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 768 139 140/* Allocate 256 KB of stack data space */ 141#define WITNESS_LO_DATA_COUNT 2048 142 143/* Prime, gives load factor of ~2 at full load */ 144#define WITNESS_LO_HASH_SIZE 1021 145 146/* 147 * XXX: This is somewhat bogus, as we assume here that at most 2048 threads 148 * will hold LOCK_NCHILDREN locks. We handle failure ok, and we should 149 * probably be safe for the most part, but it's still a SWAG. 150 */ 151#define LOCK_NCHILDREN 5 152#define LOCK_CHILDCOUNT 2048 153 154#define MAX_W_NAME 64 155 156#define BADSTACK_SBUF_SIZE (256 * WITNESS_COUNT) 157#define FULLGRAPH_SBUF_SIZE 512 158 159/* 160 * These flags go in the witness relationship matrix and describe the 161 * relationship between any two struct witness objects. 162 */ 163#define WITNESS_UNRELATED 0x00 /* No lock order relation. */ 164#define WITNESS_PARENT 0x01 /* Parent, aka direct ancestor. */ 165#define WITNESS_ANCESTOR 0x02 /* Direct or indirect ancestor. */ 166#define WITNESS_CHILD 0x04 /* Child, aka direct descendant. */ 167#define WITNESS_DESCENDANT 0x08 /* Direct or indirect descendant. */ 168#define WITNESS_ANCESTOR_MASK (WITNESS_PARENT | WITNESS_ANCESTOR) 169#define WITNESS_DESCENDANT_MASK (WITNESS_CHILD | WITNESS_DESCENDANT) 170#define WITNESS_RELATED_MASK \ 171 (WITNESS_ANCESTOR_MASK | WITNESS_DESCENDANT_MASK) 172#define WITNESS_REVERSAL 0x10 /* A lock order reversal has been 173 * observed. */ 174#define WITNESS_RESERVED1 0x20 /* Unused flag, reserved. */ 175#define WITNESS_RESERVED2 0x40 /* Unused flag, reserved. */ 176#define WITNESS_LOCK_ORDER_KNOWN 0x80 /* This lock order is known. */ 177 178/* Descendant to ancestor flags */ 179#define WITNESS_DTOA(x) (((x) & WITNESS_RELATED_MASK) >> 2) 180 181/* Ancestor to descendant flags */ 182#define WITNESS_ATOD(x) (((x) & WITNESS_RELATED_MASK) << 2) 183 184#define WITNESS_INDEX_ASSERT(i) \ 185 MPASS((i) > 0 && (i) <= w_max_used_index && (i) < WITNESS_COUNT) 186 187static MALLOC_DEFINE(M_WITNESS, "Witness", "Witness"); 188 189/* 190 * Lock instances. A lock instance is the data associated with a lock while 191 * it is held by witness. For example, a lock instance will hold the 192 * recursion count of a lock. Lock instances are held in lists. Spin locks 193 * are held in a per-cpu list while sleep locks are held in per-thread list. 194 */ 195struct lock_instance { 196 struct lock_object *li_lock; 197 const char *li_file; 198 int li_line; 199 u_int li_flags; 200}; 201 202/* 203 * A simple list type used to build the list of locks held by a thread 204 * or CPU. We can't simply embed the list in struct lock_object since a 205 * lock may be held by more than one thread if it is a shared lock. Locks 206 * are added to the head of the list, so we fill up each list entry from 207 * "the back" logically. To ease some of the arithmetic, we actually fill 208 * in each list entry the normal way (children[0] then children[1], etc.) but 209 * when we traverse the list we read children[count-1] as the first entry 210 * down to children[0] as the final entry. 211 */ 212struct lock_list_entry { 213 struct lock_list_entry *ll_next; 214 struct lock_instance ll_children[LOCK_NCHILDREN]; 215 u_int ll_count; 216}; 217 218/* 219 * The main witness structure. One of these per named lock type in the system 220 * (for example, "vnode interlock"). 221 */ 222struct witness { 223 char w_name[MAX_W_NAME]; 224 uint32_t w_index; /* Index in the relationship matrix */ 225 struct lock_class *w_class; 226 STAILQ_ENTRY(witness) w_list; /* List of all witnesses. */ 227 STAILQ_ENTRY(witness) w_typelist; /* Witnesses of a type. */ 228 struct witness *w_hash_next; /* Linked list in hash buckets. */ 229 const char *w_file; /* File where last acquired */ 230 uint32_t w_line; /* Line where last acquired */ 231 uint32_t w_refcount; 232 uint16_t w_num_ancestors; /* direct/indirect 233 * ancestor count */ 234 uint16_t w_num_descendants; /* direct/indirect 235 * descendant count */ 236 int16_t w_ddb_level; 237 unsigned w_displayed:1; 238 unsigned w_reversed:1; 239}; 240 241STAILQ_HEAD(witness_list, witness); 242 243/* 244 * The witness hash table. Keys are witness names (const char *), elements are 245 * witness objects (struct witness *). 246 */ 247struct witness_hash { 248 struct witness *wh_array[WITNESS_HASH_SIZE]; 249 uint32_t wh_size; 250 uint32_t wh_count; 251}; 252 253/* 254 * Key type for the lock order data hash table. 255 */ 256struct witness_lock_order_key { 257 uint16_t from; 258 uint16_t to; 259}; 260 261struct witness_lock_order_data { 262 struct stack wlod_stack; 263 struct witness_lock_order_key wlod_key; 264 struct witness_lock_order_data *wlod_next; 265}; 266 267/* 268 * The witness lock order data hash table. Keys are witness index tuples 269 * (struct witness_lock_order_key), elements are lock order data objects 270 * (struct witness_lock_order_data). 271 */ 272struct witness_lock_order_hash { 273 struct witness_lock_order_data *wloh_array[WITNESS_LO_HASH_SIZE]; 274 u_int wloh_size; 275 u_int wloh_count; 276}; 277 278#ifdef BLESSING 279struct witness_blessed { 280 const char *b_lock1; 281 const char *b_lock2; 282}; 283#endif 284 285struct witness_pendhelp { 286 const char *wh_type; 287 struct lock_object *wh_lock; 288}; 289 290struct witness_order_list_entry { 291 const char *w_name; 292 struct lock_class *w_class; 293}; 294 295/* 296 * Returns 0 if one of the locks is a spin lock and the other is not. 297 * Returns 1 otherwise. 298 */ 299static __inline int 300witness_lock_type_equal(struct witness *w1, struct witness *w2) 301{ 302 303 return ((w1->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) == 304 (w2->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK))); 305} 306 307static __inline int 308witness_lock_order_key_equal(const struct witness_lock_order_key *a, 309 const struct witness_lock_order_key *b) 310{ 311 312 return (a->from == b->from && a->to == b->to); 313} 314 315static int _isitmyx(struct witness *w1, struct witness *w2, int rmask, 316 const char *fname); 317#ifdef KDB 318static void _witness_debugger(int cond, const char *msg); 319#endif 320static void adopt(struct witness *parent, struct witness *child); 321#ifdef BLESSING 322static int blessed(struct witness *, struct witness *); 323#endif 324static void depart(struct witness *w); 325static struct witness *enroll(const char *description, 326 struct lock_class *lock_class); 327static struct lock_instance *find_instance(struct lock_list_entry *list, 328 struct lock_object *lock); 329static int isitmychild(struct witness *parent, struct witness *child); 330static int isitmydescendant(struct witness *parent, struct witness *child); 331static void itismychild(struct witness *parent, struct witness *child); 332static int sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS); 333static int sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS); 334static int sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS); 335static void witness_add_fullgraph(struct sbuf *sb, struct witness *parent); 336#ifdef DDB 337static void witness_ddb_compute_levels(void); 338static void witness_ddb_display(int(*)(const char *fmt, ...)); 339static void witness_ddb_display_descendants(int(*)(const char *fmt, ...), 340 struct witness *, int indent); 341static void witness_ddb_display_list(int(*prnt)(const char *fmt, ...), 342 struct witness_list *list); 343static void witness_ddb_level_descendants(struct witness *parent, int l); 344static void witness_ddb_list(struct thread *td); 345#endif 346static void witness_free(struct witness *m); 347static struct witness *witness_get(void); 348static uint32_t witness_hash_djb2(const uint8_t *key, uint32_t size); 349static struct witness *witness_hash_get(const char *key); 350static void witness_hash_put(struct witness *w); 351static void witness_init_hash_tables(void); 352static void witness_increment_graph_generation(void); 353static void witness_lock_list_free(struct lock_list_entry *lle); 354static struct lock_list_entry *witness_lock_list_get(void); 355static int witness_lock_order_add(struct witness *parent, 356 struct witness *child); 357static int witness_lock_order_check(struct witness *parent, 358 struct witness *child); 359static struct witness_lock_order_data *witness_lock_order_get( 360 struct witness *parent, 361 struct witness *child); 362static void witness_list_lock(struct lock_instance *instance, 363 int (*prnt)(const char *fmt, ...)); 364static void witness_setflag(struct lock_object *lock, int flag, int set); 365 366#ifdef KDB 367#define witness_debugger(c) _witness_debugger(c, __func__) 368#else 369#define witness_debugger(c) 370#endif 371 372static SYSCTL_NODE(_debug, OID_AUTO, witness, CTLFLAG_RW, NULL, 373 "Witness Locking"); 374 375/* 376 * If set to 0, lock order checking is disabled. If set to -1, 377 * witness is completely disabled. Otherwise witness performs full 378 * lock order checking for all locks. At runtime, lock order checking 379 * may be toggled. However, witness cannot be reenabled once it is 380 * completely disabled. 381 */ 382static int witness_watch = 1; 383TUNABLE_INT("debug.witness.watch", &witness_watch); 384SYSCTL_PROC(_debug_witness, OID_AUTO, watch, CTLFLAG_RW | CTLTYPE_INT, NULL, 0, 385 sysctl_debug_witness_watch, "I", "witness is watching lock operations"); 386 387#ifdef KDB 388/* 389 * When KDB is enabled and witness_kdb is 1, it will cause the system 390 * to drop into kdebug() when: 391 * - a lock hierarchy violation occurs 392 * - locks are held when going to sleep. 393 */ 394#ifdef WITNESS_KDB 395int witness_kdb = 1; 396#else 397int witness_kdb = 0; 398#endif 399TUNABLE_INT("debug.witness.kdb", &witness_kdb); 400SYSCTL_INT(_debug_witness, OID_AUTO, kdb, CTLFLAG_RW, &witness_kdb, 0, ""); 401 402/* 403 * When KDB is enabled and witness_trace is 1, it will cause the system 404 * to print a stack trace: 405 * - a lock hierarchy violation occurs 406 * - locks are held when going to sleep. 407 */ 408int witness_trace = 1; 409TUNABLE_INT("debug.witness.trace", &witness_trace); 410SYSCTL_INT(_debug_witness, OID_AUTO, trace, CTLFLAG_RW, &witness_trace, 0, ""); 411#endif /* KDB */ 412 413#ifdef WITNESS_SKIPSPIN 414int witness_skipspin = 1; 415#else 416int witness_skipspin = 0; 417#endif 418TUNABLE_INT("debug.witness.skipspin", &witness_skipspin); 419SYSCTL_INT(_debug_witness, OID_AUTO, skipspin, CTLFLAG_RDTUN, &witness_skipspin, 420 0, ""); 421 422/* 423 * Call this to print out the relations between locks. 424 */ 425SYSCTL_PROC(_debug_witness, OID_AUTO, fullgraph, CTLTYPE_STRING | CTLFLAG_RD, 426 NULL, 0, sysctl_debug_witness_fullgraph, "A", "Show locks relation graphs"); 427 428/* 429 * Call this to print out the witness faulty stacks. 430 */ 431SYSCTL_PROC(_debug_witness, OID_AUTO, badstacks, CTLTYPE_STRING | CTLFLAG_RD, 432 NULL, 0, sysctl_debug_witness_badstacks, "A", "Show bad witness stacks"); 433 434static struct mtx w_mtx; 435 436/* w_list */ 437static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free); 438static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all); 439 440/* w_typelist */ 441static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin); 442static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep); 443 444/* lock list */ 445static struct lock_list_entry *w_lock_list_free = NULL; 446static struct witness_pendhelp pending_locks[WITNESS_PENDLIST]; 447static u_int pending_cnt; 448 449static int w_free_cnt, w_spin_cnt, w_sleep_cnt; 450SYSCTL_INT(_debug_witness, OID_AUTO, free_cnt, CTLFLAG_RD, &w_free_cnt, 0, ""); 451SYSCTL_INT(_debug_witness, OID_AUTO, spin_cnt, CTLFLAG_RD, &w_spin_cnt, 0, ""); 452SYSCTL_INT(_debug_witness, OID_AUTO, sleep_cnt, CTLFLAG_RD, &w_sleep_cnt, 0, 453 ""); 454 455static struct witness *w_data; 456static uint8_t w_rmatrix[WITNESS_COUNT+1][WITNESS_COUNT+1]; 457static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT]; 458static struct witness_hash w_hash; /* The witness hash table. */ 459 460/* The lock order data hash */ 461static struct witness_lock_order_data w_lodata[WITNESS_LO_DATA_COUNT]; 462static struct witness_lock_order_data *w_lofree = NULL; 463static struct witness_lock_order_hash w_lohash; 464static int w_max_used_index = 0; 465static unsigned int w_generation = 0; 466static const char w_notrunning[] = "Witness not running\n"; 467static const char w_stillcold[] = "Witness is still cold\n"; 468 469 470static struct witness_order_list_entry order_lists[] = { 471 /* 472 * sx locks 473 */ 474 { "proctree", &lock_class_sx }, 475 { "allproc", &lock_class_sx }, 476 { "allprison", &lock_class_sx }, 477 { NULL, NULL }, 478 /* 479 * Various mutexes 480 */ 481 { "Giant", &lock_class_mtx_sleep }, 482 { "pipe mutex", &lock_class_mtx_sleep }, 483 { "sigio lock", &lock_class_mtx_sleep }, 484 { "process group", &lock_class_mtx_sleep }, 485 { "process lock", &lock_class_mtx_sleep }, 486 { "session", &lock_class_mtx_sleep }, 487 { "uidinfo hash", &lock_class_rw }, 488#ifdef HWPMC_HOOKS 489 { "pmc-sleep", &lock_class_mtx_sleep }, 490#endif 491 { "time lock", &lock_class_mtx_sleep }, 492 { NULL, NULL }, 493 /* 494 * Sockets 495 */ 496 { "accept", &lock_class_mtx_sleep }, 497 { "so_snd", &lock_class_mtx_sleep }, 498 { "so_rcv", &lock_class_mtx_sleep }, 499 { "sellck", &lock_class_mtx_sleep }, 500 { NULL, NULL }, 501 /* 502 * Routing 503 */ 504 { "so_rcv", &lock_class_mtx_sleep }, 505 { "radix node head", &lock_class_rw }, 506 { "rtentry", &lock_class_mtx_sleep }, 507 { "ifaddr", &lock_class_mtx_sleep }, 508 { NULL, NULL }, 509 /* 510 * IPv4 multicast: 511 * protocol locks before interface locks, after UDP locks. 512 */ 513 { "udpinp", &lock_class_rw }, 514 { "in_multi_mtx", &lock_class_mtx_sleep }, 515 { "igmp_mtx", &lock_class_mtx_sleep }, 516 { "if_addr_mtx", &lock_class_mtx_sleep }, 517 { NULL, NULL }, 518 /* 519 * IPv6 multicast: 520 * protocol locks before interface locks, after UDP locks. 521 */ 522 { "udpinp", &lock_class_rw }, 523 { "in6_multi_mtx", &lock_class_mtx_sleep }, 524 { "mld_mtx", &lock_class_mtx_sleep }, 525 { "if_addr_mtx", &lock_class_mtx_sleep }, 526 { NULL, NULL }, 527 /* 528 * UNIX Domain Sockets 529 */ 530 { "unp_global_rwlock", &lock_class_rw }, 531 { "unp_list_lock", &lock_class_mtx_sleep }, 532 { "unp", &lock_class_mtx_sleep }, 533 { "so_snd", &lock_class_mtx_sleep }, 534 { NULL, NULL }, 535 /* 536 * UDP/IP 537 */ 538 { "udp", &lock_class_rw }, 539 { "udpinp", &lock_class_rw }, 540 { "so_snd", &lock_class_mtx_sleep }, 541 { NULL, NULL }, 542 /* 543 * TCP/IP 544 */ 545 { "tcp", &lock_class_rw }, 546 { "tcpinp", &lock_class_rw }, 547 { "so_snd", &lock_class_mtx_sleep }, 548 { NULL, NULL }, 549 /* 550 * netatalk 551 */ 552 { "ddp_list_mtx", &lock_class_mtx_sleep }, 553 { "ddp_mtx", &lock_class_mtx_sleep }, 554 { NULL, NULL }, 555 /* 556 * BPF 557 */ 558 { "bpf global lock", &lock_class_mtx_sleep }, 559 { "bpf interface lock", &lock_class_rw }, 560 { "bpf cdev lock", &lock_class_mtx_sleep }, 561 { NULL, NULL }, 562 /* 563 * NFS server 564 */ 565 { "nfsd_mtx", &lock_class_mtx_sleep }, 566 { "so_snd", &lock_class_mtx_sleep }, 567 { NULL, NULL }, 568 569 /* 570 * IEEE 802.11 571 */ 572 { "802.11 com lock", &lock_class_mtx_sleep}, 573 { NULL, NULL }, 574 /* 575 * Network drivers 576 */ 577 { "network driver", &lock_class_mtx_sleep}, 578 { NULL, NULL }, 579 580 /* 581 * Netgraph 582 */ 583 { "ng_node", &lock_class_mtx_sleep }, 584 { "ng_worklist", &lock_class_mtx_sleep }, 585 { NULL, NULL }, 586 /* 587 * CDEV 588 */ 589 { "vm map (system)", &lock_class_mtx_sleep }, 590 { "vm page queue", &lock_class_mtx_sleep }, 591 { "vnode interlock", &lock_class_mtx_sleep }, 592 { "cdev", &lock_class_mtx_sleep }, 593 { NULL, NULL }, 594 /* 595 * VM 596 */ 597 { "vm map (user)", &lock_class_sx }, 598 { "vm object", &lock_class_mtx_sleep }, 599 { "vm page", &lock_class_mtx_sleep }, 600 { "vm page queue", &lock_class_mtx_sleep }, 601 { "pmap pv global", &lock_class_rw }, 602 { "pmap", &lock_class_mtx_sleep }, 603 { "pmap pv list", &lock_class_rw }, 604 { "vm page free queue", &lock_class_mtx_sleep }, 605 { NULL, NULL }, 606 /* 607 * kqueue/VFS interaction 608 */ 609 { "kqueue", &lock_class_mtx_sleep }, 610 { "struct mount mtx", &lock_class_mtx_sleep }, 611 { "vnode interlock", &lock_class_mtx_sleep }, 612 { NULL, NULL }, 613 /* 614 * ZFS locking 615 */ 616 { "dn->dn_mtx", &lock_class_sx }, 617 { "dr->dt.di.dr_mtx", &lock_class_sx }, 618 { "db->db_mtx", &lock_class_sx }, 619 { NULL, NULL }, 620 /* 621 * spin locks 622 */ 623#ifdef SMP 624 { "ap boot", &lock_class_mtx_spin }, 625#endif 626 { "rm.mutex_mtx", &lock_class_mtx_spin }, 627 { "sio", &lock_class_mtx_spin }, 628 { "scrlock", &lock_class_mtx_spin }, 629#ifdef __i386__ 630 { "cy", &lock_class_mtx_spin }, 631#endif 632#ifdef __sparc64__ 633 { "pcib_mtx", &lock_class_mtx_spin }, 634 { "rtc_mtx", &lock_class_mtx_spin }, 635#endif 636 { "scc_hwmtx", &lock_class_mtx_spin }, 637 { "uart_hwmtx", &lock_class_mtx_spin }, 638 { "fast_taskqueue", &lock_class_mtx_spin }, 639 { "intr table", &lock_class_mtx_spin }, 640#ifdef HWPMC_HOOKS 641 { "pmc-per-proc", &lock_class_mtx_spin }, 642#endif 643 { "process slock", &lock_class_mtx_spin }, 644 { "sleepq chain", &lock_class_mtx_spin }, 645 { "umtx lock", &lock_class_mtx_spin }, 646 { "rm_spinlock", &lock_class_mtx_spin }, 647 { "turnstile chain", &lock_class_mtx_spin }, 648 { "turnstile lock", &lock_class_mtx_spin }, 649 { "sched lock", &lock_class_mtx_spin }, 650 { "td_contested", &lock_class_mtx_spin }, 651 { "callout", &lock_class_mtx_spin }, 652 { "entropy harvest mutex", &lock_class_mtx_spin }, 653 { "syscons video lock", &lock_class_mtx_spin }, 654#ifdef SMP 655 { "smp rendezvous", &lock_class_mtx_spin }, 656#endif 657#ifdef __powerpc__ 658 { "tlb0", &lock_class_mtx_spin }, 659#endif 660 /* 661 * leaf locks 662 */ 663 { "intrcnt", &lock_class_mtx_spin }, 664 { "icu", &lock_class_mtx_spin }, 665#ifdef __i386__ 666 { "allpmaps", &lock_class_mtx_spin }, 667 { "descriptor tables", &lock_class_mtx_spin }, 668#endif 669 { "clk", &lock_class_mtx_spin }, 670 { "cpuset", &lock_class_mtx_spin }, 671 { "mprof lock", &lock_class_mtx_spin }, 672 { "zombie lock", &lock_class_mtx_spin }, 673 { "ALD Queue", &lock_class_mtx_spin }, 674#ifdef __ia64__ 675 { "MCA spin lock", &lock_class_mtx_spin }, 676#endif 677#if defined(__i386__) || defined(__amd64__) 678 { "pcicfg", &lock_class_mtx_spin }, 679 { "NDIS thread lock", &lock_class_mtx_spin }, 680#endif 681 { "tw_osl_io_lock", &lock_class_mtx_spin }, 682 { "tw_osl_q_lock", &lock_class_mtx_spin }, 683 { "tw_cl_io_lock", &lock_class_mtx_spin }, 684 { "tw_cl_intr_lock", &lock_class_mtx_spin }, 685 { "tw_cl_gen_lock", &lock_class_mtx_spin }, 686#ifdef HWPMC_HOOKS 687 { "pmc-leaf", &lock_class_mtx_spin }, 688#endif 689 { "blocked lock", &lock_class_mtx_spin }, 690 { NULL, NULL }, 691 { NULL, NULL } 692}; 693 694#ifdef BLESSING 695/* 696 * Pairs of locks which have been blessed 697 * Don't complain about order problems with blessed locks 698 */ 699static struct witness_blessed blessed_list[] = { 700}; 701static int blessed_count = 702 sizeof(blessed_list) / sizeof(struct witness_blessed); 703#endif 704 705/* 706 * This global is set to 0 once it becomes safe to use the witness code. 707 */ 708static int witness_cold = 1; 709 710/* 711 * This global is set to 1 once the static lock orders have been enrolled 712 * so that a warning can be issued for any spin locks enrolled later. 713 */ 714static int witness_spin_warn = 0; 715 716/* Trim useless garbage from filenames. */ 717static const char * 718fixup_filename(const char *file) 719{ 720 721 if (file == NULL) 722 return (NULL); 723 while (strncmp(file, "../", 3) == 0) 724 file += 3; 725 return (file); 726} 727 728/* 729 * The WITNESS-enabled diagnostic code. Note that the witness code does 730 * assume that the early boot is single-threaded at least until after this 731 * routine is completed. 732 */ 733static void 734witness_initialize(void *dummy __unused) 735{ 736 struct lock_object *lock; 737 struct witness_order_list_entry *order; 738 struct witness *w, *w1; 739 int i; 740 741 w_data = malloc(sizeof (struct witness) * WITNESS_COUNT, M_WITNESS, 742 M_NOWAIT | M_ZERO); 743 744 /* 745 * We have to release Giant before initializing its witness 746 * structure so that WITNESS doesn't get confused. 747 */ 748 mtx_unlock(&Giant); 749 mtx_assert(&Giant, MA_NOTOWNED); 750 751 CTR1(KTR_WITNESS, "%s: initializing witness", __func__); 752 mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET | 753 MTX_NOWITNESS | MTX_NOPROFILE); 754 for (i = WITNESS_COUNT - 1; i >= 0; i--) { 755 w = &w_data[i]; 756 memset(w, 0, sizeof(*w)); 757 w_data[i].w_index = i; /* Witness index never changes. */ 758 witness_free(w); 759 } 760 KASSERT(STAILQ_FIRST(&w_free)->w_index == 0, 761 ("%s: Invalid list of free witness objects", __func__)); 762 763 /* Witness with index 0 is not used to aid in debugging. */ 764 STAILQ_REMOVE_HEAD(&w_free, w_list); 765 w_free_cnt--; 766 767 memset(w_rmatrix, 0, 768 (sizeof(**w_rmatrix) * (WITNESS_COUNT+1) * (WITNESS_COUNT+1))); 769 770 for (i = 0; i < LOCK_CHILDCOUNT; i++) 771 witness_lock_list_free(&w_locklistdata[i]); 772 witness_init_hash_tables(); 773 774 /* First add in all the specified order lists. */ 775 for (order = order_lists; order->w_name != NULL; order++) { 776 w = enroll(order->w_name, order->w_class); 777 if (w == NULL) 778 continue; 779 w->w_file = "order list"; 780 for (order++; order->w_name != NULL; order++) { 781 w1 = enroll(order->w_name, order->w_class); 782 if (w1 == NULL) 783 continue; 784 w1->w_file = "order list"; 785 itismychild(w, w1); 786 w = w1; 787 } 788 } 789 witness_spin_warn = 1; 790 791 /* Iterate through all locks and add them to witness. */ 792 for (i = 0; pending_locks[i].wh_lock != NULL; i++) { 793 lock = pending_locks[i].wh_lock; 794 KASSERT(lock->lo_flags & LO_WITNESS, 795 ("%s: lock %s is on pending list but not LO_WITNESS", 796 __func__, lock->lo_name)); 797 lock->lo_witness = enroll(pending_locks[i].wh_type, 798 LOCK_CLASS(lock)); 799 } 800 801 /* Mark the witness code as being ready for use. */ 802 witness_cold = 0; 803 804 mtx_lock(&Giant); 805} 806SYSINIT(witness_init, SI_SUB_WITNESS, SI_ORDER_FIRST, witness_initialize, 807 NULL); 808 809void 810witness_init(struct lock_object *lock, const char *type) 811{ 812 struct lock_class *class; 813 814 /* Various sanity checks. */ 815 class = LOCK_CLASS(lock); 816 if ((lock->lo_flags & LO_RECURSABLE) != 0 && 817 (class->lc_flags & LC_RECURSABLE) == 0) 818 panic("%s: lock (%s) %s can not be recursable", __func__, 819 class->lc_name, lock->lo_name); 820 if ((lock->lo_flags & LO_SLEEPABLE) != 0 && 821 (class->lc_flags & LC_SLEEPABLE) == 0) 822 panic("%s: lock (%s) %s can not be sleepable", __func__, 823 class->lc_name, lock->lo_name); 824 if ((lock->lo_flags & LO_UPGRADABLE) != 0 && 825 (class->lc_flags & LC_UPGRADABLE) == 0) 826 panic("%s: lock (%s) %s can not be upgradable", __func__, 827 class->lc_name, lock->lo_name); 828 829 /* 830 * If we shouldn't watch this lock, then just clear lo_witness. 831 * Otherwise, if witness_cold is set, then it is too early to 832 * enroll this lock, so defer it to witness_initialize() by adding 833 * it to the pending_locks list. If it is not too early, then enroll 834 * the lock now. 835 */ 836 if (witness_watch < 1 || panicstr != NULL || 837 (lock->lo_flags & LO_WITNESS) == 0) 838 lock->lo_witness = NULL; 839 else if (witness_cold) { 840 pending_locks[pending_cnt].wh_lock = lock; 841 pending_locks[pending_cnt++].wh_type = type; 842 if (pending_cnt > WITNESS_PENDLIST) 843 panic("%s: pending locks list is too small, bump it\n", 844 __func__); 845 } else 846 lock->lo_witness = enroll(type, class); 847} 848 849void 850witness_destroy(struct lock_object *lock) 851{ 852 struct lock_class *class; 853 struct witness *w; 854 855 class = LOCK_CLASS(lock); 856 857 if (witness_cold) 858 panic("lock (%s) %s destroyed while witness_cold", 859 class->lc_name, lock->lo_name); 860 861 /* XXX: need to verify that no one holds the lock */ 862 if ((lock->lo_flags & LO_WITNESS) == 0 || lock->lo_witness == NULL) 863 return; 864 w = lock->lo_witness; 865 866 mtx_lock_spin(&w_mtx); 867 MPASS(w->w_refcount > 0); 868 w->w_refcount--; 869 870 if (w->w_refcount == 0) 871 depart(w); 872 mtx_unlock_spin(&w_mtx); 873} 874 875#ifdef DDB 876static void 877witness_ddb_compute_levels(void) 878{ 879 struct witness *w; 880 881 /* 882 * First clear all levels. 883 */ 884 STAILQ_FOREACH(w, &w_all, w_list) 885 w->w_ddb_level = -1; 886 887 /* 888 * Look for locks with no parents and level all their descendants. 889 */ 890 STAILQ_FOREACH(w, &w_all, w_list) { 891 892 /* If the witness has ancestors (is not a root), skip it. */ 893 if (w->w_num_ancestors > 0) 894 continue; 895 witness_ddb_level_descendants(w, 0); 896 } 897} 898 899static void 900witness_ddb_level_descendants(struct witness *w, int l) 901{ 902 int i; 903 904 if (w->w_ddb_level >= l) 905 return; 906 907 w->w_ddb_level = l; 908 l++; 909 910 for (i = 1; i <= w_max_used_index; i++) { 911 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT) 912 witness_ddb_level_descendants(&w_data[i], l); 913 } 914} 915 916static void 917witness_ddb_display_descendants(int(*prnt)(const char *fmt, ...), 918 struct witness *w, int indent) 919{ 920 int i; 921 922 for (i = 0; i < indent; i++) 923 prnt(" "); 924 prnt("%s (type: %s, depth: %d, active refs: %d)", 925 w->w_name, w->w_class->lc_name, 926 w->w_ddb_level, w->w_refcount); 927 if (w->w_displayed) { 928 prnt(" -- (already displayed)\n"); 929 return; 930 } 931 w->w_displayed = 1; 932 if (w->w_file != NULL && w->w_line != 0) 933 prnt(" -- last acquired @ %s:%d\n", fixup_filename(w->w_file), 934 w->w_line); 935 else 936 prnt(" -- never acquired\n"); 937 indent++; 938 WITNESS_INDEX_ASSERT(w->w_index); 939 for (i = 1; i <= w_max_used_index; i++) { 940 if (db_pager_quit) 941 return; 942 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT) 943 witness_ddb_display_descendants(prnt, &w_data[i], 944 indent); 945 } 946} 947 948static void 949witness_ddb_display_list(int(*prnt)(const char *fmt, ...), 950 struct witness_list *list) 951{ 952 struct witness *w; 953 954 STAILQ_FOREACH(w, list, w_typelist) { 955 if (w->w_file == NULL || w->w_ddb_level > 0) 956 continue; 957 958 /* This lock has no anscestors - display its descendants. */ 959 witness_ddb_display_descendants(prnt, w, 0); 960 if (db_pager_quit) 961 return; 962 } 963} 964 965static void 966witness_ddb_display(int(*prnt)(const char *fmt, ...)) 967{ 968 struct witness *w; 969 970 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__)); 971 witness_ddb_compute_levels(); 972 973 /* Clear all the displayed flags. */ 974 STAILQ_FOREACH(w, &w_all, w_list) 975 w->w_displayed = 0; 976 977 /* 978 * First, handle sleep locks which have been acquired at least 979 * once. 980 */ 981 prnt("Sleep locks:\n"); 982 witness_ddb_display_list(prnt, &w_sleep); 983 if (db_pager_quit) 984 return; 985 986 /* 987 * Now do spin locks which have been acquired at least once. 988 */ 989 prnt("\nSpin locks:\n"); 990 witness_ddb_display_list(prnt, &w_spin); 991 if (db_pager_quit) 992 return; 993 994 /* 995 * Finally, any locks which have not been acquired yet. 996 */ 997 prnt("\nLocks which were never acquired:\n"); 998 STAILQ_FOREACH(w, &w_all, w_list) { 999 if (w->w_file != NULL || w->w_refcount == 0) 1000 continue; 1001 prnt("%s (type: %s, depth: %d)\n", w->w_name, 1002 w->w_class->lc_name, w->w_ddb_level); 1003 if (db_pager_quit) 1004 return; 1005 } 1006} 1007#endif /* DDB */ 1008 1009int 1010witness_defineorder(struct lock_object *lock1, struct lock_object *lock2) 1011{ 1012 1013 if (witness_watch == -1 || panicstr != NULL) 1014 return (0); 1015 1016 /* Require locks that witness knows about. */ 1017 if (lock1 == NULL || lock1->lo_witness == NULL || lock2 == NULL || 1018 lock2->lo_witness == NULL) 1019 return (EINVAL); 1020 1021 mtx_assert(&w_mtx, MA_NOTOWNED); 1022 mtx_lock_spin(&w_mtx); 1023 1024 /* 1025 * If we already have either an explicit or implied lock order that 1026 * is the other way around, then return an error. 1027 */ 1028 if (witness_watch && 1029 isitmydescendant(lock2->lo_witness, lock1->lo_witness)) { 1030 mtx_unlock_spin(&w_mtx); 1031 return (EDOOFUS); 1032 } 1033 1034 /* Try to add the new order. */ 1035 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__, 1036 lock2->lo_witness->w_name, lock1->lo_witness->w_name); 1037 itismychild(lock1->lo_witness, lock2->lo_witness); 1038 mtx_unlock_spin(&w_mtx); 1039 return (0); 1040} 1041 1042void 1043witness_checkorder(struct lock_object *lock, int flags, const char *file, 1044 int line, struct lock_object *interlock) 1045{ 1046 struct lock_list_entry *lock_list, *lle; 1047 struct lock_instance *lock1, *lock2, *plock; 1048 struct lock_class *class; 1049 struct witness *w, *w1; 1050 struct thread *td; 1051 int i, j; 1052 1053 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL || 1054 panicstr != NULL) 1055 return; 1056 1057 w = lock->lo_witness; 1058 class = LOCK_CLASS(lock); 1059 td = curthread; 1060 1061 if (class->lc_flags & LC_SLEEPLOCK) { 1062 1063 /* 1064 * Since spin locks include a critical section, this check 1065 * implicitly enforces a lock order of all sleep locks before 1066 * all spin locks. 1067 */ 1068 if (td->td_critnest != 0 && !kdb_active) 1069 panic("blockable sleep lock (%s) %s @ %s:%d", 1070 class->lc_name, lock->lo_name, 1071 fixup_filename(file), line); 1072 1073 /* 1074 * If this is the first lock acquired then just return as 1075 * no order checking is needed. 1076 */ 1077 lock_list = td->td_sleeplocks; 1078 if (lock_list == NULL || lock_list->ll_count == 0) 1079 return; 1080 } else { 1081 1082 /* 1083 * If this is the first lock, just return as no order 1084 * checking is needed. Avoid problems with thread 1085 * migration pinning the thread while checking if 1086 * spinlocks are held. If at least one spinlock is held 1087 * the thread is in a safe path and it is allowed to 1088 * unpin it. 1089 */ 1090 sched_pin(); 1091 lock_list = PCPU_GET(spinlocks); 1092 if (lock_list == NULL || lock_list->ll_count == 0) { 1093 sched_unpin(); 1094 return; 1095 } 1096 sched_unpin(); 1097 } 1098 1099 /* 1100 * Check to see if we are recursing on a lock we already own. If 1101 * so, make sure that we don't mismatch exclusive and shared lock 1102 * acquires. 1103 */ 1104 lock1 = find_instance(lock_list, lock); 1105 if (lock1 != NULL) { 1106 if ((lock1->li_flags & LI_EXCLUSIVE) != 0 && 1107 (flags & LOP_EXCLUSIVE) == 0) { 1108 printf("shared lock of (%s) %s @ %s:%d\n", 1109 class->lc_name, lock->lo_name, 1110 fixup_filename(file), line); 1111 printf("while exclusively locked from %s:%d\n", 1112 fixup_filename(lock1->li_file), lock1->li_line); 1113 panic("share->excl"); 1114 } 1115 if ((lock1->li_flags & LI_EXCLUSIVE) == 0 && 1116 (flags & LOP_EXCLUSIVE) != 0) { 1117 printf("exclusive lock of (%s) %s @ %s:%d\n", 1118 class->lc_name, lock->lo_name, 1119 fixup_filename(file), line); 1120 printf("while share locked from %s:%d\n", 1121 fixup_filename(lock1->li_file), lock1->li_line); 1122 panic("excl->share"); 1123 } 1124 return; 1125 } 1126 1127 /* 1128 * Find the previously acquired lock, but ignore interlocks. 1129 */ 1130 plock = &lock_list->ll_children[lock_list->ll_count - 1]; 1131 if (interlock != NULL && plock->li_lock == interlock) { 1132 if (lock_list->ll_count > 1) 1133 plock = 1134 &lock_list->ll_children[lock_list->ll_count - 2]; 1135 else { 1136 lle = lock_list->ll_next; 1137 1138 /* 1139 * The interlock is the only lock we hold, so 1140 * simply return. 1141 */ 1142 if (lle == NULL) 1143 return; 1144 plock = &lle->ll_children[lle->ll_count - 1]; 1145 } 1146 } 1147 1148 /* 1149 * Try to perform most checks without a lock. If this succeeds we 1150 * can skip acquiring the lock and return success. 1151 */ 1152 w1 = plock->li_lock->lo_witness; 1153 if (witness_lock_order_check(w1, w)) 1154 return; 1155 1156 /* 1157 * Check for duplicate locks of the same type. Note that we only 1158 * have to check for this on the last lock we just acquired. Any 1159 * other cases will be caught as lock order violations. 1160 */ 1161 mtx_lock_spin(&w_mtx); 1162 witness_lock_order_add(w1, w); 1163 if (w1 == w) { 1164 i = w->w_index; 1165 if (!(lock->lo_flags & LO_DUPOK) && !(flags & LOP_DUPOK) && 1166 !(w_rmatrix[i][i] & WITNESS_REVERSAL)) { 1167 w_rmatrix[i][i] |= WITNESS_REVERSAL; 1168 w->w_reversed = 1; 1169 mtx_unlock_spin(&w_mtx); 1170 printf( 1171 "acquiring duplicate lock of same type: \"%s\"\n", 1172 w->w_name); 1173 printf(" 1st %s @ %s:%d\n", plock->li_lock->lo_name, 1174 fixup_filename(plock->li_file), plock->li_line); 1175 printf(" 2nd %s @ %s:%d\n", lock->lo_name, 1176 fixup_filename(file), line); 1177 witness_debugger(1); 1178 } else 1179 mtx_unlock_spin(&w_mtx); 1180 return; 1181 } 1182 mtx_assert(&w_mtx, MA_OWNED); 1183 1184 /* 1185 * If we know that the lock we are acquiring comes after 1186 * the lock we most recently acquired in the lock order tree, 1187 * then there is no need for any further checks. 1188 */ 1189 if (isitmychild(w1, w)) 1190 goto out; 1191 1192 for (j = 0, lle = lock_list; lle != NULL; lle = lle->ll_next) { 1193 for (i = lle->ll_count - 1; i >= 0; i--, j++) { 1194 1195 MPASS(j < WITNESS_COUNT); 1196 lock1 = &lle->ll_children[i]; 1197 1198 /* 1199 * Ignore the interlock the first time we see it. 1200 */ 1201 if (interlock != NULL && interlock == lock1->li_lock) { 1202 interlock = NULL; 1203 continue; 1204 } 1205 1206 /* 1207 * If this lock doesn't undergo witness checking, 1208 * then skip it. 1209 */ 1210 w1 = lock1->li_lock->lo_witness; 1211 if (w1 == NULL) { 1212 KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0, 1213 ("lock missing witness structure")); 1214 continue; 1215 } 1216 1217 /* 1218 * If we are locking Giant and this is a sleepable 1219 * lock, then skip it. 1220 */ 1221 if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0 && 1222 lock == &Giant.lock_object) 1223 continue; 1224 1225 /* 1226 * If we are locking a sleepable lock and this lock 1227 * is Giant, then skip it. 1228 */ 1229 if ((lock->lo_flags & LO_SLEEPABLE) != 0 && 1230 lock1->li_lock == &Giant.lock_object) 1231 continue; 1232 1233 /* 1234 * If we are locking a sleepable lock and this lock 1235 * isn't sleepable, we want to treat it as a lock 1236 * order violation to enfore a general lock order of 1237 * sleepable locks before non-sleepable locks. 1238 */ 1239 if (((lock->lo_flags & LO_SLEEPABLE) != 0 && 1240 (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0)) 1241 goto reversal; 1242 1243 /* 1244 * If we are locking Giant and this is a non-sleepable 1245 * lock, then treat it as a reversal. 1246 */ 1247 if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0 && 1248 lock == &Giant.lock_object) 1249 goto reversal; 1250 1251 /* 1252 * Check the lock order hierarchy for a reveresal. 1253 */ 1254 if (!isitmydescendant(w, w1)) 1255 continue; 1256 reversal: 1257 1258 /* 1259 * We have a lock order violation, check to see if it 1260 * is allowed or has already been yelled about. 1261 */ 1262#ifdef BLESSING 1263 1264 /* 1265 * If the lock order is blessed, just bail. We don't 1266 * look for other lock order violations though, which 1267 * may be a bug. 1268 */ 1269 if (blessed(w, w1)) 1270 goto out; 1271#endif 1272 1273 /* Bail if this violation is known */ 1274 if (w_rmatrix[w1->w_index][w->w_index] & WITNESS_REVERSAL) 1275 goto out; 1276 1277 /* Record this as a violation */ 1278 w_rmatrix[w1->w_index][w->w_index] |= WITNESS_REVERSAL; 1279 w_rmatrix[w->w_index][w1->w_index] |= WITNESS_REVERSAL; 1280 w->w_reversed = w1->w_reversed = 1; 1281 witness_increment_graph_generation(); 1282 mtx_unlock_spin(&w_mtx); 1283 1284 /* 1285 * Ok, yell about it. 1286 */ 1287 if (((lock->lo_flags & LO_SLEEPABLE) != 0 && 1288 (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0)) 1289 printf( 1290 "lock order reversal: (sleepable after non-sleepable)\n"); 1291 else if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0 1292 && lock == &Giant.lock_object) 1293 printf( 1294 "lock order reversal: (Giant after non-sleepable)\n"); 1295 else 1296 printf("lock order reversal:\n"); 1297 1298 /* 1299 * Try to locate an earlier lock with 1300 * witness w in our list. 1301 */ 1302 do { 1303 lock2 = &lle->ll_children[i]; 1304 MPASS(lock2->li_lock != NULL); 1305 if (lock2->li_lock->lo_witness == w) 1306 break; 1307 if (i == 0 && lle->ll_next != NULL) { 1308 lle = lle->ll_next; 1309 i = lle->ll_count - 1; 1310 MPASS(i >= 0 && i < LOCK_NCHILDREN); 1311 } else 1312 i--; 1313 } while (i >= 0); 1314 if (i < 0) { 1315 printf(" 1st %p %s (%s) @ %s:%d\n", 1316 lock1->li_lock, lock1->li_lock->lo_name, 1317 w1->w_name, fixup_filename(lock1->li_file), 1318 lock1->li_line); 1319 printf(" 2nd %p %s (%s) @ %s:%d\n", lock, 1320 lock->lo_name, w->w_name, 1321 fixup_filename(file), line); 1322 } else { 1323 printf(" 1st %p %s (%s) @ %s:%d\n", 1324 lock2->li_lock, lock2->li_lock->lo_name, 1325 lock2->li_lock->lo_witness->w_name, 1326 fixup_filename(lock2->li_file), 1327 lock2->li_line); 1328 printf(" 2nd %p %s (%s) @ %s:%d\n", 1329 lock1->li_lock, lock1->li_lock->lo_name, 1330 w1->w_name, fixup_filename(lock1->li_file), 1331 lock1->li_line); 1332 printf(" 3rd %p %s (%s) @ %s:%d\n", lock, 1333 lock->lo_name, w->w_name, 1334 fixup_filename(file), line); 1335 } 1336 witness_debugger(1); 1337 return; 1338 } 1339 } 1340 1341 /* 1342 * If requested, build a new lock order. However, don't build a new 1343 * relationship between a sleepable lock and Giant if it is in the 1344 * wrong direction. The correct lock order is that sleepable locks 1345 * always come before Giant. 1346 */ 1347 if (flags & LOP_NEWORDER && 1348 !(plock->li_lock == &Giant.lock_object && 1349 (lock->lo_flags & LO_SLEEPABLE) != 0)) { 1350 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__, 1351 w->w_name, plock->li_lock->lo_witness->w_name); 1352 itismychild(plock->li_lock->lo_witness, w); 1353 } 1354out: 1355 mtx_unlock_spin(&w_mtx); 1356} 1357 1358void 1359witness_lock(struct lock_object *lock, int flags, const char *file, int line) 1360{ 1361 struct lock_list_entry **lock_list, *lle; 1362 struct lock_instance *instance; 1363 struct witness *w; 1364 struct thread *td; 1365 1366 if (witness_cold || witness_watch == -1 || lock->lo_witness == NULL || 1367 panicstr != NULL) 1368 return; 1369 w = lock->lo_witness; 1370 td = curthread; 1371 1372 /* Determine lock list for this lock. */ 1373 if (LOCK_CLASS(lock)->lc_flags & LC_SLEEPLOCK) 1374 lock_list = &td->td_sleeplocks; 1375 else 1376 lock_list = PCPU_PTR(spinlocks); 1377 1378 /* Check to see if we are recursing on a lock we already own. */ 1379 instance = find_instance(*lock_list, lock); 1380 if (instance != NULL) { 1381 instance->li_flags++; 1382 CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__, 1383 td->td_proc->p_pid, lock->lo_name, 1384 instance->li_flags & LI_RECURSEMASK); 1385 instance->li_file = file; 1386 instance->li_line = line; 1387 return; 1388 } 1389 1390 /* Update per-witness last file and line acquire. */ 1391 w->w_file = file; 1392 w->w_line = line; 1393 1394 /* Find the next open lock instance in the list and fill it. */ 1395 lle = *lock_list; 1396 if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) { 1397 lle = witness_lock_list_get(); 1398 if (lle == NULL) 1399 return; 1400 lle->ll_next = *lock_list; 1401 CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__, 1402 td->td_proc->p_pid, lle); 1403 *lock_list = lle; 1404 } 1405 instance = &lle->ll_children[lle->ll_count++]; 1406 instance->li_lock = lock; 1407 instance->li_line = line; 1408 instance->li_file = file; 1409 if ((flags & LOP_EXCLUSIVE) != 0) 1410 instance->li_flags = LI_EXCLUSIVE; 1411 else 1412 instance->li_flags = 0; 1413 CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__, 1414 td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1); 1415} 1416 1417void 1418witness_upgrade(struct lock_object *lock, int flags, const char *file, int line) 1419{ 1420 struct lock_instance *instance; 1421 struct lock_class *class; 1422 1423 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__)); 1424 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL) 1425 return; 1426 class = LOCK_CLASS(lock); 1427 if (witness_watch) { 1428 if ((lock->lo_flags & LO_UPGRADABLE) == 0) 1429 panic("upgrade of non-upgradable lock (%s) %s @ %s:%d", 1430 class->lc_name, lock->lo_name, 1431 fixup_filename(file), line); 1432 if ((class->lc_flags & LC_SLEEPLOCK) == 0) 1433 panic("upgrade of non-sleep lock (%s) %s @ %s:%d", 1434 class->lc_name, lock->lo_name, 1435 fixup_filename(file), line); 1436 } 1437 instance = find_instance(curthread->td_sleeplocks, lock); 1438 if (instance == NULL) 1439 panic("upgrade of unlocked lock (%s) %s @ %s:%d", 1440 class->lc_name, lock->lo_name, 1441 fixup_filename(file), line); 1442 if (witness_watch) { 1443 if ((instance->li_flags & LI_EXCLUSIVE) != 0) 1444 panic("upgrade of exclusive lock (%s) %s @ %s:%d", 1445 class->lc_name, lock->lo_name, 1446 fixup_filename(file), line); 1447 if ((instance->li_flags & LI_RECURSEMASK) != 0) 1448 panic("upgrade of recursed lock (%s) %s r=%d @ %s:%d", 1449 class->lc_name, lock->lo_name, 1450 instance->li_flags & LI_RECURSEMASK, 1451 fixup_filename(file), line); 1452 } 1453 instance->li_flags |= LI_EXCLUSIVE; 1454} 1455 1456void 1457witness_downgrade(struct lock_object *lock, int flags, const char *file, 1458 int line) 1459{ 1460 struct lock_instance *instance; 1461 struct lock_class *class; 1462 1463 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__)); 1464 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL) 1465 return; 1466 class = LOCK_CLASS(lock); 1467 if (witness_watch) { 1468 if ((lock->lo_flags & LO_UPGRADABLE) == 0) 1469 panic("downgrade of non-upgradable lock (%s) %s @ %s:%d", 1470 class->lc_name, lock->lo_name, 1471 fixup_filename(file), line); 1472 if ((class->lc_flags & LC_SLEEPLOCK) == 0) 1473 panic("downgrade of non-sleep lock (%s) %s @ %s:%d", 1474 class->lc_name, lock->lo_name, 1475 fixup_filename(file), line); 1476 } 1477 instance = find_instance(curthread->td_sleeplocks, lock); 1478 if (instance == NULL) 1479 panic("downgrade of unlocked lock (%s) %s @ %s:%d", 1480 class->lc_name, lock->lo_name, 1481 fixup_filename(file), line); 1482 if (witness_watch) { 1483 if ((instance->li_flags & LI_EXCLUSIVE) == 0) 1484 panic("downgrade of shared lock (%s) %s @ %s:%d", 1485 class->lc_name, lock->lo_name, 1486 fixup_filename(file), line); 1487 if ((instance->li_flags & LI_RECURSEMASK) != 0) 1488 panic("downgrade of recursed lock (%s) %s r=%d @ %s:%d", 1489 class->lc_name, lock->lo_name, 1490 instance->li_flags & LI_RECURSEMASK, 1491 fixup_filename(file), line); 1492 } 1493 instance->li_flags &= ~LI_EXCLUSIVE; 1494} 1495 1496void 1497witness_unlock(struct lock_object *lock, int flags, const char *file, int line) 1498{ 1499 struct lock_list_entry **lock_list, *lle; 1500 struct lock_instance *instance; 1501 struct lock_class *class; 1502 struct thread *td; 1503 register_t s; 1504 int i, j; 1505 1506 if (witness_cold || lock->lo_witness == NULL || panicstr != NULL) 1507 return; 1508 td = curthread; 1509 class = LOCK_CLASS(lock); 1510 1511 /* Find lock instance associated with this lock. */ 1512 if (class->lc_flags & LC_SLEEPLOCK) 1513 lock_list = &td->td_sleeplocks; 1514 else 1515 lock_list = PCPU_PTR(spinlocks); 1516 lle = *lock_list; 1517 for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next) 1518 for (i = 0; i < (*lock_list)->ll_count; i++) { 1519 instance = &(*lock_list)->ll_children[i]; 1520 if (instance->li_lock == lock) 1521 goto found; 1522 } 1523 1524 /* 1525 * When disabling WITNESS through witness_watch we could end up in 1526 * having registered locks in the td_sleeplocks queue. 1527 * We have to make sure we flush these queues, so just search for 1528 * eventual register locks and remove them. 1529 */ 1530 if (witness_watch > 0) 1531 panic("lock (%s) %s not locked @ %s:%d", class->lc_name, 1532 lock->lo_name, fixup_filename(file), line); 1533 else 1534 return; 1535found: 1536 1537 /* First, check for shared/exclusive mismatches. */ 1538 if ((instance->li_flags & LI_EXCLUSIVE) != 0 && witness_watch > 0 && 1539 (flags & LOP_EXCLUSIVE) == 0) { 1540 printf("shared unlock of (%s) %s @ %s:%d\n", class->lc_name, 1541 lock->lo_name, fixup_filename(file), line); 1542 printf("while exclusively locked from %s:%d\n", 1543 fixup_filename(instance->li_file), instance->li_line); 1544 panic("excl->ushare"); 1545 } 1546 if ((instance->li_flags & LI_EXCLUSIVE) == 0 && witness_watch > 0 && 1547 (flags & LOP_EXCLUSIVE) != 0) { 1548 printf("exclusive unlock of (%s) %s @ %s:%d\n", class->lc_name, 1549 lock->lo_name, fixup_filename(file), line); 1550 printf("while share locked from %s:%d\n", 1551 fixup_filename(instance->li_file), 1552 instance->li_line); 1553 panic("share->uexcl"); 1554 } 1555 /* If we are recursed, unrecurse. */ 1556 if ((instance->li_flags & LI_RECURSEMASK) > 0) { 1557 CTR4(KTR_WITNESS, "%s: pid %d unrecursed on %s r=%d", __func__, 1558 td->td_proc->p_pid, instance->li_lock->lo_name, 1559 instance->li_flags); 1560 instance->li_flags--; 1561 return; 1562 } 1563 /* The lock is now being dropped, check for NORELEASE flag */ 1564 if ((instance->li_flags & LI_NORELEASE) != 0 && witness_watch > 0) { 1565 printf("forbidden unlock of (%s) %s @ %s:%d\n", class->lc_name, 1566 lock->lo_name, fixup_filename(file), line); 1567 panic("lock marked norelease"); 1568 } 1569 1570 /* Otherwise, remove this item from the list. */ 1571 s = intr_disable(); 1572 CTR4(KTR_WITNESS, "%s: pid %d removed %s from lle[%d]", __func__, 1573 td->td_proc->p_pid, instance->li_lock->lo_name, 1574 (*lock_list)->ll_count - 1); 1575 for (j = i; j < (*lock_list)->ll_count - 1; j++) 1576 (*lock_list)->ll_children[j] = 1577 (*lock_list)->ll_children[j + 1]; 1578 (*lock_list)->ll_count--; 1579 intr_restore(s); 1580 1581 /* 1582 * In order to reduce contention on w_mtx, we want to keep always an 1583 * head object into lists so that frequent allocation from the 1584 * free witness pool (and subsequent locking) is avoided. 1585 * In order to maintain the current code simple, when the head 1586 * object is totally unloaded it means also that we do not have 1587 * further objects in the list, so the list ownership needs to be 1588 * hand over to another object if the current head needs to be freed. 1589 */ 1590 if ((*lock_list)->ll_count == 0) { 1591 if (*lock_list == lle) { 1592 if (lle->ll_next == NULL) 1593 return; 1594 } else 1595 lle = *lock_list; 1596 *lock_list = lle->ll_next; 1597 CTR3(KTR_WITNESS, "%s: pid %d removed lle %p", __func__, 1598 td->td_proc->p_pid, lle); 1599 witness_lock_list_free(lle); 1600 } 1601} 1602 1603void 1604witness_thread_exit(struct thread *td) 1605{ 1606 struct lock_list_entry *lle; 1607 int i, n; 1608 1609 lle = td->td_sleeplocks; 1610 if (lle == NULL || panicstr != NULL) 1611 return; 1612 if (lle->ll_count != 0) { 1613 for (n = 0; lle != NULL; lle = lle->ll_next) 1614 for (i = lle->ll_count - 1; i >= 0; i--) { 1615 if (n == 0) 1616 printf("Thread %p exiting with the following locks held:\n", 1617 td); 1618 n++; 1619 witness_list_lock(&lle->ll_children[i], printf); 1620 1621 } 1622 panic("Thread %p cannot exit while holding sleeplocks\n", td); 1623 } 1624 witness_lock_list_free(lle); 1625} 1626 1627/* 1628 * Warn if any locks other than 'lock' are held. Flags can be passed in to 1629 * exempt Giant and sleepable locks from the checks as well. If any 1630 * non-exempt locks are held, then a supplied message is printed to the 1631 * console along with a list of the offending locks. If indicated in the 1632 * flags then a failure results in a panic as well. 1633 */ 1634int 1635witness_warn(int flags, struct lock_object *lock, const char *fmt, ...) 1636{ 1637 struct lock_list_entry *lock_list, *lle; 1638 struct lock_instance *lock1; 1639 struct thread *td; 1640 va_list ap; 1641 int i, n; 1642 1643 if (witness_cold || witness_watch < 1 || panicstr != NULL) 1644 return (0); 1645 n = 0; 1646 td = curthread; 1647 for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next) 1648 for (i = lle->ll_count - 1; i >= 0; i--) { 1649 lock1 = &lle->ll_children[i]; 1650 if (lock1->li_lock == lock) 1651 continue; 1652 if (flags & WARN_GIANTOK && 1653 lock1->li_lock == &Giant.lock_object) 1654 continue; 1655 if (flags & WARN_SLEEPOK && 1656 (lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0) 1657 continue; 1658 if (n == 0) { 1659 va_start(ap, fmt); 1660 vprintf(fmt, ap); 1661 va_end(ap); 1662 printf(" with the following"); 1663 if (flags & WARN_SLEEPOK) 1664 printf(" non-sleepable"); 1665 printf(" locks held:\n"); 1666 } 1667 n++; 1668 witness_list_lock(lock1, printf); 1669 } 1670 1671 /* 1672 * Pin the thread in order to avoid problems with thread migration. 1673 * Once that all verifies are passed about spinlocks ownership, 1674 * the thread is in a safe path and it can be unpinned. 1675 */ 1676 sched_pin(); 1677 lock_list = PCPU_GET(spinlocks); 1678 if (lock_list != NULL && lock_list->ll_count != 0) { 1679 sched_unpin(); 1680 1681 /* 1682 * We should only have one spinlock and as long as 1683 * the flags cannot match for this locks class, 1684 * check if the first spinlock is the one curthread 1685 * should hold. 1686 */ 1687 lock1 = &lock_list->ll_children[lock_list->ll_count - 1]; 1688 if (lock_list->ll_count == 1 && lock_list->ll_next == NULL && 1689 lock1->li_lock == lock && n == 0) 1690 return (0); 1691 1692 va_start(ap, fmt); 1693 vprintf(fmt, ap); 1694 va_end(ap); 1695 printf(" with the following"); 1696 if (flags & WARN_SLEEPOK) 1697 printf(" non-sleepable"); 1698 printf(" locks held:\n"); 1699 n += witness_list_locks(&lock_list, printf); 1700 } else 1701 sched_unpin(); 1702 if (flags & WARN_PANIC && n) 1703 panic("%s", __func__); 1704 else 1705 witness_debugger(n); 1706 return (n); 1707} 1708 1709const char * 1710witness_file(struct lock_object *lock) 1711{ 1712 struct witness *w; 1713 1714 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL) 1715 return ("?"); 1716 w = lock->lo_witness; 1717 return (w->w_file); 1718} 1719 1720int 1721witness_line(struct lock_object *lock) 1722{ 1723 struct witness *w; 1724 1725 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL) 1726 return (0); 1727 w = lock->lo_witness; 1728 return (w->w_line); 1729} 1730 1731static struct witness * 1732enroll(const char *description, struct lock_class *lock_class) 1733{ 1734 struct witness *w; 1735 struct witness_list *typelist; 1736 1737 MPASS(description != NULL); 1738 1739 if (witness_watch == -1 || panicstr != NULL) 1740 return (NULL); 1741 if ((lock_class->lc_flags & LC_SPINLOCK)) { 1742 if (witness_skipspin) 1743 return (NULL); 1744 else 1745 typelist = &w_spin; 1746 } else if ((lock_class->lc_flags & LC_SLEEPLOCK)) 1747 typelist = &w_sleep; 1748 else 1749 panic("lock class %s is not sleep or spin", 1750 lock_class->lc_name); 1751 1752 mtx_lock_spin(&w_mtx); 1753 w = witness_hash_get(description); 1754 if (w) 1755 goto found; 1756 if ((w = witness_get()) == NULL) 1757 return (NULL); 1758 MPASS(strlen(description) < MAX_W_NAME); 1759 strcpy(w->w_name, description); 1760 w->w_class = lock_class; 1761 w->w_refcount = 1; 1762 STAILQ_INSERT_HEAD(&w_all, w, w_list); 1763 if (lock_class->lc_flags & LC_SPINLOCK) { 1764 STAILQ_INSERT_HEAD(&w_spin, w, w_typelist); 1765 w_spin_cnt++; 1766 } else if (lock_class->lc_flags & LC_SLEEPLOCK) { 1767 STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist); 1768 w_sleep_cnt++; 1769 } 1770 1771 /* Insert new witness into the hash */ 1772 witness_hash_put(w); 1773 witness_increment_graph_generation(); 1774 mtx_unlock_spin(&w_mtx); 1775 return (w); 1776found: 1777 w->w_refcount++; 1778 mtx_unlock_spin(&w_mtx); 1779 if (lock_class != w->w_class) 1780 panic( 1781 "lock (%s) %s does not match earlier (%s) lock", 1782 description, lock_class->lc_name, 1783 w->w_class->lc_name); 1784 return (w); 1785} 1786 1787static void 1788depart(struct witness *w) 1789{ 1790 struct witness_list *list; 1791 1792 MPASS(w->w_refcount == 0); 1793 if (w->w_class->lc_flags & LC_SLEEPLOCK) { 1794 list = &w_sleep; 1795 w_sleep_cnt--; 1796 } else { 1797 list = &w_spin; 1798 w_spin_cnt--; 1799 } 1800 /* 1801 * Set file to NULL as it may point into a loadable module. 1802 */ 1803 w->w_file = NULL; 1804 w->w_line = 0; 1805 witness_increment_graph_generation(); 1806} 1807 1808 1809static void 1810adopt(struct witness *parent, struct witness *child) 1811{ 1812 int pi, ci, i, j; 1813 1814 if (witness_cold == 0) 1815 mtx_assert(&w_mtx, MA_OWNED); 1816 1817 /* If the relationship is already known, there's no work to be done. */ 1818 if (isitmychild(parent, child)) 1819 return; 1820 1821 /* When the structure of the graph changes, bump up the generation. */ 1822 witness_increment_graph_generation(); 1823 1824 /* 1825 * The hard part ... create the direct relationship, then propagate all 1826 * indirect relationships. 1827 */ 1828 pi = parent->w_index; 1829 ci = child->w_index; 1830 WITNESS_INDEX_ASSERT(pi); 1831 WITNESS_INDEX_ASSERT(ci); 1832 MPASS(pi != ci); 1833 w_rmatrix[pi][ci] |= WITNESS_PARENT; 1834 w_rmatrix[ci][pi] |= WITNESS_CHILD; 1835 1836 /* 1837 * If parent was not already an ancestor of child, 1838 * then we increment the descendant and ancestor counters. 1839 */ 1840 if ((w_rmatrix[pi][ci] & WITNESS_ANCESTOR) == 0) { 1841 parent->w_num_descendants++; 1842 child->w_num_ancestors++; 1843 } 1844 1845 /* 1846 * Find each ancestor of 'pi'. Note that 'pi' itself is counted as 1847 * an ancestor of 'pi' during this loop. 1848 */ 1849 for (i = 1; i <= w_max_used_index; i++) { 1850 if ((w_rmatrix[i][pi] & WITNESS_ANCESTOR_MASK) == 0 && 1851 (i != pi)) 1852 continue; 1853 1854 /* Find each descendant of 'i' and mark it as a descendant. */ 1855 for (j = 1; j <= w_max_used_index; j++) { 1856 1857 /* 1858 * Skip children that are already marked as 1859 * descendants of 'i'. 1860 */ 1861 if (w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK) 1862 continue; 1863 1864 /* 1865 * We are only interested in descendants of 'ci'. Note 1866 * that 'ci' itself is counted as a descendant of 'ci'. 1867 */ 1868 if ((w_rmatrix[ci][j] & WITNESS_ANCESTOR_MASK) == 0 && 1869 (j != ci)) 1870 continue; 1871 w_rmatrix[i][j] |= WITNESS_ANCESTOR; 1872 w_rmatrix[j][i] |= WITNESS_DESCENDANT; 1873 w_data[i].w_num_descendants++; 1874 w_data[j].w_num_ancestors++; 1875 1876 /* 1877 * Make sure we aren't marking a node as both an 1878 * ancestor and descendant. We should have caught 1879 * this as a lock order reversal earlier. 1880 */ 1881 if ((w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK) && 1882 (w_rmatrix[i][j] & WITNESS_DESCENDANT_MASK)) { 1883 printf("witness rmatrix paradox! [%d][%d]=%d " 1884 "both ancestor and descendant\n", 1885 i, j, w_rmatrix[i][j]); 1886 kdb_backtrace(); 1887 printf("Witness disabled.\n"); 1888 witness_watch = -1; 1889 } 1890 if ((w_rmatrix[j][i] & WITNESS_ANCESTOR_MASK) && 1891 (w_rmatrix[j][i] & WITNESS_DESCENDANT_MASK)) { 1892 printf("witness rmatrix paradox! [%d][%d]=%d " 1893 "both ancestor and descendant\n", 1894 j, i, w_rmatrix[j][i]); 1895 kdb_backtrace(); 1896 printf("Witness disabled.\n"); 1897 witness_watch = -1; 1898 } 1899 } 1900 } 1901} 1902 1903static void 1904itismychild(struct witness *parent, struct witness *child) 1905{ 1906 1907 MPASS(child != NULL && parent != NULL); 1908 if (witness_cold == 0) 1909 mtx_assert(&w_mtx, MA_OWNED); 1910 1911 if (!witness_lock_type_equal(parent, child)) { 1912 if (witness_cold == 0) 1913 mtx_unlock_spin(&w_mtx); 1914 panic("%s: parent \"%s\" (%s) and child \"%s\" (%s) are not " 1915 "the same lock type", __func__, parent->w_name, 1916 parent->w_class->lc_name, child->w_name, 1917 child->w_class->lc_name); 1918 } 1919 adopt(parent, child); 1920} 1921 1922/* 1923 * Generic code for the isitmy*() functions. The rmask parameter is the 1924 * expected relationship of w1 to w2. 1925 */ 1926static int 1927_isitmyx(struct witness *w1, struct witness *w2, int rmask, const char *fname) 1928{ 1929 unsigned char r1, r2; 1930 int i1, i2; 1931 1932 i1 = w1->w_index; 1933 i2 = w2->w_index; 1934 WITNESS_INDEX_ASSERT(i1); 1935 WITNESS_INDEX_ASSERT(i2); 1936 r1 = w_rmatrix[i1][i2] & WITNESS_RELATED_MASK; 1937 r2 = w_rmatrix[i2][i1] & WITNESS_RELATED_MASK; 1938 1939 /* The flags on one better be the inverse of the flags on the other */ 1940 if (!((WITNESS_ATOD(r1) == r2 && WITNESS_DTOA(r2) == r1) || 1941 (WITNESS_DTOA(r1) == r2 && WITNESS_ATOD(r2) == r1))) { 1942 printf("%s: rmatrix mismatch between %s (index %d) and %s " 1943 "(index %d): w_rmatrix[%d][%d] == %hhx but " 1944 "w_rmatrix[%d][%d] == %hhx\n", 1945 fname, w1->w_name, i1, w2->w_name, i2, i1, i2, r1, 1946 i2, i1, r2); 1947 kdb_backtrace(); 1948 printf("Witness disabled.\n"); 1949 witness_watch = -1; 1950 } 1951 return (r1 & rmask); 1952} 1953 1954/* 1955 * Checks if @child is a direct child of @parent. 1956 */ 1957static int 1958isitmychild(struct witness *parent, struct witness *child) 1959{ 1960 1961 return (_isitmyx(parent, child, WITNESS_PARENT, __func__)); 1962} 1963 1964/* 1965 * Checks if @descendant is a direct or inderect descendant of @ancestor. 1966 */ 1967static int 1968isitmydescendant(struct witness *ancestor, struct witness *descendant) 1969{ 1970 1971 return (_isitmyx(ancestor, descendant, WITNESS_ANCESTOR_MASK, 1972 __func__)); 1973} 1974 1975#ifdef BLESSING 1976static int 1977blessed(struct witness *w1, struct witness *w2) 1978{ 1979 int i; 1980 struct witness_blessed *b; 1981 1982 for (i = 0; i < blessed_count; i++) { 1983 b = &blessed_list[i]; 1984 if (strcmp(w1->w_name, b->b_lock1) == 0) { 1985 if (strcmp(w2->w_name, b->b_lock2) == 0) 1986 return (1); 1987 continue; 1988 } 1989 if (strcmp(w1->w_name, b->b_lock2) == 0) 1990 if (strcmp(w2->w_name, b->b_lock1) == 0) 1991 return (1); 1992 } 1993 return (0); 1994} 1995#endif 1996 1997static struct witness * 1998witness_get(void) 1999{ 2000 struct witness *w; 2001 int index; 2002 2003 if (witness_cold == 0) 2004 mtx_assert(&w_mtx, MA_OWNED); 2005 2006 if (witness_watch == -1) { 2007 mtx_unlock_spin(&w_mtx); 2008 return (NULL); 2009 } 2010 if (STAILQ_EMPTY(&w_free)) { 2011 witness_watch = -1; 2012 mtx_unlock_spin(&w_mtx); 2013 printf("WITNESS: unable to allocate a new witness object\n"); 2014 return (NULL); 2015 } 2016 w = STAILQ_FIRST(&w_free); 2017 STAILQ_REMOVE_HEAD(&w_free, w_list); 2018 w_free_cnt--; 2019 index = w->w_index; 2020 MPASS(index > 0 && index == w_max_used_index+1 && 2021 index < WITNESS_COUNT); 2022 bzero(w, sizeof(*w)); 2023 w->w_index = index; 2024 if (index > w_max_used_index) 2025 w_max_used_index = index; 2026 return (w); 2027} 2028 2029static void 2030witness_free(struct witness *w) 2031{ 2032 2033 STAILQ_INSERT_HEAD(&w_free, w, w_list); 2034 w_free_cnt++; 2035} 2036 2037static struct lock_list_entry * 2038witness_lock_list_get(void) 2039{ 2040 struct lock_list_entry *lle; 2041 2042 if (witness_watch == -1) 2043 return (NULL); 2044 mtx_lock_spin(&w_mtx); 2045 lle = w_lock_list_free; 2046 if (lle == NULL) { 2047 witness_watch = -1; 2048 mtx_unlock_spin(&w_mtx); 2049 printf("%s: witness exhausted\n", __func__); 2050 return (NULL); 2051 } 2052 w_lock_list_free = lle->ll_next; 2053 mtx_unlock_spin(&w_mtx); 2054 bzero(lle, sizeof(*lle)); 2055 return (lle); 2056} 2057 2058static void 2059witness_lock_list_free(struct lock_list_entry *lle) 2060{ 2061 2062 mtx_lock_spin(&w_mtx); 2063 lle->ll_next = w_lock_list_free; 2064 w_lock_list_free = lle; 2065 mtx_unlock_spin(&w_mtx); 2066} 2067 2068static struct lock_instance * 2069find_instance(struct lock_list_entry *list, struct lock_object *lock) 2070{ 2071 struct lock_list_entry *lle; 2072 struct lock_instance *instance; 2073 int i; 2074 2075 for (lle = list; lle != NULL; lle = lle->ll_next) 2076 for (i = lle->ll_count - 1; i >= 0; i--) { 2077 instance = &lle->ll_children[i]; 2078 if (instance->li_lock == lock) 2079 return (instance); 2080 } 2081 return (NULL); 2082} 2083 2084static void 2085witness_list_lock(struct lock_instance *instance, 2086 int (*prnt)(const char *fmt, ...)) 2087{ 2088 struct lock_object *lock; 2089 2090 lock = instance->li_lock; 2091 prnt("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ? 2092 "exclusive" : "shared", LOCK_CLASS(lock)->lc_name, lock->lo_name); 2093 if (lock->lo_witness->w_name != lock->lo_name) 2094 prnt(" (%s)", lock->lo_witness->w_name); 2095 prnt(" r = %d (%p) locked @ %s:%d\n", 2096 instance->li_flags & LI_RECURSEMASK, lock, 2097 fixup_filename(instance->li_file), instance->li_line); 2098} 2099 2100#ifdef DDB 2101static int 2102witness_thread_has_locks(struct thread *td) 2103{ 2104 2105 if (td->td_sleeplocks == NULL) 2106 return (0); 2107 return (td->td_sleeplocks->ll_count != 0); 2108} 2109 2110static int 2111witness_proc_has_locks(struct proc *p) 2112{ 2113 struct thread *td; 2114 2115 FOREACH_THREAD_IN_PROC(p, td) { 2116 if (witness_thread_has_locks(td)) 2117 return (1); 2118 } 2119 return (0); 2120} 2121#endif 2122 2123int 2124witness_list_locks(struct lock_list_entry **lock_list, 2125 int (*prnt)(const char *fmt, ...)) 2126{ 2127 struct lock_list_entry *lle; 2128 int i, nheld; 2129 2130 nheld = 0; 2131 for (lle = *lock_list; lle != NULL; lle = lle->ll_next) 2132 for (i = lle->ll_count - 1; i >= 0; i--) { 2133 witness_list_lock(&lle->ll_children[i], prnt); 2134 nheld++; 2135 } 2136 return (nheld); 2137} 2138 2139/* 2140 * This is a bit risky at best. We call this function when we have timed 2141 * out acquiring a spin lock, and we assume that the other CPU is stuck 2142 * with this lock held. So, we go groveling around in the other CPU's 2143 * per-cpu data to try to find the lock instance for this spin lock to 2144 * see when it was last acquired. 2145 */ 2146void 2147witness_display_spinlock(struct lock_object *lock, struct thread *owner, 2148 int (*prnt)(const char *fmt, ...)) 2149{ 2150 struct lock_instance *instance; 2151 struct pcpu *pc; 2152 2153 if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU) 2154 return; 2155 pc = pcpu_find(owner->td_oncpu); 2156 instance = find_instance(pc->pc_spinlocks, lock); 2157 if (instance != NULL) 2158 witness_list_lock(instance, prnt); 2159} 2160 2161void 2162witness_save(struct lock_object *lock, const char **filep, int *linep) 2163{ 2164 struct lock_list_entry *lock_list; 2165 struct lock_instance *instance; 2166 struct lock_class *class; 2167 2168 /* 2169 * This function is used independently in locking code to deal with 2170 * Giant, SCHEDULER_STOPPED() check can be removed here after Giant 2171 * is gone. 2172 */ 2173 if (SCHEDULER_STOPPED()) 2174 return; 2175 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__)); 2176 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL) 2177 return; 2178 class = LOCK_CLASS(lock); 2179 if (class->lc_flags & LC_SLEEPLOCK) 2180 lock_list = curthread->td_sleeplocks; 2181 else { 2182 if (witness_skipspin) 2183 return; 2184 lock_list = PCPU_GET(spinlocks); 2185 } 2186 instance = find_instance(lock_list, lock); 2187 if (instance == NULL) 2188 panic("%s: lock (%s) %s not locked", __func__, 2189 class->lc_name, lock->lo_name); 2190 *filep = instance->li_file; 2191 *linep = instance->li_line; 2192} 2193 2194void 2195witness_restore(struct lock_object *lock, const char *file, int line) 2196{ 2197 struct lock_list_entry *lock_list; 2198 struct lock_instance *instance; 2199 struct lock_class *class; 2200 2201 /* 2202 * This function is used independently in locking code to deal with 2203 * Giant, SCHEDULER_STOPPED() check can be removed here after Giant 2204 * is gone. 2205 */ 2206 if (SCHEDULER_STOPPED()) 2207 return; 2208 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__)); 2209 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL) 2210 return; 2211 class = LOCK_CLASS(lock); 2212 if (class->lc_flags & LC_SLEEPLOCK) 2213 lock_list = curthread->td_sleeplocks; 2214 else { 2215 if (witness_skipspin) 2216 return; 2217 lock_list = PCPU_GET(spinlocks); 2218 } 2219 instance = find_instance(lock_list, lock); 2220 if (instance == NULL) 2221 panic("%s: lock (%s) %s not locked", __func__, 2222 class->lc_name, lock->lo_name); 2223 lock->lo_witness->w_file = file; 2224 lock->lo_witness->w_line = line; 2225 instance->li_file = file; 2226 instance->li_line = line; 2227} 2228 2229void 2230witness_assert(struct lock_object *lock, int flags, const char *file, int line) 2231{ 2232#ifdef INVARIANT_SUPPORT 2233 struct lock_instance *instance; 2234 struct lock_class *class; 2235 2236 if (lock->lo_witness == NULL || witness_watch < 1 || panicstr != NULL) 2237 return; 2238 class = LOCK_CLASS(lock); 2239 if ((class->lc_flags & LC_SLEEPLOCK) != 0) 2240 instance = find_instance(curthread->td_sleeplocks, lock); 2241 else if ((class->lc_flags & LC_SPINLOCK) != 0) 2242 instance = find_instance(PCPU_GET(spinlocks), lock); 2243 else { 2244 panic("Lock (%s) %s is not sleep or spin!", 2245 class->lc_name, lock->lo_name); 2246 } 2247 switch (flags) { 2248 case LA_UNLOCKED: 2249 if (instance != NULL) 2250 panic("Lock (%s) %s locked @ %s:%d.", 2251 class->lc_name, lock->lo_name, 2252 fixup_filename(file), line); 2253 break; 2254 case LA_LOCKED: 2255 case LA_LOCKED | LA_RECURSED: 2256 case LA_LOCKED | LA_NOTRECURSED: 2257 case LA_SLOCKED: 2258 case LA_SLOCKED | LA_RECURSED: 2259 case LA_SLOCKED | LA_NOTRECURSED: 2260 case LA_XLOCKED: 2261 case LA_XLOCKED | LA_RECURSED: 2262 case LA_XLOCKED | LA_NOTRECURSED: 2263 if (instance == NULL) { 2264 panic("Lock (%s) %s not locked @ %s:%d.", 2265 class->lc_name, lock->lo_name, 2266 fixup_filename(file), line); 2267 break; 2268 } 2269 if ((flags & LA_XLOCKED) != 0 && 2270 (instance->li_flags & LI_EXCLUSIVE) == 0) 2271 panic("Lock (%s) %s not exclusively locked @ %s:%d.", 2272 class->lc_name, lock->lo_name, 2273 fixup_filename(file), line); 2274 if ((flags & LA_SLOCKED) != 0 && 2275 (instance->li_flags & LI_EXCLUSIVE) != 0) 2276 panic("Lock (%s) %s exclusively locked @ %s:%d.", 2277 class->lc_name, lock->lo_name, 2278 fixup_filename(file), line); 2279 if ((flags & LA_RECURSED) != 0 && 2280 (instance->li_flags & LI_RECURSEMASK) == 0) 2281 panic("Lock (%s) %s not recursed @ %s:%d.", 2282 class->lc_name, lock->lo_name, 2283 fixup_filename(file), line); 2284 if ((flags & LA_NOTRECURSED) != 0 && 2285 (instance->li_flags & LI_RECURSEMASK) != 0) 2286 panic("Lock (%s) %s recursed @ %s:%d.", 2287 class->lc_name, lock->lo_name, 2288 fixup_filename(file), line); 2289 break; 2290 default: 2291 panic("Invalid lock assertion at %s:%d.", 2292 fixup_filename(file), line); 2293 2294 } 2295#endif /* INVARIANT_SUPPORT */ 2296} 2297 2298static void 2299witness_setflag(struct lock_object *lock, int flag, int set) 2300{ 2301 struct lock_list_entry *lock_list; 2302 struct lock_instance *instance; 2303 struct lock_class *class; 2304 2305 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL) 2306 return; 2307 class = LOCK_CLASS(lock); 2308 if (class->lc_flags & LC_SLEEPLOCK) 2309 lock_list = curthread->td_sleeplocks; 2310 else { 2311 if (witness_skipspin) 2312 return; 2313 lock_list = PCPU_GET(spinlocks); 2314 } 2315 instance = find_instance(lock_list, lock); 2316 if (instance == NULL) 2317 panic("%s: lock (%s) %s not locked", __func__, 2318 class->lc_name, lock->lo_name); 2319 2320 if (set) 2321 instance->li_flags |= flag; 2322 else 2323 instance->li_flags &= ~flag; 2324} 2325 2326void 2327witness_norelease(struct lock_object *lock) 2328{ 2329 2330 witness_setflag(lock, LI_NORELEASE, 1); 2331} 2332 2333void 2334witness_releaseok(struct lock_object *lock) 2335{ 2336 2337 witness_setflag(lock, LI_NORELEASE, 0); 2338} 2339 2340#ifdef DDB 2341static void 2342witness_ddb_list(struct thread *td) 2343{ 2344 2345 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__)); 2346 KASSERT(kdb_active, ("%s: not in the debugger", __func__)); 2347 2348 if (witness_watch < 1) 2349 return; 2350 2351 witness_list_locks(&td->td_sleeplocks, db_printf); 2352 2353 /* 2354 * We only handle spinlocks if td == curthread. This is somewhat broken 2355 * if td is currently executing on some other CPU and holds spin locks 2356 * as we won't display those locks. If we had a MI way of getting 2357 * the per-cpu data for a given cpu then we could use 2358 * td->td_oncpu to get the list of spinlocks for this thread 2359 * and "fix" this. 2360 * 2361 * That still wouldn't really fix this unless we locked the scheduler 2362 * lock or stopped the other CPU to make sure it wasn't changing the 2363 * list out from under us. It is probably best to just not try to 2364 * handle threads on other CPU's for now. 2365 */ 2366 if (td == curthread && PCPU_GET(spinlocks) != NULL) 2367 witness_list_locks(PCPU_PTR(spinlocks), db_printf); 2368} 2369 2370DB_SHOW_COMMAND(locks, db_witness_list) 2371{ 2372 struct thread *td; 2373 2374 if (have_addr) 2375 td = db_lookup_thread(addr, TRUE); 2376 else 2377 td = kdb_thread; 2378 witness_ddb_list(td); 2379} 2380 2381DB_SHOW_ALL_COMMAND(locks, db_witness_list_all) 2382{ 2383 struct thread *td; 2384 struct proc *p; 2385 2386 /* 2387 * It would be nice to list only threads and processes that actually 2388 * held sleep locks, but that information is currently not exported 2389 * by WITNESS. 2390 */ 2391 FOREACH_PROC_IN_SYSTEM(p) { 2392 if (!witness_proc_has_locks(p)) 2393 continue; 2394 FOREACH_THREAD_IN_PROC(p, td) { 2395 if (!witness_thread_has_locks(td)) 2396 continue; 2397 db_printf("Process %d (%s) thread %p (%d)\n", p->p_pid, 2398 p->p_comm, td, td->td_tid); 2399 witness_ddb_list(td); 2400 if (db_pager_quit) 2401 return; 2402 } 2403 } 2404} 2405DB_SHOW_ALIAS(alllocks, db_witness_list_all) 2406 2407DB_SHOW_COMMAND(witness, db_witness_display) 2408{ 2409 2410 witness_ddb_display(db_printf); 2411} 2412#endif 2413 2414static int 2415sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS) 2416{ 2417 struct witness_lock_order_data *data1, *data2, *tmp_data1, *tmp_data2; 2418 struct witness *tmp_w1, *tmp_w2, *w1, *w2; 2419 struct sbuf *sb; 2420 u_int w_rmatrix1, w_rmatrix2; 2421 int error, generation, i, j; 2422 2423 tmp_data1 = NULL; 2424 tmp_data2 = NULL; 2425 tmp_w1 = NULL; 2426 tmp_w2 = NULL; 2427 if (witness_watch < 1) { 2428 error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning)); 2429 return (error); 2430 } 2431 if (witness_cold) { 2432 error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold)); 2433 return (error); 2434 } 2435 error = 0; 2436 sb = sbuf_new(NULL, NULL, BADSTACK_SBUF_SIZE, SBUF_AUTOEXTEND); 2437 if (sb == NULL) 2438 return (ENOMEM); 2439 2440 /* Allocate and init temporary storage space. */ 2441 tmp_w1 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO); 2442 tmp_w2 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO); 2443 tmp_data1 = malloc(sizeof(struct witness_lock_order_data), M_TEMP, 2444 M_WAITOK | M_ZERO); 2445 tmp_data2 = malloc(sizeof(struct witness_lock_order_data), M_TEMP, 2446 M_WAITOK | M_ZERO); 2447 stack_zero(&tmp_data1->wlod_stack); 2448 stack_zero(&tmp_data2->wlod_stack); 2449 2450restart: 2451 mtx_lock_spin(&w_mtx); 2452 generation = w_generation; 2453 mtx_unlock_spin(&w_mtx); 2454 sbuf_printf(sb, "Number of known direct relationships is %d\n", 2455 w_lohash.wloh_count); 2456 for (i = 1; i < w_max_used_index; i++) { 2457 mtx_lock_spin(&w_mtx); 2458 if (generation != w_generation) { 2459 mtx_unlock_spin(&w_mtx); 2460 2461 /* The graph has changed, try again. */ 2462 req->oldidx = 0; 2463 sbuf_clear(sb); 2464 goto restart; 2465 } 2466 2467 w1 = &w_data[i]; 2468 if (w1->w_reversed == 0) { 2469 mtx_unlock_spin(&w_mtx); 2470 continue; 2471 } 2472 2473 /* Copy w1 locally so we can release the spin lock. */ 2474 *tmp_w1 = *w1; 2475 mtx_unlock_spin(&w_mtx); 2476 2477 if (tmp_w1->w_reversed == 0) 2478 continue; 2479 for (j = 1; j < w_max_used_index; j++) { 2480 if ((w_rmatrix[i][j] & WITNESS_REVERSAL) == 0 || i > j) 2481 continue; 2482 2483 mtx_lock_spin(&w_mtx); 2484 if (generation != w_generation) { 2485 mtx_unlock_spin(&w_mtx); 2486 2487 /* The graph has changed, try again. */ 2488 req->oldidx = 0; 2489 sbuf_clear(sb); 2490 goto restart; 2491 } 2492 2493 w2 = &w_data[j]; 2494 data1 = witness_lock_order_get(w1, w2); 2495 data2 = witness_lock_order_get(w2, w1); 2496 2497 /* 2498 * Copy information locally so we can release the 2499 * spin lock. 2500 */ 2501 *tmp_w2 = *w2; 2502 w_rmatrix1 = (unsigned int)w_rmatrix[i][j]; 2503 w_rmatrix2 = (unsigned int)w_rmatrix[j][i]; 2504 2505 if (data1) { 2506 stack_zero(&tmp_data1->wlod_stack); 2507 stack_copy(&data1->wlod_stack, 2508 &tmp_data1->wlod_stack); 2509 } 2510 if (data2 && data2 != data1) { 2511 stack_zero(&tmp_data2->wlod_stack); 2512 stack_copy(&data2->wlod_stack, 2513 &tmp_data2->wlod_stack); 2514 } 2515 mtx_unlock_spin(&w_mtx); 2516 2517 sbuf_printf(sb, 2518 "\nLock order reversal between \"%s\"(%s) and \"%s\"(%s)!\n", 2519 tmp_w1->w_name, tmp_w1->w_class->lc_name, 2520 tmp_w2->w_name, tmp_w2->w_class->lc_name); 2521#if 0 2522 sbuf_printf(sb, 2523 "w_rmatrix[%s][%s] == %x, w_rmatrix[%s][%s] == %x\n", 2524 tmp_w1->name, tmp_w2->w_name, w_rmatrix1, 2525 tmp_w2->name, tmp_w1->w_name, w_rmatrix2); 2526#endif 2527 if (data1) { 2528 sbuf_printf(sb, 2529 "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n", 2530 tmp_w1->w_name, tmp_w1->w_class->lc_name, 2531 tmp_w2->w_name, tmp_w2->w_class->lc_name); 2532 stack_sbuf_print(sb, &tmp_data1->wlod_stack); 2533 sbuf_printf(sb, "\n"); 2534 } 2535 if (data2 && data2 != data1) { 2536 sbuf_printf(sb, 2537 "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n", 2538 tmp_w2->w_name, tmp_w2->w_class->lc_name, 2539 tmp_w1->w_name, tmp_w1->w_class->lc_name); 2540 stack_sbuf_print(sb, &tmp_data2->wlod_stack); 2541 sbuf_printf(sb, "\n"); 2542 } 2543 } 2544 } 2545 mtx_lock_spin(&w_mtx); 2546 if (generation != w_generation) { 2547 mtx_unlock_spin(&w_mtx); 2548 2549 /* 2550 * The graph changed while we were printing stack data, 2551 * try again. 2552 */ 2553 req->oldidx = 0; 2554 sbuf_clear(sb); 2555 goto restart; 2556 } 2557 mtx_unlock_spin(&w_mtx); 2558 2559 /* Free temporary storage space. */ 2560 free(tmp_data1, M_TEMP); 2561 free(tmp_data2, M_TEMP); 2562 free(tmp_w1, M_TEMP); 2563 free(tmp_w2, M_TEMP); 2564 2565 sbuf_finish(sb); 2566 error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1); 2567 sbuf_delete(sb); 2568 2569 return (error); 2570} 2571 2572static int 2573sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS) 2574{ 2575 struct witness *w; 2576 struct sbuf *sb; 2577 int error; 2578 2579 if (witness_watch < 1) { 2580 error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning)); 2581 return (error); 2582 } 2583 if (witness_cold) { 2584 error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold)); 2585 return (error); 2586 } 2587 error = 0; 2588 2589 error = sysctl_wire_old_buffer(req, 0); 2590 if (error != 0) 2591 return (error); 2592 sb = sbuf_new_for_sysctl(NULL, NULL, FULLGRAPH_SBUF_SIZE, req); 2593 if (sb == NULL) 2594 return (ENOMEM); 2595 sbuf_printf(sb, "\n"); 2596 2597 mtx_lock_spin(&w_mtx); 2598 STAILQ_FOREACH(w, &w_all, w_list) 2599 w->w_displayed = 0; 2600 STAILQ_FOREACH(w, &w_all, w_list) 2601 witness_add_fullgraph(sb, w); 2602 mtx_unlock_spin(&w_mtx); 2603 2604 /* 2605 * Close the sbuf and return to userland. 2606 */ 2607 error = sbuf_finish(sb); 2608 sbuf_delete(sb); 2609 2610 return (error); 2611} 2612 2613static int 2614sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS) 2615{ 2616 int error, value; 2617 2618 value = witness_watch; 2619 error = sysctl_handle_int(oidp, &value, 0, req); 2620 if (error != 0 || req->newptr == NULL) 2621 return (error); 2622 if (value > 1 || value < -1 || 2623 (witness_watch == -1 && value != witness_watch)) 2624 return (EINVAL); 2625 witness_watch = value; 2626 return (0); 2627} 2628 2629static void 2630witness_add_fullgraph(struct sbuf *sb, struct witness *w) 2631{ 2632 int i; 2633 2634 if (w->w_displayed != 0 || (w->w_file == NULL && w->w_line == 0)) 2635 return; 2636 w->w_displayed = 1; 2637 2638 WITNESS_INDEX_ASSERT(w->w_index); 2639 for (i = 1; i <= w_max_used_index; i++) { 2640 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT) { 2641 sbuf_printf(sb, "\"%s\",\"%s\"\n", w->w_name, 2642 w_data[i].w_name); 2643 witness_add_fullgraph(sb, &w_data[i]); 2644 } 2645 } 2646} 2647 2648/* 2649 * A simple hash function. Takes a key pointer and a key size. If size == 0, 2650 * interprets the key as a string and reads until the null 2651 * terminator. Otherwise, reads the first size bytes. Returns an unsigned 32-bit 2652 * hash value computed from the key. 2653 */ 2654static uint32_t 2655witness_hash_djb2(const uint8_t *key, uint32_t size) 2656{ 2657 unsigned int hash = 5381; 2658 int i; 2659 2660 /* hash = hash * 33 + key[i] */ 2661 if (size) 2662 for (i = 0; i < size; i++) 2663 hash = ((hash << 5) + hash) + (unsigned int)key[i]; 2664 else 2665 for (i = 0; key[i] != 0; i++) 2666 hash = ((hash << 5) + hash) + (unsigned int)key[i]; 2667 2668 return (hash); 2669} 2670 2671 2672/* 2673 * Initializes the two witness hash tables. Called exactly once from 2674 * witness_initialize(). 2675 */ 2676static void 2677witness_init_hash_tables(void) 2678{ 2679 int i; 2680 2681 MPASS(witness_cold); 2682 2683 /* Initialize the hash tables. */ 2684 for (i = 0; i < WITNESS_HASH_SIZE; i++) 2685 w_hash.wh_array[i] = NULL; 2686 2687 w_hash.wh_size = WITNESS_HASH_SIZE; 2688 w_hash.wh_count = 0; 2689 2690 /* Initialize the lock order data hash. */ 2691 w_lofree = NULL; 2692 for (i = 0; i < WITNESS_LO_DATA_COUNT; i++) { 2693 memset(&w_lodata[i], 0, sizeof(w_lodata[i])); 2694 w_lodata[i].wlod_next = w_lofree; 2695 w_lofree = &w_lodata[i]; 2696 } 2697 w_lohash.wloh_size = WITNESS_LO_HASH_SIZE; 2698 w_lohash.wloh_count = 0; 2699 for (i = 0; i < WITNESS_LO_HASH_SIZE; i++) 2700 w_lohash.wloh_array[i] = NULL; 2701} 2702 2703static struct witness * 2704witness_hash_get(const char *key) 2705{ 2706 struct witness *w; 2707 uint32_t hash; 2708 2709 MPASS(key != NULL); 2710 if (witness_cold == 0) 2711 mtx_assert(&w_mtx, MA_OWNED); 2712 hash = witness_hash_djb2(key, 0) % w_hash.wh_size; 2713 w = w_hash.wh_array[hash]; 2714 while (w != NULL) { 2715 if (strcmp(w->w_name, key) == 0) 2716 goto out; 2717 w = w->w_hash_next; 2718 } 2719 2720out: 2721 return (w); 2722} 2723 2724static void 2725witness_hash_put(struct witness *w) 2726{ 2727 uint32_t hash; 2728 2729 MPASS(w != NULL); 2730 MPASS(w->w_name != NULL); 2731 if (witness_cold == 0) 2732 mtx_assert(&w_mtx, MA_OWNED); 2733 KASSERT(witness_hash_get(w->w_name) == NULL, 2734 ("%s: trying to add a hash entry that already exists!", __func__)); 2735 KASSERT(w->w_hash_next == NULL, 2736 ("%s: w->w_hash_next != NULL", __func__)); 2737 2738 hash = witness_hash_djb2(w->w_name, 0) % w_hash.wh_size; 2739 w->w_hash_next = w_hash.wh_array[hash]; 2740 w_hash.wh_array[hash] = w; 2741 w_hash.wh_count++; 2742} 2743 2744 2745static struct witness_lock_order_data * 2746witness_lock_order_get(struct witness *parent, struct witness *child) 2747{ 2748 struct witness_lock_order_data *data = NULL; 2749 struct witness_lock_order_key key; 2750 unsigned int hash; 2751 2752 MPASS(parent != NULL && child != NULL); 2753 key.from = parent->w_index; 2754 key.to = child->w_index; 2755 WITNESS_INDEX_ASSERT(key.from); 2756 WITNESS_INDEX_ASSERT(key.to); 2757 if ((w_rmatrix[parent->w_index][child->w_index] 2758 & WITNESS_LOCK_ORDER_KNOWN) == 0) 2759 goto out; 2760 2761 hash = witness_hash_djb2((const char*)&key, 2762 sizeof(key)) % w_lohash.wloh_size; 2763 data = w_lohash.wloh_array[hash]; 2764 while (data != NULL) { 2765 if (witness_lock_order_key_equal(&data->wlod_key, &key)) 2766 break; 2767 data = data->wlod_next; 2768 } 2769 2770out: 2771 return (data); 2772} 2773 2774/* 2775 * Verify that parent and child have a known relationship, are not the same, 2776 * and child is actually a child of parent. This is done without w_mtx 2777 * to avoid contention in the common case. 2778 */ 2779static int 2780witness_lock_order_check(struct witness *parent, struct witness *child) 2781{ 2782 2783 if (parent != child && 2784 w_rmatrix[parent->w_index][child->w_index] 2785 & WITNESS_LOCK_ORDER_KNOWN && 2786 isitmychild(parent, child)) 2787 return (1); 2788 2789 return (0); 2790} 2791 2792static int 2793witness_lock_order_add(struct witness *parent, struct witness *child) 2794{ 2795 struct witness_lock_order_data *data = NULL; 2796 struct witness_lock_order_key key; 2797 unsigned int hash; 2798 2799 MPASS(parent != NULL && child != NULL); 2800 key.from = parent->w_index; 2801 key.to = child->w_index; 2802 WITNESS_INDEX_ASSERT(key.from); 2803 WITNESS_INDEX_ASSERT(key.to); 2804 if (w_rmatrix[parent->w_index][child->w_index] 2805 & WITNESS_LOCK_ORDER_KNOWN) 2806 return (1); 2807 2808 hash = witness_hash_djb2((const char*)&key, 2809 sizeof(key)) % w_lohash.wloh_size; 2810 w_rmatrix[parent->w_index][child->w_index] |= WITNESS_LOCK_ORDER_KNOWN; 2811 data = w_lofree; 2812 if (data == NULL) 2813 return (0); 2814 w_lofree = data->wlod_next; 2815 data->wlod_next = w_lohash.wloh_array[hash]; 2816 data->wlod_key = key; 2817 w_lohash.wloh_array[hash] = data; 2818 w_lohash.wloh_count++; 2819 stack_zero(&data->wlod_stack); 2820 stack_save(&data->wlod_stack); 2821 return (1); 2822} 2823 2824/* Call this whenver the structure of the witness graph changes. */ 2825static void 2826witness_increment_graph_generation(void) 2827{ 2828 2829 if (witness_cold == 0) 2830 mtx_assert(&w_mtx, MA_OWNED); 2831 w_generation++; 2832} 2833 2834#ifdef KDB 2835static void 2836_witness_debugger(int cond, const char *msg) 2837{ 2838 2839 if (witness_trace && cond) 2840 kdb_backtrace(); 2841 if (witness_kdb && cond) 2842 kdb_enter(KDB_WHY_WITNESS, msg); 2843} 2844#endif 2845