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 { "ifnet_rw", &lock_class_rw },
545 { "if_addr_lock", &lock_class_mtx_sleep },
549 * protocol locks before interface locks, after UDP locks.
551 { "in6_multi_sx", &lock_class_sx },
552 { "udpinp", &lock_class_rw },
553 { "in6_multi_list_mtx", &lock_class_mtx_sleep },
554 { "mld_mtx", &lock_class_mtx_sleep },
555 { "ifnet_rw", &lock_class_rw },
556 { "if_addr_lock", &lock_class_mtx_sleep },
559 * UNIX Domain Sockets
561 { "unp_link_rwlock", &lock_class_rw },
562 { "unp_list_lock", &lock_class_mtx_sleep },
563 { "unp", &lock_class_mtx_sleep },
564 { "so_snd", &lock_class_mtx_sleep },
569 { "udp", &lock_class_mtx_sleep },
570 { "udpinp", &lock_class_rw },
571 { "so_snd", &lock_class_mtx_sleep },
576 { "tcp", &lock_class_mtx_sleep },
577 { "tcpinp", &lock_class_rw },
578 { "so_snd", &lock_class_mtx_sleep },
583 { "bpf global lock", &lock_class_sx },
584 { "bpf cdev lock", &lock_class_mtx_sleep },
589 { "nfsd_mtx", &lock_class_mtx_sleep },
590 { "so_snd", &lock_class_mtx_sleep },
596 { "802.11 com lock", &lock_class_mtx_sleep},
601 { "network driver", &lock_class_mtx_sleep},
607 { "ng_node", &lock_class_mtx_sleep },
608 { "ng_worklist", &lock_class_mtx_sleep },
613 { "vm map (system)", &lock_class_mtx_sleep },
614 { "vnode interlock", &lock_class_mtx_sleep },
615 { "cdev", &lock_class_mtx_sleep },
616 { "devthrd", &lock_class_mtx_sleep },
621 { "vm map (user)", &lock_class_sx },
622 { "vm object", &lock_class_rw },
623 { "vm page", &lock_class_mtx_sleep },
624 { "pmap pv global", &lock_class_rw },
625 { "pmap", &lock_class_mtx_sleep },
626 { "pmap pv list", &lock_class_rw },
627 { "vm page free queue", &lock_class_mtx_sleep },
628 { "vm pagequeue", &lock_class_mtx_sleep },
631 * kqueue/VFS interaction
633 { "kqueue", &lock_class_mtx_sleep },
634 { "struct mount mtx", &lock_class_mtx_sleep },
635 { "vnode interlock", &lock_class_mtx_sleep },
640 { "ncvn", &lock_class_mtx_sleep },
641 { "ncbuc", &lock_class_mtx_sleep },
642 { "vnode interlock", &lock_class_mtx_sleep },
643 { "ncneg", &lock_class_mtx_sleep },
648 { "dn->dn_mtx", &lock_class_sx },
649 { "dr->dt.di.dr_mtx", &lock_class_sx },
650 { "db->db_mtx", &lock_class_sx },
655 { "TCP ID tree", &lock_class_rw },
656 { "tcp log id bucket", &lock_class_mtx_sleep },
657 { "tcpinp", &lock_class_rw },
658 { "TCP log expireq", &lock_class_mtx_sleep },
664 { "ap boot", &lock_class_mtx_spin },
666 { "rm.mutex_mtx", &lock_class_mtx_spin },
668 { "cy", &lock_class_mtx_spin },
670 { "scc_hwmtx", &lock_class_mtx_spin },
671 { "uart_hwmtx", &lock_class_mtx_spin },
672 { "fast_taskqueue", &lock_class_mtx_spin },
673 { "intr table", &lock_class_mtx_spin },
674 { "process slock", &lock_class_mtx_spin },
675 { "syscons video lock", &lock_class_mtx_spin },
676 { "sleepq chain", &lock_class_mtx_spin },
677 { "rm_spinlock", &lock_class_mtx_spin },
678 { "turnstile chain", &lock_class_mtx_spin },
679 { "turnstile lock", &lock_class_mtx_spin },
680 { "sched lock", &lock_class_mtx_spin },
681 { "td_contested", &lock_class_mtx_spin },
682 { "callout", &lock_class_mtx_spin },
683 { "entropy harvest mutex", &lock_class_mtx_spin },
685 { "smp rendezvous", &lock_class_mtx_spin },
688 { "tlb0", &lock_class_mtx_spin },
691 { "sched lock", &lock_class_mtx_spin },
693 { "pmc-per-proc", &lock_class_mtx_spin },
699 { "intrcnt", &lock_class_mtx_spin },
700 { "icu", &lock_class_mtx_spin },
702 { "allpmaps", &lock_class_mtx_spin },
703 { "descriptor tables", &lock_class_mtx_spin },
705 { "clk", &lock_class_mtx_spin },
706 { "cpuset", &lock_class_mtx_spin },
707 { "mprof lock", &lock_class_mtx_spin },
708 { "zombie lock", &lock_class_mtx_spin },
709 { "ALD Queue", &lock_class_mtx_spin },
710 #if defined(__i386__) || defined(__amd64__)
711 { "pcicfg", &lock_class_mtx_spin },
712 { "NDIS thread lock", &lock_class_mtx_spin },
714 { "tw_osl_io_lock", &lock_class_mtx_spin },
715 { "tw_osl_q_lock", &lock_class_mtx_spin },
716 { "tw_cl_io_lock", &lock_class_mtx_spin },
717 { "tw_cl_intr_lock", &lock_class_mtx_spin },
718 { "tw_cl_gen_lock", &lock_class_mtx_spin },
720 { "pmc-leaf", &lock_class_mtx_spin },
722 { "blocked lock", &lock_class_mtx_spin },
728 * Pairs of locks which have been blessed. Witness does not complain about
729 * order problems with blessed lock pairs. Please do not add an entry to the
730 * table without an explanatory comment.
732 static struct witness_blessed blessed_list[] = {
734 * See the comment in ufs_dirhash.c. Basically, a vnode lock serializes
735 * both lock orders, so a deadlock cannot happen as a result of this
738 { "dirhash", "bufwait" },
741 * A UFS vnode may be locked in vget() while a buffer belonging to the
742 * parent directory vnode is locked.
744 { "ufs", "bufwait" },
748 * This global is set to 0 once it becomes safe to use the witness code.
750 static int witness_cold = 1;
753 * This global is set to 1 once the static lock orders have been enrolled
754 * so that a warning can be issued for any spin locks enrolled later.
756 static int witness_spin_warn = 0;
758 /* Trim useless garbage from filenames. */
760 fixup_filename(const char *file)
765 while (strncmp(file, "../", 3) == 0)
771 * Calculate the size of early witness structures.
774 witness_startup_count(void)
778 sz = sizeof(struct witness) * witness_count;
779 sz += sizeof(*w_rmatrix) * (witness_count + 1);
780 sz += sizeof(*w_rmatrix[0]) * (witness_count + 1) *
787 * The WITNESS-enabled diagnostic code. Note that the witness code does
788 * assume that the early boot is single-threaded at least until after this
789 * routine is completed.
792 witness_startup(void *mem)
794 struct lock_object *lock;
795 struct witness_order_list_entry *order;
796 struct witness *w, *w1;
802 p += sizeof(struct witness) * witness_count;
804 w_rmatrix = (void *)p;
805 p += sizeof(*w_rmatrix) * (witness_count + 1);
807 for (i = 0; i < witness_count + 1; i++) {
808 w_rmatrix[i] = (void *)p;
809 p += sizeof(*w_rmatrix[i]) * (witness_count + 1);
811 badstack_sbuf_size = witness_count * 256;
814 * We have to release Giant before initializing its witness
815 * structure so that WITNESS doesn't get confused.
818 mtx_assert(&Giant, MA_NOTOWNED);
820 CTR1(KTR_WITNESS, "%s: initializing witness", __func__);
821 mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET |
822 MTX_NOWITNESS | MTX_NOPROFILE);
823 for (i = witness_count - 1; i >= 0; i--) {
825 memset(w, 0, sizeof(*w));
826 w_data[i].w_index = i; /* Witness index never changes. */
829 KASSERT(STAILQ_FIRST(&w_free)->w_index == 0,
830 ("%s: Invalid list of free witness objects", __func__));
832 /* Witness with index 0 is not used to aid in debugging. */
833 STAILQ_REMOVE_HEAD(&w_free, w_list);
836 for (i = 0; i < witness_count; i++) {
837 memset(w_rmatrix[i], 0, sizeof(*w_rmatrix[i]) *
838 (witness_count + 1));
841 for (i = 0; i < LOCK_CHILDCOUNT; i++)
842 witness_lock_list_free(&w_locklistdata[i]);
843 witness_init_hash_tables();
845 /* First add in all the specified order lists. */
846 for (order = order_lists; order->w_name != NULL; order++) {
847 w = enroll(order->w_name, order->w_class);
850 w->w_file = "order list";
851 for (order++; order->w_name != NULL; order++) {
852 w1 = enroll(order->w_name, order->w_class);
855 w1->w_file = "order list";
860 witness_spin_warn = 1;
862 /* Iterate through all locks and add them to witness. */
863 for (i = 0; pending_locks[i].wh_lock != NULL; i++) {
864 lock = pending_locks[i].wh_lock;
865 KASSERT(lock->lo_flags & LO_WITNESS,
866 ("%s: lock %s is on pending list but not LO_WITNESS",
867 __func__, lock->lo_name));
868 lock->lo_witness = enroll(pending_locks[i].wh_type,
872 /* Mark the witness code as being ready for use. */
879 witness_init(struct lock_object *lock, const char *type)
881 struct lock_class *class;
883 /* Various sanity checks. */
884 class = LOCK_CLASS(lock);
885 if ((lock->lo_flags & LO_RECURSABLE) != 0 &&
886 (class->lc_flags & LC_RECURSABLE) == 0)
887 kassert_panic("%s: lock (%s) %s can not be recursable",
888 __func__, class->lc_name, lock->lo_name);
889 if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
890 (class->lc_flags & LC_SLEEPABLE) == 0)
891 kassert_panic("%s: lock (%s) %s can not be sleepable",
892 __func__, class->lc_name, lock->lo_name);
893 if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
894 (class->lc_flags & LC_UPGRADABLE) == 0)
895 kassert_panic("%s: lock (%s) %s can not be upgradable",
896 __func__, class->lc_name, lock->lo_name);
899 * If we shouldn't watch this lock, then just clear lo_witness.
900 * Otherwise, if witness_cold is set, then it is too early to
901 * enroll this lock, so defer it to witness_initialize() by adding
902 * it to the pending_locks list. If it is not too early, then enroll
905 if (witness_watch < 1 || KERNEL_PANICKED() ||
906 (lock->lo_flags & LO_WITNESS) == 0)
907 lock->lo_witness = NULL;
908 else if (witness_cold) {
909 pending_locks[pending_cnt].wh_lock = lock;
910 pending_locks[pending_cnt++].wh_type = type;
911 if (pending_cnt > WITNESS_PENDLIST)
912 panic("%s: pending locks list is too small, "
913 "increase WITNESS_PENDLIST\n",
916 lock->lo_witness = enroll(type, class);
920 witness_destroy(struct lock_object *lock)
922 struct lock_class *class;
925 class = LOCK_CLASS(lock);
928 panic("lock (%s) %s destroyed while witness_cold",
929 class->lc_name, lock->lo_name);
931 /* XXX: need to verify that no one holds the lock */
932 if ((lock->lo_flags & LO_WITNESS) == 0 || lock->lo_witness == NULL)
934 w = lock->lo_witness;
936 mtx_lock_spin(&w_mtx);
937 MPASS(w->w_refcount > 0);
940 if (w->w_refcount == 0)
942 mtx_unlock_spin(&w_mtx);
947 witness_ddb_compute_levels(void)
952 * First clear all levels.
954 STAILQ_FOREACH(w, &w_all, w_list)
958 * Look for locks with no parents and level all their descendants.
960 STAILQ_FOREACH(w, &w_all, w_list) {
961 /* If the witness has ancestors (is not a root), skip it. */
962 if (w->w_num_ancestors > 0)
964 witness_ddb_level_descendants(w, 0);
969 witness_ddb_level_descendants(struct witness *w, int l)
973 if (w->w_ddb_level >= l)
979 for (i = 1; i <= w_max_used_index; i++) {
980 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
981 witness_ddb_level_descendants(&w_data[i], l);
986 witness_ddb_display_descendants(int(*prnt)(const char *fmt, ...),
987 struct witness *w, int indent)
991 for (i = 0; i < indent; i++)
993 prnt("%s (type: %s, depth: %d, active refs: %d)",
994 w->w_name, w->w_class->lc_name,
995 w->w_ddb_level, w->w_refcount);
996 if (w->w_displayed) {
997 prnt(" -- (already displayed)\n");
1001 if (w->w_file != NULL && w->w_line != 0)
1002 prnt(" -- last acquired @ %s:%d\n", fixup_filename(w->w_file),
1005 prnt(" -- never acquired\n");
1007 WITNESS_INDEX_ASSERT(w->w_index);
1008 for (i = 1; i <= w_max_used_index; i++) {
1011 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
1012 witness_ddb_display_descendants(prnt, &w_data[i],
1018 witness_ddb_display_list(int(*prnt)(const char *fmt, ...),
1019 struct witness_list *list)
1023 STAILQ_FOREACH(w, list, w_typelist) {
1024 if (w->w_file == NULL || w->w_ddb_level > 0)
1027 /* This lock has no anscestors - display its descendants. */
1028 witness_ddb_display_descendants(prnt, w, 0);
1035 witness_ddb_display(int(*prnt)(const char *fmt, ...))
1039 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1040 witness_ddb_compute_levels();
1042 /* Clear all the displayed flags. */
1043 STAILQ_FOREACH(w, &w_all, w_list)
1047 * First, handle sleep locks which have been acquired at least
1050 prnt("Sleep locks:\n");
1051 witness_ddb_display_list(prnt, &w_sleep);
1056 * Now do spin locks which have been acquired at least once.
1058 prnt("\nSpin locks:\n");
1059 witness_ddb_display_list(prnt, &w_spin);
1064 * Finally, any locks which have not been acquired yet.
1066 prnt("\nLocks which were never acquired:\n");
1067 STAILQ_FOREACH(w, &w_all, w_list) {
1068 if (w->w_file != NULL || w->w_refcount == 0)
1070 prnt("%s (type: %s, depth: %d)\n", w->w_name,
1071 w->w_class->lc_name, w->w_ddb_level);
1079 witness_defineorder(struct lock_object *lock1, struct lock_object *lock2)
1082 if (witness_watch == -1 || KERNEL_PANICKED())
1085 /* Require locks that witness knows about. */
1086 if (lock1 == NULL || lock1->lo_witness == NULL || lock2 == NULL ||
1087 lock2->lo_witness == NULL)
1090 mtx_assert(&w_mtx, MA_NOTOWNED);
1091 mtx_lock_spin(&w_mtx);
1094 * If we already have either an explicit or implied lock order that
1095 * is the other way around, then return an error.
1097 if (witness_watch &&
1098 isitmydescendant(lock2->lo_witness, lock1->lo_witness)) {
1099 mtx_unlock_spin(&w_mtx);
1103 /* Try to add the new order. */
1104 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1105 lock2->lo_witness->w_name, lock1->lo_witness->w_name);
1106 itismychild(lock1->lo_witness, lock2->lo_witness);
1107 mtx_unlock_spin(&w_mtx);
1112 witness_checkorder(struct lock_object *lock, int flags, const char *file,
1113 int line, struct lock_object *interlock)
1115 struct lock_list_entry *lock_list, *lle;
1116 struct lock_instance *lock1, *lock2, *plock;
1117 struct lock_class *class, *iclass;
1118 struct witness *w, *w1;
1122 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL ||
1126 w = lock->lo_witness;
1127 class = LOCK_CLASS(lock);
1130 if (class->lc_flags & LC_SLEEPLOCK) {
1132 * Since spin locks include a critical section, this check
1133 * implicitly enforces a lock order of all sleep locks before
1136 if (td->td_critnest != 0 && !kdb_active)
1137 kassert_panic("acquiring blockable sleep lock with "
1138 "spinlock or critical section held (%s) %s @ %s:%d",
1139 class->lc_name, lock->lo_name,
1140 fixup_filename(file), line);
1143 * If this is the first lock acquired then just return as
1144 * no order checking is needed.
1146 lock_list = td->td_sleeplocks;
1147 if (lock_list == NULL || lock_list->ll_count == 0)
1151 * If this is the first lock, just return as no order
1152 * checking is needed. Avoid problems with thread
1153 * migration pinning the thread while checking if
1154 * spinlocks are held. If at least one spinlock is held
1155 * the thread is in a safe path and it is allowed to
1159 lock_list = PCPU_GET(spinlocks);
1160 if (lock_list == NULL || lock_list->ll_count == 0) {
1168 * Check to see if we are recursing on a lock we already own. If
1169 * so, make sure that we don't mismatch exclusive and shared lock
1172 lock1 = find_instance(lock_list, lock);
1173 if (lock1 != NULL) {
1174 if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
1175 (flags & LOP_EXCLUSIVE) == 0) {
1176 witness_output("shared lock of (%s) %s @ %s:%d\n",
1177 class->lc_name, lock->lo_name,
1178 fixup_filename(file), line);
1179 witness_output("while exclusively locked from %s:%d\n",
1180 fixup_filename(lock1->li_file), lock1->li_line);
1181 kassert_panic("excl->share");
1183 if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
1184 (flags & LOP_EXCLUSIVE) != 0) {
1185 witness_output("exclusive lock of (%s) %s @ %s:%d\n",
1186 class->lc_name, lock->lo_name,
1187 fixup_filename(file), line);
1188 witness_output("while share locked from %s:%d\n",
1189 fixup_filename(lock1->li_file), lock1->li_line);
1190 kassert_panic("share->excl");
1195 /* Warn if the interlock is not locked exactly once. */
1196 if (interlock != NULL) {
1197 iclass = LOCK_CLASS(interlock);
1198 lock1 = find_instance(lock_list, interlock);
1200 kassert_panic("interlock (%s) %s not locked @ %s:%d",
1201 iclass->lc_name, interlock->lo_name,
1202 fixup_filename(file), line);
1203 else if ((lock1->li_flags & LI_RECURSEMASK) != 0)
1204 kassert_panic("interlock (%s) %s recursed @ %s:%d",
1205 iclass->lc_name, interlock->lo_name,
1206 fixup_filename(file), line);
1210 * Find the previously acquired lock, but ignore interlocks.
1212 plock = &lock_list->ll_children[lock_list->ll_count - 1];
1213 if (interlock != NULL && plock->li_lock == interlock) {
1214 if (lock_list->ll_count > 1)
1216 &lock_list->ll_children[lock_list->ll_count - 2];
1218 lle = lock_list->ll_next;
1221 * The interlock is the only lock we hold, so
1226 plock = &lle->ll_children[lle->ll_count - 1];
1231 * Try to perform most checks without a lock. If this succeeds we
1232 * can skip acquiring the lock and return success. Otherwise we redo
1233 * the check with the lock held to handle races with concurrent updates.
1235 w1 = plock->li_lock->lo_witness;
1236 if (witness_lock_order_check(w1, w))
1239 mtx_lock_spin(&w_mtx);
1240 if (witness_lock_order_check(w1, w)) {
1241 mtx_unlock_spin(&w_mtx);
1244 witness_lock_order_add(w1, w);
1247 * Check for duplicate locks of the same type. Note that we only
1248 * have to check for this on the last lock we just acquired. Any
1249 * other cases will be caught as lock order violations.
1253 if (!(lock->lo_flags & LO_DUPOK) && !(flags & LOP_DUPOK) &&
1254 !(w_rmatrix[i][i] & WITNESS_REVERSAL)) {
1255 w_rmatrix[i][i] |= WITNESS_REVERSAL;
1257 mtx_unlock_spin(&w_mtx);
1259 "acquiring duplicate lock of same type: \"%s\"\n",
1261 witness_output(" 1st %s @ %s:%d\n", plock->li_lock->lo_name,
1262 fixup_filename(plock->li_file), plock->li_line);
1263 witness_output(" 2nd %s @ %s:%d\n", lock->lo_name,
1264 fixup_filename(file), line);
1265 witness_debugger(1, __func__);
1267 mtx_unlock_spin(&w_mtx);
1270 mtx_assert(&w_mtx, MA_OWNED);
1273 * If we know that the lock we are acquiring comes after
1274 * the lock we most recently acquired in the lock order tree,
1275 * then there is no need for any further checks.
1277 if (isitmychild(w1, w))
1280 for (j = 0, lle = lock_list; lle != NULL; lle = lle->ll_next) {
1281 for (i = lle->ll_count - 1; i >= 0; i--, j++) {
1282 struct stack pstack;
1283 bool pstackv, trace;
1285 MPASS(j < LOCK_CHILDCOUNT * LOCK_NCHILDREN);
1286 lock1 = &lle->ll_children[i];
1289 * Ignore the interlock.
1291 if (interlock == lock1->li_lock)
1295 * If this lock doesn't undergo witness checking,
1298 w1 = lock1->li_lock->lo_witness;
1300 KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
1301 ("lock missing witness structure"));
1306 * If we are locking Giant and this is a sleepable
1307 * lock, then skip it.
1309 if ((lock1->li_flags & LI_SLEEPABLE) != 0 &&
1310 lock == &Giant.lock_object)
1314 * If we are locking a sleepable lock and this lock
1315 * is Giant, then skip it.
1317 if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1318 (flags & LOP_NOSLEEP) == 0 &&
1319 lock1->li_lock == &Giant.lock_object)
1323 * If we are locking a sleepable lock and this lock
1324 * isn't sleepable, we want to treat it as a lock
1325 * order violation to enfore a general lock order of
1326 * sleepable locks before non-sleepable locks.
1328 if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1329 (flags & LOP_NOSLEEP) == 0 &&
1330 (lock1->li_flags & LI_SLEEPABLE) == 0)
1334 * If we are locking Giant and this is a non-sleepable
1335 * lock, then treat it as a reversal.
1337 if ((lock1->li_flags & LI_SLEEPABLE) == 0 &&
1338 lock == &Giant.lock_object)
1342 * Check the lock order hierarchy for a reveresal.
1344 if (!isitmydescendant(w, w1))
1349 * We have a lock order violation, check to see if it
1350 * is allowed or has already been yelled about.
1353 /* Bail if this violation is known */
1354 if (w_rmatrix[w1->w_index][w->w_index] & WITNESS_REVERSAL)
1357 /* Record this as a violation */
1358 w_rmatrix[w1->w_index][w->w_index] |= WITNESS_REVERSAL;
1359 w_rmatrix[w->w_index][w1->w_index] |= WITNESS_REVERSAL;
1360 w->w_reversed = w1->w_reversed = 1;
1361 witness_increment_graph_generation();
1364 * If the lock order is blessed, bail before logging
1365 * anything. We don't look for other lock order
1366 * violations though, which may be a bug.
1371 trace = atomic_load_int(&witness_trace);
1373 struct witness_lock_order_data *data;
1376 data = witness_lock_order_get(w, w1);
1378 stack_copy(&data->wlod_stack,
1383 mtx_unlock_spin(&w_mtx);
1385 #ifdef WITNESS_NO_VNODE
1387 * There are known LORs between VNODE locks. They are
1388 * not an indication of a bug. VNODE locks are flagged
1389 * as such (LO_IS_VNODE) and we don't yell if the LOR
1390 * is between 2 VNODE locks.
1392 if ((lock->lo_flags & LO_IS_VNODE) != 0 &&
1393 (lock1->li_lock->lo_flags & LO_IS_VNODE) != 0)
1398 * Ok, yell about it.
1400 if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1401 (flags & LOP_NOSLEEP) == 0 &&
1402 (lock1->li_flags & LI_SLEEPABLE) == 0)
1404 "lock order reversal: (sleepable after non-sleepable)\n");
1405 else if ((lock1->li_flags & LI_SLEEPABLE) == 0
1406 && lock == &Giant.lock_object)
1408 "lock order reversal: (Giant after non-sleepable)\n");
1410 witness_output("lock order reversal:\n");
1413 * Try to locate an earlier lock with
1414 * witness w in our list.
1417 lock2 = &lle->ll_children[i];
1418 MPASS(lock2->li_lock != NULL);
1419 if (lock2->li_lock->lo_witness == w)
1421 if (i == 0 && lle->ll_next != NULL) {
1423 i = lle->ll_count - 1;
1424 MPASS(i >= 0 && i < LOCK_NCHILDREN);
1429 witness_output(" 1st %p %s (%s, %s) @ %s:%d\n",
1430 lock1->li_lock, lock1->li_lock->lo_name,
1431 w1->w_name, w1->w_class->lc_name,
1432 fixup_filename(lock1->li_file),
1434 witness_output(" 2nd %p %s (%s, %s) @ %s:%d\n",
1435 lock, lock->lo_name, w->w_name,
1436 w->w_class->lc_name, fixup_filename(file),
1439 struct witness *w2 = lock2->li_lock->lo_witness;
1441 witness_output(" 1st %p %s (%s, %s) @ %s:%d\n",
1442 lock2->li_lock, lock2->li_lock->lo_name,
1443 w2->w_name, w2->w_class->lc_name,
1444 fixup_filename(lock2->li_file),
1446 witness_output(" 2nd %p %s (%s, %s) @ %s:%d\n",
1447 lock1->li_lock, lock1->li_lock->lo_name,
1448 w1->w_name, w1->w_class->lc_name,
1449 fixup_filename(lock1->li_file),
1451 witness_output(" 3rd %p %s (%s, %s) @ %s:%d\n", lock,
1452 lock->lo_name, w->w_name,
1453 w->w_class->lc_name, fixup_filename(file),
1460 sbuf_new(&sb, buf, sizeof(buf), SBUF_FIXEDLEN);
1461 sbuf_set_drain(&sb, witness_output_drain,
1466 "lock order %s -> %s established at:\n",
1467 w->w_name, w1->w_name);
1468 stack_sbuf_print_flags(&sb, &pstack,
1469 M_NOWAIT, STACK_SBUF_FMT_LONG);
1473 "lock order %s -> %s attempted at:\n",
1474 w1->w_name, w->w_name);
1475 stack_save(&pstack);
1476 stack_sbuf_print_flags(&sb, &pstack, M_NOWAIT,
1477 STACK_SBUF_FMT_LONG);
1482 witness_enter_debugger(__func__);
1488 * If requested, build a new lock order. However, don't build a new
1489 * relationship between a sleepable lock and Giant if it is in the
1490 * wrong direction. The correct lock order is that sleepable locks
1491 * always come before Giant.
1493 if (flags & LOP_NEWORDER &&
1494 !(plock->li_lock == &Giant.lock_object &&
1495 (lock->lo_flags & LO_SLEEPABLE) != 0 &&
1496 (flags & LOP_NOSLEEP) == 0)) {
1497 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1498 w->w_name, plock->li_lock->lo_witness->w_name);
1499 itismychild(plock->li_lock->lo_witness, w);
1502 mtx_unlock_spin(&w_mtx);
1506 witness_lock(struct lock_object *lock, int flags, const char *file, int line)
1508 struct lock_list_entry **lock_list, *lle;
1509 struct lock_instance *instance;
1513 if (witness_cold || witness_watch == -1 || lock->lo_witness == NULL ||
1516 w = lock->lo_witness;
1519 /* Determine lock list for this lock. */
1520 if (LOCK_CLASS(lock)->lc_flags & LC_SLEEPLOCK)
1521 lock_list = &td->td_sleeplocks;
1523 lock_list = PCPU_PTR(spinlocks);
1525 /* Check to see if we are recursing on a lock we already own. */
1526 instance = find_instance(*lock_list, lock);
1527 if (instance != NULL) {
1528 instance->li_flags++;
1529 CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
1530 td->td_proc->p_pid, lock->lo_name,
1531 instance->li_flags & LI_RECURSEMASK);
1532 instance->li_file = file;
1533 instance->li_line = line;
1537 /* Update per-witness last file and line acquire. */
1541 /* Find the next open lock instance in the list and fill it. */
1543 if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
1544 lle = witness_lock_list_get();
1547 lle->ll_next = *lock_list;
1548 CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
1549 td->td_proc->p_pid, lle);
1552 instance = &lle->ll_children[lle->ll_count++];
1553 instance->li_lock = lock;
1554 instance->li_line = line;
1555 instance->li_file = file;
1556 instance->li_flags = 0;
1557 if ((flags & LOP_EXCLUSIVE) != 0)
1558 instance->li_flags |= LI_EXCLUSIVE;
1559 if ((lock->lo_flags & LO_SLEEPABLE) != 0 && (flags & LOP_NOSLEEP) == 0)
1560 instance->li_flags |= LI_SLEEPABLE;
1561 CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
1562 td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
1566 witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
1568 struct lock_instance *instance;
1569 struct lock_class *class;
1571 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1572 if (lock->lo_witness == NULL || witness_watch == -1 || KERNEL_PANICKED())
1574 class = LOCK_CLASS(lock);
1575 if (witness_watch) {
1576 if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1578 "upgrade of non-upgradable lock (%s) %s @ %s:%d",
1579 class->lc_name, lock->lo_name,
1580 fixup_filename(file), line);
1581 if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1583 "upgrade of non-sleep lock (%s) %s @ %s:%d",
1584 class->lc_name, lock->lo_name,
1585 fixup_filename(file), line);
1587 instance = find_instance(curthread->td_sleeplocks, lock);
1588 if (instance == NULL) {
1589 kassert_panic("upgrade of unlocked lock (%s) %s @ %s:%d",
1590 class->lc_name, lock->lo_name,
1591 fixup_filename(file), line);
1594 if (witness_watch) {
1595 if ((instance->li_flags & LI_EXCLUSIVE) != 0)
1597 "upgrade of exclusive lock (%s) %s @ %s:%d",
1598 class->lc_name, lock->lo_name,
1599 fixup_filename(file), line);
1600 if ((instance->li_flags & LI_RECURSEMASK) != 0)
1602 "upgrade of recursed lock (%s) %s r=%d @ %s:%d",
1603 class->lc_name, lock->lo_name,
1604 instance->li_flags & LI_RECURSEMASK,
1605 fixup_filename(file), line);
1607 instance->li_flags |= LI_EXCLUSIVE;
1611 witness_downgrade(struct lock_object *lock, int flags, const char *file,
1614 struct lock_instance *instance;
1615 struct lock_class *class;
1617 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1618 if (lock->lo_witness == NULL || witness_watch == -1 || KERNEL_PANICKED())
1620 class = LOCK_CLASS(lock);
1621 if (witness_watch) {
1622 if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1624 "downgrade of non-upgradable lock (%s) %s @ %s:%d",
1625 class->lc_name, lock->lo_name,
1626 fixup_filename(file), line);
1627 if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1629 "downgrade of non-sleep lock (%s) %s @ %s:%d",
1630 class->lc_name, lock->lo_name,
1631 fixup_filename(file), line);
1633 instance = find_instance(curthread->td_sleeplocks, lock);
1634 if (instance == NULL) {
1635 kassert_panic("downgrade of unlocked lock (%s) %s @ %s:%d",
1636 class->lc_name, lock->lo_name,
1637 fixup_filename(file), line);
1640 if (witness_watch) {
1641 if ((instance->li_flags & LI_EXCLUSIVE) == 0)
1643 "downgrade of shared lock (%s) %s @ %s:%d",
1644 class->lc_name, lock->lo_name,
1645 fixup_filename(file), line);
1646 if ((instance->li_flags & LI_RECURSEMASK) != 0)
1648 "downgrade of recursed lock (%s) %s r=%d @ %s:%d",
1649 class->lc_name, lock->lo_name,
1650 instance->li_flags & LI_RECURSEMASK,
1651 fixup_filename(file), line);
1653 instance->li_flags &= ~LI_EXCLUSIVE;
1657 witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
1659 struct lock_list_entry **lock_list, *lle;
1660 struct lock_instance *instance;
1661 struct lock_class *class;
1666 if (witness_cold || lock->lo_witness == NULL || KERNEL_PANICKED())
1669 class = LOCK_CLASS(lock);
1671 /* Find lock instance associated with this lock. */
1672 if (class->lc_flags & LC_SLEEPLOCK)
1673 lock_list = &td->td_sleeplocks;
1675 lock_list = PCPU_PTR(spinlocks);
1677 for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
1678 for (i = 0; i < (*lock_list)->ll_count; i++) {
1679 instance = &(*lock_list)->ll_children[i];
1680 if (instance->li_lock == lock)
1685 * When disabling WITNESS through witness_watch we could end up in
1686 * having registered locks in the td_sleeplocks queue.
1687 * We have to make sure we flush these queues, so just search for
1688 * eventual register locks and remove them.
1690 if (witness_watch > 0) {
1691 kassert_panic("lock (%s) %s not locked @ %s:%d", class->lc_name,
1692 lock->lo_name, fixup_filename(file), line);
1699 /* First, check for shared/exclusive mismatches. */
1700 if ((instance->li_flags & LI_EXCLUSIVE) != 0 && witness_watch > 0 &&
1701 (flags & LOP_EXCLUSIVE) == 0) {
1702 witness_output("shared unlock of (%s) %s @ %s:%d\n",
1703 class->lc_name, lock->lo_name, fixup_filename(file), line);
1704 witness_output("while exclusively locked from %s:%d\n",
1705 fixup_filename(instance->li_file), instance->li_line);
1706 kassert_panic("excl->ushare");
1708 if ((instance->li_flags & LI_EXCLUSIVE) == 0 && witness_watch > 0 &&
1709 (flags & LOP_EXCLUSIVE) != 0) {
1710 witness_output("exclusive unlock of (%s) %s @ %s:%d\n",
1711 class->lc_name, lock->lo_name, fixup_filename(file), line);
1712 witness_output("while share locked from %s:%d\n",
1713 fixup_filename(instance->li_file),
1715 kassert_panic("share->uexcl");
1717 /* If we are recursed, unrecurse. */
1718 if ((instance->li_flags & LI_RECURSEMASK) > 0) {
1719 CTR4(KTR_WITNESS, "%s: pid %d unrecursed on %s r=%d", __func__,
1720 td->td_proc->p_pid, instance->li_lock->lo_name,
1721 instance->li_flags);
1722 instance->li_flags--;
1725 /* The lock is now being dropped, check for NORELEASE flag */
1726 if ((instance->li_flags & LI_NORELEASE) != 0 && witness_watch > 0) {
1727 witness_output("forbidden unlock of (%s) %s @ %s:%d\n",
1728 class->lc_name, lock->lo_name, fixup_filename(file), line);
1729 kassert_panic("lock marked norelease");
1732 /* Otherwise, remove this item from the list. */
1734 CTR4(KTR_WITNESS, "%s: pid %d removed %s from lle[%d]", __func__,
1735 td->td_proc->p_pid, instance->li_lock->lo_name,
1736 (*lock_list)->ll_count - 1);
1737 for (j = i; j < (*lock_list)->ll_count - 1; j++)
1738 (*lock_list)->ll_children[j] =
1739 (*lock_list)->ll_children[j + 1];
1740 (*lock_list)->ll_count--;
1744 * In order to reduce contention on w_mtx, we want to keep always an
1745 * head object into lists so that frequent allocation from the
1746 * free witness pool (and subsequent locking) is avoided.
1747 * In order to maintain the current code simple, when the head
1748 * object is totally unloaded it means also that we do not have
1749 * further objects in the list, so the list ownership needs to be
1750 * hand over to another object if the current head needs to be freed.
1752 if ((*lock_list)->ll_count == 0) {
1753 if (*lock_list == lle) {
1754 if (lle->ll_next == NULL)
1758 *lock_list = lle->ll_next;
1759 CTR3(KTR_WITNESS, "%s: pid %d removed lle %p", __func__,
1760 td->td_proc->p_pid, lle);
1761 witness_lock_list_free(lle);
1766 witness_thread_exit(struct thread *td)
1768 struct lock_list_entry *lle;
1771 lle = td->td_sleeplocks;
1772 if (lle == NULL || KERNEL_PANICKED())
1774 if (lle->ll_count != 0) {
1775 for (n = 0; lle != NULL; lle = lle->ll_next)
1776 for (i = lle->ll_count - 1; i >= 0; i--) {
1779 "Thread %p exiting with the following locks held:\n", td);
1781 witness_list_lock(&lle->ll_children[i],
1786 "Thread %p cannot exit while holding sleeplocks\n", td);
1788 witness_lock_list_free(lle);
1792 * Warn if any locks other than 'lock' are held. Flags can be passed in to
1793 * exempt Giant and sleepable locks from the checks as well. If any
1794 * non-exempt locks are held, then a supplied message is printed to the
1795 * output channel along with a list of the offending locks. If indicated in the
1796 * flags then a failure results in a panic as well.
1799 witness_warn(int flags, struct lock_object *lock, const char *fmt, ...)
1801 struct lock_list_entry *lock_list, *lle;
1802 struct lock_instance *lock1;
1807 if (witness_cold || witness_watch < 1 || KERNEL_PANICKED())
1811 for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next)
1812 for (i = lle->ll_count - 1; i >= 0; i--) {
1813 lock1 = &lle->ll_children[i];
1814 if (lock1->li_lock == lock)
1816 if (flags & WARN_GIANTOK &&
1817 lock1->li_lock == &Giant.lock_object)
1819 if (flags & WARN_SLEEPOK &&
1820 (lock1->li_flags & LI_SLEEPABLE) != 0)
1826 printf(" with the following %slocks held:\n",
1827 (flags & WARN_SLEEPOK) != 0 ?
1828 "non-sleepable " : "");
1831 witness_list_lock(lock1, printf);
1835 * Pin the thread in order to avoid problems with thread migration.
1836 * Once that all verifies are passed about spinlocks ownership,
1837 * the thread is in a safe path and it can be unpinned.
1840 lock_list = PCPU_GET(spinlocks);
1841 if (lock_list != NULL && lock_list->ll_count != 0) {
1845 * We should only have one spinlock and as long as
1846 * the flags cannot match for this locks class,
1847 * check if the first spinlock is the one curthread
1850 lock1 = &lock_list->ll_children[lock_list->ll_count - 1];
1851 if (lock_list->ll_count == 1 && lock_list->ll_next == NULL &&
1852 lock1->li_lock == lock && n == 0)
1858 printf(" with the following %slocks held:\n",
1859 (flags & WARN_SLEEPOK) != 0 ? "non-sleepable " : "");
1860 n += witness_list_locks(&lock_list, printf);
1863 if (flags & WARN_PANIC && n)
1864 kassert_panic("%s", __func__);
1866 witness_debugger(n, __func__);
1871 witness_file(struct lock_object *lock)
1875 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1877 w = lock->lo_witness;
1882 witness_line(struct lock_object *lock)
1886 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1888 w = lock->lo_witness;
1892 static struct witness *
1893 enroll(const char *description, struct lock_class *lock_class)
1897 MPASS(description != NULL);
1899 if (witness_watch == -1 || KERNEL_PANICKED())
1901 if ((lock_class->lc_flags & LC_SPINLOCK)) {
1902 if (witness_skipspin)
1904 } else if ((lock_class->lc_flags & LC_SLEEPLOCK) == 0) {
1905 kassert_panic("lock class %s is not sleep or spin",
1906 lock_class->lc_name);
1910 mtx_lock_spin(&w_mtx);
1911 w = witness_hash_get(description);
1914 if ((w = witness_get()) == NULL)
1916 MPASS(strlen(description) < MAX_W_NAME);
1917 strcpy(w->w_name, description);
1918 w->w_class = lock_class;
1920 STAILQ_INSERT_HEAD(&w_all, w, w_list);
1921 if (lock_class->lc_flags & LC_SPINLOCK) {
1922 STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
1924 } else if (lock_class->lc_flags & LC_SLEEPLOCK) {
1925 STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
1929 /* Insert new witness into the hash */
1930 witness_hash_put(w);
1931 witness_increment_graph_generation();
1932 mtx_unlock_spin(&w_mtx);
1936 if (w->w_refcount == 1)
1937 w->w_class = lock_class;
1938 mtx_unlock_spin(&w_mtx);
1939 if (lock_class != w->w_class)
1941 "lock (%s) %s does not match earlier (%s) lock",
1942 description, lock_class->lc_name,
1943 w->w_class->lc_name);
1948 depart(struct witness *w)
1951 MPASS(w->w_refcount == 0);
1952 if (w->w_class->lc_flags & LC_SLEEPLOCK) {
1958 * Set file to NULL as it may point into a loadable module.
1962 witness_increment_graph_generation();
1966 adopt(struct witness *parent, struct witness *child)
1970 if (witness_cold == 0)
1971 mtx_assert(&w_mtx, MA_OWNED);
1973 /* If the relationship is already known, there's no work to be done. */
1974 if (isitmychild(parent, child))
1977 /* When the structure of the graph changes, bump up the generation. */
1978 witness_increment_graph_generation();
1981 * The hard part ... create the direct relationship, then propagate all
1982 * indirect relationships.
1984 pi = parent->w_index;
1985 ci = child->w_index;
1986 WITNESS_INDEX_ASSERT(pi);
1987 WITNESS_INDEX_ASSERT(ci);
1989 w_rmatrix[pi][ci] |= WITNESS_PARENT;
1990 w_rmatrix[ci][pi] |= WITNESS_CHILD;
1993 * If parent was not already an ancestor of child,
1994 * then we increment the descendant and ancestor counters.
1996 if ((w_rmatrix[pi][ci] & WITNESS_ANCESTOR) == 0) {
1997 parent->w_num_descendants++;
1998 child->w_num_ancestors++;
2002 * Find each ancestor of 'pi'. Note that 'pi' itself is counted as
2003 * an ancestor of 'pi' during this loop.
2005 for (i = 1; i <= w_max_used_index; i++) {
2006 if ((w_rmatrix[i][pi] & WITNESS_ANCESTOR_MASK) == 0 &&
2010 /* Find each descendant of 'i' and mark it as a descendant. */
2011 for (j = 1; j <= w_max_used_index; j++) {
2013 * Skip children that are already marked as
2014 * descendants of 'i'.
2016 if (w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK)
2020 * We are only interested in descendants of 'ci'. Note
2021 * that 'ci' itself is counted as a descendant of 'ci'.
2023 if ((w_rmatrix[ci][j] & WITNESS_ANCESTOR_MASK) == 0 &&
2026 w_rmatrix[i][j] |= WITNESS_ANCESTOR;
2027 w_rmatrix[j][i] |= WITNESS_DESCENDANT;
2028 w_data[i].w_num_descendants++;
2029 w_data[j].w_num_ancestors++;
2032 * Make sure we aren't marking a node as both an
2033 * ancestor and descendant. We should have caught
2034 * this as a lock order reversal earlier.
2036 if ((w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK) &&
2037 (w_rmatrix[i][j] & WITNESS_DESCENDANT_MASK)) {
2038 printf("witness rmatrix paradox! [%d][%d]=%d "
2039 "both ancestor and descendant\n",
2040 i, j, w_rmatrix[i][j]);
2042 printf("Witness disabled.\n");
2045 if ((w_rmatrix[j][i] & WITNESS_ANCESTOR_MASK) &&
2046 (w_rmatrix[j][i] & WITNESS_DESCENDANT_MASK)) {
2047 printf("witness rmatrix paradox! [%d][%d]=%d "
2048 "both ancestor and descendant\n",
2049 j, i, w_rmatrix[j][i]);
2051 printf("Witness disabled.\n");
2059 itismychild(struct witness *parent, struct witness *child)
2063 MPASS(child != NULL && parent != NULL);
2064 if (witness_cold == 0)
2065 mtx_assert(&w_mtx, MA_OWNED);
2067 if (!witness_lock_type_equal(parent, child)) {
2068 if (witness_cold == 0) {
2070 mtx_unlock_spin(&w_mtx);
2075 "%s: parent \"%s\" (%s) and child \"%s\" (%s) are not "
2076 "the same lock type", __func__, parent->w_name,
2077 parent->w_class->lc_name, child->w_name,
2078 child->w_class->lc_name);
2080 mtx_lock_spin(&w_mtx);
2082 adopt(parent, child);
2086 * Generic code for the isitmy*() functions. The rmask parameter is the
2087 * expected relationship of w1 to w2.
2090 _isitmyx(struct witness *w1, struct witness *w2, int rmask, const char *fname)
2092 unsigned char r1, r2;
2097 WITNESS_INDEX_ASSERT(i1);
2098 WITNESS_INDEX_ASSERT(i2);
2099 r1 = w_rmatrix[i1][i2] & WITNESS_RELATED_MASK;
2100 r2 = w_rmatrix[i2][i1] & WITNESS_RELATED_MASK;
2102 /* The flags on one better be the inverse of the flags on the other */
2103 if (!((WITNESS_ATOD(r1) == r2 && WITNESS_DTOA(r2) == r1) ||
2104 (WITNESS_DTOA(r1) == r2 && WITNESS_ATOD(r2) == r1))) {
2105 /* Don't squawk if we're potentially racing with an update. */
2106 if (!mtx_owned(&w_mtx))
2108 printf("%s: rmatrix mismatch between %s (index %d) and %s "
2109 "(index %d): w_rmatrix[%d][%d] == %hhx but "
2110 "w_rmatrix[%d][%d] == %hhx\n",
2111 fname, w1->w_name, i1, w2->w_name, i2, i1, i2, r1,
2114 printf("Witness disabled.\n");
2117 return (r1 & rmask);
2121 * Checks if @child is a direct child of @parent.
2124 isitmychild(struct witness *parent, struct witness *child)
2127 return (_isitmyx(parent, child, WITNESS_PARENT, __func__));
2131 * Checks if @descendant is a direct or inderect descendant of @ancestor.
2134 isitmydescendant(struct witness *ancestor, struct witness *descendant)
2137 return (_isitmyx(ancestor, descendant, WITNESS_ANCESTOR_MASK,
2142 blessed(struct witness *w1, struct witness *w2)
2145 struct witness_blessed *b;
2147 for (i = 0; i < nitems(blessed_list); i++) {
2148 b = &blessed_list[i];
2149 if (strcmp(w1->w_name, b->b_lock1) == 0) {
2150 if (strcmp(w2->w_name, b->b_lock2) == 0)
2154 if (strcmp(w1->w_name, b->b_lock2) == 0)
2155 if (strcmp(w2->w_name, b->b_lock1) == 0)
2161 static struct witness *
2167 if (witness_cold == 0)
2168 mtx_assert(&w_mtx, MA_OWNED);
2170 if (witness_watch == -1) {
2171 mtx_unlock_spin(&w_mtx);
2174 if (STAILQ_EMPTY(&w_free)) {
2176 mtx_unlock_spin(&w_mtx);
2177 printf("WITNESS: unable to allocate a new witness object\n");
2180 w = STAILQ_FIRST(&w_free);
2181 STAILQ_REMOVE_HEAD(&w_free, w_list);
2184 MPASS(index > 0 && index == w_max_used_index+1 &&
2185 index < witness_count);
2186 bzero(w, sizeof(*w));
2188 if (index > w_max_used_index)
2189 w_max_used_index = index;
2194 witness_free(struct witness *w)
2197 STAILQ_INSERT_HEAD(&w_free, w, w_list);
2201 static struct lock_list_entry *
2202 witness_lock_list_get(void)
2204 struct lock_list_entry *lle;
2206 if (witness_watch == -1)
2208 mtx_lock_spin(&w_mtx);
2209 lle = w_lock_list_free;
2212 mtx_unlock_spin(&w_mtx);
2213 printf("%s: witness exhausted\n", __func__);
2216 w_lock_list_free = lle->ll_next;
2217 mtx_unlock_spin(&w_mtx);
2218 bzero(lle, sizeof(*lle));
2223 witness_lock_list_free(struct lock_list_entry *lle)
2226 mtx_lock_spin(&w_mtx);
2227 lle->ll_next = w_lock_list_free;
2228 w_lock_list_free = lle;
2229 mtx_unlock_spin(&w_mtx);
2232 static struct lock_instance *
2233 find_instance(struct lock_list_entry *list, const struct lock_object *lock)
2235 struct lock_list_entry *lle;
2236 struct lock_instance *instance;
2239 for (lle = list; lle != NULL; lle = lle->ll_next)
2240 for (i = lle->ll_count - 1; i >= 0; i--) {
2241 instance = &lle->ll_children[i];
2242 if (instance->li_lock == lock)
2249 witness_list_lock(struct lock_instance *instance,
2250 int (*prnt)(const char *fmt, ...))
2252 struct lock_object *lock;
2254 lock = instance->li_lock;
2255 prnt("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ?
2256 "exclusive" : "shared", LOCK_CLASS(lock)->lc_name, lock->lo_name);
2257 if (lock->lo_witness->w_name != lock->lo_name)
2258 prnt(" (%s)", lock->lo_witness->w_name);
2259 prnt(" r = %d (%p) locked @ %s:%d\n",
2260 instance->li_flags & LI_RECURSEMASK, lock,
2261 fixup_filename(instance->li_file), instance->li_line);
2265 witness_output(const char *fmt, ...)
2271 ret = witness_voutput(fmt, ap);
2277 witness_voutput(const char *fmt, va_list ap)
2282 switch (witness_channel) {
2283 case WITNESS_CONSOLE:
2284 ret = vprintf(fmt, ap);
2287 vlog(LOG_NOTICE, fmt, ap);
2297 witness_thread_has_locks(struct thread *td)
2300 if (td->td_sleeplocks == NULL)
2302 return (td->td_sleeplocks->ll_count != 0);
2306 witness_proc_has_locks(struct proc *p)
2310 FOREACH_THREAD_IN_PROC(p, td) {
2311 if (witness_thread_has_locks(td))
2319 witness_list_locks(struct lock_list_entry **lock_list,
2320 int (*prnt)(const char *fmt, ...))
2322 struct lock_list_entry *lle;
2326 for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
2327 for (i = lle->ll_count - 1; i >= 0; i--) {
2328 witness_list_lock(&lle->ll_children[i], prnt);
2335 * This is a bit risky at best. We call this function when we have timed
2336 * out acquiring a spin lock, and we assume that the other CPU is stuck
2337 * with this lock held. So, we go groveling around in the other CPU's
2338 * per-cpu data to try to find the lock instance for this spin lock to
2339 * see when it was last acquired.
2342 witness_display_spinlock(struct lock_object *lock, struct thread *owner,
2343 int (*prnt)(const char *fmt, ...))
2345 struct lock_instance *instance;
2348 if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU)
2350 pc = pcpu_find(owner->td_oncpu);
2351 instance = find_instance(pc->pc_spinlocks, lock);
2352 if (instance != NULL)
2353 witness_list_lock(instance, prnt);
2357 witness_save(struct lock_object *lock, const char **filep, int *linep)
2359 struct lock_list_entry *lock_list;
2360 struct lock_instance *instance;
2361 struct lock_class *class;
2364 * This function is used independently in locking code to deal with
2365 * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2368 if (SCHEDULER_STOPPED())
2370 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2371 if (lock->lo_witness == NULL || witness_watch == -1 || KERNEL_PANICKED())
2373 class = LOCK_CLASS(lock);
2374 if (class->lc_flags & LC_SLEEPLOCK)
2375 lock_list = curthread->td_sleeplocks;
2377 if (witness_skipspin)
2379 lock_list = PCPU_GET(spinlocks);
2381 instance = find_instance(lock_list, lock);
2382 if (instance == NULL) {
2383 kassert_panic("%s: lock (%s) %s not locked", __func__,
2384 class->lc_name, lock->lo_name);
2387 *filep = instance->li_file;
2388 *linep = instance->li_line;
2392 witness_restore(struct lock_object *lock, const char *file, int line)
2394 struct lock_list_entry *lock_list;
2395 struct lock_instance *instance;
2396 struct lock_class *class;
2399 * This function is used independently in locking code to deal with
2400 * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2403 if (SCHEDULER_STOPPED())
2405 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2406 if (lock->lo_witness == NULL || witness_watch == -1 || KERNEL_PANICKED())
2408 class = LOCK_CLASS(lock);
2409 if (class->lc_flags & LC_SLEEPLOCK)
2410 lock_list = curthread->td_sleeplocks;
2412 if (witness_skipspin)
2414 lock_list = PCPU_GET(spinlocks);
2416 instance = find_instance(lock_list, lock);
2417 if (instance == NULL)
2418 kassert_panic("%s: lock (%s) %s not locked", __func__,
2419 class->lc_name, lock->lo_name);
2420 lock->lo_witness->w_file = file;
2421 lock->lo_witness->w_line = line;
2422 if (instance == NULL)
2424 instance->li_file = file;
2425 instance->li_line = line;
2429 witness_assert(const struct lock_object *lock, int flags, const char *file,
2432 #ifdef INVARIANT_SUPPORT
2433 struct lock_instance *instance;
2434 struct lock_class *class;
2436 if (lock->lo_witness == NULL || witness_watch < 1 || KERNEL_PANICKED())
2438 class = LOCK_CLASS(lock);
2439 if ((class->lc_flags & LC_SLEEPLOCK) != 0)
2440 instance = find_instance(curthread->td_sleeplocks, lock);
2441 else if ((class->lc_flags & LC_SPINLOCK) != 0)
2442 instance = find_instance(PCPU_GET(spinlocks), lock);
2444 kassert_panic("Lock (%s) %s is not sleep or spin!",
2445 class->lc_name, lock->lo_name);
2450 if (instance != NULL)
2451 kassert_panic("Lock (%s) %s locked @ %s:%d.",
2452 class->lc_name, lock->lo_name,
2453 fixup_filename(file), line);
2456 case LA_LOCKED | LA_RECURSED:
2457 case LA_LOCKED | LA_NOTRECURSED:
2459 case LA_SLOCKED | LA_RECURSED:
2460 case LA_SLOCKED | LA_NOTRECURSED:
2462 case LA_XLOCKED | LA_RECURSED:
2463 case LA_XLOCKED | LA_NOTRECURSED:
2464 if (instance == NULL) {
2465 kassert_panic("Lock (%s) %s not locked @ %s:%d.",
2466 class->lc_name, lock->lo_name,
2467 fixup_filename(file), line);
2470 if ((flags & LA_XLOCKED) != 0 &&
2471 (instance->li_flags & LI_EXCLUSIVE) == 0)
2473 "Lock (%s) %s not exclusively locked @ %s:%d.",
2474 class->lc_name, lock->lo_name,
2475 fixup_filename(file), line);
2476 if ((flags & LA_SLOCKED) != 0 &&
2477 (instance->li_flags & LI_EXCLUSIVE) != 0)
2479 "Lock (%s) %s exclusively locked @ %s:%d.",
2480 class->lc_name, lock->lo_name,
2481 fixup_filename(file), line);
2482 if ((flags & LA_RECURSED) != 0 &&
2483 (instance->li_flags & LI_RECURSEMASK) == 0)
2484 kassert_panic("Lock (%s) %s not recursed @ %s:%d.",
2485 class->lc_name, lock->lo_name,
2486 fixup_filename(file), line);
2487 if ((flags & LA_NOTRECURSED) != 0 &&
2488 (instance->li_flags & LI_RECURSEMASK) != 0)
2489 kassert_panic("Lock (%s) %s recursed @ %s:%d.",
2490 class->lc_name, lock->lo_name,
2491 fixup_filename(file), line);
2494 kassert_panic("Invalid lock assertion at %s:%d.",
2495 fixup_filename(file), line);
2497 #endif /* INVARIANT_SUPPORT */
2501 witness_setflag(struct lock_object *lock, int flag, int set)
2503 struct lock_list_entry *lock_list;
2504 struct lock_instance *instance;
2505 struct lock_class *class;
2507 if (lock->lo_witness == NULL || witness_watch == -1 || KERNEL_PANICKED())
2509 class = LOCK_CLASS(lock);
2510 if (class->lc_flags & LC_SLEEPLOCK)
2511 lock_list = curthread->td_sleeplocks;
2513 if (witness_skipspin)
2515 lock_list = PCPU_GET(spinlocks);
2517 instance = find_instance(lock_list, lock);
2518 if (instance == NULL) {
2519 kassert_panic("%s: lock (%s) %s not locked", __func__,
2520 class->lc_name, lock->lo_name);
2525 instance->li_flags |= flag;
2527 instance->li_flags &= ~flag;
2531 witness_norelease(struct lock_object *lock)
2534 witness_setflag(lock, LI_NORELEASE, 1);
2538 witness_releaseok(struct lock_object *lock)
2541 witness_setflag(lock, LI_NORELEASE, 0);
2546 witness_ddb_list(struct thread *td)
2549 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2550 KASSERT(kdb_active, ("%s: not in the debugger", __func__));
2552 if (witness_watch < 1)
2555 witness_list_locks(&td->td_sleeplocks, db_printf);
2558 * We only handle spinlocks if td == curthread. This is somewhat broken
2559 * if td is currently executing on some other CPU and holds spin locks
2560 * as we won't display those locks. If we had a MI way of getting
2561 * the per-cpu data for a given cpu then we could use
2562 * td->td_oncpu to get the list of spinlocks for this thread
2565 * That still wouldn't really fix this unless we locked the scheduler
2566 * lock or stopped the other CPU to make sure it wasn't changing the
2567 * list out from under us. It is probably best to just not try to
2568 * handle threads on other CPU's for now.
2570 if (td == curthread && PCPU_GET(spinlocks) != NULL)
2571 witness_list_locks(PCPU_PTR(spinlocks), db_printf);
2574 DB_SHOW_COMMAND(locks, db_witness_list)
2579 td = db_lookup_thread(addr, true);
2582 witness_ddb_list(td);
2585 DB_SHOW_ALL_COMMAND(locks, db_witness_list_all)
2591 * It would be nice to list only threads and processes that actually
2592 * held sleep locks, but that information is currently not exported
2595 FOREACH_PROC_IN_SYSTEM(p) {
2596 if (!witness_proc_has_locks(p))
2598 FOREACH_THREAD_IN_PROC(p, td) {
2599 if (!witness_thread_has_locks(td))
2601 db_printf("Process %d (%s) thread %p (%d)\n", p->p_pid,
2602 p->p_comm, td, td->td_tid);
2603 witness_ddb_list(td);
2609 DB_SHOW_ALIAS(alllocks, db_witness_list_all)
2611 DB_SHOW_COMMAND(witness, db_witness_display)
2614 witness_ddb_display(db_printf);
2619 sbuf_print_witness_badstacks(struct sbuf *sb, size_t *oldidx)
2621 struct witness_lock_order_data *data1, *data2, *tmp_data1, *tmp_data2;
2622 struct witness *tmp_w1, *tmp_w2, *w1, *w2;
2623 int generation, i, j;
2630 /* Allocate and init temporary storage space. */
2631 tmp_w1 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2632 tmp_w2 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2633 tmp_data1 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2635 tmp_data2 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2637 stack_zero(&tmp_data1->wlod_stack);
2638 stack_zero(&tmp_data2->wlod_stack);
2641 mtx_lock_spin(&w_mtx);
2642 generation = w_generation;
2643 mtx_unlock_spin(&w_mtx);
2644 sbuf_printf(sb, "Number of known direct relationships is %d\n",
2645 w_lohash.wloh_count);
2646 for (i = 1; i < w_max_used_index; i++) {
2647 mtx_lock_spin(&w_mtx);
2648 if (generation != w_generation) {
2649 mtx_unlock_spin(&w_mtx);
2651 /* The graph has changed, try again. */
2658 if (w1->w_reversed == 0) {
2659 mtx_unlock_spin(&w_mtx);
2663 /* Copy w1 locally so we can release the spin lock. */
2665 mtx_unlock_spin(&w_mtx);
2667 if (tmp_w1->w_reversed == 0)
2669 for (j = 1; j < w_max_used_index; j++) {
2670 if ((w_rmatrix[i][j] & WITNESS_REVERSAL) == 0 || i > j)
2673 mtx_lock_spin(&w_mtx);
2674 if (generation != w_generation) {
2675 mtx_unlock_spin(&w_mtx);
2677 /* The graph has changed, try again. */
2684 data1 = witness_lock_order_get(w1, w2);
2685 data2 = witness_lock_order_get(w2, w1);
2688 * Copy information locally so we can release the
2694 stack_zero(&tmp_data1->wlod_stack);
2695 stack_copy(&data1->wlod_stack,
2696 &tmp_data1->wlod_stack);
2698 if (data2 && data2 != data1) {
2699 stack_zero(&tmp_data2->wlod_stack);
2700 stack_copy(&data2->wlod_stack,
2701 &tmp_data2->wlod_stack);
2703 mtx_unlock_spin(&w_mtx);
2705 if (blessed(tmp_w1, tmp_w2))
2709 "\nLock order reversal between \"%s\"(%s) and \"%s\"(%s)!\n",
2710 tmp_w1->w_name, tmp_w1->w_class->lc_name,
2711 tmp_w2->w_name, tmp_w2->w_class->lc_name);
2714 "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2715 tmp_w1->w_name, tmp_w1->w_class->lc_name,
2716 tmp_w2->w_name, tmp_w2->w_class->lc_name);
2717 stack_sbuf_print(sb, &tmp_data1->wlod_stack);
2718 sbuf_printf(sb, "\n");
2720 if (data2 && data2 != data1) {
2722 "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2723 tmp_w2->w_name, tmp_w2->w_class->lc_name,
2724 tmp_w1->w_name, tmp_w1->w_class->lc_name);
2725 stack_sbuf_print(sb, &tmp_data2->wlod_stack);
2726 sbuf_printf(sb, "\n");
2730 mtx_lock_spin(&w_mtx);
2731 if (generation != w_generation) {
2732 mtx_unlock_spin(&w_mtx);
2735 * The graph changed while we were printing stack data,
2742 mtx_unlock_spin(&w_mtx);
2744 /* Free temporary storage space. */
2745 free(tmp_data1, M_TEMP);
2746 free(tmp_data2, M_TEMP);
2747 free(tmp_w1, M_TEMP);
2748 free(tmp_w2, M_TEMP);
2752 sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS)
2757 if (witness_watch < 1) {
2758 error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2762 error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2766 sb = sbuf_new(NULL, NULL, badstack_sbuf_size, SBUF_AUTOEXTEND);
2770 sbuf_print_witness_badstacks(sb, &req->oldidx);
2773 error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1);
2781 sbuf_db_printf_drain(void *arg __unused, const char *data, int len)
2784 return (db_printf("%.*s", len, data));
2787 DB_SHOW_COMMAND(badstacks, db_witness_badstacks)
2793 sbuf_new(&sb, buffer, sizeof(buffer), SBUF_FIXEDLEN);
2794 sbuf_set_drain(&sb, sbuf_db_printf_drain, NULL);
2795 sbuf_print_witness_badstacks(&sb, &dummy);
2801 sysctl_debug_witness_channel(SYSCTL_HANDLER_ARGS)
2803 static const struct {
2804 enum witness_channel channel;
2807 { WITNESS_CONSOLE, "console" },
2808 { WITNESS_LOG, "log" },
2809 { WITNESS_NONE, "none" },
2816 for (i = 0; i < nitems(channels); i++)
2817 if (witness_channel == channels[i].channel) {
2818 snprintf(buf, sizeof(buf), "%s", channels[i].name);
2822 error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
2823 if (error != 0 || req->newptr == NULL)
2827 for (i = 0; i < nitems(channels); i++)
2828 if (strcmp(channels[i].name, buf) == 0) {
2829 witness_channel = channels[i].channel;
2837 sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS)
2844 error = SYSCTL_OUT(req, w_notallowed, sizeof(w_notallowed));
2848 if (witness_watch < 1) {
2849 error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2853 error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2858 error = sysctl_wire_old_buffer(req, 0);
2861 sb = sbuf_new_for_sysctl(NULL, NULL, FULLGRAPH_SBUF_SIZE, req);
2864 sbuf_printf(sb, "\n");
2866 mtx_lock_spin(&w_mtx);
2867 STAILQ_FOREACH(w, &w_all, w_list)
2869 STAILQ_FOREACH(w, &w_all, w_list)
2870 witness_add_fullgraph(sb, w);
2871 mtx_unlock_spin(&w_mtx);
2874 * Close the sbuf and return to userland.
2876 error = sbuf_finish(sb);
2883 sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS)
2887 value = witness_watch;
2888 error = sysctl_handle_int(oidp, &value, 0, req);
2889 if (error != 0 || req->newptr == NULL)
2891 if (value > 1 || value < -1 ||
2892 (witness_watch == -1 && value != witness_watch))
2894 witness_watch = value;
2899 witness_add_fullgraph(struct sbuf *sb, struct witness *w)
2903 if (w->w_displayed != 0 || (w->w_file == NULL && w->w_line == 0))
2907 WITNESS_INDEX_ASSERT(w->w_index);
2908 for (i = 1; i <= w_max_used_index; i++) {
2909 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT) {
2910 sbuf_printf(sb, "\"%s\",\"%s\"\n", w->w_name,
2912 witness_add_fullgraph(sb, &w_data[i]);
2918 * A simple hash function. Takes a key pointer and a key size. If size == 0,
2919 * interprets the key as a string and reads until the null
2920 * terminator. Otherwise, reads the first size bytes. Returns an unsigned 32-bit
2921 * hash value computed from the key.
2924 witness_hash_djb2(const uint8_t *key, uint32_t size)
2926 unsigned int hash = 5381;
2929 /* hash = hash * 33 + key[i] */
2931 for (i = 0; i < size; i++)
2932 hash = ((hash << 5) + hash) + (unsigned int)key[i];
2934 for (i = 0; key[i] != 0; i++)
2935 hash = ((hash << 5) + hash) + (unsigned int)key[i];
2941 * Initializes the two witness hash tables. Called exactly once from
2942 * witness_initialize().
2945 witness_init_hash_tables(void)
2949 MPASS(witness_cold);
2951 /* Initialize the hash tables. */
2952 for (i = 0; i < WITNESS_HASH_SIZE; i++)
2953 w_hash.wh_array[i] = NULL;
2955 w_hash.wh_size = WITNESS_HASH_SIZE;
2956 w_hash.wh_count = 0;
2958 /* Initialize the lock order data hash. */
2960 for (i = 0; i < WITNESS_LO_DATA_COUNT; i++) {
2961 memset(&w_lodata[i], 0, sizeof(w_lodata[i]));
2962 w_lodata[i].wlod_next = w_lofree;
2963 w_lofree = &w_lodata[i];
2965 w_lohash.wloh_size = WITNESS_LO_HASH_SIZE;
2966 w_lohash.wloh_count = 0;
2967 for (i = 0; i < WITNESS_LO_HASH_SIZE; i++)
2968 w_lohash.wloh_array[i] = NULL;
2971 static struct witness *
2972 witness_hash_get(const char *key)
2978 if (witness_cold == 0)
2979 mtx_assert(&w_mtx, MA_OWNED);
2980 hash = witness_hash_djb2(key, 0) % w_hash.wh_size;
2981 w = w_hash.wh_array[hash];
2983 if (strcmp(w->w_name, key) == 0)
2993 witness_hash_put(struct witness *w)
2998 MPASS(w->w_name != NULL);
2999 if (witness_cold == 0)
3000 mtx_assert(&w_mtx, MA_OWNED);
3001 KASSERT(witness_hash_get(w->w_name) == NULL,
3002 ("%s: trying to add a hash entry that already exists!", __func__));
3003 KASSERT(w->w_hash_next == NULL,
3004 ("%s: w->w_hash_next != NULL", __func__));
3006 hash = witness_hash_djb2(w->w_name, 0) % w_hash.wh_size;
3007 w->w_hash_next = w_hash.wh_array[hash];
3008 w_hash.wh_array[hash] = w;
3012 static struct witness_lock_order_data *
3013 witness_lock_order_get(struct witness *parent, struct witness *child)
3015 struct witness_lock_order_data *data = NULL;
3016 struct witness_lock_order_key key;
3019 MPASS(parent != NULL && child != NULL);
3020 key.from = parent->w_index;
3021 key.to = child->w_index;
3022 WITNESS_INDEX_ASSERT(key.from);
3023 WITNESS_INDEX_ASSERT(key.to);
3024 if ((w_rmatrix[parent->w_index][child->w_index]
3025 & WITNESS_LOCK_ORDER_KNOWN) == 0)
3028 hash = witness_hash_djb2((const char*)&key,
3029 sizeof(key)) % w_lohash.wloh_size;
3030 data = w_lohash.wloh_array[hash];
3031 while (data != NULL) {
3032 if (witness_lock_order_key_equal(&data->wlod_key, &key))
3034 data = data->wlod_next;
3042 * Verify that parent and child have a known relationship, are not the same,
3043 * and child is actually a child of parent. This is done without w_mtx
3044 * to avoid contention in the common case.
3047 witness_lock_order_check(struct witness *parent, struct witness *child)
3050 if (parent != child &&
3051 w_rmatrix[parent->w_index][child->w_index]
3052 & WITNESS_LOCK_ORDER_KNOWN &&
3053 isitmychild(parent, child))
3060 witness_lock_order_add(struct witness *parent, struct witness *child)
3062 struct witness_lock_order_data *data = NULL;
3063 struct witness_lock_order_key key;
3066 MPASS(parent != NULL && child != NULL);
3067 key.from = parent->w_index;
3068 key.to = child->w_index;
3069 WITNESS_INDEX_ASSERT(key.from);
3070 WITNESS_INDEX_ASSERT(key.to);
3071 if (w_rmatrix[parent->w_index][child->w_index]
3072 & WITNESS_LOCK_ORDER_KNOWN)
3075 hash = witness_hash_djb2((const char*)&key,
3076 sizeof(key)) % w_lohash.wloh_size;
3077 w_rmatrix[parent->w_index][child->w_index] |= WITNESS_LOCK_ORDER_KNOWN;
3081 w_lofree = data->wlod_next;
3082 data->wlod_next = w_lohash.wloh_array[hash];
3083 data->wlod_key = key;
3084 w_lohash.wloh_array[hash] = data;
3085 w_lohash.wloh_count++;
3086 stack_zero(&data->wlod_stack);
3087 stack_save(&data->wlod_stack);
3091 /* Call this whenever the structure of the witness graph changes. */
3093 witness_increment_graph_generation(void)
3096 if (witness_cold == 0)
3097 mtx_assert(&w_mtx, MA_OWNED);
3102 witness_output_drain(void *arg __unused, const char *data, int len)
3105 witness_output("%.*s", len, data);
3110 witness_debugger(int cond, const char *msg)
3119 if (witness_trace) {
3120 sbuf_new(&sb, buf, sizeof(buf), SBUF_FIXEDLEN);
3121 sbuf_set_drain(&sb, witness_output_drain, NULL);
3125 witness_output("stack backtrace:\n");
3126 stack_sbuf_print_ddb(&sb, &st);
3131 witness_enter_debugger(msg);
3135 witness_enter_debugger(const char *msg)
3139 kdb_enter(KDB_WHY_WITNESS, msg);