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