1 /*-
2 * Copyright (c) 2008 Isilon Systems, Inc.
3 * Copyright (c) 2008 Ilya Maykov <ivmaykov@gmail.com>
4 * Copyright (c) 1998 Berkeley Software Design, Inc.
5 * All rights reserved.
6 *
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
9 * are met:
10 * 1. Redistributions of source code must retain the above copyright
11 * notice, this list of conditions and the following disclaimer.
12 * 2. Redistributions in binary form must reproduce the above copyright
13 * notice, this list of conditions and the following disclaimer in the
14 * documentation and/or other materials provided with the distribution.
15 * 3. Berkeley Software Design Inc's name may not be used to endorse or
16 * promote products derived from this software without specific prior
17 * written permission.
18 *
19 * THIS SOFTWARE IS PROVIDED BY BERKELEY SOFTWARE DESIGN INC ``AS IS'' AND
20 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
21 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
22 * ARE DISCLAIMED. IN NO EVENT SHALL BERKELEY SOFTWARE DESIGN INC BE LIABLE
23 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
24 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
25 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
26 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
27 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
28 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
29 * SUCH DAMAGE.
30 *
31 * from BSDI $Id: mutex_witness.c,v 1.1.2.20 2000/04/27 03:10:27 cp Exp $
32 * and BSDI $Id: synch_machdep.c,v 2.3.2.39 2000/04/27 03:10:25 cp Exp $
33 */
34
35 /*
36 * Implementation of the `witness' lock verifier. Originally implemented for
37 * mutexes in BSD/OS. Extended to handle generic lock objects and lock
38 * classes in FreeBSD.
39 */
40
41 /*
42 * Main Entry: witness
43 * Pronunciation: 'wit-n&s
44 * Function: noun
45 * Etymology: Middle English witnesse, from Old English witnes knowledge,
46 * testimony, witness, from 2wit
47 * Date: before 12th century
48 * 1 : attestation of a fact or event : TESTIMONY
49 * 2 : one that gives evidence; specifically : one who testifies in
50 * a cause or before a judicial tribunal
51 * 3 : one asked to be present at a transaction so as to be able to
52 * testify to its having taken place
53 * 4 : one who has personal knowledge of something
54 * 5 a : something serving as evidence or proof : SIGN
55 * b : public affirmation by word or example of usually
56 * religious faith or conviction <the heroic witness to divine
57 * life -- Pilot>
58 * 6 capitalized : a member of the Jehovah's Witnesses
59 */
60
61 /*
62 * Special rules concerning Giant and lock orders:
63 *
64 * 1) Giant must be acquired before any other mutexes. Stated another way,
65 * no other mutex may be held when Giant is acquired.
66 *
67 * 2) Giant must be released when blocking on a sleepable lock.
68 *
69 * This rule is less obvious, but is a result of Giant providing the same
70 * semantics as spl(). Basically, when a thread sleeps, it must release
71 * Giant. When a thread blocks on a sleepable lock, it sleeps. Hence rule
72 * 2).
73 *
74 * 3) Giant may be acquired before or after sleepable locks.
75 *
76 * This rule is also not quite as obvious. Giant may be acquired after
77 * a sleepable lock because it is a non-sleepable lock and non-sleepable
78 * locks may always be acquired while holding a sleepable lock. The second
79 * case, Giant before a sleepable lock, follows from rule 2) above. Suppose
80 * you have two threads T1 and T2 and a sleepable lock X. Suppose that T1
81 * acquires X and blocks on Giant. Then suppose that T2 acquires Giant and
82 * blocks on X. When T2 blocks on X, T2 will release Giant allowing T1 to
83 * execute. Thus, acquiring Giant both before and after a sleepable lock
84 * will not result in a lock order reversal.
85 */
86
87 #include <sys/cdefs.h>
88 __FBSDID("$FreeBSD: releng/8.0/sys/kern/subr_witness.c 192416 2009-05-20 02:51:48Z kmacy $");
89
90 #include "opt_ddb.h"
91 #include "opt_hwpmc_hooks.h"
92 #include "opt_stack.h"
93 #include "opt_witness.h"
94
95 #include <sys/param.h>
96 #include <sys/bus.h>
97 #include <sys/kdb.h>
98 #include <sys/kernel.h>
99 #include <sys/ktr.h>
100 #include <sys/lock.h>
101 #include <sys/malloc.h>
102 #include <sys/mutex.h>
103 #include <sys/priv.h>
104 #include <sys/proc.h>
105 #include <sys/sbuf.h>
106 #include <sys/sched.h>
107 #include <sys/stack.h>
108 #include <sys/sysctl.h>
109 #include <sys/systm.h>
110
111 #ifdef DDB
112 #include <ddb/ddb.h>
113 #endif
114
115 #include <machine/stdarg.h>
116
117 #if !defined(DDB) && !defined(STACK)
118 #error "DDB or STACK options are required for WITNESS"
119 #endif
120
121 /* Note that these traces do not work with KTR_ALQ. */
122 #if 0
123 #define KTR_WITNESS KTR_SUBSYS
124 #else
125 #define KTR_WITNESS 0
126 #endif
127
128 #define LI_RECURSEMASK 0x0000ffff /* Recursion depth of lock instance. */
129 #define LI_EXCLUSIVE 0x00010000 /* Exclusive lock instance. */
130 #define LI_NORELEASE 0x00020000 /* Lock not allowed to be released. */
131
132 /* Define this to check for blessed mutexes */
133 #undef BLESSING
134
135 #define WITNESS_COUNT 1024
136 #define WITNESS_CHILDCOUNT (WITNESS_COUNT * 4)
137 #define WITNESS_HASH_SIZE 251 /* Prime, gives load factor < 2 */
138 #define WITNESS_PENDLIST 512
139
140 /* Allocate 256 KB of stack data space */
141 #define WITNESS_LO_DATA_COUNT 2048
142
143 /* Prime, gives load factor of ~2 at full load */
144 #define WITNESS_LO_HASH_SIZE 1021
145
146 /*
147 * XXX: This is somewhat bogus, as we assume here that at most 2048 threads
148 * will hold LOCK_NCHILDREN locks. We handle failure ok, and we should
149 * probably be safe for the most part, but it's still a SWAG.
150 */
151 #define LOCK_NCHILDREN 5
152 #define LOCK_CHILDCOUNT 2048
153
154 #define MAX_W_NAME 64
155
156 #define BADSTACK_SBUF_SIZE (256 * WITNESS_COUNT)
157 #define CYCLEGRAPH_SBUF_SIZE 8192
158 #define FULLGRAPH_SBUF_SIZE 32768
159
160 /*
161 * These flags go in the witness relationship matrix and describe the
162 * relationship between any two struct witness objects.
163 */
164 #define WITNESS_UNRELATED 0x00 /* No lock order relation. */
165 #define WITNESS_PARENT 0x01 /* Parent, aka direct ancestor. */
166 #define WITNESS_ANCESTOR 0x02 /* Direct or indirect ancestor. */
167 #define WITNESS_CHILD 0x04 /* Child, aka direct descendant. */
168 #define WITNESS_DESCENDANT 0x08 /* Direct or indirect descendant. */
169 #define WITNESS_ANCESTOR_MASK (WITNESS_PARENT | WITNESS_ANCESTOR)
170 #define WITNESS_DESCENDANT_MASK (WITNESS_CHILD | WITNESS_DESCENDANT)
171 #define WITNESS_RELATED_MASK \
172 (WITNESS_ANCESTOR_MASK | WITNESS_DESCENDANT_MASK)
173 #define WITNESS_REVERSAL 0x10 /* A lock order reversal has been
174 * observed. */
175 #define WITNESS_RESERVED1 0x20 /* Unused flag, reserved. */
176 #define WITNESS_RESERVED2 0x40 /* Unused flag, reserved. */
177 #define WITNESS_LOCK_ORDER_KNOWN 0x80 /* This lock order is known. */
178
179 /* Descendant to ancestor flags */
180 #define WITNESS_DTOA(x) (((x) & WITNESS_RELATED_MASK) >> 2)
181
182 /* Ancestor to descendant flags */
183 #define WITNESS_ATOD(x) (((x) & WITNESS_RELATED_MASK) << 2)
184
185 #define WITNESS_INDEX_ASSERT(i) \
186 MPASS((i) > 0 && (i) <= w_max_used_index && (i) < WITNESS_COUNT)
187
188 MALLOC_DEFINE(M_WITNESS, "Witness", "Witness");
189
190 /*
191 * Lock instances. A lock instance is the data associated with a lock while
192 * it is held by witness. For example, a lock instance will hold the
193 * recursion count of a lock. Lock instances are held in lists. Spin locks
194 * are held in a per-cpu list while sleep locks are held in per-thread list.
195 */
196 struct lock_instance {
197 struct lock_object *li_lock;
198 const char *li_file;
199 int li_line;
200 u_int li_flags;
201 };
202
203 /*
204 * A simple list type used to build the list of locks held by a thread
205 * or CPU. We can't simply embed the list in struct lock_object since a
206 * lock may be held by more than one thread if it is a shared lock. Locks
207 * are added to the head of the list, so we fill up each list entry from
208 * "the back" logically. To ease some of the arithmetic, we actually fill
209 * in each list entry the normal way (children[0] then children[1], etc.) but
210 * when we traverse the list we read children[count-1] as the first entry
211 * down to children[0] as the final entry.
212 */
213 struct lock_list_entry {
214 struct lock_list_entry *ll_next;
215 struct lock_instance ll_children[LOCK_NCHILDREN];
216 u_int ll_count;
217 };
218
219 /*
220 * The main witness structure. One of these per named lock type in the system
221 * (for example, "vnode interlock").
222 */
223 struct witness {
224 char w_name[MAX_W_NAME];
225 uint32_t w_index; /* Index in the relationship matrix */
226 struct lock_class *w_class;
227 STAILQ_ENTRY(witness) w_list; /* List of all witnesses. */
228 STAILQ_ENTRY(witness) w_typelist; /* Witnesses of a type. */
229 struct witness *w_hash_next; /* Linked list in hash buckets. */
230 const char *w_file; /* File where last acquired */
231 uint32_t w_line; /* Line where last acquired */
232 uint32_t w_refcount;
233 uint16_t w_num_ancestors; /* direct/indirect
234 * ancestor count */
235 uint16_t w_num_descendants; /* direct/indirect
236 * descendant count */
237 int16_t w_ddb_level;
238 unsigned w_displayed:1;
239 unsigned w_reversed:1;
240 };
241
242 STAILQ_HEAD(witness_list, witness);
243
244 /*
245 * The witness hash table. Keys are witness names (const char *), elements are
246 * witness objects (struct witness *).
247 */
248 struct witness_hash {
249 struct witness *wh_array[WITNESS_HASH_SIZE];
250 uint32_t wh_size;
251 uint32_t wh_count;
252 };
253
254 /*
255 * Key type for the lock order data hash table.
256 */
257 struct witness_lock_order_key {
258 uint16_t from;
259 uint16_t to;
260 };
261
262 struct witness_lock_order_data {
263 struct stack wlod_stack;
264 struct witness_lock_order_key wlod_key;
265 struct witness_lock_order_data *wlod_next;
266 };
267
268 /*
269 * The witness lock order data hash table. Keys are witness index tuples
270 * (struct witness_lock_order_key), elements are lock order data objects
271 * (struct witness_lock_order_data).
272 */
273 struct witness_lock_order_hash {
274 struct witness_lock_order_data *wloh_array[WITNESS_LO_HASH_SIZE];
275 u_int wloh_size;
276 u_int wloh_count;
277 };
278
279 #ifdef BLESSING
280 struct witness_blessed {
281 const char *b_lock1;
282 const char *b_lock2;
283 };
284 #endif
285
286 struct witness_pendhelp {
287 const char *wh_type;
288 struct lock_object *wh_lock;
289 };
290
291 struct witness_order_list_entry {
292 const char *w_name;
293 struct lock_class *w_class;
294 };
295
296 /*
297 * Returns 0 if one of the locks is a spin lock and the other is not.
298 * Returns 1 otherwise.
299 */
300 static __inline int
301 witness_lock_type_equal(struct witness *w1, struct witness *w2)
302 {
303
304 return ((w1->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) ==
305 (w2->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)));
306 }
307
308 static __inline int
309 witness_lock_order_key_empty(const struct witness_lock_order_key *key)
310 {
311
312 return (key->from == 0 && key->to == 0);
313 }
314
315 static __inline int
316 witness_lock_order_key_equal(const struct witness_lock_order_key *a,
317 const struct witness_lock_order_key *b)
318 {
319
320 return (a->from == b->from && a->to == b->to);
321 }
322
323 static int _isitmyx(struct witness *w1, struct witness *w2, int rmask,
324 const char *fname);
325 #ifdef KDB
326 static void _witness_debugger(int cond, const char *msg);
327 #endif
328 static void adopt(struct witness *parent, struct witness *child);
329 #ifdef BLESSING
330 static int blessed(struct witness *, struct witness *);
331 #endif
332 static void depart(struct witness *w);
333 static struct witness *enroll(const char *description,
334 struct lock_class *lock_class);
335 static struct lock_instance *find_instance(struct lock_list_entry *list,
336 struct lock_object *lock);
337 static int isitmychild(struct witness *parent, struct witness *child);
338 static int isitmydescendant(struct witness *parent, struct witness *child);
339 static void itismychild(struct witness *parent, struct witness *child);
340 static int sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS);
341 static int sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS);
342 static int sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS);
343 static void witness_add_fullgraph(struct sbuf *sb, struct witness *parent);
344 #ifdef DDB
345 static void witness_ddb_compute_levels(void);
346 static void witness_ddb_display(void(*)(const char *fmt, ...));
347 static void witness_ddb_display_descendants(void(*)(const char *fmt, ...),
348 struct witness *, int indent);
349 static void witness_ddb_display_list(void(*prnt)(const char *fmt, ...),
350 struct witness_list *list);
351 static void witness_ddb_level_descendants(struct witness *parent, int l);
352 static void witness_ddb_list(struct thread *td);
353 #endif
354 static void witness_free(struct witness *m);
355 static struct witness *witness_get(void);
356 static uint32_t witness_hash_djb2(const uint8_t *key, uint32_t size);
357 static struct witness *witness_hash_get(const char *key);
358 static void witness_hash_put(struct witness *w);
359 static void witness_init_hash_tables(void);
360 static void witness_increment_graph_generation(void);
361 static void witness_lock_list_free(struct lock_list_entry *lle);
362 static struct lock_list_entry *witness_lock_list_get(void);
363 static int witness_lock_order_add(struct witness *parent,
364 struct witness *child);
365 static int witness_lock_order_check(struct witness *parent,
366 struct witness *child);
367 static struct witness_lock_order_data *witness_lock_order_get(
368 struct witness *parent,
369 struct witness *child);
370 static void witness_list_lock(struct lock_instance *instance);
371 static void witness_setflag(struct lock_object *lock, int flag, int set);
372
373 #ifdef KDB
374 #define witness_debugger(c) _witness_debugger(c, __func__)
375 #else
376 #define witness_debugger(c)
377 #endif
378
379 SYSCTL_NODE(_debug, OID_AUTO, witness, CTLFLAG_RW, NULL, "Witness Locking");
380
381 /*
382 * If set to 0, lock order checking is disabled. If set to -1,
383 * witness is completely disabled. Otherwise witness performs full
384 * lock order checking for all locks. At runtime, lock order checking
385 * may be toggled. However, witness cannot be reenabled once it is
386 * completely disabled.
387 */
388 static int witness_watch = 1;
389 TUNABLE_INT("debug.witness.watch", &witness_watch);
390 SYSCTL_PROC(_debug_witness, OID_AUTO, watch, CTLFLAG_RW | CTLTYPE_INT, NULL, 0,
391 sysctl_debug_witness_watch, "I", "witness is watching lock operations");
392
393 #ifdef KDB
394 /*
395 * When KDB is enabled and witness_kdb is 1, it will cause the system
396 * to drop into kdebug() when:
397 * - a lock hierarchy violation occurs
398 * - locks are held when going to sleep.
399 */
400 #ifdef WITNESS_KDB
401 int witness_kdb = 1;
402 #else
403 int witness_kdb = 0;
404 #endif
405 TUNABLE_INT("debug.witness.kdb", &witness_kdb);
406 SYSCTL_INT(_debug_witness, OID_AUTO, kdb, CTLFLAG_RW, &witness_kdb, 0, "");
407
408 /*
409 * When KDB is enabled and witness_trace is 1, it will cause the system
410 * to print a stack trace:
411 * - a lock hierarchy violation occurs
412 * - locks are held when going to sleep.
413 */
414 int witness_trace = 1;
415 TUNABLE_INT("debug.witness.trace", &witness_trace);
416 SYSCTL_INT(_debug_witness, OID_AUTO, trace, CTLFLAG_RW, &witness_trace, 0, "");
417 #endif /* KDB */
418
419 #ifdef WITNESS_SKIPSPIN
420 int witness_skipspin = 1;
421 #else
422 int witness_skipspin = 0;
423 #endif
424 TUNABLE_INT("debug.witness.skipspin", &witness_skipspin);
425 SYSCTL_INT(_debug_witness, OID_AUTO, skipspin, CTLFLAG_RDTUN, &witness_skipspin,
426 0, "");
427
428 /*
429 * Call this to print out the relations between locks.
430 */
431 SYSCTL_PROC(_debug_witness, OID_AUTO, fullgraph, CTLTYPE_STRING | CTLFLAG_RD,
432 NULL, 0, sysctl_debug_witness_fullgraph, "A", "Show locks relation graphs");
433
434 /*
435 * Call this to print out the witness faulty stacks.
436 */
437 SYSCTL_PROC(_debug_witness, OID_AUTO, badstacks, CTLTYPE_STRING | CTLFLAG_RD,
438 NULL, 0, sysctl_debug_witness_badstacks, "A", "Show bad witness stacks");
439
440 static struct mtx w_mtx;
441
442 /* w_list */
443 static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free);
444 static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all);
445
446 /* w_typelist */
447 static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin);
448 static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep);
449
450 /* lock list */
451 static struct lock_list_entry *w_lock_list_free = NULL;
452 static struct witness_pendhelp pending_locks[WITNESS_PENDLIST];
453 static u_int pending_cnt;
454
455 static int w_free_cnt, w_spin_cnt, w_sleep_cnt;
456 SYSCTL_INT(_debug_witness, OID_AUTO, free_cnt, CTLFLAG_RD, &w_free_cnt, 0, "");
457 SYSCTL_INT(_debug_witness, OID_AUTO, spin_cnt, CTLFLAG_RD, &w_spin_cnt, 0, "");
458 SYSCTL_INT(_debug_witness, OID_AUTO, sleep_cnt, CTLFLAG_RD, &w_sleep_cnt, 0,
459 "");
460
461 static struct witness *w_data;
462 static uint8_t w_rmatrix[WITNESS_COUNT+1][WITNESS_COUNT+1];
463 static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT];
464 static struct witness_hash w_hash; /* The witness hash table. */
465
466 /* The lock order data hash */
467 static struct witness_lock_order_data w_lodata[WITNESS_LO_DATA_COUNT];
468 static struct witness_lock_order_data *w_lofree = NULL;
469 static struct witness_lock_order_hash w_lohash;
470 static int w_max_used_index = 0;
471 static unsigned int w_generation = 0;
472 static const char *w_notrunning = "Witness not running\n";
473 static const char *w_stillcold = "Witness is still cold\n";
474
475
476 static struct witness_order_list_entry order_lists[] = {
477 /*
478 * sx locks
479 */
480 { "proctree", &lock_class_sx },
481 { "allproc", &lock_class_sx },
482 { "allprison", &lock_class_sx },
483 { NULL, NULL },
484 /*
485 * Various mutexes
486 */
487 { "Giant", &lock_class_mtx_sleep },
488 { "pipe mutex", &lock_class_mtx_sleep },
489 { "sigio lock", &lock_class_mtx_sleep },
490 { "process group", &lock_class_mtx_sleep },
491 { "process lock", &lock_class_mtx_sleep },
492 { "session", &lock_class_mtx_sleep },
493 { "uidinfo hash", &lock_class_rw },
494 #ifdef HWPMC_HOOKS
495 { "pmc-sleep", &lock_class_mtx_sleep },
496 #endif
497 { NULL, NULL },
498 /*
499 * Sockets
500 */
501 { "accept", &lock_class_mtx_sleep },
502 { "so_snd", &lock_class_mtx_sleep },
503 { "so_rcv", &lock_class_mtx_sleep },
504 { "sellck", &lock_class_mtx_sleep },
505 { NULL, NULL },
506 /*
507 * Routing
508 */
509 { "so_rcv", &lock_class_mtx_sleep },
510 { "radix node head", &lock_class_rw },
511 { "rtentry", &lock_class_mtx_sleep },
512 { "ifaddr", &lock_class_mtx_sleep },
513 { NULL, NULL },
514 /*
515 * IPv4 multicast:
516 * protocol locks before interface locks, after UDP locks.
517 */
518 { "udpinp", &lock_class_rw },
519 { "in_multi_mtx", &lock_class_mtx_sleep },
520 { "igmp_mtx", &lock_class_mtx_sleep },
521 { "if_addr_mtx", &lock_class_mtx_sleep },
522 { NULL, NULL },
523 /*
524 * IPv6 multicast:
525 * protocol locks before interface locks, after UDP locks.
526 */
527 { "udpinp", &lock_class_rw },
528 { "in6_multi_mtx", &lock_class_mtx_sleep },
529 { "mld_mtx", &lock_class_mtx_sleep },
530 { "if_addr_mtx", &lock_class_mtx_sleep },
531 { NULL, NULL },
532 /*
533 * UNIX Domain Sockets
534 */
535 { "unp_global_rwlock", &lock_class_rw },
536 { "unp_list_lock", &lock_class_mtx_sleep },
537 { "unp", &lock_class_mtx_sleep },
538 { "so_snd", &lock_class_mtx_sleep },
539 { NULL, NULL },
540 /*
541 * UDP/IP
542 */
543 { "udp", &lock_class_rw },
544 { "udpinp", &lock_class_rw },
545 { "so_snd", &lock_class_mtx_sleep },
546 { NULL, NULL },
547 /*
548 * TCP/IP
549 */
550 { "tcp", &lock_class_rw },
551 { "tcpinp", &lock_class_rw },
552 { "so_snd", &lock_class_mtx_sleep },
553 { NULL, NULL },
554 /*
555 * SLIP
556 */
557 { "slip_mtx", &lock_class_mtx_sleep },
558 { "slip sc_mtx", &lock_class_mtx_sleep },
559 { NULL, NULL },
560 /*
561 * netatalk
562 */
563 { "ddp_list_mtx", &lock_class_mtx_sleep },
564 { "ddp_mtx", &lock_class_mtx_sleep },
565 { NULL, NULL },
566 /*
567 * BPF
568 */
569 { "bpf global lock", &lock_class_mtx_sleep },
570 { "bpf interface lock", &lock_class_mtx_sleep },
571 { "bpf cdev lock", &lock_class_mtx_sleep },
572 { NULL, NULL },
573 /*
574 * NFS server
575 */
576 { "nfsd_mtx", &lock_class_mtx_sleep },
577 { "so_snd", &lock_class_mtx_sleep },
578 { NULL, NULL },
579
580 /*
581 * IEEE 802.11
582 */
583 { "802.11 com lock", &lock_class_mtx_sleep},
584 { NULL, NULL },
585 /*
586 * Network drivers
587 */
588 { "network driver", &lock_class_mtx_sleep},
589 { NULL, NULL },
590
591 /*
592 * Netgraph
593 */
594 { "ng_node", &lock_class_mtx_sleep },
595 { "ng_worklist", &lock_class_mtx_sleep },
596 { NULL, NULL },
597 /*
598 * CDEV
599 */
600 { "system map", &lock_class_mtx_sleep },
601 { "vm page queue mutex", &lock_class_mtx_sleep },
602 { "vnode interlock", &lock_class_mtx_sleep },
603 { "cdev", &lock_class_mtx_sleep },
604 { NULL, NULL },
605 /*
606 * kqueue/VFS interaction
607 */
608 { "kqueue", &lock_class_mtx_sleep },
609 { "struct mount mtx", &lock_class_mtx_sleep },
610 { "vnode interlock", &lock_class_mtx_sleep },
611 { NULL, NULL },
612 /*
613 * ZFS locking
614 */
615 { "dn->dn_mtx", &lock_class_sx },
616 { "dr->dt.di.dr_mtx", &lock_class_sx },
617 { "db->db_mtx", &lock_class_sx },
618 { NULL, NULL },
619 /*
620 * spin locks
621 */
622 #ifdef SMP
623 { "ap boot", &lock_class_mtx_spin },
624 #endif
625 { "rm.mutex_mtx", &lock_class_mtx_spin },
626 { "sio", &lock_class_mtx_spin },
627 { "scrlock", &lock_class_mtx_spin },
628 #ifdef __i386__
629 { "cy", &lock_class_mtx_spin },
630 #endif
631 #ifdef __sparc64__
632 { "pcib_mtx", &lock_class_mtx_spin },
633 { "rtc_mtx", &lock_class_mtx_spin },
634 #endif
635 { "scc_hwmtx", &lock_class_mtx_spin },
636 { "uart_hwmtx", &lock_class_mtx_spin },
637 { "fast_taskqueue", &lock_class_mtx_spin },
638 { "intr table", &lock_class_mtx_spin },
639 #ifdef HWPMC_HOOKS
640 { "pmc-per-proc", &lock_class_mtx_spin },
641 #endif
642 { "process slock", &lock_class_mtx_spin },
643 { "sleepq chain", &lock_class_mtx_spin },
644 { "umtx lock", &lock_class_mtx_spin },
645 { "rm_spinlock", &lock_class_mtx_spin },
646 { "turnstile chain", &lock_class_mtx_spin },
647 { "turnstile lock", &lock_class_mtx_spin },
648 { "sched lock", &lock_class_mtx_spin },
649 { "td_contested", &lock_class_mtx_spin },
650 { "callout", &lock_class_mtx_spin },
651 { "entropy harvest mutex", &lock_class_mtx_spin },
652 { "syscons video lock", &lock_class_mtx_spin },
653 { "time lock", &lock_class_mtx_spin },
654 #ifdef SMP
655 { "smp rendezvous", &lock_class_mtx_spin },
656 #endif
657 #ifdef __powerpc__
658 { "tlb0", &lock_class_mtx_spin },
659 #endif
660 /*
661 * leaf locks
662 */
663 { "intrcnt", &lock_class_mtx_spin },
664 { "icu", &lock_class_mtx_spin },
665 #if defined(SMP) && defined(__sparc64__)
666 { "ipi", &lock_class_mtx_spin },
667 #endif
668 #ifdef __i386__
669 { "allpmaps", &lock_class_mtx_spin },
670 { "descriptor tables", &lock_class_mtx_spin },
671 #endif
672 { "clk", &lock_class_mtx_spin },
673 { "cpuset", &lock_class_mtx_spin },
674 { "mprof lock", &lock_class_mtx_spin },
675 { "zombie lock", &lock_class_mtx_spin },
676 { "ALD Queue", &lock_class_mtx_spin },
677 #ifdef __ia64__
678 { "MCA spin lock", &lock_class_mtx_spin },
679 #endif
680 #if defined(__i386__) || defined(__amd64__)
681 { "pcicfg", &lock_class_mtx_spin },
682 { "NDIS thread lock", &lock_class_mtx_spin },
683 #endif
684 { "tw_osl_io_lock", &lock_class_mtx_spin },
685 { "tw_osl_q_lock", &lock_class_mtx_spin },
686 { "tw_cl_io_lock", &lock_class_mtx_spin },
687 { "tw_cl_intr_lock", &lock_class_mtx_spin },
688 { "tw_cl_gen_lock", &lock_class_mtx_spin },
689 #ifdef HWPMC_HOOKS
690 { "pmc-leaf", &lock_class_mtx_spin },
691 #endif
692 { "blocked lock", &lock_class_mtx_spin },
693 { NULL, NULL },
694 { NULL, NULL }
695 };
696
697 #ifdef BLESSING
698 /*
699 * Pairs of locks which have been blessed
700 * Don't complain about order problems with blessed locks
701 */
702 static struct witness_blessed blessed_list[] = {
703 };
704 static int blessed_count =
705 sizeof(blessed_list) / sizeof(struct witness_blessed);
706 #endif
707
708 /*
709 * This global is set to 0 once it becomes safe to use the witness code.
710 */
711 static int witness_cold = 1;
712
713 /*
714 * This global is set to 1 once the static lock orders have been enrolled
715 * so that a warning can be issued for any spin locks enrolled later.
716 */
717 static int witness_spin_warn = 0;
718
719 /*
720 * The WITNESS-enabled diagnostic code. Note that the witness code does
721 * assume that the early boot is single-threaded at least until after this
722 * routine is completed.
723 */
724 static void
725 witness_initialize(void *dummy __unused)
726 {
727 struct lock_object *lock;
728 struct witness_order_list_entry *order;
729 struct witness *w, *w1;
730 int i;
731
732 w_data = malloc(sizeof (struct witness) * WITNESS_COUNT, M_WITNESS,
733 M_NOWAIT | M_ZERO);
734
735 /*
736 * We have to release Giant before initializing its witness
737 * structure so that WITNESS doesn't get confused.
738 */
739 mtx_unlock(&Giant);
740 mtx_assert(&Giant, MA_NOTOWNED);
741
742 CTR1(KTR_WITNESS, "%s: initializing witness", __func__);
743 mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET |
744 MTX_NOWITNESS | MTX_NOPROFILE);
745 for (i = WITNESS_COUNT - 1; i >= 0; i--) {
746 w = &w_data[i];
747 memset(w, 0, sizeof(*w));
748 w_data[i].w_index = i; /* Witness index never changes. */
749 witness_free(w);
750 }
751 KASSERT(STAILQ_FIRST(&w_free)->w_index == 0,
752 ("%s: Invalid list of free witness objects", __func__));
753
754 /* Witness with index 0 is not used to aid in debugging. */
755 STAILQ_REMOVE_HEAD(&w_free, w_list);
756 w_free_cnt--;
757
758 memset(w_rmatrix, 0,
759 (sizeof(**w_rmatrix) * (WITNESS_COUNT+1) * (WITNESS_COUNT+1)));
760
761 for (i = 0; i < LOCK_CHILDCOUNT; i++)
762 witness_lock_list_free(&w_locklistdata[i]);
763 witness_init_hash_tables();
764
765 /* First add in all the specified order lists. */
766 for (order = order_lists; order->w_name != NULL; order++) {
767 w = enroll(order->w_name, order->w_class);
768 if (w == NULL)
769 continue;
770 w->w_file = "order list";
771 for (order++; order->w_name != NULL; order++) {
772 w1 = enroll(order->w_name, order->w_class);
773 if (w1 == NULL)
774 continue;
775 w1->w_file = "order list";
776 itismychild(w, w1);
777 w = w1;
778 }
779 }
780 witness_spin_warn = 1;
781
782 /* Iterate through all locks and add them to witness. */
783 for (i = 0; pending_locks[i].wh_lock != NULL; i++) {
784 lock = pending_locks[i].wh_lock;
785 KASSERT(lock->lo_flags & LO_WITNESS,
786 ("%s: lock %s is on pending list but not LO_WITNESS",
787 __func__, lock->lo_name));
788 lock->lo_witness = enroll(pending_locks[i].wh_type,
789 LOCK_CLASS(lock));
790 }
791
792 /* Mark the witness code as being ready for use. */
793 witness_cold = 0;
794
795 mtx_lock(&Giant);
796 }
797 SYSINIT(witness_init, SI_SUB_WITNESS, SI_ORDER_FIRST, witness_initialize,
798 NULL);
799
800 void
801 witness_init(struct lock_object *lock, const char *type)
802 {
803 struct lock_class *class;
804
805 /* Various sanity checks. */
806 class = LOCK_CLASS(lock);
807 if ((lock->lo_flags & LO_RECURSABLE) != 0 &&
808 (class->lc_flags & LC_RECURSABLE) == 0)
809 panic("%s: lock (%s) %s can not be recursable", __func__,
810 class->lc_name, lock->lo_name);
811 if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
812 (class->lc_flags & LC_SLEEPABLE) == 0)
813 panic("%s: lock (%s) %s can not be sleepable", __func__,
814 class->lc_name, lock->lo_name);
815 if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
816 (class->lc_flags & LC_UPGRADABLE) == 0)
817 panic("%s: lock (%s) %s can not be upgradable", __func__,
818 class->lc_name, lock->lo_name);
819
820 /*
821 * If we shouldn't watch this lock, then just clear lo_witness.
822 * Otherwise, if witness_cold is set, then it is too early to
823 * enroll this lock, so defer it to witness_initialize() by adding
824 * it to the pending_locks list. If it is not too early, then enroll
825 * the lock now.
826 */
827 if (witness_watch < 1 || panicstr != NULL ||
828 (lock->lo_flags & LO_WITNESS) == 0)
829 lock->lo_witness = NULL;
830 else if (witness_cold) {
831 pending_locks[pending_cnt].wh_lock = lock;
832 pending_locks[pending_cnt++].wh_type = type;
833 if (pending_cnt > WITNESS_PENDLIST)
834 panic("%s: pending locks list is too small, bump it\n",
835 __func__);
836 } else
837 lock->lo_witness = enroll(type, class);
838 }
839
840 void
841 witness_destroy(struct lock_object *lock)
842 {
843 struct lock_class *class;
844 struct witness *w;
845
846 class = LOCK_CLASS(lock);
847
848 if (witness_cold)
849 panic("lock (%s) %s destroyed while witness_cold",
850 class->lc_name, lock->lo_name);
851
852 /* XXX: need to verify that no one holds the lock */
853 if ((lock->lo_flags & LO_WITNESS) == 0 || lock->lo_witness == NULL)
854 return;
855 w = lock->lo_witness;
856
857 mtx_lock_spin(&w_mtx);
858 MPASS(w->w_refcount > 0);
859 w->w_refcount--;
860
861 if (w->w_refcount == 0)
862 depart(w);
863 mtx_unlock_spin(&w_mtx);
864 }
865
866 #ifdef DDB
867 static void
868 witness_ddb_compute_levels(void)
869 {
870 struct witness *w;
871
872 /*
873 * First clear all levels.
874 */
875 STAILQ_FOREACH(w, &w_all, w_list)
876 w->w_ddb_level = -1;
877
878 /*
879 * Look for locks with no parents and level all their descendants.
880 */
881 STAILQ_FOREACH(w, &w_all, w_list) {
882
883 /* If the witness has ancestors (is not a root), skip it. */
884 if (w->w_num_ancestors > 0)
885 continue;
886 witness_ddb_level_descendants(w, 0);
887 }
888 }
889
890 static void
891 witness_ddb_level_descendants(struct witness *w, int l)
892 {
893 int i;
894
895 if (w->w_ddb_level >= l)
896 return;
897
898 w->w_ddb_level = l;
899 l++;
900
901 for (i = 1; i <= w_max_used_index; i++) {
902 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
903 witness_ddb_level_descendants(&w_data[i], l);
904 }
905 }
906
907 static void
908 witness_ddb_display_descendants(void(*prnt)(const char *fmt, ...),
909 struct witness *w, int indent)
910 {
911 int i;
912
913 for (i = 0; i < indent; i++)
914 prnt(" ");
915 prnt("%s (type: %s, depth: %d, active refs: %d)",
916 w->w_name, w->w_class->lc_name,
917 w->w_ddb_level, w->w_refcount);
918 if (w->w_displayed) {
919 prnt(" -- (already displayed)\n");
920 return;
921 }
922 w->w_displayed = 1;
923 if (w->w_file != NULL && w->w_line != 0)
924 prnt(" -- last acquired @ %s:%d\n", w->w_file,
925 w->w_line);
926 else
927 prnt(" -- never acquired\n");
928 indent++;
929 WITNESS_INDEX_ASSERT(w->w_index);
930 for (i = 1; i <= w_max_used_index; i++) {
931 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
932 witness_ddb_display_descendants(prnt, &w_data[i],
933 indent);
934 }
935 }
936
937 static void
938 witness_ddb_display_list(void(*prnt)(const char *fmt, ...),
939 struct witness_list *list)
940 {
941 struct witness *w;
942
943 STAILQ_FOREACH(w, list, w_typelist) {
944 if (w->w_file == NULL || w->w_ddb_level > 0)
945 continue;
946
947 /* This lock has no anscestors - display its descendants. */
948 witness_ddb_display_descendants(prnt, w, 0);
949 }
950 }
951
952 static void
953 witness_ddb_display(void(*prnt)(const char *fmt, ...))
954 {
955 struct witness *w;
956
957 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
958 witness_ddb_compute_levels();
959
960 /* Clear all the displayed flags. */
961 STAILQ_FOREACH(w, &w_all, w_list)
962 w->w_displayed = 0;
963
964 /*
965 * First, handle sleep locks which have been acquired at least
966 * once.
967 */
968 prnt("Sleep locks:\n");
969 witness_ddb_display_list(prnt, &w_sleep);
970
971 /*
972 * Now do spin locks which have been acquired at least once.
973 */
974 prnt("\nSpin locks:\n");
975 witness_ddb_display_list(prnt, &w_spin);
976
977 /*
978 * Finally, any locks which have not been acquired yet.
979 */
980 prnt("\nLocks which were never acquired:\n");
981 STAILQ_FOREACH(w, &w_all, w_list) {
982 if (w->w_file != NULL || w->w_refcount == 0)
983 continue;
984 prnt("%s (type: %s, depth: %d)\n", w->w_name,
985 w->w_class->lc_name, w->w_ddb_level);
986 }
987 }
988 #endif /* DDB */
989
990 /* Trim useless garbage from filenames. */
991 static const char *
992 fixup_filename(const char *file)
993 {
994
995 if (file == NULL)
996 return (NULL);
997 while (strncmp(file, "../", 3) == 0)
998 file += 3;
999 return (file);
1000 }
1001
1002 int
1003 witness_defineorder(struct lock_object *lock1, struct lock_object *lock2)
1004 {
1005
1006 if (witness_watch == -1 || panicstr != NULL)
1007 return (0);
1008
1009 /* Require locks that witness knows about. */
1010 if (lock1 == NULL || lock1->lo_witness == NULL || lock2 == NULL ||
1011 lock2->lo_witness == NULL)
1012 return (EINVAL);
1013
1014 mtx_assert(&w_mtx, MA_NOTOWNED);
1015 mtx_lock_spin(&w_mtx);
1016
1017 /*
1018 * If we already have either an explicit or implied lock order that
1019 * is the other way around, then return an error.
1020 */
1021 if (witness_watch &&
1022 isitmydescendant(lock2->lo_witness, lock1->lo_witness)) {
1023 mtx_unlock_spin(&w_mtx);
1024 return (EDOOFUS);
1025 }
1026
1027 /* Try to add the new order. */
1028 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1029 lock2->lo_witness->w_name, lock1->lo_witness->w_name);
1030 itismychild(lock1->lo_witness, lock2->lo_witness);
1031 mtx_unlock_spin(&w_mtx);
1032 return (0);
1033 }
1034
1035 void
1036 witness_checkorder(struct lock_object *lock, int flags, const char *file,
1037 int line, struct lock_object *interlock)
1038 {
1039 struct lock_list_entry *lock_list, *lle;
1040 struct lock_instance *lock1, *lock2, *plock;
1041 struct lock_class *class;
1042 struct witness *w, *w1;
1043 struct thread *td;
1044 int i, j;
1045
1046 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL ||
1047 panicstr != NULL)
1048 return;
1049
1050 w = lock->lo_witness;
1051 class = LOCK_CLASS(lock);
1052 td = curthread;
1053 file = fixup_filename(file);
1054
1055 if (class->lc_flags & LC_SLEEPLOCK) {
1056
1057 /*
1058 * Since spin locks include a critical section, this check
1059 * implicitly enforces a lock order of all sleep locks before
1060 * all spin locks.
1061 */
1062 if (td->td_critnest != 0 && !kdb_active)
1063 panic("blockable sleep lock (%s) %s @ %s:%d",
1064 class->lc_name, lock->lo_name, file, line);
1065
1066 /*
1067 * If this is the first lock acquired then just return as
1068 * no order checking is needed.
1069 */
1070 lock_list = td->td_sleeplocks;
1071 if (lock_list == NULL || lock_list->ll_count == 0)
1072 return;
1073 } else {
1074
1075 /*
1076 * If this is the first lock, just return as no order
1077 * checking is needed. Avoid problems with thread
1078 * migration pinning the thread while checking if
1079 * spinlocks are held. If at least one spinlock is held
1080 * the thread is in a safe path and it is allowed to
1081 * unpin it.
1082 */
1083 sched_pin();
1084 lock_list = PCPU_GET(spinlocks);
1085 if (lock_list == NULL || lock_list->ll_count == 0) {
1086 sched_unpin();
1087 return;
1088 }
1089 sched_unpin();
1090 }
1091
1092 /*
1093 * Check to see if we are recursing on a lock we already own. If
1094 * so, make sure that we don't mismatch exclusive and shared lock
1095 * acquires.
1096 */
1097 lock1 = find_instance(lock_list, lock);
1098 if (lock1 != NULL) {
1099 if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
1100 (flags & LOP_EXCLUSIVE) == 0) {
1101 printf("shared lock of (%s) %s @ %s:%d\n",
1102 class->lc_name, lock->lo_name, file, line);
1103 printf("while exclusively locked from %s:%d\n",
1104 lock1->li_file, lock1->li_line);
1105 panic("share->excl");
1106 }
1107 if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
1108 (flags & LOP_EXCLUSIVE) != 0) {
1109 printf("exclusive lock of (%s) %s @ %s:%d\n",
1110 class->lc_name, lock->lo_name, file, line);
1111 printf("while share locked from %s:%d\n",
1112 lock1->li_file, lock1->li_line);
1113 panic("excl->share");
1114 }
1115 return;
1116 }
1117
1118 /*
1119 * Find the previously acquired lock, but ignore interlocks.
1120 */
1121 plock = &lock_list->ll_children[lock_list->ll_count - 1];
1122 if (interlock != NULL && plock->li_lock == interlock) {
1123 if (lock_list->ll_count > 1)
1124 plock =
1125 &lock_list->ll_children[lock_list->ll_count - 2];
1126 else {
1127 lle = lock_list->ll_next;
1128
1129 /*
1130 * The interlock is the only lock we hold, so
1131 * simply return.
1132 */
1133 if (lle == NULL)
1134 return;
1135 plock = &lle->ll_children[lle->ll_count - 1];
1136 }
1137 }
1138
1139 /*
1140 * Try to perform most checks without a lock. If this succeeds we
1141 * can skip acquiring the lock and return success.
1142 */
1143 w1 = plock->li_lock->lo_witness;
1144 if (witness_lock_order_check(w1, w))
1145 return;
1146
1147 /*
1148 * Check for duplicate locks of the same type. Note that we only
1149 * have to check for this on the last lock we just acquired. Any
1150 * other cases will be caught as lock order violations.
1151 */
1152 mtx_lock_spin(&w_mtx);
1153 witness_lock_order_add(w1, w);
1154 if (w1 == w) {
1155 i = w->w_index;
1156 if (!(lock->lo_flags & LO_DUPOK) && !(flags & LOP_DUPOK) &&
1157 !(w_rmatrix[i][i] & WITNESS_REVERSAL)) {
1158 w_rmatrix[i][i] |= WITNESS_REVERSAL;
1159 w->w_reversed = 1;
1160 mtx_unlock_spin(&w_mtx);
1161 printf(
1162 "acquiring duplicate lock of same type: \"%s\"\n",
1163 w->w_name);
1164 printf(" 1st %s @ %s:%d\n", plock->li_lock->lo_name,
1165 plock->li_file, plock->li_line);
1166 printf(" 2nd %s @ %s:%d\n", lock->lo_name, file, line);
1167 witness_debugger(1);
1168 } else
1169 mtx_unlock_spin(&w_mtx);
1170 return;
1171 }
1172 mtx_assert(&w_mtx, MA_OWNED);
1173
1174 /*
1175 * If we know that the the lock we are acquiring comes after
1176 * the lock we most recently acquired in the lock order tree,
1177 * then there is no need for any further checks.
1178 */
1179 if (isitmychild(w1, w))
1180 goto out;
1181
1182 for (j = 0, lle = lock_list; lle != NULL; lle = lle->ll_next) {
1183 for (i = lle->ll_count - 1; i >= 0; i--, j++) {
1184
1185 MPASS(j < WITNESS_COUNT);
1186 lock1 = &lle->ll_children[i];
1187
1188 /*
1189 * Ignore the interlock the first time we see it.
1190 */
1191 if (interlock != NULL && interlock == lock1->li_lock) {
1192 interlock = NULL;
1193 continue;
1194 }
1195
1196 /*
1197 * If this lock doesn't undergo witness checking,
1198 * then skip it.
1199 */
1200 w1 = lock1->li_lock->lo_witness;
1201 if (w1 == NULL) {
1202 KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
1203 ("lock missing witness structure"));
1204 continue;
1205 }
1206
1207 /*
1208 * If we are locking Giant and this is a sleepable
1209 * lock, then skip it.
1210 */
1211 if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0 &&
1212 lock == &Giant.lock_object)
1213 continue;
1214
1215 /*
1216 * If we are locking a sleepable lock and this lock
1217 * is Giant, then skip it.
1218 */
1219 if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1220 lock1->li_lock == &Giant.lock_object)
1221 continue;
1222
1223 /*
1224 * If we are locking a sleepable lock and this lock
1225 * isn't sleepable, we want to treat it as a lock
1226 * order violation to enfore a general lock order of
1227 * sleepable locks before non-sleepable locks.
1228 */
1229 if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1230 (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
1231 goto reversal;
1232
1233 /*
1234 * If we are locking Giant and this is a non-sleepable
1235 * lock, then treat it as a reversal.
1236 */
1237 if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0 &&
1238 lock == &Giant.lock_object)
1239 goto reversal;
1240
1241 /*
1242 * Check the lock order hierarchy for a reveresal.
1243 */
1244 if (!isitmydescendant(w, w1))
1245 continue;
1246 reversal:
1247
1248 /*
1249 * We have a lock order violation, check to see if it
1250 * is allowed or has already been yelled about.
1251 */
1252 #ifdef BLESSING
1253
1254 /*
1255 * If the lock order is blessed, just bail. We don't
1256 * look for other lock order violations though, which
1257 * may be a bug.
1258 */
1259 if (blessed(w, w1))
1260 goto out;
1261 #endif
1262
1263 /* Bail if this violation is known */
1264 if (w_rmatrix[w1->w_index][w->w_index] & WITNESS_REVERSAL)
1265 goto out;
1266
1267 /* Record this as a violation */
1268 w_rmatrix[w1->w_index][w->w_index] |= WITNESS_REVERSAL;
1269 w_rmatrix[w->w_index][w1->w_index] |= WITNESS_REVERSAL;
1270 w->w_reversed = w1->w_reversed = 1;
1271 witness_increment_graph_generation();
1272 mtx_unlock_spin(&w_mtx);
1273
1274 /*
1275 * Ok, yell about it.
1276 */
1277 if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1278 (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
1279 printf(
1280 "lock order reversal: (sleepable after non-sleepable)\n");
1281 else if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0
1282 && lock == &Giant.lock_object)
1283 printf(
1284 "lock order reversal: (Giant after non-sleepable)\n");
1285 else
1286 printf("lock order reversal:\n");
1287
1288 /*
1289 * Try to locate an earlier lock with
1290 * witness w in our list.
1291 */
1292 do {
1293 lock2 = &lle->ll_children[i];
1294 MPASS(lock2->li_lock != NULL);
1295 if (lock2->li_lock->lo_witness == w)
1296 break;
1297 if (i == 0 && lle->ll_next != NULL) {
1298 lle = lle->ll_next;
1299 i = lle->ll_count - 1;
1300 MPASS(i >= 0 && i < LOCK_NCHILDREN);
1301 } else
1302 i--;
1303 } while (i >= 0);
1304 if (i < 0) {
1305 printf(" 1st %p %s (%s) @ %s:%d\n",
1306 lock1->li_lock, lock1->li_lock->lo_name,
1307 w1->w_name, lock1->li_file, lock1->li_line);
1308 printf(" 2nd %p %s (%s) @ %s:%d\n", lock,
1309 lock->lo_name, w->w_name, file, line);
1310 } else {
1311 printf(" 1st %p %s (%s) @ %s:%d\n",
1312 lock2->li_lock, lock2->li_lock->lo_name,
1313 lock2->li_lock->lo_witness->w_name,
1314 lock2->li_file, lock2->li_line);
1315 printf(" 2nd %p %s (%s) @ %s:%d\n",
1316 lock1->li_lock, lock1->li_lock->lo_name,
1317 w1->w_name, lock1->li_file, lock1->li_line);
1318 printf(" 3rd %p %s (%s) @ %s:%d\n", lock,
1319 lock->lo_name, w->w_name, file, line);
1320 }
1321 witness_debugger(1);
1322 return;
1323 }
1324 }
1325
1326 /*
1327 * If requested, build a new lock order. However, don't build a new
1328 * relationship between a sleepable lock and Giant if it is in the
1329 * wrong direction. The correct lock order is that sleepable locks
1330 * always come before Giant.
1331 */
1332 if (flags & LOP_NEWORDER &&
1333 !(plock->li_lock == &Giant.lock_object &&
1334 (lock->lo_flags & LO_SLEEPABLE) != 0)) {
1335 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1336 w->w_name, plock->li_lock->lo_witness->w_name);
1337 itismychild(plock->li_lock->lo_witness, w);
1338 }
1339 out:
1340 mtx_unlock_spin(&w_mtx);
1341 }
1342
1343 void
1344 witness_lock(struct lock_object *lock, int flags, const char *file, int line)
1345 {
1346 struct lock_list_entry **lock_list, *lle;
1347 struct lock_instance *instance;
1348 struct witness *w;
1349 struct thread *td;
1350
1351 if (witness_cold || witness_watch == -1 || lock->lo_witness == NULL ||
1352 panicstr != NULL)
1353 return;
1354 w = lock->lo_witness;
1355 td = curthread;
1356 file = fixup_filename(file);
1357
1358 /* Determine lock list for this lock. */
1359 if (LOCK_CLASS(lock)->lc_flags & LC_SLEEPLOCK)
1360 lock_list = &td->td_sleeplocks;
1361 else
1362 lock_list = PCPU_PTR(spinlocks);
1363
1364 /* Check to see if we are recursing on a lock we already own. */
1365 instance = find_instance(*lock_list, lock);
1366 if (instance != NULL) {
1367 instance->li_flags++;
1368 CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
1369 td->td_proc->p_pid, lock->lo_name,
1370 instance->li_flags & LI_RECURSEMASK);
1371 instance->li_file = file;
1372 instance->li_line = line;
1373 return;
1374 }
1375
1376 /* Update per-witness last file and line acquire. */
1377 w->w_file = file;
1378 w->w_line = line;
1379
1380 /* Find the next open lock instance in the list and fill it. */
1381 lle = *lock_list;
1382 if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
1383 lle = witness_lock_list_get();
1384 if (lle == NULL)
1385 return;
1386 lle->ll_next = *lock_list;
1387 CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
1388 td->td_proc->p_pid, lle);
1389 *lock_list = lle;
1390 }
1391 instance = &lle->ll_children[lle->ll_count++];
1392 instance->li_lock = lock;
1393 instance->li_line = line;
1394 instance->li_file = file;
1395 if ((flags & LOP_EXCLUSIVE) != 0)
1396 instance->li_flags = LI_EXCLUSIVE;
1397 else
1398 instance->li_flags = 0;
1399 CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
1400 td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
1401 }
1402
1403 void
1404 witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
1405 {
1406 struct lock_instance *instance;
1407 struct lock_class *class;
1408
1409 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1410 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
1411 return;
1412 class = LOCK_CLASS(lock);
1413 file = fixup_filename(file);
1414 if (witness_watch) {
1415 if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1416 panic("upgrade of non-upgradable lock (%s) %s @ %s:%d",
1417 class->lc_name, lock->lo_name, file, line);
1418 if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1419 panic("upgrade of non-sleep lock (%s) %s @ %s:%d",
1420 class->lc_name, lock->lo_name, file, line);
1421 }
1422 instance = find_instance(curthread->td_sleeplocks, lock);
1423 if (instance == NULL)
1424 panic("upgrade of unlocked lock (%s) %s @ %s:%d",
1425 class->lc_name, lock->lo_name, file, line);
1426 if (witness_watch) {
1427 if ((instance->li_flags & LI_EXCLUSIVE) != 0)
1428 panic("upgrade of exclusive lock (%s) %s @ %s:%d",
1429 class->lc_name, lock->lo_name, file, line);
1430 if ((instance->li_flags & LI_RECURSEMASK) != 0)
1431 panic("upgrade of recursed lock (%s) %s r=%d @ %s:%d",
1432 class->lc_name, lock->lo_name,
1433 instance->li_flags & LI_RECURSEMASK, file, line);
1434 }
1435 instance->li_flags |= LI_EXCLUSIVE;
1436 }
1437
1438 void
1439 witness_downgrade(struct lock_object *lock, int flags, const char *file,
1440 int line)
1441 {
1442 struct lock_instance *instance;
1443 struct lock_class *class;
1444
1445 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1446 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
1447 return;
1448 class = LOCK_CLASS(lock);
1449 file = fixup_filename(file);
1450 if (witness_watch) {
1451 if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1452 panic("downgrade of non-upgradable lock (%s) %s @ %s:%d",
1453 class->lc_name, lock->lo_name, file, line);
1454 if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1455 panic("downgrade of non-sleep lock (%s) %s @ %s:%d",
1456 class->lc_name, lock->lo_name, file, line);
1457 }
1458 instance = find_instance(curthread->td_sleeplocks, lock);
1459 if (instance == NULL)
1460 panic("downgrade of unlocked lock (%s) %s @ %s:%d",
1461 class->lc_name, lock->lo_name, file, line);
1462 if (witness_watch) {
1463 if ((instance->li_flags & LI_EXCLUSIVE) == 0)
1464 panic("downgrade of shared lock (%s) %s @ %s:%d",
1465 class->lc_name, lock->lo_name, file, line);
1466 if ((instance->li_flags & LI_RECURSEMASK) != 0)
1467 panic("downgrade of recursed lock (%s) %s r=%d @ %s:%d",
1468 class->lc_name, lock->lo_name,
1469 instance->li_flags & LI_RECURSEMASK, file, line);
1470 }
1471 instance->li_flags &= ~LI_EXCLUSIVE;
1472 }
1473
1474 void
1475 witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
1476 {
1477 struct lock_list_entry **lock_list, *lle;
1478 struct lock_instance *instance;
1479 struct lock_class *class;
1480 struct thread *td;
1481 register_t s;
1482 int i, j;
1483
1484 if (witness_cold || lock->lo_witness == NULL || panicstr != NULL)
1485 return;
1486 td = curthread;
1487 class = LOCK_CLASS(lock);
1488 file = fixup_filename(file);
1489
1490 /* Find lock instance associated with this lock. */
1491 if (class->lc_flags & LC_SLEEPLOCK)
1492 lock_list = &td->td_sleeplocks;
1493 else
1494 lock_list = PCPU_PTR(spinlocks);
1495 lle = *lock_list;
1496 for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
1497 for (i = 0; i < (*lock_list)->ll_count; i++) {
1498 instance = &(*lock_list)->ll_children[i];
1499 if (instance->li_lock == lock)
1500 goto found;
1501 }
1502
1503 /*
1504 * When disabling WITNESS through witness_watch we could end up in
1505 * having registered locks in the td_sleeplocks queue.
1506 * We have to make sure we flush these queues, so just search for
1507 * eventual register locks and remove them.
1508 */
1509 if (witness_watch > 0)
1510 panic("lock (%s) %s not locked @ %s:%d", class->lc_name,
1511 lock->lo_name, file, line);
1512 else
1513 return;
1514 found:
1515
1516 /* First, check for shared/exclusive mismatches. */
1517 if ((instance->li_flags & LI_EXCLUSIVE) != 0 && witness_watch > 0 &&
1518 (flags & LOP_EXCLUSIVE) == 0) {
1519 printf("shared unlock of (%s) %s @ %s:%d\n", class->lc_name,
1520 lock->lo_name, file, line);
1521 printf("while exclusively locked from %s:%d\n",
1522 instance->li_file, instance->li_line);
1523 panic("excl->ushare");
1524 }
1525 if ((instance->li_flags & LI_EXCLUSIVE) == 0 && witness_watch > 0 &&
1526 (flags & LOP_EXCLUSIVE) != 0) {
1527 printf("exclusive unlock of (%s) %s @ %s:%d\n", class->lc_name,
1528 lock->lo_name, file, line);
1529 printf("while share locked from %s:%d\n", instance->li_file,
1530 instance->li_line);
1531 panic("share->uexcl");
1532 }
1533 /* If we are recursed, unrecurse. */
1534 if ((instance->li_flags & LI_RECURSEMASK) > 0) {
1535 CTR4(KTR_WITNESS, "%s: pid %d unrecursed on %s r=%d", __func__,
1536 td->td_proc->p_pid, instance->li_lock->lo_name,
1537 instance->li_flags);
1538 instance->li_flags--;
1539 return;
1540 }
1541 /* The lock is now being dropped, check for NORELEASE flag */
1542 if ((instance->li_flags & LI_NORELEASE) != 0 && witness_watch > 0) {
1543 printf("forbidden unlock of (%s) %s @ %s:%d\n", class->lc_name,
1544 lock->lo_name, file, line);
1545 panic("lock marked norelease");
1546 }
1547
1548 /* Otherwise, remove this item from the list. */
1549 s = intr_disable();
1550 CTR4(KTR_WITNESS, "%s: pid %d removed %s from lle[%d]", __func__,
1551 td->td_proc->p_pid, instance->li_lock->lo_name,
1552 (*lock_list)->ll_count - 1);
1553 for (j = i; j < (*lock_list)->ll_count - 1; j++)
1554 (*lock_list)->ll_children[j] =
1555 (*lock_list)->ll_children[j + 1];
1556 (*lock_list)->ll_count--;
1557 intr_restore(s);
1558
1559 /*
1560 * In order to reduce contention on w_mtx, we want to keep always an
1561 * head object into lists so that frequent allocation from the
1562 * free witness pool (and subsequent locking) is avoided.
1563 * In order to maintain the current code simple, when the head
1564 * object is totally unloaded it means also that we do not have
1565 * further objects in the list, so the list ownership needs to be
1566 * hand over to another object if the current head needs to be freed.
1567 */
1568 if ((*lock_list)->ll_count == 0) {
1569 if (*lock_list == lle) {
1570 if (lle->ll_next == NULL)
1571 return;
1572 } else
1573 lle = *lock_list;
1574 *lock_list = lle->ll_next;
1575 CTR3(KTR_WITNESS, "%s: pid %d removed lle %p", __func__,
1576 td->td_proc->p_pid, lle);
1577 witness_lock_list_free(lle);
1578 }
1579 }
1580
1581 void
1582 witness_thread_exit(struct thread *td)
1583 {
1584 struct lock_list_entry *lle;
1585 int i, n;
1586
1587 lle = td->td_sleeplocks;
1588 if (lle == NULL || panicstr != NULL)
1589 return;
1590 if (lle->ll_count != 0) {
1591 for (n = 0; lle != NULL; lle = lle->ll_next)
1592 for (i = lle->ll_count - 1; i >= 0; i--) {
1593 if (n == 0)
1594 printf("Thread %p exiting with the following locks held:\n",
1595 td);
1596 n++;
1597 witness_list_lock(&lle->ll_children[i]);
1598
1599 }
1600 panic("Thread %p cannot exit while holding sleeplocks\n", td);
1601 }
1602 witness_lock_list_free(lle);
1603 }
1604
1605 /*
1606 * Warn if any locks other than 'lock' are held. Flags can be passed in to
1607 * exempt Giant and sleepable locks from the checks as well. If any
1608 * non-exempt locks are held, then a supplied message is printed to the
1609 * console along with a list of the offending locks. If indicated in the
1610 * flags then a failure results in a panic as well.
1611 */
1612 int
1613 witness_warn(int flags, struct lock_object *lock, const char *fmt, ...)
1614 {
1615 struct lock_list_entry *lock_list, *lle;
1616 struct lock_instance *lock1;
1617 struct thread *td;
1618 va_list ap;
1619 int i, n;
1620
1621 if (witness_cold || witness_watch < 1 || panicstr != NULL)
1622 return (0);
1623 n = 0;
1624 td = curthread;
1625 for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next)
1626 for (i = lle->ll_count - 1; i >= 0; i--) {
1627 lock1 = &lle->ll_children[i];
1628 if (lock1->li_lock == lock)
1629 continue;
1630 if (flags & WARN_GIANTOK &&
1631 lock1->li_lock == &Giant.lock_object)
1632 continue;
1633 if (flags & WARN_SLEEPOK &&
1634 (lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0)
1635 continue;
1636 if (n == 0) {
1637 va_start(ap, fmt);
1638 vprintf(fmt, ap);
1639 va_end(ap);
1640 printf(" with the following");
1641 if (flags & WARN_SLEEPOK)
1642 printf(" non-sleepable");
1643 printf(" locks held:\n");
1644 }
1645 n++;
1646 witness_list_lock(lock1);
1647 }
1648
1649 /*
1650 * Pin the thread in order to avoid problems with thread migration.
1651 * Once that all verifies are passed about spinlocks ownership,
1652 * the thread is in a safe path and it can be unpinned.
1653 */
1654 sched_pin();
1655 lock_list = PCPU_GET(spinlocks);
1656 if (lock_list != NULL && lock_list->ll_count != 0) {
1657 sched_unpin();
1658
1659 /*
1660 * We should only have one spinlock and as long as
1661 * the flags cannot match for this locks class,
1662 * check if the first spinlock is the one curthread
1663 * should hold.
1664 */
1665 lock1 = &lock_list->ll_children[lock_list->ll_count - 1];
1666 if (lock_list->ll_count == 1 && lock_list->ll_next == NULL &&
1667 lock1->li_lock == lock && n == 0)
1668 return (0);
1669
1670 va_start(ap, fmt);
1671 vprintf(fmt, ap);
1672 va_end(ap);
1673 printf(" with the following");
1674 if (flags & WARN_SLEEPOK)
1675 printf(" non-sleepable");
1676 printf(" locks held:\n");
1677 n += witness_list_locks(&lock_list);
1678 } else
1679 sched_unpin();
1680 if (flags & WARN_PANIC && n)
1681 panic("%s", __func__);
1682 else
1683 witness_debugger(n);
1684 return (n);
1685 }
1686
1687 const char *
1688 witness_file(struct lock_object *lock)
1689 {
1690 struct witness *w;
1691
1692 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1693 return ("?");
1694 w = lock->lo_witness;
1695 return (w->w_file);
1696 }
1697
1698 int
1699 witness_line(struct lock_object *lock)
1700 {
1701 struct witness *w;
1702
1703 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1704 return (0);
1705 w = lock->lo_witness;
1706 return (w->w_line);
1707 }
1708
1709 static struct witness *
1710 enroll(const char *description, struct lock_class *lock_class)
1711 {
1712 struct witness *w;
1713 struct witness_list *typelist;
1714
1715 MPASS(description != NULL);
1716
1717 if (witness_watch == -1 || panicstr != NULL)
1718 return (NULL);
1719 if ((lock_class->lc_flags & LC_SPINLOCK)) {
1720 if (witness_skipspin)
1721 return (NULL);
1722 else
1723 typelist = &w_spin;
1724 } else if ((lock_class->lc_flags & LC_SLEEPLOCK))
1725 typelist = &w_sleep;
1726 else
1727 panic("lock class %s is not sleep or spin",
1728 lock_class->lc_name);
1729
1730 mtx_lock_spin(&w_mtx);
1731 w = witness_hash_get(description);
1732 if (w)
1733 goto found;
1734 if ((w = witness_get()) == NULL)
1735 return (NULL);
1736 MPASS(strlen(description) < MAX_W_NAME);
1737 strcpy(w->w_name, description);
1738 w->w_class = lock_class;
1739 w->w_refcount = 1;
1740 STAILQ_INSERT_HEAD(&w_all, w, w_list);
1741 if (lock_class->lc_flags & LC_SPINLOCK) {
1742 STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
1743 w_spin_cnt++;
1744 } else if (lock_class->lc_flags & LC_SLEEPLOCK) {
1745 STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
1746 w_sleep_cnt++;
1747 }
1748
1749 /* Insert new witness into the hash */
1750 witness_hash_put(w);
1751 witness_increment_graph_generation();
1752 mtx_unlock_spin(&w_mtx);
1753 return (w);
1754 found:
1755 w->w_refcount++;
1756 mtx_unlock_spin(&w_mtx);
1757 if (lock_class != w->w_class)
1758 panic(
1759 "lock (%s) %s does not match earlier (%s) lock",
1760 description, lock_class->lc_name,
1761 w->w_class->lc_name);
1762 return (w);
1763 }
1764
1765 static void
1766 depart(struct witness *w)
1767 {
1768 struct witness_list *list;
1769
1770 MPASS(w->w_refcount == 0);
1771 if (w->w_class->lc_flags & LC_SLEEPLOCK) {
1772 list = &w_sleep;
1773 w_sleep_cnt--;
1774 } else {
1775 list = &w_spin;
1776 w_spin_cnt--;
1777 }
1778 /*
1779 * Set file to NULL as it may point into a loadable module.
1780 */
1781 w->w_file = NULL;
1782 w->w_line = 0;
1783 witness_increment_graph_generation();
1784 }
1785
1786
1787 static void
1788 adopt(struct witness *parent, struct witness *child)
1789 {
1790 int pi, ci, i, j;
1791
1792 if (witness_cold == 0)
1793 mtx_assert(&w_mtx, MA_OWNED);
1794
1795 /* If the relationship is already known, there's no work to be done. */
1796 if (isitmychild(parent, child))
1797 return;
1798
1799 /* When the structure of the graph changes, bump up the generation. */
1800 witness_increment_graph_generation();
1801
1802 /*
1803 * The hard part ... create the direct relationship, then propagate all
1804 * indirect relationships.
1805 */
1806 pi = parent->w_index;
1807 ci = child->w_index;
1808 WITNESS_INDEX_ASSERT(pi);
1809 WITNESS_INDEX_ASSERT(ci);
1810 MPASS(pi != ci);
1811 w_rmatrix[pi][ci] |= WITNESS_PARENT;
1812 w_rmatrix[ci][pi] |= WITNESS_CHILD;
1813
1814 /*
1815 * If parent was not already an ancestor of child,
1816 * then we increment the descendant and ancestor counters.
1817 */
1818 if ((w_rmatrix[pi][ci] & WITNESS_ANCESTOR) == 0) {
1819 parent->w_num_descendants++;
1820 child->w_num_ancestors++;
1821 }
1822
1823 /*
1824 * Find each ancestor of 'pi'. Note that 'pi' itself is counted as
1825 * an ancestor of 'pi' during this loop.
1826 */
1827 for (i = 1; i <= w_max_used_index; i++) {
1828 if ((w_rmatrix[i][pi] & WITNESS_ANCESTOR_MASK) == 0 &&
1829 (i != pi))
1830 continue;
1831
1832 /* Find each descendant of 'i' and mark it as a descendant. */
1833 for (j = 1; j <= w_max_used_index; j++) {
1834
1835 /*
1836 * Skip children that are already marked as
1837 * descendants of 'i'.
1838 */
1839 if (w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK)
1840 continue;
1841
1842 /*
1843 * We are only interested in descendants of 'ci'. Note
1844 * that 'ci' itself is counted as a descendant of 'ci'.
1845 */
1846 if ((w_rmatrix[ci][j] & WITNESS_ANCESTOR_MASK) == 0 &&
1847 (j != ci))
1848 continue;
1849 w_rmatrix[i][j] |= WITNESS_ANCESTOR;
1850 w_rmatrix[j][i] |= WITNESS_DESCENDANT;
1851 w_data[i].w_num_descendants++;
1852 w_data[j].w_num_ancestors++;
1853
1854 /*
1855 * Make sure we aren't marking a node as both an
1856 * ancestor and descendant. We should have caught
1857 * this as a lock order reversal earlier.
1858 */
1859 if ((w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK) &&
1860 (w_rmatrix[i][j] & WITNESS_DESCENDANT_MASK)) {
1861 printf("witness rmatrix paradox! [%d][%d]=%d "
1862 "both ancestor and descendant\n",
1863 i, j, w_rmatrix[i][j]);
1864 kdb_backtrace();
1865 printf("Witness disabled.\n");
1866 witness_watch = -1;
1867 }
1868 if ((w_rmatrix[j][i] & WITNESS_ANCESTOR_MASK) &&
1869 (w_rmatrix[j][i] & WITNESS_DESCENDANT_MASK)) {
1870 printf("witness rmatrix paradox! [%d][%d]=%d "
1871 "both ancestor and descendant\n",
1872 j, i, w_rmatrix[j][i]);
1873 kdb_backtrace();
1874 printf("Witness disabled.\n");
1875 witness_watch = -1;
1876 }
1877 }
1878 }
1879 }
1880
1881 static void
1882 itismychild(struct witness *parent, struct witness *child)
1883 {
1884
1885 MPASS(child != NULL && parent != NULL);
1886 if (witness_cold == 0)
1887 mtx_assert(&w_mtx, MA_OWNED);
1888
1889 if (!witness_lock_type_equal(parent, child)) {
1890 if (witness_cold == 0)
1891 mtx_unlock_spin(&w_mtx);
1892 panic("%s: parent \"%s\" (%s) and child \"%s\" (%s) are not "
1893 "the same lock type", __func__, parent->w_name,
1894 parent->w_class->lc_name, child->w_name,
1895 child->w_class->lc_name);
1896 }
1897 adopt(parent, child);
1898 }
1899
1900 /*
1901 * Generic code for the isitmy*() functions. The rmask parameter is the
1902 * expected relationship of w1 to w2.
1903 */
1904 static int
1905 _isitmyx(struct witness *w1, struct witness *w2, int rmask, const char *fname)
1906 {
1907 unsigned char r1, r2;
1908 int i1, i2;
1909
1910 i1 = w1->w_index;
1911 i2 = w2->w_index;
1912 WITNESS_INDEX_ASSERT(i1);
1913 WITNESS_INDEX_ASSERT(i2);
1914 r1 = w_rmatrix[i1][i2] & WITNESS_RELATED_MASK;
1915 r2 = w_rmatrix[i2][i1] & WITNESS_RELATED_MASK;
1916
1917 /* The flags on one better be the inverse of the flags on the other */
1918 if (!((WITNESS_ATOD(r1) == r2 && WITNESS_DTOA(r2) == r1) ||
1919 (WITNESS_DTOA(r1) == r2 && WITNESS_ATOD(r2) == r1))) {
1920 printf("%s: rmatrix mismatch between %s (index %d) and %s "
1921 "(index %d): w_rmatrix[%d][%d] == %hhx but "
1922 "w_rmatrix[%d][%d] == %hhx\n",
1923 fname, w1->w_name, i1, w2->w_name, i2, i1, i2, r1,
1924 i2, i1, r2);
1925 kdb_backtrace();
1926 printf("Witness disabled.\n");
1927 witness_watch = -1;
1928 }
1929 return (r1 & rmask);
1930 }
1931
1932 /*
1933 * Checks if @child is a direct child of @parent.
1934 */
1935 static int
1936 isitmychild(struct witness *parent, struct witness *child)
1937 {
1938
1939 return (_isitmyx(parent, child, WITNESS_PARENT, __func__));
1940 }
1941
1942 /*
1943 * Checks if @descendant is a direct or inderect descendant of @ancestor.
1944 */
1945 static int
1946 isitmydescendant(struct witness *ancestor, struct witness *descendant)
1947 {
1948
1949 return (_isitmyx(ancestor, descendant, WITNESS_ANCESTOR_MASK,
1950 __func__));
1951 }
1952
1953 #ifdef BLESSING
1954 static int
1955 blessed(struct witness *w1, struct witness *w2)
1956 {
1957 int i;
1958 struct witness_blessed *b;
1959
1960 for (i = 0; i < blessed_count; i++) {
1961 b = &blessed_list[i];
1962 if (strcmp(w1->w_name, b->b_lock1) == 0) {
1963 if (strcmp(w2->w_name, b->b_lock2) == 0)
1964 return (1);
1965 continue;
1966 }
1967 if (strcmp(w1->w_name, b->b_lock2) == 0)
1968 if (strcmp(w2->w_name, b->b_lock1) == 0)
1969 return (1);
1970 }
1971 return (0);
1972 }
1973 #endif
1974
1975 static struct witness *
1976 witness_get(void)
1977 {
1978 struct witness *w;
1979 int index;
1980
1981 if (witness_cold == 0)
1982 mtx_assert(&w_mtx, MA_OWNED);
1983
1984 if (witness_watch == -1) {
1985 mtx_unlock_spin(&w_mtx);
1986 return (NULL);
1987 }
1988 if (STAILQ_EMPTY(&w_free)) {
1989 witness_watch = -1;
1990 mtx_unlock_spin(&w_mtx);
1991 printf("WITNESS: unable to allocate a new witness object\n");
1992 return (NULL);
1993 }
1994 w = STAILQ_FIRST(&w_free);
1995 STAILQ_REMOVE_HEAD(&w_free, w_list);
1996 w_free_cnt--;
1997 index = w->w_index;
1998 MPASS(index > 0 && index == w_max_used_index+1 &&
1999 index < WITNESS_COUNT);
2000 bzero(w, sizeof(*w));
2001 w->w_index = index;
2002 if (index > w_max_used_index)
2003 w_max_used_index = index;
2004 return (w);
2005 }
2006
2007 static void
2008 witness_free(struct witness *w)
2009 {
2010
2011 STAILQ_INSERT_HEAD(&w_free, w, w_list);
2012 w_free_cnt++;
2013 }
2014
2015 static struct lock_list_entry *
2016 witness_lock_list_get(void)
2017 {
2018 struct lock_list_entry *lle;
2019
2020 if (witness_watch == -1)
2021 return (NULL);
2022 mtx_lock_spin(&w_mtx);
2023 lle = w_lock_list_free;
2024 if (lle == NULL) {
2025 witness_watch = -1;
2026 mtx_unlock_spin(&w_mtx);
2027 printf("%s: witness exhausted\n", __func__);
2028 return (NULL);
2029 }
2030 w_lock_list_free = lle->ll_next;
2031 mtx_unlock_spin(&w_mtx);
2032 bzero(lle, sizeof(*lle));
2033 return (lle);
2034 }
2035
2036 static void
2037 witness_lock_list_free(struct lock_list_entry *lle)
2038 {
2039
2040 mtx_lock_spin(&w_mtx);
2041 lle->ll_next = w_lock_list_free;
2042 w_lock_list_free = lle;
2043 mtx_unlock_spin(&w_mtx);
2044 }
2045
2046 static struct lock_instance *
2047 find_instance(struct lock_list_entry *list, struct lock_object *lock)
2048 {
2049 struct lock_list_entry *lle;
2050 struct lock_instance *instance;
2051 int i;
2052
2053 for (lle = list; lle != NULL; lle = lle->ll_next)
2054 for (i = lle->ll_count - 1; i >= 0; i--) {
2055 instance = &lle->ll_children[i];
2056 if (instance->li_lock == lock)
2057 return (instance);
2058 }
2059 return (NULL);
2060 }
2061
2062 static void
2063 witness_list_lock(struct lock_instance *instance)
2064 {
2065 struct lock_object *lock;
2066
2067 lock = instance->li_lock;
2068 printf("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ?
2069 "exclusive" : "shared", LOCK_CLASS(lock)->lc_name, lock->lo_name);
2070 if (lock->lo_witness->w_name != lock->lo_name)
2071 printf(" (%s)", lock->lo_witness->w_name);
2072 printf(" r = %d (%p) locked @ %s:%d\n",
2073 instance->li_flags & LI_RECURSEMASK, lock, instance->li_file,
2074 instance->li_line);
2075 }
2076
2077 #ifdef DDB
2078 static int
2079 witness_thread_has_locks(struct thread *td)
2080 {
2081
2082 if (td->td_sleeplocks == NULL)
2083 return (0);
2084 return (td->td_sleeplocks->ll_count != 0);
2085 }
2086
2087 static int
2088 witness_proc_has_locks(struct proc *p)
2089 {
2090 struct thread *td;
2091
2092 FOREACH_THREAD_IN_PROC(p, td) {
2093 if (witness_thread_has_locks(td))
2094 return (1);
2095 }
2096 return (0);
2097 }
2098 #endif
2099
2100 int
2101 witness_list_locks(struct lock_list_entry **lock_list)
2102 {
2103 struct lock_list_entry *lle;
2104 int i, nheld;
2105
2106 nheld = 0;
2107 for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
2108 for (i = lle->ll_count - 1; i >= 0; i--) {
2109 witness_list_lock(&lle->ll_children[i]);
2110 nheld++;
2111 }
2112 return (nheld);
2113 }
2114
2115 /*
2116 * This is a bit risky at best. We call this function when we have timed
2117 * out acquiring a spin lock, and we assume that the other CPU is stuck
2118 * with this lock held. So, we go groveling around in the other CPU's
2119 * per-cpu data to try to find the lock instance for this spin lock to
2120 * see when it was last acquired.
2121 */
2122 void
2123 witness_display_spinlock(struct lock_object *lock, struct thread *owner)
2124 {
2125 struct lock_instance *instance;
2126 struct pcpu *pc;
2127
2128 if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU)
2129 return;
2130 pc = pcpu_find(owner->td_oncpu);
2131 instance = find_instance(pc->pc_spinlocks, lock);
2132 if (instance != NULL)
2133 witness_list_lock(instance);
2134 }
2135
2136 void
2137 witness_save(struct lock_object *lock, const char **filep, int *linep)
2138 {
2139 struct lock_list_entry *lock_list;
2140 struct lock_instance *instance;
2141 struct lock_class *class;
2142
2143 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2144 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2145 return;
2146 class = LOCK_CLASS(lock);
2147 if (class->lc_flags & LC_SLEEPLOCK)
2148 lock_list = curthread->td_sleeplocks;
2149 else {
2150 if (witness_skipspin)
2151 return;
2152 lock_list = PCPU_GET(spinlocks);
2153 }
2154 instance = find_instance(lock_list, lock);
2155 if (instance == NULL)
2156 panic("%s: lock (%s) %s not locked", __func__,
2157 class->lc_name, lock->lo_name);
2158 *filep = instance->li_file;
2159 *linep = instance->li_line;
2160 }
2161
2162 void
2163 witness_restore(struct lock_object *lock, const char *file, int line)
2164 {
2165 struct lock_list_entry *lock_list;
2166 struct lock_instance *instance;
2167 struct lock_class *class;
2168
2169 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2170 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2171 return;
2172 class = LOCK_CLASS(lock);
2173 if (class->lc_flags & LC_SLEEPLOCK)
2174 lock_list = curthread->td_sleeplocks;
2175 else {
2176 if (witness_skipspin)
2177 return;
2178 lock_list = PCPU_GET(spinlocks);
2179 }
2180 instance = find_instance(lock_list, lock);
2181 if (instance == NULL)
2182 panic("%s: lock (%s) %s not locked", __func__,
2183 class->lc_name, lock->lo_name);
2184 lock->lo_witness->w_file = file;
2185 lock->lo_witness->w_line = line;
2186 instance->li_file = file;
2187 instance->li_line = line;
2188 }
2189
2190 void
2191 witness_assert(struct lock_object *lock, int flags, const char *file, int line)
2192 {
2193 #ifdef INVARIANT_SUPPORT
2194 struct lock_instance *instance;
2195 struct lock_class *class;
2196
2197 if (lock->lo_witness == NULL || witness_watch < 1 || panicstr != NULL)
2198 return;
2199 class = LOCK_CLASS(lock);
2200 if ((class->lc_flags & LC_SLEEPLOCK) != 0)
2201 instance = find_instance(curthread->td_sleeplocks, lock);
2202 else if ((class->lc_flags & LC_SPINLOCK) != 0)
2203 instance = find_instance(PCPU_GET(spinlocks), lock);
2204 else {
2205 panic("Lock (%s) %s is not sleep or spin!",
2206 class->lc_name, lock->lo_name);
2207 }
2208 file = fixup_filename(file);
2209 switch (flags) {
2210 case LA_UNLOCKED:
2211 if (instance != NULL)
2212 panic("Lock (%s) %s locked @ %s:%d.",
2213 class->lc_name, lock->lo_name, file, line);
2214 break;
2215 case LA_LOCKED:
2216 case LA_LOCKED | LA_RECURSED:
2217 case LA_LOCKED | LA_NOTRECURSED:
2218 case LA_SLOCKED:
2219 case LA_SLOCKED | LA_RECURSED:
2220 case LA_SLOCKED | LA_NOTRECURSED:
2221 case LA_XLOCKED:
2222 case LA_XLOCKED | LA_RECURSED:
2223 case LA_XLOCKED | LA_NOTRECURSED:
2224 if (instance == NULL) {
2225 panic("Lock (%s) %s not locked @ %s:%d.",
2226 class->lc_name, lock->lo_name, file, line);
2227 break;
2228 }
2229 if ((flags & LA_XLOCKED) != 0 &&
2230 (instance->li_flags & LI_EXCLUSIVE) == 0)
2231 panic("Lock (%s) %s not exclusively locked @ %s:%d.",
2232 class->lc_name, lock->lo_name, file, line);
2233 if ((flags & LA_SLOCKED) != 0 &&
2234 (instance->li_flags & LI_EXCLUSIVE) != 0)
2235 panic("Lock (%s) %s exclusively locked @ %s:%d.",
2236 class->lc_name, lock->lo_name, file, line);
2237 if ((flags & LA_RECURSED) != 0 &&
2238 (instance->li_flags & LI_RECURSEMASK) == 0)
2239 panic("Lock (%s) %s not recursed @ %s:%d.",
2240 class->lc_name, lock->lo_name, file, line);
2241 if ((flags & LA_NOTRECURSED) != 0 &&
2242 (instance->li_flags & LI_RECURSEMASK) != 0)
2243 panic("Lock (%s) %s recursed @ %s:%d.",
2244 class->lc_name, lock->lo_name, file, line);
2245 break;
2246 default:
2247 panic("Invalid lock assertion at %s:%d.", file, line);
2248
2249 }
2250 #endif /* INVARIANT_SUPPORT */
2251 }
2252
2253 static void
2254 witness_setflag(struct lock_object *lock, int flag, int set)
2255 {
2256 struct lock_list_entry *lock_list;
2257 struct lock_instance *instance;
2258 struct lock_class *class;
2259
2260 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2261 return;
2262 class = LOCK_CLASS(lock);
2263 if (class->lc_flags & LC_SLEEPLOCK)
2264 lock_list = curthread->td_sleeplocks;
2265 else {
2266 if (witness_skipspin)
2267 return;
2268 lock_list = PCPU_GET(spinlocks);
2269 }
2270 instance = find_instance(lock_list, lock);
2271 if (instance == NULL)
2272 panic("%s: lock (%s) %s not locked", __func__,
2273 class->lc_name, lock->lo_name);
2274
2275 if (set)
2276 instance->li_flags |= flag;
2277 else
2278 instance->li_flags &= ~flag;
2279 }
2280
2281 void
2282 witness_norelease(struct lock_object *lock)
2283 {
2284
2285 witness_setflag(lock, LI_NORELEASE, 1);
2286 }
2287
2288 void
2289 witness_releaseok(struct lock_object *lock)
2290 {
2291
2292 witness_setflag(lock, LI_NORELEASE, 0);
2293 }
2294
2295 #ifdef DDB
2296 static void
2297 witness_ddb_list(struct thread *td)
2298 {
2299
2300 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2301 KASSERT(kdb_active, ("%s: not in the debugger", __func__));
2302
2303 if (witness_watch < 1)
2304 return;
2305
2306 witness_list_locks(&td->td_sleeplocks);
2307
2308 /*
2309 * We only handle spinlocks if td == curthread. This is somewhat broken
2310 * if td is currently executing on some other CPU and holds spin locks
2311 * as we won't display those locks. If we had a MI way of getting
2312 * the per-cpu data for a given cpu then we could use
2313 * td->td_oncpu to get the list of spinlocks for this thread
2314 * and "fix" this.
2315 *
2316 * That still wouldn't really fix this unless we locked the scheduler
2317 * lock or stopped the other CPU to make sure it wasn't changing the
2318 * list out from under us. It is probably best to just not try to
2319 * handle threads on other CPU's for now.
2320 */
2321 if (td == curthread && PCPU_GET(spinlocks) != NULL)
2322 witness_list_locks(PCPU_PTR(spinlocks));
2323 }
2324
2325 DB_SHOW_COMMAND(locks, db_witness_list)
2326 {
2327 struct thread *td;
2328
2329 if (have_addr)
2330 td = db_lookup_thread(addr, TRUE);
2331 else
2332 td = kdb_thread;
2333 witness_ddb_list(td);
2334 }
2335
2336 DB_SHOW_ALL_COMMAND(locks, db_witness_list_all)
2337 {
2338 struct thread *td;
2339 struct proc *p;
2340
2341 /*
2342 * It would be nice to list only threads and processes that actually
2343 * held sleep locks, but that information is currently not exported
2344 * by WITNESS.
2345 */
2346 FOREACH_PROC_IN_SYSTEM(p) {
2347 if (!witness_proc_has_locks(p))
2348 continue;
2349 FOREACH_THREAD_IN_PROC(p, td) {
2350 if (!witness_thread_has_locks(td))
2351 continue;
2352 db_printf("Process %d (%s) thread %p (%d)\n", p->p_pid,
2353 p->p_comm, td, td->td_tid);
2354 witness_ddb_list(td);
2355 }
2356 }
2357 }
2358 DB_SHOW_ALIAS(alllocks, db_witness_list_all)
2359
2360 DB_SHOW_COMMAND(witness, db_witness_display)
2361 {
2362
2363 witness_ddb_display(db_printf);
2364 }
2365 #endif
2366
2367 static int
2368 sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS)
2369 {
2370 struct witness_lock_order_data *data1, *data2, *tmp_data1, *tmp_data2;
2371 struct witness *tmp_w1, *tmp_w2, *w1, *w2;
2372 struct sbuf *sb;
2373 u_int w_rmatrix1, w_rmatrix2;
2374 int error, generation, i, j;
2375
2376 tmp_data1 = NULL;
2377 tmp_data2 = NULL;
2378 tmp_w1 = NULL;
2379 tmp_w2 = NULL;
2380 if (witness_watch < 1) {
2381 error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2382 return (error);
2383 }
2384 if (witness_cold) {
2385 error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2386 return (error);
2387 }
2388 error = 0;
2389 sb = sbuf_new(NULL, NULL, BADSTACK_SBUF_SIZE, SBUF_AUTOEXTEND);
2390 if (sb == NULL)
2391 return (ENOMEM);
2392
2393 /* Allocate and init temporary storage space. */
2394 tmp_w1 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2395 tmp_w2 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2396 tmp_data1 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2397 M_WAITOK | M_ZERO);
2398 tmp_data2 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2399 M_WAITOK | M_ZERO);
2400 stack_zero(&tmp_data1->wlod_stack);
2401 stack_zero(&tmp_data2->wlod_stack);
2402
2403 restart:
2404 mtx_lock_spin(&w_mtx);
2405 generation = w_generation;
2406 mtx_unlock_spin(&w_mtx);
2407 sbuf_printf(sb, "Number of known direct relationships is %d\n",
2408 w_lohash.wloh_count);
2409 for (i = 1; i < w_max_used_index; i++) {
2410 mtx_lock_spin(&w_mtx);
2411 if (generation != w_generation) {
2412 mtx_unlock_spin(&w_mtx);
2413
2414 /* The graph has changed, try again. */
2415 req->oldidx = 0;
2416 sbuf_clear(sb);
2417 goto restart;
2418 }
2419
2420 w1 = &w_data[i];
2421 if (w1->w_reversed == 0) {
2422 mtx_unlock_spin(&w_mtx);
2423 continue;
2424 }
2425
2426 /* Copy w1 locally so we can release the spin lock. */
2427 *tmp_w1 = *w1;
2428 mtx_unlock_spin(&w_mtx);
2429
2430 if (tmp_w1->w_reversed == 0)
2431 continue;
2432 for (j = 1; j < w_max_used_index; j++) {
2433 if ((w_rmatrix[i][j] & WITNESS_REVERSAL) == 0 || i > j)
2434 continue;
2435
2436 mtx_lock_spin(&w_mtx);
2437 if (generation != w_generation) {
2438 mtx_unlock_spin(&w_mtx);
2439
2440 /* The graph has changed, try again. */
2441 req->oldidx = 0;
2442 sbuf_clear(sb);
2443 goto restart;
2444 }
2445
2446 w2 = &w_data[j];
2447 data1 = witness_lock_order_get(w1, w2);
2448 data2 = witness_lock_order_get(w2, w1);
2449
2450 /*
2451 * Copy information locally so we can release the
2452 * spin lock.
2453 */
2454 *tmp_w2 = *w2;
2455 w_rmatrix1 = (unsigned int)w_rmatrix[i][j];
2456 w_rmatrix2 = (unsigned int)w_rmatrix[j][i];
2457
2458 if (data1) {
2459 stack_zero(&tmp_data1->wlod_stack);
2460 stack_copy(&data1->wlod_stack,
2461 &tmp_data1->wlod_stack);
2462 }
2463 if (data2 && data2 != data1) {
2464 stack_zero(&tmp_data2->wlod_stack);
2465 stack_copy(&data2->wlod_stack,
2466 &tmp_data2->wlod_stack);
2467 }
2468 mtx_unlock_spin(&w_mtx);
2469
2470 sbuf_printf(sb,
2471 "\nLock order reversal between \"%s\"(%s) and \"%s\"(%s)!\n",
2472 tmp_w1->w_name, tmp_w1->w_class->lc_name,
2473 tmp_w2->w_name, tmp_w2->w_class->lc_name);
2474 #if 0
2475 sbuf_printf(sb,
2476 "w_rmatrix[%s][%s] == %x, w_rmatrix[%s][%s] == %x\n",
2477 tmp_w1->name, tmp_w2->w_name, w_rmatrix1,
2478 tmp_w2->name, tmp_w1->w_name, w_rmatrix2);
2479 #endif
2480 if (data1) {
2481 sbuf_printf(sb,
2482 "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2483 tmp_w1->w_name, tmp_w1->w_class->lc_name,
2484 tmp_w2->w_name, tmp_w2->w_class->lc_name);
2485 stack_sbuf_print(sb, &tmp_data1->wlod_stack);
2486 sbuf_printf(sb, "\n");
2487 }
2488 if (data2 && data2 != data1) {
2489 sbuf_printf(sb,
2490 "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2491 tmp_w2->w_name, tmp_w2->w_class->lc_name,
2492 tmp_w1->w_name, tmp_w1->w_class->lc_name);
2493 stack_sbuf_print(sb, &tmp_data2->wlod_stack);
2494 sbuf_printf(sb, "\n");
2495 }
2496 }
2497 }
2498 mtx_lock_spin(&w_mtx);
2499 if (generation != w_generation) {
2500 mtx_unlock_spin(&w_mtx);
2501
2502 /*
2503 * The graph changed while we were printing stack data,
2504 * try again.
2505 */
2506 req->oldidx = 0;
2507 sbuf_clear(sb);
2508 goto restart;
2509 }
2510 mtx_unlock_spin(&w_mtx);
2511
2512 /* Free temporary storage space. */
2513 free(tmp_data1, M_TEMP);
2514 free(tmp_data2, M_TEMP);
2515 free(tmp_w1, M_TEMP);
2516 free(tmp_w2, M_TEMP);
2517
2518 sbuf_finish(sb);
2519 error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1);
2520 sbuf_delete(sb);
2521
2522 return (error);
2523 }
2524
2525 static int
2526 sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS)
2527 {
2528 struct witness *w;
2529 struct sbuf *sb;
2530 int error;
2531
2532 if (witness_watch < 1) {
2533 error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2534 return (error);
2535 }
2536 if (witness_cold) {
2537 error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2538 return (error);
2539 }
2540 error = 0;
2541 sb = sbuf_new(NULL, NULL, FULLGRAPH_SBUF_SIZE, SBUF_FIXEDLEN);
2542 if (sb == NULL)
2543 return (ENOMEM);
2544 sbuf_printf(sb, "\n");
2545
2546 mtx_lock_spin(&w_mtx);
2547 STAILQ_FOREACH(w, &w_all, w_list)
2548 w->w_displayed = 0;
2549 STAILQ_FOREACH(w, &w_all, w_list)
2550 witness_add_fullgraph(sb, w);
2551 mtx_unlock_spin(&w_mtx);
2552
2553 /*
2554 * While using SBUF_FIXEDLEN, check if the sbuf overflowed.
2555 */
2556 if (sbuf_overflowed(sb)) {
2557 sbuf_delete(sb);
2558 panic("%s: sbuf overflowed, bump FULLGRAPH_SBUF_SIZE value\n",
2559 __func__);
2560 }
2561
2562 /*
2563 * Close the sbuf and return to userland.
2564 */
2565 sbuf_finish(sb);
2566 error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1);
2567 sbuf_delete(sb);
2568
2569 return (error);
2570 }
2571
2572 static int
2573 sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS)
2574 {
2575 int error, value;
2576
2577 value = witness_watch;
2578 error = sysctl_handle_int(oidp, &value, 0, req);
2579 if (error != 0 || req->newptr == NULL)
2580 return (error);
2581 if (value > 1 || value < -1 ||
2582 (witness_watch == -1 && value != witness_watch))
2583 return (EINVAL);
2584 witness_watch = value;
2585 return (0);
2586 }
2587
2588 static void
2589 witness_add_fullgraph(struct sbuf *sb, struct witness *w)
2590 {
2591 int i;
2592
2593 if (w->w_displayed != 0 || (w->w_file == NULL && w->w_line == 0))
2594 return;
2595 w->w_displayed = 1;
2596
2597 WITNESS_INDEX_ASSERT(w->w_index);
2598 for (i = 1; i <= w_max_used_index; i++) {
2599 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT) {
2600 sbuf_printf(sb, "\"%s\",\"%s\"\n", w->w_name,
2601 w_data[i].w_name);
2602 witness_add_fullgraph(sb, &w_data[i]);
2603 }
2604 }
2605 }
2606
2607 /*
2608 * A simple hash function. Takes a key pointer and a key size. If size == 0,
2609 * interprets the key as a string and reads until the null
2610 * terminator. Otherwise, reads the first size bytes. Returns an unsigned 32-bit
2611 * hash value computed from the key.
2612 */
2613 static uint32_t
2614 witness_hash_djb2(const uint8_t *key, uint32_t size)
2615 {
2616 unsigned int hash = 5381;
2617 int i;
2618
2619 /* hash = hash * 33 + key[i] */
2620 if (size)
2621 for (i = 0; i < size; i++)
2622 hash = ((hash << 5) + hash) + (unsigned int)key[i];
2623 else
2624 for (i = 0; key[i] != 0; i++)
2625 hash = ((hash << 5) + hash) + (unsigned int)key[i];
2626
2627 return (hash);
2628 }
2629
2630
2631 /*
2632 * Initializes the two witness hash tables. Called exactly once from
2633 * witness_initialize().
2634 */
2635 static void
2636 witness_init_hash_tables(void)
2637 {
2638 int i;
2639
2640 MPASS(witness_cold);
2641
2642 /* Initialize the hash tables. */
2643 for (i = 0; i < WITNESS_HASH_SIZE; i++)
2644 w_hash.wh_array[i] = NULL;
2645
2646 w_hash.wh_size = WITNESS_HASH_SIZE;
2647 w_hash.wh_count = 0;
2648
2649 /* Initialize the lock order data hash. */
2650 w_lofree = NULL;
2651 for (i = 0; i < WITNESS_LO_DATA_COUNT; i++) {
2652 memset(&w_lodata[i], 0, sizeof(w_lodata[i]));
2653 w_lodata[i].wlod_next = w_lofree;
2654 w_lofree = &w_lodata[i];
2655 }
2656 w_lohash.wloh_size = WITNESS_LO_HASH_SIZE;
2657 w_lohash.wloh_count = 0;
2658 for (i = 0; i < WITNESS_LO_HASH_SIZE; i++)
2659 w_lohash.wloh_array[i] = NULL;
2660 }
2661
2662 static struct witness *
2663 witness_hash_get(const char *key)
2664 {
2665 struct witness *w;
2666 uint32_t hash;
2667
2668 MPASS(key != NULL);
2669 if (witness_cold == 0)
2670 mtx_assert(&w_mtx, MA_OWNED);
2671 hash = witness_hash_djb2(key, 0) % w_hash.wh_size;
2672 w = w_hash.wh_array[hash];
2673 while (w != NULL) {
2674 if (strcmp(w->w_name, key) == 0)
2675 goto out;
2676 w = w->w_hash_next;
2677 }
2678
2679 out:
2680 return (w);
2681 }
2682
2683 static void
2684 witness_hash_put(struct witness *w)
2685 {
2686 uint32_t hash;
2687
2688 MPASS(w != NULL);
2689 MPASS(w->w_name != NULL);
2690 if (witness_cold == 0)
2691 mtx_assert(&w_mtx, MA_OWNED);
2692 KASSERT(witness_hash_get(w->w_name) == NULL,
2693 ("%s: trying to add a hash entry that already exists!", __func__));
2694 KASSERT(w->w_hash_next == NULL,
2695 ("%s: w->w_hash_next != NULL", __func__));
2696
2697 hash = witness_hash_djb2(w->w_name, 0) % w_hash.wh_size;
2698 w->w_hash_next = w_hash.wh_array[hash];
2699 w_hash.wh_array[hash] = w;
2700 w_hash.wh_count++;
2701 }
2702
2703
2704 static struct witness_lock_order_data *
2705 witness_lock_order_get(struct witness *parent, struct witness *child)
2706 {
2707 struct witness_lock_order_data *data = NULL;
2708 struct witness_lock_order_key key;
2709 unsigned int hash;
2710
2711 MPASS(parent != NULL && child != NULL);
2712 key.from = parent->w_index;
2713 key.to = child->w_index;
2714 WITNESS_INDEX_ASSERT(key.from);
2715 WITNESS_INDEX_ASSERT(key.to);
2716 if ((w_rmatrix[parent->w_index][child->w_index]
2717 & WITNESS_LOCK_ORDER_KNOWN) == 0)
2718 goto out;
2719
2720 hash = witness_hash_djb2((const char*)&key,
2721 sizeof(key)) % w_lohash.wloh_size;
2722 data = w_lohash.wloh_array[hash];
2723 while (data != NULL) {
2724 if (witness_lock_order_key_equal(&data->wlod_key, &key))
2725 break;
2726 data = data->wlod_next;
2727 }
2728
2729 out:
2730 return (data);
2731 }
2732
2733 /*
2734 * Verify that parent and child have a known relationship, are not the same,
2735 * and child is actually a child of parent. This is done without w_mtx
2736 * to avoid contention in the common case.
2737 */
2738 static int
2739 witness_lock_order_check(struct witness *parent, struct witness *child)
2740 {
2741
2742 if (parent != child &&
2743 w_rmatrix[parent->w_index][child->w_index]
2744 & WITNESS_LOCK_ORDER_KNOWN &&
2745 isitmychild(parent, child))
2746 return (1);
2747
2748 return (0);
2749 }
2750
2751 static int
2752 witness_lock_order_add(struct witness *parent, struct witness *child)
2753 {
2754 struct witness_lock_order_data *data = NULL;
2755 struct witness_lock_order_key key;
2756 unsigned int hash;
2757
2758 MPASS(parent != NULL && child != NULL);
2759 key.from = parent->w_index;
2760 key.to = child->w_index;
2761 WITNESS_INDEX_ASSERT(key.from);
2762 WITNESS_INDEX_ASSERT(key.to);
2763 if (w_rmatrix[parent->w_index][child->w_index]
2764 & WITNESS_LOCK_ORDER_KNOWN)
2765 return (1);
2766
2767 hash = witness_hash_djb2((const char*)&key,
2768 sizeof(key)) % w_lohash.wloh_size;
2769 w_rmatrix[parent->w_index][child->w_index] |= WITNESS_LOCK_ORDER_KNOWN;
2770 data = w_lofree;
2771 if (data == NULL)
2772 return (0);
2773 w_lofree = data->wlod_next;
2774 data->wlod_next = w_lohash.wloh_array[hash];
2775 data->wlod_key = key;
2776 w_lohash.wloh_array[hash] = data;
2777 w_lohash.wloh_count++;
2778 stack_zero(&data->wlod_stack);
2779 stack_save(&data->wlod_stack);
2780 return (1);
2781 }
2782
2783 /* Call this whenver the structure of the witness graph changes. */
2784 static void
2785 witness_increment_graph_generation(void)
2786 {
2787
2788 if (witness_cold == 0)
2789 mtx_assert(&w_mtx, MA_OWNED);
2790 w_generation++;
2791 }
2792
2793 #ifdef KDB
2794 static void
2795 _witness_debugger(int cond, const char *msg)
2796 {
2797
2798 if (witness_trace && cond)
2799 kdb_backtrace();
2800 if (witness_kdb && cond)
2801 kdb_enter(KDB_WHY_WITNESS, msg);
2802 }
2803 #endif
Cache object: 493fa85b02f7d1df07570ebd0d9f6fc7
|