sys/kern/kern_rwlock.c

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