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                         return (1);
 296                 }
 297         }
 298
 299         LOCK_LOG_TRY("SLOCK", &sx->lock_object, 0, 0, file, line);
 300         return (0);
 301 }
 302
 303 int
 304 sx_try_slock_(struct sx *sx, const char *file, int line)
 305 {
 306
 307         return (sx_try_slock_int(sx LOCK_FILE_LINE_ARG));
 308 }
 309
 310 int
 311 _sx_xlock(struct sx *sx, int opts, const char *file, int line)
 312 {
 313         uintptr_t tid, x;
 314         int error = 0;
 315
 316         KASSERT(kdb_active != 0 || SCHEDULER_STOPPED() ||
 317             !TD_IS_IDLETHREAD(curthread),
 318             ("sx_xlock() by idle thread %p on sx %s @ %s:%d",
 319             curthread, sx->lock_object.lo_name, file, line));
 320         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 321             ("sx_xlock() of destroyed sx @ %s:%d", file, line));
 322         WITNESS_CHECKORDER(&sx->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE, file,
 323             line, NULL);
 324         tid = (uintptr_t)curthread;
 325         x = SX_LOCK_UNLOCKED;
 326         if (!atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid))
 327                 error = _sx_xlock_hard(sx, x, opts LOCK_FILE_LINE_ARG);
 328         else
 329                 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
 330                     0, 0, file, line, LOCKSTAT_WRITER);
 331         if (!error) {
 332                 LOCK_LOG_LOCK("XLOCK", &sx->lock_object, 0, sx->sx_recurse,
 333                     file, line);
 334                 WITNESS_LOCK(&sx->lock_object, LOP_EXCLUSIVE, file, line);
 335                 TD_LOCKS_INC(curthread);
 336         }
 337
 338         return (error);
 339 }
 340
 341 int
 342 sx_try_xlock_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
 343 {
 344         struct thread *td;
 345         uintptr_t tid, x;
 346         int rval;
 347         bool recursed;
 348
 349         td = curthread;
 350         tid = (uintptr_t)td;
 351         if (SCHEDULER_STOPPED_TD(td))
 352                 return (1);
 353
 354         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(td),
 355             ("sx_try_xlock() by idle thread %p on sx %s @ %s:%d",
 356             curthread, sx->lock_object.lo_name, file, line));
 357         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 358             ("sx_try_xlock() of destroyed sx @ %s:%d", file, line));
 359
 360         rval = 1;
 361         recursed = false;
 362         x = SX_LOCK_UNLOCKED;
 363         for (;;) {
 364                 if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid))
 365                         break;
 366                 if (x == SX_LOCK_UNLOCKED)
 367                         continue;
 368                 if (x == tid && (sx->lock_object.lo_flags & LO_RECURSABLE)) {
 369                         sx->sx_recurse++;
 370                         atomic_set_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
 371                         break;
 372                 }
 373                 rval = 0;
 374                 break;
 375         }
 376
 377         LOCK_LOG_TRY("XLOCK", &sx->lock_object, 0, rval, file, line);
 378         if (rval) {
 379                 WITNESS_LOCK(&sx->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
 380                     file, line);
 381                 if (!recursed)
 382                         LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire,
 383                             sx, 0, 0, file, line, LOCKSTAT_WRITER);
 384                 TD_LOCKS_INC(curthread);
 385         }
 386
 387         return (rval);
 388 }
 389
 390 int
 391 sx_try_xlock_(struct sx *sx, const char *file, int line)
 392 {
 393
 394         return (sx_try_xlock_int(sx LOCK_FILE_LINE_ARG));
 395 }
 396
 397 void
 398 _sx_xunlock(struct sx *sx, const char *file, int line)
 399 {
 400
 401         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 402             ("sx_xunlock() of destroyed sx @ %s:%d", file, line));
 403         _sx_assert(sx, SA_XLOCKED, file, line);
 404         WITNESS_UNLOCK(&sx->lock_object, LOP_EXCLUSIVE, file, line);
 405         LOCK_LOG_LOCK("XUNLOCK", &sx->lock_object, 0, sx->sx_recurse, file,
 406             line);
 407 #if LOCK_DEBUG > 0
 408         _sx_xunlock_hard(sx, (uintptr_t)curthread, file, line);
 409 #else
 410         __sx_xunlock(sx, curthread, file, line);
 411 #endif
 412         TD_LOCKS_DEC(curthread);
 413 }
 414
 415 /*
 416  * Try to do a non-blocking upgrade from a shared lock to an exclusive lock.
 417  * This will only succeed if this thread holds a single shared lock.
 418  * Return 1 if if the upgrade succeed, 0 otherwise.
 419  */
 420 int
 421 sx_try_upgrade_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
 422 {
 423         uintptr_t x;
 424         uintptr_t waiters;
 425         int success;
 426
 427         if (SCHEDULER_STOPPED())
 428                 return (1);
 429
 430         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 431             ("sx_try_upgrade() of destroyed sx @ %s:%d", file, line));
 432         _sx_assert(sx, SA_SLOCKED, file, line);
 433
 434         /*
 435          * Try to switch from one shared lock to an exclusive lock.  We need
 436          * to maintain the SX_LOCK_EXCLUSIVE_WAITERS flag if set so that
 437          * we will wake up the exclusive waiters when we drop the lock.
 438          */
 439         success = 0;
 440         x = SX_READ_VALUE(sx);
 441         for (;;) {
 442                 if (SX_SHARERS(x) > 1)
 443                         break;
 444                 waiters = (x & SX_LOCK_EXCLUSIVE_WAITERS);
 445                 if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x,
 446                     (uintptr_t)curthread | waiters)) {
 447                         success = 1;
 448                         break;
 449                 }
 450         }
 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(sx__upgrade, sx);
 456         }
 457         return (success);
 458 }
 459
 460 int
 461 sx_try_upgrade_(struct sx *sx, const char *file, int line)
 462 {
 463
 464         return (sx_try_upgrade_int(sx LOCK_FILE_LINE_ARG));
 465 }
 466
 467 /*
 468  * Downgrade an unrecursed exclusive lock into a single shared lock.
 469  */
 470 void
 471 sx_downgrade_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
 472 {
 473         uintptr_t x;
 474         int wakeup_swapper;
 475
 476         if (SCHEDULER_STOPPED())
 477                 return;
 478
 479         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 480             ("sx_downgrade() of destroyed sx @ %s:%d", file, line));
 481         _sx_assert(sx, SA_XLOCKED | SA_NOTRECURSED, file, line);
 482 #ifndef INVARIANTS
 483         if (sx_recursed(sx))
 484                 panic("downgrade of a recursed lock");
 485 #endif
 486
 487         WITNESS_DOWNGRADE(&sx->lock_object, 0, file, line);
 488
 489         /*
 490          * Try to switch from an exclusive lock with no shared waiters
 491          * to one sharer with no shared waiters.  If there are
 492          * exclusive waiters, we don't need to lock the sleep queue so
 493          * long as we preserve the flag.  We do one quick try and if
 494          * that fails we grab the sleepq lock to keep the flags from
 495          * changing and do it the slow way.
 496          *
 497          * We have to lock the sleep queue if there are shared waiters
 498          * so we can wake them up.
 499          */
 500         x = sx->sx_lock;
 501         if (!(x & SX_LOCK_SHARED_WAITERS) &&
 502             atomic_cmpset_rel_ptr(&sx->sx_lock, x, SX_SHARERS_LOCK(1) |
 503             (x & SX_LOCK_EXCLUSIVE_WAITERS)))
 504                 goto out;
 505
 506         /*
 507          * Lock the sleep queue so we can read the waiters bits
 508          * without any races and wakeup any shared waiters.
 509          */
 510         sleepq_lock(&sx->lock_object);
 511
 512         /*
 513          * Preserve SX_LOCK_EXCLUSIVE_WAITERS while downgraded to a single
 514          * shared lock.  If there are any shared waiters, wake them up.
 515          */
 516         wakeup_swapper = 0;
 517         x = sx->sx_lock;
 518         atomic_store_rel_ptr(&sx->sx_lock, SX_SHARERS_LOCK(1) |
 519             (x & SX_LOCK_EXCLUSIVE_WAITERS));
 520         if (x & SX_LOCK_SHARED_WAITERS)
 521                 wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX,
 522                     0, SQ_SHARED_QUEUE);
 523         sleepq_release(&sx->lock_object);
 524
 525         if (wakeup_swapper)
 526                 kick_proc0();
 527
 528 out:
 529         LOCK_LOG_LOCK("XDOWNGRADE", &sx->lock_object, 0, 0, file, line);
 530         LOCKSTAT_RECORD0(sx__downgrade, sx);
 531 }
 532
 533 void
 534 sx_downgrade_(struct sx *sx, const char *file, int line)
 535 {
 536
 537         sx_downgrade_int(sx LOCK_FILE_LINE_ARG);
 538 }
 539
 540 /*
 541  * This function represents the so-called 'hard case' for sx_xlock
 542  * operation.  All 'easy case' failures are redirected to this.  Note
 543  * that ideally this would be a static function, but it needs to be
 544  * accessible from at least sx.h.
 545  */
 546 int
 547 _sx_xlock_hard(struct sx *sx, uintptr_t x, int opts LOCK_FILE_LINE_ARG_DEF)
 548 {
 549         GIANT_DECLARE;
 550         uintptr_t tid;
 551 #ifdef ADAPTIVE_SX
 552         volatile struct thread *owner;
 553         u_int i, n, spintries = 0;
 554         enum { READERS, WRITER } sleep_reason;
 555         bool adaptive;
 556 #endif
 557 #ifdef LOCK_PROFILING
 558         uint64_t waittime = 0;
 559         int contested = 0;
 560 #endif
 561         int error = 0;
 562 #if defined(ADAPTIVE_SX) || defined(KDTRACE_HOOKS)
 563         struct lock_delay_arg lda;
 564 #endif
 565 #ifdef  KDTRACE_HOOKS
 566         u_int sleep_cnt = 0;
 567         int64_t sleep_time = 0;
 568         int64_t all_time = 0;
 569 #endif
 570 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
 571         uintptr_t state;
 572 #endif
 573         int extra_work = 0;
 574
 575         tid = (uintptr_t)curthread;
 576
 577 #ifdef KDTRACE_HOOKS
 578         if (LOCKSTAT_PROFILE_ENABLED(sx__acquire)) {
 579                 while (x == SX_LOCK_UNLOCKED) {
 580                         if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid))
 581                                 goto out_lockstat;
 582                 }
 583                 extra_work = 1;
 584                 all_time -= lockstat_nsecs(&sx->lock_object);
 585                 state = x;
 586         }
 587 #endif
 588 #ifdef LOCK_PROFILING
 589         extra_work = 1;
 590         state = x;
 591 #endif
 592
 593         if (SCHEDULER_STOPPED())
 594                 return (0);
 595
 596 #if defined(ADAPTIVE_SX)
 597         lock_delay_arg_init(&lda, &sx_delay);
 598 #elif defined(KDTRACE_HOOKS)
 599         lock_delay_arg_init(&lda, NULL);
 600 #endif
 601
 602         if (__predict_false(x == SX_LOCK_UNLOCKED))
 603                 x = SX_READ_VALUE(sx);
 604
 605         /* If we already hold an exclusive lock, then recurse. */
 606         if (__predict_false(lv_sx_owner(x) == (struct thread *)tid)) {
 607                 KASSERT((sx->lock_object.lo_flags & LO_RECURSABLE) != 0,
 608             ("_sx_xlock_hard: recursed on non-recursive sx %s @ %s:%d\n",
 609                     sx->lock_object.lo_name, file, line));
 610                 sx->sx_recurse++;
 611                 atomic_set_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
 612                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 613                         CTR2(KTR_LOCK, "%s: %p recursing", __func__, sx);
 614                 return (0);
 615         }
 616
 617         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 618                 CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__,
 619                     sx->lock_object.lo_name, (void *)sx->sx_lock, file, line);
 620
 621 #ifdef ADAPTIVE_SX
 622         adaptive = ((sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0);
 623 #endif
 624
 625 #ifdef HWPMC_HOOKS
 626         PMC_SOFT_CALL( , , lock, failed);
 627 #endif
 628         lock_profile_obtain_lock_failed(&sx->lock_object, &contested,
 629             &waittime);
 630
 631 #ifndef INVARIANTS
 632         GIANT_SAVE(extra_work);
 633 #endif
 634
 635         for (;;) {
 636                 if (x == SX_LOCK_UNLOCKED) {
 637                         if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid))
 638                                 break;
 639                         continue;
 640                 }
 641 #ifdef INVARIANTS
 642                 GIANT_SAVE(extra_work);
 643 #endif
 644 #ifdef KDTRACE_HOOKS
 645                 lda.spin_cnt++;
 646 #endif
 647 #ifdef ADAPTIVE_SX
 648                 if (__predict_false(!adaptive))
 649                         goto sleepq;
 650                 /*
 651                  * If the lock is write locked and the owner is
 652                  * running on another CPU, spin until the owner stops
 653                  * running or the state of the lock changes.
 654                  */
 655                 if ((x & SX_LOCK_SHARED) == 0) {
 656                         sleep_reason = WRITER;
 657                         owner = lv_sx_owner(x);
 658                         if (!TD_IS_RUNNING(owner))
 659                                 goto sleepq;
 660                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 661                                 CTR3(KTR_LOCK, "%s: spinning on %p held by %p",
 662                                     __func__, sx, owner);
 663                         KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
 664                             "spinning", "lockname:\"%s\"",
 665                             sx->lock_object.lo_name);
 666                         do {
 667                                 lock_delay(&lda);
 668                                 x = SX_READ_VALUE(sx);
 669                                 owner = lv_sx_owner(x);
 670                         } while (owner != NULL && TD_IS_RUNNING(owner));
 671                         KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
 672                             "running");
 673                         continue;
 674                 } else if (SX_SHARERS(x) > 0) {
 675                         sleep_reason = READERS;
 676                         if (spintries == asx_retries)
 677                                 goto sleepq;
 678                         spintries++;
 679                         KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
 680                             "spinning", "lockname:\"%s\"",
 681                             sx->lock_object.lo_name);
 682                         for (i = 0; i < asx_loops; i += n) {
 683                                 n = SX_SHARERS(x);
 684                                 lock_delay_spin(n);
 685                                 x = SX_READ_VALUE(sx);
 686                                 if ((x & SX_LOCK_SHARED) == 0 ||
 687                                     SX_SHARERS(x) == 0)
 688                                         break;
 689                         }
 690 #ifdef KDTRACE_HOOKS
 691                         lda.spin_cnt += i;
 692 #endif
 693                         KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
 694                             "running");
 695                         if (i < asx_loops)
 696                                 continue;
 697                 }
 698 sleepq:
 699 #endif
 700                 sleepq_lock(&sx->lock_object);
 701                 x = SX_READ_VALUE(sx);
 702 retry_sleepq:
 703
 704                 /*
 705                  * If the lock was released while spinning on the
 706                  * sleep queue chain lock, try again.
 707                  */
 708                 if (x == SX_LOCK_UNLOCKED) {
 709                         sleepq_release(&sx->lock_object);
 710                         continue;
 711                 }
 712
 713 #ifdef ADAPTIVE_SX
 714                 /*
 715                  * The current lock owner might have started executing
 716                  * on another CPU (or the lock could have changed
 717                  * owners) while we were waiting on the sleep queue
 718                  * chain lock.  If so, drop the sleep queue lock and try
 719                  * again.
 720                  */
 721                 if (adaptive) {
 722                         if (!(x & SX_LOCK_SHARED)) {
 723                                 owner = (struct thread *)SX_OWNER(x);
 724                                 if (TD_IS_RUNNING(owner)) {
 725                                         sleepq_release(&sx->lock_object);
 726                                         continue;
 727                                 }
 728                         } else if (SX_SHARERS(x) > 0 && sleep_reason == WRITER) {
 729                                 sleepq_release(&sx->lock_object);
 730                                 continue;
 731                         }
 732                 }
 733 #endif
 734
 735                 /*
 736                  * If an exclusive lock was released with both shared
 737                  * and exclusive waiters and a shared waiter hasn't
 738                  * woken up and acquired the lock yet, sx_lock will be
 739                  * set to SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS.
 740                  * If we see that value, try to acquire it once.  Note
 741                  * that we have to preserve SX_LOCK_EXCLUSIVE_WAITERS
 742                  * as there are other exclusive waiters still.  If we
 743                  * fail, restart the loop.
 744                  */
 745                 if (x == (SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS)) {
 746                         if (!atomic_fcmpset_acq_ptr(&sx->sx_lock, &x,
 747                             tid | SX_LOCK_EXCLUSIVE_WAITERS))
 748                                 goto retry_sleepq;
 749                         sleepq_release(&sx->lock_object);
 750                         CTR2(KTR_LOCK, "%s: %p claimed by new writer",
 751                             __func__, sx);
 752                         break;
 753                 }
 754
 755                 /*
 756                  * Try to set the SX_LOCK_EXCLUSIVE_WAITERS.  If we fail,
 757                  * than loop back and retry.
 758                  */
 759                 if (!(x & SX_LOCK_EXCLUSIVE_WAITERS)) {
 760                         if (!atomic_fcmpset_ptr(&sx->sx_lock, &x,
 761                             x | SX_LOCK_EXCLUSIVE_WAITERS)) {
 762                                 goto retry_sleepq;
 763                         }
 764                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 765                                 CTR2(KTR_LOCK, "%s: %p set excl waiters flag",
 766                                     __func__, sx);
 767                 }
 768
 769                 /*
 770                  * Since we have been unable to acquire the exclusive
 771                  * lock and the exclusive waiters flag is set, we have
 772                  * to sleep.
 773                  */
 774                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 775                         CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
 776                             __func__, sx);
 777
 778 #ifdef KDTRACE_HOOKS
 779                 sleep_time -= lockstat_nsecs(&sx->lock_object);
 780 #endif
 781                 sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
 782                     SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
 783                     SLEEPQ_INTERRUPTIBLE : 0), SQ_EXCLUSIVE_QUEUE);
 784                 if (!(opts & SX_INTERRUPTIBLE))
 785                         sleepq_wait(&sx->lock_object, 0);
 786                 else
 787                         error = sleepq_wait_sig(&sx->lock_object, 0);
 788 #ifdef KDTRACE_HOOKS
 789                 sleep_time += lockstat_nsecs(&sx->lock_object);
 790                 sleep_cnt++;
 791 #endif
 792                 if (error) {
 793                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 794                                 CTR2(KTR_LOCK,
 795                         "%s: interruptible sleep by %p suspended by signal",
 796                                     __func__, sx);
 797                         break;
 798                 }
 799                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 800                         CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
 801                             __func__, sx);
 802                 x = SX_READ_VALUE(sx);
 803         }
 804 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
 805         if (__predict_true(!extra_work))
 806                 return (error);
 807 #endif
 808 #ifdef KDTRACE_HOOKS
 809         all_time += lockstat_nsecs(&sx->lock_object);
 810         if (sleep_time)
 811                 LOCKSTAT_RECORD4(sx__block, sx, sleep_time,
 812                     LOCKSTAT_WRITER, (state & SX_LOCK_SHARED) == 0,
 813                     (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
 814         if (lda.spin_cnt > sleep_cnt)
 815                 LOCKSTAT_RECORD4(sx__spin, sx, all_time - sleep_time,
 816                     LOCKSTAT_WRITER, (state & SX_LOCK_SHARED) == 0,
 817                     (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
 818 out_lockstat:
 819 #endif
 820         if (!error)
 821                 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
 822                     contested, waittime, file, line, LOCKSTAT_WRITER);
 823         GIANT_RESTORE();
 824         return (error);
 825 }
 826
 827 /*
 828  * This function represents the so-called 'hard case' for sx_xunlock
 829  * operation.  All 'easy case' failures are redirected to this.  Note
 830  * that ideally this would be a static function, but it needs to be
 831  * accessible from at least sx.h.
 832  */
 833 void
 834 _sx_xunlock_hard(struct sx *sx, uintptr_t x LOCK_FILE_LINE_ARG_DEF)
 835 {
 836         uintptr_t tid, setx;
 837         int queue, wakeup_swapper;
 838
 839         if (SCHEDULER_STOPPED())
 840                 return;
 841
 842         tid = (uintptr_t)curthread;
 843
 844         if (__predict_false(x == tid))
 845                 x = SX_READ_VALUE(sx);
 846
 847         MPASS(!(x & SX_LOCK_SHARED));
 848
 849         if (__predict_false(x & SX_LOCK_RECURSED)) {
 850                 /* The lock is recursed, unrecurse one level. */
 851                 if ((--sx->sx_recurse) == 0)
 852                         atomic_clear_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
 853                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 854                         CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, sx);
 855                 return;
 856         }
 857
 858         LOCKSTAT_PROFILE_RELEASE_RWLOCK(sx__release, sx, LOCKSTAT_WRITER);
 859         if (x == tid &&
 860             atomic_cmpset_rel_ptr(&sx->sx_lock, tid, SX_LOCK_UNLOCKED))
 861                 return;
 862
 863         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 864                 CTR2(KTR_LOCK, "%s: %p contested", __func__, sx);
 865
 866         sleepq_lock(&sx->lock_object);
 867         x = SX_READ_VALUE(sx);
 868         MPASS(x & (SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS));
 869
 870         /*
 871          * The wake up algorithm here is quite simple and probably not
 872          * ideal.  It gives precedence to shared waiters if they are
 873          * present.  For this condition, we have to preserve the
 874          * state of the exclusive waiters flag.
 875          * If interruptible sleeps left the shared queue empty avoid a
 876          * starvation for the threads sleeping on the exclusive queue by giving
 877          * them precedence and cleaning up the shared waiters bit anyway.
 878          */
 879         setx = SX_LOCK_UNLOCKED;
 880         queue = SQ_EXCLUSIVE_QUEUE;
 881         if ((x & SX_LOCK_SHARED_WAITERS) != 0 &&
 882             sleepq_sleepcnt(&sx->lock_object, SQ_SHARED_QUEUE) != 0) {
 883                 queue = SQ_SHARED_QUEUE;
 884                 setx |= (x & SX_LOCK_EXCLUSIVE_WAITERS);
 885         }
 886         atomic_store_rel_ptr(&sx->sx_lock, setx);
 887
 888         /* Wake up all the waiters for the specific queue. */
 889         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 890                 CTR3(KTR_LOCK, "%s: %p waking up all threads on %s queue",
 891                     __func__, sx, queue == SQ_SHARED_QUEUE ? "shared" :
 892                     "exclusive");
 893
 894         wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX, 0,
 895             queue);
 896         sleepq_release(&sx->lock_object);
 897         if (wakeup_swapper)
 898                 kick_proc0();
 899 }
 900
 901 static bool __always_inline
 902 __sx_slock_try(struct sx *sx, uintptr_t *xp LOCK_FILE_LINE_ARG_DEF)
 903 {
 904
 905         /*
 906          * If no other thread has an exclusive lock then try to bump up
 907          * the count of sharers.  Since we have to preserve the state
 908          * of SX_LOCK_EXCLUSIVE_WAITERS, if we fail to acquire the
 909          * shared lock loop back and retry.
 910          */
 911         while (*xp & SX_LOCK_SHARED) {
 912                 MPASS(!(*xp & SX_LOCK_SHARED_WAITERS));
 913                 if (atomic_fcmpset_acq_ptr(&sx->sx_lock, xp,
 914                     *xp + SX_ONE_SHARER)) {
 915                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 916                                 CTR4(KTR_LOCK, "%s: %p succeed %p -> %p",
 917                                     __func__, sx, (void *)*xp,
 918                                     (void *)(*xp + SX_ONE_SHARER));
 919                         return (true);
 920                 }
 921         }
 922         return (false);
 923 }
 924
 925 static int __noinline
 926 _sx_slock_hard(struct sx *sx, int opts, uintptr_t x LOCK_FILE_LINE_ARG_DEF)
 927 {
 928         GIANT_DECLARE;
 929 #ifdef ADAPTIVE_SX
 930         volatile struct thread *owner;
 931         bool adaptive;
 932 #endif
 933 #ifdef LOCK_PROFILING
 934         uint64_t waittime = 0;
 935         int contested = 0;
 936 #endif
 937         int error = 0;
 938 #if defined(ADAPTIVE_SX) || defined(KDTRACE_HOOKS)
 939         struct lock_delay_arg lda;
 940 #endif
 941 #ifdef KDTRACE_HOOKS
 942         u_int sleep_cnt = 0;
 943         int64_t sleep_time = 0;
 944         int64_t all_time = 0;
 945 #endif
 946 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
 947         uintptr_t state;
 948 #endif
 949         int extra_work = 0;
 950
 951 #ifdef KDTRACE_HOOKS
 952         if (LOCKSTAT_PROFILE_ENABLED(sx__acquire)) {
 953                 if (__sx_slock_try(sx, &x LOCK_FILE_LINE_ARG))
 954                         goto out_lockstat;
 955                 extra_work = 1;
 956                 all_time -= lockstat_nsecs(&sx->lock_object);
 957                 state = x;
 958         }
 959 #endif
 960 #ifdef LOCK_PROFILING
 961         extra_work = 1;
 962         state = x;
 963 #endif
 964
 965         if (SCHEDULER_STOPPED())
 966                 return (0);
 967
 968 #if defined(ADAPTIVE_SX)
 969         lock_delay_arg_init(&lda, &sx_delay);
 970 #elif defined(KDTRACE_HOOKS)
 971         lock_delay_arg_init(&lda, NULL);
 972 #endif
 973
 974 #ifdef ADAPTIVE_SX
 975         adaptive = ((sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0);
 976 #endif
 977
 978 #ifdef HWPMC_HOOKS
 979         PMC_SOFT_CALL( , , lock, failed);
 980 #endif
 981         lock_profile_obtain_lock_failed(&sx->lock_object, &contested,
 982             &waittime);
 983
 984 #ifndef INVARIANTS
 985         GIANT_SAVE(extra_work);
 986 #endif
 987
 988         /*
 989          * As with rwlocks, we don't make any attempt to try to block
 990          * shared locks once there is an exclusive waiter.
 991          */
 992         for (;;) {
 993                 if (__sx_slock_try(sx, &x LOCK_FILE_LINE_ARG))
 994                         break;
 995 #ifdef INVARIANTS
 996                 GIANT_SAVE(extra_work);
 997 #endif
 998 #ifdef KDTRACE_HOOKS
 999                 lda.spin_cnt++;
1000 #endif
1001
1002 #ifdef ADAPTIVE_SX
1003                 if (__predict_false(!adaptive))
1004                         goto sleepq;
1005                 /*
1006                  * If the owner is running on another CPU, spin until
1007                  * the owner stops running or the state of the lock
1008                  * changes.
1009                  */
1010                 owner = lv_sx_owner(x);
1011                 if (TD_IS_RUNNING(owner)) {
1012                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
1013                                 CTR3(KTR_LOCK,
1014                                     "%s: spinning on %p held by %p",
1015                                     __func__, sx, owner);
1016                         KTR_STATE1(KTR_SCHED, "thread",
1017                             sched_tdname(curthread), "spinning",
1018                             "lockname:\"%s\"", sx->lock_object.lo_name);
1019                         do {
1020                                 lock_delay(&lda);
1021                                 x = SX_READ_VALUE(sx);
1022                                 owner = lv_sx_owner(x);
1023                         } while (owner != NULL && TD_IS_RUNNING(owner));
1024                         KTR_STATE0(KTR_SCHED, "thread",
1025                             sched_tdname(curthread), "running");
1026                         continue;
1027                 }
1028 sleepq:
1029 #endif
1030
1031                 /*
1032                  * Some other thread already has an exclusive lock, so
1033                  * start the process of blocking.
1034                  */
1035                 sleepq_lock(&sx->lock_object);
1036                 x = SX_READ_VALUE(sx);
1037 retry_sleepq:
1038                 /*
1039                  * The lock could have been released while we spun.
1040                  * In this case loop back and retry.
1041                  */
1042                 if (x & SX_LOCK_SHARED) {
1043                         sleepq_release(&sx->lock_object);
1044                         continue;
1045                 }
1046
1047 #ifdef ADAPTIVE_SX
1048                 /*
1049                  * If the owner is running on another CPU, spin until
1050                  * the owner stops running or the state of the lock
1051                  * changes.
1052                  */
1053                 if (!(x & SX_LOCK_SHARED) && adaptive) {
1054                         owner = (struct thread *)SX_OWNER(x);
1055                         if (TD_IS_RUNNING(owner)) {
1056                                 sleepq_release(&sx->lock_object);
1057                                 x = SX_READ_VALUE(sx);
1058                                 continue;
1059                         }
1060                 }
1061 #endif
1062
1063                 /*
1064                  * Try to set the SX_LOCK_SHARED_WAITERS flag.  If we
1065                  * fail to set it drop the sleep queue lock and loop
1066                  * back.
1067                  */
1068                 if (!(x & SX_LOCK_SHARED_WAITERS)) {
1069                         if (!atomic_fcmpset_ptr(&sx->sx_lock, &x,
1070                             x | SX_LOCK_SHARED_WAITERS))
1071                                 goto retry_sleepq;
1072                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
1073                                 CTR2(KTR_LOCK, "%s: %p set shared waiters flag",
1074                                     __func__, sx);
1075                 }
1076
1077                 /*
1078                  * Since we have been unable to acquire the shared lock,
1079                  * we have to sleep.
1080                  */
1081                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1082                         CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
1083                             __func__, sx);
1084
1085 #ifdef KDTRACE_HOOKS
1086                 sleep_time -= lockstat_nsecs(&sx->lock_object);
1087 #endif
1088                 sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
1089                     SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
1090                     SLEEPQ_INTERRUPTIBLE : 0), SQ_SHARED_QUEUE);
1091                 if (!(opts & SX_INTERRUPTIBLE))
1092                         sleepq_wait(&sx->lock_object, 0);
1093                 else
1094                         error = sleepq_wait_sig(&sx->lock_object, 0);
1095 #ifdef KDTRACE_HOOKS
1096                 sleep_time += lockstat_nsecs(&sx->lock_object);
1097                 sleep_cnt++;
1098 #endif
1099                 if (error) {
1100                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
1101                                 CTR2(KTR_LOCK,
1102                         "%s: interruptible sleep by %p suspended by signal",
1103                                     __func__, sx);
1104                         break;
1105                 }
1106                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1107                         CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
1108                             __func__, sx);
1109                 x = SX_READ_VALUE(sx);
1110         }
1111 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
1112         if (__predict_true(!extra_work))
1113                 return (error);
1114 #endif
1115 #ifdef KDTRACE_HOOKS
1116         all_time += lockstat_nsecs(&sx->lock_object);
1117         if (sleep_time)
1118                 LOCKSTAT_RECORD4(sx__block, sx, sleep_time,
1119                     LOCKSTAT_READER, (state & SX_LOCK_SHARED) == 0,
1120                     (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
1121         if (lda.spin_cnt > sleep_cnt)
1122                 LOCKSTAT_RECORD4(sx__spin, sx, all_time - sleep_time,
1123                     LOCKSTAT_READER, (state & SX_LOCK_SHARED) == 0,
1124                     (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
1125 out_lockstat:
1126 #endif
1127         if (error == 0) {
1128                 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
1129                     contested, waittime, file, line, LOCKSTAT_READER);
1130         }
1131         GIANT_RESTORE();
1132         return (error);
1133 }
1134
1135 int
1136 _sx_slock_int(struct sx *sx, int opts LOCK_FILE_LINE_ARG_DEF)
1137 {
1138         uintptr_t x;
1139         int error;
1140
1141         KASSERT(kdb_active != 0 || SCHEDULER_STOPPED() ||
1142             !TD_IS_IDLETHREAD(curthread),
1143             ("sx_slock() by idle thread %p on sx %s @ %s:%d",
1144             curthread, sx->lock_object.lo_name, file, line));
1145         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
1146             ("sx_slock() of destroyed sx @ %s:%d", file, line));
1147         WITNESS_CHECKORDER(&sx->lock_object, LOP_NEWORDER, file, line, NULL);
1148
1149         error = 0;
1150         x = SX_READ_VALUE(sx);
1151         if (__predict_false(LOCKSTAT_PROFILE_ENABLED(sx__acquire) ||
1152             !__sx_slock_try(sx, &x LOCK_FILE_LINE_ARG)))
1153                 error = _sx_slock_hard(sx, opts, x LOCK_FILE_LINE_ARG);
1154         else
1155                 lock_profile_obtain_lock_success(&sx->lock_object, 0, 0,
1156                     file, line);
1157         if (error == 0) {
1158                 LOCK_LOG_LOCK("SLOCK", &sx->lock_object, 0, 0, file, line);
1159                 WITNESS_LOCK(&sx->lock_object, 0, file, line);
1160                 TD_LOCKS_INC(curthread);
1161         }
1162         return (error);
1163 }
1164
1165 int
1166 _sx_slock(struct sx *sx, int opts, const char *file, int line)
1167 {
1168
1169         return (_sx_slock_int(sx, opts LOCK_FILE_LINE_ARG));
1170 }
1171
1172 static bool __always_inline
1173 _sx_sunlock_try(struct sx *sx, uintptr_t *xp)
1174 {
1175
1176         for (;;) {
1177                 /*
1178                  * We should never have sharers while at least one thread
1179                  * holds a shared lock.
1180                  */
1181                 KASSERT(!(*xp & SX_LOCK_SHARED_WAITERS),
1182                     ("%s: waiting sharers", __func__));
1183
1184                 /*
1185                  * See if there is more than one shared lock held.  If
1186                  * so, just drop one and return.
1187                  */
1188                 if (SX_SHARERS(*xp) > 1) {
1189                         if (atomic_fcmpset_rel_ptr(&sx->sx_lock, xp,
1190                             *xp - SX_ONE_SHARER)) {
1191                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1192                                         CTR4(KTR_LOCK,
1193                                             "%s: %p succeeded %p -> %p",
1194                                             __func__, sx, (void *)*xp,
1195                                             (void *)(*xp - SX_ONE_SHARER));
1196                                 return (true);
1197                         }
1198                         continue;
1199                 }
1200
1201                 /*
1202                  * If there aren't any waiters for an exclusive lock,
1203                  * then try to drop it quickly.
1204                  */
1205                 if (!(*xp & SX_LOCK_EXCLUSIVE_WAITERS)) {
1206                         MPASS(*xp == SX_SHARERS_LOCK(1));
1207                         *xp = SX_SHARERS_LOCK(1);
1208                         if (atomic_fcmpset_rel_ptr(&sx->sx_lock,
1209                             xp, SX_LOCK_UNLOCKED)) {
1210                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1211                                         CTR2(KTR_LOCK, "%s: %p last succeeded",
1212                                             __func__, sx);
1213                                 return (true);
1214                         }
1215                         continue;
1216                 }
1217                 break;
1218         }
1219         return (false);
1220 }
1221
1222 static void __noinline
1223 _sx_sunlock_hard(struct sx *sx, uintptr_t x LOCK_FILE_LINE_ARG_DEF)
1224 {
1225         int wakeup_swapper = 0;
1226         uintptr_t setx;
1227
1228         if (SCHEDULER_STOPPED())
1229                 return;
1230
1231         if (_sx_sunlock_try(sx, &x))
1232                 goto out_lockstat;
1233
1234         /*
1235          * At this point, there should just be one sharer with
1236          * exclusive waiters.
1237          */
1238         MPASS(x == (SX_SHARERS_LOCK(1) | SX_LOCK_EXCLUSIVE_WAITERS));
1239
1240         sleepq_lock(&sx->lock_object);
1241         x = SX_READ_VALUE(sx);
1242         for (;;) {
1243                 MPASS(x & SX_LOCK_EXCLUSIVE_WAITERS);
1244                 MPASS(!(x & SX_LOCK_SHARED_WAITERS));
1245                 if (_sx_sunlock_try(sx, &x))
1246                         break;
1247
1248                 /*
1249                  * Wake up semantic here is quite simple:
1250                  * Just wake up all the exclusive waiters.
1251                  * Note that the state of the lock could have changed,
1252                  * so if it fails loop back and retry.
1253                  */
1254                 setx = x - SX_ONE_SHARER;
1255                 setx &= ~SX_LOCK_EXCLUSIVE_WAITERS;
1256                 if (!atomic_fcmpset_rel_ptr(&sx->sx_lock, &x, setx))
1257                         continue;
1258                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1259                         CTR2(KTR_LOCK, "%s: %p waking up all thread on"
1260                             "exclusive queue", __func__, sx);
1261                 wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX,
1262                     0, SQ_EXCLUSIVE_QUEUE);
1263                 break;
1264         }
1265         sleepq_release(&sx->lock_object);
1266         if (wakeup_swapper)
1267                 kick_proc0();
1268 out_lockstat:
1269         LOCKSTAT_PROFILE_RELEASE_RWLOCK(sx__release, sx, LOCKSTAT_READER);
1270 }
1271
1272 void
1273 _sx_sunlock_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
1274 {
1275         uintptr_t x;
1276
1277         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
1278             ("sx_sunlock() of destroyed sx @ %s:%d", file, line));
1279         _sx_assert(sx, SA_SLOCKED, file, line);
1280         WITNESS_UNLOCK(&sx->lock_object, 0, file, line);
1281         LOCK_LOG_LOCK("SUNLOCK", &sx->lock_object, 0, 0, file, line);
1282
1283         x = SX_READ_VALUE(sx);
1284         if (__predict_false(LOCKSTAT_PROFILE_ENABLED(sx__release) ||
1285             !_sx_sunlock_try(sx, &x)))
1286                 _sx_sunlock_hard(sx, x LOCK_FILE_LINE_ARG);
1287         else
1288                 lock_profile_release_lock(&sx->lock_object);
1289
1290         TD_LOCKS_DEC(curthread);
1291 }
1292
1293 void
1294 _sx_sunlock(struct sx *sx, const char *file, int line)
1295 {
1296
1297         _sx_sunlock_int(sx LOCK_FILE_LINE_ARG);
1298 }
1299
1300 #ifdef INVARIANT_SUPPORT
1301 #ifndef INVARIANTS
1302 #undef  _sx_assert
1303 #endif
1304
1305 /*
1306  * In the non-WITNESS case, sx_assert() can only detect that at least
1307  * *some* thread owns an slock, but it cannot guarantee that *this*
1308  * thread owns an slock.
1309  */
1310 void
1311 _sx_assert(const struct sx *sx, int what, const char *file, int line)
1312 {
1313 #ifndef WITNESS
1314         int slocked = 0;
1315 #endif
1316
1317         if (panicstr != NULL)
1318                 return;
1319         switch (what) {
1320         case SA_SLOCKED:
1321         case SA_SLOCKED | SA_NOTRECURSED:
1322         case SA_SLOCKED | SA_RECURSED:
1323 #ifndef WITNESS
1324                 slocked = 1;
1325                 /* FALLTHROUGH */
1326 #endif
1327         case SA_LOCKED:
1328         case SA_LOCKED | SA_NOTRECURSED:
1329         case SA_LOCKED | SA_RECURSED:
1330 #ifdef WITNESS
1331                 witness_assert(&sx->lock_object, what, file, line);
1332 #else
1333                 /*
1334                  * If some other thread has an exclusive lock or we
1335                  * have one and are asserting a shared lock, fail.
1336                  * Also, if no one has a lock at all, fail.
1337                  */
1338                 if (sx->sx_lock == SX_LOCK_UNLOCKED ||
1339                     (!(sx->sx_lock & SX_LOCK_SHARED) && (slocked ||
1340                     sx_xholder(sx) != curthread)))
1341                         panic("Lock %s not %slocked @ %s:%d\n",
1342                             sx->lock_object.lo_name, slocked ? "share " : "",
1343                             file, line);
1344
1345                 if (!(sx->sx_lock & SX_LOCK_SHARED)) {
1346                         if (sx_recursed(sx)) {
1347                                 if (what & SA_NOTRECURSED)
1348                                         panic("Lock %s recursed @ %s:%d\n",
1349                                             sx->lock_object.lo_name, file,
1350                                             line);
1351                         } else if (what & SA_RECURSED)
1352                                 panic("Lock %s not recursed @ %s:%d\n",
1353                                     sx->lock_object.lo_name, file, line);
1354                 }
1355 #endif
1356                 break;
1357         case SA_XLOCKED:
1358         case SA_XLOCKED | SA_NOTRECURSED:
1359         case SA_XLOCKED | SA_RECURSED:
1360                 if (sx_xholder(sx) != curthread)
1361                         panic("Lock %s not exclusively locked @ %s:%d\n",
1362                             sx->lock_object.lo_name, file, line);
1363                 if (sx_recursed(sx)) {
1364                         if (what & SA_NOTRECURSED)
1365                                 panic("Lock %s recursed @ %s:%d\n",
1366                                     sx->lock_object.lo_name, file, line);
1367                 } else if (what & SA_RECURSED)
1368                         panic("Lock %s not recursed @ %s:%d\n",
1369                             sx->lock_object.lo_name, file, line);
1370                 break;
1371         case SA_UNLOCKED:
1372 #ifdef WITNESS
1373                 witness_assert(&sx->lock_object, what, file, line);
1374 #else
1375                 /*
1376                  * If we hold an exclusve lock fail.  We can't
1377                  * reliably check to see if we hold a shared lock or
1378                  * not.
1379                  */
1380                 if (sx_xholder(sx) == curthread)
1381                         panic("Lock %s exclusively locked @ %s:%d\n",
1382                             sx->lock_object.lo_name, file, line);
1383 #endif
1384                 break;
1385         default:
1386                 panic("Unknown sx lock assertion: %d @ %s:%d", what, file,
1387                     line);
1388         }
1389 }
1390 #endif  /* INVARIANT_SUPPORT */
1391
1392 #ifdef DDB
1393 static void
1394 db_show_sx(const struct lock_object *lock)
1395 {
1396         struct thread *td;
1397         const struct sx *sx;
1398
1399         sx = (const struct sx *)lock;
1400
1401         db_printf(" state: ");
1402         if (sx->sx_lock == SX_LOCK_UNLOCKED)
1403                 db_printf("UNLOCKED\n");
1404         else if (sx->sx_lock == SX_LOCK_DESTROYED) {
1405                 db_printf("DESTROYED\n");
1406                 return;
1407         } else if (sx->sx_lock & SX_LOCK_SHARED)
1408                 db_printf("SLOCK: %ju\n", (uintmax_t)SX_SHARERS(sx->sx_lock));
1409         else {
1410                 td = sx_xholder(sx);
1411                 db_printf("XLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1412                     td->td_tid, td->td_proc->p_pid, td->td_name);
1413                 if (sx_recursed(sx))
1414                         db_printf(" recursed: %d\n", sx->sx_recurse);
1415         }
1416
1417         db_printf(" waiters: ");
1418         switch(sx->sx_lock &
1419             (SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS)) {
1420         case SX_LOCK_SHARED_WAITERS:
1421                 db_printf("shared\n");
1422                 break;
1423         case SX_LOCK_EXCLUSIVE_WAITERS:
1424                 db_printf("exclusive\n");
1425                 break;
1426         case SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS:
1427                 db_printf("exclusive and shared\n");
1428                 break;
1429         default:
1430                 db_printf("none\n");
1431         }
1432 }
1433
1434 /*
1435  * Check to see if a thread that is blocked on a sleep queue is actually
1436  * blocked on an sx lock.  If so, output some details and return true.
1437  * If the lock has an exclusive owner, return that in *ownerp.
1438  */
1439 int
1440 sx_chain(struct thread *td, struct thread **ownerp)
1441 {
1442         struct sx *sx;
1443
1444         /*
1445          * Check to see if this thread is blocked on an sx lock.
1446          * First, we check the lock class.  If that is ok, then we
1447          * compare the lock name against the wait message.
1448          */
1449         sx = td->td_wchan;
1450         if (LOCK_CLASS(&sx->lock_object) != &lock_class_sx ||
1451             sx->lock_object.lo_name != td->td_wmesg)
1452                 return (0);
1453
1454         /* We think we have an sx lock, so output some details. */
1455         db_printf("blocked on sx \"%s\" ", td->td_wmesg);
1456         *ownerp = sx_xholder(sx);
1457         if (sx->sx_lock & SX_LOCK_SHARED)
1458                 db_printf("SLOCK (count %ju)\n",
1459                     (uintmax_t)SX_SHARERS(sx->sx_lock));
1460         else
1461                 db_printf("XLOCK\n");
1462         return (1);
1463 }
1464 #endif