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(work) do {                                           \
  92         if (__predict_false(mtx_owned(&Giant))) {                       \
  93                 work++;                                                 \
  94                 WITNESS_SAVE(&Giant.lock_object, Giant);                \
  95                 while (mtx_owned(&Giant)) {                             \
  96                         _giantcnt++;                                    \
  97                         mtx_unlock(&Giant);                             \
  98                 }                                                       \
  99         }                                                               \
 100 } while (0)
 101
 102 #define GIANT_RESTORE() do {                                            \
 103         if (_giantcnt > 0) {                                            \
 104                 mtx_assert(&Giant, MA_NOTOWNED);                        \
 105                 while (_giantcnt--)                                     \
 106                         mtx_lock(&Giant);                               \
 107                 WITNESS_RESTORE(&Giant.lock_object, Giant);             \
 108         }                                                               \
 109 } while (0)
 110
 111 /*
 112  * Returns true if an exclusive lock is recursed.  It assumes
 113  * curthread currently has an exclusive lock.
 114  */
 115 #define sx_recursed(sx)         ((sx)->sx_recurse != 0)
 116
 117 static void     assert_sx(const struct lock_object *lock, int what);
 118 #ifdef DDB
 119 static void     db_show_sx(const struct lock_object *lock);
 120 #endif
 121 static void     lock_sx(struct lock_object *lock, uintptr_t how);
 122 #ifdef KDTRACE_HOOKS
 123 static int      owner_sx(const struct lock_object *lock, struct thread **owner);
 124 #endif
 125 static uintptr_t unlock_sx(struct lock_object *lock);
 126
 127 struct lock_class lock_class_sx = {
 128         .lc_name = "sx",
 129         .lc_flags = LC_SLEEPLOCK | LC_SLEEPABLE | LC_RECURSABLE | LC_UPGRADABLE,
 130         .lc_assert = assert_sx,
 131 #ifdef DDB
 132         .lc_ddb_show = db_show_sx,
 133 #endif
 134         .lc_lock = lock_sx,
 135         .lc_unlock = unlock_sx,
 136 #ifdef KDTRACE_HOOKS
 137         .lc_owner = owner_sx,
 138 #endif
 139 };
 140
 141 #ifndef INVARIANTS
 142 #define _sx_assert(sx, what, file, line)
 143 #endif
 144
 145 #ifdef ADAPTIVE_SX
 146 static __read_frequently u_int asx_retries;
 147 static __read_frequently u_int asx_loops;
 148 static SYSCTL_NODE(_debug, OID_AUTO, sx, CTLFLAG_RD, NULL, "sxlock debugging");
 149 SYSCTL_UINT(_debug_sx, OID_AUTO, retries, CTLFLAG_RW, &asx_retries, 0, "");
 150 SYSCTL_UINT(_debug_sx, OID_AUTO, loops, CTLFLAG_RW, &asx_loops, 0, "");
 151
 152 static struct lock_delay_config __read_frequently sx_delay;
 153
 154 SYSCTL_INT(_debug_sx, OID_AUTO, delay_base, CTLFLAG_RW, &sx_delay.base,
 155     0, "");
 156 SYSCTL_INT(_debug_sx, OID_AUTO, delay_max, CTLFLAG_RW, &sx_delay.max,
 157     0, "");
 158
 159 static void
 160 sx_lock_delay_init(void *arg __unused)
 161 {
 162
 163         lock_delay_default_init(&sx_delay);
 164         asx_retries = 10;
 165         asx_loops = max(10000, sx_delay.max);
 166 }
 167 LOCK_DELAY_SYSINIT(sx_lock_delay_init);
 168 #endif
 169
 170 void
 171 assert_sx(const struct lock_object *lock, int what)
 172 {
 173
 174         sx_assert((const struct sx *)lock, what);
 175 }
 176
 177 void
 178 lock_sx(struct lock_object *lock, uintptr_t how)
 179 {
 180         struct sx *sx;
 181
 182         sx = (struct sx *)lock;
 183         if (how)
 184                 sx_slock(sx);
 185         else
 186                 sx_xlock(sx);
 187 }
 188
 189 uintptr_t
 190 unlock_sx(struct lock_object *lock)
 191 {
 192         struct sx *sx;
 193
 194         sx = (struct sx *)lock;
 195         sx_assert(sx, SA_LOCKED | SA_NOTRECURSED);
 196         if (sx_xlocked(sx)) {
 197                 sx_xunlock(sx);
 198                 return (0);
 199         } else {
 200                 sx_sunlock(sx);
 201                 return (1);
 202         }
 203 }
 204
 205 #ifdef KDTRACE_HOOKS
 206 int
 207 owner_sx(const struct lock_object *lock, struct thread **owner)
 208 {
 209         const struct sx *sx;
 210         uintptr_t x;
 211
 212         sx = (const struct sx *)lock;
 213         x = sx->sx_lock;
 214         *owner = NULL;
 215         return ((x & SX_LOCK_SHARED) != 0 ? (SX_SHARERS(x) != 0) :
 216             ((*owner = (struct thread *)SX_OWNER(x)) != NULL));
 217 }
 218 #endif
 219
 220 void
 221 sx_sysinit(void *arg)
 222 {
 223         struct sx_args *sargs = arg;
 224
 225         sx_init_flags(sargs->sa_sx, sargs->sa_desc, sargs->sa_flags);
 226 }
 227
 228 void
 229 sx_init_flags(struct sx *sx, const char *description, int opts)
 230 {
 231         int flags;
 232
 233         MPASS((opts & ~(SX_QUIET | SX_RECURSE | SX_NOWITNESS | SX_DUPOK |
 234             SX_NOPROFILE | SX_NOADAPTIVE | SX_NEW)) == 0);
 235         ASSERT_ATOMIC_LOAD_PTR(sx->sx_lock,
 236             ("%s: sx_lock not aligned for %s: %p", __func__, description,
 237             &sx->sx_lock));
 238
 239         flags = LO_SLEEPABLE | LO_UPGRADABLE;
 240         if (opts & SX_DUPOK)
 241                 flags |= LO_DUPOK;
 242         if (opts & SX_NOPROFILE)
 243                 flags |= LO_NOPROFILE;
 244         if (!(opts & SX_NOWITNESS))
 245                 flags |= LO_WITNESS;
 246         if (opts & SX_RECURSE)
 247                 flags |= LO_RECURSABLE;
 248         if (opts & SX_QUIET)
 249                 flags |= LO_QUIET;
 250         if (opts & SX_NEW)
 251                 flags |= LO_NEW;
 252
 253         flags |= opts & SX_NOADAPTIVE;
 254         lock_init(&sx->lock_object, &lock_class_sx, description, NULL, flags);
 255         sx->sx_lock = SX_LOCK_UNLOCKED;
 256         sx->sx_recurse = 0;
 257 }
 258
 259 void
 260 sx_destroy(struct sx *sx)
 261 {
 262
 263         KASSERT(sx->sx_lock == SX_LOCK_UNLOCKED, ("sx lock still held"));
 264         KASSERT(sx->sx_recurse == 0, ("sx lock still recursed"));
 265         sx->sx_lock = SX_LOCK_DESTROYED;
 266         lock_destroy(&sx->lock_object);
 267 }
 268
 269 int
 270 sx_try_slock_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
 271 {
 272         uintptr_t x;
 273
 274         if (SCHEDULER_STOPPED())
 275                 return (1);
 276
 277         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
 278             ("sx_try_slock() by idle thread %p on sx %s @ %s:%d",
 279             curthread, sx->lock_object.lo_name, file, line));
 280
 281         x = sx->sx_lock;
 282         for (;;) {
 283                 KASSERT(x != SX_LOCK_DESTROYED,
 284                     ("sx_try_slock() of destroyed sx @ %s:%d", file, line));
 285                 if (!(x & SX_LOCK_SHARED))
 286                         break;
 287                 if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, x + SX_ONE_SHARER)) {
 288                         LOCK_LOG_TRY("SLOCK", &sx->lock_object, 0, 1, file, line);
 289                         WITNESS_LOCK(&sx->lock_object, LOP_TRYLOCK, file, line);
 290                         LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire,
 291                             sx, 0, 0, file, line, LOCKSTAT_READER);
 292                         TD_LOCKS_INC(curthread);
 293                         return (1);
 294                 }
 295         }
 296
 297         LOCK_LOG_TRY("SLOCK", &sx->lock_object, 0, 0, file, line);
 298         return (0);
 299 }
 300
 301 int
 302 sx_try_slock_(struct sx *sx, const char *file, int line)
 303 {
 304
 305         return (sx_try_slock_int(sx LOCK_FILE_LINE_ARG));
 306 }
 307
 308 int
 309 _sx_xlock(struct sx *sx, int opts, const char *file, int line)
 310 {
 311         uintptr_t tid, x;
 312         int error = 0;
 313
 314         KASSERT(kdb_active != 0 || SCHEDULER_STOPPED() ||
 315             !TD_IS_IDLETHREAD(curthread),
 316             ("sx_xlock() by idle thread %p on sx %s @ %s:%d",
 317             curthread, sx->lock_object.lo_name, file, line));
 318         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 319             ("sx_xlock() of destroyed sx @ %s:%d", file, line));
 320         WITNESS_CHECKORDER(&sx->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE, file,
 321             line, NULL);
 322         tid = (uintptr_t)curthread;
 323         x = SX_LOCK_UNLOCKED;
 324         if (!atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid))
 325                 error = _sx_xlock_hard(sx, x, opts LOCK_FILE_LINE_ARG);
 326         else
 327                 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
 328                     0, 0, file, line, LOCKSTAT_WRITER);
 329         if (!error) {
 330                 LOCK_LOG_LOCK("XLOCK", &sx->lock_object, 0, sx->sx_recurse,
 331                     file, line);
 332                 WITNESS_LOCK(&sx->lock_object, LOP_EXCLUSIVE, file, line);
 333                 TD_LOCKS_INC(curthread);
 334         }
 335
 336         return (error);
 337 }
 338
 339 int
 340 sx_try_xlock_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
 341 {
 342         struct thread *td;
 343         uintptr_t tid, x;
 344         int rval;
 345         bool recursed;
 346
 347         td = curthread;
 348         tid = (uintptr_t)td;
 349         if (SCHEDULER_STOPPED_TD(td))
 350                 return (1);
 351
 352         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(td),
 353             ("sx_try_xlock() by idle thread %p on sx %s @ %s:%d",
 354             curthread, sx->lock_object.lo_name, file, line));
 355         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 356             ("sx_try_xlock() of destroyed sx @ %s:%d", file, line));
 357
 358         rval = 1;
 359         recursed = false;
 360         x = SX_LOCK_UNLOCKED;
 361         for (;;) {
 362                 if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid))
 363                         break;
 364                 if (x == SX_LOCK_UNLOCKED)
 365                         continue;
 366                 if (x == tid && (sx->lock_object.lo_flags & LO_RECURSABLE)) {
 367                         sx->sx_recurse++;
 368                         atomic_set_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
 369                         break;
 370                 }
 371                 rval = 0;
 372                 break;
 373         }
 374
 375         LOCK_LOG_TRY("XLOCK", &sx->lock_object, 0, rval, file, line);
 376         if (rval) {
 377                 WITNESS_LOCK(&sx->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
 378                     file, line);
 379                 if (!recursed)
 380                         LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire,
 381                             sx, 0, 0, file, line, LOCKSTAT_WRITER);
 382                 TD_LOCKS_INC(curthread);
 383         }
 384
 385         return (rval);
 386 }
 387
 388 int
 389 sx_try_xlock_(struct sx *sx, const char *file, int line)
 390 {
 391
 392         return (sx_try_xlock_int(sx LOCK_FILE_LINE_ARG));
 393 }
 394
 395 void
 396 _sx_xunlock(struct sx *sx, const char *file, int line)
 397 {
 398
 399         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 400             ("sx_xunlock() of destroyed sx @ %s:%d", file, line));
 401         _sx_assert(sx, SA_XLOCKED, file, line);
 402         WITNESS_UNLOCK(&sx->lock_object, LOP_EXCLUSIVE, file, line);
 403         LOCK_LOG_LOCK("XUNLOCK", &sx->lock_object, 0, sx->sx_recurse, file,
 404             line);
 405 #if LOCK_DEBUG > 0
 406         _sx_xunlock_hard(sx, (uintptr_t)curthread, file, line);
 407 #else
 408         __sx_xunlock(sx, curthread, file, line);
 409 #endif
 410         TD_LOCKS_DEC(curthread);
 411 }
 412
 413 /*
 414  * Try to do a non-blocking upgrade from a shared lock to an exclusive lock.
 415  * This will only succeed if this thread holds a single shared lock.
 416  * Return 1 if if the upgrade succeed, 0 otherwise.
 417  */
 418 int
 419 sx_try_upgrade_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
 420 {
 421         uintptr_t x;
 422         uintptr_t waiters;
 423         int success;
 424
 425         if (SCHEDULER_STOPPED())
 426                 return (1);
 427
 428         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 429             ("sx_try_upgrade() of destroyed sx @ %s:%d", file, line));
 430         _sx_assert(sx, SA_SLOCKED, file, line);
 431
 432         /*
 433          * Try to switch from one shared lock to an exclusive lock.  We need
 434          * to maintain the SX_LOCK_EXCLUSIVE_WAITERS flag if set so that
 435          * we will wake up the exclusive waiters when we drop the lock.
 436          */
 437         success = 0;
 438         x = SX_READ_VALUE(sx);
 439         for (;;) {
 440                 if (SX_SHARERS(x) > 1)
 441                         break;
 442                 waiters = (x & SX_LOCK_EXCLUSIVE_WAITERS);
 443                 if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x,
 444                     (uintptr_t)curthread | waiters)) {
 445                         success = 1;
 446                         break;
 447                 }
 448         }
 449         LOCK_LOG_TRY("XUPGRADE", &sx->lock_object, 0, success, file, line);
 450         if (success) {
 451                 WITNESS_UPGRADE(&sx->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
 452                     file, line);
 453                 LOCKSTAT_RECORD0(sx__upgrade, sx);
 454         }
 455         return (success);
 456 }
 457
 458 int
 459 sx_try_upgrade_(struct sx *sx, const char *file, int line)
 460 {
 461
 462         return (sx_try_upgrade_int(sx LOCK_FILE_LINE_ARG));
 463 }
 464
 465 /*
 466  * Downgrade an unrecursed exclusive lock into a single shared lock.
 467  */
 468 void
 469 sx_downgrade_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
 470 {
 471         uintptr_t x;
 472         int wakeup_swapper;
 473
 474         if (SCHEDULER_STOPPED())
 475                 return;
 476
 477         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 478             ("sx_downgrade() of destroyed sx @ %s:%d", file, line));
 479         _sx_assert(sx, SA_XLOCKED | SA_NOTRECURSED, file, line);
 480 #ifndef INVARIANTS
 481         if (sx_recursed(sx))
 482                 panic("downgrade of a recursed lock");
 483 #endif
 484
 485         WITNESS_DOWNGRADE(&sx->lock_object, 0, file, line);
 486
 487         /*
 488          * Try to switch from an exclusive lock with no shared waiters
 489          * to one sharer with no shared waiters.  If there are
 490          * exclusive waiters, we don't need to lock the sleep queue so
 491          * long as we preserve the flag.  We do one quick try and if
 492          * that fails we grab the sleepq lock to keep the flags from
 493          * changing and do it the slow way.
 494          *
 495          * We have to lock the sleep queue if there are shared waiters
 496          * so we can wake them up.
 497          */
 498         x = sx->sx_lock;
 499         if (!(x & SX_LOCK_SHARED_WAITERS) &&
 500             atomic_cmpset_rel_ptr(&sx->sx_lock, x, SX_SHARERS_LOCK(1) |
 501             (x & SX_LOCK_EXCLUSIVE_WAITERS)))
 502                 goto out;
 503
 504         /*
 505          * Lock the sleep queue so we can read the waiters bits
 506          * without any races and wakeup any shared waiters.
 507          */
 508         sleepq_lock(&sx->lock_object);
 509
 510         /*
 511          * Preserve SX_LOCK_EXCLUSIVE_WAITERS while downgraded to a single
 512          * shared lock.  If there are any shared waiters, wake them up.
 513          */
 514         wakeup_swapper = 0;
 515         x = sx->sx_lock;
 516         atomic_store_rel_ptr(&sx->sx_lock, SX_SHARERS_LOCK(1) |
 517             (x & SX_LOCK_EXCLUSIVE_WAITERS));
 518         if (x & SX_LOCK_SHARED_WAITERS)
 519                 wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX,
 520                     0, SQ_SHARED_QUEUE);
 521         sleepq_release(&sx->lock_object);
 522
 523         if (wakeup_swapper)
 524                 kick_proc0();
 525
 526 out:
 527         LOCK_LOG_LOCK("XDOWNGRADE", &sx->lock_object, 0, 0, file, line);
 528         LOCKSTAT_RECORD0(sx__downgrade, sx);
 529 }
 530
 531 void
 532 sx_downgrade_(struct sx *sx, const char *file, int line)
 533 {
 534
 535         sx_downgrade_int(sx LOCK_FILE_LINE_ARG);
 536 }
 537
 538 /*
 539  * This function represents the so-called 'hard case' for sx_xlock
 540  * operation.  All 'easy case' failures are redirected to this.  Note
 541  * that ideally this would be a static function, but it needs to be
 542  * accessible from at least sx.h.
 543  */
 544 int
 545 _sx_xlock_hard(struct sx *sx, uintptr_t x, int opts LOCK_FILE_LINE_ARG_DEF)
 546 {
 547         GIANT_DECLARE;
 548         uintptr_t tid;
 549 #ifdef ADAPTIVE_SX
 550         volatile struct thread *owner;
 551         u_int i, n, spintries = 0;
 552         enum { READERS, WRITER } sleep_reason = READERS;
 553         bool adaptive;
 554 #endif
 555 #ifdef LOCK_PROFILING
 556         uint64_t waittime = 0;
 557         int contested = 0;
 558 #endif
 559         int error = 0;
 560 #if defined(ADAPTIVE_SX) || defined(KDTRACE_HOOKS)
 561         struct lock_delay_arg lda;
 562 #endif
 563 #ifdef  KDTRACE_HOOKS
 564         u_int sleep_cnt = 0;
 565         int64_t sleep_time = 0;
 566         int64_t all_time = 0;
 567 #endif
 568 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
 569         uintptr_t state = 0;
 570 #endif
 571         int extra_work = 0;
 572
 573         tid = (uintptr_t)curthread;
 574
 575 #ifdef KDTRACE_HOOKS
 576         if (LOCKSTAT_PROFILE_ENABLED(sx__acquire)) {
 577                 while (x == SX_LOCK_UNLOCKED) {
 578                         if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid))
 579                                 goto out_lockstat;
 580                 }
 581                 extra_work = 1;
 582                 all_time -= lockstat_nsecs(&sx->lock_object);
 583                 state = x;
 584         }
 585 #endif
 586 #ifdef LOCK_PROFILING
 587         extra_work = 1;
 588         state = x;
 589 #endif
 590
 591         if (SCHEDULER_STOPPED())
 592                 return (0);
 593
 594 #if defined(ADAPTIVE_SX)
 595         lock_delay_arg_init(&lda, &sx_delay);
 596 #elif defined(KDTRACE_HOOKS)
 597         lock_delay_arg_init(&lda, NULL);
 598 #endif
 599
 600         if (__predict_false(x == SX_LOCK_UNLOCKED))
 601                 x = SX_READ_VALUE(sx);
 602
 603         /* If we already hold an exclusive lock, then recurse. */
 604         if (__predict_false(lv_sx_owner(x) == (struct thread *)tid)) {
 605                 KASSERT((sx->lock_object.lo_flags & LO_RECURSABLE) != 0,
 606             ("_sx_xlock_hard: recursed on non-recursive sx %s @ %s:%d\n",
 607                     sx->lock_object.lo_name, file, line));
 608                 sx->sx_recurse++;
 609                 atomic_set_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
 610                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 611                         CTR2(KTR_LOCK, "%s: %p recursing", __func__, sx);
 612                 return (0);
 613         }
 614
 615         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 616                 CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__,
 617                     sx->lock_object.lo_name, (void *)sx->sx_lock, file, line);
 618
 619 #ifdef ADAPTIVE_SX
 620         adaptive = ((sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0);
 621 #endif
 622
 623 #ifdef HWPMC_HOOKS
 624         PMC_SOFT_CALL( , , lock, failed);
 625 #endif
 626         lock_profile_obtain_lock_failed(&sx->lock_object, &contested,
 627             &waittime);
 628
 629 #ifndef INVARIANTS
 630         GIANT_SAVE(extra_work);
 631 #endif
 632
 633         for (;;) {
 634                 if (x == SX_LOCK_UNLOCKED) {
 635                         if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid))
 636                                 break;
 637                         continue;
 638                 }
 639 #ifdef INVARIANTS
 640                 GIANT_SAVE(extra_work);
 641 #endif
 642 #ifdef KDTRACE_HOOKS
 643                 lda.spin_cnt++;
 644 #endif
 645 #ifdef ADAPTIVE_SX
 646                 if (__predict_false(!adaptive))
 647                         goto sleepq;
 648                 /*
 649                  * If the lock is write locked and the owner is
 650                  * running on another CPU, spin until the owner stops
 651                  * running or the state of the lock changes.
 652                  */
 653                 if ((x & SX_LOCK_SHARED) == 0) {
 654                         sleep_reason = WRITER;
 655                         owner = lv_sx_owner(x);
 656                         if (!TD_IS_RUNNING(owner))
 657                                 goto sleepq;
 658                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 659                                 CTR3(KTR_LOCK, "%s: spinning on %p held by %p",
 660                                     __func__, sx, owner);
 661                         KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
 662                             "spinning", "lockname:\"%s\"",
 663                             sx->lock_object.lo_name);
 664                         do {
 665                                 lock_delay(&lda);
 666                                 x = SX_READ_VALUE(sx);
 667                                 owner = lv_sx_owner(x);
 668                         } while (owner != NULL && TD_IS_RUNNING(owner));
 669                         KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
 670                             "running");
 671                         continue;
 672                 } else if (SX_SHARERS(x) > 0) {
 673                         sleep_reason = READERS;
 674                         if (spintries == asx_retries)
 675                                 goto sleepq;
 676                         spintries++;
 677                         KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
 678                             "spinning", "lockname:\"%s\"",
 679                             sx->lock_object.lo_name);
 680                         for (i = 0; i < asx_loops; i += n) {
 681                                 n = SX_SHARERS(x);
 682                                 lock_delay_spin(n);
 683                                 x = SX_READ_VALUE(sx);
 684                                 if ((x & SX_LOCK_SHARED) == 0 ||
 685                                     SX_SHARERS(x) == 0)
 686                                         break;
 687                         }
 688 #ifdef KDTRACE_HOOKS
 689                         lda.spin_cnt += i;
 690 #endif
 691                         KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
 692                             "running");
 693                         if (i < asx_loops)
 694                                 continue;
 695                 }
 696 sleepq:
 697 #endif
 698                 sleepq_lock(&sx->lock_object);
 699                 x = SX_READ_VALUE(sx);
 700 retry_sleepq:
 701
 702                 /*
 703                  * If the lock was released while spinning on the
 704                  * sleep queue chain lock, try again.
 705                  */
 706                 if (x == SX_LOCK_UNLOCKED) {
 707                         sleepq_release(&sx->lock_object);
 708                         continue;
 709                 }
 710
 711 #ifdef ADAPTIVE_SX
 712                 /*
 713                  * The current lock owner might have started executing
 714                  * on another CPU (or the lock could have changed
 715                  * owners) while we were waiting on the sleep queue
 716                  * chain lock.  If so, drop the sleep queue lock and try
 717                  * again.
 718                  */
 719                 if (adaptive) {
 720                         if (!(x & SX_LOCK_SHARED)) {
 721                                 owner = (struct thread *)SX_OWNER(x);
 722                                 if (TD_IS_RUNNING(owner)) {
 723                                         sleepq_release(&sx->lock_object);
 724                                         continue;
 725                                 }
 726                         } else if (SX_SHARERS(x) > 0 && sleep_reason == WRITER) {
 727                                 sleepq_release(&sx->lock_object);
 728                                 continue;
 729                         }
 730                 }
 731 #endif
 732
 733                 /*
 734                  * If an exclusive lock was released with both shared
 735                  * and exclusive waiters and a shared waiter hasn't
 736                  * woken up and acquired the lock yet, sx_lock will be
 737                  * set to SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS.
 738                  * If we see that value, try to acquire it once.  Note
 739                  * that we have to preserve SX_LOCK_EXCLUSIVE_WAITERS
 740                  * as there are other exclusive waiters still.  If we
 741                  * fail, restart the loop.
 742                  */
 743                 if (x == (SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS)) {
 744                         if (!atomic_fcmpset_acq_ptr(&sx->sx_lock, &x,
 745                             tid | SX_LOCK_EXCLUSIVE_WAITERS))
 746                                 goto retry_sleepq;
 747                         sleepq_release(&sx->lock_object);
 748                         CTR2(KTR_LOCK, "%s: %p claimed by new writer",
 749                             __func__, sx);
 750                         break;
 751                 }
 752
 753                 /*
 754                  * Try to set the SX_LOCK_EXCLUSIVE_WAITERS.  If we fail,
 755                  * than loop back and retry.
 756                  */
 757                 if (!(x & SX_LOCK_EXCLUSIVE_WAITERS)) {
 758                         if (!atomic_fcmpset_ptr(&sx->sx_lock, &x,
 759                             x | SX_LOCK_EXCLUSIVE_WAITERS)) {
 760                                 goto retry_sleepq;
 761                         }
 762                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 763                                 CTR2(KTR_LOCK, "%s: %p set excl waiters flag",
 764                                     __func__, sx);
 765                 }
 766
 767                 /*
 768                  * Since we have been unable to acquire the exclusive
 769                  * lock and the exclusive waiters flag is set, we have
 770                  * to sleep.
 771                  */
 772                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 773                         CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
 774                             __func__, sx);
 775
 776 #ifdef KDTRACE_HOOKS
 777                 sleep_time -= lockstat_nsecs(&sx->lock_object);
 778 #endif
 779                 sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
 780                     SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
 781                     SLEEPQ_INTERRUPTIBLE : 0), SQ_EXCLUSIVE_QUEUE);
 782                 if (!(opts & SX_INTERRUPTIBLE))
 783                         sleepq_wait(&sx->lock_object, 0);
 784                 else
 785                         error = sleepq_wait_sig(&sx->lock_object, 0);
 786 #ifdef KDTRACE_HOOKS
 787                 sleep_time += lockstat_nsecs(&sx->lock_object);
 788                 sleep_cnt++;
 789 #endif
 790                 if (error) {
 791                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 792                                 CTR2(KTR_LOCK,
 793                         "%s: interruptible sleep by %p suspended by signal",
 794                                     __func__, sx);
 795                         break;
 796                 }
 797                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 798                         CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
 799                             __func__, sx);
 800                 x = SX_READ_VALUE(sx);
 801         }
 802 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
 803         if (__predict_true(!extra_work))
 804                 return (error);
 805 #endif
 806 #ifdef KDTRACE_HOOKS
 807         all_time += lockstat_nsecs(&sx->lock_object);
 808         if (sleep_time)
 809                 LOCKSTAT_RECORD4(sx__block, sx, sleep_time,
 810                     LOCKSTAT_WRITER, (state & SX_LOCK_SHARED) == 0,
 811                     (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
 812         if (lda.spin_cnt > sleep_cnt)
 813                 LOCKSTAT_RECORD4(sx__spin, sx, all_time - sleep_time,
 814                     LOCKSTAT_WRITER, (state & SX_LOCK_SHARED) == 0,
 815                     (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
 816 out_lockstat:
 817 #endif
 818         if (!error)
 819                 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
 820                     contested, waittime, file, line, LOCKSTAT_WRITER);
 821         GIANT_RESTORE();
 822         return (error);
 823 }
 824
 825 /*
 826  * This function represents the so-called 'hard case' for sx_xunlock
 827  * operation.  All 'easy case' failures are redirected to this.  Note
 828  * that ideally this would be a static function, but it needs to be
 829  * accessible from at least sx.h.
 830  */
 831 void
 832 _sx_xunlock_hard(struct sx *sx, uintptr_t x LOCK_FILE_LINE_ARG_DEF)
 833 {
 834         uintptr_t tid, setx;
 835         int queue, wakeup_swapper;
 836
 837         if (SCHEDULER_STOPPED())
 838                 return;
 839
 840         tid = (uintptr_t)curthread;
 841
 842         if (__predict_false(x == tid))
 843                 x = SX_READ_VALUE(sx);
 844
 845         MPASS(!(x & SX_LOCK_SHARED));
 846
 847         if (__predict_false(x & SX_LOCK_RECURSED)) {
 848                 /* The lock is recursed, unrecurse one level. */
 849                 if ((--sx->sx_recurse) == 0)
 850                         atomic_clear_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
 851                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 852                         CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, sx);
 853                 return;
 854         }
 855
 856         LOCKSTAT_PROFILE_RELEASE_RWLOCK(sx__release, sx, LOCKSTAT_WRITER);
 857         if (x == tid &&
 858             atomic_cmpset_rel_ptr(&sx->sx_lock, tid, SX_LOCK_UNLOCKED))
 859                 return;
 860
 861         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 862                 CTR2(KTR_LOCK, "%s: %p contested", __func__, sx);
 863
 864         sleepq_lock(&sx->lock_object);
 865         x = SX_READ_VALUE(sx);
 866         MPASS(x & (SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS));
 867
 868         /*
 869          * The wake up algorithm here is quite simple and probably not
 870          * ideal.  It gives precedence to shared waiters if they are
 871          * present.  For this condition, we have to preserve the
 872          * state of the exclusive waiters flag.
 873          * If interruptible sleeps left the shared queue empty avoid a
 874          * starvation for the threads sleeping on the exclusive queue by giving
 875          * them precedence and cleaning up the shared waiters bit anyway.
 876          */
 877         setx = SX_LOCK_UNLOCKED;
 878         queue = SQ_EXCLUSIVE_QUEUE;
 879         if ((x & SX_LOCK_SHARED_WAITERS) != 0 &&
 880             sleepq_sleepcnt(&sx->lock_object, SQ_SHARED_QUEUE) != 0) {
 881                 queue = SQ_SHARED_QUEUE;
 882                 setx |= (x & SX_LOCK_EXCLUSIVE_WAITERS);
 883         }
 884         atomic_store_rel_ptr(&sx->sx_lock, setx);
 885
 886         /* Wake up all the waiters for the specific queue. */
 887         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 888                 CTR3(KTR_LOCK, "%s: %p waking up all threads on %s queue",
 889                     __func__, sx, queue == SQ_SHARED_QUEUE ? "shared" :
 890                     "exclusive");
 891
 892         wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX, 0,
 893             queue);
 894         sleepq_release(&sx->lock_object);
 895         if (wakeup_swapper)
 896                 kick_proc0();
 897 }
 898
 899 static bool __always_inline
 900 __sx_slock_try(struct sx *sx, uintptr_t *xp LOCK_FILE_LINE_ARG_DEF)
 901 {
 902
 903         /*
 904          * If no other thread has an exclusive lock then try to bump up
 905          * the count of sharers.  Since we have to preserve the state
 906          * of SX_LOCK_EXCLUSIVE_WAITERS, if we fail to acquire the
 907          * shared lock loop back and retry.
 908          */
 909         while (*xp & SX_LOCK_SHARED) {
 910                 MPASS(!(*xp & SX_LOCK_SHARED_WAITERS));
 911                 if (atomic_fcmpset_acq_ptr(&sx->sx_lock, xp,
 912                     *xp + SX_ONE_SHARER)) {
 913                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 914                                 CTR4(KTR_LOCK, "%s: %p succeed %p -> %p",
 915                                     __func__, sx, (void *)*xp,
 916                                     (void *)(*xp + SX_ONE_SHARER));
 917                         return (true);
 918                 }
 919         }
 920         return (false);
 921 }
 922
 923 static int __noinline
 924 _sx_slock_hard(struct sx *sx, int opts, uintptr_t x LOCK_FILE_LINE_ARG_DEF)
 925 {
 926         GIANT_DECLARE;
 927 #ifdef ADAPTIVE_SX
 928         volatile struct thread *owner;
 929         bool adaptive;
 930 #endif
 931 #ifdef LOCK_PROFILING
 932         uint64_t waittime = 0;
 933         int contested = 0;
 934 #endif
 935         int error = 0;
 936 #if defined(ADAPTIVE_SX) || defined(KDTRACE_HOOKS)
 937         struct lock_delay_arg lda;
 938 #endif
 939 #ifdef KDTRACE_HOOKS
 940         u_int sleep_cnt = 0;
 941         int64_t sleep_time = 0;
 942         int64_t all_time = 0;
 943 #endif
 944 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
 945         uintptr_t state = 0;
 946 #endif
 947         int extra_work = 0;
 948
 949 #ifdef KDTRACE_HOOKS
 950         if (LOCKSTAT_PROFILE_ENABLED(sx__acquire)) {
 951                 if (__sx_slock_try(sx, &x LOCK_FILE_LINE_ARG))
 952                         goto out_lockstat;
 953                 extra_work = 1;
 954                 all_time -= lockstat_nsecs(&sx->lock_object);
 955                 state = x;
 956         }
 957 #endif
 958 #ifdef LOCK_PROFILING
 959         extra_work = 1;
 960         state = x;
 961 #endif
 962
 963         if (SCHEDULER_STOPPED())
 964                 return (0);
 965
 966 #if defined(ADAPTIVE_SX)
 967         lock_delay_arg_init(&lda, &sx_delay);
 968 #elif defined(KDTRACE_HOOKS)
 969         lock_delay_arg_init(&lda, NULL);
 970 #endif
 971
 972 #ifdef ADAPTIVE_SX
 973         adaptive = ((sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0);
 974 #endif
 975
 976 #ifdef HWPMC_HOOKS
 977         PMC_SOFT_CALL( , , lock, failed);
 978 #endif
 979         lock_profile_obtain_lock_failed(&sx->lock_object, &contested,
 980             &waittime);
 981
 982 #ifndef INVARIANTS
 983         GIANT_SAVE(extra_work);
 984 #endif
 985
 986         /*
 987          * As with rwlocks, we don't make any attempt to try to block
 988          * shared locks once there is an exclusive waiter.
 989          */
 990         for (;;) {
 991                 if (__sx_slock_try(sx, &x LOCK_FILE_LINE_ARG))
 992                         break;
 993 #ifdef INVARIANTS
 994                 GIANT_SAVE(extra_work);
 995 #endif
 996 #ifdef KDTRACE_HOOKS
 997                 lda.spin_cnt++;
 998 #endif
 999
1000 #ifdef ADAPTIVE_SX
1001                 if (__predict_false(!adaptive))
1002                         goto sleepq;
1003                 /*
1004                  * If the owner is running on another CPU, spin until
1005                  * the owner stops running or the state of the lock
1006                  * changes.
1007                  */
1008                 owner = lv_sx_owner(x);
1009                 if (TD_IS_RUNNING(owner)) {
1010                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
1011                                 CTR3(KTR_LOCK,
1012                                     "%s: spinning on %p held by %p",
1013                                     __func__, sx, owner);
1014                         KTR_STATE1(KTR_SCHED, "thread",
1015                             sched_tdname(curthread), "spinning",
1016                             "lockname:\"%s\"", sx->lock_object.lo_name);
1017                         do {
1018                                 lock_delay(&lda);
1019                                 x = SX_READ_VALUE(sx);
1020                                 owner = lv_sx_owner(x);
1021                         } while (owner != NULL && TD_IS_RUNNING(owner));
1022                         KTR_STATE0(KTR_SCHED, "thread",
1023                             sched_tdname(curthread), "running");
1024                         continue;
1025                 }
1026 sleepq:
1027 #endif
1028
1029                 /*
1030                  * Some other thread already has an exclusive lock, so
1031                  * start the process of blocking.
1032                  */
1033                 sleepq_lock(&sx->lock_object);
1034                 x = SX_READ_VALUE(sx);
1035 retry_sleepq:
1036                 /*
1037                  * The lock could have been released while we spun.
1038                  * In this case loop back and retry.
1039                  */
1040                 if (x & SX_LOCK_SHARED) {
1041                         sleepq_release(&sx->lock_object);
1042                         continue;
1043                 }
1044
1045 #ifdef ADAPTIVE_SX
1046                 /*
1047                  * If the owner is running on another CPU, spin until
1048                  * the owner stops running or the state of the lock
1049                  * changes.
1050                  */
1051                 if (!(x & SX_LOCK_SHARED) && adaptive) {
1052                         owner = (struct thread *)SX_OWNER(x);
1053                         if (TD_IS_RUNNING(owner)) {
1054                                 sleepq_release(&sx->lock_object);
1055                                 x = SX_READ_VALUE(sx);
1056                                 continue;
1057                         }
1058                 }
1059 #endif
1060
1061                 /*
1062                  * Try to set the SX_LOCK_SHARED_WAITERS flag.  If we
1063                  * fail to set it drop the sleep queue lock and loop
1064                  * back.
1065                  */
1066                 if (!(x & SX_LOCK_SHARED_WAITERS)) {
1067                         if (!atomic_fcmpset_ptr(&sx->sx_lock, &x,
1068                             x | SX_LOCK_SHARED_WAITERS))
1069                                 goto retry_sleepq;
1070                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
1071                                 CTR2(KTR_LOCK, "%s: %p set shared waiters flag",
1072                                     __func__, sx);
1073                 }
1074
1075                 /*
1076                  * Since we have been unable to acquire the shared lock,
1077                  * we have to sleep.
1078                  */
1079                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1080                         CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
1081                             __func__, sx);
1082
1083 #ifdef KDTRACE_HOOKS
1084                 sleep_time -= lockstat_nsecs(&sx->lock_object);
1085 #endif
1086                 sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
1087                     SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
1088                     SLEEPQ_INTERRUPTIBLE : 0), SQ_SHARED_QUEUE);
1089                 if (!(opts & SX_INTERRUPTIBLE))
1090                         sleepq_wait(&sx->lock_object, 0);
1091                 else
1092                         error = sleepq_wait_sig(&sx->lock_object, 0);
1093 #ifdef KDTRACE_HOOKS
1094                 sleep_time += lockstat_nsecs(&sx->lock_object);
1095                 sleep_cnt++;
1096 #endif
1097                 if (error) {
1098                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
1099                                 CTR2(KTR_LOCK,
1100                         "%s: interruptible sleep by %p suspended by signal",
1101                                     __func__, sx);
1102                         break;
1103                 }
1104                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1105                         CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
1106                             __func__, sx);
1107                 x = SX_READ_VALUE(sx);
1108         }
1109 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
1110         if (__predict_true(!extra_work))
1111                 return (error);
1112 #endif
1113 #ifdef KDTRACE_HOOKS
1114         all_time += lockstat_nsecs(&sx->lock_object);
1115         if (sleep_time)
1116                 LOCKSTAT_RECORD4(sx__block, sx, sleep_time,
1117                     LOCKSTAT_READER, (state & SX_LOCK_SHARED) == 0,
1118                     (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
1119         if (lda.spin_cnt > sleep_cnt)
1120                 LOCKSTAT_RECORD4(sx__spin, sx, all_time - sleep_time,
1121                     LOCKSTAT_READER, (state & SX_LOCK_SHARED) == 0,
1122                     (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
1123 out_lockstat:
1124 #endif
1125         if (error == 0) {
1126                 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
1127                     contested, waittime, file, line, LOCKSTAT_READER);
1128         }
1129         GIANT_RESTORE();
1130         return (error);
1131 }
1132
1133 int
1134 _sx_slock_int(struct sx *sx, int opts LOCK_FILE_LINE_ARG_DEF)
1135 {
1136         uintptr_t x;
1137         int error;
1138
1139         KASSERT(kdb_active != 0 || SCHEDULER_STOPPED() ||
1140             !TD_IS_IDLETHREAD(curthread),
1141             ("sx_slock() by idle thread %p on sx %s @ %s:%d",
1142             curthread, sx->lock_object.lo_name, file, line));
1143         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
1144             ("sx_slock() of destroyed sx @ %s:%d", file, line));
1145         WITNESS_CHECKORDER(&sx->lock_object, LOP_NEWORDER, file, line, NULL);
1146
1147         error = 0;
1148         x = SX_READ_VALUE(sx);
1149         if (__predict_false(LOCKSTAT_PROFILE_ENABLED(sx__acquire) ||
1150             !__sx_slock_try(sx, &x LOCK_FILE_LINE_ARG)))
1151                 error = _sx_slock_hard(sx, opts, x LOCK_FILE_LINE_ARG);
1152         else
1153                 lock_profile_obtain_lock_success(&sx->lock_object, 0, 0,
1154                     file, line);
1155         if (error == 0) {
1156                 LOCK_LOG_LOCK("SLOCK", &sx->lock_object, 0, 0, file, line);
1157                 WITNESS_LOCK(&sx->lock_object, 0, file, line);
1158                 TD_LOCKS_INC(curthread);
1159         }
1160         return (error);
1161 }
1162
1163 int
1164 _sx_slock(struct sx *sx, int opts, const char *file, int line)
1165 {
1166
1167         return (_sx_slock_int(sx, opts LOCK_FILE_LINE_ARG));
1168 }
1169
1170 static bool __always_inline
1171 _sx_sunlock_try(struct sx *sx, uintptr_t *xp)
1172 {
1173
1174         for (;;) {
1175                 /*
1176                  * We should never have sharers while at least one thread
1177                  * holds a shared lock.
1178                  */
1179                 KASSERT(!(*xp & SX_LOCK_SHARED_WAITERS),
1180                     ("%s: waiting sharers", __func__));
1181
1182                 /*
1183                  * See if there is more than one shared lock held.  If
1184                  * so, just drop one and return.
1185                  */
1186                 if (SX_SHARERS(*xp) > 1) {
1187                         if (atomic_fcmpset_rel_ptr(&sx->sx_lock, xp,
1188                             *xp - SX_ONE_SHARER)) {
1189                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1190                                         CTR4(KTR_LOCK,
1191                                             "%s: %p succeeded %p -> %p",
1192                                             __func__, sx, (void *)*xp,
1193                                             (void *)(*xp - SX_ONE_SHARER));
1194                                 return (true);
1195                         }
1196                         continue;
1197                 }
1198
1199                 /*
1200                  * If there aren't any waiters for an exclusive lock,
1201                  * then try to drop it quickly.
1202                  */
1203                 if (!(*xp & SX_LOCK_EXCLUSIVE_WAITERS)) {
1204                         MPASS(*xp == SX_SHARERS_LOCK(1));
1205                         *xp = SX_SHARERS_LOCK(1);
1206                         if (atomic_fcmpset_rel_ptr(&sx->sx_lock,
1207                             xp, SX_LOCK_UNLOCKED)) {
1208                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1209                                         CTR2(KTR_LOCK, "%s: %p last succeeded",
1210                                             __func__, sx);
1211                                 return (true);
1212                         }
1213                         continue;
1214                 }
1215                 break;
1216         }
1217         return (false);
1218 }
1219
1220 static void __noinline
1221 _sx_sunlock_hard(struct sx *sx, uintptr_t x LOCK_FILE_LINE_ARG_DEF)
1222 {
1223         int wakeup_swapper = 0;
1224         uintptr_t setx;
1225
1226         if (SCHEDULER_STOPPED())
1227                 return;
1228
1229         if (_sx_sunlock_try(sx, &x))
1230                 goto out_lockstat;
1231
1232         /*
1233          * At this point, there should just be one sharer with
1234          * exclusive waiters.
1235          */
1236         MPASS(x == (SX_SHARERS_LOCK(1) | SX_LOCK_EXCLUSIVE_WAITERS));
1237
1238         sleepq_lock(&sx->lock_object);
1239         x = SX_READ_VALUE(sx);
1240         for (;;) {
1241                 MPASS(x & SX_LOCK_EXCLUSIVE_WAITERS);
1242                 MPASS(!(x & SX_LOCK_SHARED_WAITERS));
1243                 if (_sx_sunlock_try(sx, &x))
1244                         break;
1245
1246                 /*
1247                  * Wake up semantic here is quite simple:
1248                  * Just wake up all the exclusive waiters.
1249                  * Note that the state of the lock could have changed,
1250                  * so if it fails loop back and retry.
1251                  */
1252                 setx = x - SX_ONE_SHARER;
1253                 setx &= ~SX_LOCK_EXCLUSIVE_WAITERS;
1254                 if (!atomic_fcmpset_rel_ptr(&sx->sx_lock, &x, setx))
1255                         continue;
1256                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1257                         CTR2(KTR_LOCK, "%s: %p waking up all thread on"
1258                             "exclusive queue", __func__, sx);
1259                 wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX,
1260                     0, SQ_EXCLUSIVE_QUEUE);
1261                 break;
1262         }
1263         sleepq_release(&sx->lock_object);
1264         if (wakeup_swapper)
1265                 kick_proc0();
1266 out_lockstat:
1267         LOCKSTAT_PROFILE_RELEASE_RWLOCK(sx__release, sx, LOCKSTAT_READER);
1268 }
1269
1270 void
1271 _sx_sunlock_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
1272 {
1273         uintptr_t x;
1274
1275         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
1276             ("sx_sunlock() of destroyed sx @ %s:%d", file, line));
1277         _sx_assert(sx, SA_SLOCKED, file, line);
1278         WITNESS_UNLOCK(&sx->lock_object, 0, file, line);
1279         LOCK_LOG_LOCK("SUNLOCK", &sx->lock_object, 0, 0, file, line);
1280
1281         x = SX_READ_VALUE(sx);
1282         if (__predict_false(LOCKSTAT_PROFILE_ENABLED(sx__release) ||
1283             !_sx_sunlock_try(sx, &x)))
1284                 _sx_sunlock_hard(sx, x LOCK_FILE_LINE_ARG);
1285         else
1286                 lock_profile_release_lock(&sx->lock_object);
1287
1288         TD_LOCKS_DEC(curthread);
1289 }
1290
1291 void
1292 _sx_sunlock(struct sx *sx, const char *file, int line)
1293 {
1294
1295         _sx_sunlock_int(sx LOCK_FILE_LINE_ARG);
1296 }
1297
1298 #ifdef INVARIANT_SUPPORT
1299 #ifndef INVARIANTS
1300 #undef  _sx_assert
1301 #endif
1302
1303 /*
1304  * In the non-WITNESS case, sx_assert() can only detect that at least
1305  * *some* thread owns an slock, but it cannot guarantee that *this*
1306  * thread owns an slock.
1307  */
1308 void
1309 _sx_assert(const struct sx *sx, int what, const char *file, int line)
1310 {
1311 #ifndef WITNESS
1312         int slocked = 0;
1313 #endif
1314
1315         if (SCHEDULER_STOPPED())
1316                 return;
1317         switch (what) {
1318         case SA_SLOCKED:
1319         case SA_SLOCKED | SA_NOTRECURSED:
1320         case SA_SLOCKED | SA_RECURSED:
1321 #ifndef WITNESS
1322                 slocked = 1;
1323                 /* FALLTHROUGH */
1324 #endif
1325         case SA_LOCKED:
1326         case SA_LOCKED | SA_NOTRECURSED:
1327         case SA_LOCKED | SA_RECURSED:
1328 #ifdef WITNESS
1329                 witness_assert(&sx->lock_object, what, file, line);
1330 #else
1331                 /*
1332                  * If some other thread has an exclusive lock or we
1333                  * have one and are asserting a shared lock, fail.
1334                  * Also, if no one has a lock at all, fail.
1335                  */
1336                 if (sx->sx_lock == SX_LOCK_UNLOCKED ||
1337                     (!(sx->sx_lock & SX_LOCK_SHARED) && (slocked ||
1338                     sx_xholder(sx) != curthread)))
1339                         panic("Lock %s not %slocked @ %s:%d\n",
1340                             sx->lock_object.lo_name, slocked ? "share " : "",
1341                             file, line);
1342
1343                 if (!(sx->sx_lock & SX_LOCK_SHARED)) {
1344                         if (sx_recursed(sx)) {
1345                                 if (what & SA_NOTRECURSED)
1346                                         panic("Lock %s recursed @ %s:%d\n",
1347                                             sx->lock_object.lo_name, file,
1348                                             line);
1349                         } else if (what & SA_RECURSED)
1350                                 panic("Lock %s not recursed @ %s:%d\n",
1351                                     sx->lock_object.lo_name, file, line);
1352                 }
1353 #endif
1354                 break;
1355         case SA_XLOCKED:
1356         case SA_XLOCKED | SA_NOTRECURSED:
1357         case SA_XLOCKED | SA_RECURSED:
1358                 if (sx_xholder(sx) != curthread)
1359                         panic("Lock %s not exclusively locked @ %s:%d\n",
1360                             sx->lock_object.lo_name, file, line);
1361                 if (sx_recursed(sx)) {
1362                         if (what & SA_NOTRECURSED)
1363                                 panic("Lock %s recursed @ %s:%d\n",
1364                                     sx->lock_object.lo_name, file, line);
1365                 } else if (what & SA_RECURSED)
1366                         panic("Lock %s not recursed @ %s:%d\n",
1367                             sx->lock_object.lo_name, file, line);
1368                 break;
1369         case SA_UNLOCKED:
1370 #ifdef WITNESS
1371                 witness_assert(&sx->lock_object, what, file, line);
1372 #else
1373                 /*
1374                  * If we hold an exclusve lock fail.  We can't
1375                  * reliably check to see if we hold a shared lock or
1376                  * not.
1377                  */
1378                 if (sx_xholder(sx) == curthread)
1379                         panic("Lock %s exclusively locked @ %s:%d\n",
1380                             sx->lock_object.lo_name, file, line);
1381 #endif
1382                 break;
1383         default:
1384                 panic("Unknown sx lock assertion: %d @ %s:%d", what, file,
1385                     line);
1386         }
1387 }
1388 #endif  /* INVARIANT_SUPPORT */
1389
1390 #ifdef DDB
1391 static void
1392 db_show_sx(const struct lock_object *lock)
1393 {
1394         struct thread *td;
1395         const struct sx *sx;
1396
1397         sx = (const struct sx *)lock;
1398
1399         db_printf(" state: ");
1400         if (sx->sx_lock == SX_LOCK_UNLOCKED)
1401                 db_printf("UNLOCKED\n");
1402         else if (sx->sx_lock == SX_LOCK_DESTROYED) {
1403                 db_printf("DESTROYED\n");
1404                 return;
1405         } else if (sx->sx_lock & SX_LOCK_SHARED)
1406                 db_printf("SLOCK: %ju\n", (uintmax_t)SX_SHARERS(sx->sx_lock));
1407         else {
1408                 td = sx_xholder(sx);
1409                 db_printf("XLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1410                     td->td_tid, td->td_proc->p_pid, td->td_name);
1411                 if (sx_recursed(sx))
1412                         db_printf(" recursed: %d\n", sx->sx_recurse);
1413         }
1414
1415         db_printf(" waiters: ");
1416         switch(sx->sx_lock &
1417             (SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS)) {
1418         case SX_LOCK_SHARED_WAITERS:
1419                 db_printf("shared\n");
1420                 break;
1421         case SX_LOCK_EXCLUSIVE_WAITERS:
1422                 db_printf("exclusive\n");
1423                 break;
1424         case SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS:
1425                 db_printf("exclusive and shared\n");
1426                 break;
1427         default:
1428                 db_printf("none\n");
1429         }
1430 }
1431
1432 /*
1433  * Check to see if a thread that is blocked on a sleep queue is actually
1434  * blocked on an sx lock.  If so, output some details and return true.
1435  * If the lock has an exclusive owner, return that in *ownerp.
1436  */
1437 int
1438 sx_chain(struct thread *td, struct thread **ownerp)
1439 {
1440         struct sx *sx;
1441
1442         /*
1443          * Check to see if this thread is blocked on an sx lock.
1444          * First, we check the lock class.  If that is ok, then we
1445          * compare the lock name against the wait message.
1446          */
1447         sx = td->td_wchan;
1448         if (LOCK_CLASS(&sx->lock_object) != &lock_class_sx ||
1449             sx->lock_object.lo_name != td->td_wmesg)
1450                 return (0);
1451
1452         /* We think we have an sx lock, so output some details. */
1453         db_printf("blocked on sx \"%s\" ", td->td_wmesg);
1454         *ownerp = sx_xholder(sx);
1455         if (sx->sx_lock & SX_LOCK_SHARED)
1456                 db_printf("SLOCK (count %ju)\n",
1457                     (uintmax_t)SX_SHARERS(sx->sx_lock));
1458         else
1459                 db_printf("XLOCK\n");
1460         return (1);
1461 }
1462 #endif