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