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         for (;;) {
 221                 x = sx->sx_lock;
 222                 KASSERT(x != SX_LOCK_DESTROYED,
 223                     ("sx_try_slock() of destroyed sx @ %s:%d", file, line));
 224                 if (!(x & SX_LOCK_SHARED))
 225                         break;
 226                 if (atomic_cmpset_acq_ptr(&sx->sx_lock, x, x + SX_ONE_SHARER)) {
 227                         LOCK_LOG_TRY("SLOCK", &sx->lock_object, 0, 1, file, line);
 228                         WITNESS_LOCK(&sx->lock_object, LOP_TRYLOCK, file, line);
 229                         curthread->td_locks++;
 230                         return (1);
 231                 }
 232         }
 233
 234         LOCK_LOG_TRY("SLOCK", &sx->lock_object, 0, 0, file, line);
 235         return (0);
 236 }
 237
 238 int
 239 _sx_xlock(struct sx *sx, int opts, const char *file, int line)
 240 {
 241         int error = 0;
 242
 243         MPASS(curthread != NULL);
 244         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 245             ("sx_xlock() of destroyed sx @ %s:%d", file, line));
 246         WITNESS_CHECKORDER(&sx->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE, file,
 247             line);
 248         error = __sx_xlock(sx, curthread, opts, file, line);
 249         if (!error) {
 250                 LOCK_LOG_LOCK("XLOCK", &sx->lock_object, 0, sx->sx_recurse,
 251                     file, line);
 252                 WITNESS_LOCK(&sx->lock_object, LOP_EXCLUSIVE, file, line);
 253                 curthread->td_locks++;
 254         }
 255
 256         return (error);
 257 }
 258
 259 int
 260 _sx_try_xlock(struct sx *sx, const char *file, int line)
 261 {
 262         int rval;
 263
 264         MPASS(curthread != NULL);
 265         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 266             ("sx_try_xlock() of destroyed sx @ %s:%d", file, line));
 267
 268         if (sx_xlocked(sx) && (sx->lock_object.lo_flags & SX_RECURSE) != 0) {
 269                 sx->sx_recurse++;
 270                 atomic_set_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
 271                 rval = 1;
 272         } else
 273                 rval = atomic_cmpset_acq_ptr(&sx->sx_lock, SX_LOCK_UNLOCKED,
 274                     (uintptr_t)curthread);
 275         LOCK_LOG_TRY("XLOCK", &sx->lock_object, 0, rval, file, line);
 276         if (rval) {
 277                 WITNESS_LOCK(&sx->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
 278                     file, line);
 279                 curthread->td_locks++;
 280         }
 281
 282         return (rval);
 283 }
 284
 285 void
 286 _sx_sunlock(struct sx *sx, const char *file, int line)
 287 {
 288
 289         MPASS(curthread != NULL);
 290         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 291             ("sx_sunlock() of destroyed sx @ %s:%d", file, line));
 292         _sx_assert(sx, SA_SLOCKED, file, line);
 293         curthread->td_locks--;
 294         WITNESS_UNLOCK(&sx->lock_object, 0, file, line);
 295         LOCK_LOG_LOCK("SUNLOCK", &sx->lock_object, 0, 0, file, line);
 296 #ifdef LOCK_PROFILING_SHARED
 297         if (SX_SHARERS(sx->sx_lock) == 1)
 298                 lock_profile_release_lock(&sx->lock_object);
 299 #endif
 300         __sx_sunlock(sx, file, line);
 301 }
 302
 303 void
 304 _sx_xunlock(struct sx *sx, const char *file, int line)
 305 {
 306
 307         MPASS(curthread != NULL);
 308         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 309             ("sx_xunlock() of destroyed sx @ %s:%d", file, line));
 310         _sx_assert(sx, SA_XLOCKED, file, line);
 311         curthread->td_locks--;
 312         WITNESS_UNLOCK(&sx->lock_object, LOP_EXCLUSIVE, file, line);
 313         LOCK_LOG_LOCK("XUNLOCK", &sx->lock_object, 0, sx->sx_recurse, file,
 314             line);
 315         if (!sx_recursed(sx))
 316                 lock_profile_release_lock(&sx->lock_object);
 317         __sx_xunlock(sx, curthread, file, line);
 318 }
 319
 320 /*
 321  * Try to do a non-blocking upgrade from a shared lock to an exclusive lock.
 322  * This will only succeed if this thread holds a single shared lock.
 323  * Return 1 if if the upgrade succeed, 0 otherwise.
 324  */
 325 int
 326 _sx_try_upgrade(struct sx *sx, const char *file, int line)
 327 {
 328         uintptr_t x;
 329         int success;
 330
 331         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 332             ("sx_try_upgrade() of destroyed sx @ %s:%d", file, line));
 333         _sx_assert(sx, SA_SLOCKED, file, line);
 334
 335         /*
 336          * Try to switch from one shared lock to an exclusive lock.  We need
 337          * to maintain the SX_LOCK_EXCLUSIVE_WAITERS flag if set so that
 338          * we will wake up the exclusive waiters when we drop the lock.
 339          */
 340         x = sx->sx_lock & SX_LOCK_EXCLUSIVE_WAITERS;
 341         success = atomic_cmpset_ptr(&sx->sx_lock, SX_SHARERS_LOCK(1) | x,
 342             (uintptr_t)curthread | x);
 343         LOCK_LOG_TRY("XUPGRADE", &sx->lock_object, 0, success, file, line);
 344         if (success)
 345                 WITNESS_UPGRADE(&sx->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
 346                     file, line);
 347         return (success);
 348 }
 349
 350 /*
 351  * Downgrade an unrecursed exclusive lock into a single shared lock.
 352  */
 353 void
 354 _sx_downgrade(struct sx *sx, const char *file, int line)
 355 {
 356         uintptr_t x;
 357         int wakeup_swapper;
 358
 359         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 360             ("sx_downgrade() of destroyed sx @ %s:%d", file, line));
 361         _sx_assert(sx, SA_XLOCKED | SA_NOTRECURSED, file, line);
 362 #ifndef INVARIANTS
 363         if (sx_recursed(sx))
 364                 panic("downgrade of a recursed lock");
 365 #endif
 366
 367         WITNESS_DOWNGRADE(&sx->lock_object, 0, file, line);
 368
 369         /*
 370          * Try to switch from an exclusive lock with no shared waiters
 371          * to one sharer with no shared waiters.  If there are
 372          * exclusive waiters, we don't need to lock the sleep queue so
 373          * long as we preserve the flag.  We do one quick try and if
 374          * that fails we grab the sleepq lock to keep the flags from
 375          * changing and do it the slow way.
 376          *
 377          * We have to lock the sleep queue if there are shared waiters
 378          * so we can wake them up.
 379          */
 380         x = sx->sx_lock;
 381         if (!(x & SX_LOCK_SHARED_WAITERS) &&
 382             atomic_cmpset_rel_ptr(&sx->sx_lock, x, SX_SHARERS_LOCK(1) |
 383             (x & SX_LOCK_EXCLUSIVE_WAITERS))) {
 384                 LOCK_LOG_LOCK("XDOWNGRADE", &sx->lock_object, 0, 0, file, line);
 385                 return;
 386         }
 387
 388         /*
 389          * Lock the sleep queue so we can read the waiters bits
 390          * without any races and wakeup any shared waiters.
 391          */
 392         sleepq_lock(&sx->lock_object);
 393
 394         /*
 395          * Preserve SX_LOCK_EXCLUSIVE_WAITERS while downgraded to a single
 396          * shared lock.  If there are any shared waiters, wake them up.
 397          */
 398         wakeup_swapper = 0;
 399         x = sx->sx_lock;
 400         atomic_store_rel_ptr(&sx->sx_lock, SX_SHARERS_LOCK(1) |
 401             (x & SX_LOCK_EXCLUSIVE_WAITERS));
 402         if (x & SX_LOCK_SHARED_WAITERS)
 403                 wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX,
 404                     -1, SQ_SHARED_QUEUE);
 405         else
 406                 sleepq_release(&sx->lock_object);
 407
 408         LOCK_LOG_LOCK("XDOWNGRADE", &sx->lock_object, 0, 0, file, line);
 409
 410         if (wakeup_swapper)
 411                 kick_proc0();
 412 }
 413
 414 /*
 415  * This function represents the so-called 'hard case' for sx_xlock
 416  * operation.  All 'easy case' failures are redirected to this.  Note
 417  * that ideally this would be a static function, but it needs to be
 418  * accessible from at least sx.h.
 419  */
 420 int
 421 _sx_xlock_hard(struct sx *sx, uintptr_t tid, int opts, const char *file,
 422     int line)
 423 {
 424         GIANT_DECLARE;
 425 #ifdef ADAPTIVE_SX
 426         volatile struct thread *owner;
 427 #endif
 428         uint64_t waittime = 0;
 429         uintptr_t x;
 430         int contested = 0, error = 0;
 431
 432         /* If we already hold an exclusive lock, then recurse. */
 433         if (sx_xlocked(sx)) {
 434                 KASSERT((sx->lock_object.lo_flags & SX_RECURSE) != 0,
 435             ("_sx_xlock_hard: recursed on non-recursive sx %s @ %s:%d\n",
 436                     sx->lock_object.lo_name, file, line));
 437                 sx->sx_recurse++;
 438                 atomic_set_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
 439                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 440                         CTR2(KTR_LOCK, "%s: %p recursing", __func__, sx);
 441                 return (0);
 442         }
 443
 444         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 445                 CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__,
 446                     sx->lock_object.lo_name, (void *)sx->sx_lock, file, line);
 447
 448         while (!atomic_cmpset_acq_ptr(&sx->sx_lock, SX_LOCK_UNLOCKED, tid)) {
 449 #ifdef ADAPTIVE_SX
 450                 /*
 451                  * If the lock is write locked and the owner is
 452                  * running on another CPU, spin until the owner stops
 453                  * running or the state of the lock changes.
 454                  */
 455                 x = sx->sx_lock;
 456                 if (!(x & SX_LOCK_SHARED) &&
 457                     (sx->lock_object.lo_flags & SX_ADAPTIVESPIN)) {
 458                         x = SX_OWNER(x);
 459                         owner = (struct thread *)x;
 460                         if (TD_IS_RUNNING(owner)) {
 461                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 462                                         CTR3(KTR_LOCK,
 463                                             "%s: spinning on %p held by %p",
 464                                             __func__, sx, owner);
 465                                 GIANT_SAVE();
 466                                 lock_profile_obtain_lock_failed(
 467                                     &sx->lock_object, &contested, &waittime);
 468                                 while (SX_OWNER(sx->sx_lock) == x &&
 469                                     TD_IS_RUNNING(owner))
 470                                         cpu_spinwait();
 471                                 continue;
 472                         }
 473                 }
 474 #endif
 475
 476                 sleepq_lock(&sx->lock_object);
 477                 x = sx->sx_lock;
 478
 479                 /*
 480                  * If the lock was released while spinning on the
 481                  * sleep queue chain lock, try again.
 482                  */
 483                 if (x == SX_LOCK_UNLOCKED) {
 484                         sleepq_release(&sx->lock_object);
 485                         continue;
 486                 }
 487
 488 #ifdef ADAPTIVE_SX
 489                 /*
 490                  * The current lock owner might have started executing
 491                  * on another CPU (or the lock could have changed
 492                  * owners) while we were waiting on the sleep queue
 493                  * chain lock.  If so, drop the sleep queue lock and try
 494                  * again.
 495                  */
 496                 if (!(x & SX_LOCK_SHARED) &&
 497                     (sx->lock_object.lo_flags & SX_ADAPTIVESPIN)) {
 498                         owner = (struct thread *)SX_OWNER(x);
 499                         if (TD_IS_RUNNING(owner)) {
 500                                 sleepq_release(&sx->lock_object);
 501                                 continue;
 502                         }
 503                 }
 504 #endif
 505
 506                 /*
 507                  * If an exclusive lock was released with both shared
 508                  * and exclusive waiters and a shared waiter hasn't
 509                  * woken up and acquired the lock yet, sx_lock will be
 510                  * set to SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS.
 511                  * If we see that value, try to acquire it once.  Note
 512                  * that we have to preserve SX_LOCK_EXCLUSIVE_WAITERS
 513                  * as there are other exclusive waiters still.  If we
 514                  * fail, restart the loop.
 515                  */
 516                 if (x == (SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS)) {
 517                         if (atomic_cmpset_acq_ptr(&sx->sx_lock,
 518                             SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS,
 519                             tid | SX_LOCK_EXCLUSIVE_WAITERS)) {
 520                                 sleepq_release(&sx->lock_object);
 521                                 CTR2(KTR_LOCK, "%s: %p claimed by new writer",
 522                                     __func__, sx);
 523                                 break;
 524                         }
 525                         sleepq_release(&sx->lock_object);
 526                         continue;
 527                 }
 528
 529                 /*
 530                  * Try to set the SX_LOCK_EXCLUSIVE_WAITERS.  If we fail,
 531                  * than loop back and retry.
 532                  */
 533                 if (!(x & SX_LOCK_EXCLUSIVE_WAITERS)) {
 534                         if (!atomic_cmpset_ptr(&sx->sx_lock, x,
 535                             x | SX_LOCK_EXCLUSIVE_WAITERS)) {
 536                                 sleepq_release(&sx->lock_object);
 537                                 continue;
 538                         }
 539                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 540                                 CTR2(KTR_LOCK, "%s: %p set excl waiters flag",
 541                                     __func__, sx);
 542                 }
 543
 544                 /*
 545                  * Since we have been unable to acquire the exclusive
 546                  * lock and the exclusive waiters flag is set, we have
 547                  * to sleep.
 548                  */
 549                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 550                         CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
 551                             __func__, sx);
 552
 553                 GIANT_SAVE();
 554                 lock_profile_obtain_lock_failed(&sx->lock_object, &contested,
 555                     &waittime);
 556                 sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
 557                     SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
 558                     SLEEPQ_INTERRUPTIBLE : 0), SQ_EXCLUSIVE_QUEUE);
 559                 if (!(opts & SX_INTERRUPTIBLE))
 560                         sleepq_wait(&sx->lock_object);
 561                 else
 562                         error = sleepq_wait_sig(&sx->lock_object);
 563
 564                 if (error) {
 565                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 566                                 CTR2(KTR_LOCK,
 567                         "%s: interruptible sleep by %p suspended by signal",
 568                                     __func__, sx);
 569                         break;
 570                 }
 571                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 572                         CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
 573                             __func__, sx);
 574         }
 575
 576         GIANT_RESTORE();
 577         if (!error)
 578                 lock_profile_obtain_lock_success(&sx->lock_object, contested,
 579                     waittime, file, line);
 580         return (error);
 581 }
 582
 583 /*
 584  * This function represents the so-called 'hard case' for sx_xunlock
 585  * operation.  All 'easy case' failures are redirected to this.  Note
 586  * that ideally this would be a static function, but it needs to be
 587  * accessible from at least sx.h.
 588  */
 589 void
 590 _sx_xunlock_hard(struct sx *sx, uintptr_t tid, const char *file, int line)
 591 {
 592         uintptr_t x;
 593         int queue, wakeup_swapper;
 594
 595         MPASS(!(sx->sx_lock & SX_LOCK_SHARED));
 596
 597         /* If the lock is recursed, then unrecurse one level. */
 598         if (sx_xlocked(sx) && sx_recursed(sx)) {
 599                 if ((--sx->sx_recurse) == 0)
 600                         atomic_clear_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
 601                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 602                         CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, sx);
 603                 return;
 604         }
 605         MPASS(sx->sx_lock & (SX_LOCK_SHARED_WAITERS |
 606             SX_LOCK_EXCLUSIVE_WAITERS));
 607         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 608                 CTR2(KTR_LOCK, "%s: %p contested", __func__, sx);
 609
 610         sleepq_lock(&sx->lock_object);
 611         x = SX_LOCK_UNLOCKED;
 612
 613         /*
 614          * The wake up algorithm here is quite simple and probably not
 615          * ideal.  It gives precedence to shared waiters if they are
 616          * present.  For this condition, we have to preserve the
 617          * state of the exclusive waiters flag.
 618          */
 619         if (sx->sx_lock & SX_LOCK_SHARED_WAITERS) {
 620                 queue = SQ_SHARED_QUEUE;
 621                 x |= (sx->sx_lock & SX_LOCK_EXCLUSIVE_WAITERS);
 622         } else
 623                 queue = SQ_EXCLUSIVE_QUEUE;
 624
 625         /* Wake up all the waiters for the specific queue. */
 626         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 627                 CTR3(KTR_LOCK, "%s: %p waking up all threads on %s queue",
 628                     __func__, sx, queue == SQ_SHARED_QUEUE ? "shared" :
 629                     "exclusive");
 630         atomic_store_rel_ptr(&sx->sx_lock, x);
 631         wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX, -1,
 632             queue);
 633         if (wakeup_swapper)
 634                 kick_proc0();
 635 }
 636
 637 /*
 638  * This function represents the so-called 'hard case' for sx_slock
 639  * operation.  All 'easy case' failures are redirected to this.  Note
 640  * that ideally this would be a static function, but it needs to be
 641  * accessible from at least sx.h.
 642  */
 643 int
 644 _sx_slock_hard(struct sx *sx, int opts, const char *file, int line)
 645 {
 646         GIANT_DECLARE;
 647 #ifdef ADAPTIVE_SX
 648         volatile struct thread *owner;
 649 #endif
 650 #ifdef LOCK_PROFILING_SHARED
 651         uint64_t waittime = 0;
 652         int contested = 0;
 653 #endif
 654         uintptr_t x;
 655         int error = 0;
 656
 657         /*
 658          * As with rwlocks, we don't make any attempt to try to block
 659          * shared locks once there is an exclusive waiter.
 660          */
 661         for (;;) {
 662                 x = sx->sx_lock;
 663
 664                 /*
 665                  * If no other thread has an exclusive lock then try to bump up
 666                  * the count of sharers.  Since we have to preserve the state
 667                  * of SX_LOCK_EXCLUSIVE_WAITERS, if we fail to acquire the
 668                  * shared lock loop back and retry.
 669                  */
 670                 if (x & SX_LOCK_SHARED) {
 671                         MPASS(!(x & SX_LOCK_SHARED_WAITERS));
 672                         if (atomic_cmpset_acq_ptr(&sx->sx_lock, x,
 673                             x + SX_ONE_SHARER)) {
 674 #ifdef LOCK_PROFILING_SHARED
 675                                 if (SX_SHARERS(x) == 0)
 676                                         lock_profile_obtain_lock_success(
 677                                             &sx->lock_object, contested,
 678                                             waittime, file, line);
 679 #endif
 680                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 681                                         CTR4(KTR_LOCK,
 682                                             "%s: %p succeed %p -> %p", __func__,
 683                                             sx, (void *)x,
 684                                             (void *)(x + SX_ONE_SHARER));
 685                                 break;
 686                         }
 687                         continue;
 688                 }
 689
 690 #ifdef ADAPTIVE_SX
 691                 /*
 692                  * If the owner is running on another CPU, spin until
 693                  * the owner stops running or the state of the lock
 694                  * changes.
 695                  */
 696                 else if (sx->lock_object.lo_flags & SX_ADAPTIVESPIN) {
 697                         x = SX_OWNER(x);
 698                         owner = (struct thread *)x;
 699                         if (TD_IS_RUNNING(owner)) {
 700                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 701                                         CTR3(KTR_LOCK,
 702                                             "%s: spinning on %p held by %p",
 703                                             __func__, sx, owner);
 704                                 GIANT_SAVE();
 705 #ifdef LOCK_PROFILING_SHARED
 706                                 lock_profile_obtain_lock_failed(
 707                                     &sx->lock_object, &contested, &waittime);
 708 #endif
 709                                 while (SX_OWNER(sx->sx_lock) == x &&
 710                                     TD_IS_RUNNING(owner))
 711                                         cpu_spinwait();
 712                                 continue;
 713                         }
 714                 }
 715 #endif
 716
 717                 /*
 718                  * Some other thread already has an exclusive lock, so
 719                  * start the process of blocking.
 720                  */
 721                 sleepq_lock(&sx->lock_object);
 722                 x = sx->sx_lock;
 723
 724                 /*
 725                  * The lock could have been released while we spun.
 726                  * In this case loop back and retry.
 727                  */
 728                 if (x & SX_LOCK_SHARED) {
 729                         sleepq_release(&sx->lock_object);
 730                         continue;
 731                 }
 732
 733 #ifdef ADAPTIVE_SX
 734                 /*
 735                  * If the owner is running on another CPU, spin until
 736                  * the owner stops running or the state of the lock
 737                  * changes.
 738                  */
 739                 if (!(x & SX_LOCK_SHARED) &&
 740                     (sx->lock_object.lo_flags & SX_ADAPTIVESPIN)) {
 741                         owner = (struct thread *)SX_OWNER(x);
 742                         if (TD_IS_RUNNING(owner)) {
 743                                 sleepq_release(&sx->lock_object);
 744                                 continue;
 745                         }
 746                 }
 747 #endif
 748
 749                 /*
 750                  * Try to set the SX_LOCK_SHARED_WAITERS flag.  If we
 751                  * fail to set it drop the sleep queue lock and loop
 752                  * back.
 753                  */
 754                 if (!(x & SX_LOCK_SHARED_WAITERS)) {
 755                         if (!atomic_cmpset_ptr(&sx->sx_lock, x,
 756                             x | SX_LOCK_SHARED_WAITERS)) {
 757                                 sleepq_release(&sx->lock_object);
 758                                 continue;
 759                         }
 760                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 761                                 CTR2(KTR_LOCK, "%s: %p set shared waiters flag",
 762                                     __func__, sx);
 763                 }
 764
 765                 /*
 766                  * Since we have been unable to acquire the shared lock,
 767                  * we have to sleep.
 768                  */
 769                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 770                         CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
 771                             __func__, sx);
 772
 773                 GIANT_SAVE();
 774 #ifdef LOCK_PROFILING_SHARED
 775                 lock_profile_obtain_lock_failed(&sx->lock_object, &contested,
 776                     &waittime);
 777 #endif
 778                 sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
 779                     SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
 780                     SLEEPQ_INTERRUPTIBLE : 0), SQ_SHARED_QUEUE);
 781                 if (!(opts & SX_INTERRUPTIBLE))
 782                         sleepq_wait(&sx->lock_object);
 783                 else
 784                         error = sleepq_wait_sig(&sx->lock_object);
 785
 786                 if (error) {
 787                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 788                                 CTR2(KTR_LOCK,
 789                         "%s: interruptible sleep by %p suspended by signal",
 790                                     __func__, sx);
 791                         break;
 792                 }
 793                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 794                         CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
 795                             __func__, sx);
 796         }
 797
 798         GIANT_RESTORE();
 799         return (error);
 800 }
 801
 802 /*
 803  * This function represents the so-called 'hard case' for sx_sunlock
 804  * operation.  All 'easy case' failures are redirected to this.  Note
 805  * that ideally this would be a static function, but it needs to be
 806  * accessible from at least sx.h.
 807  */
 808 void
 809 _sx_sunlock_hard(struct sx *sx, const char *file, int line)
 810 {
 811         uintptr_t x;
 812         int wakeup_swapper;
 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                 wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX,
 881                     -1, SQ_EXCLUSIVE_QUEUE);
 882                 if (wakeup_swapper)
 883                         kick_proc0();
 884                 break;
 885         }
 886 }
 887
 888 #ifdef INVARIANT_SUPPORT
 889 #ifndef INVARIANTS
 890 #undef  _sx_assert
 891 #endif
 892
 893 /*
 894  * In the non-WITNESS case, sx_assert() can only detect that at least
 895  * *some* thread owns an slock, but it cannot guarantee that *this*
 896  * thread owns an slock.
 897  */
 898 void
 899 _sx_assert(struct sx *sx, int what, const char *file, int line)
 900 {
 901 #ifndef WITNESS
 902         int slocked = 0;
 903 #endif
 904
 905         if (panicstr != NULL)
 906                 return;
 907         switch (what) {
 908         case SA_SLOCKED:
 909         case SA_SLOCKED | SA_NOTRECURSED:
 910         case SA_SLOCKED | SA_RECURSED:
 911 #ifndef WITNESS
 912                 slocked = 1;
 913                 /* FALLTHROUGH */
 914 #endif
 915         case SA_LOCKED:
 916         case SA_LOCKED | SA_NOTRECURSED:
 917         case SA_LOCKED | SA_RECURSED:
 918 #ifdef WITNESS
 919                 witness_assert(&sx->lock_object, what, file, line);
 920 #else
 921                 /*
 922                  * If some other thread has an exclusive lock or we
 923                  * have one and are asserting a shared lock, fail.
 924                  * Also, if no one has a lock at all, fail.
 925                  */
 926                 if (sx->sx_lock == SX_LOCK_UNLOCKED ||
 927                     (!(sx->sx_lock & SX_LOCK_SHARED) && (slocked ||
 928                     sx_xholder(sx) != curthread)))
 929                         panic("Lock %s not %slocked @ %s:%d\n",
 930                             sx->lock_object.lo_name, slocked ? "share " : "",
 931                             file, line);
 932
 933                 if (!(sx->sx_lock & SX_LOCK_SHARED)) {
 934                         if (sx_recursed(sx)) {
 935                                 if (what & SA_NOTRECURSED)
 936                                         panic("Lock %s recursed @ %s:%d\n",
 937                                             sx->lock_object.lo_name, file,
 938                                             line);
 939                         } else if (what & SA_RECURSED)
 940                                 panic("Lock %s not recursed @ %s:%d\n",
 941                                     sx->lock_object.lo_name, file, line);
 942                 }
 943 #endif
 944                 break;
 945         case SA_XLOCKED:
 946         case SA_XLOCKED | SA_NOTRECURSED:
 947         case SA_XLOCKED | SA_RECURSED:
 948                 if (sx_xholder(sx) != curthread)
 949                         panic("Lock %s not exclusively locked @ %s:%d\n",
 950                             sx->lock_object.lo_name, file, line);
 951                 if (sx_recursed(sx)) {
 952                         if (what & SA_NOTRECURSED)
 953                                 panic("Lock %s recursed @ %s:%d\n",
 954                                     sx->lock_object.lo_name, file, line);
 955                 } else if (what & SA_RECURSED)
 956                         panic("Lock %s not recursed @ %s:%d\n",
 957                             sx->lock_object.lo_name, file, line);
 958                 break;
 959         case SA_UNLOCKED:
 960 #ifdef WITNESS
 961                 witness_assert(&sx->lock_object, what, file, line);
 962 #else
 963                 /*
 964                  * If we hold an exclusve lock fail.  We can't
 965                  * reliably check to see if we hold a shared lock or
 966                  * not.
 967                  */
 968                 if (sx_xholder(sx) == curthread)
 969                         panic("Lock %s exclusively locked @ %s:%d\n",
 970                             sx->lock_object.lo_name, file, line);
 971 #endif
 972                 break;
 973         default:
 974                 panic("Unknown sx lock assertion: %d @ %s:%d", what, file,
 975                     line);
 976         }
 977 }
 978 #endif  /* INVARIANT_SUPPORT */
 979
 980 #ifdef DDB
 981 static void
 982 db_show_sx(struct lock_object *lock)
 983 {
 984         struct thread *td;
 985         struct sx *sx;
 986
 987         sx = (struct sx *)lock;
 988
 989         db_printf(" state: ");
 990         if (sx->sx_lock == SX_LOCK_UNLOCKED)
 991                 db_printf("UNLOCKED\n");
 992         else if (sx->sx_lock == SX_LOCK_DESTROYED) {
 993                 db_printf("DESTROYED\n");
 994                 return;
 995         } else if (sx->sx_lock & SX_LOCK_SHARED)
 996                 db_printf("SLOCK: %ju\n", (uintmax_t)SX_SHARERS(sx->sx_lock));
 997         else {
 998                 td = sx_xholder(sx);
 999                 db_printf("XLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1000                     td->td_tid, td->td_proc->p_pid, td->td_proc->p_comm);
1001                 if (sx_recursed(sx))
1002                         db_printf(" recursed: %d\n", sx->sx_recurse);
1003         }
1004
1005         db_printf(" waiters: ");
1006         switch(sx->sx_lock &
1007             (SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS)) {
1008         case SX_LOCK_SHARED_WAITERS:
1009                 db_printf("shared\n");
1010                 break;
1011         case SX_LOCK_EXCLUSIVE_WAITERS:
1012                 db_printf("exclusive\n");
1013                 break;
1014         case SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS:
1015                 db_printf("exclusive and shared\n");
1016                 break;
1017         default:
1018                 db_printf("none\n");
1019         }
1020 }
1021
1022 /*
1023  * Check to see if a thread that is blocked on a sleep queue is actually
1024  * blocked on an sx lock.  If so, output some details and return true.
1025  * If the lock has an exclusive owner, return that in *ownerp.
1026  */
1027 int
1028 sx_chain(struct thread *td, struct thread **ownerp)
1029 {
1030         struct sx *sx;
1031
1032         /*
1033          * Check to see if this thread is blocked on an sx lock.
1034          * First, we check the lock class.  If that is ok, then we
1035          * compare the lock name against the wait message.
1036          */
1037         sx = td->td_wchan;
1038         if (LOCK_CLASS(&sx->lock_object) != &lock_class_sx ||
1039             sx->lock_object.lo_name != td->td_wmesg)
1040                 return (0);
1041
1042         /* We think we have an sx lock, so output some details. */
1043         db_printf("blocked on sx \"%s\" ", td->td_wmesg);
1044         *ownerp = sx_xholder(sx);
1045         if (sx->sx_lock & SX_LOCK_SHARED)
1046                 db_printf("SLOCK (count %ju)\n",
1047                     (uintmax_t)SX_SHARERS(sx->sx_lock));
1048         else
1049                 db_printf("XLOCK\n");
1050         return (1);
1051 }
1052 #endif