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