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_ddb.h"
  40 #include "opt_hwpmc_hooks.h"
  41 #include "opt_no_adaptive_sx.h"
  42
  43 #include <sys/cdefs.h>
  44 __FBSDID("$FreeBSD$");
  45
  46 #include <sys/param.h>
  47 #include <sys/systm.h>
  48 #include <sys/kdb.h>
  49 #include <sys/kernel.h>
  50 #include <sys/ktr.h>
  51 #include <sys/lock.h>
  52 #include <sys/mutex.h>
  53 #include <sys/proc.h>
  54 #include <sys/sched.h>
  55 #include <sys/sleepqueue.h>
  56 #include <sys/sx.h>
  57 #include <sys/smp.h>
  58 #include <sys/sysctl.h>
  59
  60 #if defined(SMP) && !defined(NO_ADAPTIVE_SX)
  61 #include <machine/cpu.h>
  62 #endif
  63
  64 #ifdef DDB
  65 #include <ddb/ddb.h>
  66 #endif
  67
  68 #if defined(SMP) && !defined(NO_ADAPTIVE_SX)
  69 #define ADAPTIVE_SX
  70 #endif
  71
  72 CTASSERT((SX_NOADAPTIVE & LO_CLASSFLAGS) == SX_NOADAPTIVE);
  73
  74 #ifdef HWPMC_HOOKS
  75 #include <sys/pmckern.h>
  76 PMC_SOFT_DECLARE( , , lock, failed);
  77 #endif
  78
  79 /* Handy macros for sleep queues. */
  80 #define SQ_EXCLUSIVE_QUEUE      0
  81 #define SQ_SHARED_QUEUE         1
  82
  83 /*
  84  * Variations on DROP_GIANT()/PICKUP_GIANT() for use in this file.  We
  85  * drop Giant anytime we have to sleep or if we adaptively spin.
  86  */
  87 #define GIANT_DECLARE                                                   \
  88         int _giantcnt = 0;                                              \
  89         WITNESS_SAVE_DECL(Giant)                                        \
  90
  91 #define GIANT_SAVE(work) do {                                           \
  92         if (mtx_owned(&Giant)) {                                        \
  93                 work++;                                                 \
  94                 WITNESS_SAVE(&Giant.lock_object, Giant);                \
  95                 while (mtx_owned(&Giant)) {                             \
  96                         _giantcnt++;                                    \
  97                         mtx_unlock(&Giant);                             \
  98                 }                                                       \
  99         }                                                               \
 100 } while (0)
 101
 102 #define GIANT_RESTORE() do {                                            \
 103         if (_giantcnt > 0) {                                            \
 104                 mtx_assert(&Giant, MA_NOTOWNED);                        \
 105                 while (_giantcnt--)                                     \
 106                         mtx_lock(&Giant);                               \
 107                 WITNESS_RESTORE(&Giant.lock_object, Giant);             \
 108         }                                                               \
 109 } while (0)
 110
 111 /*
 112  * Returns true if an exclusive lock is recursed.  It assumes
 113  * curthread currently has an exclusive lock.
 114  */
 115 #define sx_recursed(sx)         ((sx)->sx_recurse != 0)
 116
 117 static void     assert_sx(const struct lock_object *lock, int what);
 118 #ifdef DDB
 119 static void     db_show_sx(const struct lock_object *lock);
 120 #endif
 121 static void     lock_sx(struct lock_object *lock, uintptr_t how);
 122 #ifdef KDTRACE_HOOKS
 123 static int      owner_sx(const struct lock_object *lock, struct thread **owner);
 124 #endif
 125 static uintptr_t unlock_sx(struct lock_object *lock);
 126
 127 struct lock_class lock_class_sx = {
 128         .lc_name = "sx",
 129         .lc_flags = LC_SLEEPLOCK | LC_SLEEPABLE | LC_RECURSABLE | LC_UPGRADABLE,
 130         .lc_assert = assert_sx,
 131 #ifdef DDB
 132         .lc_ddb_show = db_show_sx,
 133 #endif
 134         .lc_lock = lock_sx,
 135         .lc_unlock = unlock_sx,
 136 #ifdef KDTRACE_HOOKS
 137         .lc_owner = owner_sx,
 138 #endif
 139 };
 140
 141 #ifndef INVARIANTS
 142 #define _sx_assert(sx, what, file, line)
 143 #endif
 144
 145 #ifdef ADAPTIVE_SX
 146 static __read_frequently u_int asx_retries = 10;
 147 static __read_frequently u_int asx_loops = 10000;
 148 static SYSCTL_NODE(_debug, OID_AUTO, sx, CTLFLAG_RD, NULL, "sxlock debugging");
 149 SYSCTL_UINT(_debug_sx, OID_AUTO, retries, CTLFLAG_RW, &asx_retries, 0, "");
 150 SYSCTL_UINT(_debug_sx, OID_AUTO, loops, CTLFLAG_RW, &asx_loops, 0, "");
 151
 152 static struct lock_delay_config __read_frequently sx_delay;
 153
 154 SYSCTL_INT(_debug_sx, OID_AUTO, delay_base, CTLFLAG_RW, &sx_delay.base,
 155     0, "");
 156 SYSCTL_INT(_debug_sx, OID_AUTO, delay_max, CTLFLAG_RW, &sx_delay.max,
 157     0, "");
 158
 159 LOCK_DELAY_SYSINIT_DEFAULT(sx_delay);
 160 #endif
 161
 162 void
 163 assert_sx(const struct lock_object *lock, int what)
 164 {
 165
 166         sx_assert((const struct sx *)lock, what);
 167 }
 168
 169 void
 170 lock_sx(struct lock_object *lock, uintptr_t how)
 171 {
 172         struct sx *sx;
 173
 174         sx = (struct sx *)lock;
 175         if (how)
 176                 sx_slock(sx);
 177         else
 178                 sx_xlock(sx);
 179 }
 180
 181 uintptr_t
 182 unlock_sx(struct lock_object *lock)
 183 {
 184         struct sx *sx;
 185
 186         sx = (struct sx *)lock;
 187         sx_assert(sx, SA_LOCKED | SA_NOTRECURSED);
 188         if (sx_xlocked(sx)) {
 189                 sx_xunlock(sx);
 190                 return (0);
 191         } else {
 192                 sx_sunlock(sx);
 193                 return (1);
 194         }
 195 }
 196
 197 #ifdef KDTRACE_HOOKS
 198 int
 199 owner_sx(const struct lock_object *lock, struct thread **owner)
 200 {
 201         const struct sx *sx;
 202         uintptr_t x;
 203
 204         sx = (const struct sx *)lock;
 205         x = sx->sx_lock;
 206         *owner = NULL;
 207         return ((x & SX_LOCK_SHARED) != 0 ? (SX_SHARERS(x) != 0) :
 208             ((*owner = (struct thread *)SX_OWNER(x)) != NULL));
 209 }
 210 #endif
 211
 212 void
 213 sx_sysinit(void *arg)
 214 {
 215         struct sx_args *sargs = arg;
 216
 217         sx_init_flags(sargs->sa_sx, sargs->sa_desc, sargs->sa_flags);
 218 }
 219
 220 void
 221 sx_init_flags(struct sx *sx, const char *description, int opts)
 222 {
 223         int flags;
 224
 225         MPASS((opts & ~(SX_QUIET | SX_RECURSE | SX_NOWITNESS | SX_DUPOK |
 226             SX_NOPROFILE | SX_NOADAPTIVE | SX_NEW)) == 0);
 227         ASSERT_ATOMIC_LOAD_PTR(sx->sx_lock,
 228             ("%s: sx_lock not aligned for %s: %p", __func__, description,
 229             &sx->sx_lock));
 230
 231         flags = LO_SLEEPABLE | LO_UPGRADABLE;
 232         if (opts & SX_DUPOK)
 233                 flags |= LO_DUPOK;
 234         if (opts & SX_NOPROFILE)
 235                 flags |= LO_NOPROFILE;
 236         if (!(opts & SX_NOWITNESS))
 237                 flags |= LO_WITNESS;
 238         if (opts & SX_RECURSE)
 239                 flags |= LO_RECURSABLE;
 240         if (opts & SX_QUIET)
 241                 flags |= LO_QUIET;
 242         if (opts & SX_NEW)
 243                 flags |= LO_NEW;
 244
 245         flags |= opts & SX_NOADAPTIVE;
 246         lock_init(&sx->lock_object, &lock_class_sx, description, NULL, flags);
 247         sx->sx_lock = SX_LOCK_UNLOCKED;
 248         sx->sx_recurse = 0;
 249 }
 250
 251 void
 252 sx_destroy(struct sx *sx)
 253 {
 254
 255         KASSERT(sx->sx_lock == SX_LOCK_UNLOCKED, ("sx lock still held"));
 256         KASSERT(sx->sx_recurse == 0, ("sx lock still recursed"));
 257         sx->sx_lock = SX_LOCK_DESTROYED;
 258         lock_destroy(&sx->lock_object);
 259 }
 260
 261 int
 262 sx_try_slock_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
 263 {
 264         uintptr_t x;
 265
 266         if (SCHEDULER_STOPPED())
 267                 return (1);
 268
 269         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
 270             ("sx_try_slock() by idle thread %p on sx %s @ %s:%d",
 271             curthread, sx->lock_object.lo_name, file, line));
 272
 273         x = sx->sx_lock;
 274         for (;;) {
 275                 KASSERT(x != SX_LOCK_DESTROYED,
 276                     ("sx_try_slock() of destroyed sx @ %s:%d", file, line));
 277                 if (!(x & SX_LOCK_SHARED))
 278                         break;
 279                 if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, x + SX_ONE_SHARER)) {
 280                         LOCK_LOG_TRY("SLOCK", &sx->lock_object, 0, 1, file, line);
 281                         WITNESS_LOCK(&sx->lock_object, LOP_TRYLOCK, file, line);
 282                         LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire,
 283                             sx, 0, 0, file, line, LOCKSTAT_READER);
 284                         TD_LOCKS_INC(curthread);
 285                         return (1);
 286                 }
 287         }
 288
 289         LOCK_LOG_TRY("SLOCK", &sx->lock_object, 0, 0, file, line);
 290         return (0);
 291 }
 292
 293 int
 294 sx_try_slock_(struct sx *sx, const char *file, int line)
 295 {
 296
 297         return (sx_try_slock_int(sx LOCK_FILE_LINE_ARG));
 298 }
 299
 300 int
 301 _sx_xlock(struct sx *sx, int opts, const char *file, int line)
 302 {
 303         uintptr_t tid, x;
 304         int error = 0;
 305
 306         KASSERT(kdb_active != 0 || SCHEDULER_STOPPED() ||
 307             !TD_IS_IDLETHREAD(curthread),
 308             ("sx_xlock() by idle thread %p on sx %s @ %s:%d",
 309             curthread, sx->lock_object.lo_name, file, line));
 310         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 311             ("sx_xlock() of destroyed sx @ %s:%d", file, line));
 312         WITNESS_CHECKORDER(&sx->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE, file,
 313             line, NULL);
 314         tid = (uintptr_t)curthread;
 315         x = SX_LOCK_UNLOCKED;
 316         if (!atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid))
 317                 error = _sx_xlock_hard(sx, x, opts LOCK_FILE_LINE_ARG);
 318         else
 319                 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
 320                     0, 0, file, line, LOCKSTAT_WRITER);
 321         if (!error) {
 322                 LOCK_LOG_LOCK("XLOCK", &sx->lock_object, 0, sx->sx_recurse,
 323                     file, line);
 324                 WITNESS_LOCK(&sx->lock_object, LOP_EXCLUSIVE, file, line);
 325                 TD_LOCKS_INC(curthread);
 326         }
 327
 328         return (error);
 329 }
 330
 331 int
 332 sx_try_xlock_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
 333 {
 334         struct thread *td;
 335         uintptr_t tid, x;
 336         int rval;
 337         bool recursed;
 338
 339         td = curthread;
 340         tid = (uintptr_t)td;
 341         if (SCHEDULER_STOPPED_TD(td))
 342                 return (1);
 343
 344         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(td),
 345             ("sx_try_xlock() by idle thread %p on sx %s @ %s:%d",
 346             curthread, sx->lock_object.lo_name, file, line));
 347         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 348             ("sx_try_xlock() of destroyed sx @ %s:%d", file, line));
 349
 350         rval = 1;
 351         recursed = false;
 352         x = SX_LOCK_UNLOCKED;
 353         for (;;) {
 354                 if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid))
 355                         break;
 356                 if (x == SX_LOCK_UNLOCKED)
 357                         continue;
 358                 if (x == tid && (sx->lock_object.lo_flags & LO_RECURSABLE)) {
 359                         sx->sx_recurse++;
 360                         atomic_set_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
 361                         break;
 362                 }
 363                 rval = 0;
 364                 break;
 365         }
 366
 367         LOCK_LOG_TRY("XLOCK", &sx->lock_object, 0, rval, file, line);
 368         if (rval) {
 369                 WITNESS_LOCK(&sx->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
 370                     file, line);
 371                 if (!recursed)
 372                         LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire,
 373                             sx, 0, 0, file, line, LOCKSTAT_WRITER);
 374                 TD_LOCKS_INC(curthread);
 375         }
 376
 377         return (rval);
 378 }
 379
 380 int
 381 sx_try_xlock_(struct sx *sx, const char *file, int line)
 382 {
 383
 384         return (sx_try_xlock_int(sx LOCK_FILE_LINE_ARG));
 385 }
 386
 387 void
 388 _sx_xunlock(struct sx *sx, const char *file, int line)
 389 {
 390
 391         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 392             ("sx_xunlock() of destroyed sx @ %s:%d", file, line));
 393         _sx_assert(sx, SA_XLOCKED, file, line);
 394         WITNESS_UNLOCK(&sx->lock_object, LOP_EXCLUSIVE, file, line);
 395         LOCK_LOG_LOCK("XUNLOCK", &sx->lock_object, 0, sx->sx_recurse, file,
 396             line);
 397 #if LOCK_DEBUG > 0
 398         _sx_xunlock_hard(sx, (uintptr_t)curthread, file, line);
 399 #else
 400         __sx_xunlock(sx, curthread, file, line);
 401 #endif
 402         TD_LOCKS_DEC(curthread);
 403 }
 404
 405 /*
 406  * Try to do a non-blocking upgrade from a shared lock to an exclusive lock.
 407  * This will only succeed if this thread holds a single shared lock.
 408  * Return 1 if if the upgrade succeed, 0 otherwise.
 409  */
 410 int
 411 sx_try_upgrade_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
 412 {
 413         uintptr_t x;
 414         int success;
 415
 416         if (SCHEDULER_STOPPED())
 417                 return (1);
 418
 419         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 420             ("sx_try_upgrade() of destroyed sx @ %s:%d", file, line));
 421         _sx_assert(sx, SA_SLOCKED, file, line);
 422
 423         /*
 424          * Try to switch from one shared lock to an exclusive lock.  We need
 425          * to maintain the SX_LOCK_EXCLUSIVE_WAITERS flag if set so that
 426          * we will wake up the exclusive waiters when we drop the lock.
 427          */
 428         x = sx->sx_lock & SX_LOCK_EXCLUSIVE_WAITERS;
 429         success = atomic_cmpset_acq_ptr(&sx->sx_lock, SX_SHARERS_LOCK(1) | x,
 430             (uintptr_t)curthread | x);
 431         LOCK_LOG_TRY("XUPGRADE", &sx->lock_object, 0, success, file, line);
 432         if (success) {
 433                 WITNESS_UPGRADE(&sx->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
 434                     file, line);
 435                 LOCKSTAT_RECORD0(sx__upgrade, sx);
 436         }
 437         return (success);
 438 }
 439
 440 int
 441 sx_try_upgrade_(struct sx *sx, const char *file, int line)
 442 {
 443
 444         return (sx_try_upgrade_int(sx LOCK_FILE_LINE_ARG));
 445 }
 446
 447 /*
 448  * Downgrade an unrecursed exclusive lock into a single shared lock.
 449  */
 450 void
 451 sx_downgrade_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
 452 {
 453         uintptr_t x;
 454         int wakeup_swapper;
 455
 456         if (SCHEDULER_STOPPED())
 457                 return;
 458
 459         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 460             ("sx_downgrade() of destroyed sx @ %s:%d", file, line));
 461         _sx_assert(sx, SA_XLOCKED | SA_NOTRECURSED, file, line);
 462 #ifndef INVARIANTS
 463         if (sx_recursed(sx))
 464                 panic("downgrade of a recursed lock");
 465 #endif
 466
 467         WITNESS_DOWNGRADE(&sx->lock_object, 0, file, line);
 468
 469         /*
 470          * Try to switch from an exclusive lock with no shared waiters
 471          * to one sharer with no shared waiters.  If there are
 472          * exclusive waiters, we don't need to lock the sleep queue so
 473          * long as we preserve the flag.  We do one quick try and if
 474          * that fails we grab the sleepq lock to keep the flags from
 475          * changing and do it the slow way.
 476          *
 477          * We have to lock the sleep queue if there are shared waiters
 478          * so we can wake them up.
 479          */
 480         x = sx->sx_lock;
 481         if (!(x & SX_LOCK_SHARED_WAITERS) &&
 482             atomic_cmpset_rel_ptr(&sx->sx_lock, x, SX_SHARERS_LOCK(1) |
 483             (x & SX_LOCK_EXCLUSIVE_WAITERS))) {
 484                 LOCK_LOG_LOCK("XDOWNGRADE", &sx->lock_object, 0, 0, file, line);
 485                 return;
 486         }
 487
 488         /*
 489          * Lock the sleep queue so we can read the waiters bits
 490          * without any races and wakeup any shared waiters.
 491          */
 492         sleepq_lock(&sx->lock_object);
 493
 494         /*
 495          * Preserve SX_LOCK_EXCLUSIVE_WAITERS while downgraded to a single
 496          * shared lock.  If there are any shared waiters, wake them up.
 497          */
 498         wakeup_swapper = 0;
 499         x = sx->sx_lock;
 500         atomic_store_rel_ptr(&sx->sx_lock, SX_SHARERS_LOCK(1) |
 501             (x & SX_LOCK_EXCLUSIVE_WAITERS));
 502         if (x & SX_LOCK_SHARED_WAITERS)
 503                 wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX,
 504                     0, SQ_SHARED_QUEUE);
 505         sleepq_release(&sx->lock_object);
 506
 507         LOCK_LOG_LOCK("XDOWNGRADE", &sx->lock_object, 0, 0, file, line);
 508         LOCKSTAT_RECORD0(sx__downgrade, sx);
 509
 510         if (wakeup_swapper)
 511                 kick_proc0();
 512 }
 513
 514 void
 515 sx_downgrade_(struct sx *sx, const char *file, int line)
 516 {
 517
 518         sx_downgrade_int(sx LOCK_FILE_LINE_ARG);
 519 }
 520
 521 /*
 522  * This function represents the so-called 'hard case' for sx_xlock
 523  * operation.  All 'easy case' failures are redirected to this.  Note
 524  * that ideally this would be a static function, but it needs to be
 525  * accessible from at least sx.h.
 526  */
 527 int
 528 _sx_xlock_hard(struct sx *sx, uintptr_t x, int opts LOCK_FILE_LINE_ARG_DEF)
 529 {
 530         GIANT_DECLARE;
 531         uintptr_t tid;
 532 #ifdef ADAPTIVE_SX
 533         volatile struct thread *owner;
 534         u_int i, n, spintries = 0;
 535 #endif
 536 #ifdef LOCK_PROFILING
 537         uint64_t waittime = 0;
 538         int contested = 0;
 539 #endif
 540         int error = 0;
 541 #if defined(ADAPTIVE_SX) || defined(KDTRACE_HOOKS)
 542         struct lock_delay_arg lda;
 543 #endif
 544 #ifdef  KDTRACE_HOOKS
 545         u_int sleep_cnt = 0;
 546         int64_t sleep_time = 0;
 547         int64_t all_time = 0;
 548 #endif
 549 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
 550         uintptr_t state;
 551 #endif
 552         int extra_work = 0;
 553
 554         tid = (uintptr_t)curthread;
 555         if (SCHEDULER_STOPPED())
 556                 return (0);
 557
 558 #if defined(ADAPTIVE_SX)
 559         lock_delay_arg_init(&lda, &sx_delay);
 560 #elif defined(KDTRACE_HOOKS)
 561         lock_delay_arg_init(&lda, NULL);
 562 #endif
 563
 564         if (__predict_false(x == SX_LOCK_UNLOCKED))
 565                 x = SX_READ_VALUE(sx);
 566
 567         /* If we already hold an exclusive lock, then recurse. */
 568         if (__predict_false(lv_sx_owner(x) == (struct thread *)tid)) {
 569                 KASSERT((sx->lock_object.lo_flags & LO_RECURSABLE) != 0,
 570             ("_sx_xlock_hard: recursed on non-recursive sx %s @ %s:%d\n",
 571                     sx->lock_object.lo_name, file, line));
 572                 sx->sx_recurse++;
 573                 atomic_set_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
 574                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 575                         CTR2(KTR_LOCK, "%s: %p recursing", __func__, sx);
 576                 return (0);
 577         }
 578
 579         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 580                 CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__,
 581                     sx->lock_object.lo_name, (void *)sx->sx_lock, file, line);
 582
 583 #ifdef HWPMC_HOOKS
 584         PMC_SOFT_CALL( , , lock, failed);
 585 #endif
 586         lock_profile_obtain_lock_failed(&sx->lock_object, &contested,
 587             &waittime);
 588
 589 #ifdef LOCK_PROFILING
 590         extra_work = 1;
 591         state = x;
 592 #elif defined(KDTRACE_HOOKS)
 593         extra_work = lockstat_enabled;
 594         if (__predict_false(extra_work)) {
 595                 all_time -= lockstat_nsecs(&sx->lock_object);
 596                 state = x;
 597         }
 598 #endif
 599
 600         for (;;) {
 601                 if (x == SX_LOCK_UNLOCKED) {
 602                         if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid))
 603                                 break;
 604                         continue;
 605                 }
 606 #ifdef KDTRACE_HOOKS
 607                 lda.spin_cnt++;
 608 #endif
 609 #ifdef ADAPTIVE_SX
 610                 /*
 611                  * If the lock is write locked and the owner is
 612                  * running on another CPU, spin until the owner stops
 613                  * running or the state of the lock changes.
 614                  */
 615                 if ((sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) {
 616                         if ((x & SX_LOCK_SHARED) == 0) {
 617                                 owner = lv_sx_owner(x);
 618                                 if (TD_IS_RUNNING(owner)) {
 619                                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 620                                                 CTR3(KTR_LOCK,
 621                                             "%s: spinning on %p held by %p",
 622                                                     __func__, sx, owner);
 623                                         KTR_STATE1(KTR_SCHED, "thread",
 624                                             sched_tdname(curthread), "spinning",
 625                                             "lockname:\"%s\"",
 626                                             sx->lock_object.lo_name);
 627                                         GIANT_SAVE(extra_work);
 628                                         do {
 629                                                 lock_delay(&lda);
 630                                                 x = SX_READ_VALUE(sx);
 631                                                 owner = lv_sx_owner(x);
 632                                         } while (owner != NULL &&
 633                                                     TD_IS_RUNNING(owner));
 634                                         KTR_STATE0(KTR_SCHED, "thread",
 635                                             sched_tdname(curthread), "running");
 636                                         continue;
 637                                 }
 638                         } else if (SX_SHARERS(x) && spintries < asx_retries) {
 639                                 KTR_STATE1(KTR_SCHED, "thread",
 640                                     sched_tdname(curthread), "spinning",
 641                                     "lockname:\"%s\"", sx->lock_object.lo_name);
 642                                 GIANT_SAVE(extra_work);
 643                                 spintries++;
 644                                 for (i = 0; i < asx_loops; i += n) {
 645                                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 646                                                 CTR4(KTR_LOCK,
 647                                     "%s: shared spinning on %p with %u and %u",
 648                                                     __func__, sx, spintries, i);
 649                                         n = SX_SHARERS(x);
 650                                         lock_delay_spin(n);
 651                                         x = SX_READ_VALUE(sx);
 652                                         if ((x & SX_LOCK_SHARED) == 0 ||
 653                                             SX_SHARERS(x) == 0)
 654                                                 break;
 655                                 }
 656 #ifdef KDTRACE_HOOKS
 657                                 lda.spin_cnt += i;
 658 #endif
 659                                 KTR_STATE0(KTR_SCHED, "thread",
 660                                     sched_tdname(curthread), "running");
 661                                 if (i != asx_loops)
 662                                         continue;
 663                         }
 664                 }
 665 #endif
 666
 667                 sleepq_lock(&sx->lock_object);
 668                 x = SX_READ_VALUE(sx);
 669
 670                 /*
 671                  * If the lock was released while spinning on the
 672                  * sleep queue chain lock, try again.
 673                  */
 674                 if (x == SX_LOCK_UNLOCKED) {
 675                         sleepq_release(&sx->lock_object);
 676                         continue;
 677                 }
 678
 679 #ifdef ADAPTIVE_SX
 680                 /*
 681                  * The current lock owner might have started executing
 682                  * on another CPU (or the lock could have changed
 683                  * owners) while we were waiting on the sleep queue
 684                  * chain lock.  If so, drop the sleep queue lock and try
 685                  * again.
 686                  */
 687                 if (!(x & SX_LOCK_SHARED) &&
 688                     (sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) {
 689                         owner = (struct thread *)SX_OWNER(x);
 690                         if (TD_IS_RUNNING(owner)) {
 691                                 sleepq_release(&sx->lock_object);
 692                                 continue;
 693                         }
 694                 }
 695 #endif
 696
 697                 /*
 698                  * If an exclusive lock was released with both shared
 699                  * and exclusive waiters and a shared waiter hasn't
 700                  * woken up and acquired the lock yet, sx_lock will be
 701                  * set to SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS.
 702                  * If we see that value, try to acquire it once.  Note
 703                  * that we have to preserve SX_LOCK_EXCLUSIVE_WAITERS
 704                  * as there are other exclusive waiters still.  If we
 705                  * fail, restart the loop.
 706                  */
 707                 if (x == (SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS)) {
 708                         if (atomic_cmpset_acq_ptr(&sx->sx_lock,
 709                             SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS,
 710                             tid | SX_LOCK_EXCLUSIVE_WAITERS)) {
 711                                 sleepq_release(&sx->lock_object);
 712                                 CTR2(KTR_LOCK, "%s: %p claimed by new writer",
 713                                     __func__, sx);
 714                                 break;
 715                         }
 716                         sleepq_release(&sx->lock_object);
 717                         x = SX_READ_VALUE(sx);
 718                         continue;
 719                 }
 720
 721                 /*
 722                  * Try to set the SX_LOCK_EXCLUSIVE_WAITERS.  If we fail,
 723                  * than loop back and retry.
 724                  */
 725                 if (!(x & SX_LOCK_EXCLUSIVE_WAITERS)) {
 726                         if (!atomic_cmpset_ptr(&sx->sx_lock, x,
 727                             x | SX_LOCK_EXCLUSIVE_WAITERS)) {
 728                                 sleepq_release(&sx->lock_object);
 729                                 x = SX_READ_VALUE(sx);
 730                                 continue;
 731                         }
 732                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 733                                 CTR2(KTR_LOCK, "%s: %p set excl waiters flag",
 734                                     __func__, sx);
 735                 }
 736
 737                 /*
 738                  * Since we have been unable to acquire the exclusive
 739                  * lock and the exclusive waiters flag is set, we have
 740                  * to sleep.
 741                  */
 742                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 743                         CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
 744                             __func__, sx);
 745
 746 #ifdef KDTRACE_HOOKS
 747                 sleep_time -= lockstat_nsecs(&sx->lock_object);
 748 #endif
 749                 GIANT_SAVE(extra_work);
 750                 sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
 751                     SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
 752                     SLEEPQ_INTERRUPTIBLE : 0), SQ_EXCLUSIVE_QUEUE);
 753                 if (!(opts & SX_INTERRUPTIBLE))
 754                         sleepq_wait(&sx->lock_object, 0);
 755                 else
 756                         error = sleepq_wait_sig(&sx->lock_object, 0);
 757 #ifdef KDTRACE_HOOKS
 758                 sleep_time += lockstat_nsecs(&sx->lock_object);
 759                 sleep_cnt++;
 760 #endif
 761                 if (error) {
 762                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 763                                 CTR2(KTR_LOCK,
 764                         "%s: interruptible sleep by %p suspended by signal",
 765                                     __func__, sx);
 766                         break;
 767                 }
 768                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 769                         CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
 770                             __func__, sx);
 771                 x = SX_READ_VALUE(sx);
 772         }
 773 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
 774         if (__predict_true(!extra_work))
 775                 return (error);
 776 #endif
 777 #ifdef KDTRACE_HOOKS
 778         all_time += lockstat_nsecs(&sx->lock_object);
 779         if (sleep_time)
 780                 LOCKSTAT_RECORD4(sx__block, sx, sleep_time,
 781                     LOCKSTAT_WRITER, (state & SX_LOCK_SHARED) == 0,
 782                     (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
 783         if (lda.spin_cnt > sleep_cnt)
 784                 LOCKSTAT_RECORD4(sx__spin, sx, all_time - sleep_time,
 785                     LOCKSTAT_WRITER, (state & SX_LOCK_SHARED) == 0,
 786                     (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
 787 #endif
 788         if (!error)
 789                 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
 790                     contested, waittime, file, line, LOCKSTAT_WRITER);
 791         GIANT_RESTORE();
 792         return (error);
 793 }
 794
 795 /*
 796  * This function represents the so-called 'hard case' for sx_xunlock
 797  * operation.  All 'easy case' failures are redirected to this.  Note
 798  * that ideally this would be a static function, but it needs to be
 799  * accessible from at least sx.h.
 800  */
 801 void
 802 _sx_xunlock_hard(struct sx *sx, uintptr_t x LOCK_FILE_LINE_ARG_DEF)
 803 {
 804         uintptr_t tid, setx;
 805         int queue, wakeup_swapper;
 806
 807         if (SCHEDULER_STOPPED())
 808                 return;
 809
 810         tid = (uintptr_t)curthread;
 811
 812         if (__predict_false(x == tid))
 813                 x = SX_READ_VALUE(sx);
 814
 815         MPASS(!(x & SX_LOCK_SHARED));
 816
 817         if (__predict_false(x & SX_LOCK_RECURSED)) {
 818                 /* The lock is recursed, unrecurse one level. */
 819                 if ((--sx->sx_recurse) == 0)
 820                         atomic_clear_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
 821                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 822                         CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, sx);
 823                 return;
 824         }
 825
 826         LOCKSTAT_PROFILE_RELEASE_RWLOCK(sx__release, sx, LOCKSTAT_WRITER);
 827         if (x == tid &&
 828             atomic_cmpset_rel_ptr(&sx->sx_lock, tid, SX_LOCK_UNLOCKED))
 829                 return;
 830
 831         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 832                 CTR2(KTR_LOCK, "%s: %p contested", __func__, sx);
 833
 834         sleepq_lock(&sx->lock_object);
 835         x = SX_READ_VALUE(sx);
 836         MPASS(x & (SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS));
 837
 838         /*
 839          * The wake up algorithm here is quite simple and probably not
 840          * ideal.  It gives precedence to shared waiters if they are
 841          * present.  For this condition, we have to preserve the
 842          * state of the exclusive waiters flag.
 843          * If interruptible sleeps left the shared queue empty avoid a
 844          * starvation for the threads sleeping on the exclusive queue by giving
 845          * them precedence and cleaning up the shared waiters bit anyway.
 846          */
 847         setx = SX_LOCK_UNLOCKED;
 848         queue = SQ_EXCLUSIVE_QUEUE;
 849         if ((x & SX_LOCK_SHARED_WAITERS) != 0 &&
 850             sleepq_sleepcnt(&sx->lock_object, SQ_SHARED_QUEUE) != 0) {
 851                 queue = SQ_SHARED_QUEUE;
 852                 setx |= (x & SX_LOCK_EXCLUSIVE_WAITERS);
 853         }
 854         atomic_store_rel_ptr(&sx->sx_lock, setx);
 855
 856         /* Wake up all the waiters for the specific queue. */
 857         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 858                 CTR3(KTR_LOCK, "%s: %p waking up all threads on %s queue",
 859                     __func__, sx, queue == SQ_SHARED_QUEUE ? "shared" :
 860                     "exclusive");
 861
 862         wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX, 0,
 863             queue);
 864         sleepq_release(&sx->lock_object);
 865         if (wakeup_swapper)
 866                 kick_proc0();
 867 }
 868
 869 static bool __always_inline
 870 __sx_slock_try(struct sx *sx, uintptr_t *xp LOCK_FILE_LINE_ARG_DEF)
 871 {
 872
 873         /*
 874          * If no other thread has an exclusive lock then try to bump up
 875          * the count of sharers.  Since we have to preserve the state
 876          * of SX_LOCK_EXCLUSIVE_WAITERS, if we fail to acquire the
 877          * shared lock loop back and retry.
 878          */
 879         while (*xp & SX_LOCK_SHARED) {
 880                 MPASS(!(*xp & SX_LOCK_SHARED_WAITERS));
 881                 if (atomic_fcmpset_acq_ptr(&sx->sx_lock, xp,
 882                     *xp + SX_ONE_SHARER)) {
 883                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 884                                 CTR4(KTR_LOCK, "%s: %p succeed %p -> %p",
 885                                     __func__, sx, (void *)*xp,
 886                                     (void *)(*xp + SX_ONE_SHARER));
 887                         return (true);
 888                 }
 889         }
 890         return (false);
 891 }
 892
 893 static int __noinline
 894 _sx_slock_hard(struct sx *sx, int opts, uintptr_t x LOCK_FILE_LINE_ARG_DEF)
 895 {
 896         GIANT_DECLARE;
 897 #ifdef ADAPTIVE_SX
 898         volatile struct thread *owner;
 899 #endif
 900 #ifdef LOCK_PROFILING
 901         uint64_t waittime = 0;
 902         int contested = 0;
 903 #endif
 904         int error = 0;
 905 #if defined(ADAPTIVE_SX) || defined(KDTRACE_HOOKS)
 906         struct lock_delay_arg lda;
 907 #endif
 908 #ifdef KDTRACE_HOOKS
 909         u_int sleep_cnt = 0;
 910         int64_t sleep_time = 0;
 911         int64_t all_time = 0;
 912 #endif
 913 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
 914         uintptr_t state;
 915 #endif
 916         int extra_work = 0;
 917
 918         if (SCHEDULER_STOPPED())
 919                 return (0);
 920
 921 #if defined(ADAPTIVE_SX)
 922         lock_delay_arg_init(&lda, &sx_delay);
 923 #elif defined(KDTRACE_HOOKS)
 924         lock_delay_arg_init(&lda, NULL);
 925 #endif
 926
 927 #ifdef HWPMC_HOOKS
 928         PMC_SOFT_CALL( , , lock, failed);
 929 #endif
 930         lock_profile_obtain_lock_failed(&sx->lock_object, &contested,
 931             &waittime);
 932
 933 #ifdef LOCK_PROFILING
 934         extra_work = 1;
 935         state = x;
 936 #elif defined(KDTRACE_HOOKS)
 937         extra_work = lockstat_enabled;
 938         if (__predict_false(extra_work)) {
 939                 all_time -= lockstat_nsecs(&sx->lock_object);
 940                 state = x;
 941         }
 942 #endif
 943
 944         /*
 945          * As with rwlocks, we don't make any attempt to try to block
 946          * shared locks once there is an exclusive waiter.
 947          */
 948         for (;;) {
 949                 if (__sx_slock_try(sx, &x LOCK_FILE_LINE_ARG))
 950                         break;
 951 #ifdef KDTRACE_HOOKS
 952                 lda.spin_cnt++;
 953 #endif
 954
 955 #ifdef ADAPTIVE_SX
 956                 /*
 957                  * If the owner is running on another CPU, spin until
 958                  * the owner stops running or the state of the lock
 959                  * changes.
 960                  */
 961                 if ((sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) {
 962                         owner = lv_sx_owner(x);
 963                         if (TD_IS_RUNNING(owner)) {
 964                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 965                                         CTR3(KTR_LOCK,
 966                                             "%s: spinning on %p held by %p",
 967                                             __func__, sx, owner);
 968                                 KTR_STATE1(KTR_SCHED, "thread",
 969                                     sched_tdname(curthread), "spinning",
 970                                     "lockname:\"%s\"", sx->lock_object.lo_name);
 971                                 GIANT_SAVE(extra_work);
 972                                 do {
 973                                         lock_delay(&lda);
 974                                         x = SX_READ_VALUE(sx);
 975                                         owner = lv_sx_owner(x);
 976                                 } while (owner != NULL && TD_IS_RUNNING(owner));
 977                                 KTR_STATE0(KTR_SCHED, "thread",
 978                                     sched_tdname(curthread), "running");
 979                                 continue;
 980                         }
 981                 }
 982 #endif
 983
 984                 /*
 985                  * Some other thread already has an exclusive lock, so
 986                  * start the process of blocking.
 987                  */
 988                 sleepq_lock(&sx->lock_object);
 989                 x = SX_READ_VALUE(sx);
 990
 991                 /*
 992                  * The lock could have been released while we spun.
 993                  * In this case loop back and retry.
 994                  */
 995                 if (x & SX_LOCK_SHARED) {
 996                         sleepq_release(&sx->lock_object);
 997                         continue;
 998                 }
 999
1000 #ifdef ADAPTIVE_SX
1001                 /*
1002                  * If the owner is running on another CPU, spin until
1003                  * the owner stops running or the state of the lock
1004                  * changes.
1005                  */
1006                 if (!(x & SX_LOCK_SHARED) &&
1007                     (sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) {
1008                         owner = (struct thread *)SX_OWNER(x);
1009                         if (TD_IS_RUNNING(owner)) {
1010                                 sleepq_release(&sx->lock_object);
1011                                 x = SX_READ_VALUE(sx);
1012                                 continue;
1013                         }
1014                 }
1015 #endif
1016
1017                 /*
1018                  * Try to set the SX_LOCK_SHARED_WAITERS flag.  If we
1019                  * fail to set it drop the sleep queue lock and loop
1020                  * back.
1021                  */
1022                 if (!(x & SX_LOCK_SHARED_WAITERS)) {
1023                         if (!atomic_cmpset_ptr(&sx->sx_lock, x,
1024                             x | SX_LOCK_SHARED_WAITERS)) {
1025                                 sleepq_release(&sx->lock_object);
1026                                 x = SX_READ_VALUE(sx);
1027                                 continue;
1028                         }
1029                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
1030                                 CTR2(KTR_LOCK, "%s: %p set shared waiters flag",
1031                                     __func__, sx);
1032                 }
1033
1034                 /*
1035                  * Since we have been unable to acquire the shared lock,
1036                  * we have to sleep.
1037                  */
1038                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1039                         CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
1040                             __func__, sx);
1041
1042 #ifdef KDTRACE_HOOKS
1043                 sleep_time -= lockstat_nsecs(&sx->lock_object);
1044 #endif
1045                 GIANT_SAVE(extra_work);
1046                 sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
1047                     SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
1048                     SLEEPQ_INTERRUPTIBLE : 0), SQ_SHARED_QUEUE);
1049                 if (!(opts & SX_INTERRUPTIBLE))
1050                         sleepq_wait(&sx->lock_object, 0);
1051                 else
1052                         error = sleepq_wait_sig(&sx->lock_object, 0);
1053 #ifdef KDTRACE_HOOKS
1054                 sleep_time += lockstat_nsecs(&sx->lock_object);
1055                 sleep_cnt++;
1056 #endif
1057                 if (error) {
1058                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
1059                                 CTR2(KTR_LOCK,
1060                         "%s: interruptible sleep by %p suspended by signal",
1061                                     __func__, sx);
1062                         break;
1063                 }
1064                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1065                         CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
1066                             __func__, sx);
1067                 x = SX_READ_VALUE(sx);
1068         }
1069 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
1070         if (__predict_true(!extra_work))
1071                 return (error);
1072 #endif
1073 #ifdef KDTRACE_HOOKS
1074         all_time += lockstat_nsecs(&sx->lock_object);
1075         if (sleep_time)
1076                 LOCKSTAT_RECORD4(sx__block, sx, sleep_time,
1077                     LOCKSTAT_READER, (state & SX_LOCK_SHARED) == 0,
1078                     (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
1079         if (lda.spin_cnt > sleep_cnt)
1080                 LOCKSTAT_RECORD4(sx__spin, sx, all_time - sleep_time,
1081                     LOCKSTAT_READER, (state & SX_LOCK_SHARED) == 0,
1082                     (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
1083 #endif
1084         if (error == 0) {
1085                 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
1086                     contested, waittime, file, line, LOCKSTAT_READER);
1087         }
1088         GIANT_RESTORE();
1089         return (error);
1090 }
1091
1092 int
1093 _sx_slock_int(struct sx *sx, int opts LOCK_FILE_LINE_ARG_DEF)
1094 {
1095         uintptr_t x;
1096         int error;
1097
1098         KASSERT(kdb_active != 0 || SCHEDULER_STOPPED() ||
1099             !TD_IS_IDLETHREAD(curthread),
1100             ("sx_slock() by idle thread %p on sx %s @ %s:%d",
1101             curthread, sx->lock_object.lo_name, file, line));
1102         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
1103             ("sx_slock() of destroyed sx @ %s:%d", file, line));
1104         WITNESS_CHECKORDER(&sx->lock_object, LOP_NEWORDER, file, line, NULL);
1105
1106         error = 0;
1107         x = SX_READ_VALUE(sx);
1108         if (__predict_false(LOCKSTAT_OOL_PROFILE_ENABLED(sx__acquire) ||
1109             !__sx_slock_try(sx, &x LOCK_FILE_LINE_ARG)))
1110                 error = _sx_slock_hard(sx, opts, x LOCK_FILE_LINE_ARG);
1111         if (error == 0) {
1112                 LOCK_LOG_LOCK("SLOCK", &sx->lock_object, 0, 0, file, line);
1113                 WITNESS_LOCK(&sx->lock_object, 0, file, line);
1114                 TD_LOCKS_INC(curthread);
1115         }
1116         return (error);
1117 }
1118
1119 int
1120 _sx_slock(struct sx *sx, int opts, const char *file, int line)
1121 {
1122
1123         return (_sx_slock_int(sx, opts LOCK_FILE_LINE_ARG));
1124 }
1125
1126 static bool __always_inline
1127 _sx_sunlock_try(struct sx *sx, uintptr_t *xp)
1128 {
1129
1130         for (;;) {
1131                 /*
1132                  * We should never have sharers while at least one thread
1133                  * holds a shared lock.
1134                  */
1135                 KASSERT(!(*xp & SX_LOCK_SHARED_WAITERS),
1136                     ("%s: waiting sharers", __func__));
1137
1138                 /*
1139                  * See if there is more than one shared lock held.  If
1140                  * so, just drop one and return.
1141                  */
1142                 if (SX_SHARERS(*xp) > 1) {
1143                         if (atomic_fcmpset_rel_ptr(&sx->sx_lock, xp,
1144                             *xp - SX_ONE_SHARER)) {
1145                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1146                                         CTR4(KTR_LOCK,
1147                                             "%s: %p succeeded %p -> %p",
1148                                             __func__, sx, (void *)*xp,
1149                                             (void *)(*xp - SX_ONE_SHARER));
1150                                 return (true);
1151                         }
1152                         continue;
1153                 }
1154
1155                 /*
1156                  * If there aren't any waiters for an exclusive lock,
1157                  * then try to drop it quickly.
1158                  */
1159                 if (!(*xp & SX_LOCK_EXCLUSIVE_WAITERS)) {
1160                         MPASS(*xp == SX_SHARERS_LOCK(1));
1161                         *xp = SX_SHARERS_LOCK(1);
1162                         if (atomic_fcmpset_rel_ptr(&sx->sx_lock,
1163                             xp, SX_LOCK_UNLOCKED)) {
1164                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1165                                         CTR2(KTR_LOCK, "%s: %p last succeeded",
1166                                             __func__, sx);
1167                                 return (true);
1168                         }
1169                         continue;
1170                 }
1171                 break;
1172         }
1173         return (false);
1174 }
1175
1176 static void __noinline
1177 _sx_sunlock_hard(struct sx *sx, uintptr_t x LOCK_FILE_LINE_ARG_DEF)
1178 {
1179         int wakeup_swapper;
1180
1181         if (SCHEDULER_STOPPED())
1182                 return;
1183
1184         LOCKSTAT_PROFILE_RELEASE_RWLOCK(sx__release, sx, LOCKSTAT_READER);
1185
1186         for (;;) {
1187                 if (_sx_sunlock_try(sx, &x))
1188                         break;
1189
1190                 /*
1191                  * At this point, there should just be one sharer with
1192                  * exclusive waiters.
1193                  */
1194                 MPASS(x == (SX_SHARERS_LOCK(1) | SX_LOCK_EXCLUSIVE_WAITERS));
1195
1196                 sleepq_lock(&sx->lock_object);
1197
1198                 /*
1199                  * Wake up semantic here is quite simple:
1200                  * Just wake up all the exclusive waiters.
1201                  * Note that the state of the lock could have changed,
1202                  * so if it fails loop back and retry.
1203                  */
1204                 if (!atomic_cmpset_rel_ptr(&sx->sx_lock,
1205                     SX_SHARERS_LOCK(1) | SX_LOCK_EXCLUSIVE_WAITERS,
1206                     SX_LOCK_UNLOCKED)) {
1207                         sleepq_release(&sx->lock_object);
1208                         x = SX_READ_VALUE(sx);
1209                         continue;
1210                 }
1211                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1212                         CTR2(KTR_LOCK, "%s: %p waking up all thread on"
1213                             "exclusive queue", __func__, sx);
1214                 wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX,
1215                     0, SQ_EXCLUSIVE_QUEUE);
1216                 sleepq_release(&sx->lock_object);
1217                 if (wakeup_swapper)
1218                         kick_proc0();
1219                 break;
1220         }
1221 }
1222
1223 void
1224 _sx_sunlock_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
1225 {
1226         uintptr_t x;
1227
1228         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
1229             ("sx_sunlock() of destroyed sx @ %s:%d", file, line));
1230         _sx_assert(sx, SA_SLOCKED, file, line);
1231         WITNESS_UNLOCK(&sx->lock_object, 0, file, line);
1232         LOCK_LOG_LOCK("SUNLOCK", &sx->lock_object, 0, 0, file, line);
1233
1234         x = SX_READ_VALUE(sx);
1235         if (__predict_false(LOCKSTAT_OOL_PROFILE_ENABLED(sx__release) ||
1236             !_sx_sunlock_try(sx, &x)))
1237                 _sx_sunlock_hard(sx, x LOCK_FILE_LINE_ARG);
1238
1239         TD_LOCKS_DEC(curthread);
1240 }
1241
1242 void
1243 _sx_sunlock(struct sx *sx, const char *file, int line)
1244 {
1245
1246         _sx_sunlock_int(sx LOCK_FILE_LINE_ARG);
1247 }
1248
1249 #ifdef INVARIANT_SUPPORT
1250 #ifndef INVARIANTS
1251 #undef  _sx_assert
1252 #endif
1253
1254 /*
1255  * In the non-WITNESS case, sx_assert() can only detect that at least
1256  * *some* thread owns an slock, but it cannot guarantee that *this*
1257  * thread owns an slock.
1258  */
1259 void
1260 _sx_assert(const struct sx *sx, int what, const char *file, int line)
1261 {
1262 #ifndef WITNESS
1263         int slocked = 0;
1264 #endif
1265
1266         if (panicstr != NULL)
1267                 return;
1268         switch (what) {
1269         case SA_SLOCKED:
1270         case SA_SLOCKED | SA_NOTRECURSED:
1271         case SA_SLOCKED | SA_RECURSED:
1272 #ifndef WITNESS
1273                 slocked = 1;
1274                 /* FALLTHROUGH */
1275 #endif
1276         case SA_LOCKED:
1277         case SA_LOCKED | SA_NOTRECURSED:
1278         case SA_LOCKED | SA_RECURSED:
1279 #ifdef WITNESS
1280                 witness_assert(&sx->lock_object, what, file, line);
1281 #else
1282                 /*
1283                  * If some other thread has an exclusive lock or we
1284                  * have one and are asserting a shared lock, fail.
1285                  * Also, if no one has a lock at all, fail.
1286                  */
1287                 if (sx->sx_lock == SX_LOCK_UNLOCKED ||
1288                     (!(sx->sx_lock & SX_LOCK_SHARED) && (slocked ||
1289                     sx_xholder(sx) != curthread)))
1290                         panic("Lock %s not %slocked @ %s:%d\n",
1291                             sx->lock_object.lo_name, slocked ? "share " : "",
1292                             file, line);
1293
1294                 if (!(sx->sx_lock & SX_LOCK_SHARED)) {
1295                         if (sx_recursed(sx)) {
1296                                 if (what & SA_NOTRECURSED)
1297                                         panic("Lock %s recursed @ %s:%d\n",
1298                                             sx->lock_object.lo_name, file,
1299                                             line);
1300                         } else if (what & SA_RECURSED)
1301                                 panic("Lock %s not recursed @ %s:%d\n",
1302                                     sx->lock_object.lo_name, file, line);
1303                 }
1304 #endif
1305                 break;
1306         case SA_XLOCKED:
1307         case SA_XLOCKED | SA_NOTRECURSED:
1308         case SA_XLOCKED | SA_RECURSED:
1309                 if (sx_xholder(sx) != curthread)
1310                         panic("Lock %s not exclusively locked @ %s:%d\n",
1311                             sx->lock_object.lo_name, file, line);
1312                 if (sx_recursed(sx)) {
1313                         if (what & SA_NOTRECURSED)
1314                                 panic("Lock %s recursed @ %s:%d\n",
1315                                     sx->lock_object.lo_name, file, line);
1316                 } else if (what & SA_RECURSED)
1317                         panic("Lock %s not recursed @ %s:%d\n",
1318                             sx->lock_object.lo_name, file, line);
1319                 break;
1320         case SA_UNLOCKED:
1321 #ifdef WITNESS
1322                 witness_assert(&sx->lock_object, what, file, line);
1323 #else
1324                 /*
1325                  * If we hold an exclusve lock fail.  We can't
1326                  * reliably check to see if we hold a shared lock or
1327                  * not.
1328                  */
1329                 if (sx_xholder(sx) == curthread)
1330                         panic("Lock %s exclusively locked @ %s:%d\n",
1331                             sx->lock_object.lo_name, file, line);
1332 #endif
1333                 break;
1334         default:
1335                 panic("Unknown sx lock assertion: %d @ %s:%d", what, file,
1336                     line);
1337         }
1338 }
1339 #endif  /* INVARIANT_SUPPORT */
1340
1341 #ifdef DDB
1342 static void
1343 db_show_sx(const struct lock_object *lock)
1344 {
1345         struct thread *td;
1346         const struct sx *sx;
1347
1348         sx = (const struct sx *)lock;
1349
1350         db_printf(" state: ");
1351         if (sx->sx_lock == SX_LOCK_UNLOCKED)
1352                 db_printf("UNLOCKED\n");
1353         else if (sx->sx_lock == SX_LOCK_DESTROYED) {
1354                 db_printf("DESTROYED\n");
1355                 return;
1356         } else if (sx->sx_lock & SX_LOCK_SHARED)
1357                 db_printf("SLOCK: %ju\n", (uintmax_t)SX_SHARERS(sx->sx_lock));
1358         else {
1359                 td = sx_xholder(sx);
1360                 db_printf("XLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1361                     td->td_tid, td->td_proc->p_pid, td->td_name);
1362                 if (sx_recursed(sx))
1363                         db_printf(" recursed: %d\n", sx->sx_recurse);
1364         }
1365
1366         db_printf(" waiters: ");
1367         switch(sx->sx_lock &
1368             (SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS)) {
1369         case SX_LOCK_SHARED_WAITERS:
1370                 db_printf("shared\n");
1371                 break;
1372         case SX_LOCK_EXCLUSIVE_WAITERS:
1373                 db_printf("exclusive\n");
1374                 break;
1375         case SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS:
1376                 db_printf("exclusive and shared\n");
1377                 break;
1378         default:
1379                 db_printf("none\n");
1380         }
1381 }
1382
1383 /*
1384  * Check to see if a thread that is blocked on a sleep queue is actually
1385  * blocked on an sx lock.  If so, output some details and return true.
1386  * If the lock has an exclusive owner, return that in *ownerp.
1387  */
1388 int
1389 sx_chain(struct thread *td, struct thread **ownerp)
1390 {
1391         struct sx *sx;
1392
1393         /*
1394          * Check to see if this thread is blocked on an sx lock.
1395          * First, we check the lock class.  If that is ok, then we
1396          * compare the lock name against the wait message.
1397          */
1398         sx = td->td_wchan;
1399         if (LOCK_CLASS(&sx->lock_object) != &lock_class_sx ||
1400             sx->lock_object.lo_name != td->td_wmesg)
1401                 return (0);
1402
1403         /* We think we have an sx lock, so output some details. */
1404         db_printf("blocked on sx \"%s\" ", td->td_wmesg);
1405         *ownerp = sx_xholder(sx);
1406         if (sx->sx_lock & SX_LOCK_SHARED)
1407                 db_printf("SLOCK (count %ju)\n",
1408                     (uintmax_t)SX_SHARERS(sx->sx_lock));
1409         else
1410                 db_printf("XLOCK\n");
1411         return (1);
1412 }
1413 #endif