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>
90 __FBSDID("$FreeBSD$");
93 #include "opt_hwpmc_hooks.h"
94 #include "opt_stack.h"
95 #include "opt_witness.h"
97 #include <sys/param.h>
100 #include <sys/kernel.h>
102 #include <sys/lock.h>
103 #include <sys/malloc.h>
104 #include <sys/mutex.h>
105 #include <sys/priv.h>
106 #include <sys/proc.h>
107 #include <sys/sbuf.h>
108 #include <sys/sched.h>
109 #include <sys/stack.h>
110 #include <sys/sysctl.h>
111 #include <sys/syslog.h>
112 #include <sys/systm.h>
118 #include <machine/stdarg.h>
120 #if !defined(DDB) && !defined(STACK)
121 #error "DDB or STACK options are required for WITNESS"
124 /* Note that these traces do not work with KTR_ALQ. */
126 #define KTR_WITNESS KTR_SUBSYS
128 #define KTR_WITNESS 0
131 #define LI_RECURSEMASK 0x0000ffff /* Recursion depth of lock instance. */
132 #define LI_EXCLUSIVE 0x00010000 /* Exclusive lock instance. */
133 #define LI_NORELEASE 0x00020000 /* Lock not allowed to be released. */
135 /* Define this to check for blessed mutexes */
138 #ifndef WITNESS_COUNT
139 #define WITNESS_COUNT 1536
141 #define WITNESS_HASH_SIZE 251 /* Prime, gives load factor < 2 */
142 #define WITNESS_PENDLIST (2048 + MAXCPU)
144 /* Allocate 256 KB of stack data space */
145 #define WITNESS_LO_DATA_COUNT 2048
147 /* Prime, gives load factor of ~2 at full load */
148 #define WITNESS_LO_HASH_SIZE 1021
151 * XXX: This is somewhat bogus, as we assume here that at most 2048 threads
152 * will hold LOCK_NCHILDREN locks. We handle failure ok, and we should
153 * probably be safe for the most part, but it's still a SWAG.
155 #define LOCK_NCHILDREN 5
156 #define LOCK_CHILDCOUNT 2048
158 #define MAX_W_NAME 64
160 #define FULLGRAPH_SBUF_SIZE 512
163 * These flags go in the witness relationship matrix and describe the
164 * relationship between any two struct witness objects.
166 #define WITNESS_UNRELATED 0x00 /* No lock order relation. */
167 #define WITNESS_PARENT 0x01 /* Parent, aka direct ancestor. */
168 #define WITNESS_ANCESTOR 0x02 /* Direct or indirect ancestor. */
169 #define WITNESS_CHILD 0x04 /* Child, aka direct descendant. */
170 #define WITNESS_DESCENDANT 0x08 /* Direct or indirect descendant. */
171 #define WITNESS_ANCESTOR_MASK (WITNESS_PARENT | WITNESS_ANCESTOR)
172 #define WITNESS_DESCENDANT_MASK (WITNESS_CHILD | WITNESS_DESCENDANT)
173 #define WITNESS_RELATED_MASK \
174 (WITNESS_ANCESTOR_MASK | WITNESS_DESCENDANT_MASK)
175 #define WITNESS_REVERSAL 0x10 /* A lock order reversal has been
177 #define WITNESS_RESERVED1 0x20 /* Unused flag, reserved. */
178 #define WITNESS_RESERVED2 0x40 /* Unused flag, reserved. */
179 #define WITNESS_LOCK_ORDER_KNOWN 0x80 /* This lock order is known. */
181 /* Descendant to ancestor flags */
182 #define WITNESS_DTOA(x) (((x) & WITNESS_RELATED_MASK) >> 2)
184 /* Ancestor to descendant flags */
185 #define WITNESS_ATOD(x) (((x) & WITNESS_RELATED_MASK) << 2)
187 #define WITNESS_INDEX_ASSERT(i) \
188 MPASS((i) > 0 && (i) <= w_max_used_index && (i) < witness_count)
190 static MALLOC_DEFINE(M_WITNESS, "Witness", "Witness");
193 * Lock instances. A lock instance is the data associated with a lock while
194 * it is held by witness. For example, a lock instance will hold the
195 * recursion count of a lock. Lock instances are held in lists. Spin locks
196 * are held in a per-cpu list while sleep locks are held in per-thread list.
198 struct lock_instance {
199 struct lock_object *li_lock;
206 * A simple list type used to build the list of locks held by a thread
207 * or CPU. We can't simply embed the list in struct lock_object since a
208 * lock may be held by more than one thread if it is a shared lock. Locks
209 * are added to the head of the list, so we fill up each list entry from
210 * "the back" logically. To ease some of the arithmetic, we actually fill
211 * in each list entry the normal way (children[0] then children[1], etc.) but
212 * when we traverse the list we read children[count-1] as the first entry
213 * down to children[0] as the final entry.
215 struct lock_list_entry {
216 struct lock_list_entry *ll_next;
217 struct lock_instance ll_children[LOCK_NCHILDREN];
222 * The main witness structure. One of these per named lock type in the system
223 * (for example, "vnode interlock").
226 char w_name[MAX_W_NAME];
227 uint32_t w_index; /* Index in the relationship matrix */
228 struct lock_class *w_class;
229 STAILQ_ENTRY(witness) w_list; /* List of all witnesses. */
230 STAILQ_ENTRY(witness) w_typelist; /* Witnesses of a type. */
231 struct witness *w_hash_next; /* Linked list in hash buckets. */
232 const char *w_file; /* File where last acquired */
233 uint32_t w_line; /* Line where last acquired */
235 uint16_t w_num_ancestors; /* direct/indirect
237 uint16_t w_num_descendants; /* direct/indirect
238 * descendant count */
240 unsigned w_displayed:1;
241 unsigned w_reversed:1;
244 STAILQ_HEAD(witness_list, witness);
247 * The witness hash table. Keys are witness names (const char *), elements are
248 * witness objects (struct witness *).
250 struct witness_hash {
251 struct witness *wh_array[WITNESS_HASH_SIZE];
257 * Key type for the lock order data hash table.
259 struct witness_lock_order_key {
264 struct witness_lock_order_data {
265 struct stack wlod_stack;
266 struct witness_lock_order_key wlod_key;
267 struct witness_lock_order_data *wlod_next;
271 * The witness lock order data hash table. Keys are witness index tuples
272 * (struct witness_lock_order_key), elements are lock order data objects
273 * (struct witness_lock_order_data).
275 struct witness_lock_order_hash {
276 struct witness_lock_order_data *wloh_array[WITNESS_LO_HASH_SIZE];
282 struct witness_blessed {
288 struct witness_pendhelp {
290 struct lock_object *wh_lock;
293 struct witness_order_list_entry {
295 struct lock_class *w_class;
299 * Returns 0 if one of the locks is a spin lock and the other is not.
300 * Returns 1 otherwise.
303 witness_lock_type_equal(struct witness *w1, struct witness *w2)
306 return ((w1->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) ==
307 (w2->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)));
311 witness_lock_order_key_equal(const struct witness_lock_order_key *a,
312 const struct witness_lock_order_key *b)
315 return (a->from == b->from && a->to == b->to);
318 static int _isitmyx(struct witness *w1, struct witness *w2, int rmask,
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 int sysctl_debug_witness_channel(SYSCTL_HANDLER_ARGS);
336 static void witness_add_fullgraph(struct sbuf *sb, struct witness *parent);
338 static void witness_ddb_compute_levels(void);
339 static void witness_ddb_display(int(*)(const char *fmt, ...));
340 static void witness_ddb_display_descendants(int(*)(const char *fmt, ...),
341 struct witness *, int indent);
342 static void witness_ddb_display_list(int(*prnt)(const char *fmt, ...),
343 struct witness_list *list);
344 static void witness_ddb_level_descendants(struct witness *parent, int l);
345 static void witness_ddb_list(struct thread *td);
347 static void witness_debugger(int cond, const char *msg);
348 static void witness_free(struct witness *m);
349 static struct witness *witness_get(void);
350 static uint32_t witness_hash_djb2(const uint8_t *key, uint32_t size);
351 static struct witness *witness_hash_get(const char *key);
352 static void witness_hash_put(struct witness *w);
353 static void witness_init_hash_tables(void);
354 static void witness_increment_graph_generation(void);
355 static void witness_lock_list_free(struct lock_list_entry *lle);
356 static struct lock_list_entry *witness_lock_list_get(void);
357 static int witness_lock_order_add(struct witness *parent,
358 struct witness *child);
359 static int witness_lock_order_check(struct witness *parent,
360 struct witness *child);
361 static struct witness_lock_order_data *witness_lock_order_get(
362 struct witness *parent,
363 struct witness *child);
364 static void witness_list_lock(struct lock_instance *instance,
365 int (*prnt)(const char *fmt, ...));
366 static int witness_output(const char *fmt, ...) __printflike(1, 2);
367 static int witness_voutput(const char *fmt, va_list ap) __printflike(1, 0);
368 static void witness_setflag(struct lock_object *lock, int flag, int set);
370 static SYSCTL_NODE(_debug, OID_AUTO, witness, CTLFLAG_RW, NULL,
374 * If set to 0, lock order checking is disabled. If set to -1,
375 * witness is completely disabled. Otherwise witness performs full
376 * lock order checking for all locks. At runtime, lock order checking
377 * may be toggled. However, witness cannot be reenabled once it is
378 * completely disabled.
380 static int witness_watch = 1;
381 SYSCTL_PROC(_debug_witness, OID_AUTO, watch, CTLFLAG_RWTUN | CTLTYPE_INT, NULL, 0,
382 sysctl_debug_witness_watch, "I", "witness is watching lock operations");
386 * When KDB is enabled and witness_kdb is 1, it will cause the system
387 * to drop into kdebug() when:
388 * - a lock hierarchy violation occurs
389 * - locks are held when going to sleep.
396 SYSCTL_INT(_debug_witness, OID_AUTO, kdb, CTLFLAG_RWTUN, &witness_kdb, 0, "");
399 #if defined(DDB) || defined(KDB)
401 * When DDB or KDB is enabled and witness_trace is 1, it will cause the system
402 * to print a stack trace:
403 * - a lock hierarchy violation occurs
404 * - locks are held when going to sleep.
406 int witness_trace = 1;
407 SYSCTL_INT(_debug_witness, OID_AUTO, trace, CTLFLAG_RWTUN, &witness_trace, 0, "");
408 #endif /* DDB || KDB */
410 #ifdef WITNESS_SKIPSPIN
411 int witness_skipspin = 1;
413 int witness_skipspin = 0;
415 SYSCTL_INT(_debug_witness, OID_AUTO, skipspin, CTLFLAG_RDTUN, &witness_skipspin, 0, "");
417 int badstack_sbuf_size;
419 int witness_count = WITNESS_COUNT;
420 SYSCTL_INT(_debug_witness, OID_AUTO, witness_count, CTLFLAG_RDTUN,
421 &witness_count, 0, "");
424 * Output channel for witness messages. By default we print to the console.
426 enum witness_channel {
432 static enum witness_channel witness_channel = WITNESS_CONSOLE;
433 SYSCTL_PROC(_debug_witness, OID_AUTO, output_channel, CTLTYPE_STRING |
434 CTLFLAG_RWTUN, NULL, 0, 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, CTLTYPE_STRING | CTLFLAG_RD,
441 NULL, 0, sysctl_debug_witness_fullgraph, "A", "Show locks relation graphs");
444 * Call this to print out the witness faulty stacks.
446 SYSCTL_PROC(_debug_witness, OID_AUTO, badstacks, CTLTYPE_STRING | CTLFLAG_RD,
447 NULL, 0, sysctl_debug_witness_badstacks, "A", "Show bad witness stacks");
449 static struct mtx w_mtx;
452 static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free);
453 static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all);
456 static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin);
457 static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep);
460 static struct lock_list_entry *w_lock_list_free = NULL;
461 static struct witness_pendhelp pending_locks[WITNESS_PENDLIST];
462 static u_int pending_cnt;
464 static int w_free_cnt, w_spin_cnt, w_sleep_cnt;
465 SYSCTL_INT(_debug_witness, OID_AUTO, free_cnt, CTLFLAG_RD, &w_free_cnt, 0, "");
466 SYSCTL_INT(_debug_witness, OID_AUTO, spin_cnt, CTLFLAG_RD, &w_spin_cnt, 0, "");
467 SYSCTL_INT(_debug_witness, OID_AUTO, sleep_cnt, CTLFLAG_RD, &w_sleep_cnt, 0,
470 static struct witness *w_data;
471 static uint8_t **w_rmatrix;
472 static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT];
473 static struct witness_hash w_hash; /* The witness hash table. */
475 /* The lock order data hash */
476 static struct witness_lock_order_data w_lodata[WITNESS_LO_DATA_COUNT];
477 static struct witness_lock_order_data *w_lofree = NULL;
478 static struct witness_lock_order_hash w_lohash;
479 static int w_max_used_index = 0;
480 static unsigned int w_generation = 0;
481 static const char w_notrunning[] = "Witness not running\n";
482 static const char w_stillcold[] = "Witness is still cold\n";
485 static struct witness_order_list_entry order_lists[] = {
489 { "proctree", &lock_class_sx },
490 { "allproc", &lock_class_sx },
491 { "allprison", &lock_class_sx },
496 { "Giant", &lock_class_mtx_sleep },
497 { "pipe mutex", &lock_class_mtx_sleep },
498 { "sigio lock", &lock_class_mtx_sleep },
499 { "process group", &lock_class_mtx_sleep },
500 { "process lock", &lock_class_mtx_sleep },
501 { "session", &lock_class_mtx_sleep },
502 { "uidinfo hash", &lock_class_rw },
504 { "pmc-sleep", &lock_class_mtx_sleep },
506 { "time lock", &lock_class_mtx_sleep },
511 { "umtx lock", &lock_class_mtx_sleep },
516 { "accept", &lock_class_mtx_sleep },
517 { "so_snd", &lock_class_mtx_sleep },
518 { "so_rcv", &lock_class_mtx_sleep },
519 { "sellck", &lock_class_mtx_sleep },
524 { "so_rcv", &lock_class_mtx_sleep },
525 { "radix node head", &lock_class_rw },
526 { "rtentry", &lock_class_mtx_sleep },
527 { "ifaddr", &lock_class_mtx_sleep },
531 * protocol locks before interface locks, after UDP locks.
533 { "udpinp", &lock_class_rw },
534 { "in_multi_mtx", &lock_class_mtx_sleep },
535 { "igmp_mtx", &lock_class_mtx_sleep },
536 { "if_addr_lock", &lock_class_rw },
540 * protocol locks before interface locks, after UDP locks.
542 { "udpinp", &lock_class_rw },
543 { "in6_multi_mtx", &lock_class_mtx_sleep },
544 { "mld_mtx", &lock_class_mtx_sleep },
545 { "if_addr_lock", &lock_class_rw },
548 * UNIX Domain Sockets
550 { "unp_link_rwlock", &lock_class_rw },
551 { "unp_list_lock", &lock_class_mtx_sleep },
552 { "unp", &lock_class_mtx_sleep },
553 { "so_snd", &lock_class_mtx_sleep },
558 { "udp", &lock_class_rw },
559 { "udpinp", &lock_class_rw },
560 { "so_snd", &lock_class_mtx_sleep },
565 { "tcp", &lock_class_rw },
566 { "tcpinp", &lock_class_rw },
567 { "so_snd", &lock_class_mtx_sleep },
572 { "bpf global lock", &lock_class_mtx_sleep },
573 { "bpf interface lock", &lock_class_rw },
574 { "bpf cdev lock", &lock_class_mtx_sleep },
579 { "nfsd_mtx", &lock_class_mtx_sleep },
580 { "so_snd", &lock_class_mtx_sleep },
586 { "802.11 com lock", &lock_class_mtx_sleep},
591 { "network driver", &lock_class_mtx_sleep},
597 { "ng_node", &lock_class_mtx_sleep },
598 { "ng_worklist", &lock_class_mtx_sleep },
603 { "vm map (system)", &lock_class_mtx_sleep },
604 { "vm pagequeue", &lock_class_mtx_sleep },
605 { "vnode interlock", &lock_class_mtx_sleep },
606 { "cdev", &lock_class_mtx_sleep },
611 { "vm map (user)", &lock_class_sx },
612 { "vm object", &lock_class_rw },
613 { "vm page", &lock_class_mtx_sleep },
614 { "vm pagequeue", &lock_class_mtx_sleep },
615 { "pmap pv global", &lock_class_rw },
616 { "pmap", &lock_class_mtx_sleep },
617 { "pmap pv list", &lock_class_rw },
618 { "vm page free queue", &lock_class_mtx_sleep },
621 * kqueue/VFS interaction
623 { "kqueue", &lock_class_mtx_sleep },
624 { "struct mount mtx", &lock_class_mtx_sleep },
625 { "vnode interlock", &lock_class_mtx_sleep },
630 { "ncvn", &lock_class_mtx_sleep },
631 { "ncbuc", &lock_class_rw },
632 { "vnode interlock", &lock_class_mtx_sleep },
633 { "ncneg", &lock_class_mtx_sleep },
638 { "dn->dn_mtx", &lock_class_sx },
639 { "dr->dt.di.dr_mtx", &lock_class_sx },
640 { "db->db_mtx", &lock_class_sx },
646 { "ap boot", &lock_class_mtx_spin },
648 { "rm.mutex_mtx", &lock_class_mtx_spin },
649 { "sio", &lock_class_mtx_spin },
651 { "cy", &lock_class_mtx_spin },
654 { "pcib_mtx", &lock_class_mtx_spin },
655 { "rtc_mtx", &lock_class_mtx_spin },
657 { "scc_hwmtx", &lock_class_mtx_spin },
658 { "uart_hwmtx", &lock_class_mtx_spin },
659 { "fast_taskqueue", &lock_class_mtx_spin },
660 { "intr table", &lock_class_mtx_spin },
662 { "pmc-per-proc", &lock_class_mtx_spin },
664 { "process slock", &lock_class_mtx_spin },
665 { "syscons video lock", &lock_class_mtx_spin },
666 { "sleepq chain", &lock_class_mtx_spin },
667 { "rm_spinlock", &lock_class_mtx_spin },
668 { "turnstile chain", &lock_class_mtx_spin },
669 { "turnstile lock", &lock_class_mtx_spin },
670 { "sched lock", &lock_class_mtx_spin },
671 { "td_contested", &lock_class_mtx_spin },
672 { "callout", &lock_class_mtx_spin },
673 { "entropy harvest mutex", &lock_class_mtx_spin },
675 { "smp rendezvous", &lock_class_mtx_spin },
678 { "tlb0", &lock_class_mtx_spin },
683 { "intrcnt", &lock_class_mtx_spin },
684 { "icu", &lock_class_mtx_spin },
685 #if defined(SMP) && defined(__sparc64__)
686 { "ipi", &lock_class_mtx_spin },
689 { "allpmaps", &lock_class_mtx_spin },
690 { "descriptor tables", &lock_class_mtx_spin },
692 { "clk", &lock_class_mtx_spin },
693 { "cpuset", &lock_class_mtx_spin },
694 { "mprof lock", &lock_class_mtx_spin },
695 { "zombie lock", &lock_class_mtx_spin },
696 { "ALD Queue", &lock_class_mtx_spin },
697 #if defined(__i386__) || defined(__amd64__)
698 { "pcicfg", &lock_class_mtx_spin },
699 { "NDIS thread lock", &lock_class_mtx_spin },
701 { "tw_osl_io_lock", &lock_class_mtx_spin },
702 { "tw_osl_q_lock", &lock_class_mtx_spin },
703 { "tw_cl_io_lock", &lock_class_mtx_spin },
704 { "tw_cl_intr_lock", &lock_class_mtx_spin },
705 { "tw_cl_gen_lock", &lock_class_mtx_spin },
707 { "pmc-leaf", &lock_class_mtx_spin },
709 { "blocked lock", &lock_class_mtx_spin },
716 * Pairs of locks which have been blessed
717 * Don't complain about order problems with blessed locks
719 static struct witness_blessed blessed_list[] = {
724 * This global is set to 0 once it becomes safe to use the witness code.
726 static int witness_cold = 1;
729 * This global is set to 1 once the static lock orders have been enrolled
730 * so that a warning can be issued for any spin locks enrolled later.
732 static int witness_spin_warn = 0;
734 /* Trim useless garbage from filenames. */
736 fixup_filename(const char *file)
741 while (strncmp(file, "../", 3) == 0)
747 * The WITNESS-enabled diagnostic code. Note that the witness code does
748 * assume that the early boot is single-threaded at least until after this
749 * routine is completed.
752 witness_initialize(void *dummy __unused)
754 struct lock_object *lock;
755 struct witness_order_list_entry *order;
756 struct witness *w, *w1;
759 w_data = malloc(sizeof (struct witness) * witness_count, M_WITNESS,
762 w_rmatrix = malloc(sizeof(*w_rmatrix) * (witness_count + 1),
763 M_WITNESS, M_WAITOK | M_ZERO);
765 for (i = 0; i < witness_count + 1; i++) {
766 w_rmatrix[i] = malloc(sizeof(*w_rmatrix[i]) *
767 (witness_count + 1), M_WITNESS, M_WAITOK | M_ZERO);
769 badstack_sbuf_size = witness_count * 256;
772 * We have to release Giant before initializing its witness
773 * structure so that WITNESS doesn't get confused.
776 mtx_assert(&Giant, MA_NOTOWNED);
778 CTR1(KTR_WITNESS, "%s: initializing witness", __func__);
779 mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET |
780 MTX_NOWITNESS | MTX_NOPROFILE);
781 for (i = witness_count - 1; i >= 0; i--) {
783 memset(w, 0, sizeof(*w));
784 w_data[i].w_index = i; /* Witness index never changes. */
787 KASSERT(STAILQ_FIRST(&w_free)->w_index == 0,
788 ("%s: Invalid list of free witness objects", __func__));
790 /* Witness with index 0 is not used to aid in debugging. */
791 STAILQ_REMOVE_HEAD(&w_free, w_list);
794 for (i = 0; i < witness_count; i++) {
795 memset(w_rmatrix[i], 0, sizeof(*w_rmatrix[i]) *
796 (witness_count + 1));
799 for (i = 0; i < LOCK_CHILDCOUNT; i++)
800 witness_lock_list_free(&w_locklistdata[i]);
801 witness_init_hash_tables();
803 /* First add in all the specified order lists. */
804 for (order = order_lists; order->w_name != NULL; order++) {
805 w = enroll(order->w_name, order->w_class);
808 w->w_file = "order list";
809 for (order++; order->w_name != NULL; order++) {
810 w1 = enroll(order->w_name, order->w_class);
813 w1->w_file = "order list";
818 witness_spin_warn = 1;
820 /* Iterate through all locks and add them to witness. */
821 for (i = 0; pending_locks[i].wh_lock != NULL; i++) {
822 lock = pending_locks[i].wh_lock;
823 KASSERT(lock->lo_flags & LO_WITNESS,
824 ("%s: lock %s is on pending list but not LO_WITNESS",
825 __func__, lock->lo_name));
826 lock->lo_witness = enroll(pending_locks[i].wh_type,
830 /* Mark the witness code as being ready for use. */
835 SYSINIT(witness_init, SI_SUB_WITNESS, SI_ORDER_FIRST, witness_initialize,
839 witness_init(struct lock_object *lock, const char *type)
841 struct lock_class *class;
843 /* Various sanity checks. */
844 class = LOCK_CLASS(lock);
845 if ((lock->lo_flags & LO_RECURSABLE) != 0 &&
846 (class->lc_flags & LC_RECURSABLE) == 0)
847 kassert_panic("%s: lock (%s) %s can not be recursable",
848 __func__, class->lc_name, lock->lo_name);
849 if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
850 (class->lc_flags & LC_SLEEPABLE) == 0)
851 kassert_panic("%s: lock (%s) %s can not be sleepable",
852 __func__, class->lc_name, lock->lo_name);
853 if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
854 (class->lc_flags & LC_UPGRADABLE) == 0)
855 kassert_panic("%s: lock (%s) %s can not be upgradable",
856 __func__, class->lc_name, lock->lo_name);
859 * If we shouldn't watch this lock, then just clear lo_witness.
860 * Otherwise, if witness_cold is set, then it is too early to
861 * enroll this lock, so defer it to witness_initialize() by adding
862 * it to the pending_locks list. If it is not too early, then enroll
865 if (witness_watch < 1 || panicstr != NULL ||
866 (lock->lo_flags & LO_WITNESS) == 0)
867 lock->lo_witness = NULL;
868 else if (witness_cold) {
869 pending_locks[pending_cnt].wh_lock = lock;
870 pending_locks[pending_cnt++].wh_type = type;
871 if (pending_cnt > WITNESS_PENDLIST)
872 panic("%s: pending locks list is too small, "
873 "increase WITNESS_PENDLIST\n",
876 lock->lo_witness = enroll(type, class);
880 witness_destroy(struct lock_object *lock)
882 struct lock_class *class;
885 class = LOCK_CLASS(lock);
888 panic("lock (%s) %s destroyed while witness_cold",
889 class->lc_name, lock->lo_name);
891 /* XXX: need to verify that no one holds the lock */
892 if ((lock->lo_flags & LO_WITNESS) == 0 || lock->lo_witness == NULL)
894 w = lock->lo_witness;
896 mtx_lock_spin(&w_mtx);
897 MPASS(w->w_refcount > 0);
900 if (w->w_refcount == 0)
902 mtx_unlock_spin(&w_mtx);
907 witness_ddb_compute_levels(void)
912 * First clear all levels.
914 STAILQ_FOREACH(w, &w_all, w_list)
918 * Look for locks with no parents and level all their descendants.
920 STAILQ_FOREACH(w, &w_all, w_list) {
922 /* If the witness has ancestors (is not a root), skip it. */
923 if (w->w_num_ancestors > 0)
925 witness_ddb_level_descendants(w, 0);
930 witness_ddb_level_descendants(struct witness *w, int l)
934 if (w->w_ddb_level >= l)
940 for (i = 1; i <= w_max_used_index; i++) {
941 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
942 witness_ddb_level_descendants(&w_data[i], l);
947 witness_ddb_display_descendants(int(*prnt)(const char *fmt, ...),
948 struct witness *w, int indent)
952 for (i = 0; i < indent; i++)
954 prnt("%s (type: %s, depth: %d, active refs: %d)",
955 w->w_name, w->w_class->lc_name,
956 w->w_ddb_level, w->w_refcount);
957 if (w->w_displayed) {
958 prnt(" -- (already displayed)\n");
962 if (w->w_file != NULL && w->w_line != 0)
963 prnt(" -- last acquired @ %s:%d\n", fixup_filename(w->w_file),
966 prnt(" -- never acquired\n");
968 WITNESS_INDEX_ASSERT(w->w_index);
969 for (i = 1; i <= w_max_used_index; i++) {
972 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
973 witness_ddb_display_descendants(prnt, &w_data[i],
979 witness_ddb_display_list(int(*prnt)(const char *fmt, ...),
980 struct witness_list *list)
984 STAILQ_FOREACH(w, list, w_typelist) {
985 if (w->w_file == NULL || w->w_ddb_level > 0)
988 /* This lock has no anscestors - display its descendants. */
989 witness_ddb_display_descendants(prnt, w, 0);
996 witness_ddb_display(int(*prnt)(const char *fmt, ...))
1000 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1001 witness_ddb_compute_levels();
1003 /* Clear all the displayed flags. */
1004 STAILQ_FOREACH(w, &w_all, w_list)
1008 * First, handle sleep locks which have been acquired at least
1011 prnt("Sleep locks:\n");
1012 witness_ddb_display_list(prnt, &w_sleep);
1017 * Now do spin locks which have been acquired at least once.
1019 prnt("\nSpin locks:\n");
1020 witness_ddb_display_list(prnt, &w_spin);
1025 * Finally, any locks which have not been acquired yet.
1027 prnt("\nLocks which were never acquired:\n");
1028 STAILQ_FOREACH(w, &w_all, w_list) {
1029 if (w->w_file != NULL || w->w_refcount == 0)
1031 prnt("%s (type: %s, depth: %d)\n", w->w_name,
1032 w->w_class->lc_name, w->w_ddb_level);
1040 witness_defineorder(struct lock_object *lock1, struct lock_object *lock2)
1043 if (witness_watch == -1 || panicstr != NULL)
1046 /* Require locks that witness knows about. */
1047 if (lock1 == NULL || lock1->lo_witness == NULL || lock2 == NULL ||
1048 lock2->lo_witness == NULL)
1051 mtx_assert(&w_mtx, MA_NOTOWNED);
1052 mtx_lock_spin(&w_mtx);
1055 * If we already have either an explicit or implied lock order that
1056 * is the other way around, then return an error.
1058 if (witness_watch &&
1059 isitmydescendant(lock2->lo_witness, lock1->lo_witness)) {
1060 mtx_unlock_spin(&w_mtx);
1064 /* Try to add the new order. */
1065 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1066 lock2->lo_witness->w_name, lock1->lo_witness->w_name);
1067 itismychild(lock1->lo_witness, lock2->lo_witness);
1068 mtx_unlock_spin(&w_mtx);
1073 witness_checkorder(struct lock_object *lock, int flags, const char *file,
1074 int line, struct lock_object *interlock)
1076 struct lock_list_entry *lock_list, *lle;
1077 struct lock_instance *lock1, *lock2, *plock;
1078 struct lock_class *class, *iclass;
1079 struct witness *w, *w1;
1083 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL ||
1087 w = lock->lo_witness;
1088 class = LOCK_CLASS(lock);
1091 if (class->lc_flags & LC_SLEEPLOCK) {
1094 * Since spin locks include a critical section, this check
1095 * implicitly enforces a lock order of all sleep locks before
1098 if (td->td_critnest != 0 && !kdb_active)
1099 kassert_panic("acquiring blockable sleep lock with "
1100 "spinlock or critical section held (%s) %s @ %s:%d",
1101 class->lc_name, lock->lo_name,
1102 fixup_filename(file), line);
1105 * If this is the first lock acquired then just return as
1106 * no order checking is needed.
1108 lock_list = td->td_sleeplocks;
1109 if (lock_list == NULL || lock_list->ll_count == 0)
1114 * If this is the first lock, just return as no order
1115 * checking is needed. Avoid problems with thread
1116 * migration pinning the thread while checking if
1117 * spinlocks are held. If at least one spinlock is held
1118 * the thread is in a safe path and it is allowed to
1122 lock_list = PCPU_GET(spinlocks);
1123 if (lock_list == NULL || lock_list->ll_count == 0) {
1131 * Check to see if we are recursing on a lock we already own. If
1132 * so, make sure that we don't mismatch exclusive and shared lock
1135 lock1 = find_instance(lock_list, lock);
1136 if (lock1 != NULL) {
1137 if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
1138 (flags & LOP_EXCLUSIVE) == 0) {
1139 witness_output("shared lock of (%s) %s @ %s:%d\n",
1140 class->lc_name, lock->lo_name,
1141 fixup_filename(file), line);
1142 witness_output("while exclusively locked from %s:%d\n",
1143 fixup_filename(lock1->li_file), lock1->li_line);
1144 kassert_panic("excl->share");
1146 if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
1147 (flags & LOP_EXCLUSIVE) != 0) {
1148 witness_output("exclusive lock of (%s) %s @ %s:%d\n",
1149 class->lc_name, lock->lo_name,
1150 fixup_filename(file), line);
1151 witness_output("while share locked from %s:%d\n",
1152 fixup_filename(lock1->li_file), lock1->li_line);
1153 kassert_panic("share->excl");
1158 /* Warn if the interlock is not locked exactly once. */
1159 if (interlock != NULL) {
1160 iclass = LOCK_CLASS(interlock);
1161 lock1 = find_instance(lock_list, interlock);
1163 kassert_panic("interlock (%s) %s not locked @ %s:%d",
1164 iclass->lc_name, interlock->lo_name,
1165 fixup_filename(file), line);
1166 else if ((lock1->li_flags & LI_RECURSEMASK) != 0)
1167 kassert_panic("interlock (%s) %s recursed @ %s:%d",
1168 iclass->lc_name, interlock->lo_name,
1169 fixup_filename(file), line);
1173 * Find the previously acquired lock, but ignore interlocks.
1175 plock = &lock_list->ll_children[lock_list->ll_count - 1];
1176 if (interlock != NULL && plock->li_lock == interlock) {
1177 if (lock_list->ll_count > 1)
1179 &lock_list->ll_children[lock_list->ll_count - 2];
1181 lle = lock_list->ll_next;
1184 * The interlock is the only lock we hold, so
1189 plock = &lle->ll_children[lle->ll_count - 1];
1194 * Try to perform most checks without a lock. If this succeeds we
1195 * can skip acquiring the lock and return success. Otherwise we redo
1196 * the check with the lock held to handle races with concurrent updates.
1198 w1 = plock->li_lock->lo_witness;
1199 if (witness_lock_order_check(w1, w))
1202 mtx_lock_spin(&w_mtx);
1203 if (witness_lock_order_check(w1, w)) {
1204 mtx_unlock_spin(&w_mtx);
1207 witness_lock_order_add(w1, w);
1210 * Check for duplicate locks of the same type. Note that we only
1211 * have to check for this on the last lock we just acquired. Any
1212 * other cases will be caught as lock order violations.
1216 if (!(lock->lo_flags & LO_DUPOK) && !(flags & LOP_DUPOK) &&
1217 !(w_rmatrix[i][i] & WITNESS_REVERSAL)) {
1218 w_rmatrix[i][i] |= WITNESS_REVERSAL;
1220 mtx_unlock_spin(&w_mtx);
1222 "acquiring duplicate lock of same type: \"%s\"\n",
1224 witness_output(" 1st %s @ %s:%d\n", plock->li_lock->lo_name,
1225 fixup_filename(plock->li_file), plock->li_line);
1226 witness_output(" 2nd %s @ %s:%d\n", lock->lo_name,
1227 fixup_filename(file), line);
1228 witness_debugger(1, __func__);
1230 mtx_unlock_spin(&w_mtx);
1233 mtx_assert(&w_mtx, MA_OWNED);
1236 * If we know that the lock we are acquiring comes after
1237 * the lock we most recently acquired in the lock order tree,
1238 * then there is no need for any further checks.
1240 if (isitmychild(w1, w))
1243 for (j = 0, lle = lock_list; lle != NULL; lle = lle->ll_next) {
1244 for (i = lle->ll_count - 1; i >= 0; i--, j++) {
1246 MPASS(j < LOCK_CHILDCOUNT * LOCK_NCHILDREN);
1247 lock1 = &lle->ll_children[i];
1250 * Ignore the interlock.
1252 if (interlock == lock1->li_lock)
1256 * If this lock doesn't undergo witness checking,
1259 w1 = lock1->li_lock->lo_witness;
1261 KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
1262 ("lock missing witness structure"));
1267 * If we are locking Giant and this is a sleepable
1268 * lock, then skip it.
1270 if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0 &&
1271 lock == &Giant.lock_object)
1275 * If we are locking a sleepable lock and this lock
1276 * is Giant, then skip it.
1278 if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1279 lock1->li_lock == &Giant.lock_object)
1283 * If we are locking a sleepable lock and this lock
1284 * isn't sleepable, we want to treat it as a lock
1285 * order violation to enfore a general lock order of
1286 * sleepable locks before non-sleepable locks.
1288 if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1289 (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
1293 * If we are locking Giant and this is a non-sleepable
1294 * lock, then treat it as a reversal.
1296 if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0 &&
1297 lock == &Giant.lock_object)
1301 * Check the lock order hierarchy for a reveresal.
1303 if (!isitmydescendant(w, w1))
1308 * We have a lock order violation, check to see if it
1309 * is allowed or has already been yelled about.
1314 * If the lock order is blessed, just bail. We don't
1315 * look for other lock order violations though, which
1322 /* Bail if this violation is known */
1323 if (w_rmatrix[w1->w_index][w->w_index] & WITNESS_REVERSAL)
1326 /* Record this as a violation */
1327 w_rmatrix[w1->w_index][w->w_index] |= WITNESS_REVERSAL;
1328 w_rmatrix[w->w_index][w1->w_index] |= WITNESS_REVERSAL;
1329 w->w_reversed = w1->w_reversed = 1;
1330 witness_increment_graph_generation();
1331 mtx_unlock_spin(&w_mtx);
1333 #ifdef WITNESS_NO_VNODE
1335 * There are known LORs between VNODE locks. They are
1336 * not an indication of a bug. VNODE locks are flagged
1337 * as such (LO_IS_VNODE) and we don't yell if the LOR
1338 * is between 2 VNODE locks.
1340 if ((lock->lo_flags & LO_IS_VNODE) != 0 &&
1341 (lock1->li_lock->lo_flags & LO_IS_VNODE) != 0)
1346 * Ok, yell about it.
1348 if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1349 (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
1351 "lock order reversal: (sleepable after non-sleepable)\n");
1352 else if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0
1353 && lock == &Giant.lock_object)
1355 "lock order reversal: (Giant after non-sleepable)\n");
1357 witness_output("lock order reversal:\n");
1360 * Try to locate an earlier lock with
1361 * witness w in our list.
1364 lock2 = &lle->ll_children[i];
1365 MPASS(lock2->li_lock != NULL);
1366 if (lock2->li_lock->lo_witness == w)
1368 if (i == 0 && lle->ll_next != NULL) {
1370 i = lle->ll_count - 1;
1371 MPASS(i >= 0 && i < LOCK_NCHILDREN);
1376 witness_output(" 1st %p %s (%s) @ %s:%d\n",
1377 lock1->li_lock, lock1->li_lock->lo_name,
1378 w1->w_name, fixup_filename(lock1->li_file),
1380 witness_output(" 2nd %p %s (%s) @ %s:%d\n", lock,
1381 lock->lo_name, w->w_name,
1382 fixup_filename(file), line);
1384 witness_output(" 1st %p %s (%s) @ %s:%d\n",
1385 lock2->li_lock, lock2->li_lock->lo_name,
1386 lock2->li_lock->lo_witness->w_name,
1387 fixup_filename(lock2->li_file),
1389 witness_output(" 2nd %p %s (%s) @ %s:%d\n",
1390 lock1->li_lock, lock1->li_lock->lo_name,
1391 w1->w_name, fixup_filename(lock1->li_file),
1393 witness_output(" 3rd %p %s (%s) @ %s:%d\n", lock,
1394 lock->lo_name, w->w_name,
1395 fixup_filename(file), line);
1397 witness_debugger(1, __func__);
1403 * If requested, build a new lock order. However, don't build a new
1404 * relationship between a sleepable lock and Giant if it is in the
1405 * wrong direction. The correct lock order is that sleepable locks
1406 * always come before Giant.
1408 if (flags & LOP_NEWORDER &&
1409 !(plock->li_lock == &Giant.lock_object &&
1410 (lock->lo_flags & LO_SLEEPABLE) != 0)) {
1411 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1412 w->w_name, plock->li_lock->lo_witness->w_name);
1413 itismychild(plock->li_lock->lo_witness, w);
1416 mtx_unlock_spin(&w_mtx);
1420 witness_lock(struct lock_object *lock, int flags, const char *file, int line)
1422 struct lock_list_entry **lock_list, *lle;
1423 struct lock_instance *instance;
1427 if (witness_cold || witness_watch == -1 || lock->lo_witness == NULL ||
1430 w = lock->lo_witness;
1433 /* Determine lock list for this lock. */
1434 if (LOCK_CLASS(lock)->lc_flags & LC_SLEEPLOCK)
1435 lock_list = &td->td_sleeplocks;
1437 lock_list = PCPU_PTR(spinlocks);
1439 /* Check to see if we are recursing on a lock we already own. */
1440 instance = find_instance(*lock_list, lock);
1441 if (instance != NULL) {
1442 instance->li_flags++;
1443 CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
1444 td->td_proc->p_pid, lock->lo_name,
1445 instance->li_flags & LI_RECURSEMASK);
1446 instance->li_file = file;
1447 instance->li_line = line;
1451 /* Update per-witness last file and line acquire. */
1455 /* Find the next open lock instance in the list and fill it. */
1457 if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
1458 lle = witness_lock_list_get();
1461 lle->ll_next = *lock_list;
1462 CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
1463 td->td_proc->p_pid, lle);
1466 instance = &lle->ll_children[lle->ll_count++];
1467 instance->li_lock = lock;
1468 instance->li_line = line;
1469 instance->li_file = file;
1470 if ((flags & LOP_EXCLUSIVE) != 0)
1471 instance->li_flags = LI_EXCLUSIVE;
1473 instance->li_flags = 0;
1474 CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
1475 td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
1479 witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
1481 struct lock_instance *instance;
1482 struct lock_class *class;
1484 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1485 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
1487 class = LOCK_CLASS(lock);
1488 if (witness_watch) {
1489 if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1491 "upgrade of non-upgradable lock (%s) %s @ %s:%d",
1492 class->lc_name, lock->lo_name,
1493 fixup_filename(file), line);
1494 if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1496 "upgrade of non-sleep lock (%s) %s @ %s:%d",
1497 class->lc_name, lock->lo_name,
1498 fixup_filename(file), line);
1500 instance = find_instance(curthread->td_sleeplocks, lock);
1501 if (instance == NULL) {
1502 kassert_panic("upgrade of unlocked lock (%s) %s @ %s:%d",
1503 class->lc_name, lock->lo_name,
1504 fixup_filename(file), line);
1507 if (witness_watch) {
1508 if ((instance->li_flags & LI_EXCLUSIVE) != 0)
1510 "upgrade of exclusive lock (%s) %s @ %s:%d",
1511 class->lc_name, lock->lo_name,
1512 fixup_filename(file), line);
1513 if ((instance->li_flags & LI_RECURSEMASK) != 0)
1515 "upgrade of recursed lock (%s) %s r=%d @ %s:%d",
1516 class->lc_name, lock->lo_name,
1517 instance->li_flags & LI_RECURSEMASK,
1518 fixup_filename(file), line);
1520 instance->li_flags |= LI_EXCLUSIVE;
1524 witness_downgrade(struct lock_object *lock, int flags, const char *file,
1527 struct lock_instance *instance;
1528 struct lock_class *class;
1530 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1531 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
1533 class = LOCK_CLASS(lock);
1534 if (witness_watch) {
1535 if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1537 "downgrade of non-upgradable lock (%s) %s @ %s:%d",
1538 class->lc_name, lock->lo_name,
1539 fixup_filename(file), line);
1540 if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1542 "downgrade of non-sleep lock (%s) %s @ %s:%d",
1543 class->lc_name, lock->lo_name,
1544 fixup_filename(file), line);
1546 instance = find_instance(curthread->td_sleeplocks, lock);
1547 if (instance == NULL) {
1548 kassert_panic("downgrade of unlocked lock (%s) %s @ %s:%d",
1549 class->lc_name, lock->lo_name,
1550 fixup_filename(file), line);
1553 if (witness_watch) {
1554 if ((instance->li_flags & LI_EXCLUSIVE) == 0)
1556 "downgrade of shared lock (%s) %s @ %s:%d",
1557 class->lc_name, lock->lo_name,
1558 fixup_filename(file), line);
1559 if ((instance->li_flags & LI_RECURSEMASK) != 0)
1561 "downgrade of recursed lock (%s) %s r=%d @ %s:%d",
1562 class->lc_name, lock->lo_name,
1563 instance->li_flags & LI_RECURSEMASK,
1564 fixup_filename(file), line);
1566 instance->li_flags &= ~LI_EXCLUSIVE;
1570 witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
1572 struct lock_list_entry **lock_list, *lle;
1573 struct lock_instance *instance;
1574 struct lock_class *class;
1579 if (witness_cold || lock->lo_witness == NULL || panicstr != NULL)
1582 class = LOCK_CLASS(lock);
1584 /* Find lock instance associated with this lock. */
1585 if (class->lc_flags & LC_SLEEPLOCK)
1586 lock_list = &td->td_sleeplocks;
1588 lock_list = PCPU_PTR(spinlocks);
1590 for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
1591 for (i = 0; i < (*lock_list)->ll_count; i++) {
1592 instance = &(*lock_list)->ll_children[i];
1593 if (instance->li_lock == lock)
1598 * When disabling WITNESS through witness_watch we could end up in
1599 * having registered locks in the td_sleeplocks queue.
1600 * We have to make sure we flush these queues, so just search for
1601 * eventual register locks and remove them.
1603 if (witness_watch > 0) {
1604 kassert_panic("lock (%s) %s not locked @ %s:%d", class->lc_name,
1605 lock->lo_name, fixup_filename(file), line);
1612 /* First, check for shared/exclusive mismatches. */
1613 if ((instance->li_flags & LI_EXCLUSIVE) != 0 && witness_watch > 0 &&
1614 (flags & LOP_EXCLUSIVE) == 0) {
1615 witness_output("shared unlock of (%s) %s @ %s:%d\n",
1616 class->lc_name, lock->lo_name, fixup_filename(file), line);
1617 witness_output("while exclusively locked from %s:%d\n",
1618 fixup_filename(instance->li_file), instance->li_line);
1619 kassert_panic("excl->ushare");
1621 if ((instance->li_flags & LI_EXCLUSIVE) == 0 && witness_watch > 0 &&
1622 (flags & LOP_EXCLUSIVE) != 0) {
1623 witness_output("exclusive unlock of (%s) %s @ %s:%d\n",
1624 class->lc_name, lock->lo_name, fixup_filename(file), line);
1625 witness_output("while share locked from %s:%d\n",
1626 fixup_filename(instance->li_file),
1628 kassert_panic("share->uexcl");
1630 /* If we are recursed, unrecurse. */
1631 if ((instance->li_flags & LI_RECURSEMASK) > 0) {
1632 CTR4(KTR_WITNESS, "%s: pid %d unrecursed on %s r=%d", __func__,
1633 td->td_proc->p_pid, instance->li_lock->lo_name,
1634 instance->li_flags);
1635 instance->li_flags--;
1638 /* The lock is now being dropped, check for NORELEASE flag */
1639 if ((instance->li_flags & LI_NORELEASE) != 0 && witness_watch > 0) {
1640 witness_output("forbidden unlock of (%s) %s @ %s:%d\n",
1641 class->lc_name, lock->lo_name, fixup_filename(file), line);
1642 kassert_panic("lock marked norelease");
1645 /* Otherwise, remove this item from the list. */
1647 CTR4(KTR_WITNESS, "%s: pid %d removed %s from lle[%d]", __func__,
1648 td->td_proc->p_pid, instance->li_lock->lo_name,
1649 (*lock_list)->ll_count - 1);
1650 for (j = i; j < (*lock_list)->ll_count - 1; j++)
1651 (*lock_list)->ll_children[j] =
1652 (*lock_list)->ll_children[j + 1];
1653 (*lock_list)->ll_count--;
1657 * In order to reduce contention on w_mtx, we want to keep always an
1658 * head object into lists so that frequent allocation from the
1659 * free witness pool (and subsequent locking) is avoided.
1660 * In order to maintain the current code simple, when the head
1661 * object is totally unloaded it means also that we do not have
1662 * further objects in the list, so the list ownership needs to be
1663 * hand over to another object if the current head needs to be freed.
1665 if ((*lock_list)->ll_count == 0) {
1666 if (*lock_list == lle) {
1667 if (lle->ll_next == NULL)
1671 *lock_list = lle->ll_next;
1672 CTR3(KTR_WITNESS, "%s: pid %d removed lle %p", __func__,
1673 td->td_proc->p_pid, lle);
1674 witness_lock_list_free(lle);
1679 witness_thread_exit(struct thread *td)
1681 struct lock_list_entry *lle;
1684 lle = td->td_sleeplocks;
1685 if (lle == NULL || panicstr != NULL)
1687 if (lle->ll_count != 0) {
1688 for (n = 0; lle != NULL; lle = lle->ll_next)
1689 for (i = lle->ll_count - 1; i >= 0; i--) {
1692 "Thread %p exiting with the following locks held:\n", td);
1694 witness_list_lock(&lle->ll_children[i],
1699 "Thread %p cannot exit while holding sleeplocks\n", td);
1701 witness_lock_list_free(lle);
1705 * Warn if any locks other than 'lock' are held. Flags can be passed in to
1706 * exempt Giant and sleepable locks from the checks as well. If any
1707 * non-exempt locks are held, then a supplied message is printed to the
1708 * output channel along with a list of the offending locks. If indicated in the
1709 * flags then a failure results in a panic as well.
1712 witness_warn(int flags, struct lock_object *lock, const char *fmt, ...)
1714 struct lock_list_entry *lock_list, *lle;
1715 struct lock_instance *lock1;
1720 if (witness_cold || witness_watch < 1 || panicstr != NULL)
1724 for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next)
1725 for (i = lle->ll_count - 1; i >= 0; i--) {
1726 lock1 = &lle->ll_children[i];
1727 if (lock1->li_lock == lock)
1729 if (flags & WARN_GIANTOK &&
1730 lock1->li_lock == &Giant.lock_object)
1732 if (flags & WARN_SLEEPOK &&
1733 (lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0)
1739 printf(" with the following %slocks held:\n",
1740 (flags & WARN_SLEEPOK) != 0 ?
1741 "non-sleepable " : "");
1744 witness_list_lock(lock1, printf);
1748 * Pin the thread in order to avoid problems with thread migration.
1749 * Once that all verifies are passed about spinlocks ownership,
1750 * the thread is in a safe path and it can be unpinned.
1753 lock_list = PCPU_GET(spinlocks);
1754 if (lock_list != NULL && lock_list->ll_count != 0) {
1758 * We should only have one spinlock and as long as
1759 * the flags cannot match for this locks class,
1760 * check if the first spinlock is the one curthread
1763 lock1 = &lock_list->ll_children[lock_list->ll_count - 1];
1764 if (lock_list->ll_count == 1 && lock_list->ll_next == NULL &&
1765 lock1->li_lock == lock && n == 0)
1771 printf(" with the following %slocks held:\n",
1772 (flags & WARN_SLEEPOK) != 0 ? "non-sleepable " : "");
1773 n += witness_list_locks(&lock_list, printf);
1776 if (flags & WARN_PANIC && n)
1777 kassert_panic("%s", __func__);
1779 witness_debugger(n, __func__);
1784 witness_file(struct lock_object *lock)
1788 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1790 w = lock->lo_witness;
1795 witness_line(struct lock_object *lock)
1799 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1801 w = lock->lo_witness;
1805 static struct witness *
1806 enroll(const char *description, struct lock_class *lock_class)
1809 struct witness_list *typelist;
1811 MPASS(description != NULL);
1813 if (witness_watch == -1 || panicstr != NULL)
1815 if ((lock_class->lc_flags & LC_SPINLOCK)) {
1816 if (witness_skipspin)
1820 } else if ((lock_class->lc_flags & LC_SLEEPLOCK)) {
1821 typelist = &w_sleep;
1823 kassert_panic("lock class %s is not sleep or spin",
1824 lock_class->lc_name);
1828 mtx_lock_spin(&w_mtx);
1829 w = witness_hash_get(description);
1832 if ((w = witness_get()) == NULL)
1834 MPASS(strlen(description) < MAX_W_NAME);
1835 strcpy(w->w_name, description);
1836 w->w_class = lock_class;
1838 STAILQ_INSERT_HEAD(&w_all, w, w_list);
1839 if (lock_class->lc_flags & LC_SPINLOCK) {
1840 STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
1842 } else if (lock_class->lc_flags & LC_SLEEPLOCK) {
1843 STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
1847 /* Insert new witness into the hash */
1848 witness_hash_put(w);
1849 witness_increment_graph_generation();
1850 mtx_unlock_spin(&w_mtx);
1854 if (w->w_refcount == 1)
1855 w->w_class = lock_class;
1856 mtx_unlock_spin(&w_mtx);
1857 if (lock_class != w->w_class)
1859 "lock (%s) %s does not match earlier (%s) lock",
1860 description, lock_class->lc_name,
1861 w->w_class->lc_name);
1866 depart(struct witness *w)
1868 struct witness_list *list;
1870 MPASS(w->w_refcount == 0);
1871 if (w->w_class->lc_flags & LC_SLEEPLOCK) {
1879 * Set file to NULL as it may point into a loadable module.
1883 witness_increment_graph_generation();
1888 adopt(struct witness *parent, struct witness *child)
1892 if (witness_cold == 0)
1893 mtx_assert(&w_mtx, MA_OWNED);
1895 /* If the relationship is already known, there's no work to be done. */
1896 if (isitmychild(parent, child))
1899 /* When the structure of the graph changes, bump up the generation. */
1900 witness_increment_graph_generation();
1903 * The hard part ... create the direct relationship, then propagate all
1904 * indirect relationships.
1906 pi = parent->w_index;
1907 ci = child->w_index;
1908 WITNESS_INDEX_ASSERT(pi);
1909 WITNESS_INDEX_ASSERT(ci);
1911 w_rmatrix[pi][ci] |= WITNESS_PARENT;
1912 w_rmatrix[ci][pi] |= WITNESS_CHILD;
1915 * If parent was not already an ancestor of child,
1916 * then we increment the descendant and ancestor counters.
1918 if ((w_rmatrix[pi][ci] & WITNESS_ANCESTOR) == 0) {
1919 parent->w_num_descendants++;
1920 child->w_num_ancestors++;
1924 * Find each ancestor of 'pi'. Note that 'pi' itself is counted as
1925 * an ancestor of 'pi' during this loop.
1927 for (i = 1; i <= w_max_used_index; i++) {
1928 if ((w_rmatrix[i][pi] & WITNESS_ANCESTOR_MASK) == 0 &&
1932 /* Find each descendant of 'i' and mark it as a descendant. */
1933 for (j = 1; j <= w_max_used_index; j++) {
1936 * Skip children that are already marked as
1937 * descendants of 'i'.
1939 if (w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK)
1943 * We are only interested in descendants of 'ci'. Note
1944 * that 'ci' itself is counted as a descendant of 'ci'.
1946 if ((w_rmatrix[ci][j] & WITNESS_ANCESTOR_MASK) == 0 &&
1949 w_rmatrix[i][j] |= WITNESS_ANCESTOR;
1950 w_rmatrix[j][i] |= WITNESS_DESCENDANT;
1951 w_data[i].w_num_descendants++;
1952 w_data[j].w_num_ancestors++;
1955 * Make sure we aren't marking a node as both an
1956 * ancestor and descendant. We should have caught
1957 * this as a lock order reversal earlier.
1959 if ((w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK) &&
1960 (w_rmatrix[i][j] & WITNESS_DESCENDANT_MASK)) {
1961 printf("witness rmatrix paradox! [%d][%d]=%d "
1962 "both ancestor and descendant\n",
1963 i, j, w_rmatrix[i][j]);
1965 printf("Witness disabled.\n");
1968 if ((w_rmatrix[j][i] & WITNESS_ANCESTOR_MASK) &&
1969 (w_rmatrix[j][i] & WITNESS_DESCENDANT_MASK)) {
1970 printf("witness rmatrix paradox! [%d][%d]=%d "
1971 "both ancestor and descendant\n",
1972 j, i, w_rmatrix[j][i]);
1974 printf("Witness disabled.\n");
1982 itismychild(struct witness *parent, struct witness *child)
1986 MPASS(child != NULL && parent != NULL);
1987 if (witness_cold == 0)
1988 mtx_assert(&w_mtx, MA_OWNED);
1990 if (!witness_lock_type_equal(parent, child)) {
1991 if (witness_cold == 0) {
1993 mtx_unlock_spin(&w_mtx);
1998 "%s: parent \"%s\" (%s) and child \"%s\" (%s) are not "
1999 "the same lock type", __func__, parent->w_name,
2000 parent->w_class->lc_name, child->w_name,
2001 child->w_class->lc_name);
2003 mtx_lock_spin(&w_mtx);
2005 adopt(parent, child);
2009 * Generic code for the isitmy*() functions. The rmask parameter is the
2010 * expected relationship of w1 to w2.
2013 _isitmyx(struct witness *w1, struct witness *w2, int rmask, const char *fname)
2015 unsigned char r1, r2;
2020 WITNESS_INDEX_ASSERT(i1);
2021 WITNESS_INDEX_ASSERT(i2);
2022 r1 = w_rmatrix[i1][i2] & WITNESS_RELATED_MASK;
2023 r2 = w_rmatrix[i2][i1] & WITNESS_RELATED_MASK;
2025 /* The flags on one better be the inverse of the flags on the other */
2026 if (!((WITNESS_ATOD(r1) == r2 && WITNESS_DTOA(r2) == r1) ||
2027 (WITNESS_DTOA(r1) == r2 && WITNESS_ATOD(r2) == r1))) {
2028 /* Don't squawk if we're potentially racing with an update. */
2029 if (!mtx_owned(&w_mtx))
2031 printf("%s: rmatrix mismatch between %s (index %d) and %s "
2032 "(index %d): w_rmatrix[%d][%d] == %hhx but "
2033 "w_rmatrix[%d][%d] == %hhx\n",
2034 fname, w1->w_name, i1, w2->w_name, i2, i1, i2, r1,
2037 printf("Witness disabled.\n");
2040 return (r1 & rmask);
2044 * Checks if @child is a direct child of @parent.
2047 isitmychild(struct witness *parent, struct witness *child)
2050 return (_isitmyx(parent, child, WITNESS_PARENT, __func__));
2054 * Checks if @descendant is a direct or inderect descendant of @ancestor.
2057 isitmydescendant(struct witness *ancestor, struct witness *descendant)
2060 return (_isitmyx(ancestor, descendant, WITNESS_ANCESTOR_MASK,
2066 blessed(struct witness *w1, struct witness *w2)
2069 struct witness_blessed *b;
2071 for (i = 0; i < nitems(blessed_list); i++) {
2072 b = &blessed_list[i];
2073 if (strcmp(w1->w_name, b->b_lock1) == 0) {
2074 if (strcmp(w2->w_name, b->b_lock2) == 0)
2078 if (strcmp(w1->w_name, b->b_lock2) == 0)
2079 if (strcmp(w2->w_name, b->b_lock1) == 0)
2086 static struct witness *
2092 if (witness_cold == 0)
2093 mtx_assert(&w_mtx, MA_OWNED);
2095 if (witness_watch == -1) {
2096 mtx_unlock_spin(&w_mtx);
2099 if (STAILQ_EMPTY(&w_free)) {
2101 mtx_unlock_spin(&w_mtx);
2102 printf("WITNESS: unable to allocate a new witness object\n");
2105 w = STAILQ_FIRST(&w_free);
2106 STAILQ_REMOVE_HEAD(&w_free, w_list);
2109 MPASS(index > 0 && index == w_max_used_index+1 &&
2110 index < witness_count);
2111 bzero(w, sizeof(*w));
2113 if (index > w_max_used_index)
2114 w_max_used_index = index;
2119 witness_free(struct witness *w)
2122 STAILQ_INSERT_HEAD(&w_free, w, w_list);
2126 static struct lock_list_entry *
2127 witness_lock_list_get(void)
2129 struct lock_list_entry *lle;
2131 if (witness_watch == -1)
2133 mtx_lock_spin(&w_mtx);
2134 lle = w_lock_list_free;
2137 mtx_unlock_spin(&w_mtx);
2138 printf("%s: witness exhausted\n", __func__);
2141 w_lock_list_free = lle->ll_next;
2142 mtx_unlock_spin(&w_mtx);
2143 bzero(lle, sizeof(*lle));
2148 witness_lock_list_free(struct lock_list_entry *lle)
2151 mtx_lock_spin(&w_mtx);
2152 lle->ll_next = w_lock_list_free;
2153 w_lock_list_free = lle;
2154 mtx_unlock_spin(&w_mtx);
2157 static struct lock_instance *
2158 find_instance(struct lock_list_entry *list, const struct lock_object *lock)
2160 struct lock_list_entry *lle;
2161 struct lock_instance *instance;
2164 for (lle = list; lle != NULL; lle = lle->ll_next)
2165 for (i = lle->ll_count - 1; i >= 0; i--) {
2166 instance = &lle->ll_children[i];
2167 if (instance->li_lock == lock)
2174 witness_list_lock(struct lock_instance *instance,
2175 int (*prnt)(const char *fmt, ...))
2177 struct lock_object *lock;
2179 lock = instance->li_lock;
2180 prnt("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ?
2181 "exclusive" : "shared", LOCK_CLASS(lock)->lc_name, lock->lo_name);
2182 if (lock->lo_witness->w_name != lock->lo_name)
2183 prnt(" (%s)", lock->lo_witness->w_name);
2184 prnt(" r = %d (%p) locked @ %s:%d\n",
2185 instance->li_flags & LI_RECURSEMASK, lock,
2186 fixup_filename(instance->li_file), instance->li_line);
2190 witness_output(const char *fmt, ...)
2196 ret = witness_voutput(fmt, ap);
2202 witness_voutput(const char *fmt, va_list ap)
2207 switch (witness_channel) {
2208 case WITNESS_CONSOLE:
2209 ret = vprintf(fmt, ap);
2212 vlog(LOG_NOTICE, fmt, ap);
2222 witness_thread_has_locks(struct thread *td)
2225 if (td->td_sleeplocks == NULL)
2227 return (td->td_sleeplocks->ll_count != 0);
2231 witness_proc_has_locks(struct proc *p)
2235 FOREACH_THREAD_IN_PROC(p, td) {
2236 if (witness_thread_has_locks(td))
2244 witness_list_locks(struct lock_list_entry **lock_list,
2245 int (*prnt)(const char *fmt, ...))
2247 struct lock_list_entry *lle;
2251 for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
2252 for (i = lle->ll_count - 1; i >= 0; i--) {
2253 witness_list_lock(&lle->ll_children[i], prnt);
2260 * This is a bit risky at best. We call this function when we have timed
2261 * out acquiring a spin lock, and we assume that the other CPU is stuck
2262 * with this lock held. So, we go groveling around in the other CPU's
2263 * per-cpu data to try to find the lock instance for this spin lock to
2264 * see when it was last acquired.
2267 witness_display_spinlock(struct lock_object *lock, struct thread *owner,
2268 int (*prnt)(const char *fmt, ...))
2270 struct lock_instance *instance;
2273 if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU)
2275 pc = pcpu_find(owner->td_oncpu);
2276 instance = find_instance(pc->pc_spinlocks, lock);
2277 if (instance != NULL)
2278 witness_list_lock(instance, prnt);
2282 witness_save(struct lock_object *lock, const char **filep, int *linep)
2284 struct lock_list_entry *lock_list;
2285 struct lock_instance *instance;
2286 struct lock_class *class;
2289 * This function is used independently in locking code to deal with
2290 * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2293 if (SCHEDULER_STOPPED())
2295 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2296 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2298 class = LOCK_CLASS(lock);
2299 if (class->lc_flags & LC_SLEEPLOCK)
2300 lock_list = curthread->td_sleeplocks;
2302 if (witness_skipspin)
2304 lock_list = PCPU_GET(spinlocks);
2306 instance = find_instance(lock_list, lock);
2307 if (instance == NULL) {
2308 kassert_panic("%s: lock (%s) %s not locked", __func__,
2309 class->lc_name, lock->lo_name);
2312 *filep = instance->li_file;
2313 *linep = instance->li_line;
2317 witness_restore(struct lock_object *lock, const char *file, int line)
2319 struct lock_list_entry *lock_list;
2320 struct lock_instance *instance;
2321 struct lock_class *class;
2324 * This function is used independently in locking code to deal with
2325 * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2328 if (SCHEDULER_STOPPED())
2330 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2331 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2333 class = LOCK_CLASS(lock);
2334 if (class->lc_flags & LC_SLEEPLOCK)
2335 lock_list = curthread->td_sleeplocks;
2337 if (witness_skipspin)
2339 lock_list = PCPU_GET(spinlocks);
2341 instance = find_instance(lock_list, lock);
2342 if (instance == NULL)
2343 kassert_panic("%s: lock (%s) %s not locked", __func__,
2344 class->lc_name, lock->lo_name);
2345 lock->lo_witness->w_file = file;
2346 lock->lo_witness->w_line = line;
2347 if (instance == NULL)
2349 instance->li_file = file;
2350 instance->li_line = line;
2354 witness_assert(const struct lock_object *lock, int flags, const char *file,
2357 #ifdef INVARIANT_SUPPORT
2358 struct lock_instance *instance;
2359 struct lock_class *class;
2361 if (lock->lo_witness == NULL || witness_watch < 1 || panicstr != NULL)
2363 class = LOCK_CLASS(lock);
2364 if ((class->lc_flags & LC_SLEEPLOCK) != 0)
2365 instance = find_instance(curthread->td_sleeplocks, lock);
2366 else if ((class->lc_flags & LC_SPINLOCK) != 0)
2367 instance = find_instance(PCPU_GET(spinlocks), lock);
2369 kassert_panic("Lock (%s) %s is not sleep or spin!",
2370 class->lc_name, lock->lo_name);
2375 if (instance != NULL)
2376 kassert_panic("Lock (%s) %s locked @ %s:%d.",
2377 class->lc_name, lock->lo_name,
2378 fixup_filename(file), line);
2381 case LA_LOCKED | LA_RECURSED:
2382 case LA_LOCKED | LA_NOTRECURSED:
2384 case LA_SLOCKED | LA_RECURSED:
2385 case LA_SLOCKED | LA_NOTRECURSED:
2387 case LA_XLOCKED | LA_RECURSED:
2388 case LA_XLOCKED | LA_NOTRECURSED:
2389 if (instance == NULL) {
2390 kassert_panic("Lock (%s) %s not locked @ %s:%d.",
2391 class->lc_name, lock->lo_name,
2392 fixup_filename(file), line);
2395 if ((flags & LA_XLOCKED) != 0 &&
2396 (instance->li_flags & LI_EXCLUSIVE) == 0)
2398 "Lock (%s) %s not exclusively locked @ %s:%d.",
2399 class->lc_name, lock->lo_name,
2400 fixup_filename(file), line);
2401 if ((flags & LA_SLOCKED) != 0 &&
2402 (instance->li_flags & LI_EXCLUSIVE) != 0)
2404 "Lock (%s) %s exclusively locked @ %s:%d.",
2405 class->lc_name, lock->lo_name,
2406 fixup_filename(file), line);
2407 if ((flags & LA_RECURSED) != 0 &&
2408 (instance->li_flags & LI_RECURSEMASK) == 0)
2409 kassert_panic("Lock (%s) %s not recursed @ %s:%d.",
2410 class->lc_name, lock->lo_name,
2411 fixup_filename(file), line);
2412 if ((flags & LA_NOTRECURSED) != 0 &&
2413 (instance->li_flags & LI_RECURSEMASK) != 0)
2414 kassert_panic("Lock (%s) %s recursed @ %s:%d.",
2415 class->lc_name, lock->lo_name,
2416 fixup_filename(file), line);
2419 kassert_panic("Invalid lock assertion at %s:%d.",
2420 fixup_filename(file), line);
2423 #endif /* INVARIANT_SUPPORT */
2427 witness_setflag(struct lock_object *lock, int flag, int set)
2429 struct lock_list_entry *lock_list;
2430 struct lock_instance *instance;
2431 struct lock_class *class;
2433 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2435 class = LOCK_CLASS(lock);
2436 if (class->lc_flags & LC_SLEEPLOCK)
2437 lock_list = curthread->td_sleeplocks;
2439 if (witness_skipspin)
2441 lock_list = PCPU_GET(spinlocks);
2443 instance = find_instance(lock_list, lock);
2444 if (instance == NULL) {
2445 kassert_panic("%s: lock (%s) %s not locked", __func__,
2446 class->lc_name, lock->lo_name);
2451 instance->li_flags |= flag;
2453 instance->li_flags &= ~flag;
2457 witness_norelease(struct lock_object *lock)
2460 witness_setflag(lock, LI_NORELEASE, 1);
2464 witness_releaseok(struct lock_object *lock)
2467 witness_setflag(lock, LI_NORELEASE, 0);
2472 witness_ddb_list(struct thread *td)
2475 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2476 KASSERT(kdb_active, ("%s: not in the debugger", __func__));
2478 if (witness_watch < 1)
2481 witness_list_locks(&td->td_sleeplocks, db_printf);
2484 * We only handle spinlocks if td == curthread. This is somewhat broken
2485 * if td is currently executing on some other CPU and holds spin locks
2486 * as we won't display those locks. If we had a MI way of getting
2487 * the per-cpu data for a given cpu then we could use
2488 * td->td_oncpu to get the list of spinlocks for this thread
2491 * That still wouldn't really fix this unless we locked the scheduler
2492 * lock or stopped the other CPU to make sure it wasn't changing the
2493 * list out from under us. It is probably best to just not try to
2494 * handle threads on other CPU's for now.
2496 if (td == curthread && PCPU_GET(spinlocks) != NULL)
2497 witness_list_locks(PCPU_PTR(spinlocks), db_printf);
2500 DB_SHOW_COMMAND(locks, db_witness_list)
2505 td = db_lookup_thread(addr, true);
2508 witness_ddb_list(td);
2511 DB_SHOW_ALL_COMMAND(locks, db_witness_list_all)
2517 * It would be nice to list only threads and processes that actually
2518 * held sleep locks, but that information is currently not exported
2521 FOREACH_PROC_IN_SYSTEM(p) {
2522 if (!witness_proc_has_locks(p))
2524 FOREACH_THREAD_IN_PROC(p, td) {
2525 if (!witness_thread_has_locks(td))
2527 db_printf("Process %d (%s) thread %p (%d)\n", p->p_pid,
2528 p->p_comm, td, td->td_tid);
2529 witness_ddb_list(td);
2535 DB_SHOW_ALIAS(alllocks, db_witness_list_all)
2537 DB_SHOW_COMMAND(witness, db_witness_display)
2540 witness_ddb_display(db_printf);
2545 sbuf_print_witness_badstacks(struct sbuf *sb, size_t *oldidx)
2547 struct witness_lock_order_data *data1, *data2, *tmp_data1, *tmp_data2;
2548 struct witness *tmp_w1, *tmp_w2, *w1, *w2;
2549 u_int w_rmatrix1, w_rmatrix2;
2550 int generation, i, j;
2557 /* Allocate and init temporary storage space. */
2558 tmp_w1 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2559 tmp_w2 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2560 tmp_data1 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2562 tmp_data2 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2564 stack_zero(&tmp_data1->wlod_stack);
2565 stack_zero(&tmp_data2->wlod_stack);
2568 mtx_lock_spin(&w_mtx);
2569 generation = w_generation;
2570 mtx_unlock_spin(&w_mtx);
2571 sbuf_printf(sb, "Number of known direct relationships is %d\n",
2572 w_lohash.wloh_count);
2573 for (i = 1; i < w_max_used_index; i++) {
2574 mtx_lock_spin(&w_mtx);
2575 if (generation != w_generation) {
2576 mtx_unlock_spin(&w_mtx);
2578 /* The graph has changed, try again. */
2585 if (w1->w_reversed == 0) {
2586 mtx_unlock_spin(&w_mtx);
2590 /* Copy w1 locally so we can release the spin lock. */
2592 mtx_unlock_spin(&w_mtx);
2594 if (tmp_w1->w_reversed == 0)
2596 for (j = 1; j < w_max_used_index; j++) {
2597 if ((w_rmatrix[i][j] & WITNESS_REVERSAL) == 0 || i > j)
2600 mtx_lock_spin(&w_mtx);
2601 if (generation != w_generation) {
2602 mtx_unlock_spin(&w_mtx);
2604 /* The graph has changed, try again. */
2611 data1 = witness_lock_order_get(w1, w2);
2612 data2 = witness_lock_order_get(w2, w1);
2615 * Copy information locally so we can release the
2619 w_rmatrix1 = (unsigned int)w_rmatrix[i][j];
2620 w_rmatrix2 = (unsigned int)w_rmatrix[j][i];
2623 stack_zero(&tmp_data1->wlod_stack);
2624 stack_copy(&data1->wlod_stack,
2625 &tmp_data1->wlod_stack);
2627 if (data2 && data2 != data1) {
2628 stack_zero(&tmp_data2->wlod_stack);
2629 stack_copy(&data2->wlod_stack,
2630 &tmp_data2->wlod_stack);
2632 mtx_unlock_spin(&w_mtx);
2635 "\nLock order reversal between \"%s\"(%s) and \"%s\"(%s)!\n",
2636 tmp_w1->w_name, tmp_w1->w_class->lc_name,
2637 tmp_w2->w_name, tmp_w2->w_class->lc_name);
2640 "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2641 tmp_w1->w_name, tmp_w1->w_class->lc_name,
2642 tmp_w2->w_name, tmp_w2->w_class->lc_name);
2643 stack_sbuf_print(sb, &tmp_data1->wlod_stack);
2644 sbuf_printf(sb, "\n");
2646 if (data2 && data2 != data1) {
2648 "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2649 tmp_w2->w_name, tmp_w2->w_class->lc_name,
2650 tmp_w1->w_name, tmp_w1->w_class->lc_name);
2651 stack_sbuf_print(sb, &tmp_data2->wlod_stack);
2652 sbuf_printf(sb, "\n");
2656 mtx_lock_spin(&w_mtx);
2657 if (generation != w_generation) {
2658 mtx_unlock_spin(&w_mtx);
2661 * The graph changed while we were printing stack data,
2668 mtx_unlock_spin(&w_mtx);
2670 /* Free temporary storage space. */
2671 free(tmp_data1, M_TEMP);
2672 free(tmp_data2, M_TEMP);
2673 free(tmp_w1, M_TEMP);
2674 free(tmp_w2, M_TEMP);
2678 sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS)
2683 if (witness_watch < 1) {
2684 error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2688 error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2692 sb = sbuf_new(NULL, NULL, badstack_sbuf_size, SBUF_AUTOEXTEND);
2696 sbuf_print_witness_badstacks(sb, &req->oldidx);
2699 error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1);
2707 sbuf_db_printf_drain(void *arg __unused, const char *data, int len)
2710 return (db_printf("%.*s", len, data));
2713 DB_SHOW_COMMAND(badstacks, db_witness_badstacks)
2719 sbuf_new(&sb, buffer, sizeof(buffer), SBUF_FIXEDLEN);
2720 sbuf_set_drain(&sb, sbuf_db_printf_drain, NULL);
2721 sbuf_print_witness_badstacks(&sb, &dummy);
2727 sysctl_debug_witness_channel(SYSCTL_HANDLER_ARGS)
2729 static const struct {
2730 enum witness_channel channel;
2733 { WITNESS_CONSOLE, "console" },
2734 { WITNESS_LOG, "log" },
2735 { WITNESS_NONE, "none" },
2742 for (i = 0; i < nitems(channels); i++)
2743 if (witness_channel == channels[i].channel) {
2744 snprintf(buf, sizeof(buf), "%s", channels[i].name);
2748 error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
2749 if (error != 0 || req->newptr == NULL)
2753 for (i = 0; i < nitems(channels); i++)
2754 if (strcmp(channels[i].name, buf) == 0) {
2755 witness_channel = channels[i].channel;
2763 sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS)
2769 if (witness_watch < 1) {
2770 error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2774 error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2779 error = sysctl_wire_old_buffer(req, 0);
2782 sb = sbuf_new_for_sysctl(NULL, NULL, FULLGRAPH_SBUF_SIZE, req);
2785 sbuf_printf(sb, "\n");
2787 mtx_lock_spin(&w_mtx);
2788 STAILQ_FOREACH(w, &w_all, w_list)
2790 STAILQ_FOREACH(w, &w_all, w_list)
2791 witness_add_fullgraph(sb, w);
2792 mtx_unlock_spin(&w_mtx);
2795 * Close the sbuf and return to userland.
2797 error = sbuf_finish(sb);
2804 sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS)
2808 value = witness_watch;
2809 error = sysctl_handle_int(oidp, &value, 0, req);
2810 if (error != 0 || req->newptr == NULL)
2812 if (value > 1 || value < -1 ||
2813 (witness_watch == -1 && value != witness_watch))
2815 witness_watch = value;
2820 witness_add_fullgraph(struct sbuf *sb, struct witness *w)
2824 if (w->w_displayed != 0 || (w->w_file == NULL && w->w_line == 0))
2828 WITNESS_INDEX_ASSERT(w->w_index);
2829 for (i = 1; i <= w_max_used_index; i++) {
2830 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT) {
2831 sbuf_printf(sb, "\"%s\",\"%s\"\n", w->w_name,
2833 witness_add_fullgraph(sb, &w_data[i]);
2839 * A simple hash function. Takes a key pointer and a key size. If size == 0,
2840 * interprets the key as a string and reads until the null
2841 * terminator. Otherwise, reads the first size bytes. Returns an unsigned 32-bit
2842 * hash value computed from the key.
2845 witness_hash_djb2(const uint8_t *key, uint32_t size)
2847 unsigned int hash = 5381;
2850 /* hash = hash * 33 + key[i] */
2852 for (i = 0; i < size; i++)
2853 hash = ((hash << 5) + hash) + (unsigned int)key[i];
2855 for (i = 0; key[i] != 0; i++)
2856 hash = ((hash << 5) + hash) + (unsigned int)key[i];
2863 * Initializes the two witness hash tables. Called exactly once from
2864 * witness_initialize().
2867 witness_init_hash_tables(void)
2871 MPASS(witness_cold);
2873 /* Initialize the hash tables. */
2874 for (i = 0; i < WITNESS_HASH_SIZE; i++)
2875 w_hash.wh_array[i] = NULL;
2877 w_hash.wh_size = WITNESS_HASH_SIZE;
2878 w_hash.wh_count = 0;
2880 /* Initialize the lock order data hash. */
2882 for (i = 0; i < WITNESS_LO_DATA_COUNT; i++) {
2883 memset(&w_lodata[i], 0, sizeof(w_lodata[i]));
2884 w_lodata[i].wlod_next = w_lofree;
2885 w_lofree = &w_lodata[i];
2887 w_lohash.wloh_size = WITNESS_LO_HASH_SIZE;
2888 w_lohash.wloh_count = 0;
2889 for (i = 0; i < WITNESS_LO_HASH_SIZE; i++)
2890 w_lohash.wloh_array[i] = NULL;
2893 static struct witness *
2894 witness_hash_get(const char *key)
2900 if (witness_cold == 0)
2901 mtx_assert(&w_mtx, MA_OWNED);
2902 hash = witness_hash_djb2(key, 0) % w_hash.wh_size;
2903 w = w_hash.wh_array[hash];
2905 if (strcmp(w->w_name, key) == 0)
2915 witness_hash_put(struct witness *w)
2920 MPASS(w->w_name != NULL);
2921 if (witness_cold == 0)
2922 mtx_assert(&w_mtx, MA_OWNED);
2923 KASSERT(witness_hash_get(w->w_name) == NULL,
2924 ("%s: trying to add a hash entry that already exists!", __func__));
2925 KASSERT(w->w_hash_next == NULL,
2926 ("%s: w->w_hash_next != NULL", __func__));
2928 hash = witness_hash_djb2(w->w_name, 0) % w_hash.wh_size;
2929 w->w_hash_next = w_hash.wh_array[hash];
2930 w_hash.wh_array[hash] = w;
2935 static struct witness_lock_order_data *
2936 witness_lock_order_get(struct witness *parent, struct witness *child)
2938 struct witness_lock_order_data *data = NULL;
2939 struct witness_lock_order_key key;
2942 MPASS(parent != NULL && child != NULL);
2943 key.from = parent->w_index;
2944 key.to = child->w_index;
2945 WITNESS_INDEX_ASSERT(key.from);
2946 WITNESS_INDEX_ASSERT(key.to);
2947 if ((w_rmatrix[parent->w_index][child->w_index]
2948 & WITNESS_LOCK_ORDER_KNOWN) == 0)
2951 hash = witness_hash_djb2((const char*)&key,
2952 sizeof(key)) % w_lohash.wloh_size;
2953 data = w_lohash.wloh_array[hash];
2954 while (data != NULL) {
2955 if (witness_lock_order_key_equal(&data->wlod_key, &key))
2957 data = data->wlod_next;
2965 * Verify that parent and child have a known relationship, are not the same,
2966 * and child is actually a child of parent. This is done without w_mtx
2967 * to avoid contention in the common case.
2970 witness_lock_order_check(struct witness *parent, struct witness *child)
2973 if (parent != child &&
2974 w_rmatrix[parent->w_index][child->w_index]
2975 & WITNESS_LOCK_ORDER_KNOWN &&
2976 isitmychild(parent, child))
2983 witness_lock_order_add(struct witness *parent, struct witness *child)
2985 struct witness_lock_order_data *data = NULL;
2986 struct witness_lock_order_key key;
2989 MPASS(parent != NULL && child != NULL);
2990 key.from = parent->w_index;
2991 key.to = child->w_index;
2992 WITNESS_INDEX_ASSERT(key.from);
2993 WITNESS_INDEX_ASSERT(key.to);
2994 if (w_rmatrix[parent->w_index][child->w_index]
2995 & WITNESS_LOCK_ORDER_KNOWN)
2998 hash = witness_hash_djb2((const char*)&key,
2999 sizeof(key)) % w_lohash.wloh_size;
3000 w_rmatrix[parent->w_index][child->w_index] |= WITNESS_LOCK_ORDER_KNOWN;
3004 w_lofree = data->wlod_next;
3005 data->wlod_next = w_lohash.wloh_array[hash];
3006 data->wlod_key = key;
3007 w_lohash.wloh_array[hash] = data;
3008 w_lohash.wloh_count++;
3009 stack_zero(&data->wlod_stack);
3010 stack_save(&data->wlod_stack);
3014 /* Call this whenever the structure of the witness graph changes. */
3016 witness_increment_graph_generation(void)
3019 if (witness_cold == 0)
3020 mtx_assert(&w_mtx, MA_OWNED);
3025 witness_output_drain(void *arg __unused, const char *data, int len)
3028 witness_output("%.*s", len, data);
3033 witness_debugger(int cond, const char *msg)
3042 if (witness_trace) {
3043 sbuf_new(&sb, buf, sizeof(buf), SBUF_FIXEDLEN);
3044 sbuf_set_drain(&sb, witness_output_drain, NULL);
3048 witness_output("stack backtrace:\n");
3049 stack_sbuf_print_ddb(&sb, &st);
3056 kdb_enter(KDB_WHY_WITNESS, msg);