sys/kern/kern_sx.c

   1 /*-
   2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
   3  *
   4  * Copyright (c) 2007 Attilio Rao <attilio@freebsd.org>
   5  * Copyright (c) 2001 Jason Evans <jasone@freebsd.org>
   6  * All rights reserved.
   7  *
   8  * Redistribution and use in source and binary forms, with or without
   9  * modification, are permitted provided that the following conditions
  10  * are met:
  11  * 1. Redistributions of source code must retain the above copyright
  12  *    notice(s), this list of conditions and the following disclaimer as
  13  *    the first lines of this file unmodified other than the possible
  14  *    addition of one or more copyright notices.
  15  * 2. Redistributions in binary form must reproduce the above copyright
  16  *    notice(s), this list of conditions and the following disclaimer in the
  17  *    documentation and/or other materials provided with the distribution.
  18  *
  19  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDER(S) ``AS IS'' AND ANY
  20  * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
  21  * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
  22  * DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT HOLDER(S) BE LIABLE FOR ANY
  23  * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
  24  * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
  25  * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
  26  * CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  27  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  28  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH
  29  * DAMAGE.
  30  */
  31
  32 /*
  33  * Shared/exclusive locks.  This implementation attempts to ensure
  34  * deterministic lock granting behavior, so that slocks and xlocks are
  35  * interleaved.
  36  *
  37  * Priority propagation will not generally raise the priority of lock holders,
  38  * so should not be relied upon in combination with sx locks.
  39  */
  40
  41 #include "opt_ddb.h"
  42 #include "opt_hwpmc_hooks.h"
  43 #include "opt_no_adaptive_sx.h"
  44
  45 #include <sys/cdefs.h>
  46 __FBSDID("$FreeBSD$");
  47
  48 #include <sys/param.h>
  49 #include <sys/systm.h>
  50 #include <sys/kdb.h>
  51 #include <sys/kernel.h>
  52 #include <sys/ktr.h>
  53 #include <sys/lock.h>
  54 #include <sys/mutex.h>
  55 #include <sys/proc.h>
  56 #include <sys/sched.h>
  57 #include <sys/sleepqueue.h>
  58 #include <sys/sx.h>
  59 #include <sys/smp.h>
  60 #include <sys/sysctl.h>
  61
  62 #if defined(SMP) && !defined(NO_ADAPTIVE_SX)
  63 #include <machine/cpu.h>
  64 #endif
  65
  66 #ifdef DDB
  67 #include <ddb/ddb.h>
  68 #endif
  69
  70 #if defined(SMP) && !defined(NO_ADAPTIVE_SX)
  71 #define ADAPTIVE_SX
  72 #endif
  73
  74 CTASSERT((SX_NOADAPTIVE & LO_CLASSFLAGS) == SX_NOADAPTIVE);
  75
  76 #ifdef HWPMC_HOOKS
  77 #include <sys/pmckern.h>
  78 PMC_SOFT_DECLARE( , , lock, failed);
  79 #endif
  80
  81 /* Handy macros for sleep queues. */
  82 #define SQ_EXCLUSIVE_QUEUE      0
  83 #define SQ_SHARED_QUEUE         1
  84
  85 /*
  86  * Variations on DROP_GIANT()/PICKUP_GIANT() for use in this file.  We
  87  * drop Giant anytime we have to sleep or if we adaptively spin.
  88  */
  89 #define GIANT_DECLARE                                                   \
  90         int _giantcnt = 0;                                              \
  91         WITNESS_SAVE_DECL(Giant)                                        \
  92
  93 #define GIANT_SAVE(work) do {                                           \
  94         if (__predict_false(mtx_owned(&Giant))) {                       \
  95                 work++;                                                 \
  96                 WITNESS_SAVE(&Giant.lock_object, Giant);                \
  97                 while (mtx_owned(&Giant)) {                             \
  98                         _giantcnt++;                                    \
  99                         mtx_unlock(&Giant);                             \
 100                 }                                                       \
 101         }                                                               \
 102 } while (0)
 103
 104 #define GIANT_RESTORE() do {                                            \
 105         if (_giantcnt > 0) {                                            \
 106                 mtx_assert(&Giant, MA_NOTOWNED);                        \
 107                 while (_giantcnt--)                                     \
 108                         mtx_lock(&Giant);                               \
 109                 WITNESS_RESTORE(&Giant.lock_object, Giant);             \
 110         }                                                               \
 111 } while (0)
 112
 113 /*
 114  * Returns true if an exclusive lock is recursed.  It assumes
 115  * curthread currently has an exclusive lock.
 116  */
 117 #define sx_recursed(sx)         ((sx)->sx_recurse != 0)
 118
 119 static void     assert_sx(const struct lock_object *lock, int what);
 120 #ifdef DDB
 121 static void     db_show_sx(const struct lock_object *lock);
 122 #endif
 123 static void     lock_sx(struct lock_object *lock, uintptr_t how);
 124 #ifdef KDTRACE_HOOKS
 125 static int      owner_sx(const struct lock_object *lock, struct thread **owner);
 126 #endif
 127 static uintptr_t unlock_sx(struct lock_object *lock);
 128
 129 struct lock_class lock_class_sx = {
 130         .lc_name = "sx",
 131         .lc_flags = LC_SLEEPLOCK | LC_SLEEPABLE | LC_RECURSABLE | LC_UPGRADABLE,
 132         .lc_assert = assert_sx,
 133 #ifdef DDB
 134         .lc_ddb_show = db_show_sx,
 135 #endif
 136         .lc_lock = lock_sx,
 137         .lc_unlock = unlock_sx,
 138 #ifdef KDTRACE_HOOKS
 139         .lc_owner = owner_sx,
 140 #endif
 141 };
 142
 143 #ifndef INVARIANTS
 144 #define _sx_assert(sx, what, file, line)
 145 #endif
 146
 147 #ifdef ADAPTIVE_SX
 148 static __read_frequently u_int asx_retries = 10;
 149 static __read_frequently u_int asx_loops = 10000;
 150 static SYSCTL_NODE(_debug, OID_AUTO, sx, CTLFLAG_RD, NULL, "sxlock debugging");
 151 SYSCTL_UINT(_debug_sx, OID_AUTO, retries, CTLFLAG_RW, &asx_retries, 0, "");
 152 SYSCTL_UINT(_debug_sx, OID_AUTO, loops, CTLFLAG_RW, &asx_loops, 0, "");
 153
 154 static struct lock_delay_config __read_frequently sx_delay;
 155
 156 SYSCTL_INT(_debug_sx, OID_AUTO, delay_base, CTLFLAG_RW, &sx_delay.base,
 157     0, "");
 158 SYSCTL_INT(_debug_sx, OID_AUTO, delay_max, CTLFLAG_RW, &sx_delay.max,
 159     0, "");
 160
 161 LOCK_DELAY_SYSINIT_DEFAULT(sx_delay);
 162 #endif
 163
 164 void
 165 assert_sx(const struct lock_object *lock, int what)
 166 {
 167
 168         sx_assert((const struct sx *)lock, what);
 169 }
 170
 171 void
 172 lock_sx(struct lock_object *lock, uintptr_t how)
 173 {
 174         struct sx *sx;
 175
 176         sx = (struct sx *)lock;
 177         if (how)
 178                 sx_slock(sx);
 179         else
 180                 sx_xlock(sx);
 181 }
 182
 183 uintptr_t
 184 unlock_sx(struct lock_object *lock)
 185 {
 186         struct sx *sx;
 187
 188         sx = (struct sx *)lock;
 189         sx_assert(sx, SA_LOCKED | SA_NOTRECURSED);
 190         if (sx_xlocked(sx)) {
 191                 sx_xunlock(sx);
 192                 return (0);
 193         } else {
 194                 sx_sunlock(sx);
 195                 return (1);
 196         }
 197 }
 198
 199 #ifdef KDTRACE_HOOKS
 200 int
 201 owner_sx(const struct lock_object *lock, struct thread **owner)
 202 {
 203         const struct sx *sx;
 204         uintptr_t x;
 205
 206         sx = (const struct sx *)lock;
 207         x = sx->sx_lock;
 208         *owner = NULL;
 209         return ((x & SX_LOCK_SHARED) != 0 ? (SX_SHARERS(x) != 0) :
 210             ((*owner = (struct thread *)SX_OWNER(x)) != NULL));
 211 }
 212 #endif
 213
 214 void
 215 sx_sysinit(void *arg)
 216 {
 217         struct sx_args *sargs = arg;
 218
 219         sx_init_flags(sargs->sa_sx, sargs->sa_desc, sargs->sa_flags);
 220 }
 221
 222 void
 223 sx_init_flags(struct sx *sx, const char *description, int opts)
 224 {
 225         int flags;
 226
 227         MPASS((opts & ~(SX_QUIET | SX_RECURSE | SX_NOWITNESS | SX_DUPOK |
 228             SX_NOPROFILE | SX_NOADAPTIVE | SX_NEW)) == 0);
 229         ASSERT_ATOMIC_LOAD_PTR(sx->sx_lock,
 230             ("%s: sx_lock not aligned for %s: %p", __func__, description,
 231             &sx->sx_lock));
 232
 233         flags = LO_SLEEPABLE | LO_UPGRADABLE;
 234         if (opts & SX_DUPOK)
 235                 flags |= LO_DUPOK;
 236         if (opts & SX_NOPROFILE)
 237                 flags |= LO_NOPROFILE;
 238         if (!(opts & SX_NOWITNESS))
 239                 flags |= LO_WITNESS;
 240         if (opts & SX_RECURSE)
 241                 flags |= LO_RECURSABLE;
 242         if (opts & SX_QUIET)
 243                 flags |= LO_QUIET;
 244         if (opts & SX_NEW)
 245                 flags |= LO_NEW;
 246
 247         flags |= opts & SX_NOADAPTIVE;
 248         lock_init(&sx->lock_object, &lock_class_sx, description, NULL, flags);
 249         sx->sx_lock = SX_LOCK_UNLOCKED;
 250         sx->sx_recurse = 0;
 251 }
 252
 253 void
 254 sx_destroy(struct sx *sx)
 255 {
 256
 257         KASSERT(sx->sx_lock == SX_LOCK_UNLOCKED, ("sx lock still held"));
 258         KASSERT(sx->sx_recurse == 0, ("sx lock still recursed"));
 259         sx->sx_lock = SX_LOCK_DESTROYED;
 260         lock_destroy(&sx->lock_object);
 261 }
 262
 263 int
 264 sx_try_slock_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
 265 {
 266         uintptr_t x;
 267
 268         if (SCHEDULER_STOPPED())
 269                 return (1);
 270
 271         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
 272             ("sx_try_slock() by idle thread %p on sx %s @ %s:%d",
 273             curthread, sx->lock_object.lo_name, file, line));
 274
 275         x = sx->sx_lock;
 276         for (;;) {
 277                 KASSERT(x != SX_LOCK_DESTROYED,
 278                     ("sx_try_slock() of destroyed sx @ %s:%d", file, line));
 279                 if (!(x & SX_LOCK_SHARED))
 280                         break;
 281                 if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, x + SX_ONE_SHARER)) {
 282                         LOCK_LOG_TRY("SLOCK", &sx->lock_object, 0, 1, file, line);
 283                         WITNESS_LOCK(&sx->lock_object, LOP_TRYLOCK, file, line);
 284                         LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire,
 285                             sx, 0, 0, file, line, LOCKSTAT_READER);
 286                         TD_LOCKS_INC(curthread);
 287                         return (1);
 288                 }
 289         }
 290
 291         LOCK_LOG_TRY("SLOCK", &sx->lock_object, 0, 0, file, line);
 292         return (0);
 293 }
 294
 295 int
 296 sx_try_slock_(struct sx *sx, const char *file, int line)
 297 {
 298
 299         return (sx_try_slock_int(sx LOCK_FILE_LINE_ARG));
 300 }
 301
 302 int
 303 _sx_xlock(struct sx *sx, int opts, const char *file, int line)
 304 {
 305         uintptr_t tid, x;
 306         int error = 0;
 307
 308         KASSERT(kdb_active != 0 || SCHEDULER_STOPPED() ||
 309             !TD_IS_IDLETHREAD(curthread),
 310             ("sx_xlock() by idle thread %p on sx %s @ %s:%d",
 311             curthread, sx->lock_object.lo_name, file, line));
 312         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 313             ("sx_xlock() of destroyed sx @ %s:%d", file, line));
 314         WITNESS_CHECKORDER(&sx->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE, file,
 315             line, NULL);
 316         tid = (uintptr_t)curthread;
 317         x = SX_LOCK_UNLOCKED;
 318         if (!atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid))
 319                 error = _sx_xlock_hard(sx, x, opts LOCK_FILE_LINE_ARG);
 320         else
 321                 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
 322                     0, 0, file, line, LOCKSTAT_WRITER);
 323         if (!error) {
 324                 LOCK_LOG_LOCK("XLOCK", &sx->lock_object, 0, sx->sx_recurse,
 325                     file, line);
 326                 WITNESS_LOCK(&sx->lock_object, LOP_EXCLUSIVE, file, line);
 327                 TD_LOCKS_INC(curthread);
 328         }
 329
 330         return (error);
 331 }
 332
 333 int
 334 sx_try_xlock_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
 335 {
 336         struct thread *td;
 337         uintptr_t tid, x;
 338         int rval;
 339         bool recursed;
 340
 341         td = curthread;
 342         tid = (uintptr_t)td;
 343         if (SCHEDULER_STOPPED_TD(td))
 344                 return (1);
 345
 346         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(td),
 347             ("sx_try_xlock() by idle thread %p on sx %s @ %s:%d",
 348             curthread, sx->lock_object.lo_name, file, line));
 349         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 350             ("sx_try_xlock() of destroyed sx @ %s:%d", file, line));
 351
 352         rval = 1;
 353         recursed = false;
 354         x = SX_LOCK_UNLOCKED;
 355         for (;;) {
 356                 if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid))
 357                         break;
 358                 if (x == SX_LOCK_UNLOCKED)
 359                         continue;
 360                 if (x == tid && (sx->lock_object.lo_flags & LO_RECURSABLE)) {
 361                         sx->sx_recurse++;
 362                         atomic_set_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
 363                         break;
 364                 }
 365                 rval = 0;
 366                 break;
 367         }
 368
 369         LOCK_LOG_TRY("XLOCK", &sx->lock_object, 0, rval, file, line);
 370         if (rval) {
 371                 WITNESS_LOCK(&sx->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
 372                     file, line);
 373                 if (!recursed)
 374                         LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire,
 375                             sx, 0, 0, file, line, LOCKSTAT_WRITER);
 376                 TD_LOCKS_INC(curthread);
 377         }
 378
 379         return (rval);
 380 }
 381
 382 int
 383 sx_try_xlock_(struct sx *sx, const char *file, int line)
 384 {
 385
 386         return (sx_try_xlock_int(sx LOCK_FILE_LINE_ARG));
 387 }
 388
 389 void
 390 _sx_xunlock(struct sx *sx, const char *file, int line)
 391 {
 392
 393         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 394             ("sx_xunlock() of destroyed sx @ %s:%d", file, line));
 395         _sx_assert(sx, SA_XLOCKED, file, line);
 396         WITNESS_UNLOCK(&sx->lock_object, LOP_EXCLUSIVE, file, line);
 397         LOCK_LOG_LOCK("XUNLOCK", &sx->lock_object, 0, sx->sx_recurse, file,
 398             line);
 399 #if LOCK_DEBUG > 0
 400         _sx_xunlock_hard(sx, (uintptr_t)curthread, file, line);
 401 #else
 402         __sx_xunlock(sx, curthread, file, line);
 403 #endif
 404         TD_LOCKS_DEC(curthread);
 405 }
 406
 407 /*
 408  * Try to do a non-blocking upgrade from a shared lock to an exclusive lock.
 409  * This will only succeed if this thread holds a single shared lock.
 410  * Return 1 if if the upgrade succeed, 0 otherwise.
 411  */
 412 int
 413 sx_try_upgrade_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
 414 {
 415         uintptr_t x;
 416         int success;
 417
 418         if (SCHEDULER_STOPPED())
 419                 return (1);
 420
 421         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 422             ("sx_try_upgrade() of destroyed sx @ %s:%d", file, line));
 423         _sx_assert(sx, SA_SLOCKED, file, line);
 424
 425         /*
 426          * Try to switch from one shared lock to an exclusive lock.  We need
 427          * to maintain the SX_LOCK_EXCLUSIVE_WAITERS flag if set so that
 428          * we will wake up the exclusive waiters when we drop the lock.
 429          */
 430         x = sx->sx_lock & SX_LOCK_EXCLUSIVE_WAITERS;
 431         success = atomic_cmpset_acq_ptr(&sx->sx_lock, SX_SHARERS_LOCK(1) | x,
 432             (uintptr_t)curthread | x);
 433         LOCK_LOG_TRY("XUPGRADE", &sx->lock_object, 0, success, file, line);
 434         if (success) {
 435                 WITNESS_UPGRADE(&sx->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
 436                     file, line);
 437                 LOCKSTAT_RECORD0(sx__upgrade, sx);
 438         }
 439         return (success);
 440 }
 441
 442 int
 443 sx_try_upgrade_(struct sx *sx, const char *file, int line)
 444 {
 445
 446         return (sx_try_upgrade_int(sx LOCK_FILE_LINE_ARG));
 447 }
 448
 449 /*
 450  * Downgrade an unrecursed exclusive lock into a single shared lock.
 451  */
 452 void
 453 sx_downgrade_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
 454 {
 455         uintptr_t x;
 456         int wakeup_swapper;
 457
 458         if (SCHEDULER_STOPPED())
 459                 return;
 460
 461         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 462             ("sx_downgrade() of destroyed sx @ %s:%d", file, line));
 463         _sx_assert(sx, SA_XLOCKED | SA_NOTRECURSED, file, line);
 464 #ifndef INVARIANTS
 465         if (sx_recursed(sx))
 466                 panic("downgrade of a recursed lock");
 467 #endif
 468
 469         WITNESS_DOWNGRADE(&sx->lock_object, 0, file, line);
 470
 471         /*
 472          * Try to switch from an exclusive lock with no shared waiters
 473          * to one sharer with no shared waiters.  If there are
 474          * exclusive waiters, we don't need to lock the sleep queue so
 475          * long as we preserve the flag.  We do one quick try and if
 476          * that fails we grab the sleepq lock to keep the flags from
 477          * changing and do it the slow way.
 478          *
 479          * We have to lock the sleep queue if there are shared waiters
 480          * so we can wake them up.
 481          */
 482         x = sx->sx_lock;
 483         if (!(x & SX_LOCK_SHARED_WAITERS) &&
 484             atomic_cmpset_rel_ptr(&sx->sx_lock, x, SX_SHARERS_LOCK(1) |
 485             (x & SX_LOCK_EXCLUSIVE_WAITERS)))
 486                 goto out;
 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         if (wakeup_swapper)
 508                 kick_proc0();
 509
 510 out:
 511         LOCK_LOG_LOCK("XDOWNGRADE", &sx->lock_object, 0, 0, file, line);
 512         LOCKSTAT_RECORD0(sx__downgrade, sx);
 513 }
 514
 515 void
 516 sx_downgrade_(struct sx *sx, const char *file, int line)
 517 {
 518
 519         sx_downgrade_int(sx LOCK_FILE_LINE_ARG);
 520 }
 521
 522 /*
 523  * This function represents the so-called 'hard case' for sx_xlock
 524  * operation.  All 'easy case' failures are redirected to this.  Note
 525  * that ideally this would be a static function, but it needs to be
 526  * accessible from at least sx.h.
 527  */
 528 int
 529 _sx_xlock_hard(struct sx *sx, uintptr_t x, int opts LOCK_FILE_LINE_ARG_DEF)
 530 {
 531         GIANT_DECLARE;
 532         uintptr_t tid;
 533 #ifdef ADAPTIVE_SX
 534         volatile struct thread *owner;
 535         u_int i, n, spintries = 0;
 536         bool adaptive;
 537         int sleep_reason = 0;
 538 #endif
 539 #ifdef LOCK_PROFILING
 540         uint64_t waittime = 0;
 541         int contested = 0;
 542 #endif
 543         int error = 0;
 544 #if defined(ADAPTIVE_SX) || defined(KDTRACE_HOOKS)
 545         struct lock_delay_arg lda;
 546 #endif
 547 #ifdef  KDTRACE_HOOKS
 548         u_int sleep_cnt = 0;
 549         int64_t sleep_time = 0;
 550         int64_t all_time = 0;
 551 #endif
 552 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
 553         uintptr_t state;
 554 #endif
 555         int extra_work = 0;
 556
 557         tid = (uintptr_t)curthread;
 558         if (SCHEDULER_STOPPED())
 559                 return (0);
 560
 561 #if defined(ADAPTIVE_SX)
 562         lock_delay_arg_init(&lda, &sx_delay);
 563 #elif defined(KDTRACE_HOOKS)
 564         lock_delay_arg_init(&lda, NULL);
 565 #endif
 566
 567         if (__predict_false(x == SX_LOCK_UNLOCKED))
 568                 x = SX_READ_VALUE(sx);
 569
 570         /* If we already hold an exclusive lock, then recurse. */
 571         if (__predict_false(lv_sx_owner(x) == (struct thread *)tid)) {
 572                 KASSERT((sx->lock_object.lo_flags & LO_RECURSABLE) != 0,
 573             ("_sx_xlock_hard: recursed on non-recursive sx %s @ %s:%d\n",
 574                     sx->lock_object.lo_name, file, line));
 575                 sx->sx_recurse++;
 576                 atomic_set_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
 577                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 578                         CTR2(KTR_LOCK, "%s: %p recursing", __func__, sx);
 579                 return (0);
 580         }
 581
 582         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 583                 CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__,
 584                     sx->lock_object.lo_name, (void *)sx->sx_lock, file, line);
 585
 586 #ifdef ADAPTIVE_SX
 587         adaptive = ((sx->lock_object.lo_flags & SX_NOADAPTIVE) != 0);
 588 #endif
 589
 590 #ifdef HWPMC_HOOKS
 591         PMC_SOFT_CALL( , , lock, failed);
 592 #endif
 593         lock_profile_obtain_lock_failed(&sx->lock_object, &contested,
 594             &waittime);
 595
 596 #ifdef LOCK_PROFILING
 597         extra_work = 1;
 598         state = x;
 599 #elif defined(KDTRACE_HOOKS)
 600         extra_work = lockstat_enabled;
 601         if (__predict_false(extra_work)) {
 602                 all_time -= lockstat_nsecs(&sx->lock_object);
 603                 state = x;
 604         }
 605 #endif
 606 #ifndef INVARIANTS
 607         GIANT_SAVE(extra_work);
 608 #endif
 609
 610         for (;;) {
 611                 if (x == SX_LOCK_UNLOCKED) {
 612                         if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid))
 613                                 break;
 614                         continue;
 615                 }
 616 #ifdef INVARIANTS
 617                 GIANT_SAVE(extra_work);
 618 #endif
 619 #ifdef KDTRACE_HOOKS
 620                 lda.spin_cnt++;
 621 #endif
 622 #ifdef ADAPTIVE_SX
 623                 if (__predict_false(!adaptive))
 624                         goto sleepq;
 625                 /*
 626                  * If the lock is write locked and the owner is
 627                  * running on another CPU, spin until the owner stops
 628                  * running or the state of the lock changes.
 629                  */
 630                 if ((x & SX_LOCK_SHARED) == 0) {
 631                         owner = lv_sx_owner(x);
 632                         if (TD_IS_RUNNING(owner)) {
 633                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 634                                         CTR3(KTR_LOCK,
 635                                     "%s: spinning on %p held by %p",
 636                                             __func__, sx, owner);
 637                                 KTR_STATE1(KTR_SCHED, "thread",
 638                                     sched_tdname(curthread), "spinning",
 639                                     "lockname:\"%s\"",
 640                                     sx->lock_object.lo_name);
 641                                 do {
 642                                         lock_delay(&lda);
 643                                         x = SX_READ_VALUE(sx);
 644                                         owner = lv_sx_owner(x);
 645                                 } while (owner != NULL &&
 646                                             TD_IS_RUNNING(owner));
 647                                 KTR_STATE0(KTR_SCHED, "thread",
 648                                     sched_tdname(curthread), "running");
 649                                 continue;
 650                         }
 651                         sleep_reason = 1;
 652                 } else if (SX_SHARERS(x) && spintries < asx_retries) {
 653                         KTR_STATE1(KTR_SCHED, "thread",
 654                             sched_tdname(curthread), "spinning",
 655                             "lockname:\"%s\"", sx->lock_object.lo_name);
 656                         spintries++;
 657                         for (i = 0; i < asx_loops; i += n) {
 658                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 659                                         CTR4(KTR_LOCK,
 660                             "%s: shared spinning on %p with %u and %u",
 661                                             __func__, sx, spintries, i);
 662                                 n = SX_SHARERS(x);
 663                                 lock_delay_spin(n);
 664                                 x = SX_READ_VALUE(sx);
 665                                 if ((x & SX_LOCK_SHARED) == 0 ||
 666                                     SX_SHARERS(x) == 0)
 667                                         break;
 668                         }
 669 #ifdef KDTRACE_HOOKS
 670                         lda.spin_cnt += i;
 671 #endif
 672                         KTR_STATE0(KTR_SCHED, "thread",
 673                             sched_tdname(curthread), "running");
 674                         if (i < asx_loops)
 675                                 continue;
 676                         sleep_reason = 2;
 677                 }
 678 sleepq:
 679 #endif
 680                 sleepq_lock(&sx->lock_object);
 681                 x = SX_READ_VALUE(sx);
 682 retry_sleepq:
 683
 684                 /*
 685                  * If the lock was released while spinning on the
 686                  * sleep queue chain lock, try again.
 687                  */
 688                 if (x == SX_LOCK_UNLOCKED) {
 689                         sleepq_release(&sx->lock_object);
 690                         continue;
 691                 }
 692
 693 #ifdef ADAPTIVE_SX
 694                 /*
 695                  * The current lock owner might have started executing
 696                  * on another CPU (or the lock could have changed
 697                  * owners) while we were waiting on the sleep queue
 698                  * chain lock.  If so, drop the sleep queue lock and try
 699                  * again.
 700                  */
 701                 if (adaptive) {
 702                         if (!(x & SX_LOCK_SHARED)) {
 703                                 owner = (struct thread *)SX_OWNER(x);
 704                                 if (TD_IS_RUNNING(owner)) {
 705                                         sleepq_release(&sx->lock_object);
 706                                         continue;
 707                                 }
 708                         } else if (SX_SHARERS(x) > 0 && sleep_reason == 1) {
 709                                 sleepq_release(&sx->lock_object);
 710                                 continue;
 711                         }
 712                 }
 713 #endif
 714
 715                 /*
 716                  * If an exclusive lock was released with both shared
 717                  * and exclusive waiters and a shared waiter hasn't
 718                  * woken up and acquired the lock yet, sx_lock will be
 719                  * set to SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS.
 720                  * If we see that value, try to acquire it once.  Note
 721                  * that we have to preserve SX_LOCK_EXCLUSIVE_WAITERS
 722                  * as there are other exclusive waiters still.  If we
 723                  * fail, restart the loop.
 724                  */
 725                 if (x == (SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS)) {
 726                         if (!atomic_fcmpset_acq_ptr(&sx->sx_lock, &x,
 727                             tid | SX_LOCK_EXCLUSIVE_WAITERS))
 728                                 goto retry_sleepq;
 729                         sleepq_release(&sx->lock_object);
 730                         CTR2(KTR_LOCK, "%s: %p claimed by new writer",
 731                             __func__, sx);
 732                         break;
 733                 }
 734
 735                 /*
 736                  * Try to set the SX_LOCK_EXCLUSIVE_WAITERS.  If we fail,
 737                  * than loop back and retry.
 738                  */
 739                 if (!(x & SX_LOCK_EXCLUSIVE_WAITERS)) {
 740                         if (!atomic_fcmpset_ptr(&sx->sx_lock, &x,
 741                             x | SX_LOCK_EXCLUSIVE_WAITERS)) {
 742                                 goto retry_sleepq;
 743                         }
 744                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 745                                 CTR2(KTR_LOCK, "%s: %p set excl waiters flag",
 746                                     __func__, sx);
 747                 }
 748
 749                 /*
 750                  * Since we have been unable to acquire the exclusive
 751                  * lock and the exclusive waiters flag is set, we have
 752                  * to sleep.
 753                  */
 754                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 755                         CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
 756                             __func__, sx);
 757
 758 #ifdef KDTRACE_HOOKS
 759                 sleep_time -= lockstat_nsecs(&sx->lock_object);
 760 #endif
 761                 sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
 762                     SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
 763                     SLEEPQ_INTERRUPTIBLE : 0), SQ_EXCLUSIVE_QUEUE);
 764                 if (!(opts & SX_INTERRUPTIBLE))
 765                         sleepq_wait(&sx->lock_object, 0);
 766                 else
 767                         error = sleepq_wait_sig(&sx->lock_object, 0);
 768 #ifdef KDTRACE_HOOKS
 769                 sleep_time += lockstat_nsecs(&sx->lock_object);
 770                 sleep_cnt++;
 771 #endif
 772                 if (error) {
 773                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 774                                 CTR2(KTR_LOCK,
 775                         "%s: interruptible sleep by %p suspended by signal",
 776                                     __func__, sx);
 777                         break;
 778                 }
 779                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 780                         CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
 781                             __func__, sx);
 782                 x = SX_READ_VALUE(sx);
 783         }
 784 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
 785         if (__predict_true(!extra_work))
 786                 return (error);
 787 #endif
 788 #ifdef KDTRACE_HOOKS
 789         all_time += lockstat_nsecs(&sx->lock_object);
 790         if (sleep_time)
 791                 LOCKSTAT_RECORD4(sx__block, sx, sleep_time,
 792                     LOCKSTAT_WRITER, (state & SX_LOCK_SHARED) == 0,
 793                     (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
 794         if (lda.spin_cnt > sleep_cnt)
 795                 LOCKSTAT_RECORD4(sx__spin, sx, all_time - sleep_time,
 796                     LOCKSTAT_WRITER, (state & SX_LOCK_SHARED) == 0,
 797                     (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
 798 #endif
 799         if (!error)
 800                 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
 801                     contested, waittime, file, line, LOCKSTAT_WRITER);
 802         GIANT_RESTORE();
 803         return (error);
 804 }
 805
 806 /*
 807  * This function represents the so-called 'hard case' for sx_xunlock
 808  * operation.  All 'easy case' failures are redirected to this.  Note
 809  * that ideally this would be a static function, but it needs to be
 810  * accessible from at least sx.h.
 811  */
 812 void
 813 _sx_xunlock_hard(struct sx *sx, uintptr_t x LOCK_FILE_LINE_ARG_DEF)
 814 {
 815         uintptr_t tid, setx;
 816         int queue, wakeup_swapper;
 817
 818         if (SCHEDULER_STOPPED())
 819                 return;
 820
 821         tid = (uintptr_t)curthread;
 822
 823         if (__predict_false(x == tid))
 824                 x = SX_READ_VALUE(sx);
 825
 826         MPASS(!(x & SX_LOCK_SHARED));
 827
 828         if (__predict_false(x & SX_LOCK_RECURSED)) {
 829                 /* The lock is recursed, unrecurse one level. */
 830                 if ((--sx->sx_recurse) == 0)
 831                         atomic_clear_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
 832                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 833                         CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, sx);
 834                 return;
 835         }
 836
 837         LOCKSTAT_PROFILE_RELEASE_RWLOCK(sx__release, sx, LOCKSTAT_WRITER);
 838         if (x == tid &&
 839             atomic_cmpset_rel_ptr(&sx->sx_lock, tid, SX_LOCK_UNLOCKED))
 840                 return;
 841
 842         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 843                 CTR2(KTR_LOCK, "%s: %p contested", __func__, sx);
 844
 845         sleepq_lock(&sx->lock_object);
 846         x = SX_READ_VALUE(sx);
 847         MPASS(x & (SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS));
 848
 849         /*
 850          * The wake up algorithm here is quite simple and probably not
 851          * ideal.  It gives precedence to shared waiters if they are
 852          * present.  For this condition, we have to preserve the
 853          * state of the exclusive waiters flag.
 854          * If interruptible sleeps left the shared queue empty avoid a
 855          * starvation for the threads sleeping on the exclusive queue by giving
 856          * them precedence and cleaning up the shared waiters bit anyway.
 857          */
 858         setx = SX_LOCK_UNLOCKED;
 859         queue = SQ_EXCLUSIVE_QUEUE;
 860         if ((x & SX_LOCK_SHARED_WAITERS) != 0 &&
 861             sleepq_sleepcnt(&sx->lock_object, SQ_SHARED_QUEUE) != 0) {
 862                 queue = SQ_SHARED_QUEUE;
 863                 setx |= (x & SX_LOCK_EXCLUSIVE_WAITERS);
 864         }
 865         atomic_store_rel_ptr(&sx->sx_lock, setx);
 866
 867         /* Wake up all the waiters for the specific queue. */
 868         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 869                 CTR3(KTR_LOCK, "%s: %p waking up all threads on %s queue",
 870                     __func__, sx, queue == SQ_SHARED_QUEUE ? "shared" :
 871                     "exclusive");
 872
 873         wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX, 0,
 874             queue);
 875         sleepq_release(&sx->lock_object);
 876         if (wakeup_swapper)
 877                 kick_proc0();
 878 }
 879
 880 static bool __always_inline
 881 __sx_slock_try(struct sx *sx, uintptr_t *xp LOCK_FILE_LINE_ARG_DEF)
 882 {
 883
 884         /*
 885          * If no other thread has an exclusive lock then try to bump up
 886          * the count of sharers.  Since we have to preserve the state
 887          * of SX_LOCK_EXCLUSIVE_WAITERS, if we fail to acquire the
 888          * shared lock loop back and retry.
 889          */
 890         while (*xp & SX_LOCK_SHARED) {
 891                 MPASS(!(*xp & SX_LOCK_SHARED_WAITERS));
 892                 if (atomic_fcmpset_acq_ptr(&sx->sx_lock, xp,
 893                     *xp + SX_ONE_SHARER)) {
 894                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 895                                 CTR4(KTR_LOCK, "%s: %p succeed %p -> %p",
 896                                     __func__, sx, (void *)*xp,
 897                                     (void *)(*xp + SX_ONE_SHARER));
 898                         return (true);
 899                 }
 900         }
 901         return (false);
 902 }
 903
 904 static int __noinline
 905 _sx_slock_hard(struct sx *sx, int opts, uintptr_t x LOCK_FILE_LINE_ARG_DEF)
 906 {
 907         GIANT_DECLARE;
 908 #ifdef ADAPTIVE_SX
 909         volatile struct thread *owner;
 910         bool adaptive;
 911 #endif
 912 #ifdef LOCK_PROFILING
 913         uint64_t waittime = 0;
 914         int contested = 0;
 915 #endif
 916         int error = 0;
 917 #if defined(ADAPTIVE_SX) || defined(KDTRACE_HOOKS)
 918         struct lock_delay_arg lda;
 919 #endif
 920 #ifdef KDTRACE_HOOKS
 921         u_int sleep_cnt = 0;
 922         int64_t sleep_time = 0;
 923         int64_t all_time = 0;
 924 #endif
 925 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
 926         uintptr_t state;
 927 #endif
 928         int extra_work = 0;
 929
 930         if (SCHEDULER_STOPPED())
 931                 return (0);
 932
 933 #if defined(ADAPTIVE_SX)
 934         lock_delay_arg_init(&lda, &sx_delay);
 935 #elif defined(KDTRACE_HOOKS)
 936         lock_delay_arg_init(&lda, NULL);
 937 #endif
 938
 939 #ifdef ADAPTIVE_SX
 940         adaptive = ((sx->lock_object.lo_flags & SX_NOADAPTIVE) != 0);
 941 #endif
 942
 943 #ifdef HWPMC_HOOKS
 944         PMC_SOFT_CALL( , , lock, failed);
 945 #endif
 946         lock_profile_obtain_lock_failed(&sx->lock_object, &contested,
 947             &waittime);
 948
 949 #ifdef LOCK_PROFILING
 950         extra_work = 1;
 951         state = x;
 952 #elif defined(KDTRACE_HOOKS)
 953         extra_work = lockstat_enabled;
 954         if (__predict_false(extra_work)) {
 955                 all_time -= lockstat_nsecs(&sx->lock_object);
 956                 state = x;
 957         }
 958 #endif
 959 #ifndef INVARIANTS
 960         GIANT_SAVE(extra_work);
 961 #endif
 962
 963         /*
 964          * As with rwlocks, we don't make any attempt to try to block
 965          * shared locks once there is an exclusive waiter.
 966          */
 967         for (;;) {
 968                 if (__sx_slock_try(sx, &x LOCK_FILE_LINE_ARG))
 969                         break;
 970 #ifdef INVARIANTS
 971                 GIANT_SAVE(extra_work);
 972 #endif
 973 #ifdef KDTRACE_HOOKS
 974                 lda.spin_cnt++;
 975 #endif
 976
 977 #ifdef ADAPTIVE_SX
 978                 if (__predict_false(!adaptive))
 979                         goto sleepq;
 980                 /*
 981                  * If the owner is running on another CPU, spin until
 982                  * the owner stops running or the state of the lock
 983                  * changes.
 984                  */
 985                 owner = lv_sx_owner(x);
 986                 if (TD_IS_RUNNING(owner)) {
 987                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 988                                 CTR3(KTR_LOCK,
 989                                     "%s: spinning on %p held by %p",
 990                                     __func__, sx, owner);
 991                         KTR_STATE1(KTR_SCHED, "thread",
 992                             sched_tdname(curthread), "spinning",
 993                             "lockname:\"%s\"", sx->lock_object.lo_name);
 994                         do {
 995                                 lock_delay(&lda);
 996                                 x = SX_READ_VALUE(sx);
 997                                 owner = lv_sx_owner(x);
 998                         } while (owner != NULL && TD_IS_RUNNING(owner));
 999                         KTR_STATE0(KTR_SCHED, "thread",
1000                             sched_tdname(curthread), "running");
1001                         continue;
1002                 }
1003 sleepq:
1004 #endif
1005
1006                 /*
1007                  * Some other thread already has an exclusive lock, so
1008                  * start the process of blocking.
1009                  */
1010                 sleepq_lock(&sx->lock_object);
1011                 x = SX_READ_VALUE(sx);
1012 retry_sleepq:
1013                 /*
1014                  * The lock could have been released while we spun.
1015                  * In this case loop back and retry.
1016                  */
1017                 if (x & SX_LOCK_SHARED) {
1018                         sleepq_release(&sx->lock_object);
1019                         continue;
1020                 }
1021
1022 #ifdef ADAPTIVE_SX
1023                 /*
1024                  * If the owner is running on another CPU, spin until
1025                  * the owner stops running or the state of the lock
1026                  * changes.
1027                  */
1028                 if (!(x & SX_LOCK_SHARED) && adaptive) {
1029                         owner = (struct thread *)SX_OWNER(x);
1030                         if (TD_IS_RUNNING(owner)) {
1031                                 sleepq_release(&sx->lock_object);
1032                                 x = SX_READ_VALUE(sx);
1033                                 continue;
1034                         }
1035                 }
1036 #endif
1037
1038                 /*
1039                  * Try to set the SX_LOCK_SHARED_WAITERS flag.  If we
1040                  * fail to set it drop the sleep queue lock and loop
1041                  * back.
1042                  */
1043                 if (!(x & SX_LOCK_SHARED_WAITERS)) {
1044                         if (!atomic_fcmpset_ptr(&sx->sx_lock, &x,
1045                             x | SX_LOCK_SHARED_WAITERS))
1046                                 goto retry_sleepq;
1047                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
1048                                 CTR2(KTR_LOCK, "%s: %p set shared waiters flag",
1049                                     __func__, sx);
1050                 }
1051
1052                 /*
1053                  * Since we have been unable to acquire the shared lock,
1054                  * we have to sleep.
1055                  */
1056                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1057                         CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
1058                             __func__, sx);
1059
1060 #ifdef KDTRACE_HOOKS
1061                 sleep_time -= lockstat_nsecs(&sx->lock_object);
1062 #endif
1063                 sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
1064                     SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
1065                     SLEEPQ_INTERRUPTIBLE : 0), SQ_SHARED_QUEUE);
1066                 if (!(opts & SX_INTERRUPTIBLE))
1067                         sleepq_wait(&sx->lock_object, 0);
1068                 else
1069                         error = sleepq_wait_sig(&sx->lock_object, 0);
1070 #ifdef KDTRACE_HOOKS
1071                 sleep_time += lockstat_nsecs(&sx->lock_object);
1072                 sleep_cnt++;
1073 #endif
1074                 if (error) {
1075                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
1076                                 CTR2(KTR_LOCK,
1077                         "%s: interruptible sleep by %p suspended by signal",
1078                                     __func__, sx);
1079                         break;
1080                 }
1081                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1082                         CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
1083                             __func__, sx);
1084                 x = SX_READ_VALUE(sx);
1085         }
1086 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
1087         if (__predict_true(!extra_work))
1088                 return (error);
1089 #endif
1090 #ifdef KDTRACE_HOOKS
1091         all_time += lockstat_nsecs(&sx->lock_object);
1092         if (sleep_time)
1093                 LOCKSTAT_RECORD4(sx__block, sx, sleep_time,
1094                     LOCKSTAT_READER, (state & SX_LOCK_SHARED) == 0,
1095                     (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
1096         if (lda.spin_cnt > sleep_cnt)
1097                 LOCKSTAT_RECORD4(sx__spin, sx, all_time - sleep_time,
1098                     LOCKSTAT_READER, (state & SX_LOCK_SHARED) == 0,
1099                     (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
1100 #endif
1101         if (error == 0) {
1102                 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
1103                     contested, waittime, file, line, LOCKSTAT_READER);
1104         }
1105         GIANT_RESTORE();
1106         return (error);
1107 }
1108
1109 int
1110 _sx_slock_int(struct sx *sx, int opts LOCK_FILE_LINE_ARG_DEF)
1111 {
1112         uintptr_t x;
1113         int error;
1114
1115         KASSERT(kdb_active != 0 || SCHEDULER_STOPPED() ||
1116             !TD_IS_IDLETHREAD(curthread),
1117             ("sx_slock() by idle thread %p on sx %s @ %s:%d",
1118             curthread, sx->lock_object.lo_name, file, line));
1119         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
1120             ("sx_slock() of destroyed sx @ %s:%d", file, line));
1121         WITNESS_CHECKORDER(&sx->lock_object, LOP_NEWORDER, file, line, NULL);
1122
1123         error = 0;
1124         x = SX_READ_VALUE(sx);
1125         if (__predict_false(LOCKSTAT_OOL_PROFILE_ENABLED(sx__acquire) ||
1126             !__sx_slock_try(sx, &x LOCK_FILE_LINE_ARG)))
1127                 error = _sx_slock_hard(sx, opts, x LOCK_FILE_LINE_ARG);
1128         if (error == 0) {
1129                 LOCK_LOG_LOCK("SLOCK", &sx->lock_object, 0, 0, file, line);
1130                 WITNESS_LOCK(&sx->lock_object, 0, file, line);
1131                 TD_LOCKS_INC(curthread);
1132         }
1133         return (error);
1134 }
1135
1136 int
1137 _sx_slock(struct sx *sx, int opts, const char *file, int line)
1138 {
1139
1140         return (_sx_slock_int(sx, opts LOCK_FILE_LINE_ARG));
1141 }
1142
1143 static bool __always_inline
1144 _sx_sunlock_try(struct sx *sx, uintptr_t *xp)
1145 {
1146
1147         for (;;) {
1148                 /*
1149                  * We should never have sharers while at least one thread
1150                  * holds a shared lock.
1151                  */
1152                 KASSERT(!(*xp & SX_LOCK_SHARED_WAITERS),
1153                     ("%s: waiting sharers", __func__));
1154
1155                 /*
1156                  * See if there is more than one shared lock held.  If
1157                  * so, just drop one and return.
1158                  */
1159                 if (SX_SHARERS(*xp) > 1) {
1160                         if (atomic_fcmpset_rel_ptr(&sx->sx_lock, xp,
1161                             *xp - SX_ONE_SHARER)) {
1162                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1163                                         CTR4(KTR_LOCK,
1164                                             "%s: %p succeeded %p -> %p",
1165                                             __func__, sx, (void *)*xp,
1166                                             (void *)(*xp - SX_ONE_SHARER));
1167                                 return (true);
1168                         }
1169                         continue;
1170                 }
1171
1172                 /*
1173                  * If there aren't any waiters for an exclusive lock,
1174                  * then try to drop it quickly.
1175                  */
1176                 if (!(*xp & SX_LOCK_EXCLUSIVE_WAITERS)) {
1177                         MPASS(*xp == SX_SHARERS_LOCK(1));
1178                         *xp = SX_SHARERS_LOCK(1);
1179                         if (atomic_fcmpset_rel_ptr(&sx->sx_lock,
1180                             xp, SX_LOCK_UNLOCKED)) {
1181                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1182                                         CTR2(KTR_LOCK, "%s: %p last succeeded",
1183                                             __func__, sx);
1184                                 return (true);
1185                         }
1186                         continue;
1187                 }
1188                 break;
1189         }
1190         return (false);
1191 }
1192
1193 static void __noinline
1194 _sx_sunlock_hard(struct sx *sx, uintptr_t x LOCK_FILE_LINE_ARG_DEF)
1195 {
1196         int wakeup_swapper = 0;
1197         uintptr_t setx;
1198
1199         if (SCHEDULER_STOPPED())
1200                 return;
1201
1202         if (_sx_sunlock_try(sx, &x))
1203                 goto out_lockstat;
1204
1205         /*
1206          * At this point, there should just be one sharer with
1207          * exclusive waiters.
1208          */
1209         MPASS(x == (SX_SHARERS_LOCK(1) | SX_LOCK_EXCLUSIVE_WAITERS));
1210
1211         sleepq_lock(&sx->lock_object);
1212         x = SX_READ_VALUE(sx);
1213         for (;;) {
1214                 MPASS(x & SX_LOCK_EXCLUSIVE_WAITERS);
1215                 MPASS(!(x & SX_LOCK_SHARED_WAITERS));
1216                 if (_sx_sunlock_try(sx, &x))
1217                         break;
1218
1219                 /*
1220                  * Wake up semantic here is quite simple:
1221                  * Just wake up all the exclusive waiters.
1222                  * Note that the state of the lock could have changed,
1223                  * so if it fails loop back and retry.
1224                  */
1225                 setx = x - SX_ONE_SHARER;
1226                 setx &= ~SX_LOCK_EXCLUSIVE_WAITERS;
1227                 if (!atomic_fcmpset_rel_ptr(&sx->sx_lock, &x, setx))
1228                         continue;
1229                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1230                         CTR2(KTR_LOCK, "%s: %p waking up all thread on"
1231                             "exclusive queue", __func__, sx);
1232                 wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX,
1233                     0, SQ_EXCLUSIVE_QUEUE);
1234                 break;
1235         }
1236         sleepq_release(&sx->lock_object);
1237         if (wakeup_swapper)
1238                 kick_proc0();
1239 out_lockstat:
1240         LOCKSTAT_PROFILE_RELEASE_RWLOCK(sx__release, sx, LOCKSTAT_READER);
1241 }
1242
1243 void
1244 _sx_sunlock_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
1245 {
1246         uintptr_t x;
1247
1248         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
1249             ("sx_sunlock() of destroyed sx @ %s:%d", file, line));
1250         _sx_assert(sx, SA_SLOCKED, file, line);
1251         WITNESS_UNLOCK(&sx->lock_object, 0, file, line);
1252         LOCK_LOG_LOCK("SUNLOCK", &sx->lock_object, 0, 0, file, line);
1253
1254         x = SX_READ_VALUE(sx);
1255         if (__predict_false(LOCKSTAT_OOL_PROFILE_ENABLED(sx__release) ||
1256             !_sx_sunlock_try(sx, &x)))
1257                 _sx_sunlock_hard(sx, x LOCK_FILE_LINE_ARG);
1258
1259         TD_LOCKS_DEC(curthread);
1260 }
1261
1262 void
1263 _sx_sunlock(struct sx *sx, const char *file, int line)
1264 {
1265
1266         _sx_sunlock_int(sx LOCK_FILE_LINE_ARG);
1267 }
1268
1269 #ifdef INVARIANT_SUPPORT
1270 #ifndef INVARIANTS
1271 #undef  _sx_assert
1272 #endif
1273
1274 /*
1275  * In the non-WITNESS case, sx_assert() can only detect that at least
1276  * *some* thread owns an slock, but it cannot guarantee that *this*
1277  * thread owns an slock.
1278  */
1279 void
1280 _sx_assert(const struct sx *sx, int what, const char *file, int line)
1281 {
1282 #ifndef WITNESS
1283         int slocked = 0;
1284 #endif
1285
1286         if (panicstr != NULL)
1287                 return;
1288         switch (what) {
1289         case SA_SLOCKED:
1290         case SA_SLOCKED | SA_NOTRECURSED:
1291         case SA_SLOCKED | SA_RECURSED:
1292 #ifndef WITNESS
1293                 slocked = 1;
1294                 /* FALLTHROUGH */
1295 #endif
1296         case SA_LOCKED:
1297         case SA_LOCKED | SA_NOTRECURSED:
1298         case SA_LOCKED | SA_RECURSED:
1299 #ifdef WITNESS
1300                 witness_assert(&sx->lock_object, what, file, line);
1301 #else
1302                 /*
1303                  * If some other thread has an exclusive lock or we
1304                  * have one and are asserting a shared lock, fail.
1305                  * Also, if no one has a lock at all, fail.
1306                  */
1307                 if (sx->sx_lock == SX_LOCK_UNLOCKED ||
1308                     (!(sx->sx_lock & SX_LOCK_SHARED) && (slocked ||
1309                     sx_xholder(sx) != curthread)))
1310                         panic("Lock %s not %slocked @ %s:%d\n",
1311                             sx->lock_object.lo_name, slocked ? "share " : "",
1312                             file, line);
1313
1314                 if (!(sx->sx_lock & SX_LOCK_SHARED)) {
1315                         if (sx_recursed(sx)) {
1316                                 if (what & SA_NOTRECURSED)
1317                                         panic("Lock %s recursed @ %s:%d\n",
1318                                             sx->lock_object.lo_name, file,
1319                                             line);
1320                         } else if (what & SA_RECURSED)
1321                                 panic("Lock %s not recursed @ %s:%d\n",
1322                                     sx->lock_object.lo_name, file, line);
1323                 }
1324 #endif
1325                 break;
1326         case SA_XLOCKED:
1327         case SA_XLOCKED | SA_NOTRECURSED:
1328         case SA_XLOCKED | SA_RECURSED:
1329                 if (sx_xholder(sx) != curthread)
1330                         panic("Lock %s not exclusively locked @ %s:%d\n",
1331                             sx->lock_object.lo_name, file, line);
1332                 if (sx_recursed(sx)) {
1333                         if (what & SA_NOTRECURSED)
1334                                 panic("Lock %s recursed @ %s:%d\n",
1335                                     sx->lock_object.lo_name, file, line);
1336                 } else if (what & SA_RECURSED)
1337                         panic("Lock %s not recursed @ %s:%d\n",
1338                             sx->lock_object.lo_name, file, line);
1339                 break;
1340         case SA_UNLOCKED:
1341 #ifdef WITNESS
1342                 witness_assert(&sx->lock_object, what, file, line);
1343 #else
1344                 /*
1345                  * If we hold an exclusve lock fail.  We can't
1346                  * reliably check to see if we hold a shared lock or
1347                  * not.
1348                  */
1349                 if (sx_xholder(sx) == curthread)
1350                         panic("Lock %s exclusively locked @ %s:%d\n",
1351                             sx->lock_object.lo_name, file, line);
1352 #endif
1353                 break;
1354         default:
1355                 panic("Unknown sx lock assertion: %d @ %s:%d", what, file,
1356                     line);
1357         }
1358 }
1359 #endif  /* INVARIANT_SUPPORT */
1360
1361 #ifdef DDB
1362 static void
1363 db_show_sx(const struct lock_object *lock)
1364 {
1365         struct thread *td;
1366         const struct sx *sx;
1367
1368         sx = (const struct sx *)lock;
1369
1370         db_printf(" state: ");
1371         if (sx->sx_lock == SX_LOCK_UNLOCKED)
1372                 db_printf("UNLOCKED\n");
1373         else if (sx->sx_lock == SX_LOCK_DESTROYED) {
1374                 db_printf("DESTROYED\n");
1375                 return;
1376         } else if (sx->sx_lock & SX_LOCK_SHARED)
1377                 db_printf("SLOCK: %ju\n", (uintmax_t)SX_SHARERS(sx->sx_lock));
1378         else {
1379                 td = sx_xholder(sx);
1380                 db_printf("XLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1381                     td->td_tid, td->td_proc->p_pid, td->td_name);
1382                 if (sx_recursed(sx))
1383                         db_printf(" recursed: %d\n", sx->sx_recurse);
1384         }
1385
1386         db_printf(" waiters: ");
1387         switch(sx->sx_lock &
1388             (SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS)) {
1389         case SX_LOCK_SHARED_WAITERS:
1390                 db_printf("shared\n");
1391                 break;
1392         case SX_LOCK_EXCLUSIVE_WAITERS:
1393                 db_printf("exclusive\n");
1394                 break;
1395         case SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS:
1396                 db_printf("exclusive and shared\n");
1397                 break;
1398         default:
1399                 db_printf("none\n");
1400         }
1401 }
1402
1403 /*
1404  * Check to see if a thread that is blocked on a sleep queue is actually
1405  * blocked on an sx lock.  If so, output some details and return true.
1406  * If the lock has an exclusive owner, return that in *ownerp.
1407  */
1408 int
1409 sx_chain(struct thread *td, struct thread **ownerp)
1410 {
1411         struct sx *sx;
1412
1413         /*
1414          * Check to see if this thread is blocked on an sx lock.
1415          * First, we check the lock class.  If that is ok, then we
1416          * compare the lock name against the wait message.
1417          */
1418         sx = td->td_wchan;
1419         if (LOCK_CLASS(&sx->lock_object) != &lock_class_sx ||
1420             sx->lock_object.lo_name != td->td_wmesg)
1421                 return (0);
1422
1423         /* We think we have an sx lock, so output some details. */
1424         db_printf("blocked on sx \"%s\" ", td->td_wmesg);
1425         *ownerp = sx_xholder(sx);
1426         if (sx->sx_lock & SX_LOCK_SHARED)
1427                 db_printf("SLOCK (count %ju)\n",
1428                     (uintmax_t)SX_SHARERS(sx->sx_lock));
1429         else
1430                 db_printf("XLOCK\n");
1431         return (1);
1432 }
1433 #endif