sys/kern/kern_rwlock.c

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