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 = 0;
 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 {
 514                         if ((v & RW_LOCK_WRITE_SPINNER) && RW_READERS(v) == 0) {
 515                                 MPASS(!__rw_can_read(td, v, false));
 516                                 lock_delay_spin(2);
 517                                 v = RW_READ_VALUE(rw);
 518                                 continue;
 519                         }
 520                         if (spintries < rowner_retries) {
 521                                 spintries++;
 522                                 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
 523                                     "spinning", "lockname:\"%s\"",
 524                                     rw->lock_object.lo_name);
 525                                 n = RW_READERS(v);
 526                                 for (i = 0; i < rowner_loops; i += n) {
 527                                         lock_delay_spin(n);
 528                                         v = RW_READ_VALUE(rw);
 529                                         if (!(v & RW_LOCK_READ))
 530                                                 break;
 531                                         n = RW_READERS(v);
 532                                         if (n == 0)
 533                                                 break;
 534                                         if (__rw_can_read(td, v, false))
 535                                                 break;
 536                                 }
 537 #ifdef KDTRACE_HOOKS
 538                                 lda.spin_cnt += rowner_loops - i;
 539 #endif
 540                                 KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
 541                                     "running");
 542                                 if (i < rowner_loops)
 543                                         continue;
 544                         }
 545                 }
 546 #endif
 547
 548                 /*
 549                  * Okay, now it's the hard case.  Some other thread already
 550                  * has a write lock or there are write waiters present,
 551                  * acquire the turnstile lock so we can begin the process
 552                  * of blocking.
 553                  */
 554                 ts = turnstile_trywait(&rw->lock_object);
 555
 556                 /*
 557                  * The lock might have been released while we spun, so
 558                  * recheck its state and restart the loop if needed.
 559                  */
 560                 v = RW_READ_VALUE(rw);
 561 retry_ts:
 562                 if (((v & RW_LOCK_WRITE_SPINNER) && RW_READERS(v) == 0) ||
 563                     __rw_can_read(td, v, false)) {
 564                         turnstile_cancel(ts);
 565                         continue;
 566                 }
 567
 568                 owner = lv_rw_wowner(v);
 569
 570 #ifdef ADAPTIVE_RWLOCKS
 571                 /*
 572                  * The current lock owner might have started executing
 573                  * on another CPU (or the lock could have changed
 574                  * owners) while we were waiting on the turnstile
 575                  * chain lock.  If so, drop the turnstile lock and try
 576                  * again.
 577                  */
 578                 if (owner != NULL) {
 579                         if (TD_IS_RUNNING(owner)) {
 580                                 turnstile_cancel(ts);
 581                                 continue;
 582                         }
 583                 }
 584 #endif
 585
 586                 /*
 587                  * The lock is held in write mode or it already has waiters.
 588                  */
 589                 MPASS(!__rw_can_read(td, v, false));
 590
 591                 /*
 592                  * If the RW_LOCK_READ_WAITERS flag is already set, then
 593                  * we can go ahead and block.  If it is not set then try
 594                  * to set it.  If we fail to set it drop the turnstile
 595                  * lock and restart the loop.
 596                  */
 597                 if (!(v & RW_LOCK_READ_WAITERS)) {
 598                         if (!atomic_fcmpset_ptr(&rw->rw_lock, &v,
 599                             v | RW_LOCK_READ_WAITERS))
 600                                 goto retry_ts;
 601                         if (LOCK_LOG_TEST(&rw->lock_object, 0))
 602                                 CTR2(KTR_LOCK, "%s: %p set read waiters flag",
 603                                     __func__, rw);
 604                 }
 605
 606                 /*
 607                  * We were unable to acquire the lock and the read waiters
 608                  * flag is set, so we must block on the turnstile.
 609                  */
 610                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 611                         CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__,
 612                             rw);
 613 #ifdef KDTRACE_HOOKS
 614                 sleep_time -= lockstat_nsecs(&rw->lock_object);
 615 #endif
 616                 MPASS(owner == rw_owner(rw));
 617                 turnstile_wait(ts, owner, TS_SHARED_QUEUE);
 618 #ifdef KDTRACE_HOOKS
 619                 sleep_time += lockstat_nsecs(&rw->lock_object);
 620                 sleep_cnt++;
 621 #endif
 622                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 623                         CTR2(KTR_LOCK, "%s: %p resuming from turnstile",
 624                             __func__, rw);
 625                 v = RW_READ_VALUE(rw);
 626         }
 627 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
 628         if (__predict_true(!doing_lockprof))
 629                 return;
 630 #endif
 631 #ifdef KDTRACE_HOOKS
 632         all_time += lockstat_nsecs(&rw->lock_object);
 633         if (sleep_time)
 634                 LOCKSTAT_RECORD4(rw__block, rw, sleep_time,
 635                     LOCKSTAT_READER, (state & RW_LOCK_READ) == 0,
 636                     (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
 637
 638         /* Record only the loops spinning and not sleeping. */
 639         if (lda.spin_cnt > sleep_cnt)
 640                 LOCKSTAT_RECORD4(rw__spin, rw, all_time - sleep_time,
 641                     LOCKSTAT_READER, (state & RW_LOCK_READ) == 0,
 642                     (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
 643 out_lockstat:
 644 #endif
 645         /*
 646          * TODO: acquire "owner of record" here.  Here be turnstile dragons
 647          * however.  turnstiles don't like owners changing between calls to
 648          * turnstile_wait() currently.
 649          */
 650         LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire, rw, contested,
 651             waittime, file, line, LOCKSTAT_READER);
 652 }
 653
 654 void
 655 __rw_rlock_int(struct rwlock *rw LOCK_FILE_LINE_ARG_DEF)
 656 {
 657         struct thread *td;
 658         uintptr_t v;
 659
 660         td = curthread;
 661
 662         KASSERT(kdb_active != 0 || SCHEDULER_STOPPED_TD(td) ||
 663             !TD_IS_IDLETHREAD(td),
 664             ("rw_rlock() by idle thread %p on rwlock %s @ %s:%d",
 665             td, rw->lock_object.lo_name, file, line));
 666         KASSERT(rw->rw_lock != RW_DESTROYED,
 667             ("rw_rlock() of destroyed rwlock @ %s:%d", file, line));
 668         KASSERT(rw_wowner(rw) != td,
 669             ("rw_rlock: wlock already held for %s @ %s:%d",
 670             rw->lock_object.lo_name, file, line));
 671         WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER, file, line, NULL);
 672
 673         v = RW_READ_VALUE(rw);
 674         if (__predict_false(LOCKSTAT_PROFILE_ENABLED(rw__acquire) ||
 675             !__rw_rlock_try(rw, td, &v, true LOCK_FILE_LINE_ARG)))
 676                 __rw_rlock_hard(rw, td, v LOCK_FILE_LINE_ARG);
 677         else
 678                 lock_profile_obtain_lock_success(&rw->lock_object, 0, 0,
 679                     file, line);
 680
 681         LOCK_LOG_LOCK("RLOCK", &rw->lock_object, 0, 0, file, line);
 682         WITNESS_LOCK(&rw->lock_object, 0, file, line);
 683         TD_LOCKS_INC(curthread);
 684 }
 685
 686 void
 687 __rw_rlock(volatile uintptr_t *c, const char *file, int line)
 688 {
 689         struct rwlock *rw;
 690
 691         rw = rwlock2rw(c);
 692         __rw_rlock_int(rw LOCK_FILE_LINE_ARG);
 693 }
 694
 695 int
 696 __rw_try_rlock_int(struct rwlock *rw LOCK_FILE_LINE_ARG_DEF)
 697 {
 698         uintptr_t x;
 699
 700         if (SCHEDULER_STOPPED())
 701                 return (1);
 702
 703         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
 704             ("rw_try_rlock() by idle thread %p on rwlock %s @ %s:%d",
 705             curthread, rw->lock_object.lo_name, file, line));
 706
 707         x = rw->rw_lock;
 708         for (;;) {
 709                 KASSERT(rw->rw_lock != RW_DESTROYED,
 710                     ("rw_try_rlock() of destroyed rwlock @ %s:%d", file, line));
 711                 if (!(x & RW_LOCK_READ))
 712                         break;
 713                 if (atomic_fcmpset_acq_ptr(&rw->rw_lock, &x, x + RW_ONE_READER)) {
 714                         LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 1, file,
 715                             line);
 716                         WITNESS_LOCK(&rw->lock_object, LOP_TRYLOCK, file, line);
 717                         LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire,
 718                             rw, 0, 0, file, line, LOCKSTAT_READER);
 719                         TD_LOCKS_INC(curthread);
 720                         curthread->td_rw_rlocks++;
 721                         return (1);
 722                 }
 723         }
 724
 725         LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 0, file, line);
 726         return (0);
 727 }
 728
 729 int
 730 __rw_try_rlock(volatile uintptr_t *c, const char *file, int line)
 731 {
 732         struct rwlock *rw;
 733
 734         rw = rwlock2rw(c);
 735         return (__rw_try_rlock_int(rw LOCK_FILE_LINE_ARG));
 736 }
 737
 738 static bool __always_inline
 739 __rw_runlock_try(struct rwlock *rw, struct thread *td, uintptr_t *vp)
 740 {
 741
 742         for (;;) {
 743                 if (RW_READERS(*vp) > 1 || !(*vp & RW_LOCK_WAITERS)) {
 744                         if (atomic_fcmpset_rel_ptr(&rw->rw_lock, vp,
 745                             *vp - RW_ONE_READER)) {
 746                                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 747                                         CTR4(KTR_LOCK,
 748                                             "%s: %p succeeded %p -> %p",
 749                                             __func__, rw, (void *)*vp,
 750                                             (void *)(*vp - RW_ONE_READER));
 751                                 td->td_rw_rlocks--;
 752                                 return (true);
 753                         }
 754                         continue;
 755                 }
 756                 break;
 757         }
 758         return (false);
 759 }
 760
 761 static void __noinline
 762 __rw_runlock_hard(struct rwlock *rw, struct thread *td, uintptr_t v
 763     LOCK_FILE_LINE_ARG_DEF)
 764 {
 765         struct turnstile *ts;
 766         uintptr_t setv, queue;
 767
 768         if (SCHEDULER_STOPPED())
 769                 return;
 770
 771         if (__rw_runlock_try(rw, td, &v))
 772                 goto out_lockstat;
 773
 774         /*
 775          * Ok, we know we have waiters and we think we are the
 776          * last reader, so grab the turnstile lock.
 777          */
 778         turnstile_chain_lock(&rw->lock_object);
 779         v = RW_READ_VALUE(rw);
 780         for (;;) {
 781                 if (__rw_runlock_try(rw, td, &v))
 782                         break;
 783
 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                 setv |= (v & RW_LOCK_WRITE_SPINNER);
 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);
 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 #ifdef ADAPTIVE_RWLOCKS
 868 static inline void
 869 rw_drop_critical(uintptr_t v, bool *in_critical, int *extra_work)
 870 {
 871
 872         if (v & RW_LOCK_WRITE_SPINNER)
 873                 return;
 874         if (*in_critical) {
 875                 critical_exit();
 876                 *in_critical = false;
 877                 (*extra_work)--;
 878         }
 879 }
 880 #else
 881 #define rw_drop_critical(v, in_critical, extra_work) do { } while (0)
 882 #endif
 883
 884 /*
 885  * This function is called when we are unable to obtain a write lock on the
 886  * first try.  This means that at least one other thread holds either a
 887  * read or write lock.
 888  */
 889 void
 890 __rw_wlock_hard(volatile uintptr_t *c, uintptr_t v LOCK_FILE_LINE_ARG_DEF)
 891 {
 892         uintptr_t tid;
 893         struct rwlock *rw;
 894         struct turnstile *ts;
 895         struct thread *owner;
 896 #ifdef ADAPTIVE_RWLOCKS
 897         int spintries = 0;
 898         int i, n;
 899         enum { READERS, WRITER } sleep_reason = READERS;
 900         bool in_critical = false;
 901 #endif
 902         uintptr_t setv;
 903 #ifdef LOCK_PROFILING
 904         uint64_t waittime = 0;
 905         int contested = 0;
 906 #endif
 907 #if defined(ADAPTIVE_RWLOCKS) || defined(KDTRACE_HOOKS)
 908         struct lock_delay_arg lda;
 909 #endif
 910 #ifdef KDTRACE_HOOKS
 911         u_int sleep_cnt = 0;
 912         int64_t sleep_time = 0;
 913         int64_t all_time = 0;
 914 #endif
 915 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
 916         uintptr_t state = 0;
 917         int doing_lockprof = 0;
 918 #endif
 919         int extra_work = 0;
 920
 921         tid = (uintptr_t)curthread;
 922         rw = rwlock2rw(c);
 923
 924 #ifdef KDTRACE_HOOKS
 925         if (LOCKSTAT_PROFILE_ENABLED(rw__acquire)) {
 926                 while (v == RW_UNLOCKED) {
 927                         if (_rw_write_lock_fetch(rw, &v, tid))
 928                                 goto out_lockstat;
 929                 }
 930                 extra_work = 1;
 931                 doing_lockprof = 1;
 932                 all_time -= lockstat_nsecs(&rw->lock_object);
 933                 state = v;
 934         }
 935 #endif
 936 #ifdef LOCK_PROFILING
 937         extra_work = 1;
 938         doing_lockprof = 1;
 939         state = v;
 940 #endif
 941
 942         if (SCHEDULER_STOPPED())
 943                 return;
 944
 945 #if defined(ADAPTIVE_RWLOCKS)
 946         lock_delay_arg_init(&lda, &rw_delay);
 947 #elif defined(KDTRACE_HOOKS)
 948         lock_delay_arg_init(&lda, NULL);
 949 #endif
 950         if (__predict_false(v == RW_UNLOCKED))
 951                 v = RW_READ_VALUE(rw);
 952
 953         if (__predict_false(lv_rw_wowner(v) == (struct thread *)tid)) {
 954                 KASSERT(rw->lock_object.lo_flags & LO_RECURSABLE,
 955                     ("%s: recursing but non-recursive rw %s @ %s:%d\n",
 956                     __func__, rw->lock_object.lo_name, file, line));
 957                 rw->rw_recurse++;
 958                 atomic_set_ptr(&rw->rw_lock, RW_LOCK_WRITER_RECURSED);
 959                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 960                         CTR2(KTR_LOCK, "%s: %p recursing", __func__, rw);
 961                 return;
 962         }
 963
 964         if (LOCK_LOG_TEST(&rw->lock_object, 0))
 965                 CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__,
 966                     rw->lock_object.lo_name, (void *)rw->rw_lock, file, line);
 967
 968 #ifdef HWPMC_HOOKS
 969         PMC_SOFT_CALL( , , lock, failed);
 970 #endif
 971         lock_profile_obtain_lock_failed(&rw->lock_object,
 972             &contested, &waittime);
 973
 974         for (;;) {
 975                 if (v == RW_UNLOCKED) {
 976                         if (_rw_write_lock_fetch(rw, &v, tid))
 977                                 break;
 978                         continue;
 979                 }
 980 #ifdef KDTRACE_HOOKS
 981                 lda.spin_cnt++;
 982 #endif
 983
 984 #ifdef ADAPTIVE_RWLOCKS
 985                 if (v == (RW_LOCK_READ | RW_LOCK_WRITE_SPINNER)) {
 986                         if (atomic_fcmpset_acq_ptr(&rw->rw_lock, &v, tid))
 987                                 break;
 988                         continue;
 989                 }
 990
 991                 /*
 992                  * If the lock is write locked and the owner is
 993                  * running on another CPU, spin until the owner stops
 994                  * running or the state of the lock changes.
 995                  */
 996                 if (!(v & RW_LOCK_READ)) {
 997                         rw_drop_critical(v, &in_critical, &extra_work);
 998                         sleep_reason = WRITER;
 999                         owner = lv_rw_wowner(v);
1000                         if (!TD_IS_RUNNING(owner))
1001                                 goto ts;
1002                         if (LOCK_LOG_TEST(&rw->lock_object, 0))
1003                                 CTR3(KTR_LOCK, "%s: spinning on %p held by %p",
1004                                     __func__, rw, owner);
1005                         KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
1006                             "spinning", "lockname:\"%s\"",
1007                             rw->lock_object.lo_name);
1008                         do {
1009                                 lock_delay(&lda);
1010                                 v = RW_READ_VALUE(rw);
1011                                 owner = lv_rw_wowner(v);
1012                         } while (owner != NULL && TD_IS_RUNNING(owner));
1013                         KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
1014                             "running");
1015                         continue;
1016                 } else if (RW_READERS(v) > 0) {
1017                         sleep_reason = READERS;
1018                         if (spintries == rowner_retries)
1019                                 goto ts;
1020                         if (!(v & RW_LOCK_WRITE_SPINNER)) {
1021                                 if (!in_critical) {
1022                                         critical_enter();
1023                                         in_critical = true;
1024                                         extra_work++;
1025                                 }
1026                                 if (!atomic_fcmpset_ptr(&rw->rw_lock, &v,
1027                                     v | RW_LOCK_WRITE_SPINNER)) {
1028                                         critical_exit();
1029                                         in_critical = false;
1030                                         extra_work--;
1031                                         continue;
1032                                 }
1033                         }
1034                         spintries++;
1035                         KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
1036                             "spinning", "lockname:\"%s\"",
1037                             rw->lock_object.lo_name);
1038                         n = RW_READERS(v);
1039                         for (i = 0; i < rowner_loops; i += n) {
1040                                 lock_delay_spin(n);
1041                                 v = RW_READ_VALUE(rw);
1042                                 if (!(v & RW_LOCK_WRITE_SPINNER))
1043                                         break;
1044                                 if (!(v & RW_LOCK_READ))
1045                                         break;
1046                                 n = RW_READERS(v);
1047                                 if (n == 0)
1048                                         break;
1049                         }
1050 #ifdef KDTRACE_HOOKS
1051                         lda.spin_cnt += i;
1052 #endif
1053                         KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
1054                             "running");
1055                         if (i < rowner_loops)
1056                                 continue;
1057                 }
1058 ts:
1059 #endif
1060                 ts = turnstile_trywait(&rw->lock_object);
1061                 v = RW_READ_VALUE(rw);
1062 retry_ts:
1063                 owner = lv_rw_wowner(v);
1064
1065 #ifdef ADAPTIVE_RWLOCKS
1066                 /*
1067                  * The current lock owner might have started executing
1068                  * on another CPU (or the lock could have changed
1069                  * owners) while we were waiting on the turnstile
1070                  * chain lock.  If so, drop the turnstile lock and try
1071                  * again.
1072                  */
1073                 if (owner != NULL) {
1074                         if (TD_IS_RUNNING(owner)) {
1075                                 turnstile_cancel(ts);
1076                                 rw_drop_critical(v, &in_critical, &extra_work);
1077                                 continue;
1078                         }
1079                 } else if (RW_READERS(v) > 0 && sleep_reason == WRITER) {
1080                         turnstile_cancel(ts);
1081                         rw_drop_critical(v, &in_critical, &extra_work);
1082                         continue;
1083                 }
1084 #endif
1085                 /*
1086                  * Check for the waiters flags about this rwlock.
1087                  * If the lock was released, without maintain any pending
1088                  * waiters queue, simply try to acquire it.
1089                  * If a pending waiters queue is present, claim the lock
1090                  * ownership and maintain the pending queue.
1091                  */
1092                 setv = v & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER);
1093                 if ((v & ~setv) == RW_UNLOCKED) {
1094                         setv &= ~RW_LOCK_WRITE_SPINNER;
1095                         if (atomic_fcmpset_acq_ptr(&rw->rw_lock, &v, tid | setv)) {
1096                                 if (setv)
1097                                         turnstile_claim(ts);
1098                                 else
1099                                         turnstile_cancel(ts);
1100                                 break;
1101                         }
1102                         goto retry_ts;
1103                 }
1104
1105 #ifdef ADAPTIVE_RWLOCKS
1106                 if (in_critical) {
1107                         if ((v & RW_LOCK_WRITE_SPINNER) ||
1108                             !((v & RW_LOCK_WRITE_WAITERS))) {
1109                                 setv = v & ~RW_LOCK_WRITE_SPINNER;
1110                                 setv |= RW_LOCK_WRITE_WAITERS;
1111                                 if (!atomic_fcmpset_ptr(&rw->rw_lock, &v, setv))
1112                                         goto retry_ts;
1113                         }
1114                         critical_exit();
1115                         in_critical = false;
1116                         extra_work--;
1117                 } else {
1118 #endif
1119                         /*
1120                          * If the RW_LOCK_WRITE_WAITERS flag isn't set, then try to
1121                          * set it.  If we fail to set it, then loop back and try
1122                          * again.
1123                          */
1124                         if (!(v & RW_LOCK_WRITE_WAITERS)) {
1125                                 if (!atomic_fcmpset_ptr(&rw->rw_lock, &v,
1126                                     v | RW_LOCK_WRITE_WAITERS))
1127                                         goto retry_ts;
1128                                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
1129                                         CTR2(KTR_LOCK, "%s: %p set write waiters flag",
1130                                             __func__, rw);
1131                         }
1132 #ifdef ADAPTIVE_RWLOCKS
1133                 }
1134 #endif
1135                 /*
1136                  * We were unable to acquire the lock and the write waiters
1137                  * flag is set, so we must block on the turnstile.
1138                  */
1139                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
1140                         CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__,
1141                             rw);
1142 #ifdef KDTRACE_HOOKS
1143                 sleep_time -= lockstat_nsecs(&rw->lock_object);
1144 #endif
1145                 MPASS(owner == rw_owner(rw));
1146                 turnstile_wait(ts, owner, TS_EXCLUSIVE_QUEUE);
1147 #ifdef KDTRACE_HOOKS
1148                 sleep_time += lockstat_nsecs(&rw->lock_object);
1149                 sleep_cnt++;
1150 #endif
1151                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
1152                         CTR2(KTR_LOCK, "%s: %p resuming from turnstile",
1153                             __func__, rw);
1154 #ifdef ADAPTIVE_RWLOCKS
1155                 spintries = 0;
1156 #endif
1157                 v = RW_READ_VALUE(rw);
1158         }
1159         if (__predict_true(!extra_work))
1160                 return;
1161 #ifdef ADAPTIVE_RWLOCKS
1162         if (in_critical)
1163                 critical_exit();
1164 #endif
1165 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
1166         if (__predict_true(!doing_lockprof))
1167                 return;
1168 #endif
1169 #ifdef KDTRACE_HOOKS
1170         all_time += lockstat_nsecs(&rw->lock_object);
1171         if (sleep_time)
1172                 LOCKSTAT_RECORD4(rw__block, rw, sleep_time,
1173                     LOCKSTAT_WRITER, (state & RW_LOCK_READ) == 0,
1174                     (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
1175
1176         /* Record only the loops spinning and not sleeping. */
1177         if (lda.spin_cnt > sleep_cnt)
1178                 LOCKSTAT_RECORD4(rw__spin, rw, all_time - sleep_time,
1179                     LOCKSTAT_WRITER, (state & RW_LOCK_READ) == 0,
1180                     (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
1181 out_lockstat:
1182 #endif
1183         LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire, rw, contested,
1184             waittime, file, line, LOCKSTAT_WRITER);
1185 }
1186
1187 /*
1188  * This function is called if lockstat is active or the first try at releasing
1189  * a write lock failed.  The latter means that the lock is recursed or one of
1190  * the 2 waiter bits must be set indicating that at least one thread is waiting
1191  * on this lock.
1192  */
1193 void
1194 __rw_wunlock_hard(volatile uintptr_t *c, uintptr_t v LOCK_FILE_LINE_ARG_DEF)
1195 {
1196         struct rwlock *rw;
1197         struct turnstile *ts;
1198         uintptr_t tid, setv;
1199         int queue;
1200
1201         tid = (uintptr_t)curthread;
1202         if (SCHEDULER_STOPPED())
1203                 return;
1204
1205         rw = rwlock2rw(c);
1206         if (__predict_false(v == tid))
1207                 v = RW_READ_VALUE(rw);
1208
1209         if (v & RW_LOCK_WRITER_RECURSED) {
1210                 if (--(rw->rw_recurse) == 0)
1211                         atomic_clear_ptr(&rw->rw_lock, RW_LOCK_WRITER_RECURSED);
1212                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
1213                         CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, rw);
1214                 return;
1215         }
1216
1217         LOCKSTAT_PROFILE_RELEASE_RWLOCK(rw__release, rw, LOCKSTAT_WRITER);
1218         if (v == tid && _rw_write_unlock(rw, tid))
1219                 return;
1220
1221         KASSERT(rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS),
1222             ("%s: neither of the waiter flags are set", __func__));
1223
1224         if (LOCK_LOG_TEST(&rw->lock_object, 0))
1225                 CTR2(KTR_LOCK, "%s: %p contested", __func__, rw);
1226
1227         turnstile_chain_lock(&rw->lock_object);
1228
1229         /*
1230          * Use the same algo as sx locks for now.  Prefer waking up shared
1231          * waiters if we have any over writers.  This is probably not ideal.
1232          *
1233          * 'v' is the value we are going to write back to rw_lock.  If we
1234          * have waiters on both queues, we need to preserve the state of
1235          * the waiter flag for the queue we don't wake up.  For now this is
1236          * hardcoded for the algorithm mentioned above.
1237          *
1238          * In the case of both readers and writers waiting we wakeup the
1239          * readers but leave the RW_LOCK_WRITE_WAITERS flag set.  If a
1240          * new writer comes in before a reader it will claim the lock up
1241          * above.  There is probably a potential priority inversion in
1242          * there that could be worked around either by waking both queues
1243          * of waiters or doing some complicated lock handoff gymnastics.
1244          */
1245         setv = RW_UNLOCKED;
1246         v = RW_READ_VALUE(rw);
1247         queue = TS_SHARED_QUEUE;
1248         if (v & RW_LOCK_WRITE_WAITERS) {
1249                 queue = TS_EXCLUSIVE_QUEUE;
1250                 setv |= (v & RW_LOCK_READ_WAITERS);
1251         }
1252         atomic_store_rel_ptr(&rw->rw_lock, setv);
1253
1254         /* Wake up all waiters for the specific queue. */
1255         if (LOCK_LOG_TEST(&rw->lock_object, 0))
1256                 CTR3(KTR_LOCK, "%s: %p waking up %s waiters", __func__, rw,
1257                     queue == TS_SHARED_QUEUE ? "read" : "write");
1258
1259         ts = turnstile_lookup(&rw->lock_object);
1260         MPASS(ts != NULL);
1261         turnstile_broadcast(ts, queue);
1262         turnstile_unpend(ts);
1263         turnstile_chain_unlock(&rw->lock_object);
1264 }
1265
1266 /*
1267  * Attempt to do a non-blocking upgrade from a read lock to a write
1268  * lock.  This will only succeed if this thread holds a single read
1269  * lock.  Returns true if the upgrade succeeded and false otherwise.
1270  */
1271 int
1272 __rw_try_upgrade_int(struct rwlock *rw LOCK_FILE_LINE_ARG_DEF)
1273 {
1274         uintptr_t v, setv, tid;
1275         struct turnstile *ts;
1276         int success;
1277
1278         if (SCHEDULER_STOPPED())
1279                 return (1);
1280
1281         KASSERT(rw->rw_lock != RW_DESTROYED,
1282             ("rw_try_upgrade() of destroyed rwlock @ %s:%d", file, line));
1283         __rw_assert(&rw->rw_lock, RA_RLOCKED, file, line);
1284
1285         /*
1286          * Attempt to switch from one reader to a writer.  If there
1287          * are any write waiters, then we will have to lock the
1288          * turnstile first to prevent races with another writer
1289          * calling turnstile_wait() before we have claimed this
1290          * turnstile.  So, do the simple case of no waiters first.
1291          */
1292         tid = (uintptr_t)curthread;
1293         success = 0;
1294         v = RW_READ_VALUE(rw);
1295         for (;;) {
1296                 if (RW_READERS(v) > 1)
1297                         break;
1298                 if (!(v & RW_LOCK_WAITERS)) {
1299                         success = atomic_fcmpset_acq_ptr(&rw->rw_lock, &v, tid);
1300                         if (!success)
1301                                 continue;
1302                         break;
1303                 }
1304
1305                 /*
1306                  * Ok, we think we have waiters, so lock the turnstile.
1307                  */
1308                 ts = turnstile_trywait(&rw->lock_object);
1309                 v = RW_READ_VALUE(rw);
1310 retry_ts:
1311                 if (RW_READERS(v) > 1) {
1312                         turnstile_cancel(ts);
1313                         break;
1314                 }
1315                 /*
1316                  * Try to switch from one reader to a writer again.  This time
1317                  * we honor the current state of the waiters flags.
1318                  * If we obtain the lock with the flags set, then claim
1319                  * ownership of the turnstile.
1320                  */
1321                 setv = tid | (v & RW_LOCK_WAITERS);
1322                 success = atomic_fcmpset_ptr(&rw->rw_lock, &v, setv);
1323                 if (success) {
1324                         if (v & RW_LOCK_WAITERS)
1325                                 turnstile_claim(ts);
1326                         else
1327                                 turnstile_cancel(ts);
1328                         break;
1329                 }
1330                 goto retry_ts;
1331         }
1332         LOCK_LOG_TRY("WUPGRADE", &rw->lock_object, 0, success, file, line);
1333         if (success) {
1334                 curthread->td_rw_rlocks--;
1335                 WITNESS_UPGRADE(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
1336                     file, line);
1337                 LOCKSTAT_RECORD0(rw__upgrade, rw);
1338         }
1339         return (success);
1340 }
1341
1342 int
1343 __rw_try_upgrade(volatile uintptr_t *c, const char *file, int line)
1344 {
1345         struct rwlock *rw;
1346
1347         rw = rwlock2rw(c);
1348         return (__rw_try_upgrade_int(rw LOCK_FILE_LINE_ARG));
1349 }
1350
1351 /*
1352  * Downgrade a write lock into a single read lock.
1353  */
1354 void
1355 __rw_downgrade_int(struct rwlock *rw LOCK_FILE_LINE_ARG_DEF)
1356 {
1357         struct turnstile *ts;
1358         uintptr_t tid, v;
1359         int rwait, wwait;
1360
1361         if (SCHEDULER_STOPPED())
1362                 return;
1363
1364         KASSERT(rw->rw_lock != RW_DESTROYED,
1365             ("rw_downgrade() of destroyed rwlock @ %s:%d", file, line));
1366         __rw_assert(&rw->rw_lock, RA_WLOCKED | RA_NOTRECURSED, file, line);
1367 #ifndef INVARIANTS
1368         if (rw_recursed(rw))
1369                 panic("downgrade of a recursed lock");
1370 #endif
1371
1372         WITNESS_DOWNGRADE(&rw->lock_object, 0, file, line);
1373
1374         /*
1375          * Convert from a writer to a single reader.  First we handle
1376          * the easy case with no waiters.  If there are any waiters, we
1377          * lock the turnstile and "disown" the lock.
1378          */
1379         tid = (uintptr_t)curthread;
1380         if (atomic_cmpset_rel_ptr(&rw->rw_lock, tid, RW_READERS_LOCK(1)))
1381                 goto out;
1382
1383         /*
1384          * Ok, we think we have waiters, so lock the turnstile so we can
1385          * read the waiter flags without any races.
1386          */
1387         turnstile_chain_lock(&rw->lock_object);
1388         v = rw->rw_lock & RW_LOCK_WAITERS;
1389         rwait = v & RW_LOCK_READ_WAITERS;
1390         wwait = v & RW_LOCK_WRITE_WAITERS;
1391         MPASS(rwait | wwait);
1392
1393         /*
1394          * Downgrade from a write lock while preserving waiters flag
1395          * and give up ownership of the turnstile.
1396          */
1397         ts = turnstile_lookup(&rw->lock_object);
1398         MPASS(ts != NULL);
1399         if (!wwait)
1400                 v &= ~RW_LOCK_READ_WAITERS;
1401         atomic_store_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v);
1402         /*
1403          * Wake other readers if there are no writers pending.  Otherwise they
1404          * won't be able to acquire the lock anyway.
1405          */
1406         if (rwait && !wwait) {
1407                 turnstile_broadcast(ts, TS_SHARED_QUEUE);
1408                 turnstile_unpend(ts);
1409         } else
1410                 turnstile_disown(ts);
1411         turnstile_chain_unlock(&rw->lock_object);
1412 out:
1413         curthread->td_rw_rlocks++;
1414         LOCK_LOG_LOCK("WDOWNGRADE", &rw->lock_object, 0, 0, file, line);
1415         LOCKSTAT_RECORD0(rw__downgrade, rw);
1416 }
1417
1418 void
1419 __rw_downgrade(volatile uintptr_t *c, const char *file, int line)
1420 {
1421         struct rwlock *rw;
1422
1423         rw = rwlock2rw(c);
1424         __rw_downgrade_int(rw LOCK_FILE_LINE_ARG);
1425 }
1426
1427 #ifdef INVARIANT_SUPPORT
1428 #ifndef INVARIANTS
1429 #undef __rw_assert
1430 #endif
1431
1432 /*
1433  * In the non-WITNESS case, rw_assert() can only detect that at least
1434  * *some* thread owns an rlock, but it cannot guarantee that *this*
1435  * thread owns an rlock.
1436  */
1437 void
1438 __rw_assert(const volatile uintptr_t *c, int what, const char *file, int line)
1439 {
1440         const struct rwlock *rw;
1441
1442         if (SCHEDULER_STOPPED())
1443                 return;
1444
1445         rw = rwlock2rw(c);
1446
1447         switch (what) {
1448         case RA_LOCKED:
1449         case RA_LOCKED | RA_RECURSED:
1450         case RA_LOCKED | RA_NOTRECURSED:
1451         case RA_RLOCKED:
1452         case RA_RLOCKED | RA_RECURSED:
1453         case RA_RLOCKED | RA_NOTRECURSED:
1454 #ifdef WITNESS
1455                 witness_assert(&rw->lock_object, what, file, line);
1456 #else
1457                 /*
1458                  * If some other thread has a write lock or we have one
1459                  * and are asserting a read lock, fail.  Also, if no one
1460                  * has a lock at all, fail.
1461                  */
1462                 if (rw->rw_lock == RW_UNLOCKED ||
1463                     (!(rw->rw_lock & RW_LOCK_READ) && (what & RA_RLOCKED ||
1464                     rw_wowner(rw) != curthread)))
1465                         panic("Lock %s not %slocked @ %s:%d\n",
1466                             rw->lock_object.lo_name, (what & RA_RLOCKED) ?
1467                             "read " : "", file, line);
1468
1469                 if (!(rw->rw_lock & RW_LOCK_READ) && !(what & RA_RLOCKED)) {
1470                         if (rw_recursed(rw)) {
1471                                 if (what & RA_NOTRECURSED)
1472                                         panic("Lock %s recursed @ %s:%d\n",
1473                                             rw->lock_object.lo_name, file,
1474                                             line);
1475                         } else if (what & RA_RECURSED)
1476                                 panic("Lock %s not recursed @ %s:%d\n",
1477                                     rw->lock_object.lo_name, file, line);
1478                 }
1479 #endif
1480                 break;
1481         case RA_WLOCKED:
1482         case RA_WLOCKED | RA_RECURSED:
1483         case RA_WLOCKED | RA_NOTRECURSED:
1484                 if (rw_wowner(rw) != curthread)
1485                         panic("Lock %s not exclusively locked @ %s:%d\n",
1486                             rw->lock_object.lo_name, file, line);
1487                 if (rw_recursed(rw)) {
1488                         if (what & RA_NOTRECURSED)
1489                                 panic("Lock %s recursed @ %s:%d\n",
1490                                     rw->lock_object.lo_name, file, line);
1491                 } else if (what & RA_RECURSED)
1492                         panic("Lock %s not recursed @ %s:%d\n",
1493                             rw->lock_object.lo_name, file, line);
1494                 break;
1495         case RA_UNLOCKED:
1496 #ifdef WITNESS
1497                 witness_assert(&rw->lock_object, what, file, line);
1498 #else
1499                 /*
1500                  * If we hold a write lock fail.  We can't reliably check
1501                  * to see if we hold a read lock or not.
1502                  */
1503                 if (rw_wowner(rw) == curthread)
1504                         panic("Lock %s exclusively locked @ %s:%d\n",
1505                             rw->lock_object.lo_name, file, line);
1506 #endif
1507                 break;
1508         default:
1509                 panic("Unknown rw lock assertion: %d @ %s:%d", what, file,
1510                     line);
1511         }
1512 }
1513 #endif /* INVARIANT_SUPPORT */
1514
1515 #ifdef DDB
1516 void
1517 db_show_rwlock(const struct lock_object *lock)
1518 {
1519         const struct rwlock *rw;
1520         struct thread *td;
1521
1522         rw = (const struct rwlock *)lock;
1523
1524         db_printf(" state: ");
1525         if (rw->rw_lock == RW_UNLOCKED)
1526                 db_printf("UNLOCKED\n");
1527         else if (rw->rw_lock == RW_DESTROYED) {
1528                 db_printf("DESTROYED\n");
1529                 return;
1530         } else if (rw->rw_lock & RW_LOCK_READ)
1531                 db_printf("RLOCK: %ju locks\n",
1532                     (uintmax_t)(RW_READERS(rw->rw_lock)));
1533         else {
1534                 td = rw_wowner(rw);
1535                 db_printf("WLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1536                     td->td_tid, td->td_proc->p_pid, td->td_name);
1537                 if (rw_recursed(rw))
1538                         db_printf(" recursed: %u\n", rw->rw_recurse);
1539         }
1540         db_printf(" waiters: ");
1541         switch (rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS)) {
1542         case RW_LOCK_READ_WAITERS:
1543                 db_printf("readers\n");
1544                 break;
1545         case RW_LOCK_WRITE_WAITERS:
1546                 db_printf("writers\n");
1547                 break;
1548         case RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS:
1549                 db_printf("readers and writers\n");
1550                 break;
1551         default:
1552                 db_printf("none\n");
1553                 break;
1554         }
1555 }
1556
1557 #endif