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