2 * Copyright (c) 2008 Isilon Systems, Inc.
3 * Copyright (c) 2008 Ilya Maykov <ivmaykov@gmail.com>
4 * Copyright (c) 1998 Berkeley Software Design, Inc.
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
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
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
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 $
36 * Implementation of the `witness' lock verifier. Originally implemented for
37 * mutexes in BSD/OS. Extended to handle generic lock objects and lock
43 * Pronunciation: 'wit-n&s
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
58 * 6 capitalized : a member of the Jehovah's Witnesses
62 * Special rules concerning Giant and lock orders:
64 * 1) Giant must be acquired before any other mutexes. Stated another way,
65 * no other mutex may be held when Giant is acquired.
67 * 2) Giant must be released when blocking on a sleepable lock.
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
74 * 3) Giant may be acquired before or after sleepable locks.
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.
87 #include <sys/cdefs.h>
88 __FBSDID("$FreeBSD$");
91 #include "opt_hwpmc_hooks.h"
92 #include "opt_stack.h"
93 #include "opt_witness.h"
95 #include <sys/param.h>
98 #include <sys/kernel.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>
115 #include <machine/stdarg.h>
117 #if !defined(DDB) && !defined(STACK)
118 #error "DDB or STACK options are required for WITNESS"
121 /* Note that these traces do not work with KTR_ALQ. */
123 #define KTR_WITNESS KTR_SUBSYS
125 #define KTR_WITNESS 0
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. */
132 /* Define this to check for blessed mutexes */
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
140 /* Allocate 256 KB of stack data space */
141 #define WITNESS_LO_DATA_COUNT 2048
143 /* Prime, gives load factor of ~2 at full load */
144 #define WITNESS_LO_HASH_SIZE 1021
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.
151 #define LOCK_NCHILDREN 5
152 #define LOCK_CHILDCOUNT 2048
154 #define MAX_W_NAME 64
156 #define BADSTACK_SBUF_SIZE (256 * WITNESS_COUNT)
157 #define FULLGRAPH_SBUF_SIZE 512
160 * These flags go in the witness relationship matrix and describe the
161 * relationship between any two struct witness objects.
163 #define WITNESS_UNRELATED 0x00 /* No lock order relation. */
164 #define WITNESS_PARENT 0x01 /* Parent, aka direct ancestor. */
165 #define WITNESS_ANCESTOR 0x02 /* Direct or indirect ancestor. */
166 #define WITNESS_CHILD 0x04 /* Child, aka direct descendant. */
167 #define WITNESS_DESCENDANT 0x08 /* Direct or indirect descendant. */
168 #define WITNESS_ANCESTOR_MASK (WITNESS_PARENT | WITNESS_ANCESTOR)
169 #define WITNESS_DESCENDANT_MASK (WITNESS_CHILD | WITNESS_DESCENDANT)
170 #define WITNESS_RELATED_MASK \
171 (WITNESS_ANCESTOR_MASK | WITNESS_DESCENDANT_MASK)
172 #define WITNESS_REVERSAL 0x10 /* A lock order reversal has been
174 #define WITNESS_RESERVED1 0x20 /* Unused flag, reserved. */
175 #define WITNESS_RESERVED2 0x40 /* Unused flag, reserved. */
176 #define WITNESS_LOCK_ORDER_KNOWN 0x80 /* This lock order is known. */
178 /* Descendant to ancestor flags */
179 #define WITNESS_DTOA(x) (((x) & WITNESS_RELATED_MASK) >> 2)
181 /* Ancestor to descendant flags */
182 #define WITNESS_ATOD(x) (((x) & WITNESS_RELATED_MASK) << 2)
184 #define WITNESS_INDEX_ASSERT(i) \
185 MPASS((i) > 0 && (i) <= w_max_used_index && (i) < WITNESS_COUNT)
187 static MALLOC_DEFINE(M_WITNESS, "Witness", "Witness");
190 * Lock instances. A lock instance is the data associated with a lock while
191 * it is held by witness. For example, a lock instance will hold the
192 * recursion count of a lock. Lock instances are held in lists. Spin locks
193 * are held in a per-cpu list while sleep locks are held in per-thread list.
195 struct lock_instance {
196 struct lock_object *li_lock;
203 * A simple list type used to build the list of locks held by a thread
204 * or CPU. We can't simply embed the list in struct lock_object since a
205 * lock may be held by more than one thread if it is a shared lock. Locks
206 * are added to the head of the list, so we fill up each list entry from
207 * "the back" logically. To ease some of the arithmetic, we actually fill
208 * in each list entry the normal way (children[0] then children[1], etc.) but
209 * when we traverse the list we read children[count-1] as the first entry
210 * down to children[0] as the final entry.
212 struct lock_list_entry {
213 struct lock_list_entry *ll_next;
214 struct lock_instance ll_children[LOCK_NCHILDREN];
219 * The main witness structure. One of these per named lock type in the system
220 * (for example, "vnode interlock").
223 char w_name[MAX_W_NAME];
224 uint32_t w_index; /* Index in the relationship matrix */
225 struct lock_class *w_class;
226 STAILQ_ENTRY(witness) w_list; /* List of all witnesses. */
227 STAILQ_ENTRY(witness) w_typelist; /* Witnesses of a type. */
228 struct witness *w_hash_next; /* Linked list in hash buckets. */
229 const char *w_file; /* File where last acquired */
230 uint32_t w_line; /* Line where last acquired */
232 uint16_t w_num_ancestors; /* direct/indirect
234 uint16_t w_num_descendants; /* direct/indirect
235 * descendant count */
237 unsigned w_displayed:1;
238 unsigned w_reversed:1;
241 STAILQ_HEAD(witness_list, witness);
244 * The witness hash table. Keys are witness names (const char *), elements are
245 * witness objects (struct witness *).
247 struct witness_hash {
248 struct witness *wh_array[WITNESS_HASH_SIZE];
254 * Key type for the lock order data hash table.
256 struct witness_lock_order_key {
261 struct witness_lock_order_data {
262 struct stack wlod_stack;
263 struct witness_lock_order_key wlod_key;
264 struct witness_lock_order_data *wlod_next;
268 * The witness lock order data hash table. Keys are witness index tuples
269 * (struct witness_lock_order_key), elements are lock order data objects
270 * (struct witness_lock_order_data).
272 struct witness_lock_order_hash {
273 struct witness_lock_order_data *wloh_array[WITNESS_LO_HASH_SIZE];
279 struct witness_blessed {
285 struct witness_pendhelp {
287 struct lock_object *wh_lock;
290 struct witness_order_list_entry {
292 struct lock_class *w_class;
296 * Returns 0 if one of the locks is a spin lock and the other is not.
297 * Returns 1 otherwise.
300 witness_lock_type_equal(struct witness *w1, struct witness *w2)
303 return ((w1->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) ==
304 (w2->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)));
308 witness_lock_order_key_empty(const struct witness_lock_order_key *key)
311 return (key->from == 0 && key->to == 0);
315 witness_lock_order_key_equal(const struct witness_lock_order_key *a,
316 const struct witness_lock_order_key *b)
319 return (a->from == b->from && a->to == b->to);
322 static int _isitmyx(struct witness *w1, struct witness *w2, int rmask,
325 static void _witness_debugger(int cond, const char *msg);
327 static void adopt(struct witness *parent, struct witness *child);
329 static int blessed(struct witness *, struct witness *);
331 static void depart(struct witness *w);
332 static struct witness *enroll(const char *description,
333 struct lock_class *lock_class);
334 static struct lock_instance *find_instance(struct lock_list_entry *list,
335 struct lock_object *lock);
336 static int isitmychild(struct witness *parent, struct witness *child);
337 static int isitmydescendant(struct witness *parent, struct witness *child);
338 static void itismychild(struct witness *parent, struct witness *child);
339 static int sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS);
340 static int sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS);
341 static int sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS);
342 static void witness_add_fullgraph(struct sbuf *sb, struct witness *parent);
344 static void witness_ddb_compute_levels(void);
345 static void witness_ddb_display(int(*)(const char *fmt, ...));
346 static void witness_ddb_display_descendants(int(*)(const char *fmt, ...),
347 struct witness *, int indent);
348 static void witness_ddb_display_list(int(*prnt)(const char *fmt, ...),
349 struct witness_list *list);
350 static void witness_ddb_level_descendants(struct witness *parent, int l);
351 static void witness_ddb_list(struct thread *td);
353 static void witness_free(struct witness *m);
354 static struct witness *witness_get(void);
355 static uint32_t witness_hash_djb2(const uint8_t *key, uint32_t size);
356 static struct witness *witness_hash_get(const char *key);
357 static void witness_hash_put(struct witness *w);
358 static void witness_init_hash_tables(void);
359 static void witness_increment_graph_generation(void);
360 static void witness_lock_list_free(struct lock_list_entry *lle);
361 static struct lock_list_entry *witness_lock_list_get(void);
362 static int witness_lock_order_add(struct witness *parent,
363 struct witness *child);
364 static int witness_lock_order_check(struct witness *parent,
365 struct witness *child);
366 static struct witness_lock_order_data *witness_lock_order_get(
367 struct witness *parent,
368 struct witness *child);
369 static void witness_list_lock(struct lock_instance *instance,
370 int (*prnt)(const char *fmt, ...));
371 static void witness_setflag(struct lock_object *lock, int flag, int set);
374 #define witness_debugger(c) _witness_debugger(c, __func__)
376 #define witness_debugger(c)
379 static SYSCTL_NODE(_debug, OID_AUTO, witness, CTLFLAG_RW, NULL,
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.
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");
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.
406 TUNABLE_INT("debug.witness.kdb", &witness_kdb);
407 SYSCTL_INT(_debug_witness, OID_AUTO, kdb, CTLFLAG_RW, &witness_kdb, 0, "");
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.
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, "");
420 #ifdef WITNESS_SKIPSPIN
421 int witness_skipspin = 1;
423 int witness_skipspin = 0;
425 TUNABLE_INT("debug.witness.skipspin", &witness_skipspin);
426 SYSCTL_INT(_debug_witness, OID_AUTO, skipspin, CTLFLAG_RDTUN, &witness_skipspin,
430 * Call this to print out the relations between locks.
432 SYSCTL_PROC(_debug_witness, OID_AUTO, fullgraph, CTLTYPE_STRING | CTLFLAG_RD,
433 NULL, 0, sysctl_debug_witness_fullgraph, "A", "Show locks relation graphs");
436 * Call this to print out the witness faulty stacks.
438 SYSCTL_PROC(_debug_witness, OID_AUTO, badstacks, CTLTYPE_STRING | CTLFLAG_RD,
439 NULL, 0, sysctl_debug_witness_badstacks, "A", "Show bad witness stacks");
441 static struct mtx w_mtx;
444 static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free);
445 static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all);
448 static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin);
449 static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep);
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;
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,
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. */
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";
477 static struct witness_order_list_entry order_lists[] = {
481 { "proctree", &lock_class_sx },
482 { "allproc", &lock_class_sx },
483 { "allprison", &lock_class_sx },
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 },
496 { "pmc-sleep", &lock_class_mtx_sleep },
498 { "time lock", &lock_class_mtx_sleep },
503 { "accept", &lock_class_mtx_sleep },
504 { "so_snd", &lock_class_mtx_sleep },
505 { "so_rcv", &lock_class_mtx_sleep },
506 { "sellck", &lock_class_mtx_sleep },
511 { "so_rcv", &lock_class_mtx_sleep },
512 { "radix node head", &lock_class_rw },
513 { "rtentry", &lock_class_mtx_sleep },
514 { "ifaddr", &lock_class_mtx_sleep },
518 * protocol locks before interface locks, after UDP locks.
520 { "udpinp", &lock_class_rw },
521 { "in_multi_mtx", &lock_class_mtx_sleep },
522 { "igmp_mtx", &lock_class_mtx_sleep },
523 { "if_addr_mtx", &lock_class_mtx_sleep },
527 * protocol locks before interface locks, after UDP locks.
529 { "udpinp", &lock_class_rw },
530 { "in6_multi_mtx", &lock_class_mtx_sleep },
531 { "mld_mtx", &lock_class_mtx_sleep },
532 { "if_addr_mtx", &lock_class_mtx_sleep },
535 * UNIX Domain Sockets
537 { "unp_global_rwlock", &lock_class_rw },
538 { "unp_list_lock", &lock_class_mtx_sleep },
539 { "unp", &lock_class_mtx_sleep },
540 { "so_snd", &lock_class_mtx_sleep },
545 { "udp", &lock_class_rw },
546 { "udpinp", &lock_class_rw },
547 { "so_snd", &lock_class_mtx_sleep },
552 { "tcp", &lock_class_rw },
553 { "tcpinp", &lock_class_rw },
554 { "so_snd", &lock_class_mtx_sleep },
559 { "ddp_list_mtx", &lock_class_mtx_sleep },
560 { "ddp_mtx", &lock_class_mtx_sleep },
565 { "bpf global lock", &lock_class_mtx_sleep },
566 { "bpf interface lock", &lock_class_rw },
567 { "bpf cdev lock", &lock_class_mtx_sleep },
572 { "nfsd_mtx", &lock_class_mtx_sleep },
573 { "so_snd", &lock_class_mtx_sleep },
579 { "802.11 com lock", &lock_class_mtx_sleep},
584 { "network driver", &lock_class_mtx_sleep},
590 { "ng_node", &lock_class_mtx_sleep },
591 { "ng_worklist", &lock_class_mtx_sleep },
596 { "vm map (system)", &lock_class_mtx_sleep },
597 { "vm page queue", &lock_class_mtx_sleep },
598 { "vnode interlock", &lock_class_mtx_sleep },
599 { "cdev", &lock_class_mtx_sleep },
604 { "vm map (user)", &lock_class_sx },
605 { "vm object", &lock_class_mtx_sleep },
606 { "vm page", &lock_class_mtx_sleep },
607 { "vm page queue", &lock_class_mtx_sleep },
608 { "pmap pv global", &lock_class_rw },
609 { "pmap", &lock_class_mtx_sleep },
610 { "pmap pv list", &lock_class_rw },
611 { "vm page free queue", &lock_class_mtx_sleep },
614 * kqueue/VFS interaction
616 { "kqueue", &lock_class_mtx_sleep },
617 { "struct mount mtx", &lock_class_mtx_sleep },
618 { "vnode interlock", &lock_class_mtx_sleep },
623 { "dn->dn_mtx", &lock_class_sx },
624 { "dr->dt.di.dr_mtx", &lock_class_sx },
625 { "db->db_mtx", &lock_class_sx },
631 { "ap boot", &lock_class_mtx_spin },
633 { "rm.mutex_mtx", &lock_class_mtx_spin },
634 { "sio", &lock_class_mtx_spin },
635 { "scrlock", &lock_class_mtx_spin },
637 { "cy", &lock_class_mtx_spin },
640 { "pcib_mtx", &lock_class_mtx_spin },
641 { "rtc_mtx", &lock_class_mtx_spin },
643 { "scc_hwmtx", &lock_class_mtx_spin },
644 { "uart_hwmtx", &lock_class_mtx_spin },
645 { "fast_taskqueue", &lock_class_mtx_spin },
646 { "intr table", &lock_class_mtx_spin },
648 { "pmc-per-proc", &lock_class_mtx_spin },
650 { "process slock", &lock_class_mtx_spin },
651 { "sleepq chain", &lock_class_mtx_spin },
652 { "umtx lock", &lock_class_mtx_spin },
653 { "rm_spinlock", &lock_class_mtx_spin },
654 { "turnstile chain", &lock_class_mtx_spin },
655 { "turnstile lock", &lock_class_mtx_spin },
656 { "sched lock", &lock_class_mtx_spin },
657 { "td_contested", &lock_class_mtx_spin },
658 { "callout", &lock_class_mtx_spin },
659 { "entropy harvest mutex", &lock_class_mtx_spin },
660 { "syscons video lock", &lock_class_mtx_spin },
662 { "smp rendezvous", &lock_class_mtx_spin },
665 { "tlb0", &lock_class_mtx_spin },
670 { "intrcnt", &lock_class_mtx_spin },
671 { "icu", &lock_class_mtx_spin },
673 { "allpmaps", &lock_class_mtx_spin },
674 { "descriptor tables", &lock_class_mtx_spin },
676 { "clk", &lock_class_mtx_spin },
677 { "cpuset", &lock_class_mtx_spin },
678 { "mprof lock", &lock_class_mtx_spin },
679 { "zombie lock", &lock_class_mtx_spin },
680 { "ALD Queue", &lock_class_mtx_spin },
682 { "MCA spin lock", &lock_class_mtx_spin },
684 #if defined(__i386__) || defined(__amd64__)
685 { "pcicfg", &lock_class_mtx_spin },
686 { "NDIS thread lock", &lock_class_mtx_spin },
688 { "tw_osl_io_lock", &lock_class_mtx_spin },
689 { "tw_osl_q_lock", &lock_class_mtx_spin },
690 { "tw_cl_io_lock", &lock_class_mtx_spin },
691 { "tw_cl_intr_lock", &lock_class_mtx_spin },
692 { "tw_cl_gen_lock", &lock_class_mtx_spin },
694 { "pmc-leaf", &lock_class_mtx_spin },
696 { "blocked lock", &lock_class_mtx_spin },
703 * Pairs of locks which have been blessed
704 * Don't complain about order problems with blessed locks
706 static struct witness_blessed blessed_list[] = {
708 static int blessed_count =
709 sizeof(blessed_list) / sizeof(struct witness_blessed);
713 * This global is set to 0 once it becomes safe to use the witness code.
715 static int witness_cold = 1;
718 * This global is set to 1 once the static lock orders have been enrolled
719 * so that a warning can be issued for any spin locks enrolled later.
721 static int witness_spin_warn = 0;
723 /* Trim useless garbage from filenames. */
725 fixup_filename(const char *file)
730 while (strncmp(file, "../", 3) == 0)
736 * The WITNESS-enabled diagnostic code. Note that the witness code does
737 * assume that the early boot is single-threaded at least until after this
738 * routine is completed.
741 witness_initialize(void *dummy __unused)
743 struct lock_object *lock;
744 struct witness_order_list_entry *order;
745 struct witness *w, *w1;
748 w_data = malloc(sizeof (struct witness) * WITNESS_COUNT, M_WITNESS,
752 * We have to release Giant before initializing its witness
753 * structure so that WITNESS doesn't get confused.
756 mtx_assert(&Giant, MA_NOTOWNED);
758 CTR1(KTR_WITNESS, "%s: initializing witness", __func__);
759 mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET |
760 MTX_NOWITNESS | MTX_NOPROFILE);
761 for (i = WITNESS_COUNT - 1; i >= 0; i--) {
763 memset(w, 0, sizeof(*w));
764 w_data[i].w_index = i; /* Witness index never changes. */
767 KASSERT(STAILQ_FIRST(&w_free)->w_index == 0,
768 ("%s: Invalid list of free witness objects", __func__));
770 /* Witness with index 0 is not used to aid in debugging. */
771 STAILQ_REMOVE_HEAD(&w_free, w_list);
775 (sizeof(**w_rmatrix) * (WITNESS_COUNT+1) * (WITNESS_COUNT+1)));
777 for (i = 0; i < LOCK_CHILDCOUNT; i++)
778 witness_lock_list_free(&w_locklistdata[i]);
779 witness_init_hash_tables();
781 /* First add in all the specified order lists. */
782 for (order = order_lists; order->w_name != NULL; order++) {
783 w = enroll(order->w_name, order->w_class);
786 w->w_file = "order list";
787 for (order++; order->w_name != NULL; order++) {
788 w1 = enroll(order->w_name, order->w_class);
791 w1->w_file = "order list";
796 witness_spin_warn = 1;
798 /* Iterate through all locks and add them to witness. */
799 for (i = 0; pending_locks[i].wh_lock != NULL; i++) {
800 lock = pending_locks[i].wh_lock;
801 KASSERT(lock->lo_flags & LO_WITNESS,
802 ("%s: lock %s is on pending list but not LO_WITNESS",
803 __func__, lock->lo_name));
804 lock->lo_witness = enroll(pending_locks[i].wh_type,
808 /* Mark the witness code as being ready for use. */
813 SYSINIT(witness_init, SI_SUB_WITNESS, SI_ORDER_FIRST, witness_initialize,
817 witness_init(struct lock_object *lock, const char *type)
819 struct lock_class *class;
821 /* Various sanity checks. */
822 class = LOCK_CLASS(lock);
823 if ((lock->lo_flags & LO_RECURSABLE) != 0 &&
824 (class->lc_flags & LC_RECURSABLE) == 0)
825 panic("%s: lock (%s) %s can not be recursable", __func__,
826 class->lc_name, lock->lo_name);
827 if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
828 (class->lc_flags & LC_SLEEPABLE) == 0)
829 panic("%s: lock (%s) %s can not be sleepable", __func__,
830 class->lc_name, lock->lo_name);
831 if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
832 (class->lc_flags & LC_UPGRADABLE) == 0)
833 panic("%s: lock (%s) %s can not be upgradable", __func__,
834 class->lc_name, lock->lo_name);
837 * If we shouldn't watch this lock, then just clear lo_witness.
838 * Otherwise, if witness_cold is set, then it is too early to
839 * enroll this lock, so defer it to witness_initialize() by adding
840 * it to the pending_locks list. If it is not too early, then enroll
843 if (witness_watch < 1 || panicstr != NULL ||
844 (lock->lo_flags & LO_WITNESS) == 0)
845 lock->lo_witness = NULL;
846 else if (witness_cold) {
847 pending_locks[pending_cnt].wh_lock = lock;
848 pending_locks[pending_cnt++].wh_type = type;
849 if (pending_cnt > WITNESS_PENDLIST)
850 panic("%s: pending locks list is too small, bump it\n",
853 lock->lo_witness = enroll(type, class);
857 witness_destroy(struct lock_object *lock)
859 struct lock_class *class;
862 class = LOCK_CLASS(lock);
865 panic("lock (%s) %s destroyed while witness_cold",
866 class->lc_name, lock->lo_name);
868 /* XXX: need to verify that no one holds the lock */
869 if ((lock->lo_flags & LO_WITNESS) == 0 || lock->lo_witness == NULL)
871 w = lock->lo_witness;
873 mtx_lock_spin(&w_mtx);
874 MPASS(w->w_refcount > 0);
877 if (w->w_refcount == 0)
879 mtx_unlock_spin(&w_mtx);
884 witness_ddb_compute_levels(void)
889 * First clear all levels.
891 STAILQ_FOREACH(w, &w_all, w_list)
895 * Look for locks with no parents and level all their descendants.
897 STAILQ_FOREACH(w, &w_all, w_list) {
899 /* If the witness has ancestors (is not a root), skip it. */
900 if (w->w_num_ancestors > 0)
902 witness_ddb_level_descendants(w, 0);
907 witness_ddb_level_descendants(struct witness *w, int l)
911 if (w->w_ddb_level >= l)
917 for (i = 1; i <= w_max_used_index; i++) {
918 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
919 witness_ddb_level_descendants(&w_data[i], l);
924 witness_ddb_display_descendants(int(*prnt)(const char *fmt, ...),
925 struct witness *w, int indent)
929 for (i = 0; i < indent; i++)
931 prnt("%s (type: %s, depth: %d, active refs: %d)",
932 w->w_name, w->w_class->lc_name,
933 w->w_ddb_level, w->w_refcount);
934 if (w->w_displayed) {
935 prnt(" -- (already displayed)\n");
939 if (w->w_file != NULL && w->w_line != 0)
940 prnt(" -- last acquired @ %s:%d\n", fixup_filename(w->w_file),
943 prnt(" -- never acquired\n");
945 WITNESS_INDEX_ASSERT(w->w_index);
946 for (i = 1; i <= w_max_used_index; i++) {
949 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
950 witness_ddb_display_descendants(prnt, &w_data[i],
956 witness_ddb_display_list(int(*prnt)(const char *fmt, ...),
957 struct witness_list *list)
961 STAILQ_FOREACH(w, list, w_typelist) {
962 if (w->w_file == NULL || w->w_ddb_level > 0)
965 /* This lock has no anscestors - display its descendants. */
966 witness_ddb_display_descendants(prnt, w, 0);
973 witness_ddb_display(int(*prnt)(const char *fmt, ...))
977 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
978 witness_ddb_compute_levels();
980 /* Clear all the displayed flags. */
981 STAILQ_FOREACH(w, &w_all, w_list)
985 * First, handle sleep locks which have been acquired at least
988 prnt("Sleep locks:\n");
989 witness_ddb_display_list(prnt, &w_sleep);
994 * Now do spin locks which have been acquired at least once.
996 prnt("\nSpin locks:\n");
997 witness_ddb_display_list(prnt, &w_spin);
1002 * Finally, any locks which have not been acquired yet.
1004 prnt("\nLocks which were never acquired:\n");
1005 STAILQ_FOREACH(w, &w_all, w_list) {
1006 if (w->w_file != NULL || w->w_refcount == 0)
1008 prnt("%s (type: %s, depth: %d)\n", w->w_name,
1009 w->w_class->lc_name, w->w_ddb_level);
1017 witness_defineorder(struct lock_object *lock1, struct lock_object *lock2)
1020 if (witness_watch == -1 || panicstr != NULL)
1023 /* Require locks that witness knows about. */
1024 if (lock1 == NULL || lock1->lo_witness == NULL || lock2 == NULL ||
1025 lock2->lo_witness == NULL)
1028 mtx_assert(&w_mtx, MA_NOTOWNED);
1029 mtx_lock_spin(&w_mtx);
1032 * If we already have either an explicit or implied lock order that
1033 * is the other way around, then return an error.
1035 if (witness_watch &&
1036 isitmydescendant(lock2->lo_witness, lock1->lo_witness)) {
1037 mtx_unlock_spin(&w_mtx);
1041 /* Try to add the new order. */
1042 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1043 lock2->lo_witness->w_name, lock1->lo_witness->w_name);
1044 itismychild(lock1->lo_witness, lock2->lo_witness);
1045 mtx_unlock_spin(&w_mtx);
1050 witness_checkorder(struct lock_object *lock, int flags, const char *file,
1051 int line, struct lock_object *interlock)
1053 struct lock_list_entry *lock_list, *lle;
1054 struct lock_instance *lock1, *lock2, *plock;
1055 struct lock_class *class;
1056 struct witness *w, *w1;
1060 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL ||
1064 w = lock->lo_witness;
1065 class = LOCK_CLASS(lock);
1068 if (class->lc_flags & LC_SLEEPLOCK) {
1071 * Since spin locks include a critical section, this check
1072 * implicitly enforces a lock order of all sleep locks before
1075 if (td->td_critnest != 0 && !kdb_active)
1076 panic("blockable sleep lock (%s) %s @ %s:%d",
1077 class->lc_name, lock->lo_name,
1078 fixup_filename(file), line);
1081 * If this is the first lock acquired then just return as
1082 * no order checking is needed.
1084 lock_list = td->td_sleeplocks;
1085 if (lock_list == NULL || lock_list->ll_count == 0)
1090 * If this is the first lock, just return as no order
1091 * checking is needed. Avoid problems with thread
1092 * migration pinning the thread while checking if
1093 * spinlocks are held. If at least one spinlock is held
1094 * the thread is in a safe path and it is allowed to
1098 lock_list = PCPU_GET(spinlocks);
1099 if (lock_list == NULL || lock_list->ll_count == 0) {
1107 * Check to see if we are recursing on a lock we already own. If
1108 * so, make sure that we don't mismatch exclusive and shared lock
1111 lock1 = find_instance(lock_list, lock);
1112 if (lock1 != NULL) {
1113 if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
1114 (flags & LOP_EXCLUSIVE) == 0) {
1115 printf("shared lock of (%s) %s @ %s:%d\n",
1116 class->lc_name, lock->lo_name,
1117 fixup_filename(file), line);
1118 printf("while exclusively locked from %s:%d\n",
1119 fixup_filename(lock1->li_file), lock1->li_line);
1120 panic("share->excl");
1122 if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
1123 (flags & LOP_EXCLUSIVE) != 0) {
1124 printf("exclusive lock of (%s) %s @ %s:%d\n",
1125 class->lc_name, lock->lo_name,
1126 fixup_filename(file), line);
1127 printf("while share locked from %s:%d\n",
1128 fixup_filename(lock1->li_file), lock1->li_line);
1129 panic("excl->share");
1135 * Find the previously acquired lock, but ignore interlocks.
1137 plock = &lock_list->ll_children[lock_list->ll_count - 1];
1138 if (interlock != NULL && plock->li_lock == interlock) {
1139 if (lock_list->ll_count > 1)
1141 &lock_list->ll_children[lock_list->ll_count - 2];
1143 lle = lock_list->ll_next;
1146 * The interlock is the only lock we hold, so
1151 plock = &lle->ll_children[lle->ll_count - 1];
1156 * Try to perform most checks without a lock. If this succeeds we
1157 * can skip acquiring the lock and return success.
1159 w1 = plock->li_lock->lo_witness;
1160 if (witness_lock_order_check(w1, w))
1164 * Check for duplicate locks of the same type. Note that we only
1165 * have to check for this on the last lock we just acquired. Any
1166 * other cases will be caught as lock order violations.
1168 mtx_lock_spin(&w_mtx);
1169 witness_lock_order_add(w1, w);
1172 if (!(lock->lo_flags & LO_DUPOK) && !(flags & LOP_DUPOK) &&
1173 !(w_rmatrix[i][i] & WITNESS_REVERSAL)) {
1174 w_rmatrix[i][i] |= WITNESS_REVERSAL;
1176 mtx_unlock_spin(&w_mtx);
1178 "acquiring duplicate lock of same type: \"%s\"\n",
1180 printf(" 1st %s @ %s:%d\n", plock->li_lock->lo_name,
1181 fixup_filename(plock->li_file), plock->li_line);
1182 printf(" 2nd %s @ %s:%d\n", lock->lo_name,
1183 fixup_filename(file), line);
1184 witness_debugger(1);
1186 mtx_unlock_spin(&w_mtx);
1189 mtx_assert(&w_mtx, MA_OWNED);
1192 * If we know that the lock we are acquiring comes after
1193 * the lock we most recently acquired in the lock order tree,
1194 * then there is no need for any further checks.
1196 if (isitmychild(w1, w))
1199 for (j = 0, lle = lock_list; lle != NULL; lle = lle->ll_next) {
1200 for (i = lle->ll_count - 1; i >= 0; i--, j++) {
1202 MPASS(j < WITNESS_COUNT);
1203 lock1 = &lle->ll_children[i];
1206 * Ignore the interlock the first time we see it.
1208 if (interlock != NULL && interlock == lock1->li_lock) {
1214 * If this lock doesn't undergo witness checking,
1217 w1 = lock1->li_lock->lo_witness;
1219 KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
1220 ("lock missing witness structure"));
1225 * If we are locking Giant and this is a sleepable
1226 * lock, then skip it.
1228 if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0 &&
1229 lock == &Giant.lock_object)
1233 * If we are locking a sleepable lock and this lock
1234 * is Giant, then skip it.
1236 if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1237 lock1->li_lock == &Giant.lock_object)
1241 * If we are locking a sleepable lock and this lock
1242 * isn't sleepable, we want to treat it as a lock
1243 * order violation to enfore a general lock order of
1244 * sleepable locks before non-sleepable locks.
1246 if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1247 (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
1251 * If we are locking Giant and this is a non-sleepable
1252 * lock, then treat it as a reversal.
1254 if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0 &&
1255 lock == &Giant.lock_object)
1259 * Check the lock order hierarchy for a reveresal.
1261 if (!isitmydescendant(w, w1))
1266 * We have a lock order violation, check to see if it
1267 * is allowed or has already been yelled about.
1272 * If the lock order is blessed, just bail. We don't
1273 * look for other lock order violations though, which
1280 /* Bail if this violation is known */
1281 if (w_rmatrix[w1->w_index][w->w_index] & WITNESS_REVERSAL)
1284 /* Record this as a violation */
1285 w_rmatrix[w1->w_index][w->w_index] |= WITNESS_REVERSAL;
1286 w_rmatrix[w->w_index][w1->w_index] |= WITNESS_REVERSAL;
1287 w->w_reversed = w1->w_reversed = 1;
1288 witness_increment_graph_generation();
1289 mtx_unlock_spin(&w_mtx);
1292 * Ok, yell about it.
1294 if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1295 (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
1297 "lock order reversal: (sleepable after non-sleepable)\n");
1298 else if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0
1299 && lock == &Giant.lock_object)
1301 "lock order reversal: (Giant after non-sleepable)\n");
1303 printf("lock order reversal:\n");
1306 * Try to locate an earlier lock with
1307 * witness w in our list.
1310 lock2 = &lle->ll_children[i];
1311 MPASS(lock2->li_lock != NULL);
1312 if (lock2->li_lock->lo_witness == w)
1314 if (i == 0 && lle->ll_next != NULL) {
1316 i = lle->ll_count - 1;
1317 MPASS(i >= 0 && i < LOCK_NCHILDREN);
1322 printf(" 1st %p %s (%s) @ %s:%d\n",
1323 lock1->li_lock, lock1->li_lock->lo_name,
1324 w1->w_name, fixup_filename(lock1->li_file),
1326 printf(" 2nd %p %s (%s) @ %s:%d\n", lock,
1327 lock->lo_name, w->w_name,
1328 fixup_filename(file), line);
1330 printf(" 1st %p %s (%s) @ %s:%d\n",
1331 lock2->li_lock, lock2->li_lock->lo_name,
1332 lock2->li_lock->lo_witness->w_name,
1333 fixup_filename(lock2->li_file),
1335 printf(" 2nd %p %s (%s) @ %s:%d\n",
1336 lock1->li_lock, lock1->li_lock->lo_name,
1337 w1->w_name, fixup_filename(lock1->li_file),
1339 printf(" 3rd %p %s (%s) @ %s:%d\n", lock,
1340 lock->lo_name, w->w_name,
1341 fixup_filename(file), line);
1343 witness_debugger(1);
1349 * If requested, build a new lock order. However, don't build a new
1350 * relationship between a sleepable lock and Giant if it is in the
1351 * wrong direction. The correct lock order is that sleepable locks
1352 * always come before Giant.
1354 if (flags & LOP_NEWORDER &&
1355 !(plock->li_lock == &Giant.lock_object &&
1356 (lock->lo_flags & LO_SLEEPABLE) != 0)) {
1357 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1358 w->w_name, plock->li_lock->lo_witness->w_name);
1359 itismychild(plock->li_lock->lo_witness, w);
1362 mtx_unlock_spin(&w_mtx);
1366 witness_lock(struct lock_object *lock, int flags, const char *file, int line)
1368 struct lock_list_entry **lock_list, *lle;
1369 struct lock_instance *instance;
1373 if (witness_cold || witness_watch == -1 || lock->lo_witness == NULL ||
1376 w = lock->lo_witness;
1379 /* Determine lock list for this lock. */
1380 if (LOCK_CLASS(lock)->lc_flags & LC_SLEEPLOCK)
1381 lock_list = &td->td_sleeplocks;
1383 lock_list = PCPU_PTR(spinlocks);
1385 /* Check to see if we are recursing on a lock we already own. */
1386 instance = find_instance(*lock_list, lock);
1387 if (instance != NULL) {
1388 instance->li_flags++;
1389 CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
1390 td->td_proc->p_pid, lock->lo_name,
1391 instance->li_flags & LI_RECURSEMASK);
1392 instance->li_file = file;
1393 instance->li_line = line;
1397 /* Update per-witness last file and line acquire. */
1401 /* Find the next open lock instance in the list and fill it. */
1403 if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
1404 lle = witness_lock_list_get();
1407 lle->ll_next = *lock_list;
1408 CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
1409 td->td_proc->p_pid, lle);
1412 instance = &lle->ll_children[lle->ll_count++];
1413 instance->li_lock = lock;
1414 instance->li_line = line;
1415 instance->li_file = file;
1416 if ((flags & LOP_EXCLUSIVE) != 0)
1417 instance->li_flags = LI_EXCLUSIVE;
1419 instance->li_flags = 0;
1420 CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
1421 td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
1425 witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
1427 struct lock_instance *instance;
1428 struct lock_class *class;
1430 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1431 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
1433 class = LOCK_CLASS(lock);
1434 if (witness_watch) {
1435 if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1436 panic("upgrade of non-upgradable lock (%s) %s @ %s:%d",
1437 class->lc_name, lock->lo_name,
1438 fixup_filename(file), line);
1439 if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1440 panic("upgrade of non-sleep lock (%s) %s @ %s:%d",
1441 class->lc_name, lock->lo_name,
1442 fixup_filename(file), line);
1444 instance = find_instance(curthread->td_sleeplocks, lock);
1445 if (instance == NULL)
1446 panic("upgrade of unlocked lock (%s) %s @ %s:%d",
1447 class->lc_name, lock->lo_name,
1448 fixup_filename(file), line);
1449 if (witness_watch) {
1450 if ((instance->li_flags & LI_EXCLUSIVE) != 0)
1451 panic("upgrade of exclusive lock (%s) %s @ %s:%d",
1452 class->lc_name, lock->lo_name,
1453 fixup_filename(file), line);
1454 if ((instance->li_flags & LI_RECURSEMASK) != 0)
1455 panic("upgrade of recursed lock (%s) %s r=%d @ %s:%d",
1456 class->lc_name, lock->lo_name,
1457 instance->li_flags & LI_RECURSEMASK,
1458 fixup_filename(file), line);
1460 instance->li_flags |= LI_EXCLUSIVE;
1464 witness_downgrade(struct lock_object *lock, int flags, const char *file,
1467 struct lock_instance *instance;
1468 struct lock_class *class;
1470 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1471 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
1473 class = LOCK_CLASS(lock);
1474 if (witness_watch) {
1475 if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1476 panic("downgrade of non-upgradable lock (%s) %s @ %s:%d",
1477 class->lc_name, lock->lo_name,
1478 fixup_filename(file), line);
1479 if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1480 panic("downgrade of non-sleep lock (%s) %s @ %s:%d",
1481 class->lc_name, lock->lo_name,
1482 fixup_filename(file), line);
1484 instance = find_instance(curthread->td_sleeplocks, lock);
1485 if (instance == NULL)
1486 panic("downgrade of unlocked lock (%s) %s @ %s:%d",
1487 class->lc_name, lock->lo_name,
1488 fixup_filename(file), line);
1489 if (witness_watch) {
1490 if ((instance->li_flags & LI_EXCLUSIVE) == 0)
1491 panic("downgrade of shared lock (%s) %s @ %s:%d",
1492 class->lc_name, lock->lo_name,
1493 fixup_filename(file), line);
1494 if ((instance->li_flags & LI_RECURSEMASK) != 0)
1495 panic("downgrade of recursed lock (%s) %s r=%d @ %s:%d",
1496 class->lc_name, lock->lo_name,
1497 instance->li_flags & LI_RECURSEMASK,
1498 fixup_filename(file), line);
1500 instance->li_flags &= ~LI_EXCLUSIVE;
1504 witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
1506 struct lock_list_entry **lock_list, *lle;
1507 struct lock_instance *instance;
1508 struct lock_class *class;
1513 if (witness_cold || lock->lo_witness == NULL || panicstr != NULL)
1516 class = LOCK_CLASS(lock);
1518 /* Find lock instance associated with this lock. */
1519 if (class->lc_flags & LC_SLEEPLOCK)
1520 lock_list = &td->td_sleeplocks;
1522 lock_list = PCPU_PTR(spinlocks);
1524 for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
1525 for (i = 0; i < (*lock_list)->ll_count; i++) {
1526 instance = &(*lock_list)->ll_children[i];
1527 if (instance->li_lock == lock)
1532 * When disabling WITNESS through witness_watch we could end up in
1533 * having registered locks in the td_sleeplocks queue.
1534 * We have to make sure we flush these queues, so just search for
1535 * eventual register locks and remove them.
1537 if (witness_watch > 0)
1538 panic("lock (%s) %s not locked @ %s:%d", class->lc_name,
1539 lock->lo_name, fixup_filename(file), line);
1544 /* First, check for shared/exclusive mismatches. */
1545 if ((instance->li_flags & LI_EXCLUSIVE) != 0 && witness_watch > 0 &&
1546 (flags & LOP_EXCLUSIVE) == 0) {
1547 printf("shared unlock of (%s) %s @ %s:%d\n", class->lc_name,
1548 lock->lo_name, fixup_filename(file), line);
1549 printf("while exclusively locked from %s:%d\n",
1550 fixup_filename(instance->li_file), instance->li_line);
1551 panic("excl->ushare");
1553 if ((instance->li_flags & LI_EXCLUSIVE) == 0 && witness_watch > 0 &&
1554 (flags & LOP_EXCLUSIVE) != 0) {
1555 printf("exclusive unlock of (%s) %s @ %s:%d\n", class->lc_name,
1556 lock->lo_name, fixup_filename(file), line);
1557 printf("while share locked from %s:%d\n",
1558 fixup_filename(instance->li_file),
1560 panic("share->uexcl");
1562 /* If we are recursed, unrecurse. */
1563 if ((instance->li_flags & LI_RECURSEMASK) > 0) {
1564 CTR4(KTR_WITNESS, "%s: pid %d unrecursed on %s r=%d", __func__,
1565 td->td_proc->p_pid, instance->li_lock->lo_name,
1566 instance->li_flags);
1567 instance->li_flags--;
1570 /* The lock is now being dropped, check for NORELEASE flag */
1571 if ((instance->li_flags & LI_NORELEASE) != 0 && witness_watch > 0) {
1572 printf("forbidden unlock of (%s) %s @ %s:%d\n", class->lc_name,
1573 lock->lo_name, fixup_filename(file), line);
1574 panic("lock marked norelease");
1577 /* Otherwise, remove this item from the list. */
1579 CTR4(KTR_WITNESS, "%s: pid %d removed %s from lle[%d]", __func__,
1580 td->td_proc->p_pid, instance->li_lock->lo_name,
1581 (*lock_list)->ll_count - 1);
1582 for (j = i; j < (*lock_list)->ll_count - 1; j++)
1583 (*lock_list)->ll_children[j] =
1584 (*lock_list)->ll_children[j + 1];
1585 (*lock_list)->ll_count--;
1589 * In order to reduce contention on w_mtx, we want to keep always an
1590 * head object into lists so that frequent allocation from the
1591 * free witness pool (and subsequent locking) is avoided.
1592 * In order to maintain the current code simple, when the head
1593 * object is totally unloaded it means also that we do not have
1594 * further objects in the list, so the list ownership needs to be
1595 * hand over to another object if the current head needs to be freed.
1597 if ((*lock_list)->ll_count == 0) {
1598 if (*lock_list == lle) {
1599 if (lle->ll_next == NULL)
1603 *lock_list = lle->ll_next;
1604 CTR3(KTR_WITNESS, "%s: pid %d removed lle %p", __func__,
1605 td->td_proc->p_pid, lle);
1606 witness_lock_list_free(lle);
1611 witness_thread_exit(struct thread *td)
1613 struct lock_list_entry *lle;
1616 lle = td->td_sleeplocks;
1617 if (lle == NULL || panicstr != NULL)
1619 if (lle->ll_count != 0) {
1620 for (n = 0; lle != NULL; lle = lle->ll_next)
1621 for (i = lle->ll_count - 1; i >= 0; i--) {
1623 printf("Thread %p exiting with the following locks held:\n",
1626 witness_list_lock(&lle->ll_children[i], printf);
1629 panic("Thread %p cannot exit while holding sleeplocks\n", td);
1631 witness_lock_list_free(lle);
1635 * Warn if any locks other than 'lock' are held. Flags can be passed in to
1636 * exempt Giant and sleepable locks from the checks as well. If any
1637 * non-exempt locks are held, then a supplied message is printed to the
1638 * console along with a list of the offending locks. If indicated in the
1639 * flags then a failure results in a panic as well.
1642 witness_warn(int flags, struct lock_object *lock, const char *fmt, ...)
1644 struct lock_list_entry *lock_list, *lle;
1645 struct lock_instance *lock1;
1650 if (witness_cold || witness_watch < 1 || panicstr != NULL)
1654 for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next)
1655 for (i = lle->ll_count - 1; i >= 0; i--) {
1656 lock1 = &lle->ll_children[i];
1657 if (lock1->li_lock == lock)
1659 if (flags & WARN_GIANTOK &&
1660 lock1->li_lock == &Giant.lock_object)
1662 if (flags & WARN_SLEEPOK &&
1663 (lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0)
1669 printf(" with the following");
1670 if (flags & WARN_SLEEPOK)
1671 printf(" non-sleepable");
1672 printf(" locks held:\n");
1675 witness_list_lock(lock1, printf);
1679 * Pin the thread in order to avoid problems with thread migration.
1680 * Once that all verifies are passed about spinlocks ownership,
1681 * the thread is in a safe path and it can be unpinned.
1684 lock_list = PCPU_GET(spinlocks);
1685 if (lock_list != NULL && lock_list->ll_count != 0) {
1689 * We should only have one spinlock and as long as
1690 * the flags cannot match for this locks class,
1691 * check if the first spinlock is the one curthread
1694 lock1 = &lock_list->ll_children[lock_list->ll_count - 1];
1695 if (lock_list->ll_count == 1 && lock_list->ll_next == NULL &&
1696 lock1->li_lock == lock && n == 0)
1702 printf(" with the following");
1703 if (flags & WARN_SLEEPOK)
1704 printf(" non-sleepable");
1705 printf(" locks held:\n");
1706 n += witness_list_locks(&lock_list, printf);
1709 if (flags & WARN_PANIC && n)
1710 panic("%s", __func__);
1712 witness_debugger(n);
1717 witness_file(struct lock_object *lock)
1721 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1723 w = lock->lo_witness;
1728 witness_line(struct lock_object *lock)
1732 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1734 w = lock->lo_witness;
1738 static struct witness *
1739 enroll(const char *description, struct lock_class *lock_class)
1742 struct witness_list *typelist;
1744 MPASS(description != NULL);
1746 if (witness_watch == -1 || panicstr != NULL)
1748 if ((lock_class->lc_flags & LC_SPINLOCK)) {
1749 if (witness_skipspin)
1753 } else if ((lock_class->lc_flags & LC_SLEEPLOCK))
1754 typelist = &w_sleep;
1756 panic("lock class %s is not sleep or spin",
1757 lock_class->lc_name);
1759 mtx_lock_spin(&w_mtx);
1760 w = witness_hash_get(description);
1763 if ((w = witness_get()) == NULL)
1765 MPASS(strlen(description) < MAX_W_NAME);
1766 strcpy(w->w_name, description);
1767 w->w_class = lock_class;
1769 STAILQ_INSERT_HEAD(&w_all, w, w_list);
1770 if (lock_class->lc_flags & LC_SPINLOCK) {
1771 STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
1773 } else if (lock_class->lc_flags & LC_SLEEPLOCK) {
1774 STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
1778 /* Insert new witness into the hash */
1779 witness_hash_put(w);
1780 witness_increment_graph_generation();
1781 mtx_unlock_spin(&w_mtx);
1785 mtx_unlock_spin(&w_mtx);
1786 if (lock_class != w->w_class)
1788 "lock (%s) %s does not match earlier (%s) lock",
1789 description, lock_class->lc_name,
1790 w->w_class->lc_name);
1795 depart(struct witness *w)
1797 struct witness_list *list;
1799 MPASS(w->w_refcount == 0);
1800 if (w->w_class->lc_flags & LC_SLEEPLOCK) {
1808 * Set file to NULL as it may point into a loadable module.
1812 witness_increment_graph_generation();
1817 adopt(struct witness *parent, struct witness *child)
1821 if (witness_cold == 0)
1822 mtx_assert(&w_mtx, MA_OWNED);
1824 /* If the relationship is already known, there's no work to be done. */
1825 if (isitmychild(parent, child))
1828 /* When the structure of the graph changes, bump up the generation. */
1829 witness_increment_graph_generation();
1832 * The hard part ... create the direct relationship, then propagate all
1833 * indirect relationships.
1835 pi = parent->w_index;
1836 ci = child->w_index;
1837 WITNESS_INDEX_ASSERT(pi);
1838 WITNESS_INDEX_ASSERT(ci);
1840 w_rmatrix[pi][ci] |= WITNESS_PARENT;
1841 w_rmatrix[ci][pi] |= WITNESS_CHILD;
1844 * If parent was not already an ancestor of child,
1845 * then we increment the descendant and ancestor counters.
1847 if ((w_rmatrix[pi][ci] & WITNESS_ANCESTOR) == 0) {
1848 parent->w_num_descendants++;
1849 child->w_num_ancestors++;
1853 * Find each ancestor of 'pi'. Note that 'pi' itself is counted as
1854 * an ancestor of 'pi' during this loop.
1856 for (i = 1; i <= w_max_used_index; i++) {
1857 if ((w_rmatrix[i][pi] & WITNESS_ANCESTOR_MASK) == 0 &&
1861 /* Find each descendant of 'i' and mark it as a descendant. */
1862 for (j = 1; j <= w_max_used_index; j++) {
1865 * Skip children that are already marked as
1866 * descendants of 'i'.
1868 if (w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK)
1872 * We are only interested in descendants of 'ci'. Note
1873 * that 'ci' itself is counted as a descendant of 'ci'.
1875 if ((w_rmatrix[ci][j] & WITNESS_ANCESTOR_MASK) == 0 &&
1878 w_rmatrix[i][j] |= WITNESS_ANCESTOR;
1879 w_rmatrix[j][i] |= WITNESS_DESCENDANT;
1880 w_data[i].w_num_descendants++;
1881 w_data[j].w_num_ancestors++;
1884 * Make sure we aren't marking a node as both an
1885 * ancestor and descendant. We should have caught
1886 * this as a lock order reversal earlier.
1888 if ((w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK) &&
1889 (w_rmatrix[i][j] & WITNESS_DESCENDANT_MASK)) {
1890 printf("witness rmatrix paradox! [%d][%d]=%d "
1891 "both ancestor and descendant\n",
1892 i, j, w_rmatrix[i][j]);
1894 printf("Witness disabled.\n");
1897 if ((w_rmatrix[j][i] & WITNESS_ANCESTOR_MASK) &&
1898 (w_rmatrix[j][i] & WITNESS_DESCENDANT_MASK)) {
1899 printf("witness rmatrix paradox! [%d][%d]=%d "
1900 "both ancestor and descendant\n",
1901 j, i, w_rmatrix[j][i]);
1903 printf("Witness disabled.\n");
1911 itismychild(struct witness *parent, struct witness *child)
1914 MPASS(child != NULL && parent != NULL);
1915 if (witness_cold == 0)
1916 mtx_assert(&w_mtx, MA_OWNED);
1918 if (!witness_lock_type_equal(parent, child)) {
1919 if (witness_cold == 0)
1920 mtx_unlock_spin(&w_mtx);
1921 panic("%s: parent \"%s\" (%s) and child \"%s\" (%s) are not "
1922 "the same lock type", __func__, parent->w_name,
1923 parent->w_class->lc_name, child->w_name,
1924 child->w_class->lc_name);
1926 adopt(parent, child);
1930 * Generic code for the isitmy*() functions. The rmask parameter is the
1931 * expected relationship of w1 to w2.
1934 _isitmyx(struct witness *w1, struct witness *w2, int rmask, const char *fname)
1936 unsigned char r1, r2;
1941 WITNESS_INDEX_ASSERT(i1);
1942 WITNESS_INDEX_ASSERT(i2);
1943 r1 = w_rmatrix[i1][i2] & WITNESS_RELATED_MASK;
1944 r2 = w_rmatrix[i2][i1] & WITNESS_RELATED_MASK;
1946 /* The flags on one better be the inverse of the flags on the other */
1947 if (!((WITNESS_ATOD(r1) == r2 && WITNESS_DTOA(r2) == r1) ||
1948 (WITNESS_DTOA(r1) == r2 && WITNESS_ATOD(r2) == r1))) {
1949 printf("%s: rmatrix mismatch between %s (index %d) and %s "
1950 "(index %d): w_rmatrix[%d][%d] == %hhx but "
1951 "w_rmatrix[%d][%d] == %hhx\n",
1952 fname, w1->w_name, i1, w2->w_name, i2, i1, i2, r1,
1955 printf("Witness disabled.\n");
1958 return (r1 & rmask);
1962 * Checks if @child is a direct child of @parent.
1965 isitmychild(struct witness *parent, struct witness *child)
1968 return (_isitmyx(parent, child, WITNESS_PARENT, __func__));
1972 * Checks if @descendant is a direct or inderect descendant of @ancestor.
1975 isitmydescendant(struct witness *ancestor, struct witness *descendant)
1978 return (_isitmyx(ancestor, descendant, WITNESS_ANCESTOR_MASK,
1984 blessed(struct witness *w1, struct witness *w2)
1987 struct witness_blessed *b;
1989 for (i = 0; i < blessed_count; i++) {
1990 b = &blessed_list[i];
1991 if (strcmp(w1->w_name, b->b_lock1) == 0) {
1992 if (strcmp(w2->w_name, b->b_lock2) == 0)
1996 if (strcmp(w1->w_name, b->b_lock2) == 0)
1997 if (strcmp(w2->w_name, b->b_lock1) == 0)
2004 static struct witness *
2010 if (witness_cold == 0)
2011 mtx_assert(&w_mtx, MA_OWNED);
2013 if (witness_watch == -1) {
2014 mtx_unlock_spin(&w_mtx);
2017 if (STAILQ_EMPTY(&w_free)) {
2019 mtx_unlock_spin(&w_mtx);
2020 printf("WITNESS: unable to allocate a new witness object\n");
2023 w = STAILQ_FIRST(&w_free);
2024 STAILQ_REMOVE_HEAD(&w_free, w_list);
2027 MPASS(index > 0 && index == w_max_used_index+1 &&
2028 index < WITNESS_COUNT);
2029 bzero(w, sizeof(*w));
2031 if (index > w_max_used_index)
2032 w_max_used_index = index;
2037 witness_free(struct witness *w)
2040 STAILQ_INSERT_HEAD(&w_free, w, w_list);
2044 static struct lock_list_entry *
2045 witness_lock_list_get(void)
2047 struct lock_list_entry *lle;
2049 if (witness_watch == -1)
2051 mtx_lock_spin(&w_mtx);
2052 lle = w_lock_list_free;
2055 mtx_unlock_spin(&w_mtx);
2056 printf("%s: witness exhausted\n", __func__);
2059 w_lock_list_free = lle->ll_next;
2060 mtx_unlock_spin(&w_mtx);
2061 bzero(lle, sizeof(*lle));
2066 witness_lock_list_free(struct lock_list_entry *lle)
2069 mtx_lock_spin(&w_mtx);
2070 lle->ll_next = w_lock_list_free;
2071 w_lock_list_free = lle;
2072 mtx_unlock_spin(&w_mtx);
2075 static struct lock_instance *
2076 find_instance(struct lock_list_entry *list, struct lock_object *lock)
2078 struct lock_list_entry *lle;
2079 struct lock_instance *instance;
2082 for (lle = list; lle != NULL; lle = lle->ll_next)
2083 for (i = lle->ll_count - 1; i >= 0; i--) {
2084 instance = &lle->ll_children[i];
2085 if (instance->li_lock == lock)
2092 witness_list_lock(struct lock_instance *instance,
2093 int (*prnt)(const char *fmt, ...))
2095 struct lock_object *lock;
2097 lock = instance->li_lock;
2098 prnt("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ?
2099 "exclusive" : "shared", LOCK_CLASS(lock)->lc_name, lock->lo_name);
2100 if (lock->lo_witness->w_name != lock->lo_name)
2101 prnt(" (%s)", lock->lo_witness->w_name);
2102 prnt(" r = %d (%p) locked @ %s:%d\n",
2103 instance->li_flags & LI_RECURSEMASK, lock,
2104 fixup_filename(instance->li_file), instance->li_line);
2109 witness_thread_has_locks(struct thread *td)
2112 if (td->td_sleeplocks == NULL)
2114 return (td->td_sleeplocks->ll_count != 0);
2118 witness_proc_has_locks(struct proc *p)
2122 FOREACH_THREAD_IN_PROC(p, td) {
2123 if (witness_thread_has_locks(td))
2131 witness_list_locks(struct lock_list_entry **lock_list,
2132 int (*prnt)(const char *fmt, ...))
2134 struct lock_list_entry *lle;
2138 for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
2139 for (i = lle->ll_count - 1; i >= 0; i--) {
2140 witness_list_lock(&lle->ll_children[i], prnt);
2147 * This is a bit risky at best. We call this function when we have timed
2148 * out acquiring a spin lock, and we assume that the other CPU is stuck
2149 * with this lock held. So, we go groveling around in the other CPU's
2150 * per-cpu data to try to find the lock instance for this spin lock to
2151 * see when it was last acquired.
2154 witness_display_spinlock(struct lock_object *lock, struct thread *owner,
2155 int (*prnt)(const char *fmt, ...))
2157 struct lock_instance *instance;
2160 if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU)
2162 pc = pcpu_find(owner->td_oncpu);
2163 instance = find_instance(pc->pc_spinlocks, lock);
2164 if (instance != NULL)
2165 witness_list_lock(instance, prnt);
2169 witness_save(struct lock_object *lock, const char **filep, int *linep)
2171 struct lock_list_entry *lock_list;
2172 struct lock_instance *instance;
2173 struct lock_class *class;
2176 * This function is used independently in locking code to deal with
2177 * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2180 if (SCHEDULER_STOPPED())
2182 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2183 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2185 class = LOCK_CLASS(lock);
2186 if (class->lc_flags & LC_SLEEPLOCK)
2187 lock_list = curthread->td_sleeplocks;
2189 if (witness_skipspin)
2191 lock_list = PCPU_GET(spinlocks);
2193 instance = find_instance(lock_list, lock);
2194 if (instance == NULL)
2195 panic("%s: lock (%s) %s not locked", __func__,
2196 class->lc_name, lock->lo_name);
2197 *filep = instance->li_file;
2198 *linep = instance->li_line;
2202 witness_restore(struct lock_object *lock, const char *file, int line)
2204 struct lock_list_entry *lock_list;
2205 struct lock_instance *instance;
2206 struct lock_class *class;
2209 * This function is used independently in locking code to deal with
2210 * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2213 if (SCHEDULER_STOPPED())
2215 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2216 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2218 class = LOCK_CLASS(lock);
2219 if (class->lc_flags & LC_SLEEPLOCK)
2220 lock_list = curthread->td_sleeplocks;
2222 if (witness_skipspin)
2224 lock_list = PCPU_GET(spinlocks);
2226 instance = find_instance(lock_list, lock);
2227 if (instance == NULL)
2228 panic("%s: lock (%s) %s not locked", __func__,
2229 class->lc_name, lock->lo_name);
2230 lock->lo_witness->w_file = file;
2231 lock->lo_witness->w_line = line;
2232 instance->li_file = file;
2233 instance->li_line = line;
2237 witness_assert(struct lock_object *lock, int flags, const char *file, int line)
2239 #ifdef INVARIANT_SUPPORT
2240 struct lock_instance *instance;
2241 struct lock_class *class;
2243 if (lock->lo_witness == NULL || witness_watch < 1 || panicstr != NULL)
2245 class = LOCK_CLASS(lock);
2246 if ((class->lc_flags & LC_SLEEPLOCK) != 0)
2247 instance = find_instance(curthread->td_sleeplocks, lock);
2248 else if ((class->lc_flags & LC_SPINLOCK) != 0)
2249 instance = find_instance(PCPU_GET(spinlocks), lock);
2251 panic("Lock (%s) %s is not sleep or spin!",
2252 class->lc_name, lock->lo_name);
2256 if (instance != NULL)
2257 panic("Lock (%s) %s locked @ %s:%d.",
2258 class->lc_name, lock->lo_name,
2259 fixup_filename(file), line);
2262 case LA_LOCKED | LA_RECURSED:
2263 case LA_LOCKED | LA_NOTRECURSED:
2265 case LA_SLOCKED | LA_RECURSED:
2266 case LA_SLOCKED | LA_NOTRECURSED:
2268 case LA_XLOCKED | LA_RECURSED:
2269 case LA_XLOCKED | LA_NOTRECURSED:
2270 if (instance == NULL) {
2271 panic("Lock (%s) %s not locked @ %s:%d.",
2272 class->lc_name, lock->lo_name,
2273 fixup_filename(file), line);
2276 if ((flags & LA_XLOCKED) != 0 &&
2277 (instance->li_flags & LI_EXCLUSIVE) == 0)
2278 panic("Lock (%s) %s not exclusively locked @ %s:%d.",
2279 class->lc_name, lock->lo_name,
2280 fixup_filename(file), line);
2281 if ((flags & LA_SLOCKED) != 0 &&
2282 (instance->li_flags & LI_EXCLUSIVE) != 0)
2283 panic("Lock (%s) %s exclusively locked @ %s:%d.",
2284 class->lc_name, lock->lo_name,
2285 fixup_filename(file), line);
2286 if ((flags & LA_RECURSED) != 0 &&
2287 (instance->li_flags & LI_RECURSEMASK) == 0)
2288 panic("Lock (%s) %s not recursed @ %s:%d.",
2289 class->lc_name, lock->lo_name,
2290 fixup_filename(file), line);
2291 if ((flags & LA_NOTRECURSED) != 0 &&
2292 (instance->li_flags & LI_RECURSEMASK) != 0)
2293 panic("Lock (%s) %s recursed @ %s:%d.",
2294 class->lc_name, lock->lo_name,
2295 fixup_filename(file), line);
2298 panic("Invalid lock assertion at %s:%d.",
2299 fixup_filename(file), line);
2302 #endif /* INVARIANT_SUPPORT */
2306 witness_setflag(struct lock_object *lock, int flag, int set)
2308 struct lock_list_entry *lock_list;
2309 struct lock_instance *instance;
2310 struct lock_class *class;
2312 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2314 class = LOCK_CLASS(lock);
2315 if (class->lc_flags & LC_SLEEPLOCK)
2316 lock_list = curthread->td_sleeplocks;
2318 if (witness_skipspin)
2320 lock_list = PCPU_GET(spinlocks);
2322 instance = find_instance(lock_list, lock);
2323 if (instance == NULL)
2324 panic("%s: lock (%s) %s not locked", __func__,
2325 class->lc_name, lock->lo_name);
2328 instance->li_flags |= flag;
2330 instance->li_flags &= ~flag;
2334 witness_norelease(struct lock_object *lock)
2337 witness_setflag(lock, LI_NORELEASE, 1);
2341 witness_releaseok(struct lock_object *lock)
2344 witness_setflag(lock, LI_NORELEASE, 0);
2349 witness_ddb_list(struct thread *td)
2352 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2353 KASSERT(kdb_active, ("%s: not in the debugger", __func__));
2355 if (witness_watch < 1)
2358 witness_list_locks(&td->td_sleeplocks, db_printf);
2361 * We only handle spinlocks if td == curthread. This is somewhat broken
2362 * if td is currently executing on some other CPU and holds spin locks
2363 * as we won't display those locks. If we had a MI way of getting
2364 * the per-cpu data for a given cpu then we could use
2365 * td->td_oncpu to get the list of spinlocks for this thread
2368 * That still wouldn't really fix this unless we locked the scheduler
2369 * lock or stopped the other CPU to make sure it wasn't changing the
2370 * list out from under us. It is probably best to just not try to
2371 * handle threads on other CPU's for now.
2373 if (td == curthread && PCPU_GET(spinlocks) != NULL)
2374 witness_list_locks(PCPU_PTR(spinlocks), db_printf);
2377 DB_SHOW_COMMAND(locks, db_witness_list)
2382 td = db_lookup_thread(addr, TRUE);
2385 witness_ddb_list(td);
2388 DB_SHOW_ALL_COMMAND(locks, db_witness_list_all)
2394 * It would be nice to list only threads and processes that actually
2395 * held sleep locks, but that information is currently not exported
2398 FOREACH_PROC_IN_SYSTEM(p) {
2399 if (!witness_proc_has_locks(p))
2401 FOREACH_THREAD_IN_PROC(p, td) {
2402 if (!witness_thread_has_locks(td))
2404 db_printf("Process %d (%s) thread %p (%d)\n", p->p_pid,
2405 p->p_comm, td, td->td_tid);
2406 witness_ddb_list(td);
2412 DB_SHOW_ALIAS(alllocks, db_witness_list_all)
2414 DB_SHOW_COMMAND(witness, db_witness_display)
2417 witness_ddb_display(db_printf);
2422 sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS)
2424 struct witness_lock_order_data *data1, *data2, *tmp_data1, *tmp_data2;
2425 struct witness *tmp_w1, *tmp_w2, *w1, *w2;
2427 u_int w_rmatrix1, w_rmatrix2;
2428 int error, generation, i, j;
2434 if (witness_watch < 1) {
2435 error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2439 error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2443 sb = sbuf_new(NULL, NULL, BADSTACK_SBUF_SIZE, SBUF_AUTOEXTEND);
2447 /* Allocate and init temporary storage space. */
2448 tmp_w1 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2449 tmp_w2 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2450 tmp_data1 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2452 tmp_data2 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2454 stack_zero(&tmp_data1->wlod_stack);
2455 stack_zero(&tmp_data2->wlod_stack);
2458 mtx_lock_spin(&w_mtx);
2459 generation = w_generation;
2460 mtx_unlock_spin(&w_mtx);
2461 sbuf_printf(sb, "Number of known direct relationships is %d\n",
2462 w_lohash.wloh_count);
2463 for (i = 1; i < w_max_used_index; i++) {
2464 mtx_lock_spin(&w_mtx);
2465 if (generation != w_generation) {
2466 mtx_unlock_spin(&w_mtx);
2468 /* The graph has changed, try again. */
2475 if (w1->w_reversed == 0) {
2476 mtx_unlock_spin(&w_mtx);
2480 /* Copy w1 locally so we can release the spin lock. */
2482 mtx_unlock_spin(&w_mtx);
2484 if (tmp_w1->w_reversed == 0)
2486 for (j = 1; j < w_max_used_index; j++) {
2487 if ((w_rmatrix[i][j] & WITNESS_REVERSAL) == 0 || i > j)
2490 mtx_lock_spin(&w_mtx);
2491 if (generation != w_generation) {
2492 mtx_unlock_spin(&w_mtx);
2494 /* The graph has changed, try again. */
2501 data1 = witness_lock_order_get(w1, w2);
2502 data2 = witness_lock_order_get(w2, w1);
2505 * Copy information locally so we can release the
2509 w_rmatrix1 = (unsigned int)w_rmatrix[i][j];
2510 w_rmatrix2 = (unsigned int)w_rmatrix[j][i];
2513 stack_zero(&tmp_data1->wlod_stack);
2514 stack_copy(&data1->wlod_stack,
2515 &tmp_data1->wlod_stack);
2517 if (data2 && data2 != data1) {
2518 stack_zero(&tmp_data2->wlod_stack);
2519 stack_copy(&data2->wlod_stack,
2520 &tmp_data2->wlod_stack);
2522 mtx_unlock_spin(&w_mtx);
2525 "\nLock order reversal between \"%s\"(%s) and \"%s\"(%s)!\n",
2526 tmp_w1->w_name, tmp_w1->w_class->lc_name,
2527 tmp_w2->w_name, tmp_w2->w_class->lc_name);
2530 "w_rmatrix[%s][%s] == %x, w_rmatrix[%s][%s] == %x\n",
2531 tmp_w1->name, tmp_w2->w_name, w_rmatrix1,
2532 tmp_w2->name, tmp_w1->w_name, w_rmatrix2);
2536 "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2537 tmp_w1->w_name, tmp_w1->w_class->lc_name,
2538 tmp_w2->w_name, tmp_w2->w_class->lc_name);
2539 stack_sbuf_print(sb, &tmp_data1->wlod_stack);
2540 sbuf_printf(sb, "\n");
2542 if (data2 && data2 != data1) {
2544 "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2545 tmp_w2->w_name, tmp_w2->w_class->lc_name,
2546 tmp_w1->w_name, tmp_w1->w_class->lc_name);
2547 stack_sbuf_print(sb, &tmp_data2->wlod_stack);
2548 sbuf_printf(sb, "\n");
2552 mtx_lock_spin(&w_mtx);
2553 if (generation != w_generation) {
2554 mtx_unlock_spin(&w_mtx);
2557 * The graph changed while we were printing stack data,
2564 mtx_unlock_spin(&w_mtx);
2566 /* Free temporary storage space. */
2567 free(tmp_data1, M_TEMP);
2568 free(tmp_data2, M_TEMP);
2569 free(tmp_w1, M_TEMP);
2570 free(tmp_w2, M_TEMP);
2573 error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1);
2580 sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS)
2586 if (witness_watch < 1) {
2587 error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2591 error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2596 error = sysctl_wire_old_buffer(req, 0);
2599 sb = sbuf_new_for_sysctl(NULL, NULL, FULLGRAPH_SBUF_SIZE, req);
2602 sbuf_printf(sb, "\n");
2604 mtx_lock_spin(&w_mtx);
2605 STAILQ_FOREACH(w, &w_all, w_list)
2607 STAILQ_FOREACH(w, &w_all, w_list)
2608 witness_add_fullgraph(sb, w);
2609 mtx_unlock_spin(&w_mtx);
2612 * Close the sbuf and return to userland.
2614 error = sbuf_finish(sb);
2621 sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS)
2625 value = witness_watch;
2626 error = sysctl_handle_int(oidp, &value, 0, req);
2627 if (error != 0 || req->newptr == NULL)
2629 if (value > 1 || value < -1 ||
2630 (witness_watch == -1 && value != witness_watch))
2632 witness_watch = value;
2637 witness_add_fullgraph(struct sbuf *sb, struct witness *w)
2641 if (w->w_displayed != 0 || (w->w_file == NULL && w->w_line == 0))
2645 WITNESS_INDEX_ASSERT(w->w_index);
2646 for (i = 1; i <= w_max_used_index; i++) {
2647 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT) {
2648 sbuf_printf(sb, "\"%s\",\"%s\"\n", w->w_name,
2650 witness_add_fullgraph(sb, &w_data[i]);
2656 * A simple hash function. Takes a key pointer and a key size. If size == 0,
2657 * interprets the key as a string and reads until the null
2658 * terminator. Otherwise, reads the first size bytes. Returns an unsigned 32-bit
2659 * hash value computed from the key.
2662 witness_hash_djb2(const uint8_t *key, uint32_t size)
2664 unsigned int hash = 5381;
2667 /* hash = hash * 33 + key[i] */
2669 for (i = 0; i < size; i++)
2670 hash = ((hash << 5) + hash) + (unsigned int)key[i];
2672 for (i = 0; key[i] != 0; i++)
2673 hash = ((hash << 5) + hash) + (unsigned int)key[i];
2680 * Initializes the two witness hash tables. Called exactly once from
2681 * witness_initialize().
2684 witness_init_hash_tables(void)
2688 MPASS(witness_cold);
2690 /* Initialize the hash tables. */
2691 for (i = 0; i < WITNESS_HASH_SIZE; i++)
2692 w_hash.wh_array[i] = NULL;
2694 w_hash.wh_size = WITNESS_HASH_SIZE;
2695 w_hash.wh_count = 0;
2697 /* Initialize the lock order data hash. */
2699 for (i = 0; i < WITNESS_LO_DATA_COUNT; i++) {
2700 memset(&w_lodata[i], 0, sizeof(w_lodata[i]));
2701 w_lodata[i].wlod_next = w_lofree;
2702 w_lofree = &w_lodata[i];
2704 w_lohash.wloh_size = WITNESS_LO_HASH_SIZE;
2705 w_lohash.wloh_count = 0;
2706 for (i = 0; i < WITNESS_LO_HASH_SIZE; i++)
2707 w_lohash.wloh_array[i] = NULL;
2710 static struct witness *
2711 witness_hash_get(const char *key)
2717 if (witness_cold == 0)
2718 mtx_assert(&w_mtx, MA_OWNED);
2719 hash = witness_hash_djb2(key, 0) % w_hash.wh_size;
2720 w = w_hash.wh_array[hash];
2722 if (strcmp(w->w_name, key) == 0)
2732 witness_hash_put(struct witness *w)
2737 MPASS(w->w_name != NULL);
2738 if (witness_cold == 0)
2739 mtx_assert(&w_mtx, MA_OWNED);
2740 KASSERT(witness_hash_get(w->w_name) == NULL,
2741 ("%s: trying to add a hash entry that already exists!", __func__));
2742 KASSERT(w->w_hash_next == NULL,
2743 ("%s: w->w_hash_next != NULL", __func__));
2745 hash = witness_hash_djb2(w->w_name, 0) % w_hash.wh_size;
2746 w->w_hash_next = w_hash.wh_array[hash];
2747 w_hash.wh_array[hash] = w;
2752 static struct witness_lock_order_data *
2753 witness_lock_order_get(struct witness *parent, struct witness *child)
2755 struct witness_lock_order_data *data = NULL;
2756 struct witness_lock_order_key key;
2759 MPASS(parent != NULL && child != NULL);
2760 key.from = parent->w_index;
2761 key.to = child->w_index;
2762 WITNESS_INDEX_ASSERT(key.from);
2763 WITNESS_INDEX_ASSERT(key.to);
2764 if ((w_rmatrix[parent->w_index][child->w_index]
2765 & WITNESS_LOCK_ORDER_KNOWN) == 0)
2768 hash = witness_hash_djb2((const char*)&key,
2769 sizeof(key)) % w_lohash.wloh_size;
2770 data = w_lohash.wloh_array[hash];
2771 while (data != NULL) {
2772 if (witness_lock_order_key_equal(&data->wlod_key, &key))
2774 data = data->wlod_next;
2782 * Verify that parent and child have a known relationship, are not the same,
2783 * and child is actually a child of parent. This is done without w_mtx
2784 * to avoid contention in the common case.
2787 witness_lock_order_check(struct witness *parent, struct witness *child)
2790 if (parent != child &&
2791 w_rmatrix[parent->w_index][child->w_index]
2792 & WITNESS_LOCK_ORDER_KNOWN &&
2793 isitmychild(parent, child))
2800 witness_lock_order_add(struct witness *parent, struct witness *child)
2802 struct witness_lock_order_data *data = NULL;
2803 struct witness_lock_order_key key;
2806 MPASS(parent != NULL && child != NULL);
2807 key.from = parent->w_index;
2808 key.to = child->w_index;
2809 WITNESS_INDEX_ASSERT(key.from);
2810 WITNESS_INDEX_ASSERT(key.to);
2811 if (w_rmatrix[parent->w_index][child->w_index]
2812 & WITNESS_LOCK_ORDER_KNOWN)
2815 hash = witness_hash_djb2((const char*)&key,
2816 sizeof(key)) % w_lohash.wloh_size;
2817 w_rmatrix[parent->w_index][child->w_index] |= WITNESS_LOCK_ORDER_KNOWN;
2821 w_lofree = data->wlod_next;
2822 data->wlod_next = w_lohash.wloh_array[hash];
2823 data->wlod_key = key;
2824 w_lohash.wloh_array[hash] = data;
2825 w_lohash.wloh_count++;
2826 stack_zero(&data->wlod_stack);
2827 stack_save(&data->wlod_stack);
2831 /* Call this whenver the structure of the witness graph changes. */
2833 witness_increment_graph_generation(void)
2836 if (witness_cold == 0)
2837 mtx_assert(&w_mtx, MA_OWNED);
2843 _witness_debugger(int cond, const char *msg)
2846 if (witness_trace && cond)
2848 if (witness_kdb && cond)
2849 kdb_enter(KDB_WHY_WITNESS, msg);