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 /* Check to see if we are recursing on a lock we already own. */ 1519 instance = find_instance(*lock_list, lock); 1520 if (instance != NULL) { 1521 instance->li_flags++; 1522 CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__, 1523 td->td_proc->p_pid, lock->lo_name, 1524 instance->li_flags & LI_RECURSEMASK); 1525 instance->li_file = file; 1526 instance->li_line = line; 1527 return; 1528 } 1529 1530 /* Update per-witness last file and line acquire. */ 1531 w->w_file = file; 1532 w->w_line = line; 1533 1534 /* Find the next open lock instance in the list and fill it. */ 1535 lle = *lock_list; 1536 if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) { 1537 lle = witness_lock_list_get(); 1538 if (lle == NULL) 1539 return; 1540 lle->ll_next = *lock_list; 1541 CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__, 1542 td->td_proc->p_pid, lle); 1543 *lock_list = lle; 1544 } 1545 instance = &lle->ll_children[lle->ll_count++]; 1546 instance->li_lock = lock; 1547 instance->li_line = line; 1548 instance->li_file = file; 1549 instance->li_flags = 0; 1550 if ((flags & LOP_EXCLUSIVE) != 0) 1551 instance->li_flags |= LI_EXCLUSIVE; 1552 if ((lock->lo_flags & LO_SLEEPABLE) != 0 && (flags & LOP_NOSLEEP) == 0) 1553 instance->li_flags |= LI_SLEEPABLE; 1554 CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__, 1555 td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1); 1556 } 1557 1558 void 1559 witness_upgrade(struct lock_object *lock, int flags, const char *file, int line) 1560 { 1561 struct lock_instance *instance; 1562 struct lock_class *class; 1563 1564 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__)); 1565 if (lock->lo_witness == NULL || witness_watch == -1 || KERNEL_PANICKED()) 1566 return; 1567 class = LOCK_CLASS(lock); 1568 if (witness_watch) { 1569 if ((lock->lo_flags & LO_UPGRADABLE) == 0) 1570 kassert_panic( 1571 "upgrade of non-upgradable lock (%s) %s @ %s:%d", 1572 class->lc_name, lock->lo_name, 1573 fixup_filename(file), line); 1574 if ((class->lc_flags & LC_SLEEPLOCK) == 0) 1575 kassert_panic( 1576 "upgrade of non-sleep lock (%s) %s @ %s:%d", 1577 class->lc_name, lock->lo_name, 1578 fixup_filename(file), line); 1579 } 1580 instance = find_instance(curthread->td_sleeplocks, lock); 1581 if (instance == NULL) { 1582 kassert_panic("upgrade of unlocked lock (%s) %s @ %s:%d", 1583 class->lc_name, lock->lo_name, 1584 fixup_filename(file), line); 1585 return; 1586 } 1587 if (witness_watch) { 1588 if ((instance->li_flags & LI_EXCLUSIVE) != 0) 1589 kassert_panic( 1590 "upgrade of exclusive lock (%s) %s @ %s:%d", 1591 class->lc_name, lock->lo_name, 1592 fixup_filename(file), line); 1593 if ((instance->li_flags & LI_RECURSEMASK) != 0) 1594 kassert_panic( 1595 "upgrade of recursed lock (%s) %s r=%d @ %s:%d", 1596 class->lc_name, lock->lo_name, 1597 instance->li_flags & LI_RECURSEMASK, 1598 fixup_filename(file), line); 1599 } 1600 instance->li_flags |= LI_EXCLUSIVE; 1601 } 1602 1603 void 1604 witness_downgrade(struct lock_object *lock, int flags, const char *file, 1605 int line) 1606 { 1607 struct lock_instance *instance; 1608 struct lock_class *class; 1609 1610 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__)); 1611 if (lock->lo_witness == NULL || witness_watch == -1 || KERNEL_PANICKED()) 1612 return; 1613 class = LOCK_CLASS(lock); 1614 if (witness_watch) { 1615 if ((lock->lo_flags & LO_UPGRADABLE) == 0) 1616 kassert_panic( 1617 "downgrade of non-upgradable lock (%s) %s @ %s:%d", 1618 class->lc_name, lock->lo_name, 1619 fixup_filename(file), line); 1620 if ((class->lc_flags & LC_SLEEPLOCK) == 0) 1621 kassert_panic( 1622 "downgrade of non-sleep lock (%s) %s @ %s:%d", 1623 class->lc_name, lock->lo_name, 1624 fixup_filename(file), line); 1625 } 1626 instance = find_instance(curthread->td_sleeplocks, lock); 1627 if (instance == NULL) { 1628 kassert_panic("downgrade of unlocked lock (%s) %s @ %s:%d", 1629 class->lc_name, lock->lo_name, 1630 fixup_filename(file), line); 1631 return; 1632 } 1633 if (witness_watch) { 1634 if ((instance->li_flags & LI_EXCLUSIVE) == 0) 1635 kassert_panic( 1636 "downgrade of shared lock (%s) %s @ %s:%d", 1637 class->lc_name, lock->lo_name, 1638 fixup_filename(file), line); 1639 if ((instance->li_flags & LI_RECURSEMASK) != 0) 1640 kassert_panic( 1641 "downgrade of recursed lock (%s) %s r=%d @ %s:%d", 1642 class->lc_name, lock->lo_name, 1643 instance->li_flags & LI_RECURSEMASK, 1644 fixup_filename(file), line); 1645 } 1646 instance->li_flags &= ~LI_EXCLUSIVE; 1647 } 1648 1649 void 1650 witness_unlock(struct lock_object *lock, int flags, const char *file, int line) 1651 { 1652 struct lock_list_entry **lock_list, *lle; 1653 struct lock_instance *instance; 1654 struct lock_class *class; 1655 struct thread *td; 1656 register_t s; 1657 int i, j; 1658 1659 if (witness_cold || lock->lo_witness == NULL || KERNEL_PANICKED()) 1660 return; 1661 td = curthread; 1662 class = LOCK_CLASS(lock); 1663 1664 /* Find lock instance associated with this lock. */ 1665 if (class->lc_flags & LC_SLEEPLOCK) 1666 lock_list = &td->td_sleeplocks; 1667 else 1668 lock_list = PCPU_PTR(spinlocks); 1669 lle = *lock_list; 1670 for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next) 1671 for (i = 0; i < (*lock_list)->ll_count; i++) { 1672 instance = &(*lock_list)->ll_children[i]; 1673 if (instance->li_lock == lock) 1674 goto found; 1675 } 1676 1677 /* 1678 * When disabling WITNESS through witness_watch we could end up in 1679 * having registered locks in the td_sleeplocks queue. 1680 * We have to make sure we flush these queues, so just search for 1681 * eventual register locks and remove them. 1682 */ 1683 if (witness_watch > 0) { 1684 kassert_panic("lock (%s) %s not locked @ %s:%d", class->lc_name, 1685 lock->lo_name, fixup_filename(file), line); 1686 return; 1687 } else { 1688 return; 1689 } 1690 found: 1691 1692 /* First, check for shared/exclusive mismatches. */ 1693 if ((instance->li_flags & LI_EXCLUSIVE) != 0 && witness_watch > 0 && 1694 (flags & LOP_EXCLUSIVE) == 0) { 1695 witness_output("shared unlock of (%s) %s @ %s:%d\n", 1696 class->lc_name, lock->lo_name, fixup_filename(file), line); 1697 witness_output("while exclusively locked from %s:%d\n", 1698 fixup_filename(instance->li_file), instance->li_line); 1699 kassert_panic("excl->ushare"); 1700 } 1701 if ((instance->li_flags & LI_EXCLUSIVE) == 0 && witness_watch > 0 && 1702 (flags & LOP_EXCLUSIVE) != 0) { 1703 witness_output("exclusive unlock of (%s) %s @ %s:%d\n", 1704 class->lc_name, lock->lo_name, fixup_filename(file), line); 1705 witness_output("while share locked from %s:%d\n", 1706 fixup_filename(instance->li_file), 1707 instance->li_line); 1708 kassert_panic("share->uexcl"); 1709 } 1710 /* If we are recursed, unrecurse. */ 1711 if ((instance->li_flags & LI_RECURSEMASK) > 0) { 1712 CTR4(KTR_WITNESS, "%s: pid %d unrecursed on %s r=%d", __func__, 1713 td->td_proc->p_pid, instance->li_lock->lo_name, 1714 instance->li_flags); 1715 instance->li_flags--; 1716 return; 1717 } 1718 /* The lock is now being dropped, check for NORELEASE flag */ 1719 if ((instance->li_flags & LI_NORELEASE) != 0 && witness_watch > 0) { 1720 witness_output("forbidden unlock of (%s) %s @ %s:%d\n", 1721 class->lc_name, lock->lo_name, fixup_filename(file), line); 1722 kassert_panic("lock marked norelease"); 1723 } 1724 1725 /* Otherwise, remove this item from the list. */ 1726 s = intr_disable(); 1727 CTR4(KTR_WITNESS, "%s: pid %d removed %s from lle[%d]", __func__, 1728 td->td_proc->p_pid, instance->li_lock->lo_name, 1729 (*lock_list)->ll_count - 1); 1730 for (j = i; j < (*lock_list)->ll_count - 1; j++) 1731 (*lock_list)->ll_children[j] = 1732 (*lock_list)->ll_children[j + 1]; 1733 (*lock_list)->ll_count--; 1734 intr_restore(s); 1735 1736 /* 1737 * In order to reduce contention on w_mtx, we want to keep always an 1738 * head object into lists so that frequent allocation from the 1739 * free witness pool (and subsequent locking) is avoided. 1740 * In order to maintain the current code simple, when the head 1741 * object is totally unloaded it means also that we do not have 1742 * further objects in the list, so the list ownership needs to be 1743 * hand over to another object if the current head needs to be freed. 1744 */ 1745 if ((*lock_list)->ll_count == 0) { 1746 if (*lock_list == lle) { 1747 if (lle->ll_next == NULL) 1748 return; 1749 } else 1750 lle = *lock_list; 1751 *lock_list = lle->ll_next; 1752 CTR3(KTR_WITNESS, "%s: pid %d removed lle %p", __func__, 1753 td->td_proc->p_pid, lle); 1754 witness_lock_list_free(lle); 1755 } 1756 } 1757 1758 void 1759 witness_thread_exit(struct thread *td) 1760 { 1761 struct lock_list_entry *lle; 1762 int i, n; 1763 1764 lle = td->td_sleeplocks; 1765 if (lle == NULL || KERNEL_PANICKED()) 1766 return; 1767 if (lle->ll_count != 0) { 1768 for (n = 0; lle != NULL; lle = lle->ll_next) 1769 for (i = lle->ll_count - 1; i >= 0; i--) { 1770 if (n == 0) 1771 witness_output( 1772 "Thread %p exiting with the following locks held:\n", td); 1773 n++; 1774 witness_list_lock(&lle->ll_children[i], 1775 witness_output); 1776 1777 } 1778 kassert_panic( 1779 "Thread %p cannot exit while holding sleeplocks\n", td); 1780 } 1781 witness_lock_list_free(lle); 1782 } 1783 1784 /* 1785 * Warn if any locks other than 'lock' are held. Flags can be passed in to 1786 * exempt Giant and sleepable locks from the checks as well. If any 1787 * non-exempt locks are held, then a supplied message is printed to the 1788 * output channel along with a list of the offending locks. If indicated in the 1789 * flags then a failure results in a panic as well. 1790 */ 1791 int 1792 witness_warn(int flags, struct lock_object *lock, const char *fmt, ...) 1793 { 1794 struct lock_list_entry *lock_list, *lle; 1795 struct lock_instance *lock1; 1796 struct thread *td; 1797 va_list ap; 1798 int i, n; 1799 1800 if (witness_cold || witness_watch < 1 || KERNEL_PANICKED()) 1801 return (0); 1802 n = 0; 1803 td = curthread; 1804 for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next) 1805 for (i = lle->ll_count - 1; i >= 0; i--) { 1806 lock1 = &lle->ll_children[i]; 1807 if (lock1->li_lock == lock) 1808 continue; 1809 if (flags & WARN_GIANTOK && 1810 lock1->li_lock == &Giant.lock_object) 1811 continue; 1812 if (flags & WARN_SLEEPOK && 1813 (lock1->li_flags & LI_SLEEPABLE) != 0) 1814 continue; 1815 if (n == 0) { 1816 va_start(ap, fmt); 1817 vprintf(fmt, ap); 1818 va_end(ap); 1819 printf(" with the following %slocks held:\n", 1820 (flags & WARN_SLEEPOK) != 0 ? 1821 "non-sleepable " : ""); 1822 } 1823 n++; 1824 witness_list_lock(lock1, printf); 1825 } 1826 1827 /* 1828 * Pin the thread in order to avoid problems with thread migration. 1829 * Once that all verifies are passed about spinlocks ownership, 1830 * the thread is in a safe path and it can be unpinned. 1831 */ 1832 sched_pin(); 1833 lock_list = PCPU_GET(spinlocks); 1834 if (lock_list != NULL && lock_list->ll_count != 0) { 1835 sched_unpin(); 1836 1837 /* 1838 * We should only have one spinlock and as long as 1839 * the flags cannot match for this locks class, 1840 * check if the first spinlock is the one curthread 1841 * should hold. 1842 */ 1843 lock1 = &lock_list->ll_children[lock_list->ll_count - 1]; 1844 if (lock_list->ll_count == 1 && lock_list->ll_next == NULL && 1845 lock1->li_lock == lock && n == 0) 1846 return (0); 1847 1848 va_start(ap, fmt); 1849 vprintf(fmt, ap); 1850 va_end(ap); 1851 printf(" with the following %slocks held:\n", 1852 (flags & WARN_SLEEPOK) != 0 ? "non-sleepable " : ""); 1853 n += witness_list_locks(&lock_list, printf); 1854 } else 1855 sched_unpin(); 1856 if (flags & WARN_PANIC && n) 1857 kassert_panic("%s", __func__); 1858 else 1859 witness_debugger(n, __func__); 1860 return (n); 1861 } 1862 1863 const char * 1864 witness_file(struct lock_object *lock) 1865 { 1866 struct witness *w; 1867 1868 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL) 1869 return ("?"); 1870 w = lock->lo_witness; 1871 return (w->w_file); 1872 } 1873 1874 int 1875 witness_line(struct lock_object *lock) 1876 { 1877 struct witness *w; 1878 1879 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL) 1880 return (0); 1881 w = lock->lo_witness; 1882 return (w->w_line); 1883 } 1884 1885 static struct witness * 1886 enroll(const char *description, struct lock_class *lock_class) 1887 { 1888 struct witness *w; 1889 1890 MPASS(description != NULL); 1891 1892 if (witness_watch == -1 || KERNEL_PANICKED()) 1893 return (NULL); 1894 if ((lock_class->lc_flags & LC_SPINLOCK)) { 1895 if (witness_skipspin) 1896 return (NULL); 1897 } else if ((lock_class->lc_flags & LC_SLEEPLOCK) == 0) { 1898 kassert_panic("lock class %s is not sleep or spin", 1899 lock_class->lc_name); 1900 return (NULL); 1901 } 1902 1903 mtx_lock_spin(&w_mtx); 1904 w = witness_hash_get(description); 1905 if (w) 1906 goto found; 1907 if ((w = witness_get()) == NULL) 1908 return (NULL); 1909 MPASS(strlen(description) < MAX_W_NAME); 1910 strcpy(w->w_name, description); 1911 w->w_class = lock_class; 1912 w->w_refcount = 1; 1913 STAILQ_INSERT_HEAD(&w_all, w, w_list); 1914 if (lock_class->lc_flags & LC_SPINLOCK) { 1915 STAILQ_INSERT_HEAD(&w_spin, w, w_typelist); 1916 w_spin_cnt++; 1917 } else if (lock_class->lc_flags & LC_SLEEPLOCK) { 1918 STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist); 1919 w_sleep_cnt++; 1920 } 1921 1922 /* Insert new witness into the hash */ 1923 witness_hash_put(w); 1924 witness_increment_graph_generation(); 1925 mtx_unlock_spin(&w_mtx); 1926 return (w); 1927 found: 1928 w->w_refcount++; 1929 if (w->w_refcount == 1) 1930 w->w_class = lock_class; 1931 mtx_unlock_spin(&w_mtx); 1932 if (lock_class != w->w_class) 1933 kassert_panic( 1934 "lock (%s) %s does not match earlier (%s) lock", 1935 description, lock_class->lc_name, 1936 w->w_class->lc_name); 1937 return (w); 1938 } 1939 1940 static void 1941 depart(struct witness *w) 1942 { 1943 MPASS(w->w_refcount == 0); 1944 if (w->w_class->lc_flags & LC_SLEEPLOCK) { 1945 w_sleep_cnt--; 1946 } else { 1947 w_spin_cnt--; 1948 } 1949 /* 1950 * Set file to NULL as it may point into a loadable module. 1951 */ 1952 w->w_file = NULL; 1953 w->w_line = 0; 1954 witness_increment_graph_generation(); 1955 } 1956 1957 static void 1958 adopt(struct witness *parent, struct witness *child) 1959 { 1960 int pi, ci, i, j; 1961 1962 if (witness_cold == 0) 1963 mtx_assert(&w_mtx, MA_OWNED); 1964 1965 /* If the relationship is already known, there's no work to be done. */ 1966 if (isitmychild(parent, child)) 1967 return; 1968 1969 /* When the structure of the graph changes, bump up the generation. */ 1970 witness_increment_graph_generation(); 1971 1972 /* 1973 * The hard part ... create the direct relationship, then propagate all 1974 * indirect relationships. 1975 */ 1976 pi = parent->w_index; 1977 ci = child->w_index; 1978 WITNESS_INDEX_ASSERT(pi); 1979 WITNESS_INDEX_ASSERT(ci); 1980 MPASS(pi != ci); 1981 w_rmatrix[pi][ci] |= WITNESS_PARENT; 1982 w_rmatrix[ci][pi] |= WITNESS_CHILD; 1983 1984 /* 1985 * If parent was not already an ancestor of child, 1986 * then we increment the descendant and ancestor counters. 1987 */ 1988 if ((w_rmatrix[pi][ci] & WITNESS_ANCESTOR) == 0) { 1989 parent->w_num_descendants++; 1990 child->w_num_ancestors++; 1991 } 1992 1993 /* 1994 * Find each ancestor of 'pi'. Note that 'pi' itself is counted as 1995 * an ancestor of 'pi' during this loop. 1996 */ 1997 for (i = 1; i <= w_max_used_index; i++) { 1998 if ((w_rmatrix[i][pi] & WITNESS_ANCESTOR_MASK) == 0 && 1999 (i != pi)) 2000 continue; 2001 2002 /* Find each descendant of 'i' and mark it as a descendant. */ 2003 for (j = 1; j <= w_max_used_index; j++) { 2004 /* 2005 * Skip children that are already marked as 2006 * descendants of 'i'. 2007 */ 2008 if (w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK) 2009 continue; 2010 2011 /* 2012 * We are only interested in descendants of 'ci'. Note 2013 * that 'ci' itself is counted as a descendant of 'ci'. 2014 */ 2015 if ((w_rmatrix[ci][j] & WITNESS_ANCESTOR_MASK) == 0 && 2016 (j != ci)) 2017 continue; 2018 w_rmatrix[i][j] |= WITNESS_ANCESTOR; 2019 w_rmatrix[j][i] |= WITNESS_DESCENDANT; 2020 w_data[i].w_num_descendants++; 2021 w_data[j].w_num_ancestors++; 2022 2023 /* 2024 * Make sure we aren't marking a node as both an 2025 * ancestor and descendant. We should have caught 2026 * this as a lock order reversal earlier. 2027 */ 2028 if ((w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK) && 2029 (w_rmatrix[i][j] & WITNESS_DESCENDANT_MASK)) { 2030 printf("witness rmatrix paradox! [%d][%d]=%d " 2031 "both ancestor and descendant\n", 2032 i, j, w_rmatrix[i][j]); 2033 kdb_backtrace(); 2034 printf("Witness disabled.\n"); 2035 witness_watch = -1; 2036 } 2037 if ((w_rmatrix[j][i] & WITNESS_ANCESTOR_MASK) && 2038 (w_rmatrix[j][i] & WITNESS_DESCENDANT_MASK)) { 2039 printf("witness rmatrix paradox! [%d][%d]=%d " 2040 "both ancestor and descendant\n", 2041 j, i, w_rmatrix[j][i]); 2042 kdb_backtrace(); 2043 printf("Witness disabled.\n"); 2044 witness_watch = -1; 2045 } 2046 } 2047 } 2048 } 2049 2050 static void 2051 itismychild(struct witness *parent, struct witness *child) 2052 { 2053 int unlocked; 2054 2055 MPASS(child != NULL && parent != NULL); 2056 if (witness_cold == 0) 2057 mtx_assert(&w_mtx, MA_OWNED); 2058 2059 if (!witness_lock_type_equal(parent, child)) { 2060 if (witness_cold == 0) { 2061 unlocked = 1; 2062 mtx_unlock_spin(&w_mtx); 2063 } else { 2064 unlocked = 0; 2065 } 2066 kassert_panic( 2067 "%s: parent \"%s\" (%s) and child \"%s\" (%s) are not " 2068 "the same lock type", __func__, parent->w_name, 2069 parent->w_class->lc_name, child->w_name, 2070 child->w_class->lc_name); 2071 if (unlocked) 2072 mtx_lock_spin(&w_mtx); 2073 } 2074 adopt(parent, child); 2075 } 2076 2077 /* 2078 * Generic code for the isitmy*() functions. The rmask parameter is the 2079 * expected relationship of w1 to w2. 2080 */ 2081 static int 2082 _isitmyx(struct witness *w1, struct witness *w2, int rmask, const char *fname) 2083 { 2084 unsigned char r1, r2; 2085 int i1, i2; 2086 2087 i1 = w1->w_index; 2088 i2 = w2->w_index; 2089 WITNESS_INDEX_ASSERT(i1); 2090 WITNESS_INDEX_ASSERT(i2); 2091 r1 = w_rmatrix[i1][i2] & WITNESS_RELATED_MASK; 2092 r2 = w_rmatrix[i2][i1] & WITNESS_RELATED_MASK; 2093 2094 /* The flags on one better be the inverse of the flags on the other */ 2095 if (!((WITNESS_ATOD(r1) == r2 && WITNESS_DTOA(r2) == r1) || 2096 (WITNESS_DTOA(r1) == r2 && WITNESS_ATOD(r2) == r1))) { 2097 /* Don't squawk if we're potentially racing with an update. */ 2098 if (!mtx_owned(&w_mtx)) 2099 return (0); 2100 printf("%s: rmatrix mismatch between %s (index %d) and %s " 2101 "(index %d): w_rmatrix[%d][%d] == %hhx but " 2102 "w_rmatrix[%d][%d] == %hhx\n", 2103 fname, w1->w_name, i1, w2->w_name, i2, i1, i2, r1, 2104 i2, i1, r2); 2105 kdb_backtrace(); 2106 printf("Witness disabled.\n"); 2107 witness_watch = -1; 2108 } 2109 return (r1 & rmask); 2110 } 2111 2112 /* 2113 * Checks if @child is a direct child of @parent. 2114 */ 2115 static int 2116 isitmychild(struct witness *parent, struct witness *child) 2117 { 2118 return (_isitmyx(parent, child, WITNESS_PARENT, __func__)); 2119 } 2120 2121 /* 2122 * Checks if @descendant is a direct or inderect descendant of @ancestor. 2123 */ 2124 static int 2125 isitmydescendant(struct witness *ancestor, struct witness *descendant) 2126 { 2127 return (_isitmyx(ancestor, descendant, WITNESS_ANCESTOR_MASK, 2128 __func__)); 2129 } 2130 2131 static int 2132 blessed(struct witness *w1, struct witness *w2) 2133 { 2134 int i; 2135 struct witness_blessed *b; 2136 2137 for (i = 0; i < nitems(blessed_list); i++) { 2138 b = &blessed_list[i]; 2139 if (strcmp(w1->w_name, b->b_lock1) == 0) { 2140 if (strcmp(w2->w_name, b->b_lock2) == 0) 2141 return (1); 2142 continue; 2143 } 2144 if (strcmp(w1->w_name, b->b_lock2) == 0) 2145 if (strcmp(w2->w_name, b->b_lock1) == 0) 2146 return (1); 2147 } 2148 return (0); 2149 } 2150 2151 static struct witness * 2152 witness_get(void) 2153 { 2154 struct witness *w; 2155 int index; 2156 2157 if (witness_cold == 0) 2158 mtx_assert(&w_mtx, MA_OWNED); 2159 2160 if (witness_watch == -1) { 2161 mtx_unlock_spin(&w_mtx); 2162 return (NULL); 2163 } 2164 if (STAILQ_EMPTY(&w_free)) { 2165 witness_watch = -1; 2166 mtx_unlock_spin(&w_mtx); 2167 printf("WITNESS: unable to allocate a new witness object\n"); 2168 return (NULL); 2169 } 2170 w = STAILQ_FIRST(&w_free); 2171 STAILQ_REMOVE_HEAD(&w_free, w_list); 2172 w_free_cnt--; 2173 index = w->w_index; 2174 MPASS(index > 0 && index == w_max_used_index + 1 && 2175 index < witness_count); 2176 bzero(w, sizeof(*w)); 2177 w->w_index = index; 2178 if (index > w_max_used_index) 2179 w_max_used_index = index; 2180 return (w); 2181 } 2182 2183 static void 2184 witness_free(struct witness *w) 2185 { 2186 STAILQ_INSERT_HEAD(&w_free, w, w_list); 2187 w_free_cnt++; 2188 } 2189 2190 static struct lock_list_entry * 2191 witness_lock_list_get(void) 2192 { 2193 struct lock_list_entry *lle; 2194 2195 if (witness_watch == -1) 2196 return (NULL); 2197 mtx_lock_spin(&w_mtx); 2198 lle = w_lock_list_free; 2199 if (lle == NULL) { 2200 witness_watch = -1; 2201 mtx_unlock_spin(&w_mtx); 2202 printf("%s: witness exhausted\n", __func__); 2203 return (NULL); 2204 } 2205 w_lock_list_free = lle->ll_next; 2206 mtx_unlock_spin(&w_mtx); 2207 bzero(lle, sizeof(*lle)); 2208 return (lle); 2209 } 2210 2211 static void 2212 witness_lock_list_free(struct lock_list_entry *lle) 2213 { 2214 mtx_lock_spin(&w_mtx); 2215 lle->ll_next = w_lock_list_free; 2216 w_lock_list_free = lle; 2217 mtx_unlock_spin(&w_mtx); 2218 } 2219 2220 static struct lock_instance * 2221 find_instance(struct lock_list_entry *list, const struct lock_object *lock) 2222 { 2223 struct lock_list_entry *lle; 2224 struct lock_instance *instance; 2225 int i; 2226 2227 for (lle = list; lle != NULL; lle = lle->ll_next) 2228 for (i = lle->ll_count - 1; i >= 0; i--) { 2229 instance = &lle->ll_children[i]; 2230 if (instance->li_lock == lock) 2231 return (instance); 2232 } 2233 return (NULL); 2234 } 2235 2236 static void 2237 witness_list_lock(struct lock_instance *instance, 2238 int (*prnt)(const char *fmt, ...)) 2239 { 2240 struct lock_object *lock; 2241 2242 lock = instance->li_lock; 2243 prnt("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ? 2244 "exclusive" : "shared", LOCK_CLASS(lock)->lc_name, lock->lo_name); 2245 if (lock->lo_witness->w_name != lock->lo_name) 2246 prnt(" (%s)", lock->lo_witness->w_name); 2247 prnt(" r = %d (%p) locked @ %s:%d\n", 2248 instance->li_flags & LI_RECURSEMASK, lock, 2249 fixup_filename(instance->li_file), instance->li_line); 2250 } 2251 2252 static int 2253 witness_output(const char *fmt, ...) 2254 { 2255 va_list ap; 2256 int ret; 2257 2258 va_start(ap, fmt); 2259 ret = witness_voutput(fmt, ap); 2260 va_end(ap); 2261 return (ret); 2262 } 2263 2264 static int 2265 witness_voutput(const char *fmt, va_list ap) 2266 { 2267 int ret; 2268 2269 ret = 0; 2270 switch (witness_channel) { 2271 case WITNESS_CONSOLE: 2272 ret = vprintf(fmt, ap); 2273 break; 2274 case WITNESS_LOG: 2275 vlog(LOG_NOTICE, fmt, ap); 2276 break; 2277 case WITNESS_NONE: 2278 break; 2279 } 2280 return (ret); 2281 } 2282 2283 #ifdef DDB 2284 static int 2285 witness_thread_has_locks(struct thread *td) 2286 { 2287 if (td->td_sleeplocks == NULL) 2288 return (0); 2289 return (td->td_sleeplocks->ll_count != 0); 2290 } 2291 2292 static int 2293 witness_proc_has_locks(struct proc *p) 2294 { 2295 struct thread *td; 2296 2297 FOREACH_THREAD_IN_PROC(p, td) { 2298 if (witness_thread_has_locks(td)) 2299 return (1); 2300 } 2301 return (0); 2302 } 2303 #endif 2304 2305 int 2306 witness_list_locks(struct lock_list_entry **lock_list, 2307 int (*prnt)(const char *fmt, ...)) 2308 { 2309 struct lock_list_entry *lle; 2310 int i, nheld; 2311 2312 nheld = 0; 2313 for (lle = *lock_list; lle != NULL; lle = lle->ll_next) 2314 for (i = lle->ll_count - 1; i >= 0; i--) { 2315 witness_list_lock(&lle->ll_children[i], prnt); 2316 nheld++; 2317 } 2318 return (nheld); 2319 } 2320 2321 /* 2322 * This is a bit risky at best. We call this function when we have timed 2323 * out acquiring a spin lock, and we assume that the other CPU is stuck 2324 * with this lock held. So, we go groveling around in the other CPU's 2325 * per-cpu data to try to find the lock instance for this spin lock to 2326 * see when it was last acquired. 2327 */ 2328 void 2329 witness_display_spinlock(struct lock_object *lock, struct thread *owner, 2330 int (*prnt)(const char *fmt, ...)) 2331 { 2332 struct lock_instance *instance; 2333 struct pcpu *pc; 2334 2335 if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU) 2336 return; 2337 pc = pcpu_find(owner->td_oncpu); 2338 instance = find_instance(pc->pc_spinlocks, lock); 2339 if (instance != NULL) 2340 witness_list_lock(instance, prnt); 2341 } 2342 2343 void 2344 witness_save(struct lock_object *lock, const char **filep, int *linep) 2345 { 2346 struct lock_list_entry *lock_list; 2347 struct lock_instance *instance; 2348 struct lock_class *class; 2349 2350 /* Initialize for KMSAN's benefit. */ 2351 *filep = NULL; 2352 *linep = 0; 2353 2354 /* 2355 * This function is used independently in locking code to deal with 2356 * Giant, SCHEDULER_STOPPED() check can be removed here after Giant 2357 * is gone. 2358 */ 2359 if (SCHEDULER_STOPPED()) 2360 return; 2361 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__)); 2362 if (lock->lo_witness == NULL || witness_watch == -1 || KERNEL_PANICKED()) 2363 return; 2364 class = LOCK_CLASS(lock); 2365 if (class->lc_flags & LC_SLEEPLOCK) 2366 lock_list = curthread->td_sleeplocks; 2367 else { 2368 if (witness_skipspin) 2369 return; 2370 lock_list = PCPU_GET(spinlocks); 2371 } 2372 instance = find_instance(lock_list, lock); 2373 if (instance == NULL) { 2374 kassert_panic("%s: lock (%s) %s not locked", __func__, 2375 class->lc_name, lock->lo_name); 2376 return; 2377 } 2378 *filep = instance->li_file; 2379 *linep = instance->li_line; 2380 } 2381 2382 void 2383 witness_restore(struct lock_object *lock, const char *file, int line) 2384 { 2385 struct lock_list_entry *lock_list; 2386 struct lock_instance *instance; 2387 struct lock_class *class; 2388 2389 /* 2390 * This function is used independently in locking code to deal with 2391 * Giant, SCHEDULER_STOPPED() check can be removed here after Giant 2392 * is gone. 2393 */ 2394 if (SCHEDULER_STOPPED()) 2395 return; 2396 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__)); 2397 if (lock->lo_witness == NULL || witness_watch == -1 || KERNEL_PANICKED()) 2398 return; 2399 class = LOCK_CLASS(lock); 2400 if (class->lc_flags & LC_SLEEPLOCK) 2401 lock_list = curthread->td_sleeplocks; 2402 else { 2403 if (witness_skipspin) 2404 return; 2405 lock_list = PCPU_GET(spinlocks); 2406 } 2407 instance = find_instance(lock_list, lock); 2408 if (instance == NULL) 2409 kassert_panic("%s: lock (%s) %s not locked", __func__, 2410 class->lc_name, lock->lo_name); 2411 lock->lo_witness->w_file = file; 2412 lock->lo_witness->w_line = line; 2413 if (instance == NULL) 2414 return; 2415 instance->li_file = file; 2416 instance->li_line = line; 2417 } 2418 2419 static bool 2420 witness_find_instance(const struct lock_object *lock, 2421 struct lock_instance **instance) 2422 { 2423 #ifdef INVARIANT_SUPPORT 2424 struct lock_class *class; 2425 2426 if (lock->lo_witness == NULL || witness_watch < 1 || KERNEL_PANICKED()) 2427 return (false); 2428 class = LOCK_CLASS(lock); 2429 if ((class->lc_flags & LC_SLEEPLOCK) != 0) { 2430 *instance = find_instance(curthread->td_sleeplocks, lock); 2431 return (true); 2432 } else if ((class->lc_flags & LC_SPINLOCK) != 0) { 2433 *instance = find_instance(PCPU_GET(spinlocks), lock); 2434 return (true); 2435 } else { 2436 kassert_panic("Lock (%s) %s is not sleep or spin!", 2437 class->lc_name, lock->lo_name); 2438 return (false); 2439 } 2440 #else 2441 return (false); 2442 #endif 2443 } 2444 2445 void 2446 witness_assert(const struct lock_object *lock, int flags, const char *file, 2447 int line) 2448 { 2449 #ifdef INVARIANT_SUPPORT 2450 struct lock_instance *instance; 2451 struct lock_class *class; 2452 2453 if (!witness_find_instance(lock, &instance)) 2454 return; 2455 class = LOCK_CLASS(lock); 2456 switch (flags) { 2457 case LA_UNLOCKED: 2458 if (instance != NULL) 2459 kassert_panic("Lock (%s) %s locked @ %s:%d.", 2460 class->lc_name, lock->lo_name, 2461 fixup_filename(file), line); 2462 break; 2463 case LA_LOCKED: 2464 case LA_LOCKED | LA_RECURSED: 2465 case LA_LOCKED | LA_NOTRECURSED: 2466 case LA_SLOCKED: 2467 case LA_SLOCKED | LA_RECURSED: 2468 case LA_SLOCKED | LA_NOTRECURSED: 2469 case LA_XLOCKED: 2470 case LA_XLOCKED | LA_RECURSED: 2471 case LA_XLOCKED | LA_NOTRECURSED: 2472 if (instance == NULL) { 2473 kassert_panic("Lock (%s) %s not locked @ %s:%d.", 2474 class->lc_name, lock->lo_name, 2475 fixup_filename(file), line); 2476 break; 2477 } 2478 if ((flags & LA_XLOCKED) != 0 && 2479 (instance->li_flags & LI_EXCLUSIVE) == 0) 2480 kassert_panic( 2481 "Lock (%s) %s not exclusively locked @ %s:%d.", 2482 class->lc_name, lock->lo_name, 2483 fixup_filename(file), line); 2484 if ((flags & LA_SLOCKED) != 0 && 2485 (instance->li_flags & LI_EXCLUSIVE) != 0) 2486 kassert_panic( 2487 "Lock (%s) %s exclusively locked @ %s:%d.", 2488 class->lc_name, lock->lo_name, 2489 fixup_filename(file), line); 2490 if ((flags & LA_RECURSED) != 0 && 2491 (instance->li_flags & LI_RECURSEMASK) == 0) 2492 kassert_panic("Lock (%s) %s not recursed @ %s:%d.", 2493 class->lc_name, lock->lo_name, 2494 fixup_filename(file), line); 2495 if ((flags & LA_NOTRECURSED) != 0 && 2496 (instance->li_flags & LI_RECURSEMASK) != 0) 2497 kassert_panic("Lock (%s) %s recursed @ %s:%d.", 2498 class->lc_name, lock->lo_name, 2499 fixup_filename(file), line); 2500 break; 2501 default: 2502 kassert_panic("Invalid lock assertion at %s:%d.", 2503 fixup_filename(file), line); 2504 } 2505 #endif /* INVARIANT_SUPPORT */ 2506 } 2507 2508 /* 2509 * Checks the ownership of the lock by curthread, consulting the witness list. 2510 * Returns: 2511 * 0 if witness is disabled or did not work 2512 * -1 if not owned 2513 * 1 if owned 2514 */ 2515 int 2516 witness_is_owned(const struct lock_object *lock) 2517 { 2518 #ifdef INVARIANT_SUPPORT 2519 struct lock_instance *instance; 2520 2521 if (!witness_find_instance(lock, &instance)) 2522 return (0); 2523 return (instance == NULL ? -1 : 1); 2524 #else 2525 return (0); 2526 #endif 2527 } 2528 2529 static void 2530 witness_setflag(struct lock_object *lock, int flag, int set) 2531 { 2532 struct lock_list_entry *lock_list; 2533 struct lock_instance *instance; 2534 struct lock_class *class; 2535 2536 if (lock->lo_witness == NULL || witness_watch == -1 || KERNEL_PANICKED()) 2537 return; 2538 class = LOCK_CLASS(lock); 2539 if (class->lc_flags & LC_SLEEPLOCK) 2540 lock_list = curthread->td_sleeplocks; 2541 else { 2542 if (witness_skipspin) 2543 return; 2544 lock_list = PCPU_GET(spinlocks); 2545 } 2546 instance = find_instance(lock_list, lock); 2547 if (instance == NULL) { 2548 kassert_panic("%s: lock (%s) %s not locked", __func__, 2549 class->lc_name, lock->lo_name); 2550 return; 2551 } 2552 2553 if (set) 2554 instance->li_flags |= flag; 2555 else 2556 instance->li_flags &= ~flag; 2557 } 2558 2559 void 2560 witness_norelease(struct lock_object *lock) 2561 { 2562 witness_setflag(lock, LI_NORELEASE, 1); 2563 } 2564 2565 void 2566 witness_releaseok(struct lock_object *lock) 2567 { 2568 witness_setflag(lock, LI_NORELEASE, 0); 2569 } 2570 2571 #ifdef DDB 2572 static void 2573 witness_ddb_list(struct thread *td) 2574 { 2575 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__)); 2576 KASSERT(kdb_active, ("%s: not in the debugger", __func__)); 2577 2578 if (witness_watch < 1) 2579 return; 2580 2581 witness_list_locks(&td->td_sleeplocks, db_printf); 2582 2583 /* 2584 * We only handle spinlocks if td == curthread. This is somewhat broken 2585 * if td is currently executing on some other CPU and holds spin locks 2586 * as we won't display those locks. If we had a MI way of getting 2587 * the per-cpu data for a given cpu then we could use 2588 * td->td_oncpu to get the list of spinlocks for this thread 2589 * and "fix" this. 2590 * 2591 * That still wouldn't really fix this unless we locked the scheduler 2592 * lock or stopped the other CPU to make sure it wasn't changing the 2593 * list out from under us. It is probably best to just not try to 2594 * handle threads on other CPU's for now. 2595 */ 2596 if (td == curthread && PCPU_GET(spinlocks) != NULL) 2597 witness_list_locks(PCPU_PTR(spinlocks), db_printf); 2598 } 2599 2600 DB_SHOW_COMMAND(locks, db_witness_list) 2601 { 2602 struct thread *td; 2603 2604 if (have_addr) 2605 td = db_lookup_thread(addr, true); 2606 else 2607 td = kdb_thread; 2608 witness_ddb_list(td); 2609 } 2610 2611 DB_SHOW_ALL_COMMAND(locks, db_witness_list_all) 2612 { 2613 struct thread *td; 2614 struct proc *p; 2615 2616 /* 2617 * It would be nice to list only threads and processes that actually 2618 * held sleep locks, but that information is currently not exported 2619 * by WITNESS. 2620 */ 2621 FOREACH_PROC_IN_SYSTEM(p) { 2622 if (!witness_proc_has_locks(p)) 2623 continue; 2624 FOREACH_THREAD_IN_PROC(p, td) { 2625 if (!witness_thread_has_locks(td)) 2626 continue; 2627 db_printf("Process %d (%s) thread %p (%d)\n", p->p_pid, 2628 p->p_comm, td, td->td_tid); 2629 witness_ddb_list(td); 2630 if (db_pager_quit) 2631 return; 2632 } 2633 } 2634 } 2635 DB_SHOW_ALIAS_FLAGS(alllocks, db_witness_list_all, DB_CMD_MEMSAFE); 2636 2637 DB_SHOW_COMMAND_FLAGS(witness, db_witness_display, DB_CMD_MEMSAFE) 2638 { 2639 witness_ddb_display(db_printf); 2640 } 2641 #endif 2642 2643 static void 2644 sbuf_print_witness_badstacks(struct sbuf *sb, size_t *oldidx) 2645 { 2646 struct witness_lock_order_data *data1, *data2, *tmp_data1, *tmp_data2; 2647 struct witness *tmp_w1, *tmp_w2, *w1, *w2; 2648 int generation, i, j; 2649 2650 tmp_data1 = NULL; 2651 tmp_data2 = NULL; 2652 tmp_w1 = NULL; 2653 tmp_w2 = NULL; 2654 2655 /* Allocate and init temporary storage space. */ 2656 tmp_w1 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO); 2657 tmp_w2 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO); 2658 tmp_data1 = malloc(sizeof(struct witness_lock_order_data), M_TEMP, 2659 M_WAITOK | M_ZERO); 2660 tmp_data2 = malloc(sizeof(struct witness_lock_order_data), M_TEMP, 2661 M_WAITOK | M_ZERO); 2662 stack_zero(&tmp_data1->wlod_stack); 2663 stack_zero(&tmp_data2->wlod_stack); 2664 2665 restart: 2666 mtx_lock_spin(&w_mtx); 2667 generation = w_generation; 2668 mtx_unlock_spin(&w_mtx); 2669 sbuf_printf(sb, "Number of known direct relationships is %d\n", 2670 w_lohash.wloh_count); 2671 for (i = 1; i < w_max_used_index; i++) { 2672 mtx_lock_spin(&w_mtx); 2673 if (generation != w_generation) { 2674 mtx_unlock_spin(&w_mtx); 2675 2676 /* The graph has changed, try again. */ 2677 *oldidx = 0; 2678 sbuf_clear(sb); 2679 goto restart; 2680 } 2681 2682 w1 = &w_data[i]; 2683 if (w1->w_reversed == 0) { 2684 mtx_unlock_spin(&w_mtx); 2685 continue; 2686 } 2687 2688 /* Copy w1 locally so we can release the spin lock. */ 2689 *tmp_w1 = *w1; 2690 mtx_unlock_spin(&w_mtx); 2691 2692 if (tmp_w1->w_reversed == 0) 2693 continue; 2694 for (j = 1; j < w_max_used_index; j++) { 2695 if ((w_rmatrix[i][j] & WITNESS_REVERSAL) == 0 || i > j) 2696 continue; 2697 2698 mtx_lock_spin(&w_mtx); 2699 if (generation != w_generation) { 2700 mtx_unlock_spin(&w_mtx); 2701 2702 /* The graph has changed, try again. */ 2703 *oldidx = 0; 2704 sbuf_clear(sb); 2705 goto restart; 2706 } 2707 2708 w2 = &w_data[j]; 2709 data1 = witness_lock_order_get(w1, w2); 2710 data2 = witness_lock_order_get(w2, w1); 2711 2712 /* 2713 * Copy information locally so we can release the 2714 * spin lock. 2715 */ 2716 *tmp_w2 = *w2; 2717 2718 if (data1) { 2719 stack_zero(&tmp_data1->wlod_stack); 2720 stack_copy(&data1->wlod_stack, 2721 &tmp_data1->wlod_stack); 2722 } 2723 if (data2 && data2 != data1) { 2724 stack_zero(&tmp_data2->wlod_stack); 2725 stack_copy(&data2->wlod_stack, 2726 &tmp_data2->wlod_stack); 2727 } 2728 mtx_unlock_spin(&w_mtx); 2729 2730 if (blessed(tmp_w1, tmp_w2)) 2731 continue; 2732 2733 sbuf_printf(sb, 2734 "\nLock order reversal between \"%s\"(%s) and \"%s\"(%s)!\n", 2735 tmp_w1->w_name, tmp_w1->w_class->lc_name, 2736 tmp_w2->w_name, tmp_w2->w_class->lc_name); 2737 if (data1) { 2738 sbuf_printf(sb, 2739 "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n", 2740 tmp_w1->w_name, tmp_w1->w_class->lc_name, 2741 tmp_w2->w_name, tmp_w2->w_class->lc_name); 2742 stack_sbuf_print(sb, &tmp_data1->wlod_stack); 2743 sbuf_putc(sb, '\n'); 2744 } 2745 if (data2 && data2 != data1) { 2746 sbuf_printf(sb, 2747 "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n", 2748 tmp_w2->w_name, tmp_w2->w_class->lc_name, 2749 tmp_w1->w_name, tmp_w1->w_class->lc_name); 2750 stack_sbuf_print(sb, &tmp_data2->wlod_stack); 2751 sbuf_putc(sb, '\n'); 2752 } 2753 } 2754 } 2755 mtx_lock_spin(&w_mtx); 2756 if (generation != w_generation) { 2757 mtx_unlock_spin(&w_mtx); 2758 2759 /* 2760 * The graph changed while we were printing stack data, 2761 * try again. 2762 */ 2763 *oldidx = 0; 2764 sbuf_clear(sb); 2765 goto restart; 2766 } 2767 mtx_unlock_spin(&w_mtx); 2768 2769 /* Free temporary storage space. */ 2770 free(tmp_data1, M_TEMP); 2771 free(tmp_data2, M_TEMP); 2772 free(tmp_w1, M_TEMP); 2773 free(tmp_w2, M_TEMP); 2774 } 2775 2776 static int 2777 sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS) 2778 { 2779 struct sbuf *sb; 2780 int error; 2781 2782 if (witness_watch < 1) { 2783 error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning)); 2784 return (error); 2785 } 2786 if (witness_cold) { 2787 error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold)); 2788 return (error); 2789 } 2790 error = 0; 2791 sb = sbuf_new(NULL, NULL, badstack_sbuf_size, SBUF_AUTOEXTEND); 2792 if (sb == NULL) 2793 return (ENOMEM); 2794 2795 sbuf_print_witness_badstacks(sb, &req->oldidx); 2796 2797 sbuf_finish(sb); 2798 error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1); 2799 sbuf_delete(sb); 2800 2801 return (error); 2802 } 2803 2804 #ifdef DDB 2805 static int 2806 sbuf_db_printf_drain(void *arg __unused, const char *data, int len) 2807 { 2808 return (db_printf("%.*s", len, data)); 2809 } 2810 2811 DB_SHOW_COMMAND_FLAGS(badstacks, db_witness_badstacks, DB_CMD_MEMSAFE) 2812 { 2813 struct sbuf sb; 2814 char buffer[128]; 2815 size_t dummy; 2816 2817 sbuf_new(&sb, buffer, sizeof(buffer), SBUF_FIXEDLEN); 2818 sbuf_set_drain(&sb, sbuf_db_printf_drain, NULL); 2819 sbuf_print_witness_badstacks(&sb, &dummy); 2820 sbuf_finish(&sb); 2821 } 2822 #endif 2823 2824 static int 2825 sysctl_debug_witness_channel(SYSCTL_HANDLER_ARGS) 2826 { 2827 static const struct { 2828 enum witness_channel channel; 2829 const char *name; 2830 } channels[] = { 2831 { WITNESS_CONSOLE, "console" }, 2832 { WITNESS_LOG, "log" }, 2833 { WITNESS_NONE, "none" }, 2834 }; 2835 char buf[16]; 2836 u_int i; 2837 int error; 2838 2839 buf[0] = '\0'; 2840 for (i = 0; i < nitems(channels); i++) 2841 if (witness_channel == channels[i].channel) { 2842 snprintf(buf, sizeof(buf), "%s", channels[i].name); 2843 break; 2844 } 2845 2846 error = sysctl_handle_string(oidp, buf, sizeof(buf), req); 2847 if (error != 0 || req->newptr == NULL) 2848 return (error); 2849 2850 error = EINVAL; 2851 for (i = 0; i < nitems(channels); i++) 2852 if (strcmp(channels[i].name, buf) == 0) { 2853 witness_channel = channels[i].channel; 2854 error = 0; 2855 break; 2856 } 2857 return (error); 2858 } 2859 2860 static int 2861 sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS) 2862 { 2863 struct witness *w; 2864 struct sbuf *sb; 2865 int error; 2866 2867 #ifdef __i386__ 2868 error = SYSCTL_OUT(req, w_notallowed, sizeof(w_notallowed)); 2869 return (error); 2870 #endif 2871 2872 if (witness_watch < 1) { 2873 error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning)); 2874 return (error); 2875 } 2876 if (witness_cold) { 2877 error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold)); 2878 return (error); 2879 } 2880 error = 0; 2881 2882 error = sysctl_wire_old_buffer(req, 0); 2883 if (error != 0) 2884 return (error); 2885 sb = sbuf_new_for_sysctl(NULL, NULL, FULLGRAPH_SBUF_SIZE, req); 2886 if (sb == NULL) 2887 return (ENOMEM); 2888 sbuf_putc(sb, '\n'); 2889 2890 mtx_lock_spin(&w_mtx); 2891 STAILQ_FOREACH(w, &w_all, w_list) 2892 w->w_displayed = 0; 2893 STAILQ_FOREACH(w, &w_all, w_list) 2894 witness_add_fullgraph(sb, w); 2895 mtx_unlock_spin(&w_mtx); 2896 2897 /* 2898 * Close the sbuf and return to userland. 2899 */ 2900 error = sbuf_finish(sb); 2901 sbuf_delete(sb); 2902 2903 return (error); 2904 } 2905 2906 static int 2907 sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS) 2908 { 2909 int error, value; 2910 2911 value = witness_watch; 2912 error = sysctl_handle_int(oidp, &value, 0, req); 2913 if (error != 0 || req->newptr == NULL) 2914 return (error); 2915 if (value > 1 || value < -1 || 2916 (witness_watch == -1 && value != witness_watch)) 2917 return (EINVAL); 2918 witness_watch = value; 2919 return (0); 2920 } 2921 2922 static void 2923 witness_add_fullgraph(struct sbuf *sb, struct witness *w) 2924 { 2925 int i; 2926 2927 if (w->w_displayed != 0 || (w->w_file == NULL && w->w_line == 0)) 2928 return; 2929 w->w_displayed = 1; 2930 2931 WITNESS_INDEX_ASSERT(w->w_index); 2932 for (i = 1; i <= w_max_used_index; i++) { 2933 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT) { 2934 sbuf_printf(sb, "\"%s\",\"%s\"\n", w->w_name, 2935 w_data[i].w_name); 2936 witness_add_fullgraph(sb, &w_data[i]); 2937 } 2938 } 2939 } 2940 2941 /* 2942 * A simple hash function. Takes a key pointer and a key size. If size == 0, 2943 * interprets the key as a string and reads until the null 2944 * terminator. Otherwise, reads the first size bytes. Returns an unsigned 32-bit 2945 * hash value computed from the key. 2946 */ 2947 static uint32_t 2948 witness_hash_djb2(const uint8_t *key, uint32_t size) 2949 { 2950 unsigned int hash = 5381; 2951 int i; 2952 2953 /* hash = hash * 33 + key[i] */ 2954 if (size) 2955 for (i = 0; i < size; i++) 2956 hash = ((hash << 5) + hash) + (unsigned int)key[i]; 2957 else 2958 for (i = 0; key[i] != 0; i++) 2959 hash = ((hash << 5) + hash) + (unsigned int)key[i]; 2960 2961 return (hash); 2962 } 2963 2964 /* 2965 * Initializes the two witness hash tables. Called exactly once from 2966 * witness_initialize(). 2967 */ 2968 static void 2969 witness_init_hash_tables(void) 2970 { 2971 int i; 2972 2973 MPASS(witness_cold); 2974 2975 /* Initialize the hash tables. */ 2976 for (i = 0; i < WITNESS_HASH_SIZE; i++) 2977 w_hash.wh_array[i] = NULL; 2978 2979 w_hash.wh_size = WITNESS_HASH_SIZE; 2980 w_hash.wh_count = 0; 2981 2982 /* Initialize the lock order data hash. */ 2983 w_lofree = NULL; 2984 for (i = 0; i < WITNESS_LO_DATA_COUNT; i++) { 2985 memset(&w_lodata[i], 0, sizeof(w_lodata[i])); 2986 w_lodata[i].wlod_next = w_lofree; 2987 w_lofree = &w_lodata[i]; 2988 } 2989 w_lohash.wloh_size = WITNESS_LO_HASH_SIZE; 2990 w_lohash.wloh_count = 0; 2991 for (i = 0; i < WITNESS_LO_HASH_SIZE; i++) 2992 w_lohash.wloh_array[i] = NULL; 2993 } 2994 2995 static struct witness * 2996 witness_hash_get(const char *key) 2997 { 2998 struct witness *w; 2999 uint32_t hash; 3000 3001 MPASS(key != NULL); 3002 if (witness_cold == 0) 3003 mtx_assert(&w_mtx, MA_OWNED); 3004 hash = witness_hash_djb2(key, 0) % w_hash.wh_size; 3005 w = w_hash.wh_array[hash]; 3006 while (w != NULL) { 3007 if (strcmp(w->w_name, key) == 0) 3008 goto out; 3009 w = w->w_hash_next; 3010 } 3011 3012 out: 3013 return (w); 3014 } 3015 3016 static void 3017 witness_hash_put(struct witness *w) 3018 { 3019 uint32_t hash; 3020 3021 MPASS(w != NULL); 3022 MPASS(w->w_name != NULL); 3023 if (witness_cold == 0) 3024 mtx_assert(&w_mtx, MA_OWNED); 3025 KASSERT(witness_hash_get(w->w_name) == NULL, 3026 ("%s: trying to add a hash entry that already exists!", __func__)); 3027 KASSERT(w->w_hash_next == NULL, 3028 ("%s: w->w_hash_next != NULL", __func__)); 3029 3030 hash = witness_hash_djb2(w->w_name, 0) % w_hash.wh_size; 3031 w->w_hash_next = w_hash.wh_array[hash]; 3032 w_hash.wh_array[hash] = w; 3033 w_hash.wh_count++; 3034 } 3035 3036 static struct witness_lock_order_data * 3037 witness_lock_order_get(struct witness *parent, struct witness *child) 3038 { 3039 struct witness_lock_order_data *data = NULL; 3040 struct witness_lock_order_key key; 3041 unsigned int hash; 3042 3043 MPASS(parent != NULL && child != NULL); 3044 key.from = parent->w_index; 3045 key.to = child->w_index; 3046 WITNESS_INDEX_ASSERT(key.from); 3047 WITNESS_INDEX_ASSERT(key.to); 3048 if ((w_rmatrix[parent->w_index][child->w_index] 3049 & WITNESS_LOCK_ORDER_KNOWN) == 0) 3050 goto out; 3051 3052 hash = witness_hash_djb2((const char *)&key, 3053 sizeof(key)) % w_lohash.wloh_size; 3054 data = w_lohash.wloh_array[hash]; 3055 while (data != NULL) { 3056 if (witness_lock_order_key_equal(&data->wlod_key, &key)) 3057 break; 3058 data = data->wlod_next; 3059 } 3060 3061 out: 3062 return (data); 3063 } 3064 3065 /* 3066 * Verify that parent and child have a known relationship, are not the same, 3067 * and child is actually a child of parent. This is done without w_mtx 3068 * to avoid contention in the common case. 3069 */ 3070 static int 3071 witness_lock_order_check(struct witness *parent, struct witness *child) 3072 { 3073 if (parent != child && 3074 w_rmatrix[parent->w_index][child->w_index] 3075 & WITNESS_LOCK_ORDER_KNOWN && 3076 isitmychild(parent, child)) 3077 return (1); 3078 3079 return (0); 3080 } 3081 3082 static int 3083 witness_lock_order_add(struct witness *parent, struct witness *child) 3084 { 3085 struct witness_lock_order_data *data = NULL; 3086 struct witness_lock_order_key key; 3087 unsigned int hash; 3088 3089 MPASS(parent != NULL && child != NULL); 3090 key.from = parent->w_index; 3091 key.to = child->w_index; 3092 WITNESS_INDEX_ASSERT(key.from); 3093 WITNESS_INDEX_ASSERT(key.to); 3094 if (w_rmatrix[parent->w_index][child->w_index] 3095 & WITNESS_LOCK_ORDER_KNOWN) 3096 return (1); 3097 3098 hash = witness_hash_djb2((const char *)&key, 3099 sizeof(key)) % w_lohash.wloh_size; 3100 w_rmatrix[parent->w_index][child->w_index] |= WITNESS_LOCK_ORDER_KNOWN; 3101 data = w_lofree; 3102 if (data == NULL) 3103 return (0); 3104 w_lofree = data->wlod_next; 3105 data->wlod_next = w_lohash.wloh_array[hash]; 3106 data->wlod_key = key; 3107 w_lohash.wloh_array[hash] = data; 3108 w_lohash.wloh_count++; 3109 stack_save(&data->wlod_stack); 3110 return (1); 3111 } 3112 3113 /* Call this whenever the structure of the witness graph changes. */ 3114 static void 3115 witness_increment_graph_generation(void) 3116 { 3117 if (witness_cold == 0) 3118 mtx_assert(&w_mtx, MA_OWNED); 3119 w_generation++; 3120 } 3121 3122 static int 3123 witness_output_drain(void *arg __unused, const char *data, int len) 3124 { 3125 witness_output("%.*s", len, data); 3126 return (len); 3127 } 3128 3129 static void 3130 witness_debugger(int cond, const char *msg) 3131 { 3132 char buf[32]; 3133 struct sbuf sb; 3134 struct stack st; 3135 3136 if (!cond) 3137 return; 3138 3139 if (witness_trace) { 3140 sbuf_new(&sb, buf, sizeof(buf), SBUF_FIXEDLEN); 3141 sbuf_set_drain(&sb, witness_output_drain, NULL); 3142 3143 stack_save(&st); 3144 witness_output("stack backtrace:\n"); 3145 stack_sbuf_print_ddb(&sb, &st); 3146 3147 sbuf_finish(&sb); 3148 } 3149 3150 witness_enter_debugger(msg); 3151 } 3152 3153 static void 3154 witness_enter_debugger(const char *msg) 3155 { 3156 #ifdef KDB 3157 if (witness_kdb) 3158 kdb_enter(KDB_WHY_WITNESS, msg); 3159 #endif 3160 } 3161