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