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                 goto out;
 485
 486         /*
 487          * Lock the sleep queue so we can read the waiters bits
 488          * without any races and wakeup any shared waiters.
 489          */
 490         sleepq_lock(&sx->lock_object);
 491
 492         /*
 493          * Preserve SX_LOCK_EXCLUSIVE_WAITERS while downgraded to a single
 494          * shared lock.  If there are any shared waiters, wake them up.
 495          */
 496         wakeup_swapper = 0;
 497         x = sx->sx_lock;
 498         atomic_store_rel_ptr(&sx->sx_lock, SX_SHARERS_LOCK(1) |
 499             (x & SX_LOCK_EXCLUSIVE_WAITERS));
 500         if (x & SX_LOCK_SHARED_WAITERS)
 501                 wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX,
 502                     0, SQ_SHARED_QUEUE);
 503         sleepq_release(&sx->lock_object);
 504
 505         if (wakeup_swapper)
 506                 kick_proc0();
 507
 508 out:
 509         LOCK_LOG_LOCK("XDOWNGRADE", &sx->lock_object, 0, 0, file, line);
 510         LOCKSTAT_RECORD0(sx__downgrade, sx);
 511 }
 512
 513 void
 514 sx_downgrade_(struct sx *sx, const char *file, int line)
 515 {
 516
 517         sx_downgrade_int(sx LOCK_FILE_LINE_ARG);
 518 }
 519
 520 /*
 521  * This function represents the so-called 'hard case' for sx_xlock
 522  * operation.  All 'easy case' failures are redirected to this.  Note
 523  * that ideally this would be a static function, but it needs to be
 524  * accessible from at least sx.h.
 525  */
 526 int
 527 _sx_xlock_hard(struct sx *sx, uintptr_t x, int opts LOCK_FILE_LINE_ARG_DEF)
 528 {
 529         GIANT_DECLARE;
 530         uintptr_t tid;
 531 #ifdef ADAPTIVE_SX
 532         volatile struct thread *owner;
 533         u_int i, n, spintries = 0;
 534 #endif
 535 #ifdef LOCK_PROFILING
 536         uint64_t waittime = 0;
 537         int contested = 0;
 538 #endif
 539         int error = 0;
 540 #if defined(ADAPTIVE_SX) || defined(KDTRACE_HOOKS)
 541         struct lock_delay_arg lda;
 542 #endif
 543 #ifdef  KDTRACE_HOOKS
 544         u_int sleep_cnt = 0;
 545         int64_t sleep_time = 0;
 546         int64_t all_time = 0;
 547 #endif
 548 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
 549         uintptr_t state;
 550 #endif
 551         int extra_work = 0;
 552
 553         tid = (uintptr_t)curthread;
 554         if (SCHEDULER_STOPPED())
 555                 return (0);
 556
 557 #if defined(ADAPTIVE_SX)
 558         lock_delay_arg_init(&lda, &sx_delay);
 559 #elif defined(KDTRACE_HOOKS)
 560         lock_delay_arg_init(&lda, NULL);
 561 #endif
 562
 563         if (__predict_false(x == SX_LOCK_UNLOCKED))
 564                 x = SX_READ_VALUE(sx);
 565
 566         /* If we already hold an exclusive lock, then recurse. */
 567         if (__predict_false(lv_sx_owner(x) == (struct thread *)tid)) {
 568                 KASSERT((sx->lock_object.lo_flags & LO_RECURSABLE) != 0,
 569             ("_sx_xlock_hard: recursed on non-recursive sx %s @ %s:%d\n",
 570                     sx->lock_object.lo_name, file, line));
 571                 sx->sx_recurse++;
 572                 atomic_set_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
 573                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 574                         CTR2(KTR_LOCK, "%s: %p recursing", __func__, sx);
 575                 return (0);
 576         }
 577
 578         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 579                 CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__,
 580                     sx->lock_object.lo_name, (void *)sx->sx_lock, file, line);
 581
 582 #ifdef HWPMC_HOOKS
 583         PMC_SOFT_CALL( , , lock, failed);
 584 #endif
 585         lock_profile_obtain_lock_failed(&sx->lock_object, &contested,
 586             &waittime);
 587
 588 #ifdef LOCK_PROFILING
 589         extra_work = 1;
 590         state = x;
 591 #elif defined(KDTRACE_HOOKS)
 592         extra_work = lockstat_enabled;
 593         if (__predict_false(extra_work)) {
 594                 all_time -= lockstat_nsecs(&sx->lock_object);
 595                 state = x;
 596         }
 597 #endif
 598
 599         for (;;) {
 600                 if (x == SX_LOCK_UNLOCKED) {
 601                         if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid))
 602                                 break;
 603                         continue;
 604                 }
 605 #ifdef KDTRACE_HOOKS
 606                 lda.spin_cnt++;
 607 #endif
 608 #ifdef ADAPTIVE_SX
 609                 /*
 610                  * If the lock is write locked and the owner is
 611                  * running on another CPU, spin until the owner stops
 612                  * running or the state of the lock changes.
 613                  */
 614                 if ((sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) {
 615                         if ((x & SX_LOCK_SHARED) == 0) {
 616                                 owner = lv_sx_owner(x);
 617                                 if (TD_IS_RUNNING(owner)) {
 618                                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 619                                                 CTR3(KTR_LOCK,
 620                                             "%s: spinning on %p held by %p",
 621                                                     __func__, sx, owner);
 622                                         KTR_STATE1(KTR_SCHED, "thread",
 623                                             sched_tdname(curthread), "spinning",
 624                                             "lockname:\"%s\"",
 625                                             sx->lock_object.lo_name);
 626                                         GIANT_SAVE(extra_work);
 627                                         do {
 628                                                 lock_delay(&lda);
 629                                                 x = SX_READ_VALUE(sx);
 630                                                 owner = lv_sx_owner(x);
 631                                         } while (owner != NULL &&
 632                                                     TD_IS_RUNNING(owner));
 633                                         KTR_STATE0(KTR_SCHED, "thread",
 634                                             sched_tdname(curthread), "running");
 635                                         continue;
 636                                 }
 637                         } else if (SX_SHARERS(x) && spintries < asx_retries) {
 638                                 KTR_STATE1(KTR_SCHED, "thread",
 639                                     sched_tdname(curthread), "spinning",
 640                                     "lockname:\"%s\"", sx->lock_object.lo_name);
 641                                 GIANT_SAVE(extra_work);
 642                                 spintries++;
 643                                 for (i = 0; i < asx_loops; i += n) {
 644                                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 645                                                 CTR4(KTR_LOCK,
 646                                     "%s: shared spinning on %p with %u and %u",
 647                                                     __func__, sx, spintries, i);
 648                                         n = SX_SHARERS(x);
 649                                         lock_delay_spin(n);
 650                                         x = SX_READ_VALUE(sx);
 651                                         if ((x & SX_LOCK_SHARED) == 0 ||
 652                                             SX_SHARERS(x) == 0)
 653                                                 break;
 654                                 }
 655 #ifdef KDTRACE_HOOKS
 656                                 lda.spin_cnt += i;
 657 #endif
 658                                 KTR_STATE0(KTR_SCHED, "thread",
 659                                     sched_tdname(curthread), "running");
 660                                 if (i != asx_loops)
 661                                         continue;
 662                         }
 663                 }
 664 #endif
 665
 666                 sleepq_lock(&sx->lock_object);
 667                 x = SX_READ_VALUE(sx);
 668 retry_sleepq:
 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_fcmpset_acq_ptr(&sx->sx_lock, &x,
 709                             tid | SX_LOCK_EXCLUSIVE_WAITERS))
 710                                 goto retry_sleepq;
 711                         sleepq_release(&sx->lock_object);
 712                         CTR2(KTR_LOCK, "%s: %p claimed by new writer",
 713                             __func__, sx);
 714                         break;
 715                 }
 716
 717                 /*
 718                  * Try to set the SX_LOCK_EXCLUSIVE_WAITERS.  If we fail,
 719                  * than loop back and retry.
 720                  */
 721                 if (!(x & SX_LOCK_EXCLUSIVE_WAITERS)) {
 722                         if (!atomic_fcmpset_ptr(&sx->sx_lock, &x,
 723                             x | SX_LOCK_EXCLUSIVE_WAITERS)) {
 724                                 goto retry_sleepq;
 725                         }
 726                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 727                                 CTR2(KTR_LOCK, "%s: %p set excl waiters flag",
 728                                     __func__, sx);
 729                 }
 730
 731                 /*
 732                  * Since we have been unable to acquire the exclusive
 733                  * lock and the exclusive waiters flag is set, we have
 734                  * to sleep.
 735                  */
 736                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 737                         CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
 738                             __func__, sx);
 739
 740 #ifdef KDTRACE_HOOKS
 741                 sleep_time -= lockstat_nsecs(&sx->lock_object);
 742 #endif
 743                 GIANT_SAVE(extra_work);
 744                 sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
 745                     SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
 746                     SLEEPQ_INTERRUPTIBLE : 0), SQ_EXCLUSIVE_QUEUE);
 747                 if (!(opts & SX_INTERRUPTIBLE))
 748                         sleepq_wait(&sx->lock_object, 0);
 749                 else
 750                         error = sleepq_wait_sig(&sx->lock_object, 0);
 751 #ifdef KDTRACE_HOOKS
 752                 sleep_time += lockstat_nsecs(&sx->lock_object);
 753                 sleep_cnt++;
 754 #endif
 755                 if (error) {
 756                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 757                                 CTR2(KTR_LOCK,
 758                         "%s: interruptible sleep by %p suspended by signal",
 759                                     __func__, sx);
 760                         break;
 761                 }
 762                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 763                         CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
 764                             __func__, sx);
 765                 x = SX_READ_VALUE(sx);
 766         }
 767 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
 768         if (__predict_true(!extra_work))
 769                 return (error);
 770 #endif
 771 #ifdef KDTRACE_HOOKS
 772         all_time += lockstat_nsecs(&sx->lock_object);
 773         if (sleep_time)
 774                 LOCKSTAT_RECORD4(sx__block, sx, sleep_time,
 775                     LOCKSTAT_WRITER, (state & SX_LOCK_SHARED) == 0,
 776                     (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
 777         if (lda.spin_cnt > sleep_cnt)
 778                 LOCKSTAT_RECORD4(sx__spin, sx, all_time - sleep_time,
 779                     LOCKSTAT_WRITER, (state & SX_LOCK_SHARED) == 0,
 780                     (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
 781 #endif
 782         if (!error)
 783                 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
 784                     contested, waittime, file, line, LOCKSTAT_WRITER);
 785         GIANT_RESTORE();
 786         return (error);
 787 }
 788
 789 /*
 790  * This function represents the so-called 'hard case' for sx_xunlock
 791  * operation.  All 'easy case' failures are redirected to this.  Note
 792  * that ideally this would be a static function, but it needs to be
 793  * accessible from at least sx.h.
 794  */
 795 void
 796 _sx_xunlock_hard(struct sx *sx, uintptr_t x LOCK_FILE_LINE_ARG_DEF)
 797 {
 798         uintptr_t tid, setx;
 799         int queue, wakeup_swapper;
 800
 801         if (SCHEDULER_STOPPED())
 802                 return;
 803
 804         tid = (uintptr_t)curthread;
 805
 806         if (__predict_false(x == tid))
 807                 x = SX_READ_VALUE(sx);
 808
 809         MPASS(!(x & SX_LOCK_SHARED));
 810
 811         if (__predict_false(x & SX_LOCK_RECURSED)) {
 812                 /* The lock is recursed, unrecurse one level. */
 813                 if ((--sx->sx_recurse) == 0)
 814                         atomic_clear_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
 815                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 816                         CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, sx);
 817                 return;
 818         }
 819
 820         LOCKSTAT_PROFILE_RELEASE_RWLOCK(sx__release, sx, LOCKSTAT_WRITER);
 821         if (x == tid &&
 822             atomic_cmpset_rel_ptr(&sx->sx_lock, tid, SX_LOCK_UNLOCKED))
 823                 return;
 824
 825         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 826                 CTR2(KTR_LOCK, "%s: %p contested", __func__, sx);
 827
 828         sleepq_lock(&sx->lock_object);
 829         x = SX_READ_VALUE(sx);
 830         MPASS(x & (SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS));
 831
 832         /*
 833          * The wake up algorithm here is quite simple and probably not
 834          * ideal.  It gives precedence to shared waiters if they are
 835          * present.  For this condition, we have to preserve the
 836          * state of the exclusive waiters flag.
 837          * If interruptible sleeps left the shared queue empty avoid a
 838          * starvation for the threads sleeping on the exclusive queue by giving
 839          * them precedence and cleaning up the shared waiters bit anyway.
 840          */
 841         setx = SX_LOCK_UNLOCKED;
 842         queue = SQ_EXCLUSIVE_QUEUE;
 843         if ((x & SX_LOCK_SHARED_WAITERS) != 0 &&
 844             sleepq_sleepcnt(&sx->lock_object, SQ_SHARED_QUEUE) != 0) {
 845                 queue = SQ_SHARED_QUEUE;
 846                 setx |= (x & SX_LOCK_EXCLUSIVE_WAITERS);
 847         }
 848         atomic_store_rel_ptr(&sx->sx_lock, setx);
 849
 850         /* Wake up all the waiters for the specific queue. */
 851         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 852                 CTR3(KTR_LOCK, "%s: %p waking up all threads on %s queue",
 853                     __func__, sx, queue == SQ_SHARED_QUEUE ? "shared" :
 854                     "exclusive");
 855
 856         wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX, 0,
 857             queue);
 858         sleepq_release(&sx->lock_object);
 859         if (wakeup_swapper)
 860                 kick_proc0();
 861 }
 862
 863 static bool __always_inline
 864 __sx_slock_try(struct sx *sx, uintptr_t *xp LOCK_FILE_LINE_ARG_DEF)
 865 {
 866
 867         /*
 868          * If no other thread has an exclusive lock then try to bump up
 869          * the count of sharers.  Since we have to preserve the state
 870          * of SX_LOCK_EXCLUSIVE_WAITERS, if we fail to acquire the
 871          * shared lock loop back and retry.
 872          */
 873         while (*xp & SX_LOCK_SHARED) {
 874                 MPASS(!(*xp & SX_LOCK_SHARED_WAITERS));
 875                 if (atomic_fcmpset_acq_ptr(&sx->sx_lock, xp,
 876                     *xp + SX_ONE_SHARER)) {
 877                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 878                                 CTR4(KTR_LOCK, "%s: %p succeed %p -> %p",
 879                                     __func__, sx, (void *)*xp,
 880                                     (void *)(*xp + SX_ONE_SHARER));
 881                         return (true);
 882                 }
 883         }
 884         return (false);
 885 }
 886
 887 static int __noinline
 888 _sx_slock_hard(struct sx *sx, int opts, uintptr_t x LOCK_FILE_LINE_ARG_DEF)
 889 {
 890         GIANT_DECLARE;
 891 #ifdef ADAPTIVE_SX
 892         volatile struct thread *owner;
 893 #endif
 894 #ifdef LOCK_PROFILING
 895         uint64_t waittime = 0;
 896         int contested = 0;
 897 #endif
 898         int error = 0;
 899 #if defined(ADAPTIVE_SX) || defined(KDTRACE_HOOKS)
 900         struct lock_delay_arg lda;
 901 #endif
 902 #ifdef KDTRACE_HOOKS
 903         u_int sleep_cnt = 0;
 904         int64_t sleep_time = 0;
 905         int64_t all_time = 0;
 906 #endif
 907 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
 908         uintptr_t state;
 909 #endif
 910         int extra_work = 0;
 911
 912         if (SCHEDULER_STOPPED())
 913                 return (0);
 914
 915 #if defined(ADAPTIVE_SX)
 916         lock_delay_arg_init(&lda, &sx_delay);
 917 #elif defined(KDTRACE_HOOKS)
 918         lock_delay_arg_init(&lda, NULL);
 919 #endif
 920
 921 #ifdef HWPMC_HOOKS
 922         PMC_SOFT_CALL( , , lock, failed);
 923 #endif
 924         lock_profile_obtain_lock_failed(&sx->lock_object, &contested,
 925             &waittime);
 926
 927 #ifdef LOCK_PROFILING
 928         extra_work = 1;
 929         state = x;
 930 #elif defined(KDTRACE_HOOKS)
 931         extra_work = lockstat_enabled;
 932         if (__predict_false(extra_work)) {
 933                 all_time -= lockstat_nsecs(&sx->lock_object);
 934                 state = x;
 935         }
 936 #endif
 937
 938         /*
 939          * As with rwlocks, we don't make any attempt to try to block
 940          * shared locks once there is an exclusive waiter.
 941          */
 942         for (;;) {
 943                 if (__sx_slock_try(sx, &x LOCK_FILE_LINE_ARG))
 944                         break;
 945 #ifdef KDTRACE_HOOKS
 946                 lda.spin_cnt++;
 947 #endif
 948
 949 #ifdef ADAPTIVE_SX
 950                 /*
 951                  * If the owner is running on another CPU, spin until
 952                  * the owner stops running or the state of the lock
 953                  * changes.
 954                  */
 955                 if ((sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) {
 956                         owner = lv_sx_owner(x);
 957                         if (TD_IS_RUNNING(owner)) {
 958                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 959                                         CTR3(KTR_LOCK,
 960                                             "%s: spinning on %p held by %p",
 961                                             __func__, sx, owner);
 962                                 KTR_STATE1(KTR_SCHED, "thread",
 963                                     sched_tdname(curthread), "spinning",
 964                                     "lockname:\"%s\"", sx->lock_object.lo_name);
 965                                 GIANT_SAVE(extra_work);
 966                                 do {
 967                                         lock_delay(&lda);
 968                                         x = SX_READ_VALUE(sx);
 969                                         owner = lv_sx_owner(x);
 970                                 } while (owner != NULL && TD_IS_RUNNING(owner));
 971                                 KTR_STATE0(KTR_SCHED, "thread",
 972                                     sched_tdname(curthread), "running");
 973                                 continue;
 974                         }
 975                 }
 976 #endif
 977
 978                 /*
 979                  * Some other thread already has an exclusive lock, so
 980                  * start the process of blocking.
 981                  */
 982                 sleepq_lock(&sx->lock_object);
 983                 x = SX_READ_VALUE(sx);
 984 retry_sleepq:
 985                 /*
 986                  * The lock could have been released while we spun.
 987                  * In this case loop back and retry.
 988                  */
 989                 if (x & SX_LOCK_SHARED) {
 990                         sleepq_release(&sx->lock_object);
 991                         continue;
 992                 }
 993
 994 #ifdef ADAPTIVE_SX
 995                 /*
 996                  * If the owner is running on another CPU, spin until
 997                  * the owner stops running or the state of the lock
 998                  * changes.
 999                  */
1000                 if (!(x & SX_LOCK_SHARED) &&
1001                     (sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) {
1002                         owner = (struct thread *)SX_OWNER(x);
1003                         if (TD_IS_RUNNING(owner)) {
1004                                 sleepq_release(&sx->lock_object);
1005                                 x = SX_READ_VALUE(sx);
1006                                 continue;
1007                         }
1008                 }
1009 #endif
1010
1011                 /*
1012                  * Try to set the SX_LOCK_SHARED_WAITERS flag.  If we
1013                  * fail to set it drop the sleep queue lock and loop
1014                  * back.
1015                  */
1016                 if (!(x & SX_LOCK_SHARED_WAITERS)) {
1017                         if (!atomic_fcmpset_ptr(&sx->sx_lock, &x,
1018                             x | SX_LOCK_SHARED_WAITERS))
1019                                 goto retry_sleepq;
1020                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
1021                                 CTR2(KTR_LOCK, "%s: %p set shared waiters flag",
1022                                     __func__, sx);
1023                 }
1024
1025                 /*
1026                  * Since we have been unable to acquire the shared lock,
1027                  * we have to sleep.
1028                  */
1029                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1030                         CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
1031                             __func__, sx);
1032
1033 #ifdef KDTRACE_HOOKS
1034                 sleep_time -= lockstat_nsecs(&sx->lock_object);
1035 #endif
1036                 GIANT_SAVE(extra_work);
1037                 sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
1038                     SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
1039                     SLEEPQ_INTERRUPTIBLE : 0), SQ_SHARED_QUEUE);
1040                 if (!(opts & SX_INTERRUPTIBLE))
1041                         sleepq_wait(&sx->lock_object, 0);
1042                 else
1043                         error = sleepq_wait_sig(&sx->lock_object, 0);
1044 #ifdef KDTRACE_HOOKS
1045                 sleep_time += lockstat_nsecs(&sx->lock_object);
1046                 sleep_cnt++;
1047 #endif
1048                 if (error) {
1049                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
1050                                 CTR2(KTR_LOCK,
1051                         "%s: interruptible sleep by %p suspended by signal",
1052                                     __func__, sx);
1053                         break;
1054                 }
1055                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1056                         CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
1057                             __func__, sx);
1058                 x = SX_READ_VALUE(sx);
1059         }
1060 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
1061         if (__predict_true(!extra_work))
1062                 return (error);
1063 #endif
1064 #ifdef KDTRACE_HOOKS
1065         all_time += lockstat_nsecs(&sx->lock_object);
1066         if (sleep_time)
1067                 LOCKSTAT_RECORD4(sx__block, sx, sleep_time,
1068                     LOCKSTAT_READER, (state & SX_LOCK_SHARED) == 0,
1069                     (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
1070         if (lda.spin_cnt > sleep_cnt)
1071                 LOCKSTAT_RECORD4(sx__spin, sx, all_time - sleep_time,
1072                     LOCKSTAT_READER, (state & SX_LOCK_SHARED) == 0,
1073                     (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
1074 #endif
1075         if (error == 0) {
1076                 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
1077                     contested, waittime, file, line, LOCKSTAT_READER);
1078         }
1079         GIANT_RESTORE();
1080         return (error);
1081 }
1082
1083 int
1084 _sx_slock_int(struct sx *sx, int opts LOCK_FILE_LINE_ARG_DEF)
1085 {
1086         uintptr_t x;
1087         int error;
1088
1089         KASSERT(kdb_active != 0 || SCHEDULER_STOPPED() ||
1090             !TD_IS_IDLETHREAD(curthread),
1091             ("sx_slock() by idle thread %p on sx %s @ %s:%d",
1092             curthread, sx->lock_object.lo_name, file, line));
1093         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
1094             ("sx_slock() of destroyed sx @ %s:%d", file, line));
1095         WITNESS_CHECKORDER(&sx->lock_object, LOP_NEWORDER, file, line, NULL);
1096
1097         error = 0;
1098         x = SX_READ_VALUE(sx);
1099         if (__predict_false(LOCKSTAT_OOL_PROFILE_ENABLED(sx__acquire) ||
1100             !__sx_slock_try(sx, &x LOCK_FILE_LINE_ARG)))
1101                 error = _sx_slock_hard(sx, opts, x LOCK_FILE_LINE_ARG);
1102         if (error == 0) {
1103                 LOCK_LOG_LOCK("SLOCK", &sx->lock_object, 0, 0, file, line);
1104                 WITNESS_LOCK(&sx->lock_object, 0, file, line);
1105                 TD_LOCKS_INC(curthread);
1106         }
1107         return (error);
1108 }
1109
1110 int
1111 _sx_slock(struct sx *sx, int opts, const char *file, int line)
1112 {
1113
1114         return (_sx_slock_int(sx, opts LOCK_FILE_LINE_ARG));
1115 }
1116
1117 static bool __always_inline
1118 _sx_sunlock_try(struct sx *sx, uintptr_t *xp)
1119 {
1120
1121         for (;;) {
1122                 /*
1123                  * We should never have sharers while at least one thread
1124                  * holds a shared lock.
1125                  */
1126                 KASSERT(!(*xp & SX_LOCK_SHARED_WAITERS),
1127                     ("%s: waiting sharers", __func__));
1128
1129                 /*
1130                  * See if there is more than one shared lock held.  If
1131                  * so, just drop one and return.
1132                  */
1133                 if (SX_SHARERS(*xp) > 1) {
1134                         if (atomic_fcmpset_rel_ptr(&sx->sx_lock, xp,
1135                             *xp - SX_ONE_SHARER)) {
1136                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1137                                         CTR4(KTR_LOCK,
1138                                             "%s: %p succeeded %p -> %p",
1139                                             __func__, sx, (void *)*xp,
1140                                             (void *)(*xp - SX_ONE_SHARER));
1141                                 return (true);
1142                         }
1143                         continue;
1144                 }
1145
1146                 /*
1147                  * If there aren't any waiters for an exclusive lock,
1148                  * then try to drop it quickly.
1149                  */
1150                 if (!(*xp & SX_LOCK_EXCLUSIVE_WAITERS)) {
1151                         MPASS(*xp == SX_SHARERS_LOCK(1));
1152                         *xp = SX_SHARERS_LOCK(1);
1153                         if (atomic_fcmpset_rel_ptr(&sx->sx_lock,
1154                             xp, SX_LOCK_UNLOCKED)) {
1155                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1156                                         CTR2(KTR_LOCK, "%s: %p last succeeded",
1157                                             __func__, sx);
1158                                 return (true);
1159                         }
1160                         continue;
1161                 }
1162                 break;
1163         }
1164         return (false);
1165 }
1166
1167 static void __noinline
1168 _sx_sunlock_hard(struct sx *sx, uintptr_t x LOCK_FILE_LINE_ARG_DEF)
1169 {
1170         int wakeup_swapper;
1171         uintptr_t setx;
1172
1173         if (SCHEDULER_STOPPED())
1174                 return;
1175
1176         if (_sx_sunlock_try(sx, &x))
1177                 goto out_lockstat;
1178
1179         /*
1180          * At this point, there should just be one sharer with
1181          * exclusive waiters.
1182          */
1183         MPASS(x == (SX_SHARERS_LOCK(1) | SX_LOCK_EXCLUSIVE_WAITERS));
1184
1185         sleepq_lock(&sx->lock_object);
1186         x = SX_READ_VALUE(sx);
1187         for (;;) {
1188                 MPASS(x & SX_LOCK_EXCLUSIVE_WAITERS);
1189                 MPASS(!(x & SX_LOCK_SHARED_WAITERS));
1190                 /*
1191                  * Wake up semantic here is quite simple:
1192                  * Just wake up all the exclusive waiters.
1193                  * Note that the state of the lock could have changed,
1194                  * so if it fails loop back and retry.
1195                  */
1196                 setx = x - SX_ONE_SHARER;
1197                 setx &= ~SX_LOCK_EXCLUSIVE_WAITERS;
1198                 if (!atomic_fcmpset_rel_ptr(&sx->sx_lock, &x, setx))
1199                         continue;
1200                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1201                         CTR2(KTR_LOCK, "%s: %p waking up all thread on"
1202                             "exclusive queue", __func__, sx);
1203                 wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX,
1204                     0, SQ_EXCLUSIVE_QUEUE);
1205                 break;
1206         }
1207         sleepq_release(&sx->lock_object);
1208         if (wakeup_swapper)
1209                 kick_proc0();
1210 out_lockstat:
1211         LOCKSTAT_PROFILE_RELEASE_RWLOCK(sx__release, sx, LOCKSTAT_READER);
1212 }
1213
1214 void
1215 _sx_sunlock_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
1216 {
1217         uintptr_t x;
1218
1219         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
1220             ("sx_sunlock() of destroyed sx @ %s:%d", file, line));
1221         _sx_assert(sx, SA_SLOCKED, file, line);
1222         WITNESS_UNLOCK(&sx->lock_object, 0, file, line);
1223         LOCK_LOG_LOCK("SUNLOCK", &sx->lock_object, 0, 0, file, line);
1224
1225         x = SX_READ_VALUE(sx);
1226         if (__predict_false(LOCKSTAT_OOL_PROFILE_ENABLED(sx__release) ||
1227             !_sx_sunlock_try(sx, &x)))
1228                 _sx_sunlock_hard(sx, x LOCK_FILE_LINE_ARG);
1229
1230         TD_LOCKS_DEC(curthread);
1231 }
1232
1233 void
1234 _sx_sunlock(struct sx *sx, const char *file, int line)
1235 {
1236
1237         _sx_sunlock_int(sx LOCK_FILE_LINE_ARG);
1238 }
1239
1240 #ifdef INVARIANT_SUPPORT
1241 #ifndef INVARIANTS
1242 #undef  _sx_assert
1243 #endif
1244
1245 /*
1246  * In the non-WITNESS case, sx_assert() can only detect that at least
1247  * *some* thread owns an slock, but it cannot guarantee that *this*
1248  * thread owns an slock.
1249  */
1250 void
1251 _sx_assert(const struct sx *sx, int what, const char *file, int line)
1252 {
1253 #ifndef WITNESS
1254         int slocked = 0;
1255 #endif
1256
1257         if (panicstr != NULL)
1258                 return;
1259         switch (what) {
1260         case SA_SLOCKED:
1261         case SA_SLOCKED | SA_NOTRECURSED:
1262         case SA_SLOCKED | SA_RECURSED:
1263 #ifndef WITNESS
1264                 slocked = 1;
1265                 /* FALLTHROUGH */
1266 #endif
1267         case SA_LOCKED:
1268         case SA_LOCKED | SA_NOTRECURSED:
1269         case SA_LOCKED | SA_RECURSED:
1270 #ifdef WITNESS
1271                 witness_assert(&sx->lock_object, what, file, line);
1272 #else
1273                 /*
1274                  * If some other thread has an exclusive lock or we
1275                  * have one and are asserting a shared lock, fail.
1276                  * Also, if no one has a lock at all, fail.
1277                  */
1278                 if (sx->sx_lock == SX_LOCK_UNLOCKED ||
1279                     (!(sx->sx_lock & SX_LOCK_SHARED) && (slocked ||
1280                     sx_xholder(sx) != curthread)))
1281                         panic("Lock %s not %slocked @ %s:%d\n",
1282                             sx->lock_object.lo_name, slocked ? "share " : "",
1283                             file, line);
1284
1285                 if (!(sx->sx_lock & SX_LOCK_SHARED)) {
1286                         if (sx_recursed(sx)) {
1287                                 if (what & SA_NOTRECURSED)
1288                                         panic("Lock %s recursed @ %s:%d\n",
1289                                             sx->lock_object.lo_name, file,
1290                                             line);
1291                         } else if (what & SA_RECURSED)
1292                                 panic("Lock %s not recursed @ %s:%d\n",
1293                                     sx->lock_object.lo_name, file, line);
1294                 }
1295 #endif
1296                 break;
1297         case SA_XLOCKED:
1298         case SA_XLOCKED | SA_NOTRECURSED:
1299         case SA_XLOCKED | SA_RECURSED:
1300                 if (sx_xholder(sx) != curthread)
1301                         panic("Lock %s not exclusively locked @ %s:%d\n",
1302                             sx->lock_object.lo_name, file, line);
1303                 if (sx_recursed(sx)) {
1304                         if (what & SA_NOTRECURSED)
1305                                 panic("Lock %s recursed @ %s:%d\n",
1306                                     sx->lock_object.lo_name, file, line);
1307                 } else if (what & SA_RECURSED)
1308                         panic("Lock %s not recursed @ %s:%d\n",
1309                             sx->lock_object.lo_name, file, line);
1310                 break;
1311         case SA_UNLOCKED:
1312 #ifdef WITNESS
1313                 witness_assert(&sx->lock_object, what, file, line);
1314 #else
1315                 /*
1316                  * If we hold an exclusve lock fail.  We can't
1317                  * reliably check to see if we hold a shared lock or
1318                  * not.
1319                  */
1320                 if (sx_xholder(sx) == curthread)
1321                         panic("Lock %s exclusively locked @ %s:%d\n",
1322                             sx->lock_object.lo_name, file, line);
1323 #endif
1324                 break;
1325         default:
1326                 panic("Unknown sx lock assertion: %d @ %s:%d", what, file,
1327                     line);
1328         }
1329 }
1330 #endif  /* INVARIANT_SUPPORT */
1331
1332 #ifdef DDB
1333 static void
1334 db_show_sx(const struct lock_object *lock)
1335 {
1336         struct thread *td;
1337         const struct sx *sx;
1338
1339         sx = (const struct sx *)lock;
1340
1341         db_printf(" state: ");
1342         if (sx->sx_lock == SX_LOCK_UNLOCKED)
1343                 db_printf("UNLOCKED\n");
1344         else if (sx->sx_lock == SX_LOCK_DESTROYED) {
1345                 db_printf("DESTROYED\n");
1346                 return;
1347         } else if (sx->sx_lock & SX_LOCK_SHARED)
1348                 db_printf("SLOCK: %ju\n", (uintmax_t)SX_SHARERS(sx->sx_lock));
1349         else {
1350                 td = sx_xholder(sx);
1351                 db_printf("XLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1352                     td->td_tid, td->td_proc->p_pid, td->td_name);
1353                 if (sx_recursed(sx))
1354                         db_printf(" recursed: %d\n", sx->sx_recurse);
1355         }
1356
1357         db_printf(" waiters: ");
1358         switch(sx->sx_lock &
1359             (SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS)) {
1360         case SX_LOCK_SHARED_WAITERS:
1361                 db_printf("shared\n");
1362                 break;
1363         case SX_LOCK_EXCLUSIVE_WAITERS:
1364                 db_printf("exclusive\n");
1365                 break;
1366         case SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS:
1367                 db_printf("exclusive and shared\n");
1368                 break;
1369         default:
1370                 db_printf("none\n");
1371         }
1372 }
1373
1374 /*
1375  * Check to see if a thread that is blocked on a sleep queue is actually
1376  * blocked on an sx lock.  If so, output some details and return true.
1377  * If the lock has an exclusive owner, return that in *ownerp.
1378  */
1379 int
1380 sx_chain(struct thread *td, struct thread **ownerp)
1381 {
1382         struct sx *sx;
1383
1384         /*
1385          * Check to see if this thread is blocked on an sx lock.
1386          * First, we check the lock class.  If that is ok, then we
1387          * compare the lock name against the wait message.
1388          */
1389         sx = td->td_wchan;
1390         if (LOCK_CLASS(&sx->lock_object) != &lock_class_sx ||
1391             sx->lock_object.lo_name != td->td_wmesg)
1392                 return (0);
1393
1394         /* We think we have an sx lock, so output some details. */
1395         db_printf("blocked on sx \"%s\" ", td->td_wmesg);
1396         *ownerp = sx_xholder(sx);
1397         if (sx->sx_lock & SX_LOCK_SHARED)
1398                 db_printf("SLOCK (count %ju)\n",
1399                     (uintmax_t)SX_SHARERS(sx->sx_lock));
1400         else
1401                 db_printf("XLOCK\n");
1402         return (1);
1403 }
1404 #endif