sys/kern/kern_sx.c

   1 /*-
   2  * Copyright (c) 2007 Attilio Rao <attilio@freebsd.org>
   3  * Copyright (c) 2001 Jason Evans <jasone@freebsd.org>
   4  * All rights reserved.
   5  *
   6  * Redistribution and use in source and binary forms, with or without
   7  * modification, are permitted provided that the following conditions
   8  * are met:
   9  * 1. Redistributions of source code must retain the above copyright
  10  *    notice(s), this list of conditions and the following disclaimer as
  11  *    the first lines of this file unmodified other than the possible
  12  *    addition of one or more copyright notices.
  13  * 2. Redistributions in binary form must reproduce the above copyright
  14  *    notice(s), this list of conditions and the following disclaimer in the
  15  *    documentation and/or other materials provided with the distribution.
  16  *
  17  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDER(S) ``AS IS'' AND ANY
  18  * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
  19  * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
  20  * DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT HOLDER(S) BE LIABLE FOR ANY
  21  * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
  22  * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
  23  * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
  24  * 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 SUCH
  27  * DAMAGE.
  28  */
  29
  30 /*
  31  * Shared/exclusive locks.  This implementation attempts to ensure
  32  * deterministic lock granting behavior, so that slocks and xlocks are
  33  * interleaved.
  34  *
  35  * Priority propagation will not generally raise the priority of lock holders,
  36  * so should not be relied upon in combination with sx locks.
  37  */
  38
  39 #include "opt_adaptive_sx.h"
  40 #include "opt_ddb.h"
  41
  42 #include <sys/cdefs.h>
  43 __FBSDID("$FreeBSD$");
  44
  45 #include <sys/param.h>
  46 #include <sys/ktr.h>
  47 #include <sys/lock.h>
  48 #include <sys/mutex.h>
  49 #include <sys/proc.h>
  50 #include <sys/sleepqueue.h>
  51 #include <sys/sx.h>
  52 #include <sys/systm.h>
  53
  54 #ifdef ADAPTIVE_SX
  55 #include <machine/cpu.h>
  56 #endif
  57
  58 #ifdef DDB
  59 #include <ddb/ddb.h>
  60 #endif
  61
  62 #if !defined(SMP) && defined(ADAPTIVE_SX)
  63 #error "You must have SMP to enable the ADAPTIVE_SX option"
  64 #endif
  65
  66 CTASSERT(((SX_ADAPTIVESPIN | SX_RECURSE) & LO_CLASSFLAGS) ==
  67     (SX_ADAPTIVESPIN | SX_RECURSE));
  68
  69 /* Handy macros for sleep queues. */
  70 #define SQ_EXCLUSIVE_QUEUE      0
  71 #define SQ_SHARED_QUEUE         1
  72
  73 /*
  74  * Variations on DROP_GIANT()/PICKUP_GIANT() for use in this file.  We
  75  * drop Giant anytime we have to sleep or if we adaptively spin.
  76  */
  77 #define GIANT_DECLARE                                                   \
  78         int _giantcnt = 0;                                              \
  79         WITNESS_SAVE_DECL(Giant)                                        \
  80
  81 #define GIANT_SAVE() do {                                               \
  82         if (mtx_owned(&Giant)) {                                        \
  83                 WITNESS_SAVE(&Giant.lock_object, Giant);                \
  84                 while (mtx_owned(&Giant)) {                             \
  85                         _giantcnt++;                                    \
  86                         mtx_unlock(&Giant);                             \
  87                 }                                                       \
  88         }                                                               \
  89 } while (0)
  90
  91 #define GIANT_RESTORE() do {                                            \
  92         if (_giantcnt > 0) {                                            \
  93                 mtx_assert(&Giant, MA_NOTOWNED);                        \
  94                 while (_giantcnt--)                                     \
  95                         mtx_lock(&Giant);                               \
  96                 WITNESS_RESTORE(&Giant.lock_object, Giant);             \
  97         }                                                               \
  98 } while (0)
  99
 100 /*
 101  * Returns true if an exclusive lock is recursed.  It assumes
 102  * curthread currently has an exclusive lock.
 103  */
 104 #define sx_recursed(sx)         ((sx)->sx_recurse != 0)
 105
 106 #ifdef DDB
 107 static void     db_show_sx(struct lock_object *lock);
 108 #endif
 109 static void     lock_sx(struct lock_object *lock, int how);
 110 static int      unlock_sx(struct lock_object *lock);
 111
 112 struct lock_class lock_class_sx = {
 113         .lc_name = "sx",
 114         .lc_flags = LC_SLEEPLOCK | LC_SLEEPABLE | LC_RECURSABLE | LC_UPGRADABLE,
 115 #ifdef DDB
 116         .lc_ddb_show = db_show_sx,
 117 #endif
 118         .lc_lock = lock_sx,
 119         .lc_unlock = unlock_sx,
 120 };
 121
 122 #ifndef INVARIANTS
 123 #define _sx_assert(sx, what, file, line)
 124 #endif
 125
 126 void
 127 lock_sx(struct lock_object *lock, int how)
 128 {
 129         struct sx *sx;
 130
 131         sx = (struct sx *)lock;
 132         if (how)
 133                 sx_xlock(sx);
 134         else
 135                 sx_slock(sx);
 136 }
 137
 138 int
 139 unlock_sx(struct lock_object *lock)
 140 {
 141         struct sx *sx;
 142
 143         sx = (struct sx *)lock;
 144         sx_assert(sx, SA_LOCKED | SA_NOTRECURSED);
 145         if (sx_xlocked(sx)) {
 146                 sx_xunlock(sx);
 147                 return (1);
 148         } else {
 149                 sx_sunlock(sx);
 150                 return (0);
 151         }
 152 }
 153
 154 void
 155 sx_sysinit(void *arg)
 156 {
 157         struct sx_args *sargs = arg;
 158
 159         sx_init(sargs->sa_sx, sargs->sa_desc);
 160 }
 161
 162 void
 163 sx_init_flags(struct sx *sx, const char *description, int opts)
 164 {
 165         int flags;
 166
 167         MPASS((opts & ~(SX_QUIET | SX_RECURSE | SX_NOWITNESS | SX_DUPOK |
 168             SX_NOPROFILE | SX_ADAPTIVESPIN)) == 0);
 169
 170         flags = LO_RECURSABLE | LO_SLEEPABLE | LO_UPGRADABLE;
 171         if (opts & SX_DUPOK)
 172                 flags |= LO_DUPOK;
 173         if (opts & SX_NOPROFILE)
 174                 flags |= LO_NOPROFILE;
 175         if (!(opts & SX_NOWITNESS))
 176                 flags |= LO_WITNESS;
 177         if (opts & SX_QUIET)
 178                 flags |= LO_QUIET;
 179
 180         flags |= opts & (SX_ADAPTIVESPIN | SX_RECURSE);
 181         sx->sx_lock = SX_LOCK_UNLOCKED;
 182         sx->sx_recurse = 0;
 183         lock_init(&sx->lock_object, &lock_class_sx, description, NULL, flags);
 184 }
 185
 186 void
 187 sx_destroy(struct sx *sx)
 188 {
 189
 190         KASSERT(sx->sx_lock == SX_LOCK_UNLOCKED, ("sx lock still held"));
 191         KASSERT(sx->sx_recurse == 0, ("sx lock still recursed"));
 192         sx->sx_lock = SX_LOCK_DESTROYED;
 193         lock_destroy(&sx->lock_object);
 194 }
 195
 196 int
 197 _sx_slock(struct sx *sx, int opts, const char *file, int line)
 198 {
 199         int error = 0;
 200
 201         MPASS(curthread != NULL);
 202         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 203             ("sx_slock() of destroyed sx @ %s:%d", file, line));
 204         WITNESS_CHECKORDER(&sx->lock_object, LOP_NEWORDER, file, line);
 205         error = __sx_slock(sx, opts, file, line);
 206         if (!error) {
 207                 LOCK_LOG_LOCK("SLOCK", &sx->lock_object, 0, 0, file, line);
 208                 WITNESS_LOCK(&sx->lock_object, 0, file, line);
 209                 curthread->td_locks++;
 210         }
 211
 212         return (error);
 213 }
 214
 215 int
 216 _sx_try_slock(struct sx *sx, const char *file, int line)
 217 {
 218         uintptr_t x;
 219
 220         x = sx->sx_lock;
 221         KASSERT(x != SX_LOCK_DESTROYED,
 222             ("sx_try_slock() of destroyed sx @ %s:%d", file, line));
 223         if ((x & SX_LOCK_SHARED) && atomic_cmpset_acq_ptr(&sx->sx_lock, x,
 224             x + SX_ONE_SHARER)) {
 225                 LOCK_LOG_TRY("SLOCK", &sx->lock_object, 0, 1, file, line);
 226                 WITNESS_LOCK(&sx->lock_object, LOP_TRYLOCK, file, line);
 227                 curthread->td_locks++;
 228                 return (1);
 229         }
 230
 231         LOCK_LOG_TRY("SLOCK", &sx->lock_object, 0, 0, file, line);
 232         return (0);
 233 }
 234
 235 int
 236 _sx_xlock(struct sx *sx, int opts, const char *file, int line)
 237 {
 238         int error = 0;
 239
 240         MPASS(curthread != NULL);
 241         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 242             ("sx_xlock() of destroyed sx @ %s:%d", file, line));
 243         WITNESS_CHECKORDER(&sx->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE, file,
 244             line);
 245         error = __sx_xlock(sx, curthread, opts, file, line);
 246         if (!error) {
 247                 LOCK_LOG_LOCK("XLOCK", &sx->lock_object, 0, sx->sx_recurse,
 248                     file, line);
 249                 WITNESS_LOCK(&sx->lock_object, LOP_EXCLUSIVE, file, line);
 250                 curthread->td_locks++;
 251         }
 252
 253         return (error);
 254 }
 255
 256 int
 257 _sx_try_xlock(struct sx *sx, const char *file, int line)
 258 {
 259         int rval;
 260
 261         MPASS(curthread != NULL);
 262         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 263             ("sx_try_xlock() of destroyed sx @ %s:%d", file, line));
 264
 265         if (sx_xlocked(sx) && (sx->lock_object.lo_flags & SX_RECURSE) != 0) {
 266                 sx->sx_recurse++;
 267                 atomic_set_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
 268                 rval = 1;
 269         } else
 270                 rval = atomic_cmpset_acq_ptr(&sx->sx_lock, SX_LOCK_UNLOCKED,
 271                     (uintptr_t)curthread);
 272         LOCK_LOG_TRY("XLOCK", &sx->lock_object, 0, rval, file, line);
 273         if (rval) {
 274                 WITNESS_LOCK(&sx->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
 275                     file, line);
 276                 curthread->td_locks++;
 277         }
 278
 279         return (rval);
 280 }
 281
 282 void
 283 _sx_sunlock(struct sx *sx, const char *file, int line)
 284 {
 285
 286         MPASS(curthread != NULL);
 287         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 288             ("sx_sunlock() of destroyed sx @ %s:%d", file, line));
 289         _sx_assert(sx, SA_SLOCKED, file, line);
 290         curthread->td_locks--;
 291         WITNESS_UNLOCK(&sx->lock_object, 0, file, line);
 292         LOCK_LOG_LOCK("SUNLOCK", &sx->lock_object, 0, 0, file, line);
 293 #ifdef LOCK_PROFILING_SHARED
 294         if (SX_SHARERS(sx->sx_lock) == 1)
 295                 lock_profile_release_lock(&sx->lock_object);
 296 #endif
 297         __sx_sunlock(sx, file, line);
 298 }
 299
 300 void
 301 _sx_xunlock(struct sx *sx, const char *file, int line)
 302 {
 303
 304         MPASS(curthread != NULL);
 305         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 306             ("sx_xunlock() of destroyed sx @ %s:%d", file, line));
 307         _sx_assert(sx, SA_XLOCKED, file, line);
 308         curthread->td_locks--;
 309         WITNESS_UNLOCK(&sx->lock_object, LOP_EXCLUSIVE, file, line);
 310         LOCK_LOG_LOCK("XUNLOCK", &sx->lock_object, 0, sx->sx_recurse, file,
 311             line);
 312         if (!sx_recursed(sx))
 313                 lock_profile_release_lock(&sx->lock_object);
 314         __sx_xunlock(sx, curthread, file, line);
 315 }
 316
 317 /*
 318  * Try to do a non-blocking upgrade from a shared lock to an exclusive lock.
 319  * This will only succeed if this thread holds a single shared lock.
 320  * Return 1 if if the upgrade succeed, 0 otherwise.
 321  */
 322 int
 323 _sx_try_upgrade(struct sx *sx, const char *file, int line)
 324 {
 325         uintptr_t x;
 326         int success;
 327
 328         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 329             ("sx_try_upgrade() of destroyed sx @ %s:%d", file, line));
 330         _sx_assert(sx, SA_SLOCKED, file, line);
 331
 332         /*
 333          * Try to switch from one shared lock to an exclusive lock.  We need
 334          * to maintain the SX_LOCK_EXCLUSIVE_WAITERS flag if set so that
 335          * we will wake up the exclusive waiters when we drop the lock.
 336          */
 337         x = sx->sx_lock & SX_LOCK_EXCLUSIVE_WAITERS;
 338         success = atomic_cmpset_ptr(&sx->sx_lock, SX_SHARERS_LOCK(1) | x,
 339             (uintptr_t)curthread | x);
 340         LOCK_LOG_TRY("XUPGRADE", &sx->lock_object, 0, success, file, line);
 341         if (success)
 342                 WITNESS_UPGRADE(&sx->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
 343                     file, line);
 344         return (success);
 345 }
 346
 347 /*
 348  * Downgrade an unrecursed exclusive lock into a single shared lock.
 349  */
 350 void
 351 _sx_downgrade(struct sx *sx, const char *file, int line)
 352 {
 353         uintptr_t x;
 354
 355         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 356             ("sx_downgrade() of destroyed sx @ %s:%d", file, line));
 357         _sx_assert(sx, SA_XLOCKED | SA_NOTRECURSED, file, line);
 358 #ifndef INVARIANTS
 359         if (sx_recursed(sx))
 360                 panic("downgrade of a recursed lock");
 361 #endif
 362
 363         WITNESS_DOWNGRADE(&sx->lock_object, 0, file, line);
 364
 365         /*
 366          * Try to switch from an exclusive lock with no shared waiters
 367          * to one sharer with no shared waiters.  If there are
 368          * exclusive waiters, we don't need to lock the sleep queue so
 369          * long as we preserve the flag.  We do one quick try and if
 370          * that fails we grab the sleepq lock to keep the flags from
 371          * changing and do it the slow way.
 372          *
 373          * We have to lock the sleep queue if there are shared waiters
 374          * so we can wake them up.
 375          */
 376         x = sx->sx_lock;
 377         if (!(x & SX_LOCK_SHARED_WAITERS) &&
 378             atomic_cmpset_rel_ptr(&sx->sx_lock, x, SX_SHARERS_LOCK(1) |
 379             (x & SX_LOCK_EXCLUSIVE_WAITERS))) {
 380                 LOCK_LOG_LOCK("XDOWNGRADE", &sx->lock_object, 0, 0, file, line);
 381                 return;
 382         }
 383
 384         /*
 385          * Lock the sleep queue so we can read the waiters bits
 386          * without any races and wakeup any shared waiters.
 387          */
 388         sleepq_lock(&sx->lock_object);
 389
 390         /*
 391          * Preserve SX_LOCK_EXCLUSIVE_WAITERS while downgraded to a single
 392          * shared lock.  If there are any shared waiters, wake them up.
 393          */
 394         x = sx->sx_lock;
 395         atomic_store_rel_ptr(&sx->sx_lock, SX_SHARERS_LOCK(1) |
 396             (x & SX_LOCK_EXCLUSIVE_WAITERS));
 397         if (x & SX_LOCK_SHARED_WAITERS)
 398                 sleepq_broadcast(&sx->lock_object, SLEEPQ_SX, -1,
 399                     SQ_SHARED_QUEUE);
 400         else
 401                 sleepq_release(&sx->lock_object);
 402
 403         LOCK_LOG_LOCK("XDOWNGRADE", &sx->lock_object, 0, 0, file, line);
 404 }
 405
 406 /*
 407  * This function represents the so-called 'hard case' for sx_xlock
 408  * operation.  All 'easy case' failures are redirected to this.  Note
 409  * that ideally this would be a static function, but it needs to be
 410  * accessible from at least sx.h.
 411  */
 412 int
 413 _sx_xlock_hard(struct sx *sx, uintptr_t tid, int opts, const char *file,
 414     int line)
 415 {
 416         GIANT_DECLARE;
 417 #ifdef ADAPTIVE_SX
 418         volatile struct thread *owner;
 419 #endif
 420         uint64_t waittime = 0;
 421         uintptr_t x;
 422         int contested = 0, error = 0;
 423
 424         /* If we already hold an exclusive lock, then recurse. */
 425         if (sx_xlocked(sx)) {
 426                 KASSERT((sx->lock_object.lo_flags & SX_RECURSE) != 0,
 427             ("_sx_xlock_hard: recursed on non-recursive sx %s @ %s:%d\n",
 428                     sx->lock_object.lo_name, file, line));
 429                 sx->sx_recurse++;
 430                 atomic_set_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
 431                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 432                         CTR2(KTR_LOCK, "%s: %p recursing", __func__, sx);
 433                 return (0);
 434         }
 435
 436         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 437                 CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__,
 438                     sx->lock_object.lo_name, (void *)sx->sx_lock, file, line);
 439
 440         while (!atomic_cmpset_acq_ptr(&sx->sx_lock, SX_LOCK_UNLOCKED, tid)) {
 441 #ifdef ADAPTIVE_SX
 442                 /*
 443                  * If the lock is write locked and the owner is
 444                  * running on another CPU, spin until the owner stops
 445                  * running or the state of the lock changes.
 446                  */
 447                 x = sx->sx_lock;
 448                 if (!(x & SX_LOCK_SHARED) &&
 449                     (sx->lock_object.lo_flags & SX_ADAPTIVESPIN)) {
 450                         x = SX_OWNER(x);
 451                         owner = (struct thread *)x;
 452                         if (TD_IS_RUNNING(owner)) {
 453                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 454                                         CTR3(KTR_LOCK,
 455                                             "%s: spinning on %p held by %p",
 456                                             __func__, sx, owner);
 457                                 GIANT_SAVE();
 458                                 lock_profile_obtain_lock_failed(
 459                                     &sx->lock_object, &contested, &waittime);
 460                                 while (SX_OWNER(sx->sx_lock) == x &&
 461                                     TD_IS_RUNNING(owner))
 462                                         cpu_spinwait();
 463                                 continue;
 464                         }
 465                 }
 466 #endif
 467
 468                 sleepq_lock(&sx->lock_object);
 469                 x = sx->sx_lock;
 470
 471                 /*
 472                  * If the lock was released while spinning on the
 473                  * sleep queue chain lock, try again.
 474                  */
 475                 if (x == SX_LOCK_UNLOCKED) {
 476                         sleepq_release(&sx->lock_object);
 477                         continue;
 478                 }
 479
 480 #ifdef ADAPTIVE_SX
 481                 /*
 482                  * The current lock owner might have started executing
 483                  * on another CPU (or the lock could have changed
 484                  * owners) while we were waiting on the sleep queue
 485                  * chain lock.  If so, drop the sleep queue lock and try
 486                  * again.
 487                  */
 488                 if (!(x & SX_LOCK_SHARED) &&
 489                     (sx->lock_object.lo_flags & SX_ADAPTIVESPIN)) {
 490                         owner = (struct thread *)SX_OWNER(x);
 491                         if (TD_IS_RUNNING(owner)) {
 492                                 sleepq_release(&sx->lock_object);
 493                                 continue;
 494                         }
 495                 }
 496 #endif
 497
 498                 /*
 499                  * If an exclusive lock was released with both shared
 500                  * and exclusive waiters and a shared waiter hasn't
 501                  * woken up and acquired the lock yet, sx_lock will be
 502                  * set to SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS.
 503                  * If we see that value, try to acquire it once.  Note
 504                  * that we have to preserve SX_LOCK_EXCLUSIVE_WAITERS
 505                  * as there are other exclusive waiters still.  If we
 506                  * fail, restart the loop.
 507                  */
 508                 if (x == (SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS)) {
 509                         if (atomic_cmpset_acq_ptr(&sx->sx_lock,
 510                             SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS,
 511                             tid | SX_LOCK_EXCLUSIVE_WAITERS)) {
 512                                 sleepq_release(&sx->lock_object);
 513                                 CTR2(KTR_LOCK, "%s: %p claimed by new writer",
 514                                     __func__, sx);
 515                                 break;
 516                         }
 517                         sleepq_release(&sx->lock_object);
 518                         continue;
 519                 }
 520
 521                 /*
 522                  * Try to set the SX_LOCK_EXCLUSIVE_WAITERS.  If we fail,
 523                  * than loop back and retry.
 524                  */
 525                 if (!(x & SX_LOCK_EXCLUSIVE_WAITERS)) {
 526                         if (!atomic_cmpset_ptr(&sx->sx_lock, x,
 527                             x | SX_LOCK_EXCLUSIVE_WAITERS)) {
 528                                 sleepq_release(&sx->lock_object);
 529                                 continue;
 530                         }
 531                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 532                                 CTR2(KTR_LOCK, "%s: %p set excl waiters flag",
 533                                     __func__, sx);
 534                 }
 535
 536                 /*
 537                  * Since we have been unable to acquire the exclusive
 538                  * lock and the exclusive waiters flag is set, we have
 539                  * to sleep.
 540                  */
 541                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 542                         CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
 543                             __func__, sx);
 544
 545                 GIANT_SAVE();
 546                 lock_profile_obtain_lock_failed(&sx->lock_object, &contested,
 547                     &waittime);
 548                 sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
 549                     SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
 550                     SLEEPQ_INTERRUPTIBLE : 0), SQ_EXCLUSIVE_QUEUE);
 551                 if (!(opts & SX_INTERRUPTIBLE))
 552                         sleepq_wait(&sx->lock_object);
 553                 else
 554                         error = sleepq_wait_sig(&sx->lock_object);
 555
 556                 if (error) {
 557                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 558                                 CTR2(KTR_LOCK,
 559                         "%s: interruptible sleep by %p suspended by signal",
 560                                     __func__, sx);
 561                         break;
 562                 }
 563                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 564                         CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
 565                             __func__, sx);
 566         }
 567
 568         GIANT_RESTORE();
 569         if (!error)
 570                 lock_profile_obtain_lock_success(&sx->lock_object, contested,
 571                     waittime, file, line);
 572         return (error);
 573 }
 574
 575 /*
 576  * This function represents the so-called 'hard case' for sx_xunlock
 577  * operation.  All 'easy case' failures are redirected to this.  Note
 578  * that ideally this would be a static function, but it needs to be
 579  * accessible from at least sx.h.
 580  */
 581 void
 582 _sx_xunlock_hard(struct sx *sx, uintptr_t tid, const char *file, int line)
 583 {
 584         uintptr_t x;
 585         int queue;
 586
 587         MPASS(!(sx->sx_lock & SX_LOCK_SHARED));
 588
 589         /* If the lock is recursed, then unrecurse one level. */
 590         if (sx_xlocked(sx) && sx_recursed(sx)) {
 591                 if ((--sx->sx_recurse) == 0)
 592                         atomic_clear_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
 593                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 594                         CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, sx);
 595                 return;
 596         }
 597         MPASS(sx->sx_lock & (SX_LOCK_SHARED_WAITERS |
 598             SX_LOCK_EXCLUSIVE_WAITERS));
 599         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 600                 CTR2(KTR_LOCK, "%s: %p contested", __func__, sx);
 601
 602         sleepq_lock(&sx->lock_object);
 603         x = SX_LOCK_UNLOCKED;
 604
 605         /*
 606          * The wake up algorithm here is quite simple and probably not
 607          * ideal.  It gives precedence to shared waiters if they are
 608          * present.  For this condition, we have to preserve the
 609          * state of the exclusive waiters flag.
 610          */
 611         if (sx->sx_lock & SX_LOCK_SHARED_WAITERS) {
 612                 queue = SQ_SHARED_QUEUE;
 613                 x |= (sx->sx_lock & SX_LOCK_EXCLUSIVE_WAITERS);
 614         } else
 615                 queue = SQ_EXCLUSIVE_QUEUE;
 616
 617         /* Wake up all the waiters for the specific queue. */
 618         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 619                 CTR3(KTR_LOCK, "%s: %p waking up all threads on %s queue",
 620                     __func__, sx, queue == SQ_SHARED_QUEUE ? "shared" :
 621                     "exclusive");
 622         atomic_store_rel_ptr(&sx->sx_lock, x);
 623         sleepq_broadcast(&sx->lock_object, SLEEPQ_SX, -1, queue);
 624 }
 625
 626 /*
 627  * This function represents the so-called 'hard case' for sx_slock
 628  * operation.  All 'easy case' failures are redirected to this.  Note
 629  * that ideally this would be a static function, but it needs to be
 630  * accessible from at least sx.h.
 631  */
 632 int
 633 _sx_slock_hard(struct sx *sx, int opts, const char *file, int line)
 634 {
 635         GIANT_DECLARE;
 636 #ifdef ADAPTIVE_SX
 637         volatile struct thread *owner;
 638 #endif
 639 #ifdef LOCK_PROFILING_SHARED
 640         uint64_t waittime = 0;
 641         int contested = 0;
 642 #endif
 643         uintptr_t x;
 644         int error = 0;
 645
 646         /*
 647          * As with rwlocks, we don't make any attempt to try to block
 648          * shared locks once there is an exclusive waiter.
 649          */
 650         for (;;) {
 651                 x = sx->sx_lock;
 652
 653                 /*
 654                  * If no other thread has an exclusive lock then try to bump up
 655                  * the count of sharers.  Since we have to preserve the state
 656                  * of SX_LOCK_EXCLUSIVE_WAITERS, if we fail to acquire the
 657                  * shared lock loop back and retry.
 658                  */
 659                 if (x & SX_LOCK_SHARED) {
 660                         MPASS(!(x & SX_LOCK_SHARED_WAITERS));
 661                         if (atomic_cmpset_acq_ptr(&sx->sx_lock, x,
 662                             x + SX_ONE_SHARER)) {
 663 #ifdef LOCK_PROFILING_SHARED
 664                                 if (SX_SHARERS(x) == 0)
 665                                         lock_profile_obtain_lock_success(
 666                                             &sx->lock_object, contested,
 667                                             waittime, file, line);
 668 #endif
 669                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 670                                         CTR4(KTR_LOCK,
 671                                             "%s: %p succeed %p -> %p", __func__,
 672                                             sx, (void *)x,
 673                                             (void *)(x + SX_ONE_SHARER));
 674                                 break;
 675                         }
 676                         continue;
 677                 }
 678
 679 #ifdef ADAPTIVE_SX
 680                 /*
 681                  * If the owner is running on another CPU, spin until
 682                  * the owner stops running or the state of the lock
 683                  * changes.
 684                  */
 685                 else if (sx->lock_object.lo_flags & SX_ADAPTIVESPIN) {
 686                         x = SX_OWNER(x);
 687                         owner = (struct thread *)x;
 688                         if (TD_IS_RUNNING(owner)) {
 689                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 690                                         CTR3(KTR_LOCK,
 691                                             "%s: spinning on %p held by %p",
 692                                             __func__, sx, owner);
 693                                 GIANT_SAVE();
 694 #ifdef LOCK_PROFILING_SHARED
 695                                 lock_profile_obtain_lock_failed(
 696                                     &sx->lock_object, &contested, &waittime);
 697 #endif
 698                                 while (SX_OWNER(sx->sx_lock) == x &&
 699                                     TD_IS_RUNNING(owner))
 700                                         cpu_spinwait();
 701                                 continue;
 702                         }
 703                 }
 704 #endif
 705
 706                 /*
 707                  * Some other thread already has an exclusive lock, so
 708                  * start the process of blocking.
 709                  */
 710                 sleepq_lock(&sx->lock_object);
 711                 x = sx->sx_lock;
 712
 713                 /*
 714                  * The lock could have been released while we spun.
 715                  * In this case loop back and retry.
 716                  */
 717                 if (x & SX_LOCK_SHARED) {
 718                         sleepq_release(&sx->lock_object);
 719                         continue;
 720                 }
 721
 722 #ifdef ADAPTIVE_SX
 723                 /*
 724                  * If the owner is running on another CPU, spin until
 725                  * the owner stops running or the state of the lock
 726                  * changes.
 727                  */
 728                 if (!(x & SX_LOCK_SHARED) &&
 729                     (sx->lock_object.lo_flags & SX_ADAPTIVESPIN)) {
 730                         owner = (struct thread *)SX_OWNER(x);
 731                         if (TD_IS_RUNNING(owner)) {
 732                                 sleepq_release(&sx->lock_object);
 733                                 continue;
 734                         }
 735                 }
 736 #endif
 737
 738                 /*
 739                  * Try to set the SX_LOCK_SHARED_WAITERS flag.  If we
 740                  * fail to set it drop the sleep queue lock and loop
 741                  * back.
 742                  */
 743                 if (!(x & SX_LOCK_SHARED_WAITERS)) {
 744                         if (!atomic_cmpset_ptr(&sx->sx_lock, x,
 745                             x | SX_LOCK_SHARED_WAITERS)) {
 746                                 sleepq_release(&sx->lock_object);
 747                                 continue;
 748                         }
 749                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 750                                 CTR2(KTR_LOCK, "%s: %p set shared waiters flag",
 751                                     __func__, sx);
 752                 }
 753
 754                 /*
 755                  * Since we have been unable to acquire the shared lock,
 756                  * we have to sleep.
 757                  */
 758                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 759                         CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
 760                             __func__, sx);
 761
 762                 GIANT_SAVE();
 763 #ifdef LOCK_PROFILING_SHARED
 764                 lock_profile_obtain_lock_failed(&sx->lock_object, &contested,
 765                     &waittime);
 766 #endif
 767                 sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
 768                     SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
 769                     SLEEPQ_INTERRUPTIBLE : 0), SQ_SHARED_QUEUE);
 770                 if (!(opts & SX_INTERRUPTIBLE))
 771                         sleepq_wait(&sx->lock_object);
 772                 else
 773                         error = sleepq_wait_sig(&sx->lock_object);
 774
 775                 if (error) {
 776                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 777                                 CTR2(KTR_LOCK,
 778                         "%s: interruptible sleep by %p suspended by signal",
 779                                     __func__, sx);
 780                         break;
 781                 }
 782                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 783                         CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
 784                             __func__, sx);
 785         }
 786
 787         GIANT_RESTORE();
 788         return (error);
 789 }
 790
 791 /*
 792  * This function represents the so-called 'hard case' for sx_sunlock
 793  * operation.  All 'easy case' failures are redirected to this.  Note
 794  * that ideally this would be a static function, but it needs to be
 795  * accessible from at least sx.h.
 796  */
 797 void
 798 _sx_sunlock_hard(struct sx *sx, const char *file, int line)
 799 {
 800         uintptr_t x;
 801
 802         for (;;) {
 803                 x = sx->sx_lock;
 804
 805                 /*
 806                  * We should never have sharers while at least one thread
 807                  * holds a shared lock.
 808                  */
 809                 KASSERT(!(x & SX_LOCK_SHARED_WAITERS),
 810                     ("%s: waiting sharers", __func__));
 811
 812                 /*
 813                  * See if there is more than one shared lock held.  If
 814                  * so, just drop one and return.
 815                  */
 816                 if (SX_SHARERS(x) > 1) {
 817                         if (atomic_cmpset_ptr(&sx->sx_lock, x,
 818                             x - SX_ONE_SHARER)) {
 819                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 820                                         CTR4(KTR_LOCK,
 821                                             "%s: %p succeeded %p -> %p",
 822                                             __func__, sx, (void *)x,
 823                                             (void *)(x - SX_ONE_SHARER));
 824                                 break;
 825                         }
 826                         continue;
 827                 }
 828
 829                 /*
 830                  * If there aren't any waiters for an exclusive lock,
 831                  * then try to drop it quickly.
 832                  */
 833                 if (!(x & SX_LOCK_EXCLUSIVE_WAITERS)) {
 834                         MPASS(x == SX_SHARERS_LOCK(1));
 835                         if (atomic_cmpset_ptr(&sx->sx_lock, SX_SHARERS_LOCK(1),
 836                             SX_LOCK_UNLOCKED)) {
 837                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 838                                         CTR2(KTR_LOCK, "%s: %p last succeeded",
 839                                             __func__, sx);
 840                                 break;
 841                         }
 842                         continue;
 843                 }
 844
 845                 /*
 846                  * At this point, there should just be one sharer with
 847                  * exclusive waiters.
 848                  */
 849                 MPASS(x == (SX_SHARERS_LOCK(1) | SX_LOCK_EXCLUSIVE_WAITERS));
 850
 851                 sleepq_lock(&sx->lock_object);
 852
 853                 /*
 854                  * Wake up semantic here is quite simple:
 855                  * Just wake up all the exclusive waiters.
 856                  * Note that the state of the lock could have changed,
 857                  * so if it fails loop back and retry.
 858                  */
 859                 if (!atomic_cmpset_ptr(&sx->sx_lock,
 860                     SX_SHARERS_LOCK(1) | SX_LOCK_EXCLUSIVE_WAITERS,
 861                     SX_LOCK_UNLOCKED)) {
 862                         sleepq_release(&sx->lock_object);
 863                         continue;
 864                 }
 865                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 866                         CTR2(KTR_LOCK, "%s: %p waking up all thread on"
 867                             "exclusive queue", __func__, sx);
 868                 sleepq_broadcast(&sx->lock_object, SLEEPQ_SX, -1,
 869                     SQ_EXCLUSIVE_QUEUE);
 870                 break;
 871         }
 872 }
 873
 874 #ifdef INVARIANT_SUPPORT
 875 #ifndef INVARIANTS
 876 #undef  _sx_assert
 877 #endif
 878
 879 /*
 880  * In the non-WITNESS case, sx_assert() can only detect that at least
 881  * *some* thread owns an slock, but it cannot guarantee that *this*
 882  * thread owns an slock.
 883  */
 884 void
 885 _sx_assert(struct sx *sx, int what, const char *file, int line)
 886 {
 887 #ifndef WITNESS
 888         int slocked = 0;
 889 #endif
 890
 891         if (panicstr != NULL)
 892                 return;
 893         switch (what) {
 894         case SA_SLOCKED:
 895         case SA_SLOCKED | SA_NOTRECURSED:
 896         case SA_SLOCKED | SA_RECURSED:
 897 #ifndef WITNESS
 898                 slocked = 1;
 899                 /* FALLTHROUGH */
 900 #endif
 901         case SA_LOCKED:
 902         case SA_LOCKED | SA_NOTRECURSED:
 903         case SA_LOCKED | SA_RECURSED:
 904 #ifdef WITNESS
 905                 witness_assert(&sx->lock_object, what, file, line);
 906 #else
 907                 /*
 908                  * If some other thread has an exclusive lock or we
 909                  * have one and are asserting a shared lock, fail.
 910                  * Also, if no one has a lock at all, fail.
 911                  */
 912                 if (sx->sx_lock == SX_LOCK_UNLOCKED ||
 913                     (!(sx->sx_lock & SX_LOCK_SHARED) && (slocked ||
 914                     sx_xholder(sx) != curthread)))
 915                         panic("Lock %s not %slocked @ %s:%d\n",
 916                             sx->lock_object.lo_name, slocked ? "share " : "",
 917                             file, line);
 918
 919                 if (!(sx->sx_lock & SX_LOCK_SHARED)) {
 920                         if (sx_recursed(sx)) {
 921                                 if (what & SA_NOTRECURSED)
 922                                         panic("Lock %s recursed @ %s:%d\n",
 923                                             sx->lock_object.lo_name, file,
 924                                             line);
 925                         } else if (what & SA_RECURSED)
 926                                 panic("Lock %s not recursed @ %s:%d\n",
 927                                     sx->lock_object.lo_name, file, line);
 928                 }
 929 #endif
 930                 break;
 931         case SA_XLOCKED:
 932         case SA_XLOCKED | SA_NOTRECURSED:
 933         case SA_XLOCKED | SA_RECURSED:
 934                 if (sx_xholder(sx) != curthread)
 935                         panic("Lock %s not exclusively locked @ %s:%d\n",
 936                             sx->lock_object.lo_name, file, line);
 937                 if (sx_recursed(sx)) {
 938                         if (what & SA_NOTRECURSED)
 939                                 panic("Lock %s recursed @ %s:%d\n",
 940                                     sx->lock_object.lo_name, file, line);
 941                 } else if (what & SA_RECURSED)
 942                         panic("Lock %s not recursed @ %s:%d\n",
 943                             sx->lock_object.lo_name, file, line);
 944                 break;
 945         case SA_UNLOCKED:
 946 #ifdef WITNESS
 947                 witness_assert(&sx->lock_object, what, file, line);
 948 #else
 949                 /*
 950                  * If we hold an exclusve lock fail.  We can't
 951                  * reliably check to see if we hold a shared lock or
 952                  * not.
 953                  */
 954                 if (sx_xholder(sx) == curthread)
 955                         panic("Lock %s exclusively locked @ %s:%d\n",
 956                             sx->lock_object.lo_name, file, line);
 957 #endif
 958                 break;
 959         default:
 960                 panic("Unknown sx lock assertion: %d @ %s:%d", what, file,
 961                     line);
 962         }
 963 }
 964 #endif  /* INVARIANT_SUPPORT */
 965
 966 #ifdef DDB
 967 static void
 968 db_show_sx(struct lock_object *lock)
 969 {
 970         struct thread *td;
 971         struct sx *sx;
 972
 973         sx = (struct sx *)lock;
 974
 975         db_printf(" state: ");
 976         if (sx->sx_lock == SX_LOCK_UNLOCKED)
 977                 db_printf("UNLOCKED\n");
 978         else if (sx->sx_lock == SX_LOCK_DESTROYED) {
 979                 db_printf("DESTROYED\n");
 980                 return;
 981         } else if (sx->sx_lock & SX_LOCK_SHARED)
 982                 db_printf("SLOCK: %ju\n", (uintmax_t)SX_SHARERS(sx->sx_lock));
 983         else {
 984                 td = sx_xholder(sx);
 985                 db_printf("XLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
 986                     td->td_tid, td->td_proc->p_pid, td->td_proc->p_comm);
 987                 if (sx_recursed(sx))
 988                         db_printf(" recursed: %d\n", sx->sx_recurse);
 989         }
 990
 991         db_printf(" waiters: ");
 992         switch(sx->sx_lock &
 993             (SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS)) {
 994         case SX_LOCK_SHARED_WAITERS:
 995                 db_printf("shared\n");
 996                 break;
 997         case SX_LOCK_EXCLUSIVE_WAITERS:
 998                 db_printf("exclusive\n");
 999                 break;
1000         case SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS:
1001                 db_printf("exclusive and shared\n");
1002                 break;
1003         default:
1004                 db_printf("none\n");
1005         }
1006 }
1007
1008 /*
1009  * Check to see if a thread that is blocked on a sleep queue is actually
1010  * blocked on an sx lock.  If so, output some details and return true.
1011  * If the lock has an exclusive owner, return that in *ownerp.
1012  */
1013 int
1014 sx_chain(struct thread *td, struct thread **ownerp)
1015 {
1016         struct sx *sx;
1017
1018         /*
1019          * Check to see if this thread is blocked on an sx lock.
1020          * First, we check the lock class.  If that is ok, then we
1021          * compare the lock name against the wait message.
1022          */
1023         sx = td->td_wchan;
1024         if (LOCK_CLASS(&sx->lock_object) != &lock_class_sx ||
1025             sx->lock_object.lo_name != td->td_wmesg)
1026                 return (0);
1027
1028         /* We think we have an sx lock, so output some details. */
1029         db_printf("blocked on sx \"%s\" ", td->td_wmesg);
1030         *ownerp = sx_xholder(sx);
1031         if (sx->sx_lock & SX_LOCK_SHARED)
1032                 db_printf("SLOCK (count %ju)\n",
1033                     (uintmax_t)SX_SHARERS(sx->sx_lock));
1034         else
1035                 db_printf("XLOCK\n");
1036         return (1);
1037 }
1038 #endif