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/10.1/sys/kern/subr_witness.c 270185 2014-08-19 23:08:47Z grehan $");
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 (1024 + MAXCPU)
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 FULLGRAPH_SBUF_SIZE 512
158
159 /*
160 * These flags go in the witness relationship matrix and describe the
161 * relationship between any two struct witness objects.
162 */
163 #define WITNESS_UNRELATED 0x00 /* No lock order relation. */
164 #define WITNESS_PARENT 0x01 /* Parent, aka direct ancestor. */
165 #define WITNESS_ANCESTOR 0x02 /* Direct or indirect ancestor. */
166 #define WITNESS_CHILD 0x04 /* Child, aka direct descendant. */
167 #define WITNESS_DESCENDANT 0x08 /* Direct or indirect descendant. */
168 #define WITNESS_ANCESTOR_MASK (WITNESS_PARENT | WITNESS_ANCESTOR)
169 #define WITNESS_DESCENDANT_MASK (WITNESS_CHILD | WITNESS_DESCENDANT)
170 #define WITNESS_RELATED_MASK \
171 (WITNESS_ANCESTOR_MASK | WITNESS_DESCENDANT_MASK)
172 #define WITNESS_REVERSAL 0x10 /* A lock order reversal has been
173 * observed. */
174 #define WITNESS_RESERVED1 0x20 /* Unused flag, reserved. */
175 #define WITNESS_RESERVED2 0x40 /* Unused flag, reserved. */
176 #define WITNESS_LOCK_ORDER_KNOWN 0x80 /* This lock order is known. */
177
178 /* Descendant to ancestor flags */
179 #define WITNESS_DTOA(x) (((x) & WITNESS_RELATED_MASK) >> 2)
180
181 /* Ancestor to descendant flags */
182 #define WITNESS_ATOD(x) (((x) & WITNESS_RELATED_MASK) << 2)
183
184 #define WITNESS_INDEX_ASSERT(i) \
185 MPASS((i) > 0 && (i) <= w_max_used_index && (i) < WITNESS_COUNT)
186
187 static MALLOC_DEFINE(M_WITNESS, "Witness", "Witness");
188
189 /*
190 * Lock instances. A lock instance is the data associated with a lock while
191 * it is held by witness. For example, a lock instance will hold the
192 * recursion count of a lock. Lock instances are held in lists. Spin locks
193 * are held in a per-cpu list while sleep locks are held in per-thread list.
194 */
195 struct lock_instance {
196 struct lock_object *li_lock;
197 const char *li_file;
198 int li_line;
199 u_int li_flags;
200 };
201
202 /*
203 * A simple list type used to build the list of locks held by a thread
204 * or CPU. We can't simply embed the list in struct lock_object since a
205 * lock may be held by more than one thread if it is a shared lock. Locks
206 * are added to the head of the list, so we fill up each list entry from
207 * "the back" logically. To ease some of the arithmetic, we actually fill
208 * in each list entry the normal way (children[0] then children[1], etc.) but
209 * when we traverse the list we read children[count-1] as the first entry
210 * down to children[0] as the final entry.
211 */
212 struct lock_list_entry {
213 struct lock_list_entry *ll_next;
214 struct lock_instance ll_children[LOCK_NCHILDREN];
215 u_int ll_count;
216 };
217
218 /*
219 * The main witness structure. One of these per named lock type in the system
220 * (for example, "vnode interlock").
221 */
222 struct witness {
223 char w_name[MAX_W_NAME];
224 uint32_t w_index; /* Index in the relationship matrix */
225 struct lock_class *w_class;
226 STAILQ_ENTRY(witness) w_list; /* List of all witnesses. */
227 STAILQ_ENTRY(witness) w_typelist; /* Witnesses of a type. */
228 struct witness *w_hash_next; /* Linked list in hash buckets. */
229 const char *w_file; /* File where last acquired */
230 uint32_t w_line; /* Line where last acquired */
231 uint32_t w_refcount;
232 uint16_t w_num_ancestors; /* direct/indirect
233 * ancestor count */
234 uint16_t w_num_descendants; /* direct/indirect
235 * descendant count */
236 int16_t w_ddb_level;
237 unsigned w_displayed:1;
238 unsigned w_reversed:1;
239 };
240
241 STAILQ_HEAD(witness_list, witness);
242
243 /*
244 * The witness hash table. Keys are witness names (const char *), elements are
245 * witness objects (struct witness *).
246 */
247 struct witness_hash {
248 struct witness *wh_array[WITNESS_HASH_SIZE];
249 uint32_t wh_size;
250 uint32_t wh_count;
251 };
252
253 /*
254 * Key type for the lock order data hash table.
255 */
256 struct witness_lock_order_key {
257 uint16_t from;
258 uint16_t to;
259 };
260
261 struct witness_lock_order_data {
262 struct stack wlod_stack;
263 struct witness_lock_order_key wlod_key;
264 struct witness_lock_order_data *wlod_next;
265 };
266
267 /*
268 * The witness lock order data hash table. Keys are witness index tuples
269 * (struct witness_lock_order_key), elements are lock order data objects
270 * (struct witness_lock_order_data).
271 */
272 struct witness_lock_order_hash {
273 struct witness_lock_order_data *wloh_array[WITNESS_LO_HASH_SIZE];
274 u_int wloh_size;
275 u_int wloh_count;
276 };
277
278 #ifdef BLESSING
279 struct witness_blessed {
280 const char *b_lock1;
281 const char *b_lock2;
282 };
283 #endif
284
285 struct witness_pendhelp {
286 const char *wh_type;
287 struct lock_object *wh_lock;
288 };
289
290 struct witness_order_list_entry {
291 const char *w_name;
292 struct lock_class *w_class;
293 };
294
295 /*
296 * Returns 0 if one of the locks is a spin lock and the other is not.
297 * Returns 1 otherwise.
298 */
299 static __inline int
300 witness_lock_type_equal(struct witness *w1, struct witness *w2)
301 {
302
303 return ((w1->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) ==
304 (w2->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)));
305 }
306
307 static __inline int
308 witness_lock_order_key_equal(const struct witness_lock_order_key *a,
309 const struct witness_lock_order_key *b)
310 {
311
312 return (a->from == b->from && a->to == b->to);
313 }
314
315 static int _isitmyx(struct witness *w1, struct witness *w2, int rmask,
316 const char *fname);
317 #ifdef KDB
318 static void _witness_debugger(int cond, const char *msg);
319 #endif
320 static void adopt(struct witness *parent, struct witness *child);
321 #ifdef BLESSING
322 static int blessed(struct witness *, struct witness *);
323 #endif
324 static void depart(struct witness *w);
325 static struct witness *enroll(const char *description,
326 struct lock_class *lock_class);
327 static struct lock_instance *find_instance(struct lock_list_entry *list,
328 const struct lock_object *lock);
329 static int isitmychild(struct witness *parent, struct witness *child);
330 static int isitmydescendant(struct witness *parent, struct witness *child);
331 static void itismychild(struct witness *parent, struct witness *child);
332 static int sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS);
333 static int sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS);
334 static int sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS);
335 static void witness_add_fullgraph(struct sbuf *sb, struct witness *parent);
336 #ifdef DDB
337 static void witness_ddb_compute_levels(void);
338 static void witness_ddb_display(int(*)(const char *fmt, ...));
339 static void witness_ddb_display_descendants(int(*)(const char *fmt, ...),
340 struct witness *, int indent);
341 static void witness_ddb_display_list(int(*prnt)(const char *fmt, ...),
342 struct witness_list *list);
343 static void witness_ddb_level_descendants(struct witness *parent, int l);
344 static void witness_ddb_list(struct thread *td);
345 #endif
346 static void witness_free(struct witness *m);
347 static struct witness *witness_get(void);
348 static uint32_t witness_hash_djb2(const uint8_t *key, uint32_t size);
349 static struct witness *witness_hash_get(const char *key);
350 static void witness_hash_put(struct witness *w);
351 static void witness_init_hash_tables(void);
352 static void witness_increment_graph_generation(void);
353 static void witness_lock_list_free(struct lock_list_entry *lle);
354 static struct lock_list_entry *witness_lock_list_get(void);
355 static int witness_lock_order_add(struct witness *parent,
356 struct witness *child);
357 static int witness_lock_order_check(struct witness *parent,
358 struct witness *child);
359 static struct witness_lock_order_data *witness_lock_order_get(
360 struct witness *parent,
361 struct witness *child);
362 static void witness_list_lock(struct lock_instance *instance,
363 int (*prnt)(const char *fmt, ...));
364 static void witness_setflag(struct lock_object *lock, int flag, int set);
365
366 #ifdef KDB
367 #define witness_debugger(c) _witness_debugger(c, __func__)
368 #else
369 #define witness_debugger(c)
370 #endif
371
372 static SYSCTL_NODE(_debug, OID_AUTO, witness, CTLFLAG_RW, NULL,
373 "Witness Locking");
374
375 /*
376 * If set to 0, lock order checking is disabled. If set to -1,
377 * witness is completely disabled. Otherwise witness performs full
378 * lock order checking for all locks. At runtime, lock order checking
379 * may be toggled. However, witness cannot be reenabled once it is
380 * completely disabled.
381 */
382 static int witness_watch = 1;
383 TUNABLE_INT("debug.witness.watch", &witness_watch);
384 SYSCTL_PROC(_debug_witness, OID_AUTO, watch, CTLFLAG_RW | CTLTYPE_INT, NULL, 0,
385 sysctl_debug_witness_watch, "I", "witness is watching lock operations");
386
387 #ifdef KDB
388 /*
389 * When KDB is enabled and witness_kdb is 1, it will cause the system
390 * to drop into kdebug() when:
391 * - a lock hierarchy violation occurs
392 * - locks are held when going to sleep.
393 */
394 #ifdef WITNESS_KDB
395 int witness_kdb = 1;
396 #else
397 int witness_kdb = 0;
398 #endif
399 TUNABLE_INT("debug.witness.kdb", &witness_kdb);
400 SYSCTL_INT(_debug_witness, OID_AUTO, kdb, CTLFLAG_RW, &witness_kdb, 0, "");
401
402 /*
403 * When KDB is enabled and witness_trace is 1, it will cause the system
404 * to print a stack trace:
405 * - a lock hierarchy violation occurs
406 * - locks are held when going to sleep.
407 */
408 int witness_trace = 1;
409 TUNABLE_INT("debug.witness.trace", &witness_trace);
410 SYSCTL_INT(_debug_witness, OID_AUTO, trace, CTLFLAG_RW, &witness_trace, 0, "");
411 #endif /* KDB */
412
413 #ifdef WITNESS_SKIPSPIN
414 int witness_skipspin = 1;
415 #else
416 int witness_skipspin = 0;
417 #endif
418 TUNABLE_INT("debug.witness.skipspin", &witness_skipspin);
419 SYSCTL_INT(_debug_witness, OID_AUTO, skipspin, CTLFLAG_RDTUN, &witness_skipspin,
420 0, "");
421
422 /*
423 * Call this to print out the relations between locks.
424 */
425 SYSCTL_PROC(_debug_witness, OID_AUTO, fullgraph, CTLTYPE_STRING | CTLFLAG_RD,
426 NULL, 0, sysctl_debug_witness_fullgraph, "A", "Show locks relation graphs");
427
428 /*
429 * Call this to print out the witness faulty stacks.
430 */
431 SYSCTL_PROC(_debug_witness, OID_AUTO, badstacks, CTLTYPE_STRING | CTLFLAG_RD,
432 NULL, 0, sysctl_debug_witness_badstacks, "A", "Show bad witness stacks");
433
434 static struct mtx w_mtx;
435
436 /* w_list */
437 static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free);
438 static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all);
439
440 /* w_typelist */
441 static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin);
442 static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep);
443
444 /* lock list */
445 static struct lock_list_entry *w_lock_list_free = NULL;
446 static struct witness_pendhelp pending_locks[WITNESS_PENDLIST];
447 static u_int pending_cnt;
448
449 static int w_free_cnt, w_spin_cnt, w_sleep_cnt;
450 SYSCTL_INT(_debug_witness, OID_AUTO, free_cnt, CTLFLAG_RD, &w_free_cnt, 0, "");
451 SYSCTL_INT(_debug_witness, OID_AUTO, spin_cnt, CTLFLAG_RD, &w_spin_cnt, 0, "");
452 SYSCTL_INT(_debug_witness, OID_AUTO, sleep_cnt, CTLFLAG_RD, &w_sleep_cnt, 0,
453 "");
454
455 static struct witness *w_data;
456 static uint8_t w_rmatrix[WITNESS_COUNT+1][WITNESS_COUNT+1];
457 static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT];
458 static struct witness_hash w_hash; /* The witness hash table. */
459
460 /* The lock order data hash */
461 static struct witness_lock_order_data w_lodata[WITNESS_LO_DATA_COUNT];
462 static struct witness_lock_order_data *w_lofree = NULL;
463 static struct witness_lock_order_hash w_lohash;
464 static int w_max_used_index = 0;
465 static unsigned int w_generation = 0;
466 static const char w_notrunning[] = "Witness not running\n";
467 static const char w_stillcold[] = "Witness is still cold\n";
468
469
470 static struct witness_order_list_entry order_lists[] = {
471 /*
472 * sx locks
473 */
474 { "proctree", &lock_class_sx },
475 { "allproc", &lock_class_sx },
476 { "allprison", &lock_class_sx },
477 { NULL, NULL },
478 /*
479 * Various mutexes
480 */
481 { "Giant", &lock_class_mtx_sleep },
482 { "pipe mutex", &lock_class_mtx_sleep },
483 { "sigio lock", &lock_class_mtx_sleep },
484 { "process group", &lock_class_mtx_sleep },
485 { "process lock", &lock_class_mtx_sleep },
486 { "session", &lock_class_mtx_sleep },
487 { "uidinfo hash", &lock_class_rw },
488 #ifdef HWPMC_HOOKS
489 { "pmc-sleep", &lock_class_mtx_sleep },
490 #endif
491 { "time lock", &lock_class_mtx_sleep },
492 { NULL, NULL },
493 /*
494 * Sockets
495 */
496 { "accept", &lock_class_mtx_sleep },
497 { "so_snd", &lock_class_mtx_sleep },
498 { "so_rcv", &lock_class_mtx_sleep },
499 { "sellck", &lock_class_mtx_sleep },
500 { NULL, NULL },
501 /*
502 * Routing
503 */
504 { "so_rcv", &lock_class_mtx_sleep },
505 { "radix node head", &lock_class_rw },
506 { "rtentry", &lock_class_mtx_sleep },
507 { "ifaddr", &lock_class_mtx_sleep },
508 { NULL, NULL },
509 /*
510 * IPv4 multicast:
511 * protocol locks before interface locks, after UDP locks.
512 */
513 { "udpinp", &lock_class_rw },
514 { "in_multi_mtx", &lock_class_mtx_sleep },
515 { "igmp_mtx", &lock_class_mtx_sleep },
516 { "if_addr_lock", &lock_class_rw },
517 { NULL, NULL },
518 /*
519 * IPv6 multicast:
520 * protocol locks before interface locks, after UDP locks.
521 */
522 { "udpinp", &lock_class_rw },
523 { "in6_multi_mtx", &lock_class_mtx_sleep },
524 { "mld_mtx", &lock_class_mtx_sleep },
525 { "if_addr_lock", &lock_class_rw },
526 { NULL, NULL },
527 /*
528 * UNIX Domain Sockets
529 */
530 { "unp_global_rwlock", &lock_class_rw },
531 { "unp_list_lock", &lock_class_mtx_sleep },
532 { "unp", &lock_class_mtx_sleep },
533 { "so_snd", &lock_class_mtx_sleep },
534 { NULL, NULL },
535 /*
536 * UDP/IP
537 */
538 { "udp", &lock_class_rw },
539 { "udpinp", &lock_class_rw },
540 { "so_snd", &lock_class_mtx_sleep },
541 { NULL, NULL },
542 /*
543 * TCP/IP
544 */
545 { "tcp", &lock_class_rw },
546 { "tcpinp", &lock_class_rw },
547 { "so_snd", &lock_class_mtx_sleep },
548 { NULL, NULL },
549 /*
550 * netatalk
551 */
552 { "ddp_list_mtx", &lock_class_mtx_sleep },
553 { "ddp_mtx", &lock_class_mtx_sleep },
554 { NULL, NULL },
555 /*
556 * BPF
557 */
558 { "bpf global lock", &lock_class_mtx_sleep },
559 { "bpf interface lock", &lock_class_rw },
560 { "bpf cdev lock", &lock_class_mtx_sleep },
561 { NULL, NULL },
562 /*
563 * NFS server
564 */
565 { "nfsd_mtx", &lock_class_mtx_sleep },
566 { "so_snd", &lock_class_mtx_sleep },
567 { NULL, NULL },
568
569 /*
570 * IEEE 802.11
571 */
572 { "802.11 com lock", &lock_class_mtx_sleep},
573 { NULL, NULL },
574 /*
575 * Network drivers
576 */
577 { "network driver", &lock_class_mtx_sleep},
578 { NULL, NULL },
579
580 /*
581 * Netgraph
582 */
583 { "ng_node", &lock_class_mtx_sleep },
584 { "ng_worklist", &lock_class_mtx_sleep },
585 { NULL, NULL },
586 /*
587 * CDEV
588 */
589 { "vm map (system)", &lock_class_mtx_sleep },
590 { "vm page queue", &lock_class_mtx_sleep },
591 { "vnode interlock", &lock_class_mtx_sleep },
592 { "cdev", &lock_class_mtx_sleep },
593 { NULL, NULL },
594 /*
595 * VM
596 */
597 { "vm map (user)", &lock_class_sx },
598 { "vm object", &lock_class_rw },
599 { "vm page", &lock_class_mtx_sleep },
600 { "vm page queue", &lock_class_mtx_sleep },
601 { "pmap pv global", &lock_class_rw },
602 { "pmap", &lock_class_mtx_sleep },
603 { "pmap pv list", &lock_class_rw },
604 { "vm page free queue", &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 #ifdef SMP
655 { "smp rendezvous", &lock_class_mtx_spin },
656 #endif
657 #ifdef __powerpc__
658 { "tlb0", &lock_class_mtx_spin },
659 #endif
660 /*
661 * leaf locks
662 */
663 { "intrcnt", &lock_class_mtx_spin },
664 { "icu", &lock_class_mtx_spin },
665 #ifdef __i386__
666 { "allpmaps", &lock_class_mtx_spin },
667 { "descriptor tables", &lock_class_mtx_spin },
668 #endif
669 { "clk", &lock_class_mtx_spin },
670 { "cpuset", &lock_class_mtx_spin },
671 { "mprof lock", &lock_class_mtx_spin },
672 { "zombie lock", &lock_class_mtx_spin },
673 { "ALD Queue", &lock_class_mtx_spin },
674 #ifdef __ia64__
675 { "MCA spin lock", &lock_class_mtx_spin },
676 #endif
677 #if defined(__i386__) || defined(__amd64__)
678 { "pcicfg", &lock_class_mtx_spin },
679 { "NDIS thread lock", &lock_class_mtx_spin },
680 #endif
681 { "tw_osl_io_lock", &lock_class_mtx_spin },
682 { "tw_osl_q_lock", &lock_class_mtx_spin },
683 { "tw_cl_io_lock", &lock_class_mtx_spin },
684 { "tw_cl_intr_lock", &lock_class_mtx_spin },
685 { "tw_cl_gen_lock", &lock_class_mtx_spin },
686 #ifdef HWPMC_HOOKS
687 { "pmc-leaf", &lock_class_mtx_spin },
688 #endif
689 { "blocked lock", &lock_class_mtx_spin },
690 { NULL, NULL },
691 { NULL, NULL }
692 };
693
694 #ifdef BLESSING
695 /*
696 * Pairs of locks which have been blessed
697 * Don't complain about order problems with blessed locks
698 */
699 static struct witness_blessed blessed_list[] = {
700 };
701 static int blessed_count =
702 sizeof(blessed_list) / sizeof(struct witness_blessed);
703 #endif
704
705 /*
706 * This global is set to 0 once it becomes safe to use the witness code.
707 */
708 static int witness_cold = 1;
709
710 /*
711 * This global is set to 1 once the static lock orders have been enrolled
712 * so that a warning can be issued for any spin locks enrolled later.
713 */
714 static int witness_spin_warn = 0;
715
716 /* Trim useless garbage from filenames. */
717 static const char *
718 fixup_filename(const char *file)
719 {
720
721 if (file == NULL)
722 return (NULL);
723 while (strncmp(file, "../", 3) == 0)
724 file += 3;
725 return (file);
726 }
727
728 /*
729 * The WITNESS-enabled diagnostic code. Note that the witness code does
730 * assume that the early boot is single-threaded at least until after this
731 * routine is completed.
732 */
733 static void
734 witness_initialize(void *dummy __unused)
735 {
736 struct lock_object *lock;
737 struct witness_order_list_entry *order;
738 struct witness *w, *w1;
739 int i;
740
741 w_data = malloc(sizeof (struct witness) * WITNESS_COUNT, M_WITNESS,
742 M_NOWAIT | M_ZERO);
743
744 /*
745 * We have to release Giant before initializing its witness
746 * structure so that WITNESS doesn't get confused.
747 */
748 mtx_unlock(&Giant);
749 mtx_assert(&Giant, MA_NOTOWNED);
750
751 CTR1(KTR_WITNESS, "%s: initializing witness", __func__);
752 mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET |
753 MTX_NOWITNESS | MTX_NOPROFILE);
754 for (i = WITNESS_COUNT - 1; i >= 0; i--) {
755 w = &w_data[i];
756 memset(w, 0, sizeof(*w));
757 w_data[i].w_index = i; /* Witness index never changes. */
758 witness_free(w);
759 }
760 KASSERT(STAILQ_FIRST(&w_free)->w_index == 0,
761 ("%s: Invalid list of free witness objects", __func__));
762
763 /* Witness with index 0 is not used to aid in debugging. */
764 STAILQ_REMOVE_HEAD(&w_free, w_list);
765 w_free_cnt--;
766
767 memset(w_rmatrix, 0,
768 (sizeof(**w_rmatrix) * (WITNESS_COUNT+1) * (WITNESS_COUNT+1)));
769
770 for (i = 0; i < LOCK_CHILDCOUNT; i++)
771 witness_lock_list_free(&w_locklistdata[i]);
772 witness_init_hash_tables();
773
774 /* First add in all the specified order lists. */
775 for (order = order_lists; order->w_name != NULL; order++) {
776 w = enroll(order->w_name, order->w_class);
777 if (w == NULL)
778 continue;
779 w->w_file = "order list";
780 for (order++; order->w_name != NULL; order++) {
781 w1 = enroll(order->w_name, order->w_class);
782 if (w1 == NULL)
783 continue;
784 w1->w_file = "order list";
785 itismychild(w, w1);
786 w = w1;
787 }
788 }
789 witness_spin_warn = 1;
790
791 /* Iterate through all locks and add them to witness. */
792 for (i = 0; pending_locks[i].wh_lock != NULL; i++) {
793 lock = pending_locks[i].wh_lock;
794 KASSERT(lock->lo_flags & LO_WITNESS,
795 ("%s: lock %s is on pending list but not LO_WITNESS",
796 __func__, lock->lo_name));
797 lock->lo_witness = enroll(pending_locks[i].wh_type,
798 LOCK_CLASS(lock));
799 }
800
801 /* Mark the witness code as being ready for use. */
802 witness_cold = 0;
803
804 mtx_lock(&Giant);
805 }
806 SYSINIT(witness_init, SI_SUB_WITNESS, SI_ORDER_FIRST, witness_initialize,
807 NULL);
808
809 void
810 witness_init(struct lock_object *lock, const char *type)
811 {
812 struct lock_class *class;
813
814 /* Various sanity checks. */
815 class = LOCK_CLASS(lock);
816 if ((lock->lo_flags & LO_RECURSABLE) != 0 &&
817 (class->lc_flags & LC_RECURSABLE) == 0)
818 kassert_panic("%s: lock (%s) %s can not be recursable",
819 __func__, class->lc_name, lock->lo_name);
820 if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
821 (class->lc_flags & LC_SLEEPABLE) == 0)
822 kassert_panic("%s: lock (%s) %s can not be sleepable",
823 __func__, class->lc_name, lock->lo_name);
824 if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
825 (class->lc_flags & LC_UPGRADABLE) == 0)
826 kassert_panic("%s: lock (%s) %s can not be upgradable",
827 __func__, class->lc_name, lock->lo_name);
828
829 /*
830 * If we shouldn't watch this lock, then just clear lo_witness.
831 * Otherwise, if witness_cold is set, then it is too early to
832 * enroll this lock, so defer it to witness_initialize() by adding
833 * it to the pending_locks list. If it is not too early, then enroll
834 * the lock now.
835 */
836 if (witness_watch < 1 || panicstr != NULL ||
837 (lock->lo_flags & LO_WITNESS) == 0)
838 lock->lo_witness = NULL;
839 else if (witness_cold) {
840 pending_locks[pending_cnt].wh_lock = lock;
841 pending_locks[pending_cnt++].wh_type = type;
842 if (pending_cnt > WITNESS_PENDLIST)
843 panic("%s: pending locks list is too small, "
844 "increase WITNESS_PENDLIST\n",
845 __func__);
846 } else
847 lock->lo_witness = enroll(type, class);
848 }
849
850 void
851 witness_destroy(struct lock_object *lock)
852 {
853 struct lock_class *class;
854 struct witness *w;
855
856 class = LOCK_CLASS(lock);
857
858 if (witness_cold)
859 panic("lock (%s) %s destroyed while witness_cold",
860 class->lc_name, lock->lo_name);
861
862 /* XXX: need to verify that no one holds the lock */
863 if ((lock->lo_flags & LO_WITNESS) == 0 || lock->lo_witness == NULL)
864 return;
865 w = lock->lo_witness;
866
867 mtx_lock_spin(&w_mtx);
868 MPASS(w->w_refcount > 0);
869 w->w_refcount--;
870
871 if (w->w_refcount == 0)
872 depart(w);
873 mtx_unlock_spin(&w_mtx);
874 }
875
876 #ifdef DDB
877 static void
878 witness_ddb_compute_levels(void)
879 {
880 struct witness *w;
881
882 /*
883 * First clear all levels.
884 */
885 STAILQ_FOREACH(w, &w_all, w_list)
886 w->w_ddb_level = -1;
887
888 /*
889 * Look for locks with no parents and level all their descendants.
890 */
891 STAILQ_FOREACH(w, &w_all, w_list) {
892
893 /* If the witness has ancestors (is not a root), skip it. */
894 if (w->w_num_ancestors > 0)
895 continue;
896 witness_ddb_level_descendants(w, 0);
897 }
898 }
899
900 static void
901 witness_ddb_level_descendants(struct witness *w, int l)
902 {
903 int i;
904
905 if (w->w_ddb_level >= l)
906 return;
907
908 w->w_ddb_level = l;
909 l++;
910
911 for (i = 1; i <= w_max_used_index; i++) {
912 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
913 witness_ddb_level_descendants(&w_data[i], l);
914 }
915 }
916
917 static void
918 witness_ddb_display_descendants(int(*prnt)(const char *fmt, ...),
919 struct witness *w, int indent)
920 {
921 int i;
922
923 for (i = 0; i < indent; i++)
924 prnt(" ");
925 prnt("%s (type: %s, depth: %d, active refs: %d)",
926 w->w_name, w->w_class->lc_name,
927 w->w_ddb_level, w->w_refcount);
928 if (w->w_displayed) {
929 prnt(" -- (already displayed)\n");
930 return;
931 }
932 w->w_displayed = 1;
933 if (w->w_file != NULL && w->w_line != 0)
934 prnt(" -- last acquired @ %s:%d\n", fixup_filename(w->w_file),
935 w->w_line);
936 else
937 prnt(" -- never acquired\n");
938 indent++;
939 WITNESS_INDEX_ASSERT(w->w_index);
940 for (i = 1; i <= w_max_used_index; i++) {
941 if (db_pager_quit)
942 return;
943 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
944 witness_ddb_display_descendants(prnt, &w_data[i],
945 indent);
946 }
947 }
948
949 static void
950 witness_ddb_display_list(int(*prnt)(const char *fmt, ...),
951 struct witness_list *list)
952 {
953 struct witness *w;
954
955 STAILQ_FOREACH(w, list, w_typelist) {
956 if (w->w_file == NULL || w->w_ddb_level > 0)
957 continue;
958
959 /* This lock has no anscestors - display its descendants. */
960 witness_ddb_display_descendants(prnt, w, 0);
961 if (db_pager_quit)
962 return;
963 }
964 }
965
966 static void
967 witness_ddb_display(int(*prnt)(const char *fmt, ...))
968 {
969 struct witness *w;
970
971 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
972 witness_ddb_compute_levels();
973
974 /* Clear all the displayed flags. */
975 STAILQ_FOREACH(w, &w_all, w_list)
976 w->w_displayed = 0;
977
978 /*
979 * First, handle sleep locks which have been acquired at least
980 * once.
981 */
982 prnt("Sleep locks:\n");
983 witness_ddb_display_list(prnt, &w_sleep);
984 if (db_pager_quit)
985 return;
986
987 /*
988 * Now do spin locks which have been acquired at least once.
989 */
990 prnt("\nSpin locks:\n");
991 witness_ddb_display_list(prnt, &w_spin);
992 if (db_pager_quit)
993 return;
994
995 /*
996 * Finally, any locks which have not been acquired yet.
997 */
998 prnt("\nLocks which were never acquired:\n");
999 STAILQ_FOREACH(w, &w_all, w_list) {
1000 if (w->w_file != NULL || w->w_refcount == 0)
1001 continue;
1002 prnt("%s (type: %s, depth: %d)\n", w->w_name,
1003 w->w_class->lc_name, w->w_ddb_level);
1004 if (db_pager_quit)
1005 return;
1006 }
1007 }
1008 #endif /* DDB */
1009
1010 int
1011 witness_defineorder(struct lock_object *lock1, struct lock_object *lock2)
1012 {
1013
1014 if (witness_watch == -1 || panicstr != NULL)
1015 return (0);
1016
1017 /* Require locks that witness knows about. */
1018 if (lock1 == NULL || lock1->lo_witness == NULL || lock2 == NULL ||
1019 lock2->lo_witness == NULL)
1020 return (EINVAL);
1021
1022 mtx_assert(&w_mtx, MA_NOTOWNED);
1023 mtx_lock_spin(&w_mtx);
1024
1025 /*
1026 * If we already have either an explicit or implied lock order that
1027 * is the other way around, then return an error.
1028 */
1029 if (witness_watch &&
1030 isitmydescendant(lock2->lo_witness, lock1->lo_witness)) {
1031 mtx_unlock_spin(&w_mtx);
1032 return (EDOOFUS);
1033 }
1034
1035 /* Try to add the new order. */
1036 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1037 lock2->lo_witness->w_name, lock1->lo_witness->w_name);
1038 itismychild(lock1->lo_witness, lock2->lo_witness);
1039 mtx_unlock_spin(&w_mtx);
1040 return (0);
1041 }
1042
1043 void
1044 witness_checkorder(struct lock_object *lock, int flags, const char *file,
1045 int line, struct lock_object *interlock)
1046 {
1047 struct lock_list_entry *lock_list, *lle;
1048 struct lock_instance *lock1, *lock2, *plock;
1049 struct lock_class *class, *iclass;
1050 struct witness *w, *w1;
1051 struct thread *td;
1052 int i, j;
1053
1054 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL ||
1055 panicstr != NULL)
1056 return;
1057
1058 w = lock->lo_witness;
1059 class = LOCK_CLASS(lock);
1060 td = curthread;
1061
1062 if (class->lc_flags & LC_SLEEPLOCK) {
1063
1064 /*
1065 * Since spin locks include a critical section, this check
1066 * implicitly enforces a lock order of all sleep locks before
1067 * all spin locks.
1068 */
1069 if (td->td_critnest != 0 && !kdb_active)
1070 kassert_panic("acquiring blockable sleep lock with "
1071 "spinlock or critical section held (%s) %s @ %s:%d",
1072 class->lc_name, lock->lo_name,
1073 fixup_filename(file), line);
1074
1075 /*
1076 * If this is the first lock acquired then just return as
1077 * no order checking is needed.
1078 */
1079 lock_list = td->td_sleeplocks;
1080 if (lock_list == NULL || lock_list->ll_count == 0)
1081 return;
1082 } else {
1083
1084 /*
1085 * If this is the first lock, just return as no order
1086 * checking is needed. Avoid problems with thread
1087 * migration pinning the thread while checking if
1088 * spinlocks are held. If at least one spinlock is held
1089 * the thread is in a safe path and it is allowed to
1090 * unpin it.
1091 */
1092 sched_pin();
1093 lock_list = PCPU_GET(spinlocks);
1094 if (lock_list == NULL || lock_list->ll_count == 0) {
1095 sched_unpin();
1096 return;
1097 }
1098 sched_unpin();
1099 }
1100
1101 /*
1102 * Check to see if we are recursing on a lock we already own. If
1103 * so, make sure that we don't mismatch exclusive and shared lock
1104 * acquires.
1105 */
1106 lock1 = find_instance(lock_list, lock);
1107 if (lock1 != NULL) {
1108 if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
1109 (flags & LOP_EXCLUSIVE) == 0) {
1110 printf("shared lock of (%s) %s @ %s:%d\n",
1111 class->lc_name, lock->lo_name,
1112 fixup_filename(file), line);
1113 printf("while exclusively locked from %s:%d\n",
1114 fixup_filename(lock1->li_file), lock1->li_line);
1115 kassert_panic("excl->share");
1116 }
1117 if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
1118 (flags & LOP_EXCLUSIVE) != 0) {
1119 printf("exclusive lock of (%s) %s @ %s:%d\n",
1120 class->lc_name, lock->lo_name,
1121 fixup_filename(file), line);
1122 printf("while share locked from %s:%d\n",
1123 fixup_filename(lock1->li_file), lock1->li_line);
1124 kassert_panic("share->excl");
1125 }
1126 return;
1127 }
1128
1129 /* Warn if the interlock is not locked exactly once. */
1130 if (interlock != NULL) {
1131 iclass = LOCK_CLASS(interlock);
1132 lock1 = find_instance(lock_list, interlock);
1133 if (lock1 == NULL)
1134 kassert_panic("interlock (%s) %s not locked @ %s:%d",
1135 iclass->lc_name, interlock->lo_name,
1136 fixup_filename(file), line);
1137 else if ((lock1->li_flags & LI_RECURSEMASK) != 0)
1138 kassert_panic("interlock (%s) %s recursed @ %s:%d",
1139 iclass->lc_name, interlock->lo_name,
1140 fixup_filename(file), line);
1141 }
1142
1143 /*
1144 * Find the previously acquired lock, but ignore interlocks.
1145 */
1146 plock = &lock_list->ll_children[lock_list->ll_count - 1];
1147 if (interlock != NULL && plock->li_lock == interlock) {
1148 if (lock_list->ll_count > 1)
1149 plock =
1150 &lock_list->ll_children[lock_list->ll_count - 2];
1151 else {
1152 lle = lock_list->ll_next;
1153
1154 /*
1155 * The interlock is the only lock we hold, so
1156 * simply return.
1157 */
1158 if (lle == NULL)
1159 return;
1160 plock = &lle->ll_children[lle->ll_count - 1];
1161 }
1162 }
1163
1164 /*
1165 * Try to perform most checks without a lock. If this succeeds we
1166 * can skip acquiring the lock and return success.
1167 */
1168 w1 = plock->li_lock->lo_witness;
1169 if (witness_lock_order_check(w1, w))
1170 return;
1171
1172 /*
1173 * Check for duplicate locks of the same type. Note that we only
1174 * have to check for this on the last lock we just acquired. Any
1175 * other cases will be caught as lock order violations.
1176 */
1177 mtx_lock_spin(&w_mtx);
1178 witness_lock_order_add(w1, w);
1179 if (w1 == w) {
1180 i = w->w_index;
1181 if (!(lock->lo_flags & LO_DUPOK) && !(flags & LOP_DUPOK) &&
1182 !(w_rmatrix[i][i] & WITNESS_REVERSAL)) {
1183 w_rmatrix[i][i] |= WITNESS_REVERSAL;
1184 w->w_reversed = 1;
1185 mtx_unlock_spin(&w_mtx);
1186 printf(
1187 "acquiring duplicate lock of same type: \"%s\"\n",
1188 w->w_name);
1189 printf(" 1st %s @ %s:%d\n", plock->li_lock->lo_name,
1190 fixup_filename(plock->li_file), plock->li_line);
1191 printf(" 2nd %s @ %s:%d\n", lock->lo_name,
1192 fixup_filename(file), line);
1193 witness_debugger(1);
1194 } else
1195 mtx_unlock_spin(&w_mtx);
1196 return;
1197 }
1198 mtx_assert(&w_mtx, MA_OWNED);
1199
1200 /*
1201 * If we know that the lock we are acquiring comes after
1202 * the lock we most recently acquired in the lock order tree,
1203 * then there is no need for any further checks.
1204 */
1205 if (isitmychild(w1, w))
1206 goto out;
1207
1208 for (j = 0, lle = lock_list; lle != NULL; lle = lle->ll_next) {
1209 for (i = lle->ll_count - 1; i >= 0; i--, j++) {
1210
1211 MPASS(j < WITNESS_COUNT);
1212 lock1 = &lle->ll_children[i];
1213
1214 /*
1215 * Ignore the interlock.
1216 */
1217 if (interlock == lock1->li_lock)
1218 continue;
1219
1220 /*
1221 * If this lock doesn't undergo witness checking,
1222 * then skip it.
1223 */
1224 w1 = lock1->li_lock->lo_witness;
1225 if (w1 == NULL) {
1226 KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
1227 ("lock missing witness structure"));
1228 continue;
1229 }
1230
1231 /*
1232 * If we are locking Giant and this is a sleepable
1233 * lock, then skip it.
1234 */
1235 if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0 &&
1236 lock == &Giant.lock_object)
1237 continue;
1238
1239 /*
1240 * If we are locking a sleepable lock and this lock
1241 * is Giant, then skip it.
1242 */
1243 if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1244 lock1->li_lock == &Giant.lock_object)
1245 continue;
1246
1247 /*
1248 * If we are locking a sleepable lock and this lock
1249 * isn't sleepable, we want to treat it as a lock
1250 * order violation to enfore a general lock order of
1251 * sleepable locks before non-sleepable locks.
1252 */
1253 if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1254 (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
1255 goto reversal;
1256
1257 /*
1258 * If we are locking Giant and this is a non-sleepable
1259 * lock, then treat it as a reversal.
1260 */
1261 if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0 &&
1262 lock == &Giant.lock_object)
1263 goto reversal;
1264
1265 /*
1266 * Check the lock order hierarchy for a reveresal.
1267 */
1268 if (!isitmydescendant(w, w1))
1269 continue;
1270 reversal:
1271
1272 /*
1273 * We have a lock order violation, check to see if it
1274 * is allowed or has already been yelled about.
1275 */
1276 #ifdef BLESSING
1277
1278 /*
1279 * If the lock order is blessed, just bail. We don't
1280 * look for other lock order violations though, which
1281 * may be a bug.
1282 */
1283 if (blessed(w, w1))
1284 goto out;
1285 #endif
1286
1287 /* Bail if this violation is known */
1288 if (w_rmatrix[w1->w_index][w->w_index] & WITNESS_REVERSAL)
1289 goto out;
1290
1291 /* Record this as a violation */
1292 w_rmatrix[w1->w_index][w->w_index] |= WITNESS_REVERSAL;
1293 w_rmatrix[w->w_index][w1->w_index] |= WITNESS_REVERSAL;
1294 w->w_reversed = w1->w_reversed = 1;
1295 witness_increment_graph_generation();
1296 mtx_unlock_spin(&w_mtx);
1297
1298 #ifdef WITNESS_NO_VNODE
1299 /*
1300 * There are known LORs between VNODE locks. They are
1301 * not an indication of a bug. VNODE locks are flagged
1302 * as such (LO_IS_VNODE) and we don't yell if the LOR
1303 * is between 2 VNODE locks.
1304 */
1305 if ((lock->lo_flags & LO_IS_VNODE) != 0 &&
1306 (lock1->li_lock->lo_flags & LO_IS_VNODE) != 0)
1307 return;
1308 #endif
1309
1310 /*
1311 * Ok, yell about it.
1312 */
1313 if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1314 (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
1315 printf(
1316 "lock order reversal: (sleepable after non-sleepable)\n");
1317 else if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0
1318 && lock == &Giant.lock_object)
1319 printf(
1320 "lock order reversal: (Giant after non-sleepable)\n");
1321 else
1322 printf("lock order reversal:\n");
1323
1324 /*
1325 * Try to locate an earlier lock with
1326 * witness w in our list.
1327 */
1328 do {
1329 lock2 = &lle->ll_children[i];
1330 MPASS(lock2->li_lock != NULL);
1331 if (lock2->li_lock->lo_witness == w)
1332 break;
1333 if (i == 0 && lle->ll_next != NULL) {
1334 lle = lle->ll_next;
1335 i = lle->ll_count - 1;
1336 MPASS(i >= 0 && i < LOCK_NCHILDREN);
1337 } else
1338 i--;
1339 } while (i >= 0);
1340 if (i < 0) {
1341 printf(" 1st %p %s (%s) @ %s:%d\n",
1342 lock1->li_lock, lock1->li_lock->lo_name,
1343 w1->w_name, fixup_filename(lock1->li_file),
1344 lock1->li_line);
1345 printf(" 2nd %p %s (%s) @ %s:%d\n", lock,
1346 lock->lo_name, w->w_name,
1347 fixup_filename(file), line);
1348 } else {
1349 printf(" 1st %p %s (%s) @ %s:%d\n",
1350 lock2->li_lock, lock2->li_lock->lo_name,
1351 lock2->li_lock->lo_witness->w_name,
1352 fixup_filename(lock2->li_file),
1353 lock2->li_line);
1354 printf(" 2nd %p %s (%s) @ %s:%d\n",
1355 lock1->li_lock, lock1->li_lock->lo_name,
1356 w1->w_name, fixup_filename(lock1->li_file),
1357 lock1->li_line);
1358 printf(" 3rd %p %s (%s) @ %s:%d\n", lock,
1359 lock->lo_name, w->w_name,
1360 fixup_filename(file), line);
1361 }
1362 witness_debugger(1);
1363 return;
1364 }
1365 }
1366
1367 /*
1368 * If requested, build a new lock order. However, don't build a new
1369 * relationship between a sleepable lock and Giant if it is in the
1370 * wrong direction. The correct lock order is that sleepable locks
1371 * always come before Giant.
1372 */
1373 if (flags & LOP_NEWORDER &&
1374 !(plock->li_lock == &Giant.lock_object &&
1375 (lock->lo_flags & LO_SLEEPABLE) != 0)) {
1376 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1377 w->w_name, plock->li_lock->lo_witness->w_name);
1378 itismychild(plock->li_lock->lo_witness, w);
1379 }
1380 out:
1381 mtx_unlock_spin(&w_mtx);
1382 }
1383
1384 void
1385 witness_lock(struct lock_object *lock, int flags, const char *file, int line)
1386 {
1387 struct lock_list_entry **lock_list, *lle;
1388 struct lock_instance *instance;
1389 struct witness *w;
1390 struct thread *td;
1391
1392 if (witness_cold || witness_watch == -1 || lock->lo_witness == NULL ||
1393 panicstr != NULL)
1394 return;
1395 w = lock->lo_witness;
1396 td = curthread;
1397
1398 /* Determine lock list for this lock. */
1399 if (LOCK_CLASS(lock)->lc_flags & LC_SLEEPLOCK)
1400 lock_list = &td->td_sleeplocks;
1401 else
1402 lock_list = PCPU_PTR(spinlocks);
1403
1404 /* Check to see if we are recursing on a lock we already own. */
1405 instance = find_instance(*lock_list, lock);
1406 if (instance != NULL) {
1407 instance->li_flags++;
1408 CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
1409 td->td_proc->p_pid, lock->lo_name,
1410 instance->li_flags & LI_RECURSEMASK);
1411 instance->li_file = file;
1412 instance->li_line = line;
1413 return;
1414 }
1415
1416 /* Update per-witness last file and line acquire. */
1417 w->w_file = file;
1418 w->w_line = line;
1419
1420 /* Find the next open lock instance in the list and fill it. */
1421 lle = *lock_list;
1422 if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
1423 lle = witness_lock_list_get();
1424 if (lle == NULL)
1425 return;
1426 lle->ll_next = *lock_list;
1427 CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
1428 td->td_proc->p_pid, lle);
1429 *lock_list = lle;
1430 }
1431 instance = &lle->ll_children[lle->ll_count++];
1432 instance->li_lock = lock;
1433 instance->li_line = line;
1434 instance->li_file = file;
1435 if ((flags & LOP_EXCLUSIVE) != 0)
1436 instance->li_flags = LI_EXCLUSIVE;
1437 else
1438 instance->li_flags = 0;
1439 CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
1440 td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
1441 }
1442
1443 void
1444 witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
1445 {
1446 struct lock_instance *instance;
1447 struct lock_class *class;
1448
1449 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1450 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
1451 return;
1452 class = LOCK_CLASS(lock);
1453 if (witness_watch) {
1454 if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1455 kassert_panic(
1456 "upgrade of non-upgradable lock (%s) %s @ %s:%d",
1457 class->lc_name, lock->lo_name,
1458 fixup_filename(file), line);
1459 if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1460 kassert_panic(
1461 "upgrade of non-sleep lock (%s) %s @ %s:%d",
1462 class->lc_name, lock->lo_name,
1463 fixup_filename(file), line);
1464 }
1465 instance = find_instance(curthread->td_sleeplocks, lock);
1466 if (instance == NULL) {
1467 kassert_panic("upgrade of unlocked lock (%s) %s @ %s:%d",
1468 class->lc_name, lock->lo_name,
1469 fixup_filename(file), line);
1470 return;
1471 }
1472 if (witness_watch) {
1473 if ((instance->li_flags & LI_EXCLUSIVE) != 0)
1474 kassert_panic(
1475 "upgrade of exclusive lock (%s) %s @ %s:%d",
1476 class->lc_name, lock->lo_name,
1477 fixup_filename(file), line);
1478 if ((instance->li_flags & LI_RECURSEMASK) != 0)
1479 kassert_panic(
1480 "upgrade of recursed lock (%s) %s r=%d @ %s:%d",
1481 class->lc_name, lock->lo_name,
1482 instance->li_flags & LI_RECURSEMASK,
1483 fixup_filename(file), line);
1484 }
1485 instance->li_flags |= LI_EXCLUSIVE;
1486 }
1487
1488 void
1489 witness_downgrade(struct lock_object *lock, int flags, const char *file,
1490 int line)
1491 {
1492 struct lock_instance *instance;
1493 struct lock_class *class;
1494
1495 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1496 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
1497 return;
1498 class = LOCK_CLASS(lock);
1499 if (witness_watch) {
1500 if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1501 kassert_panic(
1502 "downgrade of non-upgradable lock (%s) %s @ %s:%d",
1503 class->lc_name, lock->lo_name,
1504 fixup_filename(file), line);
1505 if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1506 kassert_panic(
1507 "downgrade of non-sleep lock (%s) %s @ %s:%d",
1508 class->lc_name, lock->lo_name,
1509 fixup_filename(file), line);
1510 }
1511 instance = find_instance(curthread->td_sleeplocks, lock);
1512 if (instance == NULL) {
1513 kassert_panic("downgrade of unlocked lock (%s) %s @ %s:%d",
1514 class->lc_name, lock->lo_name,
1515 fixup_filename(file), line);
1516 return;
1517 }
1518 if (witness_watch) {
1519 if ((instance->li_flags & LI_EXCLUSIVE) == 0)
1520 kassert_panic(
1521 "downgrade of shared lock (%s) %s @ %s:%d",
1522 class->lc_name, lock->lo_name,
1523 fixup_filename(file), line);
1524 if ((instance->li_flags & LI_RECURSEMASK) != 0)
1525 kassert_panic(
1526 "downgrade of recursed lock (%s) %s r=%d @ %s:%d",
1527 class->lc_name, lock->lo_name,
1528 instance->li_flags & LI_RECURSEMASK,
1529 fixup_filename(file), line);
1530 }
1531 instance->li_flags &= ~LI_EXCLUSIVE;
1532 }
1533
1534 void
1535 witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
1536 {
1537 struct lock_list_entry **lock_list, *lle;
1538 struct lock_instance *instance;
1539 struct lock_class *class;
1540 struct thread *td;
1541 register_t s;
1542 int i, j;
1543
1544 if (witness_cold || lock->lo_witness == NULL || panicstr != NULL)
1545 return;
1546 td = curthread;
1547 class = LOCK_CLASS(lock);
1548
1549 /* Find lock instance associated with this lock. */
1550 if (class->lc_flags & LC_SLEEPLOCK)
1551 lock_list = &td->td_sleeplocks;
1552 else
1553 lock_list = PCPU_PTR(spinlocks);
1554 lle = *lock_list;
1555 for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
1556 for (i = 0; i < (*lock_list)->ll_count; i++) {
1557 instance = &(*lock_list)->ll_children[i];
1558 if (instance->li_lock == lock)
1559 goto found;
1560 }
1561
1562 /*
1563 * When disabling WITNESS through witness_watch we could end up in
1564 * having registered locks in the td_sleeplocks queue.
1565 * We have to make sure we flush these queues, so just search for
1566 * eventual register locks and remove them.
1567 */
1568 if (witness_watch > 0) {
1569 kassert_panic("lock (%s) %s not locked @ %s:%d", class->lc_name,
1570 lock->lo_name, fixup_filename(file), line);
1571 return;
1572 } else {
1573 return;
1574 }
1575 found:
1576
1577 /* First, check for shared/exclusive mismatches. */
1578 if ((instance->li_flags & LI_EXCLUSIVE) != 0 && witness_watch > 0 &&
1579 (flags & LOP_EXCLUSIVE) == 0) {
1580 printf("shared unlock of (%s) %s @ %s:%d\n", class->lc_name,
1581 lock->lo_name, fixup_filename(file), line);
1582 printf("while exclusively locked from %s:%d\n",
1583 fixup_filename(instance->li_file), instance->li_line);
1584 kassert_panic("excl->ushare");
1585 }
1586 if ((instance->li_flags & LI_EXCLUSIVE) == 0 && witness_watch > 0 &&
1587 (flags & LOP_EXCLUSIVE) != 0) {
1588 printf("exclusive unlock of (%s) %s @ %s:%d\n", class->lc_name,
1589 lock->lo_name, fixup_filename(file), line);
1590 printf("while share locked from %s:%d\n",
1591 fixup_filename(instance->li_file),
1592 instance->li_line);
1593 kassert_panic("share->uexcl");
1594 }
1595 /* If we are recursed, unrecurse. */
1596 if ((instance->li_flags & LI_RECURSEMASK) > 0) {
1597 CTR4(KTR_WITNESS, "%s: pid %d unrecursed on %s r=%d", __func__,
1598 td->td_proc->p_pid, instance->li_lock->lo_name,
1599 instance->li_flags);
1600 instance->li_flags--;
1601 return;
1602 }
1603 /* The lock is now being dropped, check for NORELEASE flag */
1604 if ((instance->li_flags & LI_NORELEASE) != 0 && witness_watch > 0) {
1605 printf("forbidden unlock of (%s) %s @ %s:%d\n", class->lc_name,
1606 lock->lo_name, fixup_filename(file), line);
1607 kassert_panic("lock marked norelease");
1608 }
1609
1610 /* Otherwise, remove this item from the list. */
1611 s = intr_disable();
1612 CTR4(KTR_WITNESS, "%s: pid %d removed %s from lle[%d]", __func__,
1613 td->td_proc->p_pid, instance->li_lock->lo_name,
1614 (*lock_list)->ll_count - 1);
1615 for (j = i; j < (*lock_list)->ll_count - 1; j++)
1616 (*lock_list)->ll_children[j] =
1617 (*lock_list)->ll_children[j + 1];
1618 (*lock_list)->ll_count--;
1619 intr_restore(s);
1620
1621 /*
1622 * In order to reduce contention on w_mtx, we want to keep always an
1623 * head object into lists so that frequent allocation from the
1624 * free witness pool (and subsequent locking) is avoided.
1625 * In order to maintain the current code simple, when the head
1626 * object is totally unloaded it means also that we do not have
1627 * further objects in the list, so the list ownership needs to be
1628 * hand over to another object if the current head needs to be freed.
1629 */
1630 if ((*lock_list)->ll_count == 0) {
1631 if (*lock_list == lle) {
1632 if (lle->ll_next == NULL)
1633 return;
1634 } else
1635 lle = *lock_list;
1636 *lock_list = lle->ll_next;
1637 CTR3(KTR_WITNESS, "%s: pid %d removed lle %p", __func__,
1638 td->td_proc->p_pid, lle);
1639 witness_lock_list_free(lle);
1640 }
1641 }
1642
1643 void
1644 witness_thread_exit(struct thread *td)
1645 {
1646 struct lock_list_entry *lle;
1647 int i, n;
1648
1649 lle = td->td_sleeplocks;
1650 if (lle == NULL || panicstr != NULL)
1651 return;
1652 if (lle->ll_count != 0) {
1653 for (n = 0; lle != NULL; lle = lle->ll_next)
1654 for (i = lle->ll_count - 1; i >= 0; i--) {
1655 if (n == 0)
1656 printf("Thread %p exiting with the following locks held:\n",
1657 td);
1658 n++;
1659 witness_list_lock(&lle->ll_children[i], printf);
1660
1661 }
1662 kassert_panic(
1663 "Thread %p cannot exit while holding sleeplocks\n", td);
1664 }
1665 witness_lock_list_free(lle);
1666 }
1667
1668 /*
1669 * Warn if any locks other than 'lock' are held. Flags can be passed in to
1670 * exempt Giant and sleepable locks from the checks as well. If any
1671 * non-exempt locks are held, then a supplied message is printed to the
1672 * console along with a list of the offending locks. If indicated in the
1673 * flags then a failure results in a panic as well.
1674 */
1675 int
1676 witness_warn(int flags, struct lock_object *lock, const char *fmt, ...)
1677 {
1678 struct lock_list_entry *lock_list, *lle;
1679 struct lock_instance *lock1;
1680 struct thread *td;
1681 va_list ap;
1682 int i, n;
1683
1684 if (witness_cold || witness_watch < 1 || panicstr != NULL)
1685 return (0);
1686 n = 0;
1687 td = curthread;
1688 for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next)
1689 for (i = lle->ll_count - 1; i >= 0; i--) {
1690 lock1 = &lle->ll_children[i];
1691 if (lock1->li_lock == lock)
1692 continue;
1693 if (flags & WARN_GIANTOK &&
1694 lock1->li_lock == &Giant.lock_object)
1695 continue;
1696 if (flags & WARN_SLEEPOK &&
1697 (lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0)
1698 continue;
1699 if (n == 0) {
1700 va_start(ap, fmt);
1701 vprintf(fmt, ap);
1702 va_end(ap);
1703 printf(" with the following");
1704 if (flags & WARN_SLEEPOK)
1705 printf(" non-sleepable");
1706 printf(" locks held:\n");
1707 }
1708 n++;
1709 witness_list_lock(lock1, printf);
1710 }
1711
1712 /*
1713 * Pin the thread in order to avoid problems with thread migration.
1714 * Once that all verifies are passed about spinlocks ownership,
1715 * the thread is in a safe path and it can be unpinned.
1716 */
1717 sched_pin();
1718 lock_list = PCPU_GET(spinlocks);
1719 if (lock_list != NULL && lock_list->ll_count != 0) {
1720 sched_unpin();
1721
1722 /*
1723 * We should only have one spinlock and as long as
1724 * the flags cannot match for this locks class,
1725 * check if the first spinlock is the one curthread
1726 * should hold.
1727 */
1728 lock1 = &lock_list->ll_children[lock_list->ll_count - 1];
1729 if (lock_list->ll_count == 1 && lock_list->ll_next == NULL &&
1730 lock1->li_lock == lock && n == 0)
1731 return (0);
1732
1733 va_start(ap, fmt);
1734 vprintf(fmt, ap);
1735 va_end(ap);
1736 printf(" with the following");
1737 if (flags & WARN_SLEEPOK)
1738 printf(" non-sleepable");
1739 printf(" locks held:\n");
1740 n += witness_list_locks(&lock_list, printf);
1741 } else
1742 sched_unpin();
1743 if (flags & WARN_PANIC && n)
1744 kassert_panic("%s", __func__);
1745 else
1746 witness_debugger(n);
1747 return (n);
1748 }
1749
1750 const char *
1751 witness_file(struct lock_object *lock)
1752 {
1753 struct witness *w;
1754
1755 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1756 return ("?");
1757 w = lock->lo_witness;
1758 return (w->w_file);
1759 }
1760
1761 int
1762 witness_line(struct lock_object *lock)
1763 {
1764 struct witness *w;
1765
1766 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1767 return (0);
1768 w = lock->lo_witness;
1769 return (w->w_line);
1770 }
1771
1772 static struct witness *
1773 enroll(const char *description, struct lock_class *lock_class)
1774 {
1775 struct witness *w;
1776 struct witness_list *typelist;
1777
1778 MPASS(description != NULL);
1779
1780 if (witness_watch == -1 || panicstr != NULL)
1781 return (NULL);
1782 if ((lock_class->lc_flags & LC_SPINLOCK)) {
1783 if (witness_skipspin)
1784 return (NULL);
1785 else
1786 typelist = &w_spin;
1787 } else if ((lock_class->lc_flags & LC_SLEEPLOCK)) {
1788 typelist = &w_sleep;
1789 } else {
1790 kassert_panic("lock class %s is not sleep or spin",
1791 lock_class->lc_name);
1792 return (NULL);
1793 }
1794
1795 mtx_lock_spin(&w_mtx);
1796 w = witness_hash_get(description);
1797 if (w)
1798 goto found;
1799 if ((w = witness_get()) == NULL)
1800 return (NULL);
1801 MPASS(strlen(description) < MAX_W_NAME);
1802 strcpy(w->w_name, description);
1803 w->w_class = lock_class;
1804 w->w_refcount = 1;
1805 STAILQ_INSERT_HEAD(&w_all, w, w_list);
1806 if (lock_class->lc_flags & LC_SPINLOCK) {
1807 STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
1808 w_spin_cnt++;
1809 } else if (lock_class->lc_flags & LC_SLEEPLOCK) {
1810 STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
1811 w_sleep_cnt++;
1812 }
1813
1814 /* Insert new witness into the hash */
1815 witness_hash_put(w);
1816 witness_increment_graph_generation();
1817 mtx_unlock_spin(&w_mtx);
1818 return (w);
1819 found:
1820 w->w_refcount++;
1821 mtx_unlock_spin(&w_mtx);
1822 if (lock_class != w->w_class)
1823 kassert_panic(
1824 "lock (%s) %s does not match earlier (%s) lock",
1825 description, lock_class->lc_name,
1826 w->w_class->lc_name);
1827 return (w);
1828 }
1829
1830 static void
1831 depart(struct witness *w)
1832 {
1833 struct witness_list *list;
1834
1835 MPASS(w->w_refcount == 0);
1836 if (w->w_class->lc_flags & LC_SLEEPLOCK) {
1837 list = &w_sleep;
1838 w_sleep_cnt--;
1839 } else {
1840 list = &w_spin;
1841 w_spin_cnt--;
1842 }
1843 /*
1844 * Set file to NULL as it may point into a loadable module.
1845 */
1846 w->w_file = NULL;
1847 w->w_line = 0;
1848 witness_increment_graph_generation();
1849 }
1850
1851
1852 static void
1853 adopt(struct witness *parent, struct witness *child)
1854 {
1855 int pi, ci, i, j;
1856
1857 if (witness_cold == 0)
1858 mtx_assert(&w_mtx, MA_OWNED);
1859
1860 /* If the relationship is already known, there's no work to be done. */
1861 if (isitmychild(parent, child))
1862 return;
1863
1864 /* When the structure of the graph changes, bump up the generation. */
1865 witness_increment_graph_generation();
1866
1867 /*
1868 * The hard part ... create the direct relationship, then propagate all
1869 * indirect relationships.
1870 */
1871 pi = parent->w_index;
1872 ci = child->w_index;
1873 WITNESS_INDEX_ASSERT(pi);
1874 WITNESS_INDEX_ASSERT(ci);
1875 MPASS(pi != ci);
1876 w_rmatrix[pi][ci] |= WITNESS_PARENT;
1877 w_rmatrix[ci][pi] |= WITNESS_CHILD;
1878
1879 /*
1880 * If parent was not already an ancestor of child,
1881 * then we increment the descendant and ancestor counters.
1882 */
1883 if ((w_rmatrix[pi][ci] & WITNESS_ANCESTOR) == 0) {
1884 parent->w_num_descendants++;
1885 child->w_num_ancestors++;
1886 }
1887
1888 /*
1889 * Find each ancestor of 'pi'. Note that 'pi' itself is counted as
1890 * an ancestor of 'pi' during this loop.
1891 */
1892 for (i = 1; i <= w_max_used_index; i++) {
1893 if ((w_rmatrix[i][pi] & WITNESS_ANCESTOR_MASK) == 0 &&
1894 (i != pi))
1895 continue;
1896
1897 /* Find each descendant of 'i' and mark it as a descendant. */
1898 for (j = 1; j <= w_max_used_index; j++) {
1899
1900 /*
1901 * Skip children that are already marked as
1902 * descendants of 'i'.
1903 */
1904 if (w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK)
1905 continue;
1906
1907 /*
1908 * We are only interested in descendants of 'ci'. Note
1909 * that 'ci' itself is counted as a descendant of 'ci'.
1910 */
1911 if ((w_rmatrix[ci][j] & WITNESS_ANCESTOR_MASK) == 0 &&
1912 (j != ci))
1913 continue;
1914 w_rmatrix[i][j] |= WITNESS_ANCESTOR;
1915 w_rmatrix[j][i] |= WITNESS_DESCENDANT;
1916 w_data[i].w_num_descendants++;
1917 w_data[j].w_num_ancestors++;
1918
1919 /*
1920 * Make sure we aren't marking a node as both an
1921 * ancestor and descendant. We should have caught
1922 * this as a lock order reversal earlier.
1923 */
1924 if ((w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK) &&
1925 (w_rmatrix[i][j] & WITNESS_DESCENDANT_MASK)) {
1926 printf("witness rmatrix paradox! [%d][%d]=%d "
1927 "both ancestor and descendant\n",
1928 i, j, w_rmatrix[i][j]);
1929 kdb_backtrace();
1930 printf("Witness disabled.\n");
1931 witness_watch = -1;
1932 }
1933 if ((w_rmatrix[j][i] & WITNESS_ANCESTOR_MASK) &&
1934 (w_rmatrix[j][i] & WITNESS_DESCENDANT_MASK)) {
1935 printf("witness rmatrix paradox! [%d][%d]=%d "
1936 "both ancestor and descendant\n",
1937 j, i, w_rmatrix[j][i]);
1938 kdb_backtrace();
1939 printf("Witness disabled.\n");
1940 witness_watch = -1;
1941 }
1942 }
1943 }
1944 }
1945
1946 static void
1947 itismychild(struct witness *parent, struct witness *child)
1948 {
1949 int unlocked;
1950
1951 MPASS(child != NULL && parent != NULL);
1952 if (witness_cold == 0)
1953 mtx_assert(&w_mtx, MA_OWNED);
1954
1955 if (!witness_lock_type_equal(parent, child)) {
1956 if (witness_cold == 0) {
1957 unlocked = 1;
1958 mtx_unlock_spin(&w_mtx);
1959 } else {
1960 unlocked = 0;
1961 }
1962 kassert_panic(
1963 "%s: parent \"%s\" (%s) and child \"%s\" (%s) are not "
1964 "the same lock type", __func__, parent->w_name,
1965 parent->w_class->lc_name, child->w_name,
1966 child->w_class->lc_name);
1967 if (unlocked)
1968 mtx_lock_spin(&w_mtx);
1969 }
1970 adopt(parent, child);
1971 }
1972
1973 /*
1974 * Generic code for the isitmy*() functions. The rmask parameter is the
1975 * expected relationship of w1 to w2.
1976 */
1977 static int
1978 _isitmyx(struct witness *w1, struct witness *w2, int rmask, const char *fname)
1979 {
1980 unsigned char r1, r2;
1981 int i1, i2;
1982
1983 i1 = w1->w_index;
1984 i2 = w2->w_index;
1985 WITNESS_INDEX_ASSERT(i1);
1986 WITNESS_INDEX_ASSERT(i2);
1987 r1 = w_rmatrix[i1][i2] & WITNESS_RELATED_MASK;
1988 r2 = w_rmatrix[i2][i1] & WITNESS_RELATED_MASK;
1989
1990 /* The flags on one better be the inverse of the flags on the other */
1991 if (!((WITNESS_ATOD(r1) == r2 && WITNESS_DTOA(r2) == r1) ||
1992 (WITNESS_DTOA(r1) == r2 && WITNESS_ATOD(r2) == r1))) {
1993 printf("%s: rmatrix mismatch between %s (index %d) and %s "
1994 "(index %d): w_rmatrix[%d][%d] == %hhx but "
1995 "w_rmatrix[%d][%d] == %hhx\n",
1996 fname, w1->w_name, i1, w2->w_name, i2, i1, i2, r1,
1997 i2, i1, r2);
1998 kdb_backtrace();
1999 printf("Witness disabled.\n");
2000 witness_watch = -1;
2001 }
2002 return (r1 & rmask);
2003 }
2004
2005 /*
2006 * Checks if @child is a direct child of @parent.
2007 */
2008 static int
2009 isitmychild(struct witness *parent, struct witness *child)
2010 {
2011
2012 return (_isitmyx(parent, child, WITNESS_PARENT, __func__));
2013 }
2014
2015 /*
2016 * Checks if @descendant is a direct or inderect descendant of @ancestor.
2017 */
2018 static int
2019 isitmydescendant(struct witness *ancestor, struct witness *descendant)
2020 {
2021
2022 return (_isitmyx(ancestor, descendant, WITNESS_ANCESTOR_MASK,
2023 __func__));
2024 }
2025
2026 #ifdef BLESSING
2027 static int
2028 blessed(struct witness *w1, struct witness *w2)
2029 {
2030 int i;
2031 struct witness_blessed *b;
2032
2033 for (i = 0; i < blessed_count; i++) {
2034 b = &blessed_list[i];
2035 if (strcmp(w1->w_name, b->b_lock1) == 0) {
2036 if (strcmp(w2->w_name, b->b_lock2) == 0)
2037 return (1);
2038 continue;
2039 }
2040 if (strcmp(w1->w_name, b->b_lock2) == 0)
2041 if (strcmp(w2->w_name, b->b_lock1) == 0)
2042 return (1);
2043 }
2044 return (0);
2045 }
2046 #endif
2047
2048 static struct witness *
2049 witness_get(void)
2050 {
2051 struct witness *w;
2052 int index;
2053
2054 if (witness_cold == 0)
2055 mtx_assert(&w_mtx, MA_OWNED);
2056
2057 if (witness_watch == -1) {
2058 mtx_unlock_spin(&w_mtx);
2059 return (NULL);
2060 }
2061 if (STAILQ_EMPTY(&w_free)) {
2062 witness_watch = -1;
2063 mtx_unlock_spin(&w_mtx);
2064 printf("WITNESS: unable to allocate a new witness object\n");
2065 return (NULL);
2066 }
2067 w = STAILQ_FIRST(&w_free);
2068 STAILQ_REMOVE_HEAD(&w_free, w_list);
2069 w_free_cnt--;
2070 index = w->w_index;
2071 MPASS(index > 0 && index == w_max_used_index+1 &&
2072 index < WITNESS_COUNT);
2073 bzero(w, sizeof(*w));
2074 w->w_index = index;
2075 if (index > w_max_used_index)
2076 w_max_used_index = index;
2077 return (w);
2078 }
2079
2080 static void
2081 witness_free(struct witness *w)
2082 {
2083
2084 STAILQ_INSERT_HEAD(&w_free, w, w_list);
2085 w_free_cnt++;
2086 }
2087
2088 static struct lock_list_entry *
2089 witness_lock_list_get(void)
2090 {
2091 struct lock_list_entry *lle;
2092
2093 if (witness_watch == -1)
2094 return (NULL);
2095 mtx_lock_spin(&w_mtx);
2096 lle = w_lock_list_free;
2097 if (lle == NULL) {
2098 witness_watch = -1;
2099 mtx_unlock_spin(&w_mtx);
2100 printf("%s: witness exhausted\n", __func__);
2101 return (NULL);
2102 }
2103 w_lock_list_free = lle->ll_next;
2104 mtx_unlock_spin(&w_mtx);
2105 bzero(lle, sizeof(*lle));
2106 return (lle);
2107 }
2108
2109 static void
2110 witness_lock_list_free(struct lock_list_entry *lle)
2111 {
2112
2113 mtx_lock_spin(&w_mtx);
2114 lle->ll_next = w_lock_list_free;
2115 w_lock_list_free = lle;
2116 mtx_unlock_spin(&w_mtx);
2117 }
2118
2119 static struct lock_instance *
2120 find_instance(struct lock_list_entry *list, const struct lock_object *lock)
2121 {
2122 struct lock_list_entry *lle;
2123 struct lock_instance *instance;
2124 int i;
2125
2126 for (lle = list; lle != NULL; lle = lle->ll_next)
2127 for (i = lle->ll_count - 1; i >= 0; i--) {
2128 instance = &lle->ll_children[i];
2129 if (instance->li_lock == lock)
2130 return (instance);
2131 }
2132 return (NULL);
2133 }
2134
2135 static void
2136 witness_list_lock(struct lock_instance *instance,
2137 int (*prnt)(const char *fmt, ...))
2138 {
2139 struct lock_object *lock;
2140
2141 lock = instance->li_lock;
2142 prnt("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ?
2143 "exclusive" : "shared", LOCK_CLASS(lock)->lc_name, lock->lo_name);
2144 if (lock->lo_witness->w_name != lock->lo_name)
2145 prnt(" (%s)", lock->lo_witness->w_name);
2146 prnt(" r = %d (%p) locked @ %s:%d\n",
2147 instance->li_flags & LI_RECURSEMASK, lock,
2148 fixup_filename(instance->li_file), instance->li_line);
2149 }
2150
2151 #ifdef DDB
2152 static int
2153 witness_thread_has_locks(struct thread *td)
2154 {
2155
2156 if (td->td_sleeplocks == NULL)
2157 return (0);
2158 return (td->td_sleeplocks->ll_count != 0);
2159 }
2160
2161 static int
2162 witness_proc_has_locks(struct proc *p)
2163 {
2164 struct thread *td;
2165
2166 FOREACH_THREAD_IN_PROC(p, td) {
2167 if (witness_thread_has_locks(td))
2168 return (1);
2169 }
2170 return (0);
2171 }
2172 #endif
2173
2174 int
2175 witness_list_locks(struct lock_list_entry **lock_list,
2176 int (*prnt)(const char *fmt, ...))
2177 {
2178 struct lock_list_entry *lle;
2179 int i, nheld;
2180
2181 nheld = 0;
2182 for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
2183 for (i = lle->ll_count - 1; i >= 0; i--) {
2184 witness_list_lock(&lle->ll_children[i], prnt);
2185 nheld++;
2186 }
2187 return (nheld);
2188 }
2189
2190 /*
2191 * This is a bit risky at best. We call this function when we have timed
2192 * out acquiring a spin lock, and we assume that the other CPU is stuck
2193 * with this lock held. So, we go groveling around in the other CPU's
2194 * per-cpu data to try to find the lock instance for this spin lock to
2195 * see when it was last acquired.
2196 */
2197 void
2198 witness_display_spinlock(struct lock_object *lock, struct thread *owner,
2199 int (*prnt)(const char *fmt, ...))
2200 {
2201 struct lock_instance *instance;
2202 struct pcpu *pc;
2203
2204 if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU)
2205 return;
2206 pc = pcpu_find(owner->td_oncpu);
2207 instance = find_instance(pc->pc_spinlocks, lock);
2208 if (instance != NULL)
2209 witness_list_lock(instance, prnt);
2210 }
2211
2212 void
2213 witness_save(struct lock_object *lock, const char **filep, int *linep)
2214 {
2215 struct lock_list_entry *lock_list;
2216 struct lock_instance *instance;
2217 struct lock_class *class;
2218
2219 /*
2220 * This function is used independently in locking code to deal with
2221 * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2222 * is gone.
2223 */
2224 if (SCHEDULER_STOPPED())
2225 return;
2226 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2227 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2228 return;
2229 class = LOCK_CLASS(lock);
2230 if (class->lc_flags & LC_SLEEPLOCK)
2231 lock_list = curthread->td_sleeplocks;
2232 else {
2233 if (witness_skipspin)
2234 return;
2235 lock_list = PCPU_GET(spinlocks);
2236 }
2237 instance = find_instance(lock_list, lock);
2238 if (instance == NULL) {
2239 kassert_panic("%s: lock (%s) %s not locked", __func__,
2240 class->lc_name, lock->lo_name);
2241 return;
2242 }
2243 *filep = instance->li_file;
2244 *linep = instance->li_line;
2245 }
2246
2247 void
2248 witness_restore(struct lock_object *lock, const char *file, int line)
2249 {
2250 struct lock_list_entry *lock_list;
2251 struct lock_instance *instance;
2252 struct lock_class *class;
2253
2254 /*
2255 * This function is used independently in locking code to deal with
2256 * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2257 * is gone.
2258 */
2259 if (SCHEDULER_STOPPED())
2260 return;
2261 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2262 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2263 return;
2264 class = LOCK_CLASS(lock);
2265 if (class->lc_flags & LC_SLEEPLOCK)
2266 lock_list = curthread->td_sleeplocks;
2267 else {
2268 if (witness_skipspin)
2269 return;
2270 lock_list = PCPU_GET(spinlocks);
2271 }
2272 instance = find_instance(lock_list, lock);
2273 if (instance == NULL)
2274 kassert_panic("%s: lock (%s) %s not locked", __func__,
2275 class->lc_name, lock->lo_name);
2276 lock->lo_witness->w_file = file;
2277 lock->lo_witness->w_line = line;
2278 if (instance == NULL)
2279 return;
2280 instance->li_file = file;
2281 instance->li_line = line;
2282 }
2283
2284 void
2285 witness_assert(const struct lock_object *lock, int flags, const char *file,
2286 int line)
2287 {
2288 #ifdef INVARIANT_SUPPORT
2289 struct lock_instance *instance;
2290 struct lock_class *class;
2291
2292 if (lock->lo_witness == NULL || witness_watch < 1 || panicstr != NULL)
2293 return;
2294 class = LOCK_CLASS(lock);
2295 if ((class->lc_flags & LC_SLEEPLOCK) != 0)
2296 instance = find_instance(curthread->td_sleeplocks, lock);
2297 else if ((class->lc_flags & LC_SPINLOCK) != 0)
2298 instance = find_instance(PCPU_GET(spinlocks), lock);
2299 else {
2300 kassert_panic("Lock (%s) %s is not sleep or spin!",
2301 class->lc_name, lock->lo_name);
2302 return;
2303 }
2304 switch (flags) {
2305 case LA_UNLOCKED:
2306 if (instance != NULL)
2307 kassert_panic("Lock (%s) %s locked @ %s:%d.",
2308 class->lc_name, lock->lo_name,
2309 fixup_filename(file), line);
2310 break;
2311 case LA_LOCKED:
2312 case LA_LOCKED | LA_RECURSED:
2313 case LA_LOCKED | LA_NOTRECURSED:
2314 case LA_SLOCKED:
2315 case LA_SLOCKED | LA_RECURSED:
2316 case LA_SLOCKED | LA_NOTRECURSED:
2317 case LA_XLOCKED:
2318 case LA_XLOCKED | LA_RECURSED:
2319 case LA_XLOCKED | LA_NOTRECURSED:
2320 if (instance == NULL) {
2321 kassert_panic("Lock (%s) %s not locked @ %s:%d.",
2322 class->lc_name, lock->lo_name,
2323 fixup_filename(file), line);
2324 break;
2325 }
2326 if ((flags & LA_XLOCKED) != 0 &&
2327 (instance->li_flags & LI_EXCLUSIVE) == 0)
2328 kassert_panic(
2329 "Lock (%s) %s not exclusively locked @ %s:%d.",
2330 class->lc_name, lock->lo_name,
2331 fixup_filename(file), line);
2332 if ((flags & LA_SLOCKED) != 0 &&
2333 (instance->li_flags & LI_EXCLUSIVE) != 0)
2334 kassert_panic(
2335 "Lock (%s) %s exclusively locked @ %s:%d.",
2336 class->lc_name, lock->lo_name,
2337 fixup_filename(file), line);
2338 if ((flags & LA_RECURSED) != 0 &&
2339 (instance->li_flags & LI_RECURSEMASK) == 0)
2340 kassert_panic("Lock (%s) %s not recursed @ %s:%d.",
2341 class->lc_name, lock->lo_name,
2342 fixup_filename(file), line);
2343 if ((flags & LA_NOTRECURSED) != 0 &&
2344 (instance->li_flags & LI_RECURSEMASK) != 0)
2345 kassert_panic("Lock (%s) %s recursed @ %s:%d.",
2346 class->lc_name, lock->lo_name,
2347 fixup_filename(file), line);
2348 break;
2349 default:
2350 kassert_panic("Invalid lock assertion at %s:%d.",
2351 fixup_filename(file), line);
2352
2353 }
2354 #endif /* INVARIANT_SUPPORT */
2355 }
2356
2357 static void
2358 witness_setflag(struct lock_object *lock, int flag, int set)
2359 {
2360 struct lock_list_entry *lock_list;
2361 struct lock_instance *instance;
2362 struct lock_class *class;
2363
2364 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2365 return;
2366 class = LOCK_CLASS(lock);
2367 if (class->lc_flags & LC_SLEEPLOCK)
2368 lock_list = curthread->td_sleeplocks;
2369 else {
2370 if (witness_skipspin)
2371 return;
2372 lock_list = PCPU_GET(spinlocks);
2373 }
2374 instance = find_instance(lock_list, lock);
2375 if (instance == NULL) {
2376 kassert_panic("%s: lock (%s) %s not locked", __func__,
2377 class->lc_name, lock->lo_name);
2378 return;
2379 }
2380
2381 if (set)
2382 instance->li_flags |= flag;
2383 else
2384 instance->li_flags &= ~flag;
2385 }
2386
2387 void
2388 witness_norelease(struct lock_object *lock)
2389 {
2390
2391 witness_setflag(lock, LI_NORELEASE, 1);
2392 }
2393
2394 void
2395 witness_releaseok(struct lock_object *lock)
2396 {
2397
2398 witness_setflag(lock, LI_NORELEASE, 0);
2399 }
2400
2401 #ifdef DDB
2402 static void
2403 witness_ddb_list(struct thread *td)
2404 {
2405
2406 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2407 KASSERT(kdb_active, ("%s: not in the debugger", __func__));
2408
2409 if (witness_watch < 1)
2410 return;
2411
2412 witness_list_locks(&td->td_sleeplocks, db_printf);
2413
2414 /*
2415 * We only handle spinlocks if td == curthread. This is somewhat broken
2416 * if td is currently executing on some other CPU and holds spin locks
2417 * as we won't display those locks. If we had a MI way of getting
2418 * the per-cpu data for a given cpu then we could use
2419 * td->td_oncpu to get the list of spinlocks for this thread
2420 * and "fix" this.
2421 *
2422 * That still wouldn't really fix this unless we locked the scheduler
2423 * lock or stopped the other CPU to make sure it wasn't changing the
2424 * list out from under us. It is probably best to just not try to
2425 * handle threads on other CPU's for now.
2426 */
2427 if (td == curthread && PCPU_GET(spinlocks) != NULL)
2428 witness_list_locks(PCPU_PTR(spinlocks), db_printf);
2429 }
2430
2431 DB_SHOW_COMMAND(locks, db_witness_list)
2432 {
2433 struct thread *td;
2434
2435 if (have_addr)
2436 td = db_lookup_thread(addr, TRUE);
2437 else
2438 td = kdb_thread;
2439 witness_ddb_list(td);
2440 }
2441
2442 DB_SHOW_ALL_COMMAND(locks, db_witness_list_all)
2443 {
2444 struct thread *td;
2445 struct proc *p;
2446
2447 /*
2448 * It would be nice to list only threads and processes that actually
2449 * held sleep locks, but that information is currently not exported
2450 * by WITNESS.
2451 */
2452 FOREACH_PROC_IN_SYSTEM(p) {
2453 if (!witness_proc_has_locks(p))
2454 continue;
2455 FOREACH_THREAD_IN_PROC(p, td) {
2456 if (!witness_thread_has_locks(td))
2457 continue;
2458 db_printf("Process %d (%s) thread %p (%d)\n", p->p_pid,
2459 p->p_comm, td, td->td_tid);
2460 witness_ddb_list(td);
2461 if (db_pager_quit)
2462 return;
2463 }
2464 }
2465 }
2466 DB_SHOW_ALIAS(alllocks, db_witness_list_all)
2467
2468 DB_SHOW_COMMAND(witness, db_witness_display)
2469 {
2470
2471 witness_ddb_display(db_printf);
2472 }
2473 #endif
2474
2475 static int
2476 sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS)
2477 {
2478 struct witness_lock_order_data *data1, *data2, *tmp_data1, *tmp_data2;
2479 struct witness *tmp_w1, *tmp_w2, *w1, *w2;
2480 struct sbuf *sb;
2481 u_int w_rmatrix1, w_rmatrix2;
2482 int error, generation, i, j;
2483
2484 tmp_data1 = NULL;
2485 tmp_data2 = NULL;
2486 tmp_w1 = NULL;
2487 tmp_w2 = NULL;
2488 if (witness_watch < 1) {
2489 error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2490 return (error);
2491 }
2492 if (witness_cold) {
2493 error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2494 return (error);
2495 }
2496 error = 0;
2497 sb = sbuf_new(NULL, NULL, BADSTACK_SBUF_SIZE, SBUF_AUTOEXTEND);
2498 if (sb == NULL)
2499 return (ENOMEM);
2500
2501 /* Allocate and init temporary storage space. */
2502 tmp_w1 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2503 tmp_w2 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2504 tmp_data1 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2505 M_WAITOK | M_ZERO);
2506 tmp_data2 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2507 M_WAITOK | M_ZERO);
2508 stack_zero(&tmp_data1->wlod_stack);
2509 stack_zero(&tmp_data2->wlod_stack);
2510
2511 restart:
2512 mtx_lock_spin(&w_mtx);
2513 generation = w_generation;
2514 mtx_unlock_spin(&w_mtx);
2515 sbuf_printf(sb, "Number of known direct relationships is %d\n",
2516 w_lohash.wloh_count);
2517 for (i = 1; i < w_max_used_index; i++) {
2518 mtx_lock_spin(&w_mtx);
2519 if (generation != w_generation) {
2520 mtx_unlock_spin(&w_mtx);
2521
2522 /* The graph has changed, try again. */
2523 req->oldidx = 0;
2524 sbuf_clear(sb);
2525 goto restart;
2526 }
2527
2528 w1 = &w_data[i];
2529 if (w1->w_reversed == 0) {
2530 mtx_unlock_spin(&w_mtx);
2531 continue;
2532 }
2533
2534 /* Copy w1 locally so we can release the spin lock. */
2535 *tmp_w1 = *w1;
2536 mtx_unlock_spin(&w_mtx);
2537
2538 if (tmp_w1->w_reversed == 0)
2539 continue;
2540 for (j = 1; j < w_max_used_index; j++) {
2541 if ((w_rmatrix[i][j] & WITNESS_REVERSAL) == 0 || i > j)
2542 continue;
2543
2544 mtx_lock_spin(&w_mtx);
2545 if (generation != w_generation) {
2546 mtx_unlock_spin(&w_mtx);
2547
2548 /* The graph has changed, try again. */
2549 req->oldidx = 0;
2550 sbuf_clear(sb);
2551 goto restart;
2552 }
2553
2554 w2 = &w_data[j];
2555 data1 = witness_lock_order_get(w1, w2);
2556 data2 = witness_lock_order_get(w2, w1);
2557
2558 /*
2559 * Copy information locally so we can release the
2560 * spin lock.
2561 */
2562 *tmp_w2 = *w2;
2563 w_rmatrix1 = (unsigned int)w_rmatrix[i][j];
2564 w_rmatrix2 = (unsigned int)w_rmatrix[j][i];
2565
2566 if (data1) {
2567 stack_zero(&tmp_data1->wlod_stack);
2568 stack_copy(&data1->wlod_stack,
2569 &tmp_data1->wlod_stack);
2570 }
2571 if (data2 && data2 != data1) {
2572 stack_zero(&tmp_data2->wlod_stack);
2573 stack_copy(&data2->wlod_stack,
2574 &tmp_data2->wlod_stack);
2575 }
2576 mtx_unlock_spin(&w_mtx);
2577
2578 sbuf_printf(sb,
2579 "\nLock order reversal between \"%s\"(%s) and \"%s\"(%s)!\n",
2580 tmp_w1->w_name, tmp_w1->w_class->lc_name,
2581 tmp_w2->w_name, tmp_w2->w_class->lc_name);
2582 #if 0
2583 sbuf_printf(sb,
2584 "w_rmatrix[%s][%s] == %x, w_rmatrix[%s][%s] == %x\n",
2585 tmp_w1->name, tmp_w2->w_name, w_rmatrix1,
2586 tmp_w2->name, tmp_w1->w_name, w_rmatrix2);
2587 #endif
2588 if (data1) {
2589 sbuf_printf(sb,
2590 "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2591 tmp_w1->w_name, tmp_w1->w_class->lc_name,
2592 tmp_w2->w_name, tmp_w2->w_class->lc_name);
2593 stack_sbuf_print(sb, &tmp_data1->wlod_stack);
2594 sbuf_printf(sb, "\n");
2595 }
2596 if (data2 && data2 != data1) {
2597 sbuf_printf(sb,
2598 "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2599 tmp_w2->w_name, tmp_w2->w_class->lc_name,
2600 tmp_w1->w_name, tmp_w1->w_class->lc_name);
2601 stack_sbuf_print(sb, &tmp_data2->wlod_stack);
2602 sbuf_printf(sb, "\n");
2603 }
2604 }
2605 }
2606 mtx_lock_spin(&w_mtx);
2607 if (generation != w_generation) {
2608 mtx_unlock_spin(&w_mtx);
2609
2610 /*
2611 * The graph changed while we were printing stack data,
2612 * try again.
2613 */
2614 req->oldidx = 0;
2615 sbuf_clear(sb);
2616 goto restart;
2617 }
2618 mtx_unlock_spin(&w_mtx);
2619
2620 /* Free temporary storage space. */
2621 free(tmp_data1, M_TEMP);
2622 free(tmp_data2, M_TEMP);
2623 free(tmp_w1, M_TEMP);
2624 free(tmp_w2, M_TEMP);
2625
2626 sbuf_finish(sb);
2627 error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1);
2628 sbuf_delete(sb);
2629
2630 return (error);
2631 }
2632
2633 static int
2634 sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS)
2635 {
2636 struct witness *w;
2637 struct sbuf *sb;
2638 int error;
2639
2640 if (witness_watch < 1) {
2641 error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2642 return (error);
2643 }
2644 if (witness_cold) {
2645 error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2646 return (error);
2647 }
2648 error = 0;
2649
2650 error = sysctl_wire_old_buffer(req, 0);
2651 if (error != 0)
2652 return (error);
2653 sb = sbuf_new_for_sysctl(NULL, NULL, FULLGRAPH_SBUF_SIZE, req);
2654 if (sb == NULL)
2655 return (ENOMEM);
2656 sbuf_printf(sb, "\n");
2657
2658 mtx_lock_spin(&w_mtx);
2659 STAILQ_FOREACH(w, &w_all, w_list)
2660 w->w_displayed = 0;
2661 STAILQ_FOREACH(w, &w_all, w_list)
2662 witness_add_fullgraph(sb, w);
2663 mtx_unlock_spin(&w_mtx);
2664
2665 /*
2666 * Close the sbuf and return to userland.
2667 */
2668 error = sbuf_finish(sb);
2669 sbuf_delete(sb);
2670
2671 return (error);
2672 }
2673
2674 static int
2675 sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS)
2676 {
2677 int error, value;
2678
2679 value = witness_watch;
2680 error = sysctl_handle_int(oidp, &value, 0, req);
2681 if (error != 0 || req->newptr == NULL)
2682 return (error);
2683 if (value > 1 || value < -1 ||
2684 (witness_watch == -1 && value != witness_watch))
2685 return (EINVAL);
2686 witness_watch = value;
2687 return (0);
2688 }
2689
2690 static void
2691 witness_add_fullgraph(struct sbuf *sb, struct witness *w)
2692 {
2693 int i;
2694
2695 if (w->w_displayed != 0 || (w->w_file == NULL && w->w_line == 0))
2696 return;
2697 w->w_displayed = 1;
2698
2699 WITNESS_INDEX_ASSERT(w->w_index);
2700 for (i = 1; i <= w_max_used_index; i++) {
2701 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT) {
2702 sbuf_printf(sb, "\"%s\",\"%s\"\n", w->w_name,
2703 w_data[i].w_name);
2704 witness_add_fullgraph(sb, &w_data[i]);
2705 }
2706 }
2707 }
2708
2709 /*
2710 * A simple hash function. Takes a key pointer and a key size. If size == 0,
2711 * interprets the key as a string and reads until the null
2712 * terminator. Otherwise, reads the first size bytes. Returns an unsigned 32-bit
2713 * hash value computed from the key.
2714 */
2715 static uint32_t
2716 witness_hash_djb2(const uint8_t *key, uint32_t size)
2717 {
2718 unsigned int hash = 5381;
2719 int i;
2720
2721 /* hash = hash * 33 + key[i] */
2722 if (size)
2723 for (i = 0; i < size; i++)
2724 hash = ((hash << 5) + hash) + (unsigned int)key[i];
2725 else
2726 for (i = 0; key[i] != 0; i++)
2727 hash = ((hash << 5) + hash) + (unsigned int)key[i];
2728
2729 return (hash);
2730 }
2731
2732
2733 /*
2734 * Initializes the two witness hash tables. Called exactly once from
2735 * witness_initialize().
2736 */
2737 static void
2738 witness_init_hash_tables(void)
2739 {
2740 int i;
2741
2742 MPASS(witness_cold);
2743
2744 /* Initialize the hash tables. */
2745 for (i = 0; i < WITNESS_HASH_SIZE; i++)
2746 w_hash.wh_array[i] = NULL;
2747
2748 w_hash.wh_size = WITNESS_HASH_SIZE;
2749 w_hash.wh_count = 0;
2750
2751 /* Initialize the lock order data hash. */
2752 w_lofree = NULL;
2753 for (i = 0; i < WITNESS_LO_DATA_COUNT; i++) {
2754 memset(&w_lodata[i], 0, sizeof(w_lodata[i]));
2755 w_lodata[i].wlod_next = w_lofree;
2756 w_lofree = &w_lodata[i];
2757 }
2758 w_lohash.wloh_size = WITNESS_LO_HASH_SIZE;
2759 w_lohash.wloh_count = 0;
2760 for (i = 0; i < WITNESS_LO_HASH_SIZE; i++)
2761 w_lohash.wloh_array[i] = NULL;
2762 }
2763
2764 static struct witness *
2765 witness_hash_get(const char *key)
2766 {
2767 struct witness *w;
2768 uint32_t hash;
2769
2770 MPASS(key != NULL);
2771 if (witness_cold == 0)
2772 mtx_assert(&w_mtx, MA_OWNED);
2773 hash = witness_hash_djb2(key, 0) % w_hash.wh_size;
2774 w = w_hash.wh_array[hash];
2775 while (w != NULL) {
2776 if (strcmp(w->w_name, key) == 0)
2777 goto out;
2778 w = w->w_hash_next;
2779 }
2780
2781 out:
2782 return (w);
2783 }
2784
2785 static void
2786 witness_hash_put(struct witness *w)
2787 {
2788 uint32_t hash;
2789
2790 MPASS(w != NULL);
2791 MPASS(w->w_name != NULL);
2792 if (witness_cold == 0)
2793 mtx_assert(&w_mtx, MA_OWNED);
2794 KASSERT(witness_hash_get(w->w_name) == NULL,
2795 ("%s: trying to add a hash entry that already exists!", __func__));
2796 KASSERT(w->w_hash_next == NULL,
2797 ("%s: w->w_hash_next != NULL", __func__));
2798
2799 hash = witness_hash_djb2(w->w_name, 0) % w_hash.wh_size;
2800 w->w_hash_next = w_hash.wh_array[hash];
2801 w_hash.wh_array[hash] = w;
2802 w_hash.wh_count++;
2803 }
2804
2805
2806 static struct witness_lock_order_data *
2807 witness_lock_order_get(struct witness *parent, struct witness *child)
2808 {
2809 struct witness_lock_order_data *data = NULL;
2810 struct witness_lock_order_key key;
2811 unsigned int hash;
2812
2813 MPASS(parent != NULL && child != NULL);
2814 key.from = parent->w_index;
2815 key.to = child->w_index;
2816 WITNESS_INDEX_ASSERT(key.from);
2817 WITNESS_INDEX_ASSERT(key.to);
2818 if ((w_rmatrix[parent->w_index][child->w_index]
2819 & WITNESS_LOCK_ORDER_KNOWN) == 0)
2820 goto out;
2821
2822 hash = witness_hash_djb2((const char*)&key,
2823 sizeof(key)) % w_lohash.wloh_size;
2824 data = w_lohash.wloh_array[hash];
2825 while (data != NULL) {
2826 if (witness_lock_order_key_equal(&data->wlod_key, &key))
2827 break;
2828 data = data->wlod_next;
2829 }
2830
2831 out:
2832 return (data);
2833 }
2834
2835 /*
2836 * Verify that parent and child have a known relationship, are not the same,
2837 * and child is actually a child of parent. This is done without w_mtx
2838 * to avoid contention in the common case.
2839 */
2840 static int
2841 witness_lock_order_check(struct witness *parent, struct witness *child)
2842 {
2843
2844 if (parent != child &&
2845 w_rmatrix[parent->w_index][child->w_index]
2846 & WITNESS_LOCK_ORDER_KNOWN &&
2847 isitmychild(parent, child))
2848 return (1);
2849
2850 return (0);
2851 }
2852
2853 static int
2854 witness_lock_order_add(struct witness *parent, struct witness *child)
2855 {
2856 struct witness_lock_order_data *data = NULL;
2857 struct witness_lock_order_key key;
2858 unsigned int hash;
2859
2860 MPASS(parent != NULL && child != NULL);
2861 key.from = parent->w_index;
2862 key.to = child->w_index;
2863 WITNESS_INDEX_ASSERT(key.from);
2864 WITNESS_INDEX_ASSERT(key.to);
2865 if (w_rmatrix[parent->w_index][child->w_index]
2866 & WITNESS_LOCK_ORDER_KNOWN)
2867 return (1);
2868
2869 hash = witness_hash_djb2((const char*)&key,
2870 sizeof(key)) % w_lohash.wloh_size;
2871 w_rmatrix[parent->w_index][child->w_index] |= WITNESS_LOCK_ORDER_KNOWN;
2872 data = w_lofree;
2873 if (data == NULL)
2874 return (0);
2875 w_lofree = data->wlod_next;
2876 data->wlod_next = w_lohash.wloh_array[hash];
2877 data->wlod_key = key;
2878 w_lohash.wloh_array[hash] = data;
2879 w_lohash.wloh_count++;
2880 stack_zero(&data->wlod_stack);
2881 stack_save(&data->wlod_stack);
2882 return (1);
2883 }
2884
2885 /* Call this whenver the structure of the witness graph changes. */
2886 static void
2887 witness_increment_graph_generation(void)
2888 {
2889
2890 if (witness_cold == 0)
2891 mtx_assert(&w_mtx, MA_OWNED);
2892 w_generation++;
2893 }
2894
2895 #ifdef KDB
2896 static void
2897 _witness_debugger(int cond, const char *msg)
2898 {
2899
2900 if (witness_trace && cond)
2901 kdb_backtrace();
2902 if (witness_kdb && cond)
2903 kdb_enter(KDB_WHY_WITNESS, msg);
2904 }
2905 #endif
Cache object: e484171dce7d4b3ffd3a6b3382bc517a
|