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