2 * SPDX-License-Identifier: BSD-3-Clause
4 * Copyright (c) 2008 Isilon Systems, Inc.
5 * Copyright (c) 2008 Ilya Maykov <ivmaykov@gmail.com>
6 * Copyright (c) 1998 Berkeley Software Design, Inc.
9 * Redistribution and use in source and binary forms, with or without
10 * modification, are permitted provided that the following conditions
12 * 1. Redistributions of source code must retain the above copyright
13 * notice, this list of conditions and the following disclaimer.
14 * 2. Redistributions in binary form must reproduce the above copyright
15 * notice, this list of conditions and the following disclaimer in the
16 * documentation and/or other materials provided with the distribution.
17 * 3. Berkeley Software Design Inc's name may not be used to endorse or
18 * promote products derived from this software without specific prior
21 * THIS SOFTWARE IS PROVIDED BY BERKELEY SOFTWARE DESIGN INC ``AS IS'' AND
22 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24 * ARE DISCLAIMED. IN NO EVENT SHALL BERKELEY SOFTWARE DESIGN INC BE LIABLE
25 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
33 * from BSDI $Id: mutex_witness.c,v 1.1.2.20 2000/04/27 03:10:27 cp Exp $
34 * and BSDI $Id: synch_machdep.c,v 2.3.2.39 2000/04/27 03:10:25 cp Exp $
38 * Implementation of the `witness' lock verifier. Originally implemented for
39 * mutexes in BSD/OS. Extended to handle generic lock objects and lock
45 * Pronunciation: 'wit-n&s
47 * Etymology: Middle English witnesse, from Old English witnes knowledge,
48 * testimony, witness, from 2wit
49 * Date: before 12th century
50 * 1 : attestation of a fact or event : TESTIMONY
51 * 2 : one that gives evidence; specifically : one who testifies in
52 * a cause or before a judicial tribunal
53 * 3 : one asked to be present at a transaction so as to be able to
54 * testify to its having taken place
55 * 4 : one who has personal knowledge of something
56 * 5 a : something serving as evidence or proof : SIGN
57 * b : public affirmation by word or example of usually
58 * religious faith or conviction <the heroic witness to divine
60 * 6 capitalized : a member of the Jehovah's Witnesses
64 * Special rules concerning Giant and lock orders:
66 * 1) Giant must be acquired before any other mutexes. Stated another way,
67 * no other mutex may be held when Giant is acquired.
69 * 2) Giant must be released when blocking on a sleepable lock.
71 * This rule is less obvious, but is a result of Giant providing the same
72 * semantics as spl(). Basically, when a thread sleeps, it must release
73 * Giant. When a thread blocks on a sleepable lock, it sleeps. Hence rule
76 * 3) Giant may be acquired before or after sleepable locks.
78 * This rule is also not quite as obvious. Giant may be acquired after
79 * a sleepable lock because it is a non-sleepable lock and non-sleepable
80 * locks may always be acquired while holding a sleepable lock. The second
81 * case, Giant before a sleepable lock, follows from rule 2) above. Suppose
82 * you have two threads T1 and T2 and a sleepable lock X. Suppose that T1
83 * acquires X and blocks on Giant. Then suppose that T2 acquires Giant and
84 * blocks on X. When T2 blocks on X, T2 will release Giant allowing T1 to
85 * execute. Thus, acquiring Giant both before and after a sleepable lock
86 * will not result in a lock order reversal.
89 #include <sys/cdefs.h>
90 __FBSDID("$FreeBSD$");
93 #include "opt_hwpmc_hooks.h"
94 #include "opt_stack.h"
95 #include "opt_witness.h"
97 #include <sys/param.h>
100 #include <sys/kernel.h>
102 #include <sys/lock.h>
103 #include <sys/malloc.h>
104 #include <sys/mutex.h>
105 #include <sys/priv.h>
106 #include <sys/proc.h>
107 #include <sys/sbuf.h>
108 #include <sys/sched.h>
109 #include <sys/stack.h>
110 #include <sys/sysctl.h>
111 #include <sys/syslog.h>
112 #include <sys/systm.h>
118 #include <machine/stdarg.h>
120 #if !defined(DDB) && !defined(STACK)
121 #error "DDB or STACK options are required for WITNESS"
124 /* Note that these traces do not work with KTR_ALQ. */
126 #define KTR_WITNESS KTR_SUBSYS
128 #define KTR_WITNESS 0
131 #define LI_RECURSEMASK 0x0000ffff /* Recursion depth of lock instance. */
132 #define LI_EXCLUSIVE 0x00010000 /* Exclusive lock instance. */
133 #define LI_NORELEASE 0x00020000 /* Lock not allowed to be released. */
135 /* Define this to check for blessed mutexes */
138 #ifndef WITNESS_COUNT
139 #define WITNESS_COUNT 1536
141 #define WITNESS_HASH_SIZE 251 /* Prime, gives load factor < 2 */
142 #define WITNESS_PENDLIST (512 + (MAXCPU * 4))
144 /* Allocate 256 KB of stack data space */
145 #define WITNESS_LO_DATA_COUNT 2048
147 /* Prime, gives load factor of ~2 at full load */
148 #define WITNESS_LO_HASH_SIZE 1021
151 * XXX: This is somewhat bogus, as we assume here that at most 2048 threads
152 * will hold LOCK_NCHILDREN locks. We handle failure ok, and we should
153 * probably be safe for the most part, but it's still a SWAG.
155 #define LOCK_NCHILDREN 5
156 #define LOCK_CHILDCOUNT 2048
158 #define MAX_W_NAME 64
160 #define FULLGRAPH_SBUF_SIZE 512
163 * These flags go in the witness relationship matrix and describe the
164 * relationship between any two struct witness objects.
166 #define WITNESS_UNRELATED 0x00 /* No lock order relation. */
167 #define WITNESS_PARENT 0x01 /* Parent, aka direct ancestor. */
168 #define WITNESS_ANCESTOR 0x02 /* Direct or indirect ancestor. */
169 #define WITNESS_CHILD 0x04 /* Child, aka direct descendant. */
170 #define WITNESS_DESCENDANT 0x08 /* Direct or indirect descendant. */
171 #define WITNESS_ANCESTOR_MASK (WITNESS_PARENT | WITNESS_ANCESTOR)
172 #define WITNESS_DESCENDANT_MASK (WITNESS_CHILD | WITNESS_DESCENDANT)
173 #define WITNESS_RELATED_MASK \
174 (WITNESS_ANCESTOR_MASK | WITNESS_DESCENDANT_MASK)
175 #define WITNESS_REVERSAL 0x10 /* A lock order reversal has been
177 #define WITNESS_RESERVED1 0x20 /* Unused flag, reserved. */
178 #define WITNESS_RESERVED2 0x40 /* Unused flag, reserved. */
179 #define WITNESS_LOCK_ORDER_KNOWN 0x80 /* This lock order is known. */
181 /* Descendant to ancestor flags */
182 #define WITNESS_DTOA(x) (((x) & WITNESS_RELATED_MASK) >> 2)
184 /* Ancestor to descendant flags */
185 #define WITNESS_ATOD(x) (((x) & WITNESS_RELATED_MASK) << 2)
187 #define WITNESS_INDEX_ASSERT(i) \
188 MPASS((i) > 0 && (i) <= w_max_used_index && (i) < witness_count)
190 static MALLOC_DEFINE(M_WITNESS, "Witness", "Witness");
193 * Lock instances. A lock instance is the data associated with a lock while
194 * it is held by witness. For example, a lock instance will hold the
195 * recursion count of a lock. Lock instances are held in lists. Spin locks
196 * are held in a per-cpu list while sleep locks are held in per-thread list.
198 struct lock_instance {
199 struct lock_object *li_lock;
206 * A simple list type used to build the list of locks held by a thread
207 * or CPU. We can't simply embed the list in struct lock_object since a
208 * lock may be held by more than one thread if it is a shared lock. Locks
209 * are added to the head of the list, so we fill up each list entry from
210 * "the back" logically. To ease some of the arithmetic, we actually fill
211 * in each list entry the normal way (children[0] then children[1], etc.) but
212 * when we traverse the list we read children[count-1] as the first entry
213 * down to children[0] as the final entry.
215 struct lock_list_entry {
216 struct lock_list_entry *ll_next;
217 struct lock_instance ll_children[LOCK_NCHILDREN];
222 * The main witness structure. One of these per named lock type in the system
223 * (for example, "vnode interlock").
226 char w_name[MAX_W_NAME];
227 uint32_t w_index; /* Index in the relationship matrix */
228 struct lock_class *w_class;
229 STAILQ_ENTRY(witness) w_list; /* List of all witnesses. */
230 STAILQ_ENTRY(witness) w_typelist; /* Witnesses of a type. */
231 struct witness *w_hash_next; /* Linked list in hash buckets. */
232 const char *w_file; /* File where last acquired */
233 uint32_t w_line; /* Line where last acquired */
235 uint16_t w_num_ancestors; /* direct/indirect
237 uint16_t w_num_descendants; /* direct/indirect
238 * descendant count */
240 unsigned w_displayed:1;
241 unsigned w_reversed:1;
244 STAILQ_HEAD(witness_list, witness);
247 * The witness hash table. Keys are witness names (const char *), elements are
248 * witness objects (struct witness *).
250 struct witness_hash {
251 struct witness *wh_array[WITNESS_HASH_SIZE];
257 * Key type for the lock order data hash table.
259 struct witness_lock_order_key {
264 struct witness_lock_order_data {
265 struct stack wlod_stack;
266 struct witness_lock_order_key wlod_key;
267 struct witness_lock_order_data *wlod_next;
271 * The witness lock order data hash table. Keys are witness index tuples
272 * (struct witness_lock_order_key), elements are lock order data objects
273 * (struct witness_lock_order_data).
275 struct witness_lock_order_hash {
276 struct witness_lock_order_data *wloh_array[WITNESS_LO_HASH_SIZE];
282 struct witness_blessed {
288 struct witness_pendhelp {
290 struct lock_object *wh_lock;
293 struct witness_order_list_entry {
295 struct lock_class *w_class;
299 * Returns 0 if one of the locks is a spin lock and the other is not.
300 * Returns 1 otherwise.
303 witness_lock_type_equal(struct witness *w1, struct witness *w2)
306 return ((w1->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) ==
307 (w2->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)));
311 witness_lock_order_key_equal(const struct witness_lock_order_key *a,
312 const struct witness_lock_order_key *b)
315 return (a->from == b->from && a->to == b->to);
318 static int _isitmyx(struct witness *w1, struct witness *w2, int rmask,
320 static void adopt(struct witness *parent, struct witness *child);
322 static int blessed(struct witness *, struct witness *);
324 static void depart(struct witness *w);
325 static struct witness *enroll(const char *description,
326 struct lock_class *lock_class);
327 static struct lock_instance *find_instance(struct lock_list_entry *list,
328 const struct lock_object *lock);
329 static int isitmychild(struct witness *parent, struct witness *child);
330 static int isitmydescendant(struct witness *parent, struct witness *child);
331 static void itismychild(struct witness *parent, struct witness *child);
332 static int sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS);
333 static int sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS);
334 static int sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS);
335 static int sysctl_debug_witness_channel(SYSCTL_HANDLER_ARGS);
336 static void witness_add_fullgraph(struct sbuf *sb, struct witness *parent);
338 static void witness_ddb_compute_levels(void);
339 static void witness_ddb_display(int(*)(const char *fmt, ...));
340 static void witness_ddb_display_descendants(int(*)(const char *fmt, ...),
341 struct witness *, int indent);
342 static void witness_ddb_display_list(int(*prnt)(const char *fmt, ...),
343 struct witness_list *list);
344 static void witness_ddb_level_descendants(struct witness *parent, int l);
345 static void witness_ddb_list(struct thread *td);
347 static void witness_debugger(int cond, const char *msg);
348 static void witness_free(struct witness *m);
349 static struct witness *witness_get(void);
350 static uint32_t witness_hash_djb2(const uint8_t *key, uint32_t size);
351 static struct witness *witness_hash_get(const char *key);
352 static void witness_hash_put(struct witness *w);
353 static void witness_init_hash_tables(void);
354 static void witness_increment_graph_generation(void);
355 static void witness_lock_list_free(struct lock_list_entry *lle);
356 static struct lock_list_entry *witness_lock_list_get(void);
357 static int witness_lock_order_add(struct witness *parent,
358 struct witness *child);
359 static int witness_lock_order_check(struct witness *parent,
360 struct witness *child);
361 static struct witness_lock_order_data *witness_lock_order_get(
362 struct witness *parent,
363 struct witness *child);
364 static void witness_list_lock(struct lock_instance *instance,
365 int (*prnt)(const char *fmt, ...));
366 static int witness_output(const char *fmt, ...) __printflike(1, 2);
367 static int witness_voutput(const char *fmt, va_list ap) __printflike(1, 0);
368 static void witness_setflag(struct lock_object *lock, int flag, int set);
370 static SYSCTL_NODE(_debug, OID_AUTO, witness, CTLFLAG_RW, NULL,
374 * If set to 0, lock order checking is disabled. If set to -1,
375 * witness is completely disabled. Otherwise witness performs full
376 * lock order checking for all locks. At runtime, lock order checking
377 * may be toggled. However, witness cannot be reenabled once it is
378 * completely disabled.
380 static int witness_watch = 1;
381 SYSCTL_PROC(_debug_witness, OID_AUTO, watch, CTLFLAG_RWTUN | CTLTYPE_INT, NULL, 0,
382 sysctl_debug_witness_watch, "I", "witness is watching lock operations");
386 * When KDB is enabled and witness_kdb is 1, it will cause the system
387 * to drop into kdebug() when:
388 * - a lock hierarchy violation occurs
389 * - locks are held when going to sleep.
396 SYSCTL_INT(_debug_witness, OID_AUTO, kdb, CTLFLAG_RWTUN, &witness_kdb, 0, "");
399 #if defined(DDB) || defined(KDB)
401 * When DDB or KDB is enabled and witness_trace is 1, it will cause the system
402 * to print a stack trace:
403 * - a lock hierarchy violation occurs
404 * - locks are held when going to sleep.
406 int witness_trace = 1;
407 SYSCTL_INT(_debug_witness, OID_AUTO, trace, CTLFLAG_RWTUN, &witness_trace, 0, "");
408 #endif /* DDB || KDB */
410 #ifdef WITNESS_SKIPSPIN
411 int witness_skipspin = 1;
413 int witness_skipspin = 0;
415 SYSCTL_INT(_debug_witness, OID_AUTO, skipspin, CTLFLAG_RDTUN, &witness_skipspin, 0, "");
417 int badstack_sbuf_size;
419 int witness_count = WITNESS_COUNT;
420 SYSCTL_INT(_debug_witness, OID_AUTO, witness_count, CTLFLAG_RDTUN,
421 &witness_count, 0, "");
424 * Output channel for witness messages. By default we print to the console.
426 enum witness_channel {
432 static enum witness_channel witness_channel = WITNESS_CONSOLE;
433 SYSCTL_PROC(_debug_witness, OID_AUTO, output_channel, CTLTYPE_STRING |
434 CTLFLAG_RWTUN, NULL, 0, sysctl_debug_witness_channel, "A",
435 "Output channel for warnings");
438 * Call this to print out the relations between locks.
440 SYSCTL_PROC(_debug_witness, OID_AUTO, fullgraph, CTLTYPE_STRING | CTLFLAG_RD,
441 NULL, 0, sysctl_debug_witness_fullgraph, "A", "Show locks relation graphs");
444 * Call this to print out the witness faulty stacks.
446 SYSCTL_PROC(_debug_witness, OID_AUTO, badstacks, CTLTYPE_STRING | CTLFLAG_RD,
447 NULL, 0, sysctl_debug_witness_badstacks, "A", "Show bad witness stacks");
449 static struct mtx w_mtx;
452 static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free);
453 static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all);
456 static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin);
457 static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep);
460 static struct lock_list_entry *w_lock_list_free = NULL;
461 static struct witness_pendhelp pending_locks[WITNESS_PENDLIST];
462 static u_int pending_cnt;
464 static int w_free_cnt, w_spin_cnt, w_sleep_cnt;
465 SYSCTL_INT(_debug_witness, OID_AUTO, free_cnt, CTLFLAG_RD, &w_free_cnt, 0, "");
466 SYSCTL_INT(_debug_witness, OID_AUTO, spin_cnt, CTLFLAG_RD, &w_spin_cnt, 0, "");
467 SYSCTL_INT(_debug_witness, OID_AUTO, sleep_cnt, CTLFLAG_RD, &w_sleep_cnt, 0,
470 static struct witness *w_data;
471 static uint8_t **w_rmatrix;
472 static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT];
473 static struct witness_hash w_hash; /* The witness hash table. */
475 /* The lock order data hash */
476 static struct witness_lock_order_data w_lodata[WITNESS_LO_DATA_COUNT];
477 static struct witness_lock_order_data *w_lofree = NULL;
478 static struct witness_lock_order_hash w_lohash;
479 static int w_max_used_index = 0;
480 static unsigned int w_generation = 0;
481 static const char w_notrunning[] = "Witness not running\n";
482 static const char w_stillcold[] = "Witness is still cold\n";
484 static const char w_notallowed[] = "The sysctl is disabled on the arch\n";
487 static struct witness_order_list_entry order_lists[] = {
491 { "proctree", &lock_class_sx },
492 { "allproc", &lock_class_sx },
493 { "allprison", &lock_class_sx },
498 { "Giant", &lock_class_mtx_sleep },
499 { "pipe mutex", &lock_class_mtx_sleep },
500 { "sigio lock", &lock_class_mtx_sleep },
501 { "process group", &lock_class_mtx_sleep },
502 { "process lock", &lock_class_mtx_sleep },
503 { "session", &lock_class_mtx_sleep },
504 { "uidinfo hash", &lock_class_rw },
506 { "pmc-sleep", &lock_class_mtx_sleep },
508 { "time lock", &lock_class_mtx_sleep },
513 { "umtx lock", &lock_class_mtx_sleep },
518 { "accept", &lock_class_mtx_sleep },
519 { "so_snd", &lock_class_mtx_sleep },
520 { "so_rcv", &lock_class_mtx_sleep },
521 { "sellck", &lock_class_mtx_sleep },
526 { "so_rcv", &lock_class_mtx_sleep },
527 { "radix node head", &lock_class_rw },
528 { "rtentry", &lock_class_mtx_sleep },
529 { "ifaddr", &lock_class_mtx_sleep },
533 * protocol locks before interface locks, after UDP locks.
535 { "in_multi_sx", &lock_class_sx },
536 { "udpinp", &lock_class_rw },
537 { "in_multi_list_mtx", &lock_class_mtx_sleep },
538 { "igmp_mtx", &lock_class_mtx_sleep },
539 { "ifnet_rw", &lock_class_rw },
540 { "if_addr_lock", &lock_class_rw },
544 * protocol locks before interface locks, after UDP locks.
546 { "in6_multi_sx", &lock_class_sx },
547 { "udpinp", &lock_class_rw },
548 { "in6_multi_list_mtx", &lock_class_mtx_sleep },
549 { "mld_mtx", &lock_class_mtx_sleep },
550 { "ifnet_rw", &lock_class_rw },
551 { "if_addr_lock", &lock_class_rw },
554 * UNIX Domain Sockets
556 { "unp_link_rwlock", &lock_class_rw },
557 { "unp_list_lock", &lock_class_mtx_sleep },
558 { "unp", &lock_class_mtx_sleep },
559 { "so_snd", &lock_class_mtx_sleep },
564 { "udp", &lock_class_rw },
565 { "udpinp", &lock_class_rw },
566 { "so_snd", &lock_class_mtx_sleep },
571 { "tcp", &lock_class_rw },
572 { "tcpinp", &lock_class_rw },
573 { "so_snd", &lock_class_mtx_sleep },
578 { "bpf global lock", &lock_class_sx },
579 { "bpf interface lock", &lock_class_rw },
580 { "bpf cdev lock", &lock_class_mtx_sleep },
585 { "nfsd_mtx", &lock_class_mtx_sleep },
586 { "so_snd", &lock_class_mtx_sleep },
592 { "802.11 com lock", &lock_class_mtx_sleep},
597 { "network driver", &lock_class_mtx_sleep},
603 { "ng_node", &lock_class_mtx_sleep },
604 { "ng_worklist", &lock_class_mtx_sleep },
609 { "vm map (system)", &lock_class_mtx_sleep },
610 { "vnode interlock", &lock_class_mtx_sleep },
611 { "cdev", &lock_class_mtx_sleep },
616 { "vm map (user)", &lock_class_sx },
617 { "vm object", &lock_class_rw },
618 { "vm page", &lock_class_mtx_sleep },
619 { "pmap pv global", &lock_class_rw },
620 { "pmap", &lock_class_mtx_sleep },
621 { "pmap pv list", &lock_class_rw },
622 { "vm page free queue", &lock_class_mtx_sleep },
623 { "vm pagequeue", &lock_class_mtx_sleep },
626 * kqueue/VFS interaction
628 { "kqueue", &lock_class_mtx_sleep },
629 { "struct mount mtx", &lock_class_mtx_sleep },
630 { "vnode interlock", &lock_class_mtx_sleep },
635 { "ncvn", &lock_class_mtx_sleep },
636 { "ncbuc", &lock_class_rw },
637 { "vnode interlock", &lock_class_mtx_sleep },
638 { "ncneg", &lock_class_mtx_sleep },
643 { "dn->dn_mtx", &lock_class_sx },
644 { "dr->dt.di.dr_mtx", &lock_class_sx },
645 { "db->db_mtx", &lock_class_sx },
650 { "TCP ID tree", &lock_class_rw },
651 { "tcp log id bucket", &lock_class_mtx_sleep },
652 { "tcpinp", &lock_class_rw },
653 { "TCP log expireq", &lock_class_mtx_sleep },
659 { "ap boot", &lock_class_mtx_spin },
661 { "rm.mutex_mtx", &lock_class_mtx_spin },
662 { "sio", &lock_class_mtx_spin },
664 { "cy", &lock_class_mtx_spin },
667 { "pcib_mtx", &lock_class_mtx_spin },
668 { "rtc_mtx", &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 },
675 { "pmc-per-proc", &lock_class_mtx_spin },
677 { "process slock", &lock_class_mtx_spin },
678 { "syscons video lock", &lock_class_mtx_spin },
679 { "sleepq chain", &lock_class_mtx_spin },
680 { "rm_spinlock", &lock_class_mtx_spin },
681 { "turnstile chain", &lock_class_mtx_spin },
682 { "turnstile lock", &lock_class_mtx_spin },
683 { "sched lock", &lock_class_mtx_spin },
684 { "td_contested", &lock_class_mtx_spin },
685 { "callout", &lock_class_mtx_spin },
686 { "entropy harvest mutex", &lock_class_mtx_spin },
688 { "smp rendezvous", &lock_class_mtx_spin },
691 { "tlb0", &lock_class_mtx_spin },
696 { "intrcnt", &lock_class_mtx_spin },
697 { "icu", &lock_class_mtx_spin },
698 #if defined(SMP) && defined(__sparc64__)
699 { "ipi", &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 },
729 * Pairs of locks which have been blessed
730 * Don't complain about order problems with blessed locks
732 static struct witness_blessed blessed_list[] = {
737 * This global is set to 0 once it becomes safe to use the witness code.
739 static int witness_cold = 1;
742 * This global is set to 1 once the static lock orders have been enrolled
743 * so that a warning can be issued for any spin locks enrolled later.
745 static int witness_spin_warn = 0;
747 /* Trim useless garbage from filenames. */
749 fixup_filename(const char *file)
754 while (strncmp(file, "../", 3) == 0)
760 * Calculate the size of early witness structures.
763 witness_startup_count(void)
767 sz = sizeof(struct witness) * witness_count;
768 sz += sizeof(*w_rmatrix) * (witness_count + 1);
769 sz += sizeof(*w_rmatrix[0]) * (witness_count + 1) *
776 * The WITNESS-enabled diagnostic code. Note that the witness code does
777 * assume that the early boot is single-threaded at least until after this
778 * routine is completed.
781 witness_startup(void *mem)
783 struct lock_object *lock;
784 struct witness_order_list_entry *order;
785 struct witness *w, *w1;
791 p += sizeof(struct witness) * witness_count;
793 w_rmatrix = (void *)p;
794 p += sizeof(*w_rmatrix) * (witness_count + 1);
796 for (i = 0; i < witness_count + 1; i++) {
797 w_rmatrix[i] = (void *)p;
798 p += sizeof(*w_rmatrix[i]) * (witness_count + 1);
800 badstack_sbuf_size = witness_count * 256;
803 * We have to release Giant before initializing its witness
804 * structure so that WITNESS doesn't get confused.
807 mtx_assert(&Giant, MA_NOTOWNED);
809 CTR1(KTR_WITNESS, "%s: initializing witness", __func__);
810 mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET |
811 MTX_NOWITNESS | MTX_NOPROFILE);
812 for (i = witness_count - 1; i >= 0; i--) {
814 memset(w, 0, sizeof(*w));
815 w_data[i].w_index = i; /* Witness index never changes. */
818 KASSERT(STAILQ_FIRST(&w_free)->w_index == 0,
819 ("%s: Invalid list of free witness objects", __func__));
821 /* Witness with index 0 is not used to aid in debugging. */
822 STAILQ_REMOVE_HEAD(&w_free, w_list);
825 for (i = 0; i < witness_count; i++) {
826 memset(w_rmatrix[i], 0, sizeof(*w_rmatrix[i]) *
827 (witness_count + 1));
830 for (i = 0; i < LOCK_CHILDCOUNT; i++)
831 witness_lock_list_free(&w_locklistdata[i]);
832 witness_init_hash_tables();
834 /* First add in all the specified order lists. */
835 for (order = order_lists; order->w_name != NULL; order++) {
836 w = enroll(order->w_name, order->w_class);
839 w->w_file = "order list";
840 for (order++; order->w_name != NULL; order++) {
841 w1 = enroll(order->w_name, order->w_class);
844 w1->w_file = "order list";
849 witness_spin_warn = 1;
851 /* Iterate through all locks and add them to witness. */
852 for (i = 0; pending_locks[i].wh_lock != NULL; i++) {
853 lock = pending_locks[i].wh_lock;
854 KASSERT(lock->lo_flags & LO_WITNESS,
855 ("%s: lock %s is on pending list but not LO_WITNESS",
856 __func__, lock->lo_name));
857 lock->lo_witness = enroll(pending_locks[i].wh_type,
861 /* Mark the witness code as being ready for use. */
868 witness_init(struct lock_object *lock, const char *type)
870 struct lock_class *class;
872 /* Various sanity checks. */
873 class = LOCK_CLASS(lock);
874 if ((lock->lo_flags & LO_RECURSABLE) != 0 &&
875 (class->lc_flags & LC_RECURSABLE) == 0)
876 kassert_panic("%s: lock (%s) %s can not be recursable",
877 __func__, class->lc_name, lock->lo_name);
878 if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
879 (class->lc_flags & LC_SLEEPABLE) == 0)
880 kassert_panic("%s: lock (%s) %s can not be sleepable",
881 __func__, class->lc_name, lock->lo_name);
882 if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
883 (class->lc_flags & LC_UPGRADABLE) == 0)
884 kassert_panic("%s: lock (%s) %s can not be upgradable",
885 __func__, class->lc_name, lock->lo_name);
888 * If we shouldn't watch this lock, then just clear lo_witness.
889 * Otherwise, if witness_cold is set, then it is too early to
890 * enroll this lock, so defer it to witness_initialize() by adding
891 * it to the pending_locks list. If it is not too early, then enroll
894 if (witness_watch < 1 || panicstr != NULL ||
895 (lock->lo_flags & LO_WITNESS) == 0)
896 lock->lo_witness = NULL;
897 else if (witness_cold) {
898 pending_locks[pending_cnt].wh_lock = lock;
899 pending_locks[pending_cnt++].wh_type = type;
900 if (pending_cnt > WITNESS_PENDLIST)
901 panic("%s: pending locks list is too small, "
902 "increase WITNESS_PENDLIST\n",
905 lock->lo_witness = enroll(type, class);
909 witness_destroy(struct lock_object *lock)
911 struct lock_class *class;
914 class = LOCK_CLASS(lock);
917 panic("lock (%s) %s destroyed while witness_cold",
918 class->lc_name, lock->lo_name);
920 /* XXX: need to verify that no one holds the lock */
921 if ((lock->lo_flags & LO_WITNESS) == 0 || lock->lo_witness == NULL)
923 w = lock->lo_witness;
925 mtx_lock_spin(&w_mtx);
926 MPASS(w->w_refcount > 0);
929 if (w->w_refcount == 0)
931 mtx_unlock_spin(&w_mtx);
936 witness_ddb_compute_levels(void)
941 * First clear all levels.
943 STAILQ_FOREACH(w, &w_all, w_list)
947 * Look for locks with no parents and level all their descendants.
949 STAILQ_FOREACH(w, &w_all, w_list) {
951 /* If the witness has ancestors (is not a root), skip it. */
952 if (w->w_num_ancestors > 0)
954 witness_ddb_level_descendants(w, 0);
959 witness_ddb_level_descendants(struct witness *w, int l)
963 if (w->w_ddb_level >= l)
969 for (i = 1; i <= w_max_used_index; i++) {
970 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
971 witness_ddb_level_descendants(&w_data[i], l);
976 witness_ddb_display_descendants(int(*prnt)(const char *fmt, ...),
977 struct witness *w, int indent)
981 for (i = 0; i < indent; i++)
983 prnt("%s (type: %s, depth: %d, active refs: %d)",
984 w->w_name, w->w_class->lc_name,
985 w->w_ddb_level, w->w_refcount);
986 if (w->w_displayed) {
987 prnt(" -- (already displayed)\n");
991 if (w->w_file != NULL && w->w_line != 0)
992 prnt(" -- last acquired @ %s:%d\n", fixup_filename(w->w_file),
995 prnt(" -- never acquired\n");
997 WITNESS_INDEX_ASSERT(w->w_index);
998 for (i = 1; i <= w_max_used_index; i++) {
1001 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
1002 witness_ddb_display_descendants(prnt, &w_data[i],
1008 witness_ddb_display_list(int(*prnt)(const char *fmt, ...),
1009 struct witness_list *list)
1013 STAILQ_FOREACH(w, list, w_typelist) {
1014 if (w->w_file == NULL || w->w_ddb_level > 0)
1017 /* This lock has no anscestors - display its descendants. */
1018 witness_ddb_display_descendants(prnt, w, 0);
1025 witness_ddb_display(int(*prnt)(const char *fmt, ...))
1029 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1030 witness_ddb_compute_levels();
1032 /* Clear all the displayed flags. */
1033 STAILQ_FOREACH(w, &w_all, w_list)
1037 * First, handle sleep locks which have been acquired at least
1040 prnt("Sleep locks:\n");
1041 witness_ddb_display_list(prnt, &w_sleep);
1046 * Now do spin locks which have been acquired at least once.
1048 prnt("\nSpin locks:\n");
1049 witness_ddb_display_list(prnt, &w_spin);
1054 * Finally, any locks which have not been acquired yet.
1056 prnt("\nLocks which were never acquired:\n");
1057 STAILQ_FOREACH(w, &w_all, w_list) {
1058 if (w->w_file != NULL || w->w_refcount == 0)
1060 prnt("%s (type: %s, depth: %d)\n", w->w_name,
1061 w->w_class->lc_name, w->w_ddb_level);
1069 witness_defineorder(struct lock_object *lock1, struct lock_object *lock2)
1072 if (witness_watch == -1 || panicstr != NULL)
1075 /* Require locks that witness knows about. */
1076 if (lock1 == NULL || lock1->lo_witness == NULL || lock2 == NULL ||
1077 lock2->lo_witness == NULL)
1080 mtx_assert(&w_mtx, MA_NOTOWNED);
1081 mtx_lock_spin(&w_mtx);
1084 * If we already have either an explicit or implied lock order that
1085 * is the other way around, then return an error.
1087 if (witness_watch &&
1088 isitmydescendant(lock2->lo_witness, lock1->lo_witness)) {
1089 mtx_unlock_spin(&w_mtx);
1093 /* Try to add the new order. */
1094 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1095 lock2->lo_witness->w_name, lock1->lo_witness->w_name);
1096 itismychild(lock1->lo_witness, lock2->lo_witness);
1097 mtx_unlock_spin(&w_mtx);
1102 witness_checkorder(struct lock_object *lock, int flags, const char *file,
1103 int line, struct lock_object *interlock)
1105 struct lock_list_entry *lock_list, *lle;
1106 struct lock_instance *lock1, *lock2, *plock;
1107 struct lock_class *class, *iclass;
1108 struct witness *w, *w1;
1112 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL ||
1116 w = lock->lo_witness;
1117 class = LOCK_CLASS(lock);
1120 if (class->lc_flags & LC_SLEEPLOCK) {
1123 * Since spin locks include a critical section, this check
1124 * implicitly enforces a lock order of all sleep locks before
1127 if (td->td_critnest != 0 && !kdb_active)
1128 kassert_panic("acquiring blockable sleep lock with "
1129 "spinlock or critical section held (%s) %s @ %s:%d",
1130 class->lc_name, lock->lo_name,
1131 fixup_filename(file), line);
1134 * If this is the first lock acquired then just return as
1135 * no order checking is needed.
1137 lock_list = td->td_sleeplocks;
1138 if (lock_list == NULL || lock_list->ll_count == 0)
1143 * If this is the first lock, just return as no order
1144 * checking is needed. Avoid problems with thread
1145 * migration pinning the thread while checking if
1146 * spinlocks are held. If at least one spinlock is held
1147 * the thread is in a safe path and it is allowed to
1151 lock_list = PCPU_GET(spinlocks);
1152 if (lock_list == NULL || lock_list->ll_count == 0) {
1160 * Check to see if we are recursing on a lock we already own. If
1161 * so, make sure that we don't mismatch exclusive and shared lock
1164 lock1 = find_instance(lock_list, lock);
1165 if (lock1 != NULL) {
1166 if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
1167 (flags & LOP_EXCLUSIVE) == 0) {
1168 witness_output("shared lock of (%s) %s @ %s:%d\n",
1169 class->lc_name, lock->lo_name,
1170 fixup_filename(file), line);
1171 witness_output("while exclusively locked from %s:%d\n",
1172 fixup_filename(lock1->li_file), lock1->li_line);
1173 kassert_panic("excl->share");
1175 if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
1176 (flags & LOP_EXCLUSIVE) != 0) {
1177 witness_output("exclusive lock of (%s) %s @ %s:%d\n",
1178 class->lc_name, lock->lo_name,
1179 fixup_filename(file), line);
1180 witness_output("while share locked from %s:%d\n",
1181 fixup_filename(lock1->li_file), lock1->li_line);
1182 kassert_panic("share->excl");
1187 /* Warn if the interlock is not locked exactly once. */
1188 if (interlock != NULL) {
1189 iclass = LOCK_CLASS(interlock);
1190 lock1 = find_instance(lock_list, interlock);
1192 kassert_panic("interlock (%s) %s not locked @ %s:%d",
1193 iclass->lc_name, interlock->lo_name,
1194 fixup_filename(file), line);
1195 else if ((lock1->li_flags & LI_RECURSEMASK) != 0)
1196 kassert_panic("interlock (%s) %s recursed @ %s:%d",
1197 iclass->lc_name, interlock->lo_name,
1198 fixup_filename(file), line);
1202 * Find the previously acquired lock, but ignore interlocks.
1204 plock = &lock_list->ll_children[lock_list->ll_count - 1];
1205 if (interlock != NULL && plock->li_lock == interlock) {
1206 if (lock_list->ll_count > 1)
1208 &lock_list->ll_children[lock_list->ll_count - 2];
1210 lle = lock_list->ll_next;
1213 * The interlock is the only lock we hold, so
1218 plock = &lle->ll_children[lle->ll_count - 1];
1223 * Try to perform most checks without a lock. If this succeeds we
1224 * can skip acquiring the lock and return success. Otherwise we redo
1225 * the check with the lock held to handle races with concurrent updates.
1227 w1 = plock->li_lock->lo_witness;
1228 if (witness_lock_order_check(w1, w))
1231 mtx_lock_spin(&w_mtx);
1232 if (witness_lock_order_check(w1, w)) {
1233 mtx_unlock_spin(&w_mtx);
1236 witness_lock_order_add(w1, w);
1239 * Check for duplicate locks of the same type. Note that we only
1240 * have to check for this on the last lock we just acquired. Any
1241 * other cases will be caught as lock order violations.
1245 if (!(lock->lo_flags & LO_DUPOK) && !(flags & LOP_DUPOK) &&
1246 !(w_rmatrix[i][i] & WITNESS_REVERSAL)) {
1247 w_rmatrix[i][i] |= WITNESS_REVERSAL;
1249 mtx_unlock_spin(&w_mtx);
1251 "acquiring duplicate lock of same type: \"%s\"\n",
1253 witness_output(" 1st %s @ %s:%d\n", plock->li_lock->lo_name,
1254 fixup_filename(plock->li_file), plock->li_line);
1255 witness_output(" 2nd %s @ %s:%d\n", lock->lo_name,
1256 fixup_filename(file), line);
1257 witness_debugger(1, __func__);
1259 mtx_unlock_spin(&w_mtx);
1262 mtx_assert(&w_mtx, MA_OWNED);
1265 * If we know that the lock we are acquiring comes after
1266 * the lock we most recently acquired in the lock order tree,
1267 * then there is no need for any further checks.
1269 if (isitmychild(w1, w))
1272 for (j = 0, lle = lock_list; lle != NULL; lle = lle->ll_next) {
1273 for (i = lle->ll_count - 1; i >= 0; i--, j++) {
1275 MPASS(j < LOCK_CHILDCOUNT * LOCK_NCHILDREN);
1276 lock1 = &lle->ll_children[i];
1279 * Ignore the interlock.
1281 if (interlock == lock1->li_lock)
1285 * If this lock doesn't undergo witness checking,
1288 w1 = lock1->li_lock->lo_witness;
1290 KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
1291 ("lock missing witness structure"));
1296 * If we are locking Giant and this is a sleepable
1297 * lock, then skip it.
1299 if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0 &&
1300 lock == &Giant.lock_object)
1304 * If we are locking a sleepable lock and this lock
1305 * is Giant, then skip it.
1307 if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1308 lock1->li_lock == &Giant.lock_object)
1312 * If we are locking a sleepable lock and this lock
1313 * isn't sleepable, we want to treat it as a lock
1314 * order violation to enfore a general lock order of
1315 * sleepable locks before non-sleepable locks.
1317 if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1318 (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
1322 * If we are locking Giant and this is a non-sleepable
1323 * lock, then treat it as a reversal.
1325 if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0 &&
1326 lock == &Giant.lock_object)
1330 * Check the lock order hierarchy for a reveresal.
1332 if (!isitmydescendant(w, w1))
1337 * We have a lock order violation, check to see if it
1338 * is allowed or has already been yelled about.
1343 * If the lock order is blessed, just bail. We don't
1344 * look for other lock order violations though, which
1351 /* Bail if this violation is known */
1352 if (w_rmatrix[w1->w_index][w->w_index] & WITNESS_REVERSAL)
1355 /* Record this as a violation */
1356 w_rmatrix[w1->w_index][w->w_index] |= WITNESS_REVERSAL;
1357 w_rmatrix[w->w_index][w1->w_index] |= WITNESS_REVERSAL;
1358 w->w_reversed = w1->w_reversed = 1;
1359 witness_increment_graph_generation();
1360 mtx_unlock_spin(&w_mtx);
1362 #ifdef WITNESS_NO_VNODE
1364 * There are known LORs between VNODE locks. They are
1365 * not an indication of a bug. VNODE locks are flagged
1366 * as such (LO_IS_VNODE) and we don't yell if the LOR
1367 * is between 2 VNODE locks.
1369 if ((lock->lo_flags & LO_IS_VNODE) != 0 &&
1370 (lock1->li_lock->lo_flags & LO_IS_VNODE) != 0)
1375 * Ok, yell about it.
1377 if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1378 (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
1380 "lock order reversal: (sleepable after non-sleepable)\n");
1381 else if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0
1382 && lock == &Giant.lock_object)
1384 "lock order reversal: (Giant after non-sleepable)\n");
1386 witness_output("lock order reversal:\n");
1389 * Try to locate an earlier lock with
1390 * witness w in our list.
1393 lock2 = &lle->ll_children[i];
1394 MPASS(lock2->li_lock != NULL);
1395 if (lock2->li_lock->lo_witness == w)
1397 if (i == 0 && lle->ll_next != NULL) {
1399 i = lle->ll_count - 1;
1400 MPASS(i >= 0 && i < LOCK_NCHILDREN);
1405 witness_output(" 1st %p %s (%s) @ %s:%d\n",
1406 lock1->li_lock, lock1->li_lock->lo_name,
1407 w1->w_name, fixup_filename(lock1->li_file),
1409 witness_output(" 2nd %p %s (%s) @ %s:%d\n", lock,
1410 lock->lo_name, w->w_name,
1411 fixup_filename(file), line);
1413 witness_output(" 1st %p %s (%s) @ %s:%d\n",
1414 lock2->li_lock, lock2->li_lock->lo_name,
1415 lock2->li_lock->lo_witness->w_name,
1416 fixup_filename(lock2->li_file),
1418 witness_output(" 2nd %p %s (%s) @ %s:%d\n",
1419 lock1->li_lock, lock1->li_lock->lo_name,
1420 w1->w_name, fixup_filename(lock1->li_file),
1422 witness_output(" 3rd %p %s (%s) @ %s:%d\n", lock,
1423 lock->lo_name, w->w_name,
1424 fixup_filename(file), line);
1426 witness_debugger(1, __func__);
1432 * If requested, build a new lock order. However, don't build a new
1433 * relationship between a sleepable lock and Giant if it is in the
1434 * wrong direction. The correct lock order is that sleepable locks
1435 * always come before Giant.
1437 if (flags & LOP_NEWORDER &&
1438 !(plock->li_lock == &Giant.lock_object &&
1439 (lock->lo_flags & LO_SLEEPABLE) != 0)) {
1440 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1441 w->w_name, plock->li_lock->lo_witness->w_name);
1442 itismychild(plock->li_lock->lo_witness, w);
1445 mtx_unlock_spin(&w_mtx);
1449 witness_lock(struct lock_object *lock, int flags, const char *file, int line)
1451 struct lock_list_entry **lock_list, *lle;
1452 struct lock_instance *instance;
1456 if (witness_cold || witness_watch == -1 || lock->lo_witness == NULL ||
1459 w = lock->lo_witness;
1462 /* Determine lock list for this lock. */
1463 if (LOCK_CLASS(lock)->lc_flags & LC_SLEEPLOCK)
1464 lock_list = &td->td_sleeplocks;
1466 lock_list = PCPU_PTR(spinlocks);
1468 /* Check to see if we are recursing on a lock we already own. */
1469 instance = find_instance(*lock_list, lock);
1470 if (instance != NULL) {
1471 instance->li_flags++;
1472 CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
1473 td->td_proc->p_pid, lock->lo_name,
1474 instance->li_flags & LI_RECURSEMASK);
1475 instance->li_file = file;
1476 instance->li_line = line;
1480 /* Update per-witness last file and line acquire. */
1484 /* Find the next open lock instance in the list and fill it. */
1486 if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
1487 lle = witness_lock_list_get();
1490 lle->ll_next = *lock_list;
1491 CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
1492 td->td_proc->p_pid, lle);
1495 instance = &lle->ll_children[lle->ll_count++];
1496 instance->li_lock = lock;
1497 instance->li_line = line;
1498 instance->li_file = file;
1499 if ((flags & LOP_EXCLUSIVE) != 0)
1500 instance->li_flags = LI_EXCLUSIVE;
1502 instance->li_flags = 0;
1503 CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
1504 td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
1508 witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
1510 struct lock_instance *instance;
1511 struct lock_class *class;
1513 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1514 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
1516 class = LOCK_CLASS(lock);
1517 if (witness_watch) {
1518 if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1520 "upgrade of non-upgradable lock (%s) %s @ %s:%d",
1521 class->lc_name, lock->lo_name,
1522 fixup_filename(file), line);
1523 if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1525 "upgrade of non-sleep lock (%s) %s @ %s:%d",
1526 class->lc_name, lock->lo_name,
1527 fixup_filename(file), line);
1529 instance = find_instance(curthread->td_sleeplocks, lock);
1530 if (instance == NULL) {
1531 kassert_panic("upgrade of unlocked lock (%s) %s @ %s:%d",
1532 class->lc_name, lock->lo_name,
1533 fixup_filename(file), line);
1536 if (witness_watch) {
1537 if ((instance->li_flags & LI_EXCLUSIVE) != 0)
1539 "upgrade of exclusive lock (%s) %s @ %s:%d",
1540 class->lc_name, lock->lo_name,
1541 fixup_filename(file), line);
1542 if ((instance->li_flags & LI_RECURSEMASK) != 0)
1544 "upgrade of recursed lock (%s) %s r=%d @ %s:%d",
1545 class->lc_name, lock->lo_name,
1546 instance->li_flags & LI_RECURSEMASK,
1547 fixup_filename(file), line);
1549 instance->li_flags |= LI_EXCLUSIVE;
1553 witness_downgrade(struct lock_object *lock, int flags, const char *file,
1556 struct lock_instance *instance;
1557 struct lock_class *class;
1559 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1560 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
1562 class = LOCK_CLASS(lock);
1563 if (witness_watch) {
1564 if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1566 "downgrade of non-upgradable lock (%s) %s @ %s:%d",
1567 class->lc_name, lock->lo_name,
1568 fixup_filename(file), line);
1569 if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1571 "downgrade of non-sleep lock (%s) %s @ %s:%d",
1572 class->lc_name, lock->lo_name,
1573 fixup_filename(file), line);
1575 instance = find_instance(curthread->td_sleeplocks, lock);
1576 if (instance == NULL) {
1577 kassert_panic("downgrade of unlocked lock (%s) %s @ %s:%d",
1578 class->lc_name, lock->lo_name,
1579 fixup_filename(file), line);
1582 if (witness_watch) {
1583 if ((instance->li_flags & LI_EXCLUSIVE) == 0)
1585 "downgrade of shared lock (%s) %s @ %s:%d",
1586 class->lc_name, lock->lo_name,
1587 fixup_filename(file), line);
1588 if ((instance->li_flags & LI_RECURSEMASK) != 0)
1590 "downgrade of recursed lock (%s) %s r=%d @ %s:%d",
1591 class->lc_name, lock->lo_name,
1592 instance->li_flags & LI_RECURSEMASK,
1593 fixup_filename(file), line);
1595 instance->li_flags &= ~LI_EXCLUSIVE;
1599 witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
1601 struct lock_list_entry **lock_list, *lle;
1602 struct lock_instance *instance;
1603 struct lock_class *class;
1608 if (witness_cold || lock->lo_witness == NULL || panicstr != NULL)
1611 class = LOCK_CLASS(lock);
1613 /* Find lock instance associated with this lock. */
1614 if (class->lc_flags & LC_SLEEPLOCK)
1615 lock_list = &td->td_sleeplocks;
1617 lock_list = PCPU_PTR(spinlocks);
1619 for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
1620 for (i = 0; i < (*lock_list)->ll_count; i++) {
1621 instance = &(*lock_list)->ll_children[i];
1622 if (instance->li_lock == lock)
1627 * When disabling WITNESS through witness_watch we could end up in
1628 * having registered locks in the td_sleeplocks queue.
1629 * We have to make sure we flush these queues, so just search for
1630 * eventual register locks and remove them.
1632 if (witness_watch > 0) {
1633 kassert_panic("lock (%s) %s not locked @ %s:%d", class->lc_name,
1634 lock->lo_name, fixup_filename(file), line);
1641 /* First, check for shared/exclusive mismatches. */
1642 if ((instance->li_flags & LI_EXCLUSIVE) != 0 && witness_watch > 0 &&
1643 (flags & LOP_EXCLUSIVE) == 0) {
1644 witness_output("shared unlock of (%s) %s @ %s:%d\n",
1645 class->lc_name, lock->lo_name, fixup_filename(file), line);
1646 witness_output("while exclusively locked from %s:%d\n",
1647 fixup_filename(instance->li_file), instance->li_line);
1648 kassert_panic("excl->ushare");
1650 if ((instance->li_flags & LI_EXCLUSIVE) == 0 && witness_watch > 0 &&
1651 (flags & LOP_EXCLUSIVE) != 0) {
1652 witness_output("exclusive unlock of (%s) %s @ %s:%d\n",
1653 class->lc_name, lock->lo_name, fixup_filename(file), line);
1654 witness_output("while share locked from %s:%d\n",
1655 fixup_filename(instance->li_file),
1657 kassert_panic("share->uexcl");
1659 /* If we are recursed, unrecurse. */
1660 if ((instance->li_flags & LI_RECURSEMASK) > 0) {
1661 CTR4(KTR_WITNESS, "%s: pid %d unrecursed on %s r=%d", __func__,
1662 td->td_proc->p_pid, instance->li_lock->lo_name,
1663 instance->li_flags);
1664 instance->li_flags--;
1667 /* The lock is now being dropped, check for NORELEASE flag */
1668 if ((instance->li_flags & LI_NORELEASE) != 0 && witness_watch > 0) {
1669 witness_output("forbidden unlock of (%s) %s @ %s:%d\n",
1670 class->lc_name, lock->lo_name, fixup_filename(file), line);
1671 kassert_panic("lock marked norelease");
1674 /* Otherwise, remove this item from the list. */
1676 CTR4(KTR_WITNESS, "%s: pid %d removed %s from lle[%d]", __func__,
1677 td->td_proc->p_pid, instance->li_lock->lo_name,
1678 (*lock_list)->ll_count - 1);
1679 for (j = i; j < (*lock_list)->ll_count - 1; j++)
1680 (*lock_list)->ll_children[j] =
1681 (*lock_list)->ll_children[j + 1];
1682 (*lock_list)->ll_count--;
1686 * In order to reduce contention on w_mtx, we want to keep always an
1687 * head object into lists so that frequent allocation from the
1688 * free witness pool (and subsequent locking) is avoided.
1689 * In order to maintain the current code simple, when the head
1690 * object is totally unloaded it means also that we do not have
1691 * further objects in the list, so the list ownership needs to be
1692 * hand over to another object if the current head needs to be freed.
1694 if ((*lock_list)->ll_count == 0) {
1695 if (*lock_list == lle) {
1696 if (lle->ll_next == NULL)
1700 *lock_list = lle->ll_next;
1701 CTR3(KTR_WITNESS, "%s: pid %d removed lle %p", __func__,
1702 td->td_proc->p_pid, lle);
1703 witness_lock_list_free(lle);
1708 witness_thread_exit(struct thread *td)
1710 struct lock_list_entry *lle;
1713 lle = td->td_sleeplocks;
1714 if (lle == NULL || panicstr != NULL)
1716 if (lle->ll_count != 0) {
1717 for (n = 0; lle != NULL; lle = lle->ll_next)
1718 for (i = lle->ll_count - 1; i >= 0; i--) {
1721 "Thread %p exiting with the following locks held:\n", td);
1723 witness_list_lock(&lle->ll_children[i],
1728 "Thread %p cannot exit while holding sleeplocks\n", td);
1730 witness_lock_list_free(lle);
1734 * Warn if any locks other than 'lock' are held. Flags can be passed in to
1735 * exempt Giant and sleepable locks from the checks as well. If any
1736 * non-exempt locks are held, then a supplied message is printed to the
1737 * output channel along with a list of the offending locks. If indicated in the
1738 * flags then a failure results in a panic as well.
1741 witness_warn(int flags, struct lock_object *lock, const char *fmt, ...)
1743 struct lock_list_entry *lock_list, *lle;
1744 struct lock_instance *lock1;
1749 if (witness_cold || witness_watch < 1 || panicstr != NULL)
1753 for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next)
1754 for (i = lle->ll_count - 1; i >= 0; i--) {
1755 lock1 = &lle->ll_children[i];
1756 if (lock1->li_lock == lock)
1758 if (flags & WARN_GIANTOK &&
1759 lock1->li_lock == &Giant.lock_object)
1761 if (flags & WARN_SLEEPOK &&
1762 (lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0)
1768 printf(" with the following %slocks held:\n",
1769 (flags & WARN_SLEEPOK) != 0 ?
1770 "non-sleepable " : "");
1773 witness_list_lock(lock1, printf);
1777 * Pin the thread in order to avoid problems with thread migration.
1778 * Once that all verifies are passed about spinlocks ownership,
1779 * the thread is in a safe path and it can be unpinned.
1782 lock_list = PCPU_GET(spinlocks);
1783 if (lock_list != NULL && lock_list->ll_count != 0) {
1787 * We should only have one spinlock and as long as
1788 * the flags cannot match for this locks class,
1789 * check if the first spinlock is the one curthread
1792 lock1 = &lock_list->ll_children[lock_list->ll_count - 1];
1793 if (lock_list->ll_count == 1 && lock_list->ll_next == NULL &&
1794 lock1->li_lock == lock && n == 0)
1800 printf(" with the following %slocks held:\n",
1801 (flags & WARN_SLEEPOK) != 0 ? "non-sleepable " : "");
1802 n += witness_list_locks(&lock_list, printf);
1805 if (flags & WARN_PANIC && n)
1806 kassert_panic("%s", __func__);
1808 witness_debugger(n, __func__);
1813 witness_file(struct lock_object *lock)
1817 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1819 w = lock->lo_witness;
1824 witness_line(struct lock_object *lock)
1828 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1830 w = lock->lo_witness;
1834 static struct witness *
1835 enroll(const char *description, struct lock_class *lock_class)
1839 MPASS(description != NULL);
1841 if (witness_watch == -1 || panicstr != NULL)
1843 if ((lock_class->lc_flags & LC_SPINLOCK)) {
1844 if (witness_skipspin)
1846 } else if ((lock_class->lc_flags & LC_SLEEPLOCK) == 0) {
1847 kassert_panic("lock class %s is not sleep or spin",
1848 lock_class->lc_name);
1852 mtx_lock_spin(&w_mtx);
1853 w = witness_hash_get(description);
1856 if ((w = witness_get()) == NULL)
1858 MPASS(strlen(description) < MAX_W_NAME);
1859 strcpy(w->w_name, description);
1860 w->w_class = lock_class;
1862 STAILQ_INSERT_HEAD(&w_all, w, w_list);
1863 if (lock_class->lc_flags & LC_SPINLOCK) {
1864 STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
1866 } else if (lock_class->lc_flags & LC_SLEEPLOCK) {
1867 STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
1871 /* Insert new witness into the hash */
1872 witness_hash_put(w);
1873 witness_increment_graph_generation();
1874 mtx_unlock_spin(&w_mtx);
1878 if (w->w_refcount == 1)
1879 w->w_class = lock_class;
1880 mtx_unlock_spin(&w_mtx);
1881 if (lock_class != w->w_class)
1883 "lock (%s) %s does not match earlier (%s) lock",
1884 description, lock_class->lc_name,
1885 w->w_class->lc_name);
1890 depart(struct witness *w)
1893 MPASS(w->w_refcount == 0);
1894 if (w->w_class->lc_flags & LC_SLEEPLOCK) {
1900 * Set file to NULL as it may point into a loadable module.
1904 witness_increment_graph_generation();
1909 adopt(struct witness *parent, struct witness *child)
1913 if (witness_cold == 0)
1914 mtx_assert(&w_mtx, MA_OWNED);
1916 /* If the relationship is already known, there's no work to be done. */
1917 if (isitmychild(parent, child))
1920 /* When the structure of the graph changes, bump up the generation. */
1921 witness_increment_graph_generation();
1924 * The hard part ... create the direct relationship, then propagate all
1925 * indirect relationships.
1927 pi = parent->w_index;
1928 ci = child->w_index;
1929 WITNESS_INDEX_ASSERT(pi);
1930 WITNESS_INDEX_ASSERT(ci);
1932 w_rmatrix[pi][ci] |= WITNESS_PARENT;
1933 w_rmatrix[ci][pi] |= WITNESS_CHILD;
1936 * If parent was not already an ancestor of child,
1937 * then we increment the descendant and ancestor counters.
1939 if ((w_rmatrix[pi][ci] & WITNESS_ANCESTOR) == 0) {
1940 parent->w_num_descendants++;
1941 child->w_num_ancestors++;
1945 * Find each ancestor of 'pi'. Note that 'pi' itself is counted as
1946 * an ancestor of 'pi' during this loop.
1948 for (i = 1; i <= w_max_used_index; i++) {
1949 if ((w_rmatrix[i][pi] & WITNESS_ANCESTOR_MASK) == 0 &&
1953 /* Find each descendant of 'i' and mark it as a descendant. */
1954 for (j = 1; j <= w_max_used_index; j++) {
1957 * Skip children that are already marked as
1958 * descendants of 'i'.
1960 if (w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK)
1964 * We are only interested in descendants of 'ci'. Note
1965 * that 'ci' itself is counted as a descendant of 'ci'.
1967 if ((w_rmatrix[ci][j] & WITNESS_ANCESTOR_MASK) == 0 &&
1970 w_rmatrix[i][j] |= WITNESS_ANCESTOR;
1971 w_rmatrix[j][i] |= WITNESS_DESCENDANT;
1972 w_data[i].w_num_descendants++;
1973 w_data[j].w_num_ancestors++;
1976 * Make sure we aren't marking a node as both an
1977 * ancestor and descendant. We should have caught
1978 * this as a lock order reversal earlier.
1980 if ((w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK) &&
1981 (w_rmatrix[i][j] & WITNESS_DESCENDANT_MASK)) {
1982 printf("witness rmatrix paradox! [%d][%d]=%d "
1983 "both ancestor and descendant\n",
1984 i, j, w_rmatrix[i][j]);
1986 printf("Witness disabled.\n");
1989 if ((w_rmatrix[j][i] & WITNESS_ANCESTOR_MASK) &&
1990 (w_rmatrix[j][i] & WITNESS_DESCENDANT_MASK)) {
1991 printf("witness rmatrix paradox! [%d][%d]=%d "
1992 "both ancestor and descendant\n",
1993 j, i, w_rmatrix[j][i]);
1995 printf("Witness disabled.\n");
2003 itismychild(struct witness *parent, struct witness *child)
2007 MPASS(child != NULL && parent != NULL);
2008 if (witness_cold == 0)
2009 mtx_assert(&w_mtx, MA_OWNED);
2011 if (!witness_lock_type_equal(parent, child)) {
2012 if (witness_cold == 0) {
2014 mtx_unlock_spin(&w_mtx);
2019 "%s: parent \"%s\" (%s) and child \"%s\" (%s) are not "
2020 "the same lock type", __func__, parent->w_name,
2021 parent->w_class->lc_name, child->w_name,
2022 child->w_class->lc_name);
2024 mtx_lock_spin(&w_mtx);
2026 adopt(parent, child);
2030 * Generic code for the isitmy*() functions. The rmask parameter is the
2031 * expected relationship of w1 to w2.
2034 _isitmyx(struct witness *w1, struct witness *w2, int rmask, const char *fname)
2036 unsigned char r1, r2;
2041 WITNESS_INDEX_ASSERT(i1);
2042 WITNESS_INDEX_ASSERT(i2);
2043 r1 = w_rmatrix[i1][i2] & WITNESS_RELATED_MASK;
2044 r2 = w_rmatrix[i2][i1] & WITNESS_RELATED_MASK;
2046 /* The flags on one better be the inverse of the flags on the other */
2047 if (!((WITNESS_ATOD(r1) == r2 && WITNESS_DTOA(r2) == r1) ||
2048 (WITNESS_DTOA(r1) == r2 && WITNESS_ATOD(r2) == r1))) {
2049 /* Don't squawk if we're potentially racing with an update. */
2050 if (!mtx_owned(&w_mtx))
2052 printf("%s: rmatrix mismatch between %s (index %d) and %s "
2053 "(index %d): w_rmatrix[%d][%d] == %hhx but "
2054 "w_rmatrix[%d][%d] == %hhx\n",
2055 fname, w1->w_name, i1, w2->w_name, i2, i1, i2, r1,
2058 printf("Witness disabled.\n");
2061 return (r1 & rmask);
2065 * Checks if @child is a direct child of @parent.
2068 isitmychild(struct witness *parent, struct witness *child)
2071 return (_isitmyx(parent, child, WITNESS_PARENT, __func__));
2075 * Checks if @descendant is a direct or inderect descendant of @ancestor.
2078 isitmydescendant(struct witness *ancestor, struct witness *descendant)
2081 return (_isitmyx(ancestor, descendant, WITNESS_ANCESTOR_MASK,
2087 blessed(struct witness *w1, struct witness *w2)
2090 struct witness_blessed *b;
2092 for (i = 0; i < nitems(blessed_list); i++) {
2093 b = &blessed_list[i];
2094 if (strcmp(w1->w_name, b->b_lock1) == 0) {
2095 if (strcmp(w2->w_name, b->b_lock2) == 0)
2099 if (strcmp(w1->w_name, b->b_lock2) == 0)
2100 if (strcmp(w2->w_name, b->b_lock1) == 0)
2107 static struct witness *
2113 if (witness_cold == 0)
2114 mtx_assert(&w_mtx, MA_OWNED);
2116 if (witness_watch == -1) {
2117 mtx_unlock_spin(&w_mtx);
2120 if (STAILQ_EMPTY(&w_free)) {
2122 mtx_unlock_spin(&w_mtx);
2123 printf("WITNESS: unable to allocate a new witness object\n");
2126 w = STAILQ_FIRST(&w_free);
2127 STAILQ_REMOVE_HEAD(&w_free, w_list);
2130 MPASS(index > 0 && index == w_max_used_index+1 &&
2131 index < witness_count);
2132 bzero(w, sizeof(*w));
2134 if (index > w_max_used_index)
2135 w_max_used_index = index;
2140 witness_free(struct witness *w)
2143 STAILQ_INSERT_HEAD(&w_free, w, w_list);
2147 static struct lock_list_entry *
2148 witness_lock_list_get(void)
2150 struct lock_list_entry *lle;
2152 if (witness_watch == -1)
2154 mtx_lock_spin(&w_mtx);
2155 lle = w_lock_list_free;
2158 mtx_unlock_spin(&w_mtx);
2159 printf("%s: witness exhausted\n", __func__);
2162 w_lock_list_free = lle->ll_next;
2163 mtx_unlock_spin(&w_mtx);
2164 bzero(lle, sizeof(*lle));
2169 witness_lock_list_free(struct lock_list_entry *lle)
2172 mtx_lock_spin(&w_mtx);
2173 lle->ll_next = w_lock_list_free;
2174 w_lock_list_free = lle;
2175 mtx_unlock_spin(&w_mtx);
2178 static struct lock_instance *
2179 find_instance(struct lock_list_entry *list, const struct lock_object *lock)
2181 struct lock_list_entry *lle;
2182 struct lock_instance *instance;
2185 for (lle = list; lle != NULL; lle = lle->ll_next)
2186 for (i = lle->ll_count - 1; i >= 0; i--) {
2187 instance = &lle->ll_children[i];
2188 if (instance->li_lock == lock)
2195 witness_list_lock(struct lock_instance *instance,
2196 int (*prnt)(const char *fmt, ...))
2198 struct lock_object *lock;
2200 lock = instance->li_lock;
2201 prnt("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ?
2202 "exclusive" : "shared", LOCK_CLASS(lock)->lc_name, lock->lo_name);
2203 if (lock->lo_witness->w_name != lock->lo_name)
2204 prnt(" (%s)", lock->lo_witness->w_name);
2205 prnt(" r = %d (%p) locked @ %s:%d\n",
2206 instance->li_flags & LI_RECURSEMASK, lock,
2207 fixup_filename(instance->li_file), instance->li_line);
2211 witness_output(const char *fmt, ...)
2217 ret = witness_voutput(fmt, ap);
2223 witness_voutput(const char *fmt, va_list ap)
2228 switch (witness_channel) {
2229 case WITNESS_CONSOLE:
2230 ret = vprintf(fmt, ap);
2233 vlog(LOG_NOTICE, fmt, ap);
2243 witness_thread_has_locks(struct thread *td)
2246 if (td->td_sleeplocks == NULL)
2248 return (td->td_sleeplocks->ll_count != 0);
2252 witness_proc_has_locks(struct proc *p)
2256 FOREACH_THREAD_IN_PROC(p, td) {
2257 if (witness_thread_has_locks(td))
2265 witness_list_locks(struct lock_list_entry **lock_list,
2266 int (*prnt)(const char *fmt, ...))
2268 struct lock_list_entry *lle;
2272 for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
2273 for (i = lle->ll_count - 1; i >= 0; i--) {
2274 witness_list_lock(&lle->ll_children[i], prnt);
2281 * This is a bit risky at best. We call this function when we have timed
2282 * out acquiring a spin lock, and we assume that the other CPU is stuck
2283 * with this lock held. So, we go groveling around in the other CPU's
2284 * per-cpu data to try to find the lock instance for this spin lock to
2285 * see when it was last acquired.
2288 witness_display_spinlock(struct lock_object *lock, struct thread *owner,
2289 int (*prnt)(const char *fmt, ...))
2291 struct lock_instance *instance;
2294 if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU)
2296 pc = pcpu_find(owner->td_oncpu);
2297 instance = find_instance(pc->pc_spinlocks, lock);
2298 if (instance != NULL)
2299 witness_list_lock(instance, prnt);
2303 witness_save(struct lock_object *lock, const char **filep, int *linep)
2305 struct lock_list_entry *lock_list;
2306 struct lock_instance *instance;
2307 struct lock_class *class;
2310 * This function is used independently in locking code to deal with
2311 * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2314 if (SCHEDULER_STOPPED())
2316 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2317 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2319 class = LOCK_CLASS(lock);
2320 if (class->lc_flags & LC_SLEEPLOCK)
2321 lock_list = curthread->td_sleeplocks;
2323 if (witness_skipspin)
2325 lock_list = PCPU_GET(spinlocks);
2327 instance = find_instance(lock_list, lock);
2328 if (instance == NULL) {
2329 kassert_panic("%s: lock (%s) %s not locked", __func__,
2330 class->lc_name, lock->lo_name);
2333 *filep = instance->li_file;
2334 *linep = instance->li_line;
2338 witness_restore(struct lock_object *lock, const char *file, int line)
2340 struct lock_list_entry *lock_list;
2341 struct lock_instance *instance;
2342 struct lock_class *class;
2345 * This function is used independently in locking code to deal with
2346 * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2349 if (SCHEDULER_STOPPED())
2351 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2352 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2354 class = LOCK_CLASS(lock);
2355 if (class->lc_flags & LC_SLEEPLOCK)
2356 lock_list = curthread->td_sleeplocks;
2358 if (witness_skipspin)
2360 lock_list = PCPU_GET(spinlocks);
2362 instance = find_instance(lock_list, lock);
2363 if (instance == NULL)
2364 kassert_panic("%s: lock (%s) %s not locked", __func__,
2365 class->lc_name, lock->lo_name);
2366 lock->lo_witness->w_file = file;
2367 lock->lo_witness->w_line = line;
2368 if (instance == NULL)
2370 instance->li_file = file;
2371 instance->li_line = line;
2375 witness_assert(const struct lock_object *lock, int flags, const char *file,
2378 #ifdef INVARIANT_SUPPORT
2379 struct lock_instance *instance;
2380 struct lock_class *class;
2382 if (lock->lo_witness == NULL || witness_watch < 1 || panicstr != NULL)
2384 class = LOCK_CLASS(lock);
2385 if ((class->lc_flags & LC_SLEEPLOCK) != 0)
2386 instance = find_instance(curthread->td_sleeplocks, lock);
2387 else if ((class->lc_flags & LC_SPINLOCK) != 0)
2388 instance = find_instance(PCPU_GET(spinlocks), lock);
2390 kassert_panic("Lock (%s) %s is not sleep or spin!",
2391 class->lc_name, lock->lo_name);
2396 if (instance != NULL)
2397 kassert_panic("Lock (%s) %s locked @ %s:%d.",
2398 class->lc_name, lock->lo_name,
2399 fixup_filename(file), line);
2402 case LA_LOCKED | LA_RECURSED:
2403 case LA_LOCKED | LA_NOTRECURSED:
2405 case LA_SLOCKED | LA_RECURSED:
2406 case LA_SLOCKED | LA_NOTRECURSED:
2408 case LA_XLOCKED | LA_RECURSED:
2409 case LA_XLOCKED | LA_NOTRECURSED:
2410 if (instance == NULL) {
2411 kassert_panic("Lock (%s) %s not locked @ %s:%d.",
2412 class->lc_name, lock->lo_name,
2413 fixup_filename(file), line);
2416 if ((flags & LA_XLOCKED) != 0 &&
2417 (instance->li_flags & LI_EXCLUSIVE) == 0)
2419 "Lock (%s) %s not exclusively locked @ %s:%d.",
2420 class->lc_name, lock->lo_name,
2421 fixup_filename(file), line);
2422 if ((flags & LA_SLOCKED) != 0 &&
2423 (instance->li_flags & LI_EXCLUSIVE) != 0)
2425 "Lock (%s) %s exclusively locked @ %s:%d.",
2426 class->lc_name, lock->lo_name,
2427 fixup_filename(file), line);
2428 if ((flags & LA_RECURSED) != 0 &&
2429 (instance->li_flags & LI_RECURSEMASK) == 0)
2430 kassert_panic("Lock (%s) %s not recursed @ %s:%d.",
2431 class->lc_name, lock->lo_name,
2432 fixup_filename(file), line);
2433 if ((flags & LA_NOTRECURSED) != 0 &&
2434 (instance->li_flags & LI_RECURSEMASK) != 0)
2435 kassert_panic("Lock (%s) %s recursed @ %s:%d.",
2436 class->lc_name, lock->lo_name,
2437 fixup_filename(file), line);
2440 kassert_panic("Invalid lock assertion at %s:%d.",
2441 fixup_filename(file), line);
2444 #endif /* INVARIANT_SUPPORT */
2448 witness_setflag(struct lock_object *lock, int flag, int set)
2450 struct lock_list_entry *lock_list;
2451 struct lock_instance *instance;
2452 struct lock_class *class;
2454 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2456 class = LOCK_CLASS(lock);
2457 if (class->lc_flags & LC_SLEEPLOCK)
2458 lock_list = curthread->td_sleeplocks;
2460 if (witness_skipspin)
2462 lock_list = PCPU_GET(spinlocks);
2464 instance = find_instance(lock_list, lock);
2465 if (instance == NULL) {
2466 kassert_panic("%s: lock (%s) %s not locked", __func__,
2467 class->lc_name, lock->lo_name);
2472 instance->li_flags |= flag;
2474 instance->li_flags &= ~flag;
2478 witness_norelease(struct lock_object *lock)
2481 witness_setflag(lock, LI_NORELEASE, 1);
2485 witness_releaseok(struct lock_object *lock)
2488 witness_setflag(lock, LI_NORELEASE, 0);
2493 witness_ddb_list(struct thread *td)
2496 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2497 KASSERT(kdb_active, ("%s: not in the debugger", __func__));
2499 if (witness_watch < 1)
2502 witness_list_locks(&td->td_sleeplocks, db_printf);
2505 * We only handle spinlocks if td == curthread. This is somewhat broken
2506 * if td is currently executing on some other CPU and holds spin locks
2507 * as we won't display those locks. If we had a MI way of getting
2508 * the per-cpu data for a given cpu then we could use
2509 * td->td_oncpu to get the list of spinlocks for this thread
2512 * That still wouldn't really fix this unless we locked the scheduler
2513 * lock or stopped the other CPU to make sure it wasn't changing the
2514 * list out from under us. It is probably best to just not try to
2515 * handle threads on other CPU's for now.
2517 if (td == curthread && PCPU_GET(spinlocks) != NULL)
2518 witness_list_locks(PCPU_PTR(spinlocks), db_printf);
2521 DB_SHOW_COMMAND(locks, db_witness_list)
2526 td = db_lookup_thread(addr, true);
2529 witness_ddb_list(td);
2532 DB_SHOW_ALL_COMMAND(locks, db_witness_list_all)
2538 * It would be nice to list only threads and processes that actually
2539 * held sleep locks, but that information is currently not exported
2542 FOREACH_PROC_IN_SYSTEM(p) {
2543 if (!witness_proc_has_locks(p))
2545 FOREACH_THREAD_IN_PROC(p, td) {
2546 if (!witness_thread_has_locks(td))
2548 db_printf("Process %d (%s) thread %p (%d)\n", p->p_pid,
2549 p->p_comm, td, td->td_tid);
2550 witness_ddb_list(td);
2556 DB_SHOW_ALIAS(alllocks, db_witness_list_all)
2558 DB_SHOW_COMMAND(witness, db_witness_display)
2561 witness_ddb_display(db_printf);
2566 sbuf_print_witness_badstacks(struct sbuf *sb, size_t *oldidx)
2568 struct witness_lock_order_data *data1, *data2, *tmp_data1, *tmp_data2;
2569 struct witness *tmp_w1, *tmp_w2, *w1, *w2;
2570 int generation, i, j;
2577 /* Allocate and init temporary storage space. */
2578 tmp_w1 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2579 tmp_w2 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2580 tmp_data1 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2582 tmp_data2 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2584 stack_zero(&tmp_data1->wlod_stack);
2585 stack_zero(&tmp_data2->wlod_stack);
2588 mtx_lock_spin(&w_mtx);
2589 generation = w_generation;
2590 mtx_unlock_spin(&w_mtx);
2591 sbuf_printf(sb, "Number of known direct relationships is %d\n",
2592 w_lohash.wloh_count);
2593 for (i = 1; i < w_max_used_index; i++) {
2594 mtx_lock_spin(&w_mtx);
2595 if (generation != w_generation) {
2596 mtx_unlock_spin(&w_mtx);
2598 /* The graph has changed, try again. */
2605 if (w1->w_reversed == 0) {
2606 mtx_unlock_spin(&w_mtx);
2610 /* Copy w1 locally so we can release the spin lock. */
2612 mtx_unlock_spin(&w_mtx);
2614 if (tmp_w1->w_reversed == 0)
2616 for (j = 1; j < w_max_used_index; j++) {
2617 if ((w_rmatrix[i][j] & WITNESS_REVERSAL) == 0 || i > j)
2620 mtx_lock_spin(&w_mtx);
2621 if (generation != w_generation) {
2622 mtx_unlock_spin(&w_mtx);
2624 /* The graph has changed, try again. */
2631 data1 = witness_lock_order_get(w1, w2);
2632 data2 = witness_lock_order_get(w2, w1);
2635 * Copy information locally so we can release the
2641 stack_zero(&tmp_data1->wlod_stack);
2642 stack_copy(&data1->wlod_stack,
2643 &tmp_data1->wlod_stack);
2645 if (data2 && data2 != data1) {
2646 stack_zero(&tmp_data2->wlod_stack);
2647 stack_copy(&data2->wlod_stack,
2648 &tmp_data2->wlod_stack);
2650 mtx_unlock_spin(&w_mtx);
2653 "\nLock order reversal between \"%s\"(%s) and \"%s\"(%s)!\n",
2654 tmp_w1->w_name, tmp_w1->w_class->lc_name,
2655 tmp_w2->w_name, tmp_w2->w_class->lc_name);
2658 "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2659 tmp_w1->w_name, tmp_w1->w_class->lc_name,
2660 tmp_w2->w_name, tmp_w2->w_class->lc_name);
2661 stack_sbuf_print(sb, &tmp_data1->wlod_stack);
2662 sbuf_printf(sb, "\n");
2664 if (data2 && data2 != data1) {
2666 "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2667 tmp_w2->w_name, tmp_w2->w_class->lc_name,
2668 tmp_w1->w_name, tmp_w1->w_class->lc_name);
2669 stack_sbuf_print(sb, &tmp_data2->wlod_stack);
2670 sbuf_printf(sb, "\n");
2674 mtx_lock_spin(&w_mtx);
2675 if (generation != w_generation) {
2676 mtx_unlock_spin(&w_mtx);
2679 * The graph changed while we were printing stack data,
2686 mtx_unlock_spin(&w_mtx);
2688 /* Free temporary storage space. */
2689 free(tmp_data1, M_TEMP);
2690 free(tmp_data2, M_TEMP);
2691 free(tmp_w1, M_TEMP);
2692 free(tmp_w2, M_TEMP);
2696 sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS)
2701 if (witness_watch < 1) {
2702 error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2706 error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2710 sb = sbuf_new(NULL, NULL, badstack_sbuf_size, SBUF_AUTOEXTEND);
2714 sbuf_print_witness_badstacks(sb, &req->oldidx);
2717 error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1);
2725 sbuf_db_printf_drain(void *arg __unused, const char *data, int len)
2728 return (db_printf("%.*s", len, data));
2731 DB_SHOW_COMMAND(badstacks, db_witness_badstacks)
2737 sbuf_new(&sb, buffer, sizeof(buffer), SBUF_FIXEDLEN);
2738 sbuf_set_drain(&sb, sbuf_db_printf_drain, NULL);
2739 sbuf_print_witness_badstacks(&sb, &dummy);
2745 sysctl_debug_witness_channel(SYSCTL_HANDLER_ARGS)
2747 static const struct {
2748 enum witness_channel channel;
2751 { WITNESS_CONSOLE, "console" },
2752 { WITNESS_LOG, "log" },
2753 { WITNESS_NONE, "none" },
2760 for (i = 0; i < nitems(channels); i++)
2761 if (witness_channel == channels[i].channel) {
2762 snprintf(buf, sizeof(buf), "%s", channels[i].name);
2766 error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
2767 if (error != 0 || req->newptr == NULL)
2771 for (i = 0; i < nitems(channels); i++)
2772 if (strcmp(channels[i].name, buf) == 0) {
2773 witness_channel = channels[i].channel;
2781 sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS)
2788 error = SYSCTL_OUT(req, w_notallowed, sizeof(w_notallowed));
2792 if (witness_watch < 1) {
2793 error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2797 error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2802 error = sysctl_wire_old_buffer(req, 0);
2805 sb = sbuf_new_for_sysctl(NULL, NULL, FULLGRAPH_SBUF_SIZE, req);
2808 sbuf_printf(sb, "\n");
2810 mtx_lock_spin(&w_mtx);
2811 STAILQ_FOREACH(w, &w_all, w_list)
2813 STAILQ_FOREACH(w, &w_all, w_list)
2814 witness_add_fullgraph(sb, w);
2815 mtx_unlock_spin(&w_mtx);
2818 * Close the sbuf and return to userland.
2820 error = sbuf_finish(sb);
2827 sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS)
2831 value = witness_watch;
2832 error = sysctl_handle_int(oidp, &value, 0, req);
2833 if (error != 0 || req->newptr == NULL)
2835 if (value > 1 || value < -1 ||
2836 (witness_watch == -1 && value != witness_watch))
2838 witness_watch = value;
2843 witness_add_fullgraph(struct sbuf *sb, struct witness *w)
2847 if (w->w_displayed != 0 || (w->w_file == NULL && w->w_line == 0))
2851 WITNESS_INDEX_ASSERT(w->w_index);
2852 for (i = 1; i <= w_max_used_index; i++) {
2853 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT) {
2854 sbuf_printf(sb, "\"%s\",\"%s\"\n", w->w_name,
2856 witness_add_fullgraph(sb, &w_data[i]);
2862 * A simple hash function. Takes a key pointer and a key size. If size == 0,
2863 * interprets the key as a string and reads until the null
2864 * terminator. Otherwise, reads the first size bytes. Returns an unsigned 32-bit
2865 * hash value computed from the key.
2868 witness_hash_djb2(const uint8_t *key, uint32_t size)
2870 unsigned int hash = 5381;
2873 /* hash = hash * 33 + key[i] */
2875 for (i = 0; i < size; i++)
2876 hash = ((hash << 5) + hash) + (unsigned int)key[i];
2878 for (i = 0; key[i] != 0; i++)
2879 hash = ((hash << 5) + hash) + (unsigned int)key[i];
2886 * Initializes the two witness hash tables. Called exactly once from
2887 * witness_initialize().
2890 witness_init_hash_tables(void)
2894 MPASS(witness_cold);
2896 /* Initialize the hash tables. */
2897 for (i = 0; i < WITNESS_HASH_SIZE; i++)
2898 w_hash.wh_array[i] = NULL;
2900 w_hash.wh_size = WITNESS_HASH_SIZE;
2901 w_hash.wh_count = 0;
2903 /* Initialize the lock order data hash. */
2905 for (i = 0; i < WITNESS_LO_DATA_COUNT; i++) {
2906 memset(&w_lodata[i], 0, sizeof(w_lodata[i]));
2907 w_lodata[i].wlod_next = w_lofree;
2908 w_lofree = &w_lodata[i];
2910 w_lohash.wloh_size = WITNESS_LO_HASH_SIZE;
2911 w_lohash.wloh_count = 0;
2912 for (i = 0; i < WITNESS_LO_HASH_SIZE; i++)
2913 w_lohash.wloh_array[i] = NULL;
2916 static struct witness *
2917 witness_hash_get(const char *key)
2923 if (witness_cold == 0)
2924 mtx_assert(&w_mtx, MA_OWNED);
2925 hash = witness_hash_djb2(key, 0) % w_hash.wh_size;
2926 w = w_hash.wh_array[hash];
2928 if (strcmp(w->w_name, key) == 0)
2938 witness_hash_put(struct witness *w)
2943 MPASS(w->w_name != NULL);
2944 if (witness_cold == 0)
2945 mtx_assert(&w_mtx, MA_OWNED);
2946 KASSERT(witness_hash_get(w->w_name) == NULL,
2947 ("%s: trying to add a hash entry that already exists!", __func__));
2948 KASSERT(w->w_hash_next == NULL,
2949 ("%s: w->w_hash_next != NULL", __func__));
2951 hash = witness_hash_djb2(w->w_name, 0) % w_hash.wh_size;
2952 w->w_hash_next = w_hash.wh_array[hash];
2953 w_hash.wh_array[hash] = w;
2958 static struct witness_lock_order_data *
2959 witness_lock_order_get(struct witness *parent, struct witness *child)
2961 struct witness_lock_order_data *data = NULL;
2962 struct witness_lock_order_key key;
2965 MPASS(parent != NULL && child != NULL);
2966 key.from = parent->w_index;
2967 key.to = child->w_index;
2968 WITNESS_INDEX_ASSERT(key.from);
2969 WITNESS_INDEX_ASSERT(key.to);
2970 if ((w_rmatrix[parent->w_index][child->w_index]
2971 & WITNESS_LOCK_ORDER_KNOWN) == 0)
2974 hash = witness_hash_djb2((const char*)&key,
2975 sizeof(key)) % w_lohash.wloh_size;
2976 data = w_lohash.wloh_array[hash];
2977 while (data != NULL) {
2978 if (witness_lock_order_key_equal(&data->wlod_key, &key))
2980 data = data->wlod_next;
2988 * Verify that parent and child have a known relationship, are not the same,
2989 * and child is actually a child of parent. This is done without w_mtx
2990 * to avoid contention in the common case.
2993 witness_lock_order_check(struct witness *parent, struct witness *child)
2996 if (parent != child &&
2997 w_rmatrix[parent->w_index][child->w_index]
2998 & WITNESS_LOCK_ORDER_KNOWN &&
2999 isitmychild(parent, child))
3006 witness_lock_order_add(struct witness *parent, struct witness *child)
3008 struct witness_lock_order_data *data = NULL;
3009 struct witness_lock_order_key key;
3012 MPASS(parent != NULL && child != NULL);
3013 key.from = parent->w_index;
3014 key.to = child->w_index;
3015 WITNESS_INDEX_ASSERT(key.from);
3016 WITNESS_INDEX_ASSERT(key.to);
3017 if (w_rmatrix[parent->w_index][child->w_index]
3018 & WITNESS_LOCK_ORDER_KNOWN)
3021 hash = witness_hash_djb2((const char*)&key,
3022 sizeof(key)) % w_lohash.wloh_size;
3023 w_rmatrix[parent->w_index][child->w_index] |= WITNESS_LOCK_ORDER_KNOWN;
3027 w_lofree = data->wlod_next;
3028 data->wlod_next = w_lohash.wloh_array[hash];
3029 data->wlod_key = key;
3030 w_lohash.wloh_array[hash] = data;
3031 w_lohash.wloh_count++;
3032 stack_zero(&data->wlod_stack);
3033 stack_save(&data->wlod_stack);
3037 /* Call this whenever the structure of the witness graph changes. */
3039 witness_increment_graph_generation(void)
3042 if (witness_cold == 0)
3043 mtx_assert(&w_mtx, MA_OWNED);
3048 witness_output_drain(void *arg __unused, const char *data, int len)
3051 witness_output("%.*s", len, data);
3056 witness_debugger(int cond, const char *msg)
3065 if (witness_trace) {
3066 sbuf_new(&sb, buf, sizeof(buf), SBUF_FIXEDLEN);
3067 sbuf_set_drain(&sb, witness_output_drain, NULL);
3071 witness_output("stack backtrace:\n");
3072 stack_sbuf_print_ddb(&sb, &st);
3079 kdb_enter(KDB_WHY_WITNESS, msg);