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         if (SCHEDULER_STOPPED())
 233                 return;
 234         MPASS(curthread != NULL);
 235         KASSERT(rw->rw_lock != RW_DESTROYED,
 236             ("rw_wlock() of destroyed rwlock @ %s:%d", file, line));
 237         WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE, file,
 238             line, NULL);
 239         __rw_wlock(rw, curthread, file, line);
 240         LOCK_LOG_LOCK("WLOCK", &rw->lock_object, 0, rw->rw_recurse, file, line);
 241         WITNESS_LOCK(&rw->lock_object, LOP_EXCLUSIVE, file, line);
 242         curthread->td_locks++;
 243 }
 244
 245 int
 246 _rw_try_wlock(struct rwlock *rw, const char *file, int line)
 247 {
 248         int rval;
 249
 250         if (SCHEDULER_STOPPED())
 251                 return (1);
 252
 253         KASSERT(rw->rw_lock != RW_DESTROYED,
 254             ("rw_try_wlock() of destroyed rwlock @ %s:%d", file, line));
 255
 256         if (rw_wlocked(rw) &&
 257             (rw->lock_object.lo_flags & LO_RECURSABLE) != 0) {
 258                 rw->rw_recurse++;
 259                 rval = 1;
 260         } else
 261                 rval = atomic_cmpset_acq_ptr(&rw->rw_lock, RW_UNLOCKED,
 262                     (uintptr_t)curthread);
 263
 264         LOCK_LOG_TRY("WLOCK", &rw->lock_object, 0, rval, file, line);
 265         if (rval) {
 266                 WITNESS_LOCK(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
 267                     file, line);
 268                 curthread->td_locks++;
 269         }
 270         return (rval);
 271 }
 272
 273 void
 274 _rw_wunlock(struct rwlock *rw, const char *file, int line)
 275 {
 276
 277         if (SCHEDULER_STOPPED())
 278                 return;
 279         MPASS(curthread != NULL);
 280         KASSERT(rw->rw_lock != RW_DESTROYED,
 281             ("rw_wunlock() of destroyed rwlock @ %s:%d", file, line));
 282         _rw_assert(rw, RA_WLOCKED, file, line);
 283         curthread->td_locks--;
 284         WITNESS_UNLOCK(&rw->lock_object, LOP_EXCLUSIVE, file, line);
 285         LOCK_LOG_LOCK("WUNLOCK", &rw->lock_object, 0, rw->rw_recurse, file,
 286             line);
 287         if (!rw_recursed(rw))
 288                 LOCKSTAT_PROFILE_RELEASE_LOCK(LS_RW_WUNLOCK_RELEASE, rw);
 289         __rw_wunlock(rw, curthread, file, line);
 290 }
 291 /*
 292  * Determines whether a new reader can acquire a lock.  Succeeds if the
 293  * reader already owns a read lock and the lock is locked for read to
 294  * prevent deadlock from reader recursion.  Also succeeds if the lock
 295  * is unlocked and has no writer waiters or spinners.  Failing otherwise
 296  * prioritizes writers before readers.
 297  */
 298 #define RW_CAN_READ(_rw)                                                \
 299     ((curthread->td_rw_rlocks && (_rw) & RW_LOCK_READ) || ((_rw) &      \
 300     (RW_LOCK_READ | RW_LOCK_WRITE_WAITERS | RW_LOCK_WRITE_SPINNER)) ==  \
 301     RW_LOCK_READ)
 302
 303 void
 304 _rw_rlock(struct rwlock *rw, const char *file, int line)
 305 {
 306         struct turnstile *ts;
 307 #ifdef ADAPTIVE_RWLOCKS
 308         volatile struct thread *owner;
 309         int spintries = 0;
 310         int i;
 311 #endif
 312 #ifdef LOCK_PROFILING
 313         uint64_t waittime = 0;
 314         int contested = 0;
 315 #endif
 316         uintptr_t v;
 317 #ifdef KDTRACE_HOOKS
 318         uint64_t spin_cnt = 0;
 319         uint64_t sleep_cnt = 0;
 320         int64_t sleep_time = 0;
 321 #endif
 322
 323         if (SCHEDULER_STOPPED())
 324                 return;
 325
 326         KASSERT(rw->rw_lock != RW_DESTROYED,
 327             ("rw_rlock() of destroyed rwlock @ %s:%d", file, line));
 328         KASSERT(rw_wowner(rw) != curthread,
 329             ("rw_rlock: wlock already held for %s @ %s:%d",
 330             rw->lock_object.lo_name, file, line));
 331         WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER, file, line, NULL);
 332
 333         for (;;) {
 334 #ifdef KDTRACE_HOOKS
 335                 spin_cnt++;
 336 #endif
 337                 /*
 338                  * Handle the easy case.  If no other thread has a write
 339                  * lock, then try to bump up the count of read locks.  Note
 340                  * that we have to preserve the current state of the
 341                  * RW_LOCK_WRITE_WAITERS flag.  If we fail to acquire a
 342                  * read lock, then rw_lock must have changed, so restart
 343                  * the loop.  Note that this handles the case of a
 344                  * completely unlocked rwlock since such a lock is encoded
 345                  * as a read lock with no waiters.
 346                  */
 347                 v = rw->rw_lock;
 348                 if (RW_CAN_READ(v)) {
 349                         /*
 350                          * The RW_LOCK_READ_WAITERS flag should only be set
 351                          * if the lock has been unlocked and write waiters
 352                          * were present.
 353                          */
 354                         if (atomic_cmpset_acq_ptr(&rw->rw_lock, v,
 355                             v + RW_ONE_READER)) {
 356                                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 357                                         CTR4(KTR_LOCK,
 358                                             "%s: %p succeed %p -> %p", __func__,
 359                                             rw, (void *)v,
 360                                             (void *)(v + RW_ONE_READER));
 361                                 break;
 362                         }
 363                         continue;
 364                 }
 365                 lock_profile_obtain_lock_failed(&rw->lock_object,
 366                     &contested, &waittime);
 367
 368 #ifdef ADAPTIVE_RWLOCKS
 369                 /*
 370                  * If the owner is running on another CPU, spin until
 371                  * the owner stops running or the state of the lock
 372                  * changes.
 373                  */
 374                 if ((v & RW_LOCK_READ) == 0) {
 375                         owner = (struct thread *)RW_OWNER(v);
 376                         if (TD_IS_RUNNING(owner)) {
 377                                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 378                                         CTR3(KTR_LOCK,
 379                                             "%s: spinning on %p held by %p",
 380                                             __func__, rw, owner);
 381                                 while ((struct thread*)RW_OWNER(rw->rw_lock) ==
 382                                     owner && TD_IS_RUNNING(owner)) {
 383                                         cpu_spinwait();
 384 #ifdef KDTRACE_HOOKS
 385                                         spin_cnt++;
 386 #endif
 387                                 }
 388                                 continue;
 389                         }
 390                 } else if (spintries < ROWNER_RETRIES) {
 391                         spintries++;
 392                         for (i = 0; i < ROWNER_LOOPS; i++) {
 393                                 v = rw->rw_lock;
 394                                 if ((v & RW_LOCK_READ) == 0 || RW_CAN_READ(v))
 395                                         break;
 396                                 cpu_spinwait();
 397                         }
 398                         if (i != ROWNER_LOOPS)
 399                                 continue;
 400                 }
 401 #endif
 402
 403                 /*
 404                  * Okay, now it's the hard case.  Some other thread already
 405                  * has a write lock or there are write waiters present,
 406                  * acquire the turnstile lock so we can begin the process
 407                  * of blocking.
 408                  */
 409                 ts = turnstile_trywait(&rw->lock_object);
 410
 411                 /*
 412                  * The lock might have been released while we spun, so
 413                  * recheck its state and restart the loop if needed.
 414                  */
 415                 v = rw->rw_lock;
 416                 if (RW_CAN_READ(v)) {
 417                         turnstile_cancel(ts);
 418                         continue;
 419                 }
 420
 421 #ifdef ADAPTIVE_RWLOCKS
 422                 /*
 423                  * The current lock owner might have started executing
 424                  * on another CPU (or the lock could have changed
 425                  * owners) while we were waiting on the turnstile
 426                  * chain lock.  If so, drop the turnstile lock and try
 427                  * again.
 428                  */
 429                 if ((v & RW_LOCK_READ) == 0) {
 430                         owner = (struct thread *)RW_OWNER(v);
 431                         if (TD_IS_RUNNING(owner)) {
 432                                 turnstile_cancel(ts);
 433                                 continue;
 434                         }
 435                 }
 436 #endif
 437
 438                 /*
 439                  * The lock is held in write mode or it already has waiters.
 440                  */
 441                 MPASS(!RW_CAN_READ(v));
 442
 443                 /*
 444                  * If the RW_LOCK_READ_WAITERS flag is already set, then
 445                  * we can go ahead and block.  If it is not set then try
 446                  * to set it.  If we fail to set it drop the turnstile
 447                  * lock and restart the loop.
 448                  */
 449                 if (!(v & RW_LOCK_READ_WAITERS)) {
 450                         if (!atomic_cmpset_ptr(&rw->rw_lock, v,
 451                             v | RW_LOCK_READ_WAITERS)) {
 452                                 turnstile_cancel(ts);
 453                                 continue;
 454                         }
 455                         if (LOCK_LOG_TEST(&rw->lock_object, 0))
 456                                 CTR2(KTR_LOCK, "%s: %p set read waiters flag",
 457                                     __func__, rw);
 458                 }
 459
 460                 /*
 461                  * We were unable to acquire the lock and the read waiters
 462                  * flag is set, so we must block on the turnstile.
 463                  */
 464                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 465                         CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__,
 466                             rw);
 467 #ifdef KDTRACE_HOOKS
 468                 sleep_time -= lockstat_nsecs();
 469 #endif
 470                 turnstile_wait(ts, rw_owner(rw), TS_SHARED_QUEUE);
 471 #ifdef KDTRACE_HOOKS
 472                 sleep_time += lockstat_nsecs();
 473                 sleep_cnt++;
 474 #endif
 475                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 476                         CTR2(KTR_LOCK, "%s: %p resuming from turnstile",
 477                             __func__, rw);
 478         }
 479
 480         /*
 481          * TODO: acquire "owner of record" here.  Here be turnstile dragons
 482          * however.  turnstiles don't like owners changing between calls to
 483          * turnstile_wait() currently.
 484          */
 485         LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_RW_RLOCK_ACQUIRE, rw, contested,
 486             waittime, file, line);
 487         LOCK_LOG_LOCK("RLOCK", &rw->lock_object, 0, 0, file, line);
 488         WITNESS_LOCK(&rw->lock_object, 0, file, line);
 489         curthread->td_locks++;
 490         curthread->td_rw_rlocks++;
 491 #ifdef KDTRACE_HOOKS
 492         if (sleep_time)
 493                 LOCKSTAT_RECORD1(LS_RW_RLOCK_BLOCK, rw, sleep_time);
 494
 495         /*
 496          * Record only the loops spinning and not sleeping.
 497          */
 498         if (spin_cnt > sleep_cnt)
 499                 LOCKSTAT_RECORD1(LS_RW_RLOCK_SPIN, rw, (spin_cnt - sleep_cnt));
 500 #endif
 501 }
 502
 503 int
 504 _rw_try_rlock(struct rwlock *rw, const char *file, int line)
 505 {
 506         uintptr_t x;
 507
 508         if (SCHEDULER_STOPPED())
 509                 return (1);
 510
 511         for (;;) {
 512                 x = rw->rw_lock;
 513                 KASSERT(rw->rw_lock != RW_DESTROYED,
 514                     ("rw_try_rlock() of destroyed rwlock @ %s:%d", file, line));
 515                 if (!(x & RW_LOCK_READ))
 516                         break;
 517                 if (atomic_cmpset_acq_ptr(&rw->rw_lock, x, x + RW_ONE_READER)) {
 518                         LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 1, file,
 519                             line);
 520                         WITNESS_LOCK(&rw->lock_object, LOP_TRYLOCK, file, line);
 521                         curthread->td_locks++;
 522                         curthread->td_rw_rlocks++;
 523                         return (1);
 524                 }
 525         }
 526
 527         LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 0, file, line);
 528         return (0);
 529 }
 530
 531 void
 532 _rw_runlock(struct rwlock *rw, const char *file, int line)
 533 {
 534         struct turnstile *ts;
 535         uintptr_t x, v, queue;
 536
 537         if (SCHEDULER_STOPPED())
 538                 return;
 539
 540         KASSERT(rw->rw_lock != RW_DESTROYED,
 541             ("rw_runlock() of destroyed rwlock @ %s:%d", file, line));
 542         _rw_assert(rw, RA_RLOCKED, file, line);
 543         curthread->td_locks--;
 544         curthread->td_rw_rlocks--;
 545         WITNESS_UNLOCK(&rw->lock_object, 0, file, line);
 546         LOCK_LOG_LOCK("RUNLOCK", &rw->lock_object, 0, 0, file, line);
 547
 548         /* TODO: drop "owner of record" here. */
 549
 550         for (;;) {
 551                 /*
 552                  * See if there is more than one read lock held.  If so,
 553                  * just drop one and return.
 554                  */
 555                 x = rw->rw_lock;
 556                 if (RW_READERS(x) > 1) {
 557                         if (atomic_cmpset_rel_ptr(&rw->rw_lock, x,
 558                             x - RW_ONE_READER)) {
 559                                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 560                                         CTR4(KTR_LOCK,
 561                                             "%s: %p succeeded %p -> %p",
 562                                             __func__, rw, (void *)x,
 563                                             (void *)(x - RW_ONE_READER));
 564                                 break;
 565                         }
 566                         continue;
 567                 }
 568                 /*
 569                  * If there aren't any waiters for a write lock, then try
 570                  * to drop it quickly.
 571                  */
 572                 if (!(x & RW_LOCK_WAITERS)) {
 573                         MPASS((x & ~RW_LOCK_WRITE_SPINNER) ==
 574                             RW_READERS_LOCK(1));
 575                         if (atomic_cmpset_rel_ptr(&rw->rw_lock, x,
 576                             RW_UNLOCKED)) {
 577                                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 578                                         CTR2(KTR_LOCK, "%s: %p last succeeded",
 579                                             __func__, rw);
 580                                 break;
 581                         }
 582                         continue;
 583                 }
 584                 /*
 585                  * Ok, we know we have waiters and we think we are the
 586                  * last reader, so grab the turnstile lock.
 587                  */
 588                 turnstile_chain_lock(&rw->lock_object);
 589                 v = rw->rw_lock & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER);
 590                 MPASS(v & RW_LOCK_WAITERS);
 591
 592                 /*
 593                  * Try to drop our lock leaving the lock in a unlocked
 594                  * state.
 595                  *
 596                  * If you wanted to do explicit lock handoff you'd have to
 597                  * do it here.  You'd also want to use turnstile_signal()
 598                  * and you'd have to handle the race where a higher
 599                  * priority thread blocks on the write lock before the
 600                  * thread you wakeup actually runs and have the new thread
 601                  * "steal" the lock.  For now it's a lot simpler to just
 602                  * wakeup all of the waiters.
 603                  *
 604                  * As above, if we fail, then another thread might have
 605                  * acquired a read lock, so drop the turnstile lock and
 606                  * restart.
 607                  */
 608                 x = RW_UNLOCKED;
 609                 if (v & RW_LOCK_WRITE_WAITERS) {
 610                         queue = TS_EXCLUSIVE_QUEUE;
 611                         x |= (v & RW_LOCK_READ_WAITERS);
 612                 } else
 613                         queue = TS_SHARED_QUEUE;
 614                 if (!atomic_cmpset_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v,
 615                     x)) {
 616                         turnstile_chain_unlock(&rw->lock_object);
 617                         continue;
 618                 }
 619                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 620                         CTR2(KTR_LOCK, "%s: %p last succeeded with waiters",
 621                             __func__, rw);
 622
 623                 /*
 624                  * Ok.  The lock is released and all that's left is to
 625                  * wake up the waiters.  Note that the lock might not be
 626                  * free anymore, but in that case the writers will just
 627                  * block again if they run before the new lock holder(s)
 628                  * release the lock.
 629                  */
 630                 ts = turnstile_lookup(&rw->lock_object);
 631                 MPASS(ts != NULL);
 632                 turnstile_broadcast(ts, queue);
 633                 turnstile_unpend(ts, TS_SHARED_LOCK);
 634                 turnstile_chain_unlock(&rw->lock_object);
 635                 break;
 636         }
 637         LOCKSTAT_PROFILE_RELEASE_LOCK(LS_RW_RUNLOCK_RELEASE, rw);
 638 }
 639
 640 /*
 641  * This function is called when we are unable to obtain a write lock on the
 642  * first try.  This means that at least one other thread holds either a
 643  * read or write lock.
 644  */
 645 void
 646 _rw_wlock_hard(struct rwlock *rw, uintptr_t tid, const char *file, int line)
 647 {
 648         struct turnstile *ts;
 649 #ifdef ADAPTIVE_RWLOCKS
 650         volatile struct thread *owner;
 651         int spintries = 0;
 652         int i;
 653 #endif
 654         uintptr_t v, x;
 655 #ifdef LOCK_PROFILING
 656         uint64_t waittime = 0;
 657         int contested = 0;
 658 #endif
 659 #ifdef KDTRACE_HOOKS
 660         uint64_t spin_cnt = 0;
 661         uint64_t sleep_cnt = 0;
 662         int64_t sleep_time = 0;
 663 #endif
 664
 665         if (SCHEDULER_STOPPED())
 666                 return;
 667
 668         if (rw_wlocked(rw)) {
 669                 KASSERT(rw->lock_object.lo_flags & LO_RECURSABLE,
 670                     ("%s: recursing but non-recursive rw %s @ %s:%d\n",
 671                     __func__, rw->lock_object.lo_name, file, line));
 672                 rw->rw_recurse++;
 673                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 674                         CTR2(KTR_LOCK, "%s: %p recursing", __func__, rw);
 675                 return;
 676         }
 677
 678         if (LOCK_LOG_TEST(&rw->lock_object, 0))
 679                 CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__,
 680                     rw->lock_object.lo_name, (void *)rw->rw_lock, file, line);
 681
 682         while (!_rw_write_lock(rw, tid)) {
 683 #ifdef KDTRACE_HOOKS
 684                 spin_cnt++;
 685 #endif
 686                 lock_profile_obtain_lock_failed(&rw->lock_object,
 687                     &contested, &waittime);
 688 #ifdef ADAPTIVE_RWLOCKS
 689                 /*
 690                  * If the lock is write locked and the owner is
 691                  * running on another CPU, spin until the owner stops
 692                  * running or the state of the lock changes.
 693                  */
 694                 v = rw->rw_lock;
 695                 owner = (struct thread *)RW_OWNER(v);
 696                 if (!(v & RW_LOCK_READ) && TD_IS_RUNNING(owner)) {
 697                         if (LOCK_LOG_TEST(&rw->lock_object, 0))
 698                                 CTR3(KTR_LOCK, "%s: spinning on %p held by %p",
 699                                     __func__, rw, owner);
 700                         while ((struct thread*)RW_OWNER(rw->rw_lock) == owner &&
 701                             TD_IS_RUNNING(owner)) {
 702                                 cpu_spinwait();
 703 #ifdef KDTRACE_HOOKS
 704                                 spin_cnt++;
 705 #endif
 706                         }
 707                         continue;
 708                 }
 709                 if ((v & RW_LOCK_READ) && RW_READERS(v) &&
 710                     spintries < ROWNER_RETRIES) {
 711                         if (!(v & RW_LOCK_WRITE_SPINNER)) {
 712                                 if (!atomic_cmpset_ptr(&rw->rw_lock, v,
 713                                     v | RW_LOCK_WRITE_SPINNER)) {
 714                                         continue;
 715                                 }
 716                         }
 717                         spintries++;
 718                         for (i = 0; i < ROWNER_LOOPS; i++) {
 719                                 if ((rw->rw_lock & RW_LOCK_WRITE_SPINNER) == 0)
 720                                         break;
 721                                 cpu_spinwait();
 722                         }
 723 #ifdef KDTRACE_HOOKS
 724                         spin_cnt += ROWNER_LOOPS - i;
 725 #endif
 726                         if (i != ROWNER_LOOPS)
 727                                 continue;
 728                 }
 729 #endif
 730                 ts = turnstile_trywait(&rw->lock_object);
 731                 v = rw->rw_lock;
 732
 733 #ifdef ADAPTIVE_RWLOCKS
 734                 /*
 735                  * The current lock owner might have started executing
 736                  * on another CPU (or the lock could have changed
 737                  * owners) while we were waiting on the turnstile
 738                  * chain lock.  If so, drop the turnstile lock and try
 739                  * again.
 740                  */
 741                 if (!(v & RW_LOCK_READ)) {
 742                         owner = (struct thread *)RW_OWNER(v);
 743                         if (TD_IS_RUNNING(owner)) {
 744                                 turnstile_cancel(ts);
 745                                 continue;
 746                         }
 747                 }
 748 #endif
 749                 /*
 750                  * Check for the waiters flags about this rwlock.
 751                  * If the lock was released, without maintain any pending
 752                  * waiters queue, simply try to acquire it.
 753                  * If a pending waiters queue is present, claim the lock
 754                  * ownership and maintain the pending queue.
 755                  */
 756                 x = v & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER);
 757                 if ((v & ~x) == RW_UNLOCKED) {
 758                         x &= ~RW_LOCK_WRITE_SPINNER;
 759                         if (atomic_cmpset_acq_ptr(&rw->rw_lock, v, tid | x)) {
 760                                 if (x)
 761                                         turnstile_claim(ts);
 762                                 else
 763                                         turnstile_cancel(ts);
 764                                 break;
 765                         }
 766                         turnstile_cancel(ts);
 767                         continue;
 768                 }
 769                 /*
 770                  * If the RW_LOCK_WRITE_WAITERS flag isn't set, then try to
 771                  * set it.  If we fail to set it, then loop back and try
 772                  * again.
 773                  */
 774                 if (!(v & RW_LOCK_WRITE_WAITERS)) {
 775                         if (!atomic_cmpset_ptr(&rw->rw_lock, v,
 776                             v | RW_LOCK_WRITE_WAITERS)) {
 777                                 turnstile_cancel(ts);
 778                                 continue;
 779                         }
 780                         if (LOCK_LOG_TEST(&rw->lock_object, 0))
 781                                 CTR2(KTR_LOCK, "%s: %p set write waiters flag",
 782                                     __func__, rw);
 783                 }
 784                 /*
 785                  * We were unable to acquire the lock and the write waiters
 786                  * flag is set, so we must block on the turnstile.
 787                  */
 788                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 789                         CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__,
 790                             rw);
 791 #ifdef KDTRACE_HOOKS
 792                 sleep_time -= lockstat_nsecs();
 793 #endif
 794                 turnstile_wait(ts, rw_owner(rw), TS_EXCLUSIVE_QUEUE);
 795 #ifdef KDTRACE_HOOKS
 796                 sleep_time += lockstat_nsecs();
 797                 sleep_cnt++;
 798 #endif
 799                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 800                         CTR2(KTR_LOCK, "%s: %p resuming from turnstile",
 801                             __func__, rw);
 802 #ifdef ADAPTIVE_RWLOCKS
 803                 spintries = 0;
 804 #endif
 805         }
 806         LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_RW_WLOCK_ACQUIRE, rw, contested,
 807             waittime, file, line);
 808 #ifdef KDTRACE_HOOKS
 809         if (sleep_time)
 810                 LOCKSTAT_RECORD1(LS_RW_WLOCK_BLOCK, rw, sleep_time);
 811
 812         /*
 813          * Record only the loops spinning and not sleeping.
 814          */
 815         if (spin_cnt > sleep_cnt)
 816                 LOCKSTAT_RECORD1(LS_RW_WLOCK_SPIN, rw, (spin_cnt - sleep_cnt));
 817 #endif
 818 }
 819
 820 /*
 821  * This function is called if the first try at releasing a write lock failed.
 822  * This means that one of the 2 waiter bits must be set indicating that at
 823  * least one thread is waiting on this lock.
 824  */
 825 void
 826 _rw_wunlock_hard(struct rwlock *rw, uintptr_t tid, const char *file, int line)
 827 {
 828         struct turnstile *ts;
 829         uintptr_t v;
 830         int queue;
 831
 832         if (SCHEDULER_STOPPED())
 833                 return;
 834
 835         if (rw_wlocked(rw) && rw_recursed(rw)) {
 836                 rw->rw_recurse--;
 837                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 838                         CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, rw);
 839                 return;
 840         }
 841
 842         KASSERT(rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS),
 843             ("%s: neither of the waiter flags are set", __func__));
 844
 845         if (LOCK_LOG_TEST(&rw->lock_object, 0))
 846                 CTR2(KTR_LOCK, "%s: %p contested", __func__, rw);
 847
 848         turnstile_chain_lock(&rw->lock_object);
 849         ts = turnstile_lookup(&rw->lock_object);
 850         MPASS(ts != NULL);
 851
 852         /*
 853          * Use the same algo as sx locks for now.  Prefer waking up shared
 854          * waiters if we have any over writers.  This is probably not ideal.
 855          *
 856          * 'v' is the value we are going to write back to rw_lock.  If we
 857          * have waiters on both queues, we need to preserve the state of
 858          * the waiter flag for the queue we don't wake up.  For now this is
 859          * hardcoded for the algorithm mentioned above.
 860          *
 861          * In the case of both readers and writers waiting we wakeup the
 862          * readers but leave the RW_LOCK_WRITE_WAITERS flag set.  If a
 863          * new writer comes in before a reader it will claim the lock up
 864          * above.  There is probably a potential priority inversion in
 865          * there that could be worked around either by waking both queues
 866          * of waiters or doing some complicated lock handoff gymnastics.
 867          */
 868         v = RW_UNLOCKED;
 869         if (rw->rw_lock & RW_LOCK_WRITE_WAITERS) {
 870                 queue = TS_EXCLUSIVE_QUEUE;
 871                 v |= (rw->rw_lock & RW_LOCK_READ_WAITERS);
 872         } else
 873                 queue = TS_SHARED_QUEUE;
 874
 875         /* Wake up all waiters for the specific queue. */
 876         if (LOCK_LOG_TEST(&rw->lock_object, 0))
 877                 CTR3(KTR_LOCK, "%s: %p waking up %s waiters", __func__, rw,
 878                     queue == TS_SHARED_QUEUE ? "read" : "write");
 879         turnstile_broadcast(ts, queue);
 880         atomic_store_rel_ptr(&rw->rw_lock, v);
 881         turnstile_unpend(ts, TS_EXCLUSIVE_LOCK);
 882         turnstile_chain_unlock(&rw->lock_object);
 883 }
 884
 885 /*
 886  * Attempt to do a non-blocking upgrade from a read lock to a write
 887  * lock.  This will only succeed if this thread holds a single read
 888  * lock.  Returns true if the upgrade succeeded and false otherwise.
 889  */
 890 int
 891 _rw_try_upgrade(struct rwlock *rw, const char *file, int line)
 892 {
 893         uintptr_t v, x, tid;
 894         struct turnstile *ts;
 895         int success;
 896
 897         if (SCHEDULER_STOPPED())
 898                 return (1);
 899
 900         KASSERT(rw->rw_lock != RW_DESTROYED,
 901             ("rw_try_upgrade() of destroyed rwlock @ %s:%d", file, line));
 902         _rw_assert(rw, RA_RLOCKED, file, line);
 903
 904         /*
 905          * Attempt to switch from one reader to a writer.  If there
 906          * are any write waiters, then we will have to lock the
 907          * turnstile first to prevent races with another writer
 908          * calling turnstile_wait() before we have claimed this
 909          * turnstile.  So, do the simple case of no waiters first.
 910          */
 911         tid = (uintptr_t)curthread;
 912         success = 0;
 913         for (;;) {
 914                 v = rw->rw_lock;
 915                 if (RW_READERS(v) > 1)
 916                         break;
 917                 if (!(v & RW_LOCK_WAITERS)) {
 918                         success = atomic_cmpset_ptr(&rw->rw_lock, v, tid);
 919                         if (!success)
 920                                 continue;
 921                         break;
 922                 }
 923
 924                 /*
 925                  * Ok, we think we have waiters, so lock the turnstile.
 926                  */
 927                 ts = turnstile_trywait(&rw->lock_object);
 928                 v = rw->rw_lock;
 929                 if (RW_READERS(v) > 1) {
 930                         turnstile_cancel(ts);
 931                         break;
 932                 }
 933                 /*
 934                  * Try to switch from one reader to a writer again.  This time
 935                  * we honor the current state of the waiters flags.
 936                  * If we obtain the lock with the flags set, then claim
 937                  * ownership of the turnstile.
 938                  */
 939                 x = rw->rw_lock & RW_LOCK_WAITERS;
 940                 success = atomic_cmpset_ptr(&rw->rw_lock, v, tid | x);
 941                 if (success) {
 942                         if (x)
 943                                 turnstile_claim(ts);
 944                         else
 945                                 turnstile_cancel(ts);
 946                         break;
 947                 }
 948                 turnstile_cancel(ts);
 949         }
 950         LOCK_LOG_TRY("WUPGRADE", &rw->lock_object, 0, success, file, line);
 951         if (success) {
 952                 curthread->td_rw_rlocks--;
 953                 WITNESS_UPGRADE(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
 954                     file, line);
 955                 LOCKSTAT_RECORD0(LS_RW_TRYUPGRADE_UPGRADE, rw);
 956         }
 957         return (success);
 958 }
 959
 960 /*
 961  * Downgrade a write lock into a single read lock.
 962  */
 963 void
 964 _rw_downgrade(struct rwlock *rw, const char *file, int line)
 965 {
 966         struct turnstile *ts;
 967         uintptr_t tid, v;
 968         int rwait, wwait;
 969
 970         if (SCHEDULER_STOPPED())
 971                 return;
 972
 973         KASSERT(rw->rw_lock != RW_DESTROYED,
 974             ("rw_downgrade() of destroyed rwlock @ %s:%d", file, line));
 975         _rw_assert(rw, RA_WLOCKED | RA_NOTRECURSED, file, line);
 976 #ifndef INVARIANTS
 977         if (rw_recursed(rw))
 978                 panic("downgrade of a recursed lock");
 979 #endif
 980
 981         WITNESS_DOWNGRADE(&rw->lock_object, 0, file, line);
 982
 983         /*
 984          * Convert from a writer to a single reader.  First we handle
 985          * the easy case with no waiters.  If there are any waiters, we
 986          * lock the turnstile and "disown" the lock.
 987          */
 988         tid = (uintptr_t)curthread;
 989         if (atomic_cmpset_rel_ptr(&rw->rw_lock, tid, RW_READERS_LOCK(1)))
 990                 goto out;
 991
 992         /*
 993          * Ok, we think we have waiters, so lock the turnstile so we can
 994          * read the waiter flags without any races.
 995          */
 996         turnstile_chain_lock(&rw->lock_object);
 997         v = rw->rw_lock & RW_LOCK_WAITERS;
 998         rwait = v & RW_LOCK_READ_WAITERS;
 999         wwait = v & RW_LOCK_WRITE_WAITERS;
1000         MPASS(rwait | wwait);
1001
1002         /*
1003          * Downgrade from a write lock while preserving waiters flag
1004          * and give up ownership of the turnstile.
1005          */
1006         ts = turnstile_lookup(&rw->lock_object);
1007         MPASS(ts != NULL);
1008         if (!wwait)
1009                 v &= ~RW_LOCK_READ_WAITERS;
1010         atomic_store_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v);
1011         /*
1012          * Wake other readers if there are no writers pending.  Otherwise they
1013          * won't be able to acquire the lock anyway.
1014          */
1015         if (rwait && !wwait) {
1016                 turnstile_broadcast(ts, TS_SHARED_QUEUE);
1017                 turnstile_unpend(ts, TS_EXCLUSIVE_LOCK);
1018         } else
1019                 turnstile_disown(ts);
1020         turnstile_chain_unlock(&rw->lock_object);
1021 out:
1022         curthread->td_rw_rlocks++;
1023         LOCK_LOG_LOCK("WDOWNGRADE", &rw->lock_object, 0, 0, file, line);
1024         LOCKSTAT_RECORD0(LS_RW_DOWNGRADE_DOWNGRADE, rw);
1025 }
1026
1027 #ifdef INVARIANT_SUPPORT
1028 #ifndef INVARIANTS
1029 #undef _rw_assert
1030 #endif
1031
1032 /*
1033  * In the non-WITNESS case, rw_assert() can only detect that at least
1034  * *some* thread owns an rlock, but it cannot guarantee that *this*
1035  * thread owns an rlock.
1036  */
1037 void
1038 _rw_assert(struct rwlock *rw, int what, const char *file, int line)
1039 {
1040
1041         if (panicstr != NULL)
1042                 return;
1043         switch (what) {
1044         case RA_LOCKED:
1045         case RA_LOCKED | RA_RECURSED:
1046         case RA_LOCKED | RA_NOTRECURSED:
1047         case RA_RLOCKED:
1048         case RA_RLOCKED | RA_RECURSED:
1049         case RA_RLOCKED | RA_NOTRECURSED:
1050 #ifdef WITNESS
1051                 witness_assert(&rw->lock_object, what, file, line);
1052 #else
1053                 /*
1054                  * If some other thread has a write lock or we have one
1055                  * and are asserting a read lock, fail.  Also, if no one
1056                  * has a lock at all, fail.
1057                  */
1058                 if (rw->rw_lock == RW_UNLOCKED ||
1059                     (!(rw->rw_lock & RW_LOCK_READ) && (what & RA_RLOCKED ||
1060                     rw_wowner(rw) != curthread)))
1061                         panic("Lock %s not %slocked @ %s:%d\n",
1062                             rw->lock_object.lo_name, (what & RA_RLOCKED) ?
1063                             "read " : "", file, line);
1064
1065                 if (!(rw->rw_lock & RW_LOCK_READ) && !(what & RA_RLOCKED)) {
1066                         if (rw_recursed(rw)) {
1067                                 if (what & RA_NOTRECURSED)
1068                                         panic("Lock %s recursed @ %s:%d\n",
1069                                             rw->lock_object.lo_name, file,
1070                                             line);
1071                         } else if (what & RA_RECURSED)
1072                                 panic("Lock %s not recursed @ %s:%d\n",
1073                                     rw->lock_object.lo_name, file, line);
1074                 }
1075 #endif
1076                 break;
1077         case RA_WLOCKED:
1078         case RA_WLOCKED | RA_RECURSED:
1079         case RA_WLOCKED | RA_NOTRECURSED:
1080                 if (rw_wowner(rw) != curthread)
1081                         panic("Lock %s not exclusively locked @ %s:%d\n",
1082                             rw->lock_object.lo_name, file, line);
1083                 if (rw_recursed(rw)) {
1084                         if (what & RA_NOTRECURSED)
1085                                 panic("Lock %s recursed @ %s:%d\n",
1086                                     rw->lock_object.lo_name, file, line);
1087                 } else if (what & RA_RECURSED)
1088                         panic("Lock %s not recursed @ %s:%d\n",
1089                             rw->lock_object.lo_name, file, line);
1090                 break;
1091         case RA_UNLOCKED:
1092 #ifdef WITNESS
1093                 witness_assert(&rw->lock_object, what, file, line);
1094 #else
1095                 /*
1096                  * If we hold a write lock fail.  We can't reliably check
1097                  * to see if we hold a read lock or not.
1098                  */
1099                 if (rw_wowner(rw) == curthread)
1100                         panic("Lock %s exclusively locked @ %s:%d\n",
1101                             rw->lock_object.lo_name, file, line);
1102 #endif
1103                 break;
1104         default:
1105                 panic("Unknown rw lock assertion: %d @ %s:%d", what, file,
1106                     line);
1107         }
1108 }
1109 #endif /* INVARIANT_SUPPORT */
1110
1111 #ifdef DDB
1112 void
1113 db_show_rwlock(struct lock_object *lock)
1114 {
1115         struct rwlock *rw;
1116         struct thread *td;
1117
1118         rw = (struct rwlock *)lock;
1119
1120         db_printf(" state: ");
1121         if (rw->rw_lock == RW_UNLOCKED)
1122                 db_printf("UNLOCKED\n");
1123         else if (rw->rw_lock == RW_DESTROYED) {
1124                 db_printf("DESTROYED\n");
1125                 return;
1126         } else if (rw->rw_lock & RW_LOCK_READ)
1127                 db_printf("RLOCK: %ju locks\n",
1128                     (uintmax_t)(RW_READERS(rw->rw_lock)));
1129         else {
1130                 td = rw_wowner(rw);
1131                 db_printf("WLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1132                     td->td_tid, td->td_proc->p_pid, td->td_name);
1133                 if (rw_recursed(rw))
1134                         db_printf(" recursed: %u\n", rw->rw_recurse);
1135         }
1136         db_printf(" waiters: ");
1137         switch (rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS)) {
1138         case RW_LOCK_READ_WAITERS:
1139                 db_printf("readers\n");
1140                 break;
1141         case RW_LOCK_WRITE_WAITERS:
1142                 db_printf("writers\n");
1143                 break;
1144         case RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS:
1145                 db_printf("readers and writers\n");
1146                 break;
1147         default:
1148                 db_printf("none\n");
1149                 break;
1150         }
1151 }
1152
1153 #endif