1 /*- 2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD 3 * 4 * Copyright (c) 2006 John Baldwin <jhb@FreeBSD.org> 5 * 6 * Redistribution and use in source and binary forms, with or without 7 * modification, are permitted provided that the following conditions 8 * are met: 9 * 1. Redistributions of source code must retain the above copyright 10 * notice, this list of conditions and the following disclaimer. 11 * 2. Redistributions in binary form must reproduce the above copyright 12 * notice, this list of conditions and the following disclaimer in the 13 * documentation and/or other materials provided with the distribution. 14 * 15 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 16 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 17 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 18 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 19 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 20 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 21 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 22 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 23 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 24 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 25 * SUCH DAMAGE. 26 */ 27 28 /* 29 * Machine independent bits of reader/writer lock implementation. 30 */ 31 32 #include <sys/cdefs.h> 33 __FBSDID("$FreeBSD$"); 34 35 #include "opt_ddb.h" 36 #include "opt_hwpmc_hooks.h" 37 #include "opt_no_adaptive_rwlocks.h" 38 39 #include <sys/param.h> 40 #include <sys/kdb.h> 41 #include <sys/ktr.h> 42 #include <sys/kernel.h> 43 #include <sys/lock.h> 44 #include <sys/mutex.h> 45 #include <sys/proc.h> 46 #include <sys/rwlock.h> 47 #include <sys/sched.h> 48 #include <sys/smp.h> 49 #include <sys/sysctl.h> 50 #include <sys/systm.h> 51 #include <sys/turnstile.h> 52 53 #include <machine/cpu.h> 54 55 #if defined(SMP) && !defined(NO_ADAPTIVE_RWLOCKS) 56 #define ADAPTIVE_RWLOCKS 57 #endif 58 59 #ifdef HWPMC_HOOKS 60 #include <sys/pmckern.h> 61 PMC_SOFT_DECLARE( , , lock, failed); 62 #endif 63 64 /* 65 * Return the rwlock address when the lock cookie address is provided. 66 * This functionality assumes that struct rwlock* have a member named rw_lock. 67 */ 68 #define rwlock2rw(c) (__containerof(c, struct rwlock, rw_lock)) 69 70 #ifdef DDB 71 #include <ddb/ddb.h> 72 73 static void db_show_rwlock(const struct lock_object *lock); 74 #endif 75 static void assert_rw(const struct lock_object *lock, int what); 76 static void lock_rw(struct lock_object *lock, uintptr_t how); 77 #ifdef KDTRACE_HOOKS 78 static int owner_rw(const struct lock_object *lock, struct thread **owner); 79 #endif 80 static uintptr_t unlock_rw(struct lock_object *lock); 81 82 struct lock_class lock_class_rw = { 83 .lc_name = "rw", 84 .lc_flags = LC_SLEEPLOCK | LC_RECURSABLE | LC_UPGRADABLE, 85 .lc_assert = assert_rw, 86 #ifdef DDB 87 .lc_ddb_show = db_show_rwlock, 88 #endif 89 .lc_lock = lock_rw, 90 .lc_unlock = unlock_rw, 91 #ifdef KDTRACE_HOOKS 92 .lc_owner = owner_rw, 93 #endif 94 }; 95 96 #ifdef ADAPTIVE_RWLOCKS 97 #ifdef RWLOCK_CUSTOM_BACKOFF 98 static u_short __read_frequently rowner_retries; 99 static u_short __read_frequently rowner_loops; 100 static SYSCTL_NODE(_debug, OID_AUTO, rwlock, CTLFLAG_RD, NULL, 101 "rwlock debugging"); 102 SYSCTL_U16(_debug_rwlock, OID_AUTO, retry, CTLFLAG_RW, &rowner_retries, 0, ""); 103 SYSCTL_U16(_debug_rwlock, OID_AUTO, loops, CTLFLAG_RW, &rowner_loops, 0, ""); 104 105 static struct lock_delay_config __read_frequently rw_delay; 106 107 SYSCTL_U16(_debug_rwlock, OID_AUTO, delay_base, CTLFLAG_RW, &rw_delay.base, 108 0, ""); 109 SYSCTL_U16(_debug_rwlock, OID_AUTO, delay_max, CTLFLAG_RW, &rw_delay.max, 110 0, ""); 111 112 static void 113 rw_lock_delay_init(void *arg __unused) 114 { 115 116 lock_delay_default_init(&rw_delay); 117 rowner_retries = 10; 118 rowner_loops = max(10000, rw_delay.max); 119 } 120 LOCK_DELAY_SYSINIT(rw_lock_delay_init); 121 #else 122 #define rw_delay locks_delay 123 #define rowner_retries locks_delay_retries 124 #define rowner_loops locks_delay_loops 125 #endif 126 #endif 127 128 /* 129 * Return a pointer to the owning thread if the lock is write-locked or 130 * NULL if the lock is unlocked or read-locked. 131 */ 132 133 #define lv_rw_wowner(v) \ 134 ((v) & RW_LOCK_READ ? NULL : \ 135 (struct thread *)RW_OWNER((v))) 136 137 #define rw_wowner(rw) lv_rw_wowner(RW_READ_VALUE(rw)) 138 139 /* 140 * Returns if a write owner is recursed. Write ownership is not assured 141 * here and should be previously checked. 142 */ 143 #define rw_recursed(rw) ((rw)->rw_recurse != 0) 144 145 /* 146 * Return true if curthread helds the lock. 147 */ 148 #define rw_wlocked(rw) (rw_wowner((rw)) == curthread) 149 150 /* 151 * Return a pointer to the owning thread for this lock who should receive 152 * any priority lent by threads that block on this lock. Currently this 153 * is identical to rw_wowner(). 154 */ 155 #define rw_owner(rw) rw_wowner(rw) 156 157 #ifndef INVARIANTS 158 #define __rw_assert(c, what, file, line) 159 #endif 160 161 void 162 assert_rw(const struct lock_object *lock, int what) 163 { 164 165 rw_assert((const struct rwlock *)lock, what); 166 } 167 168 void 169 lock_rw(struct lock_object *lock, uintptr_t how) 170 { 171 struct rwlock *rw; 172 173 rw = (struct rwlock *)lock; 174 if (how) 175 rw_rlock(rw); 176 else 177 rw_wlock(rw); 178 } 179 180 uintptr_t 181 unlock_rw(struct lock_object *lock) 182 { 183 struct rwlock *rw; 184 185 rw = (struct rwlock *)lock; 186 rw_assert(rw, RA_LOCKED | LA_NOTRECURSED); 187 if (rw->rw_lock & RW_LOCK_READ) { 188 rw_runlock(rw); 189 return (1); 190 } else { 191 rw_wunlock(rw); 192 return (0); 193 } 194 } 195 196 #ifdef KDTRACE_HOOKS 197 int 198 owner_rw(const struct lock_object *lock, struct thread **owner) 199 { 200 const struct rwlock *rw = (const struct rwlock *)lock; 201 uintptr_t x = rw->rw_lock; 202 203 *owner = rw_wowner(rw); 204 return ((x & RW_LOCK_READ) != 0 ? (RW_READERS(x) != 0) : 205 (*owner != NULL)); 206 } 207 #endif 208 209 void 210 _rw_init_flags(volatile uintptr_t *c, const char *name, int opts) 211 { 212 struct rwlock *rw; 213 int flags; 214 215 rw = rwlock2rw(c); 216 217 MPASS((opts & ~(RW_DUPOK | RW_NOPROFILE | RW_NOWITNESS | RW_QUIET | 218 RW_RECURSE | RW_NEW)) == 0); 219 ASSERT_ATOMIC_LOAD_PTR(rw->rw_lock, 220 ("%s: rw_lock not aligned for %s: %p", __func__, name, 221 &rw->rw_lock)); 222 223 flags = LO_UPGRADABLE; 224 if (opts & RW_DUPOK) 225 flags |= LO_DUPOK; 226 if (opts & RW_NOPROFILE) 227 flags |= LO_NOPROFILE; 228 if (!(opts & RW_NOWITNESS)) 229 flags |= LO_WITNESS; 230 if (opts & RW_RECURSE) 231 flags |= LO_RECURSABLE; 232 if (opts & RW_QUIET) 233 flags |= LO_QUIET; 234 if (opts & RW_NEW) 235 flags |= LO_NEW; 236 237 lock_init(&rw->lock_object, &lock_class_rw, name, NULL, flags); 238 rw->rw_lock = RW_UNLOCKED; 239 rw->rw_recurse = 0; 240 } 241 242 void 243 _rw_destroy(volatile uintptr_t *c) 244 { 245 struct rwlock *rw; 246 247 rw = rwlock2rw(c); 248 249 KASSERT(rw->rw_lock == RW_UNLOCKED, ("rw lock %p not unlocked", rw)); 250 KASSERT(rw->rw_recurse == 0, ("rw lock %p still recursed", rw)); 251 rw->rw_lock = RW_DESTROYED; 252 lock_destroy(&rw->lock_object); 253 } 254 255 void 256 rw_sysinit(void *arg) 257 { 258 struct rw_args *args; 259 260 args = arg; 261 rw_init_flags((struct rwlock *)args->ra_rw, args->ra_desc, 262 args->ra_flags); 263 } 264 265 int 266 _rw_wowned(const volatile uintptr_t *c) 267 { 268 269 return (rw_wowner(rwlock2rw(c)) == curthread); 270 } 271 272 void 273 _rw_wlock_cookie(volatile uintptr_t *c, const char *file, int line) 274 { 275 struct rwlock *rw; 276 uintptr_t tid, v; 277 278 rw = rwlock2rw(c); 279 280 KASSERT(kdb_active != 0 || SCHEDULER_STOPPED() || 281 !TD_IS_IDLETHREAD(curthread), 282 ("rw_wlock() by idle thread %p on rwlock %s @ %s:%d", 283 curthread, rw->lock_object.lo_name, file, line)); 284 KASSERT(rw->rw_lock != RW_DESTROYED, 285 ("rw_wlock() of destroyed rwlock @ %s:%d", file, line)); 286 WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE, file, 287 line, NULL); 288 tid = (uintptr_t)curthread; 289 v = RW_UNLOCKED; 290 if (!_rw_write_lock_fetch(rw, &v, tid)) 291 _rw_wlock_hard(rw, v, file, line); 292 else 293 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire, rw, 294 0, 0, file, line, LOCKSTAT_WRITER); 295 296 LOCK_LOG_LOCK("WLOCK", &rw->lock_object, 0, rw->rw_recurse, file, line); 297 WITNESS_LOCK(&rw->lock_object, LOP_EXCLUSIVE, file, line); 298 TD_LOCKS_INC(curthread); 299 } 300 301 int 302 __rw_try_wlock_int(struct rwlock *rw LOCK_FILE_LINE_ARG_DEF) 303 { 304 struct thread *td; 305 uintptr_t tid, v; 306 int rval; 307 bool recursed; 308 309 td = curthread; 310 tid = (uintptr_t)td; 311 if (SCHEDULER_STOPPED_TD(td)) 312 return (1); 313 314 KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(td), 315 ("rw_try_wlock() by idle thread %p on rwlock %s @ %s:%d", 316 curthread, rw->lock_object.lo_name, file, line)); 317 KASSERT(rw->rw_lock != RW_DESTROYED, 318 ("rw_try_wlock() of destroyed rwlock @ %s:%d", file, line)); 319 320 rval = 1; 321 recursed = false; 322 v = RW_UNLOCKED; 323 for (;;) { 324 if (atomic_fcmpset_acq_ptr(&rw->rw_lock, &v, tid)) 325 break; 326 if (v == RW_UNLOCKED) 327 continue; 328 if (v == tid && (rw->lock_object.lo_flags & LO_RECURSABLE)) { 329 rw->rw_recurse++; 330 atomic_set_ptr(&rw->rw_lock, RW_LOCK_WRITER_RECURSED); 331 break; 332 } 333 rval = 0; 334 break; 335 } 336 337 LOCK_LOG_TRY("WLOCK", &rw->lock_object, 0, rval, file, line); 338 if (rval) { 339 WITNESS_LOCK(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK, 340 file, line); 341 if (!recursed) 342 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire, 343 rw, 0, 0, file, line, LOCKSTAT_WRITER); 344 TD_LOCKS_INC(curthread); 345 } 346 return (rval); 347 } 348 349 int 350 __rw_try_wlock(volatile uintptr_t *c, const char *file, int line) 351 { 352 struct rwlock *rw; 353 354 rw = rwlock2rw(c); 355 return (__rw_try_wlock_int(rw LOCK_FILE_LINE_ARG)); 356 } 357 358 void 359 _rw_wunlock_cookie(volatile uintptr_t *c, const char *file, int line) 360 { 361 struct rwlock *rw; 362 363 rw = rwlock2rw(c); 364 365 KASSERT(rw->rw_lock != RW_DESTROYED, 366 ("rw_wunlock() of destroyed rwlock @ %s:%d", file, line)); 367 __rw_assert(c, RA_WLOCKED, file, line); 368 WITNESS_UNLOCK(&rw->lock_object, LOP_EXCLUSIVE, file, line); 369 LOCK_LOG_LOCK("WUNLOCK", &rw->lock_object, 0, rw->rw_recurse, file, 370 line); 371 372 #ifdef LOCK_PROFILING 373 _rw_wunlock_hard(rw, (uintptr_t)curthread, file, line); 374 #else 375 __rw_wunlock(rw, curthread, file, line); 376 #endif 377 378 TD_LOCKS_DEC(curthread); 379 } 380 381 /* 382 * Determines whether a new reader can acquire a lock. Succeeds if the 383 * reader already owns a read lock and the lock is locked for read to 384 * prevent deadlock from reader recursion. Also succeeds if the lock 385 * is unlocked and has no writer waiters or spinners. Failing otherwise 386 * prioritizes writers before readers. 387 */ 388 static bool __always_inline 389 __rw_can_read(struct thread *td, uintptr_t v, bool fp) 390 { 391 392 if ((v & (RW_LOCK_READ | RW_LOCK_WRITE_WAITERS | RW_LOCK_WRITE_SPINNER)) 393 == RW_LOCK_READ) 394 return (true); 395 if (!fp && td->td_rw_rlocks && (v & RW_LOCK_READ)) 396 return (true); 397 return (false); 398 } 399 400 static bool __always_inline 401 __rw_rlock_try(struct rwlock *rw, struct thread *td, uintptr_t *vp, bool fp 402 LOCK_FILE_LINE_ARG_DEF) 403 { 404 405 /* 406 * Handle the easy case. If no other thread has a write 407 * lock, then try to bump up the count of read locks. Note 408 * that we have to preserve the current state of the 409 * RW_LOCK_WRITE_WAITERS flag. If we fail to acquire a 410 * read lock, then rw_lock must have changed, so restart 411 * the loop. Note that this handles the case of a 412 * completely unlocked rwlock since such a lock is encoded 413 * as a read lock with no waiters. 414 */ 415 while (__rw_can_read(td, *vp, fp)) { 416 if (atomic_fcmpset_acq_ptr(&rw->rw_lock, vp, 417 *vp + RW_ONE_READER)) { 418 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 419 CTR4(KTR_LOCK, 420 "%s: %p succeed %p -> %p", __func__, 421 rw, (void *)*vp, 422 (void *)(*vp + RW_ONE_READER)); 423 td->td_rw_rlocks++; 424 return (true); 425 } 426 } 427 return (false); 428 } 429 430 static void __noinline 431 __rw_rlock_hard(struct rwlock *rw, struct thread *td, uintptr_t v 432 LOCK_FILE_LINE_ARG_DEF) 433 { 434 struct turnstile *ts; 435 struct thread *owner; 436 #ifdef ADAPTIVE_RWLOCKS 437 int spintries = 0; 438 int i, n; 439 #endif 440 #ifdef LOCK_PROFILING 441 uint64_t waittime = 0; 442 int contested = 0; 443 #endif 444 #if defined(ADAPTIVE_RWLOCKS) || defined(KDTRACE_HOOKS) 445 struct lock_delay_arg lda; 446 #endif 447 #ifdef KDTRACE_HOOKS 448 u_int sleep_cnt = 0; 449 int64_t sleep_time = 0; 450 int64_t all_time = 0; 451 #endif 452 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING) 453 uintptr_t state = 0; 454 int doing_lockprof = 0; 455 #endif 456 457 #ifdef KDTRACE_HOOKS 458 if (LOCKSTAT_PROFILE_ENABLED(rw__acquire)) { 459 if (__rw_rlock_try(rw, td, &v, false LOCK_FILE_LINE_ARG)) 460 goto out_lockstat; 461 doing_lockprof = 1; 462 all_time -= lockstat_nsecs(&rw->lock_object); 463 state = v; 464 } 465 #endif 466 #ifdef LOCK_PROFILING 467 doing_lockprof = 1; 468 state = v; 469 #endif 470 471 if (SCHEDULER_STOPPED()) 472 return; 473 474 #if defined(ADAPTIVE_RWLOCKS) 475 lock_delay_arg_init(&lda, &rw_delay); 476 #elif defined(KDTRACE_HOOKS) 477 lock_delay_arg_init(&lda, NULL); 478 #endif 479 480 #ifdef HWPMC_HOOKS 481 PMC_SOFT_CALL( , , lock, failed); 482 #endif 483 lock_profile_obtain_lock_failed(&rw->lock_object, 484 &contested, &waittime); 485 486 for (;;) { 487 if (__rw_rlock_try(rw, td, &v, false LOCK_FILE_LINE_ARG)) 488 break; 489 #ifdef KDTRACE_HOOKS 490 lda.spin_cnt++; 491 #endif 492 493 #ifdef ADAPTIVE_RWLOCKS 494 /* 495 * If the owner is running on another CPU, spin until 496 * the owner stops running or the state of the lock 497 * changes. 498 */ 499 if ((v & RW_LOCK_READ) == 0) { 500 owner = (struct thread *)RW_OWNER(v); 501 if (TD_IS_RUNNING(owner)) { 502 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 503 CTR3(KTR_LOCK, 504 "%s: spinning on %p held by %p", 505 __func__, rw, owner); 506 KTR_STATE1(KTR_SCHED, "thread", 507 sched_tdname(curthread), "spinning", 508 "lockname:\"%s\"", rw->lock_object.lo_name); 509 do { 510 lock_delay(&lda); 511 v = RW_READ_VALUE(rw); 512 owner = lv_rw_wowner(v); 513 } while (owner != NULL && TD_IS_RUNNING(owner)); 514 KTR_STATE0(KTR_SCHED, "thread", 515 sched_tdname(curthread), "running"); 516 continue; 517 } 518 } else { 519 if ((v & RW_LOCK_WRITE_SPINNER) && RW_READERS(v) == 0) { 520 MPASS(!__rw_can_read(td, v, false)); 521 lock_delay_spin(2); 522 v = RW_READ_VALUE(rw); 523 continue; 524 } 525 if (spintries < rowner_retries) { 526 spintries++; 527 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread), 528 "spinning", "lockname:\"%s\"", 529 rw->lock_object.lo_name); 530 n = RW_READERS(v); 531 for (i = 0; i < rowner_loops; i += n) { 532 lock_delay_spin(n); 533 v = RW_READ_VALUE(rw); 534 if (!(v & RW_LOCK_READ)) 535 break; 536 n = RW_READERS(v); 537 if (n == 0) 538 break; 539 if (__rw_can_read(td, v, false)) 540 break; 541 } 542 #ifdef KDTRACE_HOOKS 543 lda.spin_cnt += rowner_loops - i; 544 #endif 545 KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread), 546 "running"); 547 if (i < rowner_loops) 548 continue; 549 } 550 } 551 #endif 552 553 /* 554 * Okay, now it's the hard case. Some other thread already 555 * has a write lock or there are write waiters present, 556 * acquire the turnstile lock so we can begin the process 557 * of blocking. 558 */ 559 ts = turnstile_trywait(&rw->lock_object); 560 561 /* 562 * The lock might have been released while we spun, so 563 * recheck its state and restart the loop if needed. 564 */ 565 v = RW_READ_VALUE(rw); 566 retry_ts: 567 if (((v & RW_LOCK_WRITE_SPINNER) && RW_READERS(v) == 0) || 568 __rw_can_read(td, v, false)) { 569 turnstile_cancel(ts); 570 continue; 571 } 572 573 owner = lv_rw_wowner(v); 574 575 #ifdef ADAPTIVE_RWLOCKS 576 /* 577 * The current lock owner might have started executing 578 * on another CPU (or the lock could have changed 579 * owners) while we were waiting on the turnstile 580 * chain lock. If so, drop the turnstile lock and try 581 * again. 582 */ 583 if (owner != NULL) { 584 if (TD_IS_RUNNING(owner)) { 585 turnstile_cancel(ts); 586 continue; 587 } 588 } 589 #endif 590 591 /* 592 * The lock is held in write mode or it already has waiters. 593 */ 594 MPASS(!__rw_can_read(td, v, false)); 595 596 /* 597 * If the RW_LOCK_READ_WAITERS flag is already set, then 598 * we can go ahead and block. If it is not set then try 599 * to set it. If we fail to set it drop the turnstile 600 * lock and restart the loop. 601 */ 602 if (!(v & RW_LOCK_READ_WAITERS)) { 603 if (!atomic_fcmpset_ptr(&rw->rw_lock, &v, 604 v | RW_LOCK_READ_WAITERS)) 605 goto retry_ts; 606 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 607 CTR2(KTR_LOCK, "%s: %p set read waiters flag", 608 __func__, rw); 609 } 610 611 /* 612 * We were unable to acquire the lock and the read waiters 613 * flag is set, so we must block on the turnstile. 614 */ 615 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 616 CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__, 617 rw); 618 #ifdef KDTRACE_HOOKS 619 sleep_time -= lockstat_nsecs(&rw->lock_object); 620 #endif 621 MPASS(owner == rw_owner(rw)); 622 turnstile_wait(ts, owner, TS_SHARED_QUEUE); 623 #ifdef KDTRACE_HOOKS 624 sleep_time += lockstat_nsecs(&rw->lock_object); 625 sleep_cnt++; 626 #endif 627 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 628 CTR2(KTR_LOCK, "%s: %p resuming from turnstile", 629 __func__, rw); 630 v = RW_READ_VALUE(rw); 631 } 632 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING) 633 if (__predict_true(!doing_lockprof)) 634 return; 635 #endif 636 #ifdef KDTRACE_HOOKS 637 all_time += lockstat_nsecs(&rw->lock_object); 638 if (sleep_time) 639 LOCKSTAT_RECORD4(rw__block, rw, sleep_time, 640 LOCKSTAT_READER, (state & RW_LOCK_READ) == 0, 641 (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state)); 642 643 /* Record only the loops spinning and not sleeping. */ 644 if (lda.spin_cnt > sleep_cnt) 645 LOCKSTAT_RECORD4(rw__spin, rw, all_time - sleep_time, 646 LOCKSTAT_READER, (state & RW_LOCK_READ) == 0, 647 (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state)); 648 out_lockstat: 649 #endif 650 /* 651 * TODO: acquire "owner of record" here. Here be turnstile dragons 652 * however. turnstiles don't like owners changing between calls to 653 * turnstile_wait() currently. 654 */ 655 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire, rw, contested, 656 waittime, file, line, LOCKSTAT_READER); 657 } 658 659 void 660 __rw_rlock_int(struct rwlock *rw LOCK_FILE_LINE_ARG_DEF) 661 { 662 struct thread *td; 663 uintptr_t v; 664 665 td = curthread; 666 667 KASSERT(kdb_active != 0 || SCHEDULER_STOPPED_TD(td) || 668 !TD_IS_IDLETHREAD(td), 669 ("rw_rlock() by idle thread %p on rwlock %s @ %s:%d", 670 td, rw->lock_object.lo_name, file, line)); 671 KASSERT(rw->rw_lock != RW_DESTROYED, 672 ("rw_rlock() of destroyed rwlock @ %s:%d", file, line)); 673 KASSERT(rw_wowner(rw) != td, 674 ("rw_rlock: wlock already held for %s @ %s:%d", 675 rw->lock_object.lo_name, file, line)); 676 WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER, file, line, NULL); 677 678 v = RW_READ_VALUE(rw); 679 if (__predict_false(LOCKSTAT_PROFILE_ENABLED(rw__acquire) || 680 !__rw_rlock_try(rw, td, &v, true LOCK_FILE_LINE_ARG))) 681 __rw_rlock_hard(rw, td, v LOCK_FILE_LINE_ARG); 682 else 683 lock_profile_obtain_lock_success(&rw->lock_object, 0, 0, 684 file, line); 685 686 LOCK_LOG_LOCK("RLOCK", &rw->lock_object, 0, 0, file, line); 687 WITNESS_LOCK(&rw->lock_object, 0, file, line); 688 TD_LOCKS_INC(curthread); 689 } 690 691 void 692 __rw_rlock(volatile uintptr_t *c, const char *file, int line) 693 { 694 struct rwlock *rw; 695 696 rw = rwlock2rw(c); 697 __rw_rlock_int(rw LOCK_FILE_LINE_ARG); 698 } 699 700 int 701 __rw_try_rlock_int(struct rwlock *rw LOCK_FILE_LINE_ARG_DEF) 702 { 703 uintptr_t x; 704 705 if (SCHEDULER_STOPPED()) 706 return (1); 707 708 KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread), 709 ("rw_try_rlock() by idle thread %p on rwlock %s @ %s:%d", 710 curthread, rw->lock_object.lo_name, file, line)); 711 712 x = rw->rw_lock; 713 for (;;) { 714 KASSERT(rw->rw_lock != RW_DESTROYED, 715 ("rw_try_rlock() of destroyed rwlock @ %s:%d", file, line)); 716 if (!(x & RW_LOCK_READ)) 717 break; 718 if (atomic_fcmpset_acq_ptr(&rw->rw_lock, &x, x + RW_ONE_READER)) { 719 LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 1, file, 720 line); 721 WITNESS_LOCK(&rw->lock_object, LOP_TRYLOCK, file, line); 722 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire, 723 rw, 0, 0, file, line, LOCKSTAT_READER); 724 TD_LOCKS_INC(curthread); 725 curthread->td_rw_rlocks++; 726 return (1); 727 } 728 } 729 730 LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 0, file, line); 731 return (0); 732 } 733 734 int 735 __rw_try_rlock(volatile uintptr_t *c, const char *file, int line) 736 { 737 struct rwlock *rw; 738 739 rw = rwlock2rw(c); 740 return (__rw_try_rlock_int(rw LOCK_FILE_LINE_ARG)); 741 } 742 743 static bool __always_inline 744 __rw_runlock_try(struct rwlock *rw, struct thread *td, uintptr_t *vp) 745 { 746 747 for (;;) { 748 if (RW_READERS(*vp) > 1 || !(*vp & RW_LOCK_WAITERS)) { 749 if (atomic_fcmpset_rel_ptr(&rw->rw_lock, vp, 750 *vp - RW_ONE_READER)) { 751 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 752 CTR4(KTR_LOCK, 753 "%s: %p succeeded %p -> %p", 754 __func__, rw, (void *)*vp, 755 (void *)(*vp - RW_ONE_READER)); 756 td->td_rw_rlocks--; 757 return (true); 758 } 759 continue; 760 } 761 break; 762 } 763 return (false); 764 } 765 766 static void __noinline 767 __rw_runlock_hard(struct rwlock *rw, struct thread *td, uintptr_t v 768 LOCK_FILE_LINE_ARG_DEF) 769 { 770 struct turnstile *ts; 771 uintptr_t setv, queue; 772 773 if (SCHEDULER_STOPPED()) 774 return; 775 776 if (__rw_runlock_try(rw, td, &v)) 777 goto out_lockstat; 778 779 /* 780 * Ok, we know we have waiters and we think we are the 781 * last reader, so grab the turnstile lock. 782 */ 783 turnstile_chain_lock(&rw->lock_object); 784 v = RW_READ_VALUE(rw); 785 for (;;) { 786 if (__rw_runlock_try(rw, td, &v)) 787 break; 788 789 MPASS(v & RW_LOCK_WAITERS); 790 791 /* 792 * Try to drop our lock leaving the lock in a unlocked 793 * state. 794 * 795 * If you wanted to do explicit lock handoff you'd have to 796 * do it here. You'd also want to use turnstile_signal() 797 * and you'd have to handle the race where a higher 798 * priority thread blocks on the write lock before the 799 * thread you wakeup actually runs and have the new thread 800 * "steal" the lock. For now it's a lot simpler to just 801 * wakeup all of the waiters. 802 * 803 * As above, if we fail, then another thread might have 804 * acquired a read lock, so drop the turnstile lock and 805 * restart. 806 */ 807 setv = RW_UNLOCKED; 808 queue = TS_SHARED_QUEUE; 809 if (v & RW_LOCK_WRITE_WAITERS) { 810 queue = TS_EXCLUSIVE_QUEUE; 811 setv |= (v & RW_LOCK_READ_WAITERS); 812 } 813 setv |= (v & RW_LOCK_WRITE_SPINNER); 814 if (!atomic_fcmpset_rel_ptr(&rw->rw_lock, &v, setv)) 815 continue; 816 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 817 CTR2(KTR_LOCK, "%s: %p last succeeded with waiters", 818 __func__, rw); 819 820 /* 821 * Ok. The lock is released and all that's left is to 822 * wake up the waiters. Note that the lock might not be 823 * free anymore, but in that case the writers will just 824 * block again if they run before the new lock holder(s) 825 * release the lock. 826 */ 827 ts = turnstile_lookup(&rw->lock_object); 828 MPASS(ts != NULL); 829 turnstile_broadcast(ts, queue); 830 turnstile_unpend(ts); 831 td->td_rw_rlocks--; 832 break; 833 } 834 turnstile_chain_unlock(&rw->lock_object); 835 out_lockstat: 836 LOCKSTAT_PROFILE_RELEASE_RWLOCK(rw__release, rw, LOCKSTAT_READER); 837 } 838 839 void 840 _rw_runlock_cookie_int(struct rwlock *rw LOCK_FILE_LINE_ARG_DEF) 841 { 842 struct thread *td; 843 uintptr_t v; 844 845 KASSERT(rw->rw_lock != RW_DESTROYED, 846 ("rw_runlock() of destroyed rwlock @ %s:%d", file, line)); 847 __rw_assert(&rw->rw_lock, RA_RLOCKED, file, line); 848 WITNESS_UNLOCK(&rw->lock_object, 0, file, line); 849 LOCK_LOG_LOCK("RUNLOCK", &rw->lock_object, 0, 0, file, line); 850 851 td = curthread; 852 v = RW_READ_VALUE(rw); 853 854 if (__predict_false(LOCKSTAT_PROFILE_ENABLED(rw__release) || 855 !__rw_runlock_try(rw, td, &v))) 856 __rw_runlock_hard(rw, td, v LOCK_FILE_LINE_ARG); 857 else 858 lock_profile_release_lock(&rw->lock_object); 859 860 TD_LOCKS_DEC(curthread); 861 } 862 863 void 864 _rw_runlock_cookie(volatile uintptr_t *c, const char *file, int line) 865 { 866 struct rwlock *rw; 867 868 rw = rwlock2rw(c); 869 _rw_runlock_cookie_int(rw LOCK_FILE_LINE_ARG); 870 } 871 872 #ifdef ADAPTIVE_RWLOCKS 873 static inline void 874 rw_drop_critical(uintptr_t v, bool *in_critical, int *extra_work) 875 { 876 877 if (v & RW_LOCK_WRITE_SPINNER) 878 return; 879 if (*in_critical) { 880 critical_exit(); 881 *in_critical = false; 882 (*extra_work)--; 883 } 884 } 885 #else 886 #define rw_drop_critical(v, in_critical, extra_work) do { } while (0) 887 #endif 888 889 /* 890 * This function is called when we are unable to obtain a write lock on the 891 * first try. This means that at least one other thread holds either a 892 * read or write lock. 893 */ 894 void 895 __rw_wlock_hard(volatile uintptr_t *c, uintptr_t v LOCK_FILE_LINE_ARG_DEF) 896 { 897 uintptr_t tid; 898 struct rwlock *rw; 899 struct turnstile *ts; 900 struct thread *owner; 901 #ifdef ADAPTIVE_RWLOCKS 902 int spintries = 0; 903 int i, n; 904 enum { READERS, WRITER } sleep_reason = READERS; 905 bool in_critical = false; 906 #endif 907 uintptr_t setv; 908 #ifdef LOCK_PROFILING 909 uint64_t waittime = 0; 910 int contested = 0; 911 #endif 912 #if defined(ADAPTIVE_RWLOCKS) || defined(KDTRACE_HOOKS) 913 struct lock_delay_arg lda; 914 #endif 915 #ifdef KDTRACE_HOOKS 916 u_int sleep_cnt = 0; 917 int64_t sleep_time = 0; 918 int64_t all_time = 0; 919 #endif 920 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING) 921 uintptr_t state = 0; 922 int doing_lockprof = 0; 923 #endif 924 int extra_work = 0; 925 926 tid = (uintptr_t)curthread; 927 rw = rwlock2rw(c); 928 929 #ifdef KDTRACE_HOOKS 930 if (LOCKSTAT_PROFILE_ENABLED(rw__acquire)) { 931 while (v == RW_UNLOCKED) { 932 if (_rw_write_lock_fetch(rw, &v, tid)) 933 goto out_lockstat; 934 } 935 extra_work = 1; 936 doing_lockprof = 1; 937 all_time -= lockstat_nsecs(&rw->lock_object); 938 state = v; 939 } 940 #endif 941 #ifdef LOCK_PROFILING 942 extra_work = 1; 943 doing_lockprof = 1; 944 state = v; 945 #endif 946 947 if (SCHEDULER_STOPPED()) 948 return; 949 950 #if defined(ADAPTIVE_RWLOCKS) 951 lock_delay_arg_init(&lda, &rw_delay); 952 #elif defined(KDTRACE_HOOKS) 953 lock_delay_arg_init(&lda, NULL); 954 #endif 955 if (__predict_false(v == RW_UNLOCKED)) 956 v = RW_READ_VALUE(rw); 957 958 if (__predict_false(lv_rw_wowner(v) == (struct thread *)tid)) { 959 KASSERT(rw->lock_object.lo_flags & LO_RECURSABLE, 960 ("%s: recursing but non-recursive rw %s @ %s:%d\n", 961 __func__, rw->lock_object.lo_name, file, line)); 962 rw->rw_recurse++; 963 atomic_set_ptr(&rw->rw_lock, RW_LOCK_WRITER_RECURSED); 964 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 965 CTR2(KTR_LOCK, "%s: %p recursing", __func__, rw); 966 return; 967 } 968 969 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 970 CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__, 971 rw->lock_object.lo_name, (void *)rw->rw_lock, file, line); 972 973 #ifdef HWPMC_HOOKS 974 PMC_SOFT_CALL( , , lock, failed); 975 #endif 976 lock_profile_obtain_lock_failed(&rw->lock_object, 977 &contested, &waittime); 978 979 for (;;) { 980 if (v == RW_UNLOCKED) { 981 if (_rw_write_lock_fetch(rw, &v, tid)) 982 break; 983 continue; 984 } 985 #ifdef KDTRACE_HOOKS 986 lda.spin_cnt++; 987 #endif 988 989 #ifdef ADAPTIVE_RWLOCKS 990 if (v == (RW_LOCK_READ | RW_LOCK_WRITE_SPINNER)) { 991 if (atomic_fcmpset_acq_ptr(&rw->rw_lock, &v, tid)) 992 break; 993 continue; 994 } 995 996 /* 997 * If the lock is write locked and the owner is 998 * running on another CPU, spin until the owner stops 999 * running or the state of the lock changes. 1000 */ 1001 if (!(v & RW_LOCK_READ)) { 1002 rw_drop_critical(v, &in_critical, &extra_work); 1003 sleep_reason = WRITER; 1004 owner = lv_rw_wowner(v); 1005 if (!TD_IS_RUNNING(owner)) 1006 goto ts; 1007 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 1008 CTR3(KTR_LOCK, "%s: spinning on %p held by %p", 1009 __func__, rw, owner); 1010 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread), 1011 "spinning", "lockname:\"%s\"", 1012 rw->lock_object.lo_name); 1013 do { 1014 lock_delay(&lda); 1015 v = RW_READ_VALUE(rw); 1016 owner = lv_rw_wowner(v); 1017 } while (owner != NULL && TD_IS_RUNNING(owner)); 1018 KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread), 1019 "running"); 1020 continue; 1021 } else if (RW_READERS(v) > 0) { 1022 sleep_reason = READERS; 1023 if (spintries == rowner_retries) 1024 goto ts; 1025 if (!(v & RW_LOCK_WRITE_SPINNER)) { 1026 if (!in_critical) { 1027 critical_enter(); 1028 in_critical = true; 1029 extra_work++; 1030 } 1031 if (!atomic_fcmpset_ptr(&rw->rw_lock, &v, 1032 v | RW_LOCK_WRITE_SPINNER)) { 1033 critical_exit(); 1034 in_critical = false; 1035 extra_work--; 1036 continue; 1037 } 1038 } 1039 spintries++; 1040 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread), 1041 "spinning", "lockname:\"%s\"", 1042 rw->lock_object.lo_name); 1043 n = RW_READERS(v); 1044 for (i = 0; i < rowner_loops; i += n) { 1045 lock_delay_spin(n); 1046 v = RW_READ_VALUE(rw); 1047 if (!(v & RW_LOCK_WRITE_SPINNER)) 1048 break; 1049 if (!(v & RW_LOCK_READ)) 1050 break; 1051 n = RW_READERS(v); 1052 if (n == 0) 1053 break; 1054 } 1055 #ifdef KDTRACE_HOOKS 1056 lda.spin_cnt += i; 1057 #endif 1058 KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread), 1059 "running"); 1060 if (i < rowner_loops) 1061 continue; 1062 } 1063 ts: 1064 #endif 1065 ts = turnstile_trywait(&rw->lock_object); 1066 v = RW_READ_VALUE(rw); 1067 retry_ts: 1068 owner = lv_rw_wowner(v); 1069 1070 #ifdef ADAPTIVE_RWLOCKS 1071 /* 1072 * The current lock owner might have started executing 1073 * on another CPU (or the lock could have changed 1074 * owners) while we were waiting on the turnstile 1075 * chain lock. If so, drop the turnstile lock and try 1076 * again. 1077 */ 1078 if (owner != NULL) { 1079 if (TD_IS_RUNNING(owner)) { 1080 turnstile_cancel(ts); 1081 rw_drop_critical(v, &in_critical, &extra_work); 1082 continue; 1083 } 1084 } else if (RW_READERS(v) > 0 && sleep_reason == WRITER) { 1085 turnstile_cancel(ts); 1086 rw_drop_critical(v, &in_critical, &extra_work); 1087 continue; 1088 } 1089 #endif 1090 /* 1091 * Check for the waiters flags about this rwlock. 1092 * If the lock was released, without maintain any pending 1093 * waiters queue, simply try to acquire it. 1094 * If a pending waiters queue is present, claim the lock 1095 * ownership and maintain the pending queue. 1096 */ 1097 setv = v & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER); 1098 if ((v & ~setv) == RW_UNLOCKED) { 1099 setv &= ~RW_LOCK_WRITE_SPINNER; 1100 if (atomic_fcmpset_acq_ptr(&rw->rw_lock, &v, tid | setv)) { 1101 if (setv) 1102 turnstile_claim(ts); 1103 else 1104 turnstile_cancel(ts); 1105 break; 1106 } 1107 goto retry_ts; 1108 } 1109 1110 #ifdef ADAPTIVE_RWLOCKS 1111 if (in_critical) { 1112 if ((v & RW_LOCK_WRITE_SPINNER) || 1113 !((v & RW_LOCK_WRITE_WAITERS))) { 1114 setv = v & ~RW_LOCK_WRITE_SPINNER; 1115 setv |= RW_LOCK_WRITE_WAITERS; 1116 if (!atomic_fcmpset_ptr(&rw->rw_lock, &v, setv)) 1117 goto retry_ts; 1118 } 1119 critical_exit(); 1120 in_critical = false; 1121 extra_work--; 1122 } else { 1123 #endif 1124 /* 1125 * If the RW_LOCK_WRITE_WAITERS flag isn't set, then try to 1126 * set it. If we fail to set it, then loop back and try 1127 * again. 1128 */ 1129 if (!(v & RW_LOCK_WRITE_WAITERS)) { 1130 if (!atomic_fcmpset_ptr(&rw->rw_lock, &v, 1131 v | RW_LOCK_WRITE_WAITERS)) 1132 goto retry_ts; 1133 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 1134 CTR2(KTR_LOCK, "%s: %p set write waiters flag", 1135 __func__, rw); 1136 } 1137 #ifdef ADAPTIVE_RWLOCKS 1138 } 1139 #endif 1140 /* 1141 * We were unable to acquire the lock and the write waiters 1142 * flag is set, so we must block on the turnstile. 1143 */ 1144 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 1145 CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__, 1146 rw); 1147 #ifdef KDTRACE_HOOKS 1148 sleep_time -= lockstat_nsecs(&rw->lock_object); 1149 #endif 1150 MPASS(owner == rw_owner(rw)); 1151 turnstile_wait(ts, owner, TS_EXCLUSIVE_QUEUE); 1152 #ifdef KDTRACE_HOOKS 1153 sleep_time += lockstat_nsecs(&rw->lock_object); 1154 sleep_cnt++; 1155 #endif 1156 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 1157 CTR2(KTR_LOCK, "%s: %p resuming from turnstile", 1158 __func__, rw); 1159 #ifdef ADAPTIVE_RWLOCKS 1160 spintries = 0; 1161 #endif 1162 v = RW_READ_VALUE(rw); 1163 } 1164 if (__predict_true(!extra_work)) 1165 return; 1166 #ifdef ADAPTIVE_RWLOCKS 1167 if (in_critical) 1168 critical_exit(); 1169 #endif 1170 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING) 1171 if (__predict_true(!doing_lockprof)) 1172 return; 1173 #endif 1174 #ifdef KDTRACE_HOOKS 1175 all_time += lockstat_nsecs(&rw->lock_object); 1176 if (sleep_time) 1177 LOCKSTAT_RECORD4(rw__block, rw, sleep_time, 1178 LOCKSTAT_WRITER, (state & RW_LOCK_READ) == 0, 1179 (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state)); 1180 1181 /* Record only the loops spinning and not sleeping. */ 1182 if (lda.spin_cnt > sleep_cnt) 1183 LOCKSTAT_RECORD4(rw__spin, rw, all_time - sleep_time, 1184 LOCKSTAT_WRITER, (state & RW_LOCK_READ) == 0, 1185 (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state)); 1186 out_lockstat: 1187 #endif 1188 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire, rw, contested, 1189 waittime, file, line, LOCKSTAT_WRITER); 1190 } 1191 1192 /* 1193 * This function is called if lockstat is active or the first try at releasing 1194 * a write lock failed. The latter means that the lock is recursed or one of 1195 * the 2 waiter bits must be set indicating that at least one thread is waiting 1196 * on this lock. 1197 */ 1198 void 1199 __rw_wunlock_hard(volatile uintptr_t *c, uintptr_t v LOCK_FILE_LINE_ARG_DEF) 1200 { 1201 struct rwlock *rw; 1202 struct turnstile *ts; 1203 uintptr_t tid, setv; 1204 int queue; 1205 1206 tid = (uintptr_t)curthread; 1207 if (SCHEDULER_STOPPED()) 1208 return; 1209 1210 rw = rwlock2rw(c); 1211 if (__predict_false(v == tid)) 1212 v = RW_READ_VALUE(rw); 1213 1214 if (v & RW_LOCK_WRITER_RECURSED) { 1215 if (--(rw->rw_recurse) == 0) 1216 atomic_clear_ptr(&rw->rw_lock, RW_LOCK_WRITER_RECURSED); 1217 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 1218 CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, rw); 1219 return; 1220 } 1221 1222 LOCKSTAT_PROFILE_RELEASE_RWLOCK(rw__release, rw, LOCKSTAT_WRITER); 1223 if (v == tid && _rw_write_unlock(rw, tid)) 1224 return; 1225 1226 KASSERT(rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS), 1227 ("%s: neither of the waiter flags are set", __func__)); 1228 1229 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 1230 CTR2(KTR_LOCK, "%s: %p contested", __func__, rw); 1231 1232 turnstile_chain_lock(&rw->lock_object); 1233 1234 /* 1235 * Use the same algo as sx locks for now. Prefer waking up shared 1236 * waiters if we have any over writers. This is probably not ideal. 1237 * 1238 * 'v' is the value we are going to write back to rw_lock. If we 1239 * have waiters on both queues, we need to preserve the state of 1240 * the waiter flag for the queue we don't wake up. For now this is 1241 * hardcoded for the algorithm mentioned above. 1242 * 1243 * In the case of both readers and writers waiting we wakeup the 1244 * readers but leave the RW_LOCK_WRITE_WAITERS flag set. If a 1245 * new writer comes in before a reader it will claim the lock up 1246 * above. There is probably a potential priority inversion in 1247 * there that could be worked around either by waking both queues 1248 * of waiters or doing some complicated lock handoff gymnastics. 1249 */ 1250 setv = RW_UNLOCKED; 1251 v = RW_READ_VALUE(rw); 1252 queue = TS_SHARED_QUEUE; 1253 if (v & RW_LOCK_WRITE_WAITERS) { 1254 queue = TS_EXCLUSIVE_QUEUE; 1255 setv |= (v & RW_LOCK_READ_WAITERS); 1256 } 1257 atomic_store_rel_ptr(&rw->rw_lock, setv); 1258 1259 /* Wake up all waiters for the specific queue. */ 1260 if (LOCK_LOG_TEST(&rw->lock_object, 0)) 1261 CTR3(KTR_LOCK, "%s: %p waking up %s waiters", __func__, rw, 1262 queue == TS_SHARED_QUEUE ? "read" : "write"); 1263 1264 ts = turnstile_lookup(&rw->lock_object); 1265 MPASS(ts != NULL); 1266 turnstile_broadcast(ts, queue); 1267 turnstile_unpend(ts); 1268 turnstile_chain_unlock(&rw->lock_object); 1269 } 1270 1271 /* 1272 * Attempt to do a non-blocking upgrade from a read lock to a write 1273 * lock. This will only succeed if this thread holds a single read 1274 * lock. Returns true if the upgrade succeeded and false otherwise. 1275 */ 1276 int 1277 __rw_try_upgrade_int(struct rwlock *rw LOCK_FILE_LINE_ARG_DEF) 1278 { 1279 uintptr_t v, setv, tid; 1280 struct turnstile *ts; 1281 int success; 1282 1283 if (SCHEDULER_STOPPED()) 1284 return (1); 1285 1286 KASSERT(rw->rw_lock != RW_DESTROYED, 1287 ("rw_try_upgrade() of destroyed rwlock @ %s:%d", file, line)); 1288 __rw_assert(&rw->rw_lock, RA_RLOCKED, file, line); 1289 1290 /* 1291 * Attempt to switch from one reader to a writer. If there 1292 * are any write waiters, then we will have to lock the 1293 * turnstile first to prevent races with another writer 1294 * calling turnstile_wait() before we have claimed this 1295 * turnstile. So, do the simple case of no waiters first. 1296 */ 1297 tid = (uintptr_t)curthread; 1298 success = 0; 1299 v = RW_READ_VALUE(rw); 1300 for (;;) { 1301 if (RW_READERS(v) > 1) 1302 break; 1303 if (!(v & RW_LOCK_WAITERS)) { 1304 success = atomic_fcmpset_acq_ptr(&rw->rw_lock, &v, tid); 1305 if (!success) 1306 continue; 1307 break; 1308 } 1309 1310 /* 1311 * Ok, we think we have waiters, so lock the turnstile. 1312 */ 1313 ts = turnstile_trywait(&rw->lock_object); 1314 v = RW_READ_VALUE(rw); 1315 retry_ts: 1316 if (RW_READERS(v) > 1) { 1317 turnstile_cancel(ts); 1318 break; 1319 } 1320 /* 1321 * Try to switch from one reader to a writer again. This time 1322 * we honor the current state of the waiters flags. 1323 * If we obtain the lock with the flags set, then claim 1324 * ownership of the turnstile. 1325 */ 1326 setv = tid | (v & RW_LOCK_WAITERS); 1327 success = atomic_fcmpset_ptr(&rw->rw_lock, &v, setv); 1328 if (success) { 1329 if (v & RW_LOCK_WAITERS) 1330 turnstile_claim(ts); 1331 else 1332 turnstile_cancel(ts); 1333 break; 1334 } 1335 goto retry_ts; 1336 } 1337 LOCK_LOG_TRY("WUPGRADE", &rw->lock_object, 0, success, file, line); 1338 if (success) { 1339 curthread->td_rw_rlocks--; 1340 WITNESS_UPGRADE(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK, 1341 file, line); 1342 LOCKSTAT_RECORD0(rw__upgrade, rw); 1343 } 1344 return (success); 1345 } 1346 1347 int 1348 __rw_try_upgrade(volatile uintptr_t *c, const char *file, int line) 1349 { 1350 struct rwlock *rw; 1351 1352 rw = rwlock2rw(c); 1353 return (__rw_try_upgrade_int(rw LOCK_FILE_LINE_ARG)); 1354 } 1355 1356 /* 1357 * Downgrade a write lock into a single read lock. 1358 */ 1359 void 1360 __rw_downgrade_int(struct rwlock *rw LOCK_FILE_LINE_ARG_DEF) 1361 { 1362 struct turnstile *ts; 1363 uintptr_t tid, v; 1364 int rwait, wwait; 1365 1366 if (SCHEDULER_STOPPED()) 1367 return; 1368 1369 KASSERT(rw->rw_lock != RW_DESTROYED, 1370 ("rw_downgrade() of destroyed rwlock @ %s:%d", file, line)); 1371 __rw_assert(&rw->rw_lock, RA_WLOCKED | RA_NOTRECURSED, file, line); 1372 #ifndef INVARIANTS 1373 if (rw_recursed(rw)) 1374 panic("downgrade of a recursed lock"); 1375 #endif 1376 1377 WITNESS_DOWNGRADE(&rw->lock_object, 0, file, line); 1378 1379 /* 1380 * Convert from a writer to a single reader. First we handle 1381 * the easy case with no waiters. If there are any waiters, we 1382 * lock the turnstile and "disown" the lock. 1383 */ 1384 tid = (uintptr_t)curthread; 1385 if (atomic_cmpset_rel_ptr(&rw->rw_lock, tid, RW_READERS_LOCK(1))) 1386 goto out; 1387 1388 /* 1389 * Ok, we think we have waiters, so lock the turnstile so we can 1390 * read the waiter flags without any races. 1391 */ 1392 turnstile_chain_lock(&rw->lock_object); 1393 v = rw->rw_lock & RW_LOCK_WAITERS; 1394 rwait = v & RW_LOCK_READ_WAITERS; 1395 wwait = v & RW_LOCK_WRITE_WAITERS; 1396 MPASS(rwait | wwait); 1397 1398 /* 1399 * Downgrade from a write lock while preserving waiters flag 1400 * and give up ownership of the turnstile. 1401 */ 1402 ts = turnstile_lookup(&rw->lock_object); 1403 MPASS(ts != NULL); 1404 if (!wwait) 1405 v &= ~RW_LOCK_READ_WAITERS; 1406 atomic_store_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v); 1407 /* 1408 * Wake other readers if there are no writers pending. Otherwise they 1409 * won't be able to acquire the lock anyway. 1410 */ 1411 if (rwait && !wwait) { 1412 turnstile_broadcast(ts, TS_SHARED_QUEUE); 1413 turnstile_unpend(ts); 1414 } else 1415 turnstile_disown(ts); 1416 turnstile_chain_unlock(&rw->lock_object); 1417 out: 1418 curthread->td_rw_rlocks++; 1419 LOCK_LOG_LOCK("WDOWNGRADE", &rw->lock_object, 0, 0, file, line); 1420 LOCKSTAT_RECORD0(rw__downgrade, rw); 1421 } 1422 1423 void 1424 __rw_downgrade(volatile uintptr_t *c, const char *file, int line) 1425 { 1426 struct rwlock *rw; 1427 1428 rw = rwlock2rw(c); 1429 __rw_downgrade_int(rw LOCK_FILE_LINE_ARG); 1430 } 1431 1432 #ifdef INVARIANT_SUPPORT 1433 #ifndef INVARIANTS 1434 #undef __rw_assert 1435 #endif 1436 1437 /* 1438 * In the non-WITNESS case, rw_assert() can only detect that at least 1439 * *some* thread owns an rlock, but it cannot guarantee that *this* 1440 * thread owns an rlock. 1441 */ 1442 void 1443 __rw_assert(const volatile uintptr_t *c, int what, const char *file, int line) 1444 { 1445 const struct rwlock *rw; 1446 1447 if (SCHEDULER_STOPPED()) 1448 return; 1449 1450 rw = rwlock2rw(c); 1451 1452 switch (what) { 1453 case RA_LOCKED: 1454 case RA_LOCKED | RA_RECURSED: 1455 case RA_LOCKED | RA_NOTRECURSED: 1456 case RA_RLOCKED: 1457 case RA_RLOCKED | RA_RECURSED: 1458 case RA_RLOCKED | RA_NOTRECURSED: 1459 #ifdef WITNESS 1460 witness_assert(&rw->lock_object, what, file, line); 1461 #else 1462 /* 1463 * If some other thread has a write lock or we have one 1464 * and are asserting a read lock, fail. Also, if no one 1465 * has a lock at all, fail. 1466 */ 1467 if (rw->rw_lock == RW_UNLOCKED || 1468 (!(rw->rw_lock & RW_LOCK_READ) && (what & RA_RLOCKED || 1469 rw_wowner(rw) != curthread))) 1470 panic("Lock %s not %slocked @ %s:%d\n", 1471 rw->lock_object.lo_name, (what & RA_RLOCKED) ? 1472 "read " : "", file, line); 1473 1474 if (!(rw->rw_lock & RW_LOCK_READ) && !(what & RA_RLOCKED)) { 1475 if (rw_recursed(rw)) { 1476 if (what & RA_NOTRECURSED) 1477 panic("Lock %s recursed @ %s:%d\n", 1478 rw->lock_object.lo_name, file, 1479 line); 1480 } else if (what & RA_RECURSED) 1481 panic("Lock %s not recursed @ %s:%d\n", 1482 rw->lock_object.lo_name, file, line); 1483 } 1484 #endif 1485 break; 1486 case RA_WLOCKED: 1487 case RA_WLOCKED | RA_RECURSED: 1488 case RA_WLOCKED | RA_NOTRECURSED: 1489 if (rw_wowner(rw) != curthread) 1490 panic("Lock %s not exclusively locked @ %s:%d\n", 1491 rw->lock_object.lo_name, file, line); 1492 if (rw_recursed(rw)) { 1493 if (what & RA_NOTRECURSED) 1494 panic("Lock %s recursed @ %s:%d\n", 1495 rw->lock_object.lo_name, file, line); 1496 } else if (what & RA_RECURSED) 1497 panic("Lock %s not recursed @ %s:%d\n", 1498 rw->lock_object.lo_name, file, line); 1499 break; 1500 case RA_UNLOCKED: 1501 #ifdef WITNESS 1502 witness_assert(&rw->lock_object, what, file, line); 1503 #else 1504 /* 1505 * If we hold a write lock fail. We can't reliably check 1506 * to see if we hold a read lock or not. 1507 */ 1508 if (rw_wowner(rw) == curthread) 1509 panic("Lock %s exclusively locked @ %s:%d\n", 1510 rw->lock_object.lo_name, file, line); 1511 #endif 1512 break; 1513 default: 1514 panic("Unknown rw lock assertion: %d @ %s:%d", what, file, 1515 line); 1516 } 1517 } 1518 #endif /* INVARIANT_SUPPORT */ 1519 1520 #ifdef DDB 1521 void 1522 db_show_rwlock(const struct lock_object *lock) 1523 { 1524 const struct rwlock *rw; 1525 struct thread *td; 1526 1527 rw = (const struct rwlock *)lock; 1528 1529 db_printf(" state: "); 1530 if (rw->rw_lock == RW_UNLOCKED) 1531 db_printf("UNLOCKED\n"); 1532 else if (rw->rw_lock == RW_DESTROYED) { 1533 db_printf("DESTROYED\n"); 1534 return; 1535 } else if (rw->rw_lock & RW_LOCK_READ) 1536 db_printf("RLOCK: %ju locks\n", 1537 (uintmax_t)(RW_READERS(rw->rw_lock))); 1538 else { 1539 td = rw_wowner(rw); 1540 db_printf("WLOCK: %p (tid %d, pid %d, \"%s\")\n", td, 1541 td->td_tid, td->td_proc->p_pid, td->td_name); 1542 if (rw_recursed(rw)) 1543 db_printf(" recursed: %u\n", rw->rw_recurse); 1544 } 1545 db_printf(" waiters: "); 1546 switch (rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS)) { 1547 case RW_LOCK_READ_WAITERS: 1548 db_printf("readers\n"); 1549 break; 1550 case RW_LOCK_WRITE_WAITERS: 1551 db_printf("writers\n"); 1552 break; 1553 case RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS: 1554 db_printf("readers and writers\n"); 1555 break; 1556 default: 1557 db_printf("none\n"); 1558 break; 1559 } 1560 } 1561 1562 #endif 1563