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/systm.h>
110
111 #ifdef DDB
112 #include <ddb/ddb.h>
113 #endif
114
115 #include <machine/stdarg.h>
116
117 #if !defined(DDB) && !defined(STACK)
118 #error "DDB or STACK options are required for WITNESS"
119 #endif
120
121 /* Note that these traces do not work with KTR_ALQ. */
122 #if 0
123 #define KTR_WITNESS KTR_SUBSYS
124 #else
125 #define KTR_WITNESS 0
126 #endif
127
128 #define LI_RECURSEMASK 0x0000ffff /* Recursion depth of lock instance. */
129 #define LI_EXCLUSIVE 0x00010000 /* Exclusive lock instance. */
130 #define LI_NORELEASE 0x00020000 /* Lock not allowed to be released. */
131
132 /* Define this to check for blessed mutexes */
133 #undef BLESSING
134
135 #define WITNESS_COUNT 1024
136 #define WITNESS_CHILDCOUNT (WITNESS_COUNT * 4)
137 #define WITNESS_HASH_SIZE 251 /* Prime, gives load factor < 2 */
138 #define WITNESS_PENDLIST 768
139
140 /* Allocate 256 KB of stack data space */
141 #define WITNESS_LO_DATA_COUNT 2048
142
143 /* Prime, gives load factor of ~2 at full load */
144 #define WITNESS_LO_HASH_SIZE 1021
145
146 /*
147 * XXX: This is somewhat bogus, as we assume here that at most 2048 threads
148 * will hold LOCK_NCHILDREN locks. We handle failure ok, and we should
149 * probably be safe for the most part, but it's still a SWAG.
150 */
151 #define LOCK_NCHILDREN 5
152 #define LOCK_CHILDCOUNT 2048
153
154 #define MAX_W_NAME 64
155
156 #define BADSTACK_SBUF_SIZE (256 * WITNESS_COUNT)
157 #define CYCLEGRAPH_SBUF_SIZE 8192
158 #define FULLGRAPH_SBUF_SIZE 32768
159
160 /*
161 * These flags go in the witness relationship matrix and describe the
162 * relationship between any two struct witness objects.
163 */
164 #define WITNESS_UNRELATED 0x00 /* No lock order relation. */
165 #define WITNESS_PARENT 0x01 /* Parent, aka direct ancestor. */
166 #define WITNESS_ANCESTOR 0x02 /* Direct or indirect ancestor. */
167 #define WITNESS_CHILD 0x04 /* Child, aka direct descendant. */
168 #define WITNESS_DESCENDANT 0x08 /* Direct or indirect descendant. */
169 #define WITNESS_ANCESTOR_MASK (WITNESS_PARENT | WITNESS_ANCESTOR)
170 #define WITNESS_DESCENDANT_MASK (WITNESS_CHILD | WITNESS_DESCENDANT)
171 #define WITNESS_RELATED_MASK \
172 (WITNESS_ANCESTOR_MASK | WITNESS_DESCENDANT_MASK)
173 #define WITNESS_REVERSAL 0x10 /* A lock order reversal has been
174 * observed. */
175 #define WITNESS_RESERVED1 0x20 /* Unused flag, reserved. */
176 #define WITNESS_RESERVED2 0x40 /* Unused flag, reserved. */
177 #define WITNESS_LOCK_ORDER_KNOWN 0x80 /* This lock order is known. */
178
179 /* Descendant to ancestor flags */
180 #define WITNESS_DTOA(x) (((x) & WITNESS_RELATED_MASK) >> 2)
181
182 /* Ancestor to descendant flags */
183 #define WITNESS_ATOD(x) (((x) & WITNESS_RELATED_MASK) << 2)
184
185 #define WITNESS_INDEX_ASSERT(i) \
186 MPASS((i) > 0 && (i) <= w_max_used_index && (i) < WITNESS_COUNT)
187
188 MALLOC_DEFINE(M_WITNESS, "Witness", "Witness");
189
190 /*
191 * Lock instances. A lock instance is the data associated with a lock while
192 * it is held by witness. For example, a lock instance will hold the
193 * recursion count of a lock. Lock instances are held in lists. Spin locks
194 * are held in a per-cpu list while sleep locks are held in per-thread list.
195 */
196 struct lock_instance {
197 struct lock_object *li_lock;
198 const char *li_file;
199 int li_line;
200 u_int li_flags;
201 };
202
203 /*
204 * A simple list type used to build the list of locks held by a thread
205 * or CPU. We can't simply embed the list in struct lock_object since a
206 * lock may be held by more than one thread if it is a shared lock. Locks
207 * are added to the head of the list, so we fill up each list entry from
208 * "the back" logically. To ease some of the arithmetic, we actually fill
209 * in each list entry the normal way (children[0] then children[1], etc.) but
210 * when we traverse the list we read children[count-1] as the first entry
211 * down to children[0] as the final entry.
212 */
213 struct lock_list_entry {
214 struct lock_list_entry *ll_next;
215 struct lock_instance ll_children[LOCK_NCHILDREN];
216 u_int ll_count;
217 };
218
219 /*
220 * The main witness structure. One of these per named lock type in the system
221 * (for example, "vnode interlock").
222 */
223 struct witness {
224 char w_name[MAX_W_NAME];
225 uint32_t w_index; /* Index in the relationship matrix */
226 struct lock_class *w_class;
227 STAILQ_ENTRY(witness) w_list; /* List of all witnesses. */
228 STAILQ_ENTRY(witness) w_typelist; /* Witnesses of a type. */
229 struct witness *w_hash_next; /* Linked list in hash buckets. */
230 const char *w_file; /* File where last acquired */
231 uint32_t w_line; /* Line where last acquired */
232 uint32_t w_refcount;
233 uint16_t w_num_ancestors; /* direct/indirect
234 * ancestor count */
235 uint16_t w_num_descendants; /* direct/indirect
236 * descendant count */
237 int16_t w_ddb_level;
238 unsigned w_displayed:1;
239 unsigned w_reversed:1;
240 };
241
242 STAILQ_HEAD(witness_list, witness);
243
244 /*
245 * The witness hash table. Keys are witness names (const char *), elements are
246 * witness objects (struct witness *).
247 */
248 struct witness_hash {
249 struct witness *wh_array[WITNESS_HASH_SIZE];
250 uint32_t wh_size;
251 uint32_t wh_count;
252 };
253
254 /*
255 * Key type for the lock order data hash table.
256 */
257 struct witness_lock_order_key {
258 uint16_t from;
259 uint16_t to;
260 };
261
262 struct witness_lock_order_data {
263 struct stack wlod_stack;
264 struct witness_lock_order_key wlod_key;
265 struct witness_lock_order_data *wlod_next;
266 };
267
268 /*
269 * The witness lock order data hash table. Keys are witness index tuples
270 * (struct witness_lock_order_key), elements are lock order data objects
271 * (struct witness_lock_order_data).
272 */
273 struct witness_lock_order_hash {
274 struct witness_lock_order_data *wloh_array[WITNESS_LO_HASH_SIZE];
275 u_int wloh_size;
276 u_int wloh_count;
277 };
278
279 #ifdef BLESSING
280 struct witness_blessed {
281 const char *b_lock1;
282 const char *b_lock2;
283 };
284 #endif
285
286 struct witness_pendhelp {
287 const char *wh_type;
288 struct lock_object *wh_lock;
289 };
290
291 struct witness_order_list_entry {
292 const char *w_name;
293 struct lock_class *w_class;
294 };
295
296 /*
297 * Returns 0 if one of the locks is a spin lock and the other is not.
298 * Returns 1 otherwise.
299 */
300 static __inline int
301 witness_lock_type_equal(struct witness *w1, struct witness *w2)
302 {
303
304 return ((w1->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) ==
305 (w2->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)));
306 }
307
308 static __inline int
309 witness_lock_order_key_empty(const struct witness_lock_order_key *key)
310 {
311
312 return (key->from == 0 && key->to == 0);
313 }
314
315 static __inline int
316 witness_lock_order_key_equal(const struct witness_lock_order_key *a,
317 const struct witness_lock_order_key *b)
318 {
319
320 return (a->from == b->from && a->to == b->to);
321 }
322
323 static int _isitmyx(struct witness *w1, struct witness *w2, int rmask,
324 const char *fname);
325 #ifdef KDB
326 static void _witness_debugger(int cond, const char *msg);
327 #endif
328 static void adopt(struct witness *parent, struct witness *child);
329 #ifdef BLESSING
330 static int blessed(struct witness *, struct witness *);
331 #endif
332 static void depart(struct witness *w);
333 static struct witness *enroll(const char *description,
334 struct lock_class *lock_class);
335 static struct lock_instance *find_instance(struct lock_list_entry *list,
336 struct lock_object *lock);
337 static int isitmychild(struct witness *parent, struct witness *child);
338 static int isitmydescendant(struct witness *parent, struct witness *child);
339 static void itismychild(struct witness *parent, struct witness *child);
340 static int sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS);
341 static int sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS);
342 static int sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS);
343 static void witness_add_fullgraph(struct sbuf *sb, struct witness *parent);
344 #ifdef DDB
345 static void witness_ddb_compute_levels(void);
346 static void witness_ddb_display(int(*)(const char *fmt, ...));
347 static void witness_ddb_display_descendants(int(*)(const char *fmt, ...),
348 struct witness *, int indent);
349 static void witness_ddb_display_list(int(*prnt)(const char *fmt, ...),
350 struct witness_list *list);
351 static void witness_ddb_level_descendants(struct witness *parent, int l);
352 static void witness_ddb_list(struct thread *td);
353 #endif
354 static void witness_free(struct witness *m);
355 static struct witness *witness_get(void);
356 static uint32_t witness_hash_djb2(const uint8_t *key, uint32_t size);
357 static struct witness *witness_hash_get(const char *key);
358 static void witness_hash_put(struct witness *w);
359 static void witness_init_hash_tables(void);
360 static void witness_increment_graph_generation(void);
361 static void witness_lock_list_free(struct lock_list_entry *lle);
362 static struct lock_list_entry *witness_lock_list_get(void);
363 static int witness_lock_order_add(struct witness *parent,
364 struct witness *child);
365 static int witness_lock_order_check(struct witness *parent,
366 struct witness *child);
367 static struct witness_lock_order_data *witness_lock_order_get(
368 struct witness *parent,
369 struct witness *child);
370 static void witness_list_lock(struct lock_instance *instance,
371 int (*prnt)(const char *fmt, ...));
372 static void witness_setflag(struct lock_object *lock, int flag, int set);
373
374 #ifdef KDB
375 #define witness_debugger(c) _witness_debugger(c, __func__)
376 #else
377 #define witness_debugger(c)
378 #endif
379
380 SYSCTL_NODE(_debug, OID_AUTO, witness, CTLFLAG_RW, NULL, "Witness Locking");
381
382 /*
383 * If set to 0, lock order checking is disabled. If set to -1,
384 * witness is completely disabled. Otherwise witness performs full
385 * lock order checking for all locks. At runtime, lock order checking
386 * may be toggled. However, witness cannot be reenabled once it is
387 * completely disabled.
388 */
389 static int witness_watch = 1;
390 TUNABLE_INT("debug.witness.watch", &witness_watch);
391 SYSCTL_PROC(_debug_witness, OID_AUTO, watch, CTLFLAG_RW | CTLTYPE_INT, NULL, 0,
392 sysctl_debug_witness_watch, "I", "witness is watching lock operations");
393
394 #ifdef KDB
395 /*
396 * When KDB is enabled and witness_kdb is 1, it will cause the system
397 * to drop into kdebug() when:
398 * - a lock hierarchy violation occurs
399 * - locks are held when going to sleep.
400 */
401 #ifdef WITNESS_KDB
402 int witness_kdb = 1;
403 #else
404 int witness_kdb = 0;
405 #endif
406 TUNABLE_INT("debug.witness.kdb", &witness_kdb);
407 SYSCTL_INT(_debug_witness, OID_AUTO, kdb, CTLFLAG_RW, &witness_kdb, 0, "");
408
409 /*
410 * When KDB is enabled and witness_trace is 1, it will cause the system
411 * to print a stack trace:
412 * - a lock hierarchy violation occurs
413 * - locks are held when going to sleep.
414 */
415 int witness_trace = 1;
416 TUNABLE_INT("debug.witness.trace", &witness_trace);
417 SYSCTL_INT(_debug_witness, OID_AUTO, trace, CTLFLAG_RW, &witness_trace, 0, "");
418 #endif /* KDB */
419
420 #ifdef WITNESS_SKIPSPIN
421 int witness_skipspin = 1;
422 #else
423 int witness_skipspin = 0;
424 #endif
425 TUNABLE_INT("debug.witness.skipspin", &witness_skipspin);
426 SYSCTL_INT(_debug_witness, OID_AUTO, skipspin, CTLFLAG_RDTUN, &witness_skipspin,
427 0, "");
428
429 /*
430 * Call this to print out the relations between locks.
431 */
432 SYSCTL_PROC(_debug_witness, OID_AUTO, fullgraph, CTLTYPE_STRING | CTLFLAG_RD,
433 NULL, 0, sysctl_debug_witness_fullgraph, "A", "Show locks relation graphs");
434
435 /*
436 * Call this to print out the witness faulty stacks.
437 */
438 SYSCTL_PROC(_debug_witness, OID_AUTO, badstacks, CTLTYPE_STRING | CTLFLAG_RD,
439 NULL, 0, sysctl_debug_witness_badstacks, "A", "Show bad witness stacks");
440
441 static struct mtx w_mtx;
442
443 /* w_list */
444 static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free);
445 static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all);
446
447 /* w_typelist */
448 static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin);
449 static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep);
450
451 /* lock list */
452 static struct lock_list_entry *w_lock_list_free = NULL;
453 static struct witness_pendhelp pending_locks[WITNESS_PENDLIST];
454 static u_int pending_cnt;
455
456 static int w_free_cnt, w_spin_cnt, w_sleep_cnt;
457 SYSCTL_INT(_debug_witness, OID_AUTO, free_cnt, CTLFLAG_RD, &w_free_cnt, 0, "");
458 SYSCTL_INT(_debug_witness, OID_AUTO, spin_cnt, CTLFLAG_RD, &w_spin_cnt, 0, "");
459 SYSCTL_INT(_debug_witness, OID_AUTO, sleep_cnt, CTLFLAG_RD, &w_sleep_cnt, 0,
460 "");
461
462 static struct witness *w_data;
463 static uint8_t w_rmatrix[WITNESS_COUNT+1][WITNESS_COUNT+1];
464 static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT];
465 static struct witness_hash w_hash; /* The witness hash table. */
466
467 /* The lock order data hash */
468 static struct witness_lock_order_data w_lodata[WITNESS_LO_DATA_COUNT];
469 static struct witness_lock_order_data *w_lofree = NULL;
470 static struct witness_lock_order_hash w_lohash;
471 static int w_max_used_index = 0;
472 static unsigned int w_generation = 0;
473 static const char w_notrunning[] = "Witness not running\n";
474 static const char w_stillcold[] = "Witness is still cold\n";
475
476
477 static struct witness_order_list_entry order_lists[] = {
478 /*
479 * sx locks
480 */
481 { "proctree", &lock_class_sx },
482 { "allproc", &lock_class_sx },
483 { "allprison", &lock_class_sx },
484 { NULL, NULL },
485 /*
486 * Various mutexes
487 */
488 { "Giant", &lock_class_mtx_sleep },
489 { "pipe mutex", &lock_class_mtx_sleep },
490 { "sigio lock", &lock_class_mtx_sleep },
491 { "process group", &lock_class_mtx_sleep },
492 { "process lock", &lock_class_mtx_sleep },
493 { "session", &lock_class_mtx_sleep },
494 { "uidinfo hash", &lock_class_rw },
495 #ifdef HWPMC_HOOKS
496 { "pmc-sleep", &lock_class_mtx_sleep },
497 #endif
498 { NULL, NULL },
499 /*
500 * Sockets
501 */
502 { "accept", &lock_class_mtx_sleep },
503 { "so_snd", &lock_class_mtx_sleep },
504 { "so_rcv", &lock_class_mtx_sleep },
505 { "sellck", &lock_class_mtx_sleep },
506 { NULL, NULL },
507 /*
508 * Routing
509 */
510 { "so_rcv", &lock_class_mtx_sleep },
511 { "radix node head", &lock_class_rw },
512 { "rtentry", &lock_class_mtx_sleep },
513 { "ifaddr", &lock_class_mtx_sleep },
514 { NULL, NULL },
515 /*
516 * IPv4 multicast:
517 * protocol locks before interface locks, after UDP locks.
518 */
519 { "udpinp", &lock_class_rw },
520 { "in_multi_mtx", &lock_class_mtx_sleep },
521 { "igmp_mtx", &lock_class_mtx_sleep },
522 { "if_addr_mtx", &lock_class_mtx_sleep },
523 { NULL, NULL },
524 /*
525 * IPv6 multicast:
526 * protocol locks before interface locks, after UDP locks.
527 */
528 { "udpinp", &lock_class_rw },
529 { "in6_multi_mtx", &lock_class_mtx_sleep },
530 { "mld_mtx", &lock_class_mtx_sleep },
531 { "if_addr_mtx", &lock_class_mtx_sleep },
532 { NULL, NULL },
533 /*
534 * UNIX Domain Sockets
535 */
536 { "unp_global_rwlock", &lock_class_rw },
537 { "unp_list_lock", &lock_class_mtx_sleep },
538 { "unp", &lock_class_mtx_sleep },
539 { "so_snd", &lock_class_mtx_sleep },
540 { NULL, NULL },
541 /*
542 * UDP/IP
543 */
544 { "udp", &lock_class_rw },
545 { "udpinp", &lock_class_rw },
546 { "so_snd", &lock_class_mtx_sleep },
547 { NULL, NULL },
548 /*
549 * TCP/IP
550 */
551 { "tcp", &lock_class_rw },
552 { "tcpinp", &lock_class_rw },
553 { "so_snd", &lock_class_mtx_sleep },
554 { NULL, NULL },
555 /*
556 * SLIP
557 */
558 { "slip_mtx", &lock_class_mtx_sleep },
559 { "slip sc_mtx", &lock_class_mtx_sleep },
560 { NULL, NULL },
561 /*
562 * netatalk
563 */
564 { "ddp_list_mtx", &lock_class_mtx_sleep },
565 { "ddp_mtx", &lock_class_mtx_sleep },
566 { NULL, NULL },
567 /*
568 * BPF
569 */
570 { "bpf global lock", &lock_class_mtx_sleep },
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 { "system map", &lock_class_mtx_sleep },
602 { "vm page queue mutex", &lock_class_mtx_sleep },
603 { "vnode interlock", &lock_class_mtx_sleep },
604 { "cdev", &lock_class_mtx_sleep },
605 { NULL, NULL },
606 /*
607 * kqueue/VFS interaction
608 */
609 { "kqueue", &lock_class_mtx_sleep },
610 { "struct mount mtx", &lock_class_mtx_sleep },
611 { "vnode interlock", &lock_class_mtx_sleep },
612 { NULL, NULL },
613 /*
614 * ZFS locking
615 */
616 { "dn->dn_mtx", &lock_class_sx },
617 { "dr->dt.di.dr_mtx", &lock_class_sx },
618 { "db->db_mtx", &lock_class_sx },
619 { NULL, NULL },
620 /*
621 * spin locks
622 */
623 #ifdef SMP
624 { "ap boot", &lock_class_mtx_spin },
625 #endif
626 { "rm.mutex_mtx", &lock_class_mtx_spin },
627 { "sio", &lock_class_mtx_spin },
628 { "scrlock", &lock_class_mtx_spin },
629 #ifdef __i386__
630 { "cy", &lock_class_mtx_spin },
631 #endif
632 #ifdef __sparc64__
633 { "pcib_mtx", &lock_class_mtx_spin },
634 { "rtc_mtx", &lock_class_mtx_spin },
635 #endif
636 { "scc_hwmtx", &lock_class_mtx_spin },
637 { "uart_hwmtx", &lock_class_mtx_spin },
638 { "fast_taskqueue", &lock_class_mtx_spin },
639 { "intr table", &lock_class_mtx_spin },
640 #ifdef HWPMC_HOOKS
641 { "pmc-per-proc", &lock_class_mtx_spin },
642 #endif
643 { "process slock", &lock_class_mtx_spin },
644 { "sleepq chain", &lock_class_mtx_spin },
645 { "umtx lock", &lock_class_mtx_spin },
646 { "rm_spinlock", &lock_class_mtx_spin },
647 { "turnstile chain", &lock_class_mtx_spin },
648 { "turnstile lock", &lock_class_mtx_spin },
649 { "sched lock", &lock_class_mtx_spin },
650 { "td_contested", &lock_class_mtx_spin },
651 { "callout", &lock_class_mtx_spin },
652 { "entropy harvest mutex", &lock_class_mtx_spin },
653 { "syscons video lock", &lock_class_mtx_spin },
654 { "time lock", &lock_class_mtx_spin },
655 #ifdef SMP
656 { "smp rendezvous", &lock_class_mtx_spin },
657 #endif
658 #ifdef __powerpc__
659 { "tlb0", &lock_class_mtx_spin },
660 #endif
661 /*
662 * leaf locks
663 */
664 { "intrcnt", &lock_class_mtx_spin },
665 { "icu", &lock_class_mtx_spin },
666 #ifdef __i386__
667 { "allpmaps", &lock_class_mtx_spin },
668 { "descriptor tables", &lock_class_mtx_spin },
669 #endif
670 { "clk", &lock_class_mtx_spin },
671 { "cpuset", &lock_class_mtx_spin },
672 { "mprof lock", &lock_class_mtx_spin },
673 { "zombie lock", &lock_class_mtx_spin },
674 { "ALD Queue", &lock_class_mtx_spin },
675 #ifdef __ia64__
676 { "MCA spin lock", &lock_class_mtx_spin },
677 #endif
678 #if defined(__i386__) || defined(__amd64__)
679 { "pcicfg", &lock_class_mtx_spin },
680 { "NDIS thread lock", &lock_class_mtx_spin },
681 #endif
682 { "tw_osl_io_lock", &lock_class_mtx_spin },
683 { "tw_osl_q_lock", &lock_class_mtx_spin },
684 { "tw_cl_io_lock", &lock_class_mtx_spin },
685 { "tw_cl_intr_lock", &lock_class_mtx_spin },
686 { "tw_cl_gen_lock", &lock_class_mtx_spin },
687 #ifdef HWPMC_HOOKS
688 { "pmc-leaf", &lock_class_mtx_spin },
689 #endif
690 { "blocked lock", &lock_class_mtx_spin },
691 { NULL, NULL },
692 { NULL, NULL }
693 };
694
695 #ifdef BLESSING
696 /*
697 * Pairs of locks which have been blessed
698 * Don't complain about order problems with blessed locks
699 */
700 static struct witness_blessed blessed_list[] = {
701 };
702 static int blessed_count =
703 sizeof(blessed_list) / sizeof(struct witness_blessed);
704 #endif
705
706 /*
707 * This global is set to 0 once it becomes safe to use the witness code.
708 */
709 static int witness_cold = 1;
710
711 /*
712 * This global is set to 1 once the static lock orders have been enrolled
713 * so that a warning can be issued for any spin locks enrolled later.
714 */
715 static int witness_spin_warn = 0;
716
717 /*
718 * The WITNESS-enabled diagnostic code. Note that the witness code does
719 * assume that the early boot is single-threaded at least until after this
720 * routine is completed.
721 */
722 static void
723 witness_initialize(void *dummy __unused)
724 {
725 struct lock_object *lock;
726 struct witness_order_list_entry *order;
727 struct witness *w, *w1;
728 int i;
729
730 w_data = malloc(sizeof (struct witness) * WITNESS_COUNT, M_WITNESS,
731 M_NOWAIT | M_ZERO);
732
733 /*
734 * We have to release Giant before initializing its witness
735 * structure so that WITNESS doesn't get confused.
736 */
737 mtx_unlock(&Giant);
738 mtx_assert(&Giant, MA_NOTOWNED);
739
740 CTR1(KTR_WITNESS, "%s: initializing witness", __func__);
741 mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET |
742 MTX_NOWITNESS | MTX_NOPROFILE);
743 for (i = WITNESS_COUNT - 1; i >= 0; i--) {
744 w = &w_data[i];
745 memset(w, 0, sizeof(*w));
746 w_data[i].w_index = i; /* Witness index never changes. */
747 witness_free(w);
748 }
749 KASSERT(STAILQ_FIRST(&w_free)->w_index == 0,
750 ("%s: Invalid list of free witness objects", __func__));
751
752 /* Witness with index 0 is not used to aid in debugging. */
753 STAILQ_REMOVE_HEAD(&w_free, w_list);
754 w_free_cnt--;
755
756 memset(w_rmatrix, 0,
757 (sizeof(**w_rmatrix) * (WITNESS_COUNT+1) * (WITNESS_COUNT+1)));
758
759 for (i = 0; i < LOCK_CHILDCOUNT; i++)
760 witness_lock_list_free(&w_locklistdata[i]);
761 witness_init_hash_tables();
762
763 /* First add in all the specified order lists. */
764 for (order = order_lists; order->w_name != NULL; order++) {
765 w = enroll(order->w_name, order->w_class);
766 if (w == NULL)
767 continue;
768 w->w_file = "order list";
769 for (order++; order->w_name != NULL; order++) {
770 w1 = enroll(order->w_name, order->w_class);
771 if (w1 == NULL)
772 continue;
773 w1->w_file = "order list";
774 itismychild(w, w1);
775 w = w1;
776 }
777 }
778 witness_spin_warn = 1;
779
780 /* Iterate through all locks and add them to witness. */
781 for (i = 0; pending_locks[i].wh_lock != NULL; i++) {
782 lock = pending_locks[i].wh_lock;
783 KASSERT(lock->lo_flags & LO_WITNESS,
784 ("%s: lock %s is on pending list but not LO_WITNESS",
785 __func__, lock->lo_name));
786 lock->lo_witness = enroll(pending_locks[i].wh_type,
787 LOCK_CLASS(lock));
788 }
789
790 /* Mark the witness code as being ready for use. */
791 witness_cold = 0;
792
793 mtx_lock(&Giant);
794 }
795 SYSINIT(witness_init, SI_SUB_WITNESS, SI_ORDER_FIRST, witness_initialize,
796 NULL);
797
798 void
799 witness_init(struct lock_object *lock, const char *type)
800 {
801 struct lock_class *class;
802
803 /* Various sanity checks. */
804 class = LOCK_CLASS(lock);
805 if ((lock->lo_flags & LO_RECURSABLE) != 0 &&
806 (class->lc_flags & LC_RECURSABLE) == 0)
807 panic("%s: lock (%s) %s can not be recursable", __func__,
808 class->lc_name, lock->lo_name);
809 if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
810 (class->lc_flags & LC_SLEEPABLE) == 0)
811 panic("%s: lock (%s) %s can not be sleepable", __func__,
812 class->lc_name, lock->lo_name);
813 if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
814 (class->lc_flags & LC_UPGRADABLE) == 0)
815 panic("%s: lock (%s) %s can not be upgradable", __func__,
816 class->lc_name, lock->lo_name);
817
818 /*
819 * If we shouldn't watch this lock, then just clear lo_witness.
820 * Otherwise, if witness_cold is set, then it is too early to
821 * enroll this lock, so defer it to witness_initialize() by adding
822 * it to the pending_locks list. If it is not too early, then enroll
823 * the lock now.
824 */
825 if (witness_watch < 1 || panicstr != NULL ||
826 (lock->lo_flags & LO_WITNESS) == 0)
827 lock->lo_witness = NULL;
828 else if (witness_cold) {
829 pending_locks[pending_cnt].wh_lock = lock;
830 pending_locks[pending_cnt++].wh_type = type;
831 if (pending_cnt > WITNESS_PENDLIST)
832 panic("%s: pending locks list is too small, bump it\n",
833 __func__);
834 } else
835 lock->lo_witness = enroll(type, class);
836 }
837
838 void
839 witness_destroy(struct lock_object *lock)
840 {
841 struct lock_class *class;
842 struct witness *w;
843
844 class = LOCK_CLASS(lock);
845
846 if (witness_cold)
847 panic("lock (%s) %s destroyed while witness_cold",
848 class->lc_name, lock->lo_name);
849
850 /* XXX: need to verify that no one holds the lock */
851 if ((lock->lo_flags & LO_WITNESS) == 0 || lock->lo_witness == NULL)
852 return;
853 w = lock->lo_witness;
854
855 mtx_lock_spin(&w_mtx);
856 MPASS(w->w_refcount > 0);
857 w->w_refcount--;
858
859 if (w->w_refcount == 0)
860 depart(w);
861 mtx_unlock_spin(&w_mtx);
862 }
863
864 #ifdef DDB
865 static void
866 witness_ddb_compute_levels(void)
867 {
868 struct witness *w;
869
870 /*
871 * First clear all levels.
872 */
873 STAILQ_FOREACH(w, &w_all, w_list)
874 w->w_ddb_level = -1;
875
876 /*
877 * Look for locks with no parents and level all their descendants.
878 */
879 STAILQ_FOREACH(w, &w_all, w_list) {
880
881 /* If the witness has ancestors (is not a root), skip it. */
882 if (w->w_num_ancestors > 0)
883 continue;
884 witness_ddb_level_descendants(w, 0);
885 }
886 }
887
888 static void
889 witness_ddb_level_descendants(struct witness *w, int l)
890 {
891 int i;
892
893 if (w->w_ddb_level >= l)
894 return;
895
896 w->w_ddb_level = l;
897 l++;
898
899 for (i = 1; i <= w_max_used_index; i++) {
900 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
901 witness_ddb_level_descendants(&w_data[i], l);
902 }
903 }
904
905 static void
906 witness_ddb_display_descendants(int(*prnt)(const char *fmt, ...),
907 struct witness *w, int indent)
908 {
909 int i;
910
911 for (i = 0; i < indent; i++)
912 prnt(" ");
913 prnt("%s (type: %s, depth: %d, active refs: %d)",
914 w->w_name, w->w_class->lc_name,
915 w->w_ddb_level, w->w_refcount);
916 if (w->w_displayed) {
917 prnt(" -- (already displayed)\n");
918 return;
919 }
920 w->w_displayed = 1;
921 if (w->w_file != NULL && w->w_line != 0)
922 prnt(" -- last acquired @ %s:%d\n", w->w_file,
923 w->w_line);
924 else
925 prnt(" -- never acquired\n");
926 indent++;
927 WITNESS_INDEX_ASSERT(w->w_index);
928 for (i = 1; i <= w_max_used_index; i++) {
929 if (db_pager_quit)
930 return;
931 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
932 witness_ddb_display_descendants(prnt, &w_data[i],
933 indent);
934 }
935 }
936
937 static void
938 witness_ddb_display_list(int(*prnt)(const char *fmt, ...),
939 struct witness_list *list)
940 {
941 struct witness *w;
942
943 STAILQ_FOREACH(w, list, w_typelist) {
944 if (w->w_file == NULL || w->w_ddb_level > 0)
945 continue;
946
947 /* This lock has no anscestors - display its descendants. */
948 witness_ddb_display_descendants(prnt, w, 0);
949 if (db_pager_quit)
950 return;
951 }
952 }
953
954 static void
955 witness_ddb_display(int(*prnt)(const char *fmt, ...))
956 {
957 struct witness *w;
958
959 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
960 witness_ddb_compute_levels();
961
962 /* Clear all the displayed flags. */
963 STAILQ_FOREACH(w, &w_all, w_list)
964 w->w_displayed = 0;
965
966 /*
967 * First, handle sleep locks which have been acquired at least
968 * once.
969 */
970 prnt("Sleep locks:\n");
971 witness_ddb_display_list(prnt, &w_sleep);
972 if (db_pager_quit)
973 return;
974
975 /*
976 * Now do spin locks which have been acquired at least once.
977 */
978 prnt("\nSpin locks:\n");
979 witness_ddb_display_list(prnt, &w_spin);
980 if (db_pager_quit)
981 return;
982
983 /*
984 * Finally, any locks which have not been acquired yet.
985 */
986 prnt("\nLocks which were never acquired:\n");
987 STAILQ_FOREACH(w, &w_all, w_list) {
988 if (w->w_file != NULL || w->w_refcount == 0)
989 continue;
990 prnt("%s (type: %s, depth: %d)\n", w->w_name,
991 w->w_class->lc_name, w->w_ddb_level);
992 if (db_pager_quit)
993 return;
994 }
995 }
996 #endif /* DDB */
997
998 /* Trim useless garbage from filenames. */
999 static const char *
1000 fixup_filename(const char *file)
1001 {
1002
1003 if (file == NULL)
1004 return (NULL);
1005 while (strncmp(file, "../", 3) == 0)
1006 file += 3;
1007 return (file);
1008 }
1009
1010 int
1011 witness_defineorder(struct lock_object *lock1, struct lock_object *lock2)
1012 {
1013
1014 if (witness_watch == -1 || panicstr != NULL)
1015 return (0);
1016
1017 /* Require locks that witness knows about. */
1018 if (lock1 == NULL || lock1->lo_witness == NULL || lock2 == NULL ||
1019 lock2->lo_witness == NULL)
1020 return (EINVAL);
1021
1022 mtx_assert(&w_mtx, MA_NOTOWNED);
1023 mtx_lock_spin(&w_mtx);
1024
1025 /*
1026 * If we already have either an explicit or implied lock order that
1027 * is the other way around, then return an error.
1028 */
1029 if (witness_watch &&
1030 isitmydescendant(lock2->lo_witness, lock1->lo_witness)) {
1031 mtx_unlock_spin(&w_mtx);
1032 return (EDOOFUS);
1033 }
1034
1035 /* Try to add the new order. */
1036 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1037 lock2->lo_witness->w_name, lock1->lo_witness->w_name);
1038 itismychild(lock1->lo_witness, lock2->lo_witness);
1039 mtx_unlock_spin(&w_mtx);
1040 return (0);
1041 }
1042
1043 void
1044 witness_checkorder(struct lock_object *lock, int flags, const char *file,
1045 int line, struct lock_object *interlock)
1046 {
1047 struct lock_list_entry *lock_list, *lle;
1048 struct lock_instance *lock1, *lock2, *plock;
1049 struct lock_class *class;
1050 struct witness *w, *w1;
1051 struct thread *td;
1052 int i, j;
1053
1054 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL ||
1055 panicstr != NULL)
1056 return;
1057
1058 w = lock->lo_witness;
1059 class = LOCK_CLASS(lock);
1060 td = curthread;
1061 file = fixup_filename(file);
1062
1063 if (class->lc_flags & LC_SLEEPLOCK) {
1064
1065 /*
1066 * Since spin locks include a critical section, this check
1067 * implicitly enforces a lock order of all sleep locks before
1068 * all spin locks.
1069 */
1070 if (td->td_critnest != 0 && !kdb_active)
1071 panic("blockable sleep lock (%s) %s @ %s:%d",
1072 class->lc_name, lock->lo_name, file, line);
1073
1074 /*
1075 * If this is the first lock acquired then just return as
1076 * no order checking is needed.
1077 */
1078 lock_list = td->td_sleeplocks;
1079 if (lock_list == NULL || lock_list->ll_count == 0)
1080 return;
1081 } else {
1082
1083 /*
1084 * If this is the first lock, just return as no order
1085 * checking is needed. Avoid problems with thread
1086 * migration pinning the thread while checking if
1087 * spinlocks are held. If at least one spinlock is held
1088 * the thread is in a safe path and it is allowed to
1089 * unpin it.
1090 */
1091 sched_pin();
1092 lock_list = PCPU_GET(spinlocks);
1093 if (lock_list == NULL || lock_list->ll_count == 0) {
1094 sched_unpin();
1095 return;
1096 }
1097 sched_unpin();
1098 }
1099
1100 /*
1101 * Check to see if we are recursing on a lock we already own. If
1102 * so, make sure that we don't mismatch exclusive and shared lock
1103 * acquires.
1104 */
1105 lock1 = find_instance(lock_list, lock);
1106 if (lock1 != NULL) {
1107 if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
1108 (flags & LOP_EXCLUSIVE) == 0) {
1109 printf("shared lock of (%s) %s @ %s:%d\n",
1110 class->lc_name, lock->lo_name, file, line);
1111 printf("while exclusively locked from %s:%d\n",
1112 lock1->li_file, lock1->li_line);
1113 panic("share->excl");
1114 }
1115 if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
1116 (flags & LOP_EXCLUSIVE) != 0) {
1117 printf("exclusive lock of (%s) %s @ %s:%d\n",
1118 class->lc_name, lock->lo_name, file, line);
1119 printf("while share locked from %s:%d\n",
1120 lock1->li_file, lock1->li_line);
1121 panic("excl->share");
1122 }
1123 return;
1124 }
1125
1126 /*
1127 * Find the previously acquired lock, but ignore interlocks.
1128 */
1129 plock = &lock_list->ll_children[lock_list->ll_count - 1];
1130 if (interlock != NULL && plock->li_lock == interlock) {
1131 if (lock_list->ll_count > 1)
1132 plock =
1133 &lock_list->ll_children[lock_list->ll_count - 2];
1134 else {
1135 lle = lock_list->ll_next;
1136
1137 /*
1138 * The interlock is the only lock we hold, so
1139 * simply return.
1140 */
1141 if (lle == NULL)
1142 return;
1143 plock = &lle->ll_children[lle->ll_count - 1];
1144 }
1145 }
1146
1147 /*
1148 * Try to perform most checks without a lock. If this succeeds we
1149 * can skip acquiring the lock and return success.
1150 */
1151 w1 = plock->li_lock->lo_witness;
1152 if (witness_lock_order_check(w1, w))
1153 return;
1154
1155 /*
1156 * Check for duplicate locks of the same type. Note that we only
1157 * have to check for this on the last lock we just acquired. Any
1158 * other cases will be caught as lock order violations.
1159 */
1160 mtx_lock_spin(&w_mtx);
1161 witness_lock_order_add(w1, w);
1162 if (w1 == w) {
1163 i = w->w_index;
1164 if (!(lock->lo_flags & LO_DUPOK) && !(flags & LOP_DUPOK) &&
1165 !(w_rmatrix[i][i] & WITNESS_REVERSAL)) {
1166 w_rmatrix[i][i] |= WITNESS_REVERSAL;
1167 w->w_reversed = 1;
1168 mtx_unlock_spin(&w_mtx);
1169 printf(
1170 "acquiring duplicate lock of same type: \"%s\"\n",
1171 w->w_name);
1172 printf(" 1st %s @ %s:%d\n", plock->li_lock->lo_name,
1173 plock->li_file, plock->li_line);
1174 printf(" 2nd %s @ %s:%d\n", lock->lo_name, file, line);
1175 witness_debugger(1);
1176 } else
1177 mtx_unlock_spin(&w_mtx);
1178 return;
1179 }
1180 mtx_assert(&w_mtx, MA_OWNED);
1181
1182 /*
1183 * If we know that the lock we are acquiring comes after
1184 * the lock we most recently acquired in the lock order tree,
1185 * then there is no need for any further checks.
1186 */
1187 if (isitmychild(w1, w))
1188 goto out;
1189
1190 for (j = 0, lle = lock_list; lle != NULL; lle = lle->ll_next) {
1191 for (i = lle->ll_count - 1; i >= 0; i--, j++) {
1192
1193 MPASS(j < WITNESS_COUNT);
1194 lock1 = &lle->ll_children[i];
1195
1196 /*
1197 * Ignore the interlock the first time we see it.
1198 */
1199 if (interlock != NULL && interlock == lock1->li_lock) {
1200 interlock = NULL;
1201 continue;
1202 }
1203
1204 /*
1205 * If this lock doesn't undergo witness checking,
1206 * then skip it.
1207 */
1208 w1 = lock1->li_lock->lo_witness;
1209 if (w1 == NULL) {
1210 KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
1211 ("lock missing witness structure"));
1212 continue;
1213 }
1214
1215 /*
1216 * If we are locking Giant and this is a sleepable
1217 * lock, then skip it.
1218 */
1219 if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0 &&
1220 lock == &Giant.lock_object)
1221 continue;
1222
1223 /*
1224 * If we are locking a sleepable lock and this lock
1225 * is Giant, then skip it.
1226 */
1227 if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1228 lock1->li_lock == &Giant.lock_object)
1229 continue;
1230
1231 /*
1232 * If we are locking a sleepable lock and this lock
1233 * isn't sleepable, we want to treat it as a lock
1234 * order violation to enfore a general lock order of
1235 * sleepable locks before non-sleepable locks.
1236 */
1237 if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1238 (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
1239 goto reversal;
1240
1241 /*
1242 * If we are locking Giant and this is a non-sleepable
1243 * lock, then treat it as a reversal.
1244 */
1245 if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0 &&
1246 lock == &Giant.lock_object)
1247 goto reversal;
1248
1249 /*
1250 * Check the lock order hierarchy for a reveresal.
1251 */
1252 if (!isitmydescendant(w, w1))
1253 continue;
1254 reversal:
1255
1256 /*
1257 * We have a lock order violation, check to see if it
1258 * is allowed or has already been yelled about.
1259 */
1260 #ifdef BLESSING
1261
1262 /*
1263 * If the lock order is blessed, just bail. We don't
1264 * look for other lock order violations though, which
1265 * may be a bug.
1266 */
1267 if (blessed(w, w1))
1268 goto out;
1269 #endif
1270
1271 /* Bail if this violation is known */
1272 if (w_rmatrix[w1->w_index][w->w_index] & WITNESS_REVERSAL)
1273 goto out;
1274
1275 /* Record this as a violation */
1276 w_rmatrix[w1->w_index][w->w_index] |= WITNESS_REVERSAL;
1277 w_rmatrix[w->w_index][w1->w_index] |= WITNESS_REVERSAL;
1278 w->w_reversed = w1->w_reversed = 1;
1279 witness_increment_graph_generation();
1280 mtx_unlock_spin(&w_mtx);
1281
1282 /*
1283 * Ok, yell about it.
1284 */
1285 if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1286 (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
1287 printf(
1288 "lock order reversal: (sleepable after non-sleepable)\n");
1289 else if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0
1290 && lock == &Giant.lock_object)
1291 printf(
1292 "lock order reversal: (Giant after non-sleepable)\n");
1293 else
1294 printf("lock order reversal:\n");
1295
1296 /*
1297 * Try to locate an earlier lock with
1298 * witness w in our list.
1299 */
1300 do {
1301 lock2 = &lle->ll_children[i];
1302 MPASS(lock2->li_lock != NULL);
1303 if (lock2->li_lock->lo_witness == w)
1304 break;
1305 if (i == 0 && lle->ll_next != NULL) {
1306 lle = lle->ll_next;
1307 i = lle->ll_count - 1;
1308 MPASS(i >= 0 && i < LOCK_NCHILDREN);
1309 } else
1310 i--;
1311 } while (i >= 0);
1312 if (i < 0) {
1313 printf(" 1st %p %s (%s) @ %s:%d\n",
1314 lock1->li_lock, lock1->li_lock->lo_name,
1315 w1->w_name, lock1->li_file, lock1->li_line);
1316 printf(" 2nd %p %s (%s) @ %s:%d\n", lock,
1317 lock->lo_name, w->w_name, file, line);
1318 } else {
1319 printf(" 1st %p %s (%s) @ %s:%d\n",
1320 lock2->li_lock, lock2->li_lock->lo_name,
1321 lock2->li_lock->lo_witness->w_name,
1322 lock2->li_file, lock2->li_line);
1323 printf(" 2nd %p %s (%s) @ %s:%d\n",
1324 lock1->li_lock, lock1->li_lock->lo_name,
1325 w1->w_name, lock1->li_file, lock1->li_line);
1326 printf(" 3rd %p %s (%s) @ %s:%d\n", lock,
1327 lock->lo_name, w->w_name, file, line);
1328 }
1329 witness_debugger(1);
1330 return;
1331 }
1332 }
1333
1334 /*
1335 * If requested, build a new lock order. However, don't build a new
1336 * relationship between a sleepable lock and Giant if it is in the
1337 * wrong direction. The correct lock order is that sleepable locks
1338 * always come before Giant.
1339 */
1340 if (flags & LOP_NEWORDER &&
1341 !(plock->li_lock == &Giant.lock_object &&
1342 (lock->lo_flags & LO_SLEEPABLE) != 0)) {
1343 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1344 w->w_name, plock->li_lock->lo_witness->w_name);
1345 itismychild(plock->li_lock->lo_witness, w);
1346 }
1347 out:
1348 mtx_unlock_spin(&w_mtx);
1349 }
1350
1351 void
1352 witness_lock(struct lock_object *lock, int flags, const char *file, int line)
1353 {
1354 struct lock_list_entry **lock_list, *lle;
1355 struct lock_instance *instance;
1356 struct witness *w;
1357 struct thread *td;
1358
1359 if (witness_cold || witness_watch == -1 || lock->lo_witness == NULL ||
1360 panicstr != NULL)
1361 return;
1362 w = lock->lo_witness;
1363 td = curthread;
1364 file = fixup_filename(file);
1365
1366 /* Determine lock list for this lock. */
1367 if (LOCK_CLASS(lock)->lc_flags & LC_SLEEPLOCK)
1368 lock_list = &td->td_sleeplocks;
1369 else
1370 lock_list = PCPU_PTR(spinlocks);
1371
1372 /* Check to see if we are recursing on a lock we already own. */
1373 instance = find_instance(*lock_list, lock);
1374 if (instance != NULL) {
1375 instance->li_flags++;
1376 CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
1377 td->td_proc->p_pid, lock->lo_name,
1378 instance->li_flags & LI_RECURSEMASK);
1379 instance->li_file = file;
1380 instance->li_line = line;
1381 return;
1382 }
1383
1384 /* Update per-witness last file and line acquire. */
1385 w->w_file = file;
1386 w->w_line = line;
1387
1388 /* Find the next open lock instance in the list and fill it. */
1389 lle = *lock_list;
1390 if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
1391 lle = witness_lock_list_get();
1392 if (lle == NULL)
1393 return;
1394 lle->ll_next = *lock_list;
1395 CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
1396 td->td_proc->p_pid, lle);
1397 *lock_list = lle;
1398 }
1399 instance = &lle->ll_children[lle->ll_count++];
1400 instance->li_lock = lock;
1401 instance->li_line = line;
1402 instance->li_file = file;
1403 if ((flags & LOP_EXCLUSIVE) != 0)
1404 instance->li_flags = LI_EXCLUSIVE;
1405 else
1406 instance->li_flags = 0;
1407 CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
1408 td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
1409 }
1410
1411 void
1412 witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
1413 {
1414 struct lock_instance *instance;
1415 struct lock_class *class;
1416
1417 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1418 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
1419 return;
1420 class = LOCK_CLASS(lock);
1421 file = fixup_filename(file);
1422 if (witness_watch) {
1423 if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1424 panic("upgrade of non-upgradable lock (%s) %s @ %s:%d",
1425 class->lc_name, lock->lo_name, file, line);
1426 if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1427 panic("upgrade of non-sleep lock (%s) %s @ %s:%d",
1428 class->lc_name, lock->lo_name, file, line);
1429 }
1430 instance = find_instance(curthread->td_sleeplocks, lock);
1431 if (instance == NULL)
1432 panic("upgrade of unlocked lock (%s) %s @ %s:%d",
1433 class->lc_name, lock->lo_name, file, line);
1434 if (witness_watch) {
1435 if ((instance->li_flags & LI_EXCLUSIVE) != 0)
1436 panic("upgrade of exclusive lock (%s) %s @ %s:%d",
1437 class->lc_name, lock->lo_name, file, line);
1438 if ((instance->li_flags & LI_RECURSEMASK) != 0)
1439 panic("upgrade of recursed lock (%s) %s r=%d @ %s:%d",
1440 class->lc_name, lock->lo_name,
1441 instance->li_flags & LI_RECURSEMASK, file, line);
1442 }
1443 instance->li_flags |= LI_EXCLUSIVE;
1444 }
1445
1446 void
1447 witness_downgrade(struct lock_object *lock, int flags, const char *file,
1448 int line)
1449 {
1450 struct lock_instance *instance;
1451 struct lock_class *class;
1452
1453 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1454 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
1455 return;
1456 class = LOCK_CLASS(lock);
1457 file = fixup_filename(file);
1458 if (witness_watch) {
1459 if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1460 panic("downgrade of non-upgradable lock (%s) %s @ %s:%d",
1461 class->lc_name, lock->lo_name, file, line);
1462 if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1463 panic("downgrade of non-sleep lock (%s) %s @ %s:%d",
1464 class->lc_name, lock->lo_name, file, line);
1465 }
1466 instance = find_instance(curthread->td_sleeplocks, lock);
1467 if (instance == NULL)
1468 panic("downgrade of unlocked lock (%s) %s @ %s:%d",
1469 class->lc_name, lock->lo_name, file, line);
1470 if (witness_watch) {
1471 if ((instance->li_flags & LI_EXCLUSIVE) == 0)
1472 panic("downgrade of shared lock (%s) %s @ %s:%d",
1473 class->lc_name, lock->lo_name, file, line);
1474 if ((instance->li_flags & LI_RECURSEMASK) != 0)
1475 panic("downgrade of recursed lock (%s) %s r=%d @ %s:%d",
1476 class->lc_name, lock->lo_name,
1477 instance->li_flags & LI_RECURSEMASK, file, line);
1478 }
1479 instance->li_flags &= ~LI_EXCLUSIVE;
1480 }
1481
1482 void
1483 witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
1484 {
1485 struct lock_list_entry **lock_list, *lle;
1486 struct lock_instance *instance;
1487 struct lock_class *class;
1488 struct thread *td;
1489 register_t s;
1490 int i, j;
1491
1492 if (witness_cold || lock->lo_witness == NULL || panicstr != NULL)
1493 return;
1494 td = curthread;
1495 class = LOCK_CLASS(lock);
1496 file = fixup_filename(file);
1497
1498 /* Find lock instance associated with this lock. */
1499 if (class->lc_flags & LC_SLEEPLOCK)
1500 lock_list = &td->td_sleeplocks;
1501 else
1502 lock_list = PCPU_PTR(spinlocks);
1503 lle = *lock_list;
1504 for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
1505 for (i = 0; i < (*lock_list)->ll_count; i++) {
1506 instance = &(*lock_list)->ll_children[i];
1507 if (instance->li_lock == lock)
1508 goto found;
1509 }
1510
1511 /*
1512 * When disabling WITNESS through witness_watch we could end up in
1513 * having registered locks in the td_sleeplocks queue.
1514 * We have to make sure we flush these queues, so just search for
1515 * eventual register locks and remove them.
1516 */
1517 if (witness_watch > 0)
1518 panic("lock (%s) %s not locked @ %s:%d", class->lc_name,
1519 lock->lo_name, file, line);
1520 else
1521 return;
1522 found:
1523
1524 /* First, check for shared/exclusive mismatches. */
1525 if ((instance->li_flags & LI_EXCLUSIVE) != 0 && witness_watch > 0 &&
1526 (flags & LOP_EXCLUSIVE) == 0) {
1527 printf("shared unlock of (%s) %s @ %s:%d\n", class->lc_name,
1528 lock->lo_name, file, line);
1529 printf("while exclusively locked from %s:%d\n",
1530 instance->li_file, instance->li_line);
1531 panic("excl->ushare");
1532 }
1533 if ((instance->li_flags & LI_EXCLUSIVE) == 0 && witness_watch > 0 &&
1534 (flags & LOP_EXCLUSIVE) != 0) {
1535 printf("exclusive unlock of (%s) %s @ %s:%d\n", class->lc_name,
1536 lock->lo_name, file, line);
1537 printf("while share locked from %s:%d\n", instance->li_file,
1538 instance->li_line);
1539 panic("share->uexcl");
1540 }
1541 /* If we are recursed, unrecurse. */
1542 if ((instance->li_flags & LI_RECURSEMASK) > 0) {
1543 CTR4(KTR_WITNESS, "%s: pid %d unrecursed on %s r=%d", __func__,
1544 td->td_proc->p_pid, instance->li_lock->lo_name,
1545 instance->li_flags);
1546 instance->li_flags--;
1547 return;
1548 }
1549 /* The lock is now being dropped, check for NORELEASE flag */
1550 if ((instance->li_flags & LI_NORELEASE) != 0 && witness_watch > 0) {
1551 printf("forbidden unlock of (%s) %s @ %s:%d\n", class->lc_name,
1552 lock->lo_name, file, line);
1553 panic("lock marked norelease");
1554 }
1555
1556 /* Otherwise, remove this item from the list. */
1557 s = intr_disable();
1558 CTR4(KTR_WITNESS, "%s: pid %d removed %s from lle[%d]", __func__,
1559 td->td_proc->p_pid, instance->li_lock->lo_name,
1560 (*lock_list)->ll_count - 1);
1561 for (j = i; j < (*lock_list)->ll_count - 1; j++)
1562 (*lock_list)->ll_children[j] =
1563 (*lock_list)->ll_children[j + 1];
1564 (*lock_list)->ll_count--;
1565 intr_restore(s);
1566
1567 /*
1568 * In order to reduce contention on w_mtx, we want to keep always an
1569 * head object into lists so that frequent allocation from the
1570 * free witness pool (and subsequent locking) is avoided.
1571 * In order to maintain the current code simple, when the head
1572 * object is totally unloaded it means also that we do not have
1573 * further objects in the list, so the list ownership needs to be
1574 * hand over to another object if the current head needs to be freed.
1575 */
1576 if ((*lock_list)->ll_count == 0) {
1577 if (*lock_list == lle) {
1578 if (lle->ll_next == NULL)
1579 return;
1580 } else
1581 lle = *lock_list;
1582 *lock_list = lle->ll_next;
1583 CTR3(KTR_WITNESS, "%s: pid %d removed lle %p", __func__,
1584 td->td_proc->p_pid, lle);
1585 witness_lock_list_free(lle);
1586 }
1587 }
1588
1589 void
1590 witness_thread_exit(struct thread *td)
1591 {
1592 struct lock_list_entry *lle;
1593 int i, n;
1594
1595 lle = td->td_sleeplocks;
1596 if (lle == NULL || panicstr != NULL)
1597 return;
1598 if (lle->ll_count != 0) {
1599 for (n = 0; lle != NULL; lle = lle->ll_next)
1600 for (i = lle->ll_count - 1; i >= 0; i--) {
1601 if (n == 0)
1602 printf("Thread %p exiting with the following locks held:\n",
1603 td);
1604 n++;
1605 witness_list_lock(&lle->ll_children[i], printf);
1606
1607 }
1608 panic("Thread %p cannot exit while holding sleeplocks\n", td);
1609 }
1610 witness_lock_list_free(lle);
1611 }
1612
1613 /*
1614 * Warn if any locks other than 'lock' are held. Flags can be passed in to
1615 * exempt Giant and sleepable locks from the checks as well. If any
1616 * non-exempt locks are held, then a supplied message is printed to the
1617 * console along with a list of the offending locks. If indicated in the
1618 * flags then a failure results in a panic as well.
1619 */
1620 int
1621 witness_warn(int flags, struct lock_object *lock, const char *fmt, ...)
1622 {
1623 struct lock_list_entry *lock_list, *lle;
1624 struct lock_instance *lock1;
1625 struct thread *td;
1626 va_list ap;
1627 int i, n;
1628
1629 if (witness_cold || witness_watch < 1 || panicstr != NULL)
1630 return (0);
1631 n = 0;
1632 td = curthread;
1633 for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next)
1634 for (i = lle->ll_count - 1; i >= 0; i--) {
1635 lock1 = &lle->ll_children[i];
1636 if (lock1->li_lock == lock)
1637 continue;
1638 if (flags & WARN_GIANTOK &&
1639 lock1->li_lock == &Giant.lock_object)
1640 continue;
1641 if (flags & WARN_SLEEPOK &&
1642 (lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0)
1643 continue;
1644 if (n == 0) {
1645 va_start(ap, fmt);
1646 vprintf(fmt, ap);
1647 va_end(ap);
1648 printf(" with the following");
1649 if (flags & WARN_SLEEPOK)
1650 printf(" non-sleepable");
1651 printf(" locks held:\n");
1652 }
1653 n++;
1654 witness_list_lock(lock1, printf);
1655 }
1656
1657 /*
1658 * Pin the thread in order to avoid problems with thread migration.
1659 * Once that all verifies are passed about spinlocks ownership,
1660 * the thread is in a safe path and it can be unpinned.
1661 */
1662 sched_pin();
1663 lock_list = PCPU_GET(spinlocks);
1664 if (lock_list != NULL && lock_list->ll_count != 0) {
1665 sched_unpin();
1666
1667 /*
1668 * We should only have one spinlock and as long as
1669 * the flags cannot match for this locks class,
1670 * check if the first spinlock is the one curthread
1671 * should hold.
1672 */
1673 lock1 = &lock_list->ll_children[lock_list->ll_count - 1];
1674 if (lock_list->ll_count == 1 && lock_list->ll_next == NULL &&
1675 lock1->li_lock == lock && n == 0)
1676 return (0);
1677
1678 va_start(ap, fmt);
1679 vprintf(fmt, ap);
1680 va_end(ap);
1681 printf(" with the following");
1682 if (flags & WARN_SLEEPOK)
1683 printf(" non-sleepable");
1684 printf(" locks held:\n");
1685 n += witness_list_locks(&lock_list, printf);
1686 } else
1687 sched_unpin();
1688 if (flags & WARN_PANIC && n)
1689 panic("%s", __func__);
1690 else
1691 witness_debugger(n);
1692 return (n);
1693 }
1694
1695 const char *
1696 witness_file(struct lock_object *lock)
1697 {
1698 struct witness *w;
1699
1700 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1701 return ("?");
1702 w = lock->lo_witness;
1703 return (w->w_file);
1704 }
1705
1706 int
1707 witness_line(struct lock_object *lock)
1708 {
1709 struct witness *w;
1710
1711 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1712 return (0);
1713 w = lock->lo_witness;
1714 return (w->w_line);
1715 }
1716
1717 static struct witness *
1718 enroll(const char *description, struct lock_class *lock_class)
1719 {
1720 struct witness *w;
1721 struct witness_list *typelist;
1722
1723 MPASS(description != NULL);
1724
1725 if (witness_watch == -1 || panicstr != NULL)
1726 return (NULL);
1727 if ((lock_class->lc_flags & LC_SPINLOCK)) {
1728 if (witness_skipspin)
1729 return (NULL);
1730 else
1731 typelist = &w_spin;
1732 } else if ((lock_class->lc_flags & LC_SLEEPLOCK))
1733 typelist = &w_sleep;
1734 else
1735 panic("lock class %s is not sleep or spin",
1736 lock_class->lc_name);
1737
1738 mtx_lock_spin(&w_mtx);
1739 w = witness_hash_get(description);
1740 if (w)
1741 goto found;
1742 if ((w = witness_get()) == NULL)
1743 return (NULL);
1744 MPASS(strlen(description) < MAX_W_NAME);
1745 strcpy(w->w_name, description);
1746 w->w_class = lock_class;
1747 w->w_refcount = 1;
1748 STAILQ_INSERT_HEAD(&w_all, w, w_list);
1749 if (lock_class->lc_flags & LC_SPINLOCK) {
1750 STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
1751 w_spin_cnt++;
1752 } else if (lock_class->lc_flags & LC_SLEEPLOCK) {
1753 STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
1754 w_sleep_cnt++;
1755 }
1756
1757 /* Insert new witness into the hash */
1758 witness_hash_put(w);
1759 witness_increment_graph_generation();
1760 mtx_unlock_spin(&w_mtx);
1761 return (w);
1762 found:
1763 w->w_refcount++;
1764 mtx_unlock_spin(&w_mtx);
1765 if (lock_class != w->w_class)
1766 panic(
1767 "lock (%s) %s does not match earlier (%s) lock",
1768 description, lock_class->lc_name,
1769 w->w_class->lc_name);
1770 return (w);
1771 }
1772
1773 static void
1774 depart(struct witness *w)
1775 {
1776 struct witness_list *list;
1777
1778 MPASS(w->w_refcount == 0);
1779 if (w->w_class->lc_flags & LC_SLEEPLOCK) {
1780 list = &w_sleep;
1781 w_sleep_cnt--;
1782 } else {
1783 list = &w_spin;
1784 w_spin_cnt--;
1785 }
1786 /*
1787 * Set file to NULL as it may point into a loadable module.
1788 */
1789 w->w_file = NULL;
1790 w->w_line = 0;
1791 witness_increment_graph_generation();
1792 }
1793
1794
1795 static void
1796 adopt(struct witness *parent, struct witness *child)
1797 {
1798 int pi, ci, i, j;
1799
1800 if (witness_cold == 0)
1801 mtx_assert(&w_mtx, MA_OWNED);
1802
1803 /* If the relationship is already known, there's no work to be done. */
1804 if (isitmychild(parent, child))
1805 return;
1806
1807 /* When the structure of the graph changes, bump up the generation. */
1808 witness_increment_graph_generation();
1809
1810 /*
1811 * The hard part ... create the direct relationship, then propagate all
1812 * indirect relationships.
1813 */
1814 pi = parent->w_index;
1815 ci = child->w_index;
1816 WITNESS_INDEX_ASSERT(pi);
1817 WITNESS_INDEX_ASSERT(ci);
1818 MPASS(pi != ci);
1819 w_rmatrix[pi][ci] |= WITNESS_PARENT;
1820 w_rmatrix[ci][pi] |= WITNESS_CHILD;
1821
1822 /*
1823 * If parent was not already an ancestor of child,
1824 * then we increment the descendant and ancestor counters.
1825 */
1826 if ((w_rmatrix[pi][ci] & WITNESS_ANCESTOR) == 0) {
1827 parent->w_num_descendants++;
1828 child->w_num_ancestors++;
1829 }
1830
1831 /*
1832 * Find each ancestor of 'pi'. Note that 'pi' itself is counted as
1833 * an ancestor of 'pi' during this loop.
1834 */
1835 for (i = 1; i <= w_max_used_index; i++) {
1836 if ((w_rmatrix[i][pi] & WITNESS_ANCESTOR_MASK) == 0 &&
1837 (i != pi))
1838 continue;
1839
1840 /* Find each descendant of 'i' and mark it as a descendant. */
1841 for (j = 1; j <= w_max_used_index; j++) {
1842
1843 /*
1844 * Skip children that are already marked as
1845 * descendants of 'i'.
1846 */
1847 if (w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK)
1848 continue;
1849
1850 /*
1851 * We are only interested in descendants of 'ci'. Note
1852 * that 'ci' itself is counted as a descendant of 'ci'.
1853 */
1854 if ((w_rmatrix[ci][j] & WITNESS_ANCESTOR_MASK) == 0 &&
1855 (j != ci))
1856 continue;
1857 w_rmatrix[i][j] |= WITNESS_ANCESTOR;
1858 w_rmatrix[j][i] |= WITNESS_DESCENDANT;
1859 w_data[i].w_num_descendants++;
1860 w_data[j].w_num_ancestors++;
1861
1862 /*
1863 * Make sure we aren't marking a node as both an
1864 * ancestor and descendant. We should have caught
1865 * this as a lock order reversal earlier.
1866 */
1867 if ((w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK) &&
1868 (w_rmatrix[i][j] & WITNESS_DESCENDANT_MASK)) {
1869 printf("witness rmatrix paradox! [%d][%d]=%d "
1870 "both ancestor and descendant\n",
1871 i, j, w_rmatrix[i][j]);
1872 kdb_backtrace();
1873 printf("Witness disabled.\n");
1874 witness_watch = -1;
1875 }
1876 if ((w_rmatrix[j][i] & WITNESS_ANCESTOR_MASK) &&
1877 (w_rmatrix[j][i] & WITNESS_DESCENDANT_MASK)) {
1878 printf("witness rmatrix paradox! [%d][%d]=%d "
1879 "both ancestor and descendant\n",
1880 j, i, w_rmatrix[j][i]);
1881 kdb_backtrace();
1882 printf("Witness disabled.\n");
1883 witness_watch = -1;
1884 }
1885 }
1886 }
1887 }
1888
1889 static void
1890 itismychild(struct witness *parent, struct witness *child)
1891 {
1892
1893 MPASS(child != NULL && parent != NULL);
1894 if (witness_cold == 0)
1895 mtx_assert(&w_mtx, MA_OWNED);
1896
1897 if (!witness_lock_type_equal(parent, child)) {
1898 if (witness_cold == 0)
1899 mtx_unlock_spin(&w_mtx);
1900 panic("%s: parent \"%s\" (%s) and child \"%s\" (%s) are not "
1901 "the same lock type", __func__, parent->w_name,
1902 parent->w_class->lc_name, child->w_name,
1903 child->w_class->lc_name);
1904 }
1905 adopt(parent, child);
1906 }
1907
1908 /*
1909 * Generic code for the isitmy*() functions. The rmask parameter is the
1910 * expected relationship of w1 to w2.
1911 */
1912 static int
1913 _isitmyx(struct witness *w1, struct witness *w2, int rmask, const char *fname)
1914 {
1915 unsigned char r1, r2;
1916 int i1, i2;
1917
1918 i1 = w1->w_index;
1919 i2 = w2->w_index;
1920 WITNESS_INDEX_ASSERT(i1);
1921 WITNESS_INDEX_ASSERT(i2);
1922 r1 = w_rmatrix[i1][i2] & WITNESS_RELATED_MASK;
1923 r2 = w_rmatrix[i2][i1] & WITNESS_RELATED_MASK;
1924
1925 /* The flags on one better be the inverse of the flags on the other */
1926 if (!((WITNESS_ATOD(r1) == r2 && WITNESS_DTOA(r2) == r1) ||
1927 (WITNESS_DTOA(r1) == r2 && WITNESS_ATOD(r2) == r1))) {
1928 printf("%s: rmatrix mismatch between %s (index %d) and %s "
1929 "(index %d): w_rmatrix[%d][%d] == %hhx but "
1930 "w_rmatrix[%d][%d] == %hhx\n",
1931 fname, w1->w_name, i1, w2->w_name, i2, i1, i2, r1,
1932 i2, i1, r2);
1933 kdb_backtrace();
1934 printf("Witness disabled.\n");
1935 witness_watch = -1;
1936 }
1937 return (r1 & rmask);
1938 }
1939
1940 /*
1941 * Checks if @child is a direct child of @parent.
1942 */
1943 static int
1944 isitmychild(struct witness *parent, struct witness *child)
1945 {
1946
1947 return (_isitmyx(parent, child, WITNESS_PARENT, __func__));
1948 }
1949
1950 /*
1951 * Checks if @descendant is a direct or inderect descendant of @ancestor.
1952 */
1953 static int
1954 isitmydescendant(struct witness *ancestor, struct witness *descendant)
1955 {
1956
1957 return (_isitmyx(ancestor, descendant, WITNESS_ANCESTOR_MASK,
1958 __func__));
1959 }
1960
1961 #ifdef BLESSING
1962 static int
1963 blessed(struct witness *w1, struct witness *w2)
1964 {
1965 int i;
1966 struct witness_blessed *b;
1967
1968 for (i = 0; i < blessed_count; i++) {
1969 b = &blessed_list[i];
1970 if (strcmp(w1->w_name, b->b_lock1) == 0) {
1971 if (strcmp(w2->w_name, b->b_lock2) == 0)
1972 return (1);
1973 continue;
1974 }
1975 if (strcmp(w1->w_name, b->b_lock2) == 0)
1976 if (strcmp(w2->w_name, b->b_lock1) == 0)
1977 return (1);
1978 }
1979 return (0);
1980 }
1981 #endif
1982
1983 static struct witness *
1984 witness_get(void)
1985 {
1986 struct witness *w;
1987 int index;
1988
1989 if (witness_cold == 0)
1990 mtx_assert(&w_mtx, MA_OWNED);
1991
1992 if (witness_watch == -1) {
1993 mtx_unlock_spin(&w_mtx);
1994 return (NULL);
1995 }
1996 if (STAILQ_EMPTY(&w_free)) {
1997 witness_watch = -1;
1998 mtx_unlock_spin(&w_mtx);
1999 printf("WITNESS: unable to allocate a new witness object\n");
2000 return (NULL);
2001 }
2002 w = STAILQ_FIRST(&w_free);
2003 STAILQ_REMOVE_HEAD(&w_free, w_list);
2004 w_free_cnt--;
2005 index = w->w_index;
2006 MPASS(index > 0 && index == w_max_used_index+1 &&
2007 index < WITNESS_COUNT);
2008 bzero(w, sizeof(*w));
2009 w->w_index = index;
2010 if (index > w_max_used_index)
2011 w_max_used_index = index;
2012 return (w);
2013 }
2014
2015 static void
2016 witness_free(struct witness *w)
2017 {
2018
2019 STAILQ_INSERT_HEAD(&w_free, w, w_list);
2020 w_free_cnt++;
2021 }
2022
2023 static struct lock_list_entry *
2024 witness_lock_list_get(void)
2025 {
2026 struct lock_list_entry *lle;
2027
2028 if (witness_watch == -1)
2029 return (NULL);
2030 mtx_lock_spin(&w_mtx);
2031 lle = w_lock_list_free;
2032 if (lle == NULL) {
2033 witness_watch = -1;
2034 mtx_unlock_spin(&w_mtx);
2035 printf("%s: witness exhausted\n", __func__);
2036 return (NULL);
2037 }
2038 w_lock_list_free = lle->ll_next;
2039 mtx_unlock_spin(&w_mtx);
2040 bzero(lle, sizeof(*lle));
2041 return (lle);
2042 }
2043
2044 static void
2045 witness_lock_list_free(struct lock_list_entry *lle)
2046 {
2047
2048 mtx_lock_spin(&w_mtx);
2049 lle->ll_next = w_lock_list_free;
2050 w_lock_list_free = lle;
2051 mtx_unlock_spin(&w_mtx);
2052 }
2053
2054 static struct lock_instance *
2055 find_instance(struct lock_list_entry *list, struct lock_object *lock)
2056 {
2057 struct lock_list_entry *lle;
2058 struct lock_instance *instance;
2059 int i;
2060
2061 for (lle = list; lle != NULL; lle = lle->ll_next)
2062 for (i = lle->ll_count - 1; i >= 0; i--) {
2063 instance = &lle->ll_children[i];
2064 if (instance->li_lock == lock)
2065 return (instance);
2066 }
2067 return (NULL);
2068 }
2069
2070 static void
2071 witness_list_lock(struct lock_instance *instance,
2072 int (*prnt)(const char *fmt, ...))
2073 {
2074 struct lock_object *lock;
2075
2076 lock = instance->li_lock;
2077 prnt("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ?
2078 "exclusive" : "shared", LOCK_CLASS(lock)->lc_name, lock->lo_name);
2079 if (lock->lo_witness->w_name != lock->lo_name)
2080 prnt(" (%s)", lock->lo_witness->w_name);
2081 prnt(" r = %d (%p) locked @ %s:%d\n",
2082 instance->li_flags & LI_RECURSEMASK, lock, instance->li_file,
2083 instance->li_line);
2084 }
2085
2086 #ifdef DDB
2087 static int
2088 witness_thread_has_locks(struct thread *td)
2089 {
2090
2091 if (td->td_sleeplocks == NULL)
2092 return (0);
2093 return (td->td_sleeplocks->ll_count != 0);
2094 }
2095
2096 static int
2097 witness_proc_has_locks(struct proc *p)
2098 {
2099 struct thread *td;
2100
2101 FOREACH_THREAD_IN_PROC(p, td) {
2102 if (witness_thread_has_locks(td))
2103 return (1);
2104 }
2105 return (0);
2106 }
2107 #endif
2108
2109 int
2110 witness_list_locks(struct lock_list_entry **lock_list,
2111 int (*prnt)(const char *fmt, ...))
2112 {
2113 struct lock_list_entry *lle;
2114 int i, nheld;
2115
2116 nheld = 0;
2117 for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
2118 for (i = lle->ll_count - 1; i >= 0; i--) {
2119 witness_list_lock(&lle->ll_children[i], prnt);
2120 nheld++;
2121 }
2122 return (nheld);
2123 }
2124
2125 /*
2126 * This is a bit risky at best. We call this function when we have timed
2127 * out acquiring a spin lock, and we assume that the other CPU is stuck
2128 * with this lock held. So, we go groveling around in the other CPU's
2129 * per-cpu data to try to find the lock instance for this spin lock to
2130 * see when it was last acquired.
2131 */
2132 void
2133 witness_display_spinlock(struct lock_object *lock, struct thread *owner,
2134 int (*prnt)(const char *fmt, ...))
2135 {
2136 struct lock_instance *instance;
2137 struct pcpu *pc;
2138
2139 if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU)
2140 return;
2141 pc = pcpu_find(owner->td_oncpu);
2142 instance = find_instance(pc->pc_spinlocks, lock);
2143 if (instance != NULL)
2144 witness_list_lock(instance, prnt);
2145 }
2146
2147 void
2148 witness_save(struct lock_object *lock, const char **filep, int *linep)
2149 {
2150 struct lock_list_entry *lock_list;
2151 struct lock_instance *instance;
2152 struct lock_class *class;
2153
2154 /*
2155 * This function is used independently in locking code to deal with
2156 * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2157 * is gone.
2158 */
2159 if (SCHEDULER_STOPPED())
2160 return;
2161 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2162 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2163 return;
2164 class = LOCK_CLASS(lock);
2165 if (class->lc_flags & LC_SLEEPLOCK)
2166 lock_list = curthread->td_sleeplocks;
2167 else {
2168 if (witness_skipspin)
2169 return;
2170 lock_list = PCPU_GET(spinlocks);
2171 }
2172 instance = find_instance(lock_list, lock);
2173 if (instance == NULL)
2174 panic("%s: lock (%s) %s not locked", __func__,
2175 class->lc_name, lock->lo_name);
2176 *filep = instance->li_file;
2177 *linep = instance->li_line;
2178 }
2179
2180 void
2181 witness_restore(struct lock_object *lock, const char *file, int line)
2182 {
2183 struct lock_list_entry *lock_list;
2184 struct lock_instance *instance;
2185 struct lock_class *class;
2186
2187 /*
2188 * This function is used independently in locking code to deal with
2189 * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2190 * is gone.
2191 */
2192 if (SCHEDULER_STOPPED())
2193 return;
2194 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2195 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2196 return;
2197 class = LOCK_CLASS(lock);
2198 if (class->lc_flags & LC_SLEEPLOCK)
2199 lock_list = curthread->td_sleeplocks;
2200 else {
2201 if (witness_skipspin)
2202 return;
2203 lock_list = PCPU_GET(spinlocks);
2204 }
2205 instance = find_instance(lock_list, lock);
2206 if (instance == NULL)
2207 panic("%s: lock (%s) %s not locked", __func__,
2208 class->lc_name, lock->lo_name);
2209 lock->lo_witness->w_file = file;
2210 lock->lo_witness->w_line = line;
2211 instance->li_file = file;
2212 instance->li_line = line;
2213 }
2214
2215 void
2216 witness_assert(struct lock_object *lock, int flags, const char *file, int line)
2217 {
2218 #ifdef INVARIANT_SUPPORT
2219 struct lock_instance *instance;
2220 struct lock_class *class;
2221
2222 if (lock->lo_witness == NULL || witness_watch < 1 || panicstr != NULL)
2223 return;
2224 class = LOCK_CLASS(lock);
2225 if ((class->lc_flags & LC_SLEEPLOCK) != 0)
2226 instance = find_instance(curthread->td_sleeplocks, lock);
2227 else if ((class->lc_flags & LC_SPINLOCK) != 0)
2228 instance = find_instance(PCPU_GET(spinlocks), lock);
2229 else {
2230 panic("Lock (%s) %s is not sleep or spin!",
2231 class->lc_name, lock->lo_name);
2232 }
2233 file = fixup_filename(file);
2234 switch (flags) {
2235 case LA_UNLOCKED:
2236 if (instance != NULL)
2237 panic("Lock (%s) %s locked @ %s:%d.",
2238 class->lc_name, lock->lo_name, file, line);
2239 break;
2240 case LA_LOCKED:
2241 case LA_LOCKED | LA_RECURSED:
2242 case LA_LOCKED | LA_NOTRECURSED:
2243 case LA_SLOCKED:
2244 case LA_SLOCKED | LA_RECURSED:
2245 case LA_SLOCKED | LA_NOTRECURSED:
2246 case LA_XLOCKED:
2247 case LA_XLOCKED | LA_RECURSED:
2248 case LA_XLOCKED | LA_NOTRECURSED:
2249 if (instance == NULL) {
2250 panic("Lock (%s) %s not locked @ %s:%d.",
2251 class->lc_name, lock->lo_name, file, line);
2252 break;
2253 }
2254 if ((flags & LA_XLOCKED) != 0 &&
2255 (instance->li_flags & LI_EXCLUSIVE) == 0)
2256 panic("Lock (%s) %s not exclusively locked @ %s:%d.",
2257 class->lc_name, lock->lo_name, file, line);
2258 if ((flags & LA_SLOCKED) != 0 &&
2259 (instance->li_flags & LI_EXCLUSIVE) != 0)
2260 panic("Lock (%s) %s exclusively locked @ %s:%d.",
2261 class->lc_name, lock->lo_name, file, line);
2262 if ((flags & LA_RECURSED) != 0 &&
2263 (instance->li_flags & LI_RECURSEMASK) == 0)
2264 panic("Lock (%s) %s not recursed @ %s:%d.",
2265 class->lc_name, lock->lo_name, file, line);
2266 if ((flags & LA_NOTRECURSED) != 0 &&
2267 (instance->li_flags & LI_RECURSEMASK) != 0)
2268 panic("Lock (%s) %s recursed @ %s:%d.",
2269 class->lc_name, lock->lo_name, file, line);
2270 break;
2271 default:
2272 panic("Invalid lock assertion at %s:%d.", file, line);
2273
2274 }
2275 #endif /* INVARIANT_SUPPORT */
2276 }
2277
2278 static void
2279 witness_setflag(struct lock_object *lock, int flag, int set)
2280 {
2281 struct lock_list_entry *lock_list;
2282 struct lock_instance *instance;
2283 struct lock_class *class;
2284
2285 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2286 return;
2287 class = LOCK_CLASS(lock);
2288 if (class->lc_flags & LC_SLEEPLOCK)
2289 lock_list = curthread->td_sleeplocks;
2290 else {
2291 if (witness_skipspin)
2292 return;
2293 lock_list = PCPU_GET(spinlocks);
2294 }
2295 instance = find_instance(lock_list, lock);
2296 if (instance == NULL)
2297 panic("%s: lock (%s) %s not locked", __func__,
2298 class->lc_name, lock->lo_name);
2299
2300 if (set)
2301 instance->li_flags |= flag;
2302 else
2303 instance->li_flags &= ~flag;
2304 }
2305
2306 void
2307 witness_norelease(struct lock_object *lock)
2308 {
2309
2310 witness_setflag(lock, LI_NORELEASE, 1);
2311 }
2312
2313 void
2314 witness_releaseok(struct lock_object *lock)
2315 {
2316
2317 witness_setflag(lock, LI_NORELEASE, 0);
2318 }
2319
2320 #ifdef DDB
2321 static void
2322 witness_ddb_list(struct thread *td)
2323 {
2324
2325 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2326 KASSERT(kdb_active, ("%s: not in the debugger", __func__));
2327
2328 if (witness_watch < 1)
2329 return;
2330
2331 witness_list_locks(&td->td_sleeplocks, db_printf);
2332
2333 /*
2334 * We only handle spinlocks if td == curthread. This is somewhat broken
2335 * if td is currently executing on some other CPU and holds spin locks
2336 * as we won't display those locks. If we had a MI way of getting
2337 * the per-cpu data for a given cpu then we could use
2338 * td->td_oncpu to get the list of spinlocks for this thread
2339 * and "fix" this.
2340 *
2341 * That still wouldn't really fix this unless we locked the scheduler
2342 * lock or stopped the other CPU to make sure it wasn't changing the
2343 * list out from under us. It is probably best to just not try to
2344 * handle threads on other CPU's for now.
2345 */
2346 if (td == curthread && PCPU_GET(spinlocks) != NULL)
2347 witness_list_locks(PCPU_PTR(spinlocks), db_printf);
2348 }
2349
2350 DB_SHOW_COMMAND(locks, db_witness_list)
2351 {
2352 struct thread *td;
2353
2354 if (have_addr)
2355 td = db_lookup_thread(addr, TRUE);
2356 else
2357 td = kdb_thread;
2358 witness_ddb_list(td);
2359 }
2360
2361 DB_SHOW_ALL_COMMAND(locks, db_witness_list_all)
2362 {
2363 struct thread *td;
2364 struct proc *p;
2365
2366 /*
2367 * It would be nice to list only threads and processes that actually
2368 * held sleep locks, but that information is currently not exported
2369 * by WITNESS.
2370 */
2371 FOREACH_PROC_IN_SYSTEM(p) {
2372 if (!witness_proc_has_locks(p))
2373 continue;
2374 FOREACH_THREAD_IN_PROC(p, td) {
2375 if (!witness_thread_has_locks(td))
2376 continue;
2377 db_printf("Process %d (%s) thread %p (%d)\n", p->p_pid,
2378 p->p_comm, td, td->td_tid);
2379 witness_ddb_list(td);
2380 if (db_pager_quit)
2381 return;
2382 }
2383 }
2384 }
2385 DB_SHOW_ALIAS(alllocks, db_witness_list_all)
2386
2387 DB_SHOW_COMMAND(witness, db_witness_display)
2388 {
2389
2390 witness_ddb_display(db_printf);
2391 }
2392 #endif
2393
2394 static int
2395 sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS)
2396 {
2397 struct witness_lock_order_data *data1, *data2, *tmp_data1, *tmp_data2;
2398 struct witness *tmp_w1, *tmp_w2, *w1, *w2;
2399 struct sbuf *sb;
2400 u_int w_rmatrix1, w_rmatrix2;
2401 int error, generation, i, j;
2402
2403 tmp_data1 = NULL;
2404 tmp_data2 = NULL;
2405 tmp_w1 = NULL;
2406 tmp_w2 = NULL;
2407 if (witness_watch < 1) {
2408 error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2409 return (error);
2410 }
2411 if (witness_cold) {
2412 error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2413 return (error);
2414 }
2415 error = 0;
2416 sb = sbuf_new(NULL, NULL, BADSTACK_SBUF_SIZE, SBUF_AUTOEXTEND);
2417 if (sb == NULL)
2418 return (ENOMEM);
2419
2420 /* Allocate and init temporary storage space. */
2421 tmp_w1 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2422 tmp_w2 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2423 tmp_data1 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2424 M_WAITOK | M_ZERO);
2425 tmp_data2 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2426 M_WAITOK | M_ZERO);
2427 stack_zero(&tmp_data1->wlod_stack);
2428 stack_zero(&tmp_data2->wlod_stack);
2429
2430 restart:
2431 mtx_lock_spin(&w_mtx);
2432 generation = w_generation;
2433 mtx_unlock_spin(&w_mtx);
2434 sbuf_printf(sb, "Number of known direct relationships is %d\n",
2435 w_lohash.wloh_count);
2436 for (i = 1; i < w_max_used_index; i++) {
2437 mtx_lock_spin(&w_mtx);
2438 if (generation != w_generation) {
2439 mtx_unlock_spin(&w_mtx);
2440
2441 /* The graph has changed, try again. */
2442 req->oldidx = 0;
2443 sbuf_clear(sb);
2444 goto restart;
2445 }
2446
2447 w1 = &w_data[i];
2448 if (w1->w_reversed == 0) {
2449 mtx_unlock_spin(&w_mtx);
2450 continue;
2451 }
2452
2453 /* Copy w1 locally so we can release the spin lock. */
2454 *tmp_w1 = *w1;
2455 mtx_unlock_spin(&w_mtx);
2456
2457 if (tmp_w1->w_reversed == 0)
2458 continue;
2459 for (j = 1; j < w_max_used_index; j++) {
2460 if ((w_rmatrix[i][j] & WITNESS_REVERSAL) == 0 || i > j)
2461 continue;
2462
2463 mtx_lock_spin(&w_mtx);
2464 if (generation != w_generation) {
2465 mtx_unlock_spin(&w_mtx);
2466
2467 /* The graph has changed, try again. */
2468 req->oldidx = 0;
2469 sbuf_clear(sb);
2470 goto restart;
2471 }
2472
2473 w2 = &w_data[j];
2474 data1 = witness_lock_order_get(w1, w2);
2475 data2 = witness_lock_order_get(w2, w1);
2476
2477 /*
2478 * Copy information locally so we can release the
2479 * spin lock.
2480 */
2481 *tmp_w2 = *w2;
2482 w_rmatrix1 = (unsigned int)w_rmatrix[i][j];
2483 w_rmatrix2 = (unsigned int)w_rmatrix[j][i];
2484
2485 if (data1) {
2486 stack_zero(&tmp_data1->wlod_stack);
2487 stack_copy(&data1->wlod_stack,
2488 &tmp_data1->wlod_stack);
2489 }
2490 if (data2 && data2 != data1) {
2491 stack_zero(&tmp_data2->wlod_stack);
2492 stack_copy(&data2->wlod_stack,
2493 &tmp_data2->wlod_stack);
2494 }
2495 mtx_unlock_spin(&w_mtx);
2496
2497 sbuf_printf(sb,
2498 "\nLock order reversal between \"%s\"(%s) and \"%s\"(%s)!\n",
2499 tmp_w1->w_name, tmp_w1->w_class->lc_name,
2500 tmp_w2->w_name, tmp_w2->w_class->lc_name);
2501 #if 0
2502 sbuf_printf(sb,
2503 "w_rmatrix[%s][%s] == %x, w_rmatrix[%s][%s] == %x\n",
2504 tmp_w1->name, tmp_w2->w_name, w_rmatrix1,
2505 tmp_w2->name, tmp_w1->w_name, w_rmatrix2);
2506 #endif
2507 if (data1) {
2508 sbuf_printf(sb,
2509 "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2510 tmp_w1->w_name, tmp_w1->w_class->lc_name,
2511 tmp_w2->w_name, tmp_w2->w_class->lc_name);
2512 stack_sbuf_print(sb, &tmp_data1->wlod_stack);
2513 sbuf_printf(sb, "\n");
2514 }
2515 if (data2 && data2 != data1) {
2516 sbuf_printf(sb,
2517 "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2518 tmp_w2->w_name, tmp_w2->w_class->lc_name,
2519 tmp_w1->w_name, tmp_w1->w_class->lc_name);
2520 stack_sbuf_print(sb, &tmp_data2->wlod_stack);
2521 sbuf_printf(sb, "\n");
2522 }
2523 }
2524 }
2525 mtx_lock_spin(&w_mtx);
2526 if (generation != w_generation) {
2527 mtx_unlock_spin(&w_mtx);
2528
2529 /*
2530 * The graph changed while we were printing stack data,
2531 * try again.
2532 */
2533 req->oldidx = 0;
2534 sbuf_clear(sb);
2535 goto restart;
2536 }
2537 mtx_unlock_spin(&w_mtx);
2538
2539 /* Free temporary storage space. */
2540 free(tmp_data1, M_TEMP);
2541 free(tmp_data2, M_TEMP);
2542 free(tmp_w1, M_TEMP);
2543 free(tmp_w2, M_TEMP);
2544
2545 sbuf_finish(sb);
2546 error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1);
2547 sbuf_delete(sb);
2548
2549 return (error);
2550 }
2551
2552 static int
2553 sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS)
2554 {
2555 struct witness *w;
2556 struct sbuf *sb;
2557 int error;
2558
2559 if (witness_watch < 1) {
2560 error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2561 return (error);
2562 }
2563 if (witness_cold) {
2564 error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2565 return (error);
2566 }
2567 error = 0;
2568 sb = sbuf_new(NULL, NULL, FULLGRAPH_SBUF_SIZE, SBUF_FIXEDLEN);
2569 if (sb == NULL)
2570 return (ENOMEM);
2571 sbuf_printf(sb, "\n");
2572
2573 mtx_lock_spin(&w_mtx);
2574 STAILQ_FOREACH(w, &w_all, w_list)
2575 w->w_displayed = 0;
2576 STAILQ_FOREACH(w, &w_all, w_list)
2577 witness_add_fullgraph(sb, w);
2578 mtx_unlock_spin(&w_mtx);
2579
2580 /*
2581 * While using SBUF_FIXEDLEN, check if the sbuf overflowed.
2582 */
2583 if (sbuf_overflowed(sb)) {
2584 sbuf_delete(sb);
2585 panic("%s: sbuf overflowed, bump FULLGRAPH_SBUF_SIZE value\n",
2586 __func__);
2587 }
2588
2589 /*
2590 * Close the sbuf and return to userland.
2591 */
2592 sbuf_finish(sb);
2593 error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1);
2594 sbuf_delete(sb);
2595
2596 return (error);
2597 }
2598
2599 static int
2600 sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS)
2601 {
2602 int error, value;
2603
2604 value = witness_watch;
2605 error = sysctl_handle_int(oidp, &value, 0, req);
2606 if (error != 0 || req->newptr == NULL)
2607 return (error);
2608 if (value > 1 || value < -1 ||
2609 (witness_watch == -1 && value != witness_watch))
2610 return (EINVAL);
2611 witness_watch = value;
2612 return (0);
2613 }
2614
2615 static void
2616 witness_add_fullgraph(struct sbuf *sb, struct witness *w)
2617 {
2618 int i;
2619
2620 if (w->w_displayed != 0 || (w->w_file == NULL && w->w_line == 0))
2621 return;
2622 w->w_displayed = 1;
2623
2624 WITNESS_INDEX_ASSERT(w->w_index);
2625 for (i = 1; i <= w_max_used_index; i++) {
2626 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT) {
2627 sbuf_printf(sb, "\"%s\",\"%s\"\n", w->w_name,
2628 w_data[i].w_name);
2629 witness_add_fullgraph(sb, &w_data[i]);
2630 }
2631 }
2632 }
2633
2634 /*
2635 * A simple hash function. Takes a key pointer and a key size. If size == 0,
2636 * interprets the key as a string and reads until the null
2637 * terminator. Otherwise, reads the first size bytes. Returns an unsigned 32-bit
2638 * hash value computed from the key.
2639 */
2640 static uint32_t
2641 witness_hash_djb2(const uint8_t *key, uint32_t size)
2642 {
2643 unsigned int hash = 5381;
2644 int i;
2645
2646 /* hash = hash * 33 + key[i] */
2647 if (size)
2648 for (i = 0; i < size; i++)
2649 hash = ((hash << 5) + hash) + (unsigned int)key[i];
2650 else
2651 for (i = 0; key[i] != 0; i++)
2652 hash = ((hash << 5) + hash) + (unsigned int)key[i];
2653
2654 return (hash);
2655 }
2656
2657
2658 /*
2659 * Initializes the two witness hash tables. Called exactly once from
2660 * witness_initialize().
2661 */
2662 static void
2663 witness_init_hash_tables(void)
2664 {
2665 int i;
2666
2667 MPASS(witness_cold);
2668
2669 /* Initialize the hash tables. */
2670 for (i = 0; i < WITNESS_HASH_SIZE; i++)
2671 w_hash.wh_array[i] = NULL;
2672
2673 w_hash.wh_size = WITNESS_HASH_SIZE;
2674 w_hash.wh_count = 0;
2675
2676 /* Initialize the lock order data hash. */
2677 w_lofree = NULL;
2678 for (i = 0; i < WITNESS_LO_DATA_COUNT; i++) {
2679 memset(&w_lodata[i], 0, sizeof(w_lodata[i]));
2680 w_lodata[i].wlod_next = w_lofree;
2681 w_lofree = &w_lodata[i];
2682 }
2683 w_lohash.wloh_size = WITNESS_LO_HASH_SIZE;
2684 w_lohash.wloh_count = 0;
2685 for (i = 0; i < WITNESS_LO_HASH_SIZE; i++)
2686 w_lohash.wloh_array[i] = NULL;
2687 }
2688
2689 static struct witness *
2690 witness_hash_get(const char *key)
2691 {
2692 struct witness *w;
2693 uint32_t hash;
2694
2695 MPASS(key != NULL);
2696 if (witness_cold == 0)
2697 mtx_assert(&w_mtx, MA_OWNED);
2698 hash = witness_hash_djb2(key, 0) % w_hash.wh_size;
2699 w = w_hash.wh_array[hash];
2700 while (w != NULL) {
2701 if (strcmp(w->w_name, key) == 0)
2702 goto out;
2703 w = w->w_hash_next;
2704 }
2705
2706 out:
2707 return (w);
2708 }
2709
2710 static void
2711 witness_hash_put(struct witness *w)
2712 {
2713 uint32_t hash;
2714
2715 MPASS(w != NULL);
2716 MPASS(w->w_name != NULL);
2717 if (witness_cold == 0)
2718 mtx_assert(&w_mtx, MA_OWNED);
2719 KASSERT(witness_hash_get(w->w_name) == NULL,
2720 ("%s: trying to add a hash entry that already exists!", __func__));
2721 KASSERT(w->w_hash_next == NULL,
2722 ("%s: w->w_hash_next != NULL", __func__));
2723
2724 hash = witness_hash_djb2(w->w_name, 0) % w_hash.wh_size;
2725 w->w_hash_next = w_hash.wh_array[hash];
2726 w_hash.wh_array[hash] = w;
2727 w_hash.wh_count++;
2728 }
2729
2730
2731 static struct witness_lock_order_data *
2732 witness_lock_order_get(struct witness *parent, struct witness *child)
2733 {
2734 struct witness_lock_order_data *data = NULL;
2735 struct witness_lock_order_key key;
2736 unsigned int hash;
2737
2738 MPASS(parent != NULL && child != NULL);
2739 key.from = parent->w_index;
2740 key.to = child->w_index;
2741 WITNESS_INDEX_ASSERT(key.from);
2742 WITNESS_INDEX_ASSERT(key.to);
2743 if ((w_rmatrix[parent->w_index][child->w_index]
2744 & WITNESS_LOCK_ORDER_KNOWN) == 0)
2745 goto out;
2746
2747 hash = witness_hash_djb2((const char*)&key,
2748 sizeof(key)) % w_lohash.wloh_size;
2749 data = w_lohash.wloh_array[hash];
2750 while (data != NULL) {
2751 if (witness_lock_order_key_equal(&data->wlod_key, &key))
2752 break;
2753 data = data->wlod_next;
2754 }
2755
2756 out:
2757 return (data);
2758 }
2759
2760 /*
2761 * Verify that parent and child have a known relationship, are not the same,
2762 * and child is actually a child of parent. This is done without w_mtx
2763 * to avoid contention in the common case.
2764 */
2765 static int
2766 witness_lock_order_check(struct witness *parent, struct witness *child)
2767 {
2768
2769 if (parent != child &&
2770 w_rmatrix[parent->w_index][child->w_index]
2771 & WITNESS_LOCK_ORDER_KNOWN &&
2772 isitmychild(parent, child))
2773 return (1);
2774
2775 return (0);
2776 }
2777
2778 static int
2779 witness_lock_order_add(struct witness *parent, struct witness *child)
2780 {
2781 struct witness_lock_order_data *data = NULL;
2782 struct witness_lock_order_key key;
2783 unsigned int hash;
2784
2785 MPASS(parent != NULL && child != NULL);
2786 key.from = parent->w_index;
2787 key.to = child->w_index;
2788 WITNESS_INDEX_ASSERT(key.from);
2789 WITNESS_INDEX_ASSERT(key.to);
2790 if (w_rmatrix[parent->w_index][child->w_index]
2791 & WITNESS_LOCK_ORDER_KNOWN)
2792 return (1);
2793
2794 hash = witness_hash_djb2((const char*)&key,
2795 sizeof(key)) % w_lohash.wloh_size;
2796 w_rmatrix[parent->w_index][child->w_index] |= WITNESS_LOCK_ORDER_KNOWN;
2797 data = w_lofree;
2798 if (data == NULL)
2799 return (0);
2800 w_lofree = data->wlod_next;
2801 data->wlod_next = w_lohash.wloh_array[hash];
2802 data->wlod_key = key;
2803 w_lohash.wloh_array[hash] = data;
2804 w_lohash.wloh_count++;
2805 stack_zero(&data->wlod_stack);
2806 stack_save(&data->wlod_stack);
2807 return (1);
2808 }
2809
2810 /* Call this whenver the structure of the witness graph changes. */
2811 static void
2812 witness_increment_graph_generation(void)
2813 {
2814
2815 if (witness_cold == 0)
2816 mtx_assert(&w_mtx, MA_OWNED);
2817 w_generation++;
2818 }
2819
2820 #ifdef KDB
2821 static void
2822 _witness_debugger(int cond, const char *msg)
2823 {
2824
2825 if (witness_trace && cond)
2826 kdb_backtrace();
2827 if (witness_kdb && cond)
2828 kdb_enter(KDB_WHY_WITNESS, msg);
2829 }
2830 #endif
Cache object: 1679f60eb704d1303d9e1ae4a9edff6d
|