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         enum { READERS, WRITER } sleep_reason;
 537         bool adaptive;
 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                         sleep_reason = WRITER;
 632                         owner = lv_sx_owner(x);
 633                         if (!TD_IS_RUNNING(owner))
 634                                 goto sleepq;
 635                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 636                                 CTR3(KTR_LOCK, "%s: spinning on %p held by %p",
 637                                     __func__, sx, owner);
 638                         KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
 639                             "spinning", "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 && TD_IS_RUNNING(owner));
 646                         KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
 647                             "running");
 648                         continue;
 649                 } else if (SX_SHARERS(x) > 0) {
 650                         sleep_reason = READERS;
 651                         if (spintries == asx_retries)
 652                                 goto sleepq;
 653                         spintries++;
 654                         KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
 655                             "spinning", "lockname:\"%s\"",
 656                             sx->lock_object.lo_name);
 657                         for (i = 0; i < asx_loops; i += n) {
 658                                 n = SX_SHARERS(x);
 659                                 lock_delay_spin(n);
 660                                 x = SX_READ_VALUE(sx);
 661                                 if ((x & SX_LOCK_SHARED) == 0 ||
 662                                     SX_SHARERS(x) == 0)
 663                                         break;
 664                         }
 665 #ifdef KDTRACE_HOOKS
 666                         lda.spin_cnt += i;
 667 #endif
 668                         KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
 669                             "running");
 670                         if (i < asx_loops)
 671                                 continue;
 672                 }
 673 sleepq:
 674 #endif
 675                 sleepq_lock(&sx->lock_object);
 676                 x = SX_READ_VALUE(sx);
 677 retry_sleepq:
 678
 679                 /*
 680                  * If the lock was released while spinning on the
 681                  * sleep queue chain lock, try again.
 682                  */
 683                 if (x == SX_LOCK_UNLOCKED) {
 684                         sleepq_release(&sx->lock_object);
 685                         continue;
 686                 }
 687
 688 #ifdef ADAPTIVE_SX
 689                 /*
 690                  * The current lock owner might have started executing
 691                  * on another CPU (or the lock could have changed
 692                  * owners) while we were waiting on the sleep queue
 693                  * chain lock.  If so, drop the sleep queue lock and try
 694                  * again.
 695                  */
 696                 if (adaptive) {
 697                         if (!(x & SX_LOCK_SHARED)) {
 698                                 owner = (struct thread *)SX_OWNER(x);
 699                                 if (TD_IS_RUNNING(owner)) {
 700                                         sleepq_release(&sx->lock_object);
 701                                         continue;
 702                                 }
 703                         } else if (SX_SHARERS(x) > 0 && sleep_reason == WRITER) {
 704                                 sleepq_release(&sx->lock_object);
 705                                 continue;
 706                         }
 707                 }
 708 #endif
 709
 710                 /*
 711                  * If an exclusive lock was released with both shared
 712                  * and exclusive waiters and a shared waiter hasn't
 713                  * woken up and acquired the lock yet, sx_lock will be
 714                  * set to SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS.
 715                  * If we see that value, try to acquire it once.  Note
 716                  * that we have to preserve SX_LOCK_EXCLUSIVE_WAITERS
 717                  * as there are other exclusive waiters still.  If we
 718                  * fail, restart the loop.
 719                  */
 720                 if (x == (SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS)) {
 721                         if (!atomic_fcmpset_acq_ptr(&sx->sx_lock, &x,
 722                             tid | SX_LOCK_EXCLUSIVE_WAITERS))
 723                                 goto retry_sleepq;
 724                         sleepq_release(&sx->lock_object);
 725                         CTR2(KTR_LOCK, "%s: %p claimed by new writer",
 726                             __func__, sx);
 727                         break;
 728                 }
 729
 730                 /*
 731                  * Try to set the SX_LOCK_EXCLUSIVE_WAITERS.  If we fail,
 732                  * than loop back and retry.
 733                  */
 734                 if (!(x & SX_LOCK_EXCLUSIVE_WAITERS)) {
 735                         if (!atomic_fcmpset_ptr(&sx->sx_lock, &x,
 736                             x | SX_LOCK_EXCLUSIVE_WAITERS)) {
 737                                 goto retry_sleepq;
 738                         }
 739                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 740                                 CTR2(KTR_LOCK, "%s: %p set excl waiters flag",
 741                                     __func__, sx);
 742                 }
 743
 744                 /*
 745                  * Since we have been unable to acquire the exclusive
 746                  * lock and the exclusive waiters flag is set, we have
 747                  * to sleep.
 748                  */
 749                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 750                         CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
 751                             __func__, sx);
 752
 753 #ifdef KDTRACE_HOOKS
 754                 sleep_time -= lockstat_nsecs(&sx->lock_object);
 755 #endif
 756                 sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
 757                     SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
 758                     SLEEPQ_INTERRUPTIBLE : 0), SQ_EXCLUSIVE_QUEUE);
 759                 if (!(opts & SX_INTERRUPTIBLE))
 760                         sleepq_wait(&sx->lock_object, 0);
 761                 else
 762                         error = sleepq_wait_sig(&sx->lock_object, 0);
 763 #ifdef KDTRACE_HOOKS
 764                 sleep_time += lockstat_nsecs(&sx->lock_object);
 765                 sleep_cnt++;
 766 #endif
 767                 if (error) {
 768                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 769                                 CTR2(KTR_LOCK,
 770                         "%s: interruptible sleep by %p suspended by signal",
 771                                     __func__, sx);
 772                         break;
 773                 }
 774                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 775                         CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
 776                             __func__, sx);
 777                 x = SX_READ_VALUE(sx);
 778         }
 779 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
 780         if (__predict_true(!extra_work))
 781                 return (error);
 782 #endif
 783 #ifdef KDTRACE_HOOKS
 784         all_time += lockstat_nsecs(&sx->lock_object);
 785         if (sleep_time)
 786                 LOCKSTAT_RECORD4(sx__block, sx, sleep_time,
 787                     LOCKSTAT_WRITER, (state & SX_LOCK_SHARED) == 0,
 788                     (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
 789         if (lda.spin_cnt > sleep_cnt)
 790                 LOCKSTAT_RECORD4(sx__spin, sx, all_time - sleep_time,
 791                     LOCKSTAT_WRITER, (state & SX_LOCK_SHARED) == 0,
 792                     (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
 793 #endif
 794         if (!error)
 795                 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
 796                     contested, waittime, file, line, LOCKSTAT_WRITER);
 797         GIANT_RESTORE();
 798         return (error);
 799 }
 800
 801 /*
 802  * This function represents the so-called 'hard case' for sx_xunlock
 803  * operation.  All 'easy case' failures are redirected to this.  Note
 804  * that ideally this would be a static function, but it needs to be
 805  * accessible from at least sx.h.
 806  */
 807 void
 808 _sx_xunlock_hard(struct sx *sx, uintptr_t x LOCK_FILE_LINE_ARG_DEF)
 809 {
 810         uintptr_t tid, setx;
 811         int queue, wakeup_swapper;
 812
 813         if (SCHEDULER_STOPPED())
 814                 return;
 815
 816         tid = (uintptr_t)curthread;
 817
 818         if (__predict_false(x == tid))
 819                 x = SX_READ_VALUE(sx);
 820
 821         MPASS(!(x & SX_LOCK_SHARED));
 822
 823         if (__predict_false(x & SX_LOCK_RECURSED)) {
 824                 /* The lock is recursed, unrecurse one level. */
 825                 if ((--sx->sx_recurse) == 0)
 826                         atomic_clear_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
 827                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 828                         CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, sx);
 829                 return;
 830         }
 831
 832         LOCKSTAT_PROFILE_RELEASE_RWLOCK(sx__release, sx, LOCKSTAT_WRITER);
 833         if (x == tid &&
 834             atomic_cmpset_rel_ptr(&sx->sx_lock, tid, SX_LOCK_UNLOCKED))
 835                 return;
 836
 837         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 838                 CTR2(KTR_LOCK, "%s: %p contested", __func__, sx);
 839
 840         sleepq_lock(&sx->lock_object);
 841         x = SX_READ_VALUE(sx);
 842         MPASS(x & (SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS));
 843
 844         /*
 845          * The wake up algorithm here is quite simple and probably not
 846          * ideal.  It gives precedence to shared waiters if they are
 847          * present.  For this condition, we have to preserve the
 848          * state of the exclusive waiters flag.
 849          * If interruptible sleeps left the shared queue empty avoid a
 850          * starvation for the threads sleeping on the exclusive queue by giving
 851          * them precedence and cleaning up the shared waiters bit anyway.
 852          */
 853         setx = SX_LOCK_UNLOCKED;
 854         queue = SQ_EXCLUSIVE_QUEUE;
 855         if ((x & SX_LOCK_SHARED_WAITERS) != 0 &&
 856             sleepq_sleepcnt(&sx->lock_object, SQ_SHARED_QUEUE) != 0) {
 857                 queue = SQ_SHARED_QUEUE;
 858                 setx |= (x & SX_LOCK_EXCLUSIVE_WAITERS);
 859         }
 860         atomic_store_rel_ptr(&sx->sx_lock, setx);
 861
 862         /* Wake up all the waiters for the specific queue. */
 863         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 864                 CTR3(KTR_LOCK, "%s: %p waking up all threads on %s queue",
 865                     __func__, sx, queue == SQ_SHARED_QUEUE ? "shared" :
 866                     "exclusive");
 867
 868         wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX, 0,
 869             queue);
 870         sleepq_release(&sx->lock_object);
 871         if (wakeup_swapper)
 872                 kick_proc0();
 873 }
 874
 875 static bool __always_inline
 876 __sx_slock_try(struct sx *sx, uintptr_t *xp LOCK_FILE_LINE_ARG_DEF)
 877 {
 878
 879         /*
 880          * If no other thread has an exclusive lock then try to bump up
 881          * the count of sharers.  Since we have to preserve the state
 882          * of SX_LOCK_EXCLUSIVE_WAITERS, if we fail to acquire the
 883          * shared lock loop back and retry.
 884          */
 885         while (*xp & SX_LOCK_SHARED) {
 886                 MPASS(!(*xp & SX_LOCK_SHARED_WAITERS));
 887                 if (atomic_fcmpset_acq_ptr(&sx->sx_lock, xp,
 888                     *xp + SX_ONE_SHARER)) {
 889                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 890                                 CTR4(KTR_LOCK, "%s: %p succeed %p -> %p",
 891                                     __func__, sx, (void *)*xp,
 892                                     (void *)(*xp + SX_ONE_SHARER));
 893                         return (true);
 894                 }
 895         }
 896         return (false);
 897 }
 898
 899 static int __noinline
 900 _sx_slock_hard(struct sx *sx, int opts, uintptr_t x LOCK_FILE_LINE_ARG_DEF)
 901 {
 902         GIANT_DECLARE;
 903 #ifdef ADAPTIVE_SX
 904         volatile struct thread *owner;
 905         bool adaptive;
 906 #endif
 907 #ifdef LOCK_PROFILING
 908         uint64_t waittime = 0;
 909         int contested = 0;
 910 #endif
 911         int error = 0;
 912 #if defined(ADAPTIVE_SX) || defined(KDTRACE_HOOKS)
 913         struct lock_delay_arg lda;
 914 #endif
 915 #ifdef KDTRACE_HOOKS
 916         u_int sleep_cnt = 0;
 917         int64_t sleep_time = 0;
 918         int64_t all_time = 0;
 919 #endif
 920 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
 921         uintptr_t state;
 922 #endif
 923         int extra_work = 0;
 924
 925         if (SCHEDULER_STOPPED())
 926                 return (0);
 927
 928 #if defined(ADAPTIVE_SX)
 929         lock_delay_arg_init(&lda, &sx_delay);
 930 #elif defined(KDTRACE_HOOKS)
 931         lock_delay_arg_init(&lda, NULL);
 932 #endif
 933
 934 #ifdef ADAPTIVE_SX
 935         adaptive = ((sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0);
 936 #endif
 937
 938 #ifdef HWPMC_HOOKS
 939         PMC_SOFT_CALL( , , lock, failed);
 940 #endif
 941         lock_profile_obtain_lock_failed(&sx->lock_object, &contested,
 942             &waittime);
 943
 944 #ifdef LOCK_PROFILING
 945         extra_work = 1;
 946         state = x;
 947 #elif defined(KDTRACE_HOOKS)
 948         extra_work = lockstat_enabled;
 949         if (__predict_false(extra_work)) {
 950                 all_time -= lockstat_nsecs(&sx->lock_object);
 951                 state = x;
 952         }
 953 #endif
 954 #ifndef INVARIANTS
 955         GIANT_SAVE(extra_work);
 956 #endif
 957
 958         /*
 959          * As with rwlocks, we don't make any attempt to try to block
 960          * shared locks once there is an exclusive waiter.
 961          */
 962         for (;;) {
 963                 if (__sx_slock_try(sx, &x LOCK_FILE_LINE_ARG))
 964                         break;
 965 #ifdef INVARIANTS
 966                 GIANT_SAVE(extra_work);
 967 #endif
 968 #ifdef KDTRACE_HOOKS
 969                 lda.spin_cnt++;
 970 #endif
 971
 972 #ifdef ADAPTIVE_SX
 973                 if (__predict_false(!adaptive))
 974                         goto sleepq;
 975                 /*
 976                  * If the owner is running on another CPU, spin until
 977                  * the owner stops running or the state of the lock
 978                  * changes.
 979                  */
 980                 owner = lv_sx_owner(x);
 981                 if (TD_IS_RUNNING(owner)) {
 982                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 983                                 CTR3(KTR_LOCK,
 984                                     "%s: spinning on %p held by %p",
 985                                     __func__, sx, owner);
 986                         KTR_STATE1(KTR_SCHED, "thread",
 987                             sched_tdname(curthread), "spinning",
 988                             "lockname:\"%s\"", sx->lock_object.lo_name);
 989                         do {
 990                                 lock_delay(&lda);
 991                                 x = SX_READ_VALUE(sx);
 992                                 owner = lv_sx_owner(x);
 993                         } while (owner != NULL && TD_IS_RUNNING(owner));
 994                         KTR_STATE0(KTR_SCHED, "thread",
 995                             sched_tdname(curthread), "running");
 996                         continue;
 997                 }
 998 sleepq:
 999 #endif
1000
1001                 /*
1002                  * Some other thread already has an exclusive lock, so
1003                  * start the process of blocking.
1004                  */
1005                 sleepq_lock(&sx->lock_object);
1006                 x = SX_READ_VALUE(sx);
1007 retry_sleepq:
1008                 /*
1009                  * The lock could have been released while we spun.
1010                  * In this case loop back and retry.
1011                  */
1012                 if (x & SX_LOCK_SHARED) {
1013                         sleepq_release(&sx->lock_object);
1014                         continue;
1015                 }
1016
1017 #ifdef ADAPTIVE_SX
1018                 /*
1019                  * If the owner is running on another CPU, spin until
1020                  * the owner stops running or the state of the lock
1021                  * changes.
1022                  */
1023                 if (!(x & SX_LOCK_SHARED) && adaptive) {
1024                         owner = (struct thread *)SX_OWNER(x);
1025                         if (TD_IS_RUNNING(owner)) {
1026                                 sleepq_release(&sx->lock_object);
1027                                 x = SX_READ_VALUE(sx);
1028                                 continue;
1029                         }
1030                 }
1031 #endif
1032
1033                 /*
1034                  * Try to set the SX_LOCK_SHARED_WAITERS flag.  If we
1035                  * fail to set it drop the sleep queue lock and loop
1036                  * back.
1037                  */
1038                 if (!(x & SX_LOCK_SHARED_WAITERS)) {
1039                         if (!atomic_fcmpset_ptr(&sx->sx_lock, &x,
1040                             x | SX_LOCK_SHARED_WAITERS))
1041                                 goto retry_sleepq;
1042                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
1043                                 CTR2(KTR_LOCK, "%s: %p set shared waiters flag",
1044                                     __func__, sx);
1045                 }
1046
1047                 /*
1048                  * Since we have been unable to acquire the shared lock,
1049                  * we have to sleep.
1050                  */
1051                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1052                         CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
1053                             __func__, sx);
1054
1055 #ifdef KDTRACE_HOOKS
1056                 sleep_time -= lockstat_nsecs(&sx->lock_object);
1057 #endif
1058                 sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
1059                     SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
1060                     SLEEPQ_INTERRUPTIBLE : 0), SQ_SHARED_QUEUE);
1061                 if (!(opts & SX_INTERRUPTIBLE))
1062                         sleepq_wait(&sx->lock_object, 0);
1063                 else
1064                         error = sleepq_wait_sig(&sx->lock_object, 0);
1065 #ifdef KDTRACE_HOOKS
1066                 sleep_time += lockstat_nsecs(&sx->lock_object);
1067                 sleep_cnt++;
1068 #endif
1069                 if (error) {
1070                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
1071                                 CTR2(KTR_LOCK,
1072                         "%s: interruptible sleep by %p suspended by signal",
1073                                     __func__, sx);
1074                         break;
1075                 }
1076                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1077                         CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
1078                             __func__, sx);
1079                 x = SX_READ_VALUE(sx);
1080         }
1081 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
1082         if (__predict_true(!extra_work))
1083                 return (error);
1084 #endif
1085 #ifdef KDTRACE_HOOKS
1086         all_time += lockstat_nsecs(&sx->lock_object);
1087         if (sleep_time)
1088                 LOCKSTAT_RECORD4(sx__block, sx, sleep_time,
1089                     LOCKSTAT_READER, (state & SX_LOCK_SHARED) == 0,
1090                     (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
1091         if (lda.spin_cnt > sleep_cnt)
1092                 LOCKSTAT_RECORD4(sx__spin, sx, all_time - sleep_time,
1093                     LOCKSTAT_READER, (state & SX_LOCK_SHARED) == 0,
1094                     (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
1095 #endif
1096         if (error == 0) {
1097                 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
1098                     contested, waittime, file, line, LOCKSTAT_READER);
1099         }
1100         GIANT_RESTORE();
1101         return (error);
1102 }
1103
1104 int
1105 _sx_slock_int(struct sx *sx, int opts LOCK_FILE_LINE_ARG_DEF)
1106 {
1107         uintptr_t x;
1108         int error;
1109
1110         KASSERT(kdb_active != 0 || SCHEDULER_STOPPED() ||
1111             !TD_IS_IDLETHREAD(curthread),
1112             ("sx_slock() by idle thread %p on sx %s @ %s:%d",
1113             curthread, sx->lock_object.lo_name, file, line));
1114         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
1115             ("sx_slock() of destroyed sx @ %s:%d", file, line));
1116         WITNESS_CHECKORDER(&sx->lock_object, LOP_NEWORDER, file, line, NULL);
1117
1118         error = 0;
1119         x = SX_READ_VALUE(sx);
1120         if (__predict_false(LOCKSTAT_PROFILE_ENABLED(sx__acquire) ||
1121             !__sx_slock_try(sx, &x LOCK_FILE_LINE_ARG)))
1122                 error = _sx_slock_hard(sx, opts, x LOCK_FILE_LINE_ARG);
1123         else
1124                 lock_profile_obtain_lock_success(&sx->lock_object, 0, 0,
1125                     file, line);
1126         if (error == 0) {
1127                 LOCK_LOG_LOCK("SLOCK", &sx->lock_object, 0, 0, file, line);
1128                 WITNESS_LOCK(&sx->lock_object, 0, file, line);
1129                 TD_LOCKS_INC(curthread);
1130         }
1131         return (error);
1132 }
1133
1134 int
1135 _sx_slock(struct sx *sx, int opts, const char *file, int line)
1136 {
1137
1138         return (_sx_slock_int(sx, opts LOCK_FILE_LINE_ARG));
1139 }
1140
1141 static bool __always_inline
1142 _sx_sunlock_try(struct sx *sx, uintptr_t *xp)
1143 {
1144
1145         for (;;) {
1146                 /*
1147                  * We should never have sharers while at least one thread
1148                  * holds a shared lock.
1149                  */
1150                 KASSERT(!(*xp & SX_LOCK_SHARED_WAITERS),
1151                     ("%s: waiting sharers", __func__));
1152
1153                 /*
1154                  * See if there is more than one shared lock held.  If
1155                  * so, just drop one and return.
1156                  */
1157                 if (SX_SHARERS(*xp) > 1) {
1158                         if (atomic_fcmpset_rel_ptr(&sx->sx_lock, xp,
1159                             *xp - SX_ONE_SHARER)) {
1160                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1161                                         CTR4(KTR_LOCK,
1162                                             "%s: %p succeeded %p -> %p",
1163                                             __func__, sx, (void *)*xp,
1164                                             (void *)(*xp - SX_ONE_SHARER));
1165                                 return (true);
1166                         }
1167                         continue;
1168                 }
1169
1170                 /*
1171                  * If there aren't any waiters for an exclusive lock,
1172                  * then try to drop it quickly.
1173                  */
1174                 if (!(*xp & SX_LOCK_EXCLUSIVE_WAITERS)) {
1175                         MPASS(*xp == SX_SHARERS_LOCK(1));
1176                         *xp = SX_SHARERS_LOCK(1);
1177                         if (atomic_fcmpset_rel_ptr(&sx->sx_lock,
1178                             xp, SX_LOCK_UNLOCKED)) {
1179                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1180                                         CTR2(KTR_LOCK, "%s: %p last succeeded",
1181                                             __func__, sx);
1182                                 return (true);
1183                         }
1184                         continue;
1185                 }
1186                 break;
1187         }
1188         return (false);
1189 }
1190
1191 static void __noinline
1192 _sx_sunlock_hard(struct sx *sx, uintptr_t x LOCK_FILE_LINE_ARG_DEF)
1193 {
1194         int wakeup_swapper = 0;
1195         uintptr_t setx;
1196
1197         if (SCHEDULER_STOPPED())
1198                 return;
1199
1200         if (_sx_sunlock_try(sx, &x))
1201                 goto out_lockstat;
1202
1203         /*
1204          * At this point, there should just be one sharer with
1205          * exclusive waiters.
1206          */
1207         MPASS(x == (SX_SHARERS_LOCK(1) | SX_LOCK_EXCLUSIVE_WAITERS));
1208
1209         sleepq_lock(&sx->lock_object);
1210         x = SX_READ_VALUE(sx);
1211         for (;;) {
1212                 MPASS(x & SX_LOCK_EXCLUSIVE_WAITERS);
1213                 MPASS(!(x & SX_LOCK_SHARED_WAITERS));
1214                 if (_sx_sunlock_try(sx, &x))
1215                         break;
1216
1217                 /*
1218                  * Wake up semantic here is quite simple:
1219                  * Just wake up all the exclusive waiters.
1220                  * Note that the state of the lock could have changed,
1221                  * so if it fails loop back and retry.
1222                  */
1223                 setx = x - SX_ONE_SHARER;
1224                 setx &= ~SX_LOCK_EXCLUSIVE_WAITERS;
1225                 if (!atomic_fcmpset_rel_ptr(&sx->sx_lock, &x, setx))
1226                         continue;
1227                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1228                         CTR2(KTR_LOCK, "%s: %p waking up all thread on"
1229                             "exclusive queue", __func__, sx);
1230                 wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX,
1231                     0, SQ_EXCLUSIVE_QUEUE);
1232                 break;
1233         }
1234         sleepq_release(&sx->lock_object);
1235         if (wakeup_swapper)
1236                 kick_proc0();
1237 out_lockstat:
1238         LOCKSTAT_PROFILE_RELEASE_RWLOCK(sx__release, sx, LOCKSTAT_READER);
1239 }
1240
1241 void
1242 _sx_sunlock_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
1243 {
1244         uintptr_t x;
1245
1246         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
1247             ("sx_sunlock() of destroyed sx @ %s:%d", file, line));
1248         _sx_assert(sx, SA_SLOCKED, file, line);
1249         WITNESS_UNLOCK(&sx->lock_object, 0, file, line);
1250         LOCK_LOG_LOCK("SUNLOCK", &sx->lock_object, 0, 0, file, line);
1251
1252         x = SX_READ_VALUE(sx);
1253         if (__predict_false(LOCKSTAT_PROFILE_ENABLED(sx__release) ||
1254             !_sx_sunlock_try(sx, &x)))
1255                 _sx_sunlock_hard(sx, x LOCK_FILE_LINE_ARG);
1256         else
1257                 lock_profile_release_lock(&sx->lock_object);
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