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 static void     assert_sx(struct lock_object *lock, int what);
 107 #ifdef DDB
 108 static void     db_show_sx(struct lock_object *lock);
 109 #endif
 110 static void     lock_sx(struct lock_object *lock, int how);
 111 static int      unlock_sx(struct lock_object *lock);
 112
 113 struct lock_class lock_class_sx = {
 114         .lc_name = "sx",
 115         .lc_flags = LC_SLEEPLOCK | LC_SLEEPABLE | LC_RECURSABLE | LC_UPGRADABLE,
 116         .lc_assert = assert_sx,
 117 #ifdef DDB
 118         .lc_ddb_show = db_show_sx,
 119 #endif
 120         .lc_lock = lock_sx,
 121         .lc_unlock = unlock_sx,
 122 };
 123
 124 #ifndef INVARIANTS
 125 #define _sx_assert(sx, what, file, line)
 126 #endif
 127
 128 void
 129 assert_sx(struct lock_object *lock, int what)
 130 {
 131
 132         sx_assert((struct sx *)lock, what);
 133 }
 134
 135 void
 136 lock_sx(struct lock_object *lock, int how)
 137 {
 138         struct sx *sx;
 139
 140         sx = (struct sx *)lock;
 141         if (how)
 142                 sx_xlock(sx);
 143         else
 144                 sx_slock(sx);
 145 }
 146
 147 int
 148 unlock_sx(struct lock_object *lock)
 149 {
 150         struct sx *sx;
 151
 152         sx = (struct sx *)lock;
 153         sx_assert(sx, SA_LOCKED | SA_NOTRECURSED);
 154         if (sx_xlocked(sx)) {
 155                 sx_xunlock(sx);
 156                 return (1);
 157         } else {
 158                 sx_sunlock(sx);
 159                 return (0);
 160         }
 161 }
 162
 163 void
 164 sx_sysinit(void *arg)
 165 {
 166         struct sx_args *sargs = arg;
 167
 168         sx_init(sargs->sa_sx, sargs->sa_desc);
 169 }
 170
 171 void
 172 sx_init_flags(struct sx *sx, const char *description, int opts)
 173 {
 174         int flags;
 175
 176         MPASS((opts & ~(SX_QUIET | SX_RECURSE | SX_NOWITNESS | SX_DUPOK |
 177             SX_NOPROFILE | SX_ADAPTIVESPIN)) == 0);
 178
 179         flags = LO_RECURSABLE | LO_SLEEPABLE | LO_UPGRADABLE;
 180         if (opts & SX_DUPOK)
 181                 flags |= LO_DUPOK;
 182         if (opts & SX_NOPROFILE)
 183                 flags |= LO_NOPROFILE;
 184         if (!(opts & SX_NOWITNESS))
 185                 flags |= LO_WITNESS;
 186         if (opts & SX_QUIET)
 187                 flags |= LO_QUIET;
 188
 189         flags |= opts & (SX_ADAPTIVESPIN | SX_RECURSE);
 190         sx->sx_lock = SX_LOCK_UNLOCKED;
 191         sx->sx_recurse = 0;
 192         lock_init(&sx->lock_object, &lock_class_sx, description, NULL, flags);
 193 }
 194
 195 void
 196 sx_destroy(struct sx *sx)
 197 {
 198
 199         KASSERT(sx->sx_lock == SX_LOCK_UNLOCKED, ("sx lock still held"));
 200         KASSERT(sx->sx_recurse == 0, ("sx lock still recursed"));
 201         sx->sx_lock = SX_LOCK_DESTROYED;
 202         lock_destroy(&sx->lock_object);
 203 }
 204
 205 int
 206 _sx_slock(struct sx *sx, int opts, const char *file, int line)
 207 {
 208         int error = 0;
 209
 210         MPASS(curthread != NULL);
 211         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 212             ("sx_slock() of destroyed sx @ %s:%d", file, line));
 213         WITNESS_CHECKORDER(&sx->lock_object, LOP_NEWORDER, file, line);
 214         error = __sx_slock(sx, opts, file, line);
 215         if (!error) {
 216                 LOCK_LOG_LOCK("SLOCK", &sx->lock_object, 0, 0, file, line);
 217                 WITNESS_LOCK(&sx->lock_object, 0, file, line);
 218                 curthread->td_locks++;
 219         }
 220
 221         return (error);
 222 }
 223
 224 int
 225 _sx_try_slock(struct sx *sx, const char *file, int line)
 226 {
 227         uintptr_t x;
 228
 229         for (;;) {
 230                 x = sx->sx_lock;
 231                 KASSERT(x != SX_LOCK_DESTROYED,
 232                     ("sx_try_slock() of destroyed sx @ %s:%d", file, line));
 233                 if (!(x & SX_LOCK_SHARED))
 234                         break;
 235                 if (atomic_cmpset_acq_ptr(&sx->sx_lock, x, x + SX_ONE_SHARER)) {
 236                         LOCK_LOG_TRY("SLOCK", &sx->lock_object, 0, 1, file, line);
 237                         WITNESS_LOCK(&sx->lock_object, LOP_TRYLOCK, file, line);
 238                         curthread->td_locks++;
 239                         return (1);
 240                 }
 241         }
 242
 243         LOCK_LOG_TRY("SLOCK", &sx->lock_object, 0, 0, file, line);
 244         return (0);
 245 }
 246
 247 int
 248 _sx_xlock(struct sx *sx, int opts, const char *file, int line)
 249 {
 250         int error = 0;
 251
 252         MPASS(curthread != NULL);
 253         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 254             ("sx_xlock() of destroyed sx @ %s:%d", file, line));
 255         WITNESS_CHECKORDER(&sx->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE, file,
 256             line);
 257         error = __sx_xlock(sx, curthread, opts, file, line);
 258         if (!error) {
 259                 LOCK_LOG_LOCK("XLOCK", &sx->lock_object, 0, sx->sx_recurse,
 260                     file, line);
 261                 WITNESS_LOCK(&sx->lock_object, LOP_EXCLUSIVE, file, line);
 262                 curthread->td_locks++;
 263         }
 264
 265         return (error);
 266 }
 267
 268 int
 269 _sx_try_xlock(struct sx *sx, const char *file, int line)
 270 {
 271         int rval;
 272
 273         MPASS(curthread != NULL);
 274         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 275             ("sx_try_xlock() of destroyed sx @ %s:%d", file, line));
 276
 277         if (sx_xlocked(sx) && (sx->lock_object.lo_flags & SX_RECURSE) != 0) {
 278                 sx->sx_recurse++;
 279                 atomic_set_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
 280                 rval = 1;
 281         } else
 282                 rval = atomic_cmpset_acq_ptr(&sx->sx_lock, SX_LOCK_UNLOCKED,
 283                     (uintptr_t)curthread);
 284         LOCK_LOG_TRY("XLOCK", &sx->lock_object, 0, rval, file, line);
 285         if (rval) {
 286                 WITNESS_LOCK(&sx->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
 287                     file, line);
 288                 curthread->td_locks++;
 289         }
 290
 291         return (rval);
 292 }
 293
 294 void
 295 _sx_sunlock(struct sx *sx, const char *file, int line)
 296 {
 297
 298         MPASS(curthread != NULL);
 299         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 300             ("sx_sunlock() of destroyed sx @ %s:%d", file, line));
 301         _sx_assert(sx, SA_SLOCKED, file, line);
 302         curthread->td_locks--;
 303         WITNESS_UNLOCK(&sx->lock_object, 0, file, line);
 304         LOCK_LOG_LOCK("SUNLOCK", &sx->lock_object, 0, 0, file, line);
 305 #ifdef LOCK_PROFILING_SHARED
 306         if (SX_SHARERS(sx->sx_lock) == 1)
 307                 lock_profile_release_lock(&sx->lock_object);
 308 #endif
 309         __sx_sunlock(sx, file, line);
 310 }
 311
 312 void
 313 _sx_xunlock(struct sx *sx, const char *file, int line)
 314 {
 315
 316         MPASS(curthread != NULL);
 317         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 318             ("sx_xunlock() of destroyed sx @ %s:%d", file, line));
 319         _sx_assert(sx, SA_XLOCKED, file, line);
 320         curthread->td_locks--;
 321         WITNESS_UNLOCK(&sx->lock_object, LOP_EXCLUSIVE, file, line);
 322         LOCK_LOG_LOCK("XUNLOCK", &sx->lock_object, 0, sx->sx_recurse, file,
 323             line);
 324         if (!sx_recursed(sx))
 325                 lock_profile_release_lock(&sx->lock_object);
 326         __sx_xunlock(sx, curthread, file, line);
 327 }
 328
 329 /*
 330  * Try to do a non-blocking upgrade from a shared lock to an exclusive lock.
 331  * This will only succeed if this thread holds a single shared lock.
 332  * Return 1 if if the upgrade succeed, 0 otherwise.
 333  */
 334 int
 335 _sx_try_upgrade(struct sx *sx, const char *file, int line)
 336 {
 337         uintptr_t x;
 338         int success;
 339
 340         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 341             ("sx_try_upgrade() of destroyed sx @ %s:%d", file, line));
 342         _sx_assert(sx, SA_SLOCKED, file, line);
 343
 344         /*
 345          * Try to switch from one shared lock to an exclusive lock.  We need
 346          * to maintain the SX_LOCK_EXCLUSIVE_WAITERS flag if set so that
 347          * we will wake up the exclusive waiters when we drop the lock.
 348          */
 349         x = sx->sx_lock & SX_LOCK_EXCLUSIVE_WAITERS;
 350         success = atomic_cmpset_ptr(&sx->sx_lock, SX_SHARERS_LOCK(1) | x,
 351             (uintptr_t)curthread | x);
 352         LOCK_LOG_TRY("XUPGRADE", &sx->lock_object, 0, success, file, line);
 353         if (success)
 354                 WITNESS_UPGRADE(&sx->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
 355                     file, line);
 356         return (success);
 357 }
 358
 359 /*
 360  * Downgrade an unrecursed exclusive lock into a single shared lock.
 361  */
 362 void
 363 _sx_downgrade(struct sx *sx, const char *file, int line)
 364 {
 365         uintptr_t x;
 366
 367         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 368             ("sx_downgrade() of destroyed sx @ %s:%d", file, line));
 369         _sx_assert(sx, SA_XLOCKED | SA_NOTRECURSED, file, line);
 370 #ifndef INVARIANTS
 371         if (sx_recursed(sx))
 372                 panic("downgrade of a recursed lock");
 373 #endif
 374
 375         WITNESS_DOWNGRADE(&sx->lock_object, 0, file, line);
 376
 377         /*
 378          * Try to switch from an exclusive lock with no shared waiters
 379          * to one sharer with no shared waiters.  If there are
 380          * exclusive waiters, we don't need to lock the sleep queue so
 381          * long as we preserve the flag.  We do one quick try and if
 382          * that fails we grab the sleepq lock to keep the flags from
 383          * changing and do it the slow way.
 384          *
 385          * We have to lock the sleep queue if there are shared waiters
 386          * so we can wake them up.
 387          */
 388         x = sx->sx_lock;
 389         if (!(x & SX_LOCK_SHARED_WAITERS) &&
 390             atomic_cmpset_rel_ptr(&sx->sx_lock, x, SX_SHARERS_LOCK(1) |
 391             (x & SX_LOCK_EXCLUSIVE_WAITERS))) {
 392                 LOCK_LOG_LOCK("XDOWNGRADE", &sx->lock_object, 0, 0, file, line);
 393                 return;
 394         }
 395
 396         /*
 397          * Lock the sleep queue so we can read the waiters bits
 398          * without any races and wakeup any shared waiters.
 399          */
 400         sleepq_lock(&sx->lock_object);
 401
 402         /*
 403          * Preserve SX_LOCK_EXCLUSIVE_WAITERS while downgraded to a single
 404          * shared lock.  If there are any shared waiters, wake them up.
 405          */
 406         x = sx->sx_lock;
 407         atomic_store_rel_ptr(&sx->sx_lock, SX_SHARERS_LOCK(1) |
 408             (x & SX_LOCK_EXCLUSIVE_WAITERS));
 409         if (x & SX_LOCK_SHARED_WAITERS)
 410                 sleepq_broadcast(&sx->lock_object, SLEEPQ_SX, -1,
 411                     SQ_SHARED_QUEUE);
 412         else
 413                 sleepq_release(&sx->lock_object);
 414
 415         LOCK_LOG_LOCK("XDOWNGRADE", &sx->lock_object, 0, 0, file, line);
 416 }
 417
 418 /*
 419  * This function represents the so-called 'hard case' for sx_xlock
 420  * operation.  All 'easy case' failures are redirected to this.  Note
 421  * that ideally this would be a static function, but it needs to be
 422  * accessible from at least sx.h.
 423  */
 424 int
 425 _sx_xlock_hard(struct sx *sx, uintptr_t tid, int opts, const char *file,
 426     int line)
 427 {
 428         GIANT_DECLARE;
 429 #ifdef ADAPTIVE_SX
 430         volatile struct thread *owner;
 431 #endif
 432         uint64_t waittime = 0;
 433         uintptr_t x;
 434         int contested = 0, error = 0;
 435
 436         /* If we already hold an exclusive lock, then recurse. */
 437         if (sx_xlocked(sx)) {
 438                 KASSERT((sx->lock_object.lo_flags & SX_RECURSE) != 0,
 439             ("_sx_xlock_hard: recursed on non-recursive sx %s @ %s:%d\n",
 440                     sx->lock_object.lo_name, file, line));
 441                 sx->sx_recurse++;
 442                 atomic_set_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
 443                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 444                         CTR2(KTR_LOCK, "%s: %p recursing", __func__, sx);
 445                 return (0);
 446         }
 447
 448         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 449                 CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__,
 450                     sx->lock_object.lo_name, (void *)sx->sx_lock, file, line);
 451
 452         while (!atomic_cmpset_acq_ptr(&sx->sx_lock, SX_LOCK_UNLOCKED, tid)) {
 453 #ifdef ADAPTIVE_SX
 454                 /*
 455                  * If the lock is write locked and the owner is
 456                  * running on another CPU, spin until the owner stops
 457                  * running or the state of the lock changes.
 458                  */
 459                 x = sx->sx_lock;
 460                 if (!(x & SX_LOCK_SHARED) &&
 461                     (sx->lock_object.lo_flags & SX_ADAPTIVESPIN)) {
 462                         x = SX_OWNER(x);
 463                         owner = (struct thread *)x;
 464                         if (TD_IS_RUNNING(owner)) {
 465                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 466                                         CTR3(KTR_LOCK,
 467                                             "%s: spinning on %p held by %p",
 468                                             __func__, sx, owner);
 469                                 GIANT_SAVE();
 470                                 lock_profile_obtain_lock_failed(
 471                                     &sx->lock_object, &contested, &waittime);
 472                                 while (SX_OWNER(sx->sx_lock) == x &&
 473                                     TD_IS_RUNNING(owner))
 474                                         cpu_spinwait();
 475                                 continue;
 476                         }
 477                 }
 478 #endif
 479
 480                 sleepq_lock(&sx->lock_object);
 481                 x = sx->sx_lock;
 482
 483                 /*
 484                  * If the lock was released while spinning on the
 485                  * sleep queue chain lock, try again.
 486                  */
 487                 if (x == SX_LOCK_UNLOCKED) {
 488                         sleepq_release(&sx->lock_object);
 489                         continue;
 490                 }
 491
 492 #ifdef ADAPTIVE_SX
 493                 /*
 494                  * The current lock owner might have started executing
 495                  * on another CPU (or the lock could have changed
 496                  * owners) while we were waiting on the sleep queue
 497                  * chain lock.  If so, drop the sleep queue lock and try
 498                  * again.
 499                  */
 500                 if (!(x & SX_LOCK_SHARED) &&
 501                     (sx->lock_object.lo_flags & SX_ADAPTIVESPIN)) {
 502                         owner = (struct thread *)SX_OWNER(x);
 503                         if (TD_IS_RUNNING(owner)) {
 504                                 sleepq_release(&sx->lock_object);
 505                                 continue;
 506                         }
 507                 }
 508 #endif
 509
 510                 /*
 511                  * If an exclusive lock was released with both shared
 512                  * and exclusive waiters and a shared waiter hasn't
 513                  * woken up and acquired the lock yet, sx_lock will be
 514                  * set to SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS.
 515                  * If we see that value, try to acquire it once.  Note
 516                  * that we have to preserve SX_LOCK_EXCLUSIVE_WAITERS
 517                  * as there are other exclusive waiters still.  If we
 518                  * fail, restart the loop.
 519                  */
 520                 if (x == (SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS)) {
 521                         if (atomic_cmpset_acq_ptr(&sx->sx_lock,
 522                             SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS,
 523                             tid | SX_LOCK_EXCLUSIVE_WAITERS)) {
 524                                 sleepq_release(&sx->lock_object);
 525                                 CTR2(KTR_LOCK, "%s: %p claimed by new writer",
 526                                     __func__, sx);
 527                                 break;
 528                         }
 529                         sleepq_release(&sx->lock_object);
 530                         continue;
 531                 }
 532
 533                 /*
 534                  * Try to set the SX_LOCK_EXCLUSIVE_WAITERS.  If we fail,
 535                  * than loop back and retry.
 536                  */
 537                 if (!(x & SX_LOCK_EXCLUSIVE_WAITERS)) {
 538                         if (!atomic_cmpset_ptr(&sx->sx_lock, x,
 539                             x | SX_LOCK_EXCLUSIVE_WAITERS)) {
 540                                 sleepq_release(&sx->lock_object);
 541                                 continue;
 542                         }
 543                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 544                                 CTR2(KTR_LOCK, "%s: %p set excl waiters flag",
 545                                     __func__, sx);
 546                 }
 547
 548                 /*
 549                  * Since we have been unable to acquire the exclusive
 550                  * lock and the exclusive waiters flag is set, we have
 551                  * to sleep.
 552                  */
 553                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 554                         CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
 555                             __func__, sx);
 556
 557                 GIANT_SAVE();
 558                 lock_profile_obtain_lock_failed(&sx->lock_object, &contested,
 559                     &waittime);
 560                 sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
 561                     SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
 562                     SLEEPQ_INTERRUPTIBLE : 0), SQ_EXCLUSIVE_QUEUE);
 563                 if (!(opts & SX_INTERRUPTIBLE))
 564                         sleepq_wait(&sx->lock_object);
 565                 else
 566                         error = sleepq_wait_sig(&sx->lock_object);
 567
 568                 if (error) {
 569                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 570                                 CTR2(KTR_LOCK,
 571                         "%s: interruptible sleep by %p suspended by signal",
 572                                     __func__, sx);
 573                         break;
 574                 }
 575                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 576                         CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
 577                             __func__, sx);
 578         }
 579
 580         GIANT_RESTORE();
 581         if (!error)
 582                 lock_profile_obtain_lock_success(&sx->lock_object, contested,
 583                     waittime, file, line);
 584         return (error);
 585 }
 586
 587 /*
 588  * This function represents the so-called 'hard case' for sx_xunlock
 589  * operation.  All 'easy case' failures are redirected to this.  Note
 590  * that ideally this would be a static function, but it needs to be
 591  * accessible from at least sx.h.
 592  */
 593 void
 594 _sx_xunlock_hard(struct sx *sx, uintptr_t tid, const char *file, int line)
 595 {
 596         uintptr_t x;
 597         int queue;
 598
 599         MPASS(!(sx->sx_lock & SX_LOCK_SHARED));
 600
 601         /* If the lock is recursed, then unrecurse one level. */
 602         if (sx_xlocked(sx) && sx_recursed(sx)) {
 603                 if ((--sx->sx_recurse) == 0)
 604                         atomic_clear_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
 605                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 606                         CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, sx);
 607                 return;
 608         }
 609         MPASS(sx->sx_lock & (SX_LOCK_SHARED_WAITERS |
 610             SX_LOCK_EXCLUSIVE_WAITERS));
 611         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 612                 CTR2(KTR_LOCK, "%s: %p contested", __func__, sx);
 613
 614         sleepq_lock(&sx->lock_object);
 615         x = SX_LOCK_UNLOCKED;
 616
 617         /*
 618          * The wake up algorithm here is quite simple and probably not
 619          * ideal.  It gives precedence to shared waiters if they are
 620          * present.  For this condition, we have to preserve the
 621          * state of the exclusive waiters flag.
 622          */
 623         if (sx->sx_lock & SX_LOCK_SHARED_WAITERS) {
 624                 queue = SQ_SHARED_QUEUE;
 625                 x |= (sx->sx_lock & SX_LOCK_EXCLUSIVE_WAITERS);
 626         } else
 627                 queue = SQ_EXCLUSIVE_QUEUE;
 628
 629         /* Wake up all the waiters for the specific queue. */
 630         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 631                 CTR3(KTR_LOCK, "%s: %p waking up all threads on %s queue",
 632                     __func__, sx, queue == SQ_SHARED_QUEUE ? "shared" :
 633                     "exclusive");
 634         atomic_store_rel_ptr(&sx->sx_lock, x);
 635         sleepq_broadcast(&sx->lock_object, SLEEPQ_SX, -1, queue);
 636 }
 637
 638 /*
 639  * This function represents the so-called 'hard case' for sx_slock
 640  * operation.  All 'easy case' failures are redirected to this.  Note
 641  * that ideally this would be a static function, but it needs to be
 642  * accessible from at least sx.h.
 643  */
 644 int
 645 _sx_slock_hard(struct sx *sx, int opts, const char *file, int line)
 646 {
 647         GIANT_DECLARE;
 648 #ifdef ADAPTIVE_SX
 649         volatile struct thread *owner;
 650 #endif
 651 #ifdef LOCK_PROFILING_SHARED
 652         uint64_t waittime = 0;
 653         int contested = 0;
 654 #endif
 655         uintptr_t x;
 656         int error = 0;
 657
 658         /*
 659          * As with rwlocks, we don't make any attempt to try to block
 660          * shared locks once there is an exclusive waiter.
 661          */
 662         for (;;) {
 663                 x = sx->sx_lock;
 664
 665                 /*
 666                  * If no other thread has an exclusive lock then try to bump up
 667                  * the count of sharers.  Since we have to preserve the state
 668                  * of SX_LOCK_EXCLUSIVE_WAITERS, if we fail to acquire the
 669                  * shared lock loop back and retry.
 670                  */
 671                 if (x & SX_LOCK_SHARED) {
 672                         MPASS(!(x & SX_LOCK_SHARED_WAITERS));
 673                         if (atomic_cmpset_acq_ptr(&sx->sx_lock, x,
 674                             x + SX_ONE_SHARER)) {
 675 #ifdef LOCK_PROFILING_SHARED
 676                                 if (SX_SHARERS(x) == 0)
 677                                         lock_profile_obtain_lock_success(
 678                                             &sx->lock_object, contested,
 679                                             waittime, file, line);
 680 #endif
 681                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 682                                         CTR4(KTR_LOCK,
 683                                             "%s: %p succeed %p -> %p", __func__,
 684                                             sx, (void *)x,
 685                                             (void *)(x + SX_ONE_SHARER));
 686                                 break;
 687                         }
 688                         continue;
 689                 }
 690
 691 #ifdef ADAPTIVE_SX
 692                 /*
 693                  * If the owner is running on another CPU, spin until
 694                  * the owner stops running or the state of the lock
 695                  * changes.
 696                  */
 697                 else if (sx->lock_object.lo_flags & SX_ADAPTIVESPIN) {
 698                         x = SX_OWNER(x);
 699                         owner = (struct thread *)x;
 700                         if (TD_IS_RUNNING(owner)) {
 701                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 702                                         CTR3(KTR_LOCK,
 703                                             "%s: spinning on %p held by %p",
 704                                             __func__, sx, owner);
 705                                 GIANT_SAVE();
 706 #ifdef LOCK_PROFILING_SHARED
 707                                 lock_profile_obtain_lock_failed(
 708                                     &sx->lock_object, &contested, &waittime);
 709 #endif
 710                                 while (SX_OWNER(sx->sx_lock) == x &&
 711                                     TD_IS_RUNNING(owner))
 712                                         cpu_spinwait();
 713                                 continue;
 714                         }
 715                 }
 716 #endif
 717
 718                 /*
 719                  * Some other thread already has an exclusive lock, so
 720                  * start the process of blocking.
 721                  */
 722                 sleepq_lock(&sx->lock_object);
 723                 x = sx->sx_lock;
 724
 725                 /*
 726                  * The lock could have been released while we spun.
 727                  * In this case loop back and retry.
 728                  */
 729                 if (x & SX_LOCK_SHARED) {
 730                         sleepq_release(&sx->lock_object);
 731                         continue;
 732                 }
 733
 734 #ifdef ADAPTIVE_SX
 735                 /*
 736                  * If the owner is running on another CPU, spin until
 737                  * the owner stops running or the state of the lock
 738                  * changes.
 739                  */
 740                 if (!(x & SX_LOCK_SHARED) &&
 741                     (sx->lock_object.lo_flags & SX_ADAPTIVESPIN)) {
 742                         owner = (struct thread *)SX_OWNER(x);
 743                         if (TD_IS_RUNNING(owner)) {
 744                                 sleepq_release(&sx->lock_object);
 745                                 continue;
 746                         }
 747                 }
 748 #endif
 749
 750                 /*
 751                  * Try to set the SX_LOCK_SHARED_WAITERS flag.  If we
 752                  * fail to set it drop the sleep queue lock and loop
 753                  * back.
 754                  */
 755                 if (!(x & SX_LOCK_SHARED_WAITERS)) {
 756                         if (!atomic_cmpset_ptr(&sx->sx_lock, x,
 757                             x | SX_LOCK_SHARED_WAITERS)) {
 758                                 sleepq_release(&sx->lock_object);
 759                                 continue;
 760                         }
 761                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 762                                 CTR2(KTR_LOCK, "%s: %p set shared waiters flag",
 763                                     __func__, sx);
 764                 }
 765
 766                 /*
 767                  * Since we have been unable to acquire the shared lock,
 768                  * we have to sleep.
 769                  */
 770                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 771                         CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
 772                             __func__, sx);
 773
 774                 GIANT_SAVE();
 775 #ifdef LOCK_PROFILING_SHARED
 776                 lock_profile_obtain_lock_failed(&sx->lock_object, &contested,
 777                     &waittime);
 778 #endif
 779                 sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
 780                     SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
 781                     SLEEPQ_INTERRUPTIBLE : 0), SQ_SHARED_QUEUE);
 782                 if (!(opts & SX_INTERRUPTIBLE))
 783                         sleepq_wait(&sx->lock_object);
 784                 else
 785                         error = sleepq_wait_sig(&sx->lock_object);
 786
 787                 if (error) {
 788                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 789                                 CTR2(KTR_LOCK,
 790                         "%s: interruptible sleep by %p suspended by signal",
 791                                     __func__, sx);
 792                         break;
 793                 }
 794                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 795                         CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
 796                             __func__, sx);
 797         }
 798
 799         GIANT_RESTORE();
 800         return (error);
 801 }
 802
 803 /*
 804  * This function represents the so-called 'hard case' for sx_sunlock
 805  * operation.  All 'easy case' failures are redirected to this.  Note
 806  * that ideally this would be a static function, but it needs to be
 807  * accessible from at least sx.h.
 808  */
 809 void
 810 _sx_sunlock_hard(struct sx *sx, const char *file, int line)
 811 {
 812         uintptr_t x;
 813
 814         for (;;) {
 815                 x = sx->sx_lock;
 816
 817                 /*
 818                  * We should never have sharers while at least one thread
 819                  * holds a shared lock.
 820                  */
 821                 KASSERT(!(x & SX_LOCK_SHARED_WAITERS),
 822                     ("%s: waiting sharers", __func__));
 823
 824                 /*
 825                  * See if there is more than one shared lock held.  If
 826                  * so, just drop one and return.
 827                  */
 828                 if (SX_SHARERS(x) > 1) {
 829                         if (atomic_cmpset_ptr(&sx->sx_lock, x,
 830                             x - SX_ONE_SHARER)) {
 831                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 832                                         CTR4(KTR_LOCK,
 833                                             "%s: %p succeeded %p -> %p",
 834                                             __func__, sx, (void *)x,
 835                                             (void *)(x - SX_ONE_SHARER));
 836                                 break;
 837                         }
 838                         continue;
 839                 }
 840
 841                 /*
 842                  * If there aren't any waiters for an exclusive lock,
 843                  * then try to drop it quickly.
 844                  */
 845                 if (!(x & SX_LOCK_EXCLUSIVE_WAITERS)) {
 846                         MPASS(x == SX_SHARERS_LOCK(1));
 847                         if (atomic_cmpset_ptr(&sx->sx_lock, SX_SHARERS_LOCK(1),
 848                             SX_LOCK_UNLOCKED)) {
 849                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 850                                         CTR2(KTR_LOCK, "%s: %p last succeeded",
 851                                             __func__, sx);
 852                                 break;
 853                         }
 854                         continue;
 855                 }
 856
 857                 /*
 858                  * At this point, there should just be one sharer with
 859                  * exclusive waiters.
 860                  */
 861                 MPASS(x == (SX_SHARERS_LOCK(1) | SX_LOCK_EXCLUSIVE_WAITERS));
 862
 863                 sleepq_lock(&sx->lock_object);
 864
 865                 /*
 866                  * Wake up semantic here is quite simple:
 867                  * Just wake up all the exclusive waiters.
 868                  * Note that the state of the lock could have changed,
 869                  * so if it fails loop back and retry.
 870                  */
 871                 if (!atomic_cmpset_ptr(&sx->sx_lock,
 872                     SX_SHARERS_LOCK(1) | SX_LOCK_EXCLUSIVE_WAITERS,
 873                     SX_LOCK_UNLOCKED)) {
 874                         sleepq_release(&sx->lock_object);
 875                         continue;
 876                 }
 877                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 878                         CTR2(KTR_LOCK, "%s: %p waking up all thread on"
 879                             "exclusive queue", __func__, sx);
 880                 sleepq_broadcast(&sx->lock_object, SLEEPQ_SX, -1,
 881                     SQ_EXCLUSIVE_QUEUE);
 882                 break;
 883         }
 884 }
 885
 886 #ifdef INVARIANT_SUPPORT
 887 #ifndef INVARIANTS
 888 #undef  _sx_assert
 889 #endif
 890
 891 /*
 892  * In the non-WITNESS case, sx_assert() can only detect that at least
 893  * *some* thread owns an slock, but it cannot guarantee that *this*
 894  * thread owns an slock.
 895  */
 896 void
 897 _sx_assert(struct sx *sx, int what, const char *file, int line)
 898 {
 899 #ifndef WITNESS
 900         int slocked = 0;
 901 #endif
 902
 903         if (panicstr != NULL)
 904                 return;
 905         switch (what) {
 906         case SA_SLOCKED:
 907         case SA_SLOCKED | SA_NOTRECURSED:
 908         case SA_SLOCKED | SA_RECURSED:
 909 #ifndef WITNESS
 910                 slocked = 1;
 911                 /* FALLTHROUGH */
 912 #endif
 913         case SA_LOCKED:
 914         case SA_LOCKED | SA_NOTRECURSED:
 915         case SA_LOCKED | SA_RECURSED:
 916 #ifdef WITNESS
 917                 witness_assert(&sx->lock_object, what, file, line);
 918 #else
 919                 /*
 920                  * If some other thread has an exclusive lock or we
 921                  * have one and are asserting a shared lock, fail.
 922                  * Also, if no one has a lock at all, fail.
 923                  */
 924                 if (sx->sx_lock == SX_LOCK_UNLOCKED ||
 925                     (!(sx->sx_lock & SX_LOCK_SHARED) && (slocked ||
 926                     sx_xholder(sx) != curthread)))
 927                         panic("Lock %s not %slocked @ %s:%d\n",
 928                             sx->lock_object.lo_name, slocked ? "share " : "",
 929                             file, line);
 930
 931                 if (!(sx->sx_lock & SX_LOCK_SHARED)) {
 932                         if (sx_recursed(sx)) {
 933                                 if (what & SA_NOTRECURSED)
 934                                         panic("Lock %s recursed @ %s:%d\n",
 935                                             sx->lock_object.lo_name, file,
 936                                             line);
 937                         } else if (what & SA_RECURSED)
 938                                 panic("Lock %s not recursed @ %s:%d\n",
 939                                     sx->lock_object.lo_name, file, line);
 940                 }
 941 #endif
 942                 break;
 943         case SA_XLOCKED:
 944         case SA_XLOCKED | SA_NOTRECURSED:
 945         case SA_XLOCKED | SA_RECURSED:
 946                 if (sx_xholder(sx) != curthread)
 947                         panic("Lock %s not exclusively locked @ %s:%d\n",
 948                             sx->lock_object.lo_name, file, line);
 949                 if (sx_recursed(sx)) {
 950                         if (what & SA_NOTRECURSED)
 951                                 panic("Lock %s recursed @ %s:%d\n",
 952                                     sx->lock_object.lo_name, file, line);
 953                 } else if (what & SA_RECURSED)
 954                         panic("Lock %s not recursed @ %s:%d\n",
 955                             sx->lock_object.lo_name, file, line);
 956                 break;
 957         case SA_UNLOCKED:
 958 #ifdef WITNESS
 959                 witness_assert(&sx->lock_object, what, file, line);
 960 #else
 961                 /*
 962                  * If we hold an exclusve lock fail.  We can't
 963                  * reliably check to see if we hold a shared lock or
 964                  * not.
 965                  */
 966                 if (sx_xholder(sx) == curthread)
 967                         panic("Lock %s exclusively locked @ %s:%d\n",
 968                             sx->lock_object.lo_name, file, line);
 969 #endif
 970                 break;
 971         default:
 972                 panic("Unknown sx lock assertion: %d @ %s:%d", what, file,
 973                     line);
 974         }
 975 }
 976 #endif  /* INVARIANT_SUPPORT */
 977
 978 #ifdef DDB
 979 static void
 980 db_show_sx(struct lock_object *lock)
 981 {
 982         struct thread *td;
 983         struct sx *sx;
 984
 985         sx = (struct sx *)lock;
 986
 987         db_printf(" state: ");
 988         if (sx->sx_lock == SX_LOCK_UNLOCKED)
 989                 db_printf("UNLOCKED\n");
 990         else if (sx->sx_lock == SX_LOCK_DESTROYED) {
 991                 db_printf("DESTROYED\n");
 992                 return;
 993         } else if (sx->sx_lock & SX_LOCK_SHARED)
 994                 db_printf("SLOCK: %ju\n", (uintmax_t)SX_SHARERS(sx->sx_lock));
 995         else {
 996                 td = sx_xholder(sx);
 997                 db_printf("XLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
 998                     td->td_tid, td->td_proc->p_pid, td->td_name);
 999                 if (sx_recursed(sx))
1000                         db_printf(" recursed: %d\n", sx->sx_recurse);
1001         }
1002
1003         db_printf(" waiters: ");
1004         switch(sx->sx_lock &
1005             (SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS)) {
1006         case SX_LOCK_SHARED_WAITERS:
1007                 db_printf("shared\n");
1008                 break;
1009         case SX_LOCK_EXCLUSIVE_WAITERS:
1010                 db_printf("exclusive\n");
1011                 break;
1012         case SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS:
1013                 db_printf("exclusive and shared\n");
1014                 break;
1015         default:
1016                 db_printf("none\n");
1017         }
1018 }
1019
1020 /*
1021  * Check to see if a thread that is blocked on a sleep queue is actually
1022  * blocked on an sx lock.  If so, output some details and return true.
1023  * If the lock has an exclusive owner, return that in *ownerp.
1024  */
1025 int
1026 sx_chain(struct thread *td, struct thread **ownerp)
1027 {
1028         struct sx *sx;
1029
1030         /*
1031          * Check to see if this thread is blocked on an sx lock.
1032          * First, we check the lock class.  If that is ok, then we
1033          * compare the lock name against the wait message.
1034          */
1035         sx = td->td_wchan;
1036         if (LOCK_CLASS(&sx->lock_object) != &lock_class_sx ||
1037             sx->lock_object.lo_name != td->td_wmesg)
1038                 return (0);
1039
1040         /* We think we have an sx lock, so output some details. */
1041         db_printf("blocked on sx \"%s\" ", td->td_wmesg);
1042         *ownerp = sx_xholder(sx);
1043         if (sx->sx_lock & SX_LOCK_SHARED)
1044                 db_printf("SLOCK (count %ju)\n",
1045                     (uintmax_t)SX_SHARERS(sx->sx_lock));
1046         else
1047                 db_printf("XLOCK\n");
1048         return (1);
1049 }
1050 #endif