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         KASSERT(rw->rw_lock != RW_DESTROYED,
 241             ("rw_wlock() of destroyed rwlock @ %s:%d", file, line));
 242         WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE, file,
 243             line, NULL);
 244         __rw_wlock(rw, curthread, file, line);
 245         LOCK_LOG_LOCK("WLOCK", &rw->lock_object, 0, rw->rw_recurse, file, line);
 246         WITNESS_LOCK(&rw->lock_object, LOP_EXCLUSIVE, file, line);
 247         curthread->td_locks++;
 248 }
 249
 250 int
 251 _rw_try_wlock(struct rwlock *rw, const char *file, int line)
 252 {
 253         int rval;
 254
 255         if (SCHEDULER_STOPPED())
 256                 return (1);
 257
 258         KASSERT(rw->rw_lock != RW_DESTROYED,
 259             ("rw_try_wlock() of destroyed rwlock @ %s:%d", file, line));
 260
 261         if (rw_wlocked(rw) &&
 262             (rw->lock_object.lo_flags & LO_RECURSABLE) != 0) {
 263                 rw->rw_recurse++;
 264                 rval = 1;
 265         } else
 266                 rval = atomic_cmpset_acq_ptr(&rw->rw_lock, RW_UNLOCKED,
 267                     (uintptr_t)curthread);
 268
 269         LOCK_LOG_TRY("WLOCK", &rw->lock_object, 0, rval, file, line);
 270         if (rval) {
 271                 WITNESS_LOCK(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
 272                     file, line);
 273                 curthread->td_locks++;
 274         }
 275         return (rval);
 276 }
 277
 278 void
 279 _rw_wunlock(struct rwlock *rw, const char *file, int line)
 280 {
 281
 282         if (SCHEDULER_STOPPED())
 283                 return;
 284         KASSERT(rw->rw_lock != RW_DESTROYED,
 285             ("rw_wunlock() of destroyed rwlock @ %s:%d", file, line));
 286         _rw_assert(rw, RA_WLOCKED, file, line);
 287         curthread->td_locks--;
 288         WITNESS_UNLOCK(&rw->lock_object, LOP_EXCLUSIVE, file, line);
 289         LOCK_LOG_LOCK("WUNLOCK", &rw->lock_object, 0, rw->rw_recurse, file,
 290             line);
 291         if (!rw_recursed(rw))
 292                 LOCKSTAT_PROFILE_RELEASE_LOCK(LS_RW_WUNLOCK_RELEASE, rw);
 293         __rw_wunlock(rw, curthread, file, line);
 294 }
 295 /*
 296  * Determines whether a new reader can acquire a lock.  Succeeds if the
 297  * reader already owns a read lock and the lock is locked for read to
 298  * prevent deadlock from reader recursion.  Also succeeds if the lock
 299  * is unlocked and has no writer waiters or spinners.  Failing otherwise
 300  * prioritizes writers before readers.
 301  */
 302 #define RW_CAN_READ(_rw)                                                \
 303     ((curthread->td_rw_rlocks && (_rw) & RW_LOCK_READ) || ((_rw) &      \
 304     (RW_LOCK_READ | RW_LOCK_WRITE_WAITERS | RW_LOCK_WRITE_SPINNER)) ==  \
 305     RW_LOCK_READ)
 306
 307 void
 308 _rw_rlock(struct rwlock *rw, const char *file, int line)
 309 {
 310         struct turnstile *ts;
 311 #ifdef ADAPTIVE_RWLOCKS
 312         volatile struct thread *owner;
 313         int spintries = 0;
 314         int i;
 315 #endif
 316 #ifdef LOCK_PROFILING
 317         uint64_t waittime = 0;
 318         int contested = 0;
 319 #endif
 320         uintptr_t v;
 321 #ifdef KDTRACE_HOOKS
 322         uint64_t spin_cnt = 0;
 323         uint64_t sleep_cnt = 0;
 324         int64_t sleep_time = 0;
 325 #endif
 326
 327         if (SCHEDULER_STOPPED())
 328                 return;
 329
 330         KASSERT(rw->rw_lock != RW_DESTROYED,
 331             ("rw_rlock() of destroyed rwlock @ %s:%d", file, line));
 332         KASSERT(rw_wowner(rw) != curthread,
 333             ("rw_rlock: wlock already held for %s @ %s:%d",
 334             rw->lock_object.lo_name, file, line));
 335         WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER, file, line, NULL);
 336
 337         for (;;) {
 338 #ifdef KDTRACE_HOOKS
 339                 spin_cnt++;
 340 #endif
 341                 /*
 342                  * Handle the easy case.  If no other thread has a write
 343                  * lock, then try to bump up the count of read locks.  Note
 344                  * that we have to preserve the current state of the
 345                  * RW_LOCK_WRITE_WAITERS flag.  If we fail to acquire a
 346                  * read lock, then rw_lock must have changed, so restart
 347                  * the loop.  Note that this handles the case of a
 348                  * completely unlocked rwlock since such a lock is encoded
 349                  * as a read lock with no waiters.
 350                  */
 351                 v = rw->rw_lock;
 352                 if (RW_CAN_READ(v)) {
 353                         /*
 354                          * The RW_LOCK_READ_WAITERS flag should only be set
 355                          * if the lock has been unlocked and write waiters
 356                          * were present.
 357                          */
 358                         if (atomic_cmpset_acq_ptr(&rw->rw_lock, v,
 359                             v + RW_ONE_READER)) {
 360                                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 361                                         CTR4(KTR_LOCK,
 362                                             "%s: %p succeed %p -> %p", __func__,
 363                                             rw, (void *)v,
 364                                             (void *)(v + RW_ONE_READER));
 365                                 break;
 366                         }
 367                         continue;
 368                 }
 369 #ifdef HWPMC_HOOKS
 370                 PMC_SOFT_CALL( , , lock, failed);
 371 #endif
 372                 lock_profile_obtain_lock_failed(&rw->lock_object,
 373                     &contested, &waittime);
 374
 375 #ifdef ADAPTIVE_RWLOCKS
 376                 /*
 377                  * If the owner is running on another CPU, spin until
 378                  * the owner stops running or the state of the lock
 379                  * changes.
 380                  */
 381                 if ((v & RW_LOCK_READ) == 0) {
 382                         owner = (struct thread *)RW_OWNER(v);
 383                         if (TD_IS_RUNNING(owner)) {
 384                                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 385                                         CTR3(KTR_LOCK,
 386                                             "%s: spinning on %p held by %p",
 387                                             __func__, rw, owner);
 388                                 while ((struct thread*)RW_OWNER(rw->rw_lock) ==
 389                                     owner && TD_IS_RUNNING(owner)) {
 390                                         cpu_spinwait();
 391 #ifdef KDTRACE_HOOKS
 392                                         spin_cnt++;
 393 #endif
 394                                 }
 395                                 continue;
 396                         }
 397                 } else if (spintries < ROWNER_RETRIES) {
 398                         spintries++;
 399                         for (i = 0; i < ROWNER_LOOPS; i++) {
 400                                 v = rw->rw_lock;
 401                                 if ((v & RW_LOCK_READ) == 0 || RW_CAN_READ(v))
 402                                         break;
 403                                 cpu_spinwait();
 404                         }
 405                         if (i != ROWNER_LOOPS)
 406                                 continue;
 407                 }
 408 #endif
 409
 410                 /*
 411                  * Okay, now it's the hard case.  Some other thread already
 412                  * has a write lock or there are write waiters present,
 413                  * acquire the turnstile lock so we can begin the process
 414                  * of blocking.
 415                  */
 416                 ts = turnstile_trywait(&rw->lock_object);
 417
 418                 /*
 419                  * The lock might have been released while we spun, so
 420                  * recheck its state and restart the loop if needed.
 421                  */
 422                 v = rw->rw_lock;
 423                 if (RW_CAN_READ(v)) {
 424                         turnstile_cancel(ts);
 425                         continue;
 426                 }
 427
 428 #ifdef ADAPTIVE_RWLOCKS
 429                 /*
 430                  * The current lock owner might have started executing
 431                  * on another CPU (or the lock could have changed
 432                  * owners) while we were waiting on the turnstile
 433                  * chain lock.  If so, drop the turnstile lock and try
 434                  * again.
 435                  */
 436                 if ((v & RW_LOCK_READ) == 0) {
 437                         owner = (struct thread *)RW_OWNER(v);
 438                         if (TD_IS_RUNNING(owner)) {
 439                                 turnstile_cancel(ts);
 440                                 continue;
 441                         }
 442                 }
 443 #endif
 444
 445                 /*
 446                  * The lock is held in write mode or it already has waiters.
 447                  */
 448                 MPASS(!RW_CAN_READ(v));
 449
 450                 /*
 451                  * If the RW_LOCK_READ_WAITERS flag is already set, then
 452                  * we can go ahead and block.  If it is not set then try
 453                  * to set it.  If we fail to set it drop the turnstile
 454                  * lock and restart the loop.
 455                  */
 456                 if (!(v & RW_LOCK_READ_WAITERS)) {
 457                         if (!atomic_cmpset_ptr(&rw->rw_lock, v,
 458                             v | RW_LOCK_READ_WAITERS)) {
 459                                 turnstile_cancel(ts);
 460                                 continue;
 461                         }
 462                         if (LOCK_LOG_TEST(&rw->lock_object, 0))
 463                                 CTR2(KTR_LOCK, "%s: %p set read waiters flag",
 464                                     __func__, rw);
 465                 }
 466
 467                 /*
 468                  * We were unable to acquire the lock and the read waiters
 469                  * flag is set, so we must block on the turnstile.
 470                  */
 471                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 472                         CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__,
 473                             rw);
 474 #ifdef KDTRACE_HOOKS
 475                 sleep_time -= lockstat_nsecs();
 476 #endif
 477                 turnstile_wait(ts, rw_owner(rw), TS_SHARED_QUEUE);
 478 #ifdef KDTRACE_HOOKS
 479                 sleep_time += lockstat_nsecs();
 480                 sleep_cnt++;
 481 #endif
 482                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 483                         CTR2(KTR_LOCK, "%s: %p resuming from turnstile",
 484                             __func__, rw);
 485         }
 486
 487         /*
 488          * TODO: acquire "owner of record" here.  Here be turnstile dragons
 489          * however.  turnstiles don't like owners changing between calls to
 490          * turnstile_wait() currently.
 491          */
 492         LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_RW_RLOCK_ACQUIRE, rw, contested,
 493             waittime, file, line);
 494         LOCK_LOG_LOCK("RLOCK", &rw->lock_object, 0, 0, file, line);
 495         WITNESS_LOCK(&rw->lock_object, 0, file, line);
 496         curthread->td_locks++;
 497         curthread->td_rw_rlocks++;
 498 #ifdef KDTRACE_HOOKS
 499         if (sleep_time)
 500                 LOCKSTAT_RECORD1(LS_RW_RLOCK_BLOCK, rw, sleep_time);
 501
 502         /*
 503          * Record only the loops spinning and not sleeping.
 504          */
 505         if (spin_cnt > sleep_cnt)
 506                 LOCKSTAT_RECORD1(LS_RW_RLOCK_SPIN, rw, (spin_cnt - sleep_cnt));
 507 #endif
 508 }
 509
 510 int
 511 _rw_try_rlock(struct rwlock *rw, const char *file, int line)
 512 {
 513         uintptr_t x;
 514
 515         if (SCHEDULER_STOPPED())
 516                 return (1);
 517
 518         for (;;) {
 519                 x = rw->rw_lock;
 520                 KASSERT(rw->rw_lock != RW_DESTROYED,
 521                     ("rw_try_rlock() of destroyed rwlock @ %s:%d", file, line));
 522                 if (!(x & RW_LOCK_READ))
 523                         break;
 524                 if (atomic_cmpset_acq_ptr(&rw->rw_lock, x, x + RW_ONE_READER)) {
 525                         LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 1, file,
 526                             line);
 527                         WITNESS_LOCK(&rw->lock_object, LOP_TRYLOCK, file, line);
 528                         curthread->td_locks++;
 529                         curthread->td_rw_rlocks++;
 530                         return (1);
 531                 }
 532         }
 533
 534         LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 0, file, line);
 535         return (0);
 536 }
 537
 538 void
 539 _rw_runlock(struct rwlock *rw, const char *file, int line)
 540 {
 541         struct turnstile *ts;
 542         uintptr_t x, v, queue;
 543
 544         if (SCHEDULER_STOPPED())
 545                 return;
 546
 547         KASSERT(rw->rw_lock != RW_DESTROYED,
 548             ("rw_runlock() of destroyed rwlock @ %s:%d", file, line));
 549         _rw_assert(rw, RA_RLOCKED, file, line);
 550         curthread->td_locks--;
 551         curthread->td_rw_rlocks--;
 552         WITNESS_UNLOCK(&rw->lock_object, 0, file, line);
 553         LOCK_LOG_LOCK("RUNLOCK", &rw->lock_object, 0, 0, file, line);
 554
 555         /* TODO: drop "owner of record" here. */
 556
 557         for (;;) {
 558                 /*
 559                  * See if there is more than one read lock held.  If so,
 560                  * just drop one and return.
 561                  */
 562                 x = rw->rw_lock;
 563                 if (RW_READERS(x) > 1) {
 564                         if (atomic_cmpset_rel_ptr(&rw->rw_lock, x,
 565                             x - RW_ONE_READER)) {
 566                                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 567                                         CTR4(KTR_LOCK,
 568                                             "%s: %p succeeded %p -> %p",
 569                                             __func__, rw, (void *)x,
 570                                             (void *)(x - RW_ONE_READER));
 571                                 break;
 572                         }
 573                         continue;
 574                 }
 575                 /*
 576                  * If there aren't any waiters for a write lock, then try
 577                  * to drop it quickly.
 578                  */
 579                 if (!(x & RW_LOCK_WAITERS)) {
 580                         MPASS((x & ~RW_LOCK_WRITE_SPINNER) ==
 581                             RW_READERS_LOCK(1));
 582                         if (atomic_cmpset_rel_ptr(&rw->rw_lock, x,
 583                             RW_UNLOCKED)) {
 584                                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 585                                         CTR2(KTR_LOCK, "%s: %p last succeeded",
 586                                             __func__, rw);
 587                                 break;
 588                         }
 589                         continue;
 590                 }
 591                 /*
 592                  * Ok, we know we have waiters and we think we are the
 593                  * last reader, so grab the turnstile lock.
 594                  */
 595                 turnstile_chain_lock(&rw->lock_object);
 596                 v = rw->rw_lock & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER);
 597                 MPASS(v & RW_LOCK_WAITERS);
 598
 599                 /*
 600                  * Try to drop our lock leaving the lock in a unlocked
 601                  * state.
 602                  *
 603                  * If you wanted to do explicit lock handoff you'd have to
 604                  * do it here.  You'd also want to use turnstile_signal()
 605                  * and you'd have to handle the race where a higher
 606                  * priority thread blocks on the write lock before the
 607                  * thread you wakeup actually runs and have the new thread
 608                  * "steal" the lock.  For now it's a lot simpler to just
 609                  * wakeup all of the waiters.
 610                  *
 611                  * As above, if we fail, then another thread might have
 612                  * acquired a read lock, so drop the turnstile lock and
 613                  * restart.
 614                  */
 615                 x = RW_UNLOCKED;
 616                 if (v & RW_LOCK_WRITE_WAITERS) {
 617                         queue = TS_EXCLUSIVE_QUEUE;
 618                         x |= (v & RW_LOCK_READ_WAITERS);
 619                 } else
 620                         queue = TS_SHARED_QUEUE;
 621                 if (!atomic_cmpset_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v,
 622                     x)) {
 623                         turnstile_chain_unlock(&rw->lock_object);
 624                         continue;
 625                 }
 626                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 627                         CTR2(KTR_LOCK, "%s: %p last succeeded with waiters",
 628                             __func__, rw);
 629
 630                 /*
 631                  * Ok.  The lock is released and all that's left is to
 632                  * wake up the waiters.  Note that the lock might not be
 633                  * free anymore, but in that case the writers will just
 634                  * block again if they run before the new lock holder(s)
 635                  * release the lock.
 636                  */
 637                 ts = turnstile_lookup(&rw->lock_object);
 638                 MPASS(ts != NULL);
 639                 turnstile_broadcast(ts, queue);
 640                 turnstile_unpend(ts, TS_SHARED_LOCK);
 641                 turnstile_chain_unlock(&rw->lock_object);
 642                 break;
 643         }
 644         LOCKSTAT_PROFILE_RELEASE_LOCK(LS_RW_RUNLOCK_RELEASE, rw);
 645 }
 646
 647 /*
 648  * This function is called when we are unable to obtain a write lock on the
 649  * first try.  This means that at least one other thread holds either a
 650  * read or write lock.
 651  */
 652 void
 653 _rw_wlock_hard(struct rwlock *rw, uintptr_t tid, const char *file, int line)
 654 {
 655         struct turnstile *ts;
 656 #ifdef ADAPTIVE_RWLOCKS
 657         volatile struct thread *owner;
 658         int spintries = 0;
 659         int i;
 660 #endif
 661         uintptr_t v, x;
 662 #ifdef LOCK_PROFILING
 663         uint64_t waittime = 0;
 664         int contested = 0;
 665 #endif
 666 #ifdef KDTRACE_HOOKS
 667         uint64_t spin_cnt = 0;
 668         uint64_t sleep_cnt = 0;
 669         int64_t sleep_time = 0;
 670 #endif
 671
 672         if (SCHEDULER_STOPPED())
 673                 return;
 674
 675         if (rw_wlocked(rw)) {
 676                 KASSERT(rw->lock_object.lo_flags & LO_RECURSABLE,
 677                     ("%s: recursing but non-recursive rw %s @ %s:%d\n",
 678                     __func__, rw->lock_object.lo_name, file, line));
 679                 rw->rw_recurse++;
 680                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 681                         CTR2(KTR_LOCK, "%s: %p recursing", __func__, rw);
 682                 return;
 683         }
 684
 685         if (LOCK_LOG_TEST(&rw->lock_object, 0))
 686                 CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__,
 687                     rw->lock_object.lo_name, (void *)rw->rw_lock, file, line);
 688
 689         while (!_rw_write_lock(rw, tid)) {
 690 #ifdef KDTRACE_HOOKS
 691                 spin_cnt++;
 692 #endif
 693 #ifdef HWPMC_HOOKS
 694                 PMC_SOFT_CALL( , , lock, failed);
 695 #endif
 696                 lock_profile_obtain_lock_failed(&rw->lock_object,
 697                     &contested, &waittime);
 698 #ifdef ADAPTIVE_RWLOCKS
 699                 /*
 700                  * If the lock is write locked and the owner is
 701                  * running on another CPU, spin until the owner stops
 702                  * running or the state of the lock changes.
 703                  */
 704                 v = rw->rw_lock;
 705                 owner = (struct thread *)RW_OWNER(v);
 706                 if (!(v & RW_LOCK_READ) && TD_IS_RUNNING(owner)) {
 707                         if (LOCK_LOG_TEST(&rw->lock_object, 0))
 708                                 CTR3(KTR_LOCK, "%s: spinning on %p held by %p",
 709                                     __func__, rw, owner);
 710                         while ((struct thread*)RW_OWNER(rw->rw_lock) == owner &&
 711                             TD_IS_RUNNING(owner)) {
 712                                 cpu_spinwait();
 713 #ifdef KDTRACE_HOOKS
 714                                 spin_cnt++;
 715 #endif
 716                         }
 717                         continue;
 718                 }
 719                 if ((v & RW_LOCK_READ) && RW_READERS(v) &&
 720                     spintries < ROWNER_RETRIES) {
 721                         if (!(v & RW_LOCK_WRITE_SPINNER)) {
 722                                 if (!atomic_cmpset_ptr(&rw->rw_lock, v,
 723                                     v | RW_LOCK_WRITE_SPINNER)) {
 724                                         continue;
 725                                 }
 726                         }
 727                         spintries++;
 728                         for (i = 0; i < ROWNER_LOOPS; i++) {
 729                                 if ((rw->rw_lock & RW_LOCK_WRITE_SPINNER) == 0)
 730                                         break;
 731                                 cpu_spinwait();
 732                         }
 733 #ifdef KDTRACE_HOOKS
 734                         spin_cnt += ROWNER_LOOPS - i;
 735 #endif
 736                         if (i != ROWNER_LOOPS)
 737                                 continue;
 738                 }
 739 #endif
 740                 ts = turnstile_trywait(&rw->lock_object);
 741                 v = rw->rw_lock;
 742
 743 #ifdef ADAPTIVE_RWLOCKS
 744                 /*
 745                  * The current lock owner might have started executing
 746                  * on another CPU (or the lock could have changed
 747                  * owners) while we were waiting on the turnstile
 748                  * chain lock.  If so, drop the turnstile lock and try
 749                  * again.
 750                  */
 751                 if (!(v & RW_LOCK_READ)) {
 752                         owner = (struct thread *)RW_OWNER(v);
 753                         if (TD_IS_RUNNING(owner)) {
 754                                 turnstile_cancel(ts);
 755                                 continue;
 756                         }
 757                 }
 758 #endif
 759                 /*
 760                  * Check for the waiters flags about this rwlock.
 761                  * If the lock was released, without maintain any pending
 762                  * waiters queue, simply try to acquire it.
 763                  * If a pending waiters queue is present, claim the lock
 764                  * ownership and maintain the pending queue.
 765                  */
 766                 x = v & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER);
 767                 if ((v & ~x) == RW_UNLOCKED) {
 768                         x &= ~RW_LOCK_WRITE_SPINNER;
 769                         if (atomic_cmpset_acq_ptr(&rw->rw_lock, v, tid | x)) {
 770                                 if (x)
 771                                         turnstile_claim(ts);
 772                                 else
 773                                         turnstile_cancel(ts);
 774                                 break;
 775                         }
 776                         turnstile_cancel(ts);
 777                         continue;
 778                 }
 779                 /*
 780                  * If the RW_LOCK_WRITE_WAITERS flag isn't set, then try to
 781                  * set it.  If we fail to set it, then loop back and try
 782                  * again.
 783                  */
 784                 if (!(v & RW_LOCK_WRITE_WAITERS)) {
 785                         if (!atomic_cmpset_ptr(&rw->rw_lock, v,
 786                             v | RW_LOCK_WRITE_WAITERS)) {
 787                                 turnstile_cancel(ts);
 788                                 continue;
 789                         }
 790                         if (LOCK_LOG_TEST(&rw->lock_object, 0))
 791                                 CTR2(KTR_LOCK, "%s: %p set write waiters flag",
 792                                     __func__, rw);
 793                 }
 794                 /*
 795                  * We were unable to acquire the lock and the write waiters
 796                  * flag is set, so we must block on the turnstile.
 797                  */
 798                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 799                         CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__,
 800                             rw);
 801 #ifdef KDTRACE_HOOKS
 802                 sleep_time -= lockstat_nsecs();
 803 #endif
 804                 turnstile_wait(ts, rw_owner(rw), TS_EXCLUSIVE_QUEUE);
 805 #ifdef KDTRACE_HOOKS
 806                 sleep_time += lockstat_nsecs();
 807                 sleep_cnt++;
 808 #endif
 809                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 810                         CTR2(KTR_LOCK, "%s: %p resuming from turnstile",
 811                             __func__, rw);
 812 #ifdef ADAPTIVE_RWLOCKS
 813                 spintries = 0;
 814 #endif
 815         }
 816         LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_RW_WLOCK_ACQUIRE, rw, contested,
 817             waittime, file, line);
 818 #ifdef KDTRACE_HOOKS
 819         if (sleep_time)
 820                 LOCKSTAT_RECORD1(LS_RW_WLOCK_BLOCK, rw, sleep_time);
 821
 822         /*
 823          * Record only the loops spinning and not sleeping.
 824          */
 825         if (spin_cnt > sleep_cnt)
 826                 LOCKSTAT_RECORD1(LS_RW_WLOCK_SPIN, rw, (spin_cnt - sleep_cnt));
 827 #endif
 828 }
 829
 830 /*
 831  * This function is called if the first try at releasing a write lock failed.
 832  * This means that one of the 2 waiter bits must be set indicating that at
 833  * least one thread is waiting on this lock.
 834  */
 835 void
 836 _rw_wunlock_hard(struct rwlock *rw, uintptr_t tid, const char *file, int line)
 837 {
 838         struct turnstile *ts;
 839         uintptr_t v;
 840         int queue;
 841
 842         if (SCHEDULER_STOPPED())
 843                 return;
 844
 845         if (rw_wlocked(rw) && rw_recursed(rw)) {
 846                 rw->rw_recurse--;
 847                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 848                         CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, rw);
 849                 return;
 850         }
 851
 852         KASSERT(rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS),
 853             ("%s: neither of the waiter flags are set", __func__));
 854
 855         if (LOCK_LOG_TEST(&rw->lock_object, 0))
 856                 CTR2(KTR_LOCK, "%s: %p contested", __func__, rw);
 857
 858         turnstile_chain_lock(&rw->lock_object);
 859         ts = turnstile_lookup(&rw->lock_object);
 860         MPASS(ts != NULL);
 861
 862         /*
 863          * Use the same algo as sx locks for now.  Prefer waking up shared
 864          * waiters if we have any over writers.  This is probably not ideal.
 865          *
 866          * 'v' is the value we are going to write back to rw_lock.  If we
 867          * have waiters on both queues, we need to preserve the state of
 868          * the waiter flag for the queue we don't wake up.  For now this is
 869          * hardcoded for the algorithm mentioned above.
 870          *
 871          * In the case of both readers and writers waiting we wakeup the
 872          * readers but leave the RW_LOCK_WRITE_WAITERS flag set.  If a
 873          * new writer comes in before a reader it will claim the lock up
 874          * above.  There is probably a potential priority inversion in
 875          * there that could be worked around either by waking both queues
 876          * of waiters or doing some complicated lock handoff gymnastics.
 877          */
 878         v = RW_UNLOCKED;
 879         if (rw->rw_lock & RW_LOCK_WRITE_WAITERS) {
 880                 queue = TS_EXCLUSIVE_QUEUE;
 881                 v |= (rw->rw_lock & RW_LOCK_READ_WAITERS);
 882         } else
 883                 queue = TS_SHARED_QUEUE;
 884
 885         /* Wake up all waiters for the specific queue. */
 886         if (LOCK_LOG_TEST(&rw->lock_object, 0))
 887                 CTR3(KTR_LOCK, "%s: %p waking up %s waiters", __func__, rw,
 888                     queue == TS_SHARED_QUEUE ? "read" : "write");
 889         turnstile_broadcast(ts, queue);
 890         atomic_store_rel_ptr(&rw->rw_lock, v);
 891         turnstile_unpend(ts, TS_EXCLUSIVE_LOCK);
 892         turnstile_chain_unlock(&rw->lock_object);
 893 }
 894
 895 /*
 896  * Attempt to do a non-blocking upgrade from a read lock to a write
 897  * lock.  This will only succeed if this thread holds a single read
 898  * lock.  Returns true if the upgrade succeeded and false otherwise.
 899  */
 900 int
 901 _rw_try_upgrade(struct rwlock *rw, const char *file, int line)
 902 {
 903         uintptr_t v, x, tid;
 904         struct turnstile *ts;
 905         int success;
 906
 907         if (SCHEDULER_STOPPED())
 908                 return (1);
 909
 910         KASSERT(rw->rw_lock != RW_DESTROYED,
 911             ("rw_try_upgrade() of destroyed rwlock @ %s:%d", file, line));
 912         _rw_assert(rw, RA_RLOCKED, file, line);
 913
 914         /*
 915          * Attempt to switch from one reader to a writer.  If there
 916          * are any write waiters, then we will have to lock the
 917          * turnstile first to prevent races with another writer
 918          * calling turnstile_wait() before we have claimed this
 919          * turnstile.  So, do the simple case of no waiters first.
 920          */
 921         tid = (uintptr_t)curthread;
 922         success = 0;
 923         for (;;) {
 924                 v = rw->rw_lock;
 925                 if (RW_READERS(v) > 1)
 926                         break;
 927                 if (!(v & RW_LOCK_WAITERS)) {
 928                         success = atomic_cmpset_ptr(&rw->rw_lock, v, tid);
 929                         if (!success)
 930                                 continue;
 931                         break;
 932                 }
 933
 934                 /*
 935                  * Ok, we think we have waiters, so lock the turnstile.
 936                  */
 937                 ts = turnstile_trywait(&rw->lock_object);
 938                 v = rw->rw_lock;
 939                 if (RW_READERS(v) > 1) {
 940                         turnstile_cancel(ts);
 941                         break;
 942                 }
 943                 /*
 944                  * Try to switch from one reader to a writer again.  This time
 945                  * we honor the current state of the waiters flags.
 946                  * If we obtain the lock with the flags set, then claim
 947                  * ownership of the turnstile.
 948                  */
 949                 x = rw->rw_lock & RW_LOCK_WAITERS;
 950                 success = atomic_cmpset_ptr(&rw->rw_lock, v, tid | x);
 951                 if (success) {
 952                         if (x)
 953                                 turnstile_claim(ts);
 954                         else
 955                                 turnstile_cancel(ts);
 956                         break;
 957                 }
 958                 turnstile_cancel(ts);
 959         }
 960         LOCK_LOG_TRY("WUPGRADE", &rw->lock_object, 0, success, file, line);
 961         if (success) {
 962                 curthread->td_rw_rlocks--;
 963                 WITNESS_UPGRADE(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
 964                     file, line);
 965                 LOCKSTAT_RECORD0(LS_RW_TRYUPGRADE_UPGRADE, rw);
 966         }
 967         return (success);
 968 }
 969
 970 /*
 971  * Downgrade a write lock into a single read lock.
 972  */
 973 void
 974 _rw_downgrade(struct rwlock *rw, const char *file, int line)
 975 {
 976         struct turnstile *ts;
 977         uintptr_t tid, v;
 978         int rwait, wwait;
 979
 980         if (SCHEDULER_STOPPED())
 981                 return;
 982
 983         KASSERT(rw->rw_lock != RW_DESTROYED,
 984             ("rw_downgrade() of destroyed rwlock @ %s:%d", file, line));
 985         _rw_assert(rw, RA_WLOCKED | RA_NOTRECURSED, file, line);
 986 #ifndef INVARIANTS
 987         if (rw_recursed(rw))
 988                 panic("downgrade of a recursed lock");
 989 #endif
 990
 991         WITNESS_DOWNGRADE(&rw->lock_object, 0, file, line);
 992
 993         /*
 994          * Convert from a writer to a single reader.  First we handle
 995          * the easy case with no waiters.  If there are any waiters, we
 996          * lock the turnstile and "disown" the lock.
 997          */
 998         tid = (uintptr_t)curthread;
 999         if (atomic_cmpset_rel_ptr(&rw->rw_lock, tid, RW_READERS_LOCK(1)))
1000                 goto out;
1001
1002         /*
1003          * Ok, we think we have waiters, so lock the turnstile so we can
1004          * read the waiter flags without any races.
1005          */
1006         turnstile_chain_lock(&rw->lock_object);
1007         v = rw->rw_lock & RW_LOCK_WAITERS;
1008         rwait = v & RW_LOCK_READ_WAITERS;
1009         wwait = v & RW_LOCK_WRITE_WAITERS;
1010         MPASS(rwait | wwait);
1011
1012         /*
1013          * Downgrade from a write lock while preserving waiters flag
1014          * and give up ownership of the turnstile.
1015          */
1016         ts = turnstile_lookup(&rw->lock_object);
1017         MPASS(ts != NULL);
1018         if (!wwait)
1019                 v &= ~RW_LOCK_READ_WAITERS;
1020         atomic_store_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v);
1021         /*
1022          * Wake other readers if there are no writers pending.  Otherwise they
1023          * won't be able to acquire the lock anyway.
1024          */
1025         if (rwait && !wwait) {
1026                 turnstile_broadcast(ts, TS_SHARED_QUEUE);
1027                 turnstile_unpend(ts, TS_EXCLUSIVE_LOCK);
1028         } else
1029                 turnstile_disown(ts);
1030         turnstile_chain_unlock(&rw->lock_object);
1031 out:
1032         curthread->td_rw_rlocks++;
1033         LOCK_LOG_LOCK("WDOWNGRADE", &rw->lock_object, 0, 0, file, line);
1034         LOCKSTAT_RECORD0(LS_RW_DOWNGRADE_DOWNGRADE, rw);
1035 }
1036
1037 #ifdef INVARIANT_SUPPORT
1038 #ifndef INVARIANTS
1039 #undef _rw_assert
1040 #endif
1041
1042 /*
1043  * In the non-WITNESS case, rw_assert() can only detect that at least
1044  * *some* thread owns an rlock, but it cannot guarantee that *this*
1045  * thread owns an rlock.
1046  */
1047 void
1048 _rw_assert(struct rwlock *rw, int what, const char *file, int line)
1049 {
1050
1051         if (panicstr != NULL)
1052                 return;
1053         switch (what) {
1054         case RA_LOCKED:
1055         case RA_LOCKED | RA_RECURSED:
1056         case RA_LOCKED | RA_NOTRECURSED:
1057         case RA_RLOCKED:
1058         case RA_RLOCKED | RA_RECURSED:
1059         case RA_RLOCKED | RA_NOTRECURSED:
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) && !(what & RA_RLOCKED)) {
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