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 __read_frequently rowner_retries = 10;
  97 static int __read_frequently 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_frequently 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 || SCHEDULER_STOPPED() ||
 273             !TD_IS_IDLETHREAD(curthread),
 274             ("rw_wlock() by idle thread %p on rwlock %s @ %s:%d",
 275             curthread, rw->lock_object.lo_name, file, line));
 276         KASSERT(rw->rw_lock != RW_DESTROYED,
 277             ("rw_wlock() of destroyed rwlock @ %s:%d", file, line));
 278         WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE, file,
 279             line, NULL);
 280         tid = (uintptr_t)curthread;
 281         v = RW_UNLOCKED;
 282         if (!_rw_write_lock_fetch(rw, &v, tid))
 283                 _rw_wlock_hard(rw, v, tid, file, line);
 284         else
 285                 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire, rw,
 286                     0, 0, file, line, LOCKSTAT_WRITER);
 287
 288         LOCK_LOG_LOCK("WLOCK", &rw->lock_object, 0, rw->rw_recurse, file, line);
 289         WITNESS_LOCK(&rw->lock_object, LOP_EXCLUSIVE, file, line);
 290         TD_LOCKS_INC(curthread);
 291 }
 292
 293 int
 294 __rw_try_wlock(volatile uintptr_t *c, const char *file, int line)
 295 {
 296         struct rwlock *rw;
 297         struct thread *td;
 298         uintptr_t tid, v;
 299         int rval;
 300         bool recursed;
 301
 302         td = curthread;
 303         tid = (uintptr_t)td;
 304         if (SCHEDULER_STOPPED_TD(td))
 305                 return (1);
 306
 307         rw = rwlock2rw(c);
 308
 309         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(td),
 310             ("rw_try_wlock() by idle thread %p on rwlock %s @ %s:%d",
 311             curthread, rw->lock_object.lo_name, file, line));
 312         KASSERT(rw->rw_lock != RW_DESTROYED,
 313             ("rw_try_wlock() of destroyed rwlock @ %s:%d", file, line));
 314
 315         rval = 1;
 316         recursed = false;
 317         v = RW_UNLOCKED;
 318         for (;;) {
 319                 if (atomic_fcmpset_acq_ptr(&rw->rw_lock, &v, tid))
 320                         break;
 321                 if (v == RW_UNLOCKED)
 322                         continue;
 323                 if (v == tid && (rw->lock_object.lo_flags & LO_RECURSABLE)) {
 324                         rw->rw_recurse++;
 325                         atomic_set_ptr(&rw->rw_lock, RW_LOCK_WRITER_RECURSED);
 326                         break;
 327                 }
 328                 rval = 0;
 329                 break;
 330         }
 331
 332         LOCK_LOG_TRY("WLOCK", &rw->lock_object, 0, rval, file, line);
 333         if (rval) {
 334                 WITNESS_LOCK(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
 335                     file, line);
 336                 if (!recursed)
 337                         LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire,
 338                             rw, 0, 0, file, line, LOCKSTAT_WRITER);
 339                 TD_LOCKS_INC(curthread);
 340         }
 341         return (rval);
 342 }
 343
 344 void
 345 _rw_wunlock_cookie(volatile uintptr_t *c, const char *file, int line)
 346 {
 347         struct rwlock *rw;
 348
 349         rw = rwlock2rw(c);
 350
 351         KASSERT(rw->rw_lock != RW_DESTROYED,
 352             ("rw_wunlock() of destroyed rwlock @ %s:%d", file, line));
 353         __rw_assert(c, RA_WLOCKED, file, line);
 354         WITNESS_UNLOCK(&rw->lock_object, LOP_EXCLUSIVE, file, line);
 355         LOCK_LOG_LOCK("WUNLOCK", &rw->lock_object, 0, rw->rw_recurse, file,
 356             line);
 357
 358 #ifdef LOCK_PROFILING
 359         _rw_wunlock_hard(rw, (uintptr_t)curthread, file, line);
 360 #else
 361         __rw_wunlock(rw, curthread, file, line);
 362 #endif
 363
 364         TD_LOCKS_DEC(curthread);
 365 }
 366
 367 /*
 368  * Determines whether a new reader can acquire a lock.  Succeeds if the
 369  * reader already owns a read lock and the lock is locked for read to
 370  * prevent deadlock from reader recursion.  Also succeeds if the lock
 371  * is unlocked and has no writer waiters or spinners.  Failing otherwise
 372  * prioritizes writers before readers.
 373  */
 374 #define RW_CAN_READ(td, _rw)                                            \
 375     (((_rw) & (RW_LOCK_READ | RW_LOCK_WRITE_WAITERS | RW_LOCK_WRITE_SPINNER)) ==\
 376     RW_LOCK_READ || ((td)->td_rw_rlocks && (_rw) & RW_LOCK_READ))
 377
 378 static bool __always_inline
 379 __rw_rlock_try(struct rwlock *rw, struct thread *td, uintptr_t *vp,
 380     const char *file, int line)
 381 {
 382
 383         /*
 384          * Handle the easy case.  If no other thread has a write
 385          * lock, then try to bump up the count of read locks.  Note
 386          * that we have to preserve the current state of the
 387          * RW_LOCK_WRITE_WAITERS flag.  If we fail to acquire a
 388          * read lock, then rw_lock must have changed, so restart
 389          * the loop.  Note that this handles the case of a
 390          * completely unlocked rwlock since such a lock is encoded
 391          * as a read lock with no waiters.
 392          */
 393         while (RW_CAN_READ(td, *vp)) {
 394                 if (atomic_fcmpset_acq_ptr(&rw->rw_lock, vp,
 395                         *vp + RW_ONE_READER)) {
 396                         if (LOCK_LOG_TEST(&rw->lock_object, 0))
 397                                 CTR4(KTR_LOCK,
 398                                     "%s: %p succeed %p -> %p", __func__,
 399                                     rw, (void *)*vp,
 400                                     (void *)(*vp + RW_ONE_READER));
 401                         td->td_rw_rlocks++;
 402                         return (true);
 403                 }
 404         }
 405         return (false);
 406 }
 407
 408 static void __noinline
 409 __rw_rlock_hard(volatile uintptr_t *c, struct thread *td, uintptr_t v,
 410     const char *file, int line)
 411 {
 412         struct rwlock *rw;
 413         struct turnstile *ts;
 414 #ifdef ADAPTIVE_RWLOCKS
 415         volatile struct thread *owner;
 416         int spintries = 0;
 417         int i, n;
 418 #endif
 419 #ifdef LOCK_PROFILING
 420         uint64_t waittime = 0;
 421         int contested = 0;
 422 #endif
 423 #if defined(ADAPTIVE_RWLOCKS) || defined(KDTRACE_HOOKS)
 424         struct lock_delay_arg lda;
 425 #endif
 426 #ifdef KDTRACE_HOOKS
 427         uintptr_t state;
 428         u_int sleep_cnt = 0;
 429         int64_t sleep_time = 0;
 430         int64_t all_time = 0;
 431 #endif
 432
 433         if (SCHEDULER_STOPPED())
 434                 return;
 435
 436 #if defined(ADAPTIVE_RWLOCKS)
 437         lock_delay_arg_init(&lda, &rw_delay);
 438 #elif defined(KDTRACE_HOOKS)
 439         lock_delay_arg_init(&lda, NULL);
 440 #endif
 441         rw = rwlock2rw(c);
 442
 443 #ifdef KDTRACE_HOOKS
 444         all_time -= lockstat_nsecs(&rw->lock_object);
 445 #endif
 446 #ifdef KDTRACE_HOOKS
 447         state = v;
 448 #endif
 449         for (;;) {
 450                 if (__rw_rlock_try(rw, td, &v, file, line))
 451                         break;
 452 #ifdef KDTRACE_HOOKS
 453                 lda.spin_cnt++;
 454 #endif
 455 #ifdef HWPMC_HOOKS
 456                 PMC_SOFT_CALL( , , lock, failed);
 457 #endif
 458                 lock_profile_obtain_lock_failed(&rw->lock_object,
 459                     &contested, &waittime);
 460
 461 #ifdef ADAPTIVE_RWLOCKS
 462                 /*
 463                  * If the owner is running on another CPU, spin until
 464                  * the owner stops running or the state of the lock
 465                  * changes.
 466                  */
 467                 if ((v & RW_LOCK_READ) == 0) {
 468                         owner = (struct thread *)RW_OWNER(v);
 469                         if (TD_IS_RUNNING(owner)) {
 470                                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 471                                         CTR3(KTR_LOCK,
 472                                             "%s: spinning on %p held by %p",
 473                                             __func__, rw, owner);
 474                                 KTR_STATE1(KTR_SCHED, "thread",
 475                                     sched_tdname(curthread), "spinning",
 476                                     "lockname:\"%s\"", rw->lock_object.lo_name);
 477                                 do {
 478                                         lock_delay(&lda);
 479                                         v = RW_READ_VALUE(rw);
 480                                         owner = lv_rw_wowner(v);
 481                                 } while (owner != NULL && TD_IS_RUNNING(owner));
 482                                 KTR_STATE0(KTR_SCHED, "thread",
 483                                     sched_tdname(curthread), "running");
 484                                 continue;
 485                         }
 486                 } else if (spintries < rowner_retries) {
 487                         spintries++;
 488                         KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
 489                             "spinning", "lockname:\"%s\"",
 490                             rw->lock_object.lo_name);
 491                         for (i = 0; i < rowner_loops; i += n) {
 492                                 n = RW_READERS(v);
 493                                 lock_delay_spin(n);
 494                                 v = RW_READ_VALUE(rw);
 495                                 if ((v & RW_LOCK_READ) == 0 || RW_CAN_READ(td, v))
 496                                         break;
 497                         }
 498 #ifdef KDTRACE_HOOKS
 499                         lda.spin_cnt += rowner_loops - i;
 500 #endif
 501                         KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
 502                             "running");
 503                         if (i != rowner_loops)
 504                                 continue;
 505                 }
 506 #endif
 507
 508                 /*
 509                  * Okay, now it's the hard case.  Some other thread already
 510                  * has a write lock or there are write waiters present,
 511                  * acquire the turnstile lock so we can begin the process
 512                  * of blocking.
 513                  */
 514                 ts = turnstile_trywait(&rw->lock_object);
 515
 516                 /*
 517                  * The lock might have been released while we spun, so
 518                  * recheck its state and restart the loop if needed.
 519                  */
 520                 v = RW_READ_VALUE(rw);
 521                 if (RW_CAN_READ(td, v)) {
 522                         turnstile_cancel(ts);
 523                         continue;
 524                 }
 525
 526 #ifdef ADAPTIVE_RWLOCKS
 527                 /*
 528                  * The current lock owner might have started executing
 529                  * on another CPU (or the lock could have changed
 530                  * owners) while we were waiting on the turnstile
 531                  * chain lock.  If so, drop the turnstile lock and try
 532                  * again.
 533                  */
 534                 if ((v & RW_LOCK_READ) == 0) {
 535                         owner = (struct thread *)RW_OWNER(v);
 536                         if (TD_IS_RUNNING(owner)) {
 537                                 turnstile_cancel(ts);
 538                                 continue;
 539                         }
 540                 }
 541 #endif
 542
 543                 /*
 544                  * The lock is held in write mode or it already has waiters.
 545                  */
 546                 MPASS(!RW_CAN_READ(td, v));
 547
 548                 /*
 549                  * If the RW_LOCK_READ_WAITERS flag is already set, then
 550                  * we can go ahead and block.  If it is not set then try
 551                  * to set it.  If we fail to set it drop the turnstile
 552                  * lock and restart the loop.
 553                  */
 554                 if (!(v & RW_LOCK_READ_WAITERS)) {
 555                         if (!atomic_cmpset_ptr(&rw->rw_lock, v,
 556                             v | RW_LOCK_READ_WAITERS)) {
 557                                 turnstile_cancel(ts);
 558                                 v = RW_READ_VALUE(rw);
 559                                 continue;
 560                         }
 561                         if (LOCK_LOG_TEST(&rw->lock_object, 0))
 562                                 CTR2(KTR_LOCK, "%s: %p set read waiters flag",
 563                                     __func__, rw);
 564                 }
 565
 566                 /*
 567                  * We were unable to acquire the lock and the read waiters
 568                  * flag is set, so we must block on the turnstile.
 569                  */
 570                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 571                         CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__,
 572                             rw);
 573 #ifdef KDTRACE_HOOKS
 574                 sleep_time -= lockstat_nsecs(&rw->lock_object);
 575 #endif
 576                 turnstile_wait(ts, rw_owner(rw), TS_SHARED_QUEUE);
 577 #ifdef KDTRACE_HOOKS
 578                 sleep_time += lockstat_nsecs(&rw->lock_object);
 579                 sleep_cnt++;
 580 #endif
 581                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 582                         CTR2(KTR_LOCK, "%s: %p resuming from turnstile",
 583                             __func__, rw);
 584                 v = RW_READ_VALUE(rw);
 585         }
 586 #ifdef KDTRACE_HOOKS
 587         all_time += lockstat_nsecs(&rw->lock_object);
 588         if (sleep_time)
 589                 LOCKSTAT_RECORD4(rw__block, rw, sleep_time,
 590                     LOCKSTAT_READER, (state & RW_LOCK_READ) == 0,
 591                     (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
 592
 593         /* Record only the loops spinning and not sleeping. */
 594         if (lda.spin_cnt > sleep_cnt)
 595                 LOCKSTAT_RECORD4(rw__spin, rw, all_time - sleep_time,
 596                     LOCKSTAT_READER, (state & RW_LOCK_READ) == 0,
 597                     (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
 598 #endif
 599         /*
 600          * TODO: acquire "owner of record" here.  Here be turnstile dragons
 601          * however.  turnstiles don't like owners changing between calls to
 602          * turnstile_wait() currently.
 603          */
 604         LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire, rw, contested,
 605             waittime, file, line, LOCKSTAT_READER);
 606 }
 607
 608 void
 609 __rw_rlock(volatile uintptr_t *c, const char *file, int line)
 610 {
 611         struct rwlock *rw;
 612         struct thread *td;
 613         uintptr_t v;
 614
 615         td = curthread;
 616         rw = rwlock2rw(c);
 617
 618         KASSERT(kdb_active != 0 || SCHEDULER_STOPPED_TD(td) ||
 619             !TD_IS_IDLETHREAD(td),
 620             ("rw_rlock() by idle thread %p on rwlock %s @ %s:%d",
 621             td, rw->lock_object.lo_name, file, line));
 622         KASSERT(rw->rw_lock != RW_DESTROYED,
 623             ("rw_rlock() of destroyed rwlock @ %s:%d", file, line));
 624         KASSERT(rw_wowner(rw) != td,
 625             ("rw_rlock: wlock already held for %s @ %s:%d",
 626             rw->lock_object.lo_name, file, line));
 627         WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER, file, line, NULL);
 628
 629         v = RW_READ_VALUE(rw);
 630         if (__predict_false(LOCKSTAT_OOL_PROFILE_ENABLED(rw__acquire) ||
 631             !__rw_rlock_try(rw, td, &v, file, line)))
 632                 __rw_rlock_hard(c, td, v, file, line);
 633
 634         LOCK_LOG_LOCK("RLOCK", &rw->lock_object, 0, 0, file, line);
 635         WITNESS_LOCK(&rw->lock_object, 0, file, line);
 636         TD_LOCKS_INC(curthread);
 637 }
 638
 639 int
 640 __rw_try_rlock(volatile uintptr_t *c, const char *file, int line)
 641 {
 642         struct rwlock *rw;
 643         uintptr_t x;
 644
 645         if (SCHEDULER_STOPPED())
 646                 return (1);
 647
 648         rw = rwlock2rw(c);
 649
 650         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
 651             ("rw_try_rlock() by idle thread %p on rwlock %s @ %s:%d",
 652             curthread, rw->lock_object.lo_name, file, line));
 653
 654         x = rw->rw_lock;
 655         for (;;) {
 656                 KASSERT(rw->rw_lock != RW_DESTROYED,
 657                     ("rw_try_rlock() of destroyed rwlock @ %s:%d", file, line));
 658                 if (!(x & RW_LOCK_READ))
 659                         break;
 660                 if (atomic_fcmpset_acq_ptr(&rw->rw_lock, &x, x + RW_ONE_READER)) {
 661                         LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 1, file,
 662                             line);
 663                         WITNESS_LOCK(&rw->lock_object, LOP_TRYLOCK, file, line);
 664                         LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire,
 665                             rw, 0, 0, file, line, LOCKSTAT_READER);
 666                         TD_LOCKS_INC(curthread);
 667                         curthread->td_rw_rlocks++;
 668                         return (1);
 669                 }
 670         }
 671
 672         LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 0, file, line);
 673         return (0);
 674 }
 675
 676 static bool __always_inline
 677 __rw_runlock_try(struct rwlock *rw, struct thread *td, uintptr_t *vp)
 678 {
 679
 680         for (;;) {
 681                 /*
 682                  * See if there is more than one read lock held.  If so,
 683                  * just drop one and return.
 684                  */
 685                 if (RW_READERS(*vp) > 1) {
 686                         if (atomic_fcmpset_rel_ptr(&rw->rw_lock, vp,
 687                             *vp - RW_ONE_READER)) {
 688                                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 689                                         CTR4(KTR_LOCK,
 690                                             "%s: %p succeeded %p -> %p",
 691                                             __func__, rw, (void *)*vp,
 692                                             (void *)(*vp - RW_ONE_READER));
 693                                 td->td_rw_rlocks--;
 694                                 return (true);
 695                         }
 696                         continue;
 697                 }
 698                 /*
 699                  * If there aren't any waiters for a write lock, then try
 700                  * to drop it quickly.
 701                  */
 702                 if (!(*vp & RW_LOCK_WAITERS)) {
 703                         MPASS((*vp & ~RW_LOCK_WRITE_SPINNER) ==
 704                             RW_READERS_LOCK(1));
 705                         if (atomic_fcmpset_rel_ptr(&rw->rw_lock, vp,
 706                             RW_UNLOCKED)) {
 707                                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 708                                         CTR2(KTR_LOCK, "%s: %p last succeeded",
 709                                             __func__, rw);
 710                                 td->td_rw_rlocks--;
 711                                 return (true);
 712                         }
 713                         continue;
 714                 }
 715                 break;
 716         }
 717         return (false);
 718 }
 719
 720 static void __noinline
 721 __rw_runlock_hard(volatile uintptr_t *c, struct thread *td, uintptr_t v,
 722     const char *file, int line)
 723 {
 724         struct rwlock *rw;
 725         struct turnstile *ts;
 726         uintptr_t x, queue;
 727
 728         if (SCHEDULER_STOPPED())
 729                 return;
 730
 731         rw = rwlock2rw(c);
 732
 733         for (;;) {
 734                 if (__rw_runlock_try(rw, td, &v))
 735                         break;
 736
 737                 /*
 738                  * Ok, we know we have waiters and we think we are the
 739                  * last reader, so grab the turnstile lock.
 740                  */
 741                 turnstile_chain_lock(&rw->lock_object);
 742                 v = rw->rw_lock & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER);
 743                 MPASS(v & RW_LOCK_WAITERS);
 744
 745                 /*
 746                  * Try to drop our lock leaving the lock in a unlocked
 747                  * state.
 748                  *
 749                  * If you wanted to do explicit lock handoff you'd have to
 750                  * do it here.  You'd also want to use turnstile_signal()
 751                  * and you'd have to handle the race where a higher
 752                  * priority thread blocks on the write lock before the
 753                  * thread you wakeup actually runs and have the new thread
 754                  * "steal" the lock.  For now it's a lot simpler to just
 755                  * wakeup all of the waiters.
 756                  *
 757                  * As above, if we fail, then another thread might have
 758                  * acquired a read lock, so drop the turnstile lock and
 759                  * restart.
 760                  */
 761                 x = RW_UNLOCKED;
 762                 if (v & RW_LOCK_WRITE_WAITERS) {
 763                         queue = TS_EXCLUSIVE_QUEUE;
 764                         x |= (v & RW_LOCK_READ_WAITERS);
 765                 } else
 766                         queue = TS_SHARED_QUEUE;
 767                 if (!atomic_cmpset_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v,
 768                     x)) {
 769                         turnstile_chain_unlock(&rw->lock_object);
 770                         v = RW_READ_VALUE(rw);
 771                         continue;
 772                 }
 773                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 774                         CTR2(KTR_LOCK, "%s: %p last succeeded with waiters",
 775                             __func__, rw);
 776
 777                 /*
 778                  * Ok.  The lock is released and all that's left is to
 779                  * wake up the waiters.  Note that the lock might not be
 780                  * free anymore, but in that case the writers will just
 781                  * block again if they run before the new lock holder(s)
 782                  * release the lock.
 783                  */
 784                 ts = turnstile_lookup(&rw->lock_object);
 785                 MPASS(ts != NULL);
 786                 turnstile_broadcast(ts, queue);
 787                 turnstile_unpend(ts, TS_SHARED_LOCK);
 788                 turnstile_chain_unlock(&rw->lock_object);
 789                 td->td_rw_rlocks--;
 790                 break;
 791         }
 792         LOCKSTAT_PROFILE_RELEASE_RWLOCK(rw__release, rw, LOCKSTAT_READER);
 793 }
 794
 795 void
 796 _rw_runlock_cookie(volatile uintptr_t *c, const char *file, int line)
 797 {
 798         struct rwlock *rw;
 799         struct thread *td;
 800         uintptr_t v;
 801
 802         rw = rwlock2rw(c);
 803
 804         KASSERT(rw->rw_lock != RW_DESTROYED,
 805             ("rw_runlock() of destroyed rwlock @ %s:%d", file, line));
 806         __rw_assert(c, RA_RLOCKED, file, line);
 807         WITNESS_UNLOCK(&rw->lock_object, 0, file, line);
 808         LOCK_LOG_LOCK("RUNLOCK", &rw->lock_object, 0, 0, file, line);
 809
 810         td = curthread;
 811         v = RW_READ_VALUE(rw);
 812
 813         if (__predict_false(LOCKSTAT_OOL_PROFILE_ENABLED(rw__release) ||
 814             !__rw_runlock_try(rw, td, &v)))
 815                 __rw_runlock_hard(c, td, v, file, line);
 816
 817         TD_LOCKS_DEC(curthread);
 818 }
 819
 820 /*
 821  * This function is called when we are unable to obtain a write lock on the
 822  * first try.  This means that at least one other thread holds either a
 823  * read or write lock.
 824  */
 825 void
 826 __rw_wlock_hard(volatile uintptr_t *c, uintptr_t v, uintptr_t tid,
 827     const char *file, int line)
 828 {
 829         struct rwlock *rw;
 830         struct turnstile *ts;
 831 #ifdef ADAPTIVE_RWLOCKS
 832         volatile struct thread *owner;
 833         int spintries = 0;
 834         int i, n;
 835 #endif
 836         uintptr_t x;
 837 #ifdef LOCK_PROFILING
 838         uint64_t waittime = 0;
 839         int contested = 0;
 840 #endif
 841 #if defined(ADAPTIVE_RWLOCKS) || defined(KDTRACE_HOOKS)
 842         struct lock_delay_arg lda;
 843 #endif
 844 #ifdef KDTRACE_HOOKS
 845         uintptr_t state;
 846         u_int sleep_cnt = 0;
 847         int64_t sleep_time = 0;
 848         int64_t all_time = 0;
 849 #endif
 850
 851         if (SCHEDULER_STOPPED())
 852                 return;
 853
 854 #if defined(ADAPTIVE_RWLOCKS)
 855         lock_delay_arg_init(&lda, &rw_delay);
 856 #elif defined(KDTRACE_HOOKS)
 857         lock_delay_arg_init(&lda, NULL);
 858 #endif
 859         rw = rwlock2rw(c);
 860         if (__predict_false(v == RW_UNLOCKED))
 861                 v = RW_READ_VALUE(rw);
 862
 863         if (__predict_false(lv_rw_wowner(v) == (struct thread *)tid)) {
 864                 KASSERT(rw->lock_object.lo_flags & LO_RECURSABLE,
 865                     ("%s: recursing but non-recursive rw %s @ %s:%d\n",
 866                     __func__, rw->lock_object.lo_name, file, line));
 867                 rw->rw_recurse++;
 868                 atomic_set_ptr(&rw->rw_lock, RW_LOCK_WRITER_RECURSED);
 869                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 870                         CTR2(KTR_LOCK, "%s: %p recursing", __func__, rw);
 871                 return;
 872         }
 873
 874         if (LOCK_LOG_TEST(&rw->lock_object, 0))
 875                 CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__,
 876                     rw->lock_object.lo_name, (void *)rw->rw_lock, file, line);
 877
 878 #ifdef KDTRACE_HOOKS
 879         all_time -= lockstat_nsecs(&rw->lock_object);
 880         state = v;
 881 #endif
 882         for (;;) {
 883                 if (v == RW_UNLOCKED) {
 884                         if (_rw_write_lock_fetch(rw, &v, tid))
 885                                 break;
 886                         continue;
 887                 }
 888 #ifdef KDTRACE_HOOKS
 889                 lda.spin_cnt++;
 890 #endif
 891 #ifdef HWPMC_HOOKS
 892                 PMC_SOFT_CALL( , , lock, failed);
 893 #endif
 894                 lock_profile_obtain_lock_failed(&rw->lock_object,
 895                     &contested, &waittime);
 896 #ifdef ADAPTIVE_RWLOCKS
 897                 /*
 898                  * If the lock is write locked and the owner is
 899                  * running on another CPU, spin until the owner stops
 900                  * running or the state of the lock changes.
 901                  */
 902                 owner = lv_rw_wowner(v);
 903                 if (!(v & RW_LOCK_READ) && TD_IS_RUNNING(owner)) {
 904                         if (LOCK_LOG_TEST(&rw->lock_object, 0))
 905                                 CTR3(KTR_LOCK, "%s: spinning on %p held by %p",
 906                                     __func__, rw, owner);
 907                         KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
 908                             "spinning", "lockname:\"%s\"",
 909                             rw->lock_object.lo_name);
 910                         do {
 911                                 lock_delay(&lda);
 912                                 v = RW_READ_VALUE(rw);
 913                                 owner = lv_rw_wowner(v);
 914                         } while (owner != NULL && TD_IS_RUNNING(owner));
 915                         KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
 916                             "running");
 917                         continue;
 918                 }
 919                 if ((v & RW_LOCK_READ) && RW_READERS(v) &&
 920                     spintries < rowner_retries) {
 921                         if (!(v & RW_LOCK_WRITE_SPINNER)) {
 922                                 if (!atomic_cmpset_ptr(&rw->rw_lock, v,
 923                                     v | RW_LOCK_WRITE_SPINNER)) {
 924                                         v = RW_READ_VALUE(rw);
 925                                         continue;
 926                                 }
 927                         }
 928                         spintries++;
 929                         KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
 930                             "spinning", "lockname:\"%s\"",
 931                             rw->lock_object.lo_name);
 932                         for (i = 0; i < rowner_loops; i += n) {
 933                                 n = RW_READERS(v);
 934                                 lock_delay_spin(n);
 935                                 v = RW_READ_VALUE(rw);
 936                                 if ((v & RW_LOCK_WRITE_SPINNER) == 0)
 937                                         break;
 938                         }
 939                         KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
 940                             "running");
 941 #ifdef KDTRACE_HOOKS
 942                         lda.spin_cnt += rowner_loops - i;
 943 #endif
 944                         if (i != rowner_loops)
 945                                 continue;
 946                 }
 947 #endif
 948                 ts = turnstile_trywait(&rw->lock_object);
 949                 v = RW_READ_VALUE(rw);
 950
 951 #ifdef ADAPTIVE_RWLOCKS
 952                 /*
 953                  * The current lock owner might have started executing
 954                  * on another CPU (or the lock could have changed
 955                  * owners) while we were waiting on the turnstile
 956                  * chain lock.  If so, drop the turnstile lock and try
 957                  * again.
 958                  */
 959                 if (!(v & RW_LOCK_READ)) {
 960                         owner = (struct thread *)RW_OWNER(v);
 961                         if (TD_IS_RUNNING(owner)) {
 962                                 turnstile_cancel(ts);
 963                                 continue;
 964                         }
 965                 }
 966 #endif
 967                 /*
 968                  * Check for the waiters flags about this rwlock.
 969                  * If the lock was released, without maintain any pending
 970                  * waiters queue, simply try to acquire it.
 971                  * If a pending waiters queue is present, claim the lock
 972                  * ownership and maintain the pending queue.
 973                  */
 974                 x = v & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER);
 975                 if ((v & ~x) == RW_UNLOCKED) {
 976                         x &= ~RW_LOCK_WRITE_SPINNER;
 977                         if (atomic_cmpset_acq_ptr(&rw->rw_lock, v, tid | x)) {
 978                                 if (x)
 979                                         turnstile_claim(ts);
 980                                 else
 981                                         turnstile_cancel(ts);
 982                                 break;
 983                         }
 984                         turnstile_cancel(ts);
 985                         v = RW_READ_VALUE(rw);
 986                         continue;
 987                 }
 988                 /*
 989                  * If the RW_LOCK_WRITE_WAITERS flag isn't set, then try to
 990                  * set it.  If we fail to set it, then loop back and try
 991                  * again.
 992                  */
 993                 if (!(v & RW_LOCK_WRITE_WAITERS)) {
 994                         if (!atomic_cmpset_ptr(&rw->rw_lock, v,
 995                             v | RW_LOCK_WRITE_WAITERS)) {
 996                                 turnstile_cancel(ts);
 997                                 v = RW_READ_VALUE(rw);
 998                                 continue;
 999                         }
1000                         if (LOCK_LOG_TEST(&rw->lock_object, 0))
1001                                 CTR2(KTR_LOCK, "%s: %p set write waiters flag",
1002                                     __func__, rw);
1003                 }
1004                 /*
1005                  * We were unable to acquire the lock and the write waiters
1006                  * flag is set, so we must block on the turnstile.
1007                  */
1008                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
1009                         CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__,
1010                             rw);
1011 #ifdef KDTRACE_HOOKS
1012                 sleep_time -= lockstat_nsecs(&rw->lock_object);
1013 #endif
1014                 turnstile_wait(ts, rw_owner(rw), TS_EXCLUSIVE_QUEUE);
1015 #ifdef KDTRACE_HOOKS
1016                 sleep_time += lockstat_nsecs(&rw->lock_object);
1017                 sleep_cnt++;
1018 #endif
1019                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
1020                         CTR2(KTR_LOCK, "%s: %p resuming from turnstile",
1021                             __func__, rw);
1022 #ifdef ADAPTIVE_RWLOCKS
1023                 spintries = 0;
1024 #endif
1025                 v = RW_READ_VALUE(rw);
1026         }
1027 #ifdef KDTRACE_HOOKS
1028         all_time += lockstat_nsecs(&rw->lock_object);
1029         if (sleep_time)
1030                 LOCKSTAT_RECORD4(rw__block, rw, sleep_time,
1031                     LOCKSTAT_WRITER, (state & RW_LOCK_READ) == 0,
1032                     (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
1033
1034         /* Record only the loops spinning and not sleeping. */
1035         if (lda.spin_cnt > sleep_cnt)
1036                 LOCKSTAT_RECORD4(rw__spin, rw, all_time - sleep_time,
1037                     LOCKSTAT_WRITER, (state & RW_LOCK_READ) == 0,
1038                     (state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
1039 #endif
1040         LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire, rw, contested,
1041             waittime, file, line, LOCKSTAT_WRITER);
1042 }
1043
1044 /*
1045  * This function is called if lockstat is active or the first try at releasing
1046  * a write lock failed.  The latter means that the lock is recursed or one of
1047  * the 2 waiter bits must be set indicating that at least one thread is waiting
1048  * on this lock.
1049  */
1050 void
1051 __rw_wunlock_hard(volatile uintptr_t *c, uintptr_t tid, const char *file,
1052     int line)
1053 {
1054         struct rwlock *rw;
1055         struct turnstile *ts;
1056         uintptr_t v;
1057         int queue;
1058
1059         if (SCHEDULER_STOPPED())
1060                 return;
1061
1062         rw = rwlock2rw(c);
1063         v = RW_READ_VALUE(rw);
1064         if (v & RW_LOCK_WRITER_RECURSED) {
1065                 if (--(rw->rw_recurse) == 0)
1066                         atomic_clear_ptr(&rw->rw_lock, RW_LOCK_WRITER_RECURSED);
1067                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
1068                         CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, rw);
1069                 return;
1070         }
1071
1072         LOCKSTAT_PROFILE_RELEASE_RWLOCK(rw__release, rw, LOCKSTAT_WRITER);
1073         if (v == tid && _rw_write_unlock(rw, tid))
1074                 return;
1075
1076         KASSERT(rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS),
1077             ("%s: neither of the waiter flags are set", __func__));
1078
1079         if (LOCK_LOG_TEST(&rw->lock_object, 0))
1080                 CTR2(KTR_LOCK, "%s: %p contested", __func__, rw);
1081
1082         turnstile_chain_lock(&rw->lock_object);
1083         ts = turnstile_lookup(&rw->lock_object);
1084         MPASS(ts != NULL);
1085
1086         /*
1087          * Use the same algo as sx locks for now.  Prefer waking up shared
1088          * waiters if we have any over writers.  This is probably not ideal.
1089          *
1090          * 'v' is the value we are going to write back to rw_lock.  If we
1091          * have waiters on both queues, we need to preserve the state of
1092          * the waiter flag for the queue we don't wake up.  For now this is
1093          * hardcoded for the algorithm mentioned above.
1094          *
1095          * In the case of both readers and writers waiting we wakeup the
1096          * readers but leave the RW_LOCK_WRITE_WAITERS flag set.  If a
1097          * new writer comes in before a reader it will claim the lock up
1098          * above.  There is probably a potential priority inversion in
1099          * there that could be worked around either by waking both queues
1100          * of waiters or doing some complicated lock handoff gymnastics.
1101          */
1102         v = RW_UNLOCKED;
1103         if (rw->rw_lock & RW_LOCK_WRITE_WAITERS) {
1104                 queue = TS_EXCLUSIVE_QUEUE;
1105                 v |= (rw->rw_lock & RW_LOCK_READ_WAITERS);
1106         } else
1107                 queue = TS_SHARED_QUEUE;
1108
1109         /* Wake up all waiters for the specific queue. */
1110         if (LOCK_LOG_TEST(&rw->lock_object, 0))
1111                 CTR3(KTR_LOCK, "%s: %p waking up %s waiters", __func__, rw,
1112                     queue == TS_SHARED_QUEUE ? "read" : "write");
1113         turnstile_broadcast(ts, queue);
1114         atomic_store_rel_ptr(&rw->rw_lock, v);
1115         turnstile_unpend(ts, TS_EXCLUSIVE_LOCK);
1116         turnstile_chain_unlock(&rw->lock_object);
1117 }
1118
1119 /*
1120  * Attempt to do a non-blocking upgrade from a read lock to a write
1121  * lock.  This will only succeed if this thread holds a single read
1122  * lock.  Returns true if the upgrade succeeded and false otherwise.
1123  */
1124 int
1125 __rw_try_upgrade(volatile uintptr_t *c, const char *file, int line)
1126 {
1127         struct rwlock *rw;
1128         uintptr_t v, x, tid;
1129         struct turnstile *ts;
1130         int success;
1131
1132         if (SCHEDULER_STOPPED())
1133                 return (1);
1134
1135         rw = rwlock2rw(c);
1136
1137         KASSERT(rw->rw_lock != RW_DESTROYED,
1138             ("rw_try_upgrade() of destroyed rwlock @ %s:%d", file, line));
1139         __rw_assert(c, RA_RLOCKED, file, line);
1140
1141         /*
1142          * Attempt to switch from one reader to a writer.  If there
1143          * are any write waiters, then we will have to lock the
1144          * turnstile first to prevent races with another writer
1145          * calling turnstile_wait() before we have claimed this
1146          * turnstile.  So, do the simple case of no waiters first.
1147          */
1148         tid = (uintptr_t)curthread;
1149         success = 0;
1150         for (;;) {
1151                 v = rw->rw_lock;
1152                 if (RW_READERS(v) > 1)
1153                         break;
1154                 if (!(v & RW_LOCK_WAITERS)) {
1155                         success = atomic_cmpset_acq_ptr(&rw->rw_lock, v, tid);
1156                         if (!success)
1157                                 continue;
1158                         break;
1159                 }
1160
1161                 /*
1162                  * Ok, we think we have waiters, so lock the turnstile.
1163                  */
1164                 ts = turnstile_trywait(&rw->lock_object);
1165                 v = rw->rw_lock;
1166                 if (RW_READERS(v) > 1) {
1167                         turnstile_cancel(ts);
1168                         break;
1169                 }
1170                 /*
1171                  * Try to switch from one reader to a writer again.  This time
1172                  * we honor the current state of the waiters flags.
1173                  * If we obtain the lock with the flags set, then claim
1174                  * ownership of the turnstile.
1175                  */
1176                 x = rw->rw_lock & RW_LOCK_WAITERS;
1177                 success = atomic_cmpset_ptr(&rw->rw_lock, v, tid | x);
1178                 if (success) {
1179                         if (x)
1180                                 turnstile_claim(ts);
1181                         else
1182                                 turnstile_cancel(ts);
1183                         break;
1184                 }
1185                 turnstile_cancel(ts);
1186         }
1187         LOCK_LOG_TRY("WUPGRADE", &rw->lock_object, 0, success, file, line);
1188         if (success) {
1189                 curthread->td_rw_rlocks--;
1190                 WITNESS_UPGRADE(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
1191                     file, line);
1192                 LOCKSTAT_RECORD0(rw__upgrade, rw);
1193         }
1194         return (success);
1195 }
1196
1197 /*
1198  * Downgrade a write lock into a single read lock.
1199  */
1200 void
1201 __rw_downgrade(volatile uintptr_t *c, const char *file, int line)
1202 {
1203         struct rwlock *rw;
1204         struct turnstile *ts;
1205         uintptr_t tid, v;
1206         int rwait, wwait;
1207
1208         if (SCHEDULER_STOPPED())
1209                 return;
1210
1211         rw = rwlock2rw(c);
1212
1213         KASSERT(rw->rw_lock != RW_DESTROYED,
1214             ("rw_downgrade() of destroyed rwlock @ %s:%d", file, line));
1215         __rw_assert(c, RA_WLOCKED | RA_NOTRECURSED, file, line);
1216 #ifndef INVARIANTS
1217         if (rw_recursed(rw))
1218                 panic("downgrade of a recursed lock");
1219 #endif
1220
1221         WITNESS_DOWNGRADE(&rw->lock_object, 0, file, line);
1222
1223         /*
1224          * Convert from a writer to a single reader.  First we handle
1225          * the easy case with no waiters.  If there are any waiters, we
1226          * lock the turnstile and "disown" the lock.
1227          */
1228         tid = (uintptr_t)curthread;
1229         if (atomic_cmpset_rel_ptr(&rw->rw_lock, tid, RW_READERS_LOCK(1)))
1230                 goto out;
1231
1232         /*
1233          * Ok, we think we have waiters, so lock the turnstile so we can
1234          * read the waiter flags without any races.
1235          */
1236         turnstile_chain_lock(&rw->lock_object);
1237         v = rw->rw_lock & RW_LOCK_WAITERS;
1238         rwait = v & RW_LOCK_READ_WAITERS;
1239         wwait = v & RW_LOCK_WRITE_WAITERS;
1240         MPASS(rwait | wwait);
1241
1242         /*
1243          * Downgrade from a write lock while preserving waiters flag
1244          * and give up ownership of the turnstile.
1245          */
1246         ts = turnstile_lookup(&rw->lock_object);
1247         MPASS(ts != NULL);
1248         if (!wwait)
1249                 v &= ~RW_LOCK_READ_WAITERS;
1250         atomic_store_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v);
1251         /*
1252          * Wake other readers if there are no writers pending.  Otherwise they
1253          * won't be able to acquire the lock anyway.
1254          */
1255         if (rwait && !wwait) {
1256                 turnstile_broadcast(ts, TS_SHARED_QUEUE);
1257                 turnstile_unpend(ts, TS_EXCLUSIVE_LOCK);
1258         } else
1259                 turnstile_disown(ts);
1260         turnstile_chain_unlock(&rw->lock_object);
1261 out:
1262         curthread->td_rw_rlocks++;
1263         LOCK_LOG_LOCK("WDOWNGRADE", &rw->lock_object, 0, 0, file, line);
1264         LOCKSTAT_RECORD0(rw__downgrade, rw);
1265 }
1266
1267 #ifdef INVARIANT_SUPPORT
1268 #ifndef INVARIANTS
1269 #undef __rw_assert
1270 #endif
1271
1272 /*
1273  * In the non-WITNESS case, rw_assert() can only detect that at least
1274  * *some* thread owns an rlock, but it cannot guarantee that *this*
1275  * thread owns an rlock.
1276  */
1277 void
1278 __rw_assert(const volatile uintptr_t *c, int what, const char *file, int line)
1279 {
1280         const struct rwlock *rw;
1281
1282         if (panicstr != NULL)
1283                 return;
1284
1285         rw = rwlock2rw(c);
1286
1287         switch (what) {
1288         case RA_LOCKED:
1289         case RA_LOCKED | RA_RECURSED:
1290         case RA_LOCKED | RA_NOTRECURSED:
1291         case RA_RLOCKED:
1292         case RA_RLOCKED | RA_RECURSED:
1293         case RA_RLOCKED | RA_NOTRECURSED:
1294 #ifdef WITNESS
1295                 witness_assert(&rw->lock_object, what, file, line);
1296 #else
1297                 /*
1298                  * If some other thread has a write lock or we have one
1299                  * and are asserting a read lock, fail.  Also, if no one
1300                  * has a lock at all, fail.
1301                  */
1302                 if (rw->rw_lock == RW_UNLOCKED ||
1303                     (!(rw->rw_lock & RW_LOCK_READ) && (what & RA_RLOCKED ||
1304                     rw_wowner(rw) != curthread)))
1305                         panic("Lock %s not %slocked @ %s:%d\n",
1306                             rw->lock_object.lo_name, (what & RA_RLOCKED) ?
1307                             "read " : "", file, line);
1308
1309                 if (!(rw->rw_lock & RW_LOCK_READ) && !(what & RA_RLOCKED)) {
1310                         if (rw_recursed(rw)) {
1311                                 if (what & RA_NOTRECURSED)
1312                                         panic("Lock %s recursed @ %s:%d\n",
1313                                             rw->lock_object.lo_name, file,
1314                                             line);
1315                         } else if (what & RA_RECURSED)
1316                                 panic("Lock %s not recursed @ %s:%d\n",
1317                                     rw->lock_object.lo_name, file, line);
1318                 }
1319 #endif
1320                 break;
1321         case RA_WLOCKED:
1322         case RA_WLOCKED | RA_RECURSED:
1323         case RA_WLOCKED | RA_NOTRECURSED:
1324                 if (rw_wowner(rw) != curthread)
1325                         panic("Lock %s not exclusively locked @ %s:%d\n",
1326                             rw->lock_object.lo_name, file, line);
1327                 if (rw_recursed(rw)) {
1328                         if (what & RA_NOTRECURSED)
1329                                 panic("Lock %s recursed @ %s:%d\n",
1330                                     rw->lock_object.lo_name, file, line);
1331                 } else if (what & RA_RECURSED)
1332                         panic("Lock %s not recursed @ %s:%d\n",
1333                             rw->lock_object.lo_name, file, line);
1334                 break;
1335         case RA_UNLOCKED:
1336 #ifdef WITNESS
1337                 witness_assert(&rw->lock_object, what, file, line);
1338 #else
1339                 /*
1340                  * If we hold a write lock fail.  We can't reliably check
1341                  * to see if we hold a read lock or not.
1342                  */
1343                 if (rw_wowner(rw) == curthread)
1344                         panic("Lock %s exclusively locked @ %s:%d\n",
1345                             rw->lock_object.lo_name, file, line);
1346 #endif
1347                 break;
1348         default:
1349                 panic("Unknown rw lock assertion: %d @ %s:%d", what, file,
1350                     line);
1351         }
1352 }
1353 #endif /* INVARIANT_SUPPORT */
1354
1355 #ifdef DDB
1356 void
1357 db_show_rwlock(const struct lock_object *lock)
1358 {
1359         const struct rwlock *rw;
1360         struct thread *td;
1361
1362         rw = (const struct rwlock *)lock;
1363
1364         db_printf(" state: ");
1365         if (rw->rw_lock == RW_UNLOCKED)
1366                 db_printf("UNLOCKED\n");
1367         else if (rw->rw_lock == RW_DESTROYED) {
1368                 db_printf("DESTROYED\n");
1369                 return;
1370         } else if (rw->rw_lock & RW_LOCK_READ)
1371                 db_printf("RLOCK: %ju locks\n",
1372                     (uintmax_t)(RW_READERS(rw->rw_lock)));
1373         else {
1374                 td = rw_wowner(rw);
1375                 db_printf("WLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1376                     td->td_tid, td->td_proc->p_pid, td->td_name);
1377                 if (rw_recursed(rw))
1378                         db_printf(" recursed: %u\n", rw->rw_recurse);
1379         }
1380         db_printf(" waiters: ");
1381         switch (rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS)) {
1382         case RW_LOCK_READ_WAITERS:
1383                 db_printf("readers\n");
1384                 break;
1385         case RW_LOCK_WRITE_WAITERS:
1386                 db_printf("writers\n");
1387                 break;
1388         case RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS:
1389                 db_printf("readers and writers\n");
1390                 break;
1391         default:
1392                 db_printf("none\n");
1393                 break;
1394         }
1395 }
1396
1397 #endif