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_TD(td))
 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         for (;;) {
 486                 if (__rw_rlock_try(rw, td, &v, false LOCK_FILE_LINE_ARG))
 487                         break;
 488 #ifdef KDTRACE_HOOKS
 489                 lda.spin_cnt++;
 490 #endif
 491
 492 #ifdef ADAPTIVE_RWLOCKS
 493                 /*
 494                  * If the owner is running on another CPU, spin until
 495                  * the owner stops running or the state of the lock
 496                  * changes.
 497                  */
 498                 if ((v & RW_LOCK_READ) == 0) {
 499                         owner = (struct thread *)RW_OWNER(v);
 500                         if (TD_IS_RUNNING(owner)) {
 501                                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 502                                         CTR3(KTR_LOCK,
 503                                             "%s: spinning on %p held by %p",
 504                                             __func__, rw, owner);
 505                                 KTR_STATE1(KTR_SCHED, "thread",
 506                                     sched_tdname(curthread), "spinning",
 507                                     "lockname:\"%s\"", rw->lock_object.lo_name);
 508                                 do {
 509                                         lock_delay(&lda);
 510                                         v = RW_READ_VALUE(rw);
 511                                         owner = lv_rw_wowner(v);
 512                                 } while (owner != NULL && TD_IS_RUNNING(owner));
 513                                 KTR_STATE0(KTR_SCHED, "thread",
 514                                     sched_tdname(curthread), "running");
 515                                 continue;
 516                         }
 517                 } else {
 518                         if ((v & RW_LOCK_WRITE_SPINNER) && RW_READERS(v) == 0) {
 519                                 MPASS(!__rw_can_read(td, v, false));
 520                                 lock_delay_spin(2);
 521                                 v = RW_READ_VALUE(rw);
 522                                 continue;
 523                         }
 524                         if (spintries < rowner_retries) {
 525                                 spintries++;
 526                                 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
 527                                     "spinning", "lockname:\"%s\"",
 528                                     rw->lock_object.lo_name);
 529                                 n = RW_READERS(v);
 530                                 for (i = 0; i < rowner_loops; i += n) {
 531                                         lock_delay_spin(n);
 532                                         v = RW_READ_VALUE(rw);
 533                                         if (!(v & RW_LOCK_READ))
 534                                                 break;
 535                                         n = RW_READERS(v);
 536                                         if (n == 0)
 537                                                 break;
 538                                         if (__rw_can_read(td, v, false))
 539                                                 break;
 540                                 }
 541 #ifdef KDTRACE_HOOKS
 542                                 lda.spin_cnt += rowner_loops - i;
 543 #endif
 544                                 KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
 545                                     "running");
 546                                 if (i < rowner_loops)
 547                                         continue;
 548                         }
 549                 }
 550 #endif
 551
 552                 /*
 553                  * Okay, now it's the hard case.  Some other thread already
 554                  * has a write lock or there are write waiters present,
 555                  * acquire the turnstile lock so we can begin the process
 556                  * of blocking.
 557                  */
 558                 ts = turnstile_trywait(&rw->lock_object);
 559
 560                 /*
 561                  * The lock might have been released while we spun, so
 562                  * recheck its state and restart the loop if needed.
 563                  */
 564                 v = RW_READ_VALUE(rw);
 565 retry_ts:
 566                 if (((v & RW_LOCK_WRITE_SPINNER) && RW_READERS(v) == 0) ||
 567                     __rw_can_read(td, v, false)) {
 568                         turnstile_cancel(ts);
 569                         continue;
 570                 }
 571
 572                 owner = lv_rw_wowner(v);
 573
 574 #ifdef ADAPTIVE_RWLOCKS
 575                 /*
 576                  * The current lock owner might have started executing
 577                  * on another CPU (or the lock could have changed
 578                  * owners) while we were waiting on the turnstile
 579                  * chain lock.  If so, drop the turnstile lock and try
 580                  * again.
 581                  */
 582                 if (owner != NULL) {
 583                         if (TD_IS_RUNNING(owner)) {
 584                                 turnstile_cancel(ts);
 585                                 continue;
 586                         }
 587                 }
 588 #endif
 589
 590                 /*
 591                  * The lock is held in write mode or it already has waiters.
 592                  */
 593                 MPASS(!__rw_can_read(td, v, false));
 594
 595                 /*
 596                  * If the RW_LOCK_READ_WAITERS flag is already set, then
 597                  * we can go ahead and block.  If it is not set then try
 598                  * to set it.  If we fail to set it drop the turnstile
 599                  * lock and restart the loop.
 600                  */
 601                 if (!(v & RW_LOCK_READ_WAITERS)) {
 602                         if (!atomic_fcmpset_ptr(&rw->rw_lock, &v,
 603                             v | RW_LOCK_READ_WAITERS))
 604                                 goto retry_ts;
 605                         if (LOCK_LOG_TEST(&rw->lock_object, 0))
 606                                 CTR2(KTR_LOCK, "%s: %p set read waiters flag",
 607                                     __func__, rw);
 608                 }
 609
 610                 /*
 611                  * We were unable to acquire the lock and the read waiters
 612                  * flag is set, so we must block on the turnstile.
 613                  */
 614                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 615                         CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__,
 616                             rw);
 617 #ifdef KDTRACE_HOOKS
 618                 sleep_time -= lockstat_nsecs(&rw->lock_object);
 619 #endif
 620                 MPASS(owner == rw_owner(rw));
 621                 turnstile_wait(ts, owner, TS_SHARED_QUEUE);
 622 #ifdef KDTRACE_HOOKS
 623                 sleep_time += lockstat_nsecs(&rw->lock_object);
 624                 sleep_cnt++;
 625 #endif
 626                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 627                         CTR2(KTR_LOCK, "%s: %p resuming from turnstile",
 628                             __func__, rw);
 629                 v = RW_READ_VALUE(rw);
 630         }
 631 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
 632         if (__predict_true(!doing_lockprof))
 633                 return;
 634 #endif
 635 #ifdef KDTRACE_HOOKS
 636         all_time += lockstat_nsecs(&rw->lock_object);
 637         if (sleep_time)
 638                 LOCKSTAT_RECORD4(rw__block, rw, sleep_time,
 639                     LOCKSTAT_READER, (state & RW_LOCK_READ) == 0,
 640                     (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
 641
 642         /* Record only the loops spinning and not sleeping. */
 643         if (lda.spin_cnt > sleep_cnt)
 644                 LOCKSTAT_RECORD4(rw__spin, rw, all_time - sleep_time,
 645                     LOCKSTAT_READER, (state & RW_LOCK_READ) == 0,
 646                     (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
 647 out_lockstat:
 648 #endif
 649         /*
 650          * TODO: acquire "owner of record" here.  Here be turnstile dragons
 651          * however.  turnstiles don't like owners changing between calls to
 652          * turnstile_wait() currently.
 653          */
 654         LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire, rw, contested,
 655             waittime, file, line, LOCKSTAT_READER);
 656 }
 657
 658 void
 659 __rw_rlock_int(struct rwlock *rw LOCK_FILE_LINE_ARG_DEF)
 660 {
 661         struct thread *td;
 662         uintptr_t v;
 663
 664         td = curthread;
 665
 666         KASSERT(kdb_active != 0 || SCHEDULER_STOPPED_TD(td) ||
 667             !TD_IS_IDLETHREAD(td),
 668             ("rw_rlock() by idle thread %p on rwlock %s @ %s:%d",
 669             td, rw->lock_object.lo_name, file, line));
 670         KASSERT(rw->rw_lock != RW_DESTROYED,
 671             ("rw_rlock() of destroyed rwlock @ %s:%d", file, line));
 672         KASSERT(rw_wowner(rw) != td,
 673             ("rw_rlock: wlock already held for %s @ %s:%d",
 674             rw->lock_object.lo_name, file, line));
 675         WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER, file, line, NULL);
 676
 677         v = RW_READ_VALUE(rw);
 678         if (__predict_false(LOCKSTAT_PROFILE_ENABLED(rw__acquire) ||
 679             !__rw_rlock_try(rw, td, &v, true LOCK_FILE_LINE_ARG)))
 680                 __rw_rlock_hard(rw, td, v LOCK_FILE_LINE_ARG);
 681         else
 682                 lock_profile_obtain_lock_success(&rw->lock_object, false, 0, 0,
 683                     file, line);
 684
 685         LOCK_LOG_LOCK("RLOCK", &rw->lock_object, 0, 0, file, line);
 686         WITNESS_LOCK(&rw->lock_object, 0, file, line);
 687         TD_LOCKS_INC(curthread);
 688 }
 689
 690 void
 691 __rw_rlock(volatile uintptr_t *c, const char *file, int line)
 692 {
 693         struct rwlock *rw;
 694
 695         rw = rwlock2rw(c);
 696         __rw_rlock_int(rw LOCK_FILE_LINE_ARG);
 697 }
 698
 699 int
 700 __rw_try_rlock_int(struct rwlock *rw LOCK_FILE_LINE_ARG_DEF)
 701 {
 702         uintptr_t x;
 703
 704         if (SCHEDULER_STOPPED())
 705                 return (1);
 706
 707         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
 708             ("rw_try_rlock() by idle thread %p on rwlock %s @ %s:%d",
 709             curthread, rw->lock_object.lo_name, file, line));
 710
 711         x = rw->rw_lock;
 712         for (;;) {
 713                 KASSERT(rw->rw_lock != RW_DESTROYED,
 714                     ("rw_try_rlock() of destroyed rwlock @ %s:%d", file, line));
 715                 if (!(x & RW_LOCK_READ))
 716                         break;
 717                 if (atomic_fcmpset_acq_ptr(&rw->rw_lock, &x, x + RW_ONE_READER)) {
 718                         LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 1, file,
 719                             line);
 720                         WITNESS_LOCK(&rw->lock_object, LOP_TRYLOCK, file, line);
 721                         LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire,
 722                             rw, 0, 0, file, line, LOCKSTAT_READER);
 723                         TD_LOCKS_INC(curthread);
 724                         curthread->td_rw_rlocks++;
 725                         return (1);
 726                 }
 727         }
 728
 729         LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 0, file, line);
 730         return (0);
 731 }
 732
 733 int
 734 __rw_try_rlock(volatile uintptr_t *c, const char *file, int line)
 735 {
 736         struct rwlock *rw;
 737
 738         rw = rwlock2rw(c);
 739         return (__rw_try_rlock_int(rw LOCK_FILE_LINE_ARG));
 740 }
 741
 742 static bool __always_inline
 743 __rw_runlock_try(struct rwlock *rw, struct thread *td, uintptr_t *vp)
 744 {
 745
 746         for (;;) {
 747                 if (RW_READERS(*vp) > 1 || !(*vp & RW_LOCK_WAITERS)) {
 748                         if (atomic_fcmpset_rel_ptr(&rw->rw_lock, vp,
 749                             *vp - RW_ONE_READER)) {
 750                                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 751                                         CTR4(KTR_LOCK,
 752                                             "%s: %p succeeded %p -> %p",
 753                                             __func__, rw, (void *)*vp,
 754                                             (void *)(*vp - RW_ONE_READER));
 755                                 td->td_rw_rlocks--;
 756                                 return (true);
 757                         }
 758                         continue;
 759                 }
 760                 break;
 761         }
 762         return (false);
 763 }
 764
 765 static void __noinline
 766 __rw_runlock_hard(struct rwlock *rw, struct thread *td, uintptr_t v
 767     LOCK_FILE_LINE_ARG_DEF)
 768 {
 769         struct turnstile *ts;
 770         uintptr_t setv, queue;
 771
 772         if (SCHEDULER_STOPPED())
 773                 return;
 774
 775         if (__rw_runlock_try(rw, td, &v))
 776                 goto out_lockstat;
 777
 778         /*
 779          * Ok, we know we have waiters and we think we are the
 780          * last reader, so grab the turnstile lock.
 781          */
 782         turnstile_chain_lock(&rw->lock_object);
 783         v = RW_READ_VALUE(rw);
 784         for (;;) {
 785                 if (__rw_runlock_try(rw, td, &v))
 786                         break;
 787
 788                 MPASS(v & RW_LOCK_WAITERS);
 789
 790                 /*
 791                  * Try to drop our lock leaving the lock in a unlocked
 792                  * state.
 793                  *
 794                  * If you wanted to do explicit lock handoff you'd have to
 795                  * do it here.  You'd also want to use turnstile_signal()
 796                  * and you'd have to handle the race where a higher
 797                  * priority thread blocks on the write lock before the
 798                  * thread you wakeup actually runs and have the new thread
 799                  * "steal" the lock.  For now it's a lot simpler to just
 800                  * wakeup all of the waiters.
 801                  *
 802                  * As above, if we fail, then another thread might have
 803                  * acquired a read lock, so drop the turnstile lock and
 804                  * restart.
 805                  */
 806                 setv = RW_UNLOCKED;
 807                 queue = TS_SHARED_QUEUE;
 808                 if (v & RW_LOCK_WRITE_WAITERS) {
 809                         queue = TS_EXCLUSIVE_QUEUE;
 810                         setv |= (v & RW_LOCK_READ_WAITERS);
 811                 }
 812                 setv |= (v & RW_LOCK_WRITE_SPINNER);
 813                 if (!atomic_fcmpset_rel_ptr(&rw->rw_lock, &v, setv))
 814                         continue;
 815                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 816                         CTR2(KTR_LOCK, "%s: %p last succeeded with waiters",
 817                             __func__, rw);
 818
 819                 /*
 820                  * Ok.  The lock is released and all that's left is to
 821                  * wake up the waiters.  Note that the lock might not be
 822                  * free anymore, but in that case the writers will just
 823                  * block again if they run before the new lock holder(s)
 824                  * release the lock.
 825                  */
 826                 ts = turnstile_lookup(&rw->lock_object);
 827                 MPASS(ts != NULL);
 828                 turnstile_broadcast(ts, queue);
 829                 turnstile_unpend(ts);
 830                 td->td_rw_rlocks--;
 831                 break;
 832         }
 833         turnstile_chain_unlock(&rw->lock_object);
 834 out_lockstat:
 835         LOCKSTAT_PROFILE_RELEASE_RWLOCK(rw__release, rw, LOCKSTAT_READER);
 836 }
 837
 838 void
 839 _rw_runlock_cookie_int(struct rwlock *rw LOCK_FILE_LINE_ARG_DEF)
 840 {
 841         struct thread *td;
 842         uintptr_t v;
 843
 844         KASSERT(rw->rw_lock != RW_DESTROYED,
 845             ("rw_runlock() of destroyed rwlock @ %s:%d", file, line));
 846         __rw_assert(&rw->rw_lock, RA_RLOCKED, file, line);
 847         WITNESS_UNLOCK(&rw->lock_object, 0, file, line);
 848         LOCK_LOG_LOCK("RUNLOCK", &rw->lock_object, 0, 0, file, line);
 849
 850         td = curthread;
 851         v = RW_READ_VALUE(rw);
 852
 853         if (__predict_false(LOCKSTAT_PROFILE_ENABLED(rw__release) ||
 854             !__rw_runlock_try(rw, td, &v)))
 855                 __rw_runlock_hard(rw, td, v LOCK_FILE_LINE_ARG);
 856         else
 857                 lock_profile_release_lock(&rw->lock_object, false);
 858
 859         TD_LOCKS_DEC(curthread);
 860 }
 861
 862 void
 863 _rw_runlock_cookie(volatile uintptr_t *c, const char *file, int line)
 864 {
 865         struct rwlock *rw;
 866
 867         rw = rwlock2rw(c);
 868         _rw_runlock_cookie_int(rw LOCK_FILE_LINE_ARG);
 869 }
 870
 871 #ifdef ADAPTIVE_RWLOCKS
 872 static inline void
 873 rw_drop_critical(uintptr_t v, bool *in_critical, int *extra_work)
 874 {
 875
 876         if (v & RW_LOCK_WRITE_SPINNER)
 877                 return;
 878         if (*in_critical) {
 879                 critical_exit();
 880                 *in_critical = false;
 881                 (*extra_work)--;
 882         }
 883 }
 884 #else
 885 #define rw_drop_critical(v, in_critical, extra_work) do { } while (0)
 886 #endif
 887
 888 /*
 889  * This function is called when we are unable to obtain a write lock on the
 890  * first try.  This means that at least one other thread holds either a
 891  * read or write lock.
 892  */
 893 void
 894 __rw_wlock_hard(volatile uintptr_t *c, uintptr_t v LOCK_FILE_LINE_ARG_DEF)
 895 {
 896         uintptr_t tid;
 897         struct rwlock *rw;
 898         struct turnstile *ts;
 899         struct thread *owner;
 900 #ifdef ADAPTIVE_RWLOCKS
 901         int spintries = 0;
 902         int i, n;
 903         enum { READERS, WRITER } sleep_reason = READERS;
 904         bool in_critical = false;
 905 #endif
 906         uintptr_t setv;
 907 #ifdef LOCK_PROFILING
 908         uint64_t waittime = 0;
 909         int contested = 0;
 910 #endif
 911 #if defined(ADAPTIVE_RWLOCKS) || defined(KDTRACE_HOOKS)
 912         struct lock_delay_arg lda;
 913 #endif
 914 #ifdef KDTRACE_HOOKS
 915         u_int sleep_cnt = 0;
 916         int64_t sleep_time = 0;
 917         int64_t all_time = 0;
 918 #endif
 919 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
 920         uintptr_t state = 0;
 921         int doing_lockprof = 0;
 922 #endif
 923         int extra_work = 0;
 924
 925         tid = (uintptr_t)curthread;
 926         rw = rwlock2rw(c);
 927
 928 #ifdef KDTRACE_HOOKS
 929         if (LOCKSTAT_PROFILE_ENABLED(rw__acquire)) {
 930                 while (v == RW_UNLOCKED) {
 931                         if (_rw_write_lock_fetch(rw, &v, tid))
 932                                 goto out_lockstat;
 933                 }
 934                 extra_work = 1;
 935                 doing_lockprof = 1;
 936                 all_time -= lockstat_nsecs(&rw->lock_object);
 937                 state = v;
 938         }
 939 #endif
 940 #ifdef LOCK_PROFILING
 941         extra_work = 1;
 942         doing_lockprof = 1;
 943         state = v;
 944 #endif
 945
 946         if (SCHEDULER_STOPPED())
 947                 return;
 948
 949         if (__predict_false(v == RW_UNLOCKED))
 950                 v = RW_READ_VALUE(rw);
 951
 952         if (__predict_false(lv_rw_wowner(v) == (struct thread *)tid)) {
 953                 KASSERT(rw->lock_object.lo_flags & LO_RECURSABLE,
 954                     ("%s: recursing but non-recursive rw %s @ %s:%d\n",
 955                     __func__, rw->lock_object.lo_name, file, line));
 956                 rw->rw_recurse++;
 957                 atomic_set_ptr(&rw->rw_lock, RW_LOCK_WRITER_RECURSED);
 958                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 959                         CTR2(KTR_LOCK, "%s: %p recursing", __func__, rw);
 960                 return;
 961         }
 962
 963         if (LOCK_LOG_TEST(&rw->lock_object, 0))
 964                 CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__,
 965                     rw->lock_object.lo_name, (void *)rw->rw_lock, file, line);
 966
 967 #if defined(ADAPTIVE_RWLOCKS)
 968         lock_delay_arg_init(&lda, &rw_delay);
 969 #elif defined(KDTRACE_HOOKS)
 970         lock_delay_arg_init_noadapt(&lda);
 971 #endif
 972
 973 #ifdef HWPMC_HOOKS
 974         PMC_SOFT_CALL( , , lock, failed);
 975 #endif
 976         lock_profile_obtain_lock_failed(&rw->lock_object, false,
 977             &contested, &waittime);
 978
 979         for (;;) {
 980                 if (v == RW_UNLOCKED) {
 981                         if (_rw_write_lock_fetch(rw, &v, tid))
 982                                 break;
 983                         continue;
 984                 }
 985 #ifdef KDTRACE_HOOKS
 986                 lda.spin_cnt++;
 987 #endif
 988
 989 #ifdef ADAPTIVE_RWLOCKS
 990                 if (v == (RW_LOCK_READ | RW_LOCK_WRITE_SPINNER)) {
 991                         if (atomic_fcmpset_acq_ptr(&rw->rw_lock, &v, tid))
 992                                 break;
 993                         continue;
 994                 }
 995
 996                 /*
 997                  * If the lock is write locked and the owner is
 998                  * running on another CPU, spin until the owner stops
 999                  * running or the state of the lock changes.
1000                  */
1001                 if (!(v & RW_LOCK_READ)) {
1002                         rw_drop_critical(v, &in_critical, &extra_work);
1003                         sleep_reason = WRITER;
1004                         owner = lv_rw_wowner(v);
1005                         if (!TD_IS_RUNNING(owner))
1006                                 goto ts;
1007                         if (LOCK_LOG_TEST(&rw->lock_object, 0))
1008                                 CTR3(KTR_LOCK, "%s: spinning on %p held by %p",
1009                                     __func__, rw, owner);
1010                         KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
1011                             "spinning", "lockname:\"%s\"",
1012                             rw->lock_object.lo_name);
1013                         do {
1014                                 lock_delay(&lda);
1015                                 v = RW_READ_VALUE(rw);
1016                                 owner = lv_rw_wowner(v);
1017                         } while (owner != NULL && TD_IS_RUNNING(owner));
1018                         KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
1019                             "running");
1020                         continue;
1021                 } else if (RW_READERS(v) > 0) {
1022                         sleep_reason = READERS;
1023                         if (spintries == rowner_retries)
1024                                 goto ts;
1025                         if (!(v & RW_LOCK_WRITE_SPINNER)) {
1026                                 if (!in_critical) {
1027                                         critical_enter();
1028                                         in_critical = true;
1029                                         extra_work++;
1030                                 }
1031                                 if (!atomic_fcmpset_ptr(&rw->rw_lock, &v,
1032                                     v | RW_LOCK_WRITE_SPINNER)) {
1033                                         critical_exit();
1034                                         in_critical = false;
1035                                         extra_work--;
1036                                         continue;
1037                                 }
1038                         }
1039                         spintries++;
1040                         KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
1041                             "spinning", "lockname:\"%s\"",
1042                             rw->lock_object.lo_name);
1043                         n = RW_READERS(v);
1044                         for (i = 0; i < rowner_loops; i += n) {
1045                                 lock_delay_spin(n);
1046                                 v = RW_READ_VALUE(rw);
1047                                 if (!(v & RW_LOCK_WRITE_SPINNER))
1048                                         break;
1049                                 if (!(v & RW_LOCK_READ))
1050                                         break;
1051                                 n = RW_READERS(v);
1052                                 if (n == 0)
1053                                         break;
1054                         }
1055 #ifdef KDTRACE_HOOKS
1056                         lda.spin_cnt += i;
1057 #endif
1058                         KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
1059                             "running");
1060                         if (i < rowner_loops)
1061                                 continue;
1062                 }
1063 ts:
1064 #endif
1065                 ts = turnstile_trywait(&rw->lock_object);
1066                 v = RW_READ_VALUE(rw);
1067 retry_ts:
1068                 owner = lv_rw_wowner(v);
1069
1070 #ifdef ADAPTIVE_RWLOCKS
1071                 /*
1072                  * The current lock owner might have started executing
1073                  * on another CPU (or the lock could have changed
1074                  * owners) while we were waiting on the turnstile
1075                  * chain lock.  If so, drop the turnstile lock and try
1076                  * again.
1077                  */
1078                 if (owner != NULL) {
1079                         if (TD_IS_RUNNING(owner)) {
1080                                 turnstile_cancel(ts);
1081                                 rw_drop_critical(v, &in_critical, &extra_work);
1082                                 continue;
1083                         }
1084                 } else if (RW_READERS(v) > 0 && sleep_reason == WRITER) {
1085                         turnstile_cancel(ts);
1086                         rw_drop_critical(v, &in_critical, &extra_work);
1087                         continue;
1088                 }
1089 #endif
1090                 /*
1091                  * Check for the waiters flags about this rwlock.
1092                  * If the lock was released, without maintain any pending
1093                  * waiters queue, simply try to acquire it.
1094                  * If a pending waiters queue is present, claim the lock
1095                  * ownership and maintain the pending queue.
1096                  */
1097                 setv = v & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER);
1098                 if ((v & ~setv) == RW_UNLOCKED) {
1099                         setv &= ~RW_LOCK_WRITE_SPINNER;
1100                         if (atomic_fcmpset_acq_ptr(&rw->rw_lock, &v, tid | setv)) {
1101                                 if (setv)
1102                                         turnstile_claim(ts);
1103                                 else
1104                                         turnstile_cancel(ts);
1105                                 break;
1106                         }
1107                         goto retry_ts;
1108                 }
1109
1110 #ifdef ADAPTIVE_RWLOCKS
1111                 if (in_critical) {
1112                         if ((v & RW_LOCK_WRITE_SPINNER) ||
1113                             !((v & RW_LOCK_WRITE_WAITERS))) {
1114                                 setv = v & ~RW_LOCK_WRITE_SPINNER;
1115                                 setv |= RW_LOCK_WRITE_WAITERS;
1116                                 if (!atomic_fcmpset_ptr(&rw->rw_lock, &v, setv))
1117                                         goto retry_ts;
1118                         }
1119                         critical_exit();
1120                         in_critical = false;
1121                         extra_work--;
1122                 } else {
1123 #endif
1124                         /*
1125                          * If the RW_LOCK_WRITE_WAITERS flag isn't set, then try to
1126                          * set it.  If we fail to set it, then loop back and try
1127                          * again.
1128                          */
1129                         if (!(v & RW_LOCK_WRITE_WAITERS)) {
1130                                 if (!atomic_fcmpset_ptr(&rw->rw_lock, &v,
1131                                     v | RW_LOCK_WRITE_WAITERS))
1132                                         goto retry_ts;
1133                                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
1134                                         CTR2(KTR_LOCK, "%s: %p set write waiters flag",
1135                                             __func__, rw);
1136                         }
1137 #ifdef ADAPTIVE_RWLOCKS
1138                 }
1139 #endif
1140                 /*
1141                  * We were unable to acquire the lock and the write waiters
1142                  * flag is set, so we must block on the turnstile.
1143                  */
1144                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
1145                         CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__,
1146                             rw);
1147 #ifdef KDTRACE_HOOKS
1148                 sleep_time -= lockstat_nsecs(&rw->lock_object);
1149 #endif
1150                 MPASS(owner == rw_owner(rw));
1151                 turnstile_wait(ts, owner, TS_EXCLUSIVE_QUEUE);
1152 #ifdef KDTRACE_HOOKS
1153                 sleep_time += lockstat_nsecs(&rw->lock_object);
1154                 sleep_cnt++;
1155 #endif
1156                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
1157                         CTR2(KTR_LOCK, "%s: %p resuming from turnstile",
1158                             __func__, rw);
1159 #ifdef ADAPTIVE_RWLOCKS
1160                 spintries = 0;
1161 #endif
1162                 v = RW_READ_VALUE(rw);
1163         }
1164         if (__predict_true(!extra_work))
1165                 return;
1166 #ifdef ADAPTIVE_RWLOCKS
1167         if (in_critical)
1168                 critical_exit();
1169 #endif
1170 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
1171         if (__predict_true(!doing_lockprof))
1172                 return;
1173 #endif
1174 #ifdef KDTRACE_HOOKS
1175         all_time += lockstat_nsecs(&rw->lock_object);
1176         if (sleep_time)
1177                 LOCKSTAT_RECORD4(rw__block, rw, sleep_time,
1178                     LOCKSTAT_WRITER, (state & RW_LOCK_READ) == 0,
1179                     (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
1180
1181         /* Record only the loops spinning and not sleeping. */
1182         if (lda.spin_cnt > sleep_cnt)
1183                 LOCKSTAT_RECORD4(rw__spin, rw, all_time - sleep_time,
1184                     LOCKSTAT_WRITER, (state & RW_LOCK_READ) == 0,
1185                     (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
1186 out_lockstat:
1187 #endif
1188         LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire, rw, contested,
1189             waittime, file, line, LOCKSTAT_WRITER);
1190 }
1191
1192 /*
1193  * This function is called if lockstat is active or the first try at releasing
1194  * a write lock failed.  The latter means that the lock is recursed or one of
1195  * the 2 waiter bits must be set indicating that at least one thread is waiting
1196  * on this lock.
1197  */
1198 void
1199 __rw_wunlock_hard(volatile uintptr_t *c, uintptr_t v LOCK_FILE_LINE_ARG_DEF)
1200 {
1201         struct rwlock *rw;
1202         struct turnstile *ts;
1203         uintptr_t tid, setv;
1204         int queue;
1205
1206         tid = (uintptr_t)curthread;
1207         if (SCHEDULER_STOPPED())
1208                 return;
1209
1210         rw = rwlock2rw(c);
1211         if (__predict_false(v == tid))
1212                 v = RW_READ_VALUE(rw);
1213
1214         if (v & RW_LOCK_WRITER_RECURSED) {
1215                 if (--(rw->rw_recurse) == 0)
1216                         atomic_clear_ptr(&rw->rw_lock, RW_LOCK_WRITER_RECURSED);
1217                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
1218                         CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, rw);
1219                 return;
1220         }
1221
1222         LOCKSTAT_PROFILE_RELEASE_RWLOCK(rw__release, rw, LOCKSTAT_WRITER);
1223         if (v == tid && _rw_write_unlock(rw, tid))
1224                 return;
1225
1226         KASSERT(rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS),
1227             ("%s: neither of the waiter flags are set", __func__));
1228
1229         if (LOCK_LOG_TEST(&rw->lock_object, 0))
1230                 CTR2(KTR_LOCK, "%s: %p contested", __func__, rw);
1231
1232         turnstile_chain_lock(&rw->lock_object);
1233
1234         /*
1235          * Use the same algo as sx locks for now.  Prefer waking up shared
1236          * waiters if we have any over writers.  This is probably not ideal.
1237          *
1238          * 'v' is the value we are going to write back to rw_lock.  If we
1239          * have waiters on both queues, we need to preserve the state of
1240          * the waiter flag for the queue we don't wake up.  For now this is
1241          * hardcoded for the algorithm mentioned above.
1242          *
1243          * In the case of both readers and writers waiting we wakeup the
1244          * readers but leave the RW_LOCK_WRITE_WAITERS flag set.  If a
1245          * new writer comes in before a reader it will claim the lock up
1246          * above.  There is probably a potential priority inversion in
1247          * there that could be worked around either by waking both queues
1248          * of waiters or doing some complicated lock handoff gymnastics.
1249          */
1250         setv = RW_UNLOCKED;
1251         v = RW_READ_VALUE(rw);
1252         queue = TS_SHARED_QUEUE;
1253         if (v & RW_LOCK_WRITE_WAITERS) {
1254                 queue = TS_EXCLUSIVE_QUEUE;
1255                 setv |= (v & RW_LOCK_READ_WAITERS);
1256         }
1257         atomic_store_rel_ptr(&rw->rw_lock, setv);
1258
1259         /* Wake up all waiters for the specific queue. */
1260         if (LOCK_LOG_TEST(&rw->lock_object, 0))
1261                 CTR3(KTR_LOCK, "%s: %p waking up %s waiters", __func__, rw,
1262                     queue == TS_SHARED_QUEUE ? "read" : "write");
1263
1264         ts = turnstile_lookup(&rw->lock_object);
1265         MPASS(ts != NULL);
1266         turnstile_broadcast(ts, queue);
1267         turnstile_unpend(ts);
1268         turnstile_chain_unlock(&rw->lock_object);
1269 }
1270
1271 /*
1272  * Attempt to do a non-blocking upgrade from a read lock to a write
1273  * lock.  This will only succeed if this thread holds a single read
1274  * lock.  Returns true if the upgrade succeeded and false otherwise.
1275  */
1276 int
1277 __rw_try_upgrade_int(struct rwlock *rw LOCK_FILE_LINE_ARG_DEF)
1278 {
1279         uintptr_t v, setv, tid;
1280         struct turnstile *ts;
1281         int success;
1282
1283         if (SCHEDULER_STOPPED())
1284                 return (1);
1285
1286         KASSERT(rw->rw_lock != RW_DESTROYED,
1287             ("rw_try_upgrade() of destroyed rwlock @ %s:%d", file, line));
1288         __rw_assert(&rw->rw_lock, RA_RLOCKED, file, line);
1289
1290         /*
1291          * Attempt to switch from one reader to a writer.  If there
1292          * are any write waiters, then we will have to lock the
1293          * turnstile first to prevent races with another writer
1294          * calling turnstile_wait() before we have claimed this
1295          * turnstile.  So, do the simple case of no waiters first.
1296          */
1297         tid = (uintptr_t)curthread;
1298         success = 0;
1299         v = RW_READ_VALUE(rw);
1300         for (;;) {
1301                 if (RW_READERS(v) > 1)
1302                         break;
1303                 if (!(v & RW_LOCK_WAITERS)) {
1304                         success = atomic_fcmpset_acq_ptr(&rw->rw_lock, &v, tid);
1305                         if (!success)
1306                                 continue;
1307                         break;
1308                 }
1309
1310                 /*
1311                  * Ok, we think we have waiters, so lock the turnstile.
1312                  */
1313                 ts = turnstile_trywait(&rw->lock_object);
1314                 v = RW_READ_VALUE(rw);
1315 retry_ts:
1316                 if (RW_READERS(v) > 1) {
1317                         turnstile_cancel(ts);
1318                         break;
1319                 }
1320                 /*
1321                  * Try to switch from one reader to a writer again.  This time
1322                  * we honor the current state of the waiters flags.
1323                  * If we obtain the lock with the flags set, then claim
1324                  * ownership of the turnstile.
1325                  */
1326                 setv = tid | (v & RW_LOCK_WAITERS);
1327                 success = atomic_fcmpset_ptr(&rw->rw_lock, &v, setv);
1328                 if (success) {
1329                         if (v & RW_LOCK_WAITERS)
1330                                 turnstile_claim(ts);
1331                         else
1332                                 turnstile_cancel(ts);
1333                         break;
1334                 }
1335                 goto retry_ts;
1336         }
1337         LOCK_LOG_TRY("WUPGRADE", &rw->lock_object, 0, success, file, line);
1338         if (success) {
1339                 curthread->td_rw_rlocks--;
1340                 WITNESS_UPGRADE(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
1341                     file, line);
1342                 LOCKSTAT_RECORD0(rw__upgrade, rw);
1343         }
1344         return (success);
1345 }
1346
1347 int
1348 __rw_try_upgrade(volatile uintptr_t *c, const char *file, int line)
1349 {
1350         struct rwlock *rw;
1351
1352         rw = rwlock2rw(c);
1353         return (__rw_try_upgrade_int(rw LOCK_FILE_LINE_ARG));
1354 }
1355
1356 /*
1357  * Downgrade a write lock into a single read lock.
1358  */
1359 void
1360 __rw_downgrade_int(struct rwlock *rw LOCK_FILE_LINE_ARG_DEF)
1361 {
1362         struct turnstile *ts;
1363         uintptr_t tid, v;
1364         int rwait, wwait;
1365
1366         if (SCHEDULER_STOPPED())
1367                 return;
1368
1369         KASSERT(rw->rw_lock != RW_DESTROYED,
1370             ("rw_downgrade() of destroyed rwlock @ %s:%d", file, line));
1371         __rw_assert(&rw->rw_lock, RA_WLOCKED | RA_NOTRECURSED, file, line);
1372 #ifndef INVARIANTS
1373         if (rw_recursed(rw))
1374                 panic("downgrade of a recursed lock");
1375 #endif
1376
1377         WITNESS_DOWNGRADE(&rw->lock_object, 0, file, line);
1378
1379         /*
1380          * Convert from a writer to a single reader.  First we handle
1381          * the easy case with no waiters.  If there are any waiters, we
1382          * lock the turnstile and "disown" the lock.
1383          */
1384         tid = (uintptr_t)curthread;
1385         if (atomic_cmpset_rel_ptr(&rw->rw_lock, tid, RW_READERS_LOCK(1)))
1386                 goto out;
1387
1388         /*
1389          * Ok, we think we have waiters, so lock the turnstile so we can
1390          * read the waiter flags without any races.
1391          */
1392         turnstile_chain_lock(&rw->lock_object);
1393         v = rw->rw_lock & RW_LOCK_WAITERS;
1394         rwait = v & RW_LOCK_READ_WAITERS;
1395         wwait = v & RW_LOCK_WRITE_WAITERS;
1396         MPASS(rwait | wwait);
1397
1398         /*
1399          * Downgrade from a write lock while preserving waiters flag
1400          * and give up ownership of the turnstile.
1401          */
1402         ts = turnstile_lookup(&rw->lock_object);
1403         MPASS(ts != NULL);
1404         if (!wwait)
1405                 v &= ~RW_LOCK_READ_WAITERS;
1406         atomic_store_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v);
1407         /*
1408          * Wake other readers if there are no writers pending.  Otherwise they
1409          * won't be able to acquire the lock anyway.
1410          */
1411         if (rwait && !wwait) {
1412                 turnstile_broadcast(ts, TS_SHARED_QUEUE);
1413                 turnstile_unpend(ts);
1414         } else
1415                 turnstile_disown(ts);
1416         turnstile_chain_unlock(&rw->lock_object);
1417 out:
1418         curthread->td_rw_rlocks++;
1419         LOCK_LOG_LOCK("WDOWNGRADE", &rw->lock_object, 0, 0, file, line);
1420         LOCKSTAT_RECORD0(rw__downgrade, rw);
1421 }
1422
1423 void
1424 __rw_downgrade(volatile uintptr_t *c, const char *file, int line)
1425 {
1426         struct rwlock *rw;
1427
1428         rw = rwlock2rw(c);
1429         __rw_downgrade_int(rw LOCK_FILE_LINE_ARG);
1430 }
1431
1432 #ifdef INVARIANT_SUPPORT
1433 #ifndef INVARIANTS
1434 #undef __rw_assert
1435 #endif
1436
1437 /*
1438  * In the non-WITNESS case, rw_assert() can only detect that at least
1439  * *some* thread owns an rlock, but it cannot guarantee that *this*
1440  * thread owns an rlock.
1441  */
1442 void
1443 __rw_assert(const volatile uintptr_t *c, int what, const char *file, int line)
1444 {
1445         const struct rwlock *rw;
1446
1447         if (SCHEDULER_STOPPED())
1448                 return;
1449
1450         rw = rwlock2rw(c);
1451
1452         switch (what) {
1453         case RA_LOCKED:
1454         case RA_LOCKED | RA_RECURSED:
1455         case RA_LOCKED | RA_NOTRECURSED:
1456         case RA_RLOCKED:
1457         case RA_RLOCKED | RA_RECURSED:
1458         case RA_RLOCKED | RA_NOTRECURSED:
1459 #ifdef WITNESS
1460                 witness_assert(&rw->lock_object, what, file, line);
1461 #else
1462                 /*
1463                  * If some other thread has a write lock or we have one
1464                  * and are asserting a read lock, fail.  Also, if no one
1465                  * has a lock at all, fail.
1466                  */
1467                 if (rw->rw_lock == RW_UNLOCKED ||
1468                     (!(rw->rw_lock & RW_LOCK_READ) && (what & RA_RLOCKED ||
1469                     rw_wowner(rw) != curthread)))
1470                         panic("Lock %s not %slocked @ %s:%d\n",
1471                             rw->lock_object.lo_name, (what & RA_RLOCKED) ?
1472                             "read " : "", file, line);
1473
1474                 if (!(rw->rw_lock & RW_LOCK_READ) && !(what & RA_RLOCKED)) {
1475                         if (rw_recursed(rw)) {
1476                                 if (what & RA_NOTRECURSED)
1477                                         panic("Lock %s recursed @ %s:%d\n",
1478                                             rw->lock_object.lo_name, file,
1479                                             line);
1480                         } else if (what & RA_RECURSED)
1481                                 panic("Lock %s not recursed @ %s:%d\n",
1482                                     rw->lock_object.lo_name, file, line);
1483                 }
1484 #endif
1485                 break;
1486         case RA_WLOCKED:
1487         case RA_WLOCKED | RA_RECURSED:
1488         case RA_WLOCKED | RA_NOTRECURSED:
1489                 if (rw_wowner(rw) != curthread)
1490                         panic("Lock %s not exclusively locked @ %s:%d\n",
1491                             rw->lock_object.lo_name, file, line);
1492                 if (rw_recursed(rw)) {
1493                         if (what & RA_NOTRECURSED)
1494                                 panic("Lock %s recursed @ %s:%d\n",
1495                                     rw->lock_object.lo_name, file, line);
1496                 } else if (what & RA_RECURSED)
1497                         panic("Lock %s not recursed @ %s:%d\n",
1498                             rw->lock_object.lo_name, file, line);
1499                 break;
1500         case RA_UNLOCKED:
1501 #ifdef WITNESS
1502                 witness_assert(&rw->lock_object, what, file, line);
1503 #else
1504                 /*
1505                  * If we hold a write lock fail.  We can't reliably check
1506                  * to see if we hold a read lock or not.
1507                  */
1508                 if (rw_wowner(rw) == curthread)
1509                         panic("Lock %s exclusively locked @ %s:%d\n",
1510                             rw->lock_object.lo_name, file, line);
1511 #endif
1512                 break;
1513         default:
1514                 panic("Unknown rw lock assertion: %d @ %s:%d", what, file,
1515                     line);
1516         }
1517 }
1518 #endif /* INVARIANT_SUPPORT */
1519
1520 #ifdef DDB
1521 void
1522 db_show_rwlock(const struct lock_object *lock)
1523 {
1524         const struct rwlock *rw;
1525         struct thread *td;
1526
1527         rw = (const struct rwlock *)lock;
1528
1529         db_printf(" state: ");
1530         if (rw->rw_lock == RW_UNLOCKED)
1531                 db_printf("UNLOCKED\n");
1532         else if (rw->rw_lock == RW_DESTROYED) {
1533                 db_printf("DESTROYED\n");
1534                 return;
1535         } else if (rw->rw_lock & RW_LOCK_READ)
1536                 db_printf("RLOCK: %ju locks\n",
1537                     (uintmax_t)(RW_READERS(rw->rw_lock)));
1538         else {
1539                 td = rw_wowner(rw);
1540                 db_printf("WLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1541                     td->td_tid, td->td_proc->p_pid, td->td_name);
1542                 if (rw_recursed(rw))
1543                         db_printf(" recursed: %u\n", rw->rw_recurse);
1544         }
1545         db_printf(" waiters: ");
1546         switch (rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS)) {
1547         case RW_LOCK_READ_WAITERS:
1548                 db_printf("readers\n");
1549                 break;
1550         case RW_LOCK_WRITE_WAITERS:
1551                 db_printf("writers\n");
1552                 break;
1553         case RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS:
1554                 db_printf("readers and writers\n");
1555                 break;
1556         default:
1557                 db_printf("none\n");
1558                 break;
1559         }
1560 }
1561
1562 #endif