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