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