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/systm.h>
115 #include <machine/stdarg.h>
117 #if !defined(DDB) && !defined(STACK)
118 #error "DDB or STACK options are required for WITNESS"
121 /* Note that these traces do not work with KTR_ALQ. */
123 #define KTR_WITNESS KTR_SUBSYS
125 #define KTR_WITNESS 0
128 #define LI_RECURSEMASK 0x0000ffff /* Recursion depth of lock instance. */
129 #define LI_EXCLUSIVE 0x00010000 /* Exclusive lock instance. */
130 #define LI_NORELEASE 0x00020000 /* Lock not allowed to be released. */
132 /* Define this to check for blessed mutexes */
135 #define WITNESS_COUNT 1024
136 #define WITNESS_CHILDCOUNT (WITNESS_COUNT * 4)
137 #define WITNESS_HASH_SIZE 251 /* Prime, gives load factor < 2 */
138 #define WITNESS_PENDLIST 768
140 /* Allocate 256 KB of stack data space */
141 #define WITNESS_LO_DATA_COUNT 2048
143 /* Prime, gives load factor of ~2 at full load */
144 #define WITNESS_LO_HASH_SIZE 1021
147 * XXX: This is somewhat bogus, as we assume here that at most 2048 threads
148 * will hold LOCK_NCHILDREN locks. We handle failure ok, and we should
149 * probably be safe for the most part, but it's still a SWAG.
151 #define LOCK_NCHILDREN 5
152 #define LOCK_CHILDCOUNT 2048
154 #define MAX_W_NAME 64
156 #define BADSTACK_SBUF_SIZE (256 * WITNESS_COUNT)
157 #define CYCLEGRAPH_SBUF_SIZE 8192
158 #define FULLGRAPH_SBUF_SIZE 32768
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 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_empty(const struct witness_lock_order_key *key)
312 return (key->from == 0 && key->to == 0);
316 witness_lock_order_key_equal(const struct witness_lock_order_key *a,
317 const struct witness_lock_order_key *b)
320 return (a->from == b->from && a->to == b->to);
323 static int _isitmyx(struct witness *w1, struct witness *w2, int rmask,
326 static void _witness_debugger(int cond, const char *msg);
328 static void adopt(struct witness *parent, struct witness *child);
330 static int blessed(struct witness *, struct witness *);
332 static void depart(struct witness *w);
333 static struct witness *enroll(const char *description,
334 struct lock_class *lock_class);
335 static struct lock_instance *find_instance(struct lock_list_entry *list,
336 struct lock_object *lock);
337 static int isitmychild(struct witness *parent, struct witness *child);
338 static int isitmydescendant(struct witness *parent, struct witness *child);
339 static void itismychild(struct witness *parent, struct witness *child);
340 static int sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS);
341 static int sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS);
342 static int sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS);
343 static void witness_add_fullgraph(struct sbuf *sb, struct witness *parent);
345 static void witness_ddb_compute_levels(void);
346 static void witness_ddb_display(int(*)(const char *fmt, ...));
347 static void witness_ddb_display_descendants(int(*)(const char *fmt, ...),
348 struct witness *, int indent);
349 static void witness_ddb_display_list(int(*prnt)(const char *fmt, ...),
350 struct witness_list *list);
351 static void witness_ddb_level_descendants(struct witness *parent, int l);
352 static void witness_ddb_list(struct thread *td);
354 static void witness_free(struct witness *m);
355 static struct witness *witness_get(void);
356 static uint32_t witness_hash_djb2(const uint8_t *key, uint32_t size);
357 static struct witness *witness_hash_get(const char *key);
358 static void witness_hash_put(struct witness *w);
359 static void witness_init_hash_tables(void);
360 static void witness_increment_graph_generation(void);
361 static void witness_lock_list_free(struct lock_list_entry *lle);
362 static struct lock_list_entry *witness_lock_list_get(void);
363 static int witness_lock_order_add(struct witness *parent,
364 struct witness *child);
365 static int witness_lock_order_check(struct witness *parent,
366 struct witness *child);
367 static struct witness_lock_order_data *witness_lock_order_get(
368 struct witness *parent,
369 struct witness *child);
370 static void witness_list_lock(struct lock_instance *instance,
371 int (*prnt)(const char *fmt, ...));
372 static void witness_setflag(struct lock_object *lock, int flag, int set);
375 #define witness_debugger(c) _witness_debugger(c, __func__)
377 #define witness_debugger(c)
380 SYSCTL_NODE(_debug, OID_AUTO, witness, CTLFLAG_RW, NULL, "Witness Locking");
383 * If set to 0, lock order checking is disabled. If set to -1,
384 * witness is completely disabled. Otherwise witness performs full
385 * lock order checking for all locks. At runtime, lock order checking
386 * may be toggled. However, witness cannot be reenabled once it is
387 * completely disabled.
389 static int witness_watch = 1;
390 TUNABLE_INT("debug.witness.watch", &witness_watch);
391 SYSCTL_PROC(_debug_witness, OID_AUTO, watch, CTLFLAG_RW | CTLTYPE_INT, NULL, 0,
392 sysctl_debug_witness_watch, "I", "witness is watching lock operations");
396 * When KDB is enabled and witness_kdb is 1, it will cause the system
397 * to drop into kdebug() when:
398 * - a lock hierarchy violation occurs
399 * - locks are held when going to sleep.
406 TUNABLE_INT("debug.witness.kdb", &witness_kdb);
407 SYSCTL_INT(_debug_witness, OID_AUTO, kdb, CTLFLAG_RW, &witness_kdb, 0, "");
410 * When KDB is enabled and witness_trace is 1, it will cause the system
411 * to print a stack trace:
412 * - a lock hierarchy violation occurs
413 * - locks are held when going to sleep.
415 int witness_trace = 1;
416 TUNABLE_INT("debug.witness.trace", &witness_trace);
417 SYSCTL_INT(_debug_witness, OID_AUTO, trace, CTLFLAG_RW, &witness_trace, 0, "");
420 #ifdef WITNESS_SKIPSPIN
421 int witness_skipspin = 1;
423 int witness_skipspin = 0;
425 TUNABLE_INT("debug.witness.skipspin", &witness_skipspin);
426 SYSCTL_INT(_debug_witness, OID_AUTO, skipspin, CTLFLAG_RDTUN, &witness_skipspin,
430 * Call this to print out the relations between locks.
432 SYSCTL_PROC(_debug_witness, OID_AUTO, fullgraph, CTLTYPE_STRING | CTLFLAG_RD,
433 NULL, 0, sysctl_debug_witness_fullgraph, "A", "Show locks relation graphs");
436 * Call this to print out the witness faulty stacks.
438 SYSCTL_PROC(_debug_witness, OID_AUTO, badstacks, CTLTYPE_STRING | CTLFLAG_RD,
439 NULL, 0, sysctl_debug_witness_badstacks, "A", "Show bad witness stacks");
441 static struct mtx w_mtx;
444 static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free);
445 static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all);
448 static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin);
449 static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep);
452 static struct lock_list_entry *w_lock_list_free = NULL;
453 static struct witness_pendhelp pending_locks[WITNESS_PENDLIST];
454 static u_int pending_cnt;
456 static int w_free_cnt, w_spin_cnt, w_sleep_cnt;
457 SYSCTL_INT(_debug_witness, OID_AUTO, free_cnt, CTLFLAG_RD, &w_free_cnt, 0, "");
458 SYSCTL_INT(_debug_witness, OID_AUTO, spin_cnt, CTLFLAG_RD, &w_spin_cnt, 0, "");
459 SYSCTL_INT(_debug_witness, OID_AUTO, sleep_cnt, CTLFLAG_RD, &w_sleep_cnt, 0,
462 static struct witness *w_data;
463 static uint8_t w_rmatrix[WITNESS_COUNT+1][WITNESS_COUNT+1];
464 static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT];
465 static struct witness_hash w_hash; /* The witness hash table. */
467 /* The lock order data hash */
468 static struct witness_lock_order_data w_lodata[WITNESS_LO_DATA_COUNT];
469 static struct witness_lock_order_data *w_lofree = NULL;
470 static struct witness_lock_order_hash w_lohash;
471 static int w_max_used_index = 0;
472 static unsigned int w_generation = 0;
473 static const char w_notrunning[] = "Witness not running\n";
474 static const char w_stillcold[] = "Witness is still cold\n";
477 static struct witness_order_list_entry order_lists[] = {
481 { "proctree", &lock_class_sx },
482 { "allproc", &lock_class_sx },
483 { "allprison", &lock_class_sx },
488 { "Giant", &lock_class_mtx_sleep },
489 { "pipe mutex", &lock_class_mtx_sleep },
490 { "sigio lock", &lock_class_mtx_sleep },
491 { "process group", &lock_class_mtx_sleep },
492 { "process lock", &lock_class_mtx_sleep },
493 { "session", &lock_class_mtx_sleep },
494 { "uidinfo hash", &lock_class_rw },
496 { "pmc-sleep", &lock_class_mtx_sleep },
502 { "accept", &lock_class_mtx_sleep },
503 { "so_snd", &lock_class_mtx_sleep },
504 { "so_rcv", &lock_class_mtx_sleep },
505 { "sellck", &lock_class_mtx_sleep },
510 { "so_rcv", &lock_class_mtx_sleep },
511 { "radix node head", &lock_class_rw },
512 { "rtentry", &lock_class_mtx_sleep },
513 { "ifaddr", &lock_class_mtx_sleep },
517 * protocol locks before interface locks, after UDP locks.
519 { "udpinp", &lock_class_rw },
520 { "in_multi_mtx", &lock_class_mtx_sleep },
521 { "igmp_mtx", &lock_class_mtx_sleep },
522 { "if_addr_mtx", &lock_class_mtx_sleep },
526 * protocol locks before interface locks, after UDP locks.
528 { "udpinp", &lock_class_rw },
529 { "in6_multi_mtx", &lock_class_mtx_sleep },
530 { "mld_mtx", &lock_class_mtx_sleep },
531 { "if_addr_mtx", &lock_class_mtx_sleep },
534 * UNIX Domain Sockets
536 { "unp_global_rwlock", &lock_class_rw },
537 { "unp_list_lock", &lock_class_mtx_sleep },
538 { "unp", &lock_class_mtx_sleep },
539 { "so_snd", &lock_class_mtx_sleep },
544 { "udp", &lock_class_rw },
545 { "udpinp", &lock_class_rw },
546 { "so_snd", &lock_class_mtx_sleep },
551 { "tcp", &lock_class_rw },
552 { "tcpinp", &lock_class_rw },
553 { "so_snd", &lock_class_mtx_sleep },
558 { "slip_mtx", &lock_class_mtx_sleep },
559 { "slip sc_mtx", &lock_class_mtx_sleep },
564 { "ddp_list_mtx", &lock_class_mtx_sleep },
565 { "ddp_mtx", &lock_class_mtx_sleep },
570 { "bpf global lock", &lock_class_mtx_sleep },
571 { "bpf interface lock", &lock_class_mtx_sleep },
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 { "system map", &lock_class_mtx_sleep },
602 { "vm page queue mutex", &lock_class_mtx_sleep },
603 { "vnode interlock", &lock_class_mtx_sleep },
604 { "cdev", &lock_class_mtx_sleep },
607 * kqueue/VFS interaction
609 { "kqueue", &lock_class_mtx_sleep },
610 { "struct mount mtx", &lock_class_mtx_sleep },
611 { "vnode interlock", &lock_class_mtx_sleep },
616 { "dn->dn_mtx", &lock_class_sx },
617 { "dr->dt.di.dr_mtx", &lock_class_sx },
618 { "db->db_mtx", &lock_class_sx },
624 { "ap boot", &lock_class_mtx_spin },
626 { "rm.mutex_mtx", &lock_class_mtx_spin },
627 { "sio", &lock_class_mtx_spin },
628 { "scrlock", &lock_class_mtx_spin },
630 { "cy", &lock_class_mtx_spin },
633 { "pcib_mtx", &lock_class_mtx_spin },
634 { "rtc_mtx", &lock_class_mtx_spin },
636 { "scc_hwmtx", &lock_class_mtx_spin },
637 { "uart_hwmtx", &lock_class_mtx_spin },
638 { "fast_taskqueue", &lock_class_mtx_spin },
639 { "intr table", &lock_class_mtx_spin },
641 { "pmc-per-proc", &lock_class_mtx_spin },
643 { "process slock", &lock_class_mtx_spin },
644 { "sleepq chain", &lock_class_mtx_spin },
645 { "umtx lock", &lock_class_mtx_spin },
646 { "rm_spinlock", &lock_class_mtx_spin },
647 { "turnstile chain", &lock_class_mtx_spin },
648 { "turnstile lock", &lock_class_mtx_spin },
649 { "sched lock", &lock_class_mtx_spin },
650 { "td_contested", &lock_class_mtx_spin },
651 { "callout", &lock_class_mtx_spin },
652 { "entropy harvest mutex", &lock_class_mtx_spin },
653 { "syscons video lock", &lock_class_mtx_spin },
654 { "time lock", &lock_class_mtx_spin },
656 { "smp rendezvous", &lock_class_mtx_spin },
659 { "tlb0", &lock_class_mtx_spin },
664 { "intrcnt", &lock_class_mtx_spin },
665 { "icu", &lock_class_mtx_spin },
666 #if defined(SMP) && defined(__sparc64__)
667 { "ipi", &lock_class_mtx_spin },
670 { "allpmaps", &lock_class_mtx_spin },
671 { "descriptor tables", &lock_class_mtx_spin },
673 { "clk", &lock_class_mtx_spin },
674 { "cpuset", &lock_class_mtx_spin },
675 { "mprof lock", &lock_class_mtx_spin },
676 { "zombie lock", &lock_class_mtx_spin },
677 { "ALD Queue", &lock_class_mtx_spin },
679 { "MCA spin lock", &lock_class_mtx_spin },
681 #if defined(__i386__) || defined(__amd64__)
682 { "pcicfg", &lock_class_mtx_spin },
683 { "NDIS thread lock", &lock_class_mtx_spin },
685 { "tw_osl_io_lock", &lock_class_mtx_spin },
686 { "tw_osl_q_lock", &lock_class_mtx_spin },
687 { "tw_cl_io_lock", &lock_class_mtx_spin },
688 { "tw_cl_intr_lock", &lock_class_mtx_spin },
689 { "tw_cl_gen_lock", &lock_class_mtx_spin },
691 { "pmc-leaf", &lock_class_mtx_spin },
693 { "blocked lock", &lock_class_mtx_spin },
700 * Pairs of locks which have been blessed
701 * Don't complain about order problems with blessed locks
703 static struct witness_blessed blessed_list[] = {
705 static int blessed_count =
706 sizeof(blessed_list) / sizeof(struct witness_blessed);
710 * This global is set to 0 once it becomes safe to use the witness code.
712 static int witness_cold = 1;
715 * This global is set to 1 once the static lock orders have been enrolled
716 * so that a warning can be issued for any spin locks enrolled later.
718 static int witness_spin_warn = 0;
721 * The WITNESS-enabled diagnostic code. Note that the witness code does
722 * assume that the early boot is single-threaded at least until after this
723 * routine is completed.
726 witness_initialize(void *dummy __unused)
728 struct lock_object *lock;
729 struct witness_order_list_entry *order;
730 struct witness *w, *w1;
733 w_data = malloc(sizeof (struct witness) * WITNESS_COUNT, M_WITNESS,
737 * We have to release Giant before initializing its witness
738 * structure so that WITNESS doesn't get confused.
741 mtx_assert(&Giant, MA_NOTOWNED);
743 CTR1(KTR_WITNESS, "%s: initializing witness", __func__);
744 mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET |
745 MTX_NOWITNESS | MTX_NOPROFILE);
746 for (i = WITNESS_COUNT - 1; i >= 0; i--) {
748 memset(w, 0, sizeof(*w));
749 w_data[i].w_index = i; /* Witness index never changes. */
752 KASSERT(STAILQ_FIRST(&w_free)->w_index == 0,
753 ("%s: Invalid list of free witness objects", __func__));
755 /* Witness with index 0 is not used to aid in debugging. */
756 STAILQ_REMOVE_HEAD(&w_free, w_list);
760 (sizeof(**w_rmatrix) * (WITNESS_COUNT+1) * (WITNESS_COUNT+1)));
762 for (i = 0; i < LOCK_CHILDCOUNT; i++)
763 witness_lock_list_free(&w_locklistdata[i]);
764 witness_init_hash_tables();
766 /* First add in all the specified order lists. */
767 for (order = order_lists; order->w_name != NULL; order++) {
768 w = enroll(order->w_name, order->w_class);
771 w->w_file = "order list";
772 for (order++; order->w_name != NULL; order++) {
773 w1 = enroll(order->w_name, order->w_class);
776 w1->w_file = "order list";
781 witness_spin_warn = 1;
783 /* Iterate through all locks and add them to witness. */
784 for (i = 0; pending_locks[i].wh_lock != NULL; i++) {
785 lock = pending_locks[i].wh_lock;
786 KASSERT(lock->lo_flags & LO_WITNESS,
787 ("%s: lock %s is on pending list but not LO_WITNESS",
788 __func__, lock->lo_name));
789 lock->lo_witness = enroll(pending_locks[i].wh_type,
793 /* Mark the witness code as being ready for use. */
798 SYSINIT(witness_init, SI_SUB_WITNESS, SI_ORDER_FIRST, witness_initialize,
802 witness_init(struct lock_object *lock, const char *type)
804 struct lock_class *class;
806 /* Various sanity checks. */
807 class = LOCK_CLASS(lock);
808 if ((lock->lo_flags & LO_RECURSABLE) != 0 &&
809 (class->lc_flags & LC_RECURSABLE) == 0)
810 panic("%s: lock (%s) %s can not be recursable", __func__,
811 class->lc_name, lock->lo_name);
812 if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
813 (class->lc_flags & LC_SLEEPABLE) == 0)
814 panic("%s: lock (%s) %s can not be sleepable", __func__,
815 class->lc_name, lock->lo_name);
816 if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
817 (class->lc_flags & LC_UPGRADABLE) == 0)
818 panic("%s: lock (%s) %s can not be upgradable", __func__,
819 class->lc_name, lock->lo_name);
822 * If we shouldn't watch this lock, then just clear lo_witness.
823 * Otherwise, if witness_cold is set, then it is too early to
824 * enroll this lock, so defer it to witness_initialize() by adding
825 * it to the pending_locks list. If it is not too early, then enroll
828 if (witness_watch < 1 || panicstr != NULL ||
829 (lock->lo_flags & LO_WITNESS) == 0)
830 lock->lo_witness = NULL;
831 else if (witness_cold) {
832 pending_locks[pending_cnt].wh_lock = lock;
833 pending_locks[pending_cnt++].wh_type = type;
834 if (pending_cnt > WITNESS_PENDLIST)
835 panic("%s: pending locks list is too small, bump it\n",
838 lock->lo_witness = enroll(type, class);
842 witness_destroy(struct lock_object *lock)
844 struct lock_class *class;
847 class = LOCK_CLASS(lock);
850 panic("lock (%s) %s destroyed while witness_cold",
851 class->lc_name, lock->lo_name);
853 /* XXX: need to verify that no one holds the lock */
854 if ((lock->lo_flags & LO_WITNESS) == 0 || lock->lo_witness == NULL)
856 w = lock->lo_witness;
858 mtx_lock_spin(&w_mtx);
859 MPASS(w->w_refcount > 0);
862 if (w->w_refcount == 0)
864 mtx_unlock_spin(&w_mtx);
869 witness_ddb_compute_levels(void)
874 * First clear all levels.
876 STAILQ_FOREACH(w, &w_all, w_list)
880 * Look for locks with no parents and level all their descendants.
882 STAILQ_FOREACH(w, &w_all, w_list) {
884 /* If the witness has ancestors (is not a root), skip it. */
885 if (w->w_num_ancestors > 0)
887 witness_ddb_level_descendants(w, 0);
892 witness_ddb_level_descendants(struct witness *w, int l)
896 if (w->w_ddb_level >= l)
902 for (i = 1; i <= w_max_used_index; i++) {
903 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
904 witness_ddb_level_descendants(&w_data[i], l);
909 witness_ddb_display_descendants(int(*prnt)(const char *fmt, ...),
910 struct witness *w, int indent)
914 for (i = 0; i < indent; i++)
916 prnt("%s (type: %s, depth: %d, active refs: %d)",
917 w->w_name, w->w_class->lc_name,
918 w->w_ddb_level, w->w_refcount);
919 if (w->w_displayed) {
920 prnt(" -- (already displayed)\n");
924 if (w->w_file != NULL && w->w_line != 0)
925 prnt(" -- last acquired @ %s:%d\n", w->w_file,
928 prnt(" -- never acquired\n");
930 WITNESS_INDEX_ASSERT(w->w_index);
931 for (i = 1; i <= w_max_used_index; i++) {
934 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
935 witness_ddb_display_descendants(prnt, &w_data[i],
941 witness_ddb_display_list(int(*prnt)(const char *fmt, ...),
942 struct witness_list *list)
946 STAILQ_FOREACH(w, list, w_typelist) {
947 if (w->w_file == NULL || w->w_ddb_level > 0)
950 /* This lock has no anscestors - display its descendants. */
951 witness_ddb_display_descendants(prnt, w, 0);
958 witness_ddb_display(int(*prnt)(const char *fmt, ...))
962 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
963 witness_ddb_compute_levels();
965 /* Clear all the displayed flags. */
966 STAILQ_FOREACH(w, &w_all, w_list)
970 * First, handle sleep locks which have been acquired at least
973 prnt("Sleep locks:\n");
974 witness_ddb_display_list(prnt, &w_sleep);
979 * Now do spin locks which have been acquired at least once.
981 prnt("\nSpin locks:\n");
982 witness_ddb_display_list(prnt, &w_spin);
987 * Finally, any locks which have not been acquired yet.
989 prnt("\nLocks which were never acquired:\n");
990 STAILQ_FOREACH(w, &w_all, w_list) {
991 if (w->w_file != NULL || w->w_refcount == 0)
993 prnt("%s (type: %s, depth: %d)\n", w->w_name,
994 w->w_class->lc_name, w->w_ddb_level);
1001 /* Trim useless garbage from filenames. */
1003 fixup_filename(const char *file)
1008 while (strncmp(file, "../", 3) == 0)
1014 witness_defineorder(struct lock_object *lock1, struct lock_object *lock2)
1017 if (witness_watch == -1 || panicstr != NULL)
1020 /* Require locks that witness knows about. */
1021 if (lock1 == NULL || lock1->lo_witness == NULL || lock2 == NULL ||
1022 lock2->lo_witness == NULL)
1025 mtx_assert(&w_mtx, MA_NOTOWNED);
1026 mtx_lock_spin(&w_mtx);
1029 * If we already have either an explicit or implied lock order that
1030 * is the other way around, then return an error.
1032 if (witness_watch &&
1033 isitmydescendant(lock2->lo_witness, lock1->lo_witness)) {
1034 mtx_unlock_spin(&w_mtx);
1038 /* Try to add the new order. */
1039 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1040 lock2->lo_witness->w_name, lock1->lo_witness->w_name);
1041 itismychild(lock1->lo_witness, lock2->lo_witness);
1042 mtx_unlock_spin(&w_mtx);
1047 witness_checkorder(struct lock_object *lock, int flags, const char *file,
1048 int line, struct lock_object *interlock)
1050 struct lock_list_entry *lock_list, *lle;
1051 struct lock_instance *lock1, *lock2, *plock;
1052 struct lock_class *class;
1053 struct witness *w, *w1;
1057 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL ||
1061 w = lock->lo_witness;
1062 class = LOCK_CLASS(lock);
1064 file = fixup_filename(file);
1066 if (class->lc_flags & LC_SLEEPLOCK) {
1069 * Since spin locks include a critical section, this check
1070 * implicitly enforces a lock order of all sleep locks before
1073 if (td->td_critnest != 0 && !kdb_active)
1074 panic("blockable sleep lock (%s) %s @ %s:%d",
1075 class->lc_name, lock->lo_name, file, line);
1078 * If this is the first lock acquired then just return as
1079 * no order checking is needed.
1081 lock_list = td->td_sleeplocks;
1082 if (lock_list == NULL || lock_list->ll_count == 0)
1087 * If this is the first lock, just return as no order
1088 * checking is needed. Avoid problems with thread
1089 * migration pinning the thread while checking if
1090 * spinlocks are held. If at least one spinlock is held
1091 * the thread is in a safe path and it is allowed to
1095 lock_list = PCPU_GET(spinlocks);
1096 if (lock_list == NULL || lock_list->ll_count == 0) {
1104 * Check to see if we are recursing on a lock we already own. If
1105 * so, make sure that we don't mismatch exclusive and shared lock
1108 lock1 = find_instance(lock_list, lock);
1109 if (lock1 != NULL) {
1110 if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
1111 (flags & LOP_EXCLUSIVE) == 0) {
1112 printf("shared lock of (%s) %s @ %s:%d\n",
1113 class->lc_name, lock->lo_name, file, line);
1114 printf("while exclusively locked from %s:%d\n",
1115 lock1->li_file, lock1->li_line);
1116 panic("share->excl");
1118 if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
1119 (flags & LOP_EXCLUSIVE) != 0) {
1120 printf("exclusive lock of (%s) %s @ %s:%d\n",
1121 class->lc_name, lock->lo_name, file, line);
1122 printf("while share locked from %s:%d\n",
1123 lock1->li_file, lock1->li_line);
1124 panic("excl->share");
1130 * Find the previously acquired lock, but ignore interlocks.
1132 plock = &lock_list->ll_children[lock_list->ll_count - 1];
1133 if (interlock != NULL && plock->li_lock == interlock) {
1134 if (lock_list->ll_count > 1)
1136 &lock_list->ll_children[lock_list->ll_count - 2];
1138 lle = lock_list->ll_next;
1141 * The interlock is the only lock we hold, so
1146 plock = &lle->ll_children[lle->ll_count - 1];
1151 * Try to perform most checks without a lock. If this succeeds we
1152 * can skip acquiring the lock and return success.
1154 w1 = plock->li_lock->lo_witness;
1155 if (witness_lock_order_check(w1, w))
1159 * Check for duplicate locks of the same type. Note that we only
1160 * have to check for this on the last lock we just acquired. Any
1161 * other cases will be caught as lock order violations.
1163 mtx_lock_spin(&w_mtx);
1164 witness_lock_order_add(w1, w);
1167 if (!(lock->lo_flags & LO_DUPOK) && !(flags & LOP_DUPOK) &&
1168 !(w_rmatrix[i][i] & WITNESS_REVERSAL)) {
1169 w_rmatrix[i][i] |= WITNESS_REVERSAL;
1171 mtx_unlock_spin(&w_mtx);
1173 "acquiring duplicate lock of same type: \"%s\"\n",
1175 printf(" 1st %s @ %s:%d\n", plock->li_lock->lo_name,
1176 plock->li_file, plock->li_line);
1177 printf(" 2nd %s @ %s:%d\n", lock->lo_name, file, line);
1178 witness_debugger(1);
1180 mtx_unlock_spin(&w_mtx);
1183 mtx_assert(&w_mtx, MA_OWNED);
1186 * If we know that the lock we are acquiring comes after
1187 * the lock we most recently acquired in the lock order tree,
1188 * then there is no need for any further checks.
1190 if (isitmychild(w1, w))
1193 for (j = 0, lle = lock_list; lle != NULL; lle = lle->ll_next) {
1194 for (i = lle->ll_count - 1; i >= 0; i--, j++) {
1196 MPASS(j < WITNESS_COUNT);
1197 lock1 = &lle->ll_children[i];
1200 * Ignore the interlock the first time we see it.
1202 if (interlock != NULL && interlock == lock1->li_lock) {
1208 * If this lock doesn't undergo witness checking,
1211 w1 = lock1->li_lock->lo_witness;
1213 KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
1214 ("lock missing witness structure"));
1219 * If we are locking Giant and this is a sleepable
1220 * lock, then skip it.
1222 if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0 &&
1223 lock == &Giant.lock_object)
1227 * If we are locking a sleepable lock and this lock
1228 * is Giant, then skip it.
1230 if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1231 lock1->li_lock == &Giant.lock_object)
1235 * If we are locking a sleepable lock and this lock
1236 * isn't sleepable, we want to treat it as a lock
1237 * order violation to enfore a general lock order of
1238 * sleepable locks before non-sleepable locks.
1240 if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1241 (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
1245 * If we are locking Giant and this is a non-sleepable
1246 * lock, then treat it as a reversal.
1248 if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0 &&
1249 lock == &Giant.lock_object)
1253 * Check the lock order hierarchy for a reveresal.
1255 if (!isitmydescendant(w, w1))
1260 * We have a lock order violation, check to see if it
1261 * is allowed or has already been yelled about.
1266 * If the lock order is blessed, just bail. We don't
1267 * look for other lock order violations though, which
1274 /* Bail if this violation is known */
1275 if (w_rmatrix[w1->w_index][w->w_index] & WITNESS_REVERSAL)
1278 /* Record this as a violation */
1279 w_rmatrix[w1->w_index][w->w_index] |= WITNESS_REVERSAL;
1280 w_rmatrix[w->w_index][w1->w_index] |= WITNESS_REVERSAL;
1281 w->w_reversed = w1->w_reversed = 1;
1282 witness_increment_graph_generation();
1283 mtx_unlock_spin(&w_mtx);
1286 * Ok, yell about it.
1288 if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1289 (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
1291 "lock order reversal: (sleepable after non-sleepable)\n");
1292 else if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0
1293 && lock == &Giant.lock_object)
1295 "lock order reversal: (Giant after non-sleepable)\n");
1297 printf("lock order reversal:\n");
1300 * Try to locate an earlier lock with
1301 * witness w in our list.
1304 lock2 = &lle->ll_children[i];
1305 MPASS(lock2->li_lock != NULL);
1306 if (lock2->li_lock->lo_witness == w)
1308 if (i == 0 && lle->ll_next != NULL) {
1310 i = lle->ll_count - 1;
1311 MPASS(i >= 0 && i < LOCK_NCHILDREN);
1316 printf(" 1st %p %s (%s) @ %s:%d\n",
1317 lock1->li_lock, lock1->li_lock->lo_name,
1318 w1->w_name, lock1->li_file, lock1->li_line);
1319 printf(" 2nd %p %s (%s) @ %s:%d\n", lock,
1320 lock->lo_name, w->w_name, file, line);
1322 printf(" 1st %p %s (%s) @ %s:%d\n",
1323 lock2->li_lock, lock2->li_lock->lo_name,
1324 lock2->li_lock->lo_witness->w_name,
1325 lock2->li_file, lock2->li_line);
1326 printf(" 2nd %p %s (%s) @ %s:%d\n",
1327 lock1->li_lock, lock1->li_lock->lo_name,
1328 w1->w_name, lock1->li_file, lock1->li_line);
1329 printf(" 3rd %p %s (%s) @ %s:%d\n", lock,
1330 lock->lo_name, w->w_name, file, line);
1332 witness_debugger(1);
1338 * If requested, build a new lock order. However, don't build a new
1339 * relationship between a sleepable lock and Giant if it is in the
1340 * wrong direction. The correct lock order is that sleepable locks
1341 * always come before Giant.
1343 if (flags & LOP_NEWORDER &&
1344 !(plock->li_lock == &Giant.lock_object &&
1345 (lock->lo_flags & LO_SLEEPABLE) != 0)) {
1346 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1347 w->w_name, plock->li_lock->lo_witness->w_name);
1348 itismychild(plock->li_lock->lo_witness, w);
1351 mtx_unlock_spin(&w_mtx);
1355 witness_lock(struct lock_object *lock, int flags, const char *file, int line)
1357 struct lock_list_entry **lock_list, *lle;
1358 struct lock_instance *instance;
1362 if (witness_cold || witness_watch == -1 || lock->lo_witness == NULL ||
1365 w = lock->lo_witness;
1367 file = fixup_filename(file);
1369 /* Determine lock list for this lock. */
1370 if (LOCK_CLASS(lock)->lc_flags & LC_SLEEPLOCK)
1371 lock_list = &td->td_sleeplocks;
1373 lock_list = PCPU_PTR(spinlocks);
1375 /* Check to see if we are recursing on a lock we already own. */
1376 instance = find_instance(*lock_list, lock);
1377 if (instance != NULL) {
1378 instance->li_flags++;
1379 CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
1380 td->td_proc->p_pid, lock->lo_name,
1381 instance->li_flags & LI_RECURSEMASK);
1382 instance->li_file = file;
1383 instance->li_line = line;
1387 /* Update per-witness last file and line acquire. */
1391 /* Find the next open lock instance in the list and fill it. */
1393 if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
1394 lle = witness_lock_list_get();
1397 lle->ll_next = *lock_list;
1398 CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
1399 td->td_proc->p_pid, lle);
1402 instance = &lle->ll_children[lle->ll_count++];
1403 instance->li_lock = lock;
1404 instance->li_line = line;
1405 instance->li_file = file;
1406 if ((flags & LOP_EXCLUSIVE) != 0)
1407 instance->li_flags = LI_EXCLUSIVE;
1409 instance->li_flags = 0;
1410 CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
1411 td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
1415 witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
1417 struct lock_instance *instance;
1418 struct lock_class *class;
1420 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1421 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
1423 class = LOCK_CLASS(lock);
1424 file = fixup_filename(file);
1425 if (witness_watch) {
1426 if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1427 panic("upgrade of non-upgradable lock (%s) %s @ %s:%d",
1428 class->lc_name, lock->lo_name, file, line);
1429 if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1430 panic("upgrade of non-sleep lock (%s) %s @ %s:%d",
1431 class->lc_name, lock->lo_name, file, line);
1433 instance = find_instance(curthread->td_sleeplocks, lock);
1434 if (instance == NULL)
1435 panic("upgrade of unlocked lock (%s) %s @ %s:%d",
1436 class->lc_name, lock->lo_name, file, line);
1437 if (witness_watch) {
1438 if ((instance->li_flags & LI_EXCLUSIVE) != 0)
1439 panic("upgrade of exclusive lock (%s) %s @ %s:%d",
1440 class->lc_name, lock->lo_name, file, line);
1441 if ((instance->li_flags & LI_RECURSEMASK) != 0)
1442 panic("upgrade of recursed lock (%s) %s r=%d @ %s:%d",
1443 class->lc_name, lock->lo_name,
1444 instance->li_flags & LI_RECURSEMASK, file, line);
1446 instance->li_flags |= LI_EXCLUSIVE;
1450 witness_downgrade(struct lock_object *lock, int flags, const char *file,
1453 struct lock_instance *instance;
1454 struct lock_class *class;
1456 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1457 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
1459 class = LOCK_CLASS(lock);
1460 file = fixup_filename(file);
1461 if (witness_watch) {
1462 if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1463 panic("downgrade of non-upgradable lock (%s) %s @ %s:%d",
1464 class->lc_name, lock->lo_name, file, line);
1465 if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1466 panic("downgrade of non-sleep lock (%s) %s @ %s:%d",
1467 class->lc_name, lock->lo_name, file, line);
1469 instance = find_instance(curthread->td_sleeplocks, lock);
1470 if (instance == NULL)
1471 panic("downgrade of unlocked lock (%s) %s @ %s:%d",
1472 class->lc_name, lock->lo_name, file, line);
1473 if (witness_watch) {
1474 if ((instance->li_flags & LI_EXCLUSIVE) == 0)
1475 panic("downgrade of shared lock (%s) %s @ %s:%d",
1476 class->lc_name, lock->lo_name, file, line);
1477 if ((instance->li_flags & LI_RECURSEMASK) != 0)
1478 panic("downgrade of recursed lock (%s) %s r=%d @ %s:%d",
1479 class->lc_name, lock->lo_name,
1480 instance->li_flags & LI_RECURSEMASK, file, line);
1482 instance->li_flags &= ~LI_EXCLUSIVE;
1486 witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
1488 struct lock_list_entry **lock_list, *lle;
1489 struct lock_instance *instance;
1490 struct lock_class *class;
1495 if (witness_cold || lock->lo_witness == NULL || panicstr != NULL)
1498 class = LOCK_CLASS(lock);
1499 file = fixup_filename(file);
1501 /* Find lock instance associated with this lock. */
1502 if (class->lc_flags & LC_SLEEPLOCK)
1503 lock_list = &td->td_sleeplocks;
1505 lock_list = PCPU_PTR(spinlocks);
1507 for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
1508 for (i = 0; i < (*lock_list)->ll_count; i++) {
1509 instance = &(*lock_list)->ll_children[i];
1510 if (instance->li_lock == lock)
1515 * When disabling WITNESS through witness_watch we could end up in
1516 * having registered locks in the td_sleeplocks queue.
1517 * We have to make sure we flush these queues, so just search for
1518 * eventual register locks and remove them.
1520 if (witness_watch > 0)
1521 panic("lock (%s) %s not locked @ %s:%d", class->lc_name,
1522 lock->lo_name, file, line);
1527 /* First, check for shared/exclusive mismatches. */
1528 if ((instance->li_flags & LI_EXCLUSIVE) != 0 && witness_watch > 0 &&
1529 (flags & LOP_EXCLUSIVE) == 0) {
1530 printf("shared unlock of (%s) %s @ %s:%d\n", class->lc_name,
1531 lock->lo_name, file, line);
1532 printf("while exclusively locked from %s:%d\n",
1533 instance->li_file, instance->li_line);
1534 panic("excl->ushare");
1536 if ((instance->li_flags & LI_EXCLUSIVE) == 0 && witness_watch > 0 &&
1537 (flags & LOP_EXCLUSIVE) != 0) {
1538 printf("exclusive unlock of (%s) %s @ %s:%d\n", class->lc_name,
1539 lock->lo_name, file, line);
1540 printf("while share locked from %s:%d\n", instance->li_file,
1542 panic("share->uexcl");
1544 /* If we are recursed, unrecurse. */
1545 if ((instance->li_flags & LI_RECURSEMASK) > 0) {
1546 CTR4(KTR_WITNESS, "%s: pid %d unrecursed on %s r=%d", __func__,
1547 td->td_proc->p_pid, instance->li_lock->lo_name,
1548 instance->li_flags);
1549 instance->li_flags--;
1552 /* The lock is now being dropped, check for NORELEASE flag */
1553 if ((instance->li_flags & LI_NORELEASE) != 0 && witness_watch > 0) {
1554 printf("forbidden unlock of (%s) %s @ %s:%d\n", class->lc_name,
1555 lock->lo_name, file, line);
1556 panic("lock marked norelease");
1559 /* Otherwise, remove this item from the list. */
1561 CTR4(KTR_WITNESS, "%s: pid %d removed %s from lle[%d]", __func__,
1562 td->td_proc->p_pid, instance->li_lock->lo_name,
1563 (*lock_list)->ll_count - 1);
1564 for (j = i; j < (*lock_list)->ll_count - 1; j++)
1565 (*lock_list)->ll_children[j] =
1566 (*lock_list)->ll_children[j + 1];
1567 (*lock_list)->ll_count--;
1571 * In order to reduce contention on w_mtx, we want to keep always an
1572 * head object into lists so that frequent allocation from the
1573 * free witness pool (and subsequent locking) is avoided.
1574 * In order to maintain the current code simple, when the head
1575 * object is totally unloaded it means also that we do not have
1576 * further objects in the list, so the list ownership needs to be
1577 * hand over to another object if the current head needs to be freed.
1579 if ((*lock_list)->ll_count == 0) {
1580 if (*lock_list == lle) {
1581 if (lle->ll_next == NULL)
1585 *lock_list = lle->ll_next;
1586 CTR3(KTR_WITNESS, "%s: pid %d removed lle %p", __func__,
1587 td->td_proc->p_pid, lle);
1588 witness_lock_list_free(lle);
1593 witness_thread_exit(struct thread *td)
1595 struct lock_list_entry *lle;
1598 lle = td->td_sleeplocks;
1599 if (lle == NULL || panicstr != NULL)
1601 if (lle->ll_count != 0) {
1602 for (n = 0; lle != NULL; lle = lle->ll_next)
1603 for (i = lle->ll_count - 1; i >= 0; i--) {
1605 printf("Thread %p exiting with the following locks held:\n",
1608 witness_list_lock(&lle->ll_children[i], printf);
1611 panic("Thread %p cannot exit while holding sleeplocks\n", td);
1613 witness_lock_list_free(lle);
1617 * Warn if any locks other than 'lock' are held. Flags can be passed in to
1618 * exempt Giant and sleepable locks from the checks as well. If any
1619 * non-exempt locks are held, then a supplied message is printed to the
1620 * console along with a list of the offending locks. If indicated in the
1621 * flags then a failure results in a panic as well.
1624 witness_warn(int flags, struct lock_object *lock, const char *fmt, ...)
1626 struct lock_list_entry *lock_list, *lle;
1627 struct lock_instance *lock1;
1632 if (witness_cold || witness_watch < 1 || panicstr != NULL)
1636 for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next)
1637 for (i = lle->ll_count - 1; i >= 0; i--) {
1638 lock1 = &lle->ll_children[i];
1639 if (lock1->li_lock == lock)
1641 if (flags & WARN_GIANTOK &&
1642 lock1->li_lock == &Giant.lock_object)
1644 if (flags & WARN_SLEEPOK &&
1645 (lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0)
1651 printf(" with the following");
1652 if (flags & WARN_SLEEPOK)
1653 printf(" non-sleepable");
1654 printf(" locks held:\n");
1657 witness_list_lock(lock1, printf);
1661 * Pin the thread in order to avoid problems with thread migration.
1662 * Once that all verifies are passed about spinlocks ownership,
1663 * the thread is in a safe path and it can be unpinned.
1666 lock_list = PCPU_GET(spinlocks);
1667 if (lock_list != NULL && lock_list->ll_count != 0) {
1671 * We should only have one spinlock and as long as
1672 * the flags cannot match for this locks class,
1673 * check if the first spinlock is the one curthread
1676 lock1 = &lock_list->ll_children[lock_list->ll_count - 1];
1677 if (lock_list->ll_count == 1 && lock_list->ll_next == NULL &&
1678 lock1->li_lock == lock && n == 0)
1684 printf(" with the following");
1685 if (flags & WARN_SLEEPOK)
1686 printf(" non-sleepable");
1687 printf(" locks held:\n");
1688 n += witness_list_locks(&lock_list, printf);
1691 if (flags & WARN_PANIC && n)
1692 panic("%s", __func__);
1694 witness_debugger(n);
1699 witness_file(struct lock_object *lock)
1703 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1705 w = lock->lo_witness;
1710 witness_line(struct lock_object *lock)
1714 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1716 w = lock->lo_witness;
1720 static struct witness *
1721 enroll(const char *description, struct lock_class *lock_class)
1724 struct witness_list *typelist;
1726 MPASS(description != NULL);
1728 if (witness_watch == -1 || panicstr != NULL)
1730 if ((lock_class->lc_flags & LC_SPINLOCK)) {
1731 if (witness_skipspin)
1735 } else if ((lock_class->lc_flags & LC_SLEEPLOCK))
1736 typelist = &w_sleep;
1738 panic("lock class %s is not sleep or spin",
1739 lock_class->lc_name);
1741 mtx_lock_spin(&w_mtx);
1742 w = witness_hash_get(description);
1745 if ((w = witness_get()) == NULL)
1747 MPASS(strlen(description) < MAX_W_NAME);
1748 strcpy(w->w_name, description);
1749 w->w_class = lock_class;
1751 STAILQ_INSERT_HEAD(&w_all, w, w_list);
1752 if (lock_class->lc_flags & LC_SPINLOCK) {
1753 STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
1755 } else if (lock_class->lc_flags & LC_SLEEPLOCK) {
1756 STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
1760 /* Insert new witness into the hash */
1761 witness_hash_put(w);
1762 witness_increment_graph_generation();
1763 mtx_unlock_spin(&w_mtx);
1767 mtx_unlock_spin(&w_mtx);
1768 if (lock_class != w->w_class)
1770 "lock (%s) %s does not match earlier (%s) lock",
1771 description, lock_class->lc_name,
1772 w->w_class->lc_name);
1777 depart(struct witness *w)
1779 struct witness_list *list;
1781 MPASS(w->w_refcount == 0);
1782 if (w->w_class->lc_flags & LC_SLEEPLOCK) {
1790 * Set file to NULL as it may point into a loadable module.
1794 witness_increment_graph_generation();
1799 adopt(struct witness *parent, struct witness *child)
1803 if (witness_cold == 0)
1804 mtx_assert(&w_mtx, MA_OWNED);
1806 /* If the relationship is already known, there's no work to be done. */
1807 if (isitmychild(parent, child))
1810 /* When the structure of the graph changes, bump up the generation. */
1811 witness_increment_graph_generation();
1814 * The hard part ... create the direct relationship, then propagate all
1815 * indirect relationships.
1817 pi = parent->w_index;
1818 ci = child->w_index;
1819 WITNESS_INDEX_ASSERT(pi);
1820 WITNESS_INDEX_ASSERT(ci);
1822 w_rmatrix[pi][ci] |= WITNESS_PARENT;
1823 w_rmatrix[ci][pi] |= WITNESS_CHILD;
1826 * If parent was not already an ancestor of child,
1827 * then we increment the descendant and ancestor counters.
1829 if ((w_rmatrix[pi][ci] & WITNESS_ANCESTOR) == 0) {
1830 parent->w_num_descendants++;
1831 child->w_num_ancestors++;
1835 * Find each ancestor of 'pi'. Note that 'pi' itself is counted as
1836 * an ancestor of 'pi' during this loop.
1838 for (i = 1; i <= w_max_used_index; i++) {
1839 if ((w_rmatrix[i][pi] & WITNESS_ANCESTOR_MASK) == 0 &&
1843 /* Find each descendant of 'i' and mark it as a descendant. */
1844 for (j = 1; j <= w_max_used_index; j++) {
1847 * Skip children that are already marked as
1848 * descendants of 'i'.
1850 if (w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK)
1854 * We are only interested in descendants of 'ci'. Note
1855 * that 'ci' itself is counted as a descendant of 'ci'.
1857 if ((w_rmatrix[ci][j] & WITNESS_ANCESTOR_MASK) == 0 &&
1860 w_rmatrix[i][j] |= WITNESS_ANCESTOR;
1861 w_rmatrix[j][i] |= WITNESS_DESCENDANT;
1862 w_data[i].w_num_descendants++;
1863 w_data[j].w_num_ancestors++;
1866 * Make sure we aren't marking a node as both an
1867 * ancestor and descendant. We should have caught
1868 * this as a lock order reversal earlier.
1870 if ((w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK) &&
1871 (w_rmatrix[i][j] & WITNESS_DESCENDANT_MASK)) {
1872 printf("witness rmatrix paradox! [%d][%d]=%d "
1873 "both ancestor and descendant\n",
1874 i, j, w_rmatrix[i][j]);
1876 printf("Witness disabled.\n");
1879 if ((w_rmatrix[j][i] & WITNESS_ANCESTOR_MASK) &&
1880 (w_rmatrix[j][i] & WITNESS_DESCENDANT_MASK)) {
1881 printf("witness rmatrix paradox! [%d][%d]=%d "
1882 "both ancestor and descendant\n",
1883 j, i, w_rmatrix[j][i]);
1885 printf("Witness disabled.\n");
1893 itismychild(struct witness *parent, struct witness *child)
1896 MPASS(child != NULL && parent != NULL);
1897 if (witness_cold == 0)
1898 mtx_assert(&w_mtx, MA_OWNED);
1900 if (!witness_lock_type_equal(parent, child)) {
1901 if (witness_cold == 0)
1902 mtx_unlock_spin(&w_mtx);
1903 panic("%s: parent \"%s\" (%s) and child \"%s\" (%s) are not "
1904 "the same lock type", __func__, parent->w_name,
1905 parent->w_class->lc_name, child->w_name,
1906 child->w_class->lc_name);
1908 adopt(parent, child);
1912 * Generic code for the isitmy*() functions. The rmask parameter is the
1913 * expected relationship of w1 to w2.
1916 _isitmyx(struct witness *w1, struct witness *w2, int rmask, const char *fname)
1918 unsigned char r1, r2;
1923 WITNESS_INDEX_ASSERT(i1);
1924 WITNESS_INDEX_ASSERT(i2);
1925 r1 = w_rmatrix[i1][i2] & WITNESS_RELATED_MASK;
1926 r2 = w_rmatrix[i2][i1] & WITNESS_RELATED_MASK;
1928 /* The flags on one better be the inverse of the flags on the other */
1929 if (!((WITNESS_ATOD(r1) == r2 && WITNESS_DTOA(r2) == r1) ||
1930 (WITNESS_DTOA(r1) == r2 && WITNESS_ATOD(r2) == r1))) {
1931 printf("%s: rmatrix mismatch between %s (index %d) and %s "
1932 "(index %d): w_rmatrix[%d][%d] == %hhx but "
1933 "w_rmatrix[%d][%d] == %hhx\n",
1934 fname, w1->w_name, i1, w2->w_name, i2, i1, i2, r1,
1937 printf("Witness disabled.\n");
1940 return (r1 & rmask);
1944 * Checks if @child is a direct child of @parent.
1947 isitmychild(struct witness *parent, struct witness *child)
1950 return (_isitmyx(parent, child, WITNESS_PARENT, __func__));
1954 * Checks if @descendant is a direct or inderect descendant of @ancestor.
1957 isitmydescendant(struct witness *ancestor, struct witness *descendant)
1960 return (_isitmyx(ancestor, descendant, WITNESS_ANCESTOR_MASK,
1966 blessed(struct witness *w1, struct witness *w2)
1969 struct witness_blessed *b;
1971 for (i = 0; i < blessed_count; i++) {
1972 b = &blessed_list[i];
1973 if (strcmp(w1->w_name, b->b_lock1) == 0) {
1974 if (strcmp(w2->w_name, b->b_lock2) == 0)
1978 if (strcmp(w1->w_name, b->b_lock2) == 0)
1979 if (strcmp(w2->w_name, b->b_lock1) == 0)
1986 static struct witness *
1992 if (witness_cold == 0)
1993 mtx_assert(&w_mtx, MA_OWNED);
1995 if (witness_watch == -1) {
1996 mtx_unlock_spin(&w_mtx);
1999 if (STAILQ_EMPTY(&w_free)) {
2001 mtx_unlock_spin(&w_mtx);
2002 printf("WITNESS: unable to allocate a new witness object\n");
2005 w = STAILQ_FIRST(&w_free);
2006 STAILQ_REMOVE_HEAD(&w_free, w_list);
2009 MPASS(index > 0 && index == w_max_used_index+1 &&
2010 index < WITNESS_COUNT);
2011 bzero(w, sizeof(*w));
2013 if (index > w_max_used_index)
2014 w_max_used_index = index;
2019 witness_free(struct witness *w)
2022 STAILQ_INSERT_HEAD(&w_free, w, w_list);
2026 static struct lock_list_entry *
2027 witness_lock_list_get(void)
2029 struct lock_list_entry *lle;
2031 if (witness_watch == -1)
2033 mtx_lock_spin(&w_mtx);
2034 lle = w_lock_list_free;
2037 mtx_unlock_spin(&w_mtx);
2038 printf("%s: witness exhausted\n", __func__);
2041 w_lock_list_free = lle->ll_next;
2042 mtx_unlock_spin(&w_mtx);
2043 bzero(lle, sizeof(*lle));
2048 witness_lock_list_free(struct lock_list_entry *lle)
2051 mtx_lock_spin(&w_mtx);
2052 lle->ll_next = w_lock_list_free;
2053 w_lock_list_free = lle;
2054 mtx_unlock_spin(&w_mtx);
2057 static struct lock_instance *
2058 find_instance(struct lock_list_entry *list, struct lock_object *lock)
2060 struct lock_list_entry *lle;
2061 struct lock_instance *instance;
2064 for (lle = list; lle != NULL; lle = lle->ll_next)
2065 for (i = lle->ll_count - 1; i >= 0; i--) {
2066 instance = &lle->ll_children[i];
2067 if (instance->li_lock == lock)
2074 witness_list_lock(struct lock_instance *instance,
2075 int (*prnt)(const char *fmt, ...))
2077 struct lock_object *lock;
2079 lock = instance->li_lock;
2080 prnt("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ?
2081 "exclusive" : "shared", LOCK_CLASS(lock)->lc_name, lock->lo_name);
2082 if (lock->lo_witness->w_name != lock->lo_name)
2083 prnt(" (%s)", lock->lo_witness->w_name);
2084 prnt(" r = %d (%p) locked @ %s:%d\n",
2085 instance->li_flags & LI_RECURSEMASK, lock, instance->li_file,
2091 witness_thread_has_locks(struct thread *td)
2094 if (td->td_sleeplocks == NULL)
2096 return (td->td_sleeplocks->ll_count != 0);
2100 witness_proc_has_locks(struct proc *p)
2104 FOREACH_THREAD_IN_PROC(p, td) {
2105 if (witness_thread_has_locks(td))
2113 witness_list_locks(struct lock_list_entry **lock_list,
2114 int (*prnt)(const char *fmt, ...))
2116 struct lock_list_entry *lle;
2120 for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
2121 for (i = lle->ll_count - 1; i >= 0; i--) {
2122 witness_list_lock(&lle->ll_children[i], prnt);
2129 * This is a bit risky at best. We call this function when we have timed
2130 * out acquiring a spin lock, and we assume that the other CPU is stuck
2131 * with this lock held. So, we go groveling around in the other CPU's
2132 * per-cpu data to try to find the lock instance for this spin lock to
2133 * see when it was last acquired.
2136 witness_display_spinlock(struct lock_object *lock, struct thread *owner,
2137 int (*prnt)(const char *fmt, ...))
2139 struct lock_instance *instance;
2142 if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU)
2144 pc = pcpu_find(owner->td_oncpu);
2145 instance = find_instance(pc->pc_spinlocks, lock);
2146 if (instance != NULL)
2147 witness_list_lock(instance, prnt);
2151 witness_save(struct lock_object *lock, const char **filep, int *linep)
2153 struct lock_list_entry *lock_list;
2154 struct lock_instance *instance;
2155 struct lock_class *class;
2158 * This function is used independently in locking code to deal with
2159 * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2162 if (SCHEDULER_STOPPED())
2164 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2165 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2167 class = LOCK_CLASS(lock);
2168 if (class->lc_flags & LC_SLEEPLOCK)
2169 lock_list = curthread->td_sleeplocks;
2171 if (witness_skipspin)
2173 lock_list = PCPU_GET(spinlocks);
2175 instance = find_instance(lock_list, lock);
2176 if (instance == NULL)
2177 panic("%s: lock (%s) %s not locked", __func__,
2178 class->lc_name, lock->lo_name);
2179 *filep = instance->li_file;
2180 *linep = instance->li_line;
2184 witness_restore(struct lock_object *lock, const char *file, int line)
2186 struct lock_list_entry *lock_list;
2187 struct lock_instance *instance;
2188 struct lock_class *class;
2191 * This function is used independently in locking code to deal with
2192 * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2195 if (SCHEDULER_STOPPED())
2197 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2198 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2200 class = LOCK_CLASS(lock);
2201 if (class->lc_flags & LC_SLEEPLOCK)
2202 lock_list = curthread->td_sleeplocks;
2204 if (witness_skipspin)
2206 lock_list = PCPU_GET(spinlocks);
2208 instance = find_instance(lock_list, lock);
2209 if (instance == NULL)
2210 panic("%s: lock (%s) %s not locked", __func__,
2211 class->lc_name, lock->lo_name);
2212 lock->lo_witness->w_file = file;
2213 lock->lo_witness->w_line = line;
2214 instance->li_file = file;
2215 instance->li_line = line;
2219 witness_assert(struct lock_object *lock, int flags, const char *file, int line)
2221 #ifdef INVARIANT_SUPPORT
2222 struct lock_instance *instance;
2223 struct lock_class *class;
2225 if (lock->lo_witness == NULL || witness_watch < 1 || panicstr != NULL)
2227 class = LOCK_CLASS(lock);
2228 if ((class->lc_flags & LC_SLEEPLOCK) != 0)
2229 instance = find_instance(curthread->td_sleeplocks, lock);
2230 else if ((class->lc_flags & LC_SPINLOCK) != 0)
2231 instance = find_instance(PCPU_GET(spinlocks), lock);
2233 panic("Lock (%s) %s is not sleep or spin!",
2234 class->lc_name, lock->lo_name);
2236 file = fixup_filename(file);
2239 if (instance != NULL)
2240 panic("Lock (%s) %s locked @ %s:%d.",
2241 class->lc_name, lock->lo_name, file, line);
2244 case LA_LOCKED | LA_RECURSED:
2245 case LA_LOCKED | LA_NOTRECURSED:
2247 case LA_SLOCKED | LA_RECURSED:
2248 case LA_SLOCKED | LA_NOTRECURSED:
2250 case LA_XLOCKED | LA_RECURSED:
2251 case LA_XLOCKED | LA_NOTRECURSED:
2252 if (instance == NULL) {
2253 panic("Lock (%s) %s not locked @ %s:%d.",
2254 class->lc_name, lock->lo_name, file, line);
2257 if ((flags & LA_XLOCKED) != 0 &&
2258 (instance->li_flags & LI_EXCLUSIVE) == 0)
2259 panic("Lock (%s) %s not exclusively locked @ %s:%d.",
2260 class->lc_name, lock->lo_name, file, line);
2261 if ((flags & LA_SLOCKED) != 0 &&
2262 (instance->li_flags & LI_EXCLUSIVE) != 0)
2263 panic("Lock (%s) %s exclusively locked @ %s:%d.",
2264 class->lc_name, lock->lo_name, file, line);
2265 if ((flags & LA_RECURSED) != 0 &&
2266 (instance->li_flags & LI_RECURSEMASK) == 0)
2267 panic("Lock (%s) %s not recursed @ %s:%d.",
2268 class->lc_name, lock->lo_name, file, line);
2269 if ((flags & LA_NOTRECURSED) != 0 &&
2270 (instance->li_flags & LI_RECURSEMASK) != 0)
2271 panic("Lock (%s) %s recursed @ %s:%d.",
2272 class->lc_name, lock->lo_name, file, line);
2275 panic("Invalid lock assertion at %s:%d.", file, line);
2278 #endif /* INVARIANT_SUPPORT */
2282 witness_setflag(struct lock_object *lock, int flag, int set)
2284 struct lock_list_entry *lock_list;
2285 struct lock_instance *instance;
2286 struct lock_class *class;
2288 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2290 class = LOCK_CLASS(lock);
2291 if (class->lc_flags & LC_SLEEPLOCK)
2292 lock_list = curthread->td_sleeplocks;
2294 if (witness_skipspin)
2296 lock_list = PCPU_GET(spinlocks);
2298 instance = find_instance(lock_list, lock);
2299 if (instance == NULL)
2300 panic("%s: lock (%s) %s not locked", __func__,
2301 class->lc_name, lock->lo_name);
2304 instance->li_flags |= flag;
2306 instance->li_flags &= ~flag;
2310 witness_norelease(struct lock_object *lock)
2313 witness_setflag(lock, LI_NORELEASE, 1);
2317 witness_releaseok(struct lock_object *lock)
2320 witness_setflag(lock, LI_NORELEASE, 0);
2325 witness_ddb_list(struct thread *td)
2328 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2329 KASSERT(kdb_active, ("%s: not in the debugger", __func__));
2331 if (witness_watch < 1)
2334 witness_list_locks(&td->td_sleeplocks, db_printf);
2337 * We only handle spinlocks if td == curthread. This is somewhat broken
2338 * if td is currently executing on some other CPU and holds spin locks
2339 * as we won't display those locks. If we had a MI way of getting
2340 * the per-cpu data for a given cpu then we could use
2341 * td->td_oncpu to get the list of spinlocks for this thread
2344 * That still wouldn't really fix this unless we locked the scheduler
2345 * lock or stopped the other CPU to make sure it wasn't changing the
2346 * list out from under us. It is probably best to just not try to
2347 * handle threads on other CPU's for now.
2349 if (td == curthread && PCPU_GET(spinlocks) != NULL)
2350 witness_list_locks(PCPU_PTR(spinlocks), db_printf);
2353 DB_SHOW_COMMAND(locks, db_witness_list)
2358 td = db_lookup_thread(addr, TRUE);
2361 witness_ddb_list(td);
2364 DB_SHOW_ALL_COMMAND(locks, db_witness_list_all)
2370 * It would be nice to list only threads and processes that actually
2371 * held sleep locks, but that information is currently not exported
2374 FOREACH_PROC_IN_SYSTEM(p) {
2375 if (!witness_proc_has_locks(p))
2377 FOREACH_THREAD_IN_PROC(p, td) {
2378 if (!witness_thread_has_locks(td))
2380 db_printf("Process %d (%s) thread %p (%d)\n", p->p_pid,
2381 p->p_comm, td, td->td_tid);
2382 witness_ddb_list(td);
2388 DB_SHOW_ALIAS(alllocks, db_witness_list_all)
2390 DB_SHOW_COMMAND(witness, db_witness_display)
2393 witness_ddb_display(db_printf);
2398 sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS)
2400 struct witness_lock_order_data *data1, *data2, *tmp_data1, *tmp_data2;
2401 struct witness *tmp_w1, *tmp_w2, *w1, *w2;
2403 u_int w_rmatrix1, w_rmatrix2;
2404 int error, generation, i, j;
2410 if (witness_watch < 1) {
2411 error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2415 error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2419 sb = sbuf_new(NULL, NULL, BADSTACK_SBUF_SIZE, SBUF_AUTOEXTEND);
2423 /* Allocate and init temporary storage space. */
2424 tmp_w1 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2425 tmp_w2 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2426 tmp_data1 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2428 tmp_data2 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2430 stack_zero(&tmp_data1->wlod_stack);
2431 stack_zero(&tmp_data2->wlod_stack);
2434 mtx_lock_spin(&w_mtx);
2435 generation = w_generation;
2436 mtx_unlock_spin(&w_mtx);
2437 sbuf_printf(sb, "Number of known direct relationships is %d\n",
2438 w_lohash.wloh_count);
2439 for (i = 1; i < w_max_used_index; i++) {
2440 mtx_lock_spin(&w_mtx);
2441 if (generation != w_generation) {
2442 mtx_unlock_spin(&w_mtx);
2444 /* The graph has changed, try again. */
2451 if (w1->w_reversed == 0) {
2452 mtx_unlock_spin(&w_mtx);
2456 /* Copy w1 locally so we can release the spin lock. */
2458 mtx_unlock_spin(&w_mtx);
2460 if (tmp_w1->w_reversed == 0)
2462 for (j = 1; j < w_max_used_index; j++) {
2463 if ((w_rmatrix[i][j] & WITNESS_REVERSAL) == 0 || i > j)
2466 mtx_lock_spin(&w_mtx);
2467 if (generation != w_generation) {
2468 mtx_unlock_spin(&w_mtx);
2470 /* The graph has changed, try again. */
2477 data1 = witness_lock_order_get(w1, w2);
2478 data2 = witness_lock_order_get(w2, w1);
2481 * Copy information locally so we can release the
2485 w_rmatrix1 = (unsigned int)w_rmatrix[i][j];
2486 w_rmatrix2 = (unsigned int)w_rmatrix[j][i];
2489 stack_zero(&tmp_data1->wlod_stack);
2490 stack_copy(&data1->wlod_stack,
2491 &tmp_data1->wlod_stack);
2493 if (data2 && data2 != data1) {
2494 stack_zero(&tmp_data2->wlod_stack);
2495 stack_copy(&data2->wlod_stack,
2496 &tmp_data2->wlod_stack);
2498 mtx_unlock_spin(&w_mtx);
2501 "\nLock order reversal between \"%s\"(%s) and \"%s\"(%s)!\n",
2502 tmp_w1->w_name, tmp_w1->w_class->lc_name,
2503 tmp_w2->w_name, tmp_w2->w_class->lc_name);
2506 "w_rmatrix[%s][%s] == %x, w_rmatrix[%s][%s] == %x\n",
2507 tmp_w1->name, tmp_w2->w_name, w_rmatrix1,
2508 tmp_w2->name, tmp_w1->w_name, w_rmatrix2);
2512 "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2513 tmp_w1->w_name, tmp_w1->w_class->lc_name,
2514 tmp_w2->w_name, tmp_w2->w_class->lc_name);
2515 stack_sbuf_print(sb, &tmp_data1->wlod_stack);
2516 sbuf_printf(sb, "\n");
2518 if (data2 && data2 != data1) {
2520 "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2521 tmp_w2->w_name, tmp_w2->w_class->lc_name,
2522 tmp_w1->w_name, tmp_w1->w_class->lc_name);
2523 stack_sbuf_print(sb, &tmp_data2->wlod_stack);
2524 sbuf_printf(sb, "\n");
2528 mtx_lock_spin(&w_mtx);
2529 if (generation != w_generation) {
2530 mtx_unlock_spin(&w_mtx);
2533 * The graph changed while we were printing stack data,
2540 mtx_unlock_spin(&w_mtx);
2542 /* Free temporary storage space. */
2543 free(tmp_data1, M_TEMP);
2544 free(tmp_data2, M_TEMP);
2545 free(tmp_w1, M_TEMP);
2546 free(tmp_w2, M_TEMP);
2549 error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1);
2556 sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS)
2562 if (witness_watch < 1) {
2563 error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2567 error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2571 sb = sbuf_new(NULL, NULL, FULLGRAPH_SBUF_SIZE, SBUF_FIXEDLEN);
2574 sbuf_printf(sb, "\n");
2576 mtx_lock_spin(&w_mtx);
2577 STAILQ_FOREACH(w, &w_all, w_list)
2579 STAILQ_FOREACH(w, &w_all, w_list)
2580 witness_add_fullgraph(sb, w);
2581 mtx_unlock_spin(&w_mtx);
2584 * While using SBUF_FIXEDLEN, check if the sbuf overflowed.
2586 if (sbuf_overflowed(sb)) {
2588 panic("%s: sbuf overflowed, bump FULLGRAPH_SBUF_SIZE value\n",
2593 * Close the sbuf and return to userland.
2596 error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1);
2603 sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS)
2607 value = witness_watch;
2608 error = sysctl_handle_int(oidp, &value, 0, req);
2609 if (error != 0 || req->newptr == NULL)
2611 if (value > 1 || value < -1 ||
2612 (witness_watch == -1 && value != witness_watch))
2614 witness_watch = value;
2619 witness_add_fullgraph(struct sbuf *sb, struct witness *w)
2623 if (w->w_displayed != 0 || (w->w_file == NULL && w->w_line == 0))
2627 WITNESS_INDEX_ASSERT(w->w_index);
2628 for (i = 1; i <= w_max_used_index; i++) {
2629 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT) {
2630 sbuf_printf(sb, "\"%s\",\"%s\"\n", w->w_name,
2632 witness_add_fullgraph(sb, &w_data[i]);
2638 * A simple hash function. Takes a key pointer and a key size. If size == 0,
2639 * interprets the key as a string and reads until the null
2640 * terminator. Otherwise, reads the first size bytes. Returns an unsigned 32-bit
2641 * hash value computed from the key.
2644 witness_hash_djb2(const uint8_t *key, uint32_t size)
2646 unsigned int hash = 5381;
2649 /* hash = hash * 33 + key[i] */
2651 for (i = 0; i < size; i++)
2652 hash = ((hash << 5) + hash) + (unsigned int)key[i];
2654 for (i = 0; key[i] != 0; i++)
2655 hash = ((hash << 5) + hash) + (unsigned int)key[i];
2662 * Initializes the two witness hash tables. Called exactly once from
2663 * witness_initialize().
2666 witness_init_hash_tables(void)
2670 MPASS(witness_cold);
2672 /* Initialize the hash tables. */
2673 for (i = 0; i < WITNESS_HASH_SIZE; i++)
2674 w_hash.wh_array[i] = NULL;
2676 w_hash.wh_size = WITNESS_HASH_SIZE;
2677 w_hash.wh_count = 0;
2679 /* Initialize the lock order data hash. */
2681 for (i = 0; i < WITNESS_LO_DATA_COUNT; i++) {
2682 memset(&w_lodata[i], 0, sizeof(w_lodata[i]));
2683 w_lodata[i].wlod_next = w_lofree;
2684 w_lofree = &w_lodata[i];
2686 w_lohash.wloh_size = WITNESS_LO_HASH_SIZE;
2687 w_lohash.wloh_count = 0;
2688 for (i = 0; i < WITNESS_LO_HASH_SIZE; i++)
2689 w_lohash.wloh_array[i] = NULL;
2692 static struct witness *
2693 witness_hash_get(const char *key)
2699 if (witness_cold == 0)
2700 mtx_assert(&w_mtx, MA_OWNED);
2701 hash = witness_hash_djb2(key, 0) % w_hash.wh_size;
2702 w = w_hash.wh_array[hash];
2704 if (strcmp(w->w_name, key) == 0)
2714 witness_hash_put(struct witness *w)
2719 MPASS(w->w_name != NULL);
2720 if (witness_cold == 0)
2721 mtx_assert(&w_mtx, MA_OWNED);
2722 KASSERT(witness_hash_get(w->w_name) == NULL,
2723 ("%s: trying to add a hash entry that already exists!", __func__));
2724 KASSERT(w->w_hash_next == NULL,
2725 ("%s: w->w_hash_next != NULL", __func__));
2727 hash = witness_hash_djb2(w->w_name, 0) % w_hash.wh_size;
2728 w->w_hash_next = w_hash.wh_array[hash];
2729 w_hash.wh_array[hash] = w;
2734 static struct witness_lock_order_data *
2735 witness_lock_order_get(struct witness *parent, struct witness *child)
2737 struct witness_lock_order_data *data = NULL;
2738 struct witness_lock_order_key key;
2741 MPASS(parent != NULL && child != NULL);
2742 key.from = parent->w_index;
2743 key.to = child->w_index;
2744 WITNESS_INDEX_ASSERT(key.from);
2745 WITNESS_INDEX_ASSERT(key.to);
2746 if ((w_rmatrix[parent->w_index][child->w_index]
2747 & WITNESS_LOCK_ORDER_KNOWN) == 0)
2750 hash = witness_hash_djb2((const char*)&key,
2751 sizeof(key)) % w_lohash.wloh_size;
2752 data = w_lohash.wloh_array[hash];
2753 while (data != NULL) {
2754 if (witness_lock_order_key_equal(&data->wlod_key, &key))
2756 data = data->wlod_next;
2764 * Verify that parent and child have a known relationship, are not the same,
2765 * and child is actually a child of parent. This is done without w_mtx
2766 * to avoid contention in the common case.
2769 witness_lock_order_check(struct witness *parent, struct witness *child)
2772 if (parent != child &&
2773 w_rmatrix[parent->w_index][child->w_index]
2774 & WITNESS_LOCK_ORDER_KNOWN &&
2775 isitmychild(parent, child))
2782 witness_lock_order_add(struct witness *parent, struct witness *child)
2784 struct witness_lock_order_data *data = NULL;
2785 struct witness_lock_order_key key;
2788 MPASS(parent != NULL && child != NULL);
2789 key.from = parent->w_index;
2790 key.to = child->w_index;
2791 WITNESS_INDEX_ASSERT(key.from);
2792 WITNESS_INDEX_ASSERT(key.to);
2793 if (w_rmatrix[parent->w_index][child->w_index]
2794 & WITNESS_LOCK_ORDER_KNOWN)
2797 hash = witness_hash_djb2((const char*)&key,
2798 sizeof(key)) % w_lohash.wloh_size;
2799 w_rmatrix[parent->w_index][child->w_index] |= WITNESS_LOCK_ORDER_KNOWN;
2803 w_lofree = data->wlod_next;
2804 data->wlod_next = w_lohash.wloh_array[hash];
2805 data->wlod_key = key;
2806 w_lohash.wloh_array[hash] = data;
2807 w_lohash.wloh_count++;
2808 stack_zero(&data->wlod_stack);
2809 stack_save(&data->wlod_stack);
2813 /* Call this whenver the structure of the witness graph changes. */
2815 witness_increment_graph_generation(void)
2818 if (witness_cold == 0)
2819 mtx_assert(&w_mtx, MA_OWNED);
2825 _witness_debugger(int cond, const char *msg)
2828 if (witness_trace && cond)
2830 if (witness_kdb && cond)
2831 kdb_enter(KDB_WHY_WITNESS, msg);