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 x, uintptr_t tid, int opts,
 534     const char *file, int line)
 535 {
 536         GIANT_DECLARE;
 537 #ifdef ADAPTIVE_SX
 538         volatile struct thread *owner;
 539         u_int i, spintries = 0;
 540 #endif
 541 #ifdef LOCK_PROFILING
 542         uint64_t waittime = 0;
 543         int contested = 0;
 544 #endif
 545         int error = 0;
 546 #if defined(ADAPTIVE_SX) || defined(KDTRACE_HOOKS)
 547         struct lock_delay_arg lda;
 548 #endif
 549 #ifdef  KDTRACE_HOOKS
 550         uintptr_t state;
 551         u_int sleep_cnt = 0;
 552         int64_t sleep_time = 0;
 553         int64_t all_time = 0;
 554 #endif
 555
 556         if (SCHEDULER_STOPPED())
 557                 return (0);
 558
 559 #if defined(ADAPTIVE_SX)
 560         lock_delay_arg_init(&lda, &sx_delay);
 561 #elif defined(KDTRACE_HOOKS)
 562         lock_delay_arg_init(&lda, NULL);
 563 #endif
 564
 565         /* If we already hold an exclusive lock, then recurse. */
 566         if (__predict_false(lv_sx_owner(x) == (struct thread *)tid)) {
 567                 KASSERT((sx->lock_object.lo_flags & LO_RECURSABLE) != 0,
 568             ("_sx_xlock_hard: recursed on non-recursive sx %s @ %s:%d\n",
 569                     sx->lock_object.lo_name, file, line));
 570                 sx->sx_recurse++;
 571                 atomic_set_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
 572                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 573                         CTR2(KTR_LOCK, "%s: %p recursing", __func__, sx);
 574                 return (0);
 575         }
 576
 577         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 578                 CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__,
 579                     sx->lock_object.lo_name, (void *)sx->sx_lock, file, line);
 580
 581 #ifdef KDTRACE_HOOKS
 582         all_time -= lockstat_nsecs(&sx->lock_object);
 583         state = x;
 584 #endif
 585         for (;;) {
 586                 if (x == SX_LOCK_UNLOCKED) {
 587                         if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid))
 588                                 break;
 589                         continue;
 590                 }
 591 #ifdef KDTRACE_HOOKS
 592                 lda.spin_cnt++;
 593 #endif
 594 #ifdef HWPMC_HOOKS
 595                 PMC_SOFT_CALL( , , lock, failed);
 596 #endif
 597                 lock_profile_obtain_lock_failed(&sx->lock_object, &contested,
 598                     &waittime);
 599 #ifdef ADAPTIVE_SX
 600                 /*
 601                  * If the lock is write locked and the owner is
 602                  * running on another CPU, spin until the owner stops
 603                  * running or the state of the lock changes.
 604                  */
 605                 if ((sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) {
 606                         if ((x & SX_LOCK_SHARED) == 0) {
 607                                 owner = lv_sx_owner(x);
 608                                 if (TD_IS_RUNNING(owner)) {
 609                                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 610                                                 CTR3(KTR_LOCK,
 611                                             "%s: spinning on %p held by %p",
 612                                                     __func__, sx, owner);
 613                                         KTR_STATE1(KTR_SCHED, "thread",
 614                                             sched_tdname(curthread), "spinning",
 615                                             "lockname:\"%s\"",
 616                                             sx->lock_object.lo_name);
 617                                         GIANT_SAVE();
 618                                         do {
 619                                                 lock_delay(&lda);
 620                                                 x = SX_READ_VALUE(sx);
 621                                                 owner = lv_sx_owner(x);
 622                                         } while (owner != NULL &&
 623                                                     TD_IS_RUNNING(owner));
 624                                         KTR_STATE0(KTR_SCHED, "thread",
 625                                             sched_tdname(curthread), "running");
 626                                         continue;
 627                                 }
 628                         } else if (SX_SHARERS(x) && spintries < asx_retries) {
 629                                 KTR_STATE1(KTR_SCHED, "thread",
 630                                     sched_tdname(curthread), "spinning",
 631                                     "lockname:\"%s\"", sx->lock_object.lo_name);
 632                                 GIANT_SAVE();
 633                                 spintries++;
 634                                 for (i = 0; i < asx_loops; i++) {
 635                                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 636                                                 CTR4(KTR_LOCK,
 637                                     "%s: shared spinning on %p with %u and %u",
 638                                                     __func__, sx, spintries, i);
 639                                         x = sx->sx_lock;
 640                                         if ((x & SX_LOCK_SHARED) == 0 ||
 641                                             SX_SHARERS(x) == 0)
 642                                                 break;
 643                                         cpu_spinwait();
 644 #ifdef KDTRACE_HOOKS
 645                                         lda.spin_cnt++;
 646 #endif
 647                                 }
 648                                 KTR_STATE0(KTR_SCHED, "thread",
 649                                     sched_tdname(curthread), "running");
 650                                 x = SX_READ_VALUE(sx);
 651                                 if (i != asx_loops)
 652                                         continue;
 653                         }
 654                 }
 655 #endif
 656
 657                 sleepq_lock(&sx->lock_object);
 658                 x = SX_READ_VALUE(sx);
 659
 660                 /*
 661                  * If the lock was released while spinning on the
 662                  * sleep queue chain lock, try again.
 663                  */
 664                 if (x == SX_LOCK_UNLOCKED) {
 665                         sleepq_release(&sx->lock_object);
 666                         continue;
 667                 }
 668
 669 #ifdef ADAPTIVE_SX
 670                 /*
 671                  * The current lock owner might have started executing
 672                  * on another CPU (or the lock could have changed
 673                  * owners) while we were waiting on the sleep queue
 674                  * chain lock.  If so, drop the sleep queue lock and try
 675                  * again.
 676                  */
 677                 if (!(x & SX_LOCK_SHARED) &&
 678                     (sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) {
 679                         owner = (struct thread *)SX_OWNER(x);
 680                         if (TD_IS_RUNNING(owner)) {
 681                                 sleepq_release(&sx->lock_object);
 682                                 continue;
 683                         }
 684                 }
 685 #endif
 686
 687                 /*
 688                  * If an exclusive lock was released with both shared
 689                  * and exclusive waiters and a shared waiter hasn't
 690                  * woken up and acquired the lock yet, sx_lock will be
 691                  * set to SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS.
 692                  * If we see that value, try to acquire it once.  Note
 693                  * that we have to preserve SX_LOCK_EXCLUSIVE_WAITERS
 694                  * as there are other exclusive waiters still.  If we
 695                  * fail, restart the loop.
 696                  */
 697                 if (x == (SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS)) {
 698                         if (atomic_cmpset_acq_ptr(&sx->sx_lock,
 699                             SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS,
 700                             tid | SX_LOCK_EXCLUSIVE_WAITERS)) {
 701                                 sleepq_release(&sx->lock_object);
 702                                 CTR2(KTR_LOCK, "%s: %p claimed by new writer",
 703                                     __func__, sx);
 704                                 break;
 705                         }
 706                         sleepq_release(&sx->lock_object);
 707                         x = SX_READ_VALUE(sx);
 708                         continue;
 709                 }
 710
 711                 /*
 712                  * Try to set the SX_LOCK_EXCLUSIVE_WAITERS.  If we fail,
 713                  * than loop back and retry.
 714                  */
 715                 if (!(x & SX_LOCK_EXCLUSIVE_WAITERS)) {
 716                         if (!atomic_cmpset_ptr(&sx->sx_lock, x,
 717                             x | SX_LOCK_EXCLUSIVE_WAITERS)) {
 718                                 sleepq_release(&sx->lock_object);
 719                                 x = SX_READ_VALUE(sx);
 720                                 continue;
 721                         }
 722                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 723                                 CTR2(KTR_LOCK, "%s: %p set excl waiters flag",
 724                                     __func__, sx);
 725                 }
 726
 727                 /*
 728                  * Since we have been unable to acquire the exclusive
 729                  * lock and the exclusive waiters flag is set, we have
 730                  * to sleep.
 731                  */
 732                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 733                         CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
 734                             __func__, sx);
 735
 736 #ifdef KDTRACE_HOOKS
 737                 sleep_time -= lockstat_nsecs(&sx->lock_object);
 738 #endif
 739                 GIANT_SAVE();
 740                 sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
 741                     SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
 742                     SLEEPQ_INTERRUPTIBLE : 0), SQ_EXCLUSIVE_QUEUE);
 743                 if (!(opts & SX_INTERRUPTIBLE))
 744                         sleepq_wait(&sx->lock_object, 0);
 745                 else
 746                         error = sleepq_wait_sig(&sx->lock_object, 0);
 747 #ifdef KDTRACE_HOOKS
 748                 sleep_time += lockstat_nsecs(&sx->lock_object);
 749                 sleep_cnt++;
 750 #endif
 751                 if (error) {
 752                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 753                                 CTR2(KTR_LOCK,
 754                         "%s: interruptible sleep by %p suspended by signal",
 755                                     __func__, sx);
 756                         break;
 757                 }
 758                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 759                         CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
 760                             __func__, sx);
 761                 x = SX_READ_VALUE(sx);
 762         }
 763 #ifdef KDTRACE_HOOKS
 764         all_time += lockstat_nsecs(&sx->lock_object);
 765         if (sleep_time)
 766                 LOCKSTAT_RECORD4(sx__block, sx, sleep_time,
 767                     LOCKSTAT_WRITER, (state & SX_LOCK_SHARED) == 0,
 768                     (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
 769         if (lda.spin_cnt > sleep_cnt)
 770                 LOCKSTAT_RECORD4(sx__spin, sx, all_time - sleep_time,
 771                     LOCKSTAT_WRITER, (state & SX_LOCK_SHARED) == 0,
 772                     (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
 773 #endif
 774         if (!error)
 775                 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
 776                     contested, waittime, file, line, LOCKSTAT_WRITER);
 777         GIANT_RESTORE();
 778         return (error);
 779 }
 780
 781 /*
 782  * This function represents the so-called 'hard case' for sx_xunlock
 783  * operation.  All 'easy case' failures are redirected to this.  Note
 784  * that ideally this would be a static function, but it needs to be
 785  * accessible from at least sx.h.
 786  */
 787 void
 788 _sx_xunlock_hard(struct sx *sx, uintptr_t tid, const char *file, int line)
 789 {
 790         uintptr_t x;
 791         int queue, wakeup_swapper;
 792
 793         if (SCHEDULER_STOPPED())
 794                 return;
 795
 796         MPASS(!(sx->sx_lock & SX_LOCK_SHARED));
 797
 798         /* If the lock is recursed, then unrecurse one level. */
 799         if (sx_xlocked(sx) && sx_recursed(sx)) {
 800                 if ((--sx->sx_recurse) == 0)
 801                         atomic_clear_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
 802                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 803                         CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, sx);
 804                 return;
 805         }
 806         MPASS(sx->sx_lock & (SX_LOCK_SHARED_WAITERS |
 807             SX_LOCK_EXCLUSIVE_WAITERS));
 808         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 809                 CTR2(KTR_LOCK, "%s: %p contested", __func__, sx);
 810
 811         sleepq_lock(&sx->lock_object);
 812         x = SX_LOCK_UNLOCKED;
 813
 814         /*
 815          * The wake up algorithm here is quite simple and probably not
 816          * ideal.  It gives precedence to shared waiters if they are
 817          * present.  For this condition, we have to preserve the
 818          * state of the exclusive waiters flag.
 819          * If interruptible sleeps left the shared queue empty avoid a
 820          * starvation for the threads sleeping on the exclusive queue by giving
 821          * them precedence and cleaning up the shared waiters bit anyway.
 822          */
 823         if ((sx->sx_lock & SX_LOCK_SHARED_WAITERS) != 0 &&
 824             sleepq_sleepcnt(&sx->lock_object, SQ_SHARED_QUEUE) != 0) {
 825                 queue = SQ_SHARED_QUEUE;
 826                 x |= (sx->sx_lock & SX_LOCK_EXCLUSIVE_WAITERS);
 827         } else
 828                 queue = SQ_EXCLUSIVE_QUEUE;
 829
 830         /* Wake up all the waiters for the specific queue. */
 831         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 832                 CTR3(KTR_LOCK, "%s: %p waking up all threads on %s queue",
 833                     __func__, sx, queue == SQ_SHARED_QUEUE ? "shared" :
 834                     "exclusive");
 835         atomic_store_rel_ptr(&sx->sx_lock, x);
 836         wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX, 0,
 837             queue);
 838         sleepq_release(&sx->lock_object);
 839         if (wakeup_swapper)
 840                 kick_proc0();
 841 }
 842
 843 /*
 844  * This function represents the so-called 'hard case' for sx_slock
 845  * operation.  All 'easy case' failures are redirected to this.  Note
 846  * that ideally this would be a static function, but it needs to be
 847  * accessible from at least sx.h.
 848  */
 849 int
 850 _sx_slock_hard(struct sx *sx, int opts, const char *file, int line)
 851 {
 852         GIANT_DECLARE;
 853 #ifdef ADAPTIVE_SX
 854         volatile struct thread *owner;
 855 #endif
 856 #ifdef LOCK_PROFILING
 857         uint64_t waittime = 0;
 858         int contested = 0;
 859 #endif
 860         uintptr_t x;
 861         int error = 0;
 862 #if defined(ADAPTIVE_SX) || defined(KDTRACE_HOOKS)
 863         struct lock_delay_arg lda;
 864 #endif
 865 #ifdef KDTRACE_HOOKS
 866         uintptr_t state;
 867         u_int sleep_cnt = 0;
 868         int64_t sleep_time = 0;
 869         int64_t all_time = 0;
 870 #endif
 871
 872         if (SCHEDULER_STOPPED())
 873                 return (0);
 874
 875 #if defined(ADAPTIVE_SX)
 876         lock_delay_arg_init(&lda, &sx_delay);
 877 #elif defined(KDTRACE_HOOKS)
 878         lock_delay_arg_init(&lda, NULL);
 879 #endif
 880 #ifdef KDTRACE_HOOKS
 881         all_time -= lockstat_nsecs(&sx->lock_object);
 882 #endif
 883         x = SX_READ_VALUE(sx);
 884 #ifdef KDTRACE_HOOKS
 885         state = x;
 886 #endif
 887
 888         /*
 889          * As with rwlocks, we don't make any attempt to try to block
 890          * shared locks once there is an exclusive waiter.
 891          */
 892         for (;;) {
 893                 /*
 894                  * If no other thread has an exclusive lock then try to bump up
 895                  * the count of sharers.  Since we have to preserve the state
 896                  * of SX_LOCK_EXCLUSIVE_WAITERS, if we fail to acquire the
 897                  * shared lock loop back and retry.
 898                  */
 899                 if (x & SX_LOCK_SHARED) {
 900                         MPASS(!(x & SX_LOCK_SHARED_WAITERS));
 901                         if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x,
 902                             x + SX_ONE_SHARER)) {
 903                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 904                                         CTR4(KTR_LOCK,
 905                                             "%s: %p succeed %p -> %p", __func__,
 906                                             sx, (void *)x,
 907                                             (void *)(x + SX_ONE_SHARER));
 908                                 break;
 909                         }
 910                         continue;
 911                 }
 912 #ifdef KDTRACE_HOOKS
 913                 lda.spin_cnt++;
 914 #endif
 915
 916 #ifdef HWPMC_HOOKS
 917                 PMC_SOFT_CALL( , , lock, failed);
 918 #endif
 919                 lock_profile_obtain_lock_failed(&sx->lock_object, &contested,
 920                     &waittime);
 921
 922 #ifdef ADAPTIVE_SX
 923                 /*
 924                  * If the owner is running on another CPU, spin until
 925                  * the owner stops running or the state of the lock
 926                  * changes.
 927                  */
 928                 if ((sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) {
 929                         owner = lv_sx_owner(x);
 930                         if (TD_IS_RUNNING(owner)) {
 931                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 932                                         CTR3(KTR_LOCK,
 933                                             "%s: spinning on %p held by %p",
 934                                             __func__, sx, owner);
 935                                 KTR_STATE1(KTR_SCHED, "thread",
 936                                     sched_tdname(curthread), "spinning",
 937                                     "lockname:\"%s\"", sx->lock_object.lo_name);
 938                                 GIANT_SAVE();
 939                                 do {
 940                                         lock_delay(&lda);
 941                                         x = SX_READ_VALUE(sx);
 942                                         owner = lv_sx_owner(x);
 943                                 } while (owner != NULL && TD_IS_RUNNING(owner));
 944                                 KTR_STATE0(KTR_SCHED, "thread",
 945                                     sched_tdname(curthread), "running");
 946                                 continue;
 947                         }
 948                 }
 949 #endif
 950
 951                 /*
 952                  * Some other thread already has an exclusive lock, so
 953                  * start the process of blocking.
 954                  */
 955                 sleepq_lock(&sx->lock_object);
 956                 x = SX_READ_VALUE(sx);
 957
 958                 /*
 959                  * The lock could have been released while we spun.
 960                  * In this case loop back and retry.
 961                  */
 962                 if (x & SX_LOCK_SHARED) {
 963                         sleepq_release(&sx->lock_object);
 964                         continue;
 965                 }
 966
 967 #ifdef ADAPTIVE_SX
 968                 /*
 969                  * If the owner is running on another CPU, spin until
 970                  * the owner stops running or the state of the lock
 971                  * changes.
 972                  */
 973                 if (!(x & SX_LOCK_SHARED) &&
 974                     (sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) {
 975                         owner = (struct thread *)SX_OWNER(x);
 976                         if (TD_IS_RUNNING(owner)) {
 977                                 sleepq_release(&sx->lock_object);
 978                                 x = SX_READ_VALUE(sx);
 979                                 continue;
 980                         }
 981                 }
 982 #endif
 983
 984                 /*
 985                  * Try to set the SX_LOCK_SHARED_WAITERS flag.  If we
 986                  * fail to set it drop the sleep queue lock and loop
 987                  * back.
 988                  */
 989                 if (!(x & SX_LOCK_SHARED_WAITERS)) {
 990                         if (!atomic_cmpset_ptr(&sx->sx_lock, x,
 991                             x | SX_LOCK_SHARED_WAITERS)) {
 992                                 sleepq_release(&sx->lock_object);
 993                                 x = SX_READ_VALUE(sx);
 994                                 continue;
 995                         }
 996                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 997                                 CTR2(KTR_LOCK, "%s: %p set shared waiters flag",
 998                                     __func__, sx);
 999                 }
1000
1001                 /*
1002                  * Since we have been unable to acquire the shared lock,
1003                  * we have to sleep.
1004                  */
1005                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1006                         CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
1007                             __func__, sx);
1008
1009 #ifdef KDTRACE_HOOKS
1010                 sleep_time -= lockstat_nsecs(&sx->lock_object);
1011 #endif
1012                 GIANT_SAVE();
1013                 sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
1014                     SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
1015                     SLEEPQ_INTERRUPTIBLE : 0), SQ_SHARED_QUEUE);
1016                 if (!(opts & SX_INTERRUPTIBLE))
1017                         sleepq_wait(&sx->lock_object, 0);
1018                 else
1019                         error = sleepq_wait_sig(&sx->lock_object, 0);
1020 #ifdef KDTRACE_HOOKS
1021                 sleep_time += lockstat_nsecs(&sx->lock_object);
1022                 sleep_cnt++;
1023 #endif
1024                 if (error) {
1025                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
1026                                 CTR2(KTR_LOCK,
1027                         "%s: interruptible sleep by %p suspended by signal",
1028                                     __func__, sx);
1029                         break;
1030                 }
1031                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1032                         CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
1033                             __func__, sx);
1034                 x = SX_READ_VALUE(sx);
1035         }
1036 #ifdef KDTRACE_HOOKS
1037         all_time += lockstat_nsecs(&sx->lock_object);
1038         if (sleep_time)
1039                 LOCKSTAT_RECORD4(sx__block, sx, sleep_time,
1040                     LOCKSTAT_READER, (state & SX_LOCK_SHARED) == 0,
1041                     (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
1042         if (lda.spin_cnt > sleep_cnt)
1043                 LOCKSTAT_RECORD4(sx__spin, sx, all_time - sleep_time,
1044                     LOCKSTAT_READER, (state & SX_LOCK_SHARED) == 0,
1045                     (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
1046 #endif
1047         if (error == 0)
1048                 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
1049                     contested, waittime, file, line, LOCKSTAT_READER);
1050         GIANT_RESTORE();
1051         return (error);
1052 }
1053
1054 /*
1055  * This function represents the so-called 'hard case' for sx_sunlock
1056  * operation.  All 'easy case' failures are redirected to this.  Note
1057  * that ideally this would be a static function, but it needs to be
1058  * accessible from at least sx.h.
1059  */
1060 void
1061 _sx_sunlock_hard(struct sx *sx, const char *file, int line)
1062 {
1063         uintptr_t x;
1064         int wakeup_swapper;
1065
1066         if (SCHEDULER_STOPPED())
1067                 return;
1068
1069         x = SX_READ_VALUE(sx);
1070         for (;;) {
1071                 /*
1072                  * We should never have sharers while at least one thread
1073                  * holds a shared lock.
1074                  */
1075                 KASSERT(!(x & SX_LOCK_SHARED_WAITERS),
1076                     ("%s: waiting sharers", __func__));
1077
1078                 /*
1079                  * See if there is more than one shared lock held.  If
1080                  * so, just drop one and return.
1081                  */
1082                 if (SX_SHARERS(x) > 1) {
1083                         if (atomic_fcmpset_rel_ptr(&sx->sx_lock, &x,
1084                             x - SX_ONE_SHARER)) {
1085                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1086                                         CTR4(KTR_LOCK,
1087                                             "%s: %p succeeded %p -> %p",
1088                                             __func__, sx, (void *)x,
1089                                             (void *)(x - SX_ONE_SHARER));
1090                                 break;
1091                         }
1092                         continue;
1093                 }
1094
1095                 /*
1096                  * If there aren't any waiters for an exclusive lock,
1097                  * then try to drop it quickly.
1098                  */
1099                 if (!(x & SX_LOCK_EXCLUSIVE_WAITERS)) {
1100                         MPASS(x == SX_SHARERS_LOCK(1));
1101                         x = SX_SHARERS_LOCK(1);
1102                         if (atomic_fcmpset_rel_ptr(&sx->sx_lock,
1103                             &x, SX_LOCK_UNLOCKED)) {
1104                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1105                                         CTR2(KTR_LOCK, "%s: %p last succeeded",
1106                                             __func__, sx);
1107                                 break;
1108                         }
1109                         continue;
1110                 }
1111
1112                 /*
1113                  * At this point, there should just be one sharer with
1114                  * exclusive waiters.
1115                  */
1116                 MPASS(x == (SX_SHARERS_LOCK(1) | SX_LOCK_EXCLUSIVE_WAITERS));
1117
1118                 sleepq_lock(&sx->lock_object);
1119
1120                 /*
1121                  * Wake up semantic here is quite simple:
1122                  * Just wake up all the exclusive waiters.
1123                  * Note that the state of the lock could have changed,
1124                  * so if it fails loop back and retry.
1125                  */
1126                 if (!atomic_cmpset_rel_ptr(&sx->sx_lock,
1127                     SX_SHARERS_LOCK(1) | SX_LOCK_EXCLUSIVE_WAITERS,
1128                     SX_LOCK_UNLOCKED)) {
1129                         sleepq_release(&sx->lock_object);
1130                         x = SX_READ_VALUE(sx);
1131                         continue;
1132                 }
1133                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1134                         CTR2(KTR_LOCK, "%s: %p waking up all thread on"
1135                             "exclusive queue", __func__, sx);
1136                 wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX,
1137                     0, SQ_EXCLUSIVE_QUEUE);
1138                 sleepq_release(&sx->lock_object);
1139                 if (wakeup_swapper)
1140                         kick_proc0();
1141                 break;
1142         }
1143 }
1144
1145 #ifdef INVARIANT_SUPPORT
1146 #ifndef INVARIANTS
1147 #undef  _sx_assert
1148 #endif
1149
1150 /*
1151  * In the non-WITNESS case, sx_assert() can only detect that at least
1152  * *some* thread owns an slock, but it cannot guarantee that *this*
1153  * thread owns an slock.
1154  */
1155 void
1156 _sx_assert(const struct sx *sx, int what, const char *file, int line)
1157 {
1158 #ifndef WITNESS
1159         int slocked = 0;
1160 #endif
1161
1162         if (panicstr != NULL)
1163                 return;
1164         switch (what) {
1165         case SA_SLOCKED:
1166         case SA_SLOCKED | SA_NOTRECURSED:
1167         case SA_SLOCKED | SA_RECURSED:
1168 #ifndef WITNESS
1169                 slocked = 1;
1170                 /* FALLTHROUGH */
1171 #endif
1172         case SA_LOCKED:
1173         case SA_LOCKED | SA_NOTRECURSED:
1174         case SA_LOCKED | SA_RECURSED:
1175 #ifdef WITNESS
1176                 witness_assert(&sx->lock_object, what, file, line);
1177 #else
1178                 /*
1179                  * If some other thread has an exclusive lock or we
1180                  * have one and are asserting a shared lock, fail.
1181                  * Also, if no one has a lock at all, fail.
1182                  */
1183                 if (sx->sx_lock == SX_LOCK_UNLOCKED ||
1184                     (!(sx->sx_lock & SX_LOCK_SHARED) && (slocked ||
1185                     sx_xholder(sx) != curthread)))
1186                         panic("Lock %s not %slocked @ %s:%d\n",
1187                             sx->lock_object.lo_name, slocked ? "share " : "",
1188                             file, line);
1189
1190                 if (!(sx->sx_lock & SX_LOCK_SHARED)) {
1191                         if (sx_recursed(sx)) {
1192                                 if (what & SA_NOTRECURSED)
1193                                         panic("Lock %s recursed @ %s:%d\n",
1194                                             sx->lock_object.lo_name, file,
1195                                             line);
1196                         } else if (what & SA_RECURSED)
1197                                 panic("Lock %s not recursed @ %s:%d\n",
1198                                     sx->lock_object.lo_name, file, line);
1199                 }
1200 #endif
1201                 break;
1202         case SA_XLOCKED:
1203         case SA_XLOCKED | SA_NOTRECURSED:
1204         case SA_XLOCKED | SA_RECURSED:
1205                 if (sx_xholder(sx) != curthread)
1206                         panic("Lock %s not exclusively locked @ %s:%d\n",
1207                             sx->lock_object.lo_name, file, line);
1208                 if (sx_recursed(sx)) {
1209                         if (what & SA_NOTRECURSED)
1210                                 panic("Lock %s recursed @ %s:%d\n",
1211                                     sx->lock_object.lo_name, file, line);
1212                 } else if (what & SA_RECURSED)
1213                         panic("Lock %s not recursed @ %s:%d\n",
1214                             sx->lock_object.lo_name, file, line);
1215                 break;
1216         case SA_UNLOCKED:
1217 #ifdef WITNESS
1218                 witness_assert(&sx->lock_object, what, file, line);
1219 #else
1220                 /*
1221                  * If we hold an exclusve lock fail.  We can't
1222                  * reliably check to see if we hold a shared lock or
1223                  * not.
1224                  */
1225                 if (sx_xholder(sx) == curthread)
1226                         panic("Lock %s exclusively locked @ %s:%d\n",
1227                             sx->lock_object.lo_name, file, line);
1228 #endif
1229                 break;
1230         default:
1231                 panic("Unknown sx lock assertion: %d @ %s:%d", what, file,
1232                     line);
1233         }
1234 }
1235 #endif  /* INVARIANT_SUPPORT */
1236
1237 #ifdef DDB
1238 static void
1239 db_show_sx(const struct lock_object *lock)
1240 {
1241         struct thread *td;
1242         const struct sx *sx;
1243
1244         sx = (const struct sx *)lock;
1245
1246         db_printf(" state: ");
1247         if (sx->sx_lock == SX_LOCK_UNLOCKED)
1248                 db_printf("UNLOCKED\n");
1249         else if (sx->sx_lock == SX_LOCK_DESTROYED) {
1250                 db_printf("DESTROYED\n");
1251                 return;
1252         } else if (sx->sx_lock & SX_LOCK_SHARED)
1253                 db_printf("SLOCK: %ju\n", (uintmax_t)SX_SHARERS(sx->sx_lock));
1254         else {
1255                 td = sx_xholder(sx);
1256                 db_printf("XLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1257                     td->td_tid, td->td_proc->p_pid, td->td_name);
1258                 if (sx_recursed(sx))
1259                         db_printf(" recursed: %d\n", sx->sx_recurse);
1260         }
1261
1262         db_printf(" waiters: ");
1263         switch(sx->sx_lock &
1264             (SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS)) {
1265         case SX_LOCK_SHARED_WAITERS:
1266                 db_printf("shared\n");
1267                 break;
1268         case SX_LOCK_EXCLUSIVE_WAITERS:
1269                 db_printf("exclusive\n");
1270                 break;
1271         case SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS:
1272                 db_printf("exclusive and shared\n");
1273                 break;
1274         default:
1275                 db_printf("none\n");
1276         }
1277 }
1278
1279 /*
1280  * Check to see if a thread that is blocked on a sleep queue is actually
1281  * blocked on an sx lock.  If so, output some details and return true.
1282  * If the lock has an exclusive owner, return that in *ownerp.
1283  */
1284 int
1285 sx_chain(struct thread *td, struct thread **ownerp)
1286 {
1287         struct sx *sx;
1288
1289         /*
1290          * Check to see if this thread is blocked on an sx lock.
1291          * First, we check the lock class.  If that is ok, then we
1292          * compare the lock name against the wait message.
1293          */
1294         sx = td->td_wchan;
1295         if (LOCK_CLASS(&sx->lock_object) != &lock_class_sx ||
1296             sx->lock_object.lo_name != td->td_wmesg)
1297                 return (0);
1298
1299         /* We think we have an sx lock, so output some details. */
1300         db_printf("blocked on sx \"%s\" ", td->td_wmesg);
1301         *ownerp = sx_xholder(sx);
1302         if (sx->sx_lock & SX_LOCK_SHARED)
1303                 db_printf("SLOCK (count %ju)\n",
1304                     (uintmax_t)SX_SHARERS(sx->sx_lock));
1305         else
1306                 db_printf("XLOCK\n");
1307         return (1);
1308 }
1309 #endif