2 * Copyright (c) 2008 Isilon Systems, Inc.
3 * Copyright (c) 2008 Ilya Maykov <ivmaykov@gmail.com>
4 * Copyright (c) 1998 Berkeley Software Design, Inc.
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
10 * 1. Redistributions of source code must retain the above copyright
11 * notice, this list of conditions and the following disclaimer.
12 * 2. Redistributions in binary form must reproduce the above copyright
13 * notice, this list of conditions and the following disclaimer in the
14 * documentation and/or other materials provided with the distribution.
15 * 3. Berkeley Software Design Inc's name may not be used to endorse or
16 * promote products derived from this software without specific prior
19 * THIS SOFTWARE IS PROVIDED BY BERKELEY SOFTWARE DESIGN INC ``AS IS'' AND
20 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
21 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
22 * ARE DISCLAIMED. IN NO EVENT SHALL BERKELEY SOFTWARE DESIGN INC BE LIABLE
23 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
24 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
25 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
26 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
27 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
28 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31 * from BSDI $Id: mutex_witness.c,v 1.1.2.20 2000/04/27 03:10:27 cp Exp $
32 * and BSDI $Id: synch_machdep.c,v 2.3.2.39 2000/04/27 03:10:25 cp Exp $
36 * Implementation of the `witness' lock verifier. Originally implemented for
37 * mutexes in BSD/OS. Extended to handle generic lock objects and lock
43 * Pronunciation: 'wit-n&s
45 * Etymology: Middle English witnesse, from Old English witnes knowledge,
46 * testimony, witness, from 2wit
47 * Date: before 12th century
48 * 1 : attestation of a fact or event : TESTIMONY
49 * 2 : one that gives evidence; specifically : one who testifies in
50 * a cause or before a judicial tribunal
51 * 3 : one asked to be present at a transaction so as to be able to
52 * testify to its having taken place
53 * 4 : one who has personal knowledge of something
54 * 5 a : something serving as evidence or proof : SIGN
55 * b : public affirmation by word or example of usually
56 * religious faith or conviction <the heroic witness to divine
58 * 6 capitalized : a member of the Jehovah's Witnesses
62 * Special rules concerning Giant and lock orders:
64 * 1) Giant must be acquired before any other mutexes. Stated another way,
65 * no other mutex may be held when Giant is acquired.
67 * 2) Giant must be released when blocking on a sleepable lock.
69 * This rule is less obvious, but is a result of Giant providing the same
70 * semantics as spl(). Basically, when a thread sleeps, it must release
71 * Giant. When a thread blocks on a sleepable lock, it sleeps. Hence rule
74 * 3) Giant may be acquired before or after sleepable locks.
76 * This rule is also not quite as obvious. Giant may be acquired after
77 * a sleepable lock because it is a non-sleepable lock and non-sleepable
78 * locks may always be acquired while holding a sleepable lock. The second
79 * case, Giant before a sleepable lock, follows from rule 2) above. Suppose
80 * you have two threads T1 and T2 and a sleepable lock X. Suppose that T1
81 * acquires X and blocks on Giant. Then suppose that T2 acquires Giant and
82 * blocks on X. When T2 blocks on X, T2 will release Giant allowing T1 to
83 * execute. Thus, acquiring Giant both before and after a sleepable lock
84 * will not result in a lock order reversal.
87 #include <sys/cdefs.h>
88 __FBSDID("$FreeBSD$");
91 #include "opt_hwpmc_hooks.h"
92 #include "opt_stack.h"
93 #include "opt_witness.h"
95 #include <sys/param.h>
98 #include <sys/kernel.h>
100 #include <sys/lock.h>
101 #include <sys/malloc.h>
102 #include <sys/mutex.h>
103 #include <sys/priv.h>
104 #include <sys/proc.h>
105 #include <sys/sbuf.h>
106 #include <sys/sched.h>
107 #include <sys/stack.h>
108 #include <sys/sysctl.h>
109 #include <sys/syslog.h>
110 #include <sys/systm.h>
116 #include <machine/stdarg.h>
118 #if !defined(DDB) && !defined(STACK)
119 #error "DDB or STACK options are required for WITNESS"
122 /* Note that these traces do not work with KTR_ALQ. */
124 #define KTR_WITNESS KTR_SUBSYS
126 #define KTR_WITNESS 0
129 #define LI_RECURSEMASK 0x0000ffff /* Recursion depth of lock instance. */
130 #define LI_EXCLUSIVE 0x00010000 /* Exclusive lock instance. */
131 #define LI_NORELEASE 0x00020000 /* Lock not allowed to be released. */
133 /* Define this to check for blessed mutexes */
136 #ifndef WITNESS_COUNT
137 #define WITNESS_COUNT 1536
139 #define WITNESS_HASH_SIZE 251 /* Prime, gives load factor < 2 */
140 #define WITNESS_PENDLIST (1024 + MAXCPU)
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];
280 struct witness_blessed {
286 struct witness_pendhelp {
288 struct lock_object *wh_lock;
291 struct witness_order_list_entry {
293 struct lock_class *w_class;
297 * Returns 0 if one of the locks is a spin lock and the other is not.
298 * Returns 1 otherwise.
301 witness_lock_type_equal(struct witness *w1, struct witness *w2)
304 return ((w1->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) ==
305 (w2->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)));
309 witness_lock_order_key_equal(const struct witness_lock_order_key *a,
310 const struct witness_lock_order_key *b)
313 return (a->from == b->from && a->to == b->to);
316 static int _isitmyx(struct witness *w1, struct witness *w2, int rmask,
318 static void adopt(struct witness *parent, struct witness *child);
320 static int blessed(struct witness *, struct witness *);
322 static void depart(struct witness *w);
323 static struct witness *enroll(const char *description,
324 struct lock_class *lock_class);
325 static struct lock_instance *find_instance(struct lock_list_entry *list,
326 const struct lock_object *lock);
327 static int isitmychild(struct witness *parent, struct witness *child);
328 static int isitmydescendant(struct witness *parent, struct witness *child);
329 static void itismychild(struct witness *parent, struct witness *child);
330 static int sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS);
331 static int sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS);
332 static int sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS);
333 static int sysctl_debug_witness_channel(SYSCTL_HANDLER_ARGS);
334 static void witness_add_fullgraph(struct sbuf *sb, struct witness *parent);
336 static void witness_ddb_compute_levels(void);
337 static void witness_ddb_display(int(*)(const char *fmt, ...));
338 static void witness_ddb_display_descendants(int(*)(const char *fmt, ...),
339 struct witness *, int indent);
340 static void witness_ddb_display_list(int(*prnt)(const char *fmt, ...),
341 struct witness_list *list);
342 static void witness_ddb_level_descendants(struct witness *parent, int l);
343 static void witness_ddb_list(struct thread *td);
345 static void witness_debugger(int cond, const char *msg);
346 static void witness_free(struct witness *m);
347 static struct witness *witness_get(void);
348 static uint32_t witness_hash_djb2(const uint8_t *key, uint32_t size);
349 static struct witness *witness_hash_get(const char *key);
350 static void witness_hash_put(struct witness *w);
351 static void witness_init_hash_tables(void);
352 static void witness_increment_graph_generation(void);
353 static void witness_lock_list_free(struct lock_list_entry *lle);
354 static struct lock_list_entry *witness_lock_list_get(void);
355 static int witness_lock_order_add(struct witness *parent,
356 struct witness *child);
357 static int witness_lock_order_check(struct witness *parent,
358 struct witness *child);
359 static struct witness_lock_order_data *witness_lock_order_get(
360 struct witness *parent,
361 struct witness *child);
362 static void witness_list_lock(struct lock_instance *instance,
363 int (*prnt)(const char *fmt, ...));
364 static int witness_output(const char *fmt, ...) __printflike(1, 2);
365 static int witness_voutput(const char *fmt, va_list ap) __printflike(1, 0);
366 static void witness_setflag(struct lock_object *lock, int flag, int set);
368 static SYSCTL_NODE(_debug, OID_AUTO, witness, CTLFLAG_RW, NULL,
372 * If set to 0, lock order checking is disabled. If set to -1,
373 * witness is completely disabled. Otherwise witness performs full
374 * lock order checking for all locks. At runtime, lock order checking
375 * may be toggled. However, witness cannot be reenabled once it is
376 * completely disabled.
378 static int witness_watch = 1;
379 SYSCTL_PROC(_debug_witness, OID_AUTO, watch, CTLFLAG_RWTUN | CTLTYPE_INT, NULL, 0,
380 sysctl_debug_witness_watch, "I", "witness is watching lock operations");
384 * When KDB is enabled and witness_kdb is 1, it will cause the system
385 * to drop into kdebug() when:
386 * - a lock hierarchy violation occurs
387 * - locks are held when going to sleep.
394 SYSCTL_INT(_debug_witness, OID_AUTO, kdb, CTLFLAG_RWTUN, &witness_kdb, 0, "");
397 #if defined(DDB) || defined(KDB)
399 * When DDB or KDB is enabled and witness_trace is 1, it will cause the system
400 * to print a stack trace:
401 * - a lock hierarchy violation occurs
402 * - locks are held when going to sleep.
404 int witness_trace = 1;
405 SYSCTL_INT(_debug_witness, OID_AUTO, trace, CTLFLAG_RWTUN, &witness_trace, 0, "");
406 #endif /* DDB || KDB */
408 #ifdef WITNESS_SKIPSPIN
409 int witness_skipspin = 1;
411 int witness_skipspin = 0;
413 SYSCTL_INT(_debug_witness, OID_AUTO, skipspin, CTLFLAG_RDTUN, &witness_skipspin, 0, "");
415 int badstack_sbuf_size;
417 int witness_count = WITNESS_COUNT;
418 SYSCTL_INT(_debug_witness, OID_AUTO, witness_count, CTLFLAG_RDTUN,
419 &witness_count, 0, "");
422 * Output channel for witness messages. By default we print to the console.
424 enum witness_channel {
430 static enum witness_channel witness_channel = WITNESS_CONSOLE;
431 SYSCTL_PROC(_debug_witness, OID_AUTO, output_channel, CTLTYPE_STRING |
432 CTLFLAG_RWTUN, NULL, 0, sysctl_debug_witness_channel, "A",
433 "Output channel for warnings");
436 * Call this to print out the relations between locks.
438 SYSCTL_PROC(_debug_witness, OID_AUTO, fullgraph, CTLTYPE_STRING | CTLFLAG_RD,
439 NULL, 0, sysctl_debug_witness_fullgraph, "A", "Show locks relation graphs");
442 * Call this to print out the witness faulty stacks.
444 SYSCTL_PROC(_debug_witness, OID_AUTO, badstacks, CTLTYPE_STRING | CTLFLAG_RD,
445 NULL, 0, sysctl_debug_witness_badstacks, "A", "Show bad witness stacks");
447 static struct mtx w_mtx;
450 static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free);
451 static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all);
454 static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin);
455 static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep);
458 static struct lock_list_entry *w_lock_list_free = NULL;
459 static struct witness_pendhelp pending_locks[WITNESS_PENDLIST];
460 static u_int pending_cnt;
462 static int w_free_cnt, w_spin_cnt, w_sleep_cnt;
463 SYSCTL_INT(_debug_witness, OID_AUTO, free_cnt, CTLFLAG_RD, &w_free_cnt, 0, "");
464 SYSCTL_INT(_debug_witness, OID_AUTO, spin_cnt, CTLFLAG_RD, &w_spin_cnt, 0, "");
465 SYSCTL_INT(_debug_witness, OID_AUTO, sleep_cnt, CTLFLAG_RD, &w_sleep_cnt, 0,
468 static struct witness *w_data;
469 static uint8_t **w_rmatrix;
470 static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT];
471 static struct witness_hash w_hash; /* The witness hash table. */
473 /* The lock order data hash */
474 static struct witness_lock_order_data w_lodata[WITNESS_LO_DATA_COUNT];
475 static struct witness_lock_order_data *w_lofree = NULL;
476 static struct witness_lock_order_hash w_lohash;
477 static int w_max_used_index = 0;
478 static unsigned int w_generation = 0;
479 static const char w_notrunning[] = "Witness not running\n";
480 static const char w_stillcold[] = "Witness is still cold\n";
483 static struct witness_order_list_entry order_lists[] = {
487 { "proctree", &lock_class_sx },
488 { "allproc", &lock_class_sx },
489 { "allprison", &lock_class_sx },
494 { "Giant", &lock_class_mtx_sleep },
495 { "pipe mutex", &lock_class_mtx_sleep },
496 { "sigio lock", &lock_class_mtx_sleep },
497 { "process group", &lock_class_mtx_sleep },
498 { "process lock", &lock_class_mtx_sleep },
499 { "session", &lock_class_mtx_sleep },
500 { "uidinfo hash", &lock_class_rw },
502 { "pmc-sleep", &lock_class_mtx_sleep },
504 { "time lock", &lock_class_mtx_sleep },
509 { "umtx lock", &lock_class_mtx_sleep },
514 { "accept", &lock_class_mtx_sleep },
515 { "so_snd", &lock_class_mtx_sleep },
516 { "so_rcv", &lock_class_mtx_sleep },
517 { "sellck", &lock_class_mtx_sleep },
522 { "so_rcv", &lock_class_mtx_sleep },
523 { "radix node head", &lock_class_rw },
524 { "rtentry", &lock_class_mtx_sleep },
525 { "ifaddr", &lock_class_mtx_sleep },
529 * protocol locks before interface locks, after UDP locks.
531 { "udpinp", &lock_class_rw },
532 { "in_multi_mtx", &lock_class_mtx_sleep },
533 { "igmp_mtx", &lock_class_mtx_sleep },
534 { "if_addr_lock", &lock_class_rw },
538 * protocol locks before interface locks, after UDP locks.
540 { "udpinp", &lock_class_rw },
541 { "in6_multi_mtx", &lock_class_mtx_sleep },
542 { "mld_mtx", &lock_class_mtx_sleep },
543 { "if_addr_lock", &lock_class_rw },
546 * UNIX Domain Sockets
548 { "unp_link_rwlock", &lock_class_rw },
549 { "unp_list_lock", &lock_class_mtx_sleep },
550 { "unp", &lock_class_mtx_sleep },
551 { "so_snd", &lock_class_mtx_sleep },
556 { "udp", &lock_class_rw },
557 { "udpinp", &lock_class_rw },
558 { "so_snd", &lock_class_mtx_sleep },
563 { "tcp", &lock_class_rw },
564 { "tcpinp", &lock_class_rw },
565 { "so_snd", &lock_class_mtx_sleep },
570 { "bpf global lock", &lock_class_mtx_sleep },
571 { "bpf interface lock", &lock_class_rw },
572 { "bpf cdev lock", &lock_class_mtx_sleep },
577 { "nfsd_mtx", &lock_class_mtx_sleep },
578 { "so_snd", &lock_class_mtx_sleep },
584 { "802.11 com lock", &lock_class_mtx_sleep},
589 { "network driver", &lock_class_mtx_sleep},
595 { "ng_node", &lock_class_mtx_sleep },
596 { "ng_worklist", &lock_class_mtx_sleep },
601 { "vm map (system)", &lock_class_mtx_sleep },
602 { "vm pagequeue", &lock_class_mtx_sleep },
603 { "vnode interlock", &lock_class_mtx_sleep },
604 { "cdev", &lock_class_mtx_sleep },
609 { "vm map (user)", &lock_class_sx },
610 { "vm object", &lock_class_rw },
611 { "vm page", &lock_class_mtx_sleep },
612 { "vm pagequeue", &lock_class_mtx_sleep },
613 { "pmap pv global", &lock_class_rw },
614 { "pmap", &lock_class_mtx_sleep },
615 { "pmap pv list", &lock_class_rw },
616 { "vm page free queue", &lock_class_mtx_sleep },
619 * kqueue/VFS interaction
621 { "kqueue", &lock_class_mtx_sleep },
622 { "struct mount mtx", &lock_class_mtx_sleep },
623 { "vnode interlock", &lock_class_mtx_sleep },
628 { "ncvn", &lock_class_mtx_sleep },
629 { "ncbuc", &lock_class_rw },
630 { "vnode interlock", &lock_class_mtx_sleep },
631 { "ncneg", &lock_class_mtx_sleep },
636 { "dn->dn_mtx", &lock_class_sx },
637 { "dr->dt.di.dr_mtx", &lock_class_sx },
638 { "db->db_mtx", &lock_class_sx },
644 { "ap boot", &lock_class_mtx_spin },
646 { "rm.mutex_mtx", &lock_class_mtx_spin },
647 { "sio", &lock_class_mtx_spin },
649 { "cy", &lock_class_mtx_spin },
652 { "pcib_mtx", &lock_class_mtx_spin },
653 { "rtc_mtx", &lock_class_mtx_spin },
655 { "scc_hwmtx", &lock_class_mtx_spin },
656 { "uart_hwmtx", &lock_class_mtx_spin },
657 { "fast_taskqueue", &lock_class_mtx_spin },
658 { "intr table", &lock_class_mtx_spin },
660 { "pmc-per-proc", &lock_class_mtx_spin },
662 { "process slock", &lock_class_mtx_spin },
663 { "syscons video lock", &lock_class_mtx_spin },
664 { "sleepq chain", &lock_class_mtx_spin },
665 { "rm_spinlock", &lock_class_mtx_spin },
666 { "turnstile chain", &lock_class_mtx_spin },
667 { "turnstile lock", &lock_class_mtx_spin },
668 { "sched lock", &lock_class_mtx_spin },
669 { "td_contested", &lock_class_mtx_spin },
670 { "callout", &lock_class_mtx_spin },
671 { "entropy harvest mutex", &lock_class_mtx_spin },
673 { "smp rendezvous", &lock_class_mtx_spin },
676 { "tlb0", &lock_class_mtx_spin },
681 { "intrcnt", &lock_class_mtx_spin },
682 { "icu", &lock_class_mtx_spin },
683 #if defined(SMP) && defined(__sparc64__)
684 { "ipi", &lock_class_mtx_spin },
687 { "allpmaps", &lock_class_mtx_spin },
688 { "descriptor tables", &lock_class_mtx_spin },
690 { "clk", &lock_class_mtx_spin },
691 { "cpuset", &lock_class_mtx_spin },
692 { "mprof lock", &lock_class_mtx_spin },
693 { "zombie lock", &lock_class_mtx_spin },
694 { "ALD Queue", &lock_class_mtx_spin },
695 #if defined(__i386__) || defined(__amd64__)
696 { "pcicfg", &lock_class_mtx_spin },
697 { "NDIS thread lock", &lock_class_mtx_spin },
699 { "tw_osl_io_lock", &lock_class_mtx_spin },
700 { "tw_osl_q_lock", &lock_class_mtx_spin },
701 { "tw_cl_io_lock", &lock_class_mtx_spin },
702 { "tw_cl_intr_lock", &lock_class_mtx_spin },
703 { "tw_cl_gen_lock", &lock_class_mtx_spin },
705 { "pmc-leaf", &lock_class_mtx_spin },
707 { "blocked lock", &lock_class_mtx_spin },
714 * Pairs of locks which have been blessed
715 * Don't complain about order problems with blessed locks
717 static struct witness_blessed blessed_list[] = {
722 * This global is set to 0 once it becomes safe to use the witness code.
724 static int witness_cold = 1;
727 * This global is set to 1 once the static lock orders have been enrolled
728 * so that a warning can be issued for any spin locks enrolled later.
730 static int witness_spin_warn = 0;
732 /* Trim useless garbage from filenames. */
734 fixup_filename(const char *file)
739 while (strncmp(file, "../", 3) == 0)
745 * The WITNESS-enabled diagnostic code. Note that the witness code does
746 * assume that the early boot is single-threaded at least until after this
747 * routine is completed.
750 witness_initialize(void *dummy __unused)
752 struct lock_object *lock;
753 struct witness_order_list_entry *order;
754 struct witness *w, *w1;
757 w_data = malloc(sizeof (struct witness) * witness_count, M_WITNESS,
760 w_rmatrix = malloc(sizeof(*w_rmatrix) * (witness_count + 1),
761 M_WITNESS, M_WAITOK | M_ZERO);
763 for (i = 0; i < witness_count + 1; i++) {
764 w_rmatrix[i] = malloc(sizeof(*w_rmatrix[i]) *
765 (witness_count + 1), M_WITNESS, M_WAITOK | M_ZERO);
767 badstack_sbuf_size = witness_count * 256;
770 * We have to release Giant before initializing its witness
771 * structure so that WITNESS doesn't get confused.
774 mtx_assert(&Giant, MA_NOTOWNED);
776 CTR1(KTR_WITNESS, "%s: initializing witness", __func__);
777 mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET |
778 MTX_NOWITNESS | MTX_NOPROFILE);
779 for (i = witness_count - 1; i >= 0; i--) {
781 memset(w, 0, sizeof(*w));
782 w_data[i].w_index = i; /* Witness index never changes. */
785 KASSERT(STAILQ_FIRST(&w_free)->w_index == 0,
786 ("%s: Invalid list of free witness objects", __func__));
788 /* Witness with index 0 is not used to aid in debugging. */
789 STAILQ_REMOVE_HEAD(&w_free, w_list);
792 for (i = 0; i < witness_count; i++) {
793 memset(w_rmatrix[i], 0, sizeof(*w_rmatrix[i]) *
794 (witness_count + 1));
797 for (i = 0; i < LOCK_CHILDCOUNT; i++)
798 witness_lock_list_free(&w_locklistdata[i]);
799 witness_init_hash_tables();
801 /* First add in all the specified order lists. */
802 for (order = order_lists; order->w_name != NULL; order++) {
803 w = enroll(order->w_name, order->w_class);
806 w->w_file = "order list";
807 for (order++; order->w_name != NULL; order++) {
808 w1 = enroll(order->w_name, order->w_class);
811 w1->w_file = "order list";
816 witness_spin_warn = 1;
818 /* Iterate through all locks and add them to witness. */
819 for (i = 0; pending_locks[i].wh_lock != NULL; i++) {
820 lock = pending_locks[i].wh_lock;
821 KASSERT(lock->lo_flags & LO_WITNESS,
822 ("%s: lock %s is on pending list but not LO_WITNESS",
823 __func__, lock->lo_name));
824 lock->lo_witness = enroll(pending_locks[i].wh_type,
828 /* Mark the witness code as being ready for use. */
833 SYSINIT(witness_init, SI_SUB_WITNESS, SI_ORDER_FIRST, witness_initialize,
837 witness_init(struct lock_object *lock, const char *type)
839 struct lock_class *class;
841 /* Various sanity checks. */
842 class = LOCK_CLASS(lock);
843 if ((lock->lo_flags & LO_RECURSABLE) != 0 &&
844 (class->lc_flags & LC_RECURSABLE) == 0)
845 kassert_panic("%s: lock (%s) %s can not be recursable",
846 __func__, class->lc_name, lock->lo_name);
847 if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
848 (class->lc_flags & LC_SLEEPABLE) == 0)
849 kassert_panic("%s: lock (%s) %s can not be sleepable",
850 __func__, class->lc_name, lock->lo_name);
851 if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
852 (class->lc_flags & LC_UPGRADABLE) == 0)
853 kassert_panic("%s: lock (%s) %s can not be upgradable",
854 __func__, class->lc_name, lock->lo_name);
857 * If we shouldn't watch this lock, then just clear lo_witness.
858 * Otherwise, if witness_cold is set, then it is too early to
859 * enroll this lock, so defer it to witness_initialize() by adding
860 * it to the pending_locks list. If it is not too early, then enroll
863 if (witness_watch < 1 || panicstr != NULL ||
864 (lock->lo_flags & LO_WITNESS) == 0)
865 lock->lo_witness = NULL;
866 else if (witness_cold) {
867 pending_locks[pending_cnt].wh_lock = lock;
868 pending_locks[pending_cnt++].wh_type = type;
869 if (pending_cnt > WITNESS_PENDLIST)
870 panic("%s: pending locks list is too small, "
871 "increase WITNESS_PENDLIST\n",
874 lock->lo_witness = enroll(type, class);
878 witness_destroy(struct lock_object *lock)
880 struct lock_class *class;
883 class = LOCK_CLASS(lock);
886 panic("lock (%s) %s destroyed while witness_cold",
887 class->lc_name, lock->lo_name);
889 /* XXX: need to verify that no one holds the lock */
890 if ((lock->lo_flags & LO_WITNESS) == 0 || lock->lo_witness == NULL)
892 w = lock->lo_witness;
894 mtx_lock_spin(&w_mtx);
895 MPASS(w->w_refcount > 0);
898 if (w->w_refcount == 0)
900 mtx_unlock_spin(&w_mtx);
905 witness_ddb_compute_levels(void)
910 * First clear all levels.
912 STAILQ_FOREACH(w, &w_all, w_list)
916 * Look for locks with no parents and level all their descendants.
918 STAILQ_FOREACH(w, &w_all, w_list) {
920 /* If the witness has ancestors (is not a root), skip it. */
921 if (w->w_num_ancestors > 0)
923 witness_ddb_level_descendants(w, 0);
928 witness_ddb_level_descendants(struct witness *w, int l)
932 if (w->w_ddb_level >= l)
938 for (i = 1; i <= w_max_used_index; i++) {
939 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
940 witness_ddb_level_descendants(&w_data[i], l);
945 witness_ddb_display_descendants(int(*prnt)(const char *fmt, ...),
946 struct witness *w, int indent)
950 for (i = 0; i < indent; i++)
952 prnt("%s (type: %s, depth: %d, active refs: %d)",
953 w->w_name, w->w_class->lc_name,
954 w->w_ddb_level, w->w_refcount);
955 if (w->w_displayed) {
956 prnt(" -- (already displayed)\n");
960 if (w->w_file != NULL && w->w_line != 0)
961 prnt(" -- last acquired @ %s:%d\n", fixup_filename(w->w_file),
964 prnt(" -- never acquired\n");
966 WITNESS_INDEX_ASSERT(w->w_index);
967 for (i = 1; i <= w_max_used_index; i++) {
970 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
971 witness_ddb_display_descendants(prnt, &w_data[i],
977 witness_ddb_display_list(int(*prnt)(const char *fmt, ...),
978 struct witness_list *list)
982 STAILQ_FOREACH(w, list, w_typelist) {
983 if (w->w_file == NULL || w->w_ddb_level > 0)
986 /* This lock has no anscestors - display its descendants. */
987 witness_ddb_display_descendants(prnt, w, 0);
994 witness_ddb_display(int(*prnt)(const char *fmt, ...))
998 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
999 witness_ddb_compute_levels();
1001 /* Clear all the displayed flags. */
1002 STAILQ_FOREACH(w, &w_all, w_list)
1006 * First, handle sleep locks which have been acquired at least
1009 prnt("Sleep locks:\n");
1010 witness_ddb_display_list(prnt, &w_sleep);
1015 * Now do spin locks which have been acquired at least once.
1017 prnt("\nSpin locks:\n");
1018 witness_ddb_display_list(prnt, &w_spin);
1023 * Finally, any locks which have not been acquired yet.
1025 prnt("\nLocks which were never acquired:\n");
1026 STAILQ_FOREACH(w, &w_all, w_list) {
1027 if (w->w_file != NULL || w->w_refcount == 0)
1029 prnt("%s (type: %s, depth: %d)\n", w->w_name,
1030 w->w_class->lc_name, w->w_ddb_level);
1038 witness_defineorder(struct lock_object *lock1, struct lock_object *lock2)
1041 if (witness_watch == -1 || panicstr != NULL)
1044 /* Require locks that witness knows about. */
1045 if (lock1 == NULL || lock1->lo_witness == NULL || lock2 == NULL ||
1046 lock2->lo_witness == NULL)
1049 mtx_assert(&w_mtx, MA_NOTOWNED);
1050 mtx_lock_spin(&w_mtx);
1053 * If we already have either an explicit or implied lock order that
1054 * is the other way around, then return an error.
1056 if (witness_watch &&
1057 isitmydescendant(lock2->lo_witness, lock1->lo_witness)) {
1058 mtx_unlock_spin(&w_mtx);
1062 /* Try to add the new order. */
1063 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1064 lock2->lo_witness->w_name, lock1->lo_witness->w_name);
1065 itismychild(lock1->lo_witness, lock2->lo_witness);
1066 mtx_unlock_spin(&w_mtx);
1071 witness_checkorder(struct lock_object *lock, int flags, const char *file,
1072 int line, struct lock_object *interlock)
1074 struct lock_list_entry *lock_list, *lle;
1075 struct lock_instance *lock1, *lock2, *plock;
1076 struct lock_class *class, *iclass;
1077 struct witness *w, *w1;
1081 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL ||
1085 w = lock->lo_witness;
1086 class = LOCK_CLASS(lock);
1089 if (class->lc_flags & LC_SLEEPLOCK) {
1092 * Since spin locks include a critical section, this check
1093 * implicitly enforces a lock order of all sleep locks before
1096 if (td->td_critnest != 0 && !kdb_active)
1097 kassert_panic("acquiring blockable sleep lock with "
1098 "spinlock or critical section held (%s) %s @ %s:%d",
1099 class->lc_name, lock->lo_name,
1100 fixup_filename(file), line);
1103 * If this is the first lock acquired then just return as
1104 * no order checking is needed.
1106 lock_list = td->td_sleeplocks;
1107 if (lock_list == NULL || lock_list->ll_count == 0)
1112 * If this is the first lock, just return as no order
1113 * checking is needed. Avoid problems with thread
1114 * migration pinning the thread while checking if
1115 * spinlocks are held. If at least one spinlock is held
1116 * the thread is in a safe path and it is allowed to
1120 lock_list = PCPU_GET(spinlocks);
1121 if (lock_list == NULL || lock_list->ll_count == 0) {
1129 * Check to see if we are recursing on a lock we already own. If
1130 * so, make sure that we don't mismatch exclusive and shared lock
1133 lock1 = find_instance(lock_list, lock);
1134 if (lock1 != NULL) {
1135 if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
1136 (flags & LOP_EXCLUSIVE) == 0) {
1137 witness_output("shared lock of (%s) %s @ %s:%d\n",
1138 class->lc_name, lock->lo_name,
1139 fixup_filename(file), line);
1140 witness_output("while exclusively locked from %s:%d\n",
1141 fixup_filename(lock1->li_file), lock1->li_line);
1142 kassert_panic("excl->share");
1144 if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
1145 (flags & LOP_EXCLUSIVE) != 0) {
1146 witness_output("exclusive lock of (%s) %s @ %s:%d\n",
1147 class->lc_name, lock->lo_name,
1148 fixup_filename(file), line);
1149 witness_output("while share locked from %s:%d\n",
1150 fixup_filename(lock1->li_file), lock1->li_line);
1151 kassert_panic("share->excl");
1156 /* Warn if the interlock is not locked exactly once. */
1157 if (interlock != NULL) {
1158 iclass = LOCK_CLASS(interlock);
1159 lock1 = find_instance(lock_list, interlock);
1161 kassert_panic("interlock (%s) %s not locked @ %s:%d",
1162 iclass->lc_name, interlock->lo_name,
1163 fixup_filename(file), line);
1164 else if ((lock1->li_flags & LI_RECURSEMASK) != 0)
1165 kassert_panic("interlock (%s) %s recursed @ %s:%d",
1166 iclass->lc_name, interlock->lo_name,
1167 fixup_filename(file), line);
1171 * Find the previously acquired lock, but ignore interlocks.
1173 plock = &lock_list->ll_children[lock_list->ll_count - 1];
1174 if (interlock != NULL && plock->li_lock == interlock) {
1175 if (lock_list->ll_count > 1)
1177 &lock_list->ll_children[lock_list->ll_count - 2];
1179 lle = lock_list->ll_next;
1182 * The interlock is the only lock we hold, so
1187 plock = &lle->ll_children[lle->ll_count - 1];
1192 * Try to perform most checks without a lock. If this succeeds we
1193 * can skip acquiring the lock and return success. Otherwise we redo
1194 * the check with the lock held to handle races with concurrent updates.
1196 w1 = plock->li_lock->lo_witness;
1197 if (witness_lock_order_check(w1, w))
1200 mtx_lock_spin(&w_mtx);
1201 if (witness_lock_order_check(w1, w)) {
1202 mtx_unlock_spin(&w_mtx);
1205 witness_lock_order_add(w1, w);
1208 * Check for duplicate locks of the same type. Note that we only
1209 * have to check for this on the last lock we just acquired. Any
1210 * other cases will be caught as lock order violations.
1214 if (!(lock->lo_flags & LO_DUPOK) && !(flags & LOP_DUPOK) &&
1215 !(w_rmatrix[i][i] & WITNESS_REVERSAL)) {
1216 w_rmatrix[i][i] |= WITNESS_REVERSAL;
1218 mtx_unlock_spin(&w_mtx);
1220 "acquiring duplicate lock of same type: \"%s\"\n",
1222 witness_output(" 1st %s @ %s:%d\n", plock->li_lock->lo_name,
1223 fixup_filename(plock->li_file), plock->li_line);
1224 witness_output(" 2nd %s @ %s:%d\n", lock->lo_name,
1225 fixup_filename(file), line);
1226 witness_debugger(1, __func__);
1228 mtx_unlock_spin(&w_mtx);
1231 mtx_assert(&w_mtx, MA_OWNED);
1234 * If we know that the lock we are acquiring comes after
1235 * the lock we most recently acquired in the lock order tree,
1236 * then there is no need for any further checks.
1238 if (isitmychild(w1, w))
1241 for (j = 0, lle = lock_list; lle != NULL; lle = lle->ll_next) {
1242 for (i = lle->ll_count - 1; i >= 0; i--, j++) {
1244 MPASS(j < LOCK_CHILDCOUNT * LOCK_NCHILDREN);
1245 lock1 = &lle->ll_children[i];
1248 * Ignore the interlock.
1250 if (interlock == lock1->li_lock)
1254 * If this lock doesn't undergo witness checking,
1257 w1 = lock1->li_lock->lo_witness;
1259 KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
1260 ("lock missing witness structure"));
1265 * If we are locking Giant and this is a sleepable
1266 * lock, then skip it.
1268 if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0 &&
1269 lock == &Giant.lock_object)
1273 * If we are locking a sleepable lock and this lock
1274 * is Giant, then skip it.
1276 if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1277 lock1->li_lock == &Giant.lock_object)
1281 * If we are locking a sleepable lock and this lock
1282 * isn't sleepable, we want to treat it as a lock
1283 * order violation to enfore a general lock order of
1284 * sleepable locks before non-sleepable locks.
1286 if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1287 (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
1291 * If we are locking Giant and this is a non-sleepable
1292 * lock, then treat it as a reversal.
1294 if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0 &&
1295 lock == &Giant.lock_object)
1299 * Check the lock order hierarchy for a reveresal.
1301 if (!isitmydescendant(w, w1))
1306 * We have a lock order violation, check to see if it
1307 * is allowed or has already been yelled about.
1312 * If the lock order is blessed, just bail. We don't
1313 * look for other lock order violations though, which
1320 /* Bail if this violation is known */
1321 if (w_rmatrix[w1->w_index][w->w_index] & WITNESS_REVERSAL)
1324 /* Record this as a violation */
1325 w_rmatrix[w1->w_index][w->w_index] |= WITNESS_REVERSAL;
1326 w_rmatrix[w->w_index][w1->w_index] |= WITNESS_REVERSAL;
1327 w->w_reversed = w1->w_reversed = 1;
1328 witness_increment_graph_generation();
1329 mtx_unlock_spin(&w_mtx);
1331 #ifdef WITNESS_NO_VNODE
1333 * There are known LORs between VNODE locks. They are
1334 * not an indication of a bug. VNODE locks are flagged
1335 * as such (LO_IS_VNODE) and we don't yell if the LOR
1336 * is between 2 VNODE locks.
1338 if ((lock->lo_flags & LO_IS_VNODE) != 0 &&
1339 (lock1->li_lock->lo_flags & LO_IS_VNODE) != 0)
1344 * Ok, yell about it.
1346 if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1347 (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
1349 "lock order reversal: (sleepable after non-sleepable)\n");
1350 else if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0
1351 && lock == &Giant.lock_object)
1353 "lock order reversal: (Giant after non-sleepable)\n");
1355 witness_output("lock order reversal:\n");
1358 * Try to locate an earlier lock with
1359 * witness w in our list.
1362 lock2 = &lle->ll_children[i];
1363 MPASS(lock2->li_lock != NULL);
1364 if (lock2->li_lock->lo_witness == w)
1366 if (i == 0 && lle->ll_next != NULL) {
1368 i = lle->ll_count - 1;
1369 MPASS(i >= 0 && i < LOCK_NCHILDREN);
1374 witness_output(" 1st %p %s (%s) @ %s:%d\n",
1375 lock1->li_lock, lock1->li_lock->lo_name,
1376 w1->w_name, fixup_filename(lock1->li_file),
1378 witness_output(" 2nd %p %s (%s) @ %s:%d\n", lock,
1379 lock->lo_name, w->w_name,
1380 fixup_filename(file), line);
1382 witness_output(" 1st %p %s (%s) @ %s:%d\n",
1383 lock2->li_lock, lock2->li_lock->lo_name,
1384 lock2->li_lock->lo_witness->w_name,
1385 fixup_filename(lock2->li_file),
1387 witness_output(" 2nd %p %s (%s) @ %s:%d\n",
1388 lock1->li_lock, lock1->li_lock->lo_name,
1389 w1->w_name, fixup_filename(lock1->li_file),
1391 witness_output(" 3rd %p %s (%s) @ %s:%d\n", lock,
1392 lock->lo_name, w->w_name,
1393 fixup_filename(file), line);
1395 witness_debugger(1, __func__);
1401 * If requested, build a new lock order. However, don't build a new
1402 * relationship between a sleepable lock and Giant if it is in the
1403 * wrong direction. The correct lock order is that sleepable locks
1404 * always come before Giant.
1406 if (flags & LOP_NEWORDER &&
1407 !(plock->li_lock == &Giant.lock_object &&
1408 (lock->lo_flags & LO_SLEEPABLE) != 0)) {
1409 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1410 w->w_name, plock->li_lock->lo_witness->w_name);
1411 itismychild(plock->li_lock->lo_witness, w);
1414 mtx_unlock_spin(&w_mtx);
1418 witness_lock(struct lock_object *lock, int flags, const char *file, int line)
1420 struct lock_list_entry **lock_list, *lle;
1421 struct lock_instance *instance;
1425 if (witness_cold || witness_watch == -1 || lock->lo_witness == NULL ||
1428 w = lock->lo_witness;
1431 /* Determine lock list for this lock. */
1432 if (LOCK_CLASS(lock)->lc_flags & LC_SLEEPLOCK)
1433 lock_list = &td->td_sleeplocks;
1435 lock_list = PCPU_PTR(spinlocks);
1437 /* Check to see if we are recursing on a lock we already own. */
1438 instance = find_instance(*lock_list, lock);
1439 if (instance != NULL) {
1440 instance->li_flags++;
1441 CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
1442 td->td_proc->p_pid, lock->lo_name,
1443 instance->li_flags & LI_RECURSEMASK);
1444 instance->li_file = file;
1445 instance->li_line = line;
1449 /* Update per-witness last file and line acquire. */
1453 /* Find the next open lock instance in the list and fill it. */
1455 if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
1456 lle = witness_lock_list_get();
1459 lle->ll_next = *lock_list;
1460 CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
1461 td->td_proc->p_pid, lle);
1464 instance = &lle->ll_children[lle->ll_count++];
1465 instance->li_lock = lock;
1466 instance->li_line = line;
1467 instance->li_file = file;
1468 if ((flags & LOP_EXCLUSIVE) != 0)
1469 instance->li_flags = LI_EXCLUSIVE;
1471 instance->li_flags = 0;
1472 CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
1473 td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
1477 witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
1479 struct lock_instance *instance;
1480 struct lock_class *class;
1482 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1483 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
1485 class = LOCK_CLASS(lock);
1486 if (witness_watch) {
1487 if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1489 "upgrade of non-upgradable lock (%s) %s @ %s:%d",
1490 class->lc_name, lock->lo_name,
1491 fixup_filename(file), line);
1492 if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1494 "upgrade of non-sleep lock (%s) %s @ %s:%d",
1495 class->lc_name, lock->lo_name,
1496 fixup_filename(file), line);
1498 instance = find_instance(curthread->td_sleeplocks, lock);
1499 if (instance == NULL) {
1500 kassert_panic("upgrade of unlocked lock (%s) %s @ %s:%d",
1501 class->lc_name, lock->lo_name,
1502 fixup_filename(file), line);
1505 if (witness_watch) {
1506 if ((instance->li_flags & LI_EXCLUSIVE) != 0)
1508 "upgrade of exclusive lock (%s) %s @ %s:%d",
1509 class->lc_name, lock->lo_name,
1510 fixup_filename(file), line);
1511 if ((instance->li_flags & LI_RECURSEMASK) != 0)
1513 "upgrade of recursed lock (%s) %s r=%d @ %s:%d",
1514 class->lc_name, lock->lo_name,
1515 instance->li_flags & LI_RECURSEMASK,
1516 fixup_filename(file), line);
1518 instance->li_flags |= LI_EXCLUSIVE;
1522 witness_downgrade(struct lock_object *lock, int flags, const char *file,
1525 struct lock_instance *instance;
1526 struct lock_class *class;
1528 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1529 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
1531 class = LOCK_CLASS(lock);
1532 if (witness_watch) {
1533 if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1535 "downgrade of non-upgradable lock (%s) %s @ %s:%d",
1536 class->lc_name, lock->lo_name,
1537 fixup_filename(file), line);
1538 if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1540 "downgrade of non-sleep lock (%s) %s @ %s:%d",
1541 class->lc_name, lock->lo_name,
1542 fixup_filename(file), line);
1544 instance = find_instance(curthread->td_sleeplocks, lock);
1545 if (instance == NULL) {
1546 kassert_panic("downgrade of unlocked lock (%s) %s @ %s:%d",
1547 class->lc_name, lock->lo_name,
1548 fixup_filename(file), line);
1551 if (witness_watch) {
1552 if ((instance->li_flags & LI_EXCLUSIVE) == 0)
1554 "downgrade of shared lock (%s) %s @ %s:%d",
1555 class->lc_name, lock->lo_name,
1556 fixup_filename(file), line);
1557 if ((instance->li_flags & LI_RECURSEMASK) != 0)
1559 "downgrade of recursed lock (%s) %s r=%d @ %s:%d",
1560 class->lc_name, lock->lo_name,
1561 instance->li_flags & LI_RECURSEMASK,
1562 fixup_filename(file), line);
1564 instance->li_flags &= ~LI_EXCLUSIVE;
1568 witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
1570 struct lock_list_entry **lock_list, *lle;
1571 struct lock_instance *instance;
1572 struct lock_class *class;
1577 if (witness_cold || lock->lo_witness == NULL || panicstr != NULL)
1580 class = LOCK_CLASS(lock);
1582 /* Find lock instance associated with this lock. */
1583 if (class->lc_flags & LC_SLEEPLOCK)
1584 lock_list = &td->td_sleeplocks;
1586 lock_list = PCPU_PTR(spinlocks);
1588 for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
1589 for (i = 0; i < (*lock_list)->ll_count; i++) {
1590 instance = &(*lock_list)->ll_children[i];
1591 if (instance->li_lock == lock)
1596 * When disabling WITNESS through witness_watch we could end up in
1597 * having registered locks in the td_sleeplocks queue.
1598 * We have to make sure we flush these queues, so just search for
1599 * eventual register locks and remove them.
1601 if (witness_watch > 0) {
1602 kassert_panic("lock (%s) %s not locked @ %s:%d", class->lc_name,
1603 lock->lo_name, fixup_filename(file), line);
1610 /* First, check for shared/exclusive mismatches. */
1611 if ((instance->li_flags & LI_EXCLUSIVE) != 0 && witness_watch > 0 &&
1612 (flags & LOP_EXCLUSIVE) == 0) {
1613 witness_output("shared unlock of (%s) %s @ %s:%d\n",
1614 class->lc_name, lock->lo_name, fixup_filename(file), line);
1615 witness_output("while exclusively locked from %s:%d\n",
1616 fixup_filename(instance->li_file), instance->li_line);
1617 kassert_panic("excl->ushare");
1619 if ((instance->li_flags & LI_EXCLUSIVE) == 0 && witness_watch > 0 &&
1620 (flags & LOP_EXCLUSIVE) != 0) {
1621 witness_output("exclusive unlock of (%s) %s @ %s:%d\n",
1622 class->lc_name, lock->lo_name, fixup_filename(file), line);
1623 witness_output("while share locked from %s:%d\n",
1624 fixup_filename(instance->li_file),
1626 kassert_panic("share->uexcl");
1628 /* If we are recursed, unrecurse. */
1629 if ((instance->li_flags & LI_RECURSEMASK) > 0) {
1630 CTR4(KTR_WITNESS, "%s: pid %d unrecursed on %s r=%d", __func__,
1631 td->td_proc->p_pid, instance->li_lock->lo_name,
1632 instance->li_flags);
1633 instance->li_flags--;
1636 /* The lock is now being dropped, check for NORELEASE flag */
1637 if ((instance->li_flags & LI_NORELEASE) != 0 && witness_watch > 0) {
1638 witness_output("forbidden unlock of (%s) %s @ %s:%d\n",
1639 class->lc_name, lock->lo_name, fixup_filename(file), line);
1640 kassert_panic("lock marked norelease");
1643 /* Otherwise, remove this item from the list. */
1645 CTR4(KTR_WITNESS, "%s: pid %d removed %s from lle[%d]", __func__,
1646 td->td_proc->p_pid, instance->li_lock->lo_name,
1647 (*lock_list)->ll_count - 1);
1648 for (j = i; j < (*lock_list)->ll_count - 1; j++)
1649 (*lock_list)->ll_children[j] =
1650 (*lock_list)->ll_children[j + 1];
1651 (*lock_list)->ll_count--;
1655 * In order to reduce contention on w_mtx, we want to keep always an
1656 * head object into lists so that frequent allocation from the
1657 * free witness pool (and subsequent locking) is avoided.
1658 * In order to maintain the current code simple, when the head
1659 * object is totally unloaded it means also that we do not have
1660 * further objects in the list, so the list ownership needs to be
1661 * hand over to another object if the current head needs to be freed.
1663 if ((*lock_list)->ll_count == 0) {
1664 if (*lock_list == lle) {
1665 if (lle->ll_next == NULL)
1669 *lock_list = lle->ll_next;
1670 CTR3(KTR_WITNESS, "%s: pid %d removed lle %p", __func__,
1671 td->td_proc->p_pid, lle);
1672 witness_lock_list_free(lle);
1677 witness_thread_exit(struct thread *td)
1679 struct lock_list_entry *lle;
1682 lle = td->td_sleeplocks;
1683 if (lle == NULL || panicstr != NULL)
1685 if (lle->ll_count != 0) {
1686 for (n = 0; lle != NULL; lle = lle->ll_next)
1687 for (i = lle->ll_count - 1; i >= 0; i--) {
1690 "Thread %p exiting with the following locks held:\n", td);
1692 witness_list_lock(&lle->ll_children[i],
1697 "Thread %p cannot exit while holding sleeplocks\n", td);
1699 witness_lock_list_free(lle);
1703 * Warn if any locks other than 'lock' are held. Flags can be passed in to
1704 * exempt Giant and sleepable locks from the checks as well. If any
1705 * non-exempt locks are held, then a supplied message is printed to the
1706 * output channel along with a list of the offending locks. If indicated in the
1707 * flags then a failure results in a panic as well.
1710 witness_warn(int flags, struct lock_object *lock, const char *fmt, ...)
1712 struct lock_list_entry *lock_list, *lle;
1713 struct lock_instance *lock1;
1718 if (witness_cold || witness_watch < 1 || panicstr != NULL)
1722 for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next)
1723 for (i = lle->ll_count - 1; i >= 0; i--) {
1724 lock1 = &lle->ll_children[i];
1725 if (lock1->li_lock == lock)
1727 if (flags & WARN_GIANTOK &&
1728 lock1->li_lock == &Giant.lock_object)
1730 if (flags & WARN_SLEEPOK &&
1731 (lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0)
1737 printf(" with the following %slocks held:\n",
1738 (flags & WARN_SLEEPOK) != 0 ?
1739 "non-sleepable " : "");
1742 witness_list_lock(lock1, printf);
1746 * Pin the thread in order to avoid problems with thread migration.
1747 * Once that all verifies are passed about spinlocks ownership,
1748 * the thread is in a safe path and it can be unpinned.
1751 lock_list = PCPU_GET(spinlocks);
1752 if (lock_list != NULL && lock_list->ll_count != 0) {
1756 * We should only have one spinlock and as long as
1757 * the flags cannot match for this locks class,
1758 * check if the first spinlock is the one curthread
1761 lock1 = &lock_list->ll_children[lock_list->ll_count - 1];
1762 if (lock_list->ll_count == 1 && lock_list->ll_next == NULL &&
1763 lock1->li_lock == lock && n == 0)
1769 printf(" with the following %slocks held:\n",
1770 (flags & WARN_SLEEPOK) != 0 ? "non-sleepable " : "");
1771 n += witness_list_locks(&lock_list, printf);
1774 if (flags & WARN_PANIC && n)
1775 kassert_panic("%s", __func__);
1777 witness_debugger(n, __func__);
1782 witness_file(struct lock_object *lock)
1786 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1788 w = lock->lo_witness;
1793 witness_line(struct lock_object *lock)
1797 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1799 w = lock->lo_witness;
1803 static struct witness *
1804 enroll(const char *description, struct lock_class *lock_class)
1807 struct witness_list *typelist;
1809 MPASS(description != NULL);
1811 if (witness_watch == -1 || panicstr != NULL)
1813 if ((lock_class->lc_flags & LC_SPINLOCK)) {
1814 if (witness_skipspin)
1818 } else if ((lock_class->lc_flags & LC_SLEEPLOCK)) {
1819 typelist = &w_sleep;
1821 kassert_panic("lock class %s is not sleep or spin",
1822 lock_class->lc_name);
1826 mtx_lock_spin(&w_mtx);
1827 w = witness_hash_get(description);
1830 if ((w = witness_get()) == NULL)
1832 MPASS(strlen(description) < MAX_W_NAME);
1833 strcpy(w->w_name, description);
1834 w->w_class = lock_class;
1836 STAILQ_INSERT_HEAD(&w_all, w, w_list);
1837 if (lock_class->lc_flags & LC_SPINLOCK) {
1838 STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
1840 } else if (lock_class->lc_flags & LC_SLEEPLOCK) {
1841 STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
1845 /* Insert new witness into the hash */
1846 witness_hash_put(w);
1847 witness_increment_graph_generation();
1848 mtx_unlock_spin(&w_mtx);
1852 mtx_unlock_spin(&w_mtx);
1853 if (lock_class != w->w_class)
1855 "lock (%s) %s does not match earlier (%s) lock",
1856 description, lock_class->lc_name,
1857 w->w_class->lc_name);
1862 depart(struct witness *w)
1864 struct witness_list *list;
1866 MPASS(w->w_refcount == 0);
1867 if (w->w_class->lc_flags & LC_SLEEPLOCK) {
1875 * Set file to NULL as it may point into a loadable module.
1879 witness_increment_graph_generation();
1884 adopt(struct witness *parent, struct witness *child)
1888 if (witness_cold == 0)
1889 mtx_assert(&w_mtx, MA_OWNED);
1891 /* If the relationship is already known, there's no work to be done. */
1892 if (isitmychild(parent, child))
1895 /* When the structure of the graph changes, bump up the generation. */
1896 witness_increment_graph_generation();
1899 * The hard part ... create the direct relationship, then propagate all
1900 * indirect relationships.
1902 pi = parent->w_index;
1903 ci = child->w_index;
1904 WITNESS_INDEX_ASSERT(pi);
1905 WITNESS_INDEX_ASSERT(ci);
1907 w_rmatrix[pi][ci] |= WITNESS_PARENT;
1908 w_rmatrix[ci][pi] |= WITNESS_CHILD;
1911 * If parent was not already an ancestor of child,
1912 * then we increment the descendant and ancestor counters.
1914 if ((w_rmatrix[pi][ci] & WITNESS_ANCESTOR) == 0) {
1915 parent->w_num_descendants++;
1916 child->w_num_ancestors++;
1920 * Find each ancestor of 'pi'. Note that 'pi' itself is counted as
1921 * an ancestor of 'pi' during this loop.
1923 for (i = 1; i <= w_max_used_index; i++) {
1924 if ((w_rmatrix[i][pi] & WITNESS_ANCESTOR_MASK) == 0 &&
1928 /* Find each descendant of 'i' and mark it as a descendant. */
1929 for (j = 1; j <= w_max_used_index; j++) {
1932 * Skip children that are already marked as
1933 * descendants of 'i'.
1935 if (w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK)
1939 * We are only interested in descendants of 'ci'. Note
1940 * that 'ci' itself is counted as a descendant of 'ci'.
1942 if ((w_rmatrix[ci][j] & WITNESS_ANCESTOR_MASK) == 0 &&
1945 w_rmatrix[i][j] |= WITNESS_ANCESTOR;
1946 w_rmatrix[j][i] |= WITNESS_DESCENDANT;
1947 w_data[i].w_num_descendants++;
1948 w_data[j].w_num_ancestors++;
1951 * Make sure we aren't marking a node as both an
1952 * ancestor and descendant. We should have caught
1953 * this as a lock order reversal earlier.
1955 if ((w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK) &&
1956 (w_rmatrix[i][j] & WITNESS_DESCENDANT_MASK)) {
1957 printf("witness rmatrix paradox! [%d][%d]=%d "
1958 "both ancestor and descendant\n",
1959 i, j, w_rmatrix[i][j]);
1961 printf("Witness disabled.\n");
1964 if ((w_rmatrix[j][i] & WITNESS_ANCESTOR_MASK) &&
1965 (w_rmatrix[j][i] & WITNESS_DESCENDANT_MASK)) {
1966 printf("witness rmatrix paradox! [%d][%d]=%d "
1967 "both ancestor and descendant\n",
1968 j, i, w_rmatrix[j][i]);
1970 printf("Witness disabled.\n");
1978 itismychild(struct witness *parent, struct witness *child)
1982 MPASS(child != NULL && parent != NULL);
1983 if (witness_cold == 0)
1984 mtx_assert(&w_mtx, MA_OWNED);
1986 if (!witness_lock_type_equal(parent, child)) {
1987 if (witness_cold == 0) {
1989 mtx_unlock_spin(&w_mtx);
1994 "%s: parent \"%s\" (%s) and child \"%s\" (%s) are not "
1995 "the same lock type", __func__, parent->w_name,
1996 parent->w_class->lc_name, child->w_name,
1997 child->w_class->lc_name);
1999 mtx_lock_spin(&w_mtx);
2001 adopt(parent, child);
2005 * Generic code for the isitmy*() functions. The rmask parameter is the
2006 * expected relationship of w1 to w2.
2009 _isitmyx(struct witness *w1, struct witness *w2, int rmask, const char *fname)
2011 unsigned char r1, r2;
2016 WITNESS_INDEX_ASSERT(i1);
2017 WITNESS_INDEX_ASSERT(i2);
2018 r1 = w_rmatrix[i1][i2] & WITNESS_RELATED_MASK;
2019 r2 = w_rmatrix[i2][i1] & WITNESS_RELATED_MASK;
2021 /* The flags on one better be the inverse of the flags on the other */
2022 if (!((WITNESS_ATOD(r1) == r2 && WITNESS_DTOA(r2) == r1) ||
2023 (WITNESS_DTOA(r1) == r2 && WITNESS_ATOD(r2) == r1))) {
2024 /* Don't squawk if we're potentially racing with an update. */
2025 if (!mtx_owned(&w_mtx))
2027 printf("%s: rmatrix mismatch between %s (index %d) and %s "
2028 "(index %d): w_rmatrix[%d][%d] == %hhx but "
2029 "w_rmatrix[%d][%d] == %hhx\n",
2030 fname, w1->w_name, i1, w2->w_name, i2, i1, i2, r1,
2033 printf("Witness disabled.\n");
2036 return (r1 & rmask);
2040 * Checks if @child is a direct child of @parent.
2043 isitmychild(struct witness *parent, struct witness *child)
2046 return (_isitmyx(parent, child, WITNESS_PARENT, __func__));
2050 * Checks if @descendant is a direct or inderect descendant of @ancestor.
2053 isitmydescendant(struct witness *ancestor, struct witness *descendant)
2056 return (_isitmyx(ancestor, descendant, WITNESS_ANCESTOR_MASK,
2062 blessed(struct witness *w1, struct witness *w2)
2065 struct witness_blessed *b;
2067 for (i = 0; i < nitems(blessed_list); i++) {
2068 b = &blessed_list[i];
2069 if (strcmp(w1->w_name, b->b_lock1) == 0) {
2070 if (strcmp(w2->w_name, b->b_lock2) == 0)
2074 if (strcmp(w1->w_name, b->b_lock2) == 0)
2075 if (strcmp(w2->w_name, b->b_lock1) == 0)
2082 static struct witness *
2088 if (witness_cold == 0)
2089 mtx_assert(&w_mtx, MA_OWNED);
2091 if (witness_watch == -1) {
2092 mtx_unlock_spin(&w_mtx);
2095 if (STAILQ_EMPTY(&w_free)) {
2097 mtx_unlock_spin(&w_mtx);
2098 printf("WITNESS: unable to allocate a new witness object\n");
2101 w = STAILQ_FIRST(&w_free);
2102 STAILQ_REMOVE_HEAD(&w_free, w_list);
2105 MPASS(index > 0 && index == w_max_used_index+1 &&
2106 index < witness_count);
2107 bzero(w, sizeof(*w));
2109 if (index > w_max_used_index)
2110 w_max_used_index = index;
2115 witness_free(struct witness *w)
2118 STAILQ_INSERT_HEAD(&w_free, w, w_list);
2122 static struct lock_list_entry *
2123 witness_lock_list_get(void)
2125 struct lock_list_entry *lle;
2127 if (witness_watch == -1)
2129 mtx_lock_spin(&w_mtx);
2130 lle = w_lock_list_free;
2133 mtx_unlock_spin(&w_mtx);
2134 printf("%s: witness exhausted\n", __func__);
2137 w_lock_list_free = lle->ll_next;
2138 mtx_unlock_spin(&w_mtx);
2139 bzero(lle, sizeof(*lle));
2144 witness_lock_list_free(struct lock_list_entry *lle)
2147 mtx_lock_spin(&w_mtx);
2148 lle->ll_next = w_lock_list_free;
2149 w_lock_list_free = lle;
2150 mtx_unlock_spin(&w_mtx);
2153 static struct lock_instance *
2154 find_instance(struct lock_list_entry *list, const struct lock_object *lock)
2156 struct lock_list_entry *lle;
2157 struct lock_instance *instance;
2160 for (lle = list; lle != NULL; lle = lle->ll_next)
2161 for (i = lle->ll_count - 1; i >= 0; i--) {
2162 instance = &lle->ll_children[i];
2163 if (instance->li_lock == lock)
2170 witness_list_lock(struct lock_instance *instance,
2171 int (*prnt)(const char *fmt, ...))
2173 struct lock_object *lock;
2175 lock = instance->li_lock;
2176 prnt("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ?
2177 "exclusive" : "shared", LOCK_CLASS(lock)->lc_name, lock->lo_name);
2178 if (lock->lo_witness->w_name != lock->lo_name)
2179 prnt(" (%s)", lock->lo_witness->w_name);
2180 prnt(" r = %d (%p) locked @ %s:%d\n",
2181 instance->li_flags & LI_RECURSEMASK, lock,
2182 fixup_filename(instance->li_file), instance->li_line);
2186 witness_output(const char *fmt, ...)
2192 ret = witness_voutput(fmt, ap);
2198 witness_voutput(const char *fmt, va_list ap)
2203 switch (witness_channel) {
2204 case WITNESS_CONSOLE:
2205 ret = vprintf(fmt, ap);
2208 vlog(LOG_NOTICE, fmt, ap);
2218 witness_thread_has_locks(struct thread *td)
2221 if (td->td_sleeplocks == NULL)
2223 return (td->td_sleeplocks->ll_count != 0);
2227 witness_proc_has_locks(struct proc *p)
2231 FOREACH_THREAD_IN_PROC(p, td) {
2232 if (witness_thread_has_locks(td))
2240 witness_list_locks(struct lock_list_entry **lock_list,
2241 int (*prnt)(const char *fmt, ...))
2243 struct lock_list_entry *lle;
2247 for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
2248 for (i = lle->ll_count - 1; i >= 0; i--) {
2249 witness_list_lock(&lle->ll_children[i], prnt);
2256 * This is a bit risky at best. We call this function when we have timed
2257 * out acquiring a spin lock, and we assume that the other CPU is stuck
2258 * with this lock held. So, we go groveling around in the other CPU's
2259 * per-cpu data to try to find the lock instance for this spin lock to
2260 * see when it was last acquired.
2263 witness_display_spinlock(struct lock_object *lock, struct thread *owner,
2264 int (*prnt)(const char *fmt, ...))
2266 struct lock_instance *instance;
2269 if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU)
2271 pc = pcpu_find(owner->td_oncpu);
2272 instance = find_instance(pc->pc_spinlocks, lock);
2273 if (instance != NULL)
2274 witness_list_lock(instance, prnt);
2278 witness_save(struct lock_object *lock, const char **filep, int *linep)
2280 struct lock_list_entry *lock_list;
2281 struct lock_instance *instance;
2282 struct lock_class *class;
2285 * This function is used independently in locking code to deal with
2286 * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2289 if (SCHEDULER_STOPPED())
2291 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2292 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2294 class = LOCK_CLASS(lock);
2295 if (class->lc_flags & LC_SLEEPLOCK)
2296 lock_list = curthread->td_sleeplocks;
2298 if (witness_skipspin)
2300 lock_list = PCPU_GET(spinlocks);
2302 instance = find_instance(lock_list, lock);
2303 if (instance == NULL) {
2304 kassert_panic("%s: lock (%s) %s not locked", __func__,
2305 class->lc_name, lock->lo_name);
2308 *filep = instance->li_file;
2309 *linep = instance->li_line;
2313 witness_restore(struct lock_object *lock, const char *file, int line)
2315 struct lock_list_entry *lock_list;
2316 struct lock_instance *instance;
2317 struct lock_class *class;
2320 * This function is used independently in locking code to deal with
2321 * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2324 if (SCHEDULER_STOPPED())
2326 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2327 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2329 class = LOCK_CLASS(lock);
2330 if (class->lc_flags & LC_SLEEPLOCK)
2331 lock_list = curthread->td_sleeplocks;
2333 if (witness_skipspin)
2335 lock_list = PCPU_GET(spinlocks);
2337 instance = find_instance(lock_list, lock);
2338 if (instance == NULL)
2339 kassert_panic("%s: lock (%s) %s not locked", __func__,
2340 class->lc_name, lock->lo_name);
2341 lock->lo_witness->w_file = file;
2342 lock->lo_witness->w_line = line;
2343 if (instance == NULL)
2345 instance->li_file = file;
2346 instance->li_line = line;
2350 witness_assert(const struct lock_object *lock, int flags, const char *file,
2353 #ifdef INVARIANT_SUPPORT
2354 struct lock_instance *instance;
2355 struct lock_class *class;
2357 if (lock->lo_witness == NULL || witness_watch < 1 || panicstr != NULL)
2359 class = LOCK_CLASS(lock);
2360 if ((class->lc_flags & LC_SLEEPLOCK) != 0)
2361 instance = find_instance(curthread->td_sleeplocks, lock);
2362 else if ((class->lc_flags & LC_SPINLOCK) != 0)
2363 instance = find_instance(PCPU_GET(spinlocks), lock);
2365 kassert_panic("Lock (%s) %s is not sleep or spin!",
2366 class->lc_name, lock->lo_name);
2371 if (instance != NULL)
2372 kassert_panic("Lock (%s) %s locked @ %s:%d.",
2373 class->lc_name, lock->lo_name,
2374 fixup_filename(file), line);
2377 case LA_LOCKED | LA_RECURSED:
2378 case LA_LOCKED | LA_NOTRECURSED:
2380 case LA_SLOCKED | LA_RECURSED:
2381 case LA_SLOCKED | LA_NOTRECURSED:
2383 case LA_XLOCKED | LA_RECURSED:
2384 case LA_XLOCKED | LA_NOTRECURSED:
2385 if (instance == NULL) {
2386 kassert_panic("Lock (%s) %s not locked @ %s:%d.",
2387 class->lc_name, lock->lo_name,
2388 fixup_filename(file), line);
2391 if ((flags & LA_XLOCKED) != 0 &&
2392 (instance->li_flags & LI_EXCLUSIVE) == 0)
2394 "Lock (%s) %s not exclusively locked @ %s:%d.",
2395 class->lc_name, lock->lo_name,
2396 fixup_filename(file), line);
2397 if ((flags & LA_SLOCKED) != 0 &&
2398 (instance->li_flags & LI_EXCLUSIVE) != 0)
2400 "Lock (%s) %s exclusively locked @ %s:%d.",
2401 class->lc_name, lock->lo_name,
2402 fixup_filename(file), line);
2403 if ((flags & LA_RECURSED) != 0 &&
2404 (instance->li_flags & LI_RECURSEMASK) == 0)
2405 kassert_panic("Lock (%s) %s not recursed @ %s:%d.",
2406 class->lc_name, lock->lo_name,
2407 fixup_filename(file), line);
2408 if ((flags & LA_NOTRECURSED) != 0 &&
2409 (instance->li_flags & LI_RECURSEMASK) != 0)
2410 kassert_panic("Lock (%s) %s recursed @ %s:%d.",
2411 class->lc_name, lock->lo_name,
2412 fixup_filename(file), line);
2415 kassert_panic("Invalid lock assertion at %s:%d.",
2416 fixup_filename(file), line);
2419 #endif /* INVARIANT_SUPPORT */
2423 witness_setflag(struct lock_object *lock, int flag, int set)
2425 struct lock_list_entry *lock_list;
2426 struct lock_instance *instance;
2427 struct lock_class *class;
2429 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2431 class = LOCK_CLASS(lock);
2432 if (class->lc_flags & LC_SLEEPLOCK)
2433 lock_list = curthread->td_sleeplocks;
2435 if (witness_skipspin)
2437 lock_list = PCPU_GET(spinlocks);
2439 instance = find_instance(lock_list, lock);
2440 if (instance == NULL) {
2441 kassert_panic("%s: lock (%s) %s not locked", __func__,
2442 class->lc_name, lock->lo_name);
2447 instance->li_flags |= flag;
2449 instance->li_flags &= ~flag;
2453 witness_norelease(struct lock_object *lock)
2456 witness_setflag(lock, LI_NORELEASE, 1);
2460 witness_releaseok(struct lock_object *lock)
2463 witness_setflag(lock, LI_NORELEASE, 0);
2468 witness_ddb_list(struct thread *td)
2471 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2472 KASSERT(kdb_active, ("%s: not in the debugger", __func__));
2474 if (witness_watch < 1)
2477 witness_list_locks(&td->td_sleeplocks, db_printf);
2480 * We only handle spinlocks if td == curthread. This is somewhat broken
2481 * if td is currently executing on some other CPU and holds spin locks
2482 * as we won't display those locks. If we had a MI way of getting
2483 * the per-cpu data for a given cpu then we could use
2484 * td->td_oncpu to get the list of spinlocks for this thread
2487 * That still wouldn't really fix this unless we locked the scheduler
2488 * lock or stopped the other CPU to make sure it wasn't changing the
2489 * list out from under us. It is probably best to just not try to
2490 * handle threads on other CPU's for now.
2492 if (td == curthread && PCPU_GET(spinlocks) != NULL)
2493 witness_list_locks(PCPU_PTR(spinlocks), db_printf);
2496 DB_SHOW_COMMAND(locks, db_witness_list)
2501 td = db_lookup_thread(addr, true);
2504 witness_ddb_list(td);
2507 DB_SHOW_ALL_COMMAND(locks, db_witness_list_all)
2513 * It would be nice to list only threads and processes that actually
2514 * held sleep locks, but that information is currently not exported
2517 FOREACH_PROC_IN_SYSTEM(p) {
2518 if (!witness_proc_has_locks(p))
2520 FOREACH_THREAD_IN_PROC(p, td) {
2521 if (!witness_thread_has_locks(td))
2523 db_printf("Process %d (%s) thread %p (%d)\n", p->p_pid,
2524 p->p_comm, td, td->td_tid);
2525 witness_ddb_list(td);
2531 DB_SHOW_ALIAS(alllocks, db_witness_list_all)
2533 DB_SHOW_COMMAND(witness, db_witness_display)
2536 witness_ddb_display(db_printf);
2541 sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS)
2543 struct witness_lock_order_data *data1, *data2, *tmp_data1, *tmp_data2;
2544 struct witness *tmp_w1, *tmp_w2, *w1, *w2;
2546 u_int w_rmatrix1, w_rmatrix2;
2547 int error, generation, i, j;
2553 if (witness_watch < 1) {
2554 error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2558 error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2562 sb = sbuf_new(NULL, NULL, badstack_sbuf_size, SBUF_AUTOEXTEND);
2566 /* Allocate and init temporary storage space. */
2567 tmp_w1 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2568 tmp_w2 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2569 tmp_data1 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2571 tmp_data2 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2573 stack_zero(&tmp_data1->wlod_stack);
2574 stack_zero(&tmp_data2->wlod_stack);
2577 mtx_lock_spin(&w_mtx);
2578 generation = w_generation;
2579 mtx_unlock_spin(&w_mtx);
2580 sbuf_printf(sb, "Number of known direct relationships is %d\n",
2581 w_lohash.wloh_count);
2582 for (i = 1; i < w_max_used_index; i++) {
2583 mtx_lock_spin(&w_mtx);
2584 if (generation != w_generation) {
2585 mtx_unlock_spin(&w_mtx);
2587 /* The graph has changed, try again. */
2594 if (w1->w_reversed == 0) {
2595 mtx_unlock_spin(&w_mtx);
2599 /* Copy w1 locally so we can release the spin lock. */
2601 mtx_unlock_spin(&w_mtx);
2603 if (tmp_w1->w_reversed == 0)
2605 for (j = 1; j < w_max_used_index; j++) {
2606 if ((w_rmatrix[i][j] & WITNESS_REVERSAL) == 0 || i > j)
2609 mtx_lock_spin(&w_mtx);
2610 if (generation != w_generation) {
2611 mtx_unlock_spin(&w_mtx);
2613 /* The graph has changed, try again. */
2620 data1 = witness_lock_order_get(w1, w2);
2621 data2 = witness_lock_order_get(w2, w1);
2624 * Copy information locally so we can release the
2628 w_rmatrix1 = (unsigned int)w_rmatrix[i][j];
2629 w_rmatrix2 = (unsigned int)w_rmatrix[j][i];
2632 stack_zero(&tmp_data1->wlod_stack);
2633 stack_copy(&data1->wlod_stack,
2634 &tmp_data1->wlod_stack);
2636 if (data2 && data2 != data1) {
2637 stack_zero(&tmp_data2->wlod_stack);
2638 stack_copy(&data2->wlod_stack,
2639 &tmp_data2->wlod_stack);
2641 mtx_unlock_spin(&w_mtx);
2644 "\nLock order reversal between \"%s\"(%s) and \"%s\"(%s)!\n",
2645 tmp_w1->w_name, tmp_w1->w_class->lc_name,
2646 tmp_w2->w_name, tmp_w2->w_class->lc_name);
2649 "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2650 tmp_w1->w_name, tmp_w1->w_class->lc_name,
2651 tmp_w2->w_name, tmp_w2->w_class->lc_name);
2652 stack_sbuf_print(sb, &tmp_data1->wlod_stack);
2653 sbuf_printf(sb, "\n");
2655 if (data2 && data2 != data1) {
2657 "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2658 tmp_w2->w_name, tmp_w2->w_class->lc_name,
2659 tmp_w1->w_name, tmp_w1->w_class->lc_name);
2660 stack_sbuf_print(sb, &tmp_data2->wlod_stack);
2661 sbuf_printf(sb, "\n");
2665 mtx_lock_spin(&w_mtx);
2666 if (generation != w_generation) {
2667 mtx_unlock_spin(&w_mtx);
2670 * The graph changed while we were printing stack data,
2677 mtx_unlock_spin(&w_mtx);
2679 /* Free temporary storage space. */
2680 free(tmp_data1, M_TEMP);
2681 free(tmp_data2, M_TEMP);
2682 free(tmp_w1, M_TEMP);
2683 free(tmp_w2, M_TEMP);
2686 error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1);
2693 sysctl_debug_witness_channel(SYSCTL_HANDLER_ARGS)
2695 static const struct {
2696 enum witness_channel channel;
2699 { WITNESS_CONSOLE, "console" },
2700 { WITNESS_LOG, "log" },
2701 { WITNESS_NONE, "none" },
2708 for (i = 0; i < nitems(channels); i++)
2709 if (witness_channel == channels[i].channel) {
2710 snprintf(buf, sizeof(buf), "%s", channels[i].name);
2714 error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
2715 if (error != 0 || req->newptr == NULL)
2719 for (i = 0; i < nitems(channels); i++)
2720 if (strcmp(channels[i].name, buf) == 0) {
2721 witness_channel = channels[i].channel;
2729 sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS)
2735 if (witness_watch < 1) {
2736 error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2740 error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2745 error = sysctl_wire_old_buffer(req, 0);
2748 sb = sbuf_new_for_sysctl(NULL, NULL, FULLGRAPH_SBUF_SIZE, req);
2751 sbuf_printf(sb, "\n");
2753 mtx_lock_spin(&w_mtx);
2754 STAILQ_FOREACH(w, &w_all, w_list)
2756 STAILQ_FOREACH(w, &w_all, w_list)
2757 witness_add_fullgraph(sb, w);
2758 mtx_unlock_spin(&w_mtx);
2761 * Close the sbuf and return to userland.
2763 error = sbuf_finish(sb);
2770 sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS)
2774 value = witness_watch;
2775 error = sysctl_handle_int(oidp, &value, 0, req);
2776 if (error != 0 || req->newptr == NULL)
2778 if (value > 1 || value < -1 ||
2779 (witness_watch == -1 && value != witness_watch))
2781 witness_watch = value;
2786 witness_add_fullgraph(struct sbuf *sb, struct witness *w)
2790 if (w->w_displayed != 0 || (w->w_file == NULL && w->w_line == 0))
2794 WITNESS_INDEX_ASSERT(w->w_index);
2795 for (i = 1; i <= w_max_used_index; i++) {
2796 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT) {
2797 sbuf_printf(sb, "\"%s\",\"%s\"\n", w->w_name,
2799 witness_add_fullgraph(sb, &w_data[i]);
2805 * A simple hash function. Takes a key pointer and a key size. If size == 0,
2806 * interprets the key as a string and reads until the null
2807 * terminator. Otherwise, reads the first size bytes. Returns an unsigned 32-bit
2808 * hash value computed from the key.
2811 witness_hash_djb2(const uint8_t *key, uint32_t size)
2813 unsigned int hash = 5381;
2816 /* hash = hash * 33 + key[i] */
2818 for (i = 0; i < size; i++)
2819 hash = ((hash << 5) + hash) + (unsigned int)key[i];
2821 for (i = 0; key[i] != 0; i++)
2822 hash = ((hash << 5) + hash) + (unsigned int)key[i];
2829 * Initializes the two witness hash tables. Called exactly once from
2830 * witness_initialize().
2833 witness_init_hash_tables(void)
2837 MPASS(witness_cold);
2839 /* Initialize the hash tables. */
2840 for (i = 0; i < WITNESS_HASH_SIZE; i++)
2841 w_hash.wh_array[i] = NULL;
2843 w_hash.wh_size = WITNESS_HASH_SIZE;
2844 w_hash.wh_count = 0;
2846 /* Initialize the lock order data hash. */
2848 for (i = 0; i < WITNESS_LO_DATA_COUNT; i++) {
2849 memset(&w_lodata[i], 0, sizeof(w_lodata[i]));
2850 w_lodata[i].wlod_next = w_lofree;
2851 w_lofree = &w_lodata[i];
2853 w_lohash.wloh_size = WITNESS_LO_HASH_SIZE;
2854 w_lohash.wloh_count = 0;
2855 for (i = 0; i < WITNESS_LO_HASH_SIZE; i++)
2856 w_lohash.wloh_array[i] = NULL;
2859 static struct witness *
2860 witness_hash_get(const char *key)
2866 if (witness_cold == 0)
2867 mtx_assert(&w_mtx, MA_OWNED);
2868 hash = witness_hash_djb2(key, 0) % w_hash.wh_size;
2869 w = w_hash.wh_array[hash];
2871 if (strcmp(w->w_name, key) == 0)
2881 witness_hash_put(struct witness *w)
2886 MPASS(w->w_name != NULL);
2887 if (witness_cold == 0)
2888 mtx_assert(&w_mtx, MA_OWNED);
2889 KASSERT(witness_hash_get(w->w_name) == NULL,
2890 ("%s: trying to add a hash entry that already exists!", __func__));
2891 KASSERT(w->w_hash_next == NULL,
2892 ("%s: w->w_hash_next != NULL", __func__));
2894 hash = witness_hash_djb2(w->w_name, 0) % w_hash.wh_size;
2895 w->w_hash_next = w_hash.wh_array[hash];
2896 w_hash.wh_array[hash] = w;
2901 static struct witness_lock_order_data *
2902 witness_lock_order_get(struct witness *parent, struct witness *child)
2904 struct witness_lock_order_data *data = NULL;
2905 struct witness_lock_order_key key;
2908 MPASS(parent != NULL && child != NULL);
2909 key.from = parent->w_index;
2910 key.to = child->w_index;
2911 WITNESS_INDEX_ASSERT(key.from);
2912 WITNESS_INDEX_ASSERT(key.to);
2913 if ((w_rmatrix[parent->w_index][child->w_index]
2914 & WITNESS_LOCK_ORDER_KNOWN) == 0)
2917 hash = witness_hash_djb2((const char*)&key,
2918 sizeof(key)) % w_lohash.wloh_size;
2919 data = w_lohash.wloh_array[hash];
2920 while (data != NULL) {
2921 if (witness_lock_order_key_equal(&data->wlod_key, &key))
2923 data = data->wlod_next;
2931 * Verify that parent and child have a known relationship, are not the same,
2932 * and child is actually a child of parent. This is done without w_mtx
2933 * to avoid contention in the common case.
2936 witness_lock_order_check(struct witness *parent, struct witness *child)
2939 if (parent != child &&
2940 w_rmatrix[parent->w_index][child->w_index]
2941 & WITNESS_LOCK_ORDER_KNOWN &&
2942 isitmychild(parent, child))
2949 witness_lock_order_add(struct witness *parent, struct witness *child)
2951 struct witness_lock_order_data *data = NULL;
2952 struct witness_lock_order_key key;
2955 MPASS(parent != NULL && child != NULL);
2956 key.from = parent->w_index;
2957 key.to = child->w_index;
2958 WITNESS_INDEX_ASSERT(key.from);
2959 WITNESS_INDEX_ASSERT(key.to);
2960 if (w_rmatrix[parent->w_index][child->w_index]
2961 & WITNESS_LOCK_ORDER_KNOWN)
2964 hash = witness_hash_djb2((const char*)&key,
2965 sizeof(key)) % w_lohash.wloh_size;
2966 w_rmatrix[parent->w_index][child->w_index] |= WITNESS_LOCK_ORDER_KNOWN;
2970 w_lofree = data->wlod_next;
2971 data->wlod_next = w_lohash.wloh_array[hash];
2972 data->wlod_key = key;
2973 w_lohash.wloh_array[hash] = data;
2974 w_lohash.wloh_count++;
2975 stack_zero(&data->wlod_stack);
2976 stack_save(&data->wlod_stack);
2980 /* Call this whenever the structure of the witness graph changes. */
2982 witness_increment_graph_generation(void)
2985 if (witness_cold == 0)
2986 mtx_assert(&w_mtx, MA_OWNED);
2991 witness_output_drain(void *arg __unused, const char *data, int len)
2994 witness_output("%.*s", len, data);
2999 witness_debugger(int cond, const char *msg)
3008 if (witness_trace) {
3009 sbuf_new(&sb, buf, sizeof(buf), SBUF_FIXEDLEN);
3010 sbuf_set_drain(&sb, witness_output_drain, NULL);
3014 witness_output("stack backtrace:\n");
3015 stack_sbuf_print_ddb(&sb, &st);
3022 kdb_enter(KDB_WHY_WITNESS, msg);