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         .initial        = 1000,
 153         .step           = 500,
 154         .min            = 100,
 155         .max            = 5000,
 156 };
 157
 158 SYSCTL_INT(_debug_sx, OID_AUTO, delay_initial, CTLFLAG_RW, &sx_delay.initial,
 159     0, "");
 160 SYSCTL_INT(_debug_sx, OID_AUTO, delay_step, CTLFLAG_RW, &sx_delay.step,
 161     0, "");
 162 SYSCTL_INT(_debug_sx, OID_AUTO, delay_min, CTLFLAG_RW, &sx_delay.min,
 163     0, "");
 164 SYSCTL_INT(_debug_sx, OID_AUTO, delay_max, CTLFLAG_RW, &sx_delay.max,
 165     0, "");
 166
 167 static void
 168 sx_delay_sysinit(void *dummy)
 169 {
 170
 171         sx_delay.initial = mp_ncpus * 25;
 172         sx_delay.step = (mp_ncpus * 25) / 2;
 173         sx_delay.min = mp_ncpus * 5;
 174         sx_delay.max = mp_ncpus * 25 * 10;
 175 }
 176 LOCK_DELAY_SYSINIT(sx_delay_sysinit);
 177 #endif
 178
 179 void
 180 assert_sx(const struct lock_object *lock, int what)
 181 {
 182
 183         sx_assert((const struct sx *)lock, what);
 184 }
 185
 186 void
 187 lock_sx(struct lock_object *lock, uintptr_t how)
 188 {
 189         struct sx *sx;
 190
 191         sx = (struct sx *)lock;
 192         if (how)
 193                 sx_slock(sx);
 194         else
 195                 sx_xlock(sx);
 196 }
 197
 198 uintptr_t
 199 unlock_sx(struct lock_object *lock)
 200 {
 201         struct sx *sx;
 202
 203         sx = (struct sx *)lock;
 204         sx_assert(sx, SA_LOCKED | SA_NOTRECURSED);
 205         if (sx_xlocked(sx)) {
 206                 sx_xunlock(sx);
 207                 return (0);
 208         } else {
 209                 sx_sunlock(sx);
 210                 return (1);
 211         }
 212 }
 213
 214 #ifdef KDTRACE_HOOKS
 215 int
 216 owner_sx(const struct lock_object *lock, struct thread **owner)
 217 {
 218         const struct sx *sx;
 219         uintptr_t x;
 220
 221         sx = (const struct sx *)lock;
 222         x = sx->sx_lock;
 223         *owner = NULL;
 224         return ((x & SX_LOCK_SHARED) != 0 ? (SX_SHARERS(x) != 0) :
 225             ((*owner = (struct thread *)SX_OWNER(x)) != NULL));
 226 }
 227 #endif
 228
 229 void
 230 sx_sysinit(void *arg)
 231 {
 232         struct sx_args *sargs = arg;
 233
 234         sx_init_flags(sargs->sa_sx, sargs->sa_desc, sargs->sa_flags);
 235 }
 236
 237 void
 238 sx_init_flags(struct sx *sx, const char *description, int opts)
 239 {
 240         int flags;
 241
 242         MPASS((opts & ~(SX_QUIET | SX_RECURSE | SX_NOWITNESS | SX_DUPOK |
 243             SX_NOPROFILE | SX_NOADAPTIVE | SX_NEW)) == 0);
 244         ASSERT_ATOMIC_LOAD_PTR(sx->sx_lock,
 245             ("%s: sx_lock not aligned for %s: %p", __func__, description,
 246             &sx->sx_lock));
 247
 248         flags = LO_SLEEPABLE | LO_UPGRADABLE;
 249         if (opts & SX_DUPOK)
 250                 flags |= LO_DUPOK;
 251         if (opts & SX_NOPROFILE)
 252                 flags |= LO_NOPROFILE;
 253         if (!(opts & SX_NOWITNESS))
 254                 flags |= LO_WITNESS;
 255         if (opts & SX_RECURSE)
 256                 flags |= LO_RECURSABLE;
 257         if (opts & SX_QUIET)
 258                 flags |= LO_QUIET;
 259         if (opts & SX_NEW)
 260                 flags |= LO_NEW;
 261
 262         flags |= opts & SX_NOADAPTIVE;
 263         lock_init(&sx->lock_object, &lock_class_sx, description, NULL, flags);
 264         sx->sx_lock = SX_LOCK_UNLOCKED;
 265         sx->sx_recurse = 0;
 266 }
 267
 268 void
 269 sx_destroy(struct sx *sx)
 270 {
 271
 272         KASSERT(sx->sx_lock == SX_LOCK_UNLOCKED, ("sx lock still held"));
 273         KASSERT(sx->sx_recurse == 0, ("sx lock still recursed"));
 274         sx->sx_lock = SX_LOCK_DESTROYED;
 275         lock_destroy(&sx->lock_object);
 276 }
 277
 278 int
 279 _sx_slock(struct sx *sx, int opts, const char *file, int line)
 280 {
 281         int error = 0;
 282
 283         if (SCHEDULER_STOPPED())
 284                 return (0);
 285         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
 286             ("sx_slock() by idle thread %p on sx %s @ %s:%d",
 287             curthread, sx->lock_object.lo_name, file, line));
 288         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 289             ("sx_slock() of destroyed sx @ %s:%d", file, line));
 290         WITNESS_CHECKORDER(&sx->lock_object, LOP_NEWORDER, file, line, NULL);
 291         error = __sx_slock(sx, opts, file, line);
 292         if (!error) {
 293                 LOCK_LOG_LOCK("SLOCK", &sx->lock_object, 0, 0, file, line);
 294                 WITNESS_LOCK(&sx->lock_object, 0, file, line);
 295                 TD_LOCKS_INC(curthread);
 296         }
 297
 298         return (error);
 299 }
 300
 301 int
 302 sx_try_slock_(struct sx *sx, const char *file, int line)
 303 {
 304         uintptr_t x;
 305
 306         if (SCHEDULER_STOPPED())
 307                 return (1);
 308
 309         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
 310             ("sx_try_slock() by idle thread %p on sx %s @ %s:%d",
 311             curthread, sx->lock_object.lo_name, file, line));
 312
 313         for (;;) {
 314                 x = sx->sx_lock;
 315                 KASSERT(x != SX_LOCK_DESTROYED,
 316                     ("sx_try_slock() of destroyed sx @ %s:%d", file, line));
 317                 if (!(x & SX_LOCK_SHARED))
 318                         break;
 319                 if (atomic_cmpset_acq_ptr(&sx->sx_lock, x, x + SX_ONE_SHARER)) {
 320                         LOCK_LOG_TRY("SLOCK", &sx->lock_object, 0, 1, file, line);
 321                         WITNESS_LOCK(&sx->lock_object, LOP_TRYLOCK, file, line);
 322                         LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire,
 323                             sx, 0, 0, file, line, LOCKSTAT_READER);
 324                         TD_LOCKS_INC(curthread);
 325                         return (1);
 326                 }
 327         }
 328
 329         LOCK_LOG_TRY("SLOCK", &sx->lock_object, 0, 0, file, line);
 330         return (0);
 331 }
 332
 333 int
 334 _sx_xlock(struct sx *sx, int opts, const char *file, int line)
 335 {
 336         int error = 0;
 337
 338         if (SCHEDULER_STOPPED())
 339                 return (0);
 340         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
 341             ("sx_xlock() by idle thread %p on sx %s @ %s:%d",
 342             curthread, sx->lock_object.lo_name, file, line));
 343         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 344             ("sx_xlock() of destroyed sx @ %s:%d", file, line));
 345         WITNESS_CHECKORDER(&sx->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE, file,
 346             line, NULL);
 347         error = __sx_xlock(sx, curthread, opts, file, line);
 348         if (!error) {
 349                 LOCK_LOG_LOCK("XLOCK", &sx->lock_object, 0, sx->sx_recurse,
 350                     file, line);
 351                 WITNESS_LOCK(&sx->lock_object, LOP_EXCLUSIVE, file, line);
 352                 TD_LOCKS_INC(curthread);
 353         }
 354
 355         return (error);
 356 }
 357
 358 int
 359 sx_try_xlock_(struct sx *sx, const char *file, int line)
 360 {
 361         int rval;
 362
 363         if (SCHEDULER_STOPPED())
 364                 return (1);
 365
 366         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
 367             ("sx_try_xlock() by idle thread %p on sx %s @ %s:%d",
 368             curthread, sx->lock_object.lo_name, file, line));
 369         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 370             ("sx_try_xlock() of destroyed sx @ %s:%d", file, line));
 371
 372         if (sx_xlocked(sx) &&
 373             (sx->lock_object.lo_flags & LO_RECURSABLE) != 0) {
 374                 sx->sx_recurse++;
 375                 atomic_set_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
 376                 rval = 1;
 377         } else
 378                 rval = atomic_cmpset_acq_ptr(&sx->sx_lock, SX_LOCK_UNLOCKED,
 379                     (uintptr_t)curthread);
 380         LOCK_LOG_TRY("XLOCK", &sx->lock_object, 0, rval, file, line);
 381         if (rval) {
 382                 WITNESS_LOCK(&sx->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
 383                     file, line);
 384                 if (!sx_recursed(sx))
 385                         LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire,
 386                             sx, 0, 0, file, line, LOCKSTAT_WRITER);
 387                 TD_LOCKS_INC(curthread);
 388         }
 389
 390         return (rval);
 391 }
 392
 393 void
 394 _sx_sunlock(struct sx *sx, const char *file, int line)
 395 {
 396
 397         if (SCHEDULER_STOPPED())
 398                 return;
 399         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 400             ("sx_sunlock() of destroyed sx @ %s:%d", file, line));
 401         _sx_assert(sx, SA_SLOCKED, file, line);
 402         WITNESS_UNLOCK(&sx->lock_object, 0, file, line);
 403         LOCK_LOG_LOCK("SUNLOCK", &sx->lock_object, 0, 0, file, line);
 404         __sx_sunlock(sx, file, line);
 405         TD_LOCKS_DEC(curthread);
 406 }
 407
 408 void
 409 _sx_xunlock(struct sx *sx, const char *file, int line)
 410 {
 411
 412         if (SCHEDULER_STOPPED())
 413                 return;
 414         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 415             ("sx_xunlock() of destroyed sx @ %s:%d", file, line));
 416         _sx_assert(sx, SA_XLOCKED, file, line);
 417         WITNESS_UNLOCK(&sx->lock_object, LOP_EXCLUSIVE, file, line);
 418         LOCK_LOG_LOCK("XUNLOCK", &sx->lock_object, 0, sx->sx_recurse, file,
 419             line);
 420         __sx_xunlock(sx, curthread, file, line);
 421         TD_LOCKS_DEC(curthread);
 422 }
 423
 424 /*
 425  * Try to do a non-blocking upgrade from a shared lock to an exclusive lock.
 426  * This will only succeed if this thread holds a single shared lock.
 427  * Return 1 if if the upgrade succeed, 0 otherwise.
 428  */
 429 int
 430 sx_try_upgrade_(struct sx *sx, const char *file, int line)
 431 {
 432         uintptr_t x;
 433         int success;
 434
 435         if (SCHEDULER_STOPPED())
 436                 return (1);
 437
 438         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 439             ("sx_try_upgrade() of destroyed sx @ %s:%d", file, line));
 440         _sx_assert(sx, SA_SLOCKED, file, line);
 441
 442         /*
 443          * Try to switch from one shared lock to an exclusive lock.  We need
 444          * to maintain the SX_LOCK_EXCLUSIVE_WAITERS flag if set so that
 445          * we will wake up the exclusive waiters when we drop the lock.
 446          */
 447         x = sx->sx_lock & SX_LOCK_EXCLUSIVE_WAITERS;
 448         success = atomic_cmpset_ptr(&sx->sx_lock, SX_SHARERS_LOCK(1) | x,
 449             (uintptr_t)curthread | x);
 450         LOCK_LOG_TRY("XUPGRADE", &sx->lock_object, 0, success, file, line);
 451         if (success) {
 452                 WITNESS_UPGRADE(&sx->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
 453                     file, line);
 454                 LOCKSTAT_RECORD0(sx__upgrade, sx);
 455         }
 456         return (success);
 457 }
 458
 459 /*
 460  * Downgrade an unrecursed exclusive lock into a single shared lock.
 461  */
 462 void
 463 sx_downgrade_(struct sx *sx, const char *file, int line)
 464 {
 465         uintptr_t x;
 466         int wakeup_swapper;
 467
 468         if (SCHEDULER_STOPPED())
 469                 return;
 470
 471         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 472             ("sx_downgrade() of destroyed sx @ %s:%d", file, line));
 473         _sx_assert(sx, SA_XLOCKED | SA_NOTRECURSED, file, line);
 474 #ifndef INVARIANTS
 475         if (sx_recursed(sx))
 476                 panic("downgrade of a recursed lock");
 477 #endif
 478
 479         WITNESS_DOWNGRADE(&sx->lock_object, 0, file, line);
 480
 481         /*
 482          * Try to switch from an exclusive lock with no shared waiters
 483          * to one sharer with no shared waiters.  If there are
 484          * exclusive waiters, we don't need to lock the sleep queue so
 485          * long as we preserve the flag.  We do one quick try and if
 486          * that fails we grab the sleepq lock to keep the flags from
 487          * changing and do it the slow way.
 488          *
 489          * We have to lock the sleep queue if there are shared waiters
 490          * so we can wake them up.
 491          */
 492         x = sx->sx_lock;
 493         if (!(x & SX_LOCK_SHARED_WAITERS) &&
 494             atomic_cmpset_rel_ptr(&sx->sx_lock, x, SX_SHARERS_LOCK(1) |
 495             (x & SX_LOCK_EXCLUSIVE_WAITERS))) {
 496                 LOCK_LOG_LOCK("XDOWNGRADE", &sx->lock_object, 0, 0, file, line);
 497                 return;
 498         }
 499
 500         /*
 501          * Lock the sleep queue so we can read the waiters bits
 502          * without any races and wakeup any shared waiters.
 503          */
 504         sleepq_lock(&sx->lock_object);
 505
 506         /*
 507          * Preserve SX_LOCK_EXCLUSIVE_WAITERS while downgraded to a single
 508          * shared lock.  If there are any shared waiters, wake them up.
 509          */
 510         wakeup_swapper = 0;
 511         x = sx->sx_lock;
 512         atomic_store_rel_ptr(&sx->sx_lock, SX_SHARERS_LOCK(1) |
 513             (x & SX_LOCK_EXCLUSIVE_WAITERS));
 514         if (x & SX_LOCK_SHARED_WAITERS)
 515                 wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX,
 516                     0, SQ_SHARED_QUEUE);
 517         sleepq_release(&sx->lock_object);
 518
 519         LOCK_LOG_LOCK("XDOWNGRADE", &sx->lock_object, 0, 0, file, line);
 520         LOCKSTAT_RECORD0(sx__downgrade, sx);
 521
 522         if (wakeup_swapper)
 523                 kick_proc0();
 524 }
 525
 526 /*
 527  * This function represents the so-called 'hard case' for sx_xlock
 528  * operation.  All 'easy case' failures are redirected to this.  Note
 529  * that ideally this would be a static function, but it needs to be
 530  * accessible from at least sx.h.
 531  */
 532 int
 533 _sx_xlock_hard(struct sx *sx, uintptr_t tid, int opts, const char *file,
 534     int line)
 535 {
 536         GIANT_DECLARE;
 537 #ifdef ADAPTIVE_SX
 538         volatile struct thread *owner;
 539         u_int i, spintries = 0;
 540 #endif
 541         uintptr_t x;
 542 #ifdef LOCK_PROFILING
 543         uint64_t waittime = 0;
 544         int contested = 0;
 545 #endif
 546         int error = 0;
 547 #if defined(ADAPTIVE_SX) || defined(KDTRACE_HOOKS)
 548         struct lock_delay_arg lda;
 549 #endif
 550 #ifdef  KDTRACE_HOOKS
 551         uintptr_t state;
 552         u_int sleep_cnt = 0;
 553         int64_t sleep_time = 0;
 554         int64_t all_time = 0;
 555 #endif
 556
 557         if (SCHEDULER_STOPPED())
 558                 return (0);
 559
 560 #if defined(ADAPTIVE_SX)
 561         lock_delay_arg_init(&lda, &sx_delay);
 562 #elif defined(KDTRACE_HOOKS)
 563         lock_delay_arg_init(&lda, NULL);
 564 #endif
 565
 566         x = SX_READ_VALUE(sx);
 567
 568         /* If we already hold an exclusive lock, then recurse. */
 569         if (__predict_false(lv_sx_owner(x) == (struct thread *)tid)) {
 570                 KASSERT((sx->lock_object.lo_flags & LO_RECURSABLE) != 0,
 571             ("_sx_xlock_hard: recursed on non-recursive sx %s @ %s:%d\n",
 572                     sx->lock_object.lo_name, file, line));
 573                 sx->sx_recurse++;
 574                 atomic_set_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
 575                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 576                         CTR2(KTR_LOCK, "%s: %p recursing", __func__, sx);
 577                 return (0);
 578         }
 579
 580         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 581                 CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__,
 582                     sx->lock_object.lo_name, (void *)sx->sx_lock, file, line);
 583
 584 #ifdef KDTRACE_HOOKS
 585         all_time -= lockstat_nsecs(&sx->lock_object);
 586         state = x;
 587 #endif
 588         for (;;) {
 589                 if (x == SX_LOCK_UNLOCKED) {
 590                         if (atomic_cmpset_acq_ptr(&sx->sx_lock, x, tid))
 591                                 break;
 592                         x = SX_READ_VALUE(sx);
 593                         continue;
 594                 }
 595 #ifdef KDTRACE_HOOKS
 596                 lda.spin_cnt++;
 597 #endif
 598 #ifdef HWPMC_HOOKS
 599                 PMC_SOFT_CALL( , , lock, failed);
 600 #endif
 601                 lock_profile_obtain_lock_failed(&sx->lock_object, &contested,
 602                     &waittime);
 603 #ifdef ADAPTIVE_SX
 604                 /*
 605                  * If the lock is write locked and the owner is
 606                  * running on another CPU, spin until the owner stops
 607                  * running or the state of the lock changes.
 608                  */
 609                 if ((sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) {
 610                         if ((x & SX_LOCK_SHARED) == 0) {
 611                                 owner = lv_sx_owner(x);
 612                                 if (TD_IS_RUNNING(owner)) {
 613                                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 614                                                 CTR3(KTR_LOCK,
 615                                             "%s: spinning on %p held by %p",
 616                                                     __func__, sx, owner);
 617                                         KTR_STATE1(KTR_SCHED, "thread",
 618                                             sched_tdname(curthread), "spinning",
 619                                             "lockname:\"%s\"",
 620                                             sx->lock_object.lo_name);
 621                                         GIANT_SAVE();
 622                                         do {
 623                                                 lock_delay(&lda);
 624                                                 x = SX_READ_VALUE(sx);
 625                                                 owner = lv_sx_owner(x);
 626                                         } while (owner != NULL &&
 627                                                     TD_IS_RUNNING(owner));
 628                                         KTR_STATE0(KTR_SCHED, "thread",
 629                                             sched_tdname(curthread), "running");
 630                                         continue;
 631                                 }
 632                         } else if (SX_SHARERS(x) && spintries < asx_retries) {
 633                                 KTR_STATE1(KTR_SCHED, "thread",
 634                                     sched_tdname(curthread), "spinning",
 635                                     "lockname:\"%s\"", sx->lock_object.lo_name);
 636                                 GIANT_SAVE();
 637                                 spintries++;
 638                                 for (i = 0; i < asx_loops; i++) {
 639                                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 640                                                 CTR4(KTR_LOCK,
 641                                     "%s: shared spinning on %p with %u and %u",
 642                                                     __func__, sx, spintries, i);
 643                                         x = sx->sx_lock;
 644                                         if ((x & SX_LOCK_SHARED) == 0 ||
 645                                             SX_SHARERS(x) == 0)
 646                                                 break;
 647                                         cpu_spinwait();
 648 #ifdef KDTRACE_HOOKS
 649                                         lda.spin_cnt++;
 650 #endif
 651                                 }
 652                                 KTR_STATE0(KTR_SCHED, "thread",
 653                                     sched_tdname(curthread), "running");
 654                                 x = SX_READ_VALUE(sx);
 655                                 if (i != asx_loops)
 656                                         continue;
 657                         }
 658                 }
 659 #endif
 660
 661                 sleepq_lock(&sx->lock_object);
 662                 x = SX_READ_VALUE(sx);
 663
 664                 /*
 665                  * If the lock was released while spinning on the
 666                  * sleep queue chain lock, try again.
 667                  */
 668                 if (x == SX_LOCK_UNLOCKED) {
 669                         sleepq_release(&sx->lock_object);
 670                         continue;
 671                 }
 672
 673 #ifdef ADAPTIVE_SX
 674                 /*
 675                  * The current lock owner might have started executing
 676                  * on another CPU (or the lock could have changed
 677                  * owners) while we were waiting on the sleep queue
 678                  * chain lock.  If so, drop the sleep queue lock and try
 679                  * again.
 680                  */
 681                 if (!(x & SX_LOCK_SHARED) &&
 682                     (sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) {
 683                         owner = (struct thread *)SX_OWNER(x);
 684                         if (TD_IS_RUNNING(owner)) {
 685                                 sleepq_release(&sx->lock_object);
 686                                 continue;
 687                         }
 688                 }
 689 #endif
 690
 691                 /*
 692                  * If an exclusive lock was released with both shared
 693                  * and exclusive waiters and a shared waiter hasn't
 694                  * woken up and acquired the lock yet, sx_lock will be
 695                  * set to SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS.
 696                  * If we see that value, try to acquire it once.  Note
 697                  * that we have to preserve SX_LOCK_EXCLUSIVE_WAITERS
 698                  * as there are other exclusive waiters still.  If we
 699                  * fail, restart the loop.
 700                  */
 701                 if (x == (SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS)) {
 702                         if (atomic_cmpset_acq_ptr(&sx->sx_lock,
 703                             SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS,
 704                             tid | SX_LOCK_EXCLUSIVE_WAITERS)) {
 705                                 sleepq_release(&sx->lock_object);
 706                                 CTR2(KTR_LOCK, "%s: %p claimed by new writer",
 707                                     __func__, sx);
 708                                 break;
 709                         }
 710                         sleepq_release(&sx->lock_object);
 711                         x = SX_READ_VALUE(sx);
 712                         continue;
 713                 }
 714
 715                 /*
 716                  * Try to set the SX_LOCK_EXCLUSIVE_WAITERS.  If we fail,
 717                  * than loop back and retry.
 718                  */
 719                 if (!(x & SX_LOCK_EXCLUSIVE_WAITERS)) {
 720                         if (!atomic_cmpset_ptr(&sx->sx_lock, x,
 721                             x | SX_LOCK_EXCLUSIVE_WAITERS)) {
 722                                 sleepq_release(&sx->lock_object);
 723                                 x = SX_READ_VALUE(sx);
 724                                 continue;
 725                         }
 726                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 727                                 CTR2(KTR_LOCK, "%s: %p set excl waiters flag",
 728                                     __func__, sx);
 729                 }
 730
 731                 /*
 732                  * Since we have been unable to acquire the exclusive
 733                  * lock and the exclusive waiters flag is set, we have
 734                  * to sleep.
 735                  */
 736                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 737                         CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
 738                             __func__, sx);
 739
 740 #ifdef KDTRACE_HOOKS
 741                 sleep_time -= lockstat_nsecs(&sx->lock_object);
 742 #endif
 743                 GIANT_SAVE();
 744                 sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
 745                     SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
 746                     SLEEPQ_INTERRUPTIBLE : 0), SQ_EXCLUSIVE_QUEUE);
 747                 if (!(opts & SX_INTERRUPTIBLE))
 748                         sleepq_wait(&sx->lock_object, 0);
 749                 else
 750                         error = sleepq_wait_sig(&sx->lock_object, 0);
 751 #ifdef KDTRACE_HOOKS
 752                 sleep_time += lockstat_nsecs(&sx->lock_object);
 753                 sleep_cnt++;
 754 #endif
 755                 if (error) {
 756                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 757                                 CTR2(KTR_LOCK,
 758                         "%s: interruptible sleep by %p suspended by signal",
 759                                     __func__, sx);
 760                         break;
 761                 }
 762                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 763                         CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
 764                             __func__, sx);
 765                 x = SX_READ_VALUE(sx);
 766         }
 767 #ifdef KDTRACE_HOOKS
 768         all_time += lockstat_nsecs(&sx->lock_object);
 769         if (sleep_time)
 770                 LOCKSTAT_RECORD4(sx__block, sx, sleep_time,
 771                     LOCKSTAT_WRITER, (state & SX_LOCK_SHARED) == 0,
 772                     (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
 773         if (lda.spin_cnt > sleep_cnt)
 774                 LOCKSTAT_RECORD4(sx__spin, sx, all_time - sleep_time,
 775                     LOCKSTAT_WRITER, (state & SX_LOCK_SHARED) == 0,
 776                     (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
 777 #endif
 778         if (!error)
 779                 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
 780                     contested, waittime, file, line, LOCKSTAT_WRITER);
 781         GIANT_RESTORE();
 782         return (error);
 783 }
 784
 785 /*
 786  * This function represents the so-called 'hard case' for sx_xunlock
 787  * operation.  All 'easy case' failures are redirected to this.  Note
 788  * that ideally this would be a static function, but it needs to be
 789  * accessible from at least sx.h.
 790  */
 791 void
 792 _sx_xunlock_hard(struct sx *sx, uintptr_t tid, const char *file, int line)
 793 {
 794         uintptr_t x;
 795         int queue, wakeup_swapper;
 796
 797         if (SCHEDULER_STOPPED())
 798                 return;
 799
 800         MPASS(!(sx->sx_lock & SX_LOCK_SHARED));
 801
 802         /* If the lock is recursed, then unrecurse one level. */
 803         if (sx_xlocked(sx) && sx_recursed(sx)) {
 804                 if ((--sx->sx_recurse) == 0)
 805                         atomic_clear_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
 806                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 807                         CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, sx);
 808                 return;
 809         }
 810         MPASS(sx->sx_lock & (SX_LOCK_SHARED_WAITERS |
 811             SX_LOCK_EXCLUSIVE_WAITERS));
 812         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 813                 CTR2(KTR_LOCK, "%s: %p contested", __func__, sx);
 814
 815         sleepq_lock(&sx->lock_object);
 816         x = SX_LOCK_UNLOCKED;
 817
 818         /*
 819          * The wake up algorithm here is quite simple and probably not
 820          * ideal.  It gives precedence to shared waiters if they are
 821          * present.  For this condition, we have to preserve the
 822          * state of the exclusive waiters flag.
 823          * If interruptible sleeps left the shared queue empty avoid a
 824          * starvation for the threads sleeping on the exclusive queue by giving
 825          * them precedence and cleaning up the shared waiters bit anyway.
 826          */
 827         if ((sx->sx_lock & SX_LOCK_SHARED_WAITERS) != 0 &&
 828             sleepq_sleepcnt(&sx->lock_object, SQ_SHARED_QUEUE) != 0) {
 829                 queue = SQ_SHARED_QUEUE;
 830                 x |= (sx->sx_lock & SX_LOCK_EXCLUSIVE_WAITERS);
 831         } else
 832                 queue = SQ_EXCLUSIVE_QUEUE;
 833
 834         /* Wake up all the waiters for the specific queue. */
 835         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 836                 CTR3(KTR_LOCK, "%s: %p waking up all threads on %s queue",
 837                     __func__, sx, queue == SQ_SHARED_QUEUE ? "shared" :
 838                     "exclusive");
 839         atomic_store_rel_ptr(&sx->sx_lock, x);
 840         wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX, 0,
 841             queue);
 842         sleepq_release(&sx->lock_object);
 843         if (wakeup_swapper)
 844                 kick_proc0();
 845 }
 846
 847 /*
 848  * This function represents the so-called 'hard case' for sx_slock
 849  * operation.  All 'easy case' failures are redirected to this.  Note
 850  * that ideally this would be a static function, but it needs to be
 851  * accessible from at least sx.h.
 852  */
 853 int
 854 _sx_slock_hard(struct sx *sx, int opts, const char *file, int line)
 855 {
 856         GIANT_DECLARE;
 857 #ifdef ADAPTIVE_SX
 858         volatile struct thread *owner;
 859 #endif
 860 #ifdef LOCK_PROFILING
 861         uint64_t waittime = 0;
 862         int contested = 0;
 863 #endif
 864         uintptr_t x;
 865         int error = 0;
 866 #if defined(ADAPTIVE_SX) || defined(KDTRACE_HOOKS)
 867         struct lock_delay_arg lda;
 868 #endif
 869 #ifdef KDTRACE_HOOKS
 870         uintptr_t state;
 871         u_int sleep_cnt = 0;
 872         int64_t sleep_time = 0;
 873         int64_t all_time = 0;
 874 #endif
 875
 876         if (SCHEDULER_STOPPED())
 877                 return (0);
 878
 879 #if defined(ADAPTIVE_SX)
 880         lock_delay_arg_init(&lda, &sx_delay);
 881 #elif defined(KDTRACE_HOOKS)
 882         lock_delay_arg_init(&lda, NULL);
 883 #endif
 884 #ifdef KDTRACE_HOOKS
 885         all_time -= lockstat_nsecs(&sx->lock_object);
 886 #endif
 887         x = SX_READ_VALUE(sx);
 888 #ifdef KDTRACE_HOOKS
 889         state = x;
 890 #endif
 891
 892         /*
 893          * As with rwlocks, we don't make any attempt to try to block
 894          * shared locks once there is an exclusive waiter.
 895          */
 896         for (;;) {
 897                 /*
 898                  * If no other thread has an exclusive lock then try to bump up
 899                  * the count of sharers.  Since we have to preserve the state
 900                  * of SX_LOCK_EXCLUSIVE_WAITERS, if we fail to acquire the
 901                  * shared lock loop back and retry.
 902                  */
 903                 if (x & SX_LOCK_SHARED) {
 904                         MPASS(!(x & SX_LOCK_SHARED_WAITERS));
 905                         if (atomic_cmpset_acq_ptr(&sx->sx_lock, x,
 906                             x + SX_ONE_SHARER)) {
 907                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 908                                         CTR4(KTR_LOCK,
 909                                             "%s: %p succeed %p -> %p", __func__,
 910                                             sx, (void *)x,
 911                                             (void *)(x + SX_ONE_SHARER));
 912                                 break;
 913                         }
 914                         x = SX_READ_VALUE(sx);
 915                         continue;
 916                 }
 917 #ifdef KDTRACE_HOOKS
 918                 lda.spin_cnt++;
 919 #endif
 920
 921 #ifdef HWPMC_HOOKS
 922                 PMC_SOFT_CALL( , , lock, failed);
 923 #endif
 924                 lock_profile_obtain_lock_failed(&sx->lock_object, &contested,
 925                     &waittime);
 926
 927 #ifdef ADAPTIVE_SX
 928                 /*
 929                  * If the owner is running on another CPU, spin until
 930                  * the owner stops running or the state of the lock
 931                  * changes.
 932                  */
 933                 if ((sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) {
 934                         owner = lv_sx_owner(x);
 935                         if (TD_IS_RUNNING(owner)) {
 936                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 937                                         CTR3(KTR_LOCK,
 938                                             "%s: spinning on %p held by %p",
 939                                             __func__, sx, owner);
 940                                 KTR_STATE1(KTR_SCHED, "thread",
 941                                     sched_tdname(curthread), "spinning",
 942                                     "lockname:\"%s\"", sx->lock_object.lo_name);
 943                                 GIANT_SAVE();
 944                                 do {
 945                                         lock_delay(&lda);
 946                                         x = SX_READ_VALUE(sx);
 947                                         owner = lv_sx_owner(x);
 948                                 } while (owner != NULL && TD_IS_RUNNING(owner));
 949                                 KTR_STATE0(KTR_SCHED, "thread",
 950                                     sched_tdname(curthread), "running");
 951                                 continue;
 952                         }
 953                 }
 954 #endif
 955
 956                 /*
 957                  * Some other thread already has an exclusive lock, so
 958                  * start the process of blocking.
 959                  */
 960                 sleepq_lock(&sx->lock_object);
 961                 x = SX_READ_VALUE(sx);
 962
 963                 /*
 964                  * The lock could have been released while we spun.
 965                  * In this case loop back and retry.
 966                  */
 967                 if (x & SX_LOCK_SHARED) {
 968                         sleepq_release(&sx->lock_object);
 969                         continue;
 970                 }
 971
 972 #ifdef ADAPTIVE_SX
 973                 /*
 974                  * If the owner is running on another CPU, spin until
 975                  * the owner stops running or the state of the lock
 976                  * changes.
 977                  */
 978                 if (!(x & SX_LOCK_SHARED) &&
 979                     (sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) {
 980                         owner = (struct thread *)SX_OWNER(x);
 981                         if (TD_IS_RUNNING(owner)) {
 982                                 sleepq_release(&sx->lock_object);
 983                                 x = SX_READ_VALUE(sx);
 984                                 continue;
 985                         }
 986                 }
 987 #endif
 988
 989                 /*
 990                  * Try to set the SX_LOCK_SHARED_WAITERS flag.  If we
 991                  * fail to set it drop the sleep queue lock and loop
 992                  * back.
 993                  */
 994                 if (!(x & SX_LOCK_SHARED_WAITERS)) {
 995                         if (!atomic_cmpset_ptr(&sx->sx_lock, x,
 996                             x | SX_LOCK_SHARED_WAITERS)) {
 997                                 sleepq_release(&sx->lock_object);
 998                                 x = SX_READ_VALUE(sx);
 999                                 continue;
1000                         }
1001                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
1002                                 CTR2(KTR_LOCK, "%s: %p set shared waiters flag",
1003                                     __func__, sx);
1004                 }
1005
1006                 /*
1007                  * Since we have been unable to acquire the shared lock,
1008                  * we have to sleep.
1009                  */
1010                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1011                         CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
1012                             __func__, sx);
1013
1014 #ifdef KDTRACE_HOOKS
1015                 sleep_time -= lockstat_nsecs(&sx->lock_object);
1016 #endif
1017                 GIANT_SAVE();
1018                 sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
1019                     SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
1020                     SLEEPQ_INTERRUPTIBLE : 0), SQ_SHARED_QUEUE);
1021                 if (!(opts & SX_INTERRUPTIBLE))
1022                         sleepq_wait(&sx->lock_object, 0);
1023                 else
1024                         error = sleepq_wait_sig(&sx->lock_object, 0);
1025 #ifdef KDTRACE_HOOKS
1026                 sleep_time += lockstat_nsecs(&sx->lock_object);
1027                 sleep_cnt++;
1028 #endif
1029                 if (error) {
1030                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
1031                                 CTR2(KTR_LOCK,
1032                         "%s: interruptible sleep by %p suspended by signal",
1033                                     __func__, sx);
1034                         break;
1035                 }
1036                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1037                         CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
1038                             __func__, sx);
1039                 x = SX_READ_VALUE(sx);
1040         }
1041 #ifdef KDTRACE_HOOKS
1042         all_time += lockstat_nsecs(&sx->lock_object);
1043         if (sleep_time)
1044                 LOCKSTAT_RECORD4(sx__block, sx, sleep_time,
1045                     LOCKSTAT_READER, (state & SX_LOCK_SHARED) == 0,
1046                     (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
1047         if (lda.spin_cnt > sleep_cnt)
1048                 LOCKSTAT_RECORD4(sx__spin, sx, all_time - sleep_time,
1049                     LOCKSTAT_READER, (state & SX_LOCK_SHARED) == 0,
1050                     (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
1051 #endif
1052         if (error == 0)
1053                 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
1054                     contested, waittime, file, line, LOCKSTAT_READER);
1055         GIANT_RESTORE();
1056         return (error);
1057 }
1058
1059 /*
1060  * This function represents the so-called 'hard case' for sx_sunlock
1061  * operation.  All 'easy case' failures are redirected to this.  Note
1062  * that ideally this would be a static function, but it needs to be
1063  * accessible from at least sx.h.
1064  */
1065 void
1066 _sx_sunlock_hard(struct sx *sx, const char *file, int line)
1067 {
1068         uintptr_t x;
1069         int wakeup_swapper;
1070
1071         if (SCHEDULER_STOPPED())
1072                 return;
1073
1074         x = SX_READ_VALUE(sx);
1075         for (;;) {
1076                 /*
1077                  * We should never have sharers while at least one thread
1078                  * holds a shared lock.
1079                  */
1080                 KASSERT(!(x & SX_LOCK_SHARED_WAITERS),
1081                     ("%s: waiting sharers", __func__));
1082
1083                 /*
1084                  * See if there is more than one shared lock held.  If
1085                  * so, just drop one and return.
1086                  */
1087                 if (SX_SHARERS(x) > 1) {
1088                         if (atomic_cmpset_rel_ptr(&sx->sx_lock, x,
1089                             x - SX_ONE_SHARER)) {
1090                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1091                                         CTR4(KTR_LOCK,
1092                                             "%s: %p succeeded %p -> %p",
1093                                             __func__, sx, (void *)x,
1094                                             (void *)(x - SX_ONE_SHARER));
1095                                 break;
1096                         }
1097
1098                         x = SX_READ_VALUE(sx);
1099                         continue;
1100                 }
1101
1102                 /*
1103                  * If there aren't any waiters for an exclusive lock,
1104                  * then try to drop it quickly.
1105                  */
1106                 if (!(x & SX_LOCK_EXCLUSIVE_WAITERS)) {
1107                         MPASS(x == SX_SHARERS_LOCK(1));
1108                         if (atomic_cmpset_rel_ptr(&sx->sx_lock,
1109                             SX_SHARERS_LOCK(1), SX_LOCK_UNLOCKED)) {
1110                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1111                                         CTR2(KTR_LOCK, "%s: %p last succeeded",
1112                                             __func__, sx);
1113                                 break;
1114                         }
1115                         x = SX_READ_VALUE(sx);
1116                         continue;
1117                 }
1118
1119                 /*
1120                  * At this point, there should just be one sharer with
1121                  * exclusive waiters.
1122                  */
1123                 MPASS(x == (SX_SHARERS_LOCK(1) | SX_LOCK_EXCLUSIVE_WAITERS));
1124
1125                 sleepq_lock(&sx->lock_object);
1126
1127                 /*
1128                  * Wake up semantic here is quite simple:
1129                  * Just wake up all the exclusive waiters.
1130                  * Note that the state of the lock could have changed,
1131                  * so if it fails loop back and retry.
1132                  */
1133                 if (!atomic_cmpset_rel_ptr(&sx->sx_lock,
1134                     SX_SHARERS_LOCK(1) | SX_LOCK_EXCLUSIVE_WAITERS,
1135                     SX_LOCK_UNLOCKED)) {
1136                         sleepq_release(&sx->lock_object);
1137                         x = SX_READ_VALUE(sx);
1138                         continue;
1139                 }
1140                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1141                         CTR2(KTR_LOCK, "%s: %p waking up all thread on"
1142                             "exclusive queue", __func__, sx);
1143                 wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX,
1144                     0, SQ_EXCLUSIVE_QUEUE);
1145                 sleepq_release(&sx->lock_object);
1146                 if (wakeup_swapper)
1147                         kick_proc0();
1148                 break;
1149         }
1150 }
1151
1152 #ifdef INVARIANT_SUPPORT
1153 #ifndef INVARIANTS
1154 #undef  _sx_assert
1155 #endif
1156
1157 /*
1158  * In the non-WITNESS case, sx_assert() can only detect that at least
1159  * *some* thread owns an slock, but it cannot guarantee that *this*
1160  * thread owns an slock.
1161  */
1162 void
1163 _sx_assert(const struct sx *sx, int what, const char *file, int line)
1164 {
1165 #ifndef WITNESS
1166         int slocked = 0;
1167 #endif
1168
1169         if (panicstr != NULL)
1170                 return;
1171         switch (what) {
1172         case SA_SLOCKED:
1173         case SA_SLOCKED | SA_NOTRECURSED:
1174         case SA_SLOCKED | SA_RECURSED:
1175 #ifndef WITNESS
1176                 slocked = 1;
1177                 /* FALLTHROUGH */
1178 #endif
1179         case SA_LOCKED:
1180         case SA_LOCKED | SA_NOTRECURSED:
1181         case SA_LOCKED | SA_RECURSED:
1182 #ifdef WITNESS
1183                 witness_assert(&sx->lock_object, what, file, line);
1184 #else
1185                 /*
1186                  * If some other thread has an exclusive lock or we
1187                  * have one and are asserting a shared lock, fail.
1188                  * Also, if no one has a lock at all, fail.
1189                  */
1190                 if (sx->sx_lock == SX_LOCK_UNLOCKED ||
1191                     (!(sx->sx_lock & SX_LOCK_SHARED) && (slocked ||
1192                     sx_xholder(sx) != curthread)))
1193                         panic("Lock %s not %slocked @ %s:%d\n",
1194                             sx->lock_object.lo_name, slocked ? "share " : "",
1195                             file, line);
1196
1197                 if (!(sx->sx_lock & SX_LOCK_SHARED)) {
1198                         if (sx_recursed(sx)) {
1199                                 if (what & SA_NOTRECURSED)
1200                                         panic("Lock %s recursed @ %s:%d\n",
1201                                             sx->lock_object.lo_name, file,
1202                                             line);
1203                         } else if (what & SA_RECURSED)
1204                                 panic("Lock %s not recursed @ %s:%d\n",
1205                                     sx->lock_object.lo_name, file, line);
1206                 }
1207 #endif
1208                 break;
1209         case SA_XLOCKED:
1210         case SA_XLOCKED | SA_NOTRECURSED:
1211         case SA_XLOCKED | SA_RECURSED:
1212                 if (sx_xholder(sx) != curthread)
1213                         panic("Lock %s not exclusively locked @ %s:%d\n",
1214                             sx->lock_object.lo_name, file, line);
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, line);
1219                 } else if (what & SA_RECURSED)
1220                         panic("Lock %s not recursed @ %s:%d\n",
1221                             sx->lock_object.lo_name, file, line);
1222                 break;
1223         case SA_UNLOCKED:
1224 #ifdef WITNESS
1225                 witness_assert(&sx->lock_object, what, file, line);
1226 #else
1227                 /*
1228                  * If we hold an exclusve lock fail.  We can't
1229                  * reliably check to see if we hold a shared lock or
1230                  * not.
1231                  */
1232                 if (sx_xholder(sx) == curthread)
1233                         panic("Lock %s exclusively locked @ %s:%d\n",
1234                             sx->lock_object.lo_name, file, line);
1235 #endif
1236                 break;
1237         default:
1238                 panic("Unknown sx lock assertion: %d @ %s:%d", what, file,
1239                     line);
1240         }
1241 }
1242 #endif  /* INVARIANT_SUPPORT */
1243
1244 #ifdef DDB
1245 static void
1246 db_show_sx(const struct lock_object *lock)
1247 {
1248         struct thread *td;
1249         const struct sx *sx;
1250
1251         sx = (const struct sx *)lock;
1252
1253         db_printf(" state: ");
1254         if (sx->sx_lock == SX_LOCK_UNLOCKED)
1255                 db_printf("UNLOCKED\n");
1256         else if (sx->sx_lock == SX_LOCK_DESTROYED) {
1257                 db_printf("DESTROYED\n");
1258                 return;
1259         } else if (sx->sx_lock & SX_LOCK_SHARED)
1260                 db_printf("SLOCK: %ju\n", (uintmax_t)SX_SHARERS(sx->sx_lock));
1261         else {
1262                 td = sx_xholder(sx);
1263                 db_printf("XLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1264                     td->td_tid, td->td_proc->p_pid, td->td_name);
1265                 if (sx_recursed(sx))
1266                         db_printf(" recursed: %d\n", sx->sx_recurse);
1267         }
1268
1269         db_printf(" waiters: ");
1270         switch(sx->sx_lock &
1271             (SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS)) {
1272         case SX_LOCK_SHARED_WAITERS:
1273                 db_printf("shared\n");
1274                 break;
1275         case SX_LOCK_EXCLUSIVE_WAITERS:
1276                 db_printf("exclusive\n");
1277                 break;
1278         case SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS:
1279                 db_printf("exclusive and shared\n");
1280                 break;
1281         default:
1282                 db_printf("none\n");
1283         }
1284 }
1285
1286 /*
1287  * Check to see if a thread that is blocked on a sleep queue is actually
1288  * blocked on an sx lock.  If so, output some details and return true.
1289  * If the lock has an exclusive owner, return that in *ownerp.
1290  */
1291 int
1292 sx_chain(struct thread *td, struct thread **ownerp)
1293 {
1294         struct sx *sx;
1295
1296         /*
1297          * Check to see if this thread is blocked on an sx lock.
1298          * First, we check the lock class.  If that is ok, then we
1299          * compare the lock name against the wait message.
1300          */
1301         sx = td->td_wchan;
1302         if (LOCK_CLASS(&sx->lock_object) != &lock_class_sx ||
1303             sx->lock_object.lo_name != td->td_wmesg)
1304                 return (0);
1305
1306         /* We think we have an sx lock, so output some details. */
1307         db_printf("blocked on sx \"%s\" ", td->td_wmesg);
1308         *ownerp = sx_xholder(sx);
1309         if (sx->sx_lock & SX_LOCK_SHARED)
1310                 db_printf("SLOCK (count %ju)\n",
1311                     (uintmax_t)SX_SHARERS(sx->sx_lock));
1312         else
1313                 db_printf("XLOCK\n");
1314         return (1);
1315 }
1316 #endif