sys/kern/kern_rwlock.c

   1 /*-
   2  * SPDX-License-Identifier: BSD-2-Clause
   3  *
   4  * Copyright (c) 2006 John Baldwin <jhb@FreeBSD.org>
   5  *
   6  * Redistribution and use in source and binary forms, with or without
   7  * modification, are permitted provided that the following conditions
   8  * are met:
   9  * 1. Redistributions of source code must retain the above copyright
  10  *    notice, this list of conditions and the following disclaimer.
  11  * 2. Redistributions in binary form must reproduce the above copyright
  12  *    notice, this list of conditions and the following disclaimer in the
  13  *    documentation and/or other materials provided with the distribution.
  14  *
  15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
  16  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  18  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
  19  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  20  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  21  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  22  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  23  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  24  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  25  * SUCH DAMAGE.
  26  */
  27
  28 /*
  29  * Machine independent bits of reader/writer lock implementation.
  30  */
  31
  32 #include <sys/cdefs.h>
  33 #include "opt_ddb.h"
  34 #include "opt_hwpmc_hooks.h"
  35 #include "opt_no_adaptive_rwlocks.h"
  36
  37 #include <sys/param.h>
  38 #include <sys/kdb.h>
  39 #include <sys/ktr.h>
  40 #include <sys/kernel.h>
  41 #include <sys/lock.h>
  42 #include <sys/mutex.h>
  43 #include <sys/proc.h>
  44 #include <sys/rwlock.h>
  45 #include <sys/sched.h>
  46 #include <sys/smp.h>
  47 #include <sys/sysctl.h>
  48 #include <sys/systm.h>
  49 #include <sys/turnstile.h>
  50
  51 #include <machine/cpu.h>
  52
  53 #if defined(SMP) && !defined(NO_ADAPTIVE_RWLOCKS)
  54 #define ADAPTIVE_RWLOCKS
  55 #endif
  56
  57 #ifdef HWPMC_HOOKS
  58 #include <sys/pmckern.h>
  59 PMC_SOFT_DECLARE( , , lock, failed);
  60 #endif
  61
  62 /*
  63  * Return the rwlock address when the lock cookie address is provided.
  64  * This functionality assumes that struct rwlock* have a member named rw_lock.
  65  */
  66 #define rwlock2rw(c)    (__containerof(c, struct rwlock, rw_lock))
  67
  68 #ifdef DDB
  69 #include <ddb/ddb.h>
  70
  71 static void     db_show_rwlock(const struct lock_object *lock);
  72 #endif
  73 static void     assert_rw(const struct lock_object *lock, int what);
  74 static void     lock_rw(struct lock_object *lock, uintptr_t how);
  75 #ifdef KDTRACE_HOOKS
  76 static int      owner_rw(const struct lock_object *lock, struct thread **owner);
  77 #endif
  78 static uintptr_t unlock_rw(struct lock_object *lock);
  79
  80 struct lock_class lock_class_rw = {
  81         .lc_name = "rw",
  82         .lc_flags = LC_SLEEPLOCK | LC_RECURSABLE | LC_UPGRADABLE,
  83         .lc_assert = assert_rw,
  84 #ifdef DDB
  85         .lc_ddb_show = db_show_rwlock,
  86 #endif
  87         .lc_lock = lock_rw,
  88         .lc_unlock = unlock_rw,
  89 #ifdef KDTRACE_HOOKS
  90         .lc_owner = owner_rw,
  91 #endif
  92 };
  93
  94 #ifdef ADAPTIVE_RWLOCKS
  95 #ifdef RWLOCK_CUSTOM_BACKOFF
  96 static u_short __read_frequently rowner_retries;
  97 static u_short __read_frequently rowner_loops;
  98 static SYSCTL_NODE(_debug, OID_AUTO, rwlock,
  99     CTLFLAG_RD | CTLFLAG_MPSAFE, NULL,
 100     "rwlock debugging");
 101 SYSCTL_U16(_debug_rwlock, OID_AUTO, retry, CTLFLAG_RW, &rowner_retries, 0, "");
 102 SYSCTL_U16(_debug_rwlock, OID_AUTO, loops, CTLFLAG_RW, &rowner_loops, 0, "");
 103
 104 static struct lock_delay_config __read_frequently rw_delay;
 105
 106 SYSCTL_U16(_debug_rwlock, OID_AUTO, delay_base, CTLFLAG_RW, &rw_delay.base,
 107     0, "");
 108 SYSCTL_U16(_debug_rwlock, OID_AUTO, delay_max, CTLFLAG_RW, &rw_delay.max,
 109     0, "");
 110
 111 static void
 112 rw_lock_delay_init(void *arg __unused)
 113 {
 114
 115         lock_delay_default_init(&rw_delay);
 116         rowner_retries = 10;
 117         rowner_loops = max(10000, rw_delay.max);
 118 }
 119 LOCK_DELAY_SYSINIT(rw_lock_delay_init);
 120 #else
 121 #define rw_delay        locks_delay
 122 #define rowner_retries  locks_delay_retries
 123 #define rowner_loops    locks_delay_loops
 124 #endif
 125 #endif
 126
 127 /*
 128  * Return a pointer to the owning thread if the lock is write-locked or
 129  * NULL if the lock is unlocked or read-locked.
 130  */
 131
 132 #define lv_rw_wowner(v)                                                 \
 133         ((v) & RW_LOCK_READ ? NULL :                                    \
 134          (struct thread *)RW_OWNER((v)))
 135
 136 #define rw_wowner(rw)   lv_rw_wowner(RW_READ_VALUE(rw))
 137
 138 /*
 139  * Returns if a write owner is recursed.  Write ownership is not assured
 140  * here and should be previously checked.
 141  */
 142 #define rw_recursed(rw)         ((rw)->rw_recurse != 0)
 143
 144 /*
 145  * Return true if curthread helds the lock.
 146  */
 147 #define rw_wlocked(rw)          (rw_wowner((rw)) == curthread)
 148
 149 /*
 150  * Return a pointer to the owning thread for this lock who should receive
 151  * any priority lent by threads that block on this lock.  Currently this
 152  * is identical to rw_wowner().
 153  */
 154 #define rw_owner(rw)            rw_wowner(rw)
 155
 156 #ifndef INVARIANTS
 157 #define __rw_assert(c, what, file, line)
 158 #endif
 159
 160 void
 161 assert_rw(const struct lock_object *lock, int what)
 162 {
 163
 164         rw_assert((const struct rwlock *)lock, what);
 165 }
 166
 167 void
 168 lock_rw(struct lock_object *lock, uintptr_t how)
 169 {
 170         struct rwlock *rw;
 171
 172         rw = (struct rwlock *)lock;
 173         if (how)
 174                 rw_rlock(rw);
 175         else
 176                 rw_wlock(rw);
 177 }
 178
 179 uintptr_t
 180 unlock_rw(struct lock_object *lock)
 181 {
 182         struct rwlock *rw;
 183
 184         rw = (struct rwlock *)lock;
 185         rw_assert(rw, RA_LOCKED | LA_NOTRECURSED);
 186         if (rw->rw_lock & RW_LOCK_READ) {
 187                 rw_runlock(rw);
 188                 return (1);
 189         } else {
 190                 rw_wunlock(rw);
 191                 return (0);
 192         }
 193 }
 194
 195 #ifdef KDTRACE_HOOKS
 196 int
 197 owner_rw(const struct lock_object *lock, struct thread **owner)
 198 {
 199         const struct rwlock *rw = (const struct rwlock *)lock;
 200         uintptr_t x = rw->rw_lock;
 201
 202         *owner = rw_wowner(rw);
 203         return ((x & RW_LOCK_READ) != 0 ?  (RW_READERS(x) != 0) :
 204             (*owner != NULL));
 205 }
 206 #endif
 207
 208 void
 209 _rw_init_flags(volatile uintptr_t *c, const char *name, int opts)
 210 {
 211         struct rwlock *rw;
 212         int flags;
 213
 214         rw = rwlock2rw(c);
 215
 216         MPASS((opts & ~(RW_DUPOK | RW_NOPROFILE | RW_NOWITNESS | RW_QUIET |
 217             RW_RECURSE | RW_NEW)) == 0);
 218         ASSERT_ATOMIC_LOAD_PTR(rw->rw_lock,
 219             ("%s: rw_lock not aligned for %s: %p", __func__, name,
 220             &rw->rw_lock));
 221
 222         flags = LO_UPGRADABLE;
 223         if (opts & RW_DUPOK)
 224                 flags |= LO_DUPOK;
 225         if (opts & RW_NOPROFILE)
 226                 flags |= LO_NOPROFILE;
 227         if (!(opts & RW_NOWITNESS))
 228                 flags |= LO_WITNESS;
 229         if (opts & RW_RECURSE)
 230                 flags |= LO_RECURSABLE;
 231         if (opts & RW_QUIET)
 232                 flags |= LO_QUIET;
 233         if (opts & RW_NEW)
 234                 flags |= LO_NEW;
 235
 236         lock_init(&rw->lock_object, &lock_class_rw, name, NULL, flags);
 237         rw->rw_lock = RW_UNLOCKED;
 238         rw->rw_recurse = 0;
 239 }
 240
 241 void
 242 _rw_destroy(volatile uintptr_t *c)
 243 {
 244         struct rwlock *rw;
 245
 246         rw = rwlock2rw(c);
 247
 248         KASSERT(rw->rw_lock == RW_UNLOCKED, ("rw lock %p not unlocked", rw));
 249         KASSERT(rw->rw_recurse == 0, ("rw lock %p still recursed", rw));
 250         rw->rw_lock = RW_DESTROYED;
 251         lock_destroy(&rw->lock_object);
 252 }
 253
 254 void
 255 rw_sysinit(void *arg)
 256 {
 257         struct rw_args *args;
 258
 259         args = arg;
 260         rw_init_flags((struct rwlock *)args->ra_rw, args->ra_desc,
 261             args->ra_flags);
 262 }
 263
 264 int
 265 _rw_wowned(const volatile uintptr_t *c)
 266 {
 267
 268         return (rw_wowner(rwlock2rw(c)) == curthread);
 269 }
 270
 271 void
 272 _rw_wlock_cookie(volatile uintptr_t *c, const char *file, int line)
 273 {
 274         struct rwlock *rw;
 275         uintptr_t tid, v;
 276
 277         rw = rwlock2rw(c);
 278
 279         KASSERT(kdb_active != 0 || SCHEDULER_STOPPED() ||
 280             !TD_IS_IDLETHREAD(curthread),
 281             ("rw_wlock() by idle thread %p on rwlock %s @ %s:%d",
 282             curthread, rw->lock_object.lo_name, file, line));
 283         KASSERT(rw->rw_lock != RW_DESTROYED,
 284             ("rw_wlock() of destroyed rwlock @ %s:%d", file, line));
 285         WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE, file,
 286             line, NULL);
 287         tid = (uintptr_t)curthread;
 288         v = RW_UNLOCKED;
 289         if (!_rw_write_lock_fetch(rw, &v, tid))
 290                 _rw_wlock_hard(rw, v, file, line);
 291         else
 292                 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire, rw,
 293                     0, 0, file, line, LOCKSTAT_WRITER);
 294
 295         LOCK_LOG_LOCK("WLOCK", &rw->lock_object, 0, rw->rw_recurse, file, line);
 296         WITNESS_LOCK(&rw->lock_object, LOP_EXCLUSIVE, file, line);
 297         TD_LOCKS_INC(curthread);
 298 }
 299
 300 int
 301 __rw_try_wlock_int(struct rwlock *rw LOCK_FILE_LINE_ARG_DEF)
 302 {
 303         struct thread *td;
 304         uintptr_t tid, v;
 305         int rval;
 306         bool recursed;
 307
 308         td = curthread;
 309         tid = (uintptr_t)td;
 310         if (SCHEDULER_STOPPED())
 311                 return (1);
 312
 313         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(td),
 314             ("rw_try_wlock() by idle thread %p on rwlock %s @ %s:%d",
 315             curthread, rw->lock_object.lo_name, file, line));
 316         KASSERT(rw->rw_lock != RW_DESTROYED,
 317             ("rw_try_wlock() of destroyed rwlock @ %s:%d", file, line));
 318
 319         rval = 1;
 320         recursed = false;
 321         v = RW_UNLOCKED;
 322         for (;;) {
 323                 if (atomic_fcmpset_acq_ptr(&rw->rw_lock, &v, tid))
 324                         break;
 325                 if (v == RW_UNLOCKED)
 326                         continue;
 327                 if (v == tid && (rw->lock_object.lo_flags & LO_RECURSABLE)) {
 328                         rw->rw_recurse++;
 329                         atomic_set_ptr(&rw->rw_lock, RW_LOCK_WRITER_RECURSED);
 330                         break;
 331                 }
 332                 rval = 0;
 333                 break;
 334         }
 335
 336         LOCK_LOG_TRY("WLOCK", &rw->lock_object, 0, rval, file, line);
 337         if (rval) {
 338                 WITNESS_LOCK(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
 339                     file, line);
 340                 if (!recursed)
 341                         LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire,
 342                             rw, 0, 0, file, line, LOCKSTAT_WRITER);
 343                 TD_LOCKS_INC(curthread);
 344         }
 345         return (rval);
 346 }
 347
 348 int
 349 __rw_try_wlock(volatile uintptr_t *c, const char *file, int line)
 350 {
 351         struct rwlock *rw;
 352
 353         rw = rwlock2rw(c);
 354         return (__rw_try_wlock_int(rw LOCK_FILE_LINE_ARG));
 355 }
 356
 357 void
 358 _rw_wunlock_cookie(volatile uintptr_t *c, const char *file, int line)
 359 {
 360         struct rwlock *rw;
 361
 362         rw = rwlock2rw(c);
 363
 364         KASSERT(rw->rw_lock != RW_DESTROYED,
 365             ("rw_wunlock() of destroyed rwlock @ %s:%d", file, line));
 366         __rw_assert(c, RA_WLOCKED, file, line);
 367         WITNESS_UNLOCK(&rw->lock_object, LOP_EXCLUSIVE, file, line);
 368         LOCK_LOG_LOCK("WUNLOCK", &rw->lock_object, 0, rw->rw_recurse, file,
 369             line);
 370
 371 #ifdef LOCK_PROFILING
 372         _rw_wunlock_hard(rw, (uintptr_t)curthread, file, line);
 373 #else
 374         __rw_wunlock(rw, curthread, file, line);
 375 #endif
 376
 377         TD_LOCKS_DEC(curthread);
 378 }
 379
 380 /*
 381  * Determines whether a new reader can acquire a lock.  Succeeds if the
 382  * reader already owns a read lock and the lock is locked for read to
 383  * prevent deadlock from reader recursion.  Also succeeds if the lock
 384  * is unlocked and has no writer waiters or spinners.  Failing otherwise
 385  * prioritizes writers before readers.
 386  */
 387 static bool __always_inline
 388 __rw_can_read(struct thread *td, uintptr_t v, bool fp)
 389 {
 390
 391         if ((v & (RW_LOCK_READ | RW_LOCK_WRITE_WAITERS | RW_LOCK_WRITE_SPINNER))
 392             == RW_LOCK_READ)
 393                 return (true);
 394         if (!fp && td->td_rw_rlocks && (v & RW_LOCK_READ))
 395                 return (true);
 396         return (false);
 397 }
 398
 399 static bool __always_inline
 400 __rw_rlock_try(struct rwlock *rw, struct thread *td, uintptr_t *vp, bool fp
 401     LOCK_FILE_LINE_ARG_DEF)
 402 {
 403
 404         /*
 405          * Handle the easy case.  If no other thread has a write
 406          * lock, then try to bump up the count of read locks.  Note
 407          * that we have to preserve the current state of the
 408          * RW_LOCK_WRITE_WAITERS flag.  If we fail to acquire a
 409          * read lock, then rw_lock must have changed, so restart
 410          * the loop.  Note that this handles the case of a
 411          * completely unlocked rwlock since such a lock is encoded
 412          * as a read lock with no waiters.
 413          */
 414         while (__rw_can_read(td, *vp, fp)) {
 415                 if (atomic_fcmpset_acq_ptr(&rw->rw_lock, vp,
 416                         *vp + RW_ONE_READER)) {
 417                         if (LOCK_LOG_TEST(&rw->lock_object, 0))
 418                                 CTR4(KTR_LOCK,
 419                                     "%s: %p succeed %p -> %p", __func__,
 420                                     rw, (void *)*vp,
 421                                     (void *)(*vp + RW_ONE_READER));
 422                         td->td_rw_rlocks++;
 423                         return (true);
 424                 }
 425         }
 426         return (false);
 427 }
 428
 429 static void __noinline
 430 __rw_rlock_hard(struct rwlock *rw, struct thread *td, uintptr_t v
 431     LOCK_FILE_LINE_ARG_DEF)
 432 {
 433         struct turnstile *ts;
 434         struct thread *owner;
 435 #ifdef ADAPTIVE_RWLOCKS
 436         int spintries = 0;
 437         int i, n;
 438 #endif
 439 #ifdef LOCK_PROFILING
 440         uint64_t waittime = 0;
 441         int contested = 0;
 442 #endif
 443 #if defined(ADAPTIVE_RWLOCKS) || defined(KDTRACE_HOOKS)
 444         struct lock_delay_arg lda;
 445 #endif
 446 #ifdef KDTRACE_HOOKS
 447         u_int sleep_cnt = 0;
 448         int64_t sleep_time = 0;
 449         int64_t all_time = 0;
 450 #endif
 451 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
 452         uintptr_t state = 0;
 453         int doing_lockprof = 0;
 454 #endif
 455
 456 #ifdef KDTRACE_HOOKS
 457         if (LOCKSTAT_PROFILE_ENABLED(rw__acquire)) {
 458                 if (__rw_rlock_try(rw, td, &v, false LOCK_FILE_LINE_ARG))
 459                         goto out_lockstat;
 460                 doing_lockprof = 1;
 461                 all_time -= lockstat_nsecs(&rw->lock_object);
 462                 state = v;
 463         }
 464 #endif
 465 #ifdef LOCK_PROFILING
 466         doing_lockprof = 1;
 467         state = v;
 468 #endif
 469
 470         if (SCHEDULER_STOPPED())
 471                 return;
 472
 473 #if defined(ADAPTIVE_RWLOCKS)
 474         lock_delay_arg_init(&lda, &rw_delay);
 475 #elif defined(KDTRACE_HOOKS)
 476         lock_delay_arg_init_noadapt(&lda);
 477 #endif
 478
 479 #ifdef HWPMC_HOOKS
 480         PMC_SOFT_CALL( , , lock, failed);
 481 #endif
 482         lock_profile_obtain_lock_failed(&rw->lock_object, false,
 483             &contested, &waittime);
 484
 485         THREAD_CONTENDS_ON_LOCK(&rw->lock_object);
 486
 487         for (;;) {
 488                 if (__rw_rlock_try(rw, td, &v, false LOCK_FILE_LINE_ARG))
 489                         break;
 490 #ifdef KDTRACE_HOOKS
 491                 lda.spin_cnt++;
 492 #endif
 493
 494 #ifdef ADAPTIVE_RWLOCKS
 495                 /*
 496                  * If the owner is running on another CPU, spin until
 497                  * the owner stops running or the state of the lock
 498                  * changes.
 499                  */
 500                 if ((v & RW_LOCK_READ) == 0) {
 501                         owner = (struct thread *)RW_OWNER(v);
 502                         if (TD_IS_RUNNING(owner)) {
 503                                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 504                                         CTR3(KTR_LOCK,
 505                                             "%s: spinning on %p held by %p",
 506                                             __func__, rw, owner);
 507                                 KTR_STATE1(KTR_SCHED, "thread",
 508                                     sched_tdname(curthread), "spinning",
 509                                     "lockname:\"%s\"", rw->lock_object.lo_name);
 510                                 do {
 511                                         lock_delay(&lda);
 512                                         v = RW_READ_VALUE(rw);
 513                                         owner = lv_rw_wowner(v);
 514                                 } while (owner != NULL && TD_IS_RUNNING(owner));
 515                                 KTR_STATE0(KTR_SCHED, "thread",
 516                                     sched_tdname(curthread), "running");
 517                                 continue;
 518                         }
 519                 } else {
 520                         if ((v & RW_LOCK_WRITE_SPINNER) && RW_READERS(v) == 0) {
 521                                 MPASS(!__rw_can_read(td, v, false));
 522                                 lock_delay_spin(2);
 523                                 v = RW_READ_VALUE(rw);
 524                                 continue;
 525                         }
 526                         if (spintries < rowner_retries) {
 527                                 spintries++;
 528                                 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
 529                                     "spinning", "lockname:\"%s\"",
 530                                     rw->lock_object.lo_name);
 531                                 n = RW_READERS(v);
 532                                 for (i = 0; i < rowner_loops; i += n) {
 533                                         lock_delay_spin(n);
 534                                         v = RW_READ_VALUE(rw);
 535                                         if (!(v & RW_LOCK_READ))
 536                                                 break;
 537                                         n = RW_READERS(v);
 538                                         if (n == 0)
 539                                                 break;
 540                                         if (__rw_can_read(td, v, false))
 541                                                 break;
 542                                 }
 543 #ifdef KDTRACE_HOOKS
 544                                 lda.spin_cnt += rowner_loops - i;
 545 #endif
 546                                 KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
 547                                     "running");
 548                                 if (i < rowner_loops)
 549                                         continue;
 550                         }
 551                 }
 552 #endif
 553
 554                 /*
 555                  * Okay, now it's the hard case.  Some other thread already
 556                  * has a write lock or there are write waiters present,
 557                  * acquire the turnstile lock so we can begin the process
 558                  * of blocking.
 559                  */
 560                 ts = turnstile_trywait(&rw->lock_object);
 561
 562                 /*
 563                  * The lock might have been released while we spun, so
 564                  * recheck its state and restart the loop if needed.
 565                  */
 566                 v = RW_READ_VALUE(rw);
 567 retry_ts:
 568                 if (((v & RW_LOCK_WRITE_SPINNER) && RW_READERS(v) == 0) ||
 569                     __rw_can_read(td, v, false)) {
 570                         turnstile_cancel(ts);
 571                         continue;
 572                 }
 573
 574                 owner = lv_rw_wowner(v);
 575
 576 #ifdef ADAPTIVE_RWLOCKS
 577                 /*
 578                  * The current lock owner might have started executing
 579                  * on another CPU (or the lock could have changed
 580                  * owners) while we were waiting on the turnstile
 581                  * chain lock.  If so, drop the turnstile lock and try
 582                  * again.
 583                  */
 584                 if (owner != NULL) {
 585                         if (TD_IS_RUNNING(owner)) {
 586                                 turnstile_cancel(ts);
 587                                 continue;
 588                         }
 589                 }
 590 #endif
 591
 592                 /*
 593                  * The lock is held in write mode or it already has waiters.
 594                  */
 595                 MPASS(!__rw_can_read(td, v, false));
 596
 597                 /*
 598                  * If the RW_LOCK_READ_WAITERS flag is already set, then
 599                  * we can go ahead and block.  If it is not set then try
 600                  * to set it.  If we fail to set it drop the turnstile
 601                  * lock and restart the loop.
 602                  */
 603                 if (!(v & RW_LOCK_READ_WAITERS)) {
 604                         if (!atomic_fcmpset_ptr(&rw->rw_lock, &v,
 605                             v | RW_LOCK_READ_WAITERS))
 606                                 goto retry_ts;
 607                         if (LOCK_LOG_TEST(&rw->lock_object, 0))
 608                                 CTR2(KTR_LOCK, "%s: %p set read waiters flag",
 609                                     __func__, rw);
 610                 }
 611
 612                 /*
 613                  * We were unable to acquire the lock and the read waiters
 614                  * flag is set, so we must block on the turnstile.
 615                  */
 616                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 617                         CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__,
 618                             rw);
 619 #ifdef KDTRACE_HOOKS
 620                 sleep_time -= lockstat_nsecs(&rw->lock_object);
 621 #endif
 622                 MPASS(owner == rw_owner(rw));
 623                 turnstile_wait(ts, owner, TS_SHARED_QUEUE);
 624 #ifdef KDTRACE_HOOKS
 625                 sleep_time += lockstat_nsecs(&rw->lock_object);
 626                 sleep_cnt++;
 627 #endif
 628                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 629                         CTR2(KTR_LOCK, "%s: %p resuming from turnstile",
 630                             __func__, rw);
 631                 v = RW_READ_VALUE(rw);
 632         }
 633         THREAD_CONTENTION_DONE(&rw->lock_object);
 634 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
 635         if (__predict_true(!doing_lockprof))
 636                 return;
 637 #endif
 638 #ifdef KDTRACE_HOOKS
 639         all_time += lockstat_nsecs(&rw->lock_object);
 640         if (sleep_time)
 641                 LOCKSTAT_RECORD4(rw__block, rw, sleep_time,
 642                     LOCKSTAT_READER, (state & RW_LOCK_READ) == 0,
 643                     (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
 644
 645         /* Record only the loops spinning and not sleeping. */
 646         if (lda.spin_cnt > sleep_cnt)
 647                 LOCKSTAT_RECORD4(rw__spin, rw, all_time - sleep_time,
 648                     LOCKSTAT_READER, (state & RW_LOCK_READ) == 0,
 649                     (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
 650 out_lockstat:
 651 #endif
 652         /*
 653          * TODO: acquire "owner of record" here.  Here be turnstile dragons
 654          * however.  turnstiles don't like owners changing between calls to
 655          * turnstile_wait() currently.
 656          */
 657         LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire, rw, contested,
 658             waittime, file, line, LOCKSTAT_READER);
 659 }
 660
 661 void
 662 __rw_rlock_int(struct rwlock *rw LOCK_FILE_LINE_ARG_DEF)
 663 {
 664         struct thread *td;
 665         uintptr_t v;
 666
 667         td = curthread;
 668
 669         KASSERT(kdb_active != 0 || SCHEDULER_STOPPED() ||
 670             !TD_IS_IDLETHREAD(td),
 671             ("rw_rlock() by idle thread %p on rwlock %s @ %s:%d",
 672             td, rw->lock_object.lo_name, file, line));
 673         KASSERT(rw->rw_lock != RW_DESTROYED,
 674             ("rw_rlock() of destroyed rwlock @ %s:%d", file, line));
 675         KASSERT(rw_wowner(rw) != td,
 676             ("rw_rlock: wlock already held for %s @ %s:%d",
 677             rw->lock_object.lo_name, file, line));
 678         WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER, file, line, NULL);
 679
 680         v = RW_READ_VALUE(rw);
 681         if (__predict_false(LOCKSTAT_PROFILE_ENABLED(rw__acquire) ||
 682             !__rw_rlock_try(rw, td, &v, true LOCK_FILE_LINE_ARG)))
 683                 __rw_rlock_hard(rw, td, v LOCK_FILE_LINE_ARG);
 684         else
 685                 lock_profile_obtain_lock_success(&rw->lock_object, false, 0, 0,
 686                     file, line);
 687
 688         LOCK_LOG_LOCK("RLOCK", &rw->lock_object, 0, 0, file, line);
 689         WITNESS_LOCK(&rw->lock_object, 0, file, line);
 690         TD_LOCKS_INC(curthread);
 691 }
 692
 693 void
 694 __rw_rlock(volatile uintptr_t *c, const char *file, int line)
 695 {
 696         struct rwlock *rw;
 697
 698         rw = rwlock2rw(c);
 699         __rw_rlock_int(rw LOCK_FILE_LINE_ARG);
 700 }
 701
 702 int
 703 __rw_try_rlock_int(struct rwlock *rw LOCK_FILE_LINE_ARG_DEF)
 704 {
 705         uintptr_t x;
 706
 707         if (SCHEDULER_STOPPED())
 708                 return (1);
 709
 710         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
 711             ("rw_try_rlock() by idle thread %p on rwlock %s @ %s:%d",
 712             curthread, rw->lock_object.lo_name, file, line));
 713
 714         x = rw->rw_lock;
 715         for (;;) {
 716                 KASSERT(rw->rw_lock != RW_DESTROYED,
 717                     ("rw_try_rlock() of destroyed rwlock @ %s:%d", file, line));
 718                 if (!(x & RW_LOCK_READ))
 719                         break;
 720                 if (atomic_fcmpset_acq_ptr(&rw->rw_lock, &x, x + RW_ONE_READER)) {
 721                         LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 1, file,
 722                             line);
 723                         WITNESS_LOCK(&rw->lock_object, LOP_TRYLOCK, file, line);
 724                         LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire,
 725                             rw, 0, 0, file, line, LOCKSTAT_READER);
 726                         TD_LOCKS_INC(curthread);
 727                         curthread->td_rw_rlocks++;
 728                         return (1);
 729                 }
 730         }
 731
 732         LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 0, file, line);
 733         return (0);
 734 }
 735
 736 int
 737 __rw_try_rlock(volatile uintptr_t *c, const char *file, int line)
 738 {
 739         struct rwlock *rw;
 740
 741         rw = rwlock2rw(c);
 742         return (__rw_try_rlock_int(rw LOCK_FILE_LINE_ARG));
 743 }
 744
 745 static bool __always_inline
 746 __rw_runlock_try(struct rwlock *rw, struct thread *td, uintptr_t *vp)
 747 {
 748
 749         for (;;) {
 750                 if (RW_READERS(*vp) > 1 || !(*vp & RW_LOCK_WAITERS)) {
 751                         if (atomic_fcmpset_rel_ptr(&rw->rw_lock, vp,
 752                             *vp - RW_ONE_READER)) {
 753                                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 754                                         CTR4(KTR_LOCK,
 755                                             "%s: %p succeeded %p -> %p",
 756                                             __func__, rw, (void *)*vp,
 757                                             (void *)(*vp - RW_ONE_READER));
 758                                 td->td_rw_rlocks--;
 759                                 return (true);
 760                         }
 761                         continue;
 762                 }
 763                 break;
 764         }
 765         return (false);
 766 }
 767
 768 static void __noinline
 769 __rw_runlock_hard(struct rwlock *rw, struct thread *td, uintptr_t v
 770     LOCK_FILE_LINE_ARG_DEF)
 771 {
 772         struct turnstile *ts;
 773         uintptr_t setv, queue;
 774
 775         if (SCHEDULER_STOPPED())
 776                 return;
 777
 778         if (__rw_runlock_try(rw, td, &v))
 779                 goto out_lockstat;
 780
 781         /*
 782          * Ok, we know we have waiters and we think we are the
 783          * last reader, so grab the turnstile lock.
 784          */
 785         turnstile_chain_lock(&rw->lock_object);
 786         v = RW_READ_VALUE(rw);
 787         for (;;) {
 788                 if (__rw_runlock_try(rw, td, &v))
 789                         break;
 790
 791                 MPASS(v & RW_LOCK_WAITERS);
 792
 793                 /*
 794                  * Try to drop our lock leaving the lock in a unlocked
 795                  * state.
 796                  *
 797                  * If you wanted to do explicit lock handoff you'd have to
 798                  * do it here.  You'd also want to use turnstile_signal()
 799                  * and you'd have to handle the race where a higher
 800                  * priority thread blocks on the write lock before the
 801                  * thread you wakeup actually runs and have the new thread
 802                  * "steal" the lock.  For now it's a lot simpler to just
 803                  * wakeup all of the waiters.
 804                  *
 805                  * As above, if we fail, then another thread might have
 806                  * acquired a read lock, so drop the turnstile lock and
 807                  * restart.
 808                  */
 809                 setv = RW_UNLOCKED;
 810                 queue = TS_SHARED_QUEUE;
 811                 if (v & RW_LOCK_WRITE_WAITERS) {
 812                         queue = TS_EXCLUSIVE_QUEUE;
 813                         setv |= (v & RW_LOCK_READ_WAITERS);
 814                 }
 815                 setv |= (v & RW_LOCK_WRITE_SPINNER);
 816                 if (!atomic_fcmpset_rel_ptr(&rw->rw_lock, &v, setv))
 817                         continue;
 818                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 819                         CTR2(KTR_LOCK, "%s: %p last succeeded with waiters",
 820                             __func__, rw);
 821
 822                 /*
 823                  * Ok.  The lock is released and all that's left is to
 824                  * wake up the waiters.  Note that the lock might not be
 825                  * free anymore, but in that case the writers will just
 826                  * block again if they run before the new lock holder(s)
 827                  * release the lock.
 828                  */
 829                 ts = turnstile_lookup(&rw->lock_object);
 830                 MPASS(ts != NULL);
 831                 turnstile_broadcast(ts, queue);
 832                 turnstile_unpend(ts);
 833                 td->td_rw_rlocks--;
 834                 break;
 835         }
 836         turnstile_chain_unlock(&rw->lock_object);
 837 out_lockstat:
 838         LOCKSTAT_PROFILE_RELEASE_RWLOCK(rw__release, rw, LOCKSTAT_READER);
 839 }
 840
 841 void
 842 _rw_runlock_cookie_int(struct rwlock *rw LOCK_FILE_LINE_ARG_DEF)
 843 {
 844         struct thread *td;
 845         uintptr_t v;
 846
 847         KASSERT(rw->rw_lock != RW_DESTROYED,
 848             ("rw_runlock() of destroyed rwlock @ %s:%d", file, line));
 849         __rw_assert(&rw->rw_lock, RA_RLOCKED, file, line);
 850         WITNESS_UNLOCK(&rw->lock_object, 0, file, line);
 851         LOCK_LOG_LOCK("RUNLOCK", &rw->lock_object, 0, 0, file, line);
 852
 853         td = curthread;
 854         v = RW_READ_VALUE(rw);
 855
 856         if (__predict_false(LOCKSTAT_PROFILE_ENABLED(rw__release) ||
 857             !__rw_runlock_try(rw, td, &v)))
 858                 __rw_runlock_hard(rw, td, v LOCK_FILE_LINE_ARG);
 859         else
 860                 lock_profile_release_lock(&rw->lock_object, false);
 861
 862         TD_LOCKS_DEC(curthread);
 863 }
 864
 865 void
 866 _rw_runlock_cookie(volatile uintptr_t *c, const char *file, int line)
 867 {
 868         struct rwlock *rw;
 869
 870         rw = rwlock2rw(c);
 871         _rw_runlock_cookie_int(rw LOCK_FILE_LINE_ARG);
 872 }
 873
 874 #ifdef ADAPTIVE_RWLOCKS
 875 static inline void
 876 rw_drop_critical(uintptr_t v, bool *in_critical, int *extra_work)
 877 {
 878
 879         if (v & RW_LOCK_WRITE_SPINNER)
 880                 return;
 881         if (*in_critical) {
 882                 critical_exit();
 883                 *in_critical = false;
 884                 (*extra_work)--;
 885         }
 886 }
 887 #else
 888 #define rw_drop_critical(v, in_critical, extra_work) do { } while (0)
 889 #endif
 890
 891 /*
 892  * This function is called when we are unable to obtain a write lock on the
 893  * first try.  This means that at least one other thread holds either a
 894  * read or write lock.
 895  */
 896 void
 897 __rw_wlock_hard(volatile uintptr_t *c, uintptr_t v LOCK_FILE_LINE_ARG_DEF)
 898 {
 899         uintptr_t tid;
 900         struct rwlock *rw;
 901         struct turnstile *ts;
 902         struct thread *owner;
 903 #ifdef ADAPTIVE_RWLOCKS
 904         int spintries = 0;
 905         int i, n;
 906         enum { READERS, WRITER } sleep_reason = READERS;
 907         bool in_critical = false;
 908 #endif
 909         uintptr_t setv;
 910 #ifdef LOCK_PROFILING
 911         uint64_t waittime = 0;
 912         int contested = 0;
 913 #endif
 914 #if defined(ADAPTIVE_RWLOCKS) || defined(KDTRACE_HOOKS)
 915         struct lock_delay_arg lda;
 916 #endif
 917 #ifdef KDTRACE_HOOKS
 918         u_int sleep_cnt = 0;
 919         int64_t sleep_time = 0;
 920         int64_t all_time = 0;
 921 #endif
 922 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
 923         uintptr_t state = 0;
 924         int doing_lockprof = 0;
 925 #endif
 926         int extra_work = 0;
 927
 928         tid = (uintptr_t)curthread;
 929         rw = rwlock2rw(c);
 930
 931 #ifdef KDTRACE_HOOKS
 932         if (LOCKSTAT_PROFILE_ENABLED(rw__acquire)) {
 933                 while (v == RW_UNLOCKED) {
 934                         if (_rw_write_lock_fetch(rw, &v, tid))
 935                                 goto out_lockstat;
 936                 }
 937                 extra_work = 1;
 938                 doing_lockprof = 1;
 939                 all_time -= lockstat_nsecs(&rw->lock_object);
 940                 state = v;
 941         }
 942 #endif
 943 #ifdef LOCK_PROFILING
 944         extra_work = 1;
 945         doing_lockprof = 1;
 946         state = v;
 947 #endif
 948
 949         if (SCHEDULER_STOPPED())
 950                 return;
 951
 952         if (__predict_false(v == RW_UNLOCKED))
 953                 v = RW_READ_VALUE(rw);
 954
 955         if (__predict_false(lv_rw_wowner(v) == (struct thread *)tid)) {
 956                 KASSERT(rw->lock_object.lo_flags & LO_RECURSABLE,
 957                     ("%s: recursing but non-recursive rw %s @ %s:%d\n",
 958                     __func__, rw->lock_object.lo_name, file, line));
 959                 rw->rw_recurse++;
 960                 atomic_set_ptr(&rw->rw_lock, RW_LOCK_WRITER_RECURSED);
 961                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 962                         CTR2(KTR_LOCK, "%s: %p recursing", __func__, rw);
 963                 return;
 964         }
 965
 966         if (LOCK_LOG_TEST(&rw->lock_object, 0))
 967                 CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__,
 968                     rw->lock_object.lo_name, (void *)rw->rw_lock, file, line);
 969
 970 #if defined(ADAPTIVE_RWLOCKS)
 971         lock_delay_arg_init(&lda, &rw_delay);
 972 #elif defined(KDTRACE_HOOKS)
 973         lock_delay_arg_init_noadapt(&lda);
 974 #endif
 975
 976 #ifdef HWPMC_HOOKS
 977         PMC_SOFT_CALL( , , lock, failed);
 978 #endif
 979         lock_profile_obtain_lock_failed(&rw->lock_object, false,
 980             &contested, &waittime);
 981
 982         THREAD_CONTENDS_ON_LOCK(&rw->lock_object);
 983
 984         for (;;) {
 985                 if (v == RW_UNLOCKED) {
 986                         if (_rw_write_lock_fetch(rw, &v, tid))
 987                                 break;
 988                         continue;
 989                 }
 990 #ifdef KDTRACE_HOOKS
 991                 lda.spin_cnt++;
 992 #endif
 993
 994 #ifdef ADAPTIVE_RWLOCKS
 995                 if (v == (RW_LOCK_READ | RW_LOCK_WRITE_SPINNER)) {
 996                         if (atomic_fcmpset_acq_ptr(&rw->rw_lock, &v, tid))
 997                                 break;
 998                         continue;
 999                 }
1000
1001                 /*
1002                  * If the lock is write locked and the owner is
1003                  * running on another CPU, spin until the owner stops
1004                  * running or the state of the lock changes.
1005                  */
1006                 if (!(v & RW_LOCK_READ)) {
1007                         rw_drop_critical(v, &in_critical, &extra_work);
1008                         sleep_reason = WRITER;
1009                         owner = lv_rw_wowner(v);
1010                         if (!TD_IS_RUNNING(owner))
1011                                 goto ts;
1012                         if (LOCK_LOG_TEST(&rw->lock_object, 0))
1013                                 CTR3(KTR_LOCK, "%s: spinning on %p held by %p",
1014                                     __func__, rw, owner);
1015                         KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
1016                             "spinning", "lockname:\"%s\"",
1017                             rw->lock_object.lo_name);
1018                         do {
1019                                 lock_delay(&lda);
1020                                 v = RW_READ_VALUE(rw);
1021                                 owner = lv_rw_wowner(v);
1022                         } while (owner != NULL && TD_IS_RUNNING(owner));
1023                         KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
1024                             "running");
1025                         continue;
1026                 } else if (RW_READERS(v) > 0) {
1027                         sleep_reason = READERS;
1028                         if (spintries == rowner_retries)
1029                                 goto ts;
1030                         if (!(v & RW_LOCK_WRITE_SPINNER)) {
1031                                 if (!in_critical) {
1032                                         critical_enter();
1033                                         in_critical = true;
1034                                         extra_work++;
1035                                 }
1036                                 if (!atomic_fcmpset_ptr(&rw->rw_lock, &v,
1037                                     v | RW_LOCK_WRITE_SPINNER)) {
1038                                         critical_exit();
1039                                         in_critical = false;
1040                                         extra_work--;
1041                                         continue;
1042                                 }
1043                         }
1044                         spintries++;
1045                         KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
1046                             "spinning", "lockname:\"%s\"",
1047                             rw->lock_object.lo_name);
1048                         n = RW_READERS(v);
1049                         for (i = 0; i < rowner_loops; i += n) {
1050                                 lock_delay_spin(n);
1051                                 v = RW_READ_VALUE(rw);
1052                                 if (!(v & RW_LOCK_WRITE_SPINNER))
1053                                         break;
1054                                 if (!(v & RW_LOCK_READ))
1055                                         break;
1056                                 n = RW_READERS(v);
1057                                 if (n == 0)
1058                                         break;
1059                         }
1060 #ifdef KDTRACE_HOOKS
1061                         lda.spin_cnt += i;
1062 #endif
1063                         KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
1064                             "running");
1065                         if (i < rowner_loops)
1066                                 continue;
1067                 }
1068 ts:
1069 #endif
1070                 ts = turnstile_trywait(&rw->lock_object);
1071                 v = RW_READ_VALUE(rw);
1072 retry_ts:
1073                 owner = lv_rw_wowner(v);
1074
1075 #ifdef ADAPTIVE_RWLOCKS
1076                 /*
1077                  * The current lock owner might have started executing
1078                  * on another CPU (or the lock could have changed
1079                  * owners) while we were waiting on the turnstile
1080                  * chain lock.  If so, drop the turnstile lock and try
1081                  * again.
1082                  */
1083                 if (owner != NULL) {
1084                         if (TD_IS_RUNNING(owner)) {
1085                                 turnstile_cancel(ts);
1086                                 rw_drop_critical(v, &in_critical, &extra_work);
1087                                 continue;
1088                         }
1089                 } else if (RW_READERS(v) > 0 && sleep_reason == WRITER) {
1090                         turnstile_cancel(ts);
1091                         rw_drop_critical(v, &in_critical, &extra_work);
1092                         continue;
1093                 }
1094 #endif
1095                 /*
1096                  * Check for the waiters flags about this rwlock.
1097                  * If the lock was released, without maintain any pending
1098                  * waiters queue, simply try to acquire it.
1099                  * If a pending waiters queue is present, claim the lock
1100                  * ownership and maintain the pending queue.
1101                  */
1102                 setv = v & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER);
1103                 if ((v & ~setv) == RW_UNLOCKED) {
1104                         setv &= ~RW_LOCK_WRITE_SPINNER;
1105                         if (atomic_fcmpset_acq_ptr(&rw->rw_lock, &v, tid | setv)) {
1106                                 if (setv)
1107                                         turnstile_claim(ts);
1108                                 else
1109                                         turnstile_cancel(ts);
1110                                 break;
1111                         }
1112                         goto retry_ts;
1113                 }
1114
1115 #ifdef ADAPTIVE_RWLOCKS
1116                 if (in_critical) {
1117                         if ((v & RW_LOCK_WRITE_SPINNER) ||
1118                             !((v & RW_LOCK_WRITE_WAITERS))) {
1119                                 setv = v & ~RW_LOCK_WRITE_SPINNER;
1120                                 setv |= RW_LOCK_WRITE_WAITERS;
1121                                 if (!atomic_fcmpset_ptr(&rw->rw_lock, &v, setv))
1122                                         goto retry_ts;
1123                         }
1124                         critical_exit();
1125                         in_critical = false;
1126                         extra_work--;
1127                 } else {
1128 #endif
1129                         /*
1130                          * If the RW_LOCK_WRITE_WAITERS flag isn't set, then try to
1131                          * set it.  If we fail to set it, then loop back and try
1132                          * again.
1133                          */
1134                         if (!(v & RW_LOCK_WRITE_WAITERS)) {
1135                                 if (!atomic_fcmpset_ptr(&rw->rw_lock, &v,
1136                                     v | RW_LOCK_WRITE_WAITERS))
1137                                         goto retry_ts;
1138                                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
1139                                         CTR2(KTR_LOCK, "%s: %p set write waiters flag",
1140                                             __func__, rw);
1141                         }
1142 #ifdef ADAPTIVE_RWLOCKS
1143                 }
1144 #endif
1145                 /*
1146                  * We were unable to acquire the lock and the write waiters
1147                  * flag is set, so we must block on the turnstile.
1148                  */
1149                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
1150                         CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__,
1151                             rw);
1152 #ifdef KDTRACE_HOOKS
1153                 sleep_time -= lockstat_nsecs(&rw->lock_object);
1154 #endif
1155                 MPASS(owner == rw_owner(rw));
1156                 turnstile_wait(ts, owner, TS_EXCLUSIVE_QUEUE);
1157 #ifdef KDTRACE_HOOKS
1158                 sleep_time += lockstat_nsecs(&rw->lock_object);
1159                 sleep_cnt++;
1160 #endif
1161                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
1162                         CTR2(KTR_LOCK, "%s: %p resuming from turnstile",
1163                             __func__, rw);
1164 #ifdef ADAPTIVE_RWLOCKS
1165                 spintries = 0;
1166 #endif
1167                 v = RW_READ_VALUE(rw);
1168         }
1169         THREAD_CONTENTION_DONE(&rw->lock_object);
1170         if (__predict_true(!extra_work))
1171                 return;
1172 #ifdef ADAPTIVE_RWLOCKS
1173         if (in_critical)
1174                 critical_exit();
1175 #endif
1176 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
1177         if (__predict_true(!doing_lockprof))
1178                 return;
1179 #endif
1180 #ifdef KDTRACE_HOOKS
1181         all_time += lockstat_nsecs(&rw->lock_object);
1182         if (sleep_time)
1183                 LOCKSTAT_RECORD4(rw__block, rw, sleep_time,
1184                     LOCKSTAT_WRITER, (state & RW_LOCK_READ) == 0,
1185                     (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
1186
1187         /* Record only the loops spinning and not sleeping. */
1188         if (lda.spin_cnt > sleep_cnt)
1189                 LOCKSTAT_RECORD4(rw__spin, rw, all_time - sleep_time,
1190                     LOCKSTAT_WRITER, (state & RW_LOCK_READ) == 0,
1191                     (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
1192 out_lockstat:
1193 #endif
1194         LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire, rw, contested,
1195             waittime, file, line, LOCKSTAT_WRITER);
1196 }
1197
1198 /*
1199  * This function is called if lockstat is active or the first try at releasing
1200  * a write lock failed.  The latter means that the lock is recursed or one of
1201  * the 2 waiter bits must be set indicating that at least one thread is waiting
1202  * on this lock.
1203  */
1204 void
1205 __rw_wunlock_hard(volatile uintptr_t *c, uintptr_t v LOCK_FILE_LINE_ARG_DEF)
1206 {
1207         struct rwlock *rw;
1208         struct turnstile *ts;
1209         uintptr_t tid, setv;
1210         int queue;
1211
1212         tid = (uintptr_t)curthread;
1213         if (SCHEDULER_STOPPED())
1214                 return;
1215
1216         rw = rwlock2rw(c);
1217         if (__predict_false(v == tid))
1218                 v = RW_READ_VALUE(rw);
1219
1220         if (v & RW_LOCK_WRITER_RECURSED) {
1221                 if (--(rw->rw_recurse) == 0)
1222                         atomic_clear_ptr(&rw->rw_lock, RW_LOCK_WRITER_RECURSED);
1223                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
1224                         CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, rw);
1225                 return;
1226         }
1227
1228         LOCKSTAT_PROFILE_RELEASE_RWLOCK(rw__release, rw, LOCKSTAT_WRITER);
1229         if (v == tid && _rw_write_unlock(rw, tid))
1230                 return;
1231
1232         KASSERT(rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS),
1233             ("%s: neither of the waiter flags are set", __func__));
1234
1235         if (LOCK_LOG_TEST(&rw->lock_object, 0))
1236                 CTR2(KTR_LOCK, "%s: %p contested", __func__, rw);
1237
1238         turnstile_chain_lock(&rw->lock_object);
1239
1240         /*
1241          * Use the same algo as sx locks for now.  Prefer waking up shared
1242          * waiters if we have any over writers.  This is probably not ideal.
1243          *
1244          * 'v' is the value we are going to write back to rw_lock.  If we
1245          * have waiters on both queues, we need to preserve the state of
1246          * the waiter flag for the queue we don't wake up.  For now this is
1247          * hardcoded for the algorithm mentioned above.
1248          *
1249          * In the case of both readers and writers waiting we wakeup the
1250          * readers but leave the RW_LOCK_WRITE_WAITERS flag set.  If a
1251          * new writer comes in before a reader it will claim the lock up
1252          * above.  There is probably a potential priority inversion in
1253          * there that could be worked around either by waking both queues
1254          * of waiters or doing some complicated lock handoff gymnastics.
1255          */
1256         setv = RW_UNLOCKED;
1257         v = RW_READ_VALUE(rw);
1258         queue = TS_SHARED_QUEUE;
1259         if (v & RW_LOCK_WRITE_WAITERS) {
1260                 queue = TS_EXCLUSIVE_QUEUE;
1261                 setv |= (v & RW_LOCK_READ_WAITERS);
1262         }
1263         atomic_store_rel_ptr(&rw->rw_lock, setv);
1264
1265         /* Wake up all waiters for the specific queue. */
1266         if (LOCK_LOG_TEST(&rw->lock_object, 0))
1267                 CTR3(KTR_LOCK, "%s: %p waking up %s waiters", __func__, rw,
1268                     queue == TS_SHARED_QUEUE ? "read" : "write");
1269
1270         ts = turnstile_lookup(&rw->lock_object);
1271         MPASS(ts != NULL);
1272         turnstile_broadcast(ts, queue);
1273         turnstile_unpend(ts);
1274         turnstile_chain_unlock(&rw->lock_object);
1275 }
1276
1277 /*
1278  * Attempt to do a non-blocking upgrade from a read lock to a write
1279  * lock.  This will only succeed if this thread holds a single read
1280  * lock.  Returns true if the upgrade succeeded and false otherwise.
1281  */
1282 int
1283 __rw_try_upgrade_int(struct rwlock *rw LOCK_FILE_LINE_ARG_DEF)
1284 {
1285         uintptr_t v, setv, tid;
1286         struct turnstile *ts;
1287         int success;
1288
1289         if (SCHEDULER_STOPPED())
1290                 return (1);
1291
1292         KASSERT(rw->rw_lock != RW_DESTROYED,
1293             ("rw_try_upgrade() of destroyed rwlock @ %s:%d", file, line));
1294         __rw_assert(&rw->rw_lock, RA_RLOCKED, file, line);
1295
1296         /*
1297          * Attempt to switch from one reader to a writer.  If there
1298          * are any write waiters, then we will have to lock the
1299          * turnstile first to prevent races with another writer
1300          * calling turnstile_wait() before we have claimed this
1301          * turnstile.  So, do the simple case of no waiters first.
1302          */
1303         tid = (uintptr_t)curthread;
1304         success = 0;
1305         v = RW_READ_VALUE(rw);
1306         for (;;) {
1307                 if (RW_READERS(v) > 1)
1308                         break;
1309                 if (!(v & RW_LOCK_WAITERS)) {
1310                         success = atomic_fcmpset_acq_ptr(&rw->rw_lock, &v, tid);
1311                         if (!success)
1312                                 continue;
1313                         break;
1314                 }
1315
1316                 /*
1317                  * Ok, we think we have waiters, so lock the turnstile.
1318                  */
1319                 ts = turnstile_trywait(&rw->lock_object);
1320                 v = RW_READ_VALUE(rw);
1321 retry_ts:
1322                 if (RW_READERS(v) > 1) {
1323                         turnstile_cancel(ts);
1324                         break;
1325                 }
1326                 /*
1327                  * Try to switch from one reader to a writer again.  This time
1328                  * we honor the current state of the waiters flags.
1329                  * If we obtain the lock with the flags set, then claim
1330                  * ownership of the turnstile.
1331                  */
1332                 setv = tid | (v & RW_LOCK_WAITERS);
1333                 success = atomic_fcmpset_ptr(&rw->rw_lock, &v, setv);
1334                 if (success) {
1335                         if (v & RW_LOCK_WAITERS)
1336                                 turnstile_claim(ts);
1337                         else
1338                                 turnstile_cancel(ts);
1339                         break;
1340                 }
1341                 goto retry_ts;
1342         }
1343         LOCK_LOG_TRY("WUPGRADE", &rw->lock_object, 0, success, file, line);
1344         if (success) {
1345                 curthread->td_rw_rlocks--;
1346                 WITNESS_UPGRADE(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
1347                     file, line);
1348                 LOCKSTAT_RECORD0(rw__upgrade, rw);
1349         }
1350         return (success);
1351 }
1352
1353 int
1354 __rw_try_upgrade(volatile uintptr_t *c, const char *file, int line)
1355 {
1356         struct rwlock *rw;
1357
1358         rw = rwlock2rw(c);
1359         return (__rw_try_upgrade_int(rw LOCK_FILE_LINE_ARG));
1360 }
1361
1362 /*
1363  * Downgrade a write lock into a single read lock.
1364  */
1365 void
1366 __rw_downgrade_int(struct rwlock *rw LOCK_FILE_LINE_ARG_DEF)
1367 {
1368         struct turnstile *ts;
1369         uintptr_t tid, v;
1370         int rwait, wwait;
1371
1372         if (SCHEDULER_STOPPED())
1373                 return;
1374
1375         KASSERT(rw->rw_lock != RW_DESTROYED,
1376             ("rw_downgrade() of destroyed rwlock @ %s:%d", file, line));
1377         __rw_assert(&rw->rw_lock, RA_WLOCKED | RA_NOTRECURSED, file, line);
1378 #ifndef INVARIANTS
1379         if (rw_recursed(rw))
1380                 panic("downgrade of a recursed lock");
1381 #endif
1382
1383         WITNESS_DOWNGRADE(&rw->lock_object, 0, file, line);
1384
1385         /*
1386          * Convert from a writer to a single reader.  First we handle
1387          * the easy case with no waiters.  If there are any waiters, we
1388          * lock the turnstile and "disown" the lock.
1389          */
1390         tid = (uintptr_t)curthread;
1391         if (atomic_cmpset_rel_ptr(&rw->rw_lock, tid, RW_READERS_LOCK(1)))
1392                 goto out;
1393
1394         /*
1395          * Ok, we think we have waiters, so lock the turnstile so we can
1396          * read the waiter flags without any races.
1397          */
1398         turnstile_chain_lock(&rw->lock_object);
1399         v = rw->rw_lock & RW_LOCK_WAITERS;
1400         rwait = v & RW_LOCK_READ_WAITERS;
1401         wwait = v & RW_LOCK_WRITE_WAITERS;
1402         MPASS(rwait | wwait);
1403
1404         /*
1405          * Downgrade from a write lock while preserving waiters flag
1406          * and give up ownership of the turnstile.
1407          */
1408         ts = turnstile_lookup(&rw->lock_object);
1409         MPASS(ts != NULL);
1410         if (!wwait)
1411                 v &= ~RW_LOCK_READ_WAITERS;
1412         atomic_store_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v);
1413         /*
1414          * Wake other readers if there are no writers pending.  Otherwise they
1415          * won't be able to acquire the lock anyway.
1416          */
1417         if (rwait && !wwait) {
1418                 turnstile_broadcast(ts, TS_SHARED_QUEUE);
1419                 turnstile_unpend(ts);
1420         } else
1421                 turnstile_disown(ts);
1422         turnstile_chain_unlock(&rw->lock_object);
1423 out:
1424         curthread->td_rw_rlocks++;
1425         LOCK_LOG_LOCK("WDOWNGRADE", &rw->lock_object, 0, 0, file, line);
1426         LOCKSTAT_RECORD0(rw__downgrade, rw);
1427 }
1428
1429 void
1430 __rw_downgrade(volatile uintptr_t *c, const char *file, int line)
1431 {
1432         struct rwlock *rw;
1433
1434         rw = rwlock2rw(c);
1435         __rw_downgrade_int(rw LOCK_FILE_LINE_ARG);
1436 }
1437
1438 #ifdef INVARIANT_SUPPORT
1439 #ifndef INVARIANTS
1440 #undef __rw_assert
1441 #endif
1442
1443 /*
1444  * In the non-WITNESS case, rw_assert() can only detect that at least
1445  * *some* thread owns an rlock, but it cannot guarantee that *this*
1446  * thread owns an rlock.
1447  */
1448 void
1449 __rw_assert(const volatile uintptr_t *c, int what, const char *file, int line)
1450 {
1451         const struct rwlock *rw;
1452
1453         if (SCHEDULER_STOPPED())
1454                 return;
1455
1456         rw = rwlock2rw(c);
1457
1458         switch (what) {
1459         case RA_LOCKED:
1460         case RA_LOCKED | RA_RECURSED:
1461         case RA_LOCKED | RA_NOTRECURSED:
1462         case RA_RLOCKED:
1463         case RA_RLOCKED | RA_RECURSED:
1464         case RA_RLOCKED | RA_NOTRECURSED:
1465 #ifdef WITNESS
1466                 witness_assert(&rw->lock_object, what, file, line);
1467 #else
1468                 /*
1469                  * If some other thread has a write lock or we have one
1470                  * and are asserting a read lock, fail.  Also, if no one
1471                  * has a lock at all, fail.
1472                  */
1473                 if (rw->rw_lock == RW_UNLOCKED ||
1474                     (!(rw->rw_lock & RW_LOCK_READ) && (what & RA_RLOCKED ||
1475                     rw_wowner(rw) != curthread)))
1476                         panic("Lock %s not %slocked @ %s:%d\n",
1477                             rw->lock_object.lo_name, (what & RA_RLOCKED) ?
1478                             "read " : "", file, line);
1479
1480                 if (!(rw->rw_lock & RW_LOCK_READ) && !(what & RA_RLOCKED)) {
1481                         if (rw_recursed(rw)) {
1482                                 if (what & RA_NOTRECURSED)
1483                                         panic("Lock %s recursed @ %s:%d\n",
1484                                             rw->lock_object.lo_name, file,
1485                                             line);
1486                         } else if (what & RA_RECURSED)
1487                                 panic("Lock %s not recursed @ %s:%d\n",
1488                                     rw->lock_object.lo_name, file, line);
1489                 }
1490 #endif
1491                 break;
1492         case RA_WLOCKED:
1493         case RA_WLOCKED | RA_RECURSED:
1494         case RA_WLOCKED | RA_NOTRECURSED:
1495                 if (rw_wowner(rw) != curthread)
1496                         panic("Lock %s not exclusively locked @ %s:%d\n",
1497                             rw->lock_object.lo_name, file, line);
1498                 if (rw_recursed(rw)) {
1499                         if (what & RA_NOTRECURSED)
1500                                 panic("Lock %s recursed @ %s:%d\n",
1501                                     rw->lock_object.lo_name, file, line);
1502                 } else if (what & RA_RECURSED)
1503                         panic("Lock %s not recursed @ %s:%d\n",
1504                             rw->lock_object.lo_name, file, line);
1505                 break;
1506         case RA_UNLOCKED:
1507 #ifdef WITNESS
1508                 witness_assert(&rw->lock_object, what, file, line);
1509 #else
1510                 /*
1511                  * If we hold a write lock fail.  We can't reliably check
1512                  * to see if we hold a read lock or not.
1513                  */
1514                 if (rw_wowner(rw) == curthread)
1515                         panic("Lock %s exclusively locked @ %s:%d\n",
1516                             rw->lock_object.lo_name, file, line);
1517 #endif
1518                 break;
1519         default:
1520                 panic("Unknown rw lock assertion: %d @ %s:%d", what, file,
1521                     line);
1522         }
1523 }
1524 #endif /* INVARIANT_SUPPORT */
1525
1526 #ifdef DDB
1527 void
1528 db_show_rwlock(const struct lock_object *lock)
1529 {
1530         const struct rwlock *rw;
1531         struct thread *td;
1532
1533         rw = (const struct rwlock *)lock;
1534
1535         db_printf(" state: ");
1536         if (rw->rw_lock == RW_UNLOCKED)
1537                 db_printf("UNLOCKED\n");
1538         else if (rw->rw_lock == RW_DESTROYED) {
1539                 db_printf("DESTROYED\n");
1540                 return;
1541         } else if (rw->rw_lock & RW_LOCK_READ)
1542                 db_printf("RLOCK: %ju locks\n",
1543                     (uintmax_t)(RW_READERS(rw->rw_lock)));
1544         else {
1545                 td = rw_wowner(rw);
1546                 db_printf("WLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1547                     td->td_tid, td->td_proc->p_pid, td->td_name);
1548                 if (rw_recursed(rw))
1549                         db_printf(" recursed: %u\n", rw->rw_recurse);
1550         }
1551         db_printf(" waiters: ");
1552         switch (rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS)) {
1553         case RW_LOCK_READ_WAITERS:
1554                 db_printf("readers\n");
1555                 break;
1556         case RW_LOCK_WRITE_WAITERS:
1557                 db_printf("writers\n");
1558                 break;
1559         case RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS:
1560                 db_printf("readers and writers\n");
1561                 break;
1562         default:
1563                 db_printf("none\n");
1564                 break;
1565         }
1566 }
1567
1568 #endif