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_kdtrace.h"
  42 #include "opt_no_adaptive_sx.h"
  43
  44 #include <sys/cdefs.h>
  45 __FBSDID("$FreeBSD$");
  46
  47 #include <sys/param.h>
  48 #include <sys/systm.h>
  49 #include <sys/kdb.h>
  50 #include <sys/kernel.h>
  51 #include <sys/ktr.h>
  52 #include <sys/lock.h>
  53 #include <sys/mutex.h>
  54 #include <sys/proc.h>
  55 #include <sys/sched.h>
  56 #include <sys/sleepqueue.h>
  57 #include <sys/sx.h>
  58 #include <sys/smp.h>
  59 #include <sys/sysctl.h>
  60
  61 #if defined(SMP) && !defined(NO_ADAPTIVE_SX)
  62 #include <machine/cpu.h>
  63 #endif
  64
  65 #ifdef DDB
  66 #include <ddb/ddb.h>
  67 #endif
  68
  69 #if defined(SMP) && !defined(NO_ADAPTIVE_SX)
  70 #define ADAPTIVE_SX
  71 #endif
  72
  73 CTASSERT((SX_NOADAPTIVE & LO_CLASSFLAGS) == SX_NOADAPTIVE);
  74
  75 #ifdef HWPMC_HOOKS
  76 #include <sys/pmckern.h>
  77 PMC_SOFT_DECLARE( , , lock, failed);
  78 #endif
  79
  80 /* Handy macros for sleep queues. */
  81 #define SQ_EXCLUSIVE_QUEUE      0
  82 #define SQ_SHARED_QUEUE         1
  83
  84 /*
  85  * Variations on DROP_GIANT()/PICKUP_GIANT() for use in this file.  We
  86  * drop Giant anytime we have to sleep or if we adaptively spin.
  87  */
  88 #define GIANT_DECLARE                                                   \
  89         int _giantcnt = 0;                                              \
  90         WITNESS_SAVE_DECL(Giant)                                        \
  91
  92 #define GIANT_SAVE() do {                                               \
  93         if (mtx_owned(&Giant)) {                                        \
  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_recurse              lock_object.lo_data
 116 #define sx_recursed(sx)         ((sx)->sx_recurse != 0)
 117
 118 static void     assert_sx(const struct lock_object *lock, int what);
 119 #ifdef DDB
 120 static void     db_show_sx(const struct lock_object *lock);
 121 #endif
 122 static void     lock_sx(struct lock_object *lock, uintptr_t how);
 123 #ifdef KDTRACE_HOOKS
 124 static int      owner_sx(const struct lock_object *lock, struct thread **owner);
 125 #endif
 126 static uintptr_t unlock_sx(struct lock_object *lock);
 127
 128 struct lock_class lock_class_sx = {
 129         .lc_name = "sx",
 130         .lc_flags = LC_SLEEPLOCK | LC_SLEEPABLE | LC_RECURSABLE | LC_UPGRADABLE,
 131         .lc_assert = assert_sx,
 132 #ifdef DDB
 133         .lc_ddb_show = db_show_sx,
 134 #endif
 135         .lc_lock = lock_sx,
 136         .lc_unlock = unlock_sx,
 137 #ifdef KDTRACE_HOOKS
 138         .lc_owner = owner_sx,
 139 #endif
 140 };
 141
 142 #ifndef INVARIANTS
 143 #define _sx_assert(sx, what, file, line)
 144 #endif
 145
 146 #ifdef ADAPTIVE_SX
 147 static u_int asx_retries = 10;
 148 static u_int asx_loops = 10000;
 149 static SYSCTL_NODE(_debug, OID_AUTO, sx, CTLFLAG_RD, NULL, "sxlock debugging");
 150 SYSCTL_UINT(_debug_sx, OID_AUTO, retries, CTLFLAG_RW, &asx_retries, 0, "");
 151 SYSCTL_UINT(_debug_sx, OID_AUTO, loops, CTLFLAG_RW, &asx_loops, 0, "");
 152
 153 static struct lock_delay_config sx_delay = {
 154         .initial        = 1000,
 155         .step           = 500,
 156         .min            = 100,
 157         .max            = 5000,
 158 };
 159
 160 SYSCTL_INT(_debug_sx, OID_AUTO, delay_initial, CTLFLAG_RW, &sx_delay.initial,
 161     0, "");
 162 SYSCTL_INT(_debug_sx, OID_AUTO, delay_step, CTLFLAG_RW, &sx_delay.step,
 163     0, "");
 164 SYSCTL_INT(_debug_sx, OID_AUTO, delay_min, CTLFLAG_RW, &sx_delay.min,
 165     0, "");
 166 SYSCTL_INT(_debug_sx, OID_AUTO, delay_max, CTLFLAG_RW, &sx_delay.max,
 167     0, "");
 168
 169 static void
 170 sx_delay_sysinit(void *dummy)
 171 {
 172
 173         sx_delay.initial = mp_ncpus * 25;
 174         sx_delay.step = (mp_ncpus * 25) / 2;
 175         sx_delay.min = mp_ncpus * 5;
 176         sx_delay.max = mp_ncpus * 25 * 10;
 177 }
 178 LOCK_DELAY_SYSINIT(sx_delay_sysinit);
 179 #endif
 180
 181 void
 182 assert_sx(const struct lock_object *lock, int what)
 183 {
 184
 185         sx_assert((const struct sx *)lock, what);
 186 }
 187
 188 void
 189 lock_sx(struct lock_object *lock, uintptr_t how)
 190 {
 191         struct sx *sx;
 192
 193         sx = (struct sx *)lock;
 194         if (how)
 195                 sx_slock(sx);
 196         else
 197                 sx_xlock(sx);
 198 }
 199
 200 uintptr_t
 201 unlock_sx(struct lock_object *lock)
 202 {
 203         struct sx *sx;
 204
 205         sx = (struct sx *)lock;
 206         sx_assert(sx, SA_LOCKED | SA_NOTRECURSED);
 207         if (sx_xlocked(sx)) {
 208                 sx_xunlock(sx);
 209                 return (0);
 210         } else {
 211                 sx_sunlock(sx);
 212                 return (1);
 213         }
 214 }
 215
 216 #ifdef KDTRACE_HOOKS
 217 int
 218 owner_sx(const struct lock_object *lock, struct thread **owner)
 219 {
 220         const struct sx *sx = (const struct sx *)lock;
 221         uintptr_t x = sx->sx_lock;
 222
 223         *owner = (struct thread *)SX_OWNER(x);
 224         return ((x & SX_LOCK_SHARED) != 0 ? (SX_SHARERS(x) != 0) :
 225             (*owner != NULL));
 226 }
 227 #endif
 228
 229 void
 230 sx_sysinit(void *arg)
 231 {
 232         struct sx_args *sargs = arg;
 233
 234         sx_init_flags(sargs->sa_sx, sargs->sa_desc, sargs->sa_flags);
 235 }
 236
 237 void
 238 sx_init_flags(struct sx *sx, const char *description, int opts)
 239 {
 240         int flags;
 241
 242         MPASS((opts & ~(SX_QUIET | SX_RECURSE | SX_NOWITNESS | SX_DUPOK |
 243             SX_NOPROFILE | SX_NOADAPTIVE | SX_NEW)) == 0);
 244         ASSERT_ATOMIC_LOAD_PTR(sx->sx_lock,
 245             ("%s: sx_lock not aligned for %s: %p", __func__, description,
 246             &sx->sx_lock));
 247
 248         flags = LO_SLEEPABLE | LO_UPGRADABLE;
 249         if (opts & SX_DUPOK)
 250                 flags |= LO_DUPOK;
 251         if (opts & SX_NOPROFILE)
 252                 flags |= LO_NOPROFILE;
 253         if (!(opts & SX_NOWITNESS))
 254                 flags |= LO_WITNESS;
 255         if (opts & SX_RECURSE)
 256                 flags |= LO_RECURSABLE;
 257         if (opts & SX_QUIET)
 258                 flags |= LO_QUIET;
 259         if (opts & SX_NEW)
 260                 flags |= LO_NEW;
 261
 262         flags |= opts & SX_NOADAPTIVE;
 263         lock_init(&sx->lock_object, &lock_class_sx, description, NULL, flags);
 264         sx->sx_lock = SX_LOCK_UNLOCKED;
 265         sx->sx_recurse = 0;
 266 }
 267
 268 void
 269 sx_destroy(struct sx *sx)
 270 {
 271
 272         KASSERT(sx->sx_lock == SX_LOCK_UNLOCKED, ("sx lock still held"));
 273         KASSERT(sx->sx_recurse == 0, ("sx lock still recursed"));
 274         sx->sx_lock = SX_LOCK_DESTROYED;
 275         lock_destroy(&sx->lock_object);
 276 }
 277
 278 int
 279 _sx_slock(struct sx *sx, int opts, const char *file, int line)
 280 {
 281         int error = 0;
 282
 283         if (SCHEDULER_STOPPED())
 284                 return (0);
 285         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
 286             ("sx_slock() by idle thread %p on sx %s @ %s:%d",
 287             curthread, sx->lock_object.lo_name, file, line));
 288         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 289             ("sx_slock() of destroyed sx @ %s:%d", file, line));
 290         WITNESS_CHECKORDER(&sx->lock_object, LOP_NEWORDER, file, line, NULL);
 291         error = __sx_slock(sx, opts, file, line);
 292         if (!error) {
 293                 LOCK_LOG_LOCK("SLOCK", &sx->lock_object, 0, 0, file, line);
 294                 WITNESS_LOCK(&sx->lock_object, 0, file, line);
 295                 curthread->td_locks++;
 296         }
 297
 298         return (error);
 299 }
 300
 301 int
 302 sx_try_slock_(struct sx *sx, const char *file, int line)
 303 {
 304         uintptr_t x;
 305
 306         if (SCHEDULER_STOPPED())
 307                 return (1);
 308
 309         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
 310             ("sx_try_slock() by idle thread %p on sx %s @ %s:%d",
 311             curthread, sx->lock_object.lo_name, file, line));
 312
 313         for (;;) {
 314                 x = sx->sx_lock;
 315                 KASSERT(x != SX_LOCK_DESTROYED,
 316                     ("sx_try_slock() of destroyed sx @ %s:%d", file, line));
 317                 if (!(x & SX_LOCK_SHARED))
 318                         break;
 319                 if (atomic_cmpset_acq_ptr(&sx->sx_lock, x, x + SX_ONE_SHARER)) {
 320                         LOCK_LOG_TRY("SLOCK", &sx->lock_object, 0, 1, file, line);
 321                         WITNESS_LOCK(&sx->lock_object, LOP_TRYLOCK, file, line);
 322                         LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_SX_SLOCK_ACQUIRE,
 323                             sx, 0, 0, file, line);
 324                         curthread->td_locks++;
 325                         return (1);
 326                 }
 327         }
 328
 329         LOCK_LOG_TRY("SLOCK", &sx->lock_object, 0, 0, file, line);
 330         return (0);
 331 }
 332
 333 int
 334 _sx_xlock(struct sx *sx, int opts, const char *file, int line)
 335 {
 336         int error = 0;
 337
 338         if (SCHEDULER_STOPPED())
 339                 return (0);
 340         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
 341             ("sx_xlock() by idle thread %p on sx %s @ %s:%d",
 342             curthread, sx->lock_object.lo_name, file, line));
 343         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 344             ("sx_xlock() of destroyed sx @ %s:%d", file, line));
 345         WITNESS_CHECKORDER(&sx->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE, file,
 346             line, NULL);
 347         error = __sx_xlock(sx, curthread, opts, file, line);
 348         if (!error) {
 349                 LOCK_LOG_LOCK("XLOCK", &sx->lock_object, 0, sx->sx_recurse,
 350                     file, line);
 351                 WITNESS_LOCK(&sx->lock_object, LOP_EXCLUSIVE, file, line);
 352                 curthread->td_locks++;
 353         }
 354
 355         return (error);
 356 }
 357
 358 int
 359 sx_try_xlock_(struct sx *sx, const char *file, int line)
 360 {
 361         int rval;
 362
 363         if (SCHEDULER_STOPPED())
 364                 return (1);
 365
 366         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
 367             ("sx_try_xlock() by idle thread %p on sx %s @ %s:%d",
 368             curthread, sx->lock_object.lo_name, file, line));
 369         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 370             ("sx_try_xlock() of destroyed sx @ %s:%d", file, line));
 371
 372         if (sx_xlocked(sx) &&
 373             (sx->lock_object.lo_flags & LO_RECURSABLE) != 0) {
 374                 sx->sx_recurse++;
 375                 atomic_set_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
 376                 rval = 1;
 377         } else
 378                 rval = atomic_cmpset_acq_ptr(&sx->sx_lock, SX_LOCK_UNLOCKED,
 379                     (uintptr_t)curthread);
 380         LOCK_LOG_TRY("XLOCK", &sx->lock_object, 0, rval, file, line);
 381         if (rval) {
 382                 WITNESS_LOCK(&sx->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
 383                     file, line);
 384                 if (!sx_recursed(sx))
 385                         LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_SX_XLOCK_ACQUIRE,
 386                             sx, 0, 0, file, line);
 387                 curthread->td_locks++;
 388         }
 389
 390         return (rval);
 391 }
 392
 393 void
 394 _sx_sunlock(struct sx *sx, const char *file, int line)
 395 {
 396
 397         if (SCHEDULER_STOPPED())
 398                 return;
 399         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 400             ("sx_sunlock() of destroyed sx @ %s:%d", file, line));
 401         _sx_assert(sx, SA_SLOCKED, file, line);
 402         WITNESS_UNLOCK(&sx->lock_object, 0, file, line);
 403         LOCK_LOG_LOCK("SUNLOCK", &sx->lock_object, 0, 0, file, line);
 404         __sx_sunlock(sx, file, line);
 405         LOCKSTAT_PROFILE_RELEASE_LOCK(LS_SX_SUNLOCK_RELEASE, sx);
 406         curthread->td_locks--;
 407 }
 408
 409 void
 410 _sx_xunlock(struct sx *sx, const char *file, int line)
 411 {
 412
 413         if (SCHEDULER_STOPPED())
 414                 return;
 415         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 416             ("sx_xunlock() of destroyed sx @ %s:%d", file, line));
 417         _sx_assert(sx, SA_XLOCKED, file, line);
 418         WITNESS_UNLOCK(&sx->lock_object, LOP_EXCLUSIVE, file, line);
 419         LOCK_LOG_LOCK("XUNLOCK", &sx->lock_object, 0, sx->sx_recurse, file,
 420             line);
 421         if (!sx_recursed(sx))
 422                 LOCKSTAT_PROFILE_RELEASE_LOCK(LS_SX_XUNLOCK_RELEASE, sx);
 423         __sx_xunlock(sx, curthread, file, line);
 424         curthread->td_locks--;
 425 }
 426
 427 /*
 428  * Try to do a non-blocking upgrade from a shared lock to an exclusive lock.
 429  * This will only succeed if this thread holds a single shared lock.
 430  * Return 1 if if the upgrade succeed, 0 otherwise.
 431  */
 432 int
 433 sx_try_upgrade_(struct sx *sx, const char *file, int line)
 434 {
 435         uintptr_t x;
 436         int success;
 437
 438         if (SCHEDULER_STOPPED())
 439                 return (1);
 440
 441         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 442             ("sx_try_upgrade() of destroyed sx @ %s:%d", file, line));
 443         _sx_assert(sx, SA_SLOCKED, file, line);
 444
 445         /*
 446          * Try to switch from one shared lock to an exclusive lock.  We need
 447          * to maintain the SX_LOCK_EXCLUSIVE_WAITERS flag if set so that
 448          * we will wake up the exclusive waiters when we drop the lock.
 449          */
 450         x = sx->sx_lock & SX_LOCK_EXCLUSIVE_WAITERS;
 451         success = atomic_cmpset_ptr(&sx->sx_lock, SX_SHARERS_LOCK(1) | x,
 452             (uintptr_t)curthread | x);
 453         LOCK_LOG_TRY("XUPGRADE", &sx->lock_object, 0, success, file, line);
 454         if (success) {
 455                 WITNESS_UPGRADE(&sx->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
 456                     file, line);
 457                 LOCKSTAT_RECORD0(LS_SX_TRYUPGRADE_UPGRADE, sx);
 458         }
 459         return (success);
 460 }
 461
 462 /*
 463  * Downgrade an unrecursed exclusive lock into a single shared lock.
 464  */
 465 void
 466 sx_downgrade_(struct sx *sx, const char *file, int line)
 467 {
 468         uintptr_t x;
 469         int wakeup_swapper;
 470
 471         if (SCHEDULER_STOPPED())
 472                 return;
 473
 474         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 475             ("sx_downgrade() of destroyed sx @ %s:%d", file, line));
 476         _sx_assert(sx, SA_XLOCKED | SA_NOTRECURSED, file, line);
 477 #ifndef INVARIANTS
 478         if (sx_recursed(sx))
 479                 panic("downgrade of a recursed lock");
 480 #endif
 481
 482         WITNESS_DOWNGRADE(&sx->lock_object, 0, file, line);
 483
 484         /*
 485          * Try to switch from an exclusive lock with no shared waiters
 486          * to one sharer with no shared waiters.  If there are
 487          * exclusive waiters, we don't need to lock the sleep queue so
 488          * long as we preserve the flag.  We do one quick try and if
 489          * that fails we grab the sleepq lock to keep the flags from
 490          * changing and do it the slow way.
 491          *
 492          * We have to lock the sleep queue if there are shared waiters
 493          * so we can wake them up.
 494          */
 495         x = sx->sx_lock;
 496         if (!(x & SX_LOCK_SHARED_WAITERS) &&
 497             atomic_cmpset_rel_ptr(&sx->sx_lock, x, SX_SHARERS_LOCK(1) |
 498             (x & SX_LOCK_EXCLUSIVE_WAITERS))) {
 499                 LOCK_LOG_LOCK("XDOWNGRADE", &sx->lock_object, 0, 0, file, line);
 500                 return;
 501         }
 502
 503         /*
 504          * Lock the sleep queue so we can read the waiters bits
 505          * without any races and wakeup any shared waiters.
 506          */
 507         sleepq_lock(&sx->lock_object);
 508
 509         /*
 510          * Preserve SX_LOCK_EXCLUSIVE_WAITERS while downgraded to a single
 511          * shared lock.  If there are any shared waiters, wake them up.
 512          */
 513         wakeup_swapper = 0;
 514         x = sx->sx_lock;
 515         atomic_store_rel_ptr(&sx->sx_lock, SX_SHARERS_LOCK(1) |
 516             (x & SX_LOCK_EXCLUSIVE_WAITERS));
 517         if (x & SX_LOCK_SHARED_WAITERS)
 518                 wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX,
 519                     0, SQ_SHARED_QUEUE);
 520         sleepq_release(&sx->lock_object);
 521
 522         LOCK_LOG_LOCK("XDOWNGRADE", &sx->lock_object, 0, 0, file, line);
 523         LOCKSTAT_RECORD0(LS_SX_DOWNGRADE_DOWNGRADE, sx);
 524
 525         if (wakeup_swapper)
 526                 kick_proc0();
 527 }
 528
 529 /*
 530  * This function represents the so-called 'hard case' for sx_xlock
 531  * operation.  All 'easy case' failures are redirected to this.  Note
 532  * that ideally this would be a static function, but it needs to be
 533  * accessible from at least sx.h.
 534  */
 535 int
 536 _sx_xlock_hard(struct sx *sx, uintptr_t tid, int opts, const char *file,
 537     int line)
 538 {
 539         GIANT_DECLARE;
 540 #ifdef ADAPTIVE_SX
 541         volatile struct thread *owner;
 542         u_int i, spintries = 0;
 543 #endif
 544         uintptr_t x;
 545 #ifdef LOCK_PROFILING
 546         uint64_t waittime = 0;
 547         int contested = 0;
 548 #endif
 549         int error = 0;
 550 #if defined(ADAPTIVE_SX) || defined(KDTRACE_HOOKS)
 551         struct lock_delay_arg lda;
 552 #endif
 553 #ifdef  KDTRACE_HOOKS
 554         uintptr_t state;
 555         u_int sleep_cnt = 0;
 556         int64_t sleep_time = 0;
 557         int64_t all_time = 0;
 558 #endif
 559
 560         if (SCHEDULER_STOPPED())
 561                 return (0);
 562
 563 #if defined(ADAPTIVE_SX)
 564         lock_delay_arg_init(&lda, &sx_delay);
 565 #elif defined(KDTRACE_HOOKS)
 566         lock_delay_arg_init(&lda, NULL);
 567 #endif
 568
 569         /* If we already hold an exclusive lock, then recurse. */
 570         if (sx_xlocked(sx)) {
 571                 KASSERT((sx->lock_object.lo_flags & LO_RECURSABLE) != 0,
 572             ("_sx_xlock_hard: recursed on non-recursive sx %s @ %s:%d\n",
 573                     sx->lock_object.lo_name, file, line));
 574                 sx->sx_recurse++;
 575                 atomic_set_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
 576                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 577                         CTR2(KTR_LOCK, "%s: %p recursing", __func__, sx);
 578                 return (0);
 579         }
 580
 581         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 582                 CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__,
 583                     sx->lock_object.lo_name, (void *)sx->sx_lock, file, line);
 584
 585 #ifdef KDTRACE_HOOKS
 586         all_time -= lockstat_nsecs(&sx->lock_object);
 587         state = sx->sx_lock;
 588 #endif
 589         for (;;) {
 590                 if (sx->sx_lock == SX_LOCK_UNLOCKED &&
 591                     atomic_cmpset_acq_ptr(&sx->sx_lock, SX_LOCK_UNLOCKED, tid))
 592                         break;
 593 #ifdef KDTRACE_HOOKS
 594                 lda.spin_cnt++;
 595 #endif
 596 #ifdef HWPMC_HOOKS
 597                 PMC_SOFT_CALL( , , lock, failed);
 598 #endif
 599                 lock_profile_obtain_lock_failed(&sx->lock_object, &contested,
 600                     &waittime);
 601 #ifdef ADAPTIVE_SX
 602                 /*
 603                  * If the lock is write locked and the owner is
 604                  * running on another CPU, spin until the owner stops
 605                  * running or the state of the lock changes.
 606                  */
 607                 x = sx->sx_lock;
 608                 if ((sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) {
 609                         if ((x & SX_LOCK_SHARED) == 0) {
 610                                 x = SX_OWNER(x);
 611                                 owner = (struct thread *)x;
 612                                 if (TD_IS_RUNNING(owner)) {
 613                                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 614                                                 CTR3(KTR_LOCK,
 615                                             "%s: spinning on %p held by %p",
 616                                                     __func__, sx, owner);
 617                                         KTR_STATE1(KTR_SCHED, "thread",
 618                                             sched_tdname(curthread), "spinning",
 619                                             "lockname:\"%s\"",
 620                                             sx->lock_object.lo_name);
 621                                         GIANT_SAVE();
 622                                         while (SX_OWNER(sx->sx_lock) == x &&
 623                                             TD_IS_RUNNING(owner))
 624                                                 lock_delay(&lda);
 625                                         KTR_STATE0(KTR_SCHED, "thread",
 626                                             sched_tdname(curthread), "running");
 627                                         continue;
 628                                 }
 629                         } else if (SX_SHARERS(x) && spintries < asx_retries) {
 630                                 KTR_STATE1(KTR_SCHED, "thread",
 631                                     sched_tdname(curthread), "spinning",
 632                                     "lockname:\"%s\"", sx->lock_object.lo_name);
 633                                 GIANT_SAVE();
 634                                 spintries++;
 635                                 for (i = 0; i < asx_loops; i++) {
 636                                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 637                                                 CTR4(KTR_LOCK,
 638                                     "%s: shared spinning on %p with %u and %u",
 639                                                     __func__, sx, spintries, i);
 640                                         x = sx->sx_lock;
 641                                         if ((x & SX_LOCK_SHARED) == 0 ||
 642                                             SX_SHARERS(x) == 0)
 643                                                 break;
 644                                         cpu_spinwait();
 645 #ifdef KDTRACE_HOOKS
 646                                         lda.spin_cnt++;
 647 #endif
 648                                 }
 649                                 KTR_STATE0(KTR_SCHED, "thread",
 650                                     sched_tdname(curthread), "running");
 651                                 if (i != asx_loops)
 652                                         continue;
 653                         }
 654                 }
 655 #endif
 656
 657                 sleepq_lock(&sx->lock_object);
 658                 x = sx->sx_lock;
 659
 660                 /*
 661                  * If the lock was released while spinning on the
 662                  * sleep queue chain lock, try again.
 663                  */
 664                 if (x == SX_LOCK_UNLOCKED) {
 665                         sleepq_release(&sx->lock_object);
 666                         continue;
 667                 }
 668
 669 #ifdef ADAPTIVE_SX
 670                 /*
 671                  * The current lock owner might have started executing
 672                  * on another CPU (or the lock could have changed
 673                  * owners) while we were waiting on the sleep queue
 674                  * chain lock.  If so, drop the sleep queue lock and try
 675                  * again.
 676                  */
 677                 if (!(x & SX_LOCK_SHARED) &&
 678                     (sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) {
 679                         owner = (struct thread *)SX_OWNER(x);
 680                         if (TD_IS_RUNNING(owner)) {
 681                                 sleepq_release(&sx->lock_object);
 682                                 continue;
 683                         }
 684                 }
 685 #endif
 686
 687                 /*
 688                  * If an exclusive lock was released with both shared
 689                  * and exclusive waiters and a shared waiter hasn't
 690                  * woken up and acquired the lock yet, sx_lock will be
 691                  * set to SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS.
 692                  * If we see that value, try to acquire it once.  Note
 693                  * that we have to preserve SX_LOCK_EXCLUSIVE_WAITERS
 694                  * as there are other exclusive waiters still.  If we
 695                  * fail, restart the loop.
 696                  */
 697                 if (x == (SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS)) {
 698                         if (atomic_cmpset_acq_ptr(&sx->sx_lock,
 699                             SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS,
 700                             tid | SX_LOCK_EXCLUSIVE_WAITERS)) {
 701                                 sleepq_release(&sx->lock_object);
 702                                 CTR2(KTR_LOCK, "%s: %p claimed by new writer",
 703                                     __func__, sx);
 704                                 break;
 705                         }
 706                         sleepq_release(&sx->lock_object);
 707                         continue;
 708                 }
 709
 710                 /*
 711                  * Try to set the SX_LOCK_EXCLUSIVE_WAITERS.  If we fail,
 712                  * than loop back and retry.
 713                  */
 714                 if (!(x & SX_LOCK_EXCLUSIVE_WAITERS)) {
 715                         if (!atomic_cmpset_ptr(&sx->sx_lock, x,
 716                             x | SX_LOCK_EXCLUSIVE_WAITERS)) {
 717                                 sleepq_release(&sx->lock_object);
 718                                 continue;
 719                         }
 720                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 721                                 CTR2(KTR_LOCK, "%s: %p set excl waiters flag",
 722                                     __func__, sx);
 723                 }
 724
 725                 /*
 726                  * Since we have been unable to acquire the exclusive
 727                  * lock and the exclusive waiters flag is set, we have
 728                  * to sleep.
 729                  */
 730                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 731                         CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
 732                             __func__, sx);
 733
 734 #ifdef KDTRACE_HOOKS
 735                 sleep_time -= lockstat_nsecs(&sx->lock_object);
 736 #endif
 737                 GIANT_SAVE();
 738                 sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
 739                     SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
 740                     SLEEPQ_INTERRUPTIBLE : 0), SQ_EXCLUSIVE_QUEUE);
 741                 if (!(opts & SX_INTERRUPTIBLE))
 742                         sleepq_wait(&sx->lock_object, 0);
 743                 else
 744                         error = sleepq_wait_sig(&sx->lock_object, 0);
 745 #ifdef KDTRACE_HOOKS
 746                 sleep_time += lockstat_nsecs(&sx->lock_object);
 747                 sleep_cnt++;
 748 #endif
 749                 if (error) {
 750                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 751                                 CTR2(KTR_LOCK,
 752                         "%s: interruptible sleep by %p suspended by signal",
 753                                     __func__, sx);
 754                         break;
 755                 }
 756                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 757                         CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
 758                             __func__, sx);
 759         }
 760 #ifdef KDTRACE_HOOKS
 761         all_time += lockstat_nsecs(&sx->lock_object);
 762         if (sleep_time)
 763                 LOCKSTAT_RECORD4(LS_SX_XLOCK_BLOCK, sx, sleep_time,
 764                     LOCKSTAT_WRITER, (state & SX_LOCK_SHARED) == 0,
 765                     (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
 766         if (lda.spin_cnt > sleep_cnt)
 767                 LOCKSTAT_RECORD4(LS_SX_XLOCK_SPIN, sx, all_time - sleep_time,
 768                     LOCKSTAT_WRITER, (state & SX_LOCK_SHARED) == 0,
 769                     (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
 770 #endif
 771         if (!error)
 772                 LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_SX_XLOCK_ACQUIRE, sx,
 773                     contested, waittime, file, line);
 774         GIANT_RESTORE();
 775         return (error);
 776 }
 777
 778 /*
 779  * This function represents the so-called 'hard case' for sx_xunlock
 780  * operation.  All 'easy case' failures are redirected to this.  Note
 781  * that ideally this would be a static function, but it needs to be
 782  * accessible from at least sx.h.
 783  */
 784 void
 785 _sx_xunlock_hard(struct sx *sx, uintptr_t tid, const char *file, int line)
 786 {
 787         uintptr_t x;
 788         int queue, wakeup_swapper;
 789
 790         if (SCHEDULER_STOPPED())
 791                 return;
 792
 793         MPASS(!(sx->sx_lock & SX_LOCK_SHARED));
 794
 795         /* If the lock is recursed, then unrecurse one level. */
 796         if (sx_xlocked(sx) && sx_recursed(sx)) {
 797                 if ((--sx->sx_recurse) == 0)
 798                         atomic_clear_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
 799                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 800                         CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, sx);
 801                 return;
 802         }
 803         MPASS(sx->sx_lock & (SX_LOCK_SHARED_WAITERS |
 804             SX_LOCK_EXCLUSIVE_WAITERS));
 805         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 806                 CTR2(KTR_LOCK, "%s: %p contested", __func__, sx);
 807
 808         sleepq_lock(&sx->lock_object);
 809         x = SX_LOCK_UNLOCKED;
 810
 811         /*
 812          * The wake up algorithm here is quite simple and probably not
 813          * ideal.  It gives precedence to shared waiters if they are
 814          * present.  For this condition, we have to preserve the
 815          * state of the exclusive waiters flag.
 816          * If interruptible sleeps left the shared queue empty avoid a
 817          * starvation for the threads sleeping on the exclusive queue by giving
 818          * them precedence and cleaning up the shared waiters bit anyway.
 819          */
 820         if ((sx->sx_lock & SX_LOCK_SHARED_WAITERS) != 0 &&
 821             sleepq_sleepcnt(&sx->lock_object, SQ_SHARED_QUEUE) != 0) {
 822                 queue = SQ_SHARED_QUEUE;
 823                 x |= (sx->sx_lock & SX_LOCK_EXCLUSIVE_WAITERS);
 824         } else
 825                 queue = SQ_EXCLUSIVE_QUEUE;
 826
 827         /* Wake up all the waiters for the specific queue. */
 828         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 829                 CTR3(KTR_LOCK, "%s: %p waking up all threads on %s queue",
 830                     __func__, sx, queue == SQ_SHARED_QUEUE ? "shared" :
 831                     "exclusive");
 832         atomic_store_rel_ptr(&sx->sx_lock, x);
 833         wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX, 0,
 834             queue);
 835         sleepq_release(&sx->lock_object);
 836         if (wakeup_swapper)
 837                 kick_proc0();
 838 }
 839
 840 /*
 841  * This function represents the so-called 'hard case' for sx_slock
 842  * operation.  All 'easy case' failures are redirected to this.  Note
 843  * that ideally this would be a static function, but it needs to be
 844  * accessible from at least sx.h.
 845  */
 846 int
 847 _sx_slock_hard(struct sx *sx, int opts, const char *file, int line)
 848 {
 849         GIANT_DECLARE;
 850 #ifdef ADAPTIVE_SX
 851         volatile struct thread *owner;
 852 #endif
 853 #ifdef LOCK_PROFILING
 854         uint64_t waittime = 0;
 855         int contested = 0;
 856 #endif
 857         uintptr_t x;
 858         int error = 0;
 859 #if defined(ADAPTIVE_SX) || defined(KDTRACE_HOOKS)
 860         struct lock_delay_arg lda;
 861 #endif
 862 #ifdef KDTRACE_HOOKS
 863         uintptr_t state;
 864         u_int sleep_cnt = 0;
 865         int64_t sleep_time = 0;
 866         int64_t all_time = 0;
 867 #endif
 868
 869         if (SCHEDULER_STOPPED())
 870                 return (0);
 871
 872 #if defined(ADAPTIVE_SX)
 873         lock_delay_arg_init(&lda, &sx_delay);
 874 #elif defined(KDTRACE_HOOKS)
 875         lock_delay_arg_init(&lda, NULL);
 876 #endif
 877 #ifdef KDTRACE_HOOKS
 878         state = sx->sx_lock;
 879         all_time -= lockstat_nsecs(&sx->lock_object);
 880 #endif
 881
 882         /*
 883          * As with rwlocks, we don't make any attempt to try to block
 884          * shared locks once there is an exclusive waiter.
 885          */
 886         for (;;) {
 887 #ifdef KDTRACE_HOOKS
 888                 lda.spin_cnt++;
 889 #endif
 890                 x = sx->sx_lock;
 891
 892                 /*
 893                  * If no other thread has an exclusive lock then try to bump up
 894                  * the count of sharers.  Since we have to preserve the state
 895                  * of SX_LOCK_EXCLUSIVE_WAITERS, if we fail to acquire the
 896                  * shared lock loop back and retry.
 897                  */
 898                 if (x & SX_LOCK_SHARED) {
 899                         MPASS(!(x & SX_LOCK_SHARED_WAITERS));
 900                         if (atomic_cmpset_acq_ptr(&sx->sx_lock, x,
 901                             x + SX_ONE_SHARER)) {
 902                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 903                                         CTR4(KTR_LOCK,
 904                                             "%s: %p succeed %p -> %p", __func__,
 905                                             sx, (void *)x,
 906                                             (void *)(x + SX_ONE_SHARER));
 907                                 break;
 908                         }
 909                         continue;
 910                 }
 911 #ifdef HWPMC_HOOKS
 912                 PMC_SOFT_CALL( , , lock, failed);
 913 #endif
 914                 lock_profile_obtain_lock_failed(&sx->lock_object, &contested,
 915                     &waittime);
 916
 917 #ifdef ADAPTIVE_SX
 918                 /*
 919                  * If the owner is running on another CPU, spin until
 920                  * the owner stops running or the state of the lock
 921                  * changes.
 922                  */
 923                 if ((sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) {
 924                         x = SX_OWNER(x);
 925                         owner = (struct thread *)x;
 926                         if (TD_IS_RUNNING(owner)) {
 927                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 928                                         CTR3(KTR_LOCK,
 929                                             "%s: spinning on %p held by %p",
 930                                             __func__, sx, owner);
 931                                 KTR_STATE1(KTR_SCHED, "thread",
 932                                     sched_tdname(curthread), "spinning",
 933                                     "lockname:\"%s\"", sx->lock_object.lo_name);
 934                                 GIANT_SAVE();
 935                                 while (SX_OWNER(sx->sx_lock) == x &&
 936                                     TD_IS_RUNNING(owner))
 937                                         lock_delay(&lda);
 938                                 KTR_STATE0(KTR_SCHED, "thread",
 939                                     sched_tdname(curthread), "running");
 940                                 continue;
 941                         }
 942                 }
 943 #endif
 944
 945                 /*
 946                  * Some other thread already has an exclusive lock, so
 947                  * start the process of blocking.
 948                  */
 949                 sleepq_lock(&sx->lock_object);
 950                 x = sx->sx_lock;
 951
 952                 /*
 953                  * The lock could have been released while we spun.
 954                  * In this case loop back and retry.
 955                  */
 956                 if (x & SX_LOCK_SHARED) {
 957                         sleepq_release(&sx->lock_object);
 958                         continue;
 959                 }
 960
 961 #ifdef ADAPTIVE_SX
 962                 /*
 963                  * If the owner is running on another CPU, spin until
 964                  * the owner stops running or the state of the lock
 965                  * changes.
 966                  */
 967                 if (!(x & SX_LOCK_SHARED) &&
 968                     (sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) {
 969                         owner = (struct thread *)SX_OWNER(x);
 970                         if (TD_IS_RUNNING(owner)) {
 971                                 sleepq_release(&sx->lock_object);
 972                                 continue;
 973                         }
 974                 }
 975 #endif
 976
 977                 /*
 978                  * Try to set the SX_LOCK_SHARED_WAITERS flag.  If we
 979                  * fail to set it drop the sleep queue lock and loop
 980                  * back.
 981                  */
 982                 if (!(x & SX_LOCK_SHARED_WAITERS)) {
 983                         if (!atomic_cmpset_ptr(&sx->sx_lock, x,
 984                             x | SX_LOCK_SHARED_WAITERS)) {
 985                                 sleepq_release(&sx->lock_object);
 986                                 continue;
 987                         }
 988                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 989                                 CTR2(KTR_LOCK, "%s: %p set shared waiters flag",
 990                                     __func__, sx);
 991                 }
 992
 993                 /*
 994                  * Since we have been unable to acquire the shared lock,
 995                  * we have to sleep.
 996                  */
 997                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 998                         CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
 999                             __func__, sx);
1000
1001 #ifdef KDTRACE_HOOKS
1002                 sleep_time -= lockstat_nsecs(&sx->lock_object);
1003 #endif
1004                 GIANT_SAVE();
1005                 sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
1006                     SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
1007                     SLEEPQ_INTERRUPTIBLE : 0), SQ_SHARED_QUEUE);
1008                 if (!(opts & SX_INTERRUPTIBLE))
1009                         sleepq_wait(&sx->lock_object, 0);
1010                 else
1011                         error = sleepq_wait_sig(&sx->lock_object, 0);
1012 #ifdef KDTRACE_HOOKS
1013                 sleep_time += lockstat_nsecs(&sx->lock_object);
1014                 sleep_cnt++;
1015 #endif
1016                 if (error) {
1017                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
1018                                 CTR2(KTR_LOCK,
1019                         "%s: interruptible sleep by %p suspended by signal",
1020                                     __func__, sx);
1021                         break;
1022                 }
1023                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1024                         CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
1025                             __func__, sx);
1026         }
1027 #ifdef KDTRACE_HOOKS
1028         all_time += lockstat_nsecs(&sx->lock_object);
1029         if (sleep_time)
1030                 LOCKSTAT_RECORD4(LS_SX_SLOCK_BLOCK, sx, sleep_time,
1031                     LOCKSTAT_READER, (state & SX_LOCK_SHARED) == 0,
1032                     (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
1033         if (lda.spin_cnt > sleep_cnt)
1034                 LOCKSTAT_RECORD4(LS_SX_SLOCK_SPIN, sx, all_time - sleep_time,
1035                     LOCKSTAT_READER, (state & SX_LOCK_SHARED) == 0,
1036                     (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
1037 #endif
1038         if (error == 0)
1039                 LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_SX_SLOCK_ACQUIRE, sx,
1040                     contested, waittime, file, line);
1041         GIANT_RESTORE();
1042         return (error);
1043 }
1044
1045 /*
1046  * This function represents the so-called 'hard case' for sx_sunlock
1047  * operation.  All 'easy case' failures are redirected to this.  Note
1048  * that ideally this would be a static function, but it needs to be
1049  * accessible from at least sx.h.
1050  */
1051 void
1052 _sx_sunlock_hard(struct sx *sx, const char *file, int line)
1053 {
1054         uintptr_t x;
1055         int wakeup_swapper;
1056
1057         if (SCHEDULER_STOPPED())
1058                 return;
1059
1060         for (;;) {
1061                 x = sx->sx_lock;
1062
1063                 /*
1064                  * We should never have sharers while at least one thread
1065                  * holds a shared lock.
1066                  */
1067                 KASSERT(!(x & SX_LOCK_SHARED_WAITERS),
1068                     ("%s: waiting sharers", __func__));
1069
1070                 /*
1071                  * See if there is more than one shared lock held.  If
1072                  * so, just drop one and return.
1073                  */
1074                 if (SX_SHARERS(x) > 1) {
1075                         if (atomic_cmpset_rel_ptr(&sx->sx_lock, x,
1076                             x - SX_ONE_SHARER)) {
1077                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1078                                         CTR4(KTR_LOCK,
1079                                             "%s: %p succeeded %p -> %p",
1080                                             __func__, sx, (void *)x,
1081                                             (void *)(x - SX_ONE_SHARER));
1082                                 break;
1083                         }
1084                         continue;
1085                 }
1086
1087                 /*
1088                  * If there aren't any waiters for an exclusive lock,
1089                  * then try to drop it quickly.
1090                  */
1091                 if (!(x & SX_LOCK_EXCLUSIVE_WAITERS)) {
1092                         MPASS(x == SX_SHARERS_LOCK(1));
1093                         if (atomic_cmpset_rel_ptr(&sx->sx_lock,
1094                             SX_SHARERS_LOCK(1), SX_LOCK_UNLOCKED)) {
1095                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1096                                         CTR2(KTR_LOCK, "%s: %p last succeeded",
1097                                             __func__, sx);
1098                                 break;
1099                         }
1100                         continue;
1101                 }
1102
1103                 /*
1104                  * At this point, there should just be one sharer with
1105                  * exclusive waiters.
1106                  */
1107                 MPASS(x == (SX_SHARERS_LOCK(1) | SX_LOCK_EXCLUSIVE_WAITERS));
1108
1109                 sleepq_lock(&sx->lock_object);
1110
1111                 /*
1112                  * Wake up semantic here is quite simple:
1113                  * Just wake up all the exclusive waiters.
1114                  * Note that the state of the lock could have changed,
1115                  * so if it fails loop back and retry.
1116                  */
1117                 if (!atomic_cmpset_rel_ptr(&sx->sx_lock,
1118                     SX_SHARERS_LOCK(1) | SX_LOCK_EXCLUSIVE_WAITERS,
1119                     SX_LOCK_UNLOCKED)) {
1120                         sleepq_release(&sx->lock_object);
1121                         continue;
1122                 }
1123                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1124                         CTR2(KTR_LOCK, "%s: %p waking up all thread on"
1125                             "exclusive queue", __func__, sx);
1126                 wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX,
1127                     0, SQ_EXCLUSIVE_QUEUE);
1128                 sleepq_release(&sx->lock_object);
1129                 if (wakeup_swapper)
1130                         kick_proc0();
1131                 break;
1132         }
1133 }
1134
1135 #ifdef INVARIANT_SUPPORT
1136 #ifndef INVARIANTS
1137 #undef  _sx_assert
1138 #endif
1139
1140 /*
1141  * In the non-WITNESS case, sx_assert() can only detect that at least
1142  * *some* thread owns an slock, but it cannot guarantee that *this*
1143  * thread owns an slock.
1144  */
1145 void
1146 _sx_assert(const struct sx *sx, int what, const char *file, int line)
1147 {
1148 #ifndef WITNESS
1149         int slocked = 0;
1150 #endif
1151
1152         if (panicstr != NULL)
1153                 return;
1154         switch (what) {
1155         case SA_SLOCKED:
1156         case SA_SLOCKED | SA_NOTRECURSED:
1157         case SA_SLOCKED | SA_RECURSED:
1158 #ifndef WITNESS
1159                 slocked = 1;
1160                 /* FALLTHROUGH */
1161 #endif
1162         case SA_LOCKED:
1163         case SA_LOCKED | SA_NOTRECURSED:
1164         case SA_LOCKED | SA_RECURSED:
1165 #ifdef WITNESS
1166                 witness_assert(&sx->lock_object, what, file, line);
1167 #else
1168                 /*
1169                  * If some other thread has an exclusive lock or we
1170                  * have one and are asserting a shared lock, fail.
1171                  * Also, if no one has a lock at all, fail.
1172                  */
1173                 if (sx->sx_lock == SX_LOCK_UNLOCKED ||
1174                     (!(sx->sx_lock & SX_LOCK_SHARED) && (slocked ||
1175                     sx_xholder(sx) != curthread)))
1176                         panic("Lock %s not %slocked @ %s:%d\n",
1177                             sx->lock_object.lo_name, slocked ? "share " : "",
1178                             file, line);
1179
1180                 if (!(sx->sx_lock & SX_LOCK_SHARED)) {
1181                         if (sx_recursed(sx)) {
1182                                 if (what & SA_NOTRECURSED)
1183                                         panic("Lock %s recursed @ %s:%d\n",
1184                                             sx->lock_object.lo_name, file,
1185                                             line);
1186                         } else if (what & SA_RECURSED)
1187                                 panic("Lock %s not recursed @ %s:%d\n",
1188                                     sx->lock_object.lo_name, file, line);
1189                 }
1190 #endif
1191                 break;
1192         case SA_XLOCKED:
1193         case SA_XLOCKED | SA_NOTRECURSED:
1194         case SA_XLOCKED | SA_RECURSED:
1195                 if (sx_xholder(sx) != curthread)
1196                         panic("Lock %s not exclusively locked @ %s:%d\n",
1197                             sx->lock_object.lo_name, file, line);
1198                 if (sx_recursed(sx)) {
1199                         if (what & SA_NOTRECURSED)
1200                                 panic("Lock %s recursed @ %s:%d\n",
1201                                     sx->lock_object.lo_name, file, line);
1202                 } else if (what & SA_RECURSED)
1203                         panic("Lock %s not recursed @ %s:%d\n",
1204                             sx->lock_object.lo_name, file, line);
1205                 break;
1206         case SA_UNLOCKED:
1207 #ifdef WITNESS
1208                 witness_assert(&sx->lock_object, what, file, line);
1209 #else
1210                 /*
1211                  * If we hold an exclusve lock fail.  We can't
1212                  * reliably check to see if we hold a shared lock or
1213                  * not.
1214                  */
1215                 if (sx_xholder(sx) == curthread)
1216                         panic("Lock %s exclusively locked @ %s:%d\n",
1217                             sx->lock_object.lo_name, file, line);
1218 #endif
1219                 break;
1220         default:
1221                 panic("Unknown sx lock assertion: %d @ %s:%d", what, file,
1222                     line);
1223         }
1224 }
1225 #endif  /* INVARIANT_SUPPORT */
1226
1227 #ifdef DDB
1228 static void
1229 db_show_sx(const struct lock_object *lock)
1230 {
1231         struct thread *td;
1232         const struct sx *sx;
1233
1234         sx = (const struct sx *)lock;
1235
1236         db_printf(" state: ");
1237         if (sx->sx_lock == SX_LOCK_UNLOCKED)
1238                 db_printf("UNLOCKED\n");
1239         else if (sx->sx_lock == SX_LOCK_DESTROYED) {
1240                 db_printf("DESTROYED\n");
1241                 return;
1242         } else if (sx->sx_lock & SX_LOCK_SHARED)
1243                 db_printf("SLOCK: %ju\n", (uintmax_t)SX_SHARERS(sx->sx_lock));
1244         else {
1245                 td = sx_xholder(sx);
1246                 db_printf("XLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1247                     td->td_tid, td->td_proc->p_pid, td->td_name);
1248                 if (sx_recursed(sx))
1249                         db_printf(" recursed: %d\n", sx->sx_recurse);
1250         }
1251
1252         db_printf(" waiters: ");
1253         switch(sx->sx_lock &
1254             (SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS)) {
1255         case SX_LOCK_SHARED_WAITERS:
1256                 db_printf("shared\n");
1257                 break;
1258         case SX_LOCK_EXCLUSIVE_WAITERS:
1259                 db_printf("exclusive\n");
1260                 break;
1261         case SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS:
1262                 db_printf("exclusive and shared\n");
1263                 break;
1264         default:
1265                 db_printf("none\n");
1266         }
1267 }
1268
1269 /*
1270  * Check to see if a thread that is blocked on a sleep queue is actually
1271  * blocked on an sx lock.  If so, output some details and return true.
1272  * If the lock has an exclusive owner, return that in *ownerp.
1273  */
1274 int
1275 sx_chain(struct thread *td, struct thread **ownerp)
1276 {
1277         struct sx *sx;
1278
1279         /*
1280          * Check to see if this thread is blocked on an sx lock.
1281          * First, we check the lock class.  If that is ok, then we
1282          * compare the lock name against the wait message.
1283          */
1284         sx = td->td_wchan;
1285         if (LOCK_CLASS(&sx->lock_object) != &lock_class_sx ||
1286             sx->lock_object.lo_name != td->td_wmesg)
1287                 return (0);
1288
1289         /* We think we have an sx lock, so output some details. */
1290         db_printf("blocked on sx \"%s\" ", td->td_wmesg);
1291         *ownerp = sx_xholder(sx);
1292         if (sx->sx_lock & SX_LOCK_SHARED)
1293                 db_printf("SLOCK (count %ju)\n",
1294                     (uintmax_t)SX_SHARERS(sx->sx_lock));
1295         else
1296                 db_printf("XLOCK\n");
1297         return (1);
1298 }
1299 #endif