1 /*-
2 * Copyright (c) 2008 Isilon Systems, Inc.
3 * Copyright (c) 2008 Ilya Maykov <ivmaykov@gmail.com>
4 * Copyright (c) 1998 Berkeley Software Design, Inc.
5 * All rights reserved.
6 *
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
9 * are met:
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
17 * written permission.
18 *
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
29 * SUCH DAMAGE.
30 *
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 $
33 */
34
35 /*
36 * Implementation of the `witness' lock verifier. Originally implemented for
37 * mutexes in BSD/OS. Extended to handle generic lock objects and lock
38 * classes in FreeBSD.
39 */
40
41 /*
42 * Main Entry: witness
43 * Pronunciation: 'wit-n&s
44 * Function: noun
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
57 * life -- Pilot>
58 * 6 capitalized : a member of the Jehovah's Witnesses
59 */
60
61 /*
62 * Special rules concerning Giant and lock orders:
63 *
64 * 1) Giant must be acquired before any other mutexes. Stated another way,
65 * no other mutex may be held when Giant is acquired.
66 *
67 * 2) Giant must be released when blocking on a sleepable lock.
68 *
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
72 * 2).
73 *
74 * 3) Giant may be acquired before or after sleepable locks.
75 *
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.
85 */
86
87 #include <sys/cdefs.h>
88 __FBSDID("$FreeBSD: releng/8.3/sys/kern/subr_witness.c 219107 2011-02-28 23:28:35Z rwatson $");
89
90 #include "opt_ddb.h"
91 #include "opt_hwpmc_hooks.h"
92 #include "opt_stack.h"
93 #include "opt_witness.h"
94
95 #include <sys/param.h>
96 #include <sys/bus.h>
97 #include <sys/kdb.h>
98 #include <sys/kernel.h>
99 #include <sys/ktr.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>
110
111 #ifdef DDB
112 #include <ddb/ddb.h>
113 #endif
114
115 #include <machine/stdarg.h>
116
117 #if !defined(DDB) && !defined(STACK)
118 #error "DDB or STACK options are required for WITNESS"
119 #endif
120
121 /* Note that these traces do not work with KTR_ALQ. */
122 #if 0
123 #define KTR_WITNESS KTR_SUBSYS
124 #else
125 #define KTR_WITNESS 0
126 #endif
127
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. */
131
132 /* Define this to check for blessed mutexes */
133 #undef BLESSING
134
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
139
140 /* Allocate 256 KB of stack data space */
141 #define WITNESS_LO_DATA_COUNT 2048
142
143 /* Prime, gives load factor of ~2 at full load */
144 #define WITNESS_LO_HASH_SIZE 1021
145
146 /*
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.
150 */
151 #define LOCK_NCHILDREN 5
152 #define LOCK_CHILDCOUNT 2048
153
154 #define MAX_W_NAME 64
155
156 #define BADSTACK_SBUF_SIZE (256 * WITNESS_COUNT)
157 #define CYCLEGRAPH_SBUF_SIZE 8192
158 #define FULLGRAPH_SBUF_SIZE 32768
159
160 /*
161 * These flags go in the witness relationship matrix and describe the
162 * relationship between any two struct witness objects.
163 */
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
174 * observed. */
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. */
178
179 /* Descendant to ancestor flags */
180 #define WITNESS_DTOA(x) (((x) & WITNESS_RELATED_MASK) >> 2)
181
182 /* Ancestor to descendant flags */
183 #define WITNESS_ATOD(x) (((x) & WITNESS_RELATED_MASK) << 2)
184
185 #define WITNESS_INDEX_ASSERT(i) \
186 MPASS((i) > 0 && (i) <= w_max_used_index && (i) < WITNESS_COUNT)
187
188 MALLOC_DEFINE(M_WITNESS, "Witness", "Witness");
189
190 /*
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.
195 */
196 struct lock_instance {
197 struct lock_object *li_lock;
198 const char *li_file;
199 int li_line;
200 u_int li_flags;
201 };
202
203 /*
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.
212 */
213 struct lock_list_entry {
214 struct lock_list_entry *ll_next;
215 struct lock_instance ll_children[LOCK_NCHILDREN];
216 u_int ll_count;
217 };
218
219 /*
220 * The main witness structure. One of these per named lock type in the system
221 * (for example, "vnode interlock").
222 */
223 struct witness {
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 */
232 uint32_t w_refcount;
233 uint16_t w_num_ancestors; /* direct/indirect
234 * ancestor count */
235 uint16_t w_num_descendants; /* direct/indirect
236 * descendant count */
237 int16_t w_ddb_level;
238 unsigned w_displayed:1;
239 unsigned w_reversed:1;
240 };
241
242 STAILQ_HEAD(witness_list, witness);
243
244 /*
245 * The witness hash table. Keys are witness names (const char *), elements are
246 * witness objects (struct witness *).
247 */
248 struct witness_hash {
249 struct witness *wh_array[WITNESS_HASH_SIZE];
250 uint32_t wh_size;
251 uint32_t wh_count;
252 };
253
254 /*
255 * Key type for the lock order data hash table.
256 */
257 struct witness_lock_order_key {
258 uint16_t from;
259 uint16_t to;
260 };
261
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;
266 };
267
268 /*
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).
272 */
273 struct witness_lock_order_hash {
274 struct witness_lock_order_data *wloh_array[WITNESS_LO_HASH_SIZE];
275 u_int wloh_size;
276 u_int wloh_count;
277 };
278
279 #ifdef BLESSING
280 struct witness_blessed {
281 const char *b_lock1;
282 const char *b_lock2;
283 };
284 #endif
285
286 struct witness_pendhelp {
287 const char *wh_type;
288 struct lock_object *wh_lock;
289 };
290
291 struct witness_order_list_entry {
292 const char *w_name;
293 struct lock_class *w_class;
294 };
295
296 /*
297 * Returns 0 if one of the locks is a spin lock and the other is not.
298 * Returns 1 otherwise.
299 */
300 static __inline int
301 witness_lock_type_equal(struct witness *w1, struct witness *w2)
302 {
303
304 return ((w1->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) ==
305 (w2->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)));
306 }
307
308 static __inline int
309 witness_lock_order_key_empty(const struct witness_lock_order_key *key)
310 {
311
312 return (key->from == 0 && key->to == 0);
313 }
314
315 static __inline int
316 witness_lock_order_key_equal(const struct witness_lock_order_key *a,
317 const struct witness_lock_order_key *b)
318 {
319
320 return (a->from == b->from && a->to == b->to);
321 }
322
323 static int _isitmyx(struct witness *w1, struct witness *w2, int rmask,
324 const char *fname);
325 #ifdef KDB
326 static void _witness_debugger(int cond, const char *msg);
327 #endif
328 static void adopt(struct witness *parent, struct witness *child);
329 #ifdef BLESSING
330 static int blessed(struct witness *, struct witness *);
331 #endif
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);
344 #ifdef DDB
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);
353 #endif
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);
373
374 #ifdef KDB
375 #define witness_debugger(c) _witness_debugger(c, __func__)
376 #else
377 #define witness_debugger(c)
378 #endif
379
380 SYSCTL_NODE(_debug, OID_AUTO, witness, CTLFLAG_RW, NULL, "Witness Locking");
381
382 /*
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.
388 */
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");
393
394 #ifdef KDB
395 /*
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.
400 */
401 #ifdef WITNESS_KDB
402 int witness_kdb = 1;
403 #else
404 int witness_kdb = 0;
405 #endif
406 TUNABLE_INT("debug.witness.kdb", &witness_kdb);
407 SYSCTL_INT(_debug_witness, OID_AUTO, kdb, CTLFLAG_RW, &witness_kdb, 0, "");
408
409 /*
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.
414 */
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, "");
418 #endif /* KDB */
419
420 #ifdef WITNESS_SKIPSPIN
421 int witness_skipspin = 1;
422 #else
423 int witness_skipspin = 0;
424 #endif
425 TUNABLE_INT("debug.witness.skipspin", &witness_skipspin);
426 SYSCTL_INT(_debug_witness, OID_AUTO, skipspin, CTLFLAG_RDTUN, &witness_skipspin,
427 0, "");
428
429 /*
430 * Call this to print out the relations between locks.
431 */
432 SYSCTL_PROC(_debug_witness, OID_AUTO, fullgraph, CTLTYPE_STRING | CTLFLAG_RD,
433 NULL, 0, sysctl_debug_witness_fullgraph, "A", "Show locks relation graphs");
434
435 /*
436 * Call this to print out the witness faulty stacks.
437 */
438 SYSCTL_PROC(_debug_witness, OID_AUTO, badstacks, CTLTYPE_STRING | CTLFLAG_RD,
439 NULL, 0, sysctl_debug_witness_badstacks, "A", "Show bad witness stacks");
440
441 static struct mtx w_mtx;
442
443 /* w_list */
444 static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free);
445 static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all);
446
447 /* w_typelist */
448 static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin);
449 static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep);
450
451 /* lock list */
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;
455
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,
460 "");
461
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. */
466
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";
475
476
477 static struct witness_order_list_entry order_lists[] = {
478 /*
479 * sx locks
480 */
481 { "proctree", &lock_class_sx },
482 { "allproc", &lock_class_sx },
483 { "allprison", &lock_class_sx },
484 { NULL, NULL },
485 /*
486 * Various mutexes
487 */
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 },
495 #ifdef HWPMC_HOOKS
496 { "pmc-sleep", &lock_class_mtx_sleep },
497 #endif
498 { NULL, NULL },
499 /*
500 * Sockets
501 */
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 },
506 { NULL, NULL },
507 /*
508 * Routing
509 */
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 },
514 { NULL, NULL },
515 /*
516 * IPv4 multicast:
517 * protocol locks before interface locks, after UDP locks.
518 */
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 },
523 { NULL, NULL },
524 /*
525 * IPv6 multicast:
526 * protocol locks before interface locks, after UDP locks.
527 */
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 },
532 { NULL, NULL },
533 /*
534 * UNIX Domain Sockets
535 */
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 },
540 { NULL, NULL },
541 /*
542 * UDP/IP
543 */
544 { "udp", &lock_class_rw },
545 { "udpinp", &lock_class_rw },
546 { "so_snd", &lock_class_mtx_sleep },
547 { NULL, NULL },
548 /*
549 * TCP/IP
550 */
551 { "tcp", &lock_class_rw },
552 { "tcpinp", &lock_class_rw },
553 { "so_snd", &lock_class_mtx_sleep },
554 { NULL, NULL },
555 /*
556 * SLIP
557 */
558 { "slip_mtx", &lock_class_mtx_sleep },
559 { "slip sc_mtx", &lock_class_mtx_sleep },
560 { NULL, NULL },
561 /*
562 * netatalk
563 */
564 { "ddp_list_mtx", &lock_class_mtx_sleep },
565 { "ddp_mtx", &lock_class_mtx_sleep },
566 { NULL, NULL },
567 /*
568 * BPF
569 */
570 { "bpf global lock", &lock_class_mtx_sleep },
571 { "bpf interface lock", &lock_class_mtx_sleep },
572 { "bpf cdev lock", &lock_class_mtx_sleep },
573 { NULL, NULL },
574 /*
575 * NFS server
576 */
577 { "nfsd_mtx", &lock_class_mtx_sleep },
578 { "so_snd", &lock_class_mtx_sleep },
579 { NULL, NULL },
580
581 /*
582 * IEEE 802.11
583 */
584 { "802.11 com lock", &lock_class_mtx_sleep},
585 { NULL, NULL },
586 /*
587 * Network drivers
588 */
589 { "network driver", &lock_class_mtx_sleep},
590 { NULL, NULL },
591
592 /*
593 * Netgraph
594 */
595 { "ng_node", &lock_class_mtx_sleep },
596 { "ng_worklist", &lock_class_mtx_sleep },
597 { NULL, NULL },
598 /*
599 * CDEV
600 */
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 },
605 { NULL, NULL },
606 /*
607 * kqueue/VFS interaction
608 */
609 { "kqueue", &lock_class_mtx_sleep },
610 { "struct mount mtx", &lock_class_mtx_sleep },
611 { "vnode interlock", &lock_class_mtx_sleep },
612 { NULL, NULL },
613 /*
614 * ZFS locking
615 */
616 { "dn->dn_mtx", &lock_class_sx },
617 { "dr->dt.di.dr_mtx", &lock_class_sx },
618 { "db->db_mtx", &lock_class_sx },
619 { NULL, NULL },
620 /*
621 * spin locks
622 */
623 #ifdef SMP
624 { "ap boot", &lock_class_mtx_spin },
625 #endif
626 { "rm.mutex_mtx", &lock_class_mtx_spin },
627 { "sio", &lock_class_mtx_spin },
628 { "scrlock", &lock_class_mtx_spin },
629 #ifdef __i386__
630 { "cy", &lock_class_mtx_spin },
631 #endif
632 #ifdef __sparc64__
633 { "pcib_mtx", &lock_class_mtx_spin },
634 { "rtc_mtx", &lock_class_mtx_spin },
635 #endif
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 },
640 #ifdef HWPMC_HOOKS
641 { "pmc-per-proc", &lock_class_mtx_spin },
642 #endif
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 },
655 #ifdef SMP
656 { "smp rendezvous", &lock_class_mtx_spin },
657 #endif
658 #ifdef __powerpc__
659 { "tlb0", &lock_class_mtx_spin },
660 #endif
661 /*
662 * leaf locks
663 */
664 { "intrcnt", &lock_class_mtx_spin },
665 { "icu", &lock_class_mtx_spin },
666 #if defined(SMP) && defined(__sparc64__)
667 { "ipi", &lock_class_mtx_spin },
668 #endif
669 #ifdef __i386__
670 { "allpmaps", &lock_class_mtx_spin },
671 { "descriptor tables", &lock_class_mtx_spin },
672 #endif
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 },
678 #ifdef __ia64__
679 { "MCA spin lock", &lock_class_mtx_spin },
680 #endif
681 #if defined(__i386__) || defined(__amd64__)
682 { "pcicfg", &lock_class_mtx_spin },
683 { "NDIS thread lock", &lock_class_mtx_spin },
684 #endif
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 },
690 #ifdef HWPMC_HOOKS
691 { "pmc-leaf", &lock_class_mtx_spin },
692 #endif
693 { "blocked lock", &lock_class_mtx_spin },
694 { NULL, NULL },
695 { NULL, NULL }
696 };
697
698 #ifdef BLESSING
699 /*
700 * Pairs of locks which have been blessed
701 * Don't complain about order problems with blessed locks
702 */
703 static struct witness_blessed blessed_list[] = {
704 };
705 static int blessed_count =
706 sizeof(blessed_list) / sizeof(struct witness_blessed);
707 #endif
708
709 /*
710 * This global is set to 0 once it becomes safe to use the witness code.
711 */
712 static int witness_cold = 1;
713
714 /*
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.
717 */
718 static int witness_spin_warn = 0;
719
720 /*
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.
724 */
725 static void
726 witness_initialize(void *dummy __unused)
727 {
728 struct lock_object *lock;
729 struct witness_order_list_entry *order;
730 struct witness *w, *w1;
731 int i;
732
733 w_data = malloc(sizeof (struct witness) * WITNESS_COUNT, M_WITNESS,
734 M_NOWAIT | M_ZERO);
735
736 /*
737 * We have to release Giant before initializing its witness
738 * structure so that WITNESS doesn't get confused.
739 */
740 mtx_unlock(&Giant);
741 mtx_assert(&Giant, MA_NOTOWNED);
742
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--) {
747 w = &w_data[i];
748 memset(w, 0, sizeof(*w));
749 w_data[i].w_index = i; /* Witness index never changes. */
750 witness_free(w);
751 }
752 KASSERT(STAILQ_FIRST(&w_free)->w_index == 0,
753 ("%s: Invalid list of free witness objects", __func__));
754
755 /* Witness with index 0 is not used to aid in debugging. */
756 STAILQ_REMOVE_HEAD(&w_free, w_list);
757 w_free_cnt--;
758
759 memset(w_rmatrix, 0,
760 (sizeof(**w_rmatrix) * (WITNESS_COUNT+1) * (WITNESS_COUNT+1)));
761
762 for (i = 0; i < LOCK_CHILDCOUNT; i++)
763 witness_lock_list_free(&w_locklistdata[i]);
764 witness_init_hash_tables();
765
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);
769 if (w == NULL)
770 continue;
771 w->w_file = "order list";
772 for (order++; order->w_name != NULL; order++) {
773 w1 = enroll(order->w_name, order->w_class);
774 if (w1 == NULL)
775 continue;
776 w1->w_file = "order list";
777 itismychild(w, w1);
778 w = w1;
779 }
780 }
781 witness_spin_warn = 1;
782
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,
790 LOCK_CLASS(lock));
791 }
792
793 /* Mark the witness code as being ready for use. */
794 witness_cold = 0;
795
796 mtx_lock(&Giant);
797 }
798 SYSINIT(witness_init, SI_SUB_WITNESS, SI_ORDER_FIRST, witness_initialize,
799 NULL);
800
801 void
802 witness_init(struct lock_object *lock, const char *type)
803 {
804 struct lock_class *class;
805
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);
820
821 /*
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
826 * the lock now.
827 */
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",
836 __func__);
837 } else
838 lock->lo_witness = enroll(type, class);
839 }
840
841 void
842 witness_destroy(struct lock_object *lock)
843 {
844 struct lock_class *class;
845 struct witness *w;
846
847 class = LOCK_CLASS(lock);
848
849 if (witness_cold)
850 panic("lock (%s) %s destroyed while witness_cold",
851 class->lc_name, lock->lo_name);
852
853 /* XXX: need to verify that no one holds the lock */
854 if ((lock->lo_flags & LO_WITNESS) == 0 || lock->lo_witness == NULL)
855 return;
856 w = lock->lo_witness;
857
858 mtx_lock_spin(&w_mtx);
859 MPASS(w->w_refcount > 0);
860 w->w_refcount--;
861
862 if (w->w_refcount == 0)
863 depart(w);
864 mtx_unlock_spin(&w_mtx);
865 }
866
867 #ifdef DDB
868 static void
869 witness_ddb_compute_levels(void)
870 {
871 struct witness *w;
872
873 /*
874 * First clear all levels.
875 */
876 STAILQ_FOREACH(w, &w_all, w_list)
877 w->w_ddb_level = -1;
878
879 /*
880 * Look for locks with no parents and level all their descendants.
881 */
882 STAILQ_FOREACH(w, &w_all, w_list) {
883
884 /* If the witness has ancestors (is not a root), skip it. */
885 if (w->w_num_ancestors > 0)
886 continue;
887 witness_ddb_level_descendants(w, 0);
888 }
889 }
890
891 static void
892 witness_ddb_level_descendants(struct witness *w, int l)
893 {
894 int i;
895
896 if (w->w_ddb_level >= l)
897 return;
898
899 w->w_ddb_level = l;
900 l++;
901
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);
905 }
906 }
907
908 static void
909 witness_ddb_display_descendants(int(*prnt)(const char *fmt, ...),
910 struct witness *w, int indent)
911 {
912 int i;
913
914 for (i = 0; i < indent; i++)
915 prnt(" ");
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");
921 return;
922 }
923 w->w_displayed = 1;
924 if (w->w_file != NULL && w->w_line != 0)
925 prnt(" -- last acquired @ %s:%d\n", w->w_file,
926 w->w_line);
927 else
928 prnt(" -- never acquired\n");
929 indent++;
930 WITNESS_INDEX_ASSERT(w->w_index);
931 for (i = 1; i <= w_max_used_index; i++) {
932 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
933 witness_ddb_display_descendants(prnt, &w_data[i],
934 indent);
935 }
936 }
937
938 static void
939 witness_ddb_display_list(int(*prnt)(const char *fmt, ...),
940 struct witness_list *list)
941 {
942 struct witness *w;
943
944 STAILQ_FOREACH(w, list, w_typelist) {
945 if (w->w_file == NULL || w->w_ddb_level > 0)
946 continue;
947
948 /* This lock has no anscestors - display its descendants. */
949 witness_ddb_display_descendants(prnt, w, 0);
950 }
951 }
952
953 static void
954 witness_ddb_display(int(*prnt)(const char *fmt, ...))
955 {
956 struct witness *w;
957
958 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
959 witness_ddb_compute_levels();
960
961 /* Clear all the displayed flags. */
962 STAILQ_FOREACH(w, &w_all, w_list)
963 w->w_displayed = 0;
964
965 /*
966 * First, handle sleep locks which have been acquired at least
967 * once.
968 */
969 prnt("Sleep locks:\n");
970 witness_ddb_display_list(prnt, &w_sleep);
971
972 /*
973 * Now do spin locks which have been acquired at least once.
974 */
975 prnt("\nSpin locks:\n");
976 witness_ddb_display_list(prnt, &w_spin);
977
978 /*
979 * Finally, any locks which have not been acquired yet.
980 */
981 prnt("\nLocks which were never acquired:\n");
982 STAILQ_FOREACH(w, &w_all, w_list) {
983 if (w->w_file != NULL || w->w_refcount == 0)
984 continue;
985 prnt("%s (type: %s, depth: %d)\n", w->w_name,
986 w->w_class->lc_name, w->w_ddb_level);
987 }
988 }
989 #endif /* DDB */
990
991 /* Trim useless garbage from filenames. */
992 static const char *
993 fixup_filename(const char *file)
994 {
995
996 if (file == NULL)
997 return (NULL);
998 while (strncmp(file, "../", 3) == 0)
999 file += 3;
1000 return (file);
1001 }
1002
1003 int
1004 witness_defineorder(struct lock_object *lock1, struct lock_object *lock2)
1005 {
1006
1007 if (witness_watch == -1 || panicstr != NULL)
1008 return (0);
1009
1010 /* Require locks that witness knows about. */
1011 if (lock1 == NULL || lock1->lo_witness == NULL || lock2 == NULL ||
1012 lock2->lo_witness == NULL)
1013 return (EINVAL);
1014
1015 mtx_assert(&w_mtx, MA_NOTOWNED);
1016 mtx_lock_spin(&w_mtx);
1017
1018 /*
1019 * If we already have either an explicit or implied lock order that
1020 * is the other way around, then return an error.
1021 */
1022 if (witness_watch &&
1023 isitmydescendant(lock2->lo_witness, lock1->lo_witness)) {
1024 mtx_unlock_spin(&w_mtx);
1025 return (EDOOFUS);
1026 }
1027
1028 /* Try to add the new order. */
1029 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1030 lock2->lo_witness->w_name, lock1->lo_witness->w_name);
1031 itismychild(lock1->lo_witness, lock2->lo_witness);
1032 mtx_unlock_spin(&w_mtx);
1033 return (0);
1034 }
1035
1036 void
1037 witness_checkorder(struct lock_object *lock, int flags, const char *file,
1038 int line, struct lock_object *interlock)
1039 {
1040 struct lock_list_entry *lock_list, *lle;
1041 struct lock_instance *lock1, *lock2, *plock;
1042 struct lock_class *class;
1043 struct witness *w, *w1;
1044 struct thread *td;
1045 int i, j;
1046
1047 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL ||
1048 panicstr != NULL)
1049 return;
1050
1051 w = lock->lo_witness;
1052 class = LOCK_CLASS(lock);
1053 td = curthread;
1054 file = fixup_filename(file);
1055
1056 if (class->lc_flags & LC_SLEEPLOCK) {
1057
1058 /*
1059 * Since spin locks include a critical section, this check
1060 * implicitly enforces a lock order of all sleep locks before
1061 * all spin locks.
1062 */
1063 if (td->td_critnest != 0 && !kdb_active)
1064 panic("blockable sleep lock (%s) %s @ %s:%d",
1065 class->lc_name, lock->lo_name, file, line);
1066
1067 /*
1068 * If this is the first lock acquired then just return as
1069 * no order checking is needed.
1070 */
1071 lock_list = td->td_sleeplocks;
1072 if (lock_list == NULL || lock_list->ll_count == 0)
1073 return;
1074 } else {
1075
1076 /*
1077 * If this is the first lock, just return as no order
1078 * checking is needed. Avoid problems with thread
1079 * migration pinning the thread while checking if
1080 * spinlocks are held. If at least one spinlock is held
1081 * the thread is in a safe path and it is allowed to
1082 * unpin it.
1083 */
1084 sched_pin();
1085 lock_list = PCPU_GET(spinlocks);
1086 if (lock_list == NULL || lock_list->ll_count == 0) {
1087 sched_unpin();
1088 return;
1089 }
1090 sched_unpin();
1091 }
1092
1093 /*
1094 * Check to see if we are recursing on a lock we already own. If
1095 * so, make sure that we don't mismatch exclusive and shared lock
1096 * acquires.
1097 */
1098 lock1 = find_instance(lock_list, lock);
1099 if (lock1 != NULL) {
1100 if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
1101 (flags & LOP_EXCLUSIVE) == 0) {
1102 printf("shared lock of (%s) %s @ %s:%d\n",
1103 class->lc_name, lock->lo_name, file, line);
1104 printf("while exclusively locked from %s:%d\n",
1105 lock1->li_file, lock1->li_line);
1106 panic("share->excl");
1107 }
1108 if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
1109 (flags & LOP_EXCLUSIVE) != 0) {
1110 printf("exclusive lock of (%s) %s @ %s:%d\n",
1111 class->lc_name, lock->lo_name, file, line);
1112 printf("while share locked from %s:%d\n",
1113 lock1->li_file, lock1->li_line);
1114 panic("excl->share");
1115 }
1116 return;
1117 }
1118
1119 /*
1120 * Find the previously acquired lock, but ignore interlocks.
1121 */
1122 plock = &lock_list->ll_children[lock_list->ll_count - 1];
1123 if (interlock != NULL && plock->li_lock == interlock) {
1124 if (lock_list->ll_count > 1)
1125 plock =
1126 &lock_list->ll_children[lock_list->ll_count - 2];
1127 else {
1128 lle = lock_list->ll_next;
1129
1130 /*
1131 * The interlock is the only lock we hold, so
1132 * simply return.
1133 */
1134 if (lle == NULL)
1135 return;
1136 plock = &lle->ll_children[lle->ll_count - 1];
1137 }
1138 }
1139
1140 /*
1141 * Try to perform most checks without a lock. If this succeeds we
1142 * can skip acquiring the lock and return success.
1143 */
1144 w1 = plock->li_lock->lo_witness;
1145 if (witness_lock_order_check(w1, w))
1146 return;
1147
1148 /*
1149 * Check for duplicate locks of the same type. Note that we only
1150 * have to check for this on the last lock we just acquired. Any
1151 * other cases will be caught as lock order violations.
1152 */
1153 mtx_lock_spin(&w_mtx);
1154 witness_lock_order_add(w1, w);
1155 if (w1 == w) {
1156 i = w->w_index;
1157 if (!(lock->lo_flags & LO_DUPOK) && !(flags & LOP_DUPOK) &&
1158 !(w_rmatrix[i][i] & WITNESS_REVERSAL)) {
1159 w_rmatrix[i][i] |= WITNESS_REVERSAL;
1160 w->w_reversed = 1;
1161 mtx_unlock_spin(&w_mtx);
1162 printf(
1163 "acquiring duplicate lock of same type: \"%s\"\n",
1164 w->w_name);
1165 printf(" 1st %s @ %s:%d\n", plock->li_lock->lo_name,
1166 plock->li_file, plock->li_line);
1167 printf(" 2nd %s @ %s:%d\n", lock->lo_name, file, line);
1168 witness_debugger(1);
1169 } else
1170 mtx_unlock_spin(&w_mtx);
1171 return;
1172 }
1173 mtx_assert(&w_mtx, MA_OWNED);
1174
1175 /*
1176 * If we know that the lock we are acquiring comes after
1177 * the lock we most recently acquired in the lock order tree,
1178 * then there is no need for any further checks.
1179 */
1180 if (isitmychild(w1, w))
1181 goto out;
1182
1183 for (j = 0, lle = lock_list; lle != NULL; lle = lle->ll_next) {
1184 for (i = lle->ll_count - 1; i >= 0; i--, j++) {
1185
1186 MPASS(j < WITNESS_COUNT);
1187 lock1 = &lle->ll_children[i];
1188
1189 /*
1190 * Ignore the interlock the first time we see it.
1191 */
1192 if (interlock != NULL && interlock == lock1->li_lock) {
1193 interlock = NULL;
1194 continue;
1195 }
1196
1197 /*
1198 * If this lock doesn't undergo witness checking,
1199 * then skip it.
1200 */
1201 w1 = lock1->li_lock->lo_witness;
1202 if (w1 == NULL) {
1203 KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
1204 ("lock missing witness structure"));
1205 continue;
1206 }
1207
1208 /*
1209 * If we are locking Giant and this is a sleepable
1210 * lock, then skip it.
1211 */
1212 if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0 &&
1213 lock == &Giant.lock_object)
1214 continue;
1215
1216 /*
1217 * If we are locking a sleepable lock and this lock
1218 * is Giant, then skip it.
1219 */
1220 if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1221 lock1->li_lock == &Giant.lock_object)
1222 continue;
1223
1224 /*
1225 * If we are locking a sleepable lock and this lock
1226 * isn't sleepable, we want to treat it as a lock
1227 * order violation to enfore a general lock order of
1228 * sleepable locks before non-sleepable locks.
1229 */
1230 if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1231 (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
1232 goto reversal;
1233
1234 /*
1235 * If we are locking Giant and this is a non-sleepable
1236 * lock, then treat it as a reversal.
1237 */
1238 if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0 &&
1239 lock == &Giant.lock_object)
1240 goto reversal;
1241
1242 /*
1243 * Check the lock order hierarchy for a reveresal.
1244 */
1245 if (!isitmydescendant(w, w1))
1246 continue;
1247 reversal:
1248
1249 /*
1250 * We have a lock order violation, check to see if it
1251 * is allowed or has already been yelled about.
1252 */
1253 #ifdef BLESSING
1254
1255 /*
1256 * If the lock order is blessed, just bail. We don't
1257 * look for other lock order violations though, which
1258 * may be a bug.
1259 */
1260 if (blessed(w, w1))
1261 goto out;
1262 #endif
1263
1264 /* Bail if this violation is known */
1265 if (w_rmatrix[w1->w_index][w->w_index] & WITNESS_REVERSAL)
1266 goto out;
1267
1268 /* Record this as a violation */
1269 w_rmatrix[w1->w_index][w->w_index] |= WITNESS_REVERSAL;
1270 w_rmatrix[w->w_index][w1->w_index] |= WITNESS_REVERSAL;
1271 w->w_reversed = w1->w_reversed = 1;
1272 witness_increment_graph_generation();
1273 mtx_unlock_spin(&w_mtx);
1274
1275 /*
1276 * Ok, yell about it.
1277 */
1278 if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1279 (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
1280 printf(
1281 "lock order reversal: (sleepable after non-sleepable)\n");
1282 else if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0
1283 && lock == &Giant.lock_object)
1284 printf(
1285 "lock order reversal: (Giant after non-sleepable)\n");
1286 else
1287 printf("lock order reversal:\n");
1288
1289 /*
1290 * Try to locate an earlier lock with
1291 * witness w in our list.
1292 */
1293 do {
1294 lock2 = &lle->ll_children[i];
1295 MPASS(lock2->li_lock != NULL);
1296 if (lock2->li_lock->lo_witness == w)
1297 break;
1298 if (i == 0 && lle->ll_next != NULL) {
1299 lle = lle->ll_next;
1300 i = lle->ll_count - 1;
1301 MPASS(i >= 0 && i < LOCK_NCHILDREN);
1302 } else
1303 i--;
1304 } while (i >= 0);
1305 if (i < 0) {
1306 printf(" 1st %p %s (%s) @ %s:%d\n",
1307 lock1->li_lock, lock1->li_lock->lo_name,
1308 w1->w_name, lock1->li_file, lock1->li_line);
1309 printf(" 2nd %p %s (%s) @ %s:%d\n", lock,
1310 lock->lo_name, w->w_name, file, line);
1311 } else {
1312 printf(" 1st %p %s (%s) @ %s:%d\n",
1313 lock2->li_lock, lock2->li_lock->lo_name,
1314 lock2->li_lock->lo_witness->w_name,
1315 lock2->li_file, lock2->li_line);
1316 printf(" 2nd %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(" 3rd %p %s (%s) @ %s:%d\n", lock,
1320 lock->lo_name, w->w_name, file, line);
1321 }
1322 witness_debugger(1);
1323 return;
1324 }
1325 }
1326
1327 /*
1328 * If requested, build a new lock order. However, don't build a new
1329 * relationship between a sleepable lock and Giant if it is in the
1330 * wrong direction. The correct lock order is that sleepable locks
1331 * always come before Giant.
1332 */
1333 if (flags & LOP_NEWORDER &&
1334 !(plock->li_lock == &Giant.lock_object &&
1335 (lock->lo_flags & LO_SLEEPABLE) != 0)) {
1336 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1337 w->w_name, plock->li_lock->lo_witness->w_name);
1338 itismychild(plock->li_lock->lo_witness, w);
1339 }
1340 out:
1341 mtx_unlock_spin(&w_mtx);
1342 }
1343
1344 void
1345 witness_lock(struct lock_object *lock, int flags, const char *file, int line)
1346 {
1347 struct lock_list_entry **lock_list, *lle;
1348 struct lock_instance *instance;
1349 struct witness *w;
1350 struct thread *td;
1351
1352 if (witness_cold || witness_watch == -1 || lock->lo_witness == NULL ||
1353 panicstr != NULL)
1354 return;
1355 w = lock->lo_witness;
1356 td = curthread;
1357 file = fixup_filename(file);
1358
1359 /* Determine lock list for this lock. */
1360 if (LOCK_CLASS(lock)->lc_flags & LC_SLEEPLOCK)
1361 lock_list = &td->td_sleeplocks;
1362 else
1363 lock_list = PCPU_PTR(spinlocks);
1364
1365 /* Check to see if we are recursing on a lock we already own. */
1366 instance = find_instance(*lock_list, lock);
1367 if (instance != NULL) {
1368 instance->li_flags++;
1369 CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
1370 td->td_proc->p_pid, lock->lo_name,
1371 instance->li_flags & LI_RECURSEMASK);
1372 instance->li_file = file;
1373 instance->li_line = line;
1374 return;
1375 }
1376
1377 /* Update per-witness last file and line acquire. */
1378 w->w_file = file;
1379 w->w_line = line;
1380
1381 /* Find the next open lock instance in the list and fill it. */
1382 lle = *lock_list;
1383 if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
1384 lle = witness_lock_list_get();
1385 if (lle == NULL)
1386 return;
1387 lle->ll_next = *lock_list;
1388 CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
1389 td->td_proc->p_pid, lle);
1390 *lock_list = lle;
1391 }
1392 instance = &lle->ll_children[lle->ll_count++];
1393 instance->li_lock = lock;
1394 instance->li_line = line;
1395 instance->li_file = file;
1396 if ((flags & LOP_EXCLUSIVE) != 0)
1397 instance->li_flags = LI_EXCLUSIVE;
1398 else
1399 instance->li_flags = 0;
1400 CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
1401 td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
1402 }
1403
1404 void
1405 witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
1406 {
1407 struct lock_instance *instance;
1408 struct lock_class *class;
1409
1410 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1411 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
1412 return;
1413 class = LOCK_CLASS(lock);
1414 file = fixup_filename(file);
1415 if (witness_watch) {
1416 if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1417 panic("upgrade of non-upgradable lock (%s) %s @ %s:%d",
1418 class->lc_name, lock->lo_name, file, line);
1419 if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1420 panic("upgrade of non-sleep lock (%s) %s @ %s:%d",
1421 class->lc_name, lock->lo_name, file, line);
1422 }
1423 instance = find_instance(curthread->td_sleeplocks, lock);
1424 if (instance == NULL)
1425 panic("upgrade of unlocked lock (%s) %s @ %s:%d",
1426 class->lc_name, lock->lo_name, file, line);
1427 if (witness_watch) {
1428 if ((instance->li_flags & LI_EXCLUSIVE) != 0)
1429 panic("upgrade of exclusive lock (%s) %s @ %s:%d",
1430 class->lc_name, lock->lo_name, file, line);
1431 if ((instance->li_flags & LI_RECURSEMASK) != 0)
1432 panic("upgrade of recursed lock (%s) %s r=%d @ %s:%d",
1433 class->lc_name, lock->lo_name,
1434 instance->li_flags & LI_RECURSEMASK, file, line);
1435 }
1436 instance->li_flags |= LI_EXCLUSIVE;
1437 }
1438
1439 void
1440 witness_downgrade(struct lock_object *lock, int flags, const char *file,
1441 int line)
1442 {
1443 struct lock_instance *instance;
1444 struct lock_class *class;
1445
1446 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1447 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
1448 return;
1449 class = LOCK_CLASS(lock);
1450 file = fixup_filename(file);
1451 if (witness_watch) {
1452 if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1453 panic("downgrade of non-upgradable lock (%s) %s @ %s:%d",
1454 class->lc_name, lock->lo_name, file, line);
1455 if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1456 panic("downgrade of non-sleep lock (%s) %s @ %s:%d",
1457 class->lc_name, lock->lo_name, file, line);
1458 }
1459 instance = find_instance(curthread->td_sleeplocks, lock);
1460 if (instance == NULL)
1461 panic("downgrade of unlocked lock (%s) %s @ %s:%d",
1462 class->lc_name, lock->lo_name, file, line);
1463 if (witness_watch) {
1464 if ((instance->li_flags & LI_EXCLUSIVE) == 0)
1465 panic("downgrade of shared lock (%s) %s @ %s:%d",
1466 class->lc_name, lock->lo_name, file, line);
1467 if ((instance->li_flags & LI_RECURSEMASK) != 0)
1468 panic("downgrade of recursed lock (%s) %s r=%d @ %s:%d",
1469 class->lc_name, lock->lo_name,
1470 instance->li_flags & LI_RECURSEMASK, file, line);
1471 }
1472 instance->li_flags &= ~LI_EXCLUSIVE;
1473 }
1474
1475 void
1476 witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
1477 {
1478 struct lock_list_entry **lock_list, *lle;
1479 struct lock_instance *instance;
1480 struct lock_class *class;
1481 struct thread *td;
1482 register_t s;
1483 int i, j;
1484
1485 if (witness_cold || lock->lo_witness == NULL || panicstr != NULL)
1486 return;
1487 td = curthread;
1488 class = LOCK_CLASS(lock);
1489 file = fixup_filename(file);
1490
1491 /* Find lock instance associated with this lock. */
1492 if (class->lc_flags & LC_SLEEPLOCK)
1493 lock_list = &td->td_sleeplocks;
1494 else
1495 lock_list = PCPU_PTR(spinlocks);
1496 lle = *lock_list;
1497 for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
1498 for (i = 0; i < (*lock_list)->ll_count; i++) {
1499 instance = &(*lock_list)->ll_children[i];
1500 if (instance->li_lock == lock)
1501 goto found;
1502 }
1503
1504 /*
1505 * When disabling WITNESS through witness_watch we could end up in
1506 * having registered locks in the td_sleeplocks queue.
1507 * We have to make sure we flush these queues, so just search for
1508 * eventual register locks and remove them.
1509 */
1510 if (witness_watch > 0)
1511 panic("lock (%s) %s not locked @ %s:%d", class->lc_name,
1512 lock->lo_name, file, line);
1513 else
1514 return;
1515 found:
1516
1517 /* First, check for shared/exclusive mismatches. */
1518 if ((instance->li_flags & LI_EXCLUSIVE) != 0 && witness_watch > 0 &&
1519 (flags & LOP_EXCLUSIVE) == 0) {
1520 printf("shared unlock of (%s) %s @ %s:%d\n", class->lc_name,
1521 lock->lo_name, file, line);
1522 printf("while exclusively locked from %s:%d\n",
1523 instance->li_file, instance->li_line);
1524 panic("excl->ushare");
1525 }
1526 if ((instance->li_flags & LI_EXCLUSIVE) == 0 && witness_watch > 0 &&
1527 (flags & LOP_EXCLUSIVE) != 0) {
1528 printf("exclusive unlock of (%s) %s @ %s:%d\n", class->lc_name,
1529 lock->lo_name, file, line);
1530 printf("while share locked from %s:%d\n", instance->li_file,
1531 instance->li_line);
1532 panic("share->uexcl");
1533 }
1534 /* If we are recursed, unrecurse. */
1535 if ((instance->li_flags & LI_RECURSEMASK) > 0) {
1536 CTR4(KTR_WITNESS, "%s: pid %d unrecursed on %s r=%d", __func__,
1537 td->td_proc->p_pid, instance->li_lock->lo_name,
1538 instance->li_flags);
1539 instance->li_flags--;
1540 return;
1541 }
1542 /* The lock is now being dropped, check for NORELEASE flag */
1543 if ((instance->li_flags & LI_NORELEASE) != 0 && witness_watch > 0) {
1544 printf("forbidden unlock of (%s) %s @ %s:%d\n", class->lc_name,
1545 lock->lo_name, file, line);
1546 panic("lock marked norelease");
1547 }
1548
1549 /* Otherwise, remove this item from the list. */
1550 s = intr_disable();
1551 CTR4(KTR_WITNESS, "%s: pid %d removed %s from lle[%d]", __func__,
1552 td->td_proc->p_pid, instance->li_lock->lo_name,
1553 (*lock_list)->ll_count - 1);
1554 for (j = i; j < (*lock_list)->ll_count - 1; j++)
1555 (*lock_list)->ll_children[j] =
1556 (*lock_list)->ll_children[j + 1];
1557 (*lock_list)->ll_count--;
1558 intr_restore(s);
1559
1560 /*
1561 * In order to reduce contention on w_mtx, we want to keep always an
1562 * head object into lists so that frequent allocation from the
1563 * free witness pool (and subsequent locking) is avoided.
1564 * In order to maintain the current code simple, when the head
1565 * object is totally unloaded it means also that we do not have
1566 * further objects in the list, so the list ownership needs to be
1567 * hand over to another object if the current head needs to be freed.
1568 */
1569 if ((*lock_list)->ll_count == 0) {
1570 if (*lock_list == lle) {
1571 if (lle->ll_next == NULL)
1572 return;
1573 } else
1574 lle = *lock_list;
1575 *lock_list = lle->ll_next;
1576 CTR3(KTR_WITNESS, "%s: pid %d removed lle %p", __func__,
1577 td->td_proc->p_pid, lle);
1578 witness_lock_list_free(lle);
1579 }
1580 }
1581
1582 void
1583 witness_thread_exit(struct thread *td)
1584 {
1585 struct lock_list_entry *lle;
1586 int i, n;
1587
1588 lle = td->td_sleeplocks;
1589 if (lle == NULL || panicstr != NULL)
1590 return;
1591 if (lle->ll_count != 0) {
1592 for (n = 0; lle != NULL; lle = lle->ll_next)
1593 for (i = lle->ll_count - 1; i >= 0; i--) {
1594 if (n == 0)
1595 printf("Thread %p exiting with the following locks held:\n",
1596 td);
1597 n++;
1598 witness_list_lock(&lle->ll_children[i], printf);
1599
1600 }
1601 panic("Thread %p cannot exit while holding sleeplocks\n", td);
1602 }
1603 witness_lock_list_free(lle);
1604 }
1605
1606 /*
1607 * Warn if any locks other than 'lock' are held. Flags can be passed in to
1608 * exempt Giant and sleepable locks from the checks as well. If any
1609 * non-exempt locks are held, then a supplied message is printed to the
1610 * console along with a list of the offending locks. If indicated in the
1611 * flags then a failure results in a panic as well.
1612 */
1613 int
1614 witness_warn(int flags, struct lock_object *lock, const char *fmt, ...)
1615 {
1616 struct lock_list_entry *lock_list, *lle;
1617 struct lock_instance *lock1;
1618 struct thread *td;
1619 va_list ap;
1620 int i, n;
1621
1622 if (witness_cold || witness_watch < 1 || panicstr != NULL)
1623 return (0);
1624 n = 0;
1625 td = curthread;
1626 for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next)
1627 for (i = lle->ll_count - 1; i >= 0; i--) {
1628 lock1 = &lle->ll_children[i];
1629 if (lock1->li_lock == lock)
1630 continue;
1631 if (flags & WARN_GIANTOK &&
1632 lock1->li_lock == &Giant.lock_object)
1633 continue;
1634 if (flags & WARN_SLEEPOK &&
1635 (lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0)
1636 continue;
1637 if (n == 0) {
1638 va_start(ap, fmt);
1639 vprintf(fmt, ap);
1640 va_end(ap);
1641 printf(" with the following");
1642 if (flags & WARN_SLEEPOK)
1643 printf(" non-sleepable");
1644 printf(" locks held:\n");
1645 }
1646 n++;
1647 witness_list_lock(lock1, printf);
1648 }
1649
1650 /*
1651 * Pin the thread in order to avoid problems with thread migration.
1652 * Once that all verifies are passed about spinlocks ownership,
1653 * the thread is in a safe path and it can be unpinned.
1654 */
1655 sched_pin();
1656 lock_list = PCPU_GET(spinlocks);
1657 if (lock_list != NULL && lock_list->ll_count != 0) {
1658 sched_unpin();
1659
1660 /*
1661 * We should only have one spinlock and as long as
1662 * the flags cannot match for this locks class,
1663 * check if the first spinlock is the one curthread
1664 * should hold.
1665 */
1666 lock1 = &lock_list->ll_children[lock_list->ll_count - 1];
1667 if (lock_list->ll_count == 1 && lock_list->ll_next == NULL &&
1668 lock1->li_lock == lock && n == 0)
1669 return (0);
1670
1671 va_start(ap, fmt);
1672 vprintf(fmt, ap);
1673 va_end(ap);
1674 printf(" with the following");
1675 if (flags & WARN_SLEEPOK)
1676 printf(" non-sleepable");
1677 printf(" locks held:\n");
1678 n += witness_list_locks(&lock_list, printf);
1679 } else
1680 sched_unpin();
1681 if (flags & WARN_PANIC && n)
1682 panic("%s", __func__);
1683 else
1684 witness_debugger(n);
1685 return (n);
1686 }
1687
1688 const char *
1689 witness_file(struct lock_object *lock)
1690 {
1691 struct witness *w;
1692
1693 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1694 return ("?");
1695 w = lock->lo_witness;
1696 return (w->w_file);
1697 }
1698
1699 int
1700 witness_line(struct lock_object *lock)
1701 {
1702 struct witness *w;
1703
1704 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1705 return (0);
1706 w = lock->lo_witness;
1707 return (w->w_line);
1708 }
1709
1710 static struct witness *
1711 enroll(const char *description, struct lock_class *lock_class)
1712 {
1713 struct witness *w;
1714 struct witness_list *typelist;
1715
1716 MPASS(description != NULL);
1717
1718 if (witness_watch == -1 || panicstr != NULL)
1719 return (NULL);
1720 if ((lock_class->lc_flags & LC_SPINLOCK)) {
1721 if (witness_skipspin)
1722 return (NULL);
1723 else
1724 typelist = &w_spin;
1725 } else if ((lock_class->lc_flags & LC_SLEEPLOCK))
1726 typelist = &w_sleep;
1727 else
1728 panic("lock class %s is not sleep or spin",
1729 lock_class->lc_name);
1730
1731 mtx_lock_spin(&w_mtx);
1732 w = witness_hash_get(description);
1733 if (w)
1734 goto found;
1735 if ((w = witness_get()) == NULL)
1736 return (NULL);
1737 MPASS(strlen(description) < MAX_W_NAME);
1738 strcpy(w->w_name, description);
1739 w->w_class = lock_class;
1740 w->w_refcount = 1;
1741 STAILQ_INSERT_HEAD(&w_all, w, w_list);
1742 if (lock_class->lc_flags & LC_SPINLOCK) {
1743 STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
1744 w_spin_cnt++;
1745 } else if (lock_class->lc_flags & LC_SLEEPLOCK) {
1746 STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
1747 w_sleep_cnt++;
1748 }
1749
1750 /* Insert new witness into the hash */
1751 witness_hash_put(w);
1752 witness_increment_graph_generation();
1753 mtx_unlock_spin(&w_mtx);
1754 return (w);
1755 found:
1756 w->w_refcount++;
1757 mtx_unlock_spin(&w_mtx);
1758 if (lock_class != w->w_class)
1759 panic(
1760 "lock (%s) %s does not match earlier (%s) lock",
1761 description, lock_class->lc_name,
1762 w->w_class->lc_name);
1763 return (w);
1764 }
1765
1766 static void
1767 depart(struct witness *w)
1768 {
1769 struct witness_list *list;
1770
1771 MPASS(w->w_refcount == 0);
1772 if (w->w_class->lc_flags & LC_SLEEPLOCK) {
1773 list = &w_sleep;
1774 w_sleep_cnt--;
1775 } else {
1776 list = &w_spin;
1777 w_spin_cnt--;
1778 }
1779 /*
1780 * Set file to NULL as it may point into a loadable module.
1781 */
1782 w->w_file = NULL;
1783 w->w_line = 0;
1784 witness_increment_graph_generation();
1785 }
1786
1787
1788 static void
1789 adopt(struct witness *parent, struct witness *child)
1790 {
1791 int pi, ci, i, j;
1792
1793 if (witness_cold == 0)
1794 mtx_assert(&w_mtx, MA_OWNED);
1795
1796 /* If the relationship is already known, there's no work to be done. */
1797 if (isitmychild(parent, child))
1798 return;
1799
1800 /* When the structure of the graph changes, bump up the generation. */
1801 witness_increment_graph_generation();
1802
1803 /*
1804 * The hard part ... create the direct relationship, then propagate all
1805 * indirect relationships.
1806 */
1807 pi = parent->w_index;
1808 ci = child->w_index;
1809 WITNESS_INDEX_ASSERT(pi);
1810 WITNESS_INDEX_ASSERT(ci);
1811 MPASS(pi != ci);
1812 w_rmatrix[pi][ci] |= WITNESS_PARENT;
1813 w_rmatrix[ci][pi] |= WITNESS_CHILD;
1814
1815 /*
1816 * If parent was not already an ancestor of child,
1817 * then we increment the descendant and ancestor counters.
1818 */
1819 if ((w_rmatrix[pi][ci] & WITNESS_ANCESTOR) == 0) {
1820 parent->w_num_descendants++;
1821 child->w_num_ancestors++;
1822 }
1823
1824 /*
1825 * Find each ancestor of 'pi'. Note that 'pi' itself is counted as
1826 * an ancestor of 'pi' during this loop.
1827 */
1828 for (i = 1; i <= w_max_used_index; i++) {
1829 if ((w_rmatrix[i][pi] & WITNESS_ANCESTOR_MASK) == 0 &&
1830 (i != pi))
1831 continue;
1832
1833 /* Find each descendant of 'i' and mark it as a descendant. */
1834 for (j = 1; j <= w_max_used_index; j++) {
1835
1836 /*
1837 * Skip children that are already marked as
1838 * descendants of 'i'.
1839 */
1840 if (w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK)
1841 continue;
1842
1843 /*
1844 * We are only interested in descendants of 'ci'. Note
1845 * that 'ci' itself is counted as a descendant of 'ci'.
1846 */
1847 if ((w_rmatrix[ci][j] & WITNESS_ANCESTOR_MASK) == 0 &&
1848 (j != ci))
1849 continue;
1850 w_rmatrix[i][j] |= WITNESS_ANCESTOR;
1851 w_rmatrix[j][i] |= WITNESS_DESCENDANT;
1852 w_data[i].w_num_descendants++;
1853 w_data[j].w_num_ancestors++;
1854
1855 /*
1856 * Make sure we aren't marking a node as both an
1857 * ancestor and descendant. We should have caught
1858 * this as a lock order reversal earlier.
1859 */
1860 if ((w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK) &&
1861 (w_rmatrix[i][j] & WITNESS_DESCENDANT_MASK)) {
1862 printf("witness rmatrix paradox! [%d][%d]=%d "
1863 "both ancestor and descendant\n",
1864 i, j, w_rmatrix[i][j]);
1865 kdb_backtrace();
1866 printf("Witness disabled.\n");
1867 witness_watch = -1;
1868 }
1869 if ((w_rmatrix[j][i] & WITNESS_ANCESTOR_MASK) &&
1870 (w_rmatrix[j][i] & WITNESS_DESCENDANT_MASK)) {
1871 printf("witness rmatrix paradox! [%d][%d]=%d "
1872 "both ancestor and descendant\n",
1873 j, i, w_rmatrix[j][i]);
1874 kdb_backtrace();
1875 printf("Witness disabled.\n");
1876 witness_watch = -1;
1877 }
1878 }
1879 }
1880 }
1881
1882 static void
1883 itismychild(struct witness *parent, struct witness *child)
1884 {
1885
1886 MPASS(child != NULL && parent != NULL);
1887 if (witness_cold == 0)
1888 mtx_assert(&w_mtx, MA_OWNED);
1889
1890 if (!witness_lock_type_equal(parent, child)) {
1891 if (witness_cold == 0)
1892 mtx_unlock_spin(&w_mtx);
1893 panic("%s: parent \"%s\" (%s) and child \"%s\" (%s) are not "
1894 "the same lock type", __func__, parent->w_name,
1895 parent->w_class->lc_name, child->w_name,
1896 child->w_class->lc_name);
1897 }
1898 adopt(parent, child);
1899 }
1900
1901 /*
1902 * Generic code for the isitmy*() functions. The rmask parameter is the
1903 * expected relationship of w1 to w2.
1904 */
1905 static int
1906 _isitmyx(struct witness *w1, struct witness *w2, int rmask, const char *fname)
1907 {
1908 unsigned char r1, r2;
1909 int i1, i2;
1910
1911 i1 = w1->w_index;
1912 i2 = w2->w_index;
1913 WITNESS_INDEX_ASSERT(i1);
1914 WITNESS_INDEX_ASSERT(i2);
1915 r1 = w_rmatrix[i1][i2] & WITNESS_RELATED_MASK;
1916 r2 = w_rmatrix[i2][i1] & WITNESS_RELATED_MASK;
1917
1918 /* The flags on one better be the inverse of the flags on the other */
1919 if (!((WITNESS_ATOD(r1) == r2 && WITNESS_DTOA(r2) == r1) ||
1920 (WITNESS_DTOA(r1) == r2 && WITNESS_ATOD(r2) == r1))) {
1921 printf("%s: rmatrix mismatch between %s (index %d) and %s "
1922 "(index %d): w_rmatrix[%d][%d] == %hhx but "
1923 "w_rmatrix[%d][%d] == %hhx\n",
1924 fname, w1->w_name, i1, w2->w_name, i2, i1, i2, r1,
1925 i2, i1, r2);
1926 kdb_backtrace();
1927 printf("Witness disabled.\n");
1928 witness_watch = -1;
1929 }
1930 return (r1 & rmask);
1931 }
1932
1933 /*
1934 * Checks if @child is a direct child of @parent.
1935 */
1936 static int
1937 isitmychild(struct witness *parent, struct witness *child)
1938 {
1939
1940 return (_isitmyx(parent, child, WITNESS_PARENT, __func__));
1941 }
1942
1943 /*
1944 * Checks if @descendant is a direct or inderect descendant of @ancestor.
1945 */
1946 static int
1947 isitmydescendant(struct witness *ancestor, struct witness *descendant)
1948 {
1949
1950 return (_isitmyx(ancestor, descendant, WITNESS_ANCESTOR_MASK,
1951 __func__));
1952 }
1953
1954 #ifdef BLESSING
1955 static int
1956 blessed(struct witness *w1, struct witness *w2)
1957 {
1958 int i;
1959 struct witness_blessed *b;
1960
1961 for (i = 0; i < blessed_count; i++) {
1962 b = &blessed_list[i];
1963 if (strcmp(w1->w_name, b->b_lock1) == 0) {
1964 if (strcmp(w2->w_name, b->b_lock2) == 0)
1965 return (1);
1966 continue;
1967 }
1968 if (strcmp(w1->w_name, b->b_lock2) == 0)
1969 if (strcmp(w2->w_name, b->b_lock1) == 0)
1970 return (1);
1971 }
1972 return (0);
1973 }
1974 #endif
1975
1976 static struct witness *
1977 witness_get(void)
1978 {
1979 struct witness *w;
1980 int index;
1981
1982 if (witness_cold == 0)
1983 mtx_assert(&w_mtx, MA_OWNED);
1984
1985 if (witness_watch == -1) {
1986 mtx_unlock_spin(&w_mtx);
1987 return (NULL);
1988 }
1989 if (STAILQ_EMPTY(&w_free)) {
1990 witness_watch = -1;
1991 mtx_unlock_spin(&w_mtx);
1992 printf("WITNESS: unable to allocate a new witness object\n");
1993 return (NULL);
1994 }
1995 w = STAILQ_FIRST(&w_free);
1996 STAILQ_REMOVE_HEAD(&w_free, w_list);
1997 w_free_cnt--;
1998 index = w->w_index;
1999 MPASS(index > 0 && index == w_max_used_index+1 &&
2000 index < WITNESS_COUNT);
2001 bzero(w, sizeof(*w));
2002 w->w_index = index;
2003 if (index > w_max_used_index)
2004 w_max_used_index = index;
2005 return (w);
2006 }
2007
2008 static void
2009 witness_free(struct witness *w)
2010 {
2011
2012 STAILQ_INSERT_HEAD(&w_free, w, w_list);
2013 w_free_cnt++;
2014 }
2015
2016 static struct lock_list_entry *
2017 witness_lock_list_get(void)
2018 {
2019 struct lock_list_entry *lle;
2020
2021 if (witness_watch == -1)
2022 return (NULL);
2023 mtx_lock_spin(&w_mtx);
2024 lle = w_lock_list_free;
2025 if (lle == NULL) {
2026 witness_watch = -1;
2027 mtx_unlock_spin(&w_mtx);
2028 printf("%s: witness exhausted\n", __func__);
2029 return (NULL);
2030 }
2031 w_lock_list_free = lle->ll_next;
2032 mtx_unlock_spin(&w_mtx);
2033 bzero(lle, sizeof(*lle));
2034 return (lle);
2035 }
2036
2037 static void
2038 witness_lock_list_free(struct lock_list_entry *lle)
2039 {
2040
2041 mtx_lock_spin(&w_mtx);
2042 lle->ll_next = w_lock_list_free;
2043 w_lock_list_free = lle;
2044 mtx_unlock_spin(&w_mtx);
2045 }
2046
2047 static struct lock_instance *
2048 find_instance(struct lock_list_entry *list, struct lock_object *lock)
2049 {
2050 struct lock_list_entry *lle;
2051 struct lock_instance *instance;
2052 int i;
2053
2054 for (lle = list; lle != NULL; lle = lle->ll_next)
2055 for (i = lle->ll_count - 1; i >= 0; i--) {
2056 instance = &lle->ll_children[i];
2057 if (instance->li_lock == lock)
2058 return (instance);
2059 }
2060 return (NULL);
2061 }
2062
2063 static void
2064 witness_list_lock(struct lock_instance *instance,
2065 int (*prnt)(const char *fmt, ...))
2066 {
2067 struct lock_object *lock;
2068
2069 lock = instance->li_lock;
2070 prnt("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ?
2071 "exclusive" : "shared", LOCK_CLASS(lock)->lc_name, lock->lo_name);
2072 if (lock->lo_witness->w_name != lock->lo_name)
2073 prnt(" (%s)", lock->lo_witness->w_name);
2074 prnt(" r = %d (%p) locked @ %s:%d\n",
2075 instance->li_flags & LI_RECURSEMASK, lock, instance->li_file,
2076 instance->li_line);
2077 }
2078
2079 #ifdef DDB
2080 static int
2081 witness_thread_has_locks(struct thread *td)
2082 {
2083
2084 if (td->td_sleeplocks == NULL)
2085 return (0);
2086 return (td->td_sleeplocks->ll_count != 0);
2087 }
2088
2089 static int
2090 witness_proc_has_locks(struct proc *p)
2091 {
2092 struct thread *td;
2093
2094 FOREACH_THREAD_IN_PROC(p, td) {
2095 if (witness_thread_has_locks(td))
2096 return (1);
2097 }
2098 return (0);
2099 }
2100 #endif
2101
2102 int
2103 witness_list_locks(struct lock_list_entry **lock_list,
2104 int (*prnt)(const char *fmt, ...))
2105 {
2106 struct lock_list_entry *lle;
2107 int i, nheld;
2108
2109 nheld = 0;
2110 for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
2111 for (i = lle->ll_count - 1; i >= 0; i--) {
2112 witness_list_lock(&lle->ll_children[i], prnt);
2113 nheld++;
2114 }
2115 return (nheld);
2116 }
2117
2118 /*
2119 * This is a bit risky at best. We call this function when we have timed
2120 * out acquiring a spin lock, and we assume that the other CPU is stuck
2121 * with this lock held. So, we go groveling around in the other CPU's
2122 * per-cpu data to try to find the lock instance for this spin lock to
2123 * see when it was last acquired.
2124 */
2125 void
2126 witness_display_spinlock(struct lock_object *lock, struct thread *owner,
2127 int (*prnt)(const char *fmt, ...))
2128 {
2129 struct lock_instance *instance;
2130 struct pcpu *pc;
2131
2132 if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU)
2133 return;
2134 pc = pcpu_find(owner->td_oncpu);
2135 instance = find_instance(pc->pc_spinlocks, lock);
2136 if (instance != NULL)
2137 witness_list_lock(instance, prnt);
2138 }
2139
2140 void
2141 witness_save(struct lock_object *lock, const char **filep, int *linep)
2142 {
2143 struct lock_list_entry *lock_list;
2144 struct lock_instance *instance;
2145 struct lock_class *class;
2146
2147 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2148 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2149 return;
2150 class = LOCK_CLASS(lock);
2151 if (class->lc_flags & LC_SLEEPLOCK)
2152 lock_list = curthread->td_sleeplocks;
2153 else {
2154 if (witness_skipspin)
2155 return;
2156 lock_list = PCPU_GET(spinlocks);
2157 }
2158 instance = find_instance(lock_list, lock);
2159 if (instance == NULL)
2160 panic("%s: lock (%s) %s not locked", __func__,
2161 class->lc_name, lock->lo_name);
2162 *filep = instance->li_file;
2163 *linep = instance->li_line;
2164 }
2165
2166 void
2167 witness_restore(struct lock_object *lock, const char *file, int line)
2168 {
2169 struct lock_list_entry *lock_list;
2170 struct lock_instance *instance;
2171 struct lock_class *class;
2172
2173 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2174 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2175 return;
2176 class = LOCK_CLASS(lock);
2177 if (class->lc_flags & LC_SLEEPLOCK)
2178 lock_list = curthread->td_sleeplocks;
2179 else {
2180 if (witness_skipspin)
2181 return;
2182 lock_list = PCPU_GET(spinlocks);
2183 }
2184 instance = find_instance(lock_list, lock);
2185 if (instance == NULL)
2186 panic("%s: lock (%s) %s not locked", __func__,
2187 class->lc_name, lock->lo_name);
2188 lock->lo_witness->w_file = file;
2189 lock->lo_witness->w_line = line;
2190 instance->li_file = file;
2191 instance->li_line = line;
2192 }
2193
2194 void
2195 witness_assert(struct lock_object *lock, int flags, const char *file, int line)
2196 {
2197 #ifdef INVARIANT_SUPPORT
2198 struct lock_instance *instance;
2199 struct lock_class *class;
2200
2201 if (lock->lo_witness == NULL || witness_watch < 1 || panicstr != NULL)
2202 return;
2203 class = LOCK_CLASS(lock);
2204 if ((class->lc_flags & LC_SLEEPLOCK) != 0)
2205 instance = find_instance(curthread->td_sleeplocks, lock);
2206 else if ((class->lc_flags & LC_SPINLOCK) != 0)
2207 instance = find_instance(PCPU_GET(spinlocks), lock);
2208 else {
2209 panic("Lock (%s) %s is not sleep or spin!",
2210 class->lc_name, lock->lo_name);
2211 }
2212 file = fixup_filename(file);
2213 switch (flags) {
2214 case LA_UNLOCKED:
2215 if (instance != NULL)
2216 panic("Lock (%s) %s locked @ %s:%d.",
2217 class->lc_name, lock->lo_name, file, line);
2218 break;
2219 case LA_LOCKED:
2220 case LA_LOCKED | LA_RECURSED:
2221 case LA_LOCKED | LA_NOTRECURSED:
2222 case LA_SLOCKED:
2223 case LA_SLOCKED | LA_RECURSED:
2224 case LA_SLOCKED | LA_NOTRECURSED:
2225 case LA_XLOCKED:
2226 case LA_XLOCKED | LA_RECURSED:
2227 case LA_XLOCKED | LA_NOTRECURSED:
2228 if (instance == NULL) {
2229 panic("Lock (%s) %s not locked @ %s:%d.",
2230 class->lc_name, lock->lo_name, file, line);
2231 break;
2232 }
2233 if ((flags & LA_XLOCKED) != 0 &&
2234 (instance->li_flags & LI_EXCLUSIVE) == 0)
2235 panic("Lock (%s) %s not exclusively locked @ %s:%d.",
2236 class->lc_name, lock->lo_name, file, line);
2237 if ((flags & LA_SLOCKED) != 0 &&
2238 (instance->li_flags & LI_EXCLUSIVE) != 0)
2239 panic("Lock (%s) %s exclusively locked @ %s:%d.",
2240 class->lc_name, lock->lo_name, file, line);
2241 if ((flags & LA_RECURSED) != 0 &&
2242 (instance->li_flags & LI_RECURSEMASK) == 0)
2243 panic("Lock (%s) %s not recursed @ %s:%d.",
2244 class->lc_name, lock->lo_name, file, line);
2245 if ((flags & LA_NOTRECURSED) != 0 &&
2246 (instance->li_flags & LI_RECURSEMASK) != 0)
2247 panic("Lock (%s) %s recursed @ %s:%d.",
2248 class->lc_name, lock->lo_name, file, line);
2249 break;
2250 default:
2251 panic("Invalid lock assertion at %s:%d.", file, line);
2252
2253 }
2254 #endif /* INVARIANT_SUPPORT */
2255 }
2256
2257 static void
2258 witness_setflag(struct lock_object *lock, int flag, int set)
2259 {
2260 struct lock_list_entry *lock_list;
2261 struct lock_instance *instance;
2262 struct lock_class *class;
2263
2264 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2265 return;
2266 class = LOCK_CLASS(lock);
2267 if (class->lc_flags & LC_SLEEPLOCK)
2268 lock_list = curthread->td_sleeplocks;
2269 else {
2270 if (witness_skipspin)
2271 return;
2272 lock_list = PCPU_GET(spinlocks);
2273 }
2274 instance = find_instance(lock_list, lock);
2275 if (instance == NULL)
2276 panic("%s: lock (%s) %s not locked", __func__,
2277 class->lc_name, lock->lo_name);
2278
2279 if (set)
2280 instance->li_flags |= flag;
2281 else
2282 instance->li_flags &= ~flag;
2283 }
2284
2285 void
2286 witness_norelease(struct lock_object *lock)
2287 {
2288
2289 witness_setflag(lock, LI_NORELEASE, 1);
2290 }
2291
2292 void
2293 witness_releaseok(struct lock_object *lock)
2294 {
2295
2296 witness_setflag(lock, LI_NORELEASE, 0);
2297 }
2298
2299 #ifdef DDB
2300 static void
2301 witness_ddb_list(struct thread *td)
2302 {
2303
2304 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2305 KASSERT(kdb_active, ("%s: not in the debugger", __func__));
2306
2307 if (witness_watch < 1)
2308 return;
2309
2310 witness_list_locks(&td->td_sleeplocks, db_printf);
2311
2312 /*
2313 * We only handle spinlocks if td == curthread. This is somewhat broken
2314 * if td is currently executing on some other CPU and holds spin locks
2315 * as we won't display those locks. If we had a MI way of getting
2316 * the per-cpu data for a given cpu then we could use
2317 * td->td_oncpu to get the list of spinlocks for this thread
2318 * and "fix" this.
2319 *
2320 * That still wouldn't really fix this unless we locked the scheduler
2321 * lock or stopped the other CPU to make sure it wasn't changing the
2322 * list out from under us. It is probably best to just not try to
2323 * handle threads on other CPU's for now.
2324 */
2325 if (td == curthread && PCPU_GET(spinlocks) != NULL)
2326 witness_list_locks(PCPU_PTR(spinlocks), db_printf);
2327 }
2328
2329 DB_SHOW_COMMAND(locks, db_witness_list)
2330 {
2331 struct thread *td;
2332
2333 if (have_addr)
2334 td = db_lookup_thread(addr, TRUE);
2335 else
2336 td = kdb_thread;
2337 witness_ddb_list(td);
2338 }
2339
2340 DB_SHOW_ALL_COMMAND(locks, db_witness_list_all)
2341 {
2342 struct thread *td;
2343 struct proc *p;
2344
2345 /*
2346 * It would be nice to list only threads and processes that actually
2347 * held sleep locks, but that information is currently not exported
2348 * by WITNESS.
2349 */
2350 FOREACH_PROC_IN_SYSTEM(p) {
2351 if (!witness_proc_has_locks(p))
2352 continue;
2353 FOREACH_THREAD_IN_PROC(p, td) {
2354 if (!witness_thread_has_locks(td))
2355 continue;
2356 db_printf("Process %d (%s) thread %p (%d)\n", p->p_pid,
2357 p->p_comm, td, td->td_tid);
2358 witness_ddb_list(td);
2359 }
2360 }
2361 }
2362 DB_SHOW_ALIAS(alllocks, db_witness_list_all)
2363
2364 DB_SHOW_COMMAND(witness, db_witness_display)
2365 {
2366
2367 witness_ddb_display(db_printf);
2368 }
2369 #endif
2370
2371 static int
2372 sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS)
2373 {
2374 struct witness_lock_order_data *data1, *data2, *tmp_data1, *tmp_data2;
2375 struct witness *tmp_w1, *tmp_w2, *w1, *w2;
2376 struct sbuf *sb;
2377 u_int w_rmatrix1, w_rmatrix2;
2378 int error, generation, i, j;
2379
2380 tmp_data1 = NULL;
2381 tmp_data2 = NULL;
2382 tmp_w1 = NULL;
2383 tmp_w2 = NULL;
2384 if (witness_watch < 1) {
2385 error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2386 return (error);
2387 }
2388 if (witness_cold) {
2389 error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2390 return (error);
2391 }
2392 error = 0;
2393 sb = sbuf_new(NULL, NULL, BADSTACK_SBUF_SIZE, SBUF_AUTOEXTEND);
2394 if (sb == NULL)
2395 return (ENOMEM);
2396
2397 /* Allocate and init temporary storage space. */
2398 tmp_w1 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2399 tmp_w2 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2400 tmp_data1 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2401 M_WAITOK | M_ZERO);
2402 tmp_data2 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2403 M_WAITOK | M_ZERO);
2404 stack_zero(&tmp_data1->wlod_stack);
2405 stack_zero(&tmp_data2->wlod_stack);
2406
2407 restart:
2408 mtx_lock_spin(&w_mtx);
2409 generation = w_generation;
2410 mtx_unlock_spin(&w_mtx);
2411 sbuf_printf(sb, "Number of known direct relationships is %d\n",
2412 w_lohash.wloh_count);
2413 for (i = 1; i < w_max_used_index; i++) {
2414 mtx_lock_spin(&w_mtx);
2415 if (generation != w_generation) {
2416 mtx_unlock_spin(&w_mtx);
2417
2418 /* The graph has changed, try again. */
2419 req->oldidx = 0;
2420 sbuf_clear(sb);
2421 goto restart;
2422 }
2423
2424 w1 = &w_data[i];
2425 if (w1->w_reversed == 0) {
2426 mtx_unlock_spin(&w_mtx);
2427 continue;
2428 }
2429
2430 /* Copy w1 locally so we can release the spin lock. */
2431 *tmp_w1 = *w1;
2432 mtx_unlock_spin(&w_mtx);
2433
2434 if (tmp_w1->w_reversed == 0)
2435 continue;
2436 for (j = 1; j < w_max_used_index; j++) {
2437 if ((w_rmatrix[i][j] & WITNESS_REVERSAL) == 0 || i > j)
2438 continue;
2439
2440 mtx_lock_spin(&w_mtx);
2441 if (generation != w_generation) {
2442 mtx_unlock_spin(&w_mtx);
2443
2444 /* The graph has changed, try again. */
2445 req->oldidx = 0;
2446 sbuf_clear(sb);
2447 goto restart;
2448 }
2449
2450 w2 = &w_data[j];
2451 data1 = witness_lock_order_get(w1, w2);
2452 data2 = witness_lock_order_get(w2, w1);
2453
2454 /*
2455 * Copy information locally so we can release the
2456 * spin lock.
2457 */
2458 *tmp_w2 = *w2;
2459 w_rmatrix1 = (unsigned int)w_rmatrix[i][j];
2460 w_rmatrix2 = (unsigned int)w_rmatrix[j][i];
2461
2462 if (data1) {
2463 stack_zero(&tmp_data1->wlod_stack);
2464 stack_copy(&data1->wlod_stack,
2465 &tmp_data1->wlod_stack);
2466 }
2467 if (data2 && data2 != data1) {
2468 stack_zero(&tmp_data2->wlod_stack);
2469 stack_copy(&data2->wlod_stack,
2470 &tmp_data2->wlod_stack);
2471 }
2472 mtx_unlock_spin(&w_mtx);
2473
2474 sbuf_printf(sb,
2475 "\nLock order reversal between \"%s\"(%s) and \"%s\"(%s)!\n",
2476 tmp_w1->w_name, tmp_w1->w_class->lc_name,
2477 tmp_w2->w_name, tmp_w2->w_class->lc_name);
2478 #if 0
2479 sbuf_printf(sb,
2480 "w_rmatrix[%s][%s] == %x, w_rmatrix[%s][%s] == %x\n",
2481 tmp_w1->name, tmp_w2->w_name, w_rmatrix1,
2482 tmp_w2->name, tmp_w1->w_name, w_rmatrix2);
2483 #endif
2484 if (data1) {
2485 sbuf_printf(sb,
2486 "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2487 tmp_w1->w_name, tmp_w1->w_class->lc_name,
2488 tmp_w2->w_name, tmp_w2->w_class->lc_name);
2489 stack_sbuf_print(sb, &tmp_data1->wlod_stack);
2490 sbuf_printf(sb, "\n");
2491 }
2492 if (data2 && data2 != data1) {
2493 sbuf_printf(sb,
2494 "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2495 tmp_w2->w_name, tmp_w2->w_class->lc_name,
2496 tmp_w1->w_name, tmp_w1->w_class->lc_name);
2497 stack_sbuf_print(sb, &tmp_data2->wlod_stack);
2498 sbuf_printf(sb, "\n");
2499 }
2500 }
2501 }
2502 mtx_lock_spin(&w_mtx);
2503 if (generation != w_generation) {
2504 mtx_unlock_spin(&w_mtx);
2505
2506 /*
2507 * The graph changed while we were printing stack data,
2508 * try again.
2509 */
2510 req->oldidx = 0;
2511 sbuf_clear(sb);
2512 goto restart;
2513 }
2514 mtx_unlock_spin(&w_mtx);
2515
2516 /* Free temporary storage space. */
2517 free(tmp_data1, M_TEMP);
2518 free(tmp_data2, M_TEMP);
2519 free(tmp_w1, M_TEMP);
2520 free(tmp_w2, M_TEMP);
2521
2522 sbuf_finish(sb);
2523 error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1);
2524 sbuf_delete(sb);
2525
2526 return (error);
2527 }
2528
2529 static int
2530 sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS)
2531 {
2532 struct witness *w;
2533 struct sbuf *sb;
2534 int error;
2535
2536 if (witness_watch < 1) {
2537 error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2538 return (error);
2539 }
2540 if (witness_cold) {
2541 error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2542 return (error);
2543 }
2544 error = 0;
2545 sb = sbuf_new(NULL, NULL, FULLGRAPH_SBUF_SIZE, SBUF_FIXEDLEN);
2546 if (sb == NULL)
2547 return (ENOMEM);
2548 sbuf_printf(sb, "\n");
2549
2550 mtx_lock_spin(&w_mtx);
2551 STAILQ_FOREACH(w, &w_all, w_list)
2552 w->w_displayed = 0;
2553 STAILQ_FOREACH(w, &w_all, w_list)
2554 witness_add_fullgraph(sb, w);
2555 mtx_unlock_spin(&w_mtx);
2556
2557 /*
2558 * While using SBUF_FIXEDLEN, check if the sbuf overflowed.
2559 */
2560 if (sbuf_overflowed(sb)) {
2561 sbuf_delete(sb);
2562 panic("%s: sbuf overflowed, bump FULLGRAPH_SBUF_SIZE value\n",
2563 __func__);
2564 }
2565
2566 /*
2567 * Close the sbuf and return to userland.
2568 */
2569 sbuf_finish(sb);
2570 error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1);
2571 sbuf_delete(sb);
2572
2573 return (error);
2574 }
2575
2576 static int
2577 sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS)
2578 {
2579 int error, value;
2580
2581 value = witness_watch;
2582 error = sysctl_handle_int(oidp, &value, 0, req);
2583 if (error != 0 || req->newptr == NULL)
2584 return (error);
2585 if (value > 1 || value < -1 ||
2586 (witness_watch == -1 && value != witness_watch))
2587 return (EINVAL);
2588 witness_watch = value;
2589 return (0);
2590 }
2591
2592 static void
2593 witness_add_fullgraph(struct sbuf *sb, struct witness *w)
2594 {
2595 int i;
2596
2597 if (w->w_displayed != 0 || (w->w_file == NULL && w->w_line == 0))
2598 return;
2599 w->w_displayed = 1;
2600
2601 WITNESS_INDEX_ASSERT(w->w_index);
2602 for (i = 1; i <= w_max_used_index; i++) {
2603 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT) {
2604 sbuf_printf(sb, "\"%s\",\"%s\"\n", w->w_name,
2605 w_data[i].w_name);
2606 witness_add_fullgraph(sb, &w_data[i]);
2607 }
2608 }
2609 }
2610
2611 /*
2612 * A simple hash function. Takes a key pointer and a key size. If size == 0,
2613 * interprets the key as a string and reads until the null
2614 * terminator. Otherwise, reads the first size bytes. Returns an unsigned 32-bit
2615 * hash value computed from the key.
2616 */
2617 static uint32_t
2618 witness_hash_djb2(const uint8_t *key, uint32_t size)
2619 {
2620 unsigned int hash = 5381;
2621 int i;
2622
2623 /* hash = hash * 33 + key[i] */
2624 if (size)
2625 for (i = 0; i < size; i++)
2626 hash = ((hash << 5) + hash) + (unsigned int)key[i];
2627 else
2628 for (i = 0; key[i] != 0; i++)
2629 hash = ((hash << 5) + hash) + (unsigned int)key[i];
2630
2631 return (hash);
2632 }
2633
2634
2635 /*
2636 * Initializes the two witness hash tables. Called exactly once from
2637 * witness_initialize().
2638 */
2639 static void
2640 witness_init_hash_tables(void)
2641 {
2642 int i;
2643
2644 MPASS(witness_cold);
2645
2646 /* Initialize the hash tables. */
2647 for (i = 0; i < WITNESS_HASH_SIZE; i++)
2648 w_hash.wh_array[i] = NULL;
2649
2650 w_hash.wh_size = WITNESS_HASH_SIZE;
2651 w_hash.wh_count = 0;
2652
2653 /* Initialize the lock order data hash. */
2654 w_lofree = NULL;
2655 for (i = 0; i < WITNESS_LO_DATA_COUNT; i++) {
2656 memset(&w_lodata[i], 0, sizeof(w_lodata[i]));
2657 w_lodata[i].wlod_next = w_lofree;
2658 w_lofree = &w_lodata[i];
2659 }
2660 w_lohash.wloh_size = WITNESS_LO_HASH_SIZE;
2661 w_lohash.wloh_count = 0;
2662 for (i = 0; i < WITNESS_LO_HASH_SIZE; i++)
2663 w_lohash.wloh_array[i] = NULL;
2664 }
2665
2666 static struct witness *
2667 witness_hash_get(const char *key)
2668 {
2669 struct witness *w;
2670 uint32_t hash;
2671
2672 MPASS(key != NULL);
2673 if (witness_cold == 0)
2674 mtx_assert(&w_mtx, MA_OWNED);
2675 hash = witness_hash_djb2(key, 0) % w_hash.wh_size;
2676 w = w_hash.wh_array[hash];
2677 while (w != NULL) {
2678 if (strcmp(w->w_name, key) == 0)
2679 goto out;
2680 w = w->w_hash_next;
2681 }
2682
2683 out:
2684 return (w);
2685 }
2686
2687 static void
2688 witness_hash_put(struct witness *w)
2689 {
2690 uint32_t hash;
2691
2692 MPASS(w != NULL);
2693 MPASS(w->w_name != NULL);
2694 if (witness_cold == 0)
2695 mtx_assert(&w_mtx, MA_OWNED);
2696 KASSERT(witness_hash_get(w->w_name) == NULL,
2697 ("%s: trying to add a hash entry that already exists!", __func__));
2698 KASSERT(w->w_hash_next == NULL,
2699 ("%s: w->w_hash_next != NULL", __func__));
2700
2701 hash = witness_hash_djb2(w->w_name, 0) % w_hash.wh_size;
2702 w->w_hash_next = w_hash.wh_array[hash];
2703 w_hash.wh_array[hash] = w;
2704 w_hash.wh_count++;
2705 }
2706
2707
2708 static struct witness_lock_order_data *
2709 witness_lock_order_get(struct witness *parent, struct witness *child)
2710 {
2711 struct witness_lock_order_data *data = NULL;
2712 struct witness_lock_order_key key;
2713 unsigned int hash;
2714
2715 MPASS(parent != NULL && child != NULL);
2716 key.from = parent->w_index;
2717 key.to = child->w_index;
2718 WITNESS_INDEX_ASSERT(key.from);
2719 WITNESS_INDEX_ASSERT(key.to);
2720 if ((w_rmatrix[parent->w_index][child->w_index]
2721 & WITNESS_LOCK_ORDER_KNOWN) == 0)
2722 goto out;
2723
2724 hash = witness_hash_djb2((const char*)&key,
2725 sizeof(key)) % w_lohash.wloh_size;
2726 data = w_lohash.wloh_array[hash];
2727 while (data != NULL) {
2728 if (witness_lock_order_key_equal(&data->wlod_key, &key))
2729 break;
2730 data = data->wlod_next;
2731 }
2732
2733 out:
2734 return (data);
2735 }
2736
2737 /*
2738 * Verify that parent and child have a known relationship, are not the same,
2739 * and child is actually a child of parent. This is done without w_mtx
2740 * to avoid contention in the common case.
2741 */
2742 static int
2743 witness_lock_order_check(struct witness *parent, struct witness *child)
2744 {
2745
2746 if (parent != child &&
2747 w_rmatrix[parent->w_index][child->w_index]
2748 & WITNESS_LOCK_ORDER_KNOWN &&
2749 isitmychild(parent, child))
2750 return (1);
2751
2752 return (0);
2753 }
2754
2755 static int
2756 witness_lock_order_add(struct witness *parent, struct witness *child)
2757 {
2758 struct witness_lock_order_data *data = NULL;
2759 struct witness_lock_order_key key;
2760 unsigned int hash;
2761
2762 MPASS(parent != NULL && child != NULL);
2763 key.from = parent->w_index;
2764 key.to = child->w_index;
2765 WITNESS_INDEX_ASSERT(key.from);
2766 WITNESS_INDEX_ASSERT(key.to);
2767 if (w_rmatrix[parent->w_index][child->w_index]
2768 & WITNESS_LOCK_ORDER_KNOWN)
2769 return (1);
2770
2771 hash = witness_hash_djb2((const char*)&key,
2772 sizeof(key)) % w_lohash.wloh_size;
2773 w_rmatrix[parent->w_index][child->w_index] |= WITNESS_LOCK_ORDER_KNOWN;
2774 data = w_lofree;
2775 if (data == NULL)
2776 return (0);
2777 w_lofree = data->wlod_next;
2778 data->wlod_next = w_lohash.wloh_array[hash];
2779 data->wlod_key = key;
2780 w_lohash.wloh_array[hash] = data;
2781 w_lohash.wloh_count++;
2782 stack_zero(&data->wlod_stack);
2783 stack_save(&data->wlod_stack);
2784 return (1);
2785 }
2786
2787 /* Call this whenver the structure of the witness graph changes. */
2788 static void
2789 witness_increment_graph_generation(void)
2790 {
2791
2792 if (witness_cold == 0)
2793 mtx_assert(&w_mtx, MA_OWNED);
2794 w_generation++;
2795 }
2796
2797 #ifdef KDB
2798 static void
2799 _witness_debugger(int cond, const char *msg)
2800 {
2801
2802 if (witness_trace && cond)
2803 kdb_backtrace();
2804 if (witness_kdb && cond)
2805 kdb_enter(KDB_WHY_WITNESS, msg);
2806 }
2807 #endif
Cache object: 7e9139db773dbc5fef0fda43af3f841e
|