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