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("$FreeBSD$");
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/syslog.h>
110 #include <sys/systm.h>
111
112 #ifdef DDB
113 #include <ddb/ddb.h>
114 #endif
115
116 #include <machine/stdarg.h>
117
118 #if !defined(DDB) && !defined(STACK)
119 #error "DDB or STACK options are required for WITNESS"
120 #endif
121
122 /* Note that these traces do not work with KTR_ALQ. */
123 #if 0
124 #define KTR_WITNESS KTR_SUBSYS
125 #else
126 #define KTR_WITNESS 0
127 #endif
128
129 #define LI_RECURSEMASK 0x0000ffff /* Recursion depth of lock instance. */
130 #define LI_EXCLUSIVE 0x00010000 /* Exclusive lock instance. */
131 #define LI_NORELEASE 0x00020000 /* Lock not allowed to be released. */
132
133 /* Define this to check for blessed mutexes */
134 #undef BLESSING
135
136 #ifndef WITNESS_COUNT
137 #define WITNESS_COUNT 1536
138 #endif
139 #define WITNESS_HASH_SIZE 251 /* Prime, gives load factor < 2 */
140 #define WITNESS_PENDLIST (1024 + MAXCPU)
141
142 /* Allocate 256 KB of stack data space */
143 #define WITNESS_LO_DATA_COUNT 2048
144
145 /* Prime, gives load factor of ~2 at full load */
146 #define WITNESS_LO_HASH_SIZE 1021
147
148 /*
149 * XXX: This is somewhat bogus, as we assume here that at most 2048 threads
150 * will hold LOCK_NCHILDREN locks. We handle failure ok, and we should
151 * probably be safe for the most part, but it's still a SWAG.
152 */
153 #define LOCK_NCHILDREN 5
154 #define LOCK_CHILDCOUNT 2048
155
156 #define MAX_W_NAME 64
157
158 #define FULLGRAPH_SBUF_SIZE 512
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 static 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_equal(const struct witness_lock_order_key *a,
310 const struct witness_lock_order_key *b)
311 {
312
313 return (a->from == b->from && a->to == b->to);
314 }
315
316 static int _isitmyx(struct witness *w1, struct witness *w2, int rmask,
317 const char *fname);
318 static void adopt(struct witness *parent, struct witness *child);
319 #ifdef BLESSING
320 static int blessed(struct witness *, struct witness *);
321 #endif
322 static void depart(struct witness *w);
323 static struct witness *enroll(const char *description,
324 struct lock_class *lock_class);
325 static struct lock_instance *find_instance(struct lock_list_entry *list,
326 const struct lock_object *lock);
327 static int isitmychild(struct witness *parent, struct witness *child);
328 static int isitmydescendant(struct witness *parent, struct witness *child);
329 static void itismychild(struct witness *parent, struct witness *child);
330 static int sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS);
331 static int sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS);
332 static int sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS);
333 static int sysctl_debug_witness_channel(SYSCTL_HANDLER_ARGS);
334 static void witness_add_fullgraph(struct sbuf *sb, struct witness *parent);
335 #ifdef DDB
336 static void witness_ddb_compute_levels(void);
337 static void witness_ddb_display(int(*)(const char *fmt, ...));
338 static void witness_ddb_display_descendants(int(*)(const char *fmt, ...),
339 struct witness *, int indent);
340 static void witness_ddb_display_list(int(*prnt)(const char *fmt, ...),
341 struct witness_list *list);
342 static void witness_ddb_level_descendants(struct witness *parent, int l);
343 static void witness_ddb_list(struct thread *td);
344 #endif
345 static void witness_debugger(int cond, const char *msg);
346 static void witness_free(struct witness *m);
347 static struct witness *witness_get(void);
348 static uint32_t witness_hash_djb2(const uint8_t *key, uint32_t size);
349 static struct witness *witness_hash_get(const char *key);
350 static void witness_hash_put(struct witness *w);
351 static void witness_init_hash_tables(void);
352 static void witness_increment_graph_generation(void);
353 static void witness_lock_list_free(struct lock_list_entry *lle);
354 static struct lock_list_entry *witness_lock_list_get(void);
355 static int witness_lock_order_add(struct witness *parent,
356 struct witness *child);
357 static int witness_lock_order_check(struct witness *parent,
358 struct witness *child);
359 static struct witness_lock_order_data *witness_lock_order_get(
360 struct witness *parent,
361 struct witness *child);
362 static void witness_list_lock(struct lock_instance *instance,
363 int (*prnt)(const char *fmt, ...));
364 static int witness_output(const char *fmt, ...) __printflike(1, 2);
365 static int witness_voutput(const char *fmt, va_list ap) __printflike(1, 0);
366 static void witness_setflag(struct lock_object *lock, int flag, int set);
367
368 static SYSCTL_NODE(_debug, OID_AUTO, witness, CTLFLAG_RW, NULL,
369 "Witness Locking");
370
371 /*
372 * If set to 0, lock order checking is disabled. If set to -1,
373 * witness is completely disabled. Otherwise witness performs full
374 * lock order checking for all locks. At runtime, lock order checking
375 * may be toggled. However, witness cannot be reenabled once it is
376 * completely disabled.
377 */
378 static int witness_watch = 1;
379 SYSCTL_PROC(_debug_witness, OID_AUTO, watch, CTLFLAG_RWTUN | CTLTYPE_INT, NULL, 0,
380 sysctl_debug_witness_watch, "I", "witness is watching lock operations");
381
382 #ifdef KDB
383 /*
384 * When KDB is enabled and witness_kdb is 1, it will cause the system
385 * to drop into kdebug() when:
386 * - a lock hierarchy violation occurs
387 * - locks are held when going to sleep.
388 */
389 #ifdef WITNESS_KDB
390 int witness_kdb = 1;
391 #else
392 int witness_kdb = 0;
393 #endif
394 SYSCTL_INT(_debug_witness, OID_AUTO, kdb, CTLFLAG_RWTUN, &witness_kdb, 0, "");
395 #endif /* KDB */
396
397 #if defined(DDB) || defined(KDB)
398 /*
399 * When DDB or KDB is enabled and witness_trace is 1, it will cause the system
400 * to print a stack trace:
401 * - a lock hierarchy violation occurs
402 * - locks are held when going to sleep.
403 */
404 int witness_trace = 1;
405 SYSCTL_INT(_debug_witness, OID_AUTO, trace, CTLFLAG_RWTUN, &witness_trace, 0, "");
406 #endif /* DDB || KDB */
407
408 #ifdef WITNESS_SKIPSPIN
409 int witness_skipspin = 1;
410 #else
411 int witness_skipspin = 0;
412 #endif
413 SYSCTL_INT(_debug_witness, OID_AUTO, skipspin, CTLFLAG_RDTUN, &witness_skipspin, 0, "");
414
415 int badstack_sbuf_size;
416
417 int witness_count = WITNESS_COUNT;
418 SYSCTL_INT(_debug_witness, OID_AUTO, witness_count, CTLFLAG_RDTUN,
419 &witness_count, 0, "");
420
421 /*
422 * Output channel for witness messages. By default we print to the console.
423 */
424 enum witness_channel {
425 WITNESS_CONSOLE,
426 WITNESS_LOG,
427 WITNESS_NONE,
428 };
429
430 static enum witness_channel witness_channel = WITNESS_CONSOLE;
431 SYSCTL_PROC(_debug_witness, OID_AUTO, output_channel, CTLTYPE_STRING |
432 CTLFLAG_RWTUN, NULL, 0, sysctl_debug_witness_channel, "A",
433 "Output channel for warnings");
434
435 /*
436 * Call this to print out the relations between locks.
437 */
438 SYSCTL_PROC(_debug_witness, OID_AUTO, fullgraph, CTLTYPE_STRING | CTLFLAG_RD,
439 NULL, 0, sysctl_debug_witness_fullgraph, "A", "Show locks relation graphs");
440
441 /*
442 * Call this to print out the witness faulty stacks.
443 */
444 SYSCTL_PROC(_debug_witness, OID_AUTO, badstacks, CTLTYPE_STRING | CTLFLAG_RD,
445 NULL, 0, sysctl_debug_witness_badstacks, "A", "Show bad witness stacks");
446
447 static struct mtx w_mtx;
448
449 /* w_list */
450 static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free);
451 static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all);
452
453 /* w_typelist */
454 static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin);
455 static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep);
456
457 /* lock list */
458 static struct lock_list_entry *w_lock_list_free = NULL;
459 static struct witness_pendhelp pending_locks[WITNESS_PENDLIST];
460 static u_int pending_cnt;
461
462 static int w_free_cnt, w_spin_cnt, w_sleep_cnt;
463 SYSCTL_INT(_debug_witness, OID_AUTO, free_cnt, CTLFLAG_RD, &w_free_cnt, 0, "");
464 SYSCTL_INT(_debug_witness, OID_AUTO, spin_cnt, CTLFLAG_RD, &w_spin_cnt, 0, "");
465 SYSCTL_INT(_debug_witness, OID_AUTO, sleep_cnt, CTLFLAG_RD, &w_sleep_cnt, 0,
466 "");
467
468 static struct witness *w_data;
469 static uint8_t **w_rmatrix;
470 static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT];
471 static struct witness_hash w_hash; /* The witness hash table. */
472
473 /* The lock order data hash */
474 static struct witness_lock_order_data w_lodata[WITNESS_LO_DATA_COUNT];
475 static struct witness_lock_order_data *w_lofree = NULL;
476 static struct witness_lock_order_hash w_lohash;
477 static int w_max_used_index = 0;
478 static unsigned int w_generation = 0;
479 static const char w_notrunning[] = "Witness not running\n";
480 static const char w_stillcold[] = "Witness is still cold\n";
481
482
483 static struct witness_order_list_entry order_lists[] = {
484 /*
485 * sx locks
486 */
487 { "proctree", &lock_class_sx },
488 { "allproc", &lock_class_sx },
489 { "allprison", &lock_class_sx },
490 { NULL, NULL },
491 /*
492 * Various mutexes
493 */
494 { "Giant", &lock_class_mtx_sleep },
495 { "pipe mutex", &lock_class_mtx_sleep },
496 { "sigio lock", &lock_class_mtx_sleep },
497 { "process group", &lock_class_mtx_sleep },
498 { "process lock", &lock_class_mtx_sleep },
499 { "session", &lock_class_mtx_sleep },
500 { "uidinfo hash", &lock_class_rw },
501 #ifdef HWPMC_HOOKS
502 { "pmc-sleep", &lock_class_mtx_sleep },
503 #endif
504 { "time lock", &lock_class_mtx_sleep },
505 { NULL, NULL },
506 /*
507 * umtx
508 */
509 { "umtx lock", &lock_class_mtx_sleep },
510 { NULL, NULL },
511 /*
512 * Sockets
513 */
514 { "accept", &lock_class_mtx_sleep },
515 { "so_snd", &lock_class_mtx_sleep },
516 { "so_rcv", &lock_class_mtx_sleep },
517 { "sellck", &lock_class_mtx_sleep },
518 { NULL, NULL },
519 /*
520 * Routing
521 */
522 { "so_rcv", &lock_class_mtx_sleep },
523 { "radix node head", &lock_class_rw },
524 { "rtentry", &lock_class_mtx_sleep },
525 { "ifaddr", &lock_class_mtx_sleep },
526 { NULL, NULL },
527 /*
528 * IPv4 multicast:
529 * protocol locks before interface locks, after UDP locks.
530 */
531 { "udpinp", &lock_class_rw },
532 { "in_multi_mtx", &lock_class_mtx_sleep },
533 { "igmp_mtx", &lock_class_mtx_sleep },
534 { "if_addr_lock", &lock_class_rw },
535 { NULL, NULL },
536 /*
537 * IPv6 multicast:
538 * protocol locks before interface locks, after UDP locks.
539 */
540 { "udpinp", &lock_class_rw },
541 { "in6_multi_mtx", &lock_class_mtx_sleep },
542 { "mld_mtx", &lock_class_mtx_sleep },
543 { "if_addr_lock", &lock_class_rw },
544 { NULL, NULL },
545 /*
546 * UNIX Domain Sockets
547 */
548 { "unp_link_rwlock", &lock_class_rw },
549 { "unp_list_lock", &lock_class_mtx_sleep },
550 { "unp", &lock_class_mtx_sleep },
551 { "so_snd", &lock_class_mtx_sleep },
552 { NULL, NULL },
553 /*
554 * UDP/IP
555 */
556 { "udp", &lock_class_rw },
557 { "udpinp", &lock_class_rw },
558 { "so_snd", &lock_class_mtx_sleep },
559 { NULL, NULL },
560 /*
561 * TCP/IP
562 */
563 { "tcp", &lock_class_rw },
564 { "tcpinp", &lock_class_rw },
565 { "so_snd", &lock_class_mtx_sleep },
566 { NULL, NULL },
567 /*
568 * BPF
569 */
570 { "bpf global lock", &lock_class_sx },
571 { "bpf interface lock", &lock_class_rw },
572 { "bpf cdev lock", &lock_class_mtx_sleep },
573 { NULL, NULL },
574 /*
575 * NFS server
576 */
577 { "nfsd_mtx", &lock_class_mtx_sleep },
578 { "so_snd", &lock_class_mtx_sleep },
579 { NULL, NULL },
580
581 /*
582 * IEEE 802.11
583 */
584 { "802.11 com lock", &lock_class_mtx_sleep},
585 { NULL, NULL },
586 /*
587 * Network drivers
588 */
589 { "network driver", &lock_class_mtx_sleep},
590 { NULL, NULL },
591
592 /*
593 * Netgraph
594 */
595 { "ng_node", &lock_class_mtx_sleep },
596 { "ng_worklist", &lock_class_mtx_sleep },
597 { NULL, NULL },
598 /*
599 * CDEV
600 */
601 { "vm map (system)", &lock_class_mtx_sleep },
602 { "vm pagequeue", &lock_class_mtx_sleep },
603 { "vnode interlock", &lock_class_mtx_sleep },
604 { "cdev", &lock_class_mtx_sleep },
605 { NULL, NULL },
606 /*
607 * VM
608 */
609 { "vm map (user)", &lock_class_sx },
610 { "vm object", &lock_class_rw },
611 { "vm page", &lock_class_mtx_sleep },
612 { "vm pagequeue", &lock_class_mtx_sleep },
613 { "pmap pv global", &lock_class_rw },
614 { "pmap", &lock_class_mtx_sleep },
615 { "pmap pv list", &lock_class_rw },
616 { "vm page free queue", &lock_class_mtx_sleep },
617 { NULL, NULL },
618 /*
619 * kqueue/VFS interaction
620 */
621 { "kqueue", &lock_class_mtx_sleep },
622 { "struct mount mtx", &lock_class_mtx_sleep },
623 { "vnode interlock", &lock_class_mtx_sleep },
624 { NULL, NULL },
625 /*
626 * VFS namecache
627 */
628 { "ncvn", &lock_class_mtx_sleep },
629 { "ncbuc", &lock_class_rw },
630 { "vnode interlock", &lock_class_mtx_sleep },
631 { "ncneg", &lock_class_mtx_sleep },
632 { NULL, NULL },
633 /*
634 * ZFS locking
635 */
636 { "dn->dn_mtx", &lock_class_sx },
637 { "dr->dt.di.dr_mtx", &lock_class_sx },
638 { "db->db_mtx", &lock_class_sx },
639 { NULL, NULL },
640 /*
641 * spin locks
642 */
643 #ifdef SMP
644 { "ap boot", &lock_class_mtx_spin },
645 #endif
646 { "rm.mutex_mtx", &lock_class_mtx_spin },
647 { "sio", &lock_class_mtx_spin },
648 #ifdef __i386__
649 { "cy", &lock_class_mtx_spin },
650 #endif
651 #ifdef __sparc64__
652 { "pcib_mtx", &lock_class_mtx_spin },
653 { "rtc_mtx", &lock_class_mtx_spin },
654 #endif
655 { "scc_hwmtx", &lock_class_mtx_spin },
656 { "uart_hwmtx", &lock_class_mtx_spin },
657 { "fast_taskqueue", &lock_class_mtx_spin },
658 { "intr table", &lock_class_mtx_spin },
659 #ifdef HWPMC_HOOKS
660 { "pmc-per-proc", &lock_class_mtx_spin },
661 #endif
662 { "process slock", &lock_class_mtx_spin },
663 { "syscons video lock", &lock_class_mtx_spin },
664 { "sleepq chain", &lock_class_mtx_spin },
665 { "rm_spinlock", &lock_class_mtx_spin },
666 { "turnstile chain", &lock_class_mtx_spin },
667 { "turnstile lock", &lock_class_mtx_spin },
668 { "sched lock", &lock_class_mtx_spin },
669 { "td_contested", &lock_class_mtx_spin },
670 { "callout", &lock_class_mtx_spin },
671 { "entropy harvest mutex", &lock_class_mtx_spin },
672 #ifdef SMP
673 { "smp rendezvous", &lock_class_mtx_spin },
674 #endif
675 #ifdef __powerpc__
676 { "tlb0", &lock_class_mtx_spin },
677 #endif
678 /*
679 * leaf locks
680 */
681 { "intrcnt", &lock_class_mtx_spin },
682 { "icu", &lock_class_mtx_spin },
683 #if defined(SMP) && defined(__sparc64__)
684 { "ipi", &lock_class_mtx_spin },
685 #endif
686 #ifdef __i386__
687 { "allpmaps", &lock_class_mtx_spin },
688 { "descriptor tables", &lock_class_mtx_spin },
689 #endif
690 { "clk", &lock_class_mtx_spin },
691 { "cpuset", &lock_class_mtx_spin },
692 { "mprof lock", &lock_class_mtx_spin },
693 { "zombie lock", &lock_class_mtx_spin },
694 { "ALD Queue", &lock_class_mtx_spin },
695 #if defined(__i386__) || defined(__amd64__)
696 { "pcicfg", &lock_class_mtx_spin },
697 { "NDIS thread lock", &lock_class_mtx_spin },
698 #endif
699 { "tw_osl_io_lock", &lock_class_mtx_spin },
700 { "tw_osl_q_lock", &lock_class_mtx_spin },
701 { "tw_cl_io_lock", &lock_class_mtx_spin },
702 { "tw_cl_intr_lock", &lock_class_mtx_spin },
703 { "tw_cl_gen_lock", &lock_class_mtx_spin },
704 #ifdef HWPMC_HOOKS
705 { "pmc-leaf", &lock_class_mtx_spin },
706 #endif
707 { "blocked lock", &lock_class_mtx_spin },
708 { NULL, NULL },
709 { NULL, NULL }
710 };
711
712 #ifdef BLESSING
713 /*
714 * Pairs of locks which have been blessed
715 * Don't complain about order problems with blessed locks
716 */
717 static struct witness_blessed blessed_list[] = {
718 };
719 #endif
720
721 /*
722 * This global is set to 0 once it becomes safe to use the witness code.
723 */
724 static int witness_cold = 1;
725
726 /*
727 * This global is set to 1 once the static lock orders have been enrolled
728 * so that a warning can be issued for any spin locks enrolled later.
729 */
730 static int witness_spin_warn = 0;
731
732 /* Trim useless garbage from filenames. */
733 static const char *
734 fixup_filename(const char *file)
735 {
736
737 if (file == NULL)
738 return (NULL);
739 while (strncmp(file, "../", 3) == 0)
740 file += 3;
741 return (file);
742 }
743
744 /*
745 * The WITNESS-enabled diagnostic code. Note that the witness code does
746 * assume that the early boot is single-threaded at least until after this
747 * routine is completed.
748 */
749 static void
750 witness_initialize(void *dummy __unused)
751 {
752 struct lock_object *lock;
753 struct witness_order_list_entry *order;
754 struct witness *w, *w1;
755 int i;
756
757 w_data = malloc(sizeof (struct witness) * witness_count, M_WITNESS,
758 M_WAITOK | M_ZERO);
759
760 w_rmatrix = malloc(sizeof(*w_rmatrix) * (witness_count + 1),
761 M_WITNESS, M_WAITOK | M_ZERO);
762
763 for (i = 0; i < witness_count + 1; i++) {
764 w_rmatrix[i] = malloc(sizeof(*w_rmatrix[i]) *
765 (witness_count + 1), M_WITNESS, M_WAITOK | M_ZERO);
766 }
767 badstack_sbuf_size = witness_count * 256;
768
769 /*
770 * We have to release Giant before initializing its witness
771 * structure so that WITNESS doesn't get confused.
772 */
773 mtx_unlock(&Giant);
774 mtx_assert(&Giant, MA_NOTOWNED);
775
776 CTR1(KTR_WITNESS, "%s: initializing witness", __func__);
777 mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET |
778 MTX_NOWITNESS | MTX_NOPROFILE);
779 for (i = witness_count - 1; i >= 0; i--) {
780 w = &w_data[i];
781 memset(w, 0, sizeof(*w));
782 w_data[i].w_index = i; /* Witness index never changes. */
783 witness_free(w);
784 }
785 KASSERT(STAILQ_FIRST(&w_free)->w_index == 0,
786 ("%s: Invalid list of free witness objects", __func__));
787
788 /* Witness with index 0 is not used to aid in debugging. */
789 STAILQ_REMOVE_HEAD(&w_free, w_list);
790 w_free_cnt--;
791
792 for (i = 0; i < witness_count; i++) {
793 memset(w_rmatrix[i], 0, sizeof(*w_rmatrix[i]) *
794 (witness_count + 1));
795 }
796
797 for (i = 0; i < LOCK_CHILDCOUNT; i++)
798 witness_lock_list_free(&w_locklistdata[i]);
799 witness_init_hash_tables();
800
801 /* First add in all the specified order lists. */
802 for (order = order_lists; order->w_name != NULL; order++) {
803 w = enroll(order->w_name, order->w_class);
804 if (w == NULL)
805 continue;
806 w->w_file = "order list";
807 for (order++; order->w_name != NULL; order++) {
808 w1 = enroll(order->w_name, order->w_class);
809 if (w1 == NULL)
810 continue;
811 w1->w_file = "order list";
812 itismychild(w, w1);
813 w = w1;
814 }
815 }
816 witness_spin_warn = 1;
817
818 /* Iterate through all locks and add them to witness. */
819 for (i = 0; pending_locks[i].wh_lock != NULL; i++) {
820 lock = pending_locks[i].wh_lock;
821 KASSERT(lock->lo_flags & LO_WITNESS,
822 ("%s: lock %s is on pending list but not LO_WITNESS",
823 __func__, lock->lo_name));
824 lock->lo_witness = enroll(pending_locks[i].wh_type,
825 LOCK_CLASS(lock));
826 }
827
828 /* Mark the witness code as being ready for use. */
829 witness_cold = 0;
830
831 mtx_lock(&Giant);
832 }
833 SYSINIT(witness_init, SI_SUB_WITNESS, SI_ORDER_FIRST, witness_initialize,
834 NULL);
835
836 void
837 witness_init(struct lock_object *lock, const char *type)
838 {
839 struct lock_class *class;
840
841 /* Various sanity checks. */
842 class = LOCK_CLASS(lock);
843 if ((lock->lo_flags & LO_RECURSABLE) != 0 &&
844 (class->lc_flags & LC_RECURSABLE) == 0)
845 kassert_panic("%s: lock (%s) %s can not be recursable",
846 __func__, class->lc_name, lock->lo_name);
847 if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
848 (class->lc_flags & LC_SLEEPABLE) == 0)
849 kassert_panic("%s: lock (%s) %s can not be sleepable",
850 __func__, class->lc_name, lock->lo_name);
851 if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
852 (class->lc_flags & LC_UPGRADABLE) == 0)
853 kassert_panic("%s: lock (%s) %s can not be upgradable",
854 __func__, class->lc_name, lock->lo_name);
855
856 /*
857 * If we shouldn't watch this lock, then just clear lo_witness.
858 * Otherwise, if witness_cold is set, then it is too early to
859 * enroll this lock, so defer it to witness_initialize() by adding
860 * it to the pending_locks list. If it is not too early, then enroll
861 * the lock now.
862 */
863 if (witness_watch < 1 || panicstr != NULL ||
864 (lock->lo_flags & LO_WITNESS) == 0)
865 lock->lo_witness = NULL;
866 else if (witness_cold) {
867 pending_locks[pending_cnt].wh_lock = lock;
868 pending_locks[pending_cnt++].wh_type = type;
869 if (pending_cnt > WITNESS_PENDLIST)
870 panic("%s: pending locks list is too small, "
871 "increase WITNESS_PENDLIST\n",
872 __func__);
873 } else
874 lock->lo_witness = enroll(type, class);
875 }
876
877 void
878 witness_destroy(struct lock_object *lock)
879 {
880 struct lock_class *class;
881 struct witness *w;
882
883 class = LOCK_CLASS(lock);
884
885 if (witness_cold)
886 panic("lock (%s) %s destroyed while witness_cold",
887 class->lc_name, lock->lo_name);
888
889 /* XXX: need to verify that no one holds the lock */
890 if ((lock->lo_flags & LO_WITNESS) == 0 || lock->lo_witness == NULL)
891 return;
892 w = lock->lo_witness;
893
894 mtx_lock_spin(&w_mtx);
895 MPASS(w->w_refcount > 0);
896 w->w_refcount--;
897
898 if (w->w_refcount == 0)
899 depart(w);
900 mtx_unlock_spin(&w_mtx);
901 }
902
903 #ifdef DDB
904 static void
905 witness_ddb_compute_levels(void)
906 {
907 struct witness *w;
908
909 /*
910 * First clear all levels.
911 */
912 STAILQ_FOREACH(w, &w_all, w_list)
913 w->w_ddb_level = -1;
914
915 /*
916 * Look for locks with no parents and level all their descendants.
917 */
918 STAILQ_FOREACH(w, &w_all, w_list) {
919
920 /* If the witness has ancestors (is not a root), skip it. */
921 if (w->w_num_ancestors > 0)
922 continue;
923 witness_ddb_level_descendants(w, 0);
924 }
925 }
926
927 static void
928 witness_ddb_level_descendants(struct witness *w, int l)
929 {
930 int i;
931
932 if (w->w_ddb_level >= l)
933 return;
934
935 w->w_ddb_level = l;
936 l++;
937
938 for (i = 1; i <= w_max_used_index; i++) {
939 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
940 witness_ddb_level_descendants(&w_data[i], l);
941 }
942 }
943
944 static void
945 witness_ddb_display_descendants(int(*prnt)(const char *fmt, ...),
946 struct witness *w, int indent)
947 {
948 int i;
949
950 for (i = 0; i < indent; i++)
951 prnt(" ");
952 prnt("%s (type: %s, depth: %d, active refs: %d)",
953 w->w_name, w->w_class->lc_name,
954 w->w_ddb_level, w->w_refcount);
955 if (w->w_displayed) {
956 prnt(" -- (already displayed)\n");
957 return;
958 }
959 w->w_displayed = 1;
960 if (w->w_file != NULL && w->w_line != 0)
961 prnt(" -- last acquired @ %s:%d\n", fixup_filename(w->w_file),
962 w->w_line);
963 else
964 prnt(" -- never acquired\n");
965 indent++;
966 WITNESS_INDEX_ASSERT(w->w_index);
967 for (i = 1; i <= w_max_used_index; i++) {
968 if (db_pager_quit)
969 return;
970 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
971 witness_ddb_display_descendants(prnt, &w_data[i],
972 indent);
973 }
974 }
975
976 static void
977 witness_ddb_display_list(int(*prnt)(const char *fmt, ...),
978 struct witness_list *list)
979 {
980 struct witness *w;
981
982 STAILQ_FOREACH(w, list, w_typelist) {
983 if (w->w_file == NULL || w->w_ddb_level > 0)
984 continue;
985
986 /* This lock has no anscestors - display its descendants. */
987 witness_ddb_display_descendants(prnt, w, 0);
988 if (db_pager_quit)
989 return;
990 }
991 }
992
993 static void
994 witness_ddb_display(int(*prnt)(const char *fmt, ...))
995 {
996 struct witness *w;
997
998 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
999 witness_ddb_compute_levels();
1000
1001 /* Clear all the displayed flags. */
1002 STAILQ_FOREACH(w, &w_all, w_list)
1003 w->w_displayed = 0;
1004
1005 /*
1006 * First, handle sleep locks which have been acquired at least
1007 * once.
1008 */
1009 prnt("Sleep locks:\n");
1010 witness_ddb_display_list(prnt, &w_sleep);
1011 if (db_pager_quit)
1012 return;
1013
1014 /*
1015 * Now do spin locks which have been acquired at least once.
1016 */
1017 prnt("\nSpin locks:\n");
1018 witness_ddb_display_list(prnt, &w_spin);
1019 if (db_pager_quit)
1020 return;
1021
1022 /*
1023 * Finally, any locks which have not been acquired yet.
1024 */
1025 prnt("\nLocks which were never acquired:\n");
1026 STAILQ_FOREACH(w, &w_all, w_list) {
1027 if (w->w_file != NULL || w->w_refcount == 0)
1028 continue;
1029 prnt("%s (type: %s, depth: %d)\n", w->w_name,
1030 w->w_class->lc_name, w->w_ddb_level);
1031 if (db_pager_quit)
1032 return;
1033 }
1034 }
1035 #endif /* DDB */
1036
1037 int
1038 witness_defineorder(struct lock_object *lock1, struct lock_object *lock2)
1039 {
1040
1041 if (witness_watch == -1 || panicstr != NULL)
1042 return (0);
1043
1044 /* Require locks that witness knows about. */
1045 if (lock1 == NULL || lock1->lo_witness == NULL || lock2 == NULL ||
1046 lock2->lo_witness == NULL)
1047 return (EINVAL);
1048
1049 mtx_assert(&w_mtx, MA_NOTOWNED);
1050 mtx_lock_spin(&w_mtx);
1051
1052 /*
1053 * If we already have either an explicit or implied lock order that
1054 * is the other way around, then return an error.
1055 */
1056 if (witness_watch &&
1057 isitmydescendant(lock2->lo_witness, lock1->lo_witness)) {
1058 mtx_unlock_spin(&w_mtx);
1059 return (EDOOFUS);
1060 }
1061
1062 /* Try to add the new order. */
1063 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1064 lock2->lo_witness->w_name, lock1->lo_witness->w_name);
1065 itismychild(lock1->lo_witness, lock2->lo_witness);
1066 mtx_unlock_spin(&w_mtx);
1067 return (0);
1068 }
1069
1070 void
1071 witness_checkorder(struct lock_object *lock, int flags, const char *file,
1072 int line, struct lock_object *interlock)
1073 {
1074 struct lock_list_entry *lock_list, *lle;
1075 struct lock_instance *lock1, *lock2, *plock;
1076 struct lock_class *class, *iclass;
1077 struct witness *w, *w1;
1078 struct thread *td;
1079 int i, j;
1080
1081 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL ||
1082 panicstr != NULL)
1083 return;
1084
1085 w = lock->lo_witness;
1086 class = LOCK_CLASS(lock);
1087 td = curthread;
1088
1089 if (class->lc_flags & LC_SLEEPLOCK) {
1090
1091 /*
1092 * Since spin locks include a critical section, this check
1093 * implicitly enforces a lock order of all sleep locks before
1094 * all spin locks.
1095 */
1096 if (td->td_critnest != 0 && !kdb_active)
1097 kassert_panic("acquiring blockable sleep lock with "
1098 "spinlock or critical section held (%s) %s @ %s:%d",
1099 class->lc_name, lock->lo_name,
1100 fixup_filename(file), line);
1101
1102 /*
1103 * If this is the first lock acquired then just return as
1104 * no order checking is needed.
1105 */
1106 lock_list = td->td_sleeplocks;
1107 if (lock_list == NULL || lock_list->ll_count == 0)
1108 return;
1109 } else {
1110
1111 /*
1112 * If this is the first lock, just return as no order
1113 * checking is needed. Avoid problems with thread
1114 * migration pinning the thread while checking if
1115 * spinlocks are held. If at least one spinlock is held
1116 * the thread is in a safe path and it is allowed to
1117 * unpin it.
1118 */
1119 sched_pin();
1120 lock_list = PCPU_GET(spinlocks);
1121 if (lock_list == NULL || lock_list->ll_count == 0) {
1122 sched_unpin();
1123 return;
1124 }
1125 sched_unpin();
1126 }
1127
1128 /*
1129 * Check to see if we are recursing on a lock we already own. If
1130 * so, make sure that we don't mismatch exclusive and shared lock
1131 * acquires.
1132 */
1133 lock1 = find_instance(lock_list, lock);
1134 if (lock1 != NULL) {
1135 if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
1136 (flags & LOP_EXCLUSIVE) == 0) {
1137 witness_output("shared lock of (%s) %s @ %s:%d\n",
1138 class->lc_name, lock->lo_name,
1139 fixup_filename(file), line);
1140 witness_output("while exclusively locked from %s:%d\n",
1141 fixup_filename(lock1->li_file), lock1->li_line);
1142 kassert_panic("excl->share");
1143 }
1144 if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
1145 (flags & LOP_EXCLUSIVE) != 0) {
1146 witness_output("exclusive lock of (%s) %s @ %s:%d\n",
1147 class->lc_name, lock->lo_name,
1148 fixup_filename(file), line);
1149 witness_output("while share locked from %s:%d\n",
1150 fixup_filename(lock1->li_file), lock1->li_line);
1151 kassert_panic("share->excl");
1152 }
1153 return;
1154 }
1155
1156 /* Warn if the interlock is not locked exactly once. */
1157 if (interlock != NULL) {
1158 iclass = LOCK_CLASS(interlock);
1159 lock1 = find_instance(lock_list, interlock);
1160 if (lock1 == NULL)
1161 kassert_panic("interlock (%s) %s not locked @ %s:%d",
1162 iclass->lc_name, interlock->lo_name,
1163 fixup_filename(file), line);
1164 else if ((lock1->li_flags & LI_RECURSEMASK) != 0)
1165 kassert_panic("interlock (%s) %s recursed @ %s:%d",
1166 iclass->lc_name, interlock->lo_name,
1167 fixup_filename(file), line);
1168 }
1169
1170 /*
1171 * Find the previously acquired lock, but ignore interlocks.
1172 */
1173 plock = &lock_list->ll_children[lock_list->ll_count - 1];
1174 if (interlock != NULL && plock->li_lock == interlock) {
1175 if (lock_list->ll_count > 1)
1176 plock =
1177 &lock_list->ll_children[lock_list->ll_count - 2];
1178 else {
1179 lle = lock_list->ll_next;
1180
1181 /*
1182 * The interlock is the only lock we hold, so
1183 * simply return.
1184 */
1185 if (lle == NULL)
1186 return;
1187 plock = &lle->ll_children[lle->ll_count - 1];
1188 }
1189 }
1190
1191 /*
1192 * Try to perform most checks without a lock. If this succeeds we
1193 * can skip acquiring the lock and return success. Otherwise we redo
1194 * the check with the lock held to handle races with concurrent updates.
1195 */
1196 w1 = plock->li_lock->lo_witness;
1197 if (witness_lock_order_check(w1, w))
1198 return;
1199
1200 mtx_lock_spin(&w_mtx);
1201 if (witness_lock_order_check(w1, w)) {
1202 mtx_unlock_spin(&w_mtx);
1203 return;
1204 }
1205 witness_lock_order_add(w1, w);
1206
1207 /*
1208 * Check for duplicate locks of the same type. Note that we only
1209 * have to check for this on the last lock we just acquired. Any
1210 * other cases will be caught as lock order violations.
1211 */
1212 if (w1 == w) {
1213 i = w->w_index;
1214 if (!(lock->lo_flags & LO_DUPOK) && !(flags & LOP_DUPOK) &&
1215 !(w_rmatrix[i][i] & WITNESS_REVERSAL)) {
1216 w_rmatrix[i][i] |= WITNESS_REVERSAL;
1217 w->w_reversed = 1;
1218 mtx_unlock_spin(&w_mtx);
1219 witness_output(
1220 "acquiring duplicate lock of same type: \"%s\"\n",
1221 w->w_name);
1222 witness_output(" 1st %s @ %s:%d\n", plock->li_lock->lo_name,
1223 fixup_filename(plock->li_file), plock->li_line);
1224 witness_output(" 2nd %s @ %s:%d\n", lock->lo_name,
1225 fixup_filename(file), line);
1226 witness_debugger(1, __func__);
1227 } else
1228 mtx_unlock_spin(&w_mtx);
1229 return;
1230 }
1231 mtx_assert(&w_mtx, MA_OWNED);
1232
1233 /*
1234 * If we know that the lock we are acquiring comes after
1235 * the lock we most recently acquired in the lock order tree,
1236 * then there is no need for any further checks.
1237 */
1238 if (isitmychild(w1, w))
1239 goto out;
1240
1241 for (j = 0, lle = lock_list; lle != NULL; lle = lle->ll_next) {
1242 for (i = lle->ll_count - 1; i >= 0; i--, j++) {
1243
1244 MPASS(j < LOCK_CHILDCOUNT * LOCK_NCHILDREN);
1245 lock1 = &lle->ll_children[i];
1246
1247 /*
1248 * Ignore the interlock.
1249 */
1250 if (interlock == lock1->li_lock)
1251 continue;
1252
1253 /*
1254 * If this lock doesn't undergo witness checking,
1255 * then skip it.
1256 */
1257 w1 = lock1->li_lock->lo_witness;
1258 if (w1 == NULL) {
1259 KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
1260 ("lock missing witness structure"));
1261 continue;
1262 }
1263
1264 /*
1265 * If we are locking Giant and this is a sleepable
1266 * lock, then skip it.
1267 */
1268 if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0 &&
1269 lock == &Giant.lock_object)
1270 continue;
1271
1272 /*
1273 * If we are locking a sleepable lock and this lock
1274 * is Giant, then skip it.
1275 */
1276 if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1277 lock1->li_lock == &Giant.lock_object)
1278 continue;
1279
1280 /*
1281 * If we are locking a sleepable lock and this lock
1282 * isn't sleepable, we want to treat it as a lock
1283 * order violation to enfore a general lock order of
1284 * sleepable locks before non-sleepable locks.
1285 */
1286 if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1287 (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
1288 goto reversal;
1289
1290 /*
1291 * If we are locking Giant and this is a non-sleepable
1292 * lock, then treat it as a reversal.
1293 */
1294 if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0 &&
1295 lock == &Giant.lock_object)
1296 goto reversal;
1297
1298 /*
1299 * Check the lock order hierarchy for a reveresal.
1300 */
1301 if (!isitmydescendant(w, w1))
1302 continue;
1303 reversal:
1304
1305 /*
1306 * We have a lock order violation, check to see if it
1307 * is allowed or has already been yelled about.
1308 */
1309 #ifdef BLESSING
1310
1311 /*
1312 * If the lock order is blessed, just bail. We don't
1313 * look for other lock order violations though, which
1314 * may be a bug.
1315 */
1316 if (blessed(w, w1))
1317 goto out;
1318 #endif
1319
1320 /* Bail if this violation is known */
1321 if (w_rmatrix[w1->w_index][w->w_index] & WITNESS_REVERSAL)
1322 goto out;
1323
1324 /* Record this as a violation */
1325 w_rmatrix[w1->w_index][w->w_index] |= WITNESS_REVERSAL;
1326 w_rmatrix[w->w_index][w1->w_index] |= WITNESS_REVERSAL;
1327 w->w_reversed = w1->w_reversed = 1;
1328 witness_increment_graph_generation();
1329 mtx_unlock_spin(&w_mtx);
1330
1331 #ifdef WITNESS_NO_VNODE
1332 /*
1333 * There are known LORs between VNODE locks. They are
1334 * not an indication of a bug. VNODE locks are flagged
1335 * as such (LO_IS_VNODE) and we don't yell if the LOR
1336 * is between 2 VNODE locks.
1337 */
1338 if ((lock->lo_flags & LO_IS_VNODE) != 0 &&
1339 (lock1->li_lock->lo_flags & LO_IS_VNODE) != 0)
1340 return;
1341 #endif
1342
1343 /*
1344 * Ok, yell about it.
1345 */
1346 if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1347 (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
1348 witness_output(
1349 "lock order reversal: (sleepable after non-sleepable)\n");
1350 else if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0
1351 && lock == &Giant.lock_object)
1352 witness_output(
1353 "lock order reversal: (Giant after non-sleepable)\n");
1354 else
1355 witness_output("lock order reversal:\n");
1356
1357 /*
1358 * Try to locate an earlier lock with
1359 * witness w in our list.
1360 */
1361 do {
1362 lock2 = &lle->ll_children[i];
1363 MPASS(lock2->li_lock != NULL);
1364 if (lock2->li_lock->lo_witness == w)
1365 break;
1366 if (i == 0 && lle->ll_next != NULL) {
1367 lle = lle->ll_next;
1368 i = lle->ll_count - 1;
1369 MPASS(i >= 0 && i < LOCK_NCHILDREN);
1370 } else
1371 i--;
1372 } while (i >= 0);
1373 if (i < 0) {
1374 witness_output(" 1st %p %s (%s) @ %s:%d\n",
1375 lock1->li_lock, lock1->li_lock->lo_name,
1376 w1->w_name, fixup_filename(lock1->li_file),
1377 lock1->li_line);
1378 witness_output(" 2nd %p %s (%s) @ %s:%d\n", lock,
1379 lock->lo_name, w->w_name,
1380 fixup_filename(file), line);
1381 } else {
1382 witness_output(" 1st %p %s (%s) @ %s:%d\n",
1383 lock2->li_lock, lock2->li_lock->lo_name,
1384 lock2->li_lock->lo_witness->w_name,
1385 fixup_filename(lock2->li_file),
1386 lock2->li_line);
1387 witness_output(" 2nd %p %s (%s) @ %s:%d\n",
1388 lock1->li_lock, lock1->li_lock->lo_name,
1389 w1->w_name, fixup_filename(lock1->li_file),
1390 lock1->li_line);
1391 witness_output(" 3rd %p %s (%s) @ %s:%d\n", lock,
1392 lock->lo_name, w->w_name,
1393 fixup_filename(file), line);
1394 }
1395 witness_debugger(1, __func__);
1396 return;
1397 }
1398 }
1399
1400 /*
1401 * If requested, build a new lock order. However, don't build a new
1402 * relationship between a sleepable lock and Giant if it is in the
1403 * wrong direction. The correct lock order is that sleepable locks
1404 * always come before Giant.
1405 */
1406 if (flags & LOP_NEWORDER &&
1407 !(plock->li_lock == &Giant.lock_object &&
1408 (lock->lo_flags & LO_SLEEPABLE) != 0)) {
1409 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1410 w->w_name, plock->li_lock->lo_witness->w_name);
1411 itismychild(plock->li_lock->lo_witness, w);
1412 }
1413 out:
1414 mtx_unlock_spin(&w_mtx);
1415 }
1416
1417 void
1418 witness_lock(struct lock_object *lock, int flags, const char *file, int line)
1419 {
1420 struct lock_list_entry **lock_list, *lle;
1421 struct lock_instance *instance;
1422 struct witness *w;
1423 struct thread *td;
1424
1425 if (witness_cold || witness_watch == -1 || lock->lo_witness == NULL ||
1426 panicstr != NULL)
1427 return;
1428 w = lock->lo_witness;
1429 td = curthread;
1430
1431 /* Determine lock list for this lock. */
1432 if (LOCK_CLASS(lock)->lc_flags & LC_SLEEPLOCK)
1433 lock_list = &td->td_sleeplocks;
1434 else
1435 lock_list = PCPU_PTR(spinlocks);
1436
1437 /* Check to see if we are recursing on a lock we already own. */
1438 instance = find_instance(*lock_list, lock);
1439 if (instance != NULL) {
1440 instance->li_flags++;
1441 CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
1442 td->td_proc->p_pid, lock->lo_name,
1443 instance->li_flags & LI_RECURSEMASK);
1444 instance->li_file = file;
1445 instance->li_line = line;
1446 return;
1447 }
1448
1449 /* Update per-witness last file and line acquire. */
1450 w->w_file = file;
1451 w->w_line = line;
1452
1453 /* Find the next open lock instance in the list and fill it. */
1454 lle = *lock_list;
1455 if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
1456 lle = witness_lock_list_get();
1457 if (lle == NULL)
1458 return;
1459 lle->ll_next = *lock_list;
1460 CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
1461 td->td_proc->p_pid, lle);
1462 *lock_list = lle;
1463 }
1464 instance = &lle->ll_children[lle->ll_count++];
1465 instance->li_lock = lock;
1466 instance->li_line = line;
1467 instance->li_file = file;
1468 if ((flags & LOP_EXCLUSIVE) != 0)
1469 instance->li_flags = LI_EXCLUSIVE;
1470 else
1471 instance->li_flags = 0;
1472 CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
1473 td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
1474 }
1475
1476 void
1477 witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
1478 {
1479 struct lock_instance *instance;
1480 struct lock_class *class;
1481
1482 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1483 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
1484 return;
1485 class = LOCK_CLASS(lock);
1486 if (witness_watch) {
1487 if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1488 kassert_panic(
1489 "upgrade of non-upgradable lock (%s) %s @ %s:%d",
1490 class->lc_name, lock->lo_name,
1491 fixup_filename(file), line);
1492 if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1493 kassert_panic(
1494 "upgrade of non-sleep lock (%s) %s @ %s:%d",
1495 class->lc_name, lock->lo_name,
1496 fixup_filename(file), line);
1497 }
1498 instance = find_instance(curthread->td_sleeplocks, lock);
1499 if (instance == NULL) {
1500 kassert_panic("upgrade of unlocked lock (%s) %s @ %s:%d",
1501 class->lc_name, lock->lo_name,
1502 fixup_filename(file), line);
1503 return;
1504 }
1505 if (witness_watch) {
1506 if ((instance->li_flags & LI_EXCLUSIVE) != 0)
1507 kassert_panic(
1508 "upgrade of exclusive lock (%s) %s @ %s:%d",
1509 class->lc_name, lock->lo_name,
1510 fixup_filename(file), line);
1511 if ((instance->li_flags & LI_RECURSEMASK) != 0)
1512 kassert_panic(
1513 "upgrade of recursed lock (%s) %s r=%d @ %s:%d",
1514 class->lc_name, lock->lo_name,
1515 instance->li_flags & LI_RECURSEMASK,
1516 fixup_filename(file), line);
1517 }
1518 instance->li_flags |= LI_EXCLUSIVE;
1519 }
1520
1521 void
1522 witness_downgrade(struct lock_object *lock, int flags, const char *file,
1523 int line)
1524 {
1525 struct lock_instance *instance;
1526 struct lock_class *class;
1527
1528 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1529 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
1530 return;
1531 class = LOCK_CLASS(lock);
1532 if (witness_watch) {
1533 if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1534 kassert_panic(
1535 "downgrade of non-upgradable lock (%s) %s @ %s:%d",
1536 class->lc_name, lock->lo_name,
1537 fixup_filename(file), line);
1538 if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1539 kassert_panic(
1540 "downgrade of non-sleep lock (%s) %s @ %s:%d",
1541 class->lc_name, lock->lo_name,
1542 fixup_filename(file), line);
1543 }
1544 instance = find_instance(curthread->td_sleeplocks, lock);
1545 if (instance == NULL) {
1546 kassert_panic("downgrade of unlocked lock (%s) %s @ %s:%d",
1547 class->lc_name, lock->lo_name,
1548 fixup_filename(file), line);
1549 return;
1550 }
1551 if (witness_watch) {
1552 if ((instance->li_flags & LI_EXCLUSIVE) == 0)
1553 kassert_panic(
1554 "downgrade of shared lock (%s) %s @ %s:%d",
1555 class->lc_name, lock->lo_name,
1556 fixup_filename(file), line);
1557 if ((instance->li_flags & LI_RECURSEMASK) != 0)
1558 kassert_panic(
1559 "downgrade of recursed lock (%s) %s r=%d @ %s:%d",
1560 class->lc_name, lock->lo_name,
1561 instance->li_flags & LI_RECURSEMASK,
1562 fixup_filename(file), line);
1563 }
1564 instance->li_flags &= ~LI_EXCLUSIVE;
1565 }
1566
1567 void
1568 witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
1569 {
1570 struct lock_list_entry **lock_list, *lle;
1571 struct lock_instance *instance;
1572 struct lock_class *class;
1573 struct thread *td;
1574 register_t s;
1575 int i, j;
1576
1577 if (witness_cold || lock->lo_witness == NULL || panicstr != NULL)
1578 return;
1579 td = curthread;
1580 class = LOCK_CLASS(lock);
1581
1582 /* Find lock instance associated with this lock. */
1583 if (class->lc_flags & LC_SLEEPLOCK)
1584 lock_list = &td->td_sleeplocks;
1585 else
1586 lock_list = PCPU_PTR(spinlocks);
1587 lle = *lock_list;
1588 for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
1589 for (i = 0; i < (*lock_list)->ll_count; i++) {
1590 instance = &(*lock_list)->ll_children[i];
1591 if (instance->li_lock == lock)
1592 goto found;
1593 }
1594
1595 /*
1596 * When disabling WITNESS through witness_watch we could end up in
1597 * having registered locks in the td_sleeplocks queue.
1598 * We have to make sure we flush these queues, so just search for
1599 * eventual register locks and remove them.
1600 */
1601 if (witness_watch > 0) {
1602 kassert_panic("lock (%s) %s not locked @ %s:%d", class->lc_name,
1603 lock->lo_name, fixup_filename(file), line);
1604 return;
1605 } else {
1606 return;
1607 }
1608 found:
1609
1610 /* First, check for shared/exclusive mismatches. */
1611 if ((instance->li_flags & LI_EXCLUSIVE) != 0 && witness_watch > 0 &&
1612 (flags & LOP_EXCLUSIVE) == 0) {
1613 witness_output("shared unlock of (%s) %s @ %s:%d\n",
1614 class->lc_name, lock->lo_name, fixup_filename(file), line);
1615 witness_output("while exclusively locked from %s:%d\n",
1616 fixup_filename(instance->li_file), instance->li_line);
1617 kassert_panic("excl->ushare");
1618 }
1619 if ((instance->li_flags & LI_EXCLUSIVE) == 0 && witness_watch > 0 &&
1620 (flags & LOP_EXCLUSIVE) != 0) {
1621 witness_output("exclusive unlock of (%s) %s @ %s:%d\n",
1622 class->lc_name, lock->lo_name, fixup_filename(file), line);
1623 witness_output("while share locked from %s:%d\n",
1624 fixup_filename(instance->li_file),
1625 instance->li_line);
1626 kassert_panic("share->uexcl");
1627 }
1628 /* If we are recursed, unrecurse. */
1629 if ((instance->li_flags & LI_RECURSEMASK) > 0) {
1630 CTR4(KTR_WITNESS, "%s: pid %d unrecursed on %s r=%d", __func__,
1631 td->td_proc->p_pid, instance->li_lock->lo_name,
1632 instance->li_flags);
1633 instance->li_flags--;
1634 return;
1635 }
1636 /* The lock is now being dropped, check for NORELEASE flag */
1637 if ((instance->li_flags & LI_NORELEASE) != 0 && witness_watch > 0) {
1638 witness_output("forbidden unlock of (%s) %s @ %s:%d\n",
1639 class->lc_name, lock->lo_name, fixup_filename(file), line);
1640 kassert_panic("lock marked norelease");
1641 }
1642
1643 /* Otherwise, remove this item from the list. */
1644 s = intr_disable();
1645 CTR4(KTR_WITNESS, "%s: pid %d removed %s from lle[%d]", __func__,
1646 td->td_proc->p_pid, instance->li_lock->lo_name,
1647 (*lock_list)->ll_count - 1);
1648 for (j = i; j < (*lock_list)->ll_count - 1; j++)
1649 (*lock_list)->ll_children[j] =
1650 (*lock_list)->ll_children[j + 1];
1651 (*lock_list)->ll_count--;
1652 intr_restore(s);
1653
1654 /*
1655 * In order to reduce contention on w_mtx, we want to keep always an
1656 * head object into lists so that frequent allocation from the
1657 * free witness pool (and subsequent locking) is avoided.
1658 * In order to maintain the current code simple, when the head
1659 * object is totally unloaded it means also that we do not have
1660 * further objects in the list, so the list ownership needs to be
1661 * hand over to another object if the current head needs to be freed.
1662 */
1663 if ((*lock_list)->ll_count == 0) {
1664 if (*lock_list == lle) {
1665 if (lle->ll_next == NULL)
1666 return;
1667 } else
1668 lle = *lock_list;
1669 *lock_list = lle->ll_next;
1670 CTR3(KTR_WITNESS, "%s: pid %d removed lle %p", __func__,
1671 td->td_proc->p_pid, lle);
1672 witness_lock_list_free(lle);
1673 }
1674 }
1675
1676 void
1677 witness_thread_exit(struct thread *td)
1678 {
1679 struct lock_list_entry *lle;
1680 int i, n;
1681
1682 lle = td->td_sleeplocks;
1683 if (lle == NULL || panicstr != NULL)
1684 return;
1685 if (lle->ll_count != 0) {
1686 for (n = 0; lle != NULL; lle = lle->ll_next)
1687 for (i = lle->ll_count - 1; i >= 0; i--) {
1688 if (n == 0)
1689 witness_output(
1690 "Thread %p exiting with the following locks held:\n", td);
1691 n++;
1692 witness_list_lock(&lle->ll_children[i],
1693 witness_output);
1694
1695 }
1696 kassert_panic(
1697 "Thread %p cannot exit while holding sleeplocks\n", td);
1698 }
1699 witness_lock_list_free(lle);
1700 }
1701
1702 /*
1703 * Warn if any locks other than 'lock' are held. Flags can be passed in to
1704 * exempt Giant and sleepable locks from the checks as well. If any
1705 * non-exempt locks are held, then a supplied message is printed to the
1706 * output channel along with a list of the offending locks. If indicated in the
1707 * flags then a failure results in a panic as well.
1708 */
1709 int
1710 witness_warn(int flags, struct lock_object *lock, const char *fmt, ...)
1711 {
1712 struct lock_list_entry *lock_list, *lle;
1713 struct lock_instance *lock1;
1714 struct thread *td;
1715 va_list ap;
1716 int i, n;
1717
1718 if (witness_cold || witness_watch < 1 || panicstr != NULL)
1719 return (0);
1720 n = 0;
1721 td = curthread;
1722 for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next)
1723 for (i = lle->ll_count - 1; i >= 0; i--) {
1724 lock1 = &lle->ll_children[i];
1725 if (lock1->li_lock == lock)
1726 continue;
1727 if (flags & WARN_GIANTOK &&
1728 lock1->li_lock == &Giant.lock_object)
1729 continue;
1730 if (flags & WARN_SLEEPOK &&
1731 (lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0)
1732 continue;
1733 if (n == 0) {
1734 va_start(ap, fmt);
1735 vprintf(fmt, ap);
1736 va_end(ap);
1737 printf(" with the following %slocks held:\n",
1738 (flags & WARN_SLEEPOK) != 0 ?
1739 "non-sleepable " : "");
1740 }
1741 n++;
1742 witness_list_lock(lock1, printf);
1743 }
1744
1745 /*
1746 * Pin the thread in order to avoid problems with thread migration.
1747 * Once that all verifies are passed about spinlocks ownership,
1748 * the thread is in a safe path and it can be unpinned.
1749 */
1750 sched_pin();
1751 lock_list = PCPU_GET(spinlocks);
1752 if (lock_list != NULL && lock_list->ll_count != 0) {
1753 sched_unpin();
1754
1755 /*
1756 * We should only have one spinlock and as long as
1757 * the flags cannot match for this locks class,
1758 * check if the first spinlock is the one curthread
1759 * should hold.
1760 */
1761 lock1 = &lock_list->ll_children[lock_list->ll_count - 1];
1762 if (lock_list->ll_count == 1 && lock_list->ll_next == NULL &&
1763 lock1->li_lock == lock && n == 0)
1764 return (0);
1765
1766 va_start(ap, fmt);
1767 vprintf(fmt, ap);
1768 va_end(ap);
1769 printf(" with the following %slocks held:\n",
1770 (flags & WARN_SLEEPOK) != 0 ? "non-sleepable " : "");
1771 n += witness_list_locks(&lock_list, printf);
1772 } else
1773 sched_unpin();
1774 if (flags & WARN_PANIC && n)
1775 kassert_panic("%s", __func__);
1776 else
1777 witness_debugger(n, __func__);
1778 return (n);
1779 }
1780
1781 const char *
1782 witness_file(struct lock_object *lock)
1783 {
1784 struct witness *w;
1785
1786 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1787 return ("?");
1788 w = lock->lo_witness;
1789 return (w->w_file);
1790 }
1791
1792 int
1793 witness_line(struct lock_object *lock)
1794 {
1795 struct witness *w;
1796
1797 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1798 return (0);
1799 w = lock->lo_witness;
1800 return (w->w_line);
1801 }
1802
1803 static struct witness *
1804 enroll(const char *description, struct lock_class *lock_class)
1805 {
1806 struct witness *w;
1807 struct witness_list *typelist;
1808
1809 MPASS(description != NULL);
1810
1811 if (witness_watch == -1 || panicstr != NULL)
1812 return (NULL);
1813 if ((lock_class->lc_flags & LC_SPINLOCK)) {
1814 if (witness_skipspin)
1815 return (NULL);
1816 else
1817 typelist = &w_spin;
1818 } else if ((lock_class->lc_flags & LC_SLEEPLOCK)) {
1819 typelist = &w_sleep;
1820 } else {
1821 kassert_panic("lock class %s is not sleep or spin",
1822 lock_class->lc_name);
1823 return (NULL);
1824 }
1825
1826 mtx_lock_spin(&w_mtx);
1827 w = witness_hash_get(description);
1828 if (w)
1829 goto found;
1830 if ((w = witness_get()) == NULL)
1831 return (NULL);
1832 MPASS(strlen(description) < MAX_W_NAME);
1833 strcpy(w->w_name, description);
1834 w->w_class = lock_class;
1835 w->w_refcount = 1;
1836 STAILQ_INSERT_HEAD(&w_all, w, w_list);
1837 if (lock_class->lc_flags & LC_SPINLOCK) {
1838 STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
1839 w_spin_cnt++;
1840 } else if (lock_class->lc_flags & LC_SLEEPLOCK) {
1841 STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
1842 w_sleep_cnt++;
1843 }
1844
1845 /* Insert new witness into the hash */
1846 witness_hash_put(w);
1847 witness_increment_graph_generation();
1848 mtx_unlock_spin(&w_mtx);
1849 return (w);
1850 found:
1851 w->w_refcount++;
1852 if (w->w_refcount == 1)
1853 w->w_class = lock_class;
1854 mtx_unlock_spin(&w_mtx);
1855 if (lock_class != w->w_class)
1856 kassert_panic(
1857 "lock (%s) %s does not match earlier (%s) lock",
1858 description, lock_class->lc_name,
1859 w->w_class->lc_name);
1860 return (w);
1861 }
1862
1863 static void
1864 depart(struct witness *w)
1865 {
1866 struct witness_list *list;
1867
1868 MPASS(w->w_refcount == 0);
1869 if (w->w_class->lc_flags & LC_SLEEPLOCK) {
1870 list = &w_sleep;
1871 w_sleep_cnt--;
1872 } else {
1873 list = &w_spin;
1874 w_spin_cnt--;
1875 }
1876 /*
1877 * Set file to NULL as it may point into a loadable module.
1878 */
1879 w->w_file = NULL;
1880 w->w_line = 0;
1881 witness_increment_graph_generation();
1882 }
1883
1884
1885 static void
1886 adopt(struct witness *parent, struct witness *child)
1887 {
1888 int pi, ci, i, j;
1889
1890 if (witness_cold == 0)
1891 mtx_assert(&w_mtx, MA_OWNED);
1892
1893 /* If the relationship is already known, there's no work to be done. */
1894 if (isitmychild(parent, child))
1895 return;
1896
1897 /* When the structure of the graph changes, bump up the generation. */
1898 witness_increment_graph_generation();
1899
1900 /*
1901 * The hard part ... create the direct relationship, then propagate all
1902 * indirect relationships.
1903 */
1904 pi = parent->w_index;
1905 ci = child->w_index;
1906 WITNESS_INDEX_ASSERT(pi);
1907 WITNESS_INDEX_ASSERT(ci);
1908 MPASS(pi != ci);
1909 w_rmatrix[pi][ci] |= WITNESS_PARENT;
1910 w_rmatrix[ci][pi] |= WITNESS_CHILD;
1911
1912 /*
1913 * If parent was not already an ancestor of child,
1914 * then we increment the descendant and ancestor counters.
1915 */
1916 if ((w_rmatrix[pi][ci] & WITNESS_ANCESTOR) == 0) {
1917 parent->w_num_descendants++;
1918 child->w_num_ancestors++;
1919 }
1920
1921 /*
1922 * Find each ancestor of 'pi'. Note that 'pi' itself is counted as
1923 * an ancestor of 'pi' during this loop.
1924 */
1925 for (i = 1; i <= w_max_used_index; i++) {
1926 if ((w_rmatrix[i][pi] & WITNESS_ANCESTOR_MASK) == 0 &&
1927 (i != pi))
1928 continue;
1929
1930 /* Find each descendant of 'i' and mark it as a descendant. */
1931 for (j = 1; j <= w_max_used_index; j++) {
1932
1933 /*
1934 * Skip children that are already marked as
1935 * descendants of 'i'.
1936 */
1937 if (w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK)
1938 continue;
1939
1940 /*
1941 * We are only interested in descendants of 'ci'. Note
1942 * that 'ci' itself is counted as a descendant of 'ci'.
1943 */
1944 if ((w_rmatrix[ci][j] & WITNESS_ANCESTOR_MASK) == 0 &&
1945 (j != ci))
1946 continue;
1947 w_rmatrix[i][j] |= WITNESS_ANCESTOR;
1948 w_rmatrix[j][i] |= WITNESS_DESCENDANT;
1949 w_data[i].w_num_descendants++;
1950 w_data[j].w_num_ancestors++;
1951
1952 /*
1953 * Make sure we aren't marking a node as both an
1954 * ancestor and descendant. We should have caught
1955 * this as a lock order reversal earlier.
1956 */
1957 if ((w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK) &&
1958 (w_rmatrix[i][j] & WITNESS_DESCENDANT_MASK)) {
1959 printf("witness rmatrix paradox! [%d][%d]=%d "
1960 "both ancestor and descendant\n",
1961 i, j, w_rmatrix[i][j]);
1962 kdb_backtrace();
1963 printf("Witness disabled.\n");
1964 witness_watch = -1;
1965 }
1966 if ((w_rmatrix[j][i] & WITNESS_ANCESTOR_MASK) &&
1967 (w_rmatrix[j][i] & WITNESS_DESCENDANT_MASK)) {
1968 printf("witness rmatrix paradox! [%d][%d]=%d "
1969 "both ancestor and descendant\n",
1970 j, i, w_rmatrix[j][i]);
1971 kdb_backtrace();
1972 printf("Witness disabled.\n");
1973 witness_watch = -1;
1974 }
1975 }
1976 }
1977 }
1978
1979 static void
1980 itismychild(struct witness *parent, struct witness *child)
1981 {
1982 int unlocked;
1983
1984 MPASS(child != NULL && parent != NULL);
1985 if (witness_cold == 0)
1986 mtx_assert(&w_mtx, MA_OWNED);
1987
1988 if (!witness_lock_type_equal(parent, child)) {
1989 if (witness_cold == 0) {
1990 unlocked = 1;
1991 mtx_unlock_spin(&w_mtx);
1992 } else {
1993 unlocked = 0;
1994 }
1995 kassert_panic(
1996 "%s: parent \"%s\" (%s) and child \"%s\" (%s) are not "
1997 "the same lock type", __func__, parent->w_name,
1998 parent->w_class->lc_name, child->w_name,
1999 child->w_class->lc_name);
2000 if (unlocked)
2001 mtx_lock_spin(&w_mtx);
2002 }
2003 adopt(parent, child);
2004 }
2005
2006 /*
2007 * Generic code for the isitmy*() functions. The rmask parameter is the
2008 * expected relationship of w1 to w2.
2009 */
2010 static int
2011 _isitmyx(struct witness *w1, struct witness *w2, int rmask, const char *fname)
2012 {
2013 unsigned char r1, r2;
2014 int i1, i2;
2015
2016 i1 = w1->w_index;
2017 i2 = w2->w_index;
2018 WITNESS_INDEX_ASSERT(i1);
2019 WITNESS_INDEX_ASSERT(i2);
2020 r1 = w_rmatrix[i1][i2] & WITNESS_RELATED_MASK;
2021 r2 = w_rmatrix[i2][i1] & WITNESS_RELATED_MASK;
2022
2023 /* The flags on one better be the inverse of the flags on the other */
2024 if (!((WITNESS_ATOD(r1) == r2 && WITNESS_DTOA(r2) == r1) ||
2025 (WITNESS_DTOA(r1) == r2 && WITNESS_ATOD(r2) == r1))) {
2026 /* Don't squawk if we're potentially racing with an update. */
2027 if (!mtx_owned(&w_mtx))
2028 return (0);
2029 printf("%s: rmatrix mismatch between %s (index %d) and %s "
2030 "(index %d): w_rmatrix[%d][%d] == %hhx but "
2031 "w_rmatrix[%d][%d] == %hhx\n",
2032 fname, w1->w_name, i1, w2->w_name, i2, i1, i2, r1,
2033 i2, i1, r2);
2034 kdb_backtrace();
2035 printf("Witness disabled.\n");
2036 witness_watch = -1;
2037 }
2038 return (r1 & rmask);
2039 }
2040
2041 /*
2042 * Checks if @child is a direct child of @parent.
2043 */
2044 static int
2045 isitmychild(struct witness *parent, struct witness *child)
2046 {
2047
2048 return (_isitmyx(parent, child, WITNESS_PARENT, __func__));
2049 }
2050
2051 /*
2052 * Checks if @descendant is a direct or inderect descendant of @ancestor.
2053 */
2054 static int
2055 isitmydescendant(struct witness *ancestor, struct witness *descendant)
2056 {
2057
2058 return (_isitmyx(ancestor, descendant, WITNESS_ANCESTOR_MASK,
2059 __func__));
2060 }
2061
2062 #ifdef BLESSING
2063 static int
2064 blessed(struct witness *w1, struct witness *w2)
2065 {
2066 int i;
2067 struct witness_blessed *b;
2068
2069 for (i = 0; i < nitems(blessed_list); i++) {
2070 b = &blessed_list[i];
2071 if (strcmp(w1->w_name, b->b_lock1) == 0) {
2072 if (strcmp(w2->w_name, b->b_lock2) == 0)
2073 return (1);
2074 continue;
2075 }
2076 if (strcmp(w1->w_name, b->b_lock2) == 0)
2077 if (strcmp(w2->w_name, b->b_lock1) == 0)
2078 return (1);
2079 }
2080 return (0);
2081 }
2082 #endif
2083
2084 static struct witness *
2085 witness_get(void)
2086 {
2087 struct witness *w;
2088 int index;
2089
2090 if (witness_cold == 0)
2091 mtx_assert(&w_mtx, MA_OWNED);
2092
2093 if (witness_watch == -1) {
2094 mtx_unlock_spin(&w_mtx);
2095 return (NULL);
2096 }
2097 if (STAILQ_EMPTY(&w_free)) {
2098 witness_watch = -1;
2099 mtx_unlock_spin(&w_mtx);
2100 printf("WITNESS: unable to allocate a new witness object\n");
2101 return (NULL);
2102 }
2103 w = STAILQ_FIRST(&w_free);
2104 STAILQ_REMOVE_HEAD(&w_free, w_list);
2105 w_free_cnt--;
2106 index = w->w_index;
2107 MPASS(index > 0 && index == w_max_used_index+1 &&
2108 index < witness_count);
2109 bzero(w, sizeof(*w));
2110 w->w_index = index;
2111 if (index > w_max_used_index)
2112 w_max_used_index = index;
2113 return (w);
2114 }
2115
2116 static void
2117 witness_free(struct witness *w)
2118 {
2119
2120 STAILQ_INSERT_HEAD(&w_free, w, w_list);
2121 w_free_cnt++;
2122 }
2123
2124 static struct lock_list_entry *
2125 witness_lock_list_get(void)
2126 {
2127 struct lock_list_entry *lle;
2128
2129 if (witness_watch == -1)
2130 return (NULL);
2131 mtx_lock_spin(&w_mtx);
2132 lle = w_lock_list_free;
2133 if (lle == NULL) {
2134 witness_watch = -1;
2135 mtx_unlock_spin(&w_mtx);
2136 printf("%s: witness exhausted\n", __func__);
2137 return (NULL);
2138 }
2139 w_lock_list_free = lle->ll_next;
2140 mtx_unlock_spin(&w_mtx);
2141 bzero(lle, sizeof(*lle));
2142 return (lle);
2143 }
2144
2145 static void
2146 witness_lock_list_free(struct lock_list_entry *lle)
2147 {
2148
2149 mtx_lock_spin(&w_mtx);
2150 lle->ll_next = w_lock_list_free;
2151 w_lock_list_free = lle;
2152 mtx_unlock_spin(&w_mtx);
2153 }
2154
2155 static struct lock_instance *
2156 find_instance(struct lock_list_entry *list, const struct lock_object *lock)
2157 {
2158 struct lock_list_entry *lle;
2159 struct lock_instance *instance;
2160 int i;
2161
2162 for (lle = list; lle != NULL; lle = lle->ll_next)
2163 for (i = lle->ll_count - 1; i >= 0; i--) {
2164 instance = &lle->ll_children[i];
2165 if (instance->li_lock == lock)
2166 return (instance);
2167 }
2168 return (NULL);
2169 }
2170
2171 static void
2172 witness_list_lock(struct lock_instance *instance,
2173 int (*prnt)(const char *fmt, ...))
2174 {
2175 struct lock_object *lock;
2176
2177 lock = instance->li_lock;
2178 prnt("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ?
2179 "exclusive" : "shared", LOCK_CLASS(lock)->lc_name, lock->lo_name);
2180 if (lock->lo_witness->w_name != lock->lo_name)
2181 prnt(" (%s)", lock->lo_witness->w_name);
2182 prnt(" r = %d (%p) locked @ %s:%d\n",
2183 instance->li_flags & LI_RECURSEMASK, lock,
2184 fixup_filename(instance->li_file), instance->li_line);
2185 }
2186
2187 static int
2188 witness_output(const char *fmt, ...)
2189 {
2190 va_list ap;
2191 int ret;
2192
2193 va_start(ap, fmt);
2194 ret = witness_voutput(fmt, ap);
2195 va_end(ap);
2196 return (ret);
2197 }
2198
2199 static int
2200 witness_voutput(const char *fmt, va_list ap)
2201 {
2202 int ret;
2203
2204 ret = 0;
2205 switch (witness_channel) {
2206 case WITNESS_CONSOLE:
2207 ret = vprintf(fmt, ap);
2208 break;
2209 case WITNESS_LOG:
2210 vlog(LOG_NOTICE, fmt, ap);
2211 break;
2212 case WITNESS_NONE:
2213 break;
2214 }
2215 return (ret);
2216 }
2217
2218 #ifdef DDB
2219 static int
2220 witness_thread_has_locks(struct thread *td)
2221 {
2222
2223 if (td->td_sleeplocks == NULL)
2224 return (0);
2225 return (td->td_sleeplocks->ll_count != 0);
2226 }
2227
2228 static int
2229 witness_proc_has_locks(struct proc *p)
2230 {
2231 struct thread *td;
2232
2233 FOREACH_THREAD_IN_PROC(p, td) {
2234 if (witness_thread_has_locks(td))
2235 return (1);
2236 }
2237 return (0);
2238 }
2239 #endif
2240
2241 int
2242 witness_list_locks(struct lock_list_entry **lock_list,
2243 int (*prnt)(const char *fmt, ...))
2244 {
2245 struct lock_list_entry *lle;
2246 int i, nheld;
2247
2248 nheld = 0;
2249 for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
2250 for (i = lle->ll_count - 1; i >= 0; i--) {
2251 witness_list_lock(&lle->ll_children[i], prnt);
2252 nheld++;
2253 }
2254 return (nheld);
2255 }
2256
2257 /*
2258 * This is a bit risky at best. We call this function when we have timed
2259 * out acquiring a spin lock, and we assume that the other CPU is stuck
2260 * with this lock held. So, we go groveling around in the other CPU's
2261 * per-cpu data to try to find the lock instance for this spin lock to
2262 * see when it was last acquired.
2263 */
2264 void
2265 witness_display_spinlock(struct lock_object *lock, struct thread *owner,
2266 int (*prnt)(const char *fmt, ...))
2267 {
2268 struct lock_instance *instance;
2269 struct pcpu *pc;
2270
2271 if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU)
2272 return;
2273 pc = pcpu_find(owner->td_oncpu);
2274 instance = find_instance(pc->pc_spinlocks, lock);
2275 if (instance != NULL)
2276 witness_list_lock(instance, prnt);
2277 }
2278
2279 void
2280 witness_save(struct lock_object *lock, const char **filep, int *linep)
2281 {
2282 struct lock_list_entry *lock_list;
2283 struct lock_instance *instance;
2284 struct lock_class *class;
2285
2286 /*
2287 * This function is used independently in locking code to deal with
2288 * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2289 * is gone.
2290 */
2291 if (SCHEDULER_STOPPED())
2292 return;
2293 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2294 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2295 return;
2296 class = LOCK_CLASS(lock);
2297 if (class->lc_flags & LC_SLEEPLOCK)
2298 lock_list = curthread->td_sleeplocks;
2299 else {
2300 if (witness_skipspin)
2301 return;
2302 lock_list = PCPU_GET(spinlocks);
2303 }
2304 instance = find_instance(lock_list, lock);
2305 if (instance == NULL) {
2306 kassert_panic("%s: lock (%s) %s not locked", __func__,
2307 class->lc_name, lock->lo_name);
2308 return;
2309 }
2310 *filep = instance->li_file;
2311 *linep = instance->li_line;
2312 }
2313
2314 void
2315 witness_restore(struct lock_object *lock, const char *file, int line)
2316 {
2317 struct lock_list_entry *lock_list;
2318 struct lock_instance *instance;
2319 struct lock_class *class;
2320
2321 /*
2322 * This function is used independently in locking code to deal with
2323 * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2324 * is gone.
2325 */
2326 if (SCHEDULER_STOPPED())
2327 return;
2328 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2329 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2330 return;
2331 class = LOCK_CLASS(lock);
2332 if (class->lc_flags & LC_SLEEPLOCK)
2333 lock_list = curthread->td_sleeplocks;
2334 else {
2335 if (witness_skipspin)
2336 return;
2337 lock_list = PCPU_GET(spinlocks);
2338 }
2339 instance = find_instance(lock_list, lock);
2340 if (instance == NULL)
2341 kassert_panic("%s: lock (%s) %s not locked", __func__,
2342 class->lc_name, lock->lo_name);
2343 lock->lo_witness->w_file = file;
2344 lock->lo_witness->w_line = line;
2345 if (instance == NULL)
2346 return;
2347 instance->li_file = file;
2348 instance->li_line = line;
2349 }
2350
2351 void
2352 witness_assert(const struct lock_object *lock, int flags, const char *file,
2353 int line)
2354 {
2355 #ifdef INVARIANT_SUPPORT
2356 struct lock_instance *instance;
2357 struct lock_class *class;
2358
2359 if (lock->lo_witness == NULL || witness_watch < 1 || panicstr != NULL)
2360 return;
2361 class = LOCK_CLASS(lock);
2362 if ((class->lc_flags & LC_SLEEPLOCK) != 0)
2363 instance = find_instance(curthread->td_sleeplocks, lock);
2364 else if ((class->lc_flags & LC_SPINLOCK) != 0)
2365 instance = find_instance(PCPU_GET(spinlocks), lock);
2366 else {
2367 kassert_panic("Lock (%s) %s is not sleep or spin!",
2368 class->lc_name, lock->lo_name);
2369 return;
2370 }
2371 switch (flags) {
2372 case LA_UNLOCKED:
2373 if (instance != NULL)
2374 kassert_panic("Lock (%s) %s locked @ %s:%d.",
2375 class->lc_name, lock->lo_name,
2376 fixup_filename(file), line);
2377 break;
2378 case LA_LOCKED:
2379 case LA_LOCKED | LA_RECURSED:
2380 case LA_LOCKED | LA_NOTRECURSED:
2381 case LA_SLOCKED:
2382 case LA_SLOCKED | LA_RECURSED:
2383 case LA_SLOCKED | LA_NOTRECURSED:
2384 case LA_XLOCKED:
2385 case LA_XLOCKED | LA_RECURSED:
2386 case LA_XLOCKED | LA_NOTRECURSED:
2387 if (instance == NULL) {
2388 kassert_panic("Lock (%s) %s not locked @ %s:%d.",
2389 class->lc_name, lock->lo_name,
2390 fixup_filename(file), line);
2391 break;
2392 }
2393 if ((flags & LA_XLOCKED) != 0 &&
2394 (instance->li_flags & LI_EXCLUSIVE) == 0)
2395 kassert_panic(
2396 "Lock (%s) %s not exclusively locked @ %s:%d.",
2397 class->lc_name, lock->lo_name,
2398 fixup_filename(file), line);
2399 if ((flags & LA_SLOCKED) != 0 &&
2400 (instance->li_flags & LI_EXCLUSIVE) != 0)
2401 kassert_panic(
2402 "Lock (%s) %s exclusively locked @ %s:%d.",
2403 class->lc_name, lock->lo_name,
2404 fixup_filename(file), line);
2405 if ((flags & LA_RECURSED) != 0 &&
2406 (instance->li_flags & LI_RECURSEMASK) == 0)
2407 kassert_panic("Lock (%s) %s not recursed @ %s:%d.",
2408 class->lc_name, lock->lo_name,
2409 fixup_filename(file), line);
2410 if ((flags & LA_NOTRECURSED) != 0 &&
2411 (instance->li_flags & LI_RECURSEMASK) != 0)
2412 kassert_panic("Lock (%s) %s recursed @ %s:%d.",
2413 class->lc_name, lock->lo_name,
2414 fixup_filename(file), line);
2415 break;
2416 default:
2417 kassert_panic("Invalid lock assertion at %s:%d.",
2418 fixup_filename(file), line);
2419
2420 }
2421 #endif /* INVARIANT_SUPPORT */
2422 }
2423
2424 static void
2425 witness_setflag(struct lock_object *lock, int flag, int set)
2426 {
2427 struct lock_list_entry *lock_list;
2428 struct lock_instance *instance;
2429 struct lock_class *class;
2430
2431 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2432 return;
2433 class = LOCK_CLASS(lock);
2434 if (class->lc_flags & LC_SLEEPLOCK)
2435 lock_list = curthread->td_sleeplocks;
2436 else {
2437 if (witness_skipspin)
2438 return;
2439 lock_list = PCPU_GET(spinlocks);
2440 }
2441 instance = find_instance(lock_list, lock);
2442 if (instance == NULL) {
2443 kassert_panic("%s: lock (%s) %s not locked", __func__,
2444 class->lc_name, lock->lo_name);
2445 return;
2446 }
2447
2448 if (set)
2449 instance->li_flags |= flag;
2450 else
2451 instance->li_flags &= ~flag;
2452 }
2453
2454 void
2455 witness_norelease(struct lock_object *lock)
2456 {
2457
2458 witness_setflag(lock, LI_NORELEASE, 1);
2459 }
2460
2461 void
2462 witness_releaseok(struct lock_object *lock)
2463 {
2464
2465 witness_setflag(lock, LI_NORELEASE, 0);
2466 }
2467
2468 #ifdef DDB
2469 static void
2470 witness_ddb_list(struct thread *td)
2471 {
2472
2473 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2474 KASSERT(kdb_active, ("%s: not in the debugger", __func__));
2475
2476 if (witness_watch < 1)
2477 return;
2478
2479 witness_list_locks(&td->td_sleeplocks, db_printf);
2480
2481 /*
2482 * We only handle spinlocks if td == curthread. This is somewhat broken
2483 * if td is currently executing on some other CPU and holds spin locks
2484 * as we won't display those locks. If we had a MI way of getting
2485 * the per-cpu data for a given cpu then we could use
2486 * td->td_oncpu to get the list of spinlocks for this thread
2487 * and "fix" this.
2488 *
2489 * That still wouldn't really fix this unless we locked the scheduler
2490 * lock or stopped the other CPU to make sure it wasn't changing the
2491 * list out from under us. It is probably best to just not try to
2492 * handle threads on other CPU's for now.
2493 */
2494 if (td == curthread && PCPU_GET(spinlocks) != NULL)
2495 witness_list_locks(PCPU_PTR(spinlocks), db_printf);
2496 }
2497
2498 DB_SHOW_COMMAND(locks, db_witness_list)
2499 {
2500 struct thread *td;
2501
2502 if (have_addr)
2503 td = db_lookup_thread(addr, true);
2504 else
2505 td = kdb_thread;
2506 witness_ddb_list(td);
2507 }
2508
2509 DB_SHOW_ALL_COMMAND(locks, db_witness_list_all)
2510 {
2511 struct thread *td;
2512 struct proc *p;
2513
2514 /*
2515 * It would be nice to list only threads and processes that actually
2516 * held sleep locks, but that information is currently not exported
2517 * by WITNESS.
2518 */
2519 FOREACH_PROC_IN_SYSTEM(p) {
2520 if (!witness_proc_has_locks(p))
2521 continue;
2522 FOREACH_THREAD_IN_PROC(p, td) {
2523 if (!witness_thread_has_locks(td))
2524 continue;
2525 db_printf("Process %d (%s) thread %p (%d)\n", p->p_pid,
2526 p->p_comm, td, td->td_tid);
2527 witness_ddb_list(td);
2528 if (db_pager_quit)
2529 return;
2530 }
2531 }
2532 }
2533 DB_SHOW_ALIAS(alllocks, db_witness_list_all)
2534
2535 DB_SHOW_COMMAND(witness, db_witness_display)
2536 {
2537
2538 witness_ddb_display(db_printf);
2539 }
2540 #endif
2541
2542 static int
2543 sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS)
2544 {
2545 struct witness_lock_order_data *data1, *data2, *tmp_data1, *tmp_data2;
2546 struct witness *tmp_w1, *tmp_w2, *w1, *w2;
2547 struct sbuf *sb;
2548 u_int w_rmatrix1, w_rmatrix2;
2549 int error, generation, i, j;
2550
2551 tmp_data1 = NULL;
2552 tmp_data2 = NULL;
2553 tmp_w1 = NULL;
2554 tmp_w2 = NULL;
2555 if (witness_watch < 1) {
2556 error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2557 return (error);
2558 }
2559 if (witness_cold) {
2560 error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2561 return (error);
2562 }
2563 error = 0;
2564 sb = sbuf_new(NULL, NULL, badstack_sbuf_size, SBUF_AUTOEXTEND);
2565 if (sb == NULL)
2566 return (ENOMEM);
2567
2568 /* Allocate and init temporary storage space. */
2569 tmp_w1 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2570 tmp_w2 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2571 tmp_data1 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2572 M_WAITOK | M_ZERO);
2573 tmp_data2 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2574 M_WAITOK | M_ZERO);
2575 stack_zero(&tmp_data1->wlod_stack);
2576 stack_zero(&tmp_data2->wlod_stack);
2577
2578 restart:
2579 mtx_lock_spin(&w_mtx);
2580 generation = w_generation;
2581 mtx_unlock_spin(&w_mtx);
2582 sbuf_printf(sb, "Number of known direct relationships is %d\n",
2583 w_lohash.wloh_count);
2584 for (i = 1; i < w_max_used_index; i++) {
2585 mtx_lock_spin(&w_mtx);
2586 if (generation != w_generation) {
2587 mtx_unlock_spin(&w_mtx);
2588
2589 /* The graph has changed, try again. */
2590 req->oldidx = 0;
2591 sbuf_clear(sb);
2592 goto restart;
2593 }
2594
2595 w1 = &w_data[i];
2596 if (w1->w_reversed == 0) {
2597 mtx_unlock_spin(&w_mtx);
2598 continue;
2599 }
2600
2601 /* Copy w1 locally so we can release the spin lock. */
2602 *tmp_w1 = *w1;
2603 mtx_unlock_spin(&w_mtx);
2604
2605 if (tmp_w1->w_reversed == 0)
2606 continue;
2607 for (j = 1; j < w_max_used_index; j++) {
2608 if ((w_rmatrix[i][j] & WITNESS_REVERSAL) == 0 || i > j)
2609 continue;
2610
2611 mtx_lock_spin(&w_mtx);
2612 if (generation != w_generation) {
2613 mtx_unlock_spin(&w_mtx);
2614
2615 /* The graph has changed, try again. */
2616 req->oldidx = 0;
2617 sbuf_clear(sb);
2618 goto restart;
2619 }
2620
2621 w2 = &w_data[j];
2622 data1 = witness_lock_order_get(w1, w2);
2623 data2 = witness_lock_order_get(w2, w1);
2624
2625 /*
2626 * Copy information locally so we can release the
2627 * spin lock.
2628 */
2629 *tmp_w2 = *w2;
2630 w_rmatrix1 = (unsigned int)w_rmatrix[i][j];
2631 w_rmatrix2 = (unsigned int)w_rmatrix[j][i];
2632
2633 if (data1) {
2634 stack_zero(&tmp_data1->wlod_stack);
2635 stack_copy(&data1->wlod_stack,
2636 &tmp_data1->wlod_stack);
2637 }
2638 if (data2 && data2 != data1) {
2639 stack_zero(&tmp_data2->wlod_stack);
2640 stack_copy(&data2->wlod_stack,
2641 &tmp_data2->wlod_stack);
2642 }
2643 mtx_unlock_spin(&w_mtx);
2644
2645 sbuf_printf(sb,
2646 "\nLock order reversal between \"%s\"(%s) and \"%s\"(%s)!\n",
2647 tmp_w1->w_name, tmp_w1->w_class->lc_name,
2648 tmp_w2->w_name, tmp_w2->w_class->lc_name);
2649 if (data1) {
2650 sbuf_printf(sb,
2651 "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2652 tmp_w1->w_name, tmp_w1->w_class->lc_name,
2653 tmp_w2->w_name, tmp_w2->w_class->lc_name);
2654 stack_sbuf_print(sb, &tmp_data1->wlod_stack);
2655 sbuf_printf(sb, "\n");
2656 }
2657 if (data2 && data2 != data1) {
2658 sbuf_printf(sb,
2659 "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2660 tmp_w2->w_name, tmp_w2->w_class->lc_name,
2661 tmp_w1->w_name, tmp_w1->w_class->lc_name);
2662 stack_sbuf_print(sb, &tmp_data2->wlod_stack);
2663 sbuf_printf(sb, "\n");
2664 }
2665 }
2666 }
2667 mtx_lock_spin(&w_mtx);
2668 if (generation != w_generation) {
2669 mtx_unlock_spin(&w_mtx);
2670
2671 /*
2672 * The graph changed while we were printing stack data,
2673 * try again.
2674 */
2675 req->oldidx = 0;
2676 sbuf_clear(sb);
2677 goto restart;
2678 }
2679 mtx_unlock_spin(&w_mtx);
2680
2681 /* Free temporary storage space. */
2682 free(tmp_data1, M_TEMP);
2683 free(tmp_data2, M_TEMP);
2684 free(tmp_w1, M_TEMP);
2685 free(tmp_w2, M_TEMP);
2686
2687 sbuf_finish(sb);
2688 error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1);
2689 sbuf_delete(sb);
2690
2691 return (error);
2692 }
2693
2694 static int
2695 sysctl_debug_witness_channel(SYSCTL_HANDLER_ARGS)
2696 {
2697 static const struct {
2698 enum witness_channel channel;
2699 const char *name;
2700 } channels[] = {
2701 { WITNESS_CONSOLE, "console" },
2702 { WITNESS_LOG, "log" },
2703 { WITNESS_NONE, "none" },
2704 };
2705 char buf[16];
2706 u_int i;
2707 int error;
2708
2709 buf[0] = '\0';
2710 for (i = 0; i < nitems(channels); i++)
2711 if (witness_channel == channels[i].channel) {
2712 snprintf(buf, sizeof(buf), "%s", channels[i].name);
2713 break;
2714 }
2715
2716 error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
2717 if (error != 0 || req->newptr == NULL)
2718 return (error);
2719
2720 error = EINVAL;
2721 for (i = 0; i < nitems(channels); i++)
2722 if (strcmp(channels[i].name, buf) == 0) {
2723 witness_channel = channels[i].channel;
2724 error = 0;
2725 break;
2726 }
2727 return (error);
2728 }
2729
2730 static int
2731 sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS)
2732 {
2733 struct witness *w;
2734 struct sbuf *sb;
2735 int error;
2736
2737 if (witness_watch < 1) {
2738 error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2739 return (error);
2740 }
2741 if (witness_cold) {
2742 error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2743 return (error);
2744 }
2745 error = 0;
2746
2747 error = sysctl_wire_old_buffer(req, 0);
2748 if (error != 0)
2749 return (error);
2750 sb = sbuf_new_for_sysctl(NULL, NULL, FULLGRAPH_SBUF_SIZE, req);
2751 if (sb == NULL)
2752 return (ENOMEM);
2753 sbuf_printf(sb, "\n");
2754
2755 mtx_lock_spin(&w_mtx);
2756 STAILQ_FOREACH(w, &w_all, w_list)
2757 w->w_displayed = 0;
2758 STAILQ_FOREACH(w, &w_all, w_list)
2759 witness_add_fullgraph(sb, w);
2760 mtx_unlock_spin(&w_mtx);
2761
2762 /*
2763 * Close the sbuf and return to userland.
2764 */
2765 error = sbuf_finish(sb);
2766 sbuf_delete(sb);
2767
2768 return (error);
2769 }
2770
2771 static int
2772 sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS)
2773 {
2774 int error, value;
2775
2776 value = witness_watch;
2777 error = sysctl_handle_int(oidp, &value, 0, req);
2778 if (error != 0 || req->newptr == NULL)
2779 return (error);
2780 if (value > 1 || value < -1 ||
2781 (witness_watch == -1 && value != witness_watch))
2782 return (EINVAL);
2783 witness_watch = value;
2784 return (0);
2785 }
2786
2787 static void
2788 witness_add_fullgraph(struct sbuf *sb, struct witness *w)
2789 {
2790 int i;
2791
2792 if (w->w_displayed != 0 || (w->w_file == NULL && w->w_line == 0))
2793 return;
2794 w->w_displayed = 1;
2795
2796 WITNESS_INDEX_ASSERT(w->w_index);
2797 for (i = 1; i <= w_max_used_index; i++) {
2798 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT) {
2799 sbuf_printf(sb, "\"%s\",\"%s\"\n", w->w_name,
2800 w_data[i].w_name);
2801 witness_add_fullgraph(sb, &w_data[i]);
2802 }
2803 }
2804 }
2805
2806 /*
2807 * A simple hash function. Takes a key pointer and a key size. If size == 0,
2808 * interprets the key as a string and reads until the null
2809 * terminator. Otherwise, reads the first size bytes. Returns an unsigned 32-bit
2810 * hash value computed from the key.
2811 */
2812 static uint32_t
2813 witness_hash_djb2(const uint8_t *key, uint32_t size)
2814 {
2815 unsigned int hash = 5381;
2816 int i;
2817
2818 /* hash = hash * 33 + key[i] */
2819 if (size)
2820 for (i = 0; i < size; i++)
2821 hash = ((hash << 5) + hash) + (unsigned int)key[i];
2822 else
2823 for (i = 0; key[i] != 0; i++)
2824 hash = ((hash << 5) + hash) + (unsigned int)key[i];
2825
2826 return (hash);
2827 }
2828
2829
2830 /*
2831 * Initializes the two witness hash tables. Called exactly once from
2832 * witness_initialize().
2833 */
2834 static void
2835 witness_init_hash_tables(void)
2836 {
2837 int i;
2838
2839 MPASS(witness_cold);
2840
2841 /* Initialize the hash tables. */
2842 for (i = 0; i < WITNESS_HASH_SIZE; i++)
2843 w_hash.wh_array[i] = NULL;
2844
2845 w_hash.wh_size = WITNESS_HASH_SIZE;
2846 w_hash.wh_count = 0;
2847
2848 /* Initialize the lock order data hash. */
2849 w_lofree = NULL;
2850 for (i = 0; i < WITNESS_LO_DATA_COUNT; i++) {
2851 memset(&w_lodata[i], 0, sizeof(w_lodata[i]));
2852 w_lodata[i].wlod_next = w_lofree;
2853 w_lofree = &w_lodata[i];
2854 }
2855 w_lohash.wloh_size = WITNESS_LO_HASH_SIZE;
2856 w_lohash.wloh_count = 0;
2857 for (i = 0; i < WITNESS_LO_HASH_SIZE; i++)
2858 w_lohash.wloh_array[i] = NULL;
2859 }
2860
2861 static struct witness *
2862 witness_hash_get(const char *key)
2863 {
2864 struct witness *w;
2865 uint32_t hash;
2866
2867 MPASS(key != NULL);
2868 if (witness_cold == 0)
2869 mtx_assert(&w_mtx, MA_OWNED);
2870 hash = witness_hash_djb2(key, 0) % w_hash.wh_size;
2871 w = w_hash.wh_array[hash];
2872 while (w != NULL) {
2873 if (strcmp(w->w_name, key) == 0)
2874 goto out;
2875 w = w->w_hash_next;
2876 }
2877
2878 out:
2879 return (w);
2880 }
2881
2882 static void
2883 witness_hash_put(struct witness *w)
2884 {
2885 uint32_t hash;
2886
2887 MPASS(w != NULL);
2888 MPASS(w->w_name != NULL);
2889 if (witness_cold == 0)
2890 mtx_assert(&w_mtx, MA_OWNED);
2891 KASSERT(witness_hash_get(w->w_name) == NULL,
2892 ("%s: trying to add a hash entry that already exists!", __func__));
2893 KASSERT(w->w_hash_next == NULL,
2894 ("%s: w->w_hash_next != NULL", __func__));
2895
2896 hash = witness_hash_djb2(w->w_name, 0) % w_hash.wh_size;
2897 w->w_hash_next = w_hash.wh_array[hash];
2898 w_hash.wh_array[hash] = w;
2899 w_hash.wh_count++;
2900 }
2901
2902
2903 static struct witness_lock_order_data *
2904 witness_lock_order_get(struct witness *parent, struct witness *child)
2905 {
2906 struct witness_lock_order_data *data = NULL;
2907 struct witness_lock_order_key key;
2908 unsigned int hash;
2909
2910 MPASS(parent != NULL && child != NULL);
2911 key.from = parent->w_index;
2912 key.to = child->w_index;
2913 WITNESS_INDEX_ASSERT(key.from);
2914 WITNESS_INDEX_ASSERT(key.to);
2915 if ((w_rmatrix[parent->w_index][child->w_index]
2916 & WITNESS_LOCK_ORDER_KNOWN) == 0)
2917 goto out;
2918
2919 hash = witness_hash_djb2((const char*)&key,
2920 sizeof(key)) % w_lohash.wloh_size;
2921 data = w_lohash.wloh_array[hash];
2922 while (data != NULL) {
2923 if (witness_lock_order_key_equal(&data->wlod_key, &key))
2924 break;
2925 data = data->wlod_next;
2926 }
2927
2928 out:
2929 return (data);
2930 }
2931
2932 /*
2933 * Verify that parent and child have a known relationship, are not the same,
2934 * and child is actually a child of parent. This is done without w_mtx
2935 * to avoid contention in the common case.
2936 */
2937 static int
2938 witness_lock_order_check(struct witness *parent, struct witness *child)
2939 {
2940
2941 if (parent != child &&
2942 w_rmatrix[parent->w_index][child->w_index]
2943 & WITNESS_LOCK_ORDER_KNOWN &&
2944 isitmychild(parent, child))
2945 return (1);
2946
2947 return (0);
2948 }
2949
2950 static int
2951 witness_lock_order_add(struct witness *parent, struct witness *child)
2952 {
2953 struct witness_lock_order_data *data = NULL;
2954 struct witness_lock_order_key key;
2955 unsigned int hash;
2956
2957 MPASS(parent != NULL && child != NULL);
2958 key.from = parent->w_index;
2959 key.to = child->w_index;
2960 WITNESS_INDEX_ASSERT(key.from);
2961 WITNESS_INDEX_ASSERT(key.to);
2962 if (w_rmatrix[parent->w_index][child->w_index]
2963 & WITNESS_LOCK_ORDER_KNOWN)
2964 return (1);
2965
2966 hash = witness_hash_djb2((const char*)&key,
2967 sizeof(key)) % w_lohash.wloh_size;
2968 w_rmatrix[parent->w_index][child->w_index] |= WITNESS_LOCK_ORDER_KNOWN;
2969 data = w_lofree;
2970 if (data == NULL)
2971 return (0);
2972 w_lofree = data->wlod_next;
2973 data->wlod_next = w_lohash.wloh_array[hash];
2974 data->wlod_key = key;
2975 w_lohash.wloh_array[hash] = data;
2976 w_lohash.wloh_count++;
2977 stack_zero(&data->wlod_stack);
2978 stack_save(&data->wlod_stack);
2979 return (1);
2980 }
2981
2982 /* Call this whenever the structure of the witness graph changes. */
2983 static void
2984 witness_increment_graph_generation(void)
2985 {
2986
2987 if (witness_cold == 0)
2988 mtx_assert(&w_mtx, MA_OWNED);
2989 w_generation++;
2990 }
2991
2992 static int
2993 witness_output_drain(void *arg __unused, const char *data, int len)
2994 {
2995
2996 witness_output("%.*s", len, data);
2997 return (len);
2998 }
2999
3000 static void
3001 witness_debugger(int cond, const char *msg)
3002 {
3003 char buf[32];
3004 struct sbuf sb;
3005 struct stack st;
3006
3007 if (!cond)
3008 return;
3009
3010 if (witness_trace) {
3011 sbuf_new(&sb, buf, sizeof(buf), SBUF_FIXEDLEN);
3012 sbuf_set_drain(&sb, witness_output_drain, NULL);
3013
3014 stack_zero(&st);
3015 stack_save(&st);
3016 witness_output("stack backtrace:\n");
3017 stack_sbuf_print_ddb(&sb, &st);
3018
3019 sbuf_finish(&sb);
3020 }
3021
3022 #ifdef KDB
3023 if (witness_kdb)
3024 kdb_enter(KDB_WHY_WITNESS, msg);
3025 #endif
3026 }
Cache object: e66117eb7478582b844f234eb90bb9f1
|