subr_witness.c revision 272925
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: head/sys/kern/subr_witness.c 272925 2014-10-11 02:02:58Z marcel $"); 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#ifndef WITNESS_COUNT 136#define WITNESS_COUNT 1536 137#endif 138#define WITNESS_HASH_SIZE 251 /* Prime, gives load factor < 2 */ 139#define WITNESS_PENDLIST (1024 + MAXCPU) 140 141/* Allocate 256 KB of stack data space */ 142#define WITNESS_LO_DATA_COUNT 2048 143 144/* Prime, gives load factor of ~2 at full load */ 145#define WITNESS_LO_HASH_SIZE 1021 146 147/* 148 * XXX: This is somewhat bogus, as we assume here that at most 2048 threads 149 * will hold LOCK_NCHILDREN locks. We handle failure ok, and we should 150 * probably be safe for the most part, but it's still a SWAG. 151 */ 152#define LOCK_NCHILDREN 5 153#define LOCK_CHILDCOUNT 2048 154 155#define MAX_W_NAME 64 156 157#define FULLGRAPH_SBUF_SIZE 512 158 159/* 160 * These flags go in the witness relationship matrix and describe the 161 * relationship between any two struct witness objects. 162 */ 163#define WITNESS_UNRELATED 0x00 /* No lock order relation. */ 164#define WITNESS_PARENT 0x01 /* Parent, aka direct ancestor. */ 165#define WITNESS_ANCESTOR 0x02 /* Direct or indirect ancestor. */ 166#define WITNESS_CHILD 0x04 /* Child, aka direct descendant. */ 167#define WITNESS_DESCENDANT 0x08 /* Direct or indirect descendant. */ 168#define WITNESS_ANCESTOR_MASK (WITNESS_PARENT | WITNESS_ANCESTOR) 169#define WITNESS_DESCENDANT_MASK (WITNESS_CHILD | WITNESS_DESCENDANT) 170#define WITNESS_RELATED_MASK \ 171 (WITNESS_ANCESTOR_MASK | WITNESS_DESCENDANT_MASK) 172#define WITNESS_REVERSAL 0x10 /* A lock order reversal has been 173 * observed. */ 174#define WITNESS_RESERVED1 0x20 /* Unused flag, reserved. */ 175#define WITNESS_RESERVED2 0x40 /* Unused flag, reserved. */ 176#define WITNESS_LOCK_ORDER_KNOWN 0x80 /* This lock order is known. */ 177 178/* Descendant to ancestor flags */ 179#define WITNESS_DTOA(x) (((x) & WITNESS_RELATED_MASK) >> 2) 180 181/* Ancestor to descendant flags */ 182#define WITNESS_ATOD(x) (((x) & WITNESS_RELATED_MASK) << 2) 183 184#define WITNESS_INDEX_ASSERT(i) \ 185 MPASS((i) > 0 && (i) <= w_max_used_index && (i) < witness_count) 186 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 const 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; 383SYSCTL_PROC(_debug_witness, OID_AUTO, watch, CTLFLAG_RWTUN | CTLTYPE_INT, NULL, 0, 384 sysctl_debug_witness_watch, "I", "witness is watching lock operations"); 385 386#ifdef KDB 387/* 388 * When KDB is enabled and witness_kdb is 1, it will cause the system 389 * to drop into kdebug() when: 390 * - a lock hierarchy violation occurs 391 * - locks are held when going to sleep. 392 */ 393#ifdef WITNESS_KDB 394int witness_kdb = 1; 395#else 396int witness_kdb = 0; 397#endif 398SYSCTL_INT(_debug_witness, OID_AUTO, kdb, CTLFLAG_RWTUN, &witness_kdb, 0, ""); 399 400/* 401 * When KDB is enabled and witness_trace is 1, it will cause the system 402 * to print a stack trace: 403 * - a lock hierarchy violation occurs 404 * - locks are held when going to sleep. 405 */ 406int witness_trace = 1; 407SYSCTL_INT(_debug_witness, OID_AUTO, trace, CTLFLAG_RWTUN, &witness_trace, 0, ""); 408#endif /* KDB */ 409 410#ifdef WITNESS_SKIPSPIN 411int witness_skipspin = 1; 412#else 413int witness_skipspin = 0; 414#endif 415SYSCTL_INT(_debug_witness, OID_AUTO, skipspin, CTLFLAG_RDTUN, &witness_skipspin, 0, ""); 416 417/* tunable for Witness count */ 418int witness_count = WITNESS_COUNT; 419int badstack_sbuf_size = WITNESS_COUNT * 256; 420 421SYSCTL_INT(_debug_witness, OID_AUTO, witness_count, CTLFLAG_RDTUN, 422 &witness_count, 0, ""); 423 424/* 425 * Call this to print out the relations between locks. 426 */ 427SYSCTL_PROC(_debug_witness, OID_AUTO, fullgraph, CTLTYPE_STRING | CTLFLAG_RD, 428 NULL, 0, sysctl_debug_witness_fullgraph, "A", "Show locks relation graphs"); 429 430/* 431 * Call this to print out the witness faulty stacks. 432 */ 433SYSCTL_PROC(_debug_witness, OID_AUTO, badstacks, CTLTYPE_STRING | CTLFLAG_RD, 434 NULL, 0, sysctl_debug_witness_badstacks, "A", "Show bad witness stacks"); 435 436static struct mtx w_mtx; 437 438/* w_list */ 439static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free); 440static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all); 441 442/* w_typelist */ 443static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin); 444static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep); 445 446/* lock list */ 447static struct lock_list_entry *w_lock_list_free = NULL; 448static struct witness_pendhelp pending_locks[WITNESS_PENDLIST]; 449static u_int pending_cnt; 450 451static int w_free_cnt, w_spin_cnt, w_sleep_cnt; 452SYSCTL_INT(_debug_witness, OID_AUTO, free_cnt, CTLFLAG_RD, &w_free_cnt, 0, ""); 453SYSCTL_INT(_debug_witness, OID_AUTO, spin_cnt, CTLFLAG_RD, &w_spin_cnt, 0, ""); 454SYSCTL_INT(_debug_witness, OID_AUTO, sleep_cnt, CTLFLAG_RD, &w_sleep_cnt, 0, 455 ""); 456 457static struct witness *w_data; 458static uint8_t **w_rmatrix; 459static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT]; 460static struct witness_hash w_hash; /* The witness hash table. */ 461 462/* The lock order data hash */ 463static struct witness_lock_order_data w_lodata[WITNESS_LO_DATA_COUNT]; 464static struct witness_lock_order_data *w_lofree = NULL; 465static struct witness_lock_order_hash w_lohash; 466static int w_max_used_index = 0; 467static unsigned int w_generation = 0; 468static const char w_notrunning[] = "Witness not running\n"; 469static const char w_stillcold[] = "Witness is still cold\n"; 470 471 472static struct witness_order_list_entry order_lists[] = { 473 /* 474 * sx locks 475 */ 476 { "proctree", &lock_class_sx }, 477 { "allproc", &lock_class_sx }, 478 { "allprison", &lock_class_sx }, 479 { NULL, NULL }, 480 /* 481 * Various mutexes 482 */ 483 { "Giant", &lock_class_mtx_sleep }, 484 { "pipe mutex", &lock_class_mtx_sleep }, 485 { "sigio lock", &lock_class_mtx_sleep }, 486 { "process group", &lock_class_mtx_sleep }, 487 { "process lock", &lock_class_mtx_sleep }, 488 { "session", &lock_class_mtx_sleep }, 489 { "uidinfo hash", &lock_class_rw }, 490#ifdef HWPMC_HOOKS 491 { "pmc-sleep", &lock_class_mtx_sleep }, 492#endif 493 { "time lock", &lock_class_mtx_sleep }, 494 { NULL, NULL }, 495 /* 496 * Sockets 497 */ 498 { "accept", &lock_class_mtx_sleep }, 499 { "so_snd", &lock_class_mtx_sleep }, 500 { "so_rcv", &lock_class_mtx_sleep }, 501 { "sellck", &lock_class_mtx_sleep }, 502 { NULL, NULL }, 503 /* 504 * Routing 505 */ 506 { "so_rcv", &lock_class_mtx_sleep }, 507 { "radix node head", &lock_class_rw }, 508 { "rtentry", &lock_class_mtx_sleep }, 509 { "ifaddr", &lock_class_mtx_sleep }, 510 { NULL, NULL }, 511 /* 512 * IPv4 multicast: 513 * protocol locks before interface locks, after UDP locks. 514 */ 515 { "udpinp", &lock_class_rw }, 516 { "in_multi_mtx", &lock_class_mtx_sleep }, 517 { "igmp_mtx", &lock_class_mtx_sleep }, 518 { "if_addr_lock", &lock_class_rw }, 519 { NULL, NULL }, 520 /* 521 * IPv6 multicast: 522 * protocol locks before interface locks, after UDP locks. 523 */ 524 { "udpinp", &lock_class_rw }, 525 { "in6_multi_mtx", &lock_class_mtx_sleep }, 526 { "mld_mtx", &lock_class_mtx_sleep }, 527 { "if_addr_lock", &lock_class_rw }, 528 { NULL, NULL }, 529 /* 530 * UNIX Domain Sockets 531 */ 532 { "unp_global_rwlock", &lock_class_rw }, 533 { "unp_list_lock", &lock_class_mtx_sleep }, 534 { "unp", &lock_class_mtx_sleep }, 535 { "so_snd", &lock_class_mtx_sleep }, 536 { NULL, NULL }, 537 /* 538 * UDP/IP 539 */ 540 { "udp", &lock_class_rw }, 541 { "udpinp", &lock_class_rw }, 542 { "so_snd", &lock_class_mtx_sleep }, 543 { NULL, NULL }, 544 /* 545 * TCP/IP 546 */ 547 { "tcp", &lock_class_rw }, 548 { "tcpinp", &lock_class_rw }, 549 { "so_snd", &lock_class_mtx_sleep }, 550 { NULL, NULL }, 551 /* 552 * BPF 553 */ 554 { "bpf global lock", &lock_class_mtx_sleep }, 555 { "bpf interface lock", &lock_class_rw }, 556 { "bpf cdev lock", &lock_class_mtx_sleep }, 557 { NULL, NULL }, 558 /* 559 * NFS server 560 */ 561 { "nfsd_mtx", &lock_class_mtx_sleep }, 562 { "so_snd", &lock_class_mtx_sleep }, 563 { NULL, NULL }, 564 565 /* 566 * IEEE 802.11 567 */ 568 { "802.11 com lock", &lock_class_mtx_sleep}, 569 { NULL, NULL }, 570 /* 571 * Network drivers 572 */ 573 { "network driver", &lock_class_mtx_sleep}, 574 { NULL, NULL }, 575 576 /* 577 * Netgraph 578 */ 579 { "ng_node", &lock_class_mtx_sleep }, 580 { "ng_worklist", &lock_class_mtx_sleep }, 581 { NULL, NULL }, 582 /* 583 * CDEV 584 */ 585 { "vm map (system)", &lock_class_mtx_sleep }, 586 { "vm page queue", &lock_class_mtx_sleep }, 587 { "vnode interlock", &lock_class_mtx_sleep }, 588 { "cdev", &lock_class_mtx_sleep }, 589 { NULL, NULL }, 590 /* 591 * VM 592 */ 593 { "vm map (user)", &lock_class_sx }, 594 { "vm object", &lock_class_rw }, 595 { "vm page", &lock_class_mtx_sleep }, 596 { "vm page queue", &lock_class_mtx_sleep }, 597 { "pmap pv global", &lock_class_rw }, 598 { "pmap", &lock_class_mtx_sleep }, 599 { "pmap pv list", &lock_class_rw }, 600 { "vm page free queue", &lock_class_mtx_sleep }, 601 { NULL, NULL }, 602 /* 603 * kqueue/VFS interaction 604 */ 605 { "kqueue", &lock_class_mtx_sleep }, 606 { "struct mount mtx", &lock_class_mtx_sleep }, 607 { "vnode interlock", &lock_class_mtx_sleep }, 608 { NULL, NULL }, 609 /* 610 * ZFS locking 611 */ 612 { "dn->dn_mtx", &lock_class_sx }, 613 { "dr->dt.di.dr_mtx", &lock_class_sx }, 614 { "db->db_mtx", &lock_class_sx }, 615 { NULL, NULL }, 616 /* 617 * spin locks 618 */ 619#ifdef SMP 620 { "ap boot", &lock_class_mtx_spin }, 621#endif 622 { "rm.mutex_mtx", &lock_class_mtx_spin }, 623 { "sio", &lock_class_mtx_spin }, 624 { "scrlock", &lock_class_mtx_spin }, 625#ifdef __i386__ 626 { "cy", &lock_class_mtx_spin }, 627#endif 628#ifdef __sparc64__ 629 { "pcib_mtx", &lock_class_mtx_spin }, 630 { "rtc_mtx", &lock_class_mtx_spin }, 631#endif 632 { "scc_hwmtx", &lock_class_mtx_spin }, 633 { "uart_hwmtx", &lock_class_mtx_spin }, 634 { "fast_taskqueue", &lock_class_mtx_spin }, 635 { "intr table", &lock_class_mtx_spin }, 636#ifdef HWPMC_HOOKS 637 { "pmc-per-proc", &lock_class_mtx_spin }, 638#endif 639 { "process slock", &lock_class_mtx_spin }, 640 { "sleepq chain", &lock_class_mtx_spin }, 641 { "umtx lock", &lock_class_mtx_spin }, 642 { "rm_spinlock", &lock_class_mtx_spin }, 643 { "turnstile chain", &lock_class_mtx_spin }, 644 { "turnstile lock", &lock_class_mtx_spin }, 645 { "sched lock", &lock_class_mtx_spin }, 646 { "td_contested", &lock_class_mtx_spin }, 647 { "callout", &lock_class_mtx_spin }, 648 { "entropy harvest mutex", &lock_class_mtx_spin }, 649 { "syscons video lock", &lock_class_mtx_spin }, 650#ifdef SMP 651 { "smp rendezvous", &lock_class_mtx_spin }, 652#endif 653#ifdef __powerpc__ 654 { "tlb0", &lock_class_mtx_spin }, 655#endif 656 /* 657 * leaf locks 658 */ 659 { "intrcnt", &lock_class_mtx_spin }, 660 { "icu", &lock_class_mtx_spin }, 661#ifdef __i386__ 662 { "allpmaps", &lock_class_mtx_spin }, 663 { "descriptor tables", &lock_class_mtx_spin }, 664#endif 665 { "clk", &lock_class_mtx_spin }, 666 { "cpuset", &lock_class_mtx_spin }, 667 { "mprof lock", &lock_class_mtx_spin }, 668 { "zombie lock", &lock_class_mtx_spin }, 669 { "ALD Queue", &lock_class_mtx_spin }, 670#if defined(__i386__) || defined(__amd64__) 671 { "pcicfg", &lock_class_mtx_spin }, 672 { "NDIS thread lock", &lock_class_mtx_spin }, 673#endif 674 { "tw_osl_io_lock", &lock_class_mtx_spin }, 675 { "tw_osl_q_lock", &lock_class_mtx_spin }, 676 { "tw_cl_io_lock", &lock_class_mtx_spin }, 677 { "tw_cl_intr_lock", &lock_class_mtx_spin }, 678 { "tw_cl_gen_lock", &lock_class_mtx_spin }, 679#ifdef HWPMC_HOOKS 680 { "pmc-leaf", &lock_class_mtx_spin }, 681#endif 682 { "blocked lock", &lock_class_mtx_spin }, 683 { NULL, NULL }, 684 { NULL, NULL } 685}; 686 687#ifdef BLESSING 688/* 689 * Pairs of locks which have been blessed 690 * Don't complain about order problems with blessed locks 691 */ 692static struct witness_blessed blessed_list[] = { 693}; 694static int blessed_count = 695 sizeof(blessed_list) / sizeof(struct witness_blessed); 696#endif 697 698/* 699 * This global is set to 0 once it becomes safe to use the witness code. 700 */ 701static int witness_cold = 1; 702 703/* 704 * This global is set to 1 once the static lock orders have been enrolled 705 * so that a warning can be issued for any spin locks enrolled later. 706 */ 707static int witness_spin_warn = 0; 708 709/* Trim useless garbage from filenames. */ 710static const char * 711fixup_filename(const char *file) 712{ 713 714 if (file == NULL) 715 return (NULL); 716 while (strncmp(file, "../", 3) == 0) 717 file += 3; 718 return (file); 719} 720 721/* 722 * The WITNESS-enabled diagnostic code. Note that the witness code does 723 * assume that the early boot is single-threaded at least until after this 724 * routine is completed. 725 */ 726static void 727witness_initialize(void *dummy __unused) 728{ 729 struct lock_object *lock; 730 struct witness_order_list_entry *order; 731 struct witness *w, *w1; 732 int i; 733 734 w_data = malloc(sizeof (struct witness) * witness_count, M_WITNESS, 735 M_NOWAIT | M_ZERO); 736 737 w_rmatrix = malloc(sizeof(uint8_t *) * (witness_count+1), 738 M_WITNESS, M_NOWAIT | M_ZERO); 739 740 for(i = 0; i < witness_count+1; i++) { 741 w_rmatrix[i] = malloc(sizeof(uint8_t) * (witness_count + 1), 742 M_WITNESS, M_NOWAIT | M_ZERO); 743 } 744 badstack_sbuf_size = witness_count * 256; 745 746 /* 747 * We have to release Giant before initializing its witness 748 * structure so that WITNESS doesn't get confused. 749 */ 750 mtx_unlock(&Giant); 751 mtx_assert(&Giant, MA_NOTOWNED); 752 753 CTR1(KTR_WITNESS, "%s: initializing witness", __func__); 754 mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET | 755 MTX_NOWITNESS | MTX_NOPROFILE); 756 for (i = witness_count - 1; i >= 0; i--) { 757 w = &w_data[i]; 758 memset(w, 0, sizeof(*w)); 759 w_data[i].w_index = i; /* Witness index never changes. */ 760 witness_free(w); 761 } 762 KASSERT(STAILQ_FIRST(&w_free)->w_index == 0, 763 ("%s: Invalid list of free witness objects", __func__)); 764 765 /* Witness with index 0 is not used to aid in debugging. */ 766 STAILQ_REMOVE_HEAD(&w_free, w_list); 767 w_free_cnt--; 768 769 for(i = 0; i < witness_count; i++) { 770 memset(w_rmatrix[i], 0, sizeof(uint8_t) * 771 (witness_count + 1)); 772 } 773 774 for (i = 0; i < LOCK_CHILDCOUNT; i++) 775 witness_lock_list_free(&w_locklistdata[i]); 776 witness_init_hash_tables(); 777 778 /* First add in all the specified order lists. */ 779 for (order = order_lists; order->w_name != NULL; order++) { 780 w = enroll(order->w_name, order->w_class); 781 if (w == NULL) 782 continue; 783 w->w_file = "order list"; 784 for (order++; order->w_name != NULL; order++) { 785 w1 = enroll(order->w_name, order->w_class); 786 if (w1 == NULL) 787 continue; 788 w1->w_file = "order list"; 789 itismychild(w, w1); 790 w = w1; 791 } 792 } 793 witness_spin_warn = 1; 794 795 /* Iterate through all locks and add them to witness. */ 796 for (i = 0; pending_locks[i].wh_lock != NULL; i++) { 797 lock = pending_locks[i].wh_lock; 798 KASSERT(lock->lo_flags & LO_WITNESS, 799 ("%s: lock %s is on pending list but not LO_WITNESS", 800 __func__, lock->lo_name)); 801 lock->lo_witness = enroll(pending_locks[i].wh_type, 802 LOCK_CLASS(lock)); 803 } 804 805 /* Mark the witness code as being ready for use. */ 806 witness_cold = 0; 807 808 mtx_lock(&Giant); 809} 810SYSINIT(witness_init, SI_SUB_WITNESS, SI_ORDER_FIRST, witness_initialize, 811 NULL); 812 813void 814witness_init(struct lock_object *lock, const char *type) 815{ 816 struct lock_class *class; 817 818 /* Various sanity checks. */ 819 class = LOCK_CLASS(lock); 820 if ((lock->lo_flags & LO_RECURSABLE) != 0 && 821 (class->lc_flags & LC_RECURSABLE) == 0) 822 kassert_panic("%s: lock (%s) %s can not be recursable", 823 __func__, class->lc_name, lock->lo_name); 824 if ((lock->lo_flags & LO_SLEEPABLE) != 0 && 825 (class->lc_flags & LC_SLEEPABLE) == 0) 826 kassert_panic("%s: lock (%s) %s can not be sleepable", 827 __func__, class->lc_name, lock->lo_name); 828 if ((lock->lo_flags & LO_UPGRADABLE) != 0 && 829 (class->lc_flags & LC_UPGRADABLE) == 0) 830 kassert_panic("%s: lock (%s) %s can not be upgradable", 831 __func__, class->lc_name, lock->lo_name); 832 833 /* 834 * If we shouldn't watch this lock, then just clear lo_witness. 835 * Otherwise, if witness_cold is set, then it is too early to 836 * enroll this lock, so defer it to witness_initialize() by adding 837 * it to the pending_locks list. If it is not too early, then enroll 838 * the lock now. 839 */ 840 if (witness_watch < 1 || panicstr != NULL || 841 (lock->lo_flags & LO_WITNESS) == 0) 842 lock->lo_witness = NULL; 843 else if (witness_cold) { 844 pending_locks[pending_cnt].wh_lock = lock; 845 pending_locks[pending_cnt++].wh_type = type; 846 if (pending_cnt > WITNESS_PENDLIST) 847 panic("%s: pending locks list is too small, " 848 "increase WITNESS_PENDLIST\n", 849 __func__); 850 } else 851 lock->lo_witness = enroll(type, class); 852} 853 854void 855witness_destroy(struct lock_object *lock) 856{ 857 struct lock_class *class; 858 struct witness *w; 859 860 class = LOCK_CLASS(lock); 861 862 if (witness_cold) 863 panic("lock (%s) %s destroyed while witness_cold", 864 class->lc_name, lock->lo_name); 865 866 /* XXX: need to verify that no one holds the lock */ 867 if ((lock->lo_flags & LO_WITNESS) == 0 || lock->lo_witness == NULL) 868 return; 869 w = lock->lo_witness; 870 871 mtx_lock_spin(&w_mtx); 872 MPASS(w->w_refcount > 0); 873 w->w_refcount--; 874 875 if (w->w_refcount == 0) 876 depart(w); 877 mtx_unlock_spin(&w_mtx); 878} 879 880#ifdef DDB 881static void 882witness_ddb_compute_levels(void) 883{ 884 struct witness *w; 885 886 /* 887 * First clear all levels. 888 */ 889 STAILQ_FOREACH(w, &w_all, w_list) 890 w->w_ddb_level = -1; 891 892 /* 893 * Look for locks with no parents and level all their descendants. 894 */ 895 STAILQ_FOREACH(w, &w_all, w_list) { 896 897 /* If the witness has ancestors (is not a root), skip it. */ 898 if (w->w_num_ancestors > 0) 899 continue; 900 witness_ddb_level_descendants(w, 0); 901 } 902} 903 904static void 905witness_ddb_level_descendants(struct witness *w, int l) 906{ 907 int i; 908 909 if (w->w_ddb_level >= l) 910 return; 911 912 w->w_ddb_level = l; 913 l++; 914 915 for (i = 1; i <= w_max_used_index; i++) { 916 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT) 917 witness_ddb_level_descendants(&w_data[i], l); 918 } 919} 920 921static void 922witness_ddb_display_descendants(int(*prnt)(const char *fmt, ...), 923 struct witness *w, int indent) 924{ 925 int i; 926 927 for (i = 0; i < indent; i++) 928 prnt(" "); 929 prnt("%s (type: %s, depth: %d, active refs: %d)", 930 w->w_name, w->w_class->lc_name, 931 w->w_ddb_level, w->w_refcount); 932 if (w->w_displayed) { 933 prnt(" -- (already displayed)\n"); 934 return; 935 } 936 w->w_displayed = 1; 937 if (w->w_file != NULL && w->w_line != 0) 938 prnt(" -- last acquired @ %s:%d\n", fixup_filename(w->w_file), 939 w->w_line); 940 else 941 prnt(" -- never acquired\n"); 942 indent++; 943 WITNESS_INDEX_ASSERT(w->w_index); 944 for (i = 1; i <= w_max_used_index; i++) { 945 if (db_pager_quit) 946 return; 947 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT) 948 witness_ddb_display_descendants(prnt, &w_data[i], 949 indent); 950 } 951} 952 953static void 954witness_ddb_display_list(int(*prnt)(const char *fmt, ...), 955 struct witness_list *list) 956{ 957 struct witness *w; 958 959 STAILQ_FOREACH(w, list, w_typelist) { 960 if (w->w_file == NULL || w->w_ddb_level > 0) 961 continue; 962 963 /* This lock has no anscestors - display its descendants. */ 964 witness_ddb_display_descendants(prnt, w, 0); 965 if (db_pager_quit) 966 return; 967 } 968} 969 970static void 971witness_ddb_display(int(*prnt)(const char *fmt, ...)) 972{ 973 struct witness *w; 974 975 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__)); 976 witness_ddb_compute_levels(); 977 978 /* Clear all the displayed flags. */ 979 STAILQ_FOREACH(w, &w_all, w_list) 980 w->w_displayed = 0; 981 982 /* 983 * First, handle sleep locks which have been acquired at least 984 * once. 985 */ 986 prnt("Sleep locks:\n"); 987 witness_ddb_display_list(prnt, &w_sleep); 988 if (db_pager_quit) 989 return; 990 991 /* 992 * Now do spin locks which have been acquired at least once. 993 */ 994 prnt("\nSpin locks:\n"); 995 witness_ddb_display_list(prnt, &w_spin); 996 if (db_pager_quit) 997 return; 998 999 /* 1000 * Finally, any locks which have not been acquired yet. 1001 */ 1002 prnt("\nLocks which were never acquired:\n"); 1003 STAILQ_FOREACH(w, &w_all, w_list) { 1004 if (w->w_file != NULL || w->w_refcount == 0) 1005 continue; 1006 prnt("%s (type: %s, depth: %d)\n", w->w_name, 1007 w->w_class->lc_name, w->w_ddb_level); 1008 if (db_pager_quit) 1009 return; 1010 } 1011} 1012#endif /* DDB */ 1013 1014int 1015witness_defineorder(struct lock_object *lock1, struct lock_object *lock2) 1016{ 1017 1018 if (witness_watch == -1 || panicstr != NULL) 1019 return (0); 1020 1021 /* Require locks that witness knows about. */ 1022 if (lock1 == NULL || lock1->lo_witness == NULL || lock2 == NULL || 1023 lock2->lo_witness == NULL) 1024 return (EINVAL); 1025 1026 mtx_assert(&w_mtx, MA_NOTOWNED); 1027 mtx_lock_spin(&w_mtx); 1028 1029 /* 1030 * If we already have either an explicit or implied lock order that 1031 * is the other way around, then return an error. 1032 */ 1033 if (witness_watch && 1034 isitmydescendant(lock2->lo_witness, lock1->lo_witness)) { 1035 mtx_unlock_spin(&w_mtx); 1036 return (EDOOFUS); 1037 } 1038 1039 /* Try to add the new order. */ 1040 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__, 1041 lock2->lo_witness->w_name, lock1->lo_witness->w_name); 1042 itismychild(lock1->lo_witness, lock2->lo_witness); 1043 mtx_unlock_spin(&w_mtx); 1044 return (0); 1045} 1046 1047void 1048witness_checkorder(struct lock_object *lock, int flags, const char *file, 1049 int line, struct lock_object *interlock) 1050{ 1051 struct lock_list_entry *lock_list, *lle; 1052 struct lock_instance *lock1, *lock2, *plock; 1053 struct lock_class *class, *iclass; 1054 struct witness *w, *w1; 1055 struct thread *td; 1056 int i, j; 1057 1058 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL || 1059 panicstr != NULL) 1060 return; 1061 1062 w = lock->lo_witness; 1063 class = LOCK_CLASS(lock); 1064 td = curthread; 1065 1066 if (class->lc_flags & LC_SLEEPLOCK) { 1067 1068 /* 1069 * Since spin locks include a critical section, this check 1070 * implicitly enforces a lock order of all sleep locks before 1071 * all spin locks. 1072 */ 1073 if (td->td_critnest != 0 && !kdb_active) 1074 kassert_panic("acquiring blockable sleep lock with " 1075 "spinlock or critical section held (%s) %s @ %s:%d", 1076 class->lc_name, lock->lo_name, 1077 fixup_filename(file), line); 1078 1079 /* 1080 * If this is the first lock acquired then just return as 1081 * no order checking is needed. 1082 */ 1083 lock_list = td->td_sleeplocks; 1084 if (lock_list == NULL || lock_list->ll_count == 0) 1085 return; 1086 } else { 1087 1088 /* 1089 * If this is the first lock, just return as no order 1090 * checking is needed. Avoid problems with thread 1091 * migration pinning the thread while checking if 1092 * spinlocks are held. If at least one spinlock is held 1093 * the thread is in a safe path and it is allowed to 1094 * unpin it. 1095 */ 1096 sched_pin(); 1097 lock_list = PCPU_GET(spinlocks); 1098 if (lock_list == NULL || lock_list->ll_count == 0) { 1099 sched_unpin(); 1100 return; 1101 } 1102 sched_unpin(); 1103 } 1104 1105 /* 1106 * Check to see if we are recursing on a lock we already own. If 1107 * so, make sure that we don't mismatch exclusive and shared lock 1108 * acquires. 1109 */ 1110 lock1 = find_instance(lock_list, lock); 1111 if (lock1 != NULL) { 1112 if ((lock1->li_flags & LI_EXCLUSIVE) != 0 && 1113 (flags & LOP_EXCLUSIVE) == 0) { 1114 printf("shared lock of (%s) %s @ %s:%d\n", 1115 class->lc_name, lock->lo_name, 1116 fixup_filename(file), line); 1117 printf("while exclusively locked from %s:%d\n", 1118 fixup_filename(lock1->li_file), lock1->li_line); 1119 kassert_panic("excl->share"); 1120 } 1121 if ((lock1->li_flags & LI_EXCLUSIVE) == 0 && 1122 (flags & LOP_EXCLUSIVE) != 0) { 1123 printf("exclusive lock of (%s) %s @ %s:%d\n", 1124 class->lc_name, lock->lo_name, 1125 fixup_filename(file), line); 1126 printf("while share locked from %s:%d\n", 1127 fixup_filename(lock1->li_file), lock1->li_line); 1128 kassert_panic("share->excl"); 1129 } 1130 return; 1131 } 1132 1133 /* Warn if the interlock is not locked exactly once. */ 1134 if (interlock != NULL) { 1135 iclass = LOCK_CLASS(interlock); 1136 lock1 = find_instance(lock_list, interlock); 1137 if (lock1 == NULL) 1138 kassert_panic("interlock (%s) %s not locked @ %s:%d", 1139 iclass->lc_name, interlock->lo_name, 1140 fixup_filename(file), line); 1141 else if ((lock1->li_flags & LI_RECURSEMASK) != 0) 1142 kassert_panic("interlock (%s) %s recursed @ %s:%d", 1143 iclass->lc_name, interlock->lo_name, 1144 fixup_filename(file), line); 1145 } 1146 1147 /* 1148 * Find the previously acquired lock, but ignore interlocks. 1149 */ 1150 plock = &lock_list->ll_children[lock_list->ll_count - 1]; 1151 if (interlock != NULL && plock->li_lock == interlock) { 1152 if (lock_list->ll_count > 1) 1153 plock = 1154 &lock_list->ll_children[lock_list->ll_count - 2]; 1155 else { 1156 lle = lock_list->ll_next; 1157 1158 /* 1159 * The interlock is the only lock we hold, so 1160 * simply return. 1161 */ 1162 if (lle == NULL) 1163 return; 1164 plock = &lle->ll_children[lle->ll_count - 1]; 1165 } 1166 } 1167 1168 /* 1169 * Try to perform most checks without a lock. If this succeeds we 1170 * can skip acquiring the lock and return success. 1171 */ 1172 w1 = plock->li_lock->lo_witness; 1173 if (witness_lock_order_check(w1, w)) 1174 return; 1175 1176 /* 1177 * Check for duplicate locks of the same type. Note that we only 1178 * have to check for this on the last lock we just acquired. Any 1179 * other cases will be caught as lock order violations. 1180 */ 1181 mtx_lock_spin(&w_mtx); 1182 witness_lock_order_add(w1, w); 1183 if (w1 == w) { 1184 i = w->w_index; 1185 if (!(lock->lo_flags & LO_DUPOK) && !(flags & LOP_DUPOK) && 1186 !(w_rmatrix[i][i] & WITNESS_REVERSAL)) { 1187 w_rmatrix[i][i] |= WITNESS_REVERSAL; 1188 w->w_reversed = 1; 1189 mtx_unlock_spin(&w_mtx); 1190 printf( 1191 "acquiring duplicate lock of same type: \"%s\"\n", 1192 w->w_name); 1193 printf(" 1st %s @ %s:%d\n", plock->li_lock->lo_name, 1194 fixup_filename(plock->li_file), plock->li_line); 1195 printf(" 2nd %s @ %s:%d\n", lock->lo_name, 1196 fixup_filename(file), line); 1197 witness_debugger(1); 1198 } else 1199 mtx_unlock_spin(&w_mtx); 1200 return; 1201 } 1202 mtx_assert(&w_mtx, MA_OWNED); 1203 1204 /* 1205 * If we know that the lock we are acquiring comes after 1206 * the lock we most recently acquired in the lock order tree, 1207 * then there is no need for any further checks. 1208 */ 1209 if (isitmychild(w1, w)) 1210 goto out; 1211 1212 for (j = 0, lle = lock_list; lle != NULL; lle = lle->ll_next) { 1213 for (i = lle->ll_count - 1; i >= 0; i--, j++) { 1214 1215 MPASS(j < witness_count); 1216 lock1 = &lle->ll_children[i]; 1217 1218 /* 1219 * Ignore the interlock. 1220 */ 1221 if (interlock == lock1->li_lock) 1222 continue; 1223 1224 /* 1225 * If this lock doesn't undergo witness checking, 1226 * then skip it. 1227 */ 1228 w1 = lock1->li_lock->lo_witness; 1229 if (w1 == NULL) { 1230 KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0, 1231 ("lock missing witness structure")); 1232 continue; 1233 } 1234 1235 /* 1236 * If we are locking Giant and this is a sleepable 1237 * lock, then skip it. 1238 */ 1239 if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0 && 1240 lock == &Giant.lock_object) 1241 continue; 1242 1243 /* 1244 * If we are locking a sleepable lock and this lock 1245 * is Giant, then skip it. 1246 */ 1247 if ((lock->lo_flags & LO_SLEEPABLE) != 0 && 1248 lock1->li_lock == &Giant.lock_object) 1249 continue; 1250 1251 /* 1252 * If we are locking a sleepable lock and this lock 1253 * isn't sleepable, we want to treat it as a lock 1254 * order violation to enfore a general lock order of 1255 * sleepable locks before non-sleepable locks. 1256 */ 1257 if (((lock->lo_flags & LO_SLEEPABLE) != 0 && 1258 (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0)) 1259 goto reversal; 1260 1261 /* 1262 * If we are locking Giant and this is a non-sleepable 1263 * lock, then treat it as a reversal. 1264 */ 1265 if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0 && 1266 lock == &Giant.lock_object) 1267 goto reversal; 1268 1269 /* 1270 * Check the lock order hierarchy for a reveresal. 1271 */ 1272 if (!isitmydescendant(w, w1)) 1273 continue; 1274 reversal: 1275 1276 /* 1277 * We have a lock order violation, check to see if it 1278 * is allowed or has already been yelled about. 1279 */ 1280#ifdef BLESSING 1281 1282 /* 1283 * If the lock order is blessed, just bail. We don't 1284 * look for other lock order violations though, which 1285 * may be a bug. 1286 */ 1287 if (blessed(w, w1)) 1288 goto out; 1289#endif 1290 1291 /* Bail if this violation is known */ 1292 if (w_rmatrix[w1->w_index][w->w_index] & WITNESS_REVERSAL) 1293 goto out; 1294 1295 /* Record this as a violation */ 1296 w_rmatrix[w1->w_index][w->w_index] |= WITNESS_REVERSAL; 1297 w_rmatrix[w->w_index][w1->w_index] |= WITNESS_REVERSAL; 1298 w->w_reversed = w1->w_reversed = 1; 1299 witness_increment_graph_generation(); 1300 mtx_unlock_spin(&w_mtx); 1301 1302#ifdef WITNESS_NO_VNODE 1303 /* 1304 * There are known LORs between VNODE locks. They are 1305 * not an indication of a bug. VNODE locks are flagged 1306 * as such (LO_IS_VNODE) and we don't yell if the LOR 1307 * is between 2 VNODE locks. 1308 */ 1309 if ((lock->lo_flags & LO_IS_VNODE) != 0 && 1310 (lock1->li_lock->lo_flags & LO_IS_VNODE) != 0) 1311 return; 1312#endif 1313 1314 /* 1315 * Ok, yell about it. 1316 */ 1317 if (((lock->lo_flags & LO_SLEEPABLE) != 0 && 1318 (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0)) 1319 printf( 1320 "lock order reversal: (sleepable after non-sleepable)\n"); 1321 else if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0 1322 && lock == &Giant.lock_object) 1323 printf( 1324 "lock order reversal: (Giant after non-sleepable)\n"); 1325 else 1326 printf("lock order reversal:\n"); 1327 1328 /* 1329 * Try to locate an earlier lock with 1330 * witness w in our list. 1331 */ 1332 do { 1333 lock2 = &lle->ll_children[i]; 1334 MPASS(lock2->li_lock != NULL); 1335 if (lock2->li_lock->lo_witness == w) 1336 break; 1337 if (i == 0 && lle->ll_next != NULL) { 1338 lle = lle->ll_next; 1339 i = lle->ll_count - 1; 1340 MPASS(i >= 0 && i < LOCK_NCHILDREN); 1341 } else 1342 i--; 1343 } while (i >= 0); 1344 if (i < 0) { 1345 printf(" 1st %p %s (%s) @ %s:%d\n", 1346 lock1->li_lock, lock1->li_lock->lo_name, 1347 w1->w_name, fixup_filename(lock1->li_file), 1348 lock1->li_line); 1349 printf(" 2nd %p %s (%s) @ %s:%d\n", lock, 1350 lock->lo_name, w->w_name, 1351 fixup_filename(file), line); 1352 } else { 1353 printf(" 1st %p %s (%s) @ %s:%d\n", 1354 lock2->li_lock, lock2->li_lock->lo_name, 1355 lock2->li_lock->lo_witness->w_name, 1356 fixup_filename(lock2->li_file), 1357 lock2->li_line); 1358 printf(" 2nd %p %s (%s) @ %s:%d\n", 1359 lock1->li_lock, lock1->li_lock->lo_name, 1360 w1->w_name, fixup_filename(lock1->li_file), 1361 lock1->li_line); 1362 printf(" 3rd %p %s (%s) @ %s:%d\n", lock, 1363 lock->lo_name, w->w_name, 1364 fixup_filename(file), line); 1365 } 1366 witness_debugger(1); 1367 return; 1368 } 1369 } 1370 1371 /* 1372 * If requested, build a new lock order. However, don't build a new 1373 * relationship between a sleepable lock and Giant if it is in the 1374 * wrong direction. The correct lock order is that sleepable locks 1375 * always come before Giant. 1376 */ 1377 if (flags & LOP_NEWORDER && 1378 !(plock->li_lock == &Giant.lock_object && 1379 (lock->lo_flags & LO_SLEEPABLE) != 0)) { 1380 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__, 1381 w->w_name, plock->li_lock->lo_witness->w_name); 1382 itismychild(plock->li_lock->lo_witness, w); 1383 } 1384out: 1385 mtx_unlock_spin(&w_mtx); 1386} 1387 1388void 1389witness_lock(struct lock_object *lock, int flags, const char *file, int line) 1390{ 1391 struct lock_list_entry **lock_list, *lle; 1392 struct lock_instance *instance; 1393 struct witness *w; 1394 struct thread *td; 1395 1396 if (witness_cold || witness_watch == -1 || lock->lo_witness == NULL || 1397 panicstr != NULL) 1398 return; 1399 w = lock->lo_witness; 1400 td = curthread; 1401 1402 /* Determine lock list for this lock. */ 1403 if (LOCK_CLASS(lock)->lc_flags & LC_SLEEPLOCK) 1404 lock_list = &td->td_sleeplocks; 1405 else 1406 lock_list = PCPU_PTR(spinlocks); 1407 1408 /* Check to see if we are recursing on a lock we already own. */ 1409 instance = find_instance(*lock_list, lock); 1410 if (instance != NULL) { 1411 instance->li_flags++; 1412 CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__, 1413 td->td_proc->p_pid, lock->lo_name, 1414 instance->li_flags & LI_RECURSEMASK); 1415 instance->li_file = file; 1416 instance->li_line = line; 1417 return; 1418 } 1419 1420 /* Update per-witness last file and line acquire. */ 1421 w->w_file = file; 1422 w->w_line = line; 1423 1424 /* Find the next open lock instance in the list and fill it. */ 1425 lle = *lock_list; 1426 if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) { 1427 lle = witness_lock_list_get(); 1428 if (lle == NULL) 1429 return; 1430 lle->ll_next = *lock_list; 1431 CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__, 1432 td->td_proc->p_pid, lle); 1433 *lock_list = lle; 1434 } 1435 instance = &lle->ll_children[lle->ll_count++]; 1436 instance->li_lock = lock; 1437 instance->li_line = line; 1438 instance->li_file = file; 1439 if ((flags & LOP_EXCLUSIVE) != 0) 1440 instance->li_flags = LI_EXCLUSIVE; 1441 else 1442 instance->li_flags = 0; 1443 CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__, 1444 td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1); 1445} 1446 1447void 1448witness_upgrade(struct lock_object *lock, int flags, const char *file, int line) 1449{ 1450 struct lock_instance *instance; 1451 struct lock_class *class; 1452 1453 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__)); 1454 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL) 1455 return; 1456 class = LOCK_CLASS(lock); 1457 if (witness_watch) { 1458 if ((lock->lo_flags & LO_UPGRADABLE) == 0) 1459 kassert_panic( 1460 "upgrade of non-upgradable lock (%s) %s @ %s:%d", 1461 class->lc_name, lock->lo_name, 1462 fixup_filename(file), line); 1463 if ((class->lc_flags & LC_SLEEPLOCK) == 0) 1464 kassert_panic( 1465 "upgrade of non-sleep lock (%s) %s @ %s:%d", 1466 class->lc_name, lock->lo_name, 1467 fixup_filename(file), line); 1468 } 1469 instance = find_instance(curthread->td_sleeplocks, lock); 1470 if (instance == NULL) { 1471 kassert_panic("upgrade of unlocked lock (%s) %s @ %s:%d", 1472 class->lc_name, lock->lo_name, 1473 fixup_filename(file), line); 1474 return; 1475 } 1476 if (witness_watch) { 1477 if ((instance->li_flags & LI_EXCLUSIVE) != 0) 1478 kassert_panic( 1479 "upgrade of exclusive lock (%s) %s @ %s:%d", 1480 class->lc_name, lock->lo_name, 1481 fixup_filename(file), line); 1482 if ((instance->li_flags & LI_RECURSEMASK) != 0) 1483 kassert_panic( 1484 "upgrade of recursed lock (%s) %s r=%d @ %s:%d", 1485 class->lc_name, lock->lo_name, 1486 instance->li_flags & LI_RECURSEMASK, 1487 fixup_filename(file), line); 1488 } 1489 instance->li_flags |= LI_EXCLUSIVE; 1490} 1491 1492void 1493witness_downgrade(struct lock_object *lock, int flags, const char *file, 1494 int line) 1495{ 1496 struct lock_instance *instance; 1497 struct lock_class *class; 1498 1499 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__)); 1500 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL) 1501 return; 1502 class = LOCK_CLASS(lock); 1503 if (witness_watch) { 1504 if ((lock->lo_flags & LO_UPGRADABLE) == 0) 1505 kassert_panic( 1506 "downgrade of non-upgradable lock (%s) %s @ %s:%d", 1507 class->lc_name, lock->lo_name, 1508 fixup_filename(file), line); 1509 if ((class->lc_flags & LC_SLEEPLOCK) == 0) 1510 kassert_panic( 1511 "downgrade of non-sleep lock (%s) %s @ %s:%d", 1512 class->lc_name, lock->lo_name, 1513 fixup_filename(file), line); 1514 } 1515 instance = find_instance(curthread->td_sleeplocks, lock); 1516 if (instance == NULL) { 1517 kassert_panic("downgrade of unlocked lock (%s) %s @ %s:%d", 1518 class->lc_name, lock->lo_name, 1519 fixup_filename(file), line); 1520 return; 1521 } 1522 if (witness_watch) { 1523 if ((instance->li_flags & LI_EXCLUSIVE) == 0) 1524 kassert_panic( 1525 "downgrade of shared lock (%s) %s @ %s:%d", 1526 class->lc_name, lock->lo_name, 1527 fixup_filename(file), line); 1528 if ((instance->li_flags & LI_RECURSEMASK) != 0) 1529 kassert_panic( 1530 "downgrade of recursed lock (%s) %s r=%d @ %s:%d", 1531 class->lc_name, lock->lo_name, 1532 instance->li_flags & LI_RECURSEMASK, 1533 fixup_filename(file), line); 1534 } 1535 instance->li_flags &= ~LI_EXCLUSIVE; 1536} 1537 1538void 1539witness_unlock(struct lock_object *lock, int flags, const char *file, int line) 1540{ 1541 struct lock_list_entry **lock_list, *lle; 1542 struct lock_instance *instance; 1543 struct lock_class *class; 1544 struct thread *td; 1545 register_t s; 1546 int i, j; 1547 1548 if (witness_cold || lock->lo_witness == NULL || panicstr != NULL) 1549 return; 1550 td = curthread; 1551 class = LOCK_CLASS(lock); 1552 1553 /* Find lock instance associated with this lock. */ 1554 if (class->lc_flags & LC_SLEEPLOCK) 1555 lock_list = &td->td_sleeplocks; 1556 else 1557 lock_list = PCPU_PTR(spinlocks); 1558 lle = *lock_list; 1559 for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next) 1560 for (i = 0; i < (*lock_list)->ll_count; i++) { 1561 instance = &(*lock_list)->ll_children[i]; 1562 if (instance->li_lock == lock) 1563 goto found; 1564 } 1565 1566 /* 1567 * When disabling WITNESS through witness_watch we could end up in 1568 * having registered locks in the td_sleeplocks queue. 1569 * We have to make sure we flush these queues, so just search for 1570 * eventual register locks and remove them. 1571 */ 1572 if (witness_watch > 0) { 1573 kassert_panic("lock (%s) %s not locked @ %s:%d", class->lc_name, 1574 lock->lo_name, fixup_filename(file), line); 1575 return; 1576 } else { 1577 return; 1578 } 1579found: 1580 1581 /* First, check for shared/exclusive mismatches. */ 1582 if ((instance->li_flags & LI_EXCLUSIVE) != 0 && witness_watch > 0 && 1583 (flags & LOP_EXCLUSIVE) == 0) { 1584 printf("shared unlock of (%s) %s @ %s:%d\n", class->lc_name, 1585 lock->lo_name, fixup_filename(file), line); 1586 printf("while exclusively locked from %s:%d\n", 1587 fixup_filename(instance->li_file), instance->li_line); 1588 kassert_panic("excl->ushare"); 1589 } 1590 if ((instance->li_flags & LI_EXCLUSIVE) == 0 && witness_watch > 0 && 1591 (flags & LOP_EXCLUSIVE) != 0) { 1592 printf("exclusive unlock of (%s) %s @ %s:%d\n", class->lc_name, 1593 lock->lo_name, fixup_filename(file), line); 1594 printf("while share locked from %s:%d\n", 1595 fixup_filename(instance->li_file), 1596 instance->li_line); 1597 kassert_panic("share->uexcl"); 1598 } 1599 /* If we are recursed, unrecurse. */ 1600 if ((instance->li_flags & LI_RECURSEMASK) > 0) { 1601 CTR4(KTR_WITNESS, "%s: pid %d unrecursed on %s r=%d", __func__, 1602 td->td_proc->p_pid, instance->li_lock->lo_name, 1603 instance->li_flags); 1604 instance->li_flags--; 1605 return; 1606 } 1607 /* The lock is now being dropped, check for NORELEASE flag */ 1608 if ((instance->li_flags & LI_NORELEASE) != 0 && witness_watch > 0) { 1609 printf("forbidden unlock of (%s) %s @ %s:%d\n", class->lc_name, 1610 lock->lo_name, fixup_filename(file), line); 1611 kassert_panic("lock marked norelease"); 1612 } 1613 1614 /* Otherwise, remove this item from the list. */ 1615 s = intr_disable(); 1616 CTR4(KTR_WITNESS, "%s: pid %d removed %s from lle[%d]", __func__, 1617 td->td_proc->p_pid, instance->li_lock->lo_name, 1618 (*lock_list)->ll_count - 1); 1619 for (j = i; j < (*lock_list)->ll_count - 1; j++) 1620 (*lock_list)->ll_children[j] = 1621 (*lock_list)->ll_children[j + 1]; 1622 (*lock_list)->ll_count--; 1623 intr_restore(s); 1624 1625 /* 1626 * In order to reduce contention on w_mtx, we want to keep always an 1627 * head object into lists so that frequent allocation from the 1628 * free witness pool (and subsequent locking) is avoided. 1629 * In order to maintain the current code simple, when the head 1630 * object is totally unloaded it means also that we do not have 1631 * further objects in the list, so the list ownership needs to be 1632 * hand over to another object if the current head needs to be freed. 1633 */ 1634 if ((*lock_list)->ll_count == 0) { 1635 if (*lock_list == lle) { 1636 if (lle->ll_next == NULL) 1637 return; 1638 } else 1639 lle = *lock_list; 1640 *lock_list = lle->ll_next; 1641 CTR3(KTR_WITNESS, "%s: pid %d removed lle %p", __func__, 1642 td->td_proc->p_pid, lle); 1643 witness_lock_list_free(lle); 1644 } 1645} 1646 1647void 1648witness_thread_exit(struct thread *td) 1649{ 1650 struct lock_list_entry *lle; 1651 int i, n; 1652 1653 lle = td->td_sleeplocks; 1654 if (lle == NULL || panicstr != NULL) 1655 return; 1656 if (lle->ll_count != 0) { 1657 for (n = 0; lle != NULL; lle = lle->ll_next) 1658 for (i = lle->ll_count - 1; i >= 0; i--) { 1659 if (n == 0) 1660 printf("Thread %p exiting with the following locks held:\n", 1661 td); 1662 n++; 1663 witness_list_lock(&lle->ll_children[i], printf); 1664 1665 } 1666 kassert_panic( 1667 "Thread %p cannot exit while holding sleeplocks\n", td); 1668 } 1669 witness_lock_list_free(lle); 1670} 1671 1672/* 1673 * Warn if any locks other than 'lock' are held. Flags can be passed in to 1674 * exempt Giant and sleepable locks from the checks as well. If any 1675 * non-exempt locks are held, then a supplied message is printed to the 1676 * console along with a list of the offending locks. If indicated in the 1677 * flags then a failure results in a panic as well. 1678 */ 1679int 1680witness_warn(int flags, struct lock_object *lock, const char *fmt, ...) 1681{ 1682 struct lock_list_entry *lock_list, *lle; 1683 struct lock_instance *lock1; 1684 struct thread *td; 1685 va_list ap; 1686 int i, n; 1687 1688 if (witness_cold || witness_watch < 1 || panicstr != NULL) 1689 return (0); 1690 n = 0; 1691 td = curthread; 1692 for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next) 1693 for (i = lle->ll_count - 1; i >= 0; i--) { 1694 lock1 = &lle->ll_children[i]; 1695 if (lock1->li_lock == lock) 1696 continue; 1697 if (flags & WARN_GIANTOK && 1698 lock1->li_lock == &Giant.lock_object) 1699 continue; 1700 if (flags & WARN_SLEEPOK && 1701 (lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0) 1702 continue; 1703 if (n == 0) { 1704 va_start(ap, fmt); 1705 vprintf(fmt, ap); 1706 va_end(ap); 1707 printf(" with the following"); 1708 if (flags & WARN_SLEEPOK) 1709 printf(" non-sleepable"); 1710 printf(" locks held:\n"); 1711 } 1712 n++; 1713 witness_list_lock(lock1, printf); 1714 } 1715 1716 /* 1717 * Pin the thread in order to avoid problems with thread migration. 1718 * Once that all verifies are passed about spinlocks ownership, 1719 * the thread is in a safe path and it can be unpinned. 1720 */ 1721 sched_pin(); 1722 lock_list = PCPU_GET(spinlocks); 1723 if (lock_list != NULL && lock_list->ll_count != 0) { 1724 sched_unpin(); 1725 1726 /* 1727 * We should only have one spinlock and as long as 1728 * the flags cannot match for this locks class, 1729 * check if the first spinlock is the one curthread 1730 * should hold. 1731 */ 1732 lock1 = &lock_list->ll_children[lock_list->ll_count - 1]; 1733 if (lock_list->ll_count == 1 && lock_list->ll_next == NULL && 1734 lock1->li_lock == lock && n == 0) 1735 return (0); 1736 1737 va_start(ap, fmt); 1738 vprintf(fmt, ap); 1739 va_end(ap); 1740 printf(" with the following"); 1741 if (flags & WARN_SLEEPOK) 1742 printf(" non-sleepable"); 1743 printf(" locks held:\n"); 1744 n += witness_list_locks(&lock_list, printf); 1745 } else 1746 sched_unpin(); 1747 if (flags & WARN_PANIC && n) 1748 kassert_panic("%s", __func__); 1749 else 1750 witness_debugger(n); 1751 return (n); 1752} 1753 1754const char * 1755witness_file(struct lock_object *lock) 1756{ 1757 struct witness *w; 1758 1759 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL) 1760 return ("?"); 1761 w = lock->lo_witness; 1762 return (w->w_file); 1763} 1764 1765int 1766witness_line(struct lock_object *lock) 1767{ 1768 struct witness *w; 1769 1770 if (witness_cold || witness_watch < 1 || lock->lo_witness == NULL) 1771 return (0); 1772 w = lock->lo_witness; 1773 return (w->w_line); 1774} 1775 1776static struct witness * 1777enroll(const char *description, struct lock_class *lock_class) 1778{ 1779 struct witness *w; 1780 struct witness_list *typelist; 1781 1782 MPASS(description != NULL); 1783 1784 if (witness_watch == -1 || panicstr != NULL) 1785 return (NULL); 1786 if ((lock_class->lc_flags & LC_SPINLOCK)) { 1787 if (witness_skipspin) 1788 return (NULL); 1789 else 1790 typelist = &w_spin; 1791 } else if ((lock_class->lc_flags & LC_SLEEPLOCK)) { 1792 typelist = &w_sleep; 1793 } else { 1794 kassert_panic("lock class %s is not sleep or spin", 1795 lock_class->lc_name); 1796 return (NULL); 1797 } 1798 1799 mtx_lock_spin(&w_mtx); 1800 w = witness_hash_get(description); 1801 if (w) 1802 goto found; 1803 if ((w = witness_get()) == NULL) 1804 return (NULL); 1805 MPASS(strlen(description) < MAX_W_NAME); 1806 strcpy(w->w_name, description); 1807 w->w_class = lock_class; 1808 w->w_refcount = 1; 1809 STAILQ_INSERT_HEAD(&w_all, w, w_list); 1810 if (lock_class->lc_flags & LC_SPINLOCK) { 1811 STAILQ_INSERT_HEAD(&w_spin, w, w_typelist); 1812 w_spin_cnt++; 1813 } else if (lock_class->lc_flags & LC_SLEEPLOCK) { 1814 STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist); 1815 w_sleep_cnt++; 1816 } 1817 1818 /* Insert new witness into the hash */ 1819 witness_hash_put(w); 1820 witness_increment_graph_generation(); 1821 mtx_unlock_spin(&w_mtx); 1822 return (w); 1823found: 1824 w->w_refcount++; 1825 mtx_unlock_spin(&w_mtx); 1826 if (lock_class != w->w_class) 1827 kassert_panic( 1828 "lock (%s) %s does not match earlier (%s) lock", 1829 description, lock_class->lc_name, 1830 w->w_class->lc_name); 1831 return (w); 1832} 1833 1834static void 1835depart(struct witness *w) 1836{ 1837 struct witness_list *list; 1838 1839 MPASS(w->w_refcount == 0); 1840 if (w->w_class->lc_flags & LC_SLEEPLOCK) { 1841 list = &w_sleep; 1842 w_sleep_cnt--; 1843 } else { 1844 list = &w_spin; 1845 w_spin_cnt--; 1846 } 1847 /* 1848 * Set file to NULL as it may point into a loadable module. 1849 */ 1850 w->w_file = NULL; 1851 w->w_line = 0; 1852 witness_increment_graph_generation(); 1853} 1854 1855 1856static void 1857adopt(struct witness *parent, struct witness *child) 1858{ 1859 int pi, ci, i, j; 1860 1861 if (witness_cold == 0) 1862 mtx_assert(&w_mtx, MA_OWNED); 1863 1864 /* If the relationship is already known, there's no work to be done. */ 1865 if (isitmychild(parent, child)) 1866 return; 1867 1868 /* When the structure of the graph changes, bump up the generation. */ 1869 witness_increment_graph_generation(); 1870 1871 /* 1872 * The hard part ... create the direct relationship, then propagate all 1873 * indirect relationships. 1874 */ 1875 pi = parent->w_index; 1876 ci = child->w_index; 1877 WITNESS_INDEX_ASSERT(pi); 1878 WITNESS_INDEX_ASSERT(ci); 1879 MPASS(pi != ci); 1880 w_rmatrix[pi][ci] |= WITNESS_PARENT; 1881 w_rmatrix[ci][pi] |= WITNESS_CHILD; 1882 1883 /* 1884 * If parent was not already an ancestor of child, 1885 * then we increment the descendant and ancestor counters. 1886 */ 1887 if ((w_rmatrix[pi][ci] & WITNESS_ANCESTOR) == 0) { 1888 parent->w_num_descendants++; 1889 child->w_num_ancestors++; 1890 } 1891 1892 /* 1893 * Find each ancestor of 'pi'. Note that 'pi' itself is counted as 1894 * an ancestor of 'pi' during this loop. 1895 */ 1896 for (i = 1; i <= w_max_used_index; i++) { 1897 if ((w_rmatrix[i][pi] & WITNESS_ANCESTOR_MASK) == 0 && 1898 (i != pi)) 1899 continue; 1900 1901 /* Find each descendant of 'i' and mark it as a descendant. */ 1902 for (j = 1; j <= w_max_used_index; j++) { 1903 1904 /* 1905 * Skip children that are already marked as 1906 * descendants of 'i'. 1907 */ 1908 if (w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK) 1909 continue; 1910 1911 /* 1912 * We are only interested in descendants of 'ci'. Note 1913 * that 'ci' itself is counted as a descendant of 'ci'. 1914 */ 1915 if ((w_rmatrix[ci][j] & WITNESS_ANCESTOR_MASK) == 0 && 1916 (j != ci)) 1917 continue; 1918 w_rmatrix[i][j] |= WITNESS_ANCESTOR; 1919 w_rmatrix[j][i] |= WITNESS_DESCENDANT; 1920 w_data[i].w_num_descendants++; 1921 w_data[j].w_num_ancestors++; 1922 1923 /* 1924 * Make sure we aren't marking a node as both an 1925 * ancestor and descendant. We should have caught 1926 * this as a lock order reversal earlier. 1927 */ 1928 if ((w_rmatrix[i][j] & WITNESS_ANCESTOR_MASK) && 1929 (w_rmatrix[i][j] & WITNESS_DESCENDANT_MASK)) { 1930 printf("witness rmatrix paradox! [%d][%d]=%d " 1931 "both ancestor and descendant\n", 1932 i, j, w_rmatrix[i][j]); 1933 kdb_backtrace(); 1934 printf("Witness disabled.\n"); 1935 witness_watch = -1; 1936 } 1937 if ((w_rmatrix[j][i] & WITNESS_ANCESTOR_MASK) && 1938 (w_rmatrix[j][i] & WITNESS_DESCENDANT_MASK)) { 1939 printf("witness rmatrix paradox! [%d][%d]=%d " 1940 "both ancestor and descendant\n", 1941 j, i, w_rmatrix[j][i]); 1942 kdb_backtrace(); 1943 printf("Witness disabled.\n"); 1944 witness_watch = -1; 1945 } 1946 } 1947 } 1948} 1949 1950static void 1951itismychild(struct witness *parent, struct witness *child) 1952{ 1953 int unlocked; 1954 1955 MPASS(child != NULL && parent != NULL); 1956 if (witness_cold == 0) 1957 mtx_assert(&w_mtx, MA_OWNED); 1958 1959 if (!witness_lock_type_equal(parent, child)) { 1960 if (witness_cold == 0) { 1961 unlocked = 1; 1962 mtx_unlock_spin(&w_mtx); 1963 } else { 1964 unlocked = 0; 1965 } 1966 kassert_panic( 1967 "%s: parent \"%s\" (%s) and child \"%s\" (%s) are not " 1968 "the same lock type", __func__, parent->w_name, 1969 parent->w_class->lc_name, child->w_name, 1970 child->w_class->lc_name); 1971 if (unlocked) 1972 mtx_lock_spin(&w_mtx); 1973 } 1974 adopt(parent, child); 1975} 1976 1977/* 1978 * Generic code for the isitmy*() functions. The rmask parameter is the 1979 * expected relationship of w1 to w2. 1980 */ 1981static int 1982_isitmyx(struct witness *w1, struct witness *w2, int rmask, const char *fname) 1983{ 1984 unsigned char r1, r2; 1985 int i1, i2; 1986 1987 i1 = w1->w_index; 1988 i2 = w2->w_index; 1989 WITNESS_INDEX_ASSERT(i1); 1990 WITNESS_INDEX_ASSERT(i2); 1991 r1 = w_rmatrix[i1][i2] & WITNESS_RELATED_MASK; 1992 r2 = w_rmatrix[i2][i1] & WITNESS_RELATED_MASK; 1993 1994 /* The flags on one better be the inverse of the flags on the other */ 1995 if (!((WITNESS_ATOD(r1) == r2 && WITNESS_DTOA(r2) == r1) || 1996 (WITNESS_DTOA(r1) == r2 && WITNESS_ATOD(r2) == r1))) { 1997 printf("%s: rmatrix mismatch between %s (index %d) and %s " 1998 "(index %d): w_rmatrix[%d][%d] == %hhx but " 1999 "w_rmatrix[%d][%d] == %hhx\n", 2000 fname, w1->w_name, i1, w2->w_name, i2, i1, i2, r1, 2001 i2, i1, r2); 2002 kdb_backtrace(); 2003 printf("Witness disabled.\n"); 2004 witness_watch = -1; 2005 } 2006 return (r1 & rmask); 2007} 2008 2009/* 2010 * Checks if @child is a direct child of @parent. 2011 */ 2012static int 2013isitmychild(struct witness *parent, struct witness *child) 2014{ 2015 2016 return (_isitmyx(parent, child, WITNESS_PARENT, __func__)); 2017} 2018 2019/* 2020 * Checks if @descendant is a direct or inderect descendant of @ancestor. 2021 */ 2022static int 2023isitmydescendant(struct witness *ancestor, struct witness *descendant) 2024{ 2025 2026 return (_isitmyx(ancestor, descendant, WITNESS_ANCESTOR_MASK, 2027 __func__)); 2028} 2029 2030#ifdef BLESSING 2031static int 2032blessed(struct witness *w1, struct witness *w2) 2033{ 2034 int i; 2035 struct witness_blessed *b; 2036 2037 for (i = 0; i < blessed_count; i++) { 2038 b = &blessed_list[i]; 2039 if (strcmp(w1->w_name, b->b_lock1) == 0) { 2040 if (strcmp(w2->w_name, b->b_lock2) == 0) 2041 return (1); 2042 continue; 2043 } 2044 if (strcmp(w1->w_name, b->b_lock2) == 0) 2045 if (strcmp(w2->w_name, b->b_lock1) == 0) 2046 return (1); 2047 } 2048 return (0); 2049} 2050#endif 2051 2052static struct witness * 2053witness_get(void) 2054{ 2055 struct witness *w; 2056 int index; 2057 2058 if (witness_cold == 0) 2059 mtx_assert(&w_mtx, MA_OWNED); 2060 2061 if (witness_watch == -1) { 2062 mtx_unlock_spin(&w_mtx); 2063 return (NULL); 2064 } 2065 if (STAILQ_EMPTY(&w_free)) { 2066 witness_watch = -1; 2067 mtx_unlock_spin(&w_mtx); 2068 printf("WITNESS: unable to allocate a new witness object\n"); 2069 return (NULL); 2070 } 2071 w = STAILQ_FIRST(&w_free); 2072 STAILQ_REMOVE_HEAD(&w_free, w_list); 2073 w_free_cnt--; 2074 index = w->w_index; 2075 MPASS(index > 0 && index == w_max_used_index+1 && 2076 index < witness_count); 2077 bzero(w, sizeof(*w)); 2078 w->w_index = index; 2079 if (index > w_max_used_index) 2080 w_max_used_index = index; 2081 return (w); 2082} 2083 2084static void 2085witness_free(struct witness *w) 2086{ 2087 2088 STAILQ_INSERT_HEAD(&w_free, w, w_list); 2089 w_free_cnt++; 2090} 2091 2092static struct lock_list_entry * 2093witness_lock_list_get(void) 2094{ 2095 struct lock_list_entry *lle; 2096 2097 if (witness_watch == -1) 2098 return (NULL); 2099 mtx_lock_spin(&w_mtx); 2100 lle = w_lock_list_free; 2101 if (lle == NULL) { 2102 witness_watch = -1; 2103 mtx_unlock_spin(&w_mtx); 2104 printf("%s: witness exhausted\n", __func__); 2105 return (NULL); 2106 } 2107 w_lock_list_free = lle->ll_next; 2108 mtx_unlock_spin(&w_mtx); 2109 bzero(lle, sizeof(*lle)); 2110 return (lle); 2111} 2112 2113static void 2114witness_lock_list_free(struct lock_list_entry *lle) 2115{ 2116 2117 mtx_lock_spin(&w_mtx); 2118 lle->ll_next = w_lock_list_free; 2119 w_lock_list_free = lle; 2120 mtx_unlock_spin(&w_mtx); 2121} 2122 2123static struct lock_instance * 2124find_instance(struct lock_list_entry *list, const struct lock_object *lock) 2125{ 2126 struct lock_list_entry *lle; 2127 struct lock_instance *instance; 2128 int i; 2129 2130 for (lle = list; lle != NULL; lle = lle->ll_next) 2131 for (i = lle->ll_count - 1; i >= 0; i--) { 2132 instance = &lle->ll_children[i]; 2133 if (instance->li_lock == lock) 2134 return (instance); 2135 } 2136 return (NULL); 2137} 2138 2139static void 2140witness_list_lock(struct lock_instance *instance, 2141 int (*prnt)(const char *fmt, ...)) 2142{ 2143 struct lock_object *lock; 2144 2145 lock = instance->li_lock; 2146 prnt("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ? 2147 "exclusive" : "shared", LOCK_CLASS(lock)->lc_name, lock->lo_name); 2148 if (lock->lo_witness->w_name != lock->lo_name) 2149 prnt(" (%s)", lock->lo_witness->w_name); 2150 prnt(" r = %d (%p) locked @ %s:%d\n", 2151 instance->li_flags & LI_RECURSEMASK, lock, 2152 fixup_filename(instance->li_file), instance->li_line); 2153} 2154 2155#ifdef DDB 2156static int 2157witness_thread_has_locks(struct thread *td) 2158{ 2159 2160 if (td->td_sleeplocks == NULL) 2161 return (0); 2162 return (td->td_sleeplocks->ll_count != 0); 2163} 2164 2165static int 2166witness_proc_has_locks(struct proc *p) 2167{ 2168 struct thread *td; 2169 2170 FOREACH_THREAD_IN_PROC(p, td) { 2171 if (witness_thread_has_locks(td)) 2172 return (1); 2173 } 2174 return (0); 2175} 2176#endif 2177 2178int 2179witness_list_locks(struct lock_list_entry **lock_list, 2180 int (*prnt)(const char *fmt, ...)) 2181{ 2182 struct lock_list_entry *lle; 2183 int i, nheld; 2184 2185 nheld = 0; 2186 for (lle = *lock_list; lle != NULL; lle = lle->ll_next) 2187 for (i = lle->ll_count - 1; i >= 0; i--) { 2188 witness_list_lock(&lle->ll_children[i], prnt); 2189 nheld++; 2190 } 2191 return (nheld); 2192} 2193 2194/* 2195 * This is a bit risky at best. We call this function when we have timed 2196 * out acquiring a spin lock, and we assume that the other CPU is stuck 2197 * with this lock held. So, we go groveling around in the other CPU's 2198 * per-cpu data to try to find the lock instance for this spin lock to 2199 * see when it was last acquired. 2200 */ 2201void 2202witness_display_spinlock(struct lock_object *lock, struct thread *owner, 2203 int (*prnt)(const char *fmt, ...)) 2204{ 2205 struct lock_instance *instance; 2206 struct pcpu *pc; 2207 2208 if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU) 2209 return; 2210 pc = pcpu_find(owner->td_oncpu); 2211 instance = find_instance(pc->pc_spinlocks, lock); 2212 if (instance != NULL) 2213 witness_list_lock(instance, prnt); 2214} 2215 2216void 2217witness_save(struct lock_object *lock, const char **filep, int *linep) 2218{ 2219 struct lock_list_entry *lock_list; 2220 struct lock_instance *instance; 2221 struct lock_class *class; 2222 2223 /* 2224 * This function is used independently in locking code to deal with 2225 * Giant, SCHEDULER_STOPPED() check can be removed here after Giant 2226 * is gone. 2227 */ 2228 if (SCHEDULER_STOPPED()) 2229 return; 2230 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__)); 2231 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL) 2232 return; 2233 class = LOCK_CLASS(lock); 2234 if (class->lc_flags & LC_SLEEPLOCK) 2235 lock_list = curthread->td_sleeplocks; 2236 else { 2237 if (witness_skipspin) 2238 return; 2239 lock_list = PCPU_GET(spinlocks); 2240 } 2241 instance = find_instance(lock_list, lock); 2242 if (instance == NULL) { 2243 kassert_panic("%s: lock (%s) %s not locked", __func__, 2244 class->lc_name, lock->lo_name); 2245 return; 2246 } 2247 *filep = instance->li_file; 2248 *linep = instance->li_line; 2249} 2250 2251void 2252witness_restore(struct lock_object *lock, const char *file, int line) 2253{ 2254 struct lock_list_entry *lock_list; 2255 struct lock_instance *instance; 2256 struct lock_class *class; 2257 2258 /* 2259 * This function is used independently in locking code to deal with 2260 * Giant, SCHEDULER_STOPPED() check can be removed here after Giant 2261 * is gone. 2262 */ 2263 if (SCHEDULER_STOPPED()) 2264 return; 2265 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__)); 2266 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL) 2267 return; 2268 class = LOCK_CLASS(lock); 2269 if (class->lc_flags & LC_SLEEPLOCK) 2270 lock_list = curthread->td_sleeplocks; 2271 else { 2272 if (witness_skipspin) 2273 return; 2274 lock_list = PCPU_GET(spinlocks); 2275 } 2276 instance = find_instance(lock_list, lock); 2277 if (instance == NULL) 2278 kassert_panic("%s: lock (%s) %s not locked", __func__, 2279 class->lc_name, lock->lo_name); 2280 lock->lo_witness->w_file = file; 2281 lock->lo_witness->w_line = line; 2282 if (instance == NULL) 2283 return; 2284 instance->li_file = file; 2285 instance->li_line = line; 2286} 2287 2288void 2289witness_assert(const struct lock_object *lock, int flags, const char *file, 2290 int line) 2291{ 2292#ifdef INVARIANT_SUPPORT 2293 struct lock_instance *instance; 2294 struct lock_class *class; 2295 2296 if (lock->lo_witness == NULL || witness_watch < 1 || panicstr != NULL) 2297 return; 2298 class = LOCK_CLASS(lock); 2299 if ((class->lc_flags & LC_SLEEPLOCK) != 0) 2300 instance = find_instance(curthread->td_sleeplocks, lock); 2301 else if ((class->lc_flags & LC_SPINLOCK) != 0) 2302 instance = find_instance(PCPU_GET(spinlocks), lock); 2303 else { 2304 kassert_panic("Lock (%s) %s is not sleep or spin!", 2305 class->lc_name, lock->lo_name); 2306 return; 2307 } 2308 switch (flags) { 2309 case LA_UNLOCKED: 2310 if (instance != NULL) 2311 kassert_panic("Lock (%s) %s locked @ %s:%d.", 2312 class->lc_name, lock->lo_name, 2313 fixup_filename(file), line); 2314 break; 2315 case LA_LOCKED: 2316 case LA_LOCKED | LA_RECURSED: 2317 case LA_LOCKED | LA_NOTRECURSED: 2318 case LA_SLOCKED: 2319 case LA_SLOCKED | LA_RECURSED: 2320 case LA_SLOCKED | LA_NOTRECURSED: 2321 case LA_XLOCKED: 2322 case LA_XLOCKED | LA_RECURSED: 2323 case LA_XLOCKED | LA_NOTRECURSED: 2324 if (instance == NULL) { 2325 kassert_panic("Lock (%s) %s not locked @ %s:%d.", 2326 class->lc_name, lock->lo_name, 2327 fixup_filename(file), line); 2328 break; 2329 } 2330 if ((flags & LA_XLOCKED) != 0 && 2331 (instance->li_flags & LI_EXCLUSIVE) == 0) 2332 kassert_panic( 2333 "Lock (%s) %s not exclusively locked @ %s:%d.", 2334 class->lc_name, lock->lo_name, 2335 fixup_filename(file), line); 2336 if ((flags & LA_SLOCKED) != 0 && 2337 (instance->li_flags & LI_EXCLUSIVE) != 0) 2338 kassert_panic( 2339 "Lock (%s) %s exclusively locked @ %s:%d.", 2340 class->lc_name, lock->lo_name, 2341 fixup_filename(file), line); 2342 if ((flags & LA_RECURSED) != 0 && 2343 (instance->li_flags & LI_RECURSEMASK) == 0) 2344 kassert_panic("Lock (%s) %s not recursed @ %s:%d.", 2345 class->lc_name, lock->lo_name, 2346 fixup_filename(file), line); 2347 if ((flags & LA_NOTRECURSED) != 0 && 2348 (instance->li_flags & LI_RECURSEMASK) != 0) 2349 kassert_panic("Lock (%s) %s recursed @ %s:%d.", 2350 class->lc_name, lock->lo_name, 2351 fixup_filename(file), line); 2352 break; 2353 default: 2354 kassert_panic("Invalid lock assertion at %s:%d.", 2355 fixup_filename(file), line); 2356 2357 } 2358#endif /* INVARIANT_SUPPORT */ 2359} 2360 2361static void 2362witness_setflag(struct lock_object *lock, int flag, int set) 2363{ 2364 struct lock_list_entry *lock_list; 2365 struct lock_instance *instance; 2366 struct lock_class *class; 2367 2368 if (lock->lo_witness == NULL || witness_watch == -1 || panicstr != NULL) 2369 return; 2370 class = LOCK_CLASS(lock); 2371 if (class->lc_flags & LC_SLEEPLOCK) 2372 lock_list = curthread->td_sleeplocks; 2373 else { 2374 if (witness_skipspin) 2375 return; 2376 lock_list = PCPU_GET(spinlocks); 2377 } 2378 instance = find_instance(lock_list, lock); 2379 if (instance == NULL) { 2380 kassert_panic("%s: lock (%s) %s not locked", __func__, 2381 class->lc_name, lock->lo_name); 2382 return; 2383 } 2384 2385 if (set) 2386 instance->li_flags |= flag; 2387 else 2388 instance->li_flags &= ~flag; 2389} 2390 2391void 2392witness_norelease(struct lock_object *lock) 2393{ 2394 2395 witness_setflag(lock, LI_NORELEASE, 1); 2396} 2397 2398void 2399witness_releaseok(struct lock_object *lock) 2400{ 2401 2402 witness_setflag(lock, LI_NORELEASE, 0); 2403} 2404 2405#ifdef DDB 2406static void 2407witness_ddb_list(struct thread *td) 2408{ 2409 2410 KASSERT(witness_cold == 0, ("%s: witness_cold", __func__)); 2411 KASSERT(kdb_active, ("%s: not in the debugger", __func__)); 2412 2413 if (witness_watch < 1) 2414 return; 2415 2416 witness_list_locks(&td->td_sleeplocks, db_printf); 2417 2418 /* 2419 * We only handle spinlocks if td == curthread. This is somewhat broken 2420 * if td is currently executing on some other CPU and holds spin locks 2421 * as we won't display those locks. If we had a MI way of getting 2422 * the per-cpu data for a given cpu then we could use 2423 * td->td_oncpu to get the list of spinlocks for this thread 2424 * and "fix" this. 2425 * 2426 * That still wouldn't really fix this unless we locked the scheduler 2427 * lock or stopped the other CPU to make sure it wasn't changing the 2428 * list out from under us. It is probably best to just not try to 2429 * handle threads on other CPU's for now. 2430 */ 2431 if (td == curthread && PCPU_GET(spinlocks) != NULL) 2432 witness_list_locks(PCPU_PTR(spinlocks), db_printf); 2433} 2434 2435DB_SHOW_COMMAND(locks, db_witness_list) 2436{ 2437 struct thread *td; 2438 2439 if (have_addr) 2440 td = db_lookup_thread(addr, TRUE); 2441 else 2442 td = kdb_thread; 2443 witness_ddb_list(td); 2444} 2445 2446DB_SHOW_ALL_COMMAND(locks, db_witness_list_all) 2447{ 2448 struct thread *td; 2449 struct proc *p; 2450 2451 /* 2452 * It would be nice to list only threads and processes that actually 2453 * held sleep locks, but that information is currently not exported 2454 * by WITNESS. 2455 */ 2456 FOREACH_PROC_IN_SYSTEM(p) { 2457 if (!witness_proc_has_locks(p)) 2458 continue; 2459 FOREACH_THREAD_IN_PROC(p, td) { 2460 if (!witness_thread_has_locks(td)) 2461 continue; 2462 db_printf("Process %d (%s) thread %p (%d)\n", p->p_pid, 2463 p->p_comm, td, td->td_tid); 2464 witness_ddb_list(td); 2465 if (db_pager_quit) 2466 return; 2467 } 2468 } 2469} 2470DB_SHOW_ALIAS(alllocks, db_witness_list_all) 2471 2472DB_SHOW_COMMAND(witness, db_witness_display) 2473{ 2474 2475 witness_ddb_display(db_printf); 2476} 2477#endif 2478 2479static int 2480sysctl_debug_witness_badstacks(SYSCTL_HANDLER_ARGS) 2481{ 2482 struct witness_lock_order_data *data1, *data2, *tmp_data1, *tmp_data2; 2483 struct witness *tmp_w1, *tmp_w2, *w1, *w2; 2484 struct sbuf *sb; 2485 u_int w_rmatrix1, w_rmatrix2; 2486 int error, generation, i, j; 2487 2488 tmp_data1 = NULL; 2489 tmp_data2 = NULL; 2490 tmp_w1 = NULL; 2491 tmp_w2 = NULL; 2492 if (witness_watch < 1) { 2493 error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning)); 2494 return (error); 2495 } 2496 if (witness_cold) { 2497 error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold)); 2498 return (error); 2499 } 2500 error = 0; 2501 sb = sbuf_new(NULL, NULL, badstack_sbuf_size, SBUF_AUTOEXTEND); 2502 if (sb == NULL) 2503 return (ENOMEM); 2504 2505 /* Allocate and init temporary storage space. */ 2506 tmp_w1 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO); 2507 tmp_w2 = malloc(sizeof(struct witness), M_TEMP, M_WAITOK | M_ZERO); 2508 tmp_data1 = malloc(sizeof(struct witness_lock_order_data), M_TEMP, 2509 M_WAITOK | M_ZERO); 2510 tmp_data2 = malloc(sizeof(struct witness_lock_order_data), M_TEMP, 2511 M_WAITOK | M_ZERO); 2512 stack_zero(&tmp_data1->wlod_stack); 2513 stack_zero(&tmp_data2->wlod_stack); 2514 2515restart: 2516 mtx_lock_spin(&w_mtx); 2517 generation = w_generation; 2518 mtx_unlock_spin(&w_mtx); 2519 sbuf_printf(sb, "Number of known direct relationships is %d\n", 2520 w_lohash.wloh_count); 2521 for (i = 1; i < w_max_used_index; i++) { 2522 mtx_lock_spin(&w_mtx); 2523 if (generation != w_generation) { 2524 mtx_unlock_spin(&w_mtx); 2525 2526 /* The graph has changed, try again. */ 2527 req->oldidx = 0; 2528 sbuf_clear(sb); 2529 goto restart; 2530 } 2531 2532 w1 = &w_data[i]; 2533 if (w1->w_reversed == 0) { 2534 mtx_unlock_spin(&w_mtx); 2535 continue; 2536 } 2537 2538 /* Copy w1 locally so we can release the spin lock. */ 2539 *tmp_w1 = *w1; 2540 mtx_unlock_spin(&w_mtx); 2541 2542 if (tmp_w1->w_reversed == 0) 2543 continue; 2544 for (j = 1; j < w_max_used_index; j++) { 2545 if ((w_rmatrix[i][j] & WITNESS_REVERSAL) == 0 || i > j) 2546 continue; 2547 2548 mtx_lock_spin(&w_mtx); 2549 if (generation != w_generation) { 2550 mtx_unlock_spin(&w_mtx); 2551 2552 /* The graph has changed, try again. */ 2553 req->oldidx = 0; 2554 sbuf_clear(sb); 2555 goto restart; 2556 } 2557 2558 w2 = &w_data[j]; 2559 data1 = witness_lock_order_get(w1, w2); 2560 data2 = witness_lock_order_get(w2, w1); 2561 2562 /* 2563 * Copy information locally so we can release the 2564 * spin lock. 2565 */ 2566 *tmp_w2 = *w2; 2567 w_rmatrix1 = (unsigned int)w_rmatrix[i][j]; 2568 w_rmatrix2 = (unsigned int)w_rmatrix[j][i]; 2569 2570 if (data1) { 2571 stack_zero(&tmp_data1->wlod_stack); 2572 stack_copy(&data1->wlod_stack, 2573 &tmp_data1->wlod_stack); 2574 } 2575 if (data2 && data2 != data1) { 2576 stack_zero(&tmp_data2->wlod_stack); 2577 stack_copy(&data2->wlod_stack, 2578 &tmp_data2->wlod_stack); 2579 } 2580 mtx_unlock_spin(&w_mtx); 2581 2582 sbuf_printf(sb, 2583 "\nLock order reversal between \"%s\"(%s) and \"%s\"(%s)!\n", 2584 tmp_w1->w_name, tmp_w1->w_class->lc_name, 2585 tmp_w2->w_name, tmp_w2->w_class->lc_name); 2586#if 0 2587 sbuf_printf(sb, 2588 "w_rmatrix[%s][%s] == %x, w_rmatrix[%s][%s] == %x\n", 2589 tmp_w1->name, tmp_w2->w_name, w_rmatrix1, 2590 tmp_w2->name, tmp_w1->w_name, w_rmatrix2); 2591#endif 2592 if (data1) { 2593 sbuf_printf(sb, 2594 "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n", 2595 tmp_w1->w_name, tmp_w1->w_class->lc_name, 2596 tmp_w2->w_name, tmp_w2->w_class->lc_name); 2597 stack_sbuf_print(sb, &tmp_data1->wlod_stack); 2598 sbuf_printf(sb, "\n"); 2599 } 2600 if (data2 && data2 != data1) { 2601 sbuf_printf(sb, 2602 "Lock order \"%s\"(%s) -> \"%s\"(%s) first seen at:\n", 2603 tmp_w2->w_name, tmp_w2->w_class->lc_name, 2604 tmp_w1->w_name, tmp_w1->w_class->lc_name); 2605 stack_sbuf_print(sb, &tmp_data2->wlod_stack); 2606 sbuf_printf(sb, "\n"); 2607 } 2608 } 2609 } 2610 mtx_lock_spin(&w_mtx); 2611 if (generation != w_generation) { 2612 mtx_unlock_spin(&w_mtx); 2613 2614 /* 2615 * The graph changed while we were printing stack data, 2616 * try again. 2617 */ 2618 req->oldidx = 0; 2619 sbuf_clear(sb); 2620 goto restart; 2621 } 2622 mtx_unlock_spin(&w_mtx); 2623 2624 /* Free temporary storage space. */ 2625 free(tmp_data1, M_TEMP); 2626 free(tmp_data2, M_TEMP); 2627 free(tmp_w1, M_TEMP); 2628 free(tmp_w2, M_TEMP); 2629 2630 sbuf_finish(sb); 2631 error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1); 2632 sbuf_delete(sb); 2633 2634 return (error); 2635} 2636 2637static int 2638sysctl_debug_witness_fullgraph(SYSCTL_HANDLER_ARGS) 2639{ 2640 struct witness *w; 2641 struct sbuf *sb; 2642 int error; 2643 2644 if (witness_watch < 1) { 2645 error = SYSCTL_OUT(req, w_notrunning, sizeof(w_notrunning)); 2646 return (error); 2647 } 2648 if (witness_cold) { 2649 error = SYSCTL_OUT(req, w_stillcold, sizeof(w_stillcold)); 2650 return (error); 2651 } 2652 error = 0; 2653 2654 error = sysctl_wire_old_buffer(req, 0); 2655 if (error != 0) 2656 return (error); 2657 sb = sbuf_new_for_sysctl(NULL, NULL, FULLGRAPH_SBUF_SIZE, req); 2658 if (sb == NULL) 2659 return (ENOMEM); 2660 sbuf_printf(sb, "\n"); 2661 2662 mtx_lock_spin(&w_mtx); 2663 STAILQ_FOREACH(w, &w_all, w_list) 2664 w->w_displayed = 0; 2665 STAILQ_FOREACH(w, &w_all, w_list) 2666 witness_add_fullgraph(sb, w); 2667 mtx_unlock_spin(&w_mtx); 2668 2669 /* 2670 * Close the sbuf and return to userland. 2671 */ 2672 error = sbuf_finish(sb); 2673 sbuf_delete(sb); 2674 2675 return (error); 2676} 2677 2678static int 2679sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS) 2680{ 2681 int error, value; 2682 2683 value = witness_watch; 2684 error = sysctl_handle_int(oidp, &value, 0, req); 2685 if (error != 0 || req->newptr == NULL) 2686 return (error); 2687 if (value > 1 || value < -1 || 2688 (witness_watch == -1 && value != witness_watch)) 2689 return (EINVAL); 2690 witness_watch = value; 2691 return (0); 2692} 2693 2694static void 2695witness_add_fullgraph(struct sbuf *sb, struct witness *w) 2696{ 2697 int i; 2698 2699 if (w->w_displayed != 0 || (w->w_file == NULL && w->w_line == 0)) 2700 return; 2701 w->w_displayed = 1; 2702 2703 WITNESS_INDEX_ASSERT(w->w_index); 2704 for (i = 1; i <= w_max_used_index; i++) { 2705 if (w_rmatrix[w->w_index][i] & WITNESS_PARENT) { 2706 sbuf_printf(sb, "\"%s\",\"%s\"\n", w->w_name, 2707 w_data[i].w_name); 2708 witness_add_fullgraph(sb, &w_data[i]); 2709 } 2710 } 2711} 2712 2713/* 2714 * A simple hash function. Takes a key pointer and a key size. If size == 0, 2715 * interprets the key as a string and reads until the null 2716 * terminator. Otherwise, reads the first size bytes. Returns an unsigned 32-bit 2717 * hash value computed from the key. 2718 */ 2719static uint32_t 2720witness_hash_djb2(const uint8_t *key, uint32_t size) 2721{ 2722 unsigned int hash = 5381; 2723 int i; 2724 2725 /* hash = hash * 33 + key[i] */ 2726 if (size) 2727 for (i = 0; i < size; i++) 2728 hash = ((hash << 5) + hash) + (unsigned int)key[i]; 2729 else 2730 for (i = 0; key[i] != 0; i++) 2731 hash = ((hash << 5) + hash) + (unsigned int)key[i]; 2732 2733 return (hash); 2734} 2735 2736 2737/* 2738 * Initializes the two witness hash tables. Called exactly once from 2739 * witness_initialize(). 2740 */ 2741static void 2742witness_init_hash_tables(void) 2743{ 2744 int i; 2745 2746 MPASS(witness_cold); 2747 2748 /* Initialize the hash tables. */ 2749 for (i = 0; i < WITNESS_HASH_SIZE; i++) 2750 w_hash.wh_array[i] = NULL; 2751 2752 w_hash.wh_size = WITNESS_HASH_SIZE; 2753 w_hash.wh_count = 0; 2754 2755 /* Initialize the lock order data hash. */ 2756 w_lofree = NULL; 2757 for (i = 0; i < WITNESS_LO_DATA_COUNT; i++) { 2758 memset(&w_lodata[i], 0, sizeof(w_lodata[i])); 2759 w_lodata[i].wlod_next = w_lofree; 2760 w_lofree = &w_lodata[i]; 2761 } 2762 w_lohash.wloh_size = WITNESS_LO_HASH_SIZE; 2763 w_lohash.wloh_count = 0; 2764 for (i = 0; i < WITNESS_LO_HASH_SIZE; i++) 2765 w_lohash.wloh_array[i] = NULL; 2766} 2767 2768static struct witness * 2769witness_hash_get(const char *key) 2770{ 2771 struct witness *w; 2772 uint32_t hash; 2773 2774 MPASS(key != NULL); 2775 if (witness_cold == 0) 2776 mtx_assert(&w_mtx, MA_OWNED); 2777 hash = witness_hash_djb2(key, 0) % w_hash.wh_size; 2778 w = w_hash.wh_array[hash]; 2779 while (w != NULL) { 2780 if (strcmp(w->w_name, key) == 0) 2781 goto out; 2782 w = w->w_hash_next; 2783 } 2784 2785out: 2786 return (w); 2787} 2788 2789static void 2790witness_hash_put(struct witness *w) 2791{ 2792 uint32_t hash; 2793 2794 MPASS(w != NULL); 2795 MPASS(w->w_name != NULL); 2796 if (witness_cold == 0) 2797 mtx_assert(&w_mtx, MA_OWNED); 2798 KASSERT(witness_hash_get(w->w_name) == NULL, 2799 ("%s: trying to add a hash entry that already exists!", __func__)); 2800 KASSERT(w->w_hash_next == NULL, 2801 ("%s: w->w_hash_next != NULL", __func__)); 2802 2803 hash = witness_hash_djb2(w->w_name, 0) % w_hash.wh_size; 2804 w->w_hash_next = w_hash.wh_array[hash]; 2805 w_hash.wh_array[hash] = w; 2806 w_hash.wh_count++; 2807} 2808 2809 2810static struct witness_lock_order_data * 2811witness_lock_order_get(struct witness *parent, struct witness *child) 2812{ 2813 struct witness_lock_order_data *data = NULL; 2814 struct witness_lock_order_key key; 2815 unsigned int hash; 2816 2817 MPASS(parent != NULL && child != NULL); 2818 key.from = parent->w_index; 2819 key.to = child->w_index; 2820 WITNESS_INDEX_ASSERT(key.from); 2821 WITNESS_INDEX_ASSERT(key.to); 2822 if ((w_rmatrix[parent->w_index][child->w_index] 2823 & WITNESS_LOCK_ORDER_KNOWN) == 0) 2824 goto out; 2825 2826 hash = witness_hash_djb2((const char*)&key, 2827 sizeof(key)) % w_lohash.wloh_size; 2828 data = w_lohash.wloh_array[hash]; 2829 while (data != NULL) { 2830 if (witness_lock_order_key_equal(&data->wlod_key, &key)) 2831 break; 2832 data = data->wlod_next; 2833 } 2834 2835out: 2836 return (data); 2837} 2838 2839/* 2840 * Verify that parent and child have a known relationship, are not the same, 2841 * and child is actually a child of parent. This is done without w_mtx 2842 * to avoid contention in the common case. 2843 */ 2844static int 2845witness_lock_order_check(struct witness *parent, struct witness *child) 2846{ 2847 2848 if (parent != child && 2849 w_rmatrix[parent->w_index][child->w_index] 2850 & WITNESS_LOCK_ORDER_KNOWN && 2851 isitmychild(parent, child)) 2852 return (1); 2853 2854 return (0); 2855} 2856 2857static int 2858witness_lock_order_add(struct witness *parent, struct witness *child) 2859{ 2860 struct witness_lock_order_data *data = NULL; 2861 struct witness_lock_order_key key; 2862 unsigned int hash; 2863 2864 MPASS(parent != NULL && child != NULL); 2865 key.from = parent->w_index; 2866 key.to = child->w_index; 2867 WITNESS_INDEX_ASSERT(key.from); 2868 WITNESS_INDEX_ASSERT(key.to); 2869 if (w_rmatrix[parent->w_index][child->w_index] 2870 & WITNESS_LOCK_ORDER_KNOWN) 2871 return (1); 2872 2873 hash = witness_hash_djb2((const char*)&key, 2874 sizeof(key)) % w_lohash.wloh_size; 2875 w_rmatrix[parent->w_index][child->w_index] |= WITNESS_LOCK_ORDER_KNOWN; 2876 data = w_lofree; 2877 if (data == NULL) 2878 return (0); 2879 w_lofree = data->wlod_next; 2880 data->wlod_next = w_lohash.wloh_array[hash]; 2881 data->wlod_key = key; 2882 w_lohash.wloh_array[hash] = data; 2883 w_lohash.wloh_count++; 2884 stack_zero(&data->wlod_stack); 2885 stack_save(&data->wlod_stack); 2886 return (1); 2887} 2888 2889/* Call this whenver the structure of the witness graph changes. */ 2890static void 2891witness_increment_graph_generation(void) 2892{ 2893 2894 if (witness_cold == 0) 2895 mtx_assert(&w_mtx, MA_OWNED); 2896 w_generation++; 2897} 2898 2899#ifdef KDB 2900static void 2901_witness_debugger(int cond, const char *msg) 2902{ 2903 2904 if (witness_trace && cond) 2905 kdb_backtrace(); 2906 if (witness_kdb && cond) 2907 kdb_enter(KDB_WHY_WITNESS, msg); 2908} 2909#endif 2910