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 (1024 + MAXCPU)
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_equal(const struct witness_lock_order_key *a,
309 const struct witness_lock_order_key *b)
312 return (a->from == b->from && a->to == b->to);
315 static int _isitmyx(struct witness *w1, struct witness *w2, int rmask,
318 static void _witness_debugger(int cond, const char *msg);
320 static void adopt(struct witness *parent, struct witness *child);
322 static int blessed(struct witness *, struct witness *);
324 static void depart(struct witness *w);
325 static struct witness *enroll(const char *description,
326 struct lock_class *lock_class);
327 static struct lock_instance *find_instance(struct lock_list_entry *list,
328 const struct lock_object *lock);
329 static int isitmychild(struct witness *parent, struct witness *child);
330 static int isitmydescendant(struct witness *parent, struct witness *child);
331 static void itismychild(struct witness *parent, struct witness *child);
332 static int sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS);
333 static int sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS);
334 static int sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS);
335 static void witness_add_fullgraph(struct sbuf *sb, struct witness *parent);
337 static void witness_ddb_compute_levels(void);
338 static void witness_ddb_display(int(*)(const char *fmt, ...));
339 static void witness_ddb_display_descendants(int(*)(const char *fmt, ...),
340 struct witness *, int indent);
341 static void witness_ddb_display_list(int(*prnt)(const char *fmt, ...),
342 struct witness_list *list);
343 static void witness_ddb_level_descendants(struct witness *parent, int l);
344 static void witness_ddb_list(struct thread *td);
346 static void witness_free(struct witness *m);
347 static struct witness *witness_get(void);
348 static uint32_t witness_hash_djb2(const uint8_t *key, uint32_t size);
349 static struct witness *witness_hash_get(const char *key);
350 static void witness_hash_put(struct witness *w);
351 static void witness_init_hash_tables(void);
352 static void witness_increment_graph_generation(void);
353 static void witness_lock_list_free(struct lock_list_entry *lle);
354 static struct lock_list_entry *witness_lock_list_get(void);
355 static int witness_lock_order_add(struct witness *parent,
356 struct witness *child);
357 static int witness_lock_order_check(struct witness *parent,
358 struct witness *child);
359 static struct witness_lock_order_data *witness_lock_order_get(
360 struct witness *parent,
361 struct witness *child);
362 static void witness_list_lock(struct lock_instance *instance,
363 int (*prnt)(const char *fmt, ...));
364 static void witness_setflag(struct lock_object *lock, int flag, int set);
367 #define witness_debugger(c) _witness_debugger(c, __func__)
369 #define witness_debugger(c)
372 static SYSCTL_NODE(_debug, OID_AUTO, witness, CTLFLAG_RW, NULL,
376 * If set to 0, lock order checking is disabled. If set to -1,
377 * witness is completely disabled. Otherwise witness performs full
378 * lock order checking for all locks. At runtime, lock order checking
379 * may be toggled. However, witness cannot be reenabled once it is
380 * completely disabled.
382 static int witness_watch = 1;
383 TUNABLE_INT("debug.witness.watch", &witness_watch);
384 SYSCTL_PROC(_debug_witness, OID_AUTO, watch, CTLFLAG_RW | CTLTYPE_INT, NULL, 0,
385 sysctl_debug_witness_watch, "I", "witness is watching lock operations");
389 * When KDB is enabled and witness_kdb is 1, it will cause the system
390 * to drop into kdebug() when:
391 * - a lock hierarchy violation occurs
392 * - locks are held when going to sleep.
399 TUNABLE_INT("debug.witness.kdb", &witness_kdb);
400 SYSCTL_INT(_debug_witness, OID_AUTO, kdb, CTLFLAG_RW, &witness_kdb, 0, "");
403 * When KDB is enabled and witness_trace is 1, it will cause the system
404 * to print a stack trace:
405 * - a lock hierarchy violation occurs
406 * - locks are held when going to sleep.
408 int witness_trace = 1;
409 TUNABLE_INT("debug.witness.trace", &witness_trace);
410 SYSCTL_INT(_debug_witness, OID_AUTO, trace, CTLFLAG_RW, &witness_trace, 0, "");
413 #ifdef WITNESS_SKIPSPIN
414 int witness_skipspin = 1;
416 int witness_skipspin = 0;
418 TUNABLE_INT("debug.witness.skipspin", &witness_skipspin);
419 SYSCTL_INT(_debug_witness, OID_AUTO, skipspin, CTLFLAG_RDTUN, &witness_skipspin,
423 * Call this to print out the relations between locks.
425 SYSCTL_PROC(_debug_witness, OID_AUTO, fullgraph, CTLTYPE_STRING | CTLFLAG_RD,
426 NULL, 0, sysctl_debug_witness_fullgraph, "A", "Show locks relation graphs");
429 * Call this to print out the witness faulty stacks.
431 SYSCTL_PROC(_debug_witness, OID_AUTO, badstacks, CTLTYPE_STRING | CTLFLAG_RD,
432 NULL, 0, sysctl_debug_witness_badstacks, "A", "Show bad witness stacks");
434 static struct mtx w_mtx;
437 static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free);
438 static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all);
441 static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin);
442 static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep);
445 static struct lock_list_entry *w_lock_list_free = NULL;
446 static struct witness_pendhelp pending_locks[WITNESS_PENDLIST];
447 static u_int pending_cnt;
449 static int w_free_cnt, w_spin_cnt, w_sleep_cnt;
450 SYSCTL_INT(_debug_witness, OID_AUTO, free_cnt, CTLFLAG_RD, &w_free_cnt, 0, "");
451 SYSCTL_INT(_debug_witness, OID_AUTO, spin_cnt, CTLFLAG_RD, &w_spin_cnt, 0, "");
452 SYSCTL_INT(_debug_witness, OID_AUTO, sleep_cnt, CTLFLAG_RD, &w_sleep_cnt, 0,
455 static struct witness *w_data;
456 static uint8_t w_rmatrix[WITNESS_COUNT+1][WITNESS_COUNT+1];
457 static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT];
458 static struct witness_hash w_hash; /* The witness hash table. */
460 /* The lock order data hash */
461 static struct witness_lock_order_data w_lodata[WITNESS_LO_DATA_COUNT];
462 static struct witness_lock_order_data *w_lofree = NULL;
463 static struct witness_lock_order_hash w_lohash;
464 static int w_max_used_index = 0;
465 static unsigned int w_generation = 0;
466 static const char w_notrunning[] = "Witness not running\n";
467 static const char w_stillcold[] = "Witness is still cold\n";
470 static struct witness_order_list_entry order_lists[] = {
474 { "proctree", &lock_class_sx },
475 { "allproc", &lock_class_sx },
476 { "allprison", &lock_class_sx },
481 { "Giant", &lock_class_mtx_sleep },
482 { "pipe mutex", &lock_class_mtx_sleep },
483 { "sigio lock", &lock_class_mtx_sleep },
484 { "process group", &lock_class_mtx_sleep },
485 { "process lock", &lock_class_mtx_sleep },
486 { "session", &lock_class_mtx_sleep },
487 { "uidinfo hash", &lock_class_rw },
489 { "pmc-sleep", &lock_class_mtx_sleep },
491 { "time lock", &lock_class_mtx_sleep },
496 { "umtx lock", &lock_class_mtx_sleep },
501 { "accept", &lock_class_mtx_sleep },
502 { "so_snd", &lock_class_mtx_sleep },
503 { "so_rcv", &lock_class_mtx_sleep },
504 { "sellck", &lock_class_mtx_sleep },
509 { "so_rcv", &lock_class_mtx_sleep },
510 { "radix node head", &lock_class_rw },
511 { "rtentry", &lock_class_mtx_sleep },
512 { "ifaddr", &lock_class_mtx_sleep },
516 * protocol locks before interface locks, after UDP locks.
518 { "udpinp", &lock_class_rw },
519 { "in_multi_mtx", &lock_class_mtx_sleep },
520 { "igmp_mtx", &lock_class_mtx_sleep },
521 { "if_addr_lock", &lock_class_rw },
525 * protocol locks before interface locks, after UDP locks.
527 { "udpinp", &lock_class_rw },
528 { "in6_multi_mtx", &lock_class_mtx_sleep },
529 { "mld_mtx", &lock_class_mtx_sleep },
530 { "if_addr_lock", &lock_class_rw },
533 * UNIX Domain Sockets
535 { "unp_link_rwlock", &lock_class_rw },
536 { "unp_list_lock", &lock_class_mtx_sleep },
537 { "unp", &lock_class_mtx_sleep },
538 { "so_snd", &lock_class_mtx_sleep },
543 { "udp", &lock_class_rw },
544 { "udpinp", &lock_class_rw },
545 { "so_snd", &lock_class_mtx_sleep },
550 { "tcp", &lock_class_rw },
551 { "tcpinp", &lock_class_rw },
552 { "so_snd", &lock_class_mtx_sleep },
557 { "ddp_list_mtx", &lock_class_mtx_sleep },
558 { "ddp_mtx", &lock_class_mtx_sleep },
563 { "bpf global lock", &lock_class_mtx_sleep },
564 { "bpf interface lock", &lock_class_rw },
565 { "bpf cdev lock", &lock_class_mtx_sleep },
570 { "nfsd_mtx", &lock_class_mtx_sleep },
571 { "so_snd", &lock_class_mtx_sleep },
577 { "802.11 com lock", &lock_class_mtx_sleep},
582 { "network driver", &lock_class_mtx_sleep},
588 { "ng_node", &lock_class_mtx_sleep },
589 { "ng_worklist", &lock_class_mtx_sleep },
594 { "vm map (system)", &lock_class_mtx_sleep },
595 { "vm pagequeue", &lock_class_mtx_sleep },
596 { "vnode interlock", &lock_class_mtx_sleep },
597 { "cdev", &lock_class_mtx_sleep },
602 { "vm map (user)", &lock_class_sx },
603 { "vm object", &lock_class_rw },
604 { "vm page", &lock_class_mtx_sleep },
605 { "vm pagequeue", &lock_class_mtx_sleep },
606 { "pmap pv global", &lock_class_rw },
607 { "pmap", &lock_class_mtx_sleep },
608 { "pmap pv list", &lock_class_rw },
609 { "vm page free queue", &lock_class_mtx_sleep },
612 * kqueue/VFS interaction
614 { "kqueue", &lock_class_mtx_sleep },
615 { "struct mount mtx", &lock_class_mtx_sleep },
616 { "vnode interlock", &lock_class_mtx_sleep },
621 { "dn->dn_mtx", &lock_class_sx },
622 { "dr->dt.di.dr_mtx", &lock_class_sx },
623 { "db->db_mtx", &lock_class_sx },
629 { "ap boot", &lock_class_mtx_spin },
631 { "rm.mutex_mtx", &lock_class_mtx_spin },
632 { "sio", &lock_class_mtx_spin },
633 { "scrlock", &lock_class_mtx_spin },
635 { "cy", &lock_class_mtx_spin },
638 { "pcib_mtx", &lock_class_mtx_spin },
639 { "rtc_mtx", &lock_class_mtx_spin },
641 { "scc_hwmtx", &lock_class_mtx_spin },
642 { "uart_hwmtx", &lock_class_mtx_spin },
643 { "fast_taskqueue", &lock_class_mtx_spin },
644 { "intr table", &lock_class_mtx_spin },
646 { "pmc-per-proc", &lock_class_mtx_spin },
648 { "process slock", &lock_class_mtx_spin },
649 { "sleepq chain", &lock_class_mtx_spin },
650 { "rm_spinlock", &lock_class_mtx_spin },
651 { "turnstile chain", &lock_class_mtx_spin },
652 { "turnstile lock", &lock_class_mtx_spin },
653 { "sched lock", &lock_class_mtx_spin },
654 { "td_contested", &lock_class_mtx_spin },
655 { "callout", &lock_class_mtx_spin },
656 { "entropy harvest mutex", &lock_class_mtx_spin },
657 { "syscons video lock", &lock_class_mtx_spin },
659 { "smp rendezvous", &lock_class_mtx_spin },
662 { "tlb0", &lock_class_mtx_spin },
667 { "intrcnt", &lock_class_mtx_spin },
668 { "icu", &lock_class_mtx_spin },
669 #if defined(SMP) && defined(__sparc64__)
670 { "ipi", &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 kassert_panic("%s: lock (%s) %s can not be recursable",
826 __func__, class->lc_name, lock->lo_name);
827 if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
828 (class->lc_flags & LC_SLEEPABLE) == 0)
829 kassert_panic("%s: lock (%s) %s can not be sleepable",
830 __func__, class->lc_name, lock->lo_name);
831 if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
832 (class->lc_flags & LC_UPGRADABLE) == 0)
833 kassert_panic("%s: lock (%s) %s can not be upgradable",
834 __func__, 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, "
851 "increase WITNESS_PENDLIST\n",
854 lock->lo_witness = enroll(type, class);
858 witness_destroy(struct lock_object *lock)
860 struct lock_class *class;
863 class = LOCK_CLASS(lock);
866 panic("lock (%s) %s destroyed while witness_cold",
867 class->lc_name, lock->lo_name);
869 /* XXX: need to verify that no one holds the lock */
870 if ((lock->lo_flags & LO_WITNESS) == 0 || lock->lo_witness == NULL)
872 w = lock->lo_witness;
874 mtx_lock_spin(&w_mtx);
875 MPASS(w->w_refcount > 0);
878 if (w->w_refcount == 0)
880 mtx_unlock_spin(&w_mtx);
885 witness_ddb_compute_levels(void)
890 * First clear all levels.
892 STAILQ_FOREACH(w, &w_all, w_list)
896 * Look for locks with no parents and level all their descendants.
898 STAILQ_FOREACH(w, &w_all, w_list) {
900 /* If the witness has ancestors (is not a root), skip it. */
901 if (w->w_num_ancestors > 0)
903 witness_ddb_level_descendants(w, 0);
908 witness_ddb_level_descendants(struct witness *w, int l)
912 if (w->w_ddb_level >= l)
918 for (i = 1; i <= w_max_used_index; i++) {
919 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
920 witness_ddb_level_descendants(&w_data[i], l);
925 witness_ddb_display_descendants(int(*prnt)(const char *fmt, ...),
926 struct witness *w, int indent)
930 for (i = 0; i < indent; i++)
932 prnt("%s (type: %s, depth: %d, active refs: %d)",
933 w->w_name, w->w_class->lc_name,
934 w->w_ddb_level, w->w_refcount);
935 if (w->w_displayed) {
936 prnt(" -- (already displayed)\n");
940 if (w->w_file != NULL && w->w_line != 0)
941 prnt(" -- last acquired @ %s:%d\n", fixup_filename(w->w_file),
944 prnt(" -- never acquired\n");
946 WITNESS_INDEX_ASSERT(w->w_index);
947 for (i = 1; i <= w_max_used_index; i++) {
950 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
951 witness_ddb_display_descendants(prnt, &w_data[i],
957 witness_ddb_display_list(int(*prnt)(const char *fmt, ...),
958 struct witness_list *list)
962 STAILQ_FOREACH(w, list, w_typelist) {
963 if (w->w_file == NULL || w->w_ddb_level > 0)
966 /* This lock has no anscestors - display its descendants. */
967 witness_ddb_display_descendants(prnt, w, 0);
974 witness_ddb_display(int(*prnt)(const char *fmt, ...))
978 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
979 witness_ddb_compute_levels();
981 /* Clear all the displayed flags. */
982 STAILQ_FOREACH(w, &w_all, w_list)
986 * First, handle sleep locks which have been acquired at least
989 prnt("Sleep locks:\n");
990 witness_ddb_display_list(prnt, &w_sleep);
995 * Now do spin locks which have been acquired at least once.
997 prnt("\nSpin locks:\n");
998 witness_ddb_display_list(prnt, &w_spin);
1003 * Finally, any locks which have not been acquired yet.
1005 prnt("\nLocks which were never acquired:\n");
1006 STAILQ_FOREACH(w, &w_all, w_list) {
1007 if (w->w_file != NULL || w->w_refcount == 0)
1009 prnt("%s (type: %s, depth: %d)\n", w->w_name,
1010 w->w_class->lc_name, w->w_ddb_level);
1018 witness_defineorder(struct lock_object *lock1, struct lock_object *lock2)
1021 if (witness_watch == -1 || panicstr != NULL)
1024 /* Require locks that witness knows about. */
1025 if (lock1 == NULL || lock1->lo_witness == NULL || lock2 == NULL ||
1026 lock2->lo_witness == NULL)
1029 mtx_assert(&w_mtx, MA_NOTOWNED);
1030 mtx_lock_spin(&w_mtx);
1033 * If we already have either an explicit or implied lock order that
1034 * is the other way around, then return an error.
1036 if (witness_watch &&
1037 isitmydescendant(lock2->lo_witness, lock1->lo_witness)) {
1038 mtx_unlock_spin(&w_mtx);
1042 /* Try to add the new order. */
1043 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1044 lock2->lo_witness->w_name, lock1->lo_witness->w_name);
1045 itismychild(lock1->lo_witness, lock2->lo_witness);
1046 mtx_unlock_spin(&w_mtx);
1051 witness_checkorder(struct lock_object *lock, int flags, const char *file,
1052 int line, struct lock_object *interlock)
1054 struct lock_list_entry *lock_list, *lle;
1055 struct lock_instance *lock1, *lock2, *plock;
1056 struct lock_class *class, *iclass;
1057 struct witness *w, *w1;
1061 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL ||
1065 w = lock->lo_witness;
1066 class = LOCK_CLASS(lock);
1069 if (class->lc_flags & LC_SLEEPLOCK) {
1072 * Since spin locks include a critical section, this check
1073 * implicitly enforces a lock order of all sleep locks before
1076 if (td->td_critnest != 0 && !kdb_active)
1077 kassert_panic("acquiring blockable sleep lock with "
1078 "spinlock or critical section held (%s) %s @ %s:%d",
1079 class->lc_name, lock->lo_name,
1080 fixup_filename(file), line);
1083 * If this is the first lock acquired then just return as
1084 * no order checking is needed.
1086 lock_list = td->td_sleeplocks;
1087 if (lock_list == NULL || lock_list->ll_count == 0)
1092 * If this is the first lock, just return as no order
1093 * checking is needed. Avoid problems with thread
1094 * migration pinning the thread while checking if
1095 * spinlocks are held. If at least one spinlock is held
1096 * the thread is in a safe path and it is allowed to
1100 lock_list = PCPU_GET(spinlocks);
1101 if (lock_list == NULL || lock_list->ll_count == 0) {
1109 * Check to see if we are recursing on a lock we already own. If
1110 * so, make sure that we don't mismatch exclusive and shared lock
1113 lock1 = find_instance(lock_list, lock);
1114 if (lock1 != NULL) {
1115 if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
1116 (flags & LOP_EXCLUSIVE) == 0) {
1117 printf("shared lock of (%s) %s @ %s:%d\n",
1118 class->lc_name, lock->lo_name,
1119 fixup_filename(file), line);
1120 printf("while exclusively locked from %s:%d\n",
1121 fixup_filename(lock1->li_file), lock1->li_line);
1122 kassert_panic("excl->share");
1124 if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
1125 (flags & LOP_EXCLUSIVE) != 0) {
1126 printf("exclusive lock of (%s) %s @ %s:%d\n",
1127 class->lc_name, lock->lo_name,
1128 fixup_filename(file), line);
1129 printf("while share locked from %s:%d\n",
1130 fixup_filename(lock1->li_file), lock1->li_line);
1131 kassert_panic("share->excl");
1136 /* Warn if the interlock is not locked exactly once. */
1137 if (interlock != NULL) {
1138 iclass = LOCK_CLASS(interlock);
1139 lock1 = find_instance(lock_list, interlock);
1141 kassert_panic("interlock (%s) %s not locked @ %s:%d",
1142 iclass->lc_name, interlock->lo_name,
1143 fixup_filename(file), line);
1144 else if ((lock1->li_flags & LI_RECURSEMASK) != 0)
1145 kassert_panic("interlock (%s) %s recursed @ %s:%d",
1146 iclass->lc_name, interlock->lo_name,
1147 fixup_filename(file), line);
1151 * Find the previously acquired lock, but ignore interlocks.
1153 plock = &lock_list->ll_children[lock_list->ll_count - 1];
1154 if (interlock != NULL && plock->li_lock == interlock) {
1155 if (lock_list->ll_count > 1)
1157 &lock_list->ll_children[lock_list->ll_count - 2];
1159 lle = lock_list->ll_next;
1162 * The interlock is the only lock we hold, so
1167 plock = &lle->ll_children[lle->ll_count - 1];
1172 * Try to perform most checks without a lock. If this succeeds we
1173 * can skip acquiring the lock and return success. Otherwise we redo
1174 * the check with the lock held to handle races with concurrent updates.
1176 w1 = plock->li_lock->lo_witness;
1177 if (witness_lock_order_check(w1, w))
1180 mtx_lock_spin(&w_mtx);
1181 if (witness_lock_order_check(w1, w)) {
1182 mtx_unlock_spin(&w_mtx);
1185 witness_lock_order_add(w1, w);
1188 * Check for duplicate locks of the same type. Note that we only
1189 * have to check for this on the last lock we just acquired. Any
1190 * other cases will be caught as lock order violations.
1194 if (!(lock->lo_flags & LO_DUPOK) && !(flags & LOP_DUPOK) &&
1195 !(w_rmatrix[i][i] & WITNESS_REVERSAL)) {
1196 w_rmatrix[i][i] |= WITNESS_REVERSAL;
1198 mtx_unlock_spin(&w_mtx);
1200 "acquiring duplicate lock of same type: \"%s\"\n",
1202 printf(" 1st %s @ %s:%d\n", plock->li_lock->lo_name,
1203 fixup_filename(plock->li_file), plock->li_line);
1204 printf(" 2nd %s @ %s:%d\n", lock->lo_name,
1205 fixup_filename(file), line);
1206 witness_debugger(1);
1208 mtx_unlock_spin(&w_mtx);
1211 mtx_assert(&w_mtx, MA_OWNED);
1214 * If we know that the lock we are acquiring comes after
1215 * the lock we most recently acquired in the lock order tree,
1216 * then there is no need for any further checks.
1218 if (isitmychild(w1, w))
1221 for (j = 0, lle = lock_list; lle != NULL; lle = lle->ll_next) {
1222 for (i = lle->ll_count - 1; i >= 0; i--, j++) {
1224 MPASS(j < LOCK_CHILDCOUNT * LOCK_NCHILDREN);
1225 lock1 = &lle->ll_children[i];
1228 * Ignore the interlock.
1230 if (interlock == lock1->li_lock)
1234 * If this lock doesn't undergo witness checking,
1237 w1 = lock1->li_lock->lo_witness;
1239 KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
1240 ("lock missing witness structure"));
1245 * If we are locking Giant and this is a sleepable
1246 * lock, then skip it.
1248 if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0 &&
1249 lock == &Giant.lock_object)
1253 * If we are locking a sleepable lock and this lock
1254 * is Giant, then skip it.
1256 if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1257 lock1->li_lock == &Giant.lock_object)
1261 * If we are locking a sleepable lock and this lock
1262 * isn't sleepable, we want to treat it as a lock
1263 * order violation to enfore a general lock order of
1264 * sleepable locks before non-sleepable locks.
1266 if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1267 (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
1271 * If we are locking Giant and this is a non-sleepable
1272 * lock, then treat it as a reversal.
1274 if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0 &&
1275 lock == &Giant.lock_object)
1279 * Check the lock order hierarchy for a reveresal.
1281 if (!isitmydescendant(w, w1))
1286 * We have a lock order violation, check to see if it
1287 * is allowed or has already been yelled about.
1292 * If the lock order is blessed, just bail. We don't
1293 * look for other lock order violations though, which
1300 /* Bail if this violation is known */
1301 if (w_rmatrix[w1->w_index][w->w_index] & WITNESS_REVERSAL)
1304 /* Record this as a violation */
1305 w_rmatrix[w1->w_index][w->w_index] |= WITNESS_REVERSAL;
1306 w_rmatrix[w->w_index][w1->w_index] |= WITNESS_REVERSAL;
1307 w->w_reversed = w1->w_reversed = 1;
1308 witness_increment_graph_generation();
1309 mtx_unlock_spin(&w_mtx);
1311 #ifdef WITNESS_NO_VNODE
1313 * There are known LORs between VNODE locks. They are
1314 * not an indication of a bug. VNODE locks are flagged
1315 * as such (LO_IS_VNODE) and we don't yell if the LOR
1316 * is between 2 VNODE locks.
1318 if ((lock->lo_flags & LO_IS_VNODE) != 0 &&
1319 (lock1->li_lock->lo_flags & LO_IS_VNODE) != 0)
1324 * Ok, yell about it.
1326 if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1327 (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
1329 "lock order reversal: (sleepable after non-sleepable)\n");
1330 else if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0
1331 && lock == &Giant.lock_object)
1333 "lock order reversal: (Giant after non-sleepable)\n");
1335 printf("lock order reversal:\n");
1338 * Try to locate an earlier lock with
1339 * witness w in our list.
1342 lock2 = &lle->ll_children[i];
1343 MPASS(lock2->li_lock != NULL);
1344 if (lock2->li_lock->lo_witness == w)
1346 if (i == 0 && lle->ll_next != NULL) {
1348 i = lle->ll_count - 1;
1349 MPASS(i >= 0 && i < LOCK_NCHILDREN);
1354 printf(" 1st %p %s (%s) @ %s:%d\n",
1355 lock1->li_lock, lock1->li_lock->lo_name,
1356 w1->w_name, fixup_filename(lock1->li_file),
1358 printf(" 2nd %p %s (%s) @ %s:%d\n", lock,
1359 lock->lo_name, w->w_name,
1360 fixup_filename(file), line);
1362 printf(" 1st %p %s (%s) @ %s:%d\n",
1363 lock2->li_lock, lock2->li_lock->lo_name,
1364 lock2->li_lock->lo_witness->w_name,
1365 fixup_filename(lock2->li_file),
1367 printf(" 2nd %p %s (%s) @ %s:%d\n",
1368 lock1->li_lock, lock1->li_lock->lo_name,
1369 w1->w_name, fixup_filename(lock1->li_file),
1371 printf(" 3rd %p %s (%s) @ %s:%d\n", lock,
1372 lock->lo_name, w->w_name,
1373 fixup_filename(file), line);
1375 witness_debugger(1);
1381 * If requested, build a new lock order. However, don't build a new
1382 * relationship between a sleepable lock and Giant if it is in the
1383 * wrong direction. The correct lock order is that sleepable locks
1384 * always come before Giant.
1386 if (flags & LOP_NEWORDER &&
1387 !(plock->li_lock == &Giant.lock_object &&
1388 (lock->lo_flags & LO_SLEEPABLE) != 0)) {
1389 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1390 w->w_name, plock->li_lock->lo_witness->w_name);
1391 itismychild(plock->li_lock->lo_witness, w);
1394 mtx_unlock_spin(&w_mtx);
1398 witness_lock(struct lock_object *lock, int flags, const char *file, int line)
1400 struct lock_list_entry **lock_list, *lle;
1401 struct lock_instance *instance;
1405 if (witness_cold || witness_watch == -1 || lock->lo_witness == NULL ||
1408 w = lock->lo_witness;
1411 /* Determine lock list for this lock. */
1412 if (LOCK_CLASS(lock)->lc_flags & LC_SLEEPLOCK)
1413 lock_list = &td->td_sleeplocks;
1415 lock_list = PCPU_PTR(spinlocks);
1417 /* Check to see if we are recursing on a lock we already own. */
1418 instance = find_instance(*lock_list, lock);
1419 if (instance != NULL) {
1420 instance->li_flags++;
1421 CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
1422 td->td_proc->p_pid, lock->lo_name,
1423 instance->li_flags & LI_RECURSEMASK);
1424 instance->li_file = file;
1425 instance->li_line = line;
1429 /* Update per-witness last file and line acquire. */
1433 /* Find the next open lock instance in the list and fill it. */
1435 if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
1436 lle = witness_lock_list_get();
1439 lle->ll_next = *lock_list;
1440 CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
1441 td->td_proc->p_pid, lle);
1444 instance = &lle->ll_children[lle->ll_count++];
1445 instance->li_lock = lock;
1446 instance->li_line = line;
1447 instance->li_file = file;
1448 if ((flags & LOP_EXCLUSIVE) != 0)
1449 instance->li_flags = LI_EXCLUSIVE;
1451 instance->li_flags = 0;
1452 CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
1453 td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
1457 witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
1459 struct lock_instance *instance;
1460 struct lock_class *class;
1462 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1463 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
1465 class = LOCK_CLASS(lock);
1466 if (witness_watch) {
1467 if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1469 "upgrade of non-upgradable lock (%s) %s @ %s:%d",
1470 class->lc_name, lock->lo_name,
1471 fixup_filename(file), line);
1472 if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1474 "upgrade of non-sleep lock (%s) %s @ %s:%d",
1475 class->lc_name, lock->lo_name,
1476 fixup_filename(file), line);
1478 instance = find_instance(curthread->td_sleeplocks, lock);
1479 if (instance == NULL) {
1480 kassert_panic("upgrade of unlocked lock (%s) %s @ %s:%d",
1481 class->lc_name, lock->lo_name,
1482 fixup_filename(file), line);
1485 if (witness_watch) {
1486 if ((instance->li_flags & LI_EXCLUSIVE) != 0)
1488 "upgrade of exclusive lock (%s) %s @ %s:%d",
1489 class->lc_name, lock->lo_name,
1490 fixup_filename(file), line);
1491 if ((instance->li_flags & LI_RECURSEMASK) != 0)
1493 "upgrade of recursed lock (%s) %s r=%d @ %s:%d",
1494 class->lc_name, lock->lo_name,
1495 instance->li_flags & LI_RECURSEMASK,
1496 fixup_filename(file), line);
1498 instance->li_flags |= LI_EXCLUSIVE;
1502 witness_downgrade(struct lock_object *lock, int flags, const char *file,
1505 struct lock_instance *instance;
1506 struct lock_class *class;
1508 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1509 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
1511 class = LOCK_CLASS(lock);
1512 if (witness_watch) {
1513 if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1515 "downgrade of non-upgradable lock (%s) %s @ %s:%d",
1516 class->lc_name, lock->lo_name,
1517 fixup_filename(file), line);
1518 if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1520 "downgrade of non-sleep lock (%s) %s @ %s:%d",
1521 class->lc_name, lock->lo_name,
1522 fixup_filename(file), line);
1524 instance = find_instance(curthread->td_sleeplocks, lock);
1525 if (instance == NULL) {
1526 kassert_panic("downgrade of unlocked lock (%s) %s @ %s:%d",
1527 class->lc_name, lock->lo_name,
1528 fixup_filename(file), line);
1531 if (witness_watch) {
1532 if ((instance->li_flags & LI_EXCLUSIVE) == 0)
1534 "downgrade of shared lock (%s) %s @ %s:%d",
1535 class->lc_name, lock->lo_name,
1536 fixup_filename(file), line);
1537 if ((instance->li_flags & LI_RECURSEMASK) != 0)
1539 "downgrade of recursed lock (%s) %s r=%d @ %s:%d",
1540 class->lc_name, lock->lo_name,
1541 instance->li_flags & LI_RECURSEMASK,
1542 fixup_filename(file), line);
1544 instance->li_flags &= ~LI_EXCLUSIVE;
1548 witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
1550 struct lock_list_entry **lock_list, *lle;
1551 struct lock_instance *instance;
1552 struct lock_class *class;
1557 if (witness_cold || lock->lo_witness == NULL || panicstr != NULL)
1560 class = LOCK_CLASS(lock);
1562 /* Find lock instance associated with this lock. */
1563 if (class->lc_flags & LC_SLEEPLOCK)
1564 lock_list = &td->td_sleeplocks;
1566 lock_list = PCPU_PTR(spinlocks);
1568 for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
1569 for (i = 0; i < (*lock_list)->ll_count; i++) {
1570 instance = &(*lock_list)->ll_children[i];
1571 if (instance->li_lock == lock)
1576 * When disabling WITNESS through witness_watch we could end up in
1577 * having registered locks in the td_sleeplocks queue.
1578 * We have to make sure we flush these queues, so just search for
1579 * eventual register locks and remove them.
1581 if (witness_watch > 0) {
1582 kassert_panic("lock (%s) %s not locked @ %s:%d", class->lc_name,
1583 lock->lo_name, fixup_filename(file), line);
1590 /* First, check for shared/exclusive mismatches. */
1591 if ((instance->li_flags & LI_EXCLUSIVE) != 0 && witness_watch > 0 &&
1592 (flags & LOP_EXCLUSIVE) == 0) {
1593 printf("shared unlock of (%s) %s @ %s:%d\n", class->lc_name,
1594 lock->lo_name, fixup_filename(file), line);
1595 printf("while exclusively locked from %s:%d\n",
1596 fixup_filename(instance->li_file), instance->li_line);
1597 kassert_panic("excl->ushare");
1599 if ((instance->li_flags & LI_EXCLUSIVE) == 0 && witness_watch > 0 &&
1600 (flags & LOP_EXCLUSIVE) != 0) {
1601 printf("exclusive unlock of (%s) %s @ %s:%d\n", class->lc_name,
1602 lock->lo_name, fixup_filename(file), line);
1603 printf("while share locked from %s:%d\n",
1604 fixup_filename(instance->li_file),
1606 kassert_panic("share->uexcl");
1608 /* If we are recursed, unrecurse. */
1609 if ((instance->li_flags & LI_RECURSEMASK) > 0) {
1610 CTR4(KTR_WITNESS, "%s: pid %d unrecursed on %s r=%d", __func__,
1611 td->td_proc->p_pid, instance->li_lock->lo_name,
1612 instance->li_flags);
1613 instance->li_flags--;
1616 /* The lock is now being dropped, check for NORELEASE flag */
1617 if ((instance->li_flags & LI_NORELEASE) != 0 && witness_watch > 0) {
1618 printf("forbidden unlock of (%s) %s @ %s:%d\n", class->lc_name,
1619 lock->lo_name, fixup_filename(file), line);
1620 kassert_panic("lock marked norelease");
1623 /* Otherwise, remove this item from the list. */
1625 CTR4(KTR_WITNESS, "%s: pid %d removed %s from lle[%d]", __func__,
1626 td->td_proc->p_pid, instance->li_lock->lo_name,
1627 (*lock_list)->ll_count - 1);
1628 for (j = i; j < (*lock_list)->ll_count - 1; j++)
1629 (*lock_list)->ll_children[j] =
1630 (*lock_list)->ll_children[j + 1];
1631 (*lock_list)->ll_count--;
1635 * In order to reduce contention on w_mtx, we want to keep always an
1636 * head object into lists so that frequent allocation from the
1637 * free witness pool (and subsequent locking) is avoided.
1638 * In order to maintain the current code simple, when the head
1639 * object is totally unloaded it means also that we do not have
1640 * further objects in the list, so the list ownership needs to be
1641 * hand over to another object if the current head needs to be freed.
1643 if ((*lock_list)->ll_count == 0) {
1644 if (*lock_list == lle) {
1645 if (lle->ll_next == NULL)
1649 *lock_list = lle->ll_next;
1650 CTR3(KTR_WITNESS, "%s: pid %d removed lle %p", __func__,
1651 td->td_proc->p_pid, lle);
1652 witness_lock_list_free(lle);
1657 witness_thread_exit(struct thread *td)
1659 struct lock_list_entry *lle;
1662 lle = td->td_sleeplocks;
1663 if (lle == NULL || panicstr != NULL)
1665 if (lle->ll_count != 0) {
1666 for (n = 0; lle != NULL; lle = lle->ll_next)
1667 for (i = lle->ll_count - 1; i >= 0; i--) {
1669 printf("Thread %p exiting with the following locks held:\n",
1672 witness_list_lock(&lle->ll_children[i], printf);
1676 "Thread %p cannot exit while holding sleeplocks\n", td);
1678 witness_lock_list_free(lle);
1682 * Warn if any locks other than 'lock' are held. Flags can be passed in to
1683 * exempt Giant and sleepable locks from the checks as well. If any
1684 * non-exempt locks are held, then a supplied message is printed to the
1685 * console along with a list of the offending locks. If indicated in the
1686 * flags then a failure results in a panic as well.
1689 witness_warn(int flags, struct lock_object *lock, const char *fmt, ...)
1691 struct lock_list_entry *lock_list, *lle;
1692 struct lock_instance *lock1;
1697 if (witness_cold || witness_watch < 1 || panicstr != NULL)
1701 for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next)
1702 for (i = lle->ll_count - 1; i >= 0; i--) {
1703 lock1 = &lle->ll_children[i];
1704 if (lock1->li_lock == lock)
1706 if (flags & WARN_GIANTOK &&
1707 lock1->li_lock == &Giant.lock_object)
1709 if (flags & WARN_SLEEPOK &&
1710 (lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0)
1716 printf(" with the following");
1717 if (flags & WARN_SLEEPOK)
1718 printf(" non-sleepable");
1719 printf(" locks held:\n");
1722 witness_list_lock(lock1, printf);
1726 * Pin the thread in order to avoid problems with thread migration.
1727 * Once that all verifies are passed about spinlocks ownership,
1728 * the thread is in a safe path and it can be unpinned.
1731 lock_list = PCPU_GET(spinlocks);
1732 if (lock_list != NULL && lock_list->ll_count != 0) {
1736 * We should only have one spinlock and as long as
1737 * the flags cannot match for this locks class,
1738 * check if the first spinlock is the one curthread
1741 lock1 = &lock_list->ll_children[lock_list->ll_count - 1];
1742 if (lock_list->ll_count == 1 && lock_list->ll_next == NULL &&
1743 lock1->li_lock == lock && n == 0)
1749 printf(" with the following");
1750 if (flags & WARN_SLEEPOK)
1751 printf(" non-sleepable");
1752 printf(" locks held:\n");
1753 n += witness_list_locks(&lock_list, printf);
1756 if (flags & WARN_PANIC && n)
1757 kassert_panic("%s", __func__);
1759 witness_debugger(n);
1764 witness_file(struct lock_object *lock)
1768 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1770 w = lock->lo_witness;
1775 witness_line(struct lock_object *lock)
1779 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1781 w = lock->lo_witness;
1785 static struct witness *
1786 enroll(const char *description, struct lock_class *lock_class)
1789 struct witness_list *typelist;
1791 MPASS(description != NULL);
1793 if (witness_watch == -1 || panicstr != NULL)
1795 if ((lock_class->lc_flags & LC_SPINLOCK)) {
1796 if (witness_skipspin)
1800 } else if ((lock_class->lc_flags & LC_SLEEPLOCK)) {
1801 typelist = &w_sleep;
1803 kassert_panic("lock class %s is not sleep or spin",
1804 lock_class->lc_name);
1808 mtx_lock_spin(&w_mtx);
1809 w = witness_hash_get(description);
1812 if ((w = witness_get()) == NULL)
1814 MPASS(strlen(description) < MAX_W_NAME);
1815 strcpy(w->w_name, description);
1816 w->w_class = lock_class;
1818 STAILQ_INSERT_HEAD(&w_all, w, w_list);
1819 if (lock_class->lc_flags & LC_SPINLOCK) {
1820 STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
1822 } else if (lock_class->lc_flags & LC_SLEEPLOCK) {
1823 STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
1827 /* Insert new witness into the hash */
1828 witness_hash_put(w);
1829 witness_increment_graph_generation();
1830 mtx_unlock_spin(&w_mtx);
1834 mtx_unlock_spin(&w_mtx);
1835 if (lock_class != w->w_class)
1837 "lock (%s) %s does not match earlier (%s) lock",
1838 description, lock_class->lc_name,
1839 w->w_class->lc_name);
1844 depart(struct witness *w)
1846 struct witness_list *list;
1848 MPASS(w->w_refcount == 0);
1849 if (w->w_class->lc_flags & LC_SLEEPLOCK) {
1857 * Set file to NULL as it may point into a loadable module.
1861 witness_increment_graph_generation();
1866 adopt(struct witness *parent, struct witness *child)
1870 if (witness_cold == 0)
1871 mtx_assert(&w_mtx, MA_OWNED);
1873 /* If the relationship is already known, there's no work to be done. */
1874 if (isitmychild(parent, child))
1877 /* When the structure of the graph changes, bump up the generation. */
1878 witness_increment_graph_generation();
1881 * The hard part ... create the direct relationship, then propagate all
1882 * indirect relationships.
1884 pi = parent->w_index;
1885 ci = child->w_index;
1886 WITNESS_INDEX_ASSERT(pi);
1887 WITNESS_INDEX_ASSERT(ci);
1889 w_rmatrix[pi][ci] |= WITNESS_PARENT;
1890 w_rmatrix[ci][pi] |= WITNESS_CHILD;
1893 * If parent was not already an ancestor of child,
1894 * then we increment the descendant and ancestor counters.
1896 if ((w_rmatrix[pi][ci] & WITNESS_ANCESTOR) == 0) {
1897 parent->w_num_descendants++;
1898 child->w_num_ancestors++;
1902 * Find each ancestor of 'pi'. Note that 'pi' itself is counted as
1903 * an ancestor of 'pi' during this loop.
1905 for (i = 1; i <= w_max_used_index; i++) {
1906 if ((w_rmatrix[i][pi] & WITNESS_ANCESTOR_MASK) == 0 &&
1910 /* Find each descendant of 'i' and mark it as a descendant. */
1911 for (j = 1; j <= w_max_used_index; j++) {
1914 * Skip children that are already marked as
1915 * descendants of 'i'.
1917 if (w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK)
1921 * We are only interested in descendants of 'ci'. Note
1922 * that 'ci' itself is counted as a descendant of 'ci'.
1924 if ((w_rmatrix[ci][j] & WITNESS_ANCESTOR_MASK) == 0 &&
1927 w_rmatrix[i][j] |= WITNESS_ANCESTOR;
1928 w_rmatrix[j][i] |= WITNESS_DESCENDANT;
1929 w_data[i].w_num_descendants++;
1930 w_data[j].w_num_ancestors++;
1933 * Make sure we aren't marking a node as both an
1934 * ancestor and descendant. We should have caught
1935 * this as a lock order reversal earlier.
1937 if ((w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK) &&
1938 (w_rmatrix[i][j] & WITNESS_DESCENDANT_MASK)) {
1939 printf("witness rmatrix paradox! [%d][%d]=%d "
1940 "both ancestor and descendant\n",
1941 i, j, w_rmatrix[i][j]);
1943 printf("Witness disabled.\n");
1946 if ((w_rmatrix[j][i] & WITNESS_ANCESTOR_MASK) &&
1947 (w_rmatrix[j][i] & WITNESS_DESCENDANT_MASK)) {
1948 printf("witness rmatrix paradox! [%d][%d]=%d "
1949 "both ancestor and descendant\n",
1950 j, i, w_rmatrix[j][i]);
1952 printf("Witness disabled.\n");
1960 itismychild(struct witness *parent, struct witness *child)
1964 MPASS(child != NULL && parent != NULL);
1965 if (witness_cold == 0)
1966 mtx_assert(&w_mtx, MA_OWNED);
1968 if (!witness_lock_type_equal(parent, child)) {
1969 if (witness_cold == 0) {
1971 mtx_unlock_spin(&w_mtx);
1976 "%s: parent \"%s\" (%s) and child \"%s\" (%s) are not "
1977 "the same lock type", __func__, parent->w_name,
1978 parent->w_class->lc_name, child->w_name,
1979 child->w_class->lc_name);
1981 mtx_lock_spin(&w_mtx);
1983 adopt(parent, child);
1987 * Generic code for the isitmy*() functions. The rmask parameter is the
1988 * expected relationship of w1 to w2.
1991 _isitmyx(struct witness *w1, struct witness *w2, int rmask, const char *fname)
1993 unsigned char r1, r2;
1998 WITNESS_INDEX_ASSERT(i1);
1999 WITNESS_INDEX_ASSERT(i2);
2000 r1 = w_rmatrix[i1][i2] & WITNESS_RELATED_MASK;
2001 r2 = w_rmatrix[i2][i1] & WITNESS_RELATED_MASK;
2003 /* The flags on one better be the inverse of the flags on the other */
2004 if (!((WITNESS_ATOD(r1) == r2 && WITNESS_DTOA(r2) == r1) ||
2005 (WITNESS_DTOA(r1) == r2 && WITNESS_ATOD(r2) == r1))) {
2006 /* Don't squawk if we're potentially racing with an update. */
2007 if (!mtx_owned(&w_mtx))
2009 printf("%s: rmatrix mismatch between %s (index %d) and %s "
2010 "(index %d): w_rmatrix[%d][%d] == %hhx but "
2011 "w_rmatrix[%d][%d] == %hhx\n",
2012 fname, w1->w_name, i1, w2->w_name, i2, i1, i2, r1,
2015 printf("Witness disabled.\n");
2018 return (r1 & rmask);
2022 * Checks if @child is a direct child of @parent.
2025 isitmychild(struct witness *parent, struct witness *child)
2028 return (_isitmyx(parent, child, WITNESS_PARENT, __func__));
2032 * Checks if @descendant is a direct or inderect descendant of @ancestor.
2035 isitmydescendant(struct witness *ancestor, struct witness *descendant)
2038 return (_isitmyx(ancestor, descendant, WITNESS_ANCESTOR_MASK,
2044 blessed(struct witness *w1, struct witness *w2)
2047 struct witness_blessed *b;
2049 for (i = 0; i < blessed_count; i++) {
2050 b = &blessed_list[i];
2051 if (strcmp(w1->w_name, b->b_lock1) == 0) {
2052 if (strcmp(w2->w_name, b->b_lock2) == 0)
2056 if (strcmp(w1->w_name, b->b_lock2) == 0)
2057 if (strcmp(w2->w_name, b->b_lock1) == 0)
2064 static struct witness *
2070 if (witness_cold == 0)
2071 mtx_assert(&w_mtx, MA_OWNED);
2073 if (witness_watch == -1) {
2074 mtx_unlock_spin(&w_mtx);
2077 if (STAILQ_EMPTY(&w_free)) {
2079 mtx_unlock_spin(&w_mtx);
2080 printf("WITNESS: unable to allocate a new witness object\n");
2083 w = STAILQ_FIRST(&w_free);
2084 STAILQ_REMOVE_HEAD(&w_free, w_list);
2087 MPASS(index > 0 && index == w_max_used_index+1 &&
2088 index < WITNESS_COUNT);
2089 bzero(w, sizeof(*w));
2091 if (index > w_max_used_index)
2092 w_max_used_index = index;
2097 witness_free(struct witness *w)
2100 STAILQ_INSERT_HEAD(&w_free, w, w_list);
2104 static struct lock_list_entry *
2105 witness_lock_list_get(void)
2107 struct lock_list_entry *lle;
2109 if (witness_watch == -1)
2111 mtx_lock_spin(&w_mtx);
2112 lle = w_lock_list_free;
2115 mtx_unlock_spin(&w_mtx);
2116 printf("%s: witness exhausted\n", __func__);
2119 w_lock_list_free = lle->ll_next;
2120 mtx_unlock_spin(&w_mtx);
2121 bzero(lle, sizeof(*lle));
2126 witness_lock_list_free(struct lock_list_entry *lle)
2129 mtx_lock_spin(&w_mtx);
2130 lle->ll_next = w_lock_list_free;
2131 w_lock_list_free = lle;
2132 mtx_unlock_spin(&w_mtx);
2135 static struct lock_instance *
2136 find_instance(struct lock_list_entry *list, const struct lock_object *lock)
2138 struct lock_list_entry *lle;
2139 struct lock_instance *instance;
2142 for (lle = list; lle != NULL; lle = lle->ll_next)
2143 for (i = lle->ll_count - 1; i >= 0; i--) {
2144 instance = &lle->ll_children[i];
2145 if (instance->li_lock == lock)
2152 witness_list_lock(struct lock_instance *instance,
2153 int (*prnt)(const char *fmt, ...))
2155 struct lock_object *lock;
2157 lock = instance->li_lock;
2158 prnt("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ?
2159 "exclusive" : "shared", LOCK_CLASS(lock)->lc_name, lock->lo_name);
2160 if (lock->lo_witness->w_name != lock->lo_name)
2161 prnt(" (%s)", lock->lo_witness->w_name);
2162 prnt(" r = %d (%p) locked @ %s:%d\n",
2163 instance->li_flags & LI_RECURSEMASK, lock,
2164 fixup_filename(instance->li_file), instance->li_line);
2169 witness_thread_has_locks(struct thread *td)
2172 if (td->td_sleeplocks == NULL)
2174 return (td->td_sleeplocks->ll_count != 0);
2178 witness_proc_has_locks(struct proc *p)
2182 FOREACH_THREAD_IN_PROC(p, td) {
2183 if (witness_thread_has_locks(td))
2191 witness_list_locks(struct lock_list_entry **lock_list,
2192 int (*prnt)(const char *fmt, ...))
2194 struct lock_list_entry *lle;
2198 for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
2199 for (i = lle->ll_count - 1; i >= 0; i--) {
2200 witness_list_lock(&lle->ll_children[i], prnt);
2207 * This is a bit risky at best. We call this function when we have timed
2208 * out acquiring a spin lock, and we assume that the other CPU is stuck
2209 * with this lock held. So, we go groveling around in the other CPU's
2210 * per-cpu data to try to find the lock instance for this spin lock to
2211 * see when it was last acquired.
2214 witness_display_spinlock(struct lock_object *lock, struct thread *owner,
2215 int (*prnt)(const char *fmt, ...))
2217 struct lock_instance *instance;
2220 if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU)
2222 pc = pcpu_find(owner->td_oncpu);
2223 instance = find_instance(pc->pc_spinlocks, lock);
2224 if (instance != NULL)
2225 witness_list_lock(instance, prnt);
2229 witness_save(struct lock_object *lock, const char **filep, int *linep)
2231 struct lock_list_entry *lock_list;
2232 struct lock_instance *instance;
2233 struct lock_class *class;
2236 * This function is used independently in locking code to deal with
2237 * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2240 if (SCHEDULER_STOPPED())
2242 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2243 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2245 class = LOCK_CLASS(lock);
2246 if (class->lc_flags & LC_SLEEPLOCK)
2247 lock_list = curthread->td_sleeplocks;
2249 if (witness_skipspin)
2251 lock_list = PCPU_GET(spinlocks);
2253 instance = find_instance(lock_list, lock);
2254 if (instance == NULL) {
2255 kassert_panic("%s: lock (%s) %s not locked", __func__,
2256 class->lc_name, lock->lo_name);
2259 *filep = instance->li_file;
2260 *linep = instance->li_line;
2264 witness_restore(struct lock_object *lock, const char *file, int line)
2266 struct lock_list_entry *lock_list;
2267 struct lock_instance *instance;
2268 struct lock_class *class;
2271 * This function is used independently in locking code to deal with
2272 * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2275 if (SCHEDULER_STOPPED())
2277 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2278 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2280 class = LOCK_CLASS(lock);
2281 if (class->lc_flags & LC_SLEEPLOCK)
2282 lock_list = curthread->td_sleeplocks;
2284 if (witness_skipspin)
2286 lock_list = PCPU_GET(spinlocks);
2288 instance = find_instance(lock_list, lock);
2289 if (instance == NULL)
2290 kassert_panic("%s: lock (%s) %s not locked", __func__,
2291 class->lc_name, lock->lo_name);
2292 lock->lo_witness->w_file = file;
2293 lock->lo_witness->w_line = line;
2294 if (instance == NULL)
2296 instance->li_file = file;
2297 instance->li_line = line;
2301 witness_assert(const struct lock_object *lock, int flags, const char *file,
2304 #ifdef INVARIANT_SUPPORT
2305 struct lock_instance *instance;
2306 struct lock_class *class;
2308 if (lock->lo_witness == NULL || witness_watch < 1 || panicstr != NULL)
2310 class = LOCK_CLASS(lock);
2311 if ((class->lc_flags & LC_SLEEPLOCK) != 0)
2312 instance = find_instance(curthread->td_sleeplocks, lock);
2313 else if ((class->lc_flags & LC_SPINLOCK) != 0)
2314 instance = find_instance(PCPU_GET(spinlocks), lock);
2316 kassert_panic("Lock (%s) %s is not sleep or spin!",
2317 class->lc_name, lock->lo_name);
2322 if (instance != NULL)
2323 kassert_panic("Lock (%s) %s locked @ %s:%d.",
2324 class->lc_name, lock->lo_name,
2325 fixup_filename(file), line);
2328 case LA_LOCKED | LA_RECURSED:
2329 case LA_LOCKED | LA_NOTRECURSED:
2331 case LA_SLOCKED | LA_RECURSED:
2332 case LA_SLOCKED | LA_NOTRECURSED:
2334 case LA_XLOCKED | LA_RECURSED:
2335 case LA_XLOCKED | LA_NOTRECURSED:
2336 if (instance == NULL) {
2337 kassert_panic("Lock (%s) %s not locked @ %s:%d.",
2338 class->lc_name, lock->lo_name,
2339 fixup_filename(file), line);
2342 if ((flags & LA_XLOCKED) != 0 &&
2343 (instance->li_flags & LI_EXCLUSIVE) == 0)
2345 "Lock (%s) %s not exclusively locked @ %s:%d.",
2346 class->lc_name, lock->lo_name,
2347 fixup_filename(file), line);
2348 if ((flags & LA_SLOCKED) != 0 &&
2349 (instance->li_flags & LI_EXCLUSIVE) != 0)
2351 "Lock (%s) %s exclusively locked @ %s:%d.",
2352 class->lc_name, lock->lo_name,
2353 fixup_filename(file), line);
2354 if ((flags & LA_RECURSED) != 0 &&
2355 (instance->li_flags & LI_RECURSEMASK) == 0)
2356 kassert_panic("Lock (%s) %s not recursed @ %s:%d.",
2357 class->lc_name, lock->lo_name,
2358 fixup_filename(file), line);
2359 if ((flags & LA_NOTRECURSED) != 0 &&
2360 (instance->li_flags & LI_RECURSEMASK) != 0)
2361 kassert_panic("Lock (%s) %s recursed @ %s:%d.",
2362 class->lc_name, lock->lo_name,
2363 fixup_filename(file), line);
2366 kassert_panic("Invalid lock assertion at %s:%d.",
2367 fixup_filename(file), line);
2370 #endif /* INVARIANT_SUPPORT */
2374 witness_setflag(struct lock_object *lock, int flag, int set)
2376 struct lock_list_entry *lock_list;
2377 struct lock_instance *instance;
2378 struct lock_class *class;
2380 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2382 class = LOCK_CLASS(lock);
2383 if (class->lc_flags & LC_SLEEPLOCK)
2384 lock_list = curthread->td_sleeplocks;
2386 if (witness_skipspin)
2388 lock_list = PCPU_GET(spinlocks);
2390 instance = find_instance(lock_list, lock);
2391 if (instance == NULL) {
2392 kassert_panic("%s: lock (%s) %s not locked", __func__,
2393 class->lc_name, lock->lo_name);
2398 instance->li_flags |= flag;
2400 instance->li_flags &= ~flag;
2404 witness_norelease(struct lock_object *lock)
2407 witness_setflag(lock, LI_NORELEASE, 1);
2411 witness_releaseok(struct lock_object *lock)
2414 witness_setflag(lock, LI_NORELEASE, 0);
2419 witness_ddb_list(struct thread *td)
2422 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2423 KASSERT(kdb_active, ("%s: not in the debugger", __func__));
2425 if (witness_watch < 1)
2428 witness_list_locks(&td->td_sleeplocks, db_printf);
2431 * We only handle spinlocks if td == curthread. This is somewhat broken
2432 * if td is currently executing on some other CPU and holds spin locks
2433 * as we won't display those locks. If we had a MI way of getting
2434 * the per-cpu data for a given cpu then we could use
2435 * td->td_oncpu to get the list of spinlocks for this thread
2438 * That still wouldn't really fix this unless we locked the scheduler
2439 * lock or stopped the other CPU to make sure it wasn't changing the
2440 * list out from under us. It is probably best to just not try to
2441 * handle threads on other CPU's for now.
2443 if (td == curthread && PCPU_GET(spinlocks) != NULL)
2444 witness_list_locks(PCPU_PTR(spinlocks), db_printf);
2447 DB_SHOW_COMMAND(locks, db_witness_list)
2452 td = db_lookup_thread(addr, TRUE);
2455 witness_ddb_list(td);
2458 DB_SHOW_ALL_COMMAND(locks, db_witness_list_all)
2464 * It would be nice to list only threads and processes that actually
2465 * held sleep locks, but that information is currently not exported
2468 FOREACH_PROC_IN_SYSTEM(p) {
2469 if (!witness_proc_has_locks(p))
2471 FOREACH_THREAD_IN_PROC(p, td) {
2472 if (!witness_thread_has_locks(td))
2474 db_printf("Process %d (%s) thread %p (%d)\n", p->p_pid,
2475 p->p_comm, td, td->td_tid);
2476 witness_ddb_list(td);
2482 DB_SHOW_ALIAS(alllocks, db_witness_list_all)
2484 DB_SHOW_COMMAND(witness, db_witness_display)
2487 witness_ddb_display(db_printf);
2492 sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS)
2494 struct witness_lock_order_data *data1, *data2, *tmp_data1, *tmp_data2;
2495 struct witness *tmp_w1, *tmp_w2, *w1, *w2;
2497 u_int w_rmatrix1, w_rmatrix2;
2498 int error, generation, i, j;
2504 if (witness_watch < 1) {
2505 error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2509 error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2513 sb = sbuf_new(NULL, NULL, BADSTACK_SBUF_SIZE, SBUF_AUTOEXTEND);
2517 /* Allocate and init temporary storage space. */
2518 tmp_w1 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2519 tmp_w2 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2520 tmp_data1 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2522 tmp_data2 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2524 stack_zero(&tmp_data1->wlod_stack);
2525 stack_zero(&tmp_data2->wlod_stack);
2528 mtx_lock_spin(&w_mtx);
2529 generation = w_generation;
2530 mtx_unlock_spin(&w_mtx);
2531 sbuf_printf(sb, "Number of known direct relationships is %d\n",
2532 w_lohash.wloh_count);
2533 for (i = 1; i < w_max_used_index; i++) {
2534 mtx_lock_spin(&w_mtx);
2535 if (generation != w_generation) {
2536 mtx_unlock_spin(&w_mtx);
2538 /* The graph has changed, try again. */
2545 if (w1->w_reversed == 0) {
2546 mtx_unlock_spin(&w_mtx);
2550 /* Copy w1 locally so we can release the spin lock. */
2552 mtx_unlock_spin(&w_mtx);
2554 if (tmp_w1->w_reversed == 0)
2556 for (j = 1; j < w_max_used_index; j++) {
2557 if ((w_rmatrix[i][j] & WITNESS_REVERSAL) == 0 || i > j)
2560 mtx_lock_spin(&w_mtx);
2561 if (generation != w_generation) {
2562 mtx_unlock_spin(&w_mtx);
2564 /* The graph has changed, try again. */
2571 data1 = witness_lock_order_get(w1, w2);
2572 data2 = witness_lock_order_get(w2, w1);
2575 * Copy information locally so we can release the
2579 w_rmatrix1 = (unsigned int)w_rmatrix[i][j];
2580 w_rmatrix2 = (unsigned int)w_rmatrix[j][i];
2583 stack_zero(&tmp_data1->wlod_stack);
2584 stack_copy(&data1->wlod_stack,
2585 &tmp_data1->wlod_stack);
2587 if (data2 && data2 != data1) {
2588 stack_zero(&tmp_data2->wlod_stack);
2589 stack_copy(&data2->wlod_stack,
2590 &tmp_data2->wlod_stack);
2592 mtx_unlock_spin(&w_mtx);
2595 "\nLock order reversal between \"%s\"(%s) and \"%s\"(%s)!\n",
2596 tmp_w1->w_name, tmp_w1->w_class->lc_name,
2597 tmp_w2->w_name, tmp_w2->w_class->lc_name);
2600 "w_rmatrix[%s][%s] == %x, w_rmatrix[%s][%s] == %x\n",
2601 tmp_w1->name, tmp_w2->w_name, w_rmatrix1,
2602 tmp_w2->name, tmp_w1->w_name, w_rmatrix2);
2606 "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2607 tmp_w1->w_name, tmp_w1->w_class->lc_name,
2608 tmp_w2->w_name, tmp_w2->w_class->lc_name);
2609 stack_sbuf_print(sb, &tmp_data1->wlod_stack);
2610 sbuf_printf(sb, "\n");
2612 if (data2 && data2 != data1) {
2614 "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2615 tmp_w2->w_name, tmp_w2->w_class->lc_name,
2616 tmp_w1->w_name, tmp_w1->w_class->lc_name);
2617 stack_sbuf_print(sb, &tmp_data2->wlod_stack);
2618 sbuf_printf(sb, "\n");
2622 mtx_lock_spin(&w_mtx);
2623 if (generation != w_generation) {
2624 mtx_unlock_spin(&w_mtx);
2627 * The graph changed while we were printing stack data,
2634 mtx_unlock_spin(&w_mtx);
2636 /* Free temporary storage space. */
2637 free(tmp_data1, M_TEMP);
2638 free(tmp_data2, M_TEMP);
2639 free(tmp_w1, M_TEMP);
2640 free(tmp_w2, M_TEMP);
2643 error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1);
2650 sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS)
2656 if (witness_watch < 1) {
2657 error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2661 error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2666 error = sysctl_wire_old_buffer(req, 0);
2669 sb = sbuf_new_for_sysctl(NULL, NULL, FULLGRAPH_SBUF_SIZE, req);
2672 sbuf_printf(sb, "\n");
2674 mtx_lock_spin(&w_mtx);
2675 STAILQ_FOREACH(w, &w_all, w_list)
2677 STAILQ_FOREACH(w, &w_all, w_list)
2678 witness_add_fullgraph(sb, w);
2679 mtx_unlock_spin(&w_mtx);
2682 * Close the sbuf and return to userland.
2684 error = sbuf_finish(sb);
2691 sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS)
2695 value = witness_watch;
2696 error = sysctl_handle_int(oidp, &value, 0, req);
2697 if (error != 0 || req->newptr == NULL)
2699 if (value > 1 || value < -1 ||
2700 (witness_watch == -1 && value != witness_watch))
2702 witness_watch = value;
2707 witness_add_fullgraph(struct sbuf *sb, struct witness *w)
2711 if (w->w_displayed != 0 || (w->w_file == NULL && w->w_line == 0))
2715 WITNESS_INDEX_ASSERT(w->w_index);
2716 for (i = 1; i <= w_max_used_index; i++) {
2717 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT) {
2718 sbuf_printf(sb, "\"%s\",\"%s\"\n", w->w_name,
2720 witness_add_fullgraph(sb, &w_data[i]);
2726 * A simple hash function. Takes a key pointer and a key size. If size == 0,
2727 * interprets the key as a string and reads until the null
2728 * terminator. Otherwise, reads the first size bytes. Returns an unsigned 32-bit
2729 * hash value computed from the key.
2732 witness_hash_djb2(const uint8_t *key, uint32_t size)
2734 unsigned int hash = 5381;
2737 /* hash = hash * 33 + key[i] */
2739 for (i = 0; i < size; i++)
2740 hash = ((hash << 5) + hash) + (unsigned int)key[i];
2742 for (i = 0; key[i] != 0; i++)
2743 hash = ((hash << 5) + hash) + (unsigned int)key[i];
2750 * Initializes the two witness hash tables. Called exactly once from
2751 * witness_initialize().
2754 witness_init_hash_tables(void)
2758 MPASS(witness_cold);
2760 /* Initialize the hash tables. */
2761 for (i = 0; i < WITNESS_HASH_SIZE; i++)
2762 w_hash.wh_array[i] = NULL;
2764 w_hash.wh_size = WITNESS_HASH_SIZE;
2765 w_hash.wh_count = 0;
2767 /* Initialize the lock order data hash. */
2769 for (i = 0; i < WITNESS_LO_DATA_COUNT; i++) {
2770 memset(&w_lodata[i], 0, sizeof(w_lodata[i]));
2771 w_lodata[i].wlod_next = w_lofree;
2772 w_lofree = &w_lodata[i];
2774 w_lohash.wloh_size = WITNESS_LO_HASH_SIZE;
2775 w_lohash.wloh_count = 0;
2776 for (i = 0; i < WITNESS_LO_HASH_SIZE; i++)
2777 w_lohash.wloh_array[i] = NULL;
2780 static struct witness *
2781 witness_hash_get(const char *key)
2787 if (witness_cold == 0)
2788 mtx_assert(&w_mtx, MA_OWNED);
2789 hash = witness_hash_djb2(key, 0) % w_hash.wh_size;
2790 w = w_hash.wh_array[hash];
2792 if (strcmp(w->w_name, key) == 0)
2802 witness_hash_put(struct witness *w)
2807 MPASS(w->w_name != NULL);
2808 if (witness_cold == 0)
2809 mtx_assert(&w_mtx, MA_OWNED);
2810 KASSERT(witness_hash_get(w->w_name) == NULL,
2811 ("%s: trying to add a hash entry that already exists!", __func__));
2812 KASSERT(w->w_hash_next == NULL,
2813 ("%s: w->w_hash_next != NULL", __func__));
2815 hash = witness_hash_djb2(w->w_name, 0) % w_hash.wh_size;
2816 w->w_hash_next = w_hash.wh_array[hash];
2817 w_hash.wh_array[hash] = w;
2822 static struct witness_lock_order_data *
2823 witness_lock_order_get(struct witness *parent, struct witness *child)
2825 struct witness_lock_order_data *data = NULL;
2826 struct witness_lock_order_key key;
2829 MPASS(parent != NULL && child != NULL);
2830 key.from = parent->w_index;
2831 key.to = child->w_index;
2832 WITNESS_INDEX_ASSERT(key.from);
2833 WITNESS_INDEX_ASSERT(key.to);
2834 if ((w_rmatrix[parent->w_index][child->w_index]
2835 & WITNESS_LOCK_ORDER_KNOWN) == 0)
2838 hash = witness_hash_djb2((const char*)&key,
2839 sizeof(key)) % w_lohash.wloh_size;
2840 data = w_lohash.wloh_array[hash];
2841 while (data != NULL) {
2842 if (witness_lock_order_key_equal(&data->wlod_key, &key))
2844 data = data->wlod_next;
2852 * Verify that parent and child have a known relationship, are not the same,
2853 * and child is actually a child of parent. This is done without w_mtx
2854 * to avoid contention in the common case.
2857 witness_lock_order_check(struct witness *parent, struct witness *child)
2860 if (parent != child &&
2861 w_rmatrix[parent->w_index][child->w_index]
2862 & WITNESS_LOCK_ORDER_KNOWN &&
2863 isitmychild(parent, child))
2870 witness_lock_order_add(struct witness *parent, struct witness *child)
2872 struct witness_lock_order_data *data = NULL;
2873 struct witness_lock_order_key key;
2876 MPASS(parent != NULL && child != NULL);
2877 key.from = parent->w_index;
2878 key.to = child->w_index;
2879 WITNESS_INDEX_ASSERT(key.from);
2880 WITNESS_INDEX_ASSERT(key.to);
2881 if (w_rmatrix[parent->w_index][child->w_index]
2882 & WITNESS_LOCK_ORDER_KNOWN)
2885 hash = witness_hash_djb2((const char*)&key,
2886 sizeof(key)) % w_lohash.wloh_size;
2887 w_rmatrix[parent->w_index][child->w_index] |= WITNESS_LOCK_ORDER_KNOWN;
2891 w_lofree = data->wlod_next;
2892 data->wlod_next = w_lohash.wloh_array[hash];
2893 data->wlod_key = key;
2894 w_lohash.wloh_array[hash] = data;
2895 w_lohash.wloh_count++;
2896 stack_zero(&data->wlod_stack);
2897 stack_save(&data->wlod_stack);
2901 /* Call this whenever the structure of the witness graph changes. */
2903 witness_increment_graph_generation(void)
2906 if (witness_cold == 0)
2907 mtx_assert(&w_mtx, MA_OWNED);
2913 _witness_debugger(int cond, const char *msg)
2916 if (witness_trace && cond)
2918 if (witness_kdb && cond)
2919 kdb_enter(KDB_WHY_WITNESS, msg);