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
 105 SYSCTL_INT(_debug_rwlock, OID_AUTO, delay_base, CTLFLAG_RW, &rw_delay.base,
 106     0, "");
 107 SYSCTL_INT(_debug_rwlock, OID_AUTO, delay_max, CTLFLAG_RW, &rw_delay.max,
 108     0, "");
 109
 110 LOCK_DELAY_SYSINIT_DEFAULT(rw_delay);
 111 #endif
 112
 113 /*
 114  * Return a pointer to the owning thread if the lock is write-locked or
 115  * NULL if the lock is unlocked or read-locked.
 116  */
 117
 118 #define lv_rw_wowner(v)                                                 \
 119         ((v) & RW_LOCK_READ ? NULL :                                    \
 120          (struct thread *)RW_OWNER((v)))
 121
 122 #define rw_wowner(rw)   lv_rw_wowner(RW_READ_VALUE(rw))
 123
 124 /*
 125  * Returns if a write owner is recursed.  Write ownership is not assured
 126  * here and should be previously checked.
 127  */
 128 #define rw_recursed(rw)         ((rw)->rw_recurse != 0)
 129
 130 /*
 131  * Return true if curthread helds the lock.
 132  */
 133 #define rw_wlocked(rw)          (rw_wowner((rw)) == curthread)
 134
 135 /*
 136  * Return a pointer to the owning thread for this lock who should receive
 137  * any priority lent by threads that block on this lock.  Currently this
 138  * is identical to rw_wowner().
 139  */
 140 #define rw_owner(rw)            rw_wowner(rw)
 141
 142 #ifndef INVARIANTS
 143 #define __rw_assert(c, what, file, line)
 144 #endif
 145
 146 void
 147 assert_rw(const struct lock_object *lock, int what)
 148 {
 149
 150         rw_assert((const struct rwlock *)lock, what);
 151 }
 152
 153 void
 154 lock_rw(struct lock_object *lock, uintptr_t how)
 155 {
 156         struct rwlock *rw;
 157
 158         rw = (struct rwlock *)lock;
 159         if (how)
 160                 rw_rlock(rw);
 161         else
 162                 rw_wlock(rw);
 163 }
 164
 165 uintptr_t
 166 unlock_rw(struct lock_object *lock)
 167 {
 168         struct rwlock *rw;
 169
 170         rw = (struct rwlock *)lock;
 171         rw_assert(rw, RA_LOCKED | LA_NOTRECURSED);
 172         if (rw->rw_lock & RW_LOCK_READ) {
 173                 rw_runlock(rw);
 174                 return (1);
 175         } else {
 176                 rw_wunlock(rw);
 177                 return (0);
 178         }
 179 }
 180
 181 #ifdef KDTRACE_HOOKS
 182 int
 183 owner_rw(const struct lock_object *lock, struct thread **owner)
 184 {
 185         const struct rwlock *rw = (const struct rwlock *)lock;
 186         uintptr_t x = rw->rw_lock;
 187
 188         *owner = rw_wowner(rw);
 189         return ((x & RW_LOCK_READ) != 0 ?  (RW_READERS(x) != 0) :
 190             (*owner != NULL));
 191 }
 192 #endif
 193
 194 void
 195 _rw_init_flags(volatile uintptr_t *c, const char *name, int opts)
 196 {
 197         struct rwlock *rw;
 198         int flags;
 199
 200         rw = rwlock2rw(c);
 201
 202         MPASS((opts & ~(RW_DUPOK | RW_NOPROFILE | RW_NOWITNESS | RW_QUIET |
 203             RW_RECURSE | RW_NEW)) == 0);
 204         ASSERT_ATOMIC_LOAD_PTR(rw->rw_lock,
 205             ("%s: rw_lock not aligned for %s: %p", __func__, name,
 206             &rw->rw_lock));
 207
 208         flags = LO_UPGRADABLE;
 209         if (opts & RW_DUPOK)
 210                 flags |= LO_DUPOK;
 211         if (opts & RW_NOPROFILE)
 212                 flags |= LO_NOPROFILE;
 213         if (!(opts & RW_NOWITNESS))
 214                 flags |= LO_WITNESS;
 215         if (opts & RW_RECURSE)
 216                 flags |= LO_RECURSABLE;
 217         if (opts & RW_QUIET)
 218                 flags |= LO_QUIET;
 219         if (opts & RW_NEW)
 220                 flags |= LO_NEW;
 221
 222         lock_init(&rw->lock_object, &lock_class_rw, name, NULL, flags);
 223         rw->rw_lock = RW_UNLOCKED;
 224         rw->rw_recurse = 0;
 225 }
 226
 227 void
 228 _rw_destroy(volatile uintptr_t *c)
 229 {
 230         struct rwlock *rw;
 231
 232         rw = rwlock2rw(c);
 233
 234         KASSERT(rw->rw_lock == RW_UNLOCKED, ("rw lock %p not unlocked", rw));
 235         KASSERT(rw->rw_recurse == 0, ("rw lock %p still recursed", rw));
 236         rw->rw_lock = RW_DESTROYED;
 237         lock_destroy(&rw->lock_object);
 238 }
 239
 240 void
 241 rw_sysinit(void *arg)
 242 {
 243         struct rw_args *args = arg;
 244
 245         rw_init((struct rwlock *)args->ra_rw, args->ra_desc);
 246 }
 247
 248 void
 249 rw_sysinit_flags(void *arg)
 250 {
 251         struct rw_args_flags *args = arg;
 252
 253         rw_init_flags((struct rwlock *)args->ra_rw, args->ra_desc,
 254             args->ra_flags);
 255 }
 256
 257 int
 258 _rw_wowned(const volatile uintptr_t *c)
 259 {
 260
 261         return (rw_wowner(rwlock2rw(c)) == curthread);
 262 }
 263
 264 void
 265 _rw_wlock_cookie(volatile uintptr_t *c, const char *file, int line)
 266 {
 267         struct rwlock *rw;
 268         uintptr_t tid, v;
 269
 270         rw = rwlock2rw(c);
 271
 272         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
 273             ("rw_wlock() by idle thread %p on rwlock %s @ %s:%d",
 274             curthread, rw->lock_object.lo_name, file, line));
 275         KASSERT(rw->rw_lock != RW_DESTROYED,
 276             ("rw_wlock() of destroyed rwlock @ %s:%d", file, line));
 277         WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE, file,
 278             line, NULL);
 279         tid = (uintptr_t)curthread;
 280         v = RW_UNLOCKED;
 281         if (!_rw_write_lock_fetch(rw, &v, tid))
 282                 _rw_wlock_hard(rw, v, tid, file, line);
 283         else
 284                 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire, rw,
 285                     0, 0, file, line, LOCKSTAT_WRITER);
 286
 287         LOCK_LOG_LOCK("WLOCK", &rw->lock_object, 0, rw->rw_recurse, file, line);
 288         WITNESS_LOCK(&rw->lock_object, LOP_EXCLUSIVE, file, line);
 289         TD_LOCKS_INC(curthread);
 290 }
 291
 292 int
 293 __rw_try_wlock(volatile uintptr_t *c, const char *file, int line)
 294 {
 295         struct rwlock *rw;
 296         int rval;
 297
 298         if (SCHEDULER_STOPPED())
 299                 return (1);
 300
 301         rw = rwlock2rw(c);
 302
 303         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
 304             ("rw_try_wlock() by idle thread %p on rwlock %s @ %s:%d",
 305             curthread, rw->lock_object.lo_name, file, line));
 306         KASSERT(rw->rw_lock != RW_DESTROYED,
 307             ("rw_try_wlock() of destroyed rwlock @ %s:%d", file, line));
 308
 309         if (rw_wlocked(rw) &&
 310             (rw->lock_object.lo_flags & LO_RECURSABLE) != 0) {
 311                 rw->rw_recurse++;
 312                 atomic_set_ptr(&rw->rw_lock, RW_LOCK_WRITER_RECURSED);
 313                 rval = 1;
 314         } else
 315                 rval = atomic_cmpset_acq_ptr(&rw->rw_lock, RW_UNLOCKED,
 316                     (uintptr_t)curthread);
 317
 318         LOCK_LOG_TRY("WLOCK", &rw->lock_object, 0, rval, file, line);
 319         if (rval) {
 320                 WITNESS_LOCK(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
 321                     file, line);
 322                 if (!rw_recursed(rw))
 323                         LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire,
 324                             rw, 0, 0, file, line, LOCKSTAT_WRITER);
 325                 TD_LOCKS_INC(curthread);
 326         }
 327         return (rval);
 328 }
 329
 330 void
 331 _rw_wunlock_cookie(volatile uintptr_t *c, const char *file, int line)
 332 {
 333         struct rwlock *rw;
 334
 335         rw = rwlock2rw(c);
 336
 337         KASSERT(rw->rw_lock != RW_DESTROYED,
 338             ("rw_wunlock() of destroyed rwlock @ %s:%d", file, line));
 339         __rw_assert(c, RA_WLOCKED, file, line);
 340         WITNESS_UNLOCK(&rw->lock_object, LOP_EXCLUSIVE, file, line);
 341         LOCK_LOG_LOCK("WUNLOCK", &rw->lock_object, 0, rw->rw_recurse, file,
 342             line);
 343
 344 #ifdef LOCK_PROFILING
 345         _rw_wunlock_hard(rw, (uintptr_t)curthread, file, line);
 346 #else
 347         __rw_wunlock(rw, curthread, file, line);
 348 #endif
 349
 350         TD_LOCKS_DEC(curthread);
 351 }
 352
 353 /*
 354  * Determines whether a new reader can acquire a lock.  Succeeds if the
 355  * reader already owns a read lock and the lock is locked for read to
 356  * prevent deadlock from reader recursion.  Also succeeds if the lock
 357  * is unlocked and has no writer waiters or spinners.  Failing otherwise
 358  * prioritizes writers before readers.
 359  */
 360 #define RW_CAN_READ(td, _rw)                                            \
 361     (((td)->td_rw_rlocks && (_rw) & RW_LOCK_READ) || ((_rw) &   \
 362     (RW_LOCK_READ | RW_LOCK_WRITE_WAITERS | RW_LOCK_WRITE_SPINNER)) ==  \
 363     RW_LOCK_READ)
 364
 365 static bool __always_inline
 366 __rw_rlock_try(struct rwlock *rw, struct thread *td, uintptr_t *vp,
 367     const char *file, int line)
 368 {
 369
 370         /*
 371          * Handle the easy case.  If no other thread has a write
 372          * lock, then try to bump up the count of read locks.  Note
 373          * that we have to preserve the current state of the
 374          * RW_LOCK_WRITE_WAITERS flag.  If we fail to acquire a
 375          * read lock, then rw_lock must have changed, so restart
 376          * the loop.  Note that this handles the case of a
 377          * completely unlocked rwlock since such a lock is encoded
 378          * as a read lock with no waiters.
 379          */
 380         while (RW_CAN_READ(td, *vp)) {
 381                 if (atomic_fcmpset_acq_ptr(&rw->rw_lock, vp,
 382                         *vp + RW_ONE_READER)) {
 383                         if (LOCK_LOG_TEST(&rw->lock_object, 0))
 384                                 CTR4(KTR_LOCK,
 385                                     "%s: %p succeed %p -> %p", __func__,
 386                                     rw, (void *)*vp,
 387                                     (void *)(*vp + RW_ONE_READER));
 388                         td->td_rw_rlocks++;
 389                         return (true);
 390                 }
 391         }
 392         return (false);
 393 }
 394
 395 static void __noinline
 396 __rw_rlock_hard(volatile uintptr_t *c, struct thread *td, uintptr_t v,
 397     const char *file, int line)
 398 {
 399         struct rwlock *rw;
 400         struct turnstile *ts;
 401 #ifdef ADAPTIVE_RWLOCKS
 402         volatile struct thread *owner;
 403         int spintries = 0;
 404         int i;
 405 #endif
 406 #ifdef LOCK_PROFILING
 407         uint64_t waittime = 0;
 408         int contested = 0;
 409 #endif
 410 #if defined(ADAPTIVE_RWLOCKS) || defined(KDTRACE_HOOKS)
 411         struct lock_delay_arg lda;
 412 #endif
 413 #ifdef KDTRACE_HOOKS
 414         uintptr_t state;
 415         u_int sleep_cnt = 0;
 416         int64_t sleep_time = 0;
 417         int64_t all_time = 0;
 418 #endif
 419
 420         if (SCHEDULER_STOPPED())
 421                 return;
 422
 423 #if defined(ADAPTIVE_RWLOCKS)
 424         lock_delay_arg_init(&lda, &rw_delay);
 425 #elif defined(KDTRACE_HOOKS)
 426         lock_delay_arg_init(&lda, NULL);
 427 #endif
 428         rw = rwlock2rw(c);
 429
 430 #ifdef KDTRACE_HOOKS
 431         all_time -= lockstat_nsecs(&rw->lock_object);
 432 #endif
 433 #ifdef KDTRACE_HOOKS
 434         state = v;
 435 #endif
 436         for (;;) {
 437                 if (__rw_rlock_try(rw, td, &v, file, line))
 438                         break;
 439 #ifdef KDTRACE_HOOKS
 440                 lda.spin_cnt++;
 441 #endif
 442 #ifdef HWPMC_HOOKS
 443                 PMC_SOFT_CALL( , , lock, failed);
 444 #endif
 445                 lock_profile_obtain_lock_failed(&rw->lock_object,
 446                     &contested, &waittime);
 447
 448 #ifdef ADAPTIVE_RWLOCKS
 449                 /*
 450                  * If the owner is running on another CPU, spin until
 451                  * the owner stops running or the state of the lock
 452                  * changes.
 453                  */
 454                 if ((v & RW_LOCK_READ) == 0) {
 455                         owner = (struct thread *)RW_OWNER(v);
 456                         if (TD_IS_RUNNING(owner)) {
 457                                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 458                                         CTR3(KTR_LOCK,
 459                                             "%s: spinning on %p held by %p",
 460                                             __func__, rw, owner);
 461                                 KTR_STATE1(KTR_SCHED, "thread",
 462                                     sched_tdname(curthread), "spinning",
 463                                     "lockname:\"%s\"", rw->lock_object.lo_name);
 464                                 do {
 465                                         lock_delay(&lda);
 466                                         v = RW_READ_VALUE(rw);
 467                                         owner = lv_rw_wowner(v);
 468                                 } while (owner != NULL && TD_IS_RUNNING(owner));
 469                                 KTR_STATE0(KTR_SCHED, "thread",
 470                                     sched_tdname(curthread), "running");
 471                                 continue;
 472                         }
 473                 } else if (spintries < rowner_retries) {
 474                         spintries++;
 475                         KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
 476                             "spinning", "lockname:\"%s\"",
 477                             rw->lock_object.lo_name);
 478                         for (i = 0; i < rowner_loops; i++) {
 479                                 v = RW_READ_VALUE(rw);
 480                                 if ((v & RW_LOCK_READ) == 0 || RW_CAN_READ(td, v))
 481                                         break;
 482                                 cpu_spinwait();
 483                         }
 484                         v = RW_READ_VALUE(rw);
 485 #ifdef KDTRACE_HOOKS
 486                         lda.spin_cnt += rowner_loops - i;
 487 #endif
 488                         KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
 489                             "running");
 490                         if (i != rowner_loops)
 491                                 continue;
 492                 }
 493 #endif
 494
 495                 /*
 496                  * Okay, now it's the hard case.  Some other thread already
 497                  * has a write lock or there are write waiters present,
 498                  * acquire the turnstile lock so we can begin the process
 499                  * of blocking.
 500                  */
 501                 ts = turnstile_trywait(&rw->lock_object);
 502
 503                 /*
 504                  * The lock might have been released while we spun, so
 505                  * recheck its state and restart the loop if needed.
 506                  */
 507                 v = RW_READ_VALUE(rw);
 508                 if (RW_CAN_READ(td, v)) {
 509                         turnstile_cancel(ts);
 510                         continue;
 511                 }
 512
 513 #ifdef ADAPTIVE_RWLOCKS
 514                 /*
 515                  * The current lock owner might have started executing
 516                  * on another CPU (or the lock could have changed
 517                  * owners) while we were waiting on the turnstile
 518                  * chain lock.  If so, drop the turnstile lock and try
 519                  * again.
 520                  */
 521                 if ((v & RW_LOCK_READ) == 0) {
 522                         owner = (struct thread *)RW_OWNER(v);
 523                         if (TD_IS_RUNNING(owner)) {
 524                                 turnstile_cancel(ts);
 525                                 continue;
 526                         }
 527                 }
 528 #endif
 529
 530                 /*
 531                  * The lock is held in write mode or it already has waiters.
 532                  */
 533                 MPASS(!RW_CAN_READ(td, v));
 534
 535                 /*
 536                  * If the RW_LOCK_READ_WAITERS flag is already set, then
 537                  * we can go ahead and block.  If it is not set then try
 538                  * to set it.  If we fail to set it drop the turnstile
 539                  * lock and restart the loop.
 540                  */
 541                 if (!(v & RW_LOCK_READ_WAITERS)) {
 542                         if (!atomic_cmpset_ptr(&rw->rw_lock, v,
 543                             v | RW_LOCK_READ_WAITERS)) {
 544                                 turnstile_cancel(ts);
 545                                 v = RW_READ_VALUE(rw);
 546                                 continue;
 547                         }
 548                         if (LOCK_LOG_TEST(&rw->lock_object, 0))
 549                                 CTR2(KTR_LOCK, "%s: %p set read waiters flag",
 550                                     __func__, rw);
 551                 }
 552
 553                 /*
 554                  * We were unable to acquire the lock and the read waiters
 555                  * flag is set, so we must block on the turnstile.
 556                  */
 557                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 558                         CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__,
 559                             rw);
 560 #ifdef KDTRACE_HOOKS
 561                 sleep_time -= lockstat_nsecs(&rw->lock_object);
 562 #endif
 563                 turnstile_wait(ts, rw_owner(rw), TS_SHARED_QUEUE);
 564 #ifdef KDTRACE_HOOKS
 565                 sleep_time += lockstat_nsecs(&rw->lock_object);
 566                 sleep_cnt++;
 567 #endif
 568                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 569                         CTR2(KTR_LOCK, "%s: %p resuming from turnstile",
 570                             __func__, rw);
 571                 v = RW_READ_VALUE(rw);
 572         }
 573 #ifdef KDTRACE_HOOKS
 574         all_time += lockstat_nsecs(&rw->lock_object);
 575         if (sleep_time)
 576                 LOCKSTAT_RECORD4(rw__block, rw, sleep_time,
 577                     LOCKSTAT_READER, (state & RW_LOCK_READ) == 0,
 578                     (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
 579
 580         /* Record only the loops spinning and not sleeping. */
 581         if (lda.spin_cnt > sleep_cnt)
 582                 LOCKSTAT_RECORD4(rw__spin, rw, all_time - sleep_time,
 583                     LOCKSTAT_READER, (state & RW_LOCK_READ) == 0,
 584                     (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
 585 #endif
 586         /*
 587          * TODO: acquire "owner of record" here.  Here be turnstile dragons
 588          * however.  turnstiles don't like owners changing between calls to
 589          * turnstile_wait() currently.
 590          */
 591         LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire, rw, contested,
 592             waittime, file, line, LOCKSTAT_READER);
 593 }
 594
 595 void
 596 __rw_rlock(volatile uintptr_t *c, const char *file, int line)
 597 {
 598         struct rwlock *rw;
 599         struct thread *td;
 600         uintptr_t v;
 601
 602         td = curthread;
 603         rw = rwlock2rw(c);
 604
 605         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(td),
 606             ("rw_rlock() by idle thread %p on rwlock %s @ %s:%d",
 607             td, rw->lock_object.lo_name, file, line));
 608         KASSERT(rw->rw_lock != RW_DESTROYED,
 609             ("rw_rlock() of destroyed rwlock @ %s:%d", file, line));
 610         KASSERT(rw_wowner(rw) != td,
 611             ("rw_rlock: wlock already held for %s @ %s:%d",
 612             rw->lock_object.lo_name, file, line));
 613         WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER, file, line, NULL);
 614
 615         v = RW_READ_VALUE(rw);
 616         if (__predict_false(LOCKSTAT_OOL_PROFILE_ENABLED(rw__acquire) ||
 617             !__rw_rlock_try(rw, td, &v, file, line)))
 618                 __rw_rlock_hard(c, td, v, file, line);
 619
 620         LOCK_LOG_LOCK("RLOCK", &rw->lock_object, 0, 0, file, line);
 621         WITNESS_LOCK(&rw->lock_object, 0, file, line);
 622         TD_LOCKS_INC(curthread);
 623 }
 624
 625 int
 626 __rw_try_rlock(volatile uintptr_t *c, const char *file, int line)
 627 {
 628         struct rwlock *rw;
 629         uintptr_t x;
 630
 631         if (SCHEDULER_STOPPED())
 632                 return (1);
 633
 634         rw = rwlock2rw(c);
 635
 636         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
 637             ("rw_try_rlock() by idle thread %p on rwlock %s @ %s:%d",
 638             curthread, rw->lock_object.lo_name, file, line));
 639
 640         for (;;) {
 641                 x = rw->rw_lock;
 642                 KASSERT(rw->rw_lock != RW_DESTROYED,
 643                     ("rw_try_rlock() of destroyed rwlock @ %s:%d", file, line));
 644                 if (!(x & RW_LOCK_READ))
 645                         break;
 646                 if (atomic_cmpset_acq_ptr(&rw->rw_lock, x, x + RW_ONE_READER)) {
 647                         LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 1, file,
 648                             line);
 649                         WITNESS_LOCK(&rw->lock_object, LOP_TRYLOCK, file, line);
 650                         LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire,
 651                             rw, 0, 0, file, line, LOCKSTAT_READER);
 652                         TD_LOCKS_INC(curthread);
 653                         curthread->td_rw_rlocks++;
 654                         return (1);
 655                 }
 656         }
 657
 658         LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 0, file, line);
 659         return (0);
 660 }
 661
 662 static bool __always_inline
 663 __rw_runlock_try(struct rwlock *rw, struct thread *td, uintptr_t *vp)
 664 {
 665
 666         for (;;) {
 667                 /*
 668                  * See if there is more than one read lock held.  If so,
 669                  * just drop one and return.
 670                  */
 671                 if (RW_READERS(*vp) > 1) {
 672                         if (atomic_fcmpset_rel_ptr(&rw->rw_lock, vp,
 673                             *vp - RW_ONE_READER)) {
 674                                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 675                                         CTR4(KTR_LOCK,
 676                                             "%s: %p succeeded %p -> %p",
 677                                             __func__, rw, (void *)*vp,
 678                                             (void *)(*vp - RW_ONE_READER));
 679                                 td->td_rw_rlocks--;
 680                                 return (true);
 681                         }
 682                         continue;
 683                 }
 684                 /*
 685                  * If there aren't any waiters for a write lock, then try
 686                  * to drop it quickly.
 687                  */
 688                 if (!(*vp & RW_LOCK_WAITERS)) {
 689                         MPASS((*vp & ~RW_LOCK_WRITE_SPINNER) ==
 690                             RW_READERS_LOCK(1));
 691                         if (atomic_fcmpset_rel_ptr(&rw->rw_lock, vp,
 692                             RW_UNLOCKED)) {
 693                                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 694                                         CTR2(KTR_LOCK, "%s: %p last succeeded",
 695                                             __func__, rw);
 696                                 td->td_rw_rlocks--;
 697                                 return (true);
 698                         }
 699                         continue;
 700                 }
 701                 break;
 702         }
 703         return (false);
 704 }
 705
 706 static void __noinline
 707 __rw_runlock_hard(volatile uintptr_t *c, struct thread *td, uintptr_t v,
 708     const char *file, int line)
 709 {
 710         struct rwlock *rw;
 711         struct turnstile *ts;
 712         uintptr_t x, queue;
 713
 714         if (SCHEDULER_STOPPED())
 715                 return;
 716
 717         rw = rwlock2rw(c);
 718
 719         for (;;) {
 720                 if (__rw_runlock_try(rw, td, &v))
 721                         break;
 722
 723                 /*
 724                  * Ok, we know we have waiters and we think we are the
 725                  * last reader, so grab the turnstile lock.
 726                  */
 727                 turnstile_chain_lock(&rw->lock_object);
 728                 v = rw->rw_lock & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER);
 729                 MPASS(v & RW_LOCK_WAITERS);
 730
 731                 /*
 732                  * Try to drop our lock leaving the lock in a unlocked
 733                  * state.
 734                  *
 735                  * If you wanted to do explicit lock handoff you'd have to
 736                  * do it here.  You'd also want to use turnstile_signal()
 737                  * and you'd have to handle the race where a higher
 738                  * priority thread blocks on the write lock before the
 739                  * thread you wakeup actually runs and have the new thread
 740                  * "steal" the lock.  For now it's a lot simpler to just
 741                  * wakeup all of the waiters.
 742                  *
 743                  * As above, if we fail, then another thread might have
 744                  * acquired a read lock, so drop the turnstile lock and
 745                  * restart.
 746                  */
 747                 x = RW_UNLOCKED;
 748                 if (v & RW_LOCK_WRITE_WAITERS) {
 749                         queue = TS_EXCLUSIVE_QUEUE;
 750                         x |= (v & RW_LOCK_READ_WAITERS);
 751                 } else
 752                         queue = TS_SHARED_QUEUE;
 753                 if (!atomic_cmpset_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v,
 754                     x)) {
 755                         turnstile_chain_unlock(&rw->lock_object);
 756                         v = RW_READ_VALUE(rw);
 757                         continue;
 758                 }
 759                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 760                         CTR2(KTR_LOCK, "%s: %p last succeeded with waiters",
 761                             __func__, rw);
 762
 763                 /*
 764                  * Ok.  The lock is released and all that's left is to
 765                  * wake up the waiters.  Note that the lock might not be
 766                  * free anymore, but in that case the writers will just
 767                  * block again if they run before the new lock holder(s)
 768                  * release the lock.
 769                  */
 770                 ts = turnstile_lookup(&rw->lock_object);
 771                 MPASS(ts != NULL);
 772                 turnstile_broadcast(ts, queue);
 773                 turnstile_unpend(ts, TS_SHARED_LOCK);
 774                 turnstile_chain_unlock(&rw->lock_object);
 775                 td->td_rw_rlocks--;
 776                 break;
 777         }
 778         LOCKSTAT_PROFILE_RELEASE_RWLOCK(rw__release, rw, LOCKSTAT_READER);
 779 }
 780
 781 void
 782 _rw_runlock_cookie(volatile uintptr_t *c, const char *file, int line)
 783 {
 784         struct rwlock *rw;
 785         struct thread *td;
 786         uintptr_t v;
 787
 788         rw = rwlock2rw(c);
 789
 790         KASSERT(rw->rw_lock != RW_DESTROYED,
 791             ("rw_runlock() of destroyed rwlock @ %s:%d", file, line));
 792         __rw_assert(c, RA_RLOCKED, file, line);
 793         WITNESS_UNLOCK(&rw->lock_object, 0, file, line);
 794         LOCK_LOG_LOCK("RUNLOCK", &rw->lock_object, 0, 0, file, line);
 795
 796         td = curthread;
 797         v = RW_READ_VALUE(rw);
 798
 799         if (__predict_false(LOCKSTAT_OOL_PROFILE_ENABLED(rw__release) ||
 800             !__rw_runlock_try(rw, td, &v)))
 801                 __rw_runlock_hard(c, td, v, file, line);
 802
 803         TD_LOCKS_DEC(curthread);
 804 }
 805
 806
 807 /*
 808  * This function is called when we are unable to obtain a write lock on the
 809  * first try.  This means that at least one other thread holds either a
 810  * read or write lock.
 811  */
 812 void
 813 __rw_wlock_hard(volatile uintptr_t *c, uintptr_t v, uintptr_t tid,
 814     const char *file, int line)
 815 {
 816         struct rwlock *rw;
 817         struct turnstile *ts;
 818 #ifdef ADAPTIVE_RWLOCKS
 819         volatile struct thread *owner;
 820         int spintries = 0;
 821         int i;
 822 #endif
 823         uintptr_t x;
 824 #ifdef LOCK_PROFILING
 825         uint64_t waittime = 0;
 826         int contested = 0;
 827 #endif
 828 #if defined(ADAPTIVE_RWLOCKS) || defined(KDTRACE_HOOKS)
 829         struct lock_delay_arg lda;
 830 #endif
 831 #ifdef KDTRACE_HOOKS
 832         uintptr_t state;
 833         u_int sleep_cnt = 0;
 834         int64_t sleep_time = 0;
 835         int64_t all_time = 0;
 836 #endif
 837
 838         if (SCHEDULER_STOPPED())
 839                 return;
 840
 841 #if defined(ADAPTIVE_RWLOCKS)
 842         lock_delay_arg_init(&lda, &rw_delay);
 843 #elif defined(KDTRACE_HOOKS)
 844         lock_delay_arg_init(&lda, NULL);
 845 #endif
 846         rw = rwlock2rw(c);
 847         if (__predict_false(v == RW_UNLOCKED))
 848                 v = RW_READ_VALUE(rw);
 849
 850         if (__predict_false(lv_rw_wowner(v) == (struct thread *)tid)) {
 851                 KASSERT(rw->lock_object.lo_flags & LO_RECURSABLE,
 852                     ("%s: recursing but non-recursive rw %s @ %s:%d\n",
 853                     __func__, rw->lock_object.lo_name, file, line));
 854                 rw->rw_recurse++;
 855                 atomic_set_ptr(&rw->rw_lock, RW_LOCK_WRITER_RECURSED);
 856                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 857                         CTR2(KTR_LOCK, "%s: %p recursing", __func__, rw);
 858                 return;
 859         }
 860
 861         if (LOCK_LOG_TEST(&rw->lock_object, 0))
 862                 CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__,
 863                     rw->lock_object.lo_name, (void *)rw->rw_lock, file, line);
 864
 865 #ifdef KDTRACE_HOOKS
 866         all_time -= lockstat_nsecs(&rw->lock_object);
 867         state = v;
 868 #endif
 869         for (;;) {
 870                 if (v == RW_UNLOCKED) {
 871                         if (_rw_write_lock_fetch(rw, &v, tid))
 872                                 break;
 873                         continue;
 874                 }
 875 #ifdef KDTRACE_HOOKS
 876                 lda.spin_cnt++;
 877 #endif
 878 #ifdef HWPMC_HOOKS
 879                 PMC_SOFT_CALL( , , lock, failed);
 880 #endif
 881                 lock_profile_obtain_lock_failed(&rw->lock_object,
 882                     &contested, &waittime);
 883 #ifdef ADAPTIVE_RWLOCKS
 884                 /*
 885                  * If the lock is write locked and the owner is
 886                  * running on another CPU, spin until the owner stops
 887                  * running or the state of the lock changes.
 888                  */
 889                 owner = lv_rw_wowner(v);
 890                 if (!(v & RW_LOCK_READ) && TD_IS_RUNNING(owner)) {
 891                         if (LOCK_LOG_TEST(&rw->lock_object, 0))
 892                                 CTR3(KTR_LOCK, "%s: spinning on %p held by %p",
 893                                     __func__, rw, owner);
 894                         KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
 895                             "spinning", "lockname:\"%s\"",
 896                             rw->lock_object.lo_name);
 897                         do {
 898                                 lock_delay(&lda);
 899                                 v = RW_READ_VALUE(rw);
 900                                 owner = lv_rw_wowner(v);
 901                         } while (owner != NULL && TD_IS_RUNNING(owner));
 902                         KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
 903                             "running");
 904                         continue;
 905                 }
 906                 if ((v & RW_LOCK_READ) && RW_READERS(v) &&
 907                     spintries < rowner_retries) {
 908                         if (!(v & RW_LOCK_WRITE_SPINNER)) {
 909                                 if (!atomic_cmpset_ptr(&rw->rw_lock, v,
 910                                     v | RW_LOCK_WRITE_SPINNER)) {
 911                                         v = RW_READ_VALUE(rw);
 912                                         continue;
 913                                 }
 914                         }
 915                         spintries++;
 916                         KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
 917                             "spinning", "lockname:\"%s\"",
 918                             rw->lock_object.lo_name);
 919                         for (i = 0; i < rowner_loops; i++) {
 920                                 if ((rw->rw_lock & RW_LOCK_WRITE_SPINNER) == 0)
 921                                         break;
 922                                 cpu_spinwait();
 923                         }
 924                         KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
 925                             "running");
 926                         v = RW_READ_VALUE(rw);
 927 #ifdef KDTRACE_HOOKS
 928                         lda.spin_cnt += rowner_loops - i;
 929 #endif
 930                         if (i != rowner_loops)
 931                                 continue;
 932                 }
 933 #endif
 934                 ts = turnstile_trywait(&rw->lock_object);
 935                 v = RW_READ_VALUE(rw);
 936
 937 #ifdef ADAPTIVE_RWLOCKS
 938                 /*
 939                  * The current lock owner might have started executing
 940                  * on another CPU (or the lock could have changed
 941                  * owners) while we were waiting on the turnstile
 942                  * chain lock.  If so, drop the turnstile lock and try
 943                  * again.
 944                  */
 945                 if (!(v & RW_LOCK_READ)) {
 946                         owner = (struct thread *)RW_OWNER(v);
 947                         if (TD_IS_RUNNING(owner)) {
 948                                 turnstile_cancel(ts);
 949                                 continue;
 950                         }
 951                 }
 952 #endif
 953                 /*
 954                  * Check for the waiters flags about this rwlock.
 955                  * If the lock was released, without maintain any pending
 956                  * waiters queue, simply try to acquire it.
 957                  * If a pending waiters queue is present, claim the lock
 958                  * ownership and maintain the pending queue.
 959                  */
 960                 x = v & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER);
 961                 if ((v & ~x) == RW_UNLOCKED) {
 962                         x &= ~RW_LOCK_WRITE_SPINNER;
 963                         if (atomic_cmpset_acq_ptr(&rw->rw_lock, v, tid | x)) {
 964                                 if (x)
 965                                         turnstile_claim(ts);
 966                                 else
 967                                         turnstile_cancel(ts);
 968                                 break;
 969                         }
 970                         turnstile_cancel(ts);
 971                         v = RW_READ_VALUE(rw);
 972                         continue;
 973                 }
 974                 /*
 975                  * If the RW_LOCK_WRITE_WAITERS flag isn't set, then try to
 976                  * set it.  If we fail to set it, then loop back and try
 977                  * again.
 978                  */
 979                 if (!(v & RW_LOCK_WRITE_WAITERS)) {
 980                         if (!atomic_cmpset_ptr(&rw->rw_lock, v,
 981                             v | RW_LOCK_WRITE_WAITERS)) {
 982                                 turnstile_cancel(ts);
 983                                 v = RW_READ_VALUE(rw);
 984                                 continue;
 985                         }
 986                         if (LOCK_LOG_TEST(&rw->lock_object, 0))
 987                                 CTR2(KTR_LOCK, "%s: %p set write waiters flag",
 988                                     __func__, rw);
 989                 }
 990                 /*
 991                  * We were unable to acquire the lock and the write waiters
 992                  * flag is set, so we must block on the turnstile.
 993                  */
 994                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 995                         CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__,
 996                             rw);
 997 #ifdef KDTRACE_HOOKS
 998                 sleep_time -= lockstat_nsecs(&rw->lock_object);
 999 #endif
1000                 turnstile_wait(ts, rw_owner(rw), TS_EXCLUSIVE_QUEUE);
1001 #ifdef KDTRACE_HOOKS
1002                 sleep_time += lockstat_nsecs(&rw->lock_object);
1003                 sleep_cnt++;
1004 #endif
1005                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
1006                         CTR2(KTR_LOCK, "%s: %p resuming from turnstile",
1007                             __func__, rw);
1008 #ifdef ADAPTIVE_RWLOCKS
1009                 spintries = 0;
1010 #endif
1011                 v = RW_READ_VALUE(rw);
1012         }
1013 #ifdef KDTRACE_HOOKS
1014         all_time += lockstat_nsecs(&rw->lock_object);
1015         if (sleep_time)
1016                 LOCKSTAT_RECORD4(rw__block, rw, sleep_time,
1017                     LOCKSTAT_WRITER, (state & RW_LOCK_READ) == 0,
1018                     (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
1019
1020         /* Record only the loops spinning and not sleeping. */
1021         if (lda.spin_cnt > sleep_cnt)
1022                 LOCKSTAT_RECORD4(rw__spin, rw, all_time - sleep_time,
1023                     LOCKSTAT_WRITER, (state & RW_LOCK_READ) == 0,
1024                     (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
1025 #endif
1026         LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire, rw, contested,
1027             waittime, file, line, LOCKSTAT_WRITER);
1028 }
1029
1030 /*
1031  * This function is called if lockstat is active or the first try at releasing
1032  * a write lock failed.  The latter means that the lock is recursed or one of
1033  * the 2 waiter bits must be set indicating that at least one thread is waiting
1034  * on this lock.
1035  */
1036 void
1037 __rw_wunlock_hard(volatile uintptr_t *c, uintptr_t tid, const char *file,
1038     int line)
1039 {
1040         struct rwlock *rw;
1041         struct turnstile *ts;
1042         uintptr_t v;
1043         int queue;
1044
1045         if (SCHEDULER_STOPPED())
1046                 return;
1047
1048         rw = rwlock2rw(c);
1049         v = RW_READ_VALUE(rw);
1050         if (v & RW_LOCK_WRITER_RECURSED) {
1051                 if (--(rw->rw_recurse) == 0)
1052                         atomic_clear_ptr(&rw->rw_lock, RW_LOCK_WRITER_RECURSED);
1053                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
1054                         CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, rw);
1055                 return;
1056         }
1057
1058         LOCKSTAT_PROFILE_RELEASE_RWLOCK(rw__release, rw, LOCKSTAT_WRITER);
1059         if (v == tid && _rw_write_unlock(rw, tid))
1060                 return;
1061
1062         KASSERT(rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS),
1063             ("%s: neither of the waiter flags are set", __func__));
1064
1065         if (LOCK_LOG_TEST(&rw->lock_object, 0))
1066                 CTR2(KTR_LOCK, "%s: %p contested", __func__, rw);
1067
1068         turnstile_chain_lock(&rw->lock_object);
1069         ts = turnstile_lookup(&rw->lock_object);
1070         MPASS(ts != NULL);
1071
1072         /*
1073          * Use the same algo as sx locks for now.  Prefer waking up shared
1074          * waiters if we have any over writers.  This is probably not ideal.
1075          *
1076          * 'v' is the value we are going to write back to rw_lock.  If we
1077          * have waiters on both queues, we need to preserve the state of
1078          * the waiter flag for the queue we don't wake up.  For now this is
1079          * hardcoded for the algorithm mentioned above.
1080          *
1081          * In the case of both readers and writers waiting we wakeup the
1082          * readers but leave the RW_LOCK_WRITE_WAITERS flag set.  If a
1083          * new writer comes in before a reader it will claim the lock up
1084          * above.  There is probably a potential priority inversion in
1085          * there that could be worked around either by waking both queues
1086          * of waiters or doing some complicated lock handoff gymnastics.
1087          */
1088         v = RW_UNLOCKED;
1089         if (rw->rw_lock & RW_LOCK_WRITE_WAITERS) {
1090                 queue = TS_EXCLUSIVE_QUEUE;
1091                 v |= (rw->rw_lock & RW_LOCK_READ_WAITERS);
1092         } else
1093                 queue = TS_SHARED_QUEUE;
1094
1095         /* Wake up all waiters for the specific queue. */
1096         if (LOCK_LOG_TEST(&rw->lock_object, 0))
1097                 CTR3(KTR_LOCK, "%s: %p waking up %s waiters", __func__, rw,
1098                     queue == TS_SHARED_QUEUE ? "read" : "write");
1099         turnstile_broadcast(ts, queue);
1100         atomic_store_rel_ptr(&rw->rw_lock, v);
1101         turnstile_unpend(ts, TS_EXCLUSIVE_LOCK);
1102         turnstile_chain_unlock(&rw->lock_object);
1103 }
1104
1105 /*
1106  * Attempt to do a non-blocking upgrade from a read lock to a write
1107  * lock.  This will only succeed if this thread holds a single read
1108  * lock.  Returns true if the upgrade succeeded and false otherwise.
1109  */
1110 int
1111 __rw_try_upgrade(volatile uintptr_t *c, const char *file, int line)
1112 {
1113         struct rwlock *rw;
1114         uintptr_t v, x, tid;
1115         struct turnstile *ts;
1116         int success;
1117
1118         if (SCHEDULER_STOPPED())
1119                 return (1);
1120
1121         rw = rwlock2rw(c);
1122
1123         KASSERT(rw->rw_lock != RW_DESTROYED,
1124             ("rw_try_upgrade() of destroyed rwlock @ %s:%d", file, line));
1125         __rw_assert(c, RA_RLOCKED, file, line);
1126
1127         /*
1128          * Attempt to switch from one reader to a writer.  If there
1129          * are any write waiters, then we will have to lock the
1130          * turnstile first to prevent races with another writer
1131          * calling turnstile_wait() before we have claimed this
1132          * turnstile.  So, do the simple case of no waiters first.
1133          */
1134         tid = (uintptr_t)curthread;
1135         success = 0;
1136         for (;;) {
1137                 v = rw->rw_lock;
1138                 if (RW_READERS(v) > 1)
1139                         break;
1140                 if (!(v & RW_LOCK_WAITERS)) {
1141                         success = atomic_cmpset_ptr(&rw->rw_lock, v, tid);
1142                         if (!success)
1143                                 continue;
1144                         break;
1145                 }
1146
1147                 /*
1148                  * Ok, we think we have waiters, so lock the turnstile.
1149                  */
1150                 ts = turnstile_trywait(&rw->lock_object);
1151                 v = rw->rw_lock;
1152                 if (RW_READERS(v) > 1) {
1153                         turnstile_cancel(ts);
1154                         break;
1155                 }
1156                 /*
1157                  * Try to switch from one reader to a writer again.  This time
1158                  * we honor the current state of the waiters flags.
1159                  * If we obtain the lock with the flags set, then claim
1160                  * ownership of the turnstile.
1161                  */
1162                 x = rw->rw_lock & RW_LOCK_WAITERS;
1163                 success = atomic_cmpset_ptr(&rw->rw_lock, v, tid | x);
1164                 if (success) {
1165                         if (x)
1166                                 turnstile_claim(ts);
1167                         else
1168                                 turnstile_cancel(ts);
1169                         break;
1170                 }
1171                 turnstile_cancel(ts);
1172         }
1173         LOCK_LOG_TRY("WUPGRADE", &rw->lock_object, 0, success, file, line);
1174         if (success) {
1175                 curthread->td_rw_rlocks--;
1176                 WITNESS_UPGRADE(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
1177                     file, line);
1178                 LOCKSTAT_RECORD0(rw__upgrade, rw);
1179         }
1180         return (success);
1181 }
1182
1183 /*
1184  * Downgrade a write lock into a single read lock.
1185  */
1186 void
1187 __rw_downgrade(volatile uintptr_t *c, const char *file, int line)
1188 {
1189         struct rwlock *rw;
1190         struct turnstile *ts;
1191         uintptr_t tid, v;
1192         int rwait, wwait;
1193
1194         if (SCHEDULER_STOPPED())
1195                 return;
1196
1197         rw = rwlock2rw(c);
1198
1199         KASSERT(rw->rw_lock != RW_DESTROYED,
1200             ("rw_downgrade() of destroyed rwlock @ %s:%d", file, line));
1201         __rw_assert(c, RA_WLOCKED | RA_NOTRECURSED, file, line);
1202 #ifndef INVARIANTS
1203         if (rw_recursed(rw))
1204                 panic("downgrade of a recursed lock");
1205 #endif
1206
1207         WITNESS_DOWNGRADE(&rw->lock_object, 0, file, line);
1208
1209         /*
1210          * Convert from a writer to a single reader.  First we handle
1211          * the easy case with no waiters.  If there are any waiters, we
1212          * lock the turnstile and "disown" the lock.
1213          */
1214         tid = (uintptr_t)curthread;
1215         if (atomic_cmpset_rel_ptr(&rw->rw_lock, tid, RW_READERS_LOCK(1)))
1216                 goto out;
1217
1218         /*
1219          * Ok, we think we have waiters, so lock the turnstile so we can
1220          * read the waiter flags without any races.
1221          */
1222         turnstile_chain_lock(&rw->lock_object);
1223         v = rw->rw_lock & RW_LOCK_WAITERS;
1224         rwait = v & RW_LOCK_READ_WAITERS;
1225         wwait = v & RW_LOCK_WRITE_WAITERS;
1226         MPASS(rwait | wwait);
1227
1228         /*
1229          * Downgrade from a write lock while preserving waiters flag
1230          * and give up ownership of the turnstile.
1231          */
1232         ts = turnstile_lookup(&rw->lock_object);
1233         MPASS(ts != NULL);
1234         if (!wwait)
1235                 v &= ~RW_LOCK_READ_WAITERS;
1236         atomic_store_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v);
1237         /*
1238          * Wake other readers if there are no writers pending.  Otherwise they
1239          * won't be able to acquire the lock anyway.
1240          */
1241         if (rwait && !wwait) {
1242                 turnstile_broadcast(ts, TS_SHARED_QUEUE);
1243                 turnstile_unpend(ts, TS_EXCLUSIVE_LOCK);
1244         } else
1245                 turnstile_disown(ts);
1246         turnstile_chain_unlock(&rw->lock_object);
1247 out:
1248         curthread->td_rw_rlocks++;
1249         LOCK_LOG_LOCK("WDOWNGRADE", &rw->lock_object, 0, 0, file, line);
1250         LOCKSTAT_RECORD0(rw__downgrade, rw);
1251 }
1252
1253 #ifdef INVARIANT_SUPPORT
1254 #ifndef INVARIANTS
1255 #undef __rw_assert
1256 #endif
1257
1258 /*
1259  * In the non-WITNESS case, rw_assert() can only detect that at least
1260  * *some* thread owns an rlock, but it cannot guarantee that *this*
1261  * thread owns an rlock.
1262  */
1263 void
1264 __rw_assert(const volatile uintptr_t *c, int what, const char *file, int line)
1265 {
1266         const struct rwlock *rw;
1267
1268         if (panicstr != NULL)
1269                 return;
1270
1271         rw = rwlock2rw(c);
1272
1273         switch (what) {
1274         case RA_LOCKED:
1275         case RA_LOCKED | RA_RECURSED:
1276         case RA_LOCKED | RA_NOTRECURSED:
1277         case RA_RLOCKED:
1278         case RA_RLOCKED | RA_RECURSED:
1279         case RA_RLOCKED | RA_NOTRECURSED:
1280 #ifdef WITNESS
1281                 witness_assert(&rw->lock_object, what, file, line);
1282 #else
1283                 /*
1284                  * If some other thread has a write lock or we have one
1285                  * and are asserting a read lock, fail.  Also, if no one
1286                  * has a lock at all, fail.
1287                  */
1288                 if (rw->rw_lock == RW_UNLOCKED ||
1289                     (!(rw->rw_lock & RW_LOCK_READ) && (what & RA_RLOCKED ||
1290                     rw_wowner(rw) != curthread)))
1291                         panic("Lock %s not %slocked @ %s:%d\n",
1292                             rw->lock_object.lo_name, (what & RA_RLOCKED) ?
1293                             "read " : "", file, line);
1294
1295                 if (!(rw->rw_lock & RW_LOCK_READ) && !(what & RA_RLOCKED)) {
1296                         if (rw_recursed(rw)) {
1297                                 if (what & RA_NOTRECURSED)
1298                                         panic("Lock %s recursed @ %s:%d\n",
1299                                             rw->lock_object.lo_name, file,
1300                                             line);
1301                         } else if (what & RA_RECURSED)
1302                                 panic("Lock %s not recursed @ %s:%d\n",
1303                                     rw->lock_object.lo_name, file, line);
1304                 }
1305 #endif
1306                 break;
1307         case RA_WLOCKED:
1308         case RA_WLOCKED | RA_RECURSED:
1309         case RA_WLOCKED | RA_NOTRECURSED:
1310                 if (rw_wowner(rw) != curthread)
1311                         panic("Lock %s not exclusively locked @ %s:%d\n",
1312                             rw->lock_object.lo_name, file, line);
1313                 if (rw_recursed(rw)) {
1314                         if (what & RA_NOTRECURSED)
1315                                 panic("Lock %s recursed @ %s:%d\n",
1316                                     rw->lock_object.lo_name, file, line);
1317                 } else if (what & RA_RECURSED)
1318                         panic("Lock %s not recursed @ %s:%d\n",
1319                             rw->lock_object.lo_name, file, line);
1320                 break;
1321         case RA_UNLOCKED:
1322 #ifdef WITNESS
1323                 witness_assert(&rw->lock_object, what, file, line);
1324 #else
1325                 /*
1326                  * If we hold a write lock fail.  We can't reliably check
1327                  * to see if we hold a read lock or not.
1328                  */
1329                 if (rw_wowner(rw) == curthread)
1330                         panic("Lock %s exclusively locked @ %s:%d\n",
1331                             rw->lock_object.lo_name, file, line);
1332 #endif
1333                 break;
1334         default:
1335                 panic("Unknown rw lock assertion: %d @ %s:%d", what, file,
1336                     line);
1337         }
1338 }
1339 #endif /* INVARIANT_SUPPORT */
1340
1341 #ifdef DDB
1342 void
1343 db_show_rwlock(const struct lock_object *lock)
1344 {
1345         const struct rwlock *rw;
1346         struct thread *td;
1347
1348         rw = (const struct rwlock *)lock;
1349
1350         db_printf(" state: ");
1351         if (rw->rw_lock == RW_UNLOCKED)
1352                 db_printf("UNLOCKED\n");
1353         else if (rw->rw_lock == RW_DESTROYED) {
1354                 db_printf("DESTROYED\n");
1355                 return;
1356         } else if (rw->rw_lock & RW_LOCK_READ)
1357                 db_printf("RLOCK: %ju locks\n",
1358                     (uintmax_t)(RW_READERS(rw->rw_lock)));
1359         else {
1360                 td = rw_wowner(rw);
1361                 db_printf("WLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1362                     td->td_tid, td->td_proc->p_pid, td->td_name);
1363                 if (rw_recursed(rw))
1364                         db_printf(" recursed: %u\n", rw->rw_recurse);
1365         }
1366         db_printf(" waiters: ");
1367         switch (rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS)) {
1368         case RW_LOCK_READ_WAITERS:
1369                 db_printf("readers\n");
1370                 break;
1371         case RW_LOCK_WRITE_WAITERS:
1372                 db_printf("writers\n");
1373                 break;
1374         case RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS:
1375                 db_printf("readers and writers\n");
1376                 break;
1377         default:
1378                 db_printf("none\n");
1379                 break;
1380         }
1381 }
1382
1383 #endif