subr_witness.c revision 145425
11590Srgrimes/*- 21590Srgrimes * Copyright (c) 1998 Berkeley Software Design, Inc. All rights reserved. 31590Srgrimes * 41590Srgrimes * Redistribution and use in source and binary forms, with or without 51590Srgrimes * modification, are permitted provided that the following conditions 61590Srgrimes * are met: 71590Srgrimes * 1. Redistributions of source code must retain the above copyright 81590Srgrimes * notice, this list of conditions and the following disclaimer. 91590Srgrimes * 2. Redistributions in binary form must reproduce the above copyright 101590Srgrimes * notice, this list of conditions and the following disclaimer in the 111590Srgrimes * documentation and/or other materials provided with the distribution. 121590Srgrimes * 3. Berkeley Software Design Inc's name may not be used to endorse or 131590Srgrimes * promote products derived from this software without specific prior 141590Srgrimes * written permission. 151590Srgrimes * 161590Srgrimes * THIS SOFTWARE IS PROVIDED BY BERKELEY SOFTWARE DESIGN INC ``AS IS'' AND 171590Srgrimes * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 181590Srgrimes * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 191590Srgrimes * ARE DISCLAIMED. IN NO EVENT SHALL BERKELEY SOFTWARE DESIGN INC BE LIABLE 201590Srgrimes * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 211590Srgrimes * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 221590Srgrimes * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 231590Srgrimes * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 241590Srgrimes * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 251590Srgrimes * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 261590Srgrimes * SUCH DAMAGE. 271590Srgrimes * 281590Srgrimes * from BSDI $Id: mutex_witness.c,v 1.1.2.20 2000/04/27 03:10:27 cp Exp $ 291590Srgrimes * and BSDI $Id: synch_machdep.c,v 2.3.2.39 2000/04/27 03:10:25 cp Exp $ 301590Srgrimes */ 311590Srgrimes 321590Srgrimes/* 331590Srgrimes * Implementation of the `witness' lock verifier. Originally implemented for 341590Srgrimes * mutexes in BSD/OS. Extended to handle generic lock objects and lock 351590Srgrimes * classes in FreeBSD. 3669795Sobrien */ 371590Srgrimes 381590Srgrimes/* 391590Srgrimes * Main Entry: witness 401590Srgrimes * Pronunciation: 'wit-n&s 411590Srgrimes * Function: noun 421590Srgrimes * Etymology: Middle English witnesse, from Old English witnes knowledge, 431590Srgrimes * testimony, witness, from 2wit 441590Srgrimes * Date: before 12th century 451590Srgrimes * 1 : attestation of a fact or event : TESTIMONY 461590Srgrimes * 2 : one that gives evidence; specifically : one who testifies in 471590Srgrimes * a cause or before a judicial tribunal 481590Srgrimes * 3 : one asked to be present at a transaction so as to be able to 491590Srgrimes * testify to its having taken place 501590Srgrimes * 4 : one who has personal knowledge of something 511590Srgrimes * 5 a : something serving as evidence or proof : SIGN 521590Srgrimes * b : public affirmation by word or example of usually 531590Srgrimes * religious faith or conviction <the heroic witness to divine 541590Srgrimes * life -- Pilot> 551590Srgrimes * 6 capitalized : a member of the Jehovah's Witnesses 561590Srgrimes */ 571590Srgrimes 581590Srgrimes/* 5998771Sjmallett * Special rules concerning Giant and lock orders: 601590Srgrimes * 61116390Scharnier * 1) Giant must be acquired before any other mutexes. Stated another way, 62116390Scharnier * no other mutex may be held when Giant is acquired. 631590Srgrimes * 641590Srgrimes * 2) Giant must be released when blocking on a sleepable lock. 651590Srgrimes * 661590Srgrimes * This rule is less obvious, but is a result of Giant providing the same 671590Srgrimes * semantics as spl(). Basically, when a thread sleeps, it must release 681590Srgrimes * Giant. When a thread blocks on a sleepable lock, it sleeps. Hence rule 6998771Sjmallett * 2). 701590Srgrimes * 71116390Scharnier * 3) Giant may be acquired before or after sleepable locks. 72116390Scharnier * 731590Srgrimes * This rule is also not quite as obvious. Giant may be acquired after 741590Srgrimes * a sleepable lock because it is a non-sleepable lock and non-sleepable 751590Srgrimes * locks may always be acquired while holding a sleepable lock. The second 761590Srgrimes * case, Giant before a sleepable lock, follows from rule 2) above. Suppose 771590Srgrimes * you have two threads T1 and T2 and a sleepable lock X. Suppose that T1 781590Srgrimes * acquires X and blocks on Giant. Then suppose that T2 acquires Giant and 7998771Sjmallett * blocks on X. When T2 blocks on X, T2 will release Giant allowing T1 to 801590Srgrimes * execute. Thus, acquiring Giant both before and after a sleepable lock 81116390Scharnier * will not result in a lock order reversal. 82116390Scharnier */ 831590Srgrimes 841590Srgrimes#include <sys/cdefs.h> 851590Srgrimes__FBSDID("$FreeBSD: head/sys/kern/subr_witness.c 145425 2005-04-22 22:43:31Z jeff $"); 861590Srgrimes 871590Srgrimes#include "opt_ddb.h" 881590Srgrimes#include "opt_witness.h" 8998771Sjmallett 901590Srgrimes#include <sys/param.h> 91116390Scharnier#include <sys/bus.h> 92116390Scharnier#include <sys/kdb.h> 931590Srgrimes#include <sys/kernel.h> 941590Srgrimes#include <sys/ktr.h> 951590Srgrimes#include <sys/lock.h> 961590Srgrimes#include <sys/malloc.h> 971590Srgrimes#include <sys/mutex.h> 981590Srgrimes#include <sys/proc.h> 991590Srgrimes#include <sys/sysctl.h> 1001590Srgrimes#include <sys/systm.h> 1011590Srgrimes 1021590Srgrimes#include <ddb/ddb.h> 1031590Srgrimes 1041590Srgrimes#include <machine/stdarg.h> 1051590Srgrimes 1061590Srgrimes/* Define this to check for blessed mutexes */ 1071590Srgrimes#undef BLESSING 1081590Srgrimes 1091590Srgrimes#define WITNESS_COUNT 1024 1101590Srgrimes#define WITNESS_CHILDCOUNT (WITNESS_COUNT * 4) 1111590Srgrimes/* 1121590Srgrimes * XXX: This is somewhat bogus, as we assume here that at most 1024 threads 1131590Srgrimes * will hold LOCK_NCHILDREN * 2 locks. We handle failure ok, and we should 1141590Srgrimes * probably be safe for the most part, but it's still a SWAG. 1151590Srgrimes */ 1161590Srgrimes#define LOCK_CHILDCOUNT (MAXCPU + 1024) * 2 1171590Srgrimes 1181590Srgrimes#define WITNESS_NCHILDREN 6 1191590Srgrimes 1201590Srgrimesstruct witness_child_list_entry; 1211590Srgrimes 1221590Srgrimesstruct witness { 1231590Srgrimes const char *w_name; 1241590Srgrimes struct lock_class *w_class; 1251590Srgrimes STAILQ_ENTRY(witness) w_list; /* List of all witnesses. */ 1261590Srgrimes STAILQ_ENTRY(witness) w_typelist; /* Witnesses of a type. */ 1271590Srgrimes struct witness_child_list_entry *w_children; /* Great evilness... */ 1281590Srgrimes const char *w_file; 1291590Srgrimes int w_line; 1301590Srgrimes u_int w_level; 1311590Srgrimes u_int w_refcount; 1321590Srgrimes u_char w_Giant_squawked:1; 1331590Srgrimes u_char w_other_squawked:1; 134152395Sdwmalone u_char w_same_squawked:1; 1351590Srgrimes u_char w_displayed:1; 1361590Srgrimes}; 1371590Srgrimes 1381590Srgrimesstruct witness_child_list_entry { 1391590Srgrimes struct witness_child_list_entry *wcl_next; 1401590Srgrimes struct witness *wcl_children[WITNESS_NCHILDREN]; 1411590Srgrimes u_int wcl_count; 1421590Srgrimes}; 1431590Srgrimes 1441590SrgrimesSTAILQ_HEAD(witness_list, witness); 1451590Srgrimes 1461590Srgrimes#ifdef BLESSING 1471590Srgrimesstruct witness_blessed { 1481590Srgrimes const char *b_lock1; 1491590Srgrimes const char *b_lock2; 1501590Srgrimes}; 1511590Srgrimes#endif 1521590Srgrimes 1531590Srgrimesstruct witness_order_list_entry { 1541590Srgrimes const char *w_name; 1551590Srgrimes struct lock_class *w_class; 1561590Srgrimes}; 1571590Srgrimes 1581590Srgrimes#ifdef BLESSING 1591590Srgrimesstatic int blessed(struct witness *, struct witness *); 1601590Srgrimes#endif 1611590Srgrimesstatic int depart(struct witness *w); 1621590Srgrimesstatic struct witness *enroll(const char *description, 1631590Srgrimes struct lock_class *lock_class); 1641590Srgrimesstatic int insertchild(struct witness *parent, struct witness *child); 1651590Srgrimesstatic int isitmychild(struct witness *parent, struct witness *child); 1661590Srgrimesstatic int isitmydescendant(struct witness *parent, struct witness *child); 1671590Srgrimesstatic int itismychild(struct witness *parent, struct witness *child); 16869795Sobrienstatic int rebalancetree(struct witness_list *list); 16985632Sschweikhstatic void removechild(struct witness *parent, struct witness *child); 1701590Srgrimesstatic int reparentchildren(struct witness *newparent, 1711590Srgrimes struct witness *oldparent); 1721590Srgrimesstatic int sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS); 1731590Srgrimesstatic void witness_displaydescendants(void(*)(const char *fmt, ...), 1741590Srgrimes struct witness *, int indent); 1751590Srgrimesstatic const char *fixup_filename(const char *file); 1761590Srgrimesstatic void witness_leveldescendents(struct witness *parent, int level); 1771590Srgrimesstatic void witness_levelall(void); 1781590Srgrimesstatic struct witness *witness_get(void); 1791590Srgrimesstatic void witness_free(struct witness *m); 1801590Srgrimesstatic struct witness_child_list_entry *witness_child_get(void); 1811590Srgrimesstatic void witness_child_free(struct witness_child_list_entry *wcl); 1821590Srgrimesstatic struct lock_list_entry *witness_lock_list_get(void); 1831590Srgrimesstatic void witness_lock_list_free(struct lock_list_entry *lle); 1841590Srgrimesstatic struct lock_instance *find_instance(struct lock_list_entry *lock_list, 1851590Srgrimes struct lock_object *lock); 1861590Srgrimesstatic void witness_list_lock(struct lock_instance *instance); 1871590Srgrimes#ifdef DDB 1881590Srgrimesstatic void witness_list(struct thread *td); 1891590Srgrimesstatic void witness_display_list(void(*prnt)(const char *fmt, ...), 1901590Srgrimes struct witness_list *list); 1911590Srgrimesstatic void witness_display(void(*)(const char *fmt, ...)); 1921590Srgrimes#endif 1931590Srgrimes 1941590SrgrimesSYSCTL_NODE(_debug, OID_AUTO, witness, CTLFLAG_RW, 0, "Witness Locking"); 1951590Srgrimes 1961590Srgrimes/* 197131184Sschweikh * If set to 0, witness is disabled. If set to 1, witness performs full lock 1981590Srgrimes * order checking for all locks. If set to 2 or higher, then witness skips 1991590Srgrimes * the full lock order check if the lock being acquired is at a higher level 2001590Srgrimes * (i.e. farther down in the tree) than the current lock. This last mode is 2011590Srgrimes * somewhat experimental and not considered fully safe. At runtime, this 2021590Srgrimes * value may be set to 0 to turn off witness. witness is not allowed be 203131184Sschweikh * turned on once it is turned off, however. 204131184Sschweikh */ 205131184Sschweikhstatic int witness_watch = 1; 206131184SschweikhTUNABLE_INT("debug.witness.watch", &witness_watch); 207205989SavgSYSCTL_PROC(_debug_witness, OID_AUTO, watch, CTLFLAG_RW | CTLTYPE_INT, NULL, 0, 208205989Savg sysctl_debug_witness_watch, "I", "witness is watching lock operations"); 2091590Srgrimes 2101590Srgrimes#ifdef KDB 2111590Srgrimes/* 2121590Srgrimes * When KDB is enabled and witness_kdb is set to 1, it will cause the system 2131590Srgrimes * to drop into kdebug() when: 2141590Srgrimes * - a lock heirarchy violation occurs 2151590Srgrimes * - locks are held when going to sleep. 216244578Sandrew */ 21793440Sdwmalone#ifdef WITNESS_KDB 2181590Srgrimesint witness_kdb = 1; 2191590Srgrimes#else 2201590Srgrimesint witness_kdb = 0; 2211590Srgrimes#endif 2221590SrgrimesTUNABLE_INT("debug.witness.kdb", &witness_kdb); 2231590SrgrimesSYSCTL_INT(_debug_witness, OID_AUTO, kdb, CTLFLAG_RW, &witness_kdb, 0, ""); 2241590Srgrimes 2251590Srgrimes/* 2261590Srgrimes * When KDB is enabled and witness_trace is set to 1, it will cause the system 2271590Srgrimes * to print a stack trace: 2281590Srgrimes * - a lock heirarchy violation occurs 2291590Srgrimes * - locks are held when going to sleep. 2301590Srgrimes */ 2311590Srgrimesint witness_trace = 1; 2321590SrgrimesTUNABLE_INT("debug.witness.trace", &witness_trace); 2331590SrgrimesSYSCTL_INT(_debug_witness, OID_AUTO, trace, CTLFLAG_RW, &witness_trace, 0, ""); 2341590Srgrimes#endif /* KDB */ 2351590Srgrimes 2361590Srgrimes#ifdef WITNESS_SKIPSPIN 2371590Srgrimesint witness_skipspin = 1; 23885632Sschweikh#else 2391590Srgrimesint witness_skipspin = 0; 2401590Srgrimes#endif 2411590SrgrimesTUNABLE_INT("debug.witness.skipspin", &witness_skipspin); 2421590SrgrimesSYSCTL_INT(_debug_witness, OID_AUTO, skipspin, CTLFLAG_RDTUN, 2431590Srgrimes &witness_skipspin, 0, ""); 2441590Srgrimes 2451590Srgrimesstatic struct mtx w_mtx; 2461590Srgrimesstatic struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free); 2471590Srgrimesstatic struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all); 2481590Srgrimesstatic struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin); 2491590Srgrimesstatic struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep); 2501590Srgrimesstatic struct witness_child_list_entry *w_child_free = NULL; 2511590Srgrimesstatic struct lock_list_entry *w_lock_list_free = NULL; 2521590Srgrimes 2531590Srgrimesstatic struct witness w_data[WITNESS_COUNT]; 2541590Srgrimesstatic struct witness_child_list_entry w_childdata[WITNESS_CHILDCOUNT]; 2551590Srgrimesstatic struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT]; 2561590Srgrimes 2571590Srgrimesstatic struct witness_order_list_entry order_lists[] = { 2581590Srgrimes { "proctree", &lock_class_sx }, 259105244Scharnier { "allproc", &lock_class_sx }, 2601590Srgrimes { "Giant", &lock_class_mtx_sleep }, 2611590Srgrimes { "filedesc structure", &lock_class_mtx_sleep }, 2621590Srgrimes { "pipe mutex", &lock_class_mtx_sleep }, 2631590Srgrimes { "sigio lock", &lock_class_mtx_sleep }, 2641590Srgrimes { "process group", &lock_class_mtx_sleep }, 2651590Srgrimes { "process lock", &lock_class_mtx_sleep }, 2661590Srgrimes { "session", &lock_class_mtx_sleep }, 2671590Srgrimes { "uidinfo hash", &lock_class_mtx_sleep }, 2681590Srgrimes { "uidinfo struct", &lock_class_mtx_sleep }, 2691590Srgrimes { "allprison", &lock_class_mtx_sleep }, 2701590Srgrimes { NULL, NULL }, 2711590Srgrimes /* 2721590Srgrimes * Sockets 2731590Srgrimes */ 2741590Srgrimes { "filedesc structure", &lock_class_mtx_sleep }, 2751590Srgrimes { "accept", &lock_class_mtx_sleep }, 2761590Srgrimes { "so_snd", &lock_class_mtx_sleep }, 2771590Srgrimes { "so_rcv", &lock_class_mtx_sleep }, 2781590Srgrimes { "sellck", &lock_class_mtx_sleep }, 2791590Srgrimes { NULL, NULL }, 2801590Srgrimes /* 2811590Srgrimes * Routing 2821590Srgrimes */ 2831590Srgrimes { "so_rcv", &lock_class_mtx_sleep }, 2841590Srgrimes { "radix node head", &lock_class_mtx_sleep }, 2851590Srgrimes { "rtentry", &lock_class_mtx_sleep }, 2861590Srgrimes { "ifaddr", &lock_class_mtx_sleep }, 2871590Srgrimes { NULL, NULL }, 2881590Srgrimes /* 2891590Srgrimes * UNIX Domain Sockets 2901590Srgrimes */ 2911590Srgrimes { "unp", &lock_class_mtx_sleep }, 2921590Srgrimes { "so_snd", &lock_class_mtx_sleep }, 293105244Scharnier { NULL, NULL }, 2941590Srgrimes /* 2951590Srgrimes * UDP/IP 2961590Srgrimes */ 2971590Srgrimes { "udp", &lock_class_mtx_sleep }, 2981590Srgrimes { "udpinp", &lock_class_mtx_sleep }, 2991590Srgrimes { "so_snd", &lock_class_mtx_sleep }, 3001590Srgrimes { NULL, NULL }, 3011590Srgrimes /* 3021590Srgrimes * TCP/IP 3031590Srgrimes */ 3041590Srgrimes { "tcp", &lock_class_mtx_sleep }, 3051590Srgrimes { "tcpinp", &lock_class_mtx_sleep }, 3061590Srgrimes { "so_snd", &lock_class_mtx_sleep }, 3071590Srgrimes { NULL, NULL }, 3081590Srgrimes /* 3091590Srgrimes * SLIP 3101590Srgrimes */ 3111590Srgrimes { "slip_mtx", &lock_class_mtx_sleep }, 3121590Srgrimes { "slip sc_mtx", &lock_class_mtx_sleep }, 313125633Sbde { NULL, NULL }, 3141590Srgrimes /* 3151590Srgrimes * netatalk 3161590Srgrimes */ 3171590Srgrimes { "ddp_list_mtx", &lock_class_mtx_sleep }, 3181590Srgrimes { "ddp_mtx", &lock_class_mtx_sleep }, 3191590Srgrimes { NULL, NULL }, 3201590Srgrimes /* 3211590Srgrimes * BPF 3221590Srgrimes */ 3231590Srgrimes { "bpf global lock", &lock_class_mtx_sleep }, 3241590Srgrimes { "bpf interface lock", &lock_class_mtx_sleep }, 3251590Srgrimes { "bpf cdev lock", &lock_class_mtx_sleep }, 3261590Srgrimes { NULL, NULL }, 3271590Srgrimes /* 3281590Srgrimes * NFS server 3291590Srgrimes */ 330 { "nfsd_mtx", &lock_class_mtx_sleep }, 331 { "so_snd", &lock_class_mtx_sleep }, 332 { NULL, NULL }, 333 /* 334 * CDEV 335 */ 336 { "system map", &lock_class_mtx_sleep }, 337 { "vm page queue mutex", &lock_class_mtx_sleep }, 338 { "vnode interlock", &lock_class_mtx_sleep }, 339 { "cdev", &lock_class_mtx_sleep }, 340 { NULL, NULL }, 341 /* 342 * spin locks 343 */ 344#ifdef SMP 345 { "ap boot", &lock_class_mtx_spin }, 346#endif 347 { "sio", &lock_class_mtx_spin }, 348#ifdef __i386__ 349 { "cy", &lock_class_mtx_spin }, 350#endif 351 { "uart_hwmtx", &lock_class_mtx_spin }, 352 { "sabtty", &lock_class_mtx_spin }, 353 { "zstty", &lock_class_mtx_spin }, 354 { "ng_node", &lock_class_mtx_spin }, 355 { "ng_worklist", &lock_class_mtx_spin }, 356 { "taskqueue_fast", &lock_class_mtx_spin }, 357 { "intr table", &lock_class_mtx_spin }, 358 { "ithread table lock", &lock_class_mtx_spin }, 359 { "sleepq chain", &lock_class_mtx_spin }, 360 { "sched lock", &lock_class_mtx_spin }, 361 { "turnstile chain", &lock_class_mtx_spin }, 362 { "td_contested", &lock_class_mtx_spin }, 363 { "callout", &lock_class_mtx_spin }, 364 { "entropy harvest mutex", &lock_class_mtx_spin }, 365 /* 366 * leaf locks 367 */ 368 { "allpmaps", &lock_class_mtx_spin }, 369 { "vm page queue free mutex", &lock_class_mtx_spin }, 370 { "icu", &lock_class_mtx_spin }, 371#ifdef SMP 372 { "smp rendezvous", &lock_class_mtx_spin }, 373#if defined(__i386__) || defined(__amd64__) 374 { "tlb", &lock_class_mtx_spin }, 375#endif 376#ifdef __sparc64__ 377 { "ipi", &lock_class_mtx_spin }, 378#endif 379#endif 380 { "clk", &lock_class_mtx_spin }, 381 { "mutex profiling lock", &lock_class_mtx_spin }, 382 { "kse zombie lock", &lock_class_mtx_spin }, 383 { "ALD Queue", &lock_class_mtx_spin }, 384#ifdef __ia64__ 385 { "MCA spin lock", &lock_class_mtx_spin }, 386#endif 387#if defined(__i386__) || defined(__amd64__) 388 { "pcicfg", &lock_class_mtx_spin }, 389 { "NDIS thread lock", &lock_class_mtx_spin }, 390#endif 391 { "tw_osl_io_lock", &lock_class_mtx_spin }, 392 { "tw_osl_q_lock", &lock_class_mtx_spin }, 393 { "tw_cl_io_lock", &lock_class_mtx_spin }, 394 { "tw_cl_intr_lock", &lock_class_mtx_spin }, 395 { "tw_cl_gen_lock", &lock_class_mtx_spin }, 396 { NULL, NULL }, 397 { NULL, NULL } 398}; 399 400#ifdef BLESSING 401/* 402 * Pairs of locks which have been blessed 403 * Don't complain about order problems with blessed locks 404 */ 405static struct witness_blessed blessed_list[] = { 406}; 407static int blessed_count = 408 sizeof(blessed_list) / sizeof(struct witness_blessed); 409#endif 410 411/* 412 * List of all locks in the system. 413 */ 414TAILQ_HEAD(, lock_object) all_locks = TAILQ_HEAD_INITIALIZER(all_locks); 415 416static struct mtx all_mtx = { 417 { &lock_class_mtx_sleep, /* mtx_object.lo_class */ 418 "All locks list", /* mtx_object.lo_name */ 419 "All locks list", /* mtx_object.lo_type */ 420 LO_INITIALIZED, /* mtx_object.lo_flags */ 421 { NULL, NULL }, /* mtx_object.lo_list */ 422 NULL }, /* mtx_object.lo_witness */ 423 MTX_UNOWNED, 0 /* mtx_lock, mtx_recurse */ 424}; 425 426/* 427 * This global is set to 0 once it becomes safe to use the witness code. 428 */ 429static int witness_cold = 1; 430 431/* 432 * Global variables for book keeping. 433 */ 434static int lock_cur_cnt; 435static int lock_max_cnt; 436 437/* 438 * The WITNESS-enabled diagnostic code. 439 */ 440static void 441witness_initialize(void *dummy __unused) 442{ 443 struct lock_object *lock; 444 struct witness_order_list_entry *order; 445 struct witness *w, *w1; 446 int i; 447 448 /* 449 * We have to release Giant before initializing its witness 450 * structure so that WITNESS doesn't get confused. 451 */ 452 mtx_unlock(&Giant); 453 mtx_assert(&Giant, MA_NOTOWNED); 454 455 CTR1(KTR_WITNESS, "%s: initializing witness", __func__); 456 TAILQ_INSERT_HEAD(&all_locks, &all_mtx.mtx_object, lo_list); 457 mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET | 458 MTX_NOWITNESS); 459 for (i = 0; i < WITNESS_COUNT; i++) 460 witness_free(&w_data[i]); 461 for (i = 0; i < WITNESS_CHILDCOUNT; i++) 462 witness_child_free(&w_childdata[i]); 463 for (i = 0; i < LOCK_CHILDCOUNT; i++) 464 witness_lock_list_free(&w_locklistdata[i]); 465 466 /* First add in all the specified order lists. */ 467 for (order = order_lists; order->w_name != NULL; order++) { 468 w = enroll(order->w_name, order->w_class); 469 if (w == NULL) 470 continue; 471 w->w_file = "order list"; 472 for (order++; order->w_name != NULL; order++) { 473 w1 = enroll(order->w_name, order->w_class); 474 if (w1 == NULL) 475 continue; 476 w1->w_file = "order list"; 477 if (!itismychild(w, w1)) 478 panic("Not enough memory for static orders!"); 479 w = w1; 480 } 481 } 482 483 /* Iterate through all locks and add them to witness. */ 484 mtx_lock(&all_mtx); 485 TAILQ_FOREACH(lock, &all_locks, lo_list) { 486 if (lock->lo_flags & LO_WITNESS) 487 lock->lo_witness = enroll(lock->lo_type, 488 lock->lo_class); 489 else 490 lock->lo_witness = NULL; 491 } 492 mtx_unlock(&all_mtx); 493 494 /* Mark the witness code as being ready for use. */ 495 atomic_store_rel_int(&witness_cold, 0); 496 497 mtx_lock(&Giant); 498} 499SYSINIT(witness_init, SI_SUB_WITNESS, SI_ORDER_FIRST, witness_initialize, NULL) 500 501static int 502sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS) 503{ 504 int error, value; 505 506 value = witness_watch; 507 error = sysctl_handle_int(oidp, &value, 0, req); 508 if (error != 0 || req->newptr == NULL) 509 return (error); 510 error = suser(req->td); 511 if (error != 0) 512 return (error); 513 if (value == witness_watch) 514 return (0); 515 if (value != 0) 516 return (EINVAL); 517 witness_watch = 0; 518 return (0); 519} 520 521void 522witness_init(struct lock_object *lock) 523{ 524 struct lock_class *class; 525 526 class = lock->lo_class; 527 if (lock->lo_flags & LO_INITIALIZED) 528 panic("%s: lock (%s) %s is already initialized", __func__, 529 class->lc_name, lock->lo_name); 530 if ((lock->lo_flags & LO_RECURSABLE) != 0 && 531 (class->lc_flags & LC_RECURSABLE) == 0) 532 panic("%s: lock (%s) %s can not be recursable", __func__, 533 class->lc_name, lock->lo_name); 534 if ((lock->lo_flags & LO_SLEEPABLE) != 0 && 535 (class->lc_flags & LC_SLEEPABLE) == 0) 536 panic("%s: lock (%s) %s can not be sleepable", __func__, 537 class->lc_name, lock->lo_name); 538 if ((lock->lo_flags & LO_UPGRADABLE) != 0 && 539 (class->lc_flags & LC_UPGRADABLE) == 0) 540 panic("%s: lock (%s) %s can not be upgradable", __func__, 541 class->lc_name, lock->lo_name); 542 543 mtx_lock(&all_mtx); 544 TAILQ_INSERT_TAIL(&all_locks, lock, lo_list); 545 lock->lo_flags |= LO_INITIALIZED; 546 lock_cur_cnt++; 547 if (lock_cur_cnt > lock_max_cnt) 548 lock_max_cnt = lock_cur_cnt; 549 mtx_unlock(&all_mtx); 550 if (!witness_cold && witness_watch != 0 && panicstr == NULL && 551 (lock->lo_flags & LO_WITNESS) != 0) 552 lock->lo_witness = enroll(lock->lo_type, class); 553 else 554 lock->lo_witness = NULL; 555} 556 557void 558witness_destroy(struct lock_object *lock) 559{ 560 struct witness *w; 561 562 if (witness_cold) 563 panic("lock (%s) %s destroyed while witness_cold", 564 lock->lo_class->lc_name, lock->lo_name); 565 if ((lock->lo_flags & LO_INITIALIZED) == 0) 566 panic("%s: lock (%s) %s is not initialized", __func__, 567 lock->lo_class->lc_name, lock->lo_name); 568 569 /* XXX: need to verify that no one holds the lock */ 570 w = lock->lo_witness; 571 if (w != NULL) { 572 mtx_lock_spin(&w_mtx); 573 MPASS(w->w_refcount > 0); 574 w->w_refcount--; 575 576 /* 577 * Lock is already released if we have an allocation failure 578 * and depart() fails. 579 */ 580 if (w->w_refcount != 0 || depart(w)) 581 mtx_unlock_spin(&w_mtx); 582 } 583 584 mtx_lock(&all_mtx); 585 lock_cur_cnt--; 586 TAILQ_REMOVE(&all_locks, lock, lo_list); 587 lock->lo_flags &= ~LO_INITIALIZED; 588 mtx_unlock(&all_mtx); 589} 590 591#ifdef DDB 592static void 593witness_display_list(void(*prnt)(const char *fmt, ...), 594 struct witness_list *list) 595{ 596 struct witness *w; 597 598 STAILQ_FOREACH(w, list, w_typelist) { 599 if (w->w_file == NULL || w->w_level > 0) 600 continue; 601 /* 602 * This lock has no anscestors, display its descendants. 603 */ 604 witness_displaydescendants(prnt, w, 0); 605 } 606} 607 608static void 609witness_display(void(*prnt)(const char *fmt, ...)) 610{ 611 struct witness *w; 612 613 KASSERT(!witness_cold, ("%s: witness_cold", __func__)); 614 witness_levelall(); 615 616 /* Clear all the displayed flags. */ 617 STAILQ_FOREACH(w, &w_all, w_list) { 618 w->w_displayed = 0; 619 } 620 621 /* 622 * First, handle sleep locks which have been acquired at least 623 * once. 624 */ 625 prnt("Sleep locks:\n"); 626 witness_display_list(prnt, &w_sleep); 627 628 /* 629 * Now do spin locks which have been acquired at least once. 630 */ 631 prnt("\nSpin locks:\n"); 632 witness_display_list(prnt, &w_spin); 633 634 /* 635 * Finally, any locks which have not been acquired yet. 636 */ 637 prnt("\nLocks which were never acquired:\n"); 638 STAILQ_FOREACH(w, &w_all, w_list) { 639 if (w->w_file != NULL || w->w_refcount == 0) 640 continue; 641 prnt("%s\n", w->w_name); 642 } 643} 644#endif /* DDB */ 645 646/* Trim useless garbage from filenames. */ 647static const char * 648fixup_filename(const char *file) 649{ 650 651 if (file == NULL) 652 return (NULL); 653 while (strncmp(file, "../", 3) == 0) 654 file += 3; 655 return (file); 656} 657 658int 659witness_defineorder(struct lock_object *lock1, struct lock_object *lock2) 660{ 661 662 if (witness_watch == 0 || panicstr != NULL) 663 return (0); 664 665 /* Require locks that witness knows about. */ 666 if (lock1 == NULL || lock1->lo_witness == NULL || lock2 == NULL || 667 lock2->lo_witness == NULL) 668 return (EINVAL); 669 670 MPASS(!mtx_owned(&w_mtx)); 671 mtx_lock_spin(&w_mtx); 672 673 /* 674 * If we already have either an explicit or implied lock order that 675 * is the other way around, then return an error. 676 */ 677 if (isitmydescendant(lock2->lo_witness, lock1->lo_witness)) { 678 mtx_unlock_spin(&w_mtx); 679 return (EDOOFUS); 680 } 681 682 /* Try to add the new order. */ 683 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__, 684 lock2->lo_type, lock1->lo_type); 685 if (!itismychild(lock1->lo_witness, lock2->lo_witness)) 686 return (ENOMEM); 687 mtx_unlock_spin(&w_mtx); 688 return (0); 689} 690 691void 692witness_checkorder(struct lock_object *lock, int flags, const char *file, 693 int line) 694{ 695 struct lock_list_entry **lock_list, *lle; 696 struct lock_instance *lock1, *lock2; 697 struct lock_class *class; 698 struct witness *w, *w1; 699 struct thread *td; 700 int i, j; 701 702 if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL || 703 panicstr != NULL) 704 return; 705 706 /* 707 * Try locks do not block if they fail to acquire the lock, thus 708 * there is no danger of deadlocks or of switching while holding a 709 * spin lock if we acquire a lock via a try operation. This 710 * function shouldn't even be called for try locks, so panic if 711 * that happens. 712 */ 713 if (flags & LOP_TRYLOCK) 714 panic("%s should not be called for try lock operations", 715 __func__); 716 717 w = lock->lo_witness; 718 class = lock->lo_class; 719 td = curthread; 720 file = fixup_filename(file); 721 722 if (class->lc_flags & LC_SLEEPLOCK) { 723 /* 724 * Since spin locks include a critical section, this check 725 * implicitly enforces a lock order of all sleep locks before 726 * all spin locks. 727 */ 728 if (td->td_critnest != 0 && !kdb_active) 729 panic("blockable sleep lock (%s) %s @ %s:%d", 730 class->lc_name, lock->lo_name, file, line); 731 732 /* 733 * If this is the first lock acquired then just return as 734 * no order checking is needed. 735 */ 736 if (td->td_sleeplocks == NULL) 737 return; 738 lock_list = &td->td_sleeplocks; 739 } else { 740 /* 741 * If this is the first lock, just return as no order 742 * checking is needed. We check this in both if clauses 743 * here as unifying the check would require us to use a 744 * critical section to ensure we don't migrate while doing 745 * the check. Note that if this is not the first lock, we 746 * are already in a critical section and are safe for the 747 * rest of the check. 748 */ 749 if (PCPU_GET(spinlocks) == NULL) 750 return; 751 lock_list = PCPU_PTR(spinlocks); 752 } 753 754 /* 755 * Check to see if we are recursing on a lock we already own. If 756 * so, make sure that we don't mismatch exclusive and shared lock 757 * acquires. 758 */ 759 lock1 = find_instance(*lock_list, lock); 760 if (lock1 != NULL) { 761 if ((lock1->li_flags & LI_EXCLUSIVE) != 0 && 762 (flags & LOP_EXCLUSIVE) == 0) { 763 printf("shared lock of (%s) %s @ %s:%d\n", 764 class->lc_name, lock->lo_name, file, line); 765 printf("while exclusively locked from %s:%d\n", 766 lock1->li_file, lock1->li_line); 767 panic("share->excl"); 768 } 769 if ((lock1->li_flags & LI_EXCLUSIVE) == 0 && 770 (flags & LOP_EXCLUSIVE) != 0) { 771 printf("exclusive lock of (%s) %s @ %s:%d\n", 772 class->lc_name, lock->lo_name, file, line); 773 printf("while share locked from %s:%d\n", 774 lock1->li_file, lock1->li_line); 775 panic("excl->share"); 776 } 777 return; 778 } 779 780 /* 781 * Try locks do not block if they fail to acquire the lock, thus 782 * there is no danger of deadlocks or of switching while holding a 783 * spin lock if we acquire a lock via a try operation. 784 */ 785 if (flags & LOP_TRYLOCK) 786 return; 787 788 /* 789 * Check for duplicate locks of the same type. Note that we only 790 * have to check for this on the last lock we just acquired. Any 791 * other cases will be caught as lock order violations. 792 */ 793 lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1]; 794 w1 = lock1->li_lock->lo_witness; 795 if (w1 == w) { 796 if (w->w_same_squawked || (lock->lo_flags & LO_DUPOK) || 797 (flags & LOP_DUPOK)) 798 return; 799 w->w_same_squawked = 1; 800 printf("acquiring duplicate lock of same type: \"%s\"\n", 801 lock->lo_type); 802 printf(" 1st %s @ %s:%d\n", lock1->li_lock->lo_name, 803 lock1->li_file, lock1->li_line); 804 printf(" 2nd %s @ %s:%d\n", lock->lo_name, file, line); 805#ifdef KDB 806 goto debugger; 807#else 808 return; 809#endif 810 } 811 MPASS(!mtx_owned(&w_mtx)); 812 mtx_lock_spin(&w_mtx); 813 /* 814 * If we have a known higher number just say ok 815 */ 816 if (witness_watch > 1 && w->w_level > w1->w_level) { 817 mtx_unlock_spin(&w_mtx); 818 return; 819 } 820 /* 821 * If we know that the the lock we are acquiring comes after 822 * the lock we most recently acquired in the lock order tree, 823 * then there is no need for any further checks. 824 */ 825 if (isitmydescendant(w1, w)) { 826 mtx_unlock_spin(&w_mtx); 827 return; 828 } 829 for (j = 0, lle = *lock_list; lle != NULL; lle = lle->ll_next) { 830 for (i = lle->ll_count - 1; i >= 0; i--, j++) { 831 832 MPASS(j < WITNESS_COUNT); 833 lock1 = &lle->ll_children[i]; 834 w1 = lock1->li_lock->lo_witness; 835 836 /* 837 * If this lock doesn't undergo witness checking, 838 * then skip it. 839 */ 840 if (w1 == NULL) { 841 KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0, 842 ("lock missing witness structure")); 843 continue; 844 } 845 /* 846 * If we are locking Giant and this is a sleepable 847 * lock, then skip it. 848 */ 849 if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0 && 850 lock == &Giant.mtx_object) 851 continue; 852 /* 853 * If we are locking a sleepable lock and this lock 854 * is Giant, then skip it. 855 */ 856 if ((lock->lo_flags & LO_SLEEPABLE) != 0 && 857 lock1->li_lock == &Giant.mtx_object) 858 continue; 859 /* 860 * If we are locking a sleepable lock and this lock 861 * isn't sleepable, we want to treat it as a lock 862 * order violation to enfore a general lock order of 863 * sleepable locks before non-sleepable locks. 864 */ 865 if (!((lock->lo_flags & LO_SLEEPABLE) != 0 && 866 (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0)) 867 /* 868 * Check the lock order hierarchy for a reveresal. 869 */ 870 if (!isitmydescendant(w, w1)) 871 continue; 872 /* 873 * We have a lock order violation, check to see if it 874 * is allowed or has already been yelled about. 875 */ 876 mtx_unlock_spin(&w_mtx); 877#ifdef BLESSING 878 /* 879 * If the lock order is blessed, just bail. We don't 880 * look for other lock order violations though, which 881 * may be a bug. 882 */ 883 if (blessed(w, w1)) 884 return; 885#endif 886 if (lock1->li_lock == &Giant.mtx_object) { 887 if (w1->w_Giant_squawked) 888 return; 889 else 890 w1->w_Giant_squawked = 1; 891 } else { 892 if (w1->w_other_squawked) 893 return; 894 else 895 w1->w_other_squawked = 1; 896 } 897 /* 898 * Ok, yell about it. 899 */ 900 printf("lock order reversal\n"); 901 /* 902 * Try to locate an earlier lock with 903 * witness w in our list. 904 */ 905 do { 906 lock2 = &lle->ll_children[i]; 907 MPASS(lock2->li_lock != NULL); 908 if (lock2->li_lock->lo_witness == w) 909 break; 910 if (i == 0 && lle->ll_next != NULL) { 911 lle = lle->ll_next; 912 i = lle->ll_count - 1; 913 MPASS(i >= 0 && i < LOCK_NCHILDREN); 914 } else 915 i--; 916 } while (i >= 0); 917 if (i < 0) { 918 printf(" 1st %p %s (%s) @ %s:%d\n", 919 lock1->li_lock, lock1->li_lock->lo_name, 920 lock1->li_lock->lo_type, lock1->li_file, 921 lock1->li_line); 922 printf(" 2nd %p %s (%s) @ %s:%d\n", lock, 923 lock->lo_name, lock->lo_type, file, line); 924 } else { 925 printf(" 1st %p %s (%s) @ %s:%d\n", 926 lock2->li_lock, lock2->li_lock->lo_name, 927 lock2->li_lock->lo_type, lock2->li_file, 928 lock2->li_line); 929 printf(" 2nd %p %s (%s) @ %s:%d\n", 930 lock1->li_lock, lock1->li_lock->lo_name, 931 lock1->li_lock->lo_type, lock1->li_file, 932 lock1->li_line); 933 printf(" 3rd %p %s (%s) @ %s:%d\n", lock, 934 lock->lo_name, lock->lo_type, file, line); 935 } 936#ifdef KDB 937 goto debugger; 938#else 939 return; 940#endif 941 } 942 } 943 lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1]; 944 /* 945 * If requested, build a new lock order. However, don't build a new 946 * relationship between a sleepable lock and Giant if it is in the 947 * wrong direction. The correct lock order is that sleepable locks 948 * always come before Giant. 949 */ 950 if (flags & LOP_NEWORDER && 951 !(lock1->li_lock == &Giant.mtx_object && 952 (lock->lo_flags & LO_SLEEPABLE) != 0)) { 953 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__, 954 lock->lo_type, lock1->li_lock->lo_type); 955 if (!itismychild(lock1->li_lock->lo_witness, w)) 956 /* Witness is dead. */ 957 return; 958 } 959 mtx_unlock_spin(&w_mtx); 960 return; 961 962#ifdef KDB 963debugger: 964 if (witness_trace) 965 kdb_backtrace(); 966 if (witness_kdb) 967 kdb_enter(__func__); 968#endif 969} 970 971void 972witness_lock(struct lock_object *lock, int flags, const char *file, int line) 973{ 974 struct lock_list_entry **lock_list, *lle; 975 struct lock_instance *instance; 976 struct witness *w; 977 struct thread *td; 978 979 if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL || 980 panicstr != NULL) 981 return; 982 w = lock->lo_witness; 983 td = curthread; 984 file = fixup_filename(file); 985 986 /* Determine lock list for this lock. */ 987 if (lock->lo_class->lc_flags & LC_SLEEPLOCK) 988 lock_list = &td->td_sleeplocks; 989 else 990 lock_list = PCPU_PTR(spinlocks); 991 992 /* Check to see if we are recursing on a lock we already own. */ 993 instance = find_instance(*lock_list, lock); 994 if (instance != NULL) { 995 instance->li_flags++; 996 CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__, 997 td->td_proc->p_pid, lock->lo_name, 998 instance->li_flags & LI_RECURSEMASK); 999 instance->li_file = file; 1000 instance->li_line = line; 1001 return; 1002 } 1003 1004 /* Update per-witness last file and line acquire. */ 1005 w->w_file = file; 1006 w->w_line = line; 1007 1008 /* Find the next open lock instance in the list and fill it. */ 1009 lle = *lock_list; 1010 if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) { 1011 lle = witness_lock_list_get(); 1012 if (lle == NULL) 1013 return; 1014 lle->ll_next = *lock_list; 1015 CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__, 1016 td->td_proc->p_pid, lle); 1017 *lock_list = lle; 1018 } 1019 instance = &lle->ll_children[lle->ll_count++]; 1020 instance->li_lock = lock; 1021 instance->li_line = line; 1022 instance->li_file = file; 1023 if ((flags & LOP_EXCLUSIVE) != 0) 1024 instance->li_flags = LI_EXCLUSIVE; 1025 else 1026 instance->li_flags = 0; 1027 CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__, 1028 td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1); 1029} 1030 1031void 1032witness_upgrade(struct lock_object *lock, int flags, const char *file, int line) 1033{ 1034 struct lock_instance *instance; 1035 struct lock_class *class; 1036 1037 KASSERT(!witness_cold, ("%s: witness_cold", __func__)); 1038 if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL) 1039 return; 1040 class = lock->lo_class; 1041 file = fixup_filename(file); 1042 if ((lock->lo_flags & LO_UPGRADABLE) == 0) 1043 panic("upgrade of non-upgradable lock (%s) %s @ %s:%d", 1044 class->lc_name, lock->lo_name, file, line); 1045 if ((flags & LOP_TRYLOCK) == 0) 1046 panic("non-try upgrade of lock (%s) %s @ %s:%d", class->lc_name, 1047 lock->lo_name, file, line); 1048 if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0) 1049 panic("upgrade of non-sleep lock (%s) %s @ %s:%d", 1050 class->lc_name, lock->lo_name, file, line); 1051 instance = find_instance(curthread->td_sleeplocks, lock); 1052 if (instance == NULL) 1053 panic("upgrade of unlocked lock (%s) %s @ %s:%d", 1054 class->lc_name, lock->lo_name, file, line); 1055 if ((instance->li_flags & LI_EXCLUSIVE) != 0) 1056 panic("upgrade of exclusive lock (%s) %s @ %s:%d", 1057 class->lc_name, lock->lo_name, file, line); 1058 if ((instance->li_flags & LI_RECURSEMASK) != 0) 1059 panic("upgrade of recursed lock (%s) %s r=%d @ %s:%d", 1060 class->lc_name, lock->lo_name, 1061 instance->li_flags & LI_RECURSEMASK, file, line); 1062 instance->li_flags |= LI_EXCLUSIVE; 1063} 1064 1065void 1066witness_downgrade(struct lock_object *lock, int flags, const char *file, 1067 int line) 1068{ 1069 struct lock_instance *instance; 1070 struct lock_class *class; 1071 1072 KASSERT(!witness_cold, ("%s: witness_cold", __func__)); 1073 if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL) 1074 return; 1075 class = lock->lo_class; 1076 file = fixup_filename(file); 1077 if ((lock->lo_flags & LO_UPGRADABLE) == 0) 1078 panic("downgrade of non-upgradable lock (%s) %s @ %s:%d", 1079 class->lc_name, lock->lo_name, file, line); 1080 if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0) 1081 panic("downgrade of non-sleep lock (%s) %s @ %s:%d", 1082 class->lc_name, lock->lo_name, file, line); 1083 instance = find_instance(curthread->td_sleeplocks, lock); 1084 if (instance == NULL) 1085 panic("downgrade of unlocked lock (%s) %s @ %s:%d", 1086 class->lc_name, lock->lo_name, file, line); 1087 if ((instance->li_flags & LI_EXCLUSIVE) == 0) 1088 panic("downgrade of shared lock (%s) %s @ %s:%d", 1089 class->lc_name, lock->lo_name, file, line); 1090 if ((instance->li_flags & LI_RECURSEMASK) != 0) 1091 panic("downgrade of recursed lock (%s) %s r=%d @ %s:%d", 1092 class->lc_name, lock->lo_name, 1093 instance->li_flags & LI_RECURSEMASK, file, line); 1094 instance->li_flags &= ~LI_EXCLUSIVE; 1095} 1096 1097void 1098witness_unlock(struct lock_object *lock, int flags, const char *file, int line) 1099{ 1100 struct lock_list_entry **lock_list, *lle; 1101 struct lock_instance *instance; 1102 struct lock_class *class; 1103 struct thread *td; 1104 register_t s; 1105 int i, j; 1106 1107 if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL || 1108 panicstr != NULL) 1109 return; 1110 td = curthread; 1111 class = lock->lo_class; 1112 file = fixup_filename(file); 1113 1114 /* Find lock instance associated with this lock. */ 1115 if (class->lc_flags & LC_SLEEPLOCK) 1116 lock_list = &td->td_sleeplocks; 1117 else 1118 lock_list = PCPU_PTR(spinlocks); 1119 for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next) 1120 for (i = 0; i < (*lock_list)->ll_count; i++) { 1121 instance = &(*lock_list)->ll_children[i]; 1122 if (instance->li_lock == lock) 1123 goto found; 1124 } 1125 panic("lock (%s) %s not locked @ %s:%d", class->lc_name, lock->lo_name, 1126 file, line); 1127found: 1128 1129 /* First, check for shared/exclusive mismatches. */ 1130 if ((instance->li_flags & LI_EXCLUSIVE) != 0 && 1131 (flags & LOP_EXCLUSIVE) == 0) { 1132 printf("shared unlock of (%s) %s @ %s:%d\n", class->lc_name, 1133 lock->lo_name, file, line); 1134 printf("while exclusively locked from %s:%d\n", 1135 instance->li_file, instance->li_line); 1136 panic("excl->ushare"); 1137 } 1138 if ((instance->li_flags & LI_EXCLUSIVE) == 0 && 1139 (flags & LOP_EXCLUSIVE) != 0) { 1140 printf("exclusive unlock of (%s) %s @ %s:%d\n", class->lc_name, 1141 lock->lo_name, file, line); 1142 printf("while share locked from %s:%d\n", instance->li_file, 1143 instance->li_line); 1144 panic("share->uexcl"); 1145 } 1146 1147 /* If we are recursed, unrecurse. */ 1148 if ((instance->li_flags & LI_RECURSEMASK) > 0) { 1149 CTR4(KTR_WITNESS, "%s: pid %d unrecursed on %s r=%d", __func__, 1150 td->td_proc->p_pid, instance->li_lock->lo_name, 1151 instance->li_flags); 1152 instance->li_flags--; 1153 return; 1154 } 1155 1156 /* Otherwise, remove this item from the list. */ 1157 s = intr_disable(); 1158 CTR4(KTR_WITNESS, "%s: pid %d removed %s from lle[%d]", __func__, 1159 td->td_proc->p_pid, instance->li_lock->lo_name, 1160 (*lock_list)->ll_count - 1); 1161 for (j = i; j < (*lock_list)->ll_count - 1; j++) 1162 (*lock_list)->ll_children[j] = 1163 (*lock_list)->ll_children[j + 1]; 1164 (*lock_list)->ll_count--; 1165 intr_restore(s); 1166 1167 /* If this lock list entry is now empty, free it. */ 1168 if ((*lock_list)->ll_count == 0) { 1169 lle = *lock_list; 1170 *lock_list = lle->ll_next; 1171 CTR3(KTR_WITNESS, "%s: pid %d removed lle %p", __func__, 1172 td->td_proc->p_pid, lle); 1173 witness_lock_list_free(lle); 1174 } 1175} 1176 1177/* 1178 * Warn if any locks other than 'lock' are held. Flags can be passed in to 1179 * exempt Giant and sleepable locks from the checks as well. If any 1180 * non-exempt locks are held, then a supplied message is printed to the 1181 * console along with a list of the offending locks. If indicated in the 1182 * flags then a failure results in a panic as well. 1183 */ 1184int 1185witness_warn(int flags, struct lock_object *lock, const char *fmt, ...) 1186{ 1187 struct lock_list_entry *lle; 1188 struct lock_instance *lock1; 1189 struct thread *td; 1190 va_list ap; 1191 int i, n; 1192 1193 if (witness_cold || witness_watch == 0 || panicstr != NULL) 1194 return (0); 1195 n = 0; 1196 td = curthread; 1197 for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next) 1198 for (i = lle->ll_count - 1; i >= 0; i--) { 1199 lock1 = &lle->ll_children[i]; 1200 if (lock1->li_lock == lock) 1201 continue; 1202 if (flags & WARN_GIANTOK && 1203 lock1->li_lock == &Giant.mtx_object) 1204 continue; 1205 if (flags & WARN_SLEEPOK && 1206 (lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0) 1207 continue; 1208 if (n == 0) { 1209 va_start(ap, fmt); 1210 vprintf(fmt, ap); 1211 va_end(ap); 1212 printf(" with the following"); 1213 if (flags & WARN_SLEEPOK) 1214 printf(" non-sleepable"); 1215 printf(" locks held:\n"); 1216 } 1217 n++; 1218 witness_list_lock(lock1); 1219 } 1220 if (PCPU_GET(spinlocks) != NULL) { 1221 /* 1222 * Since we already hold a spinlock preemption is 1223 * already blocked. 1224 */ 1225 if (n == 0) { 1226 va_start(ap, fmt); 1227 vprintf(fmt, ap); 1228 va_end(ap); 1229 printf(" with the following"); 1230 if (flags & WARN_SLEEPOK) 1231 printf(" non-sleepable"); 1232 printf(" locks held:\n"); 1233 } 1234 n += witness_list_locks(PCPU_PTR(spinlocks)); 1235 } 1236 if (flags & WARN_PANIC && n) 1237 panic("witness_warn"); 1238#ifdef KDB 1239 else if (witness_kdb && n) 1240 kdb_enter(__func__); 1241 else if (witness_trace && n) 1242 kdb_backtrace(); 1243#endif 1244 return (n); 1245} 1246 1247const char * 1248witness_file(struct lock_object *lock) 1249{ 1250 struct witness *w; 1251 1252 if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL) 1253 return ("?"); 1254 w = lock->lo_witness; 1255 return (w->w_file); 1256} 1257 1258int 1259witness_line(struct lock_object *lock) 1260{ 1261 struct witness *w; 1262 1263 if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL) 1264 return (0); 1265 w = lock->lo_witness; 1266 return (w->w_line); 1267} 1268 1269static struct witness * 1270enroll(const char *description, struct lock_class *lock_class) 1271{ 1272 struct witness *w; 1273 1274 if (witness_watch == 0 || panicstr != NULL) 1275 return (NULL); 1276 if ((lock_class->lc_flags & LC_SPINLOCK) && witness_skipspin) 1277 return (NULL); 1278 mtx_lock_spin(&w_mtx); 1279 STAILQ_FOREACH(w, &w_all, w_list) { 1280 if (w->w_name == description || (w->w_refcount > 0 && 1281 strcmp(description, w->w_name) == 0)) { 1282 w->w_refcount++; 1283 mtx_unlock_spin(&w_mtx); 1284 if (lock_class != w->w_class) 1285 panic( 1286 "lock (%s) %s does not match earlier (%s) lock", 1287 description, lock_class->lc_name, 1288 w->w_class->lc_name); 1289 return (w); 1290 } 1291 } 1292 /* 1293 * This isn't quite right, as witness_cold is still 0 while we 1294 * enroll all the locks initialized before witness_initialize(). 1295 */ 1296 if ((lock_class->lc_flags & LC_SPINLOCK) && !witness_cold) { 1297 mtx_unlock_spin(&w_mtx); 1298 panic("spin lock %s not in order list", description); 1299 } 1300 if ((w = witness_get()) == NULL) 1301 return (NULL); 1302 w->w_name = description; 1303 w->w_class = lock_class; 1304 w->w_refcount = 1; 1305 STAILQ_INSERT_HEAD(&w_all, w, w_list); 1306 if (lock_class->lc_flags & LC_SPINLOCK) 1307 STAILQ_INSERT_HEAD(&w_spin, w, w_typelist); 1308 else if (lock_class->lc_flags & LC_SLEEPLOCK) 1309 STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist); 1310 else { 1311 mtx_unlock_spin(&w_mtx); 1312 panic("lock class %s is not sleep or spin", 1313 lock_class->lc_name); 1314 } 1315 mtx_unlock_spin(&w_mtx); 1316 return (w); 1317} 1318 1319/* Don't let the door bang you on the way out... */ 1320static int 1321depart(struct witness *w) 1322{ 1323 struct witness_child_list_entry *wcl, *nwcl; 1324 struct witness_list *list; 1325 struct witness *parent; 1326 1327 MPASS(w->w_refcount == 0); 1328 if (w->w_class->lc_flags & LC_SLEEPLOCK) 1329 list = &w_sleep; 1330 else 1331 list = &w_spin; 1332 /* 1333 * First, we run through the entire tree looking for any 1334 * witnesses that the outgoing witness is a child of. For 1335 * each parent that we find, we reparent all the direct 1336 * children of the outgoing witness to its parent. 1337 */ 1338 STAILQ_FOREACH(parent, list, w_typelist) { 1339 if (!isitmychild(parent, w)) 1340 continue; 1341 removechild(parent, w); 1342 if (!reparentchildren(parent, w)) 1343 return (0); 1344 } 1345 1346 /* 1347 * Now we go through and free up the child list of the 1348 * outgoing witness. 1349 */ 1350 for (wcl = w->w_children; wcl != NULL; wcl = nwcl) { 1351 nwcl = wcl->wcl_next; 1352 witness_child_free(wcl); 1353 } 1354 1355 /* 1356 * Detach from various lists and free. 1357 */ 1358 STAILQ_REMOVE(list, w, witness, w_typelist); 1359 STAILQ_REMOVE(&w_all, w, witness, w_list); 1360 witness_free(w); 1361 1362 /* Finally, fixup the tree. */ 1363 return (rebalancetree(list)); 1364} 1365 1366/* 1367 * Prune an entire lock order tree. We look for cases where a lock 1368 * is now both a descendant and a direct child of a given lock. In 1369 * that case, we want to remove the direct child link from the tree. 1370 * 1371 * Returns false if insertchild() fails. 1372 */ 1373static int 1374rebalancetree(struct witness_list *list) 1375{ 1376 struct witness *child, *parent; 1377 1378 STAILQ_FOREACH(child, list, w_typelist) { 1379 STAILQ_FOREACH(parent, list, w_typelist) { 1380 if (!isitmychild(parent, child)) 1381 continue; 1382 removechild(parent, child); 1383 if (isitmydescendant(parent, child)) 1384 continue; 1385 if (!insertchild(parent, child)) 1386 return (0); 1387 } 1388 } 1389 witness_levelall(); 1390 return (1); 1391} 1392 1393/* 1394 * Add "child" as a direct child of "parent". Returns false if 1395 * we fail due to out of memory. 1396 */ 1397static int 1398insertchild(struct witness *parent, struct witness *child) 1399{ 1400 struct witness_child_list_entry **wcl; 1401 1402 MPASS(child != NULL && parent != NULL); 1403 1404 /* 1405 * Insert "child" after "parent" 1406 */ 1407 wcl = &parent->w_children; 1408 while (*wcl != NULL && (*wcl)->wcl_count == WITNESS_NCHILDREN) 1409 wcl = &(*wcl)->wcl_next; 1410 if (*wcl == NULL) { 1411 *wcl = witness_child_get(); 1412 if (*wcl == NULL) 1413 return (0); 1414 } 1415 (*wcl)->wcl_children[(*wcl)->wcl_count++] = child; 1416 1417 return (1); 1418} 1419 1420/* 1421 * Make all the direct descendants of oldparent be direct descendants 1422 * of newparent. 1423 */ 1424static int 1425reparentchildren(struct witness *newparent, struct witness *oldparent) 1426{ 1427 struct witness_child_list_entry *wcl; 1428 int i; 1429 1430 /* Avoid making a witness a child of itself. */ 1431 MPASS(!isitmychild(oldparent, newparent)); 1432 1433 for (wcl = oldparent->w_children; wcl != NULL; wcl = wcl->wcl_next) 1434 for (i = 0; i < wcl->wcl_count; i++) 1435 if (!insertchild(newparent, wcl->wcl_children[i])) 1436 return (0); 1437 return (1); 1438} 1439 1440static int 1441itismychild(struct witness *parent, struct witness *child) 1442{ 1443 struct witness_list *list; 1444 1445 MPASS(child != NULL && parent != NULL); 1446 if ((parent->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) != 1447 (child->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK))) 1448 panic( 1449 "%s: parent (%s) and child (%s) are not the same lock type", 1450 __func__, parent->w_class->lc_name, 1451 child->w_class->lc_name); 1452 1453 if (!insertchild(parent, child)) 1454 return (0); 1455 1456 if (parent->w_class->lc_flags & LC_SLEEPLOCK) 1457 list = &w_sleep; 1458 else 1459 list = &w_spin; 1460 return (rebalancetree(list)); 1461} 1462 1463static void 1464removechild(struct witness *parent, struct witness *child) 1465{ 1466 struct witness_child_list_entry **wcl, *wcl1; 1467 int i; 1468 1469 for (wcl = &parent->w_children; *wcl != NULL; wcl = &(*wcl)->wcl_next) 1470 for (i = 0; i < (*wcl)->wcl_count; i++) 1471 if ((*wcl)->wcl_children[i] == child) 1472 goto found; 1473 return; 1474found: 1475 (*wcl)->wcl_count--; 1476 if ((*wcl)->wcl_count > i) 1477 (*wcl)->wcl_children[i] = 1478 (*wcl)->wcl_children[(*wcl)->wcl_count]; 1479 MPASS((*wcl)->wcl_children[i] != NULL); 1480 if ((*wcl)->wcl_count != 0) 1481 return; 1482 wcl1 = *wcl; 1483 *wcl = wcl1->wcl_next; 1484 witness_child_free(wcl1); 1485} 1486 1487static int 1488isitmychild(struct witness *parent, struct witness *child) 1489{ 1490 struct witness_child_list_entry *wcl; 1491 int i; 1492 1493 for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) { 1494 for (i = 0; i < wcl->wcl_count; i++) { 1495 if (wcl->wcl_children[i] == child) 1496 return (1); 1497 } 1498 } 1499 return (0); 1500} 1501 1502static int 1503isitmydescendant(struct witness *parent, struct witness *child) 1504{ 1505 struct witness_child_list_entry *wcl; 1506 int i, j; 1507 1508 if (isitmychild(parent, child)) 1509 return (1); 1510 j = 0; 1511 for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) { 1512 MPASS(j < 1000); 1513 for (i = 0; i < wcl->wcl_count; i++) { 1514 if (isitmydescendant(wcl->wcl_children[i], child)) 1515 return (1); 1516 } 1517 j++; 1518 } 1519 return (0); 1520} 1521 1522static void 1523witness_levelall (void) 1524{ 1525 struct witness_list *list; 1526 struct witness *w, *w1; 1527 1528 /* 1529 * First clear all levels. 1530 */ 1531 STAILQ_FOREACH(w, &w_all, w_list) { 1532 w->w_level = 0; 1533 } 1534 1535 /* 1536 * Look for locks with no parent and level all their descendants. 1537 */ 1538 STAILQ_FOREACH(w, &w_all, w_list) { 1539 /* 1540 * This is just an optimization, technically we could get 1541 * away just walking the all list each time. 1542 */ 1543 if (w->w_class->lc_flags & LC_SLEEPLOCK) 1544 list = &w_sleep; 1545 else 1546 list = &w_spin; 1547 STAILQ_FOREACH(w1, list, w_typelist) { 1548 if (isitmychild(w1, w)) 1549 goto skip; 1550 } 1551 witness_leveldescendents(w, 0); 1552 skip: 1553 ; /* silence GCC 3.x */ 1554 } 1555} 1556 1557static void 1558witness_leveldescendents(struct witness *parent, int level) 1559{ 1560 struct witness_child_list_entry *wcl; 1561 int i; 1562 1563 if (parent->w_level < level) 1564 parent->w_level = level; 1565 level++; 1566 for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) 1567 for (i = 0; i < wcl->wcl_count; i++) 1568 witness_leveldescendents(wcl->wcl_children[i], level); 1569} 1570 1571static void 1572witness_displaydescendants(void(*prnt)(const char *fmt, ...), 1573 struct witness *parent, int indent) 1574{ 1575 struct witness_child_list_entry *wcl; 1576 int i, level; 1577 1578 level = parent->w_level; 1579 prnt("%-2d", level); 1580 for (i = 0; i < indent; i++) 1581 prnt(" "); 1582 if (parent->w_refcount > 0) 1583 prnt("%s", parent->w_name); 1584 else 1585 prnt("(dead)"); 1586 if (parent->w_displayed) { 1587 prnt(" -- (already displayed)\n"); 1588 return; 1589 } 1590 parent->w_displayed = 1; 1591 if (parent->w_refcount > 0) { 1592 if (parent->w_file != NULL) 1593 prnt(" -- last acquired @ %s:%d", parent->w_file, 1594 parent->w_line); 1595 } 1596 prnt("\n"); 1597 for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) 1598 for (i = 0; i < wcl->wcl_count; i++) 1599 witness_displaydescendants(prnt, 1600 wcl->wcl_children[i], indent + 1); 1601} 1602 1603#ifdef BLESSING 1604static int 1605blessed(struct witness *w1, struct witness *w2) 1606{ 1607 int i; 1608 struct witness_blessed *b; 1609 1610 for (i = 0; i < blessed_count; i++) { 1611 b = &blessed_list[i]; 1612 if (strcmp(w1->w_name, b->b_lock1) == 0) { 1613 if (strcmp(w2->w_name, b->b_lock2) == 0) 1614 return (1); 1615 continue; 1616 } 1617 if (strcmp(w1->w_name, b->b_lock2) == 0) 1618 if (strcmp(w2->w_name, b->b_lock1) == 0) 1619 return (1); 1620 } 1621 return (0); 1622} 1623#endif 1624 1625static struct witness * 1626witness_get(void) 1627{ 1628 struct witness *w; 1629 1630 if (witness_watch == 0) { 1631 mtx_unlock_spin(&w_mtx); 1632 return (NULL); 1633 } 1634 if (STAILQ_EMPTY(&w_free)) { 1635 witness_watch = 0; 1636 mtx_unlock_spin(&w_mtx); 1637 printf("%s: witness exhausted\n", __func__); 1638 return (NULL); 1639 } 1640 w = STAILQ_FIRST(&w_free); 1641 STAILQ_REMOVE_HEAD(&w_free, w_list); 1642 bzero(w, sizeof(*w)); 1643 return (w); 1644} 1645 1646static void 1647witness_free(struct witness *w) 1648{ 1649 1650 STAILQ_INSERT_HEAD(&w_free, w, w_list); 1651} 1652 1653static struct witness_child_list_entry * 1654witness_child_get(void) 1655{ 1656 struct witness_child_list_entry *wcl; 1657 1658 if (witness_watch == 0) { 1659 mtx_unlock_spin(&w_mtx); 1660 return (NULL); 1661 } 1662 wcl = w_child_free; 1663 if (wcl == NULL) { 1664 witness_watch = 0; 1665 mtx_unlock_spin(&w_mtx); 1666 printf("%s: witness exhausted\n", __func__); 1667 return (NULL); 1668 } 1669 w_child_free = wcl->wcl_next; 1670 bzero(wcl, sizeof(*wcl)); 1671 return (wcl); 1672} 1673 1674static void 1675witness_child_free(struct witness_child_list_entry *wcl) 1676{ 1677 1678 wcl->wcl_next = w_child_free; 1679 w_child_free = wcl; 1680} 1681 1682static struct lock_list_entry * 1683witness_lock_list_get(void) 1684{ 1685 struct lock_list_entry *lle; 1686 1687 if (witness_watch == 0) 1688 return (NULL); 1689 mtx_lock_spin(&w_mtx); 1690 lle = w_lock_list_free; 1691 if (lle == NULL) { 1692 witness_watch = 0; 1693 mtx_unlock_spin(&w_mtx); 1694 printf("%s: witness exhausted\n", __func__); 1695 return (NULL); 1696 } 1697 w_lock_list_free = lle->ll_next; 1698 mtx_unlock_spin(&w_mtx); 1699 bzero(lle, sizeof(*lle)); 1700 return (lle); 1701} 1702 1703static void 1704witness_lock_list_free(struct lock_list_entry *lle) 1705{ 1706 1707 mtx_lock_spin(&w_mtx); 1708 lle->ll_next = w_lock_list_free; 1709 w_lock_list_free = lle; 1710 mtx_unlock_spin(&w_mtx); 1711} 1712 1713static struct lock_instance * 1714find_instance(struct lock_list_entry *lock_list, struct lock_object *lock) 1715{ 1716 struct lock_list_entry *lle; 1717 struct lock_instance *instance; 1718 int i; 1719 1720 for (lle = lock_list; lle != NULL; lle = lle->ll_next) 1721 for (i = lle->ll_count - 1; i >= 0; i--) { 1722 instance = &lle->ll_children[i]; 1723 if (instance->li_lock == lock) 1724 return (instance); 1725 } 1726 return (NULL); 1727} 1728 1729static void 1730witness_list_lock(struct lock_instance *instance) 1731{ 1732 struct lock_object *lock; 1733 1734 lock = instance->li_lock; 1735 printf("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ? 1736 "exclusive" : "shared", lock->lo_class->lc_name, lock->lo_name); 1737 if (lock->lo_type != lock->lo_name) 1738 printf(" (%s)", lock->lo_type); 1739 printf(" r = %d (%p) locked @ %s:%d\n", 1740 instance->li_flags & LI_RECURSEMASK, lock, instance->li_file, 1741 instance->li_line); 1742} 1743 1744#ifdef DDB 1745static int 1746witness_thread_has_locks(struct thread *td) 1747{ 1748 1749 return (td->td_sleeplocks != NULL); 1750} 1751 1752static int 1753witness_proc_has_locks(struct proc *p) 1754{ 1755 struct thread *td; 1756 1757 FOREACH_THREAD_IN_PROC(p, td) { 1758 if (witness_thread_has_locks(td)) 1759 return (1); 1760 } 1761 return (0); 1762} 1763#endif 1764 1765int 1766witness_list_locks(struct lock_list_entry **lock_list) 1767{ 1768 struct lock_list_entry *lle; 1769 int i, nheld; 1770 1771 nheld = 0; 1772 for (lle = *lock_list; lle != NULL; lle = lle->ll_next) 1773 for (i = lle->ll_count - 1; i >= 0; i--) { 1774 witness_list_lock(&lle->ll_children[i]); 1775 nheld++; 1776 } 1777 return (nheld); 1778} 1779 1780/* 1781 * This is a bit risky at best. We call this function when we have timed 1782 * out acquiring a spin lock, and we assume that the other CPU is stuck 1783 * with this lock held. So, we go groveling around in the other CPU's 1784 * per-cpu data to try to find the lock instance for this spin lock to 1785 * see when it was last acquired. 1786 */ 1787void 1788witness_display_spinlock(struct lock_object *lock, struct thread *owner) 1789{ 1790 struct lock_instance *instance; 1791 struct pcpu *pc; 1792 1793 if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU) 1794 return; 1795 pc = pcpu_find(owner->td_oncpu); 1796 instance = find_instance(pc->pc_spinlocks, lock); 1797 if (instance != NULL) 1798 witness_list_lock(instance); 1799} 1800 1801void 1802witness_save(struct lock_object *lock, const char **filep, int *linep) 1803{ 1804 struct lock_instance *instance; 1805 1806 KASSERT(!witness_cold, ("%s: witness_cold", __func__)); 1807 if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL) 1808 return; 1809 if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0) 1810 panic("%s: lock (%s) %s is not a sleep lock", __func__, 1811 lock->lo_class->lc_name, lock->lo_name); 1812 instance = find_instance(curthread->td_sleeplocks, lock); 1813 if (instance == NULL) 1814 panic("%s: lock (%s) %s not locked", __func__, 1815 lock->lo_class->lc_name, lock->lo_name); 1816 *filep = instance->li_file; 1817 *linep = instance->li_line; 1818} 1819 1820void 1821witness_restore(struct lock_object *lock, const char *file, int line) 1822{ 1823 struct lock_instance *instance; 1824 1825 KASSERT(!witness_cold, ("%s: witness_cold", __func__)); 1826 if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL) 1827 return; 1828 if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) == 0) 1829 panic("%s: lock (%s) %s is not a sleep lock", __func__, 1830 lock->lo_class->lc_name, lock->lo_name); 1831 instance = find_instance(curthread->td_sleeplocks, lock); 1832 if (instance == NULL) 1833 panic("%s: lock (%s) %s not locked", __func__, 1834 lock->lo_class->lc_name, lock->lo_name); 1835 lock->lo_witness->w_file = file; 1836 lock->lo_witness->w_line = line; 1837 instance->li_file = file; 1838 instance->li_line = line; 1839} 1840 1841void 1842witness_assert(struct lock_object *lock, int flags, const char *file, int line) 1843{ 1844#ifdef INVARIANT_SUPPORT 1845 struct lock_instance *instance; 1846 1847 if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL) 1848 return; 1849 if ((lock->lo_class->lc_flags & LC_SLEEPLOCK) != 0) 1850 instance = find_instance(curthread->td_sleeplocks, lock); 1851 else if ((lock->lo_class->lc_flags & LC_SPINLOCK) != 0) 1852 instance = find_instance(PCPU_GET(spinlocks), lock); 1853 else { 1854 panic("Lock (%s) %s is not sleep or spin!", 1855 lock->lo_class->lc_name, lock->lo_name); 1856 } 1857 file = fixup_filename(file); 1858 switch (flags) { 1859 case LA_UNLOCKED: 1860 if (instance != NULL) 1861 panic("Lock (%s) %s locked @ %s:%d.", 1862 lock->lo_class->lc_name, lock->lo_name, file, line); 1863 break; 1864 case LA_LOCKED: 1865 case LA_LOCKED | LA_RECURSED: 1866 case LA_LOCKED | LA_NOTRECURSED: 1867 case LA_SLOCKED: 1868 case LA_SLOCKED | LA_RECURSED: 1869 case LA_SLOCKED | LA_NOTRECURSED: 1870 case LA_XLOCKED: 1871 case LA_XLOCKED | LA_RECURSED: 1872 case LA_XLOCKED | LA_NOTRECURSED: 1873 if (instance == NULL) { 1874 panic("Lock (%s) %s not locked @ %s:%d.", 1875 lock->lo_class->lc_name, lock->lo_name, file, line); 1876 break; 1877 } 1878 if ((flags & LA_XLOCKED) != 0 && 1879 (instance->li_flags & LI_EXCLUSIVE) == 0) 1880 panic("Lock (%s) %s not exclusively locked @ %s:%d.", 1881 lock->lo_class->lc_name, lock->lo_name, file, line); 1882 if ((flags & LA_SLOCKED) != 0 && 1883 (instance->li_flags & LI_EXCLUSIVE) != 0) 1884 panic("Lock (%s) %s exclusively locked @ %s:%d.", 1885 lock->lo_class->lc_name, lock->lo_name, file, line); 1886 if ((flags & LA_RECURSED) != 0 && 1887 (instance->li_flags & LI_RECURSEMASK) == 0) 1888 panic("Lock (%s) %s not recursed @ %s:%d.", 1889 lock->lo_class->lc_name, lock->lo_name, file, line); 1890 if ((flags & LA_NOTRECURSED) != 0 && 1891 (instance->li_flags & LI_RECURSEMASK) != 0) 1892 panic("Lock (%s) %s recursed @ %s:%d.", 1893 lock->lo_class->lc_name, lock->lo_name, file, line); 1894 break; 1895 default: 1896 panic("Invalid lock assertion at %s:%d.", file, line); 1897 1898 } 1899#endif /* INVARIANT_SUPPORT */ 1900} 1901 1902#ifdef DDB 1903static void 1904witness_list(struct thread *td) 1905{ 1906 1907 KASSERT(!witness_cold, ("%s: witness_cold", __func__)); 1908 KASSERT(kdb_active, ("%s: not in the debugger", __func__)); 1909 1910 if (witness_watch == 0) 1911 return; 1912 1913 witness_list_locks(&td->td_sleeplocks); 1914 1915 /* 1916 * We only handle spinlocks if td == curthread. This is somewhat broken 1917 * if td is currently executing on some other CPU and holds spin locks 1918 * as we won't display those locks. If we had a MI way of getting 1919 * the per-cpu data for a given cpu then we could use 1920 * td->td_oncpu to get the list of spinlocks for this thread 1921 * and "fix" this. 1922 * 1923 * That still wouldn't really fix this unless we locked sched_lock 1924 * or stopped the other CPU to make sure it wasn't changing the list 1925 * out from under us. It is probably best to just not try to handle 1926 * threads on other CPU's for now. 1927 */ 1928 if (td == curthread && PCPU_GET(spinlocks) != NULL) 1929 witness_list_locks(PCPU_PTR(spinlocks)); 1930} 1931 1932DB_SHOW_COMMAND(locks, db_witness_list) 1933{ 1934 struct thread *td; 1935 pid_t pid; 1936 struct proc *p; 1937 1938 if (have_addr) { 1939 pid = (addr % 16) + ((addr >> 4) % 16) * 10 + 1940 ((addr >> 8) % 16) * 100 + ((addr >> 12) % 16) * 1000 + 1941 ((addr >> 16) % 16) * 10000; 1942 /* sx_slock(&allproc_lock); */ 1943 FOREACH_PROC_IN_SYSTEM(p) { 1944 if (p->p_pid == pid) 1945 break; 1946 } 1947 /* sx_sunlock(&allproc_lock); */ 1948 if (p == NULL) { 1949 db_printf("pid %d not found\n", pid); 1950 return; 1951 } 1952 FOREACH_THREAD_IN_PROC(p, td) { 1953 witness_list(td); 1954 } 1955 } else { 1956 td = curthread; 1957 witness_list(td); 1958 } 1959} 1960 1961DB_SHOW_COMMAND(alllocks, db_witness_list_all) 1962{ 1963 struct thread *td; 1964 struct proc *p; 1965 1966 /* 1967 * It would be nice to list only threads and processes that actually 1968 * held sleep locks, but that information is currently not exported 1969 * by WITNESS. 1970 */ 1971 FOREACH_PROC_IN_SYSTEM(p) { 1972 if (!witness_proc_has_locks(p)) 1973 continue; 1974 FOREACH_THREAD_IN_PROC(p, td) { 1975 if (!witness_thread_has_locks(td)) 1976 continue; 1977 printf("Process %d (%s) thread %p (%d)\n", p->p_pid, 1978 p->p_comm, td, td->td_tid); 1979 witness_list(td); 1980 } 1981 } 1982} 1983 1984DB_SHOW_COMMAND(witness, db_witness_display) 1985{ 1986 1987 witness_display(db_printf); 1988} 1989#endif 1990