1 /*-
2 * SPDX-License-Identifier: BSD-3-Clause
3 *
4 * Copyright (c) 2008 Isilon Systems, Inc.
5 * Copyright (c) 2008 Ilya Maykov <ivmaykov@gmail.com>
6 * Copyright (c) 1998 Berkeley Software Design, Inc.
7 * All rights reserved.
8 *
9 * Redistribution and use in source and binary forms, with or without
10 * modification, are permitted provided that the following conditions
11 * are met:
12 * 1. Redistributions of source code must retain the above copyright
13 * notice, this list of conditions and the following disclaimer.
14 * 2. Redistributions in binary form must reproduce the above copyright
15 * notice, this list of conditions and the following disclaimer in the
16 * documentation and/or other materials provided with the distribution.
17 * 3. Berkeley Software Design Inc's name may not be used to endorse or
18 * promote products derived from this software without specific prior
19 * written permission.
20 *
21 * THIS SOFTWARE IS PROVIDED BY BERKELEY SOFTWARE DESIGN INC ``AS IS'' AND
22 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24 * ARE DISCLAIMED. IN NO EVENT SHALL BERKELEY SOFTWARE DESIGN INC BE LIABLE
25 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31 * SUCH DAMAGE.
32 *
33 * from BSDI $Id: mutex_witness.c,v 1.1.2.20 2000/04/27 03:10:27 cp Exp $
34 * and BSDI $Id: synch_machdep.c,v 2.3.2.39 2000/04/27 03:10:25 cp Exp $
35 */
36
37 /*
38 * Implementation of the `witness' lock verifier. Originally implemented for
39 * mutexes in BSD/OS. Extended to handle generic lock objects and lock
40 * classes in FreeBSD.
41 */
42
43 /*
44 * Main Entry: witness
45 * Pronunciation: 'wit-n&s
46 * Function: noun
47 * Etymology: Middle English witnesse, from Old English witnes knowledge,
48 * testimony, witness, from 2wit
49 * Date: before 12th century
50 * 1 : attestation of a fact or event : TESTIMONY
51 * 2 : one that gives evidence; specifically : one who testifies in
52 * a cause or before a judicial tribunal
53 * 3 : one asked to be present at a transaction so as to be able to
54 * testify to its having taken place
55 * 4 : one who has personal knowledge of something
56 * 5 a : something serving as evidence or proof : SIGN
57 * b : public affirmation by word or example of usually
58 * religious faith or conviction <the heroic witness to divine
59 * life -- Pilot>
60 * 6 capitalized : a member of the Jehovah's Witnesses
61 */
62
63 /*
64 * Special rules concerning Giant and lock orders:
65 *
66 * 1) Giant must be acquired before any other mutexes. Stated another way,
67 * no other mutex may be held when Giant is acquired.
68 *
69 * 2) Giant must be released when blocking on a sleepable lock.
70 *
71 * This rule is less obvious, but is a result of Giant providing the same
72 * semantics as spl(). Basically, when a thread sleeps, it must release
73 * Giant. When a thread blocks on a sleepable lock, it sleeps. Hence rule
74 * 2).
75 *
76 * 3) Giant may be acquired before or after sleepable locks.
77 *
78 * This rule is also not quite as obvious. Giant may be acquired after
79 * a sleepable lock because it is a non-sleepable lock and non-sleepable
80 * locks may always be acquired while holding a sleepable lock. The second
81 * case, Giant before a sleepable lock, follows from rule 2) above. Suppose
82 * you have two threads T1 and T2 and a sleepable lock X. Suppose that T1
83 * acquires X and blocks on Giant. Then suppose that T2 acquires Giant and
84 * blocks on X. When T2 blocks on X, T2 will release Giant allowing T1 to
85 * execute. Thus, acquiring Giant both before and after a sleepable lock
86 * will not result in a lock order reversal.
87 */
88
89 #include <sys/cdefs.h>
90 __FBSDID("$FreeBSD$");
91
92 #include "opt_ddb.h"
93 #include "opt_hwpmc_hooks.h"
94 #include "opt_stack.h"
95 #include "opt_witness.h"
96
97 #include <sys/param.h>
98 #include <sys/bus.h>
99 #include <sys/kdb.h>
100 #include <sys/kernel.h>
101 #include <sys/ktr.h>
102 #include <sys/lock.h>
103 #include <sys/malloc.h>
104 #include <sys/mutex.h>
105 #include <sys/priv.h>
106 #include <sys/proc.h>
107 #include <sys/sbuf.h>
108 #include <sys/sched.h>
109 #include <sys/stack.h>
110 #include <sys/sysctl.h>
111 #include <sys/syslog.h>
112 #include <sys/systm.h>
113
114 #ifdef DDB
115 #include <ddb/ddb.h>
116 #endif
117
118 #include <machine/stdarg.h>
119
120 #if !defined(DDB) && !defined(STACK)
121 #error "DDB or STACK options are required for WITNESS"
122 #endif
123
124 /* Note that these traces do not work with KTR_ALQ. */
125 #if 0
126 #define KTR_WITNESS KTR_SUBSYS
127 #else
128 #define KTR_WITNESS 0
129 #endif
130
131 #define LI_RECURSEMASK 0x0000ffff /* Recursion depth of lock instance. */
132 #define LI_EXCLUSIVE 0x00010000 /* Exclusive lock instance. */
133 #define LI_NORELEASE 0x00020000 /* Lock not allowed to be released. */
134
135 #ifndef WITNESS_COUNT
136 #define WITNESS_COUNT 1536
137 #endif
138 #define WITNESS_HASH_SIZE 251 /* Prime, gives load factor < 2 */
139 #define WITNESS_PENDLIST (512 + (MAXCPU * 4))
140
141 /* Allocate 256 KB of stack data space */
142 #define WITNESS_LO_DATA_COUNT 2048
143
144 /* Prime, gives load factor of ~2 at full load */
145 #define WITNESS_LO_HASH_SIZE 1021
146
147 /*
148 * XXX: This is somewhat bogus, as we assume here that at most 2048 threads
149 * will hold LOCK_NCHILDREN locks. We handle failure ok, and we should
150 * probably be safe for the most part, but it's still a SWAG.
151 */
152 #define LOCK_NCHILDREN 5
153 #define LOCK_CHILDCOUNT 2048
154
155 #define MAX_W_NAME 64
156
157 #define FULLGRAPH_SBUF_SIZE 512
158
159 /*
160 * These flags go in the witness relationship matrix and describe the
161 * relationship between any two struct witness objects.
162 */
163 #define WITNESS_UNRELATED 0x00 /* No lock order relation. */
164 #define WITNESS_PARENT 0x01 /* Parent, aka direct ancestor. */
165 #define WITNESS_ANCESTOR 0x02 /* Direct or indirect ancestor. */
166 #define WITNESS_CHILD 0x04 /* Child, aka direct descendant. */
167 #define WITNESS_DESCENDANT 0x08 /* Direct or indirect descendant. */
168 #define WITNESS_ANCESTOR_MASK (WITNESS_PARENT | WITNESS_ANCESTOR)
169 #define WITNESS_DESCENDANT_MASK (WITNESS_CHILD | WITNESS_DESCENDANT)
170 #define WITNESS_RELATED_MASK \
171 (WITNESS_ANCESTOR_MASK | WITNESS_DESCENDANT_MASK)
172 #define WITNESS_REVERSAL 0x10 /* A lock order reversal has been
173 * observed. */
174 #define WITNESS_RESERVED1 0x20 /* Unused flag, reserved. */
175 #define WITNESS_RESERVED2 0x40 /* Unused flag, reserved. */
176 #define WITNESS_LOCK_ORDER_KNOWN 0x80 /* This lock order is known. */
177
178 /* Descendant to ancestor flags */
179 #define WITNESS_DTOA(x) (((x) & WITNESS_RELATED_MASK) >> 2)
180
181 /* Ancestor to descendant flags */
182 #define WITNESS_ATOD(x) (((x) & WITNESS_RELATED_MASK) << 2)
183
184 #define WITNESS_INDEX_ASSERT(i) \
185 MPASS((i) > 0 && (i) <= w_max_used_index && (i) < witness_count)
186
187 static MALLOC_DEFINE(M_WITNESS, "Witness", "Witness");
188
189 /*
190 * Lock instances. A lock instance is the data associated with a lock while
191 * it is held by witness. For example, a lock instance will hold the
192 * recursion count of a lock. Lock instances are held in lists. Spin locks
193 * are held in a per-cpu list while sleep locks are held in per-thread list.
194 */
195 struct lock_instance {
196 struct lock_object *li_lock;
197 const char *li_file;
198 int li_line;
199 u_int li_flags;
200 };
201
202 /*
203 * A simple list type used to build the list of locks held by a thread
204 * or CPU. We can't simply embed the list in struct lock_object since a
205 * lock may be held by more than one thread if it is a shared lock. Locks
206 * are added to the head of the list, so we fill up each list entry from
207 * "the back" logically. To ease some of the arithmetic, we actually fill
208 * in each list entry the normal way (children[0] then children[1], etc.) but
209 * when we traverse the list we read children[count-1] as the first entry
210 * down to children[0] as the final entry.
211 */
212 struct lock_list_entry {
213 struct lock_list_entry *ll_next;
214 struct lock_instance ll_children[LOCK_NCHILDREN];
215 u_int ll_count;
216 };
217
218 /*
219 * The main witness structure. One of these per named lock type in the system
220 * (for example, "vnode interlock").
221 */
222 struct witness {
223 char w_name[MAX_W_NAME];
224 uint32_t w_index; /* Index in the relationship matrix */
225 struct lock_class *w_class;
226 STAILQ_ENTRY(witness) w_list; /* List of all witnesses. */
227 STAILQ_ENTRY(witness) w_typelist; /* Witnesses of a type. */
228 struct witness *w_hash_next; /* Linked list in hash buckets. */
229 const char *w_file; /* File where last acquired */
230 uint32_t w_line; /* Line where last acquired */
231 uint32_t w_refcount;
232 uint16_t w_num_ancestors; /* direct/indirect
233 * ancestor count */
234 uint16_t w_num_descendants; /* direct/indirect
235 * descendant count */
236 int16_t w_ddb_level;
237 unsigned w_displayed:1;
238 unsigned w_reversed:1;
239 };
240
241 STAILQ_HEAD(witness_list, witness);
242
243 /*
244 * The witness hash table. Keys are witness names (const char *), elements are
245 * witness objects (struct witness *).
246 */
247 struct witness_hash {
248 struct witness *wh_array[WITNESS_HASH_SIZE];
249 uint32_t wh_size;
250 uint32_t wh_count;
251 };
252
253 /*
254 * Key type for the lock order data hash table.
255 */
256 struct witness_lock_order_key {
257 uint16_t from;
258 uint16_t to;
259 };
260
261 struct witness_lock_order_data {
262 struct stack wlod_stack;
263 struct witness_lock_order_key wlod_key;
264 struct witness_lock_order_data *wlod_next;
265 };
266
267 /*
268 * The witness lock order data hash table. Keys are witness index tuples
269 * (struct witness_lock_order_key), elements are lock order data objects
270 * (struct witness_lock_order_data).
271 */
272 struct witness_lock_order_hash {
273 struct witness_lock_order_data *wloh_array[WITNESS_LO_HASH_SIZE];
274 u_int wloh_size;
275 u_int wloh_count;
276 };
277
278 struct witness_blessed {
279 const char *b_lock1;
280 const char *b_lock2;
281 };
282
283 struct witness_pendhelp {
284 const char *wh_type;
285 struct lock_object *wh_lock;
286 };
287
288 struct witness_order_list_entry {
289 const char *w_name;
290 struct lock_class *w_class;
291 };
292
293 /*
294 * Returns 0 if one of the locks is a spin lock and the other is not.
295 * Returns 1 otherwise.
296 */
297 static __inline int
298 witness_lock_type_equal(struct witness *w1, struct witness *w2)
299 {
300
301 return ((w1->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) ==
302 (w2->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)));
303 }
304
305 static __inline int
306 witness_lock_order_key_equal(const struct witness_lock_order_key *a,
307 const struct witness_lock_order_key *b)
308 {
309
310 return (a->from == b->from && a->to == b->to);
311 }
312
313 static int _isitmyx(struct witness *w1, struct witness *w2, int rmask,
314 const char *fname);
315 static void adopt(struct witness *parent, struct witness *child);
316 static int blessed(struct witness *, struct witness *);
317 static void depart(struct witness *w);
318 static struct witness *enroll(const char *description,
319 struct lock_class *lock_class);
320 static struct lock_instance *find_instance(struct lock_list_entry *list,
321 const struct lock_object *lock);
322 static int isitmychild(struct witness *parent, struct witness *child);
323 static int isitmydescendant(struct witness *parent, struct witness *child);
324 static void itismychild(struct witness *parent, struct witness *child);
325 static int sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS);
326 static int sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS);
327 static int sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS);
328 static int sysctl_debug_witness_channel(SYSCTL_HANDLER_ARGS);
329 static void witness_add_fullgraph(struct sbuf *sb, struct witness *parent);
330 #ifdef DDB
331 static void witness_ddb_compute_levels(void);
332 static void witness_ddb_display(int(*)(const char *fmt, ...));
333 static void witness_ddb_display_descendants(int(*)(const char *fmt, ...),
334 struct witness *, int indent);
335 static void witness_ddb_display_list(int(*prnt)(const char *fmt, ...),
336 struct witness_list *list);
337 static void witness_ddb_level_descendants(struct witness *parent, int l);
338 static void witness_ddb_list(struct thread *td);
339 #endif
340 static void witness_debugger(int cond, const char *msg);
341 static void witness_free(struct witness *m);
342 static struct witness *witness_get(void);
343 static uint32_t witness_hash_djb2(const uint8_t *key, uint32_t size);
344 static struct witness *witness_hash_get(const char *key);
345 static void witness_hash_put(struct witness *w);
346 static void witness_init_hash_tables(void);
347 static void witness_increment_graph_generation(void);
348 static void witness_lock_list_free(struct lock_list_entry *lle);
349 static struct lock_list_entry *witness_lock_list_get(void);
350 static int witness_lock_order_add(struct witness *parent,
351 struct witness *child);
352 static int witness_lock_order_check(struct witness *parent,
353 struct witness *child);
354 static struct witness_lock_order_data *witness_lock_order_get(
355 struct witness *parent,
356 struct witness *child);
357 static void witness_list_lock(struct lock_instance *instance,
358 int (*prnt)(const char *fmt, ...));
359 static int witness_output(const char *fmt, ...) __printflike(1, 2);
360 static int witness_voutput(const char *fmt, va_list ap) __printflike(1, 0);
361 static void witness_setflag(struct lock_object *lock, int flag, int set);
362
363 FEATURE(witness, "kernel has witness(9) support");
364
365 static SYSCTL_NODE(_debug, OID_AUTO, witness, CTLFLAG_RW, NULL,
366 "Witness Locking");
367
368 /*
369 * If set to 0, lock order checking is disabled. If set to -1,
370 * witness is completely disabled. Otherwise witness performs full
371 * lock order checking for all locks. At runtime, lock order checking
372 * may be toggled. However, witness cannot be reenabled once it is
373 * completely disabled.
374 */
375 static int witness_watch = 1;
376 SYSCTL_PROC(_debug_witness, OID_AUTO, watch, CTLFLAG_RWTUN | CTLTYPE_INT, NULL, 0,
377 sysctl_debug_witness_watch, "I", "witness is watching lock operations");
378
379 #ifdef KDB
380 /*
381 * When KDB is enabled and witness_kdb is 1, it will cause the system
382 * to drop into kdebug() when:
383 * - a lock hierarchy violation occurs
384 * - locks are held when going to sleep.
385 */
386 #ifdef WITNESS_KDB
387 int witness_kdb = 1;
388 #else
389 int witness_kdb = 0;
390 #endif
391 SYSCTL_INT(_debug_witness, OID_AUTO, kdb, CTLFLAG_RWTUN, &witness_kdb, 0, "");
392 #endif /* KDB */
393
394 #if defined(DDB) || defined(KDB)
395 /*
396 * When DDB or KDB is enabled and witness_trace is 1, it will cause the system
397 * to print a stack trace:
398 * - a lock hierarchy violation occurs
399 * - locks are held when going to sleep.
400 */
401 int witness_trace = 1;
402 SYSCTL_INT(_debug_witness, OID_AUTO, trace, CTLFLAG_RWTUN, &witness_trace, 0, "");
403 #endif /* DDB || KDB */
404
405 #ifdef WITNESS_SKIPSPIN
406 int witness_skipspin = 1;
407 #else
408 int witness_skipspin = 0;
409 #endif
410 SYSCTL_INT(_debug_witness, OID_AUTO, skipspin, CTLFLAG_RDTUN, &witness_skipspin, 0, "");
411
412 int badstack_sbuf_size;
413
414 int witness_count = WITNESS_COUNT;
415 SYSCTL_INT(_debug_witness, OID_AUTO, witness_count, CTLFLAG_RDTUN,
416 &witness_count, 0, "");
417
418 /*
419 * Output channel for witness messages. By default we print to the console.
420 */
421 enum witness_channel {
422 WITNESS_CONSOLE,
423 WITNESS_LOG,
424 WITNESS_NONE,
425 };
426
427 static enum witness_channel witness_channel = WITNESS_CONSOLE;
428 SYSCTL_PROC(_debug_witness, OID_AUTO, output_channel, CTLTYPE_STRING |
429 CTLFLAG_RWTUN, NULL, 0, sysctl_debug_witness_channel, "A",
430 "Output channel for warnings");
431
432 /*
433 * Call this to print out the relations between locks.
434 */
435 SYSCTL_PROC(_debug_witness, OID_AUTO, fullgraph, CTLTYPE_STRING | CTLFLAG_RD,
436 NULL, 0, sysctl_debug_witness_fullgraph, "A", "Show locks relation graphs");
437
438 /*
439 * Call this to print out the witness faulty stacks.
440 */
441 SYSCTL_PROC(_debug_witness, OID_AUTO, badstacks, CTLTYPE_STRING | CTLFLAG_RD,
442 NULL, 0, sysctl_debug_witness_badstacks, "A", "Show bad witness stacks");
443
444 static struct mtx w_mtx;
445
446 /* w_list */
447 static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free);
448 static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all);
449
450 /* w_typelist */
451 static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin);
452 static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep);
453
454 /* lock list */
455 static struct lock_list_entry *w_lock_list_free = NULL;
456 static struct witness_pendhelp pending_locks[WITNESS_PENDLIST];
457 static u_int pending_cnt;
458
459 static int w_free_cnt, w_spin_cnt, w_sleep_cnt;
460 SYSCTL_INT(_debug_witness, OID_AUTO, free_cnt, CTLFLAG_RD, &w_free_cnt, 0, "");
461 SYSCTL_INT(_debug_witness, OID_AUTO, spin_cnt, CTLFLAG_RD, &w_spin_cnt, 0, "");
462 SYSCTL_INT(_debug_witness, OID_AUTO, sleep_cnt, CTLFLAG_RD, &w_sleep_cnt, 0,
463 "");
464
465 static struct witness *w_data;
466 static uint8_t **w_rmatrix;
467 static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT];
468 static struct witness_hash w_hash; /* The witness hash table. */
469
470 /* The lock order data hash */
471 static struct witness_lock_order_data w_lodata[WITNESS_LO_DATA_COUNT];
472 static struct witness_lock_order_data *w_lofree = NULL;
473 static struct witness_lock_order_hash w_lohash;
474 static int w_max_used_index = 0;
475 static unsigned int w_generation = 0;
476 static const char w_notrunning[] = "Witness not running\n";
477 static const char w_stillcold[] = "Witness is still cold\n";
478 #ifdef __i386__
479 static const char w_notallowed[] = "The sysctl is disabled on the arch\n";
480 #endif
481
482 static struct witness_order_list_entry order_lists[] = {
483 /*
484 * sx locks
485 */
486 { "proctree", &lock_class_sx },
487 { "allproc", &lock_class_sx },
488 { "allprison", &lock_class_sx },
489 { NULL, NULL },
490 /*
491 * Various mutexes
492 */
493 { "Giant", &lock_class_mtx_sleep },
494 { "pipe mutex", &lock_class_mtx_sleep },
495 { "sigio lock", &lock_class_mtx_sleep },
496 { "process group", &lock_class_mtx_sleep },
497 #ifdef HWPMC_HOOKS
498 { "pmc-sleep", &lock_class_mtx_sleep },
499 #endif
500 { "process lock", &lock_class_mtx_sleep },
501 { "session", &lock_class_mtx_sleep },
502 { "uidinfo hash", &lock_class_rw },
503 { "time lock", &lock_class_mtx_sleep },
504 { NULL, NULL },
505 /*
506 * umtx
507 */
508 { "umtx lock", &lock_class_mtx_sleep },
509 { NULL, NULL },
510 /*
511 * Sockets
512 */
513 { "accept", &lock_class_mtx_sleep },
514 { "so_snd", &lock_class_mtx_sleep },
515 { "so_rcv", &lock_class_mtx_sleep },
516 { "sellck", &lock_class_mtx_sleep },
517 { NULL, NULL },
518 /*
519 * Routing
520 */
521 { "so_rcv", &lock_class_mtx_sleep },
522 { "radix node head", &lock_class_rm },
523 { "rtentry", &lock_class_mtx_sleep },
524 { "ifaddr", &lock_class_mtx_sleep },
525 { NULL, NULL },
526 /*
527 * IPv4 multicast:
528 * protocol locks before interface locks, after UDP locks.
529 */
530 { "in_multi_sx", &lock_class_sx },
531 { "udpinp", &lock_class_rw },
532 { "in_multi_list_mtx", &lock_class_mtx_sleep },
533 { "igmp_mtx", &lock_class_mtx_sleep },
534 { "ifnet_rw", &lock_class_rw },
535 { "if_addr_lock", &lock_class_mtx_sleep },
536 { NULL, NULL },
537 /*
538 * IPv6 multicast:
539 * protocol locks before interface locks, after UDP locks.
540 */
541 { "in6_multi_sx", &lock_class_sx },
542 { "udpinp", &lock_class_rw },
543 { "in6_multi_list_mtx", &lock_class_mtx_sleep },
544 { "mld_mtx", &lock_class_mtx_sleep },
545 { "ifnet_rw", &lock_class_rw },
546 { "if_addr_lock", &lock_class_mtx_sleep },
547 { NULL, NULL },
548 /*
549 * UNIX Domain Sockets
550 */
551 { "unp_link_rwlock", &lock_class_rw },
552 { "unp_list_lock", &lock_class_mtx_sleep },
553 { "unp", &lock_class_mtx_sleep },
554 { "so_snd", &lock_class_mtx_sleep },
555 { NULL, NULL },
556 /*
557 * UDP/IP
558 */
559 { "udp", &lock_class_mtx_sleep },
560 { "udpinp", &lock_class_rw },
561 { "so_snd", &lock_class_mtx_sleep },
562 { NULL, NULL },
563 /*
564 * TCP/IP
565 */
566 { "tcp", &lock_class_mtx_sleep },
567 { "tcpinp", &lock_class_rw },
568 { "so_snd", &lock_class_mtx_sleep },
569 { NULL, NULL },
570 /*
571 * BPF
572 */
573 { "bpf global lock", &lock_class_sx },
574 { "bpf interface lock", &lock_class_rw },
575 { "bpf cdev lock", &lock_class_mtx_sleep },
576 { NULL, NULL },
577 /*
578 * NFS server
579 */
580 { "nfsd_mtx", &lock_class_mtx_sleep },
581 { "so_snd", &lock_class_mtx_sleep },
582 { NULL, NULL },
583
584 /*
585 * IEEE 802.11
586 */
587 { "802.11 com lock", &lock_class_mtx_sleep},
588 { NULL, NULL },
589 /*
590 * Network drivers
591 */
592 { "network driver", &lock_class_mtx_sleep},
593 { NULL, NULL },
594
595 /*
596 * Netgraph
597 */
598 { "ng_node", &lock_class_mtx_sleep },
599 { "ng_worklist", &lock_class_mtx_sleep },
600 { NULL, NULL },
601 /*
602 * CDEV
603 */
604 { "vm map (system)", &lock_class_mtx_sleep },
605 { "vnode interlock", &lock_class_mtx_sleep },
606 { "cdev", &lock_class_mtx_sleep },
607 { "devthrd", &lock_class_mtx_sleep },
608 { NULL, NULL },
609 /*
610 * VM
611 */
612 { "vm map (user)", &lock_class_sx },
613 { "vm object", &lock_class_rw },
614 { "vm page", &lock_class_mtx_sleep },
615 { "pmap pv global", &lock_class_rw },
616 { "pmap", &lock_class_mtx_sleep },
617 { "pmap pv list", &lock_class_rw },
618 { "vm page free queue", &lock_class_mtx_sleep },
619 { "vm pagequeue", &lock_class_mtx_sleep },
620 { NULL, NULL },
621 /*
622 * kqueue/VFS interaction
623 */
624 { "kqueue", &lock_class_mtx_sleep },
625 { "struct mount mtx", &lock_class_mtx_sleep },
626 { "vnode interlock", &lock_class_mtx_sleep },
627 { NULL, NULL },
628 /*
629 * VFS namecache
630 */
631 { "ncvn", &lock_class_mtx_sleep },
632 { "ncbuc", &lock_class_rw },
633 { "vnode interlock", &lock_class_mtx_sleep },
634 { "ncneg", &lock_class_mtx_sleep },
635 { NULL, NULL },
636 /*
637 * ZFS locking
638 */
639 { "dn->dn_mtx", &lock_class_sx },
640 { "dr->dt.di.dr_mtx", &lock_class_sx },
641 { "db->db_mtx", &lock_class_sx },
642 { NULL, NULL },
643 /*
644 * TCP log locks
645 */
646 { "TCP ID tree", &lock_class_rw },
647 { "tcp log id bucket", &lock_class_mtx_sleep },
648 { "tcpinp", &lock_class_rw },
649 { "TCP log expireq", &lock_class_mtx_sleep },
650 { NULL, NULL },
651 /*
652 * spin locks
653 */
654 #ifdef SMP
655 { "ap boot", &lock_class_mtx_spin },
656 #endif
657 { "rm.mutex_mtx", &lock_class_mtx_spin },
658 { "sio", &lock_class_mtx_spin },
659 #ifdef __i386__
660 { "cy", &lock_class_mtx_spin },
661 #endif
662 #ifdef __sparc64__
663 { "pcib_mtx", &lock_class_mtx_spin },
664 { "rtc_mtx", &lock_class_mtx_spin },
665 #endif
666 { "scc_hwmtx", &lock_class_mtx_spin },
667 { "uart_hwmtx", &lock_class_mtx_spin },
668 { "fast_taskqueue", &lock_class_mtx_spin },
669 { "intr table", &lock_class_mtx_spin },
670 { "process slock", &lock_class_mtx_spin },
671 { "syscons video lock", &lock_class_mtx_spin },
672 { "sleepq chain", &lock_class_mtx_spin },
673 { "rm_spinlock", &lock_class_mtx_spin },
674 { "turnstile chain", &lock_class_mtx_spin },
675 { "turnstile lock", &lock_class_mtx_spin },
676 { "sched lock", &lock_class_mtx_spin },
677 { "td_contested", &lock_class_mtx_spin },
678 { "callout", &lock_class_mtx_spin },
679 { "entropy harvest mutex", &lock_class_mtx_spin },
680 #ifdef SMP
681 { "smp rendezvous", &lock_class_mtx_spin },
682 #endif
683 #ifdef __powerpc__
684 { "tlb0", &lock_class_mtx_spin },
685 #endif
686 { NULL, NULL },
687 { "sched lock", &lock_class_mtx_spin },
688 #ifdef HWPMC_HOOKS
689 { "pmc-per-proc", &lock_class_mtx_spin },
690 #endif
691 { NULL, NULL },
692 /*
693 * leaf locks
694 */
695 { "intrcnt", &lock_class_mtx_spin },
696 { "icu", &lock_class_mtx_spin },
697 #if defined(SMP) && defined(__sparc64__)
698 { "ipi", &lock_class_mtx_spin },
699 #endif
700 #ifdef __i386__
701 { "allpmaps", &lock_class_mtx_spin },
702 { "descriptor tables", &lock_class_mtx_spin },
703 #endif
704 { "clk", &lock_class_mtx_spin },
705 { "cpuset", &lock_class_mtx_spin },
706 { "mprof lock", &lock_class_mtx_spin },
707 { "zombie lock", &lock_class_mtx_spin },
708 { "ALD Queue", &lock_class_mtx_spin },
709 #if defined(__i386__) || defined(__amd64__)
710 { "pcicfg", &lock_class_mtx_spin },
711 { "NDIS thread lock", &lock_class_mtx_spin },
712 #endif
713 { "tw_osl_io_lock", &lock_class_mtx_spin },
714 { "tw_osl_q_lock", &lock_class_mtx_spin },
715 { "tw_cl_io_lock", &lock_class_mtx_spin },
716 { "tw_cl_intr_lock", &lock_class_mtx_spin },
717 { "tw_cl_gen_lock", &lock_class_mtx_spin },
718 #ifdef HWPMC_HOOKS
719 { "pmc-leaf", &lock_class_mtx_spin },
720 #endif
721 { "blocked lock", &lock_class_mtx_spin },
722 { NULL, NULL },
723 { NULL, NULL }
724 };
725
726 /*
727 * Pairs of locks which have been blessed. Witness does not complain about
728 * order problems with blessed lock pairs. Please do not add an entry to the
729 * table without an explanatory comment.
730 */
731 static struct witness_blessed blessed_list[] = {
732 /*
733 * See the comment in ufs_dirhash.c. Basically, a vnode lock serializes
734 * both lock orders, so a deadlock cannot happen as a result of this
735 * LOR.
736 */
737 { "dirhash", "bufwait" },
738
739 /*
740 * A UFS vnode may be locked in vget() while a buffer belonging to the
741 * parent directory vnode is locked.
742 */
743 { "ufs", "bufwait" },
744 };
745
746 /*
747 * This global is set to 0 once it becomes safe to use the witness code.
748 */
749 static int witness_cold = 1;
750
751 /*
752 * This global is set to 1 once the static lock orders have been enrolled
753 * so that a warning can be issued for any spin locks enrolled later.
754 */
755 static int witness_spin_warn = 0;
756
757 /* Trim useless garbage from filenames. */
758 static const char *
759 fixup_filename(const char *file)
760 {
761
762 if (file == NULL)
763 return (NULL);
764 while (strncmp(file, "../", 3) == 0)
765 file += 3;
766 return (file);
767 }
768
769 /*
770 * Calculate the size of early witness structures.
771 */
772 int
773 witness_startup_count(void)
774 {
775 int sz;
776
777 sz = sizeof(struct witness) * witness_count;
778 sz += sizeof(*w_rmatrix) * (witness_count + 1);
779 sz += sizeof(*w_rmatrix[0]) * (witness_count + 1) *
780 (witness_count + 1);
781
782 return (sz);
783 }
784
785 /*
786 * The WITNESS-enabled diagnostic code. Note that the witness code does
787 * assume that the early boot is single-threaded at least until after this
788 * routine is completed.
789 */
790 void
791 witness_startup(void *mem)
792 {
793 struct lock_object *lock;
794 struct witness_order_list_entry *order;
795 struct witness *w, *w1;
796 uintptr_t p;
797 int i;
798
799 p = (uintptr_t)mem;
800 w_data = (void *)p;
801 p += sizeof(struct witness) * witness_count;
802
803 w_rmatrix = (void *)p;
804 p += sizeof(*w_rmatrix) * (witness_count + 1);
805
806 for (i = 0; i < witness_count + 1; i++) {
807 w_rmatrix[i] = (void *)p;
808 p += sizeof(*w_rmatrix[i]) * (witness_count + 1);
809 }
810 badstack_sbuf_size = witness_count * 256;
811
812 /*
813 * We have to release Giant before initializing its witness
814 * structure so that WITNESS doesn't get confused.
815 */
816 mtx_unlock(&Giant);
817 mtx_assert(&Giant, MA_NOTOWNED);
818
819 CTR1(KTR_WITNESS, "%s: initializing witness", __func__);
820 mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET |
821 MTX_NOWITNESS | MTX_NOPROFILE);
822 for (i = witness_count - 1; i >= 0; i--) {
823 w = &w_data[i];
824 memset(w, 0, sizeof(*w));
825 w_data[i].w_index = i; /* Witness index never changes. */
826 witness_free(w);
827 }
828 KASSERT(STAILQ_FIRST(&w_free)->w_index == 0,
829 ("%s: Invalid list of free witness objects", __func__));
830
831 /* Witness with index 0 is not used to aid in debugging. */
832 STAILQ_REMOVE_HEAD(&w_free, w_list);
833 w_free_cnt--;
834
835 for (i = 0; i < witness_count; i++) {
836 memset(w_rmatrix[i], 0, sizeof(*w_rmatrix[i]) *
837 (witness_count + 1));
838 }
839
840 for (i = 0; i < LOCK_CHILDCOUNT; i++)
841 witness_lock_list_free(&w_locklistdata[i]);
842 witness_init_hash_tables();
843
844 /* First add in all the specified order lists. */
845 for (order = order_lists; order->w_name != NULL; order++) {
846 w = enroll(order->w_name, order->w_class);
847 if (w == NULL)
848 continue;
849 w->w_file = "order list";
850 for (order++; order->w_name != NULL; order++) {
851 w1 = enroll(order->w_name, order->w_class);
852 if (w1 == NULL)
853 continue;
854 w1->w_file = "order list";
855 itismychild(w, w1);
856 w = w1;
857 }
858 }
859 witness_spin_warn = 1;
860
861 /* Iterate through all locks and add them to witness. */
862 for (i = 0; pending_locks[i].wh_lock != NULL; i++) {
863 lock = pending_locks[i].wh_lock;
864 KASSERT(lock->lo_flags & LO_WITNESS,
865 ("%s: lock %s is on pending list but not LO_WITNESS",
866 __func__, lock->lo_name));
867 lock->lo_witness = enroll(pending_locks[i].wh_type,
868 LOCK_CLASS(lock));
869 }
870
871 /* Mark the witness code as being ready for use. */
872 witness_cold = 0;
873
874 mtx_lock(&Giant);
875 }
876
877 void
878 witness_init(struct lock_object *lock, const char *type)
879 {
880 struct lock_class *class;
881
882 /* Various sanity checks. */
883 class = LOCK_CLASS(lock);
884 if ((lock->lo_flags & LO_RECURSABLE) != 0 &&
885 (class->lc_flags & LC_RECURSABLE) == 0)
886 kassert_panic("%s: lock (%s) %s can not be recursable",
887 __func__, class->lc_name, lock->lo_name);
888 if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
889 (class->lc_flags & LC_SLEEPABLE) == 0)
890 kassert_panic("%s: lock (%s) %s can not be sleepable",
891 __func__, class->lc_name, lock->lo_name);
892 if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
893 (class->lc_flags & LC_UPGRADABLE) == 0)
894 kassert_panic("%s: lock (%s) %s can not be upgradable",
895 __func__, class->lc_name, lock->lo_name);
896
897 /*
898 * If we shouldn't watch this lock, then just clear lo_witness.
899 * Otherwise, if witness_cold is set, then it is too early to
900 * enroll this lock, so defer it to witness_initialize() by adding
901 * it to the pending_locks list. If it is not too early, then enroll
902 * the lock now.
903 */
904 if (witness_watch < 1 || panicstr != NULL ||
905 (lock->lo_flags & LO_WITNESS) == 0)
906 lock->lo_witness = NULL;
907 else if (witness_cold) {
908 pending_locks[pending_cnt].wh_lock = lock;
909 pending_locks[pending_cnt++].wh_type = type;
910 if (pending_cnt > WITNESS_PENDLIST)
911 panic("%s: pending locks list is too small, "
912 "increase WITNESS_PENDLIST\n",
913 __func__);
914 } else
915 lock->lo_witness = enroll(type, class);
916 }
917
918 void
919 witness_destroy(struct lock_object *lock)
920 {
921 struct lock_class *class;
922 struct witness *w;
923
924 class = LOCK_CLASS(lock);
925
926 if (witness_cold)
927 panic("lock (%s) %s destroyed while witness_cold",
928 class->lc_name, lock->lo_name);
929
930 /* XXX: need to verify that no one holds the lock */
931 if ((lock->lo_flags & LO_WITNESS) == 0 || lock->lo_witness == NULL)
932 return;
933 w = lock->lo_witness;
934
935 mtx_lock_spin(&w_mtx);
936 MPASS(w->w_refcount > 0);
937 w->w_refcount--;
938
939 if (w->w_refcount == 0)
940 depart(w);
941 mtx_unlock_spin(&w_mtx);
942 }
943
944 #ifdef DDB
945 static void
946 witness_ddb_compute_levels(void)
947 {
948 struct witness *w;
949
950 /*
951 * First clear all levels.
952 */
953 STAILQ_FOREACH(w, &w_all, w_list)
954 w->w_ddb_level = -1;
955
956 /*
957 * Look for locks with no parents and level all their descendants.
958 */
959 STAILQ_FOREACH(w, &w_all, w_list) {
960
961 /* If the witness has ancestors (is not a root), skip it. */
962 if (w->w_num_ancestors > 0)
963 continue;
964 witness_ddb_level_descendants(w, 0);
965 }
966 }
967
968 static void
969 witness_ddb_level_descendants(struct witness *w, int l)
970 {
971 int i;
972
973 if (w->w_ddb_level >= l)
974 return;
975
976 w->w_ddb_level = l;
977 l++;
978
979 for (i = 1; i <= w_max_used_index; i++) {
980 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
981 witness_ddb_level_descendants(&w_data[i], l);
982 }
983 }
984
985 static void
986 witness_ddb_display_descendants(int(*prnt)(const char *fmt, ...),
987 struct witness *w, int indent)
988 {
989 int i;
990
991 for (i = 0; i < indent; i++)
992 prnt(" ");
993 prnt("%s (type: %s, depth: %d, active refs: %d)",
994 w->w_name, w->w_class->lc_name,
995 w->w_ddb_level, w->w_refcount);
996 if (w->w_displayed) {
997 prnt(" -- (already displayed)\n");
998 return;
999 }
1000 w->w_displayed = 1;
1001 if (w->w_file != NULL && w->w_line != 0)
1002 prnt(" -- last acquired @ %s:%d\n", fixup_filename(w->w_file),
1003 w->w_line);
1004 else
1005 prnt(" -- never acquired\n");
1006 indent++;
1007 WITNESS_INDEX_ASSERT(w->w_index);
1008 for (i = 1; i <= w_max_used_index; i++) {
1009 if (db_pager_quit)
1010 return;
1011 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
1012 witness_ddb_display_descendants(prnt, &w_data[i],
1013 indent);
1014 }
1015 }
1016
1017 static void
1018 witness_ddb_display_list(int(*prnt)(const char *fmt, ...),
1019 struct witness_list *list)
1020 {
1021 struct witness *w;
1022
1023 STAILQ_FOREACH(w, list, w_typelist) {
1024 if (w->w_file == NULL || w->w_ddb_level > 0)
1025 continue;
1026
1027 /* This lock has no anscestors - display its descendants. */
1028 witness_ddb_display_descendants(prnt, w, 0);
1029 if (db_pager_quit)
1030 return;
1031 }
1032 }
1033
1034 static void
1035 witness_ddb_display(int(*prnt)(const char *fmt, ...))
1036 {
1037 struct witness *w;
1038
1039 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1040 witness_ddb_compute_levels();
1041
1042 /* Clear all the displayed flags. */
1043 STAILQ_FOREACH(w, &w_all, w_list)
1044 w->w_displayed = 0;
1045
1046 /*
1047 * First, handle sleep locks which have been acquired at least
1048 * once.
1049 */
1050 prnt("Sleep locks:\n");
1051 witness_ddb_display_list(prnt, &w_sleep);
1052 if (db_pager_quit)
1053 return;
1054
1055 /*
1056 * Now do spin locks which have been acquired at least once.
1057 */
1058 prnt("\nSpin locks:\n");
1059 witness_ddb_display_list(prnt, &w_spin);
1060 if (db_pager_quit)
1061 return;
1062
1063 /*
1064 * Finally, any locks which have not been acquired yet.
1065 */
1066 prnt("\nLocks which were never acquired:\n");
1067 STAILQ_FOREACH(w, &w_all, w_list) {
1068 if (w->w_file != NULL || w->w_refcount == 0)
1069 continue;
1070 prnt("%s (type: %s, depth: %d)\n", w->w_name,
1071 w->w_class->lc_name, w->w_ddb_level);
1072 if (db_pager_quit)
1073 return;
1074 }
1075 }
1076 #endif /* DDB */
1077
1078 int
1079 witness_defineorder(struct lock_object *lock1, struct lock_object *lock2)
1080 {
1081
1082 if (witness_watch == -1 || panicstr != NULL)
1083 return (0);
1084
1085 /* Require locks that witness knows about. */
1086 if (lock1 == NULL || lock1->lo_witness == NULL || lock2 == NULL ||
1087 lock2->lo_witness == NULL)
1088 return (EINVAL);
1089
1090 mtx_assert(&w_mtx, MA_NOTOWNED);
1091 mtx_lock_spin(&w_mtx);
1092
1093 /*
1094 * If we already have either an explicit or implied lock order that
1095 * is the other way around, then return an error.
1096 */
1097 if (witness_watch &&
1098 isitmydescendant(lock2->lo_witness, lock1->lo_witness)) {
1099 mtx_unlock_spin(&w_mtx);
1100 return (EDOOFUS);
1101 }
1102
1103 /* Try to add the new order. */
1104 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1105 lock2->lo_witness->w_name, lock1->lo_witness->w_name);
1106 itismychild(lock1->lo_witness, lock2->lo_witness);
1107 mtx_unlock_spin(&w_mtx);
1108 return (0);
1109 }
1110
1111 void
1112 witness_checkorder(struct lock_object *lock, int flags, const char *file,
1113 int line, struct lock_object *interlock)
1114 {
1115 struct lock_list_entry *lock_list, *lle;
1116 struct lock_instance *lock1, *lock2, *plock;
1117 struct lock_class *class, *iclass;
1118 struct witness *w, *w1;
1119 struct thread *td;
1120 int i, j;
1121
1122 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL ||
1123 panicstr != NULL)
1124 return;
1125
1126 w = lock->lo_witness;
1127 class = LOCK_CLASS(lock);
1128 td = curthread;
1129
1130 if (class->lc_flags & LC_SLEEPLOCK) {
1131
1132 /*
1133 * Since spin locks include a critical section, this check
1134 * implicitly enforces a lock order of all sleep locks before
1135 * all spin locks.
1136 */
1137 if (td->td_critnest != 0 && !kdb_active)
1138 kassert_panic("acquiring blockable sleep lock with "
1139 "spinlock or critical section held (%s) %s @ %s:%d",
1140 class->lc_name, lock->lo_name,
1141 fixup_filename(file), line);
1142
1143 /*
1144 * If this is the first lock acquired then just return as
1145 * no order checking is needed.
1146 */
1147 lock_list = td->td_sleeplocks;
1148 if (lock_list == NULL || lock_list->ll_count == 0)
1149 return;
1150 } else {
1151
1152 /*
1153 * If this is the first lock, just return as no order
1154 * checking is needed. Avoid problems with thread
1155 * migration pinning the thread while checking if
1156 * spinlocks are held. If at least one spinlock is held
1157 * the thread is in a safe path and it is allowed to
1158 * unpin it.
1159 */
1160 sched_pin();
1161 lock_list = PCPU_GET(spinlocks);
1162 if (lock_list == NULL || lock_list->ll_count == 0) {
1163 sched_unpin();
1164 return;
1165 }
1166 sched_unpin();
1167 }
1168
1169 /*
1170 * Check to see if we are recursing on a lock we already own. If
1171 * so, make sure that we don't mismatch exclusive and shared lock
1172 * acquires.
1173 */
1174 lock1 = find_instance(lock_list, lock);
1175 if (lock1 != NULL) {
1176 if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
1177 (flags & LOP_EXCLUSIVE) == 0) {
1178 witness_output("shared lock of (%s) %s @ %s:%d\n",
1179 class->lc_name, lock->lo_name,
1180 fixup_filename(file), line);
1181 witness_output("while exclusively locked from %s:%d\n",
1182 fixup_filename(lock1->li_file), lock1->li_line);
1183 kassert_panic("excl->share");
1184 }
1185 if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
1186 (flags & LOP_EXCLUSIVE) != 0) {
1187 witness_output("exclusive lock of (%s) %s @ %s:%d\n",
1188 class->lc_name, lock->lo_name,
1189 fixup_filename(file), line);
1190 witness_output("while share locked from %s:%d\n",
1191 fixup_filename(lock1->li_file), lock1->li_line);
1192 kassert_panic("share->excl");
1193 }
1194 return;
1195 }
1196
1197 /* Warn if the interlock is not locked exactly once. */
1198 if (interlock != NULL) {
1199 iclass = LOCK_CLASS(interlock);
1200 lock1 = find_instance(lock_list, interlock);
1201 if (lock1 == NULL)
1202 kassert_panic("interlock (%s) %s not locked @ %s:%d",
1203 iclass->lc_name, interlock->lo_name,
1204 fixup_filename(file), line);
1205 else if ((lock1->li_flags & LI_RECURSEMASK) != 0)
1206 kassert_panic("interlock (%s) %s recursed @ %s:%d",
1207 iclass->lc_name, interlock->lo_name,
1208 fixup_filename(file), line);
1209 }
1210
1211 /*
1212 * Find the previously acquired lock, but ignore interlocks.
1213 */
1214 plock = &lock_list->ll_children[lock_list->ll_count - 1];
1215 if (interlock != NULL && plock->li_lock == interlock) {
1216 if (lock_list->ll_count > 1)
1217 plock =
1218 &lock_list->ll_children[lock_list->ll_count - 2];
1219 else {
1220 lle = lock_list->ll_next;
1221
1222 /*
1223 * The interlock is the only lock we hold, so
1224 * simply return.
1225 */
1226 if (lle == NULL)
1227 return;
1228 plock = &lle->ll_children[lle->ll_count - 1];
1229 }
1230 }
1231
1232 /*
1233 * Try to perform most checks without a lock. If this succeeds we
1234 * can skip acquiring the lock and return success. Otherwise we redo
1235 * the check with the lock held to handle races with concurrent updates.
1236 */
1237 w1 = plock->li_lock->lo_witness;
1238 if (witness_lock_order_check(w1, w))
1239 return;
1240
1241 mtx_lock_spin(&w_mtx);
1242 if (witness_lock_order_check(w1, w)) {
1243 mtx_unlock_spin(&w_mtx);
1244 return;
1245 }
1246 witness_lock_order_add(w1, w);
1247
1248 /*
1249 * Check for duplicate locks of the same type. Note that we only
1250 * have to check for this on the last lock we just acquired. Any
1251 * other cases will be caught as lock order violations.
1252 */
1253 if (w1 == w) {
1254 i = w->w_index;
1255 if (!(lock->lo_flags & LO_DUPOK) && !(flags & LOP_DUPOK) &&
1256 !(w_rmatrix[i][i] & WITNESS_REVERSAL)) {
1257 w_rmatrix[i][i] |= WITNESS_REVERSAL;
1258 w->w_reversed = 1;
1259 mtx_unlock_spin(&w_mtx);
1260 witness_output(
1261 "acquiring duplicate lock of same type: \"%s\"\n",
1262 w->w_name);
1263 witness_output(" 1st %s @ %s:%d\n", plock->li_lock->lo_name,
1264 fixup_filename(plock->li_file), plock->li_line);
1265 witness_output(" 2nd %s @ %s:%d\n", lock->lo_name,
1266 fixup_filename(file), line);
1267 witness_debugger(1, __func__);
1268 } else
1269 mtx_unlock_spin(&w_mtx);
1270 return;
1271 }
1272 mtx_assert(&w_mtx, MA_OWNED);
1273
1274 /*
1275 * If we know that the lock we are acquiring comes after
1276 * the lock we most recently acquired in the lock order tree,
1277 * then there is no need for any further checks.
1278 */
1279 if (isitmychild(w1, w))
1280 goto out;
1281
1282 for (j = 0, lle = lock_list; lle != NULL; lle = lle->ll_next) {
1283 for (i = lle->ll_count - 1; i >= 0; i--, j++) {
1284
1285 MPASS(j < LOCK_CHILDCOUNT * LOCK_NCHILDREN);
1286 lock1 = &lle->ll_children[i];
1287
1288 /*
1289 * Ignore the interlock.
1290 */
1291 if (interlock == lock1->li_lock)
1292 continue;
1293
1294 /*
1295 * If this lock doesn't undergo witness checking,
1296 * then skip it.
1297 */
1298 w1 = lock1->li_lock->lo_witness;
1299 if (w1 == NULL) {
1300 KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
1301 ("lock missing witness structure"));
1302 continue;
1303 }
1304
1305 /*
1306 * If we are locking Giant and this is a sleepable
1307 * lock, then skip it.
1308 */
1309 if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0 &&
1310 lock == &Giant.lock_object)
1311 continue;
1312
1313 /*
1314 * If we are locking a sleepable lock and this lock
1315 * is Giant, then skip it.
1316 */
1317 if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1318 lock1->li_lock == &Giant.lock_object)
1319 continue;
1320
1321 /*
1322 * If we are locking a sleepable lock and this lock
1323 * isn't sleepable, we want to treat it as a lock
1324 * order violation to enfore a general lock order of
1325 * sleepable locks before non-sleepable locks.
1326 */
1327 if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1328 (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
1329 goto reversal;
1330
1331 /*
1332 * If we are locking Giant and this is a non-sleepable
1333 * lock, then treat it as a reversal.
1334 */
1335 if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0 &&
1336 lock == &Giant.lock_object)
1337 goto reversal;
1338
1339 /*
1340 * Check the lock order hierarchy for a reveresal.
1341 */
1342 if (!isitmydescendant(w, w1))
1343 continue;
1344 reversal:
1345
1346 /*
1347 * We have a lock order violation, check to see if it
1348 * is allowed or has already been yelled about.
1349 */
1350
1351 /* Bail if this violation is known */
1352 if (w_rmatrix[w1->w_index][w->w_index] & WITNESS_REVERSAL)
1353 goto out;
1354
1355 /* Record this as a violation */
1356 w_rmatrix[w1->w_index][w->w_index] |= WITNESS_REVERSAL;
1357 w_rmatrix[w->w_index][w1->w_index] |= WITNESS_REVERSAL;
1358 w->w_reversed = w1->w_reversed = 1;
1359 witness_increment_graph_generation();
1360
1361 /*
1362 * If the lock order is blessed, bail before logging
1363 * anything. We don't look for other lock order
1364 * violations though, which may be a bug.
1365 */
1366 if (blessed(w, w1))
1367 goto out;
1368 mtx_unlock_spin(&w_mtx);
1369
1370 #ifdef WITNESS_NO_VNODE
1371 /*
1372 * There are known LORs between VNODE locks. They are
1373 * not an indication of a bug. VNODE locks are flagged
1374 * as such (LO_IS_VNODE) and we don't yell if the LOR
1375 * is between 2 VNODE locks.
1376 */
1377 if ((lock->lo_flags & LO_IS_VNODE) != 0 &&
1378 (lock1->li_lock->lo_flags & LO_IS_VNODE) != 0)
1379 return;
1380 #endif
1381
1382 /*
1383 * Ok, yell about it.
1384 */
1385 if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1386 (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
1387 witness_output(
1388 "lock order reversal: (sleepable after non-sleepable)\n");
1389 else if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0
1390 && lock == &Giant.lock_object)
1391 witness_output(
1392 "lock order reversal: (Giant after non-sleepable)\n");
1393 else
1394 witness_output("lock order reversal:\n");
1395
1396 /*
1397 * Try to locate an earlier lock with
1398 * witness w in our list.
1399 */
1400 do {
1401 lock2 = &lle->ll_children[i];
1402 MPASS(lock2->li_lock != NULL);
1403 if (lock2->li_lock->lo_witness == w)
1404 break;
1405 if (i == 0 && lle->ll_next != NULL) {
1406 lle = lle->ll_next;
1407 i = lle->ll_count - 1;
1408 MPASS(i >= 0 && i < LOCK_NCHILDREN);
1409 } else
1410 i--;
1411 } while (i >= 0);
1412 if (i < 0) {
1413 witness_output(" 1st %p %s (%s) @ %s:%d\n",
1414 lock1->li_lock, lock1->li_lock->lo_name,
1415 w1->w_name, fixup_filename(lock1->li_file),
1416 lock1->li_line);
1417 witness_output(" 2nd %p %s (%s) @ %s:%d\n", lock,
1418 lock->lo_name, w->w_name,
1419 fixup_filename(file), line);
1420 } else {
1421 witness_output(" 1st %p %s (%s) @ %s:%d\n",
1422 lock2->li_lock, lock2->li_lock->lo_name,
1423 lock2->li_lock->lo_witness->w_name,
1424 fixup_filename(lock2->li_file),
1425 lock2->li_line);
1426 witness_output(" 2nd %p %s (%s) @ %s:%d\n",
1427 lock1->li_lock, lock1->li_lock->lo_name,
1428 w1->w_name, fixup_filename(lock1->li_file),
1429 lock1->li_line);
1430 witness_output(" 3rd %p %s (%s) @ %s:%d\n", lock,
1431 lock->lo_name, w->w_name,
1432 fixup_filename(file), line);
1433 }
1434 witness_debugger(1, __func__);
1435 return;
1436 }
1437 }
1438
1439 /*
1440 * If requested, build a new lock order. However, don't build a new
1441 * relationship between a sleepable lock and Giant if it is in the
1442 * wrong direction. The correct lock order is that sleepable locks
1443 * always come before Giant.
1444 */
1445 if (flags & LOP_NEWORDER &&
1446 !(plock->li_lock == &Giant.lock_object &&
1447 (lock->lo_flags & LO_SLEEPABLE) != 0)) {
1448 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1449 w->w_name, plock->li_lock->lo_witness->w_name);
1450 itismychild(plock->li_lock->lo_witness, w);
1451 }
1452 out:
1453 mtx_unlock_spin(&w_mtx);
1454 }
1455
1456 void
1457 witness_lock(struct lock_object *lock, int flags, const char *file, int line)
1458 {
1459 struct lock_list_entry **lock_list, *lle;
1460 struct lock_instance *instance;
1461 struct witness *w;
1462 struct thread *td;
1463
1464 if (witness_cold || witness_watch == -1 || lock->lo_witness == NULL ||
1465 panicstr != NULL)
1466 return;
1467 w = lock->lo_witness;
1468 td = curthread;
1469
1470 /* Determine lock list for this lock. */
1471 if (LOCK_CLASS(lock)->lc_flags & LC_SLEEPLOCK)
1472 lock_list = &td->td_sleeplocks;
1473 else
1474 lock_list = PCPU_PTR(spinlocks);
1475
1476 /* Check to see if we are recursing on a lock we already own. */
1477 instance = find_instance(*lock_list, lock);
1478 if (instance != NULL) {
1479 instance->li_flags++;
1480 CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
1481 td->td_proc->p_pid, lock->lo_name,
1482 instance->li_flags & LI_RECURSEMASK);
1483 instance->li_file = file;
1484 instance->li_line = line;
1485 return;
1486 }
1487
1488 /* Update per-witness last file and line acquire. */
1489 w->w_file = file;
1490 w->w_line = line;
1491
1492 /* Find the next open lock instance in the list and fill it. */
1493 lle = *lock_list;
1494 if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
1495 lle = witness_lock_list_get();
1496 if (lle == NULL)
1497 return;
1498 lle->ll_next = *lock_list;
1499 CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
1500 td->td_proc->p_pid, lle);
1501 *lock_list = lle;
1502 }
1503 instance = &lle->ll_children[lle->ll_count++];
1504 instance->li_lock = lock;
1505 instance->li_line = line;
1506 instance->li_file = file;
1507 if ((flags & LOP_EXCLUSIVE) != 0)
1508 instance->li_flags = LI_EXCLUSIVE;
1509 else
1510 instance->li_flags = 0;
1511 CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
1512 td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
1513 }
1514
1515 void
1516 witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
1517 {
1518 struct lock_instance *instance;
1519 struct lock_class *class;
1520
1521 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1522 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
1523 return;
1524 class = LOCK_CLASS(lock);
1525 if (witness_watch) {
1526 if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1527 kassert_panic(
1528 "upgrade of non-upgradable lock (%s) %s @ %s:%d",
1529 class->lc_name, lock->lo_name,
1530 fixup_filename(file), line);
1531 if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1532 kassert_panic(
1533 "upgrade of non-sleep lock (%s) %s @ %s:%d",
1534 class->lc_name, lock->lo_name,
1535 fixup_filename(file), line);
1536 }
1537 instance = find_instance(curthread->td_sleeplocks, lock);
1538 if (instance == NULL) {
1539 kassert_panic("upgrade of unlocked lock (%s) %s @ %s:%d",
1540 class->lc_name, lock->lo_name,
1541 fixup_filename(file), line);
1542 return;
1543 }
1544 if (witness_watch) {
1545 if ((instance->li_flags & LI_EXCLUSIVE) != 0)
1546 kassert_panic(
1547 "upgrade of exclusive lock (%s) %s @ %s:%d",
1548 class->lc_name, lock->lo_name,
1549 fixup_filename(file), line);
1550 if ((instance->li_flags & LI_RECURSEMASK) != 0)
1551 kassert_panic(
1552 "upgrade of recursed lock (%s) %s r=%d @ %s:%d",
1553 class->lc_name, lock->lo_name,
1554 instance->li_flags & LI_RECURSEMASK,
1555 fixup_filename(file), line);
1556 }
1557 instance->li_flags |= LI_EXCLUSIVE;
1558 }
1559
1560 void
1561 witness_downgrade(struct lock_object *lock, int flags, const char *file,
1562 int line)
1563 {
1564 struct lock_instance *instance;
1565 struct lock_class *class;
1566
1567 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1568 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
1569 return;
1570 class = LOCK_CLASS(lock);
1571 if (witness_watch) {
1572 if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1573 kassert_panic(
1574 "downgrade of non-upgradable lock (%s) %s @ %s:%d",
1575 class->lc_name, lock->lo_name,
1576 fixup_filename(file), line);
1577 if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1578 kassert_panic(
1579 "downgrade of non-sleep lock (%s) %s @ %s:%d",
1580 class->lc_name, lock->lo_name,
1581 fixup_filename(file), line);
1582 }
1583 instance = find_instance(curthread->td_sleeplocks, lock);
1584 if (instance == NULL) {
1585 kassert_panic("downgrade of unlocked lock (%s) %s @ %s:%d",
1586 class->lc_name, lock->lo_name,
1587 fixup_filename(file), line);
1588 return;
1589 }
1590 if (witness_watch) {
1591 if ((instance->li_flags & LI_EXCLUSIVE) == 0)
1592 kassert_panic(
1593 "downgrade of shared lock (%s) %s @ %s:%d",
1594 class->lc_name, lock->lo_name,
1595 fixup_filename(file), line);
1596 if ((instance->li_flags & LI_RECURSEMASK) != 0)
1597 kassert_panic(
1598 "downgrade of recursed lock (%s) %s r=%d @ %s:%d",
1599 class->lc_name, lock->lo_name,
1600 instance->li_flags & LI_RECURSEMASK,
1601 fixup_filename(file), line);
1602 }
1603 instance->li_flags &= ~LI_EXCLUSIVE;
1604 }
1605
1606 void
1607 witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
1608 {
1609 struct lock_list_entry **lock_list, *lle;
1610 struct lock_instance *instance;
1611 struct lock_class *class;
1612 struct thread *td;
1613 register_t s;
1614 int i, j;
1615
1616 if (witness_cold || lock->lo_witness == NULL || panicstr != NULL)
1617 return;
1618 td = curthread;
1619 class = LOCK_CLASS(lock);
1620
1621 /* Find lock instance associated with this lock. */
1622 if (class->lc_flags & LC_SLEEPLOCK)
1623 lock_list = &td->td_sleeplocks;
1624 else
1625 lock_list = PCPU_PTR(spinlocks);
1626 lle = *lock_list;
1627 for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
1628 for (i = 0; i < (*lock_list)->ll_count; i++) {
1629 instance = &(*lock_list)->ll_children[i];
1630 if (instance->li_lock == lock)
1631 goto found;
1632 }
1633
1634 /*
1635 * When disabling WITNESS through witness_watch we could end up in
1636 * having registered locks in the td_sleeplocks queue.
1637 * We have to make sure we flush these queues, so just search for
1638 * eventual register locks and remove them.
1639 */
1640 if (witness_watch > 0) {
1641 kassert_panic("lock (%s) %s not locked @ %s:%d", class->lc_name,
1642 lock->lo_name, fixup_filename(file), line);
1643 return;
1644 } else {
1645 return;
1646 }
1647 found:
1648
1649 /* First, check for shared/exclusive mismatches. */
1650 if ((instance->li_flags & LI_EXCLUSIVE) != 0 && witness_watch > 0 &&
1651 (flags & LOP_EXCLUSIVE) == 0) {
1652 witness_output("shared unlock of (%s) %s @ %s:%d\n",
1653 class->lc_name, lock->lo_name, fixup_filename(file), line);
1654 witness_output("while exclusively locked from %s:%d\n",
1655 fixup_filename(instance->li_file), instance->li_line);
1656 kassert_panic("excl->ushare");
1657 }
1658 if ((instance->li_flags & LI_EXCLUSIVE) == 0 && witness_watch > 0 &&
1659 (flags & LOP_EXCLUSIVE) != 0) {
1660 witness_output("exclusive unlock of (%s) %s @ %s:%d\n",
1661 class->lc_name, lock->lo_name, fixup_filename(file), line);
1662 witness_output("while share locked from %s:%d\n",
1663 fixup_filename(instance->li_file),
1664 instance->li_line);
1665 kassert_panic("share->uexcl");
1666 }
1667 /* If we are recursed, unrecurse. */
1668 if ((instance->li_flags & LI_RECURSEMASK) > 0) {
1669 CTR4(KTR_WITNESS, "%s: pid %d unrecursed on %s r=%d", __func__,
1670 td->td_proc->p_pid, instance->li_lock->lo_name,
1671 instance->li_flags);
1672 instance->li_flags--;
1673 return;
1674 }
1675 /* The lock is now being dropped, check for NORELEASE flag */
1676 if ((instance->li_flags & LI_NORELEASE) != 0 && witness_watch > 0) {
1677 witness_output("forbidden unlock of (%s) %s @ %s:%d\n",
1678 class->lc_name, lock->lo_name, fixup_filename(file), line);
1679 kassert_panic("lock marked norelease");
1680 }
1681
1682 /* Otherwise, remove this item from the list. */
1683 s = intr_disable();
1684 CTR4(KTR_WITNESS, "%s: pid %d removed %s from lle[%d]", __func__,
1685 td->td_proc->p_pid, instance->li_lock->lo_name,
1686 (*lock_list)->ll_count - 1);
1687 for (j = i; j < (*lock_list)->ll_count - 1; j++)
1688 (*lock_list)->ll_children[j] =
1689 (*lock_list)->ll_children[j + 1];
1690 (*lock_list)->ll_count--;
1691 intr_restore(s);
1692
1693 /*
1694 * In order to reduce contention on w_mtx, we want to keep always an
1695 * head object into lists so that frequent allocation from the
1696 * free witness pool (and subsequent locking) is avoided.
1697 * In order to maintain the current code simple, when the head
1698 * object is totally unloaded it means also that we do not have
1699 * further objects in the list, so the list ownership needs to be
1700 * hand over to another object if the current head needs to be freed.
1701 */
1702 if ((*lock_list)->ll_count == 0) {
1703 if (*lock_list == lle) {
1704 if (lle->ll_next == NULL)
1705 return;
1706 } else
1707 lle = *lock_list;
1708 *lock_list = lle->ll_next;
1709 CTR3(KTR_WITNESS, "%s: pid %d removed lle %p", __func__,
1710 td->td_proc->p_pid, lle);
1711 witness_lock_list_free(lle);
1712 }
1713 }
1714
1715 void
1716 witness_thread_exit(struct thread *td)
1717 {
1718 struct lock_list_entry *lle;
1719 int i, n;
1720
1721 lle = td->td_sleeplocks;
1722 if (lle == NULL || panicstr != NULL)
1723 return;
1724 if (lle->ll_count != 0) {
1725 for (n = 0; lle != NULL; lle = lle->ll_next)
1726 for (i = lle->ll_count - 1; i >= 0; i--) {
1727 if (n == 0)
1728 witness_output(
1729 "Thread %p exiting with the following locks held:\n", td);
1730 n++;
1731 witness_list_lock(&lle->ll_children[i],
1732 witness_output);
1733
1734 }
1735 kassert_panic(
1736 "Thread %p cannot exit while holding sleeplocks\n", td);
1737 }
1738 witness_lock_list_free(lle);
1739 }
1740
1741 /*
1742 * Warn if any locks other than 'lock' are held. Flags can be passed in to
1743 * exempt Giant and sleepable locks from the checks as well. If any
1744 * non-exempt locks are held, then a supplied message is printed to the
1745 * output channel along with a list of the offending locks. If indicated in the
1746 * flags then a failure results in a panic as well.
1747 */
1748 int
1749 witness_warn(int flags, struct lock_object *lock, const char *fmt, ...)
1750 {
1751 struct lock_list_entry *lock_list, *lle;
1752 struct lock_instance *lock1;
1753 struct thread *td;
1754 va_list ap;
1755 int i, n;
1756
1757 if (witness_cold || witness_watch < 1 || panicstr != NULL)
1758 return (0);
1759 n = 0;
1760 td = curthread;
1761 for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next)
1762 for (i = lle->ll_count - 1; i >= 0; i--) {
1763 lock1 = &lle->ll_children[i];
1764 if (lock1->li_lock == lock)
1765 continue;
1766 if (flags & WARN_GIANTOK &&
1767 lock1->li_lock == &Giant.lock_object)
1768 continue;
1769 if (flags & WARN_SLEEPOK &&
1770 (lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0)
1771 continue;
1772 if (n == 0) {
1773 va_start(ap, fmt);
1774 vprintf(fmt, ap);
1775 va_end(ap);
1776 printf(" with the following %slocks held:\n",
1777 (flags & WARN_SLEEPOK) != 0 ?
1778 "non-sleepable " : "");
1779 }
1780 n++;
1781 witness_list_lock(lock1, printf);
1782 }
1783
1784 /*
1785 * Pin the thread in order to avoid problems with thread migration.
1786 * Once that all verifies are passed about spinlocks ownership,
1787 * the thread is in a safe path and it can be unpinned.
1788 */
1789 sched_pin();
1790 lock_list = PCPU_GET(spinlocks);
1791 if (lock_list != NULL && lock_list->ll_count != 0) {
1792 sched_unpin();
1793
1794 /*
1795 * We should only have one spinlock and as long as
1796 * the flags cannot match for this locks class,
1797 * check if the first spinlock is the one curthread
1798 * should hold.
1799 */
1800 lock1 = &lock_list->ll_children[lock_list->ll_count - 1];
1801 if (lock_list->ll_count == 1 && lock_list->ll_next == NULL &&
1802 lock1->li_lock == lock && n == 0)
1803 return (0);
1804
1805 va_start(ap, fmt);
1806 vprintf(fmt, ap);
1807 va_end(ap);
1808 printf(" with the following %slocks held:\n",
1809 (flags & WARN_SLEEPOK) != 0 ? "non-sleepable " : "");
1810 n += witness_list_locks(&lock_list, printf);
1811 } else
1812 sched_unpin();
1813 if (flags & WARN_PANIC && n)
1814 kassert_panic("%s", __func__);
1815 else
1816 witness_debugger(n, __func__);
1817 return (n);
1818 }
1819
1820 const char *
1821 witness_file(struct lock_object *lock)
1822 {
1823 struct witness *w;
1824
1825 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1826 return ("?");
1827 w = lock->lo_witness;
1828 return (w->w_file);
1829 }
1830
1831 int
1832 witness_line(struct lock_object *lock)
1833 {
1834 struct witness *w;
1835
1836 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1837 return (0);
1838 w = lock->lo_witness;
1839 return (w->w_line);
1840 }
1841
1842 static struct witness *
1843 enroll(const char *description, struct lock_class *lock_class)
1844 {
1845 struct witness *w;
1846
1847 MPASS(description != NULL);
1848
1849 if (witness_watch == -1 || panicstr != NULL)
1850 return (NULL);
1851 if ((lock_class->lc_flags & LC_SPINLOCK)) {
1852 if (witness_skipspin)
1853 return (NULL);
1854 } else if ((lock_class->lc_flags & LC_SLEEPLOCK) == 0) {
1855 kassert_panic("lock class %s is not sleep or spin",
1856 lock_class->lc_name);
1857 return (NULL);
1858 }
1859
1860 mtx_lock_spin(&w_mtx);
1861 w = witness_hash_get(description);
1862 if (w)
1863 goto found;
1864 if ((w = witness_get()) == NULL)
1865 return (NULL);
1866 MPASS(strlen(description) < MAX_W_NAME);
1867 strcpy(w->w_name, description);
1868 w->w_class = lock_class;
1869 w->w_refcount = 1;
1870 STAILQ_INSERT_HEAD(&w_all, w, w_list);
1871 if (lock_class->lc_flags & LC_SPINLOCK) {
1872 STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
1873 w_spin_cnt++;
1874 } else if (lock_class->lc_flags & LC_SLEEPLOCK) {
1875 STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
1876 w_sleep_cnt++;
1877 }
1878
1879 /* Insert new witness into the hash */
1880 witness_hash_put(w);
1881 witness_increment_graph_generation();
1882 mtx_unlock_spin(&w_mtx);
1883 return (w);
1884 found:
1885 w->w_refcount++;
1886 if (w->w_refcount == 1)
1887 w->w_class = lock_class;
1888 mtx_unlock_spin(&w_mtx);
1889 if (lock_class != w->w_class)
1890 kassert_panic(
1891 "lock (%s) %s does not match earlier (%s) lock",
1892 description, lock_class->lc_name,
1893 w->w_class->lc_name);
1894 return (w);
1895 }
1896
1897 static void
1898 depart(struct witness *w)
1899 {
1900
1901 MPASS(w->w_refcount == 0);
1902 if (w->w_class->lc_flags & LC_SLEEPLOCK) {
1903 w_sleep_cnt--;
1904 } else {
1905 w_spin_cnt--;
1906 }
1907 /*
1908 * Set file to NULL as it may point into a loadable module.
1909 */
1910 w->w_file = NULL;
1911 w->w_line = 0;
1912 witness_increment_graph_generation();
1913 }
1914
1915
1916 static void
1917 adopt(struct witness *parent, struct witness *child)
1918 {
1919 int pi, ci, i, j;
1920
1921 if (witness_cold == 0)
1922 mtx_assert(&w_mtx, MA_OWNED);
1923
1924 /* If the relationship is already known, there's no work to be done. */
1925 if (isitmychild(parent, child))
1926 return;
1927
1928 /* When the structure of the graph changes, bump up the generation. */
1929 witness_increment_graph_generation();
1930
1931 /*
1932 * The hard part ... create the direct relationship, then propagate all
1933 * indirect relationships.
1934 */
1935 pi = parent->w_index;
1936 ci = child->w_index;
1937 WITNESS_INDEX_ASSERT(pi);
1938 WITNESS_INDEX_ASSERT(ci);
1939 MPASS(pi != ci);
1940 w_rmatrix[pi][ci] |= WITNESS_PARENT;
1941 w_rmatrix[ci][pi] |= WITNESS_CHILD;
1942
1943 /*
1944 * If parent was not already an ancestor of child,
1945 * then we increment the descendant and ancestor counters.
1946 */
1947 if ((w_rmatrix[pi][ci] & WITNESS_ANCESTOR) == 0) {
1948 parent->w_num_descendants++;
1949 child->w_num_ancestors++;
1950 }
1951
1952 /*
1953 * Find each ancestor of 'pi'. Note that 'pi' itself is counted as
1954 * an ancestor of 'pi' during this loop.
1955 */
1956 for (i = 1; i <= w_max_used_index; i++) {
1957 if ((w_rmatrix[i][pi] & WITNESS_ANCESTOR_MASK) == 0 &&
1958 (i != pi))
1959 continue;
1960
1961 /* Find each descendant of 'i' and mark it as a descendant. */
1962 for (j = 1; j <= w_max_used_index; j++) {
1963
1964 /*
1965 * Skip children that are already marked as
1966 * descendants of 'i'.
1967 */
1968 if (w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK)
1969 continue;
1970
1971 /*
1972 * We are only interested in descendants of 'ci'. Note
1973 * that 'ci' itself is counted as a descendant of 'ci'.
1974 */
1975 if ((w_rmatrix[ci][j] & WITNESS_ANCESTOR_MASK) == 0 &&
1976 (j != ci))
1977 continue;
1978 w_rmatrix[i][j] |= WITNESS_ANCESTOR;
1979 w_rmatrix[j][i] |= WITNESS_DESCENDANT;
1980 w_data[i].w_num_descendants++;
1981 w_data[j].w_num_ancestors++;
1982
1983 /*
1984 * Make sure we aren't marking a node as both an
1985 * ancestor and descendant. We should have caught
1986 * this as a lock order reversal earlier.
1987 */
1988 if ((w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK) &&
1989 (w_rmatrix[i][j] & WITNESS_DESCENDANT_MASK)) {
1990 printf("witness rmatrix paradox! [%d][%d]=%d "
1991 "both ancestor and descendant\n",
1992 i, j, w_rmatrix[i][j]);
1993 kdb_backtrace();
1994 printf("Witness disabled.\n");
1995 witness_watch = -1;
1996 }
1997 if ((w_rmatrix[j][i] & WITNESS_ANCESTOR_MASK) &&
1998 (w_rmatrix[j][i] & WITNESS_DESCENDANT_MASK)) {
1999 printf("witness rmatrix paradox! [%d][%d]=%d "
2000 "both ancestor and descendant\n",
2001 j, i, w_rmatrix[j][i]);
2002 kdb_backtrace();
2003 printf("Witness disabled.\n");
2004 witness_watch = -1;
2005 }
2006 }
2007 }
2008 }
2009
2010 static void
2011 itismychild(struct witness *parent, struct witness *child)
2012 {
2013 int unlocked;
2014
2015 MPASS(child != NULL && parent != NULL);
2016 if (witness_cold == 0)
2017 mtx_assert(&w_mtx, MA_OWNED);
2018
2019 if (!witness_lock_type_equal(parent, child)) {
2020 if (witness_cold == 0) {
2021 unlocked = 1;
2022 mtx_unlock_spin(&w_mtx);
2023 } else {
2024 unlocked = 0;
2025 }
2026 kassert_panic(
2027 "%s: parent \"%s\" (%s) and child \"%s\" (%s) are not "
2028 "the same lock type", __func__, parent->w_name,
2029 parent->w_class->lc_name, child->w_name,
2030 child->w_class->lc_name);
2031 if (unlocked)
2032 mtx_lock_spin(&w_mtx);
2033 }
2034 adopt(parent, child);
2035 }
2036
2037 /*
2038 * Generic code for the isitmy*() functions. The rmask parameter is the
2039 * expected relationship of w1 to w2.
2040 */
2041 static int
2042 _isitmyx(struct witness *w1, struct witness *w2, int rmask, const char *fname)
2043 {
2044 unsigned char r1, r2;
2045 int i1, i2;
2046
2047 i1 = w1->w_index;
2048 i2 = w2->w_index;
2049 WITNESS_INDEX_ASSERT(i1);
2050 WITNESS_INDEX_ASSERT(i2);
2051 r1 = w_rmatrix[i1][i2] & WITNESS_RELATED_MASK;
2052 r2 = w_rmatrix[i2][i1] & WITNESS_RELATED_MASK;
2053
2054 /* The flags on one better be the inverse of the flags on the other */
2055 if (!((WITNESS_ATOD(r1) == r2 && WITNESS_DTOA(r2) == r1) ||
2056 (WITNESS_DTOA(r1) == r2 && WITNESS_ATOD(r2) == r1))) {
2057 /* Don't squawk if we're potentially racing with an update. */
2058 if (!mtx_owned(&w_mtx))
2059 return (0);
2060 printf("%s: rmatrix mismatch between %s (index %d) and %s "
2061 "(index %d): w_rmatrix[%d][%d] == %hhx but "
2062 "w_rmatrix[%d][%d] == %hhx\n",
2063 fname, w1->w_name, i1, w2->w_name, i2, i1, i2, r1,
2064 i2, i1, r2);
2065 kdb_backtrace();
2066 printf("Witness disabled.\n");
2067 witness_watch = -1;
2068 }
2069 return (r1 & rmask);
2070 }
2071
2072 /*
2073 * Checks if @child is a direct child of @parent.
2074 */
2075 static int
2076 isitmychild(struct witness *parent, struct witness *child)
2077 {
2078
2079 return (_isitmyx(parent, child, WITNESS_PARENT, __func__));
2080 }
2081
2082 /*
2083 * Checks if @descendant is a direct or inderect descendant of @ancestor.
2084 */
2085 static int
2086 isitmydescendant(struct witness *ancestor, struct witness *descendant)
2087 {
2088
2089 return (_isitmyx(ancestor, descendant, WITNESS_ANCESTOR_MASK,
2090 __func__));
2091 }
2092
2093 static int
2094 blessed(struct witness *w1, struct witness *w2)
2095 {
2096 int i;
2097 struct witness_blessed *b;
2098
2099 for (i = 0; i < nitems(blessed_list); i++) {
2100 b = &blessed_list[i];
2101 if (strcmp(w1->w_name, b->b_lock1) == 0) {
2102 if (strcmp(w2->w_name, b->b_lock2) == 0)
2103 return (1);
2104 continue;
2105 }
2106 if (strcmp(w1->w_name, b->b_lock2) == 0)
2107 if (strcmp(w2->w_name, b->b_lock1) == 0)
2108 return (1);
2109 }
2110 return (0);
2111 }
2112
2113 static struct witness *
2114 witness_get(void)
2115 {
2116 struct witness *w;
2117 int index;
2118
2119 if (witness_cold == 0)
2120 mtx_assert(&w_mtx, MA_OWNED);
2121
2122 if (witness_watch == -1) {
2123 mtx_unlock_spin(&w_mtx);
2124 return (NULL);
2125 }
2126 if (STAILQ_EMPTY(&w_free)) {
2127 witness_watch = -1;
2128 mtx_unlock_spin(&w_mtx);
2129 printf("WITNESS: unable to allocate a new witness object\n");
2130 return (NULL);
2131 }
2132 w = STAILQ_FIRST(&w_free);
2133 STAILQ_REMOVE_HEAD(&w_free, w_list);
2134 w_free_cnt--;
2135 index = w->w_index;
2136 MPASS(index > 0 && index == w_max_used_index+1 &&
2137 index < witness_count);
2138 bzero(w, sizeof(*w));
2139 w->w_index = index;
2140 if (index > w_max_used_index)
2141 w_max_used_index = index;
2142 return (w);
2143 }
2144
2145 static void
2146 witness_free(struct witness *w)
2147 {
2148
2149 STAILQ_INSERT_HEAD(&w_free, w, w_list);
2150 w_free_cnt++;
2151 }
2152
2153 static struct lock_list_entry *
2154 witness_lock_list_get(void)
2155 {
2156 struct lock_list_entry *lle;
2157
2158 if (witness_watch == -1)
2159 return (NULL);
2160 mtx_lock_spin(&w_mtx);
2161 lle = w_lock_list_free;
2162 if (lle == NULL) {
2163 witness_watch = -1;
2164 mtx_unlock_spin(&w_mtx);
2165 printf("%s: witness exhausted\n", __func__);
2166 return (NULL);
2167 }
2168 w_lock_list_free = lle->ll_next;
2169 mtx_unlock_spin(&w_mtx);
2170 bzero(lle, sizeof(*lle));
2171 return (lle);
2172 }
2173
2174 static void
2175 witness_lock_list_free(struct lock_list_entry *lle)
2176 {
2177
2178 mtx_lock_spin(&w_mtx);
2179 lle->ll_next = w_lock_list_free;
2180 w_lock_list_free = lle;
2181 mtx_unlock_spin(&w_mtx);
2182 }
2183
2184 static struct lock_instance *
2185 find_instance(struct lock_list_entry *list, const struct lock_object *lock)
2186 {
2187 struct lock_list_entry *lle;
2188 struct lock_instance *instance;
2189 int i;
2190
2191 for (lle = list; lle != NULL; lle = lle->ll_next)
2192 for (i = lle->ll_count - 1; i >= 0; i--) {
2193 instance = &lle->ll_children[i];
2194 if (instance->li_lock == lock)
2195 return (instance);
2196 }
2197 return (NULL);
2198 }
2199
2200 static void
2201 witness_list_lock(struct lock_instance *instance,
2202 int (*prnt)(const char *fmt, ...))
2203 {
2204 struct lock_object *lock;
2205
2206 lock = instance->li_lock;
2207 prnt("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ?
2208 "exclusive" : "shared", LOCK_CLASS(lock)->lc_name, lock->lo_name);
2209 if (lock->lo_witness->w_name != lock->lo_name)
2210 prnt(" (%s)", lock->lo_witness->w_name);
2211 prnt(" r = %d (%p) locked @ %s:%d\n",
2212 instance->li_flags & LI_RECURSEMASK, lock,
2213 fixup_filename(instance->li_file), instance->li_line);
2214 }
2215
2216 static int
2217 witness_output(const char *fmt, ...)
2218 {
2219 va_list ap;
2220 int ret;
2221
2222 va_start(ap, fmt);
2223 ret = witness_voutput(fmt, ap);
2224 va_end(ap);
2225 return (ret);
2226 }
2227
2228 static int
2229 witness_voutput(const char *fmt, va_list ap)
2230 {
2231 int ret;
2232
2233 ret = 0;
2234 switch (witness_channel) {
2235 case WITNESS_CONSOLE:
2236 ret = vprintf(fmt, ap);
2237 break;
2238 case WITNESS_LOG:
2239 vlog(LOG_NOTICE, fmt, ap);
2240 break;
2241 case WITNESS_NONE:
2242 break;
2243 }
2244 return (ret);
2245 }
2246
2247 #ifdef DDB
2248 static int
2249 witness_thread_has_locks(struct thread *td)
2250 {
2251
2252 if (td->td_sleeplocks == NULL)
2253 return (0);
2254 return (td->td_sleeplocks->ll_count != 0);
2255 }
2256
2257 static int
2258 witness_proc_has_locks(struct proc *p)
2259 {
2260 struct thread *td;
2261
2262 FOREACH_THREAD_IN_PROC(p, td) {
2263 if (witness_thread_has_locks(td))
2264 return (1);
2265 }
2266 return (0);
2267 }
2268 #endif
2269
2270 int
2271 witness_list_locks(struct lock_list_entry **lock_list,
2272 int (*prnt)(const char *fmt, ...))
2273 {
2274 struct lock_list_entry *lle;
2275 int i, nheld;
2276
2277 nheld = 0;
2278 for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
2279 for (i = lle->ll_count - 1; i >= 0; i--) {
2280 witness_list_lock(&lle->ll_children[i], prnt);
2281 nheld++;
2282 }
2283 return (nheld);
2284 }
2285
2286 /*
2287 * This is a bit risky at best. We call this function when we have timed
2288 * out acquiring a spin lock, and we assume that the other CPU is stuck
2289 * with this lock held. So, we go groveling around in the other CPU's
2290 * per-cpu data to try to find the lock instance for this spin lock to
2291 * see when it was last acquired.
2292 */
2293 void
2294 witness_display_spinlock(struct lock_object *lock, struct thread *owner,
2295 int (*prnt)(const char *fmt, ...))
2296 {
2297 struct lock_instance *instance;
2298 struct pcpu *pc;
2299
2300 if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU)
2301 return;
2302 pc = pcpu_find(owner->td_oncpu);
2303 instance = find_instance(pc->pc_spinlocks, lock);
2304 if (instance != NULL)
2305 witness_list_lock(instance, prnt);
2306 }
2307
2308 void
2309 witness_save(struct lock_object *lock, const char **filep, int *linep)
2310 {
2311 struct lock_list_entry *lock_list;
2312 struct lock_instance *instance;
2313 struct lock_class *class;
2314
2315 /*
2316 * This function is used independently in locking code to deal with
2317 * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2318 * is gone.
2319 */
2320 if (SCHEDULER_STOPPED())
2321 return;
2322 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2323 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2324 return;
2325 class = LOCK_CLASS(lock);
2326 if (class->lc_flags & LC_SLEEPLOCK)
2327 lock_list = curthread->td_sleeplocks;
2328 else {
2329 if (witness_skipspin)
2330 return;
2331 lock_list = PCPU_GET(spinlocks);
2332 }
2333 instance = find_instance(lock_list, lock);
2334 if (instance == NULL) {
2335 kassert_panic("%s: lock (%s) %s not locked", __func__,
2336 class->lc_name, lock->lo_name);
2337 return;
2338 }
2339 *filep = instance->li_file;
2340 *linep = instance->li_line;
2341 }
2342
2343 void
2344 witness_restore(struct lock_object *lock, const char *file, int line)
2345 {
2346 struct lock_list_entry *lock_list;
2347 struct lock_instance *instance;
2348 struct lock_class *class;
2349
2350 /*
2351 * This function is used independently in locking code to deal with
2352 * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2353 * is gone.
2354 */
2355 if (SCHEDULER_STOPPED())
2356 return;
2357 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2358 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2359 return;
2360 class = LOCK_CLASS(lock);
2361 if (class->lc_flags & LC_SLEEPLOCK)
2362 lock_list = curthread->td_sleeplocks;
2363 else {
2364 if (witness_skipspin)
2365 return;
2366 lock_list = PCPU_GET(spinlocks);
2367 }
2368 instance = find_instance(lock_list, lock);
2369 if (instance == NULL)
2370 kassert_panic("%s: lock (%s) %s not locked", __func__,
2371 class->lc_name, lock->lo_name);
2372 lock->lo_witness->w_file = file;
2373 lock->lo_witness->w_line = line;
2374 if (instance == NULL)
2375 return;
2376 instance->li_file = file;
2377 instance->li_line = line;
2378 }
2379
2380 void
2381 witness_assert(const struct lock_object *lock, int flags, const char *file,
2382 int line)
2383 {
2384 #ifdef INVARIANT_SUPPORT
2385 struct lock_instance *instance;
2386 struct lock_class *class;
2387
2388 if (lock->lo_witness == NULL || witness_watch < 1 || panicstr != NULL)
2389 return;
2390 class = LOCK_CLASS(lock);
2391 if ((class->lc_flags & LC_SLEEPLOCK) != 0)
2392 instance = find_instance(curthread->td_sleeplocks, lock);
2393 else if ((class->lc_flags & LC_SPINLOCK) != 0)
2394 instance = find_instance(PCPU_GET(spinlocks), lock);
2395 else {
2396 kassert_panic("Lock (%s) %s is not sleep or spin!",
2397 class->lc_name, lock->lo_name);
2398 return;
2399 }
2400 switch (flags) {
2401 case LA_UNLOCKED:
2402 if (instance != NULL)
2403 kassert_panic("Lock (%s) %s locked @ %s:%d.",
2404 class->lc_name, lock->lo_name,
2405 fixup_filename(file), line);
2406 break;
2407 case LA_LOCKED:
2408 case LA_LOCKED | LA_RECURSED:
2409 case LA_LOCKED | LA_NOTRECURSED:
2410 case LA_SLOCKED:
2411 case LA_SLOCKED | LA_RECURSED:
2412 case LA_SLOCKED | LA_NOTRECURSED:
2413 case LA_XLOCKED:
2414 case LA_XLOCKED | LA_RECURSED:
2415 case LA_XLOCKED | LA_NOTRECURSED:
2416 if (instance == NULL) {
2417 kassert_panic("Lock (%s) %s not locked @ %s:%d.",
2418 class->lc_name, lock->lo_name,
2419 fixup_filename(file), line);
2420 break;
2421 }
2422 if ((flags & LA_XLOCKED) != 0 &&
2423 (instance->li_flags & LI_EXCLUSIVE) == 0)
2424 kassert_panic(
2425 "Lock (%s) %s not exclusively locked @ %s:%d.",
2426 class->lc_name, lock->lo_name,
2427 fixup_filename(file), line);
2428 if ((flags & LA_SLOCKED) != 0 &&
2429 (instance->li_flags & LI_EXCLUSIVE) != 0)
2430 kassert_panic(
2431 "Lock (%s) %s exclusively locked @ %s:%d.",
2432 class->lc_name, lock->lo_name,
2433 fixup_filename(file), line);
2434 if ((flags & LA_RECURSED) != 0 &&
2435 (instance->li_flags & LI_RECURSEMASK) == 0)
2436 kassert_panic("Lock (%s) %s not recursed @ %s:%d.",
2437 class->lc_name, lock->lo_name,
2438 fixup_filename(file), line);
2439 if ((flags & LA_NOTRECURSED) != 0 &&
2440 (instance->li_flags & LI_RECURSEMASK) != 0)
2441 kassert_panic("Lock (%s) %s recursed @ %s:%d.",
2442 class->lc_name, lock->lo_name,
2443 fixup_filename(file), line);
2444 break;
2445 default:
2446 kassert_panic("Invalid lock assertion at %s:%d.",
2447 fixup_filename(file), line);
2448
2449 }
2450 #endif /* INVARIANT_SUPPORT */
2451 }
2452
2453 static void
2454 witness_setflag(struct lock_object *lock, int flag, int set)
2455 {
2456 struct lock_list_entry *lock_list;
2457 struct lock_instance *instance;
2458 struct lock_class *class;
2459
2460 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2461 return;
2462 class = LOCK_CLASS(lock);
2463 if (class->lc_flags & LC_SLEEPLOCK)
2464 lock_list = curthread->td_sleeplocks;
2465 else {
2466 if (witness_skipspin)
2467 return;
2468 lock_list = PCPU_GET(spinlocks);
2469 }
2470 instance = find_instance(lock_list, lock);
2471 if (instance == NULL) {
2472 kassert_panic("%s: lock (%s) %s not locked", __func__,
2473 class->lc_name, lock->lo_name);
2474 return;
2475 }
2476
2477 if (set)
2478 instance->li_flags |= flag;
2479 else
2480 instance->li_flags &= ~flag;
2481 }
2482
2483 void
2484 witness_norelease(struct lock_object *lock)
2485 {
2486
2487 witness_setflag(lock, LI_NORELEASE, 1);
2488 }
2489
2490 void
2491 witness_releaseok(struct lock_object *lock)
2492 {
2493
2494 witness_setflag(lock, LI_NORELEASE, 0);
2495 }
2496
2497 #ifdef DDB
2498 static void
2499 witness_ddb_list(struct thread *td)
2500 {
2501
2502 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2503 KASSERT(kdb_active, ("%s: not in the debugger", __func__));
2504
2505 if (witness_watch < 1)
2506 return;
2507
2508 witness_list_locks(&td->td_sleeplocks, db_printf);
2509
2510 /*
2511 * We only handle spinlocks if td == curthread. This is somewhat broken
2512 * if td is currently executing on some other CPU and holds spin locks
2513 * as we won't display those locks. If we had a MI way of getting
2514 * the per-cpu data for a given cpu then we could use
2515 * td->td_oncpu to get the list of spinlocks for this thread
2516 * and "fix" this.
2517 *
2518 * That still wouldn't really fix this unless we locked the scheduler
2519 * lock or stopped the other CPU to make sure it wasn't changing the
2520 * list out from under us. It is probably best to just not try to
2521 * handle threads on other CPU's for now.
2522 */
2523 if (td == curthread && PCPU_GET(spinlocks) != NULL)
2524 witness_list_locks(PCPU_PTR(spinlocks), db_printf);
2525 }
2526
2527 DB_SHOW_COMMAND(locks, db_witness_list)
2528 {
2529 struct thread *td;
2530
2531 if (have_addr)
2532 td = db_lookup_thread(addr, true);
2533 else
2534 td = kdb_thread;
2535 witness_ddb_list(td);
2536 }
2537
2538 DB_SHOW_ALL_COMMAND(locks, db_witness_list_all)
2539 {
2540 struct thread *td;
2541 struct proc *p;
2542
2543 /*
2544 * It would be nice to list only threads and processes that actually
2545 * held sleep locks, but that information is currently not exported
2546 * by WITNESS.
2547 */
2548 FOREACH_PROC_IN_SYSTEM(p) {
2549 if (!witness_proc_has_locks(p))
2550 continue;
2551 FOREACH_THREAD_IN_PROC(p, td) {
2552 if (!witness_thread_has_locks(td))
2553 continue;
2554 db_printf("Process %d (%s) thread %p (%d)\n", p->p_pid,
2555 p->p_comm, td, td->td_tid);
2556 witness_ddb_list(td);
2557 if (db_pager_quit)
2558 return;
2559 }
2560 }
2561 }
2562 DB_SHOW_ALIAS(alllocks, db_witness_list_all)
2563
2564 DB_SHOW_COMMAND(witness, db_witness_display)
2565 {
2566
2567 witness_ddb_display(db_printf);
2568 }
2569 #endif
2570
2571 static void
2572 sbuf_print_witness_badstacks(struct sbuf *sb, size_t *oldidx)
2573 {
2574 struct witness_lock_order_data *data1, *data2, *tmp_data1, *tmp_data2;
2575 struct witness *tmp_w1, *tmp_w2, *w1, *w2;
2576 int generation, i, j;
2577
2578 tmp_data1 = NULL;
2579 tmp_data2 = NULL;
2580 tmp_w1 = NULL;
2581 tmp_w2 = NULL;
2582
2583 /* Allocate and init temporary storage space. */
2584 tmp_w1 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2585 tmp_w2 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2586 tmp_data1 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2587 M_WAITOK | M_ZERO);
2588 tmp_data2 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2589 M_WAITOK | M_ZERO);
2590 stack_zero(&tmp_data1->wlod_stack);
2591 stack_zero(&tmp_data2->wlod_stack);
2592
2593 restart:
2594 mtx_lock_spin(&w_mtx);
2595 generation = w_generation;
2596 mtx_unlock_spin(&w_mtx);
2597 sbuf_printf(sb, "Number of known direct relationships is %d\n",
2598 w_lohash.wloh_count);
2599 for (i = 1; i < w_max_used_index; i++) {
2600 mtx_lock_spin(&w_mtx);
2601 if (generation != w_generation) {
2602 mtx_unlock_spin(&w_mtx);
2603
2604 /* The graph has changed, try again. */
2605 *oldidx = 0;
2606 sbuf_clear(sb);
2607 goto restart;
2608 }
2609
2610 w1 = &w_data[i];
2611 if (w1->w_reversed == 0) {
2612 mtx_unlock_spin(&w_mtx);
2613 continue;
2614 }
2615
2616 /* Copy w1 locally so we can release the spin lock. */
2617 *tmp_w1 = *w1;
2618 mtx_unlock_spin(&w_mtx);
2619
2620 if (tmp_w1->w_reversed == 0)
2621 continue;
2622 for (j = 1; j < w_max_used_index; j++) {
2623 if ((w_rmatrix[i][j] & WITNESS_REVERSAL) == 0 || i > j)
2624 continue;
2625
2626 mtx_lock_spin(&w_mtx);
2627 if (generation != w_generation) {
2628 mtx_unlock_spin(&w_mtx);
2629
2630 /* The graph has changed, try again. */
2631 *oldidx = 0;
2632 sbuf_clear(sb);
2633 goto restart;
2634 }
2635
2636 w2 = &w_data[j];
2637 data1 = witness_lock_order_get(w1, w2);
2638 data2 = witness_lock_order_get(w2, w1);
2639
2640 /*
2641 * Copy information locally so we can release the
2642 * spin lock.
2643 */
2644 *tmp_w2 = *w2;
2645
2646 if (data1) {
2647 stack_zero(&tmp_data1->wlod_stack);
2648 stack_copy(&data1->wlod_stack,
2649 &tmp_data1->wlod_stack);
2650 }
2651 if (data2 && data2 != data1) {
2652 stack_zero(&tmp_data2->wlod_stack);
2653 stack_copy(&data2->wlod_stack,
2654 &tmp_data2->wlod_stack);
2655 }
2656 mtx_unlock_spin(&w_mtx);
2657
2658 if (blessed(tmp_w1, tmp_w2))
2659 continue;
2660
2661 sbuf_printf(sb,
2662 "\nLock order reversal between \"%s\"(%s) and \"%s\"(%s)!\n",
2663 tmp_w1->w_name, tmp_w1->w_class->lc_name,
2664 tmp_w2->w_name, tmp_w2->w_class->lc_name);
2665 if (data1) {
2666 sbuf_printf(sb,
2667 "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2668 tmp_w1->w_name, tmp_w1->w_class->lc_name,
2669 tmp_w2->w_name, tmp_w2->w_class->lc_name);
2670 stack_sbuf_print(sb, &tmp_data1->wlod_stack);
2671 sbuf_printf(sb, "\n");
2672 }
2673 if (data2 && data2 != data1) {
2674 sbuf_printf(sb,
2675 "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2676 tmp_w2->w_name, tmp_w2->w_class->lc_name,
2677 tmp_w1->w_name, tmp_w1->w_class->lc_name);
2678 stack_sbuf_print(sb, &tmp_data2->wlod_stack);
2679 sbuf_printf(sb, "\n");
2680 }
2681 }
2682 }
2683 mtx_lock_spin(&w_mtx);
2684 if (generation != w_generation) {
2685 mtx_unlock_spin(&w_mtx);
2686
2687 /*
2688 * The graph changed while we were printing stack data,
2689 * try again.
2690 */
2691 *oldidx = 0;
2692 sbuf_clear(sb);
2693 goto restart;
2694 }
2695 mtx_unlock_spin(&w_mtx);
2696
2697 /* Free temporary storage space. */
2698 free(tmp_data1, M_TEMP);
2699 free(tmp_data2, M_TEMP);
2700 free(tmp_w1, M_TEMP);
2701 free(tmp_w2, M_TEMP);
2702 }
2703
2704 static int
2705 sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS)
2706 {
2707 struct sbuf *sb;
2708 int error;
2709
2710 if (witness_watch < 1) {
2711 error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2712 return (error);
2713 }
2714 if (witness_cold) {
2715 error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2716 return (error);
2717 }
2718 error = 0;
2719 sb = sbuf_new(NULL, NULL, badstack_sbuf_size, SBUF_AUTOEXTEND);
2720 if (sb == NULL)
2721 return (ENOMEM);
2722
2723 sbuf_print_witness_badstacks(sb, &req->oldidx);
2724
2725 sbuf_finish(sb);
2726 error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1);
2727 sbuf_delete(sb);
2728
2729 return (error);
2730 }
2731
2732 #ifdef DDB
2733 static int
2734 sbuf_db_printf_drain(void *arg __unused, const char *data, int len)
2735 {
2736
2737 return (db_printf("%.*s", len, data));
2738 }
2739
2740 DB_SHOW_COMMAND(badstacks, db_witness_badstacks)
2741 {
2742 struct sbuf sb;
2743 char buffer[128];
2744 size_t dummy;
2745
2746 sbuf_new(&sb, buffer, sizeof(buffer), SBUF_FIXEDLEN);
2747 sbuf_set_drain(&sb, sbuf_db_printf_drain, NULL);
2748 sbuf_print_witness_badstacks(&sb, &dummy);
2749 sbuf_finish(&sb);
2750 }
2751 #endif
2752
2753 static int
2754 sysctl_debug_witness_channel(SYSCTL_HANDLER_ARGS)
2755 {
2756 static const struct {
2757 enum witness_channel channel;
2758 const char *name;
2759 } channels[] = {
2760 { WITNESS_CONSOLE, "console" },
2761 { WITNESS_LOG, "log" },
2762 { WITNESS_NONE, "none" },
2763 };
2764 char buf[16];
2765 u_int i;
2766 int error;
2767
2768 buf[0] = '\0';
2769 for (i = 0; i < nitems(channels); i++)
2770 if (witness_channel == channels[i].channel) {
2771 snprintf(buf, sizeof(buf), "%s", channels[i].name);
2772 break;
2773 }
2774
2775 error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
2776 if (error != 0 || req->newptr == NULL)
2777 return (error);
2778
2779 error = EINVAL;
2780 for (i = 0; i < nitems(channels); i++)
2781 if (strcmp(channels[i].name, buf) == 0) {
2782 witness_channel = channels[i].channel;
2783 error = 0;
2784 break;
2785 }
2786 return (error);
2787 }
2788
2789 static int
2790 sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS)
2791 {
2792 struct witness *w;
2793 struct sbuf *sb;
2794 int error;
2795
2796 #ifdef __i386__
2797 error = SYSCTL_OUT(req, w_notallowed, sizeof(w_notallowed));
2798 return (error);
2799 #endif
2800
2801 if (witness_watch < 1) {
2802 error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2803 return (error);
2804 }
2805 if (witness_cold) {
2806 error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2807 return (error);
2808 }
2809 error = 0;
2810
2811 error = sysctl_wire_old_buffer(req, 0);
2812 if (error != 0)
2813 return (error);
2814 sb = sbuf_new_for_sysctl(NULL, NULL, FULLGRAPH_SBUF_SIZE, req);
2815 if (sb == NULL)
2816 return (ENOMEM);
2817 sbuf_printf(sb, "\n");
2818
2819 mtx_lock_spin(&w_mtx);
2820 STAILQ_FOREACH(w, &w_all, w_list)
2821 w->w_displayed = 0;
2822 STAILQ_FOREACH(w, &w_all, w_list)
2823 witness_add_fullgraph(sb, w);
2824 mtx_unlock_spin(&w_mtx);
2825
2826 /*
2827 * Close the sbuf and return to userland.
2828 */
2829 error = sbuf_finish(sb);
2830 sbuf_delete(sb);
2831
2832 return (error);
2833 }
2834
2835 static int
2836 sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS)
2837 {
2838 int error, value;
2839
2840 value = witness_watch;
2841 error = sysctl_handle_int(oidp, &value, 0, req);
2842 if (error != 0 || req->newptr == NULL)
2843 return (error);
2844 if (value > 1 || value < -1 ||
2845 (witness_watch == -1 && value != witness_watch))
2846 return (EINVAL);
2847 witness_watch = value;
2848 return (0);
2849 }
2850
2851 static void
2852 witness_add_fullgraph(struct sbuf *sb, struct witness *w)
2853 {
2854 int i;
2855
2856 if (w->w_displayed != 0 || (w->w_file == NULL && w->w_line == 0))
2857 return;
2858 w->w_displayed = 1;
2859
2860 WITNESS_INDEX_ASSERT(w->w_index);
2861 for (i = 1; i <= w_max_used_index; i++) {
2862 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT) {
2863 sbuf_printf(sb, "\"%s\",\"%s\"\n", w->w_name,
2864 w_data[i].w_name);
2865 witness_add_fullgraph(sb, &w_data[i]);
2866 }
2867 }
2868 }
2869
2870 /*
2871 * A simple hash function. Takes a key pointer and a key size. If size == 0,
2872 * interprets the key as a string and reads until the null
2873 * terminator. Otherwise, reads the first size bytes. Returns an unsigned 32-bit
2874 * hash value computed from the key.
2875 */
2876 static uint32_t
2877 witness_hash_djb2(const uint8_t *key, uint32_t size)
2878 {
2879 unsigned int hash = 5381;
2880 int i;
2881
2882 /* hash = hash * 33 + key[i] */
2883 if (size)
2884 for (i = 0; i < size; i++)
2885 hash = ((hash << 5) + hash) + (unsigned int)key[i];
2886 else
2887 for (i = 0; key[i] != 0; i++)
2888 hash = ((hash << 5) + hash) + (unsigned int)key[i];
2889
2890 return (hash);
2891 }
2892
2893
2894 /*
2895 * Initializes the two witness hash tables. Called exactly once from
2896 * witness_initialize().
2897 */
2898 static void
2899 witness_init_hash_tables(void)
2900 {
2901 int i;
2902
2903 MPASS(witness_cold);
2904
2905 /* Initialize the hash tables. */
2906 for (i = 0; i < WITNESS_HASH_SIZE; i++)
2907 w_hash.wh_array[i] = NULL;
2908
2909 w_hash.wh_size = WITNESS_HASH_SIZE;
2910 w_hash.wh_count = 0;
2911
2912 /* Initialize the lock order data hash. */
2913 w_lofree = NULL;
2914 for (i = 0; i < WITNESS_LO_DATA_COUNT; i++) {
2915 memset(&w_lodata[i], 0, sizeof(w_lodata[i]));
2916 w_lodata[i].wlod_next = w_lofree;
2917 w_lofree = &w_lodata[i];
2918 }
2919 w_lohash.wloh_size = WITNESS_LO_HASH_SIZE;
2920 w_lohash.wloh_count = 0;
2921 for (i = 0; i < WITNESS_LO_HASH_SIZE; i++)
2922 w_lohash.wloh_array[i] = NULL;
2923 }
2924
2925 static struct witness *
2926 witness_hash_get(const char *key)
2927 {
2928 struct witness *w;
2929 uint32_t hash;
2930
2931 MPASS(key != NULL);
2932 if (witness_cold == 0)
2933 mtx_assert(&w_mtx, MA_OWNED);
2934 hash = witness_hash_djb2(key, 0) % w_hash.wh_size;
2935 w = w_hash.wh_array[hash];
2936 while (w != NULL) {
2937 if (strcmp(w->w_name, key) == 0)
2938 goto out;
2939 w = w->w_hash_next;
2940 }
2941
2942 out:
2943 return (w);
2944 }
2945
2946 static void
2947 witness_hash_put(struct witness *w)
2948 {
2949 uint32_t hash;
2950
2951 MPASS(w != NULL);
2952 MPASS(w->w_name != NULL);
2953 if (witness_cold == 0)
2954 mtx_assert(&w_mtx, MA_OWNED);
2955 KASSERT(witness_hash_get(w->w_name) == NULL,
2956 ("%s: trying to add a hash entry that already exists!", __func__));
2957 KASSERT(w->w_hash_next == NULL,
2958 ("%s: w->w_hash_next != NULL", __func__));
2959
2960 hash = witness_hash_djb2(w->w_name, 0) % w_hash.wh_size;
2961 w->w_hash_next = w_hash.wh_array[hash];
2962 w_hash.wh_array[hash] = w;
2963 w_hash.wh_count++;
2964 }
2965
2966
2967 static struct witness_lock_order_data *
2968 witness_lock_order_get(struct witness *parent, struct witness *child)
2969 {
2970 struct witness_lock_order_data *data = NULL;
2971 struct witness_lock_order_key key;
2972 unsigned int hash;
2973
2974 MPASS(parent != NULL && child != NULL);
2975 key.from = parent->w_index;
2976 key.to = child->w_index;
2977 WITNESS_INDEX_ASSERT(key.from);
2978 WITNESS_INDEX_ASSERT(key.to);
2979 if ((w_rmatrix[parent->w_index][child->w_index]
2980 & WITNESS_LOCK_ORDER_KNOWN) == 0)
2981 goto out;
2982
2983 hash = witness_hash_djb2((const char*)&key,
2984 sizeof(key)) % w_lohash.wloh_size;
2985 data = w_lohash.wloh_array[hash];
2986 while (data != NULL) {
2987 if (witness_lock_order_key_equal(&data->wlod_key, &key))
2988 break;
2989 data = data->wlod_next;
2990 }
2991
2992 out:
2993 return (data);
2994 }
2995
2996 /*
2997 * Verify that parent and child have a known relationship, are not the same,
2998 * and child is actually a child of parent. This is done without w_mtx
2999 * to avoid contention in the common case.
3000 */
3001 static int
3002 witness_lock_order_check(struct witness *parent, struct witness *child)
3003 {
3004
3005 if (parent != child &&
3006 w_rmatrix[parent->w_index][child->w_index]
3007 & WITNESS_LOCK_ORDER_KNOWN &&
3008 isitmychild(parent, child))
3009 return (1);
3010
3011 return (0);
3012 }
3013
3014 static int
3015 witness_lock_order_add(struct witness *parent, struct witness *child)
3016 {
3017 struct witness_lock_order_data *data = NULL;
3018 struct witness_lock_order_key key;
3019 unsigned int hash;
3020
3021 MPASS(parent != NULL && child != NULL);
3022 key.from = parent->w_index;
3023 key.to = child->w_index;
3024 WITNESS_INDEX_ASSERT(key.from);
3025 WITNESS_INDEX_ASSERT(key.to);
3026 if (w_rmatrix[parent->w_index][child->w_index]
3027 & WITNESS_LOCK_ORDER_KNOWN)
3028 return (1);
3029
3030 hash = witness_hash_djb2((const char*)&key,
3031 sizeof(key)) % w_lohash.wloh_size;
3032 w_rmatrix[parent->w_index][child->w_index] |= WITNESS_LOCK_ORDER_KNOWN;
3033 data = w_lofree;
3034 if (data == NULL)
3035 return (0);
3036 w_lofree = data->wlod_next;
3037 data->wlod_next = w_lohash.wloh_array[hash];
3038 data->wlod_key = key;
3039 w_lohash.wloh_array[hash] = data;
3040 w_lohash.wloh_count++;
3041 stack_zero(&data->wlod_stack);
3042 stack_save(&data->wlod_stack);
3043 return (1);
3044 }
3045
3046 /* Call this whenever the structure of the witness graph changes. */
3047 static void
3048 witness_increment_graph_generation(void)
3049 {
3050
3051 if (witness_cold == 0)
3052 mtx_assert(&w_mtx, MA_OWNED);
3053 w_generation++;
3054 }
3055
3056 static int
3057 witness_output_drain(void *arg __unused, const char *data, int len)
3058 {
3059
3060 witness_output("%.*s", len, data);
3061 return (len);
3062 }
3063
3064 static void
3065 witness_debugger(int cond, const char *msg)
3066 {
3067 char buf[32];
3068 struct sbuf sb;
3069 struct stack st;
3070
3071 if (!cond)
3072 return;
3073
3074 if (witness_trace) {
3075 sbuf_new(&sb, buf, sizeof(buf), SBUF_FIXEDLEN);
3076 sbuf_set_drain(&sb, witness_output_drain, NULL);
3077
3078 stack_zero(&st);
3079 stack_save(&st);
3080 witness_output("stack backtrace:\n");
3081 stack_sbuf_print_ddb(&sb, &st);
3082
3083 sbuf_finish(&sb);
3084 }
3085
3086 #ifdef KDB
3087 if (witness_kdb)
3088 kdb_enter(KDB_WHY_WITNESS, msg);
3089 #endif
3090 }
Cache object: 2ce8a5c05e2f77fc222d209a5b1a448e
|