1 /*- 2 * Copyright (c) 2008 Isilon Systems, Inc. 3 * Copyright (c) 2008 Ilya Maykov <ivmaykov@gmail.com> 4 * Copyright (c) 1998 Berkeley Software Design, Inc. 5 * All rights reserved. 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 1. Redistributions of source code must retain the above copyright 11 * notice, this list of conditions and the following disclaimer. 12 * 2. Redistributions in binary form must reproduce the above copyright 13 * notice, this list of conditions and the following disclaimer in the 14 * documentation and/or other materials provided with the distribution. 15 * 3. Berkeley Software Design Inc's name may not be used to endorse or 16 * promote products derived from this software without specific prior 17 * written permission. 18 * 19 * THIS SOFTWARE IS PROVIDED BY BERKELEY SOFTWARE DESIGN INC ``AS IS'' AND 20 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 21 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 22 * ARE DISCLAIMED. IN NO EVENT SHALL BERKELEY SOFTWARE DESIGN INC BE LIABLE 23 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 24 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 25 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 26 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 27 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 28 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 29 * SUCH DAMAGE. 30 * 31 * from BSDI $Id: mutex_witness.c,v 1.1.2.20 2000/04/27 03:10:27 cp Exp $ 32 * and BSDI $Id: synch_machdep.c,v 2.3.2.39 2000/04/27 03:10:25 cp Exp $ 33 */ 34 35 /* 36 * Implementation of the `witness' lock verifier. Originally implemented for 37 * mutexes in BSD/OS. Extended to handle generic lock objects and lock 38 * classes in FreeBSD. 39 */ 40 41 /* 42 * Main Entry: witness 43 * Pronunciation: 'wit-n&s 44 * Function: noun 45 * Etymology: Middle English witnesse, from Old English witnes knowledge, 46 * testimony, witness, from 2wit 47 * Date: before 12th century 48 * 1 : attestation of a fact or event : TESTIMONY 49 * 2 : one that gives evidence; specifically : one who testifies in 50 * a cause or before a judicial tribunal 51 * 3 : one asked to be present at a transaction so as to be able to 52 * testify to its having taken place 53 * 4 : one who has personal knowledge of something 54 * 5 a : something serving as evidence or proof : SIGN 55 * b : public affirmation by word or example of usually 56 * religious faith or conviction <the heroic witness to divine 57 * life -- Pilot> 58 * 6 capitalized : a member of the Jehovah's Witnesses 59 */ 60 61 /* 62 * Special rules concerning Giant and lock orders: 63 * 64 * 1) Giant must be acquired before any other mutexes. Stated another way, 65 * no other mutex may be held when Giant is acquired. 66 * 67 * 2) Giant must be released when blocking on a sleepable lock. 68 * 69 * This rule is less obvious, but is a result of Giant providing the same 70 * semantics as spl(). Basically, when a thread sleeps, it must release 71 * Giant. When a thread blocks on a sleepable lock, it sleeps. Hence rule 72 * 2). 73 * 74 * 3) Giant may be acquired before or after sleepable locks. 75 * 76 * This rule is also not quite as obvious. Giant may be acquired after 77 * a sleepable lock because it is a non-sleepable lock and non-sleepable 78 * locks may always be acquired while holding a sleepable lock. The second 79 * case, Giant before a sleepable lock, follows from rule 2) above. Suppose 80 * you have two threads T1 and T2 and a sleepable lock X. Suppose that T1 81 * acquires X and blocks on Giant. Then suppose that T2 acquires Giant and 82 * blocks on X. When T2 blocks on X, T2 will release Giant allowing T1 to 83 * execute. Thus, acquiring Giant both before and after a sleepable lock 84 * will not result in a lock order reversal. 85 */ 86 87 #include <sys/cdefs.h> 88 __FBSDID("$FreeBSD$"); 89 90 #include "opt_ddb.h" 91 #include "opt_hwpmc_hooks.h" 92 #include "opt_stack.h" 93 #include "opt_witness.h" 94 95 #include <sys/param.h> 96 #include <sys/bus.h> 97 #include <sys/kdb.h> 98 #include <sys/kernel.h> 99 #include <sys/ktr.h> 100 #include <sys/lock.h> 101 #include <sys/malloc.h> 102 #include <sys/mutex.h> 103 #include <sys/priv.h> 104 #include <sys/proc.h> 105 #include <sys/sbuf.h> 106 #include <sys/sched.h> 107 #include <sys/stack.h> 108 #include <sys/sysctl.h> 109 #include <sys/systm.h> 110 111 #ifdef DDB 112 #include <ddb/ddb.h> 113 #endif 114 115 #include <machine/stdarg.h> 116 117 #if !defined(DDB) && !defined(STACK) 118 #error "DDB or STACK options are required for WITNESS" 119 #endif 120 121 /* Note that these traces do not work with KTR_ALQ. */ 122 #if 0 123 #define KTR_WITNESS KTR_SUBSYS 124 #else 125 #define KTR_WITNESS 0 126 #endif 127 128 #define LI_RECURSEMASK 0x0000ffff /* Recursion depth of lock instance. */ 129 #define LI_EXCLUSIVE 0x00010000 /* Exclusive lock instance. */ 130 #define LI_NORELEASE 0x00020000 /* Lock not allowed to be released. */ 131 132 /* Define this to check for blessed mutexes */ 133 #undef BLESSING 134 135 #define WITNESS_COUNT 1024 136 #define WITNESS_CHILDCOUNT (WITNESS_COUNT * 4) 137 #define WITNESS_HASH_SIZE 251 /* Prime, gives load factor < 2 */ 138 #define WITNESS_PENDLIST 512 139 140 /* Allocate 256 KB of stack data space */ 141 #define WITNESS_LO_DATA_COUNT 2048 142 143 /* Prime, gives load factor of ~2 at full load */ 144 #define WITNESS_LO_HASH_SIZE 1021 145 146 /* 147 * XXX: This is somewhat bogus, as we assume here that at most 2048 threads 148 * will hold LOCK_NCHILDREN locks. We handle failure ok, and we should 149 * probably be safe for the most part, but it's still a SWAG. 150 */ 151 #define LOCK_NCHILDREN 5 152 #define LOCK_CHILDCOUNT 2048 153 154 #define MAX_W_NAME 64 155 156 #define BADSTACK_SBUF_SIZE (256 * WITNESS_COUNT) 157 #define CYCLEGRAPH_SBUF_SIZE 8192 158 #define FULLGRAPH_SBUF_SIZE 32768 159 160 /* 161 * These flags go in the witness relationship matrix and describe the 162 * relationship between any two struct witness objects. 163 */ 164 #define WITNESS_UNRELATED 0x00 /* No lock order relation. */ 165 #define WITNESS_PARENT 0x01 /* Parent, aka direct ancestor. */ 166 #define WITNESS_ANCESTOR 0x02 /* Direct or indirect ancestor. */ 167 #define WITNESS_CHILD 0x04 /* Child, aka direct descendant. */ 168 #define WITNESS_DESCENDANT 0x08 /* Direct or indirect descendant. */ 169 #define WITNESS_ANCESTOR_MASK (WITNESS_PARENT | WITNESS_ANCESTOR) 170 #define WITNESS_DESCENDANT_MASK (WITNESS_CHILD | WITNESS_DESCENDANT) 171 #define WITNESS_RELATED_MASK \ 172 (WITNESS_ANCESTOR_MASK | WITNESS_DESCENDANT_MASK) 173 #define WITNESS_REVERSAL 0x10 /* A lock order reversal has been 174 * observed. */ 175 #define WITNESS_RESERVED1 0x20 /* Unused flag, reserved. */ 176 #define WITNESS_RESERVED2 0x40 /* Unused flag, reserved. */ 177 #define WITNESS_LOCK_ORDER_KNOWN 0x80 /* This lock order is known. */ 178 179 /* Descendant to ancestor flags */ 180 #define WITNESS_DTOA(x) (((x) & WITNESS_RELATED_MASK) >> 2) 181 182 /* Ancestor to descendant flags */ 183 #define WITNESS_ATOD(x) (((x) & WITNESS_RELATED_MASK) << 2) 184 185 #define WITNESS_INDEX_ASSERT(i) \ 186 MPASS((i) > 0 && (i) <= w_max_used_index && (i) < WITNESS_COUNT) 187 188 MALLOC_DEFINE(M_WITNESS, "Witness", "Witness"); 189 190 /* 191 * Lock instances. A lock instance is the data associated with a lock while 192 * it is held by witness. For example, a lock instance will hold the 193 * recursion count of a lock. Lock instances are held in lists. Spin locks 194 * are held in a per-cpu list while sleep locks are held in per-thread list. 195 */ 196 struct lock_instance { 197 struct lock_object *li_lock; 198 const char *li_file; 199 int li_line; 200 u_int li_flags; 201 }; 202 203 /* 204 * A simple list type used to build the list of locks held by a thread 205 * or CPU. We can't simply embed the list in struct lock_object since a 206 * lock may be held by more than one thread if it is a shared lock. Locks 207 * are added to the head of the list, so we fill up each list entry from 208 * "the back" logically. To ease some of the arithmetic, we actually fill 209 * in each list entry the normal way (children[0] then children[1], etc.) but 210 * when we traverse the list we read children[count-1] as the first entry 211 * down to children[0] as the final entry. 212 */ 213 struct lock_list_entry { 214 struct lock_list_entry *ll_next; 215 struct lock_instance ll_children[LOCK_NCHILDREN]; 216 u_int ll_count; 217 }; 218 219 /* 220 * The main witness structure. One of these per named lock type in the system 221 * (for example, "vnode interlock"). 222 */ 223 struct witness { 224 char w_name[MAX_W_NAME]; 225 uint32_t w_index; /* Index in the relationship matrix */ 226 struct lock_class *w_class; 227 STAILQ_ENTRY(witness) w_list; /* List of all witnesses. */ 228 STAILQ_ENTRY(witness) w_typelist; /* Witnesses of a type. */ 229 struct witness *w_hash_next; /* Linked list in hash buckets. */ 230 const char *w_file; /* File where last acquired */ 231 uint32_t w_line; /* Line where last acquired */ 232 uint32_t w_refcount; 233 uint16_t w_num_ancestors; /* direct/indirect 234 * ancestor count */ 235 uint16_t w_num_descendants; /* direct/indirect 236 * descendant count */ 237 int16_t w_ddb_level; 238 unsigned w_displayed:1; 239 unsigned w_reversed:1; 240 }; 241 242 STAILQ_HEAD(witness_list, witness); 243 244 /* 245 * The witness hash table. Keys are witness names (const char *), elements are 246 * witness objects (struct witness *). 247 */ 248 struct witness_hash { 249 struct witness *wh_array[WITNESS_HASH_SIZE]; 250 uint32_t wh_size; 251 uint32_t wh_count; 252 }; 253 254 /* 255 * Key type for the lock order data hash table. 256 */ 257 struct witness_lock_order_key { 258 uint16_t from; 259 uint16_t to; 260 }; 261 262 struct witness_lock_order_data { 263 struct stack wlod_stack; 264 struct witness_lock_order_key wlod_key; 265 struct witness_lock_order_data *wlod_next; 266 }; 267 268 /* 269 * The witness lock order data hash table. Keys are witness index tuples 270 * (struct witness_lock_order_key), elements are lock order data objects 271 * (struct witness_lock_order_data). 272 */ 273 struct witness_lock_order_hash { 274 struct witness_lock_order_data *wloh_array[WITNESS_LO_HASH_SIZE]; 275 u_int wloh_size; 276 u_int wloh_count; 277 }; 278 279 #ifdef BLESSING 280 struct witness_blessed { 281 const char *b_lock1; 282 const char *b_lock2; 283 }; 284 #endif 285 286 struct witness_pendhelp { 287 const char *wh_type; 288 struct lock_object *wh_lock; 289 }; 290 291 struct witness_order_list_entry { 292 const char *w_name; 293 struct lock_class *w_class; 294 }; 295 296 /* 297 * Returns 0 if one of the locks is a spin lock and the other is not. 298 * Returns 1 otherwise. 299 */ 300 static __inline int 301 witness_lock_type_equal(struct witness *w1, struct witness *w2) 302 { 303 304 return ((w1->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) == 305 (w2->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK))); 306 } 307 308 static __inline int 309 witness_lock_order_key_empty(const struct witness_lock_order_key *key) 310 { 311 312 return (key->from == 0 && key->to == 0); 313 } 314 315 static __inline int 316 witness_lock_order_key_equal(const struct witness_lock_order_key *a, 317 const struct witness_lock_order_key *b) 318 { 319 320 return (a->from == b->from && a->to == b->to); 321 } 322 323 static int _isitmyx(struct witness *w1, struct witness *w2, int rmask, 324 const char *fname); 325 #ifdef KDB 326 static void _witness_debugger(int cond, const char *msg); 327 #endif 328 static void adopt(struct witness *parent, struct witness *child); 329 #ifdef BLESSING 330 static int blessed(struct witness *, struct witness *); 331 #endif 332 static void depart(struct witness *w); 333 static struct witness *enroll(const char *description, 334 struct lock_class *lock_class); 335 static struct lock_instance *find_instance(struct lock_list_entry *list, 336 struct lock_object *lock); 337 static int isitmychild(struct witness *parent, struct witness *child); 338 static int isitmydescendant(struct witness *parent, struct witness *child); 339 static void itismychild(struct witness *parent, struct witness *child); 340 static int sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS); 341 static int sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS); 342 static int sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS); 343 static void witness_add_fullgraph(struct sbuf *sb, struct witness *parent); 344 #ifdef DDB 345 static void witness_ddb_compute_levels(void); 346 static void witness_ddb_display(void(*)(const char *fmt, ...)); 347 static void witness_ddb_display_descendants(void(*)(const char *fmt, ...), 348 struct witness *, int indent); 349 static void witness_ddb_display_list(void(*prnt)(const char *fmt, ...), 350 struct witness_list *list); 351 static void witness_ddb_level_descendants(struct witness *parent, int l); 352 static void witness_ddb_list(struct thread *td); 353 #endif 354 static void witness_free(struct witness *m); 355 static struct witness *witness_get(void); 356 static uint32_t witness_hash_djb2(const uint8_t *key, uint32_t size); 357 static struct witness *witness_hash_get(const char *key); 358 static void witness_hash_put(struct witness *w); 359 static void witness_init_hash_tables(void); 360 static void witness_increment_graph_generation(void); 361 static void witness_lock_list_free(struct lock_list_entry *lle); 362 static struct lock_list_entry *witness_lock_list_get(void); 363 static int witness_lock_order_add(struct witness *parent, 364 struct witness *child); 365 static int witness_lock_order_check(struct witness *parent, 366 struct witness *child); 367 static struct witness_lock_order_data *witness_lock_order_get( 368 struct witness *parent, 369 struct witness *child); 370 static void witness_list_lock(struct lock_instance *instance); 371 static void witness_setflag(struct lock_object *lock, int flag, int set); 372 373 #ifdef KDB 374 #define witness_debugger(c) _witness_debugger(c, __func__) 375 #else 376 #define witness_debugger(c) 377 #endif 378 379 SYSCTL_NODE(_debug, OID_AUTO, witness, CTLFLAG_RW, NULL, "Witness Locking"); 380 381 /* 382 * If set to 0, lock order checking is disabled. If set to -1, 383 * witness is completely disabled. Otherwise witness performs full 384 * lock order checking for all locks. At runtime, lock order checking 385 * may be toggled. However, witness cannot be reenabled once it is 386 * completely disabled. 387 */ 388 static int witness_watch = 1; 389 TUNABLE_INT("debug.witness.watch", &witness_watch); 390 SYSCTL_PROC(_debug_witness, OID_AUTO, watch, CTLFLAG_RW | CTLTYPE_INT, NULL, 0, 391 sysctl_debug_witness_watch, "I", "witness is watching lock operations"); 392 393 #ifdef KDB 394 /* 395 * When KDB is enabled and witness_kdb is 1, it will cause the system 396 * to drop into kdebug() when: 397 * - a lock hierarchy violation occurs 398 * - locks are held when going to sleep. 399 */ 400 #ifdef WITNESS_KDB 401 int witness_kdb = 1; 402 #else 403 int witness_kdb = 0; 404 #endif 405 TUNABLE_INT("debug.witness.kdb", &witness_kdb); 406 SYSCTL_INT(_debug_witness, OID_AUTO, kdb, CTLFLAG_RW, &witness_kdb, 0, ""); 407 408 /* 409 * When KDB is enabled and witness_trace is 1, it will cause the system 410 * to print a stack trace: 411 * - a lock hierarchy violation occurs 412 * - locks are held when going to sleep. 413 */ 414 int witness_trace = 1; 415 TUNABLE_INT("debug.witness.trace", &witness_trace); 416 SYSCTL_INT(_debug_witness, OID_AUTO, trace, CTLFLAG_RW, &witness_trace, 0, ""); 417 #endif /* KDB */ 418 419 #ifdef WITNESS_SKIPSPIN 420 int witness_skipspin = 1; 421 #else 422 int witness_skipspin = 0; 423 #endif 424 TUNABLE_INT("debug.witness.skipspin", &witness_skipspin); 425 SYSCTL_INT(_debug_witness, OID_AUTO, skipspin, CTLFLAG_RDTUN, &witness_skipspin, 426 0, ""); 427 428 /* 429 * Call this to print out the relations between locks. 430 */ 431 SYSCTL_PROC(_debug_witness, OID_AUTO, fullgraph, CTLTYPE_STRING | CTLFLAG_RD, 432 NULL, 0, sysctl_debug_witness_fullgraph, "A", "Show locks relation graphs"); 433 434 /* 435 * Call this to print out the witness faulty stacks. 436 */ 437 SYSCTL_PROC(_debug_witness, OID_AUTO, badstacks, CTLTYPE_STRING | CTLFLAG_RD, 438 NULL, 0, sysctl_debug_witness_badstacks, "A", "Show bad witness stacks"); 439 440 static struct mtx w_mtx; 441 442 /* w_list */ 443 static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free); 444 static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all); 445 446 /* w_typelist */ 447 static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin); 448 static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep); 449 450 /* lock list */ 451 static struct lock_list_entry *w_lock_list_free = NULL; 452 static struct witness_pendhelp pending_locks[WITNESS_PENDLIST]; 453 static u_int pending_cnt; 454 455 static int w_free_cnt, w_spin_cnt, w_sleep_cnt; 456 SYSCTL_INT(_debug_witness, OID_AUTO, free_cnt, CTLFLAG_RD, &w_free_cnt, 0, ""); 457 SYSCTL_INT(_debug_witness, OID_AUTO, spin_cnt, CTLFLAG_RD, &w_spin_cnt, 0, ""); 458 SYSCTL_INT(_debug_witness, OID_AUTO, sleep_cnt, CTLFLAG_RD, &w_sleep_cnt, 0, 459 ""); 460 461 static struct witness *w_data; 462 static uint8_t w_rmatrix[WITNESS_COUNT+1][WITNESS_COUNT+1]; 463 static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT]; 464 static struct witness_hash w_hash; /* The witness hash table. */ 465 466 /* The lock order data hash */ 467 static struct witness_lock_order_data w_lodata[WITNESS_LO_DATA_COUNT]; 468 static struct witness_lock_order_data *w_lofree = NULL; 469 static struct witness_lock_order_hash w_lohash; 470 static int w_max_used_index = 0; 471 static unsigned int w_generation = 0; 472 static const char *w_notrunning = "Witness not running\n"; 473 static const char *w_stillcold = "Witness is still cold\n"; 474 475 476 static struct witness_order_list_entry order_lists[] = { 477 /* 478 * sx locks 479 */ 480 { "proctree", &lock_class_sx }, 481 { "allproc", &lock_class_sx }, 482 { "allprison", &lock_class_sx }, 483 { NULL, NULL }, 484 /* 485 * Various mutexes 486 */ 487 { "Giant", &lock_class_mtx_sleep }, 488 { "pipe mutex", &lock_class_mtx_sleep }, 489 { "sigio lock", &lock_class_mtx_sleep }, 490 { "process group", &lock_class_mtx_sleep }, 491 { "process lock", &lock_class_mtx_sleep }, 492 { "session", &lock_class_mtx_sleep }, 493 { "uidinfo hash", &lock_class_rw }, 494 #ifdef HWPMC_HOOKS 495 { "pmc-sleep", &lock_class_mtx_sleep }, 496 #endif 497 { NULL, NULL }, 498 /* 499 * Sockets 500 */ 501 { "accept", &lock_class_mtx_sleep }, 502 { "so_snd", &lock_class_mtx_sleep }, 503 { "so_rcv", &lock_class_mtx_sleep }, 504 { "sellck", &lock_class_mtx_sleep }, 505 { NULL, NULL }, 506 /* 507 * Routing 508 */ 509 { "so_rcv", &lock_class_mtx_sleep }, 510 { "radix node head", &lock_class_rw }, 511 { "rtentry", &lock_class_mtx_sleep }, 512 { "ifaddr", &lock_class_mtx_sleep }, 513 { NULL, NULL }, 514 /* 515 * Multicast - protocol locks before interface locks, after UDP locks. 516 */ 517 { "udpinp", &lock_class_rw }, 518 { "in_multi_mtx", &lock_class_mtx_sleep }, 519 { "igmp_mtx", &lock_class_mtx_sleep }, 520 { "if_addr_mtx", &lock_class_mtx_sleep }, 521 { NULL, NULL }, 522 /* 523 * UNIX Domain Sockets 524 */ 525 { "unp_global_rwlock", &lock_class_rw }, 526 { "unp_list_lock", &lock_class_mtx_sleep }, 527 { "unp", &lock_class_mtx_sleep }, 528 { "so_snd", &lock_class_mtx_sleep }, 529 { NULL, NULL }, 530 /* 531 * UDP/IP 532 */ 533 { "udp", &lock_class_rw }, 534 { "udpinp", &lock_class_rw }, 535 { "so_snd", &lock_class_mtx_sleep }, 536 { NULL, NULL }, 537 /* 538 * TCP/IP 539 */ 540 { "tcp", &lock_class_rw }, 541 { "tcpinp", &lock_class_rw }, 542 { "so_snd", &lock_class_mtx_sleep }, 543 { NULL, NULL }, 544 /* 545 * SLIP 546 */ 547 { "slip_mtx", &lock_class_mtx_sleep }, 548 { "slip sc_mtx", &lock_class_mtx_sleep }, 549 { NULL, NULL }, 550 /* 551 * netatalk 552 */ 553 { "ddp_list_mtx", &lock_class_mtx_sleep }, 554 { "ddp_mtx", &lock_class_mtx_sleep }, 555 { NULL, NULL }, 556 /* 557 * BPF 558 */ 559 { "bpf global lock", &lock_class_mtx_sleep }, 560 { "bpf interface lock", &lock_class_mtx_sleep }, 561 { "bpf cdev lock", &lock_class_mtx_sleep }, 562 { NULL, NULL }, 563 /* 564 * NFS server 565 */ 566 { "nfsd_mtx", &lock_class_mtx_sleep }, 567 { "so_snd", &lock_class_mtx_sleep }, 568 { NULL, NULL }, 569 570 /* 571 * IEEE 802.11 572 */ 573 { "802.11 com lock", &lock_class_mtx_sleep}, 574 { NULL, NULL }, 575 /* 576 * Network drivers 577 */ 578 { "network driver", &lock_class_mtx_sleep}, 579 { NULL, NULL }, 580 581 /* 582 * Netgraph 583 */ 584 { "ng_node", &lock_class_mtx_sleep }, 585 { "ng_worklist", &lock_class_mtx_sleep }, 586 { NULL, NULL }, 587 /* 588 * CDEV 589 */ 590 { "system map", &lock_class_mtx_sleep }, 591 { "vm page queue mutex", &lock_class_mtx_sleep }, 592 { "vnode interlock", &lock_class_mtx_sleep }, 593 { "cdev", &lock_class_mtx_sleep }, 594 { NULL, NULL }, 595 /* 596 * kqueue/VFS interaction 597 */ 598 { "kqueue", &lock_class_mtx_sleep }, 599 { "struct mount mtx", &lock_class_mtx_sleep }, 600 { "vnode interlock", &lock_class_mtx_sleep }, 601 { NULL, NULL }, 602 /* 603 * spin locks 604 */ 605 #ifdef SMP 606 { "ap boot", &lock_class_mtx_spin }, 607 #endif 608 { "rm.mutex_mtx", &lock_class_mtx_spin }, 609 { "sio", &lock_class_mtx_spin }, 610 { "scrlock", &lock_class_mtx_spin }, 611 #ifdef __i386__ 612 { "cy", &lock_class_mtx_spin }, 613 #endif 614 #ifdef __sparc64__ 615 { "pcib_mtx", &lock_class_mtx_spin }, 616 { "rtc_mtx", &lock_class_mtx_spin }, 617 #endif 618 { "scc_hwmtx", &lock_class_mtx_spin }, 619 { "uart_hwmtx", &lock_class_mtx_spin }, 620 { "fast_taskqueue", &lock_class_mtx_spin }, 621 { "intr table", &lock_class_mtx_spin }, 622 #ifdef HWPMC_HOOKS 623 { "pmc-per-proc", &lock_class_mtx_spin }, 624 #endif 625 { "process slock", &lock_class_mtx_spin }, 626 { "sleepq chain", &lock_class_mtx_spin }, 627 { "umtx lock", &lock_class_mtx_spin }, 628 { "rm_spinlock", &lock_class_mtx_spin }, 629 { "turnstile chain", &lock_class_mtx_spin }, 630 { "turnstile lock", &lock_class_mtx_spin }, 631 { "sched lock", &lock_class_mtx_spin }, 632 { "td_contested", &lock_class_mtx_spin }, 633 { "callout", &lock_class_mtx_spin }, 634 { "entropy harvest mutex", &lock_class_mtx_spin }, 635 { "syscons video lock", &lock_class_mtx_spin }, 636 { "time lock", &lock_class_mtx_spin }, 637 #ifdef SMP 638 { "smp rendezvous", &lock_class_mtx_spin }, 639 #endif 640 #ifdef __powerpc__ 641 { "tlb0", &lock_class_mtx_spin }, 642 #endif 643 /* 644 * leaf locks 645 */ 646 { "intrcnt", &lock_class_mtx_spin }, 647 { "icu", &lock_class_mtx_spin }, 648 #if defined(SMP) && defined(__sparc64__) 649 { "ipi", &lock_class_mtx_spin }, 650 #endif 651 #ifdef __i386__ 652 { "allpmaps", &lock_class_mtx_spin }, 653 { "descriptor tables", &lock_class_mtx_spin }, 654 #endif 655 { "clk", &lock_class_mtx_spin }, 656 { "cpuset", &lock_class_mtx_spin }, 657 { "mprof lock", &lock_class_mtx_spin }, 658 { "zombie lock", &lock_class_mtx_spin }, 659 { "ALD Queue", &lock_class_mtx_spin }, 660 #ifdef __ia64__ 661 { "MCA spin lock", &lock_class_mtx_spin }, 662 #endif 663 #if defined(__i386__) || defined(__amd64__) 664 { "pcicfg", &lock_class_mtx_spin }, 665 { "NDIS thread lock", &lock_class_mtx_spin }, 666 #endif 667 { "tw_osl_io_lock", &lock_class_mtx_spin }, 668 { "tw_osl_q_lock", &lock_class_mtx_spin }, 669 { "tw_cl_io_lock", &lock_class_mtx_spin }, 670 { "tw_cl_intr_lock", &lock_class_mtx_spin }, 671 { "tw_cl_gen_lock", &lock_class_mtx_spin }, 672 #ifdef HWPMC_HOOKS 673 { "pmc-leaf", &lock_class_mtx_spin }, 674 #endif 675 { "blocked lock", &lock_class_mtx_spin }, 676 { NULL, NULL }, 677 { NULL, NULL } 678 }; 679 680 #ifdef BLESSING 681 /* 682 * Pairs of locks which have been blessed 683 * Don't complain about order problems with blessed locks 684 */ 685 static struct witness_blessed blessed_list[] = { 686 }; 687 static int blessed_count = 688 sizeof(blessed_list) / sizeof(struct witness_blessed); 689 #endif 690 691 /* 692 * This global is set to 0 once it becomes safe to use the witness code. 693 */ 694 static int witness_cold = 1; 695 696 /* 697 * This global is set to 1 once the static lock orders have been enrolled 698 * so that a warning can be issued for any spin locks enrolled later. 699 */ 700 static int witness_spin_warn = 0; 701 702 /* 703 * The WITNESS-enabled diagnostic code. Note that the witness code does 704 * assume that the early boot is single-threaded at least until after this 705 * routine is completed. 706 */ 707 static void 708 witness_initialize(void *dummy __unused) 709 { 710 struct lock_object *lock; 711 struct witness_order_list_entry *order; 712 struct witness *w, *w1; 713 int i; 714 715 w_data = malloc(sizeof (struct witness) * WITNESS_COUNT, M_WITNESS, 716 M_NOWAIT | M_ZERO); 717 718 /* 719 * We have to release Giant before initializing its witness 720 * structure so that WITNESS doesn't get confused. 721 */ 722 mtx_unlock(&Giant); 723 mtx_assert(&Giant, MA_NOTOWNED); 724 725 CTR1(KTR_WITNESS, "%s: initializing witness", __func__); 726 mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET | 727 MTX_NOWITNESS | MTX_NOPROFILE); 728 for (i = WITNESS_COUNT - 1; i >= 0; i--) { 729 w = &w_data[i]; 730 memset(w, 0, sizeof(*w)); 731 w_data[i].w_index = i; /* Witness index never changes. */ 732 witness_free(w); 733 } 734 KASSERT(STAILQ_FIRST(&w_free)->w_index == 0, 735 ("%s: Invalid list of free witness objects", __func__)); 736 737 /* Witness with index 0 is not used to aid in debugging. */ 738 STAILQ_REMOVE_HEAD(&w_free, w_list); 739 w_free_cnt--; 740 741 memset(w_rmatrix, 0, 742 (sizeof(**w_rmatrix) * (WITNESS_COUNT+1) * (WITNESS_COUNT+1))); 743 744 for (i = 0; i < LOCK_CHILDCOUNT; i++) 745 witness_lock_list_free(&w_locklistdata[i]); 746 witness_init_hash_tables(); 747 748 /* First add in all the specified order lists. */ 749 for (order = order_lists; order->w_name != NULL; order++) { 750 w = enroll(order->w_name, order->w_class); 751 if (w == NULL) 752 continue; 753 w->w_file = "order list"; 754 for (order++; order->w_name != NULL; order++) { 755 w1 = enroll(order->w_name, order->w_class); 756 if (w1 == NULL) 757 continue; 758 w1->w_file = "order list"; 759 itismychild(w, w1); 760 w = w1; 761 } 762 } 763 witness_spin_warn = 1; 764 765 /* Iterate through all locks and add them to witness. */ 766 for (i = 0; pending_locks[i].wh_lock != NULL; i++) { 767 lock = pending_locks[i].wh_lock; 768 KASSERT(lock->lo_flags & LO_WITNESS, 769 ("%s: lock %s is on pending list but not LO_WITNESS", 770 __func__, lock->lo_name)); 771 lock->lo_witness = enroll(pending_locks[i].wh_type, 772 LOCK_CLASS(lock)); 773 } 774 775 /* Mark the witness code as being ready for use. */ 776 witness_cold = 0; 777 778 mtx_lock(&Giant); 779 } 780 SYSINIT(witness_init, SI_SUB_WITNESS, SI_ORDER_FIRST, witness_initialize, 781 NULL); 782 783 void 784 witness_init(struct lock_object *lock, const char *type) 785 { 786 struct lock_class *class; 787 788 /* Various sanity checks. */ 789 class = LOCK_CLASS(lock); 790 if ((lock->lo_flags & LO_RECURSABLE) != 0 && 791 (class->lc_flags & LC_RECURSABLE) == 0) 792 panic("%s: lock (%s) %s can not be recursable", __func__, 793 class->lc_name, lock->lo_name); 794 if ((lock->lo_flags & LO_SLEEPABLE) != 0 && 795 (class->lc_flags & LC_SLEEPABLE) == 0) 796 panic("%s: lock (%s) %s can not be sleepable", __func__, 797 class->lc_name, lock->lo_name); 798 if ((lock->lo_flags & LO_UPGRADABLE) != 0 && 799 (class->lc_flags & LC_UPGRADABLE) == 0) 800 panic("%s: lock (%s) %s can not be upgradable", __func__, 801 class->lc_name, lock->lo_name); 802 803 /* 804 * If we shouldn't watch this lock, then just clear lo_witness. 805 * Otherwise, if witness_cold is set, then it is too early to 806 * enroll this lock, so defer it to witness_initialize() by adding 807 * it to the pending_locks list. If it is not too early, then enroll 808 * the lock now. 809 */ 810 if (witness_watch < 1 || panicstr != NULL || 811 (lock->lo_flags & LO_WITNESS) == 0) 812 lock->lo_witness = NULL; 813 else if (witness_cold) { 814 pending_locks[pending_cnt].wh_lock = lock; 815 pending_locks[pending_cnt++].wh_type = type; 816 if (pending_cnt > WITNESS_PENDLIST) 817 panic("%s: pending locks list is too small, bump it\n", 818 __func__); 819 } else 820 lock->lo_witness = enroll(type, class); 821 } 822 823 void 824 witness_destroy(struct lock_object *lock) 825 { 826 struct lock_class *class; 827 struct witness *w; 828 829 class = LOCK_CLASS(lock); 830 831 if (witness_cold) 832 panic("lock (%s) %s destroyed while witness_cold", 833 class->lc_name, lock->lo_name); 834 835 /* XXX: need to verify that no one holds the lock */ 836 if ((lock->lo_flags & LO_WITNESS) == 0 || lock->lo_witness == NULL) 837 return; 838 w = lock->lo_witness; 839 840 mtx_lock_spin(&w_mtx); 841 MPASS(w->w_refcount > 0); 842 w->w_refcount--; 843 844 if (w->w_refcount == 0) 845 depart(w); 846 mtx_unlock_spin(&w_mtx); 847 } 848 849 #ifdef DDB 850 static void 851 witness_ddb_compute_levels(void) 852 { 853 struct witness *w; 854 855 /* 856 * First clear all levels. 857 */ 858 STAILQ_FOREACH(w, &w_all, w_list) 859 w->w_ddb_level = -1; 860 861 /* 862 * Look for locks with no parents and level all their descendants. 863 */ 864 STAILQ_FOREACH(w, &w_all, w_list) { 865 866 /* If the witness has ancestors (is not a root), skip it. */ 867 if (w->w_num_ancestors > 0) 868 continue; 869 witness_ddb_level_descendants(w, 0); 870 } 871 } 872 873 static void 874 witness_ddb_level_descendants(struct witness *w, int l) 875 { 876 int i; 877 878 if (w->w_ddb_level >= l) 879 return; 880 881 w->w_ddb_level = l; 882 l++; 883 884 for (i = 1; i <= w_max_used_index; i++) { 885 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT) 886 witness_ddb_level_descendants(&w_data[i], l); 887 } 888 } 889 890 static void 891 witness_ddb_display_descendants(void(*prnt)(const char *fmt, ...), 892 struct witness *w, int indent) 893 { 894 int i; 895 896 for (i = 0; i < indent; i++) 897 prnt(" "); 898 prnt("%s (type: %s, depth: %d, active refs: %d)", 899 w->w_name, w->w_class->lc_name, 900 w->w_ddb_level, w->w_refcount); 901 if (w->w_displayed) { 902 prnt(" -- (already displayed)\n"); 903 return; 904 } 905 w->w_displayed = 1; 906 if (w->w_file != NULL && w->w_line != 0) 907 prnt(" -- last acquired @ %s:%d\n", w->w_file, 908 w->w_line); 909 else 910 prnt(" -- never acquired\n"); 911 indent++; 912 WITNESS_INDEX_ASSERT(w->w_index); 913 for (i = 1; i <= w_max_used_index; i++) { 914 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT) 915 witness_ddb_display_descendants(prnt, &w_data[i], 916 indent); 917 } 918 } 919 920 static void 921 witness_ddb_display_list(void(*prnt)(const char *fmt, ...), 922 struct witness_list *list) 923 { 924 struct witness *w; 925 926 STAILQ_FOREACH(w, list, w_typelist) { 927 if (w->w_file == NULL || w->w_ddb_level > 0) 928 continue; 929 930 /* This lock has no anscestors - display its descendants. */ 931 witness_ddb_display_descendants(prnt, w, 0); 932 } 933 } 934 935 static void 936 witness_ddb_display(void(*prnt)(const char *fmt, ...)) 937 { 938 struct witness *w; 939 940 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__)); 941 witness_ddb_compute_levels(); 942 943 /* Clear all the displayed flags. */ 944 STAILQ_FOREACH(w, &w_all, w_list) 945 w->w_displayed = 0; 946 947 /* 948 * First, handle sleep locks which have been acquired at least 949 * once. 950 */ 951 prnt("Sleep locks:\n"); 952 witness_ddb_display_list(prnt, &w_sleep); 953 954 /* 955 * Now do spin locks which have been acquired at least once. 956 */ 957 prnt("\nSpin locks:\n"); 958 witness_ddb_display_list(prnt, &w_spin); 959 960 /* 961 * Finally, any locks which have not been acquired yet. 962 */ 963 prnt("\nLocks which were never acquired:\n"); 964 STAILQ_FOREACH(w, &w_all, w_list) { 965 if (w->w_file != NULL || w->w_refcount == 0) 966 continue; 967 prnt("%s (type: %s, depth: %d)\n", w->w_name, 968 w->w_class->lc_name, w->w_ddb_level); 969 } 970 } 971 #endif /* DDB */ 972 973 /* Trim useless garbage from filenames. */ 974 static const char * 975 fixup_filename(const char *file) 976 { 977 978 if (file == NULL) 979 return (NULL); 980 while (strncmp(file, "../", 3) == 0) 981 file += 3; 982 return (file); 983 } 984 985 int 986 witness_defineorder(struct lock_object *lock1, struct lock_object *lock2) 987 { 988 989 if (witness_watch == -1 || panicstr != NULL) 990 return (0); 991 992 /* Require locks that witness knows about. */ 993 if (lock1 == NULL || lock1->lo_witness == NULL || lock2 == NULL || 994 lock2->lo_witness == NULL) 995 return (EINVAL); 996 997 mtx_assert(&w_mtx, MA_NOTOWNED); 998 mtx_lock_spin(&w_mtx); 999 1000 /* 1001 * If we already have either an explicit or implied lock order that 1002 * is the other way around, then return an error. 1003 */ 1004 if (witness_watch && 1005 isitmydescendant(lock2->lo_witness, lock1->lo_witness)) { 1006 mtx_unlock_spin(&w_mtx); 1007 return (EDOOFUS); 1008 } 1009 1010 /* Try to add the new order. */ 1011 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__, 1012 lock2->lo_witness->w_name, lock1->lo_witness->w_name); 1013 itismychild(lock1->lo_witness, lock2->lo_witness); 1014 mtx_unlock_spin(&w_mtx); 1015 return (0); 1016 } 1017 1018 void 1019 witness_checkorder(struct lock_object *lock, int flags, const char *file, 1020 int line, struct lock_object *interlock) 1021 { 1022 struct lock_list_entry *lock_list, *lle; 1023 struct lock_instance *lock1, *lock2, *plock; 1024 struct lock_class *class; 1025 struct witness *w, *w1; 1026 struct thread *td; 1027 int i, j; 1028 1029 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL || 1030 panicstr != NULL) 1031 return; 1032 1033 w = lock->lo_witness; 1034 class = LOCK_CLASS(lock); 1035 td = curthread; 1036 file = fixup_filename(file); 1037 1038 if (class->lc_flags & LC_SLEEPLOCK) { 1039 1040 /* 1041 * Since spin locks include a critical section, this check 1042 * implicitly enforces a lock order of all sleep locks before 1043 * all spin locks. 1044 */ 1045 if (td->td_critnest != 0 && !kdb_active) 1046 panic("blockable sleep lock (%s) %s @ %s:%d", 1047 class->lc_name, lock->lo_name, file, line); 1048 1049 /* 1050 * If this is the first lock acquired then just return as 1051 * no order checking is needed. 1052 */ 1053 lock_list = td->td_sleeplocks; 1054 if (lock_list == NULL || lock_list->ll_count == 0) 1055 return; 1056 } else { 1057 1058 /* 1059 * If this is the first lock, just return as no order 1060 * checking is needed. Avoid problems with thread 1061 * migration pinning the thread while checking if 1062 * spinlocks are held. If at least one spinlock is held 1063 * the thread is in a safe path and it is allowed to 1064 * unpin it. 1065 */ 1066 sched_pin(); 1067 lock_list = PCPU_GET(spinlocks); 1068 if (lock_list == NULL || lock_list->ll_count == 0) { 1069 sched_unpin(); 1070 return; 1071 } 1072 sched_unpin(); 1073 } 1074 1075 /* 1076 * Check to see if we are recursing on a lock we already own. If 1077 * so, make sure that we don't mismatch exclusive and shared lock 1078 * acquires. 1079 */ 1080 lock1 = find_instance(lock_list, lock); 1081 if (lock1 != NULL) { 1082 if ((lock1->li_flags & LI_EXCLUSIVE) != 0 && 1083 (flags & LOP_EXCLUSIVE) == 0) { 1084 printf("shared lock of (%s) %s @ %s:%d\n", 1085 class->lc_name, lock->lo_name, file, line); 1086 printf("while exclusively locked from %s:%d\n", 1087 lock1->li_file, lock1->li_line); 1088 panic("share->excl"); 1089 } 1090 if ((lock1->li_flags & LI_EXCLUSIVE) == 0 && 1091 (flags & LOP_EXCLUSIVE) != 0) { 1092 printf("exclusive lock of (%s) %s @ %s:%d\n", 1093 class->lc_name, lock->lo_name, file, line); 1094 printf("while share locked from %s:%d\n", 1095 lock1->li_file, lock1->li_line); 1096 panic("excl->share"); 1097 } 1098 return; 1099 } 1100 1101 /* 1102 * Find the previously acquired lock, but ignore interlocks. 1103 */ 1104 plock = &lock_list->ll_children[lock_list->ll_count - 1]; 1105 if (interlock != NULL && plock->li_lock == interlock) { 1106 if (lock_list->ll_count > 1) 1107 plock = 1108 &lock_list->ll_children[lock_list->ll_count - 2]; 1109 else { 1110 lle = lock_list->ll_next; 1111 1112 /* 1113 * The interlock is the only lock we hold, so 1114 * simply return. 1115 */ 1116 if (lle == NULL) 1117 return; 1118 plock = &lle->ll_children[lle->ll_count - 1]; 1119 } 1120 } 1121 1122 /* 1123 * Try to perform most checks without a lock. If this succeeds we 1124 * can skip acquiring the lock and return success. 1125 */ 1126 w1 = plock->li_lock->lo_witness; 1127 if (witness_lock_order_check(w1, w)) 1128 return; 1129 1130 /* 1131 * Check for duplicate locks of the same type. Note that we only 1132 * have to check for this on the last lock we just acquired. Any 1133 * other cases will be caught as lock order violations. 1134 */ 1135 mtx_lock_spin(&w_mtx); 1136 witness_lock_order_add(w1, w); 1137 if (w1 == w) { 1138 i = w->w_index; 1139 if (!(lock->lo_flags & LO_DUPOK) && !(flags & LOP_DUPOK) && 1140 !(w_rmatrix[i][i] & WITNESS_REVERSAL)) { 1141 w_rmatrix[i][i] |= WITNESS_REVERSAL; 1142 w->w_reversed = 1; 1143 mtx_unlock_spin(&w_mtx); 1144 printf( 1145 "acquiring duplicate lock of same type: \"%s\"\n", 1146 w->w_name); 1147 printf(" 1st %s @ %s:%d\n", plock->li_lock->lo_name, 1148 plock->li_file, plock->li_line); 1149 printf(" 2nd %s @ %s:%d\n", lock->lo_name, file, line); 1150 witness_debugger(1); 1151 } else 1152 mtx_unlock_spin(&w_mtx); 1153 return; 1154 } 1155 mtx_assert(&w_mtx, MA_OWNED); 1156 1157 /* 1158 * If we know that the the lock we are acquiring comes after 1159 * the lock we most recently acquired in the lock order tree, 1160 * then there is no need for any further checks. 1161 */ 1162 if (isitmychild(w1, w)) 1163 goto out; 1164 1165 for (j = 0, lle = lock_list; lle != NULL; lle = lle->ll_next) { 1166 for (i = lle->ll_count - 1; i >= 0; i--, j++) { 1167 1168 MPASS(j < WITNESS_COUNT); 1169 lock1 = &lle->ll_children[i]; 1170 1171 /* 1172 * Ignore the interlock the first time we see it. 1173 */ 1174 if (interlock != NULL && interlock == lock1->li_lock) { 1175 interlock = NULL; 1176 continue; 1177 } 1178 1179 /* 1180 * If this lock doesn't undergo witness checking, 1181 * then skip it. 1182 */ 1183 w1 = lock1->li_lock->lo_witness; 1184 if (w1 == NULL) { 1185 KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0, 1186 ("lock missing witness structure")); 1187 continue; 1188 } 1189 1190 /* 1191 * If we are locking Giant and this is a sleepable 1192 * lock, then skip it. 1193 */ 1194 if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0 && 1195 lock == &Giant.lock_object) 1196 continue; 1197 1198 /* 1199 * If we are locking a sleepable lock and this lock 1200 * is Giant, then skip it. 1201 */ 1202 if ((lock->lo_flags & LO_SLEEPABLE) != 0 && 1203 lock1->li_lock == &Giant.lock_object) 1204 continue; 1205 1206 /* 1207 * If we are locking a sleepable lock and this lock 1208 * isn't sleepable, we want to treat it as a lock 1209 * order violation to enfore a general lock order of 1210 * sleepable locks before non-sleepable locks. 1211 */ 1212 if (((lock->lo_flags & LO_SLEEPABLE) != 0 && 1213 (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0)) 1214 goto reversal; 1215 1216 /* 1217 * If we are locking Giant and this is a non-sleepable 1218 * lock, then treat it as a reversal. 1219 */ 1220 if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0 && 1221 lock == &Giant.lock_object) 1222 goto reversal; 1223 1224 /* 1225 * Check the lock order hierarchy for a reveresal. 1226 */ 1227 if (!isitmydescendant(w, w1)) 1228 continue; 1229 reversal: 1230 1231 /* 1232 * We have a lock order violation, check to see if it 1233 * is allowed or has already been yelled about. 1234 */ 1235 #ifdef BLESSING 1236 1237 /* 1238 * If the lock order is blessed, just bail. We don't 1239 * look for other lock order violations though, which 1240 * may be a bug. 1241 */ 1242 if (blessed(w, w1)) 1243 goto out; 1244 #endif 1245 1246 /* Bail if this violation is known */ 1247 if (w_rmatrix[w1->w_index][w->w_index] & WITNESS_REVERSAL) 1248 goto out; 1249 1250 /* Record this as a violation */ 1251 w_rmatrix[w1->w_index][w->w_index] |= WITNESS_REVERSAL; 1252 w_rmatrix[w->w_index][w1->w_index] |= WITNESS_REVERSAL; 1253 w->w_reversed = w1->w_reversed = 1; 1254 witness_increment_graph_generation(); 1255 mtx_unlock_spin(&w_mtx); 1256 1257 /* 1258 * Ok, yell about it. 1259 */ 1260 if (((lock->lo_flags & LO_SLEEPABLE) != 0 && 1261 (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0)) 1262 printf( 1263 "lock order reversal: (sleepable after non-sleepable)\n"); 1264 else if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0 1265 && lock == &Giant.lock_object) 1266 printf( 1267 "lock order reversal: (Giant after non-sleepable)\n"); 1268 else 1269 printf("lock order reversal:\n"); 1270 1271 /* 1272 * Try to locate an earlier lock with 1273 * witness w in our list. 1274 */ 1275 do { 1276 lock2 = &lle->ll_children[i]; 1277 MPASS(lock2->li_lock != NULL); 1278 if (lock2->li_lock->lo_witness == w) 1279 break; 1280 if (i == 0 && lle->ll_next != NULL) { 1281 lle = lle->ll_next; 1282 i = lle->ll_count - 1; 1283 MPASS(i >= 0 && i < LOCK_NCHILDREN); 1284 } else 1285 i--; 1286 } while (i >= 0); 1287 if (i < 0) { 1288 printf(" 1st %p %s (%s) @ %s:%d\n", 1289 lock1->li_lock, lock1->li_lock->lo_name, 1290 w1->w_name, lock1->li_file, lock1->li_line); 1291 printf(" 2nd %p %s (%s) @ %s:%d\n", lock, 1292 lock->lo_name, w->w_name, file, line); 1293 } else { 1294 printf(" 1st %p %s (%s) @ %s:%d\n", 1295 lock2->li_lock, lock2->li_lock->lo_name, 1296 lock2->li_lock->lo_witness->w_name, 1297 lock2->li_file, lock2->li_line); 1298 printf(" 2nd %p %s (%s) @ %s:%d\n", 1299 lock1->li_lock, lock1->li_lock->lo_name, 1300 w1->w_name, lock1->li_file, lock1->li_line); 1301 printf(" 3rd %p %s (%s) @ %s:%d\n", lock, 1302 lock->lo_name, w->w_name, file, line); 1303 } 1304 witness_debugger(1); 1305 return; 1306 } 1307 } 1308 1309 /* 1310 * If requested, build a new lock order. However, don't build a new 1311 * relationship between a sleepable lock and Giant if it is in the 1312 * wrong direction. The correct lock order is that sleepable locks 1313 * always come before Giant. 1314 */ 1315 if (flags & LOP_NEWORDER && 1316 !(plock->li_lock == &Giant.lock_object && 1317 (lock->lo_flags & LO_SLEEPABLE) != 0)) { 1318 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__, 1319 w->w_name, plock->li_lock->lo_witness->w_name); 1320 itismychild(plock->li_lock->lo_witness, w); 1321 } 1322 out: 1323 mtx_unlock_spin(&w_mtx); 1324 } 1325 1326 void 1327 witness_lock(struct lock_object *lock, int flags, const char *file, int line) 1328 { 1329 struct lock_list_entry **lock_list, *lle; 1330 struct lock_instance *instance; 1331 struct witness *w; 1332 struct thread *td; 1333 1334 if (witness_cold || witness_watch == -1 || lock->lo_witness == NULL || 1335 panicstr != NULL) 1336 return; 1337 w = lock->lo_witness; 1338 td = curthread; 1339 file = fixup_filename(file); 1340 1341 /* Determine lock list for this lock. */ 1342 if (LOCK_CLASS(lock)->lc_flags & LC_SLEEPLOCK) 1343 lock_list = &td->td_sleeplocks; 1344 else 1345 lock_list = PCPU_PTR(spinlocks); 1346 1347 /* Check to see if we are recursing on a lock we already own. */ 1348 instance = find_instance(*lock_list, lock); 1349 if (instance != NULL) { 1350 instance->li_flags++; 1351 CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__, 1352 td->td_proc->p_pid, lock->lo_name, 1353 instance->li_flags & LI_RECURSEMASK); 1354 instance->li_file = file; 1355 instance->li_line = line; 1356 return; 1357 } 1358 1359 /* Update per-witness last file and line acquire. */ 1360 w->w_file = file; 1361 w->w_line = line; 1362 1363 /* Find the next open lock instance in the list and fill it. */ 1364 lle = *lock_list; 1365 if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) { 1366 lle = witness_lock_list_get(); 1367 if (lle == NULL) 1368 return; 1369 lle->ll_next = *lock_list; 1370 CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__, 1371 td->td_proc->p_pid, lle); 1372 *lock_list = lle; 1373 } 1374 instance = &lle->ll_children[lle->ll_count++]; 1375 instance->li_lock = lock; 1376 instance->li_line = line; 1377 instance->li_file = file; 1378 if ((flags & LOP_EXCLUSIVE) != 0) 1379 instance->li_flags = LI_EXCLUSIVE; 1380 else 1381 instance->li_flags = 0; 1382 CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__, 1383 td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1); 1384 } 1385 1386 void 1387 witness_upgrade(struct lock_object *lock, int flags, const char *file, int line) 1388 { 1389 struct lock_instance *instance; 1390 struct lock_class *class; 1391 1392 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__)); 1393 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL) 1394 return; 1395 class = LOCK_CLASS(lock); 1396 file = fixup_filename(file); 1397 if (witness_watch) { 1398 if ((lock->lo_flags & LO_UPGRADABLE) == 0) 1399 panic("upgrade of non-upgradable lock (%s) %s @ %s:%d", 1400 class->lc_name, lock->lo_name, file, line); 1401 if ((class->lc_flags & LC_SLEEPLOCK) == 0) 1402 panic("upgrade of non-sleep lock (%s) %s @ %s:%d", 1403 class->lc_name, lock->lo_name, file, line); 1404 } 1405 instance = find_instance(curthread->td_sleeplocks, lock); 1406 if (instance == NULL) 1407 panic("upgrade of unlocked lock (%s) %s @ %s:%d", 1408 class->lc_name, lock->lo_name, file, line); 1409 if (witness_watch) { 1410 if ((instance->li_flags & LI_EXCLUSIVE) != 0) 1411 panic("upgrade of exclusive lock (%s) %s @ %s:%d", 1412 class->lc_name, lock->lo_name, file, line); 1413 if ((instance->li_flags & LI_RECURSEMASK) != 0) 1414 panic("upgrade of recursed lock (%s) %s r=%d @ %s:%d", 1415 class->lc_name, lock->lo_name, 1416 instance->li_flags & LI_RECURSEMASK, file, line); 1417 } 1418 instance->li_flags |= LI_EXCLUSIVE; 1419 } 1420 1421 void 1422 witness_downgrade(struct lock_object *lock, int flags, const char *file, 1423 int line) 1424 { 1425 struct lock_instance *instance; 1426 struct lock_class *class; 1427 1428 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__)); 1429 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL) 1430 return; 1431 class = LOCK_CLASS(lock); 1432 file = fixup_filename(file); 1433 if (witness_watch) { 1434 if ((lock->lo_flags & LO_UPGRADABLE) == 0) 1435 panic("downgrade of non-upgradable lock (%s) %s @ %s:%d", 1436 class->lc_name, lock->lo_name, file, line); 1437 if ((class->lc_flags & LC_SLEEPLOCK) == 0) 1438 panic("downgrade of non-sleep lock (%s) %s @ %s:%d", 1439 class->lc_name, lock->lo_name, file, line); 1440 } 1441 instance = find_instance(curthread->td_sleeplocks, lock); 1442 if (instance == NULL) 1443 panic("downgrade of unlocked lock (%s) %s @ %s:%d", 1444 class->lc_name, lock->lo_name, file, line); 1445 if (witness_watch) { 1446 if ((instance->li_flags & LI_EXCLUSIVE) == 0) 1447 panic("downgrade of shared lock (%s) %s @ %s:%d", 1448 class->lc_name, lock->lo_name, file, line); 1449 if ((instance->li_flags & LI_RECURSEMASK) != 0) 1450 panic("downgrade of recursed lock (%s) %s r=%d @ %s:%d", 1451 class->lc_name, lock->lo_name, 1452 instance->li_flags & LI_RECURSEMASK, file, line); 1453 } 1454 instance->li_flags &= ~LI_EXCLUSIVE; 1455 } 1456 1457 void 1458 witness_unlock(struct lock_object *lock, int flags, const char *file, int line) 1459 { 1460 struct lock_list_entry **lock_list, *lle; 1461 struct lock_instance *instance; 1462 struct lock_class *class; 1463 struct thread *td; 1464 register_t s; 1465 int i, j; 1466 1467 if (witness_cold || lock->lo_witness == NULL || panicstr != NULL) 1468 return; 1469 td = curthread; 1470 class = LOCK_CLASS(lock); 1471 file = fixup_filename(file); 1472 1473 /* Find lock instance associated with this lock. */ 1474 if (class->lc_flags & LC_SLEEPLOCK) 1475 lock_list = &td->td_sleeplocks; 1476 else 1477 lock_list = PCPU_PTR(spinlocks); 1478 lle = *lock_list; 1479 for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next) 1480 for (i = 0; i < (*lock_list)->ll_count; i++) { 1481 instance = &(*lock_list)->ll_children[i]; 1482 if (instance->li_lock == lock) 1483 goto found; 1484 } 1485 1486 /* 1487 * When disabling WITNESS through witness_watch we could end up in 1488 * having registered locks in the td_sleeplocks queue. 1489 * We have to make sure we flush these queues, so just search for 1490 * eventual register locks and remove them. 1491 */ 1492 if (witness_watch > 0) 1493 panic("lock (%s) %s not locked @ %s:%d", class->lc_name, 1494 lock->lo_name, file, line); 1495 else 1496 return; 1497 found: 1498 1499 /* First, check for shared/exclusive mismatches. */ 1500 if ((instance->li_flags & LI_EXCLUSIVE) != 0 && witness_watch > 0 && 1501 (flags & LOP_EXCLUSIVE) == 0) { 1502 printf("shared unlock of (%s) %s @ %s:%d\n", class->lc_name, 1503 lock->lo_name, file, line); 1504 printf("while exclusively locked from %s:%d\n", 1505 instance->li_file, instance->li_line); 1506 panic("excl->ushare"); 1507 } 1508 if ((instance->li_flags & LI_EXCLUSIVE) == 0 && witness_watch > 0 && 1509 (flags & LOP_EXCLUSIVE) != 0) { 1510 printf("exclusive unlock of (%s) %s @ %s:%d\n", class->lc_name, 1511 lock->lo_name, file, line); 1512 printf("while share locked from %s:%d\n", instance->li_file, 1513 instance->li_line); 1514 panic("share->uexcl"); 1515 } 1516 /* If we are recursed, unrecurse. */ 1517 if ((instance->li_flags & LI_RECURSEMASK) > 0) { 1518 CTR4(KTR_WITNESS, "%s: pid %d unrecursed on %s r=%d", __func__, 1519 td->td_proc->p_pid, instance->li_lock->lo_name, 1520 instance->li_flags); 1521 instance->li_flags--; 1522 return; 1523 } 1524 /* The lock is now being dropped, check for NORELEASE flag */ 1525 if ((instance->li_flags & LI_NORELEASE) != 0 && witness_watch > 0) { 1526 printf("forbidden unlock of (%s) %s @ %s:%d\n", class->lc_name, 1527 lock->lo_name, file, line); 1528 panic("lock marked norelease"); 1529 } 1530 1531 /* Otherwise, remove this item from the list. */ 1532 s = intr_disable(); 1533 CTR4(KTR_WITNESS, "%s: pid %d removed %s from lle[%d]", __func__, 1534 td->td_proc->p_pid, instance->li_lock->lo_name, 1535 (*lock_list)->ll_count - 1); 1536 for (j = i; j < (*lock_list)->ll_count - 1; j++) 1537 (*lock_list)->ll_children[j] = 1538 (*lock_list)->ll_children[j + 1]; 1539 (*lock_list)->ll_count--; 1540 intr_restore(s); 1541 1542 /* 1543 * In order to reduce contention on w_mtx, we want to keep always an 1544 * head object into lists so that frequent allocation from the 1545 * free witness pool (and subsequent locking) is avoided. 1546 * In order to maintain the current code simple, when the head 1547 * object is totally unloaded it means also that we do not have 1548 * further objects in the list, so the list ownership needs to be 1549 * hand over to another object if the current head needs to be freed. 1550 */ 1551 if ((*lock_list)->ll_count == 0) { 1552 if (*lock_list == lle) { 1553 if (lle->ll_next == NULL) 1554 return; 1555 } else 1556 lle = *lock_list; 1557 *lock_list = lle->ll_next; 1558 CTR3(KTR_WITNESS, "%s: pid %d removed lle %p", __func__, 1559 td->td_proc->p_pid, lle); 1560 witness_lock_list_free(lle); 1561 } 1562 } 1563 1564 void 1565 witness_thread_exit(struct thread *td) 1566 { 1567 struct lock_list_entry *lle; 1568 int i, n; 1569 1570 lle = td->td_sleeplocks; 1571 if (lle == NULL || panicstr != NULL) 1572 return; 1573 if (lle->ll_count != 0) { 1574 for (n = 0; lle != NULL; lle = lle->ll_next) 1575 for (i = lle->ll_count - 1; i >= 0; i--) { 1576 if (n == 0) 1577 printf("Thread %p exiting with the following locks held:\n", 1578 td); 1579 n++; 1580 witness_list_lock(&lle->ll_children[i]); 1581 1582 } 1583 panic("Thread %p cannot exit while holding sleeplocks\n", td); 1584 } 1585 witness_lock_list_free(lle); 1586 } 1587 1588 /* 1589 * Warn if any locks other than 'lock' are held. Flags can be passed in to 1590 * exempt Giant and sleepable locks from the checks as well. If any 1591 * non-exempt locks are held, then a supplied message is printed to the 1592 * console along with a list of the offending locks. If indicated in the 1593 * flags then a failure results in a panic as well. 1594 */ 1595 int 1596 witness_warn(int flags, struct lock_object *lock, const char *fmt, ...) 1597 { 1598 struct lock_list_entry *lock_list, *lle; 1599 struct lock_instance *lock1; 1600 struct thread *td; 1601 va_list ap; 1602 int i, n; 1603 1604 if (witness_cold || witness_watch < 1 || panicstr != NULL) 1605 return (0); 1606 n = 0; 1607 td = curthread; 1608 for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next) 1609 for (i = lle->ll_count - 1; i >= 0; i--) { 1610 lock1 = &lle->ll_children[i]; 1611 if (lock1->li_lock == lock) 1612 continue; 1613 if (flags & WARN_GIANTOK && 1614 lock1->li_lock == &Giant.lock_object) 1615 continue; 1616 if (flags & WARN_SLEEPOK && 1617 (lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0) 1618 continue; 1619 if (n == 0) { 1620 va_start(ap, fmt); 1621 vprintf(fmt, ap); 1622 va_end(ap); 1623 printf(" with the following"); 1624 if (flags & WARN_SLEEPOK) 1625 printf(" non-sleepable"); 1626 printf(" locks held:\n"); 1627 } 1628 n++; 1629 witness_list_lock(lock1); 1630 } 1631 1632 /* 1633 * Pin the thread in order to avoid problems with thread migration. 1634 * Once that all verifies are passed about spinlocks ownership, 1635 * the thread is in a safe path and it can be unpinned. 1636 */ 1637 sched_pin(); 1638 lock_list = PCPU_GET(spinlocks); 1639 if (lock_list != NULL && lock_list->ll_count != 0) { 1640 sched_unpin(); 1641 1642 /* 1643 * We should only have one spinlock and as long as 1644 * the flags cannot match for this locks class, 1645 * check if the first spinlock is the one curthread 1646 * should hold. 1647 */ 1648 lock1 = &lock_list->ll_children[lock_list->ll_count - 1]; 1649 if (lock_list->ll_count == 1 && lock_list->ll_next == NULL && 1650 lock1->li_lock == lock && n == 0) 1651 return (0); 1652 1653 va_start(ap, fmt); 1654 vprintf(fmt, ap); 1655 va_end(ap); 1656 printf(" with the following"); 1657 if (flags & WARN_SLEEPOK) 1658 printf(" non-sleepable"); 1659 printf(" locks held:\n"); 1660 n += witness_list_locks(&lock_list); 1661 } else 1662 sched_unpin(); 1663 if (flags & WARN_PANIC && n) 1664 panic("%s", __func__); 1665 else 1666 witness_debugger(n); 1667 return (n); 1668 } 1669 1670 const char * 1671 witness_file(struct lock_object *lock) 1672 { 1673 struct witness *w; 1674 1675 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL) 1676 return ("?"); 1677 w = lock->lo_witness; 1678 return (w->w_file); 1679 } 1680 1681 int 1682 witness_line(struct lock_object *lock) 1683 { 1684 struct witness *w; 1685 1686 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL) 1687 return (0); 1688 w = lock->lo_witness; 1689 return (w->w_line); 1690 } 1691 1692 static struct witness * 1693 enroll(const char *description, struct lock_class *lock_class) 1694 { 1695 struct witness *w; 1696 struct witness_list *typelist; 1697 1698 MPASS(description != NULL); 1699 1700 if (witness_watch == -1 || panicstr != NULL) 1701 return (NULL); 1702 if ((lock_class->lc_flags & LC_SPINLOCK)) { 1703 if (witness_skipspin) 1704 return (NULL); 1705 else 1706 typelist = &w_spin; 1707 } else if ((lock_class->lc_flags & LC_SLEEPLOCK)) 1708 typelist = &w_sleep; 1709 else 1710 panic("lock class %s is not sleep or spin", 1711 lock_class->lc_name); 1712 1713 mtx_lock_spin(&w_mtx); 1714 w = witness_hash_get(description); 1715 if (w) 1716 goto found; 1717 if ((w = witness_get()) == NULL) 1718 return (NULL); 1719 MPASS(strlen(description) < MAX_W_NAME); 1720 strcpy(w->w_name, description); 1721 w->w_class = lock_class; 1722 w->w_refcount = 1; 1723 STAILQ_INSERT_HEAD(&w_all, w, w_list); 1724 if (lock_class->lc_flags & LC_SPINLOCK) { 1725 STAILQ_INSERT_HEAD(&w_spin, w, w_typelist); 1726 w_spin_cnt++; 1727 } else if (lock_class->lc_flags & LC_SLEEPLOCK) { 1728 STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist); 1729 w_sleep_cnt++; 1730 } 1731 1732 /* Insert new witness into the hash */ 1733 witness_hash_put(w); 1734 witness_increment_graph_generation(); 1735 mtx_unlock_spin(&w_mtx); 1736 return (w); 1737 found: 1738 w->w_refcount++; 1739 mtx_unlock_spin(&w_mtx); 1740 if (lock_class != w->w_class) 1741 panic( 1742 "lock (%s) %s does not match earlier (%s) lock", 1743 description, lock_class->lc_name, 1744 w->w_class->lc_name); 1745 return (w); 1746 } 1747 1748 static void 1749 depart(struct witness *w) 1750 { 1751 struct witness_list *list; 1752 1753 MPASS(w->w_refcount == 0); 1754 if (w->w_class->lc_flags & LC_SLEEPLOCK) { 1755 list = &w_sleep; 1756 w_sleep_cnt--; 1757 } else { 1758 list = &w_spin; 1759 w_spin_cnt--; 1760 } 1761 /* 1762 * Set file to NULL as it may point into a loadable module. 1763 */ 1764 w->w_file = NULL; 1765 w->w_line = 0; 1766 witness_increment_graph_generation(); 1767 } 1768 1769 1770 static void 1771 adopt(struct witness *parent, struct witness *child) 1772 { 1773 int pi, ci, i, j; 1774 1775 if (witness_cold == 0) 1776 mtx_assert(&w_mtx, MA_OWNED); 1777 1778 /* If the relationship is already known, there's no work to be done. */ 1779 if (isitmychild(parent, child)) 1780 return; 1781 1782 /* When the structure of the graph changes, bump up the generation. */ 1783 witness_increment_graph_generation(); 1784 1785 /* 1786 * The hard part ... create the direct relationship, then propagate all 1787 * indirect relationships. 1788 */ 1789 pi = parent->w_index; 1790 ci = child->w_index; 1791 WITNESS_INDEX_ASSERT(pi); 1792 WITNESS_INDEX_ASSERT(ci); 1793 MPASS(pi != ci); 1794 w_rmatrix[pi][ci] |= WITNESS_PARENT; 1795 w_rmatrix[ci][pi] |= WITNESS_CHILD; 1796 1797 /* 1798 * If parent was not already an ancestor of child, 1799 * then we increment the descendant and ancestor counters. 1800 */ 1801 if ((w_rmatrix[pi][ci] & WITNESS_ANCESTOR) == 0) { 1802 parent->w_num_descendants++; 1803 child->w_num_ancestors++; 1804 } 1805 1806 /* 1807 * Find each ancestor of 'pi'. Note that 'pi' itself is counted as 1808 * an ancestor of 'pi' during this loop. 1809 */ 1810 for (i = 1; i <= w_max_used_index; i++) { 1811 if ((w_rmatrix[i][pi] & WITNESS_ANCESTOR_MASK) == 0 && 1812 (i != pi)) 1813 continue; 1814 1815 /* Find each descendant of 'i' and mark it as a descendant. */ 1816 for (j = 1; j <= w_max_used_index; j++) { 1817 1818 /* 1819 * Skip children that are already marked as 1820 * descendants of 'i'. 1821 */ 1822 if (w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK) 1823 continue; 1824 1825 /* 1826 * We are only interested in descendants of 'ci'. Note 1827 * that 'ci' itself is counted as a descendant of 'ci'. 1828 */ 1829 if ((w_rmatrix[ci][j] & WITNESS_ANCESTOR_MASK) == 0 && 1830 (j != ci)) 1831 continue; 1832 w_rmatrix[i][j] |= WITNESS_ANCESTOR; 1833 w_rmatrix[j][i] |= WITNESS_DESCENDANT; 1834 w_data[i].w_num_descendants++; 1835 w_data[j].w_num_ancestors++; 1836 1837 /* 1838 * Make sure we aren't marking a node as both an 1839 * ancestor and descendant. We should have caught 1840 * this as a lock order reversal earlier. 1841 */ 1842 if ((w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK) && 1843 (w_rmatrix[i][j] & WITNESS_DESCENDANT_MASK)) { 1844 printf("witness rmatrix paradox! [%d][%d]=%d " 1845 "both ancestor and descendant\n", 1846 i, j, w_rmatrix[i][j]); 1847 kdb_backtrace(); 1848 printf("Witness disabled.\n"); 1849 witness_watch = -1; 1850 } 1851 if ((w_rmatrix[j][i] & WITNESS_ANCESTOR_MASK) && 1852 (w_rmatrix[j][i] & WITNESS_DESCENDANT_MASK)) { 1853 printf("witness rmatrix paradox! [%d][%d]=%d " 1854 "both ancestor and descendant\n", 1855 j, i, w_rmatrix[j][i]); 1856 kdb_backtrace(); 1857 printf("Witness disabled.\n"); 1858 witness_watch = -1; 1859 } 1860 } 1861 } 1862 } 1863 1864 static void 1865 itismychild(struct witness *parent, struct witness *child) 1866 { 1867 1868 MPASS(child != NULL && parent != NULL); 1869 if (witness_cold == 0) 1870 mtx_assert(&w_mtx, MA_OWNED); 1871 1872 if (!witness_lock_type_equal(parent, child)) { 1873 if (witness_cold == 0) 1874 mtx_unlock_spin(&w_mtx); 1875 panic("%s: parent \"%s\" (%s) and child \"%s\" (%s) are not " 1876 "the same lock type", __func__, parent->w_name, 1877 parent->w_class->lc_name, child->w_name, 1878 child->w_class->lc_name); 1879 } 1880 adopt(parent, child); 1881 } 1882 1883 /* 1884 * Generic code for the isitmy*() functions. The rmask parameter is the 1885 * expected relationship of w1 to w2. 1886 */ 1887 static int 1888 _isitmyx(struct witness *w1, struct witness *w2, int rmask, const char *fname) 1889 { 1890 unsigned char r1, r2; 1891 int i1, i2; 1892 1893 i1 = w1->w_index; 1894 i2 = w2->w_index; 1895 WITNESS_INDEX_ASSERT(i1); 1896 WITNESS_INDEX_ASSERT(i2); 1897 r1 = w_rmatrix[i1][i2] & WITNESS_RELATED_MASK; 1898 r2 = w_rmatrix[i2][i1] & WITNESS_RELATED_MASK; 1899 1900 /* The flags on one better be the inverse of the flags on the other */ 1901 if (!((WITNESS_ATOD(r1) == r2 && WITNESS_DTOA(r2) == r1) || 1902 (WITNESS_DTOA(r1) == r2 && WITNESS_ATOD(r2) == r1))) { 1903 printf("%s: rmatrix mismatch between %s (index %d) and %s " 1904 "(index %d): w_rmatrix[%d][%d] == %hhx but " 1905 "w_rmatrix[%d][%d] == %hhx\n", 1906 fname, w1->w_name, i1, w2->w_name, i2, i1, i2, r1, 1907 i2, i1, r2); 1908 kdb_backtrace(); 1909 printf("Witness disabled.\n"); 1910 witness_watch = -1; 1911 } 1912 return (r1 & rmask); 1913 } 1914 1915 /* 1916 * Checks if @child is a direct child of @parent. 1917 */ 1918 static int 1919 isitmychild(struct witness *parent, struct witness *child) 1920 { 1921 1922 return (_isitmyx(parent, child, WITNESS_PARENT, __func__)); 1923 } 1924 1925 /* 1926 * Checks if @descendant is a direct or inderect descendant of @ancestor. 1927 */ 1928 static int 1929 isitmydescendant(struct witness *ancestor, struct witness *descendant) 1930 { 1931 1932 return (_isitmyx(ancestor, descendant, WITNESS_ANCESTOR_MASK, 1933 __func__)); 1934 } 1935 1936 #ifdef BLESSING 1937 static int 1938 blessed(struct witness *w1, struct witness *w2) 1939 { 1940 int i; 1941 struct witness_blessed *b; 1942 1943 for (i = 0; i < blessed_count; i++) { 1944 b = &blessed_list[i]; 1945 if (strcmp(w1->w_name, b->b_lock1) == 0) { 1946 if (strcmp(w2->w_name, b->b_lock2) == 0) 1947 return (1); 1948 continue; 1949 } 1950 if (strcmp(w1->w_name, b->b_lock2) == 0) 1951 if (strcmp(w2->w_name, b->b_lock1) == 0) 1952 return (1); 1953 } 1954 return (0); 1955 } 1956 #endif 1957 1958 static struct witness * 1959 witness_get(void) 1960 { 1961 struct witness *w; 1962 int index; 1963 1964 if (witness_cold == 0) 1965 mtx_assert(&w_mtx, MA_OWNED); 1966 1967 if (witness_watch == -1) { 1968 mtx_unlock_spin(&w_mtx); 1969 return (NULL); 1970 } 1971 if (STAILQ_EMPTY(&w_free)) { 1972 witness_watch = -1; 1973 mtx_unlock_spin(&w_mtx); 1974 printf("WITNESS: unable to allocate a new witness object\n"); 1975 return (NULL); 1976 } 1977 w = STAILQ_FIRST(&w_free); 1978 STAILQ_REMOVE_HEAD(&w_free, w_list); 1979 w_free_cnt--; 1980 index = w->w_index; 1981 MPASS(index > 0 && index == w_max_used_index+1 && 1982 index < WITNESS_COUNT); 1983 bzero(w, sizeof(*w)); 1984 w->w_index = index; 1985 if (index > w_max_used_index) 1986 w_max_used_index = index; 1987 return (w); 1988 } 1989 1990 static void 1991 witness_free(struct witness *w) 1992 { 1993 1994 STAILQ_INSERT_HEAD(&w_free, w, w_list); 1995 w_free_cnt++; 1996 } 1997 1998 static struct lock_list_entry * 1999 witness_lock_list_get(void) 2000 { 2001 struct lock_list_entry *lle; 2002 2003 if (witness_watch == -1) 2004 return (NULL); 2005 mtx_lock_spin(&w_mtx); 2006 lle = w_lock_list_free; 2007 if (lle == NULL) { 2008 witness_watch = -1; 2009 mtx_unlock_spin(&w_mtx); 2010 printf("%s: witness exhausted\n", __func__); 2011 return (NULL); 2012 } 2013 w_lock_list_free = lle->ll_next; 2014 mtx_unlock_spin(&w_mtx); 2015 bzero(lle, sizeof(*lle)); 2016 return (lle); 2017 } 2018 2019 static void 2020 witness_lock_list_free(struct lock_list_entry *lle) 2021 { 2022 2023 mtx_lock_spin(&w_mtx); 2024 lle->ll_next = w_lock_list_free; 2025 w_lock_list_free = lle; 2026 mtx_unlock_spin(&w_mtx); 2027 } 2028 2029 static struct lock_instance * 2030 find_instance(struct lock_list_entry *list, struct lock_object *lock) 2031 { 2032 struct lock_list_entry *lle; 2033 struct lock_instance *instance; 2034 int i; 2035 2036 for (lle = list; lle != NULL; lle = lle->ll_next) 2037 for (i = lle->ll_count - 1; i >= 0; i--) { 2038 instance = &lle->ll_children[i]; 2039 if (instance->li_lock == lock) 2040 return (instance); 2041 } 2042 return (NULL); 2043 } 2044 2045 static void 2046 witness_list_lock(struct lock_instance *instance) 2047 { 2048 struct lock_object *lock; 2049 2050 lock = instance->li_lock; 2051 printf("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ? 2052 "exclusive" : "shared", LOCK_CLASS(lock)->lc_name, lock->lo_name); 2053 if (lock->lo_witness->w_name != lock->lo_name) 2054 printf(" (%s)", lock->lo_witness->w_name); 2055 printf(" r = %d (%p) locked @ %s:%d\n", 2056 instance->li_flags & LI_RECURSEMASK, lock, instance->li_file, 2057 instance->li_line); 2058 } 2059 2060 #ifdef DDB 2061 static int 2062 witness_thread_has_locks(struct thread *td) 2063 { 2064 2065 if (td->td_sleeplocks == NULL) 2066 return (0); 2067 return (td->td_sleeplocks->ll_count != 0); 2068 } 2069 2070 static int 2071 witness_proc_has_locks(struct proc *p) 2072 { 2073 struct thread *td; 2074 2075 FOREACH_THREAD_IN_PROC(p, td) { 2076 if (witness_thread_has_locks(td)) 2077 return (1); 2078 } 2079 return (0); 2080 } 2081 #endif 2082 2083 int 2084 witness_list_locks(struct lock_list_entry **lock_list) 2085 { 2086 struct lock_list_entry *lle; 2087 int i, nheld; 2088 2089 nheld = 0; 2090 for (lle = *lock_list; lle != NULL; lle = lle->ll_next) 2091 for (i = lle->ll_count - 1; i >= 0; i--) { 2092 witness_list_lock(&lle->ll_children[i]); 2093 nheld++; 2094 } 2095 return (nheld); 2096 } 2097 2098 /* 2099 * This is a bit risky at best. We call this function when we have timed 2100 * out acquiring a spin lock, and we assume that the other CPU is stuck 2101 * with this lock held. So, we go groveling around in the other CPU's 2102 * per-cpu data to try to find the lock instance for this spin lock to 2103 * see when it was last acquired. 2104 */ 2105 void 2106 witness_display_spinlock(struct lock_object *lock, struct thread *owner) 2107 { 2108 struct lock_instance *instance; 2109 struct pcpu *pc; 2110 2111 if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU) 2112 return; 2113 pc = pcpu_find(owner->td_oncpu); 2114 instance = find_instance(pc->pc_spinlocks, lock); 2115 if (instance != NULL) 2116 witness_list_lock(instance); 2117 } 2118 2119 void 2120 witness_save(struct lock_object *lock, const char **filep, int *linep) 2121 { 2122 struct lock_list_entry *lock_list; 2123 struct lock_instance *instance; 2124 struct lock_class *class; 2125 2126 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__)); 2127 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL) 2128 return; 2129 class = LOCK_CLASS(lock); 2130 if (class->lc_flags & LC_SLEEPLOCK) 2131 lock_list = curthread->td_sleeplocks; 2132 else { 2133 if (witness_skipspin) 2134 return; 2135 lock_list = PCPU_GET(spinlocks); 2136 } 2137 instance = find_instance(lock_list, lock); 2138 if (instance == NULL) 2139 panic("%s: lock (%s) %s not locked", __func__, 2140 class->lc_name, lock->lo_name); 2141 *filep = instance->li_file; 2142 *linep = instance->li_line; 2143 } 2144 2145 void 2146 witness_restore(struct lock_object *lock, const char *file, int line) 2147 { 2148 struct lock_list_entry *lock_list; 2149 struct lock_instance *instance; 2150 struct lock_class *class; 2151 2152 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__)); 2153 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL) 2154 return; 2155 class = LOCK_CLASS(lock); 2156 if (class->lc_flags & LC_SLEEPLOCK) 2157 lock_list = curthread->td_sleeplocks; 2158 else { 2159 if (witness_skipspin) 2160 return; 2161 lock_list = PCPU_GET(spinlocks); 2162 } 2163 instance = find_instance(lock_list, lock); 2164 if (instance == NULL) 2165 panic("%s: lock (%s) %s not locked", __func__, 2166 class->lc_name, lock->lo_name); 2167 lock->lo_witness->w_file = file; 2168 lock->lo_witness->w_line = line; 2169 instance->li_file = file; 2170 instance->li_line = line; 2171 } 2172 2173 void 2174 witness_assert(struct lock_object *lock, int flags, const char *file, int line) 2175 { 2176 #ifdef INVARIANT_SUPPORT 2177 struct lock_instance *instance; 2178 struct lock_class *class; 2179 2180 if (lock->lo_witness == NULL || witness_watch < 1 || panicstr != NULL) 2181 return; 2182 class = LOCK_CLASS(lock); 2183 if ((class->lc_flags & LC_SLEEPLOCK) != 0) 2184 instance = find_instance(curthread->td_sleeplocks, lock); 2185 else if ((class->lc_flags & LC_SPINLOCK) != 0) 2186 instance = find_instance(PCPU_GET(spinlocks), lock); 2187 else { 2188 panic("Lock (%s) %s is not sleep or spin!", 2189 class->lc_name, lock->lo_name); 2190 } 2191 file = fixup_filename(file); 2192 switch (flags) { 2193 case LA_UNLOCKED: 2194 if (instance != NULL) 2195 panic("Lock (%s) %s locked @ %s:%d.", 2196 class->lc_name, lock->lo_name, file, line); 2197 break; 2198 case LA_LOCKED: 2199 case LA_LOCKED | LA_RECURSED: 2200 case LA_LOCKED | LA_NOTRECURSED: 2201 case LA_SLOCKED: 2202 case LA_SLOCKED | LA_RECURSED: 2203 case LA_SLOCKED | LA_NOTRECURSED: 2204 case LA_XLOCKED: 2205 case LA_XLOCKED | LA_RECURSED: 2206 case LA_XLOCKED | LA_NOTRECURSED: 2207 if (instance == NULL) { 2208 panic("Lock (%s) %s not locked @ %s:%d.", 2209 class->lc_name, lock->lo_name, file, line); 2210 break; 2211 } 2212 if ((flags & LA_XLOCKED) != 0 && 2213 (instance->li_flags & LI_EXCLUSIVE) == 0) 2214 panic("Lock (%s) %s not exclusively locked @ %s:%d.", 2215 class->lc_name, lock->lo_name, file, line); 2216 if ((flags & LA_SLOCKED) != 0 && 2217 (instance->li_flags & LI_EXCLUSIVE) != 0) 2218 panic("Lock (%s) %s exclusively locked @ %s:%d.", 2219 class->lc_name, lock->lo_name, file, line); 2220 if ((flags & LA_RECURSED) != 0 && 2221 (instance->li_flags & LI_RECURSEMASK) == 0) 2222 panic("Lock (%s) %s not recursed @ %s:%d.", 2223 class->lc_name, lock->lo_name, file, line); 2224 if ((flags & LA_NOTRECURSED) != 0 && 2225 (instance->li_flags & LI_RECURSEMASK) != 0) 2226 panic("Lock (%s) %s recursed @ %s:%d.", 2227 class->lc_name, lock->lo_name, file, line); 2228 break; 2229 default: 2230 panic("Invalid lock assertion at %s:%d.", file, line); 2231 2232 } 2233 #endif /* INVARIANT_SUPPORT */ 2234 } 2235 2236 static void 2237 witness_setflag(struct lock_object *lock, int flag, int set) 2238 { 2239 struct lock_list_entry *lock_list; 2240 struct lock_instance *instance; 2241 struct lock_class *class; 2242 2243 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL) 2244 return; 2245 class = LOCK_CLASS(lock); 2246 if (class->lc_flags & LC_SLEEPLOCK) 2247 lock_list = curthread->td_sleeplocks; 2248 else { 2249 if (witness_skipspin) 2250 return; 2251 lock_list = PCPU_GET(spinlocks); 2252 } 2253 instance = find_instance(lock_list, lock); 2254 if (instance == NULL) 2255 panic("%s: lock (%s) %s not locked", __func__, 2256 class->lc_name, lock->lo_name); 2257 2258 if (set) 2259 instance->li_flags |= flag; 2260 else 2261 instance->li_flags &= ~flag; 2262 } 2263 2264 void 2265 witness_norelease(struct lock_object *lock) 2266 { 2267 2268 witness_setflag(lock, LI_NORELEASE, 1); 2269 } 2270 2271 void 2272 witness_releaseok(struct lock_object *lock) 2273 { 2274 2275 witness_setflag(lock, LI_NORELEASE, 0); 2276 } 2277 2278 #ifdef DDB 2279 static void 2280 witness_ddb_list(struct thread *td) 2281 { 2282 2283 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__)); 2284 KASSERT(kdb_active, ("%s: not in the debugger", __func__)); 2285 2286 if (witness_watch < 1) 2287 return; 2288 2289 witness_list_locks(&td->td_sleeplocks); 2290 2291 /* 2292 * We only handle spinlocks if td == curthread. This is somewhat broken 2293 * if td is currently executing on some other CPU and holds spin locks 2294 * as we won't display those locks. If we had a MI way of getting 2295 * the per-cpu data for a given cpu then we could use 2296 * td->td_oncpu to get the list of spinlocks for this thread 2297 * and "fix" this. 2298 * 2299 * That still wouldn't really fix this unless we locked the scheduler 2300 * lock or stopped the other CPU to make sure it wasn't changing the 2301 * list out from under us. It is probably best to just not try to 2302 * handle threads on other CPU's for now. 2303 */ 2304 if (td == curthread && PCPU_GET(spinlocks) != NULL) 2305 witness_list_locks(PCPU_PTR(spinlocks)); 2306 } 2307 2308 DB_SHOW_COMMAND(locks, db_witness_list) 2309 { 2310 struct thread *td; 2311 2312 if (have_addr) 2313 td = db_lookup_thread(addr, TRUE); 2314 else 2315 td = kdb_thread; 2316 witness_ddb_list(td); 2317 } 2318 2319 DB_SHOW_ALL_COMMAND(locks, db_witness_list_all) 2320 { 2321 struct thread *td; 2322 struct proc *p; 2323 2324 /* 2325 * It would be nice to list only threads and processes that actually 2326 * held sleep locks, but that information is currently not exported 2327 * by WITNESS. 2328 */ 2329 FOREACH_PROC_IN_SYSTEM(p) { 2330 if (!witness_proc_has_locks(p)) 2331 continue; 2332 FOREACH_THREAD_IN_PROC(p, td) { 2333 if (!witness_thread_has_locks(td)) 2334 continue; 2335 db_printf("Process %d (%s) thread %p (%d)\n", p->p_pid, 2336 p->p_comm, td, td->td_tid); 2337 witness_ddb_list(td); 2338 } 2339 } 2340 } 2341 DB_SHOW_ALIAS(alllocks, db_witness_list_all) 2342 2343 DB_SHOW_COMMAND(witness, db_witness_display) 2344 { 2345 2346 witness_ddb_display(db_printf); 2347 } 2348 #endif 2349 2350 static int 2351 sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS) 2352 { 2353 struct witness_lock_order_data *data1, *data2, *tmp_data1, *tmp_data2; 2354 struct witness *tmp_w1, *tmp_w2, *w1, *w2; 2355 struct sbuf *sb; 2356 u_int w_rmatrix1, w_rmatrix2; 2357 int error, generation, i, j; 2358 2359 tmp_data1 = NULL; 2360 tmp_data2 = NULL; 2361 tmp_w1 = NULL; 2362 tmp_w2 = NULL; 2363 if (witness_watch < 1) { 2364 error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning)); 2365 return (error); 2366 } 2367 if (witness_cold) { 2368 error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold)); 2369 return (error); 2370 } 2371 error = 0; 2372 sb = sbuf_new(NULL, NULL, BADSTACK_SBUF_SIZE, SBUF_AUTOEXTEND); 2373 if (sb == NULL) 2374 return (ENOMEM); 2375 2376 /* Allocate and init temporary storage space. */ 2377 tmp_w1 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO); 2378 tmp_w2 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO); 2379 tmp_data1 = malloc(sizeof(struct witness_lock_order_data), M_TEMP, 2380 M_WAITOK | M_ZERO); 2381 tmp_data2 = malloc(sizeof(struct witness_lock_order_data), M_TEMP, 2382 M_WAITOK | M_ZERO); 2383 stack_zero(&tmp_data1->wlod_stack); 2384 stack_zero(&tmp_data2->wlod_stack); 2385 2386 restart: 2387 mtx_lock_spin(&w_mtx); 2388 generation = w_generation; 2389 mtx_unlock_spin(&w_mtx); 2390 sbuf_printf(sb, "Number of known direct relationships is %d\n", 2391 w_lohash.wloh_count); 2392 for (i = 1; i < w_max_used_index; i++) { 2393 mtx_lock_spin(&w_mtx); 2394 if (generation != w_generation) { 2395 mtx_unlock_spin(&w_mtx); 2396 2397 /* The graph has changed, try again. */ 2398 req->oldidx = 0; 2399 sbuf_clear(sb); 2400 goto restart; 2401 } 2402 2403 w1 = &w_data[i]; 2404 if (w1->w_reversed == 0) { 2405 mtx_unlock_spin(&w_mtx); 2406 continue; 2407 } 2408 2409 /* Copy w1 locally so we can release the spin lock. */ 2410 *tmp_w1 = *w1; 2411 mtx_unlock_spin(&w_mtx); 2412 2413 if (tmp_w1->w_reversed == 0) 2414 continue; 2415 for (j = 1; j < w_max_used_index; j++) { 2416 if ((w_rmatrix[i][j] & WITNESS_REVERSAL) == 0 || i > j) 2417 continue; 2418 2419 mtx_lock_spin(&w_mtx); 2420 if (generation != w_generation) { 2421 mtx_unlock_spin(&w_mtx); 2422 2423 /* The graph has changed, try again. */ 2424 req->oldidx = 0; 2425 sbuf_clear(sb); 2426 goto restart; 2427 } 2428 2429 w2 = &w_data[j]; 2430 data1 = witness_lock_order_get(w1, w2); 2431 data2 = witness_lock_order_get(w2, w1); 2432 2433 /* 2434 * Copy information locally so we can release the 2435 * spin lock. 2436 */ 2437 *tmp_w2 = *w2; 2438 w_rmatrix1 = (unsigned int)w_rmatrix[i][j]; 2439 w_rmatrix2 = (unsigned int)w_rmatrix[j][i]; 2440 2441 if (data1) { 2442 stack_zero(&tmp_data1->wlod_stack); 2443 stack_copy(&data1->wlod_stack, 2444 &tmp_data1->wlod_stack); 2445 } 2446 if (data2 && data2 != data1) { 2447 stack_zero(&tmp_data2->wlod_stack); 2448 stack_copy(&data2->wlod_stack, 2449 &tmp_data2->wlod_stack); 2450 } 2451 mtx_unlock_spin(&w_mtx); 2452 2453 sbuf_printf(sb, 2454 "\nLock order reversal between \"%s\"(%s) and \"%s\"(%s)!\n", 2455 tmp_w1->w_name, tmp_w1->w_class->lc_name, 2456 tmp_w2->w_name, tmp_w2->w_class->lc_name); 2457 #if 0 2458 sbuf_printf(sb, 2459 "w_rmatrix[%s][%s] == %x, w_rmatrix[%s][%s] == %x\n", 2460 tmp_w1->name, tmp_w2->w_name, w_rmatrix1, 2461 tmp_w2->name, tmp_w1->w_name, w_rmatrix2); 2462 #endif 2463 if (data1) { 2464 sbuf_printf(sb, 2465 "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n", 2466 tmp_w1->w_name, tmp_w1->w_class->lc_name, 2467 tmp_w2->w_name, tmp_w2->w_class->lc_name); 2468 stack_sbuf_print(sb, &tmp_data1->wlod_stack); 2469 sbuf_printf(sb, "\n"); 2470 } 2471 if (data2 && data2 != data1) { 2472 sbuf_printf(sb, 2473 "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n", 2474 tmp_w2->w_name, tmp_w2->w_class->lc_name, 2475 tmp_w1->w_name, tmp_w1->w_class->lc_name); 2476 stack_sbuf_print(sb, &tmp_data2->wlod_stack); 2477 sbuf_printf(sb, "\n"); 2478 } 2479 } 2480 } 2481 mtx_lock_spin(&w_mtx); 2482 if (generation != w_generation) { 2483 mtx_unlock_spin(&w_mtx); 2484 2485 /* 2486 * The graph changed while we were printing stack data, 2487 * try again. 2488 */ 2489 req->oldidx = 0; 2490 sbuf_clear(sb); 2491 goto restart; 2492 } 2493 mtx_unlock_spin(&w_mtx); 2494 2495 /* Free temporary storage space. */ 2496 free(tmp_data1, M_TEMP); 2497 free(tmp_data2, M_TEMP); 2498 free(tmp_w1, M_TEMP); 2499 free(tmp_w2, M_TEMP); 2500 2501 sbuf_finish(sb); 2502 error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1); 2503 sbuf_delete(sb); 2504 2505 return (error); 2506 } 2507 2508 static int 2509 sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS) 2510 { 2511 struct witness *w; 2512 struct sbuf *sb; 2513 int error; 2514 2515 if (witness_watch < 1) { 2516 error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning)); 2517 return (error); 2518 } 2519 if (witness_cold) { 2520 error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold)); 2521 return (error); 2522 } 2523 error = 0; 2524 sb = sbuf_new(NULL, NULL, FULLGRAPH_SBUF_SIZE, SBUF_FIXEDLEN); 2525 if (sb == NULL) 2526 return (ENOMEM); 2527 sbuf_printf(sb, "\n"); 2528 2529 mtx_lock_spin(&w_mtx); 2530 STAILQ_FOREACH(w, &w_all, w_list) 2531 w->w_displayed = 0; 2532 STAILQ_FOREACH(w, &w_all, w_list) 2533 witness_add_fullgraph(sb, w); 2534 mtx_unlock_spin(&w_mtx); 2535 2536 /* 2537 * While using SBUF_FIXEDLEN, check if the sbuf overflowed. 2538 */ 2539 if (sbuf_overflowed(sb)) { 2540 sbuf_delete(sb); 2541 panic("%s: sbuf overflowed, bump FULLGRAPH_SBUF_SIZE value\n", 2542 __func__); 2543 } 2544 2545 /* 2546 * Close the sbuf and return to userland. 2547 */ 2548 sbuf_finish(sb); 2549 error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1); 2550 sbuf_delete(sb); 2551 2552 return (error); 2553 } 2554 2555 static int 2556 sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS) 2557 { 2558 int error, value; 2559 2560 value = witness_watch; 2561 error = sysctl_handle_int(oidp, &value, 0, req); 2562 if (error != 0 || req->newptr == NULL) 2563 return (error); 2564 if (value > 1 || value < -1 || 2565 (witness_watch == -1 && value != witness_watch)) 2566 return (EINVAL); 2567 witness_watch = value; 2568 return (0); 2569 } 2570 2571 static void 2572 witness_add_fullgraph(struct sbuf *sb, struct witness *w) 2573 { 2574 int i; 2575 2576 if (w->w_displayed != 0 || (w->w_file == NULL && w->w_line == 0)) 2577 return; 2578 w->w_displayed = 1; 2579 2580 WITNESS_INDEX_ASSERT(w->w_index); 2581 for (i = 1; i <= w_max_used_index; i++) { 2582 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT) { 2583 sbuf_printf(sb, "\"%s\",\"%s\"\n", w->w_name, 2584 w_data[i].w_name); 2585 witness_add_fullgraph(sb, &w_data[i]); 2586 } 2587 } 2588 } 2589 2590 /* 2591 * A simple hash function. Takes a key pointer and a key size. If size == 0, 2592 * interprets the key as a string and reads until the null 2593 * terminator. Otherwise, reads the first size bytes. Returns an unsigned 32-bit 2594 * hash value computed from the key. 2595 */ 2596 static uint32_t 2597 witness_hash_djb2(const uint8_t *key, uint32_t size) 2598 { 2599 unsigned int hash = 5381; 2600 int i; 2601 2602 /* hash = hash * 33 + key[i] */ 2603 if (size) 2604 for (i = 0; i < size; i++) 2605 hash = ((hash << 5) + hash) + (unsigned int)key[i]; 2606 else 2607 for (i = 0; key[i] != 0; i++) 2608 hash = ((hash << 5) + hash) + (unsigned int)key[i]; 2609 2610 return (hash); 2611 } 2612 2613 2614 /* 2615 * Initializes the two witness hash tables. Called exactly once from 2616 * witness_initialize(). 2617 */ 2618 static void 2619 witness_init_hash_tables(void) 2620 { 2621 int i; 2622 2623 MPASS(witness_cold); 2624 2625 /* Initialize the hash tables. */ 2626 for (i = 0; i < WITNESS_HASH_SIZE; i++) 2627 w_hash.wh_array[i] = NULL; 2628 2629 w_hash.wh_size = WITNESS_HASH_SIZE; 2630 w_hash.wh_count = 0; 2631 2632 /* Initialize the lock order data hash. */ 2633 w_lofree = NULL; 2634 for (i = 0; i < WITNESS_LO_DATA_COUNT; i++) { 2635 memset(&w_lodata[i], 0, sizeof(w_lodata[i])); 2636 w_lodata[i].wlod_next = w_lofree; 2637 w_lofree = &w_lodata[i]; 2638 } 2639 w_lohash.wloh_size = WITNESS_LO_HASH_SIZE; 2640 w_lohash.wloh_count = 0; 2641 for (i = 0; i < WITNESS_LO_HASH_SIZE; i++) 2642 w_lohash.wloh_array[i] = NULL; 2643 } 2644 2645 static struct witness * 2646 witness_hash_get(const char *key) 2647 { 2648 struct witness *w; 2649 uint32_t hash; 2650 2651 MPASS(key != NULL); 2652 if (witness_cold == 0) 2653 mtx_assert(&w_mtx, MA_OWNED); 2654 hash = witness_hash_djb2(key, 0) % w_hash.wh_size; 2655 w = w_hash.wh_array[hash]; 2656 while (w != NULL) { 2657 if (strcmp(w->w_name, key) == 0) 2658 goto out; 2659 w = w->w_hash_next; 2660 } 2661 2662 out: 2663 return (w); 2664 } 2665 2666 static void 2667 witness_hash_put(struct witness *w) 2668 { 2669 uint32_t hash; 2670 2671 MPASS(w != NULL); 2672 MPASS(w->w_name != NULL); 2673 if (witness_cold == 0) 2674 mtx_assert(&w_mtx, MA_OWNED); 2675 KASSERT(witness_hash_get(w->w_name) == NULL, 2676 ("%s: trying to add a hash entry that already exists!", __func__)); 2677 KASSERT(w->w_hash_next == NULL, 2678 ("%s: w->w_hash_next != NULL", __func__)); 2679 2680 hash = witness_hash_djb2(w->w_name, 0) % w_hash.wh_size; 2681 w->w_hash_next = w_hash.wh_array[hash]; 2682 w_hash.wh_array[hash] = w; 2683 w_hash.wh_count++; 2684 } 2685 2686 2687 static struct witness_lock_order_data * 2688 witness_lock_order_get(struct witness *parent, struct witness *child) 2689 { 2690 struct witness_lock_order_data *data = NULL; 2691 struct witness_lock_order_key key; 2692 unsigned int hash; 2693 2694 MPASS(parent != NULL && child != NULL); 2695 key.from = parent->w_index; 2696 key.to = child->w_index; 2697 WITNESS_INDEX_ASSERT(key.from); 2698 WITNESS_INDEX_ASSERT(key.to); 2699 if ((w_rmatrix[parent->w_index][child->w_index] 2700 & WITNESS_LOCK_ORDER_KNOWN) == 0) 2701 goto out; 2702 2703 hash = witness_hash_djb2((const char*)&key, 2704 sizeof(key)) % w_lohash.wloh_size; 2705 data = w_lohash.wloh_array[hash]; 2706 while (data != NULL) { 2707 if (witness_lock_order_key_equal(&data->wlod_key, &key)) 2708 break; 2709 data = data->wlod_next; 2710 } 2711 2712 out: 2713 return (data); 2714 } 2715 2716 /* 2717 * Verify that parent and child have a known relationship, are not the same, 2718 * and child is actually a child of parent. This is done without w_mtx 2719 * to avoid contention in the common case. 2720 */ 2721 static int 2722 witness_lock_order_check(struct witness *parent, struct witness *child) 2723 { 2724 2725 if (parent != child && 2726 w_rmatrix[parent->w_index][child->w_index] 2727 & WITNESS_LOCK_ORDER_KNOWN && 2728 isitmychild(parent, child)) 2729 return (1); 2730 2731 return (0); 2732 } 2733 2734 static int 2735 witness_lock_order_add(struct witness *parent, struct witness *child) 2736 { 2737 struct witness_lock_order_data *data = NULL; 2738 struct witness_lock_order_key key; 2739 unsigned int hash; 2740 2741 MPASS(parent != NULL && child != NULL); 2742 key.from = parent->w_index; 2743 key.to = child->w_index; 2744 WITNESS_INDEX_ASSERT(key.from); 2745 WITNESS_INDEX_ASSERT(key.to); 2746 if (w_rmatrix[parent->w_index][child->w_index] 2747 & WITNESS_LOCK_ORDER_KNOWN) 2748 return (1); 2749 2750 hash = witness_hash_djb2((const char*)&key, 2751 sizeof(key)) % w_lohash.wloh_size; 2752 w_rmatrix[parent->w_index][child->w_index] |= WITNESS_LOCK_ORDER_KNOWN; 2753 data = w_lofree; 2754 if (data == NULL) 2755 return (0); 2756 w_lofree = data->wlod_next; 2757 data->wlod_next = w_lohash.wloh_array[hash]; 2758 data->wlod_key = key; 2759 w_lohash.wloh_array[hash] = data; 2760 w_lohash.wloh_count++; 2761 stack_zero(&data->wlod_stack); 2762 stack_save(&data->wlod_stack); 2763 return (1); 2764 } 2765 2766 /* Call this whenver the structure of the witness graph changes. */ 2767 static void 2768 witness_increment_graph_generation(void) 2769 { 2770 2771 if (witness_cold == 0) 2772 mtx_assert(&w_mtx, MA_OWNED); 2773 w_generation++; 2774 } 2775 2776 #ifdef KDB 2777 static void 2778 _witness_debugger(int cond, const char *msg) 2779 { 2780 2781 if (witness_trace && cond) 2782 kdb_backtrace(); 2783 if (witness_kdb && cond) 2784 kdb_enter(KDB_WHY_WITNESS, msg); 2785 } 2786 #endif 2787