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_no_adaptive_rwlocks.h"
  37
  38 #include <sys/param.h>
  39 #include <sys/kdb.h>
  40 #include <sys/ktr.h>
  41 #include <sys/kernel.h>
  42 #include <sys/lock.h>
  43 #include <sys/mutex.h>
  44 #include <sys/proc.h>
  45 #include <sys/rwlock.h>
  46 #include <sys/sched.h>
  47 #include <sys/smp.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 HWPMC_HOOKS
  59 #include <sys/pmckern.h>
  60 PMC_SOFT_DECLARE( , , lock, failed);
  61 #endif
  62
  63 /*
  64  * Return the rwlock address when the lock cookie address is provided.
  65  * This functionality assumes that struct rwlock* have a member named rw_lock.
  66  */
  67 #define rwlock2rw(c)    (__containerof(c, struct rwlock, rw_lock))
  68
  69 #ifdef DDB
  70 #include <ddb/ddb.h>
  71
  72 static void     db_show_rwlock(const struct lock_object *lock);
  73 #endif
  74 static void     assert_rw(const struct lock_object *lock, int what);
  75 static void     lock_rw(struct lock_object *lock, uintptr_t how);
  76 #ifdef KDTRACE_HOOKS
  77 static int      owner_rw(const struct lock_object *lock, struct thread **owner);
  78 #endif
  79 static uintptr_t unlock_rw(struct lock_object *lock);
  80
  81 struct lock_class lock_class_rw = {
  82         .lc_name = "rw",
  83         .lc_flags = LC_SLEEPLOCK | LC_RECURSABLE | LC_UPGRADABLE,
  84         .lc_assert = assert_rw,
  85 #ifdef DDB
  86         .lc_ddb_show = db_show_rwlock,
  87 #endif
  88         .lc_lock = lock_rw,
  89         .lc_unlock = unlock_rw,
  90 #ifdef KDTRACE_HOOKS
  91         .lc_owner = owner_rw,
  92 #endif
  93 };
  94
  95 #ifdef ADAPTIVE_RWLOCKS
  96 static int rowner_retries = 10;
  97 static int rowner_loops = 10000;
  98 static SYSCTL_NODE(_debug, OID_AUTO, rwlock, CTLFLAG_RD, NULL,
  99     "rwlock debugging");
 100 SYSCTL_INT(_debug_rwlock, OID_AUTO, retry, CTLFLAG_RW, &rowner_retries, 0, "");
 101 SYSCTL_INT(_debug_rwlock, OID_AUTO, loops, CTLFLAG_RW, &rowner_loops, 0, "");
 102
 103 static struct lock_delay_config __read_mostly rw_delay = {
 104         .initial        = 1000,
 105         .step           = 500,
 106         .min            = 100,
 107         .max            = 5000,
 108 };
 109
 110 SYSCTL_INT(_debug_rwlock, OID_AUTO, delay_initial, CTLFLAG_RW, &rw_delay.initial,
 111     0, "");
 112 SYSCTL_INT(_debug_rwlock, OID_AUTO, delay_step, CTLFLAG_RW, &rw_delay.step,
 113     0, "");
 114 SYSCTL_INT(_debug_rwlock, OID_AUTO, delay_min, CTLFLAG_RW, &rw_delay.min,
 115     0, "");
 116 SYSCTL_INT(_debug_rwlock, OID_AUTO, delay_max, CTLFLAG_RW, &rw_delay.max,
 117     0, "");
 118
 119 static void
 120 rw_delay_sysinit(void *dummy)
 121 {
 122
 123         rw_delay.initial = mp_ncpus * 25;
 124         rw_delay.step = (mp_ncpus * 25) / 2;
 125         rw_delay.min = mp_ncpus * 5;
 126         rw_delay.max = mp_ncpus * 25 * 10;
 127 }
 128 LOCK_DELAY_SYSINIT(rw_delay_sysinit);
 129 #endif
 130
 131 /*
 132  * Return a pointer to the owning thread if the lock is write-locked or
 133  * NULL if the lock is unlocked or read-locked.
 134  */
 135
 136 #define lv_rw_wowner(v)                                                 \
 137         ((v) & RW_LOCK_READ ? NULL :                                    \
 138          (struct thread *)RW_OWNER((v)))
 139
 140 #define rw_wowner(rw)   lv_rw_wowner(RW_READ_VALUE(rw))
 141
 142 /*
 143  * Returns if a write owner is recursed.  Write ownership is not assured
 144  * here and should be previously checked.
 145  */
 146 #define rw_recursed(rw)         ((rw)->rw_recurse != 0)
 147
 148 /*
 149  * Return true if curthread helds the lock.
 150  */
 151 #define rw_wlocked(rw)          (rw_wowner((rw)) == curthread)
 152
 153 /*
 154  * Return a pointer to the owning thread for this lock who should receive
 155  * any priority lent by threads that block on this lock.  Currently this
 156  * is identical to rw_wowner().
 157  */
 158 #define rw_owner(rw)            rw_wowner(rw)
 159
 160 #ifndef INVARIANTS
 161 #define __rw_assert(c, what, file, line)
 162 #endif
 163
 164 void
 165 assert_rw(const struct lock_object *lock, int what)
 166 {
 167
 168         rw_assert((const struct rwlock *)lock, what);
 169 }
 170
 171 void
 172 lock_rw(struct lock_object *lock, uintptr_t how)
 173 {
 174         struct rwlock *rw;
 175
 176         rw = (struct rwlock *)lock;
 177         if (how)
 178                 rw_rlock(rw);
 179         else
 180                 rw_wlock(rw);
 181 }
 182
 183 uintptr_t
 184 unlock_rw(struct lock_object *lock)
 185 {
 186         struct rwlock *rw;
 187
 188         rw = (struct rwlock *)lock;
 189         rw_assert(rw, RA_LOCKED | LA_NOTRECURSED);
 190         if (rw->rw_lock & RW_LOCK_READ) {
 191                 rw_runlock(rw);
 192                 return (1);
 193         } else {
 194                 rw_wunlock(rw);
 195                 return (0);
 196         }
 197 }
 198
 199 #ifdef KDTRACE_HOOKS
 200 int
 201 owner_rw(const struct lock_object *lock, struct thread **owner)
 202 {
 203         const struct rwlock *rw = (const struct rwlock *)lock;
 204         uintptr_t x = rw->rw_lock;
 205
 206         *owner = rw_wowner(rw);
 207         return ((x & RW_LOCK_READ) != 0 ?  (RW_READERS(x) != 0) :
 208             (*owner != NULL));
 209 }
 210 #endif
 211
 212 void
 213 _rw_init_flags(volatile uintptr_t *c, const char *name, int opts)
 214 {
 215         struct rwlock *rw;
 216         int flags;
 217
 218         rw = rwlock2rw(c);
 219
 220         MPASS((opts & ~(RW_DUPOK | RW_NOPROFILE | RW_NOWITNESS | RW_QUIET |
 221             RW_RECURSE | RW_NEW)) == 0);
 222         ASSERT_ATOMIC_LOAD_PTR(rw->rw_lock,
 223             ("%s: rw_lock not aligned for %s: %p", __func__, name,
 224             &rw->rw_lock));
 225
 226         flags = LO_UPGRADABLE;
 227         if (opts & RW_DUPOK)
 228                 flags |= LO_DUPOK;
 229         if (opts & RW_NOPROFILE)
 230                 flags |= LO_NOPROFILE;
 231         if (!(opts & RW_NOWITNESS))
 232                 flags |= LO_WITNESS;
 233         if (opts & RW_RECURSE)
 234                 flags |= LO_RECURSABLE;
 235         if (opts & RW_QUIET)
 236                 flags |= LO_QUIET;
 237         if (opts & RW_NEW)
 238                 flags |= LO_NEW;
 239
 240         lock_init(&rw->lock_object, &lock_class_rw, name, NULL, flags);
 241         rw->rw_lock = RW_UNLOCKED;
 242         rw->rw_recurse = 0;
 243 }
 244
 245 void
 246 _rw_destroy(volatile uintptr_t *c)
 247 {
 248         struct rwlock *rw;
 249
 250         rw = rwlock2rw(c);
 251
 252         KASSERT(rw->rw_lock == RW_UNLOCKED, ("rw lock %p not unlocked", rw));
 253         KASSERT(rw->rw_recurse == 0, ("rw lock %p still recursed", rw));
 254         rw->rw_lock = RW_DESTROYED;
 255         lock_destroy(&rw->lock_object);
 256 }
 257
 258 void
 259 rw_sysinit(void *arg)
 260 {
 261         struct rw_args *args = arg;
 262
 263         rw_init((struct rwlock *)args->ra_rw, args->ra_desc);
 264 }
 265
 266 void
 267 rw_sysinit_flags(void *arg)
 268 {
 269         struct rw_args_flags *args = arg;
 270
 271         rw_init_flags((struct rwlock *)args->ra_rw, args->ra_desc,
 272             args->ra_flags);
 273 }
 274
 275 int
 276 _rw_wowned(const volatile uintptr_t *c)
 277 {
 278
 279         return (rw_wowner(rwlock2rw(c)) == curthread);
 280 }
 281
 282 void
 283 _rw_wlock_cookie(volatile uintptr_t *c, const char *file, int line)
 284 {
 285         struct rwlock *rw;
 286
 287         if (SCHEDULER_STOPPED())
 288                 return;
 289
 290         rw = rwlock2rw(c);
 291
 292         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
 293             ("rw_wlock() by idle thread %p on rwlock %s @ %s:%d",
 294             curthread, rw->lock_object.lo_name, file, line));
 295         KASSERT(rw->rw_lock != RW_DESTROYED,
 296             ("rw_wlock() of destroyed rwlock @ %s:%d", file, line));
 297         WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE, file,
 298             line, NULL);
 299         __rw_wlock(rw, curthread, file, line);
 300         LOCK_LOG_LOCK("WLOCK", &rw->lock_object, 0, rw->rw_recurse, file, line);
 301         WITNESS_LOCK(&rw->lock_object, LOP_EXCLUSIVE, file, line);
 302         TD_LOCKS_INC(curthread);
 303 }
 304
 305 int
 306 __rw_try_wlock(volatile uintptr_t *c, const char *file, int line)
 307 {
 308         struct rwlock *rw;
 309         int rval;
 310
 311         if (SCHEDULER_STOPPED())
 312                 return (1);
 313
 314         rw = rwlock2rw(c);
 315
 316         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
 317             ("rw_try_wlock() by idle thread %p on rwlock %s @ %s:%d",
 318             curthread, rw->lock_object.lo_name, file, line));
 319         KASSERT(rw->rw_lock != RW_DESTROYED,
 320             ("rw_try_wlock() of destroyed rwlock @ %s:%d", file, line));
 321
 322         if (rw_wlocked(rw) &&
 323             (rw->lock_object.lo_flags & LO_RECURSABLE) != 0) {
 324                 rw->rw_recurse++;
 325                 rval = 1;
 326         } else
 327                 rval = atomic_cmpset_acq_ptr(&rw->rw_lock, RW_UNLOCKED,
 328                     (uintptr_t)curthread);
 329
 330         LOCK_LOG_TRY("WLOCK", &rw->lock_object, 0, rval, file, line);
 331         if (rval) {
 332                 WITNESS_LOCK(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
 333                     file, line);
 334                 if (!rw_recursed(rw))
 335                         LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire,
 336                             rw, 0, 0, file, line, LOCKSTAT_WRITER);
 337                 TD_LOCKS_INC(curthread);
 338         }
 339         return (rval);
 340 }
 341
 342 void
 343 _rw_wunlock_cookie(volatile uintptr_t *c, const char *file, int line)
 344 {
 345         struct rwlock *rw;
 346
 347         if (SCHEDULER_STOPPED())
 348                 return;
 349
 350         rw = rwlock2rw(c);
 351
 352         KASSERT(rw->rw_lock != RW_DESTROYED,
 353             ("rw_wunlock() of destroyed rwlock @ %s:%d", file, line));
 354         __rw_assert(c, RA_WLOCKED, file, line);
 355         WITNESS_UNLOCK(&rw->lock_object, LOP_EXCLUSIVE, file, line);
 356         LOCK_LOG_LOCK("WUNLOCK", &rw->lock_object, 0, rw->rw_recurse, file,
 357             line);
 358         __rw_wunlock(rw, curthread, file, line);
 359         TD_LOCKS_DEC(curthread);
 360 }
 361
 362 /*
 363  * Determines whether a new reader can acquire a lock.  Succeeds if the
 364  * reader already owns a read lock and the lock is locked for read to
 365  * prevent deadlock from reader recursion.  Also succeeds if the lock
 366  * is unlocked and has no writer waiters or spinners.  Failing otherwise
 367  * prioritizes writers before readers.
 368  */
 369 #define RW_CAN_READ(_rw)                                                \
 370     ((curthread->td_rw_rlocks && (_rw) & RW_LOCK_READ) || ((_rw) &      \
 371     (RW_LOCK_READ | RW_LOCK_WRITE_WAITERS | RW_LOCK_WRITE_SPINNER)) ==  \
 372     RW_LOCK_READ)
 373
 374 void
 375 __rw_rlock(volatile uintptr_t *c, const char *file, int line)
 376 {
 377         struct rwlock *rw;
 378         struct turnstile *ts;
 379 #ifdef ADAPTIVE_RWLOCKS
 380         volatile struct thread *owner;
 381         int spintries = 0;
 382         int i;
 383 #endif
 384 #ifdef LOCK_PROFILING
 385         uint64_t waittime = 0;
 386         int contested = 0;
 387 #endif
 388         uintptr_t v;
 389 #if defined(ADAPTIVE_RWLOCKS) || defined(KDTRACE_HOOKS)
 390         struct lock_delay_arg lda;
 391 #endif
 392 #ifdef KDTRACE_HOOKS
 393         uintptr_t state;
 394         u_int sleep_cnt = 0;
 395         int64_t sleep_time = 0;
 396         int64_t all_time = 0;
 397 #endif
 398
 399         if (SCHEDULER_STOPPED())
 400                 return;
 401
 402 #if defined(ADAPTIVE_RWLOCKS)
 403         lock_delay_arg_init(&lda, &rw_delay);
 404 #elif defined(KDTRACE_HOOKS)
 405         lock_delay_arg_init(&lda, NULL);
 406 #endif
 407         rw = rwlock2rw(c);
 408
 409         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
 410             ("rw_rlock() by idle thread %p on rwlock %s @ %s:%d",
 411             curthread, rw->lock_object.lo_name, file, line));
 412         KASSERT(rw->rw_lock != RW_DESTROYED,
 413             ("rw_rlock() of destroyed rwlock @ %s:%d", file, line));
 414         KASSERT(rw_wowner(rw) != curthread,
 415             ("rw_rlock: wlock already held for %s @ %s:%d",
 416             rw->lock_object.lo_name, file, line));
 417         WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER, file, line, NULL);
 418
 419 #ifdef KDTRACE_HOOKS
 420         all_time -= lockstat_nsecs(&rw->lock_object);
 421 #endif
 422         v = RW_READ_VALUE(rw);
 423 #ifdef KDTRACE_HOOKS
 424         state = v;
 425 #endif
 426         for (;;) {
 427                 /*
 428                  * Handle the easy case.  If no other thread has a write
 429                  * lock, then try to bump up the count of read locks.  Note
 430                  * that we have to preserve the current state of the
 431                  * RW_LOCK_WRITE_WAITERS flag.  If we fail to acquire a
 432                  * read lock, then rw_lock must have changed, so restart
 433                  * the loop.  Note that this handles the case of a
 434                  * completely unlocked rwlock since such a lock is encoded
 435                  * as a read lock with no waiters.
 436                  */
 437                 if (RW_CAN_READ(v)) {
 438                         /*
 439                          * The RW_LOCK_READ_WAITERS flag should only be set
 440                          * if the lock has been unlocked and write waiters
 441                          * were present.
 442                          */
 443                         if (atomic_cmpset_acq_ptr(&rw->rw_lock, v,
 444                             v + RW_ONE_READER)) {
 445                                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 446                                         CTR4(KTR_LOCK,
 447                                             "%s: %p succeed %p -> %p", __func__,
 448                                             rw, (void *)v,
 449                                             (void *)(v + RW_ONE_READER));
 450                                 break;
 451                         }
 452                         v = RW_READ_VALUE(rw);
 453                         continue;
 454                 }
 455 #ifdef KDTRACE_HOOKS
 456                 lda.spin_cnt++;
 457 #endif
 458 #ifdef HWPMC_HOOKS
 459                 PMC_SOFT_CALL( , , lock, failed);
 460 #endif
 461                 lock_profile_obtain_lock_failed(&rw->lock_object,
 462                     &contested, &waittime);
 463
 464 #ifdef ADAPTIVE_RWLOCKS
 465                 /*
 466                  * If the owner is running on another CPU, spin until
 467                  * the owner stops running or the state of the lock
 468                  * changes.
 469                  */
 470                 if ((v & RW_LOCK_READ) == 0) {
 471                         owner = (struct thread *)RW_OWNER(v);
 472                         if (TD_IS_RUNNING(owner)) {
 473                                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 474                                         CTR3(KTR_LOCK,
 475                                             "%s: spinning on %p held by %p",
 476                                             __func__, rw, owner);
 477                                 KTR_STATE1(KTR_SCHED, "thread",
 478                                     sched_tdname(curthread), "spinning",
 479                                     "lockname:\"%s\"", rw->lock_object.lo_name);
 480                                 do {
 481                                         lock_delay(&lda);
 482                                         v = RW_READ_VALUE(rw);
 483                                         owner = lv_rw_wowner(v);
 484                                 } while (owner != NULL && TD_IS_RUNNING(owner));
 485                                 KTR_STATE0(KTR_SCHED, "thread",
 486                                     sched_tdname(curthread), "running");
 487                                 continue;
 488                         }
 489                 } else if (spintries < rowner_retries) {
 490                         spintries++;
 491                         KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
 492                             "spinning", "lockname:\"%s\"",
 493                             rw->lock_object.lo_name);
 494                         for (i = 0; i < rowner_loops; i++) {
 495                                 v = RW_READ_VALUE(rw);
 496                                 if ((v & RW_LOCK_READ) == 0 || RW_CAN_READ(v))
 497                                         break;
 498                                 cpu_spinwait();
 499                         }
 500                         v = RW_READ_VALUE(rw);
 501 #ifdef KDTRACE_HOOKS
 502                         lda.spin_cnt += rowner_loops - i;
 503 #endif
 504                         KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
 505                             "running");
 506                         if (i != rowner_loops)
 507                                 continue;
 508                 }
 509 #endif
 510
 511                 /*
 512                  * Okay, now it's the hard case.  Some other thread already
 513                  * has a write lock or there are write waiters present,
 514                  * acquire the turnstile lock so we can begin the process
 515                  * of blocking.
 516                  */
 517                 ts = turnstile_trywait(&rw->lock_object);
 518
 519                 /*
 520                  * The lock might have been released while we spun, so
 521                  * recheck its state and restart the loop if needed.
 522                  */
 523                 v = RW_READ_VALUE(rw);
 524                 if (RW_CAN_READ(v)) {
 525                         turnstile_cancel(ts);
 526                         continue;
 527                 }
 528
 529 #ifdef ADAPTIVE_RWLOCKS
 530                 /*
 531                  * The current lock owner might have started executing
 532                  * on another CPU (or the lock could have changed
 533                  * owners) while we were waiting on the turnstile
 534                  * chain lock.  If so, drop the turnstile lock and try
 535                  * again.
 536                  */
 537                 if ((v & RW_LOCK_READ) == 0) {
 538                         owner = (struct thread *)RW_OWNER(v);
 539                         if (TD_IS_RUNNING(owner)) {
 540                                 turnstile_cancel(ts);
 541                                 continue;
 542                         }
 543                 }
 544 #endif
 545
 546                 /*
 547                  * The lock is held in write mode or it already has waiters.
 548                  */
 549                 MPASS(!RW_CAN_READ(v));
 550
 551                 /*
 552                  * If the RW_LOCK_READ_WAITERS flag is already set, then
 553                  * we can go ahead and block.  If it is not set then try
 554                  * to set it.  If we fail to set it drop the turnstile
 555                  * lock and restart the loop.
 556                  */
 557                 if (!(v & RW_LOCK_READ_WAITERS)) {
 558                         if (!atomic_cmpset_ptr(&rw->rw_lock, v,
 559                             v | RW_LOCK_READ_WAITERS)) {
 560                                 turnstile_cancel(ts);
 561                                 v = RW_READ_VALUE(rw);
 562                                 continue;
 563                         }
 564                         if (LOCK_LOG_TEST(&rw->lock_object, 0))
 565                                 CTR2(KTR_LOCK, "%s: %p set read waiters flag",
 566                                     __func__, rw);
 567                 }
 568
 569                 /*
 570                  * We were unable to acquire the lock and the read waiters
 571                  * flag is set, so we must block on the turnstile.
 572                  */
 573                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 574                         CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__,
 575                             rw);
 576 #ifdef KDTRACE_HOOKS
 577                 sleep_time -= lockstat_nsecs(&rw->lock_object);
 578 #endif
 579                 turnstile_wait(ts, rw_owner(rw), TS_SHARED_QUEUE);
 580 #ifdef KDTRACE_HOOKS
 581                 sleep_time += lockstat_nsecs(&rw->lock_object);
 582                 sleep_cnt++;
 583 #endif
 584                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 585                         CTR2(KTR_LOCK, "%s: %p resuming from turnstile",
 586                             __func__, rw);
 587                 v = RW_READ_VALUE(rw);
 588         }
 589 #ifdef KDTRACE_HOOKS
 590         all_time += lockstat_nsecs(&rw->lock_object);
 591         if (sleep_time)
 592                 LOCKSTAT_RECORD4(rw__block, rw, sleep_time,
 593                     LOCKSTAT_READER, (state & RW_LOCK_READ) == 0,
 594                     (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
 595
 596         /* Record only the loops spinning and not sleeping. */
 597         if (lda.spin_cnt > sleep_cnt)
 598                 LOCKSTAT_RECORD4(rw__spin, rw, all_time - sleep_time,
 599                     LOCKSTAT_READER, (state & RW_LOCK_READ) == 0,
 600                     (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
 601 #endif
 602         /*
 603          * TODO: acquire "owner of record" here.  Here be turnstile dragons
 604          * however.  turnstiles don't like owners changing between calls to
 605          * turnstile_wait() currently.
 606          */
 607         LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire, rw, contested,
 608             waittime, file, line, LOCKSTAT_READER);
 609         LOCK_LOG_LOCK("RLOCK", &rw->lock_object, 0, 0, file, line);
 610         WITNESS_LOCK(&rw->lock_object, 0, file, line);
 611         TD_LOCKS_INC(curthread);
 612         curthread->td_rw_rlocks++;
 613 }
 614
 615 int
 616 __rw_try_rlock(volatile uintptr_t *c, const char *file, int line)
 617 {
 618         struct rwlock *rw;
 619         uintptr_t x;
 620
 621         if (SCHEDULER_STOPPED())
 622                 return (1);
 623
 624         rw = rwlock2rw(c);
 625
 626         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
 627             ("rw_try_rlock() by idle thread %p on rwlock %s @ %s:%d",
 628             curthread, rw->lock_object.lo_name, file, line));
 629
 630         for (;;) {
 631                 x = rw->rw_lock;
 632                 KASSERT(rw->rw_lock != RW_DESTROYED,
 633                     ("rw_try_rlock() of destroyed rwlock @ %s:%d", file, line));
 634                 if (!(x & RW_LOCK_READ))
 635                         break;
 636                 if (atomic_cmpset_acq_ptr(&rw->rw_lock, x, x + RW_ONE_READER)) {
 637                         LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 1, file,
 638                             line);
 639                         WITNESS_LOCK(&rw->lock_object, LOP_TRYLOCK, file, line);
 640                         LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire,
 641                             rw, 0, 0, file, line, LOCKSTAT_READER);
 642                         TD_LOCKS_INC(curthread);
 643                         curthread->td_rw_rlocks++;
 644                         return (1);
 645                 }
 646         }
 647
 648         LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 0, file, line);
 649         return (0);
 650 }
 651
 652 void
 653 _rw_runlock_cookie(volatile uintptr_t *c, const char *file, int line)
 654 {
 655         struct rwlock *rw;
 656         struct turnstile *ts;
 657         uintptr_t x, v, queue;
 658
 659         if (SCHEDULER_STOPPED())
 660                 return;
 661
 662         rw = rwlock2rw(c);
 663
 664         KASSERT(rw->rw_lock != RW_DESTROYED,
 665             ("rw_runlock() of destroyed rwlock @ %s:%d", file, line));
 666         __rw_assert(c, RA_RLOCKED, file, line);
 667         WITNESS_UNLOCK(&rw->lock_object, 0, file, line);
 668         LOCK_LOG_LOCK("RUNLOCK", &rw->lock_object, 0, 0, file, line);
 669
 670         /* TODO: drop "owner of record" here. */
 671         x = RW_READ_VALUE(rw);
 672         for (;;) {
 673                 /*
 674                  * See if there is more than one read lock held.  If so,
 675                  * just drop one and return.
 676                  */
 677                 if (RW_READERS(x) > 1) {
 678                         if (atomic_cmpset_rel_ptr(&rw->rw_lock, x,
 679                             x - RW_ONE_READER)) {
 680                                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 681                                         CTR4(KTR_LOCK,
 682                                             "%s: %p succeeded %p -> %p",
 683                                             __func__, rw, (void *)x,
 684                                             (void *)(x - RW_ONE_READER));
 685                                 break;
 686                         }
 687                         x = RW_READ_VALUE(rw);
 688                         continue;
 689                 }
 690                 /*
 691                  * If there aren't any waiters for a write lock, then try
 692                  * to drop it quickly.
 693                  */
 694                 if (!(x & RW_LOCK_WAITERS)) {
 695                         MPASS((x & ~RW_LOCK_WRITE_SPINNER) ==
 696                             RW_READERS_LOCK(1));
 697                         if (atomic_cmpset_rel_ptr(&rw->rw_lock, x,
 698                             RW_UNLOCKED)) {
 699                                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 700                                         CTR2(KTR_LOCK, "%s: %p last succeeded",
 701                                             __func__, rw);
 702                                 break;
 703                         }
 704                         x = RW_READ_VALUE(rw);
 705                         continue;
 706                 }
 707                 /*
 708                  * Ok, we know we have waiters and we think we are the
 709                  * last reader, so grab the turnstile lock.
 710                  */
 711                 turnstile_chain_lock(&rw->lock_object);
 712                 v = rw->rw_lock & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER);
 713                 MPASS(v & RW_LOCK_WAITERS);
 714
 715                 /*
 716                  * Try to drop our lock leaving the lock in a unlocked
 717                  * state.
 718                  *
 719                  * If you wanted to do explicit lock handoff you'd have to
 720                  * do it here.  You'd also want to use turnstile_signal()
 721                  * and you'd have to handle the race where a higher
 722                  * priority thread blocks on the write lock before the
 723                  * thread you wakeup actually runs and have the new thread
 724                  * "steal" the lock.  For now it's a lot simpler to just
 725                  * wakeup all of the waiters.
 726                  *
 727                  * As above, if we fail, then another thread might have
 728                  * acquired a read lock, so drop the turnstile lock and
 729                  * restart.
 730                  */
 731                 x = RW_UNLOCKED;
 732                 if (v & RW_LOCK_WRITE_WAITERS) {
 733                         queue = TS_EXCLUSIVE_QUEUE;
 734                         x |= (v & RW_LOCK_READ_WAITERS);
 735                 } else
 736                         queue = TS_SHARED_QUEUE;
 737                 if (!atomic_cmpset_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v,
 738                     x)) {
 739                         turnstile_chain_unlock(&rw->lock_object);
 740                         x = RW_READ_VALUE(rw);
 741                         continue;
 742                 }
 743                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 744                         CTR2(KTR_LOCK, "%s: %p last succeeded with waiters",
 745                             __func__, rw);
 746
 747                 /*
 748                  * Ok.  The lock is released and all that's left is to
 749                  * wake up the waiters.  Note that the lock might not be
 750                  * free anymore, but in that case the writers will just
 751                  * block again if they run before the new lock holder(s)
 752                  * release the lock.
 753                  */
 754                 ts = turnstile_lookup(&rw->lock_object);
 755                 MPASS(ts != NULL);
 756                 turnstile_broadcast(ts, queue);
 757                 turnstile_unpend(ts, TS_SHARED_LOCK);
 758                 turnstile_chain_unlock(&rw->lock_object);
 759                 break;
 760         }
 761         LOCKSTAT_PROFILE_RELEASE_RWLOCK(rw__release, rw, LOCKSTAT_READER);
 762         TD_LOCKS_DEC(curthread);
 763         curthread->td_rw_rlocks--;
 764 }
 765
 766 /*
 767  * This function is called when we are unable to obtain a write lock on the
 768  * first try.  This means that at least one other thread holds either a
 769  * read or write lock.
 770  */
 771 void
 772 __rw_wlock_hard(volatile uintptr_t *c, uintptr_t tid, const char *file,
 773     int line)
 774 {
 775         struct rwlock *rw;
 776         struct turnstile *ts;
 777 #ifdef ADAPTIVE_RWLOCKS
 778         volatile struct thread *owner;
 779         int spintries = 0;
 780         int i;
 781 #endif
 782         uintptr_t v, x;
 783 #ifdef LOCK_PROFILING
 784         uint64_t waittime = 0;
 785         int contested = 0;
 786 #endif
 787 #if defined(ADAPTIVE_RWLOCKS) || defined(KDTRACE_HOOKS)
 788         struct lock_delay_arg lda;
 789 #endif
 790 #ifdef KDTRACE_HOOKS
 791         uintptr_t state;
 792         u_int sleep_cnt = 0;
 793         int64_t sleep_time = 0;
 794         int64_t all_time = 0;
 795 #endif
 796
 797         if (SCHEDULER_STOPPED())
 798                 return;
 799
 800 #if defined(ADAPTIVE_RWLOCKS)
 801         lock_delay_arg_init(&lda, &rw_delay);
 802 #elif defined(KDTRACE_HOOKS)
 803         lock_delay_arg_init(&lda, NULL);
 804 #endif
 805         rw = rwlock2rw(c);
 806         v = RW_READ_VALUE(rw);
 807
 808         if (__predict_false(lv_rw_wowner(v) == (struct thread *)tid)) {
 809                 KASSERT(rw->lock_object.lo_flags & LO_RECURSABLE,
 810                     ("%s: recursing but non-recursive rw %s @ %s:%d\n",
 811                     __func__, rw->lock_object.lo_name, file, line));
 812                 rw->rw_recurse++;
 813                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 814                         CTR2(KTR_LOCK, "%s: %p recursing", __func__, rw);
 815                 return;
 816         }
 817
 818         if (LOCK_LOG_TEST(&rw->lock_object, 0))
 819                 CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__,
 820                     rw->lock_object.lo_name, (void *)rw->rw_lock, file, line);
 821
 822 #ifdef KDTRACE_HOOKS
 823         all_time -= lockstat_nsecs(&rw->lock_object);
 824         state = v;
 825 #endif
 826         for (;;) {
 827                 if (v == RW_UNLOCKED) {
 828                         if (_rw_write_lock(rw, tid))
 829                                 break;
 830                         v = RW_READ_VALUE(rw);
 831                         continue;
 832                 }
 833 #ifdef KDTRACE_HOOKS
 834                 lda.spin_cnt++;
 835 #endif
 836 #ifdef HWPMC_HOOKS
 837                 PMC_SOFT_CALL( , , lock, failed);
 838 #endif
 839                 lock_profile_obtain_lock_failed(&rw->lock_object,
 840                     &contested, &waittime);
 841 #ifdef ADAPTIVE_RWLOCKS
 842                 /*
 843                  * If the lock is write locked and the owner is
 844                  * running on another CPU, spin until the owner stops
 845                  * running or the state of the lock changes.
 846                  */
 847                 owner = lv_rw_wowner(v);
 848                 if (!(v & RW_LOCK_READ) && TD_IS_RUNNING(owner)) {
 849                         if (LOCK_LOG_TEST(&rw->lock_object, 0))
 850                                 CTR3(KTR_LOCK, "%s: spinning on %p held by %p",
 851                                     __func__, rw, owner);
 852                         KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
 853                             "spinning", "lockname:\"%s\"",
 854                             rw->lock_object.lo_name);
 855                         do {
 856                                 lock_delay(&lda);
 857                                 v = RW_READ_VALUE(rw);
 858                                 owner = lv_rw_wowner(v);
 859                         } while (owner != NULL && TD_IS_RUNNING(owner));
 860                         KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
 861                             "running");
 862                         continue;
 863                 }
 864                 if ((v & RW_LOCK_READ) && RW_READERS(v) &&
 865                     spintries < rowner_retries) {
 866                         if (!(v & RW_LOCK_WRITE_SPINNER)) {
 867                                 if (!atomic_cmpset_ptr(&rw->rw_lock, v,
 868                                     v | RW_LOCK_WRITE_SPINNER)) {
 869                                         v = RW_READ_VALUE(rw);
 870                                         continue;
 871                                 }
 872                         }
 873                         spintries++;
 874                         KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
 875                             "spinning", "lockname:\"%s\"",
 876                             rw->lock_object.lo_name);
 877                         for (i = 0; i < rowner_loops; i++) {
 878                                 if ((rw->rw_lock & RW_LOCK_WRITE_SPINNER) == 0)
 879                                         break;
 880                                 cpu_spinwait();
 881                         }
 882                         KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
 883                             "running");
 884                         v = RW_READ_VALUE(rw);
 885 #ifdef KDTRACE_HOOKS
 886                         lda.spin_cnt += rowner_loops - i;
 887 #endif
 888                         if (i != rowner_loops)
 889                                 continue;
 890                 }
 891 #endif
 892                 ts = turnstile_trywait(&rw->lock_object);
 893                 v = RW_READ_VALUE(rw);
 894
 895 #ifdef ADAPTIVE_RWLOCKS
 896                 /*
 897                  * The current lock owner might have started executing
 898                  * on another CPU (or the lock could have changed
 899                  * owners) while we were waiting on the turnstile
 900                  * chain lock.  If so, drop the turnstile lock and try
 901                  * again.
 902                  */
 903                 if (!(v & RW_LOCK_READ)) {
 904                         owner = (struct thread *)RW_OWNER(v);
 905                         if (TD_IS_RUNNING(owner)) {
 906                                 turnstile_cancel(ts);
 907                                 continue;
 908                         }
 909                 }
 910 #endif
 911                 /*
 912                  * Check for the waiters flags about this rwlock.
 913                  * If the lock was released, without maintain any pending
 914                  * waiters queue, simply try to acquire it.
 915                  * If a pending waiters queue is present, claim the lock
 916                  * ownership and maintain the pending queue.
 917                  */
 918                 x = v & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER);
 919                 if ((v & ~x) == RW_UNLOCKED) {
 920                         x &= ~RW_LOCK_WRITE_SPINNER;
 921                         if (atomic_cmpset_acq_ptr(&rw->rw_lock, v, tid | x)) {
 922                                 if (x)
 923                                         turnstile_claim(ts);
 924                                 else
 925                                         turnstile_cancel(ts);
 926                                 break;
 927                         }
 928                         turnstile_cancel(ts);
 929                         v = RW_READ_VALUE(rw);
 930                         continue;
 931                 }
 932                 /*
 933                  * If the RW_LOCK_WRITE_WAITERS flag isn't set, then try to
 934                  * set it.  If we fail to set it, then loop back and try
 935                  * again.
 936                  */
 937                 if (!(v & RW_LOCK_WRITE_WAITERS)) {
 938                         if (!atomic_cmpset_ptr(&rw->rw_lock, v,
 939                             v | RW_LOCK_WRITE_WAITERS)) {
 940                                 turnstile_cancel(ts);
 941                                 v = RW_READ_VALUE(rw);
 942                                 continue;
 943                         }
 944                         if (LOCK_LOG_TEST(&rw->lock_object, 0))
 945                                 CTR2(KTR_LOCK, "%s: %p set write waiters flag",
 946                                     __func__, rw);
 947                 }
 948                 /*
 949                  * We were unable to acquire the lock and the write waiters
 950                  * flag is set, so we must block on the turnstile.
 951                  */
 952                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 953                         CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__,
 954                             rw);
 955 #ifdef KDTRACE_HOOKS
 956                 sleep_time -= lockstat_nsecs(&rw->lock_object);
 957 #endif
 958                 turnstile_wait(ts, rw_owner(rw), TS_EXCLUSIVE_QUEUE);
 959 #ifdef KDTRACE_HOOKS
 960                 sleep_time += lockstat_nsecs(&rw->lock_object);
 961                 sleep_cnt++;
 962 #endif
 963                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 964                         CTR2(KTR_LOCK, "%s: %p resuming from turnstile",
 965                             __func__, rw);
 966 #ifdef ADAPTIVE_RWLOCKS
 967                 spintries = 0;
 968 #endif
 969                 v = RW_READ_VALUE(rw);
 970         }
 971 #ifdef KDTRACE_HOOKS
 972         all_time += lockstat_nsecs(&rw->lock_object);
 973         if (sleep_time)
 974                 LOCKSTAT_RECORD4(rw__block, rw, sleep_time,
 975                     LOCKSTAT_WRITER, (state & RW_LOCK_READ) == 0,
 976                     (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
 977
 978         /* Record only the loops spinning and not sleeping. */
 979         if (lda.spin_cnt > sleep_cnt)
 980                 LOCKSTAT_RECORD4(rw__spin, rw, all_time - sleep_time,
 981                     LOCKSTAT_WRITER, (state & RW_LOCK_READ) == 0,
 982                     (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
 983 #endif
 984         LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire, rw, contested,
 985             waittime, file, line, LOCKSTAT_WRITER);
 986 }
 987
 988 /*
 989  * This function is called if the first try at releasing a write lock failed.
 990  * This means that one of the 2 waiter bits must be set indicating that at
 991  * least one thread is waiting on this lock.
 992  */
 993 void
 994 __rw_wunlock_hard(volatile uintptr_t *c, uintptr_t tid, const char *file,
 995     int line)
 996 {
 997         struct rwlock *rw;
 998         struct turnstile *ts;
 999         uintptr_t v;
1000         int queue;
1001
1002         if (SCHEDULER_STOPPED())
1003                 return;
1004
1005         rw = rwlock2rw(c);
1006
1007         if (rw_wlocked(rw) && rw_recursed(rw)) {
1008                 rw->rw_recurse--;
1009                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
1010                         CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, rw);
1011                 return;
1012         }
1013
1014         KASSERT(rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS),
1015             ("%s: neither of the waiter flags are set", __func__));
1016
1017         if (LOCK_LOG_TEST(&rw->lock_object, 0))
1018                 CTR2(KTR_LOCK, "%s: %p contested", __func__, rw);
1019
1020         turnstile_chain_lock(&rw->lock_object);
1021         ts = turnstile_lookup(&rw->lock_object);
1022         MPASS(ts != NULL);
1023
1024         /*
1025          * Use the same algo as sx locks for now.  Prefer waking up shared
1026          * waiters if we have any over writers.  This is probably not ideal.
1027          *
1028          * 'v' is the value we are going to write back to rw_lock.  If we
1029          * have waiters on both queues, we need to preserve the state of
1030          * the waiter flag for the queue we don't wake up.  For now this is
1031          * hardcoded for the algorithm mentioned above.
1032          *
1033          * In the case of both readers and writers waiting we wakeup the
1034          * readers but leave the RW_LOCK_WRITE_WAITERS flag set.  If a
1035          * new writer comes in before a reader it will claim the lock up
1036          * above.  There is probably a potential priority inversion in
1037          * there that could be worked around either by waking both queues
1038          * of waiters or doing some complicated lock handoff gymnastics.
1039          */
1040         v = RW_UNLOCKED;
1041         if (rw->rw_lock & RW_LOCK_WRITE_WAITERS) {
1042                 queue = TS_EXCLUSIVE_QUEUE;
1043                 v |= (rw->rw_lock & RW_LOCK_READ_WAITERS);
1044         } else
1045                 queue = TS_SHARED_QUEUE;
1046
1047         /* Wake up all waiters for the specific queue. */
1048         if (LOCK_LOG_TEST(&rw->lock_object, 0))
1049                 CTR3(KTR_LOCK, "%s: %p waking up %s waiters", __func__, rw,
1050                     queue == TS_SHARED_QUEUE ? "read" : "write");
1051         turnstile_broadcast(ts, queue);
1052         atomic_store_rel_ptr(&rw->rw_lock, v);
1053         turnstile_unpend(ts, TS_EXCLUSIVE_LOCK);
1054         turnstile_chain_unlock(&rw->lock_object);
1055 }
1056
1057 /*
1058  * Attempt to do a non-blocking upgrade from a read lock to a write
1059  * lock.  This will only succeed if this thread holds a single read
1060  * lock.  Returns true if the upgrade succeeded and false otherwise.
1061  */
1062 int
1063 __rw_try_upgrade(volatile uintptr_t *c, const char *file, int line)
1064 {
1065         struct rwlock *rw;
1066         uintptr_t v, x, tid;
1067         struct turnstile *ts;
1068         int success;
1069
1070         if (SCHEDULER_STOPPED())
1071                 return (1);
1072
1073         rw = rwlock2rw(c);
1074
1075         KASSERT(rw->rw_lock != RW_DESTROYED,
1076             ("rw_try_upgrade() of destroyed rwlock @ %s:%d", file, line));
1077         __rw_assert(c, RA_RLOCKED, file, line);
1078
1079         /*
1080          * Attempt to switch from one reader to a writer.  If there
1081          * are any write waiters, then we will have to lock the
1082          * turnstile first to prevent races with another writer
1083          * calling turnstile_wait() before we have claimed this
1084          * turnstile.  So, do the simple case of no waiters first.
1085          */
1086         tid = (uintptr_t)curthread;
1087         success = 0;
1088         for (;;) {
1089                 v = rw->rw_lock;
1090                 if (RW_READERS(v) > 1)
1091                         break;
1092                 if (!(v & RW_LOCK_WAITERS)) {
1093                         success = atomic_cmpset_ptr(&rw->rw_lock, v, tid);
1094                         if (!success)
1095                                 continue;
1096                         break;
1097                 }
1098
1099                 /*
1100                  * Ok, we think we have waiters, so lock the turnstile.
1101                  */
1102                 ts = turnstile_trywait(&rw->lock_object);
1103                 v = rw->rw_lock;
1104                 if (RW_READERS(v) > 1) {
1105                         turnstile_cancel(ts);
1106                         break;
1107                 }
1108                 /*
1109                  * Try to switch from one reader to a writer again.  This time
1110                  * we honor the current state of the waiters flags.
1111                  * If we obtain the lock with the flags set, then claim
1112                  * ownership of the turnstile.
1113                  */
1114                 x = rw->rw_lock & RW_LOCK_WAITERS;
1115                 success = atomic_cmpset_ptr(&rw->rw_lock, v, tid | x);
1116                 if (success) {
1117                         if (x)
1118                                 turnstile_claim(ts);
1119                         else
1120                                 turnstile_cancel(ts);
1121                         break;
1122                 }
1123                 turnstile_cancel(ts);
1124         }
1125         LOCK_LOG_TRY("WUPGRADE", &rw->lock_object, 0, success, file, line);
1126         if (success) {
1127                 curthread->td_rw_rlocks--;
1128                 WITNESS_UPGRADE(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
1129                     file, line);
1130                 LOCKSTAT_RECORD0(rw__upgrade, rw);
1131         }
1132         return (success);
1133 }
1134
1135 /*
1136  * Downgrade a write lock into a single read lock.
1137  */
1138 void
1139 __rw_downgrade(volatile uintptr_t *c, const char *file, int line)
1140 {
1141         struct rwlock *rw;
1142         struct turnstile *ts;
1143         uintptr_t tid, v;
1144         int rwait, wwait;
1145
1146         if (SCHEDULER_STOPPED())
1147                 return;
1148
1149         rw = rwlock2rw(c);
1150
1151         KASSERT(rw->rw_lock != RW_DESTROYED,
1152             ("rw_downgrade() of destroyed rwlock @ %s:%d", file, line));
1153         __rw_assert(c, RA_WLOCKED | RA_NOTRECURSED, file, line);
1154 #ifndef INVARIANTS
1155         if (rw_recursed(rw))
1156                 panic("downgrade of a recursed lock");
1157 #endif
1158
1159         WITNESS_DOWNGRADE(&rw->lock_object, 0, file, line);
1160
1161         /*
1162          * Convert from a writer to a single reader.  First we handle
1163          * the easy case with no waiters.  If there are any waiters, we
1164          * lock the turnstile and "disown" the lock.
1165          */
1166         tid = (uintptr_t)curthread;
1167         if (atomic_cmpset_rel_ptr(&rw->rw_lock, tid, RW_READERS_LOCK(1)))
1168                 goto out;
1169
1170         /*
1171          * Ok, we think we have waiters, so lock the turnstile so we can
1172          * read the waiter flags without any races.
1173          */
1174         turnstile_chain_lock(&rw->lock_object);
1175         v = rw->rw_lock & RW_LOCK_WAITERS;
1176         rwait = v & RW_LOCK_READ_WAITERS;
1177         wwait = v & RW_LOCK_WRITE_WAITERS;
1178         MPASS(rwait | wwait);
1179
1180         /*
1181          * Downgrade from a write lock while preserving waiters flag
1182          * and give up ownership of the turnstile.
1183          */
1184         ts = turnstile_lookup(&rw->lock_object);
1185         MPASS(ts != NULL);
1186         if (!wwait)
1187                 v &= ~RW_LOCK_READ_WAITERS;
1188         atomic_store_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v);
1189         /*
1190          * Wake other readers if there are no writers pending.  Otherwise they
1191          * won't be able to acquire the lock anyway.
1192          */
1193         if (rwait && !wwait) {
1194                 turnstile_broadcast(ts, TS_SHARED_QUEUE);
1195                 turnstile_unpend(ts, TS_EXCLUSIVE_LOCK);
1196         } else
1197                 turnstile_disown(ts);
1198         turnstile_chain_unlock(&rw->lock_object);
1199 out:
1200         curthread->td_rw_rlocks++;
1201         LOCK_LOG_LOCK("WDOWNGRADE", &rw->lock_object, 0, 0, file, line);
1202         LOCKSTAT_RECORD0(rw__downgrade, rw);
1203 }
1204
1205 #ifdef INVARIANT_SUPPORT
1206 #ifndef INVARIANTS
1207 #undef __rw_assert
1208 #endif
1209
1210 /*
1211  * In the non-WITNESS case, rw_assert() can only detect that at least
1212  * *some* thread owns an rlock, but it cannot guarantee that *this*
1213  * thread owns an rlock.
1214  */
1215 void
1216 __rw_assert(const volatile uintptr_t *c, int what, const char *file, int line)
1217 {
1218         const struct rwlock *rw;
1219
1220         if (panicstr != NULL)
1221                 return;
1222
1223         rw = rwlock2rw(c);
1224
1225         switch (what) {
1226         case RA_LOCKED:
1227         case RA_LOCKED | RA_RECURSED:
1228         case RA_LOCKED | RA_NOTRECURSED:
1229         case RA_RLOCKED:
1230         case RA_RLOCKED | RA_RECURSED:
1231         case RA_RLOCKED | RA_NOTRECURSED:
1232 #ifdef WITNESS
1233                 witness_assert(&rw->lock_object, what, file, line);
1234 #else
1235                 /*
1236                  * If some other thread has a write lock or we have one
1237                  * and are asserting a read lock, fail.  Also, if no one
1238                  * has a lock at all, fail.
1239                  */
1240                 if (rw->rw_lock == RW_UNLOCKED ||
1241                     (!(rw->rw_lock & RW_LOCK_READ) && (what & RA_RLOCKED ||
1242                     rw_wowner(rw) != curthread)))
1243                         panic("Lock %s not %slocked @ %s:%d\n",
1244                             rw->lock_object.lo_name, (what & RA_RLOCKED) ?
1245                             "read " : "", file, line);
1246
1247                 if (!(rw->rw_lock & RW_LOCK_READ) && !(what & RA_RLOCKED)) {
1248                         if (rw_recursed(rw)) {
1249                                 if (what & RA_NOTRECURSED)
1250                                         panic("Lock %s recursed @ %s:%d\n",
1251                                             rw->lock_object.lo_name, file,
1252                                             line);
1253                         } else if (what & RA_RECURSED)
1254                                 panic("Lock %s not recursed @ %s:%d\n",
1255                                     rw->lock_object.lo_name, file, line);
1256                 }
1257 #endif
1258                 break;
1259         case RA_WLOCKED:
1260         case RA_WLOCKED | RA_RECURSED:
1261         case RA_WLOCKED | RA_NOTRECURSED:
1262                 if (rw_wowner(rw) != curthread)
1263                         panic("Lock %s not exclusively locked @ %s:%d\n",
1264                             rw->lock_object.lo_name, file, line);
1265                 if (rw_recursed(rw)) {
1266                         if (what & RA_NOTRECURSED)
1267                                 panic("Lock %s recursed @ %s:%d\n",
1268                                     rw->lock_object.lo_name, file, line);
1269                 } else if (what & RA_RECURSED)
1270                         panic("Lock %s not recursed @ %s:%d\n",
1271                             rw->lock_object.lo_name, file, line);
1272                 break;
1273         case RA_UNLOCKED:
1274 #ifdef WITNESS
1275                 witness_assert(&rw->lock_object, what, file, line);
1276 #else
1277                 /*
1278                  * If we hold a write lock fail.  We can't reliably check
1279                  * to see if we hold a read lock or not.
1280                  */
1281                 if (rw_wowner(rw) == curthread)
1282                         panic("Lock %s exclusively locked @ %s:%d\n",
1283                             rw->lock_object.lo_name, file, line);
1284 #endif
1285                 break;
1286         default:
1287                 panic("Unknown rw lock assertion: %d @ %s:%d", what, file,
1288                     line);
1289         }
1290 }
1291 #endif /* INVARIANT_SUPPORT */
1292
1293 #ifdef DDB
1294 void
1295 db_show_rwlock(const struct lock_object *lock)
1296 {
1297         const struct rwlock *rw;
1298         struct thread *td;
1299
1300         rw = (const struct rwlock *)lock;
1301
1302         db_printf(" state: ");
1303         if (rw->rw_lock == RW_UNLOCKED)
1304                 db_printf("UNLOCKED\n");
1305         else if (rw->rw_lock == RW_DESTROYED) {
1306                 db_printf("DESTROYED\n");
1307                 return;
1308         } else if (rw->rw_lock & RW_LOCK_READ)
1309                 db_printf("RLOCK: %ju locks\n",
1310                     (uintmax_t)(RW_READERS(rw->rw_lock)));
1311         else {
1312                 td = rw_wowner(rw);
1313                 db_printf("WLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1314                     td->td_tid, td->td_proc->p_pid, td->td_name);
1315                 if (rw_recursed(rw))
1316                         db_printf(" recursed: %u\n", rw->rw_recurse);
1317         }
1318         db_printf(" waiters: ");
1319         switch (rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS)) {
1320         case RW_LOCK_READ_WAITERS:
1321                 db_printf("readers\n");
1322                 break;
1323         case RW_LOCK_WRITE_WAITERS:
1324                 db_printf("writers\n");
1325                 break;
1326         case RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS:
1327                 db_printf("readers and writers\n");
1328                 break;
1329         default:
1330                 db_printf("none\n");
1331                 break;
1332         }
1333 }
1334
1335 #endif