2 * SPDX-License-Identifier: BSD-3-Clause
4 * Copyright (c) 2008 Isilon Systems, Inc.
5 * Copyright (c) 2008 Ilya Maykov <ivmaykov@gmail.com>
6 * Copyright (c) 1998 Berkeley Software Design, Inc.
9 * Redistribution and use in source and binary forms, with or without
10 * modification, are permitted provided that the following conditions
12 * 1. Redistributions of source code must retain the above copyright
13 * notice, this list of conditions and the following disclaimer.
14 * 2. Redistributions in binary form must reproduce the above copyright
15 * notice, this list of conditions and the following disclaimer in the
16 * documentation and/or other materials provided with the distribution.
17 * 3. Berkeley Software Design Inc's name may not be used to endorse or
18 * promote products derived from this software without specific prior
21 * THIS SOFTWARE IS PROVIDED BY BERKELEY SOFTWARE DESIGN INC ``AS IS'' AND
22 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24 * ARE DISCLAIMED. IN NO EVENT SHALL BERKELEY SOFTWARE DESIGN INC BE LIABLE
25 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
33 * from BSDI $Id: mutex_witness.c,v 1.1.2.20 2000/04/27 03:10:27 cp Exp $
34 * and BSDI $Id: synch_machdep.c,v 2.3.2.39 2000/04/27 03:10:25 cp Exp $
38 * Implementation of the `witness' lock verifier. Originally implemented for
39 * mutexes in BSD/OS. Extended to handle generic lock objects and lock
45 * Pronunciation: 'wit-n&s
47 * Etymology: Middle English witnesse, from Old English witnes knowledge,
48 * testimony, witness, from 2wit
49 * Date: before 12th century
50 * 1 : attestation of a fact or event : TESTIMONY
51 * 2 : one that gives evidence; specifically : one who testifies in
52 * a cause or before a judicial tribunal
53 * 3 : one asked to be present at a transaction so as to be able to
54 * testify to its having taken place
55 * 4 : one who has personal knowledge of something
56 * 5 a : something serving as evidence or proof : SIGN
57 * b : public affirmation by word or example of usually
58 * religious faith or conviction <the heroic witness to divine
60 * 6 capitalized : a member of the Jehovah's Witnesses
64 * Special rules concerning Giant and lock orders:
66 * 1) Giant must be acquired before any other mutexes. Stated another way,
67 * no other mutex may be held when Giant is acquired.
69 * 2) Giant must be released when blocking on a sleepable lock.
71 * This rule is less obvious, but is a result of Giant providing the same
72 * semantics as spl(). Basically, when a thread sleeps, it must release
73 * Giant. When a thread blocks on a sleepable lock, it sleeps. Hence rule
76 * 3) Giant may be acquired before or after sleepable locks.
78 * This rule is also not quite as obvious. Giant may be acquired after
79 * a sleepable lock because it is a non-sleepable lock and non-sleepable
80 * locks may always be acquired while holding a sleepable lock. The second
81 * case, Giant before a sleepable lock, follows from rule 2) above. Suppose
82 * you have two threads T1 and T2 and a sleepable lock X. Suppose that T1
83 * acquires X and blocks on Giant. Then suppose that T2 acquires Giant and
84 * blocks on X. When T2 blocks on X, T2 will release Giant allowing T1 to
85 * execute. Thus, acquiring Giant both before and after a sleepable lock
86 * will not result in a lock order reversal.
89 #include <sys/cdefs.h>
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/syslog.h>
110 #include <sys/systm.h>
116 #include <machine/stdarg.h>
118 #if !defined(DDB) && !defined(STACK)
119 #error "DDB or STACK options are required for WITNESS"
122 /* Note that these traces do not work with KTR_ALQ. */
124 #define KTR_WITNESS KTR_SUBSYS
126 #define KTR_WITNESS 0
129 #define LI_RECURSEMASK 0x0000ffff /* Recursion depth of lock instance. */
130 #define LI_EXCLUSIVE 0x00010000 /* Exclusive lock instance. */
131 #define LI_NORELEASE 0x00020000 /* Lock not allowed to be released. */
132 #define LI_SLEEPABLE 0x00040000 /* Lock may be held while sleeping. */
134 #ifndef WITNESS_COUNT
135 #define WITNESS_COUNT 1536
137 #define WITNESS_HASH_SIZE 251 /* Prime, gives load factor < 2 */
138 #define WITNESS_PENDLIST (512 + (MAXCPU * 4))
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 FULLGRAPH_SBUF_SIZE 512
159 * These flags go in the witness relationship matrix and describe the
160 * relationship between any two struct witness objects.
162 #define WITNESS_UNRELATED 0x00 /* No lock order relation. */
163 #define WITNESS_PARENT 0x01 /* Parent, aka direct ancestor. */
164 #define WITNESS_ANCESTOR 0x02 /* Direct or indirect ancestor. */
165 #define WITNESS_CHILD 0x04 /* Child, aka direct descendant. */
166 #define WITNESS_DESCENDANT 0x08 /* Direct or indirect descendant. */
167 #define WITNESS_ANCESTOR_MASK (WITNESS_PARENT | WITNESS_ANCESTOR)
168 #define WITNESS_DESCENDANT_MASK (WITNESS_CHILD | WITNESS_DESCENDANT)
169 #define WITNESS_RELATED_MASK \
170 (WITNESS_ANCESTOR_MASK | WITNESS_DESCENDANT_MASK)
171 #define WITNESS_REVERSAL 0x10 /* A lock order reversal has been
173 #define WITNESS_RESERVED1 0x20 /* Unused flag, reserved. */
174 #define WITNESS_RESERVED2 0x40 /* Unused flag, reserved. */
175 #define WITNESS_LOCK_ORDER_KNOWN 0x80 /* This lock order is known. */
177 /* Descendant to ancestor flags */
178 #define WITNESS_DTOA(x) (((x) & WITNESS_RELATED_MASK) >> 2)
180 /* Ancestor to descendant flags */
181 #define WITNESS_ATOD(x) (((x) & WITNESS_RELATED_MASK) << 2)
183 #define WITNESS_INDEX_ASSERT(i) \
184 MPASS((i) > 0 && (i) <= w_max_used_index && (i) < witness_count)
186 static MALLOC_DEFINE(M_WITNESS, "Witness", "Witness");
189 * Lock instances. A lock instance is the data associated with a lock while
190 * it is held by witness. For example, a lock instance will hold the
191 * recursion count of a lock. Lock instances are held in lists. Spin locks
192 * are held in a per-cpu list while sleep locks are held in per-thread list.
194 struct lock_instance {
195 struct lock_object *li_lock;
202 * A simple list type used to build the list of locks held by a thread
203 * or CPU. We can't simply embed the list in struct lock_object since a
204 * lock may be held by more than one thread if it is a shared lock. Locks
205 * are added to the head of the list, so we fill up each list entry from
206 * "the back" logically. To ease some of the arithmetic, we actually fill
207 * in each list entry the normal way (children[0] then children[1], etc.) but
208 * when we traverse the list we read children[count-1] as the first entry
209 * down to children[0] as the final entry.
211 struct lock_list_entry {
212 struct lock_list_entry *ll_next;
213 struct lock_instance ll_children[LOCK_NCHILDREN];
218 * The main witness structure. One of these per named lock type in the system
219 * (for example, "vnode interlock").
222 char w_name[MAX_W_NAME];
223 uint32_t w_index; /* Index in the relationship matrix */
224 struct lock_class *w_class;
225 STAILQ_ENTRY(witness) w_list; /* List of all witnesses. */
226 STAILQ_ENTRY(witness) w_typelist; /* Witnesses of a type. */
227 struct witness *w_hash_next; /* Linked list in hash buckets. */
228 const char *w_file; /* File where last acquired */
229 uint32_t w_line; /* Line where last acquired */
231 uint16_t w_num_ancestors; /* direct/indirect
233 uint16_t w_num_descendants; /* direct/indirect
234 * descendant count */
236 unsigned w_displayed:1;
237 unsigned w_reversed:1;
240 STAILQ_HEAD(witness_list, witness);
243 * The witness hash table. Keys are witness names (const char *), elements are
244 * witness objects (struct witness *).
246 struct witness_hash {
247 struct witness *wh_array[WITNESS_HASH_SIZE];
253 * Key type for the lock order data hash table.
255 struct witness_lock_order_key {
260 struct witness_lock_order_data {
261 struct stack wlod_stack;
262 struct witness_lock_order_key wlod_key;
263 struct witness_lock_order_data *wlod_next;
267 * The witness lock order data hash table. Keys are witness index tuples
268 * (struct witness_lock_order_key), elements are lock order data objects
269 * (struct witness_lock_order_data).
271 struct witness_lock_order_hash {
272 struct witness_lock_order_data *wloh_array[WITNESS_LO_HASH_SIZE];
277 struct witness_blessed {
282 struct witness_pendhelp {
284 struct lock_object *wh_lock;
287 struct witness_order_list_entry {
289 struct lock_class *w_class;
293 * Returns 0 if one of the locks is a spin lock and the other is not.
294 * Returns 1 otherwise.
297 witness_lock_type_equal(struct witness *w1, struct witness *w2)
300 return ((w1->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) ==
301 (w2->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)));
305 witness_lock_order_key_equal(const struct witness_lock_order_key *a,
306 const struct witness_lock_order_key *b)
309 return (a->from == b->from && a->to == b->to);
312 static int _isitmyx(struct witness *w1, struct witness *w2, int rmask,
314 static void adopt(struct witness *parent, struct witness *child);
315 static int blessed(struct witness *, struct witness *);
316 static void depart(struct witness *w);
317 static struct witness *enroll(const char *description,
318 struct lock_class *lock_class);
319 static struct lock_instance *find_instance(struct lock_list_entry *list,
320 const struct lock_object *lock);
321 static int isitmychild(struct witness *parent, struct witness *child);
322 static int isitmydescendant(struct witness *parent, struct witness *child);
323 static void itismychild(struct witness *parent, struct witness *child);
324 static int sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS);
325 static int sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS);
326 static int sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS);
327 static int sysctl_debug_witness_channel(SYSCTL_HANDLER_ARGS);
328 static void witness_add_fullgraph(struct sbuf *sb, struct witness *parent);
330 static void witness_ddb_compute_levels(void);
331 static void witness_ddb_display(int(*)(const char *fmt, ...));
332 static void witness_ddb_display_descendants(int(*)(const char *fmt, ...),
333 struct witness *, int indent);
334 static void witness_ddb_display_list(int(*prnt)(const char *fmt, ...),
335 struct witness_list *list);
336 static void witness_ddb_level_descendants(struct witness *parent, int l);
337 static void witness_ddb_list(struct thread *td);
339 static void witness_enter_debugger(const char *msg);
340 static void witness_debugger(int cond, const char *msg);
341 static void witness_free(struct witness *m);
342 static struct witness *witness_get(void);
343 static uint32_t witness_hash_djb2(const uint8_t *key, uint32_t size);
344 static struct witness *witness_hash_get(const char *key);
345 static void witness_hash_put(struct witness *w);
346 static void witness_init_hash_tables(void);
347 static void witness_increment_graph_generation(void);
348 static void witness_lock_list_free(struct lock_list_entry *lle);
349 static struct lock_list_entry *witness_lock_list_get(void);
350 static int witness_lock_order_add(struct witness *parent,
351 struct witness *child);
352 static int witness_lock_order_check(struct witness *parent,
353 struct witness *child);
354 static struct witness_lock_order_data *witness_lock_order_get(
355 struct witness *parent,
356 struct witness *child);
357 static void witness_list_lock(struct lock_instance *instance,
358 int (*prnt)(const char *fmt, ...));
359 static int witness_output(const char *fmt, ...) __printflike(1, 2);
360 static int witness_output_drain(void *arg __unused, const char *data,
362 static int witness_voutput(const char *fmt, va_list ap) __printflike(1, 0);
363 static void witness_setflag(struct lock_object *lock, int flag, int set);
365 FEATURE(witness, "kernel has witness(9) support");
367 static SYSCTL_NODE(_debug, OID_AUTO, witness, CTLFLAG_RW | CTLFLAG_MPSAFE, NULL,
371 * If set to 0, lock order checking is disabled. If set to -1,
372 * witness is completely disabled. Otherwise witness performs full
373 * lock order checking for all locks. At runtime, lock order checking
374 * may be toggled. However, witness cannot be reenabled once it is
375 * completely disabled.
377 static int witness_watch = 1;
378 SYSCTL_PROC(_debug_witness, OID_AUTO, watch,
379 CTLFLAG_RWTUN | CTLTYPE_INT | CTLFLAG_MPSAFE, NULL, 0,
380 sysctl_debug_witness_watch, "I",
381 "witness is watching lock operations");
385 * When KDB is enabled and witness_kdb is 1, it will cause the system
386 * to drop into kdebug() when:
387 * - a lock hierarchy violation occurs
388 * - locks are held when going to sleep.
395 SYSCTL_INT(_debug_witness, OID_AUTO, kdb, CTLFLAG_RWTUN, &witness_kdb, 0, "");
398 #if defined(DDB) || defined(KDB)
400 * When DDB or KDB is enabled and witness_trace is 1, it will cause the system
401 * to print a stack trace:
402 * - a lock hierarchy violation occurs
403 * - locks are held when going to sleep.
405 int witness_trace = 1;
406 SYSCTL_INT(_debug_witness, OID_AUTO, trace, CTLFLAG_RWTUN, &witness_trace, 0, "");
407 #endif /* DDB || KDB */
409 #ifdef WITNESS_SKIPSPIN
410 int witness_skipspin = 1;
412 int witness_skipspin = 0;
414 SYSCTL_INT(_debug_witness, OID_AUTO, skipspin, CTLFLAG_RDTUN, &witness_skipspin, 0, "");
416 int badstack_sbuf_size;
418 int witness_count = WITNESS_COUNT;
419 SYSCTL_INT(_debug_witness, OID_AUTO, witness_count, CTLFLAG_RDTUN,
420 &witness_count, 0, "");
423 * Output channel for witness messages. By default we print to the console.
425 enum witness_channel {
431 static enum witness_channel witness_channel = WITNESS_CONSOLE;
432 SYSCTL_PROC(_debug_witness, OID_AUTO, output_channel,
433 CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_MPSAFE, NULL, 0,
434 sysctl_debug_witness_channel, "A",
435 "Output channel for warnings");
438 * Call this to print out the relations between locks.
440 SYSCTL_PROC(_debug_witness, OID_AUTO, fullgraph,
441 CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
442 sysctl_debug_witness_fullgraph, "A",
443 "Show locks relation graphs");
446 * Call this to print out the witness faulty stacks.
448 SYSCTL_PROC(_debug_witness, OID_AUTO, badstacks,
449 CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
450 sysctl_debug_witness_badstacks, "A",
451 "Show bad witness stacks");
453 static struct mtx w_mtx;
456 static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free);
457 static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all);
460 static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin);
461 static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep);
464 static struct lock_list_entry *w_lock_list_free = NULL;
465 static struct witness_pendhelp pending_locks[WITNESS_PENDLIST];
466 static u_int pending_cnt;
468 static int w_free_cnt, w_spin_cnt, w_sleep_cnt;
469 SYSCTL_INT(_debug_witness, OID_AUTO, free_cnt, CTLFLAG_RD, &w_free_cnt, 0, "");
470 SYSCTL_INT(_debug_witness, OID_AUTO, spin_cnt, CTLFLAG_RD, &w_spin_cnt, 0, "");
471 SYSCTL_INT(_debug_witness, OID_AUTO, sleep_cnt, CTLFLAG_RD, &w_sleep_cnt, 0,
474 static struct witness *w_data;
475 static uint8_t **w_rmatrix;
476 static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT];
477 static struct witness_hash w_hash; /* The witness hash table. */
479 /* The lock order data hash */
480 static struct witness_lock_order_data w_lodata[WITNESS_LO_DATA_COUNT];
481 static struct witness_lock_order_data *w_lofree = NULL;
482 static struct witness_lock_order_hash w_lohash;
483 static int w_max_used_index = 0;
484 static unsigned int w_generation = 0;
485 static const char w_notrunning[] = "Witness not running\n";
486 static const char w_stillcold[] = "Witness is still cold\n";
488 static const char w_notallowed[] = "The sysctl is disabled on the arch\n";
491 static struct witness_order_list_entry order_lists[] = {
495 { "proctree", &lock_class_sx },
496 { "allproc", &lock_class_sx },
497 { "allprison", &lock_class_sx },
502 { "Giant", &lock_class_mtx_sleep },
503 { "pipe mutex", &lock_class_mtx_sleep },
504 { "sigio lock", &lock_class_mtx_sleep },
505 { "process group", &lock_class_mtx_sleep },
507 { "pmc-sleep", &lock_class_mtx_sleep },
509 { "process lock", &lock_class_mtx_sleep },
510 { "session", &lock_class_mtx_sleep },
511 { "uidinfo hash", &lock_class_rw },
512 { "time lock", &lock_class_mtx_sleep },
517 { "umtx lock", &lock_class_mtx_sleep },
522 { "accept", &lock_class_mtx_sleep },
523 { "so_snd", &lock_class_mtx_sleep },
524 { "so_rcv", &lock_class_mtx_sleep },
525 { "sellck", &lock_class_mtx_sleep },
530 { "so_rcv", &lock_class_mtx_sleep },
531 { "radix node head", &lock_class_rm },
532 { "ifaddr", &lock_class_mtx_sleep },
536 * protocol locks before interface locks, after UDP locks.
538 { "in_multi_sx", &lock_class_sx },
539 { "udpinp", &lock_class_rw },
540 { "in_multi_list_mtx", &lock_class_mtx_sleep },
541 { "igmp_mtx", &lock_class_mtx_sleep },
542 { "ifnet_rw", &lock_class_rw },
543 { "if_addr_lock", &lock_class_mtx_sleep },
547 * protocol locks before interface locks, after UDP locks.
549 { "in6_multi_sx", &lock_class_sx },
550 { "udpinp", &lock_class_rw },
551 { "in6_multi_list_mtx", &lock_class_mtx_sleep },
552 { "mld_mtx", &lock_class_mtx_sleep },
553 { "ifnet_rw", &lock_class_rw },
554 { "if_addr_lock", &lock_class_mtx_sleep },
557 * UNIX Domain Sockets
559 { "unp_link_rwlock", &lock_class_rw },
560 { "unp_list_lock", &lock_class_mtx_sleep },
561 { "unp", &lock_class_mtx_sleep },
562 { "so_snd", &lock_class_mtx_sleep },
567 { "udp", &lock_class_mtx_sleep },
568 { "udpinp", &lock_class_rw },
569 { "so_snd", &lock_class_mtx_sleep },
574 { "tcp", &lock_class_mtx_sleep },
575 { "tcpinp", &lock_class_rw },
576 { "so_snd", &lock_class_mtx_sleep },
581 { "bpf global lock", &lock_class_sx },
582 { "bpf cdev lock", &lock_class_mtx_sleep },
587 { "nfsd_mtx", &lock_class_mtx_sleep },
588 { "so_snd", &lock_class_mtx_sleep },
594 { "802.11 com lock", &lock_class_mtx_sleep},
599 { "network driver", &lock_class_mtx_sleep},
605 { "ng_node", &lock_class_mtx_sleep },
606 { "ng_worklist", &lock_class_mtx_sleep },
611 { "vm map (system)", &lock_class_mtx_sleep },
612 { "vnode interlock", &lock_class_mtx_sleep },
613 { "cdev", &lock_class_mtx_sleep },
614 { "devthrd", &lock_class_mtx_sleep },
619 { "vm map (user)", &lock_class_sx },
620 { "vm object", &lock_class_rw },
621 { "vm page", &lock_class_mtx_sleep },
622 { "pmap pv global", &lock_class_rw },
623 { "pmap", &lock_class_mtx_sleep },
624 { "pmap pv list", &lock_class_rw },
625 { "vm page free queue", &lock_class_mtx_sleep },
626 { "vm pagequeue", &lock_class_mtx_sleep },
629 * kqueue/VFS interaction
631 { "kqueue", &lock_class_mtx_sleep },
632 { "struct mount mtx", &lock_class_mtx_sleep },
633 { "vnode interlock", &lock_class_mtx_sleep },
638 { "ncvn", &lock_class_mtx_sleep },
639 { "ncbuc", &lock_class_mtx_sleep },
640 { "vnode interlock", &lock_class_mtx_sleep },
641 { "ncneg", &lock_class_mtx_sleep },
646 { "dn->dn_mtx", &lock_class_sx },
647 { "dr->dt.di.dr_mtx", &lock_class_sx },
648 { "db->db_mtx", &lock_class_sx },
653 { "TCP ID tree", &lock_class_rw },
654 { "tcp log id bucket", &lock_class_mtx_sleep },
655 { "tcpinp", &lock_class_rw },
656 { "TCP log expireq", &lock_class_mtx_sleep },
662 { "ap boot", &lock_class_mtx_spin },
664 { "rm.mutex_mtx", &lock_class_mtx_spin },
666 { "cy", &lock_class_mtx_spin },
668 { "scc_hwmtx", &lock_class_mtx_spin },
669 { "uart_hwmtx", &lock_class_mtx_spin },
670 { "fast_taskqueue", &lock_class_mtx_spin },
671 { "intr table", &lock_class_mtx_spin },
672 { "process slock", &lock_class_mtx_spin },
673 { "syscons video lock", &lock_class_mtx_spin },
674 { "sleepq chain", &lock_class_mtx_spin },
675 { "rm_spinlock", &lock_class_mtx_spin },
676 { "turnstile chain", &lock_class_mtx_spin },
677 { "turnstile lock", &lock_class_mtx_spin },
678 { "sched lock", &lock_class_mtx_spin },
679 { "td_contested", &lock_class_mtx_spin },
680 { "callout", &lock_class_mtx_spin },
681 { "entropy harvest mutex", &lock_class_mtx_spin },
683 { "smp rendezvous", &lock_class_mtx_spin },
686 { "tlb0", &lock_class_mtx_spin },
689 { "sched lock", &lock_class_mtx_spin },
691 { "pmc-per-proc", &lock_class_mtx_spin },
697 { "intrcnt", &lock_class_mtx_spin },
698 { "icu", &lock_class_mtx_spin },
700 { "allpmaps", &lock_class_mtx_spin },
701 { "descriptor tables", &lock_class_mtx_spin },
703 { "clk", &lock_class_mtx_spin },
704 { "cpuset", &lock_class_mtx_spin },
705 { "mprof lock", &lock_class_mtx_spin },
706 { "zombie lock", &lock_class_mtx_spin },
707 { "ALD Queue", &lock_class_mtx_spin },
708 #if defined(__i386__) || defined(__amd64__)
709 { "pcicfg", &lock_class_mtx_spin },
710 { "NDIS thread lock", &lock_class_mtx_spin },
712 { "tw_osl_io_lock", &lock_class_mtx_spin },
713 { "tw_osl_q_lock", &lock_class_mtx_spin },
714 { "tw_cl_io_lock", &lock_class_mtx_spin },
715 { "tw_cl_intr_lock", &lock_class_mtx_spin },
716 { "tw_cl_gen_lock", &lock_class_mtx_spin },
718 { "pmc-leaf", &lock_class_mtx_spin },
720 { "blocked lock", &lock_class_mtx_spin },
726 * Pairs of locks which have been blessed. Witness does not complain about
727 * order problems with blessed lock pairs. Please do not add an entry to the
728 * table without an explanatory comment.
730 static struct witness_blessed blessed_list[] = {
732 * See the comment in ufs_dirhash.c. Basically, a vnode lock serializes
733 * both lock orders, so a deadlock cannot happen as a result of this
736 { "dirhash", "bufwait" },
739 * A UFS vnode may be locked in vget() while a buffer belonging to the
740 * parent directory vnode is locked.
742 { "ufs", "bufwait" },
746 * This global is set to 0 once it becomes safe to use the witness code.
748 static int witness_cold = 1;
751 * This global is set to 1 once the static lock orders have been enrolled
752 * so that a warning can be issued for any spin locks enrolled later.
754 static int witness_spin_warn = 0;
756 /* Trim useless garbage from filenames. */
758 fixup_filename(const char *file)
763 while (strncmp(file, "../", 3) == 0)
769 * Calculate the size of early witness structures.
772 witness_startup_count(void)
776 sz = sizeof(struct witness) * witness_count;
777 sz += sizeof(*w_rmatrix) * (witness_count + 1);
778 sz += sizeof(*w_rmatrix[0]) * (witness_count + 1) *
785 * The WITNESS-enabled diagnostic code. Note that the witness code does
786 * assume that the early boot is single-threaded at least until after this
787 * routine is completed.
790 witness_startup(void *mem)
792 struct lock_object *lock;
793 struct witness_order_list_entry *order;
794 struct witness *w, *w1;
800 p += sizeof(struct witness) * witness_count;
802 w_rmatrix = (void *)p;
803 p += sizeof(*w_rmatrix) * (witness_count + 1);
805 for (i = 0; i < witness_count + 1; i++) {
806 w_rmatrix[i] = (void *)p;
807 p += sizeof(*w_rmatrix[i]) * (witness_count + 1);
809 badstack_sbuf_size = witness_count * 256;
812 * We have to release Giant before initializing its witness
813 * structure so that WITNESS doesn't get confused.
816 mtx_assert(&Giant, MA_NOTOWNED);
818 CTR1(KTR_WITNESS, "%s: initializing witness", __func__);
819 mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET |
820 MTX_NOWITNESS | MTX_NOPROFILE);
821 for (i = witness_count - 1; i >= 0; i--) {
823 memset(w, 0, sizeof(*w));
824 w_data[i].w_index = i; /* Witness index never changes. */
827 KASSERT(STAILQ_FIRST(&w_free)->w_index == 0,
828 ("%s: Invalid list of free witness objects", __func__));
830 /* Witness with index 0 is not used to aid in debugging. */
831 STAILQ_REMOVE_HEAD(&w_free, w_list);
834 for (i = 0; i < witness_count; i++) {
835 memset(w_rmatrix[i], 0, sizeof(*w_rmatrix[i]) *
836 (witness_count + 1));
839 for (i = 0; i < LOCK_CHILDCOUNT; i++)
840 witness_lock_list_free(&w_locklistdata[i]);
841 witness_init_hash_tables();
843 /* First add in all the specified order lists. */
844 for (order = order_lists; order->w_name != NULL; order++) {
845 w = enroll(order->w_name, order->w_class);
848 w->w_file = "order list";
849 for (order++; order->w_name != NULL; order++) {
850 w1 = enroll(order->w_name, order->w_class);
853 w1->w_file = "order list";
858 witness_spin_warn = 1;
860 /* Iterate through all locks and add them to witness. */
861 for (i = 0; pending_locks[i].wh_lock != NULL; i++) {
862 lock = pending_locks[i].wh_lock;
863 KASSERT(lock->lo_flags & LO_WITNESS,
864 ("%s: lock %s is on pending list but not LO_WITNESS",
865 __func__, lock->lo_name));
866 lock->lo_witness = enroll(pending_locks[i].wh_type,
870 /* Mark the witness code as being ready for use. */
877 witness_init(struct lock_object *lock, const char *type)
879 struct lock_class *class;
881 /* Various sanity checks. */
882 class = LOCK_CLASS(lock);
883 if ((lock->lo_flags & LO_RECURSABLE) != 0 &&
884 (class->lc_flags & LC_RECURSABLE) == 0)
885 kassert_panic("%s: lock (%s) %s can not be recursable",
886 __func__, class->lc_name, lock->lo_name);
887 if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
888 (class->lc_flags & LC_SLEEPABLE) == 0)
889 kassert_panic("%s: lock (%s) %s can not be sleepable",
890 __func__, class->lc_name, lock->lo_name);
891 if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
892 (class->lc_flags & LC_UPGRADABLE) == 0)
893 kassert_panic("%s: lock (%s) %s can not be upgradable",
894 __func__, class->lc_name, lock->lo_name);
897 * If we shouldn't watch this lock, then just clear lo_witness.
898 * Otherwise, if witness_cold is set, then it is too early to
899 * enroll this lock, so defer it to witness_initialize() by adding
900 * it to the pending_locks list. If it is not too early, then enroll
903 if (witness_watch < 1 || KERNEL_PANICKED() ||
904 (lock->lo_flags & LO_WITNESS) == 0)
905 lock->lo_witness = NULL;
906 else if (witness_cold) {
907 pending_locks[pending_cnt].wh_lock = lock;
908 pending_locks[pending_cnt++].wh_type = type;
909 if (pending_cnt > WITNESS_PENDLIST)
910 panic("%s: pending locks list is too small, "
911 "increase WITNESS_PENDLIST\n",
914 lock->lo_witness = enroll(type, class);
918 witness_destroy(struct lock_object *lock)
920 struct lock_class *class;
923 class = LOCK_CLASS(lock);
926 panic("lock (%s) %s destroyed while witness_cold",
927 class->lc_name, lock->lo_name);
929 /* XXX: need to verify that no one holds the lock */
930 if ((lock->lo_flags & LO_WITNESS) == 0 || lock->lo_witness == NULL)
932 w = lock->lo_witness;
934 mtx_lock_spin(&w_mtx);
935 MPASS(w->w_refcount > 0);
938 if (w->w_refcount == 0)
940 mtx_unlock_spin(&w_mtx);
945 witness_ddb_compute_levels(void)
950 * First clear all levels.
952 STAILQ_FOREACH(w, &w_all, w_list)
956 * Look for locks with no parents and level all their descendants.
958 STAILQ_FOREACH(w, &w_all, w_list) {
959 /* If the witness has ancestors (is not a root), skip it. */
960 if (w->w_num_ancestors > 0)
962 witness_ddb_level_descendants(w, 0);
967 witness_ddb_level_descendants(struct witness *w, int l)
971 if (w->w_ddb_level >= l)
977 for (i = 1; i <= w_max_used_index; i++) {
978 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
979 witness_ddb_level_descendants(&w_data[i], l);
984 witness_ddb_display_descendants(int(*prnt)(const char *fmt, ...),
985 struct witness *w, int indent)
989 for (i = 0; i < indent; i++)
991 prnt("%s (type: %s, depth: %d, active refs: %d)",
992 w->w_name, w->w_class->lc_name,
993 w->w_ddb_level, w->w_refcount);
994 if (w->w_displayed) {
995 prnt(" -- (already displayed)\n");
999 if (w->w_file != NULL && w->w_line != 0)
1000 prnt(" -- last acquired @ %s:%d\n", fixup_filename(w->w_file),
1003 prnt(" -- never acquired\n");
1005 WITNESS_INDEX_ASSERT(w->w_index);
1006 for (i = 1; i <= w_max_used_index; i++) {
1009 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
1010 witness_ddb_display_descendants(prnt, &w_data[i],
1016 witness_ddb_display_list(int(*prnt)(const char *fmt, ...),
1017 struct witness_list *list)
1021 STAILQ_FOREACH(w, list, w_typelist) {
1022 if (w->w_file == NULL || w->w_ddb_level > 0)
1025 /* This lock has no anscestors - display its descendants. */
1026 witness_ddb_display_descendants(prnt, w, 0);
1033 witness_ddb_display(int(*prnt)(const char *fmt, ...))
1037 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1038 witness_ddb_compute_levels();
1040 /* Clear all the displayed flags. */
1041 STAILQ_FOREACH(w, &w_all, w_list)
1045 * First, handle sleep locks which have been acquired at least
1048 prnt("Sleep locks:\n");
1049 witness_ddb_display_list(prnt, &w_sleep);
1054 * Now do spin locks which have been acquired at least once.
1056 prnt("\nSpin locks:\n");
1057 witness_ddb_display_list(prnt, &w_spin);
1062 * Finally, any locks which have not been acquired yet.
1064 prnt("\nLocks which were never acquired:\n");
1065 STAILQ_FOREACH(w, &w_all, w_list) {
1066 if (w->w_file != NULL || w->w_refcount == 0)
1068 prnt("%s (type: %s, depth: %d)\n", w->w_name,
1069 w->w_class->lc_name, w->w_ddb_level);
1077 witness_defineorder(struct lock_object *lock1, struct lock_object *lock2)
1080 if (witness_watch == -1 || KERNEL_PANICKED())
1083 /* Require locks that witness knows about. */
1084 if (lock1 == NULL || lock1->lo_witness == NULL || lock2 == NULL ||
1085 lock2->lo_witness == NULL)
1088 mtx_assert(&w_mtx, MA_NOTOWNED);
1089 mtx_lock_spin(&w_mtx);
1092 * If we already have either an explicit or implied lock order that
1093 * is the other way around, then return an error.
1095 if (witness_watch &&
1096 isitmydescendant(lock2->lo_witness, lock1->lo_witness)) {
1097 mtx_unlock_spin(&w_mtx);
1101 /* Try to add the new order. */
1102 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1103 lock2->lo_witness->w_name, lock1->lo_witness->w_name);
1104 itismychild(lock1->lo_witness, lock2->lo_witness);
1105 mtx_unlock_spin(&w_mtx);
1110 witness_checkorder(struct lock_object *lock, int flags, const char *file,
1111 int line, struct lock_object *interlock)
1113 struct lock_list_entry *lock_list, *lle;
1114 struct lock_instance *lock1, *lock2, *plock;
1115 struct lock_class *class, *iclass;
1116 struct witness *w, *w1;
1120 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL ||
1124 w = lock->lo_witness;
1125 class = LOCK_CLASS(lock);
1128 if (class->lc_flags & LC_SLEEPLOCK) {
1130 * Since spin locks include a critical section, this check
1131 * implicitly enforces a lock order of all sleep locks before
1134 if (td->td_critnest != 0 && !kdb_active)
1135 kassert_panic("acquiring blockable sleep lock with "
1136 "spinlock or critical section held (%s) %s @ %s:%d",
1137 class->lc_name, lock->lo_name,
1138 fixup_filename(file), line);
1141 * If this is the first lock acquired then just return as
1142 * no order checking is needed.
1144 lock_list = td->td_sleeplocks;
1145 if (lock_list == NULL || lock_list->ll_count == 0)
1149 * If this is the first lock, just return as no order
1150 * checking is needed. Avoid problems with thread
1151 * migration pinning the thread while checking if
1152 * spinlocks are held. If at least one spinlock is held
1153 * the thread is in a safe path and it is allowed to
1157 lock_list = PCPU_GET(spinlocks);
1158 if (lock_list == NULL || lock_list->ll_count == 0) {
1166 * Check to see if we are recursing on a lock we already own. If
1167 * so, make sure that we don't mismatch exclusive and shared lock
1170 lock1 = find_instance(lock_list, lock);
1171 if (lock1 != NULL) {
1172 if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
1173 (flags & LOP_EXCLUSIVE) == 0) {
1174 witness_output("shared lock of (%s) %s @ %s:%d\n",
1175 class->lc_name, lock->lo_name,
1176 fixup_filename(file), line);
1177 witness_output("while exclusively locked from %s:%d\n",
1178 fixup_filename(lock1->li_file), lock1->li_line);
1179 kassert_panic("excl->share");
1181 if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
1182 (flags & LOP_EXCLUSIVE) != 0) {
1183 witness_output("exclusive lock of (%s) %s @ %s:%d\n",
1184 class->lc_name, lock->lo_name,
1185 fixup_filename(file), line);
1186 witness_output("while share locked from %s:%d\n",
1187 fixup_filename(lock1->li_file), lock1->li_line);
1188 kassert_panic("share->excl");
1193 /* Warn if the interlock is not locked exactly once. */
1194 if (interlock != NULL) {
1195 iclass = LOCK_CLASS(interlock);
1196 lock1 = find_instance(lock_list, interlock);
1198 kassert_panic("interlock (%s) %s not locked @ %s:%d",
1199 iclass->lc_name, interlock->lo_name,
1200 fixup_filename(file), line);
1201 else if ((lock1->li_flags & LI_RECURSEMASK) != 0)
1202 kassert_panic("interlock (%s) %s recursed @ %s:%d",
1203 iclass->lc_name, interlock->lo_name,
1204 fixup_filename(file), line);
1208 * Find the previously acquired lock, but ignore interlocks.
1210 plock = &lock_list->ll_children[lock_list->ll_count - 1];
1211 if (interlock != NULL && plock->li_lock == interlock) {
1212 if (lock_list->ll_count > 1)
1214 &lock_list->ll_children[lock_list->ll_count - 2];
1216 lle = lock_list->ll_next;
1219 * The interlock is the only lock we hold, so
1224 plock = &lle->ll_children[lle->ll_count - 1];
1229 * Try to perform most checks without a lock. If this succeeds we
1230 * can skip acquiring the lock and return success. Otherwise we redo
1231 * the check with the lock held to handle races with concurrent updates.
1233 w1 = plock->li_lock->lo_witness;
1234 if (witness_lock_order_check(w1, w))
1237 mtx_lock_spin(&w_mtx);
1238 if (witness_lock_order_check(w1, w)) {
1239 mtx_unlock_spin(&w_mtx);
1242 witness_lock_order_add(w1, w);
1245 * Check for duplicate locks of the same type. Note that we only
1246 * have to check for this on the last lock we just acquired. Any
1247 * other cases will be caught as lock order violations.
1251 if (!(lock->lo_flags & LO_DUPOK) && !(flags & LOP_DUPOK) &&
1252 !(w_rmatrix[i][i] & WITNESS_REVERSAL)) {
1253 w_rmatrix[i][i] |= WITNESS_REVERSAL;
1255 mtx_unlock_spin(&w_mtx);
1257 "acquiring duplicate lock of same type: \"%s\"\n",
1259 witness_output(" 1st %s @ %s:%d\n", plock->li_lock->lo_name,
1260 fixup_filename(plock->li_file), plock->li_line);
1261 witness_output(" 2nd %s @ %s:%d\n", lock->lo_name,
1262 fixup_filename(file), line);
1263 witness_debugger(1, __func__);
1265 mtx_unlock_spin(&w_mtx);
1268 mtx_assert(&w_mtx, MA_OWNED);
1271 * If we know that the lock we are acquiring comes after
1272 * the lock we most recently acquired in the lock order tree,
1273 * then there is no need for any further checks.
1275 if (isitmychild(w1, w))
1278 for (j = 0, lle = lock_list; lle != NULL; lle = lle->ll_next) {
1279 for (i = lle->ll_count - 1; i >= 0; i--, j++) {
1280 struct stack pstack;
1281 bool pstackv, trace;
1283 MPASS(j < LOCK_CHILDCOUNT * LOCK_NCHILDREN);
1284 lock1 = &lle->ll_children[i];
1287 * Ignore the interlock.
1289 if (interlock == lock1->li_lock)
1293 * If this lock doesn't undergo witness checking,
1296 w1 = lock1->li_lock->lo_witness;
1298 KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
1299 ("lock missing witness structure"));
1304 * If we are locking Giant and this is a sleepable
1305 * lock, then skip it.
1307 if ((lock1->li_flags & LI_SLEEPABLE) != 0 &&
1308 lock == &Giant.lock_object)
1312 * If we are locking a sleepable lock and this lock
1313 * is Giant, then skip it.
1315 if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1316 (flags & LOP_NOSLEEP) == 0 &&
1317 lock1->li_lock == &Giant.lock_object)
1321 * If we are locking a sleepable lock and this lock
1322 * isn't sleepable, we want to treat it as a lock
1323 * order violation to enfore a general lock order of
1324 * sleepable locks before non-sleepable locks.
1326 if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1327 (flags & LOP_NOSLEEP) == 0 &&
1328 (lock1->li_flags & LI_SLEEPABLE) == 0)
1332 * If we are locking Giant and this is a non-sleepable
1333 * lock, then treat it as a reversal.
1335 if ((lock1->li_flags & LI_SLEEPABLE) == 0 &&
1336 lock == &Giant.lock_object)
1340 * Check the lock order hierarchy for a reveresal.
1342 if (!isitmydescendant(w, w1))
1347 * We have a lock order violation, check to see if it
1348 * is allowed or has already been yelled about.
1351 /* Bail if this violation is known */
1352 if (w_rmatrix[w1->w_index][w->w_index] & WITNESS_REVERSAL)
1355 /* Record this as a violation */
1356 w_rmatrix[w1->w_index][w->w_index] |= WITNESS_REVERSAL;
1357 w_rmatrix[w->w_index][w1->w_index] |= WITNESS_REVERSAL;
1358 w->w_reversed = w1->w_reversed = 1;
1359 witness_increment_graph_generation();
1362 * If the lock order is blessed, bail before logging
1363 * anything. We don't look for other lock order
1364 * violations though, which may be a bug.
1369 trace = atomic_load_int(&witness_trace);
1371 struct witness_lock_order_data *data;
1374 data = witness_lock_order_get(w, w1);
1376 stack_copy(&data->wlod_stack,
1381 mtx_unlock_spin(&w_mtx);
1383 #ifdef WITNESS_NO_VNODE
1385 * There are known LORs between VNODE locks. They are
1386 * not an indication of a bug. VNODE locks are flagged
1387 * as such (LO_IS_VNODE) and we don't yell if the LOR
1388 * is between 2 VNODE locks.
1390 if ((lock->lo_flags & LO_IS_VNODE) != 0 &&
1391 (lock1->li_lock->lo_flags & LO_IS_VNODE) != 0)
1396 * Ok, yell about it.
1398 if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1399 (flags & LOP_NOSLEEP) == 0 &&
1400 (lock1->li_flags & LI_SLEEPABLE) == 0)
1402 "lock order reversal: (sleepable after non-sleepable)\n");
1403 else if ((lock1->li_flags & LI_SLEEPABLE) == 0
1404 && lock == &Giant.lock_object)
1406 "lock order reversal: (Giant after non-sleepable)\n");
1408 witness_output("lock order reversal:\n");
1411 * Try to locate an earlier lock with
1412 * witness w in our list.
1415 lock2 = &lle->ll_children[i];
1416 MPASS(lock2->li_lock != NULL);
1417 if (lock2->li_lock->lo_witness == w)
1419 if (i == 0 && lle->ll_next != NULL) {
1421 i = lle->ll_count - 1;
1422 MPASS(i >= 0 && i < LOCK_NCHILDREN);
1427 witness_output(" 1st %p %s (%s, %s) @ %s:%d\n",
1428 lock1->li_lock, lock1->li_lock->lo_name,
1429 w1->w_name, w1->w_class->lc_name,
1430 fixup_filename(lock1->li_file),
1432 witness_output(" 2nd %p %s (%s, %s) @ %s:%d\n",
1433 lock, lock->lo_name, w->w_name,
1434 w->w_class->lc_name, fixup_filename(file),
1437 struct witness *w2 = lock2->li_lock->lo_witness;
1439 witness_output(" 1st %p %s (%s, %s) @ %s:%d\n",
1440 lock2->li_lock, lock2->li_lock->lo_name,
1441 w2->w_name, w2->w_class->lc_name,
1442 fixup_filename(lock2->li_file),
1444 witness_output(" 2nd %p %s (%s, %s) @ %s:%d\n",
1445 lock1->li_lock, lock1->li_lock->lo_name,
1446 w1->w_name, w1->w_class->lc_name,
1447 fixup_filename(lock1->li_file),
1449 witness_output(" 3rd %p %s (%s, %s) @ %s:%d\n", lock,
1450 lock->lo_name, w->w_name,
1451 w->w_class->lc_name, fixup_filename(file),
1458 sbuf_new(&sb, buf, sizeof(buf), SBUF_FIXEDLEN);
1459 sbuf_set_drain(&sb, witness_output_drain,
1464 "lock order %s -> %s established at:\n",
1465 w->w_name, w1->w_name);
1466 stack_sbuf_print_flags(&sb, &pstack,
1467 M_NOWAIT, STACK_SBUF_FMT_LONG);
1471 "lock order %s -> %s attempted at:\n",
1472 w1->w_name, w->w_name);
1473 stack_save(&pstack);
1474 stack_sbuf_print_flags(&sb, &pstack, M_NOWAIT,
1475 STACK_SBUF_FMT_LONG);
1480 witness_enter_debugger(__func__);
1486 * If requested, build a new lock order. However, don't build a new
1487 * relationship between a sleepable lock and Giant if it is in the
1488 * wrong direction. The correct lock order is that sleepable locks
1489 * always come before Giant.
1491 if (flags & LOP_NEWORDER &&
1492 !(plock->li_lock == &Giant.lock_object &&
1493 (lock->lo_flags & LO_SLEEPABLE) != 0 &&
1494 (flags & LOP_NOSLEEP) == 0)) {
1495 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1496 w->w_name, plock->li_lock->lo_witness->w_name);
1497 itismychild(plock->li_lock->lo_witness, w);
1500 mtx_unlock_spin(&w_mtx);
1504 witness_lock(struct lock_object *lock, int flags, const char *file, int line)
1506 struct lock_list_entry **lock_list, *lle;
1507 struct lock_instance *instance;
1511 if (witness_cold || witness_watch == -1 || lock->lo_witness == NULL ||
1514 w = lock->lo_witness;
1517 /* Determine lock list for this lock. */
1518 if (LOCK_CLASS(lock)->lc_flags & LC_SLEEPLOCK)
1519 lock_list = &td->td_sleeplocks;
1521 lock_list = PCPU_PTR(spinlocks);
1523 /* Check to see if we are recursing on a lock we already own. */
1524 instance = find_instance(*lock_list, lock);
1525 if (instance != NULL) {
1526 instance->li_flags++;
1527 CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
1528 td->td_proc->p_pid, lock->lo_name,
1529 instance->li_flags & LI_RECURSEMASK);
1530 instance->li_file = file;
1531 instance->li_line = line;
1535 /* Update per-witness last file and line acquire. */
1539 /* Find the next open lock instance in the list and fill it. */
1541 if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
1542 lle = witness_lock_list_get();
1545 lle->ll_next = *lock_list;
1546 CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
1547 td->td_proc->p_pid, lle);
1550 instance = &lle->ll_children[lle->ll_count++];
1551 instance->li_lock = lock;
1552 instance->li_line = line;
1553 instance->li_file = file;
1554 instance->li_flags = 0;
1555 if ((flags & LOP_EXCLUSIVE) != 0)
1556 instance->li_flags |= LI_EXCLUSIVE;
1557 if ((lock->lo_flags & LO_SLEEPABLE) != 0 && (flags & LOP_NOSLEEP) == 0)
1558 instance->li_flags |= LI_SLEEPABLE;
1559 CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
1560 td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
1564 witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
1566 struct lock_instance *instance;
1567 struct lock_class *class;
1569 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1570 if (lock->lo_witness == NULL || witness_watch == -1 || KERNEL_PANICKED())
1572 class = LOCK_CLASS(lock);
1573 if (witness_watch) {
1574 if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1576 "upgrade of non-upgradable lock (%s) %s @ %s:%d",
1577 class->lc_name, lock->lo_name,
1578 fixup_filename(file), line);
1579 if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1581 "upgrade of non-sleep lock (%s) %s @ %s:%d",
1582 class->lc_name, lock->lo_name,
1583 fixup_filename(file), line);
1585 instance = find_instance(curthread->td_sleeplocks, lock);
1586 if (instance == NULL) {
1587 kassert_panic("upgrade of unlocked lock (%s) %s @ %s:%d",
1588 class->lc_name, lock->lo_name,
1589 fixup_filename(file), line);
1592 if (witness_watch) {
1593 if ((instance->li_flags & LI_EXCLUSIVE) != 0)
1595 "upgrade of exclusive lock (%s) %s @ %s:%d",
1596 class->lc_name, lock->lo_name,
1597 fixup_filename(file), line);
1598 if ((instance->li_flags & LI_RECURSEMASK) != 0)
1600 "upgrade of recursed lock (%s) %s r=%d @ %s:%d",
1601 class->lc_name, lock->lo_name,
1602 instance->li_flags & LI_RECURSEMASK,
1603 fixup_filename(file), line);
1605 instance->li_flags |= LI_EXCLUSIVE;
1609 witness_downgrade(struct lock_object *lock, int flags, const char *file,
1612 struct lock_instance *instance;
1613 struct lock_class *class;
1615 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1616 if (lock->lo_witness == NULL || witness_watch == -1 || KERNEL_PANICKED())
1618 class = LOCK_CLASS(lock);
1619 if (witness_watch) {
1620 if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1622 "downgrade of non-upgradable lock (%s) %s @ %s:%d",
1623 class->lc_name, lock->lo_name,
1624 fixup_filename(file), line);
1625 if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1627 "downgrade of non-sleep lock (%s) %s @ %s:%d",
1628 class->lc_name, lock->lo_name,
1629 fixup_filename(file), line);
1631 instance = find_instance(curthread->td_sleeplocks, lock);
1632 if (instance == NULL) {
1633 kassert_panic("downgrade of unlocked lock (%s) %s @ %s:%d",
1634 class->lc_name, lock->lo_name,
1635 fixup_filename(file), line);
1638 if (witness_watch) {
1639 if ((instance->li_flags & LI_EXCLUSIVE) == 0)
1641 "downgrade of shared lock (%s) %s @ %s:%d",
1642 class->lc_name, lock->lo_name,
1643 fixup_filename(file), line);
1644 if ((instance->li_flags & LI_RECURSEMASK) != 0)
1646 "downgrade of recursed lock (%s) %s r=%d @ %s:%d",
1647 class->lc_name, lock->lo_name,
1648 instance->li_flags & LI_RECURSEMASK,
1649 fixup_filename(file), line);
1651 instance->li_flags &= ~LI_EXCLUSIVE;
1655 witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
1657 struct lock_list_entry **lock_list, *lle;
1658 struct lock_instance *instance;
1659 struct lock_class *class;
1664 if (witness_cold || lock->lo_witness == NULL || KERNEL_PANICKED())
1667 class = LOCK_CLASS(lock);
1669 /* Find lock instance associated with this lock. */
1670 if (class->lc_flags & LC_SLEEPLOCK)
1671 lock_list = &td->td_sleeplocks;
1673 lock_list = PCPU_PTR(spinlocks);
1675 for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
1676 for (i = 0; i < (*lock_list)->ll_count; i++) {
1677 instance = &(*lock_list)->ll_children[i];
1678 if (instance->li_lock == lock)
1683 * When disabling WITNESS through witness_watch we could end up in
1684 * having registered locks in the td_sleeplocks queue.
1685 * We have to make sure we flush these queues, so just search for
1686 * eventual register locks and remove them.
1688 if (witness_watch > 0) {
1689 kassert_panic("lock (%s) %s not locked @ %s:%d", class->lc_name,
1690 lock->lo_name, fixup_filename(file), line);
1697 /* First, check for shared/exclusive mismatches. */
1698 if ((instance->li_flags & LI_EXCLUSIVE) != 0 && witness_watch > 0 &&
1699 (flags & LOP_EXCLUSIVE) == 0) {
1700 witness_output("shared unlock of (%s) %s @ %s:%d\n",
1701 class->lc_name, lock->lo_name, fixup_filename(file), line);
1702 witness_output("while exclusively locked from %s:%d\n",
1703 fixup_filename(instance->li_file), instance->li_line);
1704 kassert_panic("excl->ushare");
1706 if ((instance->li_flags & LI_EXCLUSIVE) == 0 && witness_watch > 0 &&
1707 (flags & LOP_EXCLUSIVE) != 0) {
1708 witness_output("exclusive unlock of (%s) %s @ %s:%d\n",
1709 class->lc_name, lock->lo_name, fixup_filename(file), line);
1710 witness_output("while share locked from %s:%d\n",
1711 fixup_filename(instance->li_file),
1713 kassert_panic("share->uexcl");
1715 /* If we are recursed, unrecurse. */
1716 if ((instance->li_flags & LI_RECURSEMASK) > 0) {
1717 CTR4(KTR_WITNESS, "%s: pid %d unrecursed on %s r=%d", __func__,
1718 td->td_proc->p_pid, instance->li_lock->lo_name,
1719 instance->li_flags);
1720 instance->li_flags--;
1723 /* The lock is now being dropped, check for NORELEASE flag */
1724 if ((instance->li_flags & LI_NORELEASE) != 0 && witness_watch > 0) {
1725 witness_output("forbidden unlock of (%s) %s @ %s:%d\n",
1726 class->lc_name, lock->lo_name, fixup_filename(file), line);
1727 kassert_panic("lock marked norelease");
1730 /* Otherwise, remove this item from the list. */
1732 CTR4(KTR_WITNESS, "%s: pid %d removed %s from lle[%d]", __func__,
1733 td->td_proc->p_pid, instance->li_lock->lo_name,
1734 (*lock_list)->ll_count - 1);
1735 for (j = i; j < (*lock_list)->ll_count - 1; j++)
1736 (*lock_list)->ll_children[j] =
1737 (*lock_list)->ll_children[j + 1];
1738 (*lock_list)->ll_count--;
1742 * In order to reduce contention on w_mtx, we want to keep always an
1743 * head object into lists so that frequent allocation from the
1744 * free witness pool (and subsequent locking) is avoided.
1745 * In order to maintain the current code simple, when the head
1746 * object is totally unloaded it means also that we do not have
1747 * further objects in the list, so the list ownership needs to be
1748 * hand over to another object if the current head needs to be freed.
1750 if ((*lock_list)->ll_count == 0) {
1751 if (*lock_list == lle) {
1752 if (lle->ll_next == NULL)
1756 *lock_list = lle->ll_next;
1757 CTR3(KTR_WITNESS, "%s: pid %d removed lle %p", __func__,
1758 td->td_proc->p_pid, lle);
1759 witness_lock_list_free(lle);
1764 witness_thread_exit(struct thread *td)
1766 struct lock_list_entry *lle;
1769 lle = td->td_sleeplocks;
1770 if (lle == NULL || KERNEL_PANICKED())
1772 if (lle->ll_count != 0) {
1773 for (n = 0; lle != NULL; lle = lle->ll_next)
1774 for (i = lle->ll_count - 1; i >= 0; i--) {
1777 "Thread %p exiting with the following locks held:\n", td);
1779 witness_list_lock(&lle->ll_children[i],
1784 "Thread %p cannot exit while holding sleeplocks\n", td);
1786 witness_lock_list_free(lle);
1790 * Warn if any locks other than 'lock' are held. Flags can be passed in to
1791 * exempt Giant and sleepable locks from the checks as well. If any
1792 * non-exempt locks are held, then a supplied message is printed to the
1793 * output channel along with a list of the offending locks. If indicated in the
1794 * flags then a failure results in a panic as well.
1797 witness_warn(int flags, struct lock_object *lock, const char *fmt, ...)
1799 struct lock_list_entry *lock_list, *lle;
1800 struct lock_instance *lock1;
1805 if (witness_cold || witness_watch < 1 || KERNEL_PANICKED())
1809 for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next)
1810 for (i = lle->ll_count - 1; i >= 0; i--) {
1811 lock1 = &lle->ll_children[i];
1812 if (lock1->li_lock == lock)
1814 if (flags & WARN_GIANTOK &&
1815 lock1->li_lock == &Giant.lock_object)
1817 if (flags & WARN_SLEEPOK &&
1818 (lock1->li_flags & LI_SLEEPABLE) != 0)
1824 printf(" with the following %slocks held:\n",
1825 (flags & WARN_SLEEPOK) != 0 ?
1826 "non-sleepable " : "");
1829 witness_list_lock(lock1, printf);
1833 * Pin the thread in order to avoid problems with thread migration.
1834 * Once that all verifies are passed about spinlocks ownership,
1835 * the thread is in a safe path and it can be unpinned.
1838 lock_list = PCPU_GET(spinlocks);
1839 if (lock_list != NULL && lock_list->ll_count != 0) {
1843 * We should only have one spinlock and as long as
1844 * the flags cannot match for this locks class,
1845 * check if the first spinlock is the one curthread
1848 lock1 = &lock_list->ll_children[lock_list->ll_count - 1];
1849 if (lock_list->ll_count == 1 && lock_list->ll_next == NULL &&
1850 lock1->li_lock == lock && n == 0)
1856 printf(" with the following %slocks held:\n",
1857 (flags & WARN_SLEEPOK) != 0 ? "non-sleepable " : "");
1858 n += witness_list_locks(&lock_list, printf);
1861 if (flags & WARN_PANIC && n)
1862 kassert_panic("%s", __func__);
1864 witness_debugger(n, __func__);
1869 witness_file(struct lock_object *lock)
1873 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1875 w = lock->lo_witness;
1880 witness_line(struct lock_object *lock)
1884 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1886 w = lock->lo_witness;
1890 static struct witness *
1891 enroll(const char *description, struct lock_class *lock_class)
1895 MPASS(description != NULL);
1897 if (witness_watch == -1 || KERNEL_PANICKED())
1899 if ((lock_class->lc_flags & LC_SPINLOCK)) {
1900 if (witness_skipspin)
1902 } else if ((lock_class->lc_flags & LC_SLEEPLOCK) == 0) {
1903 kassert_panic("lock class %s is not sleep or spin",
1904 lock_class->lc_name);
1908 mtx_lock_spin(&w_mtx);
1909 w = witness_hash_get(description);
1912 if ((w = witness_get()) == NULL)
1914 MPASS(strlen(description) < MAX_W_NAME);
1915 strcpy(w->w_name, description);
1916 w->w_class = lock_class;
1918 STAILQ_INSERT_HEAD(&w_all, w, w_list);
1919 if (lock_class->lc_flags & LC_SPINLOCK) {
1920 STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
1922 } else if (lock_class->lc_flags & LC_SLEEPLOCK) {
1923 STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
1927 /* Insert new witness into the hash */
1928 witness_hash_put(w);
1929 witness_increment_graph_generation();
1930 mtx_unlock_spin(&w_mtx);
1934 if (w->w_refcount == 1)
1935 w->w_class = lock_class;
1936 mtx_unlock_spin(&w_mtx);
1937 if (lock_class != w->w_class)
1939 "lock (%s) %s does not match earlier (%s) lock",
1940 description, lock_class->lc_name,
1941 w->w_class->lc_name);
1946 depart(struct witness *w)
1949 MPASS(w->w_refcount == 0);
1950 if (w->w_class->lc_flags & LC_SLEEPLOCK) {
1956 * Set file to NULL as it may point into a loadable module.
1960 witness_increment_graph_generation();
1964 adopt(struct witness *parent, struct witness *child)
1968 if (witness_cold == 0)
1969 mtx_assert(&w_mtx, MA_OWNED);
1971 /* If the relationship is already known, there's no work to be done. */
1972 if (isitmychild(parent, child))
1975 /* When the structure of the graph changes, bump up the generation. */
1976 witness_increment_graph_generation();
1979 * The hard part ... create the direct relationship, then propagate all
1980 * indirect relationships.
1982 pi = parent->w_index;
1983 ci = child->w_index;
1984 WITNESS_INDEX_ASSERT(pi);
1985 WITNESS_INDEX_ASSERT(ci);
1987 w_rmatrix[pi][ci] |= WITNESS_PARENT;
1988 w_rmatrix[ci][pi] |= WITNESS_CHILD;
1991 * If parent was not already an ancestor of child,
1992 * then we increment the descendant and ancestor counters.
1994 if ((w_rmatrix[pi][ci] & WITNESS_ANCESTOR) == 0) {
1995 parent->w_num_descendants++;
1996 child->w_num_ancestors++;
2000 * Find each ancestor of 'pi'. Note that 'pi' itself is counted as
2001 * an ancestor of 'pi' during this loop.
2003 for (i = 1; i <= w_max_used_index; i++) {
2004 if ((w_rmatrix[i][pi] & WITNESS_ANCESTOR_MASK) == 0 &&
2008 /* Find each descendant of 'i' and mark it as a descendant. */
2009 for (j = 1; j <= w_max_used_index; j++) {
2011 * Skip children that are already marked as
2012 * descendants of 'i'.
2014 if (w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK)
2018 * We are only interested in descendants of 'ci'. Note
2019 * that 'ci' itself is counted as a descendant of 'ci'.
2021 if ((w_rmatrix[ci][j] & WITNESS_ANCESTOR_MASK) == 0 &&
2024 w_rmatrix[i][j] |= WITNESS_ANCESTOR;
2025 w_rmatrix[j][i] |= WITNESS_DESCENDANT;
2026 w_data[i].w_num_descendants++;
2027 w_data[j].w_num_ancestors++;
2030 * Make sure we aren't marking a node as both an
2031 * ancestor and descendant. We should have caught
2032 * this as a lock order reversal earlier.
2034 if ((w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK) &&
2035 (w_rmatrix[i][j] & WITNESS_DESCENDANT_MASK)) {
2036 printf("witness rmatrix paradox! [%d][%d]=%d "
2037 "both ancestor and descendant\n",
2038 i, j, w_rmatrix[i][j]);
2040 printf("Witness disabled.\n");
2043 if ((w_rmatrix[j][i] & WITNESS_ANCESTOR_MASK) &&
2044 (w_rmatrix[j][i] & WITNESS_DESCENDANT_MASK)) {
2045 printf("witness rmatrix paradox! [%d][%d]=%d "
2046 "both ancestor and descendant\n",
2047 j, i, w_rmatrix[j][i]);
2049 printf("Witness disabled.\n");
2057 itismychild(struct witness *parent, struct witness *child)
2061 MPASS(child != NULL && parent != NULL);
2062 if (witness_cold == 0)
2063 mtx_assert(&w_mtx, MA_OWNED);
2065 if (!witness_lock_type_equal(parent, child)) {
2066 if (witness_cold == 0) {
2068 mtx_unlock_spin(&w_mtx);
2073 "%s: parent \"%s\" (%s) and child \"%s\" (%s) are not "
2074 "the same lock type", __func__, parent->w_name,
2075 parent->w_class->lc_name, child->w_name,
2076 child->w_class->lc_name);
2078 mtx_lock_spin(&w_mtx);
2080 adopt(parent, child);
2084 * Generic code for the isitmy*() functions. The rmask parameter is the
2085 * expected relationship of w1 to w2.
2088 _isitmyx(struct witness *w1, struct witness *w2, int rmask, const char *fname)
2090 unsigned char r1, r2;
2095 WITNESS_INDEX_ASSERT(i1);
2096 WITNESS_INDEX_ASSERT(i2);
2097 r1 = w_rmatrix[i1][i2] & WITNESS_RELATED_MASK;
2098 r2 = w_rmatrix[i2][i1] & WITNESS_RELATED_MASK;
2100 /* The flags on one better be the inverse of the flags on the other */
2101 if (!((WITNESS_ATOD(r1) == r2 && WITNESS_DTOA(r2) == r1) ||
2102 (WITNESS_DTOA(r1) == r2 && WITNESS_ATOD(r2) == r1))) {
2103 /* Don't squawk if we're potentially racing with an update. */
2104 if (!mtx_owned(&w_mtx))
2106 printf("%s: rmatrix mismatch between %s (index %d) and %s "
2107 "(index %d): w_rmatrix[%d][%d] == %hhx but "
2108 "w_rmatrix[%d][%d] == %hhx\n",
2109 fname, w1->w_name, i1, w2->w_name, i2, i1, i2, r1,
2112 printf("Witness disabled.\n");
2115 return (r1 & rmask);
2119 * Checks if @child is a direct child of @parent.
2122 isitmychild(struct witness *parent, struct witness *child)
2125 return (_isitmyx(parent, child, WITNESS_PARENT, __func__));
2129 * Checks if @descendant is a direct or inderect descendant of @ancestor.
2132 isitmydescendant(struct witness *ancestor, struct witness *descendant)
2135 return (_isitmyx(ancestor, descendant, WITNESS_ANCESTOR_MASK,
2140 blessed(struct witness *w1, struct witness *w2)
2143 struct witness_blessed *b;
2145 for (i = 0; i < nitems(blessed_list); i++) {
2146 b = &blessed_list[i];
2147 if (strcmp(w1->w_name, b->b_lock1) == 0) {
2148 if (strcmp(w2->w_name, b->b_lock2) == 0)
2152 if (strcmp(w1->w_name, b->b_lock2) == 0)
2153 if (strcmp(w2->w_name, b->b_lock1) == 0)
2159 static struct witness *
2165 if (witness_cold == 0)
2166 mtx_assert(&w_mtx, MA_OWNED);
2168 if (witness_watch == -1) {
2169 mtx_unlock_spin(&w_mtx);
2172 if (STAILQ_EMPTY(&w_free)) {
2174 mtx_unlock_spin(&w_mtx);
2175 printf("WITNESS: unable to allocate a new witness object\n");
2178 w = STAILQ_FIRST(&w_free);
2179 STAILQ_REMOVE_HEAD(&w_free, w_list);
2182 MPASS(index > 0 && index == w_max_used_index+1 &&
2183 index < witness_count);
2184 bzero(w, sizeof(*w));
2186 if (index > w_max_used_index)
2187 w_max_used_index = index;
2192 witness_free(struct witness *w)
2195 STAILQ_INSERT_HEAD(&w_free, w, w_list);
2199 static struct lock_list_entry *
2200 witness_lock_list_get(void)
2202 struct lock_list_entry *lle;
2204 if (witness_watch == -1)
2206 mtx_lock_spin(&w_mtx);
2207 lle = w_lock_list_free;
2210 mtx_unlock_spin(&w_mtx);
2211 printf("%s: witness exhausted\n", __func__);
2214 w_lock_list_free = lle->ll_next;
2215 mtx_unlock_spin(&w_mtx);
2216 bzero(lle, sizeof(*lle));
2221 witness_lock_list_free(struct lock_list_entry *lle)
2224 mtx_lock_spin(&w_mtx);
2225 lle->ll_next = w_lock_list_free;
2226 w_lock_list_free = lle;
2227 mtx_unlock_spin(&w_mtx);
2230 static struct lock_instance *
2231 find_instance(struct lock_list_entry *list, const struct lock_object *lock)
2233 struct lock_list_entry *lle;
2234 struct lock_instance *instance;
2237 for (lle = list; lle != NULL; lle = lle->ll_next)
2238 for (i = lle->ll_count - 1; i >= 0; i--) {
2239 instance = &lle->ll_children[i];
2240 if (instance->li_lock == lock)
2247 witness_list_lock(struct lock_instance *instance,
2248 int (*prnt)(const char *fmt, ...))
2250 struct lock_object *lock;
2252 lock = instance->li_lock;
2253 prnt("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ?
2254 "exclusive" : "shared", LOCK_CLASS(lock)->lc_name, lock->lo_name);
2255 if (lock->lo_witness->w_name != lock->lo_name)
2256 prnt(" (%s)", lock->lo_witness->w_name);
2257 prnt(" r = %d (%p) locked @ %s:%d\n",
2258 instance->li_flags & LI_RECURSEMASK, lock,
2259 fixup_filename(instance->li_file), instance->li_line);
2263 witness_output(const char *fmt, ...)
2269 ret = witness_voutput(fmt, ap);
2275 witness_voutput(const char *fmt, va_list ap)
2280 switch (witness_channel) {
2281 case WITNESS_CONSOLE:
2282 ret = vprintf(fmt, ap);
2285 vlog(LOG_NOTICE, fmt, ap);
2295 witness_thread_has_locks(struct thread *td)
2298 if (td->td_sleeplocks == NULL)
2300 return (td->td_sleeplocks->ll_count != 0);
2304 witness_proc_has_locks(struct proc *p)
2308 FOREACH_THREAD_IN_PROC(p, td) {
2309 if (witness_thread_has_locks(td))
2317 witness_list_locks(struct lock_list_entry **lock_list,
2318 int (*prnt)(const char *fmt, ...))
2320 struct lock_list_entry *lle;
2324 for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
2325 for (i = lle->ll_count - 1; i >= 0; i--) {
2326 witness_list_lock(&lle->ll_children[i], prnt);
2333 * This is a bit risky at best. We call this function when we have timed
2334 * out acquiring a spin lock, and we assume that the other CPU is stuck
2335 * with this lock held. So, we go groveling around in the other CPU's
2336 * per-cpu data to try to find the lock instance for this spin lock to
2337 * see when it was last acquired.
2340 witness_display_spinlock(struct lock_object *lock, struct thread *owner,
2341 int (*prnt)(const char *fmt, ...))
2343 struct lock_instance *instance;
2346 if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU)
2348 pc = pcpu_find(owner->td_oncpu);
2349 instance = find_instance(pc->pc_spinlocks, lock);
2350 if (instance != NULL)
2351 witness_list_lock(instance, prnt);
2355 witness_save(struct lock_object *lock, const char **filep, int *linep)
2357 struct lock_list_entry *lock_list;
2358 struct lock_instance *instance;
2359 struct lock_class *class;
2362 * This function is used independently in locking code to deal with
2363 * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2366 if (SCHEDULER_STOPPED())
2368 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2369 if (lock->lo_witness == NULL || witness_watch == -1 || KERNEL_PANICKED())
2371 class = LOCK_CLASS(lock);
2372 if (class->lc_flags & LC_SLEEPLOCK)
2373 lock_list = curthread->td_sleeplocks;
2375 if (witness_skipspin)
2377 lock_list = PCPU_GET(spinlocks);
2379 instance = find_instance(lock_list, lock);
2380 if (instance == NULL) {
2381 kassert_panic("%s: lock (%s) %s not locked", __func__,
2382 class->lc_name, lock->lo_name);
2385 *filep = instance->li_file;
2386 *linep = instance->li_line;
2390 witness_restore(struct lock_object *lock, const char *file, int line)
2392 struct lock_list_entry *lock_list;
2393 struct lock_instance *instance;
2394 struct lock_class *class;
2397 * This function is used independently in locking code to deal with
2398 * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2401 if (SCHEDULER_STOPPED())
2403 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2404 if (lock->lo_witness == NULL || witness_watch == -1 || KERNEL_PANICKED())
2406 class = LOCK_CLASS(lock);
2407 if (class->lc_flags & LC_SLEEPLOCK)
2408 lock_list = curthread->td_sleeplocks;
2410 if (witness_skipspin)
2412 lock_list = PCPU_GET(spinlocks);
2414 instance = find_instance(lock_list, lock);
2415 if (instance == NULL)
2416 kassert_panic("%s: lock (%s) %s not locked", __func__,
2417 class->lc_name, lock->lo_name);
2418 lock->lo_witness->w_file = file;
2419 lock->lo_witness->w_line = line;
2420 if (instance == NULL)
2422 instance->li_file = file;
2423 instance->li_line = line;
2427 witness_find_instance(const struct lock_object *lock,
2428 struct lock_instance **instance)
2430 #ifdef INVARIANT_SUPPORT
2431 struct lock_class *class;
2433 if (lock->lo_witness == NULL || witness_watch < 1 || KERNEL_PANICKED())
2435 class = LOCK_CLASS(lock);
2436 if ((class->lc_flags & LC_SLEEPLOCK) != 0) {
2437 *instance = find_instance(curthread->td_sleeplocks, lock);
2439 } else if ((class->lc_flags & LC_SPINLOCK) != 0) {
2440 *instance = find_instance(PCPU_GET(spinlocks), lock);
2443 kassert_panic("Lock (%s) %s is not sleep or spin!",
2444 class->lc_name, lock->lo_name);
2453 witness_assert(const struct lock_object *lock, int flags, const char *file,
2456 #ifdef INVARIANT_SUPPORT
2457 struct lock_instance *instance;
2458 struct lock_class *class;
2460 if (!witness_find_instance(lock, &instance))
2462 class = LOCK_CLASS(lock);
2465 if (instance != NULL)
2466 kassert_panic("Lock (%s) %s locked @ %s:%d.",
2467 class->lc_name, lock->lo_name,
2468 fixup_filename(file), line);
2471 case LA_LOCKED | LA_RECURSED:
2472 case LA_LOCKED | LA_NOTRECURSED:
2474 case LA_SLOCKED | LA_RECURSED:
2475 case LA_SLOCKED | LA_NOTRECURSED:
2477 case LA_XLOCKED | LA_RECURSED:
2478 case LA_XLOCKED | LA_NOTRECURSED:
2479 if (instance == NULL) {
2480 kassert_panic("Lock (%s) %s not locked @ %s:%d.",
2481 class->lc_name, lock->lo_name,
2482 fixup_filename(file), line);
2485 if ((flags & LA_XLOCKED) != 0 &&
2486 (instance->li_flags & LI_EXCLUSIVE) == 0)
2488 "Lock (%s) %s not exclusively locked @ %s:%d.",
2489 class->lc_name, lock->lo_name,
2490 fixup_filename(file), line);
2491 if ((flags & LA_SLOCKED) != 0 &&
2492 (instance->li_flags & LI_EXCLUSIVE) != 0)
2494 "Lock (%s) %s exclusively locked @ %s:%d.",
2495 class->lc_name, lock->lo_name,
2496 fixup_filename(file), line);
2497 if ((flags & LA_RECURSED) != 0 &&
2498 (instance->li_flags & LI_RECURSEMASK) == 0)
2499 kassert_panic("Lock (%s) %s not recursed @ %s:%d.",
2500 class->lc_name, lock->lo_name,
2501 fixup_filename(file), line);
2502 if ((flags & LA_NOTRECURSED) != 0 &&
2503 (instance->li_flags & LI_RECURSEMASK) != 0)
2504 kassert_panic("Lock (%s) %s recursed @ %s:%d.",
2505 class->lc_name, lock->lo_name,
2506 fixup_filename(file), line);
2509 kassert_panic("Invalid lock assertion at %s:%d.",
2510 fixup_filename(file), line);
2512 #endif /* INVARIANT_SUPPORT */
2516 * Checks the ownership of the lock by curthread, consulting the witness list.
2518 * 0 if witness is disabled or did not work
2523 witness_is_owned(const struct lock_object *lock)
2525 #ifdef INVARIANT_SUPPORT
2526 struct lock_instance *instance;
2528 if (!witness_find_instance(lock, &instance))
2530 return (instance == NULL ? -1 : 1);
2537 witness_setflag(struct lock_object *lock, int flag, int set)
2539 struct lock_list_entry *lock_list;
2540 struct lock_instance *instance;
2541 struct lock_class *class;
2543 if (lock->lo_witness == NULL || witness_watch == -1 || KERNEL_PANICKED())
2545 class = LOCK_CLASS(lock);
2546 if (class->lc_flags & LC_SLEEPLOCK)
2547 lock_list = curthread->td_sleeplocks;
2549 if (witness_skipspin)
2551 lock_list = PCPU_GET(spinlocks);
2553 instance = find_instance(lock_list, lock);
2554 if (instance == NULL) {
2555 kassert_panic("%s: lock (%s) %s not locked", __func__,
2556 class->lc_name, lock->lo_name);
2561 instance->li_flags |= flag;
2563 instance->li_flags &= ~flag;
2567 witness_norelease(struct lock_object *lock)
2570 witness_setflag(lock, LI_NORELEASE, 1);
2574 witness_releaseok(struct lock_object *lock)
2577 witness_setflag(lock, LI_NORELEASE, 0);
2582 witness_ddb_list(struct thread *td)
2585 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2586 KASSERT(kdb_active, ("%s: not in the debugger", __func__));
2588 if (witness_watch < 1)
2591 witness_list_locks(&td->td_sleeplocks, db_printf);
2594 * We only handle spinlocks if td == curthread. This is somewhat broken
2595 * if td is currently executing on some other CPU and holds spin locks
2596 * as we won't display those locks. If we had a MI way of getting
2597 * the per-cpu data for a given cpu then we could use
2598 * td->td_oncpu to get the list of spinlocks for this thread
2601 * That still wouldn't really fix this unless we locked the scheduler
2602 * lock or stopped the other CPU to make sure it wasn't changing the
2603 * list out from under us. It is probably best to just not try to
2604 * handle threads on other CPU's for now.
2606 if (td == curthread && PCPU_GET(spinlocks) != NULL)
2607 witness_list_locks(PCPU_PTR(spinlocks), db_printf);
2610 DB_SHOW_COMMAND(locks, db_witness_list)
2615 td = db_lookup_thread(addr, true);
2618 witness_ddb_list(td);
2621 DB_SHOW_ALL_COMMAND(locks, db_witness_list_all)
2627 * It would be nice to list only threads and processes that actually
2628 * held sleep locks, but that information is currently not exported
2631 FOREACH_PROC_IN_SYSTEM(p) {
2632 if (!witness_proc_has_locks(p))
2634 FOREACH_THREAD_IN_PROC(p, td) {
2635 if (!witness_thread_has_locks(td))
2637 db_printf("Process %d (%s) thread %p (%d)\n", p->p_pid,
2638 p->p_comm, td, td->td_tid);
2639 witness_ddb_list(td);
2645 DB_SHOW_ALIAS(alllocks, db_witness_list_all)
2647 DB_SHOW_COMMAND(witness, db_witness_display)
2650 witness_ddb_display(db_printf);
2655 sbuf_print_witness_badstacks(struct sbuf *sb, size_t *oldidx)
2657 struct witness_lock_order_data *data1, *data2, *tmp_data1, *tmp_data2;
2658 struct witness *tmp_w1, *tmp_w2, *w1, *w2;
2659 int generation, i, j;
2666 /* Allocate and init temporary storage space. */
2667 tmp_w1 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2668 tmp_w2 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2669 tmp_data1 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2671 tmp_data2 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2673 stack_zero(&tmp_data1->wlod_stack);
2674 stack_zero(&tmp_data2->wlod_stack);
2677 mtx_lock_spin(&w_mtx);
2678 generation = w_generation;
2679 mtx_unlock_spin(&w_mtx);
2680 sbuf_printf(sb, "Number of known direct relationships is %d\n",
2681 w_lohash.wloh_count);
2682 for (i = 1; i < w_max_used_index; i++) {
2683 mtx_lock_spin(&w_mtx);
2684 if (generation != w_generation) {
2685 mtx_unlock_spin(&w_mtx);
2687 /* The graph has changed, try again. */
2694 if (w1->w_reversed == 0) {
2695 mtx_unlock_spin(&w_mtx);
2699 /* Copy w1 locally so we can release the spin lock. */
2701 mtx_unlock_spin(&w_mtx);
2703 if (tmp_w1->w_reversed == 0)
2705 for (j = 1; j < w_max_used_index; j++) {
2706 if ((w_rmatrix[i][j] & WITNESS_REVERSAL) == 0 || i > j)
2709 mtx_lock_spin(&w_mtx);
2710 if (generation != w_generation) {
2711 mtx_unlock_spin(&w_mtx);
2713 /* The graph has changed, try again. */
2720 data1 = witness_lock_order_get(w1, w2);
2721 data2 = witness_lock_order_get(w2, w1);
2724 * Copy information locally so we can release the
2730 stack_zero(&tmp_data1->wlod_stack);
2731 stack_copy(&data1->wlod_stack,
2732 &tmp_data1->wlod_stack);
2734 if (data2 && data2 != data1) {
2735 stack_zero(&tmp_data2->wlod_stack);
2736 stack_copy(&data2->wlod_stack,
2737 &tmp_data2->wlod_stack);
2739 mtx_unlock_spin(&w_mtx);
2741 if (blessed(tmp_w1, tmp_w2))
2745 "\nLock order reversal between \"%s\"(%s) and \"%s\"(%s)!\n",
2746 tmp_w1->w_name, tmp_w1->w_class->lc_name,
2747 tmp_w2->w_name, tmp_w2->w_class->lc_name);
2750 "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2751 tmp_w1->w_name, tmp_w1->w_class->lc_name,
2752 tmp_w2->w_name, tmp_w2->w_class->lc_name);
2753 stack_sbuf_print(sb, &tmp_data1->wlod_stack);
2754 sbuf_printf(sb, "\n");
2756 if (data2 && data2 != data1) {
2758 "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2759 tmp_w2->w_name, tmp_w2->w_class->lc_name,
2760 tmp_w1->w_name, tmp_w1->w_class->lc_name);
2761 stack_sbuf_print(sb, &tmp_data2->wlod_stack);
2762 sbuf_printf(sb, "\n");
2766 mtx_lock_spin(&w_mtx);
2767 if (generation != w_generation) {
2768 mtx_unlock_spin(&w_mtx);
2771 * The graph changed while we were printing stack data,
2778 mtx_unlock_spin(&w_mtx);
2780 /* Free temporary storage space. */
2781 free(tmp_data1, M_TEMP);
2782 free(tmp_data2, M_TEMP);
2783 free(tmp_w1, M_TEMP);
2784 free(tmp_w2, M_TEMP);
2788 sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS)
2793 if (witness_watch < 1) {
2794 error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2798 error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2802 sb = sbuf_new(NULL, NULL, badstack_sbuf_size, SBUF_AUTOEXTEND);
2806 sbuf_print_witness_badstacks(sb, &req->oldidx);
2809 error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1);
2817 sbuf_db_printf_drain(void *arg __unused, const char *data, int len)
2820 return (db_printf("%.*s", len, data));
2823 DB_SHOW_COMMAND(badstacks, db_witness_badstacks)
2829 sbuf_new(&sb, buffer, sizeof(buffer), SBUF_FIXEDLEN);
2830 sbuf_set_drain(&sb, sbuf_db_printf_drain, NULL);
2831 sbuf_print_witness_badstacks(&sb, &dummy);
2837 sysctl_debug_witness_channel(SYSCTL_HANDLER_ARGS)
2839 static const struct {
2840 enum witness_channel channel;
2843 { WITNESS_CONSOLE, "console" },
2844 { WITNESS_LOG, "log" },
2845 { WITNESS_NONE, "none" },
2852 for (i = 0; i < nitems(channels); i++)
2853 if (witness_channel == channels[i].channel) {
2854 snprintf(buf, sizeof(buf), "%s", channels[i].name);
2858 error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
2859 if (error != 0 || req->newptr == NULL)
2863 for (i = 0; i < nitems(channels); i++)
2864 if (strcmp(channels[i].name, buf) == 0) {
2865 witness_channel = channels[i].channel;
2873 sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS)
2880 error = SYSCTL_OUT(req, w_notallowed, sizeof(w_notallowed));
2884 if (witness_watch < 1) {
2885 error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2889 error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2894 error = sysctl_wire_old_buffer(req, 0);
2897 sb = sbuf_new_for_sysctl(NULL, NULL, FULLGRAPH_SBUF_SIZE, req);
2900 sbuf_printf(sb, "\n");
2902 mtx_lock_spin(&w_mtx);
2903 STAILQ_FOREACH(w, &w_all, w_list)
2905 STAILQ_FOREACH(w, &w_all, w_list)
2906 witness_add_fullgraph(sb, w);
2907 mtx_unlock_spin(&w_mtx);
2910 * Close the sbuf and return to userland.
2912 error = sbuf_finish(sb);
2919 sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS)
2923 value = witness_watch;
2924 error = sysctl_handle_int(oidp, &value, 0, req);
2925 if (error != 0 || req->newptr == NULL)
2927 if (value > 1 || value < -1 ||
2928 (witness_watch == -1 && value != witness_watch))
2930 witness_watch = value;
2935 witness_add_fullgraph(struct sbuf *sb, struct witness *w)
2939 if (w->w_displayed != 0 || (w->w_file == NULL && w->w_line == 0))
2943 WITNESS_INDEX_ASSERT(w->w_index);
2944 for (i = 1; i <= w_max_used_index; i++) {
2945 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT) {
2946 sbuf_printf(sb, "\"%s\",\"%s\"\n", w->w_name,
2948 witness_add_fullgraph(sb, &w_data[i]);
2954 * A simple hash function. Takes a key pointer and a key size. If size == 0,
2955 * interprets the key as a string and reads until the null
2956 * terminator. Otherwise, reads the first size bytes. Returns an unsigned 32-bit
2957 * hash value computed from the key.
2960 witness_hash_djb2(const uint8_t *key, uint32_t size)
2962 unsigned int hash = 5381;
2965 /* hash = hash * 33 + key[i] */
2967 for (i = 0; i < size; i++)
2968 hash = ((hash << 5) + hash) + (unsigned int)key[i];
2970 for (i = 0; key[i] != 0; i++)
2971 hash = ((hash << 5) + hash) + (unsigned int)key[i];
2977 * Initializes the two witness hash tables. Called exactly once from
2978 * witness_initialize().
2981 witness_init_hash_tables(void)
2985 MPASS(witness_cold);
2987 /* Initialize the hash tables. */
2988 for (i = 0; i < WITNESS_HASH_SIZE; i++)
2989 w_hash.wh_array[i] = NULL;
2991 w_hash.wh_size = WITNESS_HASH_SIZE;
2992 w_hash.wh_count = 0;
2994 /* Initialize the lock order data hash. */
2996 for (i = 0; i < WITNESS_LO_DATA_COUNT; i++) {
2997 memset(&w_lodata[i], 0, sizeof(w_lodata[i]));
2998 w_lodata[i].wlod_next = w_lofree;
2999 w_lofree = &w_lodata[i];
3001 w_lohash.wloh_size = WITNESS_LO_HASH_SIZE;
3002 w_lohash.wloh_count = 0;
3003 for (i = 0; i < WITNESS_LO_HASH_SIZE; i++)
3004 w_lohash.wloh_array[i] = NULL;
3007 static struct witness *
3008 witness_hash_get(const char *key)
3014 if (witness_cold == 0)
3015 mtx_assert(&w_mtx, MA_OWNED);
3016 hash = witness_hash_djb2(key, 0) % w_hash.wh_size;
3017 w = w_hash.wh_array[hash];
3019 if (strcmp(w->w_name, key) == 0)
3029 witness_hash_put(struct witness *w)
3034 MPASS(w->w_name != NULL);
3035 if (witness_cold == 0)
3036 mtx_assert(&w_mtx, MA_OWNED);
3037 KASSERT(witness_hash_get(w->w_name) == NULL,
3038 ("%s: trying to add a hash entry that already exists!", __func__));
3039 KASSERT(w->w_hash_next == NULL,
3040 ("%s: w->w_hash_next != NULL", __func__));
3042 hash = witness_hash_djb2(w->w_name, 0) % w_hash.wh_size;
3043 w->w_hash_next = w_hash.wh_array[hash];
3044 w_hash.wh_array[hash] = w;
3048 static struct witness_lock_order_data *
3049 witness_lock_order_get(struct witness *parent, struct witness *child)
3051 struct witness_lock_order_data *data = NULL;
3052 struct witness_lock_order_key key;
3055 MPASS(parent != NULL && child != NULL);
3056 key.from = parent->w_index;
3057 key.to = child->w_index;
3058 WITNESS_INDEX_ASSERT(key.from);
3059 WITNESS_INDEX_ASSERT(key.to);
3060 if ((w_rmatrix[parent->w_index][child->w_index]
3061 & WITNESS_LOCK_ORDER_KNOWN) == 0)
3064 hash = witness_hash_djb2((const char*)&key,
3065 sizeof(key)) % w_lohash.wloh_size;
3066 data = w_lohash.wloh_array[hash];
3067 while (data != NULL) {
3068 if (witness_lock_order_key_equal(&data->wlod_key, &key))
3070 data = data->wlod_next;
3078 * Verify that parent and child have a known relationship, are not the same,
3079 * and child is actually a child of parent. This is done without w_mtx
3080 * to avoid contention in the common case.
3083 witness_lock_order_check(struct witness *parent, struct witness *child)
3086 if (parent != child &&
3087 w_rmatrix[parent->w_index][child->w_index]
3088 & WITNESS_LOCK_ORDER_KNOWN &&
3089 isitmychild(parent, child))
3096 witness_lock_order_add(struct witness *parent, struct witness *child)
3098 struct witness_lock_order_data *data = NULL;
3099 struct witness_lock_order_key key;
3102 MPASS(parent != NULL && child != NULL);
3103 key.from = parent->w_index;
3104 key.to = child->w_index;
3105 WITNESS_INDEX_ASSERT(key.from);
3106 WITNESS_INDEX_ASSERT(key.to);
3107 if (w_rmatrix[parent->w_index][child->w_index]
3108 & WITNESS_LOCK_ORDER_KNOWN)
3111 hash = witness_hash_djb2((const char*)&key,
3112 sizeof(key)) % w_lohash.wloh_size;
3113 w_rmatrix[parent->w_index][child->w_index] |= WITNESS_LOCK_ORDER_KNOWN;
3117 w_lofree = data->wlod_next;
3118 data->wlod_next = w_lohash.wloh_array[hash];
3119 data->wlod_key = key;
3120 w_lohash.wloh_array[hash] = data;
3121 w_lohash.wloh_count++;
3122 stack_zero(&data->wlod_stack);
3123 stack_save(&data->wlod_stack);
3127 /* Call this whenever the structure of the witness graph changes. */
3129 witness_increment_graph_generation(void)
3132 if (witness_cold == 0)
3133 mtx_assert(&w_mtx, MA_OWNED);
3138 witness_output_drain(void *arg __unused, const char *data, int len)
3141 witness_output("%.*s", len, data);
3146 witness_debugger(int cond, const char *msg)
3155 if (witness_trace) {
3156 sbuf_new(&sb, buf, sizeof(buf), SBUF_FIXEDLEN);
3157 sbuf_set_drain(&sb, witness_output_drain, NULL);
3161 witness_output("stack backtrace:\n");
3162 stack_sbuf_print_ddb(&sb, &st);
3167 witness_enter_debugger(msg);
3171 witness_enter_debugger(const char *msg)
3175 kdb_enter(KDB_WHY_WITNESS, msg);