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  *
  14  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
  15  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  16  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  17  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
  18  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  19  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  20  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  21  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  22  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  23  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  24  * SUCH DAMAGE.
  25  */
  26
  27 /*
  28  * Machine independent bits of reader/writer lock implementation.
  29  */
  30
  31 #include <sys/cdefs.h>
  32 __FBSDID("$FreeBSD$");
  33
  34 #include "opt_ddb.h"
  35 #include "opt_hwpmc_hooks.h"
  36 #include "opt_kdtrace.h"
  37 #include "opt_no_adaptive_rwlocks.h"
  38
  39 #include <sys/param.h>
  40 #include <sys/kdb.h>
  41 #include <sys/ktr.h>
  42 #include <sys/kernel.h>
  43 #include <sys/lock.h>
  44 #include <sys/mutex.h>
  45 #include <sys/proc.h>
  46 #include <sys/rwlock.h>
  47 #include <sys/sched.h>
  48 #include <sys/smp.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 /*
  65  * Return the rwlock address when the lock cookie address is provided.
  66  * This functionality assumes that struct rwlock* have a member named rw_lock.
  67  */
  68 #define rwlock2rw(c)    (__containerof(c, struct rwlock, rw_lock))
  69
  70 #ifdef DDB
  71 #include <ddb/ddb.h>
  72
  73 static void     db_show_rwlock(const struct lock_object *lock);
  74 #endif
  75 static void     assert_rw(const struct lock_object *lock, int what);
  76 static void     lock_rw(struct lock_object *lock, uintptr_t how);
  77 #ifdef KDTRACE_HOOKS
  78 static int      owner_rw(const struct lock_object *lock, struct thread **owner);
  79 #endif
  80 static uintptr_t unlock_rw(struct lock_object *lock);
  81
  82 struct lock_class lock_class_rw = {
  83         .lc_name = "rw",
  84         .lc_flags = LC_SLEEPLOCK | LC_RECURSABLE | LC_UPGRADABLE,
  85         .lc_assert = assert_rw,
  86 #ifdef DDB
  87         .lc_ddb_show = db_show_rwlock,
  88 #endif
  89         .lc_lock = lock_rw,
  90         .lc_unlock = unlock_rw,
  91 #ifdef KDTRACE_HOOKS
  92         .lc_owner = owner_rw,
  93 #endif
  94 };
  95
  96 #ifdef ADAPTIVE_RWLOCKS
  97 static int rowner_retries = 10;
  98 static int rowner_loops = 10000;
  99 static SYSCTL_NODE(_debug, OID_AUTO, rwlock, CTLFLAG_RD, NULL,
 100     "rwlock debugging");
 101 SYSCTL_INT(_debug_rwlock, OID_AUTO, retry, CTLFLAG_RW, &rowner_retries, 0, "");
 102 SYSCTL_INT(_debug_rwlock, OID_AUTO, loops, CTLFLAG_RW, &rowner_loops, 0, "");
 103
 104 static struct lock_delay_config rw_delay = {
 105         .initial        = 1000,
 106         .step           = 500,
 107         .min            = 100,
 108         .max            = 5000,
 109 };
 110
 111 SYSCTL_INT(_debug_rwlock, OID_AUTO, delay_initial, CTLFLAG_RW, &rw_delay.initial,
 112     0, "");
 113 SYSCTL_INT(_debug_rwlock, OID_AUTO, delay_step, CTLFLAG_RW, &rw_delay.step,
 114     0, "");
 115 SYSCTL_INT(_debug_rwlock, OID_AUTO, delay_min, CTLFLAG_RW, &rw_delay.min,
 116     0, "");
 117 SYSCTL_INT(_debug_rwlock, OID_AUTO, delay_max, CTLFLAG_RW, &rw_delay.max,
 118     0, "");
 119
 120 static void
 121 rw_delay_sysinit(void *dummy)
 122 {
 123
 124         rw_delay.initial = mp_ncpus * 25;
 125         rw_delay.step = (mp_ncpus * 25) / 2;
 126         rw_delay.min = mp_ncpus * 5;
 127         rw_delay.max = mp_ncpus * 25 * 10;
 128 }
 129 LOCK_DELAY_SYSINIT(rw_delay_sysinit);
 130 #endif
 131
 132 /*
 133  * Return a pointer to the owning thread if the lock is write-locked or
 134  * NULL if the lock is unlocked or read-locked.
 135  */
 136 #define rw_wowner(rw)                                                   \
 137         ((rw)->rw_lock & RW_LOCK_READ ? NULL :                          \
 138             (struct thread *)RW_OWNER((rw)->rw_lock))
 139
 140 /*
 141  * Returns if a write owner is recursed.  Write ownership is not assured
 142  * here and should be previously checked.
 143  */
 144 #define rw_recursed(rw)         ((rw)->rw_recurse != 0)
 145
 146 /*
 147  * Return true if curthread helds the lock.
 148  */
 149 #define rw_wlocked(rw)          (rw_wowner((rw)) == curthread)
 150
 151 /*
 152  * Return a pointer to the owning thread for this lock who should receive
 153  * any priority lent by threads that block on this lock.  Currently this
 154  * is identical to rw_wowner().
 155  */
 156 #define rw_owner(rw)            rw_wowner(rw)
 157
 158 #ifndef INVARIANTS
 159 #define __rw_assert(c, what, file, line)
 160 #endif
 161
 162 void
 163 assert_rw(const struct lock_object *lock, int what)
 164 {
 165
 166         rw_assert((const struct rwlock *)lock, what);
 167 }
 168
 169 void
 170 lock_rw(struct lock_object *lock, uintptr_t how)
 171 {
 172         struct rwlock *rw;
 173
 174         rw = (struct rwlock *)lock;
 175         if (how)
 176                 rw_rlock(rw);
 177         else
 178                 rw_wlock(rw);
 179 }
 180
 181 uintptr_t
 182 unlock_rw(struct lock_object *lock)
 183 {
 184         struct rwlock *rw;
 185
 186         rw = (struct rwlock *)lock;
 187         rw_assert(rw, RA_LOCKED | LA_NOTRECURSED);
 188         if (rw->rw_lock & RW_LOCK_READ) {
 189                 rw_runlock(rw);
 190                 return (1);
 191         } else {
 192                 rw_wunlock(rw);
 193                 return (0);
 194         }
 195 }
 196
 197 #ifdef KDTRACE_HOOKS
 198 int
 199 owner_rw(const struct lock_object *lock, struct thread **owner)
 200 {
 201         const struct rwlock *rw = (const struct rwlock *)lock;
 202         uintptr_t x = rw->rw_lock;
 203
 204         *owner = rw_wowner(rw);
 205         return ((x & RW_LOCK_READ) != 0 ?  (RW_READERS(x) != 0) :
 206             (*owner != NULL));
 207 }
 208 #endif
 209
 210 void
 211 _rw_init_flags(volatile uintptr_t *c, const char *name, int opts)
 212 {
 213         struct rwlock *rw;
 214         int flags;
 215
 216         rw = rwlock2rw(c);
 217
 218         MPASS((opts & ~(RW_DUPOK | RW_NOPROFILE | RW_NOWITNESS | RW_QUIET |
 219             RW_RECURSE)) == 0);
 220         ASSERT_ATOMIC_LOAD_PTR(rw->rw_lock,
 221             ("%s: rw_lock not aligned for %s: %p", __func__, name,
 222             &rw->rw_lock));
 223
 224         flags = LO_UPGRADABLE;
 225         if (opts & RW_DUPOK)
 226                 flags |= LO_DUPOK;
 227         if (opts & RW_NOPROFILE)
 228                 flags |= LO_NOPROFILE;
 229         if (!(opts & RW_NOWITNESS))
 230                 flags |= LO_WITNESS;
 231         if (opts & RW_RECURSE)
 232                 flags |= LO_RECURSABLE;
 233         if (opts & RW_QUIET)
 234                 flags |= LO_QUIET;
 235
 236         lock_init(&rw->lock_object, &lock_class_rw, name, NULL, flags);
 237         rw->rw_lock = RW_UNLOCKED;
 238         rw->rw_recurse = 0;
 239 }
 240
 241 void
 242 _rw_destroy(volatile uintptr_t *c)
 243 {
 244         struct rwlock *rw;
 245
 246         rw = rwlock2rw(c);
 247
 248         KASSERT(rw->rw_lock == RW_UNLOCKED, ("rw lock %p not unlocked", rw));
 249         KASSERT(rw->rw_recurse == 0, ("rw lock %p still recursed", rw));
 250         rw->rw_lock = RW_DESTROYED;
 251         lock_destroy(&rw->lock_object);
 252 }
 253
 254 void
 255 rw_sysinit(void *arg)
 256 {
 257         struct rw_args *args = arg;
 258
 259         rw_init((struct rwlock *)args->ra_rw, args->ra_desc);
 260 }
 261
 262 void
 263 rw_sysinit_flags(void *arg)
 264 {
 265         struct rw_args_flags *args = arg;
 266
 267         rw_init_flags((struct rwlock *)args->ra_rw, args->ra_desc,
 268             args->ra_flags);
 269 }
 270
 271 int
 272 _rw_wowned(const volatile uintptr_t *c)
 273 {
 274
 275         return (rw_wowner(rwlock2rw(c)) == curthread);
 276 }
 277
 278 void
 279 _rw_wlock_cookie(volatile uintptr_t *c, const char *file, int line)
 280 {
 281         struct rwlock *rw;
 282
 283         if (SCHEDULER_STOPPED())
 284                 return;
 285
 286         rw = rwlock2rw(c);
 287
 288         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
 289             ("rw_wlock() by idle thread %p on rwlock %s @ %s:%d",
 290             curthread, rw->lock_object.lo_name, file, line));
 291         KASSERT(rw->rw_lock != RW_DESTROYED,
 292             ("rw_wlock() of destroyed rwlock @ %s:%d", file, line));
 293         WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE, file,
 294             line, NULL);
 295         __rw_wlock(rw, curthread, file, line);
 296         LOCK_LOG_LOCK("WLOCK", &rw->lock_object, 0, rw->rw_recurse, file, line);
 297         WITNESS_LOCK(&rw->lock_object, LOP_EXCLUSIVE, file, line);
 298         curthread->td_locks++;
 299 }
 300
 301 int
 302 __rw_try_wlock(volatile uintptr_t *c, const char *file, int line)
 303 {
 304         struct rwlock *rw;
 305         int rval;
 306
 307         if (SCHEDULER_STOPPED())
 308                 return (1);
 309
 310         rw = rwlock2rw(c);
 311
 312         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
 313             ("rw_try_wlock() by idle thread %p on rwlock %s @ %s:%d",
 314             curthread, rw->lock_object.lo_name, file, line));
 315         KASSERT(rw->rw_lock != RW_DESTROYED,
 316             ("rw_try_wlock() of destroyed rwlock @ %s:%d", file, line));
 317
 318         if (rw_wlocked(rw) &&
 319             (rw->lock_object.lo_flags & LO_RECURSABLE) != 0) {
 320                 rw->rw_recurse++;
 321                 rval = 1;
 322         } else
 323                 rval = atomic_cmpset_acq_ptr(&rw->rw_lock, RW_UNLOCKED,
 324                     (uintptr_t)curthread);
 325
 326         LOCK_LOG_TRY("WLOCK", &rw->lock_object, 0, rval, file, line);
 327         if (rval) {
 328                 WITNESS_LOCK(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
 329                     file, line);
 330                 if (!rw_recursed(rw))
 331                         LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_RW_WLOCK_ACQUIRE,
 332                             rw, 0, 0, file, line);
 333                 curthread->td_locks++;
 334         }
 335         return (rval);
 336 }
 337
 338 void
 339 _rw_wunlock_cookie(volatile uintptr_t *c, const char *file, int line)
 340 {
 341         struct rwlock *rw;
 342
 343         if (SCHEDULER_STOPPED())
 344                 return;
 345
 346         rw = rwlock2rw(c);
 347
 348         KASSERT(rw->rw_lock != RW_DESTROYED,
 349             ("rw_wunlock() of destroyed rwlock @ %s:%d", file, line));
 350         __rw_assert(c, RA_WLOCKED, file, line);
 351         WITNESS_UNLOCK(&rw->lock_object, LOP_EXCLUSIVE, file, line);
 352         LOCK_LOG_LOCK("WUNLOCK", &rw->lock_object, 0, rw->rw_recurse, file,
 353             line);
 354         if (!rw_recursed(rw))
 355                 LOCKSTAT_PROFILE_RELEASE_LOCK(LS_RW_WUNLOCK_RELEASE, rw);
 356         __rw_wunlock(rw, curthread, file, line);
 357         curthread->td_locks--;
 358 }
 359 /*
 360  * Determines whether a new reader can acquire a lock.  Succeeds if the
 361  * reader already owns a read lock and the lock is locked for read to
 362  * prevent deadlock from reader recursion.  Also succeeds if the lock
 363  * is unlocked and has no writer waiters or spinners.  Failing otherwise
 364  * prioritizes writers before readers.
 365  */
 366 #define RW_CAN_READ(_rw)                                                \
 367     ((curthread->td_rw_rlocks && (_rw) & RW_LOCK_READ) || ((_rw) &      \
 368     (RW_LOCK_READ | RW_LOCK_WRITE_WAITERS | RW_LOCK_WRITE_SPINNER)) ==  \
 369     RW_LOCK_READ)
 370
 371 void
 372 __rw_rlock(volatile uintptr_t *c, const char *file, int line)
 373 {
 374         struct rwlock *rw;
 375         struct turnstile *ts;
 376 #ifdef ADAPTIVE_RWLOCKS
 377         volatile struct thread *owner;
 378         int spintries = 0;
 379         int i;
 380 #endif
 381 #ifdef LOCK_PROFILING
 382         uint64_t waittime = 0;
 383         int contested = 0;
 384 #endif
 385         uintptr_t v;
 386 #if defined(ADAPTIVE_RWLOCKS) || defined(KDTRACE_HOOKS)
 387         struct lock_delay_arg lda;
 388 #endif
 389 #ifdef KDTRACE_HOOKS
 390         uintptr_t state;
 391         u_int sleep_cnt = 0;
 392         int64_t sleep_time = 0;
 393         int64_t all_time = 0;
 394 #endif
 395
 396         if (SCHEDULER_STOPPED())
 397                 return;
 398
 399 #if defined(ADAPTIVE_RWLOCKS)
 400         lock_delay_arg_init(&lda, &rw_delay);
 401 #elif defined(KDTRACE_HOOKS)
 402         lock_delay_arg_init(&lda, NULL);
 403 #endif
 404         rw = rwlock2rw(c);
 405
 406         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
 407             ("rw_rlock() by idle thread %p on rwlock %s @ %s:%d",
 408             curthread, rw->lock_object.lo_name, file, line));
 409         KASSERT(rw->rw_lock != RW_DESTROYED,
 410             ("rw_rlock() of destroyed rwlock @ %s:%d", file, line));
 411         KASSERT(rw_wowner(rw) != curthread,
 412             ("rw_rlock: wlock already held for %s @ %s:%d",
 413             rw->lock_object.lo_name, file, line));
 414         WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER, file, line, NULL);
 415
 416 #ifdef KDTRACE_HOOKS
 417         all_time -= lockstat_nsecs(&rw->lock_object);
 418         state = rw->rw_lock;
 419 #endif
 420         for (;;) {
 421                 /*
 422                  * Handle the easy case.  If no other thread has a write
 423                  * lock, then try to bump up the count of read locks.  Note
 424                  * that we have to preserve the current state of the
 425                  * RW_LOCK_WRITE_WAITERS flag.  If we fail to acquire a
 426                  * read lock, then rw_lock must have changed, so restart
 427                  * the loop.  Note that this handles the case of a
 428                  * completely unlocked rwlock since such a lock is encoded
 429                  * as a read lock with no waiters.
 430                  */
 431                 v = rw->rw_lock;
 432                 if (RW_CAN_READ(v)) {
 433                         /*
 434                          * The RW_LOCK_READ_WAITERS flag should only be set
 435                          * if the lock has been unlocked and write waiters
 436                          * were present.
 437                          */
 438                         if (atomic_cmpset_acq_ptr(&rw->rw_lock, v,
 439                             v + RW_ONE_READER)) {
 440                                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 441                                         CTR4(KTR_LOCK,
 442                                             "%s: %p succeed %p -> %p", __func__,
 443                                             rw, (void *)v,
 444                                             (void *)(v + RW_ONE_READER));
 445                                 break;
 446                         }
 447                         continue;
 448                 }
 449 #ifdef KDTRACE_HOOKS
 450                 lda.spin_cnt++;
 451 #endif
 452 #ifdef HWPMC_HOOKS
 453                 PMC_SOFT_CALL( , , lock, failed);
 454 #endif
 455                 lock_profile_obtain_lock_failed(&rw->lock_object,
 456                     &contested, &waittime);
 457
 458 #ifdef ADAPTIVE_RWLOCKS
 459                 /*
 460                  * If the owner is running on another CPU, spin until
 461                  * the owner stops running or the state of the lock
 462                  * changes.
 463                  */
 464                 if ((v & RW_LOCK_READ) == 0) {
 465                         owner = (struct thread *)RW_OWNER(v);
 466                         if (TD_IS_RUNNING(owner)) {
 467                                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 468                                         CTR3(KTR_LOCK,
 469                                             "%s: spinning on %p held by %p",
 470                                             __func__, rw, owner);
 471                                 KTR_STATE1(KTR_SCHED, "thread",
 472                                     sched_tdname(curthread), "spinning",
 473                                     "lockname:\"%s\"", rw->lock_object.lo_name);
 474                                 while ((struct thread*)RW_OWNER(rw->rw_lock) ==
 475                                     owner && TD_IS_RUNNING(owner))
 476                                         lock_delay(&lda);
 477                                 KTR_STATE0(KTR_SCHED, "thread",
 478                                     sched_tdname(curthread), "running");
 479                                 continue;
 480                         }
 481                 } else if (spintries < rowner_retries) {
 482                         spintries++;
 483                         KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
 484                             "spinning", "lockname:\"%s\"",
 485                             rw->lock_object.lo_name);
 486                         for (i = 0; i < rowner_loops; i++) {
 487                                 v = rw->rw_lock;
 488                                 if ((v & RW_LOCK_READ) == 0 || RW_CAN_READ(v))
 489                                         break;
 490                                 cpu_spinwait();
 491                         }
 492 #ifdef KDTRACE_HOOKS
 493                         lda.spin_cnt += rowner_loops - i;
 494 #endif
 495                         KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
 496                             "running");
 497                         if (i != rowner_loops)
 498                                 continue;
 499                 }
 500 #endif
 501
 502                 /*
 503                  * Okay, now it's the hard case.  Some other thread already
 504                  * has a write lock or there are write waiters present,
 505                  * acquire the turnstile lock so we can begin the process
 506                  * of blocking.
 507                  */
 508                 ts = turnstile_trywait(&rw->lock_object);
 509
 510                 /*
 511                  * The lock might have been released while we spun, so
 512                  * recheck its state and restart the loop if needed.
 513                  */
 514                 v = rw->rw_lock;
 515                 if (RW_CAN_READ(v)) {
 516                         turnstile_cancel(ts);
 517                         continue;
 518                 }
 519
 520 #ifdef ADAPTIVE_RWLOCKS
 521                 /*
 522                  * The current lock owner might have started executing
 523                  * on another CPU (or the lock could have changed
 524                  * owners) while we were waiting on the turnstile
 525                  * chain lock.  If so, drop the turnstile lock and try
 526                  * again.
 527                  */
 528                 if ((v & RW_LOCK_READ) == 0) {
 529                         owner = (struct thread *)RW_OWNER(v);
 530                         if (TD_IS_RUNNING(owner)) {
 531                                 turnstile_cancel(ts);
 532                                 continue;
 533                         }
 534                 }
 535 #endif
 536
 537                 /*
 538                  * The lock is held in write mode or it already has waiters.
 539                  */
 540                 MPASS(!RW_CAN_READ(v));
 541
 542                 /*
 543                  * If the RW_LOCK_READ_WAITERS flag is already set, then
 544                  * we can go ahead and block.  If it is not set then try
 545                  * to set it.  If we fail to set it drop the turnstile
 546                  * lock and restart the loop.
 547                  */
 548                 if (!(v & RW_LOCK_READ_WAITERS)) {
 549                         if (!atomic_cmpset_ptr(&rw->rw_lock, v,
 550                             v | RW_LOCK_READ_WAITERS)) {
 551                                 turnstile_cancel(ts);
 552                                 continue;
 553                         }
 554                         if (LOCK_LOG_TEST(&rw->lock_object, 0))
 555                                 CTR2(KTR_LOCK, "%s: %p set read waiters flag",
 556                                     __func__, rw);
 557                 }
 558
 559                 /*
 560                  * We were unable to acquire the lock and the read waiters
 561                  * flag is set, so we must block on the turnstile.
 562                  */
 563                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 564                         CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__,
 565                             rw);
 566 #ifdef KDTRACE_HOOKS
 567                 sleep_time -= lockstat_nsecs(&rw->lock_object);
 568 #endif
 569                 turnstile_wait(ts, rw_owner(rw), TS_SHARED_QUEUE);
 570 #ifdef KDTRACE_HOOKS
 571                 sleep_time += lockstat_nsecs(&rw->lock_object);
 572                 sleep_cnt++;
 573 #endif
 574                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 575                         CTR2(KTR_LOCK, "%s: %p resuming from turnstile",
 576                             __func__, rw);
 577         }
 578 #ifdef KDTRACE_HOOKS
 579         all_time += lockstat_nsecs(&rw->lock_object);
 580         if (sleep_time)
 581                 LOCKSTAT_RECORD4(LS_RW_RLOCK_BLOCK, rw, sleep_time,
 582                     LOCKSTAT_READER, (state & RW_LOCK_READ) == 0,
 583                     (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
 584
 585         /* Record only the loops spinning and not sleeping. */
 586         if (lda.spin_cnt > sleep_cnt)
 587                 LOCKSTAT_RECORD4(LS_RW_RLOCK_SPIN, rw, all_time - sleep_time,
 588                     LOCKSTAT_READER, (state & RW_LOCK_READ) == 0,
 589                     (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
 590 #endif
 591         /*
 592          * TODO: acquire "owner of record" here.  Here be turnstile dragons
 593          * however.  turnstiles don't like owners changing between calls to
 594          * turnstile_wait() currently.
 595          */
 596         LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_RW_RLOCK_ACQUIRE, rw, contested,
 597             waittime, file, line);
 598         LOCK_LOG_LOCK("RLOCK", &rw->lock_object, 0, 0, file, line);
 599         WITNESS_LOCK(&rw->lock_object, 0, file, line);
 600         curthread->td_locks++;
 601         curthread->td_rw_rlocks++;
 602 }
 603
 604 int
 605 __rw_try_rlock(volatile uintptr_t *c, const char *file, int line)
 606 {
 607         struct rwlock *rw;
 608         uintptr_t x;
 609
 610         if (SCHEDULER_STOPPED())
 611                 return (1);
 612
 613         rw = rwlock2rw(c);
 614
 615         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
 616             ("rw_try_rlock() by idle thread %p on rwlock %s @ %s:%d",
 617             curthread, rw->lock_object.lo_name, file, line));
 618
 619         for (;;) {
 620                 x = rw->rw_lock;
 621                 KASSERT(rw->rw_lock != RW_DESTROYED,
 622                     ("rw_try_rlock() of destroyed rwlock @ %s:%d", file, line));
 623                 if (!(x & RW_LOCK_READ))
 624                         break;
 625                 if (atomic_cmpset_acq_ptr(&rw->rw_lock, x, x + RW_ONE_READER)) {
 626                         LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 1, file,
 627                             line);
 628                         WITNESS_LOCK(&rw->lock_object, LOP_TRYLOCK, file, line);
 629                         LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_RW_RLOCK_ACQUIRE,
 630                             rw, 0, 0, file, line);
 631                         curthread->td_locks++;
 632                         curthread->td_rw_rlocks++;
 633                         return (1);
 634                 }
 635         }
 636
 637         LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 0, file, line);
 638         return (0);
 639 }
 640
 641 void
 642 _rw_runlock_cookie(volatile uintptr_t *c, const char *file, int line)
 643 {
 644         struct rwlock *rw;
 645         struct turnstile *ts;
 646         uintptr_t x, v, queue;
 647
 648         if (SCHEDULER_STOPPED())
 649                 return;
 650
 651         rw = rwlock2rw(c);
 652
 653         KASSERT(rw->rw_lock != RW_DESTROYED,
 654             ("rw_runlock() of destroyed rwlock @ %s:%d", file, line));
 655         __rw_assert(c, RA_RLOCKED, file, line);
 656         WITNESS_UNLOCK(&rw->lock_object, 0, file, line);
 657         LOCK_LOG_LOCK("RUNLOCK", &rw->lock_object, 0, 0, file, line);
 658
 659         /* TODO: drop "owner of record" here. */
 660
 661         for (;;) {
 662                 /*
 663                  * See if there is more than one read lock held.  If so,
 664                  * just drop one and return.
 665                  */
 666                 x = rw->rw_lock;
 667                 if (RW_READERS(x) > 1) {
 668                         if (atomic_cmpset_rel_ptr(&rw->rw_lock, x,
 669                             x - RW_ONE_READER)) {
 670                                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 671                                         CTR4(KTR_LOCK,
 672                                             "%s: %p succeeded %p -> %p",
 673                                             __func__, rw, (void *)x,
 674                                             (void *)(x - RW_ONE_READER));
 675                                 break;
 676                         }
 677                         continue;
 678                 }
 679                 /*
 680                  * If there aren't any waiters for a write lock, then try
 681                  * to drop it quickly.
 682                  */
 683                 if (!(x & RW_LOCK_WAITERS)) {
 684                         MPASS((x & ~RW_LOCK_WRITE_SPINNER) ==
 685                             RW_READERS_LOCK(1));
 686                         if (atomic_cmpset_rel_ptr(&rw->rw_lock, x,
 687                             RW_UNLOCKED)) {
 688                                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 689                                         CTR2(KTR_LOCK, "%s: %p last succeeded",
 690                                             __func__, rw);
 691                                 break;
 692                         }
 693                         continue;
 694                 }
 695                 /*
 696                  * Ok, we know we have waiters and we think we are the
 697                  * last reader, so grab the turnstile lock.
 698                  */
 699                 turnstile_chain_lock(&rw->lock_object);
 700                 v = rw->rw_lock & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER);
 701                 MPASS(v & RW_LOCK_WAITERS);
 702
 703                 /*
 704                  * Try to drop our lock leaving the lock in a unlocked
 705                  * state.
 706                  *
 707                  * If you wanted to do explicit lock handoff you'd have to
 708                  * do it here.  You'd also want to use turnstile_signal()
 709                  * and you'd have to handle the race where a higher
 710                  * priority thread blocks on the write lock before the
 711                  * thread you wakeup actually runs and have the new thread
 712                  * "steal" the lock.  For now it's a lot simpler to just
 713                  * wakeup all of the waiters.
 714                  *
 715                  * As above, if we fail, then another thread might have
 716                  * acquired a read lock, so drop the turnstile lock and
 717                  * restart.
 718                  */
 719                 x = RW_UNLOCKED;
 720                 if (v & RW_LOCK_WRITE_WAITERS) {
 721                         queue = TS_EXCLUSIVE_QUEUE;
 722                         x |= (v & RW_LOCK_READ_WAITERS);
 723                 } else
 724                         queue = TS_SHARED_QUEUE;
 725                 if (!atomic_cmpset_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v,
 726                     x)) {
 727                         turnstile_chain_unlock(&rw->lock_object);
 728                         continue;
 729                 }
 730                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 731                         CTR2(KTR_LOCK, "%s: %p last succeeded with waiters",
 732                             __func__, rw);
 733
 734                 /*
 735                  * Ok.  The lock is released and all that's left is to
 736                  * wake up the waiters.  Note that the lock might not be
 737                  * free anymore, but in that case the writers will just
 738                  * block again if they run before the new lock holder(s)
 739                  * release the lock.
 740                  */
 741                 ts = turnstile_lookup(&rw->lock_object);
 742                 MPASS(ts != NULL);
 743                 turnstile_broadcast(ts, queue);
 744                 turnstile_unpend(ts, TS_SHARED_LOCK);
 745                 turnstile_chain_unlock(&rw->lock_object);
 746                 break;
 747         }
 748         LOCKSTAT_PROFILE_RELEASE_LOCK(LS_RW_RUNLOCK_RELEASE, rw);
 749         curthread->td_locks--;
 750         curthread->td_rw_rlocks--;
 751 }
 752
 753 /*
 754  * This function is called when we are unable to obtain a write lock on the
 755  * first try.  This means that at least one other thread holds either a
 756  * read or write lock.
 757  */
 758 void
 759 __rw_wlock_hard(volatile uintptr_t *c, uintptr_t tid, const char *file,
 760     int line)
 761 {
 762         struct rwlock *rw;
 763         struct turnstile *ts;
 764 #ifdef ADAPTIVE_RWLOCKS
 765         volatile struct thread *owner;
 766         int spintries = 0;
 767         int i;
 768 #endif
 769         uintptr_t v, x;
 770 #ifdef LOCK_PROFILING
 771         uint64_t waittime = 0;
 772         int contested = 0;
 773 #endif
 774 #if defined(ADAPTIVE_RWLOCKS) || defined(KDTRACE_HOOKS)
 775         struct lock_delay_arg lda;
 776 #endif
 777 #ifdef KDTRACE_HOOKS
 778         uintptr_t state;
 779         u_int sleep_cnt = 0;
 780         int64_t sleep_time = 0;
 781         int64_t all_time = 0;
 782 #endif
 783
 784         if (SCHEDULER_STOPPED())
 785                 return;
 786
 787 #if defined(ADAPTIVE_RWLOCKS)
 788         lock_delay_arg_init(&lda, &rw_delay);
 789 #elif defined(KDTRACE_HOOKS)
 790         lock_delay_arg_init(&lda, NULL);
 791 #endif
 792         rw = rwlock2rw(c);
 793
 794         if (rw_wlocked(rw)) {
 795                 KASSERT(rw->lock_object.lo_flags & LO_RECURSABLE,
 796                     ("%s: recursing but non-recursive rw %s @ %s:%d\n",
 797                     __func__, rw->lock_object.lo_name, file, line));
 798                 rw->rw_recurse++;
 799                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 800                         CTR2(KTR_LOCK, "%s: %p recursing", __func__, rw);
 801                 return;
 802         }
 803
 804         if (LOCK_LOG_TEST(&rw->lock_object, 0))
 805                 CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__,
 806                     rw->lock_object.lo_name, (void *)rw->rw_lock, file, line);
 807
 808 #ifdef KDTRACE_HOOKS
 809         all_time -= lockstat_nsecs(&rw->lock_object);
 810         state = rw->rw_lock;
 811 #endif
 812         for (;;) {
 813                 if (rw->rw_lock == RW_UNLOCKED && _rw_write_lock(rw, tid))
 814                         break;
 815 #ifdef KDTRACE_HOOKS
 816                 lda.spin_cnt++;
 817 #endif
 818 #ifdef HWPMC_HOOKS
 819                 PMC_SOFT_CALL( , , lock, failed);
 820 #endif
 821                 lock_profile_obtain_lock_failed(&rw->lock_object,
 822                     &contested, &waittime);
 823 #ifdef ADAPTIVE_RWLOCKS
 824                 /*
 825                  * If the lock is write locked and the owner is
 826                  * running on another CPU, spin until the owner stops
 827                  * running or the state of the lock changes.
 828                  */
 829                 v = rw->rw_lock;
 830                 owner = (struct thread *)RW_OWNER(v);
 831                 if (!(v & RW_LOCK_READ) && TD_IS_RUNNING(owner)) {
 832                         if (LOCK_LOG_TEST(&rw->lock_object, 0))
 833                                 CTR3(KTR_LOCK, "%s: spinning on %p held by %p",
 834                                     __func__, rw, owner);
 835                         KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
 836                             "spinning", "lockname:\"%s\"",
 837                             rw->lock_object.lo_name);
 838                         while ((struct thread*)RW_OWNER(rw->rw_lock) == owner &&
 839                             TD_IS_RUNNING(owner))
 840                                 lock_delay(&lda);
 841                         KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
 842                             "running");
 843                         continue;
 844                 }
 845                 if ((v & RW_LOCK_READ) && RW_READERS(v) &&
 846                     spintries < rowner_retries) {
 847                         if (!(v & RW_LOCK_WRITE_SPINNER)) {
 848                                 if (!atomic_cmpset_ptr(&rw->rw_lock, v,
 849                                     v | RW_LOCK_WRITE_SPINNER)) {
 850                                         continue;
 851                                 }
 852                         }
 853                         spintries++;
 854                         KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
 855                             "spinning", "lockname:\"%s\"",
 856                             rw->lock_object.lo_name);
 857                         for (i = 0; i < rowner_loops; i++) {
 858                                 if ((rw->rw_lock & RW_LOCK_WRITE_SPINNER) == 0)
 859                                         break;
 860                                 cpu_spinwait();
 861                         }
 862                         KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
 863                             "running");
 864 #ifdef KDTRACE_HOOKS
 865                         lda.spin_cnt += rowner_loops - i;
 866 #endif
 867                         if (i != rowner_loops)
 868                                 continue;
 869                 }
 870 #endif
 871                 ts = turnstile_trywait(&rw->lock_object);
 872                 v = rw->rw_lock;
 873
 874 #ifdef ADAPTIVE_RWLOCKS
 875                 /*
 876                  * The current lock owner might have started executing
 877                  * on another CPU (or the lock could have changed
 878                  * owners) while we were waiting on the turnstile
 879                  * chain lock.  If so, drop the turnstile lock and try
 880                  * again.
 881                  */
 882                 if (!(v & RW_LOCK_READ)) {
 883                         owner = (struct thread *)RW_OWNER(v);
 884                         if (TD_IS_RUNNING(owner)) {
 885                                 turnstile_cancel(ts);
 886                                 continue;
 887                         }
 888                 }
 889 #endif
 890                 /*
 891                  * Check for the waiters flags about this rwlock.
 892                  * If the lock was released, without maintain any pending
 893                  * waiters queue, simply try to acquire it.
 894                  * If a pending waiters queue is present, claim the lock
 895                  * ownership and maintain the pending queue.
 896                  */
 897                 x = v & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER);
 898                 if ((v & ~x) == RW_UNLOCKED) {
 899                         x &= ~RW_LOCK_WRITE_SPINNER;
 900                         if (atomic_cmpset_acq_ptr(&rw->rw_lock, v, tid | x)) {
 901                                 if (x)
 902                                         turnstile_claim(ts);
 903                                 else
 904                                         turnstile_cancel(ts);
 905                                 break;
 906                         }
 907                         turnstile_cancel(ts);
 908                         continue;
 909                 }
 910                 /*
 911                  * If the RW_LOCK_WRITE_WAITERS flag isn't set, then try to
 912                  * set it.  If we fail to set it, then loop back and try
 913                  * again.
 914                  */
 915                 if (!(v & RW_LOCK_WRITE_WAITERS)) {
 916                         if (!atomic_cmpset_ptr(&rw->rw_lock, v,
 917                             v | RW_LOCK_WRITE_WAITERS)) {
 918                                 turnstile_cancel(ts);
 919                                 continue;
 920                         }
 921                         if (LOCK_LOG_TEST(&rw->lock_object, 0))
 922                                 CTR2(KTR_LOCK, "%s: %p set write waiters flag",
 923                                     __func__, rw);
 924                 }
 925                 /*
 926                  * We were unable to acquire the lock and the write waiters
 927                  * flag is set, so we must block on the turnstile.
 928                  */
 929                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 930                         CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__,
 931                             rw);
 932 #ifdef KDTRACE_HOOKS
 933                 sleep_time -= lockstat_nsecs(&rw->lock_object);
 934 #endif
 935                 turnstile_wait(ts, rw_owner(rw), TS_EXCLUSIVE_QUEUE);
 936 #ifdef KDTRACE_HOOKS
 937                 sleep_time += lockstat_nsecs(&rw->lock_object);
 938                 sleep_cnt++;
 939 #endif
 940                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 941                         CTR2(KTR_LOCK, "%s: %p resuming from turnstile",
 942                             __func__, rw);
 943 #ifdef ADAPTIVE_RWLOCKS
 944                 spintries = 0;
 945 #endif
 946         }
 947 #ifdef KDTRACE_HOOKS
 948         all_time += lockstat_nsecs(&rw->lock_object);
 949         if (sleep_time)
 950                 LOCKSTAT_RECORD4(LS_RW_WLOCK_BLOCK, rw, sleep_time,
 951                     LOCKSTAT_WRITER, (state & RW_LOCK_READ) == 0,
 952                     (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
 953
 954         /* Record only the loops spinning and not sleeping. */
 955         if (lda.spin_cnt > sleep_cnt)
 956                 LOCKSTAT_RECORD4(LS_RW_WLOCK_SPIN, rw, all_time - sleep_time,
 957                     LOCKSTAT_WRITER, (state & RW_LOCK_READ) == 0,
 958                     (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
 959 #endif
 960         LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_RW_WLOCK_ACQUIRE, rw, contested,
 961             waittime, file, line);
 962 }
 963
 964 /*
 965  * This function is called if the first try at releasing a write lock failed.
 966  * This means that one of the 2 waiter bits must be set indicating that at
 967  * least one thread is waiting on this lock.
 968  */
 969 void
 970 __rw_wunlock_hard(volatile uintptr_t *c, uintptr_t tid, const char *file,
 971     int line)
 972 {
 973         struct rwlock *rw;
 974         struct turnstile *ts;
 975         uintptr_t v;
 976         int queue;
 977
 978         if (SCHEDULER_STOPPED())
 979                 return;
 980
 981         rw = rwlock2rw(c);
 982
 983         if (rw_wlocked(rw) && rw_recursed(rw)) {
 984                 rw->rw_recurse--;
 985                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 986                         CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, rw);
 987                 return;
 988         }
 989
 990         KASSERT(rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS),
 991             ("%s: neither of the waiter flags are set", __func__));
 992
 993         if (LOCK_LOG_TEST(&rw->lock_object, 0))
 994                 CTR2(KTR_LOCK, "%s: %p contested", __func__, rw);
 995
 996         turnstile_chain_lock(&rw->lock_object);
 997         ts = turnstile_lookup(&rw->lock_object);
 998         MPASS(ts != NULL);
 999
1000         /*
1001          * Use the same algo as sx locks for now.  Prefer waking up shared
1002          * waiters if we have any over writers.  This is probably not ideal.
1003          *
1004          * 'v' is the value we are going to write back to rw_lock.  If we
1005          * have waiters on both queues, we need to preserve the state of
1006          * the waiter flag for the queue we don't wake up.  For now this is
1007          * hardcoded for the algorithm mentioned above.
1008          *
1009          * In the case of both readers and writers waiting we wakeup the
1010          * readers but leave the RW_LOCK_WRITE_WAITERS flag set.  If a
1011          * new writer comes in before a reader it will claim the lock up
1012          * above.  There is probably a potential priority inversion in
1013          * there that could be worked around either by waking both queues
1014          * of waiters or doing some complicated lock handoff gymnastics.
1015          */
1016         v = RW_UNLOCKED;
1017         if (rw->rw_lock & RW_LOCK_WRITE_WAITERS) {
1018                 queue = TS_EXCLUSIVE_QUEUE;
1019                 v |= (rw->rw_lock & RW_LOCK_READ_WAITERS);
1020         } else
1021                 queue = TS_SHARED_QUEUE;
1022
1023         /* Wake up all waiters for the specific queue. */
1024         if (LOCK_LOG_TEST(&rw->lock_object, 0))
1025                 CTR3(KTR_LOCK, "%s: %p waking up %s waiters", __func__, rw,
1026                     queue == TS_SHARED_QUEUE ? "read" : "write");
1027         turnstile_broadcast(ts, queue);
1028         atomic_store_rel_ptr(&rw->rw_lock, v);
1029         turnstile_unpend(ts, TS_EXCLUSIVE_LOCK);
1030         turnstile_chain_unlock(&rw->lock_object);
1031 }
1032
1033 /*
1034  * Attempt to do a non-blocking upgrade from a read lock to a write
1035  * lock.  This will only succeed if this thread holds a single read
1036  * lock.  Returns true if the upgrade succeeded and false otherwise.
1037  */
1038 int
1039 __rw_try_upgrade(volatile uintptr_t *c, const char *file, int line)
1040 {
1041         struct rwlock *rw;
1042         uintptr_t v, x, tid;
1043         struct turnstile *ts;
1044         int success;
1045
1046         if (SCHEDULER_STOPPED())
1047                 return (1);
1048
1049         rw = rwlock2rw(c);
1050
1051         KASSERT(rw->rw_lock != RW_DESTROYED,
1052             ("rw_try_upgrade() of destroyed rwlock @ %s:%d", file, line));
1053         __rw_assert(c, RA_RLOCKED, file, line);
1054
1055         /*
1056          * Attempt to switch from one reader to a writer.  If there
1057          * are any write waiters, then we will have to lock the
1058          * turnstile first to prevent races with another writer
1059          * calling turnstile_wait() before we have claimed this
1060          * turnstile.  So, do the simple case of no waiters first.
1061          */
1062         tid = (uintptr_t)curthread;
1063         success = 0;
1064         for (;;) {
1065                 v = rw->rw_lock;
1066                 if (RW_READERS(v) > 1)
1067                         break;
1068                 if (!(v & RW_LOCK_WAITERS)) {
1069                         success = atomic_cmpset_ptr(&rw->rw_lock, v, tid);
1070                         if (!success)
1071                                 continue;
1072                         break;
1073                 }
1074
1075                 /*
1076                  * Ok, we think we have waiters, so lock the turnstile.
1077                  */
1078                 ts = turnstile_trywait(&rw->lock_object);
1079                 v = rw->rw_lock;
1080                 if (RW_READERS(v) > 1) {
1081                         turnstile_cancel(ts);
1082                         break;
1083                 }
1084                 /*
1085                  * Try to switch from one reader to a writer again.  This time
1086                  * we honor the current state of the waiters flags.
1087                  * If we obtain the lock with the flags set, then claim
1088                  * ownership of the turnstile.
1089                  */
1090                 x = rw->rw_lock & RW_LOCK_WAITERS;
1091                 success = atomic_cmpset_ptr(&rw->rw_lock, v, tid | x);
1092                 if (success) {
1093                         if (x)
1094                                 turnstile_claim(ts);
1095                         else
1096                                 turnstile_cancel(ts);
1097                         break;
1098                 }
1099                 turnstile_cancel(ts);
1100         }
1101         LOCK_LOG_TRY("WUPGRADE", &rw->lock_object, 0, success, file, line);
1102         if (success) {
1103                 curthread->td_rw_rlocks--;
1104                 WITNESS_UPGRADE(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
1105                     file, line);
1106                 LOCKSTAT_RECORD0(LS_RW_TRYUPGRADE_UPGRADE, rw);
1107         }
1108         return (success);
1109 }
1110
1111 /*
1112  * Downgrade a write lock into a single read lock.
1113  */
1114 void
1115 __rw_downgrade(volatile uintptr_t *c, const char *file, int line)
1116 {
1117         struct rwlock *rw;
1118         struct turnstile *ts;
1119         uintptr_t tid, v;
1120         int rwait, wwait;
1121
1122         if (SCHEDULER_STOPPED())
1123                 return;
1124
1125         rw = rwlock2rw(c);
1126
1127         KASSERT(rw->rw_lock != RW_DESTROYED,
1128             ("rw_downgrade() of destroyed rwlock @ %s:%d", file, line));
1129         __rw_assert(c, RA_WLOCKED | RA_NOTRECURSED, file, line);
1130 #ifndef INVARIANTS
1131         if (rw_recursed(rw))
1132                 panic("downgrade of a recursed lock");
1133 #endif
1134
1135         WITNESS_DOWNGRADE(&rw->lock_object, 0, file, line);
1136
1137         /*
1138          * Convert from a writer to a single reader.  First we handle
1139          * the easy case with no waiters.  If there are any waiters, we
1140          * lock the turnstile and "disown" the lock.
1141          */
1142         tid = (uintptr_t)curthread;
1143         if (atomic_cmpset_rel_ptr(&rw->rw_lock, tid, RW_READERS_LOCK(1)))
1144                 goto out;
1145
1146         /*
1147          * Ok, we think we have waiters, so lock the turnstile so we can
1148          * read the waiter flags without any races.
1149          */
1150         turnstile_chain_lock(&rw->lock_object);
1151         v = rw->rw_lock & RW_LOCK_WAITERS;
1152         rwait = v & RW_LOCK_READ_WAITERS;
1153         wwait = v & RW_LOCK_WRITE_WAITERS;
1154         MPASS(rwait | wwait);
1155
1156         /*
1157          * Downgrade from a write lock while preserving waiters flag
1158          * and give up ownership of the turnstile.
1159          */
1160         ts = turnstile_lookup(&rw->lock_object);
1161         MPASS(ts != NULL);
1162         if (!wwait)
1163                 v &= ~RW_LOCK_READ_WAITERS;
1164         atomic_store_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v);
1165         /*
1166          * Wake other readers if there are no writers pending.  Otherwise they
1167          * won't be able to acquire the lock anyway.
1168          */
1169         if (rwait && !wwait) {
1170                 turnstile_broadcast(ts, TS_SHARED_QUEUE);
1171                 turnstile_unpend(ts, TS_EXCLUSIVE_LOCK);
1172         } else
1173                 turnstile_disown(ts);
1174         turnstile_chain_unlock(&rw->lock_object);
1175 out:
1176         curthread->td_rw_rlocks++;
1177         LOCK_LOG_LOCK("WDOWNGRADE", &rw->lock_object, 0, 0, file, line);
1178         LOCKSTAT_RECORD0(LS_RW_DOWNGRADE_DOWNGRADE, rw);
1179 }
1180
1181 #ifdef INVARIANT_SUPPORT
1182 #ifndef INVARIANTS
1183 #undef __rw_assert
1184 #endif
1185
1186 /*
1187  * In the non-WITNESS case, rw_assert() can only detect that at least
1188  * *some* thread owns an rlock, but it cannot guarantee that *this*
1189  * thread owns an rlock.
1190  */
1191 void
1192 __rw_assert(const volatile uintptr_t *c, int what, const char *file, int line)
1193 {
1194         const struct rwlock *rw;
1195
1196         if (panicstr != NULL)
1197                 return;
1198
1199         rw = rwlock2rw(c);
1200
1201         switch (what) {
1202         case RA_LOCKED:
1203         case RA_LOCKED | RA_RECURSED:
1204         case RA_LOCKED | RA_NOTRECURSED:
1205         case RA_RLOCKED:
1206         case RA_RLOCKED | RA_RECURSED:
1207         case RA_RLOCKED | RA_NOTRECURSED:
1208 #ifdef WITNESS
1209                 witness_assert(&rw->lock_object, what, file, line);
1210 #else
1211                 /*
1212                  * If some other thread has a write lock or we have one
1213                  * and are asserting a read lock, fail.  Also, if no one
1214                  * has a lock at all, fail.
1215                  */
1216                 if (rw->rw_lock == RW_UNLOCKED ||
1217                     (!(rw->rw_lock & RW_LOCK_READ) && (what & RA_RLOCKED ||
1218                     rw_wowner(rw) != curthread)))
1219                         panic("Lock %s not %slocked @ %s:%d\n",
1220                             rw->lock_object.lo_name, (what & RA_RLOCKED) ?
1221                             "read " : "", file, line);
1222
1223                 if (!(rw->rw_lock & RW_LOCK_READ) && !(what & RA_RLOCKED)) {
1224                         if (rw_recursed(rw)) {
1225                                 if (what & RA_NOTRECURSED)
1226                                         panic("Lock %s recursed @ %s:%d\n",
1227                                             rw->lock_object.lo_name, file,
1228                                             line);
1229                         } else if (what & RA_RECURSED)
1230                                 panic("Lock %s not recursed @ %s:%d\n",
1231                                     rw->lock_object.lo_name, file, line);
1232                 }
1233 #endif
1234                 break;
1235         case RA_WLOCKED:
1236         case RA_WLOCKED | RA_RECURSED:
1237         case RA_WLOCKED | RA_NOTRECURSED:
1238                 if (rw_wowner(rw) != curthread)
1239                         panic("Lock %s not exclusively locked @ %s:%d\n",
1240                             rw->lock_object.lo_name, file, line);
1241                 if (rw_recursed(rw)) {
1242                         if (what & RA_NOTRECURSED)
1243                                 panic("Lock %s recursed @ %s:%d\n",
1244                                     rw->lock_object.lo_name, file, line);
1245                 } else if (what & RA_RECURSED)
1246                         panic("Lock %s not recursed @ %s:%d\n",
1247                             rw->lock_object.lo_name, file, line);
1248                 break;
1249         case RA_UNLOCKED:
1250 #ifdef WITNESS
1251                 witness_assert(&rw->lock_object, what, file, line);
1252 #else
1253                 /*
1254                  * If we hold a write lock fail.  We can't reliably check
1255                  * to see if we hold a read lock or not.
1256                  */
1257                 if (rw_wowner(rw) == curthread)
1258                         panic("Lock %s exclusively locked @ %s:%d\n",
1259                             rw->lock_object.lo_name, file, line);
1260 #endif
1261                 break;
1262         default:
1263                 panic("Unknown rw lock assertion: %d @ %s:%d", what, file,
1264                     line);
1265         }
1266 }
1267 #endif /* INVARIANT_SUPPORT */
1268
1269 #ifdef DDB
1270 void
1271 db_show_rwlock(const struct lock_object *lock)
1272 {
1273         const struct rwlock *rw;
1274         struct thread *td;
1275
1276         rw = (const struct rwlock *)lock;
1277
1278         db_printf(" state: ");
1279         if (rw->rw_lock == RW_UNLOCKED)
1280                 db_printf("UNLOCKED\n");
1281         else if (rw->rw_lock == RW_DESTROYED) {
1282                 db_printf("DESTROYED\n");
1283                 return;
1284         } else if (rw->rw_lock & RW_LOCK_READ)
1285                 db_printf("RLOCK: %ju locks\n",
1286                     (uintmax_t)(RW_READERS(rw->rw_lock)));
1287         else {
1288                 td = rw_wowner(rw);
1289                 db_printf("WLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1290                     td->td_tid, td->td_proc->p_pid, td->td_name);
1291                 if (rw_recursed(rw))
1292                         db_printf(" recursed: %u\n", rw->rw_recurse);
1293         }
1294         db_printf(" waiters: ");
1295         switch (rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS)) {
1296         case RW_LOCK_READ_WAITERS:
1297                 db_printf("readers\n");
1298                 break;
1299         case RW_LOCK_WRITE_WAITERS:
1300                 db_printf("writers\n");
1301                 break;
1302         case RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS:
1303                 db_printf("readers and writers\n");
1304                 break;
1305         default:
1306                 db_printf("none\n");
1307                 break;
1308         }
1309 }
1310
1311 #endif