sys/kern/kern_rwlock.c

   1 /*-
   2  * Copyright (c) 2006 John Baldwin <jhb@FreeBSD.org>
   3  * All rights reserved.
   4  *
   5  * Redistribution and use in source and binary forms, with or without
   6  * modification, are permitted provided that the following conditions
   7  * are met:
   8  * 1. Redistributions of source code must retain the above copyright
   9  *    notice, this list of conditions and the following disclaimer.
  10  * 2. Redistributions in binary form must reproduce the above copyright
  11  *    notice, this list of conditions and the following disclaimer in the
  12  *    documentation and/or other materials provided with the distribution.
  13  * 3. Neither the name of the author nor the names of any co-contributors
  14  *    may be used to endorse or promote products derived from this software
  15  *    without specific prior written permission.
  16  *
  17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
  18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  20  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
  21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  27  * SUCH DAMAGE.
  28  */
  29
  30 /*
  31  * Machine independent bits of reader/writer lock implementation.
  32  */
  33
  34 #include <sys/cdefs.h>
  35 __FBSDID("$FreeBSD$");
  36
  37 #include "opt_ddb.h"
  38 #include "opt_hwpmc_hooks.h"
  39 #include "opt_kdtrace.h"
  40 #include "opt_no_adaptive_rwlocks.h"
  41
  42 #include <sys/param.h>
  43 #include <sys/ktr.h>
  44 #include <sys/kernel.h>
  45 #include <sys/lock.h>
  46 #include <sys/mutex.h>
  47 #include <sys/proc.h>
  48 #include <sys/rwlock.h>
  49 #include <sys/sysctl.h>
  50 #include <sys/systm.h>
  51 #include <sys/turnstile.h>
  52
  53 #include <machine/cpu.h>
  54
  55 #if defined(SMP) && !defined(NO_ADAPTIVE_RWLOCKS)
  56 #define ADAPTIVE_RWLOCKS
  57 #endif
  58
  59 #ifdef HWPMC_HOOKS
  60 #include <sys/pmckern.h>
  61 PMC_SOFT_DECLARE( , , lock, failed);
  62 #endif
  63
  64 #ifdef ADAPTIVE_RWLOCKS
  65 #define ROWNER_RETRIES  10
  66 #define ROWNER_LOOPS    10000
  67 #endif
  68
  69 #ifdef DDB
  70 #include <ddb/ddb.h>
  71
  72 static void     db_show_rwlock(struct lock_object *lock);
  73 #endif
  74 static void     assert_rw(struct lock_object *lock, int what);
  75 static void     lock_rw(struct lock_object *lock, int how);
  76 #ifdef KDTRACE_HOOKS
  77 static int      owner_rw(struct lock_object *lock, struct thread **owner);
  78 #endif
  79 static int      unlock_rw(struct lock_object *lock);
  80
  81 struct lock_class lock_class_rw = {
  82         .lc_name = "rw",
  83         .lc_flags = LC_SLEEPLOCK | LC_RECURSABLE | LC_UPGRADABLE,
  84         .lc_assert = assert_rw,
  85 #ifdef DDB
  86         .lc_ddb_show = db_show_rwlock,
  87 #endif
  88         .lc_lock = lock_rw,
  89         .lc_unlock = unlock_rw,
  90 #ifdef KDTRACE_HOOKS
  91         .lc_owner = owner_rw,
  92 #endif
  93 };
  94
  95 /*
  96  * Return a pointer to the owning thread if the lock is write-locked or
  97  * NULL if the lock is unlocked or read-locked.
  98  */
  99 #define rw_wowner(rw)                                                   \
 100         ((rw)->rw_lock & RW_LOCK_READ ? NULL :                          \
 101             (struct thread *)RW_OWNER((rw)->rw_lock))
 102
 103 /*
 104  * Returns if a write owner is recursed.  Write ownership is not assured
 105  * here and should be previously checked.
 106  */
 107 #define rw_recursed(rw)         ((rw)->rw_recurse != 0)
 108
 109 /*
 110  * Return true if curthread helds the lock.
 111  */
 112 #define rw_wlocked(rw)          (rw_wowner((rw)) == curthread)
 113
 114 /*
 115  * Return a pointer to the owning thread for this lock who should receive
 116  * any priority lent by threads that block on this lock.  Currently this
 117  * is identical to rw_wowner().
 118  */
 119 #define rw_owner(rw)            rw_wowner(rw)
 120
 121 #ifndef INVARIANTS
 122 #define _rw_assert(rw, what, file, line)
 123 #endif
 124
 125 void
 126 assert_rw(struct lock_object *lock, int what)
 127 {
 128
 129         rw_assert((struct rwlock *)lock, what);
 130 }
 131
 132 void
 133 lock_rw(struct lock_object *lock, int how)
 134 {
 135         struct rwlock *rw;
 136
 137         rw = (struct rwlock *)lock;
 138         if (how)
 139                 rw_wlock(rw);
 140         else
 141                 rw_rlock(rw);
 142 }
 143
 144 int
 145 unlock_rw(struct lock_object *lock)
 146 {
 147         struct rwlock *rw;
 148
 149         rw = (struct rwlock *)lock;
 150         rw_assert(rw, RA_LOCKED | LA_NOTRECURSED);
 151         if (rw->rw_lock & RW_LOCK_READ) {
 152                 rw_runlock(rw);
 153                 return (0);
 154         } else {
 155                 rw_wunlock(rw);
 156                 return (1);
 157         }
 158 }
 159
 160 #ifdef KDTRACE_HOOKS
 161 int
 162 owner_rw(struct lock_object *lock, struct thread **owner)
 163 {
 164         struct rwlock *rw = (struct rwlock *)lock;
 165         uintptr_t x = rw->rw_lock;
 166
 167         *owner = rw_wowner(rw);
 168         return ((x & RW_LOCK_READ) != 0 ?  (RW_READERS(x) != 0) :
 169             (*owner != NULL));
 170 }
 171 #endif
 172
 173 void
 174 rw_init_flags(struct rwlock *rw, const char *name, int opts)
 175 {
 176         int flags;
 177
 178         MPASS((opts & ~(RW_DUPOK | RW_NOPROFILE | RW_NOWITNESS | RW_QUIET |
 179             RW_RECURSE)) == 0);
 180         ASSERT_ATOMIC_LOAD_PTR(rw->rw_lock,
 181             ("%s: rw_lock not aligned for %s: %p", __func__, name,
 182             &rw->rw_lock));
 183
 184         flags = LO_UPGRADABLE;
 185         if (opts & RW_DUPOK)
 186                 flags |= LO_DUPOK;
 187         if (opts & RW_NOPROFILE)
 188                 flags |= LO_NOPROFILE;
 189         if (!(opts & RW_NOWITNESS))
 190                 flags |= LO_WITNESS;
 191         if (opts & RW_RECURSE)
 192                 flags |= LO_RECURSABLE;
 193         if (opts & RW_QUIET)
 194                 flags |= LO_QUIET;
 195
 196         rw->rw_lock = RW_UNLOCKED;
 197         rw->rw_recurse = 0;
 198         lock_init(&rw->lock_object, &lock_class_rw, name, NULL, flags);
 199 }
 200
 201 void
 202 rw_destroy(struct rwlock *rw)
 203 {
 204
 205         KASSERT(rw->rw_lock == RW_UNLOCKED, ("rw lock %p not unlocked", rw));
 206         KASSERT(rw->rw_recurse == 0, ("rw lock %p still recursed", rw));
 207         rw->rw_lock = RW_DESTROYED;
 208         lock_destroy(&rw->lock_object);
 209 }
 210
 211 void
 212 rw_sysinit(void *arg)
 213 {
 214         struct rw_args *args = arg;
 215
 216         rw_init(args->ra_rw, args->ra_desc);
 217 }
 218
 219 void
 220 rw_sysinit_flags(void *arg)
 221 {
 222         struct rw_args_flags *args = arg;
 223
 224         rw_init_flags(args->ra_rw, args->ra_desc, args->ra_flags);
 225 }
 226
 227 int
 228 rw_wowned(struct rwlock *rw)
 229 {
 230
 231         return (rw_wowner(rw) == curthread);
 232 }
 233
 234 void
 235 _rw_wlock(struct rwlock *rw, const char *file, int line)
 236 {
 237
 238         if (SCHEDULER_STOPPED())
 239                 return;
 240         MPASS(curthread != NULL);
 241         KASSERT(rw->rw_lock != RW_DESTROYED,
 242             ("rw_wlock() of destroyed rwlock @ %s:%d", file, line));
 243         WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE, file,
 244             line, NULL);
 245         __rw_wlock(rw, curthread, file, line);
 246         LOCK_LOG_LOCK("WLOCK", &rw->lock_object, 0, rw->rw_recurse, file, line);
 247         WITNESS_LOCK(&rw->lock_object, LOP_EXCLUSIVE, file, line);
 248         curthread->td_locks++;
 249 }
 250
 251 int
 252 _rw_try_wlock(struct rwlock *rw, const char *file, int line)
 253 {
 254         int rval;
 255
 256         if (SCHEDULER_STOPPED())
 257                 return (1);
 258
 259         KASSERT(rw->rw_lock != RW_DESTROYED,
 260             ("rw_try_wlock() of destroyed rwlock @ %s:%d", file, line));
 261
 262         if (rw_wlocked(rw) &&
 263             (rw->lock_object.lo_flags & LO_RECURSABLE) != 0) {
 264                 rw->rw_recurse++;
 265                 rval = 1;
 266         } else
 267                 rval = atomic_cmpset_acq_ptr(&rw->rw_lock, RW_UNLOCKED,
 268                     (uintptr_t)curthread);
 269
 270         LOCK_LOG_TRY("WLOCK", &rw->lock_object, 0, rval, file, line);
 271         if (rval) {
 272                 WITNESS_LOCK(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
 273                     file, line);
 274                 curthread->td_locks++;
 275         }
 276         return (rval);
 277 }
 278
 279 void
 280 _rw_wunlock(struct rwlock *rw, const char *file, int line)
 281 {
 282
 283         if (SCHEDULER_STOPPED())
 284                 return;
 285         MPASS(curthread != NULL);
 286         KASSERT(rw->rw_lock != RW_DESTROYED,
 287             ("rw_wunlock() of destroyed rwlock @ %s:%d", file, line));
 288         _rw_assert(rw, RA_WLOCKED, file, line);
 289         curthread->td_locks--;
 290         WITNESS_UNLOCK(&rw->lock_object, LOP_EXCLUSIVE, file, line);
 291         LOCK_LOG_LOCK("WUNLOCK", &rw->lock_object, 0, rw->rw_recurse, file,
 292             line);
 293         if (!rw_recursed(rw))
 294                 LOCKSTAT_PROFILE_RELEASE_LOCK(LS_RW_WUNLOCK_RELEASE, rw);
 295         __rw_wunlock(rw, curthread, file, line);
 296 }
 297 /*
 298  * Determines whether a new reader can acquire a lock.  Succeeds if the
 299  * reader already owns a read lock and the lock is locked for read to
 300  * prevent deadlock from reader recursion.  Also succeeds if the lock
 301  * is unlocked and has no writer waiters or spinners.  Failing otherwise
 302  * prioritizes writers before readers.
 303  */
 304 #define RW_CAN_READ(_rw)                                                \
 305     ((curthread->td_rw_rlocks && (_rw) & RW_LOCK_READ) || ((_rw) &      \
 306     (RW_LOCK_READ | RW_LOCK_WRITE_WAITERS | RW_LOCK_WRITE_SPINNER)) ==  \
 307     RW_LOCK_READ)
 308
 309 void
 310 _rw_rlock(struct rwlock *rw, const char *file, int line)
 311 {
 312         struct turnstile *ts;
 313 #ifdef ADAPTIVE_RWLOCKS
 314         volatile struct thread *owner;
 315         int spintries = 0;
 316         int i;
 317 #endif
 318 #ifdef LOCK_PROFILING
 319         uint64_t waittime = 0;
 320         int contested = 0;
 321 #endif
 322         uintptr_t v;
 323 #ifdef KDTRACE_HOOKS
 324         uint64_t spin_cnt = 0;
 325         uint64_t sleep_cnt = 0;
 326         int64_t sleep_time = 0;
 327 #endif
 328
 329         if (SCHEDULER_STOPPED())
 330                 return;
 331
 332         KASSERT(rw->rw_lock != RW_DESTROYED,
 333             ("rw_rlock() of destroyed rwlock @ %s:%d", file, line));
 334         KASSERT(rw_wowner(rw) != curthread,
 335             ("%s (%s): wlock already held @ %s:%d", __func__,
 336             rw->lock_object.lo_name, file, line));
 337         WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER, file, line, NULL);
 338
 339         for (;;) {
 340 #ifdef KDTRACE_HOOKS
 341                 spin_cnt++;
 342 #endif
 343                 /*
 344                  * Handle the easy case.  If no other thread has a write
 345                  * lock, then try to bump up the count of read locks.  Note
 346                  * that we have to preserve the current state of the
 347                  * RW_LOCK_WRITE_WAITERS flag.  If we fail to acquire a
 348                  * read lock, then rw_lock must have changed, so restart
 349                  * the loop.  Note that this handles the case of a
 350                  * completely unlocked rwlock since such a lock is encoded
 351                  * as a read lock with no waiters.
 352                  */
 353                 v = rw->rw_lock;
 354                 if (RW_CAN_READ(v)) {
 355                         /*
 356                          * The RW_LOCK_READ_WAITERS flag should only be set
 357                          * if the lock has been unlocked and write waiters
 358                          * were present.
 359                          */
 360                         if (atomic_cmpset_acq_ptr(&rw->rw_lock, v,
 361                             v + RW_ONE_READER)) {
 362                                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 363                                         CTR4(KTR_LOCK,
 364                                             "%s: %p succeed %p -> %p", __func__,
 365                                             rw, (void *)v,
 366                                             (void *)(v + RW_ONE_READER));
 367                                 break;
 368                         }
 369                         continue;
 370                 }
 371 #ifdef HWPMC_HOOKS
 372                 PMC_SOFT_CALL( , , lock, failed);
 373 #endif
 374                 lock_profile_obtain_lock_failed(&rw->lock_object,
 375                     &contested, &waittime);
 376
 377 #ifdef ADAPTIVE_RWLOCKS
 378                 /*
 379                  * If the owner is running on another CPU, spin until
 380                  * the owner stops running or the state of the lock
 381                  * changes.
 382                  */
 383                 if ((v & RW_LOCK_READ) == 0) {
 384                         owner = (struct thread *)RW_OWNER(v);
 385                         if (TD_IS_RUNNING(owner)) {
 386                                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 387                                         CTR3(KTR_LOCK,
 388                                             "%s: spinning on %p held by %p",
 389                                             __func__, rw, owner);
 390                                 while ((struct thread*)RW_OWNER(rw->rw_lock) ==
 391                                     owner && TD_IS_RUNNING(owner)) {
 392                                         cpu_spinwait();
 393 #ifdef KDTRACE_HOOKS
 394                                         spin_cnt++;
 395 #endif
 396                                 }
 397                                 continue;
 398                         }
 399                 } else if (spintries < ROWNER_RETRIES) {
 400                         spintries++;
 401                         for (i = 0; i < ROWNER_LOOPS; i++) {
 402                                 v = rw->rw_lock;
 403                                 if ((v & RW_LOCK_READ) == 0 || RW_CAN_READ(v))
 404                                         break;
 405                                 cpu_spinwait();
 406                         }
 407                         if (i != ROWNER_LOOPS)
 408                                 continue;
 409                 }
 410 #endif
 411
 412                 /*
 413                  * Okay, now it's the hard case.  Some other thread already
 414                  * has a write lock or there are write waiters present,
 415                  * acquire the turnstile lock so we can begin the process
 416                  * of blocking.
 417                  */
 418                 ts = turnstile_trywait(&rw->lock_object);
 419
 420                 /*
 421                  * The lock might have been released while we spun, so
 422                  * recheck its state and restart the loop if needed.
 423                  */
 424                 v = rw->rw_lock;
 425                 if (RW_CAN_READ(v)) {
 426                         turnstile_cancel(ts);
 427                         continue;
 428                 }
 429
 430 #ifdef ADAPTIVE_RWLOCKS
 431                 /*
 432                  * The current lock owner might have started executing
 433                  * on another CPU (or the lock could have changed
 434                  * owners) while we were waiting on the turnstile
 435                  * chain lock.  If so, drop the turnstile lock and try
 436                  * again.
 437                  */
 438                 if ((v & RW_LOCK_READ) == 0) {
 439                         owner = (struct thread *)RW_OWNER(v);
 440                         if (TD_IS_RUNNING(owner)) {
 441                                 turnstile_cancel(ts);
 442                                 continue;
 443                         }
 444                 }
 445 #endif
 446
 447                 /*
 448                  * The lock is held in write mode or it already has waiters.
 449                  */
 450                 MPASS(!RW_CAN_READ(v));
 451
 452                 /*
 453                  * If the RW_LOCK_READ_WAITERS flag is already set, then
 454                  * we can go ahead and block.  If it is not set then try
 455                  * to set it.  If we fail to set it drop the turnstile
 456                  * lock and restart the loop.
 457                  */
 458                 if (!(v & RW_LOCK_READ_WAITERS)) {
 459                         if (!atomic_cmpset_ptr(&rw->rw_lock, v,
 460                             v | RW_LOCK_READ_WAITERS)) {
 461                                 turnstile_cancel(ts);
 462                                 continue;
 463                         }
 464                         if (LOCK_LOG_TEST(&rw->lock_object, 0))
 465                                 CTR2(KTR_LOCK, "%s: %p set read waiters flag",
 466                                     __func__, rw);
 467                 }
 468
 469                 /*
 470                  * We were unable to acquire the lock and the read waiters
 471                  * flag is set, so we must block on the turnstile.
 472                  */
 473                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 474                         CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__,
 475                             rw);
 476 #ifdef KDTRACE_HOOKS
 477                 sleep_time -= lockstat_nsecs();
 478 #endif
 479                 turnstile_wait(ts, rw_owner(rw), TS_SHARED_QUEUE);
 480 #ifdef KDTRACE_HOOKS
 481                 sleep_time += lockstat_nsecs();
 482                 sleep_cnt++;
 483 #endif
 484                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 485                         CTR2(KTR_LOCK, "%s: %p resuming from turnstile",
 486                             __func__, rw);
 487         }
 488
 489         /*
 490          * TODO: acquire "owner of record" here.  Here be turnstile dragons
 491          * however.  turnstiles don't like owners changing between calls to
 492          * turnstile_wait() currently.
 493          */
 494         LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_RW_RLOCK_ACQUIRE, rw, contested,
 495             waittime, file, line);
 496         LOCK_LOG_LOCK("RLOCK", &rw->lock_object, 0, 0, file, line);
 497         WITNESS_LOCK(&rw->lock_object, 0, file, line);
 498         curthread->td_locks++;
 499         curthread->td_rw_rlocks++;
 500 #ifdef KDTRACE_HOOKS
 501         if (sleep_time)
 502                 LOCKSTAT_RECORD1(LS_RW_RLOCK_BLOCK, rw, sleep_time);
 503
 504         /*
 505          * Record only the loops spinning and not sleeping.
 506          */
 507         if (spin_cnt > sleep_cnt)
 508                 LOCKSTAT_RECORD1(LS_RW_RLOCK_SPIN, rw, (spin_cnt - sleep_cnt));
 509 #endif
 510 }
 511
 512 int
 513 _rw_try_rlock(struct rwlock *rw, const char *file, int line)
 514 {
 515         uintptr_t x;
 516
 517         if (SCHEDULER_STOPPED())
 518                 return (1);
 519
 520         for (;;) {
 521                 x = rw->rw_lock;
 522                 KASSERT(rw->rw_lock != RW_DESTROYED,
 523                     ("rw_try_rlock() of destroyed rwlock @ %s:%d", file, line));
 524                 if (!(x & RW_LOCK_READ))
 525                         break;
 526                 if (atomic_cmpset_acq_ptr(&rw->rw_lock, x, x + RW_ONE_READER)) {
 527                         LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 1, file,
 528                             line);
 529                         WITNESS_LOCK(&rw->lock_object, LOP_TRYLOCK, file, line);
 530                         curthread->td_locks++;
 531                         curthread->td_rw_rlocks++;
 532                         return (1);
 533                 }
 534         }
 535
 536         LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 0, file, line);
 537         return (0);
 538 }
 539
 540 void
 541 _rw_runlock(struct rwlock *rw, const char *file, int line)
 542 {
 543         struct turnstile *ts;
 544         uintptr_t x, v, queue;
 545
 546         if (SCHEDULER_STOPPED())
 547                 return;
 548
 549         KASSERT(rw->rw_lock != RW_DESTROYED,
 550             ("rw_runlock() of destroyed rwlock @ %s:%d", file, line));
 551         _rw_assert(rw, RA_RLOCKED, file, line);
 552         curthread->td_locks--;
 553         curthread->td_rw_rlocks--;
 554         WITNESS_UNLOCK(&rw->lock_object, 0, file, line);
 555         LOCK_LOG_LOCK("RUNLOCK", &rw->lock_object, 0, 0, file, line);
 556
 557         /* TODO: drop "owner of record" here. */
 558
 559         for (;;) {
 560                 /*
 561                  * See if there is more than one read lock held.  If so,
 562                  * just drop one and return.
 563                  */
 564                 x = rw->rw_lock;
 565                 if (RW_READERS(x) > 1) {
 566                         if (atomic_cmpset_rel_ptr(&rw->rw_lock, x,
 567                             x - RW_ONE_READER)) {
 568                                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 569                                         CTR4(KTR_LOCK,
 570                                             "%s: %p succeeded %p -> %p",
 571                                             __func__, rw, (void *)x,
 572                                             (void *)(x - RW_ONE_READER));
 573                                 break;
 574                         }
 575                         continue;
 576                 }
 577                 /*
 578                  * If there aren't any waiters for a write lock, then try
 579                  * to drop it quickly.
 580                  */
 581                 if (!(x & RW_LOCK_WAITERS)) {
 582                         MPASS((x & ~RW_LOCK_WRITE_SPINNER) ==
 583                             RW_READERS_LOCK(1));
 584                         if (atomic_cmpset_rel_ptr(&rw->rw_lock, x,
 585                             RW_UNLOCKED)) {
 586                                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 587                                         CTR2(KTR_LOCK, "%s: %p last succeeded",
 588                                             __func__, rw);
 589                                 break;
 590                         }
 591                         continue;
 592                 }
 593                 /*
 594                  * Ok, we know we have waiters and we think we are the
 595                  * last reader, so grab the turnstile lock.
 596                  */
 597                 turnstile_chain_lock(&rw->lock_object);
 598                 v = rw->rw_lock & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER);
 599                 MPASS(v & RW_LOCK_WAITERS);
 600
 601                 /*
 602                  * Try to drop our lock leaving the lock in a unlocked
 603                  * state.
 604                  *
 605                  * If you wanted to do explicit lock handoff you'd have to
 606                  * do it here.  You'd also want to use turnstile_signal()
 607                  * and you'd have to handle the race where a higher
 608                  * priority thread blocks on the write lock before the
 609                  * thread you wakeup actually runs and have the new thread
 610                  * "steal" the lock.  For now it's a lot simpler to just
 611                  * wakeup all of the waiters.
 612                  *
 613                  * As above, if we fail, then another thread might have
 614                  * acquired a read lock, so drop the turnstile lock and
 615                  * restart.
 616                  */
 617                 x = RW_UNLOCKED;
 618                 if (v & RW_LOCK_WRITE_WAITERS) {
 619                         queue = TS_EXCLUSIVE_QUEUE;
 620                         x |= (v & RW_LOCK_READ_WAITERS);
 621                 } else
 622                         queue = TS_SHARED_QUEUE;
 623                 if (!atomic_cmpset_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v,
 624                     x)) {
 625                         turnstile_chain_unlock(&rw->lock_object);
 626                         continue;
 627                 }
 628                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 629                         CTR2(KTR_LOCK, "%s: %p last succeeded with waiters",
 630                             __func__, rw);
 631
 632                 /*
 633                  * Ok.  The lock is released and all that's left is to
 634                  * wake up the waiters.  Note that the lock might not be
 635                  * free anymore, but in that case the writers will just
 636                  * block again if they run before the new lock holder(s)
 637                  * release the lock.
 638                  */
 639                 ts = turnstile_lookup(&rw->lock_object);
 640                 MPASS(ts != NULL);
 641                 turnstile_broadcast(ts, queue);
 642                 turnstile_unpend(ts, TS_SHARED_LOCK);
 643                 turnstile_chain_unlock(&rw->lock_object);
 644                 break;
 645         }
 646         LOCKSTAT_PROFILE_RELEASE_LOCK(LS_RW_RUNLOCK_RELEASE, rw);
 647 }
 648
 649 /*
 650  * This function is called when we are unable to obtain a write lock on the
 651  * first try.  This means that at least one other thread holds either a
 652  * read or write lock.
 653  */
 654 void
 655 _rw_wlock_hard(struct rwlock *rw, uintptr_t tid, const char *file, int line)
 656 {
 657         struct turnstile *ts;
 658 #ifdef ADAPTIVE_RWLOCKS
 659         volatile struct thread *owner;
 660         int spintries = 0;
 661         int i;
 662 #endif
 663         uintptr_t v, x;
 664 #ifdef LOCK_PROFILING
 665         uint64_t waittime = 0;
 666         int contested = 0;
 667 #endif
 668 #ifdef KDTRACE_HOOKS
 669         uint64_t spin_cnt = 0;
 670         uint64_t sleep_cnt = 0;
 671         int64_t sleep_time = 0;
 672 #endif
 673
 674         if (SCHEDULER_STOPPED())
 675                 return;
 676
 677         if (rw_wlocked(rw)) {
 678                 KASSERT(rw->lock_object.lo_flags & LO_RECURSABLE,
 679                     ("%s: recursing but non-recursive rw %s @ %s:%d\n",
 680                     __func__, rw->lock_object.lo_name, file, line));
 681                 rw->rw_recurse++;
 682                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 683                         CTR2(KTR_LOCK, "%s: %p recursing", __func__, rw);
 684                 return;
 685         }
 686
 687         if (LOCK_LOG_TEST(&rw->lock_object, 0))
 688                 CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__,
 689                     rw->lock_object.lo_name, (void *)rw->rw_lock, file, line);
 690
 691         while (!_rw_write_lock(rw, tid)) {
 692 #ifdef KDTRACE_HOOKS
 693                 spin_cnt++;
 694 #endif
 695 #ifdef HWPMC_HOOKS
 696                 PMC_SOFT_CALL( , , lock, failed);
 697 #endif
 698                 lock_profile_obtain_lock_failed(&rw->lock_object,
 699                     &contested, &waittime);
 700 #ifdef ADAPTIVE_RWLOCKS
 701                 /*
 702                  * If the lock is write locked and the owner is
 703                  * running on another CPU, spin until the owner stops
 704                  * running or the state of the lock changes.
 705                  */
 706                 v = rw->rw_lock;
 707                 owner = (struct thread *)RW_OWNER(v);
 708                 if (!(v & RW_LOCK_READ) && TD_IS_RUNNING(owner)) {
 709                         if (LOCK_LOG_TEST(&rw->lock_object, 0))
 710                                 CTR3(KTR_LOCK, "%s: spinning on %p held by %p",
 711                                     __func__, rw, owner);
 712                         while ((struct thread*)RW_OWNER(rw->rw_lock) == owner &&
 713                             TD_IS_RUNNING(owner)) {
 714                                 cpu_spinwait();
 715 #ifdef KDTRACE_HOOKS
 716                                 spin_cnt++;
 717 #endif
 718                         }
 719                         continue;
 720                 }
 721                 if ((v & RW_LOCK_READ) && RW_READERS(v) &&
 722                     spintries < ROWNER_RETRIES) {
 723                         if (!(v & RW_LOCK_WRITE_SPINNER)) {
 724                                 if (!atomic_cmpset_ptr(&rw->rw_lock, v,
 725                                     v | RW_LOCK_WRITE_SPINNER)) {
 726                                         continue;
 727                                 }
 728                         }
 729                         spintries++;
 730                         for (i = 0; i < ROWNER_LOOPS; i++) {
 731                                 if ((rw->rw_lock & RW_LOCK_WRITE_SPINNER) == 0)
 732                                         break;
 733                                 cpu_spinwait();
 734                         }
 735 #ifdef KDTRACE_HOOKS
 736                         spin_cnt += ROWNER_LOOPS - i;
 737 #endif
 738                         if (i != ROWNER_LOOPS)
 739                                 continue;
 740                 }
 741 #endif
 742                 ts = turnstile_trywait(&rw->lock_object);
 743                 v = rw->rw_lock;
 744
 745 #ifdef ADAPTIVE_RWLOCKS
 746                 /*
 747                  * The current lock owner might have started executing
 748                  * on another CPU (or the lock could have changed
 749                  * owners) while we were waiting on the turnstile
 750                  * chain lock.  If so, drop the turnstile lock and try
 751                  * again.
 752                  */
 753                 if (!(v & RW_LOCK_READ)) {
 754                         owner = (struct thread *)RW_OWNER(v);
 755                         if (TD_IS_RUNNING(owner)) {
 756                                 turnstile_cancel(ts);
 757                                 continue;
 758                         }
 759                 }
 760 #endif
 761                 /*
 762                  * Check for the waiters flags about this rwlock.
 763                  * If the lock was released, without maintain any pending
 764                  * waiters queue, simply try to acquire it.
 765                  * If a pending waiters queue is present, claim the lock
 766                  * ownership and maintain the pending queue.
 767                  */
 768                 x = v & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER);
 769                 if ((v & ~x) == RW_UNLOCKED) {
 770                         x &= ~RW_LOCK_WRITE_SPINNER;
 771                         if (atomic_cmpset_acq_ptr(&rw->rw_lock, v, tid | x)) {
 772                                 if (x)
 773                                         turnstile_claim(ts);
 774                                 else
 775                                         turnstile_cancel(ts);
 776                                 break;
 777                         }
 778                         turnstile_cancel(ts);
 779                         continue;
 780                 }
 781                 /*
 782                  * If the RW_LOCK_WRITE_WAITERS flag isn't set, then try to
 783                  * set it.  If we fail to set it, then loop back and try
 784                  * again.
 785                  */
 786                 if (!(v & RW_LOCK_WRITE_WAITERS)) {
 787                         if (!atomic_cmpset_ptr(&rw->rw_lock, v,
 788                             v | RW_LOCK_WRITE_WAITERS)) {
 789                                 turnstile_cancel(ts);
 790                                 continue;
 791                         }
 792                         if (LOCK_LOG_TEST(&rw->lock_object, 0))
 793                                 CTR2(KTR_LOCK, "%s: %p set write waiters flag",
 794                                     __func__, rw);
 795                 }
 796                 /*
 797                  * We were unable to acquire the lock and the write waiters
 798                  * flag is set, so we must block on the turnstile.
 799                  */
 800                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 801                         CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__,
 802                             rw);
 803 #ifdef KDTRACE_HOOKS
 804                 sleep_time -= lockstat_nsecs();
 805 #endif
 806                 turnstile_wait(ts, rw_owner(rw), TS_EXCLUSIVE_QUEUE);
 807 #ifdef KDTRACE_HOOKS
 808                 sleep_time += lockstat_nsecs();
 809                 sleep_cnt++;
 810 #endif
 811                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 812                         CTR2(KTR_LOCK, "%s: %p resuming from turnstile",
 813                             __func__, rw);
 814 #ifdef ADAPTIVE_RWLOCKS
 815                 spintries = 0;
 816 #endif
 817         }
 818         LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_RW_WLOCK_ACQUIRE, rw, contested,
 819             waittime, file, line);
 820 #ifdef KDTRACE_HOOKS
 821         if (sleep_time)
 822                 LOCKSTAT_RECORD1(LS_RW_WLOCK_BLOCK, rw, sleep_time);
 823
 824         /*
 825          * Record only the loops spinning and not sleeping.
 826          */
 827         if (spin_cnt > sleep_cnt)
 828                 LOCKSTAT_RECORD1(LS_RW_WLOCK_SPIN, rw, (spin_cnt - sleep_cnt));
 829 #endif
 830 }
 831
 832 /*
 833  * This function is called if the first try at releasing a write lock failed.
 834  * This means that one of the 2 waiter bits must be set indicating that at
 835  * least one thread is waiting on this lock.
 836  */
 837 void
 838 _rw_wunlock_hard(struct rwlock *rw, uintptr_t tid, const char *file, int line)
 839 {
 840         struct turnstile *ts;
 841         uintptr_t v;
 842         int queue;
 843
 844         if (SCHEDULER_STOPPED())
 845                 return;
 846
 847         if (rw_wlocked(rw) && rw_recursed(rw)) {
 848                 rw->rw_recurse--;
 849                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 850                         CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, rw);
 851                 return;
 852         }
 853
 854         KASSERT(rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS),
 855             ("%s: neither of the waiter flags are set", __func__));
 856
 857         if (LOCK_LOG_TEST(&rw->lock_object, 0))
 858                 CTR2(KTR_LOCK, "%s: %p contested", __func__, rw);
 859
 860         turnstile_chain_lock(&rw->lock_object);
 861         ts = turnstile_lookup(&rw->lock_object);
 862         MPASS(ts != NULL);
 863
 864         /*
 865          * Use the same algo as sx locks for now.  Prefer waking up shared
 866          * waiters if we have any over writers.  This is probably not ideal.
 867          *
 868          * 'v' is the value we are going to write back to rw_lock.  If we
 869          * have waiters on both queues, we need to preserve the state of
 870          * the waiter flag for the queue we don't wake up.  For now this is
 871          * hardcoded for the algorithm mentioned above.
 872          *
 873          * In the case of both readers and writers waiting we wakeup the
 874          * readers but leave the RW_LOCK_WRITE_WAITERS flag set.  If a
 875          * new writer comes in before a reader it will claim the lock up
 876          * above.  There is probably a potential priority inversion in
 877          * there that could be worked around either by waking both queues
 878          * of waiters or doing some complicated lock handoff gymnastics.
 879          */
 880         v = RW_UNLOCKED;
 881         if (rw->rw_lock & RW_LOCK_WRITE_WAITERS) {
 882                 queue = TS_EXCLUSIVE_QUEUE;
 883                 v |= (rw->rw_lock & RW_LOCK_READ_WAITERS);
 884         } else
 885                 queue = TS_SHARED_QUEUE;
 886
 887         /* Wake up all waiters for the specific queue. */
 888         if (LOCK_LOG_TEST(&rw->lock_object, 0))
 889                 CTR3(KTR_LOCK, "%s: %p waking up %s waiters", __func__, rw,
 890                     queue == TS_SHARED_QUEUE ? "read" : "write");
 891         turnstile_broadcast(ts, queue);
 892         atomic_store_rel_ptr(&rw->rw_lock, v);
 893         turnstile_unpend(ts, TS_EXCLUSIVE_LOCK);
 894         turnstile_chain_unlock(&rw->lock_object);
 895 }
 896
 897 /*
 898  * Attempt to do a non-blocking upgrade from a read lock to a write
 899  * lock.  This will only succeed if this thread holds a single read
 900  * lock.  Returns true if the upgrade succeeded and false otherwise.
 901  */
 902 int
 903 _rw_try_upgrade(struct rwlock *rw, const char *file, int line)
 904 {
 905         uintptr_t v, x, tid;
 906         struct turnstile *ts;
 907         int success;
 908
 909         if (SCHEDULER_STOPPED())
 910                 return (1);
 911
 912         KASSERT(rw->rw_lock != RW_DESTROYED,
 913             ("rw_try_upgrade() of destroyed rwlock @ %s:%d", file, line));
 914         _rw_assert(rw, RA_RLOCKED, file, line);
 915
 916         /*
 917          * Attempt to switch from one reader to a writer.  If there
 918          * are any write waiters, then we will have to lock the
 919          * turnstile first to prevent races with another writer
 920          * calling turnstile_wait() before we have claimed this
 921          * turnstile.  So, do the simple case of no waiters first.
 922          */
 923         tid = (uintptr_t)curthread;
 924         success = 0;
 925         for (;;) {
 926                 v = rw->rw_lock;
 927                 if (RW_READERS(v) > 1)
 928                         break;
 929                 if (!(v & RW_LOCK_WAITERS)) {
 930                         success = atomic_cmpset_ptr(&rw->rw_lock, v, tid);
 931                         if (!success)
 932                                 continue;
 933                         break;
 934                 }
 935
 936                 /*
 937                  * Ok, we think we have waiters, so lock the turnstile.
 938                  */
 939                 ts = turnstile_trywait(&rw->lock_object);
 940                 v = rw->rw_lock;
 941                 if (RW_READERS(v) > 1) {
 942                         turnstile_cancel(ts);
 943                         break;
 944                 }
 945                 /*
 946                  * Try to switch from one reader to a writer again.  This time
 947                  * we honor the current state of the waiters flags.
 948                  * If we obtain the lock with the flags set, then claim
 949                  * ownership of the turnstile.
 950                  */
 951                 x = rw->rw_lock & RW_LOCK_WAITERS;
 952                 success = atomic_cmpset_ptr(&rw->rw_lock, v, tid | x);
 953                 if (success) {
 954                         if (x)
 955                                 turnstile_claim(ts);
 956                         else
 957                                 turnstile_cancel(ts);
 958                         break;
 959                 }
 960                 turnstile_cancel(ts);
 961         }
 962         LOCK_LOG_TRY("WUPGRADE", &rw->lock_object, 0, success, file, line);
 963         if (success) {
 964                 curthread->td_rw_rlocks--;
 965                 WITNESS_UPGRADE(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
 966                     file, line);
 967                 LOCKSTAT_RECORD0(LS_RW_TRYUPGRADE_UPGRADE, rw);
 968         }
 969         return (success);
 970 }
 971
 972 /*
 973  * Downgrade a write lock into a single read lock.
 974  */
 975 void
 976 _rw_downgrade(struct rwlock *rw, const char *file, int line)
 977 {
 978         struct turnstile *ts;
 979         uintptr_t tid, v;
 980         int rwait, wwait;
 981
 982         if (SCHEDULER_STOPPED())
 983                 return;
 984
 985         KASSERT(rw->rw_lock != RW_DESTROYED,
 986             ("rw_downgrade() of destroyed rwlock @ %s:%d", file, line));
 987         _rw_assert(rw, RA_WLOCKED | RA_NOTRECURSED, file, line);
 988 #ifndef INVARIANTS
 989         if (rw_recursed(rw))
 990                 panic("downgrade of a recursed lock");
 991 #endif
 992
 993         WITNESS_DOWNGRADE(&rw->lock_object, 0, file, line);
 994
 995         /*
 996          * Convert from a writer to a single reader.  First we handle
 997          * the easy case with no waiters.  If there are any waiters, we
 998          * lock the turnstile and "disown" the lock.
 999          */
1000         tid = (uintptr_t)curthread;
1001         if (atomic_cmpset_rel_ptr(&rw->rw_lock, tid, RW_READERS_LOCK(1)))
1002                 goto out;
1003
1004         /*
1005          * Ok, we think we have waiters, so lock the turnstile so we can
1006          * read the waiter flags without any races.
1007          */
1008         turnstile_chain_lock(&rw->lock_object);
1009         v = rw->rw_lock & RW_LOCK_WAITERS;
1010         rwait = v & RW_LOCK_READ_WAITERS;
1011         wwait = v & RW_LOCK_WRITE_WAITERS;
1012         MPASS(rwait | wwait);
1013
1014         /*
1015          * Downgrade from a write lock while preserving waiters flag
1016          * and give up ownership of the turnstile.
1017          */
1018         ts = turnstile_lookup(&rw->lock_object);
1019         MPASS(ts != NULL);
1020         if (!wwait)
1021                 v &= ~RW_LOCK_READ_WAITERS;
1022         atomic_store_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v);
1023         /*
1024          * Wake other readers if there are no writers pending.  Otherwise they
1025          * won't be able to acquire the lock anyway.
1026          */
1027         if (rwait && !wwait) {
1028                 turnstile_broadcast(ts, TS_SHARED_QUEUE);
1029                 turnstile_unpend(ts, TS_EXCLUSIVE_LOCK);
1030         } else
1031                 turnstile_disown(ts);
1032         turnstile_chain_unlock(&rw->lock_object);
1033 out:
1034         curthread->td_rw_rlocks++;
1035         LOCK_LOG_LOCK("WDOWNGRADE", &rw->lock_object, 0, 0, file, line);
1036         LOCKSTAT_RECORD0(LS_RW_DOWNGRADE_DOWNGRADE, rw);
1037 }
1038
1039 #ifdef INVARIANT_SUPPORT
1040 #ifndef INVARIANTS
1041 #undef _rw_assert
1042 #endif
1043
1044 /*
1045  * In the non-WITNESS case, rw_assert() can only detect that at least
1046  * *some* thread owns an rlock, but it cannot guarantee that *this*
1047  * thread owns an rlock.
1048  */
1049 void
1050 _rw_assert(struct rwlock *rw, int what, const char *file, int line)
1051 {
1052
1053         if (panicstr != NULL)
1054                 return;
1055         switch (what) {
1056         case RA_LOCKED:
1057         case RA_LOCKED | RA_RECURSED:
1058         case RA_LOCKED | RA_NOTRECURSED:
1059         case RA_RLOCKED:
1060 #ifdef WITNESS
1061                 witness_assert(&rw->lock_object, what, file, line);
1062 #else
1063                 /*
1064                  * If some other thread has a write lock or we have one
1065                  * and are asserting a read lock, fail.  Also, if no one
1066                  * has a lock at all, fail.
1067                  */
1068                 if (rw->rw_lock == RW_UNLOCKED ||
1069                     (!(rw->rw_lock & RW_LOCK_READ) && (what == RA_RLOCKED ||
1070                     rw_wowner(rw) != curthread)))
1071                         panic("Lock %s not %slocked @ %s:%d\n",
1072                             rw->lock_object.lo_name, (what == RA_RLOCKED) ?
1073                             "read " : "", file, line);
1074
1075                 if (!(rw->rw_lock & RW_LOCK_READ)) {
1076                         if (rw_recursed(rw)) {
1077                                 if (what & RA_NOTRECURSED)
1078                                         panic("Lock %s recursed @ %s:%d\n",
1079                                             rw->lock_object.lo_name, file,
1080                                             line);
1081                         } else if (what & RA_RECURSED)
1082                                 panic("Lock %s not recursed @ %s:%d\n",
1083                                     rw->lock_object.lo_name, file, line);
1084                 }
1085 #endif
1086                 break;
1087         case RA_WLOCKED:
1088         case RA_WLOCKED | RA_RECURSED:
1089         case RA_WLOCKED | RA_NOTRECURSED:
1090                 if (rw_wowner(rw) != curthread)
1091                         panic("Lock %s not exclusively locked @ %s:%d\n",
1092                             rw->lock_object.lo_name, file, line);
1093                 if (rw_recursed(rw)) {
1094                         if (what & RA_NOTRECURSED)
1095                                 panic("Lock %s recursed @ %s:%d\n",
1096                                     rw->lock_object.lo_name, file, line);
1097                 } else if (what & RA_RECURSED)
1098                         panic("Lock %s not recursed @ %s:%d\n",
1099                             rw->lock_object.lo_name, file, line);
1100                 break;
1101         case RA_UNLOCKED:
1102 #ifdef WITNESS
1103                 witness_assert(&rw->lock_object, what, file, line);
1104 #else
1105                 /*
1106                  * If we hold a write lock fail.  We can't reliably check
1107                  * to see if we hold a read lock or not.
1108                  */
1109                 if (rw_wowner(rw) == curthread)
1110                         panic("Lock %s exclusively locked @ %s:%d\n",
1111                             rw->lock_object.lo_name, file, line);
1112 #endif
1113                 break;
1114         default:
1115                 panic("Unknown rw lock assertion: %d @ %s:%d", what, file,
1116                     line);
1117         }
1118 }
1119 #endif /* INVARIANT_SUPPORT */
1120
1121 #ifdef DDB
1122 void
1123 db_show_rwlock(struct lock_object *lock)
1124 {
1125         struct rwlock *rw;
1126         struct thread *td;
1127
1128         rw = (struct rwlock *)lock;
1129
1130         db_printf(" state: ");
1131         if (rw->rw_lock == RW_UNLOCKED)
1132                 db_printf("UNLOCKED\n");
1133         else if (rw->rw_lock == RW_DESTROYED) {
1134                 db_printf("DESTROYED\n");
1135                 return;
1136         } else if (rw->rw_lock & RW_LOCK_READ)
1137                 db_printf("RLOCK: %ju locks\n",
1138                     (uintmax_t)(RW_READERS(rw->rw_lock)));
1139         else {
1140                 td = rw_wowner(rw);
1141                 db_printf("WLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1142                     td->td_tid, td->td_proc->p_pid, td->td_name);
1143                 if (rw_recursed(rw))
1144                         db_printf(" recursed: %u\n", rw->rw_recurse);
1145         }
1146         db_printf(" waiters: ");
1147         switch (rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS)) {
1148         case RW_LOCK_READ_WAITERS:
1149                 db_printf("readers\n");
1150                 break;
1151         case RW_LOCK_WRITE_WAITERS:
1152                 db_printf("writers\n");
1153                 break;
1154         case RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS:
1155                 db_printf("readers and writers\n");
1156                 break;
1157         default:
1158                 db_printf("none\n");
1159                 break;
1160         }
1161 }
1162
1163 #endif