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  * 3. Neither the name of the author nor the names of any co-contributors
  14  *    may be used to endorse or promote products derived from this software
  15  *    without specific prior written permission.
  16  *
  17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
  18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  20  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
  21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  27  * SUCH DAMAGE.
  28  */
  29
  30 /*
  31  * Machine independent bits of reader/writer lock implementation.
  32  */
  33
  34 #include <sys/cdefs.h>
  35 __FBSDID("$FreeBSD$");
  36
  37 #include "opt_ddb.h"
  38 #include "opt_hwpmc_hooks.h"
  39 #include "opt_kdtrace.h"
  40 #include "opt_no_adaptive_rwlocks.h"
  41
  42 #include <sys/param.h>
  43 #include <sys/ktr.h>
  44 #include <sys/kernel.h>
  45 #include <sys/lock.h>
  46 #include <sys/mutex.h>
  47 #include <sys/proc.h>
  48 #include <sys/rwlock.h>
  49 #include <sys/sysctl.h>
  50 #include <sys/systm.h>
  51 #include <sys/turnstile.h>
  52
  53 #include <machine/cpu.h>
  54
  55 #if defined(SMP) && !defined(NO_ADAPTIVE_RWLOCKS)
  56 #define ADAPTIVE_RWLOCKS
  57 #endif
  58
  59 #ifdef HWPMC_HOOKS
  60 #include <sys/pmckern.h>
  61 PMC_SOFT_DECLARE( , , lock, failed);
  62 #endif
  63
  64 #ifdef ADAPTIVE_RWLOCKS
  65 static int rowner_retries = 10;
  66 static int rowner_loops = 10000;
  67 static SYSCTL_NODE(_debug, OID_AUTO, rwlock, CTLFLAG_RD, NULL,
  68     "rwlock debugging");
  69 SYSCTL_INT(_debug_rwlock, OID_AUTO, retry, CTLFLAG_RW, &rowner_retries, 0, "");
  70 SYSCTL_INT(_debug_rwlock, OID_AUTO, loops, CTLFLAG_RW, &rowner_loops, 0, "");
  71 #endif
  72
  73 #ifdef DDB
  74 #include <ddb/ddb.h>
  75
  76 static void     db_show_rwlock(const struct lock_object *lock);
  77 #endif
  78 static void     assert_rw(const struct lock_object *lock, int what);
  79 static void     lock_rw(struct lock_object *lock, int how);
  80 #ifdef KDTRACE_HOOKS
  81 static int      owner_rw(const struct lock_object *lock, struct thread **owner);
  82 #endif
  83 static int      unlock_rw(struct lock_object *lock);
  84
  85 struct lock_class lock_class_rw = {
  86         .lc_name = "rw",
  87         .lc_flags = LC_SLEEPLOCK | LC_RECURSABLE | LC_UPGRADABLE,
  88         .lc_assert = assert_rw,
  89 #ifdef DDB
  90         .lc_ddb_show = db_show_rwlock,
  91 #endif
  92         .lc_lock = lock_rw,
  93         .lc_unlock = unlock_rw,
  94 #ifdef KDTRACE_HOOKS
  95         .lc_owner = owner_rw,
  96 #endif
  97 };
  98
  99 /*
 100  * Return a pointer to the owning thread if the lock is write-locked or
 101  * NULL if the lock is unlocked or read-locked.
 102  */
 103 #define rw_wowner(rw)                                                   \
 104         ((rw)->rw_lock & RW_LOCK_READ ? NULL :                          \
 105             (struct thread *)RW_OWNER((rw)->rw_lock))
 106
 107 /*
 108  * Returns if a write owner is recursed.  Write ownership is not assured
 109  * here and should be previously checked.
 110  */
 111 #define rw_recursed(rw)         ((rw)->rw_recurse != 0)
 112
 113 /*
 114  * Return true if curthread helds the lock.
 115  */
 116 #define rw_wlocked(rw)          (rw_wowner((rw)) == curthread)
 117
 118 /*
 119  * Return a pointer to the owning thread for this lock who should receive
 120  * any priority lent by threads that block on this lock.  Currently this
 121  * is identical to rw_wowner().
 122  */
 123 #define rw_owner(rw)            rw_wowner(rw)
 124
 125 #ifndef INVARIANTS
 126 #define _rw_assert(rw, what, file, line)
 127 #endif
 128
 129 void
 130 assert_rw(const struct lock_object *lock, int what)
 131 {
 132
 133         rw_assert((const struct rwlock *)lock, what);
 134 }
 135
 136 void
 137 lock_rw(struct lock_object *lock, int how)
 138 {
 139         struct rwlock *rw;
 140
 141         rw = (struct rwlock *)lock;
 142         if (how)
 143                 rw_wlock(rw);
 144         else
 145                 rw_rlock(rw);
 146 }
 147
 148 int
 149 unlock_rw(struct lock_object *lock)
 150 {
 151         struct rwlock *rw;
 152
 153         rw = (struct rwlock *)lock;
 154         rw_assert(rw, RA_LOCKED | LA_NOTRECURSED);
 155         if (rw->rw_lock & RW_LOCK_READ) {
 156                 rw_runlock(rw);
 157                 return (0);
 158         } else {
 159                 rw_wunlock(rw);
 160                 return (1);
 161         }
 162 }
 163
 164 #ifdef KDTRACE_HOOKS
 165 int
 166 owner_rw(const struct lock_object *lock, struct thread **owner)
 167 {
 168         const struct rwlock *rw = (const struct rwlock *)lock;
 169         uintptr_t x = rw->rw_lock;
 170
 171         *owner = rw_wowner(rw);
 172         return ((x & RW_LOCK_READ) != 0 ?  (RW_READERS(x) != 0) :
 173             (*owner != NULL));
 174 }
 175 #endif
 176
 177 void
 178 rw_init_flags(struct rwlock *rw, const char *name, int opts)
 179 {
 180         int flags;
 181
 182         MPASS((opts & ~(RW_DUPOK | RW_NOPROFILE | RW_NOWITNESS | RW_QUIET |
 183             RW_RECURSE)) == 0);
 184         ASSERT_ATOMIC_LOAD_PTR(rw->rw_lock,
 185             ("%s: rw_lock not aligned for %s: %p", __func__, name,
 186             &rw->rw_lock));
 187
 188         flags = LO_UPGRADABLE;
 189         if (opts & RW_DUPOK)
 190                 flags |= LO_DUPOK;
 191         if (opts & RW_NOPROFILE)
 192                 flags |= LO_NOPROFILE;
 193         if (!(opts & RW_NOWITNESS))
 194                 flags |= LO_WITNESS;
 195         if (opts & RW_RECURSE)
 196                 flags |= LO_RECURSABLE;
 197         if (opts & RW_QUIET)
 198                 flags |= LO_QUIET;
 199
 200         rw->rw_lock = RW_UNLOCKED;
 201         rw->rw_recurse = 0;
 202         lock_init(&rw->lock_object, &lock_class_rw, name, NULL, flags);
 203 }
 204
 205 void
 206 rw_destroy(struct rwlock *rw)
 207 {
 208
 209         KASSERT(rw->rw_lock == RW_UNLOCKED, ("rw lock %p not unlocked", rw));
 210         KASSERT(rw->rw_recurse == 0, ("rw lock %p still recursed", rw));
 211         rw->rw_lock = RW_DESTROYED;
 212         lock_destroy(&rw->lock_object);
 213 }
 214
 215 void
 216 rw_sysinit(void *arg)
 217 {
 218         struct rw_args *args = arg;
 219
 220         rw_init(args->ra_rw, args->ra_desc);
 221 }
 222
 223 void
 224 rw_sysinit_flags(void *arg)
 225 {
 226         struct rw_args_flags *args = arg;
 227
 228         rw_init_flags(args->ra_rw, args->ra_desc, args->ra_flags);
 229 }
 230
 231 int
 232 rw_wowned(const struct rwlock *rw)
 233 {
 234
 235         return (rw_wowner(rw) == curthread);
 236 }
 237
 238 void
 239 _rw_wlock(struct rwlock *rw, const char *file, int line)
 240 {
 241
 242         if (SCHEDULER_STOPPED())
 243                 return;
 244         MPASS(curthread != NULL);
 245         KASSERT(rw->rw_lock != RW_DESTROYED,
 246             ("rw_wlock() of destroyed rwlock @ %s:%d", file, line));
 247         WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE, file,
 248             line, NULL);
 249         __rw_wlock(rw, curthread, file, line);
 250         LOCK_LOG_LOCK("WLOCK", &rw->lock_object, 0, rw->rw_recurse, file, line);
 251         WITNESS_LOCK(&rw->lock_object, LOP_EXCLUSIVE, file, line);
 252         curthread->td_locks++;
 253 }
 254
 255 int
 256 _rw_try_wlock(struct rwlock *rw, const char *file, int line)
 257 {
 258         int rval;
 259
 260         if (SCHEDULER_STOPPED())
 261                 return (1);
 262
 263         KASSERT(rw->rw_lock != RW_DESTROYED,
 264             ("rw_try_wlock() of destroyed rwlock @ %s:%d", file, line));
 265
 266         if (rw_wlocked(rw) &&
 267             (rw->lock_object.lo_flags & LO_RECURSABLE) != 0) {
 268                 rw->rw_recurse++;
 269                 rval = 1;
 270         } else
 271                 rval = atomic_cmpset_acq_ptr(&rw->rw_lock, RW_UNLOCKED,
 272                     (uintptr_t)curthread);
 273
 274         LOCK_LOG_TRY("WLOCK", &rw->lock_object, 0, rval, file, line);
 275         if (rval) {
 276                 WITNESS_LOCK(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
 277                     file, line);
 278                 curthread->td_locks++;
 279         }
 280         return (rval);
 281 }
 282
 283 void
 284 _rw_wunlock(struct rwlock *rw, const char *file, int line)
 285 {
 286
 287         if (SCHEDULER_STOPPED())
 288                 return;
 289         MPASS(curthread != NULL);
 290         KASSERT(rw->rw_lock != RW_DESTROYED,
 291             ("rw_wunlock() of destroyed rwlock @ %s:%d", file, line));
 292         _rw_assert(rw, RA_WLOCKED, file, line);
 293         curthread->td_locks--;
 294         WITNESS_UNLOCK(&rw->lock_object, LOP_EXCLUSIVE, file, line);
 295         LOCK_LOG_LOCK("WUNLOCK", &rw->lock_object, 0, rw->rw_recurse, file,
 296             line);
 297         if (!rw_recursed(rw))
 298                 LOCKSTAT_PROFILE_RELEASE_LOCK(LS_RW_WUNLOCK_RELEASE, rw);
 299         __rw_wunlock(rw, curthread, file, line);
 300 }
 301 /*
 302  * Determines whether a new reader can acquire a lock.  Succeeds if the
 303  * reader already owns a read lock and the lock is locked for read to
 304  * prevent deadlock from reader recursion.  Also succeeds if the lock
 305  * is unlocked and has no writer waiters or spinners.  Failing otherwise
 306  * prioritizes writers before readers.
 307  */
 308 #define RW_CAN_READ(_rw)                                                \
 309     ((curthread->td_rw_rlocks && (_rw) & RW_LOCK_READ) || ((_rw) &      \
 310     (RW_LOCK_READ | RW_LOCK_WRITE_WAITERS | RW_LOCK_WRITE_SPINNER)) ==  \
 311     RW_LOCK_READ)
 312
 313 void
 314 _rw_rlock(struct rwlock *rw, const char *file, int line)
 315 {
 316         struct turnstile *ts;
 317 #ifdef ADAPTIVE_RWLOCKS
 318         volatile struct thread *owner;
 319         int spintries = 0;
 320         int i;
 321 #endif
 322 #ifdef LOCK_PROFILING
 323         uint64_t waittime = 0;
 324         int contested = 0;
 325 #endif
 326         uintptr_t v;
 327 #ifdef KDTRACE_HOOKS
 328         uint64_t spin_cnt = 0;
 329         uint64_t sleep_cnt = 0;
 330         int64_t sleep_time = 0;
 331 #endif
 332
 333         if (SCHEDULER_STOPPED())
 334                 return;
 335
 336         KASSERT(rw->rw_lock != RW_DESTROYED,
 337             ("rw_rlock() of destroyed rwlock @ %s:%d", file, line));
 338         KASSERT(rw_wowner(rw) != curthread,
 339             ("%s (%s): wlock already held @ %s:%d", __func__,
 340             rw->lock_object.lo_name, file, line));
 341         WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER, file, line, NULL);
 342
 343         for (;;) {
 344 #ifdef KDTRACE_HOOKS
 345                 spin_cnt++;
 346 #endif
 347                 /*
 348                  * Handle the easy case.  If no other thread has a write
 349                  * lock, then try to bump up the count of read locks.  Note
 350                  * that we have to preserve the current state of the
 351                  * RW_LOCK_WRITE_WAITERS flag.  If we fail to acquire a
 352                  * read lock, then rw_lock must have changed, so restart
 353                  * the loop.  Note that this handles the case of a
 354                  * completely unlocked rwlock since such a lock is encoded
 355                  * as a read lock with no waiters.
 356                  */
 357                 v = rw->rw_lock;
 358                 if (RW_CAN_READ(v)) {
 359                         /*
 360                          * The RW_LOCK_READ_WAITERS flag should only be set
 361                          * if the lock has been unlocked and write waiters
 362                          * were present.
 363                          */
 364                         if (atomic_cmpset_acq_ptr(&rw->rw_lock, v,
 365                             v + RW_ONE_READER)) {
 366                                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 367                                         CTR4(KTR_LOCK,
 368                                             "%s: %p succeed %p -> %p", __func__,
 369                                             rw, (void *)v,
 370                                             (void *)(v + RW_ONE_READER));
 371                                 break;
 372                         }
 373                         continue;
 374                 }
 375 #ifdef HWPMC_HOOKS
 376                 PMC_SOFT_CALL( , , lock, failed);
 377 #endif
 378                 lock_profile_obtain_lock_failed(&rw->lock_object,
 379                     &contested, &waittime);
 380
 381 #ifdef ADAPTIVE_RWLOCKS
 382                 /*
 383                  * If the owner is running on another CPU, spin until
 384                  * the owner stops running or the state of the lock
 385                  * changes.
 386                  */
 387                 if ((v & RW_LOCK_READ) == 0) {
 388                         owner = (struct thread *)RW_OWNER(v);
 389                         if (TD_IS_RUNNING(owner)) {
 390                                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 391                                         CTR3(KTR_LOCK,
 392                                             "%s: spinning on %p held by %p",
 393                                             __func__, rw, owner);
 394                                 while ((struct thread*)RW_OWNER(rw->rw_lock) ==
 395                                     owner && TD_IS_RUNNING(owner)) {
 396                                         cpu_spinwait();
 397 #ifdef KDTRACE_HOOKS
 398                                         spin_cnt++;
 399 #endif
 400                                 }
 401                                 continue;
 402                         }
 403                 } else if (spintries < rowner_retries) {
 404                         spintries++;
 405                         for (i = 0; i < rowner_loops; i++) {
 406                                 v = rw->rw_lock;
 407                                 if ((v & RW_LOCK_READ) == 0 || RW_CAN_READ(v))
 408                                         break;
 409                                 cpu_spinwait();
 410                         }
 411                         if (i != rowner_loops)
 412                                 continue;
 413                 }
 414 #endif
 415
 416                 /*
 417                  * Okay, now it's the hard case.  Some other thread already
 418                  * has a write lock or there are write waiters present,
 419                  * acquire the turnstile lock so we can begin the process
 420                  * of blocking.
 421                  */
 422                 ts = turnstile_trywait(&rw->lock_object);
 423
 424                 /*
 425                  * The lock might have been released while we spun, so
 426                  * recheck its state and restart the loop if needed.
 427                  */
 428                 v = rw->rw_lock;
 429                 if (RW_CAN_READ(v)) {
 430                         turnstile_cancel(ts);
 431                         continue;
 432                 }
 433
 434 #ifdef ADAPTIVE_RWLOCKS
 435                 /*
 436                  * The current lock owner might have started executing
 437                  * on another CPU (or the lock could have changed
 438                  * owners) while we were waiting on the turnstile
 439                  * chain lock.  If so, drop the turnstile lock and try
 440                  * again.
 441                  */
 442                 if ((v & RW_LOCK_READ) == 0) {
 443                         owner = (struct thread *)RW_OWNER(v);
 444                         if (TD_IS_RUNNING(owner)) {
 445                                 turnstile_cancel(ts);
 446                                 continue;
 447                         }
 448                 }
 449 #endif
 450
 451                 /*
 452                  * The lock is held in write mode or it already has waiters.
 453                  */
 454                 MPASS(!RW_CAN_READ(v));
 455
 456                 /*
 457                  * If the RW_LOCK_READ_WAITERS flag is already set, then
 458                  * we can go ahead and block.  If it is not set then try
 459                  * to set it.  If we fail to set it drop the turnstile
 460                  * lock and restart the loop.
 461                  */
 462                 if (!(v & RW_LOCK_READ_WAITERS)) {
 463                         if (!atomic_cmpset_ptr(&rw->rw_lock, v,
 464                             v | RW_LOCK_READ_WAITERS)) {
 465                                 turnstile_cancel(ts);
 466                                 continue;
 467                         }
 468                         if (LOCK_LOG_TEST(&rw->lock_object, 0))
 469                                 CTR2(KTR_LOCK, "%s: %p set read waiters flag",
 470                                     __func__, rw);
 471                 }
 472
 473                 /*
 474                  * We were unable to acquire the lock and the read waiters
 475                  * flag is set, so we must block on the turnstile.
 476                  */
 477                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 478                         CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__,
 479                             rw);
 480 #ifdef KDTRACE_HOOKS
 481                 sleep_time -= lockstat_nsecs();
 482 #endif
 483                 turnstile_wait(ts, rw_owner(rw), TS_SHARED_QUEUE);
 484 #ifdef KDTRACE_HOOKS
 485                 sleep_time += lockstat_nsecs();
 486                 sleep_cnt++;
 487 #endif
 488                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 489                         CTR2(KTR_LOCK, "%s: %p resuming from turnstile",
 490                             __func__, rw);
 491         }
 492
 493         /*
 494          * TODO: acquire "owner of record" here.  Here be turnstile dragons
 495          * however.  turnstiles don't like owners changing between calls to
 496          * turnstile_wait() currently.
 497          */
 498         LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_RW_RLOCK_ACQUIRE, rw, contested,
 499             waittime, file, line);
 500         LOCK_LOG_LOCK("RLOCK", &rw->lock_object, 0, 0, file, line);
 501         WITNESS_LOCK(&rw->lock_object, 0, file, line);
 502         curthread->td_locks++;
 503         curthread->td_rw_rlocks++;
 504 #ifdef KDTRACE_HOOKS
 505         if (sleep_time)
 506                 LOCKSTAT_RECORD1(LS_RW_RLOCK_BLOCK, rw, sleep_time);
 507
 508         /*
 509          * Record only the loops spinning and not sleeping.
 510          */
 511         if (spin_cnt > sleep_cnt)
 512                 LOCKSTAT_RECORD1(LS_RW_RLOCK_SPIN, rw, (spin_cnt - sleep_cnt));
 513 #endif
 514 }
 515
 516 int
 517 _rw_try_rlock(struct rwlock *rw, const char *file, int line)
 518 {
 519         uintptr_t x;
 520
 521         if (SCHEDULER_STOPPED())
 522                 return (1);
 523
 524         for (;;) {
 525                 x = rw->rw_lock;
 526                 KASSERT(rw->rw_lock != RW_DESTROYED,
 527                     ("rw_try_rlock() of destroyed rwlock @ %s:%d", file, line));
 528                 if (!(x & RW_LOCK_READ))
 529                         break;
 530                 if (atomic_cmpset_acq_ptr(&rw->rw_lock, x, x + RW_ONE_READER)) {
 531                         LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 1, file,
 532                             line);
 533                         WITNESS_LOCK(&rw->lock_object, LOP_TRYLOCK, file, line);
 534                         curthread->td_locks++;
 535                         curthread->td_rw_rlocks++;
 536                         return (1);
 537                 }
 538         }
 539
 540         LOCK_LOG_TRY("RLOCK", &rw->lock_object, 0, 0, file, line);
 541         return (0);
 542 }
 543
 544 void
 545 _rw_runlock(struct rwlock *rw, const char *file, int line)
 546 {
 547         struct turnstile *ts;
 548         uintptr_t x, v, queue;
 549
 550         if (SCHEDULER_STOPPED())
 551                 return;
 552
 553         KASSERT(rw->rw_lock != RW_DESTROYED,
 554             ("rw_runlock() of destroyed rwlock @ %s:%d", file, line));
 555         _rw_assert(rw, RA_RLOCKED, file, line);
 556         curthread->td_locks--;
 557         curthread->td_rw_rlocks--;
 558         WITNESS_UNLOCK(&rw->lock_object, 0, file, line);
 559         LOCK_LOG_LOCK("RUNLOCK", &rw->lock_object, 0, 0, file, line);
 560
 561         /* TODO: drop "owner of record" here. */
 562
 563         for (;;) {
 564                 /*
 565                  * See if there is more than one read lock held.  If so,
 566                  * just drop one and return.
 567                  */
 568                 x = rw->rw_lock;
 569                 if (RW_READERS(x) > 1) {
 570                         if (atomic_cmpset_rel_ptr(&rw->rw_lock, x,
 571                             x - RW_ONE_READER)) {
 572                                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 573                                         CTR4(KTR_LOCK,
 574                                             "%s: %p succeeded %p -> %p",
 575                                             __func__, rw, (void *)x,
 576                                             (void *)(x - RW_ONE_READER));
 577                                 break;
 578                         }
 579                         continue;
 580                 }
 581                 /*
 582                  * If there aren't any waiters for a write lock, then try
 583                  * to drop it quickly.
 584                  */
 585                 if (!(x & RW_LOCK_WAITERS)) {
 586                         MPASS((x & ~RW_LOCK_WRITE_SPINNER) ==
 587                             RW_READERS_LOCK(1));
 588                         if (atomic_cmpset_rel_ptr(&rw->rw_lock, x,
 589                             RW_UNLOCKED)) {
 590                                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 591                                         CTR2(KTR_LOCK, "%s: %p last succeeded",
 592                                             __func__, rw);
 593                                 break;
 594                         }
 595                         continue;
 596                 }
 597                 /*
 598                  * Ok, we know we have waiters and we think we are the
 599                  * last reader, so grab the turnstile lock.
 600                  */
 601                 turnstile_chain_lock(&rw->lock_object);
 602                 v = rw->rw_lock & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER);
 603                 MPASS(v & RW_LOCK_WAITERS);
 604
 605                 /*
 606                  * Try to drop our lock leaving the lock in a unlocked
 607                  * state.
 608                  *
 609                  * If you wanted to do explicit lock handoff you'd have to
 610                  * do it here.  You'd also want to use turnstile_signal()
 611                  * and you'd have to handle the race where a higher
 612                  * priority thread blocks on the write lock before the
 613                  * thread you wakeup actually runs and have the new thread
 614                  * "steal" the lock.  For now it's a lot simpler to just
 615                  * wakeup all of the waiters.
 616                  *
 617                  * As above, if we fail, then another thread might have
 618                  * acquired a read lock, so drop the turnstile lock and
 619                  * restart.
 620                  */
 621                 x = RW_UNLOCKED;
 622                 if (v & RW_LOCK_WRITE_WAITERS) {
 623                         queue = TS_EXCLUSIVE_QUEUE;
 624                         x |= (v & RW_LOCK_READ_WAITERS);
 625                 } else
 626                         queue = TS_SHARED_QUEUE;
 627                 if (!atomic_cmpset_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v,
 628                     x)) {
 629                         turnstile_chain_unlock(&rw->lock_object);
 630                         continue;
 631                 }
 632                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 633                         CTR2(KTR_LOCK, "%s: %p last succeeded with waiters",
 634                             __func__, rw);
 635
 636                 /*
 637                  * Ok.  The lock is released and all that's left is to
 638                  * wake up the waiters.  Note that the lock might not be
 639                  * free anymore, but in that case the writers will just
 640                  * block again if they run before the new lock holder(s)
 641                  * release the lock.
 642                  */
 643                 ts = turnstile_lookup(&rw->lock_object);
 644                 MPASS(ts != NULL);
 645                 turnstile_broadcast(ts, queue);
 646                 turnstile_unpend(ts, TS_SHARED_LOCK);
 647                 turnstile_chain_unlock(&rw->lock_object);
 648                 break;
 649         }
 650         LOCKSTAT_PROFILE_RELEASE_LOCK(LS_RW_RUNLOCK_RELEASE, rw);
 651 }
 652
 653 /*
 654  * This function is called when we are unable to obtain a write lock on the
 655  * first try.  This means that at least one other thread holds either a
 656  * read or write lock.
 657  */
 658 void
 659 _rw_wlock_hard(struct rwlock *rw, uintptr_t tid, const char *file, int line)
 660 {
 661         struct turnstile *ts;
 662 #ifdef ADAPTIVE_RWLOCKS
 663         volatile struct thread *owner;
 664         int spintries = 0;
 665         int i;
 666 #endif
 667         uintptr_t v, x;
 668 #ifdef LOCK_PROFILING
 669         uint64_t waittime = 0;
 670         int contested = 0;
 671 #endif
 672 #ifdef KDTRACE_HOOKS
 673         uint64_t spin_cnt = 0;
 674         uint64_t sleep_cnt = 0;
 675         int64_t sleep_time = 0;
 676 #endif
 677
 678         if (SCHEDULER_STOPPED())
 679                 return;
 680
 681         if (rw_wlocked(rw)) {
 682                 KASSERT(rw->lock_object.lo_flags & LO_RECURSABLE,
 683                     ("%s: recursing but non-recursive rw %s @ %s:%d\n",
 684                     __func__, rw->lock_object.lo_name, file, line));
 685                 rw->rw_recurse++;
 686                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 687                         CTR2(KTR_LOCK, "%s: %p recursing", __func__, rw);
 688                 return;
 689         }
 690
 691         if (LOCK_LOG_TEST(&rw->lock_object, 0))
 692                 CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__,
 693                     rw->lock_object.lo_name, (void *)rw->rw_lock, file, line);
 694
 695         while (!_rw_write_lock(rw, tid)) {
 696 #ifdef KDTRACE_HOOKS
 697                 spin_cnt++;
 698 #endif
 699 #ifdef HWPMC_HOOKS
 700                 PMC_SOFT_CALL( , , lock, failed);
 701 #endif
 702                 lock_profile_obtain_lock_failed(&rw->lock_object,
 703                     &contested, &waittime);
 704 #ifdef ADAPTIVE_RWLOCKS
 705                 /*
 706                  * If the lock is write locked and the owner is
 707                  * running on another CPU, spin until the owner stops
 708                  * running or the state of the lock changes.
 709                  */
 710                 v = rw->rw_lock;
 711                 owner = (struct thread *)RW_OWNER(v);
 712                 if (!(v & RW_LOCK_READ) && TD_IS_RUNNING(owner)) {
 713                         if (LOCK_LOG_TEST(&rw->lock_object, 0))
 714                                 CTR3(KTR_LOCK, "%s: spinning on %p held by %p",
 715                                     __func__, rw, owner);
 716                         while ((struct thread*)RW_OWNER(rw->rw_lock) == owner &&
 717                             TD_IS_RUNNING(owner)) {
 718                                 cpu_spinwait();
 719 #ifdef KDTRACE_HOOKS
 720                                 spin_cnt++;
 721 #endif
 722                         }
 723                         continue;
 724                 }
 725                 if ((v & RW_LOCK_READ) && RW_READERS(v) &&
 726                     spintries < rowner_retries) {
 727                         if (!(v & RW_LOCK_WRITE_SPINNER)) {
 728                                 if (!atomic_cmpset_ptr(&rw->rw_lock, v,
 729                                     v | RW_LOCK_WRITE_SPINNER)) {
 730                                         continue;
 731                                 }
 732                         }
 733                         spintries++;
 734                         for (i = 0; i < rowner_loops; i++) {
 735                                 if ((rw->rw_lock & RW_LOCK_WRITE_SPINNER) == 0)
 736                                         break;
 737                                 cpu_spinwait();
 738                         }
 739 #ifdef KDTRACE_HOOKS
 740                         spin_cnt += rowner_loops - i;
 741 #endif
 742                         if (i != rowner_loops)
 743                                 continue;
 744                 }
 745 #endif
 746                 ts = turnstile_trywait(&rw->lock_object);
 747                 v = rw->rw_lock;
 748
 749 #ifdef ADAPTIVE_RWLOCKS
 750                 /*
 751                  * The current lock owner might have started executing
 752                  * on another CPU (or the lock could have changed
 753                  * owners) while we were waiting on the turnstile
 754                  * chain lock.  If so, drop the turnstile lock and try
 755                  * again.
 756                  */
 757                 if (!(v & RW_LOCK_READ)) {
 758                         owner = (struct thread *)RW_OWNER(v);
 759                         if (TD_IS_RUNNING(owner)) {
 760                                 turnstile_cancel(ts);
 761                                 continue;
 762                         }
 763                 }
 764 #endif
 765                 /*
 766                  * Check for the waiters flags about this rwlock.
 767                  * If the lock was released, without maintain any pending
 768                  * waiters queue, simply try to acquire it.
 769                  * If a pending waiters queue is present, claim the lock
 770                  * ownership and maintain the pending queue.
 771                  */
 772                 x = v & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER);
 773                 if ((v & ~x) == RW_UNLOCKED) {
 774                         x &= ~RW_LOCK_WRITE_SPINNER;
 775                         if (atomic_cmpset_acq_ptr(&rw->rw_lock, v, tid | x)) {
 776                                 if (x)
 777                                         turnstile_claim(ts);
 778                                 else
 779                                         turnstile_cancel(ts);
 780                                 break;
 781                         }
 782                         turnstile_cancel(ts);
 783                         continue;
 784                 }
 785                 /*
 786                  * If the RW_LOCK_WRITE_WAITERS flag isn't set, then try to
 787                  * set it.  If we fail to set it, then loop back and try
 788                  * again.
 789                  */
 790                 if (!(v & RW_LOCK_WRITE_WAITERS)) {
 791                         if (!atomic_cmpset_ptr(&rw->rw_lock, v,
 792                             v | RW_LOCK_WRITE_WAITERS)) {
 793                                 turnstile_cancel(ts);
 794                                 continue;
 795                         }
 796                         if (LOCK_LOG_TEST(&rw->lock_object, 0))
 797                                 CTR2(KTR_LOCK, "%s: %p set write waiters flag",
 798                                     __func__, rw);
 799                 }
 800                 /*
 801                  * We were unable to acquire the lock and the write waiters
 802                  * flag is set, so we must block on the turnstile.
 803                  */
 804                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 805                         CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__,
 806                             rw);
 807 #ifdef KDTRACE_HOOKS
 808                 sleep_time -= lockstat_nsecs();
 809 #endif
 810                 turnstile_wait(ts, rw_owner(rw), TS_EXCLUSIVE_QUEUE);
 811 #ifdef KDTRACE_HOOKS
 812                 sleep_time += lockstat_nsecs();
 813                 sleep_cnt++;
 814 #endif
 815                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 816                         CTR2(KTR_LOCK, "%s: %p resuming from turnstile",
 817                             __func__, rw);
 818 #ifdef ADAPTIVE_RWLOCKS
 819                 spintries = 0;
 820 #endif
 821         }
 822         LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_RW_WLOCK_ACQUIRE, rw, contested,
 823             waittime, file, line);
 824 #ifdef KDTRACE_HOOKS
 825         if (sleep_time)
 826                 LOCKSTAT_RECORD1(LS_RW_WLOCK_BLOCK, rw, sleep_time);
 827
 828         /*
 829          * Record only the loops spinning and not sleeping.
 830          */
 831         if (spin_cnt > sleep_cnt)
 832                 LOCKSTAT_RECORD1(LS_RW_WLOCK_SPIN, rw, (spin_cnt - sleep_cnt));
 833 #endif
 834 }
 835
 836 /*
 837  * This function is called if the first try at releasing a write lock failed.
 838  * This means that one of the 2 waiter bits must be set indicating that at
 839  * least one thread is waiting on this lock.
 840  */
 841 void
 842 _rw_wunlock_hard(struct rwlock *rw, uintptr_t tid, const char *file, int line)
 843 {
 844         struct turnstile *ts;
 845         uintptr_t v;
 846         int queue;
 847
 848         if (SCHEDULER_STOPPED())
 849                 return;
 850
 851         if (rw_wlocked(rw) && rw_recursed(rw)) {
 852                 rw->rw_recurse--;
 853                 if (LOCK_LOG_TEST(&rw->lock_object, 0))
 854                         CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, rw);
 855                 return;
 856         }
 857
 858         KASSERT(rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS),
 859             ("%s: neither of the waiter flags are set", __func__));
 860
 861         if (LOCK_LOG_TEST(&rw->lock_object, 0))
 862                 CTR2(KTR_LOCK, "%s: %p contested", __func__, rw);
 863
 864         turnstile_chain_lock(&rw->lock_object);
 865         ts = turnstile_lookup(&rw->lock_object);
 866         MPASS(ts != NULL);
 867
 868         /*
 869          * Use the same algo as sx locks for now.  Prefer waking up shared
 870          * waiters if we have any over writers.  This is probably not ideal.
 871          *
 872          * 'v' is the value we are going to write back to rw_lock.  If we
 873          * have waiters on both queues, we need to preserve the state of
 874          * the waiter flag for the queue we don't wake up.  For now this is
 875          * hardcoded for the algorithm mentioned above.
 876          *
 877          * In the case of both readers and writers waiting we wakeup the
 878          * readers but leave the RW_LOCK_WRITE_WAITERS flag set.  If a
 879          * new writer comes in before a reader it will claim the lock up
 880          * above.  There is probably a potential priority inversion in
 881          * there that could be worked around either by waking both queues
 882          * of waiters or doing some complicated lock handoff gymnastics.
 883          */
 884         v = RW_UNLOCKED;
 885         if (rw->rw_lock & RW_LOCK_WRITE_WAITERS) {
 886                 queue = TS_EXCLUSIVE_QUEUE;
 887                 v |= (rw->rw_lock & RW_LOCK_READ_WAITERS);
 888         } else
 889                 queue = TS_SHARED_QUEUE;
 890
 891         /* Wake up all waiters for the specific queue. */
 892         if (LOCK_LOG_TEST(&rw->lock_object, 0))
 893                 CTR3(KTR_LOCK, "%s: %p waking up %s waiters", __func__, rw,
 894                     queue == TS_SHARED_QUEUE ? "read" : "write");
 895         turnstile_broadcast(ts, queue);
 896         atomic_store_rel_ptr(&rw->rw_lock, v);
 897         turnstile_unpend(ts, TS_EXCLUSIVE_LOCK);
 898         turnstile_chain_unlock(&rw->lock_object);
 899 }
 900
 901 /*
 902  * Attempt to do a non-blocking upgrade from a read lock to a write
 903  * lock.  This will only succeed if this thread holds a single read
 904  * lock.  Returns true if the upgrade succeeded and false otherwise.
 905  */
 906 int
 907 _rw_try_upgrade(struct rwlock *rw, const char *file, int line)
 908 {
 909         uintptr_t v, x, tid;
 910         struct turnstile *ts;
 911         int success;
 912
 913         if (SCHEDULER_STOPPED())
 914                 return (1);
 915
 916         KASSERT(rw->rw_lock != RW_DESTROYED,
 917             ("rw_try_upgrade() of destroyed rwlock @ %s:%d", file, line));
 918         _rw_assert(rw, RA_RLOCKED, file, line);
 919
 920         /*
 921          * Attempt to switch from one reader to a writer.  If there
 922          * are any write waiters, then we will have to lock the
 923          * turnstile first to prevent races with another writer
 924          * calling turnstile_wait() before we have claimed this
 925          * turnstile.  So, do the simple case of no waiters first.
 926          */
 927         tid = (uintptr_t)curthread;
 928         success = 0;
 929         for (;;) {
 930                 v = rw->rw_lock;
 931                 if (RW_READERS(v) > 1)
 932                         break;
 933                 if (!(v & RW_LOCK_WAITERS)) {
 934                         success = atomic_cmpset_ptr(&rw->rw_lock, v, tid);
 935                         if (!success)
 936                                 continue;
 937                         break;
 938                 }
 939
 940                 /*
 941                  * Ok, we think we have waiters, so lock the turnstile.
 942                  */
 943                 ts = turnstile_trywait(&rw->lock_object);
 944                 v = rw->rw_lock;
 945                 if (RW_READERS(v) > 1) {
 946                         turnstile_cancel(ts);
 947                         break;
 948                 }
 949                 /*
 950                  * Try to switch from one reader to a writer again.  This time
 951                  * we honor the current state of the waiters flags.
 952                  * If we obtain the lock with the flags set, then claim
 953                  * ownership of the turnstile.
 954                  */
 955                 x = rw->rw_lock & RW_LOCK_WAITERS;
 956                 success = atomic_cmpset_ptr(&rw->rw_lock, v, tid | x);
 957                 if (success) {
 958                         if (x)
 959                                 turnstile_claim(ts);
 960                         else
 961                                 turnstile_cancel(ts);
 962                         break;
 963                 }
 964                 turnstile_cancel(ts);
 965         }
 966         LOCK_LOG_TRY("WUPGRADE", &rw->lock_object, 0, success, file, line);
 967         if (success) {
 968                 curthread->td_rw_rlocks--;
 969                 WITNESS_UPGRADE(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
 970                     file, line);
 971                 LOCKSTAT_RECORD0(LS_RW_TRYUPGRADE_UPGRADE, rw);
 972         }
 973         return (success);
 974 }
 975
 976 /*
 977  * Downgrade a write lock into a single read lock.
 978  */
 979 void
 980 _rw_downgrade(struct rwlock *rw, const char *file, int line)
 981 {
 982         struct turnstile *ts;
 983         uintptr_t tid, v;
 984         int rwait, wwait;
 985
 986         if (SCHEDULER_STOPPED())
 987                 return;
 988
 989         KASSERT(rw->rw_lock != RW_DESTROYED,
 990             ("rw_downgrade() of destroyed rwlock @ %s:%d", file, line));
 991         _rw_assert(rw, RA_WLOCKED | RA_NOTRECURSED, file, line);
 992 #ifndef INVARIANTS
 993         if (rw_recursed(rw))
 994                 panic("downgrade of a recursed lock");
 995 #endif
 996
 997         WITNESS_DOWNGRADE(&rw->lock_object, 0, file, line);
 998
 999         /*
1000          * Convert from a writer to a single reader.  First we handle
1001          * the easy case with no waiters.  If there are any waiters, we
1002          * lock the turnstile and "disown" the lock.
1003          */
1004         tid = (uintptr_t)curthread;
1005         if (atomic_cmpset_rel_ptr(&rw->rw_lock, tid, RW_READERS_LOCK(1)))
1006                 goto out;
1007
1008         /*
1009          * Ok, we think we have waiters, so lock the turnstile so we can
1010          * read the waiter flags without any races.
1011          */
1012         turnstile_chain_lock(&rw->lock_object);
1013         v = rw->rw_lock & RW_LOCK_WAITERS;
1014         rwait = v & RW_LOCK_READ_WAITERS;
1015         wwait = v & RW_LOCK_WRITE_WAITERS;
1016         MPASS(rwait | wwait);
1017
1018         /*
1019          * Downgrade from a write lock while preserving waiters flag
1020          * and give up ownership of the turnstile.
1021          */
1022         ts = turnstile_lookup(&rw->lock_object);
1023         MPASS(ts != NULL);
1024         if (!wwait)
1025                 v &= ~RW_LOCK_READ_WAITERS;
1026         atomic_store_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v);
1027         /*
1028          * Wake other readers if there are no writers pending.  Otherwise they
1029          * won't be able to acquire the lock anyway.
1030          */
1031         if (rwait && !wwait) {
1032                 turnstile_broadcast(ts, TS_SHARED_QUEUE);
1033                 turnstile_unpend(ts, TS_EXCLUSIVE_LOCK);
1034         } else
1035                 turnstile_disown(ts);
1036         turnstile_chain_unlock(&rw->lock_object);
1037 out:
1038         curthread->td_rw_rlocks++;
1039         LOCK_LOG_LOCK("WDOWNGRADE", &rw->lock_object, 0, 0, file, line);
1040         LOCKSTAT_RECORD0(LS_RW_DOWNGRADE_DOWNGRADE, rw);
1041 }
1042
1043 #ifdef INVARIANT_SUPPORT
1044 #ifndef INVARIANTS
1045 #undef _rw_assert
1046 #endif
1047
1048 /*
1049  * In the non-WITNESS case, rw_assert() can only detect that at least
1050  * *some* thread owns an rlock, but it cannot guarantee that *this*
1051  * thread owns an rlock.
1052  */
1053 void
1054 _rw_assert(const struct rwlock *rw, int what, const char *file, int line)
1055 {
1056
1057         if (panicstr != NULL)
1058                 return;
1059         switch (what) {
1060         case RA_LOCKED:
1061         case RA_LOCKED | RA_RECURSED:
1062         case RA_LOCKED | RA_NOTRECURSED:
1063         case RA_RLOCKED:
1064 #ifdef WITNESS
1065                 witness_assert(&rw->lock_object, what, file, line);
1066 #else
1067                 /*
1068                  * If some other thread has a write lock or we have one
1069                  * and are asserting a read lock, fail.  Also, if no one
1070                  * has a lock at all, fail.
1071                  */
1072                 if (rw->rw_lock == RW_UNLOCKED ||
1073                     (!(rw->rw_lock & RW_LOCK_READ) && (what == RA_RLOCKED ||
1074                     rw_wowner(rw) != curthread)))
1075                         panic("Lock %s not %slocked @ %s:%d\n",
1076                             rw->lock_object.lo_name, (what == RA_RLOCKED) ?
1077                             "read " : "", file, line);
1078
1079                 if (!(rw->rw_lock & RW_LOCK_READ)) {
1080                         if (rw_recursed(rw)) {
1081                                 if (what & RA_NOTRECURSED)
1082                                         panic("Lock %s recursed @ %s:%d\n",
1083                                             rw->lock_object.lo_name, file,
1084                                             line);
1085                         } else if (what & RA_RECURSED)
1086                                 panic("Lock %s not recursed @ %s:%d\n",
1087                                     rw->lock_object.lo_name, file, line);
1088                 }
1089 #endif
1090                 break;
1091         case RA_WLOCKED:
1092         case RA_WLOCKED | RA_RECURSED:
1093         case RA_WLOCKED | RA_NOTRECURSED:
1094                 if (rw_wowner(rw) != curthread)
1095                         panic("Lock %s not exclusively locked @ %s:%d\n",
1096                             rw->lock_object.lo_name, file, line);
1097                 if (rw_recursed(rw)) {
1098                         if (what & RA_NOTRECURSED)
1099                                 panic("Lock %s recursed @ %s:%d\n",
1100                                     rw->lock_object.lo_name, file, line);
1101                 } else if (what & RA_RECURSED)
1102                         panic("Lock %s not recursed @ %s:%d\n",
1103                             rw->lock_object.lo_name, file, line);
1104                 break;
1105         case RA_UNLOCKED:
1106 #ifdef WITNESS
1107                 witness_assert(&rw->lock_object, what, file, line);
1108 #else
1109                 /*
1110                  * If we hold a write lock fail.  We can't reliably check
1111                  * to see if we hold a read lock or not.
1112                  */
1113                 if (rw_wowner(rw) == curthread)
1114                         panic("Lock %s exclusively locked @ %s:%d\n",
1115                             rw->lock_object.lo_name, file, line);
1116 #endif
1117                 break;
1118         default:
1119                 panic("Unknown rw lock assertion: %d @ %s:%d", what, file,
1120                     line);
1121         }
1122 }
1123 #endif /* INVARIANT_SUPPORT */
1124
1125 #ifdef DDB
1126 void
1127 db_show_rwlock(const struct lock_object *lock)
1128 {
1129         const struct rwlock *rw;
1130         struct thread *td;
1131
1132         rw = (const struct rwlock *)lock;
1133
1134         db_printf(" state: ");
1135         if (rw->rw_lock == RW_UNLOCKED)
1136                 db_printf("UNLOCKED\n");
1137         else if (rw->rw_lock == RW_DESTROYED) {
1138                 db_printf("DESTROYED\n");
1139                 return;
1140         } else if (rw->rw_lock & RW_LOCK_READ)
1141                 db_printf("RLOCK: %ju locks\n",
1142                     (uintmax_t)(RW_READERS(rw->rw_lock)));
1143         else {
1144                 td = rw_wowner(rw);
1145                 db_printf("WLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1146                     td->td_tid, td->td_proc->p_pid, td->td_name);
1147                 if (rw_recursed(rw))
1148                         db_printf(" recursed: %u\n", rw->rw_recurse);
1149         }
1150         db_printf(" waiters: ");
1151         switch (rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS)) {
1152         case RW_LOCK_READ_WAITERS:
1153                 db_printf("readers\n");
1154                 break;
1155         case RW_LOCK_WRITE_WAITERS:
1156                 db_printf("writers\n");
1157                 break;
1158         case RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS:
1159                 db_printf("readers and writers\n");
1160                 break;
1161         default:
1162                 db_printf("none\n");
1163                 break;
1164         }
1165 }
1166
1167 #endif