subr_witness.c revision 178841
1/*- 2 * Copyright (c) 1998 Berkeley Software Design, Inc. All rights reserved. 3 * 4 * Redistribution and use in source and binary forms, with or without 5 * modification, are permitted provided that the following conditions 6 * are met: 7 * 1. Redistributions of source code must retain the above copyright 8 * notice, this list of conditions and the following disclaimer. 9 * 2. Redistributions in binary form must reproduce the above copyright 10 * notice, this list of conditions and the following disclaimer in the 11 * documentation and/or other materials provided with the distribution. 12 * 3. Berkeley Software Design Inc's name may not be used to endorse or 13 * promote products derived from this software without specific prior 14 * written permission. 15 * 16 * THIS SOFTWARE IS PROVIDED BY BERKELEY SOFTWARE DESIGN INC ``AS IS'' AND 17 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 18 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 19 * ARE DISCLAIMED. IN NO EVENT SHALL BERKELEY SOFTWARE DESIGN INC BE LIABLE 20 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 21 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 22 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 23 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 24 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 25 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 26 * SUCH DAMAGE. 27 * 28 * from BSDI $Id: mutex_witness.c,v 1.1.2.20 2000/04/27 03:10:27 cp Exp $ 29 * and BSDI $Id: synch_machdep.c,v 2.3.2.39 2000/04/27 03:10:25 cp Exp $ 30 */ 31 32/* 33 * Implementation of the `witness' lock verifier. Originally implemented for 34 * mutexes in BSD/OS. Extended to handle generic lock objects and lock 35 * classes in FreeBSD. 36 */ 37 38/* 39 * Main Entry: witness 40 * Pronunciation: 'wit-n&s 41 * Function: noun 42 * Etymology: Middle English witnesse, from Old English witnes knowledge, 43 * testimony, witness, from 2wit 44 * Date: before 12th century 45 * 1 : attestation of a fact or event : TESTIMONY 46 * 2 : one that gives evidence; specifically : one who testifies in 47 * a cause or before a judicial tribunal 48 * 3 : one asked to be present at a transaction so as to be able to 49 * testify to its having taken place 50 * 4 : one who has personal knowledge of something 51 * 5 a : something serving as evidence or proof : SIGN 52 * b : public affirmation by word or example of usually 53 * religious faith or conviction <the heroic witness to divine 54 * life -- Pilot> 55 * 6 capitalized : a member of the Jehovah's Witnesses 56 */ 57 58/* 59 * Special rules concerning Giant and lock orders: 60 * 61 * 1) Giant must be acquired before any other mutexes. Stated another way, 62 * no other mutex may be held when Giant is acquired. 63 * 64 * 2) Giant must be released when blocking on a sleepable lock. 65 * 66 * This rule is less obvious, but is a result of Giant providing the same 67 * semantics as spl(). Basically, when a thread sleeps, it must release 68 * Giant. When a thread blocks on a sleepable lock, it sleeps. Hence rule 69 * 2). 70 * 71 * 3) Giant may be acquired before or after sleepable locks. 72 * 73 * This rule is also not quite as obvious. Giant may be acquired after 74 * a sleepable lock because it is a non-sleepable lock and non-sleepable 75 * locks may always be acquired while holding a sleepable lock. The second 76 * case, Giant before a sleepable lock, follows from rule 2) above. Suppose 77 * you have two threads T1 and T2 and a sleepable lock X. Suppose that T1 78 * acquires X and blocks on Giant. Then suppose that T2 acquires Giant and 79 * blocks on X. When T2 blocks on X, T2 will release Giant allowing T1 to 80 * execute. Thus, acquiring Giant both before and after a sleepable lock 81 * will not result in a lock order reversal. 82 */ 83 84#include <sys/cdefs.h> 85__FBSDID("$FreeBSD: head/sys/kern/subr_witness.c 178841 2008-05-07 21:41:36Z attilio $"); 86 87#include "opt_ddb.h" 88#include "opt_hwpmc_hooks.h" 89#include "opt_witness.h" 90 91#include <sys/param.h> 92#include <sys/bus.h> 93#include <sys/kdb.h> 94#include <sys/kernel.h> 95#include <sys/ktr.h> 96#include <sys/lock.h> 97#include <sys/malloc.h> 98#include <sys/mutex.h> 99#include <sys/priv.h> 100#include <sys/proc.h> 101#include <sys/sbuf.h> 102#include <sys/sysctl.h> 103#include <sys/systm.h> 104 105#include <ddb/ddb.h> 106 107#include <machine/stdarg.h> 108 109/* Note that these traces do not work with KTR_ALQ. */ 110#if 0 111#define KTR_WITNESS KTR_SUBSYS 112#else 113#define KTR_WITNESS 0 114#endif 115 116/* Easier to stay with the old names. */ 117#define lo_list lo_witness_data.lod_list 118#define lo_witness lo_witness_data.lod_witness 119 120#define LI_RECURSEMASK 0x0000ffff /* Recursion depth of lock instance. */ 121#define LI_EXCLUSIVE 0x00010000 /* Exclusive lock instance. */ 122 123/* Define this to check for blessed mutexes */ 124#undef BLESSING 125 126#define WITNESS_COUNT 1024 127#define WITNESS_CHILDCOUNT (WITNESS_COUNT * 4) 128#define WITNESS_SBUFSIZE 32768 129/* 130 * XXX: This is somewhat bogus, as we assume here that at most 1024 threads 131 * will hold LOCK_NCHILDREN * 2 locks. We handle failure ok, and we should 132 * probably be safe for the most part, but it's still a SWAG. 133 */ 134#define LOCK_CHILDCOUNT (MAXCPU + 1024) * 2 135 136#define WITNESS_NCHILDREN 6 137 138#define LOCK_NCHILDREN 3 139 140struct witness_child_list_entry; 141 142/* 143 * Lock instances. A lock instance is the data associated with a lock while 144 * it is held by witness. For example, a lock instance will hold the 145 * recursion count of a lock. Lock instances are held in lists. Spin locks 146 * are held in a per-cpu list while sleep locks are held in per-thread list. 147 */ 148struct lock_instance { 149 struct lock_object *li_lock; 150 const char *li_file; 151 int li_line; 152 u_int li_flags; /* Recursion count and LI_* flags. */ 153}; 154 155/* 156 * A simple list type used to build the list of locks held by a thread 157 * or CPU. We can't simply embed the list in struct lock_object since a 158 * lock may be held by more than one thread if it is a shared lock. Locks 159 * are added to the head of the list, so we fill up each list entry from 160 * "the back" logically. To ease some of the arithmetic, we actually fill 161 * in each list entry the normal way (children[0] then children[1], etc.) but 162 * when we traverse the list we read children[count-1] as the first entry 163 * down to children[0] as the final entry. 164 */ 165struct lock_list_entry { 166 struct lock_list_entry *ll_next; 167 struct lock_instance ll_children[LOCK_NCHILDREN]; 168 u_int ll_count; 169}; 170 171struct witness { 172 const char *w_name; 173 struct lock_class *w_class; 174 STAILQ_ENTRY(witness) w_list; /* List of all witnesses. */ 175 STAILQ_ENTRY(witness) w_typelist; /* Witnesses of a type. */ 176 struct witness_child_list_entry *w_children; /* Great evilness... */ 177 const char *w_file; 178 int w_line; 179 u_int w_level; 180 u_int w_refcount; 181 u_char w_Giant_squawked:1; 182 u_char w_other_squawked:1; 183 u_char w_same_squawked:1; 184 u_char w_displayed:1; 185}; 186 187struct witness_child_list_entry { 188 struct witness_child_list_entry *wcl_next; 189 struct witness *wcl_children[WITNESS_NCHILDREN]; 190 u_int wcl_count; 191}; 192 193STAILQ_HEAD(witness_list, witness); 194 195#ifdef BLESSING 196struct witness_blessed { 197 const char *b_lock1; 198 const char *b_lock2; 199}; 200#endif 201 202struct witness_order_list_entry { 203 const char *w_name; 204 struct lock_class *w_class; 205}; 206 207#ifdef BLESSING 208static int blessed(struct witness *, struct witness *); 209#endif 210static int depart(struct witness *w); 211static struct witness *enroll(const char *description, 212 struct lock_class *lock_class); 213static int insertchild(struct witness *parent, struct witness *child); 214static int isitmychild(struct witness *parent, struct witness *child); 215static int isitmydescendant(struct witness *parent, struct witness *child); 216static int itismychild(struct witness *parent, struct witness *child); 217static void removechild(struct witness *parent, struct witness *child); 218static int sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS); 219static int sysctl_debug_witness_graphs(SYSCTL_HANDLER_ARGS); 220static const char *fixup_filename(const char *file); 221static void witness_addgraph(struct sbuf *sb, struct witness *w); 222static struct witness *witness_get(void); 223static void witness_free(struct witness *m); 224static struct witness_child_list_entry *witness_child_get(void); 225static void witness_child_free(struct witness_child_list_entry *wcl); 226static struct lock_list_entry *witness_lock_list_get(void); 227static void witness_lock_list_free(struct lock_list_entry *lle); 228static struct lock_instance *find_instance(struct lock_list_entry *lock_list, 229 struct lock_object *lock); 230static void witness_list_lock(struct lock_instance *instance); 231#ifdef DDB 232static void witness_leveldescendents(struct witness *parent, int level); 233static void witness_levelall(void); 234static void witness_displaydescendants(void(*)(const char *fmt, ...), 235 struct witness *, int indent); 236static void witness_display_list(void(*prnt)(const char *fmt, ...), 237 struct witness_list *list); 238static void witness_display(void(*)(const char *fmt, ...)); 239static void witness_list(struct thread *td); 240#endif 241 242SYSCTL_NODE(_debug, OID_AUTO, witness, CTLFLAG_RW, 0, "Witness Locking"); 243 244/* 245 * If set to 0, witness is disabled. If set to a non-zero value, witness 246 * performs full lock order checking for all locks. At runtime, this 247 * value may be set to 0 to turn off witness. witness is not allowed be 248 * turned on once it is turned off, however. 249 */ 250static int witness_watch = 1; 251TUNABLE_INT("debug.witness.watch", &witness_watch); 252SYSCTL_PROC(_debug_witness, OID_AUTO, watch, CTLFLAG_RW | CTLTYPE_INT, NULL, 0, 253 sysctl_debug_witness_watch, "I", "witness is watching lock operations"); 254SYSCTL_PROC(_debug_witness, OID_AUTO, graphs, CTLTYPE_STRING | CTLFLAG_RD, 255 NULL, 0, sysctl_debug_witness_graphs, "A", "Show locks relation graphs"); 256 257#ifdef KDB 258/* 259 * When KDB is enabled and witness_kdb is set to 1, it will cause the system 260 * to drop into kdebug() when: 261 * - a lock hierarchy violation occurs 262 * - locks are held when going to sleep. 263 */ 264#ifdef WITNESS_KDB 265int witness_kdb = 1; 266#else 267int witness_kdb = 0; 268#endif 269TUNABLE_INT("debug.witness.kdb", &witness_kdb); 270SYSCTL_INT(_debug_witness, OID_AUTO, kdb, CTLFLAG_RW, &witness_kdb, 0, ""); 271 272/* 273 * When KDB is enabled and witness_trace is set to 1, it will cause the system 274 * to print a stack trace: 275 * - a lock hierarchy violation occurs 276 * - locks are held when going to sleep. 277 */ 278int witness_trace = 1; 279TUNABLE_INT("debug.witness.trace", &witness_trace); 280SYSCTL_INT(_debug_witness, OID_AUTO, trace, CTLFLAG_RW, &witness_trace, 0, ""); 281#endif /* KDB */ 282 283#ifdef WITNESS_SKIPSPIN 284int witness_skipspin = 1; 285#else 286int witness_skipspin = 0; 287#endif 288TUNABLE_INT("debug.witness.skipspin", &witness_skipspin); 289SYSCTL_INT(_debug_witness, OID_AUTO, skipspin, CTLFLAG_RDTUN, 290 &witness_skipspin, 0, ""); 291 292static struct mtx w_mtx; 293static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free); 294static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all); 295static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin); 296static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep); 297static struct witness_child_list_entry *w_child_free = NULL; 298static struct lock_list_entry *w_lock_list_free = NULL; 299 300static int w_free_cnt, w_spin_cnt, w_sleep_cnt, w_child_free_cnt, w_child_cnt; 301SYSCTL_INT(_debug_witness, OID_AUTO, free_cnt, CTLFLAG_RD, &w_free_cnt, 0, ""); 302SYSCTL_INT(_debug_witness, OID_AUTO, spin_cnt, CTLFLAG_RD, &w_spin_cnt, 0, ""); 303SYSCTL_INT(_debug_witness, OID_AUTO, sleep_cnt, CTLFLAG_RD, &w_sleep_cnt, 0, 304 ""); 305SYSCTL_INT(_debug_witness, OID_AUTO, child_free_cnt, CTLFLAG_RD, 306 &w_child_free_cnt, 0, ""); 307SYSCTL_INT(_debug_witness, OID_AUTO, child_cnt, CTLFLAG_RD, &w_child_cnt, 0, 308 ""); 309 310static struct witness w_data[WITNESS_COUNT]; 311static struct witness_child_list_entry w_childdata[WITNESS_CHILDCOUNT]; 312static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT]; 313 314static struct witness_order_list_entry order_lists[] = { 315 /* 316 * sx locks 317 */ 318 { "proctree", &lock_class_sx }, 319 { "allproc", &lock_class_sx }, 320 { "allprison", &lock_class_sx }, 321 { NULL, NULL }, 322 /* 323 * Various mutexes 324 */ 325 { "Giant", &lock_class_mtx_sleep }, 326 { "pipe mutex", &lock_class_mtx_sleep }, 327 { "sigio lock", &lock_class_mtx_sleep }, 328 { "process group", &lock_class_mtx_sleep }, 329 { "process lock", &lock_class_mtx_sleep }, 330 { "session", &lock_class_mtx_sleep }, 331 { "uidinfo hash", &lock_class_rw }, 332#ifdef HWPMC_HOOKS 333 { "pmc-sleep", &lock_class_mtx_sleep }, 334#endif 335 { NULL, NULL }, 336 /* 337 * Sockets 338 */ 339 { "accept", &lock_class_mtx_sleep }, 340 { "so_snd", &lock_class_mtx_sleep }, 341 { "so_rcv", &lock_class_mtx_sleep }, 342 { "sellck", &lock_class_mtx_sleep }, 343 { NULL, NULL }, 344 /* 345 * Routing 346 */ 347 { "so_rcv", &lock_class_mtx_sleep }, 348 { "radix node head", &lock_class_mtx_sleep }, 349 { "rtentry", &lock_class_mtx_sleep }, 350 { "ifaddr", &lock_class_mtx_sleep }, 351 { NULL, NULL }, 352 /* 353 * Multicast - protocol locks before interface locks, after UDP locks. 354 */ 355 { "udpinp", &lock_class_rw }, 356 { "in_multi_mtx", &lock_class_mtx_sleep }, 357 { "igmp_mtx", &lock_class_mtx_sleep }, 358 { "if_addr_mtx", &lock_class_mtx_sleep }, 359 { NULL, NULL }, 360 /* 361 * UNIX Domain Sockets 362 */ 363 { "unp", &lock_class_mtx_sleep }, 364 { "so_snd", &lock_class_mtx_sleep }, 365 { NULL, NULL }, 366 /* 367 * UDP/IP 368 */ 369 { "udp", &lock_class_rw }, 370 { "udpinp", &lock_class_rw }, 371 { "so_snd", &lock_class_mtx_sleep }, 372 { NULL, NULL }, 373 /* 374 * TCP/IP 375 */ 376 { "tcp", &lock_class_rw }, 377 { "tcpinp", &lock_class_rw }, 378 { "so_snd", &lock_class_mtx_sleep }, 379 { NULL, NULL }, 380 /* 381 * SLIP 382 */ 383 { "slip_mtx", &lock_class_mtx_sleep }, 384 { "slip sc_mtx", &lock_class_mtx_sleep }, 385 { NULL, NULL }, 386 /* 387 * netatalk 388 */ 389 { "ddp_list_mtx", &lock_class_mtx_sleep }, 390 { "ddp_mtx", &lock_class_mtx_sleep }, 391 { NULL, NULL }, 392 /* 393 * BPF 394 */ 395 { "bpf global lock", &lock_class_mtx_sleep }, 396 { "bpf interface lock", &lock_class_mtx_sleep }, 397 { "bpf cdev lock", &lock_class_mtx_sleep }, 398 { NULL, NULL }, 399 /* 400 * NFS server 401 */ 402 { "nfsd_mtx", &lock_class_mtx_sleep }, 403 { "so_snd", &lock_class_mtx_sleep }, 404 { NULL, NULL }, 405 406 /* 407 * IEEE 802.11 408 */ 409 { "802.11 com lock", &lock_class_mtx_sleep}, 410 { NULL, NULL }, 411 /* 412 * Network drivers 413 */ 414 { "network driver", &lock_class_mtx_sleep}, 415 { NULL, NULL }, 416 417 /* 418 * Netgraph 419 */ 420 { "ng_node", &lock_class_mtx_sleep }, 421 { "ng_worklist", &lock_class_mtx_sleep }, 422 { NULL, NULL }, 423 /* 424 * CDEV 425 */ 426 { "system map", &lock_class_mtx_sleep }, 427 { "vm page queue mutex", &lock_class_mtx_sleep }, 428 { "vnode interlock", &lock_class_mtx_sleep }, 429 { "cdev", &lock_class_mtx_sleep }, 430 { NULL, NULL }, 431 /* 432 * kqueue/VFS interaction 433 */ 434 { "kqueue", &lock_class_mtx_sleep }, 435 { "struct mount mtx", &lock_class_mtx_sleep }, 436 { "vnode interlock", &lock_class_mtx_sleep }, 437 { NULL, NULL }, 438 /* 439 * spin locks 440 */ 441#ifdef SMP 442 { "ap boot", &lock_class_mtx_spin }, 443#endif 444 { "rm.mutex_mtx", &lock_class_mtx_spin }, 445 { "sio", &lock_class_mtx_spin }, 446 { "scrlock", &lock_class_mtx_spin }, 447#ifdef __i386__ 448 { "cy", &lock_class_mtx_spin }, 449#endif 450#ifdef __sparc64__ 451 { "pcib_mtx", &lock_class_mtx_spin }, 452 { "rtc_mtx", &lock_class_mtx_spin }, 453#endif 454 { "scc_hwmtx", &lock_class_mtx_spin }, 455 { "uart_hwmtx", &lock_class_mtx_spin }, 456 { "fast_taskqueue", &lock_class_mtx_spin }, 457 { "intr table", &lock_class_mtx_spin }, 458#ifdef HWPMC_HOOKS 459 { "pmc-per-proc", &lock_class_mtx_spin }, 460#endif 461 { "process slock", &lock_class_mtx_spin }, 462 { "sleepq chain", &lock_class_mtx_spin }, 463 { "umtx lock", &lock_class_mtx_spin }, 464 { "rm_spinlock", &lock_class_mtx_spin }, 465 { "turnstile chain", &lock_class_mtx_spin }, 466 { "turnstile lock", &lock_class_mtx_spin }, 467 { "sched lock", &lock_class_mtx_spin }, 468 { "td_contested", &lock_class_mtx_spin }, 469 { "callout", &lock_class_mtx_spin }, 470 { "entropy harvest mutex", &lock_class_mtx_spin }, 471 { "syscons video lock", &lock_class_mtx_spin }, 472 { "time lock", &lock_class_mtx_spin }, 473#ifdef SMP 474 { "smp rendezvous", &lock_class_mtx_spin }, 475#endif 476#ifdef __powerpc__ 477 { "tlb0", &lock_class_mtx_spin }, 478#endif 479 /* 480 * leaf locks 481 */ 482 { "intrcnt", &lock_class_mtx_spin }, 483 { "icu", &lock_class_mtx_spin }, 484#if defined(SMP) && defined(__sparc64__) 485 { "ipi", &lock_class_mtx_spin }, 486#endif 487#ifdef __i386__ 488 { "allpmaps", &lock_class_mtx_spin }, 489 { "descriptor tables", &lock_class_mtx_spin }, 490#endif 491 { "clk", &lock_class_mtx_spin }, 492 { "cpuset", &lock_class_mtx_spin }, 493 { "mprof lock", &lock_class_mtx_spin }, 494 { "zombie lock", &lock_class_mtx_spin }, 495 { "ALD Queue", &lock_class_mtx_spin }, 496#ifdef __ia64__ 497 { "MCA spin lock", &lock_class_mtx_spin }, 498#endif 499#if defined(__i386__) || defined(__amd64__) 500 { "pcicfg", &lock_class_mtx_spin }, 501 { "NDIS thread lock", &lock_class_mtx_spin }, 502#endif 503 { "tw_osl_io_lock", &lock_class_mtx_spin }, 504 { "tw_osl_q_lock", &lock_class_mtx_spin }, 505 { "tw_cl_io_lock", &lock_class_mtx_spin }, 506 { "tw_cl_intr_lock", &lock_class_mtx_spin }, 507 { "tw_cl_gen_lock", &lock_class_mtx_spin }, 508#ifdef HWPMC_HOOKS 509 { "pmc-leaf", &lock_class_mtx_spin }, 510#endif 511 { "blocked lock", &lock_class_mtx_spin }, 512 { NULL, NULL }, 513 { NULL, NULL } 514}; 515 516#ifdef BLESSING 517/* 518 * Pairs of locks which have been blessed 519 * Don't complain about order problems with blessed locks 520 */ 521static struct witness_blessed blessed_list[] = { 522}; 523static int blessed_count = 524 sizeof(blessed_list) / sizeof(struct witness_blessed); 525#endif 526 527/* 528 * List of locks initialized prior to witness being initialized whose 529 * enrollment is currently deferred. 530 */ 531STAILQ_HEAD(, lock_object) pending_locks = 532 STAILQ_HEAD_INITIALIZER(pending_locks); 533 534/* 535 * This global is set to 0 once it becomes safe to use the witness code. 536 */ 537static int witness_cold = 1; 538 539/* 540 * This global is set to 1 once the static lock orders have been enrolled 541 * so that a warning can be issued for any spin locks enrolled later. 542 */ 543static int witness_spin_warn = 0; 544 545/* 546 * The WITNESS-enabled diagnostic code. Note that the witness code does 547 * assume that the early boot is single-threaded at least until after this 548 * routine is completed. 549 */ 550static void 551witness_initialize(void *dummy __unused) 552{ 553 struct lock_object *lock; 554 struct witness_order_list_entry *order; 555 struct witness *w, *w1; 556 int i; 557 558 /* 559 * We have to release Giant before initializing its witness 560 * structure so that WITNESS doesn't get confused. 561 */ 562 mtx_unlock(&Giant); 563 mtx_assert(&Giant, MA_NOTOWNED); 564 565 CTR1(KTR_WITNESS, "%s: initializing witness", __func__); 566 mtx_init(&w_mtx, "witness lock", NULL, MTX_SPIN | MTX_QUIET | 567 MTX_NOWITNESS | MTX_NOPROFILE); 568 for (i = 0; i < WITNESS_COUNT; i++) 569 witness_free(&w_data[i]); 570 for (i = 0; i < WITNESS_CHILDCOUNT; i++) 571 witness_child_free(&w_childdata[i]); 572 for (i = 0; i < LOCK_CHILDCOUNT; i++) 573 witness_lock_list_free(&w_locklistdata[i]); 574 575 /* First add in all the specified order lists. */ 576 for (order = order_lists; order->w_name != NULL; order++) { 577 w = enroll(order->w_name, order->w_class); 578 if (w == NULL) 579 continue; 580 w->w_file = "order list"; 581 for (order++; order->w_name != NULL; order++) { 582 w1 = enroll(order->w_name, order->w_class); 583 if (w1 == NULL) 584 continue; 585 w1->w_file = "order list"; 586 if (!itismychild(w, w1)) 587 panic("Not enough memory for static orders!"); 588 w = w1; 589 } 590 } 591 witness_spin_warn = 1; 592 593 /* Iterate through all locks and add them to witness. */ 594 while (!STAILQ_EMPTY(&pending_locks)) { 595 lock = STAILQ_FIRST(&pending_locks); 596 STAILQ_REMOVE_HEAD(&pending_locks, lo_list); 597 KASSERT(lock->lo_flags & LO_WITNESS, 598 ("%s: lock %s is on pending list but not LO_WITNESS", 599 __func__, lock->lo_name)); 600 lock->lo_witness = enroll(lock->lo_type, LOCK_CLASS(lock)); 601 } 602 603 /* Mark the witness code as being ready for use. */ 604 witness_cold = 0; 605 606 mtx_lock(&Giant); 607} 608SYSINIT(witness_init, SI_SUB_WITNESS, SI_ORDER_FIRST, witness_initialize, 609 NULL); 610 611static int 612sysctl_debug_witness_watch(SYSCTL_HANDLER_ARGS) 613{ 614 int error, value; 615 616 value = witness_watch; 617 error = sysctl_handle_int(oidp, &value, 0, req); 618 if (error != 0 || req->newptr == NULL) 619 return (error); 620 if (value == witness_watch) 621 return (0); 622 if (value != 0) 623 return (EINVAL); 624 witness_watch = 0; 625 return (0); 626} 627 628static int 629sysctl_debug_witness_graphs(SYSCTL_HANDLER_ARGS) 630{ 631 struct witness *w; 632 struct sbuf *sb; 633 int error; 634 635 KASSERT(witness_cold == 0, ("%s: witness is still cold", __func__)); 636 637 sb = sbuf_new(NULL, NULL, WITNESS_SBUFSIZE, SBUF_FIXEDLEN); 638 if (sb == NULL) 639 return (ENOMEM); 640 641 mtx_lock_spin(&w_mtx); 642 STAILQ_FOREACH(w, &w_all, w_list) 643 w->w_displayed = 0; 644 STAILQ_FOREACH(w, &w_all, w_list) 645 witness_addgraph(sb, w); 646 mtx_unlock_spin(&w_mtx); 647 648 if (sbuf_overflowed(sb)) { 649 sbuf_delete(sb); 650 panic("%s: sbuf overflowed, bump the static buffer size\n", 651 __func__); 652 } 653 654 sbuf_finish(sb); 655 error = SYSCTL_OUT(req, sbuf_data(sb), sbuf_len(sb) + 1); 656 sbuf_delete(sb); 657 658 return (error); 659} 660 661void 662witness_init(struct lock_object *lock) 663{ 664 struct lock_class *class; 665 666 /* Various sanity checks. */ 667 class = LOCK_CLASS(lock); 668 if ((lock->lo_flags & LO_RECURSABLE) != 0 && 669 (class->lc_flags & LC_RECURSABLE) == 0) 670 panic("%s: lock (%s) %s can not be recursable", __func__, 671 class->lc_name, lock->lo_name); 672 if ((lock->lo_flags & LO_SLEEPABLE) != 0 && 673 (class->lc_flags & LC_SLEEPABLE) == 0) 674 panic("%s: lock (%s) %s can not be sleepable", __func__, 675 class->lc_name, lock->lo_name); 676 if ((lock->lo_flags & LO_UPGRADABLE) != 0 && 677 (class->lc_flags & LC_UPGRADABLE) == 0) 678 panic("%s: lock (%s) %s can not be upgradable", __func__, 679 class->lc_name, lock->lo_name); 680 681 /* 682 * If we shouldn't watch this lock, then just clear lo_witness. 683 * Otherwise, if witness_cold is set, then it is too early to 684 * enroll this lock, so defer it to witness_initialize() by adding 685 * it to the pending_locks list. If it is not too early, then enroll 686 * the lock now. 687 */ 688 if (witness_watch == 0 || panicstr != NULL || 689 (lock->lo_flags & LO_WITNESS) == 0) 690 lock->lo_witness = NULL; 691 else if (witness_cold) { 692 STAILQ_INSERT_TAIL(&pending_locks, lock, lo_list); 693 lock->lo_flags |= LO_ENROLLPEND; 694 } else 695 lock->lo_witness = enroll(lock->lo_type, class); 696} 697 698void 699witness_destroy(struct lock_object *lock) 700{ 701 struct lock_class *class; 702 struct witness *w; 703 704 class = LOCK_CLASS(lock); 705 if (witness_cold) 706 panic("lock (%s) %s destroyed while witness_cold", 707 class->lc_name, lock->lo_name); 708 709 /* XXX: need to verify that no one holds the lock */ 710 if ((lock->lo_flags & (LO_WITNESS | LO_ENROLLPEND)) == LO_WITNESS && 711 lock->lo_witness != NULL) { 712 w = lock->lo_witness; 713 mtx_lock_spin(&w_mtx); 714 MPASS(w->w_refcount > 0); 715 w->w_refcount--; 716 717 /* 718 * Lock is already released if we have an allocation failure 719 * and depart() fails. 720 */ 721 if (w->w_refcount != 0 || depart(w)) 722 mtx_unlock_spin(&w_mtx); 723 } 724 725 /* 726 * If this lock is destroyed before witness is up and running, 727 * remove it from the pending list. 728 */ 729 if (lock->lo_flags & LO_ENROLLPEND) { 730 STAILQ_REMOVE(&pending_locks, lock, lock_object, lo_list); 731 lock->lo_flags &= ~LO_ENROLLPEND; 732 } 733} 734 735#ifdef DDB 736static void 737witness_levelall (void) 738{ 739 struct witness_list *list; 740 struct witness *w, *w1; 741 742 /* 743 * First clear all levels. 744 */ 745 STAILQ_FOREACH(w, &w_all, w_list) { 746 w->w_level = 0; 747 } 748 749 /* 750 * Look for locks with no parent and level all their descendants. 751 */ 752 STAILQ_FOREACH(w, &w_all, w_list) { 753 /* 754 * This is just an optimization, technically we could get 755 * away just walking the all list each time. 756 */ 757 if (w->w_class->lc_flags & LC_SLEEPLOCK) 758 list = &w_sleep; 759 else 760 list = &w_spin; 761 STAILQ_FOREACH(w1, list, w_typelist) { 762 if (isitmychild(w1, w)) 763 goto skip; 764 } 765 witness_leveldescendents(w, 0); 766 skip: 767 ; /* silence GCC 3.x */ 768 } 769} 770 771static void 772witness_leveldescendents(struct witness *parent, int level) 773{ 774 struct witness_child_list_entry *wcl; 775 int i; 776 777 if (parent->w_level < level) 778 parent->w_level = level; 779 level++; 780 for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) 781 for (i = 0; i < wcl->wcl_count; i++) 782 witness_leveldescendents(wcl->wcl_children[i], level); 783} 784 785static void 786witness_displaydescendants(void(*prnt)(const char *fmt, ...), 787 struct witness *parent, int indent) 788{ 789 struct witness_child_list_entry *wcl; 790 int i, level; 791 792 level = parent->w_level; 793 prnt("%-2d", level); 794 for (i = 0; i < indent; i++) 795 prnt(" "); 796 if (parent->w_refcount > 0) 797 prnt("%s", parent->w_name); 798 else 799 prnt("(dead)"); 800 if (parent->w_displayed) { 801 prnt(" -- (already displayed)\n"); 802 return; 803 } 804 parent->w_displayed = 1; 805 if (parent->w_refcount > 0) { 806 if (parent->w_file != NULL) 807 prnt(" -- last acquired @ %s:%d", parent->w_file, 808 parent->w_line); 809 } 810 prnt("\n"); 811 for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) 812 for (i = 0; i < wcl->wcl_count; i++) 813 witness_displaydescendants(prnt, 814 wcl->wcl_children[i], indent + 1); 815} 816 817static void 818witness_display_list(void(*prnt)(const char *fmt, ...), 819 struct witness_list *list) 820{ 821 struct witness *w; 822 823 STAILQ_FOREACH(w, list, w_typelist) { 824 if (w->w_file == NULL || w->w_level > 0) 825 continue; 826 /* 827 * This lock has no anscestors, display its descendants. 828 */ 829 witness_displaydescendants(prnt, w, 0); 830 } 831} 832 833static void 834witness_addgraph(struct sbuf *sb, struct witness *parent) 835{ 836 struct witness_child_list_entry *wcl; 837 int i; 838 839 if (parent->w_displayed != 0 || parent->w_refcount == 0 || 840 parent->w_file == NULL) 841 return; 842 843 parent->w_displayed = 1; 844 for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) 845 for (i = 0; i < wcl->wcl_count; i++) { 846 sbuf_printf(sb, "\"%s\",\"%s\"\n", parent->w_name, 847 wcl->wcl_children[i]->w_name); 848 witness_addgraph(sb, wcl->wcl_children[i]); 849 } 850} 851 852static void 853witness_display(void(*prnt)(const char *fmt, ...)) 854{ 855 struct witness *w; 856 857 KASSERT(!witness_cold, ("%s: witness_cold", __func__)); 858 witness_levelall(); 859 860 /* Clear all the displayed flags. */ 861 STAILQ_FOREACH(w, &w_all, w_list) { 862 w->w_displayed = 0; 863 } 864 865 /* 866 * First, handle sleep locks which have been acquired at least 867 * once. 868 */ 869 prnt("Sleep locks:\n"); 870 witness_display_list(prnt, &w_sleep); 871 872 /* 873 * Now do spin locks which have been acquired at least once. 874 */ 875 prnt("\nSpin locks:\n"); 876 witness_display_list(prnt, &w_spin); 877 878 /* 879 * Finally, any locks which have not been acquired yet. 880 */ 881 prnt("\nLocks which were never acquired:\n"); 882 STAILQ_FOREACH(w, &w_all, w_list) { 883 if (w->w_file != NULL || w->w_refcount == 0) 884 continue; 885 prnt("%s\n", w->w_name); 886 } 887} 888#endif /* DDB */ 889 890/* Trim useless garbage from filenames. */ 891static const char * 892fixup_filename(const char *file) 893{ 894 895 if (file == NULL) 896 return (NULL); 897 while (strncmp(file, "../", 3) == 0) 898 file += 3; 899 return (file); 900} 901 902int 903witness_defineorder(struct lock_object *lock1, struct lock_object *lock2) 904{ 905 906 if (witness_watch == 0 || panicstr != NULL) 907 return (0); 908 909 /* Require locks that witness knows about. */ 910 if (lock1 == NULL || lock1->lo_witness == NULL || lock2 == NULL || 911 lock2->lo_witness == NULL) 912 return (EINVAL); 913 914 MPASS(!mtx_owned(&w_mtx)); 915 mtx_lock_spin(&w_mtx); 916 917 /* 918 * If we already have either an explicit or implied lock order that 919 * is the other way around, then return an error. 920 */ 921 if (isitmydescendant(lock2->lo_witness, lock1->lo_witness)) { 922 mtx_unlock_spin(&w_mtx); 923 return (EDOOFUS); 924 } 925 926 /* Try to add the new order. */ 927 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__, 928 lock2->lo_type, lock1->lo_type); 929 if (!itismychild(lock1->lo_witness, lock2->lo_witness)) 930 return (ENOMEM); 931 mtx_unlock_spin(&w_mtx); 932 return (0); 933} 934 935void 936witness_checkorder(struct lock_object *lock, int flags, const char *file, 937 int line) 938{ 939 struct lock_list_entry **lock_list, *lle; 940 struct lock_instance *lock1, *lock2; 941 struct lock_class *class; 942 struct witness *w, *w1; 943 struct thread *td; 944 int i, j; 945 946 if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL || 947 panicstr != NULL) 948 return; 949 950 /* 951 * Try locks do not block if they fail to acquire the lock, thus 952 * there is no danger of deadlocks or of switching while holding a 953 * spin lock if we acquire a lock via a try operation. This 954 * function shouldn't even be called for try locks, so panic if 955 * that happens. 956 */ 957 if (flags & LOP_TRYLOCK) 958 panic("%s should not be called for try lock operations", 959 __func__); 960 961 w = lock->lo_witness; 962 class = LOCK_CLASS(lock); 963 td = curthread; 964 file = fixup_filename(file); 965 966 if (class->lc_flags & LC_SLEEPLOCK) { 967 /* 968 * Since spin locks include a critical section, this check 969 * implicitly enforces a lock order of all sleep locks before 970 * all spin locks. 971 */ 972 if (td->td_critnest != 0 && !kdb_active) 973 panic("blockable sleep lock (%s) %s @ %s:%d", 974 class->lc_name, lock->lo_name, file, line); 975 976 /* 977 * If this is the first lock acquired then just return as 978 * no order checking is needed. 979 */ 980 if (td->td_sleeplocks == NULL) 981 return; 982 lock_list = &td->td_sleeplocks; 983 } else { 984 /* 985 * If this is the first lock, just return as no order 986 * checking is needed. We check this in both if clauses 987 * here as unifying the check would require us to use a 988 * critical section to ensure we don't migrate while doing 989 * the check. Note that if this is not the first lock, we 990 * are already in a critical section and are safe for the 991 * rest of the check. 992 */ 993 if (PCPU_GET(spinlocks) == NULL) 994 return; 995 lock_list = PCPU_PTR(spinlocks); 996 } 997 998 /* 999 * Check to see if we are recursing on a lock we already own. If 1000 * so, make sure that we don't mismatch exclusive and shared lock 1001 * acquires. 1002 */ 1003 lock1 = find_instance(*lock_list, lock); 1004 if (lock1 != NULL) { 1005 if ((lock1->li_flags & LI_EXCLUSIVE) != 0 && 1006 (flags & LOP_EXCLUSIVE) == 0) { 1007 printf("shared lock of (%s) %s @ %s:%d\n", 1008 class->lc_name, lock->lo_name, file, line); 1009 printf("while exclusively locked from %s:%d\n", 1010 lock1->li_file, lock1->li_line); 1011 panic("share->excl"); 1012 } 1013 if ((lock1->li_flags & LI_EXCLUSIVE) == 0 && 1014 (flags & LOP_EXCLUSIVE) != 0) { 1015 printf("exclusive lock of (%s) %s @ %s:%d\n", 1016 class->lc_name, lock->lo_name, file, line); 1017 printf("while share locked from %s:%d\n", 1018 lock1->li_file, lock1->li_line); 1019 panic("excl->share"); 1020 } 1021 return; 1022 } 1023 1024 /* 1025 * Check for duplicate locks of the same type. Note that we only 1026 * have to check for this on the last lock we just acquired. Any 1027 * other cases will be caught as lock order violations. 1028 */ 1029 lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1]; 1030 w1 = lock1->li_lock->lo_witness; 1031 if (w1 == w) { 1032 if (w->w_same_squawked || (lock->lo_flags & LO_DUPOK) || 1033 (flags & LOP_DUPOK)) 1034 return; 1035 w->w_same_squawked = 1; 1036 printf("acquiring duplicate lock of same type: \"%s\"\n", 1037 lock->lo_type); 1038 printf(" 1st %s @ %s:%d\n", lock1->li_lock->lo_name, 1039 lock1->li_file, lock1->li_line); 1040 printf(" 2nd %s @ %s:%d\n", lock->lo_name, file, line); 1041#ifdef KDB 1042 goto debugger; 1043#else 1044 return; 1045#endif 1046 } 1047 MPASS(!mtx_owned(&w_mtx)); 1048 mtx_lock_spin(&w_mtx); 1049 /* 1050 * If we know that the the lock we are acquiring comes after 1051 * the lock we most recently acquired in the lock order tree, 1052 * then there is no need for any further checks. 1053 */ 1054 if (isitmychild(w1, w)) { 1055 mtx_unlock_spin(&w_mtx); 1056 return; 1057 } 1058 for (j = 0, lle = *lock_list; lle != NULL; lle = lle->ll_next) { 1059 for (i = lle->ll_count - 1; i >= 0; i--, j++) { 1060 1061 MPASS(j < WITNESS_COUNT); 1062 lock1 = &lle->ll_children[i]; 1063 w1 = lock1->li_lock->lo_witness; 1064 1065 /* 1066 * If this lock doesn't undergo witness checking, 1067 * then skip it. 1068 */ 1069 if (w1 == NULL) { 1070 KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0, 1071 ("lock missing witness structure")); 1072 continue; 1073 } 1074 /* 1075 * If we are locking Giant and this is a sleepable 1076 * lock, then skip it. 1077 */ 1078 if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0 && 1079 lock == &Giant.lock_object) 1080 continue; 1081 /* 1082 * If we are locking a sleepable lock and this lock 1083 * is Giant, then skip it. 1084 */ 1085 if ((lock->lo_flags & LO_SLEEPABLE) != 0 && 1086 lock1->li_lock == &Giant.lock_object) 1087 continue; 1088 /* 1089 * If we are locking a sleepable lock and this lock 1090 * isn't sleepable, we want to treat it as a lock 1091 * order violation to enfore a general lock order of 1092 * sleepable locks before non-sleepable locks. 1093 */ 1094 if (((lock->lo_flags & LO_SLEEPABLE) != 0 && 1095 (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0)) 1096 goto reversal; 1097 /* 1098 * If we are locking Giant and this is a non-sleepable 1099 * lock, then treat it as a reversal. 1100 */ 1101 if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0 && 1102 lock == &Giant.lock_object) 1103 goto reversal; 1104 /* 1105 * Check the lock order hierarchy for a reveresal. 1106 */ 1107 if (!isitmydescendant(w, w1)) 1108 continue; 1109 reversal: 1110 /* 1111 * We have a lock order violation, check to see if it 1112 * is allowed or has already been yelled about. 1113 */ 1114 mtx_unlock_spin(&w_mtx); 1115#ifdef BLESSING 1116 /* 1117 * If the lock order is blessed, just bail. We don't 1118 * look for other lock order violations though, which 1119 * may be a bug. 1120 */ 1121 if (blessed(w, w1)) 1122 return; 1123#endif 1124 if (lock1->li_lock == &Giant.lock_object) { 1125 if (w1->w_Giant_squawked) 1126 return; 1127 else 1128 w1->w_Giant_squawked = 1; 1129 } else { 1130 if (w1->w_other_squawked) 1131 return; 1132 else 1133 w1->w_other_squawked = 1; 1134 } 1135 /* 1136 * Ok, yell about it. 1137 */ 1138 if (((lock->lo_flags & LO_SLEEPABLE) != 0 && 1139 (lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0)) 1140 printf( 1141 "lock order reversal: (sleepable after non-sleepable)\n"); 1142 else if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) == 0 1143 && lock == &Giant.lock_object) 1144 printf( 1145 "lock order reversal: (Giant after non-sleepable)\n"); 1146 else 1147 printf("lock order reversal:\n"); 1148 /* 1149 * Try to locate an earlier lock with 1150 * witness w in our list. 1151 */ 1152 do { 1153 lock2 = &lle->ll_children[i]; 1154 MPASS(lock2->li_lock != NULL); 1155 if (lock2->li_lock->lo_witness == w) 1156 break; 1157 if (i == 0 && lle->ll_next != NULL) { 1158 lle = lle->ll_next; 1159 i = lle->ll_count - 1; 1160 MPASS(i >= 0 && i < LOCK_NCHILDREN); 1161 } else 1162 i--; 1163 } while (i >= 0); 1164 if (i < 0) { 1165 printf(" 1st %p %s (%s) @ %s:%d\n", 1166 lock1->li_lock, lock1->li_lock->lo_name, 1167 lock1->li_lock->lo_type, lock1->li_file, 1168 lock1->li_line); 1169 printf(" 2nd %p %s (%s) @ %s:%d\n", lock, 1170 lock->lo_name, lock->lo_type, file, line); 1171 } else { 1172 printf(" 1st %p %s (%s) @ %s:%d\n", 1173 lock2->li_lock, lock2->li_lock->lo_name, 1174 lock2->li_lock->lo_type, lock2->li_file, 1175 lock2->li_line); 1176 printf(" 2nd %p %s (%s) @ %s:%d\n", 1177 lock1->li_lock, lock1->li_lock->lo_name, 1178 lock1->li_lock->lo_type, lock1->li_file, 1179 lock1->li_line); 1180 printf(" 3rd %p %s (%s) @ %s:%d\n", lock, 1181 lock->lo_name, lock->lo_type, file, line); 1182 } 1183#ifdef KDB 1184 goto debugger; 1185#else 1186 return; 1187#endif 1188 } 1189 } 1190 lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1]; 1191 /* 1192 * If requested, build a new lock order. However, don't build a new 1193 * relationship between a sleepable lock and Giant if it is in the 1194 * wrong direction. The correct lock order is that sleepable locks 1195 * always come before Giant. 1196 */ 1197 if (flags & LOP_NEWORDER && 1198 !(lock1->li_lock == &Giant.lock_object && 1199 (lock->lo_flags & LO_SLEEPABLE) != 0)) { 1200 CTR3(KTR_WITNESS, "%s: adding %s as a child of %s", __func__, 1201 lock->lo_type, lock1->li_lock->lo_type); 1202 if (!itismychild(lock1->li_lock->lo_witness, w)) 1203 /* Witness is dead. */ 1204 return; 1205 } 1206 mtx_unlock_spin(&w_mtx); 1207 return; 1208 1209#ifdef KDB 1210debugger: 1211 if (witness_trace) 1212 kdb_backtrace(); 1213 if (witness_kdb) 1214 kdb_enter(KDB_WHY_WITNESS, __func__); 1215#endif 1216} 1217 1218void 1219witness_lock(struct lock_object *lock, int flags, const char *file, int line) 1220{ 1221 struct lock_list_entry **lock_list, *lle; 1222 struct lock_instance *instance; 1223 struct witness *w; 1224 struct thread *td; 1225 1226 if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL || 1227 panicstr != NULL) 1228 return; 1229 w = lock->lo_witness; 1230 td = curthread; 1231 file = fixup_filename(file); 1232 1233 /* Determine lock list for this lock. */ 1234 if (LOCK_CLASS(lock)->lc_flags & LC_SLEEPLOCK) 1235 lock_list = &td->td_sleeplocks; 1236 else 1237 lock_list = PCPU_PTR(spinlocks); 1238 1239 /* Check to see if we are recursing on a lock we already own. */ 1240 instance = find_instance(*lock_list, lock); 1241 if (instance != NULL) { 1242 instance->li_flags++; 1243 CTR4(KTR_WITNESS, "%s: pid %d recursed on %s r=%d", __func__, 1244 td->td_proc->p_pid, lock->lo_name, 1245 instance->li_flags & LI_RECURSEMASK); 1246 instance->li_file = file; 1247 instance->li_line = line; 1248 return; 1249 } 1250 1251 /* Update per-witness last file and line acquire. */ 1252 w->w_file = file; 1253 w->w_line = line; 1254 1255 /* Find the next open lock instance in the list and fill it. */ 1256 lle = *lock_list; 1257 if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) { 1258 lle = witness_lock_list_get(); 1259 if (lle == NULL) 1260 return; 1261 lle->ll_next = *lock_list; 1262 CTR3(KTR_WITNESS, "%s: pid %d added lle %p", __func__, 1263 td->td_proc->p_pid, lle); 1264 *lock_list = lle; 1265 } 1266 instance = &lle->ll_children[lle->ll_count++]; 1267 instance->li_lock = lock; 1268 instance->li_line = line; 1269 instance->li_file = file; 1270 if ((flags & LOP_EXCLUSIVE) != 0) 1271 instance->li_flags = LI_EXCLUSIVE; 1272 else 1273 instance->li_flags = 0; 1274 CTR4(KTR_WITNESS, "%s: pid %d added %s as lle[%d]", __func__, 1275 td->td_proc->p_pid, lock->lo_name, lle->ll_count - 1); 1276} 1277 1278void 1279witness_upgrade(struct lock_object *lock, int flags, const char *file, int line) 1280{ 1281 struct lock_instance *instance; 1282 struct lock_class *class; 1283 1284 KASSERT(!witness_cold, ("%s: witness_cold", __func__)); 1285 if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL) 1286 return; 1287 class = LOCK_CLASS(lock); 1288 file = fixup_filename(file); 1289 if ((lock->lo_flags & LO_UPGRADABLE) == 0) 1290 panic("upgrade of non-upgradable lock (%s) %s @ %s:%d", 1291 class->lc_name, lock->lo_name, file, line); 1292 if ((class->lc_flags & LC_SLEEPLOCK) == 0) 1293 panic("upgrade of non-sleep lock (%s) %s @ %s:%d", 1294 class->lc_name, lock->lo_name, file, line); 1295 instance = find_instance(curthread->td_sleeplocks, lock); 1296 if (instance == NULL) 1297 panic("upgrade of unlocked lock (%s) %s @ %s:%d", 1298 class->lc_name, lock->lo_name, file, line); 1299 if ((instance->li_flags & LI_EXCLUSIVE) != 0) 1300 panic("upgrade of exclusive lock (%s) %s @ %s:%d", 1301 class->lc_name, lock->lo_name, file, line); 1302 if ((instance->li_flags & LI_RECURSEMASK) != 0) 1303 panic("upgrade of recursed lock (%s) %s r=%d @ %s:%d", 1304 class->lc_name, lock->lo_name, 1305 instance->li_flags & LI_RECURSEMASK, file, line); 1306 instance->li_flags |= LI_EXCLUSIVE; 1307} 1308 1309void 1310witness_downgrade(struct lock_object *lock, int flags, const char *file, 1311 int line) 1312{ 1313 struct lock_instance *instance; 1314 struct lock_class *class; 1315 1316 KASSERT(!witness_cold, ("%s: witness_cold", __func__)); 1317 if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL) 1318 return; 1319 class = LOCK_CLASS(lock); 1320 file = fixup_filename(file); 1321 if ((lock->lo_flags & LO_UPGRADABLE) == 0) 1322 panic("downgrade of non-upgradable lock (%s) %s @ %s:%d", 1323 class->lc_name, lock->lo_name, file, line); 1324 if ((class->lc_flags & LC_SLEEPLOCK) == 0) 1325 panic("downgrade of non-sleep lock (%s) %s @ %s:%d", 1326 class->lc_name, lock->lo_name, file, line); 1327 instance = find_instance(curthread->td_sleeplocks, lock); 1328 if (instance == NULL) 1329 panic("downgrade of unlocked lock (%s) %s @ %s:%d", 1330 class->lc_name, lock->lo_name, file, line); 1331 if ((instance->li_flags & LI_EXCLUSIVE) == 0) 1332 panic("downgrade of shared lock (%s) %s @ %s:%d", 1333 class->lc_name, lock->lo_name, file, line); 1334 if ((instance->li_flags & LI_RECURSEMASK) != 0) 1335 panic("downgrade of recursed lock (%s) %s r=%d @ %s:%d", 1336 class->lc_name, lock->lo_name, 1337 instance->li_flags & LI_RECURSEMASK, file, line); 1338 instance->li_flags &= ~LI_EXCLUSIVE; 1339} 1340 1341void 1342witness_unlock(struct lock_object *lock, int flags, const char *file, int line) 1343{ 1344 struct lock_list_entry **lock_list, *lle; 1345 struct lock_instance *instance; 1346 struct lock_class *class; 1347 struct thread *td; 1348 register_t s; 1349 int i, j; 1350 1351 if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL || 1352 panicstr != NULL) 1353 return; 1354 td = curthread; 1355 class = LOCK_CLASS(lock); 1356 file = fixup_filename(file); 1357 1358 /* Find lock instance associated with this lock. */ 1359 if (class->lc_flags & LC_SLEEPLOCK) 1360 lock_list = &td->td_sleeplocks; 1361 else 1362 lock_list = PCPU_PTR(spinlocks); 1363 for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next) 1364 for (i = 0; i < (*lock_list)->ll_count; i++) { 1365 instance = &(*lock_list)->ll_children[i]; 1366 if (instance->li_lock == lock) 1367 goto found; 1368 } 1369 panic("lock (%s) %s not locked @ %s:%d", class->lc_name, lock->lo_name, 1370 file, line); 1371found: 1372 1373 /* First, check for shared/exclusive mismatches. */ 1374 if ((instance->li_flags & LI_EXCLUSIVE) != 0 && 1375 (flags & LOP_EXCLUSIVE) == 0) { 1376 printf("shared unlock of (%s) %s @ %s:%d\n", class->lc_name, 1377 lock->lo_name, file, line); 1378 printf("while exclusively locked from %s:%d\n", 1379 instance->li_file, instance->li_line); 1380 panic("excl->ushare"); 1381 } 1382 if ((instance->li_flags & LI_EXCLUSIVE) == 0 && 1383 (flags & LOP_EXCLUSIVE) != 0) { 1384 printf("exclusive unlock of (%s) %s @ %s:%d\n", class->lc_name, 1385 lock->lo_name, file, line); 1386 printf("while share locked from %s:%d\n", instance->li_file, 1387 instance->li_line); 1388 panic("share->uexcl"); 1389 } 1390 1391 /* If we are recursed, unrecurse. */ 1392 if ((instance->li_flags & LI_RECURSEMASK) > 0) { 1393 CTR4(KTR_WITNESS, "%s: pid %d unrecursed on %s r=%d", __func__, 1394 td->td_proc->p_pid, instance->li_lock->lo_name, 1395 instance->li_flags); 1396 instance->li_flags--; 1397 return; 1398 } 1399 1400 /* Otherwise, remove this item from the list. */ 1401 s = intr_disable(); 1402 CTR4(KTR_WITNESS, "%s: pid %d removed %s from lle[%d]", __func__, 1403 td->td_proc->p_pid, instance->li_lock->lo_name, 1404 (*lock_list)->ll_count - 1); 1405 for (j = i; j < (*lock_list)->ll_count - 1; j++) 1406 (*lock_list)->ll_children[j] = 1407 (*lock_list)->ll_children[j + 1]; 1408 (*lock_list)->ll_count--; 1409 intr_restore(s); 1410 1411 /* If this lock list entry is now empty, free it. */ 1412 if ((*lock_list)->ll_count == 0) { 1413 lle = *lock_list; 1414 *lock_list = lle->ll_next; 1415 CTR3(KTR_WITNESS, "%s: pid %d removed lle %p", __func__, 1416 td->td_proc->p_pid, lle); 1417 witness_lock_list_free(lle); 1418 } 1419} 1420 1421/* 1422 * Warn if any locks other than 'lock' are held. Flags can be passed in to 1423 * exempt Giant and sleepable locks from the checks as well. If any 1424 * non-exempt locks are held, then a supplied message is printed to the 1425 * console along with a list of the offending locks. If indicated in the 1426 * flags then a failure results in a panic as well. 1427 */ 1428int 1429witness_warn(int flags, struct lock_object *lock, const char *fmt, ...) 1430{ 1431 struct lock_list_entry *lle; 1432 struct lock_instance *lock1; 1433 struct thread *td; 1434 va_list ap; 1435 int i, n; 1436 1437 if (witness_cold || witness_watch == 0 || panicstr != NULL) 1438 return (0); 1439 n = 0; 1440 td = curthread; 1441 for (lle = td->td_sleeplocks; lle != NULL; lle = lle->ll_next) 1442 for (i = lle->ll_count - 1; i >= 0; i--) { 1443 lock1 = &lle->ll_children[i]; 1444 if (lock1->li_lock == lock) 1445 continue; 1446 if (flags & WARN_GIANTOK && 1447 lock1->li_lock == &Giant.lock_object) 1448 continue; 1449 if (flags & WARN_SLEEPOK && 1450 (lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0) 1451 continue; 1452 if (n == 0) { 1453 va_start(ap, fmt); 1454 vprintf(fmt, ap); 1455 va_end(ap); 1456 printf(" with the following"); 1457 if (flags & WARN_SLEEPOK) 1458 printf(" non-sleepable"); 1459 printf(" locks held:\n"); 1460 } 1461 n++; 1462 witness_list_lock(lock1); 1463 } 1464 if (PCPU_GET(spinlocks) != NULL) { 1465 /* 1466 * Since we already hold a spinlock preemption is 1467 * already blocked. 1468 */ 1469 if (n == 0) { 1470 va_start(ap, fmt); 1471 vprintf(fmt, ap); 1472 va_end(ap); 1473 printf(" with the following"); 1474 if (flags & WARN_SLEEPOK) 1475 printf(" non-sleepable"); 1476 printf(" locks held:\n"); 1477 } 1478 n += witness_list_locks(PCPU_PTR(spinlocks)); 1479 } 1480 if (flags & WARN_PANIC && n) 1481 panic("witness_warn"); 1482#ifdef KDB 1483 else if (witness_kdb && n) 1484 kdb_enter(KDB_WHY_WITNESS, __func__); 1485 else if (witness_trace && n) 1486 kdb_backtrace(); 1487#endif 1488 return (n); 1489} 1490 1491const char * 1492witness_file(struct lock_object *lock) 1493{ 1494 struct witness *w; 1495 1496 if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL) 1497 return ("?"); 1498 w = lock->lo_witness; 1499 return (w->w_file); 1500} 1501 1502int 1503witness_line(struct lock_object *lock) 1504{ 1505 struct witness *w; 1506 1507 if (witness_cold || witness_watch == 0 || lock->lo_witness == NULL) 1508 return (0); 1509 w = lock->lo_witness; 1510 return (w->w_line); 1511} 1512 1513static struct witness * 1514enroll(const char *description, struct lock_class *lock_class) 1515{ 1516 struct witness *w; 1517 1518 if (witness_watch == 0 || panicstr != NULL) 1519 return (NULL); 1520 if ((lock_class->lc_flags & LC_SPINLOCK) && witness_skipspin) 1521 return (NULL); 1522 mtx_lock_spin(&w_mtx); 1523 STAILQ_FOREACH(w, &w_all, w_list) { 1524 if (w->w_name == description || (w->w_refcount > 0 && 1525 strcmp(description, w->w_name) == 0)) { 1526 w->w_refcount++; 1527 mtx_unlock_spin(&w_mtx); 1528 if (lock_class != w->w_class) 1529 panic( 1530 "lock (%s) %s does not match earlier (%s) lock", 1531 description, lock_class->lc_name, 1532 w->w_class->lc_name); 1533 return (w); 1534 } 1535 } 1536 if ((w = witness_get()) == NULL) 1537 goto out; 1538 w->w_name = description; 1539 w->w_class = lock_class; 1540 w->w_refcount = 1; 1541 STAILQ_INSERT_HEAD(&w_all, w, w_list); 1542 if (lock_class->lc_flags & LC_SPINLOCK) { 1543 STAILQ_INSERT_HEAD(&w_spin, w, w_typelist); 1544 w_spin_cnt++; 1545 } else if (lock_class->lc_flags & LC_SLEEPLOCK) { 1546 STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist); 1547 w_sleep_cnt++; 1548 } else { 1549 mtx_unlock_spin(&w_mtx); 1550 panic("lock class %s is not sleep or spin", 1551 lock_class->lc_name); 1552 } 1553 mtx_unlock_spin(&w_mtx); 1554out: 1555 /* 1556 * We issue a warning for any spin locks not defined in the static 1557 * order list as a way to discourage their use (folks should really 1558 * be using non-spin mutexes most of the time). However, several 1559 * 3rd part device drivers use spin locks because that is all they 1560 * have available on Windows and Linux and they think that normal 1561 * mutexes are insufficient. 1562 */ 1563 if ((lock_class->lc_flags & LC_SPINLOCK) && witness_spin_warn) 1564 printf("WITNESS: spin lock %s not in order list\n", 1565 description); 1566 return (w); 1567} 1568 1569/* Don't let the door bang you on the way out... */ 1570static int 1571depart(struct witness *w) 1572{ 1573 struct witness_child_list_entry *wcl, *nwcl; 1574 struct witness_list *list; 1575 struct witness *parent; 1576 1577 MPASS(w->w_refcount == 0); 1578 if (w->w_class->lc_flags & LC_SLEEPLOCK) { 1579 list = &w_sleep; 1580 w_sleep_cnt--; 1581 } else { 1582 list = &w_spin; 1583 w_spin_cnt--; 1584 } 1585 /* 1586 * First, we run through the entire tree looking for any 1587 * witnesses that the outgoing witness is a child of. For 1588 * each parent that we find, we reparent all the direct 1589 * children of the outgoing witness to its parent. 1590 */ 1591 STAILQ_FOREACH(parent, list, w_typelist) { 1592 if (!isitmychild(parent, w)) 1593 continue; 1594 removechild(parent, w); 1595 } 1596 1597 /* 1598 * Now we go through and free up the child list of the 1599 * outgoing witness. 1600 */ 1601 for (wcl = w->w_children; wcl != NULL; wcl = nwcl) { 1602 nwcl = wcl->wcl_next; 1603 w_child_cnt--; 1604 witness_child_free(wcl); 1605 } 1606 1607 /* 1608 * Detach from various lists and free. 1609 */ 1610 STAILQ_REMOVE(list, w, witness, w_typelist); 1611 STAILQ_REMOVE(&w_all, w, witness, w_list); 1612 witness_free(w); 1613 1614 return (1); 1615} 1616 1617/* 1618 * Add "child" as a direct child of "parent". Returns false if 1619 * we fail due to out of memory. 1620 */ 1621static int 1622insertchild(struct witness *parent, struct witness *child) 1623{ 1624 struct witness_child_list_entry **wcl; 1625 1626 MPASS(child != NULL && parent != NULL); 1627 1628 /* 1629 * Insert "child" after "parent" 1630 */ 1631 wcl = &parent->w_children; 1632 while (*wcl != NULL && (*wcl)->wcl_count == WITNESS_NCHILDREN) 1633 wcl = &(*wcl)->wcl_next; 1634 if (*wcl == NULL) { 1635 *wcl = witness_child_get(); 1636 if (*wcl == NULL) 1637 return (0); 1638 w_child_cnt++; 1639 } 1640 (*wcl)->wcl_children[(*wcl)->wcl_count++] = child; 1641 1642 return (1); 1643} 1644 1645 1646static int 1647itismychild(struct witness *parent, struct witness *child) 1648{ 1649 struct witness_list *list; 1650 1651 MPASS(child != NULL && parent != NULL); 1652 if ((parent->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) != 1653 (child->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK))) 1654 panic( 1655 "%s: parent (%s) and child (%s) are not the same lock type", 1656 __func__, parent->w_class->lc_name, 1657 child->w_class->lc_name); 1658 1659 if (!insertchild(parent, child)) 1660 return (0); 1661 1662 if (parent->w_class->lc_flags & LC_SLEEPLOCK) 1663 list = &w_sleep; 1664 else 1665 list = &w_spin; 1666 return (1); 1667} 1668 1669static void 1670removechild(struct witness *parent, struct witness *child) 1671{ 1672 struct witness_child_list_entry **wcl, *wcl1; 1673 int i; 1674 1675 for (wcl = &parent->w_children; *wcl != NULL; wcl = &(*wcl)->wcl_next) 1676 for (i = 0; i < (*wcl)->wcl_count; i++) 1677 if ((*wcl)->wcl_children[i] == child) 1678 goto found; 1679 return; 1680found: 1681 (*wcl)->wcl_count--; 1682 if ((*wcl)->wcl_count > i) 1683 (*wcl)->wcl_children[i] = 1684 (*wcl)->wcl_children[(*wcl)->wcl_count]; 1685 MPASS((*wcl)->wcl_children[i] != NULL); 1686 if ((*wcl)->wcl_count != 0) 1687 return; 1688 wcl1 = *wcl; 1689 *wcl = wcl1->wcl_next; 1690 w_child_cnt--; 1691 witness_child_free(wcl1); 1692} 1693 1694static int 1695isitmychild(struct witness *parent, struct witness *child) 1696{ 1697 struct witness_child_list_entry *wcl; 1698 int i; 1699 1700 for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) { 1701 for (i = 0; i < wcl->wcl_count; i++) { 1702 if (wcl->wcl_children[i] == child) 1703 return (1); 1704 } 1705 } 1706 return (0); 1707} 1708 1709static int 1710isitmydescendant(struct witness *parent, struct witness *child) 1711{ 1712 struct witness_child_list_entry *wcl; 1713 int i, j; 1714 1715 if (isitmychild(parent, child)) 1716 return (1); 1717 j = 0; 1718 for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) { 1719 MPASS(j < 1000); 1720 for (i = 0; i < wcl->wcl_count; i++) { 1721 if (isitmydescendant(wcl->wcl_children[i], child)) 1722 return (1); 1723 } 1724 j++; 1725 } 1726 return (0); 1727} 1728 1729#ifdef BLESSING 1730static int 1731blessed(struct witness *w1, struct witness *w2) 1732{ 1733 int i; 1734 struct witness_blessed *b; 1735 1736 for (i = 0; i < blessed_count; i++) { 1737 b = &blessed_list[i]; 1738 if (strcmp(w1->w_name, b->b_lock1) == 0) { 1739 if (strcmp(w2->w_name, b->b_lock2) == 0) 1740 return (1); 1741 continue; 1742 } 1743 if (strcmp(w1->w_name, b->b_lock2) == 0) 1744 if (strcmp(w2->w_name, b->b_lock1) == 0) 1745 return (1); 1746 } 1747 return (0); 1748} 1749#endif 1750 1751static struct witness * 1752witness_get(void) 1753{ 1754 struct witness *w; 1755 1756 if (witness_watch == 0) { 1757 mtx_unlock_spin(&w_mtx); 1758 return (NULL); 1759 } 1760 if (STAILQ_EMPTY(&w_free)) { 1761 witness_watch = 0; 1762 mtx_unlock_spin(&w_mtx); 1763 printf("%s: witness exhausted\n", __func__); 1764 return (NULL); 1765 } 1766 w = STAILQ_FIRST(&w_free); 1767 STAILQ_REMOVE_HEAD(&w_free, w_list); 1768 w_free_cnt--; 1769 bzero(w, sizeof(*w)); 1770 return (w); 1771} 1772 1773static void 1774witness_free(struct witness *w) 1775{ 1776 1777 STAILQ_INSERT_HEAD(&w_free, w, w_list); 1778 w_free_cnt++; 1779} 1780 1781static struct witness_child_list_entry * 1782witness_child_get(void) 1783{ 1784 struct witness_child_list_entry *wcl; 1785 1786 if (witness_watch == 0) { 1787 mtx_unlock_spin(&w_mtx); 1788 return (NULL); 1789 } 1790 wcl = w_child_free; 1791 if (wcl == NULL) { 1792 witness_watch = 0; 1793 mtx_unlock_spin(&w_mtx); 1794 printf("%s: witness exhausted\n", __func__); 1795 return (NULL); 1796 } 1797 w_child_free = wcl->wcl_next; 1798 w_child_free_cnt--; 1799 bzero(wcl, sizeof(*wcl)); 1800 return (wcl); 1801} 1802 1803static void 1804witness_child_free(struct witness_child_list_entry *wcl) 1805{ 1806 1807 wcl->wcl_next = w_child_free; 1808 w_child_free = wcl; 1809 w_child_free_cnt++; 1810} 1811 1812static struct lock_list_entry * 1813witness_lock_list_get(void) 1814{ 1815 struct lock_list_entry *lle; 1816 1817 if (witness_watch == 0) 1818 return (NULL); 1819 mtx_lock_spin(&w_mtx); 1820 lle = w_lock_list_free; 1821 if (lle == NULL) { 1822 witness_watch = 0; 1823 mtx_unlock_spin(&w_mtx); 1824 printf("%s: witness exhausted\n", __func__); 1825 return (NULL); 1826 } 1827 w_lock_list_free = lle->ll_next; 1828 mtx_unlock_spin(&w_mtx); 1829 bzero(lle, sizeof(*lle)); 1830 return (lle); 1831} 1832 1833static void 1834witness_lock_list_free(struct lock_list_entry *lle) 1835{ 1836 1837 mtx_lock_spin(&w_mtx); 1838 lle->ll_next = w_lock_list_free; 1839 w_lock_list_free = lle; 1840 mtx_unlock_spin(&w_mtx); 1841} 1842 1843static struct lock_instance * 1844find_instance(struct lock_list_entry *lock_list, struct lock_object *lock) 1845{ 1846 struct lock_list_entry *lle; 1847 struct lock_instance *instance; 1848 int i; 1849 1850 for (lle = lock_list; lle != NULL; lle = lle->ll_next) 1851 for (i = lle->ll_count - 1; i >= 0; i--) { 1852 instance = &lle->ll_children[i]; 1853 if (instance->li_lock == lock) 1854 return (instance); 1855 } 1856 return (NULL); 1857} 1858 1859static void 1860witness_list_lock(struct lock_instance *instance) 1861{ 1862 struct lock_object *lock; 1863 1864 lock = instance->li_lock; 1865 printf("%s %s %s", (instance->li_flags & LI_EXCLUSIVE) != 0 ? 1866 "exclusive" : "shared", LOCK_CLASS(lock)->lc_name, lock->lo_name); 1867 if (lock->lo_type != lock->lo_name) 1868 printf(" (%s)", lock->lo_type); 1869 printf(" r = %d (%p) locked @ %s:%d\n", 1870 instance->li_flags & LI_RECURSEMASK, lock, instance->li_file, 1871 instance->li_line); 1872} 1873 1874#ifdef DDB 1875static int 1876witness_thread_has_locks(struct thread *td) 1877{ 1878 1879 return (td->td_sleeplocks != NULL); 1880} 1881 1882static int 1883witness_proc_has_locks(struct proc *p) 1884{ 1885 struct thread *td; 1886 1887 FOREACH_THREAD_IN_PROC(p, td) { 1888 if (witness_thread_has_locks(td)) 1889 return (1); 1890 } 1891 return (0); 1892} 1893#endif 1894 1895int 1896witness_list_locks(struct lock_list_entry **lock_list) 1897{ 1898 struct lock_list_entry *lle; 1899 int i, nheld; 1900 1901 nheld = 0; 1902 for (lle = *lock_list; lle != NULL; lle = lle->ll_next) 1903 for (i = lle->ll_count - 1; i >= 0; i--) { 1904 witness_list_lock(&lle->ll_children[i]); 1905 nheld++; 1906 } 1907 return (nheld); 1908} 1909 1910/* 1911 * This is a bit risky at best. We call this function when we have timed 1912 * out acquiring a spin lock, and we assume that the other CPU is stuck 1913 * with this lock held. So, we go groveling around in the other CPU's 1914 * per-cpu data to try to find the lock instance for this spin lock to 1915 * see when it was last acquired. 1916 */ 1917void 1918witness_display_spinlock(struct lock_object *lock, struct thread *owner) 1919{ 1920 struct lock_instance *instance; 1921 struct pcpu *pc; 1922 1923 if (owner->td_critnest == 0 || owner->td_oncpu == NOCPU) 1924 return; 1925 pc = pcpu_find(owner->td_oncpu); 1926 instance = find_instance(pc->pc_spinlocks, lock); 1927 if (instance != NULL) 1928 witness_list_lock(instance); 1929} 1930 1931void 1932witness_save(struct lock_object *lock, const char **filep, int *linep) 1933{ 1934 struct lock_list_entry *lock_list; 1935 struct lock_instance *instance; 1936 struct lock_class *class; 1937 1938 KASSERT(!witness_cold, ("%s: witness_cold", __func__)); 1939 if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL) 1940 return; 1941 class = LOCK_CLASS(lock); 1942 if (class->lc_flags & LC_SLEEPLOCK) 1943 lock_list = curthread->td_sleeplocks; 1944 else { 1945 if (witness_skipspin) 1946 return; 1947 lock_list = PCPU_GET(spinlocks); 1948 } 1949 instance = find_instance(lock_list, lock); 1950 if (instance == NULL) 1951 panic("%s: lock (%s) %s not locked", __func__, 1952 class->lc_name, lock->lo_name); 1953 *filep = instance->li_file; 1954 *linep = instance->li_line; 1955} 1956 1957void 1958witness_restore(struct lock_object *lock, const char *file, int line) 1959{ 1960 struct lock_list_entry *lock_list; 1961 struct lock_instance *instance; 1962 struct lock_class *class; 1963 1964 KASSERT(!witness_cold, ("%s: witness_cold", __func__)); 1965 if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL) 1966 return; 1967 class = LOCK_CLASS(lock); 1968 if (class->lc_flags & LC_SLEEPLOCK) 1969 lock_list = curthread->td_sleeplocks; 1970 else { 1971 if (witness_skipspin) 1972 return; 1973 lock_list = PCPU_GET(spinlocks); 1974 } 1975 instance = find_instance(lock_list, lock); 1976 if (instance == NULL) 1977 panic("%s: lock (%s) %s not locked", __func__, 1978 class->lc_name, lock->lo_name); 1979 lock->lo_witness->w_file = file; 1980 lock->lo_witness->w_line = line; 1981 instance->li_file = file; 1982 instance->li_line = line; 1983} 1984 1985void 1986witness_assert(struct lock_object *lock, int flags, const char *file, int line) 1987{ 1988#ifdef INVARIANT_SUPPORT 1989 struct lock_instance *instance; 1990 struct lock_class *class; 1991 1992 if (lock->lo_witness == NULL || witness_watch == 0 || panicstr != NULL) 1993 return; 1994 class = LOCK_CLASS(lock); 1995 if ((class->lc_flags & LC_SLEEPLOCK) != 0) 1996 instance = find_instance(curthread->td_sleeplocks, lock); 1997 else if ((class->lc_flags & LC_SPINLOCK) != 0) 1998 instance = find_instance(PCPU_GET(spinlocks), lock); 1999 else { 2000 panic("Lock (%s) %s is not sleep or spin!", 2001 class->lc_name, lock->lo_name); 2002 } 2003 file = fixup_filename(file); 2004 switch (flags) { 2005 case LA_UNLOCKED: 2006 if (instance != NULL) 2007 panic("Lock (%s) %s locked @ %s:%d.", 2008 class->lc_name, lock->lo_name, file, line); 2009 break; 2010 case LA_LOCKED: 2011 case LA_LOCKED | LA_RECURSED: 2012 case LA_LOCKED | LA_NOTRECURSED: 2013 case LA_SLOCKED: 2014 case LA_SLOCKED | LA_RECURSED: 2015 case LA_SLOCKED | LA_NOTRECURSED: 2016 case LA_XLOCKED: 2017 case LA_XLOCKED | LA_RECURSED: 2018 case LA_XLOCKED | LA_NOTRECURSED: 2019 if (instance == NULL) { 2020 panic("Lock (%s) %s not locked @ %s:%d.", 2021 class->lc_name, lock->lo_name, file, line); 2022 break; 2023 } 2024 if ((flags & LA_XLOCKED) != 0 && 2025 (instance->li_flags & LI_EXCLUSIVE) == 0) 2026 panic("Lock (%s) %s not exclusively locked @ %s:%d.", 2027 class->lc_name, lock->lo_name, file, line); 2028 if ((flags & LA_SLOCKED) != 0 && 2029 (instance->li_flags & LI_EXCLUSIVE) != 0) 2030 panic("Lock (%s) %s exclusively locked @ %s:%d.", 2031 class->lc_name, lock->lo_name, file, line); 2032 if ((flags & LA_RECURSED) != 0 && 2033 (instance->li_flags & LI_RECURSEMASK) == 0) 2034 panic("Lock (%s) %s not recursed @ %s:%d.", 2035 class->lc_name, lock->lo_name, file, line); 2036 if ((flags & LA_NOTRECURSED) != 0 && 2037 (instance->li_flags & LI_RECURSEMASK) != 0) 2038 panic("Lock (%s) %s recursed @ %s:%d.", 2039 class->lc_name, lock->lo_name, file, line); 2040 break; 2041 default: 2042 panic("Invalid lock assertion at %s:%d.", file, line); 2043 2044 } 2045#endif /* INVARIANT_SUPPORT */ 2046} 2047 2048#ifdef DDB 2049static void 2050witness_list(struct thread *td) 2051{ 2052 2053 KASSERT(!witness_cold, ("%s: witness_cold", __func__)); 2054 KASSERT(kdb_active, ("%s: not in the debugger", __func__)); 2055 2056 if (witness_watch == 0) 2057 return; 2058 2059 witness_list_locks(&td->td_sleeplocks); 2060 2061 /* 2062 * We only handle spinlocks if td == curthread. This is somewhat broken 2063 * if td is currently executing on some other CPU and holds spin locks 2064 * as we won't display those locks. If we had a MI way of getting 2065 * the per-cpu data for a given cpu then we could use 2066 * td->td_oncpu to get the list of spinlocks for this thread 2067 * and "fix" this. 2068 * 2069 * That still wouldn't really fix this unless we locked the scheduler 2070 * lock or stopped the other CPU to make sure it wasn't changing the 2071 * list out from under us. It is probably best to just not try to 2072 * handle threads on other CPU's for now. 2073 */ 2074 if (td == curthread && PCPU_GET(spinlocks) != NULL) 2075 witness_list_locks(PCPU_PTR(spinlocks)); 2076} 2077 2078DB_SHOW_COMMAND(locks, db_witness_list) 2079{ 2080 struct thread *td; 2081 2082 if (have_addr) 2083 td = db_lookup_thread(addr, TRUE); 2084 else 2085 td = kdb_thread; 2086 witness_list(td); 2087} 2088 2089DB_SHOW_COMMAND(alllocks, db_witness_list_all) 2090{ 2091 struct thread *td; 2092 struct proc *p; 2093 2094 /* 2095 * It would be nice to list only threads and processes that actually 2096 * held sleep locks, but that information is currently not exported 2097 * by WITNESS. 2098 */ 2099 FOREACH_PROC_IN_SYSTEM(p) { 2100 if (!witness_proc_has_locks(p)) 2101 continue; 2102 FOREACH_THREAD_IN_PROC(p, td) { 2103 if (!witness_thread_has_locks(td)) 2104 continue; 2105 db_printf("Process %d (%s) thread %p (%d)\n", p->p_pid, 2106 td->td_name, td, td->td_tid); 2107 witness_list(td); 2108 } 2109 } 2110} 2111 2112DB_SHOW_COMMAND(witness, db_witness_display) 2113{ 2114 2115 witness_display(db_printf); 2116} 2117#endif 2118