIF_SUBR_SX_EXCLUSIVE_HELD:
4417 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
4418 mstate, vstate)) {
4419 regs[rd] = 0;
4420 break;
4421 }
4422 l.lx = dtrace_loadptr(tupregs[0].dttk_value);
4423 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4424 regs[rd] = LOCK_CLASS(l.li)->lc_owner(l.li, &lowner) &&
4425 lowner != NULL;
4426 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4427 break;
4428
4429 case DIF_SUBR_RW_ISWRITER:
4430 case DIF_SUBR_SX_ISEXCLUSIVE:
4431 if (!dtrace_canload(tupregs[0].dttk_value, sizeof (uintptr_t),
4432 mstate, vstate)) {
4433 regs[rd] = 0;
4434 break;
4435 }
4436 l.lx = dtrace_loadptr(tupregs[0].dttk_value);
4437 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4438 LOCK_CLASS(l.li)->lc_owner(l.li, &lowner);
4439 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4440 regs[rd] = (lowner == curthread);
4441 break;
4442 #endif /* illumos */
4443
4444 case DIF_SUBR_BCOPY: {
4445 /*
4446 * We need to be sure that the destination is in the scratch
4447 * region -- no other region is allowed.
4448 */
4449 uintptr_t src = tupregs[0].dttk_value;
4450 uintptr_t dest = tupregs[1].dttk_value;
4451 size_t size = tupregs[2].dttk_value;
4452
4453 if (!dtrace_inscratch(dest, size, mstate)) {
4454 *flags |= CPU_DTRACE_BADADDR;
4455 *illval = regs[rd];
4456 break;
4457 }
4458
4459 if (!dtrace_canload(src, size, mstate, vstate)) {
4460 regs[rd] = 0;
4461 break;
4462 }
4463
4464 dtrace_bcopy((void *)src, (void *)dest, size);
4465 break;
4466 }
4467
4468 case DIF_SUBR_ALLOCA:
4469 case DIF_SUBR_COPYIN: {
4470 uintptr_t dest = P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
4471 uint64_t size =
4472 tupregs[subr == DIF_SUBR_ALLOCA ? 0 : 1].dttk_value;
4473 size_t scratch_size = (dest - mstate->dtms_scratch_ptr) + size;
4474
4475 /*
4476 * This action doesn't require any credential checks since
4477 * probes will not activate in user contexts to which the
4478 * enabling user does not have permissions.
4479 */
4480
4481 /*
4482 * Rounding up the user allocation size could have overflowed
4483 * a large, bogus allocation (like -1ULL) to 0.
4484 */
4485 if (scratch_size < size ||
4486 !DTRACE_INSCRATCH(mstate, scratch_size)) {
4487 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4488 regs[rd] = 0;
4489 break;
4490 }
4491
4492 if (subr == DIF_SUBR_COPYIN) {
4493 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4494 dtrace_copyin(tupregs[0].dttk_value, dest, size, flags);
4495 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4496 }
4497
4498 mstate->dtms_scratch_ptr += scratch_size;
4499 regs[rd] = dest;
4500 break;
4501 }
4502
4503 case DIF_SUBR_COPYINTO: {
4504 uint64_t size = tupregs[1].dttk_value;
4505 uintptr_t dest = tupregs[2].dttk_value;
4506
4507 /*
4508 * This action doesn't require any credential checks since
4509 * probes will not activate in user contexts to which the
4510 * enabling user does not have permissions.
4511 */
4512 if (!dtrace_inscratch(dest, size, mstate)) {
4513 *flags |= CPU_DTRACE_BADADDR;
4514 *illval = regs[rd];
4515 break;
4516 }
4517
4518 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4519 dtrace_copyin(tupregs[0].dttk_value, dest, size, flags);
4520 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4521 break;
4522 }
4523
4524 case DIF_SUBR_COPYINSTR: {
4525 uintptr_t dest = mstate->dtms_scratch_ptr;
4526 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4527
4528 if (nargs > 1 && tupregs[1].dttk_value < size)
4529 size = tupregs[1].dttk_value + 1;
4530
4531 /*
4532 * This action doesn't require any credential checks since
4533 * probes will not activate in user contexts to which the
4534 * enabling user does not have permissions.
4535 */
4536 if (!DTRACE_INSCRATCH(mstate, size)) {
4537 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4538 regs[rd] = 0;
4539 break;
4540 }
4541
4542 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4543 dtrace_copyinstr(tupregs[0].dttk_value, dest, size, flags);
4544 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4545
4546 ((char *)dest)[size - 1] = '\0';
4547 mstate->dtms_scratch_ptr += size;
4548 regs[rd] = dest;
4549 break;
4550 }
4551
4552 #ifdef illumos
4553 case DIF_SUBR_MSGSIZE:
4554 case DIF_SUBR_MSGDSIZE: {
4555 uintptr_t baddr = tupregs[0].dttk_value, daddr;
4556 uintptr_t wptr, rptr;
4557 size_t count = 0;
4558 int cont = 0;
4559
4560 while (baddr != 0 && !(*flags & CPU_DTRACE_FAULT)) {
4561
4562 if (!dtrace_canload(baddr, sizeof (mblk_t), mstate,
4563 vstate)) {
4564 regs[rd] = 0;
4565 break;
4566 }
4567
4568 wptr = dtrace_loadptr(baddr +
4569 offsetof(mblk_t, b_wptr));
4570
4571 rptr = dtrace_loadptr(baddr +
4572 offsetof(mblk_t, b_rptr));
4573
4574 if (wptr < rptr) {
4575 *flags |= CPU_DTRACE_BADADDR;
4576 *illval = tupregs[0].dttk_value;
4577 break;
4578 }
4579
4580 daddr = dtrace_loadptr(baddr +
4581 offsetof(mblk_t, b_datap));
4582
4583 baddr = dtrace_loadptr(baddr +
4584 offsetof(mblk_t, b_cont));
4585
4586 /*
4587 * We want to prevent against denial-of-service here,
4588 * so we're only going to search the list for
4589 * dtrace_msgdsize_max mblks.
4590 */
4591 if (cont++ > dtrace_msgdsize_max) {
4592 *flags |= CPU_DTRACE_ILLOP;
4593 break;
4594 }
4595
4596 if (subr == DIF_SUBR_MSGDSIZE) {
4597 if (dtrace_load8(daddr +
4598 offsetof(dblk_t, db_type)) != M_DATA)
4599 continue;
4600 }
4601
4602 count += wptr - rptr;
4603 }
4604
4605 if (!(*flags & CPU_DTRACE_FAULT))
4606 regs[rd] = count;
4607
4608 break;
4609 }
4610 #endif
4611
4612 case DIF_SUBR_PROGENYOF: {
4613 pid_t pid = tupregs[0].dttk_value;
4614 proc_t *p;
4615 int rval = 0;
4616
4617 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4618
4619 for (p = curthread->t_procp; p != NULL; p = p->p_parent) {
4620 #ifdef illumos
4621 if (p->p_pidp->pid_id == pid) {
4622 #else
4623 if (p->p_pid == pid) {
4624 #endif
4625 rval = 1;
4626 break;
4627 }
4628 }
4629
4630 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4631
4632 regs[rd] = rval;
4633 break;
4634 }
4635
4636 case DIF_SUBR_SPECULATION:
4637 regs[rd] = dtrace_speculation(state);
4638 break;
4639
4640 case DIF_SUBR_COPYOUT: {
4641 uintptr_t kaddr = tupregs[0].dttk_value;
4642 uintptr_t uaddr = tupregs[1].dttk_value;
4643 uint64_t size = tupregs[2].dttk_value;
4644
4645 if (!dtrace_destructive_disallow &&
4646 dtrace_priv_proc_control(state) &&
4647 !dtrace_istoxic(kaddr, size) &&
4648 dtrace_canload(kaddr, size, mstate, vstate)) {
4649 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4650 dtrace_copyout(kaddr, uaddr, size, flags);
4651 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4652 }
4653 break;
4654 }
4655
4656 case DIF_SUBR_COPYOUTSTR: {
4657 uintptr_t kaddr = tupregs[0].dttk_value;
4658 uintptr_t uaddr = tupregs[1].dttk_value;
4659 uint64_t size = tupregs[2].dttk_value;
4660 size_t lim;
4661
4662 if (!dtrace_destructive_disallow &&
4663 dtrace_priv_proc_control(state) &&
4664 !dtrace_istoxic(kaddr, size) &&
4665 dtrace_strcanload(kaddr, size, &lim, mstate, vstate)) {
4666 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
4667 dtrace_copyoutstr(kaddr, uaddr, lim, flags);
4668 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
4669 }
4670 break;
4671 }
4672
4673 case DIF_SUBR_STRLEN: {
4674 size_t size = state->dts_options[DTRACEOPT_STRSIZE];
4675 uintptr_t addr = (uintptr_t)tupregs[0].dttk_value;
4676 size_t lim;
4677
4678 if (!dtrace_strcanload(addr, size, &lim, mstate, vstate)) {
4679 regs[rd] = 0;
4680 break;
4681 }
4682
4683 regs[rd] = dtrace_strlen((char *)addr, lim);
4684 break;
4685 }
4686
4687 case DIF_SUBR_STRCHR:
4688 case DIF_SUBR_STRRCHR: {
4689 /*
4690 * We're going to iterate over the string looking for the
4691 * specified character. We will iterate until we have reached
4692 * the string length or we have found the character. If this
4693 * is DIF_SUBR_STRRCHR, we will look for the last occurrence
4694 * of the specified character instead of the first.
4695 */
4696 uintptr_t addr = tupregs[0].dttk_value;
4697 uintptr_t addr_limit;
4698 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4699 size_t lim;
4700 char c, target = (char)tupregs[1].dttk_value;
4701
4702 if (!dtrace_strcanload(addr, size, &lim, mstate, vstate)) {
4703 regs[rd] = 0;
4704 break;
4705 }
4706 addr_limit = addr + lim;
4707
4708 for (regs[rd] = 0; addr < addr_limit; addr++) {
4709 if ((c = dtrace_load8(addr)) == target) {
4710 regs[rd] = addr;
4711
4712 if (subr == DIF_SUBR_STRCHR)
4713 break;
4714 }
4715
4716 if (c == '\0')
4717 break;
4718 }
4719 break;
4720 }
4721
4722 case DIF_SUBR_STRSTR:
4723 case DIF_SUBR_INDEX:
4724 case DIF_SUBR_RINDEX: {
4725 /*
4726 * We're going to iterate over the string looking for the
4727 * specified string. We will iterate until we have reached
4728 * the string length or we have found the string. (Yes, this
4729 * is done in the most naive way possible -- but considering
4730 * that the string we're searching for is likely to be
4731 * relatively short, the complexity of Rabin-Karp or similar
4732 * hardly seems merited.)
4733 */
4734 char *addr = (char *)(uintptr_t)tupregs[0].dttk_value;
4735 char *substr = (char *)(uintptr_t)tupregs[1].dttk_value;
4736 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4737 size_t len = dtrace_strlen(addr, size);
4738 size_t sublen = dtrace_strlen(substr, size);
4739 char *limit = addr + len, *orig = addr;
4740 int notfound = subr == DIF_SUBR_STRSTR ? 0 : -1;
4741 int inc = 1;
4742
4743 regs[rd] = notfound;
4744
4745 if (!dtrace_canload((uintptr_t)addr, len + 1, mstate, vstate)) {
4746 regs[rd] = 0;
4747 break;
4748 }
4749
4750 if (!dtrace_canload((uintptr_t)substr, sublen + 1, mstate,
4751 vstate)) {
4752 regs[rd] = 0;
4753 break;
4754 }
4755
4756 /*
4757 * strstr() and index()/rindex() have similar semantics if
4758 * both strings are the empty string: strstr() returns a
4759 * pointer to the (empty) string, and index() and rindex()
4760 * both return index 0 (regardless of any position argument).
4761 */
4762 if (sublen == 0 && len == 0) {
4763 if (subr == DIF_SUBR_STRSTR)
4764 regs[rd] = (uintptr_t)addr;
4765 else
4766 regs[rd] = 0;
4767 break;
4768 }
4769
4770 if (subr != DIF_SUBR_STRSTR) {
4771 if (subr == DIF_SUBR_RINDEX) {
4772 limit = orig - 1;
4773 addr += len;
4774 inc = -1;
4775 }
4776
4777 /*
4778 * Both index() and rindex() take an optional position
4779 * argument that denotes the starting position.
4780 */
4781 if (nargs == 3) {
4782 int64_t pos = (int64_t)tupregs[2].dttk_value;
4783
4784 /*
4785 * If the position argument to index() is
4786 * negative, Perl implicitly clamps it at
4787 * zero. This semantic is a little surprising
4788 * given the special meaning of negative
4789 * positions to similar Perl functions like
4790 * substr(), but it appears to reflect a
4791 * notion that index() can start from a
4792 * negative index and increment its way up to
4793 * the string. Given this notion, Perl's
4794 * rindex() is at least self-consistent in
4795 * that it implicitly clamps positions greater
4796 * than the string length to be the string
4797 * length. Where Perl completely loses
4798 * coherence, however, is when the specified
4799 * substring is the empty string (""). In
4800 * this case, even if the position is
4801 * negative, rindex() returns 0 -- and even if
4802 * the position is greater than the length,
4803 * index() returns the string length. These
4804 * semantics violate the notion that index()
4805 * should never return a value less than the
4806 * specified position and that rindex() should
4807 * never return a value greater than the
4808 * specified position. (One assumes that
4809 * these semantics are artifacts of Perl's
4810 * implementation and not the results of
4811 * deliberate design -- it beggars belief that
4812 * even Larry Wall could desire such oddness.)
4813 * While in the abstract one would wish for
4814 * consistent position semantics across
4815 * substr(), index() and rindex() -- or at the
4816 * very least self-consistent position
4817 * semantics for index() and rindex() -- we
4818 * instead opt to keep with the extant Perl
4819 * semantics, in all their broken glory. (Do
4820 * we have more desire to maintain Perl's
4821 * semantics than Perl does? Probably.)
4822 */
4823 if (subr == DIF_SUBR_RINDEX) {
4824 if (pos < 0) {
4825 if (sublen == 0)
4826 regs[rd] = 0;
4827 break;
4828 }
4829
4830 if (pos > len)
4831 pos = len;
4832 } else {
4833 if (pos < 0)
4834 pos = 0;
4835
4836 if (pos >= len) {
4837 if (sublen == 0)
4838 regs[rd] = len;
4839 break;
4840 }
4841 }
4842
4843 addr = orig + pos;
4844 }
4845 }
4846
4847 for (regs[rd] = notfound; addr != limit; addr += inc) {
4848 if (dtrace_strncmp(addr, substr, sublen) == 0) {
4849 if (subr != DIF_SUBR_STRSTR) {
4850 /*
4851 * As D index() and rindex() are
4852 * modeled on Perl (and not on awk),
4853 * we return a zero-based (and not a
4854 * one-based) index. (For you Perl
4855 * weenies: no, we're not going to add
4856 * $[ -- and shouldn't you be at a con
4857 * or something?)
4858 */
4859 regs[rd] = (uintptr_t)(addr - orig);
4860 break;
4861 }
4862
4863 ASSERT(subr == DIF_SUBR_STRSTR);
4864 regs[rd] = (uintptr_t)addr;
4865 break;
4866 }
4867 }
4868
4869 break;
4870 }
4871
4872 case DIF_SUBR_STRTOK: {
4873 uintptr_t addr = tupregs[0].dttk_value;
4874 uintptr_t tokaddr = tupregs[1].dttk_value;
4875 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4876 uintptr_t limit, toklimit;
4877 size_t clim;
4878 uint8_t c = 0, tokmap[32]; /* 256 / 8 */
4879 char *dest = (char *)mstate->dtms_scratch_ptr;
4880 int i;
4881
4882 /*
4883 * Check both the token buffer and (later) the input buffer,
4884 * since both could be non-scratch addresses.
4885 */
4886 if (!dtrace_strcanload(tokaddr, size, &clim, mstate, vstate)) {
4887 regs[rd] = 0;
4888 break;
4889 }
4890 toklimit = tokaddr + clim;
4891
4892 if (!DTRACE_INSCRATCH(mstate, size)) {
4893 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
4894 regs[rd] = 0;
4895 break;
4896 }
4897
4898 if (addr == 0) {
4899 /*
4900 * If the address specified is NULL, we use our saved
4901 * strtok pointer from the mstate. Note that this
4902 * means that the saved strtok pointer is _only_
4903 * valid within multiple enablings of the same probe --
4904 * it behaves like an implicit clause-local variable.
4905 */
4906 addr = mstate->dtms_strtok;
4907 limit = mstate->dtms_strtok_limit;
4908 } else {
4909 /*
4910 * If the user-specified address is non-NULL we must
4911 * access check it. This is the only time we have
4912 * a chance to do so, since this address may reside
4913 * in the string table of this clause-- future calls
4914 * (when we fetch addr from mstate->dtms_strtok)
4915 * would fail this access check.
4916 */
4917 if (!dtrace_strcanload(addr, size, &clim, mstate,
4918 vstate)) {
4919 regs[rd] = 0;
4920 break;
4921 }
4922 limit = addr + clim;
4923 }
4924
4925 /*
4926 * First, zero the token map, and then process the token
4927 * string -- setting a bit in the map for every character
4928 * found in the token string.
4929 */
4930 for (i = 0; i < sizeof (tokmap); i++)
4931 tokmap[i] = 0;
4932
4933 for (; tokaddr < toklimit; tokaddr++) {
4934 if ((c = dtrace_load8(tokaddr)) == '\0')
4935 break;
4936
4937 ASSERT((c >> 3) < sizeof (tokmap));
4938 tokmap[c >> 3] |= (1 << (c & 0x7));
4939 }
4940
4941 for (; addr < limit; addr++) {
4942 /*
4943 * We're looking for a character that is _not_
4944 * contained in the token string.
4945 */
4946 if ((c = dtrace_load8(addr)) == '\0')
4947 break;
4948
4949 if (!(tokmap[c >> 3] & (1 << (c & 0x7))))
4950 break;
4951 }
4952
4953 if (c == '\0') {
4954 /*
4955 * We reached the end of the string without finding
4956 * any character that was not in the token string.
4957 * We return NULL in this case, and we set the saved
4958 * address to NULL as well.
4959 */
4960 regs[rd] = 0;
4961 mstate->dtms_strtok = 0;
4962 mstate->dtms_strtok_limit = 0;
4963 break;
4964 }
4965
4966 /*
4967 * From here on, we're copying into the destination string.
4968 */
4969 for (i = 0; addr < limit && i < size - 1; addr++) {
4970 if ((c = dtrace_load8(addr)) == '\0')
4971 break;
4972
4973 if (tokmap[c >> 3] & (1 << (c & 0x7)))
4974 break;
4975
4976 ASSERT(i < size);
4977 dest[i++] = c;
4978 }
4979
4980 ASSERT(i < size);
4981 dest[i] = '\0';
4982 regs[rd] = (uintptr_t)dest;
4983 mstate->dtms_scratch_ptr += size;
4984 mstate->dtms_strtok = addr;
4985 mstate->dtms_strtok_limit = limit;
4986 break;
4987 }
4988
4989 case DIF_SUBR_SUBSTR: {
4990 uintptr_t s = tupregs[0].dttk_value;
4991 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
4992 char *d = (char *)mstate->dtms_scratch_ptr;
4993 int64_t index = (int64_t)tupregs[1].dttk_value;
4994 int64_t remaining = (int64_t)tupregs[2].dttk_value;
4995 size_t len = dtrace_strlen((char *)s, size);
4996 int64_t i;
4997
4998 if (!dtrace_canload(s, len + 1, mstate, vstate)) {
4999 regs[rd] = 0;
5000 break;
5001 }
5002
5003 if (!DTRACE_INSCRATCH(mstate, size)) {
5004 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5005 regs[rd] = 0;
5006 break;
5007 }
5008
5009 if (nargs <= 2)
5010 remaining = (int64_t)size;
5011
5012 if (index < 0) {
5013 index += len;
5014
5015 if (index < 0 && index + remaining > 0) {
5016 remaining += index;
5017 index = 0;
5018 }
5019 }
5020
5021 if (index >= len || index < 0) {
5022 remaining = 0;
5023 } else if (remaining < 0) {
5024 remaining += len - index;
5025 } else if (index + remaining > size) {
5026 remaining = size - index;
5027 }
5028
5029 for (i = 0; i < remaining; i++) {
5030 if ((d[i] = dtrace_load8(s + index + i)) == '\0')
5031 break;
5032 }
5033
5034 d[i] = '\0';
5035
5036 mstate->dtms_scratch_ptr += size;
5037 regs[rd] = (uintptr_t)d;
5038 break;
5039 }
5040
5041 case DIF_SUBR_JSON: {
5042 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
5043 uintptr_t json = tupregs[0].dttk_value;
5044 size_t jsonlen = dtrace_strlen((char *)json, size);
5045 uintptr_t elem = tupregs[1].dttk_value;
5046 size_t elemlen = dtrace_strlen((char *)elem, size);
5047
5048 char *dest = (char *)mstate->dtms_scratch_ptr;
5049 char *elemlist = (char *)mstate->dtms_scratch_ptr + jsonlen + 1;
5050 char *ee = elemlist;
5051 int nelems = 1;
5052 uintptr_t cur;
5053
5054 if (!dtrace_canload(json, jsonlen + 1, mstate, vstate) ||
5055 !dtrace_canload(elem, elemlen + 1, mstate, vstate)) {
5056 regs[rd] = 0;
5057 break;
5058 }
5059
5060 if (!DTRACE_INSCRATCH(mstate, jsonlen + 1 + elemlen + 1)) {
5061 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5062 regs[rd] = 0;
5063 break;
5064 }
5065
5066 /*
5067 * Read the element selector and split it up into a packed list
5068 * of strings.
5069 */
5070 for (cur = elem; cur < elem + elemlen; cur++) {
5071 char cc = dtrace_load8(cur);
5072
5073 if (cur == elem && cc == '[') {
5074 /*
5075 * If the first element selector key is
5076 * actually an array index then ignore the
5077 * bracket.
5078 */
5079 continue;
5080 }
5081
5082 if (cc == ']')
5083 continue;
5084
5085 if (cc == '.' || cc == '[') {
5086 nelems++;
5087 cc = '\0';
5088 }
5089
5090 *ee++ = cc;
5091 }
5092 *ee++ = '\0';
5093
5094 if ((regs[rd] = (uintptr_t)dtrace_json(size, json, elemlist,
5095 nelems, dest)) != 0)
5096 mstate->dtms_scratch_ptr += jsonlen + 1;
5097 break;
5098 }
5099
5100 case DIF_SUBR_TOUPPER:
5101 case DIF_SUBR_TOLOWER: {
5102 uintptr_t s = tupregs[0].dttk_value;
5103 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
5104 char *dest = (char *)mstate->dtms_scratch_ptr, c;
5105 size_t len = dtrace_strlen((char *)s, size);
5106 char lower, upper, convert;
5107 int64_t i;
5108
5109 if (subr == DIF_SUBR_TOUPPER) {
5110 lower = 'a';
5111 upper = 'z';
5112 convert = 'A';
5113 } else {
5114 lower = 'A';
5115 upper = 'Z';
5116 convert = 'a';
5117 }
5118
5119 if (!dtrace_canload(s, len + 1, mstate, vstate)) {
5120 regs[rd] = 0;
5121 break;
5122 }
5123
5124 if (!DTRACE_INSCRATCH(mstate, size)) {
5125 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5126 regs[rd] = 0;
5127 break;
5128 }
5129
5130 for (i = 0; i < size - 1; i++) {
5131 if ((c = dtrace_load8(s + i)) == '\0')
5132 break;
5133
5134 if (c >= lower && c <= upper)
5135 c = convert + (c - lower);
5136
5137 dest[i] = c;
5138 }
5139
5140 ASSERT(i < size);
5141 dest[i] = '\0';
5142 regs[rd] = (uintptr_t)dest;
5143 mstate->dtms_scratch_ptr += size;
5144 break;
5145 }
5146
5147 #ifdef illumos
5148 case DIF_SUBR_GETMAJOR:
5149 #ifdef _LP64
5150 regs[rd] = (tupregs[0].dttk_value >> NBITSMINOR64) & MAXMAJ64;
5151 #else
5152 regs[rd] = (tupregs[0].dttk_value >> NBITSMINOR) & MAXMAJ;
5153 #endif
5154 break;
5155
5156 case DIF_SUBR_GETMINOR:
5157 #ifdef _LP64
5158 regs[rd] = tupregs[0].dttk_value & MAXMIN64;
5159 #else
5160 regs[rd] = tupregs[0].dttk_value & MAXMIN;
5161 #endif
5162 break;
5163
5164 case DIF_SUBR_DDI_PATHNAME: {
5165 /*
5166 * This one is a galactic mess. We are going to roughly
5167 * emulate ddi_pathname(), but it's made more complicated
5168 * by the fact that we (a) want to include the minor name and
5169 * (b) must proceed iteratively instead of recursively.
5170 */
5171 uintptr_t dest = mstate->dtms_scratch_ptr;
5172 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
5173 char *start = (char *)dest, *end = start + size - 1;
5174 uintptr_t daddr = tupregs[0].dttk_value;
5175 int64_t minor = (int64_t)tupregs[1].dttk_value;
5176 char *s;
5177 int i, len, depth = 0;
5178
5179 /*
5180 * Due to all the pointer jumping we do and context we must
5181 * rely upon, we just mandate that the user must have kernel
5182 * read privileges to use this routine.
5183 */
5184 if ((mstate->dtms_access & DTRACE_ACCESS_KERNEL) == 0) {
5185 *flags |= CPU_DTRACE_KPRIV;
5186 *illval = daddr;
5187 regs[rd] = 0;
5188 }
5189
5190 if (!DTRACE_INSCRATCH(mstate, size)) {
5191 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5192 regs[rd] = 0;
5193 break;
5194 }
5195
5196 *end = '\0';
5197
5198 /*
5199 * We want to have a name for the minor. In order to do this,
5200 * we need to walk the minor list from the devinfo. We want
5201 * to be sure that we don't infinitely walk a circular list,
5202 * so we check for circularity by sending a scout pointer
5203 * ahead two elements for every element that we iterate over;
5204 * if the list is circular, these will ultimately point to the
5205 * same element. You may recognize this little trick as the
5206 * answer to a stupid interview question -- one that always
5207 * seems to be asked by those who had to have it laboriously
5208 * explained to them, and who can't even concisely describe
5209 * the conditions under which one would be forced to resort to
5210 * this technique. Needless to say, those conditions are
5211 * found here -- and probably only here. Is this the only use
5212 * of this infamous trick in shipping, production code? If it
5213 * isn't, it probably should be...
5214 */
5215 if (minor != -1) {
5216 uintptr_t maddr = dtrace_loadptr(daddr +
5217 offsetof(struct dev_info, devi_minor));
5218
5219 uintptr_t next = offsetof(struct ddi_minor_data, next);
5220 uintptr_t name = offsetof(struct ddi_minor_data,
5221 d_minor) + offsetof(struct ddi_minor, name);
5222 uintptr_t dev = offsetof(struct ddi_minor_data,
5223 d_minor) + offsetof(struct ddi_minor, dev);
5224 uintptr_t scout;
5225
5226 if (maddr != NULL)
5227 scout = dtrace_loadptr(maddr + next);
5228
5229 while (maddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
5230 uint64_t m;
5231 #ifdef _LP64
5232 m = dtrace_load64(maddr + dev) & MAXMIN64;
5233 #else
5234 m = dtrace_load32(maddr + dev) & MAXMIN;
5235 #endif
5236 if (m != minor) {
5237 maddr = dtrace_loadptr(maddr + next);
5238
5239 if (scout == NULL)
5240 continue;
5241
5242 scout = dtrace_loadptr(scout + next);
5243
5244 if (scout == NULL)
5245 continue;
5246
5247 scout = dtrace_loadptr(scout + next);
5248
5249 if (scout == NULL)
5250 continue;
5251
5252 if (scout == maddr) {
5253 *flags |= CPU_DTRACE_ILLOP;
5254 break;
5255 }
5256
5257 continue;
5258 }
5259
5260 /*
5261 * We have the minor data. Now we need to
5262 * copy the minor's name into the end of the
5263 * pathname.
5264 */
5265 s = (char *)dtrace_loadptr(maddr + name);
5266 len = dtrace_strlen(s, size);
5267
5268 if (*flags & CPU_DTRACE_FAULT)
5269 break;
5270
5271 if (len != 0) {
5272 if ((end -= (len + 1)) < start)
5273 break;
5274
5275 *end = ':';
5276 }
5277
5278 for (i = 1; i <= len; i++)
5279 end[i] = dtrace_load8((uintptr_t)s++);
5280 break;
5281 }
5282 }
5283
5284 while (daddr != NULL && !(*flags & CPU_DTRACE_FAULT)) {
5285 ddi_node_state_t devi_state;
5286
5287 devi_state = dtrace_load32(daddr +
5288 offsetof(struct dev_info, devi_node_state));
5289
5290 if (*flags & CPU_DTRACE_FAULT)
5291 break;
5292
5293 if (devi_state >= DS_INITIALIZED) {
5294 s = (char *)dtrace_loadptr(daddr +
5295 offsetof(struct dev_info, devi_addr));
5296 len = dtrace_strlen(s, size);
5297
5298 if (*flags & CPU_DTRACE_FAULT)
5299 break;
5300
5301 if (len != 0) {
5302 if ((end -= (len + 1)) < start)
5303 break;
5304
5305 *end = '@';
5306 }
5307
5308 for (i = 1; i <= len; i++)
5309 end[i] = dtrace_load8((uintptr_t)s++);
5310 }
5311
5312 /*
5313 * Now for the node name...
5314 */
5315 s = (char *)dtrace_loadptr(daddr +
5316 offsetof(struct dev_info, devi_node_name));
5317
5318 daddr = dtrace_loadptr(daddr +
5319 offsetof(struct dev_info, devi_parent));
5320
5321 /*
5322 * If our parent is NULL (that is, if we're the root
5323 * node), we're going to use the special path
5324 * "devices".
5325 */
5326 if (daddr == 0)
5327 s = "devices";
5328
5329 len = dtrace_strlen(s, size);
5330 if (*flags & CPU_DTRACE_FAULT)
5331 break;
5332
5333 if ((end -= (len + 1)) < start)
5334 break;
5335
5336 for (i = 1; i <= len; i++)
5337 end[i] = dtrace_load8((uintptr_t)s++);
5338 *end = '/';
5339
5340 if (depth++ > dtrace_devdepth_max) {
5341 *flags |= CPU_DTRACE_ILLOP;
5342 break;
5343 }
5344 }
5345
5346 if (end < start)
5347 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5348
5349 if (daddr == 0) {
5350 regs[rd] = (uintptr_t)end;
5351 mstate->dtms_scratch_ptr += size;
5352 }
5353
5354 break;
5355 }
5356 #endif
5357
5358 case DIF_SUBR_STRJOIN: {
5359 char *d = (char *)mstate->dtms_scratch_ptr;
5360 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
5361 uintptr_t s1 = tupregs[0].dttk_value;
5362 uintptr_t s2 = tupregs[1].dttk_value;
5363 int i = 0, j = 0;
5364 size_t lim1, lim2;
5365 char c;
5366
5367 if (!dtrace_strcanload(s1, size, &lim1, mstate, vstate) ||
5368 !dtrace_strcanload(s2, size, &lim2, mstate, vstate)) {
5369 regs[rd] = 0;
5370 break;
5371 }
5372
5373 if (!DTRACE_INSCRATCH(mstate, size)) {
5374 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5375 regs[rd] = 0;
5376 break;
5377 }
5378
5379 for (;;) {
5380 if (i >= size) {
5381 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5382 regs[rd] = 0;
5383 break;
5384 }
5385 c = (i >= lim1) ? '\0' : dtrace_load8(s1++);
5386 if ((d[i++] = c) == '\0') {
5387 i--;
5388 break;
5389 }
5390 }
5391
5392 for (;;) {
5393 if (i >= size) {
5394 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5395 regs[rd] = 0;
5396 break;
5397 }
5398
5399 c = (j++ >= lim2) ? '\0' : dtrace_load8(s2++);
5400 if ((d[i++] = c) == '\0')
5401 break;
5402 }
5403
5404 if (i < size) {
5405 mstate->dtms_scratch_ptr += i;
5406 regs[rd] = (uintptr_t)d;
5407 }
5408
5409 break;
5410 }
5411
5412 case DIF_SUBR_STRTOLL: {
5413 uintptr_t s = tupregs[0].dttk_value;
5414 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
5415 size_t lim;
5416 int base = 10;
5417
5418 if (nargs > 1) {
5419 if ((base = tupregs[1].dttk_value) <= 1 ||
5420 base > ('z' - 'a' + 1) + ('9' - '' + 1)) {
5421 *flags |= CPU_DTRACE_ILLOP;
5422 break;
5423 }
5424 }
5425
5426 if (!dtrace_strcanload(s, size, &lim, mstate, vstate)) {
5427 regs[rd] = INT64_MIN;
5428 break;
5429 }
5430
5431 regs[rd] = dtrace_strtoll((char *)s, base, lim);
5432 break;
5433 }
5434
5435 case DIF_SUBR_LLTOSTR: {
5436 int64_t i = (int64_t)tupregs[0].dttk_value;
5437 uint64_t val, digit;
5438 uint64_t size = 65; /* enough room for 2^64 in binary */
5439 char *end = (char *)mstate->dtms_scratch_ptr + size - 1;
5440 int base = 10;
5441
5442 if (nargs > 1) {
5443 if ((base = tupregs[1].dttk_value) <= 1 ||
5444 base > ('z' - 'a' + 1) + ('9' - '' + 1)) {
5445 *flags |= CPU_DTRACE_ILLOP;
5446 break;
5447 }
5448 }
5449
5450 val = (base == 10 && i < 0) ? i * -1 : i;
5451
5452 if (!DTRACE_INSCRATCH(mstate, size)) {
5453 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5454 regs[rd] = 0;
5455 break;
5456 }
5457
5458 for (*end-- = '\0'; val; val /= base) {
5459 if ((digit = val % base) <= '9' - '') {
5460 *end-- = '' + digit;
5461 } else {
5462 *end-- = 'a' + (digit - ('9' - '') - 1);
5463 }
5464 }
5465
5466 if (i == 0 && base == 16)
5467 *end-- = '';
5468
5469 if (base == 16)
5470 *end-- = 'x';
5471
5472 if (i == 0 || base == 8 || base == 16)
5473 *end-- = '';
5474
5475 if (i < 0 && base == 10)
5476 *end-- = '-';
5477
5478 regs[rd] = (uintptr_t)end + 1;
5479 mstate->dtms_scratch_ptr += size;
5480 break;
5481 }
5482
5483 case DIF_SUBR_HTONS:
5484 case DIF_SUBR_NTOHS:
5485 #if BYTE_ORDER == BIG_ENDIAN
5486 regs[rd] = (uint16_t)tupregs[0].dttk_value;
5487 #else
5488 regs[rd] = DT_BSWAP_16((uint16_t)tupregs[0].dttk_value);
5489 #endif
5490 break;
5491
5492
5493 case DIF_SUBR_HTONL:
5494 case DIF_SUBR_NTOHL:
5495 #if BYTE_ORDER == BIG_ENDIAN
5496 regs[rd] = (uint32_t)tupregs[0].dttk_value;
5497 #else
5498 regs[rd] = DT_BSWAP_32((uint32_t)tupregs[0].dttk_value);
5499 #endif
5500 break;
5501
5502
5503 case DIF_SUBR_HTONLL:
5504 case DIF_SUBR_NTOHLL:
5505 #if BYTE_ORDER == BIG_ENDIAN
5506 regs[rd] = (uint64_t)tupregs[0].dttk_value;
5507 #else
5508 regs[rd] = DT_BSWAP_64((uint64_t)tupregs[0].dttk_value);
5509 #endif
5510 break;
5511
5512
5513 case DIF_SUBR_DIRNAME:
5514 case DIF_SUBR_BASENAME: {
5515 char *dest = (char *)mstate->dtms_scratch_ptr;
5516 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
5517 uintptr_t src = tupregs[0].dttk_value;
5518 int i, j, len = dtrace_strlen((char *)src, size);
5519 int lastbase = -1, firstbase = -1, lastdir = -1;
5520 int start, end;
5521
5522 if (!dtrace_canload(src, len + 1, mstate, vstate)) {
5523 regs[rd] = 0;
5524 break;
5525 }
5526
5527 if (!DTRACE_INSCRATCH(mstate, size)) {
5528 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5529 regs[rd] = 0;
5530 break;
5531 }
5532
5533 /*
5534 * The basename and dirname for a zero-length string is
5535 * defined to be "."
5536 */
5537 if (len == 0) {
5538 len = 1;
5539 src = (uintptr_t)".";
5540 }
5541
5542 /*
5543 * Start from the back of the string, moving back toward the
5544 * front until we see a character that isn't a slash. That
5545 * character is the last character in the basename.
5546 */
5547 for (i = len - 1; i >= 0; i--) {
5548 if (dtrace_load8(src + i) != '/')
5549 break;
5550 }
5551
5552 if (i >= 0)
5553 lastbase = i;
5554
5555 /*
5556 * Starting from the last character in the basename, move
5557 * towards the front until we find a slash. The character
5558 * that we processed immediately before that is the first
5559 * character in the basename.
5560 */
5561 for (; i >= 0; i--) {
5562 if (dtrace_load8(src + i) == '/')
5563 break;
5564 }
5565
5566 if (i >= 0)
5567 firstbase = i + 1;
5568
5569 /*
5570 * Now keep going until we find a non-slash character. That
5571 * character is the last character in the dirname.
5572 */
5573 for (; i >= 0; i--) {
5574 if (dtrace_load8(src + i) != '/')
5575 break;
5576 }
5577
5578 if (i >= 0)
5579 lastdir = i;
5580
5581 ASSERT(!(lastbase == -1 && firstbase != -1));
5582 ASSERT(!(firstbase == -1 && lastdir != -1));
5583
5584 if (lastbase == -1) {
5585 /*
5586 * We didn't find a non-slash character. We know that
5587 * the length is non-zero, so the whole string must be
5588 * slashes. In either the dirname or the basename
5589 * case, we return '/'.
5590 */
5591 ASSERT(firstbase == -1);
5592 firstbase = lastbase = lastdir = 0;
5593 }
5594
5595 if (firstbase == -1) {
5596 /*
5597 * The entire string consists only of a basename
5598 * component. If we're looking for dirname, we need
5599 * to change our string to be just "."; if we're
5600 * looking for a basename, we'll just set the first
5601 * character of the basename to be 0.
5602 */
5603 if (subr == DIF_SUBR_DIRNAME) {
5604 ASSERT(lastdir == -1);
5605 src = (uintptr_t)".";
5606 lastdir = 0;
5607 } else {
5608 firstbase = 0;
5609 }
5610 }
5611
5612 if (subr == DIF_SUBR_DIRNAME) {
5613 if (lastdir == -1) {
5614 /*
5615 * We know that we have a slash in the name --
5616 * or lastdir would be set to 0, above. And
5617 * because lastdir is -1, we know that this
5618 * slash must be the first character. (That
5619 * is, the full string must be of the form
5620 * "/basename".) In this case, the last
5621 * character of the directory name is 0.
5622 */
5623 lastdir = 0;
5624 }
5625
5626 start = 0;
5627 end = lastdir;
5628 } else {
5629 ASSERT(subr == DIF_SUBR_BASENAME);
5630 ASSERT(firstbase != -1 && lastbase != -1);
5631 start = firstbase;
5632 end = lastbase;
5633 }
5634
5635 for (i = start, j = 0; i <= end && j < size - 1; i++, j++)
5636 dest[j] = dtrace_load8(src + i);
5637
5638 dest[j] = '\0';
5639 regs[rd] = (uintptr_t)dest;
5640 mstate->dtms_scratch_ptr += size;
5641 break;
5642 }
5643
5644 case DIF_SUBR_GETF: {
5645 uintptr_t fd = tupregs[0].dttk_value;
5646 struct filedesc *fdp;
5647 file_t *fp;
5648
5649 if (!dtrace_priv_proc(state)) {
5650 regs[rd] = 0;
5651 break;
5652 }
5653 fdp = curproc->p_fd;
5654 FILEDESC_SLOCK(fdp);
5655 /*
5656 * XXXMJG this looks broken as no ref is taken.
5657 */
5658 fp = fget_noref(fdp, fd);
5659 mstate->dtms_getf = fp;
5660 regs[rd] = (uintptr_t)fp;
5661 FILEDESC_SUNLOCK(fdp);
5662 break;
5663 }
5664
5665 case DIF_SUBR_CLEANPATH: {
5666 char *dest = (char *)mstate->dtms_scratch_ptr, c;
5667 uint64_t size = state->dts_options[DTRACEOPT_STRSIZE];
5668 uintptr_t src = tupregs[0].dttk_value;
5669 size_t lim;
5670 int i = 0, j = 0;
5671 #ifdef illumos
5672 zone_t *z;
5673 #endif
5674
5675 if (!dtrace_strcanload(src, size, &lim, mstate, vstate)) {
5676 regs[rd] = 0;
5677 break;
5678 }
5679
5680 if (!DTRACE_INSCRATCH(mstate, size)) {
5681 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5682 regs[rd] = 0;
5683 break;
5684 }
5685
5686 /*
5687 * Move forward, loading each character.
5688 */
5689 do {
5690 c = (i >= lim) ? '\0' : dtrace_load8(src + i++);
5691 next:
5692 if (j + 5 >= size) /* 5 = strlen("/..c\0") */
5693 break;
5694
5695 if (c != '/') {
5696 dest[j++] = c;
5697 continue;
5698 }
5699
5700 c = (i >= lim) ? '\0' : dtrace_load8(src + i++);
5701
5702 if (c == '/') {
5703 /*
5704 * We have two slashes -- we can just advance
5705 * to the next character.
5706 */
5707 goto next;
5708 }
5709
5710 if (c != '.') {
5711 /*
5712 * This is not "." and it's not ".." -- we can
5713 * just store the "/" and this character and
5714 * drive on.
5715 */
5716 dest[j++] = '/';
5717 dest[j++] = c;
5718 continue;
5719 }
5720
5721 c = (i >= lim) ? '\0' : dtrace_load8(src + i++);
5722
5723 if (c == '/') {
5724 /*
5725 * This is a "/./" component. We're not going
5726 * to store anything in the destination buffer;
5727 * we're just going to go to the next component.
5728 */
5729 goto next;
5730 }
5731
5732 if (c != '.') {
5733 /*
5734 * This is not ".." -- we can just store the
5735 * "/." and this character and continue
5736 * processing.
5737 */
5738 dest[j++] = '/';
5739 dest[j++] = '.';
5740 dest[j++] = c;
5741 continue;
5742 }
5743
5744 c = (i >= lim) ? '\0' : dtrace_load8(src + i++);
5745
5746 if (c != '/' && c != '\0') {
5747 /*
5748 * This is not ".." -- it's "..[mumble]".
5749 * We'll store the "/.." and this character
5750 * and continue processing.
5751 */
5752 dest[j++] = '/';
5753 dest[j++] = '.';
5754 dest[j++] = '.';
5755 dest[j++] = c;
5756 continue;
5757 }
5758
5759 /*
5760 * This is "/../" or "/..\0". We need to back up
5761 * our destination pointer until we find a "/".
5762 */
5763 i--;
5764 while (j != 0 && dest[--j] != '/')
5765 continue;
5766
5767 if (c == '\0')
5768 dest[++j] = '/';
5769 } while (c != '\0');
5770
5771 dest[j] = '\0';
5772
5773 #ifdef illumos
5774 if (mstate->dtms_getf != NULL &&
5775 !(mstate->dtms_access & DTRACE_ACCESS_KERNEL) &&
5776 (z = state->dts_cred.dcr_cred->cr_zone) != kcred->cr_zone) {
5777 /*
5778 * If we've done a getf() as a part of this ECB and we
5779 * don't have kernel access (and we're not in the global
5780 * zone), check if the path we cleaned up begins with
5781 * the zone's root path, and trim it off if so. Note
5782 * that this is an output cleanliness issue, not a
5783 * security issue: knowing one's zone root path does
5784 * not enable privilege escalation.
5785 */
5786 if (strstr(dest, z->zone_rootpath) == dest)
5787 dest += strlen(z->zone_rootpath) - 1;
5788 }
5789 #endif
5790
5791 regs[rd] = (uintptr_t)dest;
5792 mstate->dtms_scratch_ptr += size;
5793 break;
5794 }
5795
5796 case DIF_SUBR_INET_NTOA:
5797 case DIF_SUBR_INET_NTOA6:
5798 case DIF_SUBR_INET_NTOP: {
5799 size_t size;
5800 int af, argi, i;
5801 char *base, *end;
5802
5803 if (subr == DIF_SUBR_INET_NTOP) {
5804 af = (int)tupregs[0].dttk_value;
5805 argi = 1;
5806 } else {
5807 af = subr == DIF_SUBR_INET_NTOA ? AF_INET: AF_INET6;
5808 argi = 0;
5809 }
5810
5811 if (af == AF_INET) {
5812 ipaddr_t ip4;
5813 uint8_t *ptr8, val;
5814
5815 if (!dtrace_canload(tupregs[argi].dttk_value,
5816 sizeof (ipaddr_t), mstate, vstate)) {
5817 regs[rd] = 0;
5818 break;
5819 }
5820
5821 /*
5822 * Safely load the IPv4 address.
5823 */
5824 ip4 = dtrace_load32(tupregs[argi].dttk_value);
5825
5826 /*
5827 * Check an IPv4 string will fit in scratch.
5828 */
5829 size = INET_ADDRSTRLEN;
5830 if (!DTRACE_INSCRATCH(mstate, size)) {
5831 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5832 regs[rd] = 0;
5833 break;
5834 }
5835 base = (char *)mstate->dtms_scratch_ptr;
5836 end = (char *)mstate->dtms_scratch_ptr + size - 1;
5837
5838 /*
5839 * Stringify as a dotted decimal quad.
5840 */
5841 *end-- = '\0';
5842 ptr8 = (uint8_t *)&ip4;
5843 for (i = 3; i >= 0; i--) {
5844 val = ptr8[i];
5845
5846 if (val == 0) {
5847 *end-- = '';
5848 } else {
5849 for (; val; val /= 10) {
5850 *end-- = '' + (val % 10);
5851 }
5852 }
5853
5854 if (i > 0)
5855 *end-- = '.';
5856 }
5857 ASSERT(end + 1 >= base);
5858
5859 } else if (af == AF_INET6) {
5860 struct in6_addr ip6;
5861 int firstzero, tryzero, numzero, v6end;
5862 uint16_t val;
5863 const char digits[] = "0123456789abcdef";
5864
5865 /*
5866 * Stringify using RFC 1884 convention 2 - 16 bit
5867 * hexadecimal values with a zero-run compression.
5868 * Lower case hexadecimal digits are used.
5869 * eg, fe80::214:4fff:fe0b:76c8.
5870 * The IPv4 embedded form is returned for inet_ntop,
5871 * just the IPv4 string is returned for inet_ntoa6.
5872 */
5873
5874 if (!dtrace_canload(tupregs[argi].dttk_value,
5875 sizeof (struct in6_addr), mstate, vstate)) {
5876 regs[rd] = 0;
5877 break;
5878 }
5879
5880 /*
5881 * Safely load the IPv6 address.
5882 */
5883 dtrace_bcopy(
5884 (void *)(uintptr_t)tupregs[argi].dttk_value,
5885 (void *)(uintptr_t)&ip6, sizeof (struct in6_addr));
5886
5887 /*
5888 * Check an IPv6 string will fit in scratch.
5889 */
5890 size = INET6_ADDRSTRLEN;
5891 if (!DTRACE_INSCRATCH(mstate, size)) {
5892 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
5893 regs[rd] = 0;
5894 break;
5895 }
5896 base = (char *)mstate->dtms_scratch_ptr;
5897 end = (char *)mstate->dtms_scratch_ptr + size - 1;
5898 *end-- = '\0';
5899
5900 /*
5901 * Find the longest run of 16 bit zero values
5902 * for the single allowed zero compression - "::".
5903 */
5904 firstzero = -1;
5905 tryzero = -1;
5906 numzero = 1;
5907 for (i = 0; i < sizeof (struct in6_addr); i++) {
5908 #ifdef illumos
5909 if (ip6._S6_un._S6_u8[i] == 0 &&
5910 #else
5911 if (ip6.__u6_addr.__u6_addr8[i] == 0 &&
5912 #endif
5913 tryzero == -1 && i % 2 == 0) {
5914 tryzero = i;
5915 continue;
5916 }
5917
5918 if (tryzero != -1 &&
5919 #ifdef illumos
5920 (ip6._S6_un._S6_u8[i] != 0 ||
5921 #else
5922 (ip6.__u6_addr.__u6_addr8[i] != 0 ||
5923 #endif
5924 i == sizeof (struct in6_addr) - 1)) {
5925
5926 if (i - tryzero <= numzero) {
5927 tryzero = -1;
5928 continue;
5929 }
5930
5931 firstzero = tryzero;
5932 numzero = i - i % 2 - tryzero;
5933 tryzero = -1;
5934
5935 #ifdef illumos
5936 if (ip6._S6_un._S6_u8[i] == 0 &&
5937 #else
5938 if (ip6.__u6_addr.__u6_addr8[i] == 0 &&
5939 #endif
5940 i == sizeof (struct in6_addr) - 1)
5941 numzero += 2;
5942 }
5943 }
5944 ASSERT(firstzero + numzero <= sizeof (struct in6_addr));
5945
5946 /*
5947 * Check for an IPv4 embedded address.
5948 */
5949 v6end = sizeof (struct in6_addr) - 2;
5950 if (IN6_IS_ADDR_V4MAPPED(&ip6) ||
5951 IN6_IS_ADDR_V4COMPAT(&ip6)) {
5952 for (i = sizeof (struct in6_addr) - 1;
5953 i >= DTRACE_V4MAPPED_OFFSET; i--) {
5954 ASSERT(end >= base);
5955
5956 #ifdef illumos
5957 val = ip6._S6_un._S6_u8[i];
5958 #else
5959 val = ip6.__u6_addr.__u6_addr8[i];
5960 #endif
5961
5962 if (val == 0) {
5963 *end-- = '';
5964 } else {
5965 for (; val; val /= 10) {
5966 *end-- = '' + val % 10;
5967 }
5968 }
5969
5970 if (i > DTRACE_V4MAPPED_OFFSET)
5971 *end-- = '.';
5972 }
5973
5974 if (subr == DIF_SUBR_INET_NTOA6)
5975 goto inetout;
5976
5977 /*
5978 * Set v6end to skip the IPv4 address that
5979 * we have already stringified.
5980 */
5981 v6end = 10;
5982 }
5983
5984 /*
5985 * Build the IPv6 string by working through the
5986 * address in reverse.
5987 */
5988 for (i = v6end; i >= 0; i -= 2) {
5989 ASSERT(end >= base);
5990
5991 if (i == firstzero + numzero - 2) {
5992 *end-- = ':';
5993 *end-- = ':';
5994 i -= numzero - 2;
5995 continue;
5996 }
5997
5998 if (i < 14 && i != firstzero - 2)
5999 *end-- = ':';
6000
6001 #ifdef illumos
6002 val = (ip6._S6_un._S6_u8[i] << 8) +
6003 ip6._S6_un._S6_u8[i + 1];
6004 #else
6005 val = (ip6.__u6_addr.__u6_addr8[i] << 8) +
6006 ip6.__u6_addr.__u6_addr8[i + 1];
6007 #endif
6008
6009 if (val == 0) {
6010 *end-- = '';
6011 } else {
6012 for (; val; val /= 16) {
6013 *end-- = digits[val % 16];
6014 }
6015 }
6016 }
6017 ASSERT(end + 1 >= base);
6018
6019 } else {
6020 /*
6021 * The user didn't use AH_INET or AH_INET6.
6022 */
6023 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
6024 regs[rd] = 0;
6025 break;
6026 }
6027
6028 inetout: regs[rd] = (uintptr_t)end + 1;
6029 mstate->dtms_scratch_ptr += size;
6030 break;
6031 }
6032
6033 case DIF_SUBR_MEMREF: {
6034 uintptr_t size = 2 * sizeof(uintptr_t);
6035 uintptr_t *memref = (uintptr_t *) P2ROUNDUP(mstate->dtms_scratch_ptr, sizeof(uintptr_t));
6036 size_t scratch_size = ((uintptr_t) memref - mstate->dtms_scratch_ptr) + size;
6037
6038 /* address and length */
6039 memref[0] = tupregs[0].dttk_value;
6040 memref[1] = tupregs[1].dttk_value;
6041
6042 regs[rd] = (uintptr_t) memref;
6043 mstate->dtms_scratch_ptr += scratch_size;
6044 break;
6045 }
6046
6047 #ifndef illumos
6048 case DIF_SUBR_MEMSTR: {
6049 char *str = (char *)mstate->dtms_scratch_ptr;
6050 uintptr_t mem = tupregs[0].dttk_value;
6051 char c = tupregs[1].dttk_value;
6052 size_t size = tupregs[2].dttk_value;
6053 uint8_t n;
6054 int i;
6055
6056 regs[rd] = 0;
6057
6058 if (size == 0)
6059 break;
6060
6061 if (!dtrace_canload(mem, size - 1, mstate, vstate))
6062 break;
6063
6064 if (!DTRACE_INSCRATCH(mstate, size)) {
6065 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
6066 break;
6067 }
6068
6069 if (dtrace_memstr_max != 0 && size > dtrace_memstr_max) {
6070 *flags |= CPU_DTRACE_ILLOP;
6071 break;
6072 }
6073
6074 for (i = 0; i < size - 1; i++) {
6075 n = dtrace_load8(mem++);
6076 str[i] = (n == 0) ? c : n;
6077 }
6078 str[size - 1] = 0;
6079
6080 regs[rd] = (uintptr_t)str;
6081 mstate->dtms_scratch_ptr += size;
6082 break;
6083 }
6084 #endif
6085 }
6086 }
6087
6088 /*
6089 * Emulate the execution of DTrace IR instructions specified by the given
6090 * DIF object. This function is deliberately void of assertions as all of
6091 * the necessary checks are handled by a call to dtrace_difo_validate().
6092 */
6093 static uint64_t
6094 dtrace_dif_emulate(dtrace_difo_t *difo, dtrace_mstate_t *mstate,
6095 dtrace_vstate_t *vstate, dtrace_state_t *state)
6096 {
6097 const dif_instr_t *text = difo->dtdo_buf;
6098 const uint_t textlen = difo->dtdo_len;
6099 const char *strtab = difo->dtdo_strtab;
6100 const uint64_t *inttab = difo->dtdo_inttab;
6101
6102 uint64_t rval = 0;
6103 dtrace_statvar_t *svar;
6104 dtrace_dstate_t *dstate = &vstate->dtvs_dynvars;
6105 dtrace_difv_t *v;
6106 volatile uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags;
6107 volatile uintptr_t *illval = &cpu_core[curcpu].cpuc_dtrace_illval;
6108
6109 dtrace_key_t tupregs[DIF_DTR_NREGS + 2]; /* +2 for thread and id */
6110 uint64_t regs[DIF_DIR_NREGS];
6111 uint64_t *tmp;
6112
6113 uint8_t cc_n = 0, cc_z = 0, cc_v = 0, cc_c = 0;
6114 int64_t cc_r;
6115 uint_t pc = 0, id, opc = 0;
6116 uint8_t ttop = 0;
6117 dif_instr_t instr;
6118 uint_t r1, r2, rd;
6119
6120 /*
6121 * We stash the current DIF object into the machine state: we need it
6122 * for subsequent access checking.
6123 */
6124 mstate->dtms_difo = difo;
6125
6126 regs[DIF_REG_R0] = 0; /* %r0 is fixed at zero */
6127
6128 while (pc < textlen && !(*flags & CPU_DTRACE_FAULT)) {
6129 opc = pc;
6130
6131 instr = text[pc++];
6132 r1 = DIF_INSTR_R1(instr);
6133 r2 = DIF_INSTR_R2(instr);
6134 rd = DIF_INSTR_RD(instr);
6135
6136 switch (DIF_INSTR_OP(instr)) {
6137 case DIF_OP_OR:
6138 regs[rd] = regs[r1] | regs[r2];
6139 break;
6140 case DIF_OP_XOR:
6141 regs[rd] = regs[r1] ^ regs[r2];
6142 break;
6143 case DIF_OP_AND:
6144 regs[rd] = regs[r1] & regs[r2];
6145 break;
6146 case DIF_OP_SLL:
6147 regs[rd] = regs[r1] << regs[r2];
6148 break;
6149 case DIF_OP_SRL:
6150 regs[rd] = regs[r1] >> regs[r2];
6151 break;
6152 case DIF_OP_SUB:
6153 regs[rd] = regs[r1] - regs[r2];
6154 break;
6155 case DIF_OP_ADD:
6156 regs[rd] = regs[r1] + regs[r2];
6157 break;
6158 case DIF_OP_MUL:
6159 regs[rd] = regs[r1] * regs[r2];
6160 break;
6161 case DIF_OP_SDIV:
6162 if (regs[r2] == 0) {
6163 regs[rd] = 0;
6164 *flags |= CPU_DTRACE_DIVZERO;
6165 } else {
6166 regs[rd] = (int64_t)regs[r1] /
6167 (int64_t)regs[r2];
6168 }
6169 break;
6170
6171 case DIF_OP_UDIV:
6172 if (regs[r2] == 0) {
6173 regs[rd] = 0;
6174 *flags |= CPU_DTRACE_DIVZERO;
6175 } else {
6176 regs[rd] = regs[r1] / regs[r2];
6177 }
6178 break;
6179
6180 case DIF_OP_SREM:
6181 if (regs[r2] == 0) {
6182 regs[rd] = 0;
6183 *flags |= CPU_DTRACE_DIVZERO;
6184 } else {
6185 regs[rd] = (int64_t)regs[r1] %
6186 (int64_t)regs[r2];
6187 }
6188 break;
6189
6190 case DIF_OP_UREM:
6191 if (regs[r2] == 0) {
6192 regs[rd] = 0;
6193 *flags |= CPU_DTRACE_DIVZERO;
6194 } else {
6195 regs[rd] = regs[r1] % regs[r2];
6196 }
6197 break;
6198
6199 case DIF_OP_NOT:
6200 regs[rd] = ~regs[r1];
6201 break;
6202 case DIF_OP_MOV:
6203 regs[rd] = regs[r1];
6204 break;
6205 case DIF_OP_CMP:
6206 cc_r = regs[r1] - regs[r2];
6207 cc_n = cc_r < 0;
6208 cc_z = cc_r == 0;
6209 cc_v = 0;
6210 cc_c = regs[r1] < regs[r2];
6211 break;
6212 case DIF_OP_TST:
6213 cc_n = cc_v = cc_c = 0;
6214 cc_z = regs[r1] == 0;
6215 break;
6216 case DIF_OP_BA:
6217 pc = DIF_INSTR_LABEL(instr);
6218 break;
6219 case DIF_OP_BE:
6220 if (cc_z)
6221 pc = DIF_INSTR_LABEL(instr);
6222 break;
6223 case DIF_OP_BNE:
6224 if (cc_z == 0)
6225 pc = DIF_INSTR_LABEL(instr);
6226 break;
6227 case DIF_OP_BG:
6228 if ((cc_z | (cc_n ^ cc_v)) == 0)
6229 pc = DIF_INSTR_LABEL(instr);
6230 break;
6231 case DIF_OP_BGU:
6232 if ((cc_c | cc_z) == 0)
6233 pc = DIF_INSTR_LABEL(instr);
6234 break;
6235 case DIF_OP_BGE:
6236 if ((cc_n ^ cc_v) == 0)
6237 pc = DIF_INSTR_LABEL(instr);
6238 break;
6239 case DIF_OP_BGEU:
6240 if (cc_c == 0)
6241 pc = DIF_INSTR_LABEL(instr);
6242 break;
6243 case DIF_OP_BL:
6244 if (cc_n ^ cc_v)
6245 pc = DIF_INSTR_LABEL(instr);
6246 break;
6247 case DIF_OP_BLU:
6248 if (cc_c)
6249 pc = DIF_INSTR_LABEL(instr);
6250 break;
6251 case DIF_OP_BLE:
6252 if (cc_z | (cc_n ^ cc_v))
6253 pc = DIF_INSTR_LABEL(instr);
6254 break;
6255 case DIF_OP_BLEU:
6256 if (cc_c | cc_z)
6257 pc = DIF_INSTR_LABEL(instr);
6258 break;
6259 case DIF_OP_RLDSB:
6260 if (!dtrace_canload(regs[r1], 1, mstate, vstate))
6261 break;
6262 /*FALLTHROUGH*/
6263 case DIF_OP_LDSB:
6264 regs[rd] = (int8_t)dtrace_load8(regs[r1]);
6265 break;
6266 case DIF_OP_RLDSH:
6267 if (!dtrace_canload(regs[r1], 2, mstate, vstate))
6268 break;
6269 /*FALLTHROUGH*/
6270 case DIF_OP_LDSH:
6271 regs[rd] = (int16_t)dtrace_load16(regs[r1]);
6272 break;
6273 case DIF_OP_RLDSW:
6274 if (!dtrace_canload(regs[r1], 4, mstate, vstate))
6275 break;
6276 /*FALLTHROUGH*/
6277 case DIF_OP_LDSW:
6278 regs[rd] = (int32_t)dtrace_load32(regs[r1]);
6279 break;
6280 case DIF_OP_RLDUB:
6281 if (!dtrace_canload(regs[r1], 1, mstate, vstate))
6282 break;
6283 /*FALLTHROUGH*/
6284 case DIF_OP_LDUB:
6285 regs[rd] = dtrace_load8(regs[r1]);
6286 break;
6287 case DIF_OP_RLDUH:
6288 if (!dtrace_canload(regs[r1], 2, mstate, vstate))
6289 break;
6290 /*FALLTHROUGH*/
6291 case DIF_OP_LDUH:
6292 regs[rd] = dtrace_load16(regs[r1]);
6293 break;
6294 case DIF_OP_RLDUW:
6295 if (!dtrace_canload(regs[r1], 4, mstate, vstate))
6296 break;
6297 /*FALLTHROUGH*/
6298 case DIF_OP_LDUW:
6299 regs[rd] = dtrace_load32(regs[r1]);
6300 break;
6301 case DIF_OP_RLDX:
6302 if (!dtrace_canload(regs[r1], 8, mstate, vstate))
6303 break;
6304 /*FALLTHROUGH*/
6305 case DIF_OP_LDX:
6306 regs[rd] = dtrace_load64(regs[r1]);
6307 break;
6308 case DIF_OP_ULDSB:
6309 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6310 regs[rd] = (int8_t)
6311 dtrace_fuword8((void *)(uintptr_t)regs[r1]);
6312 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6313 break;
6314 case DIF_OP_ULDSH:
6315 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6316 regs[rd] = (int16_t)
6317 dtrace_fuword16((void *)(uintptr_t)regs[r1]);
6318 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6319 break;
6320 case DIF_OP_ULDSW:
6321 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6322 regs[rd] = (int32_t)
6323 dtrace_fuword32((void *)(uintptr_t)regs[r1]);
6324 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6325 break;
6326 case DIF_OP_ULDUB:
6327 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6328 regs[rd] =
6329 dtrace_fuword8((void *)(uintptr_t)regs[r1]);
6330 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6331 break;
6332 case DIF_OP_ULDUH:
6333 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6334 regs[rd] =
6335 dtrace_fuword16((void *)(uintptr_t)regs[r1]);
6336 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6337 break;
6338 case DIF_OP_ULDUW:
6339 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6340 regs[rd] =
6341 dtrace_fuword32((void *)(uintptr_t)regs[r1]);
6342 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6343 break;
6344 case DIF_OP_ULDX:
6345 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
6346 regs[rd] =
6347 dtrace_fuword64((void *)(uintptr_t)regs[r1]);
6348 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
6349 break;
6350 case DIF_OP_RET:
6351 rval = regs[rd];
6352 pc = textlen;
6353 break;
6354 case DIF_OP_NOP:
6355 break;
6356 case DIF_OP_SETX:
6357 regs[rd] = inttab[DIF_INSTR_INTEGER(instr)];
6358 break;
6359 case DIF_OP_SETS:
6360 regs[rd] = (uint64_t)(uintptr_t)
6361 (strtab + DIF_INSTR_STRING(instr));
6362 break;
6363 case DIF_OP_SCMP: {
6364 size_t sz = state->dts_options[DTRACEOPT_STRSIZE];
6365 uintptr_t s1 = regs[r1];
6366 uintptr_t s2 = regs[r2];
6367 size_t lim1, lim2;
6368
6369 /*
6370 * If one of the strings is NULL then the limit becomes
6371 * 0 which compares 0 characters in dtrace_strncmp()
6372 * resulting in a false positive. dtrace_strncmp()
6373 * treats a NULL as an empty 1-char string.
6374 */
6375 lim1 = lim2 = 1;
6376
6377 if (s1 != 0 &&
6378 !dtrace_strcanload(s1, sz, &lim1, mstate, vstate))
6379 break;
6380 if (s2 != 0 &&
6381 !dtrace_strcanload(s2, sz, &lim2, mstate, vstate))
6382 break;
6383
6384 cc_r = dtrace_strncmp((char *)s1, (char *)s2,
6385 MIN(lim1, lim2));
6386
6387 cc_n = cc_r < 0;
6388 cc_z = cc_r == 0;
6389 cc_v = cc_c = 0;
6390 break;
6391 }
6392 case DIF_OP_LDGA:
6393 regs[rd] = dtrace_dif_variable(mstate, state,
6394 r1, regs[r2]);
6395 break;
6396 case DIF_OP_LDGS:
6397 id = DIF_INSTR_VAR(instr);
6398
6399 if (id >= DIF_VAR_OTHER_UBASE) {
6400 uintptr_t a;
6401
6402 id -= DIF_VAR_OTHER_UBASE;
6403 svar = vstate->dtvs_globals[id];
6404 ASSERT(svar != NULL);
6405 v = &svar->dtsv_var;
6406
6407 if (!(v->dtdv_type.dtdt_flags & DIF_TF_BYREF)) {
6408 regs[rd] = svar->dtsv_data;
6409 break;
6410 }
6411
6412 a = (uintptr_t)svar->dtsv_data;
6413
6414 if (*(uint8_t *)a == UINT8_MAX) {
6415 /*
6416 * If the 0th byte is set to UINT8_MAX
6417 * then this is to be treated as a
6418 * reference to a NULL variable.
6419 */
6420 regs[rd] = 0;
6421 } else {
6422 regs[rd] = a + sizeof (uint64_t);
6423 }
6424
6425 break;
6426 }
6427
6428 regs[rd] = dtrace_dif_variable(mstate, state, id, 0);
6429 break;
6430
6431 case DIF_OP_STGS:
6432 id = DIF_INSTR_VAR(instr);
6433
6434 ASSERT(id >= DIF_VAR_OTHER_UBASE);
6435 id -= DIF_VAR_OTHER_UBASE;
6436
6437 VERIFY(id < vstate->dtvs_nglobals);
6438 svar = vstate->dtvs_globals[id];
6439 ASSERT(svar != NULL);
6440 v = &svar->dtsv_var;
6441
6442 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6443 uintptr_t a = (uintptr_t)svar->dtsv_data;
6444 size_t lim;
6445
6446 ASSERT(a != 0);
6447 ASSERT(svar->dtsv_size != 0);
6448
6449 if (regs[rd] == 0) {
6450 *(uint8_t *)a = UINT8_MAX;
6451 break;
6452 } else {
6453 *(uint8_t *)a = 0;
6454 a += sizeof (uint64_t);
6455 }
6456 if (!dtrace_vcanload(
6457 (void *)(uintptr_t)regs[rd], &v->dtdv_type,
6458 &lim, mstate, vstate))
6459 break;
6460
6461 dtrace_vcopy((void *)(uintptr_t)regs[rd],
6462 (void *)a, &v->dtdv_type, lim);
6463 break;
6464 }
6465
6466 svar->dtsv_data = regs[rd];
6467 break;
6468
6469 case DIF_OP_LDTA:
6470 /*
6471 * There are no DTrace built-in thread-local arrays at
6472 * present. This opcode is saved for future work.
6473 */
6474 *flags |= CPU_DTRACE_ILLOP;
6475 regs[rd] = 0;
6476 break;
6477
6478 case DIF_OP_LDLS:
6479 id = DIF_INSTR_VAR(instr);
6480
6481 if (id < DIF_VAR_OTHER_UBASE) {
6482 /*
6483 * For now, this has no meaning.
6484 */
6485 regs[rd] = 0;
6486 break;
6487 }
6488
6489 id -= DIF_VAR_OTHER_UBASE;
6490
6491 ASSERT(id < vstate->dtvs_nlocals);
6492 ASSERT(vstate->dtvs_locals != NULL);
6493
6494 svar = vstate->dtvs_locals[id];
6495 ASSERT(svar != NULL);
6496 v = &svar->dtsv_var;
6497
6498 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6499 uintptr_t a = (uintptr_t)svar->dtsv_data;
6500 size_t sz = v->dtdv_type.dtdt_size;
6501 size_t lim;
6502
6503 sz += sizeof (uint64_t);
6504 ASSERT(svar->dtsv_size == NCPU * sz);
6505 a += curcpu * sz;
6506
6507 if (*(uint8_t *)a == UINT8_MAX) {
6508 /*
6509 * If the 0th byte is set to UINT8_MAX
6510 * then this is to be treated as a
6511 * reference to a NULL variable.
6512 */
6513 regs[rd] = 0;
6514 } else {
6515 regs[rd] = a + sizeof (uint64_t);
6516 }
6517
6518 break;
6519 }
6520
6521 ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t));
6522 tmp = (uint64_t *)(uintptr_t)svar->dtsv_data;
6523 regs[rd] = tmp[curcpu];
6524 break;
6525
6526 case DIF_OP_STLS:
6527 id = DIF_INSTR_VAR(instr);
6528
6529 ASSERT(id >= DIF_VAR_OTHER_UBASE);
6530 id -= DIF_VAR_OTHER_UBASE;
6531 VERIFY(id < vstate->dtvs_nlocals);
6532
6533 ASSERT(vstate->dtvs_locals != NULL);
6534 svar = vstate->dtvs_locals[id];
6535 ASSERT(svar != NULL);
6536 v = &svar->dtsv_var;
6537
6538 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6539 uintptr_t a = (uintptr_t)svar->dtsv_data;
6540 size_t sz = v->dtdv_type.dtdt_size;
6541 size_t lim;
6542
6543 sz += sizeof (uint64_t);
6544 ASSERT(svar->dtsv_size == NCPU * sz);
6545 a += curcpu * sz;
6546
6547 if (regs[rd] == 0) {
6548 *(uint8_t *)a = UINT8_MAX;
6549 break;
6550 } else {
6551 *(uint8_t *)a = 0;
6552 a += sizeof (uint64_t);
6553 }
6554
6555 if (!dtrace_vcanload(
6556 (void *)(uintptr_t)regs[rd], &v->dtdv_type,
6557 &lim, mstate, vstate))
6558 break;
6559
6560 dtrace_vcopy((void *)(uintptr_t)regs[rd],
6561 (void *)a, &v->dtdv_type, lim);
6562 break;
6563 }
6564
6565 ASSERT(svar->dtsv_size == NCPU * sizeof (uint64_t));
6566 tmp = (uint64_t *)(uintptr_t)svar->dtsv_data;
6567 tmp[curcpu] = regs[rd];
6568 break;
6569
6570 case DIF_OP_LDTS: {
6571 dtrace_dynvar_t *dvar;
6572 dtrace_key_t *key;
6573
6574 id = DIF_INSTR_VAR(instr);
6575 ASSERT(id >= DIF_VAR_OTHER_UBASE);
6576 id -= DIF_VAR_OTHER_UBASE;
6577 v = &vstate->dtvs_tlocals[id];
6578
6579 key = &tupregs[DIF_DTR_NREGS];
6580 key[0].dttk_value = (uint64_t)id;
6581 key[0].dttk_size = 0;
6582 DTRACE_TLS_THRKEY(key[1].dttk_value);
6583 key[1].dttk_size = 0;
6584
6585 dvar = dtrace_dynvar(dstate, 2, key,
6586 sizeof (uint64_t), DTRACE_DYNVAR_NOALLOC,
6587 mstate, vstate);
6588
6589 if (dvar == NULL) {
6590 regs[rd] = 0;
6591 break;
6592 }
6593
6594 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6595 regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data;
6596 } else {
6597 regs[rd] = *((uint64_t *)dvar->dtdv_data);
6598 }
6599
6600 break;
6601 }
6602
6603 case DIF_OP_STTS: {
6604 dtrace_dynvar_t *dvar;
6605 dtrace_key_t *key;
6606
6607 id = DIF_INSTR_VAR(instr);
6608 ASSERT(id >= DIF_VAR_OTHER_UBASE);
6609 id -= DIF_VAR_OTHER_UBASE;
6610 VERIFY(id < vstate->dtvs_ntlocals);
6611
6612 key = &tupregs[DIF_DTR_NREGS];
6613 key[0].dttk_value = (uint64_t)id;
6614 key[0].dttk_size = 0;
6615 DTRACE_TLS_THRKEY(key[1].dttk_value);
6616 key[1].dttk_size = 0;
6617 v = &vstate->dtvs_tlocals[id];
6618
6619 dvar = dtrace_dynvar(dstate, 2, key,
6620 v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
6621 v->dtdv_type.dtdt_size : sizeof (uint64_t),
6622 regs[rd] ? DTRACE_DYNVAR_ALLOC :
6623 DTRACE_DYNVAR_DEALLOC, mstate, vstate);
6624
6625 /*
6626 * Given that we're storing to thread-local data,
6627 * we need to flush our predicate cache.
6628 */
6629 curthread->t_predcache = 0;
6630
6631 if (dvar == NULL)
6632 break;
6633
6634 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6635 size_t lim;
6636
6637 if (!dtrace_vcanload(
6638 (void *)(uintptr_t)regs[rd],
6639 &v->dtdv_type, &lim, mstate, vstate))
6640 break;
6641
6642 dtrace_vcopy((void *)(uintptr_t)regs[rd],
6643 dvar->dtdv_data, &v->dtdv_type, lim);
6644 } else {
6645 *((uint64_t *)dvar->dtdv_data) = regs[rd];
6646 }
6647
6648 break;
6649 }
6650
6651 case DIF_OP_SRA:
6652 regs[rd] = (int64_t)regs[r1] >> regs[r2];
6653 break;
6654
6655 case DIF_OP_CALL:
6656 dtrace_dif_subr(DIF_INSTR_SUBR(instr), rd,
6657 regs, tupregs, ttop, mstate, state);
6658 break;
6659
6660 case DIF_OP_PUSHTR:
6661 if (ttop == DIF_DTR_NREGS) {
6662 *flags |= CPU_DTRACE_TUPOFLOW;
6663 break;
6664 }
6665
6666 if (r1 == DIF_TYPE_STRING) {
6667 /*
6668 * If this is a string type and the size is 0,
6669 * we'll use the system-wide default string
6670 * size. Note that we are _not_ looking at
6671 * the value of the DTRACEOPT_STRSIZE option;
6672 * had this been set, we would expect to have
6673 * a non-zero size value in the "pushtr".
6674 */
6675 tupregs[ttop].dttk_size =
6676 dtrace_strlen((char *)(uintptr_t)regs[rd],
6677 regs[r2] ? regs[r2] :
6678 dtrace_strsize_default) + 1;
6679 } else {
6680 if (regs[r2] > LONG_MAX) {
6681 *flags |= CPU_DTRACE_ILLOP;
6682 break;
6683 }
6684
6685 tupregs[ttop].dttk_size = regs[r2];
6686 }
6687
6688 tupregs[ttop++].dttk_value = regs[rd];
6689 break;
6690
6691 case DIF_OP_PUSHTV:
6692 if (ttop == DIF_DTR_NREGS) {
6693 *flags |= CPU_DTRACE_TUPOFLOW;
6694 break;
6695 }
6696
6697 tupregs[ttop].dttk_value = regs[rd];
6698 tupregs[ttop++].dttk_size = 0;
6699 break;
6700
6701 case DIF_OP_POPTS:
6702 if (ttop != 0)
6703 ttop--;
6704 break;
6705
6706 case DIF_OP_FLUSHTS:
6707 ttop = 0;
6708 break;
6709
6710 case DIF_OP_LDGAA:
6711 case DIF_OP_LDTAA: {
6712 dtrace_dynvar_t *dvar;
6713 dtrace_key_t *key = tupregs;
6714 uint_t nkeys = ttop;
6715
6716 id = DIF_INSTR_VAR(instr);
6717 ASSERT(id >= DIF_VAR_OTHER_UBASE);
6718 id -= DIF_VAR_OTHER_UBASE;
6719
6720 key[nkeys].dttk_value = (uint64_t)id;
6721 key[nkeys++].dttk_size = 0;
6722
6723 if (DIF_INSTR_OP(instr) == DIF_OP_LDTAA) {
6724 DTRACE_TLS_THRKEY(key[nkeys].dttk_value);
6725 key[nkeys++].dttk_size = 0;
6726 VERIFY(id < vstate->dtvs_ntlocals);
6727 v = &vstate->dtvs_tlocals[id];
6728 } else {
6729 VERIFY(id < vstate->dtvs_nglobals);
6730 v = &vstate->dtvs_globals[id]->dtsv_var;
6731 }
6732
6733 dvar = dtrace_dynvar(dstate, nkeys, key,
6734 v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
6735 v->dtdv_type.dtdt_size : sizeof (uint64_t),
6736 DTRACE_DYNVAR_NOALLOC, mstate, vstate);
6737
6738 if (dvar == NULL) {
6739 regs[rd] = 0;
6740 break;
6741 }
6742
6743 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6744 regs[rd] = (uint64_t)(uintptr_t)dvar->dtdv_data;
6745 } else {
6746 regs[rd] = *((uint64_t *)dvar->dtdv_data);
6747 }
6748
6749 break;
6750 }
6751
6752 case DIF_OP_STGAA:
6753 case DIF_OP_STTAA: {
6754 dtrace_dynvar_t *dvar;
6755 dtrace_key_t *key = tupregs;
6756 uint_t nkeys = ttop;
6757
6758 id = DIF_INSTR_VAR(instr);
6759 ASSERT(id >= DIF_VAR_OTHER_UBASE);
6760 id -= DIF_VAR_OTHER_UBASE;
6761
6762 key[nkeys].dttk_value = (uint64_t)id;
6763 key[nkeys++].dttk_size = 0;
6764
6765 if (DIF_INSTR_OP(instr) == DIF_OP_STTAA) {
6766 DTRACE_TLS_THRKEY(key[nkeys].dttk_value);
6767 key[nkeys++].dttk_size = 0;
6768 VERIFY(id < vstate->dtvs_ntlocals);
6769 v = &vstate->dtvs_tlocals[id];
6770 } else {
6771 VERIFY(id < vstate->dtvs_nglobals);
6772 v = &vstate->dtvs_globals[id]->dtsv_var;
6773 }
6774
6775 dvar = dtrace_dynvar(dstate, nkeys, key,
6776 v->dtdv_type.dtdt_size > sizeof (uint64_t) ?
6777 v->dtdv_type.dtdt_size : sizeof (uint64_t),
6778 regs[rd] ? DTRACE_DYNVAR_ALLOC :
6779 DTRACE_DYNVAR_DEALLOC, mstate, vstate);
6780
6781 if (dvar == NULL)
6782 break;
6783
6784 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF) {
6785 size_t lim;
6786
6787 if (!dtrace_vcanload(
6788 (void *)(uintptr_t)regs[rd], &v->dtdv_type,
6789 &lim, mstate, vstate))
6790 break;
6791
6792 dtrace_vcopy((void *)(uintptr_t)regs[rd],
6793 dvar->dtdv_data, &v->dtdv_type, lim);
6794 } else {
6795 *((uint64_t *)dvar->dtdv_data) = regs[rd];
6796 }
6797
6798 break;
6799 }
6800
6801 case DIF_OP_ALLOCS: {
6802 uintptr_t ptr = P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
6803 size_t size = ptr - mstate->dtms_scratch_ptr + regs[r1];
6804
6805 /*
6806 * Rounding up the user allocation size could have
6807 * overflowed large, bogus allocations (like -1ULL) to
6808 * 0.
6809 */
6810 if (size < regs[r1] ||
6811 !DTRACE_INSCRATCH(mstate, size)) {
6812 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
6813 regs[rd] = 0;
6814 break;
6815 }
6816
6817 dtrace_bzero((void *) mstate->dtms_scratch_ptr, size);
6818 mstate->dtms_scratch_ptr += size;
6819 regs[rd] = ptr;
6820 break;
6821 }
6822
6823 case DIF_OP_COPYS:
6824 if (!dtrace_canstore(regs[rd], regs[r2],
6825 mstate, vstate)) {
6826 *flags |= CPU_DTRACE_BADADDR;
6827 *illval = regs[rd];
6828 break;
6829 }
6830
6831 if (!dtrace_canload(regs[r1], regs[r2], mstate, vstate))
6832 break;
6833
6834 dtrace_bcopy((void *)(uintptr_t)regs[r1],
6835 (void *)(uintptr_t)regs[rd], (size_t)regs[r2]);
6836 break;
6837
6838 case DIF_OP_STB:
6839 if (!dtrace_canstore(regs[rd], 1, mstate, vstate)) {
6840 *flags |= CPU_DTRACE_BADADDR;
6841 *illval = regs[rd];
6842 break;
6843 }
6844 *((uint8_t *)(uintptr_t)regs[rd]) = (uint8_t)regs[r1];
6845 break;
6846
6847 case DIF_OP_STH:
6848 if (!dtrace_canstore(regs[rd], 2, mstate, vstate)) {
6849 *flags |= CPU_DTRACE_BADADDR;
6850 *illval = regs[rd];
6851 break;
6852 }
6853 if (regs[rd] & 1) {
6854 *flags |= CPU_DTRACE_BADALIGN;
6855 *illval = regs[rd];
6856 break;
6857 }
6858 *((uint16_t *)(uintptr_t)regs[rd]) = (uint16_t)regs[r1];
6859 break;
6860
6861 case DIF_OP_STW:
6862 if (!dtrace_canstore(regs[rd], 4, mstate, vstate)) {
6863 *flags |= CPU_DTRACE_BADADDR;
6864 *illval = regs[rd];
6865 break;
6866 }
6867 if (regs[rd] & 3) {
6868 *flags |= CPU_DTRACE_BADALIGN;
6869 *illval = regs[rd];
6870 break;
6871 }
6872 *((uint32_t *)(uintptr_t)regs[rd]) = (uint32_t)regs[r1];
6873 break;
6874
6875 case DIF_OP_STX:
6876 if (!dtrace_canstore(regs[rd], 8, mstate, vstate)) {
6877 *flags |= CPU_DTRACE_BADADDR;
6878 *illval = regs[rd];
6879 break;
6880 }
6881 if (regs[rd] & 7) {
6882 *flags |= CPU_DTRACE_BADALIGN;
6883 *illval = regs[rd];
6884 break;
6885 }
6886 *((uint64_t *)(uintptr_t)regs[rd]) = regs[r1];
6887 break;
6888 }
6889 }
6890
6891 if (!(*flags & CPU_DTRACE_FAULT))
6892 return (rval);
6893
6894 mstate->dtms_fltoffs = opc * sizeof (dif_instr_t);
6895 mstate->dtms_present |= DTRACE_MSTATE_FLTOFFS;
6896
6897 return (0);
6898 }
6899
6900 static void
6901 dtrace_action_breakpoint(dtrace_ecb_t *ecb)
6902 {
6903 dtrace_probe_t *probe = ecb->dte_probe;
6904 dtrace_provider_t *prov = probe->dtpr_provider;
6905 char c[DTRACE_FULLNAMELEN + 80], *str;
6906 char *msg = "dtrace: breakpoint action at probe ";
6907 char *ecbmsg = " (ecb ";
6908 uintptr_t mask = (0xf << (sizeof (uintptr_t) * NBBY / 4));
6909 uintptr_t val = (uintptr_t)ecb;
6910 int shift = (sizeof (uintptr_t) * NBBY) - 4, i = 0;
6911
6912 if (dtrace_destructive_disallow)
6913 return;
6914
6915 /*
6916 * It's impossible to be taking action on the NULL probe.
6917 */
6918 ASSERT(probe != NULL);
6919
6920 /*
6921 * This is a poor man's (destitute man's?) sprintf(): we want to
6922 * print the provider name, module name, function name and name of
6923 * the probe, along with the hex address of the ECB with the breakpoint
6924 * action -- all of which we must place in the character buffer by
6925 * hand.
6926 */
6927 while (*msg != '\0')
6928 c[i++] = *msg++;
6929
6930 for (str = prov->dtpv_name; *str != '\0'; str++)
6931 c[i++] = *str;
6932 c[i++] = ':';
6933
6934 for (str = probe->dtpr_mod; *str != '\0'; str++)
6935 c[i++] = *str;
6936 c[i++] = ':';
6937
6938 for (str = probe->dtpr_func; *str != '\0'; str++)
6939 c[i++] = *str;
6940 c[i++] = ':';
6941
6942 for (str = probe->dtpr_name; *str != '\0'; str++)
6943 c[i++] = *str;
6944
6945 while (*ecbmsg != '\0')
6946 c[i++] = *ecbmsg++;
6947
6948 while (shift >= 0) {
6949 mask = (uintptr_t)0xf << shift;
6950
6951 if (val >= ((uintptr_t)1 << shift))
6952 c[i++] = "0123456789abcdef"[(val & mask) >> shift];
6953 shift -= 4;
6954 }
6955
6956 c[i++] = ')';
6957 c[i] = '\0';
6958
6959 #ifdef illumos
6960 debug_enter(c);
6961 #else
6962 kdb_enter(KDB_WHY_DTRACE, "breakpoint action");
6963 #endif
6964 }
6965
6966 static void
6967 dtrace_action_panic(dtrace_ecb_t *ecb)
6968 {
6969 dtrace_probe_t *probe = ecb->dte_probe;
6970
6971 /*
6972 * It's impossible to be taking action on the NULL probe.
6973 */
6974 ASSERT(probe != NULL);
6975
6976 if (dtrace_destructive_disallow)
6977 return;
6978
6979 if (dtrace_panicked != NULL)
6980 return;
6981
6982 if (dtrace_casptr(&dtrace_panicked, NULL, curthread) != NULL)
6983 return;
6984
6985 /*
6986 * We won the right to panic. (We want to be sure that only one
6987 * thread calls panic() from dtrace_probe(), and that panic() is
6988 * called exactly once.)
6989 */
6990 dtrace_panic("dtrace: panic action at probe %s:%s:%s:%s (ecb %p)",
6991 probe->dtpr_provider->dtpv_name, probe->dtpr_mod,
6992 probe->dtpr_func, probe->dtpr_name, (void *)ecb);
6993 }
6994
6995 static void
6996 dtrace_action_raise(uint64_t sig)
6997 {
6998 if (dtrace_destructive_disallow)
6999 return;
7000
7001 if (sig >= NSIG) {
7002 DTRACE_CPUFLAG_SET(CPU_DTRACE_ILLOP);
7003 return;
7004 }
7005
7006 #ifdef illumos
7007 /*
7008 * raise() has a queue depth of 1 -- we ignore all subsequent
7009 * invocations of the raise() action.
7010 */
7011 if (curthread->t_dtrace_sig == 0)
7012 curthread->t_dtrace_sig = (uint8_t)sig;
7013
7014 curthread->t_sig_check = 1;
7015 aston(curthread);
7016 #else
7017 struct proc *p = curproc;
7018 PROC_LOCK(p);
7019 kern_psignal(p, sig);
7020 PROC_UNLOCK(p);
7021 #endif
7022 }
7023
7024 static void
7025 dtrace_action_stop(void)
7026 {
7027 if (dtrace_destructive_disallow)
7028 return;
7029
7030 #ifdef illumos
7031 if (!curthread->t_dtrace_stop) {
7032 curthread->t_dtrace_stop = 1;
7033 curthread->t_sig_check = 1;
7034 aston(curthread);
7035 }
7036 #else
7037 struct proc *p = curproc;
7038 PROC_LOCK(p);
7039 kern_psignal(p, SIGSTOP);
7040 PROC_UNLOCK(p);
7041 #endif
7042 }
7043
7044 static void
7045 dtrace_action_chill(dtrace_mstate_t *mstate, hrtime_t val)
7046 {
7047 hrtime_t now;
7048 volatile uint16_t *flags;
7049 #ifdef illumos
7050 cpu_t *cpu = CPU;
7051 #else
7052 cpu_t *cpu = &solaris_cpu[curcpu];
7053 #endif
7054
7055 if (dtrace_destructive_disallow)
7056 return;
7057
7058 flags = (volatile uint16_t *)&cpu_core[curcpu].cpuc_dtrace_flags;
7059
7060 now = dtrace_gethrtime();
7061
7062 if (now - cpu->cpu_dtrace_chillmark > dtrace_chill_interval) {
7063 /*
7064 * We need to advance the mark to the current time.
7065 */
7066 cpu->cpu_dtrace_chillmark = now;
7067 cpu->cpu_dtrace_chilled = 0;
7068 }
7069
7070 /*
7071 * Now check to see if the requested chill time would take us over
7072 * the maximum amount of time allowed in the chill interval. (Or
7073 * worse, if the calculation itself induces overflow.)
7074 */
7075 if (cpu->cpu_dtrace_chilled + val > dtrace_chill_max ||
7076 cpu->cpu_dtrace_chilled + val < cpu->cpu_dtrace_chilled) {
7077 *flags |= CPU_DTRACE_ILLOP;
7078 return;
7079 }
7080
7081 while (dtrace_gethrtime() - now < val)
7082 continue;
7083
7084 /*
7085 * Normally, we assure that the value of the variable "timestamp" does
7086 * not change within an ECB. The presence of chill() represents an
7087 * exception to this rule, however.
7088 */
7089 mstate->dtms_present &= ~DTRACE_MSTATE_TIMESTAMP;
7090 cpu->cpu_dtrace_chilled += val;
7091 }
7092
7093 static void
7094 dtrace_action_ustack(dtrace_mstate_t *mstate, dtrace_state_t *state,
7095 uint64_t *buf, uint64_t arg)
7096 {
7097 int nframes = DTRACE_USTACK_NFRAMES(arg);
7098 int strsize = DTRACE_USTACK_STRSIZE(arg);
7099 uint64_t *pcs = &buf[1], *fps;
7100 char *str = (char *)&pcs[nframes];
7101 int size, offs = 0, i, j;
7102 size_t rem;
7103 uintptr_t old = mstate->dtms_scratch_ptr, saved;
7104 uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags;
7105 char *sym;
7106
7107 /*
7108 * Should be taking a faster path if string space has not been
7109 * allocated.
7110 */
7111 ASSERT(strsize != 0);
7112
7113 /*
7114 * We will first allocate some temporary space for the frame pointers.
7115 */
7116 fps = (uint64_t *)P2ROUNDUP(mstate->dtms_scratch_ptr, 8);
7117 size = (uintptr_t)fps - mstate->dtms_scratch_ptr +
7118 (nframes * sizeof (uint64_t));
7119
7120 if (!DTRACE_INSCRATCH(mstate, size)) {
7121 /*
7122 * Not enough room for our frame pointers -- need to indicate
7123 * that we ran out of scratch space.
7124 */
7125 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOSCRATCH);
7126 return;
7127 }
7128
7129 mstate->dtms_scratch_ptr += size;
7130 saved = mstate->dtms_scratch_ptr;
7131
7132 /*
7133 * Now get a stack with both program counters and frame pointers.
7134 */
7135 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
7136 dtrace_getufpstack(buf, fps, nframes + 1);
7137 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
7138
7139 /*
7140 * If that faulted, we're cooked.
7141 */
7142 if (*flags & CPU_DTRACE_FAULT)
7143 goto out;
7144
7145 /*
7146 * Now we want to walk up the stack, calling the USTACK helper. For
7147 * each iteration, we restore the scratch pointer.
7148 */
7149 for (i = 0; i < nframes; i++) {
7150 mstate->dtms_scratch_ptr = saved;
7151
7152 if (offs >= strsize)
7153 break;
7154
7155 sym = (char *)(uintptr_t)dtrace_helper(
7156 DTRACE_HELPER_ACTION_USTACK,
7157 mstate, state, pcs[i], fps[i]);
7158
7159 /*
7160 * If we faulted while running the helper, we're going to
7161 * clear the fault and null out the corresponding string.
7162 */
7163 if (*flags & CPU_DTRACE_FAULT) {
7164 *flags &= ~CPU_DTRACE_FAULT;
7165 str[offs++] = '\0';
7166 continue;
7167 }
7168
7169 if (sym == NULL) {
7170 str[offs++] = '\0';
7171 continue;
7172 }
7173
7174 if (!dtrace_strcanload((uintptr_t)sym, strsize, &rem, mstate,
7175 &(state->dts_vstate))) {
7176 str[offs++] = '\0';
7177 continue;
7178 }
7179
7180 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
7181
7182 /*
7183 * Now copy in the string that the helper returned to us.
7184 */
7185 for (j = 0; offs + j < strsize && j < rem; j++) {
7186 if ((str[offs + j] = sym[j]) == '\0')
7187 break;
7188 }
7189
7190 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
7191
7192 offs += j + 1;
7193 }
7194
7195 if (offs >= strsize) {
7196 /*
7197 * If we didn't have room for all of the strings, we don't
7198 * abort processing -- this needn't be a fatal error -- but we
7199 * still want to increment a counter (dts_stkstroverflows) to
7200 * allow this condition to be warned about. (If this is from
7201 * a jstack() action, it is easily tuned via jstackstrsize.)
7202 */
7203 dtrace_error(&state->dts_stkstroverflows);
7204 }
7205
7206 while (offs < strsize)
7207 str[offs++] = '\0';
7208
7209 out:
7210 mstate->dtms_scratch_ptr = old;
7211 }
7212
7213 static void
7214 dtrace_store_by_ref(dtrace_difo_t *dp, caddr_t tomax, size_t size,
7215 size_t *valoffsp, uint64_t *valp, uint64_t end, int intuple, int dtkind)
7216 {
7217 volatile uint16_t *flags;
7218 uint64_t val = *valp;
7219 size_t valoffs = *valoffsp;
7220
7221 flags = (volatile uint16_t *)&cpu_core[curcpu].cpuc_dtrace_flags;
7222 ASSERT(dtkind == DIF_TF_BYREF || dtkind == DIF_TF_BYUREF);
7223
7224 /*
7225 * If this is a string, we're going to only load until we find the zero
7226 * byte -- after which we'll store zero bytes.
7227 */
7228 if (dp->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING) {
7229 char c = '\0' + 1;
7230 size_t s;
7231
7232 for (s = 0; s < size; s++) {
7233 if (c != '\0' && dtkind == DIF_TF_BYREF) {
7234 c = dtrace_load8(val++);
7235 } else if (c != '\0' && dtkind == DIF_TF_BYUREF) {
7236 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
7237 c = dtrace_fuword8((void *)(uintptr_t)val++);
7238 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
7239 if (*flags & CPU_DTRACE_FAULT)
7240 break;
7241 }
7242
7243 DTRACE_STORE(uint8_t, tomax, valoffs++, c);
7244
7245 if (c == '\0' && intuple)
7246 break;
7247 }
7248 } else {
7249 uint8_t c;
7250 while (valoffs < end) {
7251 if (dtkind == DIF_TF_BYREF) {
7252 c = dtrace_load8(val++);
7253 } else if (dtkind == DIF_TF_BYUREF) {
7254 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
7255 c = dtrace_fuword8((void *)(uintptr_t)val++);
7256 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
7257 if (*flags & CPU_DTRACE_FAULT)
7258 break;
7259 }
7260
7261 DTRACE_STORE(uint8_t, tomax,
7262 valoffs++, c);
7263 }
7264 }
7265
7266 *valp = val;
7267 *valoffsp = valoffs;
7268 }
7269
7270 /*
7271 * Disables interrupts and sets the per-thread inprobe flag. When DEBUG is
7272 * defined, we also assert that we are not recursing unless the probe ID is an
7273 * error probe.
7274 */
7275 static dtrace_icookie_t
7276 dtrace_probe_enter(dtrace_id_t id)
7277 {
7278 dtrace_icookie_t cookie;
7279
7280 cookie = dtrace_interrupt_disable();
7281
7282 /*
7283 * Unless this is an ERROR probe, we are not allowed to recurse in
7284 * dtrace_probe(). Recursing into DTrace probe usually means that a
7285 * function is instrumented that should not have been instrumented or
7286 * that the ordering guarantee of the records will be violated,
7287 * resulting in unexpected output. If there is an exception to this
7288 * assertion, a new case should be added.
7289 */
7290 ASSERT(curthread->t_dtrace_inprobe == 0 ||
7291 id == dtrace_probeid_error);
7292 curthread->t_dtrace_inprobe = 1;
7293
7294 return (cookie);
7295 }
7296
7297 /*
7298 * Clears the per-thread inprobe flag and enables interrupts.
7299 */
7300 static void
7301 dtrace_probe_exit(dtrace_icookie_t cookie)
7302 {
7303
7304 curthread->t_dtrace_inprobe = 0;
7305 dtrace_interrupt_enable(cookie);
7306 }
7307
7308 /*
7309 * If you're looking for the epicenter of DTrace, you just found it. This
7310 * is the function called by the provider to fire a probe -- from which all
7311 * subsequent probe-context DTrace activity emanates.
7312 */
7313 void
7314 dtrace_probe(dtrace_id_t id, uintptr_t arg0, uintptr_t arg1,
7315 uintptr_t arg2, uintptr_t arg3, uintptr_t arg4)
7316 {
7317 processorid_t cpuid;
7318 dtrace_icookie_t cookie;
7319 dtrace_probe_t *probe;
7320 dtrace_mstate_t mstate;
7321 dtrace_ecb_t *ecb;
7322 dtrace_action_t *act;
7323 intptr_t offs;
7324 size_t size;
7325 int vtime, onintr;
7326 volatile uint16_t *flags;
7327 hrtime_t now;
7328
7329 if (KERNEL_PANICKED())
7330 return;
7331
7332 #ifdef illumos
7333 /*
7334 * Kick out immediately if this CPU is still being born (in which case
7335 * curthread will be set to -1) or the current thread can't allow
7336 * probes in its current context.
7337 */
7338 if (((uintptr_t)curthread & 1) || (curthread->t_flag & T_DONTDTRACE))
7339 return;
7340 #endif
7341
7342 cookie = dtrace_probe_enter(id);
7343 probe = dtrace_probes[id - 1];
7344 cpuid = curcpu;
7345 onintr = CPU_ON_INTR(CPU);
7346
7347 if (!onintr && probe->dtpr_predcache != DTRACE_CACHEIDNONE &&
7348 probe->dtpr_predcache == curthread->t_predcache) {
7349 /*
7350 * We have hit in the predicate cache; we know that
7351 * this predicate would evaluate to be false.
7352 */
7353 dtrace_probe_exit(cookie);
7354 return;
7355 }
7356
7357 #ifdef illumos
7358 if (panic_quiesce) {
7359 #else
7360 if (KERNEL_PANICKED()) {
7361 #endif
7362 /*
7363 * We don't trace anything if we're panicking.
7364 */
7365 dtrace_probe_exit(cookie);
7366 return;
7367 }
7368
7369 now = mstate.dtms_timestamp = dtrace_gethrtime();
7370 mstate.dtms_present = DTRACE_MSTATE_TIMESTAMP;
7371 vtime = dtrace_vtime_references != 0;
7372
7373 if (vtime && curthread->t_dtrace_start)
7374 curthread->t_dtrace_vtime += now - curthread->t_dtrace_start;
7375
7376 mstate.dtms_difo = NULL;
7377 mstate.dtms_probe = probe;
7378 mstate.dtms_strtok = 0;
7379 mstate.dtms_arg[0] = arg0;
7380 mstate.dtms_arg[1] = arg1;
7381 mstate.dtms_arg[2] = arg2;
7382 mstate.dtms_arg[3] = arg3;
7383 mstate.dtms_arg[4] = arg4;
7384
7385 flags = (volatile uint16_t *)&cpu_core[cpuid].cpuc_dtrace_flags;
7386
7387 for (ecb = probe->dtpr_ecb; ecb != NULL; ecb = ecb->dte_next) {
7388 dtrace_predicate_t *pred = ecb->dte_predicate;
7389 dtrace_state_t *state = ecb->dte_state;
7390 dtrace_buffer_t *buf = &state->dts_buffer[cpuid];
7391 dtrace_buffer_t *aggbuf = &state->dts_aggbuffer[cpuid];
7392 dtrace_vstate_t *vstate = &state->dts_vstate;
7393 dtrace_provider_t *prov = probe->dtpr_provider;
7394 uint64_t tracememsize = 0;
7395 int committed = 0;
7396 caddr_t tomax;
7397
7398 /*
7399 * A little subtlety with the following (seemingly innocuous)
7400 * declaration of the automatic 'val': by looking at the
7401 * code, you might think that it could be declared in the
7402 * action processing loop, below. (That is, it's only used in
7403 * the action processing loop.) However, it must be declared
7404 * out of that scope because in the case of DIF expression
7405 * arguments to aggregating actions, one iteration of the
7406 * action loop will use the last iteration's value.
7407 */
7408 uint64_t val = 0;
7409
7410 mstate.dtms_present = DTRACE_MSTATE_ARGS | DTRACE_MSTATE_PROBE;
7411 mstate.dtms_getf = NULL;
7412
7413 *flags &= ~CPU_DTRACE_ERROR;
7414
7415 if (prov == dtrace_provider) {
7416 /*
7417 * If dtrace itself is the provider of this probe,
7418 * we're only going to continue processing the ECB if
7419 * arg0 (the dtrace_state_t) is equal to the ECB's
7420 * creating state. (This prevents disjoint consumers
7421 * from seeing one another's metaprobes.)
7422 */
7423 if (arg0 != (uint64_t)(uintptr_t)state)
7424 continue;
7425 }
7426
7427 if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE) {
7428 /*
7429 * We're not currently active. If our provider isn't
7430 * the dtrace pseudo provider, we're not interested.
7431 */
7432 if (prov != dtrace_provider)
7433 continue;
7434
7435 /*
7436 * Now we must further check if we are in the BEGIN
7437 * probe. If we are, we will only continue processing
7438 * if we're still in WARMUP -- if one BEGIN enabling
7439 * has invoked the exit() action, we don't want to
7440 * evaluate subsequent BEGIN enablings.
7441 */
7442 if (probe->dtpr_id == dtrace_probeid_begin &&
7443 state->dts_activity != DTRACE_ACTIVITY_WARMUP) {
7444 ASSERT(state->dts_activity ==
7445 DTRACE_ACTIVITY_DRAINING);
7446 continue;
7447 }
7448 }
7449
7450 if (ecb->dte_cond) {
7451 /*
7452 * If the dte_cond bits indicate that this
7453 * consumer is only allowed to see user-mode firings
7454 * of this probe, call the provider's dtps_usermode()
7455 * entry point to check that the probe was fired
7456 * while in a user context. Skip this ECB if that's
7457 * not the case.
7458 */
7459 if ((ecb->dte_cond & DTRACE_COND_USERMODE) &&
7460 prov->dtpv_pops.dtps_usermode(prov->dtpv_arg,
7461 probe->dtpr_id, probe->dtpr_arg) == 0)
7462 continue;
7463
7464 #ifdef illumos
7465 /*
7466 * This is more subtle than it looks. We have to be
7467 * absolutely certain that CRED() isn't going to
7468 * change out from under us so it's only legit to
7469 * examine that structure if we're in constrained
7470 * situations. Currently, the only times we'll this
7471 * check is if a non-super-user has enabled the
7472 * profile or syscall providers -- providers that
7473 * allow visibility of all processes. For the
7474 * profile case, the check above will ensure that
7475 * we're examining a user context.
7476 */
7477 if (ecb->dte_cond & DTRACE_COND_OWNER) {
7478 cred_t *cr;
7479 cred_t *s_cr =
7480 ecb->dte_state->dts_cred.dcr_cred;
7481 proc_t *proc;
7482
7483 ASSERT(s_cr != NULL);
7484
7485 if ((cr = CRED()) == NULL ||
7486 s_cr->cr_uid != cr->cr_uid ||
7487 s_cr->cr_uid != cr->cr_ruid ||
7488 s_cr->cr_uid != cr->cr_suid ||
7489 s_cr->cr_gid != cr->cr_gid ||
7490 s_cr->cr_gid != cr->cr_rgid ||
7491 s_cr->cr_gid != cr->cr_sgid ||
7492 (proc = ttoproc(curthread)) == NULL ||
7493 (proc->p_flag & SNOCD))
7494 continue;
7495 }
7496
7497 if (ecb->dte_cond & DTRACE_COND_ZONEOWNER) {
7498 cred_t *cr;
7499 cred_t *s_cr =
7500 ecb->dte_state->dts_cred.dcr_cred;
7501
7502 ASSERT(s_cr != NULL);
7503
7504 if ((cr = CRED()) == NULL ||
7505 s_cr->cr_zone->zone_id !=
7506 cr->cr_zone->zone_id)
7507 continue;
7508 }
7509 #endif
7510 }
7511
7512 if (now - state->dts_alive > dtrace_deadman_timeout) {
7513 /*
7514 * We seem to be dead. Unless we (a) have kernel
7515 * destructive permissions (b) have explicitly enabled
7516 * destructive actions and (c) destructive actions have
7517 * not been disabled, we're going to transition into
7518 * the KILLED state, from which no further processing
7519 * on this state will be performed.
7520 */
7521 if (!dtrace_priv_kernel_destructive(state) ||
7522 !state->dts_cred.dcr_destructive ||
7523 dtrace_destructive_disallow) {
7524 void *activity = &state->dts_activity;
7525 dtrace_activity_t curstate;
7526
7527 do {
7528 curstate = state->dts_activity;
7529 } while (dtrace_cas32(activity, curstate,
7530 DTRACE_ACTIVITY_KILLED) != curstate);
7531
7532 continue;
7533 }
7534 }
7535
7536 if ((offs = dtrace_buffer_reserve(buf, ecb->dte_needed,
7537 ecb->dte_alignment, state, &mstate)) < 0)
7538 continue;
7539
7540 tomax = buf->dtb_tomax;
7541 ASSERT(tomax != NULL);
7542
7543 if (ecb->dte_size != 0) {
7544 dtrace_rechdr_t dtrh;
7545 if (!(mstate.dtms_present & DTRACE_MSTATE_TIMESTAMP)) {
7546 mstate.dtms_timestamp = dtrace_gethrtime();
7547 mstate.dtms_present |= DTRACE_MSTATE_TIMESTAMP;
7548 }
7549 ASSERT3U(ecb->dte_size, >=, sizeof (dtrace_rechdr_t));
7550 dtrh.dtrh_epid = ecb->dte_epid;
7551 DTRACE_RECORD_STORE_TIMESTAMP(&dtrh,
7552 mstate.dtms_timestamp);
7553 *((dtrace_rechdr_t *)(tomax + offs)) = dtrh;
7554 }
7555
7556 mstate.dtms_epid = ecb->dte_epid;
7557 mstate.dtms_present |= DTRACE_MSTATE_EPID;
7558
7559 if (state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL)
7560 mstate.dtms_access = DTRACE_ACCESS_KERNEL;
7561 else
7562 mstate.dtms_access = 0;
7563
7564 if (pred != NULL) {
7565 dtrace_difo_t *dp = pred->dtp_difo;
7566 uint64_t rval;
7567
7568 rval = dtrace_dif_emulate(dp, &mstate, vstate, state);
7569
7570 if (!(*flags & CPU_DTRACE_ERROR) && !rval) {
7571 dtrace_cacheid_t cid = probe->dtpr_predcache;
7572
7573 if (cid != DTRACE_CACHEIDNONE && !onintr) {
7574 /*
7575 * Update the predicate cache...
7576 */
7577 ASSERT(cid == pred->dtp_cacheid);
7578 curthread->t_predcache = cid;
7579 }
7580
7581 continue;
7582 }
7583 }
7584
7585 for (act = ecb->dte_action; !(*flags & CPU_DTRACE_ERROR) &&
7586 act != NULL; act = act->dta_next) {
7587 size_t valoffs;
7588 dtrace_difo_t *dp;
7589 dtrace_recdesc_t *rec = &act->dta_rec;
7590
7591 size = rec->dtrd_size;
7592 valoffs = offs + rec->dtrd_offset;
7593
7594 if (DTRACEACT_ISAGG(act->dta_kind)) {
7595 uint64_t v = 0xbad;
7596 dtrace_aggregation_t *agg;
7597
7598 agg = (dtrace_aggregation_t *)act;
7599
7600 if ((dp = act->dta_difo) != NULL)
7601 v = dtrace_dif_emulate(dp,
7602 &mstate, vstate, state);
7603
7604 if (*flags & CPU_DTRACE_ERROR)
7605 continue;
7606
7607 /*
7608 * Note that we always pass the expression
7609 * value from the previous iteration of the
7610 * action loop. This value will only be used
7611 * if there is an expression argument to the
7612 * aggregating action, denoted by the
7613 * dtag_hasarg field.
7614 */
7615 dtrace_aggregate(agg, buf,
7616 offs, aggbuf, v, val);
7617 continue;
7618 }
7619
7620 switch (act->dta_kind) {
7621 case DTRACEACT_STOP:
7622 if (dtrace_priv_proc_destructive(state))
7623 dtrace_action_stop();
7624 continue;
7625
7626 case DTRACEACT_BREAKPOINT:
7627 if (dtrace_priv_kernel_destructive(state))
7628 dtrace_action_breakpoint(ecb);
7629 continue;
7630
7631 case DTRACEACT_PANIC:
7632 if (dtrace_priv_kernel_destructive(state))
7633 dtrace_action_panic(ecb);
7634 continue;
7635
7636 case DTRACEACT_STACK:
7637 if (!dtrace_priv_kernel(state))
7638 continue;
7639
7640 dtrace_getpcstack((pc_t *)(tomax + valoffs),
7641 size / sizeof (pc_t), probe->dtpr_aframes,
7642 DTRACE_ANCHORED(probe) ? NULL :
7643 (uint32_t *)arg0);
7644 continue;
7645
7646 case DTRACEACT_JSTACK:
7647 case DTRACEACT_USTACK:
7648 if (!dtrace_priv_proc(state))
7649 continue;
7650
7651 /*
7652 * See comment in DIF_VAR_PID.
7653 */
7654 if (DTRACE_ANCHORED(mstate.dtms_probe) &&
7655 CPU_ON_INTR(CPU)) {
7656 int depth = DTRACE_USTACK_NFRAMES(
7657 rec->dtrd_arg) + 1;
7658
7659 dtrace_bzero((void *)(tomax + valoffs),
7660 DTRACE_USTACK_STRSIZE(rec->dtrd_arg)
7661 + depth * sizeof (uint64_t));
7662
7663 continue;
7664 }
7665
7666 if (DTRACE_USTACK_STRSIZE(rec->dtrd_arg) != 0 &&
7667 curproc->p_dtrace_helpers != NULL) {
7668 /*
7669 * This is the slow path -- we have
7670 * allocated string space, and we're
7671 * getting the stack of a process that
7672 * has helpers. Call into a separate
7673 * routine to perform this processing.
7674 */
7675 dtrace_action_ustack(&mstate, state,
7676 (uint64_t *)(tomax + valoffs),
7677 rec->dtrd_arg);
7678 continue;
7679 }
7680
7681 DTRACE_CPUFLAG_SET(CPU_DTRACE_NOFAULT);
7682 dtrace_getupcstack((uint64_t *)
7683 (tomax + valoffs),
7684 DTRACE_USTACK_NFRAMES(rec->dtrd_arg) + 1);
7685 DTRACE_CPUFLAG_CLEAR(CPU_DTRACE_NOFAULT);
7686 continue;
7687
7688 default:
7689 break;
7690 }
7691
7692 dp = act->dta_difo;
7693 ASSERT(dp != NULL);
7694
7695 val = dtrace_dif_emulate(dp, &mstate, vstate, state);
7696
7697 if (*flags & CPU_DTRACE_ERROR)
7698 continue;
7699
7700 switch (act->dta_kind) {
7701 case DTRACEACT_SPECULATE: {
7702 dtrace_rechdr_t *dtrh;
7703
7704 ASSERT(buf == &state->dts_buffer[cpuid]);
7705 buf = dtrace_speculation_buffer(state,
7706 cpuid, val);
7707
7708 if (buf == NULL) {
7709 *flags |= CPU_DTRACE_DROP;
7710 continue;
7711 }
7712
7713 offs = dtrace_buffer_reserve(buf,
7714 ecb->dte_needed, ecb->dte_alignment,
7715 state, NULL);
7716
7717 if (offs < 0) {
7718 *flags |= CPU_DTRACE_DROP;
7719 continue;
7720 }
7721
7722 tomax = buf->dtb_tomax;
7723 ASSERT(tomax != NULL);
7724
7725 if (ecb->dte_size == 0)
7726 continue;
7727
7728 ASSERT3U(ecb->dte_size, >=,
7729 sizeof (dtrace_rechdr_t));
7730 dtrh = ((void *)(tomax + offs));
7731 dtrh->dtrh_epid = ecb->dte_epid;
7732 /*
7733 * When the speculation is committed, all of
7734 * the records in the speculative buffer will
7735 * have their timestamps set to the commit
7736 * time. Until then, it is set to a sentinel
7737 * value, for debugability.
7738 */
7739 DTRACE_RECORD_STORE_TIMESTAMP(dtrh, UINT64_MAX);
7740 continue;
7741 }
7742
7743 case DTRACEACT_PRINTM: {
7744 /* The DIF returns a 'memref'. */
7745 uintptr_t *memref = (uintptr_t *)(uintptr_t) val;
7746
7747 /* Get the size from the memref. */
7748 size = memref[1];
7749
7750 /*
7751 * Check if the size exceeds the allocated
7752 * buffer size.
7753 */
7754 if (size + sizeof(uintptr_t) > dp->dtdo_rtype.dtdt_size) {
7755 /* Flag a drop! */
7756 *flags |= CPU_DTRACE_DROP;
7757 continue;
7758 }
7759
7760 /* Store the size in the buffer first. */
7761 DTRACE_STORE(uintptr_t, tomax,
7762 valoffs, size);
7763
7764 /*
7765 * Offset the buffer address to the start
7766 * of the data.
7767 */
7768 valoffs += sizeof(uintptr_t);
7769
7770 /*
7771 * Reset to the memory address rather than
7772 * the memref array, then let the BYREF
7773 * code below do the work to store the
7774 * memory data in the buffer.
7775 */
7776 val = memref[0];
7777 break;
7778 }
7779
7780 case DTRACEACT_CHILL:
7781 if (dtrace_priv_kernel_destructive(state))
7782 dtrace_action_chill(&mstate, val);
7783 continue;
7784
7785 case DTRACEACT_RAISE:
7786 if (dtrace_priv_proc_destructive(state))
7787 dtrace_action_raise(val);
7788 continue;
7789
7790 case DTRACEACT_COMMIT:
7791 ASSERT(!committed);
7792
7793 /*
7794 * We need to commit our buffer state.
7795 */
7796 if (ecb->dte_size)
7797 buf->dtb_offset = offs + ecb->dte_size;
7798 buf = &state->dts_buffer[cpuid];
7799 dtrace_speculation_commit(state, cpuid, val);
7800 committed = 1;
7801 continue;
7802
7803 case DTRACEACT_DISCARD:
7804 dtrace_speculation_discard(state, cpuid, val);
7805 continue;
7806
7807 case DTRACEACT_DIFEXPR:
7808 case DTRACEACT_LIBACT:
7809 case DTRACEACT_PRINTF:
7810 case DTRACEACT_PRINTA:
7811 case DTRACEACT_SYSTEM:
7812 case DTRACEACT_FREOPEN:
7813 case DTRACEACT_TRACEMEM:
7814 break;
7815
7816 case DTRACEACT_TRACEMEM_DYNSIZE:
7817 tracememsize = val;
7818 break;
7819
7820 case DTRACEACT_SYM:
7821 case DTRACEACT_MOD:
7822 if (!dtrace_priv_kernel(state))
7823 continue;
7824 break;
7825
7826 case DTRACEACT_USYM:
7827 case DTRACEACT_UMOD:
7828 case DTRACEACT_UADDR: {
7829 #ifdef illumos
7830 struct pid *pid = curthread->t_procp->p_pidp;
7831 #endif
7832
7833 if (!dtrace_priv_proc(state))
7834 continue;
7835
7836 DTRACE_STORE(uint64_t, tomax,
7837 #ifdef illumos
7838 valoffs, (uint64_t)pid->pid_id);
7839 #else
7840 valoffs, (uint64_t) curproc->p_pid);
7841 #endif
7842 DTRACE_STORE(uint64_t, tomax,
7843 valoffs + sizeof (uint64_t), val);
7844
7845 continue;
7846 }
7847
7848 case DTRACEACT_EXIT: {
7849 /*
7850 * For the exit action, we are going to attempt
7851 * to atomically set our activity to be
7852 * draining. If this fails (either because
7853 * another CPU has beat us to the exit action,
7854 * or because our current activity is something
7855 * other than ACTIVE or WARMUP), we will
7856 * continue. This assures that the exit action
7857 * can be successfully recorded at most once
7858 * when we're in the ACTIVE state. If we're
7859 * encountering the exit() action while in
7860 * COOLDOWN, however, we want to honor the new
7861 * status code. (We know that we're the only
7862 * thread in COOLDOWN, so there is no race.)
7863 */
7864 void *activity = &state->dts_activity;
7865 dtrace_activity_t curstate = state->dts_activity;
7866
7867 if (curstate == DTRACE_ACTIVITY_COOLDOWN)
7868 break;
7869
7870 if (curstate != DTRACE_ACTIVITY_WARMUP)
7871 curstate = DTRACE_ACTIVITY_ACTIVE;
7872
7873 if (dtrace_cas32(activity, curstate,
7874 DTRACE_ACTIVITY_DRAINING) != curstate) {
7875 *flags |= CPU_DTRACE_DROP;
7876 continue;
7877 }
7878
7879 break;
7880 }
7881
7882 default:
7883 ASSERT(0);
7884 }
7885
7886 if (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF ||
7887 dp->dtdo_rtype.dtdt_flags & DIF_TF_BYUREF) {
7888 uintptr_t end = valoffs + size;
7889
7890 if (tracememsize != 0 &&
7891 valoffs + tracememsize < end) {
7892 end = valoffs + tracememsize;
7893 tracememsize = 0;
7894 }
7895
7896 if (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF &&
7897 !dtrace_vcanload((void *)(uintptr_t)val,
7898 &dp->dtdo_rtype, NULL, &mstate, vstate))
7899 continue;
7900
7901 dtrace_store_by_ref(dp, tomax, size, &valoffs,
7902 &val, end, act->dta_intuple,
7903 dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF ?
7904 DIF_TF_BYREF: DIF_TF_BYUREF);
7905 continue;
7906 }
7907
7908 switch (size) {
7909 case 0:
7910 break;
7911
7912 case sizeof (uint8_t):
7913 DTRACE_STORE(uint8_t, tomax, valoffs, val);
7914 break;
7915 case sizeof (uint16_t):
7916 DTRACE_STORE(uint16_t, tomax, valoffs, val);
7917 break;
7918 case sizeof (uint32_t):
7919 DTRACE_STORE(uint32_t, tomax, valoffs, val);
7920 break;
7921 case sizeof (uint64_t):
7922 DTRACE_STORE(uint64_t, tomax, valoffs, val);
7923 break;
7924 default:
7925 /*
7926 * Any other size should have been returned by
7927 * reference, not by value.
7928 */
7929 ASSERT(0);
7930 break;
7931 }
7932 }
7933
7934 if (*flags & CPU_DTRACE_DROP)
7935 continue;
7936
7937 if (*flags & CPU_DTRACE_FAULT) {
7938 int ndx;
7939 dtrace_action_t *err;
7940
7941 buf->dtb_errors++;
7942
7943 if (probe->dtpr_id == dtrace_probeid_error) {
7944 /*
7945 * There's nothing we can do -- we had an
7946 * error on the error probe. We bump an
7947 * error counter to at least indicate that
7948 * this condition happened.
7949 */
7950 dtrace_error(&state->dts_dblerrors);
7951 continue;
7952 }
7953
7954 if (vtime) {
7955 /*
7956 * Before recursing on dtrace_probe(), we
7957 * need to explicitly clear out our start
7958 * time to prevent it from being accumulated
7959 * into t_dtrace_vtime.
7960 */
7961 curthread->t_dtrace_start = 0;
7962 }
7963
7964 /*
7965 * Iterate over the actions to figure out which action
7966 * we were processing when we experienced the error.
7967 * Note that act points _past_ the faulting action; if
7968 * act is ecb->dte_action, the fault was in the
7969 * predicate, if it's ecb->dte_action->dta_next it's
7970 * in action #1, and so on.
7971 */
7972 for (err = ecb->dte_action, ndx = 0;
7973 err != act; err = err->dta_next, ndx++)
7974 continue;
7975
7976 dtrace_probe_error(state, ecb->dte_epid, ndx,
7977 (mstate.dtms_present & DTRACE_MSTATE_FLTOFFS) ?
7978 mstate.dtms_fltoffs : -1, DTRACE_FLAGS2FLT(*flags),
7979 cpu_core[cpuid].cpuc_dtrace_illval);
7980
7981 continue;
7982 }
7983
7984 if (!committed)
7985 buf->dtb_offset = offs + ecb->dte_size;
7986 }
7987
7988 if (vtime)
7989 curthread->t_dtrace_start = dtrace_gethrtime();
7990
7991 dtrace_probe_exit(cookie);
7992 }
7993
7994 /*
7995 * DTrace Probe Hashing Functions
7996 *
7997 * The functions in this section (and indeed, the functions in remaining
7998 * sections) are not _called_ from probe context. (Any exceptions to this are
7999 * marked with a "Note:".) Rather, they are called from elsewhere in the
8000 * DTrace framework to look-up probes in, add probes to and remove probes from
8001 * the DTrace probe hashes. (Each probe is hashed by each element of the
8002 * probe tuple -- allowing for fast lookups, regardless of what was
8003 * specified.)
8004 */
8005 static uint_t
8006 dtrace_hash_str(const char *p)
8007 {
8008 unsigned int g;
8009 uint_t hval = 0;
8010
8011 while (*p) {
8012 hval = (hval << 4) + *p++;
8013 if ((g = (hval & 0xf0000000)) != 0)
8014 hval ^= g >> 24;
8015 hval &= ~g;
8016 }
8017 return (hval);
8018 }
8019
8020 static dtrace_hash_t *
8021 dtrace_hash_create(uintptr_t stroffs, uintptr_t nextoffs, uintptr_t prevoffs)
8022 {
8023 dtrace_hash_t *hash = kmem_zalloc(sizeof (dtrace_hash_t), KM_SLEEP);
8024
8025 hash->dth_stroffs = stroffs;
8026 hash->dth_nextoffs = nextoffs;
8027 hash->dth_prevoffs = prevoffs;
8028
8029 hash->dth_size = 1;
8030 hash->dth_mask = hash->dth_size - 1;
8031
8032 hash->dth_tab = kmem_zalloc(hash->dth_size *
8033 sizeof (dtrace_hashbucket_t *), KM_SLEEP);
8034
8035 return (hash);
8036 }
8037
8038 static void
8039 dtrace_hash_destroy(dtrace_hash_t *hash)
8040 {
8041 #ifdef DEBUG
8042 int i;
8043
8044 for (i = 0; i < hash->dth_size; i++)
8045 ASSERT(hash->dth_tab[i] == NULL);
8046 #endif
8047
8048 kmem_free(hash->dth_tab,
8049 hash->dth_size * sizeof (dtrace_hashbucket_t *));
8050 kmem_free(hash, sizeof (dtrace_hash_t));
8051 }
8052
8053 static void
8054 dtrace_hash_resize(dtrace_hash_t *hash)
8055 {
8056 int size = hash->dth_size, i, ndx;
8057 int new_size = hash->dth_size << 1;
8058 int new_mask = new_size - 1;
8059 dtrace_hashbucket_t **new_tab, *bucket, *next;
8060
8061 ASSERT((new_size & new_mask) == 0);
8062
8063 new_tab = kmem_zalloc(new_size * sizeof (void *), KM_SLEEP);
8064
8065 for (i = 0; i < size; i++) {
8066 for (bucket = hash->dth_tab[i]; bucket != NULL; bucket = next) {
8067 dtrace_probe_t *probe = bucket->dthb_chain;
8068
8069 ASSERT(probe != NULL);
8070 ndx = DTRACE_HASHSTR(hash, probe) & new_mask;
8071
8072 next = bucket->dthb_next;
8073 bucket->dthb_next = new_tab[ndx];
8074 new_tab[ndx] = bucket;
8075 }
8076 }
8077
8078 kmem_free(hash->dth_tab, hash->dth_size * sizeof (void *));
8079 hash->dth_tab = new_tab;
8080 hash->dth_size = new_size;
8081 hash->dth_mask = new_mask;
8082 }
8083
8084 static void
8085 dtrace_hash_add(dtrace_hash_t *hash, dtrace_probe_t *new)
8086 {
8087 int hashval = DTRACE_HASHSTR(hash, new);
8088 int ndx = hashval & hash->dth_mask;
8089 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
8090 dtrace_probe_t **nextp, **prevp;
8091
8092 for (; bucket != NULL; bucket = bucket->dthb_next) {
8093 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, new))
8094 goto add;
8095 }
8096
8097 if ((hash->dth_nbuckets >> 1) > hash->dth_size) {
8098 dtrace_hash_resize(hash);
8099 dtrace_hash_add(hash, new);
8100 return;
8101 }
8102
8103 bucket = kmem_zalloc(sizeof (dtrace_hashbucket_t), KM_SLEEP);
8104 bucket->dthb_next = hash->dth_tab[ndx];
8105 hash->dth_tab[ndx] = bucket;
8106 hash->dth_nbuckets++;
8107
8108 add:
8109 nextp = DTRACE_HASHNEXT(hash, new);
8110 ASSERT(*nextp == NULL && *(DTRACE_HASHPREV(hash, new)) == NULL);
8111 *nextp = bucket->dthb_chain;
8112
8113 if (bucket->dthb_chain != NULL) {
8114 prevp = DTRACE_HASHPREV(hash, bucket->dthb_chain);
8115 ASSERT(*prevp == NULL);
8116 *prevp = new;
8117 }
8118
8119 bucket->dthb_chain = new;
8120 bucket->dthb_len++;
8121 }
8122
8123 static dtrace_probe_t *
8124 dtrace_hash_lookup(dtrace_hash_t *hash, dtrace_probe_t *template)
8125 {
8126 int hashval = DTRACE_HASHSTR(hash, template);
8127 int ndx = hashval & hash->dth_mask;
8128 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
8129
8130 for (; bucket != NULL; bucket = bucket->dthb_next) {
8131 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template))
8132 return (bucket->dthb_chain);
8133 }
8134
8135 return (NULL);
8136 }
8137
8138 static int
8139 dtrace_hash_collisions(dtrace_hash_t *hash, dtrace_probe_t *template)
8140 {
8141 int hashval = DTRACE_HASHSTR(hash, template);
8142 int ndx = hashval & hash->dth_mask;
8143 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
8144
8145 for (; bucket != NULL; bucket = bucket->dthb_next) {
8146 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, template))
8147 return (bucket->dthb_len);
8148 }
8149
8150 return (0);
8151 }
8152
8153 static void
8154 dtrace_hash_remove(dtrace_hash_t *hash, dtrace_probe_t *probe)
8155 {
8156 int ndx = DTRACE_HASHSTR(hash, probe) & hash->dth_mask;
8157 dtrace_hashbucket_t *bucket = hash->dth_tab[ndx];
8158
8159 dtrace_probe_t **prevp = DTRACE_HASHPREV(hash, probe);
8160 dtrace_probe_t **nextp = DTRACE_HASHNEXT(hash, probe);
8161
8162 /*
8163 * Find the bucket that we're removing this probe from.
8164 */
8165 for (; bucket != NULL; bucket = bucket->dthb_next) {
8166 if (DTRACE_HASHEQ(hash, bucket->dthb_chain, probe))
8167 break;
8168 }
8169
8170 ASSERT(bucket != NULL);
8171
8172 if (*prevp == NULL) {
8173 if (*nextp == NULL) {
8174 /*
8175 * The removed probe was the only probe on this
8176 * bucket; we need to remove the bucket.
8177 */
8178 dtrace_hashbucket_t *b = hash->dth_tab[ndx];
8179
8180 ASSERT(bucket->dthb_chain == probe);
8181 ASSERT(b != NULL);
8182
8183 if (b == bucket) {
8184 hash->dth_tab[ndx] = bucket->dthb_next;
8185 } else {
8186 while (b->dthb_next != bucket)
8187 b = b->dthb_next;
8188 b->dthb_next = bucket->dthb_next;
8189 }
8190
8191 ASSERT(hash->dth_nbuckets > 0);
8192 hash->dth_nbuckets--;
8193 kmem_free(bucket, sizeof (dtrace_hashbucket_t));
8194 return;
8195 }
8196
8197 bucket->dthb_chain = *nextp;
8198 } else {
8199 *(DTRACE_HASHNEXT(hash, *prevp)) = *nextp;
8200 }
8201
8202 if (*nextp != NULL)
8203 *(DTRACE_HASHPREV(hash, *nextp)) = *prevp;
8204 }
8205
8206 /*
8207 * DTrace Utility Functions
8208 *
8209 * These are random utility functions that are _not_ called from probe context.
8210 */
8211 static int
8212 dtrace_badattr(const dtrace_attribute_t *a)
8213 {
8214 return (a->dtat_name > DTRACE_STABILITY_MAX ||
8215 a->dtat_data > DTRACE_STABILITY_MAX ||
8216 a->dtat_class > DTRACE_CLASS_MAX);
8217 }
8218
8219 /*
8220 * Return a duplicate copy of a string. If the specified string is NULL,
8221 * this function returns a zero-length string.
8222 */
8223 static char *
8224 dtrace_strdup(const char *str)
8225 {
8226 char *new = kmem_zalloc((str != NULL ? strlen(str) : 0) + 1, KM_SLEEP);
8227
8228 if (str != NULL)
8229 (void) strcpy(new, str);
8230
8231 return (new);
8232 }
8233
8234 #define DTRACE_ISALPHA(c) \
8235 (((c) >= 'a' && (c) <= 'z') || ((c) >= 'A' && (c) <= 'Z'))
8236
8237 static int
8238 dtrace_badname(const char *s)
8239 {
8240 char c;
8241
8242 if (s == NULL || (c = *s++) == '\0')
8243 return (0);
8244
8245 if (!DTRACE_ISALPHA(c) && c != '-' && c != '_' && c != '.')
8246 return (1);
8247
8248 while ((c = *s++) != '\0') {
8249 if (!DTRACE_ISALPHA(c) && (c < '' || c > '9') &&
8250 c != '-' && c != '_' && c != '.' && c != '`')
8251 return (1);
8252 }
8253
8254 return (0);
8255 }
8256
8257 static void
8258 dtrace_cred2priv(cred_t *cr, uint32_t *privp, uid_t *uidp, zoneid_t *zoneidp)
8259 {
8260 uint32_t priv;
8261
8262 #ifdef illumos
8263 if (cr == NULL || PRIV_POLICY_ONLY(cr, PRIV_ALL, B_FALSE)) {
8264 /*
8265 * For DTRACE_PRIV_ALL, the uid and zoneid don't matter.
8266 */
8267 priv = DTRACE_PRIV_ALL;
8268 } else {
8269 *uidp = crgetuid(cr);
8270 *zoneidp = crgetzoneid(cr);
8271
8272 priv = 0;
8273 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE))
8274 priv |= DTRACE_PRIV_KERNEL | DTRACE_PRIV_USER;
8275 else if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE))
8276 priv |= DTRACE_PRIV_USER;
8277 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE))
8278 priv |= DTRACE_PRIV_PROC;
8279 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
8280 priv |= DTRACE_PRIV_OWNER;
8281 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
8282 priv |= DTRACE_PRIV_ZONEOWNER;
8283 }
8284 #else
8285 priv = DTRACE_PRIV_ALL;
8286 #endif
8287
8288 *privp = priv;
8289 }
8290
8291 #ifdef DTRACE_ERRDEBUG
8292 static void
8293 dtrace_errdebug(const char *str)
8294 {
8295 int hval = dtrace_hash_str(str) % DTRACE_ERRHASHSZ;
8296 int occupied = 0;
8297
8298 mutex_enter(&dtrace_errlock);
8299 dtrace_errlast = str;
8300 dtrace_errthread = curthread;
8301
8302 while (occupied++ < DTRACE_ERRHASHSZ) {
8303 if (dtrace_errhash[hval].dter_msg == str) {
8304 dtrace_errhash[hval].dter_count++;
8305 goto out;
8306 }
8307
8308 if (dtrace_errhash[hval].dter_msg != NULL) {
8309 hval = (hval + 1) % DTRACE_ERRHASHSZ;
8310 continue;
8311 }
8312
8313 dtrace_errhash[hval].dter_msg = str;
8314 dtrace_errhash[hval].dter_count = 1;
8315 goto out;
8316 }
8317
8318 panic("dtrace: undersized error hash");
8319 out:
8320 mutex_exit(&dtrace_errlock);
8321 }
8322 #endif
8323
8324 /*
8325 * DTrace Matching Functions
8326 *
8327 * These functions are used to match groups of probes, given some elements of
8328 * a probe tuple, or some globbed expressions for elements of a probe tuple.
8329 */
8330 static int
8331 dtrace_match_priv(const dtrace_probe_t *prp, uint32_t priv, uid_t uid,
8332 zoneid_t zoneid)
8333 {
8334 if (priv != DTRACE_PRIV_ALL) {
8335 uint32_t ppriv = prp->dtpr_provider->dtpv_priv.dtpp_flags;
8336 uint32_t match = priv & ppriv;
8337
8338 /*
8339 * No PRIV_DTRACE_* privileges...
8340 */
8341 if ((priv & (DTRACE_PRIV_PROC | DTRACE_PRIV_USER |
8342 DTRACE_PRIV_KERNEL)) == 0)
8343 return (0);
8344
8345 /*
8346 * No matching bits, but there were bits to match...
8347 */
8348 if (match == 0 && ppriv != 0)
8349 return (0);
8350
8351 /*
8352 * Need to have permissions to the process, but don't...
8353 */
8354 if (((ppriv & ~match) & DTRACE_PRIV_OWNER) != 0 &&
8355 uid != prp->dtpr_provider->dtpv_priv.dtpp_uid) {
8356 return (0);
8357 }
8358
8359 /*
8360 * Need to be in the same zone unless we possess the
8361 * privilege to examine all zones.
8362 */
8363 if (((ppriv & ~match) & DTRACE_PRIV_ZONEOWNER) != 0 &&
8364 zoneid != prp->dtpr_provider->dtpv_priv.dtpp_zoneid) {
8365 return (0);
8366 }
8367 }
8368
8369 return (1);
8370 }
8371
8372 /*
8373 * dtrace_match_probe compares a dtrace_probe_t to a pre-compiled key, which
8374 * consists of input pattern strings and an ops-vector to evaluate them.
8375 * This function returns >0 for match, 0 for no match, and <0 for error.
8376 */
8377 static int
8378 dtrace_match_probe(const dtrace_probe_t *prp, const dtrace_probekey_t *pkp,
8379 uint32_t priv, uid_t uid, zoneid_t zoneid)
8380 {
8381 dtrace_provider_t *pvp = prp->dtpr_provider;
8382 int rv;
8383
8384 if (pvp->dtpv_defunct)
8385 return (0);
8386
8387 if ((rv = pkp->dtpk_pmatch(pvp->dtpv_name, pkp->dtpk_prov, 0)) <= 0)
8388 return (rv);
8389
8390 if ((rv = pkp->dtpk_mmatch(prp->dtpr_mod, pkp->dtpk_mod, 0)) <= 0)
8391 return (rv);
8392
8393 if ((rv = pkp->dtpk_fmatch(prp->dtpr_func, pkp->dtpk_func, 0)) <= 0)
8394 return (rv);
8395
8396 if ((rv = pkp->dtpk_nmatch(prp->dtpr_name, pkp->dtpk_name, 0)) <= 0)
8397 return (rv);
8398
8399 if (dtrace_match_priv(prp, priv, uid, zoneid) == 0)
8400 return (0);
8401
8402 return (rv);
8403 }
8404
8405 /*
8406 * dtrace_match_glob() is a safe kernel implementation of the gmatch(3GEN)
8407 * interface for matching a glob pattern 'p' to an input string 's'. Unlike
8408 * libc's version, the kernel version only applies to 8-bit ASCII strings.
8409 * In addition, all of the recursion cases except for '*' matching have been
8410 * unwound. For '*', we still implement recursive evaluation, but a depth
8411 * counter is maintained and matching is aborted if we recurse too deep.
8412 * The function returns 0 if no match, >0 if match, and <0 if recursion error.
8413 */
8414 static int
8415 dtrace_match_glob(const char *s, const char *p, int depth)
8416 {
8417 const char *olds;
8418 char s1, c;
8419 int gs;
8420
8421 if (depth > DTRACE_PROBEKEY_MAXDEPTH)
8422 return (-1);
8423
8424 if (s == NULL)
8425 s = ""; /* treat NULL as empty string */
8426
8427 top:
8428 olds = s;
8429 s1 = *s++;
8430
8431 if (p == NULL)
8432 return (0);
8433
8434 if ((c = *p++) == '\0')
8435 return (s1 == '\0');
8436
8437 switch (c) {
8438 case '[': {
8439 int ok = 0, notflag = 0;
8440 char lc = '\0';
8441
8442 if (s1 == '\0')
8443 return (0);
8444
8445 if (*p == '!') {
8446 notflag = 1;
8447 p++;
8448 }
8449
8450 if ((c = *p++) == '\0')
8451 return (0);
8452
8453 do {
8454 if (c == '-' && lc != '\0' && *p != ']') {
8455 if ((c = *p++) == '\0')
8456 return (0);
8457 if (c == '\\' && (c = *p++) == '\0')
8458 return (0);
8459
8460 if (notflag) {
8461 if (s1 < lc || s1 > c)
8462 ok++;
8463 else
8464 return (0);
8465 } else if (lc <= s1 && s1 <= c)
8466 ok++;
8467
8468 } else if (c == '\\' && (c = *p++) == '\0')
8469 return (0);
8470
8471 lc = c; /* save left-hand 'c' for next iteration */
8472
8473 if (notflag) {
8474 if (s1 != c)
8475 ok++;
8476 else
8477 return (0);
8478 } else if (s1 == c)
8479 ok++;
8480
8481 if ((c = *p++) == '\0')
8482 return (0);
8483
8484 } while (c != ']');
8485
8486 if (ok)
8487 goto top;
8488
8489 return (0);
8490 }
8491
8492 case '\\':
8493 if ((c = *p++) == '\0')
8494 return (0);
8495 /*FALLTHRU*/
8496
8497 default:
8498 if (c != s1)
8499 return (0);
8500 /*FALLTHRU*/
8501
8502 case '?':
8503 if (s1 != '\0')
8504 goto top;
8505 return (0);
8506
8507 case '*':
8508 while (*p == '*')
8509 p++; /* consecutive *'s are identical to a single one */
8510
8511 if (*p == '\0')
8512 return (1);
8513
8514 for (s = olds; *s != '\0'; s++) {
8515 if ((gs = dtrace_match_glob(s, p, depth + 1)) != 0)
8516 return (gs);
8517 }
8518
8519 return (0);
8520 }
8521 }
8522
8523 /*ARGSUSED*/
8524 static int
8525 dtrace_match_string(const char *s, const char *p, int depth)
8526 {
8527 return (s != NULL && strcmp(s, p) == 0);
8528 }
8529
8530 /*ARGSUSED*/
8531 static int
8532 dtrace_match_nul(const char *s, const char *p, int depth)
8533 {
8534 return (1); /* always match the empty pattern */
8535 }
8536
8537 /*ARGSUSED*/
8538 static int
8539 dtrace_match_nonzero(const char *s, const char *p, int depth)
8540 {
8541 return (s != NULL && s[0] != '\0');
8542 }
8543
8544 static int
8545 dtrace_match(const dtrace_probekey_t *pkp, uint32_t priv, uid_t uid,
8546 zoneid_t zoneid, int (*matched)(dtrace_probe_t *, void *), void *arg)
8547 {
8548 dtrace_probe_t template, *probe;
8549 dtrace_hash_t *hash = NULL;
8550 int len, best = INT_MAX, nmatched = 0;
8551 dtrace_id_t i;
8552
8553 ASSERT(MUTEX_HELD(&dtrace_lock));
8554
8555 /*
8556 * If the probe ID is specified in the key, just lookup by ID and
8557 * invoke the match callback once if a matching probe is found.
8558 */
8559 if (pkp->dtpk_id != DTRACE_IDNONE) {
8560 if ((probe = dtrace_probe_lookup_id(pkp->dtpk_id)) != NULL &&
8561 dtrace_match_probe(probe, pkp, priv, uid, zoneid) > 0) {
8562 (void) (*matched)(probe, arg);
8563 nmatched++;
8564 }
8565 return (nmatched);
8566 }
8567
8568 template.dtpr_mod = (char *)pkp->dtpk_mod;
8569 template.dtpr_func = (char *)pkp->dtpk_func;
8570 template.dtpr_name = (char *)pkp->dtpk_name;
8571
8572 /*
8573 * We want to find the most distinct of the module name, function
8574 * name, and name. So for each one that is not a glob pattern or
8575 * empty string, we perform a lookup in the corresponding hash and
8576 * use the hash table with the fewest collisions to do our search.
8577 */
8578 if (pkp->dtpk_mmatch == &dtrace_match_string &&
8579 (len = dtrace_hash_collisions(dtrace_bymod, &template)) < best) {
8580 best = len;
8581 hash = dtrace_bymod;
8582 }
8583
8584 if (pkp->dtpk_fmatch == &dtrace_match_string &&
8585 (len = dtrace_hash_collisions(dtrace_byfunc, &template)) < best) {
8586 best = len;
8587 hash = dtrace_byfunc;
8588 }
8589
8590 if (pkp->dtpk_nmatch == &dtrace_match_string &&
8591 (len = dtrace_hash_collisions(dtrace_byname, &template)) < best) {
8592 best = len;
8593 hash = dtrace_byname;
8594 }
8595
8596 /*
8597 * If we did not select a hash table, iterate over every probe and
8598 * invoke our callback for each one that matches our input probe key.
8599 */
8600 if (hash == NULL) {
8601 for (i = 0; i < dtrace_nprobes; i++) {
8602 if ((probe = dtrace_probes[i]) == NULL ||
8603 dtrace_match_probe(probe, pkp, priv, uid,
8604 zoneid) <= 0)
8605 continue;
8606
8607 nmatched++;
8608
8609 if ((*matched)(probe, arg) != DTRACE_MATCH_NEXT)
8610 break;
8611 }
8612
8613 return (nmatched);
8614 }
8615
8616 /*
8617 * If we selected a hash table, iterate over each probe of the same key
8618 * name and invoke the callback for every probe that matches the other
8619 * attributes of our input probe key.
8620 */
8621 for (probe = dtrace_hash_lookup(hash, &template); probe != NULL;
8622 probe = *(DTRACE_HASHNEXT(hash, probe))) {
8623
8624 if (dtrace_match_probe(probe, pkp, priv, uid, zoneid) <= 0)
8625 continue;
8626
8627 nmatched++;
8628
8629 if ((*matched)(probe, arg) != DTRACE_MATCH_NEXT)
8630 break;
8631 }
8632
8633 return (nmatched);
8634 }
8635
8636 /*
8637 * Return the function pointer dtrace_probecmp() should use to compare the
8638 * specified pattern with a string. For NULL or empty patterns, we select
8639 * dtrace_match_nul(). For glob pattern strings, we use dtrace_match_glob().
8640 * For non-empty non-glob strings, we use dtrace_match_string().
8641 */
8642 static dtrace_probekey_f *
8643 dtrace_probekey_func(const char *p)
8644 {
8645 char c;
8646
8647 if (p == NULL || *p == '\0')
8648 return (&dtrace_match_nul);
8649
8650 while ((c = *p++) != '\0') {
8651 if (c == '[' || c == '?' || c == '*' || c == '\\')
8652 return (&dtrace_match_glob);
8653 }
8654
8655 return (&dtrace_match_string);
8656 }
8657
8658 /*
8659 * Build a probe comparison key for use with dtrace_match_probe() from the
8660 * given probe description. By convention, a null key only matches anchored
8661 * probes: if each field is the empty string, reset dtpk_fmatch to
8662 * dtrace_match_nonzero().
8663 */
8664 static void
8665 dtrace_probekey(dtrace_probedesc_t *pdp, dtrace_probekey_t *pkp)
8666 {
8667 pkp->dtpk_prov = pdp->dtpd_provider;
8668 pkp->dtpk_pmatch = dtrace_probekey_func(pdp->dtpd_provider);
8669
8670 pkp->dtpk_mod = pdp->dtpd_mod;
8671 pkp->dtpk_mmatch = dtrace_probekey_func(pdp->dtpd_mod);
8672
8673 pkp->dtpk_func = pdp->dtpd_func;
8674 pkp->dtpk_fmatch = dtrace_probekey_func(pdp->dtpd_func);
8675
8676 pkp->dtpk_name = pdp->dtpd_name;
8677 pkp->dtpk_nmatch = dtrace_probekey_func(pdp->dtpd_name);
8678
8679 pkp->dtpk_id = pdp->dtpd_id;
8680
8681 if (pkp->dtpk_id == DTRACE_IDNONE &&
8682 pkp->dtpk_pmatch == &dtrace_match_nul &&
8683 pkp->dtpk_mmatch == &dtrace_match_nul &&
8684 pkp->dtpk_fmatch == &dtrace_match_nul &&
8685 pkp->dtpk_nmatch == &dtrace_match_nul)
8686 pkp->dtpk_fmatch = &dtrace_match_nonzero;
8687 }
8688
8689 /*
8690 * DTrace Provider-to-Framework API Functions
8691 *
8692 * These functions implement much of the Provider-to-Framework API, as
8693 * described in <sys/dtrace.h>. The parts of the API not in this section are
8694 * the functions in the API for probe management (found below), and
8695 * dtrace_probe() itself (found above).
8696 */
8697
8698 /*
8699 * Register the calling provider with the DTrace framework. This should
8700 * generally be called by DTrace providers in their attach(9E) entry point.
8701 */
8702 int
8703 dtrace_register(const char *name, const dtrace_pattr_t *pap, uint32_t priv,
8704 cred_t *cr, const dtrace_pops_t *pops, void *arg, dtrace_provider_id_t *idp)
8705 {
8706 dtrace_provider_t *provider;
8707
8708 if (name == NULL || pap == NULL || pops == NULL || idp == NULL) {
8709 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
8710 "arguments", name ? name : "<NULL>");
8711 return (EINVAL);
8712 }
8713
8714 if (name[0] == '\0' || dtrace_badname(name)) {
8715 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
8716 "provider name", name);
8717 return (EINVAL);
8718 }
8719
8720 if ((pops->dtps_provide == NULL && pops->dtps_provide_module == NULL) ||
8721 pops->dtps_enable == NULL || pops->dtps_disable == NULL ||
8722 pops->dtps_destroy == NULL ||
8723 ((pops->dtps_resume == NULL) != (pops->dtps_suspend == NULL))) {
8724 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
8725 "provider ops", name);
8726 return (EINVAL);
8727 }
8728
8729 if (dtrace_badattr(&pap->dtpa_provider) ||
8730 dtrace_badattr(&pap->dtpa_mod) ||
8731 dtrace_badattr(&pap->dtpa_func) ||
8732 dtrace_badattr(&pap->dtpa_name) ||
8733 dtrace_badattr(&pap->dtpa_args)) {
8734 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
8735 "provider attributes", name);
8736 return (EINVAL);
8737 }
8738
8739 if (priv & ~DTRACE_PRIV_ALL) {
8740 cmn_err(CE_WARN, "failed to register provider '%s': invalid "
8741 "privilege attributes", name);
8742 return (EINVAL);
8743 }
8744
8745 if ((priv & DTRACE_PRIV_KERNEL) &&
8746 (priv & (DTRACE_PRIV_USER | DTRACE_PRIV_OWNER)) &&
8747 pops->dtps_usermode == NULL) {
8748 cmn_err(CE_WARN, "failed to register provider '%s': need "
8749 "dtps_usermode() op for given privilege attributes", name);
8750 return (EINVAL);
8751 }
8752
8753 provider = kmem_zalloc(sizeof (dtrace_provider_t), KM_SLEEP);
8754 provider->dtpv_name = kmem_alloc(strlen(name) + 1, KM_SLEEP);
8755 (void) strcpy(provider->dtpv_name, name);
8756
8757 provider->dtpv_attr = *pap;
8758 provider->dtpv_priv.dtpp_flags = priv;
8759 if (cr != NULL) {
8760 provider->dtpv_priv.dtpp_uid = crgetuid(cr);
8761 provider->dtpv_priv.dtpp_zoneid = crgetzoneid(cr);
8762 }
8763 provider->dtpv_pops = *pops;
8764
8765 if (pops->dtps_provide == NULL) {
8766 ASSERT(pops->dtps_provide_module != NULL);
8767 provider->dtpv_pops.dtps_provide =
8768 (void (*)(void *, dtrace_probedesc_t *))dtrace_nullop;
8769 }
8770
8771 if (pops->dtps_provide_module == NULL) {
8772 ASSERT(pops->dtps_provide != NULL);
8773 provider->dtpv_pops.dtps_provide_module =
8774 (void (*)(void *, modctl_t *))dtrace_nullop;
8775 }
8776
8777 if (pops->dtps_suspend == NULL) {
8778 ASSERT(pops->dtps_resume == NULL);
8779 provider->dtpv_pops.dtps_suspend =
8780 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop;
8781 provider->dtpv_pops.dtps_resume =
8782 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop;
8783 }
8784
8785 provider->dtpv_arg = arg;
8786 *idp = (dtrace_provider_id_t)provider;
8787
8788 if (pops == &dtrace_provider_ops) {
8789 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
8790 ASSERT(MUTEX_HELD(&dtrace_lock));
8791 ASSERT(dtrace_anon.dta_enabling == NULL);
8792
8793 /*
8794 * We make sure that the DTrace provider is at the head of
8795 * the provider chain.
8796 */
8797 provider->dtpv_next = dtrace_provider;
8798 dtrace_provider = provider;
8799 return (0);
8800 }
8801
8802 mutex_enter(&dtrace_provider_lock);
8803 mutex_enter(&dtrace_lock);
8804
8805 /*
8806 * If there is at least one provider registered, we'll add this
8807 * provider after the first provider.
8808 */
8809 if (dtrace_provider != NULL) {
8810 provider->dtpv_next = dtrace_provider->dtpv_next;
8811 dtrace_provider->dtpv_next = provider;
8812 } else {
8813 dtrace_provider = provider;
8814 }
8815
8816 if (dtrace_retained != NULL) {
8817 dtrace_enabling_provide(provider);
8818
8819 /*
8820 * Now we need to call dtrace_enabling_matchall() -- which
8821 * will acquire cpu_lock and dtrace_lock. We therefore need
8822 * to drop all of our locks before calling into it...
8823 */
8824 mutex_exit(&dtrace_lock);
8825 mutex_exit(&dtrace_provider_lock);
8826 dtrace_enabling_matchall();
8827
8828 return (0);
8829 }
8830
8831 mutex_exit(&dtrace_lock);
8832 mutex_exit(&dtrace_provider_lock);
8833
8834 return (0);
8835 }
8836
8837 /*
8838 * Unregister the specified provider from the DTrace framework. This should
8839 * generally be called by DTrace providers in their detach(9E) entry point.
8840 */
8841 int
8842 dtrace_unregister(dtrace_provider_id_t id)
8843 {
8844 dtrace_provider_t *old = (dtrace_provider_t *)id;
8845 dtrace_provider_t *prev = NULL;
8846 int i, self = 0, noreap = 0;
8847 dtrace_probe_t *probe, *first = NULL;
8848
8849 if (old->dtpv_pops.dtps_enable ==
8850 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop) {
8851 /*
8852 * If DTrace itself is the provider, we're called with locks
8853 * already held.
8854 */
8855 ASSERT(old == dtrace_provider);
8856 #ifdef illumos
8857 ASSERT(dtrace_devi != NULL);
8858 #endif
8859 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
8860 ASSERT(MUTEX_HELD(&dtrace_lock));
8861 self = 1;
8862
8863 if (dtrace_provider->dtpv_next != NULL) {
8864 /*
8865 * There's another provider here; return failure.
8866 */
8867 return (EBUSY);
8868 }
8869 } else {
8870 mutex_enter(&dtrace_provider_lock);
8871 #ifdef illumos
8872 mutex_enter(&mod_lock);
8873 #endif
8874 mutex_enter(&dtrace_lock);
8875 }
8876
8877 /*
8878 * If anyone has /dev/dtrace open, or if there are anonymous enabled
8879 * probes, we refuse to let providers slither away, unless this
8880 * provider has already been explicitly invalidated.
8881 */
8882 if (!old->dtpv_defunct &&
8883 (dtrace_opens || (dtrace_anon.dta_state != NULL &&
8884 dtrace_anon.dta_state->dts_necbs > 0))) {
8885 if (!self) {
8886 mutex_exit(&dtrace_lock);
8887 #ifdef illumos
8888 mutex_exit(&mod_lock);
8889 #endif
8890 mutex_exit(&dtrace_provider_lock);
8891 }
8892 return (EBUSY);
8893 }
8894
8895 /*
8896 * Attempt to destroy the probes associated with this provider.
8897 */
8898 for (i = 0; i < dtrace_nprobes; i++) {
8899 if ((probe = dtrace_probes[i]) == NULL)
8900 continue;
8901
8902 if (probe->dtpr_provider != old)
8903 continue;
8904
8905 if (probe->dtpr_ecb == NULL)
8906 continue;
8907
8908 /*
8909 * If we are trying to unregister a defunct provider, and the
8910 * provider was made defunct within the interval dictated by
8911 * dtrace_unregister_defunct_reap, we'll (asynchronously)
8912 * attempt to reap our enablings. To denote that the provider
8913 * should reattempt to unregister itself at some point in the
8914 * future, we will return a differentiable error code (EAGAIN
8915 * instead of EBUSY) in this case.
8916 */
8917 if (dtrace_gethrtime() - old->dtpv_defunct >
8918 dtrace_unregister_defunct_reap)
8919 noreap = 1;
8920
8921 if (!self) {
8922 mutex_exit(&dtrace_lock);
8923 #ifdef illumos
8924 mutex_exit(&mod_lock);
8925 #endif
8926 mutex_exit(&dtrace_provider_lock);
8927 }
8928
8929 if (noreap)
8930 return (EBUSY);
8931
8932 (void) taskq_dispatch(dtrace_taskq,
8933 (task_func_t *)dtrace_enabling_reap, NULL, TQ_SLEEP);
8934
8935 return (EAGAIN);
8936 }
8937
8938 /*
8939 * All of the probes for this provider are disabled; we can safely
8940 * remove all of them from their hash chains and from the probe array.
8941 */
8942 for (i = 0; i < dtrace_nprobes; i++) {
8943 if ((probe = dtrace_probes[i]) == NULL)
8944 continue;
8945
8946 if (probe->dtpr_provider != old)
8947 continue;
8948
8949 dtrace_probes[i] = NULL;
8950
8951 dtrace_hash_remove(dtrace_bymod, probe);
8952 dtrace_hash_remove(dtrace_byfunc, probe);
8953 dtrace_hash_remove(dtrace_byname, probe);
8954
8955 if (first == NULL) {
8956 first = probe;
8957 probe->dtpr_nextmod = NULL;
8958 } else {
8959 probe->dtpr_nextmod = first;
8960 first = probe;
8961 }
8962 }
8963
8964 /*
8965 * The provider's probes have been removed from the hash chains and
8966 * from the probe array. Now issue a dtrace_sync() to be sure that
8967 * everyone has cleared out from any probe array processing.
8968 */
8969 dtrace_sync();
8970
8971 for (probe = first; probe != NULL; probe = first) {
8972 first = probe->dtpr_nextmod;
8973
8974 old->dtpv_pops.dtps_destroy(old->dtpv_arg, probe->dtpr_id,
8975 probe->dtpr_arg);
8976 kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
8977 kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
8978 kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
8979 #ifdef illumos
8980 vmem_free(dtrace_arena, (void *)(uintptr_t)(probe->dtpr_id), 1);
8981 #else
8982 free_unr(dtrace_arena, probe->dtpr_id);
8983 #endif
8984 kmem_free(probe, sizeof (dtrace_probe_t));
8985 }
8986
8987 if ((prev = dtrace_provider) == old) {
8988 #ifdef illumos
8989 ASSERT(self || dtrace_devi == NULL);
8990 ASSERT(old->dtpv_next == NULL || dtrace_devi == NULL);
8991 #endif
8992 dtrace_provider = old->dtpv_next;
8993 } else {
8994 while (prev != NULL && prev->dtpv_next != old)
8995 prev = prev->dtpv_next;
8996
8997 if (prev == NULL) {
8998 panic("attempt to unregister non-existent "
8999 "dtrace provider %p\n", (void *)id);
9000 }
9001
9002 prev->dtpv_next = old->dtpv_next;
9003 }
9004
9005 if (!self) {
9006 mutex_exit(&dtrace_lock);
9007 #ifdef illumos
9008 mutex_exit(&mod_lock);
9009 #endif
9010 mutex_exit(&dtrace_provider_lock);
9011 }
9012
9013 kmem_free(old->dtpv_name, strlen(old->dtpv_name) + 1);
9014 kmem_free(old, sizeof (dtrace_provider_t));
9015
9016 return (0);
9017 }
9018
9019 /*
9020 * Invalidate the specified provider. All subsequent probe lookups for the
9021 * specified provider will fail, but its probes will not be removed.
9022 */
9023 void
9024 dtrace_invalidate(dtrace_provider_id_t id)
9025 {
9026 dtrace_provider_t *pvp = (dtrace_provider_t *)id;
9027
9028 ASSERT(pvp->dtpv_pops.dtps_enable !=
9029 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop);
9030
9031 mutex_enter(&dtrace_provider_lock);
9032 mutex_enter(&dtrace_lock);
9033
9034 pvp->dtpv_defunct = dtrace_gethrtime();
9035
9036 mutex_exit(&dtrace_lock);
9037 mutex_exit(&dtrace_provider_lock);
9038 }
9039
9040 /*
9041 * Indicate whether or not DTrace has attached.
9042 */
9043 int
9044 dtrace_attached(void)
9045 {
9046 /*
9047 * dtrace_provider will be non-NULL iff the DTrace driver has
9048 * attached. (It's non-NULL because DTrace is always itself a
9049 * provider.)
9050 */
9051 return (dtrace_provider != NULL);
9052 }
9053
9054 /*
9055 * Remove all the unenabled probes for the given provider. This function is
9056 * not unlike dtrace_unregister(), except that it doesn't remove the provider
9057 * -- just as many of its associated probes as it can.
9058 */
9059 int
9060 dtrace_condense(dtrace_provider_id_t id)
9061 {
9062 dtrace_provider_t *prov = (dtrace_provider_t *)id;
9063 int i;
9064 dtrace_probe_t *probe;
9065
9066 /*
9067 * Make sure this isn't the dtrace provider itself.
9068 */
9069 ASSERT(prov->dtpv_pops.dtps_enable !=
9070 (void (*)(void *, dtrace_id_t, void *))dtrace_nullop);
9071
9072 mutex_enter(&dtrace_provider_lock);
9073 mutex_enter(&dtrace_lock);
9074
9075 /*
9076 * Attempt to destroy the probes associated with this provider.
9077 */
9078 for (i = 0; i < dtrace_nprobes; i++) {
9079 if ((probe = dtrace_probes[i]) == NULL)
9080 continue;
9081
9082 if (probe->dtpr_provider != prov)
9083 continue;
9084
9085 if (probe->dtpr_ecb != NULL)
9086 continue;
9087
9088 dtrace_probes[i] = NULL;
9089
9090 dtrace_hash_remove(dtrace_bymod, probe);
9091 dtrace_hash_remove(dtrace_byfunc, probe);
9092 dtrace_hash_remove(dtrace_byname, probe);
9093
9094 prov->dtpv_pops.dtps_destroy(prov->dtpv_arg, i + 1,
9095 probe->dtpr_arg);
9096 kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
9097 kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
9098 kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
9099 kmem_free(probe, sizeof (dtrace_probe_t));
9100 #ifdef illumos
9101 vmem_free(dtrace_arena, (void *)((uintptr_t)i + 1), 1);
9102 #else
9103 free_unr(dtrace_arena, i + 1);
9104 #endif
9105 }
9106
9107 mutex_exit(&dtrace_lock);
9108 mutex_exit(&dtrace_provider_lock);
9109
9110 return (0);
9111 }
9112
9113 /*
9114 * DTrace Probe Management Functions
9115 *
9116 * The functions in this section perform the DTrace probe management,
9117 * including functions to create probes, look-up probes, and call into the
9118 * providers to request that probes be provided. Some of these functions are
9119 * in the Provider-to-Framework API; these functions can be identified by the
9120 * fact that they are not declared "static".
9121 */
9122
9123 /*
9124 * Create a probe with the specified module name, function name, and name.
9125 */
9126 dtrace_id_t
9127 dtrace_probe_create(dtrace_provider_id_t prov, const char *mod,
9128 const char *func, const char *name, int aframes, void *arg)
9129 {
9130 dtrace_probe_t *probe, **probes;
9131 dtrace_provider_t *provider = (dtrace_provider_t *)prov;
9132 dtrace_id_t id;
9133
9134 if (provider == dtrace_provider) {
9135 ASSERT(MUTEX_HELD(&dtrace_lock));
9136 } else {
9137 mutex_enter(&dtrace_lock);
9138 }
9139
9140 #ifdef illumos
9141 id = (dtrace_id_t)(uintptr_t)vmem_alloc(dtrace_arena, 1,
9142 VM_BESTFIT | VM_SLEEP);
9143 #else
9144 id = alloc_unr(dtrace_arena);
9145 #endif
9146 probe = kmem_zalloc(sizeof (dtrace_probe_t), KM_SLEEP);
9147
9148 probe->dtpr_id = id;
9149 probe->dtpr_gen = dtrace_probegen++;
9150 probe->dtpr_mod = dtrace_strdup(mod);
9151 probe->dtpr_func = dtrace_strdup(func);
9152 probe->dtpr_name = dtrace_strdup(name);
9153 probe->dtpr_arg = arg;
9154 probe->dtpr_aframes = aframes;
9155 probe->dtpr_provider = provider;
9156
9157 dtrace_hash_add(dtrace_bymod, probe);
9158 dtrace_hash_add(dtrace_byfunc, probe);
9159 dtrace_hash_add(dtrace_byname, probe);
9160
9161 if (id - 1 >= dtrace_nprobes) {
9162 size_t osize = dtrace_nprobes * sizeof (dtrace_probe_t *);
9163 size_t nsize = osize << 1;
9164
9165 if (nsize == 0) {
9166 ASSERT(osize == 0);
9167 ASSERT(dtrace_probes == NULL);
9168 nsize = sizeof (dtrace_probe_t *);
9169 }
9170
9171 probes = kmem_zalloc(nsize, KM_SLEEP);
9172
9173 if (dtrace_probes == NULL) {
9174 ASSERT(osize == 0);
9175 dtrace_probes = probes;
9176 dtrace_nprobes = 1;
9177 } else {
9178 dtrace_probe_t **oprobes = dtrace_probes;
9179
9180 bcopy(oprobes, probes, osize);
9181 dtrace_membar_producer();
9182 dtrace_probes = probes;
9183
9184 dtrace_sync();
9185
9186 /*
9187 * All CPUs are now seeing the new probes array; we can
9188 * safely free the old array.
9189 */
9190 kmem_free(oprobes, osize);
9191 dtrace_nprobes <<= 1;
9192 }
9193
9194 ASSERT(id - 1 < dtrace_nprobes);
9195 }
9196
9197 ASSERT(dtrace_probes[id - 1] == NULL);
9198 dtrace_probes[id - 1] = probe;
9199
9200 if (provider != dtrace_provider)
9201 mutex_exit(&dtrace_lock);
9202
9203 return (id);
9204 }
9205
9206 static dtrace_probe_t *
9207 dtrace_probe_lookup_id(dtrace_id_t id)
9208 {
9209 ASSERT(MUTEX_HELD(&dtrace_lock));
9210
9211 if (id == 0 || id > dtrace_nprobes)
9212 return (NULL);
9213
9214 return (dtrace_probes[id - 1]);
9215 }
9216
9217 static int
9218 dtrace_probe_lookup_match(dtrace_probe_t *probe, void *arg)
9219 {
9220 *((dtrace_id_t *)arg) = probe->dtpr_id;
9221
9222 return (DTRACE_MATCH_DONE);
9223 }
9224
9225 /*
9226 * Look up a probe based on provider and one or more of module name, function
9227 * name and probe name.
9228 */
9229 dtrace_id_t
9230 dtrace_probe_lookup(dtrace_provider_id_t prid, char *mod,
9231 char *func, char *name)
9232 {
9233 dtrace_probekey_t pkey;
9234 dtrace_id_t id;
9235 int match;
9236
9237 pkey.dtpk_prov = ((dtrace_provider_t *)prid)->dtpv_name;
9238 pkey.dtpk_pmatch = &dtrace_match_string;
9239 pkey.dtpk_mod = mod;
9240 pkey.dtpk_mmatch = mod ? &dtrace_match_string : &dtrace_match_nul;
9241 pkey.dtpk_func = func;
9242 pkey.dtpk_fmatch = func ? &dtrace_match_string : &dtrace_match_nul;
9243 pkey.dtpk_name = name;
9244 pkey.dtpk_nmatch = name ? &dtrace_match_string : &dtrace_match_nul;
9245 pkey.dtpk_id = DTRACE_IDNONE;
9246
9247 mutex_enter(&dtrace_lock);
9248 match = dtrace_match(&pkey, DTRACE_PRIV_ALL, 0, 0,
9249 dtrace_probe_lookup_match, &id);
9250 mutex_exit(&dtrace_lock);
9251
9252 ASSERT(match == 1 || match == 0);
9253 return (match ? id : 0);
9254 }
9255
9256 /*
9257 * Returns the probe argument associated with the specified probe.
9258 */
9259 void *
9260 dtrace_probe_arg(dtrace_provider_id_t id, dtrace_id_t pid)
9261 {
9262 dtrace_probe_t *probe;
9263 void *rval = NULL;
9264
9265 mutex_enter(&dtrace_lock);
9266
9267 if ((probe = dtrace_probe_lookup_id(pid)) != NULL &&
9268 probe->dtpr_provider == (dtrace_provider_t *)id)
9269 rval = probe->dtpr_arg;
9270
9271 mutex_exit(&dtrace_lock);
9272
9273 return (rval);
9274 }
9275
9276 /*
9277 * Copy a probe into a probe description.
9278 */
9279 static void
9280 dtrace_probe_description(const dtrace_probe_t *prp, dtrace_probedesc_t *pdp)
9281 {
9282 bzero(pdp, sizeof (dtrace_probedesc_t));
9283 pdp->dtpd_id = prp->dtpr_id;
9284
9285 (void) strncpy(pdp->dtpd_provider,
9286 prp->dtpr_provider->dtpv_name, DTRACE_PROVNAMELEN - 1);
9287
9288 (void) strncpy(pdp->dtpd_mod, prp->dtpr_mod, DTRACE_MODNAMELEN - 1);
9289 (void) strncpy(pdp->dtpd_func, prp->dtpr_func, DTRACE_FUNCNAMELEN - 1);
9290 (void) strncpy(pdp->dtpd_name, prp->dtpr_name, DTRACE_NAMELEN - 1);
9291 }
9292
9293 /*
9294 * Called to indicate that a probe -- or probes -- should be provided by a
9295 * specfied provider. If the specified description is NULL, the provider will
9296 * be told to provide all of its probes. (This is done whenever a new
9297 * consumer comes along, or whenever a retained enabling is to be matched.) If
9298 * the specified description is non-NULL, the provider is given the
9299 * opportunity to dynamically provide the specified probe, allowing providers
9300 * to support the creation of probes on-the-fly. (So-called _autocreated_
9301 * probes.) If the provider is NULL, the operations will be applied to all
9302 * providers; if the provider is non-NULL the operations will only be applied
9303 * to the specified provider. The dtrace_provider_lock must be held, and the
9304 * dtrace_lock must _not_ be held -- the provider's dtps_provide() operation
9305 * will need to grab the dtrace_lock when it reenters the framework through
9306 * dtrace_probe_lookup(), dtrace_probe_create(), etc.
9307 */
9308 static void
9309 dtrace_probe_provide(dtrace_probedesc_t *desc, dtrace_provider_t *prv)
9310 {
9311 #ifdef illumos
9312 modctl_t *ctl;
9313 #endif
9314 int all = 0;
9315
9316 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
9317
9318 if (prv == NULL) {
9319 all = 1;
9320 prv = dtrace_provider;
9321 }
9322
9323 do {
9324 /*
9325 * First, call the blanket provide operation.
9326 */
9327 prv->dtpv_pops.dtps_provide(prv->dtpv_arg, desc);
9328
9329 #ifdef illumos
9330 /*
9331 * Now call the per-module provide operation. We will grab
9332 * mod_lock to prevent the list from being modified. Note
9333 * that this also prevents the mod_busy bits from changing.
9334 * (mod_busy can only be changed with mod_lock held.)
9335 */
9336 mutex_enter(&mod_lock);
9337
9338 ctl = &modules;
9339 do {
9340 if (ctl->mod_busy || ctl->mod_mp == NULL)
9341 continue;
9342
9343 prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, ctl);
9344
9345 } while ((ctl = ctl->mod_next) != &modules);
9346
9347 mutex_exit(&mod_lock);
9348 #endif
9349 } while (all && (prv = prv->dtpv_next) != NULL);
9350 }
9351
9352 #ifdef illumos
9353 /*
9354 * Iterate over each probe, and call the Framework-to-Provider API function
9355 * denoted by offs.
9356 */
9357 static void
9358 dtrace_probe_foreach(uintptr_t offs)
9359 {
9360 dtrace_provider_t *prov;
9361 void (*func)(void *, dtrace_id_t, void *);
9362 dtrace_probe_t *probe;
9363 dtrace_icookie_t cookie;
9364 int i;
9365
9366 /*
9367 * We disable interrupts to walk through the probe array. This is
9368 * safe -- the dtrace_sync() in dtrace_unregister() assures that we
9369 * won't see stale data.
9370 */
9371 cookie = dtrace_interrupt_disable();
9372
9373 for (i = 0; i < dtrace_nprobes; i++) {
9374 if ((probe = dtrace_probes[i]) == NULL)
9375 continue;
9376
9377 if (probe->dtpr_ecb == NULL) {
9378 /*
9379 * This probe isn't enabled -- don't call the function.
9380 */
9381 continue;
9382 }
9383
9384 prov = probe->dtpr_provider;
9385 func = *((void(**)(void *, dtrace_id_t, void *))
9386 ((uintptr_t)&prov->dtpv_pops + offs));
9387
9388 func(prov->dtpv_arg, i + 1, probe->dtpr_arg);
9389 }
9390
9391 dtrace_interrupt_enable(cookie);
9392 }
9393 #endif
9394
9395 static int
9396 dtrace_probe_enable(dtrace_probedesc_t *desc, dtrace_enabling_t *enab)
9397 {
9398 dtrace_probekey_t pkey;
9399 uint32_t priv;
9400 uid_t uid;
9401 zoneid_t zoneid;
9402
9403 ASSERT(MUTEX_HELD(&dtrace_lock));
9404 dtrace_ecb_create_cache = NULL;
9405
9406 if (desc == NULL) {
9407 /*
9408 * If we're passed a NULL description, we're being asked to
9409 * create an ECB with a NULL probe.
9410 */
9411 (void) dtrace_ecb_create_enable(NULL, enab);
9412 return (0);
9413 }
9414
9415 dtrace_probekey(desc, &pkey);
9416 dtrace_cred2priv(enab->dten_vstate->dtvs_state->dts_cred.dcr_cred,
9417 &priv, &uid, &zoneid);
9418
9419 return (dtrace_match(&pkey, priv, uid, zoneid, dtrace_ecb_create_enable,
9420 enab));
9421 }
9422
9423 /*
9424 * DTrace Helper Provider Functions
9425 */
9426 static void
9427 dtrace_dofattr2attr(dtrace_attribute_t *attr, const dof_attr_t dofattr)
9428 {
9429 attr->dtat_name = DOF_ATTR_NAME(dofattr);
9430 attr->dtat_data = DOF_ATTR_DATA(dofattr);
9431 attr->dtat_class = DOF_ATTR_CLASS(dofattr);
9432 }
9433
9434 static void
9435 dtrace_dofprov2hprov(dtrace_helper_provdesc_t *hprov,
9436 const dof_provider_t *dofprov, char *strtab)
9437 {
9438 hprov->dthpv_provname = strtab + dofprov->dofpv_name;
9439 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_provider,
9440 dofprov->dofpv_provattr);
9441 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_mod,
9442 dofprov->dofpv_modattr);
9443 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_func,
9444 dofprov->dofpv_funcattr);
9445 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_name,
9446 dofprov->dofpv_nameattr);
9447 dtrace_dofattr2attr(&hprov->dthpv_pattr.dtpa_args,
9448 dofprov->dofpv_argsattr);
9449 }
9450
9451 static void
9452 dtrace_helper_provide_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid)
9453 {
9454 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
9455 dof_hdr_t *dof = (dof_hdr_t *)daddr;
9456 dof_sec_t *str_sec, *prb_sec, *arg_sec, *off_sec, *enoff_sec;
9457 dof_provider_t *provider;
9458 dof_probe_t *probe;
9459 uint32_t *off, *enoff;
9460 uint8_t *arg;
9461 char *strtab;
9462 uint_t i, nprobes;
9463 dtrace_helper_provdesc_t dhpv;
9464 dtrace_helper_probedesc_t dhpb;
9465 dtrace_meta_t *meta = dtrace_meta_pid;
9466 dtrace_mops_t *mops = &meta->dtm_mops;
9467 void *parg;
9468
9469 provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
9470 str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
9471 provider->dofpv_strtab * dof->dofh_secsize);
9472 prb_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
9473 provider->dofpv_probes * dof->dofh_secsize);
9474 arg_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
9475 provider->dofpv_prargs * dof->dofh_secsize);
9476 off_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
9477 provider->dofpv_proffs * dof->dofh_secsize);
9478
9479 strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
9480 off = (uint32_t *)(uintptr_t)(daddr + off_sec->dofs_offset);
9481 arg = (uint8_t *)(uintptr_t)(daddr + arg_sec->dofs_offset);
9482 enoff = NULL;
9483
9484 /*
9485 * See dtrace_helper_provider_validate().
9486 */
9487 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
9488 provider->dofpv_prenoffs != DOF_SECT_NONE) {
9489 enoff_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
9490 provider->dofpv_prenoffs * dof->dofh_secsize);
9491 enoff = (uint32_t *)(uintptr_t)(daddr + enoff_sec->dofs_offset);
9492 }
9493
9494 nprobes = prb_sec->dofs_size / prb_sec->dofs_entsize;
9495
9496 /*
9497 * Create the provider.
9498 */
9499 dtrace_dofprov2hprov(&dhpv, provider, strtab);
9500
9501 if ((parg = mops->dtms_provide_pid(meta->dtm_arg, &dhpv, pid)) == NULL)
9502 return;
9503
9504 meta->dtm_count++;
9505
9506 /*
9507 * Create the probes.
9508 */
9509 for (i = 0; i < nprobes; i++) {
9510 probe = (dof_probe_t *)(uintptr_t)(daddr +
9511 prb_sec->dofs_offset + i * prb_sec->dofs_entsize);
9512
9513 /* See the check in dtrace_helper_provider_validate(). */
9514 if (strlen(strtab + probe->dofpr_func) >= DTRACE_FUNCNAMELEN)
9515 continue;
9516
9517 dhpb.dthpb_mod = dhp->dofhp_mod;
9518 dhpb.dthpb_func = strtab + probe->dofpr_func;
9519 dhpb.dthpb_name = strtab + probe->dofpr_name;
9520 dhpb.dthpb_base = probe->dofpr_addr;
9521 dhpb.dthpb_offs = off + probe->dofpr_offidx;
9522 dhpb.dthpb_noffs = probe->dofpr_noffs;
9523 if (enoff != NULL) {
9524 dhpb.dthpb_enoffs = enoff + probe->dofpr_enoffidx;
9525 dhpb.dthpb_nenoffs = probe->dofpr_nenoffs;
9526 } else {
9527 dhpb.dthpb_enoffs = NULL;
9528 dhpb.dthpb_nenoffs = 0;
9529 }
9530 dhpb.dthpb_args = arg + probe->dofpr_argidx;
9531 dhpb.dthpb_nargc = probe->dofpr_nargc;
9532 dhpb.dthpb_xargc = probe->dofpr_xargc;
9533 dhpb.dthpb_ntypes = strtab + probe->dofpr_nargv;
9534 dhpb.dthpb_xtypes = strtab + probe->dofpr_xargv;
9535
9536 mops->dtms_create_probe(meta->dtm_arg, parg, &dhpb);
9537 }
9538 }
9539
9540 static void
9541 dtrace_helper_provide(dof_helper_t *dhp, pid_t pid)
9542 {
9543 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
9544 dof_hdr_t *dof = (dof_hdr_t *)daddr;
9545 int i;
9546
9547 ASSERT(MUTEX_HELD(&dtrace_meta_lock));
9548
9549 for (i = 0; i < dof->dofh_secnum; i++) {
9550 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
9551 dof->dofh_secoff + i * dof->dofh_secsize);
9552
9553 if (sec->dofs_type != DOF_SECT_PROVIDER)
9554 continue;
9555
9556 dtrace_helper_provide_one(dhp, sec, pid);
9557 }
9558
9559 /*
9560 * We may have just created probes, so we must now rematch against
9561 * any retained enablings. Note that this call will acquire both
9562 * cpu_lock and dtrace_lock; the fact that we are holding
9563 * dtrace_meta_lock now is what defines the ordering with respect to
9564 * these three locks.
9565 */
9566 dtrace_enabling_matchall();
9567 }
9568
9569 static void
9570 dtrace_helper_provider_remove_one(dof_helper_t *dhp, dof_sec_t *sec, pid_t pid)
9571 {
9572 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
9573 dof_hdr_t *dof = (dof_hdr_t *)daddr;
9574 dof_sec_t *str_sec;
9575 dof_provider_t *provider;
9576 char *strtab;
9577 dtrace_helper_provdesc_t dhpv;
9578 dtrace_meta_t *meta = dtrace_meta_pid;
9579 dtrace_mops_t *mops = &meta->dtm_mops;
9580
9581 provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
9582 str_sec = (dof_sec_t *)(uintptr_t)(daddr + dof->dofh_secoff +
9583 provider->dofpv_strtab * dof->dofh_secsize);
9584
9585 strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
9586
9587 /*
9588 * Create the provider.
9589 */
9590 dtrace_dofprov2hprov(&dhpv, provider, strtab);
9591
9592 mops->dtms_remove_pid(meta->dtm_arg, &dhpv, pid);
9593
9594 meta->dtm_count--;
9595 }
9596
9597 static void
9598 dtrace_helper_provider_remove(dof_helper_t *dhp, pid_t pid)
9599 {
9600 uintptr_t daddr = (uintptr_t)dhp->dofhp_dof;
9601 dof_hdr_t *dof = (dof_hdr_t *)daddr;
9602 int i;
9603
9604 ASSERT(MUTEX_HELD(&dtrace_meta_lock));
9605
9606 for (i = 0; i < dof->dofh_secnum; i++) {
9607 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
9608 dof->dofh_secoff + i * dof->dofh_secsize);
9609
9610 if (sec->dofs_type != DOF_SECT_PROVIDER)
9611 continue;
9612
9613 dtrace_helper_provider_remove_one(dhp, sec, pid);
9614 }
9615 }
9616
9617 /*
9618 * DTrace Meta Provider-to-Framework API Functions
9619 *
9620 * These functions implement the Meta Provider-to-Framework API, as described
9621 * in <sys/dtrace.h>.
9622 */
9623 int
9624 dtrace_meta_register(const char *name, const dtrace_mops_t *mops, void *arg,
9625 dtrace_meta_provider_id_t *idp)
9626 {
9627 dtrace_meta_t *meta;
9628 dtrace_helpers_t *help, *next;
9629 int i;
9630
9631 *idp = DTRACE_METAPROVNONE;
9632
9633 /*
9634 * We strictly don't need the name, but we hold onto it for
9635 * debuggability. All hail error queues!
9636 */
9637 if (name == NULL) {
9638 cmn_err(CE_WARN, "failed to register meta-provider: "
9639 "invalid name");
9640 return (EINVAL);
9641 }
9642
9643 if (mops == NULL ||
9644 mops->dtms_create_probe == NULL ||
9645 mops->dtms_provide_pid == NULL ||
9646 mops->dtms_remove_pid == NULL) {
9647 cmn_err(CE_WARN, "failed to register meta-register %s: "
9648 "invalid ops", name);
9649 return (EINVAL);
9650 }
9651
9652 meta = kmem_zalloc(sizeof (dtrace_meta_t), KM_SLEEP);
9653 meta->dtm_mops = *mops;
9654 meta->dtm_name = kmem_alloc(strlen(name) + 1, KM_SLEEP);
9655 (void) strcpy(meta->dtm_name, name);
9656 meta->dtm_arg = arg;
9657
9658 mutex_enter(&dtrace_meta_lock);
9659 mutex_enter(&dtrace_lock);
9660
9661 if (dtrace_meta_pid != NULL) {
9662 mutex_exit(&dtrace_lock);
9663 mutex_exit(&dtrace_meta_lock);
9664 cmn_err(CE_WARN, "failed to register meta-register %s: "
9665 "user-land meta-provider exists", name);
9666 kmem_free(meta->dtm_name, strlen(meta->dtm_name) + 1);
9667 kmem_free(meta, sizeof (dtrace_meta_t));
9668 return (EINVAL);
9669 }
9670
9671 dtrace_meta_pid = meta;
9672 *idp = (dtrace_meta_provider_id_t)meta;
9673
9674 /*
9675 * If there are providers and probes ready to go, pass them
9676 * off to the new meta provider now.
9677 */
9678
9679 help = dtrace_deferred_pid;
9680 dtrace_deferred_pid = NULL;
9681
9682 mutex_exit(&dtrace_lock);
9683
9684 while (help != NULL) {
9685 for (i = 0; i < help->dthps_nprovs; i++) {
9686 dtrace_helper_provide(&help->dthps_provs[i]->dthp_prov,
9687 help->dthps_pid);
9688 }
9689
9690 next = help->dthps_next;
9691 help->dthps_next = NULL;
9692 help->dthps_prev = NULL;
9693 help->dthps_deferred = 0;
9694 help = next;
9695 }
9696
9697 mutex_exit(&dtrace_meta_lock);
9698
9699 return (0);
9700 }
9701
9702 int
9703 dtrace_meta_unregister(dtrace_meta_provider_id_t id)
9704 {
9705 dtrace_meta_t **pp, *old = (dtrace_meta_t *)id;
9706
9707 mutex_enter(&dtrace_meta_lock);
9708 mutex_enter(&dtrace_lock);
9709
9710 if (old == dtrace_meta_pid) {
9711 pp = &dtrace_meta_pid;
9712 } else {
9713 panic("attempt to unregister non-existent "
9714 "dtrace meta-provider %p\n", (void *)old);
9715 }
9716
9717 if (old->dtm_count != 0) {
9718 mutex_exit(&dtrace_lock);
9719 mutex_exit(&dtrace_meta_lock);
9720 return (EBUSY);
9721 }
9722
9723 *pp = NULL;
9724
9725 mutex_exit(&dtrace_lock);
9726 mutex_exit(&dtrace_meta_lock);
9727
9728 kmem_free(old->dtm_name, strlen(old->dtm_name) + 1);
9729 kmem_free(old, sizeof (dtrace_meta_t));
9730
9731 return (0);
9732 }
9733
9734
9735 /*
9736 * DTrace DIF Object Functions
9737 */
9738 static int
9739 dtrace_difo_err(uint_t pc, const char *format, ...)
9740 {
9741 if (dtrace_err_verbose) {
9742 va_list alist;
9743
9744 (void) uprintf("dtrace DIF object error: [%u]: ", pc);
9745 va_start(alist, format);
9746 (void) vuprintf(format, alist);
9747 va_end(alist);
9748 }
9749
9750 #ifdef DTRACE_ERRDEBUG
9751 dtrace_errdebug(format);
9752 #endif
9753 return (1);
9754 }
9755
9756 /*
9757 * Validate a DTrace DIF object by checking the IR instructions. The following
9758 * rules are currently enforced by dtrace_difo_validate():
9759 *
9760 * 1. Each instruction must have a valid opcode
9761 * 2. Each register, string, variable, or subroutine reference must be valid
9762 * 3. No instruction can modify register %r0 (must be zero)
9763 * 4. All instruction reserved bits must be set to zero
9764 * 5. The last instruction must be a "ret" instruction
9765 * 6. All branch targets must reference a valid instruction _after_ the branch
9766 */
9767 static int
9768 dtrace_difo_validate(dtrace_difo_t *dp, dtrace_vstate_t *vstate, uint_t nregs,
9769 cred_t *cr)
9770 {
9771 int err = 0, i;
9772 int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err;
9773 int kcheckload;
9774 uint_t pc;
9775 int maxglobal = -1, maxlocal = -1, maxtlocal = -1;
9776
9777 kcheckload = cr == NULL ||
9778 (vstate->dtvs_state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL) == 0;
9779
9780 dp->dtdo_destructive = 0;
9781
9782 for (pc = 0; pc < dp->dtdo_len && err == 0; pc++) {
9783 dif_instr_t instr = dp->dtdo_buf[pc];
9784
9785 uint_t r1 = DIF_INSTR_R1(instr);
9786 uint_t r2 = DIF_INSTR_R2(instr);
9787 uint_t rd = DIF_INSTR_RD(instr);
9788 uint_t rs = DIF_INSTR_RS(instr);
9789 uint_t label = DIF_INSTR_LABEL(instr);
9790 uint_t v = DIF_INSTR_VAR(instr);
9791 uint_t subr = DIF_INSTR_SUBR(instr);
9792 uint_t type = DIF_INSTR_TYPE(instr);
9793 uint_t op = DIF_INSTR_OP(instr);
9794
9795 switch (op) {
9796 case DIF_OP_OR:
9797 case DIF_OP_XOR:
9798 case DIF_OP_AND:
9799 case DIF_OP_SLL:
9800 case DIF_OP_SRL:
9801 case DIF_OP_SRA:
9802 case DIF_OP_SUB:
9803 case DIF_OP_ADD:
9804 case DIF_OP_MUL:
9805 case DIF_OP_SDIV:
9806 case DIF_OP_UDIV:
9807 case DIF_OP_SREM:
9808 case DIF_OP_UREM:
9809 case DIF_OP_COPYS:
9810 if (r1 >= nregs)
9811 err += efunc(pc, "invalid register %u\n", r1);
9812 if (r2 >= nregs)
9813 err += efunc(pc, "invalid register %u\n", r2);
9814 if (rd >= nregs)
9815 err += efunc(pc, "invalid register %u\n", rd);
9816 if (rd == 0)
9817 err += efunc(pc, "cannot write to %%r0\n");
9818 break;
9819 case DIF_OP_NOT:
9820 case DIF_OP_MOV:
9821 case DIF_OP_ALLOCS:
9822 if (r1 >= nregs)
9823 err += efunc(pc, "invalid register %u\n", r1);
9824 if (r2 != 0)
9825 err += efunc(pc, "non-zero reserved bits\n");
9826 if (rd >= nregs)
9827 err += efunc(pc, "invalid register %u\n", rd);
9828 if (rd == 0)
9829 err += efunc(pc, "cannot write to %%r0\n");
9830 break;
9831 case DIF_OP_LDSB:
9832 case DIF_OP_LDSH:
9833 case DIF_OP_LDSW:
9834 case DIF_OP_LDUB:
9835 case DIF_OP_LDUH:
9836 case DIF_OP_LDUW:
9837 case DIF_OP_LDX:
9838 if (r1 >= nregs)
9839 err += efunc(pc, "invalid register %u\n", r1);
9840 if (r2 != 0)
9841 err += efunc(pc, "non-zero reserved bits\n");
9842 if (rd >= nregs)
9843 err += efunc(pc, "invalid register %u\n", rd);
9844 if (rd == 0)
9845 err += efunc(pc, "cannot write to %%r0\n");
9846 if (kcheckload)
9847 dp->dtdo_buf[pc] = DIF_INSTR_LOAD(op +
9848 DIF_OP_RLDSB - DIF_OP_LDSB, r1, rd);
9849 break;
9850 case DIF_OP_RLDSB:
9851 case DIF_OP_RLDSH:
9852 case DIF_OP_RLDSW:
9853 case DIF_OP_RLDUB:
9854 case DIF_OP_RLDUH:
9855 case DIF_OP_RLDUW:
9856 case DIF_OP_RLDX:
9857 if (r1 >= nregs)
9858 err += efunc(pc, "invalid register %u\n", r1);
9859 if (r2 != 0)
9860 err += efunc(pc, "non-zero reserved bits\n");
9861 if (rd >= nregs)
9862 err += efunc(pc, "invalid register %u\n", rd);
9863 if (rd == 0)
9864 err += efunc(pc, "cannot write to %%r0\n");
9865 break;
9866 case DIF_OP_ULDSB:
9867 case DIF_OP_ULDSH:
9868 case DIF_OP_ULDSW:
9869 case DIF_OP_ULDUB:
9870 case DIF_OP_ULDUH:
9871 case DIF_OP_ULDUW:
9872 case DIF_OP_ULDX:
9873 if (r1 >= nregs)
9874 err += efunc(pc, "invalid register %u\n", r1);
9875 if (r2 != 0)
9876 err += efunc(pc, "non-zero reserved bits\n");
9877 if (rd >= nregs)
9878 err += efunc(pc, "invalid register %u\n", rd);
9879 if (rd == 0)
9880 err += efunc(pc, "cannot write to %%r0\n");
9881 break;
9882 case DIF_OP_STB:
9883 case DIF_OP_STH:
9884 case DIF_OP_STW:
9885 case DIF_OP_STX:
9886 if (r1 >= nregs)
9887 err += efunc(pc, "invalid register %u\n", r1);
9888 if (r2 != 0)
9889 err += efunc(pc, "non-zero reserved bits\n");
9890 if (rd >= nregs)
9891 err += efunc(pc, "invalid register %u\n", rd);
9892 if (rd == 0)
9893 err += efunc(pc, "cannot write to 0 address\n");
9894 break;
9895 case DIF_OP_CMP:
9896 case DIF_OP_SCMP:
9897 if (r1 >= nregs)
9898 err += efunc(pc, "invalid register %u\n", r1);
9899 if (r2 >= nregs)
9900 err += efunc(pc, "invalid register %u\n", r2);
9901 if (rd != 0)
9902 err += efunc(pc, "non-zero reserved bits\n");
9903 break;
9904 case DIF_OP_TST:
9905 if (r1 >= nregs)
9906 err += efunc(pc, "invalid register %u\n", r1);
9907 if (r2 != 0 || rd != 0)
9908 err += efunc(pc, "non-zero reserved bits\n");
9909 break;
9910 case DIF_OP_BA:
9911 case DIF_OP_BE:
9912 case DIF_OP_BNE:
9913 case DIF_OP_BG:
9914 case DIF_OP_BGU:
9915 case DIF_OP_BGE:
9916 case DIF_OP_BGEU:
9917 case DIF_OP_BL:
9918 case DIF_OP_BLU:
9919 case DIF_OP_BLE:
9920 case DIF_OP_BLEU:
9921 if (label >= dp->dtdo_len) {
9922 err += efunc(pc, "invalid branch target %u\n",
9923 label);
9924 }
9925 if (label <= pc) {
9926 err += efunc(pc, "backward branch to %u\n",
9927 label);
9928 }
9929 break;
9930 case DIF_OP_RET:
9931 if (r1 != 0 || r2 != 0)
9932 err += efunc(pc, "non-zero reserved bits\n");
9933 if (rd >= nregs)
9934 err += efunc(pc, "invalid register %u\n", rd);
9935 break;
9936 case DIF_OP_NOP:
9937 case DIF_OP_POPTS:
9938 case DIF_OP_FLUSHTS:
9939 if (r1 != 0 || r2 != 0 || rd != 0)
9940 err += efunc(pc, "non-zero reserved bits\n");
9941 break;
9942 case DIF_OP_SETX:
9943 if (DIF_INSTR_INTEGER(instr) >= dp->dtdo_intlen) {
9944 err += efunc(pc, "invalid integer ref %u\n",
9945 DIF_INSTR_INTEGER(instr));
9946 }
9947 if (rd >= nregs)
9948 err += efunc(pc, "invalid register %u\n", rd);
9949 if (rd == 0)
9950 err += efunc(pc, "cannot write to %%r0\n");
9951 break;
9952 case DIF_OP_SETS:
9953 if (DIF_INSTR_STRING(instr) >= dp->dtdo_strlen) {
9954 err += efunc(pc, "invalid string ref %u\n",
9955 DIF_INSTR_STRING(instr));
9956 }
9957 if (rd >= nregs)
9958 err += efunc(pc, "invalid register %u\n", rd);
9959 if (rd == 0)
9960 err += efunc(pc, "cannot write to %%r0\n");
9961 break;
9962 case DIF_OP_LDGA:
9963 case DIF_OP_LDTA:
9964 if (r1 > DIF_VAR_ARRAY_MAX)
9965 err += efunc(pc, "invalid array %u\n", r1);
9966 if (r2 >= nregs)
9967 err += efunc(pc, "invalid register %u\n", r2);
9968 if (rd >= nregs)
9969 err += efunc(pc, "invalid register %u\n", rd);
9970 if (rd == 0)
9971 err += efunc(pc, "cannot write to %%r0\n");
9972 break;
9973 case DIF_OP_LDGS:
9974 case DIF_OP_LDTS:
9975 case DIF_OP_LDLS:
9976 case DIF_OP_LDGAA:
9977 case DIF_OP_LDTAA:
9978 if (v < DIF_VAR_OTHER_MIN || v > DIF_VAR_OTHER_MAX)
9979 err += efunc(pc, "invalid variable %u\n", v);
9980 if (rd >= nregs)
9981 err += efunc(pc, "invalid register %u\n", rd);
9982 if (rd == 0)
9983 err += efunc(pc, "cannot write to %%r0\n");
9984 break;
9985 case DIF_OP_STGS:
9986 case DIF_OP_STTS:
9987 case DIF_OP_STLS:
9988 case DIF_OP_STGAA:
9989 case DIF_OP_STTAA:
9990 if (v < DIF_VAR_OTHER_UBASE || v > DIF_VAR_OTHER_MAX)
9991 err += efunc(pc, "invalid variable %u\n", v);
9992 if (rs >= nregs)
9993 err += efunc(pc, "invalid register %u\n", rd);
9994 break;
9995 case DIF_OP_CALL:
9996 if (subr > DIF_SUBR_MAX)
9997 err += efunc(pc, "invalid subr %u\n", subr);
9998 if (rd >= nregs)
9999 err += efunc(pc, "invalid register %u\n", rd);
10000 if (rd == 0)
10001 err += efunc(pc, "cannot write to %%r0\n");
10002
10003 if (subr == DIF_SUBR_COPYOUT ||
10004 subr == DIF_SUBR_COPYOUTSTR) {
10005 dp->dtdo_destructive = 1;
10006 }
10007
10008 if (subr == DIF_SUBR_GETF) {
10009 #ifdef __FreeBSD__
10010 err += efunc(pc, "getf() not supported");
10011 #else
10012 /*
10013 * If we have a getf() we need to record that
10014 * in our state. Note that our state can be
10015 * NULL if this is a helper -- but in that
10016 * case, the call to getf() is itself illegal,
10017 * and will be caught (slightly later) when
10018 * the helper is validated.
10019 */
10020 if (vstate->dtvs_state != NULL)
10021 vstate->dtvs_state->dts_getf++;
10022 #endif
10023 }
10024
10025 break;
10026 case DIF_OP_PUSHTR:
10027 if (type != DIF_TYPE_STRING && type != DIF_TYPE_CTF)
10028 err += efunc(pc, "invalid ref type %u\n", type);
10029 if (r2 >= nregs)
10030 err += efunc(pc, "invalid register %u\n", r2);
10031 if (rs >= nregs)
10032 err += efunc(pc, "invalid register %u\n", rs);
10033 break;
10034 case DIF_OP_PUSHTV:
10035 if (type != DIF_TYPE_CTF)
10036 err += efunc(pc, "invalid val type %u\n", type);
10037 if (r2 >= nregs)
10038 err += efunc(pc, "invalid register %u\n", r2);
10039 if (rs >= nregs)
10040 err += efunc(pc, "invalid register %u\n", rs);
10041 break;
10042 default:
10043 err += efunc(pc, "invalid opcode %u\n",
10044 DIF_INSTR_OP(instr));
10045 }
10046 }
10047
10048 if (dp->dtdo_len != 0 &&
10049 DIF_INSTR_OP(dp->dtdo_buf[dp->dtdo_len - 1]) != DIF_OP_RET) {
10050 err += efunc(dp->dtdo_len - 1,
10051 "expected 'ret' as last DIF instruction\n");
10052 }
10053
10054 if (!(dp->dtdo_rtype.dtdt_flags & (DIF_TF_BYREF | DIF_TF_BYUREF))) {
10055 /*
10056 * If we're not returning by reference, the size must be either
10057 * 0 or the size of one of the base types.
10058 */
10059 switch (dp->dtdo_rtype.dtdt_size) {
10060 case 0:
10061 case sizeof (uint8_t):
10062 case sizeof (uint16_t):
10063 case sizeof (uint32_t):
10064 case sizeof (uint64_t):
10065 break;
10066
10067 default:
10068 err += efunc(dp->dtdo_len - 1, "bad return size\n");
10069 }
10070 }
10071
10072 for (i = 0; i < dp->dtdo_varlen && err == 0; i++) {
10073 dtrace_difv_t *v = &dp->dtdo_vartab[i], *existing = NULL;
10074 dtrace_diftype_t *vt, *et;
10075 uint_t id, ndx;
10076
10077 if (v->dtdv_scope != DIFV_SCOPE_GLOBAL &&
10078 v->dtdv_scope != DIFV_SCOPE_THREAD &&
10079 v->dtdv_scope != DIFV_SCOPE_LOCAL) {
10080 err += efunc(i, "unrecognized variable scope %d\n",
10081 v->dtdv_scope);
10082 break;
10083 }
10084
10085 if (v->dtdv_kind != DIFV_KIND_ARRAY &&
10086 v->dtdv_kind != DIFV_KIND_SCALAR) {
10087 err += efunc(i, "unrecognized variable type %d\n",
10088 v->dtdv_kind);
10089 break;
10090 }
10091
10092 if ((id = v->dtdv_id) > DIF_VARIABLE_MAX) {
10093 err += efunc(i, "%d exceeds variable id limit\n", id);
10094 break;
10095 }
10096
10097 if (id < DIF_VAR_OTHER_UBASE)
10098 continue;
10099
10100 /*
10101 * For user-defined variables, we need to check that this
10102 * definition is identical to any previous definition that we
10103 * encountered.
10104 */
10105 ndx = id - DIF_VAR_OTHER_UBASE;
10106
10107 switch (v->dtdv_scope) {
10108 case DIFV_SCOPE_GLOBAL:
10109 if (maxglobal == -1 || ndx > maxglobal)
10110 maxglobal = ndx;
10111
10112 if (ndx < vstate->dtvs_nglobals) {
10113 dtrace_statvar_t *svar;
10114
10115 if ((svar = vstate->dtvs_globals[ndx]) != NULL)
10116 existing = &svar->dtsv_var;
10117 }
10118
10119 break;
10120
10121 case DIFV_SCOPE_THREAD:
10122 if (maxtlocal == -1 || ndx > maxtlocal)
10123 maxtlocal = ndx;
10124
10125 if (ndx < vstate->dtvs_ntlocals)
10126 existing = &vstate->dtvs_tlocals[ndx];
10127 break;
10128
10129 case DIFV_SCOPE_LOCAL:
10130 if (maxlocal == -1 || ndx > maxlocal)
10131 maxlocal = ndx;
10132
10133 if (ndx < vstate->dtvs_nlocals) {
10134 dtrace_statvar_t *svar;
10135
10136 if ((svar = vstate->dtvs_locals[ndx]) != NULL)
10137 existing = &svar->dtsv_var;
10138 }
10139
10140 break;
10141 }
10142
10143 vt = &v->dtdv_type;
10144
10145 if (vt->dtdt_flags & DIF_TF_BYREF) {
10146 if (vt->dtdt_size == 0) {
10147 err += efunc(i, "zero-sized variable\n");
10148 break;
10149 }
10150
10151 if ((v->dtdv_scope == DIFV_SCOPE_GLOBAL ||
10152 v->dtdv_scope == DIFV_SCOPE_LOCAL) &&
10153 vt->dtdt_size > dtrace_statvar_maxsize) {
10154 err += efunc(i, "oversized by-ref static\n");
10155 break;
10156 }
10157 }
10158
10159 if (existing == NULL || existing->dtdv_id == 0)
10160 continue;
10161
10162 ASSERT(existing->dtdv_id == v->dtdv_id);
10163 ASSERT(existing->dtdv_scope == v->dtdv_scope);
10164
10165 if (existing->dtdv_kind != v->dtdv_kind)
10166 err += efunc(i, "%d changed variable kind\n", id);
10167
10168 et = &existing->dtdv_type;
10169
10170 if (vt->dtdt_flags != et->dtdt_flags) {
10171 err += efunc(i, "%d changed variable type flags\n", id);
10172 break;
10173 }
10174
10175 if (vt->dtdt_size != 0 && vt->dtdt_size != et->dtdt_size) {
10176 err += efunc(i, "%d changed variable type size\n", id);
10177 break;
10178 }
10179 }
10180
10181 for (pc = 0; pc < dp->dtdo_len && err == 0; pc++) {
10182 dif_instr_t instr = dp->dtdo_buf[pc];
10183
10184 uint_t v = DIF_INSTR_VAR(instr);
10185 uint_t op = DIF_INSTR_OP(instr);
10186
10187 switch (op) {
10188 case DIF_OP_LDGS:
10189 case DIF_OP_LDGAA:
10190 case DIF_OP_STGS:
10191 case DIF_OP_STGAA:
10192 if (v > DIF_VAR_OTHER_UBASE + maxglobal)
10193 err += efunc(pc, "invalid variable %u\n", v);
10194 break;
10195 case DIF_OP_LDTS:
10196 case DIF_OP_LDTAA:
10197 case DIF_OP_STTS:
10198 case DIF_OP_STTAA:
10199 if (v > DIF_VAR_OTHER_UBASE + maxtlocal)
10200 err += efunc(pc, "invalid variable %u\n", v);
10201 break;
10202 case DIF_OP_LDLS:
10203 case DIF_OP_STLS:
10204 if (v > DIF_VAR_OTHER_UBASE + maxlocal)
10205 err += efunc(pc, "invalid variable %u\n", v);
10206 break;
10207 default:
10208 break;
10209 }
10210 }
10211
10212 return (err);
10213 }
10214
10215 /*
10216 * Validate a DTrace DIF object that it is to be used as a helper. Helpers
10217 * are much more constrained than normal DIFOs. Specifically, they may
10218 * not:
10219 *
10220 * 1. Make calls to subroutines other than copyin(), copyinstr() or
10221 * miscellaneous string routines
10222 * 2. Access DTrace variables other than the args[] array, and the
10223 * curthread, pid, ppid, tid, execname, zonename, uid and gid variables.
10224 * 3. Have thread-local variables.
10225 * 4. Have dynamic variables.
10226 */
10227 static int
10228 dtrace_difo_validate_helper(dtrace_difo_t *dp)
10229 {
10230 int (*efunc)(uint_t pc, const char *, ...) = dtrace_difo_err;
10231 int err = 0;
10232 uint_t pc;
10233
10234 for (pc = 0; pc < dp->dtdo_len; pc++) {
10235 dif_instr_t instr = dp->dtdo_buf[pc];
10236
10237 uint_t v = DIF_INSTR_VAR(instr);
10238 uint_t subr = DIF_INSTR_SUBR(instr);
10239 uint_t op = DIF_INSTR_OP(instr);
10240
10241 switch (op) {
10242 case DIF_OP_OR:
10243 case DIF_OP_XOR:
10244 case DIF_OP_AND:
10245 case DIF_OP_SLL:
10246 case DIF_OP_SRL:
10247 case DIF_OP_SRA:
10248 case DIF_OP_SUB:
10249 case DIF_OP_ADD:
10250 case DIF_OP_MUL:
10251 case DIF_OP_SDIV:
10252 case DIF_OP_UDIV:
10253 case DIF_OP_SREM:
10254 case DIF_OP_UREM:
10255 case DIF_OP_COPYS:
10256 case DIF_OP_NOT:
10257 case DIF_OP_MOV:
10258 case DIF_OP_RLDSB:
10259 case DIF_OP_RLDSH:
10260 case DIF_OP_RLDSW:
10261 case DIF_OP_RLDUB:
10262 case DIF_OP_RLDUH:
10263 case DIF_OP_RLDUW:
10264 case DIF_OP_RLDX:
10265 case DIF_OP_ULDSB:
10266 case DIF_OP_ULDSH:
10267 case DIF_OP_ULDSW:
10268 case DIF_OP_ULDUB:
10269 case DIF_OP_ULDUH:
10270 case DIF_OP_ULDUW:
10271 case DIF_OP_ULDX:
10272 case DIF_OP_STB:
10273 case DIF_OP_STH:
10274 case DIF_OP_STW:
10275 case DIF_OP_STX:
10276 case DIF_OP_ALLOCS:
10277 case DIF_OP_CMP:
10278 case DIF_OP_SCMP:
10279 case DIF_OP_TST:
10280 case DIF_OP_BA:
10281 case DIF_OP_BE:
10282 case DIF_OP_BNE:
10283 case DIF_OP_BG:
10284 case DIF_OP_BGU:
10285 case DIF_OP_BGE:
10286 case DIF_OP_BGEU:
10287 case DIF_OP_BL:
10288 case DIF_OP_BLU:
10289 case DIF_OP_BLE:
10290 case DIF_OP_BLEU:
10291 case DIF_OP_RET:
10292 case DIF_OP_NOP:
10293 case DIF_OP_POPTS:
10294 case DIF_OP_FLUSHTS:
10295 case DIF_OP_SETX:
10296 case DIF_OP_SETS:
10297 case DIF_OP_LDGA:
10298 case DIF_OP_LDLS:
10299 case DIF_OP_STGS:
10300 case DIF_OP_STLS:
10301 case DIF_OP_PUSHTR:
10302 case DIF_OP_PUSHTV:
10303 break;
10304
10305 case DIF_OP_LDGS:
10306 if (v >= DIF_VAR_OTHER_UBASE)
10307 break;
10308
10309 if (v >= DIF_VAR_ARG0 && v <= DIF_VAR_ARG9)
10310 break;
10311
10312 if (v == DIF_VAR_CURTHREAD || v == DIF_VAR_PID ||
10313 v == DIF_VAR_PPID || v == DIF_VAR_TID ||
10314 v == DIF_VAR_EXECARGS ||
10315 v == DIF_VAR_EXECNAME || v == DIF_VAR_ZONENAME ||
10316 v == DIF_VAR_UID || v == DIF_VAR_GID)
10317 break;
10318
10319 err += efunc(pc, "illegal variable %u\n", v);
10320 break;
10321
10322 case DIF_OP_LDTA:
10323 case DIF_OP_LDTS:
10324 case DIF_OP_LDGAA:
10325 case DIF_OP_LDTAA:
10326 err += efunc(pc, "illegal dynamic variable load\n");
10327 break;
10328
10329 case DIF_OP_STTS:
10330 case DIF_OP_STGAA:
10331 case DIF_OP_STTAA:
10332 err += efunc(pc, "illegal dynamic variable store\n");
10333 break;
10334
10335 case DIF_OP_CALL:
10336 if (subr == DIF_SUBR_ALLOCA ||
10337 subr == DIF_SUBR_BCOPY ||
10338 subr == DIF_SUBR_COPYIN ||
10339 subr == DIF_SUBR_COPYINTO ||
10340 subr == DIF_SUBR_COPYINSTR ||
10341 subr == DIF_SUBR_INDEX ||
10342 subr == DIF_SUBR_INET_NTOA ||
10343 subr == DIF_SUBR_INET_NTOA6 ||
10344 subr == DIF_SUBR_INET_NTOP ||
10345 subr == DIF_SUBR_JSON ||
10346 subr == DIF_SUBR_LLTOSTR ||
10347 subr == DIF_SUBR_STRTOLL ||
10348 subr == DIF_SUBR_RINDEX ||
10349 subr == DIF_SUBR_STRCHR ||
10350 subr == DIF_SUBR_STRJOIN ||
10351 subr == DIF_SUBR_STRRCHR ||
10352 subr == DIF_SUBR_STRSTR ||
10353 subr == DIF_SUBR_HTONS ||
10354 subr == DIF_SUBR_HTONL ||
10355 subr == DIF_SUBR_HTONLL ||
10356 subr == DIF_SUBR_NTOHS ||
10357 subr == DIF_SUBR_NTOHL ||
10358 subr == DIF_SUBR_NTOHLL ||
10359 subr == DIF_SUBR_MEMREF)
10360 break;
10361 #ifdef __FreeBSD__
10362 if (subr == DIF_SUBR_MEMSTR)
10363 break;
10364 #endif
10365
10366 err += efunc(pc, "invalid subr %u\n", subr);
10367 break;
10368
10369 default:
10370 err += efunc(pc, "invalid opcode %u\n",
10371 DIF_INSTR_OP(instr));
10372 }
10373 }
10374
10375 return (err);
10376 }
10377
10378 /*
10379 * Returns 1 if the expression in the DIF object can be cached on a per-thread
10380 * basis; 0 if not.
10381 */
10382 static int
10383 dtrace_difo_cacheable(dtrace_difo_t *dp)
10384 {
10385 int i;
10386
10387 if (dp == NULL)
10388 return (0);
10389
10390 for (i = 0; i < dp->dtdo_varlen; i++) {
10391 dtrace_difv_t *v = &dp->dtdo_vartab[i];
10392
10393 if (v->dtdv_scope != DIFV_SCOPE_GLOBAL)
10394 continue;
10395
10396 switch (v->dtdv_id) {
10397 case DIF_VAR_CURTHREAD:
10398 case DIF_VAR_PID:
10399 case DIF_VAR_TID:
10400 case DIF_VAR_EXECARGS:
10401 case DIF_VAR_EXECNAME:
10402 case DIF_VAR_ZONENAME:
10403 break;
10404
10405 default:
10406 return (0);
10407 }
10408 }
10409
10410 /*
10411 * This DIF object may be cacheable. Now we need to look for any
10412 * array loading instructions, any memory loading instructions, or
10413 * any stores to thread-local variables.
10414 */
10415 for (i = 0; i < dp->dtdo_len; i++) {
10416 uint_t op = DIF_INSTR_OP(dp->dtdo_buf[i]);
10417
10418 if ((op >= DIF_OP_LDSB && op <= DIF_OP_LDX) ||
10419 (op >= DIF_OP_ULDSB && op <= DIF_OP_ULDX) ||
10420 (op >= DIF_OP_RLDSB && op <= DIF_OP_RLDX) ||
10421 op == DIF_OP_LDGA || op == DIF_OP_STTS)
10422 return (0);
10423 }
10424
10425 return (1);
10426 }
10427
10428 static void
10429 dtrace_difo_hold(dtrace_difo_t *dp)
10430 {
10431 int i;
10432
10433 ASSERT(MUTEX_HELD(&dtrace_lock));
10434
10435 dp->dtdo_refcnt++;
10436 ASSERT(dp->dtdo_refcnt != 0);
10437
10438 /*
10439 * We need to check this DIF object for references to the variable
10440 * DIF_VAR_VTIMESTAMP.
10441 */
10442 for (i = 0; i < dp->dtdo_varlen; i++) {
10443 dtrace_difv_t *v = &dp->dtdo_vartab[i];
10444
10445 if (v->dtdv_id != DIF_VAR_VTIMESTAMP)
10446 continue;
10447
10448 if (dtrace_vtime_references++ == 0)
10449 dtrace_vtime_enable();
10450 }
10451 }
10452
10453 /*
10454 * This routine calculates the dynamic variable chunksize for a given DIF
10455 * object. The calculation is not fool-proof, and can probably be tricked by
10456 * malicious DIF -- but it works for all compiler-generated DIF. Because this
10457 * calculation is likely imperfect, dtrace_dynvar() is able to gracefully fail
10458 * if a dynamic variable size exceeds the chunksize.
10459 */
10460 static void
10461 dtrace_difo_chunksize(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
10462 {
10463 uint64_t sval = 0;
10464 dtrace_key_t tupregs[DIF_DTR_NREGS + 2]; /* +2 for thread and id */
10465 const dif_instr_t *text = dp->dtdo_buf;
10466 uint_t pc, srd = 0;
10467 uint_t ttop = 0;
10468 size_t size, ksize;
10469 uint_t id, i;
10470
10471 for (pc = 0; pc < dp->dtdo_len; pc++) {
10472 dif_instr_t instr = text[pc];
10473 uint_t op = DIF_INSTR_OP(instr);
10474 uint_t rd = DIF_INSTR_RD(instr);
10475 uint_t r1 = DIF_INSTR_R1(instr);
10476 uint_t nkeys = 0;
10477 uchar_t scope = 0;
10478
10479 dtrace_key_t *key = tupregs;
10480
10481 switch (op) {
10482 case DIF_OP_SETX:
10483 sval = dp->dtdo_inttab[DIF_INSTR_INTEGER(instr)];
10484 srd = rd;
10485 continue;
10486
10487 case DIF_OP_STTS:
10488 key = &tupregs[DIF_DTR_NREGS];
10489 key[0].dttk_size = 0;
10490 key[1].dttk_size = 0;
10491 nkeys = 2;
10492 scope = DIFV_SCOPE_THREAD;
10493 break;
10494
10495 case DIF_OP_STGAA:
10496 case DIF_OP_STTAA:
10497 nkeys = ttop;
10498
10499 if (DIF_INSTR_OP(instr) == DIF_OP_STTAA)
10500 key[nkeys++].dttk_size = 0;
10501
10502 key[nkeys++].dttk_size = 0;
10503
10504 if (op == DIF_OP_STTAA) {
10505 scope = DIFV_SCOPE_THREAD;
10506 } else {
10507 scope = DIFV_SCOPE_GLOBAL;
10508 }
10509
10510 break;
10511
10512 case DIF_OP_PUSHTR:
10513 if (ttop == DIF_DTR_NREGS)
10514 return;
10515
10516 if ((srd == 0 || sval == 0) && r1 == DIF_TYPE_STRING) {
10517 /*
10518 * If the register for the size of the "pushtr"
10519 * is %r0 (or the value is 0) and the type is
10520 * a string, we'll use the system-wide default
10521 * string size.
10522 */
10523 tupregs[ttop++].dttk_size =
10524 dtrace_strsize_default;
10525 } else {
10526 if (srd == 0)
10527 return;
10528
10529 if (sval > LONG_MAX)
10530 return;
10531
10532 tupregs[ttop++].dttk_size = sval;
10533 }
10534
10535 break;
10536
10537 case DIF_OP_PUSHTV:
10538 if (ttop == DIF_DTR_NREGS)
10539 return;
10540
10541 tupregs[ttop++].dttk_size = 0;
10542 break;
10543
10544 case DIF_OP_FLUSHTS:
10545 ttop = 0;
10546 break;
10547
10548 case DIF_OP_POPTS:
10549 if (ttop != 0)
10550 ttop--;
10551 break;
10552 }
10553
10554 sval = 0;
10555 srd = 0;
10556
10557 if (nkeys == 0)
10558 continue;
10559
10560 /*
10561 * We have a dynamic variable allocation; calculate its size.
10562 */
10563 for (ksize = 0, i = 0; i < nkeys; i++)
10564 ksize += P2ROUNDUP(key[i].dttk_size, sizeof (uint64_t));
10565
10566 size = sizeof (dtrace_dynvar_t);
10567 size += sizeof (dtrace_key_t) * (nkeys - 1);
10568 size += ksize;
10569
10570 /*
10571 * Now we need to determine the size of the stored data.
10572 */
10573 id = DIF_INSTR_VAR(instr);
10574
10575 for (i = 0; i < dp->dtdo_varlen; i++) {
10576 dtrace_difv_t *v = &dp->dtdo_vartab[i];
10577
10578 if (v->dtdv_id == id && v->dtdv_scope == scope) {
10579 size += v->dtdv_type.dtdt_size;
10580 break;
10581 }
10582 }
10583
10584 if (i == dp->dtdo_varlen)
10585 return;
10586
10587 /*
10588 * We have the size. If this is larger than the chunk size
10589 * for our dynamic variable state, reset the chunk size.
10590 */
10591 size = P2ROUNDUP(size, sizeof (uint64_t));
10592
10593 /*
10594 * Before setting the chunk size, check that we're not going
10595 * to set it to a negative value...
10596 */
10597 if (size > LONG_MAX)
10598 return;
10599
10600 /*
10601 * ...and make certain that we didn't badly overflow.
10602 */
10603 if (size < ksize || size < sizeof (dtrace_dynvar_t))
10604 return;
10605
10606 if (size > vstate->dtvs_dynvars.dtds_chunksize)
10607 vstate->dtvs_dynvars.dtds_chunksize = size;
10608 }
10609 }
10610
10611 static void
10612 dtrace_difo_init(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
10613 {
10614 int i, oldsvars, osz, nsz, otlocals, ntlocals;
10615 uint_t id;
10616
10617 ASSERT(MUTEX_HELD(&dtrace_lock));
10618 ASSERT(dp->dtdo_buf != NULL && dp->dtdo_len != 0);
10619
10620 for (i = 0; i < dp->dtdo_varlen; i++) {
10621 dtrace_difv_t *v = &dp->dtdo_vartab[i];
10622 dtrace_statvar_t *svar, ***svarp = NULL;
10623 size_t dsize = 0;
10624 uint8_t scope = v->dtdv_scope;
10625 int *np = NULL;
10626
10627 if ((id = v->dtdv_id) < DIF_VAR_OTHER_UBASE)
10628 continue;
10629
10630 id -= DIF_VAR_OTHER_UBASE;
10631
10632 switch (scope) {
10633 case DIFV_SCOPE_THREAD:
10634 while (id >= (otlocals = vstate->dtvs_ntlocals)) {
10635 dtrace_difv_t *tlocals;
10636
10637 if ((ntlocals = (otlocals << 1)) == 0)
10638 ntlocals = 1;
10639
10640 osz = otlocals * sizeof (dtrace_difv_t);
10641 nsz = ntlocals * sizeof (dtrace_difv_t);
10642
10643 tlocals = kmem_zalloc(nsz, KM_SLEEP);
10644
10645 if (osz != 0) {
10646 bcopy(vstate->dtvs_tlocals,
10647 tlocals, osz);
10648 kmem_free(vstate->dtvs_tlocals, osz);
10649 }
10650
10651 vstate->dtvs_tlocals = tlocals;
10652 vstate->dtvs_ntlocals = ntlocals;
10653 }
10654
10655 vstate->dtvs_tlocals[id] = *v;
10656 continue;
10657
10658 case DIFV_SCOPE_LOCAL:
10659 np = &vstate->dtvs_nlocals;
10660 svarp = &vstate->dtvs_locals;
10661
10662 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF)
10663 dsize = NCPU * (v->dtdv_type.dtdt_size +
10664 sizeof (uint64_t));
10665 else
10666 dsize = NCPU * sizeof (uint64_t);
10667
10668 break;
10669
10670 case DIFV_SCOPE_GLOBAL:
10671 np = &vstate->dtvs_nglobals;
10672 svarp = &vstate->dtvs_globals;
10673
10674 if (v->dtdv_type.dtdt_flags & DIF_TF_BYREF)
10675 dsize = v->dtdv_type.dtdt_size +
10676 sizeof (uint64_t);
10677
10678 break;
10679
10680 default:
10681 ASSERT(0);
10682 }
10683
10684 while (id >= (oldsvars = *np)) {
10685 dtrace_statvar_t **statics;
10686 int newsvars, oldsize, newsize;
10687
10688 if ((newsvars = (oldsvars << 1)) == 0)
10689 newsvars = 1;
10690
10691 oldsize = oldsvars * sizeof (dtrace_statvar_t *);
10692 newsize = newsvars * sizeof (dtrace_statvar_t *);
10693
10694 statics = kmem_zalloc(newsize, KM_SLEEP);
10695
10696 if (oldsize != 0) {
10697 bcopy(*svarp, statics, oldsize);
10698 kmem_free(*svarp, oldsize);
10699 }
10700
10701 *svarp = statics;
10702 *np = newsvars;
10703 }
10704
10705 if ((svar = (*svarp)[id]) == NULL) {
10706 svar = kmem_zalloc(sizeof (dtrace_statvar_t), KM_SLEEP);
10707 svar->dtsv_var = *v;
10708
10709 if ((svar->dtsv_size = dsize) != 0) {
10710 svar->dtsv_data = (uint64_t)(uintptr_t)
10711 kmem_zalloc(dsize, KM_SLEEP);
10712 }
10713
10714 (*svarp)[id] = svar;
10715 }
10716
10717 svar->dtsv_refcnt++;
10718 }
10719
10720 dtrace_difo_chunksize(dp, vstate);
10721 dtrace_difo_hold(dp);
10722 }
10723
10724 static dtrace_difo_t *
10725 dtrace_difo_duplicate(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
10726 {
10727 dtrace_difo_t *new;
10728 size_t sz;
10729
10730 ASSERT(dp->dtdo_buf != NULL);
10731 ASSERT(dp->dtdo_refcnt != 0);
10732
10733 new = kmem_zalloc(sizeof (dtrace_difo_t), KM_SLEEP);
10734
10735 ASSERT(dp->dtdo_buf != NULL);
10736 sz = dp->dtdo_len * sizeof (dif_instr_t);
10737 new->dtdo_buf = kmem_alloc(sz, KM_SLEEP);
10738 bcopy(dp->dtdo_buf, new->dtdo_buf, sz);
10739 new->dtdo_len = dp->dtdo_len;
10740
10741 if (dp->dtdo_strtab != NULL) {
10742 ASSERT(dp->dtdo_strlen != 0);
10743 new->dtdo_strtab = kmem_alloc(dp->dtdo_strlen, KM_SLEEP);
10744 bcopy(dp->dtdo_strtab, new->dtdo_strtab, dp->dtdo_strlen);
10745 new->dtdo_strlen = dp->dtdo_strlen;
10746 }
10747
10748 if (dp->dtdo_inttab != NULL) {
10749 ASSERT(dp->dtdo_intlen != 0);
10750 sz = dp->dtdo_intlen * sizeof (uint64_t);
10751 new->dtdo_inttab = kmem_alloc(sz, KM_SLEEP);
10752 bcopy(dp->dtdo_inttab, new->dtdo_inttab, sz);
10753 new->dtdo_intlen = dp->dtdo_intlen;
10754 }
10755
10756 if (dp->dtdo_vartab != NULL) {
10757 ASSERT(dp->dtdo_varlen != 0);
10758 sz = dp->dtdo_varlen * sizeof (dtrace_difv_t);
10759 new->dtdo_vartab = kmem_alloc(sz, KM_SLEEP);
10760 bcopy(dp->dtdo_vartab, new->dtdo_vartab, sz);
10761 new->dtdo_varlen = dp->dtdo_varlen;
10762 }
10763
10764 dtrace_difo_init(new, vstate);
10765 return (new);
10766 }
10767
10768 static void
10769 dtrace_difo_destroy(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
10770 {
10771 int i;
10772
10773 ASSERT(dp->dtdo_refcnt == 0);
10774
10775 for (i = 0; i < dp->dtdo_varlen; i++) {
10776 dtrace_difv_t *v = &dp->dtdo_vartab[i];
10777 dtrace_statvar_t *svar, **svarp = NULL;
10778 uint_t id;
10779 uint8_t scope = v->dtdv_scope;
10780 int *np = NULL;
10781
10782 switch (scope) {
10783 case DIFV_SCOPE_THREAD:
10784 continue;
10785
10786 case DIFV_SCOPE_LOCAL:
10787 np = &vstate->dtvs_nlocals;
10788 svarp = vstate->dtvs_locals;
10789 break;
10790
10791 case DIFV_SCOPE_GLOBAL:
10792 np = &vstate->dtvs_nglobals;
10793 svarp = vstate->dtvs_globals;
10794 break;
10795
10796 default:
10797 ASSERT(0);
10798 }
10799
10800 if ((id = v->dtdv_id) < DIF_VAR_OTHER_UBASE)
10801 continue;
10802
10803 id -= DIF_VAR_OTHER_UBASE;
10804 ASSERT(id < *np);
10805
10806 svar = svarp[id];
10807 ASSERT(svar != NULL);
10808 ASSERT(svar->dtsv_refcnt > 0);
10809
10810 if (--svar->dtsv_refcnt > 0)
10811 continue;
10812
10813 if (svar->dtsv_size != 0) {
10814 ASSERT(svar->dtsv_data != 0);
10815 kmem_free((void *)(uintptr_t)svar->dtsv_data,
10816 svar->dtsv_size);
10817 }
10818
10819 kmem_free(svar, sizeof (dtrace_statvar_t));
10820 svarp[id] = NULL;
10821 }
10822
10823 if (dp->dtdo_buf != NULL)
10824 kmem_free(dp->dtdo_buf, dp->dtdo_len * sizeof (dif_instr_t));
10825 if (dp->dtdo_inttab != NULL)
10826 kmem_free(dp->dtdo_inttab, dp->dtdo_intlen * sizeof (uint64_t));
10827 if (dp->dtdo_strtab != NULL)
10828 kmem_free(dp->dtdo_strtab, dp->dtdo_strlen);
10829 if (dp->dtdo_vartab != NULL)
10830 kmem_free(dp->dtdo_vartab, dp->dtdo_varlen * sizeof (dtrace_difv_t));
10831
10832 kmem_free(dp, sizeof (dtrace_difo_t));
10833 }
10834
10835 static void
10836 dtrace_difo_release(dtrace_difo_t *dp, dtrace_vstate_t *vstate)
10837 {
10838 int i;
10839
10840 ASSERT(MUTEX_HELD(&dtrace_lock));
10841 ASSERT(dp->dtdo_refcnt != 0);
10842
10843 for (i = 0; i < dp->dtdo_varlen; i++) {
10844 dtrace_difv_t *v = &dp->dtdo_vartab[i];
10845
10846 if (v->dtdv_id != DIF_VAR_VTIMESTAMP)
10847 continue;
10848
10849 ASSERT(dtrace_vtime_references > 0);
10850 if (--dtrace_vtime_references == 0)
10851 dtrace_vtime_disable();
10852 }
10853
10854 if (--dp->dtdo_refcnt == 0)
10855 dtrace_difo_destroy(dp, vstate);
10856 }
10857
10858 /*
10859 * DTrace Format Functions
10860 */
10861 static uint16_t
10862 dtrace_format_add(dtrace_state_t *state, char *str)
10863 {
10864 char *fmt, **new;
10865 uint16_t ndx, len = strlen(str) + 1;
10866
10867 fmt = kmem_zalloc(len, KM_SLEEP);
10868 bcopy(str, fmt, len);
10869
10870 for (ndx = 0; ndx < state->dts_nformats; ndx++) {
10871 if (state->dts_formats[ndx] == NULL) {
10872 state->dts_formats[ndx] = fmt;
10873 return (ndx + 1);
10874 }
10875 }
10876
10877 if (state->dts_nformats == USHRT_MAX) {
10878 /*
10879 * This is only likely if a denial-of-service attack is being
10880 * attempted. As such, it's okay to fail silently here.
10881 */
10882 kmem_free(fmt, len);
10883 return (0);
10884 }
10885
10886 /*
10887 * For simplicity, we always resize the formats array to be exactly the
10888 * number of formats.
10889 */
10890 ndx = state->dts_nformats++;
10891 new = kmem_alloc((ndx + 1) * sizeof (char *), KM_SLEEP);
10892
10893 if (state->dts_formats != NULL) {
10894 ASSERT(ndx != 0);
10895 bcopy(state->dts_formats, new, ndx * sizeof (char *));
10896 kmem_free(state->dts_formats, ndx * sizeof (char *));
10897 }
10898
10899 state->dts_formats = new;
10900 state->dts_formats[ndx] = fmt;
10901
10902 return (ndx + 1);
10903 }
10904
10905 static void
10906 dtrace_format_remove(dtrace_state_t *state, uint16_t format)
10907 {
10908 char *fmt;
10909
10910 ASSERT(state->dts_formats != NULL);
10911 ASSERT(format <= state->dts_nformats);
10912 ASSERT(state->dts_formats[format - 1] != NULL);
10913
10914 fmt = state->dts_formats[format - 1];
10915 kmem_free(fmt, strlen(fmt) + 1);
10916 state->dts_formats[format - 1] = NULL;
10917 }
10918
10919 static void
10920 dtrace_format_destroy(dtrace_state_t *state)
10921 {
10922 int i;
10923
10924 if (state->dts_nformats == 0) {
10925 ASSERT(state->dts_formats == NULL);
10926 return;
10927 }
10928
10929 ASSERT(state->dts_formats != NULL);
10930
10931 for (i = 0; i < state->dts_nformats; i++) {
10932 char *fmt = state->dts_formats[i];
10933
10934 if (fmt == NULL)
10935 continue;
10936
10937 kmem_free(fmt, strlen(fmt) + 1);
10938 }
10939
10940 kmem_free(state->dts_formats, state->dts_nformats * sizeof (char *));
10941 state->dts_nformats = 0;
10942 state->dts_formats = NULL;
10943 }
10944
10945 /*
10946 * DTrace Predicate Functions
10947 */
10948 static dtrace_predicate_t *
10949 dtrace_predicate_create(dtrace_difo_t *dp)
10950 {
10951 dtrace_predicate_t *pred;
10952
10953 ASSERT(MUTEX_HELD(&dtrace_lock));
10954 ASSERT(dp->dtdo_refcnt != 0);
10955
10956 pred = kmem_zalloc(sizeof (dtrace_predicate_t), KM_SLEEP);
10957 pred->dtp_difo = dp;
10958 pred->dtp_refcnt = 1;
10959
10960 if (!dtrace_difo_cacheable(dp))
10961 return (pred);
10962
10963 if (dtrace_predcache_id == DTRACE_CACHEIDNONE) {
10964 /*
10965 * This is only theoretically possible -- we have had 2^32
10966 * cacheable predicates on this machine. We cannot allow any
10967 * more predicates to become cacheable: as unlikely as it is,
10968 * there may be a thread caching a (now stale) predicate cache
10969 * ID. (N.B.: the temptation is being successfully resisted to
10970 * have this cmn_err() "Holy shit -- we executed this code!")
10971 */
10972 return (pred);
10973 }
10974
10975 pred->dtp_cacheid = dtrace_predcache_id++;
10976
10977 return (pred);
10978 }
10979
10980 static void
10981 dtrace_predicate_hold(dtrace_predicate_t *pred)
10982 {
10983 ASSERT(MUTEX_HELD(&dtrace_lock));
10984 ASSERT(pred->dtp_difo != NULL && pred->dtp_difo->dtdo_refcnt != 0);
10985 ASSERT(pred->dtp_refcnt > 0);
10986
10987 pred->dtp_refcnt++;
10988 }
10989
10990 static void
10991 dtrace_predicate_release(dtrace_predicate_t *pred, dtrace_vstate_t *vstate)
10992 {
10993 dtrace_difo_t *dp = pred->dtp_difo;
10994
10995 ASSERT(MUTEX_HELD(&dtrace_lock));
10996 ASSERT(dp != NULL && dp->dtdo_refcnt != 0);
10997 ASSERT(pred->dtp_refcnt > 0);
10998
10999 if (--pred->dtp_refcnt == 0) {
11000 dtrace_difo_release(pred->dtp_difo, vstate);
11001 kmem_free(pred, sizeof (dtrace_predicate_t));
11002 }
11003 }
11004
11005 /*
11006 * DTrace Action Description Functions
11007 */
11008 static dtrace_actdesc_t *
11009 dtrace_actdesc_create(dtrace_actkind_t kind, uint32_t ntuple,
11010 uint64_t uarg, uint64_t arg)
11011 {
11012 dtrace_actdesc_t *act;
11013
11014 #ifdef illumos
11015 ASSERT(!DTRACEACT_ISPRINTFLIKE(kind) || (arg != NULL &&
11016 arg >= KERNELBASE) || (arg == NULL && kind == DTRACEACT_PRINTA));
11017 #endif
11018
11019 act = kmem_zalloc(sizeof (dtrace_actdesc_t), KM_SLEEP);
11020 act->dtad_kind = kind;
11021 act->dtad_ntuple = ntuple;
11022 act->dtad_uarg = uarg;
11023 act->dtad_arg = arg;
11024 act->dtad_refcnt = 1;
11025
11026 return (act);
11027 }
11028
11029 static void
11030 dtrace_actdesc_hold(dtrace_actdesc_t *act)
11031 {
11032 ASSERT(act->dtad_refcnt >= 1);
11033 act->dtad_refcnt++;
11034 }
11035
11036 static void
11037 dtrace_actdesc_release(dtrace_actdesc_t *act, dtrace_vstate_t *vstate)
11038 {
11039 dtrace_actkind_t kind = act->dtad_kind;
11040 dtrace_difo_t *dp;
11041
11042 ASSERT(act->dtad_refcnt >= 1);
11043
11044 if (--act->dtad_refcnt != 0)
11045 return;
11046
11047 if ((dp = act->dtad_difo) != NULL)
11048 dtrace_difo_release(dp, vstate);
11049
11050 if (DTRACEACT_ISPRINTFLIKE(kind)) {
11051 char *str = (char *)(uintptr_t)act->dtad_arg;
11052
11053 #ifdef illumos
11054 ASSERT((str != NULL && (uintptr_t)str >= KERNELBASE) ||
11055 (str == NULL && act->dtad_kind == DTRACEACT_PRINTA));
11056 #endif
11057
11058 if (str != NULL)
11059 kmem_free(str, strlen(str) + 1);
11060 }
11061
11062 kmem_free(act, sizeof (dtrace_actdesc_t));
11063 }
11064
11065 /*
11066 * DTrace ECB Functions
11067 */
11068 static dtrace_ecb_t *
11069 dtrace_ecb_add(dtrace_state_t *state, dtrace_probe_t *probe)
11070 {
11071 dtrace_ecb_t *ecb;
11072 dtrace_epid_t epid;
11073
11074 ASSERT(MUTEX_HELD(&dtrace_lock));
11075
11076 ecb = kmem_zalloc(sizeof (dtrace_ecb_t), KM_SLEEP);
11077 ecb->dte_predicate = NULL;
11078 ecb->dte_probe = probe;
11079
11080 /*
11081 * The default size is the size of the default action: recording
11082 * the header.
11083 */
11084 ecb->dte_size = ecb->dte_needed = sizeof (dtrace_rechdr_t);
11085 ecb->dte_alignment = sizeof (dtrace_epid_t);
11086
11087 epid = state->dts_epid++;
11088
11089 if (epid - 1 >= state->dts_necbs) {
11090 dtrace_ecb_t **oecbs = state->dts_ecbs, **ecbs;
11091 int necbs = state->dts_necbs << 1;
11092
11093 ASSERT(epid == state->dts_necbs + 1);
11094
11095 if (necbs == 0) {
11096 ASSERT(oecbs == NULL);
11097 necbs = 1;
11098 }
11099
11100 ecbs = kmem_zalloc(necbs * sizeof (*ecbs), KM_SLEEP);
11101
11102 if (oecbs != NULL)
11103 bcopy(oecbs, ecbs, state->dts_necbs * sizeof (*ecbs));
11104
11105 dtrace_membar_producer();
11106 state->dts_ecbs = ecbs;
11107
11108 if (oecbs != NULL) {
11109 /*
11110 * If this state is active, we must dtrace_sync()
11111 * before we can free the old dts_ecbs array: we're
11112 * coming in hot, and there may be active ring
11113 * buffer processing (which indexes into the dts_ecbs
11114 * array) on another CPU.
11115 */
11116 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
11117 dtrace_sync();
11118
11119 kmem_free(oecbs, state->dts_necbs * sizeof (*ecbs));
11120 }
11121
11122 dtrace_membar_producer();
11123 state->dts_necbs = necbs;
11124 }
11125
11126 ecb->dte_state = state;
11127
11128 ASSERT(state->dts_ecbs[epid - 1] == NULL);
11129 dtrace_membar_producer();
11130 state->dts_ecbs[(ecb->dte_epid = epid) - 1] = ecb;
11131
11132 return (ecb);
11133 }
11134
11135 static void
11136 dtrace_ecb_enable(dtrace_ecb_t *ecb)
11137 {
11138 dtrace_probe_t *probe = ecb->dte_probe;
11139
11140 ASSERT(MUTEX_HELD(&cpu_lock));
11141 ASSERT(MUTEX_HELD(&dtrace_lock));
11142 ASSERT(ecb->dte_next == NULL);
11143
11144 if (probe == NULL) {
11145 /*
11146 * This is the NULL probe -- there's nothing to do.
11147 */
11148 return;
11149 }
11150
11151 if (probe->dtpr_ecb == NULL) {
11152 dtrace_provider_t *prov = probe->dtpr_provider;
11153
11154 /*
11155 * We're the first ECB on this probe.
11156 */
11157 probe->dtpr_ecb = probe->dtpr_ecb_last = ecb;
11158
11159 if (ecb->dte_predicate != NULL)
11160 probe->dtpr_predcache = ecb->dte_predicate->dtp_cacheid;
11161
11162 prov->dtpv_pops.dtps_enable(prov->dtpv_arg,
11163 probe->dtpr_id, probe->dtpr_arg);
11164 } else {
11165 /*
11166 * This probe is already active. Swing the last pointer to
11167 * point to the new ECB, and issue a dtrace_sync() to assure
11168 * that all CPUs have seen the change.
11169 */
11170 ASSERT(probe->dtpr_ecb_last != NULL);
11171 probe->dtpr_ecb_last->dte_next = ecb;
11172 probe->dtpr_ecb_last = ecb;
11173 probe->dtpr_predcache = 0;
11174
11175 dtrace_sync();
11176 }
11177 }
11178
11179 static int
11180 dtrace_ecb_resize(dtrace_ecb_t *ecb)
11181 {
11182 dtrace_action_t *act;
11183 uint32_t curneeded = UINT32_MAX;
11184 uint32_t aggbase = UINT32_MAX;
11185
11186 /*
11187 * If we record anything, we always record the dtrace_rechdr_t. (And
11188 * we always record it first.)
11189 */
11190 ecb->dte_size = sizeof (dtrace_rechdr_t);
11191 ecb->dte_alignment = sizeof (dtrace_epid_t);
11192
11193 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
11194 dtrace_recdesc_t *rec = &act->dta_rec;
11195 ASSERT(rec->dtrd_size > 0 || rec->dtrd_alignment == 1);
11196
11197 ecb->dte_alignment = MAX(ecb->dte_alignment,
11198 rec->dtrd_alignment);
11199
11200 if (DTRACEACT_ISAGG(act->dta_kind)) {
11201 dtrace_aggregation_t *agg = (dtrace_aggregation_t *)act;
11202
11203 ASSERT(rec->dtrd_size != 0);
11204 ASSERT(agg->dtag_first != NULL);
11205 ASSERT(act->dta_prev->dta_intuple);
11206 ASSERT(aggbase != UINT32_MAX);
11207 ASSERT(curneeded != UINT32_MAX);
11208
11209 agg->dtag_base = aggbase;
11210
11211 curneeded = P2ROUNDUP(curneeded, rec->dtrd_alignment);
11212 rec->dtrd_offset = curneeded;
11213 if (curneeded + rec->dtrd_size < curneeded)
11214 return (EINVAL);
11215 curneeded += rec->dtrd_size;
11216 ecb->dte_needed = MAX(ecb->dte_needed, curneeded);
11217
11218 aggbase = UINT32_MAX;
11219 curneeded = UINT32_MAX;
11220 } else if (act->dta_intuple) {
11221 if (curneeded == UINT32_MAX) {
11222 /*
11223 * This is the first record in a tuple. Align
11224 * curneeded to be at offset 4 in an 8-byte
11225 * aligned block.
11226 */
11227 ASSERT(act->dta_prev == NULL ||
11228 !act->dta_prev->dta_intuple);
11229 ASSERT3U(aggbase, ==, UINT32_MAX);
11230 curneeded = P2PHASEUP(ecb->dte_size,
11231 sizeof (uint64_t), sizeof (dtrace_aggid_t));
11232
11233 aggbase = curneeded - sizeof (dtrace_aggid_t);
11234 ASSERT(IS_P2ALIGNED(aggbase,
11235 sizeof (uint64_t)));
11236 }
11237 curneeded = P2ROUNDUP(curneeded, rec->dtrd_alignment);
11238 rec->dtrd_offset = curneeded;
11239 if (curneeded + rec->dtrd_size < curneeded)
11240 return (EINVAL);
11241 curneeded += rec->dtrd_size;
11242 } else {
11243 /* tuples must be followed by an aggregation */
11244 ASSERT(act->dta_prev == NULL ||
11245 !act->dta_prev->dta_intuple);
11246
11247 ecb->dte_size = P2ROUNDUP(ecb->dte_size,
11248 rec->dtrd_alignment);
11249 rec->dtrd_offset = ecb->dte_size;
11250 if (ecb->dte_size + rec->dtrd_size < ecb->dte_size)
11251 return (EINVAL);
11252 ecb->dte_size += rec->dtrd_size;
11253 ecb->dte_needed = MAX(ecb->dte_needed, ecb->dte_size);
11254 }
11255 }
11256
11257 if ((act = ecb->dte_action) != NULL &&
11258 !(act->dta_kind == DTRACEACT_SPECULATE && act->dta_next == NULL) &&
11259 ecb->dte_size == sizeof (dtrace_rechdr_t)) {
11260 /*
11261 * If the size is still sizeof (dtrace_rechdr_t), then all
11262 * actions store no data; set the size to 0.
11263 */
11264 ecb->dte_size = 0;
11265 }
11266
11267 ecb->dte_size = P2ROUNDUP(ecb->dte_size, sizeof (dtrace_epid_t));
11268 ecb->dte_needed = P2ROUNDUP(ecb->dte_needed, (sizeof (dtrace_epid_t)));
11269 ecb->dte_state->dts_needed = MAX(ecb->dte_state->dts_needed,
11270 ecb->dte_needed);
11271 return (0);
11272 }
11273
11274 static dtrace_action_t *
11275 dtrace_ecb_aggregation_create(dtrace_ecb_t *ecb, dtrace_actdesc_t *desc)
11276 {
11277 dtrace_aggregation_t *agg;
11278 size_t size = sizeof (uint64_t);
11279 int ntuple = desc->dtad_ntuple;
11280 dtrace_action_t *act;
11281 dtrace_recdesc_t *frec;
11282 dtrace_aggid_t aggid;
11283 dtrace_state_t *state = ecb->dte_state;
11284
11285 agg = kmem_zalloc(sizeof (dtrace_aggregation_t), KM_SLEEP);
11286 agg->dtag_ecb = ecb;
11287
11288 ASSERT(DTRACEACT_ISAGG(desc->dtad_kind));
11289
11290 switch (desc->dtad_kind) {
11291 case DTRACEAGG_MIN:
11292 agg->dtag_initial = INT64_MAX;
11293 agg->dtag_aggregate = dtrace_aggregate_min;
11294 break;
11295
11296 case DTRACEAGG_MAX:
11297 agg->dtag_initial = INT64_MIN;
11298 agg->dtag_aggregate = dtrace_aggregate_max;
11299 break;
11300
11301 case DTRACEAGG_COUNT:
11302 agg->dtag_aggregate = dtrace_aggregate_count;
11303 break;
11304
11305 case DTRACEAGG_QUANTIZE:
11306 agg->dtag_aggregate = dtrace_aggregate_quantize;
11307 size = (((sizeof (uint64_t) * NBBY) - 1) * 2 + 1) *
11308 sizeof (uint64_t);
11309 break;
11310
11311 case DTRACEAGG_LQUANTIZE: {
11312 uint16_t step = DTRACE_LQUANTIZE_STEP(desc->dtad_arg);
11313 uint16_t levels = DTRACE_LQUANTIZE_LEVELS(desc->dtad_arg);
11314
11315 agg->dtag_initial = desc->dtad_arg;
11316 agg->dtag_aggregate = dtrace_aggregate_lquantize;
11317
11318 if (step == 0 || levels == 0)
11319 goto err;
11320
11321 size = levels * sizeof (uint64_t) + 3 * sizeof (uint64_t);
11322 break;
11323 }
11324
11325 case DTRACEAGG_LLQUANTIZE: {
11326 uint16_t factor = DTRACE_LLQUANTIZE_FACTOR(desc->dtad_arg);
11327 uint16_t low = DTRACE_LLQUANTIZE_LOW(desc->dtad_arg);
11328 uint16_t high = DTRACE_LLQUANTIZE_HIGH(desc->dtad_arg);
11329 uint16_t nsteps = DTRACE_LLQUANTIZE_NSTEP(desc->dtad_arg);
11330 int64_t v;
11331
11332 agg->dtag_initial = desc->dtad_arg;
11333 agg->dtag_aggregate = dtrace_aggregate_llquantize;
11334
11335 if (factor < 2 || low >= high || nsteps < factor)
11336 goto err;
11337
11338 /*
11339 * Now check that the number of steps evenly divides a power
11340 * of the factor. (This assures both integer bucket size and
11341 * linearity within each magnitude.)
11342 */
11343 for (v = factor; v < nsteps; v *= factor)
11344 continue;
11345
11346 if ((v % nsteps) || (nsteps % factor))
11347 goto err;
11348
11349 size = (dtrace_aggregate_llquantize_bucket(factor,
11350 low, high, nsteps, INT64_MAX) + 2) * sizeof (uint64_t);
11351 break;
11352 }
11353
11354 case DTRACEAGG_AVG:
11355 agg->dtag_aggregate = dtrace_aggregate_avg;
11356 size = sizeof (uint64_t) * 2;
11357 break;
11358
11359 case DTRACEAGG_STDDEV:
11360 agg->dtag_aggregate = dtrace_aggregate_stddev;
11361 size = sizeof (uint64_t) * 4;
11362 break;
11363
11364 case DTRACEAGG_SUM:
11365 agg->dtag_aggregate = dtrace_aggregate_sum;
11366 break;
11367
11368 default:
11369 goto err;
11370 }
11371
11372 agg->dtag_action.dta_rec.dtrd_size = size;
11373
11374 if (ntuple == 0)
11375 goto err;
11376
11377 /*
11378 * We must make sure that we have enough actions for the n-tuple.
11379 */
11380 for (act = ecb->dte_action_last; act != NULL; act = act->dta_prev) {
11381 if (DTRACEACT_ISAGG(act->dta_kind))
11382 break;
11383
11384 if (--ntuple == 0) {
11385 /*
11386 * This is the action with which our n-tuple begins.
11387 */
11388 agg->dtag_first = act;
11389 goto success;
11390 }
11391 }
11392
11393 /*
11394 * This n-tuple is short by ntuple elements. Return failure.
11395 */
11396 ASSERT(ntuple != 0);
11397 err:
11398 kmem_free(agg, sizeof (dtrace_aggregation_t));
11399 return (NULL);
11400
11401 success:
11402 /*
11403 * If the last action in the tuple has a size of zero, it's actually
11404 * an expression argument for the aggregating action.
11405 */
11406 ASSERT(ecb->dte_action_last != NULL);
11407 act = ecb->dte_action_last;
11408
11409 if (act->dta_kind == DTRACEACT_DIFEXPR) {
11410 ASSERT(act->dta_difo != NULL);
11411
11412 if (act->dta_difo->dtdo_rtype.dtdt_size == 0)
11413 agg->dtag_hasarg = 1;
11414 }
11415
11416 /*
11417 * We need to allocate an id for this aggregation.
11418 */
11419 #ifdef illumos
11420 aggid = (dtrace_aggid_t)(uintptr_t)vmem_alloc(state->dts_aggid_arena, 1,
11421 VM_BESTFIT | VM_SLEEP);
11422 #else
11423 aggid = alloc_unr(state->dts_aggid_arena);
11424 #endif
11425
11426 if (aggid - 1 >= state->dts_naggregations) {
11427 dtrace_aggregation_t **oaggs = state->dts_aggregations;
11428 dtrace_aggregation_t **aggs;
11429 int naggs = state->dts_naggregations << 1;
11430 int onaggs = state->dts_naggregations;
11431
11432 ASSERT(aggid == state->dts_naggregations + 1);
11433
11434 if (naggs == 0) {
11435 ASSERT(oaggs == NULL);
11436 naggs = 1;
11437 }
11438
11439 aggs = kmem_zalloc(naggs * sizeof (*aggs), KM_SLEEP);
11440
11441 if (oaggs != NULL) {
11442 bcopy(oaggs, aggs, onaggs * sizeof (*aggs));
11443 kmem_free(oaggs, onaggs * sizeof (*aggs));
11444 }
11445
11446 state->dts_aggregations = aggs;
11447 state->dts_naggregations = naggs;
11448 }
11449
11450 ASSERT(state->dts_aggregations[aggid - 1] == NULL);
11451 state->dts_aggregations[(agg->dtag_id = aggid) - 1] = agg;
11452
11453 frec = &agg->dtag_first->dta_rec;
11454 if (frec->dtrd_alignment < sizeof (dtrace_aggid_t))
11455 frec->dtrd_alignment = sizeof (dtrace_aggid_t);
11456
11457 for (act = agg->dtag_first; act != NULL; act = act->dta_next) {
11458 ASSERT(!act->dta_intuple);
11459 act->dta_intuple = 1;
11460 }
11461
11462 return (&agg->dtag_action);
11463 }
11464
11465 static void
11466 dtrace_ecb_aggregation_destroy(dtrace_ecb_t *ecb, dtrace_action_t *act)
11467 {
11468 dtrace_aggregation_t *agg = (dtrace_aggregation_t *)act;
11469 dtrace_state_t *state = ecb->dte_state;
11470 dtrace_aggid_t aggid = agg->dtag_id;
11471
11472 ASSERT(DTRACEACT_ISAGG(act->dta_kind));
11473 #ifdef illumos
11474 vmem_free(state->dts_aggid_arena, (void *)(uintptr_t)aggid, 1);
11475 #else
11476 free_unr(state->dts_aggid_arena, aggid);
11477 #endif
11478
11479 ASSERT(state->dts_aggregations[aggid - 1] == agg);
11480 state->dts_aggregations[aggid - 1] = NULL;
11481
11482 kmem_free(agg, sizeof (dtrace_aggregation_t));
11483 }
11484
11485 static int
11486 dtrace_ecb_action_add(dtrace_ecb_t *ecb, dtrace_actdesc_t *desc)
11487 {
11488 dtrace_action_t *action, *last;
11489 dtrace_difo_t *dp = desc->dtad_difo;
11490 uint32_t size = 0, align = sizeof (uint8_t), mask;
11491 uint16_t format = 0;
11492 dtrace_recdesc_t *rec;
11493 dtrace_state_t *state = ecb->dte_state;
11494 dtrace_optval_t *opt = state->dts_options, nframes = 0, strsize;
11495 uint64_t arg = desc->dtad_arg;
11496
11497 ASSERT(MUTEX_HELD(&dtrace_lock));
11498 ASSERT(ecb->dte_action == NULL || ecb->dte_action->dta_refcnt == 1);
11499
11500 if (DTRACEACT_ISAGG(desc->dtad_kind)) {
11501 /*
11502 * If this is an aggregating action, there must be neither
11503 * a speculate nor a commit on the action chain.
11504 */
11505 dtrace_action_t *act;
11506
11507 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
11508 if (act->dta_kind == DTRACEACT_COMMIT)
11509 return (EINVAL);
11510
11511 if (act->dta_kind == DTRACEACT_SPECULATE)
11512 return (EINVAL);
11513 }
11514
11515 action = dtrace_ecb_aggregation_create(ecb, desc);
11516
11517 if (action == NULL)
11518 return (EINVAL);
11519 } else {
11520 if (DTRACEACT_ISDESTRUCTIVE(desc->dtad_kind) ||
11521 (desc->dtad_kind == DTRACEACT_DIFEXPR &&
11522 dp != NULL && dp->dtdo_destructive)) {
11523 state->dts_destructive = 1;
11524 }
11525
11526 switch (desc->dtad_kind) {
11527 case DTRACEACT_PRINTF:
11528 case DTRACEACT_PRINTA:
11529 case DTRACEACT_SYSTEM:
11530 case DTRACEACT_FREOPEN:
11531 case DTRACEACT_DIFEXPR:
11532 /*
11533 * We know that our arg is a string -- turn it into a
11534 * format.
11535 */
11536 if (arg == 0) {
11537 ASSERT(desc->dtad_kind == DTRACEACT_PRINTA ||
11538 desc->dtad_kind == DTRACEACT_DIFEXPR);
11539 format = 0;
11540 } else {
11541 ASSERT(arg != 0);
11542 #ifdef illumos
11543 ASSERT(arg > KERNELBASE);
11544 #endif
11545 format = dtrace_format_add(state,
11546 (char *)(uintptr_t)arg);
11547 }
11548
11549 /*FALLTHROUGH*/
11550 case DTRACEACT_LIBACT:
11551 case DTRACEACT_TRACEMEM:
11552 case DTRACEACT_TRACEMEM_DYNSIZE:
11553 if (dp == NULL)
11554 return (EINVAL);
11555
11556 if ((size = dp->dtdo_rtype.dtdt_size) != 0)
11557 break;
11558
11559 if (dp->dtdo_rtype.dtdt_kind == DIF_TYPE_STRING) {
11560 if (!(dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
11561 return (EINVAL);
11562
11563 size = opt[DTRACEOPT_STRSIZE];
11564 }
11565
11566 break;
11567
11568 case DTRACEACT_STACK:
11569 if ((nframes = arg) == 0) {
11570 nframes = opt[DTRACEOPT_STACKFRAMES];
11571 ASSERT(nframes > 0);
11572 arg = nframes;
11573 }
11574
11575 size = nframes * sizeof (pc_t);
11576 break;
11577
11578 case DTRACEACT_JSTACK:
11579 if ((strsize = DTRACE_USTACK_STRSIZE(arg)) == 0)
11580 strsize = opt[DTRACEOPT_JSTACKSTRSIZE];
11581
11582 if ((nframes = DTRACE_USTACK_NFRAMES(arg)) == 0)
11583 nframes = opt[DTRACEOPT_JSTACKFRAMES];
11584
11585 arg = DTRACE_USTACK_ARG(nframes, strsize);
11586
11587 /*FALLTHROUGH*/
11588 case DTRACEACT_USTACK:
11589 if (desc->dtad_kind != DTRACEACT_JSTACK &&
11590 (nframes = DTRACE_USTACK_NFRAMES(arg)) == 0) {
11591 strsize = DTRACE_USTACK_STRSIZE(arg);
11592 nframes = opt[DTRACEOPT_USTACKFRAMES];
11593 ASSERT(nframes > 0);
11594 arg = DTRACE_USTACK_ARG(nframes, strsize);
11595 }
11596
11597 /*
11598 * Save a slot for the pid.
11599 */
11600 size = (nframes + 1) * sizeof (uint64_t);
11601 size += DTRACE_USTACK_STRSIZE(arg);
11602 size = P2ROUNDUP(size, (uint32_t)(sizeof (uintptr_t)));
11603
11604 break;
11605
11606 case DTRACEACT_SYM:
11607 case DTRACEACT_MOD:
11608 if (dp == NULL || ((size = dp->dtdo_rtype.dtdt_size) !=
11609 sizeof (uint64_t)) ||
11610 (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
11611 return (EINVAL);
11612 break;
11613
11614 case DTRACEACT_USYM:
11615 case DTRACEACT_UMOD:
11616 case DTRACEACT_UADDR:
11617 if (dp == NULL ||
11618 (dp->dtdo_rtype.dtdt_size != sizeof (uint64_t)) ||
11619 (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
11620 return (EINVAL);
11621
11622 /*
11623 * We have a slot for the pid, plus a slot for the
11624 * argument. To keep things simple (aligned with
11625 * bitness-neutral sizing), we store each as a 64-bit
11626 * quantity.
11627 */
11628 size = 2 * sizeof (uint64_t);
11629 break;
11630
11631 case DTRACEACT_STOP:
11632 case DTRACEACT_BREAKPOINT:
11633 case DTRACEACT_PANIC:
11634 break;
11635
11636 case DTRACEACT_CHILL:
11637 case DTRACEACT_DISCARD:
11638 case DTRACEACT_RAISE:
11639 if (dp == NULL)
11640 return (EINVAL);
11641 break;
11642
11643 case DTRACEACT_EXIT:
11644 if (dp == NULL ||
11645 (size = dp->dtdo_rtype.dtdt_size) != sizeof (int) ||
11646 (dp->dtdo_rtype.dtdt_flags & DIF_TF_BYREF))
11647 return (EINVAL);
11648 break;
11649
11650 case DTRACEACT_SPECULATE:
11651 if (ecb->dte_size > sizeof (dtrace_rechdr_t))
11652 return (EINVAL);
11653
11654 if (dp == NULL)
11655 return (EINVAL);
11656
11657 state->dts_speculates = 1;
11658 break;
11659
11660 case DTRACEACT_PRINTM:
11661 size = dp->dtdo_rtype.dtdt_size;
11662 break;
11663
11664 case DTRACEACT_COMMIT: {
11665 dtrace_action_t *act = ecb->dte_action;
11666
11667 for (; act != NULL; act = act->dta_next) {
11668 if (act->dta_kind == DTRACEACT_COMMIT)
11669 return (EINVAL);
11670 }
11671
11672 if (dp == NULL)
11673 return (EINVAL);
11674 break;
11675 }
11676
11677 default:
11678 return (EINVAL);
11679 }
11680
11681 if (size != 0 || desc->dtad_kind == DTRACEACT_SPECULATE) {
11682 /*
11683 * If this is a data-storing action or a speculate,
11684 * we must be sure that there isn't a commit on the
11685 * action chain.
11686 */
11687 dtrace_action_t *act = ecb->dte_action;
11688
11689 for (; act != NULL; act = act->dta_next) {
11690 if (act->dta_kind == DTRACEACT_COMMIT)
11691 return (EINVAL);
11692 }
11693 }
11694
11695 action = kmem_zalloc(sizeof (dtrace_action_t), KM_SLEEP);
11696 action->dta_rec.dtrd_size = size;
11697 }
11698
11699 action->dta_refcnt = 1;
11700 rec = &action->dta_rec;
11701 size = rec->dtrd_size;
11702
11703 for (mask = sizeof (uint64_t) - 1; size != 0 && mask > 0; mask >>= 1) {
11704 if (!(size & mask)) {
11705 align = mask + 1;
11706 break;
11707 }
11708 }
11709
11710 action->dta_kind = desc->dtad_kind;
11711
11712 if ((action->dta_difo = dp) != NULL)
11713 dtrace_difo_hold(dp);
11714
11715 rec->dtrd_action = action->dta_kind;
11716 rec->dtrd_arg = arg;
11717 rec->dtrd_uarg = desc->dtad_uarg;
11718 rec->dtrd_alignment = (uint16_t)align;
11719 rec->dtrd_format = format;
11720
11721 if ((last = ecb->dte_action_last) != NULL) {
11722 ASSERT(ecb->dte_action != NULL);
11723 action->dta_prev = last;
11724 last->dta_next = action;
11725 } else {
11726 ASSERT(ecb->dte_action == NULL);
11727 ecb->dte_action = action;
11728 }
11729
11730 ecb->dte_action_last = action;
11731
11732 return (0);
11733 }
11734
11735 static void
11736 dtrace_ecb_action_remove(dtrace_ecb_t *ecb)
11737 {
11738 dtrace_action_t *act = ecb->dte_action, *next;
11739 dtrace_vstate_t *vstate = &ecb->dte_state->dts_vstate;
11740 dtrace_difo_t *dp;
11741 uint16_t format;
11742
11743 if (act != NULL && act->dta_refcnt > 1) {
11744 ASSERT(act->dta_next == NULL || act->dta_next->dta_refcnt == 1);
11745 act->dta_refcnt--;
11746 } else {
11747 for (; act != NULL; act = next) {
11748 next = act->dta_next;
11749 ASSERT(next != NULL || act == ecb->dte_action_last);
11750 ASSERT(act->dta_refcnt == 1);
11751
11752 if ((format = act->dta_rec.dtrd_format) != 0)
11753 dtrace_format_remove(ecb->dte_state, format);
11754
11755 if ((dp = act->dta_difo) != NULL)
11756 dtrace_difo_release(dp, vstate);
11757
11758 if (DTRACEACT_ISAGG(act->dta_kind)) {
11759 dtrace_ecb_aggregation_destroy(ecb, act);
11760 } else {
11761 kmem_free(act, sizeof (dtrace_action_t));
11762 }
11763 }
11764 }
11765
11766 ecb->dte_action = NULL;
11767 ecb->dte_action_last = NULL;
11768 ecb->dte_size = 0;
11769 }
11770
11771 static void
11772 dtrace_ecb_disable(dtrace_ecb_t *ecb)
11773 {
11774 /*
11775 * We disable the ECB by removing it from its probe.
11776 */
11777 dtrace_ecb_t *pecb, *prev = NULL;
11778 dtrace_probe_t *probe = ecb->dte_probe;
11779
11780 ASSERT(MUTEX_HELD(&dtrace_lock));
11781
11782 if (probe == NULL) {
11783 /*
11784 * This is the NULL probe; there is nothing to disable.
11785 */
11786 return;
11787 }
11788
11789 for (pecb = probe->dtpr_ecb; pecb != NULL; pecb = pecb->dte_next) {
11790 if (pecb == ecb)
11791 break;
11792 prev = pecb;
11793 }
11794
11795 ASSERT(pecb != NULL);
11796
11797 if (prev == NULL) {
11798 probe->dtpr_ecb = ecb->dte_next;
11799 } else {
11800 prev->dte_next = ecb->dte_next;
11801 }
11802
11803 if (ecb == probe->dtpr_ecb_last) {
11804 ASSERT(ecb->dte_next == NULL);
11805 probe->dtpr_ecb_last = prev;
11806 }
11807
11808 /*
11809 * The ECB has been disconnected from the probe; now sync to assure
11810 * that all CPUs have seen the change before returning.
11811 */
11812 dtrace_sync();
11813
11814 if (probe->dtpr_ecb == NULL) {
11815 /*
11816 * That was the last ECB on the probe; clear the predicate
11817 * cache ID for the probe, disable it and sync one more time
11818 * to assure that we'll never hit it again.
11819 */
11820 dtrace_provider_t *prov = probe->dtpr_provider;
11821
11822 ASSERT(ecb->dte_next == NULL);
11823 ASSERT(probe->dtpr_ecb_last == NULL);
11824 probe->dtpr_predcache = DTRACE_CACHEIDNONE;
11825 prov->dtpv_pops.dtps_disable(prov->dtpv_arg,
11826 probe->dtpr_id, probe->dtpr_arg);
11827 dtrace_sync();
11828 } else {
11829 /*
11830 * There is at least one ECB remaining on the probe. If there
11831 * is _exactly_ one, set the probe's predicate cache ID to be
11832 * the predicate cache ID of the remaining ECB.
11833 */
11834 ASSERT(probe->dtpr_ecb_last != NULL);
11835 ASSERT(probe->dtpr_predcache == DTRACE_CACHEIDNONE);
11836
11837 if (probe->dtpr_ecb == probe->dtpr_ecb_last) {
11838 dtrace_predicate_t *p = probe->dtpr_ecb->dte_predicate;
11839
11840 ASSERT(probe->dtpr_ecb->dte_next == NULL);
11841
11842 if (p != NULL)
11843 probe->dtpr_predcache = p->dtp_cacheid;
11844 }
11845
11846 ecb->dte_next = NULL;
11847 }
11848 }
11849
11850 static void
11851 dtrace_ecb_destroy(dtrace_ecb_t *ecb)
11852 {
11853 dtrace_state_t *state = ecb->dte_state;
11854 dtrace_vstate_t *vstate = &state->dts_vstate;
11855 dtrace_predicate_t *pred;
11856 dtrace_epid_t epid = ecb->dte_epid;
11857
11858 ASSERT(MUTEX_HELD(&dtrace_lock));
11859 ASSERT(ecb->dte_next == NULL);
11860 ASSERT(ecb->dte_probe == NULL || ecb->dte_probe->dtpr_ecb != ecb);
11861
11862 if ((pred = ecb->dte_predicate) != NULL)
11863 dtrace_predicate_release(pred, vstate);
11864
11865 dtrace_ecb_action_remove(ecb);
11866
11867 ASSERT(state->dts_ecbs[epid - 1] == ecb);
11868 state->dts_ecbs[epid - 1] = NULL;
11869
11870 kmem_free(ecb, sizeof (dtrace_ecb_t));
11871 }
11872
11873 static dtrace_ecb_t *
11874 dtrace_ecb_create(dtrace_state_t *state, dtrace_probe_t *probe,
11875 dtrace_enabling_t *enab)
11876 {
11877 dtrace_ecb_t *ecb;
11878 dtrace_predicate_t *pred;
11879 dtrace_actdesc_t *act;
11880 dtrace_provider_t *prov;
11881 dtrace_ecbdesc_t *desc = enab->dten_current;
11882
11883 ASSERT(MUTEX_HELD(&dtrace_lock));
11884 ASSERT(state != NULL);
11885
11886 ecb = dtrace_ecb_add(state, probe);
11887 ecb->dte_uarg = desc->dted_uarg;
11888
11889 if ((pred = desc->dted_pred.dtpdd_predicate) != NULL) {
11890 dtrace_predicate_hold(pred);
11891 ecb->dte_predicate = pred;
11892 }
11893
11894 if (probe != NULL) {
11895 /*
11896 * If the provider shows more leg than the consumer is old
11897 * enough to see, we need to enable the appropriate implicit
11898 * predicate bits to prevent the ecb from activating at
11899 * revealing times.
11900 *
11901 * Providers specifying DTRACE_PRIV_USER at register time
11902 * are stating that they need the /proc-style privilege
11903 * model to be enforced, and this is what DTRACE_COND_OWNER
11904 * and DTRACE_COND_ZONEOWNER will then do at probe time.
11905 */
11906 prov = probe->dtpr_provider;
11907 if (!(state->dts_cred.dcr_visible & DTRACE_CRV_ALLPROC) &&
11908 (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_USER))
11909 ecb->dte_cond |= DTRACE_COND_OWNER;
11910
11911 if (!(state->dts_cred.dcr_visible & DTRACE_CRV_ALLZONE) &&
11912 (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_USER))
11913 ecb->dte_cond |= DTRACE_COND_ZONEOWNER;
11914
11915 /*
11916 * If the provider shows us kernel innards and the user
11917 * is lacking sufficient privilege, enable the
11918 * DTRACE_COND_USERMODE implicit predicate.
11919 */
11920 if (!(state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL) &&
11921 (prov->dtpv_priv.dtpp_flags & DTRACE_PRIV_KERNEL))
11922 ecb->dte_cond |= DTRACE_COND_USERMODE;
11923 }
11924
11925 if (dtrace_ecb_create_cache != NULL) {
11926 /*
11927 * If we have a cached ecb, we'll use its action list instead
11928 * of creating our own (saving both time and space).
11929 */
11930 dtrace_ecb_t *cached = dtrace_ecb_create_cache;
11931 dtrace_action_t *act = cached->dte_action;
11932
11933 if (act != NULL) {
11934 ASSERT(act->dta_refcnt > 0);
11935 act->dta_refcnt++;
11936 ecb->dte_action = act;
11937 ecb->dte_action_last = cached->dte_action_last;
11938 ecb->dte_needed = cached->dte_needed;
11939 ecb->dte_size = cached->dte_size;
11940 ecb->dte_alignment = cached->dte_alignment;
11941 }
11942
11943 return (ecb);
11944 }
11945
11946 for (act = desc->dted_action; act != NULL; act = act->dtad_next) {
11947 if ((enab->dten_error = dtrace_ecb_action_add(ecb, act)) != 0) {
11948 dtrace_ecb_destroy(ecb);
11949 return (NULL);
11950 }
11951 }
11952
11953 if ((enab->dten_error = dtrace_ecb_resize(ecb)) != 0) {
11954 dtrace_ecb_destroy(ecb);
11955 return (NULL);
11956 }
11957
11958 return (dtrace_ecb_create_cache = ecb);
11959 }
11960
11961 static int
11962 dtrace_ecb_create_enable(dtrace_probe_t *probe, void *arg)
11963 {
11964 dtrace_ecb_t *ecb;
11965 dtrace_enabling_t *enab = arg;
11966 dtrace_state_t *state = enab->dten_vstate->dtvs_state;
11967
11968 ASSERT(state != NULL);
11969
11970 if (probe != NULL && probe->dtpr_gen < enab->dten_probegen) {
11971 /*
11972 * This probe was created in a generation for which this
11973 * enabling has previously created ECBs; we don't want to
11974 * enable it again, so just kick out.
11975 */
11976 return (DTRACE_MATCH_NEXT);
11977 }
11978
11979 if ((ecb = dtrace_ecb_create(state, probe, enab)) == NULL)
11980 return (DTRACE_MATCH_DONE);
11981
11982 dtrace_ecb_enable(ecb);
11983 return (DTRACE_MATCH_NEXT);
11984 }
11985
11986 static dtrace_ecb_t *
11987 dtrace_epid2ecb(dtrace_state_t *state, dtrace_epid_t id)
11988 {
11989 dtrace_ecb_t *ecb;
11990
11991 ASSERT(MUTEX_HELD(&dtrace_lock));
11992
11993 if (id == 0 || id > state->dts_necbs)
11994 return (NULL);
11995
11996 ASSERT(state->dts_necbs > 0 && state->dts_ecbs != NULL);
11997 ASSERT((ecb = state->dts_ecbs[id - 1]) == NULL || ecb->dte_epid == id);
11998
11999 return (state->dts_ecbs[id - 1]);
12000 }
12001
12002 static dtrace_aggregation_t *
12003 dtrace_aggid2agg(dtrace_state_t *state, dtrace_aggid_t id)
12004 {
12005 dtrace_aggregation_t *agg;
12006
12007 ASSERT(MUTEX_HELD(&dtrace_lock));
12008
12009 if (id == 0 || id > state->dts_naggregations)
12010 return (NULL);
12011
12012 ASSERT(state->dts_naggregations > 0 && state->dts_aggregations != NULL);
12013 ASSERT((agg = state->dts_aggregations[id - 1]) == NULL ||
12014 agg->dtag_id == id);
12015
12016 return (state->dts_aggregations[id - 1]);
12017 }
12018
12019 /*
12020 * DTrace Buffer Functions
12021 *
12022 * The following functions manipulate DTrace buffers. Most of these functions
12023 * are called in the context of establishing or processing consumer state;
12024 * exceptions are explicitly noted.
12025 */
12026
12027 /*
12028 * Note: called from cross call context. This function switches the two
12029 * buffers on a given CPU. The atomicity of this operation is assured by
12030 * disabling interrupts while the actual switch takes place; the disabling of
12031 * interrupts serializes the execution with any execution of dtrace_probe() on
12032 * the same CPU.
12033 */
12034 static void
12035 dtrace_buffer_switch(dtrace_buffer_t *buf)
12036 {
12037 caddr_t tomax = buf->dtb_tomax;
12038 caddr_t xamot = buf->dtb_xamot;
12039 dtrace_icookie_t cookie;
12040 hrtime_t now;
12041
12042 ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
12043 ASSERT(!(buf->dtb_flags & DTRACEBUF_RING));
12044
12045 cookie = dtrace_interrupt_disable();
12046 now = dtrace_gethrtime();
12047 buf->dtb_tomax = xamot;
12048 buf->dtb_xamot = tomax;
12049 buf->dtb_xamot_drops = buf->dtb_drops;
12050 buf->dtb_xamot_offset = buf->dtb_offset;
12051 buf->dtb_xamot_errors = buf->dtb_errors;
12052 buf->dtb_xamot_flags = buf->dtb_flags;
12053 buf->dtb_offset = 0;
12054 buf->dtb_drops = 0;
12055 buf->dtb_errors = 0;
12056 buf->dtb_flags &= ~(DTRACEBUF_ERROR | DTRACEBUF_DROPPED);
12057 buf->dtb_interval = now - buf->dtb_switched;
12058 buf->dtb_switched = now;
12059 dtrace_interrupt_enable(cookie);
12060 }
12061
12062 /*
12063 * Note: called from cross call context. This function activates a buffer
12064 * on a CPU. As with dtrace_buffer_switch(), the atomicity of the operation
12065 * is guaranteed by the disabling of interrupts.
12066 */
12067 static void
12068 dtrace_buffer_activate(dtrace_state_t *state)
12069 {
12070 dtrace_buffer_t *buf;
12071 dtrace_icookie_t cookie = dtrace_interrupt_disable();
12072
12073 buf = &state->dts_buffer[curcpu];
12074
12075 if (buf->dtb_tomax != NULL) {
12076 /*
12077 * We might like to assert that the buffer is marked inactive,
12078 * but this isn't necessarily true: the buffer for the CPU
12079 * that processes the BEGIN probe has its buffer activated
12080 * manually. In this case, we take the (harmless) action
12081 * re-clearing the bit INACTIVE bit.
12082 */
12083 buf->dtb_flags &= ~DTRACEBUF_INACTIVE;
12084 }
12085
12086 dtrace_interrupt_enable(cookie);
12087 }
12088
12089 #ifdef __FreeBSD__
12090 /*
12091 * Activate the specified per-CPU buffer. This is used instead of
12092 * dtrace_buffer_activate() when APs have not yet started, i.e. when
12093 * activating anonymous state.
12094 */
12095 static void
12096 dtrace_buffer_activate_cpu(dtrace_state_t *state, int cpu)
12097 {
12098
12099 if (state->dts_buffer[cpu].dtb_tomax != NULL)
12100 state->dts_buffer[cpu].dtb_flags &= ~DTRACEBUF_INACTIVE;
12101 }
12102 #endif
12103
12104 static int
12105 dtrace_buffer_alloc(dtrace_buffer_t *bufs, size_t size, int flags,
12106 processorid_t cpu, int *factor)
12107 {
12108 #ifdef illumos
12109 cpu_t *cp;
12110 #endif
12111 dtrace_buffer_t *buf;
12112 int allocated = 0, desired = 0;
12113
12114 #ifdef illumos
12115 ASSERT(MUTEX_HELD(&cpu_lock));
12116 ASSERT(MUTEX_HELD(&dtrace_lock));
12117
12118 *factor = 1;
12119
12120 if (size > dtrace_nonroot_maxsize &&
12121 !PRIV_POLICY_CHOICE(CRED(), PRIV_ALL, B_FALSE))
12122 return (EFBIG);
12123
12124 cp = cpu_list;
12125
12126 do {
12127 if (cpu != DTRACE_CPUALL && cpu != cp->cpu_id)
12128 continue;
12129
12130 buf = &bufs[cp->cpu_id];
12131
12132 /*
12133 * If there is already a buffer allocated for this CPU, it
12134 * is only possible that this is a DR event. In this case,
12135 */
12136 if (buf->dtb_tomax != NULL) {
12137 ASSERT(buf->dtb_size == size);
12138 continue;
12139 }
12140
12141 ASSERT(buf->dtb_xamot == NULL);
12142
12143 if ((buf->dtb_tomax = kmem_zalloc(size,
12144 KM_NOSLEEP | KM_NORMALPRI)) == NULL)
12145 goto err;
12146
12147 buf->dtb_size = size;
12148 buf->dtb_flags = flags;
12149 buf->dtb_offset = 0;
12150 buf->dtb_drops = 0;
12151
12152 if (flags & DTRACEBUF_NOSWITCH)
12153 continue;
12154
12155 if ((buf->dtb_xamot = kmem_zalloc(size,
12156 KM_NOSLEEP | KM_NORMALPRI)) == NULL)
12157 goto err;
12158 } while ((cp = cp->cpu_next) != cpu_list);
12159
12160 return (0);
12161
12162 err:
12163 cp = cpu_list;
12164
12165 do {
12166 if (cpu != DTRACE_CPUALL && cpu != cp->cpu_id)
12167 continue;
12168
12169 buf = &bufs[cp->cpu_id];
12170 desired += 2;
12171
12172 if (buf->dtb_xamot != NULL) {
12173 ASSERT(buf->dtb_tomax != NULL);
12174 ASSERT(buf->dtb_size == size);
12175 kmem_free(buf->dtb_xamot, size);
12176 allocated++;
12177 }
12178
12179 if (buf->dtb_tomax != NULL) {
12180 ASSERT(buf->dtb_size == size);
12181 kmem_free(buf->dtb_tomax, size);
12182 allocated++;
12183 }
12184
12185 buf->dtb_tomax = NULL;
12186 buf->dtb_xamot = NULL;
12187 buf->dtb_size = 0;
12188 } while ((cp = cp->cpu_next) != cpu_list);
12189 #else
12190 int i;
12191
12192 *factor = 1;
12193 #if defined(__aarch64__) || defined(__amd64__) || defined(__arm__) || \
12194 defined(__mips__) || defined(__powerpc__) || defined(__riscv)
12195 /*
12196 * FreeBSD isn't good at limiting the amount of memory we
12197 * ask to malloc, so let's place a limit here before trying
12198 * to do something that might well end in tears at bedtime.
12199 */
12200 int bufsize_percpu_frac = dtrace_bufsize_max_frac * mp_ncpus;
12201 if (size > physmem * PAGE_SIZE / bufsize_percpu_frac)
12202 return (ENOMEM);
12203 #endif
12204
12205 ASSERT(MUTEX_HELD(&dtrace_lock));
12206 CPU_FOREACH(i) {
12207 if (cpu != DTRACE_CPUALL && cpu != i)
12208 continue;
12209
12210 buf = &bufs[i];
12211
12212 /*
12213 * If there is already a buffer allocated for this CPU, it
12214 * is only possible that this is a DR event. In this case,
12215 * the buffer size must match our specified size.
12216 */
12217 if (buf->dtb_tomax != NULL) {
12218 ASSERT(buf->dtb_size == size);
12219 continue;
12220 }
12221
12222 ASSERT(buf->dtb_xamot == NULL);
12223
12224 if ((buf->dtb_tomax = kmem_zalloc(size,
12225 KM_NOSLEEP | KM_NORMALPRI)) == NULL)
12226 goto err;
12227
12228 buf->dtb_size = size;
12229 buf->dtb_flags = flags;
12230 buf->dtb_offset = 0;
12231 buf->dtb_drops = 0;
12232
12233 if (flags & DTRACEBUF_NOSWITCH)
12234 continue;
12235
12236 if ((buf->dtb_xamot = kmem_zalloc(size,
12237 KM_NOSLEEP | KM_NORMALPRI)) == NULL)
12238 goto err;
12239 }
12240
12241 return (0);
12242
12243 err:
12244 /*
12245 * Error allocating memory, so free the buffers that were
12246 * allocated before the failed allocation.
12247 */
12248 CPU_FOREACH(i) {
12249 if (cpu != DTRACE_CPUALL && cpu != i)
12250 continue;
12251
12252 buf = &bufs[i];
12253 desired += 2;
12254
12255 if (buf->dtb_xamot != NULL) {
12256 ASSERT(buf->dtb_tomax != NULL);
12257 ASSERT(buf->dtb_size == size);
12258 kmem_free(buf->dtb_xamot, size);
12259 allocated++;
12260 }
12261
12262 if (buf->dtb_tomax != NULL) {
12263 ASSERT(buf->dtb_size == size);
12264 kmem_free(buf->dtb_tomax, size);
12265 allocated++;
12266 }
12267
12268 buf->dtb_tomax = NULL;
12269 buf->dtb_xamot = NULL;
12270 buf->dtb_size = 0;
12271
12272 }
12273 #endif
12274 *factor = desired / (allocated > 0 ? allocated : 1);
12275
12276 return (ENOMEM);
12277 }
12278
12279 /*
12280 * Note: called from probe context. This function just increments the drop
12281 * count on a buffer. It has been made a function to allow for the
12282 * possibility of understanding the source of mysterious drop counts. (A
12283 * problem for which one may be particularly disappointed that DTrace cannot
12284 * be used to understand DTrace.)
12285 */
12286 static void
12287 dtrace_buffer_drop(dtrace_buffer_t *buf)
12288 {
12289 buf->dtb_drops++;
12290 }
12291
12292 /*
12293 * Note: called from probe context. This function is called to reserve space
12294 * in a buffer. If mstate is non-NULL, sets the scratch base and size in the
12295 * mstate. Returns the new offset in the buffer, or a negative value if an
12296 * error has occurred.
12297 */
12298 static intptr_t
12299 dtrace_buffer_reserve(dtrace_buffer_t *buf, size_t needed, size_t align,
12300 dtrace_state_t *state, dtrace_mstate_t *mstate)
12301 {
12302 intptr_t offs = buf->dtb_offset, soffs;
12303 intptr_t woffs;
12304 caddr_t tomax;
12305 size_t total;
12306
12307 if (buf->dtb_flags & DTRACEBUF_INACTIVE)
12308 return (-1);
12309
12310 if ((tomax = buf->dtb_tomax) == NULL) {
12311 dtrace_buffer_drop(buf);
12312 return (-1);
12313 }
12314
12315 if (!(buf->dtb_flags & (DTRACEBUF_RING | DTRACEBUF_FILL))) {
12316 while (offs & (align - 1)) {
12317 /*
12318 * Assert that our alignment is off by a number which
12319 * is itself sizeof (uint32_t) aligned.
12320 */
12321 ASSERT(!((align - (offs & (align - 1))) &
12322 (sizeof (uint32_t) - 1)));
12323 DTRACE_STORE(uint32_t, tomax, offs, DTRACE_EPIDNONE);
12324 offs += sizeof (uint32_t);
12325 }
12326
12327 if ((soffs = offs + needed) > buf->dtb_size) {
12328 dtrace_buffer_drop(buf);
12329 return (-1);
12330 }
12331
12332 if (mstate == NULL)
12333 return (offs);
12334
12335 mstate->dtms_scratch_base = (uintptr_t)tomax + soffs;
12336 mstate->dtms_scratch_size = buf->dtb_size - soffs;
12337 mstate->dtms_scratch_ptr = mstate->dtms_scratch_base;
12338
12339 return (offs);
12340 }
12341
12342 if (buf->dtb_flags & DTRACEBUF_FILL) {
12343 if (state->dts_activity != DTRACE_ACTIVITY_COOLDOWN &&
12344 (buf->dtb_flags & DTRACEBUF_FULL))
12345 return (-1);
12346 goto out;
12347 }
12348
12349 total = needed + (offs & (align - 1));
12350
12351 /*
12352 * For a ring buffer, life is quite a bit more complicated. Before
12353 * we can store any padding, we need to adjust our wrapping offset.
12354 * (If we've never before wrapped or we're not about to, no adjustment
12355 * is required.)
12356 */
12357 if ((buf->dtb_flags & DTRACEBUF_WRAPPED) ||
12358 offs + total > buf->dtb_size) {
12359 woffs = buf->dtb_xamot_offset;
12360
12361 if (offs + total > buf->dtb_size) {
12362 /*
12363 * We can't fit in the end of the buffer. First, a
12364 * sanity check that we can fit in the buffer at all.
12365 */
12366 if (total > buf->dtb_size) {
12367 dtrace_buffer_drop(buf);
12368 return (-1);
12369 }
12370
12371 /*
12372 * We're going to be storing at the top of the buffer,
12373 * so now we need to deal with the wrapped offset. We
12374 * only reset our wrapped offset to 0 if it is
12375 * currently greater than the current offset. If it
12376 * is less than the current offset, it is because a
12377 * previous allocation induced a wrap -- but the
12378 * allocation didn't subsequently take the space due
12379 * to an error or false predicate evaluation. In this
12380 * case, we'll just leave the wrapped offset alone: if
12381 * the wrapped offset hasn't been advanced far enough
12382 * for this allocation, it will be adjusted in the
12383 * lower loop.
12384 */
12385 if (buf->dtb_flags & DTRACEBUF_WRAPPED) {
12386 if (woffs >= offs)
12387 woffs = 0;
12388 } else {
12389 woffs = 0;
12390 }
12391
12392 /*
12393 * Now we know that we're going to be storing to the
12394 * top of the buffer and that there is room for us
12395 * there. We need to clear the buffer from the current
12396 * offset to the end (there may be old gunk there).
12397 */
12398 while (offs < buf->dtb_size)
12399 tomax[offs++] = 0;
12400
12401 /*
12402 * We need to set our offset to zero. And because we
12403 * are wrapping, we need to set the bit indicating as
12404 * much. We can also adjust our needed space back
12405 * down to the space required by the ECB -- we know
12406 * that the top of the buffer is aligned.
12407 */
12408 offs = 0;
12409 total = needed;
12410 buf->dtb_flags |= DTRACEBUF_WRAPPED;
12411 } else {
12412 /*
12413 * There is room for us in the buffer, so we simply
12414 * need to check the wrapped offset.
12415 */
12416 if (woffs < offs) {
12417 /*
12418 * The wrapped offset is less than the offset.
12419 * This can happen if we allocated buffer space
12420 * that induced a wrap, but then we didn't
12421 * subsequently take the space due to an error
12422 * or false predicate evaluation. This is
12423 * okay; we know that _this_ allocation isn't
12424 * going to induce a wrap. We still can't
12425 * reset the wrapped offset to be zero,
12426 * however: the space may have been trashed in
12427 * the previous failed probe attempt. But at
12428 * least the wrapped offset doesn't need to
12429 * be adjusted at all...
12430 */
12431 goto out;
12432 }
12433 }
12434
12435 while (offs + total > woffs) {
12436 dtrace_epid_t epid = *(uint32_t *)(tomax + woffs);
12437 size_t size;
12438
12439 if (epid == DTRACE_EPIDNONE) {
12440 size = sizeof (uint32_t);
12441 } else {
12442 ASSERT3U(epid, <=, state->dts_necbs);
12443 ASSERT(state->dts_ecbs[epid - 1] != NULL);
12444
12445 size = state->dts_ecbs[epid - 1]->dte_size;
12446 }
12447
12448 ASSERT(woffs + size <= buf->dtb_size);
12449 ASSERT(size != 0);
12450
12451 if (woffs + size == buf->dtb_size) {
12452 /*
12453 * We've reached the end of the buffer; we want
12454 * to set the wrapped offset to 0 and break
12455 * out. However, if the offs is 0, then we're
12456 * in a strange edge-condition: the amount of
12457 * space that we want to reserve plus the size
12458 * of the record that we're overwriting is
12459 * greater than the size of the buffer. This
12460 * is problematic because if we reserve the
12461 * space but subsequently don't consume it (due
12462 * to a failed predicate or error) the wrapped
12463 * offset will be 0 -- yet the EPID at offset 0
12464 * will not be committed. This situation is
12465 * relatively easy to deal with: if we're in
12466 * this case, the buffer is indistinguishable
12467 * from one that hasn't wrapped; we need only
12468 * finish the job by clearing the wrapped bit,
12469 * explicitly setting the offset to be 0, and
12470 * zero'ing out the old data in the buffer.
12471 */
12472 if (offs == 0) {
12473 buf->dtb_flags &= ~DTRACEBUF_WRAPPED;
12474 buf->dtb_offset = 0;
12475 woffs = total;
12476
12477 while (woffs < buf->dtb_size)
12478 tomax[woffs++] = 0;
12479 }
12480
12481 woffs = 0;
12482 break;
12483 }
12484
12485 woffs += size;
12486 }
12487
12488 /*
12489 * We have a wrapped offset. It may be that the wrapped offset
12490 * has become zero -- that's okay.
12491 */
12492 buf->dtb_xamot_offset = woffs;
12493 }
12494
12495 out:
12496 /*
12497 * Now we can plow the buffer with any necessary padding.
12498 */
12499 while (offs & (align - 1)) {
12500 /*
12501 * Assert that our alignment is off by a number which
12502 * is itself sizeof (uint32_t) aligned.
12503 */
12504 ASSERT(!((align - (offs & (align - 1))) &
12505 (sizeof (uint32_t) - 1)));
12506 DTRACE_STORE(uint32_t, tomax, offs, DTRACE_EPIDNONE);
12507 offs += sizeof (uint32_t);
12508 }
12509
12510 if (buf->dtb_flags & DTRACEBUF_FILL) {
12511 if (offs + needed > buf->dtb_size - state->dts_reserve) {
12512 buf->dtb_flags |= DTRACEBUF_FULL;
12513 return (-1);
12514 }
12515 }
12516
12517 if (mstate == NULL)
12518 return (offs);
12519
12520 /*
12521 * For ring buffers and fill buffers, the scratch space is always
12522 * the inactive buffer.
12523 */
12524 mstate->dtms_scratch_base = (uintptr_t)buf->dtb_xamot;
12525 mstate->dtms_scratch_size = buf->dtb_size;
12526 mstate->dtms_scratch_ptr = mstate->dtms_scratch_base;
12527
12528 return (offs);
12529 }
12530
12531 static void
12532 dtrace_buffer_polish(dtrace_buffer_t *buf)
12533 {
12534 ASSERT(buf->dtb_flags & DTRACEBUF_RING);
12535 ASSERT(MUTEX_HELD(&dtrace_lock));
12536
12537 if (!(buf->dtb_flags & DTRACEBUF_WRAPPED))
12538 return;
12539
12540 /*
12541 * We need to polish the ring buffer. There are three cases:
12542 *
12543 * - The first (and presumably most common) is that there is no gap
12544 * between the buffer offset and the wrapped offset. In this case,
12545 * there is nothing in the buffer that isn't valid data; we can
12546 * mark the buffer as polished and return.
12547 *
12548 * - The second (less common than the first but still more common
12549 * than the third) is that there is a gap between the buffer offset
12550 * and the wrapped offset, and the wrapped offset is larger than the
12551 * buffer offset. This can happen because of an alignment issue, or
12552 * can happen because of a call to dtrace_buffer_reserve() that
12553 * didn't subsequently consume the buffer space. In this case,
12554 * we need to zero the data from the buffer offset to the wrapped
12555 * offset.
12556 *
12557 * - The third (and least common) is that there is a gap between the
12558 * buffer offset and the wrapped offset, but the wrapped offset is
12559 * _less_ than the buffer offset. This can only happen because a
12560 * call to dtrace_buffer_reserve() induced a wrap, but the space
12561 * was not subsequently consumed. In this case, we need to zero the
12562 * space from the offset to the end of the buffer _and_ from the
12563 * top of the buffer to the wrapped offset.
12564 */
12565 if (buf->dtb_offset < buf->dtb_xamot_offset) {
12566 bzero(buf->dtb_tomax + buf->dtb_offset,
12567 buf->dtb_xamot_offset - buf->dtb_offset);
12568 }
12569
12570 if (buf->dtb_offset > buf->dtb_xamot_offset) {
12571 bzero(buf->dtb_tomax + buf->dtb_offset,
12572 buf->dtb_size - buf->dtb_offset);
12573 bzero(buf->dtb_tomax, buf->dtb_xamot_offset);
12574 }
12575 }
12576
12577 /*
12578 * This routine determines if data generated at the specified time has likely
12579 * been entirely consumed at user-level. This routine is called to determine
12580 * if an ECB on a defunct probe (but for an active enabling) can be safely
12581 * disabled and destroyed.
12582 */
12583 static int
12584 dtrace_buffer_consumed(dtrace_buffer_t *bufs, hrtime_t when)
12585 {
12586 int i;
12587
12588 for (i = 0; i < NCPU; i++) {
12589 dtrace_buffer_t *buf = &bufs[i];
12590
12591 if (buf->dtb_size == 0)
12592 continue;
12593
12594 if (buf->dtb_flags & DTRACEBUF_RING)
12595 return (0);
12596
12597 if (!buf->dtb_switched && buf->dtb_offset != 0)
12598 return (0);
12599
12600 if (buf->dtb_switched - buf->dtb_interval < when)
12601 return (0);
12602 }
12603
12604 return (1);
12605 }
12606
12607 static void
12608 dtrace_buffer_free(dtrace_buffer_t *bufs)
12609 {
12610 int i;
12611
12612 for (i = 0; i < NCPU; i++) {
12613 dtrace_buffer_t *buf = &bufs[i];
12614
12615 if (buf->dtb_tomax == NULL) {
12616 ASSERT(buf->dtb_xamot == NULL);
12617 ASSERT(buf->dtb_size == 0);
12618 continue;
12619 }
12620
12621 if (buf->dtb_xamot != NULL) {
12622 ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
12623 kmem_free(buf->dtb_xamot, buf->dtb_size);
12624 }
12625
12626 kmem_free(buf->dtb_tomax, buf->dtb_size);
12627 buf->dtb_size = 0;
12628 buf->dtb_tomax = NULL;
12629 buf->dtb_xamot = NULL;
12630 }
12631 }
12632
12633 /*
12634 * DTrace Enabling Functions
12635 */
12636 static dtrace_enabling_t *
12637 dtrace_enabling_create(dtrace_vstate_t *vstate)
12638 {
12639 dtrace_enabling_t *enab;
12640
12641 enab = kmem_zalloc(sizeof (dtrace_enabling_t), KM_SLEEP);
12642 enab->dten_vstate = vstate;
12643
12644 return (enab);
12645 }
12646
12647 static void
12648 dtrace_enabling_add(dtrace_enabling_t *enab, dtrace_ecbdesc_t *ecb)
12649 {
12650 dtrace_ecbdesc_t **ndesc;
12651 size_t osize, nsize;
12652
12653 /*
12654 * We can't add to enablings after we've enabled them, or after we've
12655 * retained them.
12656 */
12657 ASSERT(enab->dten_probegen == 0);
12658 ASSERT(enab->dten_next == NULL && enab->dten_prev == NULL);
12659
12660 if (enab->dten_ndesc < enab->dten_maxdesc) {
12661 enab->dten_desc[enab->dten_ndesc++] = ecb;
12662 return;
12663 }
12664
12665 osize = enab->dten_maxdesc * sizeof (dtrace_enabling_t *);
12666
12667 if (enab->dten_maxdesc == 0) {
12668 enab->dten_maxdesc = 1;
12669 } else {
12670 enab->dten_maxdesc <<= 1;
12671 }
12672
12673 ASSERT(enab->dten_ndesc < enab->dten_maxdesc);
12674
12675 nsize = enab->dten_maxdesc * sizeof (dtrace_enabling_t *);
12676 ndesc = kmem_zalloc(nsize, KM_SLEEP);
12677 bcopy(enab->dten_desc, ndesc, osize);
12678 if (enab->dten_desc != NULL)
12679 kmem_free(enab->dten_desc, osize);
12680
12681 enab->dten_desc = ndesc;
12682 enab->dten_desc[enab->dten_ndesc++] = ecb;
12683 }
12684
12685 static void
12686 dtrace_enabling_addlike(dtrace_enabling_t *enab, dtrace_ecbdesc_t *ecb,
12687 dtrace_probedesc_t *pd)
12688 {
12689 dtrace_ecbdesc_t *new;
12690 dtrace_predicate_t *pred;
12691 dtrace_actdesc_t *act;
12692
12693 /*
12694 * We're going to create a new ECB description that matches the
12695 * specified ECB in every way, but has the specified probe description.
12696 */
12697 new = kmem_zalloc(sizeof (dtrace_ecbdesc_t), KM_SLEEP);
12698
12699 if ((pred = ecb->dted_pred.dtpdd_predicate) != NULL)
12700 dtrace_predicate_hold(pred);
12701
12702 for (act = ecb->dted_action; act != NULL; act = act->dtad_next)
12703 dtrace_actdesc_hold(act);
12704
12705 new->dted_action = ecb->dted_action;
12706 new->dted_pred = ecb->dted_pred;
12707 new->dted_probe = *pd;
12708 new->dted_uarg = ecb->dted_uarg;
12709
12710 dtrace_enabling_add(enab, new);
12711 }
12712
12713 static void
12714 dtrace_enabling_dump(dtrace_enabling_t *enab)
12715 {
12716 int i;
12717
12718 for (i = 0; i < enab->dten_ndesc; i++) {
12719 dtrace_probedesc_t *desc = &enab->dten_desc[i]->dted_probe;
12720
12721 #ifdef __FreeBSD__
12722 printf("dtrace: enabling probe %d (%s:%s:%s:%s)\n", i,
12723 desc->dtpd_provider, desc->dtpd_mod,
12724 desc->dtpd_func, desc->dtpd_name);
12725 #else
12726 cmn_err(CE_NOTE, "enabling probe %d (%s:%s:%s:%s)", i,
12727 desc->dtpd_provider, desc->dtpd_mod,
12728 desc->dtpd_func, desc->dtpd_name);
12729 #endif
12730 }
12731 }
12732
12733 static void
12734 dtrace_enabling_destroy(dtrace_enabling_t *enab)
12735 {
12736 int i;
12737 dtrace_ecbdesc_t *ep;
12738 dtrace_vstate_t *vstate = enab->dten_vstate;
12739
12740 ASSERT(MUTEX_HELD(&dtrace_lock));
12741
12742 for (i = 0; i < enab->dten_ndesc; i++) {
12743 dtrace_actdesc_t *act, *next;
12744 dtrace_predicate_t *pred;
12745
12746 ep = enab->dten_desc[i];
12747
12748 if ((pred = ep->dted_pred.dtpdd_predicate) != NULL)
12749 dtrace_predicate_release(pred, vstate);
12750
12751 for (act = ep->dted_action; act != NULL; act = next) {
12752 next = act->dtad_next;
12753 dtrace_actdesc_release(act, vstate);
12754 }
12755
12756 kmem_free(ep, sizeof (dtrace_ecbdesc_t));
12757 }
12758
12759 if (enab->dten_desc != NULL)
12760 kmem_free(enab->dten_desc,
12761 enab->dten_maxdesc * sizeof (dtrace_enabling_t *));
12762
12763 /*
12764 * If this was a retained enabling, decrement the dts_nretained count
12765 * and take it off of the dtrace_retained list.
12766 */
12767 if (enab->dten_prev != NULL || enab->dten_next != NULL ||
12768 dtrace_retained == enab) {
12769 ASSERT(enab->dten_vstate->dtvs_state != NULL);
12770 ASSERT(enab->dten_vstate->dtvs_state->dts_nretained > 0);
12771 enab->dten_vstate->dtvs_state->dts_nretained--;
12772 dtrace_retained_gen++;
12773 }
12774
12775 if (enab->dten_prev == NULL) {
12776 if (dtrace_retained == enab) {
12777 dtrace_retained = enab->dten_next;
12778
12779 if (dtrace_retained != NULL)
12780 dtrace_retained->dten_prev = NULL;
12781 }
12782 } else {
12783 ASSERT(enab != dtrace_retained);
12784 ASSERT(dtrace_retained != NULL);
12785 enab->dten_prev->dten_next = enab->dten_next;
12786 }
12787
12788 if (enab->dten_next != NULL) {
12789 ASSERT(dtrace_retained != NULL);
12790 enab->dten_next->dten_prev = enab->dten_prev;
12791 }
12792
12793 kmem_free(enab, sizeof (dtrace_enabling_t));
12794 }
12795
12796 static int
12797 dtrace_enabling_retain(dtrace_enabling_t *enab)
12798 {
12799 dtrace_state_t *state;
12800
12801 ASSERT(MUTEX_HELD(&dtrace_lock));
12802 ASSERT(enab->dten_next == NULL && enab->dten_prev == NULL);
12803 ASSERT(enab->dten_vstate != NULL);
12804
12805 state = enab->dten_vstate->dtvs_state;
12806 ASSERT(state != NULL);
12807
12808 /*
12809 * We only allow each state to retain dtrace_retain_max enablings.
12810 */
12811 if (state->dts_nretained >= dtrace_retain_max)
12812 return (ENOSPC);
12813
12814 state->dts_nretained++;
12815 dtrace_retained_gen++;
12816
12817 if (dtrace_retained == NULL) {
12818 dtrace_retained = enab;
12819 return (0);
12820 }
12821
12822 enab->dten_next = dtrace_retained;
12823 dtrace_retained->dten_prev = enab;
12824 dtrace_retained = enab;
12825
12826 return (0);
12827 }
12828
12829 static int
12830 dtrace_enabling_replicate(dtrace_state_t *state, dtrace_probedesc_t *match,
12831 dtrace_probedesc_t *create)
12832 {
12833 dtrace_enabling_t *new, *enab;
12834 int found = 0, err = ENOENT;
12835
12836 ASSERT(MUTEX_HELD(&dtrace_lock));
12837 ASSERT(strlen(match->dtpd_provider) < DTRACE_PROVNAMELEN);
12838 ASSERT(strlen(match->dtpd_mod) < DTRACE_MODNAMELEN);
12839 ASSERT(strlen(match->dtpd_func) < DTRACE_FUNCNAMELEN);
12840 ASSERT(strlen(match->dtpd_name) < DTRACE_NAMELEN);
12841
12842 new = dtrace_enabling_create(&state->dts_vstate);
12843
12844 /*
12845 * Iterate over all retained enablings, looking for enablings that
12846 * match the specified state.
12847 */
12848 for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
12849 int i;
12850
12851 /*
12852 * dtvs_state can only be NULL for helper enablings -- and
12853 * helper enablings can't be retained.
12854 */
12855 ASSERT(enab->dten_vstate->dtvs_state != NULL);
12856
12857 if (enab->dten_vstate->dtvs_state != state)
12858 continue;
12859
12860 /*
12861 * Now iterate over each probe description; we're looking for
12862 * an exact match to the specified probe description.
12863 */
12864 for (i = 0; i < enab->dten_ndesc; i++) {
12865 dtrace_ecbdesc_t *ep = enab->dten_desc[i];
12866 dtrace_probedesc_t *pd = &ep->dted_probe;
12867
12868 if (strcmp(pd->dtpd_provider, match->dtpd_provider))
12869 continue;
12870
12871 if (strcmp(pd->dtpd_mod, match->dtpd_mod))
12872 continue;
12873
12874 if (strcmp(pd->dtpd_func, match->dtpd_func))
12875 continue;
12876
12877 if (strcmp(pd->dtpd_name, match->dtpd_name))
12878 continue;
12879
12880 /*
12881 * We have a winning probe! Add it to our growing
12882 * enabling.
12883 */
12884 found = 1;
12885 dtrace_enabling_addlike(new, ep, create);
12886 }
12887 }
12888
12889 if (!found || (err = dtrace_enabling_retain(new)) != 0) {
12890 dtrace_enabling_destroy(new);
12891 return (err);
12892 }
12893
12894 return (0);
12895 }
12896
12897 static void
12898 dtrace_enabling_retract(dtrace_state_t *state)
12899 {
12900 dtrace_enabling_t *enab, *next;
12901
12902 ASSERT(MUTEX_HELD(&dtrace_lock));
12903
12904 /*
12905 * Iterate over all retained enablings, destroy the enablings retained
12906 * for the specified state.
12907 */
12908 for (enab = dtrace_retained; enab != NULL; enab = next) {
12909 next = enab->dten_next;
12910
12911 /*
12912 * dtvs_state can only be NULL for helper enablings -- and
12913 * helper enablings can't be retained.
12914 */
12915 ASSERT(enab->dten_vstate->dtvs_state != NULL);
12916
12917 if (enab->dten_vstate->dtvs_state == state) {
12918 ASSERT(state->dts_nretained > 0);
12919 dtrace_enabling_destroy(enab);
12920 }
12921 }
12922
12923 ASSERT(state->dts_nretained == 0);
12924 }
12925
12926 static int
12927 dtrace_enabling_match(dtrace_enabling_t *enab, int *nmatched)
12928 {
12929 int i = 0;
12930 int matched = 0;
12931
12932 ASSERT(MUTEX_HELD(&cpu_lock));
12933 ASSERT(MUTEX_HELD(&dtrace_lock));
12934
12935 for (i = 0; i < enab->dten_ndesc; i++) {
12936 dtrace_ecbdesc_t *ep = enab->dten_desc[i];
12937
12938 enab->dten_current = ep;
12939 enab->dten_error = 0;
12940
12941 matched += dtrace_probe_enable(&ep->dted_probe, enab);
12942
12943 if (enab->dten_error != 0) {
12944 /*
12945 * If we get an error half-way through enabling the
12946 * probes, we kick out -- perhaps with some number of
12947 * them enabled. Leaving enabled probes enabled may
12948 * be slightly confusing for user-level, but we expect
12949 * that no one will attempt to actually drive on in
12950 * the face of such errors. If this is an anonymous
12951 * enabling (indicated with a NULL nmatched pointer),
12952 * we cmn_err() a message. We aren't expecting to
12953 * get such an error -- such as it can exist at all,
12954 * it would be a result of corrupted DOF in the driver
12955 * properties.
12956 */
12957 if (nmatched == NULL) {
12958 cmn_err(CE_WARN, "dtrace_enabling_match() "
12959 "error on %p: %d", (void *)ep,
12960 enab->dten_error);
12961 }
12962
12963 return (enab->dten_error);
12964 }
12965 }
12966
12967 enab->dten_probegen = dtrace_probegen;
12968 if (nmatched != NULL)
12969 *nmatched = matched;
12970
12971 return (0);
12972 }
12973
12974 static void
12975 dtrace_enabling_matchall(void)
12976 {
12977 dtrace_enabling_t *enab;
12978
12979 mutex_enter(&cpu_lock);
12980 mutex_enter(&dtrace_lock);
12981
12982 /*
12983 * Iterate over all retained enablings to see if any probes match
12984 * against them. We only perform this operation on enablings for which
12985 * we have sufficient permissions by virtue of being in the global zone
12986 * or in the same zone as the DTrace client. Because we can be called
12987 * after dtrace_detach() has been called, we cannot assert that there
12988 * are retained enablings. We can safely load from dtrace_retained,
12989 * however: the taskq_destroy() at the end of dtrace_detach() will
12990 * block pending our completion.
12991 */
12992 for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
12993 #ifdef illumos
12994 cred_t *cr = enab->dten_vstate->dtvs_state->dts_cred.dcr_cred;
12995
12996 if (INGLOBALZONE(curproc) ||
12997 cr != NULL && getzoneid() == crgetzoneid(cr))
12998 #endif
12999 (void) dtrace_enabling_match(enab, NULL);
13000 }
13001
13002 mutex_exit(&dtrace_lock);
13003 mutex_exit(&cpu_lock);
13004 }
13005
13006 /*
13007 * If an enabling is to be enabled without having matched probes (that is, if
13008 * dtrace_state_go() is to be called on the underlying dtrace_state_t), the
13009 * enabling must be _primed_ by creating an ECB for every ECB description.
13010 * This must be done to assure that we know the number of speculations, the
13011 * number of aggregations, the minimum buffer size needed, etc. before we
13012 * transition out of DTRACE_ACTIVITY_INACTIVE. To do this without actually
13013 * enabling any probes, we create ECBs for every ECB decription, but with a
13014 * NULL probe -- which is exactly what this function does.
13015 */
13016 static void
13017 dtrace_enabling_prime(dtrace_state_t *state)
13018 {
13019 dtrace_enabling_t *enab;
13020 int i;
13021
13022 for (enab = dtrace_retained; enab != NULL; enab = enab->dten_next) {
13023 ASSERT(enab->dten_vstate->dtvs_state != NULL);
13024
13025 if (enab->dten_vstate->dtvs_state != state)
13026 continue;
13027
13028 /*
13029 * We don't want to prime an enabling more than once, lest
13030 * we allow a malicious user to induce resource exhaustion.
13031 * (The ECBs that result from priming an enabling aren't
13032 * leaked -- but they also aren't deallocated until the
13033 * consumer state is destroyed.)
13034 */
13035 if (enab->dten_primed)
13036 continue;
13037
13038 for (i = 0; i < enab->dten_ndesc; i++) {
13039 enab->dten_current = enab->dten_desc[i];
13040 (void) dtrace_probe_enable(NULL, enab);
13041 }
13042
13043 enab->dten_primed = 1;
13044 }
13045 }
13046
13047 /*
13048 * Called to indicate that probes should be provided due to retained
13049 * enablings. This is implemented in terms of dtrace_probe_provide(), but it
13050 * must take an initial lap through the enabling calling the dtps_provide()
13051 * entry point explicitly to allow for autocreated probes.
13052 */
13053 static void
13054 dtrace_enabling_provide(dtrace_provider_t *prv)
13055 {
13056 int i, all = 0;
13057 dtrace_probedesc_t desc;
13058 dtrace_genid_t gen;
13059
13060 ASSERT(MUTEX_HELD(&dtrace_lock));
13061 ASSERT(MUTEX_HELD(&dtrace_provider_lock));
13062
13063 if (prv == NULL) {
13064 all = 1;
13065 prv = dtrace_provider;
13066 }
13067
13068 do {
13069 dtrace_enabling_t *enab;
13070 void *parg = prv->dtpv_arg;
13071
13072 retry:
13073 gen = dtrace_retained_gen;
13074 for (enab = dtrace_retained; enab != NULL;
13075 enab = enab->dten_next) {
13076 for (i = 0; i < enab->dten_ndesc; i++) {
13077 desc = enab->dten_desc[i]->dted_probe;
13078 mutex_exit(&dtrace_lock);
13079 prv->dtpv_pops.dtps_provide(parg, &desc);
13080 mutex_enter(&dtrace_lock);
13081 /*
13082 * Process the retained enablings again if
13083 * they have changed while we weren't holding
13084 * dtrace_lock.
13085 */
13086 if (gen != dtrace_retained_gen)
13087 goto retry;
13088 }
13089 }
13090 } while (all && (prv = prv->dtpv_next) != NULL);
13091
13092 mutex_exit(&dtrace_lock);
13093 dtrace_probe_provide(NULL, all ? NULL : prv);
13094 mutex_enter(&dtrace_lock);
13095 }
13096
13097 /*
13098 * Called to reap ECBs that are attached to probes from defunct providers.
13099 */
13100 static void
13101 dtrace_enabling_reap(void)
13102 {
13103 dtrace_provider_t *prov;
13104 dtrace_probe_t *probe;
13105 dtrace_ecb_t *ecb;
13106 hrtime_t when;
13107 int i;
13108
13109 mutex_enter(&cpu_lock);
13110 mutex_enter(&dtrace_lock);
13111
13112 for (i = 0; i < dtrace_nprobes; i++) {
13113 if ((probe = dtrace_probes[i]) == NULL)
13114 continue;
13115
13116 if (probe->dtpr_ecb == NULL)
13117 continue;
13118
13119 prov = probe->dtpr_provider;
13120
13121 if ((when = prov->dtpv_defunct) == 0)
13122 continue;
13123
13124 /*
13125 * We have ECBs on a defunct provider: we want to reap these
13126 * ECBs to allow the provider to unregister. The destruction
13127 * of these ECBs must be done carefully: if we destroy the ECB
13128 * and the consumer later wishes to consume an EPID that
13129 * corresponds to the destroyed ECB (and if the EPID metadata
13130 * has not been previously consumed), the consumer will abort
13131 * processing on the unknown EPID. To reduce (but not, sadly,
13132 * eliminate) the possibility of this, we will only destroy an
13133 * ECB for a defunct provider if, for the state that
13134 * corresponds to the ECB:
13135 *
13136 * (a) There is no speculative tracing (which can effectively
13137 * cache an EPID for an arbitrary amount of time).
13138 *
13139 * (b) The principal buffers have been switched twice since the
13140 * provider became defunct.
13141 *
13142 * (c) The aggregation buffers are of zero size or have been
13143 * switched twice since the provider became defunct.
13144 *
13145 * We use dts_speculates to determine (a) and call a function
13146 * (dtrace_buffer_consumed()) to determine (b) and (c). Note
13147 * that as soon as we've been unable to destroy one of the ECBs
13148 * associated with the probe, we quit trying -- reaping is only
13149 * fruitful in as much as we can destroy all ECBs associated
13150 * with the defunct provider's probes.
13151 */
13152 while ((ecb = probe->dtpr_ecb) != NULL) {
13153 dtrace_state_t *state = ecb->dte_state;
13154 dtrace_buffer_t *buf = state->dts_buffer;
13155 dtrace_buffer_t *aggbuf = state->dts_aggbuffer;
13156
13157 if (state->dts_speculates)
13158 break;
13159
13160 if (!dtrace_buffer_consumed(buf, when))
13161 break;
13162
13163 if (!dtrace_buffer_consumed(aggbuf, when))
13164 break;
13165
13166 dtrace_ecb_disable(ecb);
13167 ASSERT(probe->dtpr_ecb != ecb);
13168 dtrace_ecb_destroy(ecb);
13169 }
13170 }
13171
13172 mutex_exit(&dtrace_lock);
13173 mutex_exit(&cpu_lock);
13174 }
13175
13176 /*
13177 * DTrace DOF Functions
13178 */
13179 /*ARGSUSED*/
13180 static void
13181 dtrace_dof_error(dof_hdr_t *dof, const char *str)
13182 {
13183 if (dtrace_err_verbose)
13184 cmn_err(CE_WARN, "failed to process DOF: %s", str);
13185
13186 #ifdef DTRACE_ERRDEBUG
13187 dtrace_errdebug(str);
13188 #endif
13189 }
13190
13191 /*
13192 * Create DOF out of a currently enabled state. Right now, we only create
13193 * DOF containing the run-time options -- but this could be expanded to create
13194 * complete DOF representing the enabled state.
13195 */
13196 static dof_hdr_t *
13197 dtrace_dof_create(dtrace_state_t *state)
13198 {
13199 dof_hdr_t *dof;
13200 dof_sec_t *sec;
13201 dof_optdesc_t *opt;
13202 int i, len = sizeof (dof_hdr_t) +
13203 roundup(sizeof (dof_sec_t), sizeof (uint64_t)) +
13204 sizeof (dof_optdesc_t) * DTRACEOPT_MAX;
13205
13206 ASSERT(MUTEX_HELD(&dtrace_lock));
13207
13208 dof = kmem_zalloc(len, KM_SLEEP);
13209 dof->dofh_ident[DOF_ID_MAG0] = DOF_MAG_MAG0;
13210 dof->dofh_ident[DOF_ID_MAG1] = DOF_MAG_MAG1;
13211 dof->dofh_ident[DOF_ID_MAG2] = DOF_MAG_MAG2;
13212 dof->dofh_ident[DOF_ID_MAG3] = DOF_MAG_MAG3;
13213
13214 dof->dofh_ident[DOF_ID_MODEL] = DOF_MODEL_NATIVE;
13215 dof->dofh_ident[DOF_ID_ENCODING] = DOF_ENCODE_NATIVE;
13216 dof->dofh_ident[DOF_ID_VERSION] = DOF_VERSION;
13217 dof->dofh_ident[DOF_ID_DIFVERS] = DIF_VERSION;
13218 dof->dofh_ident[DOF_ID_DIFIREG] = DIF_DIR_NREGS;
13219 dof->dofh_ident[DOF_ID_DIFTREG] = DIF_DTR_NREGS;
13220
13221 dof->dofh_flags = 0;
13222 dof->dofh_hdrsize = sizeof (dof_hdr_t);
13223 dof->dofh_secsize = sizeof (dof_sec_t);
13224 dof->dofh_secnum = 1; /* only DOF_SECT_OPTDESC */
13225 dof->dofh_secoff = sizeof (dof_hdr_t);
13226 dof->dofh_loadsz = len;
13227 dof->dofh_filesz = len;
13228 dof->dofh_pad = 0;
13229
13230 /*
13231 * Fill in the option section header...
13232 */
13233 sec = (dof_sec_t *)((uintptr_t)dof + sizeof (dof_hdr_t));
13234 sec->dofs_type = DOF_SECT_OPTDESC;
13235 sec->dofs_align = sizeof (uint64_t);
13236 sec->dofs_flags = DOF_SECF_LOAD;
13237 sec->dofs_entsize = sizeof (dof_optdesc_t);
13238
13239 opt = (dof_optdesc_t *)((uintptr_t)sec +
13240 roundup(sizeof (dof_sec_t), sizeof (uint64_t)));
13241
13242 sec->dofs_offset = (uintptr_t)opt - (uintptr_t)dof;
13243 sec->dofs_size = sizeof (dof_optdesc_t) * DTRACEOPT_MAX;
13244
13245 for (i = 0; i < DTRACEOPT_MAX; i++) {
13246 opt[i].dofo_option = i;
13247 opt[i].dofo_strtab = DOF_SECIDX_NONE;
13248 opt[i].dofo_value = state->dts_options[i];
13249 }
13250
13251 return (dof);
13252 }
13253
13254 static dof_hdr_t *
13255 dtrace_dof_copyin(uintptr_t uarg, int *errp)
13256 {
13257 dof_hdr_t hdr, *dof;
13258
13259 ASSERT(!MUTEX_HELD(&dtrace_lock));
13260
13261 /*
13262 * First, we're going to copyin() the sizeof (dof_hdr_t).
13263 */
13264 if (copyin((void *)uarg, &hdr, sizeof (hdr)) != 0) {
13265 dtrace_dof_error(NULL, "failed to copyin DOF header");
13266 *errp = EFAULT;
13267 return (NULL);
13268 }
13269
13270 /*
13271 * Now we'll allocate the entire DOF and copy it in -- provided
13272 * that the length isn't outrageous.
13273 */
13274 if (hdr.dofh_loadsz >= dtrace_dof_maxsize) {
13275 dtrace_dof_error(&hdr, "load size exceeds maximum");
13276 *errp = E2BIG;
13277 return (NULL);
13278 }
13279
13280 if (hdr.dofh_loadsz < sizeof (hdr)) {
13281 dtrace_dof_error(&hdr, "invalid load size");
13282 *errp = EINVAL;
13283 return (NULL);
13284 }
13285
13286 dof = kmem_alloc(hdr.dofh_loadsz, KM_SLEEP);
13287
13288 if (copyin((void *)uarg, dof, hdr.dofh_loadsz) != 0 ||
13289 dof->dofh_loadsz != hdr.dofh_loadsz) {
13290 kmem_free(dof, hdr.dofh_loadsz);
13291 *errp = EFAULT;
13292 return (NULL);
13293 }
13294
13295 return (dof);
13296 }
13297
13298 #ifdef __FreeBSD__
13299 static dof_hdr_t *
13300 dtrace_dof_copyin_proc(struct proc *p, uintptr_t uarg, int *errp)
13301 {
13302 dof_hdr_t hdr, *dof;
13303 struct thread *td;
13304 size_t loadsz;
13305
13306 ASSERT(!MUTEX_HELD(&dtrace_lock));
13307
13308 td = curthread;
13309
13310 /*
13311 * First, we're going to copyin() the sizeof (dof_hdr_t).
13312 */
13313 if (proc_readmem(td, p, uarg, &hdr, sizeof(hdr)) != sizeof(hdr)) {
13314 dtrace_dof_error(NULL, "failed to copyin DOF header");
13315 *errp = EFAULT;
13316 return (NULL);
13317 }
13318
13319 /*
13320 * Now we'll allocate the entire DOF and copy it in -- provided
13321 * that the length isn't outrageous.
13322 */
13323 if (hdr.dofh_loadsz >= dtrace_dof_maxsize) {
13324 dtrace_dof_error(&hdr, "load size exceeds maximum");
13325 *errp = E2BIG;
13326 return (NULL);
13327 }
13328 loadsz = (size_t)hdr.dofh_loadsz;
13329
13330 if (loadsz < sizeof (hdr)) {
13331 dtrace_dof_error(&hdr, "invalid load size");
13332 *errp = EINVAL;
13333 return (NULL);
13334 }
13335
13336 dof = kmem_alloc(loadsz, KM_SLEEP);
13337
13338 if (proc_readmem(td, p, uarg, dof, loadsz) != loadsz ||
13339 dof->dofh_loadsz != loadsz) {
13340 kmem_free(dof, hdr.dofh_loadsz);
13341 *errp = EFAULT;
13342 return (NULL);
13343 }
13344
13345 return (dof);
13346 }
13347
13348 static __inline uchar_t
13349 dtrace_dof_char(char c)
13350 {
13351
13352 switch (c) {
13353 case '':
13354 case '1':
13355 case '2':
13356 case '3':
13357 case '4':
13358 case '5':
13359 case '6':
13360 case '7':
13361 case '8':
13362 case '9':
13363 return (c - '');
13364 case 'A':
13365 case 'B':
13366 case 'C':
13367 case 'D':
13368 case 'E':
13369 case 'F':
13370 return (c - 'A' + 10);
13371 case 'a':
13372 case 'b':
13373 case 'c':
13374 case 'd':
13375 case 'e':
13376 case 'f':
13377 return (c - 'a' + 10);
13378 }
13379 /* Should not reach here. */
13380 return (UCHAR_MAX);
13381 }
13382 #endif /* __FreeBSD__ */
13383
13384 static dof_hdr_t *
13385 dtrace_dof_property(const char *name)
13386 {
13387 #ifdef __FreeBSD__
13388 uint8_t *dofbuf;
13389 u_char *data, *eol;
13390 caddr_t doffile;
13391 size_t bytes, len, i;
13392 dof_hdr_t *dof;
13393 u_char c1, c2;
13394
13395 dof = NULL;
13396
13397 doffile = preload_search_by_type("dtrace_dof");
13398 if (doffile == NULL)
13399 return (NULL);
13400
13401 data = preload_fetch_addr(doffile);
13402 len = preload_fetch_size(doffile);
13403 for (;;) {
13404 /* Look for the end of the line. All lines end in a newline. */
13405 eol = memchr(data, '\n', len);
13406 if (eol == NULL)
13407 return (NULL);
13408
13409 if (strncmp(name, data, strlen(name)) == 0)
13410 break;
13411
13412 eol++; /* skip past the newline */
13413 len -= eol - data;
13414 data = eol;
13415 }
13416
13417 /* We've found the data corresponding to the specified key. */
13418
13419 data += strlen(name) + 1; /* skip past the '=' */
13420 len = eol - data;
13421 if (len % 2 != 0) {
13422 dtrace_dof_error(NULL, "invalid DOF encoding length");
13423 goto doferr;
13424 }
13425 bytes = len / 2;
13426 if (bytes < sizeof(dof_hdr_t)) {
13427 dtrace_dof_error(NULL, "truncated header");
13428 goto doferr;
13429 }
13430
13431 /*
13432 * Each byte is represented by the two ASCII characters in its hex
13433 * representation.
13434 */
13435 dofbuf = malloc(bytes, M_SOLARIS, M_WAITOK);
13436 for (i = 0; i < bytes; i++) {
13437 c1 = dtrace_dof_char(data[i * 2]);
13438 c2 = dtrace_dof_char(data[i * 2 + 1]);
13439 if (c1 == UCHAR_MAX || c2 == UCHAR_MAX) {
13440 dtrace_dof_error(NULL, "invalid hex char in DOF");
13441 goto doferr;
13442 }
13443 dofbuf[i] = c1 * 16 + c2;
13444 }
13445
13446 dof = (dof_hdr_t *)dofbuf;
13447 if (bytes < dof->dofh_loadsz) {
13448 dtrace_dof_error(NULL, "truncated DOF");
13449 goto doferr;
13450 }
13451
13452 if (dof->dofh_loadsz >= dtrace_dof_maxsize) {
13453 dtrace_dof_error(NULL, "oversized DOF");
13454 goto doferr;
13455 }
13456
13457 return (dof);
13458
13459 doferr:
13460 free(dof, M_SOLARIS);
13461 return (NULL);
13462 #else /* __FreeBSD__ */
13463 uchar_t *buf;
13464 uint64_t loadsz;
13465 unsigned int len, i;
13466 dof_hdr_t *dof;
13467
13468 /*
13469 * Unfortunately, array of values in .conf files are always (and
13470 * only) interpreted to be integer arrays. We must read our DOF
13471 * as an integer array, and then squeeze it into a byte array.
13472 */
13473 if (ddi_prop_lookup_int_array(DDI_DEV_T_ANY, dtrace_devi, 0,
13474 (char *)name, (int **)&buf, &len) != DDI_PROP_SUCCESS)
13475 return (NULL);
13476
13477 for (i = 0; i < len; i++)
13478 buf[i] = (uchar_t)(((int *)buf)[i]);
13479
13480 if (len < sizeof (dof_hdr_t)) {
13481 ddi_prop_free(buf);
13482 dtrace_dof_error(NULL, "truncated header");
13483 return (NULL);
13484 }
13485
13486 if (len < (loadsz = ((dof_hdr_t *)buf)->dofh_loadsz)) {
13487 ddi_prop_free(buf);
13488 dtrace_dof_error(NULL, "truncated DOF");
13489 return (NULL);
13490 }
13491
13492 if (loadsz >= dtrace_dof_maxsize) {
13493 ddi_prop_free(buf);
13494 dtrace_dof_error(NULL, "oversized DOF");
13495 return (NULL);
13496 }
13497
13498 dof = kmem_alloc(loadsz, KM_SLEEP);
13499 bcopy(buf, dof, loadsz);
13500 ddi_prop_free(buf);
13501
13502 return (dof);
13503 #endif /* !__FreeBSD__ */
13504 }
13505
13506 static void
13507 dtrace_dof_destroy(dof_hdr_t *dof)
13508 {
13509 kmem_free(dof, dof->dofh_loadsz);
13510 }
13511
13512 /*
13513 * Return the dof_sec_t pointer corresponding to a given section index. If the
13514 * index is not valid, dtrace_dof_error() is called and NULL is returned. If
13515 * a type other than DOF_SECT_NONE is specified, the header is checked against
13516 * this type and NULL is returned if the types do not match.
13517 */
13518 static dof_sec_t *
13519 dtrace_dof_sect(dof_hdr_t *dof, uint32_t type, dof_secidx_t i)
13520 {
13521 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)
13522 ((uintptr_t)dof + dof->dofh_secoff + i * dof->dofh_secsize);
13523
13524 if (i >= dof->dofh_secnum) {
13525 dtrace_dof_error(dof, "referenced section index is invalid");
13526 return (NULL);
13527 }
13528
13529 if (!(sec->dofs_flags & DOF_SECF_LOAD)) {
13530 dtrace_dof_error(dof, "referenced section is not loadable");
13531 return (NULL);
13532 }
13533
13534 if (type != DOF_SECT_NONE && type != sec->dofs_type) {
13535 dtrace_dof_error(dof, "referenced section is the wrong type");
13536 return (NULL);
13537 }
13538
13539 return (sec);
13540 }
13541
13542 static dtrace_probedesc_t *
13543 dtrace_dof_probedesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_probedesc_t *desc)
13544 {
13545 dof_probedesc_t *probe;
13546 dof_sec_t *strtab;
13547 uintptr_t daddr = (uintptr_t)dof;
13548 uintptr_t str;
13549 size_t size;
13550
13551 if (sec->dofs_type != DOF_SECT_PROBEDESC) {
13552 dtrace_dof_error(dof, "invalid probe section");
13553 return (NULL);
13554 }
13555
13556 if (sec->dofs_align != sizeof (dof_secidx_t)) {
13557 dtrace_dof_error(dof, "bad alignment in probe description");
13558 return (NULL);
13559 }
13560
13561 if (sec->dofs_offset + sizeof (dof_probedesc_t) > dof->dofh_loadsz) {
13562 dtrace_dof_error(dof, "truncated probe description");
13563 return (NULL);
13564 }
13565
13566 probe = (dof_probedesc_t *)(uintptr_t)(daddr + sec->dofs_offset);
13567 strtab = dtrace_dof_sect(dof, DOF_SECT_STRTAB, probe->dofp_strtab);
13568
13569 if (strtab == NULL)
13570 return (NULL);
13571
13572 str = daddr + strtab->dofs_offset;
13573 size = strtab->dofs_size;
13574
13575 if (probe->dofp_provider >= strtab->dofs_size) {
13576 dtrace_dof_error(dof, "corrupt probe provider");
13577 return (NULL);
13578 }
13579
13580 (void) strncpy(desc->dtpd_provider,
13581 (char *)(str + probe->dofp_provider),
13582 MIN(DTRACE_PROVNAMELEN - 1, size - probe->dofp_provider));
13583
13584 if (probe->dofp_mod >= strtab->dofs_size) {
13585 dtrace_dof_error(dof, "corrupt probe module");
13586 return (NULL);
13587 }
13588
13589 (void) strncpy(desc->dtpd_mod, (char *)(str + probe->dofp_mod),
13590 MIN(DTRACE_MODNAMELEN - 1, size - probe->dofp_mod));
13591
13592 if (probe->dofp_func >= strtab->dofs_size) {
13593 dtrace_dof_error(dof, "corrupt probe function");
13594 return (NULL);
13595 }
13596
13597 (void) strncpy(desc->dtpd_func, (char *)(str + probe->dofp_func),
13598 MIN(DTRACE_FUNCNAMELEN - 1, size - probe->dofp_func));
13599
13600 if (probe->dofp_name >= strtab->dofs_size) {
13601 dtrace_dof_error(dof, "corrupt probe name");
13602 return (NULL);
13603 }
13604
13605 (void) strncpy(desc->dtpd_name, (char *)(str + probe->dofp_name),
13606 MIN(DTRACE_NAMELEN - 1, size - probe->dofp_name));
13607
13608 return (desc);
13609 }
13610
13611 static dtrace_difo_t *
13612 dtrace_dof_difo(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
13613 cred_t *cr)
13614 {
13615 dtrace_difo_t *dp;
13616 size_t ttl = 0;
13617 dof_difohdr_t *dofd;
13618 uintptr_t daddr = (uintptr_t)dof;
13619 size_t max = dtrace_difo_maxsize;
13620 int i, l, n;
13621
13622 static const struct {
13623 int section;
13624 int bufoffs;
13625 int lenoffs;
13626 int entsize;
13627 int align;
13628 const char *msg;
13629 } difo[] = {
13630 { DOF_SECT_DIF, offsetof(dtrace_difo_t, dtdo_buf),
13631 offsetof(dtrace_difo_t, dtdo_len), sizeof (dif_instr_t),
13632 sizeof (dif_instr_t), "multiple DIF sections" },
13633
13634 { DOF_SECT_INTTAB, offsetof(dtrace_difo_t, dtdo_inttab),
13635 offsetof(dtrace_difo_t, dtdo_intlen), sizeof (uint64_t),
13636 sizeof (uint64_t), "multiple integer tables" },
13637
13638 { DOF_SECT_STRTAB, offsetof(dtrace_difo_t, dtdo_strtab),
13639 offsetof(dtrace_difo_t, dtdo_strlen), 0,
13640 sizeof (char), "multiple string tables" },
13641
13642 { DOF_SECT_VARTAB, offsetof(dtrace_difo_t, dtdo_vartab),
13643 offsetof(dtrace_difo_t, dtdo_varlen), sizeof (dtrace_difv_t),
13644 sizeof (uint_t), "multiple variable tables" },
13645
13646 { DOF_SECT_NONE, 0, 0, 0, 0, NULL }
13647 };
13648
13649 if (sec->dofs_type != DOF_SECT_DIFOHDR) {
13650 dtrace_dof_error(dof, "invalid DIFO header section");
13651 return (NULL);
13652 }
13653
13654 if (sec->dofs_align != sizeof (dof_secidx_t)) {
13655 dtrace_dof_error(dof, "bad alignment in DIFO header");
13656 return (NULL);
13657 }
13658
13659 if (sec->dofs_size < sizeof (dof_difohdr_t) ||
13660 sec->dofs_size % sizeof (dof_secidx_t)) {
13661 dtrace_dof_error(dof, "bad size in DIFO header");
13662 return (NULL);
13663 }
13664
13665 dofd = (dof_difohdr_t *)(uintptr_t)(daddr + sec->dofs_offset);
13666 n = (sec->dofs_size - sizeof (*dofd)) / sizeof (dof_secidx_t) + 1;
13667
13668 dp = kmem_zalloc(sizeof (dtrace_difo_t), KM_SLEEP);
13669 dp->dtdo_rtype = dofd->dofd_rtype;
13670
13671 for (l = 0; l < n; l++) {
13672 dof_sec_t *subsec;
13673 void **bufp;
13674 uint32_t *lenp;
13675
13676 if ((subsec = dtrace_dof_sect(dof, DOF_SECT_NONE,
13677 dofd->dofd_links[l])) == NULL)
13678 goto err; /* invalid section link */
13679
13680 if (ttl + subsec->dofs_size > max) {
13681 dtrace_dof_error(dof, "exceeds maximum size");
13682 goto err;
13683 }
13684
13685 ttl += subsec->dofs_size;
13686
13687 for (i = 0; difo[i].section != DOF_SECT_NONE; i++) {
13688 if (subsec->dofs_type != difo[i].section)
13689 continue;
13690
13691 if (!(subsec->dofs_flags & DOF_SECF_LOAD)) {
13692 dtrace_dof_error(dof, "section not loaded");
13693 goto err;
13694 }
13695
13696 if (subsec->dofs_align != difo[i].align) {
13697 dtrace_dof_error(dof, "bad alignment");
13698 goto err;
13699 }
13700
13701 bufp = (void **)((uintptr_t)dp + difo[i].bufoffs);
13702 lenp = (uint32_t *)((uintptr_t)dp + difo[i].lenoffs);
13703
13704 if (*bufp != NULL) {
13705 dtrace_dof_error(dof, difo[i].msg);
13706 goto err;
13707 }
13708
13709 if (difo[i].entsize != subsec->dofs_entsize) {
13710 dtrace_dof_error(dof, "entry size mismatch");
13711 goto err;
13712 }
13713
13714 if (subsec->dofs_entsize != 0 &&
13715 (subsec->dofs_size % subsec->dofs_entsize) != 0) {
13716 dtrace_dof_error(dof, "corrupt entry size");
13717 goto err;
13718 }
13719
13720 *lenp = subsec->dofs_size;
13721 *bufp = kmem_alloc(subsec->dofs_size, KM_SLEEP);
13722 bcopy((char *)(uintptr_t)(daddr + subsec->dofs_offset),
13723 *bufp, subsec->dofs_size);
13724
13725 if (subsec->dofs_entsize != 0)
13726 *lenp /= subsec->dofs_entsize;
13727
13728 break;
13729 }
13730
13731 /*
13732 * If we encounter a loadable DIFO sub-section that is not
13733 * known to us, assume this is a broken program and fail.
13734 */
13735 if (difo[i].section == DOF_SECT_NONE &&
13736 (subsec->dofs_flags & DOF_SECF_LOAD)) {
13737 dtrace_dof_error(dof, "unrecognized DIFO subsection");
13738 goto err;
13739 }
13740 }
13741
13742 if (dp->dtdo_buf == NULL) {
13743 /*
13744 * We can't have a DIF object without DIF text.
13745 */
13746 dtrace_dof_error(dof, "missing DIF text");
13747 goto err;
13748 }
13749
13750 /*
13751 * Before we validate the DIF object, run through the variable table
13752 * looking for the strings -- if any of their size are under, we'll set
13753 * their size to be the system-wide default string size. Note that
13754 * this should _not_ happen if the "strsize" option has been set --
13755 * in this case, the compiler should have set the size to reflect the
13756 * setting of the option.
13757 */
13758 for (i = 0; i < dp->dtdo_varlen; i++) {
13759 dtrace_difv_t *v = &dp->dtdo_vartab[i];
13760 dtrace_diftype_t *t = &v->dtdv_type;
13761
13762 if (v->dtdv_id < DIF_VAR_OTHER_UBASE)
13763 continue;
13764
13765 if (t->dtdt_kind == DIF_TYPE_STRING && t->dtdt_size == 0)
13766 t->dtdt_size = dtrace_strsize_default;
13767 }
13768
13769 if (dtrace_difo_validate(dp, vstate, DIF_DIR_NREGS, cr) != 0)
13770 goto err;
13771
13772 dtrace_difo_init(dp, vstate);
13773 return (dp);
13774
13775 err:
13776 kmem_free(dp->dtdo_buf, dp->dtdo_len * sizeof (dif_instr_t));
13777 kmem_free(dp->dtdo_inttab, dp->dtdo_intlen * sizeof (uint64_t));
13778 kmem_free(dp->dtdo_strtab, dp->dtdo_strlen);
13779 kmem_free(dp->dtdo_vartab, dp->dtdo_varlen * sizeof (dtrace_difv_t));
13780
13781 kmem_free(dp, sizeof (dtrace_difo_t));
13782 return (NULL);
13783 }
13784
13785 static dtrace_predicate_t *
13786 dtrace_dof_predicate(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
13787 cred_t *cr)
13788 {
13789 dtrace_difo_t *dp;
13790
13791 if ((dp = dtrace_dof_difo(dof, sec, vstate, cr)) == NULL)
13792 return (NULL);
13793
13794 return (dtrace_predicate_create(dp));
13795 }
13796
13797 static dtrace_actdesc_t *
13798 dtrace_dof_actdesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
13799 cred_t *cr)
13800 {
13801 dtrace_actdesc_t *act, *first = NULL, *last = NULL, *next;
13802 dof_actdesc_t *desc;
13803 dof_sec_t *difosec;
13804 size_t offs;
13805 uintptr_t daddr = (uintptr_t)dof;
13806 uint64_t arg;
13807 dtrace_actkind_t kind;
13808
13809 if (sec->dofs_type != DOF_SECT_ACTDESC) {
13810 dtrace_dof_error(dof, "invalid action section");
13811 return (NULL);
13812 }
13813
13814 if (sec->dofs_offset + sizeof (dof_actdesc_t) > dof->dofh_loadsz) {
13815 dtrace_dof_error(dof, "truncated action description");
13816 return (NULL);
13817 }
13818
13819 if (sec->dofs_align != sizeof (uint64_t)) {
13820 dtrace_dof_error(dof, "bad alignment in action description");
13821 return (NULL);
13822 }
13823
13824 if (sec->dofs_size < sec->dofs_entsize) {
13825 dtrace_dof_error(dof, "section entry size exceeds total size");
13826 return (NULL);
13827 }
13828
13829 if (sec->dofs_entsize != sizeof (dof_actdesc_t)) {
13830 dtrace_dof_error(dof, "bad entry size in action description");
13831 return (NULL);
13832 }
13833
13834 if (sec->dofs_size / sec->dofs_entsize > dtrace_actions_max) {
13835 dtrace_dof_error(dof, "actions exceed dtrace_actions_max");
13836 return (NULL);
13837 }
13838
13839 for (offs = 0; offs < sec->dofs_size; offs += sec->dofs_entsize) {
13840 desc = (dof_actdesc_t *)(daddr +
13841 (uintptr_t)sec->dofs_offset + offs);
13842 kind = (dtrace_actkind_t)desc->dofa_kind;
13843
13844 if ((DTRACEACT_ISPRINTFLIKE(kind) &&
13845 (kind != DTRACEACT_PRINTA ||
13846 desc->dofa_strtab != DOF_SECIDX_NONE)) ||
13847 (kind == DTRACEACT_DIFEXPR &&
13848 desc->dofa_strtab != DOF_SECIDX_NONE)) {
13849 dof_sec_t *strtab;
13850 char *str, *fmt;
13851 uint64_t i;
13852
13853 /*
13854 * The argument to these actions is an index into the
13855 * DOF string table. For printf()-like actions, this
13856 * is the format string. For print(), this is the
13857 * CTF type of the expression result.
13858 */
13859 if ((strtab = dtrace_dof_sect(dof,
13860 DOF_SECT_STRTAB, desc->dofa_strtab)) == NULL)
13861 goto err;
13862
13863 str = (char *)((uintptr_t)dof +
13864 (uintptr_t)strtab->dofs_offset);
13865
13866 for (i = desc->dofa_arg; i < strtab->dofs_size; i++) {
13867 if (str[i] == '\0')
13868 break;
13869 }
13870
13871 if (i >= strtab->dofs_size) {
13872 dtrace_dof_error(dof, "bogus format string");
13873 goto err;
13874 }
13875
13876 if (i == desc->dofa_arg) {
13877 dtrace_dof_error(dof, "empty format string");
13878 goto err;
13879 }
13880
13881 i -= desc->dofa_arg;
13882 fmt = kmem_alloc(i + 1, KM_SLEEP);
13883 bcopy(&str[desc->dofa_arg], fmt, i + 1);
13884 arg = (uint64_t)(uintptr_t)fmt;
13885 } else {
13886 if (kind == DTRACEACT_PRINTA) {
13887 ASSERT(desc->dofa_strtab == DOF_SECIDX_NONE);
13888 arg = 0;
13889 } else {
13890 arg = desc->dofa_arg;
13891 }
13892 }
13893
13894 act = dtrace_actdesc_create(kind, desc->dofa_ntuple,
13895 desc->dofa_uarg, arg);
13896
13897 if (last != NULL) {
13898 last->dtad_next = act;
13899 } else {
13900 first = act;
13901 }
13902
13903 last = act;
13904
13905 if (desc->dofa_difo == DOF_SECIDX_NONE)
13906 continue;
13907
13908 if ((difosec = dtrace_dof_sect(dof,
13909 DOF_SECT_DIFOHDR, desc->dofa_difo)) == NULL)
13910 goto err;
13911
13912 act->dtad_difo = dtrace_dof_difo(dof, difosec, vstate, cr);
13913
13914 if (act->dtad_difo == NULL)
13915 goto err;
13916 }
13917
13918 ASSERT(first != NULL);
13919 return (first);
13920
13921 err:
13922 for (act = first; act != NULL; act = next) {
13923 next = act->dtad_next;
13924 dtrace_actdesc_release(act, vstate);
13925 }
13926
13927 return (NULL);
13928 }
13929
13930 static dtrace_ecbdesc_t *
13931 dtrace_dof_ecbdesc(dof_hdr_t *dof, dof_sec_t *sec, dtrace_vstate_t *vstate,
13932 cred_t *cr)
13933 {
13934 dtrace_ecbdesc_t *ep;
13935 dof_ecbdesc_t *ecb;
13936 dtrace_probedesc_t *desc;
13937 dtrace_predicate_t *pred = NULL;
13938
13939 if (sec->dofs_size < sizeof (dof_ecbdesc_t)) {
13940 dtrace_dof_error(dof, "truncated ECB description");
13941 return (NULL);
13942 }
13943
13944 if (sec->dofs_align != sizeof (uint64_t)) {
13945 dtrace_dof_error(dof, "bad alignment in ECB description");
13946 return (NULL);
13947 }
13948
13949 ecb = (dof_ecbdesc_t *)((uintptr_t)dof + (uintptr_t)sec->dofs_offset);
13950 sec = dtrace_dof_sect(dof, DOF_SECT_PROBEDESC, ecb->dofe_probes);
13951
13952 if (sec == NULL)
13953 return (NULL);
13954
13955 ep = kmem_zalloc(sizeof (dtrace_ecbdesc_t), KM_SLEEP);
13956 ep->dted_uarg = ecb->dofe_uarg;
13957 desc = &ep->dted_probe;
13958
13959 if (dtrace_dof_probedesc(dof, sec, desc) == NULL)
13960 goto err;
13961
13962 if (ecb->dofe_pred != DOF_SECIDX_NONE) {
13963 if ((sec = dtrace_dof_sect(dof,
13964 DOF_SECT_DIFOHDR, ecb->dofe_pred)) == NULL)
13965 goto err;
13966
13967 if ((pred = dtrace_dof_predicate(dof, sec, vstate, cr)) == NULL)
13968 goto err;
13969
13970 ep->dted_pred.dtpdd_predicate = pred;
13971 }
13972
13973 if (ecb->dofe_actions != DOF_SECIDX_NONE) {
13974 if ((sec = dtrace_dof_sect(dof,
13975 DOF_SECT_ACTDESC, ecb->dofe_actions)) == NULL)
13976 goto err;
13977
13978 ep->dted_action = dtrace_dof_actdesc(dof, sec, vstate, cr);
13979
13980 if (ep->dted_action == NULL)
13981 goto err;
13982 }
13983
13984 return (ep);
13985
13986 err:
13987 if (pred != NULL)
13988 dtrace_predicate_release(pred, vstate);
13989 kmem_free(ep, sizeof (dtrace_ecbdesc_t));
13990 return (NULL);
13991 }
13992
13993 /*
13994 * Apply the relocations from the specified 'sec' (a DOF_SECT_URELHDR) to the
13995 * specified DOF. SETX relocations are computed using 'ubase', the base load
13996 * address of the object containing the DOF, and DOFREL relocations are relative
13997 * to the relocation offset within the DOF.
13998 */
13999 static int
14000 dtrace_dof_relocate(dof_hdr_t *dof, dof_sec_t *sec, uint64_t ubase,
14001 uint64_t udaddr)
14002 {
14003 uintptr_t daddr = (uintptr_t)dof;
14004 uintptr_t ts_end;
14005 dof_relohdr_t *dofr =
14006 (dof_relohdr_t *)(uintptr_t)(daddr + sec->dofs_offset);
14007 dof_sec_t *ss, *rs, *ts;
14008 dof_relodesc_t *r;
14009 uint_t i, n;
14010
14011 if (sec->dofs_size < sizeof (dof_relohdr_t) ||
14012 sec->dofs_align != sizeof (dof_secidx_t)) {
14013 dtrace_dof_error(dof, "invalid relocation header");
14014 return (-1);
14015 }
14016
14017 ss = dtrace_dof_sect(dof, DOF_SECT_STRTAB, dofr->dofr_strtab);
14018 rs = dtrace_dof_sect(dof, DOF_SECT_RELTAB, dofr->dofr_relsec);
14019 ts = dtrace_dof_sect(dof, DOF_SECT_NONE, dofr->dofr_tgtsec);
14020 ts_end = (uintptr_t)ts + sizeof (dof_sec_t);
14021
14022 if (ss == NULL || rs == NULL || ts == NULL)
14023 return (-1); /* dtrace_dof_error() has been called already */
14024
14025 if (rs->dofs_entsize < sizeof (dof_relodesc_t) ||
14026 rs->dofs_align != sizeof (uint64_t)) {
14027 dtrace_dof_error(dof, "invalid relocation section");
14028 return (-1);
14029 }
14030
14031 r = (dof_relodesc_t *)(uintptr_t)(daddr + rs->dofs_offset);
14032 n = rs->dofs_size / rs->dofs_entsize;
14033
14034 for (i = 0; i < n; i++) {
14035 uintptr_t taddr = daddr + ts->dofs_offset + r->dofr_offset;
14036
14037 switch (r->dofr_type) {
14038 case DOF_RELO_NONE:
14039 break;
14040 case DOF_RELO_SETX:
14041 case DOF_RELO_DOFREL:
14042 if (r->dofr_offset >= ts->dofs_size || r->dofr_offset +
14043 sizeof (uint64_t) > ts->dofs_size) {
14044 dtrace_dof_error(dof, "bad relocation offset");
14045 return (-1);
14046 }
14047
14048 if (taddr >= (uintptr_t)ts && taddr < ts_end) {
14049 dtrace_dof_error(dof, "bad relocation offset");
14050 return (-1);
14051 }
14052
14053 if (!IS_P2ALIGNED(taddr, sizeof (uint64_t))) {
14054 dtrace_dof_error(dof, "misaligned setx relo");
14055 return (-1);
14056 }
14057
14058 if (r->dofr_type == DOF_RELO_SETX)
14059 *(uint64_t *)taddr += ubase;
14060 else
14061 *(uint64_t *)taddr +=
14062 udaddr + ts->dofs_offset + r->dofr_offset;
14063 break;
14064 default:
14065 dtrace_dof_error(dof, "invalid relocation type");
14066 return (-1);
14067 }
14068
14069 r = (dof_relodesc_t *)((uintptr_t)r + rs->dofs_entsize);
14070 }
14071
14072 return (0);
14073 }
14074
14075 /*
14076 * The dof_hdr_t passed to dtrace_dof_slurp() should be a partially validated
14077 * header: it should be at the front of a memory region that is at least
14078 * sizeof (dof_hdr_t) in size -- and then at least dof_hdr.dofh_loadsz in
14079 * size. It need not be validated in any other way.
14080 */
14081 static int
14082 dtrace_dof_slurp(dof_hdr_t *dof, dtrace_vstate_t *vstate, cred_t *cr,
14083 dtrace_enabling_t **enabp, uint64_t ubase, uint64_t udaddr, int noprobes)
14084 {
14085 uint64_t len = dof->dofh_loadsz, seclen;
14086 uintptr_t daddr = (uintptr_t)dof;
14087 dtrace_ecbdesc_t *ep;
14088 dtrace_enabling_t *enab;
14089 uint_t i;
14090
14091 ASSERT(MUTEX_HELD(&dtrace_lock));
14092 ASSERT(dof->dofh_loadsz >= sizeof (dof_hdr_t));
14093
14094 /*
14095 * Check the DOF header identification bytes. In addition to checking
14096 * valid settings, we also verify that unused bits/bytes are zeroed so
14097 * we can use them later without fear of regressing existing binaries.
14098 */
14099 if (bcmp(&dof->dofh_ident[DOF_ID_MAG0],
14100 DOF_MAG_STRING, DOF_MAG_STRLEN) != 0) {
14101 dtrace_dof_error(dof, "DOF magic string mismatch");
14102 return (-1);
14103 }
14104
14105 if (dof->dofh_ident[DOF_ID_MODEL] != DOF_MODEL_ILP32 &&
14106 dof->dofh_ident[DOF_ID_MODEL] != DOF_MODEL_LP64) {
14107 dtrace_dof_error(dof, "DOF has invalid data model");
14108 return (-1);
14109 }
14110
14111 if (dof->dofh_ident[DOF_ID_ENCODING] != DOF_ENCODE_NATIVE) {
14112 dtrace_dof_error(dof, "DOF encoding mismatch");
14113 return (-1);
14114 }
14115
14116 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
14117 dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_2) {
14118 dtrace_dof_error(dof, "DOF version mismatch");
14119 return (-1);
14120 }
14121
14122 if (dof->dofh_ident[DOF_ID_DIFVERS] != DIF_VERSION_2) {
14123 dtrace_dof_error(dof, "DOF uses unsupported instruction set");
14124 return (-1);
14125 }
14126
14127 if (dof->dofh_ident[DOF_ID_DIFIREG] > DIF_DIR_NREGS) {
14128 dtrace_dof_error(dof, "DOF uses too many integer registers");
14129 return (-1);
14130 }
14131
14132 if (dof->dofh_ident[DOF_ID_DIFTREG] > DIF_DTR_NREGS) {
14133 dtrace_dof_error(dof, "DOF uses too many tuple registers");
14134 return (-1);
14135 }
14136
14137 for (i = DOF_ID_PAD; i < DOF_ID_SIZE; i++) {
14138 if (dof->dofh_ident[i] != 0) {
14139 dtrace_dof_error(dof, "DOF has invalid ident byte set");
14140 return (-1);
14141 }
14142 }
14143
14144 if (dof->dofh_flags & ~DOF_FL_VALID) {
14145 dtrace_dof_error(dof, "DOF has invalid flag bits set");
14146 return (-1);
14147 }
14148
14149 if (dof->dofh_secsize == 0) {
14150 dtrace_dof_error(dof, "zero section header size");
14151 return (-1);
14152 }
14153
14154 /*
14155 * Check that the section headers don't exceed the amount of DOF
14156 * data. Note that we cast the section size and number of sections
14157 * to uint64_t's to prevent possible overflow in the multiplication.
14158 */
14159 seclen = (uint64_t)dof->dofh_secnum * (uint64_t)dof->dofh_secsize;
14160
14161 if (dof->dofh_secoff > len || seclen > len ||
14162 dof->dofh_secoff + seclen > len) {
14163 dtrace_dof_error(dof, "truncated section headers");
14164 return (-1);
14165 }
14166
14167 if (!IS_P2ALIGNED(dof->dofh_secoff, sizeof (uint64_t))) {
14168 dtrace_dof_error(dof, "misaligned section headers");
14169 return (-1);
14170 }
14171
14172 if (!IS_P2ALIGNED(dof->dofh_secsize, sizeof (uint64_t))) {
14173 dtrace_dof_error(dof, "misaligned section size");
14174 return (-1);
14175 }
14176
14177 /*
14178 * Take an initial pass through the section headers to be sure that
14179 * the headers don't have stray offsets. If the 'noprobes' flag is
14180 * set, do not permit sections relating to providers, probes, or args.
14181 */
14182 for (i = 0; i < dof->dofh_secnum; i++) {
14183 dof_sec_t *sec = (dof_sec_t *)(daddr +
14184 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
14185
14186 if (noprobes) {
14187 switch (sec->dofs_type) {
14188 case DOF_SECT_PROVIDER:
14189 case DOF_SECT_PROBES:
14190 case DOF_SECT_PRARGS:
14191 case DOF_SECT_PROFFS:
14192 dtrace_dof_error(dof, "illegal sections "
14193 "for enabling");
14194 return (-1);
14195 }
14196 }
14197
14198 if (DOF_SEC_ISLOADABLE(sec->dofs_type) &&
14199 !(sec->dofs_flags & DOF_SECF_LOAD)) {
14200 dtrace_dof_error(dof, "loadable section with load "
14201 "flag unset");
14202 return (-1);
14203 }
14204
14205 if (!(sec->dofs_flags & DOF_SECF_LOAD))
14206 continue; /* just ignore non-loadable sections */
14207
14208 if (!ISP2(sec->dofs_align)) {
14209 dtrace_dof_error(dof, "bad section alignment");
14210 return (-1);
14211 }
14212
14213 if (sec->dofs_offset & (sec->dofs_align - 1)) {
14214 dtrace_dof_error(dof, "misaligned section");
14215 return (-1);
14216 }
14217
14218 if (sec->dofs_offset > len || sec->dofs_size > len ||
14219 sec->dofs_offset + sec->dofs_size > len) {
14220 dtrace_dof_error(dof, "corrupt section header");
14221 return (-1);
14222 }
14223
14224 if (sec->dofs_type == DOF_SECT_STRTAB && *((char *)daddr +
14225 sec->dofs_offset + sec->dofs_size - 1) != '\0') {
14226 dtrace_dof_error(dof, "non-terminating string table");
14227 return (-1);
14228 }
14229 }
14230
14231 /*
14232 * Take a second pass through the sections and locate and perform any
14233 * relocations that are present. We do this after the first pass to
14234 * be sure that all sections have had their headers validated.
14235 */
14236 for (i = 0; i < dof->dofh_secnum; i++) {
14237 dof_sec_t *sec = (dof_sec_t *)(daddr +
14238 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
14239
14240 if (!(sec->dofs_flags & DOF_SECF_LOAD))
14241 continue; /* skip sections that are not loadable */
14242
14243 switch (sec->dofs_type) {
14244 case DOF_SECT_URELHDR:
14245 if (dtrace_dof_relocate(dof, sec, ubase, udaddr) != 0)
14246 return (-1);
14247 break;
14248 }
14249 }
14250
14251 if ((enab = *enabp) == NULL)
14252 enab = *enabp = dtrace_enabling_create(vstate);
14253
14254 for (i = 0; i < dof->dofh_secnum; i++) {
14255 dof_sec_t *sec = (dof_sec_t *)(daddr +
14256 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
14257
14258 if (sec->dofs_type != DOF_SECT_ECBDESC)
14259 continue;
14260
14261 if ((ep = dtrace_dof_ecbdesc(dof, sec, vstate, cr)) == NULL) {
14262 dtrace_enabling_destroy(enab);
14263 *enabp = NULL;
14264 return (-1);
14265 }
14266
14267 dtrace_enabling_add(enab, ep);
14268 }
14269
14270 return (0);
14271 }
14272
14273 /*
14274 * Process DOF for any options. This routine assumes that the DOF has been
14275 * at least processed by dtrace_dof_slurp().
14276 */
14277 static int
14278 dtrace_dof_options(dof_hdr_t *dof, dtrace_state_t *state)
14279 {
14280 int i, rval;
14281 uint32_t entsize;
14282 size_t offs;
14283 dof_optdesc_t *desc;
14284
14285 for (i = 0; i < dof->dofh_secnum; i++) {
14286 dof_sec_t *sec = (dof_sec_t *)((uintptr_t)dof +
14287 (uintptr_t)dof->dofh_secoff + i * dof->dofh_secsize);
14288
14289 if (sec->dofs_type != DOF_SECT_OPTDESC)
14290 continue;
14291
14292 if (sec->dofs_align != sizeof (uint64_t)) {
14293 dtrace_dof_error(dof, "bad alignment in "
14294 "option description");
14295 return (EINVAL);
14296 }
14297
14298 if ((entsize = sec->dofs_entsize) == 0) {
14299 dtrace_dof_error(dof, "zeroed option entry size");
14300 return (EINVAL);
14301 }
14302
14303 if (entsize < sizeof (dof_optdesc_t)) {
14304 dtrace_dof_error(dof, "bad option entry size");
14305 return (EINVAL);
14306 }
14307
14308 for (offs = 0; offs < sec->dofs_size; offs += entsize) {
14309 desc = (dof_optdesc_t *)((uintptr_t)dof +
14310 (uintptr_t)sec->dofs_offset + offs);
14311
14312 if (desc->dofo_strtab != DOF_SECIDX_NONE) {
14313 dtrace_dof_error(dof, "non-zero option string");
14314 return (EINVAL);
14315 }
14316
14317 if (desc->dofo_value == DTRACEOPT_UNSET) {
14318 dtrace_dof_error(dof, "unset option");
14319 return (EINVAL);
14320 }
14321
14322 if ((rval = dtrace_state_option(state,
14323 desc->dofo_option, desc->dofo_value)) != 0) {
14324 dtrace_dof_error(dof, "rejected option");
14325 return (rval);
14326 }
14327 }
14328 }
14329
14330 return (0);
14331 }
14332
14333 /*
14334 * DTrace Consumer State Functions
14335 */
14336 static int
14337 dtrace_dstate_init(dtrace_dstate_t *dstate, size_t size)
14338 {
14339 size_t hashsize, maxper, min, chunksize = dstate->dtds_chunksize;
14340 void *base;
14341 uintptr_t limit;
14342 dtrace_dynvar_t *dvar, *next, *start;
14343 int i;
14344
14345 ASSERT(MUTEX_HELD(&dtrace_lock));
14346 ASSERT(dstate->dtds_base == NULL && dstate->dtds_percpu == NULL);
14347
14348 bzero(dstate, sizeof (dtrace_dstate_t));
14349
14350 if ((dstate->dtds_chunksize = chunksize) == 0)
14351 dstate->dtds_chunksize = DTRACE_DYNVAR_CHUNKSIZE;
14352
14353 VERIFY(dstate->dtds_chunksize < LONG_MAX);
14354
14355 if (size < (min = dstate->dtds_chunksize + sizeof (dtrace_dynhash_t)))
14356 size = min;
14357
14358 if ((base = kmem_zalloc(size, KM_NOSLEEP | KM_NORMALPRI)) == NULL)
14359 return (ENOMEM);
14360
14361 dstate->dtds_size = size;
14362 dstate->dtds_base = base;
14363 dstate->dtds_percpu = kmem_cache_alloc(dtrace_state_cache, KM_SLEEP);
14364 bzero(dstate->dtds_percpu, NCPU * sizeof (dtrace_dstate_percpu_t));
14365
14366 hashsize = size / (dstate->dtds_chunksize + sizeof (dtrace_dynhash_t));
14367
14368 if (hashsize != 1 && (hashsize & 1))
14369 hashsize--;
14370
14371 dstate->dtds_hashsize = hashsize;
14372 dstate->dtds_hash = dstate->dtds_base;
14373
14374 /*
14375 * Set all of our hash buckets to point to the single sink, and (if
14376 * it hasn't already been set), set the sink's hash value to be the
14377 * sink sentinel value. The sink is needed for dynamic variable
14378 * lookups to know that they have iterated over an entire, valid hash
14379 * chain.
14380 */
14381 for (i = 0; i < hashsize; i++)
14382 dstate->dtds_hash[i].dtdh_chain = &dtrace_dynhash_sink;
14383
14384 if (dtrace_dynhash_sink.dtdv_hashval != DTRACE_DYNHASH_SINK)
14385 dtrace_dynhash_sink.dtdv_hashval = DTRACE_DYNHASH_SINK;
14386
14387 /*
14388 * Determine number of active CPUs. Divide free list evenly among
14389 * active CPUs.
14390 */
14391 start = (dtrace_dynvar_t *)
14392 ((uintptr_t)base + hashsize * sizeof (dtrace_dynhash_t));
14393 limit = (uintptr_t)base + size;
14394
14395 VERIFY((uintptr_t)start < limit);
14396 VERIFY((uintptr_t)start >= (uintptr_t)base);
14397
14398 maxper = (limit - (uintptr_t)start) / NCPU;
14399 maxper = (maxper / dstate->dtds_chunksize) * dstate->dtds_chunksize;
14400
14401 #ifndef illumos
14402 CPU_FOREACH(i) {
14403 #else
14404 for (i = 0; i < NCPU; i++) {
14405 #endif
14406 dstate->dtds_percpu[i].dtdsc_free = dvar = start;
14407
14408 /*
14409 * If we don't even have enough chunks to make it once through
14410 * NCPUs, we're just going to allocate everything to the first
14411 * CPU. And if we're on the last CPU, we're going to allocate
14412 * whatever is left over. In either case, we set the limit to
14413 * be the limit of the dynamic variable space.
14414 */
14415 if (maxper == 0 || i == NCPU - 1) {
14416 limit = (uintptr_t)base + size;
14417 start = NULL;
14418 } else {
14419 limit = (uintptr_t)start + maxper;
14420 start = (dtrace_dynvar_t *)limit;
14421 }
14422
14423 VERIFY(limit <= (uintptr_t)base + size);
14424
14425 for (;;) {
14426 next = (dtrace_dynvar_t *)((uintptr_t)dvar +
14427 dstate->dtds_chunksize);
14428
14429 if ((uintptr_t)next + dstate->dtds_chunksize >= limit)
14430 break;
14431
14432 VERIFY((uintptr_t)dvar >= (uintptr_t)base &&
14433 (uintptr_t)dvar <= (uintptr_t)base + size);
14434 dvar->dtdv_next = next;
14435 dvar = next;
14436 }
14437
14438 if (maxper == 0)
14439 break;
14440 }
14441
14442 return (0);
14443 }
14444
14445 static void
14446 dtrace_dstate_fini(dtrace_dstate_t *dstate)
14447 {
14448 ASSERT(MUTEX_HELD(&cpu_lock));
14449
14450 if (dstate->dtds_base == NULL)
14451 return;
14452
14453 kmem_free(dstate->dtds_base, dstate->dtds_size);
14454 kmem_cache_free(dtrace_state_cache, dstate->dtds_percpu);
14455 }
14456
14457 static void
14458 dtrace_vstate_fini(dtrace_vstate_t *vstate)
14459 {
14460 /*
14461 * Logical XOR, where are you?
14462 */
14463 ASSERT((vstate->dtvs_nglobals == 0) ^ (vstate->dtvs_globals != NULL));
14464
14465 if (vstate->dtvs_nglobals > 0) {
14466 kmem_free(vstate->dtvs_globals, vstate->dtvs_nglobals *
14467 sizeof (dtrace_statvar_t *));
14468 }
14469
14470 if (vstate->dtvs_ntlocals > 0) {
14471 kmem_free(vstate->dtvs_tlocals, vstate->dtvs_ntlocals *
14472 sizeof (dtrace_difv_t));
14473 }
14474
14475 ASSERT((vstate->dtvs_nlocals == 0) ^ (vstate->dtvs_locals != NULL));
14476
14477 if (vstate->dtvs_nlocals > 0) {
14478 kmem_free(vstate->dtvs_locals, vstate->dtvs_nlocals *
14479 sizeof (dtrace_statvar_t *));
14480 }
14481 }
14482
14483 #ifdef illumos
14484 static void
14485 dtrace_state_clean(dtrace_state_t *state)
14486 {
14487 if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE)
14488 return;
14489
14490 dtrace_dynvar_clean(&state->dts_vstate.dtvs_dynvars);
14491 dtrace_speculation_clean(state);
14492 }
14493
14494 static void
14495 dtrace_state_deadman(dtrace_state_t *state)
14496 {
14497 hrtime_t now;
14498
14499 dtrace_sync();
14500
14501 now = dtrace_gethrtime();
14502
14503 if (state != dtrace_anon.dta_state &&
14504 now - state->dts_laststatus >= dtrace_deadman_user)
14505 return;
14506
14507 /*
14508 * We must be sure that dts_alive never appears to be less than the
14509 * value upon entry to dtrace_state_deadman(), and because we lack a
14510 * dtrace_cas64(), we cannot store to it atomically. We thus instead
14511 * store INT64_MAX to it, followed by a memory barrier, followed by
14512 * the new value. This assures that dts_alive never appears to be
14513 * less than its true value, regardless of the order in which the
14514 * stores to the underlying storage are issued.
14515 */
14516 state->dts_alive = INT64_MAX;
14517 dtrace_membar_producer();
14518 state->dts_alive = now;
14519 }
14520 #else /* !illumos */
14521 static void
14522 dtrace_state_clean(void *arg)
14523 {
14524 dtrace_state_t *state = arg;
14525 dtrace_optval_t *opt = state->dts_options;
14526
14527 if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE)
14528 return;
14529
14530 dtrace_dynvar_clean(&state->dts_vstate.dtvs_dynvars);
14531 dtrace_speculation_clean(state);
14532
14533 callout_reset(&state->dts_cleaner, hz * opt[DTRACEOPT_CLEANRATE] / NANOSEC,
14534 dtrace_state_clean, state);
14535 }
14536
14537 static void
14538 dtrace_state_deadman(void *arg)
14539 {
14540 dtrace_state_t *state = arg;
14541 hrtime_t now;
14542
14543 dtrace_sync();
14544
14545 dtrace_debug_output();
14546
14547 now = dtrace_gethrtime();
14548
14549 if (state != dtrace_anon.dta_state &&
14550 now - state->dts_laststatus >= dtrace_deadman_user)
14551 return;
14552
14553 /*
14554 * We must be sure that dts_alive never appears to be less than the
14555 * value upon entry to dtrace_state_deadman(), and because we lack a
14556 * dtrace_cas64(), we cannot store to it atomically. We thus instead
14557 * store INT64_MAX to it, followed by a memory barrier, followed by
14558 * the new value. This assures that dts_alive never appears to be
14559 * less than its true value, regardless of the order in which the
14560 * stores to the underlying storage are issued.
14561 */
14562 state->dts_alive = INT64_MAX;
14563 dtrace_membar_producer();
14564 state->dts_alive = now;
14565
14566 callout_reset(&state->dts_deadman, hz * dtrace_deadman_interval / NANOSEC,
14567 dtrace_state_deadman, state);
14568 }
14569 #endif /* illumos */
14570
14571 static dtrace_state_t *
14572 #ifdef illumos
14573 dtrace_state_create(dev_t *devp, cred_t *cr)
14574 #else
14575 dtrace_state_create(struct cdev *dev, struct ucred *cred __unused)
14576 #endif
14577 {
14578 #ifdef illumos
14579 minor_t minor;
14580 major_t major;
14581 #else
14582 cred_t *cr = NULL;
14583 int m = 0;
14584 #endif
14585 char c[30];
14586 dtrace_state_t *state;
14587 dtrace_optval_t *opt;
14588 int bufsize = NCPU * sizeof (dtrace_buffer_t), i;
14589 int cpu_it;
14590
14591 ASSERT(MUTEX_HELD(&dtrace_lock));
14592 ASSERT(MUTEX_HELD(&cpu_lock));
14593
14594 #ifdef illumos
14595 minor = (minor_t)(uintptr_t)vmem_alloc(dtrace_minor, 1,
14596 VM_BESTFIT | VM_SLEEP);
14597
14598 if (ddi_soft_state_zalloc(dtrace_softstate, minor) != DDI_SUCCESS) {
14599 vmem_free(dtrace_minor, (void *)(uintptr_t)minor, 1);
14600 return (NULL);
14601 }
14602
14603 state = ddi_get_soft_state(dtrace_softstate, minor);
14604 #else
14605 if (dev != NULL) {
14606 cr = dev->si_cred;
14607 m = dev2unit(dev);
14608 }
14609
14610 /* Allocate memory for the state. */
14611 state = kmem_zalloc(sizeof(dtrace_state_t), KM_SLEEP);
14612 #endif
14613
14614 state->dts_epid = DTRACE_EPIDNONE + 1;
14615
14616 (void) snprintf(c, sizeof (c), "dtrace_aggid_%d", m);
14617 #ifdef illumos
14618 state->dts_aggid_arena = vmem_create(c, (void *)1, UINT32_MAX, 1,
14619 NULL, NULL, NULL, 0, VM_SLEEP | VMC_IDENTIFIER);
14620
14621 if (devp != NULL) {
14622 major = getemajor(*devp);
14623 } else {
14624 major = ddi_driver_major(dtrace_devi);
14625 }
14626
14627 state->dts_dev = makedevice(major, minor);
14628
14629 if (devp != NULL)
14630 *devp = state->dts_dev;
14631 #else
14632 state->dts_aggid_arena = new_unrhdr(1, INT_MAX, &dtrace_unr_mtx);
14633 state->dts_dev = dev;
14634 #endif
14635
14636 /*
14637 * We allocate NCPU buffers. On the one hand, this can be quite
14638 * a bit of memory per instance (nearly 36K on a Starcat). On the
14639 * other hand, it saves an additional memory reference in the probe
14640 * path.
14641 */
14642 state->dts_buffer = kmem_zalloc(bufsize, KM_SLEEP);
14643 state->dts_aggbuffer = kmem_zalloc(bufsize, KM_SLEEP);
14644
14645 /*
14646 * Allocate and initialise the per-process per-CPU random state.
14647 * SI_SUB_RANDOM < SI_SUB_DTRACE_ANON therefore entropy device is
14648 * assumed to be seeded at this point (if from Fortuna seed file).
14649 */
14650 arc4random_buf(&state->dts_rstate[0], 2 * sizeof(uint64_t));
14651 for (cpu_it = 1; cpu_it < NCPU; cpu_it++) {
14652 /*
14653 * Each CPU is assigned a 2^64 period, non-overlapping
14654 * subsequence.
14655 */
14656 dtrace_xoroshiro128_plus_jump(state->dts_rstate[cpu_it-1],
14657 state->dts_rstate[cpu_it]);
14658 }
14659
14660 #ifdef illumos
14661 state->dts_cleaner = CYCLIC_NONE;
14662 state->dts_deadman = CYCLIC_NONE;
14663 #else
14664 callout_init(&state->dts_cleaner, 1);
14665 callout_init(&state->dts_deadman, 1);
14666 #endif
14667 state->dts_vstate.dtvs_state = state;
14668
14669 for (i = 0; i < DTRACEOPT_MAX; i++)
14670 state->dts_options[i] = DTRACEOPT_UNSET;
14671
14672 /*
14673 * Set the default options.
14674 */
14675 opt = state->dts_options;
14676 opt[DTRACEOPT_BUFPOLICY] = DTRACEOPT_BUFPOLICY_SWITCH;
14677 opt[DTRACEOPT_BUFRESIZE] = DTRACEOPT_BUFRESIZE_AUTO;
14678 opt[DTRACEOPT_NSPEC] = dtrace_nspec_default;
14679 opt[DTRACEOPT_SPECSIZE] = dtrace_specsize_default;
14680 opt[DTRACEOPT_CPU] = (dtrace_optval_t)DTRACE_CPUALL;
14681 opt[DTRACEOPT_STRSIZE] = dtrace_strsize_default;
14682 opt[DTRACEOPT_STACKFRAMES] = dtrace_stackframes_default;
14683 opt[DTRACEOPT_USTACKFRAMES] = dtrace_ustackframes_default;
14684 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_default;
14685 opt[DTRACEOPT_AGGRATE] = dtrace_aggrate_default;
14686 opt[DTRACEOPT_SWITCHRATE] = dtrace_switchrate_default;
14687 opt[DTRACEOPT_STATUSRATE] = dtrace_statusrate_default;
14688 opt[DTRACEOPT_JSTACKFRAMES] = dtrace_jstackframes_default;
14689 opt[DTRACEOPT_JSTACKSTRSIZE] = dtrace_jstackstrsize_default;
14690
14691 state->dts_activity = DTRACE_ACTIVITY_INACTIVE;
14692
14693 /*
14694 * Depending on the user credentials, we set flag bits which alter probe
14695 * visibility or the amount of destructiveness allowed. In the case of
14696 * actual anonymous tracing, or the possession of all privileges, all of
14697 * the normal checks are bypassed.
14698 */
14699 if (cr == NULL || PRIV_POLICY_ONLY(cr, PRIV_ALL, B_FALSE)) {
14700 state->dts_cred.dcr_visible = DTRACE_CRV_ALL;
14701 state->dts_cred.dcr_action = DTRACE_CRA_ALL;
14702 } else {
14703 /*
14704 * Set up the credentials for this instantiation. We take a
14705 * hold on the credential to prevent it from disappearing on
14706 * us; this in turn prevents the zone_t referenced by this
14707 * credential from disappearing. This means that we can
14708 * examine the credential and the zone from probe context.
14709 */
14710 crhold(cr);
14711 state->dts_cred.dcr_cred = cr;
14712
14713 /*
14714 * CRA_PROC means "we have *some* privilege for dtrace" and
14715 * unlocks the use of variables like pid, zonename, etc.
14716 */
14717 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE) ||
14718 PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE)) {
14719 state->dts_cred.dcr_action |= DTRACE_CRA_PROC;
14720 }
14721
14722 /*
14723 * dtrace_user allows use of syscall and profile providers.
14724 * If the user also has proc_owner and/or proc_zone, we
14725 * extend the scope to include additional visibility and
14726 * destructive power.
14727 */
14728 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_USER, B_FALSE)) {
14729 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE)) {
14730 state->dts_cred.dcr_visible |=
14731 DTRACE_CRV_ALLPROC;
14732
14733 state->dts_cred.dcr_action |=
14734 DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
14735 }
14736
14737 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE)) {
14738 state->dts_cred.dcr_visible |=
14739 DTRACE_CRV_ALLZONE;
14740
14741 state->dts_cred.dcr_action |=
14742 DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
14743 }
14744
14745 /*
14746 * If we have all privs in whatever zone this is,
14747 * we can do destructive things to processes which
14748 * have altered credentials.
14749 */
14750 #ifdef illumos
14751 if (priv_isequalset(priv_getset(cr, PRIV_EFFECTIVE),
14752 cr->cr_zone->zone_privset)) {
14753 state->dts_cred.dcr_action |=
14754 DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG;
14755 }
14756 #endif
14757 }
14758
14759 /*
14760 * Holding the dtrace_kernel privilege also implies that
14761 * the user has the dtrace_user privilege from a visibility
14762 * perspective. But without further privileges, some
14763 * destructive actions are not available.
14764 */
14765 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_KERNEL, B_FALSE)) {
14766 /*
14767 * Make all probes in all zones visible. However,
14768 * this doesn't mean that all actions become available
14769 * to all zones.
14770 */
14771 state->dts_cred.dcr_visible |= DTRACE_CRV_KERNEL |
14772 DTRACE_CRV_ALLPROC | DTRACE_CRV_ALLZONE;
14773
14774 state->dts_cred.dcr_action |= DTRACE_CRA_KERNEL |
14775 DTRACE_CRA_PROC;
14776 /*
14777 * Holding proc_owner means that destructive actions
14778 * for *this* zone are allowed.
14779 */
14780 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
14781 state->dts_cred.dcr_action |=
14782 DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
14783
14784 /*
14785 * Holding proc_zone means that destructive actions
14786 * for this user/group ID in all zones is allowed.
14787 */
14788 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
14789 state->dts_cred.dcr_action |=
14790 DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
14791
14792 #ifdef illumos
14793 /*
14794 * If we have all privs in whatever zone this is,
14795 * we can do destructive things to processes which
14796 * have altered credentials.
14797 */
14798 if (priv_isequalset(priv_getset(cr, PRIV_EFFECTIVE),
14799 cr->cr_zone->zone_privset)) {
14800 state->dts_cred.dcr_action |=
14801 DTRACE_CRA_PROC_DESTRUCTIVE_CREDCHG;
14802 }
14803 #endif
14804 }
14805
14806 /*
14807 * Holding the dtrace_proc privilege gives control over fasttrap
14808 * and pid providers. We need to grant wider destructive
14809 * privileges in the event that the user has proc_owner and/or
14810 * proc_zone.
14811 */
14812 if (PRIV_POLICY_ONLY(cr, PRIV_DTRACE_PROC, B_FALSE)) {
14813 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_OWNER, B_FALSE))
14814 state->dts_cred.dcr_action |=
14815 DTRACE_CRA_PROC_DESTRUCTIVE_ALLUSER;
14816
14817 if (PRIV_POLICY_ONLY(cr, PRIV_PROC_ZONE, B_FALSE))
14818 state->dts_cred.dcr_action |=
14819 DTRACE_CRA_PROC_DESTRUCTIVE_ALLZONE;
14820 }
14821 }
14822
14823 return (state);
14824 }
14825
14826 static int
14827 dtrace_state_buffer(dtrace_state_t *state, dtrace_buffer_t *buf, int which)
14828 {
14829 dtrace_optval_t *opt = state->dts_options, size;
14830 processorid_t cpu = 0;
14831 int flags = 0, rval, factor, divisor = 1;
14832
14833 ASSERT(MUTEX_HELD(&dtrace_lock));
14834 ASSERT(MUTEX_HELD(&cpu_lock));
14835 ASSERT(which < DTRACEOPT_MAX);
14836 ASSERT(state->dts_activity == DTRACE_ACTIVITY_INACTIVE ||
14837 (state == dtrace_anon.dta_state &&
14838 state->dts_activity == DTRACE_ACTIVITY_ACTIVE));
14839
14840 if (opt[which] == DTRACEOPT_UNSET || opt[which] == 0)
14841 return (0);
14842
14843 if (opt[DTRACEOPT_CPU] != DTRACEOPT_UNSET)
14844 cpu = opt[DTRACEOPT_CPU];
14845
14846 if (which == DTRACEOPT_SPECSIZE)
14847 flags |= DTRACEBUF_NOSWITCH;
14848
14849 if (which == DTRACEOPT_BUFSIZE) {
14850 if (opt[DTRACEOPT_BUFPOLICY] == DTRACEOPT_BUFPOLICY_RING)
14851 flags |= DTRACEBUF_RING;
14852
14853 if (opt[DTRACEOPT_BUFPOLICY] == DTRACEOPT_BUFPOLICY_FILL)
14854 flags |= DTRACEBUF_FILL;
14855
14856 if (state != dtrace_anon.dta_state ||
14857 state->dts_activity != DTRACE_ACTIVITY_ACTIVE)
14858 flags |= DTRACEBUF_INACTIVE;
14859 }
14860
14861 for (size = opt[which]; size >= sizeof (uint64_t); size /= divisor) {
14862 /*
14863 * The size must be 8-byte aligned. If the size is not 8-byte
14864 * aligned, drop it down by the difference.
14865 */
14866 if (size & (sizeof (uint64_t) - 1))
14867 size -= size & (sizeof (uint64_t) - 1);
14868
14869 if (size < state->dts_reserve) {
14870 /*
14871 * Buffers always must be large enough to accommodate
14872 * their prereserved space. We return E2BIG instead
14873 * of ENOMEM in this case to allow for user-level
14874 * software to differentiate the cases.
14875 */
14876 return (E2BIG);
14877 }
14878
14879 rval = dtrace_buffer_alloc(buf, size, flags, cpu, &factor);
14880
14881 if (rval != ENOMEM) {
14882 opt[which] = size;
14883 return (rval);
14884 }
14885
14886 if (opt[DTRACEOPT_BUFRESIZE] == DTRACEOPT_BUFRESIZE_MANUAL)
14887 return (rval);
14888
14889 for (divisor = 2; divisor < factor; divisor <<= 1)
14890 continue;
14891 }
14892
14893 return (ENOMEM);
14894 }
14895
14896 static int
14897 dtrace_state_buffers(dtrace_state_t *state)
14898 {
14899 dtrace_speculation_t *spec = state->dts_speculations;
14900 int rval, i;
14901
14902 if ((rval = dtrace_state_buffer(state, state->dts_buffer,
14903 DTRACEOPT_BUFSIZE)) != 0)
14904 return (rval);
14905
14906 if ((rval = dtrace_state_buffer(state, state->dts_aggbuffer,
14907 DTRACEOPT_AGGSIZE)) != 0)
14908 return (rval);
14909
14910 for (i = 0; i < state->dts_nspeculations; i++) {
14911 if ((rval = dtrace_state_buffer(state,
14912 spec[i].dtsp_buffer, DTRACEOPT_SPECSIZE)) != 0)
14913 return (rval);
14914 }
14915
14916 return (0);
14917 }
14918
14919 static void
14920 dtrace_state_prereserve(dtrace_state_t *state)
14921 {
14922 dtrace_ecb_t *ecb;
14923 dtrace_probe_t *probe;
14924
14925 state->dts_reserve = 0;
14926
14927 if (state->dts_options[DTRACEOPT_BUFPOLICY] != DTRACEOPT_BUFPOLICY_FILL)
14928 return;
14929
14930 /*
14931 * If our buffer policy is a "fill" buffer policy, we need to set the
14932 * prereserved space to be the space required by the END probes.
14933 */
14934 probe = dtrace_probes[dtrace_probeid_end - 1];
14935 ASSERT(probe != NULL);
14936
14937 for (ecb = probe->dtpr_ecb; ecb != NULL; ecb = ecb->dte_next) {
14938 if (ecb->dte_state != state)
14939 continue;
14940
14941 state->dts_reserve += ecb->dte_needed + ecb->dte_alignment;
14942 }
14943 }
14944
14945 static int
14946 dtrace_state_go(dtrace_state_t *state, processorid_t *cpu)
14947 {
14948 dtrace_optval_t *opt = state->dts_options, sz, nspec;
14949 dtrace_speculation_t *spec;
14950 dtrace_buffer_t *buf;
14951 #ifdef illumos
14952 cyc_handler_t hdlr;
14953 cyc_time_t when;
14954 #endif
14955 int rval = 0, i, bufsize = NCPU * sizeof (dtrace_buffer_t);
14956 dtrace_icookie_t cookie;
14957
14958 mutex_enter(&cpu_lock);
14959 mutex_enter(&dtrace_lock);
14960
14961 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) {
14962 rval = EBUSY;
14963 goto out;
14964 }
14965
14966 /*
14967 * Before we can perform any checks, we must prime all of the
14968 * retained enablings that correspond to this state.
14969 */
14970 dtrace_enabling_prime(state);
14971
14972 if (state->dts_destructive && !state->dts_cred.dcr_destructive) {
14973 rval = EACCES;
14974 goto out;
14975 }
14976
14977 dtrace_state_prereserve(state);
14978
14979 /*
14980 * Now we want to do is try to allocate our speculations.
14981 * We do not automatically resize the number of speculations; if
14982 * this fails, we will fail the operation.
14983 */
14984 nspec = opt[DTRACEOPT_NSPEC];
14985 ASSERT(nspec != DTRACEOPT_UNSET);
14986
14987 if (nspec > INT_MAX) {
14988 rval = ENOMEM;
14989 goto out;
14990 }
14991
14992 spec = kmem_zalloc(nspec * sizeof (dtrace_speculation_t),
14993 KM_NOSLEEP | KM_NORMALPRI);
14994
14995 if (spec == NULL) {
14996 rval = ENOMEM;
14997 goto out;
14998 }
14999
15000 state->dts_speculations = spec;
15001 state->dts_nspeculations = (int)nspec;
15002
15003 for (i = 0; i < nspec; i++) {
15004 if ((buf = kmem_zalloc(bufsize,
15005 KM_NOSLEEP | KM_NORMALPRI)) == NULL) {
15006 rval = ENOMEM;
15007 goto err;
15008 }
15009
15010 spec[i].dtsp_buffer = buf;
15011 }
15012
15013 if (opt[DTRACEOPT_GRABANON] != DTRACEOPT_UNSET) {
15014 if (dtrace_anon.dta_state == NULL) {
15015 rval = ENOENT;
15016 goto out;
15017 }
15018
15019 if (state->dts_necbs != 0) {
15020 rval = EALREADY;
15021 goto out;
15022 }
15023
15024 state->dts_anon = dtrace_anon_grab();
15025 ASSERT(state->dts_anon != NULL);
15026 state = state->dts_anon;
15027
15028 /*
15029 * We want "grabanon" to be set in the grabbed state, so we'll
15030 * copy that option value from the grabbing state into the
15031 * grabbed state.
15032 */
15033 state->dts_options[DTRACEOPT_GRABANON] =
15034 opt[DTRACEOPT_GRABANON];
15035
15036 *cpu = dtrace_anon.dta_beganon;
15037
15038 /*
15039 * If the anonymous state is active (as it almost certainly
15040 * is if the anonymous enabling ultimately matched anything),
15041 * we don't allow any further option processing -- but we
15042 * don't return failure.
15043 */
15044 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
15045 goto out;
15046 }
15047
15048 if (opt[DTRACEOPT_AGGSIZE] != DTRACEOPT_UNSET &&
15049 opt[DTRACEOPT_AGGSIZE] != 0) {
15050 if (state->dts_aggregations == NULL) {
15051 /*
15052 * We're not going to create an aggregation buffer
15053 * because we don't have any ECBs that contain
15054 * aggregations -- set this option to 0.
15055 */
15056 opt[DTRACEOPT_AGGSIZE] = 0;
15057 } else {
15058 /*
15059 * If we have an aggregation buffer, we must also have
15060 * a buffer to use as scratch.
15061 */
15062 if (opt[DTRACEOPT_BUFSIZE] == DTRACEOPT_UNSET ||
15063 opt[DTRACEOPT_BUFSIZE] < state->dts_needed) {
15064 opt[DTRACEOPT_BUFSIZE] = state->dts_needed;
15065 }
15066 }
15067 }
15068
15069 if (opt[DTRACEOPT_SPECSIZE] != DTRACEOPT_UNSET &&
15070 opt[DTRACEOPT_SPECSIZE] != 0) {
15071 if (!state->dts_speculates) {
15072 /*
15073 * We're not going to create speculation buffers
15074 * because we don't have any ECBs that actually
15075 * speculate -- set the speculation size to 0.
15076 */
15077 opt[DTRACEOPT_SPECSIZE] = 0;
15078 }
15079 }
15080
15081 /*
15082 * The bare minimum size for any buffer that we're actually going to
15083 * do anything to is sizeof (uint64_t).
15084 */
15085 sz = sizeof (uint64_t);
15086
15087 if ((state->dts_needed != 0 && opt[DTRACEOPT_BUFSIZE] < sz) ||
15088 (state->dts_speculates && opt[DTRACEOPT_SPECSIZE] < sz) ||
15089 (state->dts_aggregations != NULL && opt[DTRACEOPT_AGGSIZE] < sz)) {
15090 /*
15091 * A buffer size has been explicitly set to 0 (or to a size
15092 * that will be adjusted to 0) and we need the space -- we
15093 * need to return failure. We return ENOSPC to differentiate
15094 * it from failing to allocate a buffer due to failure to meet
15095 * the reserve (for which we return E2BIG).
15096 */
15097 rval = ENOSPC;
15098 goto out;
15099 }
15100
15101 if ((rval = dtrace_state_buffers(state)) != 0)
15102 goto err;
15103
15104 if ((sz = opt[DTRACEOPT_DYNVARSIZE]) == DTRACEOPT_UNSET)
15105 sz = dtrace_dstate_defsize;
15106
15107 do {
15108 rval = dtrace_dstate_init(&state->dts_vstate.dtvs_dynvars, sz);
15109
15110 if (rval == 0)
15111 break;
15112
15113 if (opt[DTRACEOPT_BUFRESIZE] == DTRACEOPT_BUFRESIZE_MANUAL)
15114 goto err;
15115 } while (sz >>= 1);
15116
15117 opt[DTRACEOPT_DYNVARSIZE] = sz;
15118
15119 if (rval != 0)
15120 goto err;
15121
15122 if (opt[DTRACEOPT_STATUSRATE] > dtrace_statusrate_max)
15123 opt[DTRACEOPT_STATUSRATE] = dtrace_statusrate_max;
15124
15125 if (opt[DTRACEOPT_CLEANRATE] == 0)
15126 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_max;
15127
15128 if (opt[DTRACEOPT_CLEANRATE] < dtrace_cleanrate_min)
15129 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_min;
15130
15131 if (opt[DTRACEOPT_CLEANRATE] > dtrace_cleanrate_max)
15132 opt[DTRACEOPT_CLEANRATE] = dtrace_cleanrate_max;
15133
15134 state->dts_alive = state->dts_laststatus = dtrace_gethrtime();
15135 #ifdef illumos
15136 hdlr.cyh_func = (cyc_func_t)dtrace_state_clean;
15137 hdlr.cyh_arg = state;
15138 hdlr.cyh_level = CY_LOW_LEVEL;
15139
15140 when.cyt_when = 0;
15141 when.cyt_interval = opt[DTRACEOPT_CLEANRATE];
15142
15143 state->dts_cleaner = cyclic_add(&hdlr, &when);
15144
15145 hdlr.cyh_func = (cyc_func_t)dtrace_state_deadman;
15146 hdlr.cyh_arg = state;
15147 hdlr.cyh_level = CY_LOW_LEVEL;
15148
15149 when.cyt_when = 0;
15150 when.cyt_interval = dtrace_deadman_interval;
15151
15152 state->dts_deadman = cyclic_add(&hdlr, &when);
15153 #else
15154 callout_reset(&state->dts_cleaner, hz * opt[DTRACEOPT_CLEANRATE] / NANOSEC,
15155 dtrace_state_clean, state);
15156 callout_reset(&state->dts_deadman, hz * dtrace_deadman_interval / NANOSEC,
15157 dtrace_state_deadman, state);
15158 #endif
15159
15160 state->dts_activity = DTRACE_ACTIVITY_WARMUP;
15161
15162 #ifdef illumos
15163 if (state->dts_getf != 0 &&
15164 !(state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL)) {
15165 /*
15166 * We don't have kernel privs but we have at least one call
15167 * to getf(); we need to bump our zone's count, and (if
15168 * this is the first enabling to have an unprivileged call
15169 * to getf()) we need to hook into closef().
15170 */
15171 state->dts_cred.dcr_cred->cr_zone->zone_dtrace_getf++;
15172
15173 if (dtrace_getf++ == 0) {
15174 ASSERT(dtrace_closef == NULL);
15175 dtrace_closef = dtrace_getf_barrier;
15176 }
15177 }
15178 #endif
15179
15180 /*
15181 * Now it's time to actually fire the BEGIN probe. We need to disable
15182 * interrupts here both to record the CPU on which we fired the BEGIN
15183 * probe (the data from this CPU will be processed first at user
15184 * level) and to manually activate the buffer for this CPU.
15185 */
15186 cookie = dtrace_interrupt_disable();
15187 *cpu = curcpu;
15188 ASSERT(state->dts_buffer[*cpu].dtb_flags & DTRACEBUF_INACTIVE);
15189 state->dts_buffer[*cpu].dtb_flags &= ~DTRACEBUF_INACTIVE;
15190
15191 dtrace_probe(dtrace_probeid_begin,
15192 (uint64_t)(uintptr_t)state, 0, 0, 0, 0);
15193 dtrace_interrupt_enable(cookie);
15194 /*
15195 * We may have had an exit action from a BEGIN probe; only change our
15196 * state to ACTIVE if we're still in WARMUP.
15197 */
15198 ASSERT(state->dts_activity == DTRACE_ACTIVITY_WARMUP ||
15199 state->dts_activity == DTRACE_ACTIVITY_DRAINING);
15200
15201 if (state->dts_activity == DTRACE_ACTIVITY_WARMUP)
15202 state->dts_activity = DTRACE_ACTIVITY_ACTIVE;
15203
15204 #ifdef __FreeBSD__
15205 /*
15206 * We enable anonymous tracing before APs are started, so we must
15207 * activate buffers using the current CPU.
15208 */
15209 if (state == dtrace_anon.dta_state)
15210 for (int i = 0; i < NCPU; i++)
15211 dtrace_buffer_activate_cpu(state, i);
15212 else
15213 dtrace_xcall(DTRACE_CPUALL,
15214 (dtrace_xcall_t)dtrace_buffer_activate, state);
15215 #else
15216 /*
15217 * Regardless of whether or not now we're in ACTIVE or DRAINING, we
15218 * want each CPU to transition its principal buffer out of the
15219 * INACTIVE state. Doing this assures that no CPU will suddenly begin
15220 * processing an ECB halfway down a probe's ECB chain; all CPUs will
15221 * atomically transition from processing none of a state's ECBs to
15222 * processing all of them.
15223 */
15224 dtrace_xcall(DTRACE_CPUALL,
15225 (dtrace_xcall_t)dtrace_buffer_activate, state);
15226 #endif
15227 goto out;
15228
15229 err:
15230 dtrace_buffer_free(state->dts_buffer);
15231 dtrace_buffer_free(state->dts_aggbuffer);
15232
15233 if ((nspec = state->dts_nspeculations) == 0) {
15234 ASSERT(state->dts_speculations == NULL);
15235 goto out;
15236 }
15237
15238 spec = state->dts_speculations;
15239 ASSERT(spec != NULL);
15240
15241 for (i = 0; i < state->dts_nspeculations; i++) {
15242 if ((buf = spec[i].dtsp_buffer) == NULL)
15243 break;
15244
15245 dtrace_buffer_free(buf);
15246 kmem_free(buf, bufsize);
15247 }
15248
15249 kmem_free(spec, nspec * sizeof (dtrace_speculation_t));
15250 state->dts_nspeculations = 0;
15251 state->dts_speculations = NULL;
15252
15253 out:
15254 mutex_exit(&dtrace_lock);
15255 mutex_exit(&cpu_lock);
15256
15257 return (rval);
15258 }
15259
15260 static int
15261 dtrace_state_stop(dtrace_state_t *state, processorid_t *cpu)
15262 {
15263 dtrace_icookie_t cookie;
15264
15265 ASSERT(MUTEX_HELD(&dtrace_lock));
15266
15267 if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE &&
15268 state->dts_activity != DTRACE_ACTIVITY_DRAINING)
15269 return (EINVAL);
15270
15271 /*
15272 * We'll set the activity to DTRACE_ACTIVITY_DRAINING, and issue a sync
15273 * to be sure that every CPU has seen it. See below for the details
15274 * on why this is done.
15275 */
15276 state->dts_activity = DTRACE_ACTIVITY_DRAINING;
15277 dtrace_sync();
15278
15279 /*
15280 * By this point, it is impossible for any CPU to be still processing
15281 * with DTRACE_ACTIVITY_ACTIVE. We can thus set our activity to
15282 * DTRACE_ACTIVITY_COOLDOWN and know that we're not racing with any
15283 * other CPU in dtrace_buffer_reserve(). This allows dtrace_probe()
15284 * and callees to know that the activity is DTRACE_ACTIVITY_COOLDOWN
15285 * iff we're in the END probe.
15286 */
15287 state->dts_activity = DTRACE_ACTIVITY_COOLDOWN;
15288 dtrace_sync();
15289 ASSERT(state->dts_activity == DTRACE_ACTIVITY_COOLDOWN);
15290
15291 /*
15292 * Finally, we can release the reserve and call the END probe. We
15293 * disable interrupts across calling the END probe to allow us to
15294 * return the CPU on which we actually called the END probe. This
15295 * allows user-land to be sure that this CPU's principal buffer is
15296 * processed last.
15297 */
15298 state->dts_reserve = 0;
15299
15300 cookie = dtrace_interrupt_disable();
15301 *cpu = curcpu;
15302 dtrace_probe(dtrace_probeid_end,
15303 (uint64_t)(uintptr_t)state, 0, 0, 0, 0);
15304 dtrace_interrupt_enable(cookie);
15305
15306 state->dts_activity = DTRACE_ACTIVITY_STOPPED;
15307 dtrace_sync();
15308
15309 #ifdef illumos
15310 if (state->dts_getf != 0 &&
15311 !(state->dts_cred.dcr_visible & DTRACE_CRV_KERNEL)) {
15312 /*
15313 * We don't have kernel privs but we have at least one call
15314 * to getf(); we need to lower our zone's count, and (if
15315 * this is the last enabling to have an unprivileged call
15316 * to getf()) we need to clear the closef() hook.
15317 */
15318 ASSERT(state->dts_cred.dcr_cred->cr_zone->zone_dtrace_getf > 0);
15319 ASSERT(dtrace_closef == dtrace_getf_barrier);
15320 ASSERT(dtrace_getf > 0);
15321
15322 state->dts_cred.dcr_cred->cr_zone->zone_dtrace_getf--;
15323
15324 if (--dtrace_getf == 0)
15325 dtrace_closef = NULL;
15326 }
15327 #endif
15328
15329 return (0);
15330 }
15331
15332 static int
15333 dtrace_state_option(dtrace_state_t *state, dtrace_optid_t option,
15334 dtrace_optval_t val)
15335 {
15336 ASSERT(MUTEX_HELD(&dtrace_lock));
15337
15338 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE)
15339 return (EBUSY);
15340
15341 if (option >= DTRACEOPT_MAX)
15342 return (EINVAL);
15343
15344 if (option != DTRACEOPT_CPU && val < 0)
15345 return (EINVAL);
15346
15347 switch (option) {
15348 case DTRACEOPT_DESTRUCTIVE:
15349 if (dtrace_destructive_disallow)
15350 return (EACCES);
15351
15352 state->dts_cred.dcr_destructive = 1;
15353 break;
15354
15355 case DTRACEOPT_BUFSIZE:
15356 case DTRACEOPT_DYNVARSIZE:
15357 case DTRACEOPT_AGGSIZE:
15358 case DTRACEOPT_SPECSIZE:
15359 case DTRACEOPT_STRSIZE:
15360 if (val < 0)
15361 return (EINVAL);
15362
15363 if (val >= LONG_MAX) {
15364 /*
15365 * If this is an otherwise negative value, set it to
15366 * the highest multiple of 128m less than LONG_MAX.
15367 * Technically, we're adjusting the size without
15368 * regard to the buffer resizing policy, but in fact,
15369 * this has no effect -- if we set the buffer size to
15370 * ~LONG_MAX and the buffer policy is ultimately set to
15371 * be "manual", the buffer allocation is guaranteed to
15372 * fail, if only because the allocation requires two
15373 * buffers. (We set the the size to the highest
15374 * multiple of 128m because it ensures that the size
15375 * will remain a multiple of a megabyte when
15376 * repeatedly halved -- all the way down to 15m.)
15377 */
15378 val = LONG_MAX - (1 << 27) + 1;
15379 }
15380 }
15381
15382 state->dts_options[option] = val;
15383
15384 return (0);
15385 }
15386
15387 static void
15388 dtrace_state_destroy(dtrace_state_t *state)
15389 {
15390 dtrace_ecb_t *ecb;
15391 dtrace_vstate_t *vstate = &state->dts_vstate;
15392 #ifdef illumos
15393 minor_t minor = getminor(state->dts_dev);
15394 #endif
15395 int i, bufsize = NCPU * sizeof (dtrace_buffer_t);
15396 dtrace_speculation_t *spec = state->dts_speculations;
15397 int nspec = state->dts_nspeculations;
15398 uint32_t match;
15399
15400 ASSERT(MUTEX_HELD(&dtrace_lock));
15401 ASSERT(MUTEX_HELD(&cpu_lock));
15402
15403 /*
15404 * First, retract any retained enablings for this state.
15405 */
15406 dtrace_enabling_retract(state);
15407 ASSERT(state->dts_nretained == 0);
15408
15409 if (state->dts_activity == DTRACE_ACTIVITY_ACTIVE ||
15410 state->dts_activity == DTRACE_ACTIVITY_DRAINING) {
15411 /*
15412 * We have managed to come into dtrace_state_destroy() on a
15413 * hot enabling -- almost certainly because of a disorderly
15414 * shutdown of a consumer. (That is, a consumer that is
15415 * exiting without having called dtrace_stop().) In this case,
15416 * we're going to set our activity to be KILLED, and then
15417 * issue a sync to be sure that everyone is out of probe
15418 * context before we start blowing away ECBs.
15419 */
15420 state->dts_activity = DTRACE_ACTIVITY_KILLED;
15421 dtrace_sync();
15422 }
15423
15424 /*
15425 * Release the credential hold we took in dtrace_state_create().
15426 */
15427 if (state->dts_cred.dcr_cred != NULL)
15428 crfree(state->dts_cred.dcr_cred);
15429
15430 /*
15431 * Now we can safely disable and destroy any enabled probes. Because
15432 * any DTRACE_PRIV_KERNEL probes may actually be slowing our progress
15433 * (especially if they're all enabled), we take two passes through the
15434 * ECBs: in the first, we disable just DTRACE_PRIV_KERNEL probes, and
15435 * in the second we disable whatever is left over.
15436 */
15437 for (match = DTRACE_PRIV_KERNEL; ; match = 0) {
15438 for (i = 0; i < state->dts_necbs; i++) {
15439 if ((ecb = state->dts_ecbs[i]) == NULL)
15440 continue;
15441
15442 if (match && ecb->dte_probe != NULL) {
15443 dtrace_probe_t *probe = ecb->dte_probe;
15444 dtrace_provider_t *prov = probe->dtpr_provider;
15445
15446 if (!(prov->dtpv_priv.dtpp_flags & match))
15447 continue;
15448 }
15449
15450 dtrace_ecb_disable(ecb);
15451 dtrace_ecb_destroy(ecb);
15452 }
15453
15454 if (!match)
15455 break;
15456 }
15457
15458 /*
15459 * Before we free the buffers, perform one more sync to assure that
15460 * every CPU is out of probe context.
15461 */
15462 dtrace_sync();
15463
15464 dtrace_buffer_free(state->dts_buffer);
15465 dtrace_buffer_free(state->dts_aggbuffer);
15466
15467 for (i = 0; i < nspec; i++)
15468 dtrace_buffer_free(spec[i].dtsp_buffer);
15469
15470 #ifdef illumos
15471 if (state->dts_cleaner != CYCLIC_NONE)
15472 cyclic_remove(state->dts_cleaner);
15473
15474 if (state->dts_deadman != CYCLIC_NONE)
15475 cyclic_remove(state->dts_deadman);
15476 #else
15477 callout_stop(&state->dts_cleaner);
15478 callout_drain(&state->dts_cleaner);
15479 callout_stop(&state->dts_deadman);
15480 callout_drain(&state->dts_deadman);
15481 #endif
15482
15483 dtrace_dstate_fini(&vstate->dtvs_dynvars);
15484 dtrace_vstate_fini(vstate);
15485 if (state->dts_ecbs != NULL)
15486 kmem_free(state->dts_ecbs, state->dts_necbs * sizeof (dtrace_ecb_t *));
15487
15488 if (state->dts_aggregations != NULL) {
15489 #ifdef DEBUG
15490 for (i = 0; i < state->dts_naggregations; i++)
15491 ASSERT(state->dts_aggregations[i] == NULL);
15492 #endif
15493 ASSERT(state->dts_naggregations > 0);
15494 kmem_free(state->dts_aggregations,
15495 state->dts_naggregations * sizeof (dtrace_aggregation_t *));
15496 }
15497
15498 kmem_free(state->dts_buffer, bufsize);
15499 kmem_free(state->dts_aggbuffer, bufsize);
15500
15501 for (i = 0; i < nspec; i++)
15502 kmem_free(spec[i].dtsp_buffer, bufsize);
15503
15504 if (spec != NULL)
15505 kmem_free(spec, nspec * sizeof (dtrace_speculation_t));
15506
15507 dtrace_format_destroy(state);
15508
15509 if (state->dts_aggid_arena != NULL) {
15510 #ifdef illumos
15511 vmem_destroy(state->dts_aggid_arena);
15512 #else
15513 delete_unrhdr(state->dts_aggid_arena);
15514 #endif
15515 state->dts_aggid_arena = NULL;
15516 }
15517 #ifdef illumos
15518 ddi_soft_state_free(dtrace_softstate, minor);
15519 vmem_free(dtrace_minor, (void *)(uintptr_t)minor, 1);
15520 #endif
15521 }
15522
15523 /*
15524 * DTrace Anonymous Enabling Functions
15525 */
15526 static dtrace_state_t *
15527 dtrace_anon_grab(void)
15528 {
15529 dtrace_state_t *state;
15530
15531 ASSERT(MUTEX_HELD(&dtrace_lock));
15532
15533 if ((state = dtrace_anon.dta_state) == NULL) {
15534 ASSERT(dtrace_anon.dta_enabling == NULL);
15535 return (NULL);
15536 }
15537
15538 ASSERT(dtrace_anon.dta_enabling != NULL);
15539 ASSERT(dtrace_retained != NULL);
15540
15541 dtrace_enabling_destroy(dtrace_anon.dta_enabling);
15542 dtrace_anon.dta_enabling = NULL;
15543 dtrace_anon.dta_state = NULL;
15544
15545 return (state);
15546 }
15547
15548 static void
15549 dtrace_anon_property(void)
15550 {
15551 int i, rv;
15552 dtrace_state_t *state;
15553 dof_hdr_t *dof;
15554 char c[32]; /* enough for "dof-data-" + digits */
15555
15556 ASSERT(MUTEX_HELD(&dtrace_lock));
15557 ASSERT(MUTEX_HELD(&cpu_lock));
15558
15559 for (i = 0; ; i++) {
15560 (void) snprintf(c, sizeof (c), "dof-data-%d", i);
15561
15562 dtrace_err_verbose = 1;
15563
15564 if ((dof = dtrace_dof_property(c)) == NULL) {
15565 dtrace_err_verbose = 0;
15566 break;
15567 }
15568
15569 #ifdef illumos
15570 /*
15571 * We want to create anonymous state, so we need to transition
15572 * the kernel debugger to indicate that DTrace is active. If
15573 * this fails (e.g. because the debugger has modified text in
15574 * some way), we won't continue with the processing.
15575 */
15576 if (kdi_dtrace_set(KDI_DTSET_DTRACE_ACTIVATE) != 0) {
15577 cmn_err(CE_NOTE, "kernel debugger active; anonymous "
15578 "enabling ignored.");
15579 dtrace_dof_destroy(dof);
15580 break;
15581 }
15582 #endif
15583
15584 /*
15585 * If we haven't allocated an anonymous state, we'll do so now.
15586 */
15587 if ((state = dtrace_anon.dta_state) == NULL) {
15588 state = dtrace_state_create(NULL, NULL);
15589 dtrace_anon.dta_state = state;
15590
15591 if (state == NULL) {
15592 /*
15593 * This basically shouldn't happen: the only
15594 * failure mode from dtrace_state_create() is a
15595 * failure of ddi_soft_state_zalloc() that
15596 * itself should never happen. Still, the
15597 * interface allows for a failure mode, and
15598 * we want to fail as gracefully as possible:
15599 * we'll emit an error message and cease
15600 * processing anonymous state in this case.
15601 */
15602 cmn_err(CE_WARN, "failed to create "
15603 "anonymous state");
15604 dtrace_dof_destroy(dof);
15605 break;
15606 }
15607 }
15608
15609 rv = dtrace_dof_slurp(dof, &state->dts_vstate, CRED(),
15610 &dtrace_anon.dta_enabling, 0, 0, B_TRUE);
15611
15612 if (rv == 0)
15613 rv = dtrace_dof_options(dof, state);
15614
15615 dtrace_err_verbose = 0;
15616 dtrace_dof_destroy(dof);
15617
15618 if (rv != 0) {
15619 /*
15620 * This is malformed DOF; chuck any anonymous state
15621 * that we created.
15622 */
15623 ASSERT(dtrace_anon.dta_enabling == NULL);
15624 dtrace_state_destroy(state);
15625 dtrace_anon.dta_state = NULL;
15626 break;
15627 }
15628
15629 ASSERT(dtrace_anon.dta_enabling != NULL);
15630 }
15631
15632 if (dtrace_anon.dta_enabling != NULL) {
15633 int rval;
15634
15635 /*
15636 * dtrace_enabling_retain() can only fail because we are
15637 * trying to retain more enablings than are allowed -- but
15638 * we only have one anonymous enabling, and we are guaranteed
15639 * to be allowed at least one retained enabling; we assert
15640 * that dtrace_enabling_retain() returns success.
15641 */
15642 rval = dtrace_enabling_retain(dtrace_anon.dta_enabling);
15643 ASSERT(rval == 0);
15644
15645 dtrace_enabling_dump(dtrace_anon.dta_enabling);
15646 }
15647 }
15648
15649 /*
15650 * DTrace Helper Functions
15651 */
15652 static void
15653 dtrace_helper_trace(dtrace_helper_action_t *helper,
15654 dtrace_mstate_t *mstate, dtrace_vstate_t *vstate, int where)
15655 {
15656 uint32_t size, next, nnext, i;
15657 dtrace_helptrace_t *ent, *buffer;
15658 uint16_t flags = cpu_core[curcpu].cpuc_dtrace_flags;
15659
15660 if ((buffer = dtrace_helptrace_buffer) == NULL)
15661 return;
15662
15663 ASSERT(vstate->dtvs_nlocals <= dtrace_helptrace_nlocals);
15664
15665 /*
15666 * What would a tracing framework be without its own tracing
15667 * framework? (Well, a hell of a lot simpler, for starters...)
15668 */
15669 size = sizeof (dtrace_helptrace_t) + dtrace_helptrace_nlocals *
15670 sizeof (uint64_t) - sizeof (uint64_t);
15671
15672 /*
15673 * Iterate until we can allocate a slot in the trace buffer.
15674 */
15675 do {
15676 next = dtrace_helptrace_next;
15677
15678 if (next + size < dtrace_helptrace_bufsize) {
15679 nnext = next + size;
15680 } else {
15681 nnext = size;
15682 }
15683 } while (dtrace_cas32(&dtrace_helptrace_next, next, nnext) != next);
15684
15685 /*
15686 * We have our slot; fill it in.
15687 */
15688 if (nnext == size) {
15689 dtrace_helptrace_wrapped++;
15690 next = 0;
15691 }
15692
15693 ent = (dtrace_helptrace_t *)((uintptr_t)buffer + next);
15694 ent->dtht_helper = helper;
15695 ent->dtht_where = where;
15696 ent->dtht_nlocals = vstate->dtvs_nlocals;
15697
15698 ent->dtht_fltoffs = (mstate->dtms_present & DTRACE_MSTATE_FLTOFFS) ?
15699 mstate->dtms_fltoffs : -1;
15700 ent->dtht_fault = DTRACE_FLAGS2FLT(flags);
15701 ent->dtht_illval = cpu_core[curcpu].cpuc_dtrace_illval;
15702
15703 for (i = 0; i < vstate->dtvs_nlocals; i++) {
15704 dtrace_statvar_t *svar;
15705
15706 if ((svar = vstate->dtvs_locals[i]) == NULL)
15707 continue;
15708
15709 ASSERT(svar->dtsv_size >= NCPU * sizeof (uint64_t));
15710 ent->dtht_locals[i] =
15711 ((uint64_t *)(uintptr_t)svar->dtsv_data)[curcpu];
15712 }
15713 }
15714
15715 static uint64_t
15716 dtrace_helper(int which, dtrace_mstate_t *mstate,
15717 dtrace_state_t *state, uint64_t arg0, uint64_t arg1)
15718 {
15719 uint16_t *flags = &cpu_core[curcpu].cpuc_dtrace_flags;
15720 uint64_t sarg0 = mstate->dtms_arg[0];
15721 uint64_t sarg1 = mstate->dtms_arg[1];
15722 uint64_t rval = 0;
15723 dtrace_helpers_t *helpers = curproc->p_dtrace_helpers;
15724 dtrace_helper_action_t *helper;
15725 dtrace_vstate_t *vstate;
15726 dtrace_difo_t *pred;
15727 int i, trace = dtrace_helptrace_buffer != NULL;
15728
15729 ASSERT(which >= 0 && which < DTRACE_NHELPER_ACTIONS);
15730
15731 if (helpers == NULL)
15732 return (0);
15733
15734 if ((helper = helpers->dthps_actions[which]) == NULL)
15735 return (0);
15736
15737 vstate = &helpers->dthps_vstate;
15738 mstate->dtms_arg[0] = arg0;
15739 mstate->dtms_arg[1] = arg1;
15740
15741 /*
15742 * Now iterate over each helper. If its predicate evaluates to 'true',
15743 * we'll call the corresponding actions. Note that the below calls
15744 * to dtrace_dif_emulate() may set faults in machine state. This is
15745 * okay: our caller (the outer dtrace_dif_emulate()) will simply plow
15746 * the stored DIF offset with its own (which is the desired behavior).
15747 * Also, note the calls to dtrace_dif_emulate() may allocate scratch
15748 * from machine state; this is okay, too.
15749 */
15750 for (; helper != NULL; helper = helper->dtha_next) {
15751 if ((pred = helper->dtha_predicate) != NULL) {
15752 if (trace)
15753 dtrace_helper_trace(helper, mstate, vstate, 0);
15754
15755 if (!dtrace_dif_emulate(pred, mstate, vstate, state))
15756 goto next;
15757
15758 if (*flags & CPU_DTRACE_FAULT)
15759 goto err;
15760 }
15761
15762 for (i = 0; i < helper->dtha_nactions; i++) {
15763 if (trace)
15764 dtrace_helper_trace(helper,
15765 mstate, vstate, i + 1);
15766
15767 rval = dtrace_dif_emulate(helper->dtha_actions[i],
15768 mstate, vstate, state);
15769
15770 if (*flags & CPU_DTRACE_FAULT)
15771 goto err;
15772 }
15773
15774 next:
15775 if (trace)
15776 dtrace_helper_trace(helper, mstate, vstate,
15777 DTRACE_HELPTRACE_NEXT);
15778 }
15779
15780 if (trace)
15781 dtrace_helper_trace(helper, mstate, vstate,
15782 DTRACE_HELPTRACE_DONE);
15783
15784 /*
15785 * Restore the arg0 that we saved upon entry.
15786 */
15787 mstate->dtms_arg[0] = sarg0;
15788 mstate->dtms_arg[1] = sarg1;
15789
15790 return (rval);
15791
15792 err:
15793 if (trace)
15794 dtrace_helper_trace(helper, mstate, vstate,
15795 DTRACE_HELPTRACE_ERR);
15796
15797 /*
15798 * Restore the arg0 that we saved upon entry.
15799 */
15800 mstate->dtms_arg[0] = sarg0;
15801 mstate->dtms_arg[1] = sarg1;
15802
15803 return (0);
15804 }
15805
15806 static void
15807 dtrace_helper_action_destroy(dtrace_helper_action_t *helper,
15808 dtrace_vstate_t *vstate)
15809 {
15810 int i;
15811
15812 if (helper->dtha_predicate != NULL)
15813 dtrace_difo_release(helper->dtha_predicate, vstate);
15814
15815 for (i = 0; i < helper->dtha_nactions; i++) {
15816 ASSERT(helper->dtha_actions[i] != NULL);
15817 dtrace_difo_release(helper->dtha_actions[i], vstate);
15818 }
15819
15820 kmem_free(helper->dtha_actions,
15821 helper->dtha_nactions * sizeof (dtrace_difo_t *));
15822 kmem_free(helper, sizeof (dtrace_helper_action_t));
15823 }
15824
15825 static int
15826 dtrace_helper_destroygen(dtrace_helpers_t *help, int gen)
15827 {
15828 proc_t *p = curproc;
15829 dtrace_vstate_t *vstate;
15830 int i;
15831
15832 if (help == NULL)
15833 help = p->p_dtrace_helpers;
15834
15835 ASSERT(MUTEX_HELD(&dtrace_lock));
15836
15837 if (help == NULL || gen > help->dthps_generation)
15838 return (EINVAL);
15839
15840 vstate = &help->dthps_vstate;
15841
15842 for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
15843 dtrace_helper_action_t *last = NULL, *h, *next;
15844
15845 for (h = help->dthps_actions[i]; h != NULL; h = next) {
15846 next = h->dtha_next;
15847
15848 if (h->dtha_generation == gen) {
15849 if (last != NULL) {
15850 last->dtha_next = next;
15851 } else {
15852 help->dthps_actions[i] = next;
15853 }
15854
15855 dtrace_helper_action_destroy(h, vstate);
15856 } else {
15857 last = h;
15858 }
15859 }
15860 }
15861
15862 /*
15863 * Interate until we've cleared out all helper providers with the
15864 * given generation number.
15865 */
15866 for (;;) {
15867 dtrace_helper_provider_t *prov;
15868
15869 /*
15870 * Look for a helper provider with the right generation. We
15871 * have to start back at the beginning of the list each time
15872 * because we drop dtrace_lock. It's unlikely that we'll make
15873 * more than two passes.
15874 */
15875 for (i = 0; i < help->dthps_nprovs; i++) {
15876 prov = help->dthps_provs[i];
15877
15878 if (prov->dthp_generation == gen)
15879 break;
15880 }
15881
15882 /*
15883 * If there were no matches, we're done.
15884 */
15885 if (i == help->dthps_nprovs)
15886 break;
15887
15888 /*
15889 * Move the last helper provider into this slot.
15890 */
15891 help->dthps_nprovs--;
15892 help->dthps_provs[i] = help->dthps_provs[help->dthps_nprovs];
15893 help->dthps_provs[help->dthps_nprovs] = NULL;
15894
15895 mutex_exit(&dtrace_lock);
15896
15897 /*
15898 * If we have a meta provider, remove this helper provider.
15899 */
15900 mutex_enter(&dtrace_meta_lock);
15901 if (dtrace_meta_pid != NULL) {
15902 ASSERT(dtrace_deferred_pid == NULL);
15903 dtrace_helper_provider_remove(&prov->dthp_prov,
15904 p->p_pid);
15905 }
15906 mutex_exit(&dtrace_meta_lock);
15907
15908 dtrace_helper_provider_destroy(prov);
15909
15910 mutex_enter(&dtrace_lock);
15911 }
15912
15913 return (0);
15914 }
15915
15916 static int
15917 dtrace_helper_validate(dtrace_helper_action_t *helper)
15918 {
15919 int err = 0, i;
15920 dtrace_difo_t *dp;
15921
15922 if ((dp = helper->dtha_predicate) != NULL)
15923 err += dtrace_difo_validate_helper(dp);
15924
15925 for (i = 0; i < helper->dtha_nactions; i++)
15926 err += dtrace_difo_validate_helper(helper->dtha_actions[i]);
15927
15928 return (err == 0);
15929 }
15930
15931 static int
15932 dtrace_helper_action_add(int which, dtrace_ecbdesc_t *ep,
15933 dtrace_helpers_t *help)
15934 {
15935 dtrace_helper_action_t *helper, *last;
15936 dtrace_actdesc_t *act;
15937 dtrace_vstate_t *vstate;
15938 dtrace_predicate_t *pred;
15939 int count = 0, nactions = 0, i;
15940
15941 if (which < 0 || which >= DTRACE_NHELPER_ACTIONS)
15942 return (EINVAL);
15943
15944 last = help->dthps_actions[which];
15945 vstate = &help->dthps_vstate;
15946
15947 for (count = 0; last != NULL; last = last->dtha_next) {
15948 count++;
15949 if (last->dtha_next == NULL)
15950 break;
15951 }
15952
15953 /*
15954 * If we already have dtrace_helper_actions_max helper actions for this
15955 * helper action type, we'll refuse to add a new one.
15956 */
15957 if (count >= dtrace_helper_actions_max)
15958 return (ENOSPC);
15959
15960 helper = kmem_zalloc(sizeof (dtrace_helper_action_t), KM_SLEEP);
15961 helper->dtha_generation = help->dthps_generation;
15962
15963 if ((pred = ep->dted_pred.dtpdd_predicate) != NULL) {
15964 ASSERT(pred->dtp_difo != NULL);
15965 dtrace_difo_hold(pred->dtp_difo);
15966 helper->dtha_predicate = pred->dtp_difo;
15967 }
15968
15969 for (act = ep->dted_action; act != NULL; act = act->dtad_next) {
15970 if (act->dtad_kind != DTRACEACT_DIFEXPR)
15971 goto err;
15972
15973 if (act->dtad_difo == NULL)
15974 goto err;
15975
15976 nactions++;
15977 }
15978
15979 helper->dtha_actions = kmem_zalloc(sizeof (dtrace_difo_t *) *
15980 (helper->dtha_nactions = nactions), KM_SLEEP);
15981
15982 for (act = ep->dted_action, i = 0; act != NULL; act = act->dtad_next) {
15983 dtrace_difo_hold(act->dtad_difo);
15984 helper->dtha_actions[i++] = act->dtad_difo;
15985 }
15986
15987 if (!dtrace_helper_validate(helper))
15988 goto err;
15989
15990 if (last == NULL) {
15991 help->dthps_actions[which] = helper;
15992 } else {
15993 last->dtha_next = helper;
15994 }
15995
15996 if (vstate->dtvs_nlocals > dtrace_helptrace_nlocals) {
15997 dtrace_helptrace_nlocals = vstate->dtvs_nlocals;
15998 dtrace_helptrace_next = 0;
15999 }
16000
16001 return (0);
16002 err:
16003 dtrace_helper_action_destroy(helper, vstate);
16004 return (EINVAL);
16005 }
16006
16007 static void
16008 dtrace_helper_provider_register(proc_t *p, dtrace_helpers_t *help,
16009 dof_helper_t *dofhp)
16010 {
16011 ASSERT(MUTEX_NOT_HELD(&dtrace_lock));
16012
16013 mutex_enter(&dtrace_meta_lock);
16014 mutex_enter(&dtrace_lock);
16015
16016 if (!dtrace_attached() || dtrace_meta_pid == NULL) {
16017 /*
16018 * If the dtrace module is loaded but not attached, or if
16019 * there aren't isn't a meta provider registered to deal with
16020 * these provider descriptions, we need to postpone creating
16021 * the actual providers until later.
16022 */
16023
16024 if (help->dthps_next == NULL && help->dthps_prev == NULL &&
16025 dtrace_deferred_pid != help) {
16026 help->dthps_deferred = 1;
16027 help->dthps_pid = p->p_pid;
16028 help->dthps_next = dtrace_deferred_pid;
16029 help->dthps_prev = NULL;
16030 if (dtrace_deferred_pid != NULL)
16031 dtrace_deferred_pid->dthps_prev = help;
16032 dtrace_deferred_pid = help;
16033 }
16034
16035 mutex_exit(&dtrace_lock);
16036
16037 } else if (dofhp != NULL) {
16038 /*
16039 * If the dtrace module is loaded and we have a particular
16040 * helper provider description, pass that off to the
16041 * meta provider.
16042 */
16043
16044 mutex_exit(&dtrace_lock);
16045
16046 dtrace_helper_provide(dofhp, p->p_pid);
16047
16048 } else {
16049 /*
16050 * Otherwise, just pass all the helper provider descriptions
16051 * off to the meta provider.
16052 */
16053
16054 int i;
16055 mutex_exit(&dtrace_lock);
16056
16057 for (i = 0; i < help->dthps_nprovs; i++) {
16058 dtrace_helper_provide(&help->dthps_provs[i]->dthp_prov,
16059 p->p_pid);
16060 }
16061 }
16062
16063 mutex_exit(&dtrace_meta_lock);
16064 }
16065
16066 static int
16067 dtrace_helper_provider_add(dof_helper_t *dofhp, dtrace_helpers_t *help, int gen)
16068 {
16069 dtrace_helper_provider_t *hprov, **tmp_provs;
16070 uint_t tmp_maxprovs, i;
16071
16072 ASSERT(MUTEX_HELD(&dtrace_lock));
16073 ASSERT(help != NULL);
16074
16075 /*
16076 * If we already have dtrace_helper_providers_max helper providers,
16077 * we're refuse to add a new one.
16078 */
16079 if (help->dthps_nprovs >= dtrace_helper_providers_max)
16080 return (ENOSPC);
16081
16082 /*
16083 * Check to make sure this isn't a duplicate.
16084 */
16085 for (i = 0; i < help->dthps_nprovs; i++) {
16086 if (dofhp->dofhp_addr ==
16087 help->dthps_provs[i]->dthp_prov.dofhp_addr)
16088 return (EALREADY);
16089 }
16090
16091 hprov = kmem_zalloc(sizeof (dtrace_helper_provider_t), KM_SLEEP);
16092 hprov->dthp_prov = *dofhp;
16093 hprov->dthp_ref = 1;
16094 hprov->dthp_generation = gen;
16095
16096 /*
16097 * Allocate a bigger table for helper providers if it's already full.
16098 */
16099 if (help->dthps_maxprovs == help->dthps_nprovs) {
16100 tmp_maxprovs = help->dthps_maxprovs;
16101 tmp_provs = help->dthps_provs;
16102
16103 if (help->dthps_maxprovs == 0)
16104 help->dthps_maxprovs = 2;
16105 else
16106 help->dthps_maxprovs *= 2;
16107 if (help->dthps_maxprovs > dtrace_helper_providers_max)
16108 help->dthps_maxprovs = dtrace_helper_providers_max;
16109
16110 ASSERT(tmp_maxprovs < help->dthps_maxprovs);
16111
16112 help->dthps_provs = kmem_zalloc(help->dthps_maxprovs *
16113 sizeof (dtrace_helper_provider_t *), KM_SLEEP);
16114
16115 if (tmp_provs != NULL) {
16116 bcopy(tmp_provs, help->dthps_provs, tmp_maxprovs *
16117 sizeof (dtrace_helper_provider_t *));
16118 kmem_free(tmp_provs, tmp_maxprovs *
16119 sizeof (dtrace_helper_provider_t *));
16120 }
16121 }
16122
16123 help->dthps_provs[help->dthps_nprovs] = hprov;
16124 help->dthps_nprovs++;
16125
16126 return (0);
16127 }
16128
16129 static void
16130 dtrace_helper_provider_destroy(dtrace_helper_provider_t *hprov)
16131 {
16132 mutex_enter(&dtrace_lock);
16133
16134 if (--hprov->dthp_ref == 0) {
16135 dof_hdr_t *dof;
16136 mutex_exit(&dtrace_lock);
16137 dof = (dof_hdr_t *)(uintptr_t)hprov->dthp_prov.dofhp_dof;
16138 dtrace_dof_destroy(dof);
16139 kmem_free(hprov, sizeof (dtrace_helper_provider_t));
16140 } else {
16141 mutex_exit(&dtrace_lock);
16142 }
16143 }
16144
16145 static int
16146 dtrace_helper_provider_validate(dof_hdr_t *dof, dof_sec_t *sec)
16147 {
16148 uintptr_t daddr = (uintptr_t)dof;
16149 dof_sec_t *str_sec, *prb_sec, *arg_sec, *off_sec, *enoff_sec;
16150 dof_provider_t *provider;
16151 dof_probe_t *probe;
16152 uint8_t *arg;
16153 char *strtab, *typestr;
16154 dof_stridx_t typeidx;
16155 size_t typesz;
16156 uint_t nprobes, j, k;
16157
16158 ASSERT(sec->dofs_type == DOF_SECT_PROVIDER);
16159
16160 if (sec->dofs_offset & (sizeof (uint_t) - 1)) {
16161 dtrace_dof_error(dof, "misaligned section offset");
16162 return (-1);
16163 }
16164
16165 /*
16166 * The section needs to be large enough to contain the DOF provider
16167 * structure appropriate for the given version.
16168 */
16169 if (sec->dofs_size <
16170 ((dof->dofh_ident[DOF_ID_VERSION] == DOF_VERSION_1) ?
16171 offsetof(dof_provider_t, dofpv_prenoffs) :
16172 sizeof (dof_provider_t))) {
16173 dtrace_dof_error(dof, "provider section too small");
16174 return (-1);
16175 }
16176
16177 provider = (dof_provider_t *)(uintptr_t)(daddr + sec->dofs_offset);
16178 str_sec = dtrace_dof_sect(dof, DOF_SECT_STRTAB, provider->dofpv_strtab);
16179 prb_sec = dtrace_dof_sect(dof, DOF_SECT_PROBES, provider->dofpv_probes);
16180 arg_sec = dtrace_dof_sect(dof, DOF_SECT_PRARGS, provider->dofpv_prargs);
16181 off_sec = dtrace_dof_sect(dof, DOF_SECT_PROFFS, provider->dofpv_proffs);
16182
16183 if (str_sec == NULL || prb_sec == NULL ||
16184 arg_sec == NULL || off_sec == NULL)
16185 return (-1);
16186
16187 enoff_sec = NULL;
16188
16189 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1 &&
16190 provider->dofpv_prenoffs != DOF_SECT_NONE &&
16191 (enoff_sec = dtrace_dof_sect(dof, DOF_SECT_PRENOFFS,
16192 provider->dofpv_prenoffs)) == NULL)
16193 return (-1);
16194
16195 strtab = (char *)(uintptr_t)(daddr + str_sec->dofs_offset);
16196
16197 if (provider->dofpv_name >= str_sec->dofs_size ||
16198 strlen(strtab + provider->dofpv_name) >= DTRACE_PROVNAMELEN) {
16199 dtrace_dof_error(dof, "invalid provider name");
16200 return (-1);
16201 }
16202
16203 if (prb_sec->dofs_entsize == 0 ||
16204 prb_sec->dofs_entsize > prb_sec->dofs_size) {
16205 dtrace_dof_error(dof, "invalid entry size");
16206 return (-1);
16207 }
16208
16209 if (prb_sec->dofs_entsize & (sizeof (uintptr_t) - 1)) {
16210 dtrace_dof_error(dof, "misaligned entry size");
16211 return (-1);
16212 }
16213
16214 if (off_sec->dofs_entsize != sizeof (uint32_t)) {
16215 dtrace_dof_error(dof, "invalid entry size");
16216 return (-1);
16217 }
16218
16219 if (off_sec->dofs_offset & (sizeof (uint32_t) - 1)) {
16220 dtrace_dof_error(dof, "misaligned section offset");
16221 return (-1);
16222 }
16223
16224 if (arg_sec->dofs_entsize != sizeof (uint8_t)) {
16225 dtrace_dof_error(dof, "invalid entry size");
16226 return (-1);
16227 }
16228
16229 arg = (uint8_t *)(uintptr_t)(daddr + arg_sec->dofs_offset);
16230
16231 nprobes = prb_sec->dofs_size / prb_sec->dofs_entsize;
16232
16233 /*
16234 * Take a pass through the probes to check for errors.
16235 */
16236 for (j = 0; j < nprobes; j++) {
16237 probe = (dof_probe_t *)(uintptr_t)(daddr +
16238 prb_sec->dofs_offset + j * prb_sec->dofs_entsize);
16239
16240 if (probe->dofpr_func >= str_sec->dofs_size) {
16241 dtrace_dof_error(dof, "invalid function name");
16242 return (-1);
16243 }
16244
16245 if (strlen(strtab + probe->dofpr_func) >= DTRACE_FUNCNAMELEN) {
16246 dtrace_dof_error(dof, "function name too long");
16247 /*
16248 * Keep going if the function name is too long.
16249 * Unlike provider and probe names, we cannot reasonably
16250 * impose restrictions on function names, since they're
16251 * a property of the code being instrumented. We will
16252 * skip this probe in dtrace_helper_provide_one().
16253 */
16254 }
16255
16256 if (probe->dofpr_name >= str_sec->dofs_size ||
16257 strlen(strtab + probe->dofpr_name) >= DTRACE_NAMELEN) {
16258 dtrace_dof_error(dof, "invalid probe name");
16259 return (-1);
16260 }
16261
16262 /*
16263 * The offset count must not wrap the index, and the offsets
16264 * must also not overflow the section's data.
16265 */
16266 if (probe->dofpr_offidx + probe->dofpr_noffs <
16267 probe->dofpr_offidx ||
16268 (probe->dofpr_offidx + probe->dofpr_noffs) *
16269 off_sec->dofs_entsize > off_sec->dofs_size) {
16270 dtrace_dof_error(dof, "invalid probe offset");
16271 return (-1);
16272 }
16273
16274 if (dof->dofh_ident[DOF_ID_VERSION] != DOF_VERSION_1) {
16275 /*
16276 * If there's no is-enabled offset section, make sure
16277 * there aren't any is-enabled offsets. Otherwise
16278 * perform the same checks as for probe offsets
16279 * (immediately above).
16280 */
16281 if (enoff_sec == NULL) {
16282 if (probe->dofpr_enoffidx != 0 ||
16283 probe->dofpr_nenoffs != 0) {
16284 dtrace_dof_error(dof, "is-enabled "
16285 "offsets with null section");
16286 return (-1);
16287 }
16288 } else if (probe->dofpr_enoffidx +
16289 probe->dofpr_nenoffs < probe->dofpr_enoffidx ||
16290 (probe->dofpr_enoffidx + probe->dofpr_nenoffs) *
16291 enoff_sec->dofs_entsize > enoff_sec->dofs_size) {
16292 dtrace_dof_error(dof, "invalid is-enabled "
16293 "offset");
16294 return (-1);
16295 }
16296
16297 if (probe->dofpr_noffs + probe->dofpr_nenoffs == 0) {
16298 dtrace_dof_error(dof, "zero probe and "
16299 "is-enabled offsets");
16300 return (-1);
16301 }
16302 } else if (probe->dofpr_noffs == 0) {
16303 dtrace_dof_error(dof, "zero probe offsets");
16304 return (-1);
16305 }
16306
16307 if (probe->dofpr_argidx + probe->dofpr_xargc <
16308 probe->dofpr_argidx ||
16309 (probe->dofpr_argidx + probe->dofpr_xargc) *
16310 arg_sec->dofs_entsize > arg_sec->dofs_size) {
16311 dtrace_dof_error(dof, "invalid args");
16312 return (-1);
16313 }
16314
16315 typeidx = probe->dofpr_nargv;
16316 typestr = strtab + probe->dofpr_nargv;
16317 for (k = 0; k < probe->dofpr_nargc; k++) {
16318 if (typeidx >= str_sec->dofs_size) {
16319 dtrace_dof_error(dof, "bad "
16320 "native argument type");
16321 return (-1);
16322 }
16323
16324 typesz = strlen(typestr) + 1;
16325 if (typesz > DTRACE_ARGTYPELEN) {
16326 dtrace_dof_error(dof, "native "
16327 "argument type too long");
16328 return (-1);
16329 }
16330 typeidx += typesz;
16331 typestr += typesz;
16332 }
16333
16334 typeidx = probe->dofpr_xargv;
16335 typestr = strtab + probe->dofpr_xargv;
16336 for (k = 0; k < probe->dofpr_xargc; k++) {
16337 if (arg[probe->dofpr_argidx + k] > probe->dofpr_nargc) {
16338 dtrace_dof_error(dof, "bad "
16339 "native argument index");
16340 return (-1);
16341 }
16342
16343 if (typeidx >= str_sec->dofs_size) {
16344 dtrace_dof_error(dof, "bad "
16345 "translated argument type");
16346 return (-1);
16347 }
16348
16349 typesz = strlen(typestr) + 1;
16350 if (typesz > DTRACE_ARGTYPELEN) {
16351 dtrace_dof_error(dof, "translated argument "
16352 "type too long");
16353 return (-1);
16354 }
16355
16356 typeidx += typesz;
16357 typestr += typesz;
16358 }
16359 }
16360
16361 return (0);
16362 }
16363
16364 static int
16365 dtrace_helper_slurp(dof_hdr_t *dof, dof_helper_t *dhp, struct proc *p)
16366 {
16367 dtrace_helpers_t *help;
16368 dtrace_vstate_t *vstate;
16369 dtrace_enabling_t *enab = NULL;
16370 int i, gen, rv, nhelpers = 0, nprovs = 0, destroy = 1;
16371 uintptr_t daddr = (uintptr_t)dof;
16372
16373 ASSERT(MUTEX_HELD(&dtrace_lock));
16374
16375 if ((help = p->p_dtrace_helpers) == NULL)
16376 help = dtrace_helpers_create(p);
16377
16378 vstate = &help->dthps_vstate;
16379
16380 if ((rv = dtrace_dof_slurp(dof, vstate, NULL, &enab, dhp->dofhp_addr,
16381 dhp->dofhp_dof, B_FALSE)) != 0) {
16382 dtrace_dof_destroy(dof);
16383 return (rv);
16384 }
16385
16386 /*
16387 * Look for helper providers and validate their descriptions.
16388 */
16389 for (i = 0; i < dof->dofh_secnum; i++) {
16390 dof_sec_t *sec = (dof_sec_t *)(uintptr_t)(daddr +
16391 dof->dofh_secoff + i * dof->dofh_secsize);
16392
16393 if (sec->dofs_type != DOF_SECT_PROVIDER)
16394 continue;
16395
16396 if (dtrace_helper_provider_validate(dof, sec) != 0) {
16397 dtrace_enabling_destroy(enab);
16398 dtrace_dof_destroy(dof);
16399 return (-1);
16400 }
16401
16402 nprovs++;
16403 }
16404
16405 /*
16406 * Now we need to walk through the ECB descriptions in the enabling.
16407 */
16408 for (i = 0; i < enab->dten_ndesc; i++) {
16409 dtrace_ecbdesc_t *ep = enab->dten_desc[i];
16410 dtrace_probedesc_t *desc = &ep->dted_probe;
16411
16412 if (strcmp(desc->dtpd_provider, "dtrace") != 0)
16413 continue;
16414
16415 if (strcmp(desc->dtpd_mod, "helper") != 0)
16416 continue;
16417
16418 if (strcmp(desc->dtpd_func, "ustack") != 0)
16419 continue;
16420
16421 if ((rv = dtrace_helper_action_add(DTRACE_HELPER_ACTION_USTACK,
16422 ep, help)) != 0) {
16423 /*
16424 * Adding this helper action failed -- we are now going
16425 * to rip out the entire generation and return failure.
16426 */
16427 (void) dtrace_helper_destroygen(help,
16428 help->dthps_generation);
16429 dtrace_enabling_destroy(enab);
16430 dtrace_dof_destroy(dof);
16431 return (-1);
16432 }
16433
16434 nhelpers++;
16435 }
16436
16437 if (nhelpers < enab->dten_ndesc)
16438 dtrace_dof_error(dof, "unmatched helpers");
16439
16440 gen = help->dthps_generation++;
16441 dtrace_enabling_destroy(enab);
16442
16443 if (nprovs > 0) {
16444 /*
16445 * Now that this is in-kernel, we change the sense of the
16446 * members: dofhp_dof denotes the in-kernel copy of the DOF
16447 * and dofhp_addr denotes the address at user-level.
16448 */
16449 dhp->dofhp_addr = dhp->dofhp_dof;
16450 dhp->dofhp_dof = (uint64_t)(uintptr_t)dof;
16451
16452 if (dtrace_helper_provider_add(dhp, help, gen) == 0) {
16453 mutex_exit(&dtrace_lock);
16454 dtrace_helper_provider_register(p, help, dhp);
16455 mutex_enter(&dtrace_lock);
16456
16457 destroy = 0;
16458 }
16459 }
16460
16461 if (destroy)
16462 dtrace_dof_destroy(dof);
16463
16464 return (gen);
16465 }
16466
16467 static dtrace_helpers_t *
16468 dtrace_helpers_create(proc_t *p)
16469 {
16470 dtrace_helpers_t *help;
16471
16472 ASSERT(MUTEX_HELD(&dtrace_lock));
16473 ASSERT(p->p_dtrace_helpers == NULL);
16474
16475 help = kmem_zalloc(sizeof (dtrace_helpers_t), KM_SLEEP);
16476 help->dthps_actions = kmem_zalloc(sizeof (dtrace_helper_action_t *) *
16477 DTRACE_NHELPER_ACTIONS, KM_SLEEP);
16478
16479 p->p_dtrace_helpers = help;
16480 dtrace_helpers++;
16481
16482 return (help);
16483 }
16484
16485 #ifdef illumos
16486 static
16487 #endif
16488 void
16489 dtrace_helpers_destroy(proc_t *p)
16490 {
16491 dtrace_helpers_t *help;
16492 dtrace_vstate_t *vstate;
16493 #ifdef illumos
16494 proc_t *p = curproc;
16495 #endif
16496 int i;
16497
16498 mutex_enter(&dtrace_lock);
16499
16500 ASSERT(p->p_dtrace_helpers != NULL);
16501 ASSERT(dtrace_helpers > 0);
16502
16503 help = p->p_dtrace_helpers;
16504 vstate = &help->dthps_vstate;
16505
16506 /*
16507 * We're now going to lose the help from this process.
16508 */
16509 p->p_dtrace_helpers = NULL;
16510 dtrace_sync();
16511
16512 /*
16513 * Destory the helper actions.
16514 */
16515 for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
16516 dtrace_helper_action_t *h, *next;
16517
16518 for (h = help->dthps_actions[i]; h != NULL; h = next) {
16519 next = h->dtha_next;
16520 dtrace_helper_action_destroy(h, vstate);
16521 h = next;
16522 }
16523 }
16524
16525 mutex_exit(&dtrace_lock);
16526
16527 /*
16528 * Destroy the helper providers.
16529 */
16530 if (help->dthps_maxprovs > 0) {
16531 mutex_enter(&dtrace_meta_lock);
16532 if (dtrace_meta_pid != NULL) {
16533 ASSERT(dtrace_deferred_pid == NULL);
16534
16535 for (i = 0; i < help->dthps_nprovs; i++) {
16536 dtrace_helper_provider_remove(
16537 &help->dthps_provs[i]->dthp_prov, p->p_pid);
16538 }
16539 } else {
16540 mutex_enter(&dtrace_lock);
16541 ASSERT(help->dthps_deferred == 0 ||
16542 help->dthps_next != NULL ||
16543 help->dthps_prev != NULL ||
16544 help == dtrace_deferred_pid);
16545
16546 /*
16547 * Remove the helper from the deferred list.
16548 */
16549 if (help->dthps_next != NULL)
16550 help->dthps_next->dthps_prev = help->dthps_prev;
16551 if (help->dthps_prev != NULL)
16552 help->dthps_prev->dthps_next = help->dthps_next;
16553 if (dtrace_deferred_pid == help) {
16554 dtrace_deferred_pid = help->dthps_next;
16555 ASSERT(help->dthps_prev == NULL);
16556 }
16557
16558 mutex_exit(&dtrace_lock);
16559 }
16560
16561 mutex_exit(&dtrace_meta_lock);
16562
16563 for (i = 0; i < help->dthps_nprovs; i++) {
16564 dtrace_helper_provider_destroy(help->dthps_provs[i]);
16565 }
16566
16567 kmem_free(help->dthps_provs, help->dthps_maxprovs *
16568 sizeof (dtrace_helper_provider_t *));
16569 }
16570
16571 mutex_enter(&dtrace_lock);
16572
16573 dtrace_vstate_fini(&help->dthps_vstate);
16574 kmem_free(help->dthps_actions,
16575 sizeof (dtrace_helper_action_t *) * DTRACE_NHELPER_ACTIONS);
16576 kmem_free(help, sizeof (dtrace_helpers_t));
16577
16578 --dtrace_helpers;
16579 mutex_exit(&dtrace_lock);
16580 }
16581
16582 #ifdef illumos
16583 static
16584 #endif
16585 void
16586 dtrace_helpers_duplicate(proc_t *from, proc_t *to)
16587 {
16588 dtrace_helpers_t *help, *newhelp;
16589 dtrace_helper_action_t *helper, *new, *last;
16590 dtrace_difo_t *dp;
16591 dtrace_vstate_t *vstate;
16592 int i, j, sz, hasprovs = 0;
16593
16594 mutex_enter(&dtrace_lock);
16595 ASSERT(from->p_dtrace_helpers != NULL);
16596 ASSERT(dtrace_helpers > 0);
16597
16598 help = from->p_dtrace_helpers;
16599 newhelp = dtrace_helpers_create(to);
16600 ASSERT(to->p_dtrace_helpers != NULL);
16601
16602 newhelp->dthps_generation = help->dthps_generation;
16603 vstate = &newhelp->dthps_vstate;
16604
16605 /*
16606 * Duplicate the helper actions.
16607 */
16608 for (i = 0; i < DTRACE_NHELPER_ACTIONS; i++) {
16609 if ((helper = help->dthps_actions[i]) == NULL)
16610 continue;
16611
16612 for (last = NULL; helper != NULL; helper = helper->dtha_next) {
16613 new = kmem_zalloc(sizeof (dtrace_helper_action_t),
16614 KM_SLEEP);
16615 new->dtha_generation = helper->dtha_generation;
16616
16617 if ((dp = helper->dtha_predicate) != NULL) {
16618 dp = dtrace_difo_duplicate(dp, vstate);
16619 new->dtha_predicate = dp;
16620 }
16621
16622 new->dtha_nactions = helper->dtha_nactions;
16623 sz = sizeof (dtrace_difo_t *) * new->dtha_nactions;
16624 new->dtha_actions = kmem_alloc(sz, KM_SLEEP);
16625
16626 for (j = 0; j < new->dtha_nactions; j++) {
16627 dtrace_difo_t *dp = helper->dtha_actions[j];
16628
16629 ASSERT(dp != NULL);
16630 dp = dtrace_difo_duplicate(dp, vstate);
16631 new->dtha_actions[j] = dp;
16632 }
16633
16634 if (last != NULL) {
16635 last->dtha_next = new;
16636 } else {
16637 newhelp->dthps_actions[i] = new;
16638 }
16639
16640 last = new;
16641 }
16642 }
16643
16644 /*
16645 * Duplicate the helper providers and register them with the
16646 * DTrace framework.
16647 */
16648 if (help->dthps_nprovs > 0) {
16649 newhelp->dthps_nprovs = help->dthps_nprovs;
16650 newhelp->dthps_maxprovs = help->dthps_nprovs;
16651 newhelp->dthps_provs = kmem_alloc(newhelp->dthps_nprovs *
16652 sizeof (dtrace_helper_provider_t *), KM_SLEEP);
16653 for (i = 0; i < newhelp->dthps_nprovs; i++) {
16654 newhelp->dthps_provs[i] = help->dthps_provs[i];
16655 newhelp->dthps_provs[i]->dthp_ref++;
16656 }
16657
16658 hasprovs = 1;
16659 }
16660
16661 mutex_exit(&dtrace_lock);
16662
16663 if (hasprovs)
16664 dtrace_helper_provider_register(to, newhelp, NULL);
16665 }
16666
16667 /*
16668 * DTrace Hook Functions
16669 */
16670 static void
16671 dtrace_module_loaded(modctl_t *ctl)
16672 {
16673 dtrace_provider_t *prv;
16674
16675 mutex_enter(&dtrace_provider_lock);
16676 #ifdef illumos
16677 mutex_enter(&mod_lock);
16678 #endif
16679
16680 #ifdef illumos
16681 ASSERT(ctl->mod_busy);
16682 #endif
16683
16684 /*
16685 * We're going to call each providers per-module provide operation
16686 * specifying only this module.
16687 */
16688 for (prv = dtrace_provider; prv != NULL; prv = prv->dtpv_next)
16689 prv->dtpv_pops.dtps_provide_module(prv->dtpv_arg, ctl);
16690
16691 #ifdef illumos
16692 mutex_exit(&mod_lock);
16693 #endif
16694 mutex_exit(&dtrace_provider_lock);
16695
16696 /*
16697 * If we have any retained enablings, we need to match against them.
16698 * Enabling probes requires that cpu_lock be held, and we cannot hold
16699 * cpu_lock here -- it is legal for cpu_lock to be held when loading a
16700 * module. (In particular, this happens when loading scheduling
16701 * classes.) So if we have any retained enablings, we need to dispatch
16702 * our task queue to do the match for us.
16703 */
16704 mutex_enter(&dtrace_lock);
16705
16706 if (dtrace_retained == NULL) {
16707 mutex_exit(&dtrace_lock);
16708 return;
16709 }
16710
16711 (void) taskq_dispatch(dtrace_taskq,
16712 (task_func_t *)dtrace_enabling_matchall, NULL, TQ_SLEEP);
16713
16714 mutex_exit(&dtrace_lock);
16715
16716 /*
16717 * And now, for a little heuristic sleaze: in general, we want to
16718 * match modules as soon as they load. However, we cannot guarantee
16719 * this, because it would lead us to the lock ordering violation
16720 * outlined above. The common case, of course, is that cpu_lock is
16721 * _not_ held -- so we delay here for a clock tick, hoping that that's
16722 * long enough for the task queue to do its work. If it's not, it's
16723 * not a serious problem -- it just means that the module that we
16724 * just loaded may not be immediately instrumentable.
16725 */
16726 delay(1);
16727 }
16728
16729 static void
16730 #ifdef illumos
16731 dtrace_module_unloaded(modctl_t *ctl)
16732 #else
16733 dtrace_module_unloaded(modctl_t *ctl, int *error)
16734 #endif
16735 {
16736 dtrace_probe_t template, *probe, *first, *next;
16737 dtrace_provider_t *prov;
16738 #ifndef illumos
16739 char modname[DTRACE_MODNAMELEN];
16740 size_t len;
16741 #endif
16742
16743 #ifdef illumos
16744 template.dtpr_mod = ctl->mod_modname;
16745 #else
16746 /* Handle the fact that ctl->filename may end in ".ko". */
16747 strlcpy(modname, ctl->filename, sizeof(modname));
16748 len = strlen(ctl->filename);
16749 if (len > 3 && strcmp(modname + len - 3, ".ko") == 0)
16750 modname[len - 3] = '\0';
16751 template.dtpr_mod = modname;
16752 #endif
16753
16754 mutex_enter(&dtrace_provider_lock);
16755 #ifdef illumos
16756 mutex_enter(&mod_lock);
16757 #endif
16758 mutex_enter(&dtrace_lock);
16759
16760 #ifndef illumos
16761 if (ctl->nenabled > 0) {
16762 /* Don't allow unloads if a probe is enabled. */
16763 mutex_exit(&dtrace_provider_lock);
16764 mutex_exit(&dtrace_lock);
16765 *error = -1;
16766 printf(
16767 "kldunload: attempt to unload module that has DTrace probes enabled\n");
16768 return;
16769 }
16770 #endif
16771
16772 if (dtrace_bymod == NULL) {
16773 /*
16774 * The DTrace module is loaded (obviously) but not attached;
16775 * we don't have any work to do.
16776 */
16777 mutex_exit(&dtrace_provider_lock);
16778 #ifdef illumos
16779 mutex_exit(&mod_lock);
16780 #endif
16781 mutex_exit(&dtrace_lock);
16782 return;
16783 }
16784
16785 for (probe = first = dtrace_hash_lookup(dtrace_bymod, &template);
16786 probe != NULL; probe = probe->dtpr_nextmod) {
16787 if (probe->dtpr_ecb != NULL) {
16788 mutex_exit(&dtrace_provider_lock);
16789 #ifdef illumos
16790 mutex_exit(&mod_lock);
16791 #endif
16792 mutex_exit(&dtrace_lock);
16793
16794 /*
16795 * This shouldn't _actually_ be possible -- we're
16796 * unloading a module that has an enabled probe in it.
16797 * (It's normally up to the provider to make sure that
16798 * this can't happen.) However, because dtps_enable()
16799 * doesn't have a failure mode, there can be an
16800 * enable/unload race. Upshot: we don't want to
16801 * assert, but we're not going to disable the
16802 * probe, either.
16803 */
16804 if (dtrace_err_verbose) {
16805 #ifdef illumos
16806 cmn_err(CE_WARN, "unloaded module '%s' had "
16807 "enabled probes", ctl->mod_modname);
16808 #else
16809 cmn_err(CE_WARN, "unloaded module '%s' had "
16810 "enabled probes", modname);
16811 #endif
16812 }
16813
16814 return;
16815 }
16816 }
16817
16818 probe = first;
16819
16820 for (first = NULL; probe != NULL; probe = next) {
16821 ASSERT(dtrace_probes[probe->dtpr_id - 1] == probe);
16822
16823 dtrace_probes[probe->dtpr_id - 1] = NULL;
16824
16825 next = probe->dtpr_nextmod;
16826 dtrace_hash_remove(dtrace_bymod, probe);
16827 dtrace_hash_remove(dtrace_byfunc, probe);
16828 dtrace_hash_remove(dtrace_byname, probe);
16829
16830 if (first == NULL) {
16831 first = probe;
16832 probe->dtpr_nextmod = NULL;
16833 } else {
16834 probe->dtpr_nextmod = first;
16835 first = probe;
16836 }
16837 }
16838
16839 /*
16840 * We've removed all of the module's probes from the hash chains and
16841 * from the probe array. Now issue a dtrace_sync() to be sure that
16842 * everyone has cleared out from any probe array processing.
16843 */
16844 dtrace_sync();
16845
16846 for (probe = first; probe != NULL; probe = first) {
16847 first = probe->dtpr_nextmod;
16848 prov = probe->dtpr_provider;
16849 prov->dtpv_pops.dtps_destroy(prov->dtpv_arg, probe->dtpr_id,
16850 probe->dtpr_arg);
16851 kmem_free(probe->dtpr_mod, strlen(probe->dtpr_mod) + 1);
16852 kmem_free(probe->dtpr_func, strlen(probe->dtpr_func) + 1);
16853 kmem_free(probe->dtpr_name, strlen(probe->dtpr_name) + 1);
16854 #ifdef illumos
16855 vmem_free(dtrace_arena, (void *)(uintptr_t)probe->dtpr_id, 1);
16856 #else
16857 free_unr(dtrace_arena, probe->dtpr_id);
16858 #endif
16859 kmem_free(probe, sizeof (dtrace_probe_t));
16860 }
16861
16862 mutex_exit(&dtrace_lock);
16863 #ifdef illumos
16864 mutex_exit(&mod_lock);
16865 #endif
16866 mutex_exit(&dtrace_provider_lock);
16867 }
16868
16869 #ifndef illumos
16870 static void
16871 dtrace_kld_load(void *arg __unused, linker_file_t lf)
16872 {
16873
16874 dtrace_module_loaded(lf);
16875 }
16876
16877 static void
16878 dtrace_kld_unload_try(void *arg __unused, linker_file_t lf, int *error)
16879 {
16880
16881 if (*error != 0)
16882 /* We already have an error, so don't do anything. */
16883 return;
16884 dtrace_module_unloaded(lf, error);
16885 }
16886 #endif
16887
16888 #ifdef illumos
16889 static void
16890 dtrace_suspend(void)
16891 {
16892 dtrace_probe_foreach(offsetof(dtrace_pops_t, dtps_suspend));
16893 }
16894
16895 static void
16896 dtrace_resume(void)
16897 {
16898 dtrace_probe_foreach(offsetof(dtrace_pops_t, dtps_resume));
16899 }
16900 #endif
16901
16902 static int
16903 dtrace_cpu_setup(cpu_setup_t what, processorid_t cpu)
16904 {
16905 ASSERT(MUTEX_HELD(&cpu_lock));
16906 mutex_enter(&dtrace_lock);
16907
16908 switch (what) {
16909 case CPU_CONFIG: {
16910 dtrace_state_t *state;
16911 dtrace_optval_t *opt, rs, c;
16912
16913 /*
16914 * For now, we only allocate a new buffer for anonymous state.
16915 */
16916 if ((state = dtrace_anon.dta_state) == NULL)
16917 break;
16918
16919 if (state->dts_activity != DTRACE_ACTIVITY_ACTIVE)
16920 break;
16921
16922 opt = state->dts_options;
16923 c = opt[DTRACEOPT_CPU];
16924
16925 if (c != DTRACE_CPUALL && c != DTRACEOPT_UNSET && c != cpu)
16926 break;
16927
16928 /*
16929 * Regardless of what the actual policy is, we're going to
16930 * temporarily set our resize policy to be manual. We're
16931 * also going to temporarily set our CPU option to denote
16932 * the newly configured CPU.
16933 */
16934 rs = opt[DTRACEOPT_BUFRESIZE];
16935 opt[DTRACEOPT_BUFRESIZE] = DTRACEOPT_BUFRESIZE_MANUAL;
16936 opt[DTRACEOPT_CPU] = (dtrace_optval_t)cpu;
16937
16938 (void) dtrace_state_buffers(state);
16939
16940 opt[DTRACEOPT_BUFRESIZE] = rs;
16941 opt[DTRACEOPT_CPU] = c;
16942
16943 break;
16944 }
16945
16946 case CPU_UNCONFIG:
16947 /*
16948 * We don't free the buffer in the CPU_UNCONFIG case. (The
16949 * buffer will be freed when the consumer exits.)
16950 */
16951 break;
16952
16953 default:
16954 break;
16955 }
16956
16957 mutex_exit(&dtrace_lock);
16958 return (0);
16959 }
16960
16961 #ifdef illumos
16962 static void
16963 dtrace_cpu_setup_initial(processorid_t cpu)
16964 {
16965 (void) dtrace_cpu_setup(CPU_CONFIG, cpu);
16966 }
16967 #endif
16968
16969 static void
16970 dtrace_toxrange_add(uintptr_t base, uintptr_t limit)
16971 {
16972 if (dtrace_toxranges >= dtrace_toxranges_max) {
16973 int osize, nsize;
16974 dtrace_toxrange_t *range;
16975
16976 osize = dtrace_toxranges_max * sizeof (dtrace_toxrange_t);
16977
16978 if (osize == 0) {
16979 ASSERT(dtrace_toxrange == NULL);
16980 ASSERT(dtrace_toxranges_max == 0);
16981 dtrace_toxranges_max = 1;
16982 } else {
16983 dtrace_toxranges_max <<= 1;
16984 }
16985
16986 nsize = dtrace_toxranges_max * sizeof (dtrace_toxrange_t);
16987 range = kmem_zalloc(nsize, KM_SLEEP);
16988
16989 if (dtrace_toxrange != NULL) {
16990 ASSERT(osize != 0);
16991 bcopy(dtrace_toxrange, range, osize);
16992 kmem_free(dtrace_toxrange, osize);
16993 }
16994
16995 dtrace_toxrange = range;
16996 }
16997
16998 ASSERT(dtrace_toxrange[dtrace_toxranges].dtt_base == 0);
16999 ASSERT(dtrace_toxrange[dtrace_toxranges].dtt_limit == 0);
17000
17001 dtrace_toxrange[dtrace_toxranges].dtt_base = base;
17002 dtrace_toxrange[dtrace_toxranges].dtt_limit = limit;
17003 dtrace_toxranges++;
17004 }
17005
17006 static void
17007 dtrace_getf_barrier(void)
17008 {
17009 #ifdef illumos
17010 /*
17011 * When we have unprivileged (that is, non-DTRACE_CRV_KERNEL) enablings
17012 * that contain calls to getf(), this routine will be called on every
17013 * closef() before either the underlying vnode is released or the
17014 * file_t itself is freed. By the time we are here, it is essential
17015 * that the file_t can no longer be accessed from a call to getf()
17016 * in probe context -- that assures that a dtrace_sync() can be used
17017 * to clear out any enablings referring to the old structures.
17018 */
17019 if (curthread->t_procp->p_zone->zone_dtrace_getf != 0 ||
17020 kcred->cr_zone->zone_dtrace_getf != 0)
17021 dtrace_sync();
17022 #endif
17023 }
17024
17025 /*
17026 * DTrace Driver Cookbook Functions
17027 */
17028 #ifdef illumos
17029 /*ARGSUSED*/
17030 static int
17031 dtrace_attach(dev_info_t *devi, ddi_attach_cmd_t cmd)
17032 {
17033 dtrace_provider_id_t id;
17034 dtrace_state_t *state = NULL;
17035 dtrace_enabling_t *enab;
17036
17037 mutex_enter(&cpu_lock);
17038 mutex_enter(&dtrace_provider_lock);
17039 mutex_enter(&dtrace_lock);
17040
17041 if (ddi_soft_state_init(&dtrace_softstate,
17042 sizeof (dtrace_state_t), 0) != 0) {
17043 cmn_err(CE_NOTE, "/dev/dtrace failed to initialize soft state");
17044 mutex_exit(&cpu_lock);
17045 mutex_exit(&dtrace_provider_lock);
17046 mutex_exit(&dtrace_lock);
17047 return (DDI_FAILURE);
17048 }
17049
17050 if (ddi_create_minor_node(devi, DTRACEMNR_DTRACE, S_IFCHR,
17051 DTRACEMNRN_DTRACE, DDI_PSEUDO, NULL) == DDI_FAILURE ||
17052 ddi_create_minor_node(devi, DTRACEMNR_HELPER, S_IFCHR,
17053 DTRACEMNRN_HELPER, DDI_PSEUDO, NULL) == DDI_FAILURE) {
17054 cmn_err(CE_NOTE, "/dev/dtrace couldn't create minor nodes");
17055 ddi_remove_minor_node(devi, NULL);
17056 ddi_soft_state_fini(&dtrace_softstate);
17057 mutex_exit(&cpu_lock);
17058 mutex_exit(&dtrace_provider_lock);
17059 mutex_exit(&dtrace_lock);
17060 return (DDI_FAILURE);
17061 }
17062
17063 ddi_report_dev(devi);
17064 dtrace_devi = devi;
17065
17066 dtrace_modload = dtrace_module_loaded;
17067 dtrace_modunload = dtrace_module_unloaded;
17068 dtrace_cpu_init = dtrace_cpu_setup_initial;
17069 dtrace_helpers_cleanup = dtrace_helpers_destroy;
17070 dtrace_helpers_fork = dtrace_helpers_duplicate;
17071 dtrace_cpustart_init = dtrace_suspend;
17072 dtrace_cpustart_fini = dtrace_resume;
17073 dtrace_debugger_init = dtrace_suspend;
17074 dtrace_debugger_fini = dtrace_resume;
17075
17076 register_cpu_setup_func((cpu_setup_func_t *)dtrace_cpu_setup, NULL);
17077
17078 ASSERT(MUTEX_HELD(&cpu_lock));
17079
17080 dtrace_arena = vmem_create("dtrace", (void *)1, UINT32_MAX, 1,
17081 NULL, NULL, NULL, 0, VM_SLEEP | VMC_IDENTIFIER);
17082 dtrace_minor = vmem_create("dtrace_minor", (void *)DTRACEMNRN_CLONE,
17083 UINT32_MAX - DTRACEMNRN_CLONE, 1, NULL, NULL, NULL, 0,
17084 VM_SLEEP | VMC_IDENTIFIER);
17085 dtrace_taskq = taskq_create("dtrace_taskq", 1, maxclsyspri,
17086 1, INT_MAX, 0);
17087
17088 dtrace_state_cache = kmem_cache_create("dtrace_state_cache",
17089 sizeof (dtrace_dstate_percpu_t) * NCPU, DTRACE_STATE_ALIGN,
17090 NULL, NULL, NULL, NULL, NULL, 0);
17091
17092 ASSERT(MUTEX_HELD(&cpu_lock));
17093 dtrace_bymod = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_mod),
17094 offsetof(dtrace_probe_t, dtpr_nextmod),
17095 offsetof(dtrace_probe_t, dtpr_prevmod));
17096
17097 dtrace_byfunc = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_func),
17098 offsetof(dtrace_probe_t, dtpr_nextfunc),
17099 offsetof(dtrace_probe_t, dtpr_prevfunc));
17100
17101 dtrace_byname = dtrace_hash_create(offsetof(dtrace_probe_t, dtpr_name),
17102 offsetof(dtrace_probe_t, dtpr_nextname),
17103 offsetof(dtrace_probe_t, dtpr_prevname));
17104
17105 if (dtrace_retain_max < 1) {
17106 cmn_err(CE_WARN, "illegal value (%zu) for dtrace_retain_max; "
17107 "setting to 1", dtrace_retain_max);
17108 dtrace_retain_max = 1;
17109 }
17110
17111 /*
17112 * Now discover our toxic ranges.
17113 */
17114 dtrace_toxic_ranges(dtrace_toxrange_add);
17115
17116 /*
17117 * Before we register ourselves as a provider to our own framework,
17118 * we would like to assert that dtrace_provider is NULL -- but that's
17119 * not true if we were loaded as a dependency of a DTrace provider.
17120 * Once we've registered, we can assert that dtrace_provider is our
17121 * pseudo provider.
17122 */
17123 (void) dtrace_register("dtrace", &dtrace_provider_attr,
17124 DTRACE_PRIV_NONE, 0, &dtrace_provider_ops, NULL, &id);
17125
17126 ASSERT(dtrace_provider != NULL);
17127 ASSERT((dtrace_provider_id_t)dtrace_provider == id);
17128
17129 dtrace_probeid_begin = dtrace_probe_create((dtrace_provider_id_t)
17130 dtrace_provider, NULL, NULL, "BEGIN", 0, NULL);
17131 dtrace_probeid_end = dtrace_probe_create((dtrace_provider_id_t)
17132 dtrace_provider, NULL, NULL, "END", 0, NULL);
17133 dtrace_probeid_error = dtrace_probe_create((dtrace_provider_id_t)
17134 dtrace_provider, NULL, NULL, "ERROR", 1, NULL);
17135
17136 dtrace_anon_property();
17137 mutex_exit(&cpu_lock);
17138
17139 /*
17140 * If there are already providers, we must ask them to provide their
17141 * probes, and then match any anonymous enabling against them. Note
17142 * that there should be no other retained enablings at this time:
17143 * the only retained enablings at this time should be the anonymous
17144 * enabling.
17145 */
17146 if (dtrace_anon.dta_enabling != NULL) {
17147 ASSERT(dtrace_retained == dtrace_anon.dta_enabling);
17148
17149 dtrace_enabling_provide(NULL);
17150 state = dtrace_anon.dta_state;
17151
17152 /*
17153 * We couldn't hold cpu_lock across the above call to
17154 * dtrace_enabling_provide(), but we must hold it to actually
17155 * enable the probes. We have to drop all of our locks, pick
17156 * up cpu_lock, and regain our locks before matching the
17157 * retained anonymous enabling.
17158 */
17159 mutex_exit(&dtrace_lock);
17160 mutex_exit(&dtrace_provider_lock);
17161
17162 mutex_enter(&cpu_lock);
17163 mutex_enter(&dtrace_provider_lock);
17164 mutex_enter(&dtrace_lock);
17165
17166 if ((enab = dtrace_anon.dta_enabling) != NULL)
17167 (void) dtrace_enabling_match(enab, NULL);
17168
17169 mutex_exit(&cpu_lock);
17170 }
17171
17172 mutex_exit(&dtrace_lock);
17173 mutex_exit(&dtrace_provider_lock);
17174
17175 if (state != NULL) {
17176 /*
17177 * If we created any anonymous state, set it going now.
17178 */
17179 (void) dtrace_state_go(state, &dtrace_anon.dta_beganon);
17180 }
17181
17182 return (DDI_SUCCESS);
17183 }
17184 #endif /* illumos */
17185
17186 #ifndef illumos
17187 static void dtrace_dtr(void *);
17188 #endif
17189
17190 /*ARGSUSED*/
17191 static int
17192 #ifdef illumos
17193 dtrace_open(dev_t *devp, int flag, int otyp, cred_t *cred_p)
17194 #else
17195 dtrace_open(struct cdev *dev, int oflags, int devtype, struct thread *td)
17196 #endif
17197 {
17198 dtrace_state_t *state;
17199 uint32_t priv;
17200 uid_t uid;
17201 zoneid_t zoneid;
17202
17203 #ifdef illumos
17204 if (getminor(*devp) == DTRACEMNRN_HELPER)
17205 return (0);
17206
17207 /*
17208 * If this wasn't an open with the "helper" minor, then it must be
17209 * the "dtrace" minor.
17210 */
17211 if (getminor(*devp) == DTRACEMNRN_DTRACE)
17212 return (ENXIO);
17213 #else
17214 cred_t *cred_p = NULL;
17215 cred_p = dev->si_cred;
17216
17217 /*
17218 * If no DTRACE_PRIV_* bits are set in the credential, then the
17219 * caller lacks sufficient permission to do anything with DTrace.
17220 */
17221 dtrace_cred2priv(cred_p, &priv, &uid, &zoneid);
17222 if (priv == DTRACE_PRIV_NONE) {
17223 #endif
17224
17225 return (EACCES);
17226 }
17227
17228 /*
17229 * Ask all providers to provide all their probes.
17230 */
17231 mutex_enter(&dtrace_provider_lock);
17232 dtrace_probe_provide(NULL, NULL);
17233 mutex_exit(&dtrace_provider_lock);
17234
17235 mutex_enter(&cpu_lock);
17236 mutex_enter(&dtrace_lock);
17237 dtrace_opens++;
17238 dtrace_membar_producer();
17239
17240 #ifdef illumos
17241 /*
17242 * If the kernel debugger is active (that is, if the kernel debugger
17243 * modified text in some way), we won't allow the open.
17244 */
17245 if (kdi_dtrace_set(KDI_DTSET_DTRACE_ACTIVATE) != 0) {
17246 dtrace_opens--;
17247 mutex_exit(&cpu_lock);
17248 mutex_exit(&dtrace_lock);
17249 return (EBUSY);
17250 }
17251
17252 if (dtrace_helptrace_enable && dtrace_helptrace_buffer == NULL) {
17253 /*
17254 * If DTrace helper tracing is enabled, we need to allocate the
17255 * trace buffer and initialize the values.
17256 */
17257 dtrace_helptrace_buffer =
17258 kmem_zalloc(dtrace_helptrace_bufsize, KM_SLEEP);
17259 dtrace_helptrace_next = 0;
17260 dtrace_helptrace_wrapped = 0;
17261 dtrace_helptrace_enable = 0;
17262 }
17263
17264 state = dtrace_state_create(devp, cred_p);
17265 #else
17266 state = dtrace_state_create(dev, NULL);
17267 devfs_set_cdevpriv(state, dtrace_dtr);
17268 #endif
17269
17270 mutex_exit(&cpu_lock);
17271
17272 if (state == NULL) {
17273 #ifdef illumos
17274 if (--dtrace_opens == 0 && dtrace_anon.dta_enabling == NULL)
17275 (void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
17276 #else
17277 --dtrace_opens;
17278 #endif
17279 mutex_exit(&dtrace_lock);
17280 return (EAGAIN);
17281 }
17282
17283 mutex_exit(&dtrace_lock);
17284
17285 return (0);
17286 }
17287
17288 /*ARGSUSED*/
17289 #ifdef illumos
17290 static int
17291 dtrace_close(dev_t dev, int flag, int otyp, cred_t *cred_p)
17292 #else
17293 static void
17294 dtrace_dtr(void *data)
17295 #endif
17296 {
17297 #ifdef illumos
17298 minor_t minor = getminor(dev);
17299 dtrace_state_t *state;
17300 #endif
17301 dtrace_helptrace_t *buf = NULL;
17302
17303 #ifdef illumos
17304 if (minor == DTRACEMNRN_HELPER)
17305 return (0);
17306
17307 state = ddi_get_soft_state(dtrace_softstate, minor);
17308 #else
17309 dtrace_state_t *state = data;
17310 #endif
17311
17312 mutex_enter(&cpu_lock);
17313 mutex_enter(&dtrace_lock);
17314
17315 #ifdef illumos
17316 if (state->dts_anon)
17317 #else
17318 if (state != NULL && state->dts_anon)
17319 #endif
17320 {
17321 /*
17322 * There is anonymous state. Destroy that first.
17323 */
17324 ASSERT(dtrace_anon.dta_state == NULL);
17325 dtrace_state_destroy(state->dts_anon);
17326 }
17327
17328 if (dtrace_helptrace_disable) {
17329 /*
17330 * If we have been told to disable helper tracing, set the
17331 * buffer to NULL before calling into dtrace_state_destroy();
17332 * we take advantage of its dtrace_sync() to know that no
17333 * CPU is in probe context with enabled helper tracing
17334 * after it returns.
17335 */
17336 buf = dtrace_helptrace_buffer;
17337 dtrace_helptrace_buffer = NULL;
17338 }
17339
17340 #ifdef illumos
17341 dtrace_state_destroy(state);
17342 #else
17343 if (state != NULL) {
17344 dtrace_state_destroy(state);
17345 kmem_free(state, 0);
17346 }
17347 #endif
17348 ASSERT(dtrace_opens > 0);
17349
17350 #ifdef illumos
17351 /*
17352 * Only relinquish control of the kernel debugger interface when there
17353 * are no consumers and no anonymous enablings.
17354 */
17355 if (--dtrace_opens == 0 && dtrace_anon.dta_enabling == NULL)
17356 (void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
17357 #else
17358 --dtrace_opens;
17359 #endif
17360
17361 if (buf != NULL) {
17362 kmem_free(buf, dtrace_helptrace_bufsize);
17363 dtrace_helptrace_disable = 0;
17364 }
17365
17366 mutex_exit(&dtrace_lock);
17367 mutex_exit(&cpu_lock);
17368
17369 #ifdef illumos
17370 return (0);
17371 #endif
17372 }
17373
17374 #ifdef illumos
17375 /*ARGSUSED*/
17376 static int
17377 dtrace_ioctl_helper(int cmd, intptr_t arg, int *rv)
17378 {
17379 int rval;
17380 dof_helper_t help, *dhp = NULL;
17381
17382 switch (cmd) {
17383 case DTRACEHIOC_ADDDOF:
17384 if (copyin((void *)arg, &help, sizeof (help)) != 0) {
17385 dtrace_dof_error(NULL, "failed to copyin DOF helper");
17386 return (EFAULT);
17387 }
17388
17389 dhp = &help;
17390 arg = (intptr_t)help.dofhp_dof;
17391 /*FALLTHROUGH*/
17392
17393 case DTRACEHIOC_ADD: {
17394 dof_hdr_t *dof = dtrace_dof_copyin(arg, &rval);
17395
17396 if (dof == NULL)
17397 return (rval);
17398
17399 mutex_enter(&dtrace_lock);
17400
17401 /*
17402 * dtrace_helper_slurp() takes responsibility for the dof --
17403 * it may free it now or it may save it and free it later.
17404 */
17405 if ((rval = dtrace_helper_slurp(dof, dhp)) != -1) {
17406 *rv = rval;
17407 rval = 0;
17408 } else {
17409 rval = EINVAL;
17410 }
17411
17412 mutex_exit(&dtrace_lock);
17413 return (rval);
17414 }
17415
17416 case DTRACEHIOC_REMOVE: {
17417 mutex_enter(&dtrace_lock);
17418 rval = dtrace_helper_destroygen(NULL, arg);
17419 mutex_exit(&dtrace_lock);
17420
17421 return (rval);
17422 }
17423
17424 default:
17425 break;
17426 }
17427
17428 return (ENOTTY);
17429 }
17430
17431 /*ARGSUSED*/
17432 static int
17433 dtrace_ioctl(dev_t dev, int cmd, intptr_t arg, int md, cred_t *cr, int *rv)
17434 {
17435 minor_t minor = getminor(dev);
17436 dtrace_state_t *state;
17437 int rval;
17438
17439 if (minor == DTRACEMNRN_HELPER)
17440 return (dtrace_ioctl_helper(cmd, arg, rv));
17441
17442 state = ddi_get_soft_state(dtrace_softstate, minor);
17443
17444 if (state->dts_anon) {
17445 ASSERT(dtrace_anon.dta_state == NULL);
17446 state = state->dts_anon;
17447 }
17448
17449 switch (cmd) {
17450 case DTRACEIOC_PROVIDER: {
17451 dtrace_providerdesc_t pvd;
17452 dtrace_provider_t *pvp;
17453
17454 if (copyin((void *)arg, &pvd, sizeof (pvd)) != 0)
17455 return (EFAULT);
17456
17457 pvd.dtvd_name[DTRACE_PROVNAMELEN - 1] = '\0';
17458 mutex_enter(&dtrace_provider_lock);
17459
17460 for (pvp = dtrace_provider; pvp != NULL; pvp = pvp->dtpv_next) {
17461 if (strcmp(pvp->dtpv_name, pvd.dtvd_name) == 0)
17462 break;
17463 }
17464
17465 mutex_exit(&dtrace_provider_lock);
17466
17467 if (pvp == NULL)
17468 return (ESRCH);
17469
17470 bcopy(&pvp->dtpv_priv, &pvd.dtvd_priv, sizeof (dtrace_ppriv_t));
17471 bcopy(&pvp->dtpv_attr, &pvd.dtvd_attr, sizeof (dtrace_pattr_t));
17472
17473 if (copyout(&pvd, (void *)arg, sizeof (pvd)) != 0)
17474 return (EFAULT);
17475
17476 return (0);
17477 }
17478
17479 case DTRACEIOC_EPROBE: {
17480 dtrace_eprobedesc_t epdesc;
17481 dtrace_ecb_t *ecb;
17482 dtrace_action_t *act;
17483 void *buf;
17484 size_t size;
17485 uintptr_t dest;
17486 int nrecs;
17487
17488 if (copyin((void *)arg, &epdesc, sizeof (epdesc)) != 0)
17489 return (EFAULT);
17490
17491 mutex_enter(&dtrace_lock);
17492
17493 if ((ecb = dtrace_epid2ecb(state, epdesc.dtepd_epid)) == NULL) {
17494 mutex_exit(&dtrace_lock);
17495 return (EINVAL);
17496 }
17497
17498 if (ecb->dte_probe == NULL) {
17499 mutex_exit(&dtrace_lock);
17500 return (EINVAL);
17501 }
17502
17503 epdesc.dtepd_probeid = ecb->dte_probe->dtpr_id;
17504 epdesc.dtepd_uarg = ecb->dte_uarg;
17505 epdesc.dtepd_size = ecb->dte_size;
17506
17507 nrecs = epdesc.dtepd_nrecs;
17508 epdesc.dtepd_nrecs = 0;
17509 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
17510 if (DTRACEACT_ISAGG(act->dta_kind) || act->dta_intuple)
17511 continue;
17512
17513 epdesc.dtepd_nrecs++;
17514 }
17515
17516 /*
17517 * Now that we have the size, we need to allocate a temporary
17518 * buffer in which to store the complete description. We need
17519 * the temporary buffer to be able to drop dtrace_lock()
17520 * across the copyout(), below.
17521 */
17522 size = sizeof (dtrace_eprobedesc_t) +
17523 (epdesc.dtepd_nrecs * sizeof (dtrace_recdesc_t));
17524
17525 buf = kmem_alloc(size, KM_SLEEP);
17526 dest = (uintptr_t)buf;
17527
17528 bcopy(&epdesc, (void *)dest, sizeof (epdesc));
17529 dest += offsetof(dtrace_eprobedesc_t, dtepd_rec[0]);
17530
17531 for (act = ecb->dte_action; act != NULL; act = act->dta_next) {
17532 if (DTRACEACT_ISAGG(act->dta_kind) || act->dta_intuple)
17533 continue;
17534
17535 if (nrecs-- == 0)
17536 break;
17537
17538 bcopy(&act->dta_rec, (void *)dest,
17539 sizeof (dtrace_recdesc_t));
17540 dest += sizeof (dtrace_recdesc_t);
17541 }
17542
17543 mutex_exit(&dtrace_lock);
17544
17545 if (copyout(buf, (void *)arg, dest - (uintptr_t)buf) != 0) {
17546 kmem_free(buf, size);
17547 return (EFAULT);
17548 }
17549
17550 kmem_free(buf, size);
17551 return (0);
17552 }
17553
17554 case DTRACEIOC_AGGDESC: {
17555 dtrace_aggdesc_t aggdesc;
17556 dtrace_action_t *act;
17557 dtrace_aggregation_t *agg;
17558 int nrecs;
17559 uint32_t offs;
17560 dtrace_recdesc_t *lrec;
17561 void *buf;
17562 size_t size;
17563 uintptr_t dest;
17564
17565 if (copyin((void *)arg, &aggdesc, sizeof (aggdesc)) != 0)
17566 return (EFAULT);
17567
17568 mutex_enter(&dtrace_lock);
17569
17570 if ((agg = dtrace_aggid2agg(state, aggdesc.dtagd_id)) == NULL) {
17571 mutex_exit(&dtrace_lock);
17572 return (EINVAL);
17573 }
17574
17575 aggdesc.dtagd_epid = agg->dtag_ecb->dte_epid;
17576
17577 nrecs = aggdesc.dtagd_nrecs;
17578 aggdesc.dtagd_nrecs = 0;
17579
17580 offs = agg->dtag_base;
17581 lrec = &agg->dtag_action.dta_rec;
17582 aggdesc.dtagd_size = lrec->dtrd_offset + lrec->dtrd_size - offs;
17583
17584 for (act = agg->dtag_first; ; act = act->dta_next) {
17585 ASSERT(act->dta_intuple ||
17586 DTRACEACT_ISAGG(act->dta_kind));
17587
17588 /*
17589 * If this action has a record size of zero, it
17590 * denotes an argument to the aggregating action.
17591 * Because the presence of this record doesn't (or
17592 * shouldn't) affect the way the data is interpreted,
17593 * we don't copy it out to save user-level the
17594 * confusion of dealing with a zero-length record.
17595 */
17596 if (act->dta_rec.dtrd_size == 0) {
17597 ASSERT(agg->dtag_hasarg);
17598 continue;
17599 }
17600
17601 aggdesc.dtagd_nrecs++;
17602
17603 if (act == &agg->dtag_action)
17604 break;
17605 }
17606
17607 /*
17608 * Now that we have the size, we need to allocate a temporary
17609 * buffer in which to store the complete description. We need
17610 * the temporary buffer to be able to drop dtrace_lock()
17611 * across the copyout(), below.
17612 */
17613 size = sizeof (dtrace_aggdesc_t) +
17614 (aggdesc.dtagd_nrecs * sizeof (dtrace_recdesc_t));
17615
17616 buf = kmem_alloc(size, KM_SLEEP);
17617 dest = (uintptr_t)buf;
17618
17619 bcopy(&aggdesc, (void *)dest, sizeof (aggdesc));
17620 dest += offsetof(dtrace_aggdesc_t, dtagd_rec[0]);
17621
17622 for (act = agg->dtag_first; ; act = act->dta_next) {
17623 dtrace_recdesc_t rec = act->dta_rec;
17624
17625 /*
17626 * See the comment in the above loop for why we pass
17627 * over zero-length records.
17628 */
17629 if (rec.dtrd_size == 0) {
17630 ASSERT(agg->dtag_hasarg);
17631 continue;
17632 }
17633
17634 if (nrecs-- == 0)
17635 break;
17636
17637 rec.dtrd_offset -= offs;
17638 bcopy(&rec, (void *)dest, sizeof (rec));
17639 dest += sizeof (dtrace_recdesc_t);
17640
17641 if (act == &agg->dtag_action)
17642 break;
17643 }
17644
17645 mutex_exit(&dtrace_lock);
17646
17647 if (copyout(buf, (void *)arg, dest - (uintptr_t)buf) != 0) {
17648 kmem_free(buf, size);
17649 return (EFAULT);
17650 }
17651
17652 kmem_free(buf, size);
17653 return (0);
17654 }
17655
17656 case DTRACEIOC_ENABLE: {
17657 dof_hdr_t *dof;
17658 dtrace_enabling_t *enab = NULL;
17659 dtrace_vstate_t *vstate;
17660 int err = 0;
17661
17662 *rv = 0;
17663
17664 /*
17665 * If a NULL argument has been passed, we take this as our
17666 * cue to reevaluate our enablings.
17667 */
17668 if (arg == NULL) {
17669 dtrace_enabling_matchall();
17670
17671 return (0);
17672 }
17673
17674 if ((dof = dtrace_dof_copyin(arg, &rval)) == NULL)
17675 return (rval);
17676
17677 mutex_enter(&cpu_lock);
17678 mutex_enter(&dtrace_lock);
17679 vstate = &state->dts_vstate;
17680
17681 if (state->dts_activity != DTRACE_ACTIVITY_INACTIVE) {
17682 mutex_exit(&dtrace_lock);
17683 mutex_exit(&cpu_lock);
17684 dtrace_dof_destroy(dof);
17685 return (EBUSY);
17686 }
17687
17688 if (dtrace_dof_slurp(dof, vstate, cr, &enab, 0, B_TRUE) != 0) {
17689 mutex_exit(&dtrace_lock);
17690 mutex_exit(&cpu_lock);
17691 dtrace_dof_destroy(dof);
17692 return (EINVAL);
17693 }
17694
17695 if ((rval = dtrace_dof_options(dof, state)) != 0) {
17696 dtrace_enabling_destroy(enab);
17697 mutex_exit(&dtrace_lock);
17698 mutex_exit(&cpu_lock);
17699 dtrace_dof_destroy(dof);
17700 return (rval);
17701 }
17702
17703 if ((err = dtrace_enabling_match(enab, rv)) == 0) {
17704 err = dtrace_enabling_retain(enab);
17705 } else {
17706 dtrace_enabling_destroy(enab);
17707 }
17708
17709 mutex_exit(&cpu_lock);
17710 mutex_exit(&dtrace_lock);
17711 dtrace_dof_destroy(dof);
17712
17713 return (err);
17714 }
17715
17716 case DTRACEIOC_REPLICATE: {
17717 dtrace_repldesc_t desc;
17718 dtrace_probedesc_t *match = &desc.dtrpd_match;
17719 dtrace_probedesc_t *create = &desc.dtrpd_create;
17720 int err;
17721
17722 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
17723 return (EFAULT);
17724
17725 match->dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
17726 match->dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
17727 match->dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
17728 match->dtpd_name[DTRACE_NAMELEN - 1] = '\0';
17729
17730 create->dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
17731 create->dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
17732 create->dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
17733 create->dtpd_name[DTRACE_NAMELEN - 1] = '\0';
17734
17735 mutex_enter(&dtrace_lock);
17736 err = dtrace_enabling_replicate(state, match, create);
17737 mutex_exit(&dtrace_lock);
17738
17739 return (err);
17740 }
17741
17742 case DTRACEIOC_PROBEMATCH:
17743 case DTRACEIOC_PROBES: {
17744 dtrace_probe_t *probe = NULL;
17745 dtrace_probedesc_t desc;
17746 dtrace_probekey_t pkey;
17747 dtrace_id_t i;
17748 int m = 0;
17749 uint32_t priv;
17750 uid_t uid;
17751 zoneid_t zoneid;
17752
17753 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
17754 return (EFAULT);
17755
17756 desc.dtpd_provider[DTRACE_PROVNAMELEN - 1] = '\0';
17757 desc.dtpd_mod[DTRACE_MODNAMELEN - 1] = '\0';
17758 desc.dtpd_func[DTRACE_FUNCNAMELEN - 1] = '\0';
17759 desc.dtpd_name[DTRACE_NAMELEN - 1] = '\0';
17760
17761 /*
17762 * Before we attempt to match this probe, we want to give
17763 * all providers the opportunity to provide it.
17764 */
17765 if (desc.dtpd_id == DTRACE_IDNONE) {
17766 mutex_enter(&dtrace_provider_lock);
17767 dtrace_probe_provide(&desc, NULL);
17768 mutex_exit(&dtrace_provider_lock);
17769 desc.dtpd_id++;
17770 }
17771
17772 if (cmd == DTRACEIOC_PROBEMATCH) {
17773 dtrace_probekey(&desc, &pkey);
17774 pkey.dtpk_id = DTRACE_IDNONE;
17775 }
17776
17777 dtrace_cred2priv(cr, &priv, &uid, &zoneid);
17778
17779 mutex_enter(&dtrace_lock);
17780
17781 if (cmd == DTRACEIOC_PROBEMATCH) {
17782 for (i = desc.dtpd_id; i <= dtrace_nprobes; i++) {
17783 if ((probe = dtrace_probes[i - 1]) != NULL &&
17784 (m = dtrace_match_probe(probe, &pkey,
17785 priv, uid, zoneid)) != 0)
17786 break;
17787 }
17788
17789 if (m < 0) {
17790 mutex_exit(&dtrace_lock);
17791 return (EINVAL);
17792 }
17793
17794 } else {
17795 for (i = desc.dtpd_id; i <= dtrace_nprobes; i++) {
17796 if ((probe = dtrace_probes[i - 1]) != NULL &&
17797 dtrace_match_priv(probe, priv, uid, zoneid))
17798 break;
17799 }
17800 }
17801
17802 if (probe == NULL) {
17803 mutex_exit(&dtrace_lock);
17804 return (ESRCH);
17805 }
17806
17807 dtrace_probe_description(probe, &desc);
17808 mutex_exit(&dtrace_lock);
17809
17810 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
17811 return (EFAULT);
17812
17813 return (0);
17814 }
17815
17816 case DTRACEIOC_PROBEARG: {
17817 dtrace_argdesc_t desc;
17818 dtrace_probe_t *probe;
17819 dtrace_provider_t *prov;
17820
17821 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
17822 return (EFAULT);
17823
17824 if (desc.dtargd_id == DTRACE_IDNONE)
17825 return (EINVAL);
17826
17827 if (desc.dtargd_ndx == DTRACE_ARGNONE)
17828 return (EINVAL);
17829
17830 mutex_enter(&dtrace_provider_lock);
17831 mutex_enter(&mod_lock);
17832 mutex_enter(&dtrace_lock);
17833
17834 if (desc.dtargd_id > dtrace_nprobes) {
17835 mutex_exit(&dtrace_lock);
17836 mutex_exit(&mod_lock);
17837 mutex_exit(&dtrace_provider_lock);
17838 return (EINVAL);
17839 }
17840
17841 if ((probe = dtrace_probes[desc.dtargd_id - 1]) == NULL) {
17842 mutex_exit(&dtrace_lock);
17843 mutex_exit(&mod_lock);
17844 mutex_exit(&dtrace_provider_lock);
17845 return (EINVAL);
17846 }
17847
17848 mutex_exit(&dtrace_lock);
17849
17850 prov = probe->dtpr_provider;
17851
17852 if (prov->dtpv_pops.dtps_getargdesc == NULL) {
17853 /*
17854 * There isn't any typed information for this probe.
17855 * Set the argument number to DTRACE_ARGNONE.
17856 */
17857 desc.dtargd_ndx = DTRACE_ARGNONE;
17858 } else {
17859 desc.dtargd_native[0] = '\0';
17860 desc.dtargd_xlate[0] = '\0';
17861 desc.dtargd_mapping = desc.dtargd_ndx;
17862
17863 prov->dtpv_pops.dtps_getargdesc(prov->dtpv_arg,
17864 probe->dtpr_id, probe->dtpr_arg, &desc);
17865 }
17866
17867 mutex_exit(&mod_lock);
17868 mutex_exit(&dtrace_provider_lock);
17869
17870 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
17871 return (EFAULT);
17872
17873 return (0);
17874 }
17875
17876 case DTRACEIOC_GO: {
17877 processorid_t cpuid;
17878 rval = dtrace_state_go(state, &cpuid);
17879
17880 if (rval != 0)
17881 return (rval);
17882
17883 if (copyout(&cpuid, (void *)arg, sizeof (cpuid)) != 0)
17884 return (EFAULT);
17885
17886 return (0);
17887 }
17888
17889 case DTRACEIOC_STOP: {
17890 processorid_t cpuid;
17891
17892 mutex_enter(&dtrace_lock);
17893 rval = dtrace_state_stop(state, &cpuid);
17894 mutex_exit(&dtrace_lock);
17895
17896 if (rval != 0)
17897 return (rval);
17898
17899 if (copyout(&cpuid, (void *)arg, sizeof (cpuid)) != 0)
17900 return (EFAULT);
17901
17902 return (0);
17903 }
17904
17905 case DTRACEIOC_DOFGET: {
17906 dof_hdr_t hdr, *dof;
17907 uint64_t len;
17908
17909 if (copyin((void *)arg, &hdr, sizeof (hdr)) != 0)
17910 return (EFAULT);
17911
17912 mutex_enter(&dtrace_lock);
17913 dof = dtrace_dof_create(state);
17914 mutex_exit(&dtrace_lock);
17915
17916 len = MIN(hdr.dofh_loadsz, dof->dofh_loadsz);
17917 rval = copyout(dof, (void *)arg, len);
17918 dtrace_dof_destroy(dof);
17919
17920 return (rval == 0 ? 0 : EFAULT);
17921 }
17922
17923 case DTRACEIOC_AGGSNAP:
17924 case DTRACEIOC_BUFSNAP: {
17925 dtrace_bufdesc_t desc;
17926 caddr_t cached;
17927 dtrace_buffer_t *buf;
17928
17929 if (copyin((void *)arg, &desc, sizeof (desc)) != 0)
17930 return (EFAULT);
17931
17932 if (desc.dtbd_cpu < 0 || desc.dtbd_cpu >= NCPU)
17933 return (EINVAL);
17934
17935 mutex_enter(&dtrace_lock);
17936
17937 if (cmd == DTRACEIOC_BUFSNAP) {
17938 buf = &state->dts_buffer[desc.dtbd_cpu];
17939 } else {
17940 buf = &state->dts_aggbuffer[desc.dtbd_cpu];
17941 }
17942
17943 if (buf->dtb_flags & (DTRACEBUF_RING | DTRACEBUF_FILL)) {
17944 size_t sz = buf->dtb_offset;
17945
17946 if (state->dts_activity != DTRACE_ACTIVITY_STOPPED) {
17947 mutex_exit(&dtrace_lock);
17948 return (EBUSY);
17949 }
17950
17951 /*
17952 * If this buffer has already been consumed, we're
17953 * going to indicate that there's nothing left here
17954 * to consume.
17955 */
17956 if (buf->dtb_flags & DTRACEBUF_CONSUMED) {
17957 mutex_exit(&dtrace_lock);
17958
17959 desc.dtbd_size = 0;
17960 desc.dtbd_drops = 0;
17961 desc.dtbd_errors = 0;
17962 desc.dtbd_oldest = 0;
17963 sz = sizeof (desc);
17964
17965 if (copyout(&desc, (void *)arg, sz) != 0)
17966 return (EFAULT);
17967
17968 return (0);
17969 }
17970
17971 /*
17972 * If this is a ring buffer that has wrapped, we want
17973 * to copy the whole thing out.
17974 */
17975 if (buf->dtb_flags & DTRACEBUF_WRAPPED) {
17976 dtrace_buffer_polish(buf);
17977 sz = buf->dtb_size;
17978 }
17979
17980 if (copyout(buf->dtb_tomax, desc.dtbd_data, sz) != 0) {
17981 mutex_exit(&dtrace_lock);
17982 return (EFAULT);
17983 }
17984
17985 desc.dtbd_size = sz;
17986 desc.dtbd_drops = buf->dtb_drops;
17987 desc.dtbd_errors = buf->dtb_errors;
17988 desc.dtbd_oldest = buf->dtb_xamot_offset;
17989 desc.dtbd_timestamp = dtrace_gethrtime();
17990
17991 mutex_exit(&dtrace_lock);
17992
17993 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
17994 return (EFAULT);
17995
17996 buf->dtb_flags |= DTRACEBUF_CONSUMED;
17997
17998 return (0);
17999 }
18000
18001 if (buf->dtb_tomax == NULL) {
18002 ASSERT(buf->dtb_xamot == NULL);
18003 mutex_exit(&dtrace_lock);
18004 return (ENOENT);
18005 }
18006
18007 cached = buf->dtb_tomax;
18008 ASSERT(!(buf->dtb_flags & DTRACEBUF_NOSWITCH));
18009
18010 dtrace_xcall(desc.dtbd_cpu,
18011 (dtrace_xcall_t)dtrace_buffer_switch, buf);
18012
18013 state->dts_errors += buf->dtb_xamot_errors;
18014
18015 /*
18016 * If the buffers did not actually switch, then the cross call
18017 * did not take place -- presumably because the given CPU is
18018 * not in the ready set. If this is the case, we'll return
18019 * ENOENT.
18020 */
18021 if (buf->dtb_tomax == cached) {
18022 ASSERT(buf->dtb_xamot != cached);
18023 mutex_exit(&dtrace_lock);
18024 return (ENOENT);
18025 }
18026
18027 ASSERT(cached == buf->dtb_xamot);
18028
18029 /*
18030 * We have our snapshot; now copy it out.
18031 */
18032 if (copyout(buf->dtb_xamot, desc.dtbd_data,
18033 buf->dtb_xamot_offset) != 0) {
18034 mutex_exit(&dtrace_lock);
18035 return (EFAULT);
18036 }
18037
18038 desc.dtbd_size = buf->dtb_xamot_offset;
18039 desc.dtbd_drops = buf->dtb_xamot_drops;
18040 desc.dtbd_errors = buf->dtb_xamot_errors;
18041 desc.dtbd_oldest = 0;
18042 desc.dtbd_timestamp = buf->dtb_switched;
18043
18044 mutex_exit(&dtrace_lock);
18045
18046 /*
18047 * Finally, copy out the buffer description.
18048 */
18049 if (copyout(&desc, (void *)arg, sizeof (desc)) != 0)
18050 return (EFAULT);
18051
18052 return (0);
18053 }
18054
18055 case DTRACEIOC_CONF: {
18056 dtrace_conf_t conf;
18057
18058 bzero(&conf, sizeof (conf));
18059 conf.dtc_difversion = DIF_VERSION;
18060 conf.dtc_difintregs = DIF_DIR_NREGS;
18061 conf.dtc_diftupregs = DIF_DTR_NREGS;
18062 conf.dtc_ctfmodel = CTF_MODEL_NATIVE;
18063
18064 if (copyout(&conf, (void *)arg, sizeof (conf)) != 0)
18065 return (EFAULT);
18066
18067 return (0);
18068 }
18069
18070 case DTRACEIOC_STATUS: {
18071 dtrace_status_t stat;
18072 dtrace_dstate_t *dstate;
18073 int i, j;
18074 uint64_t nerrs;
18075
18076 /*
18077 * See the comment in dtrace_state_deadman() for the reason
18078 * for setting dts_laststatus to INT64_MAX before setting
18079 * it to the correct value.
18080 */
18081 state->dts_laststatus = INT64_MAX;
18082 dtrace_membar_producer();
18083 state->dts_laststatus = dtrace_gethrtime();
18084
18085 bzero(&stat, sizeof (stat));
18086
18087 mutex_enter(&dtrace_lock);
18088
18089 if (state->dts_activity == DTRACE_ACTIVITY_INACTIVE) {
18090 mutex_exit(&dtrace_lock);
18091 return (ENOENT);
18092 }
18093
18094 if (state->dts_activity == DTRACE_ACTIVITY_DRAINING)
18095 stat.dtst_exiting = 1;
18096
18097 nerrs = state->dts_errors;
18098 dstate = &state->dts_vstate.dtvs_dynvars;
18099
18100 for (i = 0; i < NCPU; i++) {
18101 dtrace_dstate_percpu_t *dcpu = &dstate->dtds_percpu[i];
18102
18103 stat.dtst_dyndrops += dcpu->dtdsc_drops;
18104 stat.dtst_dyndrops_dirty += dcpu->dtdsc_dirty_drops;
18105 stat.dtst_dyndrops_rinsing += dcpu->dtdsc_rinsing_drops;
18106
18107 if (state->dts_buffer[i].dtb_flags & DTRACEBUF_FULL)
18108 stat.dtst_filled++;
18109
18110 nerrs += state->dts_buffer[i].dtb_errors;
18111
18112 for (j = 0; j < state->dts_nspeculations; j++) {
18113 dtrace_speculation_t *spec;
18114 dtrace_buffer_t *buf;
18115
18116 spec = &state->dts_speculations[j];
18117 buf = &spec->dtsp_buffer[i];
18118 stat.dtst_specdrops += buf->dtb_xamot_drops;
18119 }
18120 }
18121
18122 stat.dtst_specdrops_busy = state->dts_speculations_busy;
18123 stat.dtst_specdrops_unavail = state->dts_speculations_unavail;
18124 stat.dtst_stkstroverflows = state->dts_stkstroverflows;
18125 stat.dtst_dblerrors = state->dts_dblerrors;
18126 stat.dtst_killed =
18127 (state->dts_activity == DTRACE_ACTIVITY_KILLED);
18128 stat.dtst_errors = nerrs;
18129
18130 mutex_exit(&dtrace_lock);
18131
18132 if (copyout(&stat, (void *)arg, sizeof (stat)) != 0)
18133 return (EFAULT);
18134
18135 return (0);
18136 }
18137
18138 case DTRACEIOC_FORMAT: {
18139 dtrace_fmtdesc_t fmt;
18140 char *str;
18141 int len;
18142
18143 if (copyin((void *)arg, &fmt, sizeof (fmt)) != 0)
18144 return (EFAULT);
18145
18146 mutex_enter(&dtrace_lock);
18147
18148 if (fmt.dtfd_format == 0 ||
18149 fmt.dtfd_format > state->dts_nformats) {
18150 mutex_exit(&dtrace_lock);
18151 return (EINVAL);
18152 }
18153
18154 /*
18155 * Format strings are allocated contiguously and they are
18156 * never freed; if a format index is less than the number
18157 * of formats, we can assert that the format map is non-NULL
18158 * and that the format for the specified index is non-NULL.
18159 */
18160 ASSERT(state->dts_formats != NULL);
18161 str = state->dts_formats[fmt.dtfd_format - 1];
18162 ASSERT(str != NULL);
18163
18164 len = strlen(str) + 1;
18165
18166 if (len > fmt.dtfd_length) {
18167 fmt.dtfd_length = len;
18168
18169 if (copyout(&fmt, (void *)arg, sizeof (fmt)) != 0) {
18170 mutex_exit(&dtrace_lock);
18171 return (EINVAL);
18172 }
18173 } else {
18174 if (copyout(str, fmt.dtfd_string, len) != 0) {
18175 mutex_exit(&dtrace_lock);
18176 return (EINVAL);
18177 }
18178 }
18179
18180 mutex_exit(&dtrace_lock);
18181 return (0);
18182 }
18183
18184 default:
18185 break;
18186 }
18187
18188 return (ENOTTY);
18189 }
18190
18191 /*ARGSUSED*/
18192 static int
18193 dtrace_detach(dev_info_t *dip, ddi_detach_cmd_t cmd)
18194 {
18195 dtrace_state_t *state;
18196
18197 switch (cmd) {
18198 case DDI_DETACH:
18199 break;
18200
18201 case DDI_SUSPEND:
18202 return (DDI_SUCCESS);
18203
18204 default:
18205 return (DDI_FAILURE);
18206 }
18207
18208 mutex_enter(&cpu_lock);
18209 mutex_enter(&dtrace_provider_lock);
18210 mutex_enter(&dtrace_lock);
18211
18212 ASSERT(dtrace_opens == 0);
18213
18214 if (dtrace_helpers > 0) {
18215 mutex_exit(&dtrace_provider_lock);
18216 mutex_exit(&dtrace_lock);
18217 mutex_exit(&cpu_lock);
18218 return (DDI_FAILURE);
18219 }
18220
18221 if (dtrace_unregister((dtrace_provider_id_t)dtrace_provider) != 0) {
18222 mutex_exit(&dtrace_provider_lock);
18223 mutex_exit(&dtrace_lock);
18224 mutex_exit(&cpu_lock);
18225 return (DDI_FAILURE);
18226 }
18227
18228 dtrace_provider = NULL;
18229
18230 if ((state = dtrace_anon_grab()) != NULL) {
18231 /*
18232 * If there were ECBs on this state, the provider should
18233 * have not been allowed to detach; assert that there is
18234 * none.
18235 */
18236 ASSERT(state->dts_necbs == 0);
18237 dtrace_state_destroy(state);
18238
18239 /*
18240 * If we're being detached with anonymous state, we need to
18241 * indicate to the kernel debugger that DTrace is now inactive.
18242 */
18243 (void) kdi_dtrace_set(KDI_DTSET_DTRACE_DEACTIVATE);
18244 }
18245
18246 bzero(&dtrace_anon, sizeof (dtrace_anon_t));
18247 unregister_cpu_setup_func((cpu_setup_func_t *)dtrace_cpu_setup, NULL);
18248 dtrace_cpu_init = NULL;
18249 dtrace_helpers_cleanup = NULL;
18250 dtrace_helpers_fork = NULL;
18251 dtrace_cpustart_init = NULL;
18252 dtrace_cpustart_fini = NULL;
18253 dtrace_debugger_init = NULL;
18254 dtrace_debugger_fini = NULL;
18255 dtrace_modload = NULL;
18256 dtrace_modunload = NULL;
18257
18258 ASSERT(dtrace_getf == 0);
18259 ASSERT(dtrace_closef == NULL);
18260
18261 mutex_exit(&cpu_lock);
18262
18263 kmem_free(dtrace_probes, dtrace_nprobes * sizeof (dtrace_probe_t *));
18264 dtrace_probes = NULL;
18265 dtrace_nprobes = 0;
18266
18267 dtrace_hash_destroy(dtrace_bymod);
18268 dtrace_hash_destroy(dtrace_byfunc);
18269 dtrace_hash_destroy(dtrace_byname);
18270 dtrace_bymod = NULL;
18271 dtrace_byfunc = NULL;
18272 dtrace_byname = NULL;
18273
18274 kmem_cache_destroy(dtrace_state_cache);
18275 vmem_destroy(dtrace_minor);
18276 vmem_destroy(dtrace_arena);
18277
18278 if (dtrace_toxrange != NULL) {
18279 kmem_free(dtrace_toxrange,
18280 dtrace_toxranges_max * sizeof (dtrace_toxrange_t));
18281 dtrace_toxrange = NULL;
18282 dtrace_toxranges = 0;
18283 dtrace_toxranges_max = 0;
18284 }
18285
18286 ddi_remove_minor_node(dtrace_devi, NULL);
18287 dtrace_devi = NULL;
18288
18289 ddi_soft_state_fini(&dtrace_softstate);
18290
18291 ASSERT(dtrace_vtime_references == 0);
18292 ASSERT(dtrace_opens == 0);
18293 ASSERT(dtrace_retained == NULL);
18294
18295 mutex_exit(&dtrace_lock);
18296 mutex_exit(&dtrace_provider_lock);
18297
18298 /*
18299 * We don't destroy the task queue until after we have dropped our
18300 * locks (taskq_destroy() may block on running tasks). To prevent
18301 * attempting to do work after we have effectively detached but before
18302 * the task queue has been destroyed, all tasks dispatched via the
18303 * task queue must check that DTrace is still attached before
18304 * performing any operation.
18305 */
18306 taskq_destroy(dtrace_taskq);
18307 dtrace_taskq = NULL;
18308
18309 return (DDI_SUCCESS);
18310 }
18311 #endif
18312
18313 #ifdef illumos
18314 /*ARGSUSED*/
18315 static int
18316 dtrace_info(dev_info_t *dip, ddi_info_cmd_t infocmd, void *arg, void **result)
18317 {
18318 int error;
18319
18320 switch (infocmd) {
18321 case DDI_INFO_DEVT2DEVINFO:
18322 *result = (void *)dtrace_devi;
18323 error = DDI_SUCCESS;
18324 break;
18325 case DDI_INFO_DEVT2INSTANCE:
18326 *result = (void *)0;
18327 error = DDI_SUCCESS;
18328 break;
18329 default:
18330 error = DDI_FAILURE;
18331 }
18332 return (error);
18333 }
18334 #endif
18335
18336 #ifdef illumos
18337 static struct cb_ops dtrace_cb_ops = {
18338 dtrace_open, /* open */
18339 dtrace_close, /* close */
18340 nulldev, /* strategy */
18341 nulldev, /* print */
18342 nodev, /* dump */
18343 nodev, /* read */
18344 nodev, /* write */
18345 dtrace_ioctl, /* ioctl */
18346 nodev, /* devmap */
18347 nodev, /* mmap */
18348 nodev, /* segmap */
18349 nochpoll, /* poll */
18350 ddi_prop_op, /* cb_prop_op */
18351 0, /* streamtab */
18352 D_NEW | D_MP /* Driver compatibility flag */
18353 };
18354
18355 static struct dev_ops dtrace_ops = {
18356 DEVO_REV, /* devo_rev */
18357 0, /* refcnt */
18358 dtrace_info, /* get_dev_info */
18359 nulldev, /* identify */
18360 nulldev, /* probe */
18361 dtrace_attach, /* attach */
18362 dtrace_detach, /* detach */
18363 nodev, /* reset */
18364 &dtrace_cb_ops, /* driver operations */
18365 NULL, /* bus operations */
18366 nodev /* dev power */
18367 };
18368
18369 static struct modldrv modldrv = {
18370 &mod_driverops, /* module type (this is a pseudo driver) */
18371 "Dynamic Tracing", /* name of module */
18372 &dtrace_ops, /* driver ops */
18373 };
18374
18375 static struct modlinkage modlinkage = {
18376 MODREV_1,
18377 (void *)&modldrv,
18378 NULL
18379 };
18380
18381 int
18382 _init(void)
18383 {
18384 return (mod_install(&modlinkage));
18385 }
18386
18387 int
18388 _info(struct modinfo *modinfop)
18389 {
18390 return (mod_info(&modlinkage, modinfop));
18391 }
18392
18393 int
18394 _fini(void)
18395 {
18396 return (mod_remove(&modlinkage));
18397 }
18398 #else
18399
18400 static d_ioctl_t dtrace_ioctl;
18401 static d_ioctl_t dtrace_ioctl_helper;
18402 static void dtrace_load(void *);
18403 static int dtrace_unload(void);
18404 static struct cdev *dtrace_dev;
18405 static struct cdev *helper_dev;
18406
18407 void dtrace_invop_init(void);
18408 void dtrace_invop_uninit(void);
18409
18410 static struct cdevsw dtrace_cdevsw = {
18411 .d_version = D_VERSION,
18412 .d_ioctl = dtrace_ioctl,
18413 .d_open = dtrace_open,
18414 .d_name = "dtrace",
18415 };
18416
18417 static struct cdevsw helper_cdevsw = {
18418 .d_version = D_VERSION,
18419 .d_ioctl = dtrace_ioctl_helper,
18420 .d_name = "helper",
18421 };
18422
18423 #include <dtrace_anon.c>
18424 #include <dtrace_ioctl.c>
18425 #include <dtrace_load.c>
18426 #include <dtrace_modevent.c>
18427 #include <dtrace_sysctl.c>
18428 #include <dtrace_unload.c>
18429 #include <dtrace_vtime.c>
18430 #include <dtrace_hacks.c>
18431 #include <dtrace_isa.c>
18432
18433 SYSINIT(dtrace_load, SI_SUB_DTRACE, SI_ORDER_FIRST, dtrace_load, NULL);
18434 SYSUNINIT(dtrace_unload, SI_SUB_DTRACE, SI_ORDER_FIRST, dtrace_unload, NULL);
18435 SYSINIT(dtrace_anon_init, SI_SUB_DTRACE_ANON, SI_ORDER_FIRST, dtrace_anon_init, NULL);
18436
18437 DEV_MODULE(dtrace, dtrace_modevent, NULL);
18438 MODULE_VERSION(dtrace, 1);
18439 MODULE_DEPEND(dtrace, opensolaris, 1, 1, 1);
18440 #endif
Cache object: 54a8e0a001e90856ecfda828fa344e06
|