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 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 = (const struct sx *)lock;
 219         uintptr_t x = sx->sx_lock;
 220
 221         *owner = (struct thread *)SX_OWNER(x);
 222         return ((x & SX_LOCK_SHARED) != 0 ? (SX_SHARERS(x) != 0) :
 223             (*owner != NULL));
 224 }
 225 #endif
 226
 227 void
 228 sx_sysinit(void *arg)
 229 {
 230         struct sx_args *sargs = arg;
 231
 232         sx_init_flags(sargs->sa_sx, sargs->sa_desc, sargs->sa_flags);
 233 }
 234
 235 void
 236 sx_init_flags(struct sx *sx, const char *description, int opts)
 237 {
 238         int flags;
 239
 240         MPASS((opts & ~(SX_QUIET | SX_RECURSE | SX_NOWITNESS | SX_DUPOK |
 241             SX_NOPROFILE | SX_NOADAPTIVE | SX_NEW)) == 0);
 242         ASSERT_ATOMIC_LOAD_PTR(sx->sx_lock,
 243             ("%s: sx_lock not aligned for %s: %p", __func__, description,
 244             &sx->sx_lock));
 245
 246         flags = LO_SLEEPABLE | LO_UPGRADABLE;
 247         if (opts & SX_DUPOK)
 248                 flags |= LO_DUPOK;
 249         if (opts & SX_NOPROFILE)
 250                 flags |= LO_NOPROFILE;
 251         if (!(opts & SX_NOWITNESS))
 252                 flags |= LO_WITNESS;
 253         if (opts & SX_RECURSE)
 254                 flags |= LO_RECURSABLE;
 255         if (opts & SX_QUIET)
 256                 flags |= LO_QUIET;
 257         if (opts & SX_NEW)
 258                 flags |= LO_NEW;
 259
 260         flags |= opts & SX_NOADAPTIVE;
 261         lock_init(&sx->lock_object, &lock_class_sx, description, NULL, flags);
 262         sx->sx_lock = SX_LOCK_UNLOCKED;
 263         sx->sx_recurse = 0;
 264 }
 265
 266 void
 267 sx_destroy(struct sx *sx)
 268 {
 269
 270         KASSERT(sx->sx_lock == SX_LOCK_UNLOCKED, ("sx lock still held"));
 271         KASSERT(sx->sx_recurse == 0, ("sx lock still recursed"));
 272         sx->sx_lock = SX_LOCK_DESTROYED;
 273         lock_destroy(&sx->lock_object);
 274 }
 275
 276 int
 277 _sx_slock(struct sx *sx, int opts, const char *file, int line)
 278 {
 279         int error = 0;
 280
 281         if (SCHEDULER_STOPPED())
 282                 return (0);
 283         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
 284             ("sx_slock() by idle thread %p on sx %s @ %s:%d",
 285             curthread, sx->lock_object.lo_name, file, line));
 286         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 287             ("sx_slock() of destroyed sx @ %s:%d", file, line));
 288         WITNESS_CHECKORDER(&sx->lock_object, LOP_NEWORDER, file, line, NULL);
 289         error = __sx_slock(sx, opts, file, line);
 290         if (!error) {
 291                 LOCK_LOG_LOCK("SLOCK", &sx->lock_object, 0, 0, file, line);
 292                 WITNESS_LOCK(&sx->lock_object, 0, file, line);
 293                 TD_LOCKS_INC(curthread);
 294         }
 295
 296         return (error);
 297 }
 298
 299 int
 300 sx_try_slock_(struct sx *sx, const char *file, int line)
 301 {
 302         uintptr_t x;
 303
 304         if (SCHEDULER_STOPPED())
 305                 return (1);
 306
 307         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
 308             ("sx_try_slock() by idle thread %p on sx %s @ %s:%d",
 309             curthread, sx->lock_object.lo_name, file, line));
 310
 311         for (;;) {
 312                 x = sx->sx_lock;
 313                 KASSERT(x != SX_LOCK_DESTROYED,
 314                     ("sx_try_slock() of destroyed sx @ %s:%d", file, line));
 315                 if (!(x & SX_LOCK_SHARED))
 316                         break;
 317                 if (atomic_cmpset_acq_ptr(&sx->sx_lock, x, x + SX_ONE_SHARER)) {
 318                         LOCK_LOG_TRY("SLOCK", &sx->lock_object, 0, 1, file, line);
 319                         WITNESS_LOCK(&sx->lock_object, LOP_TRYLOCK, file, line);
 320                         LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire,
 321                             sx, 0, 0, file, line, LOCKSTAT_READER);
 322                         TD_LOCKS_INC(curthread);
 323                         return (1);
 324                 }
 325         }
 326
 327         LOCK_LOG_TRY("SLOCK", &sx->lock_object, 0, 0, file, line);
 328         return (0);
 329 }
 330
 331 int
 332 _sx_xlock(struct sx *sx, int opts, const char *file, int line)
 333 {
 334         int error = 0;
 335
 336         if (SCHEDULER_STOPPED())
 337                 return (0);
 338         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
 339             ("sx_xlock() by idle thread %p on sx %s @ %s:%d",
 340             curthread, sx->lock_object.lo_name, file, line));
 341         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 342             ("sx_xlock() of destroyed sx @ %s:%d", file, line));
 343         WITNESS_CHECKORDER(&sx->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE, file,
 344             line, NULL);
 345         error = __sx_xlock(sx, curthread, opts, file, line);
 346         if (!error) {
 347                 LOCK_LOG_LOCK("XLOCK", &sx->lock_object, 0, sx->sx_recurse,
 348                     file, line);
 349                 WITNESS_LOCK(&sx->lock_object, LOP_EXCLUSIVE, file, line);
 350                 TD_LOCKS_INC(curthread);
 351         }
 352
 353         return (error);
 354 }
 355
 356 int
 357 sx_try_xlock_(struct sx *sx, const char *file, int line)
 358 {
 359         int rval;
 360
 361         if (SCHEDULER_STOPPED())
 362                 return (1);
 363
 364         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
 365             ("sx_try_xlock() by idle thread %p on sx %s @ %s:%d",
 366             curthread, sx->lock_object.lo_name, file, line));
 367         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 368             ("sx_try_xlock() of destroyed sx @ %s:%d", file, line));
 369
 370         if (sx_xlocked(sx) &&
 371             (sx->lock_object.lo_flags & LO_RECURSABLE) != 0) {
 372                 sx->sx_recurse++;
 373                 atomic_set_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
 374                 rval = 1;
 375         } else
 376                 rval = atomic_cmpset_acq_ptr(&sx->sx_lock, SX_LOCK_UNLOCKED,
 377                     (uintptr_t)curthread);
 378         LOCK_LOG_TRY("XLOCK", &sx->lock_object, 0, rval, file, line);
 379         if (rval) {
 380                 WITNESS_LOCK(&sx->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
 381                     file, line);
 382                 if (!sx_recursed(sx))
 383                         LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire,
 384                             sx, 0, 0, file, line, LOCKSTAT_WRITER);
 385                 TD_LOCKS_INC(curthread);
 386         }
 387
 388         return (rval);
 389 }
 390
 391 void
 392 _sx_sunlock(struct sx *sx, const char *file, int line)
 393 {
 394
 395         if (SCHEDULER_STOPPED())
 396                 return;
 397         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 398             ("sx_sunlock() of destroyed sx @ %s:%d", file, line));
 399         _sx_assert(sx, SA_SLOCKED, file, line);
 400         WITNESS_UNLOCK(&sx->lock_object, 0, file, line);
 401         LOCK_LOG_LOCK("SUNLOCK", &sx->lock_object, 0, 0, file, line);
 402         __sx_sunlock(sx, file, line);
 403         TD_LOCKS_DEC(curthread);
 404 }
 405
 406 void
 407 _sx_xunlock(struct sx *sx, const char *file, int line)
 408 {
 409
 410         if (SCHEDULER_STOPPED())
 411                 return;
 412         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 413             ("sx_xunlock() of destroyed sx @ %s:%d", file, line));
 414         _sx_assert(sx, SA_XLOCKED, file, line);
 415         WITNESS_UNLOCK(&sx->lock_object, LOP_EXCLUSIVE, file, line);
 416         LOCK_LOG_LOCK("XUNLOCK", &sx->lock_object, 0, sx->sx_recurse, file,
 417             line);
 418         __sx_xunlock(sx, curthread, file, line);
 419         TD_LOCKS_DEC(curthread);
 420 }
 421
 422 /*
 423  * Try to do a non-blocking upgrade from a shared lock to an exclusive lock.
 424  * This will only succeed if this thread holds a single shared lock.
 425  * Return 1 if if the upgrade succeed, 0 otherwise.
 426  */
 427 int
 428 sx_try_upgrade_(struct sx *sx, const char *file, int line)
 429 {
 430         uintptr_t x;
 431         int success;
 432
 433         if (SCHEDULER_STOPPED())
 434                 return (1);
 435
 436         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 437             ("sx_try_upgrade() of destroyed sx @ %s:%d", file, line));
 438         _sx_assert(sx, SA_SLOCKED, file, line);
 439
 440         /*
 441          * Try to switch from one shared lock to an exclusive lock.  We need
 442          * to maintain the SX_LOCK_EXCLUSIVE_WAITERS flag if set so that
 443          * we will wake up the exclusive waiters when we drop the lock.
 444          */
 445         x = sx->sx_lock & SX_LOCK_EXCLUSIVE_WAITERS;
 446         success = atomic_cmpset_ptr(&sx->sx_lock, SX_SHARERS_LOCK(1) | x,
 447             (uintptr_t)curthread | x);
 448         LOCK_LOG_TRY("XUPGRADE", &sx->lock_object, 0, success, file, line);
 449         if (success) {
 450                 WITNESS_UPGRADE(&sx->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
 451                     file, line);
 452                 LOCKSTAT_RECORD0(sx__upgrade, sx);
 453         }
 454         return (success);
 455 }
 456
 457 /*
 458  * Downgrade an unrecursed exclusive lock into a single shared lock.
 459  */
 460 void
 461 sx_downgrade_(struct sx *sx, const char *file, int line)
 462 {
 463         uintptr_t x;
 464         int wakeup_swapper;
 465
 466         if (SCHEDULER_STOPPED())
 467                 return;
 468
 469         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 470             ("sx_downgrade() of destroyed sx @ %s:%d", file, line));
 471         _sx_assert(sx, SA_XLOCKED | SA_NOTRECURSED, file, line);
 472 #ifndef INVARIANTS
 473         if (sx_recursed(sx))
 474                 panic("downgrade of a recursed lock");
 475 #endif
 476
 477         WITNESS_DOWNGRADE(&sx->lock_object, 0, file, line);
 478
 479         /*
 480          * Try to switch from an exclusive lock with no shared waiters
 481          * to one sharer with no shared waiters.  If there are
 482          * exclusive waiters, we don't need to lock the sleep queue so
 483          * long as we preserve the flag.  We do one quick try and if
 484          * that fails we grab the sleepq lock to keep the flags from
 485          * changing and do it the slow way.
 486          *
 487          * We have to lock the sleep queue if there are shared waiters
 488          * so we can wake them up.
 489          */
 490         x = sx->sx_lock;
 491         if (!(x & SX_LOCK_SHARED_WAITERS) &&
 492             atomic_cmpset_rel_ptr(&sx->sx_lock, x, SX_SHARERS_LOCK(1) |
 493             (x & SX_LOCK_EXCLUSIVE_WAITERS))) {
 494                 LOCK_LOG_LOCK("XDOWNGRADE", &sx->lock_object, 0, 0, file, line);
 495                 return;
 496         }
 497
 498         /*
 499          * Lock the sleep queue so we can read the waiters bits
 500          * without any races and wakeup any shared waiters.
 501          */
 502         sleepq_lock(&sx->lock_object);
 503
 504         /*
 505          * Preserve SX_LOCK_EXCLUSIVE_WAITERS while downgraded to a single
 506          * shared lock.  If there are any shared waiters, wake them up.
 507          */
 508         wakeup_swapper = 0;
 509         x = sx->sx_lock;
 510         atomic_store_rel_ptr(&sx->sx_lock, SX_SHARERS_LOCK(1) |
 511             (x & SX_LOCK_EXCLUSIVE_WAITERS));
 512         if (x & SX_LOCK_SHARED_WAITERS)
 513                 wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX,
 514                     0, SQ_SHARED_QUEUE);
 515         sleepq_release(&sx->lock_object);
 516
 517         LOCK_LOG_LOCK("XDOWNGRADE", &sx->lock_object, 0, 0, file, line);
 518         LOCKSTAT_RECORD0(sx__downgrade, sx);
 519
 520         if (wakeup_swapper)
 521                 kick_proc0();
 522 }
 523
 524 /*
 525  * This function represents the so-called 'hard case' for sx_xlock
 526  * operation.  All 'easy case' failures are redirected to this.  Note
 527  * that ideally this would be a static function, but it needs to be
 528  * accessible from at least sx.h.
 529  */
 530 int
 531 _sx_xlock_hard(struct sx *sx, uintptr_t tid, int opts, const char *file,
 532     int line)
 533 {
 534         GIANT_DECLARE;
 535 #ifdef ADAPTIVE_SX
 536         volatile struct thread *owner;
 537         u_int i, spintries = 0;
 538 #endif
 539         uintptr_t x;
 540 #ifdef LOCK_PROFILING
 541         uint64_t waittime = 0;
 542         int contested = 0;
 543 #endif
 544         int error = 0;
 545 #if defined(ADAPTIVE_SX) || defined(KDTRACE_HOOKS)
 546         struct lock_delay_arg lda;
 547 #endif
 548 #ifdef  KDTRACE_HOOKS
 549         uintptr_t state;
 550         u_int sleep_cnt = 0;
 551         int64_t sleep_time = 0;
 552         int64_t all_time = 0;
 553 #endif
 554
 555         if (SCHEDULER_STOPPED())
 556                 return (0);
 557
 558 #if defined(ADAPTIVE_SX)
 559         lock_delay_arg_init(&lda, &sx_delay);
 560 #elif defined(KDTRACE_HOOKS)
 561         lock_delay_arg_init(&lda, NULL);
 562 #endif
 563
 564         /* If we already hold an exclusive lock, then recurse. */
 565         if (sx_xlocked(sx)) {
 566                 KASSERT((sx->lock_object.lo_flags & LO_RECURSABLE) != 0,
 567             ("_sx_xlock_hard: recursed on non-recursive sx %s @ %s:%d\n",
 568                     sx->lock_object.lo_name, file, line));
 569                 sx->sx_recurse++;
 570                 atomic_set_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
 571                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 572                         CTR2(KTR_LOCK, "%s: %p recursing", __func__, sx);
 573                 return (0);
 574         }
 575
 576         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 577                 CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__,
 578                     sx->lock_object.lo_name, (void *)sx->sx_lock, file, line);
 579
 580 #ifdef KDTRACE_HOOKS
 581         all_time -= lockstat_nsecs(&sx->lock_object);
 582         state = sx->sx_lock;
 583 #endif
 584         for (;;) {
 585                 if (sx->sx_lock == SX_LOCK_UNLOCKED &&
 586                     atomic_cmpset_acq_ptr(&sx->sx_lock, SX_LOCK_UNLOCKED, tid))
 587                         break;
 588 #ifdef KDTRACE_HOOKS
 589                 lda.spin_cnt++;
 590 #endif
 591 #ifdef HWPMC_HOOKS
 592                 PMC_SOFT_CALL( , , lock, failed);
 593 #endif
 594                 lock_profile_obtain_lock_failed(&sx->lock_object, &contested,
 595                     &waittime);
 596 #ifdef ADAPTIVE_SX
 597                 /*
 598                  * If the lock is write locked and the owner is
 599                  * running on another CPU, spin until the owner stops
 600                  * running or the state of the lock changes.
 601                  */
 602                 x = sx->sx_lock;
 603                 if ((sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) {
 604                         if ((x & SX_LOCK_SHARED) == 0) {
 605                                 x = SX_OWNER(x);
 606                                 owner = (struct thread *)x;
 607                                 if (TD_IS_RUNNING(owner)) {
 608                                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 609                                                 CTR3(KTR_LOCK,
 610                                             "%s: spinning on %p held by %p",
 611                                                     __func__, sx, owner);
 612                                         KTR_STATE1(KTR_SCHED, "thread",
 613                                             sched_tdname(curthread), "spinning",
 614                                             "lockname:\"%s\"",
 615                                             sx->lock_object.lo_name);
 616                                         GIANT_SAVE();
 617                                         while (SX_OWNER(sx->sx_lock) == x &&
 618                                             TD_IS_RUNNING(owner))
 619                                                 lock_delay(&lda);
 620                                         KTR_STATE0(KTR_SCHED, "thread",
 621                                             sched_tdname(curthread), "running");
 622                                         continue;
 623                                 }
 624                         } else if (SX_SHARERS(x) && spintries < asx_retries) {
 625                                 KTR_STATE1(KTR_SCHED, "thread",
 626                                     sched_tdname(curthread), "spinning",
 627                                     "lockname:\"%s\"", sx->lock_object.lo_name);
 628                                 GIANT_SAVE();
 629                                 spintries++;
 630                                 for (i = 0; i < asx_loops; i++) {
 631                                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 632                                                 CTR4(KTR_LOCK,
 633                                     "%s: shared spinning on %p with %u and %u",
 634                                                     __func__, sx, spintries, i);
 635                                         x = sx->sx_lock;
 636                                         if ((x & SX_LOCK_SHARED) == 0 ||
 637                                             SX_SHARERS(x) == 0)
 638                                                 break;
 639                                         cpu_spinwait();
 640 #ifdef KDTRACE_HOOKS
 641                                         lda.spin_cnt++;
 642 #endif
 643                                 }
 644                                 KTR_STATE0(KTR_SCHED, "thread",
 645                                     sched_tdname(curthread), "running");
 646                                 if (i != asx_loops)
 647                                         continue;
 648                         }
 649                 }
 650 #endif
 651
 652                 sleepq_lock(&sx->lock_object);
 653                 x = sx->sx_lock;
 654
 655                 /*
 656                  * If the lock was released while spinning on the
 657                  * sleep queue chain lock, try again.
 658                  */
 659                 if (x == SX_LOCK_UNLOCKED) {
 660                         sleepq_release(&sx->lock_object);
 661                         continue;
 662                 }
 663
 664 #ifdef ADAPTIVE_SX
 665                 /*
 666                  * The current lock owner might have started executing
 667                  * on another CPU (or the lock could have changed
 668                  * owners) while we were waiting on the sleep queue
 669                  * chain lock.  If so, drop the sleep queue lock and try
 670                  * again.
 671                  */
 672                 if (!(x & SX_LOCK_SHARED) &&
 673                     (sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) {
 674                         owner = (struct thread *)SX_OWNER(x);
 675                         if (TD_IS_RUNNING(owner)) {
 676                                 sleepq_release(&sx->lock_object);
 677                                 continue;
 678                         }
 679                 }
 680 #endif
 681
 682                 /*
 683                  * If an exclusive lock was released with both shared
 684                  * and exclusive waiters and a shared waiter hasn't
 685                  * woken up and acquired the lock yet, sx_lock will be
 686                  * set to SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS.
 687                  * If we see that value, try to acquire it once.  Note
 688                  * that we have to preserve SX_LOCK_EXCLUSIVE_WAITERS
 689                  * as there are other exclusive waiters still.  If we
 690                  * fail, restart the loop.
 691                  */
 692                 if (x == (SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS)) {
 693                         if (atomic_cmpset_acq_ptr(&sx->sx_lock,
 694                             SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS,
 695                             tid | SX_LOCK_EXCLUSIVE_WAITERS)) {
 696                                 sleepq_release(&sx->lock_object);
 697                                 CTR2(KTR_LOCK, "%s: %p claimed by new writer",
 698                                     __func__, sx);
 699                                 break;
 700                         }
 701                         sleepq_release(&sx->lock_object);
 702                         continue;
 703                 }
 704
 705                 /*
 706                  * Try to set the SX_LOCK_EXCLUSIVE_WAITERS.  If we fail,
 707                  * than loop back and retry.
 708                  */
 709                 if (!(x & SX_LOCK_EXCLUSIVE_WAITERS)) {
 710                         if (!atomic_cmpset_ptr(&sx->sx_lock, x,
 711                             x | SX_LOCK_EXCLUSIVE_WAITERS)) {
 712                                 sleepq_release(&sx->lock_object);
 713                                 continue;
 714                         }
 715                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 716                                 CTR2(KTR_LOCK, "%s: %p set excl waiters flag",
 717                                     __func__, sx);
 718                 }
 719
 720                 /*
 721                  * Since we have been unable to acquire the exclusive
 722                  * lock and the exclusive waiters flag is set, we have
 723                  * to sleep.
 724                  */
 725                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 726                         CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
 727                             __func__, sx);
 728
 729 #ifdef KDTRACE_HOOKS
 730                 sleep_time -= lockstat_nsecs(&sx->lock_object);
 731 #endif
 732                 GIANT_SAVE();
 733                 sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
 734                     SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
 735                     SLEEPQ_INTERRUPTIBLE : 0), SQ_EXCLUSIVE_QUEUE);
 736                 if (!(opts & SX_INTERRUPTIBLE))
 737                         sleepq_wait(&sx->lock_object, 0);
 738                 else
 739                         error = sleepq_wait_sig(&sx->lock_object, 0);
 740 #ifdef KDTRACE_HOOKS
 741                 sleep_time += lockstat_nsecs(&sx->lock_object);
 742                 sleep_cnt++;
 743 #endif
 744                 if (error) {
 745                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 746                                 CTR2(KTR_LOCK,
 747                         "%s: interruptible sleep by %p suspended by signal",
 748                                     __func__, sx);
 749                         break;
 750                 }
 751                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 752                         CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
 753                             __func__, sx);
 754         }
 755 #ifdef KDTRACE_HOOKS
 756         all_time += lockstat_nsecs(&sx->lock_object);
 757         if (sleep_time)
 758                 LOCKSTAT_RECORD4(sx__block, sx, sleep_time,
 759                     LOCKSTAT_WRITER, (state & SX_LOCK_SHARED) == 0,
 760                     (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
 761         if (lda.spin_cnt > sleep_cnt)
 762                 LOCKSTAT_RECORD4(sx__spin, sx, all_time - sleep_time,
 763                     LOCKSTAT_WRITER, (state & SX_LOCK_SHARED) == 0,
 764                     (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
 765 #endif
 766         if (!error)
 767                 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
 768                     contested, waittime, file, line, LOCKSTAT_WRITER);
 769         GIANT_RESTORE();
 770         return (error);
 771 }
 772
 773 /*
 774  * This function represents the so-called 'hard case' for sx_xunlock
 775  * operation.  All 'easy case' failures are redirected to this.  Note
 776  * that ideally this would be a static function, but it needs to be
 777  * accessible from at least sx.h.
 778  */
 779 void
 780 _sx_xunlock_hard(struct sx *sx, uintptr_t tid, const char *file, int line)
 781 {
 782         uintptr_t x;
 783         int queue, wakeup_swapper;
 784
 785         if (SCHEDULER_STOPPED())
 786                 return;
 787
 788         MPASS(!(sx->sx_lock & SX_LOCK_SHARED));
 789
 790         /* If the lock is recursed, then unrecurse one level. */
 791         if (sx_xlocked(sx) && sx_recursed(sx)) {
 792                 if ((--sx->sx_recurse) == 0)
 793                         atomic_clear_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
 794                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 795                         CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, sx);
 796                 return;
 797         }
 798         MPASS(sx->sx_lock & (SX_LOCK_SHARED_WAITERS |
 799             SX_LOCK_EXCLUSIVE_WAITERS));
 800         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 801                 CTR2(KTR_LOCK, "%s: %p contested", __func__, sx);
 802
 803         sleepq_lock(&sx->lock_object);
 804         x = SX_LOCK_UNLOCKED;
 805
 806         /*
 807          * The wake up algorithm here is quite simple and probably not
 808          * ideal.  It gives precedence to shared waiters if they are
 809          * present.  For this condition, we have to preserve the
 810          * state of the exclusive waiters flag.
 811          * If interruptible sleeps left the shared queue empty avoid a
 812          * starvation for the threads sleeping on the exclusive queue by giving
 813          * them precedence and cleaning up the shared waiters bit anyway.
 814          */
 815         if ((sx->sx_lock & SX_LOCK_SHARED_WAITERS) != 0 &&
 816             sleepq_sleepcnt(&sx->lock_object, SQ_SHARED_QUEUE) != 0) {
 817                 queue = SQ_SHARED_QUEUE;
 818                 x |= (sx->sx_lock & SX_LOCK_EXCLUSIVE_WAITERS);
 819         } else
 820                 queue = SQ_EXCLUSIVE_QUEUE;
 821
 822         /* Wake up all the waiters for the specific queue. */
 823         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 824                 CTR3(KTR_LOCK, "%s: %p waking up all threads on %s queue",
 825                     __func__, sx, queue == SQ_SHARED_QUEUE ? "shared" :
 826                     "exclusive");
 827         atomic_store_rel_ptr(&sx->sx_lock, x);
 828         wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX, 0,
 829             queue);
 830         sleepq_release(&sx->lock_object);
 831         if (wakeup_swapper)
 832                 kick_proc0();
 833 }
 834
 835 /*
 836  * This function represents the so-called 'hard case' for sx_slock
 837  * operation.  All 'easy case' failures are redirected to this.  Note
 838  * that ideally this would be a static function, but it needs to be
 839  * accessible from at least sx.h.
 840  */
 841 int
 842 _sx_slock_hard(struct sx *sx, int opts, const char *file, int line)
 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         uintptr_t x;
 853         int error = 0;
 854 #if defined(ADAPTIVE_SX) || defined(KDTRACE_HOOKS)
 855         struct lock_delay_arg lda;
 856 #endif
 857 #ifdef KDTRACE_HOOKS
 858         uintptr_t state;
 859         u_int sleep_cnt = 0;
 860         int64_t sleep_time = 0;
 861         int64_t all_time = 0;
 862 #endif
 863
 864         if (SCHEDULER_STOPPED())
 865                 return (0);
 866
 867 #if defined(ADAPTIVE_SX)
 868         lock_delay_arg_init(&lda, &sx_delay);
 869 #elif defined(KDTRACE_HOOKS)
 870         lock_delay_arg_init(&lda, NULL);
 871 #endif
 872 #ifdef KDTRACE_HOOKS
 873         state = sx->sx_lock;
 874         all_time -= lockstat_nsecs(&sx->lock_object);
 875 #endif
 876
 877         /*
 878          * As with rwlocks, we don't make any attempt to try to block
 879          * shared locks once there is an exclusive waiter.
 880          */
 881         for (;;) {
 882 #ifdef KDTRACE_HOOKS
 883                 lda.spin_cnt++;
 884 #endif
 885                 x = sx->sx_lock;
 886
 887                 /*
 888                  * If no other thread has an exclusive lock then try to bump up
 889                  * the count of sharers.  Since we have to preserve the state
 890                  * of SX_LOCK_EXCLUSIVE_WAITERS, if we fail to acquire the
 891                  * shared lock loop back and retry.
 892                  */
 893                 if (x & SX_LOCK_SHARED) {
 894                         MPASS(!(x & SX_LOCK_SHARED_WAITERS));
 895                         if (atomic_cmpset_acq_ptr(&sx->sx_lock, x,
 896                             x + SX_ONE_SHARER)) {
 897                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 898                                         CTR4(KTR_LOCK,
 899                                             "%s: %p succeed %p -> %p", __func__,
 900                                             sx, (void *)x,
 901                                             (void *)(x + SX_ONE_SHARER));
 902                                 break;
 903                         }
 904                         continue;
 905                 }
 906 #ifdef HWPMC_HOOKS
 907                 PMC_SOFT_CALL( , , lock, failed);
 908 #endif
 909                 lock_profile_obtain_lock_failed(&sx->lock_object, &contested,
 910                     &waittime);
 911
 912 #ifdef ADAPTIVE_SX
 913                 /*
 914                  * If the owner is running on another CPU, spin until
 915                  * the owner stops running or the state of the lock
 916                  * changes.
 917                  */
 918                 if ((sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) {
 919                         x = SX_OWNER(x);
 920                         owner = (struct thread *)x;
 921                         if (TD_IS_RUNNING(owner)) {
 922                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 923                                         CTR3(KTR_LOCK,
 924                                             "%s: spinning on %p held by %p",
 925                                             __func__, sx, owner);
 926                                 KTR_STATE1(KTR_SCHED, "thread",
 927                                     sched_tdname(curthread), "spinning",
 928                                     "lockname:\"%s\"", sx->lock_object.lo_name);
 929                                 GIANT_SAVE();
 930                                 while (SX_OWNER(sx->sx_lock) == x &&
 931                                     TD_IS_RUNNING(owner))
 932                                         lock_delay(&lda);
 933                                 KTR_STATE0(KTR_SCHED, "thread",
 934                                     sched_tdname(curthread), "running");
 935                                 continue;
 936                         }
 937                 }
 938 #endif
 939
 940                 /*
 941                  * Some other thread already has an exclusive lock, so
 942                  * start the process of blocking.
 943                  */
 944                 sleepq_lock(&sx->lock_object);
 945                 x = sx->sx_lock;
 946
 947                 /*
 948                  * The lock could have been released while we spun.
 949                  * In this case loop back and retry.
 950                  */
 951                 if (x & SX_LOCK_SHARED) {
 952                         sleepq_release(&sx->lock_object);
 953                         continue;
 954                 }
 955
 956 #ifdef ADAPTIVE_SX
 957                 /*
 958                  * If the owner is running on another CPU, spin until
 959                  * the owner stops running or the state of the lock
 960                  * changes.
 961                  */
 962                 if (!(x & SX_LOCK_SHARED) &&
 963                     (sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) {
 964                         owner = (struct thread *)SX_OWNER(x);
 965                         if (TD_IS_RUNNING(owner)) {
 966                                 sleepq_release(&sx->lock_object);
 967                                 continue;
 968                         }
 969                 }
 970 #endif
 971
 972                 /*
 973                  * Try to set the SX_LOCK_SHARED_WAITERS flag.  If we
 974                  * fail to set it drop the sleep queue lock and loop
 975                  * back.
 976                  */
 977                 if (!(x & SX_LOCK_SHARED_WAITERS)) {
 978                         if (!atomic_cmpset_ptr(&sx->sx_lock, x,
 979                             x | SX_LOCK_SHARED_WAITERS)) {
 980                                 sleepq_release(&sx->lock_object);
 981                                 continue;
 982                         }
 983                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 984                                 CTR2(KTR_LOCK, "%s: %p set shared waiters flag",
 985                                     __func__, sx);
 986                 }
 987
 988                 /*
 989                  * Since we have been unable to acquire the shared lock,
 990                  * we have to sleep.
 991                  */
 992                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 993                         CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
 994                             __func__, sx);
 995
 996 #ifdef KDTRACE_HOOKS
 997                 sleep_time -= lockstat_nsecs(&sx->lock_object);
 998 #endif
 999                 GIANT_SAVE();
1000                 sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
1001                     SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
1002                     SLEEPQ_INTERRUPTIBLE : 0), SQ_SHARED_QUEUE);
1003                 if (!(opts & SX_INTERRUPTIBLE))
1004                         sleepq_wait(&sx->lock_object, 0);
1005                 else
1006                         error = sleepq_wait_sig(&sx->lock_object, 0);
1007 #ifdef KDTRACE_HOOKS
1008                 sleep_time += lockstat_nsecs(&sx->lock_object);
1009                 sleep_cnt++;
1010 #endif
1011                 if (error) {
1012                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
1013                                 CTR2(KTR_LOCK,
1014                         "%s: interruptible sleep by %p suspended by signal",
1015                                     __func__, sx);
1016                         break;
1017                 }
1018                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1019                         CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
1020                             __func__, sx);
1021         }
1022 #ifdef KDTRACE_HOOKS
1023         all_time += lockstat_nsecs(&sx->lock_object);
1024         if (sleep_time)
1025                 LOCKSTAT_RECORD4(sx__block, sx, sleep_time,
1026                     LOCKSTAT_READER, (state & SX_LOCK_SHARED) == 0,
1027                     (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
1028         if (lda.spin_cnt > sleep_cnt)
1029                 LOCKSTAT_RECORD4(sx__spin, sx, all_time - sleep_time,
1030                     LOCKSTAT_READER, (state & SX_LOCK_SHARED) == 0,
1031                     (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
1032 #endif
1033         if (error == 0)
1034                 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
1035                     contested, waittime, file, line, LOCKSTAT_READER);
1036         GIANT_RESTORE();
1037         return (error);
1038 }
1039
1040 /*
1041  * This function represents the so-called 'hard case' for sx_sunlock
1042  * operation.  All 'easy case' failures are redirected to this.  Note
1043  * that ideally this would be a static function, but it needs to be
1044  * accessible from at least sx.h.
1045  */
1046 void
1047 _sx_sunlock_hard(struct sx *sx, const char *file, int line)
1048 {
1049         uintptr_t x;
1050         int wakeup_swapper;
1051
1052         if (SCHEDULER_STOPPED())
1053                 return;
1054
1055         for (;;) {
1056                 x = sx->sx_lock;
1057
1058                 /*
1059                  * We should never have sharers while at least one thread
1060                  * holds a shared lock.
1061                  */
1062                 KASSERT(!(x & 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(x) > 1) {
1070                         if (atomic_cmpset_rel_ptr(&sx->sx_lock, x,
1071                             x - 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 *)x,
1076                                             (void *)(x - SX_ONE_SHARER));
1077                                 break;
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 (!(x & SX_LOCK_EXCLUSIVE_WAITERS)) {
1087                         MPASS(x == SX_SHARERS_LOCK(1));
1088                         if (atomic_cmpset_rel_ptr(&sx->sx_lock,
1089                             SX_SHARERS_LOCK(1), SX_LOCK_UNLOCKED)) {
1090                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1091                                         CTR2(KTR_LOCK, "%s: %p last succeeded",
1092                                             __func__, sx);
1093                                 break;
1094                         }
1095                         continue;
1096                 }
1097
1098                 /*
1099                  * At this point, there should just be one sharer with
1100                  * exclusive waiters.
1101                  */
1102                 MPASS(x == (SX_SHARERS_LOCK(1) | SX_LOCK_EXCLUSIVE_WAITERS));
1103
1104                 sleepq_lock(&sx->lock_object);
1105
1106                 /*
1107                  * Wake up semantic here is quite simple:
1108                  * Just wake up all the exclusive waiters.
1109                  * Note that the state of the lock could have changed,
1110                  * so if it fails loop back and retry.
1111                  */
1112                 if (!atomic_cmpset_rel_ptr(&sx->sx_lock,
1113                     SX_SHARERS_LOCK(1) | SX_LOCK_EXCLUSIVE_WAITERS,
1114                     SX_LOCK_UNLOCKED)) {
1115                         sleepq_release(&sx->lock_object);
1116                         continue;
1117                 }
1118                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1119                         CTR2(KTR_LOCK, "%s: %p waking up all thread on"
1120                             "exclusive queue", __func__, sx);
1121                 wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX,
1122                     0, SQ_EXCLUSIVE_QUEUE);
1123                 sleepq_release(&sx->lock_object);
1124                 if (wakeup_swapper)
1125                         kick_proc0();
1126                 break;
1127         }
1128 }
1129
1130 #ifdef INVARIANT_SUPPORT
1131 #ifndef INVARIANTS
1132 #undef  _sx_assert
1133 #endif
1134
1135 /*
1136  * In the non-WITNESS case, sx_assert() can only detect that at least
1137  * *some* thread owns an slock, but it cannot guarantee that *this*
1138  * thread owns an slock.
1139  */
1140 void
1141 _sx_assert(const struct sx *sx, int what, const char *file, int line)
1142 {
1143 #ifndef WITNESS
1144         int slocked = 0;
1145 #endif
1146
1147         if (panicstr != NULL)
1148                 return;
1149         switch (what) {
1150         case SA_SLOCKED:
1151         case SA_SLOCKED | SA_NOTRECURSED:
1152         case SA_SLOCKED | SA_RECURSED:
1153 #ifndef WITNESS
1154                 slocked = 1;
1155                 /* FALLTHROUGH */
1156 #endif
1157         case SA_LOCKED:
1158         case SA_LOCKED | SA_NOTRECURSED:
1159         case SA_LOCKED | SA_RECURSED:
1160 #ifdef WITNESS
1161                 witness_assert(&sx->lock_object, what, file, line);
1162 #else
1163                 /*
1164                  * If some other thread has an exclusive lock or we
1165                  * have one and are asserting a shared lock, fail.
1166                  * Also, if no one has a lock at all, fail.
1167                  */
1168                 if (sx->sx_lock == SX_LOCK_UNLOCKED ||
1169                     (!(sx->sx_lock & SX_LOCK_SHARED) && (slocked ||
1170                     sx_xholder(sx) != curthread)))
1171                         panic("Lock %s not %slocked @ %s:%d\n",
1172                             sx->lock_object.lo_name, slocked ? "share " : "",
1173                             file, line);
1174
1175                 if (!(sx->sx_lock & SX_LOCK_SHARED)) {
1176                         if (sx_recursed(sx)) {
1177                                 if (what & SA_NOTRECURSED)
1178                                         panic("Lock %s recursed @ %s:%d\n",
1179                                             sx->lock_object.lo_name, file,
1180                                             line);
1181                         } else if (what & SA_RECURSED)
1182                                 panic("Lock %s not recursed @ %s:%d\n",
1183                                     sx->lock_object.lo_name, file, line);
1184                 }
1185 #endif
1186                 break;
1187         case SA_XLOCKED:
1188         case SA_XLOCKED | SA_NOTRECURSED:
1189         case SA_XLOCKED | SA_RECURSED:
1190                 if (sx_xholder(sx) != curthread)
1191                         panic("Lock %s not exclusively locked @ %s:%d\n",
1192                             sx->lock_object.lo_name, file, line);
1193                 if (sx_recursed(sx)) {
1194                         if (what & SA_NOTRECURSED)
1195                                 panic("Lock %s recursed @ %s:%d\n",
1196                                     sx->lock_object.lo_name, file, line);
1197                 } else if (what & SA_RECURSED)
1198                         panic("Lock %s not recursed @ %s:%d\n",
1199                             sx->lock_object.lo_name, file, line);
1200                 break;
1201         case SA_UNLOCKED:
1202 #ifdef WITNESS
1203                 witness_assert(&sx->lock_object, what, file, line);
1204 #else
1205                 /*
1206                  * If we hold an exclusve lock fail.  We can't
1207                  * reliably check to see if we hold a shared lock or
1208                  * not.
1209                  */
1210                 if (sx_xholder(sx) == curthread)
1211                         panic("Lock %s exclusively locked @ %s:%d\n",
1212                             sx->lock_object.lo_name, file, line);
1213 #endif
1214                 break;
1215         default:
1216                 panic("Unknown sx lock assertion: %d @ %s:%d", what, file,
1217                     line);
1218         }
1219 }
1220 #endif  /* INVARIANT_SUPPORT */
1221
1222 #ifdef DDB
1223 static void
1224 db_show_sx(const struct lock_object *lock)
1225 {
1226         struct thread *td;
1227         const struct sx *sx;
1228
1229         sx = (const struct sx *)lock;
1230
1231         db_printf(" state: ");
1232         if (sx->sx_lock == SX_LOCK_UNLOCKED)
1233                 db_printf("UNLOCKED\n");
1234         else if (sx->sx_lock == SX_LOCK_DESTROYED) {
1235                 db_printf("DESTROYED\n");
1236                 return;
1237         } else if (sx->sx_lock & SX_LOCK_SHARED)
1238                 db_printf("SLOCK: %ju\n", (uintmax_t)SX_SHARERS(sx->sx_lock));
1239         else {
1240                 td = sx_xholder(sx);
1241                 db_printf("XLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1242                     td->td_tid, td->td_proc->p_pid, td->td_name);
1243                 if (sx_recursed(sx))
1244                         db_printf(" recursed: %d\n", sx->sx_recurse);
1245         }
1246
1247         db_printf(" waiters: ");
1248         switch(sx->sx_lock &
1249             (SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS)) {
1250         case SX_LOCK_SHARED_WAITERS:
1251                 db_printf("shared\n");
1252                 break;
1253         case SX_LOCK_EXCLUSIVE_WAITERS:
1254                 db_printf("exclusive\n");
1255                 break;
1256         case SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS:
1257                 db_printf("exclusive and shared\n");
1258                 break;
1259         default:
1260                 db_printf("none\n");
1261         }
1262 }
1263
1264 /*
1265  * Check to see if a thread that is blocked on a sleep queue is actually
1266  * blocked on an sx lock.  If so, output some details and return true.
1267  * If the lock has an exclusive owner, return that in *ownerp.
1268  */
1269 int
1270 sx_chain(struct thread *td, struct thread **ownerp)
1271 {
1272         struct sx *sx;
1273
1274         /*
1275          * Check to see if this thread is blocked on an sx lock.
1276          * First, we check the lock class.  If that is ok, then we
1277          * compare the lock name against the wait message.
1278          */
1279         sx = td->td_wchan;
1280         if (LOCK_CLASS(&sx->lock_object) != &lock_class_sx ||
1281             sx->lock_object.lo_name != td->td_wmesg)
1282                 return (0);
1283
1284         /* We think we have an sx lock, so output some details. */
1285         db_printf("blocked on sx \"%s\" ", td->td_wmesg);
1286         *ownerp = sx_xholder(sx);
1287         if (sx->sx_lock & SX_LOCK_SHARED)
1288                 db_printf("SLOCK (count %ju)\n",
1289                     (uintmax_t)SX_SHARERS(sx->sx_lock));
1290         else
1291                 db_printf("XLOCK\n");
1292         return (1);
1293 }
1294 #endif