1 /*- 2 * SPDX-License-Identifier: BSD-2-Clause 3 * 4 * Copyright (c) 2008 Attilio Rao <attilio@FreeBSD.org> 5 * All rights reserved. 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 1. Redistributions of source code must retain the above copyright 11 * notice(s), this list of conditions and the following disclaimer as 12 * the first lines of this file unmodified other than the possible 13 * addition of one or more copyright notices. 14 * 2. Redistributions in binary form must reproduce the above copyright 15 * notice(s), this list of conditions and the following disclaimer in the 16 * documentation and/or other materials provided with the distribution. 17 * 18 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDER(S) ``AS IS'' AND ANY 19 * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED 20 * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE 21 * DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER(S) BE LIABLE FOR ANY 22 * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES 23 * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR 24 * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER 25 * CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 26 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 27 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH 28 * DAMAGE. 29 */ 30 31 #include "opt_ddb.h" 32 #include "opt_hwpmc_hooks.h" 33 34 #include <sys/param.h> 35 #include <sys/kdb.h> 36 #include <sys/ktr.h> 37 #include <sys/limits.h> 38 #include <sys/lock.h> 39 #include <sys/lock_profile.h> 40 #include <sys/lockmgr.h> 41 #include <sys/lockstat.h> 42 #include <sys/mutex.h> 43 #include <sys/proc.h> 44 #include <sys/sleepqueue.h> 45 #ifdef DEBUG_LOCKS 46 #include <sys/stack.h> 47 #endif 48 #include <sys/sysctl.h> 49 #include <sys/systm.h> 50 51 #include <machine/cpu.h> 52 53 #ifdef DDB 54 #include <ddb/ddb.h> 55 #endif 56 57 #ifdef HWPMC_HOOKS 58 #include <sys/pmckern.h> 59 PMC_SOFT_DECLARE( , , lock, failed); 60 #endif 61 62 /* 63 * Hack. There should be prio_t or similar so that this is not necessary. 64 */ 65 _Static_assert((PRILASTFLAG * 2) - 1 <= USHRT_MAX, 66 "prio flags wont fit in u_short pri in struct lock"); 67 68 CTASSERT(LK_UNLOCKED == (LK_UNLOCKED & 69 ~(LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS))); 70 71 #define SQ_EXCLUSIVE_QUEUE 0 72 #define SQ_SHARED_QUEUE 1 73 74 #ifndef INVARIANTS 75 #define _lockmgr_assert(lk, what, file, line) 76 #endif 77 78 #define TD_SLOCKS_INC(td) ((td)->td_lk_slocks++) 79 #define TD_SLOCKS_DEC(td) ((td)->td_lk_slocks--) 80 81 #ifndef DEBUG_LOCKS 82 #define STACK_PRINT(lk) 83 #define STACK_SAVE(lk) 84 #define STACK_ZERO(lk) 85 #else 86 #define STACK_PRINT(lk) stack_print_ddb(&(lk)->lk_stack) 87 #define STACK_SAVE(lk) stack_save(&(lk)->lk_stack) 88 #define STACK_ZERO(lk) stack_zero(&(lk)->lk_stack) 89 #endif 90 91 #define LOCK_LOG2(lk, string, arg1, arg2) \ 92 if (LOCK_LOG_TEST(&(lk)->lock_object, 0)) \ 93 CTR2(KTR_LOCK, (string), (arg1), (arg2)) 94 #define LOCK_LOG3(lk, string, arg1, arg2, arg3) \ 95 if (LOCK_LOG_TEST(&(lk)->lock_object, 0)) \ 96 CTR3(KTR_LOCK, (string), (arg1), (arg2), (arg3)) 97 98 #define GIANT_DECLARE \ 99 int _i = 0; \ 100 WITNESS_SAVE_DECL(Giant) 101 #define GIANT_RESTORE() do { \ 102 if (__predict_false(_i > 0)) { \ 103 while (_i--) \ 104 mtx_lock(&Giant); \ 105 WITNESS_RESTORE(&Giant.lock_object, Giant); \ 106 } \ 107 } while (0) 108 #define GIANT_SAVE() do { \ 109 if (__predict_false(mtx_owned(&Giant))) { \ 110 WITNESS_SAVE(&Giant.lock_object, Giant); \ 111 while (mtx_owned(&Giant)) { \ 112 _i++; \ 113 mtx_unlock(&Giant); \ 114 } \ 115 } \ 116 } while (0) 117 118 static __always_inline bool 119 LK_CAN_SHARE(uintptr_t x, int flags, bool fp) 120 { 121 122 if ((x & (LK_SHARE | LK_EXCLUSIVE_WAITERS | LK_EXCLUSIVE_SPINNERS)) == 123 LK_SHARE) 124 return (true); 125 if (fp || (!(x & LK_SHARE))) 126 return (false); 127 if ((curthread->td_lk_slocks != 0 && !(flags & LK_NODDLKTREAT)) || 128 (curthread->td_pflags & TDP_DEADLKTREAT)) 129 return (true); 130 return (false); 131 } 132 133 #define LK_TRYOP(x) \ 134 ((x) & LK_NOWAIT) 135 136 #define LK_CAN_WITNESS(x) \ 137 (((x) & LK_NOWITNESS) == 0 && !LK_TRYOP(x)) 138 #define LK_TRYWIT(x) \ 139 (LK_TRYOP(x) ? LOP_TRYLOCK : 0) 140 141 #define lockmgr_xlocked_v(v) \ 142 (((v) & ~(LK_FLAGMASK & ~LK_SHARE)) == (uintptr_t)curthread) 143 144 #define lockmgr_xlocked(lk) lockmgr_xlocked_v(lockmgr_read_value(lk)) 145 146 static void assert_lockmgr(const struct lock_object *lock, int how); 147 #ifdef DDB 148 static void db_show_lockmgr(const struct lock_object *lock); 149 #endif 150 static void lock_lockmgr(struct lock_object *lock, uintptr_t how); 151 #ifdef KDTRACE_HOOKS 152 static int owner_lockmgr(const struct lock_object *lock, 153 struct thread **owner); 154 #endif 155 static uintptr_t unlock_lockmgr(struct lock_object *lock); 156 157 struct lock_class lock_class_lockmgr = { 158 .lc_name = "lockmgr", 159 .lc_flags = LC_RECURSABLE | LC_SLEEPABLE | LC_SLEEPLOCK | LC_UPGRADABLE, 160 .lc_assert = assert_lockmgr, 161 #ifdef DDB 162 .lc_ddb_show = db_show_lockmgr, 163 #endif 164 .lc_lock = lock_lockmgr, 165 .lc_unlock = unlock_lockmgr, 166 #ifdef KDTRACE_HOOKS 167 .lc_owner = owner_lockmgr, 168 #endif 169 }; 170 171 static __read_mostly bool lk_adaptive = true; 172 static SYSCTL_NODE(_debug, OID_AUTO, lockmgr, CTLFLAG_RD, NULL, "lockmgr debugging"); 173 SYSCTL_BOOL(_debug_lockmgr, OID_AUTO, adaptive_spinning, CTLFLAG_RW, &lk_adaptive, 174 0, ""); 175 #define lockmgr_delay locks_delay 176 177 struct lockmgr_wait { 178 const char *iwmesg; 179 int ipri; 180 int itimo; 181 }; 182 183 static __always_inline bool lockmgr_slock_try(struct lock *lk, uintptr_t *xp, 184 int flags, bool fp); 185 static __always_inline bool lockmgr_sunlock_try(struct lock *lk, 186 uintptr_t *xp); 187 188 static void 189 lockmgr_exit(u_int flags, struct lock_object *ilk) 190 { 191 struct lock_class *class; 192 193 if (flags & LK_INTERLOCK) { 194 class = LOCK_CLASS(ilk); 195 class->lc_unlock(ilk); 196 } 197 } 198 199 static void 200 lockmgr_note_shared_acquire(struct lock *lk, int contested, 201 uint64_t waittime, const char *file, int line, int flags) 202 { 203 204 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(lockmgr__acquire, lk, contested, 205 waittime, file, line, LOCKSTAT_READER); 206 LOCK_LOG_LOCK("SLOCK", &lk->lock_object, 0, 0, file, line); 207 WITNESS_LOCK(&lk->lock_object, LK_TRYWIT(flags), file, line); 208 TD_LOCKS_INC(curthread); 209 TD_SLOCKS_INC(curthread); 210 STACK_SAVE(lk); 211 } 212 213 static void 214 lockmgr_note_shared_release(struct lock *lk, const char *file, int line) 215 { 216 217 WITNESS_UNLOCK(&lk->lock_object, 0, file, line); 218 LOCK_LOG_LOCK("SUNLOCK", &lk->lock_object, 0, 0, file, line); 219 TD_LOCKS_DEC(curthread); 220 TD_SLOCKS_DEC(curthread); 221 } 222 223 static void 224 lockmgr_note_exclusive_acquire(struct lock *lk, int contested, 225 uint64_t waittime, const char *file, int line, int flags) 226 { 227 228 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(lockmgr__acquire, lk, contested, 229 waittime, file, line, LOCKSTAT_WRITER); 230 LOCK_LOG_LOCK("XLOCK", &lk->lock_object, 0, lk->lk_recurse, file, line); 231 WITNESS_LOCK(&lk->lock_object, LOP_EXCLUSIVE | LK_TRYWIT(flags), file, 232 line); 233 TD_LOCKS_INC(curthread); 234 STACK_SAVE(lk); 235 } 236 237 static void 238 lockmgr_note_exclusive_release(struct lock *lk, const char *file, int line) 239 { 240 241 if (!lockmgr_disowned(lk)) { 242 WITNESS_UNLOCK(&lk->lock_object, LOP_EXCLUSIVE, file, line); 243 TD_LOCKS_DEC(curthread); 244 } 245 LOCK_LOG_LOCK("XUNLOCK", &lk->lock_object, 0, lk->lk_recurse, file, 246 line); 247 } 248 249 static __inline struct thread * 250 lockmgr_xholder(const struct lock *lk) 251 { 252 uintptr_t x; 253 254 x = lockmgr_read_value(lk); 255 return ((x & LK_SHARE) ? NULL : (struct thread *)LK_HOLDER(x)); 256 } 257 258 /* 259 * It assumes sleepq_lock held and returns with this one unheld. 260 * It also assumes the generic interlock is sane and previously checked. 261 * If LK_INTERLOCK is specified the interlock is not reacquired after the 262 * sleep. 263 */ 264 static __inline int 265 sleeplk(struct lock *lk, u_int flags, struct lock_object *ilk, 266 const char *wmesg, int pri, int timo, int queue) 267 { 268 GIANT_DECLARE; 269 struct lock_class *class; 270 int catch, error; 271 272 class = (flags & LK_INTERLOCK) ? LOCK_CLASS(ilk) : NULL; 273 catch = pri & PCATCH; 274 pri &= PRIMASK; 275 error = 0; 276 277 LOCK_LOG3(lk, "%s: %p blocking on the %s sleepqueue", __func__, lk, 278 (queue == SQ_EXCLUSIVE_QUEUE) ? "exclusive" : "shared"); 279 280 if (flags & LK_INTERLOCK) 281 class->lc_unlock(ilk); 282 if (queue == SQ_EXCLUSIVE_QUEUE && (flags & LK_SLEEPFAIL) != 0) { 283 if (lk->lk_exslpfail < USHRT_MAX) 284 lk->lk_exslpfail++; 285 } 286 GIANT_SAVE(); 287 sleepq_add(&lk->lock_object, NULL, wmesg, SLEEPQ_LK | (catch ? 288 SLEEPQ_INTERRUPTIBLE : 0), queue); 289 if ((flags & LK_TIMELOCK) && timo) 290 sleepq_set_timeout(&lk->lock_object, timo); 291 292 /* 293 * Decisional switch for real sleeping. 294 */ 295 if ((flags & LK_TIMELOCK) && timo && catch) 296 error = sleepq_timedwait_sig(&lk->lock_object, pri); 297 else if ((flags & LK_TIMELOCK) && timo) 298 error = sleepq_timedwait(&lk->lock_object, pri); 299 else if (catch) 300 error = sleepq_wait_sig(&lk->lock_object, pri); 301 else 302 sleepq_wait(&lk->lock_object, pri); 303 GIANT_RESTORE(); 304 if ((flags & LK_SLEEPFAIL) && error == 0) 305 error = ENOLCK; 306 307 return (error); 308 } 309 310 static __inline void 311 wakeupshlk(struct lock *lk, const char *file, int line) 312 { 313 uintptr_t v, x, orig_x; 314 u_int realexslp; 315 int queue; 316 317 for (;;) { 318 x = lockmgr_read_value(lk); 319 if (lockmgr_sunlock_try(lk, &x)) 320 break; 321 322 /* 323 * We should have a sharer with waiters, so enter the hard 324 * path in order to handle wakeups correctly. 325 */ 326 sleepq_lock(&lk->lock_object); 327 orig_x = lockmgr_read_value(lk); 328 retry_sleepq: 329 x = orig_x & (LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS); 330 v = LK_UNLOCKED; 331 332 /* 333 * If the lock has exclusive waiters, give them preference in 334 * order to avoid deadlock with shared runners up. 335 * If interruptible sleeps left the exclusive queue empty 336 * avoid a starvation for the threads sleeping on the shared 337 * queue by giving them precedence and cleaning up the 338 * exclusive waiters bit anyway. 339 * Please note that lk_exslpfail count may be lying about 340 * the real number of waiters with the LK_SLEEPFAIL flag on 341 * because they may be used in conjunction with interruptible 342 * sleeps so lk_exslpfail might be considered an 'upper limit' 343 * bound, including the edge cases. 344 */ 345 realexslp = sleepq_sleepcnt(&lk->lock_object, 346 SQ_EXCLUSIVE_QUEUE); 347 if ((x & LK_EXCLUSIVE_WAITERS) != 0 && realexslp != 0) { 348 if (lk->lk_exslpfail != USHRT_MAX && lk->lk_exslpfail < realexslp) { 349 lk->lk_exslpfail = 0; 350 queue = SQ_EXCLUSIVE_QUEUE; 351 v |= (x & LK_SHARED_WAITERS); 352 } else { 353 lk->lk_exslpfail = 0; 354 LOCK_LOG2(lk, 355 "%s: %p has only LK_SLEEPFAIL sleepers", 356 __func__, lk); 357 LOCK_LOG2(lk, 358 "%s: %p waking up threads on the exclusive queue", 359 __func__, lk); 360 sleepq_broadcast(&lk->lock_object, SLEEPQ_LK, 0, 361 SQ_EXCLUSIVE_QUEUE); 362 queue = SQ_SHARED_QUEUE; 363 } 364 } else { 365 /* 366 * Exclusive waiters sleeping with LK_SLEEPFAIL on 367 * and using interruptible sleeps/timeout may have 368 * left spourious lk_exslpfail counts on, so clean 369 * it up anyway. 370 */ 371 lk->lk_exslpfail = 0; 372 queue = SQ_SHARED_QUEUE; 373 } 374 375 if (lockmgr_sunlock_try(lk, &orig_x)) { 376 sleepq_release(&lk->lock_object); 377 break; 378 } 379 380 x |= LK_SHARERS_LOCK(1); 381 if (!atomic_fcmpset_rel_ptr(&lk->lk_lock, &x, v)) { 382 orig_x = x; 383 goto retry_sleepq; 384 } 385 LOCK_LOG3(lk, "%s: %p waking up threads on the %s queue", 386 __func__, lk, queue == SQ_SHARED_QUEUE ? "shared" : 387 "exclusive"); 388 sleepq_broadcast(&lk->lock_object, SLEEPQ_LK, 0, queue); 389 sleepq_release(&lk->lock_object); 390 break; 391 } 392 393 LOCKSTAT_PROFILE_RELEASE_RWLOCK(lockmgr__release, lk, LOCKSTAT_READER); 394 } 395 396 static void 397 assert_lockmgr(const struct lock_object *lock, int what) 398 { 399 400 panic("lockmgr locks do not support assertions"); 401 } 402 403 static void 404 lock_lockmgr(struct lock_object *lock, uintptr_t how) 405 { 406 407 panic("lockmgr locks do not support sleep interlocking"); 408 } 409 410 static uintptr_t 411 unlock_lockmgr(struct lock_object *lock) 412 { 413 414 panic("lockmgr locks do not support sleep interlocking"); 415 } 416 417 #ifdef KDTRACE_HOOKS 418 static int 419 owner_lockmgr(const struct lock_object *lock, struct thread **owner) 420 { 421 422 panic("lockmgr locks do not support owner inquiring"); 423 } 424 #endif 425 426 void 427 lockinit(struct lock *lk, int pri, const char *wmesg, int timo, int flags) 428 { 429 int iflags; 430 431 MPASS((flags & ~LK_INIT_MASK) == 0); 432 ASSERT_ATOMIC_LOAD_PTR(lk->lk_lock, 433 ("%s: lockmgr not aligned for %s: %p", __func__, wmesg, 434 &lk->lk_lock)); 435 436 iflags = LO_SLEEPABLE | LO_UPGRADABLE; 437 if (flags & LK_CANRECURSE) 438 iflags |= LO_RECURSABLE; 439 if ((flags & LK_NODUP) == 0) 440 iflags |= LO_DUPOK; 441 if (flags & LK_NOPROFILE) 442 iflags |= LO_NOPROFILE; 443 if ((flags & LK_NOWITNESS) == 0) 444 iflags |= LO_WITNESS; 445 if (flags & LK_QUIET) 446 iflags |= LO_QUIET; 447 if (flags & LK_IS_VNODE) 448 iflags |= LO_IS_VNODE; 449 if (flags & LK_NEW) 450 iflags |= LO_NEW; 451 iflags |= flags & LK_NOSHARE; 452 453 lock_init(&lk->lock_object, &lock_class_lockmgr, wmesg, NULL, iflags); 454 lk->lk_lock = LK_UNLOCKED; 455 lk->lk_recurse = 0; 456 lk->lk_exslpfail = 0; 457 lk->lk_timo = timo; 458 lk->lk_pri = pri; 459 STACK_ZERO(lk); 460 } 461 462 /* 463 * XXX: Gross hacks to manipulate external lock flags after 464 * initialization. Used for certain vnode and buf locks. 465 */ 466 void 467 lockallowshare(struct lock *lk) 468 { 469 470 lockmgr_assert(lk, KA_XLOCKED); 471 lk->lock_object.lo_flags &= ~LK_NOSHARE; 472 } 473 474 void 475 lockdisableshare(struct lock *lk) 476 { 477 478 lockmgr_assert(lk, KA_XLOCKED); 479 lk->lock_object.lo_flags |= LK_NOSHARE; 480 } 481 482 void 483 lockallowrecurse(struct lock *lk) 484 { 485 486 lockmgr_assert(lk, KA_XLOCKED); 487 lk->lock_object.lo_flags |= LO_RECURSABLE; 488 } 489 490 void 491 lockdisablerecurse(struct lock *lk) 492 { 493 494 lockmgr_assert(lk, KA_XLOCKED); 495 lk->lock_object.lo_flags &= ~LO_RECURSABLE; 496 } 497 498 void 499 lockdestroy(struct lock *lk) 500 { 501 502 KASSERT(lk->lk_lock == LK_UNLOCKED, ("lockmgr still held")); 503 KASSERT(lk->lk_recurse == 0, ("lockmgr still recursed")); 504 KASSERT(lk->lk_exslpfail == 0, ("lockmgr still exclusive waiters")); 505 lock_destroy(&lk->lock_object); 506 } 507 508 static __always_inline bool 509 lockmgr_slock_try(struct lock *lk, uintptr_t *xp, int flags, bool fp) 510 { 511 512 /* 513 * If no other thread has an exclusive lock, or 514 * no exclusive waiter is present, bump the count of 515 * sharers. Since we have to preserve the state of 516 * waiters, if we fail to acquire the shared lock 517 * loop back and retry. 518 */ 519 while (LK_CAN_SHARE(*xp, flags, fp)) { 520 if (atomic_fcmpset_acq_ptr(&lk->lk_lock, xp, 521 *xp + LK_ONE_SHARER)) { 522 return (true); 523 } 524 } 525 return (false); 526 } 527 528 static __always_inline bool 529 lockmgr_sunlock_try(struct lock *lk, uintptr_t *xp) 530 { 531 532 for (;;) { 533 if (LK_SHARERS(*xp) > 1 || !(*xp & LK_ALL_WAITERS)) { 534 if (atomic_fcmpset_rel_ptr(&lk->lk_lock, xp, 535 *xp - LK_ONE_SHARER)) 536 return (true); 537 continue; 538 } 539 break; 540 } 541 return (false); 542 } 543 544 static bool 545 lockmgr_slock_adaptive(struct lock_delay_arg *lda, struct lock *lk, uintptr_t *xp, 546 int flags) 547 { 548 struct thread *owner; 549 uintptr_t x; 550 551 x = *xp; 552 MPASS(x != LK_UNLOCKED); 553 owner = (struct thread *)LK_HOLDER(x); 554 for (;;) { 555 MPASS(owner != curthread); 556 if (owner == (struct thread *)LK_KERNPROC) 557 return (false); 558 if ((x & LK_SHARE) && LK_SHARERS(x) > 0) 559 return (false); 560 if (owner == NULL) 561 return (false); 562 if (!TD_IS_RUNNING(owner)) 563 return (false); 564 if ((x & LK_ALL_WAITERS) != 0) 565 return (false); 566 lock_delay(lda); 567 x = lockmgr_read_value(lk); 568 if (LK_CAN_SHARE(x, flags, false)) { 569 *xp = x; 570 return (true); 571 } 572 owner = (struct thread *)LK_HOLDER(x); 573 } 574 } 575 576 static __noinline int 577 lockmgr_slock_hard(struct lock *lk, u_int flags, struct lock_object *ilk, 578 const char *file, int line, struct lockmgr_wait *lwa) 579 { 580 uintptr_t tid, x; 581 int error = 0; 582 const char *iwmesg; 583 int ipri, itimo; 584 585 #ifdef KDTRACE_HOOKS 586 uint64_t sleep_time = 0; 587 #endif 588 #ifdef LOCK_PROFILING 589 uint64_t waittime = 0; 590 int contested = 0; 591 #endif 592 struct lock_delay_arg lda; 593 594 if (SCHEDULER_STOPPED()) 595 goto out; 596 597 tid = (uintptr_t)curthread; 598 599 if (LK_CAN_WITNESS(flags)) 600 WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER, 601 file, line, flags & LK_INTERLOCK ? ilk : NULL); 602 x = lockmgr_read_value(lk); 603 lock_delay_arg_init(&lda, &lockmgr_delay); 604 if (!lk_adaptive) 605 flags &= ~LK_ADAPTIVE; 606 /* 607 * The lock may already be locked exclusive by curthread, 608 * avoid deadlock. 609 */ 610 if (LK_HOLDER(x) == tid) { 611 LOCK_LOG2(lk, 612 "%s: %p already held in exclusive mode", 613 __func__, lk); 614 error = EDEADLK; 615 goto out; 616 } 617 618 for (;;) { 619 if (lockmgr_slock_try(lk, &x, flags, false)) 620 break; 621 622 lock_profile_obtain_lock_failed(&lk->lock_object, false, 623 &contested, &waittime); 624 625 if ((flags & (LK_ADAPTIVE | LK_INTERLOCK)) == LK_ADAPTIVE) { 626 if (lockmgr_slock_adaptive(&lda, lk, &x, flags)) 627 continue; 628 } 629 630 #ifdef HWPMC_HOOKS 631 PMC_SOFT_CALL( , , lock, failed); 632 #endif 633 634 /* 635 * If the lock is expected to not sleep just give up 636 * and return. 637 */ 638 if (LK_TRYOP(flags)) { 639 LOCK_LOG2(lk, "%s: %p fails the try operation", 640 __func__, lk); 641 error = EBUSY; 642 break; 643 } 644 645 /* 646 * Acquire the sleepqueue chain lock because we 647 * probabilly will need to manipulate waiters flags. 648 */ 649 sleepq_lock(&lk->lock_object); 650 x = lockmgr_read_value(lk); 651 retry_sleepq: 652 653 /* 654 * if the lock can be acquired in shared mode, try 655 * again. 656 */ 657 if (LK_CAN_SHARE(x, flags, false)) { 658 sleepq_release(&lk->lock_object); 659 continue; 660 } 661 662 /* 663 * Try to set the LK_SHARED_WAITERS flag. If we fail, 664 * loop back and retry. 665 */ 666 if ((x & LK_SHARED_WAITERS) == 0) { 667 if (!atomic_fcmpset_acq_ptr(&lk->lk_lock, &x, 668 x | LK_SHARED_WAITERS)) { 669 goto retry_sleepq; 670 } 671 LOCK_LOG2(lk, "%s: %p set shared waiters flag", 672 __func__, lk); 673 } 674 675 if (lwa == NULL) { 676 iwmesg = lk->lock_object.lo_name; 677 ipri = lk->lk_pri; 678 itimo = lk->lk_timo; 679 } else { 680 iwmesg = lwa->iwmesg; 681 ipri = lwa->ipri; 682 itimo = lwa->itimo; 683 } 684 685 /* 686 * As far as we have been unable to acquire the 687 * shared lock and the shared waiters flag is set, 688 * we will sleep. 689 */ 690 #ifdef KDTRACE_HOOKS 691 sleep_time -= lockstat_nsecs(&lk->lock_object); 692 #endif 693 error = sleeplk(lk, flags, ilk, iwmesg, ipri, itimo, 694 SQ_SHARED_QUEUE); 695 #ifdef KDTRACE_HOOKS 696 sleep_time += lockstat_nsecs(&lk->lock_object); 697 #endif 698 flags &= ~LK_INTERLOCK; 699 if (error) { 700 LOCK_LOG3(lk, 701 "%s: interrupted sleep for %p with %d", 702 __func__, lk, error); 703 break; 704 } 705 LOCK_LOG2(lk, "%s: %p resuming from the sleep queue", 706 __func__, lk); 707 x = lockmgr_read_value(lk); 708 } 709 if (error == 0) { 710 #ifdef KDTRACE_HOOKS 711 if (sleep_time != 0) 712 LOCKSTAT_RECORD4(lockmgr__block, lk, sleep_time, 713 LOCKSTAT_READER, (x & LK_SHARE) == 0, 714 (x & LK_SHARE) == 0 ? 0 : LK_SHARERS(x)); 715 #endif 716 #ifdef LOCK_PROFILING 717 lockmgr_note_shared_acquire(lk, contested, waittime, 718 file, line, flags); 719 #else 720 lockmgr_note_shared_acquire(lk, 0, 0, file, line, 721 flags); 722 #endif 723 } 724 725 out: 726 lockmgr_exit(flags, ilk); 727 return (error); 728 } 729 730 static bool 731 lockmgr_xlock_adaptive(struct lock_delay_arg *lda, struct lock *lk, uintptr_t *xp) 732 { 733 struct thread *owner; 734 uintptr_t x; 735 736 x = *xp; 737 MPASS(x != LK_UNLOCKED); 738 owner = (struct thread *)LK_HOLDER(x); 739 for (;;) { 740 MPASS(owner != curthread); 741 if (owner == NULL) 742 return (false); 743 if ((x & LK_SHARE) && LK_SHARERS(x) > 0) 744 return (false); 745 if (owner == (struct thread *)LK_KERNPROC) 746 return (false); 747 if (!TD_IS_RUNNING(owner)) 748 return (false); 749 if ((x & LK_ALL_WAITERS) != 0) 750 return (false); 751 lock_delay(lda); 752 x = lockmgr_read_value(lk); 753 if (x == LK_UNLOCKED) { 754 *xp = x; 755 return (true); 756 } 757 owner = (struct thread *)LK_HOLDER(x); 758 } 759 } 760 761 static __noinline int 762 lockmgr_xlock_hard(struct lock *lk, u_int flags, struct lock_object *ilk, 763 const char *file, int line, struct lockmgr_wait *lwa) 764 { 765 struct lock_class *class; 766 uintptr_t tid, x, v; 767 int error = 0; 768 const char *iwmesg; 769 int ipri, itimo; 770 771 #ifdef KDTRACE_HOOKS 772 uint64_t sleep_time = 0; 773 #endif 774 #ifdef LOCK_PROFILING 775 uint64_t waittime = 0; 776 int contested = 0; 777 #endif 778 struct lock_delay_arg lda; 779 780 if (SCHEDULER_STOPPED()) 781 goto out; 782 783 tid = (uintptr_t)curthread; 784 785 if (LK_CAN_WITNESS(flags)) 786 WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER | 787 LOP_EXCLUSIVE, file, line, flags & LK_INTERLOCK ? 788 ilk : NULL); 789 790 /* 791 * If curthread already holds the lock and this one is 792 * allowed to recurse, simply recurse on it. 793 */ 794 if (lockmgr_xlocked(lk)) { 795 if ((flags & LK_CANRECURSE) == 0 && 796 (lk->lock_object.lo_flags & LO_RECURSABLE) == 0) { 797 /* 798 * If the lock is expected to not panic just 799 * give up and return. 800 */ 801 if (LK_TRYOP(flags)) { 802 LOCK_LOG2(lk, 803 "%s: %p fails the try operation", 804 __func__, lk); 805 error = EBUSY; 806 goto out; 807 } 808 if (flags & LK_INTERLOCK) { 809 class = LOCK_CLASS(ilk); 810 class->lc_unlock(ilk); 811 } 812 STACK_PRINT(lk); 813 panic("%s: recursing on non recursive lockmgr %p " 814 "@ %s:%d\n", __func__, lk, file, line); 815 } 816 atomic_set_ptr(&lk->lk_lock, LK_WRITER_RECURSED); 817 lk->lk_recurse++; 818 LOCK_LOG2(lk, "%s: %p recursing", __func__, lk); 819 LOCK_LOG_LOCK("XLOCK", &lk->lock_object, 0, 820 lk->lk_recurse, file, line); 821 WITNESS_LOCK(&lk->lock_object, LOP_EXCLUSIVE | 822 LK_TRYWIT(flags), file, line); 823 TD_LOCKS_INC(curthread); 824 goto out; 825 } 826 827 x = LK_UNLOCKED; 828 lock_delay_arg_init(&lda, &lockmgr_delay); 829 if (!lk_adaptive) 830 flags &= ~LK_ADAPTIVE; 831 for (;;) { 832 if (x == LK_UNLOCKED) { 833 if (atomic_fcmpset_acq_ptr(&lk->lk_lock, &x, tid)) 834 break; 835 continue; 836 } 837 838 lock_profile_obtain_lock_failed(&lk->lock_object, false, 839 &contested, &waittime); 840 841 if ((flags & (LK_ADAPTIVE | LK_INTERLOCK)) == LK_ADAPTIVE) { 842 if (lockmgr_xlock_adaptive(&lda, lk, &x)) 843 continue; 844 } 845 #ifdef HWPMC_HOOKS 846 PMC_SOFT_CALL( , , lock, failed); 847 #endif 848 849 /* 850 * If the lock is expected to not sleep just give up 851 * and return. 852 */ 853 if (LK_TRYOP(flags)) { 854 LOCK_LOG2(lk, "%s: %p fails the try operation", 855 __func__, lk); 856 error = EBUSY; 857 break; 858 } 859 860 /* 861 * Acquire the sleepqueue chain lock because we 862 * probabilly will need to manipulate waiters flags. 863 */ 864 sleepq_lock(&lk->lock_object); 865 x = lockmgr_read_value(lk); 866 retry_sleepq: 867 868 /* 869 * if the lock has been released while we spun on 870 * the sleepqueue chain lock just try again. 871 */ 872 if (x == LK_UNLOCKED) { 873 sleepq_release(&lk->lock_object); 874 continue; 875 } 876 877 /* 878 * The lock can be in the state where there is a 879 * pending queue of waiters, but still no owner. 880 * This happens when the lock is contested and an 881 * owner is going to claim the lock. 882 * If curthread is the one successfully acquiring it 883 * claim lock ownership and return, preserving waiters 884 * flags. 885 */ 886 v = x & (LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS); 887 if ((x & ~v) == LK_UNLOCKED) { 888 v &= ~LK_EXCLUSIVE_SPINNERS; 889 if (atomic_fcmpset_acq_ptr(&lk->lk_lock, &x, 890 tid | v)) { 891 sleepq_release(&lk->lock_object); 892 LOCK_LOG2(lk, 893 "%s: %p claimed by a new writer", 894 __func__, lk); 895 break; 896 } 897 goto retry_sleepq; 898 } 899 900 /* 901 * Try to set the LK_EXCLUSIVE_WAITERS flag. If we 902 * fail, loop back and retry. 903 */ 904 if ((x & LK_EXCLUSIVE_WAITERS) == 0) { 905 if (!atomic_fcmpset_ptr(&lk->lk_lock, &x, 906 x | LK_EXCLUSIVE_WAITERS)) { 907 goto retry_sleepq; 908 } 909 LOCK_LOG2(lk, "%s: %p set excl waiters flag", 910 __func__, lk); 911 } 912 913 if (lwa == NULL) { 914 iwmesg = lk->lock_object.lo_name; 915 ipri = lk->lk_pri; 916 itimo = lk->lk_timo; 917 } else { 918 iwmesg = lwa->iwmesg; 919 ipri = lwa->ipri; 920 itimo = lwa->itimo; 921 } 922 923 /* 924 * As far as we have been unable to acquire the 925 * exclusive lock and the exclusive waiters flag 926 * is set, we will sleep. 927 */ 928 #ifdef KDTRACE_HOOKS 929 sleep_time -= lockstat_nsecs(&lk->lock_object); 930 #endif 931 error = sleeplk(lk, flags, ilk, iwmesg, ipri, itimo, 932 SQ_EXCLUSIVE_QUEUE); 933 #ifdef KDTRACE_HOOKS 934 sleep_time += lockstat_nsecs(&lk->lock_object); 935 #endif 936 flags &= ~LK_INTERLOCK; 937 if (error) { 938 LOCK_LOG3(lk, 939 "%s: interrupted sleep for %p with %d", 940 __func__, lk, error); 941 break; 942 } 943 LOCK_LOG2(lk, "%s: %p resuming from the sleep queue", 944 __func__, lk); 945 x = lockmgr_read_value(lk); 946 } 947 if (error == 0) { 948 #ifdef KDTRACE_HOOKS 949 if (sleep_time != 0) 950 LOCKSTAT_RECORD4(lockmgr__block, lk, sleep_time, 951 LOCKSTAT_WRITER, (x & LK_SHARE) == 0, 952 (x & LK_SHARE) == 0 ? 0 : LK_SHARERS(x)); 953 #endif 954 #ifdef LOCK_PROFILING 955 lockmgr_note_exclusive_acquire(lk, contested, waittime, 956 file, line, flags); 957 #else 958 lockmgr_note_exclusive_acquire(lk, 0, 0, file, line, 959 flags); 960 #endif 961 } 962 963 out: 964 lockmgr_exit(flags, ilk); 965 return (error); 966 } 967 968 static __noinline int 969 lockmgr_upgrade(struct lock *lk, u_int flags, struct lock_object *ilk, 970 const char *file, int line, struct lockmgr_wait *lwa) 971 { 972 uintptr_t tid, v, setv; 973 int error = 0; 974 int op; 975 976 if (SCHEDULER_STOPPED()) 977 goto out; 978 979 tid = (uintptr_t)curthread; 980 981 _lockmgr_assert(lk, KA_SLOCKED, file, line); 982 983 op = flags & LK_TYPE_MASK; 984 v = lockmgr_read_value(lk); 985 for (;;) { 986 if (LK_SHARERS(v) > 1) { 987 if (op == LK_TRYUPGRADE) { 988 LOCK_LOG2(lk, "%s: %p failed the nowait upgrade", 989 __func__, lk); 990 error = EBUSY; 991 goto out; 992 } 993 if (atomic_fcmpset_rel_ptr(&lk->lk_lock, &v, 994 v - LK_ONE_SHARER)) { 995 lockmgr_note_shared_release(lk, file, line); 996 goto out_xlock; 997 } 998 continue; 999 } 1000 MPASS((v & ~LK_ALL_WAITERS) == LK_SHARERS_LOCK(1)); 1001 1002 setv = tid; 1003 setv |= (v & LK_ALL_WAITERS); 1004 1005 /* 1006 * Try to switch from one shared lock to an exclusive one. 1007 * We need to preserve waiters flags during the operation. 1008 */ 1009 if (atomic_fcmpset_ptr(&lk->lk_lock, &v, setv)) { 1010 LOCK_LOG_LOCK("XUPGRADE", &lk->lock_object, 0, 0, file, 1011 line); 1012 WITNESS_UPGRADE(&lk->lock_object, LOP_EXCLUSIVE | 1013 LK_TRYWIT(flags), file, line); 1014 LOCKSTAT_RECORD0(lockmgr__upgrade, lk); 1015 TD_SLOCKS_DEC(curthread); 1016 goto out; 1017 } 1018 } 1019 1020 out_xlock: 1021 error = lockmgr_xlock_hard(lk, flags, ilk, file, line, lwa); 1022 flags &= ~LK_INTERLOCK; 1023 out: 1024 lockmgr_exit(flags, ilk); 1025 return (error); 1026 } 1027 1028 int 1029 lockmgr_lock_flags(struct lock *lk, u_int flags, struct lock_object *ilk, 1030 const char *file, int line) 1031 { 1032 struct lock_class *class; 1033 uintptr_t x, tid; 1034 u_int op; 1035 bool locked; 1036 1037 if (SCHEDULER_STOPPED()) 1038 return (0); 1039 1040 op = flags & LK_TYPE_MASK; 1041 locked = false; 1042 switch (op) { 1043 case LK_SHARED: 1044 if (LK_CAN_WITNESS(flags)) 1045 WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER, 1046 file, line, flags & LK_INTERLOCK ? ilk : NULL); 1047 if (__predict_false(lk->lock_object.lo_flags & LK_NOSHARE)) 1048 break; 1049 x = lockmgr_read_value(lk); 1050 if (lockmgr_slock_try(lk, &x, flags, true)) { 1051 lockmgr_note_shared_acquire(lk, 0, 0, 1052 file, line, flags); 1053 locked = true; 1054 } else { 1055 return (lockmgr_slock_hard(lk, flags, ilk, file, line, 1056 NULL)); 1057 } 1058 break; 1059 case LK_EXCLUSIVE: 1060 if (LK_CAN_WITNESS(flags)) 1061 WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER | 1062 LOP_EXCLUSIVE, file, line, flags & LK_INTERLOCK ? 1063 ilk : NULL); 1064 tid = (uintptr_t)curthread; 1065 if (lockmgr_read_value(lk) == LK_UNLOCKED && 1066 atomic_cmpset_acq_ptr(&lk->lk_lock, LK_UNLOCKED, tid)) { 1067 lockmgr_note_exclusive_acquire(lk, 0, 0, file, line, 1068 flags); 1069 locked = true; 1070 } else { 1071 return (lockmgr_xlock_hard(lk, flags, ilk, file, line, 1072 NULL)); 1073 } 1074 break; 1075 case LK_UPGRADE: 1076 case LK_TRYUPGRADE: 1077 return (lockmgr_upgrade(lk, flags, ilk, file, line, NULL)); 1078 default: 1079 break; 1080 } 1081 if (__predict_true(locked)) { 1082 if (__predict_false(flags & LK_INTERLOCK)) { 1083 class = LOCK_CLASS(ilk); 1084 class->lc_unlock(ilk); 1085 } 1086 return (0); 1087 } else { 1088 return (__lockmgr_args(lk, flags, ilk, LK_WMESG_DEFAULT, 1089 LK_PRIO_DEFAULT, LK_TIMO_DEFAULT, file, line)); 1090 } 1091 } 1092 1093 static __noinline int 1094 lockmgr_sunlock_hard(struct lock *lk, uintptr_t x, u_int flags, struct lock_object *ilk, 1095 const char *file, int line) 1096 { 1097 if (!SCHEDULER_STOPPED()) 1098 wakeupshlk(lk, file, line); 1099 lockmgr_exit(flags, ilk); 1100 return (0); 1101 } 1102 1103 static __noinline int 1104 lockmgr_xunlock_hard(struct lock *lk, uintptr_t x, u_int flags, struct lock_object *ilk, 1105 const char *file, int line) 1106 { 1107 uintptr_t tid, v; 1108 u_int realexslp; 1109 int queue; 1110 1111 if (SCHEDULER_STOPPED()) 1112 goto out; 1113 1114 tid = (uintptr_t)curthread; 1115 1116 /* 1117 * As first option, treact the lock as if it has not 1118 * any waiter. 1119 * Fix-up the tid var if the lock has been disowned. 1120 */ 1121 if (lockmgr_disowned_v(x)) 1122 tid = LK_KERNPROC; 1123 1124 /* 1125 * The lock is held in exclusive mode. 1126 * If the lock is recursed also, then unrecurse it. 1127 */ 1128 if (lockmgr_recursed_v(x)) { 1129 LOCK_LOG2(lk, "%s: %p unrecursing", __func__, lk); 1130 lk->lk_recurse--; 1131 if (lk->lk_recurse == 0) 1132 atomic_clear_ptr(&lk->lk_lock, LK_WRITER_RECURSED); 1133 goto out; 1134 } 1135 if (tid != LK_KERNPROC) 1136 LOCKSTAT_PROFILE_RELEASE_RWLOCK(lockmgr__release, lk, 1137 LOCKSTAT_WRITER); 1138 1139 if (x == tid && atomic_cmpset_rel_ptr(&lk->lk_lock, tid, LK_UNLOCKED)) 1140 goto out; 1141 1142 sleepq_lock(&lk->lock_object); 1143 x = lockmgr_read_value(lk); 1144 v = LK_UNLOCKED; 1145 1146 /* 1147 * If the lock has exclusive waiters, give them 1148 * preference in order to avoid deadlock with 1149 * shared runners up. 1150 * If interruptible sleeps left the exclusive queue 1151 * empty avoid a starvation for the threads sleeping 1152 * on the shared queue by giving them precedence 1153 * and cleaning up the exclusive waiters bit anyway. 1154 * Please note that lk_exslpfail count may be lying 1155 * about the real number of waiters with the 1156 * LK_SLEEPFAIL flag on because they may be used in 1157 * conjunction with interruptible sleeps so 1158 * lk_exslpfail might be considered an 'upper limit' 1159 * bound, including the edge cases. 1160 */ 1161 MPASS((x & LK_EXCLUSIVE_SPINNERS) == 0); 1162 realexslp = sleepq_sleepcnt(&lk->lock_object, SQ_EXCLUSIVE_QUEUE); 1163 if ((x & LK_EXCLUSIVE_WAITERS) != 0 && realexslp != 0) { 1164 if (lk->lk_exslpfail != USHRT_MAX && lk->lk_exslpfail < realexslp) { 1165 lk->lk_exslpfail = 0; 1166 queue = SQ_EXCLUSIVE_QUEUE; 1167 v |= (x & LK_SHARED_WAITERS); 1168 } else { 1169 lk->lk_exslpfail = 0; 1170 LOCK_LOG2(lk, 1171 "%s: %p has only LK_SLEEPFAIL sleepers", 1172 __func__, lk); 1173 LOCK_LOG2(lk, 1174 "%s: %p waking up threads on the exclusive queue", 1175 __func__, lk); 1176 sleepq_broadcast(&lk->lock_object, SLEEPQ_LK, 0, 1177 SQ_EXCLUSIVE_QUEUE); 1178 queue = SQ_SHARED_QUEUE; 1179 } 1180 } else { 1181 /* 1182 * Exclusive waiters sleeping with LK_SLEEPFAIL 1183 * on and using interruptible sleeps/timeout 1184 * may have left spourious lk_exslpfail counts 1185 * on, so clean it up anyway. 1186 */ 1187 lk->lk_exslpfail = 0; 1188 queue = SQ_SHARED_QUEUE; 1189 } 1190 1191 LOCK_LOG3(lk, "%s: %p waking up threads on the %s queue", 1192 __func__, lk, queue == SQ_SHARED_QUEUE ? "shared" : 1193 "exclusive"); 1194 atomic_store_rel_ptr(&lk->lk_lock, v); 1195 sleepq_broadcast(&lk->lock_object, SLEEPQ_LK, 0, queue); 1196 sleepq_release(&lk->lock_object); 1197 1198 out: 1199 lockmgr_exit(flags, ilk); 1200 return (0); 1201 } 1202 1203 /* 1204 * Lightweight entry points for common operations. 1205 * 1206 * Functionality is similar to sx locks, in that none of the additional lockmgr 1207 * features are supported. To be clear, these are NOT supported: 1208 * 1. shared locking disablement 1209 * 2. returning with an error after sleep 1210 * 3. unlocking the interlock 1211 * 1212 * If in doubt, use lockmgr_lock_flags. 1213 */ 1214 int 1215 lockmgr_slock(struct lock *lk, u_int flags, const char *file, int line) 1216 { 1217 uintptr_t x; 1218 1219 MPASS((flags & LK_TYPE_MASK) == LK_SHARED); 1220 MPASS((flags & LK_INTERLOCK) == 0); 1221 MPASS((lk->lock_object.lo_flags & LK_NOSHARE) == 0); 1222 1223 if (LK_CAN_WITNESS(flags)) 1224 WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER, 1225 file, line, NULL); 1226 x = lockmgr_read_value(lk); 1227 if (__predict_true(lockmgr_slock_try(lk, &x, flags, true))) { 1228 lockmgr_note_shared_acquire(lk, 0, 0, file, line, flags); 1229 return (0); 1230 } 1231 1232 return (lockmgr_slock_hard(lk, flags | LK_ADAPTIVE, NULL, file, line, NULL)); 1233 } 1234 1235 int 1236 lockmgr_xlock(struct lock *lk, u_int flags, const char *file, int line) 1237 { 1238 uintptr_t tid; 1239 1240 MPASS((flags & LK_TYPE_MASK) == LK_EXCLUSIVE); 1241 MPASS((flags & LK_INTERLOCK) == 0); 1242 1243 if (LK_CAN_WITNESS(flags)) 1244 WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER | 1245 LOP_EXCLUSIVE, file, line, NULL); 1246 tid = (uintptr_t)curthread; 1247 if (atomic_cmpset_acq_ptr(&lk->lk_lock, LK_UNLOCKED, tid)) { 1248 lockmgr_note_exclusive_acquire(lk, 0, 0, file, line, 1249 flags); 1250 return (0); 1251 } 1252 1253 return (lockmgr_xlock_hard(lk, flags | LK_ADAPTIVE, NULL, file, line, NULL)); 1254 } 1255 1256 int 1257 lockmgr_unlock(struct lock *lk) 1258 { 1259 uintptr_t x, tid; 1260 const char *file; 1261 int line; 1262 1263 file = __FILE__; 1264 line = __LINE__; 1265 1266 _lockmgr_assert(lk, KA_LOCKED, file, line); 1267 x = lockmgr_read_value(lk); 1268 if (__predict_true(x & LK_SHARE) != 0) { 1269 lockmgr_note_shared_release(lk, file, line); 1270 if (lockmgr_sunlock_try(lk, &x)) { 1271 LOCKSTAT_PROFILE_RELEASE_RWLOCK(lockmgr__release, lk, LOCKSTAT_READER); 1272 } else { 1273 return (lockmgr_sunlock_hard(lk, x, LK_RELEASE, NULL, file, line)); 1274 } 1275 } else { 1276 tid = (uintptr_t)curthread; 1277 lockmgr_note_exclusive_release(lk, file, line); 1278 if (x == tid && atomic_cmpset_rel_ptr(&lk->lk_lock, tid, LK_UNLOCKED)) { 1279 LOCKSTAT_PROFILE_RELEASE_RWLOCK(lockmgr__release, lk,LOCKSTAT_WRITER); 1280 } else { 1281 return (lockmgr_xunlock_hard(lk, x, LK_RELEASE, NULL, file, line)); 1282 } 1283 } 1284 return (0); 1285 } 1286 1287 int 1288 __lockmgr_args(struct lock *lk, u_int flags, struct lock_object *ilk, 1289 const char *wmesg, int pri, int timo, const char *file, int line) 1290 { 1291 GIANT_DECLARE; 1292 struct lockmgr_wait lwa; 1293 struct lock_class *class; 1294 const char *iwmesg; 1295 uintptr_t tid, v, x; 1296 u_int op, realexslp; 1297 int error, ipri, itimo, queue; 1298 #ifdef LOCK_PROFILING 1299 uint64_t waittime = 0; 1300 int contested = 0; 1301 #endif 1302 1303 if (SCHEDULER_STOPPED()) 1304 return (0); 1305 1306 error = 0; 1307 tid = (uintptr_t)curthread; 1308 op = (flags & LK_TYPE_MASK); 1309 iwmesg = (wmesg == LK_WMESG_DEFAULT) ? lk->lock_object.lo_name : wmesg; 1310 ipri = (pri == LK_PRIO_DEFAULT) ? lk->lk_pri : pri; 1311 itimo = (timo == LK_TIMO_DEFAULT) ? lk->lk_timo : timo; 1312 1313 lwa.iwmesg = iwmesg; 1314 lwa.ipri = ipri; 1315 lwa.itimo = itimo; 1316 1317 MPASS((flags & ~LK_TOTAL_MASK) == 0); 1318 KASSERT((op & (op - 1)) == 0, 1319 ("%s: Invalid requested operation @ %s:%d", __func__, file, line)); 1320 KASSERT((flags & (LK_NOWAIT | LK_SLEEPFAIL)) == 0 || 1321 (op != LK_DOWNGRADE && op != LK_RELEASE), 1322 ("%s: Invalid flags in regard of the operation desired @ %s:%d", 1323 __func__, file, line)); 1324 KASSERT((flags & LK_INTERLOCK) == 0 || ilk != NULL, 1325 ("%s: LK_INTERLOCK passed without valid interlock @ %s:%d", 1326 __func__, file, line)); 1327 KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread), 1328 ("%s: idle thread %p on lockmgr %s @ %s:%d", __func__, curthread, 1329 lk->lock_object.lo_name, file, line)); 1330 1331 class = (flags & LK_INTERLOCK) ? LOCK_CLASS(ilk) : NULL; 1332 1333 if (lk->lock_object.lo_flags & LK_NOSHARE) { 1334 switch (op) { 1335 case LK_SHARED: 1336 op = LK_EXCLUSIVE; 1337 break; 1338 case LK_UPGRADE: 1339 case LK_TRYUPGRADE: 1340 case LK_DOWNGRADE: 1341 _lockmgr_assert(lk, KA_XLOCKED | KA_NOTRECURSED, 1342 file, line); 1343 if (flags & LK_INTERLOCK) 1344 class->lc_unlock(ilk); 1345 return (0); 1346 } 1347 } 1348 1349 switch (op) { 1350 case LK_SHARED: 1351 return (lockmgr_slock_hard(lk, flags, ilk, file, line, &lwa)); 1352 break; 1353 case LK_UPGRADE: 1354 case LK_TRYUPGRADE: 1355 return (lockmgr_upgrade(lk, flags, ilk, file, line, &lwa)); 1356 break; 1357 case LK_EXCLUSIVE: 1358 return (lockmgr_xlock_hard(lk, flags, ilk, file, line, &lwa)); 1359 break; 1360 case LK_DOWNGRADE: 1361 _lockmgr_assert(lk, KA_XLOCKED, file, line); 1362 WITNESS_DOWNGRADE(&lk->lock_object, 0, file, line); 1363 1364 /* 1365 * Panic if the lock is recursed. 1366 */ 1367 if (lockmgr_xlocked(lk) && lockmgr_recursed(lk)) { 1368 if (flags & LK_INTERLOCK) 1369 class->lc_unlock(ilk); 1370 panic("%s: downgrade a recursed lockmgr %s @ %s:%d\n", 1371 __func__, iwmesg, file, line); 1372 } 1373 TD_SLOCKS_INC(curthread); 1374 1375 /* 1376 * In order to preserve waiters flags, just spin. 1377 */ 1378 for (;;) { 1379 x = lockmgr_read_value(lk); 1380 MPASS((x & LK_EXCLUSIVE_SPINNERS) == 0); 1381 x &= LK_ALL_WAITERS; 1382 if (atomic_cmpset_rel_ptr(&lk->lk_lock, tid | x, 1383 LK_SHARERS_LOCK(1) | x)) 1384 break; 1385 cpu_spinwait(); 1386 } 1387 LOCK_LOG_LOCK("XDOWNGRADE", &lk->lock_object, 0, 0, file, line); 1388 LOCKSTAT_RECORD0(lockmgr__downgrade, lk); 1389 break; 1390 case LK_RELEASE: 1391 _lockmgr_assert(lk, KA_LOCKED, file, line); 1392 x = lockmgr_read_value(lk); 1393 1394 if (__predict_true(x & LK_SHARE) != 0) { 1395 lockmgr_note_shared_release(lk, file, line); 1396 return (lockmgr_sunlock_hard(lk, x, flags, ilk, file, line)); 1397 } else { 1398 lockmgr_note_exclusive_release(lk, file, line); 1399 return (lockmgr_xunlock_hard(lk, x, flags, ilk, file, line)); 1400 } 1401 break; 1402 case LK_DRAIN: 1403 if (LK_CAN_WITNESS(flags)) 1404 WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER | 1405 LOP_EXCLUSIVE, file, line, flags & LK_INTERLOCK ? 1406 ilk : NULL); 1407 1408 /* 1409 * Trying to drain a lock we already own will result in a 1410 * deadlock. 1411 */ 1412 if (lockmgr_xlocked(lk)) { 1413 if (flags & LK_INTERLOCK) 1414 class->lc_unlock(ilk); 1415 panic("%s: draining %s with the lock held @ %s:%d\n", 1416 __func__, iwmesg, file, line); 1417 } 1418 1419 for (;;) { 1420 if (lk->lk_lock == LK_UNLOCKED && 1421 atomic_cmpset_acq_ptr(&lk->lk_lock, LK_UNLOCKED, tid)) 1422 break; 1423 1424 #ifdef HWPMC_HOOKS 1425 PMC_SOFT_CALL( , , lock, failed); 1426 #endif 1427 lock_profile_obtain_lock_failed(&lk->lock_object, false, 1428 &contested, &waittime); 1429 1430 /* 1431 * If the lock is expected to not sleep just give up 1432 * and return. 1433 */ 1434 if (LK_TRYOP(flags)) { 1435 LOCK_LOG2(lk, "%s: %p fails the try operation", 1436 __func__, lk); 1437 error = EBUSY; 1438 break; 1439 } 1440 1441 /* 1442 * Acquire the sleepqueue chain lock because we 1443 * probabilly will need to manipulate waiters flags. 1444 */ 1445 sleepq_lock(&lk->lock_object); 1446 x = lockmgr_read_value(lk); 1447 1448 /* 1449 * if the lock has been released while we spun on 1450 * the sleepqueue chain lock just try again. 1451 */ 1452 if (x == LK_UNLOCKED) { 1453 sleepq_release(&lk->lock_object); 1454 continue; 1455 } 1456 1457 v = x & (LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS); 1458 if ((x & ~v) == LK_UNLOCKED) { 1459 v = (x & ~LK_EXCLUSIVE_SPINNERS); 1460 1461 /* 1462 * If interruptible sleeps left the exclusive 1463 * queue empty avoid a starvation for the 1464 * threads sleeping on the shared queue by 1465 * giving them precedence and cleaning up the 1466 * exclusive waiters bit anyway. 1467 * Please note that lk_exslpfail count may be 1468 * lying about the real number of waiters with 1469 * the LK_SLEEPFAIL flag on because they may 1470 * be used in conjunction with interruptible 1471 * sleeps so lk_exslpfail might be considered 1472 * an 'upper limit' bound, including the edge 1473 * cases. 1474 */ 1475 if (v & LK_EXCLUSIVE_WAITERS) { 1476 queue = SQ_EXCLUSIVE_QUEUE; 1477 v &= ~LK_EXCLUSIVE_WAITERS; 1478 } else { 1479 /* 1480 * Exclusive waiters sleeping with 1481 * LK_SLEEPFAIL on and using 1482 * interruptible sleeps/timeout may 1483 * have left spourious lk_exslpfail 1484 * counts on, so clean it up anyway. 1485 */ 1486 MPASS(v & LK_SHARED_WAITERS); 1487 lk->lk_exslpfail = 0; 1488 queue = SQ_SHARED_QUEUE; 1489 v &= ~LK_SHARED_WAITERS; 1490 } 1491 if (queue == SQ_EXCLUSIVE_QUEUE) { 1492 realexslp = 1493 sleepq_sleepcnt(&lk->lock_object, 1494 SQ_EXCLUSIVE_QUEUE); 1495 if (lk->lk_exslpfail >= realexslp) { 1496 lk->lk_exslpfail = 0; 1497 queue = SQ_SHARED_QUEUE; 1498 v &= ~LK_SHARED_WAITERS; 1499 if (realexslp != 0) { 1500 LOCK_LOG2(lk, 1501 "%s: %p has only LK_SLEEPFAIL sleepers", 1502 __func__, lk); 1503 LOCK_LOG2(lk, 1504 "%s: %p waking up threads on the exclusive queue", 1505 __func__, lk); 1506 sleepq_broadcast( 1507 &lk->lock_object, 1508 SLEEPQ_LK, 0, 1509 SQ_EXCLUSIVE_QUEUE); 1510 } 1511 } else 1512 lk->lk_exslpfail = 0; 1513 } 1514 if (!atomic_cmpset_ptr(&lk->lk_lock, x, v)) { 1515 sleepq_release(&lk->lock_object); 1516 continue; 1517 } 1518 LOCK_LOG3(lk, 1519 "%s: %p waking up all threads on the %s queue", 1520 __func__, lk, queue == SQ_SHARED_QUEUE ? 1521 "shared" : "exclusive"); 1522 sleepq_broadcast(&lk->lock_object, SLEEPQ_LK, 0, 1523 queue); 1524 1525 /* 1526 * If shared waiters have been woken up we need 1527 * to wait for one of them to acquire the lock 1528 * before to set the exclusive waiters in 1529 * order to avoid a deadlock. 1530 */ 1531 if (queue == SQ_SHARED_QUEUE) { 1532 for (v = lk->lk_lock; 1533 (v & LK_SHARE) && !LK_SHARERS(v); 1534 v = lk->lk_lock) 1535 cpu_spinwait(); 1536 } 1537 } 1538 1539 /* 1540 * Try to set the LK_EXCLUSIVE_WAITERS flag. If we 1541 * fail, loop back and retry. 1542 */ 1543 if ((x & LK_EXCLUSIVE_WAITERS) == 0) { 1544 if (!atomic_cmpset_ptr(&lk->lk_lock, x, 1545 x | LK_EXCLUSIVE_WAITERS)) { 1546 sleepq_release(&lk->lock_object); 1547 continue; 1548 } 1549 LOCK_LOG2(lk, "%s: %p set drain waiters flag", 1550 __func__, lk); 1551 } 1552 1553 /* 1554 * As far as we have been unable to acquire the 1555 * exclusive lock and the exclusive waiters flag 1556 * is set, we will sleep. 1557 */ 1558 if (flags & LK_INTERLOCK) { 1559 class->lc_unlock(ilk); 1560 flags &= ~LK_INTERLOCK; 1561 } 1562 GIANT_SAVE(); 1563 sleepq_add(&lk->lock_object, NULL, iwmesg, SLEEPQ_LK, 1564 SQ_EXCLUSIVE_QUEUE); 1565 sleepq_wait(&lk->lock_object, ipri & PRIMASK); 1566 GIANT_RESTORE(); 1567 LOCK_LOG2(lk, "%s: %p resuming from the sleep queue", 1568 __func__, lk); 1569 } 1570 1571 if (error == 0) { 1572 lock_profile_obtain_lock_success(&lk->lock_object, 1573 false, contested, waittime, file, line); 1574 LOCK_LOG_LOCK("DRAIN", &lk->lock_object, 0, 1575 lk->lk_recurse, file, line); 1576 WITNESS_LOCK(&lk->lock_object, LOP_EXCLUSIVE | 1577 LK_TRYWIT(flags), file, line); 1578 TD_LOCKS_INC(curthread); 1579 STACK_SAVE(lk); 1580 } 1581 break; 1582 default: 1583 if (flags & LK_INTERLOCK) 1584 class->lc_unlock(ilk); 1585 panic("%s: unknown lockmgr request 0x%x\n", __func__, op); 1586 } 1587 1588 if (flags & LK_INTERLOCK) 1589 class->lc_unlock(ilk); 1590 1591 return (error); 1592 } 1593 1594 void 1595 _lockmgr_disown(struct lock *lk, const char *file, int line) 1596 { 1597 uintptr_t tid, x; 1598 1599 if (SCHEDULER_STOPPED()) 1600 return; 1601 1602 tid = (uintptr_t)curthread; 1603 _lockmgr_assert(lk, KA_XLOCKED, file, line); 1604 1605 /* 1606 * Panic if the lock is recursed. 1607 */ 1608 if (lockmgr_xlocked(lk) && lockmgr_recursed(lk)) 1609 panic("%s: disown a recursed lockmgr @ %s:%d\n", 1610 __func__, file, line); 1611 1612 /* 1613 * If the owner is already LK_KERNPROC just skip the whole operation. 1614 */ 1615 if (LK_HOLDER(lk->lk_lock) != tid) 1616 return; 1617 lock_profile_release_lock(&lk->lock_object, false); 1618 LOCKSTAT_RECORD1(lockmgr__disown, lk, LOCKSTAT_WRITER); 1619 LOCK_LOG_LOCK("XDISOWN", &lk->lock_object, 0, 0, file, line); 1620 WITNESS_UNLOCK(&lk->lock_object, LOP_EXCLUSIVE, file, line); 1621 TD_LOCKS_DEC(curthread); 1622 STACK_SAVE(lk); 1623 1624 /* 1625 * In order to preserve waiters flags, just spin. 1626 */ 1627 for (;;) { 1628 x = lockmgr_read_value(lk); 1629 MPASS((x & LK_EXCLUSIVE_SPINNERS) == 0); 1630 x &= LK_ALL_WAITERS; 1631 if (atomic_cmpset_rel_ptr(&lk->lk_lock, tid | x, 1632 LK_KERNPROC | x)) 1633 return; 1634 cpu_spinwait(); 1635 } 1636 } 1637 1638 void 1639 lockmgr_printinfo(const struct lock *lk) 1640 { 1641 struct thread *td; 1642 uintptr_t x; 1643 1644 if (lk->lk_lock == LK_UNLOCKED) 1645 printf("lock type %s: UNLOCKED\n", lk->lock_object.lo_name); 1646 else if (lk->lk_lock & LK_SHARE) 1647 printf("lock type %s: SHARED (count %ju)\n", 1648 lk->lock_object.lo_name, 1649 (uintmax_t)LK_SHARERS(lk->lk_lock)); 1650 else { 1651 td = lockmgr_xholder(lk); 1652 if (td == (struct thread *)LK_KERNPROC) 1653 printf("lock type %s: EXCL by KERNPROC\n", 1654 lk->lock_object.lo_name); 1655 else 1656 printf("lock type %s: EXCL by thread %p " 1657 "(pid %d, %s, tid %d)\n", lk->lock_object.lo_name, 1658 td, td->td_proc->p_pid, td->td_proc->p_comm, 1659 td->td_tid); 1660 } 1661 1662 x = lk->lk_lock; 1663 if (x & LK_EXCLUSIVE_WAITERS) 1664 printf(" with exclusive waiters pending\n"); 1665 if (x & LK_SHARED_WAITERS) 1666 printf(" with shared waiters pending\n"); 1667 if (x & LK_EXCLUSIVE_SPINNERS) 1668 printf(" with exclusive spinners pending\n"); 1669 1670 STACK_PRINT(lk); 1671 } 1672 1673 int 1674 lockstatus(const struct lock *lk) 1675 { 1676 uintptr_t v, x; 1677 int ret; 1678 1679 ret = LK_SHARED; 1680 x = lockmgr_read_value(lk); 1681 v = LK_HOLDER(x); 1682 1683 if ((x & LK_SHARE) == 0) { 1684 if (v == (uintptr_t)curthread || v == LK_KERNPROC) 1685 ret = LK_EXCLUSIVE; 1686 else 1687 ret = LK_EXCLOTHER; 1688 } else if (x == LK_UNLOCKED) 1689 ret = 0; 1690 1691 return (ret); 1692 } 1693 1694 #ifdef INVARIANT_SUPPORT 1695 1696 FEATURE(invariant_support, 1697 "Support for modules compiled with INVARIANTS option"); 1698 1699 #ifndef INVARIANTS 1700 #undef _lockmgr_assert 1701 #endif 1702 1703 void 1704 _lockmgr_assert(const struct lock *lk, int what, const char *file, int line) 1705 { 1706 int slocked = 0; 1707 1708 if (SCHEDULER_STOPPED()) 1709 return; 1710 switch (what) { 1711 case KA_SLOCKED: 1712 case KA_SLOCKED | KA_NOTRECURSED: 1713 case KA_SLOCKED | KA_RECURSED: 1714 slocked = 1; 1715 case KA_LOCKED: 1716 case KA_LOCKED | KA_NOTRECURSED: 1717 case KA_LOCKED | KA_RECURSED: 1718 #ifdef WITNESS 1719 1720 /* 1721 * We cannot trust WITNESS if the lock is held in exclusive 1722 * mode and a call to lockmgr_disown() happened. 1723 * Workaround this skipping the check if the lock is held in 1724 * exclusive mode even for the KA_LOCKED case. 1725 */ 1726 if (slocked || (lk->lk_lock & LK_SHARE)) { 1727 witness_assert(&lk->lock_object, what, file, line); 1728 break; 1729 } 1730 #endif 1731 if (lk->lk_lock == LK_UNLOCKED || 1732 ((lk->lk_lock & LK_SHARE) == 0 && (slocked || 1733 (!lockmgr_xlocked(lk) && !lockmgr_disowned(lk))))) 1734 panic("Lock %s not %slocked @ %s:%d\n", 1735 lk->lock_object.lo_name, slocked ? "share" : "", 1736 file, line); 1737 1738 if ((lk->lk_lock & LK_SHARE) == 0) { 1739 if (lockmgr_recursed(lk)) { 1740 if (what & KA_NOTRECURSED) 1741 panic("Lock %s recursed @ %s:%d\n", 1742 lk->lock_object.lo_name, file, 1743 line); 1744 } else if (what & KA_RECURSED) 1745 panic("Lock %s not recursed @ %s:%d\n", 1746 lk->lock_object.lo_name, file, line); 1747 } 1748 break; 1749 case KA_XLOCKED: 1750 case KA_XLOCKED | KA_NOTRECURSED: 1751 case KA_XLOCKED | KA_RECURSED: 1752 if (!lockmgr_xlocked(lk) && !lockmgr_disowned(lk)) 1753 panic("Lock %s not exclusively locked @ %s:%d\n", 1754 lk->lock_object.lo_name, file, line); 1755 if (lockmgr_recursed(lk)) { 1756 if (what & KA_NOTRECURSED) 1757 panic("Lock %s recursed @ %s:%d\n", 1758 lk->lock_object.lo_name, file, line); 1759 } else if (what & KA_RECURSED) 1760 panic("Lock %s not recursed @ %s:%d\n", 1761 lk->lock_object.lo_name, file, line); 1762 break; 1763 case KA_UNLOCKED: 1764 if (lockmgr_xlocked(lk) || lockmgr_disowned(lk)) 1765 panic("Lock %s exclusively locked @ %s:%d\n", 1766 lk->lock_object.lo_name, file, line); 1767 break; 1768 default: 1769 panic("Unknown lockmgr assertion: %d @ %s:%d\n", what, file, 1770 line); 1771 } 1772 } 1773 #endif 1774 1775 #ifdef DDB 1776 int 1777 lockmgr_chain(struct thread *td, struct thread **ownerp) 1778 { 1779 const struct lock *lk; 1780 1781 lk = td->td_wchan; 1782 1783 if (LOCK_CLASS(&lk->lock_object) != &lock_class_lockmgr) 1784 return (0); 1785 db_printf("blocked on lockmgr %s", lk->lock_object.lo_name); 1786 if (lk->lk_lock & LK_SHARE) 1787 db_printf("SHARED (count %ju)\n", 1788 (uintmax_t)LK_SHARERS(lk->lk_lock)); 1789 else 1790 db_printf("EXCL\n"); 1791 *ownerp = lockmgr_xholder(lk); 1792 1793 return (1); 1794 } 1795 1796 static void 1797 db_show_lockmgr(const struct lock_object *lock) 1798 { 1799 struct thread *td; 1800 const struct lock *lk; 1801 1802 lk = (const struct lock *)lock; 1803 1804 db_printf(" state: "); 1805 if (lk->lk_lock == LK_UNLOCKED) 1806 db_printf("UNLOCKED\n"); 1807 else if (lk->lk_lock & LK_SHARE) 1808 db_printf("SLOCK: %ju\n", (uintmax_t)LK_SHARERS(lk->lk_lock)); 1809 else { 1810 td = lockmgr_xholder(lk); 1811 if (td == (struct thread *)LK_KERNPROC) 1812 db_printf("XLOCK: LK_KERNPROC\n"); 1813 else 1814 db_printf("XLOCK: %p (tid %d, pid %d, \"%s\")\n", td, 1815 td->td_tid, td->td_proc->p_pid, 1816 td->td_proc->p_comm); 1817 if (lockmgr_recursed(lk)) 1818 db_printf(" recursed: %d\n", lk->lk_recurse); 1819 } 1820 db_printf(" waiters: "); 1821 switch (lk->lk_lock & LK_ALL_WAITERS) { 1822 case LK_SHARED_WAITERS: 1823 db_printf("shared\n"); 1824 break; 1825 case LK_EXCLUSIVE_WAITERS: 1826 db_printf("exclusive\n"); 1827 break; 1828 case LK_ALL_WAITERS: 1829 db_printf("shared and exclusive\n"); 1830 break; 1831 default: 1832 db_printf("none\n"); 1833 } 1834 db_printf(" spinners: "); 1835 if (lk->lk_lock & LK_EXCLUSIVE_SPINNERS) 1836 db_printf("exclusive\n"); 1837 else 1838 db_printf("none\n"); 1839 } 1840 #endif 1841