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 #ifndef WITNESS_COUNT
136 #define WITNESS_COUNT 1536
138 #define WITNESS_HASH_SIZE 251 /* Prime, gives load factor < 2 */
139 #define WITNESS_PENDLIST (512 + (MAXCPU * 4))
141 /* Allocate 256 KB of stack data space */
142 #define WITNESS_LO_DATA_COUNT 2048
144 /* Prime, gives load factor of ~2 at full load */
145 #define WITNESS_LO_HASH_SIZE 1021
148 * XXX: This is somewhat bogus, as we assume here that at most 2048 threads
149 * will hold LOCK_NCHILDREN locks. We handle failure ok, and we should
150 * probably be safe for the most part, but it's still a SWAG.
152 #define LOCK_NCHILDREN 5
153 #define LOCK_CHILDCOUNT 2048
155 #define MAX_W_NAME 64
157 #define FULLGRAPH_SBUF_SIZE 512
160 * These flags go in the witness relationship matrix and describe the
161 * relationship between any two struct witness objects.
163 #define WITNESS_UNRELATED 0x00 /* No lock order relation. */
164 #define WITNESS_PARENT 0x01 /* Parent, aka direct ancestor. */
165 #define WITNESS_ANCESTOR 0x02 /* Direct or indirect ancestor. */
166 #define WITNESS_CHILD 0x04 /* Child, aka direct descendant. */
167 #define WITNESS_DESCENDANT 0x08 /* Direct or indirect descendant. */
168 #define WITNESS_ANCESTOR_MASK (WITNESS_PARENT | WITNESS_ANCESTOR)
169 #define WITNESS_DESCENDANT_MASK (WITNESS_CHILD | WITNESS_DESCENDANT)
170 #define WITNESS_RELATED_MASK \
171 (WITNESS_ANCESTOR_MASK | WITNESS_DESCENDANT_MASK)
172 #define WITNESS_REVERSAL 0x10 /* A lock order reversal has been
174 #define WITNESS_RESERVED1 0x20 /* Unused flag, reserved. */
175 #define WITNESS_RESERVED2 0x40 /* Unused flag, reserved. */
176 #define WITNESS_LOCK_ORDER_KNOWN 0x80 /* This lock order is known. */
178 /* Descendant to ancestor flags */
179 #define WITNESS_DTOA(x) (((x) & WITNESS_RELATED_MASK) >> 2)
181 /* Ancestor to descendant flags */
182 #define WITNESS_ATOD(x) (((x) & WITNESS_RELATED_MASK) << 2)
184 #define WITNESS_INDEX_ASSERT(i) \
185 MPASS((i) > 0 && (i) <= w_max_used_index && (i) < witness_count)
187 static MALLOC_DEFINE(M_WITNESS, "Witness", "Witness");
190 * Lock instances. A lock instance is the data associated with a lock while
191 * it is held by witness. For example, a lock instance will hold the
192 * recursion count of a lock. Lock instances are held in lists. Spin locks
193 * are held in a per-cpu list while sleep locks are held in per-thread list.
195 struct lock_instance {
196 struct lock_object *li_lock;
203 * A simple list type used to build the list of locks held by a thread
204 * or CPU. We can't simply embed the list in struct lock_object since a
205 * lock may be held by more than one thread if it is a shared lock. Locks
206 * are added to the head of the list, so we fill up each list entry from
207 * "the back" logically. To ease some of the arithmetic, we actually fill
208 * in each list entry the normal way (children[0] then children[1], etc.) but
209 * when we traverse the list we read children[count-1] as the first entry
210 * down to children[0] as the final entry.
212 struct lock_list_entry {
213 struct lock_list_entry *ll_next;
214 struct lock_instance ll_children[LOCK_NCHILDREN];
219 * The main witness structure. One of these per named lock type in the system
220 * (for example, "vnode interlock").
223 char w_name[MAX_W_NAME];
224 uint32_t w_index; /* Index in the relationship matrix */
225 struct lock_class *w_class;
226 STAILQ_ENTRY(witness) w_list; /* List of all witnesses. */
227 STAILQ_ENTRY(witness) w_typelist; /* Witnesses of a type. */
228 struct witness *w_hash_next; /* Linked list in hash buckets. */
229 const char *w_file; /* File where last acquired */
230 uint32_t w_line; /* Line where last acquired */
232 uint16_t w_num_ancestors; /* direct/indirect
234 uint16_t w_num_descendants; /* direct/indirect
235 * descendant count */
237 unsigned w_displayed:1;
238 unsigned w_reversed:1;
241 STAILQ_HEAD(witness_list, witness);
244 * The witness hash table. Keys are witness names (const char *), elements are
245 * witness objects (struct witness *).
247 struct witness_hash {
248 struct witness *wh_array[WITNESS_HASH_SIZE];
254 * Key type for the lock order data hash table.
256 struct witness_lock_order_key {
261 struct witness_lock_order_data {
262 struct stack wlod_stack;
263 struct witness_lock_order_key wlod_key;
264 struct witness_lock_order_data *wlod_next;
268 * The witness lock order data hash table. Keys are witness index tuples
269 * (struct witness_lock_order_key), elements are lock order data objects
270 * (struct witness_lock_order_data).
272 struct witness_lock_order_hash {
273 struct witness_lock_order_data *wloh_array[WITNESS_LO_HASH_SIZE];
278 struct witness_blessed {
283 struct witness_pendhelp {
285 struct lock_object *wh_lock;
288 struct witness_order_list_entry {
290 struct lock_class *w_class;
294 * Returns 0 if one of the locks is a spin lock and the other is not.
295 * Returns 1 otherwise.
298 witness_lock_type_equal(struct witness *w1, struct witness *w2)
301 return ((w1->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) ==
302 (w2->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)));
306 witness_lock_order_key_equal(const struct witness_lock_order_key *a,
307 const struct witness_lock_order_key *b)
310 return (a->from == b->from && a->to == b->to);
313 static int _isitmyx(struct witness *w1, struct witness *w2, int rmask,
315 static void adopt(struct witness *parent, struct witness *child);
316 static int blessed(struct witness *, struct witness *);
317 static void depart(struct witness *w);
318 static struct witness *enroll(const char *description,
319 struct lock_class *lock_class);
320 static struct lock_instance *find_instance(struct lock_list_entry *list,
321 const struct lock_object *lock);
322 static int isitmychild(struct witness *parent, struct witness *child);
323 static int isitmydescendant(struct witness *parent, struct witness *child);
324 static void itismychild(struct witness *parent, struct witness *child);
325 static int sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS);
326 static int sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS);
327 static int sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS);
328 static int sysctl_debug_witness_channel(SYSCTL_HANDLER_ARGS);
329 static void witness_add_fullgraph(struct sbuf *sb, struct witness *parent);
331 static void witness_ddb_compute_levels(void);
332 static void witness_ddb_display(int(*)(const char *fmt, ...));
333 static void witness_ddb_display_descendants(int(*)(const char *fmt, ...),
334 struct witness *, int indent);
335 static void witness_ddb_display_list(int(*prnt)(const char *fmt, ...),
336 struct witness_list *list);
337 static void witness_ddb_level_descendants(struct witness *parent, int l);
338 static void witness_ddb_list(struct thread *td);
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_voutput(const char *fmt, va_list ap) __printflike(1, 0);
361 static void witness_setflag(struct lock_object *lock, int flag, int set);
363 static SYSCTL_NODE(_debug, OID_AUTO, witness, CTLFLAG_RW, NULL,
367 * If set to 0, lock order checking is disabled. If set to -1,
368 * witness is completely disabled. Otherwise witness performs full
369 * lock order checking for all locks. At runtime, lock order checking
370 * may be toggled. However, witness cannot be reenabled once it is
371 * completely disabled.
373 static int witness_watch = 1;
374 SYSCTL_PROC(_debug_witness, OID_AUTO, watch, CTLFLAG_RWTUN | CTLTYPE_INT, NULL, 0,
375 sysctl_debug_witness_watch, "I", "witness is watching lock operations");
379 * When KDB is enabled and witness_kdb is 1, it will cause the system
380 * to drop into kdebug() when:
381 * - a lock hierarchy violation occurs
382 * - locks are held when going to sleep.
389 SYSCTL_INT(_debug_witness, OID_AUTO, kdb, CTLFLAG_RWTUN, &witness_kdb, 0, "");
392 #if defined(DDB) || defined(KDB)
394 * When DDB or KDB is enabled and witness_trace is 1, it will cause the system
395 * to print a stack trace:
396 * - a lock hierarchy violation occurs
397 * - locks are held when going to sleep.
399 int witness_trace = 1;
400 SYSCTL_INT(_debug_witness, OID_AUTO, trace, CTLFLAG_RWTUN, &witness_trace, 0, "");
401 #endif /* DDB || KDB */
403 #ifdef WITNESS_SKIPSPIN
404 int witness_skipspin = 1;
406 int witness_skipspin = 0;
408 SYSCTL_INT(_debug_witness, OID_AUTO, skipspin, CTLFLAG_RDTUN, &witness_skipspin, 0, "");
410 int badstack_sbuf_size;
412 int witness_count = WITNESS_COUNT;
413 SYSCTL_INT(_debug_witness, OID_AUTO, witness_count, CTLFLAG_RDTUN,
414 &witness_count, 0, "");
417 * Output channel for witness messages. By default we print to the console.
419 enum witness_channel {
425 static enum witness_channel witness_channel = WITNESS_CONSOLE;
426 SYSCTL_PROC(_debug_witness, OID_AUTO, output_channel, CTLTYPE_STRING |
427 CTLFLAG_RWTUN, NULL, 0, sysctl_debug_witness_channel, "A",
428 "Output channel for warnings");
431 * Call this to print out the relations between locks.
433 SYSCTL_PROC(_debug_witness, OID_AUTO, fullgraph, CTLTYPE_STRING | CTLFLAG_RD,
434 NULL, 0, sysctl_debug_witness_fullgraph, "A", "Show locks relation graphs");
437 * Call this to print out the witness faulty stacks.
439 SYSCTL_PROC(_debug_witness, OID_AUTO, badstacks, CTLTYPE_STRING | CTLFLAG_RD,
440 NULL, 0, sysctl_debug_witness_badstacks, "A", "Show bad witness stacks");
442 static struct mtx w_mtx;
445 static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free);
446 static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all);
449 static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin);
450 static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep);
453 static struct lock_list_entry *w_lock_list_free = NULL;
454 static struct witness_pendhelp pending_locks[WITNESS_PENDLIST];
455 static u_int pending_cnt;
457 static int w_free_cnt, w_spin_cnt, w_sleep_cnt;
458 SYSCTL_INT(_debug_witness, OID_AUTO, free_cnt, CTLFLAG_RD, &w_free_cnt, 0, "");
459 SYSCTL_INT(_debug_witness, OID_AUTO, spin_cnt, CTLFLAG_RD, &w_spin_cnt, 0, "");
460 SYSCTL_INT(_debug_witness, OID_AUTO, sleep_cnt, CTLFLAG_RD, &w_sleep_cnt, 0,
463 static struct witness *w_data;
464 static uint8_t **w_rmatrix;
465 static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT];
466 static struct witness_hash w_hash; /* The witness hash table. */
468 /* The lock order data hash */
469 static struct witness_lock_order_data w_lodata[WITNESS_LO_DATA_COUNT];
470 static struct witness_lock_order_data *w_lofree = NULL;
471 static struct witness_lock_order_hash w_lohash;
472 static int w_max_used_index = 0;
473 static unsigned int w_generation = 0;
474 static const char w_notrunning[] = "Witness not running\n";
475 static const char w_stillcold[] = "Witness is still cold\n";
477 static const char w_notallowed[] = "The sysctl is disabled on the arch\n";
480 static struct witness_order_list_entry order_lists[] = {
484 { "proctree", &lock_class_sx },
485 { "allproc", &lock_class_sx },
486 { "allprison", &lock_class_sx },
491 { "Giant", &lock_class_mtx_sleep },
492 { "pipe mutex", &lock_class_mtx_sleep },
493 { "sigio lock", &lock_class_mtx_sleep },
494 { "process group", &lock_class_mtx_sleep },
496 { "pmc-sleep", &lock_class_mtx_sleep },
498 { "process lock", &lock_class_mtx_sleep },
499 { "session", &lock_class_mtx_sleep },
500 { "uidinfo hash", &lock_class_rw },
501 { "time lock", &lock_class_mtx_sleep },
506 { "umtx lock", &lock_class_mtx_sleep },
511 { "accept", &lock_class_mtx_sleep },
512 { "so_snd", &lock_class_mtx_sleep },
513 { "so_rcv", &lock_class_mtx_sleep },
514 { "sellck", &lock_class_mtx_sleep },
519 { "so_rcv", &lock_class_mtx_sleep },
520 { "radix node head", &lock_class_rm },
521 { "rtentry", &lock_class_mtx_sleep },
522 { "ifaddr", &lock_class_mtx_sleep },
526 * protocol locks before interface locks, after UDP locks.
528 { "in_multi_sx", &lock_class_sx },
529 { "udpinp", &lock_class_rw },
530 { "in_multi_list_mtx", &lock_class_mtx_sleep },
531 { "igmp_mtx", &lock_class_mtx_sleep },
532 { "ifnet_rw", &lock_class_rw },
533 { "if_addr_lock", &lock_class_mtx_sleep },
537 * protocol locks before interface locks, after UDP locks.
539 { "in6_multi_sx", &lock_class_sx },
540 { "udpinp", &lock_class_rw },
541 { "in6_multi_list_mtx", &lock_class_mtx_sleep },
542 { "mld_mtx", &lock_class_mtx_sleep },
543 { "ifnet_rw", &lock_class_rw },
544 { "if_addr_lock", &lock_class_mtx_sleep },
547 * UNIX Domain Sockets
549 { "unp_link_rwlock", &lock_class_rw },
550 { "unp_list_lock", &lock_class_mtx_sleep },
551 { "unp", &lock_class_mtx_sleep },
552 { "so_snd", &lock_class_mtx_sleep },
557 { "udp", &lock_class_mtx_sleep },
558 { "udpinp", &lock_class_rw },
559 { "so_snd", &lock_class_mtx_sleep },
564 { "tcp", &lock_class_mtx_sleep },
565 { "tcpinp", &lock_class_rw },
566 { "so_snd", &lock_class_mtx_sleep },
571 { "bpf global lock", &lock_class_sx },
572 { "bpf cdev lock", &lock_class_mtx_sleep },
577 { "nfsd_mtx", &lock_class_mtx_sleep },
578 { "so_snd", &lock_class_mtx_sleep },
584 { "802.11 com lock", &lock_class_mtx_sleep},
589 { "network driver", &lock_class_mtx_sleep},
595 { "ng_node", &lock_class_mtx_sleep },
596 { "ng_worklist", &lock_class_mtx_sleep },
601 { "vm map (system)", &lock_class_mtx_sleep },
602 { "vnode interlock", &lock_class_mtx_sleep },
603 { "cdev", &lock_class_mtx_sleep },
608 { "vm map (user)", &lock_class_sx },
609 { "vm object", &lock_class_rw },
610 { "vm page", &lock_class_mtx_sleep },
611 { "pmap pv global", &lock_class_rw },
612 { "pmap", &lock_class_mtx_sleep },
613 { "pmap pv list", &lock_class_rw },
614 { "vm page free queue", &lock_class_mtx_sleep },
615 { "vm pagequeue", &lock_class_mtx_sleep },
618 * kqueue/VFS interaction
620 { "kqueue", &lock_class_mtx_sleep },
621 { "struct mount mtx", &lock_class_mtx_sleep },
622 { "vnode interlock", &lock_class_mtx_sleep },
627 { "ncvn", &lock_class_mtx_sleep },
628 { "ncbuc", &lock_class_rw },
629 { "vnode interlock", &lock_class_mtx_sleep },
630 { "ncneg", &lock_class_mtx_sleep },
635 { "dn->dn_mtx", &lock_class_sx },
636 { "dr->dt.di.dr_mtx", &lock_class_sx },
637 { "db->db_mtx", &lock_class_sx },
642 { "TCP ID tree", &lock_class_rw },
643 { "tcp log id bucket", &lock_class_mtx_sleep },
644 { "tcpinp", &lock_class_rw },
645 { "TCP log expireq", &lock_class_mtx_sleep },
651 { "ap boot", &lock_class_mtx_spin },
653 { "rm.mutex_mtx", &lock_class_mtx_spin },
654 { "sio", &lock_class_mtx_spin },
656 { "cy", &lock_class_mtx_spin },
659 { "pcib_mtx", &lock_class_mtx_spin },
660 { "rtc_mtx", &lock_class_mtx_spin },
662 { "scc_hwmtx", &lock_class_mtx_spin },
663 { "uart_hwmtx", &lock_class_mtx_spin },
664 { "fast_taskqueue", &lock_class_mtx_spin },
665 { "intr table", &lock_class_mtx_spin },
666 { "process slock", &lock_class_mtx_spin },
667 { "syscons video lock", &lock_class_mtx_spin },
668 { "sleepq chain", &lock_class_mtx_spin },
669 { "rm_spinlock", &lock_class_mtx_spin },
670 { "turnstile chain", &lock_class_mtx_spin },
671 { "turnstile lock", &lock_class_mtx_spin },
672 { "sched lock", &lock_class_mtx_spin },
673 { "td_contested", &lock_class_mtx_spin },
674 { "callout", &lock_class_mtx_spin },
675 { "entropy harvest mutex", &lock_class_mtx_spin },
677 { "smp rendezvous", &lock_class_mtx_spin },
680 { "tlb0", &lock_class_mtx_spin },
683 { "sched lock", &lock_class_mtx_spin },
685 { "pmc-per-proc", &lock_class_mtx_spin },
691 { "intrcnt", &lock_class_mtx_spin },
692 { "icu", &lock_class_mtx_spin },
693 #if defined(SMP) && defined(__sparc64__)
694 { "ipi", &lock_class_mtx_spin },
697 { "allpmaps", &lock_class_mtx_spin },
698 { "descriptor tables", &lock_class_mtx_spin },
700 { "clk", &lock_class_mtx_spin },
701 { "cpuset", &lock_class_mtx_spin },
702 { "mprof lock", &lock_class_mtx_spin },
703 { "zombie lock", &lock_class_mtx_spin },
704 { "ALD Queue", &lock_class_mtx_spin },
705 #if defined(__i386__) || defined(__amd64__)
706 { "pcicfg", &lock_class_mtx_spin },
707 { "NDIS thread lock", &lock_class_mtx_spin },
709 { "tw_osl_io_lock", &lock_class_mtx_spin },
710 { "tw_osl_q_lock", &lock_class_mtx_spin },
711 { "tw_cl_io_lock", &lock_class_mtx_spin },
712 { "tw_cl_intr_lock", &lock_class_mtx_spin },
713 { "tw_cl_gen_lock", &lock_class_mtx_spin },
715 { "pmc-leaf", &lock_class_mtx_spin },
717 { "blocked lock", &lock_class_mtx_spin },
723 * Pairs of locks which have been blessed. Witness does not complain about
724 * order problems with blessed lock pairs. Please do not add an entry to the
725 * table without an explanatory comment.
727 static struct witness_blessed blessed_list[] = {
729 * See the comment in ufs_dirhash.c. Basically, a vnode lock serializes
730 * both lock orders, so a deadlock cannot happen as a result of this
733 { "dirhash", "bufwait" },
736 * A UFS vnode may be locked in vget() while a buffer belonging to the
737 * parent directory vnode is locked.
739 { "ufs", "bufwait" },
743 * This global is set to 0 once it becomes safe to use the witness code.
745 static int witness_cold = 1;
748 * This global is set to 1 once the static lock orders have been enrolled
749 * so that a warning can be issued for any spin locks enrolled later.
751 static int witness_spin_warn = 0;
753 /* Trim useless garbage from filenames. */
755 fixup_filename(const char *file)
760 while (strncmp(file, "../", 3) == 0)
766 * Calculate the size of early witness structures.
769 witness_startup_count(void)
773 sz = sizeof(struct witness) * witness_count;
774 sz += sizeof(*w_rmatrix) * (witness_count + 1);
775 sz += sizeof(*w_rmatrix[0]) * (witness_count + 1) *
782 * The WITNESS-enabled diagnostic code. Note that the witness code does
783 * assume that the early boot is single-threaded at least until after this
784 * routine is completed.
787 witness_startup(void *mem)
789 struct lock_object *lock;
790 struct witness_order_list_entry *order;
791 struct witness *w, *w1;
797 p += sizeof(struct witness) * witness_count;
799 w_rmatrix = (void *)p;
800 p += sizeof(*w_rmatrix) * (witness_count + 1);
802 for (i = 0; i < witness_count + 1; i++) {
803 w_rmatrix[i] = (void *)p;
804 p += sizeof(*w_rmatrix[i]) * (witness_count + 1);
806 badstack_sbuf_size = witness_count * 256;
809 * We have to release Giant before initializing its witness
810 * structure so that WITNESS doesn't get confused.
813 mtx_assert(&Giant, MA_NOTOWNED);
815 CTR1(KTR_WITNESS, "%s: initializing witness", __func__);
816 mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET |
817 MTX_NOWITNESS | MTX_NOPROFILE);
818 for (i = witness_count - 1; i >= 0; i--) {
820 memset(w, 0, sizeof(*w));
821 w_data[i].w_index = i; /* Witness index never changes. */
824 KASSERT(STAILQ_FIRST(&w_free)->w_index == 0,
825 ("%s: Invalid list of free witness objects", __func__));
827 /* Witness with index 0 is not used to aid in debugging. */
828 STAILQ_REMOVE_HEAD(&w_free, w_list);
831 for (i = 0; i < witness_count; i++) {
832 memset(w_rmatrix[i], 0, sizeof(*w_rmatrix[i]) *
833 (witness_count + 1));
836 for (i = 0; i < LOCK_CHILDCOUNT; i++)
837 witness_lock_list_free(&w_locklistdata[i]);
838 witness_init_hash_tables();
840 /* First add in all the specified order lists. */
841 for (order = order_lists; order->w_name != NULL; order++) {
842 w = enroll(order->w_name, order->w_class);
845 w->w_file = "order list";
846 for (order++; order->w_name != NULL; order++) {
847 w1 = enroll(order->w_name, order->w_class);
850 w1->w_file = "order list";
855 witness_spin_warn = 1;
857 /* Iterate through all locks and add them to witness. */
858 for (i = 0; pending_locks[i].wh_lock != NULL; i++) {
859 lock = pending_locks[i].wh_lock;
860 KASSERT(lock->lo_flags & LO_WITNESS,
861 ("%s: lock %s is on pending list but not LO_WITNESS",
862 __func__, lock->lo_name));
863 lock->lo_witness = enroll(pending_locks[i].wh_type,
867 /* Mark the witness code as being ready for use. */
874 witness_init(struct lock_object *lock, const char *type)
876 struct lock_class *class;
878 /* Various sanity checks. */
879 class = LOCK_CLASS(lock);
880 if ((lock->lo_flags & LO_RECURSABLE) != 0 &&
881 (class->lc_flags & LC_RECURSABLE) == 0)
882 kassert_panic("%s: lock (%s) %s can not be recursable",
883 __func__, class->lc_name, lock->lo_name);
884 if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
885 (class->lc_flags & LC_SLEEPABLE) == 0)
886 kassert_panic("%s: lock (%s) %s can not be sleepable",
887 __func__, class->lc_name, lock->lo_name);
888 if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
889 (class->lc_flags & LC_UPGRADABLE) == 0)
890 kassert_panic("%s: lock (%s) %s can not be upgradable",
891 __func__, class->lc_name, lock->lo_name);
894 * If we shouldn't watch this lock, then just clear lo_witness.
895 * Otherwise, if witness_cold is set, then it is too early to
896 * enroll this lock, so defer it to witness_initialize() by adding
897 * it to the pending_locks list. If it is not too early, then enroll
900 if (witness_watch < 1 || panicstr != NULL ||
901 (lock->lo_flags & LO_WITNESS) == 0)
902 lock->lo_witness = NULL;
903 else if (witness_cold) {
904 pending_locks[pending_cnt].wh_lock = lock;
905 pending_locks[pending_cnt++].wh_type = type;
906 if (pending_cnt > WITNESS_PENDLIST)
907 panic("%s: pending locks list is too small, "
908 "increase WITNESS_PENDLIST\n",
911 lock->lo_witness = enroll(type, class);
915 witness_destroy(struct lock_object *lock)
917 struct lock_class *class;
920 class = LOCK_CLASS(lock);
923 panic("lock (%s) %s destroyed while witness_cold",
924 class->lc_name, lock->lo_name);
926 /* XXX: need to verify that no one holds the lock */
927 if ((lock->lo_flags & LO_WITNESS) == 0 || lock->lo_witness == NULL)
929 w = lock->lo_witness;
931 mtx_lock_spin(&w_mtx);
932 MPASS(w->w_refcount > 0);
935 if (w->w_refcount == 0)
937 mtx_unlock_spin(&w_mtx);
942 witness_ddb_compute_levels(void)
947 * First clear all levels.
949 STAILQ_FOREACH(w, &w_all, w_list)
953 * Look for locks with no parents and level all their descendants.
955 STAILQ_FOREACH(w, &w_all, w_list) {
957 /* If the witness has ancestors (is not a root), skip it. */
958 if (w->w_num_ancestors > 0)
960 witness_ddb_level_descendants(w, 0);
965 witness_ddb_level_descendants(struct witness *w, int l)
969 if (w->w_ddb_level >= l)
975 for (i = 1; i <= w_max_used_index; i++) {
976 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
977 witness_ddb_level_descendants(&w_data[i], l);
982 witness_ddb_display_descendants(int(*prnt)(const char *fmt, ...),
983 struct witness *w, int indent)
987 for (i = 0; i < indent; i++)
989 prnt("%s (type: %s, depth: %d, active refs: %d)",
990 w->w_name, w->w_class->lc_name,
991 w->w_ddb_level, w->w_refcount);
992 if (w->w_displayed) {
993 prnt(" -- (already displayed)\n");
997 if (w->w_file != NULL && w->w_line != 0)
998 prnt(" -- last acquired @ %s:%d\n", fixup_filename(w->w_file),
1001 prnt(" -- never acquired\n");
1003 WITNESS_INDEX_ASSERT(w->w_index);
1004 for (i = 1; i <= w_max_used_index; i++) {
1007 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
1008 witness_ddb_display_descendants(prnt, &w_data[i],
1014 witness_ddb_display_list(int(*prnt)(const char *fmt, ...),
1015 struct witness_list *list)
1019 STAILQ_FOREACH(w, list, w_typelist) {
1020 if (w->w_file == NULL || w->w_ddb_level > 0)
1023 /* This lock has no anscestors - display its descendants. */
1024 witness_ddb_display_descendants(prnt, w, 0);
1031 witness_ddb_display(int(*prnt)(const char *fmt, ...))
1035 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1036 witness_ddb_compute_levels();
1038 /* Clear all the displayed flags. */
1039 STAILQ_FOREACH(w, &w_all, w_list)
1043 * First, handle sleep locks which have been acquired at least
1046 prnt("Sleep locks:\n");
1047 witness_ddb_display_list(prnt, &w_sleep);
1052 * Now do spin locks which have been acquired at least once.
1054 prnt("\nSpin locks:\n");
1055 witness_ddb_display_list(prnt, &w_spin);
1060 * Finally, any locks which have not been acquired yet.
1062 prnt("\nLocks which were never acquired:\n");
1063 STAILQ_FOREACH(w, &w_all, w_list) {
1064 if (w->w_file != NULL || w->w_refcount == 0)
1066 prnt("%s (type: %s, depth: %d)\n", w->w_name,
1067 w->w_class->lc_name, w->w_ddb_level);
1075 witness_defineorder(struct lock_object *lock1, struct lock_object *lock2)
1078 if (witness_watch == -1 || panicstr != NULL)
1081 /* Require locks that witness knows about. */
1082 if (lock1 == NULL || lock1->lo_witness == NULL || lock2 == NULL ||
1083 lock2->lo_witness == NULL)
1086 mtx_assert(&w_mtx, MA_NOTOWNED);
1087 mtx_lock_spin(&w_mtx);
1090 * If we already have either an explicit or implied lock order that
1091 * is the other way around, then return an error.
1093 if (witness_watch &&
1094 isitmydescendant(lock2->lo_witness, lock1->lo_witness)) {
1095 mtx_unlock_spin(&w_mtx);
1099 /* Try to add the new order. */
1100 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1101 lock2->lo_witness->w_name, lock1->lo_witness->w_name);
1102 itismychild(lock1->lo_witness, lock2->lo_witness);
1103 mtx_unlock_spin(&w_mtx);
1108 witness_checkorder(struct lock_object *lock, int flags, const char *file,
1109 int line, struct lock_object *interlock)
1111 struct lock_list_entry *lock_list, *lle;
1112 struct lock_instance *lock1, *lock2, *plock;
1113 struct lock_class *class, *iclass;
1114 struct witness *w, *w1;
1118 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL ||
1122 w = lock->lo_witness;
1123 class = LOCK_CLASS(lock);
1126 if (class->lc_flags & LC_SLEEPLOCK) {
1129 * Since spin locks include a critical section, this check
1130 * implicitly enforces a lock order of all sleep locks before
1133 if (td->td_critnest != 0 && !kdb_active)
1134 kassert_panic("acquiring blockable sleep lock with "
1135 "spinlock or critical section held (%s) %s @ %s:%d",
1136 class->lc_name, lock->lo_name,
1137 fixup_filename(file), line);
1140 * If this is the first lock acquired then just return as
1141 * no order checking is needed.
1143 lock_list = td->td_sleeplocks;
1144 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++) {
1281 MPASS(j < LOCK_CHILDCOUNT * LOCK_NCHILDREN);
1282 lock1 = &lle->ll_children[i];
1285 * Ignore the interlock.
1287 if (interlock == lock1->li_lock)
1291 * If this lock doesn't undergo witness checking,
1294 w1 = lock1->li_lock->lo_witness;
1296 KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
1297 ("lock missing witness structure"));
1302 * If we are locking Giant and this is a sleepable
1303 * lock, then skip it.
1305 if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0 &&
1306 lock == &Giant.lock_object)
1310 * If we are locking a sleepable lock and this lock
1311 * is Giant, then skip it.
1313 if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1314 lock1->li_lock == &Giant.lock_object)
1318 * If we are locking a sleepable lock and this lock
1319 * isn't sleepable, we want to treat it as a lock
1320 * order violation to enfore a general lock order of
1321 * sleepable locks before non-sleepable locks.
1323 if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1324 (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
1328 * If we are locking Giant and this is a non-sleepable
1329 * lock, then treat it as a reversal.
1331 if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0 &&
1332 lock == &Giant.lock_object)
1336 * Check the lock order hierarchy for a reveresal.
1338 if (!isitmydescendant(w, w1))
1343 * We have a lock order violation, check to see if it
1344 * is allowed or has already been yelled about.
1347 /* Bail if this violation is known */
1348 if (w_rmatrix[w1->w_index][w->w_index] & WITNESS_REVERSAL)
1351 /* Record this as a violation */
1352 w_rmatrix[w1->w_index][w->w_index] |= WITNESS_REVERSAL;
1353 w_rmatrix[w->w_index][w1->w_index] |= WITNESS_REVERSAL;
1354 w->w_reversed = w1->w_reversed = 1;
1355 witness_increment_graph_generation();
1358 * If the lock order is blessed, bail before logging
1359 * anything. We don't look for other lock order
1360 * violations though, which may be a bug.
1364 mtx_unlock_spin(&w_mtx);
1366 #ifdef WITNESS_NO_VNODE
1368 * There are known LORs between VNODE locks. They are
1369 * not an indication of a bug. VNODE locks are flagged
1370 * as such (LO_IS_VNODE) and we don't yell if the LOR
1371 * is between 2 VNODE locks.
1373 if ((lock->lo_flags & LO_IS_VNODE) != 0 &&
1374 (lock1->li_lock->lo_flags & LO_IS_VNODE) != 0)
1379 * Ok, yell about it.
1381 if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1382 (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
1384 "lock order reversal: (sleepable after non-sleepable)\n");
1385 else if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0
1386 && lock == &Giant.lock_object)
1388 "lock order reversal: (Giant after non-sleepable)\n");
1390 witness_output("lock order reversal:\n");
1393 * Try to locate an earlier lock with
1394 * witness w in our list.
1397 lock2 = &lle->ll_children[i];
1398 MPASS(lock2->li_lock != NULL);
1399 if (lock2->li_lock->lo_witness == w)
1401 if (i == 0 && lle->ll_next != NULL) {
1403 i = lle->ll_count - 1;
1404 MPASS(i >= 0 && i < LOCK_NCHILDREN);
1409 witness_output(" 1st %p %s (%s) @ %s:%d\n",
1410 lock1->li_lock, lock1->li_lock->lo_name,
1411 w1->w_name, fixup_filename(lock1->li_file),
1413 witness_output(" 2nd %p %s (%s) @ %s:%d\n", lock,
1414 lock->lo_name, w->w_name,
1415 fixup_filename(file), line);
1417 witness_output(" 1st %p %s (%s) @ %s:%d\n",
1418 lock2->li_lock, lock2->li_lock->lo_name,
1419 lock2->li_lock->lo_witness->w_name,
1420 fixup_filename(lock2->li_file),
1422 witness_output(" 2nd %p %s (%s) @ %s:%d\n",
1423 lock1->li_lock, lock1->li_lock->lo_name,
1424 w1->w_name, fixup_filename(lock1->li_file),
1426 witness_output(" 3rd %p %s (%s) @ %s:%d\n", lock,
1427 lock->lo_name, w->w_name,
1428 fixup_filename(file), line);
1430 witness_debugger(1, __func__);
1436 * If requested, build a new lock order. However, don't build a new
1437 * relationship between a sleepable lock and Giant if it is in the
1438 * wrong direction. The correct lock order is that sleepable locks
1439 * always come before Giant.
1441 if (flags & LOP_NEWORDER &&
1442 !(plock->li_lock == &Giant.lock_object &&
1443 (lock->lo_flags & LO_SLEEPABLE) != 0)) {
1444 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1445 w->w_name, plock->li_lock->lo_witness->w_name);
1446 itismychild(plock->li_lock->lo_witness, w);
1449 mtx_unlock_spin(&w_mtx);
1453 witness_lock(struct lock_object *lock, int flags, const char *file, int line)
1455 struct lock_list_entry **lock_list, *lle;
1456 struct lock_instance *instance;
1460 if (witness_cold || witness_watch == -1 || lock->lo_witness == NULL ||
1463 w = lock->lo_witness;
1466 /* Determine lock list for this lock. */
1467 if (LOCK_CLASS(lock)->lc_flags & LC_SLEEPLOCK)
1468 lock_list = &td->td_sleeplocks;
1470 lock_list = PCPU_PTR(spinlocks);
1472 /* Check to see if we are recursing on a lock we already own. */
1473 instance = find_instance(*lock_list, lock);
1474 if (instance != NULL) {
1475 instance->li_flags++;
1476 CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
1477 td->td_proc->p_pid, lock->lo_name,
1478 instance->li_flags & LI_RECURSEMASK);
1479 instance->li_file = file;
1480 instance->li_line = line;
1484 /* Update per-witness last file and line acquire. */
1488 /* Find the next open lock instance in the list and fill it. */
1490 if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
1491 lle = witness_lock_list_get();
1494 lle->ll_next = *lock_list;
1495 CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
1496 td->td_proc->p_pid, lle);
1499 instance = &lle->ll_children[lle->ll_count++];
1500 instance->li_lock = lock;
1501 instance->li_line = line;
1502 instance->li_file = file;
1503 if ((flags & LOP_EXCLUSIVE) != 0)
1504 instance->li_flags = LI_EXCLUSIVE;
1506 instance->li_flags = 0;
1507 CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
1508 td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
1512 witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
1514 struct lock_instance *instance;
1515 struct lock_class *class;
1517 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1518 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
1520 class = LOCK_CLASS(lock);
1521 if (witness_watch) {
1522 if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1524 "upgrade of non-upgradable lock (%s) %s @ %s:%d",
1525 class->lc_name, lock->lo_name,
1526 fixup_filename(file), line);
1527 if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1529 "upgrade of non-sleep lock (%s) %s @ %s:%d",
1530 class->lc_name, lock->lo_name,
1531 fixup_filename(file), line);
1533 instance = find_instance(curthread->td_sleeplocks, lock);
1534 if (instance == NULL) {
1535 kassert_panic("upgrade of unlocked lock (%s) %s @ %s:%d",
1536 class->lc_name, lock->lo_name,
1537 fixup_filename(file), line);
1540 if (witness_watch) {
1541 if ((instance->li_flags & LI_EXCLUSIVE) != 0)
1543 "upgrade of exclusive lock (%s) %s @ %s:%d",
1544 class->lc_name, lock->lo_name,
1545 fixup_filename(file), line);
1546 if ((instance->li_flags & LI_RECURSEMASK) != 0)
1548 "upgrade of recursed lock (%s) %s r=%d @ %s:%d",
1549 class->lc_name, lock->lo_name,
1550 instance->li_flags & LI_RECURSEMASK,
1551 fixup_filename(file), line);
1553 instance->li_flags |= LI_EXCLUSIVE;
1557 witness_downgrade(struct lock_object *lock, int flags, const char *file,
1560 struct lock_instance *instance;
1561 struct lock_class *class;
1563 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1564 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
1566 class = LOCK_CLASS(lock);
1567 if (witness_watch) {
1568 if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1570 "downgrade of non-upgradable lock (%s) %s @ %s:%d",
1571 class->lc_name, lock->lo_name,
1572 fixup_filename(file), line);
1573 if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1575 "downgrade of non-sleep lock (%s) %s @ %s:%d",
1576 class->lc_name, lock->lo_name,
1577 fixup_filename(file), line);
1579 instance = find_instance(curthread->td_sleeplocks, lock);
1580 if (instance == NULL) {
1581 kassert_panic("downgrade of unlocked lock (%s) %s @ %s:%d",
1582 class->lc_name, lock->lo_name,
1583 fixup_filename(file), line);
1586 if (witness_watch) {
1587 if ((instance->li_flags & LI_EXCLUSIVE) == 0)
1589 "downgrade of shared lock (%s) %s @ %s:%d",
1590 class->lc_name, lock->lo_name,
1591 fixup_filename(file), line);
1592 if ((instance->li_flags & LI_RECURSEMASK) != 0)
1594 "downgrade of recursed lock (%s) %s r=%d @ %s:%d",
1595 class->lc_name, lock->lo_name,
1596 instance->li_flags & LI_RECURSEMASK,
1597 fixup_filename(file), line);
1599 instance->li_flags &= ~LI_EXCLUSIVE;
1603 witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
1605 struct lock_list_entry **lock_list, *lle;
1606 struct lock_instance *instance;
1607 struct lock_class *class;
1612 if (witness_cold || lock->lo_witness == NULL || panicstr != NULL)
1615 class = LOCK_CLASS(lock);
1617 /* Find lock instance associated with this lock. */
1618 if (class->lc_flags & LC_SLEEPLOCK)
1619 lock_list = &td->td_sleeplocks;
1621 lock_list = PCPU_PTR(spinlocks);
1623 for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
1624 for (i = 0; i < (*lock_list)->ll_count; i++) {
1625 instance = &(*lock_list)->ll_children[i];
1626 if (instance->li_lock == lock)
1631 * When disabling WITNESS through witness_watch we could end up in
1632 * having registered locks in the td_sleeplocks queue.
1633 * We have to make sure we flush these queues, so just search for
1634 * eventual register locks and remove them.
1636 if (witness_watch > 0) {
1637 kassert_panic("lock (%s) %s not locked @ %s:%d", class->lc_name,
1638 lock->lo_name, fixup_filename(file), line);
1645 /* First, check for shared/exclusive mismatches. */
1646 if ((instance->li_flags & LI_EXCLUSIVE) != 0 && witness_watch > 0 &&
1647 (flags & LOP_EXCLUSIVE) == 0) {
1648 witness_output("shared unlock of (%s) %s @ %s:%d\n",
1649 class->lc_name, lock->lo_name, fixup_filename(file), line);
1650 witness_output("while exclusively locked from %s:%d\n",
1651 fixup_filename(instance->li_file), instance->li_line);
1652 kassert_panic("excl->ushare");
1654 if ((instance->li_flags & LI_EXCLUSIVE) == 0 && witness_watch > 0 &&
1655 (flags & LOP_EXCLUSIVE) != 0) {
1656 witness_output("exclusive unlock of (%s) %s @ %s:%d\n",
1657 class->lc_name, lock->lo_name, fixup_filename(file), line);
1658 witness_output("while share locked from %s:%d\n",
1659 fixup_filename(instance->li_file),
1661 kassert_panic("share->uexcl");
1663 /* If we are recursed, unrecurse. */
1664 if ((instance->li_flags & LI_RECURSEMASK) > 0) {
1665 CTR4(KTR_WITNESS, "%s: pid %d unrecursed on %s r=%d", __func__,
1666 td->td_proc->p_pid, instance->li_lock->lo_name,
1667 instance->li_flags);
1668 instance->li_flags--;
1671 /* The lock is now being dropped, check for NORELEASE flag */
1672 if ((instance->li_flags & LI_NORELEASE) != 0 && witness_watch > 0) {
1673 witness_output("forbidden unlock of (%s) %s @ %s:%d\n",
1674 class->lc_name, lock->lo_name, fixup_filename(file), line);
1675 kassert_panic("lock marked norelease");
1678 /* Otherwise, remove this item from the list. */
1680 CTR4(KTR_WITNESS, "%s: pid %d removed %s from lle[%d]", __func__,
1681 td->td_proc->p_pid, instance->li_lock->lo_name,
1682 (*lock_list)->ll_count - 1);
1683 for (j = i; j < (*lock_list)->ll_count - 1; j++)
1684 (*lock_list)->ll_children[j] =
1685 (*lock_list)->ll_children[j + 1];
1686 (*lock_list)->ll_count--;
1690 * In order to reduce contention on w_mtx, we want to keep always an
1691 * head object into lists so that frequent allocation from the
1692 * free witness pool (and subsequent locking) is avoided.
1693 * In order to maintain the current code simple, when the head
1694 * object is totally unloaded it means also that we do not have
1695 * further objects in the list, so the list ownership needs to be
1696 * hand over to another object if the current head needs to be freed.
1698 if ((*lock_list)->ll_count == 0) {
1699 if (*lock_list == lle) {
1700 if (lle->ll_next == NULL)
1704 *lock_list = lle->ll_next;
1705 CTR3(KTR_WITNESS, "%s: pid %d removed lle %p", __func__,
1706 td->td_proc->p_pid, lle);
1707 witness_lock_list_free(lle);
1712 witness_thread_exit(struct thread *td)
1714 struct lock_list_entry *lle;
1717 lle = td->td_sleeplocks;
1718 if (lle == NULL || panicstr != NULL)
1720 if (lle->ll_count != 0) {
1721 for (n = 0; lle != NULL; lle = lle->ll_next)
1722 for (i = lle->ll_count - 1; i >= 0; i--) {
1725 "Thread %p exiting with the following locks held:\n", td);
1727 witness_list_lock(&lle->ll_children[i],
1732 "Thread %p cannot exit while holding sleeplocks\n", td);
1734 witness_lock_list_free(lle);
1738 * Warn if any locks other than 'lock' are held. Flags can be passed in to
1739 * exempt Giant and sleepable locks from the checks as well. If any
1740 * non-exempt locks are held, then a supplied message is printed to the
1741 * output channel along with a list of the offending locks. If indicated in the
1742 * flags then a failure results in a panic as well.
1745 witness_warn(int flags, struct lock_object *lock, const char *fmt, ...)
1747 struct lock_list_entry *lock_list, *lle;
1748 struct lock_instance *lock1;
1753 if (witness_cold || witness_watch < 1 || panicstr != NULL)
1757 for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next)
1758 for (i = lle->ll_count - 1; i >= 0; i--) {
1759 lock1 = &lle->ll_children[i];
1760 if (lock1->li_lock == lock)
1762 if (flags & WARN_GIANTOK &&
1763 lock1->li_lock == &Giant.lock_object)
1765 if (flags & WARN_SLEEPOK &&
1766 (lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0)
1772 printf(" with the following %slocks held:\n",
1773 (flags & WARN_SLEEPOK) != 0 ?
1774 "non-sleepable " : "");
1777 witness_list_lock(lock1, printf);
1781 * Pin the thread in order to avoid problems with thread migration.
1782 * Once that all verifies are passed about spinlocks ownership,
1783 * the thread is in a safe path and it can be unpinned.
1786 lock_list = PCPU_GET(spinlocks);
1787 if (lock_list != NULL && lock_list->ll_count != 0) {
1791 * We should only have one spinlock and as long as
1792 * the flags cannot match for this locks class,
1793 * check if the first spinlock is the one curthread
1796 lock1 = &lock_list->ll_children[lock_list->ll_count - 1];
1797 if (lock_list->ll_count == 1 && lock_list->ll_next == NULL &&
1798 lock1->li_lock == lock && n == 0)
1804 printf(" with the following %slocks held:\n",
1805 (flags & WARN_SLEEPOK) != 0 ? "non-sleepable " : "");
1806 n += witness_list_locks(&lock_list, printf);
1809 if (flags & WARN_PANIC && n)
1810 kassert_panic("%s", __func__);
1812 witness_debugger(n, __func__);
1817 witness_file(struct lock_object *lock)
1821 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1823 w = lock->lo_witness;
1828 witness_line(struct lock_object *lock)
1832 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1834 w = lock->lo_witness;
1838 static struct witness *
1839 enroll(const char *description, struct lock_class *lock_class)
1843 MPASS(description != NULL);
1845 if (witness_watch == -1 || panicstr != NULL)
1847 if ((lock_class->lc_flags & LC_SPINLOCK)) {
1848 if (witness_skipspin)
1850 } else if ((lock_class->lc_flags & LC_SLEEPLOCK) == 0) {
1851 kassert_panic("lock class %s is not sleep or spin",
1852 lock_class->lc_name);
1856 mtx_lock_spin(&w_mtx);
1857 w = witness_hash_get(description);
1860 if ((w = witness_get()) == NULL)
1862 MPASS(strlen(description) < MAX_W_NAME);
1863 strcpy(w->w_name, description);
1864 w->w_class = lock_class;
1866 STAILQ_INSERT_HEAD(&w_all, w, w_list);
1867 if (lock_class->lc_flags & LC_SPINLOCK) {
1868 STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
1870 } else if (lock_class->lc_flags & LC_SLEEPLOCK) {
1871 STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
1875 /* Insert new witness into the hash */
1876 witness_hash_put(w);
1877 witness_increment_graph_generation();
1878 mtx_unlock_spin(&w_mtx);
1882 if (w->w_refcount == 1)
1883 w->w_class = lock_class;
1884 mtx_unlock_spin(&w_mtx);
1885 if (lock_class != w->w_class)
1887 "lock (%s) %s does not match earlier (%s) lock",
1888 description, lock_class->lc_name,
1889 w->w_class->lc_name);
1894 depart(struct witness *w)
1897 MPASS(w->w_refcount == 0);
1898 if (w->w_class->lc_flags & LC_SLEEPLOCK) {
1904 * Set file to NULL as it may point into a loadable module.
1908 witness_increment_graph_generation();
1913 adopt(struct witness *parent, struct witness *child)
1917 if (witness_cold == 0)
1918 mtx_assert(&w_mtx, MA_OWNED);
1920 /* If the relationship is already known, there's no work to be done. */
1921 if (isitmychild(parent, child))
1924 /* When the structure of the graph changes, bump up the generation. */
1925 witness_increment_graph_generation();
1928 * The hard part ... create the direct relationship, then propagate all
1929 * indirect relationships.
1931 pi = parent->w_index;
1932 ci = child->w_index;
1933 WITNESS_INDEX_ASSERT(pi);
1934 WITNESS_INDEX_ASSERT(ci);
1936 w_rmatrix[pi][ci] |= WITNESS_PARENT;
1937 w_rmatrix[ci][pi] |= WITNESS_CHILD;
1940 * If parent was not already an ancestor of child,
1941 * then we increment the descendant and ancestor counters.
1943 if ((w_rmatrix[pi][ci] & WITNESS_ANCESTOR) == 0) {
1944 parent->w_num_descendants++;
1945 child->w_num_ancestors++;
1949 * Find each ancestor of 'pi'. Note that 'pi' itself is counted as
1950 * an ancestor of 'pi' during this loop.
1952 for (i = 1; i <= w_max_used_index; i++) {
1953 if ((w_rmatrix[i][pi] & WITNESS_ANCESTOR_MASK) == 0 &&
1957 /* Find each descendant of 'i' and mark it as a descendant. */
1958 for (j = 1; j <= w_max_used_index; j++) {
1961 * Skip children that are already marked as
1962 * descendants of 'i'.
1964 if (w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK)
1968 * We are only interested in descendants of 'ci'. Note
1969 * that 'ci' itself is counted as a descendant of 'ci'.
1971 if ((w_rmatrix[ci][j] & WITNESS_ANCESTOR_MASK) == 0 &&
1974 w_rmatrix[i][j] |= WITNESS_ANCESTOR;
1975 w_rmatrix[j][i] |= WITNESS_DESCENDANT;
1976 w_data[i].w_num_descendants++;
1977 w_data[j].w_num_ancestors++;
1980 * Make sure we aren't marking a node as both an
1981 * ancestor and descendant. We should have caught
1982 * this as a lock order reversal earlier.
1984 if ((w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK) &&
1985 (w_rmatrix[i][j] & WITNESS_DESCENDANT_MASK)) {
1986 printf("witness rmatrix paradox! [%d][%d]=%d "
1987 "both ancestor and descendant\n",
1988 i, j, w_rmatrix[i][j]);
1990 printf("Witness disabled.\n");
1993 if ((w_rmatrix[j][i] & WITNESS_ANCESTOR_MASK) &&
1994 (w_rmatrix[j][i] & WITNESS_DESCENDANT_MASK)) {
1995 printf("witness rmatrix paradox! [%d][%d]=%d "
1996 "both ancestor and descendant\n",
1997 j, i, w_rmatrix[j][i]);
1999 printf("Witness disabled.\n");
2007 itismychild(struct witness *parent, struct witness *child)
2011 MPASS(child != NULL && parent != NULL);
2012 if (witness_cold == 0)
2013 mtx_assert(&w_mtx, MA_OWNED);
2015 if (!witness_lock_type_equal(parent, child)) {
2016 if (witness_cold == 0) {
2018 mtx_unlock_spin(&w_mtx);
2023 "%s: parent \"%s\" (%s) and child \"%s\" (%s) are not "
2024 "the same lock type", __func__, parent->w_name,
2025 parent->w_class->lc_name, child->w_name,
2026 child->w_class->lc_name);
2028 mtx_lock_spin(&w_mtx);
2030 adopt(parent, child);
2034 * Generic code for the isitmy*() functions. The rmask parameter is the
2035 * expected relationship of w1 to w2.
2038 _isitmyx(struct witness *w1, struct witness *w2, int rmask, const char *fname)
2040 unsigned char r1, r2;
2045 WITNESS_INDEX_ASSERT(i1);
2046 WITNESS_INDEX_ASSERT(i2);
2047 r1 = w_rmatrix[i1][i2] & WITNESS_RELATED_MASK;
2048 r2 = w_rmatrix[i2][i1] & WITNESS_RELATED_MASK;
2050 /* The flags on one better be the inverse of the flags on the other */
2051 if (!((WITNESS_ATOD(r1) == r2 && WITNESS_DTOA(r2) == r1) ||
2052 (WITNESS_DTOA(r1) == r2 && WITNESS_ATOD(r2) == r1))) {
2053 /* Don't squawk if we're potentially racing with an update. */
2054 if (!mtx_owned(&w_mtx))
2056 printf("%s: rmatrix mismatch between %s (index %d) and %s "
2057 "(index %d): w_rmatrix[%d][%d] == %hhx but "
2058 "w_rmatrix[%d][%d] == %hhx\n",
2059 fname, w1->w_name, i1, w2->w_name, i2, i1, i2, r1,
2062 printf("Witness disabled.\n");
2065 return (r1 & rmask);
2069 * Checks if @child is a direct child of @parent.
2072 isitmychild(struct witness *parent, struct witness *child)
2075 return (_isitmyx(parent, child, WITNESS_PARENT, __func__));
2079 * Checks if @descendant is a direct or inderect descendant of @ancestor.
2082 isitmydescendant(struct witness *ancestor, struct witness *descendant)
2085 return (_isitmyx(ancestor, descendant, WITNESS_ANCESTOR_MASK,
2090 blessed(struct witness *w1, struct witness *w2)
2093 struct witness_blessed *b;
2095 for (i = 0; i < nitems(blessed_list); i++) {
2096 b = &blessed_list[i];
2097 if (strcmp(w1->w_name, b->b_lock1) == 0) {
2098 if (strcmp(w2->w_name, b->b_lock2) == 0)
2102 if (strcmp(w1->w_name, b->b_lock2) == 0)
2103 if (strcmp(w2->w_name, b->b_lock1) == 0)
2109 static struct witness *
2115 if (witness_cold == 0)
2116 mtx_assert(&w_mtx, MA_OWNED);
2118 if (witness_watch == -1) {
2119 mtx_unlock_spin(&w_mtx);
2122 if (STAILQ_EMPTY(&w_free)) {
2124 mtx_unlock_spin(&w_mtx);
2125 printf("WITNESS: unable to allocate a new witness object\n");
2128 w = STAILQ_FIRST(&w_free);
2129 STAILQ_REMOVE_HEAD(&w_free, w_list);
2132 MPASS(index > 0 && index == w_max_used_index+1 &&
2133 index < witness_count);
2134 bzero(w, sizeof(*w));
2136 if (index > w_max_used_index)
2137 w_max_used_index = index;
2142 witness_free(struct witness *w)
2145 STAILQ_INSERT_HEAD(&w_free, w, w_list);
2149 static struct lock_list_entry *
2150 witness_lock_list_get(void)
2152 struct lock_list_entry *lle;
2154 if (witness_watch == -1)
2156 mtx_lock_spin(&w_mtx);
2157 lle = w_lock_list_free;
2160 mtx_unlock_spin(&w_mtx);
2161 printf("%s: witness exhausted\n", __func__);
2164 w_lock_list_free = lle->ll_next;
2165 mtx_unlock_spin(&w_mtx);
2166 bzero(lle, sizeof(*lle));
2171 witness_lock_list_free(struct lock_list_entry *lle)
2174 mtx_lock_spin(&w_mtx);
2175 lle->ll_next = w_lock_list_free;
2176 w_lock_list_free = lle;
2177 mtx_unlock_spin(&w_mtx);
2180 static struct lock_instance *
2181 find_instance(struct lock_list_entry *list, const struct lock_object *lock)
2183 struct lock_list_entry *lle;
2184 struct lock_instance *instance;
2187 for (lle = list; lle != NULL; lle = lle->ll_next)
2188 for (i = lle->ll_count - 1; i >= 0; i--) {
2189 instance = &lle->ll_children[i];
2190 if (instance->li_lock == lock)
2197 witness_list_lock(struct lock_instance *instance,
2198 int (*prnt)(const char *fmt, ...))
2200 struct lock_object *lock;
2202 lock = instance->li_lock;
2203 prnt("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ?
2204 "exclusive" : "shared", LOCK_CLASS(lock)->lc_name, lock->lo_name);
2205 if (lock->lo_witness->w_name != lock->lo_name)
2206 prnt(" (%s)", lock->lo_witness->w_name);
2207 prnt(" r = %d (%p) locked @ %s:%d\n",
2208 instance->li_flags & LI_RECURSEMASK, lock,
2209 fixup_filename(instance->li_file), instance->li_line);
2213 witness_output(const char *fmt, ...)
2219 ret = witness_voutput(fmt, ap);
2225 witness_voutput(const char *fmt, va_list ap)
2230 switch (witness_channel) {
2231 case WITNESS_CONSOLE:
2232 ret = vprintf(fmt, ap);
2235 vlog(LOG_NOTICE, fmt, ap);
2245 witness_thread_has_locks(struct thread *td)
2248 if (td->td_sleeplocks == NULL)
2250 return (td->td_sleeplocks->ll_count != 0);
2254 witness_proc_has_locks(struct proc *p)
2258 FOREACH_THREAD_IN_PROC(p, td) {
2259 if (witness_thread_has_locks(td))
2267 witness_list_locks(struct lock_list_entry **lock_list,
2268 int (*prnt)(const char *fmt, ...))
2270 struct lock_list_entry *lle;
2274 for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
2275 for (i = lle->ll_count - 1; i >= 0; i--) {
2276 witness_list_lock(&lle->ll_children[i], prnt);
2283 * This is a bit risky at best. We call this function when we have timed
2284 * out acquiring a spin lock, and we assume that the other CPU is stuck
2285 * with this lock held. So, we go groveling around in the other CPU's
2286 * per-cpu data to try to find the lock instance for this spin lock to
2287 * see when it was last acquired.
2290 witness_display_spinlock(struct lock_object *lock, struct thread *owner,
2291 int (*prnt)(const char *fmt, ...))
2293 struct lock_instance *instance;
2296 if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU)
2298 pc = pcpu_find(owner->td_oncpu);
2299 instance = find_instance(pc->pc_spinlocks, lock);
2300 if (instance != NULL)
2301 witness_list_lock(instance, prnt);
2305 witness_save(struct lock_object *lock, const char **filep, int *linep)
2307 struct lock_list_entry *lock_list;
2308 struct lock_instance *instance;
2309 struct lock_class *class;
2312 * This function is used independently in locking code to deal with
2313 * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2316 if (SCHEDULER_STOPPED())
2318 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2319 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2321 class = LOCK_CLASS(lock);
2322 if (class->lc_flags & LC_SLEEPLOCK)
2323 lock_list = curthread->td_sleeplocks;
2325 if (witness_skipspin)
2327 lock_list = PCPU_GET(spinlocks);
2329 instance = find_instance(lock_list, lock);
2330 if (instance == NULL) {
2331 kassert_panic("%s: lock (%s) %s not locked", __func__,
2332 class->lc_name, lock->lo_name);
2335 *filep = instance->li_file;
2336 *linep = instance->li_line;
2340 witness_restore(struct lock_object *lock, const char *file, int line)
2342 struct lock_list_entry *lock_list;
2343 struct lock_instance *instance;
2344 struct lock_class *class;
2347 * This function is used independently in locking code to deal with
2348 * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2351 if (SCHEDULER_STOPPED())
2353 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2354 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2356 class = LOCK_CLASS(lock);
2357 if (class->lc_flags & LC_SLEEPLOCK)
2358 lock_list = curthread->td_sleeplocks;
2360 if (witness_skipspin)
2362 lock_list = PCPU_GET(spinlocks);
2364 instance = find_instance(lock_list, lock);
2365 if (instance == NULL)
2366 kassert_panic("%s: lock (%s) %s not locked", __func__,
2367 class->lc_name, lock->lo_name);
2368 lock->lo_witness->w_file = file;
2369 lock->lo_witness->w_line = line;
2370 if (instance == NULL)
2372 instance->li_file = file;
2373 instance->li_line = line;
2377 witness_assert(const struct lock_object *lock, int flags, const char *file,
2380 #ifdef INVARIANT_SUPPORT
2381 struct lock_instance *instance;
2382 struct lock_class *class;
2384 if (lock->lo_witness == NULL || witness_watch < 1 || panicstr != NULL)
2386 class = LOCK_CLASS(lock);
2387 if ((class->lc_flags & LC_SLEEPLOCK) != 0)
2388 instance = find_instance(curthread->td_sleeplocks, lock);
2389 else if ((class->lc_flags & LC_SPINLOCK) != 0)
2390 instance = find_instance(PCPU_GET(spinlocks), lock);
2392 kassert_panic("Lock (%s) %s is not sleep or spin!",
2393 class->lc_name, lock->lo_name);
2398 if (instance != NULL)
2399 kassert_panic("Lock (%s) %s locked @ %s:%d.",
2400 class->lc_name, lock->lo_name,
2401 fixup_filename(file), line);
2404 case LA_LOCKED | LA_RECURSED:
2405 case LA_LOCKED | LA_NOTRECURSED:
2407 case LA_SLOCKED | LA_RECURSED:
2408 case LA_SLOCKED | LA_NOTRECURSED:
2410 case LA_XLOCKED | LA_RECURSED:
2411 case LA_XLOCKED | LA_NOTRECURSED:
2412 if (instance == NULL) {
2413 kassert_panic("Lock (%s) %s not locked @ %s:%d.",
2414 class->lc_name, lock->lo_name,
2415 fixup_filename(file), line);
2418 if ((flags & LA_XLOCKED) != 0 &&
2419 (instance->li_flags & LI_EXCLUSIVE) == 0)
2421 "Lock (%s) %s not exclusively locked @ %s:%d.",
2422 class->lc_name, lock->lo_name,
2423 fixup_filename(file), line);
2424 if ((flags & LA_SLOCKED) != 0 &&
2425 (instance->li_flags & LI_EXCLUSIVE) != 0)
2427 "Lock (%s) %s exclusively locked @ %s:%d.",
2428 class->lc_name, lock->lo_name,
2429 fixup_filename(file), line);
2430 if ((flags & LA_RECURSED) != 0 &&
2431 (instance->li_flags & LI_RECURSEMASK) == 0)
2432 kassert_panic("Lock (%s) %s not recursed @ %s:%d.",
2433 class->lc_name, lock->lo_name,
2434 fixup_filename(file), line);
2435 if ((flags & LA_NOTRECURSED) != 0 &&
2436 (instance->li_flags & LI_RECURSEMASK) != 0)
2437 kassert_panic("Lock (%s) %s recursed @ %s:%d.",
2438 class->lc_name, lock->lo_name,
2439 fixup_filename(file), line);
2442 kassert_panic("Invalid lock assertion at %s:%d.",
2443 fixup_filename(file), line);
2446 #endif /* INVARIANT_SUPPORT */
2450 witness_setflag(struct lock_object *lock, int flag, int set)
2452 struct lock_list_entry *lock_list;
2453 struct lock_instance *instance;
2454 struct lock_class *class;
2456 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2458 class = LOCK_CLASS(lock);
2459 if (class->lc_flags & LC_SLEEPLOCK)
2460 lock_list = curthread->td_sleeplocks;
2462 if (witness_skipspin)
2464 lock_list = PCPU_GET(spinlocks);
2466 instance = find_instance(lock_list, lock);
2467 if (instance == NULL) {
2468 kassert_panic("%s: lock (%s) %s not locked", __func__,
2469 class->lc_name, lock->lo_name);
2474 instance->li_flags |= flag;
2476 instance->li_flags &= ~flag;
2480 witness_norelease(struct lock_object *lock)
2483 witness_setflag(lock, LI_NORELEASE, 1);
2487 witness_releaseok(struct lock_object *lock)
2490 witness_setflag(lock, LI_NORELEASE, 0);
2495 witness_ddb_list(struct thread *td)
2498 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2499 KASSERT(kdb_active, ("%s: not in the debugger", __func__));
2501 if (witness_watch < 1)
2504 witness_list_locks(&td->td_sleeplocks, db_printf);
2507 * We only handle spinlocks if td == curthread. This is somewhat broken
2508 * if td is currently executing on some other CPU and holds spin locks
2509 * as we won't display those locks. If we had a MI way of getting
2510 * the per-cpu data for a given cpu then we could use
2511 * td->td_oncpu to get the list of spinlocks for this thread
2514 * That still wouldn't really fix this unless we locked the scheduler
2515 * lock or stopped the other CPU to make sure it wasn't changing the
2516 * list out from under us. It is probably best to just not try to
2517 * handle threads on other CPU's for now.
2519 if (td == curthread && PCPU_GET(spinlocks) != NULL)
2520 witness_list_locks(PCPU_PTR(spinlocks), db_printf);
2523 DB_SHOW_COMMAND(locks, db_witness_list)
2528 td = db_lookup_thread(addr, true);
2531 witness_ddb_list(td);
2534 DB_SHOW_ALL_COMMAND(locks, db_witness_list_all)
2540 * It would be nice to list only threads and processes that actually
2541 * held sleep locks, but that information is currently not exported
2544 FOREACH_PROC_IN_SYSTEM(p) {
2545 if (!witness_proc_has_locks(p))
2547 FOREACH_THREAD_IN_PROC(p, td) {
2548 if (!witness_thread_has_locks(td))
2550 db_printf("Process %d (%s) thread %p (%d)\n", p->p_pid,
2551 p->p_comm, td, td->td_tid);
2552 witness_ddb_list(td);
2558 DB_SHOW_ALIAS(alllocks, db_witness_list_all)
2560 DB_SHOW_COMMAND(witness, db_witness_display)
2563 witness_ddb_display(db_printf);
2568 sbuf_print_witness_badstacks(struct sbuf *sb, size_t *oldidx)
2570 struct witness_lock_order_data *data1, *data2, *tmp_data1, *tmp_data2;
2571 struct witness *tmp_w1, *tmp_w2, *w1, *w2;
2572 int generation, i, j;
2579 /* Allocate and init temporary storage space. */
2580 tmp_w1 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2581 tmp_w2 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2582 tmp_data1 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2584 tmp_data2 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2586 stack_zero(&tmp_data1->wlod_stack);
2587 stack_zero(&tmp_data2->wlod_stack);
2590 mtx_lock_spin(&w_mtx);
2591 generation = w_generation;
2592 mtx_unlock_spin(&w_mtx);
2593 sbuf_printf(sb, "Number of known direct relationships is %d\n",
2594 w_lohash.wloh_count);
2595 for (i = 1; i < w_max_used_index; i++) {
2596 mtx_lock_spin(&w_mtx);
2597 if (generation != w_generation) {
2598 mtx_unlock_spin(&w_mtx);
2600 /* The graph has changed, try again. */
2607 if (w1->w_reversed == 0) {
2608 mtx_unlock_spin(&w_mtx);
2612 /* Copy w1 locally so we can release the spin lock. */
2614 mtx_unlock_spin(&w_mtx);
2616 if (tmp_w1->w_reversed == 0)
2618 for (j = 1; j < w_max_used_index; j++) {
2619 if ((w_rmatrix[i][j] & WITNESS_REVERSAL) == 0 || i > j)
2622 mtx_lock_spin(&w_mtx);
2623 if (generation != w_generation) {
2624 mtx_unlock_spin(&w_mtx);
2626 /* The graph has changed, try again. */
2633 data1 = witness_lock_order_get(w1, w2);
2634 data2 = witness_lock_order_get(w2, w1);
2637 * Copy information locally so we can release the
2643 stack_zero(&tmp_data1->wlod_stack);
2644 stack_copy(&data1->wlod_stack,
2645 &tmp_data1->wlod_stack);
2647 if (data2 && data2 != data1) {
2648 stack_zero(&tmp_data2->wlod_stack);
2649 stack_copy(&data2->wlod_stack,
2650 &tmp_data2->wlod_stack);
2652 mtx_unlock_spin(&w_mtx);
2654 if (blessed(tmp_w1, tmp_w2))
2658 "\nLock order reversal between \"%s\"(%s) and \"%s\"(%s)!\n",
2659 tmp_w1->w_name, tmp_w1->w_class->lc_name,
2660 tmp_w2->w_name, tmp_w2->w_class->lc_name);
2663 "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2664 tmp_w1->w_name, tmp_w1->w_class->lc_name,
2665 tmp_w2->w_name, tmp_w2->w_class->lc_name);
2666 stack_sbuf_print(sb, &tmp_data1->wlod_stack);
2667 sbuf_printf(sb, "\n");
2669 if (data2 && data2 != data1) {
2671 "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2672 tmp_w2->w_name, tmp_w2->w_class->lc_name,
2673 tmp_w1->w_name, tmp_w1->w_class->lc_name);
2674 stack_sbuf_print(sb, &tmp_data2->wlod_stack);
2675 sbuf_printf(sb, "\n");
2679 mtx_lock_spin(&w_mtx);
2680 if (generation != w_generation) {
2681 mtx_unlock_spin(&w_mtx);
2684 * The graph changed while we were printing stack data,
2691 mtx_unlock_spin(&w_mtx);
2693 /* Free temporary storage space. */
2694 free(tmp_data1, M_TEMP);
2695 free(tmp_data2, M_TEMP);
2696 free(tmp_w1, M_TEMP);
2697 free(tmp_w2, M_TEMP);
2701 sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS)
2706 if (witness_watch < 1) {
2707 error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2711 error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2715 sb = sbuf_new(NULL, NULL, badstack_sbuf_size, SBUF_AUTOEXTEND);
2719 sbuf_print_witness_badstacks(sb, &req->oldidx);
2722 error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1);
2730 sbuf_db_printf_drain(void *arg __unused, const char *data, int len)
2733 return (db_printf("%.*s", len, data));
2736 DB_SHOW_COMMAND(badstacks, db_witness_badstacks)
2742 sbuf_new(&sb, buffer, sizeof(buffer), SBUF_FIXEDLEN);
2743 sbuf_set_drain(&sb, sbuf_db_printf_drain, NULL);
2744 sbuf_print_witness_badstacks(&sb, &dummy);
2750 sysctl_debug_witness_channel(SYSCTL_HANDLER_ARGS)
2752 static const struct {
2753 enum witness_channel channel;
2756 { WITNESS_CONSOLE, "console" },
2757 { WITNESS_LOG, "log" },
2758 { WITNESS_NONE, "none" },
2765 for (i = 0; i < nitems(channels); i++)
2766 if (witness_channel == channels[i].channel) {
2767 snprintf(buf, sizeof(buf), "%s", channels[i].name);
2771 error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
2772 if (error != 0 || req->newptr == NULL)
2776 for (i = 0; i < nitems(channels); i++)
2777 if (strcmp(channels[i].name, buf) == 0) {
2778 witness_channel = channels[i].channel;
2786 sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS)
2793 error = SYSCTL_OUT(req, w_notallowed, sizeof(w_notallowed));
2797 if (witness_watch < 1) {
2798 error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2802 error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2807 error = sysctl_wire_old_buffer(req, 0);
2810 sb = sbuf_new_for_sysctl(NULL, NULL, FULLGRAPH_SBUF_SIZE, req);
2813 sbuf_printf(sb, "\n");
2815 mtx_lock_spin(&w_mtx);
2816 STAILQ_FOREACH(w, &w_all, w_list)
2818 STAILQ_FOREACH(w, &w_all, w_list)
2819 witness_add_fullgraph(sb, w);
2820 mtx_unlock_spin(&w_mtx);
2823 * Close the sbuf and return to userland.
2825 error = sbuf_finish(sb);
2832 sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS)
2836 value = witness_watch;
2837 error = sysctl_handle_int(oidp, &value, 0, req);
2838 if (error != 0 || req->newptr == NULL)
2840 if (value > 1 || value < -1 ||
2841 (witness_watch == -1 && value != witness_watch))
2843 witness_watch = value;
2848 witness_add_fullgraph(struct sbuf *sb, struct witness *w)
2852 if (w->w_displayed != 0 || (w->w_file == NULL && w->w_line == 0))
2856 WITNESS_INDEX_ASSERT(w->w_index);
2857 for (i = 1; i <= w_max_used_index; i++) {
2858 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT) {
2859 sbuf_printf(sb, "\"%s\",\"%s\"\n", w->w_name,
2861 witness_add_fullgraph(sb, &w_data[i]);
2867 * A simple hash function. Takes a key pointer and a key size. If size == 0,
2868 * interprets the key as a string and reads until the null
2869 * terminator. Otherwise, reads the first size bytes. Returns an unsigned 32-bit
2870 * hash value computed from the key.
2873 witness_hash_djb2(const uint8_t *key, uint32_t size)
2875 unsigned int hash = 5381;
2878 /* hash = hash * 33 + key[i] */
2880 for (i = 0; i < size; i++)
2881 hash = ((hash << 5) + hash) + (unsigned int)key[i];
2883 for (i = 0; key[i] != 0; i++)
2884 hash = ((hash << 5) + hash) + (unsigned int)key[i];
2891 * Initializes the two witness hash tables. Called exactly once from
2892 * witness_initialize().
2895 witness_init_hash_tables(void)
2899 MPASS(witness_cold);
2901 /* Initialize the hash tables. */
2902 for (i = 0; i < WITNESS_HASH_SIZE; i++)
2903 w_hash.wh_array[i] = NULL;
2905 w_hash.wh_size = WITNESS_HASH_SIZE;
2906 w_hash.wh_count = 0;
2908 /* Initialize the lock order data hash. */
2910 for (i = 0; i < WITNESS_LO_DATA_COUNT; i++) {
2911 memset(&w_lodata[i], 0, sizeof(w_lodata[i]));
2912 w_lodata[i].wlod_next = w_lofree;
2913 w_lofree = &w_lodata[i];
2915 w_lohash.wloh_size = WITNESS_LO_HASH_SIZE;
2916 w_lohash.wloh_count = 0;
2917 for (i = 0; i < WITNESS_LO_HASH_SIZE; i++)
2918 w_lohash.wloh_array[i] = NULL;
2921 static struct witness *
2922 witness_hash_get(const char *key)
2928 if (witness_cold == 0)
2929 mtx_assert(&w_mtx, MA_OWNED);
2930 hash = witness_hash_djb2(key, 0) % w_hash.wh_size;
2931 w = w_hash.wh_array[hash];
2933 if (strcmp(w->w_name, key) == 0)
2943 witness_hash_put(struct witness *w)
2948 MPASS(w->w_name != NULL);
2949 if (witness_cold == 0)
2950 mtx_assert(&w_mtx, MA_OWNED);
2951 KASSERT(witness_hash_get(w->w_name) == NULL,
2952 ("%s: trying to add a hash entry that already exists!", __func__));
2953 KASSERT(w->w_hash_next == NULL,
2954 ("%s: w->w_hash_next != NULL", __func__));
2956 hash = witness_hash_djb2(w->w_name, 0) % w_hash.wh_size;
2957 w->w_hash_next = w_hash.wh_array[hash];
2958 w_hash.wh_array[hash] = w;
2963 static struct witness_lock_order_data *
2964 witness_lock_order_get(struct witness *parent, struct witness *child)
2966 struct witness_lock_order_data *data = NULL;
2967 struct witness_lock_order_key key;
2970 MPASS(parent != NULL && child != NULL);
2971 key.from = parent->w_index;
2972 key.to = child->w_index;
2973 WITNESS_INDEX_ASSERT(key.from);
2974 WITNESS_INDEX_ASSERT(key.to);
2975 if ((w_rmatrix[parent->w_index][child->w_index]
2976 & WITNESS_LOCK_ORDER_KNOWN) == 0)
2979 hash = witness_hash_djb2((const char*)&key,
2980 sizeof(key)) % w_lohash.wloh_size;
2981 data = w_lohash.wloh_array[hash];
2982 while (data != NULL) {
2983 if (witness_lock_order_key_equal(&data->wlod_key, &key))
2985 data = data->wlod_next;
2993 * Verify that parent and child have a known relationship, are not the same,
2994 * and child is actually a child of parent. This is done without w_mtx
2995 * to avoid contention in the common case.
2998 witness_lock_order_check(struct witness *parent, struct witness *child)
3001 if (parent != child &&
3002 w_rmatrix[parent->w_index][child->w_index]
3003 & WITNESS_LOCK_ORDER_KNOWN &&
3004 isitmychild(parent, child))
3011 witness_lock_order_add(struct witness *parent, struct witness *child)
3013 struct witness_lock_order_data *data = NULL;
3014 struct witness_lock_order_key key;
3017 MPASS(parent != NULL && child != NULL);
3018 key.from = parent->w_index;
3019 key.to = child->w_index;
3020 WITNESS_INDEX_ASSERT(key.from);
3021 WITNESS_INDEX_ASSERT(key.to);
3022 if (w_rmatrix[parent->w_index][child->w_index]
3023 & WITNESS_LOCK_ORDER_KNOWN)
3026 hash = witness_hash_djb2((const char*)&key,
3027 sizeof(key)) % w_lohash.wloh_size;
3028 w_rmatrix[parent->w_index][child->w_index] |= WITNESS_LOCK_ORDER_KNOWN;
3032 w_lofree = data->wlod_next;
3033 data->wlod_next = w_lohash.wloh_array[hash];
3034 data->wlod_key = key;
3035 w_lohash.wloh_array[hash] = data;
3036 w_lohash.wloh_count++;
3037 stack_zero(&data->wlod_stack);
3038 stack_save(&data->wlod_stack);
3042 /* Call this whenever the structure of the witness graph changes. */
3044 witness_increment_graph_generation(void)
3047 if (witness_cold == 0)
3048 mtx_assert(&w_mtx, MA_OWNED);
3053 witness_output_drain(void *arg __unused, const char *data, int len)
3056 witness_output("%.*s", len, data);
3061 witness_debugger(int cond, const char *msg)
3070 if (witness_trace) {
3071 sbuf_new(&sb, buf, sizeof(buf), SBUF_FIXEDLEN);
3072 sbuf_set_drain(&sb, witness_output_drain, NULL);
3076 witness_output("stack backtrace:\n");
3077 stack_sbuf_print_ddb(&sb, &st);
3084 kdb_enter(KDB_WHY_WITNESS, msg);