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/cdefs.h> 35 #include <sys/param.h> 36 #include <sys/kdb.h> 37 #include <sys/ktr.h> 38 #include <sys/limits.h> 39 #include <sys/lock.h> 40 #include <sys/lock_profile.h> 41 #include <sys/lockmgr.h> 42 #include <sys/lockstat.h> 43 #include <sys/mutex.h> 44 #include <sys/proc.h> 45 #include <sys/sleepqueue.h> 46 #ifdef DEBUG_LOCKS 47 #include <sys/stack.h> 48 #endif 49 #include <sys/sysctl.h> 50 #include <sys/systm.h> 51 52 #include <machine/cpu.h> 53 54 #ifdef DDB 55 #include <ddb/ddb.h> 56 #endif 57 58 #ifdef HWPMC_HOOKS 59 #include <sys/pmckern.h> 60 PMC_SOFT_DECLARE( , , lock, failed); 61 #endif 62 63 /* 64 * Hack. There should be prio_t or similar so that this is not necessary. 65 */ 66 _Static_assert((PRILASTFLAG * 2) - 1 <= USHRT_MAX, 67 "prio flags wont fit in u_short pri in struct lock"); 68 69 CTASSERT(LK_UNLOCKED == (LK_UNLOCKED & 70 ~(LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS))); 71 72 #define SQ_EXCLUSIVE_QUEUE 0 73 #define SQ_SHARED_QUEUE 1 74 75 #ifndef INVARIANTS 76 #define _lockmgr_assert(lk, what, file, line) 77 #endif 78 79 #define TD_SLOCKS_INC(td) ((td)->td_lk_slocks++) 80 #define TD_SLOCKS_DEC(td) ((td)->td_lk_slocks--) 81 82 #ifndef DEBUG_LOCKS 83 #define STACK_PRINT(lk) 84 #define STACK_SAVE(lk) 85 #define STACK_ZERO(lk) 86 #else 87 #define STACK_PRINT(lk) stack_print_ddb(&(lk)->lk_stack) 88 #define STACK_SAVE(lk) stack_save(&(lk)->lk_stack) 89 #define STACK_ZERO(lk) stack_zero(&(lk)->lk_stack) 90 #endif 91 92 #define LOCK_LOG2(lk, string, arg1, arg2) \ 93 if (LOCK_LOG_TEST(&(lk)->lock_object, 0)) \ 94 CTR2(KTR_LOCK, (string), (arg1), (arg2)) 95 #define LOCK_LOG3(lk, string, arg1, arg2, arg3) \ 96 if (LOCK_LOG_TEST(&(lk)->lock_object, 0)) \ 97 CTR3(KTR_LOCK, (string), (arg1), (arg2), (arg3)) 98 99 #define GIANT_DECLARE \ 100 int _i = 0; \ 101 WITNESS_SAVE_DECL(Giant) 102 #define GIANT_RESTORE() do { \ 103 if (__predict_false(_i > 0)) { \ 104 while (_i--) \ 105 mtx_lock(&Giant); \ 106 WITNESS_RESTORE(&Giant.lock_object, Giant); \ 107 } \ 108 } while (0) 109 #define GIANT_SAVE() do { \ 110 if (__predict_false(mtx_owned(&Giant))) { \ 111 WITNESS_SAVE(&Giant.lock_object, Giant); \ 112 while (mtx_owned(&Giant)) { \ 113 _i++; \ 114 mtx_unlock(&Giant); \ 115 } \ 116 } \ 117 } while (0) 118 119 static bool __always_inline 120 LK_CAN_SHARE(uintptr_t x, int flags, bool fp) 121 { 122 123 if ((x & (LK_SHARE | LK_EXCLUSIVE_WAITERS | LK_EXCLUSIVE_SPINNERS)) == 124 LK_SHARE) 125 return (true); 126 if (fp || (!(x & LK_SHARE))) 127 return (false); 128 if ((curthread->td_lk_slocks != 0 && !(flags & LK_NODDLKTREAT)) || 129 (curthread->td_pflags & TDP_DEADLKTREAT)) 130 return (true); 131 return (false); 132 } 133 134 #define LK_TRYOP(x) \ 135 ((x) & LK_NOWAIT) 136 137 #define LK_CAN_WITNESS(x) \ 138 (((x) & LK_NOWITNESS) == 0 && !LK_TRYOP(x)) 139 #define LK_TRYWIT(x) \ 140 (LK_TRYOP(x) ? LOP_TRYLOCK : 0) 141 142 #define lockmgr_disowned(lk) \ 143 (((lk)->lk_lock & ~(LK_FLAGMASK & ~LK_SHARE)) == LK_KERNPROC) 144 145 #define lockmgr_xlocked_v(v) \ 146 (((v) & ~(LK_FLAGMASK & ~LK_SHARE)) == (uintptr_t)curthread) 147 148 #define lockmgr_xlocked(lk) lockmgr_xlocked_v(lockmgr_read_value(lk)) 149 150 static void assert_lockmgr(const struct lock_object *lock, int how); 151 #ifdef DDB 152 static void db_show_lockmgr(const struct lock_object *lock); 153 #endif 154 static void lock_lockmgr(struct lock_object *lock, uintptr_t how); 155 #ifdef KDTRACE_HOOKS 156 static int owner_lockmgr(const struct lock_object *lock, 157 struct thread **owner); 158 #endif 159 static uintptr_t unlock_lockmgr(struct lock_object *lock); 160 161 struct lock_class lock_class_lockmgr = { 162 .lc_name = "lockmgr", 163 .lc_flags = LC_RECURSABLE | LC_SLEEPABLE | LC_SLEEPLOCK | LC_UPGRADABLE, 164 .lc_assert = assert_lockmgr, 165 #ifdef DDB 166 .lc_ddb_show = db_show_lockmgr, 167 #endif 168 .lc_lock = lock_lockmgr, 169 .lc_unlock = unlock_lockmgr, 170 #ifdef KDTRACE_HOOKS 171 .lc_owner = owner_lockmgr, 172 #endif 173 }; 174 175 static __read_mostly bool lk_adaptive = true; 176 static SYSCTL_NODE(_debug, OID_AUTO, lockmgr, CTLFLAG_RD, NULL, "lockmgr debugging"); 177 SYSCTL_BOOL(_debug_lockmgr, OID_AUTO, adaptive_spinning, CTLFLAG_RW, &lk_adaptive, 178 0, ""); 179 #define lockmgr_delay locks_delay 180 181 struct lockmgr_wait { 182 const char *iwmesg; 183 int ipri; 184 int itimo; 185 }; 186 187 static bool __always_inline lockmgr_slock_try(struct lock *lk, uintptr_t *xp, 188 int flags, bool fp); 189 static bool __always_inline lockmgr_sunlock_try(struct lock *lk, uintptr_t *xp); 190 191 static void 192 lockmgr_exit(u_int flags, struct lock_object *ilk, int wakeup_swapper) 193 { 194 struct lock_class *class; 195 196 if (flags & LK_INTERLOCK) { 197 class = LOCK_CLASS(ilk); 198 class->lc_unlock(ilk); 199 } 200 201 if (__predict_false(wakeup_swapper)) 202 kick_proc0(); 203 } 204 205 static void 206 lockmgr_note_shared_acquire(struct lock *lk, int contested, 207 uint64_t waittime, const char *file, int line, int flags) 208 { 209 210 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(lockmgr__acquire, lk, contested, 211 waittime, file, line, LOCKSTAT_READER); 212 LOCK_LOG_LOCK("SLOCK", &lk->lock_object, 0, 0, file, line); 213 WITNESS_LOCK(&lk->lock_object, LK_TRYWIT(flags), file, line); 214 TD_LOCKS_INC(curthread); 215 TD_SLOCKS_INC(curthread); 216 STACK_SAVE(lk); 217 } 218 219 static void 220 lockmgr_note_shared_release(struct lock *lk, const char *file, int line) 221 { 222 223 WITNESS_UNLOCK(&lk->lock_object, 0, file, line); 224 LOCK_LOG_LOCK("SUNLOCK", &lk->lock_object, 0, 0, file, line); 225 TD_LOCKS_DEC(curthread); 226 TD_SLOCKS_DEC(curthread); 227 } 228 229 static void 230 lockmgr_note_exclusive_acquire(struct lock *lk, int contested, 231 uint64_t waittime, const char *file, int line, int flags) 232 { 233 234 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(lockmgr__acquire, lk, contested, 235 waittime, file, line, LOCKSTAT_WRITER); 236 LOCK_LOG_LOCK("XLOCK", &lk->lock_object, 0, lk->lk_recurse, file, line); 237 WITNESS_LOCK(&lk->lock_object, LOP_EXCLUSIVE | LK_TRYWIT(flags), file, 238 line); 239 TD_LOCKS_INC(curthread); 240 STACK_SAVE(lk); 241 } 242 243 static void 244 lockmgr_note_exclusive_release(struct lock *lk, const char *file, int line) 245 { 246 247 if (LK_HOLDER(lockmgr_read_value(lk)) != LK_KERNPROC) { 248 WITNESS_UNLOCK(&lk->lock_object, LOP_EXCLUSIVE, file, line); 249 TD_LOCKS_DEC(curthread); 250 } 251 LOCK_LOG_LOCK("XUNLOCK", &lk->lock_object, 0, lk->lk_recurse, file, 252 line); 253 } 254 255 static __inline struct thread * 256 lockmgr_xholder(const struct lock *lk) 257 { 258 uintptr_t x; 259 260 x = lockmgr_read_value(lk); 261 return ((x & LK_SHARE) ? NULL : (struct thread *)LK_HOLDER(x)); 262 } 263 264 /* 265 * It assumes sleepq_lock held and returns with this one unheld. 266 * It also assumes the generic interlock is sane and previously checked. 267 * If LK_INTERLOCK is specified the interlock is not reacquired after the 268 * sleep. 269 */ 270 static __inline int 271 sleeplk(struct lock *lk, u_int flags, struct lock_object *ilk, 272 const char *wmesg, int pri, int timo, int queue) 273 { 274 GIANT_DECLARE; 275 struct lock_class *class; 276 int catch, error; 277 278 class = (flags & LK_INTERLOCK) ? LOCK_CLASS(ilk) : NULL; 279 catch = pri & PCATCH; 280 pri &= PRIMASK; 281 error = 0; 282 283 LOCK_LOG3(lk, "%s: %p blocking on the %s sleepqueue", __func__, lk, 284 (queue == SQ_EXCLUSIVE_QUEUE) ? "exclusive" : "shared"); 285 286 if (flags & LK_INTERLOCK) 287 class->lc_unlock(ilk); 288 if (queue == SQ_EXCLUSIVE_QUEUE && (flags & LK_SLEEPFAIL) != 0) { 289 if (lk->lk_exslpfail < USHRT_MAX) 290 lk->lk_exslpfail++; 291 } 292 GIANT_SAVE(); 293 sleepq_add(&lk->lock_object, NULL, wmesg, SLEEPQ_LK | (catch ? 294 SLEEPQ_INTERRUPTIBLE : 0), queue); 295 if ((flags & LK_TIMELOCK) && timo) 296 sleepq_set_timeout(&lk->lock_object, timo); 297 298 /* 299 * Decisional switch for real sleeping. 300 */ 301 if ((flags & LK_TIMELOCK) && timo && catch) 302 error = sleepq_timedwait_sig(&lk->lock_object, pri); 303 else if ((flags & LK_TIMELOCK) && timo) 304 error = sleepq_timedwait(&lk->lock_object, pri); 305 else if (catch) 306 error = sleepq_wait_sig(&lk->lock_object, pri); 307 else 308 sleepq_wait(&lk->lock_object, pri); 309 GIANT_RESTORE(); 310 if ((flags & LK_SLEEPFAIL) && error == 0) 311 error = ENOLCK; 312 313 return (error); 314 } 315 316 static __inline int 317 wakeupshlk(struct lock *lk, const char *file, int line) 318 { 319 uintptr_t v, x, orig_x; 320 u_int realexslp; 321 int queue, wakeup_swapper; 322 323 wakeup_swapper = 0; 324 for (;;) { 325 x = lockmgr_read_value(lk); 326 if (lockmgr_sunlock_try(lk, &x)) 327 break; 328 329 /* 330 * We should have a sharer with waiters, so enter the hard 331 * path in order to handle wakeups correctly. 332 */ 333 sleepq_lock(&lk->lock_object); 334 orig_x = lockmgr_read_value(lk); 335 retry_sleepq: 336 x = orig_x & (LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS); 337 v = LK_UNLOCKED; 338 339 /* 340 * If the lock has exclusive waiters, give them preference in 341 * order to avoid deadlock with shared runners up. 342 * If interruptible sleeps left the exclusive queue empty 343 * avoid a starvation for the threads sleeping on the shared 344 * queue by giving them precedence and cleaning up the 345 * exclusive waiters bit anyway. 346 * Please note that lk_exslpfail count may be lying about 347 * the real number of waiters with the LK_SLEEPFAIL flag on 348 * because they may be used in conjunction with interruptible 349 * sleeps so lk_exslpfail might be considered an 'upper limit' 350 * bound, including the edge cases. 351 */ 352 realexslp = sleepq_sleepcnt(&lk->lock_object, 353 SQ_EXCLUSIVE_QUEUE); 354 if ((x & LK_EXCLUSIVE_WAITERS) != 0 && realexslp != 0) { 355 if (lk->lk_exslpfail != USHRT_MAX && lk->lk_exslpfail < realexslp) { 356 lk->lk_exslpfail = 0; 357 queue = SQ_EXCLUSIVE_QUEUE; 358 v |= (x & LK_SHARED_WAITERS); 359 } else { 360 lk->lk_exslpfail = 0; 361 LOCK_LOG2(lk, 362 "%s: %p has only LK_SLEEPFAIL sleepers", 363 __func__, lk); 364 LOCK_LOG2(lk, 365 "%s: %p waking up threads on the exclusive queue", 366 __func__, lk); 367 wakeup_swapper = 368 sleepq_broadcast(&lk->lock_object, 369 SLEEPQ_LK, 0, SQ_EXCLUSIVE_QUEUE); 370 queue = SQ_SHARED_QUEUE; 371 } 372 } else { 373 /* 374 * Exclusive waiters sleeping with LK_SLEEPFAIL on 375 * and using interruptible sleeps/timeout may have 376 * left spourious lk_exslpfail counts on, so clean 377 * it up anyway. 378 */ 379 lk->lk_exslpfail = 0; 380 queue = SQ_SHARED_QUEUE; 381 } 382 383 if (lockmgr_sunlock_try(lk, &orig_x)) { 384 sleepq_release(&lk->lock_object); 385 break; 386 } 387 388 x |= LK_SHARERS_LOCK(1); 389 if (!atomic_fcmpset_rel_ptr(&lk->lk_lock, &x, v)) { 390 orig_x = x; 391 goto retry_sleepq; 392 } 393 LOCK_LOG3(lk, "%s: %p waking up threads on the %s queue", 394 __func__, lk, queue == SQ_SHARED_QUEUE ? "shared" : 395 "exclusive"); 396 wakeup_swapper |= sleepq_broadcast(&lk->lock_object, SLEEPQ_LK, 397 0, queue); 398 sleepq_release(&lk->lock_object); 399 break; 400 } 401 402 LOCKSTAT_PROFILE_RELEASE_RWLOCK(lockmgr__release, lk, LOCKSTAT_READER); 403 return (wakeup_swapper); 404 } 405 406 static void 407 assert_lockmgr(const struct lock_object *lock, int what) 408 { 409 410 panic("lockmgr locks do not support assertions"); 411 } 412 413 static void 414 lock_lockmgr(struct lock_object *lock, uintptr_t how) 415 { 416 417 panic("lockmgr locks do not support sleep interlocking"); 418 } 419 420 static uintptr_t 421 unlock_lockmgr(struct lock_object *lock) 422 { 423 424 panic("lockmgr locks do not support sleep interlocking"); 425 } 426 427 #ifdef KDTRACE_HOOKS 428 static int 429 owner_lockmgr(const struct lock_object *lock, struct thread **owner) 430 { 431 432 panic("lockmgr locks do not support owner inquiring"); 433 } 434 #endif 435 436 void 437 lockinit(struct lock *lk, int pri, const char *wmesg, int timo, int flags) 438 { 439 int iflags; 440 441 MPASS((flags & ~LK_INIT_MASK) == 0); 442 ASSERT_ATOMIC_LOAD_PTR(lk->lk_lock, 443 ("%s: lockmgr not aligned for %s: %p", __func__, wmesg, 444 &lk->lk_lock)); 445 446 iflags = LO_SLEEPABLE | LO_UPGRADABLE; 447 if (flags & LK_CANRECURSE) 448 iflags |= LO_RECURSABLE; 449 if ((flags & LK_NODUP) == 0) 450 iflags |= LO_DUPOK; 451 if (flags & LK_NOPROFILE) 452 iflags |= LO_NOPROFILE; 453 if ((flags & LK_NOWITNESS) == 0) 454 iflags |= LO_WITNESS; 455 if (flags & LK_QUIET) 456 iflags |= LO_QUIET; 457 if (flags & LK_IS_VNODE) 458 iflags |= LO_IS_VNODE; 459 if (flags & LK_NEW) 460 iflags |= LO_NEW; 461 iflags |= flags & LK_NOSHARE; 462 463 lock_init(&lk->lock_object, &lock_class_lockmgr, wmesg, NULL, iflags); 464 lk->lk_lock = LK_UNLOCKED; 465 lk->lk_recurse = 0; 466 lk->lk_exslpfail = 0; 467 lk->lk_timo = timo; 468 lk->lk_pri = pri; 469 STACK_ZERO(lk); 470 } 471 472 /* 473 * XXX: Gross hacks to manipulate external lock flags after 474 * initialization. Used for certain vnode and buf locks. 475 */ 476 void 477 lockallowshare(struct lock *lk) 478 { 479 480 lockmgr_assert(lk, KA_XLOCKED); 481 lk->lock_object.lo_flags &= ~LK_NOSHARE; 482 } 483 484 void 485 lockdisableshare(struct lock *lk) 486 { 487 488 lockmgr_assert(lk, KA_XLOCKED); 489 lk->lock_object.lo_flags |= LK_NOSHARE; 490 } 491 492 void 493 lockallowrecurse(struct lock *lk) 494 { 495 496 lockmgr_assert(lk, KA_XLOCKED); 497 lk->lock_object.lo_flags |= LO_RECURSABLE; 498 } 499 500 void 501 lockdisablerecurse(struct lock *lk) 502 { 503 504 lockmgr_assert(lk, KA_XLOCKED); 505 lk->lock_object.lo_flags &= ~LO_RECURSABLE; 506 } 507 508 void 509 lockdestroy(struct lock *lk) 510 { 511 512 KASSERT(lk->lk_lock == LK_UNLOCKED, ("lockmgr still held")); 513 KASSERT(lk->lk_recurse == 0, ("lockmgr still recursed")); 514 KASSERT(lk->lk_exslpfail == 0, ("lockmgr still exclusive waiters")); 515 lock_destroy(&lk->lock_object); 516 } 517 518 static bool __always_inline 519 lockmgr_slock_try(struct lock *lk, uintptr_t *xp, int flags, bool fp) 520 { 521 522 /* 523 * If no other thread has an exclusive lock, or 524 * no exclusive waiter is present, bump the count of 525 * sharers. Since we have to preserve the state of 526 * waiters, if we fail to acquire the shared lock 527 * loop back and retry. 528 */ 529 while (LK_CAN_SHARE(*xp, flags, fp)) { 530 if (atomic_fcmpset_acq_ptr(&lk->lk_lock, xp, 531 *xp + LK_ONE_SHARER)) { 532 return (true); 533 } 534 } 535 return (false); 536 } 537 538 static bool __always_inline 539 lockmgr_sunlock_try(struct lock *lk, uintptr_t *xp) 540 { 541 542 for (;;) { 543 if (LK_SHARERS(*xp) > 1 || !(*xp & LK_ALL_WAITERS)) { 544 if (atomic_fcmpset_rel_ptr(&lk->lk_lock, xp, 545 *xp - LK_ONE_SHARER)) 546 return (true); 547 continue; 548 } 549 break; 550 } 551 return (false); 552 } 553 554 static bool 555 lockmgr_slock_adaptive(struct lock_delay_arg *lda, struct lock *lk, uintptr_t *xp, 556 int flags) 557 { 558 struct thread *owner; 559 uintptr_t x; 560 561 x = *xp; 562 MPASS(x != LK_UNLOCKED); 563 owner = (struct thread *)LK_HOLDER(x); 564 for (;;) { 565 MPASS(owner != curthread); 566 if (owner == (struct thread *)LK_KERNPROC) 567 return (false); 568 if ((x & LK_SHARE) && LK_SHARERS(x) > 0) 569 return (false); 570 if (owner == NULL) 571 return (false); 572 if (!TD_IS_RUNNING(owner)) 573 return (false); 574 if ((x & LK_ALL_WAITERS) != 0) 575 return (false); 576 lock_delay(lda); 577 x = lockmgr_read_value(lk); 578 if (LK_CAN_SHARE(x, flags, false)) { 579 *xp = x; 580 return (true); 581 } 582 owner = (struct thread *)LK_HOLDER(x); 583 } 584 } 585 586 static __noinline int 587 lockmgr_slock_hard(struct lock *lk, u_int flags, struct lock_object *ilk, 588 const char *file, int line, struct lockmgr_wait *lwa) 589 { 590 uintptr_t tid, x; 591 int error = 0; 592 const char *iwmesg; 593 int ipri, itimo; 594 595 #ifdef KDTRACE_HOOKS 596 uint64_t sleep_time = 0; 597 #endif 598 #ifdef LOCK_PROFILING 599 uint64_t waittime = 0; 600 int contested = 0; 601 #endif 602 struct lock_delay_arg lda; 603 604 if (SCHEDULER_STOPPED()) 605 goto out; 606 607 tid = (uintptr_t)curthread; 608 609 if (LK_CAN_WITNESS(flags)) 610 WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER, 611 file, line, flags & LK_INTERLOCK ? ilk : NULL); 612 x = lockmgr_read_value(lk); 613 lock_delay_arg_init(&lda, &lockmgr_delay); 614 if (!lk_adaptive) 615 flags &= ~LK_ADAPTIVE; 616 /* 617 * The lock may already be locked exclusive by curthread, 618 * avoid deadlock. 619 */ 620 if (LK_HOLDER(x) == tid) { 621 LOCK_LOG2(lk, 622 "%s: %p already held in exclusive mode", 623 __func__, lk); 624 error = EDEADLK; 625 goto out; 626 } 627 628 for (;;) { 629 if (lockmgr_slock_try(lk, &x, flags, false)) 630 break; 631 632 lock_profile_obtain_lock_failed(&lk->lock_object, false, 633 &contested, &waittime); 634 635 if ((flags & (LK_ADAPTIVE | LK_INTERLOCK)) == LK_ADAPTIVE) { 636 if (lockmgr_slock_adaptive(&lda, lk, &x, flags)) 637 continue; 638 } 639 640 #ifdef HWPMC_HOOKS 641 PMC_SOFT_CALL( , , lock, failed); 642 #endif 643 644 /* 645 * If the lock is expected to not sleep just give up 646 * and return. 647 */ 648 if (LK_TRYOP(flags)) { 649 LOCK_LOG2(lk, "%s: %p fails the try operation", 650 __func__, lk); 651 error = EBUSY; 652 break; 653 } 654 655 /* 656 * Acquire the sleepqueue chain lock because we 657 * probabilly will need to manipulate waiters flags. 658 */ 659 sleepq_lock(&lk->lock_object); 660 x = lockmgr_read_value(lk); 661 retry_sleepq: 662 663 /* 664 * if the lock can be acquired in shared mode, try 665 * again. 666 */ 667 if (LK_CAN_SHARE(x, flags, false)) { 668 sleepq_release(&lk->lock_object); 669 continue; 670 } 671 672 /* 673 * Try to set the LK_SHARED_WAITERS flag. If we fail, 674 * loop back and retry. 675 */ 676 if ((x & LK_SHARED_WAITERS) == 0) { 677 if (!atomic_fcmpset_acq_ptr(&lk->lk_lock, &x, 678 x | LK_SHARED_WAITERS)) { 679 goto retry_sleepq; 680 } 681 LOCK_LOG2(lk, "%s: %p set shared waiters flag", 682 __func__, lk); 683 } 684 685 if (lwa == NULL) { 686 iwmesg = lk->lock_object.lo_name; 687 ipri = lk->lk_pri; 688 itimo = lk->lk_timo; 689 } else { 690 iwmesg = lwa->iwmesg; 691 ipri = lwa->ipri; 692 itimo = lwa->itimo; 693 } 694 695 /* 696 * As far as we have been unable to acquire the 697 * shared lock and the shared waiters flag is set, 698 * we will sleep. 699 */ 700 #ifdef KDTRACE_HOOKS 701 sleep_time -= lockstat_nsecs(&lk->lock_object); 702 #endif 703 error = sleeplk(lk, flags, ilk, iwmesg, ipri, itimo, 704 SQ_SHARED_QUEUE); 705 #ifdef KDTRACE_HOOKS 706 sleep_time += lockstat_nsecs(&lk->lock_object); 707 #endif 708 flags &= ~LK_INTERLOCK; 709 if (error) { 710 LOCK_LOG3(lk, 711 "%s: interrupted sleep for %p with %d", 712 __func__, lk, error); 713 break; 714 } 715 LOCK_LOG2(lk, "%s: %p resuming from the sleep queue", 716 __func__, lk); 717 x = lockmgr_read_value(lk); 718 } 719 if (error == 0) { 720 #ifdef KDTRACE_HOOKS 721 if (sleep_time != 0) 722 LOCKSTAT_RECORD4(lockmgr__block, lk, sleep_time, 723 LOCKSTAT_READER, (x & LK_SHARE) == 0, 724 (x & LK_SHARE) == 0 ? 0 : LK_SHARERS(x)); 725 #endif 726 #ifdef LOCK_PROFILING 727 lockmgr_note_shared_acquire(lk, contested, waittime, 728 file, line, flags); 729 #else 730 lockmgr_note_shared_acquire(lk, 0, 0, file, line, 731 flags); 732 #endif 733 } 734 735 out: 736 lockmgr_exit(flags, ilk, 0); 737 return (error); 738 } 739 740 static bool 741 lockmgr_xlock_adaptive(struct lock_delay_arg *lda, struct lock *lk, uintptr_t *xp) 742 { 743 struct thread *owner; 744 uintptr_t x; 745 746 x = *xp; 747 MPASS(x != LK_UNLOCKED); 748 owner = (struct thread *)LK_HOLDER(x); 749 for (;;) { 750 MPASS(owner != curthread); 751 if (owner == NULL) 752 return (false); 753 if ((x & LK_SHARE) && LK_SHARERS(x) > 0) 754 return (false); 755 if (owner == (struct thread *)LK_KERNPROC) 756 return (false); 757 if (!TD_IS_RUNNING(owner)) 758 return (false); 759 if ((x & LK_ALL_WAITERS) != 0) 760 return (false); 761 lock_delay(lda); 762 x = lockmgr_read_value(lk); 763 if (x == LK_UNLOCKED) { 764 *xp = x; 765 return (true); 766 } 767 owner = (struct thread *)LK_HOLDER(x); 768 } 769 } 770 771 static __noinline int 772 lockmgr_xlock_hard(struct lock *lk, u_int flags, struct lock_object *ilk, 773 const char *file, int line, struct lockmgr_wait *lwa) 774 { 775 struct lock_class *class; 776 uintptr_t tid, x, v; 777 int error = 0; 778 const char *iwmesg; 779 int ipri, itimo; 780 781 #ifdef KDTRACE_HOOKS 782 uint64_t sleep_time = 0; 783 #endif 784 #ifdef LOCK_PROFILING 785 uint64_t waittime = 0; 786 int contested = 0; 787 #endif 788 struct lock_delay_arg lda; 789 790 if (SCHEDULER_STOPPED()) 791 goto out; 792 793 tid = (uintptr_t)curthread; 794 795 if (LK_CAN_WITNESS(flags)) 796 WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER | 797 LOP_EXCLUSIVE, file, line, flags & LK_INTERLOCK ? 798 ilk : NULL); 799 800 /* 801 * If curthread already holds the lock and this one is 802 * allowed to recurse, simply recurse on it. 803 */ 804 if (lockmgr_xlocked(lk)) { 805 if ((flags & LK_CANRECURSE) == 0 && 806 (lk->lock_object.lo_flags & LO_RECURSABLE) == 0) { 807 /* 808 * If the lock is expected to not panic just 809 * give up and return. 810 */ 811 if (LK_TRYOP(flags)) { 812 LOCK_LOG2(lk, 813 "%s: %p fails the try operation", 814 __func__, lk); 815 error = EBUSY; 816 goto out; 817 } 818 if (flags & LK_INTERLOCK) { 819 class = LOCK_CLASS(ilk); 820 class->lc_unlock(ilk); 821 } 822 STACK_PRINT(lk); 823 panic("%s: recursing on non recursive lockmgr %p " 824 "@ %s:%d\n", __func__, lk, file, line); 825 } 826 atomic_set_ptr(&lk->lk_lock, LK_WRITER_RECURSED); 827 lk->lk_recurse++; 828 LOCK_LOG2(lk, "%s: %p recursing", __func__, lk); 829 LOCK_LOG_LOCK("XLOCK", &lk->lock_object, 0, 830 lk->lk_recurse, file, line); 831 WITNESS_LOCK(&lk->lock_object, LOP_EXCLUSIVE | 832 LK_TRYWIT(flags), file, line); 833 TD_LOCKS_INC(curthread); 834 goto out; 835 } 836 837 x = LK_UNLOCKED; 838 lock_delay_arg_init(&lda, &lockmgr_delay); 839 if (!lk_adaptive) 840 flags &= ~LK_ADAPTIVE; 841 for (;;) { 842 if (x == LK_UNLOCKED) { 843 if (atomic_fcmpset_acq_ptr(&lk->lk_lock, &x, tid)) 844 break; 845 continue; 846 } 847 848 lock_profile_obtain_lock_failed(&lk->lock_object, false, 849 &contested, &waittime); 850 851 if ((flags & (LK_ADAPTIVE | LK_INTERLOCK)) == LK_ADAPTIVE) { 852 if (lockmgr_xlock_adaptive(&lda, lk, &x)) 853 continue; 854 } 855 #ifdef HWPMC_HOOKS 856 PMC_SOFT_CALL( , , lock, failed); 857 #endif 858 859 /* 860 * If the lock is expected to not sleep just give up 861 * and return. 862 */ 863 if (LK_TRYOP(flags)) { 864 LOCK_LOG2(lk, "%s: %p fails the try operation", 865 __func__, lk); 866 error = EBUSY; 867 break; 868 } 869 870 /* 871 * Acquire the sleepqueue chain lock because we 872 * probabilly will need to manipulate waiters flags. 873 */ 874 sleepq_lock(&lk->lock_object); 875 x = lockmgr_read_value(lk); 876 retry_sleepq: 877 878 /* 879 * if the lock has been released while we spun on 880 * the sleepqueue chain lock just try again. 881 */ 882 if (x == LK_UNLOCKED) { 883 sleepq_release(&lk->lock_object); 884 continue; 885 } 886 887 /* 888 * The lock can be in the state where there is a 889 * pending queue of waiters, but still no owner. 890 * This happens when the lock is contested and an 891 * owner is going to claim the lock. 892 * If curthread is the one successfully acquiring it 893 * claim lock ownership and return, preserving waiters 894 * flags. 895 */ 896 v = x & (LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS); 897 if ((x & ~v) == LK_UNLOCKED) { 898 v &= ~LK_EXCLUSIVE_SPINNERS; 899 if (atomic_fcmpset_acq_ptr(&lk->lk_lock, &x, 900 tid | v)) { 901 sleepq_release(&lk->lock_object); 902 LOCK_LOG2(lk, 903 "%s: %p claimed by a new writer", 904 __func__, lk); 905 break; 906 } 907 goto retry_sleepq; 908 } 909 910 /* 911 * Try to set the LK_EXCLUSIVE_WAITERS flag. If we 912 * fail, loop back and retry. 913 */ 914 if ((x & LK_EXCLUSIVE_WAITERS) == 0) { 915 if (!atomic_fcmpset_ptr(&lk->lk_lock, &x, 916 x | LK_EXCLUSIVE_WAITERS)) { 917 goto retry_sleepq; 918 } 919 LOCK_LOG2(lk, "%s: %p set excl waiters flag", 920 __func__, lk); 921 } 922 923 if (lwa == NULL) { 924 iwmesg = lk->lock_object.lo_name; 925 ipri = lk->lk_pri; 926 itimo = lk->lk_timo; 927 } else { 928 iwmesg = lwa->iwmesg; 929 ipri = lwa->ipri; 930 itimo = lwa->itimo; 931 } 932 933 /* 934 * As far as we have been unable to acquire the 935 * exclusive lock and the exclusive waiters flag 936 * is set, we will sleep. 937 */ 938 #ifdef KDTRACE_HOOKS 939 sleep_time -= lockstat_nsecs(&lk->lock_object); 940 #endif 941 error = sleeplk(lk, flags, ilk, iwmesg, ipri, itimo, 942 SQ_EXCLUSIVE_QUEUE); 943 #ifdef KDTRACE_HOOKS 944 sleep_time += lockstat_nsecs(&lk->lock_object); 945 #endif 946 flags &= ~LK_INTERLOCK; 947 if (error) { 948 LOCK_LOG3(lk, 949 "%s: interrupted sleep for %p with %d", 950 __func__, lk, error); 951 break; 952 } 953 LOCK_LOG2(lk, "%s: %p resuming from the sleep queue", 954 __func__, lk); 955 x = lockmgr_read_value(lk); 956 } 957 if (error == 0) { 958 #ifdef KDTRACE_HOOKS 959 if (sleep_time != 0) 960 LOCKSTAT_RECORD4(lockmgr__block, lk, sleep_time, 961 LOCKSTAT_WRITER, (x & LK_SHARE) == 0, 962 (x & LK_SHARE) == 0 ? 0 : LK_SHARERS(x)); 963 #endif 964 #ifdef LOCK_PROFILING 965 lockmgr_note_exclusive_acquire(lk, contested, waittime, 966 file, line, flags); 967 #else 968 lockmgr_note_exclusive_acquire(lk, 0, 0, file, line, 969 flags); 970 #endif 971 } 972 973 out: 974 lockmgr_exit(flags, ilk, 0); 975 return (error); 976 } 977 978 static __noinline int 979 lockmgr_upgrade(struct lock *lk, u_int flags, struct lock_object *ilk, 980 const char *file, int line, struct lockmgr_wait *lwa) 981 { 982 uintptr_t tid, v, setv; 983 int error = 0; 984 int op; 985 986 if (SCHEDULER_STOPPED()) 987 goto out; 988 989 tid = (uintptr_t)curthread; 990 991 _lockmgr_assert(lk, KA_SLOCKED, file, line); 992 993 op = flags & LK_TYPE_MASK; 994 v = lockmgr_read_value(lk); 995 for (;;) { 996 if (LK_SHARERS(v) > 1) { 997 if (op == LK_TRYUPGRADE) { 998 LOCK_LOG2(lk, "%s: %p failed the nowait upgrade", 999 __func__, lk); 1000 error = EBUSY; 1001 goto out; 1002 } 1003 if (atomic_fcmpset_rel_ptr(&lk->lk_lock, &v, 1004 v - LK_ONE_SHARER)) { 1005 lockmgr_note_shared_release(lk, file, line); 1006 goto out_xlock; 1007 } 1008 continue; 1009 } 1010 MPASS((v & ~LK_ALL_WAITERS) == LK_SHARERS_LOCK(1)); 1011 1012 setv = tid; 1013 setv |= (v & LK_ALL_WAITERS); 1014 1015 /* 1016 * Try to switch from one shared lock to an exclusive one. 1017 * We need to preserve waiters flags during the operation. 1018 */ 1019 if (atomic_fcmpset_ptr(&lk->lk_lock, &v, setv)) { 1020 LOCK_LOG_LOCK("XUPGRADE", &lk->lock_object, 0, 0, file, 1021 line); 1022 WITNESS_UPGRADE(&lk->lock_object, LOP_EXCLUSIVE | 1023 LK_TRYWIT(flags), file, line); 1024 LOCKSTAT_RECORD0(lockmgr__upgrade, lk); 1025 TD_SLOCKS_DEC(curthread); 1026 goto out; 1027 } 1028 } 1029 1030 out_xlock: 1031 error = lockmgr_xlock_hard(lk, flags, ilk, file, line, lwa); 1032 flags &= ~LK_INTERLOCK; 1033 out: 1034 lockmgr_exit(flags, ilk, 0); 1035 return (error); 1036 } 1037 1038 int 1039 lockmgr_lock_flags(struct lock *lk, u_int flags, struct lock_object *ilk, 1040 const char *file, int line) 1041 { 1042 struct lock_class *class; 1043 uintptr_t x, tid; 1044 u_int op; 1045 bool locked; 1046 1047 if (SCHEDULER_STOPPED()) 1048 return (0); 1049 1050 op = flags & LK_TYPE_MASK; 1051 locked = false; 1052 switch (op) { 1053 case LK_SHARED: 1054 if (LK_CAN_WITNESS(flags)) 1055 WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER, 1056 file, line, flags & LK_INTERLOCK ? ilk : NULL); 1057 if (__predict_false(lk->lock_object.lo_flags & LK_NOSHARE)) 1058 break; 1059 x = lockmgr_read_value(lk); 1060 if (lockmgr_slock_try(lk, &x, flags, true)) { 1061 lockmgr_note_shared_acquire(lk, 0, 0, 1062 file, line, flags); 1063 locked = true; 1064 } else { 1065 return (lockmgr_slock_hard(lk, flags, ilk, file, line, 1066 NULL)); 1067 } 1068 break; 1069 case LK_EXCLUSIVE: 1070 if (LK_CAN_WITNESS(flags)) 1071 WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER | 1072 LOP_EXCLUSIVE, file, line, flags & LK_INTERLOCK ? 1073 ilk : NULL); 1074 tid = (uintptr_t)curthread; 1075 if (lockmgr_read_value(lk) == LK_UNLOCKED && 1076 atomic_cmpset_acq_ptr(&lk->lk_lock, LK_UNLOCKED, tid)) { 1077 lockmgr_note_exclusive_acquire(lk, 0, 0, file, line, 1078 flags); 1079 locked = true; 1080 } else { 1081 return (lockmgr_xlock_hard(lk, flags, ilk, file, line, 1082 NULL)); 1083 } 1084 break; 1085 case LK_UPGRADE: 1086 case LK_TRYUPGRADE: 1087 return (lockmgr_upgrade(lk, flags, ilk, file, line, NULL)); 1088 default: 1089 break; 1090 } 1091 if (__predict_true(locked)) { 1092 if (__predict_false(flags & LK_INTERLOCK)) { 1093 class = LOCK_CLASS(ilk); 1094 class->lc_unlock(ilk); 1095 } 1096 return (0); 1097 } else { 1098 return (__lockmgr_args(lk, flags, ilk, LK_WMESG_DEFAULT, 1099 LK_PRIO_DEFAULT, LK_TIMO_DEFAULT, file, line)); 1100 } 1101 } 1102 1103 static __noinline int 1104 lockmgr_sunlock_hard(struct lock *lk, uintptr_t x, u_int flags, struct lock_object *ilk, 1105 const char *file, int line) 1106 1107 { 1108 int wakeup_swapper = 0; 1109 1110 if (SCHEDULER_STOPPED()) 1111 goto out; 1112 1113 wakeup_swapper = wakeupshlk(lk, file, line); 1114 1115 out: 1116 lockmgr_exit(flags, ilk, wakeup_swapper); 1117 return (0); 1118 } 1119 1120 static __noinline int 1121 lockmgr_xunlock_hard(struct lock *lk, uintptr_t x, u_int flags, struct lock_object *ilk, 1122 const char *file, int line) 1123 { 1124 uintptr_t tid, v; 1125 int wakeup_swapper = 0; 1126 u_int realexslp; 1127 int queue; 1128 1129 if (SCHEDULER_STOPPED()) 1130 goto out; 1131 1132 tid = (uintptr_t)curthread; 1133 1134 /* 1135 * As first option, treact the lock as if it has not 1136 * any waiter. 1137 * Fix-up the tid var if the lock has been disowned. 1138 */ 1139 if (LK_HOLDER(x) == LK_KERNPROC) 1140 tid = LK_KERNPROC; 1141 1142 /* 1143 * The lock is held in exclusive mode. 1144 * If the lock is recursed also, then unrecurse it. 1145 */ 1146 if (lockmgr_recursed_v(x)) { 1147 LOCK_LOG2(lk, "%s: %p unrecursing", __func__, lk); 1148 lk->lk_recurse--; 1149 if (lk->lk_recurse == 0) 1150 atomic_clear_ptr(&lk->lk_lock, LK_WRITER_RECURSED); 1151 goto out; 1152 } 1153 if (tid != LK_KERNPROC) 1154 LOCKSTAT_PROFILE_RELEASE_RWLOCK(lockmgr__release, lk, 1155 LOCKSTAT_WRITER); 1156 1157 if (x == tid && atomic_cmpset_rel_ptr(&lk->lk_lock, tid, LK_UNLOCKED)) 1158 goto out; 1159 1160 sleepq_lock(&lk->lock_object); 1161 x = lockmgr_read_value(lk); 1162 v = LK_UNLOCKED; 1163 1164 /* 1165 * If the lock has exclusive waiters, give them 1166 * preference in order to avoid deadlock with 1167 * shared runners up. 1168 * If interruptible sleeps left the exclusive queue 1169 * empty avoid a starvation for the threads sleeping 1170 * on the shared queue by giving them precedence 1171 * and cleaning up the exclusive waiters bit anyway. 1172 * Please note that lk_exslpfail count may be lying 1173 * about the real number of waiters with the 1174 * LK_SLEEPFAIL flag on because they may be used in 1175 * conjunction with interruptible sleeps so 1176 * lk_exslpfail might be considered an 'upper limit' 1177 * bound, including the edge cases. 1178 */ 1179 MPASS((x & LK_EXCLUSIVE_SPINNERS) == 0); 1180 realexslp = sleepq_sleepcnt(&lk->lock_object, SQ_EXCLUSIVE_QUEUE); 1181 if ((x & LK_EXCLUSIVE_WAITERS) != 0 && realexslp != 0) { 1182 if (lk->lk_exslpfail != USHRT_MAX && lk->lk_exslpfail < realexslp) { 1183 lk->lk_exslpfail = 0; 1184 queue = SQ_EXCLUSIVE_QUEUE; 1185 v |= (x & LK_SHARED_WAITERS); 1186 } else { 1187 lk->lk_exslpfail = 0; 1188 LOCK_LOG2(lk, 1189 "%s: %p has only LK_SLEEPFAIL sleepers", 1190 __func__, lk); 1191 LOCK_LOG2(lk, 1192 "%s: %p waking up threads on the exclusive queue", 1193 __func__, lk); 1194 wakeup_swapper = sleepq_broadcast(&lk->lock_object, 1195 SLEEPQ_LK, 0, SQ_EXCLUSIVE_QUEUE); 1196 queue = SQ_SHARED_QUEUE; 1197 } 1198 } else { 1199 /* 1200 * Exclusive waiters sleeping with LK_SLEEPFAIL 1201 * on and using interruptible sleeps/timeout 1202 * may have left spourious lk_exslpfail counts 1203 * on, so clean it up anyway. 1204 */ 1205 lk->lk_exslpfail = 0; 1206 queue = SQ_SHARED_QUEUE; 1207 } 1208 1209 LOCK_LOG3(lk, "%s: %p waking up threads on the %s queue", 1210 __func__, lk, queue == SQ_SHARED_QUEUE ? "shared" : 1211 "exclusive"); 1212 atomic_store_rel_ptr(&lk->lk_lock, v); 1213 wakeup_swapper |= sleepq_broadcast(&lk->lock_object, SLEEPQ_LK, 0, queue); 1214 sleepq_release(&lk->lock_object); 1215 1216 out: 1217 lockmgr_exit(flags, ilk, wakeup_swapper); 1218 return (0); 1219 } 1220 1221 /* 1222 * Lightweight entry points for common operations. 1223 * 1224 * Functionality is similar to sx locks, in that none of the additional lockmgr 1225 * features are supported. To be clear, these are NOT supported: 1226 * 1. shared locking disablement 1227 * 2. returning with an error after sleep 1228 * 3. unlocking the interlock 1229 * 1230 * If in doubt, use lockmgr_lock_flags. 1231 */ 1232 int 1233 lockmgr_slock(struct lock *lk, u_int flags, const char *file, int line) 1234 { 1235 uintptr_t x; 1236 1237 MPASS((flags & LK_TYPE_MASK) == LK_SHARED); 1238 MPASS((flags & LK_INTERLOCK) == 0); 1239 MPASS((lk->lock_object.lo_flags & LK_NOSHARE) == 0); 1240 1241 if (LK_CAN_WITNESS(flags)) 1242 WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER, 1243 file, line, NULL); 1244 x = lockmgr_read_value(lk); 1245 if (__predict_true(lockmgr_slock_try(lk, &x, flags, true))) { 1246 lockmgr_note_shared_acquire(lk, 0, 0, file, line, flags); 1247 return (0); 1248 } 1249 1250 return (lockmgr_slock_hard(lk, flags | LK_ADAPTIVE, NULL, file, line, NULL)); 1251 } 1252 1253 int 1254 lockmgr_xlock(struct lock *lk, u_int flags, const char *file, int line) 1255 { 1256 uintptr_t tid; 1257 1258 MPASS((flags & LK_TYPE_MASK) == LK_EXCLUSIVE); 1259 MPASS((flags & LK_INTERLOCK) == 0); 1260 1261 if (LK_CAN_WITNESS(flags)) 1262 WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER | 1263 LOP_EXCLUSIVE, file, line, NULL); 1264 tid = (uintptr_t)curthread; 1265 if (atomic_cmpset_acq_ptr(&lk->lk_lock, LK_UNLOCKED, tid)) { 1266 lockmgr_note_exclusive_acquire(lk, 0, 0, file, line, 1267 flags); 1268 return (0); 1269 } 1270 1271 return (lockmgr_xlock_hard(lk, flags | LK_ADAPTIVE, NULL, file, line, NULL)); 1272 } 1273 1274 int 1275 lockmgr_unlock(struct lock *lk) 1276 { 1277 uintptr_t x, tid; 1278 const char *file; 1279 int line; 1280 1281 file = __FILE__; 1282 line = __LINE__; 1283 1284 _lockmgr_assert(lk, KA_LOCKED, file, line); 1285 x = lockmgr_read_value(lk); 1286 if (__predict_true(x & LK_SHARE) != 0) { 1287 lockmgr_note_shared_release(lk, file, line); 1288 if (lockmgr_sunlock_try(lk, &x)) { 1289 LOCKSTAT_PROFILE_RELEASE_RWLOCK(lockmgr__release, lk, LOCKSTAT_READER); 1290 } else { 1291 return (lockmgr_sunlock_hard(lk, x, LK_RELEASE, NULL, file, line)); 1292 } 1293 } else { 1294 tid = (uintptr_t)curthread; 1295 lockmgr_note_exclusive_release(lk, file, line); 1296 if (x == tid && atomic_cmpset_rel_ptr(&lk->lk_lock, tid, LK_UNLOCKED)) { 1297 LOCKSTAT_PROFILE_RELEASE_RWLOCK(lockmgr__release, lk,LOCKSTAT_WRITER); 1298 } else { 1299 return (lockmgr_xunlock_hard(lk, x, LK_RELEASE, NULL, file, line)); 1300 } 1301 } 1302 return (0); 1303 } 1304 1305 int 1306 __lockmgr_args(struct lock *lk, u_int flags, struct lock_object *ilk, 1307 const char *wmesg, int pri, int timo, const char *file, int line) 1308 { 1309 GIANT_DECLARE; 1310 struct lockmgr_wait lwa; 1311 struct lock_class *class; 1312 const char *iwmesg; 1313 uintptr_t tid, v, x; 1314 u_int op, realexslp; 1315 int error, ipri, itimo, queue, wakeup_swapper; 1316 #ifdef LOCK_PROFILING 1317 uint64_t waittime = 0; 1318 int contested = 0; 1319 #endif 1320 1321 if (SCHEDULER_STOPPED()) 1322 return (0); 1323 1324 error = 0; 1325 tid = (uintptr_t)curthread; 1326 op = (flags & LK_TYPE_MASK); 1327 iwmesg = (wmesg == LK_WMESG_DEFAULT) ? lk->lock_object.lo_name : wmesg; 1328 ipri = (pri == LK_PRIO_DEFAULT) ? lk->lk_pri : pri; 1329 itimo = (timo == LK_TIMO_DEFAULT) ? lk->lk_timo : timo; 1330 1331 lwa.iwmesg = iwmesg; 1332 lwa.ipri = ipri; 1333 lwa.itimo = itimo; 1334 1335 MPASS((flags & ~LK_TOTAL_MASK) == 0); 1336 KASSERT((op & (op - 1)) == 0, 1337 ("%s: Invalid requested operation @ %s:%d", __func__, file, line)); 1338 KASSERT((flags & (LK_NOWAIT | LK_SLEEPFAIL)) == 0 || 1339 (op != LK_DOWNGRADE && op != LK_RELEASE), 1340 ("%s: Invalid flags in regard of the operation desired @ %s:%d", 1341 __func__, file, line)); 1342 KASSERT((flags & LK_INTERLOCK) == 0 || ilk != NULL, 1343 ("%s: LK_INTERLOCK passed without valid interlock @ %s:%d", 1344 __func__, file, line)); 1345 KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread), 1346 ("%s: idle thread %p on lockmgr %s @ %s:%d", __func__, curthread, 1347 lk->lock_object.lo_name, file, line)); 1348 1349 class = (flags & LK_INTERLOCK) ? LOCK_CLASS(ilk) : NULL; 1350 1351 if (lk->lock_object.lo_flags & LK_NOSHARE) { 1352 switch (op) { 1353 case LK_SHARED: 1354 op = LK_EXCLUSIVE; 1355 break; 1356 case LK_UPGRADE: 1357 case LK_TRYUPGRADE: 1358 case LK_DOWNGRADE: 1359 _lockmgr_assert(lk, KA_XLOCKED | KA_NOTRECURSED, 1360 file, line); 1361 if (flags & LK_INTERLOCK) 1362 class->lc_unlock(ilk); 1363 return (0); 1364 } 1365 } 1366 1367 wakeup_swapper = 0; 1368 switch (op) { 1369 case LK_SHARED: 1370 return (lockmgr_slock_hard(lk, flags, ilk, file, line, &lwa)); 1371 break; 1372 case LK_UPGRADE: 1373 case LK_TRYUPGRADE: 1374 return (lockmgr_upgrade(lk, flags, ilk, file, line, &lwa)); 1375 break; 1376 case LK_EXCLUSIVE: 1377 return (lockmgr_xlock_hard(lk, flags, ilk, file, line, &lwa)); 1378 break; 1379 case LK_DOWNGRADE: 1380 _lockmgr_assert(lk, KA_XLOCKED, file, line); 1381 WITNESS_DOWNGRADE(&lk->lock_object, 0, file, line); 1382 1383 /* 1384 * Panic if the lock is recursed. 1385 */ 1386 if (lockmgr_xlocked(lk) && lockmgr_recursed(lk)) { 1387 if (flags & LK_INTERLOCK) 1388 class->lc_unlock(ilk); 1389 panic("%s: downgrade a recursed lockmgr %s @ %s:%d\n", 1390 __func__, iwmesg, file, line); 1391 } 1392 TD_SLOCKS_INC(curthread); 1393 1394 /* 1395 * In order to preserve waiters flags, just spin. 1396 */ 1397 for (;;) { 1398 x = lockmgr_read_value(lk); 1399 MPASS((x & LK_EXCLUSIVE_SPINNERS) == 0); 1400 x &= LK_ALL_WAITERS; 1401 if (atomic_cmpset_rel_ptr(&lk->lk_lock, tid | x, 1402 LK_SHARERS_LOCK(1) | x)) 1403 break; 1404 cpu_spinwait(); 1405 } 1406 LOCK_LOG_LOCK("XDOWNGRADE", &lk->lock_object, 0, 0, file, line); 1407 LOCKSTAT_RECORD0(lockmgr__downgrade, lk); 1408 break; 1409 case LK_RELEASE: 1410 _lockmgr_assert(lk, KA_LOCKED, file, line); 1411 x = lockmgr_read_value(lk); 1412 1413 if (__predict_true(x & LK_SHARE) != 0) { 1414 lockmgr_note_shared_release(lk, file, line); 1415 return (lockmgr_sunlock_hard(lk, x, flags, ilk, file, line)); 1416 } else { 1417 lockmgr_note_exclusive_release(lk, file, line); 1418 return (lockmgr_xunlock_hard(lk, x, flags, ilk, file, line)); 1419 } 1420 break; 1421 case LK_DRAIN: 1422 if (LK_CAN_WITNESS(flags)) 1423 WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER | 1424 LOP_EXCLUSIVE, file, line, flags & LK_INTERLOCK ? 1425 ilk : NULL); 1426 1427 /* 1428 * Trying to drain a lock we already own will result in a 1429 * deadlock. 1430 */ 1431 if (lockmgr_xlocked(lk)) { 1432 if (flags & LK_INTERLOCK) 1433 class->lc_unlock(ilk); 1434 panic("%s: draining %s with the lock held @ %s:%d\n", 1435 __func__, iwmesg, file, line); 1436 } 1437 1438 for (;;) { 1439 if (lk->lk_lock == LK_UNLOCKED && 1440 atomic_cmpset_acq_ptr(&lk->lk_lock, LK_UNLOCKED, tid)) 1441 break; 1442 1443 #ifdef HWPMC_HOOKS 1444 PMC_SOFT_CALL( , , lock, failed); 1445 #endif 1446 lock_profile_obtain_lock_failed(&lk->lock_object, false, 1447 &contested, &waittime); 1448 1449 /* 1450 * If the lock is expected to not sleep just give up 1451 * and return. 1452 */ 1453 if (LK_TRYOP(flags)) { 1454 LOCK_LOG2(lk, "%s: %p fails the try operation", 1455 __func__, lk); 1456 error = EBUSY; 1457 break; 1458 } 1459 1460 /* 1461 * Acquire the sleepqueue chain lock because we 1462 * probabilly will need to manipulate waiters flags. 1463 */ 1464 sleepq_lock(&lk->lock_object); 1465 x = lockmgr_read_value(lk); 1466 1467 /* 1468 * if the lock has been released while we spun on 1469 * the sleepqueue chain lock just try again. 1470 */ 1471 if (x == LK_UNLOCKED) { 1472 sleepq_release(&lk->lock_object); 1473 continue; 1474 } 1475 1476 v = x & (LK_ALL_WAITERS | LK_EXCLUSIVE_SPINNERS); 1477 if ((x & ~v) == LK_UNLOCKED) { 1478 v = (x & ~LK_EXCLUSIVE_SPINNERS); 1479 1480 /* 1481 * If interruptible sleeps left the exclusive 1482 * queue empty avoid a starvation for the 1483 * threads sleeping on the shared queue by 1484 * giving them precedence and cleaning up the 1485 * exclusive waiters bit anyway. 1486 * Please note that lk_exslpfail count may be 1487 * lying about the real number of waiters with 1488 * the LK_SLEEPFAIL flag on because they may 1489 * be used in conjunction with interruptible 1490 * sleeps so lk_exslpfail might be considered 1491 * an 'upper limit' bound, including the edge 1492 * cases. 1493 */ 1494 if (v & LK_EXCLUSIVE_WAITERS) { 1495 queue = SQ_EXCLUSIVE_QUEUE; 1496 v &= ~LK_EXCLUSIVE_WAITERS; 1497 } else { 1498 /* 1499 * Exclusive waiters sleeping with 1500 * LK_SLEEPFAIL on and using 1501 * interruptible sleeps/timeout may 1502 * have left spourious lk_exslpfail 1503 * counts on, so clean it up anyway. 1504 */ 1505 MPASS(v & LK_SHARED_WAITERS); 1506 lk->lk_exslpfail = 0; 1507 queue = SQ_SHARED_QUEUE; 1508 v &= ~LK_SHARED_WAITERS; 1509 } 1510 if (queue == SQ_EXCLUSIVE_QUEUE) { 1511 realexslp = 1512 sleepq_sleepcnt(&lk->lock_object, 1513 SQ_EXCLUSIVE_QUEUE); 1514 if (lk->lk_exslpfail >= realexslp) { 1515 lk->lk_exslpfail = 0; 1516 queue = SQ_SHARED_QUEUE; 1517 v &= ~LK_SHARED_WAITERS; 1518 if (realexslp != 0) { 1519 LOCK_LOG2(lk, 1520 "%s: %p has only LK_SLEEPFAIL sleepers", 1521 __func__, lk); 1522 LOCK_LOG2(lk, 1523 "%s: %p waking up threads on the exclusive queue", 1524 __func__, lk); 1525 wakeup_swapper = 1526 sleepq_broadcast( 1527 &lk->lock_object, 1528 SLEEPQ_LK, 0, 1529 SQ_EXCLUSIVE_QUEUE); 1530 } 1531 } else 1532 lk->lk_exslpfail = 0; 1533 } 1534 if (!atomic_cmpset_ptr(&lk->lk_lock, x, v)) { 1535 sleepq_release(&lk->lock_object); 1536 continue; 1537 } 1538 LOCK_LOG3(lk, 1539 "%s: %p waking up all threads on the %s queue", 1540 __func__, lk, queue == SQ_SHARED_QUEUE ? 1541 "shared" : "exclusive"); 1542 wakeup_swapper |= sleepq_broadcast( 1543 &lk->lock_object, SLEEPQ_LK, 0, queue); 1544 1545 /* 1546 * If shared waiters have been woken up we need 1547 * to wait for one of them to acquire the lock 1548 * before to set the exclusive waiters in 1549 * order to avoid a deadlock. 1550 */ 1551 if (queue == SQ_SHARED_QUEUE) { 1552 for (v = lk->lk_lock; 1553 (v & LK_SHARE) && !LK_SHARERS(v); 1554 v = lk->lk_lock) 1555 cpu_spinwait(); 1556 } 1557 } 1558 1559 /* 1560 * Try to set the LK_EXCLUSIVE_WAITERS flag. If we 1561 * fail, loop back and retry. 1562 */ 1563 if ((x & LK_EXCLUSIVE_WAITERS) == 0) { 1564 if (!atomic_cmpset_ptr(&lk->lk_lock, x, 1565 x | LK_EXCLUSIVE_WAITERS)) { 1566 sleepq_release(&lk->lock_object); 1567 continue; 1568 } 1569 LOCK_LOG2(lk, "%s: %p set drain waiters flag", 1570 __func__, lk); 1571 } 1572 1573 /* 1574 * As far as we have been unable to acquire the 1575 * exclusive lock and the exclusive waiters flag 1576 * is set, we will sleep. 1577 */ 1578 if (flags & LK_INTERLOCK) { 1579 class->lc_unlock(ilk); 1580 flags &= ~LK_INTERLOCK; 1581 } 1582 GIANT_SAVE(); 1583 sleepq_add(&lk->lock_object, NULL, iwmesg, SLEEPQ_LK, 1584 SQ_EXCLUSIVE_QUEUE); 1585 sleepq_wait(&lk->lock_object, ipri & PRIMASK); 1586 GIANT_RESTORE(); 1587 LOCK_LOG2(lk, "%s: %p resuming from the sleep queue", 1588 __func__, lk); 1589 } 1590 1591 if (error == 0) { 1592 lock_profile_obtain_lock_success(&lk->lock_object, 1593 false, contested, waittime, file, line); 1594 LOCK_LOG_LOCK("DRAIN", &lk->lock_object, 0, 1595 lk->lk_recurse, file, line); 1596 WITNESS_LOCK(&lk->lock_object, LOP_EXCLUSIVE | 1597 LK_TRYWIT(flags), file, line); 1598 TD_LOCKS_INC(curthread); 1599 STACK_SAVE(lk); 1600 } 1601 break; 1602 default: 1603 if (flags & LK_INTERLOCK) 1604 class->lc_unlock(ilk); 1605 panic("%s: unknown lockmgr request 0x%x\n", __func__, op); 1606 } 1607 1608 if (flags & LK_INTERLOCK) 1609 class->lc_unlock(ilk); 1610 if (wakeup_swapper) 1611 kick_proc0(); 1612 1613 return (error); 1614 } 1615 1616 void 1617 _lockmgr_disown(struct lock *lk, const char *file, int line) 1618 { 1619 uintptr_t tid, x; 1620 1621 if (SCHEDULER_STOPPED()) 1622 return; 1623 1624 tid = (uintptr_t)curthread; 1625 _lockmgr_assert(lk, KA_XLOCKED, file, line); 1626 1627 /* 1628 * Panic if the lock is recursed. 1629 */ 1630 if (lockmgr_xlocked(lk) && lockmgr_recursed(lk)) 1631 panic("%s: disown a recursed lockmgr @ %s:%d\n", 1632 __func__, file, line); 1633 1634 /* 1635 * If the owner is already LK_KERNPROC just skip the whole operation. 1636 */ 1637 if (LK_HOLDER(lk->lk_lock) != tid) 1638 return; 1639 lock_profile_release_lock(&lk->lock_object, false); 1640 LOCKSTAT_RECORD1(lockmgr__disown, lk, LOCKSTAT_WRITER); 1641 LOCK_LOG_LOCK("XDISOWN", &lk->lock_object, 0, 0, file, line); 1642 WITNESS_UNLOCK(&lk->lock_object, LOP_EXCLUSIVE, file, line); 1643 TD_LOCKS_DEC(curthread); 1644 STACK_SAVE(lk); 1645 1646 /* 1647 * In order to preserve waiters flags, just spin. 1648 */ 1649 for (;;) { 1650 x = lockmgr_read_value(lk); 1651 MPASS((x & LK_EXCLUSIVE_SPINNERS) == 0); 1652 x &= LK_ALL_WAITERS; 1653 if (atomic_cmpset_rel_ptr(&lk->lk_lock, tid | x, 1654 LK_KERNPROC | x)) 1655 return; 1656 cpu_spinwait(); 1657 } 1658 } 1659 1660 void 1661 lockmgr_printinfo(const struct lock *lk) 1662 { 1663 struct thread *td; 1664 uintptr_t x; 1665 1666 if (lk->lk_lock == LK_UNLOCKED) 1667 printf("lock type %s: UNLOCKED\n", lk->lock_object.lo_name); 1668 else if (lk->lk_lock & LK_SHARE) 1669 printf("lock type %s: SHARED (count %ju)\n", 1670 lk->lock_object.lo_name, 1671 (uintmax_t)LK_SHARERS(lk->lk_lock)); 1672 else { 1673 td = lockmgr_xholder(lk); 1674 if (td == (struct thread *)LK_KERNPROC) 1675 printf("lock type %s: EXCL by KERNPROC\n", 1676 lk->lock_object.lo_name); 1677 else 1678 printf("lock type %s: EXCL by thread %p " 1679 "(pid %d, %s, tid %d)\n", lk->lock_object.lo_name, 1680 td, td->td_proc->p_pid, td->td_proc->p_comm, 1681 td->td_tid); 1682 } 1683 1684 x = lk->lk_lock; 1685 if (x & LK_EXCLUSIVE_WAITERS) 1686 printf(" with exclusive waiters pending\n"); 1687 if (x & LK_SHARED_WAITERS) 1688 printf(" with shared waiters pending\n"); 1689 if (x & LK_EXCLUSIVE_SPINNERS) 1690 printf(" with exclusive spinners pending\n"); 1691 1692 STACK_PRINT(lk); 1693 } 1694 1695 int 1696 lockstatus(const struct lock *lk) 1697 { 1698 uintptr_t v, x; 1699 int ret; 1700 1701 ret = LK_SHARED; 1702 x = lockmgr_read_value(lk); 1703 v = LK_HOLDER(x); 1704 1705 if ((x & LK_SHARE) == 0) { 1706 if (v == (uintptr_t)curthread || v == LK_KERNPROC) 1707 ret = LK_EXCLUSIVE; 1708 else 1709 ret = LK_EXCLOTHER; 1710 } else if (x == LK_UNLOCKED) 1711 ret = 0; 1712 1713 return (ret); 1714 } 1715 1716 #ifdef INVARIANT_SUPPORT 1717 1718 FEATURE(invariant_support, 1719 "Support for modules compiled with INVARIANTS option"); 1720 1721 #ifndef INVARIANTS 1722 #undef _lockmgr_assert 1723 #endif 1724 1725 void 1726 _lockmgr_assert(const struct lock *lk, int what, const char *file, int line) 1727 { 1728 int slocked = 0; 1729 1730 if (SCHEDULER_STOPPED()) 1731 return; 1732 switch (what) { 1733 case KA_SLOCKED: 1734 case KA_SLOCKED | KA_NOTRECURSED: 1735 case KA_SLOCKED | KA_RECURSED: 1736 slocked = 1; 1737 case KA_LOCKED: 1738 case KA_LOCKED | KA_NOTRECURSED: 1739 case KA_LOCKED | KA_RECURSED: 1740 #ifdef WITNESS 1741 1742 /* 1743 * We cannot trust WITNESS if the lock is held in exclusive 1744 * mode and a call to lockmgr_disown() happened. 1745 * Workaround this skipping the check if the lock is held in 1746 * exclusive mode even for the KA_LOCKED case. 1747 */ 1748 if (slocked || (lk->lk_lock & LK_SHARE)) { 1749 witness_assert(&lk->lock_object, what, file, line); 1750 break; 1751 } 1752 #endif 1753 if (lk->lk_lock == LK_UNLOCKED || 1754 ((lk->lk_lock & LK_SHARE) == 0 && (slocked || 1755 (!lockmgr_xlocked(lk) && !lockmgr_disowned(lk))))) 1756 panic("Lock %s not %slocked @ %s:%d\n", 1757 lk->lock_object.lo_name, slocked ? "share" : "", 1758 file, line); 1759 1760 if ((lk->lk_lock & LK_SHARE) == 0) { 1761 if (lockmgr_recursed(lk)) { 1762 if (what & KA_NOTRECURSED) 1763 panic("Lock %s recursed @ %s:%d\n", 1764 lk->lock_object.lo_name, file, 1765 line); 1766 } else if (what & KA_RECURSED) 1767 panic("Lock %s not recursed @ %s:%d\n", 1768 lk->lock_object.lo_name, file, line); 1769 } 1770 break; 1771 case KA_XLOCKED: 1772 case KA_XLOCKED | KA_NOTRECURSED: 1773 case KA_XLOCKED | KA_RECURSED: 1774 if (!lockmgr_xlocked(lk) && !lockmgr_disowned(lk)) 1775 panic("Lock %s not exclusively locked @ %s:%d\n", 1776 lk->lock_object.lo_name, file, line); 1777 if (lockmgr_recursed(lk)) { 1778 if (what & KA_NOTRECURSED) 1779 panic("Lock %s recursed @ %s:%d\n", 1780 lk->lock_object.lo_name, file, line); 1781 } else if (what & KA_RECURSED) 1782 panic("Lock %s not recursed @ %s:%d\n", 1783 lk->lock_object.lo_name, file, line); 1784 break; 1785 case KA_UNLOCKED: 1786 if (lockmgr_xlocked(lk) || lockmgr_disowned(lk)) 1787 panic("Lock %s exclusively locked @ %s:%d\n", 1788 lk->lock_object.lo_name, file, line); 1789 break; 1790 default: 1791 panic("Unknown lockmgr assertion: %d @ %s:%d\n", what, file, 1792 line); 1793 } 1794 } 1795 #endif 1796 1797 #ifdef DDB 1798 int 1799 lockmgr_chain(struct thread *td, struct thread **ownerp) 1800 { 1801 const struct lock *lk; 1802 1803 lk = td->td_wchan; 1804 1805 if (LOCK_CLASS(&lk->lock_object) != &lock_class_lockmgr) 1806 return (0); 1807 db_printf("blocked on lockmgr %s", lk->lock_object.lo_name); 1808 if (lk->lk_lock & LK_SHARE) 1809 db_printf("SHARED (count %ju)\n", 1810 (uintmax_t)LK_SHARERS(lk->lk_lock)); 1811 else 1812 db_printf("EXCL\n"); 1813 *ownerp = lockmgr_xholder(lk); 1814 1815 return (1); 1816 } 1817 1818 static void 1819 db_show_lockmgr(const struct lock_object *lock) 1820 { 1821 struct thread *td; 1822 const struct lock *lk; 1823 1824 lk = (const struct lock *)lock; 1825 1826 db_printf(" state: "); 1827 if (lk->lk_lock == LK_UNLOCKED) 1828 db_printf("UNLOCKED\n"); 1829 else if (lk->lk_lock & LK_SHARE) 1830 db_printf("SLOCK: %ju\n", (uintmax_t)LK_SHARERS(lk->lk_lock)); 1831 else { 1832 td = lockmgr_xholder(lk); 1833 if (td == (struct thread *)LK_KERNPROC) 1834 db_printf("XLOCK: LK_KERNPROC\n"); 1835 else 1836 db_printf("XLOCK: %p (tid %d, pid %d, \"%s\")\n", td, 1837 td->td_tid, td->td_proc->p_pid, 1838 td->td_proc->p_comm); 1839 if (lockmgr_recursed(lk)) 1840 db_printf(" recursed: %d\n", lk->lk_recurse); 1841 } 1842 db_printf(" waiters: "); 1843 switch (lk->lk_lock & LK_ALL_WAITERS) { 1844 case LK_SHARED_WAITERS: 1845 db_printf("shared\n"); 1846 break; 1847 case LK_EXCLUSIVE_WAITERS: 1848 db_printf("exclusive\n"); 1849 break; 1850 case LK_ALL_WAITERS: 1851 db_printf("shared and exclusive\n"); 1852 break; 1853 default: 1854 db_printf("none\n"); 1855 } 1856 db_printf(" spinners: "); 1857 if (lk->lk_lock & LK_EXCLUSIVE_SPINNERS) 1858 db_printf("exclusive\n"); 1859 else 1860 db_printf("none\n"); 1861 } 1862 #endif 1863