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 (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 #endif
 537 #ifdef LOCK_PROFILING
 538         uint64_t waittime = 0;
 539         int contested = 0;
 540 #endif
 541         int error = 0;
 542 #if defined(ADAPTIVE_SX) || defined(KDTRACE_HOOKS)
 543         struct lock_delay_arg lda;
 544 #endif
 545 #ifdef  KDTRACE_HOOKS
 546         u_int sleep_cnt = 0;
 547         int64_t sleep_time = 0;
 548         int64_t all_time = 0;
 549 #endif
 550 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
 551         uintptr_t state;
 552 #endif
 553         int extra_work = 0;
 554
 555         tid = (uintptr_t)curthread;
 556         if (SCHEDULER_STOPPED())
 557                 return (0);
 558
 559 #if defined(ADAPTIVE_SX)
 560         lock_delay_arg_init(&lda, &sx_delay);
 561 #elif defined(KDTRACE_HOOKS)
 562         lock_delay_arg_init(&lda, NULL);
 563 #endif
 564
 565         if (__predict_false(x == SX_LOCK_UNLOCKED))
 566                 x = SX_READ_VALUE(sx);
 567
 568         /* If we already hold an exclusive lock, then recurse. */
 569         if (__predict_false(lv_sx_owner(x) == (struct thread *)tid)) {
 570                 KASSERT((sx->lock_object.lo_flags & LO_RECURSABLE) != 0,
 571             ("_sx_xlock_hard: recursed on non-recursive sx %s @ %s:%d\n",
 572                     sx->lock_object.lo_name, file, line));
 573                 sx->sx_recurse++;
 574                 atomic_set_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
 575                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 576                         CTR2(KTR_LOCK, "%s: %p recursing", __func__, sx);
 577                 return (0);
 578         }
 579
 580         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 581                 CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__,
 582                     sx->lock_object.lo_name, (void *)sx->sx_lock, file, line);
 583
 584 #ifdef HWPMC_HOOKS
 585         PMC_SOFT_CALL( , , lock, failed);
 586 #endif
 587         lock_profile_obtain_lock_failed(&sx->lock_object, &contested,
 588             &waittime);
 589
 590 #ifdef LOCK_PROFILING
 591         extra_work = 1;
 592         state = x;
 593 #elif defined(KDTRACE_HOOKS)
 594         extra_work = lockstat_enabled;
 595         if (__predict_false(extra_work)) {
 596                 all_time -= lockstat_nsecs(&sx->lock_object);
 597                 state = x;
 598         }
 599 #endif
 600
 601         for (;;) {
 602                 if (x == SX_LOCK_UNLOCKED) {
 603                         if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid))
 604                                 break;
 605                         continue;
 606                 }
 607 #ifdef KDTRACE_HOOKS
 608                 lda.spin_cnt++;
 609 #endif
 610 #ifdef ADAPTIVE_SX
 611                 /*
 612                  * If the lock is write locked and the owner is
 613                  * running on another CPU, spin until the owner stops
 614                  * running or the state of the lock changes.
 615                  */
 616                 if ((sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) {
 617                         if ((x & SX_LOCK_SHARED) == 0) {
 618                                 owner = lv_sx_owner(x);
 619                                 if (TD_IS_RUNNING(owner)) {
 620                                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 621                                                 CTR3(KTR_LOCK,
 622                                             "%s: spinning on %p held by %p",
 623                                                     __func__, sx, owner);
 624                                         KTR_STATE1(KTR_SCHED, "thread",
 625                                             sched_tdname(curthread), "spinning",
 626                                             "lockname:\"%s\"",
 627                                             sx->lock_object.lo_name);
 628                                         GIANT_SAVE(extra_work);
 629                                         do {
 630                                                 lock_delay(&lda);
 631                                                 x = SX_READ_VALUE(sx);
 632                                                 owner = lv_sx_owner(x);
 633                                         } while (owner != NULL &&
 634                                                     TD_IS_RUNNING(owner));
 635                                         KTR_STATE0(KTR_SCHED, "thread",
 636                                             sched_tdname(curthread), "running");
 637                                         continue;
 638                                 }
 639                         } else if (SX_SHARERS(x) && spintries < asx_retries) {
 640                                 KTR_STATE1(KTR_SCHED, "thread",
 641                                     sched_tdname(curthread), "spinning",
 642                                     "lockname:\"%s\"", sx->lock_object.lo_name);
 643                                 GIANT_SAVE(extra_work);
 644                                 spintries++;
 645                                 for (i = 0; i < asx_loops; i += n) {
 646                                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 647                                                 CTR4(KTR_LOCK,
 648                                     "%s: shared spinning on %p with %u and %u",
 649                                                     __func__, sx, spintries, i);
 650                                         n = SX_SHARERS(x);
 651                                         lock_delay_spin(n);
 652                                         x = SX_READ_VALUE(sx);
 653                                         if ((x & SX_LOCK_SHARED) == 0 ||
 654                                             SX_SHARERS(x) == 0)
 655                                                 break;
 656                                 }
 657 #ifdef KDTRACE_HOOKS
 658                                 lda.spin_cnt += i;
 659 #endif
 660                                 KTR_STATE0(KTR_SCHED, "thread",
 661                                     sched_tdname(curthread), "running");
 662                                 if (i != asx_loops)
 663                                         continue;
 664                         }
 665                 }
 666 #endif
 667
 668                 sleepq_lock(&sx->lock_object);
 669                 x = SX_READ_VALUE(sx);
 670 retry_sleepq:
 671
 672                 /*
 673                  * If the lock was released while spinning on the
 674                  * sleep queue chain lock, try again.
 675                  */
 676                 if (x == SX_LOCK_UNLOCKED) {
 677                         sleepq_release(&sx->lock_object);
 678                         continue;
 679                 }
 680
 681 #ifdef ADAPTIVE_SX
 682                 /*
 683                  * The current lock owner might have started executing
 684                  * on another CPU (or the lock could have changed
 685                  * owners) while we were waiting on the sleep queue
 686                  * chain lock.  If so, drop the sleep queue lock and try
 687                  * again.
 688                  */
 689                 if (!(x & SX_LOCK_SHARED) &&
 690                     (sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) {
 691                         owner = (struct thread *)SX_OWNER(x);
 692                         if (TD_IS_RUNNING(owner)) {
 693                                 sleepq_release(&sx->lock_object);
 694                                 continue;
 695                         }
 696                 }
 697 #endif
 698
 699                 /*
 700                  * If an exclusive lock was released with both shared
 701                  * and exclusive waiters and a shared waiter hasn't
 702                  * woken up and acquired the lock yet, sx_lock will be
 703                  * set to SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS.
 704                  * If we see that value, try to acquire it once.  Note
 705                  * that we have to preserve SX_LOCK_EXCLUSIVE_WAITERS
 706                  * as there are other exclusive waiters still.  If we
 707                  * fail, restart the loop.
 708                  */
 709                 if (x == (SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS)) {
 710                         if (!atomic_fcmpset_acq_ptr(&sx->sx_lock, &x,
 711                             tid | SX_LOCK_EXCLUSIVE_WAITERS))
 712                                 goto retry_sleepq;
 713                         sleepq_release(&sx->lock_object);
 714                         CTR2(KTR_LOCK, "%s: %p claimed by new writer",
 715                             __func__, sx);
 716                         break;
 717                 }
 718
 719                 /*
 720                  * Try to set the SX_LOCK_EXCLUSIVE_WAITERS.  If we fail,
 721                  * than loop back and retry.
 722                  */
 723                 if (!(x & SX_LOCK_EXCLUSIVE_WAITERS)) {
 724                         if (!atomic_fcmpset_ptr(&sx->sx_lock, &x,
 725                             x | SX_LOCK_EXCLUSIVE_WAITERS)) {
 726                                 goto retry_sleepq;
 727                         }
 728                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 729                                 CTR2(KTR_LOCK, "%s: %p set excl waiters flag",
 730                                     __func__, sx);
 731                 }
 732
 733                 /*
 734                  * Since we have been unable to acquire the exclusive
 735                  * lock and the exclusive waiters flag is set, we have
 736                  * to sleep.
 737                  */
 738                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 739                         CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
 740                             __func__, sx);
 741
 742 #ifdef KDTRACE_HOOKS
 743                 sleep_time -= lockstat_nsecs(&sx->lock_object);
 744 #endif
 745                 GIANT_SAVE(extra_work);
 746                 sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
 747                     SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
 748                     SLEEPQ_INTERRUPTIBLE : 0), SQ_EXCLUSIVE_QUEUE);
 749                 if (!(opts & SX_INTERRUPTIBLE))
 750                         sleepq_wait(&sx->lock_object, 0);
 751                 else
 752                         error = sleepq_wait_sig(&sx->lock_object, 0);
 753 #ifdef KDTRACE_HOOKS
 754                 sleep_time += lockstat_nsecs(&sx->lock_object);
 755                 sleep_cnt++;
 756 #endif
 757                 if (error) {
 758                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 759                                 CTR2(KTR_LOCK,
 760                         "%s: interruptible sleep by %p suspended by signal",
 761                                     __func__, sx);
 762                         break;
 763                 }
 764                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 765                         CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
 766                             __func__, sx);
 767                 x = SX_READ_VALUE(sx);
 768         }
 769 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
 770         if (__predict_true(!extra_work))
 771                 return (error);
 772 #endif
 773 #ifdef KDTRACE_HOOKS
 774         all_time += lockstat_nsecs(&sx->lock_object);
 775         if (sleep_time)
 776                 LOCKSTAT_RECORD4(sx__block, sx, sleep_time,
 777                     LOCKSTAT_WRITER, (state & SX_LOCK_SHARED) == 0,
 778                     (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
 779         if (lda.spin_cnt > sleep_cnt)
 780                 LOCKSTAT_RECORD4(sx__spin, sx, all_time - sleep_time,
 781                     LOCKSTAT_WRITER, (state & SX_LOCK_SHARED) == 0,
 782                     (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
 783 #endif
 784         if (!error)
 785                 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
 786                     contested, waittime, file, line, LOCKSTAT_WRITER);
 787         GIANT_RESTORE();
 788         return (error);
 789 }
 790
 791 /*
 792  * This function represents the so-called 'hard case' for sx_xunlock
 793  * operation.  All 'easy case' failures are redirected to this.  Note
 794  * that ideally this would be a static function, but it needs to be
 795  * accessible from at least sx.h.
 796  */
 797 void
 798 _sx_xunlock_hard(struct sx *sx, uintptr_t x LOCK_FILE_LINE_ARG_DEF)
 799 {
 800         uintptr_t tid, setx;
 801         int queue, wakeup_swapper;
 802
 803         if (SCHEDULER_STOPPED())
 804                 return;
 805
 806         tid = (uintptr_t)curthread;
 807
 808         if (__predict_false(x == tid))
 809                 x = SX_READ_VALUE(sx);
 810
 811         MPASS(!(x & SX_LOCK_SHARED));
 812
 813         if (__predict_false(x & SX_LOCK_RECURSED)) {
 814                 /* The lock is recursed, unrecurse one level. */
 815                 if ((--sx->sx_recurse) == 0)
 816                         atomic_clear_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
 817                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 818                         CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, sx);
 819                 return;
 820         }
 821
 822         LOCKSTAT_PROFILE_RELEASE_RWLOCK(sx__release, sx, LOCKSTAT_WRITER);
 823         if (x == tid &&
 824             atomic_cmpset_rel_ptr(&sx->sx_lock, tid, SX_LOCK_UNLOCKED))
 825                 return;
 826
 827         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 828                 CTR2(KTR_LOCK, "%s: %p contested", __func__, sx);
 829
 830         sleepq_lock(&sx->lock_object);
 831         x = SX_READ_VALUE(sx);
 832         MPASS(x & (SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS));
 833
 834         /*
 835          * The wake up algorithm here is quite simple and probably not
 836          * ideal.  It gives precedence to shared waiters if they are
 837          * present.  For this condition, we have to preserve the
 838          * state of the exclusive waiters flag.
 839          * If interruptible sleeps left the shared queue empty avoid a
 840          * starvation for the threads sleeping on the exclusive queue by giving
 841          * them precedence and cleaning up the shared waiters bit anyway.
 842          */
 843         setx = SX_LOCK_UNLOCKED;
 844         queue = SQ_EXCLUSIVE_QUEUE;
 845         if ((x & SX_LOCK_SHARED_WAITERS) != 0 &&
 846             sleepq_sleepcnt(&sx->lock_object, SQ_SHARED_QUEUE) != 0) {
 847                 queue = SQ_SHARED_QUEUE;
 848                 setx |= (x & SX_LOCK_EXCLUSIVE_WAITERS);
 849         }
 850         atomic_store_rel_ptr(&sx->sx_lock, setx);
 851
 852         /* Wake up all the waiters for the specific queue. */
 853         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 854                 CTR3(KTR_LOCK, "%s: %p waking up all threads on %s queue",
 855                     __func__, sx, queue == SQ_SHARED_QUEUE ? "shared" :
 856                     "exclusive");
 857
 858         wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX, 0,
 859             queue);
 860         sleepq_release(&sx->lock_object);
 861         if (wakeup_swapper)
 862                 kick_proc0();
 863 }
 864
 865 static bool __always_inline
 866 __sx_slock_try(struct sx *sx, uintptr_t *xp LOCK_FILE_LINE_ARG_DEF)
 867 {
 868
 869         /*
 870          * If no other thread has an exclusive lock then try to bump up
 871          * the count of sharers.  Since we have to preserve the state
 872          * of SX_LOCK_EXCLUSIVE_WAITERS, if we fail to acquire the
 873          * shared lock loop back and retry.
 874          */
 875         while (*xp & SX_LOCK_SHARED) {
 876                 MPASS(!(*xp & SX_LOCK_SHARED_WAITERS));
 877                 if (atomic_fcmpset_acq_ptr(&sx->sx_lock, xp,
 878                     *xp + SX_ONE_SHARER)) {
 879                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 880                                 CTR4(KTR_LOCK, "%s: %p succeed %p -> %p",
 881                                     __func__, sx, (void *)*xp,
 882                                     (void *)(*xp + SX_ONE_SHARER));
 883                         return (true);
 884                 }
 885         }
 886         return (false);
 887 }
 888
 889 static int __noinline
 890 _sx_slock_hard(struct sx *sx, int opts, uintptr_t x LOCK_FILE_LINE_ARG_DEF)
 891 {
 892         GIANT_DECLARE;
 893 #ifdef ADAPTIVE_SX
 894         volatile struct thread *owner;
 895 #endif
 896 #ifdef LOCK_PROFILING
 897         uint64_t waittime = 0;
 898         int contested = 0;
 899 #endif
 900         int error = 0;
 901 #if defined(ADAPTIVE_SX) || defined(KDTRACE_HOOKS)
 902         struct lock_delay_arg lda;
 903 #endif
 904 #ifdef KDTRACE_HOOKS
 905         u_int sleep_cnt = 0;
 906         int64_t sleep_time = 0;
 907         int64_t all_time = 0;
 908 #endif
 909 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
 910         uintptr_t state;
 911 #endif
 912         int extra_work = 0;
 913
 914         if (SCHEDULER_STOPPED())
 915                 return (0);
 916
 917 #if defined(ADAPTIVE_SX)
 918         lock_delay_arg_init(&lda, &sx_delay);
 919 #elif defined(KDTRACE_HOOKS)
 920         lock_delay_arg_init(&lda, NULL);
 921 #endif
 922
 923 #ifdef HWPMC_HOOKS
 924         PMC_SOFT_CALL( , , lock, failed);
 925 #endif
 926         lock_profile_obtain_lock_failed(&sx->lock_object, &contested,
 927             &waittime);
 928
 929 #ifdef LOCK_PROFILING
 930         extra_work = 1;
 931         state = x;
 932 #elif defined(KDTRACE_HOOKS)
 933         extra_work = lockstat_enabled;
 934         if (__predict_false(extra_work)) {
 935                 all_time -= lockstat_nsecs(&sx->lock_object);
 936                 state = x;
 937         }
 938 #endif
 939
 940         /*
 941          * As with rwlocks, we don't make any attempt to try to block
 942          * shared locks once there is an exclusive waiter.
 943          */
 944         for (;;) {
 945                 if (__sx_slock_try(sx, &x LOCK_FILE_LINE_ARG))
 946                         break;
 947 #ifdef KDTRACE_HOOKS
 948                 lda.spin_cnt++;
 949 #endif
 950
 951 #ifdef ADAPTIVE_SX
 952                 /*
 953                  * If the owner is running on another CPU, spin until
 954                  * the owner stops running or the state of the lock
 955                  * changes.
 956                  */
 957                 if ((sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) {
 958                         owner = lv_sx_owner(x);
 959                         if (TD_IS_RUNNING(owner)) {
 960                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 961                                         CTR3(KTR_LOCK,
 962                                             "%s: spinning on %p held by %p",
 963                                             __func__, sx, owner);
 964                                 KTR_STATE1(KTR_SCHED, "thread",
 965                                     sched_tdname(curthread), "spinning",
 966                                     "lockname:\"%s\"", sx->lock_object.lo_name);
 967                                 GIANT_SAVE(extra_work);
 968                                 do {
 969                                         lock_delay(&lda);
 970                                         x = SX_READ_VALUE(sx);
 971                                         owner = lv_sx_owner(x);
 972                                 } while (owner != NULL && TD_IS_RUNNING(owner));
 973                                 KTR_STATE0(KTR_SCHED, "thread",
 974                                     sched_tdname(curthread), "running");
 975                                 continue;
 976                         }
 977                 }
 978 #endif
 979
 980                 /*
 981                  * Some other thread already has an exclusive lock, so
 982                  * start the process of blocking.
 983                  */
 984                 sleepq_lock(&sx->lock_object);
 985                 x = SX_READ_VALUE(sx);
 986 retry_sleepq:
 987                 /*
 988                  * The lock could have been released while we spun.
 989                  * In this case loop back and retry.
 990                  */
 991                 if (x & SX_LOCK_SHARED) {
 992                         sleepq_release(&sx->lock_object);
 993                         continue;
 994                 }
 995
 996 #ifdef ADAPTIVE_SX
 997                 /*
 998                  * If the owner is running on another CPU, spin until
 999                  * the owner stops running or the state of the lock
1000                  * changes.
1001                  */
1002                 if (!(x & SX_LOCK_SHARED) &&
1003                     (sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) {
1004                         owner = (struct thread *)SX_OWNER(x);
1005                         if (TD_IS_RUNNING(owner)) {
1006                                 sleepq_release(&sx->lock_object);
1007                                 x = SX_READ_VALUE(sx);
1008                                 continue;
1009                         }
1010                 }
1011 #endif
1012
1013                 /*
1014                  * Try to set the SX_LOCK_SHARED_WAITERS flag.  If we
1015                  * fail to set it drop the sleep queue lock and loop
1016                  * back.
1017                  */
1018                 if (!(x & SX_LOCK_SHARED_WAITERS)) {
1019                         if (!atomic_fcmpset_ptr(&sx->sx_lock, &x,
1020                             x | SX_LOCK_SHARED_WAITERS))
1021                                 goto retry_sleepq;
1022                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
1023                                 CTR2(KTR_LOCK, "%s: %p set shared waiters flag",
1024                                     __func__, sx);
1025                 }
1026
1027                 /*
1028                  * Since we have been unable to acquire the shared lock,
1029                  * we have to sleep.
1030                  */
1031                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1032                         CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
1033                             __func__, sx);
1034
1035 #ifdef KDTRACE_HOOKS
1036                 sleep_time -= lockstat_nsecs(&sx->lock_object);
1037 #endif
1038                 GIANT_SAVE(extra_work);
1039                 sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
1040                     SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
1041                     SLEEPQ_INTERRUPTIBLE : 0), SQ_SHARED_QUEUE);
1042                 if (!(opts & SX_INTERRUPTIBLE))
1043                         sleepq_wait(&sx->lock_object, 0);
1044                 else
1045                         error = sleepq_wait_sig(&sx->lock_object, 0);
1046 #ifdef KDTRACE_HOOKS
1047                 sleep_time += lockstat_nsecs(&sx->lock_object);
1048                 sleep_cnt++;
1049 #endif
1050                 if (error) {
1051                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
1052                                 CTR2(KTR_LOCK,
1053                         "%s: interruptible sleep by %p suspended by signal",
1054                                     __func__, sx);
1055                         break;
1056                 }
1057                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1058                         CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
1059                             __func__, sx);
1060                 x = SX_READ_VALUE(sx);
1061         }
1062 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
1063         if (__predict_true(!extra_work))
1064                 return (error);
1065 #endif
1066 #ifdef KDTRACE_HOOKS
1067         all_time += lockstat_nsecs(&sx->lock_object);
1068         if (sleep_time)
1069                 LOCKSTAT_RECORD4(sx__block, sx, sleep_time,
1070                     LOCKSTAT_READER, (state & SX_LOCK_SHARED) == 0,
1071                     (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
1072         if (lda.spin_cnt > sleep_cnt)
1073                 LOCKSTAT_RECORD4(sx__spin, sx, all_time - sleep_time,
1074                     LOCKSTAT_READER, (state & SX_LOCK_SHARED) == 0,
1075                     (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
1076 #endif
1077         if (error == 0) {
1078                 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
1079                     contested, waittime, file, line, LOCKSTAT_READER);
1080         }
1081         GIANT_RESTORE();
1082         return (error);
1083 }
1084
1085 int
1086 _sx_slock_int(struct sx *sx, int opts LOCK_FILE_LINE_ARG_DEF)
1087 {
1088         uintptr_t x;
1089         int error;
1090
1091         KASSERT(kdb_active != 0 || SCHEDULER_STOPPED() ||
1092             !TD_IS_IDLETHREAD(curthread),
1093             ("sx_slock() by idle thread %p on sx %s @ %s:%d",
1094             curthread, sx->lock_object.lo_name, file, line));
1095         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
1096             ("sx_slock() of destroyed sx @ %s:%d", file, line));
1097         WITNESS_CHECKORDER(&sx->lock_object, LOP_NEWORDER, file, line, NULL);
1098
1099         error = 0;
1100         x = SX_READ_VALUE(sx);
1101         if (__predict_false(LOCKSTAT_OOL_PROFILE_ENABLED(sx__acquire) ||
1102             !__sx_slock_try(sx, &x LOCK_FILE_LINE_ARG)))
1103                 error = _sx_slock_hard(sx, opts, x LOCK_FILE_LINE_ARG);
1104         if (error == 0) {
1105                 LOCK_LOG_LOCK("SLOCK", &sx->lock_object, 0, 0, file, line);
1106                 WITNESS_LOCK(&sx->lock_object, 0, file, line);
1107                 TD_LOCKS_INC(curthread);
1108         }
1109         return (error);
1110 }
1111
1112 int
1113 _sx_slock(struct sx *sx, int opts, const char *file, int line)
1114 {
1115
1116         return (_sx_slock_int(sx, opts LOCK_FILE_LINE_ARG));
1117 }
1118
1119 static bool __always_inline
1120 _sx_sunlock_try(struct sx *sx, uintptr_t *xp)
1121 {
1122
1123         for (;;) {
1124                 /*
1125                  * We should never have sharers while at least one thread
1126                  * holds a shared lock.
1127                  */
1128                 KASSERT(!(*xp & SX_LOCK_SHARED_WAITERS),
1129                     ("%s: waiting sharers", __func__));
1130
1131                 /*
1132                  * See if there is more than one shared lock held.  If
1133                  * so, just drop one and return.
1134                  */
1135                 if (SX_SHARERS(*xp) > 1) {
1136                         if (atomic_fcmpset_rel_ptr(&sx->sx_lock, xp,
1137                             *xp - SX_ONE_SHARER)) {
1138                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1139                                         CTR4(KTR_LOCK,
1140                                             "%s: %p succeeded %p -> %p",
1141                                             __func__, sx, (void *)*xp,
1142                                             (void *)(*xp - SX_ONE_SHARER));
1143                                 return (true);
1144                         }
1145                         continue;
1146                 }
1147
1148                 /*
1149                  * If there aren't any waiters for an exclusive lock,
1150                  * then try to drop it quickly.
1151                  */
1152                 if (!(*xp & SX_LOCK_EXCLUSIVE_WAITERS)) {
1153                         MPASS(*xp == SX_SHARERS_LOCK(1));
1154                         *xp = SX_SHARERS_LOCK(1);
1155                         if (atomic_fcmpset_rel_ptr(&sx->sx_lock,
1156                             xp, SX_LOCK_UNLOCKED)) {
1157                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1158                                         CTR2(KTR_LOCK, "%s: %p last succeeded",
1159                                             __func__, sx);
1160                                 return (true);
1161                         }
1162                         continue;
1163                 }
1164                 break;
1165         }
1166         return (false);
1167 }
1168
1169 static void __noinline
1170 _sx_sunlock_hard(struct sx *sx, uintptr_t x LOCK_FILE_LINE_ARG_DEF)
1171 {
1172         int wakeup_swapper;
1173         uintptr_t setx;
1174
1175         if (SCHEDULER_STOPPED())
1176                 return;
1177
1178         if (_sx_sunlock_try(sx, &x))
1179                 goto out_lockstat;
1180
1181         /*
1182          * At this point, there should just be one sharer with
1183          * exclusive waiters.
1184          */
1185         MPASS(x == (SX_SHARERS_LOCK(1) | SX_LOCK_EXCLUSIVE_WAITERS));
1186
1187         sleepq_lock(&sx->lock_object);
1188         x = SX_READ_VALUE(sx);
1189         for (;;) {
1190                 MPASS(x & SX_LOCK_EXCLUSIVE_WAITERS);
1191                 MPASS(!(x & SX_LOCK_SHARED_WAITERS));
1192                 /*
1193                  * Wake up semantic here is quite simple:
1194                  * Just wake up all the exclusive waiters.
1195                  * Note that the state of the lock could have changed,
1196                  * so if it fails loop back and retry.
1197                  */
1198                 setx = x - SX_ONE_SHARER;
1199                 setx &= ~SX_LOCK_EXCLUSIVE_WAITERS;
1200                 if (!atomic_fcmpset_rel_ptr(&sx->sx_lock, &x, setx))
1201                         continue;
1202                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1203                         CTR2(KTR_LOCK, "%s: %p waking up all thread on"
1204                             "exclusive queue", __func__, sx);
1205                 wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX,
1206                     0, SQ_EXCLUSIVE_QUEUE);
1207                 break;
1208         }
1209         sleepq_release(&sx->lock_object);
1210         if (wakeup_swapper)
1211                 kick_proc0();
1212 out_lockstat:
1213         LOCKSTAT_PROFILE_RELEASE_RWLOCK(sx__release, sx, LOCKSTAT_READER);
1214 }
1215
1216 void
1217 _sx_sunlock_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
1218 {
1219         uintptr_t x;
1220
1221         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
1222             ("sx_sunlock() of destroyed sx @ %s:%d", file, line));
1223         _sx_assert(sx, SA_SLOCKED, file, line);
1224         WITNESS_UNLOCK(&sx->lock_object, 0, file, line);
1225         LOCK_LOG_LOCK("SUNLOCK", &sx->lock_object, 0, 0, file, line);
1226
1227         x = SX_READ_VALUE(sx);
1228         if (__predict_false(LOCKSTAT_OOL_PROFILE_ENABLED(sx__release) ||
1229             !_sx_sunlock_try(sx, &x)))
1230                 _sx_sunlock_hard(sx, x LOCK_FILE_LINE_ARG);
1231
1232         TD_LOCKS_DEC(curthread);
1233 }
1234
1235 void
1236 _sx_sunlock(struct sx *sx, const char *file, int line)
1237 {
1238
1239         _sx_sunlock_int(sx LOCK_FILE_LINE_ARG);
1240 }
1241
1242 #ifdef INVARIANT_SUPPORT
1243 #ifndef INVARIANTS
1244 #undef  _sx_assert
1245 #endif
1246
1247 /*
1248  * In the non-WITNESS case, sx_assert() can only detect that at least
1249  * *some* thread owns an slock, but it cannot guarantee that *this*
1250  * thread owns an slock.
1251  */
1252 void
1253 _sx_assert(const struct sx *sx, int what, const char *file, int line)
1254 {
1255 #ifndef WITNESS
1256         int slocked = 0;
1257 #endif
1258
1259         if (panicstr != NULL)
1260                 return;
1261         switch (what) {
1262         case SA_SLOCKED:
1263         case SA_SLOCKED | SA_NOTRECURSED:
1264         case SA_SLOCKED | SA_RECURSED:
1265 #ifndef WITNESS
1266                 slocked = 1;
1267                 /* FALLTHROUGH */
1268 #endif
1269         case SA_LOCKED:
1270         case SA_LOCKED | SA_NOTRECURSED:
1271         case SA_LOCKED | SA_RECURSED:
1272 #ifdef WITNESS
1273                 witness_assert(&sx->lock_object, what, file, line);
1274 #else
1275                 /*
1276                  * If some other thread has an exclusive lock or we
1277                  * have one and are asserting a shared lock, fail.
1278                  * Also, if no one has a lock at all, fail.
1279                  */
1280                 if (sx->sx_lock == SX_LOCK_UNLOCKED ||
1281                     (!(sx->sx_lock & SX_LOCK_SHARED) && (slocked ||
1282                     sx_xholder(sx) != curthread)))
1283                         panic("Lock %s not %slocked @ %s:%d\n",
1284                             sx->lock_object.lo_name, slocked ? "share " : "",
1285                             file, line);
1286
1287                 if (!(sx->sx_lock & SX_LOCK_SHARED)) {
1288                         if (sx_recursed(sx)) {
1289                                 if (what & SA_NOTRECURSED)
1290                                         panic("Lock %s recursed @ %s:%d\n",
1291                                             sx->lock_object.lo_name, file,
1292                                             line);
1293                         } else if (what & SA_RECURSED)
1294                                 panic("Lock %s not recursed @ %s:%d\n",
1295                                     sx->lock_object.lo_name, file, line);
1296                 }
1297 #endif
1298                 break;
1299         case SA_XLOCKED:
1300         case SA_XLOCKED | SA_NOTRECURSED:
1301         case SA_XLOCKED | SA_RECURSED:
1302                 if (sx_xholder(sx) != curthread)
1303                         panic("Lock %s not exclusively locked @ %s:%d\n",
1304                             sx->lock_object.lo_name, file, line);
1305                 if (sx_recursed(sx)) {
1306                         if (what & SA_NOTRECURSED)
1307                                 panic("Lock %s recursed @ %s:%d\n",
1308                                     sx->lock_object.lo_name, file, line);
1309                 } else if (what & SA_RECURSED)
1310                         panic("Lock %s not recursed @ %s:%d\n",
1311                             sx->lock_object.lo_name, file, line);
1312                 break;
1313         case SA_UNLOCKED:
1314 #ifdef WITNESS
1315                 witness_assert(&sx->lock_object, what, file, line);
1316 #else
1317                 /*
1318                  * If we hold an exclusve lock fail.  We can't
1319                  * reliably check to see if we hold a shared lock or
1320                  * not.
1321                  */
1322                 if (sx_xholder(sx) == curthread)
1323                         panic("Lock %s exclusively locked @ %s:%d\n",
1324                             sx->lock_object.lo_name, file, line);
1325 #endif
1326                 break;
1327         default:
1328                 panic("Unknown sx lock assertion: %d @ %s:%d", what, file,
1329                     line);
1330         }
1331 }
1332 #endif  /* INVARIANT_SUPPORT */
1333
1334 #ifdef DDB
1335 static void
1336 db_show_sx(const struct lock_object *lock)
1337 {
1338         struct thread *td;
1339         const struct sx *sx;
1340
1341         sx = (const struct sx *)lock;
1342
1343         db_printf(" state: ");
1344         if (sx->sx_lock == SX_LOCK_UNLOCKED)
1345                 db_printf("UNLOCKED\n");
1346         else if (sx->sx_lock == SX_LOCK_DESTROYED) {
1347                 db_printf("DESTROYED\n");
1348                 return;
1349         } else if (sx->sx_lock & SX_LOCK_SHARED)
1350                 db_printf("SLOCK: %ju\n", (uintmax_t)SX_SHARERS(sx->sx_lock));
1351         else {
1352                 td = sx_xholder(sx);
1353                 db_printf("XLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1354                     td->td_tid, td->td_proc->p_pid, td->td_name);
1355                 if (sx_recursed(sx))
1356                         db_printf(" recursed: %d\n", sx->sx_recurse);
1357         }
1358
1359         db_printf(" waiters: ");
1360         switch(sx->sx_lock &
1361             (SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS)) {
1362         case SX_LOCK_SHARED_WAITERS:
1363                 db_printf("shared\n");
1364                 break;
1365         case SX_LOCK_EXCLUSIVE_WAITERS:
1366                 db_printf("exclusive\n");
1367                 break;
1368         case SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS:
1369                 db_printf("exclusive and shared\n");
1370                 break;
1371         default:
1372                 db_printf("none\n");
1373         }
1374 }
1375
1376 /*
1377  * Check to see if a thread that is blocked on a sleep queue is actually
1378  * blocked on an sx lock.  If so, output some details and return true.
1379  * If the lock has an exclusive owner, return that in *ownerp.
1380  */
1381 int
1382 sx_chain(struct thread *td, struct thread **ownerp)
1383 {
1384         struct sx *sx;
1385
1386         /*
1387          * Check to see if this thread is blocked on an sx lock.
1388          * First, we check the lock class.  If that is ok, then we
1389          * compare the lock name against the wait message.
1390          */
1391         sx = td->td_wchan;
1392         if (LOCK_CLASS(&sx->lock_object) != &lock_class_sx ||
1393             sx->lock_object.lo_name != td->td_wmesg)
1394                 return (0);
1395
1396         /* We think we have an sx lock, so output some details. */
1397         db_printf("blocked on sx \"%s\" ", td->td_wmesg);
1398         *ownerp = sx_xholder(sx);
1399         if (sx->sx_lock & SX_LOCK_SHARED)
1400                 db_printf("SLOCK (count %ju)\n",
1401                     (uintmax_t)SX_SHARERS(sx->sx_lock));
1402         else
1403                 db_printf("XLOCK\n");
1404         return (1);
1405 }
1406 #endif