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. */
134 #define LI_SLEEPABLE 0x00040000 /* Lock may be held while sleeping. */
136 #ifndef WITNESS_COUNT
137 #define WITNESS_COUNT 1536
139 #define WITNESS_HASH_SIZE 251 /* Prime, gives load factor < 2 */
140 #define WITNESS_PENDLIST (512 + (MAXCPU * 4))
142 /* Allocate 256 KB of stack data space */
143 #define WITNESS_LO_DATA_COUNT 2048
145 /* Prime, gives load factor of ~2 at full load */
146 #define WITNESS_LO_HASH_SIZE 1021
149 * XXX: This is somewhat bogus, as we assume here that at most 2048 threads
150 * will hold LOCK_NCHILDREN locks. We handle failure ok, and we should
151 * probably be safe for the most part, but it's still a SWAG.
153 #define LOCK_NCHILDREN 5
154 #define LOCK_CHILDCOUNT 2048
156 #define MAX_W_NAME 64
158 #define FULLGRAPH_SBUF_SIZE 512
161 * These flags go in the witness relationship matrix and describe the
162 * relationship between any two struct witness objects.
164 #define WITNESS_UNRELATED 0x00 /* No lock order relation. */
165 #define WITNESS_PARENT 0x01 /* Parent, aka direct ancestor. */
166 #define WITNESS_ANCESTOR 0x02 /* Direct or indirect ancestor. */
167 #define WITNESS_CHILD 0x04 /* Child, aka direct descendant. */
168 #define WITNESS_DESCENDANT 0x08 /* Direct or indirect descendant. */
169 #define WITNESS_ANCESTOR_MASK (WITNESS_PARENT | WITNESS_ANCESTOR)
170 #define WITNESS_DESCENDANT_MASK (WITNESS_CHILD | WITNESS_DESCENDANT)
171 #define WITNESS_RELATED_MASK \
172 (WITNESS_ANCESTOR_MASK | WITNESS_DESCENDANT_MASK)
173 #define WITNESS_REVERSAL 0x10 /* A lock order reversal has been
175 #define WITNESS_RESERVED1 0x20 /* Unused flag, reserved. */
176 #define WITNESS_RESERVED2 0x40 /* Unused flag, reserved. */
177 #define WITNESS_LOCK_ORDER_KNOWN 0x80 /* This lock order is known. */
179 /* Descendant to ancestor flags */
180 #define WITNESS_DTOA(x) (((x) & WITNESS_RELATED_MASK) >> 2)
182 /* Ancestor to descendant flags */
183 #define WITNESS_ATOD(x) (((x) & WITNESS_RELATED_MASK) << 2)
185 #define WITNESS_INDEX_ASSERT(i) \
186 MPASS((i) > 0 && (i) <= w_max_used_index && (i) < witness_count)
188 static MALLOC_DEFINE(M_WITNESS, "Witness", "Witness");
191 * Lock instances. A lock instance is the data associated with a lock while
192 * it is held by witness. For example, a lock instance will hold the
193 * recursion count of a lock. Lock instances are held in lists. Spin locks
194 * are held in a per-cpu list while sleep locks are held in per-thread list.
196 struct lock_instance {
197 struct lock_object *li_lock;
204 * A simple list type used to build the list of locks held by a thread
205 * or CPU. We can't simply embed the list in struct lock_object since a
206 * lock may be held by more than one thread if it is a shared lock. Locks
207 * are added to the head of the list, so we fill up each list entry from
208 * "the back" logically. To ease some of the arithmetic, we actually fill
209 * in each list entry the normal way (children[0] then children[1], etc.) but
210 * when we traverse the list we read children[count-1] as the first entry
211 * down to children[0] as the final entry.
213 struct lock_list_entry {
214 struct lock_list_entry *ll_next;
215 struct lock_instance ll_children[LOCK_NCHILDREN];
220 * The main witness structure. One of these per named lock type in the system
221 * (for example, "vnode interlock").
224 char w_name[MAX_W_NAME];
225 uint32_t w_index; /* Index in the relationship matrix */
226 struct lock_class *w_class;
227 STAILQ_ENTRY(witness) w_list; /* List of all witnesses. */
228 STAILQ_ENTRY(witness) w_typelist; /* Witnesses of a type. */
229 struct witness *w_hash_next; /* Linked list in hash buckets. */
230 const char *w_file; /* File where last acquired */
231 uint32_t w_line; /* Line where last acquired */
233 uint16_t w_num_ancestors; /* direct/indirect
235 uint16_t w_num_descendants; /* direct/indirect
236 * descendant count */
238 unsigned w_displayed:1;
239 unsigned w_reversed:1;
242 STAILQ_HEAD(witness_list, witness);
245 * The witness hash table. Keys are witness names (const char *), elements are
246 * witness objects (struct witness *).
248 struct witness_hash {
249 struct witness *wh_array[WITNESS_HASH_SIZE];
255 * Key type for the lock order data hash table.
257 struct witness_lock_order_key {
262 struct witness_lock_order_data {
263 struct stack wlod_stack;
264 struct witness_lock_order_key wlod_key;
265 struct witness_lock_order_data *wlod_next;
269 * The witness lock order data hash table. Keys are witness index tuples
270 * (struct witness_lock_order_key), elements are lock order data objects
271 * (struct witness_lock_order_data).
273 struct witness_lock_order_hash {
274 struct witness_lock_order_data *wloh_array[WITNESS_LO_HASH_SIZE];
279 struct witness_blessed {
284 struct witness_pendhelp {
286 struct lock_object *wh_lock;
289 struct witness_order_list_entry {
291 struct lock_class *w_class;
295 * Returns 0 if one of the locks is a spin lock and the other is not.
296 * Returns 1 otherwise.
299 witness_lock_type_equal(struct witness *w1, struct witness *w2)
302 return ((w1->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) ==
303 (w2->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)));
307 witness_lock_order_key_equal(const struct witness_lock_order_key *a,
308 const struct witness_lock_order_key *b)
311 return (a->from == b->from && a->to == b->to);
314 static int _isitmyx(struct witness *w1, struct witness *w2, int rmask,
316 static void adopt(struct witness *parent, struct witness *child);
317 static int blessed(struct witness *, struct witness *);
318 static void depart(struct witness *w);
319 static struct witness *enroll(const char *description,
320 struct lock_class *lock_class);
321 static struct lock_instance *find_instance(struct lock_list_entry *list,
322 const struct lock_object *lock);
323 static int isitmychild(struct witness *parent, struct witness *child);
324 static int isitmydescendant(struct witness *parent, struct witness *child);
325 static void itismychild(struct witness *parent, struct witness *child);
326 static int sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS);
327 static int sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS);
328 static int sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS);
329 static int sysctl_debug_witness_channel(SYSCTL_HANDLER_ARGS);
330 static void witness_add_fullgraph(struct sbuf *sb, struct witness *parent);
332 static void witness_ddb_compute_levels(void);
333 static void witness_ddb_display(int(*)(const char *fmt, ...));
334 static void witness_ddb_display_descendants(int(*)(const char *fmt, ...),
335 struct witness *, int indent);
336 static void witness_ddb_display_list(int(*prnt)(const char *fmt, ...),
337 struct witness_list *list);
338 static void witness_ddb_level_descendants(struct witness *parent, int l);
339 static void witness_ddb_list(struct thread *td);
341 static void witness_enter_debugger(const char *msg);
342 static void witness_debugger(int cond, const char *msg);
343 static void witness_free(struct witness *m);
344 static struct witness *witness_get(void);
345 static uint32_t witness_hash_djb2(const uint8_t *key, uint32_t size);
346 static struct witness *witness_hash_get(const char *key);
347 static void witness_hash_put(struct witness *w);
348 static void witness_init_hash_tables(void);
349 static void witness_increment_graph_generation(void);
350 static void witness_lock_list_free(struct lock_list_entry *lle);
351 static struct lock_list_entry *witness_lock_list_get(void);
352 static int witness_lock_order_add(struct witness *parent,
353 struct witness *child);
354 static int witness_lock_order_check(struct witness *parent,
355 struct witness *child);
356 static struct witness_lock_order_data *witness_lock_order_get(
357 struct witness *parent,
358 struct witness *child);
359 static void witness_list_lock(struct lock_instance *instance,
360 int (*prnt)(const char *fmt, ...));
361 static int witness_output(const char *fmt, ...) __printflike(1, 2);
362 static int witness_output_drain(void *arg __unused, const char *data,
364 static int witness_voutput(const char *fmt, va_list ap) __printflike(1, 0);
365 static void witness_setflag(struct lock_object *lock, int flag, int set);
367 FEATURE(witness, "kernel has witness(9) support");
369 static SYSCTL_NODE(_debug, OID_AUTO, witness, CTLFLAG_RW | CTLFLAG_MPSAFE, NULL,
373 * If set to 0, lock order checking is disabled. If set to -1,
374 * witness is completely disabled. Otherwise witness performs full
375 * lock order checking for all locks. At runtime, lock order checking
376 * may be toggled. However, witness cannot be reenabled once it is
377 * completely disabled.
379 static int witness_watch = 1;
380 SYSCTL_PROC(_debug_witness, OID_AUTO, watch,
381 CTLFLAG_RWTUN | CTLTYPE_INT | CTLFLAG_MPSAFE, NULL, 0,
382 sysctl_debug_witness_watch, "I",
383 "witness is watching lock operations");
387 * When KDB is enabled and witness_kdb is 1, it will cause the system
388 * to drop into kdebug() when:
389 * - a lock hierarchy violation occurs
390 * - locks are held when going to sleep.
397 SYSCTL_INT(_debug_witness, OID_AUTO, kdb, CTLFLAG_RWTUN, &witness_kdb, 0, "");
400 #if defined(DDB) || defined(KDB)
402 * When DDB or KDB is enabled and witness_trace is 1, it will cause the system
403 * to print a stack trace:
404 * - a lock hierarchy violation occurs
405 * - locks are held when going to sleep.
407 int witness_trace = 1;
408 SYSCTL_INT(_debug_witness, OID_AUTO, trace, CTLFLAG_RWTUN, &witness_trace, 0, "");
409 #endif /* DDB || KDB */
411 #ifdef WITNESS_SKIPSPIN
412 int witness_skipspin = 1;
414 int witness_skipspin = 0;
416 SYSCTL_INT(_debug_witness, OID_AUTO, skipspin, CTLFLAG_RDTUN, &witness_skipspin, 0, "");
418 int badstack_sbuf_size;
420 int witness_count = WITNESS_COUNT;
421 SYSCTL_INT(_debug_witness, OID_AUTO, witness_count, CTLFLAG_RDTUN,
422 &witness_count, 0, "");
425 * Output channel for witness messages. By default we print to the console.
427 enum witness_channel {
433 static enum witness_channel witness_channel = WITNESS_CONSOLE;
434 SYSCTL_PROC(_debug_witness, OID_AUTO, output_channel,
435 CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_MPSAFE, NULL, 0,
436 sysctl_debug_witness_channel, "A",
437 "Output channel for warnings");
440 * Call this to print out the relations between locks.
442 SYSCTL_PROC(_debug_witness, OID_AUTO, fullgraph,
443 CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
444 sysctl_debug_witness_fullgraph, "A",
445 "Show locks relation graphs");
448 * Call this to print out the witness faulty stacks.
450 SYSCTL_PROC(_debug_witness, OID_AUTO, badstacks,
451 CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
452 sysctl_debug_witness_badstacks, "A",
453 "Show bad witness stacks");
455 static struct mtx w_mtx;
458 static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free);
459 static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all);
462 static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin);
463 static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep);
466 static struct lock_list_entry *w_lock_list_free = NULL;
467 static struct witness_pendhelp pending_locks[WITNESS_PENDLIST];
468 static u_int pending_cnt;
470 static int w_free_cnt, w_spin_cnt, w_sleep_cnt;
471 SYSCTL_INT(_debug_witness, OID_AUTO, free_cnt, CTLFLAG_RD, &w_free_cnt, 0, "");
472 SYSCTL_INT(_debug_witness, OID_AUTO, spin_cnt, CTLFLAG_RD, &w_spin_cnt, 0, "");
473 SYSCTL_INT(_debug_witness, OID_AUTO, sleep_cnt, CTLFLAG_RD, &w_sleep_cnt, 0,
476 static struct witness *w_data;
477 static uint8_t **w_rmatrix;
478 static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT];
479 static struct witness_hash w_hash; /* The witness hash table. */
481 /* The lock order data hash */
482 static struct witness_lock_order_data w_lodata[WITNESS_LO_DATA_COUNT];
483 static struct witness_lock_order_data *w_lofree = NULL;
484 static struct witness_lock_order_hash w_lohash;
485 static int w_max_used_index = 0;
486 static unsigned int w_generation = 0;
487 static const char w_notrunning[] = "Witness not running\n";
488 static const char w_stillcold[] = "Witness is still cold\n";
490 static const char w_notallowed[] = "The sysctl is disabled on the arch\n";
493 static struct witness_order_list_entry order_lists[] = {
497 { "proctree", &lock_class_sx },
498 { "allproc", &lock_class_sx },
499 { "allprison", &lock_class_sx },
504 { "Giant", &lock_class_mtx_sleep },
505 { "pipe mutex", &lock_class_mtx_sleep },
506 { "sigio lock", &lock_class_mtx_sleep },
507 { "process group", &lock_class_mtx_sleep },
509 { "pmc-sleep", &lock_class_mtx_sleep },
511 { "process lock", &lock_class_mtx_sleep },
512 { "session", &lock_class_mtx_sleep },
513 { "uidinfo hash", &lock_class_rw },
514 { "time lock", &lock_class_mtx_sleep },
519 { "umtx lock", &lock_class_mtx_sleep },
524 { "accept", &lock_class_mtx_sleep },
525 { "so_snd", &lock_class_mtx_sleep },
526 { "so_rcv", &lock_class_mtx_sleep },
527 { "sellck", &lock_class_mtx_sleep },
532 { "so_rcv", &lock_class_mtx_sleep },
533 { "radix node head", &lock_class_rm },
534 { "ifaddr", &lock_class_mtx_sleep },
538 * protocol locks before interface locks, after UDP locks.
540 { "in_multi_sx", &lock_class_sx },
541 { "udpinp", &lock_class_rw },
542 { "in_multi_list_mtx", &lock_class_mtx_sleep },
543 { "igmp_mtx", &lock_class_mtx_sleep },
544 { "if_addr_lock", &lock_class_mtx_sleep },
548 * protocol locks before interface locks, after UDP locks.
550 { "in6_multi_sx", &lock_class_sx },
551 { "udpinp", &lock_class_rw },
552 { "in6_multi_list_mtx", &lock_class_mtx_sleep },
553 { "mld_mtx", &lock_class_mtx_sleep },
554 { "if_addr_lock", &lock_class_mtx_sleep },
557 * UNIX Domain Sockets
559 { "unp_link_rwlock", &lock_class_rw },
560 { "unp_list_lock", &lock_class_mtx_sleep },
561 { "unp", &lock_class_mtx_sleep },
562 { "so_snd", &lock_class_mtx_sleep },
567 { "udpinp", &lock_class_rw },
568 { "udp", &lock_class_mtx_sleep },
569 { "so_snd", &lock_class_mtx_sleep },
574 { "tcpinp", &lock_class_rw },
575 { "tcp", &lock_class_mtx_sleep },
576 { "so_snd", &lock_class_mtx_sleep },
581 { "bpf global lock", &lock_class_sx },
582 { "bpf cdev lock", &lock_class_mtx_sleep },
587 { "nfsd_mtx", &lock_class_mtx_sleep },
588 { "so_snd", &lock_class_mtx_sleep },
594 { "802.11 com lock", &lock_class_mtx_sleep},
599 { "network driver", &lock_class_mtx_sleep},
605 { "ng_node", &lock_class_mtx_sleep },
606 { "ng_worklist", &lock_class_mtx_sleep },
611 { "vm map (system)", &lock_class_mtx_sleep },
612 { "vnode interlock", &lock_class_mtx_sleep },
613 { "cdev", &lock_class_mtx_sleep },
614 { "devthrd", &lock_class_mtx_sleep },
619 { "vm map (user)", &lock_class_sx },
620 { "vm object", &lock_class_rw },
621 { "vm page", &lock_class_mtx_sleep },
622 { "pmap pv global", &lock_class_rw },
623 { "pmap", &lock_class_mtx_sleep },
624 { "pmap pv list", &lock_class_rw },
625 { "vm page free queue", &lock_class_mtx_sleep },
626 { "vm pagequeue", &lock_class_mtx_sleep },
629 * kqueue/VFS interaction
631 { "kqueue", &lock_class_mtx_sleep },
632 { "struct mount mtx", &lock_class_mtx_sleep },
633 { "vnode interlock", &lock_class_mtx_sleep },
638 { "ncvn", &lock_class_mtx_sleep },
639 { "ncbuc", &lock_class_mtx_sleep },
640 { "vnode interlock", &lock_class_mtx_sleep },
641 { "ncneg", &lock_class_mtx_sleep },
646 { "dn->dn_mtx", &lock_class_sx },
647 { "dr->dt.di.dr_mtx", &lock_class_sx },
648 { "db->db_mtx", &lock_class_sx },
653 { "TCP ID tree", &lock_class_rw },
654 { "tcp log id bucket", &lock_class_mtx_sleep },
655 { "tcpinp", &lock_class_rw },
656 { "TCP log expireq", &lock_class_mtx_sleep },
662 { "ap boot", &lock_class_mtx_spin },
664 { "rm.mutex_mtx", &lock_class_mtx_spin },
666 { "cy", &lock_class_mtx_spin },
668 { "scc_hwmtx", &lock_class_mtx_spin },
669 { "uart_hwmtx", &lock_class_mtx_spin },
670 { "fast_taskqueue", &lock_class_mtx_spin },
671 { "intr table", &lock_class_mtx_spin },
672 { "process slock", &lock_class_mtx_spin },
673 { "syscons video lock", &lock_class_mtx_spin },
674 { "sleepq chain", &lock_class_mtx_spin },
675 { "rm_spinlock", &lock_class_mtx_spin },
676 { "turnstile chain", &lock_class_mtx_spin },
677 { "turnstile lock", &lock_class_mtx_spin },
678 { "sched lock", &lock_class_mtx_spin },
679 { "td_contested", &lock_class_mtx_spin },
680 { "callout", &lock_class_mtx_spin },
681 { "entropy harvest mutex", &lock_class_mtx_spin },
683 { "smp rendezvous", &lock_class_mtx_spin },
686 { "tlb0", &lock_class_mtx_spin },
689 { "sched lock", &lock_class_mtx_spin },
691 { "pmc-per-proc", &lock_class_mtx_spin },
697 { "intrcnt", &lock_class_mtx_spin },
698 { "icu", &lock_class_mtx_spin },
700 { "allpmaps", &lock_class_mtx_spin },
701 { "descriptor tables", &lock_class_mtx_spin },
703 { "clk", &lock_class_mtx_spin },
704 { "cpuset", &lock_class_mtx_spin },
705 { "mprof lock", &lock_class_mtx_spin },
706 { "zombie lock", &lock_class_mtx_spin },
707 { "ALD Queue", &lock_class_mtx_spin },
708 #if defined(__i386__) || defined(__amd64__)
709 { "pcicfg", &lock_class_mtx_spin },
710 { "NDIS thread lock", &lock_class_mtx_spin },
712 { "tw_osl_io_lock", &lock_class_mtx_spin },
713 { "tw_osl_q_lock", &lock_class_mtx_spin },
714 { "tw_cl_io_lock", &lock_class_mtx_spin },
715 { "tw_cl_intr_lock", &lock_class_mtx_spin },
716 { "tw_cl_gen_lock", &lock_class_mtx_spin },
718 { "pmc-leaf", &lock_class_mtx_spin },
720 { "blocked lock", &lock_class_mtx_spin },
726 * Pairs of locks which have been blessed. Witness does not complain about
727 * order problems with blessed lock pairs. Please do not add an entry to the
728 * table without an explanatory comment.
730 static struct witness_blessed blessed_list[] = {
732 * See the comment in ufs_dirhash.c. Basically, a vnode lock serializes
733 * both lock orders, so a deadlock cannot happen as a result of this
736 { "dirhash", "bufwait" },
739 * A UFS vnode may be locked in vget() while a buffer belonging to the
740 * parent directory vnode is locked.
742 { "ufs", "bufwait" },
745 * The tarfs decompression stream vnode may be locked while a
746 * buffer belonging to a tarfs data vnode is locked.
748 { "tarfs", "bufwait" },
752 * This global is set to 0 once it becomes safe to use the witness code.
754 static int witness_cold = 1;
757 * This global is set to 1 once the static lock orders have been enrolled
758 * so that a warning can be issued for any spin locks enrolled later.
760 static int witness_spin_warn = 0;
762 /* Trim useless garbage from filenames. */
764 fixup_filename(const char *file)
769 while (strncmp(file, "../", 3) == 0)
775 * Calculate the size of early witness structures.
778 witness_startup_count(void)
782 sz = sizeof(struct witness) * witness_count;
783 sz += sizeof(*w_rmatrix) * (witness_count + 1);
784 sz += sizeof(*w_rmatrix[0]) * (witness_count + 1) *
791 * The WITNESS-enabled diagnostic code. Note that the witness code does
792 * assume that the early boot is single-threaded at least until after this
793 * routine is completed.
796 witness_startup(void *mem)
798 struct lock_object *lock;
799 struct witness_order_list_entry *order;
800 struct witness *w, *w1;
806 p += sizeof(struct witness) * witness_count;
808 w_rmatrix = (void *)p;
809 p += sizeof(*w_rmatrix) * (witness_count + 1);
811 for (i = 0; i < witness_count + 1; i++) {
812 w_rmatrix[i] = (void *)p;
813 p += sizeof(*w_rmatrix[i]) * (witness_count + 1);
815 badstack_sbuf_size = witness_count * 256;
818 * We have to release Giant before initializing its witness
819 * structure so that WITNESS doesn't get confused.
822 mtx_assert(&Giant, MA_NOTOWNED);
824 CTR1(KTR_WITNESS, "%s: initializing witness", __func__);
825 mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET |
826 MTX_NOWITNESS | MTX_NOPROFILE);
827 for (i = witness_count - 1; i >= 0; i--) {
829 memset(w, 0, sizeof(*w));
830 w_data[i].w_index = i; /* Witness index never changes. */
833 KASSERT(STAILQ_FIRST(&w_free)->w_index == 0,
834 ("%s: Invalid list of free witness objects", __func__));
836 /* Witness with index 0 is not used to aid in debugging. */
837 STAILQ_REMOVE_HEAD(&w_free, w_list);
840 for (i = 0; i < witness_count; i++) {
841 memset(w_rmatrix[i], 0, sizeof(*w_rmatrix[i]) *
842 (witness_count + 1));
845 for (i = 0; i < LOCK_CHILDCOUNT; i++)
846 witness_lock_list_free(&w_locklistdata[i]);
847 witness_init_hash_tables();
849 /* First add in all the specified order lists. */
850 for (order = order_lists; order->w_name != NULL; order++) {
851 w = enroll(order->w_name, order->w_class);
854 w->w_file = "order list";
855 for (order++; order->w_name != NULL; order++) {
856 w1 = enroll(order->w_name, order->w_class);
859 w1->w_file = "order list";
864 witness_spin_warn = 1;
866 /* Iterate through all locks and add them to witness. */
867 for (i = 0; pending_locks[i].wh_lock != NULL; i++) {
868 lock = pending_locks[i].wh_lock;
869 KASSERT(lock->lo_flags & LO_WITNESS,
870 ("%s: lock %s is on pending list but not LO_WITNESS",
871 __func__, lock->lo_name));
872 lock->lo_witness = enroll(pending_locks[i].wh_type,
876 /* Mark the witness code as being ready for use. */
883 witness_init(struct lock_object *lock, const char *type)
885 struct lock_class *class;
887 /* Various sanity checks. */
888 class = LOCK_CLASS(lock);
889 if ((lock->lo_flags & LO_RECURSABLE) != 0 &&
890 (class->lc_flags & LC_RECURSABLE) == 0)
891 kassert_panic("%s: lock (%s) %s can not be recursable",
892 __func__, class->lc_name, lock->lo_name);
893 if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
894 (class->lc_flags & LC_SLEEPABLE) == 0)
895 kassert_panic("%s: lock (%s) %s can not be sleepable",
896 __func__, class->lc_name, lock->lo_name);
897 if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
898 (class->lc_flags & LC_UPGRADABLE) == 0)
899 kassert_panic("%s: lock (%s) %s can not be upgradable",
900 __func__, class->lc_name, lock->lo_name);
903 * If we shouldn't watch this lock, then just clear lo_witness.
904 * Otherwise, if witness_cold is set, then it is too early to
905 * enroll this lock, so defer it to witness_initialize() by adding
906 * it to the pending_locks list. If it is not too early, then enroll
909 if (witness_watch < 1 || KERNEL_PANICKED() ||
910 (lock->lo_flags & LO_WITNESS) == 0)
911 lock->lo_witness = NULL;
912 else if (witness_cold) {
913 pending_locks[pending_cnt].wh_lock = lock;
914 pending_locks[pending_cnt++].wh_type = type;
915 if (pending_cnt > WITNESS_PENDLIST)
916 panic("%s: pending locks list is too small, "
917 "increase WITNESS_PENDLIST\n",
920 lock->lo_witness = enroll(type, class);
924 witness_destroy(struct lock_object *lock)
926 struct lock_class *class;
929 class = LOCK_CLASS(lock);
932 panic("lock (%s) %s destroyed while witness_cold",
933 class->lc_name, lock->lo_name);
935 /* XXX: need to verify that no one holds the lock */
936 if ((lock->lo_flags & LO_WITNESS) == 0 || lock->lo_witness == NULL)
938 w = lock->lo_witness;
940 mtx_lock_spin(&w_mtx);
941 MPASS(w->w_refcount > 0);
944 if (w->w_refcount == 0)
946 mtx_unlock_spin(&w_mtx);
951 witness_ddb_compute_levels(void)
956 * First clear all levels.
958 STAILQ_FOREACH(w, &w_all, w_list)
962 * Look for locks with no parents and level all their descendants.
964 STAILQ_FOREACH(w, &w_all, w_list) {
965 /* If the witness has ancestors (is not a root), skip it. */
966 if (w->w_num_ancestors > 0)
968 witness_ddb_level_descendants(w, 0);
973 witness_ddb_level_descendants(struct witness *w, int l)
977 if (w->w_ddb_level >= l)
983 for (i = 1; i <= w_max_used_index; i++) {
984 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
985 witness_ddb_level_descendants(&w_data[i], l);
990 witness_ddb_display_descendants(int(*prnt)(const char *fmt, ...),
991 struct witness *w, int indent)
995 for (i = 0; i < indent; i++)
997 prnt("%s (type: %s, depth: %d, active refs: %d)",
998 w->w_name, w->w_class->lc_name,
999 w->w_ddb_level, w->w_refcount);
1000 if (w->w_displayed) {
1001 prnt(" -- (already displayed)\n");
1005 if (w->w_file != NULL && w->w_line != 0)
1006 prnt(" -- last acquired @ %s:%d\n", fixup_filename(w->w_file),
1009 prnt(" -- never acquired\n");
1011 WITNESS_INDEX_ASSERT(w->w_index);
1012 for (i = 1; i <= w_max_used_index; i++) {
1015 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
1016 witness_ddb_display_descendants(prnt, &w_data[i],
1022 witness_ddb_display_list(int(*prnt)(const char *fmt, ...),
1023 struct witness_list *list)
1027 STAILQ_FOREACH(w, list, w_typelist) {
1028 if (w->w_file == NULL || w->w_ddb_level > 0)
1031 /* This lock has no anscestors - display its descendants. */
1032 witness_ddb_display_descendants(prnt, w, 0);
1039 witness_ddb_display(int(*prnt)(const char *fmt, ...))
1043 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1044 witness_ddb_compute_levels();
1046 /* Clear all the displayed flags. */
1047 STAILQ_FOREACH(w, &w_all, w_list)
1051 * First, handle sleep locks which have been acquired at least
1054 prnt("Sleep locks:\n");
1055 witness_ddb_display_list(prnt, &w_sleep);
1060 * Now do spin locks which have been acquired at least once.
1062 prnt("\nSpin locks:\n");
1063 witness_ddb_display_list(prnt, &w_spin);
1068 * Finally, any locks which have not been acquired yet.
1070 prnt("\nLocks which were never acquired:\n");
1071 STAILQ_FOREACH(w, &w_all, w_list) {
1072 if (w->w_file != NULL || w->w_refcount == 0)
1074 prnt("%s (type: %s, depth: %d)\n", w->w_name,
1075 w->w_class->lc_name, w->w_ddb_level);
1083 witness_defineorder(struct lock_object *lock1, struct lock_object *lock2)
1086 if (witness_watch == -1 || KERNEL_PANICKED())
1089 /* Require locks that witness knows about. */
1090 if (lock1 == NULL || lock1->lo_witness == NULL || lock2 == NULL ||
1091 lock2->lo_witness == NULL)
1094 mtx_assert(&w_mtx, MA_NOTOWNED);
1095 mtx_lock_spin(&w_mtx);
1098 * If we already have either an explicit or implied lock order that
1099 * is the other way around, then return an error.
1101 if (witness_watch &&
1102 isitmydescendant(lock2->lo_witness, lock1->lo_witness)) {
1103 mtx_unlock_spin(&w_mtx);
1107 /* Try to add the new order. */
1108 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1109 lock2->lo_witness->w_name, lock1->lo_witness->w_name);
1110 itismychild(lock1->lo_witness, lock2->lo_witness);
1111 mtx_unlock_spin(&w_mtx);
1116 witness_checkorder(struct lock_object *lock, int flags, const char *file,
1117 int line, struct lock_object *interlock)
1119 struct lock_list_entry *lock_list, *lle;
1120 struct lock_instance *lock1, *lock2, *plock;
1121 struct lock_class *class, *iclass;
1122 struct witness *w, *w1;
1126 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL ||
1130 w = lock->lo_witness;
1131 class = LOCK_CLASS(lock);
1134 if (class->lc_flags & LC_SLEEPLOCK) {
1136 * Since spin locks include a critical section, this check
1137 * implicitly enforces a lock order of all sleep locks before
1140 if (td->td_critnest != 0 && !kdb_active)
1141 kassert_panic("acquiring blockable sleep lock with "
1142 "spinlock or critical section held (%s) %s @ %s:%d",
1143 class->lc_name, lock->lo_name,
1144 fixup_filename(file), line);
1147 * If this is the first lock acquired then just return as
1148 * no order checking is needed.
1150 lock_list = td->td_sleeplocks;
1151 if (lock_list == NULL || lock_list->ll_count == 0)
1155 * If this is the first lock, just return as no order
1156 * checking is needed. Avoid problems with thread
1157 * migration pinning the thread while checking if
1158 * spinlocks are held. If at least one spinlock is held
1159 * the thread is in a safe path and it is allowed to
1163 lock_list = PCPU_GET(spinlocks);
1164 if (lock_list == NULL || lock_list->ll_count == 0) {
1172 * Check to see if we are recursing on a lock we already own. If
1173 * so, make sure that we don't mismatch exclusive and shared lock
1176 lock1 = find_instance(lock_list, lock);
1177 if (lock1 != NULL) {
1178 if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
1179 (flags & LOP_EXCLUSIVE) == 0) {
1180 witness_output("shared lock of (%s) %s @ %s:%d\n",
1181 class->lc_name, lock->lo_name,
1182 fixup_filename(file), line);
1183 witness_output("while exclusively locked from %s:%d\n",
1184 fixup_filename(lock1->li_file), lock1->li_line);
1185 kassert_panic("excl->share");
1187 if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
1188 (flags & LOP_EXCLUSIVE) != 0) {
1189 witness_output("exclusive lock of (%s) %s @ %s:%d\n",
1190 class->lc_name, lock->lo_name,
1191 fixup_filename(file), line);
1192 witness_output("while share locked from %s:%d\n",
1193 fixup_filename(lock1->li_file), lock1->li_line);
1194 kassert_panic("share->excl");
1199 /* Warn if the interlock is not locked exactly once. */
1200 if (interlock != NULL) {
1201 iclass = LOCK_CLASS(interlock);
1202 lock1 = find_instance(lock_list, interlock);
1204 kassert_panic("interlock (%s) %s not locked @ %s:%d",
1205 iclass->lc_name, interlock->lo_name,
1206 fixup_filename(file), line);
1207 else if ((lock1->li_flags & LI_RECURSEMASK) != 0)
1208 kassert_panic("interlock (%s) %s recursed @ %s:%d",
1209 iclass->lc_name, interlock->lo_name,
1210 fixup_filename(file), line);
1214 * Find the previously acquired lock, but ignore interlocks.
1216 plock = &lock_list->ll_children[lock_list->ll_count - 1];
1217 if (interlock != NULL && plock->li_lock == interlock) {
1218 if (lock_list->ll_count > 1)
1220 &lock_list->ll_children[lock_list->ll_count - 2];
1222 lle = lock_list->ll_next;
1225 * The interlock is the only lock we hold, so
1230 plock = &lle->ll_children[lle->ll_count - 1];
1235 * Try to perform most checks without a lock. If this succeeds we
1236 * can skip acquiring the lock and return success. Otherwise we redo
1237 * the check with the lock held to handle races with concurrent updates.
1239 w1 = plock->li_lock->lo_witness;
1240 if (witness_lock_order_check(w1, w))
1243 mtx_lock_spin(&w_mtx);
1244 if (witness_lock_order_check(w1, w)) {
1245 mtx_unlock_spin(&w_mtx);
1248 witness_lock_order_add(w1, w);
1251 * Check for duplicate locks of the same type. Note that we only
1252 * have to check for this on the last lock we just acquired. Any
1253 * other cases will be caught as lock order violations.
1257 if (!(lock->lo_flags & LO_DUPOK) && !(flags & LOP_DUPOK) &&
1258 !(w_rmatrix[i][i] & WITNESS_REVERSAL)) {
1259 w_rmatrix[i][i] |= WITNESS_REVERSAL;
1261 mtx_unlock_spin(&w_mtx);
1263 "acquiring duplicate lock of same type: \"%s\"\n",
1265 witness_output(" 1st %s @ %s:%d\n", plock->li_lock->lo_name,
1266 fixup_filename(plock->li_file), plock->li_line);
1267 witness_output(" 2nd %s @ %s:%d\n", lock->lo_name,
1268 fixup_filename(file), line);
1269 witness_debugger(1, __func__);
1271 mtx_unlock_spin(&w_mtx);
1274 mtx_assert(&w_mtx, MA_OWNED);
1277 * If we know that the lock we are acquiring comes after
1278 * the lock we most recently acquired in the lock order tree,
1279 * then there is no need for any further checks.
1281 if (isitmychild(w1, w))
1284 for (j = 0, lle = lock_list; lle != NULL; lle = lle->ll_next) {
1285 for (i = lle->ll_count - 1; i >= 0; i--, j++) {
1286 struct stack pstack;
1287 bool pstackv, trace;
1289 MPASS(j < LOCK_CHILDCOUNT * LOCK_NCHILDREN);
1290 lock1 = &lle->ll_children[i];
1293 * Ignore the interlock.
1295 if (interlock == lock1->li_lock)
1299 * If this lock doesn't undergo witness checking,
1302 w1 = lock1->li_lock->lo_witness;
1304 KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
1305 ("lock missing witness structure"));
1310 * If we are locking Giant and this is a sleepable
1311 * lock, then skip it.
1313 if ((lock1->li_flags & LI_SLEEPABLE) != 0 &&
1314 lock == &Giant.lock_object)
1318 * If we are locking a sleepable lock and this lock
1319 * is Giant, then skip it.
1321 if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1322 (flags & LOP_NOSLEEP) == 0 &&
1323 lock1->li_lock == &Giant.lock_object)
1327 * If we are locking a sleepable lock and this lock
1328 * isn't sleepable, we want to treat it as a lock
1329 * order violation to enfore a general lock order of
1330 * sleepable locks before non-sleepable locks.
1332 if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1333 (flags & LOP_NOSLEEP) == 0 &&
1334 (lock1->li_flags & LI_SLEEPABLE) == 0)
1338 * If we are locking Giant and this is a non-sleepable
1339 * lock, then treat it as a reversal.
1341 if ((lock1->li_flags & LI_SLEEPABLE) == 0 &&
1342 lock == &Giant.lock_object)
1346 * Check the lock order hierarchy for a reveresal.
1348 if (!isitmydescendant(w, w1))
1353 * We have a lock order violation, check to see if it
1354 * is allowed or has already been yelled about.
1357 /* Bail if this violation is known */
1358 if (w_rmatrix[w1->w_index][w->w_index] & WITNESS_REVERSAL)
1361 /* Record this as a violation */
1362 w_rmatrix[w1->w_index][w->w_index] |= WITNESS_REVERSAL;
1363 w_rmatrix[w->w_index][w1->w_index] |= WITNESS_REVERSAL;
1364 w->w_reversed = w1->w_reversed = 1;
1365 witness_increment_graph_generation();
1368 * If the lock order is blessed, bail before logging
1369 * anything. We don't look for other lock order
1370 * violations though, which may be a bug.
1375 trace = atomic_load_int(&witness_trace);
1377 struct witness_lock_order_data *data;
1380 data = witness_lock_order_get(w, w1);
1382 stack_copy(&data->wlod_stack,
1387 mtx_unlock_spin(&w_mtx);
1389 #ifdef WITNESS_NO_VNODE
1391 * There are known LORs between VNODE locks. They are
1392 * not an indication of a bug. VNODE locks are flagged
1393 * as such (LO_IS_VNODE) and we don't yell if the LOR
1394 * is between 2 VNODE locks.
1396 if ((lock->lo_flags & LO_IS_VNODE) != 0 &&
1397 (lock1->li_lock->lo_flags & LO_IS_VNODE) != 0)
1402 * Ok, yell about it.
1404 if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1405 (flags & LOP_NOSLEEP) == 0 &&
1406 (lock1->li_flags & LI_SLEEPABLE) == 0)
1408 "lock order reversal: (sleepable after non-sleepable)\n");
1409 else if ((lock1->li_flags & LI_SLEEPABLE) == 0
1410 && lock == &Giant.lock_object)
1412 "lock order reversal: (Giant after non-sleepable)\n");
1414 witness_output("lock order reversal:\n");
1417 * Try to locate an earlier lock with
1418 * witness w in our list.
1421 lock2 = &lle->ll_children[i];
1422 MPASS(lock2->li_lock != NULL);
1423 if (lock2->li_lock->lo_witness == w)
1425 if (i == 0 && lle->ll_next != NULL) {
1427 i = lle->ll_count - 1;
1428 MPASS(i >= 0 && i < LOCK_NCHILDREN);
1433 witness_output(" 1st %p %s (%s, %s) @ %s:%d\n",
1434 lock1->li_lock, lock1->li_lock->lo_name,
1435 w1->w_name, w1->w_class->lc_name,
1436 fixup_filename(lock1->li_file),
1438 witness_output(" 2nd %p %s (%s, %s) @ %s:%d\n",
1439 lock, lock->lo_name, w->w_name,
1440 w->w_class->lc_name, fixup_filename(file),
1443 struct witness *w2 = lock2->li_lock->lo_witness;
1445 witness_output(" 1st %p %s (%s, %s) @ %s:%d\n",
1446 lock2->li_lock, lock2->li_lock->lo_name,
1447 w2->w_name, w2->w_class->lc_name,
1448 fixup_filename(lock2->li_file),
1450 witness_output(" 2nd %p %s (%s, %s) @ %s:%d\n",
1451 lock1->li_lock, lock1->li_lock->lo_name,
1452 w1->w_name, w1->w_class->lc_name,
1453 fixup_filename(lock1->li_file),
1455 witness_output(" 3rd %p %s (%s, %s) @ %s:%d\n", lock,
1456 lock->lo_name, w->w_name,
1457 w->w_class->lc_name, fixup_filename(file),
1464 sbuf_new(&sb, buf, sizeof(buf), SBUF_FIXEDLEN);
1465 sbuf_set_drain(&sb, witness_output_drain,
1470 "lock order %s -> %s established at:\n",
1471 w->w_name, w1->w_name);
1472 stack_sbuf_print_flags(&sb, &pstack,
1473 M_NOWAIT, STACK_SBUF_FMT_LONG);
1477 "lock order %s -> %s attempted at:\n",
1478 w1->w_name, w->w_name);
1479 stack_save(&pstack);
1480 stack_sbuf_print_flags(&sb, &pstack, M_NOWAIT,
1481 STACK_SBUF_FMT_LONG);
1486 witness_enter_debugger(__func__);
1492 * If requested, build a new lock order. However, don't build a new
1493 * relationship between a sleepable lock and Giant if it is in the
1494 * wrong direction. The correct lock order is that sleepable locks
1495 * always come before Giant.
1497 if (flags & LOP_NEWORDER &&
1498 !(plock->li_lock == &Giant.lock_object &&
1499 (lock->lo_flags & LO_SLEEPABLE) != 0 &&
1500 (flags & LOP_NOSLEEP) == 0)) {
1501 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1502 w->w_name, plock->li_lock->lo_witness->w_name);
1503 itismychild(plock->li_lock->lo_witness, w);
1506 mtx_unlock_spin(&w_mtx);
1510 witness_lock(struct lock_object *lock, int flags, const char *file, int line)
1512 struct lock_list_entry **lock_list, *lle;
1513 struct lock_instance *instance;
1517 if (witness_cold || witness_watch == -1 || lock->lo_witness == NULL ||
1520 w = lock->lo_witness;
1523 /* Determine lock list for this lock. */
1524 if (LOCK_CLASS(lock)->lc_flags & LC_SLEEPLOCK)
1525 lock_list = &td->td_sleeplocks;
1527 lock_list = PCPU_PTR(spinlocks);
1529 /* Check to see if we are recursing on a lock we already own. */
1530 instance = find_instance(*lock_list, lock);
1531 if (instance != NULL) {
1532 instance->li_flags++;
1533 CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
1534 td->td_proc->p_pid, lock->lo_name,
1535 instance->li_flags & LI_RECURSEMASK);
1536 instance->li_file = file;
1537 instance->li_line = line;
1541 /* Update per-witness last file and line acquire. */
1545 /* Find the next open lock instance in the list and fill it. */
1547 if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
1548 lle = witness_lock_list_get();
1551 lle->ll_next = *lock_list;
1552 CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
1553 td->td_proc->p_pid, lle);
1556 instance = &lle->ll_children[lle->ll_count++];
1557 instance->li_lock = lock;
1558 instance->li_line = line;
1559 instance->li_file = file;
1560 instance->li_flags = 0;
1561 if ((flags & LOP_EXCLUSIVE) != 0)
1562 instance->li_flags |= LI_EXCLUSIVE;
1563 if ((lock->lo_flags & LO_SLEEPABLE) != 0 && (flags & LOP_NOSLEEP) == 0)
1564 instance->li_flags |= LI_SLEEPABLE;
1565 CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
1566 td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
1570 witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
1572 struct lock_instance *instance;
1573 struct lock_class *class;
1575 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1576 if (lock->lo_witness == NULL || witness_watch == -1 || KERNEL_PANICKED())
1578 class = LOCK_CLASS(lock);
1579 if (witness_watch) {
1580 if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1582 "upgrade of non-upgradable lock (%s) %s @ %s:%d",
1583 class->lc_name, lock->lo_name,
1584 fixup_filename(file), line);
1585 if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1587 "upgrade of non-sleep lock (%s) %s @ %s:%d",
1588 class->lc_name, lock->lo_name,
1589 fixup_filename(file), line);
1591 instance = find_instance(curthread->td_sleeplocks, lock);
1592 if (instance == NULL) {
1593 kassert_panic("upgrade of unlocked lock (%s) %s @ %s:%d",
1594 class->lc_name, lock->lo_name,
1595 fixup_filename(file), line);
1598 if (witness_watch) {
1599 if ((instance->li_flags & LI_EXCLUSIVE) != 0)
1601 "upgrade of exclusive lock (%s) %s @ %s:%d",
1602 class->lc_name, lock->lo_name,
1603 fixup_filename(file), line);
1604 if ((instance->li_flags & LI_RECURSEMASK) != 0)
1606 "upgrade of recursed lock (%s) %s r=%d @ %s:%d",
1607 class->lc_name, lock->lo_name,
1608 instance->li_flags & LI_RECURSEMASK,
1609 fixup_filename(file), line);
1611 instance->li_flags |= LI_EXCLUSIVE;
1615 witness_downgrade(struct lock_object *lock, int flags, const char *file,
1618 struct lock_instance *instance;
1619 struct lock_class *class;
1621 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1622 if (lock->lo_witness == NULL || witness_watch == -1 || KERNEL_PANICKED())
1624 class = LOCK_CLASS(lock);
1625 if (witness_watch) {
1626 if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1628 "downgrade of non-upgradable lock (%s) %s @ %s:%d",
1629 class->lc_name, lock->lo_name,
1630 fixup_filename(file), line);
1631 if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1633 "downgrade of non-sleep lock (%s) %s @ %s:%d",
1634 class->lc_name, lock->lo_name,
1635 fixup_filename(file), line);
1637 instance = find_instance(curthread->td_sleeplocks, lock);
1638 if (instance == NULL) {
1639 kassert_panic("downgrade of unlocked lock (%s) %s @ %s:%d",
1640 class->lc_name, lock->lo_name,
1641 fixup_filename(file), line);
1644 if (witness_watch) {
1645 if ((instance->li_flags & LI_EXCLUSIVE) == 0)
1647 "downgrade of shared lock (%s) %s @ %s:%d",
1648 class->lc_name, lock->lo_name,
1649 fixup_filename(file), line);
1650 if ((instance->li_flags & LI_RECURSEMASK) != 0)
1652 "downgrade of recursed lock (%s) %s r=%d @ %s:%d",
1653 class->lc_name, lock->lo_name,
1654 instance->li_flags & LI_RECURSEMASK,
1655 fixup_filename(file), line);
1657 instance->li_flags &= ~LI_EXCLUSIVE;
1661 witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
1663 struct lock_list_entry **lock_list, *lle;
1664 struct lock_instance *instance;
1665 struct lock_class *class;
1670 if (witness_cold || lock->lo_witness == NULL || KERNEL_PANICKED())
1673 class = LOCK_CLASS(lock);
1675 /* Find lock instance associated with this lock. */
1676 if (class->lc_flags & LC_SLEEPLOCK)
1677 lock_list = &td->td_sleeplocks;
1679 lock_list = PCPU_PTR(spinlocks);
1681 for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
1682 for (i = 0; i < (*lock_list)->ll_count; i++) {
1683 instance = &(*lock_list)->ll_children[i];
1684 if (instance->li_lock == lock)
1689 * When disabling WITNESS through witness_watch we could end up in
1690 * having registered locks in the td_sleeplocks queue.
1691 * We have to make sure we flush these queues, so just search for
1692 * eventual register locks and remove them.
1694 if (witness_watch > 0) {
1695 kassert_panic("lock (%s) %s not locked @ %s:%d", class->lc_name,
1696 lock->lo_name, fixup_filename(file), line);
1703 /* First, check for shared/exclusive mismatches. */
1704 if ((instance->li_flags & LI_EXCLUSIVE) != 0 && witness_watch > 0 &&
1705 (flags & LOP_EXCLUSIVE) == 0) {
1706 witness_output("shared unlock of (%s) %s @ %s:%d\n",
1707 class->lc_name, lock->lo_name, fixup_filename(file), line);
1708 witness_output("while exclusively locked from %s:%d\n",
1709 fixup_filename(instance->li_file), instance->li_line);
1710 kassert_panic("excl->ushare");
1712 if ((instance->li_flags & LI_EXCLUSIVE) == 0 && witness_watch > 0 &&
1713 (flags & LOP_EXCLUSIVE) != 0) {
1714 witness_output("exclusive unlock of (%s) %s @ %s:%d\n",
1715 class->lc_name, lock->lo_name, fixup_filename(file), line);
1716 witness_output("while share locked from %s:%d\n",
1717 fixup_filename(instance->li_file),
1719 kassert_panic("share->uexcl");
1721 /* If we are recursed, unrecurse. */
1722 if ((instance->li_flags & LI_RECURSEMASK) > 0) {
1723 CTR4(KTR_WITNESS, "%s: pid %d unrecursed on %s r=%d", __func__,
1724 td->td_proc->p_pid, instance->li_lock->lo_name,
1725 instance->li_flags);
1726 instance->li_flags--;
1729 /* The lock is now being dropped, check for NORELEASE flag */
1730 if ((instance->li_flags & LI_NORELEASE) != 0 && witness_watch > 0) {
1731 witness_output("forbidden unlock of (%s) %s @ %s:%d\n",
1732 class->lc_name, lock->lo_name, fixup_filename(file), line);
1733 kassert_panic("lock marked norelease");
1736 /* Otherwise, remove this item from the list. */
1738 CTR4(KTR_WITNESS, "%s: pid %d removed %s from lle[%d]", __func__,
1739 td->td_proc->p_pid, instance->li_lock->lo_name,
1740 (*lock_list)->ll_count - 1);
1741 for (j = i; j < (*lock_list)->ll_count - 1; j++)
1742 (*lock_list)->ll_children[j] =
1743 (*lock_list)->ll_children[j + 1];
1744 (*lock_list)->ll_count--;
1748 * In order to reduce contention on w_mtx, we want to keep always an
1749 * head object into lists so that frequent allocation from the
1750 * free witness pool (and subsequent locking) is avoided.
1751 * In order to maintain the current code simple, when the head
1752 * object is totally unloaded it means also that we do not have
1753 * further objects in the list, so the list ownership needs to be
1754 * hand over to another object if the current head needs to be freed.
1756 if ((*lock_list)->ll_count == 0) {
1757 if (*lock_list == lle) {
1758 if (lle->ll_next == NULL)
1762 *lock_list = lle->ll_next;
1763 CTR3(KTR_WITNESS, "%s: pid %d removed lle %p", __func__,
1764 td->td_proc->p_pid, lle);
1765 witness_lock_list_free(lle);
1770 witness_thread_exit(struct thread *td)
1772 struct lock_list_entry *lle;
1775 lle = td->td_sleeplocks;
1776 if (lle == NULL || KERNEL_PANICKED())
1778 if (lle->ll_count != 0) {
1779 for (n = 0; lle != NULL; lle = lle->ll_next)
1780 for (i = lle->ll_count - 1; i >= 0; i--) {
1783 "Thread %p exiting with the following locks held:\n", td);
1785 witness_list_lock(&lle->ll_children[i],
1790 "Thread %p cannot exit while holding sleeplocks\n", td);
1792 witness_lock_list_free(lle);
1796 * Warn if any locks other than 'lock' are held. Flags can be passed in to
1797 * exempt Giant and sleepable locks from the checks as well. If any
1798 * non-exempt locks are held, then a supplied message is printed to the
1799 * output channel along with a list of the offending locks. If indicated in the
1800 * flags then a failure results in a panic as well.
1803 witness_warn(int flags, struct lock_object *lock, const char *fmt, ...)
1805 struct lock_list_entry *lock_list, *lle;
1806 struct lock_instance *lock1;
1811 if (witness_cold || witness_watch < 1 || KERNEL_PANICKED())
1815 for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next)
1816 for (i = lle->ll_count - 1; i >= 0; i--) {
1817 lock1 = &lle->ll_children[i];
1818 if (lock1->li_lock == lock)
1820 if (flags & WARN_GIANTOK &&
1821 lock1->li_lock == &Giant.lock_object)
1823 if (flags & WARN_SLEEPOK &&
1824 (lock1->li_flags & LI_SLEEPABLE) != 0)
1830 printf(" with the following %slocks held:\n",
1831 (flags & WARN_SLEEPOK) != 0 ?
1832 "non-sleepable " : "");
1835 witness_list_lock(lock1, printf);
1839 * Pin the thread in order to avoid problems with thread migration.
1840 * Once that all verifies are passed about spinlocks ownership,
1841 * the thread is in a safe path and it can be unpinned.
1844 lock_list = PCPU_GET(spinlocks);
1845 if (lock_list != NULL && lock_list->ll_count != 0) {
1849 * We should only have one spinlock and as long as
1850 * the flags cannot match for this locks class,
1851 * check if the first spinlock is the one curthread
1854 lock1 = &lock_list->ll_children[lock_list->ll_count - 1];
1855 if (lock_list->ll_count == 1 && lock_list->ll_next == NULL &&
1856 lock1->li_lock == lock && n == 0)
1862 printf(" with the following %slocks held:\n",
1863 (flags & WARN_SLEEPOK) != 0 ? "non-sleepable " : "");
1864 n += witness_list_locks(&lock_list, printf);
1867 if (flags & WARN_PANIC && n)
1868 kassert_panic("%s", __func__);
1870 witness_debugger(n, __func__);
1875 witness_file(struct lock_object *lock)
1879 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1881 w = lock->lo_witness;
1886 witness_line(struct lock_object *lock)
1890 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1892 w = lock->lo_witness;
1896 static struct witness *
1897 enroll(const char *description, struct lock_class *lock_class)
1901 MPASS(description != NULL);
1903 if (witness_watch == -1 || KERNEL_PANICKED())
1905 if ((lock_class->lc_flags & LC_SPINLOCK)) {
1906 if (witness_skipspin)
1908 } else if ((lock_class->lc_flags & LC_SLEEPLOCK) == 0) {
1909 kassert_panic("lock class %s is not sleep or spin",
1910 lock_class->lc_name);
1914 mtx_lock_spin(&w_mtx);
1915 w = witness_hash_get(description);
1918 if ((w = witness_get()) == NULL)
1920 MPASS(strlen(description) < MAX_W_NAME);
1921 strcpy(w->w_name, description);
1922 w->w_class = lock_class;
1924 STAILQ_INSERT_HEAD(&w_all, w, w_list);
1925 if (lock_class->lc_flags & LC_SPINLOCK) {
1926 STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
1928 } else if (lock_class->lc_flags & LC_SLEEPLOCK) {
1929 STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
1933 /* Insert new witness into the hash */
1934 witness_hash_put(w);
1935 witness_increment_graph_generation();
1936 mtx_unlock_spin(&w_mtx);
1940 if (w->w_refcount == 1)
1941 w->w_class = lock_class;
1942 mtx_unlock_spin(&w_mtx);
1943 if (lock_class != w->w_class)
1945 "lock (%s) %s does not match earlier (%s) lock",
1946 description, lock_class->lc_name,
1947 w->w_class->lc_name);
1952 depart(struct witness *w)
1955 MPASS(w->w_refcount == 0);
1956 if (w->w_class->lc_flags & LC_SLEEPLOCK) {
1962 * Set file to NULL as it may point into a loadable module.
1966 witness_increment_graph_generation();
1970 adopt(struct witness *parent, struct witness *child)
1974 if (witness_cold == 0)
1975 mtx_assert(&w_mtx, MA_OWNED);
1977 /* If the relationship is already known, there's no work to be done. */
1978 if (isitmychild(parent, child))
1981 /* When the structure of the graph changes, bump up the generation. */
1982 witness_increment_graph_generation();
1985 * The hard part ... create the direct relationship, then propagate all
1986 * indirect relationships.
1988 pi = parent->w_index;
1989 ci = child->w_index;
1990 WITNESS_INDEX_ASSERT(pi);
1991 WITNESS_INDEX_ASSERT(ci);
1993 w_rmatrix[pi][ci] |= WITNESS_PARENT;
1994 w_rmatrix[ci][pi] |= WITNESS_CHILD;
1997 * If parent was not already an ancestor of child,
1998 * then we increment the descendant and ancestor counters.
2000 if ((w_rmatrix[pi][ci] & WITNESS_ANCESTOR) == 0) {
2001 parent->w_num_descendants++;
2002 child->w_num_ancestors++;
2006 * Find each ancestor of 'pi'. Note that 'pi' itself is counted as
2007 * an ancestor of 'pi' during this loop.
2009 for (i = 1; i <= w_max_used_index; i++) {
2010 if ((w_rmatrix[i][pi] & WITNESS_ANCESTOR_MASK) == 0 &&
2014 /* Find each descendant of 'i' and mark it as a descendant. */
2015 for (j = 1; j <= w_max_used_index; j++) {
2017 * Skip children that are already marked as
2018 * descendants of 'i'.
2020 if (w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK)
2024 * We are only interested in descendants of 'ci'. Note
2025 * that 'ci' itself is counted as a descendant of 'ci'.
2027 if ((w_rmatrix[ci][j] & WITNESS_ANCESTOR_MASK) == 0 &&
2030 w_rmatrix[i][j] |= WITNESS_ANCESTOR;
2031 w_rmatrix[j][i] |= WITNESS_DESCENDANT;
2032 w_data[i].w_num_descendants++;
2033 w_data[j].w_num_ancestors++;
2036 * Make sure we aren't marking a node as both an
2037 * ancestor and descendant. We should have caught
2038 * this as a lock order reversal earlier.
2040 if ((w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK) &&
2041 (w_rmatrix[i][j] & WITNESS_DESCENDANT_MASK)) {
2042 printf("witness rmatrix paradox! [%d][%d]=%d "
2043 "both ancestor and descendant\n",
2044 i, j, w_rmatrix[i][j]);
2046 printf("Witness disabled.\n");
2049 if ((w_rmatrix[j][i] & WITNESS_ANCESTOR_MASK) &&
2050 (w_rmatrix[j][i] & WITNESS_DESCENDANT_MASK)) {
2051 printf("witness rmatrix paradox! [%d][%d]=%d "
2052 "both ancestor and descendant\n",
2053 j, i, w_rmatrix[j][i]);
2055 printf("Witness disabled.\n");
2063 itismychild(struct witness *parent, struct witness *child)
2067 MPASS(child != NULL && parent != NULL);
2068 if (witness_cold == 0)
2069 mtx_assert(&w_mtx, MA_OWNED);
2071 if (!witness_lock_type_equal(parent, child)) {
2072 if (witness_cold == 0) {
2074 mtx_unlock_spin(&w_mtx);
2079 "%s: parent \"%s\" (%s) and child \"%s\" (%s) are not "
2080 "the same lock type", __func__, parent->w_name,
2081 parent->w_class->lc_name, child->w_name,
2082 child->w_class->lc_name);
2084 mtx_lock_spin(&w_mtx);
2086 adopt(parent, child);
2090 * Generic code for the isitmy*() functions. The rmask parameter is the
2091 * expected relationship of w1 to w2.
2094 _isitmyx(struct witness *w1, struct witness *w2, int rmask, const char *fname)
2096 unsigned char r1, r2;
2101 WITNESS_INDEX_ASSERT(i1);
2102 WITNESS_INDEX_ASSERT(i2);
2103 r1 = w_rmatrix[i1][i2] & WITNESS_RELATED_MASK;
2104 r2 = w_rmatrix[i2][i1] & WITNESS_RELATED_MASK;
2106 /* The flags on one better be the inverse of the flags on the other */
2107 if (!((WITNESS_ATOD(r1) == r2 && WITNESS_DTOA(r2) == r1) ||
2108 (WITNESS_DTOA(r1) == r2 && WITNESS_ATOD(r2) == r1))) {
2109 /* Don't squawk if we're potentially racing with an update. */
2110 if (!mtx_owned(&w_mtx))
2112 printf("%s: rmatrix mismatch between %s (index %d) and %s "
2113 "(index %d): w_rmatrix[%d][%d] == %hhx but "
2114 "w_rmatrix[%d][%d] == %hhx\n",
2115 fname, w1->w_name, i1, w2->w_name, i2, i1, i2, r1,
2118 printf("Witness disabled.\n");
2121 return (r1 & rmask);
2125 * Checks if @child is a direct child of @parent.
2128 isitmychild(struct witness *parent, struct witness *child)
2131 return (_isitmyx(parent, child, WITNESS_PARENT, __func__));
2135 * Checks if @descendant is a direct or inderect descendant of @ancestor.
2138 isitmydescendant(struct witness *ancestor, struct witness *descendant)
2141 return (_isitmyx(ancestor, descendant, WITNESS_ANCESTOR_MASK,
2146 blessed(struct witness *w1, struct witness *w2)
2149 struct witness_blessed *b;
2151 for (i = 0; i < nitems(blessed_list); i++) {
2152 b = &blessed_list[i];
2153 if (strcmp(w1->w_name, b->b_lock1) == 0) {
2154 if (strcmp(w2->w_name, b->b_lock2) == 0)
2158 if (strcmp(w1->w_name, b->b_lock2) == 0)
2159 if (strcmp(w2->w_name, b->b_lock1) == 0)
2165 static struct witness *
2171 if (witness_cold == 0)
2172 mtx_assert(&w_mtx, MA_OWNED);
2174 if (witness_watch == -1) {
2175 mtx_unlock_spin(&w_mtx);
2178 if (STAILQ_EMPTY(&w_free)) {
2180 mtx_unlock_spin(&w_mtx);
2181 printf("WITNESS: unable to allocate a new witness object\n");
2184 w = STAILQ_FIRST(&w_free);
2185 STAILQ_REMOVE_HEAD(&w_free, w_list);
2188 MPASS(index > 0 && index == w_max_used_index+1 &&
2189 index < witness_count);
2190 bzero(w, sizeof(*w));
2192 if (index > w_max_used_index)
2193 w_max_used_index = index;
2198 witness_free(struct witness *w)
2201 STAILQ_INSERT_HEAD(&w_free, w, w_list);
2205 static struct lock_list_entry *
2206 witness_lock_list_get(void)
2208 struct lock_list_entry *lle;
2210 if (witness_watch == -1)
2212 mtx_lock_spin(&w_mtx);
2213 lle = w_lock_list_free;
2216 mtx_unlock_spin(&w_mtx);
2217 printf("%s: witness exhausted\n", __func__);
2220 w_lock_list_free = lle->ll_next;
2221 mtx_unlock_spin(&w_mtx);
2222 bzero(lle, sizeof(*lle));
2227 witness_lock_list_free(struct lock_list_entry *lle)
2230 mtx_lock_spin(&w_mtx);
2231 lle->ll_next = w_lock_list_free;
2232 w_lock_list_free = lle;
2233 mtx_unlock_spin(&w_mtx);
2236 static struct lock_instance *
2237 find_instance(struct lock_list_entry *list, const struct lock_object *lock)
2239 struct lock_list_entry *lle;
2240 struct lock_instance *instance;
2243 for (lle = list; lle != NULL; lle = lle->ll_next)
2244 for (i = lle->ll_count - 1; i >= 0; i--) {
2245 instance = &lle->ll_children[i];
2246 if (instance->li_lock == lock)
2253 witness_list_lock(struct lock_instance *instance,
2254 int (*prnt)(const char *fmt, ...))
2256 struct lock_object *lock;
2258 lock = instance->li_lock;
2259 prnt("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ?
2260 "exclusive" : "shared", LOCK_CLASS(lock)->lc_name, lock->lo_name);
2261 if (lock->lo_witness->w_name != lock->lo_name)
2262 prnt(" (%s)", lock->lo_witness->w_name);
2263 prnt(" r = %d (%p) locked @ %s:%d\n",
2264 instance->li_flags & LI_RECURSEMASK, lock,
2265 fixup_filename(instance->li_file), instance->li_line);
2269 witness_output(const char *fmt, ...)
2275 ret = witness_voutput(fmt, ap);
2281 witness_voutput(const char *fmt, va_list ap)
2286 switch (witness_channel) {
2287 case WITNESS_CONSOLE:
2288 ret = vprintf(fmt, ap);
2291 vlog(LOG_NOTICE, fmt, ap);
2301 witness_thread_has_locks(struct thread *td)
2304 if (td->td_sleeplocks == NULL)
2306 return (td->td_sleeplocks->ll_count != 0);
2310 witness_proc_has_locks(struct proc *p)
2314 FOREACH_THREAD_IN_PROC(p, td) {
2315 if (witness_thread_has_locks(td))
2323 witness_list_locks(struct lock_list_entry **lock_list,
2324 int (*prnt)(const char *fmt, ...))
2326 struct lock_list_entry *lle;
2330 for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
2331 for (i = lle->ll_count - 1; i >= 0; i--) {
2332 witness_list_lock(&lle->ll_children[i], prnt);
2339 * This is a bit risky at best. We call this function when we have timed
2340 * out acquiring a spin lock, and we assume that the other CPU is stuck
2341 * with this lock held. So, we go groveling around in the other CPU's
2342 * per-cpu data to try to find the lock instance for this spin lock to
2343 * see when it was last acquired.
2346 witness_display_spinlock(struct lock_object *lock, struct thread *owner,
2347 int (*prnt)(const char *fmt, ...))
2349 struct lock_instance *instance;
2352 if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU)
2354 pc = pcpu_find(owner->td_oncpu);
2355 instance = find_instance(pc->pc_spinlocks, lock);
2356 if (instance != NULL)
2357 witness_list_lock(instance, prnt);
2361 witness_save(struct lock_object *lock, const char **filep, int *linep)
2363 struct lock_list_entry *lock_list;
2364 struct lock_instance *instance;
2365 struct lock_class *class;
2368 * This function is used independently in locking code to deal with
2369 * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2372 if (SCHEDULER_STOPPED())
2374 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2375 if (lock->lo_witness == NULL || witness_watch == -1 || KERNEL_PANICKED())
2377 class = LOCK_CLASS(lock);
2378 if (class->lc_flags & LC_SLEEPLOCK)
2379 lock_list = curthread->td_sleeplocks;
2381 if (witness_skipspin)
2383 lock_list = PCPU_GET(spinlocks);
2385 instance = find_instance(lock_list, lock);
2386 if (instance == NULL) {
2387 kassert_panic("%s: lock (%s) %s not locked", __func__,
2388 class->lc_name, lock->lo_name);
2391 *filep = instance->li_file;
2392 *linep = instance->li_line;
2396 witness_restore(struct lock_object *lock, const char *file, int line)
2398 struct lock_list_entry *lock_list;
2399 struct lock_instance *instance;
2400 struct lock_class *class;
2403 * This function is used independently in locking code to deal with
2404 * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2407 if (SCHEDULER_STOPPED())
2409 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2410 if (lock->lo_witness == NULL || witness_watch == -1 || KERNEL_PANICKED())
2412 class = LOCK_CLASS(lock);
2413 if (class->lc_flags & LC_SLEEPLOCK)
2414 lock_list = curthread->td_sleeplocks;
2416 if (witness_skipspin)
2418 lock_list = PCPU_GET(spinlocks);
2420 instance = find_instance(lock_list, lock);
2421 if (instance == NULL)
2422 kassert_panic("%s: lock (%s) %s not locked", __func__,
2423 class->lc_name, lock->lo_name);
2424 lock->lo_witness->w_file = file;
2425 lock->lo_witness->w_line = line;
2426 if (instance == NULL)
2428 instance->li_file = file;
2429 instance->li_line = line;
2433 witness_find_instance(const struct lock_object *lock,
2434 struct lock_instance **instance)
2436 #ifdef INVARIANT_SUPPORT
2437 struct lock_class *class;
2439 if (lock->lo_witness == NULL || witness_watch < 1 || KERNEL_PANICKED())
2441 class = LOCK_CLASS(lock);
2442 if ((class->lc_flags & LC_SLEEPLOCK) != 0) {
2443 *instance = find_instance(curthread->td_sleeplocks, lock);
2445 } else if ((class->lc_flags & LC_SPINLOCK) != 0) {
2446 *instance = find_instance(PCPU_GET(spinlocks), lock);
2449 kassert_panic("Lock (%s) %s is not sleep or spin!",
2450 class->lc_name, lock->lo_name);
2459 witness_assert(const struct lock_object *lock, int flags, const char *file,
2462 #ifdef INVARIANT_SUPPORT
2463 struct lock_instance *instance;
2464 struct lock_class *class;
2466 if (!witness_find_instance(lock, &instance))
2468 class = LOCK_CLASS(lock);
2471 if (instance != NULL)
2472 kassert_panic("Lock (%s) %s locked @ %s:%d.",
2473 class->lc_name, lock->lo_name,
2474 fixup_filename(file), line);
2477 case LA_LOCKED | LA_RECURSED:
2478 case LA_LOCKED | LA_NOTRECURSED:
2480 case LA_SLOCKED | LA_RECURSED:
2481 case LA_SLOCKED | LA_NOTRECURSED:
2483 case LA_XLOCKED | LA_RECURSED:
2484 case LA_XLOCKED | LA_NOTRECURSED:
2485 if (instance == NULL) {
2486 kassert_panic("Lock (%s) %s not locked @ %s:%d.",
2487 class->lc_name, lock->lo_name,
2488 fixup_filename(file), line);
2491 if ((flags & LA_XLOCKED) != 0 &&
2492 (instance->li_flags & LI_EXCLUSIVE) == 0)
2494 "Lock (%s) %s not exclusively locked @ %s:%d.",
2495 class->lc_name, lock->lo_name,
2496 fixup_filename(file), line);
2497 if ((flags & LA_SLOCKED) != 0 &&
2498 (instance->li_flags & LI_EXCLUSIVE) != 0)
2500 "Lock (%s) %s exclusively locked @ %s:%d.",
2501 class->lc_name, lock->lo_name,
2502 fixup_filename(file), line);
2503 if ((flags & LA_RECURSED) != 0 &&
2504 (instance->li_flags & LI_RECURSEMASK) == 0)
2505 kassert_panic("Lock (%s) %s not recursed @ %s:%d.",
2506 class->lc_name, lock->lo_name,
2507 fixup_filename(file), line);
2508 if ((flags & LA_NOTRECURSED) != 0 &&
2509 (instance->li_flags & LI_RECURSEMASK) != 0)
2510 kassert_panic("Lock (%s) %s recursed @ %s:%d.",
2511 class->lc_name, lock->lo_name,
2512 fixup_filename(file), line);
2515 kassert_panic("Invalid lock assertion at %s:%d.",
2516 fixup_filename(file), line);
2518 #endif /* INVARIANT_SUPPORT */
2522 * Checks the ownership of the lock by curthread, consulting the witness list.
2524 * 0 if witness is disabled or did not work
2529 witness_is_owned(const struct lock_object *lock)
2531 #ifdef INVARIANT_SUPPORT
2532 struct lock_instance *instance;
2534 if (!witness_find_instance(lock, &instance))
2536 return (instance == NULL ? -1 : 1);
2543 witness_setflag(struct lock_object *lock, int flag, int set)
2545 struct lock_list_entry *lock_list;
2546 struct lock_instance *instance;
2547 struct lock_class *class;
2549 if (lock->lo_witness == NULL || witness_watch == -1 || KERNEL_PANICKED())
2551 class = LOCK_CLASS(lock);
2552 if (class->lc_flags & LC_SLEEPLOCK)
2553 lock_list = curthread->td_sleeplocks;
2555 if (witness_skipspin)
2557 lock_list = PCPU_GET(spinlocks);
2559 instance = find_instance(lock_list, lock);
2560 if (instance == NULL) {
2561 kassert_panic("%s: lock (%s) %s not locked", __func__,
2562 class->lc_name, lock->lo_name);
2567 instance->li_flags |= flag;
2569 instance->li_flags &= ~flag;
2573 witness_norelease(struct lock_object *lock)
2576 witness_setflag(lock, LI_NORELEASE, 1);
2580 witness_releaseok(struct lock_object *lock)
2583 witness_setflag(lock, LI_NORELEASE, 0);
2588 witness_ddb_list(struct thread *td)
2591 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2592 KASSERT(kdb_active, ("%s: not in the debugger", __func__));
2594 if (witness_watch < 1)
2597 witness_list_locks(&td->td_sleeplocks, db_printf);
2600 * We only handle spinlocks if td == curthread. This is somewhat broken
2601 * if td is currently executing on some other CPU and holds spin locks
2602 * as we won't display those locks. If we had a MI way of getting
2603 * the per-cpu data for a given cpu then we could use
2604 * td->td_oncpu to get the list of spinlocks for this thread
2607 * That still wouldn't really fix this unless we locked the scheduler
2608 * lock or stopped the other CPU to make sure it wasn't changing the
2609 * list out from under us. It is probably best to just not try to
2610 * handle threads on other CPU's for now.
2612 if (td == curthread && PCPU_GET(spinlocks) != NULL)
2613 witness_list_locks(PCPU_PTR(spinlocks), db_printf);
2616 DB_SHOW_COMMAND(locks, db_witness_list)
2621 td = db_lookup_thread(addr, true);
2624 witness_ddb_list(td);
2627 DB_SHOW_ALL_COMMAND(locks, db_witness_list_all)
2633 * It would be nice to list only threads and processes that actually
2634 * held sleep locks, but that information is currently not exported
2637 FOREACH_PROC_IN_SYSTEM(p) {
2638 if (!witness_proc_has_locks(p))
2640 FOREACH_THREAD_IN_PROC(p, td) {
2641 if (!witness_thread_has_locks(td))
2643 db_printf("Process %d (%s) thread %p (%d)\n", p->p_pid,
2644 p->p_comm, td, td->td_tid);
2645 witness_ddb_list(td);
2651 DB_SHOW_ALIAS_FLAGS(alllocks, db_witness_list_all, DB_CMD_MEMSAFE)
2653 DB_SHOW_COMMAND_FLAGS(witness, db_witness_display, DB_CMD_MEMSAFE)
2656 witness_ddb_display(db_printf);
2661 sbuf_print_witness_badstacks(struct sbuf *sb, size_t *oldidx)
2663 struct witness_lock_order_data *data1, *data2, *tmp_data1, *tmp_data2;
2664 struct witness *tmp_w1, *tmp_w2, *w1, *w2;
2665 int generation, i, j;
2672 /* Allocate and init temporary storage space. */
2673 tmp_w1 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2674 tmp_w2 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2675 tmp_data1 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2677 tmp_data2 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2679 stack_zero(&tmp_data1->wlod_stack);
2680 stack_zero(&tmp_data2->wlod_stack);
2683 mtx_lock_spin(&w_mtx);
2684 generation = w_generation;
2685 mtx_unlock_spin(&w_mtx);
2686 sbuf_printf(sb, "Number of known direct relationships is %d\n",
2687 w_lohash.wloh_count);
2688 for (i = 1; i < w_max_used_index; i++) {
2689 mtx_lock_spin(&w_mtx);
2690 if (generation != w_generation) {
2691 mtx_unlock_spin(&w_mtx);
2693 /* The graph has changed, try again. */
2700 if (w1->w_reversed == 0) {
2701 mtx_unlock_spin(&w_mtx);
2705 /* Copy w1 locally so we can release the spin lock. */
2707 mtx_unlock_spin(&w_mtx);
2709 if (tmp_w1->w_reversed == 0)
2711 for (j = 1; j < w_max_used_index; j++) {
2712 if ((w_rmatrix[i][j] & WITNESS_REVERSAL) == 0 || i > j)
2715 mtx_lock_spin(&w_mtx);
2716 if (generation != w_generation) {
2717 mtx_unlock_spin(&w_mtx);
2719 /* The graph has changed, try again. */
2726 data1 = witness_lock_order_get(w1, w2);
2727 data2 = witness_lock_order_get(w2, w1);
2730 * Copy information locally so we can release the
2736 stack_zero(&tmp_data1->wlod_stack);
2737 stack_copy(&data1->wlod_stack,
2738 &tmp_data1->wlod_stack);
2740 if (data2 && data2 != data1) {
2741 stack_zero(&tmp_data2->wlod_stack);
2742 stack_copy(&data2->wlod_stack,
2743 &tmp_data2->wlod_stack);
2745 mtx_unlock_spin(&w_mtx);
2747 if (blessed(tmp_w1, tmp_w2))
2751 "\nLock order reversal between \"%s\"(%s) and \"%s\"(%s)!\n",
2752 tmp_w1->w_name, tmp_w1->w_class->lc_name,
2753 tmp_w2->w_name, tmp_w2->w_class->lc_name);
2756 "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2757 tmp_w1->w_name, tmp_w1->w_class->lc_name,
2758 tmp_w2->w_name, tmp_w2->w_class->lc_name);
2759 stack_sbuf_print(sb, &tmp_data1->wlod_stack);
2760 sbuf_printf(sb, "\n");
2762 if (data2 && data2 != data1) {
2764 "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2765 tmp_w2->w_name, tmp_w2->w_class->lc_name,
2766 tmp_w1->w_name, tmp_w1->w_class->lc_name);
2767 stack_sbuf_print(sb, &tmp_data2->wlod_stack);
2768 sbuf_printf(sb, "\n");
2772 mtx_lock_spin(&w_mtx);
2773 if (generation != w_generation) {
2774 mtx_unlock_spin(&w_mtx);
2777 * The graph changed while we were printing stack data,
2784 mtx_unlock_spin(&w_mtx);
2786 /* Free temporary storage space. */
2787 free(tmp_data1, M_TEMP);
2788 free(tmp_data2, M_TEMP);
2789 free(tmp_w1, M_TEMP);
2790 free(tmp_w2, M_TEMP);
2794 sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS)
2799 if (witness_watch < 1) {
2800 error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2804 error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2808 sb = sbuf_new(NULL, NULL, badstack_sbuf_size, SBUF_AUTOEXTEND);
2812 sbuf_print_witness_badstacks(sb, &req->oldidx);
2815 error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1);
2823 sbuf_db_printf_drain(void *arg __unused, const char *data, int len)
2826 return (db_printf("%.*s", len, data));
2829 DB_SHOW_COMMAND_FLAGS(badstacks, db_witness_badstacks, DB_CMD_MEMSAFE)
2835 sbuf_new(&sb, buffer, sizeof(buffer), SBUF_FIXEDLEN);
2836 sbuf_set_drain(&sb, sbuf_db_printf_drain, NULL);
2837 sbuf_print_witness_badstacks(&sb, &dummy);
2843 sysctl_debug_witness_channel(SYSCTL_HANDLER_ARGS)
2845 static const struct {
2846 enum witness_channel channel;
2849 { WITNESS_CONSOLE, "console" },
2850 { WITNESS_LOG, "log" },
2851 { WITNESS_NONE, "none" },
2858 for (i = 0; i < nitems(channels); i++)
2859 if (witness_channel == channels[i].channel) {
2860 snprintf(buf, sizeof(buf), "%s", channels[i].name);
2864 error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
2865 if (error != 0 || req->newptr == NULL)
2869 for (i = 0; i < nitems(channels); i++)
2870 if (strcmp(channels[i].name, buf) == 0) {
2871 witness_channel = channels[i].channel;
2879 sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS)
2886 error = SYSCTL_OUT(req, w_notallowed, sizeof(w_notallowed));
2890 if (witness_watch < 1) {
2891 error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2895 error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2900 error = sysctl_wire_old_buffer(req, 0);
2903 sb = sbuf_new_for_sysctl(NULL, NULL, FULLGRAPH_SBUF_SIZE, req);
2906 sbuf_printf(sb, "\n");
2908 mtx_lock_spin(&w_mtx);
2909 STAILQ_FOREACH(w, &w_all, w_list)
2911 STAILQ_FOREACH(w, &w_all, w_list)
2912 witness_add_fullgraph(sb, w);
2913 mtx_unlock_spin(&w_mtx);
2916 * Close the sbuf and return to userland.
2918 error = sbuf_finish(sb);
2925 sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS)
2929 value = witness_watch;
2930 error = sysctl_handle_int(oidp, &value, 0, req);
2931 if (error != 0 || req->newptr == NULL)
2933 if (value > 1 || value < -1 ||
2934 (witness_watch == -1 && value != witness_watch))
2936 witness_watch = value;
2941 witness_add_fullgraph(struct sbuf *sb, struct witness *w)
2945 if (w->w_displayed != 0 || (w->w_file == NULL && w->w_line == 0))
2949 WITNESS_INDEX_ASSERT(w->w_index);
2950 for (i = 1; i <= w_max_used_index; i++) {
2951 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT) {
2952 sbuf_printf(sb, "\"%s\",\"%s\"\n", w->w_name,
2954 witness_add_fullgraph(sb, &w_data[i]);
2960 * A simple hash function. Takes a key pointer and a key size. If size == 0,
2961 * interprets the key as a string and reads until the null
2962 * terminator. Otherwise, reads the first size bytes. Returns an unsigned 32-bit
2963 * hash value computed from the key.
2966 witness_hash_djb2(const uint8_t *key, uint32_t size)
2968 unsigned int hash = 5381;
2971 /* hash = hash * 33 + key[i] */
2973 for (i = 0; i < size; i++)
2974 hash = ((hash << 5) + hash) + (unsigned int)key[i];
2976 for (i = 0; key[i] != 0; i++)
2977 hash = ((hash << 5) + hash) + (unsigned int)key[i];
2983 * Initializes the two witness hash tables. Called exactly once from
2984 * witness_initialize().
2987 witness_init_hash_tables(void)
2991 MPASS(witness_cold);
2993 /* Initialize the hash tables. */
2994 for (i = 0; i < WITNESS_HASH_SIZE; i++)
2995 w_hash.wh_array[i] = NULL;
2997 w_hash.wh_size = WITNESS_HASH_SIZE;
2998 w_hash.wh_count = 0;
3000 /* Initialize the lock order data hash. */
3002 for (i = 0; i < WITNESS_LO_DATA_COUNT; i++) {
3003 memset(&w_lodata[i], 0, sizeof(w_lodata[i]));
3004 w_lodata[i].wlod_next = w_lofree;
3005 w_lofree = &w_lodata[i];
3007 w_lohash.wloh_size = WITNESS_LO_HASH_SIZE;
3008 w_lohash.wloh_count = 0;
3009 for (i = 0; i < WITNESS_LO_HASH_SIZE; i++)
3010 w_lohash.wloh_array[i] = NULL;
3013 static struct witness *
3014 witness_hash_get(const char *key)
3020 if (witness_cold == 0)
3021 mtx_assert(&w_mtx, MA_OWNED);
3022 hash = witness_hash_djb2(key, 0) % w_hash.wh_size;
3023 w = w_hash.wh_array[hash];
3025 if (strcmp(w->w_name, key) == 0)
3035 witness_hash_put(struct witness *w)
3040 MPASS(w->w_name != NULL);
3041 if (witness_cold == 0)
3042 mtx_assert(&w_mtx, MA_OWNED);
3043 KASSERT(witness_hash_get(w->w_name) == NULL,
3044 ("%s: trying to add a hash entry that already exists!", __func__));
3045 KASSERT(w->w_hash_next == NULL,
3046 ("%s: w->w_hash_next != NULL", __func__));
3048 hash = witness_hash_djb2(w->w_name, 0) % w_hash.wh_size;
3049 w->w_hash_next = w_hash.wh_array[hash];
3050 w_hash.wh_array[hash] = w;
3054 static struct witness_lock_order_data *
3055 witness_lock_order_get(struct witness *parent, struct witness *child)
3057 struct witness_lock_order_data *data = NULL;
3058 struct witness_lock_order_key key;
3061 MPASS(parent != NULL && child != NULL);
3062 key.from = parent->w_index;
3063 key.to = child->w_index;
3064 WITNESS_INDEX_ASSERT(key.from);
3065 WITNESS_INDEX_ASSERT(key.to);
3066 if ((w_rmatrix[parent->w_index][child->w_index]
3067 & WITNESS_LOCK_ORDER_KNOWN) == 0)
3070 hash = witness_hash_djb2((const char*)&key,
3071 sizeof(key)) % w_lohash.wloh_size;
3072 data = w_lohash.wloh_array[hash];
3073 while (data != NULL) {
3074 if (witness_lock_order_key_equal(&data->wlod_key, &key))
3076 data = data->wlod_next;
3084 * Verify that parent and child have a known relationship, are not the same,
3085 * and child is actually a child of parent. This is done without w_mtx
3086 * to avoid contention in the common case.
3089 witness_lock_order_check(struct witness *parent, struct witness *child)
3092 if (parent != child &&
3093 w_rmatrix[parent->w_index][child->w_index]
3094 & WITNESS_LOCK_ORDER_KNOWN &&
3095 isitmychild(parent, child))
3102 witness_lock_order_add(struct witness *parent, struct witness *child)
3104 struct witness_lock_order_data *data = NULL;
3105 struct witness_lock_order_key key;
3108 MPASS(parent != NULL && child != NULL);
3109 key.from = parent->w_index;
3110 key.to = child->w_index;
3111 WITNESS_INDEX_ASSERT(key.from);
3112 WITNESS_INDEX_ASSERT(key.to);
3113 if (w_rmatrix[parent->w_index][child->w_index]
3114 & WITNESS_LOCK_ORDER_KNOWN)
3117 hash = witness_hash_djb2((const char*)&key,
3118 sizeof(key)) % w_lohash.wloh_size;
3119 w_rmatrix[parent->w_index][child->w_index] |= WITNESS_LOCK_ORDER_KNOWN;
3123 w_lofree = data->wlod_next;
3124 data->wlod_next = w_lohash.wloh_array[hash];
3125 data->wlod_key = key;
3126 w_lohash.wloh_array[hash] = data;
3127 w_lohash.wloh_count++;
3128 stack_save(&data->wlod_stack);
3132 /* Call this whenever the structure of the witness graph changes. */
3134 witness_increment_graph_generation(void)
3137 if (witness_cold == 0)
3138 mtx_assert(&w_mtx, MA_OWNED);
3143 witness_output_drain(void *arg __unused, const char *data, int len)
3146 witness_output("%.*s", len, data);
3151 witness_debugger(int cond, const char *msg)
3160 if (witness_trace) {
3161 sbuf_new(&sb, buf, sizeof(buf), SBUF_FIXEDLEN);
3162 sbuf_set_drain(&sb, witness_output_drain, NULL);
3165 witness_output("stack backtrace:\n");
3166 stack_sbuf_print_ddb(&sb, &st);
3171 witness_enter_debugger(msg);
3175 witness_enter_debugger(const char *msg)
3179 kdb_enter(KDB_WHY_WITNESS, msg);