sys/kern/kern_rwlock.c

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