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