subr_witness.c revision 76272
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 * $FreeBSD: head/sys/kern/subr_witness.c 76272 2001-05-04 17:15:16Z jhb $ 31 */ 32 33/* 34 * Implementation of the `witness' lock verifier. Originally implemented for 35 * mutexes in BSD/OS. Extended to handle generic lock objects and lock 36 * classes in FreeBSD. 37 */ 38 39/* 40 * Main Entry: witness 41 * Pronunciation: 'wit-n&s 42 * Function: noun 43 * Etymology: Middle English witnesse, from Old English witnes knowledge, 44 * testimony, witness, from 2wit 45 * Date: before 12th century 46 * 1 : attestation of a fact or event : TESTIMONY 47 * 2 : one that gives evidence; specifically : one who testifies in 48 * a cause or before a judicial tribunal 49 * 3 : one asked to be present at a transaction so as to be able to 50 * testify to its having taken place 51 * 4 : one who has personal knowledge of something 52 * 5 a : something serving as evidence or proof : SIGN 53 * b : public affirmation by word or example of usually 54 * religious faith or conviction <the heroic witness to divine 55 * life -- Pilot> 56 * 6 capitalized : a member of the Jehovah's Witnesses 57 */ 58 59#include "opt_ddb.h" 60#include "opt_witness.h" 61 62#include <sys/param.h> 63#include <sys/bus.h> 64#include <sys/kernel.h> 65#include <sys/ktr.h> 66#include <sys/lock.h> 67#include <sys/malloc.h> 68#include <sys/mutex.h> 69#include <sys/proc.h> 70#include <sys/sysctl.h> 71#include <sys/systm.h> 72 73#include <ddb/ddb.h> 74 75#define WITNESS_COUNT 200 76#define WITNESS_CHILDCOUNT (WITNESS_COUNT * 4) 77/* 78 * XXX: This is somewhat bogus, as we assume here that at most 1024 processes 79 * will hold LOCK_NCHILDREN * 2 locks. We handle failure ok, and we should 80 * probably be safe for the most part, but it's still a SWAG. 81 */ 82#define LOCK_CHILDCOUNT (MAXCPU + 1024) * 2 83 84#define WITNESS_NCHILDREN 6 85 86struct witness_child_list_entry; 87 88struct witness { 89 const char *w_name; 90 struct lock_class *w_class; 91 STAILQ_ENTRY(witness) w_list; /* List of all witnesses. */ 92 STAILQ_ENTRY(witness) w_typelist; /* Witnesses of a type. */ 93 struct witness_child_list_entry *w_children; /* Great evilness... */ 94 const char *w_file; 95 int w_line; 96 u_int w_level; 97 u_int w_refcount; 98 u_char w_Giant_squawked:1; 99 u_char w_other_squawked:1; 100 u_char w_same_squawked:1; 101}; 102 103struct witness_child_list_entry { 104 struct witness_child_list_entry *wcl_next; 105 struct witness *wcl_children[WITNESS_NCHILDREN]; 106 u_int wcl_count; 107}; 108 109STAILQ_HEAD(witness_list, witness); 110 111struct witness_blessed { 112 const char *b_lock1; 113 const char *b_lock2; 114}; 115 116struct witness_order_list_entry { 117 const char *w_name; 118 struct lock_class *w_class; 119}; 120 121static struct witness *enroll(const char *description, 122 struct lock_class *lock_class); 123static int itismychild(struct witness *parent, struct witness *child); 124static void removechild(struct witness *parent, struct witness *child); 125static int isitmychild(struct witness *parent, struct witness *child); 126static int isitmydescendant(struct witness *parent, struct witness *child); 127static int dup_ok(struct witness *); 128static int blessed(struct witness *, struct witness *); 129static void witness_display_list(void(*prnt)(const char *fmt, ...), 130 struct witness_list *list); 131static void witness_displaydescendants(void(*)(const char *fmt, ...), 132 struct witness *); 133static void witness_leveldescendents(struct witness *parent, int level); 134static void witness_levelall(void); 135static struct witness *witness_get(void); 136static void witness_free(struct witness *m); 137static struct witness_child_list_entry *witness_child_get(void); 138static void witness_child_free(struct witness_child_list_entry *wcl); 139static struct lock_list_entry *witness_lock_list_get(void); 140static void witness_lock_list_free(struct lock_list_entry *lle); 141static void witness_display(void(*)(const char *fmt, ...)); 142static struct lock_instance *find_instance(struct lock_list_entry *lock_list, 143 struct lock_object *lock); 144 145MALLOC_DEFINE(M_WITNESS, "witness", "witness structure"); 146 147static int witness_watch; 148TUNABLE_INT_DECL("debug.witness_watch", 1, witness_watch); 149SYSCTL_INT(_debug, OID_AUTO, witness_watch, CTLFLAG_RD, &witness_watch, 0, ""); 150 151#ifdef DDB 152/* 153 * When DDB is enabled and witness_ddb is set to 1, it will cause the system to 154 * drop into kdebug() when: 155 * - a lock heirarchy violation occurs 156 * - locks are held when going to sleep. 157 */ 158int witness_ddb; 159#ifdef WITNESS_DDB 160TUNABLE_INT_DECL("debug.witness_ddb", 1, witness_ddb); 161#else 162TUNABLE_INT_DECL("debug.witness_ddb", 0, witness_ddb); 163#endif 164SYSCTL_INT(_debug, OID_AUTO, witness_ddb, CTLFLAG_RW, &witness_ddb, 0, ""); 165#endif /* DDB */ 166 167int witness_skipspin; 168#ifdef WITNESS_SKIPSPIN 169TUNABLE_INT_DECL("debug.witness_skipspin", 1, witness_skipspin); 170#else 171TUNABLE_INT_DECL("debug.witness_skipspin", 0, witness_skipspin); 172#endif 173SYSCTL_INT(_debug, OID_AUTO, witness_skipspin, CTLFLAG_RD, &witness_skipspin, 0, 174 ""); 175 176static struct mtx w_mtx; 177static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free); 178static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all); 179static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin); 180static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep); 181static struct witness_child_list_entry *w_child_free = NULL; 182static struct lock_list_entry *w_lock_list_free = NULL; 183static int witness_dead; /* fatal error, probably no memory */ 184 185static struct witness w_data[WITNESS_COUNT]; 186static struct witness_child_list_entry w_childdata[WITNESS_CHILDCOUNT]; 187static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT]; 188 189static struct witness_order_list_entry order_lists[] = { 190 { "Giant", &lock_class_mtx_sleep }, 191 { "proctree", &lock_class_sx }, 192 { "allproc", &lock_class_sx }, 193 { "process lock", &lock_class_mtx_sleep }, 194 { "uidinfo hash", &lock_class_mtx_sleep }, 195 { "uidinfo struct", &lock_class_mtx_sleep }, 196 { NULL, NULL }, 197 /* 198 * spin locks 199 */ 200#if defined(__i386__) && defined (SMP) 201 { "com", &lock_class_mtx_spin }, 202#endif 203 { "sio", &lock_class_mtx_spin }, 204#ifdef __i386__ 205 { "cy", &lock_class_mtx_spin }, 206#endif 207 { "ng_node", &lock_class_mtx_spin }, 208 { "ng_worklist", &lock_class_mtx_spin }, 209 { "ithread table lock", &lock_class_mtx_spin }, 210 { "ithread list lock", &lock_class_mtx_spin }, 211 { "sched lock", &lock_class_mtx_spin }, 212 { "clk", &lock_class_mtx_spin }, 213 { "callout", &lock_class_mtx_spin }, 214 /* 215 * leaf locks 216 */ 217#ifdef SMP 218 { "ap boot", &lock_class_mtx_spin }, 219#ifdef __i386__ 220 { "imen", &lock_class_mtx_spin }, 221#endif 222 { "smp rendezvous", &lock_class_mtx_spin }, 223#endif 224 { NULL, NULL }, 225 { NULL, NULL } 226}; 227 228static const char *dup_list[] = { 229 "process lock", 230 NULL 231}; 232 233/* 234 * Pairs of locks which have been blessed 235 * Don't complain about order problems with blessed locks 236 */ 237static struct witness_blessed blessed_list[] = { 238}; 239static int blessed_count = 240 sizeof(blessed_list) / sizeof(struct witness_blessed); 241 242/* 243 * List of all locks in the system. 244 */ 245STAILQ_HEAD(, lock_object) all_locks = STAILQ_HEAD_INITIALIZER(all_locks); 246 247static struct mtx all_mtx = { 248 { &lock_class_mtx_sleep, /* mtx_object.lo_class */ 249 "All locks list", /* mtx_object.lo_name */ 250 LO_INITIALIZED, /* mtx_object.lo_flags */ 251 { NULL }, /* mtx_object.lo_list */ 252 NULL }, /* mtx_object.lo_witness */ 253 MTX_UNOWNED, 0, /* mtx_lock, mtx_recurse */ 254 0, /* mtx_savecrit */ 255 TAILQ_HEAD_INITIALIZER(all_mtx.mtx_blocked), 256 { NULL, NULL } /* mtx_contested */ 257}; 258 259/* 260 * This global is set to 0 once it becomes safe to use the witness code. 261 */ 262static int witness_cold = 1; 263 264/* 265 * Global variables for book keeping. 266 */ 267static int lock_cur_cnt; 268static int lock_max_cnt; 269 270/* 271 * The WITNESS-enabled diagnostic code. 272 */ 273static void 274witness_initialize(void *dummy __unused) 275{ 276 struct lock_object *lock; 277 struct witness_order_list_entry *order; 278 struct witness *w, *w1; 279 int i; 280 281 /* 282 * We have to release Giant before initializing its witness 283 * structure so that WITNESS doesn't get confused. 284 */ 285 mtx_unlock(&Giant); 286 mtx_assert(&Giant, MA_NOTOWNED); 287 288 STAILQ_INSERT_HEAD(&all_locks, &all_mtx.mtx_object, lo_list); 289 mtx_init(&w_mtx, "witness lock", MTX_SPIN | MTX_QUIET | MTX_NOWITNESS); 290 for (i = 0; i < WITNESS_COUNT; i++) 291 witness_free(&w_data[i]); 292 for (i = 0; i < WITNESS_CHILDCOUNT; i++) 293 witness_child_free(&w_childdata[i]); 294 for (i = 0; i < LOCK_CHILDCOUNT; i++) 295 witness_lock_list_free(&w_locklistdata[i]); 296 297 /* First add in all the specified order lists. */ 298 for (order = order_lists; order->w_name != NULL; order++) { 299 w = enroll(order->w_name, order->w_class); 300 if (w == NULL) 301 continue; 302 w->w_file = "order list"; 303 for (order++; order->w_name != NULL; order++) { 304 w1 = enroll(order->w_name, order->w_class); 305 if (w1 == NULL) 306 continue; 307 w1->w_file = "order list"; 308 itismychild(w, w1); 309 w = w1; 310 } 311 } 312 313 /* Iterate through all locks and add them to witness. */ 314 mtx_lock(&all_mtx); 315 STAILQ_FOREACH(lock, &all_locks, lo_list) { 316 if (lock->lo_flags & LO_WITNESS) 317 lock->lo_witness = enroll(lock->lo_name, 318 lock->lo_class); 319 else 320 lock->lo_witness = NULL; 321 } 322 mtx_unlock(&all_mtx); 323 324 /* Mark the witness code as being ready for use. */ 325 atomic_store_rel_int(&witness_cold, 0); 326 327 mtx_lock(&Giant); 328} 329SYSINIT(witness_init, SI_SUB_WITNESS, SI_ORDER_FIRST, witness_initialize, NULL) 330 331void 332witness_init(struct lock_object *lock) 333{ 334 struct lock_class *class; 335 336 class = lock->lo_class; 337 if (lock->lo_flags & LO_INITIALIZED) 338 panic("%s: lock (%s) %s is already initialized!\n", __func__, 339 class->lc_name, lock->lo_name); 340 341 if ((lock->lo_flags & LO_RECURSABLE) != 0 && 342 (class->lc_flags & LC_RECURSABLE) == 0) 343 panic("%s: lock (%s) %s can not be recursable!\n", __func__, 344 class->lc_name, lock->lo_name); 345 346 if ((lock->lo_flags & LO_SLEEPABLE) != 0 && 347 (class->lc_flags & LC_SLEEPABLE) == 0) 348 panic("%s: lock (%s) %s can not be sleepable!\n", __func__, 349 class->lc_name, lock->lo_name); 350 351 mtx_lock(&all_mtx); 352 STAILQ_INSERT_TAIL(&all_locks, lock, lo_list); 353 lock->lo_flags |= LO_INITIALIZED; 354 lock_cur_cnt++; 355 if (lock_cur_cnt > lock_max_cnt) 356 lock_max_cnt = lock_cur_cnt; 357 mtx_unlock(&all_mtx); 358 if (!witness_cold && !witness_dead && 359 (lock->lo_flags & LO_WITNESS) != 0) 360 lock->lo_witness = enroll(lock->lo_name, class); 361 else 362 lock->lo_witness = NULL; 363} 364 365void 366witness_destroy(struct lock_object *lock) 367{ 368 struct witness *w; 369 370 if (witness_cold) 371 panic("lock (%s) %s destroyed while witness_cold", 372 lock->lo_class->lc_name, lock->lo_name); 373 374 if ((lock->lo_flags & LO_INITIALIZED) == 0) 375 panic("%s: lock (%s) %s is not initialized!\n", __func__, 376 lock->lo_class->lc_name, lock->lo_name); 377 378 /* XXX: need to verify that no one holds the lock */ 379 w = lock->lo_witness; 380 if (w != NULL) { 381 mtx_lock_spin(&w_mtx); 382 w->w_refcount--; 383 if (w->w_refcount == 0) { 384 w->w_name = "(dead)"; 385 w->w_file = "(dead)"; 386 w->w_line = 0; 387 } 388 mtx_unlock_spin(&w_mtx); 389 } 390 391 mtx_lock(&all_mtx); 392 lock_cur_cnt--; 393 STAILQ_REMOVE(&all_locks, lock, lock_object, lo_list); 394 lock->lo_flags &= LO_INITIALIZED; 395 mtx_unlock(&all_mtx); 396} 397 398static void 399witness_display_list(void(*prnt)(const char *fmt, ...), 400 struct witness_list *list) 401{ 402 struct witness *w, *w1; 403 int found; 404 405 STAILQ_FOREACH(w, list, w_typelist) { 406 if (w->w_file == NULL) 407 continue; 408 found = 0; 409 STAILQ_FOREACH(w1, list, w_typelist) { 410 if (isitmychild(w1, w)) { 411 found++; 412 break; 413 } 414 } 415 if (found) 416 continue; 417 /* 418 * This lock has no anscestors, display its descendants. 419 */ 420 witness_displaydescendants(prnt, w); 421 } 422} 423 424static void 425witness_display(void(*prnt)(const char *fmt, ...)) 426{ 427 struct witness *w; 428 429 KASSERT(!witness_cold, ("%s: witness_cold\n", __func__)); 430 witness_levelall(); 431 432 /* 433 * First, handle sleep locks which have been acquired at least 434 * once. 435 */ 436 prnt("Sleep locks:\n"); 437 witness_display_list(prnt, &w_sleep); 438 439 /* 440 * Now do spin locks which have been acquired at least once. 441 */ 442 prnt("\nSpin locks:\n"); 443 witness_display_list(prnt, &w_spin); 444 445 /* 446 * Finally, any locks which have not been acquired yet. 447 */ 448 prnt("\nLocks which were never acquired:\n"); 449 STAILQ_FOREACH(w, &w_all, w_list) { 450 if (w->w_file != NULL) 451 continue; 452 prnt("%s\n", w->w_name); 453 } 454} 455 456void 457witness_lock(struct lock_object *lock, int flags, const char *file, int line) 458{ 459 struct lock_list_entry **lock_list, *lle; 460 struct lock_instance *lock1, *lock2; 461 struct lock_class *class; 462 struct witness *w, *w1; 463 struct proc *p; 464 int i, j; 465#ifdef DDB 466 int go_into_ddb = 0; 467#endif /* DDB */ 468 469 if (witness_cold || witness_dead || lock->lo_witness == NULL || 470 panicstr) 471 return; 472 w = lock->lo_witness; 473 class = lock->lo_class; 474 p = curproc; 475 476 /* 477 * We have to hold a spinlock to keep lock_list valid across the check 478 * in the LC_SLEEPLOCK case. In the LC_SPINLOCK case, it is already 479 * protected by the spinlock we are currently performing the witness 480 * checks on, so it is ok to release the lock after performing this 481 * check. All we have to protect is the LC_SLEEPLOCK case when no 482 * spinlocks are held as we may get preempted during this check and 483 * lock_list could end up pointing to some other CPU's spinlock list. 484 */ 485 mtx_lock_spin(&w_mtx); 486 lock_list = PCPU_PTR(spinlocks); 487 if (class->lc_flags & LC_SLEEPLOCK) { 488 if (*lock_list != NULL) { 489 mtx_unlock_spin(&w_mtx); 490 panic("blockable sleep lock (%s) %s @ %s:%d", 491 class->lc_name, lock->lo_name, file, line); 492 } 493 lock_list = &p->p_sleeplocks; 494 } 495 mtx_unlock_spin(&w_mtx); 496 497 if (flags & LOP_TRYLOCK) 498 goto out; 499 500 /* 501 * Is this the first lock acquired? If so, then no order checking 502 * is needed. 503 */ 504 if (*lock_list == NULL) 505 goto out; 506 507 /* 508 * Check to see if we are recursing on a lock we already own. 509 */ 510 lock1 = find_instance(*lock_list, lock); 511 if (lock1 != NULL) { 512 if ((lock1->li_flags & LI_EXCLUSIVE) != 0 && 513 (flags & LOP_EXCLUSIVE) == 0) { 514 printf("shared lock of (%s) %s @ %s:%d\n", 515 class->lc_name, lock->lo_name, file, line); 516 printf("while exclusively locked from %s:%d\n", 517 lock1->li_file, lock1->li_line); 518 panic("share->excl"); 519 } 520 if ((lock1->li_flags & LI_EXCLUSIVE) == 0 && 521 (flags & LOP_EXCLUSIVE) != 0) { 522 printf("exclusive lock of (%s) %s @ %s:%d\n", 523 class->lc_name, lock->lo_name, file, line); 524 printf("while share locked from %s:%d\n", 525 lock1->li_file, lock1->li_line); 526 panic("excl->share"); 527 } 528 lock1->li_flags++; 529 if ((lock->lo_flags & LO_RECURSABLE) == 0) { 530 printf( 531 "recursed on non-recursive lock (%s) %s @ %s:%d\n", 532 class->lc_name, lock->lo_name, file, line); 533 printf("first acquired @ %s:%d\n", lock1->li_file, 534 lock1->li_line); 535 panic("recurse"); 536 } 537 lock1->li_file = file; 538 lock1->li_line = line; 539 return; 540 } 541 542 /* 543 * Check for duplicate locks of the same type. Note that we only 544 * have to check for this on the last lock we just acquired. Any 545 * other cases will be caught as lock order violations. 546 */ 547 lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1]; 548 w1 = lock1->li_lock->lo_witness; 549 if (w1 == w) { 550 if (w->w_same_squawked || dup_ok(w)) 551 goto out; 552 w->w_same_squawked = 1; 553 printf("acquiring duplicate lock of same type: \"%s\"\n", 554 lock->lo_name); 555 printf(" 1st @ %s:%d\n", lock1->li_file, lock1->li_line); 556 printf(" 2nd @ %s:%d\n", file, line); 557#ifdef DDB 558 go_into_ddb = 1; 559#endif /* DDB */ 560 goto out; 561 } 562 MPASS(!mtx_owned(&w_mtx)); 563 mtx_lock_spin(&w_mtx); 564 /* 565 * If we have a known higher number just say ok 566 */ 567 if (witness_watch > 1 && w->w_level > w1->w_level) { 568 mtx_unlock_spin(&w_mtx); 569 goto out; 570 } 571 if (isitmydescendant(w1, w)) { 572 mtx_unlock_spin(&w_mtx); 573 goto out; 574 } 575 for (j = 0, lle = *lock_list; lle != NULL; lle = lle->ll_next) { 576 for (i = lle->ll_count - 1; i >= 0; i--, j++) { 577 578 MPASS(j < WITNESS_COUNT); 579 lock1 = &lle->ll_children[i]; 580 w1 = lock1->li_lock->lo_witness; 581 582 /* 583 * If this lock doesn't undergo witness checking, 584 * then skip it. 585 */ 586 if (w1 == NULL) { 587 KASSERT((lock1->li_lock->lo_flags & LO_WITNESS) == 0, 588 ("lock missing witness structure")); 589 continue; 590 } 591 /* 592 * If we are locking Giant and we slept with this 593 * lock, then skip it. 594 */ 595 if ((lock1->li_flags & LI_SLEPT) != 0 && 596 lock == &Giant.mtx_object) 597 continue; 598 if (!isitmydescendant(w, w1)) 599 continue; 600 /* 601 * We have a lock order violation, check to see if it 602 * is allowed or has already been yelled about. 603 */ 604 mtx_unlock_spin(&w_mtx); 605 if (blessed(w, w1)) 606 goto out; 607 if (lock1->li_lock == &Giant.mtx_object) { 608 if (w1->w_Giant_squawked) 609 goto out; 610 else 611 w1->w_Giant_squawked = 1; 612 } else { 613 if (w1->w_other_squawked) 614 goto out; 615 else 616 w1->w_other_squawked = 1; 617 } 618 /* 619 * Ok, yell about it. 620 */ 621 printf("lock order reversal\n"); 622 /* 623 * Try to locate an earlier lock with 624 * witness w in our list. 625 */ 626 do { 627 lock2 = &lle->ll_children[i]; 628 MPASS(lock2->li_lock != NULL); 629 if (lock2->li_lock->lo_witness == w) 630 break; 631 i--; 632 if (i == 0 && lle->ll_next != NULL) { 633 lle = lle->ll_next; 634 i = lle->ll_count - 1; 635 MPASS(i != 0); 636 } 637 } while (i >= 0); 638 if (i < 0) { 639 printf(" 1st %p %s @ %s:%d\n", lock1->li_lock, 640 lock1->li_lock->lo_name, lock1->li_file, 641 lock1->li_line); 642 printf(" 2nd %p %s @ %s:%d\n", lock, 643 lock->lo_name, file, line); 644 } else { 645 printf(" 1st %p %s @ %s:%d\n", lock2->li_lock, 646 lock2->li_lock->lo_name, lock2->li_file, 647 lock2->li_line); 648 printf(" 2nd %p %s @ %s:%d\n", lock1->li_lock, 649 lock1->li_lock->lo_name, lock1->li_file, 650 lock1->li_line); 651 printf(" 3rd %p %s @ %s:%d\n", lock, 652 lock->lo_name, file, line); 653 } 654#ifdef DDB 655 go_into_ddb = 1; 656#endif /* DDB */ 657 goto out; 658 } 659 } 660 lock1 = &(*lock_list)->ll_children[(*lock_list)->ll_count - 1]; 661 if (!itismychild(lock1->li_lock->lo_witness, w)) 662 mtx_unlock_spin(&w_mtx); 663 664out: 665#ifdef DDB 666 if (witness_ddb && go_into_ddb) 667 Debugger(__func__); 668#endif /* DDB */ 669 w->w_file = file; 670 w->w_line = line; 671 672 lle = *lock_list; 673 if (lle == NULL || lle->ll_count == LOCK_NCHILDREN) { 674 *lock_list = witness_lock_list_get(); 675 if (*lock_list == NULL) 676 return; 677 (*lock_list)->ll_next = lle; 678 lle = *lock_list; 679 } 680 lock1 = &lle->ll_children[lle->ll_count++]; 681 lock1->li_lock = lock; 682 lock1->li_line = line; 683 lock1->li_file = file; 684 if ((flags & LOP_EXCLUSIVE) != 0) 685 lock1->li_flags = LI_EXCLUSIVE; 686 else 687 lock1->li_flags = 0; 688} 689 690void 691witness_unlock(struct lock_object *lock, int flags, const char *file, int line) 692{ 693 struct lock_list_entry **lock_list, *lle; 694 struct lock_instance *instance; 695 struct lock_class *class; 696 struct proc *p; 697 int i, j; 698 699 if (witness_cold || witness_dead || lock->lo_witness == NULL || 700 panicstr) 701 return; 702 p = curproc; 703 class = lock->lo_class; 704 if (class->lc_flags & LC_SLEEPLOCK) 705 lock_list = &p->p_sleeplocks; 706 else 707 lock_list = PCPU_PTR(spinlocks); 708 for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next) 709 for (i = 0; i < (*lock_list)->ll_count; i++) { 710 instance = &(*lock_list)->ll_children[i]; 711 if (instance->li_lock == lock) { 712 if ((instance->li_flags & LI_EXCLUSIVE) != 0 && 713 (flags & LOP_EXCLUSIVE) == 0) { 714 printf( 715 "shared unlock of (%s) %s @ %s:%d\n", 716 class->lc_name, lock->lo_name, 717 file, line); 718 printf( 719 "while exclusively locked from %s:%d\n", 720 instance->li_file, 721 instance->li_line); 722 panic("excl->ushare"); 723 } 724 if ((instance->li_flags & LI_EXCLUSIVE) == 0 && 725 (flags & LOP_EXCLUSIVE) != 0) { 726 printf( 727 "exclusive unlock of (%s) %s @ %s:%d\n", 728 class->lc_name, lock->lo_name, 729 file, line); 730 printf( 731 "while share locked from %s:%d\n", 732 instance->li_file, 733 instance->li_line); 734 panic("share->uexcl"); 735 } 736 /* If we are recursed, unrecurse. */ 737 if ((instance->li_flags & LI_RECURSEMASK) > 0) { 738 instance->li_flags--; 739 goto out; 740 } 741 (*lock_list)->ll_count--; 742 for (j = i; j < (*lock_list)->ll_count; j++) 743 (*lock_list)->ll_children[j] = 744 (*lock_list)->ll_children[j + 1]; 745 if ((*lock_list)->ll_count == 0) { 746 lle = *lock_list; 747 *lock_list = lle->ll_next; 748 witness_lock_list_free(lle); 749 } 750 goto out; 751 } 752 } 753 panic("lock (%s) %s not locked @ %s:%d", class->lc_name, lock->lo_name, 754 file, line); 755out: 756 /* 757 * We don't need to protect this PCPU_GET() here against preemption 758 * because if we hold any spinlocks then we are already protected, 759 * and if we don't we will get NULL if we hold no spinlocks even if 760 * we switch CPU's while reading it. 761 */ 762 if (class->lc_flags & LC_SLEEPLOCK) { 763 if ((flags & LOP_NOSWITCH) == 0 && PCPU_GET(spinlocks) != NULL) 764 panic("switchable sleep unlock (%s) %s @ %s:%d", 765 class->lc_name, lock->lo_name, file, line); 766 } 767} 768 769/* 770 * Warn if any held locks are not sleepable. Note that Giant and the lock 771 * passed in are both special cases since they are both released during the 772 * sleep process and aren't actually held while the process is asleep. 773 */ 774int 775witness_sleep(int check_only, struct lock_object *lock, const char *file, 776 int line) 777{ 778 struct lock_list_entry **lock_list, *lle; 779 struct lock_instance *lock1; 780 struct proc *p; 781 critical_t savecrit; 782 int i, n; 783 784 if (witness_dead || panicstr) 785 return (0); 786 KASSERT(!witness_cold, ("%s: witness_cold\n", __func__)); 787 n = 0; 788 /* 789 * Preemption bad because we need PCPU_PTR(spinlocks) to not change. 790 */ 791 savecrit = critical_enter(); 792 p = curproc; 793 lock_list = &p->p_sleeplocks; 794again: 795 for (lle = *lock_list; lle != NULL; lle = lle->ll_next) 796 for (i = lle->ll_count - 1; i >= 0; i--) { 797 lock1 = &lle->ll_children[i]; 798 if (lock1->li_lock == lock || 799 lock1->li_lock == &Giant.mtx_object) 800 continue; 801 if ((lock1->li_lock->lo_flags & LO_SLEEPABLE) != 0) { 802 if (check_only == 0) 803 lock1->li_flags |= LI_SLEPT; 804 continue; 805 } 806 n++; 807 printf("%s:%d: %s with \"%s\" locked from %s:%d\n", 808 file, line, check_only ? "could sleep" : "sleeping", 809 lock1->li_lock->lo_name, lock1->li_file, 810 lock1->li_line); 811 } 812 if (lock_list == &p->p_sleeplocks) { 813 lock_list = PCPU_PTR(spinlocks); 814 goto again; 815 } 816#ifdef DDB 817 if (witness_ddb && n) 818 Debugger(__func__); 819#endif /* DDB */ 820 critical_exit(savecrit); 821 return (n); 822} 823 824static struct witness * 825enroll(const char *description, struct lock_class *lock_class) 826{ 827 struct witness *w; 828 829 if (!witness_watch) 830 return (NULL); 831 832 if ((lock_class->lc_flags & LC_SPINLOCK) && witness_skipspin) 833 return (NULL); 834 mtx_lock_spin(&w_mtx); 835 STAILQ_FOREACH(w, &w_all, w_list) { 836 if (strcmp(description, w->w_name) == 0) { 837 w->w_refcount++; 838 mtx_unlock_spin(&w_mtx); 839 if (lock_class != w->w_class) 840 panic( 841 "lock (%s) %s does not match earlier (%s) lock", 842 description, lock_class->lc_name, 843 w->w_class->lc_name); 844 return (w); 845 } 846 } 847 /* 848 * This isn't quite right, as witness_cold is still 0 while we 849 * enroll all the locks initialized before witness_initialize(). 850 */ 851 if ((lock_class->lc_flags & LC_SPINLOCK) && !witness_cold) { 852 mtx_unlock_spin(&w_mtx); 853 panic("spin lock %s not in order list", description); 854 } 855 if ((w = witness_get()) == NULL) 856 return (NULL); 857 w->w_name = description; 858 w->w_class = lock_class; 859 w->w_refcount = 1; 860 STAILQ_INSERT_HEAD(&w_all, w, w_list); 861 if (lock_class->lc_flags & LC_SPINLOCK) 862 STAILQ_INSERT_HEAD(&w_spin, w, w_typelist); 863 else if (lock_class->lc_flags & LC_SLEEPLOCK) 864 STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist); 865 else { 866 mtx_unlock_spin(&w_mtx); 867 panic("lock class %s is not sleep or spin", 868 lock_class->lc_name); 869 } 870 mtx_unlock_spin(&w_mtx); 871 872 return (w); 873} 874 875static int 876itismychild(struct witness *parent, struct witness *child) 877{ 878 static int recursed; 879 struct witness_child_list_entry **wcl; 880 struct witness_list *list; 881 882 MPASS(child != NULL && parent != NULL); 883 if ((parent->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) != 884 (child->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK))) 885 panic( 886 "%s: parent (%s) and child (%s) are not the same lock type", 887 __func__, parent->w_class->lc_name, 888 child->w_class->lc_name); 889 890 /* 891 * Insert "child" after "parent" 892 */ 893 wcl = &parent->w_children; 894 while (*wcl != NULL && (*wcl)->wcl_count == WITNESS_NCHILDREN) 895 wcl = &(*wcl)->wcl_next; 896 897 if (*wcl == NULL) { 898 *wcl = witness_child_get(); 899 if (*wcl == NULL) 900 return (1); 901 } 902 903 (*wcl)->wcl_children[(*wcl)->wcl_count++] = child; 904 905 /* 906 * Now prune whole tree. We look for cases where a lock is now 907 * both a descendant and a direct child of a given lock. In that 908 * case, we want to remove the direct child link from the tree. 909 */ 910 if (recursed) 911 return (0); 912 recursed = 1; 913 if (parent->w_class->lc_flags & LC_SLEEPLOCK) 914 list = &w_sleep; 915 else 916 list = &w_spin; 917 STAILQ_FOREACH(child, list, w_typelist) { 918 STAILQ_FOREACH(parent, list, w_typelist) { 919 if (!isitmychild(parent, child)) 920 continue; 921 removechild(parent, child); 922 if (isitmydescendant(parent, child)) 923 continue; 924 itismychild(parent, child); 925 } 926 } 927 recursed = 0; 928 witness_levelall(); 929 return (0); 930} 931 932static void 933removechild(struct witness *parent, struct witness *child) 934{ 935 struct witness_child_list_entry **wcl, *wcl1; 936 int i; 937 938 for (wcl = &parent->w_children; *wcl != NULL; wcl = &(*wcl)->wcl_next) 939 for (i = 0; i < (*wcl)->wcl_count; i++) 940 if ((*wcl)->wcl_children[i] == child) 941 goto found; 942 return; 943found: 944 (*wcl)->wcl_count--; 945 if ((*wcl)->wcl_count > i) 946 (*wcl)->wcl_children[i] = 947 (*wcl)->wcl_children[(*wcl)->wcl_count]; 948 MPASS((*wcl)->wcl_children[i] != NULL); 949 950 if ((*wcl)->wcl_count != 0) 951 return; 952 953 wcl1 = *wcl; 954 *wcl = wcl1->wcl_next; 955 witness_child_free(wcl1); 956} 957 958static int 959isitmychild(struct witness *parent, struct witness *child) 960{ 961 struct witness_child_list_entry *wcl; 962 int i; 963 964 for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) { 965 for (i = 0; i < wcl->wcl_count; i++) { 966 if (wcl->wcl_children[i] == child) 967 return (1); 968 } 969 } 970 return (0); 971} 972 973static int 974isitmydescendant(struct witness *parent, struct witness *child) 975{ 976 struct witness_child_list_entry *wcl; 977 int i, j; 978 979 if (isitmychild(parent, child)) 980 return (1); 981 j = 0; 982 for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) { 983 MPASS(j < 1000); 984 for (i = 0; i < wcl->wcl_count; i++) { 985 if (isitmydescendant(wcl->wcl_children[i], child)) 986 return (1); 987 } 988 j++; 989 } 990 return (0); 991} 992 993void 994witness_levelall (void) 995{ 996 struct witness_list *list; 997 struct witness *w, *w1; 998 999 /* 1000 * First clear all levels. 1001 */ 1002 STAILQ_FOREACH(w, &w_all, w_list) { 1003 w->w_level = 0; 1004 } 1005 1006 /* 1007 * Look for locks with no parent and level all their descendants. 1008 */ 1009 STAILQ_FOREACH(w, &w_all, w_list) { 1010 /* 1011 * This is just an optimization, technically we could get 1012 * away just walking the all list each time. 1013 */ 1014 if (w->w_class->lc_flags & LC_SLEEPLOCK) 1015 list = &w_sleep; 1016 else 1017 list = &w_spin; 1018 STAILQ_FOREACH(w1, list, w_typelist) { 1019 if (isitmychild(w1, w)) 1020 goto skip; 1021 } 1022 witness_leveldescendents(w, 0); 1023 skip: 1024 } 1025} 1026 1027static void 1028witness_leveldescendents(struct witness *parent, int level) 1029{ 1030 struct witness_child_list_entry *wcl; 1031 int i; 1032 1033 if (parent->w_level < level) 1034 parent->w_level = level; 1035 level++; 1036 for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) 1037 for (i = 0; i < wcl->wcl_count; i++) 1038 witness_leveldescendents(wcl->wcl_children[i], level); 1039} 1040 1041static void 1042witness_displaydescendants(void(*prnt)(const char *fmt, ...), 1043 struct witness *parent) 1044{ 1045 struct witness_child_list_entry *wcl; 1046 int i, level; 1047 1048 level = parent->w_level; 1049 1050 prnt("%-2d", level); 1051 for (i = 0; i < level; i++) 1052 prnt(" "); 1053 prnt("%s", parent->w_name); 1054 if (parent->w_file != NULL) 1055 prnt(" -- last acquired @ %s:%d\n", parent->w_file, 1056 parent->w_line); 1057 1058 for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) 1059 for (i = 0; i < wcl->wcl_count; i++) 1060 witness_displaydescendants(prnt, 1061 wcl->wcl_children[i]); 1062} 1063 1064static int 1065dup_ok(struct witness *w) 1066{ 1067 const char **dup; 1068 1069 for (dup = dup_list; *dup != NULL; dup++) 1070 if (strcmp(w->w_name, *dup) == 0) 1071 return (1); 1072 return (0); 1073} 1074 1075static int 1076blessed(struct witness *w1, struct witness *w2) 1077{ 1078 int i; 1079 struct witness_blessed *b; 1080 1081 for (i = 0; i < blessed_count; i++) { 1082 b = &blessed_list[i]; 1083 if (strcmp(w1->w_name, b->b_lock1) == 0) { 1084 if (strcmp(w2->w_name, b->b_lock2) == 0) 1085 return (1); 1086 continue; 1087 } 1088 if (strcmp(w1->w_name, b->b_lock2) == 0) 1089 if (strcmp(w2->w_name, b->b_lock1) == 0) 1090 return (1); 1091 } 1092 return (0); 1093} 1094 1095static struct witness * 1096witness_get(void) 1097{ 1098 struct witness *w; 1099 1100 if (STAILQ_EMPTY(&w_free)) { 1101 witness_dead = 1; 1102 mtx_unlock_spin(&w_mtx); 1103 printf("%s: witness exhausted\n", __func__); 1104 return (NULL); 1105 } 1106 w = STAILQ_FIRST(&w_free); 1107 STAILQ_REMOVE_HEAD(&w_free, w_list); 1108 bzero(w, sizeof(*w)); 1109 return (w); 1110} 1111 1112static void 1113witness_free(struct witness *w) 1114{ 1115 1116 STAILQ_INSERT_HEAD(&w_free, w, w_list); 1117} 1118 1119static struct witness_child_list_entry * 1120witness_child_get(void) 1121{ 1122 struct witness_child_list_entry *wcl; 1123 1124 wcl = w_child_free; 1125 if (wcl == NULL) { 1126 witness_dead = 1; 1127 mtx_unlock_spin(&w_mtx); 1128 printf("%s: witness exhausted\n", __func__); 1129 return (NULL); 1130 } 1131 w_child_free = wcl->wcl_next; 1132 bzero(wcl, sizeof(*wcl)); 1133 return (wcl); 1134} 1135 1136static void 1137witness_child_free(struct witness_child_list_entry *wcl) 1138{ 1139 1140 wcl->wcl_next = w_child_free; 1141 w_child_free = wcl; 1142} 1143 1144static struct lock_list_entry * 1145witness_lock_list_get(void) 1146{ 1147 struct lock_list_entry *lle; 1148 1149 mtx_lock_spin(&w_mtx); 1150 lle = w_lock_list_free; 1151 if (lle == NULL) { 1152 witness_dead = 1; 1153 mtx_unlock_spin(&w_mtx); 1154 printf("%s: witness exhausted\n", __func__); 1155 return (NULL); 1156 } 1157 w_lock_list_free = lle->ll_next; 1158 mtx_unlock_spin(&w_mtx); 1159 bzero(lle, sizeof(*lle)); 1160 return (lle); 1161} 1162 1163static void 1164witness_lock_list_free(struct lock_list_entry *lle) 1165{ 1166 1167 mtx_lock_spin(&w_mtx); 1168 lle->ll_next = w_lock_list_free; 1169 w_lock_list_free = lle; 1170 mtx_unlock_spin(&w_mtx); 1171} 1172 1173static struct lock_instance * 1174find_instance(struct lock_list_entry *lock_list, struct lock_object *lock) 1175{ 1176 struct lock_list_entry *lle; 1177 struct lock_instance *instance; 1178 int i; 1179 1180 for (lle = lock_list; lle != NULL; lle = lle->ll_next) 1181 for (i = lle->ll_count - 1; i >= 0; i--) { 1182 instance = &lle->ll_children[i]; 1183 if (instance->li_lock == lock) 1184 return (instance); 1185 } 1186 return (NULL); 1187} 1188 1189int 1190witness_list_locks(struct lock_list_entry **lock_list) 1191{ 1192 struct lock_list_entry *lle; 1193 struct lock_instance *instance; 1194 struct lock_object *lock; 1195 int i, nheld; 1196 1197 nheld = 0; 1198 for (lle = *lock_list; lle != NULL; lle = lle->ll_next) 1199 for (i = lle->ll_count - 1; i >= 0; i--) { 1200 instance = &lle->ll_children[i]; 1201 lock = instance->li_lock; 1202 printf("%s (%s) %s (%p) locked @ %s:%d\n", 1203 (instance->li_flags & LI_EXCLUSIVE) != 0 ? 1204 "exclusive" : "shared", 1205 lock->lo_class->lc_name, lock->lo_name, lock, 1206 instance->li_file, instance->li_line); 1207 nheld++; 1208 } 1209 return (nheld); 1210} 1211 1212/* 1213 * Calling this on p != curproc is bad unless we are in ddb. 1214 */ 1215int 1216witness_list(struct proc *p) 1217{ 1218 critical_t savecrit; 1219 int nheld; 1220 1221 KASSERT(p == curproc || db_active, 1222 ("%s: p != curproc and we aren't in the debugger", __func__)); 1223 KASSERT(!witness_cold, ("%s: witness_cold", __func__)); 1224 1225 nheld = witness_list_locks(&p->p_sleeplocks); 1226 1227 /* 1228 * We only handle spinlocks if p == curproc. This is somewhat broken 1229 * if p is currently executing on some other CPU and holds spin locks 1230 * as we won't display those locks. If we had a MI way of getting 1231 * the per-cpu data for a given cpu then we could use p->p_oncpu to 1232 * get the list of spinlocks for this process and "fix" this. 1233 */ 1234 if (p == curproc) { 1235 /* 1236 * Preemption bad because we need PCPU_PTR(spinlocks) to not 1237 * change. 1238 */ 1239 savecrit = critical_enter(); 1240 nheld += witness_list_locks(PCPU_PTR(spinlocks)); 1241 critical_exit(savecrit); 1242 } 1243 1244 return (nheld); 1245} 1246 1247void 1248witness_save(struct lock_object *lock, const char **filep, int *linep) 1249{ 1250 struct lock_instance *instance; 1251 1252 KASSERT(!witness_cold, ("%s: witness_cold\n", __func__)); 1253 if (lock->lo_witness == NULL) 1254 return; 1255 1256 KASSERT(lock->lo_class->lc_flags & LC_SLEEPLOCK, 1257 ("%s: lock (%s) %s is not a sleep lock", __func__, 1258 lock->lo_class->lc_name, lock->lo_name)); 1259 instance = find_instance(curproc->p_sleeplocks, lock); 1260 KASSERT(instance != NULL, ("%s: lock (%s) %s not locked", __func__, 1261 lock->lo_class->lc_name, lock->lo_name)); 1262 1263 *filep = instance->li_file; 1264 *linep = instance->li_line; 1265} 1266 1267void 1268witness_restore(struct lock_object *lock, const char *file, int line) 1269{ 1270 struct lock_instance *instance; 1271 1272 KASSERT(!witness_cold, ("%s: witness_cold\n", __func__)); 1273 if (lock->lo_witness == NULL) 1274 return; 1275 1276 KASSERT(lock->lo_class->lc_flags & LC_SLEEPLOCK, 1277 ("%s: lock (%s) %s is not a sleep lock", __func__, 1278 lock->lo_class->lc_name, lock->lo_name)); 1279 instance = find_instance(curproc->p_sleeplocks, lock); 1280 KASSERT(instance != NULL, ("%s: lock (%s) %s not locked", __func__, 1281 lock->lo_class->lc_name, lock->lo_name)); 1282 1283 lock->lo_witness->w_file = file; 1284 lock->lo_witness->w_line = line; 1285 instance->li_file = file; 1286 instance->li_line = line; 1287} 1288 1289#ifdef DDB 1290 1291DB_SHOW_COMMAND(locks, db_witness_list) 1292{ 1293 struct proc *p; 1294 pid_t pid; 1295 1296 if (have_addr) { 1297 pid = (addr % 16) + ((addr >> 4) % 16) * 10 + 1298 ((addr >> 8) % 16) * 100 + ((addr >> 12) % 16) * 1000 + 1299 ((addr >> 16) % 16) * 10000; 1300 1301 /* sx_slock(&allproc_lock); */ 1302 LIST_FOREACH(p, &allproc, p_list) { 1303 if (p->p_pid == pid) 1304 break; 1305 } 1306 /* sx_sunlock(&allproc_lock); */ 1307 if (p == NULL) { 1308 db_printf("pid %d not found\n", pid); 1309 return; 1310 } 1311 } else 1312 p = curproc; 1313 1314 witness_list(p); 1315} 1316 1317DB_SHOW_COMMAND(witness, db_witness_display) 1318{ 1319 1320 witness_display(db_printf); 1321} 1322#endif 1323