1 /*- 2 * SPDX-License-Identifier: BSD-3-Clause 3 * 4 * Copyright (c) 2008 Isilon Systems, Inc. 5 * Copyright (c) 2008 Ilya Maykov <ivmaykov@gmail.com> 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 #include "opt_ddb.h" 91 #include "opt_hwpmc_hooks.h" 92 #include "opt_stack.h" 93 #include "opt_witness.h" 94 95 #include <sys/param.h> 96 #include <sys/bus.h> 97 #include <sys/kdb.h> 98 #include <sys/kernel.h> 99 #include <sys/ktr.h> 100 #include <sys/lock.h> 101 #include <sys/malloc.h> 102 #include <sys/mutex.h> 103 #include <sys/priv.h> 104 #include <sys/proc.h> 105 #include <sys/sbuf.h> 106 #include <sys/sched.h> 107 #include <sys/stack.h> 108 #include <sys/stdarg.h> 109 #include <sys/sysctl.h> 110 #include <sys/syslog.h> 111 #include <sys/systm.h> 112 113 #ifdef DDB 114 #include <ddb/ddb.h> 115 #endif 116 117 #if !defined(DDB) && !defined(STACK) 118 #error "DDB or STACK options are required for WITNESS" 119 #endif 120 121 /* Note that these traces do not work with KTR_ALQ. */ 122 #if 0 123 #define KTR_WITNESS KTR_SUBSYS 124 #else 125 #define KTR_WITNESS 0 126 #endif 127 128 #define LI_RECURSEMASK 0x0000ffff /* Recursion depth of lock instance. */ 129 #define LI_EXCLUSIVE 0x00010000 /* Exclusive lock instance. */ 130 #define LI_NORELEASE 0x00020000 /* Lock not allowed to be released. */ 131 #define LI_SLEEPABLE 0x00040000 /* Lock may be held while sleeping. */ 132 133 #ifndef WITNESS_COUNT 134 #define WITNESS_COUNT 1536 135 #endif 136 #define WITNESS_HASH_SIZE 251 /* Prime, gives load factor < 2 */ 137 #define WITNESS_PENDLIST (512 + (MAXCPU * 4)) 138 139 /* Allocate 256 KB of stack data space */ 140 #define WITNESS_LO_DATA_COUNT 2048 141 142 /* Prime, gives load factor of ~2 at full load */ 143 #define WITNESS_LO_HASH_SIZE 1021 144 145 /* 146 * XXX: This is somewhat bogus, as we assume here that at most 2048 threads 147 * will hold LOCK_NCHILDREN locks. We handle failure ok, and we should 148 * probably be safe for the most part, but it's still a SWAG. 149 */ 150 #define LOCK_NCHILDREN 5 151 #define LOCK_CHILDCOUNT 2048 152 153 #define MAX_W_NAME 64 154 155 #define FULLGRAPH_SBUF_SIZE 512 156 157 /* 158 * These flags go in the witness relationship matrix and describe the 159 * relationship between any two struct witness objects. 160 */ 161 #define WITNESS_UNRELATED 0x00 /* No lock order relation. */ 162 #define WITNESS_PARENT 0x01 /* Parent, aka direct ancestor. */ 163 #define WITNESS_ANCESTOR 0x02 /* Direct or indirect ancestor. */ 164 #define WITNESS_CHILD 0x04 /* Child, aka direct descendant. */ 165 #define WITNESS_DESCENDANT 0x08 /* Direct or indirect descendant. */ 166 #define WITNESS_ANCESTOR_MASK (WITNESS_PARENT | WITNESS_ANCESTOR) 167 #define WITNESS_DESCENDANT_MASK (WITNESS_CHILD | WITNESS_DESCENDANT) 168 #define WITNESS_RELATED_MASK (WITNESS_ANCESTOR_MASK | WITNESS_DESCENDANT_MASK) 169 #define WITNESS_REVERSAL 0x10 /* A lock order reversal has been observed. */ 170 #define WITNESS_RESERVED1 0x20 /* Unused flag, reserved. */ 171 #define WITNESS_RESERVED2 0x40 /* Unused flag, reserved. */ 172 #define WITNESS_LOCK_ORDER_KNOWN 0x80 /* This lock order is known. */ 173 174 /* Descendant to ancestor flags */ 175 #define WITNESS_DTOA(x) (((x) & WITNESS_RELATED_MASK) >> 2) 176 177 /* Ancestor to descendant flags */ 178 #define WITNESS_ATOD(x) (((x) & WITNESS_RELATED_MASK) << 2) 179 180 #define WITNESS_INDEX_ASSERT(i) \ 181 MPASS((i) > 0 && (i) <= w_max_used_index && (i) < witness_count) 182 183 static MALLOC_DEFINE(M_WITNESS, "Witness", "Witness"); 184 185 /* 186 * Lock instances. A lock instance is the data associated with a lock while 187 * it is held by witness. For example, a lock instance will hold the 188 * recursion count of a lock. Lock instances are held in lists. Spin locks 189 * are held in a per-cpu list while sleep locks are held in per-thread list. 190 */ 191 struct lock_instance { 192 struct lock_object *li_lock; 193 const char *li_file; 194 int li_line; 195 u_int li_flags; 196 }; 197 198 /* 199 * A simple list type used to build the list of locks held by a thread 200 * or CPU. We can't simply embed the list in struct lock_object since a 201 * lock may be held by more than one thread if it is a shared lock. Locks 202 * are added to the head of the list, so we fill up each list entry from 203 * "the back" logically. To ease some of the arithmetic, we actually fill 204 * in each list entry the normal way (children[0] then children[1], etc.) but 205 * when we traverse the list we read children[count-1] as the first entry 206 * down to children[0] as the final entry. 207 */ 208 struct lock_list_entry { 209 struct lock_list_entry *ll_next; 210 struct lock_instance ll_children[LOCK_NCHILDREN]; 211 u_int ll_count; 212 }; 213 214 /* 215 * The main witness structure. One of these per named lock type in the system 216 * (for example, "vnode interlock"). 217 */ 218 struct witness { 219 char w_name[MAX_W_NAME]; 220 uint32_t w_index; /* Index in the relationship matrix */ 221 struct lock_class *w_class; 222 STAILQ_ENTRY(witness) w_list; /* List of all witnesses. */ 223 STAILQ_ENTRY(witness) w_typelist; /* Witnesses of a type. */ 224 struct witness *w_hash_next; /* Linked list in hash buckets. */ 225 const char *w_file; /* File where last acquired */ 226 uint32_t w_line; /* Line where last acquired */ 227 uint32_t w_refcount; 228 uint16_t w_num_ancestors; /* direct/indirect ancestor count */ 229 uint16_t w_num_descendants; /* direct/indirect descendant count */ 230 int16_t w_ddb_level; 231 unsigned w_displayed:1; 232 unsigned w_reversed:1; 233 }; 234 235 STAILQ_HEAD(witness_list, witness); 236 237 /* 238 * The witness hash table. Keys are witness names (const char *), elements are 239 * witness objects (struct witness *). 240 */ 241 struct witness_hash { 242 struct witness *wh_array[WITNESS_HASH_SIZE]; 243 uint32_t wh_size; 244 uint32_t wh_count; 245 }; 246 247 /* 248 * Key type for the lock order data hash table. 249 */ 250 struct witness_lock_order_key { 251 uint16_t from; 252 uint16_t to; 253 }; 254 255 struct witness_lock_order_data { 256 struct stack wlod_stack; 257 struct witness_lock_order_key wlod_key; 258 struct witness_lock_order_data *wlod_next; 259 }; 260 261 /* 262 * The witness lock order data hash table. Keys are witness index tuples 263 * (struct witness_lock_order_key), elements are lock order data objects 264 * (struct witness_lock_order_data). 265 */ 266 struct witness_lock_order_hash { 267 struct witness_lock_order_data *wloh_array[WITNESS_LO_HASH_SIZE]; 268 u_int wloh_size; 269 u_int wloh_count; 270 }; 271 272 struct witness_blessed { 273 const char *b_lock1; 274 const char *b_lock2; 275 }; 276 277 struct witness_pendhelp { 278 const char *wh_type; 279 struct lock_object *wh_lock; 280 }; 281 282 struct witness_order_list_entry { 283 const char *w_name; 284 struct lock_class *w_class; 285 }; 286 287 /* 288 * Returns 0 if one of the locks is a spin lock and the other is not. 289 * Returns 1 otherwise. 290 */ 291 static __inline int 292 witness_lock_type_equal(struct witness *w1, struct witness *w2) 293 { 294 return ((w1->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) == 295 (w2->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK))); 296 } 297 298 static __inline int 299 witness_lock_order_key_equal(const struct witness_lock_order_key *a, 300 const struct witness_lock_order_key *b) 301 { 302 return (a->from == b->from && a->to == b->to); 303 } 304 305 static int _isitmyx(struct witness *w1, struct witness *w2, int rmask, 306 const char *fname); 307 static void adopt(struct witness *parent, struct witness *child); 308 static int blessed(struct witness *, struct witness *); 309 static void depart(struct witness *w); 310 static struct witness *enroll(const char *description, 311 struct lock_class *lock_class); 312 static struct lock_instance *find_instance(struct lock_list_entry *list, 313 const struct lock_object *lock); 314 static int isitmychild(struct witness *parent, struct witness *child); 315 static int isitmydescendant(struct witness *parent, struct witness *child); 316 static void itismychild(struct witness *parent, struct witness *child); 317 static int sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS); 318 static int sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS); 319 static int sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS); 320 static int sysctl_debug_witness_channel(SYSCTL_HANDLER_ARGS); 321 static void witness_add_fullgraph(struct sbuf *sb, struct witness *parent); 322 #ifdef DDB 323 static void witness_ddb_compute_levels(void); 324 static void witness_ddb_display(int(*)(const char *fmt, ...)); 325 static void witness_ddb_display_descendants(int(*)(const char *fmt, ...), 326 struct witness *, int indent); 327 static void witness_ddb_display_list(int(*prnt)(const char *fmt, ...), 328 struct witness_list *list); 329 static void witness_ddb_level_descendants(struct witness *parent, int l); 330 static void witness_ddb_list(struct thread *td); 331 #endif 332 static void witness_enter_debugger(const char *msg); 333 static void witness_debugger(int cond, const char *msg); 334 static void witness_free(struct witness *m); 335 static struct witness *witness_get(void); 336 static uint32_t witness_hash_djb2(const uint8_t *key, uint32_t size); 337 static struct witness *witness_hash_get(const char *key); 338 static void witness_hash_put(struct witness *w); 339 static void witness_init_hash_tables(void); 340 static void witness_increment_graph_generation(void); 341 static void witness_lock_list_free(struct lock_list_entry *lle); 342 static struct lock_list_entry *witness_lock_list_get(void); 343 static int witness_lock_order_add(struct witness *parent, 344 struct witness *child); 345 static int witness_lock_order_check(struct witness *parent, 346 struct witness *child); 347 static struct witness_lock_order_data *witness_lock_order_get( 348 struct witness *parent, 349 struct witness *child); 350 static void witness_list_lock(struct lock_instance *instance, 351 int (*prnt)(const char *fmt, ...)); 352 static int witness_output(const char *fmt, ...) __printflike(1, 2); 353 static int witness_output_drain(void *arg __unused, const char *data, 354 int len); 355 static int witness_voutput(const char *fmt, va_list ap) __printflike(1, 0); 356 static void witness_setflag(struct lock_object *lock, int flag, int set); 357 358 FEATURE(witness, "kernel has witness(9) support"); 359 360 static SYSCTL_NODE(_debug, OID_AUTO, witness, CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, 361 "Witness Locking"); 362 363 /* 364 * If set to 0, lock order checking is disabled. If set to -1, 365 * witness is completely disabled. Otherwise witness performs full 366 * lock order checking for all locks. At runtime, lock order checking 367 * may be toggled. However, witness cannot be reenabled once it is 368 * completely disabled. 369 */ 370 static int witness_watch = 1; 371 SYSCTL_PROC(_debug_witness, OID_AUTO, watch, 372 CTLFLAG_RWTUN | CTLTYPE_INT | CTLFLAG_MPSAFE, NULL, 0, 373 sysctl_debug_witness_watch, "I", 374 "witness is watching lock operations"); 375 376 #ifdef KDB 377 /* 378 * When KDB is enabled and witness_kdb is 1, it will cause the system 379 * to drop into kdebug() when: 380 * - a lock hierarchy violation occurs 381 * - locks are held when going to sleep. 382 */ 383 #ifdef WITNESS_KDB 384 int witness_kdb = 1; 385 #else 386 int witness_kdb = 0; 387 #endif 388 SYSCTL_INT(_debug_witness, OID_AUTO, kdb, CTLFLAG_RWTUN, &witness_kdb, 0, ""); 389 #endif /* KDB */ 390 391 #if defined(DDB) || defined(KDB) 392 /* 393 * When DDB or KDB is enabled and witness_trace is 1, it will cause the system 394 * to print a stack trace: 395 * - a lock hierarchy violation occurs 396 * - locks are held when going to sleep. 397 */ 398 int witness_trace = 1; 399 SYSCTL_INT(_debug_witness, OID_AUTO, trace, CTLFLAG_RWTUN, &witness_trace, 0, ""); 400 #endif /* DDB || KDB */ 401 402 #ifdef WITNESS_SKIPSPIN 403 int witness_skipspin = 1; 404 #else 405 int witness_skipspin = 0; 406 #endif 407 SYSCTL_INT(_debug_witness, OID_AUTO, skipspin, CTLFLAG_RDTUN, &witness_skipspin, 0, ""); 408 409 int badstack_sbuf_size; 410 411 int witness_count = WITNESS_COUNT; 412 SYSCTL_INT(_debug_witness, OID_AUTO, witness_count, CTLFLAG_RDTUN, 413 &witness_count, 0, ""); 414 415 /* 416 * Output channel for witness messages. By default we print to the console. 417 */ 418 enum witness_channel { 419 WITNESS_CONSOLE, 420 WITNESS_LOG, 421 WITNESS_NONE, 422 }; 423 424 static enum witness_channel witness_channel = WITNESS_CONSOLE; 425 SYSCTL_PROC(_debug_witness, OID_AUTO, output_channel, 426 CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_MPSAFE, NULL, 0, 427 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, 434 CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0, 435 sysctl_debug_witness_fullgraph, "A", 436 "Show locks relation graphs"); 437 438 /* 439 * Call this to print out the witness faulty stacks. 440 */ 441 SYSCTL_PROC(_debug_witness, OID_AUTO, badstacks, 442 CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0, 443 sysctl_debug_witness_badstacks, "A", 444 "Show bad witness stacks"); 445 446 static struct mtx w_mtx; 447 448 /* w_list */ 449 static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free); 450 static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all); 451 452 /* w_typelist */ 453 static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin); 454 static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep); 455 456 /* lock list */ 457 static struct lock_list_entry *w_lock_list_free = NULL; 458 static struct witness_pendhelp pending_locks[WITNESS_PENDLIST]; 459 static u_int pending_cnt; 460 461 static int w_free_cnt, w_spin_cnt, w_sleep_cnt; 462 SYSCTL_INT(_debug_witness, OID_AUTO, free_cnt, CTLFLAG_RD, &w_free_cnt, 0, ""); 463 SYSCTL_INT(_debug_witness, OID_AUTO, spin_cnt, CTLFLAG_RD, &w_spin_cnt, 0, ""); 464 SYSCTL_INT(_debug_witness, OID_AUTO, sleep_cnt, CTLFLAG_RD, &w_sleep_cnt, 0, 465 ""); 466 467 static struct witness *w_data; 468 static uint8_t **w_rmatrix; 469 static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT]; 470 static struct witness_hash w_hash; /* The witness hash table. */ 471 472 /* The lock order data hash */ 473 static struct witness_lock_order_data w_lodata[WITNESS_LO_DATA_COUNT]; 474 static struct witness_lock_order_data *w_lofree = NULL; 475 static struct witness_lock_order_hash w_lohash; 476 static int w_max_used_index = 0; 477 static unsigned int w_generation = 0; 478 static const char w_notrunning[] = "Witness not running\n"; 479 static const char w_stillcold[] = "Witness is still cold\n"; 480 #ifdef __i386__ 481 static const char w_notallowed[] = "The sysctl is disabled on the arch\n"; 482 #endif 483 484 static struct witness_order_list_entry order_lists[] = { 485 /* 486 * sx locks 487 */ 488 { "proctree", &lock_class_sx }, 489 { "allproc", &lock_class_sx }, 490 { "allprison", &lock_class_sx }, 491 { NULL, NULL }, 492 /* 493 * Various mutexes 494 */ 495 { "Giant", &lock_class_mtx_sleep }, 496 { "pipe mutex", &lock_class_mtx_sleep }, 497 { "sigio lock", &lock_class_mtx_sleep }, 498 { "process group", &lock_class_mtx_sleep }, 499 #ifdef HWPMC_HOOKS 500 { "pmc-sleep", &lock_class_mtx_sleep }, 501 #endif 502 { "process lock", &lock_class_mtx_sleep }, 503 { "session", &lock_class_mtx_sleep }, 504 { "uidinfo hash", &lock_class_rw }, 505 { "time lock", &lock_class_mtx_sleep }, 506 { NULL, NULL }, 507 /* 508 * umtx 509 */ 510 { "umtx lock", &lock_class_mtx_sleep }, 511 { NULL, NULL }, 512 /* 513 * Sockets 514 */ 515 { "accept", &lock_class_mtx_sleep }, 516 { "so_snd", &lock_class_mtx_sleep }, 517 { "so_rcv", &lock_class_mtx_sleep }, 518 { "sellck", &lock_class_mtx_sleep }, 519 { NULL, NULL }, 520 /* 521 * Routing 522 */ 523 { "so_rcv", &lock_class_mtx_sleep }, 524 { "radix node head", &lock_class_rm }, 525 { "ifaddr", &lock_class_mtx_sleep }, 526 { NULL, NULL }, 527 /* 528 * IPv4 multicast: 529 * protocol locks before interface locks, after UDP locks. 530 */ 531 { "in_multi_sx", &lock_class_sx }, 532 { "udpinp", &lock_class_rw }, 533 { "in_multi_list_mtx", &lock_class_mtx_sleep }, 534 { "igmp_mtx", &lock_class_mtx_sleep }, 535 { "if_addr_lock", &lock_class_mtx_sleep }, 536 { NULL, NULL }, 537 /* 538 * IPv6 multicast: 539 * protocol locks before interface locks, after UDP locks. 540 */ 541 { "in6_multi_sx", &lock_class_sx }, 542 { "udpinp", &lock_class_rw }, 543 { "in6_multi_list_mtx", &lock_class_mtx_sleep }, 544 { "mld_mtx", &lock_class_mtx_sleep }, 545 { "if_addr_lock", &lock_class_mtx_sleep }, 546 { NULL, NULL }, 547 /* 548 * UNIX Domain Sockets 549 */ 550 { "unp_link_rwlock", &lock_class_rw }, 551 { "unp_list_lock", &lock_class_mtx_sleep }, 552 { "unp", &lock_class_mtx_sleep }, 553 { "so_snd", &lock_class_mtx_sleep }, 554 { NULL, NULL }, 555 /* 556 * UDP/IP 557 */ 558 { "udpinp", &lock_class_rw }, 559 { "udp", &lock_class_mtx_sleep }, 560 { "so_snd", &lock_class_mtx_sleep }, 561 { NULL, NULL }, 562 /* 563 * TCP/IP 564 */ 565 { "tcpinp", &lock_class_rw }, 566 { "tcp", &lock_class_mtx_sleep }, 567 { "so_snd", &lock_class_mtx_sleep }, 568 { NULL, NULL }, 569 /* 570 * BPF 571 */ 572 { "bpf global lock", &lock_class_sx }, 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 { "devthrd", &lock_class_mtx_sleep }, 606 { NULL, NULL }, 607 /* 608 * VM 609 */ 610 { "vm map (user)", &lock_class_sx }, 611 { "vm object", &lock_class_rw }, 612 { "vm page", &lock_class_mtx_sleep }, 613 { "pmap pv global", &lock_class_rw }, 614 { "pmap", &lock_class_mtx_sleep }, 615 { "pmap pv list", &lock_class_rw }, 616 { "vm page free queue", &lock_class_mtx_sleep }, 617 { "vm pagequeue", &lock_class_mtx_sleep }, 618 { NULL, NULL }, 619 /* 620 * kqueue/VFS interaction 621 */ 622 { "kqueue", &lock_class_mtx_sleep }, 623 { "struct mount mtx", &lock_class_mtx_sleep }, 624 { "vnode interlock", &lock_class_mtx_sleep }, 625 { NULL, NULL }, 626 /* 627 * VFS namecache 628 */ 629 { "ncvn", &lock_class_mtx_sleep }, 630 { "ncbuc", &lock_class_mtx_sleep }, 631 { "vnode interlock", &lock_class_mtx_sleep }, 632 { "ncneg", &lock_class_mtx_sleep }, 633 { NULL, NULL }, 634 /* 635 * ZFS locking 636 */ 637 { "dn->dn_mtx", &lock_class_sx }, 638 { "dr->dt.di.dr_mtx", &lock_class_sx }, 639 { "db->db_mtx", &lock_class_sx }, 640 { NULL, NULL }, 641 /* 642 * TCP log locks 643 */ 644 { "TCP ID tree", &lock_class_rw }, 645 { "tcp log id bucket", &lock_class_mtx_sleep }, 646 { "tcpinp", &lock_class_rw }, 647 { "TCP log expireq", &lock_class_mtx_sleep }, 648 { NULL, NULL }, 649 /* 650 * spin locks 651 */ 652 #ifdef SMP 653 { "ap boot", &lock_class_mtx_spin }, 654 #endif 655 { "rm.mutex_mtx", &lock_class_mtx_spin }, 656 #ifdef __i386__ 657 { "cy", &lock_class_mtx_spin }, 658 #endif 659 { "scc_hwmtx", &lock_class_mtx_spin }, 660 { "uart_hwmtx", &lock_class_mtx_spin }, 661 { "fast_taskqueue", &lock_class_mtx_spin }, 662 { "intr table", &lock_class_mtx_spin }, 663 { "process slock", &lock_class_mtx_spin }, 664 { "syscons video lock", &lock_class_mtx_spin }, 665 { "sleepq chain", &lock_class_mtx_spin }, 666 { "rm_spinlock", &lock_class_mtx_spin }, 667 { "turnstile chain", &lock_class_mtx_spin }, 668 { "turnstile lock", &lock_class_mtx_spin }, 669 { "sched lock", &lock_class_mtx_spin }, 670 { "td_contested", &lock_class_mtx_spin }, 671 { "callout", &lock_class_mtx_spin }, 672 { "entropy harvest mutex", &lock_class_mtx_spin }, 673 #ifdef SMP 674 { "smp rendezvous", &lock_class_mtx_spin }, 675 #endif 676 #ifdef __powerpc__ 677 { "tlb0", &lock_class_mtx_spin }, 678 #endif 679 { NULL, NULL }, 680 { "sched lock", &lock_class_mtx_spin }, 681 #ifdef HWPMC_HOOKS 682 { "pmc-per-proc", &lock_class_mtx_spin }, 683 #endif 684 { NULL, NULL }, 685 /* 686 * leaf locks 687 */ 688 { "intrcnt", &lock_class_mtx_spin }, 689 { "icu", &lock_class_mtx_spin }, 690 #ifdef __i386__ 691 { "allpmaps", &lock_class_mtx_spin }, 692 { "descriptor tables", &lock_class_mtx_spin }, 693 #endif 694 { "clk", &lock_class_mtx_spin }, 695 { "cpuset", &lock_class_mtx_spin }, 696 { "mprof lock", &lock_class_mtx_spin }, 697 { "zombie lock", &lock_class_mtx_spin }, 698 { "ALD Queue", &lock_class_mtx_spin }, 699 #if defined(__i386__) || defined(__amd64__) 700 { "pcicfg", &lock_class_mtx_spin }, 701 { "NDIS thread lock", &lock_class_mtx_spin }, 702 #endif 703 { "tw_osl_io_lock", &lock_class_mtx_spin }, 704 { "tw_osl_q_lock", &lock_class_mtx_spin }, 705 { "tw_cl_io_lock", &lock_class_mtx_spin }, 706 { "tw_cl_intr_lock", &lock_class_mtx_spin }, 707 { "tw_cl_gen_lock", &lock_class_mtx_spin }, 708 #ifdef HWPMC_HOOKS 709 { "pmc-leaf", &lock_class_mtx_spin }, 710 #endif 711 { "blocked lock", &lock_class_mtx_spin }, 712 { NULL, NULL }, 713 { NULL, NULL } 714 }; 715 716 /* 717 * Pairs of locks which have been blessed. Witness does not complain about 718 * order problems with blessed lock pairs. Please do not add an entry to the 719 * table without an explanatory comment. 720 */ 721 static struct witness_blessed blessed_list[] = { 722 /* 723 * See the comment in ufs_dirhash.c. Basically, a vnode lock serializes 724 * both lock orders, so a deadlock cannot happen as a result of this 725 * LOR. 726 */ 727 { "dirhash", "bufwait" }, 728 729 /* 730 * A UFS vnode may be locked in vget() while a buffer belonging to the 731 * parent directory vnode is locked. 732 */ 733 { "ufs", "bufwait" }, 734 735 /* 736 * The tarfs decompression stream vnode may be locked while a 737 * buffer belonging to a tarfs data vnode is locked. 738 */ 739 { "tarfs", "bufwait" }, 740 }; 741 742 /* 743 * This global is set to 0 once it becomes safe to use the witness code. 744 */ 745 static int witness_cold = 1; 746 747 /* 748 * This global is set to 1 once the static lock orders have been enrolled 749 * so that a warning can be issued for any spin locks enrolled later. 750 */ 751 static int witness_spin_warn = 0; 752 753 /* Trim useless garbage from filenames. */ 754 static const char * 755 fixup_filename(const char *file) 756 { 757 if (file == NULL) 758 return (NULL); 759 while (strncmp(file, "../", 3) == 0) 760 file += 3; 761 return (file); 762 } 763 764 /* 765 * Calculate the size of early witness structures. 766 */ 767 int 768 witness_startup_count(void) 769 { 770 int sz; 771 772 sz = sizeof(struct witness) * witness_count; 773 sz += sizeof(*w_rmatrix) * (witness_count + 1); 774 sz += sizeof(*w_rmatrix[0]) * (witness_count + 1) * 775 (witness_count + 1); 776 777 return (sz); 778 } 779 780 /* 781 * The WITNESS-enabled diagnostic code. Note that the witness code does 782 * assume that the early boot is single-threaded at least until after this 783 * routine is completed. 784 */ 785 void 786 witness_startup(void *mem) 787 { 788 struct lock_object *lock; 789 struct witness_order_list_entry *order; 790 struct witness *w, *w1; 791 uintptr_t p; 792 int i; 793 794 p = (uintptr_t)mem; 795 w_data = (void *)p; 796 p += sizeof(struct witness) * witness_count; 797 798 w_rmatrix = (void *)p; 799 p += sizeof(*w_rmatrix) * (witness_count + 1); 800 801 for (i = 0; i < witness_count + 1; i++) { 802 w_rmatrix[i] = (void *)p; 803 p += sizeof(*w_rmatrix[i]) * (witness_count + 1); 804 } 805 badstack_sbuf_size = witness_count * 256; 806 807 /* 808 * We have to release Giant before initializing its witness 809 * structure so that WITNESS doesn't get confused. 810 */ 811 mtx_unlock(&Giant); 812 mtx_assert(&Giant, MA_NOTOWNED); 813 814 CTR1(KTR_WITNESS, "%s: initializing witness", __func__); 815 mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET | 816 MTX_NOWITNESS | MTX_NOPROFILE); 817 for (i = witness_count - 1; i >= 0; i--) { 818 w = &w_data[i]; 819 memset(w, 0, sizeof(*w)); 820 w_data[i].w_index = i; /* Witness index never changes. */ 821 witness_free(w); 822 } 823 KASSERT(STAILQ_FIRST(&w_free)->w_index == 0, 824 ("%s: Invalid list of free witness objects", __func__)); 825 826 /* Witness with index 0 is not used to aid in debugging. */ 827 STAILQ_REMOVE_HEAD(&w_free, w_list); 828 w_free_cnt--; 829 830 for (i = 0; i < witness_count; i++) { 831 memset(w_rmatrix[i], 0, sizeof(*w_rmatrix[i]) * 832 (witness_count + 1)); 833 } 834 835 for (i = 0; i < LOCK_CHILDCOUNT; i++) 836 witness_lock_list_free(&w_locklistdata[i]); 837 witness_init_hash_tables(); 838 839 /* First add in all the specified order lists. */ 840 for (order = order_lists; order->w_name != NULL; order++) { 841 w = enroll(order->w_name, order->w_class); 842 if (w == NULL) 843 continue; 844 w->w_file = "order list"; 845 for (order++; order->w_name != NULL; order++) { 846 w1 = enroll(order->w_name, order->w_class); 847 if (w1 == NULL) 848 continue; 849 w1->w_file = "order list"; 850 itismychild(w, w1); 851 w = w1; 852 } 853 } 854 witness_spin_warn = 1; 855 856 /* Iterate through all locks and add them to witness. */ 857 for (i = 0; pending_locks[i].wh_lock != NULL; i++) { 858 lock = pending_locks[i].wh_lock; 859 KASSERT(lock->lo_flags & LO_WITNESS, 860 ("%s: lock %s is on pending list but not LO_WITNESS", 861 __func__, lock->lo_name)); 862 lock->lo_witness = enroll(pending_locks[i].wh_type, 863 LOCK_CLASS(lock)); 864 } 865 866 /* Mark the witness code as being ready for use. */ 867 witness_cold = 0; 868 869 mtx_lock(&Giant); 870 } 871 872 void 873 witness_init(struct lock_object *lock, const char *type) 874 { 875 struct lock_class *class; 876 877 /* Various sanity checks. */ 878 class = LOCK_CLASS(lock); 879 if ((lock->lo_flags & LO_RECURSABLE) != 0 && 880 (class->lc_flags & LC_RECURSABLE) == 0) 881 kassert_panic("%s: lock (%s) %s can not be recursable", 882 __func__, class->lc_name, lock->lo_name); 883 if ((lock->lo_flags & LO_SLEEPABLE) != 0 && 884 (class->lc_flags & LC_SLEEPABLE) == 0) 885 kassert_panic("%s: lock (%s) %s can not be sleepable", 886 __func__, class->lc_name, lock->lo_name); 887 if ((lock->lo_flags & LO_UPGRADABLE) != 0 && 888 (class->lc_flags & LC_UPGRADABLE) == 0) 889 kassert_panic("%s: lock (%s) %s can not be upgradable", 890 __func__, class->lc_name, lock->lo_name); 891 892 /* 893 * If we shouldn't watch this lock, then just clear lo_witness. 894 * Otherwise, if witness_cold is set, then it is too early to 895 * enroll this lock, so defer it to witness_initialize() by adding 896 * it to the pending_locks list. If it is not too early, then enroll 897 * the lock now. 898 */ 899 if (witness_watch < 1 || KERNEL_PANICKED() || 900 (lock->lo_flags & LO_WITNESS) == 0) 901 lock->lo_witness = NULL; 902 else if (witness_cold) { 903 pending_locks[pending_cnt].wh_lock = lock; 904 pending_locks[pending_cnt++].wh_type = type; 905 if (pending_cnt > WITNESS_PENDLIST) 906 panic("%s: pending locks list is too small, " 907 "increase WITNESS_PENDLIST\n", 908 __func__); 909 } else 910 lock->lo_witness = enroll(type, class); 911 } 912 913 void 914 witness_destroy(struct lock_object *lock) 915 { 916 struct lock_class *class; 917 struct witness *w; 918 919 class = LOCK_CLASS(lock); 920 921 if (witness_cold) 922 panic("lock (%s) %s destroyed while witness_cold", 923 class->lc_name, lock->lo_name); 924 925 /* XXX: need to verify that no one holds the lock */ 926 if ((lock->lo_flags & LO_WITNESS) == 0 || lock->lo_witness == NULL) 927 return; 928 w = lock->lo_witness; 929 930 mtx_lock_spin(&w_mtx); 931 MPASS(w->w_refcount > 0); 932 w->w_refcount--; 933 934 if (w->w_refcount == 0) 935 depart(w); 936 mtx_unlock_spin(&w_mtx); 937 } 938 939 #ifdef DDB 940 static void 941 witness_ddb_compute_levels(void) 942 { 943 struct witness *w; 944 945 /* 946 * First clear all levels. 947 */ 948 STAILQ_FOREACH(w, &w_all, w_list) 949 w->w_ddb_level = -1; 950 951 /* 952 * Look for locks with no parents and level all their descendants. 953 */ 954 STAILQ_FOREACH(w, &w_all, w_list) { 955 /* If the witness has ancestors (is not a root), skip it. */ 956 if (w->w_num_ancestors > 0) 957 continue; 958 witness_ddb_level_descendants(w, 0); 959 } 960 } 961 962 static void 963 witness_ddb_level_descendants(struct witness *w, int l) 964 { 965 int i; 966 967 if (w->w_ddb_level >= l) 968 return; 969 970 w->w_ddb_level = l; 971 l++; 972 973 for (i = 1; i <= w_max_used_index; i++) { 974 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT) 975 witness_ddb_level_descendants(&w_data[i], l); 976 } 977 } 978 979 static void 980 witness_ddb_display_descendants(int(*prnt)(const char *fmt, ...), 981 struct witness *w, int indent) 982 { 983 int i; 984 985 for (i = 0; i < indent; i++) 986 prnt(" "); 987 prnt("%s (type: %s, depth: %d, active refs: %d)", 988 w->w_name, w->w_class->lc_name, 989 w->w_ddb_level, w->w_refcount); 990 if (w->w_displayed) { 991 prnt(" -- (already displayed)\n"); 992 return; 993 } 994 w->w_displayed = 1; 995 if (w->w_file != NULL && w->w_line != 0) 996 prnt(" -- last acquired @ %s:%d\n", fixup_filename(w->w_file), 997 w->w_line); 998 else 999 prnt(" -- never acquired\n"); 1000 indent++; 1001 WITNESS_INDEX_ASSERT(w->w_index); 1002 for (i = 1; i <= w_max_used_index; i++) { 1003 if (db_pager_quit) 1004 return; 1005 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT) 1006 witness_ddb_display_descendants(prnt, &w_data[i], 1007 indent); 1008 } 1009 } 1010 1011 static void 1012 witness_ddb_display_list(int(*prnt)(const char *fmt, ...), 1013 struct witness_list *list) 1014 { 1015 struct witness *w; 1016 1017 STAILQ_FOREACH(w, list, w_typelist) { 1018 if (w->w_file == NULL || w->w_ddb_level > 0) 1019 continue; 1020 1021 /* This lock has no anscestors - display its descendants. */ 1022 witness_ddb_display_descendants(prnt, w, 0); 1023 if (db_pager_quit) 1024 return; 1025 } 1026 } 1027 1028 static void 1029 witness_ddb_display(int(*prnt)(const char *fmt, ...)) 1030 { 1031 struct witness *w; 1032 1033 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__)); 1034 witness_ddb_compute_levels(); 1035 1036 /* Clear all the displayed flags. */ 1037 STAILQ_FOREACH(w, &w_all, w_list) 1038 w->w_displayed = 0; 1039 1040 /* 1041 * First, handle sleep locks which have been acquired at least 1042 * once. 1043 */ 1044 prnt("Sleep locks:\n"); 1045 witness_ddb_display_list(prnt, &w_sleep); 1046 if (db_pager_quit) 1047 return; 1048 1049 /* 1050 * Now do spin locks which have been acquired at least once. 1051 */ 1052 prnt("\nSpin locks:\n"); 1053 witness_ddb_display_list(prnt, &w_spin); 1054 if (db_pager_quit) 1055 return; 1056 1057 /* 1058 * Finally, any locks which have not been acquired yet. 1059 */ 1060 prnt("\nLocks which were never acquired:\n"); 1061 STAILQ_FOREACH(w, &w_all, w_list) { 1062 if (w->w_file != NULL || w->w_refcount == 0) 1063 continue; 1064 prnt("%s (type: %s, depth: %d)\n", w->w_name, 1065 w->w_class->lc_name, w->w_ddb_level); 1066 if (db_pager_quit) 1067 return; 1068 } 1069 } 1070 #endif /* DDB */ 1071 1072 int 1073 witness_defineorder(struct lock_object *lock1, struct lock_object *lock2) 1074 { 1075 if (witness_watch == -1 || KERNEL_PANICKED()) 1076 return (0); 1077 1078 /* Require locks that witness knows about. */ 1079 if (lock1 == NULL || lock1->lo_witness == NULL || lock2 == NULL || 1080 lock2->lo_witness == NULL) 1081 return (EINVAL); 1082 1083 mtx_assert(&w_mtx, MA_NOTOWNED); 1084 mtx_lock_spin(&w_mtx); 1085 1086 /* 1087 * If we already have either an explicit or implied lock order that 1088 * is the other way around, then return an error. 1089 */ 1090 if (witness_watch && 1091 isitmydescendant(lock2->lo_witness, lock1->lo_witness)) { 1092 mtx_unlock_spin(&w_mtx); 1093 return (EDOOFUS); 1094 } 1095 1096 /* Try to add the new order. */ 1097 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__, 1098 lock2->lo_witness->w_name, lock1->lo_witness->w_name); 1099 itismychild(lock1->lo_witness, lock2->lo_witness); 1100 mtx_unlock_spin(&w_mtx); 1101 return (0); 1102 } 1103 1104 void 1105 witness_checkorder(struct lock_object *lock, int flags, const char *file, 1106 int line, struct lock_object *interlock) 1107 { 1108 struct lock_list_entry *lock_list, *lle; 1109 struct lock_instance *lock1, *lock2, *plock; 1110 struct lock_class *class, *iclass; 1111 struct witness *w, *w1; 1112 struct thread *td; 1113 int i, j; 1114 1115 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL || 1116 KERNEL_PANICKED()) 1117 return; 1118 1119 w = lock->lo_witness; 1120 class = LOCK_CLASS(lock); 1121 td = curthread; 1122 1123 if (class->lc_flags & LC_SLEEPLOCK) { 1124 /* 1125 * Since spin locks include a critical section, this check 1126 * implicitly enforces a lock order of all sleep locks before 1127 * all spin locks. 1128 */ 1129 if (td->td_critnest != 0 && !kdb_active) 1130 kassert_panic("acquiring blockable sleep lock with " 1131 "spinlock or critical section held (%s) %s @ %s:%d", 1132 class->lc_name, lock->lo_name, 1133 fixup_filename(file), line); 1134 1135 /* 1136 * If this is the first lock acquired then just return as 1137 * no order checking is needed. 1138 */ 1139 lock_list = td->td_sleeplocks; 1140 if (lock_list == NULL || lock_list->ll_count == 0) 1141 return; 1142 } else { 1143 /* 1144 * If this is the first lock, just return as no order 1145 * checking is needed. Avoid problems with thread 1146 * migration pinning the thread while checking if 1147 * spinlocks are held. If at least one spinlock is held 1148 * the thread is in a safe path and it is allowed to 1149 * unpin it. 1150 */ 1151 sched_pin(); 1152 lock_list = PCPU_GET(spinlocks); 1153 if (lock_list == NULL || lock_list->ll_count == 0) { 1154 sched_unpin(); 1155 return; 1156 } 1157 sched_unpin(); 1158 } 1159 1160 /* 1161 * Check to see if we are recursing on a lock we already own. If 1162 * so, make sure that we don't mismatch exclusive and shared lock 1163 * acquires. 1164 */ 1165 lock1 = find_instance(lock_list, lock); 1166 if (lock1 != NULL) { 1167 if ((lock1->li_flags & LI_EXCLUSIVE) != 0 && 1168 (flags & LOP_EXCLUSIVE) == 0) { 1169 witness_output("shared lock of (%s) %s @ %s:%d\n", 1170 class->lc_name, lock->lo_name, 1171 fixup_filename(file), line); 1172 witness_output("while exclusively locked from %s:%d\n", 1173 fixup_filename(lock1->li_file), lock1->li_line); 1174 kassert_panic("excl->share"); 1175 } 1176 if ((lock1->li_flags & LI_EXCLUSIVE) == 0 && 1177 (flags & LOP_EXCLUSIVE) != 0) { 1178 witness_output("exclusive lock of (%s) %s @ %s:%d\n", 1179 class->lc_name, lock->lo_name, 1180 fixup_filename(file), line); 1181 witness_output("while share locked from %s:%d\n", 1182 fixup_filename(lock1->li_file), lock1->li_line); 1183 kassert_panic("share->excl"); 1184 } 1185 return; 1186 } 1187 1188 /* Warn if the interlock is not locked exactly once. */ 1189 if (interlock != NULL) { 1190 iclass = LOCK_CLASS(interlock); 1191 lock1 = find_instance(lock_list, interlock); 1192 if (lock1 == NULL) 1193 kassert_panic("interlock (%s) %s not locked @ %s:%d", 1194 iclass->lc_name, interlock->lo_name, 1195 fixup_filename(file), line); 1196 else if ((lock1->li_flags & LI_RECURSEMASK) != 0) 1197 kassert_panic("interlock (%s) %s recursed @ %s:%d", 1198 iclass->lc_name, interlock->lo_name, 1199 fixup_filename(file), line); 1200 } 1201 1202 /* 1203 * Find the previously acquired lock, but ignore interlocks. 1204 */ 1205 plock = &lock_list->ll_children[lock_list->ll_count - 1]; 1206 if (interlock != NULL && plock->li_lock == interlock) { 1207 if (lock_list->ll_count > 1) 1208 plock = 1209 &lock_list->ll_children[lock_list->ll_count - 2]; 1210 else { 1211 lle = lock_list->ll_next; 1212 1213 /* 1214 * The interlock is the only lock we hold, so 1215 * simply return. 1216 */ 1217 if (lle == NULL) 1218 return; 1219 plock = &lle->ll_children[lle->ll_count - 1]; 1220 } 1221 } 1222 1223 /* 1224 * Try to perform most checks without a lock. If this succeeds we 1225 * can skip acquiring the lock and return success. Otherwise we redo 1226 * the check with the lock held to handle races with concurrent updates. 1227 */ 1228 w1 = plock->li_lock->lo_witness; 1229 if (witness_lock_order_check(w1, w)) 1230 return; 1231 1232 mtx_lock_spin(&w_mtx); 1233 if (witness_lock_order_check(w1, w)) { 1234 mtx_unlock_spin(&w_mtx); 1235 return; 1236 } 1237 witness_lock_order_add(w1, w); 1238 1239 /* 1240 * Check for duplicate locks of the same type. Note that we only 1241 * have to check for this on the last lock we just acquired. Any 1242 * other cases will be caught as lock order violations. 1243 */ 1244 if (w1 == w) { 1245 i = w->w_index; 1246 if (!(lock->lo_flags & LO_DUPOK) && !(flags & LOP_DUPOK) && 1247 !(w_rmatrix[i][i] & WITNESS_REVERSAL)) { 1248 w_rmatrix[i][i] |= WITNESS_REVERSAL; 1249 w->w_reversed = 1; 1250 mtx_unlock_spin(&w_mtx); 1251 witness_output( 1252 "acquiring duplicate lock of same type: \"%s\"\n", 1253 w->w_name); 1254 witness_output(" 1st %s @ %s:%d\n", plock->li_lock->lo_name, 1255 fixup_filename(plock->li_file), plock->li_line); 1256 witness_output(" 2nd %s @ %s:%d\n", lock->lo_name, 1257 fixup_filename(file), line); 1258 witness_debugger(1, __func__); 1259 } else 1260 mtx_unlock_spin(&w_mtx); 1261 return; 1262 } 1263 mtx_assert(&w_mtx, MA_OWNED); 1264 1265 /* 1266 * If we know that the lock we are acquiring comes after 1267 * the lock we most recently acquired in the lock order tree, 1268 * then there is no need for any further checks. 1269 */ 1270 if (isitmychild(w1, w)) 1271 goto out; 1272 1273 for (j = 0, lle = lock_list; lle != NULL; lle = lle->ll_next) { 1274 for (i = lle->ll_count - 1; i >= 0; i--, j++) { 1275 struct stack pstack; 1276 bool pstackv, trace; 1277 1278 MPASS(j < LOCK_CHILDCOUNT * LOCK_NCHILDREN); 1279 lock1 = &lle->ll_children[i]; 1280 1281 /* 1282 * Ignore the interlock. 1283 */ 1284 if (interlock == lock1->li_lock) 1285 continue; 1286 1287 /* 1288 * If this lock doesn't undergo witness checking, 1289 * then skip it. 1290 */ 1291 w1 = lock1->li_lock->lo_witness; 1292 if (w1 == NULL) { 1293 KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0, 1294 ("lock missing witness structure")); 1295 continue; 1296 } 1297 1298 /* 1299 * If we are locking Giant and this is a sleepable 1300 * lock, then skip it. 1301 */ 1302 if ((lock1->li_flags & LI_SLEEPABLE) != 0 && 1303 lock == &Giant.lock_object) 1304 continue; 1305 1306 /* 1307 * If we are locking a sleepable lock and this lock 1308 * is Giant, then skip it. 1309 */ 1310 if ((lock->lo_flags & LO_SLEEPABLE) != 0 && 1311 (flags & LOP_NOSLEEP) == 0 && 1312 lock1->li_lock == &Giant.lock_object) 1313 continue; 1314 1315 /* 1316 * If we are locking a sleepable lock and this lock 1317 * isn't sleepable, we want to treat it as a lock 1318 * order violation to enfore a general lock order of 1319 * sleepable locks before non-sleepable locks. 1320 */ 1321 if ((lock->lo_flags & LO_SLEEPABLE) != 0 && 1322 (flags & LOP_NOSLEEP) == 0 && 1323 (lock1->li_flags & LI_SLEEPABLE) == 0) 1324 goto reversal; 1325 1326 /* 1327 * If we are locking Giant and this is a non-sleepable 1328 * lock, then treat it as a reversal. 1329 */ 1330 if ((lock1->li_flags & LI_SLEEPABLE) == 0 && 1331 lock == &Giant.lock_object) 1332 goto reversal; 1333 1334 /* 1335 * Check the lock order hierarchy for a reveresal. 1336 */ 1337 if (!isitmydescendant(w, w1)) 1338 continue; 1339 reversal: 1340 1341 /* 1342 * We have a lock order violation, check to see if it 1343 * is allowed or has already been yelled about. 1344 */ 1345 1346 /* Bail if this violation is known */ 1347 if (w_rmatrix[w1->w_index][w->w_index] & WITNESS_REVERSAL) 1348 goto out; 1349 1350 /* Record this as a violation */ 1351 w_rmatrix[w1->w_index][w->w_index] |= WITNESS_REVERSAL; 1352 w_rmatrix[w->w_index][w1->w_index] |= WITNESS_REVERSAL; 1353 w->w_reversed = w1->w_reversed = 1; 1354 witness_increment_graph_generation(); 1355 1356 /* 1357 * If the lock order is blessed, bail before logging 1358 * anything. We don't look for other lock order 1359 * violations though, which may be a bug. 1360 */ 1361 if (blessed(w, w1)) 1362 goto out; 1363 1364 trace = atomic_load_int(&witness_trace); 1365 if (trace) { 1366 struct witness_lock_order_data *data; 1367 1368 pstackv = false; 1369 data = witness_lock_order_get(w, w1); 1370 if (data != NULL) { 1371 stack_copy(&data->wlod_stack, 1372 &pstack); 1373 pstackv = true; 1374 } 1375 } 1376 mtx_unlock_spin(&w_mtx); 1377 1378 #ifdef WITNESS_NO_VNODE 1379 /* 1380 * There are known LORs between VNODE locks. They are 1381 * not an indication of a bug. VNODE locks are flagged 1382 * as such (LO_IS_VNODE) and we don't yell if the LOR 1383 * is between 2 VNODE locks. 1384 */ 1385 if ((lock->lo_flags & LO_IS_VNODE) != 0 && 1386 (lock1->li_lock->lo_flags & LO_IS_VNODE) != 0) 1387 return; 1388 #endif 1389 1390 /* 1391 * Ok, yell about it. 1392 */ 1393 if ((lock->lo_flags & LO_SLEEPABLE) != 0 && 1394 (flags & LOP_NOSLEEP) == 0 && 1395 (lock1->li_flags & LI_SLEEPABLE) == 0) 1396 witness_output( 1397 "lock order reversal: (sleepable after non-sleepable)\n"); 1398 else if ((lock1->li_flags & LI_SLEEPABLE) == 0 1399 && lock == &Giant.lock_object) 1400 witness_output( 1401 "lock order reversal: (Giant after non-sleepable)\n"); 1402 else 1403 witness_output("lock order reversal:\n"); 1404 1405 /* 1406 * Try to locate an earlier lock with 1407 * witness w in our list. 1408 */ 1409 do { 1410 lock2 = &lle->ll_children[i]; 1411 MPASS(lock2->li_lock != NULL); 1412 if (lock2->li_lock->lo_witness == w) 1413 break; 1414 if (i == 0 && lle->ll_next != NULL) { 1415 lle = lle->ll_next; 1416 i = lle->ll_count - 1; 1417 MPASS(i >= 0 && i < LOCK_NCHILDREN); 1418 } else 1419 i--; 1420 } while (i >= 0); 1421 if (i < 0) { 1422 witness_output(" 1st %p %s (%s, %s) @ %s:%d\n", 1423 lock1->li_lock, lock1->li_lock->lo_name, 1424 w1->w_name, w1->w_class->lc_name, 1425 fixup_filename(lock1->li_file), 1426 lock1->li_line); 1427 witness_output(" 2nd %p %s (%s, %s) @ %s:%d\n", 1428 lock, lock->lo_name, w->w_name, 1429 w->w_class->lc_name, fixup_filename(file), 1430 line); 1431 } else { 1432 struct witness *w2 = lock2->li_lock->lo_witness; 1433 1434 witness_output(" 1st %p %s (%s, %s) @ %s:%d\n", 1435 lock2->li_lock, lock2->li_lock->lo_name, 1436 w2->w_name, w2->w_class->lc_name, 1437 fixup_filename(lock2->li_file), 1438 lock2->li_line); 1439 witness_output(" 2nd %p %s (%s, %s) @ %s:%d\n", 1440 lock1->li_lock, lock1->li_lock->lo_name, 1441 w1->w_name, w1->w_class->lc_name, 1442 fixup_filename(lock1->li_file), 1443 lock1->li_line); 1444 witness_output(" 3rd %p %s (%s, %s) @ %s:%d\n", lock, 1445 lock->lo_name, w->w_name, 1446 w->w_class->lc_name, fixup_filename(file), 1447 line); 1448 } 1449 if (trace) { 1450 char buf[64]; 1451 struct sbuf sb; 1452 1453 sbuf_new(&sb, buf, sizeof(buf), SBUF_FIXEDLEN); 1454 sbuf_set_drain(&sb, witness_output_drain, 1455 NULL); 1456 1457 if (pstackv) { 1458 sbuf_printf(&sb, 1459 "lock order %s -> %s established at:\n", 1460 w->w_name, w1->w_name); 1461 stack_sbuf_print_flags(&sb, &pstack, 1462 M_NOWAIT, STACK_SBUF_FMT_LONG); 1463 } 1464 1465 sbuf_printf(&sb, 1466 "lock order %s -> %s attempted at:\n", 1467 w1->w_name, w->w_name); 1468 stack_save(&pstack); 1469 stack_sbuf_print_flags(&sb, &pstack, M_NOWAIT, 1470 STACK_SBUF_FMT_LONG); 1471 1472 sbuf_finish(&sb); 1473 sbuf_delete(&sb); 1474 } 1475 witness_enter_debugger(__func__); 1476 return; 1477 } 1478 } 1479 1480 /* 1481 * If requested, build a new lock order. However, don't build a new 1482 * relationship between a sleepable lock and Giant if it is in the 1483 * wrong direction. The correct lock order is that sleepable locks 1484 * always come before Giant. 1485 */ 1486 if (flags & LOP_NEWORDER && 1487 !(plock->li_lock == &Giant.lock_object && 1488 (lock->lo_flags & LO_SLEEPABLE) != 0 && 1489 (flags & LOP_NOSLEEP) == 0)) { 1490 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__, 1491 w->w_name, plock->li_lock->lo_witness->w_name); 1492 itismychild(plock->li_lock->lo_witness, w); 1493 } 1494 out: 1495 mtx_unlock_spin(&w_mtx); 1496 } 1497 1498 void 1499 witness_lock(struct lock_object *lock, int flags, const char *file, int line) 1500 { 1501 struct lock_list_entry **lock_list, *lle; 1502 struct lock_instance *instance; 1503 struct witness *w; 1504 struct thread *td; 1505 1506 if (witness_cold || witness_watch == -1 || lock->lo_witness == NULL || 1507 KERNEL_PANICKED()) 1508 return; 1509 w = lock->lo_witness; 1510 td = curthread; 1511 1512 /* Determine lock list for this lock. */ 1513 if (LOCK_CLASS(lock)->lc_flags & LC_SLEEPLOCK) 1514 lock_list = &td->td_sleeplocks; 1515 else 1516 lock_list = PCPU_PTR(spinlocks); 1517 1518 /* Update per-witness last file and line acquire. */ 1519 w->w_file = file; 1520 w->w_line = line; 1521 1522 /* Check to see if we are recursing on a lock we already own. */ 1523 instance = find_instance(*lock_list, lock); 1524 if (instance != NULL) { 1525 instance->li_flags++; 1526 CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__, 1527 td->td_proc->p_pid, lock->lo_name, 1528 instance->li_flags & LI_RECURSEMASK); 1529 return; 1530 } 1531 1532 /* Find the next open lock instance in the list and fill it. */ 1533 lle = *lock_list; 1534 if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) { 1535 lle = witness_lock_list_get(); 1536 if (lle == NULL) 1537 return; 1538 lle->ll_next = *lock_list; 1539 CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__, 1540 td->td_proc->p_pid, lle); 1541 *lock_list = lle; 1542 } 1543 instance = &lle->ll_children[lle->ll_count++]; 1544 instance->li_lock = lock; 1545 instance->li_line = line; 1546 instance->li_file = file; 1547 instance->li_flags = 0; 1548 if ((flags & LOP_EXCLUSIVE) != 0) 1549 instance->li_flags |= LI_EXCLUSIVE; 1550 if ((lock->lo_flags & LO_SLEEPABLE) != 0 && (flags & LOP_NOSLEEP) == 0) 1551 instance->li_flags |= LI_SLEEPABLE; 1552 CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__, 1553 td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1); 1554 } 1555 1556 void 1557 witness_upgrade(struct lock_object *lock, int flags, const char *file, int line) 1558 { 1559 struct lock_instance *instance; 1560 struct lock_class *class; 1561 1562 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__)); 1563 if (lock->lo_witness == NULL || witness_watch == -1 || KERNEL_PANICKED()) 1564 return; 1565 class = LOCK_CLASS(lock); 1566 if (witness_watch) { 1567 if ((lock->lo_flags & LO_UPGRADABLE) == 0) 1568 kassert_panic( 1569 "upgrade of non-upgradable lock (%s) %s @ %s:%d", 1570 class->lc_name, lock->lo_name, 1571 fixup_filename(file), line); 1572 if ((class->lc_flags & LC_SLEEPLOCK) == 0) 1573 kassert_panic( 1574 "upgrade of non-sleep lock (%s) %s @ %s:%d", 1575 class->lc_name, lock->lo_name, 1576 fixup_filename(file), line); 1577 } 1578 instance = find_instance(curthread->td_sleeplocks, lock); 1579 if (instance == NULL) { 1580 kassert_panic("upgrade of unlocked lock (%s) %s @ %s:%d", 1581 class->lc_name, lock->lo_name, 1582 fixup_filename(file), line); 1583 return; 1584 } 1585 if (witness_watch) { 1586 if ((instance->li_flags & LI_EXCLUSIVE) != 0) 1587 kassert_panic( 1588 "upgrade of exclusive lock (%s) %s @ %s:%d", 1589 class->lc_name, lock->lo_name, 1590 fixup_filename(file), line); 1591 if ((instance->li_flags & LI_RECURSEMASK) != 0) 1592 kassert_panic( 1593 "upgrade of recursed lock (%s) %s r=%d @ %s:%d", 1594 class->lc_name, lock->lo_name, 1595 instance->li_flags & LI_RECURSEMASK, 1596 fixup_filename(file), line); 1597 } 1598 instance->li_flags |= LI_EXCLUSIVE; 1599 } 1600 1601 void 1602 witness_downgrade(struct lock_object *lock, int flags, const char *file, 1603 int line) 1604 { 1605 struct lock_instance *instance; 1606 struct lock_class *class; 1607 1608 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__)); 1609 if (lock->lo_witness == NULL || witness_watch == -1 || KERNEL_PANICKED()) 1610 return; 1611 class = LOCK_CLASS(lock); 1612 if (witness_watch) { 1613 if ((lock->lo_flags & LO_UPGRADABLE) == 0) 1614 kassert_panic( 1615 "downgrade of non-upgradable lock (%s) %s @ %s:%d", 1616 class->lc_name, lock->lo_name, 1617 fixup_filename(file), line); 1618 if ((class->lc_flags & LC_SLEEPLOCK) == 0) 1619 kassert_panic( 1620 "downgrade of non-sleep lock (%s) %s @ %s:%d", 1621 class->lc_name, lock->lo_name, 1622 fixup_filename(file), line); 1623 } 1624 instance = find_instance(curthread->td_sleeplocks, lock); 1625 if (instance == NULL) { 1626 kassert_panic("downgrade of unlocked lock (%s) %s @ %s:%d", 1627 class->lc_name, lock->lo_name, 1628 fixup_filename(file), line); 1629 return; 1630 } 1631 if (witness_watch) { 1632 if ((instance->li_flags & LI_EXCLUSIVE) == 0) 1633 kassert_panic( 1634 "downgrade of shared lock (%s) %s @ %s:%d", 1635 class->lc_name, lock->lo_name, 1636 fixup_filename(file), line); 1637 if ((instance->li_flags & LI_RECURSEMASK) != 0) 1638 kassert_panic( 1639 "downgrade of recursed lock (%s) %s r=%d @ %s:%d", 1640 class->lc_name, lock->lo_name, 1641 instance->li_flags & LI_RECURSEMASK, 1642 fixup_filename(file), line); 1643 } 1644 instance->li_flags &= ~LI_EXCLUSIVE; 1645 } 1646 1647 void 1648 witness_unlock(struct lock_object *lock, int flags, const char *file, int line) 1649 { 1650 struct lock_list_entry **lock_list, *lle; 1651 struct lock_instance *instance; 1652 struct lock_class *class; 1653 struct thread *td; 1654 register_t s; 1655 int i, j; 1656 1657 if (witness_cold || lock->lo_witness == NULL || KERNEL_PANICKED()) 1658 return; 1659 td = curthread; 1660 class = LOCK_CLASS(lock); 1661 1662 /* Find lock instance associated with this lock. */ 1663 if (class->lc_flags & LC_SLEEPLOCK) 1664 lock_list = &td->td_sleeplocks; 1665 else 1666 lock_list = PCPU_PTR(spinlocks); 1667 lle = *lock_list; 1668 for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next) 1669 for (i = 0; i < (*lock_list)->ll_count; i++) { 1670 instance = &(*lock_list)->ll_children[i]; 1671 if (instance->li_lock == lock) 1672 goto found; 1673 } 1674 1675 /* 1676 * When disabling WITNESS through witness_watch we could end up in 1677 * having registered locks in the td_sleeplocks queue. 1678 * We have to make sure we flush these queues, so just search for 1679 * eventual register locks and remove them. 1680 */ 1681 if (witness_watch > 0) { 1682 kassert_panic("lock (%s) %s not locked @ %s:%d", class->lc_name, 1683 lock->lo_name, fixup_filename(file), line); 1684 return; 1685 } else { 1686 return; 1687 } 1688 found: 1689 1690 /* First, check for shared/exclusive mismatches. */ 1691 if ((instance->li_flags & LI_EXCLUSIVE) != 0 && witness_watch > 0 && 1692 (flags & LOP_EXCLUSIVE) == 0) { 1693 witness_output("shared unlock of (%s) %s @ %s:%d\n", 1694 class->lc_name, lock->lo_name, fixup_filename(file), line); 1695 witness_output("while exclusively locked from %s:%d\n", 1696 fixup_filename(instance->li_file), instance->li_line); 1697 kassert_panic("excl->ushare"); 1698 } 1699 if ((instance->li_flags & LI_EXCLUSIVE) == 0 && witness_watch > 0 && 1700 (flags & LOP_EXCLUSIVE) != 0) { 1701 witness_output("exclusive unlock of (%s) %s @ %s:%d\n", 1702 class->lc_name, lock->lo_name, fixup_filename(file), line); 1703 witness_output("while share locked from %s:%d\n", 1704 fixup_filename(instance->li_file), 1705 instance->li_line); 1706 kassert_panic("share->uexcl"); 1707 } 1708 /* If we are recursed, unrecurse. */ 1709 if ((instance->li_flags & LI_RECURSEMASK) > 0) { 1710 CTR4(KTR_WITNESS, "%s: pid %d unrecursed on %s r=%d", __func__, 1711 td->td_proc->p_pid, instance->li_lock->lo_name, 1712 instance->li_flags); 1713 instance->li_flags--; 1714 return; 1715 } 1716 /* The lock is now being dropped, check for NORELEASE flag */ 1717 if ((instance->li_flags & LI_NORELEASE) != 0 && witness_watch > 0) { 1718 witness_output("forbidden unlock of (%s) %s @ %s:%d\n", 1719 class->lc_name, lock->lo_name, fixup_filename(file), line); 1720 kassert_panic("lock marked norelease"); 1721 } 1722 1723 /* Otherwise, remove this item from the list. */ 1724 s = intr_disable(); 1725 CTR4(KTR_WITNESS, "%s: pid %d removed %s from lle[%d]", __func__, 1726 td->td_proc->p_pid, instance->li_lock->lo_name, 1727 (*lock_list)->ll_count - 1); 1728 for (j = i; j < (*lock_list)->ll_count - 1; j++) 1729 (*lock_list)->ll_children[j] = 1730 (*lock_list)->ll_children[j + 1]; 1731 (*lock_list)->ll_count--; 1732 intr_restore(s); 1733 1734 /* 1735 * In order to reduce contention on w_mtx, we want to keep always an 1736 * head object into lists so that frequent allocation from the 1737 * free witness pool (and subsequent locking) is avoided. 1738 * In order to maintain the current code simple, when the head 1739 * object is totally unloaded it means also that we do not have 1740 * further objects in the list, so the list ownership needs to be 1741 * hand over to another object if the current head needs to be freed. 1742 */ 1743 if ((*lock_list)->ll_count == 0) { 1744 if (*lock_list == lle) { 1745 if (lle->ll_next == NULL) 1746 return; 1747 } else 1748 lle = *lock_list; 1749 *lock_list = lle->ll_next; 1750 CTR3(KTR_WITNESS, "%s: pid %d removed lle %p", __func__, 1751 td->td_proc->p_pid, lle); 1752 witness_lock_list_free(lle); 1753 } 1754 } 1755 1756 void 1757 witness_thread_exit(struct thread *td) 1758 { 1759 struct lock_list_entry *lle; 1760 int i, n; 1761 1762 lle = td->td_sleeplocks; 1763 if (lle == NULL || KERNEL_PANICKED()) 1764 return; 1765 if (lle->ll_count != 0) { 1766 for (n = 0; lle != NULL; lle = lle->ll_next) 1767 for (i = lle->ll_count - 1; i >= 0; i--) { 1768 if (n == 0) 1769 witness_output( 1770 "Thread %p exiting with the following locks held:\n", td); 1771 n++; 1772 witness_list_lock(&lle->ll_children[i], 1773 witness_output); 1774 1775 } 1776 kassert_panic( 1777 "Thread %p cannot exit while holding sleeplocks\n", td); 1778 } 1779 witness_lock_list_free(lle); 1780 } 1781 1782 /* 1783 * Warn if any locks other than 'lock' are held. Flags can be passed in to 1784 * exempt Giant and sleepable locks from the checks as well. If any 1785 * non-exempt locks are held, then a supplied message is printed to the 1786 * output channel along with a list of the offending locks. If indicated in the 1787 * flags then a failure results in a panic as well. 1788 */ 1789 int 1790 witness_warn(int flags, struct lock_object *lock, const char *fmt, ...) 1791 { 1792 struct lock_list_entry *lock_list, *lle; 1793 struct lock_instance *lock1; 1794 struct thread *td; 1795 va_list ap; 1796 int i, n; 1797 1798 if (witness_cold || witness_watch < 1 || KERNEL_PANICKED()) 1799 return (0); 1800 n = 0; 1801 td = curthread; 1802 for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next) 1803 for (i = lle->ll_count - 1; i >= 0; i--) { 1804 lock1 = &lle->ll_children[i]; 1805 if (lock1->li_lock == lock) 1806 continue; 1807 if (flags & WARN_GIANTOK && 1808 lock1->li_lock == &Giant.lock_object) 1809 continue; 1810 if (flags & WARN_SLEEPOK && 1811 (lock1->li_flags & LI_SLEEPABLE) != 0) 1812 continue; 1813 if (n == 0) { 1814 va_start(ap, fmt); 1815 vprintf(fmt, ap); 1816 va_end(ap); 1817 printf(" with the following %slocks held:\n", 1818 (flags & WARN_SLEEPOK) != 0 ? 1819 "non-sleepable " : ""); 1820 } 1821 n++; 1822 witness_list_lock(lock1, printf); 1823 } 1824 1825 /* 1826 * Pin the thread in order to avoid problems with thread migration. 1827 * Once that all verifies are passed about spinlocks ownership, 1828 * the thread is in a safe path and it can be unpinned. 1829 */ 1830 sched_pin(); 1831 lock_list = PCPU_GET(spinlocks); 1832 if (lock_list != NULL && lock_list->ll_count != 0) { 1833 sched_unpin(); 1834 1835 /* 1836 * We should only have one spinlock and as long as 1837 * the flags cannot match for this locks class, 1838 * check if the first spinlock is the one curthread 1839 * should hold. 1840 */ 1841 lock1 = &lock_list->ll_children[lock_list->ll_count - 1]; 1842 if (lock_list->ll_count == 1 && lock_list->ll_next == NULL && 1843 lock1->li_lock == lock && n == 0) 1844 return (0); 1845 1846 va_start(ap, fmt); 1847 vprintf(fmt, ap); 1848 va_end(ap); 1849 printf(" with the following %slocks held:\n", 1850 (flags & WARN_SLEEPOK) != 0 ? "non-sleepable " : ""); 1851 n += witness_list_locks(&lock_list, printf); 1852 } else 1853 sched_unpin(); 1854 if (flags & WARN_PANIC && n) 1855 kassert_panic("%s", __func__); 1856 else 1857 witness_debugger(n, __func__); 1858 return (n); 1859 } 1860 1861 const char * 1862 witness_file(struct lock_object *lock) 1863 { 1864 struct witness *w; 1865 1866 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL) 1867 return ("?"); 1868 w = lock->lo_witness; 1869 return (w->w_file); 1870 } 1871 1872 int 1873 witness_line(struct lock_object *lock) 1874 { 1875 struct witness *w; 1876 1877 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL) 1878 return (0); 1879 w = lock->lo_witness; 1880 return (w->w_line); 1881 } 1882 1883 static struct witness * 1884 enroll(const char *description, struct lock_class *lock_class) 1885 { 1886 struct witness *w; 1887 1888 MPASS(description != NULL); 1889 1890 if (witness_watch == -1 || KERNEL_PANICKED()) 1891 return (NULL); 1892 if ((lock_class->lc_flags & LC_SPINLOCK)) { 1893 if (witness_skipspin) 1894 return (NULL); 1895 } else if ((lock_class->lc_flags & LC_SLEEPLOCK) == 0) { 1896 kassert_panic("lock class %s is not sleep or spin", 1897 lock_class->lc_name); 1898 return (NULL); 1899 } 1900 1901 mtx_lock_spin(&w_mtx); 1902 w = witness_hash_get(description); 1903 if (w) 1904 goto found; 1905 if ((w = witness_get()) == NULL) 1906 return (NULL); 1907 MPASS(strlen(description) < MAX_W_NAME); 1908 strcpy(w->w_name, description); 1909 w->w_class = lock_class; 1910 w->w_refcount = 1; 1911 STAILQ_INSERT_HEAD(&w_all, w, w_list); 1912 if (lock_class->lc_flags & LC_SPINLOCK) { 1913 STAILQ_INSERT_HEAD(&w_spin, w, w_typelist); 1914 w_spin_cnt++; 1915 } else if (lock_class->lc_flags & LC_SLEEPLOCK) { 1916 STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist); 1917 w_sleep_cnt++; 1918 } 1919 1920 /* Insert new witness into the hash */ 1921 witness_hash_put(w); 1922 witness_increment_graph_generation(); 1923 mtx_unlock_spin(&w_mtx); 1924 return (w); 1925 found: 1926 w->w_refcount++; 1927 if (w->w_refcount == 1) 1928 w->w_class = lock_class; 1929 mtx_unlock_spin(&w_mtx); 1930 if (lock_class != w->w_class) 1931 kassert_panic( 1932 "lock (%s) %s does not match earlier (%s) lock", 1933 description, lock_class->lc_name, 1934 w->w_class->lc_name); 1935 return (w); 1936 } 1937 1938 static void 1939 depart(struct witness *w) 1940 { 1941 MPASS(w->w_refcount == 0); 1942 if (w->w_class->lc_flags & LC_SLEEPLOCK) { 1943 w_sleep_cnt--; 1944 } else { 1945 w_spin_cnt--; 1946 } 1947 /* 1948 * Set file to NULL as it may point into a loadable module. 1949 */ 1950 w->w_file = NULL; 1951 w->w_line = 0; 1952 witness_increment_graph_generation(); 1953 } 1954 1955 static void 1956 adopt(struct witness *parent, struct witness *child) 1957 { 1958 int pi, ci, i, j; 1959 1960 if (witness_cold == 0) 1961 mtx_assert(&w_mtx, MA_OWNED); 1962 1963 /* If the relationship is already known, there's no work to be done. */ 1964 if (isitmychild(parent, child)) 1965 return; 1966 1967 /* When the structure of the graph changes, bump up the generation. */ 1968 witness_increment_graph_generation(); 1969 1970 /* 1971 * The hard part ... create the direct relationship, then propagate all 1972 * indirect relationships. 1973 */ 1974 pi = parent->w_index; 1975 ci = child->w_index; 1976 WITNESS_INDEX_ASSERT(pi); 1977 WITNESS_INDEX_ASSERT(ci); 1978 MPASS(pi != ci); 1979 w_rmatrix[pi][ci] |= WITNESS_PARENT; 1980 w_rmatrix[ci][pi] |= WITNESS_CHILD; 1981 1982 /* 1983 * If parent was not already an ancestor of child, 1984 * then we increment the descendant and ancestor counters. 1985 */ 1986 if ((w_rmatrix[pi][ci] & WITNESS_ANCESTOR) == 0) { 1987 parent->w_num_descendants++; 1988 child->w_num_ancestors++; 1989 } 1990 1991 /* 1992 * Find each ancestor of 'pi'. Note that 'pi' itself is counted as 1993 * an ancestor of 'pi' during this loop. 1994 */ 1995 for (i = 1; i <= w_max_used_index; i++) { 1996 if ((w_rmatrix[i][pi] & WITNESS_ANCESTOR_MASK) == 0 && 1997 (i != pi)) 1998 continue; 1999 2000 /* Find each descendant of 'i' and mark it as a descendant. */ 2001 for (j = 1; j <= w_max_used_index; j++) { 2002 /* 2003 * Skip children that are already marked as 2004 * descendants of 'i'. 2005 */ 2006 if (w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK) 2007 continue; 2008 2009 /* 2010 * We are only interested in descendants of 'ci'. Note 2011 * that 'ci' itself is counted as a descendant of 'ci'. 2012 */ 2013 if ((w_rmatrix[ci][j] & WITNESS_ANCESTOR_MASK) == 0 && 2014 (j != ci)) 2015 continue; 2016 w_rmatrix[i][j] |= WITNESS_ANCESTOR; 2017 w_rmatrix[j][i] |= WITNESS_DESCENDANT; 2018 w_data[i].w_num_descendants++; 2019 w_data[j].w_num_ancestors++; 2020 2021 /* 2022 * Make sure we aren't marking a node as both an 2023 * ancestor and descendant. We should have caught 2024 * this as a lock order reversal earlier. 2025 */ 2026 if ((w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK) && 2027 (w_rmatrix[i][j] & WITNESS_DESCENDANT_MASK)) { 2028 printf("witness rmatrix paradox! [%d][%d]=%d " 2029 "both ancestor and descendant\n", 2030 i, j, w_rmatrix[i][j]); 2031 kdb_backtrace(); 2032 printf("Witness disabled.\n"); 2033 witness_watch = -1; 2034 } 2035 if ((w_rmatrix[j][i] & WITNESS_ANCESTOR_MASK) && 2036 (w_rmatrix[j][i] & WITNESS_DESCENDANT_MASK)) { 2037 printf("witness rmatrix paradox! [%d][%d]=%d " 2038 "both ancestor and descendant\n", 2039 j, i, w_rmatrix[j][i]); 2040 kdb_backtrace(); 2041 printf("Witness disabled.\n"); 2042 witness_watch = -1; 2043 } 2044 } 2045 } 2046 } 2047 2048 static void 2049 itismychild(struct witness *parent, struct witness *child) 2050 { 2051 int unlocked; 2052 2053 MPASS(child != NULL && parent != NULL); 2054 if (witness_cold == 0) 2055 mtx_assert(&w_mtx, MA_OWNED); 2056 2057 if (!witness_lock_type_equal(parent, child)) { 2058 if (witness_cold == 0) { 2059 unlocked = 1; 2060 mtx_unlock_spin(&w_mtx); 2061 } else { 2062 unlocked = 0; 2063 } 2064 kassert_panic( 2065 "%s: parent \"%s\" (%s) and child \"%s\" (%s) are not " 2066 "the same lock type", __func__, parent->w_name, 2067 parent->w_class->lc_name, child->w_name, 2068 child->w_class->lc_name); 2069 if (unlocked) 2070 mtx_lock_spin(&w_mtx); 2071 } 2072 adopt(parent, child); 2073 } 2074 2075 /* 2076 * Generic code for the isitmy*() functions. The rmask parameter is the 2077 * expected relationship of w1 to w2. 2078 */ 2079 static int 2080 _isitmyx(struct witness *w1, struct witness *w2, int rmask, const char *fname) 2081 { 2082 unsigned char r1, r2; 2083 int i1, i2; 2084 2085 i1 = w1->w_index; 2086 i2 = w2->w_index; 2087 WITNESS_INDEX_ASSERT(i1); 2088 WITNESS_INDEX_ASSERT(i2); 2089 r1 = w_rmatrix[i1][i2] & WITNESS_RELATED_MASK; 2090 r2 = w_rmatrix[i2][i1] & WITNESS_RELATED_MASK; 2091 2092 /* The flags on one better be the inverse of the flags on the other */ 2093 if (!((WITNESS_ATOD(r1) == r2 && WITNESS_DTOA(r2) == r1) || 2094 (WITNESS_DTOA(r1) == r2 && WITNESS_ATOD(r2) == r1))) { 2095 /* Don't squawk if we're potentially racing with an update. */ 2096 if (!mtx_owned(&w_mtx)) 2097 return (0); 2098 printf("%s: rmatrix mismatch between %s (index %d) and %s " 2099 "(index %d): w_rmatrix[%d][%d] == %hhx but " 2100 "w_rmatrix[%d][%d] == %hhx\n", 2101 fname, w1->w_name, i1, w2->w_name, i2, i1, i2, r1, 2102 i2, i1, r2); 2103 kdb_backtrace(); 2104 printf("Witness disabled.\n"); 2105 witness_watch = -1; 2106 } 2107 return (r1 & rmask); 2108 } 2109 2110 /* 2111 * Checks if @child is a direct child of @parent. 2112 */ 2113 static int 2114 isitmychild(struct witness *parent, struct witness *child) 2115 { 2116 return (_isitmyx(parent, child, WITNESS_PARENT, __func__)); 2117 } 2118 2119 /* 2120 * Checks if @descendant is a direct or inderect descendant of @ancestor. 2121 */ 2122 static int 2123 isitmydescendant(struct witness *ancestor, struct witness *descendant) 2124 { 2125 return (_isitmyx(ancestor, descendant, WITNESS_ANCESTOR_MASK, 2126 __func__)); 2127 } 2128 2129 static int 2130 blessed(struct witness *w1, struct witness *w2) 2131 { 2132 int i; 2133 struct witness_blessed *b; 2134 2135 for (i = 0; i < nitems(blessed_list); i++) { 2136 b = &blessed_list[i]; 2137 if (strcmp(w1->w_name, b->b_lock1) == 0) { 2138 if (strcmp(w2->w_name, b->b_lock2) == 0) 2139 return (1); 2140 continue; 2141 } 2142 if (strcmp(w1->w_name, b->b_lock2) == 0) 2143 if (strcmp(w2->w_name, b->b_lock1) == 0) 2144 return (1); 2145 } 2146 return (0); 2147 } 2148 2149 static struct witness * 2150 witness_get(void) 2151 { 2152 struct witness *w; 2153 int index; 2154 2155 if (witness_cold == 0) 2156 mtx_assert(&w_mtx, MA_OWNED); 2157 2158 if (witness_watch == -1) { 2159 mtx_unlock_spin(&w_mtx); 2160 return (NULL); 2161 } 2162 if (STAILQ_EMPTY(&w_free)) { 2163 witness_watch = -1; 2164 mtx_unlock_spin(&w_mtx); 2165 printf("WITNESS: unable to allocate a new witness object\n"); 2166 return (NULL); 2167 } 2168 w = STAILQ_FIRST(&w_free); 2169 STAILQ_REMOVE_HEAD(&w_free, w_list); 2170 w_free_cnt--; 2171 index = w->w_index; 2172 MPASS(index > 0 && index == w_max_used_index + 1 && 2173 index < witness_count); 2174 bzero(w, sizeof(*w)); 2175 w->w_index = index; 2176 if (index > w_max_used_index) 2177 w_max_used_index = index; 2178 return (w); 2179 } 2180 2181 static void 2182 witness_free(struct witness *w) 2183 { 2184 STAILQ_INSERT_HEAD(&w_free, w, w_list); 2185 w_free_cnt++; 2186 } 2187 2188 static struct lock_list_entry * 2189 witness_lock_list_get(void) 2190 { 2191 struct lock_list_entry *lle; 2192 2193 if (witness_watch == -1) 2194 return (NULL); 2195 mtx_lock_spin(&w_mtx); 2196 lle = w_lock_list_free; 2197 if (lle == NULL) { 2198 witness_watch = -1; 2199 mtx_unlock_spin(&w_mtx); 2200 printf("%s: witness exhausted\n", __func__); 2201 return (NULL); 2202 } 2203 w_lock_list_free = lle->ll_next; 2204 mtx_unlock_spin(&w_mtx); 2205 bzero(lle, sizeof(*lle)); 2206 return (lle); 2207 } 2208 2209 static void 2210 witness_lock_list_free(struct lock_list_entry *lle) 2211 { 2212 mtx_lock_spin(&w_mtx); 2213 lle->ll_next = w_lock_list_free; 2214 w_lock_list_free = lle; 2215 mtx_unlock_spin(&w_mtx); 2216 } 2217 2218 static struct lock_instance * 2219 find_instance(struct lock_list_entry *list, const struct lock_object *lock) 2220 { 2221 struct lock_list_entry *lle; 2222 struct lock_instance *instance; 2223 int i; 2224 2225 for (lle = list; lle != NULL; lle = lle->ll_next) 2226 for (i = lle->ll_count - 1; i >= 0; i--) { 2227 instance = &lle->ll_children[i]; 2228 if (instance->li_lock == lock) 2229 return (instance); 2230 } 2231 return (NULL); 2232 } 2233 2234 static void 2235 witness_list_lock(struct lock_instance *instance, 2236 int (*prnt)(const char *fmt, ...)) 2237 { 2238 struct lock_object *lock; 2239 2240 lock = instance->li_lock; 2241 prnt("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ? 2242 "exclusive" : "shared", LOCK_CLASS(lock)->lc_name, lock->lo_name); 2243 if (lock->lo_witness->w_name != lock->lo_name) 2244 prnt(" (%s)", lock->lo_witness->w_name); 2245 prnt(" r = %d (%p) locked @ %s:%d\n", 2246 instance->li_flags & LI_RECURSEMASK, lock, 2247 fixup_filename(instance->li_file), instance->li_line); 2248 } 2249 2250 static int 2251 witness_output(const char *fmt, ...) 2252 { 2253 va_list ap; 2254 int ret; 2255 2256 va_start(ap, fmt); 2257 ret = witness_voutput(fmt, ap); 2258 va_end(ap); 2259 return (ret); 2260 } 2261 2262 static int 2263 witness_voutput(const char *fmt, va_list ap) 2264 { 2265 int ret; 2266 2267 ret = 0; 2268 switch (witness_channel) { 2269 case WITNESS_CONSOLE: 2270 ret = vprintf(fmt, ap); 2271 break; 2272 case WITNESS_LOG: 2273 vlog(LOG_NOTICE, fmt, ap); 2274 break; 2275 case WITNESS_NONE: 2276 break; 2277 } 2278 return (ret); 2279 } 2280 2281 #ifdef DDB 2282 static int 2283 witness_thread_has_locks(struct thread *td) 2284 { 2285 if (td->td_sleeplocks == NULL) 2286 return (0); 2287 return (td->td_sleeplocks->ll_count != 0); 2288 } 2289 2290 static int 2291 witness_proc_has_locks(struct proc *p) 2292 { 2293 struct thread *td; 2294 2295 FOREACH_THREAD_IN_PROC(p, td) { 2296 if (witness_thread_has_locks(td)) 2297 return (1); 2298 } 2299 return (0); 2300 } 2301 #endif 2302 2303 int 2304 witness_list_locks(struct lock_list_entry **lock_list, 2305 int (*prnt)(const char *fmt, ...)) 2306 { 2307 struct lock_list_entry *lle; 2308 int i, nheld; 2309 2310 nheld = 0; 2311 for (lle = *lock_list; lle != NULL; lle = lle->ll_next) 2312 for (i = lle->ll_count - 1; i >= 0; i--) { 2313 witness_list_lock(&lle->ll_children[i], prnt); 2314 nheld++; 2315 } 2316 return (nheld); 2317 } 2318 2319 /* 2320 * This is a bit risky at best. We call this function when we have timed 2321 * out acquiring a spin lock, and we assume that the other CPU is stuck 2322 * with this lock held. So, we go groveling around in the other CPU's 2323 * per-cpu data to try to find the lock instance for this spin lock to 2324 * see when it was last acquired. 2325 */ 2326 void 2327 witness_display_spinlock(struct lock_object *lock, struct thread *owner, 2328 int (*prnt)(const char *fmt, ...)) 2329 { 2330 struct lock_instance *instance; 2331 struct pcpu *pc; 2332 2333 if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU) 2334 return; 2335 pc = pcpu_find(owner->td_oncpu); 2336 instance = find_instance(pc->pc_spinlocks, lock); 2337 if (instance != NULL) 2338 witness_list_lock(instance, prnt); 2339 } 2340 2341 void 2342 witness_save(struct lock_object *lock, const char **filep, int *linep) 2343 { 2344 struct lock_list_entry *lock_list; 2345 struct lock_instance *instance; 2346 struct lock_class *class; 2347 2348 /* Initialize for KMSAN's benefit. */ 2349 *filep = NULL; 2350 *linep = 0; 2351 2352 /* 2353 * This function is used independently in locking code to deal with 2354 * Giant, SCHEDULER_STOPPED() check can be removed here after Giant 2355 * is gone. 2356 */ 2357 if (SCHEDULER_STOPPED()) 2358 return; 2359 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__)); 2360 if (lock->lo_witness == NULL || witness_watch == -1 || KERNEL_PANICKED()) 2361 return; 2362 class = LOCK_CLASS(lock); 2363 if (class->lc_flags & LC_SLEEPLOCK) 2364 lock_list = curthread->td_sleeplocks; 2365 else { 2366 if (witness_skipspin) 2367 return; 2368 lock_list = PCPU_GET(spinlocks); 2369 } 2370 instance = find_instance(lock_list, lock); 2371 if (instance == NULL) { 2372 kassert_panic("%s: lock (%s) %s not locked", __func__, 2373 class->lc_name, lock->lo_name); 2374 return; 2375 } 2376 *filep = instance->li_file; 2377 *linep = instance->li_line; 2378 } 2379 2380 void 2381 witness_restore(struct lock_object *lock, const char *file, int line) 2382 { 2383 struct lock_list_entry *lock_list; 2384 struct lock_instance *instance; 2385 struct lock_class *class; 2386 2387 /* 2388 * This function is used independently in locking code to deal with 2389 * Giant, SCHEDULER_STOPPED() check can be removed here after Giant 2390 * is gone. 2391 */ 2392 if (SCHEDULER_STOPPED()) 2393 return; 2394 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__)); 2395 if (lock->lo_witness == NULL || witness_watch == -1 || KERNEL_PANICKED()) 2396 return; 2397 class = LOCK_CLASS(lock); 2398 if (class->lc_flags & LC_SLEEPLOCK) 2399 lock_list = curthread->td_sleeplocks; 2400 else { 2401 if (witness_skipspin) 2402 return; 2403 lock_list = PCPU_GET(spinlocks); 2404 } 2405 instance = find_instance(lock_list, lock); 2406 if (instance == NULL) 2407 kassert_panic("%s: lock (%s) %s not locked", __func__, 2408 class->lc_name, lock->lo_name); 2409 lock->lo_witness->w_file = file; 2410 lock->lo_witness->w_line = line; 2411 if (instance == NULL) 2412 return; 2413 instance->li_file = file; 2414 instance->li_line = line; 2415 } 2416 2417 static bool 2418 witness_find_instance(const struct lock_object *lock, 2419 struct lock_instance **instance) 2420 { 2421 #ifdef INVARIANT_SUPPORT 2422 struct lock_class *class; 2423 2424 if (lock->lo_witness == NULL || witness_watch < 1 || KERNEL_PANICKED()) 2425 return (false); 2426 class = LOCK_CLASS(lock); 2427 if ((class->lc_flags & LC_SLEEPLOCK) != 0) { 2428 *instance = find_instance(curthread->td_sleeplocks, lock); 2429 return (true); 2430 } else if ((class->lc_flags & LC_SPINLOCK) != 0) { 2431 *instance = find_instance(PCPU_GET(spinlocks), lock); 2432 return (true); 2433 } else { 2434 kassert_panic("Lock (%s) %s is not sleep or spin!", 2435 class->lc_name, lock->lo_name); 2436 return (false); 2437 } 2438 #else 2439 return (false); 2440 #endif 2441 } 2442 2443 void 2444 witness_assert(const struct lock_object *lock, int flags, const char *file, 2445 int line) 2446 { 2447 #ifdef INVARIANT_SUPPORT 2448 struct lock_instance *instance; 2449 struct lock_class *class; 2450 2451 if (!witness_find_instance(lock, &instance)) 2452 return; 2453 class = LOCK_CLASS(lock); 2454 switch (flags) { 2455 case LA_UNLOCKED: 2456 if (instance != NULL) 2457 kassert_panic("Lock (%s) %s locked @ %s:%d.", 2458 class->lc_name, lock->lo_name, 2459 fixup_filename(file), line); 2460 break; 2461 case LA_LOCKED: 2462 case LA_LOCKED | LA_RECURSED: 2463 case LA_LOCKED | LA_NOTRECURSED: 2464 case LA_SLOCKED: 2465 case LA_SLOCKED | LA_RECURSED: 2466 case LA_SLOCKED | LA_NOTRECURSED: 2467 case LA_XLOCKED: 2468 case LA_XLOCKED | LA_RECURSED: 2469 case LA_XLOCKED | LA_NOTRECURSED: 2470 if (instance == NULL) { 2471 kassert_panic("Lock (%s) %s not locked @ %s:%d.", 2472 class->lc_name, lock->lo_name, 2473 fixup_filename(file), line); 2474 break; 2475 } 2476 if ((flags & LA_XLOCKED) != 0 && 2477 (instance->li_flags & LI_EXCLUSIVE) == 0) 2478 kassert_panic( 2479 "Lock (%s) %s not exclusively locked @ %s:%d.", 2480 class->lc_name, lock->lo_name, 2481 fixup_filename(file), line); 2482 if ((flags & LA_SLOCKED) != 0 && 2483 (instance->li_flags & LI_EXCLUSIVE) != 0) 2484 kassert_panic( 2485 "Lock (%s) %s exclusively locked @ %s:%d.", 2486 class->lc_name, lock->lo_name, 2487 fixup_filename(file), line); 2488 if ((flags & LA_RECURSED) != 0 && 2489 (instance->li_flags & LI_RECURSEMASK) == 0) 2490 kassert_panic("Lock (%s) %s not recursed @ %s:%d.", 2491 class->lc_name, lock->lo_name, 2492 fixup_filename(file), line); 2493 if ((flags & LA_NOTRECURSED) != 0 && 2494 (instance->li_flags & LI_RECURSEMASK) != 0) 2495 kassert_panic("Lock (%s) %s recursed @ %s:%d.", 2496 class->lc_name, lock->lo_name, 2497 fixup_filename(file), line); 2498 break; 2499 default: 2500 kassert_panic("Invalid lock assertion at %s:%d.", 2501 fixup_filename(file), line); 2502 } 2503 #endif /* INVARIANT_SUPPORT */ 2504 } 2505 2506 /* 2507 * Checks the ownership of the lock by curthread, consulting the witness list. 2508 * Returns: 2509 * 0 if witness is disabled or did not work 2510 * -1 if not owned 2511 * 1 if owned 2512 */ 2513 int 2514 witness_is_owned(const struct lock_object *lock) 2515 { 2516 #ifdef INVARIANT_SUPPORT 2517 struct lock_instance *instance; 2518 2519 if (!witness_find_instance(lock, &instance)) 2520 return (0); 2521 return (instance == NULL ? -1 : 1); 2522 #else 2523 return (0); 2524 #endif 2525 } 2526 2527 static void 2528 witness_setflag(struct lock_object *lock, int flag, int set) 2529 { 2530 struct lock_list_entry *lock_list; 2531 struct lock_instance *instance; 2532 struct lock_class *class; 2533 2534 if (lock->lo_witness == NULL || witness_watch == -1 || KERNEL_PANICKED()) 2535 return; 2536 class = LOCK_CLASS(lock); 2537 if (class->lc_flags & LC_SLEEPLOCK) 2538 lock_list = curthread->td_sleeplocks; 2539 else { 2540 if (witness_skipspin) 2541 return; 2542 lock_list = PCPU_GET(spinlocks); 2543 } 2544 instance = find_instance(lock_list, lock); 2545 if (instance == NULL) { 2546 kassert_panic("%s: lock (%s) %s not locked", __func__, 2547 class->lc_name, lock->lo_name); 2548 return; 2549 } 2550 2551 if (set) 2552 instance->li_flags |= flag; 2553 else 2554 instance->li_flags &= ~flag; 2555 } 2556 2557 void 2558 witness_norelease(struct lock_object *lock) 2559 { 2560 witness_setflag(lock, LI_NORELEASE, 1); 2561 } 2562 2563 void 2564 witness_releaseok(struct lock_object *lock) 2565 { 2566 witness_setflag(lock, LI_NORELEASE, 0); 2567 } 2568 2569 #ifdef DDB 2570 static void 2571 witness_ddb_list(struct thread *td) 2572 { 2573 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__)); 2574 KASSERT(kdb_active, ("%s: not in the debugger", __func__)); 2575 2576 if (witness_watch < 1) 2577 return; 2578 2579 witness_list_locks(&td->td_sleeplocks, db_printf); 2580 2581 /* 2582 * We only handle spinlocks if td == curthread. This is somewhat broken 2583 * if td is currently executing on some other CPU and holds spin locks 2584 * as we won't display those locks. If we had a MI way of getting 2585 * the per-cpu data for a given cpu then we could use 2586 * td->td_oncpu to get the list of spinlocks for this thread 2587 * and "fix" this. 2588 * 2589 * That still wouldn't really fix this unless we locked the scheduler 2590 * lock or stopped the other CPU to make sure it wasn't changing the 2591 * list out from under us. It is probably best to just not try to 2592 * handle threads on other CPU's for now. 2593 */ 2594 if (td == curthread && PCPU_GET(spinlocks) != NULL) 2595 witness_list_locks(PCPU_PTR(spinlocks), db_printf); 2596 } 2597 2598 DB_SHOW_COMMAND(locks, db_witness_list) 2599 { 2600 struct thread *td; 2601 2602 if (have_addr) 2603 td = db_lookup_thread(addr, true); 2604 else 2605 td = kdb_thread; 2606 witness_ddb_list(td); 2607 } 2608 2609 DB_SHOW_ALL_COMMAND(locks, db_witness_list_all) 2610 { 2611 struct thread *td; 2612 struct proc *p; 2613 2614 /* 2615 * It would be nice to list only threads and processes that actually 2616 * held sleep locks, but that information is currently not exported 2617 * by WITNESS. 2618 */ 2619 FOREACH_PROC_IN_SYSTEM(p) { 2620 if (!witness_proc_has_locks(p)) 2621 continue; 2622 FOREACH_THREAD_IN_PROC(p, td) { 2623 if (!witness_thread_has_locks(td)) 2624 continue; 2625 db_printf("Process %d (%s) thread %p (%d)\n", p->p_pid, 2626 p->p_comm, td, td->td_tid); 2627 witness_ddb_list(td); 2628 if (db_pager_quit) 2629 return; 2630 } 2631 } 2632 } 2633 DB_SHOW_ALIAS_FLAGS(alllocks, db_witness_list_all, DB_CMD_MEMSAFE); 2634 2635 DB_SHOW_COMMAND_FLAGS(witness, db_witness_display, DB_CMD_MEMSAFE) 2636 { 2637 witness_ddb_display(db_printf); 2638 } 2639 #endif 2640 2641 static void 2642 sbuf_print_witness_badstacks(struct sbuf *sb, size_t *oldidx) 2643 { 2644 struct witness_lock_order_data *data1, *data2, *tmp_data1, *tmp_data2; 2645 struct witness *tmp_w1, *tmp_w2, *w1, *w2; 2646 int generation, i, j; 2647 2648 tmp_data1 = NULL; 2649 tmp_data2 = NULL; 2650 tmp_w1 = NULL; 2651 tmp_w2 = NULL; 2652 2653 /* Allocate and init temporary storage space. */ 2654 tmp_w1 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO); 2655 tmp_w2 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO); 2656 tmp_data1 = malloc(sizeof(struct witness_lock_order_data), M_TEMP, 2657 M_WAITOK | M_ZERO); 2658 tmp_data2 = malloc(sizeof(struct witness_lock_order_data), M_TEMP, 2659 M_WAITOK | M_ZERO); 2660 stack_zero(&tmp_data1->wlod_stack); 2661 stack_zero(&tmp_data2->wlod_stack); 2662 2663 restart: 2664 mtx_lock_spin(&w_mtx); 2665 generation = w_generation; 2666 mtx_unlock_spin(&w_mtx); 2667 sbuf_printf(sb, "Number of known direct relationships is %d\n", 2668 w_lohash.wloh_count); 2669 for (i = 1; i < w_max_used_index; i++) { 2670 mtx_lock_spin(&w_mtx); 2671 if (generation != w_generation) { 2672 mtx_unlock_spin(&w_mtx); 2673 2674 /* The graph has changed, try again. */ 2675 *oldidx = 0; 2676 sbuf_clear(sb); 2677 goto restart; 2678 } 2679 2680 w1 = &w_data[i]; 2681 if (w1->w_reversed == 0) { 2682 mtx_unlock_spin(&w_mtx); 2683 continue; 2684 } 2685 2686 /* Copy w1 locally so we can release the spin lock. */ 2687 *tmp_w1 = *w1; 2688 mtx_unlock_spin(&w_mtx); 2689 2690 if (tmp_w1->w_reversed == 0) 2691 continue; 2692 for (j = 1; j < w_max_used_index; j++) { 2693 if ((w_rmatrix[i][j] & WITNESS_REVERSAL) == 0 || i > j) 2694 continue; 2695 2696 mtx_lock_spin(&w_mtx); 2697 if (generation != w_generation) { 2698 mtx_unlock_spin(&w_mtx); 2699 2700 /* The graph has changed, try again. */ 2701 *oldidx = 0; 2702 sbuf_clear(sb); 2703 goto restart; 2704 } 2705 2706 w2 = &w_data[j]; 2707 data1 = witness_lock_order_get(w1, w2); 2708 data2 = witness_lock_order_get(w2, w1); 2709 2710 /* 2711 * Copy information locally so we can release the 2712 * spin lock. 2713 */ 2714 *tmp_w2 = *w2; 2715 2716 if (data1) { 2717 stack_zero(&tmp_data1->wlod_stack); 2718 stack_copy(&data1->wlod_stack, 2719 &tmp_data1->wlod_stack); 2720 } 2721 if (data2 && data2 != data1) { 2722 stack_zero(&tmp_data2->wlod_stack); 2723 stack_copy(&data2->wlod_stack, 2724 &tmp_data2->wlod_stack); 2725 } 2726 mtx_unlock_spin(&w_mtx); 2727 2728 if (blessed(tmp_w1, tmp_w2)) 2729 continue; 2730 2731 sbuf_printf(sb, 2732 "\nLock order reversal between \"%s\"(%s) and \"%s\"(%s)!\n", 2733 tmp_w1->w_name, tmp_w1->w_class->lc_name, 2734 tmp_w2->w_name, tmp_w2->w_class->lc_name); 2735 if (data1) { 2736 sbuf_printf(sb, 2737 "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n", 2738 tmp_w1->w_name, tmp_w1->w_class->lc_name, 2739 tmp_w2->w_name, tmp_w2->w_class->lc_name); 2740 stack_sbuf_print(sb, &tmp_data1->wlod_stack); 2741 sbuf_putc(sb, '\n'); 2742 } 2743 if (data2 && data2 != data1) { 2744 sbuf_printf(sb, 2745 "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n", 2746 tmp_w2->w_name, tmp_w2->w_class->lc_name, 2747 tmp_w1->w_name, tmp_w1->w_class->lc_name); 2748 stack_sbuf_print(sb, &tmp_data2->wlod_stack); 2749 sbuf_putc(sb, '\n'); 2750 } 2751 } 2752 } 2753 mtx_lock_spin(&w_mtx); 2754 if (generation != w_generation) { 2755 mtx_unlock_spin(&w_mtx); 2756 2757 /* 2758 * The graph changed while we were printing stack data, 2759 * try again. 2760 */ 2761 *oldidx = 0; 2762 sbuf_clear(sb); 2763 goto restart; 2764 } 2765 mtx_unlock_spin(&w_mtx); 2766 2767 /* Free temporary storage space. */ 2768 free(tmp_data1, M_TEMP); 2769 free(tmp_data2, M_TEMP); 2770 free(tmp_w1, M_TEMP); 2771 free(tmp_w2, M_TEMP); 2772 } 2773 2774 static int 2775 sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS) 2776 { 2777 struct sbuf *sb; 2778 int error; 2779 2780 if (witness_watch < 1) { 2781 error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning)); 2782 return (error); 2783 } 2784 if (witness_cold) { 2785 error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold)); 2786 return (error); 2787 } 2788 error = 0; 2789 sb = sbuf_new(NULL, NULL, badstack_sbuf_size, SBUF_AUTOEXTEND); 2790 if (sb == NULL) 2791 return (ENOMEM); 2792 2793 sbuf_print_witness_badstacks(sb, &req->oldidx); 2794 2795 sbuf_finish(sb); 2796 error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1); 2797 sbuf_delete(sb); 2798 2799 return (error); 2800 } 2801 2802 #ifdef DDB 2803 static int 2804 sbuf_db_printf_drain(void *arg __unused, const char *data, int len) 2805 { 2806 return (db_printf("%.*s", len, data)); 2807 } 2808 2809 DB_SHOW_COMMAND_FLAGS(badstacks, db_witness_badstacks, DB_CMD_MEMSAFE) 2810 { 2811 struct sbuf sb; 2812 char buffer[128]; 2813 size_t dummy; 2814 2815 sbuf_new(&sb, buffer, sizeof(buffer), SBUF_FIXEDLEN); 2816 sbuf_set_drain(&sb, sbuf_db_printf_drain, NULL); 2817 sbuf_print_witness_badstacks(&sb, &dummy); 2818 sbuf_finish(&sb); 2819 } 2820 #endif 2821 2822 static int 2823 sysctl_debug_witness_channel(SYSCTL_HANDLER_ARGS) 2824 { 2825 static const struct { 2826 enum witness_channel channel; 2827 const char *name; 2828 } channels[] = { 2829 { WITNESS_CONSOLE, "console" }, 2830 { WITNESS_LOG, "log" }, 2831 { WITNESS_NONE, "none" }, 2832 }; 2833 char buf[16]; 2834 u_int i; 2835 int error; 2836 2837 buf[0] = '\0'; 2838 for (i = 0; i < nitems(channels); i++) 2839 if (witness_channel == channels[i].channel) { 2840 snprintf(buf, sizeof(buf), "%s", channels[i].name); 2841 break; 2842 } 2843 2844 error = sysctl_handle_string(oidp, buf, sizeof(buf), req); 2845 if (error != 0 || req->newptr == NULL) 2846 return (error); 2847 2848 error = EINVAL; 2849 for (i = 0; i < nitems(channels); i++) 2850 if (strcmp(channels[i].name, buf) == 0) { 2851 witness_channel = channels[i].channel; 2852 error = 0; 2853 break; 2854 } 2855 return (error); 2856 } 2857 2858 static int 2859 sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS) 2860 { 2861 struct witness *w; 2862 struct sbuf *sb; 2863 int error; 2864 2865 #ifdef __i386__ 2866 error = SYSCTL_OUT(req, w_notallowed, sizeof(w_notallowed)); 2867 return (error); 2868 #endif 2869 2870 if (witness_watch < 1) { 2871 error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning)); 2872 return (error); 2873 } 2874 if (witness_cold) { 2875 error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold)); 2876 return (error); 2877 } 2878 error = 0; 2879 2880 error = sysctl_wire_old_buffer(req, 0); 2881 if (error != 0) 2882 return (error); 2883 sb = sbuf_new_for_sysctl(NULL, NULL, FULLGRAPH_SBUF_SIZE, req); 2884 if (sb == NULL) 2885 return (ENOMEM); 2886 sbuf_putc(sb, '\n'); 2887 2888 mtx_lock_spin(&w_mtx); 2889 STAILQ_FOREACH(w, &w_all, w_list) 2890 w->w_displayed = 0; 2891 STAILQ_FOREACH(w, &w_all, w_list) 2892 witness_add_fullgraph(sb, w); 2893 mtx_unlock_spin(&w_mtx); 2894 2895 /* 2896 * Close the sbuf and return to userland. 2897 */ 2898 error = sbuf_finish(sb); 2899 sbuf_delete(sb); 2900 2901 return (error); 2902 } 2903 2904 static int 2905 sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS) 2906 { 2907 int error, value; 2908 2909 value = witness_watch; 2910 error = sysctl_handle_int(oidp, &value, 0, req); 2911 if (error != 0 || req->newptr == NULL) 2912 return (error); 2913 if (value > 1 || value < -1 || 2914 (witness_watch == -1 && value != witness_watch)) 2915 return (EINVAL); 2916 witness_watch = value; 2917 return (0); 2918 } 2919 2920 static void 2921 witness_add_fullgraph(struct sbuf *sb, struct witness *w) 2922 { 2923 int i; 2924 2925 if (w->w_displayed != 0 || (w->w_file == NULL && w->w_line == 0)) 2926 return; 2927 w->w_displayed = 1; 2928 2929 WITNESS_INDEX_ASSERT(w->w_index); 2930 for (i = 1; i <= w_max_used_index; i++) { 2931 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT) { 2932 sbuf_printf(sb, "\"%s\",\"%s\"\n", w->w_name, 2933 w_data[i].w_name); 2934 witness_add_fullgraph(sb, &w_data[i]); 2935 } 2936 } 2937 } 2938 2939 /* 2940 * A simple hash function. Takes a key pointer and a key size. If size == 0, 2941 * interprets the key as a string and reads until the null 2942 * terminator. Otherwise, reads the first size bytes. Returns an unsigned 32-bit 2943 * hash value computed from the key. 2944 */ 2945 static uint32_t 2946 witness_hash_djb2(const uint8_t *key, uint32_t size) 2947 { 2948 unsigned int hash = 5381; 2949 int i; 2950 2951 /* hash = hash * 33 + key[i] */ 2952 if (size) 2953 for (i = 0; i < size; i++) 2954 hash = ((hash << 5) + hash) + (unsigned int)key[i]; 2955 else 2956 for (i = 0; key[i] != 0; i++) 2957 hash = ((hash << 5) + hash) + (unsigned int)key[i]; 2958 2959 return (hash); 2960 } 2961 2962 /* 2963 * Initializes the two witness hash tables. Called exactly once from 2964 * witness_initialize(). 2965 */ 2966 static void 2967 witness_init_hash_tables(void) 2968 { 2969 int i; 2970 2971 MPASS(witness_cold); 2972 2973 /* Initialize the hash tables. */ 2974 for (i = 0; i < WITNESS_HASH_SIZE; i++) 2975 w_hash.wh_array[i] = NULL; 2976 2977 w_hash.wh_size = WITNESS_HASH_SIZE; 2978 w_hash.wh_count = 0; 2979 2980 /* Initialize the lock order data hash. */ 2981 w_lofree = NULL; 2982 for (i = 0; i < WITNESS_LO_DATA_COUNT; i++) { 2983 memset(&w_lodata[i], 0, sizeof(w_lodata[i])); 2984 w_lodata[i].wlod_next = w_lofree; 2985 w_lofree = &w_lodata[i]; 2986 } 2987 w_lohash.wloh_size = WITNESS_LO_HASH_SIZE; 2988 w_lohash.wloh_count = 0; 2989 for (i = 0; i < WITNESS_LO_HASH_SIZE; i++) 2990 w_lohash.wloh_array[i] = NULL; 2991 } 2992 2993 static struct witness * 2994 witness_hash_get(const char *key) 2995 { 2996 struct witness *w; 2997 uint32_t hash; 2998 2999 MPASS(key != NULL); 3000 if (witness_cold == 0) 3001 mtx_assert(&w_mtx, MA_OWNED); 3002 hash = witness_hash_djb2(key, 0) % w_hash.wh_size; 3003 w = w_hash.wh_array[hash]; 3004 while (w != NULL) { 3005 if (strcmp(w->w_name, key) == 0) 3006 goto out; 3007 w = w->w_hash_next; 3008 } 3009 3010 out: 3011 return (w); 3012 } 3013 3014 static void 3015 witness_hash_put(struct witness *w) 3016 { 3017 uint32_t hash; 3018 3019 MPASS(w != NULL); 3020 MPASS(w->w_name != NULL); 3021 if (witness_cold == 0) 3022 mtx_assert(&w_mtx, MA_OWNED); 3023 KASSERT(witness_hash_get(w->w_name) == NULL, 3024 ("%s: trying to add a hash entry that already exists!", __func__)); 3025 KASSERT(w->w_hash_next == NULL, 3026 ("%s: w->w_hash_next != NULL", __func__)); 3027 3028 hash = witness_hash_djb2(w->w_name, 0) % w_hash.wh_size; 3029 w->w_hash_next = w_hash.wh_array[hash]; 3030 w_hash.wh_array[hash] = w; 3031 w_hash.wh_count++; 3032 } 3033 3034 static struct witness_lock_order_data * 3035 witness_lock_order_get(struct witness *parent, struct witness *child) 3036 { 3037 struct witness_lock_order_data *data = NULL; 3038 struct witness_lock_order_key key; 3039 unsigned int hash; 3040 3041 MPASS(parent != NULL && child != NULL); 3042 key.from = parent->w_index; 3043 key.to = child->w_index; 3044 WITNESS_INDEX_ASSERT(key.from); 3045 WITNESS_INDEX_ASSERT(key.to); 3046 if ((w_rmatrix[parent->w_index][child->w_index] 3047 & WITNESS_LOCK_ORDER_KNOWN) == 0) 3048 goto out; 3049 3050 hash = witness_hash_djb2((const char *)&key, 3051 sizeof(key)) % w_lohash.wloh_size; 3052 data = w_lohash.wloh_array[hash]; 3053 while (data != NULL) { 3054 if (witness_lock_order_key_equal(&data->wlod_key, &key)) 3055 break; 3056 data = data->wlod_next; 3057 } 3058 3059 out: 3060 return (data); 3061 } 3062 3063 /* 3064 * Verify that parent and child have a known relationship, are not the same, 3065 * and child is actually a child of parent. This is done without w_mtx 3066 * to avoid contention in the common case. 3067 */ 3068 static int 3069 witness_lock_order_check(struct witness *parent, struct witness *child) 3070 { 3071 if (parent != child && 3072 w_rmatrix[parent->w_index][child->w_index] 3073 & WITNESS_LOCK_ORDER_KNOWN && 3074 isitmychild(parent, child)) 3075 return (1); 3076 3077 return (0); 3078 } 3079 3080 static int 3081 witness_lock_order_add(struct witness *parent, struct witness *child) 3082 { 3083 struct witness_lock_order_data *data = NULL; 3084 struct witness_lock_order_key key; 3085 unsigned int hash; 3086 3087 MPASS(parent != NULL && child != NULL); 3088 key.from = parent->w_index; 3089 key.to = child->w_index; 3090 WITNESS_INDEX_ASSERT(key.from); 3091 WITNESS_INDEX_ASSERT(key.to); 3092 if (w_rmatrix[parent->w_index][child->w_index] 3093 & WITNESS_LOCK_ORDER_KNOWN) 3094 return (1); 3095 3096 hash = witness_hash_djb2((const char *)&key, 3097 sizeof(key)) % w_lohash.wloh_size; 3098 w_rmatrix[parent->w_index][child->w_index] |= WITNESS_LOCK_ORDER_KNOWN; 3099 data = w_lofree; 3100 if (data == NULL) 3101 return (0); 3102 w_lofree = data->wlod_next; 3103 data->wlod_next = w_lohash.wloh_array[hash]; 3104 data->wlod_key = key; 3105 w_lohash.wloh_array[hash] = data; 3106 w_lohash.wloh_count++; 3107 stack_save(&data->wlod_stack); 3108 return (1); 3109 } 3110 3111 /* Call this whenever the structure of the witness graph changes. */ 3112 static void 3113 witness_increment_graph_generation(void) 3114 { 3115 if (witness_cold == 0) 3116 mtx_assert(&w_mtx, MA_OWNED); 3117 w_generation++; 3118 } 3119 3120 static int 3121 witness_output_drain(void *arg __unused, const char *data, int len) 3122 { 3123 witness_output("%.*s", len, data); 3124 return (len); 3125 } 3126 3127 static void 3128 witness_debugger(int cond, const char *msg) 3129 { 3130 char buf[32]; 3131 struct sbuf sb; 3132 struct stack st; 3133 3134 if (!cond) 3135 return; 3136 3137 if (witness_trace) { 3138 sbuf_new(&sb, buf, sizeof(buf), SBUF_FIXEDLEN); 3139 sbuf_set_drain(&sb, witness_output_drain, NULL); 3140 3141 stack_save(&st); 3142 witness_output("stack backtrace:\n"); 3143 stack_sbuf_print_ddb(&sb, &st); 3144 3145 sbuf_finish(&sb); 3146 } 3147 3148 witness_enter_debugger(msg); 3149 } 3150 3151 static void 3152 witness_enter_debugger(const char *msg) 3153 { 3154 #ifdef KDB 3155 if (witness_kdb) 3156 kdb_enter(KDB_WHY_WITNESS, msg); 3157 #endif 3158 } 3159