sys/kern/kern_sx.c

   1 /*-
   2  * Copyright (c) 2007 Attilio Rao <attilio@freebsd.org>
   3  * Copyright (c) 2001 Jason Evans <jasone@freebsd.org>
   4  * All rights reserved.
   5  *
   6  * Redistribution and use in source and binary forms, with or without
   7  * modification, are permitted provided that the following conditions
   8  * are met:
   9  * 1. Redistributions of source code must retain the above copyright
  10  *    notice(s), this list of conditions and the following disclaimer as
  11  *    the first lines of this file unmodified other than the possible
  12  *    addition of one or more copyright notices.
  13  * 2. Redistributions in binary form must reproduce the above copyright
  14  *    notice(s), this list of conditions and the following disclaimer in the
  15  *    documentation and/or other materials provided with the distribution.
  16  *
  17  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDER(S) ``AS IS'' AND ANY
  18  * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
  19  * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
  20  * DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT HOLDER(S) BE LIABLE FOR ANY
  21  * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
  22  * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
  23  * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
  24  * CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH
  27  * DAMAGE.
  28  */
  29
  30 /*
  31  * Shared/exclusive locks.  This implementation attempts to ensure
  32  * deterministic lock granting behavior, so that slocks and xlocks are
  33  * interleaved.
  34  *
  35  * Priority propagation will not generally raise the priority of lock holders,
  36  * so should not be relied upon in combination with sx locks.
  37  */
  38
  39 #include "opt_ddb.h"
  40 #include "opt_hwpmc_hooks.h"
  41 #include "opt_no_adaptive_sx.h"
  42
  43 #include <sys/cdefs.h>
  44 __FBSDID("$FreeBSD$");
  45
  46 #include <sys/param.h>
  47 #include <sys/systm.h>
  48 #include <sys/kdb.h>
  49 #include <sys/kernel.h>
  50 #include <sys/ktr.h>
  51 #include <sys/lock.h>
  52 #include <sys/mutex.h>
  53 #include <sys/proc.h>
  54 #include <sys/sched.h>
  55 #include <sys/sleepqueue.h>
  56 #include <sys/sx.h>
  57 #include <sys/smp.h>
  58 #include <sys/sysctl.h>
  59
  60 #if defined(SMP) && !defined(NO_ADAPTIVE_SX)
  61 #include <machine/cpu.h>
  62 #endif
  63
  64 #ifdef DDB
  65 #include <ddb/ddb.h>
  66 #endif
  67
  68 #if defined(SMP) && !defined(NO_ADAPTIVE_SX)
  69 #define ADAPTIVE_SX
  70 #endif
  71
  72 CTASSERT((SX_NOADAPTIVE & LO_CLASSFLAGS) == SX_NOADAPTIVE);
  73
  74 #ifdef HWPMC_HOOKS
  75 #include <sys/pmckern.h>
  76 PMC_SOFT_DECLARE( , , lock, failed);
  77 #endif
  78
  79 /* Handy macros for sleep queues. */
  80 #define SQ_EXCLUSIVE_QUEUE      0
  81 #define SQ_SHARED_QUEUE         1
  82
  83 /*
  84  * Variations on DROP_GIANT()/PICKUP_GIANT() for use in this file.  We
  85  * drop Giant anytime we have to sleep or if we adaptively spin.
  86  */
  87 #define GIANT_DECLARE                                                   \
  88         int _giantcnt = 0;                                              \
  89         WITNESS_SAVE_DECL(Giant)                                        \
  90
  91 #define GIANT_SAVE() do {                                               \
  92         if (mtx_owned(&Giant)) {                                        \
  93                 WITNESS_SAVE(&Giant.lock_object, Giant);                \
  94                 while (mtx_owned(&Giant)) {                             \
  95                         _giantcnt++;                                    \
  96                         mtx_unlock(&Giant);                             \
  97                 }                                                       \
  98         }                                                               \
  99 } while (0)
 100
 101 #define GIANT_RESTORE() do {                                            \
 102         if (_giantcnt > 0) {                                            \
 103                 mtx_assert(&Giant, MA_NOTOWNED);                        \
 104                 while (_giantcnt--)                                     \
 105                         mtx_lock(&Giant);                               \
 106                 WITNESS_RESTORE(&Giant.lock_object, Giant);             \
 107         }                                                               \
 108 } while (0)
 109
 110 /*
 111  * Returns true if an exclusive lock is recursed.  It assumes
 112  * curthread currently has an exclusive lock.
 113  */
 114 #define sx_recursed(sx)         ((sx)->sx_recurse != 0)
 115
 116 static void     assert_sx(const struct lock_object *lock, int what);
 117 #ifdef DDB
 118 static void     db_show_sx(const struct lock_object *lock);
 119 #endif
 120 static void     lock_sx(struct lock_object *lock, uintptr_t how);
 121 #ifdef KDTRACE_HOOKS
 122 static int      owner_sx(const struct lock_object *lock, struct thread **owner);
 123 #endif
 124 static uintptr_t unlock_sx(struct lock_object *lock);
 125
 126 struct lock_class lock_class_sx = {
 127         .lc_name = "sx",
 128         .lc_flags = LC_SLEEPLOCK | LC_SLEEPABLE | LC_RECURSABLE | LC_UPGRADABLE,
 129         .lc_assert = assert_sx,
 130 #ifdef DDB
 131         .lc_ddb_show = db_show_sx,
 132 #endif
 133         .lc_lock = lock_sx,
 134         .lc_unlock = unlock_sx,
 135 #ifdef KDTRACE_HOOKS
 136         .lc_owner = owner_sx,
 137 #endif
 138 };
 139
 140 #ifndef INVARIANTS
 141 #define _sx_assert(sx, what, file, line)
 142 #endif
 143
 144 #ifdef ADAPTIVE_SX
 145 static u_int asx_retries = 10;
 146 static u_int asx_loops = 10000;
 147 static SYSCTL_NODE(_debug, OID_AUTO, sx, CTLFLAG_RD, NULL, "sxlock debugging");
 148 SYSCTL_UINT(_debug_sx, OID_AUTO, retries, CTLFLAG_RW, &asx_retries, 0, "");
 149 SYSCTL_UINT(_debug_sx, OID_AUTO, loops, CTLFLAG_RW, &asx_loops, 0, "");
 150
 151 static struct lock_delay_config __read_mostly sx_delay;
 152
 153 SYSCTL_INT(_debug_sx, OID_AUTO, delay_base, CTLFLAG_RW, &sx_delay.base,
 154     0, "");
 155 SYSCTL_INT(_debug_sx, OID_AUTO, delay_max, CTLFLAG_RW, &sx_delay.max,
 156     0, "");
 157
 158 LOCK_DELAY_SYSINIT_DEFAULT(sx_delay);
 159 #endif
 160
 161 void
 162 assert_sx(const struct lock_object *lock, int what)
 163 {
 164
 165         sx_assert((const struct sx *)lock, what);
 166 }
 167
 168 void
 169 lock_sx(struct lock_object *lock, uintptr_t how)
 170 {
 171         struct sx *sx;
 172
 173         sx = (struct sx *)lock;
 174         if (how)
 175                 sx_slock(sx);
 176         else
 177                 sx_xlock(sx);
 178 }
 179
 180 uintptr_t
 181 unlock_sx(struct lock_object *lock)
 182 {
 183         struct sx *sx;
 184
 185         sx = (struct sx *)lock;
 186         sx_assert(sx, SA_LOCKED | SA_NOTRECURSED);
 187         if (sx_xlocked(sx)) {
 188                 sx_xunlock(sx);
 189                 return (0);
 190         } else {
 191                 sx_sunlock(sx);
 192                 return (1);
 193         }
 194 }
 195
 196 #ifdef KDTRACE_HOOKS
 197 int
 198 owner_sx(const struct lock_object *lock, struct thread **owner)
 199 {
 200         const struct sx *sx = (const struct sx *)lock;
 201         uintptr_t x = sx->sx_lock;
 202
 203         *owner = (struct thread *)SX_OWNER(x);
 204         return ((x & SX_LOCK_SHARED) != 0 ? (SX_SHARERS(x) != 0) :
 205             (*owner != NULL));
 206 }
 207 #endif
 208
 209 void
 210 sx_sysinit(void *arg)
 211 {
 212         struct sx_args *sargs = arg;
 213
 214         sx_init_flags(sargs->sa_sx, sargs->sa_desc, sargs->sa_flags);
 215 }
 216
 217 void
 218 sx_init_flags(struct sx *sx, const char *description, int opts)
 219 {
 220         int flags;
 221
 222         MPASS((opts & ~(SX_QUIET | SX_RECURSE | SX_NOWITNESS | SX_DUPOK |
 223             SX_NOPROFILE | SX_NOADAPTIVE | SX_NEW)) == 0);
 224         ASSERT_ATOMIC_LOAD_PTR(sx->sx_lock,
 225             ("%s: sx_lock not aligned for %s: %p", __func__, description,
 226             &sx->sx_lock));
 227
 228         flags = LO_SLEEPABLE | LO_UPGRADABLE;
 229         if (opts & SX_DUPOK)
 230                 flags |= LO_DUPOK;
 231         if (opts & SX_NOPROFILE)
 232                 flags |= LO_NOPROFILE;
 233         if (!(opts & SX_NOWITNESS))
 234                 flags |= LO_WITNESS;
 235         if (opts & SX_RECURSE)
 236                 flags |= LO_RECURSABLE;
 237         if (opts & SX_QUIET)
 238                 flags |= LO_QUIET;
 239         if (opts & SX_NEW)
 240                 flags |= LO_NEW;
 241
 242         flags |= opts & SX_NOADAPTIVE;
 243         lock_init(&sx->lock_object, &lock_class_sx, description, NULL, flags);
 244         sx->sx_lock = SX_LOCK_UNLOCKED;
 245         sx->sx_recurse = 0;
 246 }
 247
 248 void
 249 sx_destroy(struct sx *sx)
 250 {
 251
 252         KASSERT(sx->sx_lock == SX_LOCK_UNLOCKED, ("sx lock still held"));
 253         KASSERT(sx->sx_recurse == 0, ("sx lock still recursed"));
 254         sx->sx_lock = SX_LOCK_DESTROYED;
 255         lock_destroy(&sx->lock_object);
 256 }
 257
 258 int
 259 sx_try_slock_(struct sx *sx, const char *file, int line)
 260 {
 261         uintptr_t x;
 262
 263         if (SCHEDULER_STOPPED())
 264                 return (1);
 265
 266         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
 267             ("sx_try_slock() by idle thread %p on sx %s @ %s:%d",
 268             curthread, sx->lock_object.lo_name, file, line));
 269
 270         x = sx->sx_lock;
 271         for (;;) {
 272                 KASSERT(x != SX_LOCK_DESTROYED,
 273                     ("sx_try_slock() of destroyed sx @ %s:%d", file, line));
 274                 if (!(x & SX_LOCK_SHARED))
 275                         break;
 276                 if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, x + SX_ONE_SHARER)) {
 277                         LOCK_LOG_TRY("SLOCK", &sx->lock_object, 0, 1, file, line);
 278                         WITNESS_LOCK(&sx->lock_object, LOP_TRYLOCK, file, line);
 279                         LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire,
 280                             sx, 0, 0, file, line, LOCKSTAT_READER);
 281                         TD_LOCKS_INC(curthread);
 282                         return (1);
 283                 }
 284         }
 285
 286         LOCK_LOG_TRY("SLOCK", &sx->lock_object, 0, 0, file, line);
 287         return (0);
 288 }
 289
 290 int
 291 _sx_xlock(struct sx *sx, int opts, const char *file, int line)
 292 {
 293         uintptr_t tid, x;
 294         int error = 0;
 295
 296         KASSERT(kdb_active != 0 || SCHEDULER_STOPPED() ||
 297             !TD_IS_IDLETHREAD(curthread),
 298             ("sx_xlock() by idle thread %p on sx %s @ %s:%d",
 299             curthread, sx->lock_object.lo_name, file, line));
 300         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 301             ("sx_xlock() of destroyed sx @ %s:%d", file, line));
 302         WITNESS_CHECKORDER(&sx->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE, file,
 303             line, NULL);
 304         tid = (uintptr_t)curthread;
 305         x = SX_LOCK_UNLOCKED;
 306         if (!atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid))
 307                 error = _sx_xlock_hard(sx, x, tid, opts, file, line);
 308         else
 309                 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
 310                     0, 0, file, line, LOCKSTAT_WRITER);
 311         if (!error) {
 312                 LOCK_LOG_LOCK("XLOCK", &sx->lock_object, 0, sx->sx_recurse,
 313                     file, line);
 314                 WITNESS_LOCK(&sx->lock_object, LOP_EXCLUSIVE, file, line);
 315                 TD_LOCKS_INC(curthread);
 316         }
 317
 318         return (error);
 319 }
 320
 321 int
 322 sx_try_xlock_(struct sx *sx, const char *file, int line)
 323 {
 324         struct thread *td;
 325         uintptr_t tid, x;
 326         int rval;
 327         bool recursed;
 328
 329         td = curthread;
 330         tid = (uintptr_t)td;
 331         if (SCHEDULER_STOPPED_TD(td))
 332                 return (1);
 333
 334         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(td),
 335             ("sx_try_xlock() by idle thread %p on sx %s @ %s:%d",
 336             curthread, sx->lock_object.lo_name, file, line));
 337         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 338             ("sx_try_xlock() of destroyed sx @ %s:%d", file, line));
 339
 340         rval = 1;
 341         recursed = false;
 342         x = SX_LOCK_UNLOCKED;
 343         for (;;) {
 344                 if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid))
 345                         break;
 346                 if (x == SX_LOCK_UNLOCKED)
 347                         continue;
 348                 if (x == tid && (sx->lock_object.lo_flags & LO_RECURSABLE)) {
 349                         sx->sx_recurse++;
 350                         atomic_set_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
 351                         break;
 352                 }
 353                 rval = 0;
 354                 break;
 355         }
 356
 357         LOCK_LOG_TRY("XLOCK", &sx->lock_object, 0, rval, file, line);
 358         if (rval) {
 359                 WITNESS_LOCK(&sx->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
 360                     file, line);
 361                 if (!recursed)
 362                         LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire,
 363                             sx, 0, 0, file, line, LOCKSTAT_WRITER);
 364                 TD_LOCKS_INC(curthread);
 365         }
 366
 367         return (rval);
 368 }
 369
 370 void
 371 _sx_xunlock(struct sx *sx, const char *file, int line)
 372 {
 373
 374         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 375             ("sx_xunlock() of destroyed sx @ %s:%d", file, line));
 376         _sx_assert(sx, SA_XLOCKED, file, line);
 377         WITNESS_UNLOCK(&sx->lock_object, LOP_EXCLUSIVE, file, line);
 378         LOCK_LOG_LOCK("XUNLOCK", &sx->lock_object, 0, sx->sx_recurse, file,
 379             line);
 380 #if LOCK_DEBUG > 0
 381         _sx_xunlock_hard(sx, (uintptr_t)curthread, file, line);
 382 #else
 383         __sx_xunlock(sx, curthread, file, line);
 384 #endif
 385         TD_LOCKS_DEC(curthread);
 386 }
 387
 388 /*
 389  * Try to do a non-blocking upgrade from a shared lock to an exclusive lock.
 390  * This will only succeed if this thread holds a single shared lock.
 391  * Return 1 if if the upgrade succeed, 0 otherwise.
 392  */
 393 int
 394 sx_try_upgrade_(struct sx *sx, const char *file, int line)
 395 {
 396         uintptr_t x;
 397         int success;
 398
 399         if (SCHEDULER_STOPPED())
 400                 return (1);
 401
 402         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 403             ("sx_try_upgrade() of destroyed sx @ %s:%d", file, line));
 404         _sx_assert(sx, SA_SLOCKED, file, line);
 405
 406         /*
 407          * Try to switch from one shared lock to an exclusive lock.  We need
 408          * to maintain the SX_LOCK_EXCLUSIVE_WAITERS flag if set so that
 409          * we will wake up the exclusive waiters when we drop the lock.
 410          */
 411         x = sx->sx_lock & SX_LOCK_EXCLUSIVE_WAITERS;
 412         success = atomic_cmpset_acq_ptr(&sx->sx_lock, SX_SHARERS_LOCK(1) | x,
 413             (uintptr_t)curthread | x);
 414         LOCK_LOG_TRY("XUPGRADE", &sx->lock_object, 0, success, file, line);
 415         if (success) {
 416                 WITNESS_UPGRADE(&sx->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
 417                     file, line);
 418                 LOCKSTAT_RECORD0(sx__upgrade, sx);
 419         }
 420         return (success);
 421 }
 422
 423 /*
 424  * Downgrade an unrecursed exclusive lock into a single shared lock.
 425  */
 426 void
 427 sx_downgrade_(struct sx *sx, const char *file, int line)
 428 {
 429         uintptr_t x;
 430         int wakeup_swapper;
 431
 432         if (SCHEDULER_STOPPED())
 433                 return;
 434
 435         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 436             ("sx_downgrade() of destroyed sx @ %s:%d", file, line));
 437         _sx_assert(sx, SA_XLOCKED | SA_NOTRECURSED, file, line);
 438 #ifndef INVARIANTS
 439         if (sx_recursed(sx))
 440                 panic("downgrade of a recursed lock");
 441 #endif
 442
 443         WITNESS_DOWNGRADE(&sx->lock_object, 0, file, line);
 444
 445         /*
 446          * Try to switch from an exclusive lock with no shared waiters
 447          * to one sharer with no shared waiters.  If there are
 448          * exclusive waiters, we don't need to lock the sleep queue so
 449          * long as we preserve the flag.  We do one quick try and if
 450          * that fails we grab the sleepq lock to keep the flags from
 451          * changing and do it the slow way.
 452          *
 453          * We have to lock the sleep queue if there are shared waiters
 454          * so we can wake them up.
 455          */
 456         x = sx->sx_lock;
 457         if (!(x & SX_LOCK_SHARED_WAITERS) &&
 458             atomic_cmpset_rel_ptr(&sx->sx_lock, x, SX_SHARERS_LOCK(1) |
 459             (x & SX_LOCK_EXCLUSIVE_WAITERS))) {
 460                 LOCK_LOG_LOCK("XDOWNGRADE", &sx->lock_object, 0, 0, file, line);
 461                 return;
 462         }
 463
 464         /*
 465          * Lock the sleep queue so we can read the waiters bits
 466          * without any races and wakeup any shared waiters.
 467          */
 468         sleepq_lock(&sx->lock_object);
 469
 470         /*
 471          * Preserve SX_LOCK_EXCLUSIVE_WAITERS while downgraded to a single
 472          * shared lock.  If there are any shared waiters, wake them up.
 473          */
 474         wakeup_swapper = 0;
 475         x = sx->sx_lock;
 476         atomic_store_rel_ptr(&sx->sx_lock, SX_SHARERS_LOCK(1) |
 477             (x & SX_LOCK_EXCLUSIVE_WAITERS));
 478         if (x & SX_LOCK_SHARED_WAITERS)
 479                 wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX,
 480                     0, SQ_SHARED_QUEUE);
 481         sleepq_release(&sx->lock_object);
 482
 483         LOCK_LOG_LOCK("XDOWNGRADE", &sx->lock_object, 0, 0, file, line);
 484         LOCKSTAT_RECORD0(sx__downgrade, sx);
 485
 486         if (wakeup_swapper)
 487                 kick_proc0();
 488 }
 489
 490 /*
 491  * This function represents the so-called 'hard case' for sx_xlock
 492  * operation.  All 'easy case' failures are redirected to this.  Note
 493  * that ideally this would be a static function, but it needs to be
 494  * accessible from at least sx.h.
 495  */
 496 int
 497 _sx_xlock_hard(struct sx *sx, uintptr_t x, uintptr_t tid, int opts,
 498     const char *file, int line)
 499 {
 500         GIANT_DECLARE;
 501 #ifdef ADAPTIVE_SX
 502         volatile struct thread *owner;
 503         u_int i, spintries = 0;
 504 #endif
 505 #ifdef LOCK_PROFILING
 506         uint64_t waittime = 0;
 507         int contested = 0;
 508 #endif
 509         int error = 0;
 510 #if defined(ADAPTIVE_SX) || defined(KDTRACE_HOOKS)
 511         struct lock_delay_arg lda;
 512 #endif
 513 #ifdef  KDTRACE_HOOKS
 514         uintptr_t state;
 515         u_int sleep_cnt = 0;
 516         int64_t sleep_time = 0;
 517         int64_t all_time = 0;
 518 #endif
 519
 520         if (SCHEDULER_STOPPED())
 521                 return (0);
 522
 523 #if defined(ADAPTIVE_SX)
 524         lock_delay_arg_init(&lda, &sx_delay);
 525 #elif defined(KDTRACE_HOOKS)
 526         lock_delay_arg_init(&lda, NULL);
 527 #endif
 528
 529         if (__predict_false(x == SX_LOCK_UNLOCKED))
 530                 x = SX_READ_VALUE(sx);
 531
 532         /* If we already hold an exclusive lock, then recurse. */
 533         if (__predict_false(lv_sx_owner(x) == (struct thread *)tid)) {
 534                 KASSERT((sx->lock_object.lo_flags & LO_RECURSABLE) != 0,
 535             ("_sx_xlock_hard: recursed on non-recursive sx %s @ %s:%d\n",
 536                     sx->lock_object.lo_name, file, line));
 537                 sx->sx_recurse++;
 538                 atomic_set_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
 539                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 540                         CTR2(KTR_LOCK, "%s: %p recursing", __func__, sx);
 541                 return (0);
 542         }
 543
 544         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 545                 CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__,
 546                     sx->lock_object.lo_name, (void *)sx->sx_lock, file, line);
 547
 548 #ifdef KDTRACE_HOOKS
 549         all_time -= lockstat_nsecs(&sx->lock_object);
 550         state = x;
 551 #endif
 552         for (;;) {
 553                 if (x == SX_LOCK_UNLOCKED) {
 554                         if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid))
 555                                 break;
 556                         continue;
 557                 }
 558 #ifdef KDTRACE_HOOKS
 559                 lda.spin_cnt++;
 560 #endif
 561 #ifdef HWPMC_HOOKS
 562                 PMC_SOFT_CALL( , , lock, failed);
 563 #endif
 564                 lock_profile_obtain_lock_failed(&sx->lock_object, &contested,
 565                     &waittime);
 566 #ifdef ADAPTIVE_SX
 567                 /*
 568                  * If the lock is write locked and the owner is
 569                  * running on another CPU, spin until the owner stops
 570                  * running or the state of the lock changes.
 571                  */
 572                 if ((sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) {
 573                         if ((x & SX_LOCK_SHARED) == 0) {
 574                                 owner = lv_sx_owner(x);
 575                                 if (TD_IS_RUNNING(owner)) {
 576                                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 577                                                 CTR3(KTR_LOCK,
 578                                             "%s: spinning on %p held by %p",
 579                                                     __func__, sx, owner);
 580                                         KTR_STATE1(KTR_SCHED, "thread",
 581                                             sched_tdname(curthread), "spinning",
 582                                             "lockname:\"%s\"",
 583                                             sx->lock_object.lo_name);
 584                                         GIANT_SAVE();
 585                                         do {
 586                                                 lock_delay(&lda);
 587                                                 x = SX_READ_VALUE(sx);
 588                                                 owner = lv_sx_owner(x);
 589                                         } while (owner != NULL &&
 590                                                     TD_IS_RUNNING(owner));
 591                                         KTR_STATE0(KTR_SCHED, "thread",
 592                                             sched_tdname(curthread), "running");
 593                                         continue;
 594                                 }
 595                         } else if (SX_SHARERS(x) && spintries < asx_retries) {
 596                                 KTR_STATE1(KTR_SCHED, "thread",
 597                                     sched_tdname(curthread), "spinning",
 598                                     "lockname:\"%s\"", sx->lock_object.lo_name);
 599                                 GIANT_SAVE();
 600                                 spintries++;
 601                                 for (i = 0; i < asx_loops; i++) {
 602                                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 603                                                 CTR4(KTR_LOCK,
 604                                     "%s: shared spinning on %p with %u and %u",
 605                                                     __func__, sx, spintries, i);
 606                                         x = sx->sx_lock;
 607                                         if ((x & SX_LOCK_SHARED) == 0 ||
 608                                             SX_SHARERS(x) == 0)
 609                                                 break;
 610                                         cpu_spinwait();
 611 #ifdef KDTRACE_HOOKS
 612                                         lda.spin_cnt++;
 613 #endif
 614                                 }
 615                                 KTR_STATE0(KTR_SCHED, "thread",
 616                                     sched_tdname(curthread), "running");
 617                                 x = SX_READ_VALUE(sx);
 618                                 if (i != asx_loops)
 619                                         continue;
 620                         }
 621                 }
 622 #endif
 623
 624                 sleepq_lock(&sx->lock_object);
 625                 x = SX_READ_VALUE(sx);
 626
 627                 /*
 628                  * If the lock was released while spinning on the
 629                  * sleep queue chain lock, try again.
 630                  */
 631                 if (x == SX_LOCK_UNLOCKED) {
 632                         sleepq_release(&sx->lock_object);
 633                         continue;
 634                 }
 635
 636 #ifdef ADAPTIVE_SX
 637                 /*
 638                  * The current lock owner might have started executing
 639                  * on another CPU (or the lock could have changed
 640                  * owners) while we were waiting on the sleep queue
 641                  * chain lock.  If so, drop the sleep queue lock and try
 642                  * again.
 643                  */
 644                 if (!(x & SX_LOCK_SHARED) &&
 645                     (sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) {
 646                         owner = (struct thread *)SX_OWNER(x);
 647                         if (TD_IS_RUNNING(owner)) {
 648                                 sleepq_release(&sx->lock_object);
 649                                 continue;
 650                         }
 651                 }
 652 #endif
 653
 654                 /*
 655                  * If an exclusive lock was released with both shared
 656                  * and exclusive waiters and a shared waiter hasn't
 657                  * woken up and acquired the lock yet, sx_lock will be
 658                  * set to SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS.
 659                  * If we see that value, try to acquire it once.  Note
 660                  * that we have to preserve SX_LOCK_EXCLUSIVE_WAITERS
 661                  * as there are other exclusive waiters still.  If we
 662                  * fail, restart the loop.
 663                  */
 664                 if (x == (SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS)) {
 665                         if (atomic_cmpset_acq_ptr(&sx->sx_lock,
 666                             SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS,
 667                             tid | SX_LOCK_EXCLUSIVE_WAITERS)) {
 668                                 sleepq_release(&sx->lock_object);
 669                                 CTR2(KTR_LOCK, "%s: %p claimed by new writer",
 670                                     __func__, sx);
 671                                 break;
 672                         }
 673                         sleepq_release(&sx->lock_object);
 674                         x = SX_READ_VALUE(sx);
 675                         continue;
 676                 }
 677
 678                 /*
 679                  * Try to set the SX_LOCK_EXCLUSIVE_WAITERS.  If we fail,
 680                  * than loop back and retry.
 681                  */
 682                 if (!(x & SX_LOCK_EXCLUSIVE_WAITERS)) {
 683                         if (!atomic_cmpset_ptr(&sx->sx_lock, x,
 684                             x | SX_LOCK_EXCLUSIVE_WAITERS)) {
 685                                 sleepq_release(&sx->lock_object);
 686                                 x = SX_READ_VALUE(sx);
 687                                 continue;
 688                         }
 689                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 690                                 CTR2(KTR_LOCK, "%s: %p set excl waiters flag",
 691                                     __func__, sx);
 692                 }
 693
 694                 /*
 695                  * Since we have been unable to acquire the exclusive
 696                  * lock and the exclusive waiters flag is set, we have
 697                  * to sleep.
 698                  */
 699                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 700                         CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
 701                             __func__, sx);
 702
 703 #ifdef KDTRACE_HOOKS
 704                 sleep_time -= lockstat_nsecs(&sx->lock_object);
 705 #endif
 706                 GIANT_SAVE();
 707                 sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
 708                     SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
 709                     SLEEPQ_INTERRUPTIBLE : 0), SQ_EXCLUSIVE_QUEUE);
 710                 if (!(opts & SX_INTERRUPTIBLE))
 711                         sleepq_wait(&sx->lock_object, 0);
 712                 else
 713                         error = sleepq_wait_sig(&sx->lock_object, 0);
 714 #ifdef KDTRACE_HOOKS
 715                 sleep_time += lockstat_nsecs(&sx->lock_object);
 716                 sleep_cnt++;
 717 #endif
 718                 if (error) {
 719                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 720                                 CTR2(KTR_LOCK,
 721                         "%s: interruptible sleep by %p suspended by signal",
 722                                     __func__, sx);
 723                         break;
 724                 }
 725                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 726                         CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
 727                             __func__, sx);
 728                 x = SX_READ_VALUE(sx);
 729         }
 730 #ifdef KDTRACE_HOOKS
 731         all_time += lockstat_nsecs(&sx->lock_object);
 732         if (sleep_time)
 733                 LOCKSTAT_RECORD4(sx__block, sx, sleep_time,
 734                     LOCKSTAT_WRITER, (state & SX_LOCK_SHARED) == 0,
 735                     (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
 736         if (lda.spin_cnt > sleep_cnt)
 737                 LOCKSTAT_RECORD4(sx__spin, sx, all_time - sleep_time,
 738                     LOCKSTAT_WRITER, (state & SX_LOCK_SHARED) == 0,
 739                     (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
 740 #endif
 741         if (!error)
 742                 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
 743                     contested, waittime, file, line, LOCKSTAT_WRITER);
 744         GIANT_RESTORE();
 745         return (error);
 746 }
 747
 748 /*
 749  * This function represents the so-called 'hard case' for sx_xunlock
 750  * operation.  All 'easy case' failures are redirected to this.  Note
 751  * that ideally this would be a static function, but it needs to be
 752  * accessible from at least sx.h.
 753  */
 754 void
 755 _sx_xunlock_hard(struct sx *sx, uintptr_t tid, const char *file, int line)
 756 {
 757         uintptr_t x;
 758         int queue, wakeup_swapper;
 759
 760         if (SCHEDULER_STOPPED())
 761                 return;
 762
 763         MPASS(!(sx->sx_lock & SX_LOCK_SHARED));
 764
 765         x = SX_READ_VALUE(sx);
 766         if (x & SX_LOCK_RECURSED) {
 767                 /* The lock is recursed, unrecurse one level. */
 768                 if ((--sx->sx_recurse) == 0)
 769                         atomic_clear_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
 770                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 771                         CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, sx);
 772                 return;
 773         }
 774
 775         LOCKSTAT_PROFILE_RELEASE_RWLOCK(sx__release, sx, LOCKSTAT_WRITER);
 776         if (x == tid &&
 777             atomic_cmpset_rel_ptr(&sx->sx_lock, tid, SX_LOCK_UNLOCKED))
 778                 return;
 779
 780         MPASS(sx->sx_lock & (SX_LOCK_SHARED_WAITERS |
 781             SX_LOCK_EXCLUSIVE_WAITERS));
 782         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 783                 CTR2(KTR_LOCK, "%s: %p contested", __func__, sx);
 784
 785         sleepq_lock(&sx->lock_object);
 786         x = SX_LOCK_UNLOCKED;
 787
 788         /*
 789          * The wake up algorithm here is quite simple and probably not
 790          * ideal.  It gives precedence to shared waiters if they are
 791          * present.  For this condition, we have to preserve the
 792          * state of the exclusive waiters flag.
 793          * If interruptible sleeps left the shared queue empty avoid a
 794          * starvation for the threads sleeping on the exclusive queue by giving
 795          * them precedence and cleaning up the shared waiters bit anyway.
 796          */
 797         if ((sx->sx_lock & SX_LOCK_SHARED_WAITERS) != 0 &&
 798             sleepq_sleepcnt(&sx->lock_object, SQ_SHARED_QUEUE) != 0) {
 799                 queue = SQ_SHARED_QUEUE;
 800                 x |= (sx->sx_lock & SX_LOCK_EXCLUSIVE_WAITERS);
 801         } else
 802                 queue = SQ_EXCLUSIVE_QUEUE;
 803
 804         /* Wake up all the waiters for the specific queue. */
 805         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 806                 CTR3(KTR_LOCK, "%s: %p waking up all threads on %s queue",
 807                     __func__, sx, queue == SQ_SHARED_QUEUE ? "shared" :
 808                     "exclusive");
 809         atomic_store_rel_ptr(&sx->sx_lock, x);
 810         wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX, 0,
 811             queue);
 812         sleepq_release(&sx->lock_object);
 813         if (wakeup_swapper)
 814                 kick_proc0();
 815 }
 816
 817 static bool __always_inline
 818 __sx_slock_try(struct sx *sx, uintptr_t *xp, const char *file, int line)
 819 {
 820
 821         /*
 822          * If no other thread has an exclusive lock then try to bump up
 823          * the count of sharers.  Since we have to preserve the state
 824          * of SX_LOCK_EXCLUSIVE_WAITERS, if we fail to acquire the
 825          * shared lock loop back and retry.
 826          */
 827         while (*xp & SX_LOCK_SHARED) {
 828                 MPASS(!(*xp & SX_LOCK_SHARED_WAITERS));
 829                 if (atomic_fcmpset_acq_ptr(&sx->sx_lock, xp,
 830                     *xp + SX_ONE_SHARER)) {
 831                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 832                                 CTR4(KTR_LOCK, "%s: %p succeed %p -> %p",
 833                                     __func__, sx, (void *)*xp,
 834                                     (void *)(*xp + SX_ONE_SHARER));
 835                         return (true);
 836                 }
 837         }
 838         return (false);
 839 }
 840
 841 static int __noinline
 842 _sx_slock_hard(struct sx *sx, int opts, const char *file, int line, uintptr_t x)
 843 {
 844         GIANT_DECLARE;
 845 #ifdef ADAPTIVE_SX
 846         volatile struct thread *owner;
 847 #endif
 848 #ifdef LOCK_PROFILING
 849         uint64_t waittime = 0;
 850         int contested = 0;
 851 #endif
 852         int error = 0;
 853 #if defined(ADAPTIVE_SX) || defined(KDTRACE_HOOKS)
 854         struct lock_delay_arg lda;
 855 #endif
 856 #ifdef KDTRACE_HOOKS
 857         uintptr_t state;
 858         u_int sleep_cnt = 0;
 859         int64_t sleep_time = 0;
 860         int64_t all_time = 0;
 861 #endif
 862
 863         if (SCHEDULER_STOPPED())
 864                 return (0);
 865
 866 #if defined(ADAPTIVE_SX)
 867         lock_delay_arg_init(&lda, &sx_delay);
 868 #elif defined(KDTRACE_HOOKS)
 869         lock_delay_arg_init(&lda, NULL);
 870 #endif
 871 #ifdef KDTRACE_HOOKS
 872         all_time -= lockstat_nsecs(&sx->lock_object);
 873         state = x;
 874 #endif
 875
 876         /*
 877          * As with rwlocks, we don't make any attempt to try to block
 878          * shared locks once there is an exclusive waiter.
 879          */
 880         for (;;) {
 881                 if (__sx_slock_try(sx, &x, file, line))
 882                         break;
 883 #ifdef KDTRACE_HOOKS
 884                 lda.spin_cnt++;
 885 #endif
 886
 887 #ifdef HWPMC_HOOKS
 888                 PMC_SOFT_CALL( , , lock, failed);
 889 #endif
 890                 lock_profile_obtain_lock_failed(&sx->lock_object, &contested,
 891                     &waittime);
 892
 893 #ifdef ADAPTIVE_SX
 894                 /*
 895                  * If the owner is running on another CPU, spin until
 896                  * the owner stops running or the state of the lock
 897                  * changes.
 898                  */
 899                 if ((sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) {
 900                         owner = lv_sx_owner(x);
 901                         if (TD_IS_RUNNING(owner)) {
 902                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 903                                         CTR3(KTR_LOCK,
 904                                             "%s: spinning on %p held by %p",
 905                                             __func__, sx, owner);
 906                                 KTR_STATE1(KTR_SCHED, "thread",
 907                                     sched_tdname(curthread), "spinning",
 908                                     "lockname:\"%s\"", sx->lock_object.lo_name);
 909                                 GIANT_SAVE();
 910                                 do {
 911                                         lock_delay(&lda);
 912                                         x = SX_READ_VALUE(sx);
 913                                         owner = lv_sx_owner(x);
 914                                 } while (owner != NULL && TD_IS_RUNNING(owner));
 915                                 KTR_STATE0(KTR_SCHED, "thread",
 916                                     sched_tdname(curthread), "running");
 917                                 continue;
 918                         }
 919                 }
 920 #endif
 921
 922                 /*
 923                  * Some other thread already has an exclusive lock, so
 924                  * start the process of blocking.
 925                  */
 926                 sleepq_lock(&sx->lock_object);
 927                 x = SX_READ_VALUE(sx);
 928
 929                 /*
 930                  * The lock could have been released while we spun.
 931                  * In this case loop back and retry.
 932                  */
 933                 if (x & SX_LOCK_SHARED) {
 934                         sleepq_release(&sx->lock_object);
 935                         continue;
 936                 }
 937
 938 #ifdef ADAPTIVE_SX
 939                 /*
 940                  * If the owner is running on another CPU, spin until
 941                  * the owner stops running or the state of the lock
 942                  * changes.
 943                  */
 944                 if (!(x & SX_LOCK_SHARED) &&
 945                     (sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) {
 946                         owner = (struct thread *)SX_OWNER(x);
 947                         if (TD_IS_RUNNING(owner)) {
 948                                 sleepq_release(&sx->lock_object);
 949                                 x = SX_READ_VALUE(sx);
 950                                 continue;
 951                         }
 952                 }
 953 #endif
 954
 955                 /*
 956                  * Try to set the SX_LOCK_SHARED_WAITERS flag.  If we
 957                  * fail to set it drop the sleep queue lock and loop
 958                  * back.
 959                  */
 960                 if (!(x & SX_LOCK_SHARED_WAITERS)) {
 961                         if (!atomic_cmpset_ptr(&sx->sx_lock, x,
 962                             x | SX_LOCK_SHARED_WAITERS)) {
 963                                 sleepq_release(&sx->lock_object);
 964                                 x = SX_READ_VALUE(sx);
 965                                 continue;
 966                         }
 967                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 968                                 CTR2(KTR_LOCK, "%s: %p set shared waiters flag",
 969                                     __func__, sx);
 970                 }
 971
 972                 /*
 973                  * Since we have been unable to acquire the shared lock,
 974                  * we have to sleep.
 975                  */
 976                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 977                         CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
 978                             __func__, sx);
 979
 980 #ifdef KDTRACE_HOOKS
 981                 sleep_time -= lockstat_nsecs(&sx->lock_object);
 982 #endif
 983                 GIANT_SAVE();
 984                 sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
 985                     SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
 986                     SLEEPQ_INTERRUPTIBLE : 0), SQ_SHARED_QUEUE);
 987                 if (!(opts & SX_INTERRUPTIBLE))
 988                         sleepq_wait(&sx->lock_object, 0);
 989                 else
 990                         error = sleepq_wait_sig(&sx->lock_object, 0);
 991 #ifdef KDTRACE_HOOKS
 992                 sleep_time += lockstat_nsecs(&sx->lock_object);
 993                 sleep_cnt++;
 994 #endif
 995                 if (error) {
 996                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 997                                 CTR2(KTR_LOCK,
 998                         "%s: interruptible sleep by %p suspended by signal",
 999                                     __func__, sx);
1000                         break;
1001                 }
1002                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1003                         CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
1004                             __func__, sx);
1005                 x = SX_READ_VALUE(sx);
1006         }
1007 #ifdef KDTRACE_HOOKS
1008         all_time += lockstat_nsecs(&sx->lock_object);
1009         if (sleep_time)
1010                 LOCKSTAT_RECORD4(sx__block, sx, sleep_time,
1011                     LOCKSTAT_READER, (state & SX_LOCK_SHARED) == 0,
1012                     (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
1013         if (lda.spin_cnt > sleep_cnt)
1014                 LOCKSTAT_RECORD4(sx__spin, sx, all_time - sleep_time,
1015                     LOCKSTAT_READER, (state & SX_LOCK_SHARED) == 0,
1016                     (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
1017 #endif
1018         if (error == 0) {
1019                 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
1020                     contested, waittime, file, line, LOCKSTAT_READER);
1021         }
1022         GIANT_RESTORE();
1023         return (error);
1024 }
1025
1026 int
1027 _sx_slock(struct sx *sx, int opts, const char *file, int line)
1028 {
1029         uintptr_t x;
1030         int error;
1031
1032         KASSERT(kdb_active != 0 || SCHEDULER_STOPPED() ||
1033             !TD_IS_IDLETHREAD(curthread),
1034             ("sx_slock() by idle thread %p on sx %s @ %s:%d",
1035             curthread, sx->lock_object.lo_name, file, line));
1036         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
1037             ("sx_slock() of destroyed sx @ %s:%d", file, line));
1038         WITNESS_CHECKORDER(&sx->lock_object, LOP_NEWORDER, file, line, NULL);
1039
1040         error = 0;
1041         x = SX_READ_VALUE(sx);
1042         if (__predict_false(LOCKSTAT_OOL_PROFILE_ENABLED(sx__acquire) ||
1043             !__sx_slock_try(sx, &x, file, line)))
1044                 error = _sx_slock_hard(sx, opts, file, line, x);
1045         if (error == 0) {
1046                 LOCK_LOG_LOCK("SLOCK", &sx->lock_object, 0, 0, file, line);
1047                 WITNESS_LOCK(&sx->lock_object, 0, file, line);
1048                 TD_LOCKS_INC(curthread);
1049         }
1050         return (error);
1051 }
1052
1053 static bool __always_inline
1054 _sx_sunlock_try(struct sx *sx, uintptr_t *xp)
1055 {
1056
1057         for (;;) {
1058                 /*
1059                  * We should never have sharers while at least one thread
1060                  * holds a shared lock.
1061                  */
1062                 KASSERT(!(*xp & SX_LOCK_SHARED_WAITERS),
1063                     ("%s: waiting sharers", __func__));
1064
1065                 /*
1066                  * See if there is more than one shared lock held.  If
1067                  * so, just drop one and return.
1068                  */
1069                 if (SX_SHARERS(*xp) > 1) {
1070                         if (atomic_fcmpset_rel_ptr(&sx->sx_lock, xp,
1071                             *xp - SX_ONE_SHARER)) {
1072                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1073                                         CTR4(KTR_LOCK,
1074                                             "%s: %p succeeded %p -> %p",
1075                                             __func__, sx, (void *)*xp,
1076                                             (void *)(*xp - SX_ONE_SHARER));
1077                                 return (true);
1078                         }
1079                         continue;
1080                 }
1081
1082                 /*
1083                  * If there aren't any waiters for an exclusive lock,
1084                  * then try to drop it quickly.
1085                  */
1086                 if (!(*xp & SX_LOCK_EXCLUSIVE_WAITERS)) {
1087                         MPASS(*xp == SX_SHARERS_LOCK(1));
1088                         *xp = SX_SHARERS_LOCK(1);
1089                         if (atomic_fcmpset_rel_ptr(&sx->sx_lock,
1090                             xp, SX_LOCK_UNLOCKED)) {
1091                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1092                                         CTR2(KTR_LOCK, "%s: %p last succeeded",
1093                                             __func__, sx);
1094                                 return (true);
1095                         }
1096                         continue;
1097                 }
1098                 break;
1099         }
1100         return (false);
1101 }
1102
1103 static void __noinline
1104 _sx_sunlock_hard(struct sx *sx, uintptr_t x, const char *file, int line)
1105 {
1106         int wakeup_swapper;
1107
1108         if (SCHEDULER_STOPPED())
1109                 return;
1110
1111         LOCKSTAT_PROFILE_RELEASE_RWLOCK(sx__release, sx, LOCKSTAT_READER);
1112
1113         for (;;) {
1114                 if (_sx_sunlock_try(sx, &x))
1115                         break;
1116
1117                 /*
1118                  * At this point, there should just be one sharer with
1119                  * exclusive waiters.
1120                  */
1121                 MPASS(x == (SX_SHARERS_LOCK(1) | SX_LOCK_EXCLUSIVE_WAITERS));
1122
1123                 sleepq_lock(&sx->lock_object);
1124
1125                 /*
1126                  * Wake up semantic here is quite simple:
1127                  * Just wake up all the exclusive waiters.
1128                  * Note that the state of the lock could have changed,
1129                  * so if it fails loop back and retry.
1130                  */
1131                 if (!atomic_cmpset_rel_ptr(&sx->sx_lock,
1132                     SX_SHARERS_LOCK(1) | SX_LOCK_EXCLUSIVE_WAITERS,
1133                     SX_LOCK_UNLOCKED)) {
1134                         sleepq_release(&sx->lock_object);
1135                         x = SX_READ_VALUE(sx);
1136                         continue;
1137                 }
1138                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1139                         CTR2(KTR_LOCK, "%s: %p waking up all thread on"
1140                             "exclusive queue", __func__, sx);
1141                 wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX,
1142                     0, SQ_EXCLUSIVE_QUEUE);
1143                 sleepq_release(&sx->lock_object);
1144                 if (wakeup_swapper)
1145                         kick_proc0();
1146                 break;
1147         }
1148 }
1149
1150 void
1151 _sx_sunlock(struct sx *sx, const char *file, int line)
1152 {
1153         uintptr_t x;
1154
1155         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
1156             ("sx_sunlock() of destroyed sx @ %s:%d", file, line));
1157         _sx_assert(sx, SA_SLOCKED, file, line);
1158         WITNESS_UNLOCK(&sx->lock_object, 0, file, line);
1159         LOCK_LOG_LOCK("SUNLOCK", &sx->lock_object, 0, 0, file, line);
1160
1161         x = SX_READ_VALUE(sx);
1162         if (__predict_false(LOCKSTAT_OOL_PROFILE_ENABLED(sx__release) ||
1163             !_sx_sunlock_try(sx, &x)))
1164                 _sx_sunlock_hard(sx, x, file, line);
1165
1166         TD_LOCKS_DEC(curthread);
1167 }
1168
1169 #ifdef INVARIANT_SUPPORT
1170 #ifndef INVARIANTS
1171 #undef  _sx_assert
1172 #endif
1173
1174 /*
1175  * In the non-WITNESS case, sx_assert() can only detect that at least
1176  * *some* thread owns an slock, but it cannot guarantee that *this*
1177  * thread owns an slock.
1178  */
1179 void
1180 _sx_assert(const struct sx *sx, int what, const char *file, int line)
1181 {
1182 #ifndef WITNESS
1183         int slocked = 0;
1184 #endif
1185
1186         if (panicstr != NULL)
1187                 return;
1188         switch (what) {
1189         case SA_SLOCKED:
1190         case SA_SLOCKED | SA_NOTRECURSED:
1191         case SA_SLOCKED | SA_RECURSED:
1192 #ifndef WITNESS
1193                 slocked = 1;
1194                 /* FALLTHROUGH */
1195 #endif
1196         case SA_LOCKED:
1197         case SA_LOCKED | SA_NOTRECURSED:
1198         case SA_LOCKED | SA_RECURSED:
1199 #ifdef WITNESS
1200                 witness_assert(&sx->lock_object, what, file, line);
1201 #else
1202                 /*
1203                  * If some other thread has an exclusive lock or we
1204                  * have one and are asserting a shared lock, fail.
1205                  * Also, if no one has a lock at all, fail.
1206                  */
1207                 if (sx->sx_lock == SX_LOCK_UNLOCKED ||
1208                     (!(sx->sx_lock & SX_LOCK_SHARED) && (slocked ||
1209                     sx_xholder(sx) != curthread)))
1210                         panic("Lock %s not %slocked @ %s:%d\n",
1211                             sx->lock_object.lo_name, slocked ? "share " : "",
1212                             file, line);
1213
1214                 if (!(sx->sx_lock & SX_LOCK_SHARED)) {
1215                         if (sx_recursed(sx)) {
1216                                 if (what & SA_NOTRECURSED)
1217                                         panic("Lock %s recursed @ %s:%d\n",
1218                                             sx->lock_object.lo_name, file,
1219                                             line);
1220                         } else if (what & SA_RECURSED)
1221                                 panic("Lock %s not recursed @ %s:%d\n",
1222                                     sx->lock_object.lo_name, file, line);
1223                 }
1224 #endif
1225                 break;
1226         case SA_XLOCKED:
1227         case SA_XLOCKED | SA_NOTRECURSED:
1228         case SA_XLOCKED | SA_RECURSED:
1229                 if (sx_xholder(sx) != curthread)
1230                         panic("Lock %s not exclusively locked @ %s:%d\n",
1231                             sx->lock_object.lo_name, file, line);
1232                 if (sx_recursed(sx)) {
1233                         if (what & SA_NOTRECURSED)
1234                                 panic("Lock %s recursed @ %s:%d\n",
1235                                     sx->lock_object.lo_name, file, line);
1236                 } else if (what & SA_RECURSED)
1237                         panic("Lock %s not recursed @ %s:%d\n",
1238                             sx->lock_object.lo_name, file, line);
1239                 break;
1240         case SA_UNLOCKED:
1241 #ifdef WITNESS
1242                 witness_assert(&sx->lock_object, what, file, line);
1243 #else
1244                 /*
1245                  * If we hold an exclusve lock fail.  We can't
1246                  * reliably check to see if we hold a shared lock or
1247                  * not.
1248                  */
1249                 if (sx_xholder(sx) == curthread)
1250                         panic("Lock %s exclusively locked @ %s:%d\n",
1251                             sx->lock_object.lo_name, file, line);
1252 #endif
1253                 break;
1254         default:
1255                 panic("Unknown sx lock assertion: %d @ %s:%d", what, file,
1256                     line);
1257         }
1258 }
1259 #endif  /* INVARIANT_SUPPORT */
1260
1261 #ifdef DDB
1262 static void
1263 db_show_sx(const struct lock_object *lock)
1264 {
1265         struct thread *td;
1266         const struct sx *sx;
1267
1268         sx = (const struct sx *)lock;
1269
1270         db_printf(" state: ");
1271         if (sx->sx_lock == SX_LOCK_UNLOCKED)
1272                 db_printf("UNLOCKED\n");
1273         else if (sx->sx_lock == SX_LOCK_DESTROYED) {
1274                 db_printf("DESTROYED\n");
1275                 return;
1276         } else if (sx->sx_lock & SX_LOCK_SHARED)
1277                 db_printf("SLOCK: %ju\n", (uintmax_t)SX_SHARERS(sx->sx_lock));
1278         else {
1279                 td = sx_xholder(sx);
1280                 db_printf("XLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1281                     td->td_tid, td->td_proc->p_pid, td->td_name);
1282                 if (sx_recursed(sx))
1283                         db_printf(" recursed: %d\n", sx->sx_recurse);
1284         }
1285
1286         db_printf(" waiters: ");
1287         switch(sx->sx_lock &
1288             (SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS)) {
1289         case SX_LOCK_SHARED_WAITERS:
1290                 db_printf("shared\n");
1291                 break;
1292         case SX_LOCK_EXCLUSIVE_WAITERS:
1293                 db_printf("exclusive\n");
1294                 break;
1295         case SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS:
1296                 db_printf("exclusive and shared\n");
1297                 break;
1298         default:
1299                 db_printf("none\n");
1300         }
1301 }
1302
1303 /*
1304  * Check to see if a thread that is blocked on a sleep queue is actually
1305  * blocked on an sx lock.  If so, output some details and return true.
1306  * If the lock has an exclusive owner, return that in *ownerp.
1307  */
1308 int
1309 sx_chain(struct thread *td, struct thread **ownerp)
1310 {
1311         struct sx *sx;
1312
1313         /*
1314          * Check to see if this thread is blocked on an sx lock.
1315          * First, we check the lock class.  If that is ok, then we
1316          * compare the lock name against the wait message.
1317          */
1318         sx = td->td_wchan;
1319         if (LOCK_CLASS(&sx->lock_object) != &lock_class_sx ||
1320             sx->lock_object.lo_name != td->td_wmesg)
1321                 return (0);
1322
1323         /* We think we have an sx lock, so output some details. */
1324         db_printf("blocked on sx \"%s\" ", td->td_wmesg);
1325         *ownerp = sx_xholder(sx);
1326         if (sx->sx_lock & SX_LOCK_SHARED)
1327                 db_printf("SLOCK (count %ju)\n",
1328                     (uintmax_t)SX_SHARERS(sx->sx_lock));
1329         else
1330                 db_printf("XLOCK\n");
1331         return (1);
1332 }
1333 #endif