1 /*-
2 * SPDX-License-Identifier: BSD-3-Clause
3 *
4 * Copyright (c) 2008 Isilon Systems, Inc.
5 * Copyright (c) 2008 Ilya Maykov <[email protected]>
6 * Copyright (c) 1998 Berkeley Software Design, Inc.
7 * All rights reserved.
8 *
9 * Redistribution and use in source and binary forms, with or without
10 * modification, are permitted provided that the following conditions
11 * are met:
12 * 1. Redistributions of source code must retain the above copyright
13 * notice, this list of conditions and the following disclaimer.
14 * 2. Redistributions in binary form must reproduce the above copyright
15 * notice, this list of conditions and the following disclaimer in the
16 * documentation and/or other materials provided with the distribution.
17 * 3. Berkeley Software Design Inc's name may not be used to endorse or
18 * promote products derived from this software without specific prior
19 * written permission.
20 *
21 * THIS SOFTWARE IS PROVIDED BY BERKELEY SOFTWARE DESIGN INC ``AS IS'' AND
22 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24 * ARE DISCLAIMED. IN NO EVENT SHALL BERKELEY SOFTWARE DESIGN INC BE LIABLE
25 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31 * SUCH DAMAGE.
32 *
33 * from BSDI $Id: mutex_witness.c,v 1.1.2.20 2000/04/27 03:10:27 cp Exp $
34 * and BSDI $Id: synch_machdep.c,v 2.3.2.39 2000/04/27 03:10:25 cp Exp $
35 */
36
37 /*
38 * Implementation of the `witness' lock verifier. Originally implemented for
39 * mutexes in BSD/OS. Extended to handle generic lock objects and lock
40 * classes in FreeBSD.
41 */
42
43 /*
44 * Main Entry: witness
45 * Pronunciation: 'wit-n&s
46 * Function: noun
47 * Etymology: Middle English witnesse, from Old English witnes knowledge,
48 * testimony, witness, from 2wit
49 * Date: before 12th century
50 * 1 : attestation of a fact or event : TESTIMONY
51 * 2 : one that gives evidence; specifically : one who testifies in
52 * a cause or before a judicial tribunal
53 * 3 : one asked to be present at a transaction so as to be able to
54 * testify to its having taken place
55 * 4 : one who has personal knowledge of something
56 * 5 a : something serving as evidence or proof : SIGN
57 * b : public affirmation by word or example of usually
58 * religious faith or conviction <the heroic witness to divine
59 * life -- Pilot>
60 * 6 capitalized : a member of the Jehovah's Witnesses
61 */
62
63 /*
64 * Special rules concerning Giant and lock orders:
65 *
66 * 1) Giant must be acquired before any other mutexes. Stated another way,
67 * no other mutex may be held when Giant is acquired.
68 *
69 * 2) Giant must be released when blocking on a sleepable lock.
70 *
71 * This rule is less obvious, but is a result of Giant providing the same
72 * semantics as spl(). Basically, when a thread sleeps, it must release
73 * Giant. When a thread blocks on a sleepable lock, it sleeps. Hence rule
74 * 2).
75 *
76 * 3) Giant may be acquired before or after sleepable locks.
77 *
78 * This rule is also not quite as obvious. Giant may be acquired after
79 * a sleepable lock because it is a non-sleepable lock and non-sleepable
80 * locks may always be acquired while holding a sleepable lock. The second
81 * case, Giant before a sleepable lock, follows from rule 2) above. Suppose
82 * you have two threads T1 and T2 and a sleepable lock X. Suppose that T1
83 * acquires X and blocks on Giant. Then suppose that T2 acquires Giant and
84 * blocks on X. When T2 blocks on X, T2 will release Giant allowing T1 to
85 * execute. Thus, acquiring Giant both before and after a sleepable lock
86 * will not result in a lock order reversal.
87 */
88
89 #include <sys/cdefs.h>
90 __FBSDID("$FreeBSD$");
91
92 #include "opt_ddb.h"
93 #include "opt_hwpmc_hooks.h"
94 #include "opt_stack.h"
95 #include "opt_witness.h"
96
97 #include <sys/param.h>
98 #include <sys/bus.h>
99 #include <sys/kdb.h>
100 #include <sys/kernel.h>
101 #include <sys/ktr.h>
102 #include <sys/lock.h>
103 #include <sys/malloc.h>
104 #include <sys/mutex.h>
105 #include <sys/priv.h>
106 #include <sys/proc.h>
107 #include <sys/sbuf.h>
108 #include <sys/sched.h>
109 #include <sys/stack.h>
110 #include <sys/sysctl.h>
111 #include <sys/syslog.h>
112 #include <sys/systm.h>
113
114 #ifdef DDB
115 #include <ddb/ddb.h>
116 #endif
117
118 #include <machine/stdarg.h>
119
120 #if !defined(DDB) && !defined(STACK)
121 #error "DDB or STACK options are required for WITNESS"
122 #endif
123
124 /* Note that these traces do not work with KTR_ALQ. */
125 #if 0
126 #define KTR_WITNESS KTR_SUBSYS
127 #else
128 #define KTR_WITNESS 0
129 #endif
130
131 #define LI_RECURSEMASK 0x0000ffff /* Recursion depth of lock instance. */
132 #define LI_EXCLUSIVE 0x00010000 /* Exclusive lock instance. */
133 #define LI_NORELEASE 0x00020000 /* Lock not allowed to be released. */
134
135 #ifndef WITNESS_COUNT
136 #define WITNESS_COUNT 1536
137 #endif
138 #define WITNESS_HASH_SIZE 251 /* Prime, gives load factor < 2 */
139 #define WITNESS_PENDLIST (512 + (MAXCPU * 4))
140
141 /* Allocate 256 KB of stack data space */
142 #define WITNESS_LO_DATA_COUNT 2048
143
144 /* Prime, gives load factor of ~2 at full load */
145 #define WITNESS_LO_HASH_SIZE 1021
146
147 /*
148 * XXX: This is somewhat bogus, as we assume here that at most 2048 threads
149 * will hold LOCK_NCHILDREN locks. We handle failure ok, and we should
150 * probably be safe for the most part, but it's still a SWAG.
151 */
152 #define LOCK_NCHILDREN 5
153 #define LOCK_CHILDCOUNT 2048
154
155 #define MAX_W_NAME 64
156
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 struct witness_blessed {
279 const char *b_lock1;
280 const char *b_lock2;
281 };
282
283 struct witness_pendhelp {
284 const char *wh_type;
285 struct lock_object *wh_lock;
286 };
287
288 struct witness_order_list_entry {
289 const char *w_name;
290 struct lock_class *w_class;
291 };
292
293 /*
294 * Returns 0 if one of the locks is a spin lock and the other is not.
295 * Returns 1 otherwise.
296 */
297 static __inline int
witness_lock_type_equal(struct witness * w1,struct witness * w2)298 witness_lock_type_equal(struct witness *w1, struct witness *w2)
299 {
300
301 return ((w1->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) ==
302 (w2->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)));
303 }
304
305 static __inline int
witness_lock_order_key_equal(const struct witness_lock_order_key * a,const struct witness_lock_order_key * b)306 witness_lock_order_key_equal(const struct witness_lock_order_key *a,
307 const struct witness_lock_order_key *b)
308 {
309
310 return (a->from == b->from && a->to == b->to);
311 }
312
313 static int _isitmyx(struct witness *w1, struct witness *w2, int rmask,
314 const char *fname);
315 static void adopt(struct witness *parent, struct witness *child);
316 static int blessed(struct witness *, struct witness *);
317 static void depart(struct witness *w);
318 static struct witness *enroll(const char *description,
319 struct lock_class *lock_class);
320 static struct lock_instance *find_instance(struct lock_list_entry *list,
321 const struct lock_object *lock);
322 static int isitmychild(struct witness *parent, struct witness *child);
323 static int isitmydescendant(struct witness *parent, struct witness *child);
324 static void itismychild(struct witness *parent, struct witness *child);
325 static int sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS);
326 static int sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS);
327 static int sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS);
328 static int sysctl_debug_witness_channel(SYSCTL_HANDLER_ARGS);
329 static void witness_add_fullgraph(struct sbuf *sb, struct witness *parent);
330 #ifdef DDB
331 static void witness_ddb_compute_levels(void);
332 static void witness_ddb_display(int(*)(const char *fmt, ...));
333 static void witness_ddb_display_descendants(int(*)(const char *fmt, ...),
334 struct witness *, int indent);
335 static void witness_ddb_display_list(int(*prnt)(const char *fmt, ...),
336 struct witness_list *list);
337 static void witness_ddb_level_descendants(struct witness *parent, int l);
338 static void witness_ddb_list(struct thread *td);
339 #endif
340 static void witness_debugger(int cond, const char *msg);
341 static void witness_free(struct witness *m);
342 static struct witness *witness_get(void);
343 static uint32_t witness_hash_djb2(const uint8_t *key, uint32_t size);
344 static struct witness *witness_hash_get(const char *key);
345 static void witness_hash_put(struct witness *w);
346 static void witness_init_hash_tables(void);
347 static void witness_increment_graph_generation(void);
348 static void witness_lock_list_free(struct lock_list_entry *lle);
349 static struct lock_list_entry *witness_lock_list_get(void);
350 static int witness_lock_order_add(struct witness *parent,
351 struct witness *child);
352 static int witness_lock_order_check(struct witness *parent,
353 struct witness *child);
354 static struct witness_lock_order_data *witness_lock_order_get(
355 struct witness *parent,
356 struct witness *child);
357 static void witness_list_lock(struct lock_instance *instance,
358 int (*prnt)(const char *fmt, ...));
359 static int witness_output(const char *fmt, ...) __printflike(1, 2);
360 static int witness_voutput(const char *fmt, va_list ap) __printflike(1, 0);
361 static void witness_setflag(struct lock_object *lock, int flag, int set);
362
363 static SYSCTL_NODE(_debug, OID_AUTO, witness, CTLFLAG_RW, NULL,
364 "Witness Locking");
365
366 /*
367 * If set to 0, lock order checking is disabled. If set to -1,
368 * witness is completely disabled. Otherwise witness performs full
369 * lock order checking for all locks. At runtime, lock order checking
370 * may be toggled. However, witness cannot be reenabled once it is
371 * completely disabled.
372 */
373 static int witness_watch = 1;
374 SYSCTL_PROC(_debug_witness, OID_AUTO, watch, CTLFLAG_RWTUN | CTLTYPE_INT, NULL, 0,
375 sysctl_debug_witness_watch, "I", "witness is watching lock operations");
376
377 #ifdef KDB
378 /*
379 * When KDB is enabled and witness_kdb is 1, it will cause the system
380 * to drop into kdebug() when:
381 * - a lock hierarchy violation occurs
382 * - locks are held when going to sleep.
383 */
384 #ifdef WITNESS_KDB
385 int witness_kdb = 1;
386 #else
387 int witness_kdb = 0;
388 #endif
389 SYSCTL_INT(_debug_witness, OID_AUTO, kdb, CTLFLAG_RWTUN, &witness_kdb, 0, "");
390 #endif /* KDB */
391
392 #if defined(DDB) || defined(KDB)
393 /*
394 * When DDB or KDB is enabled and witness_trace is 1, it will cause the system
395 * to print a stack trace:
396 * - a lock hierarchy violation occurs
397 * - locks are held when going to sleep.
398 */
399 int witness_trace = 1;
400 SYSCTL_INT(_debug_witness, OID_AUTO, trace, CTLFLAG_RWTUN, &witness_trace, 0, "");
401 #endif /* DDB || KDB */
402
403 #ifdef WITNESS_SKIPSPIN
404 int witness_skipspin = 1;
405 #else
406 int witness_skipspin = 0;
407 #endif
408 SYSCTL_INT(_debug_witness, OID_AUTO, skipspin, CTLFLAG_RDTUN, &witness_skipspin, 0, "");
409
410 int badstack_sbuf_size;
411
412 int witness_count = WITNESS_COUNT;
413 SYSCTL_INT(_debug_witness, OID_AUTO, witness_count, CTLFLAG_RDTUN,
414 &witness_count, 0, "");
415
416 /*
417 * Output channel for witness messages. By default we print to the console.
418 */
419 enum witness_channel {
420 WITNESS_CONSOLE,
421 WITNESS_LOG,
422 WITNESS_NONE,
423 };
424
425 static enum witness_channel witness_channel = WITNESS_CONSOLE;
426 SYSCTL_PROC(_debug_witness, OID_AUTO, output_channel, CTLTYPE_STRING |
427 CTLFLAG_RWTUN, NULL, 0, sysctl_debug_witness_channel, "A",
428 "Output channel for warnings");
429
430 /*
431 * Call this to print out the relations between locks.
432 */
433 SYSCTL_PROC(_debug_witness, OID_AUTO, fullgraph, CTLTYPE_STRING | CTLFLAG_RD,
434 NULL, 0, sysctl_debug_witness_fullgraph, "A", "Show locks relation graphs");
435
436 /*
437 * Call this to print out the witness faulty stacks.
438 */
439 SYSCTL_PROC(_debug_witness, OID_AUTO, badstacks, CTLTYPE_STRING | CTLFLAG_RD,
440 NULL, 0, sysctl_debug_witness_badstacks, "A", "Show bad witness stacks");
441
442 static struct mtx w_mtx;
443
444 /* w_list */
445 static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free);
446 static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all);
447
448 /* w_typelist */
449 static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin);
450 static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep);
451
452 /* lock list */
453 static struct lock_list_entry *w_lock_list_free = NULL;
454 static struct witness_pendhelp pending_locks[WITNESS_PENDLIST];
455 static u_int pending_cnt;
456
457 static int w_free_cnt, w_spin_cnt, w_sleep_cnt;
458 SYSCTL_INT(_debug_witness, OID_AUTO, free_cnt, CTLFLAG_RD, &w_free_cnt, 0, "");
459 SYSCTL_INT(_debug_witness, OID_AUTO, spin_cnt, CTLFLAG_RD, &w_spin_cnt, 0, "");
460 SYSCTL_INT(_debug_witness, OID_AUTO, sleep_cnt, CTLFLAG_RD, &w_sleep_cnt, 0,
461 "");
462
463 static struct witness *w_data;
464 static uint8_t **w_rmatrix;
465 static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT];
466 static struct witness_hash w_hash; /* The witness hash table. */
467
468 /* The lock order data hash */
469 static struct witness_lock_order_data w_lodata[WITNESS_LO_DATA_COUNT];
470 static struct witness_lock_order_data *w_lofree = NULL;
471 static struct witness_lock_order_hash w_lohash;
472 static int w_max_used_index = 0;
473 static unsigned int w_generation = 0;
474 static const char w_notrunning[] = "Witness not running\n";
475 static const char w_stillcold[] = "Witness is still cold\n";
476 #ifdef __i386__
477 static const char w_notallowed[] = "The sysctl is disabled on the arch\n";
478 #endif
479
480 static struct witness_order_list_entry order_lists[] = {
481 /*
482 * sx locks
483 */
484 { "proctree", &lock_class_sx },
485 { "allproc", &lock_class_sx },
486 { "allprison", &lock_class_sx },
487 { NULL, NULL },
488 /*
489 * Various mutexes
490 */
491 { "Giant", &lock_class_mtx_sleep },
492 { "pipe mutex", &lock_class_mtx_sleep },
493 { "sigio lock", &lock_class_mtx_sleep },
494 { "process group", &lock_class_mtx_sleep },
495 #ifdef HWPMC_HOOKS
496 { "pmc-sleep", &lock_class_mtx_sleep },
497 #endif
498 { "process lock", &lock_class_mtx_sleep },
499 { "session", &lock_class_mtx_sleep },
500 { "uidinfo hash", &lock_class_rw },
501 { "time lock", &lock_class_mtx_sleep },
502 { NULL, NULL },
503 /*
504 * umtx
505 */
506 { "umtx lock", &lock_class_mtx_sleep },
507 { NULL, NULL },
508 /*
509 * Sockets
510 */
511 { "accept", &lock_class_mtx_sleep },
512 { "so_snd", &lock_class_mtx_sleep },
513 { "so_rcv", &lock_class_mtx_sleep },
514 { "sellck", &lock_class_mtx_sleep },
515 { NULL, NULL },
516 /*
517 * Routing
518 */
519 { "so_rcv", &lock_class_mtx_sleep },
520 { "radix node head", &lock_class_rm },
521 { "rtentry", &lock_class_mtx_sleep },
522 { "ifaddr", &lock_class_mtx_sleep },
523 { NULL, NULL },
524 /*
525 * IPv4 multicast:
526 * protocol locks before interface locks, after UDP locks.
527 */
528 { "in_multi_sx", &lock_class_sx },
529 { "udpinp", &lock_class_rw },
530 { "in_multi_list_mtx", &lock_class_mtx_sleep },
531 { "igmp_mtx", &lock_class_mtx_sleep },
532 { "ifnet_rw", &lock_class_rw },
533 { "if_addr_lock", &lock_class_mtx_sleep },
534 { NULL, NULL },
535 /*
536 * IPv6 multicast:
537 * protocol locks before interface locks, after UDP locks.
538 */
539 { "in6_multi_sx", &lock_class_sx },
540 { "udpinp", &lock_class_rw },
541 { "in6_multi_list_mtx", &lock_class_mtx_sleep },
542 { "mld_mtx", &lock_class_mtx_sleep },
543 { "ifnet_rw", &lock_class_rw },
544 { "if_addr_lock", &lock_class_mtx_sleep },
545 { NULL, NULL },
546 /*
547 * UNIX Domain Sockets
548 */
549 { "unp_link_rwlock", &lock_class_rw },
550 { "unp_list_lock", &lock_class_mtx_sleep },
551 { "unp", &lock_class_mtx_sleep },
552 { "so_snd", &lock_class_mtx_sleep },
553 { NULL, NULL },
554 /*
555 * UDP/IP
556 */
557 { "udp", &lock_class_mtx_sleep },
558 { "udpinp", &lock_class_rw },
559 { "so_snd", &lock_class_mtx_sleep },
560 { NULL, NULL },
561 /*
562 * TCP/IP
563 */
564 { "tcp", &lock_class_mtx_sleep },
565 { "tcpinp", &lock_class_rw },
566 { "so_snd", &lock_class_mtx_sleep },
567 { NULL, NULL },
568 /*
569 * BPF
570 */
571 { "bpf global lock", &lock_class_sx },
572 { "bpf interface lock", &lock_class_rw },
573 { "bpf cdev lock", &lock_class_mtx_sleep },
574 { NULL, NULL },
575 /*
576 * NFS server
577 */
578 { "nfsd_mtx", &lock_class_mtx_sleep },
579 { "so_snd", &lock_class_mtx_sleep },
580 { NULL, NULL },
581
582 /*
583 * IEEE 802.11
584 */
585 { "802.11 com lock", &lock_class_mtx_sleep},
586 { NULL, NULL },
587 /*
588 * Network drivers
589 */
590 { "network driver", &lock_class_mtx_sleep},
591 { NULL, NULL },
592
593 /*
594 * Netgraph
595 */
596 { "ng_node", &lock_class_mtx_sleep },
597 { "ng_worklist", &lock_class_mtx_sleep },
598 { NULL, NULL },
599 /*
600 * CDEV
601 */
602 { "vm map (system)", &lock_class_mtx_sleep },
603 { "vnode interlock", &lock_class_mtx_sleep },
604 { "cdev", &lock_class_mtx_sleep },
605 { NULL, NULL },
606 /*
607 * VM
608 */
609 { "vm map (user)", &lock_class_sx },
610 { "vm object", &lock_class_rw },
611 { "vm page", &lock_class_mtx_sleep },
612 { "pmap pv global", &lock_class_rw },
613 { "pmap", &lock_class_mtx_sleep },
614 { "pmap pv list", &lock_class_rw },
615 { "vm page free queue", &lock_class_mtx_sleep },
616 { "vm pagequeue", &lock_class_mtx_sleep },
617 { NULL, NULL },
618 /*
619 * kqueue/VFS interaction
620 */
621 { "kqueue", &lock_class_mtx_sleep },
622 { "struct mount mtx", &lock_class_mtx_sleep },
623 { "vnode interlock", &lock_class_mtx_sleep },
624 { NULL, NULL },
625 /*
626 * VFS namecache
627 */
628 { "ncvn", &lock_class_mtx_sleep },
629 { "ncbuc", &lock_class_rw },
630 { "vnode interlock", &lock_class_mtx_sleep },
631 { "ncneg", &lock_class_mtx_sleep },
632 { NULL, NULL },
633 /*
634 * ZFS locking
635 */
636 { "dn->dn_mtx", &lock_class_sx },
637 { "dr->dt.di.dr_mtx", &lock_class_sx },
638 { "db->db_mtx", &lock_class_sx },
639 { NULL, NULL },
640 /*
641 * TCP log locks
642 */
643 { "TCP ID tree", &lock_class_rw },
644 { "tcp log id bucket", &lock_class_mtx_sleep },
645 { "tcpinp", &lock_class_rw },
646 { "TCP log expireq", &lock_class_mtx_sleep },
647 { NULL, NULL },
648 /*
649 * spin locks
650 */
651 #ifdef SMP
652 { "ap boot", &lock_class_mtx_spin },
653 #endif
654 { "rm.mutex_mtx", &lock_class_mtx_spin },
655 { "sio", &lock_class_mtx_spin },
656 #ifdef __i386__
657 { "cy", &lock_class_mtx_spin },
658 #endif
659 #ifdef __sparc64__
660 { "pcib_mtx", &lock_class_mtx_spin },
661 { "rtc_mtx", &lock_class_mtx_spin },
662 #endif
663 { "scc_hwmtx", &lock_class_mtx_spin },
664 { "uart_hwmtx", &lock_class_mtx_spin },
665 { "fast_taskqueue", &lock_class_mtx_spin },
666 { "intr table", &lock_class_mtx_spin },
667 { "process slock", &lock_class_mtx_spin },
668 { "syscons video lock", &lock_class_mtx_spin },
669 { "sleepq chain", &lock_class_mtx_spin },
670 { "rm_spinlock", &lock_class_mtx_spin },
671 { "turnstile chain", &lock_class_mtx_spin },
672 { "turnstile lock", &lock_class_mtx_spin },
673 { "sched lock", &lock_class_mtx_spin },
674 { "td_contested", &lock_class_mtx_spin },
675 { "callout", &lock_class_mtx_spin },
676 { "entropy harvest mutex", &lock_class_mtx_spin },
677 #ifdef SMP
678 { "smp rendezvous", &lock_class_mtx_spin },
679 #endif
680 #ifdef __powerpc__
681 { "tlb0", &lock_class_mtx_spin },
682 #endif
683 { NULL, NULL },
684 { "sched lock", &lock_class_mtx_spin },
685 #ifdef HWPMC_HOOKS
686 { "pmc-per-proc", &lock_class_mtx_spin },
687 #endif
688 { NULL, NULL },
689 /*
690 * leaf locks
691 */
692 { "intrcnt", &lock_class_mtx_spin },
693 { "icu", &lock_class_mtx_spin },
694 #if defined(SMP) && defined(__sparc64__)
695 { "ipi", &lock_class_mtx_spin },
696 #endif
697 #ifdef __i386__
698 { "allpmaps", &lock_class_mtx_spin },
699 { "descriptor tables", &lock_class_mtx_spin },
700 #endif
701 { "clk", &lock_class_mtx_spin },
702 { "cpuset", &lock_class_mtx_spin },
703 { "mprof lock", &lock_class_mtx_spin },
704 { "zombie lock", &lock_class_mtx_spin },
705 { "ALD Queue", &lock_class_mtx_spin },
706 #if defined(__i386__) || defined(__amd64__)
707 { "pcicfg", &lock_class_mtx_spin },
708 { "NDIS thread lock", &lock_class_mtx_spin },
709 #endif
710 { "tw_osl_io_lock", &lock_class_mtx_spin },
711 { "tw_osl_q_lock", &lock_class_mtx_spin },
712 { "tw_cl_io_lock", &lock_class_mtx_spin },
713 { "tw_cl_intr_lock", &lock_class_mtx_spin },
714 { "tw_cl_gen_lock", &lock_class_mtx_spin },
715 #ifdef HWPMC_HOOKS
716 { "pmc-leaf", &lock_class_mtx_spin },
717 #endif
718 { "blocked lock", &lock_class_mtx_spin },
719 { NULL, NULL },
720 { NULL, NULL }
721 };
722
723 /*
724 * Pairs of locks which have been blessed. Witness does not complain about
725 * order problems with blessed lock pairs. Please do not add an entry to the
726 * table without an explanatory comment.
727 */
728 static struct witness_blessed blessed_list[] = {
729 /*
730 * See the comment in ufs_dirhash.c. Basically, a vnode lock serializes
731 * both lock orders, so a deadlock cannot happen as a result of this
732 * LOR.
733 */
734 { "dirhash", "bufwait" },
735
736 /*
737 * A UFS vnode may be locked in vget() while a buffer belonging to the
738 * parent directory vnode is locked.
739 */
740 { "ufs", "bufwait" },
741 };
742
743 /*
744 * This global is set to 0 once it becomes safe to use the witness code.
745 */
746 static int witness_cold = 1;
747
748 /*
749 * This global is set to 1 once the static lock orders have been enrolled
750 * so that a warning can be issued for any spin locks enrolled later.
751 */
752 static int witness_spin_warn = 0;
753
754 /* Trim useless garbage from filenames. */
755 static const char *
fixup_filename(const char * file)756 fixup_filename(const char *file)
757 {
758
759 if (file == NULL)
760 return (NULL);
761 while (strncmp(file, "../", 3) == 0)
762 file += 3;
763 return (file);
764 }
765
766 /*
767 * Calculate the size of early witness structures.
768 */
769 int
witness_startup_count(void)770 witness_startup_count(void)
771 {
772 int sz;
773
774 sz = sizeof(struct witness) * witness_count;
775 sz += sizeof(*w_rmatrix) * (witness_count + 1);
776 sz += sizeof(*w_rmatrix[0]) * (witness_count + 1) *
777 (witness_count + 1);
778
779 return (sz);
780 }
781
782 /*
783 * The WITNESS-enabled diagnostic code. Note that the witness code does
784 * assume that the early boot is single-threaded at least until after this
785 * routine is completed.
786 */
787 void
witness_startup(void * mem)788 witness_startup(void *mem)
789 {
790 struct lock_object *lock;
791 struct witness_order_list_entry *order;
792 struct witness *w, *w1;
793 uintptr_t p;
794 int i;
795
796 p = (uintptr_t)mem;
797 w_data = (void *)p;
798 p += sizeof(struct witness) * witness_count;
799
800 w_rmatrix = (void *)p;
801 p += sizeof(*w_rmatrix) * (witness_count + 1);
802
803 for (i = 0; i < witness_count + 1; i++) {
804 w_rmatrix[i] = (void *)p;
805 p += sizeof(*w_rmatrix[i]) * (witness_count + 1);
806 }
807 badstack_sbuf_size = witness_count * 256;
808
809 /*
810 * We have to release Giant before initializing its witness
811 * structure so that WITNESS doesn't get confused.
812 */
813 mtx_unlock(&Giant);
814 mtx_assert(&Giant, MA_NOTOWNED);
815
816 CTR1(KTR_WITNESS, "%s: initializing witness", __func__);
817 mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET |
818 MTX_NOWITNESS | MTX_NOPROFILE);
819 for (i = witness_count - 1; i >= 0; i--) {
820 w = &w_data[i];
821 memset(w, 0, sizeof(*w));
822 w_data[i].w_index = i; /* Witness index never changes. */
823 witness_free(w);
824 }
825 KASSERT(STAILQ_FIRST(&w_free)->w_index == 0,
826 ("%s: Invalid list of free witness objects", __func__));
827
828 /* Witness with index 0 is not used to aid in debugging. */
829 STAILQ_REMOVE_HEAD(&w_free, w_list);
830 w_free_cnt--;
831
832 for (i = 0; i < witness_count; i++) {
833 memset(w_rmatrix[i], 0, sizeof(*w_rmatrix[i]) *
834 (witness_count + 1));
835 }
836
837 for (i = 0; i < LOCK_CHILDCOUNT; i++)
838 witness_lock_list_free(&w_locklistdata[i]);
839 witness_init_hash_tables();
840
841 /* First add in all the specified order lists. */
842 for (order = order_lists; order->w_name != NULL; order++) {
843 w = enroll(order->w_name, order->w_class);
844 if (w == NULL)
845 continue;
846 w->w_file = "order list";
847 for (order++; order->w_name != NULL; order++) {
848 w1 = enroll(order->w_name, order->w_class);
849 if (w1 == NULL)
850 continue;
851 w1->w_file = "order list";
852 itismychild(w, w1);
853 w = w1;
854 }
855 }
856 witness_spin_warn = 1;
857
858 /* Iterate through all locks and add them to witness. */
859 for (i = 0; pending_locks[i].wh_lock != NULL; i++) {
860 lock = pending_locks[i].wh_lock;
861 KASSERT(lock->lo_flags & LO_WITNESS,
862 ("%s: lock %s is on pending list but not LO_WITNESS",
863 __func__, lock->lo_name));
864 lock->lo_witness = enroll(pending_locks[i].wh_type,
865 LOCK_CLASS(lock));
866 }
867
868 /* Mark the witness code as being ready for use. */
869 witness_cold = 0;
870
871 mtx_lock(&Giant);
872 }
873
874 void
witness_init(struct lock_object * lock,const char * type)875 witness_init(struct lock_object *lock, const char *type)
876 {
877 struct lock_class *class;
878
879 /* Various sanity checks. */
880 class = LOCK_CLASS(lock);
881 if ((lock->lo_flags & LO_RECURSABLE) != 0 &&
882 (class->lc_flags & LC_RECURSABLE) == 0)
883 kassert_panic("%s: lock (%s) %s can not be recursable",
884 __func__, class->lc_name, lock->lo_name);
885 if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
886 (class->lc_flags & LC_SLEEPABLE) == 0)
887 kassert_panic("%s: lock (%s) %s can not be sleepable",
888 __func__, class->lc_name, lock->lo_name);
889 if ((lock->lo_flags & LO_UPGRADABLE) != 0 &&
890 (class->lc_flags & LC_UPGRADABLE) == 0)
891 kassert_panic("%s: lock (%s) %s can not be upgradable",
892 __func__, class->lc_name, lock->lo_name);
893
894 /*
895 * If we shouldn't watch this lock, then just clear lo_witness.
896 * Otherwise, if witness_cold is set, then it is too early to
897 * enroll this lock, so defer it to witness_initialize() by adding
898 * it to the pending_locks list. If it is not too early, then enroll
899 * the lock now.
900 */
901 if (witness_watch < 1 || panicstr != NULL ||
902 (lock->lo_flags & LO_WITNESS) == 0)
903 lock->lo_witness = NULL;
904 else if (witness_cold) {
905 pending_locks[pending_cnt].wh_lock = lock;
906 pending_locks[pending_cnt++].wh_type = type;
907 if (pending_cnt > WITNESS_PENDLIST)
908 panic("%s: pending locks list is too small, "
909 "increase WITNESS_PENDLIST\n",
910 __func__);
911 } else
912 lock->lo_witness = enroll(type, class);
913 }
914
915 void
witness_destroy(struct lock_object * lock)916 witness_destroy(struct lock_object *lock)
917 {
918 struct lock_class *class;
919 struct witness *w;
920
921 class = LOCK_CLASS(lock);
922
923 if (witness_cold)
924 panic("lock (%s) %s destroyed while witness_cold",
925 class->lc_name, lock->lo_name);
926
927 /* XXX: need to verify that no one holds the lock */
928 if ((lock->lo_flags & LO_WITNESS) == 0 || lock->lo_witness == NULL)
929 return;
930 w = lock->lo_witness;
931
932 mtx_lock_spin(&w_mtx);
933 MPASS(w->w_refcount > 0);
934 w->w_refcount--;
935
936 if (w->w_refcount == 0)
937 depart(w);
938 mtx_unlock_spin(&w_mtx);
939 }
940
941 #ifdef DDB
942 static void
witness_ddb_compute_levels(void)943 witness_ddb_compute_levels(void)
944 {
945 struct witness *w;
946
947 /*
948 * First clear all levels.
949 */
950 STAILQ_FOREACH(w, &w_all, w_list)
951 w->w_ddb_level = -1;
952
953 /*
954 * Look for locks with no parents and level all their descendants.
955 */
956 STAILQ_FOREACH(w, &w_all, w_list) {
957
958 /* If the witness has ancestors (is not a root), skip it. */
959 if (w->w_num_ancestors > 0)
960 continue;
961 witness_ddb_level_descendants(w, 0);
962 }
963 }
964
965 static void
witness_ddb_level_descendants(struct witness * w,int l)966 witness_ddb_level_descendants(struct witness *w, int l)
967 {
968 int i;
969
970 if (w->w_ddb_level >= l)
971 return;
972
973 w->w_ddb_level = l;
974 l++;
975
976 for (i = 1; i <= w_max_used_index; i++) {
977 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
978 witness_ddb_level_descendants(&w_data[i], l);
979 }
980 }
981
982 static void
witness_ddb_display_descendants(int (* prnt)(const char * fmt,...),struct witness * w,int indent)983 witness_ddb_display_descendants(int(*prnt)(const char *fmt, ...),
984 struct witness *w, int indent)
985 {
986 int i;
987
988 for (i = 0; i < indent; i++)
989 prnt(" ");
990 prnt("%s (type: %s, depth: %d, active refs: %d)",
991 w->w_name, w->w_class->lc_name,
992 w->w_ddb_level, w->w_refcount);
993 if (w->w_displayed) {
994 prnt(" -- (already displayed)\n");
995 return;
996 }
997 w->w_displayed = 1;
998 if (w->w_file != NULL && w->w_line != 0)
999 prnt(" -- last acquired @ %s:%d\n", fixup_filename(w->w_file),
1000 w->w_line);
1001 else
1002 prnt(" -- never acquired\n");
1003 indent++;
1004 WITNESS_INDEX_ASSERT(w->w_index);
1005 for (i = 1; i <= w_max_used_index; i++) {
1006 if (db_pager_quit)
1007 return;
1008 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT)
1009 witness_ddb_display_descendants(prnt, &w_data[i],
1010 indent);
1011 }
1012 }
1013
1014 static void
witness_ddb_display_list(int (* prnt)(const char * fmt,...),struct witness_list * list)1015 witness_ddb_display_list(int(*prnt)(const char *fmt, ...),
1016 struct witness_list *list)
1017 {
1018 struct witness *w;
1019
1020 STAILQ_FOREACH(w, list, w_typelist) {
1021 if (w->w_file == NULL || w->w_ddb_level > 0)
1022 continue;
1023
1024 /* This lock has no anscestors - display its descendants. */
1025 witness_ddb_display_descendants(prnt, w, 0);
1026 if (db_pager_quit)
1027 return;
1028 }
1029 }
1030
1031 static void
witness_ddb_display(int (* prnt)(const char * fmt,...))1032 witness_ddb_display(int(*prnt)(const char *fmt, ...))
1033 {
1034 struct witness *w;
1035
1036 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1037 witness_ddb_compute_levels();
1038
1039 /* Clear all the displayed flags. */
1040 STAILQ_FOREACH(w, &w_all, w_list)
1041 w->w_displayed = 0;
1042
1043 /*
1044 * First, handle sleep locks which have been acquired at least
1045 * once.
1046 */
1047 prnt("Sleep locks:\n");
1048 witness_ddb_display_list(prnt, &w_sleep);
1049 if (db_pager_quit)
1050 return;
1051
1052 /*
1053 * Now do spin locks which have been acquired at least once.
1054 */
1055 prnt("\nSpin locks:\n");
1056 witness_ddb_display_list(prnt, &w_spin);
1057 if (db_pager_quit)
1058 return;
1059
1060 /*
1061 * Finally, any locks which have not been acquired yet.
1062 */
1063 prnt("\nLocks which were never acquired:\n");
1064 STAILQ_FOREACH(w, &w_all, w_list) {
1065 if (w->w_file != NULL || w->w_refcount == 0)
1066 continue;
1067 prnt("%s (type: %s, depth: %d)\n", w->w_name,
1068 w->w_class->lc_name, w->w_ddb_level);
1069 if (db_pager_quit)
1070 return;
1071 }
1072 }
1073 #endif /* DDB */
1074
1075 int
witness_defineorder(struct lock_object * lock1,struct lock_object * lock2)1076 witness_defineorder(struct lock_object *lock1, struct lock_object *lock2)
1077 {
1078
1079 if (witness_watch == -1 || panicstr != NULL)
1080 return (0);
1081
1082 /* Require locks that witness knows about. */
1083 if (lock1 == NULL || lock1->lo_witness == NULL || lock2 == NULL ||
1084 lock2->lo_witness == NULL)
1085 return (EINVAL);
1086
1087 mtx_assert(&w_mtx, MA_NOTOWNED);
1088 mtx_lock_spin(&w_mtx);
1089
1090 /*
1091 * If we already have either an explicit or implied lock order that
1092 * is the other way around, then return an error.
1093 */
1094 if (witness_watch &&
1095 isitmydescendant(lock2->lo_witness, lock1->lo_witness)) {
1096 mtx_unlock_spin(&w_mtx);
1097 return (EDOOFUS);
1098 }
1099
1100 /* Try to add the new order. */
1101 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1102 lock2->lo_witness->w_name, lock1->lo_witness->w_name);
1103 itismychild(lock1->lo_witness, lock2->lo_witness);
1104 mtx_unlock_spin(&w_mtx);
1105 return (0);
1106 }
1107
1108 void
witness_checkorder(struct lock_object * lock,int flags,const char * file,int line,struct lock_object * interlock)1109 witness_checkorder(struct lock_object *lock, int flags, const char *file,
1110 int line, struct lock_object *interlock)
1111 {
1112 struct lock_list_entry *lock_list, *lle;
1113 struct lock_instance *lock1, *lock2, *plock;
1114 struct lock_class *class, *iclass;
1115 struct witness *w, *w1;
1116 struct thread *td;
1117 int i, j;
1118
1119 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL ||
1120 panicstr != NULL)
1121 return;
1122
1123 w = lock->lo_witness;
1124 class = LOCK_CLASS(lock);
1125 td = curthread;
1126
1127 if (class->lc_flags & LC_SLEEPLOCK) {
1128
1129 /*
1130 * Since spin locks include a critical section, this check
1131 * implicitly enforces a lock order of all sleep locks before
1132 * all spin locks.
1133 */
1134 if (td->td_critnest != 0 && !kdb_active)
1135 kassert_panic("acquiring blockable sleep lock with "
1136 "spinlock or critical section held (%s) %s @ %s:%d",
1137 class->lc_name, lock->lo_name,
1138 fixup_filename(file), line);
1139
1140 /*
1141 * If this is the first lock acquired then just return as
1142 * no order checking is needed.
1143 */
1144 lock_list = td->td_sleeplocks;
1145 if (lock_list == NULL || lock_list->ll_count == 0)
1146 return;
1147 } else {
1148
1149 /*
1150 * If this is the first lock, just return as no order
1151 * checking is needed. Avoid problems with thread
1152 * migration pinning the thread while checking if
1153 * spinlocks are held. If at least one spinlock is held
1154 * the thread is in a safe path and it is allowed to
1155 * unpin it.
1156 */
1157 sched_pin();
1158 lock_list = PCPU_GET(spinlocks);
1159 if (lock_list == NULL || lock_list->ll_count == 0) {
1160 sched_unpin();
1161 return;
1162 }
1163 sched_unpin();
1164 }
1165
1166 /*
1167 * Check to see if we are recursing on a lock we already own. If
1168 * so, make sure that we don't mismatch exclusive and shared lock
1169 * acquires.
1170 */
1171 lock1 = find_instance(lock_list, lock);
1172 if (lock1 != NULL) {
1173 if ((lock1->li_flags & LI_EXCLUSIVE) != 0 &&
1174 (flags & LOP_EXCLUSIVE) == 0) {
1175 witness_output("shared lock of (%s) %s @ %s:%d\n",
1176 class->lc_name, lock->lo_name,
1177 fixup_filename(file), line);
1178 witness_output("while exclusively locked from %s:%d\n",
1179 fixup_filename(lock1->li_file), lock1->li_line);
1180 kassert_panic("excl->share");
1181 }
1182 if ((lock1->li_flags & LI_EXCLUSIVE) == 0 &&
1183 (flags & LOP_EXCLUSIVE) != 0) {
1184 witness_output("exclusive lock of (%s) %s @ %s:%d\n",
1185 class->lc_name, lock->lo_name,
1186 fixup_filename(file), line);
1187 witness_output("while share locked from %s:%d\n",
1188 fixup_filename(lock1->li_file), lock1->li_line);
1189 kassert_panic("share->excl");
1190 }
1191 return;
1192 }
1193
1194 /* Warn if the interlock is not locked exactly once. */
1195 if (interlock != NULL) {
1196 iclass = LOCK_CLASS(interlock);
1197 lock1 = find_instance(lock_list, interlock);
1198 if (lock1 == NULL)
1199 kassert_panic("interlock (%s) %s not locked @ %s:%d",
1200 iclass->lc_name, interlock->lo_name,
1201 fixup_filename(file), line);
1202 else if ((lock1->li_flags & LI_RECURSEMASK) != 0)
1203 kassert_panic("interlock (%s) %s recursed @ %s:%d",
1204 iclass->lc_name, interlock->lo_name,
1205 fixup_filename(file), line);
1206 }
1207
1208 /*
1209 * Find the previously acquired lock, but ignore interlocks.
1210 */
1211 plock = &lock_list->ll_children[lock_list->ll_count - 1];
1212 if (interlock != NULL && plock->li_lock == interlock) {
1213 if (lock_list->ll_count > 1)
1214 plock =
1215 &lock_list->ll_children[lock_list->ll_count - 2];
1216 else {
1217 lle = lock_list->ll_next;
1218
1219 /*
1220 * The interlock is the only lock we hold, so
1221 * simply return.
1222 */
1223 if (lle == NULL)
1224 return;
1225 plock = &lle->ll_children[lle->ll_count - 1];
1226 }
1227 }
1228
1229 /*
1230 * Try to perform most checks without a lock. If this succeeds we
1231 * can skip acquiring the lock and return success. Otherwise we redo
1232 * the check with the lock held to handle races with concurrent updates.
1233 */
1234 w1 = plock->li_lock->lo_witness;
1235 if (witness_lock_order_check(w1, w))
1236 return;
1237
1238 mtx_lock_spin(&w_mtx);
1239 if (witness_lock_order_check(w1, w)) {
1240 mtx_unlock_spin(&w_mtx);
1241 return;
1242 }
1243 witness_lock_order_add(w1, w);
1244
1245 /*
1246 * Check for duplicate locks of the same type. Note that we only
1247 * have to check for this on the last lock we just acquired. Any
1248 * other cases will be caught as lock order violations.
1249 */
1250 if (w1 == w) {
1251 i = w->w_index;
1252 if (!(lock->lo_flags & LO_DUPOK) && !(flags & LOP_DUPOK) &&
1253 !(w_rmatrix[i][i] & WITNESS_REVERSAL)) {
1254 w_rmatrix[i][i] |= WITNESS_REVERSAL;
1255 w->w_reversed = 1;
1256 mtx_unlock_spin(&w_mtx);
1257 witness_output(
1258 "acquiring duplicate lock of same type: \"%s\"\n",
1259 w->w_name);
1260 witness_output(" 1st %s @ %s:%d\n", plock->li_lock->lo_name,
1261 fixup_filename(plock->li_file), plock->li_line);
1262 witness_output(" 2nd %s @ %s:%d\n", lock->lo_name,
1263 fixup_filename(file), line);
1264 witness_debugger(1, __func__);
1265 } else
1266 mtx_unlock_spin(&w_mtx);
1267 return;
1268 }
1269 mtx_assert(&w_mtx, MA_OWNED);
1270
1271 /*
1272 * If we know that the lock we are acquiring comes after
1273 * the lock we most recently acquired in the lock order tree,
1274 * then there is no need for any further checks.
1275 */
1276 if (isitmychild(w1, w))
1277 goto out;
1278
1279 for (j = 0, lle = lock_list; lle != NULL; lle = lle->ll_next) {
1280 for (i = lle->ll_count - 1; i >= 0; i--, j++) {
1281
1282 MPASS(j < LOCK_CHILDCOUNT * LOCK_NCHILDREN);
1283 lock1 = &lle->ll_children[i];
1284
1285 /*
1286 * Ignore the interlock.
1287 */
1288 if (interlock == lock1->li_lock)
1289 continue;
1290
1291 /*
1292 * If this lock doesn't undergo witness checking,
1293 * then skip it.
1294 */
1295 w1 = lock1->li_lock->lo_witness;
1296 if (w1 == NULL) {
1297 KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0,
1298 ("lock missing witness structure"));
1299 continue;
1300 }
1301
1302 /*
1303 * If we are locking Giant and this is a sleepable
1304 * lock, then skip it.
1305 */
1306 if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0 &&
1307 lock == &Giant.lock_object)
1308 continue;
1309
1310 /*
1311 * If we are locking a sleepable lock and this lock
1312 * is Giant, then skip it.
1313 */
1314 if ((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1315 lock1->li_lock == &Giant.lock_object)
1316 continue;
1317
1318 /*
1319 * If we are locking a sleepable lock and this lock
1320 * isn't sleepable, we want to treat it as a lock
1321 * order violation to enfore a general lock order of
1322 * sleepable locks before non-sleepable locks.
1323 */
1324 if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1325 (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
1326 goto reversal;
1327
1328 /*
1329 * If we are locking Giant and this is a non-sleepable
1330 * lock, then treat it as a reversal.
1331 */
1332 if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0 &&
1333 lock == &Giant.lock_object)
1334 goto reversal;
1335
1336 /*
1337 * Check the lock order hierarchy for a reveresal.
1338 */
1339 if (!isitmydescendant(w, w1))
1340 continue;
1341 reversal:
1342
1343 /*
1344 * We have a lock order violation, check to see if it
1345 * is allowed or has already been yelled about.
1346 */
1347
1348 /* Bail if this violation is known */
1349 if (w_rmatrix[w1->w_index][w->w_index] & WITNESS_REVERSAL)
1350 goto out;
1351
1352 /* Record this as a violation */
1353 w_rmatrix[w1->w_index][w->w_index] |= WITNESS_REVERSAL;
1354 w_rmatrix[w->w_index][w1->w_index] |= WITNESS_REVERSAL;
1355 w->w_reversed = w1->w_reversed = 1;
1356 witness_increment_graph_generation();
1357
1358 /*
1359 * If the lock order is blessed, bail before logging
1360 * anything. We don't look for other lock order
1361 * violations though, which may be a bug.
1362 */
1363 if (blessed(w, w1))
1364 goto out;
1365 mtx_unlock_spin(&w_mtx);
1366
1367 #ifdef WITNESS_NO_VNODE
1368 /*
1369 * There are known LORs between VNODE locks. They are
1370 * not an indication of a bug. VNODE locks are flagged
1371 * as such (LO_IS_VNODE) and we don't yell if the LOR
1372 * is between 2 VNODE locks.
1373 */
1374 if ((lock->lo_flags & LO_IS_VNODE) != 0 &&
1375 (lock1->li_lock->lo_flags & LO_IS_VNODE) != 0)
1376 return;
1377 #endif
1378
1379 /*
1380 * Ok, yell about it.
1381 */
1382 if (((lock->lo_flags & LO_SLEEPABLE) != 0 &&
1383 (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0))
1384 witness_output(
1385 "lock order reversal: (sleepable after non-sleepable)\n");
1386 else if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0
1387 && lock == &Giant.lock_object)
1388 witness_output(
1389 "lock order reversal: (Giant after non-sleepable)\n");
1390 else
1391 witness_output("lock order reversal:\n");
1392
1393 /*
1394 * Try to locate an earlier lock with
1395 * witness w in our list.
1396 */
1397 do {
1398 lock2 = &lle->ll_children[i];
1399 MPASS(lock2->li_lock != NULL);
1400 if (lock2->li_lock->lo_witness == w)
1401 break;
1402 if (i == 0 && lle->ll_next != NULL) {
1403 lle = lle->ll_next;
1404 i = lle->ll_count - 1;
1405 MPASS(i >= 0 && i < LOCK_NCHILDREN);
1406 } else
1407 i--;
1408 } while (i >= 0);
1409 if (i < 0) {
1410 witness_output(" 1st %p %s (%s) @ %s:%d\n",
1411 lock1->li_lock, lock1->li_lock->lo_name,
1412 w1->w_name, fixup_filename(lock1->li_file),
1413 lock1->li_line);
1414 witness_output(" 2nd %p %s (%s) @ %s:%d\n", lock,
1415 lock->lo_name, w->w_name,
1416 fixup_filename(file), line);
1417 } else {
1418 witness_output(" 1st %p %s (%s) @ %s:%d\n",
1419 lock2->li_lock, lock2->li_lock->lo_name,
1420 lock2->li_lock->lo_witness->w_name,
1421 fixup_filename(lock2->li_file),
1422 lock2->li_line);
1423 witness_output(" 2nd %p %s (%s) @ %s:%d\n",
1424 lock1->li_lock, lock1->li_lock->lo_name,
1425 w1->w_name, fixup_filename(lock1->li_file),
1426 lock1->li_line);
1427 witness_output(" 3rd %p %s (%s) @ %s:%d\n", lock,
1428 lock->lo_name, w->w_name,
1429 fixup_filename(file), line);
1430 }
1431 witness_debugger(1, __func__);
1432 return;
1433 }
1434 }
1435
1436 /*
1437 * If requested, build a new lock order. However, don't build a new
1438 * relationship between a sleepable lock and Giant if it is in the
1439 * wrong direction. The correct lock order is that sleepable locks
1440 * always come before Giant.
1441 */
1442 if (flags & LOP_NEWORDER &&
1443 !(plock->li_lock == &Giant.lock_object &&
1444 (lock->lo_flags & LO_SLEEPABLE) != 0)) {
1445 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__,
1446 w->w_name, plock->li_lock->lo_witness->w_name);
1447 itismychild(plock->li_lock->lo_witness, w);
1448 }
1449 out:
1450 mtx_unlock_spin(&w_mtx);
1451 }
1452
1453 void
witness_lock(struct lock_object * lock,int flags,const char * file,int line)1454 witness_lock(struct lock_object *lock, int flags, const char *file, int line)
1455 {
1456 struct lock_list_entry **lock_list, *lle;
1457 struct lock_instance *instance;
1458 struct witness *w;
1459 struct thread *td;
1460
1461 if (witness_cold || witness_watch == -1 || lock->lo_witness == NULL ||
1462 panicstr != NULL)
1463 return;
1464 w = lock->lo_witness;
1465 td = curthread;
1466
1467 /* Determine lock list for this lock. */
1468 if (LOCK_CLASS(lock)->lc_flags & LC_SLEEPLOCK)
1469 lock_list = &td->td_sleeplocks;
1470 else
1471 lock_list = PCPU_PTR(spinlocks);
1472
1473 /* Check to see if we are recursing on a lock we already own. */
1474 instance = find_instance(*lock_list, lock);
1475 if (instance != NULL) {
1476 instance->li_flags++;
1477 CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__,
1478 td->td_proc->p_pid, lock->lo_name,
1479 instance->li_flags & LI_RECURSEMASK);
1480 instance->li_file = file;
1481 instance->li_line = line;
1482 return;
1483 }
1484
1485 /* Update per-witness last file and line acquire. */
1486 w->w_file = file;
1487 w->w_line = line;
1488
1489 /* Find the next open lock instance in the list and fill it. */
1490 lle = *lock_list;
1491 if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) {
1492 lle = witness_lock_list_get();
1493 if (lle == NULL)
1494 return;
1495 lle->ll_next = *lock_list;
1496 CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__,
1497 td->td_proc->p_pid, lle);
1498 *lock_list = lle;
1499 }
1500 instance = &lle->ll_children[lle->ll_count++];
1501 instance->li_lock = lock;
1502 instance->li_line = line;
1503 instance->li_file = file;
1504 if ((flags & LOP_EXCLUSIVE) != 0)
1505 instance->li_flags = LI_EXCLUSIVE;
1506 else
1507 instance->li_flags = 0;
1508 CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__,
1509 td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1);
1510 }
1511
1512 void
witness_upgrade(struct lock_object * lock,int flags,const char * file,int line)1513 witness_upgrade(struct lock_object *lock, int flags, const char *file, int line)
1514 {
1515 struct lock_instance *instance;
1516 struct lock_class *class;
1517
1518 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1519 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
1520 return;
1521 class = LOCK_CLASS(lock);
1522 if (witness_watch) {
1523 if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1524 kassert_panic(
1525 "upgrade of non-upgradable lock (%s) %s @ %s:%d",
1526 class->lc_name, lock->lo_name,
1527 fixup_filename(file), line);
1528 if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1529 kassert_panic(
1530 "upgrade of non-sleep lock (%s) %s @ %s:%d",
1531 class->lc_name, lock->lo_name,
1532 fixup_filename(file), line);
1533 }
1534 instance = find_instance(curthread->td_sleeplocks, lock);
1535 if (instance == NULL) {
1536 kassert_panic("upgrade of unlocked lock (%s) %s @ %s:%d",
1537 class->lc_name, lock->lo_name,
1538 fixup_filename(file), line);
1539 return;
1540 }
1541 if (witness_watch) {
1542 if ((instance->li_flags & LI_EXCLUSIVE) != 0)
1543 kassert_panic(
1544 "upgrade of exclusive lock (%s) %s @ %s:%d",
1545 class->lc_name, lock->lo_name,
1546 fixup_filename(file), line);
1547 if ((instance->li_flags & LI_RECURSEMASK) != 0)
1548 kassert_panic(
1549 "upgrade of recursed lock (%s) %s r=%d @ %s:%d",
1550 class->lc_name, lock->lo_name,
1551 instance->li_flags & LI_RECURSEMASK,
1552 fixup_filename(file), line);
1553 }
1554 instance->li_flags |= LI_EXCLUSIVE;
1555 }
1556
1557 void
witness_downgrade(struct lock_object * lock,int flags,const char * file,int line)1558 witness_downgrade(struct lock_object *lock, int flags, const char *file,
1559 int line)
1560 {
1561 struct lock_instance *instance;
1562 struct lock_class *class;
1563
1564 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
1565 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
1566 return;
1567 class = LOCK_CLASS(lock);
1568 if (witness_watch) {
1569 if ((lock->lo_flags & LO_UPGRADABLE) == 0)
1570 kassert_panic(
1571 "downgrade of non-upgradable lock (%s) %s @ %s:%d",
1572 class->lc_name, lock->lo_name,
1573 fixup_filename(file), line);
1574 if ((class->lc_flags & LC_SLEEPLOCK) == 0)
1575 kassert_panic(
1576 "downgrade of non-sleep lock (%s) %s @ %s:%d",
1577 class->lc_name, lock->lo_name,
1578 fixup_filename(file), line);
1579 }
1580 instance = find_instance(curthread->td_sleeplocks, lock);
1581 if (instance == NULL) {
1582 kassert_panic("downgrade of unlocked lock (%s) %s @ %s:%d",
1583 class->lc_name, lock->lo_name,
1584 fixup_filename(file), line);
1585 return;
1586 }
1587 if (witness_watch) {
1588 if ((instance->li_flags & LI_EXCLUSIVE) == 0)
1589 kassert_panic(
1590 "downgrade of shared lock (%s) %s @ %s:%d",
1591 class->lc_name, lock->lo_name,
1592 fixup_filename(file), line);
1593 if ((instance->li_flags & LI_RECURSEMASK) != 0)
1594 kassert_panic(
1595 "downgrade of recursed lock (%s) %s r=%d @ %s:%d",
1596 class->lc_name, lock->lo_name,
1597 instance->li_flags & LI_RECURSEMASK,
1598 fixup_filename(file), line);
1599 }
1600 instance->li_flags &= ~LI_EXCLUSIVE;
1601 }
1602
1603 void
witness_unlock(struct lock_object * lock,int flags,const char * file,int line)1604 witness_unlock(struct lock_object *lock, int flags, const char *file, int line)
1605 {
1606 struct lock_list_entry **lock_list, *lle;
1607 struct lock_instance *instance;
1608 struct lock_class *class;
1609 struct thread *td;
1610 register_t s;
1611 int i, j;
1612
1613 if (witness_cold || lock->lo_witness == NULL || panicstr != NULL)
1614 return;
1615 td = curthread;
1616 class = LOCK_CLASS(lock);
1617
1618 /* Find lock instance associated with this lock. */
1619 if (class->lc_flags & LC_SLEEPLOCK)
1620 lock_list = &td->td_sleeplocks;
1621 else
1622 lock_list = PCPU_PTR(spinlocks);
1623 lle = *lock_list;
1624 for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next)
1625 for (i = 0; i < (*lock_list)->ll_count; i++) {
1626 instance = &(*lock_list)->ll_children[i];
1627 if (instance->li_lock == lock)
1628 goto found;
1629 }
1630
1631 /*
1632 * When disabling WITNESS through witness_watch we could end up in
1633 * having registered locks in the td_sleeplocks queue.
1634 * We have to make sure we flush these queues, so just search for
1635 * eventual register locks and remove them.
1636 */
1637 if (witness_watch > 0) {
1638 kassert_panic("lock (%s) %s not locked @ %s:%d", class->lc_name,
1639 lock->lo_name, fixup_filename(file), line);
1640 return;
1641 } else {
1642 return;
1643 }
1644 found:
1645
1646 /* First, check for shared/exclusive mismatches. */
1647 if ((instance->li_flags & LI_EXCLUSIVE) != 0 && witness_watch > 0 &&
1648 (flags & LOP_EXCLUSIVE) == 0) {
1649 witness_output("shared unlock of (%s) %s @ %s:%d\n",
1650 class->lc_name, lock->lo_name, fixup_filename(file), line);
1651 witness_output("while exclusively locked from %s:%d\n",
1652 fixup_filename(instance->li_file), instance->li_line);
1653 kassert_panic("excl->ushare");
1654 }
1655 if ((instance->li_flags & LI_EXCLUSIVE) == 0 && witness_watch > 0 &&
1656 (flags & LOP_EXCLUSIVE) != 0) {
1657 witness_output("exclusive unlock of (%s) %s @ %s:%d\n",
1658 class->lc_name, lock->lo_name, fixup_filename(file), line);
1659 witness_output("while share locked from %s:%d\n",
1660 fixup_filename(instance->li_file),
1661 instance->li_line);
1662 kassert_panic("share->uexcl");
1663 }
1664 /* If we are recursed, unrecurse. */
1665 if ((instance->li_flags & LI_RECURSEMASK) > 0) {
1666 CTR4(KTR_WITNESS, "%s: pid %d unrecursed on %s r=%d", __func__,
1667 td->td_proc->p_pid, instance->li_lock->lo_name,
1668 instance->li_flags);
1669 instance->li_flags--;
1670 return;
1671 }
1672 /* The lock is now being dropped, check for NORELEASE flag */
1673 if ((instance->li_flags & LI_NORELEASE) != 0 && witness_watch > 0) {
1674 witness_output("forbidden unlock of (%s) %s @ %s:%d\n",
1675 class->lc_name, lock->lo_name, fixup_filename(file), line);
1676 kassert_panic("lock marked norelease");
1677 }
1678
1679 /* Otherwise, remove this item from the list. */
1680 s = intr_disable();
1681 CTR4(KTR_WITNESS, "%s: pid %d removed %s from lle[%d]", __func__,
1682 td->td_proc->p_pid, instance->li_lock->lo_name,
1683 (*lock_list)->ll_count - 1);
1684 for (j = i; j < (*lock_list)->ll_count - 1; j++)
1685 (*lock_list)->ll_children[j] =
1686 (*lock_list)->ll_children[j + 1];
1687 (*lock_list)->ll_count--;
1688 intr_restore(s);
1689
1690 /*
1691 * In order to reduce contention on w_mtx, we want to keep always an
1692 * head object into lists so that frequent allocation from the
1693 * free witness pool (and subsequent locking) is avoided.
1694 * In order to maintain the current code simple, when the head
1695 * object is totally unloaded it means also that we do not have
1696 * further objects in the list, so the list ownership needs to be
1697 * hand over to another object if the current head needs to be freed.
1698 */
1699 if ((*lock_list)->ll_count == 0) {
1700 if (*lock_list == lle) {
1701 if (lle->ll_next == NULL)
1702 return;
1703 } else
1704 lle = *lock_list;
1705 *lock_list = lle->ll_next;
1706 CTR3(KTR_WITNESS, "%s: pid %d removed lle %p", __func__,
1707 td->td_proc->p_pid, lle);
1708 witness_lock_list_free(lle);
1709 }
1710 }
1711
1712 void
witness_thread_exit(struct thread * td)1713 witness_thread_exit(struct thread *td)
1714 {
1715 struct lock_list_entry *lle;
1716 int i, n;
1717
1718 lle = td->td_sleeplocks;
1719 if (lle == NULL || panicstr != NULL)
1720 return;
1721 if (lle->ll_count != 0) {
1722 for (n = 0; lle != NULL; lle = lle->ll_next)
1723 for (i = lle->ll_count - 1; i >= 0; i--) {
1724 if (n == 0)
1725 witness_output(
1726 "Thread %p exiting with the following locks held:\n", td);
1727 n++;
1728 witness_list_lock(&lle->ll_children[i],
1729 witness_output);
1730
1731 }
1732 kassert_panic(
1733 "Thread %p cannot exit while holding sleeplocks\n", td);
1734 }
1735 witness_lock_list_free(lle);
1736 }
1737
1738 /*
1739 * Warn if any locks other than 'lock' are held. Flags can be passed in to
1740 * exempt Giant and sleepable locks from the checks as well. If any
1741 * non-exempt locks are held, then a supplied message is printed to the
1742 * output channel along with a list of the offending locks. If indicated in the
1743 * flags then a failure results in a panic as well.
1744 */
1745 int
witness_warn(int flags,struct lock_object * lock,const char * fmt,...)1746 witness_warn(int flags, struct lock_object *lock, const char *fmt, ...)
1747 {
1748 struct lock_list_entry *lock_list, *lle;
1749 struct lock_instance *lock1;
1750 struct thread *td;
1751 va_list ap;
1752 int i, n;
1753
1754 if (witness_cold || witness_watch < 1 || panicstr != NULL)
1755 return (0);
1756 n = 0;
1757 td = curthread;
1758 for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next)
1759 for (i = lle->ll_count - 1; i >= 0; i--) {
1760 lock1 = &lle->ll_children[i];
1761 if (lock1->li_lock == lock)
1762 continue;
1763 if (flags & WARN_GIANTOK &&
1764 lock1->li_lock == &Giant.lock_object)
1765 continue;
1766 if (flags & WARN_SLEEPOK &&
1767 (lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0)
1768 continue;
1769 if (n == 0) {
1770 va_start(ap, fmt);
1771 vprintf(fmt, ap);
1772 va_end(ap);
1773 printf(" with the following %slocks held:\n",
1774 (flags & WARN_SLEEPOK) != 0 ?
1775 "non-sleepable " : "");
1776 }
1777 n++;
1778 witness_list_lock(lock1, printf);
1779 }
1780
1781 /*
1782 * Pin the thread in order to avoid problems with thread migration.
1783 * Once that all verifies are passed about spinlocks ownership,
1784 * the thread is in a safe path and it can be unpinned.
1785 */
1786 sched_pin();
1787 lock_list = PCPU_GET(spinlocks);
1788 if (lock_list != NULL && lock_list->ll_count != 0) {
1789 sched_unpin();
1790
1791 /*
1792 * We should only have one spinlock and as long as
1793 * the flags cannot match for this locks class,
1794 * check if the first spinlock is the one curthread
1795 * should hold.
1796 */
1797 lock1 = &lock_list->ll_children[lock_list->ll_count - 1];
1798 if (lock_list->ll_count == 1 && lock_list->ll_next == NULL &&
1799 lock1->li_lock == lock && n == 0)
1800 return (0);
1801
1802 va_start(ap, fmt);
1803 vprintf(fmt, ap);
1804 va_end(ap);
1805 printf(" with the following %slocks held:\n",
1806 (flags & WARN_SLEEPOK) != 0 ? "non-sleepable " : "");
1807 n += witness_list_locks(&lock_list, printf);
1808 } else
1809 sched_unpin();
1810 if (flags & WARN_PANIC && n)
1811 kassert_panic("%s", __func__);
1812 else
1813 witness_debugger(n, __func__);
1814 return (n);
1815 }
1816
1817 const char *
witness_file(struct lock_object * lock)1818 witness_file(struct lock_object *lock)
1819 {
1820 struct witness *w;
1821
1822 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1823 return ("?");
1824 w = lock->lo_witness;
1825 return (w->w_file);
1826 }
1827
1828 int
witness_line(struct lock_object * lock)1829 witness_line(struct lock_object *lock)
1830 {
1831 struct witness *w;
1832
1833 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL)
1834 return (0);
1835 w = lock->lo_witness;
1836 return (w->w_line);
1837 }
1838
1839 static struct witness *
enroll(const char * description,struct lock_class * lock_class)1840 enroll(const char *description, struct lock_class *lock_class)
1841 {
1842 struct witness *w;
1843
1844 MPASS(description != NULL);
1845
1846 if (witness_watch == -1 || panicstr != NULL)
1847 return (NULL);
1848 if ((lock_class->lc_flags & LC_SPINLOCK)) {
1849 if (witness_skipspin)
1850 return (NULL);
1851 } else if ((lock_class->lc_flags & LC_SLEEPLOCK) == 0) {
1852 kassert_panic("lock class %s is not sleep or spin",
1853 lock_class->lc_name);
1854 return (NULL);
1855 }
1856
1857 mtx_lock_spin(&w_mtx);
1858 w = witness_hash_get(description);
1859 if (w)
1860 goto found;
1861 if ((w = witness_get()) == NULL)
1862 return (NULL);
1863 MPASS(strlen(description) < MAX_W_NAME);
1864 strcpy(w->w_name, description);
1865 w->w_class = lock_class;
1866 w->w_refcount = 1;
1867 STAILQ_INSERT_HEAD(&w_all, w, w_list);
1868 if (lock_class->lc_flags & LC_SPINLOCK) {
1869 STAILQ_INSERT_HEAD(&w_spin, w, w_typelist);
1870 w_spin_cnt++;
1871 } else if (lock_class->lc_flags & LC_SLEEPLOCK) {
1872 STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist);
1873 w_sleep_cnt++;
1874 }
1875
1876 /* Insert new witness into the hash */
1877 witness_hash_put(w);
1878 witness_increment_graph_generation();
1879 mtx_unlock_spin(&w_mtx);
1880 return (w);
1881 found:
1882 w->w_refcount++;
1883 if (w->w_refcount == 1)
1884 w->w_class = lock_class;
1885 mtx_unlock_spin(&w_mtx);
1886 if (lock_class != w->w_class)
1887 kassert_panic(
1888 "lock (%s) %s does not match earlier (%s) lock",
1889 description, lock_class->lc_name,
1890 w->w_class->lc_name);
1891 return (w);
1892 }
1893
1894 static void
depart(struct witness * w)1895 depart(struct witness *w)
1896 {
1897
1898 MPASS(w->w_refcount == 0);
1899 if (w->w_class->lc_flags & LC_SLEEPLOCK) {
1900 w_sleep_cnt--;
1901 } else {
1902 w_spin_cnt--;
1903 }
1904 /*
1905 * Set file to NULL as it may point into a loadable module.
1906 */
1907 w->w_file = NULL;
1908 w->w_line = 0;
1909 witness_increment_graph_generation();
1910 }
1911
1912
1913 static void
adopt(struct witness * parent,struct witness * child)1914 adopt(struct witness *parent, struct witness *child)
1915 {
1916 int pi, ci, i, j;
1917
1918 if (witness_cold == 0)
1919 mtx_assert(&w_mtx, MA_OWNED);
1920
1921 /* If the relationship is already known, there's no work to be done. */
1922 if (isitmychild(parent, child))
1923 return;
1924
1925 /* When the structure of the graph changes, bump up the generation. */
1926 witness_increment_graph_generation();
1927
1928 /*
1929 * The hard part ... create the direct relationship, then propagate all
1930 * indirect relationships.
1931 */
1932 pi = parent->w_index;
1933 ci = child->w_index;
1934 WITNESS_INDEX_ASSERT(pi);
1935 WITNESS_INDEX_ASSERT(ci);
1936 MPASS(pi != ci);
1937 w_rmatrix[pi][ci] |= WITNESS_PARENT;
1938 w_rmatrix[ci][pi] |= WITNESS_CHILD;
1939
1940 /*
1941 * If parent was not already an ancestor of child,
1942 * then we increment the descendant and ancestor counters.
1943 */
1944 if ((w_rmatrix[pi][ci] & WITNESS_ANCESTOR) == 0) {
1945 parent->w_num_descendants++;
1946 child->w_num_ancestors++;
1947 }
1948
1949 /*
1950 * Find each ancestor of 'pi'. Note that 'pi' itself is counted as
1951 * an ancestor of 'pi' during this loop.
1952 */
1953 for (i = 1; i <= w_max_used_index; i++) {
1954 if ((w_rmatrix[i][pi] & WITNESS_ANCESTOR_MASK) == 0 &&
1955 (i != pi))
1956 continue;
1957
1958 /* Find each descendant of 'i' and mark it as a descendant. */
1959 for (j = 1; j <= w_max_used_index; j++) {
1960
1961 /*
1962 * Skip children that are already marked as
1963 * descendants of 'i'.
1964 */
1965 if (w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK)
1966 continue;
1967
1968 /*
1969 * We are only interested in descendants of 'ci'. Note
1970 * that 'ci' itself is counted as a descendant of 'ci'.
1971 */
1972 if ((w_rmatrix[ci][j] & WITNESS_ANCESTOR_MASK) == 0 &&
1973 (j != ci))
1974 continue;
1975 w_rmatrix[i][j] |= WITNESS_ANCESTOR;
1976 w_rmatrix[j][i] |= WITNESS_DESCENDANT;
1977 w_data[i].w_num_descendants++;
1978 w_data[j].w_num_ancestors++;
1979
1980 /*
1981 * Make sure we aren't marking a node as both an
1982 * ancestor and descendant. We should have caught
1983 * this as a lock order reversal earlier.
1984 */
1985 if ((w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK) &&
1986 (w_rmatrix[i][j] & WITNESS_DESCENDANT_MASK)) {
1987 printf("witness rmatrix paradox! [%d][%d]=%d "
1988 "both ancestor and descendant\n",
1989 i, j, w_rmatrix[i][j]);
1990 kdb_backtrace();
1991 printf("Witness disabled.\n");
1992 witness_watch = -1;
1993 }
1994 if ((w_rmatrix[j][i] & WITNESS_ANCESTOR_MASK) &&
1995 (w_rmatrix[j][i] & WITNESS_DESCENDANT_MASK)) {
1996 printf("witness rmatrix paradox! [%d][%d]=%d "
1997 "both ancestor and descendant\n",
1998 j, i, w_rmatrix[j][i]);
1999 kdb_backtrace();
2000 printf("Witness disabled.\n");
2001 witness_watch = -1;
2002 }
2003 }
2004 }
2005 }
2006
2007 static void
itismychild(struct witness * parent,struct witness * child)2008 itismychild(struct witness *parent, struct witness *child)
2009 {
2010 int unlocked;
2011
2012 MPASS(child != NULL && parent != NULL);
2013 if (witness_cold == 0)
2014 mtx_assert(&w_mtx, MA_OWNED);
2015
2016 if (!witness_lock_type_equal(parent, child)) {
2017 if (witness_cold == 0) {
2018 unlocked = 1;
2019 mtx_unlock_spin(&w_mtx);
2020 } else {
2021 unlocked = 0;
2022 }
2023 kassert_panic(
2024 "%s: parent \"%s\" (%s) and child \"%s\" (%s) are not "
2025 "the same lock type", __func__, parent->w_name,
2026 parent->w_class->lc_name, child->w_name,
2027 child->w_class->lc_name);
2028 if (unlocked)
2029 mtx_lock_spin(&w_mtx);
2030 }
2031 adopt(parent, child);
2032 }
2033
2034 /*
2035 * Generic code for the isitmy*() functions. The rmask parameter is the
2036 * expected relationship of w1 to w2.
2037 */
2038 static int
_isitmyx(struct witness * w1,struct witness * w2,int rmask,const char * fname)2039 _isitmyx(struct witness *w1, struct witness *w2, int rmask, const char *fname)
2040 {
2041 unsigned char r1, r2;
2042 int i1, i2;
2043
2044 i1 = w1->w_index;
2045 i2 = w2->w_index;
2046 WITNESS_INDEX_ASSERT(i1);
2047 WITNESS_INDEX_ASSERT(i2);
2048 r1 = w_rmatrix[i1][i2] & WITNESS_RELATED_MASK;
2049 r2 = w_rmatrix[i2][i1] & WITNESS_RELATED_MASK;
2050
2051 /* The flags on one better be the inverse of the flags on the other */
2052 if (!((WITNESS_ATOD(r1) == r2 && WITNESS_DTOA(r2) == r1) ||
2053 (WITNESS_DTOA(r1) == r2 && WITNESS_ATOD(r2) == r1))) {
2054 /* Don't squawk if we're potentially racing with an update. */
2055 if (!mtx_owned(&w_mtx))
2056 return (0);
2057 printf("%s: rmatrix mismatch between %s (index %d) and %s "
2058 "(index %d): w_rmatrix[%d][%d] == %hhx but "
2059 "w_rmatrix[%d][%d] == %hhx\n",
2060 fname, w1->w_name, i1, w2->w_name, i2, i1, i2, r1,
2061 i2, i1, r2);
2062 kdb_backtrace();
2063 printf("Witness disabled.\n");
2064 witness_watch = -1;
2065 }
2066 return (r1 & rmask);
2067 }
2068
2069 /*
2070 * Checks if @child is a direct child of @parent.
2071 */
2072 static int
isitmychild(struct witness * parent,struct witness * child)2073 isitmychild(struct witness *parent, struct witness *child)
2074 {
2075
2076 return (_isitmyx(parent, child, WITNESS_PARENT, __func__));
2077 }
2078
2079 /*
2080 * Checks if @descendant is a direct or inderect descendant of @ancestor.
2081 */
2082 static int
isitmydescendant(struct witness * ancestor,struct witness * descendant)2083 isitmydescendant(struct witness *ancestor, struct witness *descendant)
2084 {
2085
2086 return (_isitmyx(ancestor, descendant, WITNESS_ANCESTOR_MASK,
2087 __func__));
2088 }
2089
2090 static int
blessed(struct witness * w1,struct witness * w2)2091 blessed(struct witness *w1, struct witness *w2)
2092 {
2093 int i;
2094 struct witness_blessed *b;
2095
2096 for (i = 0; i < nitems(blessed_list); i++) {
2097 b = &blessed_list[i];
2098 if (strcmp(w1->w_name, b->b_lock1) == 0) {
2099 if (strcmp(w2->w_name, b->b_lock2) == 0)
2100 return (1);
2101 continue;
2102 }
2103 if (strcmp(w1->w_name, b->b_lock2) == 0)
2104 if (strcmp(w2->w_name, b->b_lock1) == 0)
2105 return (1);
2106 }
2107 return (0);
2108 }
2109
2110 static struct witness *
witness_get(void)2111 witness_get(void)
2112 {
2113 struct witness *w;
2114 int index;
2115
2116 if (witness_cold == 0)
2117 mtx_assert(&w_mtx, MA_OWNED);
2118
2119 if (witness_watch == -1) {
2120 mtx_unlock_spin(&w_mtx);
2121 return (NULL);
2122 }
2123 if (STAILQ_EMPTY(&w_free)) {
2124 witness_watch = -1;
2125 mtx_unlock_spin(&w_mtx);
2126 printf("WITNESS: unable to allocate a new witness object\n");
2127 return (NULL);
2128 }
2129 w = STAILQ_FIRST(&w_free);
2130 STAILQ_REMOVE_HEAD(&w_free, w_list);
2131 w_free_cnt--;
2132 index = w->w_index;
2133 MPASS(index > 0 && index == w_max_used_index+1 &&
2134 index < witness_count);
2135 bzero(w, sizeof(*w));
2136 w->w_index = index;
2137 if (index > w_max_used_index)
2138 w_max_used_index = index;
2139 return (w);
2140 }
2141
2142 static void
witness_free(struct witness * w)2143 witness_free(struct witness *w)
2144 {
2145
2146 STAILQ_INSERT_HEAD(&w_free, w, w_list);
2147 w_free_cnt++;
2148 }
2149
2150 static struct lock_list_entry *
witness_lock_list_get(void)2151 witness_lock_list_get(void)
2152 {
2153 struct lock_list_entry *lle;
2154
2155 if (witness_watch == -1)
2156 return (NULL);
2157 mtx_lock_spin(&w_mtx);
2158 lle = w_lock_list_free;
2159 if (lle == NULL) {
2160 witness_watch = -1;
2161 mtx_unlock_spin(&w_mtx);
2162 printf("%s: witness exhausted\n", __func__);
2163 return (NULL);
2164 }
2165 w_lock_list_free = lle->ll_next;
2166 mtx_unlock_spin(&w_mtx);
2167 bzero(lle, sizeof(*lle));
2168 return (lle);
2169 }
2170
2171 static void
witness_lock_list_free(struct lock_list_entry * lle)2172 witness_lock_list_free(struct lock_list_entry *lle)
2173 {
2174
2175 mtx_lock_spin(&w_mtx);
2176 lle->ll_next = w_lock_list_free;
2177 w_lock_list_free = lle;
2178 mtx_unlock_spin(&w_mtx);
2179 }
2180
2181 static struct lock_instance *
find_instance(struct lock_list_entry * list,const struct lock_object * lock)2182 find_instance(struct lock_list_entry *list, const struct lock_object *lock)
2183 {
2184 struct lock_list_entry *lle;
2185 struct lock_instance *instance;
2186 int i;
2187
2188 for (lle = list; lle != NULL; lle = lle->ll_next)
2189 for (i = lle->ll_count - 1; i >= 0; i--) {
2190 instance = &lle->ll_children[i];
2191 if (instance->li_lock == lock)
2192 return (instance);
2193 }
2194 return (NULL);
2195 }
2196
2197 static void
witness_list_lock(struct lock_instance * instance,int (* prnt)(const char * fmt,...))2198 witness_list_lock(struct lock_instance *instance,
2199 int (*prnt)(const char *fmt, ...))
2200 {
2201 struct lock_object *lock;
2202
2203 lock = instance->li_lock;
2204 prnt("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ?
2205 "exclusive" : "shared", LOCK_CLASS(lock)->lc_name, lock->lo_name);
2206 if (lock->lo_witness->w_name != lock->lo_name)
2207 prnt(" (%s)", lock->lo_witness->w_name);
2208 prnt(" r = %d (%p) locked @ %s:%d\n",
2209 instance->li_flags & LI_RECURSEMASK, lock,
2210 fixup_filename(instance->li_file), instance->li_line);
2211 }
2212
2213 static int
witness_output(const char * fmt,...)2214 witness_output(const char *fmt, ...)
2215 {
2216 va_list ap;
2217 int ret;
2218
2219 va_start(ap, fmt);
2220 ret = witness_voutput(fmt, ap);
2221 va_end(ap);
2222 return (ret);
2223 }
2224
2225 static int
witness_voutput(const char * fmt,va_list ap)2226 witness_voutput(const char *fmt, va_list ap)
2227 {
2228 int ret;
2229
2230 ret = 0;
2231 switch (witness_channel) {
2232 case WITNESS_CONSOLE:
2233 ret = vprintf(fmt, ap);
2234 break;
2235 case WITNESS_LOG:
2236 vlog(LOG_NOTICE, fmt, ap);
2237 break;
2238 case WITNESS_NONE:
2239 break;
2240 }
2241 return (ret);
2242 }
2243
2244 #ifdef DDB
2245 static int
witness_thread_has_locks(struct thread * td)2246 witness_thread_has_locks(struct thread *td)
2247 {
2248
2249 if (td->td_sleeplocks == NULL)
2250 return (0);
2251 return (td->td_sleeplocks->ll_count != 0);
2252 }
2253
2254 static int
witness_proc_has_locks(struct proc * p)2255 witness_proc_has_locks(struct proc *p)
2256 {
2257 struct thread *td;
2258
2259 FOREACH_THREAD_IN_PROC(p, td) {
2260 if (witness_thread_has_locks(td))
2261 return (1);
2262 }
2263 return (0);
2264 }
2265 #endif
2266
2267 int
witness_list_locks(struct lock_list_entry ** lock_list,int (* prnt)(const char * fmt,...))2268 witness_list_locks(struct lock_list_entry **lock_list,
2269 int (*prnt)(const char *fmt, ...))
2270 {
2271 struct lock_list_entry *lle;
2272 int i, nheld;
2273
2274 nheld = 0;
2275 for (lle = *lock_list; lle != NULL; lle = lle->ll_next)
2276 for (i = lle->ll_count - 1; i >= 0; i--) {
2277 witness_list_lock(&lle->ll_children[i], prnt);
2278 nheld++;
2279 }
2280 return (nheld);
2281 }
2282
2283 /*
2284 * This is a bit risky at best. We call this function when we have timed
2285 * out acquiring a spin lock, and we assume that the other CPU is stuck
2286 * with this lock held. So, we go groveling around in the other CPU's
2287 * per-cpu data to try to find the lock instance for this spin lock to
2288 * see when it was last acquired.
2289 */
2290 void
witness_display_spinlock(struct lock_object * lock,struct thread * owner,int (* prnt)(const char * fmt,...))2291 witness_display_spinlock(struct lock_object *lock, struct thread *owner,
2292 int (*prnt)(const char *fmt, ...))
2293 {
2294 struct lock_instance *instance;
2295 struct pcpu *pc;
2296
2297 if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU)
2298 return;
2299 pc = pcpu_find(owner->td_oncpu);
2300 instance = find_instance(pc->pc_spinlocks, lock);
2301 if (instance != NULL)
2302 witness_list_lock(instance, prnt);
2303 }
2304
2305 void
witness_save(struct lock_object * lock,const char ** filep,int * linep)2306 witness_save(struct lock_object *lock, const char **filep, int *linep)
2307 {
2308 struct lock_list_entry *lock_list;
2309 struct lock_instance *instance;
2310 struct lock_class *class;
2311
2312 /*
2313 * This function is used independently in locking code to deal with
2314 * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2315 * is gone.
2316 */
2317 if (SCHEDULER_STOPPED())
2318 return;
2319 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2320 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2321 return;
2322 class = LOCK_CLASS(lock);
2323 if (class->lc_flags & LC_SLEEPLOCK)
2324 lock_list = curthread->td_sleeplocks;
2325 else {
2326 if (witness_skipspin)
2327 return;
2328 lock_list = PCPU_GET(spinlocks);
2329 }
2330 instance = find_instance(lock_list, lock);
2331 if (instance == NULL) {
2332 kassert_panic("%s: lock (%s) %s not locked", __func__,
2333 class->lc_name, lock->lo_name);
2334 return;
2335 }
2336 *filep = instance->li_file;
2337 *linep = instance->li_line;
2338 }
2339
2340 void
witness_restore(struct lock_object * lock,const char * file,int line)2341 witness_restore(struct lock_object *lock, const char *file, int line)
2342 {
2343 struct lock_list_entry *lock_list;
2344 struct lock_instance *instance;
2345 struct lock_class *class;
2346
2347 /*
2348 * This function is used independently in locking code to deal with
2349 * Giant, SCHEDULER_STOPPED() check can be removed here after Giant
2350 * is gone.
2351 */
2352 if (SCHEDULER_STOPPED())
2353 return;
2354 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2355 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2356 return;
2357 class = LOCK_CLASS(lock);
2358 if (class->lc_flags & LC_SLEEPLOCK)
2359 lock_list = curthread->td_sleeplocks;
2360 else {
2361 if (witness_skipspin)
2362 return;
2363 lock_list = PCPU_GET(spinlocks);
2364 }
2365 instance = find_instance(lock_list, lock);
2366 if (instance == NULL)
2367 kassert_panic("%s: lock (%s) %s not locked", __func__,
2368 class->lc_name, lock->lo_name);
2369 lock->lo_witness->w_file = file;
2370 lock->lo_witness->w_line = line;
2371 if (instance == NULL)
2372 return;
2373 instance->li_file = file;
2374 instance->li_line = line;
2375 }
2376
2377 void
witness_assert(const struct lock_object * lock,int flags,const char * file,int line)2378 witness_assert(const struct lock_object *lock, int flags, const char *file,
2379 int line)
2380 {
2381 #ifdef INVARIANT_SUPPORT
2382 struct lock_instance *instance;
2383 struct lock_class *class;
2384
2385 if (lock->lo_witness == NULL || witness_watch < 1 || panicstr != NULL)
2386 return;
2387 class = LOCK_CLASS(lock);
2388 if ((class->lc_flags & LC_SLEEPLOCK) != 0)
2389 instance = find_instance(curthread->td_sleeplocks, lock);
2390 else if ((class->lc_flags & LC_SPINLOCK) != 0)
2391 instance = find_instance(PCPU_GET(spinlocks), lock);
2392 else {
2393 kassert_panic("Lock (%s) %s is not sleep or spin!",
2394 class->lc_name, lock->lo_name);
2395 return;
2396 }
2397 switch (flags) {
2398 case LA_UNLOCKED:
2399 if (instance != NULL)
2400 kassert_panic("Lock (%s) %s locked @ %s:%d.",
2401 class->lc_name, lock->lo_name,
2402 fixup_filename(file), line);
2403 break;
2404 case LA_LOCKED:
2405 case LA_LOCKED | LA_RECURSED:
2406 case LA_LOCKED | LA_NOTRECURSED:
2407 case LA_SLOCKED:
2408 case LA_SLOCKED | LA_RECURSED:
2409 case LA_SLOCKED | LA_NOTRECURSED:
2410 case LA_XLOCKED:
2411 case LA_XLOCKED | LA_RECURSED:
2412 case LA_XLOCKED | LA_NOTRECURSED:
2413 if (instance == NULL) {
2414 kassert_panic("Lock (%s) %s not locked @ %s:%d.",
2415 class->lc_name, lock->lo_name,
2416 fixup_filename(file), line);
2417 break;
2418 }
2419 if ((flags & LA_XLOCKED) != 0 &&
2420 (instance->li_flags & LI_EXCLUSIVE) == 0)
2421 kassert_panic(
2422 "Lock (%s) %s not exclusively locked @ %s:%d.",
2423 class->lc_name, lock->lo_name,
2424 fixup_filename(file), line);
2425 if ((flags & LA_SLOCKED) != 0 &&
2426 (instance->li_flags & LI_EXCLUSIVE) != 0)
2427 kassert_panic(
2428 "Lock (%s) %s exclusively locked @ %s:%d.",
2429 class->lc_name, lock->lo_name,
2430 fixup_filename(file), line);
2431 if ((flags & LA_RECURSED) != 0 &&
2432 (instance->li_flags & LI_RECURSEMASK) == 0)
2433 kassert_panic("Lock (%s) %s not recursed @ %s:%d.",
2434 class->lc_name, lock->lo_name,
2435 fixup_filename(file), line);
2436 if ((flags & LA_NOTRECURSED) != 0 &&
2437 (instance->li_flags & LI_RECURSEMASK) != 0)
2438 kassert_panic("Lock (%s) %s recursed @ %s:%d.",
2439 class->lc_name, lock->lo_name,
2440 fixup_filename(file), line);
2441 break;
2442 default:
2443 kassert_panic("Invalid lock assertion at %s:%d.",
2444 fixup_filename(file), line);
2445
2446 }
2447 #endif /* INVARIANT_SUPPORT */
2448 }
2449
2450 static void
witness_setflag(struct lock_object * lock,int flag,int set)2451 witness_setflag(struct lock_object *lock, int flag, int set)
2452 {
2453 struct lock_list_entry *lock_list;
2454 struct lock_instance *instance;
2455 struct lock_class *class;
2456
2457 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL)
2458 return;
2459 class = LOCK_CLASS(lock);
2460 if (class->lc_flags & LC_SLEEPLOCK)
2461 lock_list = curthread->td_sleeplocks;
2462 else {
2463 if (witness_skipspin)
2464 return;
2465 lock_list = PCPU_GET(spinlocks);
2466 }
2467 instance = find_instance(lock_list, lock);
2468 if (instance == NULL) {
2469 kassert_panic("%s: lock (%s) %s not locked", __func__,
2470 class->lc_name, lock->lo_name);
2471 return;
2472 }
2473
2474 if (set)
2475 instance->li_flags |= flag;
2476 else
2477 instance->li_flags &= ~flag;
2478 }
2479
2480 void
witness_norelease(struct lock_object * lock)2481 witness_norelease(struct lock_object *lock)
2482 {
2483
2484 witness_setflag(lock, LI_NORELEASE, 1);
2485 }
2486
2487 void
witness_releaseok(struct lock_object * lock)2488 witness_releaseok(struct lock_object *lock)
2489 {
2490
2491 witness_setflag(lock, LI_NORELEASE, 0);
2492 }
2493
2494 #ifdef DDB
2495 static void
witness_ddb_list(struct thread * td)2496 witness_ddb_list(struct thread *td)
2497 {
2498
2499 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__));
2500 KASSERT(kdb_active, ("%s: not in the debugger", __func__));
2501
2502 if (witness_watch < 1)
2503 return;
2504
2505 witness_list_locks(&td->td_sleeplocks, db_printf);
2506
2507 /*
2508 * We only handle spinlocks if td == curthread. This is somewhat broken
2509 * if td is currently executing on some other CPU and holds spin locks
2510 * as we won't display those locks. If we had a MI way of getting
2511 * the per-cpu data for a given cpu then we could use
2512 * td->td_oncpu to get the list of spinlocks for this thread
2513 * and "fix" this.
2514 *
2515 * That still wouldn't really fix this unless we locked the scheduler
2516 * lock or stopped the other CPU to make sure it wasn't changing the
2517 * list out from under us. It is probably best to just not try to
2518 * handle threads on other CPU's for now.
2519 */
2520 if (td == curthread && PCPU_GET(spinlocks) != NULL)
2521 witness_list_locks(PCPU_PTR(spinlocks), db_printf);
2522 }
2523
DB_SHOW_COMMAND(locks,db_witness_list)2524 DB_SHOW_COMMAND(locks, db_witness_list)
2525 {
2526 struct thread *td;
2527
2528 if (have_addr)
2529 td = db_lookup_thread(addr, true);
2530 else
2531 td = kdb_thread;
2532 witness_ddb_list(td);
2533 }
2534
DB_SHOW_ALL_COMMAND(locks,db_witness_list_all)2535 DB_SHOW_ALL_COMMAND(locks, db_witness_list_all)
2536 {
2537 struct thread *td;
2538 struct proc *p;
2539
2540 /*
2541 * It would be nice to list only threads and processes that actually
2542 * held sleep locks, but that information is currently not exported
2543 * by WITNESS.
2544 */
2545 FOREACH_PROC_IN_SYSTEM(p) {
2546 if (!witness_proc_has_locks(p))
2547 continue;
2548 FOREACH_THREAD_IN_PROC(p, td) {
2549 if (!witness_thread_has_locks(td))
2550 continue;
2551 db_printf("Process %d (%s) thread %p (%d)\n", p->p_pid,
2552 p->p_comm, td, td->td_tid);
2553 witness_ddb_list(td);
2554 if (db_pager_quit)
2555 return;
2556 }
2557 }
2558 }
DB_SHOW_ALIAS(alllocks,db_witness_list_all)2559 DB_SHOW_ALIAS(alllocks, db_witness_list_all)
2560
2561 DB_SHOW_COMMAND(witness, db_witness_display)
2562 {
2563
2564 witness_ddb_display(db_printf);
2565 }
2566 #endif
2567
2568 static void
sbuf_print_witness_badstacks(struct sbuf * sb,size_t * oldidx)2569 sbuf_print_witness_badstacks(struct sbuf *sb, size_t *oldidx)
2570 {
2571 struct witness_lock_order_data *data1, *data2, *tmp_data1, *tmp_data2;
2572 struct witness *tmp_w1, *tmp_w2, *w1, *w2;
2573 int generation, i, j;
2574
2575 tmp_data1 = NULL;
2576 tmp_data2 = NULL;
2577 tmp_w1 = NULL;
2578 tmp_w2 = NULL;
2579
2580 /* Allocate and init temporary storage space. */
2581 tmp_w1 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2582 tmp_w2 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO);
2583 tmp_data1 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2584 M_WAITOK | M_ZERO);
2585 tmp_data2 = malloc(sizeof(struct witness_lock_order_data), M_TEMP,
2586 M_WAITOK | M_ZERO);
2587 stack_zero(&tmp_data1->wlod_stack);
2588 stack_zero(&tmp_data2->wlod_stack);
2589
2590 restart:
2591 mtx_lock_spin(&w_mtx);
2592 generation = w_generation;
2593 mtx_unlock_spin(&w_mtx);
2594 sbuf_printf(sb, "Number of known direct relationships is %d\n",
2595 w_lohash.wloh_count);
2596 for (i = 1; i < w_max_used_index; i++) {
2597 mtx_lock_spin(&w_mtx);
2598 if (generation != w_generation) {
2599 mtx_unlock_spin(&w_mtx);
2600
2601 /* The graph has changed, try again. */
2602 *oldidx = 0;
2603 sbuf_clear(sb);
2604 goto restart;
2605 }
2606
2607 w1 = &w_data[i];
2608 if (w1->w_reversed == 0) {
2609 mtx_unlock_spin(&w_mtx);
2610 continue;
2611 }
2612
2613 /* Copy w1 locally so we can release the spin lock. */
2614 *tmp_w1 = *w1;
2615 mtx_unlock_spin(&w_mtx);
2616
2617 if (tmp_w1->w_reversed == 0)
2618 continue;
2619 for (j = 1; j < w_max_used_index; j++) {
2620 if ((w_rmatrix[i][j] & WITNESS_REVERSAL) == 0 || i > j)
2621 continue;
2622
2623 mtx_lock_spin(&w_mtx);
2624 if (generation != w_generation) {
2625 mtx_unlock_spin(&w_mtx);
2626
2627 /* The graph has changed, try again. */
2628 *oldidx = 0;
2629 sbuf_clear(sb);
2630 goto restart;
2631 }
2632
2633 w2 = &w_data[j];
2634 data1 = witness_lock_order_get(w1, w2);
2635 data2 = witness_lock_order_get(w2, w1);
2636
2637 /*
2638 * Copy information locally so we can release the
2639 * spin lock.
2640 */
2641 *tmp_w2 = *w2;
2642
2643 if (data1) {
2644 stack_zero(&tmp_data1->wlod_stack);
2645 stack_copy(&data1->wlod_stack,
2646 &tmp_data1->wlod_stack);
2647 }
2648 if (data2 && data2 != data1) {
2649 stack_zero(&tmp_data2->wlod_stack);
2650 stack_copy(&data2->wlod_stack,
2651 &tmp_data2->wlod_stack);
2652 }
2653 mtx_unlock_spin(&w_mtx);
2654
2655 if (blessed(tmp_w1, tmp_w2))
2656 continue;
2657
2658 sbuf_printf(sb,
2659 "\nLock order reversal between \"%s\"(%s) and \"%s\"(%s)!\n",
2660 tmp_w1->w_name, tmp_w1->w_class->lc_name,
2661 tmp_w2->w_name, tmp_w2->w_class->lc_name);
2662 if (data1) {
2663 sbuf_printf(sb,
2664 "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2665 tmp_w1->w_name, tmp_w1->w_class->lc_name,
2666 tmp_w2->w_name, tmp_w2->w_class->lc_name);
2667 stack_sbuf_print(sb, &tmp_data1->wlod_stack);
2668 sbuf_printf(sb, "\n");
2669 }
2670 if (data2 && data2 != data1) {
2671 sbuf_printf(sb,
2672 "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n",
2673 tmp_w2->w_name, tmp_w2->w_class->lc_name,
2674 tmp_w1->w_name, tmp_w1->w_class->lc_name);
2675 stack_sbuf_print(sb, &tmp_data2->wlod_stack);
2676 sbuf_printf(sb, "\n");
2677 }
2678 }
2679 }
2680 mtx_lock_spin(&w_mtx);
2681 if (generation != w_generation) {
2682 mtx_unlock_spin(&w_mtx);
2683
2684 /*
2685 * The graph changed while we were printing stack data,
2686 * try again.
2687 */
2688 *oldidx = 0;
2689 sbuf_clear(sb);
2690 goto restart;
2691 }
2692 mtx_unlock_spin(&w_mtx);
2693
2694 /* Free temporary storage space. */
2695 free(tmp_data1, M_TEMP);
2696 free(tmp_data2, M_TEMP);
2697 free(tmp_w1, M_TEMP);
2698 free(tmp_w2, M_TEMP);
2699 }
2700
2701 static int
sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS)2702 sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS)
2703 {
2704 struct sbuf *sb;
2705 int error;
2706
2707 if (witness_watch < 1) {
2708 error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2709 return (error);
2710 }
2711 if (witness_cold) {
2712 error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2713 return (error);
2714 }
2715 error = 0;
2716 sb = sbuf_new(NULL, NULL, badstack_sbuf_size, SBUF_AUTOEXTEND);
2717 if (sb == NULL)
2718 return (ENOMEM);
2719
2720 sbuf_print_witness_badstacks(sb, &req->oldidx);
2721
2722 sbuf_finish(sb);
2723 error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1);
2724 sbuf_delete(sb);
2725
2726 return (error);
2727 }
2728
2729 #ifdef DDB
2730 static int
sbuf_db_printf_drain(void * arg __unused,const char * data,int len)2731 sbuf_db_printf_drain(void *arg __unused, const char *data, int len)
2732 {
2733
2734 return (db_printf("%.*s", len, data));
2735 }
2736
DB_SHOW_COMMAND(badstacks,db_witness_badstacks)2737 DB_SHOW_COMMAND(badstacks, db_witness_badstacks)
2738 {
2739 struct sbuf sb;
2740 char buffer[128];
2741 size_t dummy;
2742
2743 sbuf_new(&sb, buffer, sizeof(buffer), SBUF_FIXEDLEN);
2744 sbuf_set_drain(&sb, sbuf_db_printf_drain, NULL);
2745 sbuf_print_witness_badstacks(&sb, &dummy);
2746 sbuf_finish(&sb);
2747 }
2748 #endif
2749
2750 static int
sysctl_debug_witness_channel(SYSCTL_HANDLER_ARGS)2751 sysctl_debug_witness_channel(SYSCTL_HANDLER_ARGS)
2752 {
2753 static const struct {
2754 enum witness_channel channel;
2755 const char *name;
2756 } channels[] = {
2757 { WITNESS_CONSOLE, "console" },
2758 { WITNESS_LOG, "log" },
2759 { WITNESS_NONE, "none" },
2760 };
2761 char buf[16];
2762 u_int i;
2763 int error;
2764
2765 buf[0] = '\0';
2766 for (i = 0; i < nitems(channels); i++)
2767 if (witness_channel == channels[i].channel) {
2768 snprintf(buf, sizeof(buf), "%s", channels[i].name);
2769 break;
2770 }
2771
2772 error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
2773 if (error != 0 || req->newptr == NULL)
2774 return (error);
2775
2776 error = EINVAL;
2777 for (i = 0; i < nitems(channels); i++)
2778 if (strcmp(channels[i].name, buf) == 0) {
2779 witness_channel = channels[i].channel;
2780 error = 0;
2781 break;
2782 }
2783 return (error);
2784 }
2785
2786 static int
sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS)2787 sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS)
2788 {
2789 struct witness *w;
2790 struct sbuf *sb;
2791 int error;
2792
2793 #ifdef __i386__
2794 error = SYSCTL_OUT(req, w_notallowed, sizeof(w_notallowed));
2795 return (error);
2796 #endif
2797
2798 if (witness_watch < 1) {
2799 error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning));
2800 return (error);
2801 }
2802 if (witness_cold) {
2803 error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold));
2804 return (error);
2805 }
2806 error = 0;
2807
2808 error = sysctl_wire_old_buffer(req, 0);
2809 if (error != 0)
2810 return (error);
2811 sb = sbuf_new_for_sysctl(NULL, NULL, FULLGRAPH_SBUF_SIZE, req);
2812 if (sb == NULL)
2813 return (ENOMEM);
2814 sbuf_printf(sb, "\n");
2815
2816 mtx_lock_spin(&w_mtx);
2817 STAILQ_FOREACH(w, &w_all, w_list)
2818 w->w_displayed = 0;
2819 STAILQ_FOREACH(w, &w_all, w_list)
2820 witness_add_fullgraph(sb, w);
2821 mtx_unlock_spin(&w_mtx);
2822
2823 /*
2824 * Close the sbuf and return to userland.
2825 */
2826 error = sbuf_finish(sb);
2827 sbuf_delete(sb);
2828
2829 return (error);
2830 }
2831
2832 static int
sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS)2833 sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS)
2834 {
2835 int error, value;
2836
2837 value = witness_watch;
2838 error = sysctl_handle_int(oidp, &value, 0, req);
2839 if (error != 0 || req->newptr == NULL)
2840 return (error);
2841 if (value > 1 || value < -1 ||
2842 (witness_watch == -1 && value != witness_watch))
2843 return (EINVAL);
2844 witness_watch = value;
2845 return (0);
2846 }
2847
2848 static void
witness_add_fullgraph(struct sbuf * sb,struct witness * w)2849 witness_add_fullgraph(struct sbuf *sb, struct witness *w)
2850 {
2851 int i;
2852
2853 if (w->w_displayed != 0 || (w->w_file == NULL && w->w_line == 0))
2854 return;
2855 w->w_displayed = 1;
2856
2857 WITNESS_INDEX_ASSERT(w->w_index);
2858 for (i = 1; i <= w_max_used_index; i++) {
2859 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT) {
2860 sbuf_printf(sb, "\"%s\",\"%s\"\n", w->w_name,
2861 w_data[i].w_name);
2862 witness_add_fullgraph(sb, &w_data[i]);
2863 }
2864 }
2865 }
2866
2867 /*
2868 * A simple hash function. Takes a key pointer and a key size. If size == 0,
2869 * interprets the key as a string and reads until the null
2870 * terminator. Otherwise, reads the first size bytes. Returns an unsigned 32-bit
2871 * hash value computed from the key.
2872 */
2873 static uint32_t
witness_hash_djb2(const uint8_t * key,uint32_t size)2874 witness_hash_djb2(const uint8_t *key, uint32_t size)
2875 {
2876 unsigned int hash = 5381;
2877 int i;
2878
2879 /* hash = hash * 33 + key[i] */
2880 if (size)
2881 for (i = 0; i < size; i++)
2882 hash = ((hash << 5) + hash) + (unsigned int)key[i];
2883 else
2884 for (i = 0; key[i] != 0; i++)
2885 hash = ((hash << 5) + hash) + (unsigned int)key[i];
2886
2887 return (hash);
2888 }
2889
2890
2891 /*
2892 * Initializes the two witness hash tables. Called exactly once from
2893 * witness_initialize().
2894 */
2895 static void
witness_init_hash_tables(void)2896 witness_init_hash_tables(void)
2897 {
2898 int i;
2899
2900 MPASS(witness_cold);
2901
2902 /* Initialize the hash tables. */
2903 for (i = 0; i < WITNESS_HASH_SIZE; i++)
2904 w_hash.wh_array[i] = NULL;
2905
2906 w_hash.wh_size = WITNESS_HASH_SIZE;
2907 w_hash.wh_count = 0;
2908
2909 /* Initialize the lock order data hash. */
2910 w_lofree = NULL;
2911 for (i = 0; i < WITNESS_LO_DATA_COUNT; i++) {
2912 memset(&w_lodata[i], 0, sizeof(w_lodata[i]));
2913 w_lodata[i].wlod_next = w_lofree;
2914 w_lofree = &w_lodata[i];
2915 }
2916 w_lohash.wloh_size = WITNESS_LO_HASH_SIZE;
2917 w_lohash.wloh_count = 0;
2918 for (i = 0; i < WITNESS_LO_HASH_SIZE; i++)
2919 w_lohash.wloh_array[i] = NULL;
2920 }
2921
2922 static struct witness *
witness_hash_get(const char * key)2923 witness_hash_get(const char *key)
2924 {
2925 struct witness *w;
2926 uint32_t hash;
2927
2928 MPASS(key != NULL);
2929 if (witness_cold == 0)
2930 mtx_assert(&w_mtx, MA_OWNED);
2931 hash = witness_hash_djb2(key, 0) % w_hash.wh_size;
2932 w = w_hash.wh_array[hash];
2933 while (w != NULL) {
2934 if (strcmp(w->w_name, key) == 0)
2935 goto out;
2936 w = w->w_hash_next;
2937 }
2938
2939 out:
2940 return (w);
2941 }
2942
2943 static void
witness_hash_put(struct witness * w)2944 witness_hash_put(struct witness *w)
2945 {
2946 uint32_t hash;
2947
2948 MPASS(w != NULL);
2949 MPASS(w->w_name != NULL);
2950 if (witness_cold == 0)
2951 mtx_assert(&w_mtx, MA_OWNED);
2952 KASSERT(witness_hash_get(w->w_name) == NULL,
2953 ("%s: trying to add a hash entry that already exists!", __func__));
2954 KASSERT(w->w_hash_next == NULL,
2955 ("%s: w->w_hash_next != NULL", __func__));
2956
2957 hash = witness_hash_djb2(w->w_name, 0) % w_hash.wh_size;
2958 w->w_hash_next = w_hash.wh_array[hash];
2959 w_hash.wh_array[hash] = w;
2960 w_hash.wh_count++;
2961 }
2962
2963
2964 static struct witness_lock_order_data *
witness_lock_order_get(struct witness * parent,struct witness * child)2965 witness_lock_order_get(struct witness *parent, struct witness *child)
2966 {
2967 struct witness_lock_order_data *data = NULL;
2968 struct witness_lock_order_key key;
2969 unsigned int hash;
2970
2971 MPASS(parent != NULL && child != NULL);
2972 key.from = parent->w_index;
2973 key.to = child->w_index;
2974 WITNESS_INDEX_ASSERT(key.from);
2975 WITNESS_INDEX_ASSERT(key.to);
2976 if ((w_rmatrix[parent->w_index][child->w_index]
2977 & WITNESS_LOCK_ORDER_KNOWN) == 0)
2978 goto out;
2979
2980 hash = witness_hash_djb2((const char*)&key,
2981 sizeof(key)) % w_lohash.wloh_size;
2982 data = w_lohash.wloh_array[hash];
2983 while (data != NULL) {
2984 if (witness_lock_order_key_equal(&data->wlod_key, &key))
2985 break;
2986 data = data->wlod_next;
2987 }
2988
2989 out:
2990 return (data);
2991 }
2992
2993 /*
2994 * Verify that parent and child have a known relationship, are not the same,
2995 * and child is actually a child of parent. This is done without w_mtx
2996 * to avoid contention in the common case.
2997 */
2998 static int
witness_lock_order_check(struct witness * parent,struct witness * child)2999 witness_lock_order_check(struct witness *parent, struct witness *child)
3000 {
3001
3002 if (parent != child &&
3003 w_rmatrix[parent->w_index][child->w_index]
3004 & WITNESS_LOCK_ORDER_KNOWN &&
3005 isitmychild(parent, child))
3006 return (1);
3007
3008 return (0);
3009 }
3010
3011 static int
witness_lock_order_add(struct witness * parent,struct witness * child)3012 witness_lock_order_add(struct witness *parent, struct witness *child)
3013 {
3014 struct witness_lock_order_data *data = NULL;
3015 struct witness_lock_order_key key;
3016 unsigned int hash;
3017
3018 MPASS(parent != NULL && child != NULL);
3019 key.from = parent->w_index;
3020 key.to = child->w_index;
3021 WITNESS_INDEX_ASSERT(key.from);
3022 WITNESS_INDEX_ASSERT(key.to);
3023 if (w_rmatrix[parent->w_index][child->w_index]
3024 & WITNESS_LOCK_ORDER_KNOWN)
3025 return (1);
3026
3027 hash = witness_hash_djb2((const char*)&key,
3028 sizeof(key)) % w_lohash.wloh_size;
3029 w_rmatrix[parent->w_index][child->w_index] |= WITNESS_LOCK_ORDER_KNOWN;
3030 data = w_lofree;
3031 if (data == NULL)
3032 return (0);
3033 w_lofree = data->wlod_next;
3034 data->wlod_next = w_lohash.wloh_array[hash];
3035 data->wlod_key = key;
3036 w_lohash.wloh_array[hash] = data;
3037 w_lohash.wloh_count++;
3038 stack_zero(&data->wlod_stack);
3039 stack_save(&data->wlod_stack);
3040 return (1);
3041 }
3042
3043 /* Call this whenever the structure of the witness graph changes. */
3044 static void
witness_increment_graph_generation(void)3045 witness_increment_graph_generation(void)
3046 {
3047
3048 if (witness_cold == 0)
3049 mtx_assert(&w_mtx, MA_OWNED);
3050 w_generation++;
3051 }
3052
3053 static int
witness_output_drain(void * arg __unused,const char * data,int len)3054 witness_output_drain(void *arg __unused, const char *data, int len)
3055 {
3056
3057 witness_output("%.*s", len, data);
3058 return (len);
3059 }
3060
3061 static void
witness_debugger(int cond,const char * msg)3062 witness_debugger(int cond, const char *msg)
3063 {
3064 char buf[32];
3065 struct sbuf sb;
3066 struct stack st;
3067
3068 if (!cond)
3069 return;
3070
3071 if (witness_trace) {
3072 sbuf_new(&sb, buf, sizeof(buf), SBUF_FIXEDLEN);
3073 sbuf_set_drain(&sb, witness_output_drain, NULL);
3074
3075 stack_zero(&st);
3076 stack_save(&st);
3077 witness_output("stack backtrace:\n");
3078 stack_sbuf_print_ddb(&sb, &st);
3079
3080 sbuf_finish(&sb);
3081 }
3082
3083 #ifdef KDB
3084 if (witness_kdb)
3085 kdb_enter(KDB_WHY_WITNESS, msg);
3086 #endif
3087 }
3088