subr_witness.c revision 75569
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 75569 2001-04-17 03:35:38Z 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, ...)); 142 143MALLOC_DEFINE(M_WITNESS, "witness", "witness structure"); 144 145static int witness_watch; 146TUNABLE_INT_DECL("debug.witness_watch", 1, witness_watch); 147SYSCTL_INT(_debug, OID_AUTO, witness_watch, CTLFLAG_RD, &witness_watch, 0, ""); 148 149#ifdef DDB 150/* 151 * When DDB is enabled and witness_ddb is set to 1, it will cause the system to 152 * drop into kdebug() when: 153 * - a lock heirarchy violation occurs 154 * - locks are held when going to sleep. 155 */ 156int witness_ddb; 157#ifdef WITNESS_DDB 158TUNABLE_INT_DECL("debug.witness_ddb", 1, witness_ddb); 159#else 160TUNABLE_INT_DECL("debug.witness_ddb", 0, witness_ddb); 161#endif 162SYSCTL_INT(_debug, OID_AUTO, witness_ddb, CTLFLAG_RW, &witness_ddb, 0, ""); 163#endif /* DDB */ 164 165int witness_skipspin; 166#ifdef WITNESS_SKIPSPIN 167TUNABLE_INT_DECL("debug.witness_skipspin", 1, witness_skipspin); 168#else 169TUNABLE_INT_DECL("debug.witness_skipspin", 0, witness_skipspin); 170#endif 171SYSCTL_INT(_debug, OID_AUTO, witness_skipspin, CTLFLAG_RD, &witness_skipspin, 0, 172 ""); 173 174static struct mtx w_mtx; 175static struct witness_list w_free = STAILQ_HEAD_INITIALIZER(w_free); 176static struct witness_list w_all = STAILQ_HEAD_INITIALIZER(w_all); 177static struct witness_list w_spin = STAILQ_HEAD_INITIALIZER(w_spin); 178static struct witness_list w_sleep = STAILQ_HEAD_INITIALIZER(w_sleep); 179static struct witness_child_list_entry *w_child_free = NULL; 180static struct lock_list_entry *w_lock_list_free = NULL; 181static int witness_dead; /* fatal error, probably no memory */ 182 183static struct witness w_data[WITNESS_COUNT]; 184static struct witness_child_list_entry w_childdata[WITNESS_CHILDCOUNT]; 185static struct lock_list_entry w_locklistdata[LOCK_CHILDCOUNT]; 186 187static struct witness_order_list_entry order_lists[] = { 188 { "Giant", &lock_class_mtx_sleep }, 189 { "proctree", &lock_class_sx }, 190 { "allproc", &lock_class_sx }, 191 { "process lock", &lock_class_mtx_sleep }, 192 { "uidinfo hash", &lock_class_mtx_sleep }, 193 { "uidinfo struct", &lock_class_mtx_sleep }, 194 { NULL, NULL }, 195 /* 196 * spin locks 197 */ 198#if defined(__i386__) && defined (SMP) 199 { "com", &lock_class_mtx_spin }, 200#endif 201 { "sio", &lock_class_mtx_spin }, 202#ifdef __i386__ 203 { "cy", &lock_class_mtx_spin }, 204#endif 205 { "ng_node", &lock_class_mtx_spin }, 206 { "ng_worklist", &lock_class_mtx_spin }, 207 { "ithread table lock", &lock_class_mtx_spin }, 208 { "ithread list lock", &lock_class_mtx_spin }, 209 { "sched lock", &lock_class_mtx_spin }, 210#ifdef __i386__ 211 { "clk", &lock_class_mtx_spin }, 212#endif 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 NULL, /* mtx_object.lo_file */ 251 0, /* mtx_object.lo_line */ 252 LO_INITIALIZED, /* mtx_object.lo_flags */ 253 { NULL }, /* mtx_object.lo_list */ 254 NULL }, /* mtx_object.lo_witness */ 255 MTX_UNOWNED, 0, /* mtx_lock, mtx_recurse */ 256 0, /* mtx_savecrit */ 257 TAILQ_HEAD_INITIALIZER(all_mtx.mtx_blocked), 258 { NULL, NULL } /* mtx_contested */ 259}; 260 261/* 262 * This global is set to 0 once it becomes safe to use the witness code. 263 */ 264static int witness_cold = 1; 265 266/* 267 * Global variables for book keeping. 268 */ 269static int lock_cur_cnt; 270static int lock_max_cnt; 271 272/* 273 * The WITNESS-enabled diagnostic code. 274 */ 275static void 276witness_initialize(void *dummy __unused) 277{ 278 struct lock_object *lock; 279 struct witness_order_list_entry *order; 280 struct witness *w, *w1; 281 int i; 282 283 /* 284 * We have to release Giant before initializing its witness 285 * structure so that WITNESS doesn't get confused. 286 */ 287 mtx_unlock(&Giant); 288 mtx_assert(&Giant, MA_NOTOWNED); 289 290 STAILQ_INSERT_HEAD(&all_locks, &all_mtx.mtx_object, lo_list); 291 mtx_init(&w_mtx, "witness lock", MTX_SPIN | MTX_QUIET | MTX_NOWITNESS); 292 for (i = 0; i < WITNESS_COUNT; i++) 293 witness_free(&w_data[i]); 294 for (i = 0; i < WITNESS_CHILDCOUNT; i++) 295 witness_child_free(&w_childdata[i]); 296 for (i = 0; i < LOCK_CHILDCOUNT; i++) 297 witness_lock_list_free(&w_locklistdata[i]); 298 299 /* First add in all the specified order lists. */ 300 for (order = order_lists; order->w_name != NULL; order++) { 301 w = enroll(order->w_name, order->w_class); 302 if (w == NULL) 303 continue; 304 w->w_file = "order list"; 305 for (order++; order->w_name != NULL; order++) { 306 w1 = enroll(order->w_name, order->w_class); 307 if (w1 == NULL) 308 continue; 309 w1->w_file = "order list"; 310 itismychild(w, w1); 311 w = w1; 312 } 313 } 314 315 /* Iterate through all locks and add them to witness. */ 316 mtx_lock(&all_mtx); 317 STAILQ_FOREACH(lock, &all_locks, lo_list) { 318 if (lock->lo_flags & LO_WITNESS) 319 lock->lo_witness = enroll(lock->lo_name, 320 lock->lo_class); 321 else 322 lock->lo_witness = NULL; 323 } 324 mtx_unlock(&all_mtx); 325 326 /* Mark the witness code as being ready for use. */ 327 atomic_store_rel_int(&witness_cold, 0); 328 329 mtx_lock(&Giant); 330} 331SYSINIT(witness_init, SI_SUB_WITNESS, SI_ORDER_FIRST, witness_initialize, NULL) 332 333void 334witness_init(struct lock_object *lock) 335{ 336 struct lock_class *class; 337 338 class = lock->lo_class; 339 if (lock->lo_flags & LO_INITIALIZED) 340 panic("%s: lock (%s) %s is already initialized!\n", __func__, 341 class->lc_name, lock->lo_name); 342 343 if ((lock->lo_flags & LO_RECURSABLE) != 0 && 344 (class->lc_flags & LC_RECURSABLE) == 0) 345 panic("%s: lock (%s) %s can not be recursable!\n", __func__, 346 class->lc_name, lock->lo_name); 347 348 if ((lock->lo_flags & LO_SLEEPABLE) != 0 && 349 (class->lc_flags & LC_SLEEPABLE) == 0) 350 panic("%s: lock (%s) %s can not be sleepable!\n", __func__, 351 class->lc_name, lock->lo_name); 352 353 mtx_lock(&all_mtx); 354 STAILQ_INSERT_TAIL(&all_locks, lock, lo_list); 355 lock->lo_flags |= LO_INITIALIZED; 356 lock_cur_cnt++; 357 if (lock_cur_cnt > lock_max_cnt) 358 lock_max_cnt = lock_cur_cnt; 359 mtx_unlock(&all_mtx); 360 if (!witness_cold && !witness_dead && 361 (lock->lo_flags & LO_WITNESS) != 0) 362 lock->lo_witness = enroll(lock->lo_name, class); 363 else 364 lock->lo_witness = NULL; 365} 366 367void 368witness_destroy(struct lock_object *lock) 369{ 370 struct witness *w; 371 372 if (witness_cold) 373 panic("lock (%s) %s destroyed while witness_cold", 374 lock->lo_class->lc_name, lock->lo_name); 375 376 if ((lock->lo_flags & LO_INITIALIZED) == 0) 377 panic("%s: lock (%s) %s is not initialized!\n", __func__, 378 lock->lo_class->lc_name, lock->lo_name); 379 380 if (lock->lo_flags & LO_LOCKED) 381 panic("lock (%s) %s destroyed while held", 382 lock->lo_class->lc_name, lock->lo_name); 383 384 w = lock->lo_witness; 385 if (w != NULL) { 386 mtx_lock_spin(&w_mtx); 387 w->w_refcount--; 388 if (w->w_refcount == 0) { 389 w->w_name = "(dead)"; 390 w->w_file = "(dead)"; 391 w->w_line = 0; 392 } 393 mtx_unlock_spin(&w_mtx); 394 } 395 396 mtx_lock(&all_mtx); 397 lock_cur_cnt--; 398 STAILQ_REMOVE(&all_locks, lock, lock_object, lo_list); 399 lock->lo_flags &= LO_INITIALIZED; 400 mtx_unlock(&all_mtx); 401} 402 403static void 404witness_display_list(void(*prnt)(const char *fmt, ...), 405 struct witness_list *list) 406{ 407 struct witness *w, *w1; 408 int found; 409 410 STAILQ_FOREACH(w, list, w_typelist) { 411 if (w->w_file == NULL) 412 continue; 413 found = 0; 414 STAILQ_FOREACH(w1, list, w_typelist) { 415 if (isitmychild(w1, w)) { 416 found++; 417 break; 418 } 419 } 420 if (found) 421 continue; 422 /* 423 * This lock has no anscestors, display its descendants. 424 */ 425 witness_displaydescendants(prnt, w); 426 } 427} 428 429static void 430witness_display(void(*prnt)(const char *fmt, ...)) 431{ 432 struct witness *w; 433 434 KASSERT(!witness_cold, ("%s: witness_cold\n", __func__)); 435 witness_levelall(); 436 437 /* 438 * First, handle sleep locks which have been acquired at least 439 * once. 440 */ 441 prnt("Sleep locks:\n"); 442 witness_display_list(prnt, &w_sleep); 443 444 /* 445 * Now do spin locks which have been acquired at least once. 446 */ 447 prnt("\nSpin locks:\n"); 448 witness_display_list(prnt, &w_spin); 449 450 /* 451 * Finally, any locks which have not been acquired yet. 452 */ 453 prnt("\nLocks which were never acquired:\n"); 454 STAILQ_FOREACH(w, &w_all, w_list) { 455 if (w->w_file != NULL) 456 continue; 457 prnt("%s\n", w->w_name); 458 } 459} 460 461void 462witness_lock(struct lock_object *lock, int flags, const char *file, int line) 463{ 464 struct lock_list_entry **lock_list, *lle; 465 struct lock_object *lock1, *lock2; 466 struct lock_class *class; 467 struct witness *w, *w1; 468 struct proc *p; 469 int i, j; 470#ifdef DDB 471 int go_into_ddb = 0; 472#endif /* DDB */ 473 474 if (witness_cold || witness_dead || lock->lo_witness == NULL || 475 panicstr) 476 return; 477 w = lock->lo_witness; 478 class = lock->lo_class; 479 p = curproc; 480 481 if ((lock->lo_flags & LO_LOCKED) == 0) 482 panic("%s: lock (%s) %s is not locked @ %s:%d", __func__, 483 class->lc_name, lock->lo_name, file, line); 484 485 if ((lock->lo_flags & LO_RECURSED) != 0) { 486 if ((lock->lo_flags & LO_RECURSABLE) == 0) 487 panic( 488 "%s: recursed on non-recursive lock (%s) %s @ %s:%d", 489 __func__, class->lc_name, lock->lo_name, file, 490 line); 491 return; 492 } 493 494 /* 495 * We have to hold a spinlock to keep lock_list valid across the check 496 * in the LC_SLEEPLOCK case. In the LC_SPINLOCK case, it is already 497 * protected by the spinlock we are currently performing the witness 498 * checks on, so it is ok to release the lock after performing this 499 * check. All we have to protect is the LC_SLEEPLOCK case when no 500 * spinlocks are held as we may get preempted during this check and 501 * lock_list could end up pointing to some other CPU's spinlock list. 502 */ 503 mtx_lock_spin(&w_mtx); 504 lock_list = PCPU_PTR(spinlocks); 505 if (class->lc_flags & LC_SLEEPLOCK) { 506 if (*lock_list != NULL) { 507 mtx_unlock_spin(&w_mtx); 508 panic("blockable sleep lock (%s) %s @ %s:%d", 509 class->lc_name, lock->lo_name, file, line); 510 } 511 lock_list = &p->p_sleeplocks; 512 } 513 mtx_unlock_spin(&w_mtx); 514 515 if (flags & LOP_TRYLOCK) 516 goto out; 517 518 /* 519 * Is this the first lock acquired? If so, then no order checking 520 * is needed. 521 */ 522 if (*lock_list == NULL) 523 goto out; 524 525 /* 526 * Check for duplicate locks of the same type. Note that we only 527 * have to check for this on the last lock we just acquired. Any 528 * other cases will be caught as lock order violations. 529 */ 530 lock1 = (*lock_list)->ll_children[(*lock_list)->ll_count - 1]; 531 w1 = lock1->lo_witness; 532 if (w1 == w) { 533 if (w->w_same_squawked || dup_ok(w)) 534 goto out; 535 w->w_same_squawked = 1; 536 printf("acquring duplicate lock of same type: \"%s\"\n", 537 lock->lo_name); 538 printf(" 1st @ %s:%d\n", w->w_file, w->w_line); 539 printf(" 2nd @ %s:%d\n", file, line); 540#ifdef DDB 541 go_into_ddb = 1; 542#endif /* DDB */ 543 goto out; 544 } 545 MPASS(!mtx_owned(&w_mtx)); 546 mtx_lock_spin(&w_mtx); 547 /* 548 * If we have a known higher number just say ok 549 */ 550 if (witness_watch > 1 && w->w_level > w1->w_level) { 551 mtx_unlock_spin(&w_mtx); 552 goto out; 553 } 554 if (isitmydescendant(w1, w)) { 555 mtx_unlock_spin(&w_mtx); 556 goto out; 557 } 558 for (j = 0, lle = *lock_list; lle != NULL; lle = lle->ll_next) { 559 for (i = lle->ll_count - 1; i >= 0; i--, j++) { 560 561 MPASS(j < WITNESS_COUNT); 562 lock1 = lle->ll_children[i]; 563 w1 = lock1->lo_witness; 564 565 /* 566 * If this lock doesn't undergo witness checking, 567 * then skip it. 568 */ 569 if (w1 == NULL) { 570 KASSERT((lock1->lo_flags & LO_WITNESS) == 0, 571 ("lock missing witness structure")); 572 continue; 573 } 574 if (!isitmydescendant(w, w1)) 575 continue; 576 /* 577 * We have a lock order violation, check to see if it 578 * is allowed or has already been yelled about. 579 */ 580 mtx_unlock_spin(&w_mtx); 581 if (blessed(w, w1)) 582 goto out; 583 if (lock1 == &Giant.mtx_object) { 584 if (w1->w_Giant_squawked) 585 goto out; 586 else 587 w1->w_Giant_squawked = 1; 588 } else { 589 if (w1->w_other_squawked) 590 goto out; 591 else 592 w1->w_other_squawked = 1; 593 } 594 /* 595 * Ok, yell about it. 596 */ 597 printf("lock order reversal\n"); 598 /* 599 * Try to locate an earlier lock with 600 * witness w in our list. 601 */ 602 do { 603 lock2 = lle->ll_children[i]; 604 MPASS(lock2 != NULL); 605 if (lock2->lo_witness == w) 606 break; 607 i--; 608 if (i == 0 && lle->ll_next != NULL) { 609 lle = lle->ll_next; 610 i = lle->ll_count - 1; 611 MPASS(i != 0); 612 } 613 } while (i >= 0); 614 if (i < 0) 615 /* 616 * We are very likely bogus in this case. 617 */ 618 printf(" 1st %s last acquired @ %s:%d\n", 619 w->w_name, w->w_file, w->w_line); 620 else 621 printf(" 1st %p %s @ %s:%d\n", lock2, 622 lock2->lo_name, lock2->lo_file, 623 lock2->lo_line); 624 printf(" 2nd %p %s @ %s:%d\n", 625 lock1, lock1->lo_name, lock1->lo_file, 626 lock1->lo_line); 627 printf(" 3rd %p %s @ %s:%d\n", 628 lock, lock->lo_name, file, line); 629#ifdef DDB 630 go_into_ddb = 1; 631#endif /* DDB */ 632 goto out; 633 } 634 } 635 lock1 = (*lock_list)->ll_children[(*lock_list)->ll_count - 1]; 636 if (!itismychild(lock1->lo_witness, w)) 637 mtx_unlock_spin(&w_mtx); 638 639out: 640#ifdef DDB 641 if (witness_ddb && go_into_ddb) 642 Debugger("witness_enter"); 643#endif /* DDB */ 644 w->w_file = file; 645 w->w_line = line; 646 lock->lo_line = line; 647 lock->lo_file = file; 648 649 lle = *lock_list; 650 if (lle == NULL || lle->ll_count == LOCK_CHILDCOUNT) { 651 *lock_list = witness_lock_list_get(); 652 if (*lock_list == NULL) 653 return; 654 (*lock_list)->ll_next = lle; 655 lle = *lock_list; 656 } 657 lle->ll_children[lle->ll_count++] = lock; 658} 659 660void 661witness_unlock(struct lock_object *lock, int flags, const char *file, int line) 662{ 663 struct lock_list_entry **lock_list, *lle; 664 struct lock_class *class; 665 struct proc *p; 666 int i, j; 667 668 if (witness_cold || witness_dead || lock->lo_witness == NULL || 669 panicstr) 670 return; 671 p = curproc; 672 class = lock->lo_class; 673 674 if (lock->lo_flags & LO_RECURSED) { 675 if ((lock->lo_flags & LO_LOCKED) == 0) 676 panic("%s: recursed lock (%s) %s is not locked @ %s:%d", 677 __func__, class->lc_name, lock->lo_name, file, 678 line); 679 return; 680 } 681 682 /* 683 * We don't need to protect this PCPU_GET() here against preemption 684 * because if we hold any spinlocks then we are already protected, 685 * and if we don't we will get NULL if we hold no spinlocks even if 686 * we switch CPU's while reading it. 687 */ 688 if (class->lc_flags & LC_SLEEPLOCK) { 689 if ((flags & LOP_NOSWITCH) == 0 && PCPU_GET(spinlocks) != NULL) 690 panic("switchable sleep unlock (%s) %s @ %s:%d", 691 class->lc_name, lock->lo_name, file, line); 692 lock_list = &p->p_sleeplocks; 693 } else 694 lock_list = PCPU_PTR(spinlocks); 695 696 for (; *lock_list != NULL; lock_list = &(*lock_list)->ll_next) 697 for (i = 0; i < (*lock_list)->ll_count; i++) 698 if ((*lock_list)->ll_children[i] == lock) { 699 (*lock_list)->ll_count--; 700 for (j = i; j < (*lock_list)->ll_count; j++) 701 (*lock_list)->ll_children[j] = 702 (*lock_list)->ll_children[j + 1]; 703 if ((*lock_list)->ll_count == 0) { 704 lle = *lock_list; 705 *lock_list = lle->ll_next; 706 witness_lock_list_free(lle); 707 } 708 return; 709 } 710} 711 712/* 713 * Warn if any held locks are not sleepable. Note that Giant and the lock 714 * passed in are both special cases since they are both released during the 715 * sleep process and aren't actually held while the process is asleep. 716 */ 717int 718witness_sleep(int check_only, struct lock_object *lock, const char *file, 719 int line) 720{ 721 struct lock_list_entry **lock_list, *lle; 722 struct lock_object *lock1; 723 struct proc *p; 724 critical_t savecrit; 725 int i, n; 726 727 if (witness_dead || panicstr) 728 return (0); 729 KASSERT(!witness_cold, ("%s: witness_cold\n", __func__)); 730 n = 0; 731 /* 732 * Preemption bad because we need PCPU_PTR(spinlocks) to not change. 733 */ 734 savecrit = critical_enter(); 735 p = curproc; 736 lock_list = &p->p_sleeplocks; 737again: 738 for (lle = *lock_list; lle != NULL; lle = lle->ll_next) 739 for (i = lle->ll_count - 1; i >= 0; i--) { 740 lock1 = lle->ll_children[i]; 741 if (lock1 == lock || lock1 == &Giant.mtx_object || 742 (lock1->lo_flags & LO_SLEEPABLE)) 743 continue; 744 n++; 745 printf("%s:%d: %s with \"%s\" locked from %s:%d\n", 746 file, line, check_only ? "could sleep" : "sleeping", 747 lock1->lo_name, lock1->lo_file, lock1->lo_line); 748 } 749 if (lock_list == &p->p_sleeplocks) { 750 lock_list = PCPU_PTR(spinlocks); 751 goto again; 752 } 753#ifdef DDB 754 if (witness_ddb && n) 755 Debugger("witness_sleep"); 756#endif /* DDB */ 757 critical_exit(savecrit); 758 return (n); 759} 760 761static struct witness * 762enroll(const char *description, struct lock_class *lock_class) 763{ 764 struct witness *w; 765 766 if (!witness_watch) 767 return (NULL); 768 769 if ((lock_class->lc_flags & LC_SPINLOCK) && witness_skipspin) 770 return (NULL); 771 mtx_lock_spin(&w_mtx); 772 STAILQ_FOREACH(w, &w_all, w_list) { 773 if (strcmp(description, w->w_name) == 0) { 774 w->w_refcount++; 775 mtx_unlock_spin(&w_mtx); 776 if (lock_class != w->w_class) 777 panic( 778 "lock (%s) %s does not match earlier (%s) lock", 779 description, lock_class->lc_name, 780 w->w_class->lc_name); 781 return (w); 782 } 783 } 784 /* 785 * This isn't quite right, as witness_cold is still 0 while we 786 * enroll all the locks initialized before witness_initialize(). 787 */ 788 if ((lock_class->lc_flags & LC_SPINLOCK) && !witness_cold) { 789 mtx_unlock_spin(&w_mtx); 790 panic("spin lock %s not in order list", description); 791 } 792 if ((w = witness_get()) == NULL) 793 return (NULL); 794 w->w_name = description; 795 w->w_class = lock_class; 796 w->w_refcount = 1; 797 STAILQ_INSERT_HEAD(&w_all, w, w_list); 798 if (lock_class->lc_flags & LC_SPINLOCK) 799 STAILQ_INSERT_HEAD(&w_spin, w, w_typelist); 800 else if (lock_class->lc_flags & LC_SLEEPLOCK) 801 STAILQ_INSERT_HEAD(&w_sleep, w, w_typelist); 802 else { 803 mtx_unlock_spin(&w_mtx); 804 panic("lock class %s is not sleep or spin", 805 lock_class->lc_name); 806 } 807 mtx_unlock_spin(&w_mtx); 808 809 return (w); 810} 811 812static int 813itismychild(struct witness *parent, struct witness *child) 814{ 815 static int recursed; 816 struct witness_child_list_entry **wcl; 817 struct witness_list *list; 818 819 MPASS(child != NULL && parent != NULL); 820 if ((parent->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK)) != 821 (child->w_class->lc_flags & (LC_SLEEPLOCK | LC_SPINLOCK))) 822 panic( 823 "%s: parent (%s) and child (%s) are not the same lock type", 824 __func__, parent->w_class->lc_name, 825 child->w_class->lc_name); 826 827 /* 828 * Insert "child" after "parent" 829 */ 830 wcl = &parent->w_children; 831 while (*wcl != NULL && (*wcl)->wcl_count == WITNESS_NCHILDREN) 832 wcl = &(*wcl)->wcl_next; 833 834 if (*wcl == NULL) { 835 *wcl = witness_child_get(); 836 if (*wcl == NULL) 837 return (1); 838 } 839 840 (*wcl)->wcl_children[(*wcl)->wcl_count++] = child; 841 842 /* 843 * Now prune whole tree. We look for cases where a lock is now 844 * both a descendant and a direct child of a given lock. In that 845 * case, we want to remove the direct child link from the tree. 846 */ 847 if (recursed) 848 return (0); 849 recursed = 1; 850 if (parent->w_class->lc_flags & LC_SLEEPLOCK) 851 list = &w_sleep; 852 else 853 list = &w_spin; 854 STAILQ_FOREACH(child, list, w_typelist) { 855 STAILQ_FOREACH(parent, list, w_typelist) { 856 if (!isitmychild(parent, child)) 857 continue; 858 removechild(parent, child); 859 if (isitmydescendant(parent, child)) 860 continue; 861 itismychild(parent, child); 862 } 863 } 864 recursed = 0; 865 witness_levelall(); 866 return (0); 867} 868 869static void 870removechild(struct witness *parent, struct witness *child) 871{ 872 struct witness_child_list_entry **wcl, *wcl1; 873 int i; 874 875 for (wcl = &parent->w_children; *wcl != NULL; wcl = &(*wcl)->wcl_next) 876 for (i = 0; i < (*wcl)->wcl_count; i++) 877 if ((*wcl)->wcl_children[i] == child) 878 goto found; 879 return; 880found: 881 (*wcl)->wcl_count--; 882 if ((*wcl)->wcl_count > i) 883 (*wcl)->wcl_children[i] = 884 (*wcl)->wcl_children[(*wcl)->wcl_count]; 885 MPASS((*wcl)->wcl_children[i] != NULL); 886 887 if ((*wcl)->wcl_count != 0) 888 return; 889 890 wcl1 = *wcl; 891 *wcl = wcl1->wcl_next; 892 witness_child_free(wcl1); 893} 894 895static int 896isitmychild(struct witness *parent, struct witness *child) 897{ 898 struct witness_child_list_entry *wcl; 899 int i; 900 901 for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) { 902 for (i = 0; i < wcl->wcl_count; i++) { 903 if (wcl->wcl_children[i] == child) 904 return (1); 905 } 906 } 907 return (0); 908} 909 910static int 911isitmydescendant(struct witness *parent, struct witness *child) 912{ 913 struct witness_child_list_entry *wcl; 914 int i, j; 915 916 if (isitmychild(parent, child)) 917 return (1); 918 j = 0; 919 for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) { 920 MPASS(j < 1000); 921 for (i = 0; i < wcl->wcl_count; i++) { 922 if (isitmydescendant(wcl->wcl_children[i], child)) 923 return (1); 924 } 925 j++; 926 } 927 return (0); 928} 929 930void 931witness_levelall (void) 932{ 933 struct witness_list *list; 934 struct witness *w, *w1; 935 936 /* 937 * First clear all levels. 938 */ 939 STAILQ_FOREACH(w, &w_all, w_list) { 940 w->w_level = 0; 941 } 942 943 /* 944 * Look for locks with no parent and level all their descendants. 945 */ 946 STAILQ_FOREACH(w, &w_all, w_list) { 947 /* 948 * This is just an optimization, technically we could get 949 * away just walking the all list each time. 950 */ 951 if (w->w_class->lc_flags & LC_SLEEPLOCK) 952 list = &w_sleep; 953 else 954 list = &w_spin; 955 STAILQ_FOREACH(w1, list, w_typelist) { 956 if (isitmychild(w1, w)) 957 goto skip; 958 } 959 witness_leveldescendents(w, 0); 960 skip: 961 } 962} 963 964static void 965witness_leveldescendents(struct witness *parent, int level) 966{ 967 struct witness_child_list_entry *wcl; 968 int i; 969 970 if (parent->w_level < level) 971 parent->w_level = level; 972 level++; 973 for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) 974 for (i = 0; i < wcl->wcl_count; i++) 975 witness_leveldescendents(wcl->wcl_children[i], level); 976} 977 978static void 979witness_displaydescendants(void(*prnt)(const char *fmt, ...), 980 struct witness *parent) 981{ 982 struct witness_child_list_entry *wcl; 983 int i, level; 984 985 level = parent->w_level; 986 987 prnt("%-2d", level); 988 for (i = 0; i < level; i++) 989 prnt(" "); 990 prnt("%s", parent->w_name); 991 if (parent->w_file != NULL) 992 prnt(" -- last acquired @ %s:%d\n", parent->w_file, 993 parent->w_line); 994 995 for (wcl = parent->w_children; wcl != NULL; wcl = wcl->wcl_next) 996 for (i = 0; i < wcl->wcl_count; i++) 997 witness_displaydescendants(prnt, 998 wcl->wcl_children[i]); 999} 1000 1001static int 1002dup_ok(struct witness *w) 1003{ 1004 const char **dup; 1005 1006 for (dup = dup_list; *dup != NULL; dup++) 1007 if (strcmp(w->w_name, *dup) == 0) 1008 return (1); 1009 return (0); 1010} 1011 1012static int 1013blessed(struct witness *w1, struct witness *w2) 1014{ 1015 int i; 1016 struct witness_blessed *b; 1017 1018 for (i = 0; i < blessed_count; i++) { 1019 b = &blessed_list[i]; 1020 if (strcmp(w1->w_name, b->b_lock1) == 0) { 1021 if (strcmp(w2->w_name, b->b_lock2) == 0) 1022 return (1); 1023 continue; 1024 } 1025 if (strcmp(w1->w_name, b->b_lock2) == 0) 1026 if (strcmp(w2->w_name, b->b_lock1) == 0) 1027 return (1); 1028 } 1029 return (0); 1030} 1031 1032static struct witness * 1033witness_get(void) 1034{ 1035 struct witness *w; 1036 1037 if (STAILQ_EMPTY(&w_free)) { 1038 witness_dead = 1; 1039 mtx_unlock_spin(&w_mtx); 1040 printf("%s: witness exhausted\n", __func__); 1041 return (NULL); 1042 } 1043 w = STAILQ_FIRST(&w_free); 1044 STAILQ_REMOVE_HEAD(&w_free, w_list); 1045 bzero(w, sizeof(*w)); 1046 return (w); 1047} 1048 1049static void 1050witness_free(struct witness *w) 1051{ 1052 1053 STAILQ_INSERT_HEAD(&w_free, w, w_list); 1054} 1055 1056static struct witness_child_list_entry * 1057witness_child_get(void) 1058{ 1059 struct witness_child_list_entry *wcl; 1060 1061 wcl = w_child_free; 1062 if (wcl == NULL) { 1063 witness_dead = 1; 1064 mtx_unlock_spin(&w_mtx); 1065 printf("%s: witness exhausted\n", __func__); 1066 return (NULL); 1067 } 1068 w_child_free = wcl->wcl_next; 1069 bzero(wcl, sizeof(*wcl)); 1070 return (wcl); 1071} 1072 1073static void 1074witness_child_free(struct witness_child_list_entry *wcl) 1075{ 1076 1077 wcl->wcl_next = w_child_free; 1078 w_child_free = wcl; 1079} 1080 1081static struct lock_list_entry * 1082witness_lock_list_get(void) 1083{ 1084 struct lock_list_entry *lle; 1085 1086 mtx_lock_spin(&w_mtx); 1087 lle = w_lock_list_free; 1088 if (lle == NULL) { 1089 witness_dead = 1; 1090 mtx_unlock_spin(&w_mtx); 1091 printf("%s: witness exhausted\n", __func__); 1092 return (NULL); 1093 } 1094 w_lock_list_free = lle->ll_next; 1095 mtx_unlock_spin(&w_mtx); 1096 bzero(lle, sizeof(*lle)); 1097 return (lle); 1098} 1099 1100static void 1101witness_lock_list_free(struct lock_list_entry *lle) 1102{ 1103 1104 mtx_lock_spin(&w_mtx); 1105 lle->ll_next = w_lock_list_free; 1106 w_lock_list_free = lle; 1107 mtx_unlock_spin(&w_mtx); 1108} 1109 1110int 1111witness_list_locks(struct lock_list_entry **lock_list) 1112{ 1113 struct lock_list_entry *lle; 1114 struct lock_object *lock; 1115 int i, nheld; 1116 1117 nheld = 0; 1118 for (lle = *lock_list; lle != NULL; lle = lle->ll_next) 1119 for (i = lle->ll_count - 1; i >= 0; i--) { 1120 lock = lle->ll_children[i]; 1121 printf("\t(%s) %s (%p) locked at %s:%d\n", 1122 lock->lo_class->lc_name, lock->lo_name, lock, 1123 lock->lo_file, lock->lo_line); 1124 nheld++; 1125 } 1126 return (nheld); 1127} 1128 1129/* 1130 * Calling this on p != curproc is bad unless we are in ddb. 1131 */ 1132int 1133witness_list(struct proc *p) 1134{ 1135 critical_t savecrit; 1136 int nheld; 1137 1138 KASSERT(p == curproc || db_active, 1139 ("%s: p != curproc and we aren't in the debugger", __func__)); 1140 KASSERT(!witness_cold, ("%s: witness_cold", __func__)); 1141 1142 nheld = witness_list_locks(&p->p_sleeplocks); 1143 1144 /* 1145 * We only handle spinlocks if p == curproc. This is somewhat broken 1146 * if p is currently executing on some other CPU and holds spin locks 1147 * as we won't display those locks. If we had a MI way of getting 1148 * the per-cpu data for a given cpu then we could use p->p_oncpu to 1149 * get the list of spinlocks for this process and "fix" this. 1150 */ 1151 if (p == curproc) { 1152 /* 1153 * Preemption bad because we need PCPU_PTR(spinlocks) to not 1154 * change. 1155 */ 1156 savecrit = critical_enter(); 1157 nheld += witness_list_locks(PCPU_PTR(spinlocks)); 1158 critical_exit(savecrit); 1159 } 1160 1161 return (nheld); 1162} 1163 1164void 1165witness_save(struct lock_object *lock, const char **filep, int *linep) 1166{ 1167 1168 KASSERT(!witness_cold, ("%s: witness_cold\n", __func__)); 1169 if (lock->lo_witness == NULL) 1170 return; 1171 1172 *filep = lock->lo_file; 1173 *linep = lock->lo_line; 1174} 1175 1176void 1177witness_restore(struct lock_object *lock, const char *file, int line) 1178{ 1179 1180 KASSERT(!witness_cold, ("%s: witness_cold\n", __func__)); 1181 if (lock->lo_witness == NULL) 1182 return; 1183 1184 lock->lo_witness->w_file = file; 1185 lock->lo_witness->w_line = line; 1186 lock->lo_file = file; 1187 lock->lo_line = line; 1188} 1189 1190#ifdef DDB 1191 1192DB_SHOW_COMMAND(locks, db_witness_list) 1193{ 1194 struct proc *p; 1195 pid_t pid; 1196 1197 if (have_addr) { 1198 pid = (addr % 16) + ((addr >> 4) % 16) * 10 + 1199 ((addr >> 8) % 16) * 100 + ((addr >> 12) % 16) * 1000 + 1200 ((addr >> 16) % 16) * 10000; 1201 1202 /* sx_slock(&allproc_lock); */ 1203 LIST_FOREACH(p, &allproc, p_list) { 1204 if (p->p_pid == pid) 1205 break; 1206 } 1207 /* sx_sunlock(&allproc_lock); */ 1208 if (p == NULL) { 1209 db_printf("pid %d not found\n", pid); 1210 return; 1211 } 1212 } else 1213 p = curproc; 1214 1215 witness_list(p); 1216} 1217 1218DB_SHOW_COMMAND(witness, db_witness_display) 1219{ 1220 1221 witness_display(db_printf); 1222} 1223#endif 1224