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;
 219         uintptr_t x;
 220
 221         sx = (const struct sx *)lock;
 222         x = sx->sx_lock;
 223         *owner = NULL;
 224         return ((x & SX_LOCK_SHARED) != 0 ? (SX_SHARERS(x) != 0) :
 225             ((*owner = (struct thread *)SX_OWNER(x)) != NULL));
 226 }
 227 #endif
 228
 229 void
 230 sx_sysinit(void *arg)
 231 {
 232         struct sx_args *sargs = arg;
 233
 234         sx_init_flags(sargs->sa_sx, sargs->sa_desc, sargs->sa_flags);
 235 }
 236
 237 void
 238 sx_init_flags(struct sx *sx, const char *description, int opts)
 239 {
 240         int flags;
 241
 242         MPASS((opts & ~(SX_QUIET | SX_RECURSE | SX_NOWITNESS | SX_DUPOK |
 243             SX_NOPROFILE | SX_NOADAPTIVE | SX_NEW)) == 0);
 244         ASSERT_ATOMIC_LOAD_PTR(sx->sx_lock,
 245             ("%s: sx_lock not aligned for %s: %p", __func__, description,
 246             &sx->sx_lock));
 247
 248         flags = LO_SLEEPABLE | LO_UPGRADABLE;
 249         if (opts & SX_DUPOK)
 250                 flags |= LO_DUPOK;
 251         if (opts & SX_NOPROFILE)
 252                 flags |= LO_NOPROFILE;
 253         if (!(opts & SX_NOWITNESS))
 254                 flags |= LO_WITNESS;
 255         if (opts & SX_RECURSE)
 256                 flags |= LO_RECURSABLE;
 257         if (opts & SX_QUIET)
 258                 flags |= LO_QUIET;
 259         if (opts & SX_NEW)
 260                 flags |= LO_NEW;
 261
 262         flags |= opts & SX_NOADAPTIVE;
 263         lock_init(&sx->lock_object, &lock_class_sx, description, NULL, flags);
 264         sx->sx_lock = SX_LOCK_UNLOCKED;
 265         sx->sx_recurse = 0;
 266 }
 267
 268 void
 269 sx_destroy(struct sx *sx)
 270 {
 271
 272         KASSERT(sx->sx_lock == SX_LOCK_UNLOCKED, ("sx lock still held"));
 273         KASSERT(sx->sx_recurse == 0, ("sx lock still recursed"));
 274         sx->sx_lock = SX_LOCK_DESTROYED;
 275         lock_destroy(&sx->lock_object);
 276 }
 277
 278 int
 279 _sx_slock(struct sx *sx, int opts, const char *file, int line)
 280 {
 281         int error = 0;
 282
 283         if (SCHEDULER_STOPPED())
 284                 return (0);
 285         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
 286             ("sx_slock() by idle thread %p on sx %s @ %s:%d",
 287             curthread, sx->lock_object.lo_name, file, line));
 288         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 289             ("sx_slock() of destroyed sx @ %s:%d", file, line));
 290         WITNESS_CHECKORDER(&sx->lock_object, LOP_NEWORDER, file, line, NULL);
 291         error = __sx_slock(sx, opts, file, line);
 292         if (!error) {
 293                 LOCK_LOG_LOCK("SLOCK", &sx->lock_object, 0, 0, file, line);
 294                 WITNESS_LOCK(&sx->lock_object, 0, file, line);
 295                 TD_LOCKS_INC(curthread);
 296         }
 297
 298         return (error);
 299 }
 300
 301 int
 302 sx_try_slock_(struct sx *sx, const char *file, int line)
 303 {
 304         uintptr_t x;
 305
 306         if (SCHEDULER_STOPPED())
 307                 return (1);
 308
 309         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
 310             ("sx_try_slock() by idle thread %p on sx %s @ %s:%d",
 311             curthread, sx->lock_object.lo_name, file, line));
 312
 313         for (;;) {
 314                 x = sx->sx_lock;
 315                 KASSERT(x != SX_LOCK_DESTROYED,
 316                     ("sx_try_slock() of destroyed sx @ %s:%d", file, line));
 317                 if (!(x & SX_LOCK_SHARED))
 318                         break;
 319                 if (atomic_cmpset_acq_ptr(&sx->sx_lock, x, x + SX_ONE_SHARER)) {
 320                         LOCK_LOG_TRY("SLOCK", &sx->lock_object, 0, 1, file, line);
 321                         WITNESS_LOCK(&sx->lock_object, LOP_TRYLOCK, file, line);
 322                         LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire,
 323                             sx, 0, 0, file, line, LOCKSTAT_READER);
 324                         TD_LOCKS_INC(curthread);
 325                         return (1);
 326                 }
 327         }
 328
 329         LOCK_LOG_TRY("SLOCK", &sx->lock_object, 0, 0, file, line);
 330         return (0);
 331 }
 332
 333 int
 334 _sx_xlock(struct sx *sx, int opts, const char *file, int line)
 335 {
 336         int error = 0;
 337
 338         if (SCHEDULER_STOPPED())
 339                 return (0);
 340         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
 341             ("sx_xlock() by idle thread %p on sx %s @ %s:%d",
 342             curthread, sx->lock_object.lo_name, file, line));
 343         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 344             ("sx_xlock() of destroyed sx @ %s:%d", file, line));
 345         WITNESS_CHECKORDER(&sx->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE, file,
 346             line, NULL);
 347         error = __sx_xlock(sx, curthread, opts, file, line);
 348         if (!error) {
 349                 LOCK_LOG_LOCK("XLOCK", &sx->lock_object, 0, sx->sx_recurse,
 350                     file, line);
 351                 WITNESS_LOCK(&sx->lock_object, LOP_EXCLUSIVE, file, line);
 352                 TD_LOCKS_INC(curthread);
 353         }
 354
 355         return (error);
 356 }
 357
 358 int
 359 sx_try_xlock_(struct sx *sx, const char *file, int line)
 360 {
 361         int rval;
 362
 363         if (SCHEDULER_STOPPED())
 364                 return (1);
 365
 366         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
 367             ("sx_try_xlock() by idle thread %p on sx %s @ %s:%d",
 368             curthread, sx->lock_object.lo_name, file, line));
 369         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 370             ("sx_try_xlock() of destroyed sx @ %s:%d", file, line));
 371
 372         if (sx_xlocked(sx) &&
 373             (sx->lock_object.lo_flags & LO_RECURSABLE) != 0) {
 374                 sx->sx_recurse++;
 375                 atomic_set_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
 376                 rval = 1;
 377         } else
 378                 rval = atomic_cmpset_acq_ptr(&sx->sx_lock, SX_LOCK_UNLOCKED,
 379                     (uintptr_t)curthread);
 380         LOCK_LOG_TRY("XLOCK", &sx->lock_object, 0, rval, file, line);
 381         if (rval) {
 382                 WITNESS_LOCK(&sx->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
 383                     file, line);
 384                 if (!sx_recursed(sx))
 385                         LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire,
 386                             sx, 0, 0, file, line, LOCKSTAT_WRITER);
 387                 TD_LOCKS_INC(curthread);
 388         }
 389
 390         return (rval);
 391 }
 392
 393 void
 394 _sx_sunlock(struct sx *sx, const char *file, int line)
 395 {
 396
 397         if (SCHEDULER_STOPPED())
 398                 return;
 399         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 400             ("sx_sunlock() of destroyed sx @ %s:%d", file, line));
 401         _sx_assert(sx, SA_SLOCKED, file, line);
 402         WITNESS_UNLOCK(&sx->lock_object, 0, file, line);
 403         LOCK_LOG_LOCK("SUNLOCK", &sx->lock_object, 0, 0, file, line);
 404         __sx_sunlock(sx, file, line);
 405         TD_LOCKS_DEC(curthread);
 406 }
 407
 408 void
 409 _sx_xunlock(struct sx *sx, const char *file, int line)
 410 {
 411
 412         if (SCHEDULER_STOPPED())
 413                 return;
 414         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 415             ("sx_xunlock() of destroyed sx @ %s:%d", file, line));
 416         _sx_assert(sx, SA_XLOCKED, file, line);
 417         WITNESS_UNLOCK(&sx->lock_object, LOP_EXCLUSIVE, file, line);
 418         LOCK_LOG_LOCK("XUNLOCK", &sx->lock_object, 0, sx->sx_recurse, file,
 419             line);
 420         __sx_xunlock(sx, curthread, file, line);
 421         TD_LOCKS_DEC(curthread);
 422 }
 423
 424 /*
 425  * Try to do a non-blocking upgrade from a shared lock to an exclusive lock.
 426  * This will only succeed if this thread holds a single shared lock.
 427  * Return 1 if if the upgrade succeed, 0 otherwise.
 428  */
 429 int
 430 sx_try_upgrade_(struct sx *sx, const char *file, int line)
 431 {
 432         uintptr_t x;
 433         int success;
 434
 435         if (SCHEDULER_STOPPED())
 436                 return (1);
 437
 438         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 439             ("sx_try_upgrade() of destroyed sx @ %s:%d", file, line));
 440         _sx_assert(sx, SA_SLOCKED, file, line);
 441
 442         /*
 443          * Try to switch from one shared lock to an exclusive lock.  We need
 444          * to maintain the SX_LOCK_EXCLUSIVE_WAITERS flag if set so that
 445          * we will wake up the exclusive waiters when we drop the lock.
 446          */
 447         x = sx->sx_lock & SX_LOCK_EXCLUSIVE_WAITERS;
 448         success = atomic_cmpset_ptr(&sx->sx_lock, SX_SHARERS_LOCK(1) | x,
 449             (uintptr_t)curthread | x);
 450         LOCK_LOG_TRY("XUPGRADE", &sx->lock_object, 0, success, file, line);
 451         if (success) {
 452                 WITNESS_UPGRADE(&sx->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
 453                     file, line);
 454                 LOCKSTAT_RECORD0(sx__upgrade, sx);
 455         }
 456         return (success);
 457 }
 458
 459 /*
 460  * Downgrade an unrecursed exclusive lock into a single shared lock.
 461  */
 462 void
 463 sx_downgrade_(struct sx *sx, const char *file, int line)
 464 {
 465         uintptr_t x;
 466         int wakeup_swapper;
 467
 468         if (SCHEDULER_STOPPED())
 469                 return;
 470
 471         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 472             ("sx_downgrade() of destroyed sx @ %s:%d", file, line));
 473         _sx_assert(sx, SA_XLOCKED | SA_NOTRECURSED, file, line);
 474 #ifndef INVARIANTS
 475         if (sx_recursed(sx))
 476                 panic("downgrade of a recursed lock");
 477 #endif
 478
 479         WITNESS_DOWNGRADE(&sx->lock_object, 0, file, line);
 480
 481         /*
 482          * Try to switch from an exclusive lock with no shared waiters
 483          * to one sharer with no shared waiters.  If there are
 484          * exclusive waiters, we don't need to lock the sleep queue so
 485          * long as we preserve the flag.  We do one quick try and if
 486          * that fails we grab the sleepq lock to keep the flags from
 487          * changing and do it the slow way.
 488          *
 489          * We have to lock the sleep queue if there are shared waiters
 490          * so we can wake them up.
 491          */
 492         x = sx->sx_lock;
 493         if (!(x & SX_LOCK_SHARED_WAITERS) &&
 494             atomic_cmpset_rel_ptr(&sx->sx_lock, x, SX_SHARERS_LOCK(1) |
 495             (x & SX_LOCK_EXCLUSIVE_WAITERS))) {
 496                 LOCK_LOG_LOCK("XDOWNGRADE", &sx->lock_object, 0, 0, file, line);
 497                 return;
 498         }
 499
 500         /*
 501          * Lock the sleep queue so we can read the waiters bits
 502          * without any races and wakeup any shared waiters.
 503          */
 504         sleepq_lock(&sx->lock_object);
 505
 506         /*
 507          * Preserve SX_LOCK_EXCLUSIVE_WAITERS while downgraded to a single
 508          * shared lock.  If there are any shared waiters, wake them up.
 509          */
 510         wakeup_swapper = 0;
 511         x = sx->sx_lock;
 512         atomic_store_rel_ptr(&sx->sx_lock, SX_SHARERS_LOCK(1) |
 513             (x & SX_LOCK_EXCLUSIVE_WAITERS));
 514         if (x & SX_LOCK_SHARED_WAITERS)
 515                 wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX,
 516                     0, SQ_SHARED_QUEUE);
 517         sleepq_release(&sx->lock_object);
 518
 519         LOCK_LOG_LOCK("XDOWNGRADE", &sx->lock_object, 0, 0, file, line);
 520         LOCKSTAT_RECORD0(sx__downgrade, sx);
 521
 522         if (wakeup_swapper)
 523                 kick_proc0();
 524 }
 525
 526 /*
 527  * This function represents the so-called 'hard case' for sx_xlock
 528  * operation.  All 'easy case' failures are redirected to this.  Note
 529  * that ideally this would be a static function, but it needs to be
 530  * accessible from at least sx.h.
 531  */
 532 int
 533 _sx_xlock_hard(struct sx *sx, uintptr_t tid, int opts, const char *file,
 534     int line)
 535 {
 536         GIANT_DECLARE;
 537 #ifdef ADAPTIVE_SX
 538         volatile struct thread *owner;
 539         u_int i, spintries = 0;
 540 #endif
 541         uintptr_t x;
 542 #ifdef LOCK_PROFILING
 543         uint64_t waittime = 0;
 544         int contested = 0;
 545 #endif
 546         int error = 0;
 547 #if defined(ADAPTIVE_SX) || defined(KDTRACE_HOOKS)
 548         struct lock_delay_arg lda;
 549 #endif
 550 #ifdef  KDTRACE_HOOKS
 551         uintptr_t state;
 552         u_int sleep_cnt = 0;
 553         int64_t sleep_time = 0;
 554         int64_t all_time = 0;
 555 #endif
 556
 557         if (SCHEDULER_STOPPED())
 558                 return (0);
 559
 560 #if defined(ADAPTIVE_SX)
 561         lock_delay_arg_init(&lda, &sx_delay);
 562 #elif defined(KDTRACE_HOOKS)
 563         lock_delay_arg_init(&lda, NULL);
 564 #endif
 565
 566         /* If we already hold an exclusive lock, then recurse. */
 567         if (sx_xlocked(sx)) {
 568                 KASSERT((sx->lock_object.lo_flags & LO_RECURSABLE) != 0,
 569             ("_sx_xlock_hard: recursed on non-recursive sx %s @ %s:%d\n",
 570                     sx->lock_object.lo_name, file, line));
 571                 sx->sx_recurse++;
 572                 atomic_set_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
 573                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 574                         CTR2(KTR_LOCK, "%s: %p recursing", __func__, sx);
 575                 return (0);
 576         }
 577
 578         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 579                 CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__,
 580                     sx->lock_object.lo_name, (void *)sx->sx_lock, file, line);
 581
 582 #ifdef KDTRACE_HOOKS
 583         all_time -= lockstat_nsecs(&sx->lock_object);
 584         state = sx->sx_lock;
 585 #endif
 586         for (;;) {
 587                 if (sx->sx_lock == SX_LOCK_UNLOCKED &&
 588                     atomic_cmpset_acq_ptr(&sx->sx_lock, SX_LOCK_UNLOCKED, tid))
 589                         break;
 590 #ifdef KDTRACE_HOOKS
 591                 lda.spin_cnt++;
 592 #endif
 593 #ifdef HWPMC_HOOKS
 594                 PMC_SOFT_CALL( , , lock, failed);
 595 #endif
 596                 lock_profile_obtain_lock_failed(&sx->lock_object, &contested,
 597                     &waittime);
 598 #ifdef ADAPTIVE_SX
 599                 /*
 600                  * If the lock is write locked and the owner is
 601                  * running on another CPU, spin until the owner stops
 602                  * running or the state of the lock changes.
 603                  */
 604                 x = sx->sx_lock;
 605                 if ((sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) {
 606                         if ((x & SX_LOCK_SHARED) == 0) {
 607                                 x = SX_OWNER(x);
 608                                 owner = (struct thread *)x;
 609                                 if (TD_IS_RUNNING(owner)) {
 610                                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 611                                                 CTR3(KTR_LOCK,
 612                                             "%s: spinning on %p held by %p",
 613                                                     __func__, sx, owner);
 614                                         KTR_STATE1(KTR_SCHED, "thread",
 615                                             sched_tdname(curthread), "spinning",
 616                                             "lockname:\"%s\"",
 617                                             sx->lock_object.lo_name);
 618                                         GIANT_SAVE();
 619                                         while (SX_OWNER(sx->sx_lock) == x &&
 620                                             TD_IS_RUNNING(owner))
 621                                                 lock_delay(&lda);
 622                                         KTR_STATE0(KTR_SCHED, "thread",
 623                                             sched_tdname(curthread), "running");
 624                                         continue;
 625                                 }
 626                         } else if (SX_SHARERS(x) && spintries < asx_retries) {
 627                                 KTR_STATE1(KTR_SCHED, "thread",
 628                                     sched_tdname(curthread), "spinning",
 629                                     "lockname:\"%s\"", sx->lock_object.lo_name);
 630                                 GIANT_SAVE();
 631                                 spintries++;
 632                                 for (i = 0; i < asx_loops; i++) {
 633                                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 634                                                 CTR4(KTR_LOCK,
 635                                     "%s: shared spinning on %p with %u and %u",
 636                                                     __func__, sx, spintries, i);
 637                                         x = sx->sx_lock;
 638                                         if ((x & SX_LOCK_SHARED) == 0 ||
 639                                             SX_SHARERS(x) == 0)
 640                                                 break;
 641                                         cpu_spinwait();
 642 #ifdef KDTRACE_HOOKS
 643                                         lda.spin_cnt++;
 644 #endif
 645                                 }
 646                                 KTR_STATE0(KTR_SCHED, "thread",
 647                                     sched_tdname(curthread), "running");
 648                                 if (i != asx_loops)
 649                                         continue;
 650                         }
 651                 }
 652 #endif
 653
 654                 sleepq_lock(&sx->lock_object);
 655                 x = sx->sx_lock;
 656
 657                 /*
 658                  * If the lock was released while spinning on the
 659                  * sleep queue chain lock, try again.
 660                  */
 661                 if (x == SX_LOCK_UNLOCKED) {
 662                         sleepq_release(&sx->lock_object);
 663                         continue;
 664                 }
 665
 666 #ifdef ADAPTIVE_SX
 667                 /*
 668                  * The current lock owner might have started executing
 669                  * on another CPU (or the lock could have changed
 670                  * owners) while we were waiting on the sleep queue
 671                  * chain lock.  If so, drop the sleep queue lock and try
 672                  * again.
 673                  */
 674                 if (!(x & SX_LOCK_SHARED) &&
 675                     (sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) {
 676                         owner = (struct thread *)SX_OWNER(x);
 677                         if (TD_IS_RUNNING(owner)) {
 678                                 sleepq_release(&sx->lock_object);
 679                                 continue;
 680                         }
 681                 }
 682 #endif
 683
 684                 /*
 685                  * If an exclusive lock was released with both shared
 686                  * and exclusive waiters and a shared waiter hasn't
 687                  * woken up and acquired the lock yet, sx_lock will be
 688                  * set to SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS.
 689                  * If we see that value, try to acquire it once.  Note
 690                  * that we have to preserve SX_LOCK_EXCLUSIVE_WAITERS
 691                  * as there are other exclusive waiters still.  If we
 692                  * fail, restart the loop.
 693                  */
 694                 if (x == (SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS)) {
 695                         if (atomic_cmpset_acq_ptr(&sx->sx_lock,
 696                             SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS,
 697                             tid | SX_LOCK_EXCLUSIVE_WAITERS)) {
 698                                 sleepq_release(&sx->lock_object);
 699                                 CTR2(KTR_LOCK, "%s: %p claimed by new writer",
 700                                     __func__, sx);
 701                                 break;
 702                         }
 703                         sleepq_release(&sx->lock_object);
 704                         continue;
 705                 }
 706
 707                 /*
 708                  * Try to set the SX_LOCK_EXCLUSIVE_WAITERS.  If we fail,
 709                  * than loop back and retry.
 710                  */
 711                 if (!(x & SX_LOCK_EXCLUSIVE_WAITERS)) {
 712                         if (!atomic_cmpset_ptr(&sx->sx_lock, x,
 713                             x | SX_LOCK_EXCLUSIVE_WAITERS)) {
 714                                 sleepq_release(&sx->lock_object);
 715                                 continue;
 716                         }
 717                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 718                                 CTR2(KTR_LOCK, "%s: %p set excl waiters flag",
 719                                     __func__, sx);
 720                 }
 721
 722                 /*
 723                  * Since we have been unable to acquire the exclusive
 724                  * lock and the exclusive waiters flag is set, we have
 725                  * to sleep.
 726                  */
 727                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 728                         CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
 729                             __func__, sx);
 730
 731 #ifdef KDTRACE_HOOKS
 732                 sleep_time -= lockstat_nsecs(&sx->lock_object);
 733 #endif
 734                 GIANT_SAVE();
 735                 sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
 736                     SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
 737                     SLEEPQ_INTERRUPTIBLE : 0), SQ_EXCLUSIVE_QUEUE);
 738                 if (!(opts & SX_INTERRUPTIBLE))
 739                         sleepq_wait(&sx->lock_object, 0);
 740                 else
 741                         error = sleepq_wait_sig(&sx->lock_object, 0);
 742 #ifdef KDTRACE_HOOKS
 743                 sleep_time += lockstat_nsecs(&sx->lock_object);
 744                 sleep_cnt++;
 745 #endif
 746                 if (error) {
 747                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 748                                 CTR2(KTR_LOCK,
 749                         "%s: interruptible sleep by %p suspended by signal",
 750                                     __func__, sx);
 751                         break;
 752                 }
 753                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 754                         CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
 755                             __func__, sx);
 756         }
 757 #ifdef KDTRACE_HOOKS
 758         all_time += lockstat_nsecs(&sx->lock_object);
 759         if (sleep_time)
 760                 LOCKSTAT_RECORD4(sx__block, sx, sleep_time,
 761                     LOCKSTAT_WRITER, (state & SX_LOCK_SHARED) == 0,
 762                     (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
 763         if (lda.spin_cnt > sleep_cnt)
 764                 LOCKSTAT_RECORD4(sx__spin, sx, all_time - sleep_time,
 765                     LOCKSTAT_WRITER, (state & SX_LOCK_SHARED) == 0,
 766                     (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
 767 #endif
 768         if (!error)
 769                 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
 770                     contested, waittime, file, line, LOCKSTAT_WRITER);
 771         GIANT_RESTORE();
 772         return (error);
 773 }
 774
 775 /*
 776  * This function represents the so-called 'hard case' for sx_xunlock
 777  * operation.  All 'easy case' failures are redirected to this.  Note
 778  * that ideally this would be a static function, but it needs to be
 779  * accessible from at least sx.h.
 780  */
 781 void
 782 _sx_xunlock_hard(struct sx *sx, uintptr_t tid, const char *file, int line)
 783 {
 784         uintptr_t x;
 785         int queue, wakeup_swapper;
 786
 787         if (SCHEDULER_STOPPED())
 788                 return;
 789
 790         MPASS(!(sx->sx_lock & SX_LOCK_SHARED));
 791
 792         /* If the lock is recursed, then unrecurse one level. */
 793         if (sx_xlocked(sx) && sx_recursed(sx)) {
 794                 if ((--sx->sx_recurse) == 0)
 795                         atomic_clear_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
 796                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 797                         CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, sx);
 798                 return;
 799         }
 800         MPASS(sx->sx_lock & (SX_LOCK_SHARED_WAITERS |
 801             SX_LOCK_EXCLUSIVE_WAITERS));
 802         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 803                 CTR2(KTR_LOCK, "%s: %p contested", __func__, sx);
 804
 805         sleepq_lock(&sx->lock_object);
 806         x = SX_LOCK_UNLOCKED;
 807
 808         /*
 809          * The wake up algorithm here is quite simple and probably not
 810          * ideal.  It gives precedence to shared waiters if they are
 811          * present.  For this condition, we have to preserve the
 812          * state of the exclusive waiters flag.
 813          * If interruptible sleeps left the shared queue empty avoid a
 814          * starvation for the threads sleeping on the exclusive queue by giving
 815          * them precedence and cleaning up the shared waiters bit anyway.
 816          */
 817         if ((sx->sx_lock & SX_LOCK_SHARED_WAITERS) != 0 &&
 818             sleepq_sleepcnt(&sx->lock_object, SQ_SHARED_QUEUE) != 0) {
 819                 queue = SQ_SHARED_QUEUE;
 820                 x |= (sx->sx_lock & SX_LOCK_EXCLUSIVE_WAITERS);
 821         } else
 822                 queue = SQ_EXCLUSIVE_QUEUE;
 823
 824         /* Wake up all the waiters for the specific queue. */
 825         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 826                 CTR3(KTR_LOCK, "%s: %p waking up all threads on %s queue",
 827                     __func__, sx, queue == SQ_SHARED_QUEUE ? "shared" :
 828                     "exclusive");
 829         atomic_store_rel_ptr(&sx->sx_lock, x);
 830         wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX, 0,
 831             queue);
 832         sleepq_release(&sx->lock_object);
 833         if (wakeup_swapper)
 834                 kick_proc0();
 835 }
 836
 837 /*
 838  * This function represents the so-called 'hard case' for sx_slock
 839  * operation.  All 'easy case' failures are redirected to this.  Note
 840  * that ideally this would be a static function, but it needs to be
 841  * accessible from at least sx.h.
 842  */
 843 int
 844 _sx_slock_hard(struct sx *sx, int opts, const char *file, int line)
 845 {
 846         GIANT_DECLARE;
 847 #ifdef ADAPTIVE_SX
 848         volatile struct thread *owner;
 849 #endif
 850 #ifdef LOCK_PROFILING
 851         uint64_t waittime = 0;
 852         int contested = 0;
 853 #endif
 854         uintptr_t x;
 855         int error = 0;
 856 #if defined(ADAPTIVE_SX) || defined(KDTRACE_HOOKS)
 857         struct lock_delay_arg lda;
 858 #endif
 859 #ifdef KDTRACE_HOOKS
 860         uintptr_t state;
 861         u_int sleep_cnt = 0;
 862         int64_t sleep_time = 0;
 863         int64_t all_time = 0;
 864 #endif
 865
 866         if (SCHEDULER_STOPPED())
 867                 return (0);
 868
 869 #if defined(ADAPTIVE_SX)
 870         lock_delay_arg_init(&lda, &sx_delay);
 871 #elif defined(KDTRACE_HOOKS)
 872         lock_delay_arg_init(&lda, NULL);
 873 #endif
 874 #ifdef KDTRACE_HOOKS
 875         state = sx->sx_lock;
 876         all_time -= lockstat_nsecs(&sx->lock_object);
 877 #endif
 878
 879         /*
 880          * As with rwlocks, we don't make any attempt to try to block
 881          * shared locks once there is an exclusive waiter.
 882          */
 883         for (;;) {
 884 #ifdef KDTRACE_HOOKS
 885                 lda.spin_cnt++;
 886 #endif
 887                 x = sx->sx_lock;
 888
 889                 /*
 890                  * If no other thread has an exclusive lock then try to bump up
 891                  * the count of sharers.  Since we have to preserve the state
 892                  * of SX_LOCK_EXCLUSIVE_WAITERS, if we fail to acquire the
 893                  * shared lock loop back and retry.
 894                  */
 895                 if (x & SX_LOCK_SHARED) {
 896                         MPASS(!(x & SX_LOCK_SHARED_WAITERS));
 897                         if (atomic_cmpset_acq_ptr(&sx->sx_lock, x,
 898                             x + SX_ONE_SHARER)) {
 899                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 900                                         CTR4(KTR_LOCK,
 901                                             "%s: %p succeed %p -> %p", __func__,
 902                                             sx, (void *)x,
 903                                             (void *)(x + SX_ONE_SHARER));
 904                                 break;
 905                         }
 906                         continue;
 907                 }
 908 #ifdef HWPMC_HOOKS
 909                 PMC_SOFT_CALL( , , lock, failed);
 910 #endif
 911                 lock_profile_obtain_lock_failed(&sx->lock_object, &contested,
 912                     &waittime);
 913
 914 #ifdef ADAPTIVE_SX
 915                 /*
 916                  * If the owner is running on another CPU, spin until
 917                  * the owner stops running or the state of the lock
 918                  * changes.
 919                  */
 920                 if ((sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) {
 921                         x = SX_OWNER(x);
 922                         owner = (struct thread *)x;
 923                         if (TD_IS_RUNNING(owner)) {
 924                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 925                                         CTR3(KTR_LOCK,
 926                                             "%s: spinning on %p held by %p",
 927                                             __func__, sx, owner);
 928                                 KTR_STATE1(KTR_SCHED, "thread",
 929                                     sched_tdname(curthread), "spinning",
 930                                     "lockname:\"%s\"", sx->lock_object.lo_name);
 931                                 GIANT_SAVE();
 932                                 while (SX_OWNER(sx->sx_lock) == x &&
 933                                     TD_IS_RUNNING(owner))
 934                                         lock_delay(&lda);
 935                                 KTR_STATE0(KTR_SCHED, "thread",
 936                                     sched_tdname(curthread), "running");
 937                                 continue;
 938                         }
 939                 }
 940 #endif
 941
 942                 /*
 943                  * Some other thread already has an exclusive lock, so
 944                  * start the process of blocking.
 945                  */
 946                 sleepq_lock(&sx->lock_object);
 947                 x = sx->sx_lock;
 948
 949                 /*
 950                  * The lock could have been released while we spun.
 951                  * In this case loop back and retry.
 952                  */
 953                 if (x & SX_LOCK_SHARED) {
 954                         sleepq_release(&sx->lock_object);
 955                         continue;
 956                 }
 957
 958 #ifdef ADAPTIVE_SX
 959                 /*
 960                  * If the owner is running on another CPU, spin until
 961                  * the owner stops running or the state of the lock
 962                  * changes.
 963                  */
 964                 if (!(x & SX_LOCK_SHARED) &&
 965                     (sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) {
 966                         owner = (struct thread *)SX_OWNER(x);
 967                         if (TD_IS_RUNNING(owner)) {
 968                                 sleepq_release(&sx->lock_object);
 969                                 continue;
 970                         }
 971                 }
 972 #endif
 973
 974                 /*
 975                  * Try to set the SX_LOCK_SHARED_WAITERS flag.  If we
 976                  * fail to set it drop the sleep queue lock and loop
 977                  * back.
 978                  */
 979                 if (!(x & SX_LOCK_SHARED_WAITERS)) {
 980                         if (!atomic_cmpset_ptr(&sx->sx_lock, x,
 981                             x | SX_LOCK_SHARED_WAITERS)) {
 982                                 sleepq_release(&sx->lock_object);
 983                                 continue;
 984                         }
 985                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 986                                 CTR2(KTR_LOCK, "%s: %p set shared waiters flag",
 987                                     __func__, sx);
 988                 }
 989
 990                 /*
 991                  * Since we have been unable to acquire the shared lock,
 992                  * we have to sleep.
 993                  */
 994                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 995                         CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
 996                             __func__, sx);
 997
 998 #ifdef KDTRACE_HOOKS
 999                 sleep_time -= lockstat_nsecs(&sx->lock_object);
1000 #endif
1001                 GIANT_SAVE();
1002                 sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
1003                     SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
1004                     SLEEPQ_INTERRUPTIBLE : 0), SQ_SHARED_QUEUE);
1005                 if (!(opts & SX_INTERRUPTIBLE))
1006                         sleepq_wait(&sx->lock_object, 0);
1007                 else
1008                         error = sleepq_wait_sig(&sx->lock_object, 0);
1009 #ifdef KDTRACE_HOOKS
1010                 sleep_time += lockstat_nsecs(&sx->lock_object);
1011                 sleep_cnt++;
1012 #endif
1013                 if (error) {
1014                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
1015                                 CTR2(KTR_LOCK,
1016                         "%s: interruptible sleep by %p suspended by signal",
1017                                     __func__, sx);
1018                         break;
1019                 }
1020                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1021                         CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
1022                             __func__, sx);
1023         }
1024 #ifdef KDTRACE_HOOKS
1025         all_time += lockstat_nsecs(&sx->lock_object);
1026         if (sleep_time)
1027                 LOCKSTAT_RECORD4(sx__block, sx, sleep_time,
1028                     LOCKSTAT_READER, (state & SX_LOCK_SHARED) == 0,
1029                     (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
1030         if (lda.spin_cnt > sleep_cnt)
1031                 LOCKSTAT_RECORD4(sx__spin, sx, all_time - sleep_time,
1032                     LOCKSTAT_READER, (state & SX_LOCK_SHARED) == 0,
1033                     (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
1034 #endif
1035         if (error == 0)
1036                 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
1037                     contested, waittime, file, line, LOCKSTAT_READER);
1038         GIANT_RESTORE();
1039         return (error);
1040 }
1041
1042 /*
1043  * This function represents the so-called 'hard case' for sx_sunlock
1044  * operation.  All 'easy case' failures are redirected to this.  Note
1045  * that ideally this would be a static function, but it needs to be
1046  * accessible from at least sx.h.
1047  */
1048 void
1049 _sx_sunlock_hard(struct sx *sx, const char *file, int line)
1050 {
1051         uintptr_t x;
1052         int wakeup_swapper;
1053
1054         if (SCHEDULER_STOPPED())
1055                 return;
1056
1057         for (;;) {
1058                 x = sx->sx_lock;
1059
1060                 /*
1061                  * We should never have sharers while at least one thread
1062                  * holds a shared lock.
1063                  */
1064                 KASSERT(!(x & SX_LOCK_SHARED_WAITERS),
1065                     ("%s: waiting sharers", __func__));
1066
1067                 /*
1068                  * See if there is more than one shared lock held.  If
1069                  * so, just drop one and return.
1070                  */
1071                 if (SX_SHARERS(x) > 1) {
1072                         if (atomic_cmpset_rel_ptr(&sx->sx_lock, x,
1073                             x - SX_ONE_SHARER)) {
1074                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1075                                         CTR4(KTR_LOCK,
1076                                             "%s: %p succeeded %p -> %p",
1077                                             __func__, sx, (void *)x,
1078                                             (void *)(x - SX_ONE_SHARER));
1079                                 break;
1080                         }
1081                         continue;
1082                 }
1083
1084                 /*
1085                  * If there aren't any waiters for an exclusive lock,
1086                  * then try to drop it quickly.
1087                  */
1088                 if (!(x & SX_LOCK_EXCLUSIVE_WAITERS)) {
1089                         MPASS(x == SX_SHARERS_LOCK(1));
1090                         if (atomic_cmpset_rel_ptr(&sx->sx_lock,
1091                             SX_SHARERS_LOCK(1), SX_LOCK_UNLOCKED)) {
1092                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1093                                         CTR2(KTR_LOCK, "%s: %p last succeeded",
1094                                             __func__, sx);
1095                                 break;
1096                         }
1097                         continue;
1098                 }
1099
1100                 /*
1101                  * At this point, there should just be one sharer with
1102                  * exclusive waiters.
1103                  */
1104                 MPASS(x == (SX_SHARERS_LOCK(1) | SX_LOCK_EXCLUSIVE_WAITERS));
1105
1106                 sleepq_lock(&sx->lock_object);
1107
1108                 /*
1109                  * Wake up semantic here is quite simple:
1110                  * Just wake up all the exclusive waiters.
1111                  * Note that the state of the lock could have changed,
1112                  * so if it fails loop back and retry.
1113                  */
1114                 if (!atomic_cmpset_rel_ptr(&sx->sx_lock,
1115                     SX_SHARERS_LOCK(1) | SX_LOCK_EXCLUSIVE_WAITERS,
1116                     SX_LOCK_UNLOCKED)) {
1117                         sleepq_release(&sx->lock_object);
1118                         continue;
1119                 }
1120                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1121                         CTR2(KTR_LOCK, "%s: %p waking up all thread on"
1122                             "exclusive queue", __func__, sx);
1123                 wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX,
1124                     0, SQ_EXCLUSIVE_QUEUE);
1125                 sleepq_release(&sx->lock_object);
1126                 if (wakeup_swapper)
1127                         kick_proc0();
1128                 break;
1129         }
1130 }
1131
1132 #ifdef INVARIANT_SUPPORT
1133 #ifndef INVARIANTS
1134 #undef  _sx_assert
1135 #endif
1136
1137 /*
1138  * In the non-WITNESS case, sx_assert() can only detect that at least
1139  * *some* thread owns an slock, but it cannot guarantee that *this*
1140  * thread owns an slock.
1141  */
1142 void
1143 _sx_assert(const struct sx *sx, int what, const char *file, int line)
1144 {
1145 #ifndef WITNESS
1146         int slocked = 0;
1147 #endif
1148
1149         if (panicstr != NULL)
1150                 return;
1151         switch (what) {
1152         case SA_SLOCKED:
1153         case SA_SLOCKED | SA_NOTRECURSED:
1154         case SA_SLOCKED | SA_RECURSED:
1155 #ifndef WITNESS
1156                 slocked = 1;
1157                 /* FALLTHROUGH */
1158 #endif
1159         case SA_LOCKED:
1160         case SA_LOCKED | SA_NOTRECURSED:
1161         case SA_LOCKED | SA_RECURSED:
1162 #ifdef WITNESS
1163                 witness_assert(&sx->lock_object, what, file, line);
1164 #else
1165                 /*
1166                  * If some other thread has an exclusive lock or we
1167                  * have one and are asserting a shared lock, fail.
1168                  * Also, if no one has a lock at all, fail.
1169                  */
1170                 if (sx->sx_lock == SX_LOCK_UNLOCKED ||
1171                     (!(sx->sx_lock & SX_LOCK_SHARED) && (slocked ||
1172                     sx_xholder(sx) != curthread)))
1173                         panic("Lock %s not %slocked @ %s:%d\n",
1174                             sx->lock_object.lo_name, slocked ? "share " : "",
1175                             file, line);
1176
1177                 if (!(sx->sx_lock & SX_LOCK_SHARED)) {
1178                         if (sx_recursed(sx)) {
1179                                 if (what & SA_NOTRECURSED)
1180                                         panic("Lock %s recursed @ %s:%d\n",
1181                                             sx->lock_object.lo_name, file,
1182                                             line);
1183                         } else if (what & SA_RECURSED)
1184                                 panic("Lock %s not recursed @ %s:%d\n",
1185                                     sx->lock_object.lo_name, file, line);
1186                 }
1187 #endif
1188                 break;
1189         case SA_XLOCKED:
1190         case SA_XLOCKED | SA_NOTRECURSED:
1191         case SA_XLOCKED | SA_RECURSED:
1192                 if (sx_xholder(sx) != curthread)
1193                         panic("Lock %s not exclusively locked @ %s:%d\n",
1194                             sx->lock_object.lo_name, file, line);
1195                 if (sx_recursed(sx)) {
1196                         if (what & SA_NOTRECURSED)
1197                                 panic("Lock %s recursed @ %s:%d\n",
1198                                     sx->lock_object.lo_name, file, line);
1199                 } else if (what & SA_RECURSED)
1200                         panic("Lock %s not recursed @ %s:%d\n",
1201                             sx->lock_object.lo_name, file, line);
1202                 break;
1203         case SA_UNLOCKED:
1204 #ifdef WITNESS
1205                 witness_assert(&sx->lock_object, what, file, line);
1206 #else
1207                 /*
1208                  * If we hold an exclusve lock fail.  We can't
1209                  * reliably check to see if we hold a shared lock or
1210                  * not.
1211                  */
1212                 if (sx_xholder(sx) == curthread)
1213                         panic("Lock %s exclusively locked @ %s:%d\n",
1214                             sx->lock_object.lo_name, file, line);
1215 #endif
1216                 break;
1217         default:
1218                 panic("Unknown sx lock assertion: %d @ %s:%d", what, file,
1219                     line);
1220         }
1221 }
1222 #endif  /* INVARIANT_SUPPORT */
1223
1224 #ifdef DDB
1225 static void
1226 db_show_sx(const struct lock_object *lock)
1227 {
1228         struct thread *td;
1229         const struct sx *sx;
1230
1231         sx = (const struct sx *)lock;
1232
1233         db_printf(" state: ");
1234         if (sx->sx_lock == SX_LOCK_UNLOCKED)
1235                 db_printf("UNLOCKED\n");
1236         else if (sx->sx_lock == SX_LOCK_DESTROYED) {
1237                 db_printf("DESTROYED\n");
1238                 return;
1239         } else if (sx->sx_lock & SX_LOCK_SHARED)
1240                 db_printf("SLOCK: %ju\n", (uintmax_t)SX_SHARERS(sx->sx_lock));
1241         else {
1242                 td = sx_xholder(sx);
1243                 db_printf("XLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1244                     td->td_tid, td->td_proc->p_pid, td->td_name);
1245                 if (sx_recursed(sx))
1246                         db_printf(" recursed: %d\n", sx->sx_recurse);
1247         }
1248
1249         db_printf(" waiters: ");
1250         switch(sx->sx_lock &
1251             (SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS)) {
1252         case SX_LOCK_SHARED_WAITERS:
1253                 db_printf("shared\n");
1254                 break;
1255         case SX_LOCK_EXCLUSIVE_WAITERS:
1256                 db_printf("exclusive\n");
1257                 break;
1258         case SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS:
1259                 db_printf("exclusive and shared\n");
1260                 break;
1261         default:
1262                 db_printf("none\n");
1263         }
1264 }
1265
1266 /*
1267  * Check to see if a thread that is blocked on a sleep queue is actually
1268  * blocked on an sx lock.  If so, output some details and return true.
1269  * If the lock has an exclusive owner, return that in *ownerp.
1270  */
1271 int
1272 sx_chain(struct thread *td, struct thread **ownerp)
1273 {
1274         struct sx *sx;
1275
1276         /*
1277          * Check to see if this thread is blocked on an sx lock.
1278          * First, we check the lock class.  If that is ok, then we
1279          * compare the lock name against the wait message.
1280          */
1281         sx = td->td_wchan;
1282         if (LOCK_CLASS(&sx->lock_object) != &lock_class_sx ||
1283             sx->lock_object.lo_name != td->td_wmesg)
1284                 return (0);
1285
1286         /* We think we have an sx lock, so output some details. */
1287         db_printf("blocked on sx \"%s\" ", td->td_wmesg);
1288         *ownerp = sx_xholder(sx);
1289         if (sx->sx_lock & SX_LOCK_SHARED)
1290                 db_printf("SLOCK (count %ju)\n",
1291                     (uintmax_t)SX_SHARERS(sx->sx_lock));
1292         else
1293                 db_printf("XLOCK\n");
1294         return (1);
1295 }
1296 #endif