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