sys/kern/kern_sx.c

   1 /*-
   2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
   3  *
   4  * Copyright (c) 2007 Attilio Rao <attilio@freebsd.org>
   5  * Copyright (c) 2001 Jason Evans <jasone@freebsd.org>
   6  * All rights reserved.
   7  *
   8  * Redistribution and use in source and binary forms, with or without
   9  * modification, are permitted provided that the following conditions
  10  * are met:
  11  * 1. Redistributions of source code must retain the above copyright
  12  *    notice(s), this list of conditions and the following disclaimer as
  13  *    the first lines of this file unmodified other than the possible
  14  *    addition of one or more copyright notices.
  15  * 2. Redistributions in binary form must reproduce the above copyright
  16  *    notice(s), this list of conditions and the following disclaimer in the
  17  *    documentation and/or other materials provided with the distribution.
  18  *
  19  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDER(S) ``AS IS'' AND ANY
  20  * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
  21  * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
  22  * DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT HOLDER(S) BE LIABLE FOR ANY
  23  * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
  24  * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
  25  * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
  26  * CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  27  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  28  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH
  29  * DAMAGE.
  30  */
  31
  32 /*
  33  * Shared/exclusive locks.  This implementation attempts to ensure
  34  * deterministic lock granting behavior, so that slocks and xlocks are
  35  * interleaved.
  36  *
  37  * Priority propagation will not generally raise the priority of lock holders,
  38  * so should not be relied upon in combination with sx locks.
  39  */
  40
  41 #include "opt_ddb.h"
  42 #include "opt_hwpmc_hooks.h"
  43 #include "opt_no_adaptive_sx.h"
  44
  45 #include <sys/cdefs.h>
  46 __FBSDID("$FreeBSD$");
  47
  48 #include <sys/param.h>
  49 #include <sys/systm.h>
  50 #include <sys/kdb.h>
  51 #include <sys/kernel.h>
  52 #include <sys/ktr.h>
  53 #include <sys/lock.h>
  54 #include <sys/mutex.h>
  55 #include <sys/proc.h>
  56 #include <sys/sched.h>
  57 #include <sys/sleepqueue.h>
  58 #include <sys/sx.h>
  59 #include <sys/smp.h>
  60 #include <sys/sysctl.h>
  61
  62 #if defined(SMP) && !defined(NO_ADAPTIVE_SX)
  63 #include <machine/cpu.h>
  64 #endif
  65
  66 #ifdef DDB
  67 #include <ddb/ddb.h>
  68 #endif
  69
  70 #if defined(SMP) && !defined(NO_ADAPTIVE_SX)
  71 #define ADAPTIVE_SX
  72 #endif
  73
  74 CTASSERT((SX_NOADAPTIVE & LO_CLASSFLAGS) == SX_NOADAPTIVE);
  75
  76 #ifdef HWPMC_HOOKS
  77 #include <sys/pmckern.h>
  78 PMC_SOFT_DECLARE( , , lock, failed);
  79 #endif
  80
  81 /* Handy macros for sleep queues. */
  82 #define SQ_EXCLUSIVE_QUEUE      0
  83 #define SQ_SHARED_QUEUE         1
  84
  85 /*
  86  * Variations on DROP_GIANT()/PICKUP_GIANT() for use in this file.  We
  87  * drop Giant anytime we have to sleep or if we adaptively spin.
  88  */
  89 #define GIANT_DECLARE                                                   \
  90         int _giantcnt = 0;                                              \
  91         WITNESS_SAVE_DECL(Giant)                                        \
  92
  93 #define GIANT_SAVE(work) do {                                           \
  94         if (__predict_false(mtx_owned(&Giant))) {                       \
  95                 work++;                                                 \
  96                 WITNESS_SAVE(&Giant.lock_object, Giant);                \
  97                 while (mtx_owned(&Giant)) {                             \
  98                         _giantcnt++;                                    \
  99                         mtx_unlock(&Giant);                             \
 100                 }                                                       \
 101         }                                                               \
 102 } while (0)
 103
 104 #define GIANT_RESTORE() do {                                            \
 105         if (_giantcnt > 0) {                                            \
 106                 mtx_assert(&Giant, MA_NOTOWNED);                        \
 107                 while (_giantcnt--)                                     \
 108                         mtx_lock(&Giant);                               \
 109                 WITNESS_RESTORE(&Giant.lock_object, Giant);             \
 110         }                                                               \
 111 } while (0)
 112
 113 /*
 114  * Returns true if an exclusive lock is recursed.  It assumes
 115  * curthread currently has an exclusive lock.
 116  */
 117 #define sx_recursed(sx)         ((sx)->sx_recurse != 0)
 118
 119 static void     assert_sx(const struct lock_object *lock, int what);
 120 #ifdef DDB
 121 static void     db_show_sx(const struct lock_object *lock);
 122 #endif
 123 static void     lock_sx(struct lock_object *lock, uintptr_t how);
 124 #ifdef KDTRACE_HOOKS
 125 static int      owner_sx(const struct lock_object *lock, struct thread **owner);
 126 #endif
 127 static uintptr_t unlock_sx(struct lock_object *lock);
 128
 129 struct lock_class lock_class_sx = {
 130         .lc_name = "sx",
 131         .lc_flags = LC_SLEEPLOCK | LC_SLEEPABLE | LC_RECURSABLE | LC_UPGRADABLE,
 132         .lc_assert = assert_sx,
 133 #ifdef DDB
 134         .lc_ddb_show = db_show_sx,
 135 #endif
 136         .lc_lock = lock_sx,
 137         .lc_unlock = unlock_sx,
 138 #ifdef KDTRACE_HOOKS
 139         .lc_owner = owner_sx,
 140 #endif
 141 };
 142
 143 #ifndef INVARIANTS
 144 #define _sx_assert(sx, what, file, line)
 145 #endif
 146
 147 #ifdef ADAPTIVE_SX
 148 static __read_frequently u_int asx_retries;
 149 static __read_frequently u_int asx_loops;
 150 static SYSCTL_NODE(_debug, OID_AUTO, sx, CTLFLAG_RD, NULL, "sxlock debugging");
 151 SYSCTL_UINT(_debug_sx, OID_AUTO, retries, CTLFLAG_RW, &asx_retries, 0, "");
 152 SYSCTL_UINT(_debug_sx, OID_AUTO, loops, CTLFLAG_RW, &asx_loops, 0, "");
 153
 154 static struct lock_delay_config __read_frequently sx_delay;
 155
 156 SYSCTL_INT(_debug_sx, OID_AUTO, delay_base, CTLFLAG_RW, &sx_delay.base,
 157     0, "");
 158 SYSCTL_INT(_debug_sx, OID_AUTO, delay_max, CTLFLAG_RW, &sx_delay.max,
 159     0, "");
 160
 161 static void
 162 sx_lock_delay_init(void *arg __unused)
 163 {
 164
 165         lock_delay_default_init(&sx_delay);
 166         asx_retries = 10;
 167         asx_loops = max(10000, sx_delay.max);
 168 }
 169 LOCK_DELAY_SYSINIT(sx_lock_delay_init);
 170 #endif
 171
 172 void
 173 assert_sx(const struct lock_object *lock, int what)
 174 {
 175
 176         sx_assert((const struct sx *)lock, what);
 177 }
 178
 179 void
 180 lock_sx(struct lock_object *lock, uintptr_t how)
 181 {
 182         struct sx *sx;
 183
 184         sx = (struct sx *)lock;
 185         if (how)
 186                 sx_slock(sx);
 187         else
 188                 sx_xlock(sx);
 189 }
 190
 191 uintptr_t
 192 unlock_sx(struct lock_object *lock)
 193 {
 194         struct sx *sx;
 195
 196         sx = (struct sx *)lock;
 197         sx_assert(sx, SA_LOCKED | SA_NOTRECURSED);
 198         if (sx_xlocked(sx)) {
 199                 sx_xunlock(sx);
 200                 return (0);
 201         } else {
 202                 sx_sunlock(sx);
 203                 return (1);
 204         }
 205 }
 206
 207 #ifdef KDTRACE_HOOKS
 208 int
 209 owner_sx(const struct lock_object *lock, struct thread **owner)
 210 {
 211         const struct sx *sx;
 212         uintptr_t x;
 213
 214         sx = (const struct sx *)lock;
 215         x = sx->sx_lock;
 216         *owner = NULL;
 217         return ((x & SX_LOCK_SHARED) != 0 ? (SX_SHARERS(x) != 0) :
 218             ((*owner = (struct thread *)SX_OWNER(x)) != NULL));
 219 }
 220 #endif
 221
 222 void
 223 sx_sysinit(void *arg)
 224 {
 225         struct sx_args *sargs = arg;
 226
 227         sx_init_flags(sargs->sa_sx, sargs->sa_desc, sargs->sa_flags);
 228 }
 229
 230 void
 231 sx_init_flags(struct sx *sx, const char *description, int opts)
 232 {
 233         int flags;
 234
 235         MPASS((opts & ~(SX_QUIET | SX_RECURSE | SX_NOWITNESS | SX_DUPOK |
 236             SX_NOPROFILE | SX_NOADAPTIVE | SX_NEW)) == 0);
 237         ASSERT_ATOMIC_LOAD_PTR(sx->sx_lock,
 238             ("%s: sx_lock not aligned for %s: %p", __func__, description,
 239             &sx->sx_lock));
 240
 241         flags = LO_SLEEPABLE | LO_UPGRADABLE;
 242         if (opts & SX_DUPOK)
 243                 flags |= LO_DUPOK;
 244         if (opts & SX_NOPROFILE)
 245                 flags |= LO_NOPROFILE;
 246         if (!(opts & SX_NOWITNESS))
 247                 flags |= LO_WITNESS;
 248         if (opts & SX_RECURSE)
 249                 flags |= LO_RECURSABLE;
 250         if (opts & SX_QUIET)
 251                 flags |= LO_QUIET;
 252         if (opts & SX_NEW)
 253                 flags |= LO_NEW;
 254
 255         flags |= opts & SX_NOADAPTIVE;
 256         lock_init(&sx->lock_object, &lock_class_sx, description, NULL, flags);
 257         sx->sx_lock = SX_LOCK_UNLOCKED;
 258         sx->sx_recurse = 0;
 259 }
 260
 261 void
 262 sx_destroy(struct sx *sx)
 263 {
 264
 265         KASSERT(sx->sx_lock == SX_LOCK_UNLOCKED, ("sx lock still held"));
 266         KASSERT(sx->sx_recurse == 0, ("sx lock still recursed"));
 267         sx->sx_lock = SX_LOCK_DESTROYED;
 268         lock_destroy(&sx->lock_object);
 269 }
 270
 271 int
 272 sx_try_slock_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
 273 {
 274         uintptr_t x;
 275
 276         if (SCHEDULER_STOPPED())
 277                 return (1);
 278
 279         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
 280             ("sx_try_slock() by idle thread %p on sx %s @ %s:%d",
 281             curthread, sx->lock_object.lo_name, file, line));
 282
 283         x = sx->sx_lock;
 284         for (;;) {
 285                 KASSERT(x != SX_LOCK_DESTROYED,
 286                     ("sx_try_slock() of destroyed sx @ %s:%d", file, line));
 287                 if (!(x & SX_LOCK_SHARED))
 288                         break;
 289                 if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, x + SX_ONE_SHARER)) {
 290                         LOCK_LOG_TRY("SLOCK", &sx->lock_object, 0, 1, file, line);
 291                         WITNESS_LOCK(&sx->lock_object, LOP_TRYLOCK, file, line);
 292                         LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire,
 293                             sx, 0, 0, file, line, LOCKSTAT_READER);
 294                         TD_LOCKS_INC(curthread);
 295                         curthread->td_sx_slocks++;
 296                         return (1);
 297                 }
 298         }
 299
 300         LOCK_LOG_TRY("SLOCK", &sx->lock_object, 0, 0, file, line);
 301         return (0);
 302 }
 303
 304 int
 305 sx_try_slock_(struct sx *sx, const char *file, int line)
 306 {
 307
 308         return (sx_try_slock_int(sx LOCK_FILE_LINE_ARG));
 309 }
 310
 311 int
 312 _sx_xlock(struct sx *sx, int opts, const char *file, int line)
 313 {
 314         uintptr_t tid, x;
 315         int error = 0;
 316
 317         KASSERT(kdb_active != 0 || SCHEDULER_STOPPED() ||
 318             !TD_IS_IDLETHREAD(curthread),
 319             ("sx_xlock() by idle thread %p on sx %s @ %s:%d",
 320             curthread, sx->lock_object.lo_name, file, line));
 321         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 322             ("sx_xlock() of destroyed sx @ %s:%d", file, line));
 323         WITNESS_CHECKORDER(&sx->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE, file,
 324             line, NULL);
 325         tid = (uintptr_t)curthread;
 326         x = SX_LOCK_UNLOCKED;
 327         if (!atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid))
 328                 error = _sx_xlock_hard(sx, x, opts LOCK_FILE_LINE_ARG);
 329         else
 330                 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
 331                     0, 0, file, line, LOCKSTAT_WRITER);
 332         if (!error) {
 333                 LOCK_LOG_LOCK("XLOCK", &sx->lock_object, 0, sx->sx_recurse,
 334                     file, line);
 335                 WITNESS_LOCK(&sx->lock_object, LOP_EXCLUSIVE, file, line);
 336                 TD_LOCKS_INC(curthread);
 337         }
 338
 339         return (error);
 340 }
 341
 342 int
 343 sx_try_xlock_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
 344 {
 345         struct thread *td;
 346         uintptr_t tid, x;
 347         int rval;
 348         bool recursed;
 349
 350         td = curthread;
 351         tid = (uintptr_t)td;
 352         if (SCHEDULER_STOPPED_TD(td))
 353                 return (1);
 354
 355         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(td),
 356             ("sx_try_xlock() by idle thread %p on sx %s @ %s:%d",
 357             curthread, sx->lock_object.lo_name, file, line));
 358         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 359             ("sx_try_xlock() of destroyed sx @ %s:%d", file, line));
 360
 361         rval = 1;
 362         recursed = false;
 363         x = SX_LOCK_UNLOCKED;
 364         for (;;) {
 365                 if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid))
 366                         break;
 367                 if (x == SX_LOCK_UNLOCKED)
 368                         continue;
 369                 if (x == tid && (sx->lock_object.lo_flags & LO_RECURSABLE)) {
 370                         sx->sx_recurse++;
 371                         atomic_set_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
 372                         break;
 373                 }
 374                 rval = 0;
 375                 break;
 376         }
 377
 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 (!recursed)
 383                         LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire,
 384                             sx, 0, 0, file, line, LOCKSTAT_WRITER);
 385                 TD_LOCKS_INC(curthread);
 386         }
 387
 388         return (rval);
 389 }
 390
 391 int
 392 sx_try_xlock_(struct sx *sx, const char *file, int line)
 393 {
 394
 395         return (sx_try_xlock_int(sx LOCK_FILE_LINE_ARG));
 396 }
 397
 398 void
 399 _sx_xunlock(struct sx *sx, const char *file, int line)
 400 {
 401
 402         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 403             ("sx_xunlock() of destroyed sx @ %s:%d", file, line));
 404         _sx_assert(sx, SA_XLOCKED, file, line);
 405         WITNESS_UNLOCK(&sx->lock_object, LOP_EXCLUSIVE, file, line);
 406         LOCK_LOG_LOCK("XUNLOCK", &sx->lock_object, 0, sx->sx_recurse, file,
 407             line);
 408 #if LOCK_DEBUG > 0
 409         _sx_xunlock_hard(sx, (uintptr_t)curthread, file, line);
 410 #else
 411         __sx_xunlock(sx, curthread, file, line);
 412 #endif
 413         TD_LOCKS_DEC(curthread);
 414 }
 415
 416 /*
 417  * Try to do a non-blocking upgrade from a shared lock to an exclusive lock.
 418  * This will only succeed if this thread holds a single shared lock.
 419  * Return 1 if if the upgrade succeed, 0 otherwise.
 420  */
 421 int
 422 sx_try_upgrade_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
 423 {
 424         uintptr_t x;
 425         uintptr_t waiters;
 426         int success;
 427
 428         if (SCHEDULER_STOPPED())
 429                 return (1);
 430
 431         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 432             ("sx_try_upgrade() of destroyed sx @ %s:%d", file, line));
 433         _sx_assert(sx, SA_SLOCKED, file, line);
 434
 435         /*
 436          * Try to switch from one shared lock to an exclusive lock.  We need
 437          * to maintain the SX_LOCK_EXCLUSIVE_WAITERS flag if set so that
 438          * we will wake up the exclusive waiters when we drop the lock.
 439          */
 440         success = 0;
 441         x = SX_READ_VALUE(sx);
 442         for (;;) {
 443                 if (SX_SHARERS(x) > 1)
 444                         break;
 445                 waiters = (x & SX_LOCK_WAITERS);
 446                 if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x,
 447                     (uintptr_t)curthread | waiters)) {
 448                         success = 1;
 449                         break;
 450                 }
 451         }
 452         LOCK_LOG_TRY("XUPGRADE", &sx->lock_object, 0, success, file, line);
 453         if (success) {
 454                 curthread->td_sx_slocks--;
 455                 WITNESS_UPGRADE(&sx->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
 456                     file, line);
 457                 LOCKSTAT_RECORD0(sx__upgrade, sx);
 458         }
 459         return (success);
 460 }
 461
 462 int
 463 sx_try_upgrade_(struct sx *sx, const char *file, int line)
 464 {
 465
 466         return (sx_try_upgrade_int(sx LOCK_FILE_LINE_ARG));
 467 }
 468
 469 /*
 470  * Downgrade an unrecursed exclusive lock into a single shared lock.
 471  */
 472 void
 473 sx_downgrade_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
 474 {
 475         uintptr_t x;
 476         int wakeup_swapper;
 477
 478         if (SCHEDULER_STOPPED())
 479                 return;
 480
 481         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 482             ("sx_downgrade() of destroyed sx @ %s:%d", file, line));
 483         _sx_assert(sx, SA_XLOCKED | SA_NOTRECURSED, file, line);
 484 #ifndef INVARIANTS
 485         if (sx_recursed(sx))
 486                 panic("downgrade of a recursed lock");
 487 #endif
 488
 489         WITNESS_DOWNGRADE(&sx->lock_object, 0, file, line);
 490
 491         /*
 492          * Try to switch from an exclusive lock with no shared waiters
 493          * to one sharer with no shared waiters.  If there are
 494          * exclusive waiters, we don't need to lock the sleep queue so
 495          * long as we preserve the flag.  We do one quick try and if
 496          * that fails we grab the sleepq lock to keep the flags from
 497          * changing and do it the slow way.
 498          *
 499          * We have to lock the sleep queue if there are shared waiters
 500          * so we can wake them up.
 501          */
 502         x = sx->sx_lock;
 503         if (!(x & SX_LOCK_SHARED_WAITERS) &&
 504             atomic_cmpset_rel_ptr(&sx->sx_lock, x, SX_SHARERS_LOCK(1) |
 505             (x & SX_LOCK_EXCLUSIVE_WAITERS)))
 506                 goto out;
 507
 508         /*
 509          * Lock the sleep queue so we can read the waiters bits
 510          * without any races and wakeup any shared waiters.
 511          */
 512         sleepq_lock(&sx->lock_object);
 513
 514         /*
 515          * Preserve SX_LOCK_EXCLUSIVE_WAITERS while downgraded to a single
 516          * shared lock.  If there are any shared waiters, wake them up.
 517          */
 518         wakeup_swapper = 0;
 519         x = sx->sx_lock;
 520         atomic_store_rel_ptr(&sx->sx_lock, SX_SHARERS_LOCK(1) |
 521             (x & SX_LOCK_EXCLUSIVE_WAITERS));
 522         if (x & SX_LOCK_SHARED_WAITERS)
 523                 wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX,
 524                     0, SQ_SHARED_QUEUE);
 525         sleepq_release(&sx->lock_object);
 526
 527         if (wakeup_swapper)
 528                 kick_proc0();
 529
 530 out:
 531         curthread->td_sx_slocks++;
 532         LOCK_LOG_LOCK("XDOWNGRADE", &sx->lock_object, 0, 0, file, line);
 533         LOCKSTAT_RECORD0(sx__downgrade, sx);
 534 }
 535
 536 void
 537 sx_downgrade_(struct sx *sx, const char *file, int line)
 538 {
 539
 540         sx_downgrade_int(sx LOCK_FILE_LINE_ARG);
 541 }
 542
 543 #ifdef  ADAPTIVE_SX
 544 static inline void
 545 sx_drop_critical(uintptr_t x, bool *in_critical, int *extra_work)
 546 {
 547
 548         if (x & SX_LOCK_WRITE_SPINNER)
 549                 return;
 550         if (*in_critical) {
 551                 critical_exit();
 552                 *in_critical = false;
 553                 (*extra_work)--;
 554         }
 555 }
 556 #else
 557 #define sx_drop_critical(x, in_critical, extra_work) do { } while(0)
 558 #endif
 559
 560 /*
 561  * This function represents the so-called 'hard case' for sx_xlock
 562  * operation.  All 'easy case' failures are redirected to this.  Note
 563  * that ideally this would be a static function, but it needs to be
 564  * accessible from at least sx.h.
 565  */
 566 int
 567 _sx_xlock_hard(struct sx *sx, uintptr_t x, int opts LOCK_FILE_LINE_ARG_DEF)
 568 {
 569         GIANT_DECLARE;
 570         uintptr_t tid, setx;
 571 #ifdef ADAPTIVE_SX
 572         volatile struct thread *owner;
 573         u_int i, n, spintries = 0;
 574         enum { READERS, WRITER } sleep_reason = READERS;
 575         bool adaptive;
 576         bool in_critical = false;
 577 #endif
 578 #ifdef LOCK_PROFILING
 579         uint64_t waittime = 0;
 580         int contested = 0;
 581 #endif
 582         int error = 0;
 583 #if defined(ADAPTIVE_SX) || defined(KDTRACE_HOOKS)
 584         struct lock_delay_arg lda;
 585 #endif
 586 #ifdef  KDTRACE_HOOKS
 587         u_int sleep_cnt = 0;
 588         int64_t sleep_time = 0;
 589         int64_t all_time = 0;
 590 #endif
 591 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
 592         uintptr_t state = 0;
 593         int doing_lockprof = 0;
 594 #endif
 595         int extra_work = 0;
 596
 597         tid = (uintptr_t)curthread;
 598
 599 #ifdef KDTRACE_HOOKS
 600         if (LOCKSTAT_PROFILE_ENABLED(sx__acquire)) {
 601                 while (x == SX_LOCK_UNLOCKED) {
 602                         if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid))
 603                                 goto out_lockstat;
 604                 }
 605                 extra_work = 1;
 606                 doing_lockprof = 1;
 607                 all_time -= lockstat_nsecs(&sx->lock_object);
 608                 state = x;
 609         }
 610 #endif
 611 #ifdef LOCK_PROFILING
 612         extra_work = 1;
 613         doing_lockprof = 1;
 614         state = x;
 615 #endif
 616
 617         if (SCHEDULER_STOPPED())
 618                 return (0);
 619
 620 #if defined(ADAPTIVE_SX)
 621         lock_delay_arg_init(&lda, &sx_delay);
 622 #elif defined(KDTRACE_HOOKS)
 623         lock_delay_arg_init(&lda, NULL);
 624 #endif
 625
 626         if (__predict_false(x == SX_LOCK_UNLOCKED))
 627                 x = SX_READ_VALUE(sx);
 628
 629         /* If we already hold an exclusive lock, then recurse. */
 630         if (__predict_false(lv_sx_owner(x) == (struct thread *)tid)) {
 631                 KASSERT((sx->lock_object.lo_flags & LO_RECURSABLE) != 0,
 632             ("_sx_xlock_hard: recursed on non-recursive sx %s @ %s:%d\n",
 633                     sx->lock_object.lo_name, file, line));
 634                 sx->sx_recurse++;
 635                 atomic_set_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
 636                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 637                         CTR2(KTR_LOCK, "%s: %p recursing", __func__, sx);
 638                 return (0);
 639         }
 640
 641         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 642                 CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__,
 643                     sx->lock_object.lo_name, (void *)sx->sx_lock, file, line);
 644
 645 #ifdef ADAPTIVE_SX
 646         adaptive = ((sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0);
 647 #endif
 648
 649 #ifdef HWPMC_HOOKS
 650         PMC_SOFT_CALL( , , lock, failed);
 651 #endif
 652         lock_profile_obtain_lock_failed(&sx->lock_object, &contested,
 653             &waittime);
 654
 655 #ifndef INVARIANTS
 656         GIANT_SAVE(extra_work);
 657 #endif
 658
 659         for (;;) {
 660                 if (x == SX_LOCK_UNLOCKED) {
 661                         if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid))
 662                                 break;
 663                         continue;
 664                 }
 665 #ifdef INVARIANTS
 666                 GIANT_SAVE(extra_work);
 667 #endif
 668 #ifdef KDTRACE_HOOKS
 669                 lda.spin_cnt++;
 670 #endif
 671 #ifdef ADAPTIVE_SX
 672                 if (__predict_false(!adaptive))
 673                         goto sleepq;
 674                 /*
 675                  * If the lock is write locked and the owner is
 676                  * running on another CPU, spin until the owner stops
 677                  * running or the state of the lock changes.
 678                  */
 679                 if ((x & SX_LOCK_SHARED) == 0) {
 680                         sx_drop_critical(x, &in_critical, &extra_work);
 681                         sleep_reason = WRITER;
 682                         owner = lv_sx_owner(x);
 683                         if (!TD_IS_RUNNING(owner))
 684                                 goto sleepq;
 685                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 686                                 CTR3(KTR_LOCK, "%s: spinning on %p held by %p",
 687                                     __func__, sx, owner);
 688                         KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
 689                             "spinning", "lockname:\"%s\"",
 690                             sx->lock_object.lo_name);
 691                         do {
 692                                 lock_delay(&lda);
 693                                 x = SX_READ_VALUE(sx);
 694                                 owner = lv_sx_owner(x);
 695                         } while (owner != NULL && TD_IS_RUNNING(owner));
 696                         KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
 697                             "running");
 698                         continue;
 699                 } else if (SX_SHARERS(x) > 0) {
 700                         sleep_reason = READERS;
 701                         if (spintries == asx_retries)
 702                                 goto sleepq;
 703                         if (!(x & SX_LOCK_WRITE_SPINNER)) {
 704                                 if (!in_critical) {
 705                                         critical_enter();
 706                                         in_critical = true;
 707                                         extra_work++;
 708                                 }
 709                                 if (!atomic_fcmpset_ptr(&sx->sx_lock, &x,
 710                                     x | SX_LOCK_WRITE_SPINNER)) {
 711                                         critical_exit();
 712                                         in_critical = false;
 713                                         extra_work--;
 714                                         continue;
 715                                 }
 716                         }
 717                         spintries++;
 718                         KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
 719                             "spinning", "lockname:\"%s\"",
 720                             sx->lock_object.lo_name);
 721                         n = SX_SHARERS(x);
 722                         for (i = 0; i < asx_loops; i += n) {
 723                                 lock_delay_spin(n);
 724                                 x = SX_READ_VALUE(sx);
 725                                 if (!(x & SX_LOCK_WRITE_SPINNER))
 726                                         break;
 727                                 if (!(x & SX_LOCK_SHARED))
 728                                         break;
 729                                 n = SX_SHARERS(x);
 730                                 if (n == 0)
 731                                         break;
 732                         }
 733 #ifdef KDTRACE_HOOKS
 734                         lda.spin_cnt += i;
 735 #endif
 736                         KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
 737                             "running");
 738                         if (i < asx_loops)
 739                                 continue;
 740                 }
 741 sleepq:
 742 #endif
 743                 sleepq_lock(&sx->lock_object);
 744                 x = SX_READ_VALUE(sx);
 745 retry_sleepq:
 746
 747                 /*
 748                  * If the lock was released while spinning on the
 749                  * sleep queue chain lock, try again.
 750                  */
 751                 if (x == SX_LOCK_UNLOCKED) {
 752                         sleepq_release(&sx->lock_object);
 753                         sx_drop_critical(x, &in_critical, &extra_work);
 754                         continue;
 755                 }
 756
 757 #ifdef ADAPTIVE_SX
 758                 /*
 759                  * The current lock owner might have started executing
 760                  * on another CPU (or the lock could have changed
 761                  * owners) while we were waiting on the sleep queue
 762                  * chain lock.  If so, drop the sleep queue lock and try
 763                  * again.
 764                  */
 765                 if (adaptive) {
 766                         if (!(x & SX_LOCK_SHARED)) {
 767                                 owner = (struct thread *)SX_OWNER(x);
 768                                 if (TD_IS_RUNNING(owner)) {
 769                                         sleepq_release(&sx->lock_object);
 770                                         sx_drop_critical(x, &in_critical,
 771                                             &extra_work);
 772                                         continue;
 773                                 }
 774                         } else if (SX_SHARERS(x) > 0 && sleep_reason == WRITER) {
 775                                 sleepq_release(&sx->lock_object);
 776                                 sx_drop_critical(x, &in_critical, &extra_work);
 777                                 continue;
 778                         }
 779                 }
 780 #endif
 781
 782                 /*
 783                  * If an exclusive lock was released with both shared
 784                  * and exclusive waiters and a shared waiter hasn't
 785                  * woken up and acquired the lock yet, sx_lock will be
 786                  * set to SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS.
 787                  * If we see that value, try to acquire it once.  Note
 788                  * that we have to preserve SX_LOCK_EXCLUSIVE_WAITERS
 789                  * as there are other exclusive waiters still.  If we
 790                  * fail, restart the loop.
 791                  */
 792                 setx = x & (SX_LOCK_WAITERS | SX_LOCK_WRITE_SPINNER);
 793                 if ((x & ~setx) == SX_LOCK_SHARED) {
 794                         setx &= ~SX_LOCK_WRITE_SPINNER;
 795                         if (!atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid | setx))
 796                                 goto retry_sleepq;
 797                         sleepq_release(&sx->lock_object);
 798                         CTR2(KTR_LOCK, "%s: %p claimed by new writer",
 799                             __func__, sx);
 800                         break;
 801                 }
 802
 803 #ifdef ADAPTIVE_SX
 804                 /*
 805                  * It is possible we set the SX_LOCK_WRITE_SPINNER bit.
 806                  * It is an invariant that when the bit is set, there is
 807                  * a writer ready to grab the lock. Thus clear the bit since
 808                  * we are going to sleep.
 809                  */
 810                 if (in_critical) {
 811                         if ((x & SX_LOCK_WRITE_SPINNER) ||
 812                             !((x & SX_LOCK_EXCLUSIVE_WAITERS))) {
 813                                 setx = x & ~SX_LOCK_WRITE_SPINNER;
 814                                 setx |= SX_LOCK_EXCLUSIVE_WAITERS;
 815                                 if (!atomic_fcmpset_ptr(&sx->sx_lock, &x,
 816                                     setx)) {
 817                                         goto retry_sleepq;
 818                                 }
 819                         }
 820                         critical_exit();
 821                         in_critical = false;
 822                 } else {
 823 #endif
 824                         /*
 825                          * Try to set the SX_LOCK_EXCLUSIVE_WAITERS.  If we fail,
 826                          * than loop back and retry.
 827                          */
 828                         if (!(x & SX_LOCK_EXCLUSIVE_WAITERS)) {
 829                                 if (!atomic_fcmpset_ptr(&sx->sx_lock, &x,
 830                                     x | SX_LOCK_EXCLUSIVE_WAITERS)) {
 831                                         goto retry_sleepq;
 832                                 }
 833                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 834                                         CTR2(KTR_LOCK, "%s: %p set excl waiters flag",
 835                                             __func__, sx);
 836                         }
 837 #ifdef ADAPTIVE_SX
 838                 }
 839 #endif
 840
 841                 /*
 842                  * Since we have been unable to acquire the exclusive
 843                  * lock and the exclusive waiters flag is set, we have
 844                  * to sleep.
 845                  */
 846                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 847                         CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
 848                             __func__, sx);
 849
 850 #ifdef KDTRACE_HOOKS
 851                 sleep_time -= lockstat_nsecs(&sx->lock_object);
 852 #endif
 853                 sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
 854                     SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
 855                     SLEEPQ_INTERRUPTIBLE : 0), SQ_EXCLUSIVE_QUEUE);
 856                 if (!(opts & SX_INTERRUPTIBLE))
 857                         sleepq_wait(&sx->lock_object, 0);
 858                 else
 859                         error = sleepq_wait_sig(&sx->lock_object, 0);
 860 #ifdef KDTRACE_HOOKS
 861                 sleep_time += lockstat_nsecs(&sx->lock_object);
 862                 sleep_cnt++;
 863 #endif
 864                 if (error) {
 865                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 866                                 CTR2(KTR_LOCK,
 867                         "%s: interruptible sleep by %p suspended by signal",
 868                                     __func__, sx);
 869                         break;
 870                 }
 871                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 872                         CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
 873                             __func__, sx);
 874                 x = SX_READ_VALUE(sx);
 875         }
 876         if (__predict_true(!extra_work))
 877                 return (error);
 878 #ifdef ADAPTIVE_SX
 879         if (in_critical)
 880                 critical_exit();
 881 #endif
 882         GIANT_RESTORE();
 883 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
 884         if (__predict_true(!doing_lockprof))
 885                 return (error);
 886 #endif
 887 #ifdef KDTRACE_HOOKS
 888         all_time += lockstat_nsecs(&sx->lock_object);
 889         if (sleep_time)
 890                 LOCKSTAT_RECORD4(sx__block, sx, sleep_time,
 891                     LOCKSTAT_WRITER, (state & SX_LOCK_SHARED) == 0,
 892                     (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
 893         if (lda.spin_cnt > sleep_cnt)
 894                 LOCKSTAT_RECORD4(sx__spin, sx, all_time - sleep_time,
 895                     LOCKSTAT_WRITER, (state & SX_LOCK_SHARED) == 0,
 896                     (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
 897 out_lockstat:
 898 #endif
 899         if (!error)
 900                 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
 901                     contested, waittime, file, line, LOCKSTAT_WRITER);
 902         return (error);
 903 }
 904
 905 /*
 906  * This function represents the so-called 'hard case' for sx_xunlock
 907  * operation.  All 'easy case' failures are redirected to this.  Note
 908  * that ideally this would be a static function, but it needs to be
 909  * accessible from at least sx.h.
 910  */
 911 void
 912 _sx_xunlock_hard(struct sx *sx, uintptr_t x LOCK_FILE_LINE_ARG_DEF)
 913 {
 914         uintptr_t tid, setx;
 915         int queue, wakeup_swapper;
 916
 917         if (SCHEDULER_STOPPED())
 918                 return;
 919
 920         tid = (uintptr_t)curthread;
 921
 922         if (__predict_false(x == tid))
 923                 x = SX_READ_VALUE(sx);
 924
 925         MPASS(!(x & SX_LOCK_SHARED));
 926
 927         if (__predict_false(x & SX_LOCK_RECURSED)) {
 928                 /* The lock is recursed, unrecurse one level. */
 929                 if ((--sx->sx_recurse) == 0)
 930                         atomic_clear_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
 931                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 932                         CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, sx);
 933                 return;
 934         }
 935
 936         LOCKSTAT_PROFILE_RELEASE_RWLOCK(sx__release, sx, LOCKSTAT_WRITER);
 937         if (x == tid &&
 938             atomic_cmpset_rel_ptr(&sx->sx_lock, tid, SX_LOCK_UNLOCKED))
 939                 return;
 940
 941         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 942                 CTR2(KTR_LOCK, "%s: %p contested", __func__, sx);
 943
 944         sleepq_lock(&sx->lock_object);
 945         x = SX_READ_VALUE(sx);
 946         MPASS(x & (SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS));
 947
 948         /*
 949          * The wake up algorithm here is quite simple and probably not
 950          * ideal.  It gives precedence to shared waiters if they are
 951          * present.  For this condition, we have to preserve the
 952          * state of the exclusive waiters flag.
 953          * If interruptible sleeps left the shared queue empty avoid a
 954          * starvation for the threads sleeping on the exclusive queue by giving
 955          * them precedence and cleaning up the shared waiters bit anyway.
 956          */
 957         setx = SX_LOCK_UNLOCKED;
 958         queue = SQ_SHARED_QUEUE;
 959         if ((x & SX_LOCK_EXCLUSIVE_WAITERS) != 0 &&
 960             sleepq_sleepcnt(&sx->lock_object, SQ_EXCLUSIVE_QUEUE) != 0) {
 961                 queue = SQ_EXCLUSIVE_QUEUE;
 962                 setx |= (x & SX_LOCK_SHARED_WAITERS);
 963         }
 964         atomic_store_rel_ptr(&sx->sx_lock, setx);
 965
 966         /* Wake up all the waiters for the specific queue. */
 967         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 968                 CTR3(KTR_LOCK, "%s: %p waking up all threads on %s queue",
 969                     __func__, sx, queue == SQ_SHARED_QUEUE ? "shared" :
 970                     "exclusive");
 971
 972         wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX, 0,
 973             queue);
 974         sleepq_release(&sx->lock_object);
 975         if (wakeup_swapper)
 976                 kick_proc0();
 977 }
 978
 979 static bool __always_inline
 980 __sx_can_read(struct thread *td, uintptr_t x, bool fp)
 981 {
 982
 983         if ((x & (SX_LOCK_SHARED | SX_LOCK_EXCLUSIVE_WAITERS | SX_LOCK_WRITE_SPINNER))
 984                         == SX_LOCK_SHARED)
 985                 return (true);
 986         if (!fp && td->td_sx_slocks && (x & SX_LOCK_SHARED))
 987                 return (true);
 988         return (false);
 989 }
 990
 991 static bool __always_inline
 992 __sx_slock_try(struct sx *sx, struct thread *td, uintptr_t *xp, bool fp
 993     LOCK_FILE_LINE_ARG_DEF)
 994 {
 995
 996         /*
 997          * If no other thread has an exclusive lock then try to bump up
 998          * the count of sharers.  Since we have to preserve the state
 999          * of SX_LOCK_EXCLUSIVE_WAITERS, if we fail to acquire the
1000          * shared lock loop back and retry.
1001          */
1002         while (__sx_can_read(td, *xp, fp)) {
1003                 if (atomic_fcmpset_acq_ptr(&sx->sx_lock, xp,
1004                     *xp + SX_ONE_SHARER)) {
1005                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
1006                                 CTR4(KTR_LOCK, "%s: %p succeed %p -> %p",
1007                                     __func__, sx, (void *)*xp,
1008                                     (void *)(*xp + SX_ONE_SHARER));
1009                         td->td_sx_slocks++;
1010                         return (true);
1011                 }
1012         }
1013         return (false);
1014 }
1015
1016 static int __noinline
1017 _sx_slock_hard(struct sx *sx, int opts, uintptr_t x LOCK_FILE_LINE_ARG_DEF)
1018 {
1019         GIANT_DECLARE;
1020         struct thread *td;
1021 #ifdef ADAPTIVE_SX
1022         volatile struct thread *owner;
1023         u_int i, n, spintries = 0;
1024         bool adaptive;
1025 #endif
1026 #ifdef LOCK_PROFILING
1027         uint64_t waittime = 0;
1028         int contested = 0;
1029 #endif
1030         int error = 0;
1031 #if defined(ADAPTIVE_SX) || defined(KDTRACE_HOOKS)
1032         struct lock_delay_arg lda;
1033 #endif
1034 #ifdef KDTRACE_HOOKS
1035         u_int sleep_cnt = 0;
1036         int64_t sleep_time = 0;
1037         int64_t all_time = 0;
1038 #endif
1039 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
1040         uintptr_t state = 0;
1041 #endif
1042         int extra_work = 0;
1043
1044         td = curthread;
1045
1046 #ifdef KDTRACE_HOOKS
1047         if (LOCKSTAT_PROFILE_ENABLED(sx__acquire)) {
1048                 if (__sx_slock_try(sx, td, &x, false LOCK_FILE_LINE_ARG))
1049                         goto out_lockstat;
1050                 extra_work = 1;
1051                 all_time -= lockstat_nsecs(&sx->lock_object);
1052                 state = x;
1053         }
1054 #endif
1055 #ifdef LOCK_PROFILING
1056         extra_work = 1;
1057         state = x;
1058 #endif
1059
1060         if (SCHEDULER_STOPPED())
1061                 return (0);
1062
1063 #if defined(ADAPTIVE_SX)
1064         lock_delay_arg_init(&lda, &sx_delay);
1065 #elif defined(KDTRACE_HOOKS)
1066         lock_delay_arg_init(&lda, NULL);
1067 #endif
1068
1069 #ifdef ADAPTIVE_SX
1070         adaptive = ((sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0);
1071 #endif
1072
1073 #ifdef HWPMC_HOOKS
1074         PMC_SOFT_CALL( , , lock, failed);
1075 #endif
1076         lock_profile_obtain_lock_failed(&sx->lock_object, &contested,
1077             &waittime);
1078
1079 #ifndef INVARIANTS
1080         GIANT_SAVE(extra_work);
1081 #endif
1082
1083         /*
1084          * As with rwlocks, we don't make any attempt to try to block
1085          * shared locks once there is an exclusive waiter.
1086          */
1087         for (;;) {
1088                 if (__sx_slock_try(sx, td, &x, false LOCK_FILE_LINE_ARG))
1089                         break;
1090 #ifdef INVARIANTS
1091                 GIANT_SAVE(extra_work);
1092 #endif
1093 #ifdef KDTRACE_HOOKS
1094                 lda.spin_cnt++;
1095 #endif
1096
1097 #ifdef ADAPTIVE_SX
1098                 if (__predict_false(!adaptive))
1099                         goto sleepq;
1100
1101                 /*
1102                  * If the owner is running on another CPU, spin until
1103                  * the owner stops running or the state of the lock
1104                  * changes.
1105                  */
1106                 if ((x & SX_LOCK_SHARED) == 0) {
1107                         owner = lv_sx_owner(x);
1108                         if (TD_IS_RUNNING(owner)) {
1109                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1110                                         CTR3(KTR_LOCK,
1111                                             "%s: spinning on %p held by %p",
1112                                             __func__, sx, owner);
1113                                 KTR_STATE1(KTR_SCHED, "thread",
1114                                     sched_tdname(curthread), "spinning",
1115                                     "lockname:\"%s\"", sx->lock_object.lo_name);
1116                                 do {
1117                                         lock_delay(&lda);
1118                                         x = SX_READ_VALUE(sx);
1119                                         owner = lv_sx_owner(x);
1120                                 } while (owner != NULL && TD_IS_RUNNING(owner));
1121                                 KTR_STATE0(KTR_SCHED, "thread",
1122                                     sched_tdname(curthread), "running");
1123                                 continue;
1124                         }
1125                 } else {
1126                         if ((x & SX_LOCK_WRITE_SPINNER) && SX_SHARERS(x) == 0) {
1127                                 MPASS(!__sx_can_read(td, x, false));
1128                                 lock_delay_spin(2);
1129                                 x = SX_READ_VALUE(sx);
1130                                 continue;
1131                         }
1132                         if (spintries < asx_retries) {
1133                                 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
1134                                     "spinning", "lockname:\"%s\"",
1135                                     sx->lock_object.lo_name);
1136                                 n = SX_SHARERS(x);
1137                                 for (i = 0; i < asx_loops; i += n) {
1138                                         lock_delay_spin(n);
1139                                         x = SX_READ_VALUE(sx);
1140                                         if (!(x & SX_LOCK_SHARED))
1141                                                 break;
1142                                         n = SX_SHARERS(x);
1143                                         if (n == 0)
1144                                                 break;
1145                                         if (__sx_can_read(td, x, false))
1146                                                 break;
1147                                 }
1148 #ifdef KDTRACE_HOOKS
1149                                 lda.spin_cnt += i;
1150 #endif
1151                                 KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
1152                                     "running");
1153                                 if (i < asx_loops)
1154                                         continue;
1155                         }
1156                 }
1157 sleepq:
1158 #endif
1159
1160                 /*
1161                  * Some other thread already has an exclusive lock, so
1162                  * start the process of blocking.
1163                  */
1164                 sleepq_lock(&sx->lock_object);
1165                 x = SX_READ_VALUE(sx);
1166 retry_sleepq:
1167                 if (((x & SX_LOCK_WRITE_SPINNER) && SX_SHARERS(x) == 0) ||
1168                     __sx_can_read(td, x, false)) {
1169                         sleepq_release(&sx->lock_object);
1170                         continue;
1171                 }
1172
1173 #ifdef ADAPTIVE_SX
1174                 /*
1175                  * If the owner is running on another CPU, spin until
1176                  * the owner stops running or the state of the lock
1177                  * changes.
1178                  */
1179                 if (!(x & SX_LOCK_SHARED) && adaptive) {
1180                         owner = (struct thread *)SX_OWNER(x);
1181                         if (TD_IS_RUNNING(owner)) {
1182                                 sleepq_release(&sx->lock_object);
1183                                 x = SX_READ_VALUE(sx);
1184                                 continue;
1185                         }
1186                 }
1187 #endif
1188
1189                 /*
1190                  * Try to set the SX_LOCK_SHARED_WAITERS flag.  If we
1191                  * fail to set it drop the sleep queue lock and loop
1192                  * back.
1193                  */
1194                 if (!(x & SX_LOCK_SHARED_WAITERS)) {
1195                         if (!atomic_fcmpset_ptr(&sx->sx_lock, &x,
1196                             x | SX_LOCK_SHARED_WAITERS))
1197                                 goto retry_sleepq;
1198                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
1199                                 CTR2(KTR_LOCK, "%s: %p set shared waiters flag",
1200                                     __func__, sx);
1201                 }
1202
1203                 /*
1204                  * Since we have been unable to acquire the shared lock,
1205                  * we have to sleep.
1206                  */
1207                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1208                         CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
1209                             __func__, sx);
1210
1211 #ifdef KDTRACE_HOOKS
1212                 sleep_time -= lockstat_nsecs(&sx->lock_object);
1213 #endif
1214                 sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
1215                     SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
1216                     SLEEPQ_INTERRUPTIBLE : 0), SQ_SHARED_QUEUE);
1217                 if (!(opts & SX_INTERRUPTIBLE))
1218                         sleepq_wait(&sx->lock_object, 0);
1219                 else
1220                         error = sleepq_wait_sig(&sx->lock_object, 0);
1221 #ifdef KDTRACE_HOOKS
1222                 sleep_time += lockstat_nsecs(&sx->lock_object);
1223                 sleep_cnt++;
1224 #endif
1225                 if (error) {
1226                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
1227                                 CTR2(KTR_LOCK,
1228                         "%s: interruptible sleep by %p suspended by signal",
1229                                     __func__, sx);
1230                         break;
1231                 }
1232                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1233                         CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
1234                             __func__, sx);
1235                 x = SX_READ_VALUE(sx);
1236         }
1237 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
1238         if (__predict_true(!extra_work))
1239                 return (error);
1240 #endif
1241 #ifdef KDTRACE_HOOKS
1242         all_time += lockstat_nsecs(&sx->lock_object);
1243         if (sleep_time)
1244                 LOCKSTAT_RECORD4(sx__block, sx, sleep_time,
1245                     LOCKSTAT_READER, (state & SX_LOCK_SHARED) == 0,
1246                     (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
1247         if (lda.spin_cnt > sleep_cnt)
1248                 LOCKSTAT_RECORD4(sx__spin, sx, all_time - sleep_time,
1249                     LOCKSTAT_READER, (state & SX_LOCK_SHARED) == 0,
1250                     (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
1251 out_lockstat:
1252 #endif
1253         if (error == 0) {
1254                 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
1255                     contested, waittime, file, line, LOCKSTAT_READER);
1256         }
1257         GIANT_RESTORE();
1258         return (error);
1259 }
1260
1261 int
1262 _sx_slock_int(struct sx *sx, int opts LOCK_FILE_LINE_ARG_DEF)
1263 {
1264         struct thread *td;
1265         uintptr_t x;
1266         int error;
1267
1268         KASSERT(kdb_active != 0 || SCHEDULER_STOPPED() ||
1269             !TD_IS_IDLETHREAD(curthread),
1270             ("sx_slock() by idle thread %p on sx %s @ %s:%d",
1271             curthread, sx->lock_object.lo_name, file, line));
1272         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
1273             ("sx_slock() of destroyed sx @ %s:%d", file, line));
1274         WITNESS_CHECKORDER(&sx->lock_object, LOP_NEWORDER, file, line, NULL);
1275
1276         error = 0;
1277         td = curthread;
1278         x = SX_READ_VALUE(sx);
1279         if (__predict_false(LOCKSTAT_PROFILE_ENABLED(sx__acquire) ||
1280             !__sx_slock_try(sx, td, &x, true LOCK_FILE_LINE_ARG)))
1281                 error = _sx_slock_hard(sx, opts, x LOCK_FILE_LINE_ARG);
1282         else
1283                 lock_profile_obtain_lock_success(&sx->lock_object, 0, 0,
1284                     file, line);
1285         if (error == 0) {
1286                 LOCK_LOG_LOCK("SLOCK", &sx->lock_object, 0, 0, file, line);
1287                 WITNESS_LOCK(&sx->lock_object, 0, file, line);
1288                 TD_LOCKS_INC(curthread);
1289         }
1290         return (error);
1291 }
1292
1293 int
1294 _sx_slock(struct sx *sx, int opts, const char *file, int line)
1295 {
1296
1297         return (_sx_slock_int(sx, opts LOCK_FILE_LINE_ARG));
1298 }
1299
1300 static bool __always_inline
1301 _sx_sunlock_try(struct sx *sx, struct thread *td, uintptr_t *xp)
1302 {
1303
1304         for (;;) {
1305                 if (SX_SHARERS(*xp) > 1 || !(*xp & SX_LOCK_WAITERS)) {
1306                         if (atomic_fcmpset_rel_ptr(&sx->sx_lock, xp,
1307                             *xp - SX_ONE_SHARER)) {
1308                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1309                                         CTR4(KTR_LOCK,
1310                                             "%s: %p succeeded %p -> %p",
1311                                             __func__, sx, (void *)*xp,
1312                                             (void *)(*xp - SX_ONE_SHARER));
1313                                 td->td_sx_slocks--;
1314                                 return (true);
1315                         }
1316                         continue;
1317                 }
1318                 break;
1319         }
1320         return (false);
1321 }
1322
1323 static void __noinline
1324 _sx_sunlock_hard(struct sx *sx, struct thread *td, uintptr_t x
1325     LOCK_FILE_LINE_ARG_DEF)
1326 {
1327         int wakeup_swapper = 0;
1328         uintptr_t setx, queue;
1329
1330         if (SCHEDULER_STOPPED())
1331                 return;
1332
1333         if (_sx_sunlock_try(sx, td, &x))
1334                 goto out_lockstat;
1335
1336         sleepq_lock(&sx->lock_object);
1337         x = SX_READ_VALUE(sx);
1338         for (;;) {
1339                 if (_sx_sunlock_try(sx, td, &x))
1340                         break;
1341
1342                 /*
1343                  * Wake up semantic here is quite simple:
1344                  * Just wake up all the exclusive waiters.
1345                  * Note that the state of the lock could have changed,
1346                  * so if it fails loop back and retry.
1347                  */
1348                 setx = SX_LOCK_UNLOCKED;
1349                 queue = SQ_SHARED_QUEUE;
1350                 if (x & SX_LOCK_EXCLUSIVE_WAITERS) {
1351                         setx |= (x & SX_LOCK_SHARED_WAITERS);
1352                         queue = SQ_EXCLUSIVE_QUEUE;
1353                 }
1354                 setx |= (x & SX_LOCK_WRITE_SPINNER);
1355                 if (!atomic_fcmpset_rel_ptr(&sx->sx_lock, &x, setx))
1356                         continue;
1357                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1358                         CTR2(KTR_LOCK, "%s: %p waking up all thread on"
1359                             "exclusive queue", __func__, sx);
1360                 wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX,
1361                     0, queue);
1362                 td->td_sx_slocks--;
1363                 break;
1364         }
1365         sleepq_release(&sx->lock_object);
1366         if (wakeup_swapper)
1367                 kick_proc0();
1368 out_lockstat:
1369         LOCKSTAT_PROFILE_RELEASE_RWLOCK(sx__release, sx, LOCKSTAT_READER);
1370 }
1371
1372 void
1373 _sx_sunlock_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
1374 {
1375         struct thread *td;
1376         uintptr_t x;
1377
1378         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
1379             ("sx_sunlock() of destroyed sx @ %s:%d", file, line));
1380         _sx_assert(sx, SA_SLOCKED, file, line);
1381         WITNESS_UNLOCK(&sx->lock_object, 0, file, line);
1382         LOCK_LOG_LOCK("SUNLOCK", &sx->lock_object, 0, 0, file, line);
1383
1384         td = curthread;
1385         x = SX_READ_VALUE(sx);
1386         if (__predict_false(LOCKSTAT_PROFILE_ENABLED(sx__release) ||
1387             !_sx_sunlock_try(sx, td, &x)))
1388                 _sx_sunlock_hard(sx, td, x LOCK_FILE_LINE_ARG);
1389         else
1390                 lock_profile_release_lock(&sx->lock_object);
1391
1392         TD_LOCKS_DEC(curthread);
1393 }
1394
1395 void
1396 _sx_sunlock(struct sx *sx, const char *file, int line)
1397 {
1398
1399         _sx_sunlock_int(sx LOCK_FILE_LINE_ARG);
1400 }
1401
1402 #ifdef INVARIANT_SUPPORT
1403 #ifndef INVARIANTS
1404 #undef  _sx_assert
1405 #endif
1406
1407 /*
1408  * In the non-WITNESS case, sx_assert() can only detect that at least
1409  * *some* thread owns an slock, but it cannot guarantee that *this*
1410  * thread owns an slock.
1411  */
1412 void
1413 _sx_assert(const struct sx *sx, int what, const char *file, int line)
1414 {
1415 #ifndef WITNESS
1416         int slocked = 0;
1417 #endif
1418
1419         if (SCHEDULER_STOPPED())
1420                 return;
1421         switch (what) {
1422         case SA_SLOCKED:
1423         case SA_SLOCKED | SA_NOTRECURSED:
1424         case SA_SLOCKED | SA_RECURSED:
1425 #ifndef WITNESS
1426                 slocked = 1;
1427                 /* FALLTHROUGH */
1428 #endif
1429         case SA_LOCKED:
1430         case SA_LOCKED | SA_NOTRECURSED:
1431         case SA_LOCKED | SA_RECURSED:
1432 #ifdef WITNESS
1433                 witness_assert(&sx->lock_object, what, file, line);
1434 #else
1435                 /*
1436                  * If some other thread has an exclusive lock or we
1437                  * have one and are asserting a shared lock, fail.
1438                  * Also, if no one has a lock at all, fail.
1439                  */
1440                 if (sx->sx_lock == SX_LOCK_UNLOCKED ||
1441                     (!(sx->sx_lock & SX_LOCK_SHARED) && (slocked ||
1442                     sx_xholder(sx) != curthread)))
1443                         panic("Lock %s not %slocked @ %s:%d\n",
1444                             sx->lock_object.lo_name, slocked ? "share " : "",
1445                             file, line);
1446
1447                 if (!(sx->sx_lock & SX_LOCK_SHARED)) {
1448                         if (sx_recursed(sx)) {
1449                                 if (what & SA_NOTRECURSED)
1450                                         panic("Lock %s recursed @ %s:%d\n",
1451                                             sx->lock_object.lo_name, file,
1452                                             line);
1453                         } else if (what & SA_RECURSED)
1454                                 panic("Lock %s not recursed @ %s:%d\n",
1455                                     sx->lock_object.lo_name, file, line);
1456                 }
1457 #endif
1458                 break;
1459         case SA_XLOCKED:
1460         case SA_XLOCKED | SA_NOTRECURSED:
1461         case SA_XLOCKED | SA_RECURSED:
1462                 if (sx_xholder(sx) != curthread)
1463                         panic("Lock %s not exclusively locked @ %s:%d\n",
1464                             sx->lock_object.lo_name, file, line);
1465                 if (sx_recursed(sx)) {
1466                         if (what & SA_NOTRECURSED)
1467                                 panic("Lock %s recursed @ %s:%d\n",
1468                                     sx->lock_object.lo_name, file, line);
1469                 } else if (what & SA_RECURSED)
1470                         panic("Lock %s not recursed @ %s:%d\n",
1471                             sx->lock_object.lo_name, file, line);
1472                 break;
1473         case SA_UNLOCKED:
1474 #ifdef WITNESS
1475                 witness_assert(&sx->lock_object, what, file, line);
1476 #else
1477                 /*
1478                  * If we hold an exclusve lock fail.  We can't
1479                  * reliably check to see if we hold a shared lock or
1480                  * not.
1481                  */
1482                 if (sx_xholder(sx) == curthread)
1483                         panic("Lock %s exclusively locked @ %s:%d\n",
1484                             sx->lock_object.lo_name, file, line);
1485 #endif
1486                 break;
1487         default:
1488                 panic("Unknown sx lock assertion: %d @ %s:%d", what, file,
1489                     line);
1490         }
1491 }
1492 #endif  /* INVARIANT_SUPPORT */
1493
1494 #ifdef DDB
1495 static void
1496 db_show_sx(const struct lock_object *lock)
1497 {
1498         struct thread *td;
1499         const struct sx *sx;
1500
1501         sx = (const struct sx *)lock;
1502
1503         db_printf(" state: ");
1504         if (sx->sx_lock == SX_LOCK_UNLOCKED)
1505                 db_printf("UNLOCKED\n");
1506         else if (sx->sx_lock == SX_LOCK_DESTROYED) {
1507                 db_printf("DESTROYED\n");
1508                 return;
1509         } else if (sx->sx_lock & SX_LOCK_SHARED)
1510                 db_printf("SLOCK: %ju\n", (uintmax_t)SX_SHARERS(sx->sx_lock));
1511         else {
1512                 td = sx_xholder(sx);
1513                 db_printf("XLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1514                     td->td_tid, td->td_proc->p_pid, td->td_name);
1515                 if (sx_recursed(sx))
1516                         db_printf(" recursed: %d\n", sx->sx_recurse);
1517         }
1518
1519         db_printf(" waiters: ");
1520         switch(sx->sx_lock &
1521             (SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS)) {
1522         case SX_LOCK_SHARED_WAITERS:
1523                 db_printf("shared\n");
1524                 break;
1525         case SX_LOCK_EXCLUSIVE_WAITERS:
1526                 db_printf("exclusive\n");
1527                 break;
1528         case SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS:
1529                 db_printf("exclusive and shared\n");
1530                 break;
1531         default:
1532                 db_printf("none\n");
1533         }
1534 }
1535
1536 /*
1537  * Check to see if a thread that is blocked on a sleep queue is actually
1538  * blocked on an sx lock.  If so, output some details and return true.
1539  * If the lock has an exclusive owner, return that in *ownerp.
1540  */
1541 int
1542 sx_chain(struct thread *td, struct thread **ownerp)
1543 {
1544         struct sx *sx;
1545
1546         /*
1547          * Check to see if this thread is blocked on an sx lock.
1548          * First, we check the lock class.  If that is ok, then we
1549          * compare the lock name against the wait message.
1550          */
1551         sx = td->td_wchan;
1552         if (LOCK_CLASS(&sx->lock_object) != &lock_class_sx ||
1553             sx->lock_object.lo_name != td->td_wmesg)
1554                 return (0);
1555
1556         /* We think we have an sx lock, so output some details. */
1557         db_printf("blocked on sx \"%s\" ", td->td_wmesg);
1558         *ownerp = sx_xholder(sx);
1559         if (sx->sx_lock & SX_LOCK_SHARED)
1560                 db_printf("SLOCK (count %ju)\n",
1561                     (uintmax_t)SX_SHARERS(sx->sx_lock));
1562         else
1563                 db_printf("XLOCK\n");
1564         return (1);
1565 }
1566 #endif