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