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