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 = 10;
 149 static __read_frequently u_int asx_loops = 10000;
 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 LOCK_DELAY_SYSINIT_DEFAULT(sx_delay);
 162 #endif
 163
 164 void
 165 assert_sx(const struct lock_object *lock, int what)
 166 {
 167
 168         sx_assert((const struct sx *)lock, what);
 169 }
 170
 171 void
 172 lock_sx(struct lock_object *lock, uintptr_t how)
 173 {
 174         struct sx *sx;
 175
 176         sx = (struct sx *)lock;
 177         if (how)
 178                 sx_slock(sx);
 179         else
 180                 sx_xlock(sx);
 181 }
 182
 183 uintptr_t
 184 unlock_sx(struct lock_object *lock)
 185 {
 186         struct sx *sx;
 187
 188         sx = (struct sx *)lock;
 189         sx_assert(sx, SA_LOCKED | SA_NOTRECURSED);
 190         if (sx_xlocked(sx)) {
 191                 sx_xunlock(sx);
 192                 return (0);
 193         } else {
 194                 sx_sunlock(sx);
 195                 return (1);
 196         }
 197 }
 198
 199 #ifdef KDTRACE_HOOKS
 200 int
 201 owner_sx(const struct lock_object *lock, struct thread **owner)
 202 {
 203         const struct sx *sx;
 204         uintptr_t x;
 205
 206         sx = (const struct sx *)lock;
 207         x = sx->sx_lock;
 208         *owner = NULL;
 209         return ((x & SX_LOCK_SHARED) != 0 ? (SX_SHARERS(x) != 0) :
 210             ((*owner = (struct thread *)SX_OWNER(x)) != NULL));
 211 }
 212 #endif
 213
 214 void
 215 sx_sysinit(void *arg)
 216 {
 217         struct sx_args *sargs = arg;
 218
 219         sx_init_flags(sargs->sa_sx, sargs->sa_desc, sargs->sa_flags);
 220 }
 221
 222 void
 223 sx_init_flags(struct sx *sx, const char *description, int opts)
 224 {
 225         int flags;
 226
 227         MPASS((opts & ~(SX_QUIET | SX_RECURSE | SX_NOWITNESS | SX_DUPOK |
 228             SX_NOPROFILE | SX_NOADAPTIVE | SX_NEW)) == 0);
 229         ASSERT_ATOMIC_LOAD_PTR(sx->sx_lock,
 230             ("%s: sx_lock not aligned for %s: %p", __func__, description,
 231             &sx->sx_lock));
 232
 233         flags = LO_SLEEPABLE | LO_UPGRADABLE;
 234         if (opts & SX_DUPOK)
 235                 flags |= LO_DUPOK;
 236         if (opts & SX_NOPROFILE)
 237                 flags |= LO_NOPROFILE;
 238         if (!(opts & SX_NOWITNESS))
 239                 flags |= LO_WITNESS;
 240         if (opts & SX_RECURSE)
 241                 flags |= LO_RECURSABLE;
 242         if (opts & SX_QUIET)
 243                 flags |= LO_QUIET;
 244         if (opts & SX_NEW)
 245                 flags |= LO_NEW;
 246
 247         flags |= opts & SX_NOADAPTIVE;
 248         lock_init(&sx->lock_object, &lock_class_sx, description, NULL, flags);
 249         sx->sx_lock = SX_LOCK_UNLOCKED;
 250         sx->sx_recurse = 0;
 251 }
 252
 253 void
 254 sx_destroy(struct sx *sx)
 255 {
 256
 257         KASSERT(sx->sx_lock == SX_LOCK_UNLOCKED, ("sx lock still held"));
 258         KASSERT(sx->sx_recurse == 0, ("sx lock still recursed"));
 259         sx->sx_lock = SX_LOCK_DESTROYED;
 260         lock_destroy(&sx->lock_object);
 261 }
 262
 263 int
 264 sx_try_slock_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
 265 {
 266         uintptr_t x;
 267
 268         if (SCHEDULER_STOPPED())
 269                 return (1);
 270
 271         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
 272             ("sx_try_slock() by idle thread %p on sx %s @ %s:%d",
 273             curthread, sx->lock_object.lo_name, file, line));
 274
 275         x = sx->sx_lock;
 276         for (;;) {
 277                 KASSERT(x != SX_LOCK_DESTROYED,
 278                     ("sx_try_slock() of destroyed sx @ %s:%d", file, line));
 279                 if (!(x & SX_LOCK_SHARED))
 280                         break;
 281                 if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, x + SX_ONE_SHARER)) {
 282                         LOCK_LOG_TRY("SLOCK", &sx->lock_object, 0, 1, file, line);
 283                         WITNESS_LOCK(&sx->lock_object, LOP_TRYLOCK, file, line);
 284                         LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire,
 285                             sx, 0, 0, file, line, LOCKSTAT_READER);
 286                         TD_LOCKS_INC(curthread);
 287                         return (1);
 288                 }
 289         }
 290
 291         LOCK_LOG_TRY("SLOCK", &sx->lock_object, 0, 0, file, line);
 292         return (0);
 293 }
 294
 295 int
 296 sx_try_slock_(struct sx *sx, const char *file, int line)
 297 {
 298
 299         return (sx_try_slock_int(sx LOCK_FILE_LINE_ARG));
 300 }
 301
 302 int
 303 _sx_xlock(struct sx *sx, int opts, const char *file, int line)
 304 {
 305         uintptr_t tid, x;
 306         int error = 0;
 307
 308         KASSERT(kdb_active != 0 || SCHEDULER_STOPPED() ||
 309             !TD_IS_IDLETHREAD(curthread),
 310             ("sx_xlock() by idle thread %p on sx %s @ %s:%d",
 311             curthread, sx->lock_object.lo_name, file, line));
 312         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 313             ("sx_xlock() of destroyed sx @ %s:%d", file, line));
 314         WITNESS_CHECKORDER(&sx->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE, file,
 315             line, NULL);
 316         tid = (uintptr_t)curthread;
 317         x = SX_LOCK_UNLOCKED;
 318         if (!atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid))
 319                 error = _sx_xlock_hard(sx, x, opts LOCK_FILE_LINE_ARG);
 320         else
 321                 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
 322                     0, 0, file, line, LOCKSTAT_WRITER);
 323         if (!error) {
 324                 LOCK_LOG_LOCK("XLOCK", &sx->lock_object, 0, sx->sx_recurse,
 325                     file, line);
 326                 WITNESS_LOCK(&sx->lock_object, LOP_EXCLUSIVE, file, line);
 327                 TD_LOCKS_INC(curthread);
 328         }
 329
 330         return (error);
 331 }
 332
 333 int
 334 sx_try_xlock_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
 335 {
 336         struct thread *td;
 337         uintptr_t tid, x;
 338         int rval;
 339         bool recursed;
 340
 341         td = curthread;
 342         tid = (uintptr_t)td;
 343         if (SCHEDULER_STOPPED_TD(td))
 344                 return (1);
 345
 346         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(td),
 347             ("sx_try_xlock() by idle thread %p on sx %s @ %s:%d",
 348             curthread, sx->lock_object.lo_name, file, line));
 349         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 350             ("sx_try_xlock() of destroyed sx @ %s:%d", file, line));
 351
 352         rval = 1;
 353         recursed = false;
 354         x = SX_LOCK_UNLOCKED;
 355         for (;;) {
 356                 if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid))
 357                         break;
 358                 if (x == SX_LOCK_UNLOCKED)
 359                         continue;
 360                 if (x == tid && (sx->lock_object.lo_flags & LO_RECURSABLE)) {
 361                         sx->sx_recurse++;
 362                         atomic_set_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
 363                         break;
 364                 }
 365                 rval = 0;
 366                 break;
 367         }
 368
 369         LOCK_LOG_TRY("XLOCK", &sx->lock_object, 0, rval, file, line);
 370         if (rval) {
 371                 WITNESS_LOCK(&sx->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
 372                     file, line);
 373                 if (!recursed)
 374                         LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire,
 375                             sx, 0, 0, file, line, LOCKSTAT_WRITER);
 376                 TD_LOCKS_INC(curthread);
 377         }
 378
 379         return (rval);
 380 }
 381
 382 int
 383 sx_try_xlock_(struct sx *sx, const char *file, int line)
 384 {
 385
 386         return (sx_try_xlock_int(sx LOCK_FILE_LINE_ARG));
 387 }
 388
 389 void
 390 _sx_xunlock(struct sx *sx, const char *file, int line)
 391 {
 392
 393         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 394             ("sx_xunlock() of destroyed sx @ %s:%d", file, line));
 395         _sx_assert(sx, SA_XLOCKED, file, line);
 396         WITNESS_UNLOCK(&sx->lock_object, LOP_EXCLUSIVE, file, line);
 397         LOCK_LOG_LOCK("XUNLOCK", &sx->lock_object, 0, sx->sx_recurse, file,
 398             line);
 399 #if LOCK_DEBUG > 0
 400         _sx_xunlock_hard(sx, (uintptr_t)curthread, file, line);
 401 #else
 402         __sx_xunlock(sx, curthread, file, line);
 403 #endif
 404         TD_LOCKS_DEC(curthread);
 405 }
 406
 407 /*
 408  * Try to do a non-blocking upgrade from a shared lock to an exclusive lock.
 409  * This will only succeed if this thread holds a single shared lock.
 410  * Return 1 if if the upgrade succeed, 0 otherwise.
 411  */
 412 int
 413 sx_try_upgrade_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
 414 {
 415         uintptr_t x;
 416         uintptr_t waiters;
 417         int success;
 418
 419         if (SCHEDULER_STOPPED())
 420                 return (1);
 421
 422         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 423             ("sx_try_upgrade() of destroyed sx @ %s:%d", file, line));
 424         _sx_assert(sx, SA_SLOCKED, file, line);
 425
 426         /*
 427          * Try to switch from one shared lock to an exclusive lock.  We need
 428          * to maintain the SX_LOCK_EXCLUSIVE_WAITERS flag if set so that
 429          * we will wake up the exclusive waiters when we drop the lock.
 430          */
 431         success = 0;
 432         x = SX_READ_VALUE(sx);
 433         for (;;) {
 434                 if (SX_SHARERS(x) > 1)
 435                         break;
 436                 waiters = (x & SX_LOCK_EXCLUSIVE_WAITERS);
 437                 if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x,
 438                     (uintptr_t)curthread | waiters)) {
 439                         success = 1;
 440                         break;
 441                 }
 442         }
 443         LOCK_LOG_TRY("XUPGRADE", &sx->lock_object, 0, success, file, line);
 444         if (success) {
 445                 WITNESS_UPGRADE(&sx->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
 446                     file, line);
 447                 LOCKSTAT_RECORD0(sx__upgrade, sx);
 448         }
 449         return (success);
 450 }
 451
 452 int
 453 sx_try_upgrade_(struct sx *sx, const char *file, int line)
 454 {
 455
 456         return (sx_try_upgrade_int(sx LOCK_FILE_LINE_ARG));
 457 }
 458
 459 /*
 460  * Downgrade an unrecursed exclusive lock into a single shared lock.
 461  */
 462 void
 463 sx_downgrade_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
 464 {
 465         uintptr_t x;
 466         int wakeup_swapper;
 467
 468         if (SCHEDULER_STOPPED())
 469                 return;
 470
 471         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 472             ("sx_downgrade() of destroyed sx @ %s:%d", file, line));
 473         _sx_assert(sx, SA_XLOCKED | SA_NOTRECURSED, file, line);
 474 #ifndef INVARIANTS
 475         if (sx_recursed(sx))
 476                 panic("downgrade of a recursed lock");
 477 #endif
 478
 479         WITNESS_DOWNGRADE(&sx->lock_object, 0, file, line);
 480
 481         /*
 482          * Try to switch from an exclusive lock with no shared waiters
 483          * to one sharer with no shared waiters.  If there are
 484          * exclusive waiters, we don't need to lock the sleep queue so
 485          * long as we preserve the flag.  We do one quick try and if
 486          * that fails we grab the sleepq lock to keep the flags from
 487          * changing and do it the slow way.
 488          *
 489          * We have to lock the sleep queue if there are shared waiters
 490          * so we can wake them up.
 491          */
 492         x = sx->sx_lock;
 493         if (!(x & SX_LOCK_SHARED_WAITERS) &&
 494             atomic_cmpset_rel_ptr(&sx->sx_lock, x, SX_SHARERS_LOCK(1) |
 495             (x & SX_LOCK_EXCLUSIVE_WAITERS)))
 496                 goto out;
 497
 498         /*
 499          * Lock the sleep queue so we can read the waiters bits
 500          * without any races and wakeup any shared waiters.
 501          */
 502         sleepq_lock(&sx->lock_object);
 503
 504         /*
 505          * Preserve SX_LOCK_EXCLUSIVE_WAITERS while downgraded to a single
 506          * shared lock.  If there are any shared waiters, wake them up.
 507          */
 508         wakeup_swapper = 0;
 509         x = sx->sx_lock;
 510         atomic_store_rel_ptr(&sx->sx_lock, SX_SHARERS_LOCK(1) |
 511             (x & SX_LOCK_EXCLUSIVE_WAITERS));
 512         if (x & SX_LOCK_SHARED_WAITERS)
 513                 wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX,
 514                     0, SQ_SHARED_QUEUE);
 515         sleepq_release(&sx->lock_object);
 516
 517         if (wakeup_swapper)
 518                 kick_proc0();
 519
 520 out:
 521         LOCK_LOG_LOCK("XDOWNGRADE", &sx->lock_object, 0, 0, file, line);
 522         LOCKSTAT_RECORD0(sx__downgrade, sx);
 523 }
 524
 525 void
 526 sx_downgrade_(struct sx *sx, const char *file, int line)
 527 {
 528
 529         sx_downgrade_int(sx LOCK_FILE_LINE_ARG);
 530 }
 531
 532 /*
 533  * This function represents the so-called 'hard case' for sx_xlock
 534  * operation.  All 'easy case' failures are redirected to this.  Note
 535  * that ideally this would be a static function, but it needs to be
 536  * accessible from at least sx.h.
 537  */
 538 int
 539 _sx_xlock_hard(struct sx *sx, uintptr_t x, int opts LOCK_FILE_LINE_ARG_DEF)
 540 {
 541         GIANT_DECLARE;
 542         uintptr_t tid;
 543 #ifdef ADAPTIVE_SX
 544         volatile struct thread *owner;
 545         u_int i, n, spintries = 0;
 546         enum { READERS, WRITER } sleep_reason;
 547         bool adaptive;
 548 #endif
 549 #ifdef LOCK_PROFILING
 550         uint64_t waittime = 0;
 551         int contested = 0;
 552 #endif
 553         int error = 0;
 554 #if defined(ADAPTIVE_SX) || defined(KDTRACE_HOOKS)
 555         struct lock_delay_arg lda;
 556 #endif
 557 #ifdef  KDTRACE_HOOKS
 558         u_int sleep_cnt = 0;
 559         int64_t sleep_time = 0;
 560         int64_t all_time = 0;
 561 #endif
 562 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
 563         uintptr_t state;
 564 #endif
 565         int extra_work = 0;
 566
 567         tid = (uintptr_t)curthread;
 568
 569 #ifdef KDTRACE_HOOKS
 570         if (LOCKSTAT_PROFILE_ENABLED(sx__acquire)) {
 571                 while (x == SX_LOCK_UNLOCKED) {
 572                         if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid))
 573                                 goto out_lockstat;
 574                 }
 575                 extra_work = 1;
 576                 all_time -= lockstat_nsecs(&sx->lock_object);
 577                 state = x;
 578         }
 579 #endif
 580 #ifdef LOCK_PROFILING
 581         extra_work = 1;
 582         state = x;
 583 #endif
 584
 585         if (SCHEDULER_STOPPED())
 586                 return (0);
 587
 588 #if defined(ADAPTIVE_SX)
 589         lock_delay_arg_init(&lda, &sx_delay);
 590 #elif defined(KDTRACE_HOOKS)
 591         lock_delay_arg_init(&lda, NULL);
 592 #endif
 593
 594         if (__predict_false(x == SX_LOCK_UNLOCKED))
 595                 x = SX_READ_VALUE(sx);
 596
 597         /* If we already hold an exclusive lock, then recurse. */
 598         if (__predict_false(lv_sx_owner(x) == (struct thread *)tid)) {
 599                 KASSERT((sx->lock_object.lo_flags & LO_RECURSABLE) != 0,
 600             ("_sx_xlock_hard: recursed on non-recursive sx %s @ %s:%d\n",
 601                     sx->lock_object.lo_name, file, line));
 602                 sx->sx_recurse++;
 603                 atomic_set_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
 604                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 605                         CTR2(KTR_LOCK, "%s: %p recursing", __func__, sx);
 606                 return (0);
 607         }
 608
 609         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 610                 CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__,
 611                     sx->lock_object.lo_name, (void *)sx->sx_lock, file, line);
 612
 613 #ifdef ADAPTIVE_SX
 614         adaptive = ((sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0);
 615 #endif
 616
 617 #ifdef HWPMC_HOOKS
 618         PMC_SOFT_CALL( , , lock, failed);
 619 #endif
 620         lock_profile_obtain_lock_failed(&sx->lock_object, &contested,
 621             &waittime);
 622
 623 #ifndef INVARIANTS
 624         GIANT_SAVE(extra_work);
 625 #endif
 626
 627         for (;;) {
 628                 if (x == SX_LOCK_UNLOCKED) {
 629                         if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid))
 630                                 break;
 631                         continue;
 632                 }
 633 #ifdef INVARIANTS
 634                 GIANT_SAVE(extra_work);
 635 #endif
 636 #ifdef KDTRACE_HOOKS
 637                 lda.spin_cnt++;
 638 #endif
 639 #ifdef ADAPTIVE_SX
 640                 if (__predict_false(!adaptive))
 641                         goto sleepq;
 642                 /*
 643                  * If the lock is write locked and the owner is
 644                  * running on another CPU, spin until the owner stops
 645                  * running or the state of the lock changes.
 646                  */
 647                 if ((x & SX_LOCK_SHARED) == 0) {
 648                         sleep_reason = WRITER;
 649                         owner = lv_sx_owner(x);
 650                         if (!TD_IS_RUNNING(owner))
 651                                 goto sleepq;
 652                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 653                                 CTR3(KTR_LOCK, "%s: spinning on %p held by %p",
 654                                     __func__, sx, owner);
 655                         KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
 656                             "spinning", "lockname:\"%s\"",
 657                             sx->lock_object.lo_name);
 658                         do {
 659                                 lock_delay(&lda);
 660                                 x = SX_READ_VALUE(sx);
 661                                 owner = lv_sx_owner(x);
 662                         } while (owner != NULL && TD_IS_RUNNING(owner));
 663                         KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
 664                             "running");
 665                         continue;
 666                 } else if (SX_SHARERS(x) > 0) {
 667                         sleep_reason = READERS;
 668                         if (spintries == asx_retries)
 669                                 goto sleepq;
 670                         spintries++;
 671                         KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
 672                             "spinning", "lockname:\"%s\"",
 673                             sx->lock_object.lo_name);
 674                         for (i = 0; i < asx_loops; i += n) {
 675                                 n = SX_SHARERS(x);
 676                                 lock_delay_spin(n);
 677                                 x = SX_READ_VALUE(sx);
 678                                 if ((x & SX_LOCK_SHARED) == 0 ||
 679                                     SX_SHARERS(x) == 0)
 680                                         break;
 681                         }
 682 #ifdef KDTRACE_HOOKS
 683                         lda.spin_cnt += i;
 684 #endif
 685                         KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
 686                             "running");
 687                         if (i < asx_loops)
 688                                 continue;
 689                 }
 690 sleepq:
 691 #endif
 692                 sleepq_lock(&sx->lock_object);
 693                 x = SX_READ_VALUE(sx);
 694 retry_sleepq:
 695
 696                 /*
 697                  * If the lock was released while spinning on the
 698                  * sleep queue chain lock, try again.
 699                  */
 700                 if (x == SX_LOCK_UNLOCKED) {
 701                         sleepq_release(&sx->lock_object);
 702                         continue;
 703                 }
 704
 705 #ifdef ADAPTIVE_SX
 706                 /*
 707                  * The current lock owner might have started executing
 708                  * on another CPU (or the lock could have changed
 709                  * owners) while we were waiting on the sleep queue
 710                  * chain lock.  If so, drop the sleep queue lock and try
 711                  * again.
 712                  */
 713                 if (adaptive) {
 714                         if (!(x & SX_LOCK_SHARED)) {
 715                                 owner = (struct thread *)SX_OWNER(x);
 716                                 if (TD_IS_RUNNING(owner)) {
 717                                         sleepq_release(&sx->lock_object);
 718                                         continue;
 719                                 }
 720                         } else if (SX_SHARERS(x) > 0 && sleep_reason == WRITER) {
 721                                 sleepq_release(&sx->lock_object);
 722                                 continue;
 723                         }
 724                 }
 725 #endif
 726
 727                 /*
 728                  * If an exclusive lock was released with both shared
 729                  * and exclusive waiters and a shared waiter hasn't
 730                  * woken up and acquired the lock yet, sx_lock will be
 731                  * set to SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS.
 732                  * If we see that value, try to acquire it once.  Note
 733                  * that we have to preserve SX_LOCK_EXCLUSIVE_WAITERS
 734                  * as there are other exclusive waiters still.  If we
 735                  * fail, restart the loop.
 736                  */
 737                 if (x == (SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS)) {
 738                         if (!atomic_fcmpset_acq_ptr(&sx->sx_lock, &x,
 739                             tid | SX_LOCK_EXCLUSIVE_WAITERS))
 740                                 goto retry_sleepq;
 741                         sleepq_release(&sx->lock_object);
 742                         CTR2(KTR_LOCK, "%s: %p claimed by new writer",
 743                             __func__, sx);
 744                         break;
 745                 }
 746
 747                 /*
 748                  * Try to set the SX_LOCK_EXCLUSIVE_WAITERS.  If we fail,
 749                  * than loop back and retry.
 750                  */
 751                 if (!(x & SX_LOCK_EXCLUSIVE_WAITERS)) {
 752                         if (!atomic_fcmpset_ptr(&sx->sx_lock, &x,
 753                             x | SX_LOCK_EXCLUSIVE_WAITERS)) {
 754                                 goto retry_sleepq;
 755                         }
 756                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 757                                 CTR2(KTR_LOCK, "%s: %p set excl waiters flag",
 758                                     __func__, sx);
 759                 }
 760
 761                 /*
 762                  * Since we have been unable to acquire the exclusive
 763                  * lock and the exclusive waiters flag is set, we have
 764                  * to sleep.
 765                  */
 766                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 767                         CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
 768                             __func__, sx);
 769
 770 #ifdef KDTRACE_HOOKS
 771                 sleep_time -= lockstat_nsecs(&sx->lock_object);
 772 #endif
 773                 sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
 774                     SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
 775                     SLEEPQ_INTERRUPTIBLE : 0), SQ_EXCLUSIVE_QUEUE);
 776                 if (!(opts & SX_INTERRUPTIBLE))
 777                         sleepq_wait(&sx->lock_object, 0);
 778                 else
 779                         error = sleepq_wait_sig(&sx->lock_object, 0);
 780 #ifdef KDTRACE_HOOKS
 781                 sleep_time += lockstat_nsecs(&sx->lock_object);
 782                 sleep_cnt++;
 783 #endif
 784                 if (error) {
 785                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 786                                 CTR2(KTR_LOCK,
 787                         "%s: interruptible sleep by %p suspended by signal",
 788                                     __func__, sx);
 789                         break;
 790                 }
 791                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 792                         CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
 793                             __func__, sx);
 794                 x = SX_READ_VALUE(sx);
 795         }
 796 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
 797         if (__predict_true(!extra_work))
 798                 return (error);
 799 #endif
 800 #ifdef KDTRACE_HOOKS
 801         all_time += lockstat_nsecs(&sx->lock_object);
 802         if (sleep_time)
 803                 LOCKSTAT_RECORD4(sx__block, sx, sleep_time,
 804                     LOCKSTAT_WRITER, (state & SX_LOCK_SHARED) == 0,
 805                     (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
 806         if (lda.spin_cnt > sleep_cnt)
 807                 LOCKSTAT_RECORD4(sx__spin, sx, all_time - sleep_time,
 808                     LOCKSTAT_WRITER, (state & SX_LOCK_SHARED) == 0,
 809                     (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
 810 out_lockstat:
 811 #endif
 812         if (!error)
 813                 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
 814                     contested, waittime, file, line, LOCKSTAT_WRITER);
 815         GIANT_RESTORE();
 816         return (error);
 817 }
 818
 819 /*
 820  * This function represents the so-called 'hard case' for sx_xunlock
 821  * operation.  All 'easy case' failures are redirected to this.  Note
 822  * that ideally this would be a static function, but it needs to be
 823  * accessible from at least sx.h.
 824  */
 825 void
 826 _sx_xunlock_hard(struct sx *sx, uintptr_t x LOCK_FILE_LINE_ARG_DEF)
 827 {
 828         uintptr_t tid, setx;
 829         int queue, wakeup_swapper;
 830
 831         if (SCHEDULER_STOPPED())
 832                 return;
 833
 834         tid = (uintptr_t)curthread;
 835
 836         if (__predict_false(x == tid))
 837                 x = SX_READ_VALUE(sx);
 838
 839         MPASS(!(x & SX_LOCK_SHARED));
 840
 841         if (__predict_false(x & SX_LOCK_RECURSED)) {
 842                 /* The lock is recursed, unrecurse one level. */
 843                 if ((--sx->sx_recurse) == 0)
 844                         atomic_clear_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
 845                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 846                         CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, sx);
 847                 return;
 848         }
 849
 850         LOCKSTAT_PROFILE_RELEASE_RWLOCK(sx__release, sx, LOCKSTAT_WRITER);
 851         if (x == tid &&
 852             atomic_cmpset_rel_ptr(&sx->sx_lock, tid, SX_LOCK_UNLOCKED))
 853                 return;
 854
 855         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 856                 CTR2(KTR_LOCK, "%s: %p contested", __func__, sx);
 857
 858         sleepq_lock(&sx->lock_object);
 859         x = SX_READ_VALUE(sx);
 860         MPASS(x & (SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS));
 861
 862         /*
 863          * The wake up algorithm here is quite simple and probably not
 864          * ideal.  It gives precedence to shared waiters if they are
 865          * present.  For this condition, we have to preserve the
 866          * state of the exclusive waiters flag.
 867          * If interruptible sleeps left the shared queue empty avoid a
 868          * starvation for the threads sleeping on the exclusive queue by giving
 869          * them precedence and cleaning up the shared waiters bit anyway.
 870          */
 871         setx = SX_LOCK_UNLOCKED;
 872         queue = SQ_EXCLUSIVE_QUEUE;
 873         if ((x & SX_LOCK_SHARED_WAITERS) != 0 &&
 874             sleepq_sleepcnt(&sx->lock_object, SQ_SHARED_QUEUE) != 0) {
 875                 queue = SQ_SHARED_QUEUE;
 876                 setx |= (x & SX_LOCK_EXCLUSIVE_WAITERS);
 877         }
 878         atomic_store_rel_ptr(&sx->sx_lock, setx);
 879
 880         /* Wake up all the waiters for the specific queue. */
 881         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 882                 CTR3(KTR_LOCK, "%s: %p waking up all threads on %s queue",
 883                     __func__, sx, queue == SQ_SHARED_QUEUE ? "shared" :
 884                     "exclusive");
 885
 886         wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX, 0,
 887             queue);
 888         sleepq_release(&sx->lock_object);
 889         if (wakeup_swapper)
 890                 kick_proc0();
 891 }
 892
 893 static bool __always_inline
 894 __sx_slock_try(struct sx *sx, uintptr_t *xp LOCK_FILE_LINE_ARG_DEF)
 895 {
 896
 897         /*
 898          * If no other thread has an exclusive lock then try to bump up
 899          * the count of sharers.  Since we have to preserve the state
 900          * of SX_LOCK_EXCLUSIVE_WAITERS, if we fail to acquire the
 901          * shared lock loop back and retry.
 902          */
 903         while (*xp & SX_LOCK_SHARED) {
 904                 MPASS(!(*xp & SX_LOCK_SHARED_WAITERS));
 905                 if (atomic_fcmpset_acq_ptr(&sx->sx_lock, xp,
 906                     *xp + SX_ONE_SHARER)) {
 907                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 908                                 CTR4(KTR_LOCK, "%s: %p succeed %p -> %p",
 909                                     __func__, sx, (void *)*xp,
 910                                     (void *)(*xp + SX_ONE_SHARER));
 911                         return (true);
 912                 }
 913         }
 914         return (false);
 915 }
 916
 917 static int __noinline
 918 _sx_slock_hard(struct sx *sx, int opts, uintptr_t x LOCK_FILE_LINE_ARG_DEF)
 919 {
 920         GIANT_DECLARE;
 921 #ifdef ADAPTIVE_SX
 922         volatile struct thread *owner;
 923         bool adaptive;
 924 #endif
 925 #ifdef LOCK_PROFILING
 926         uint64_t waittime = 0;
 927         int contested = 0;
 928 #endif
 929         int error = 0;
 930 #if defined(ADAPTIVE_SX) || defined(KDTRACE_HOOKS)
 931         struct lock_delay_arg lda;
 932 #endif
 933 #ifdef KDTRACE_HOOKS
 934         u_int sleep_cnt = 0;
 935         int64_t sleep_time = 0;
 936         int64_t all_time = 0;
 937 #endif
 938 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
 939         uintptr_t state;
 940 #endif
 941         int extra_work = 0;
 942
 943 #ifdef KDTRACE_HOOKS
 944         if (LOCKSTAT_PROFILE_ENABLED(sx__acquire)) {
 945                 if (__sx_slock_try(sx, &x LOCK_FILE_LINE_ARG))
 946                         goto out_lockstat;
 947                 extra_work = 1;
 948                 all_time -= lockstat_nsecs(&sx->lock_object);
 949                 state = x;
 950         }
 951 #endif
 952 #ifdef LOCK_PROFILING
 953         extra_work = 1;
 954         state = x;
 955 #endif
 956
 957         if (SCHEDULER_STOPPED())
 958                 return (0);
 959
 960 #if defined(ADAPTIVE_SX)
 961         lock_delay_arg_init(&lda, &sx_delay);
 962 #elif defined(KDTRACE_HOOKS)
 963         lock_delay_arg_init(&lda, NULL);
 964 #endif
 965
 966 #ifdef ADAPTIVE_SX
 967         adaptive = ((sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0);
 968 #endif
 969
 970 #ifdef HWPMC_HOOKS
 971         PMC_SOFT_CALL( , , lock, failed);
 972 #endif
 973         lock_profile_obtain_lock_failed(&sx->lock_object, &contested,
 974             &waittime);
 975
 976 #ifndef INVARIANTS
 977         GIANT_SAVE(extra_work);
 978 #endif
 979
 980         /*
 981          * As with rwlocks, we don't make any attempt to try to block
 982          * shared locks once there is an exclusive waiter.
 983          */
 984         for (;;) {
 985                 if (__sx_slock_try(sx, &x LOCK_FILE_LINE_ARG))
 986                         break;
 987 #ifdef INVARIANTS
 988                 GIANT_SAVE(extra_work);
 989 #endif
 990 #ifdef KDTRACE_HOOKS
 991                 lda.spin_cnt++;
 992 #endif
 993
 994 #ifdef ADAPTIVE_SX
 995                 if (__predict_false(!adaptive))
 996                         goto sleepq;
 997                 /*
 998                  * If the owner is running on another CPU, spin until
 999                  * the owner stops running or the state of the lock
1000                  * changes.
1001                  */
1002                 owner = lv_sx_owner(x);
1003                 if (TD_IS_RUNNING(owner)) {
1004                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
1005                                 CTR3(KTR_LOCK,
1006                                     "%s: spinning on %p held by %p",
1007                                     __func__, sx, owner);
1008                         KTR_STATE1(KTR_SCHED, "thread",
1009                             sched_tdname(curthread), "spinning",
1010                             "lockname:\"%s\"", sx->lock_object.lo_name);
1011                         do {
1012                                 lock_delay(&lda);
1013                                 x = SX_READ_VALUE(sx);
1014                                 owner = lv_sx_owner(x);
1015                         } while (owner != NULL && TD_IS_RUNNING(owner));
1016                         KTR_STATE0(KTR_SCHED, "thread",
1017                             sched_tdname(curthread), "running");
1018                         continue;
1019                 }
1020 sleepq:
1021 #endif
1022
1023                 /*
1024                  * Some other thread already has an exclusive lock, so
1025                  * start the process of blocking.
1026                  */
1027                 sleepq_lock(&sx->lock_object);
1028                 x = SX_READ_VALUE(sx);
1029 retry_sleepq:
1030                 /*
1031                  * The lock could have been released while we spun.
1032                  * In this case loop back and retry.
1033                  */
1034                 if (x & SX_LOCK_SHARED) {
1035                         sleepq_release(&sx->lock_object);
1036                         continue;
1037                 }
1038
1039 #ifdef ADAPTIVE_SX
1040                 /*
1041                  * If the owner is running on another CPU, spin until
1042                  * the owner stops running or the state of the lock
1043                  * changes.
1044                  */
1045                 if (!(x & SX_LOCK_SHARED) && adaptive) {
1046                         owner = (struct thread *)SX_OWNER(x);
1047                         if (TD_IS_RUNNING(owner)) {
1048                                 sleepq_release(&sx->lock_object);
1049                                 x = SX_READ_VALUE(sx);
1050                                 continue;
1051                         }
1052                 }
1053 #endif
1054
1055                 /*
1056                  * Try to set the SX_LOCK_SHARED_WAITERS flag.  If we
1057                  * fail to set it drop the sleep queue lock and loop
1058                  * back.
1059                  */
1060                 if (!(x & SX_LOCK_SHARED_WAITERS)) {
1061                         if (!atomic_fcmpset_ptr(&sx->sx_lock, &x,
1062                             x | SX_LOCK_SHARED_WAITERS))
1063                                 goto retry_sleepq;
1064                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
1065                                 CTR2(KTR_LOCK, "%s: %p set shared waiters flag",
1066                                     __func__, sx);
1067                 }
1068
1069                 /*
1070                  * Since we have been unable to acquire the shared lock,
1071                  * we have to sleep.
1072                  */
1073                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1074                         CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
1075                             __func__, sx);
1076
1077 #ifdef KDTRACE_HOOKS
1078                 sleep_time -= lockstat_nsecs(&sx->lock_object);
1079 #endif
1080                 sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
1081                     SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
1082                     SLEEPQ_INTERRUPTIBLE : 0), SQ_SHARED_QUEUE);
1083                 if (!(opts & SX_INTERRUPTIBLE))
1084                         sleepq_wait(&sx->lock_object, 0);
1085                 else
1086                         error = sleepq_wait_sig(&sx->lock_object, 0);
1087 #ifdef KDTRACE_HOOKS
1088                 sleep_time += lockstat_nsecs(&sx->lock_object);
1089                 sleep_cnt++;
1090 #endif
1091                 if (error) {
1092                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
1093                                 CTR2(KTR_LOCK,
1094                         "%s: interruptible sleep by %p suspended by signal",
1095                                     __func__, sx);
1096                         break;
1097                 }
1098                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1099                         CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
1100                             __func__, sx);
1101                 x = SX_READ_VALUE(sx);
1102         }
1103 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
1104         if (__predict_true(!extra_work))
1105                 return (error);
1106 #endif
1107 #ifdef KDTRACE_HOOKS
1108         all_time += lockstat_nsecs(&sx->lock_object);
1109         if (sleep_time)
1110                 LOCKSTAT_RECORD4(sx__block, sx, sleep_time,
1111                     LOCKSTAT_READER, (state & SX_LOCK_SHARED) == 0,
1112                     (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
1113         if (lda.spin_cnt > sleep_cnt)
1114                 LOCKSTAT_RECORD4(sx__spin, sx, all_time - sleep_time,
1115                     LOCKSTAT_READER, (state & SX_LOCK_SHARED) == 0,
1116                     (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
1117 out_lockstat:
1118 #endif
1119         if (error == 0) {
1120                 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
1121                     contested, waittime, file, line, LOCKSTAT_READER);
1122         }
1123         GIANT_RESTORE();
1124         return (error);
1125 }
1126
1127 int
1128 _sx_slock_int(struct sx *sx, int opts LOCK_FILE_LINE_ARG_DEF)
1129 {
1130         uintptr_t x;
1131         int error;
1132
1133         KASSERT(kdb_active != 0 || SCHEDULER_STOPPED() ||
1134             !TD_IS_IDLETHREAD(curthread),
1135             ("sx_slock() by idle thread %p on sx %s @ %s:%d",
1136             curthread, sx->lock_object.lo_name, file, line));
1137         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
1138             ("sx_slock() of destroyed sx @ %s:%d", file, line));
1139         WITNESS_CHECKORDER(&sx->lock_object, LOP_NEWORDER, file, line, NULL);
1140
1141         error = 0;
1142         x = SX_READ_VALUE(sx);
1143         if (__predict_false(LOCKSTAT_PROFILE_ENABLED(sx__acquire) ||
1144             !__sx_slock_try(sx, &x LOCK_FILE_LINE_ARG)))
1145                 error = _sx_slock_hard(sx, opts, x LOCK_FILE_LINE_ARG);
1146         else
1147                 lock_profile_obtain_lock_success(&sx->lock_object, 0, 0,
1148                     file, line);
1149         if (error == 0) {
1150                 LOCK_LOG_LOCK("SLOCK", &sx->lock_object, 0, 0, file, line);
1151                 WITNESS_LOCK(&sx->lock_object, 0, file, line);
1152                 TD_LOCKS_INC(curthread);
1153         }
1154         return (error);
1155 }
1156
1157 int
1158 _sx_slock(struct sx *sx, int opts, const char *file, int line)
1159 {
1160
1161         return (_sx_slock_int(sx, opts LOCK_FILE_LINE_ARG));
1162 }
1163
1164 static bool __always_inline
1165 _sx_sunlock_try(struct sx *sx, uintptr_t *xp)
1166 {
1167
1168         for (;;) {
1169                 /*
1170                  * We should never have sharers while at least one thread
1171                  * holds a shared lock.
1172                  */
1173                 KASSERT(!(*xp & SX_LOCK_SHARED_WAITERS),
1174                     ("%s: waiting sharers", __func__));
1175
1176                 /*
1177                  * See if there is more than one shared lock held.  If
1178                  * so, just drop one and return.
1179                  */
1180                 if (SX_SHARERS(*xp) > 1) {
1181                         if (atomic_fcmpset_rel_ptr(&sx->sx_lock, xp,
1182                             *xp - SX_ONE_SHARER)) {
1183                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1184                                         CTR4(KTR_LOCK,
1185                                             "%s: %p succeeded %p -> %p",
1186                                             __func__, sx, (void *)*xp,
1187                                             (void *)(*xp - SX_ONE_SHARER));
1188                                 return (true);
1189                         }
1190                         continue;
1191                 }
1192
1193                 /*
1194                  * If there aren't any waiters for an exclusive lock,
1195                  * then try to drop it quickly.
1196                  */
1197                 if (!(*xp & SX_LOCK_EXCLUSIVE_WAITERS)) {
1198                         MPASS(*xp == SX_SHARERS_LOCK(1));
1199                         *xp = SX_SHARERS_LOCK(1);
1200                         if (atomic_fcmpset_rel_ptr(&sx->sx_lock,
1201                             xp, SX_LOCK_UNLOCKED)) {
1202                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1203                                         CTR2(KTR_LOCK, "%s: %p last succeeded",
1204                                             __func__, sx);
1205                                 return (true);
1206                         }
1207                         continue;
1208                 }
1209                 break;
1210         }
1211         return (false);
1212 }
1213
1214 static void __noinline
1215 _sx_sunlock_hard(struct sx *sx, uintptr_t x LOCK_FILE_LINE_ARG_DEF)
1216 {
1217         int wakeup_swapper = 0;
1218         uintptr_t setx;
1219
1220         if (SCHEDULER_STOPPED())
1221                 return;
1222
1223         if (_sx_sunlock_try(sx, &x))
1224                 goto out_lockstat;
1225
1226         /*
1227          * At this point, there should just be one sharer with
1228          * exclusive waiters.
1229          */
1230         MPASS(x == (SX_SHARERS_LOCK(1) | SX_LOCK_EXCLUSIVE_WAITERS));
1231
1232         sleepq_lock(&sx->lock_object);
1233         x = SX_READ_VALUE(sx);
1234         for (;;) {
1235                 MPASS(x & SX_LOCK_EXCLUSIVE_WAITERS);
1236                 MPASS(!(x & SX_LOCK_SHARED_WAITERS));
1237                 if (_sx_sunlock_try(sx, &x))
1238                         break;
1239
1240                 /*
1241                  * Wake up semantic here is quite simple:
1242                  * Just wake up all the exclusive waiters.
1243                  * Note that the state of the lock could have changed,
1244                  * so if it fails loop back and retry.
1245                  */
1246                 setx = x - SX_ONE_SHARER;
1247                 setx &= ~SX_LOCK_EXCLUSIVE_WAITERS;
1248                 if (!atomic_fcmpset_rel_ptr(&sx->sx_lock, &x, setx))
1249                         continue;
1250                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1251                         CTR2(KTR_LOCK, "%s: %p waking up all thread on"
1252                             "exclusive queue", __func__, sx);
1253                 wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX,
1254                     0, SQ_EXCLUSIVE_QUEUE);
1255                 break;
1256         }
1257         sleepq_release(&sx->lock_object);
1258         if (wakeup_swapper)
1259                 kick_proc0();
1260 out_lockstat:
1261         LOCKSTAT_PROFILE_RELEASE_RWLOCK(sx__release, sx, LOCKSTAT_READER);
1262 }
1263
1264 void
1265 _sx_sunlock_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
1266 {
1267         uintptr_t x;
1268
1269         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
1270             ("sx_sunlock() of destroyed sx @ %s:%d", file, line));
1271         _sx_assert(sx, SA_SLOCKED, file, line);
1272         WITNESS_UNLOCK(&sx->lock_object, 0, file, line);
1273         LOCK_LOG_LOCK("SUNLOCK", &sx->lock_object, 0, 0, file, line);
1274
1275         x = SX_READ_VALUE(sx);
1276         if (__predict_false(LOCKSTAT_PROFILE_ENABLED(sx__release) ||
1277             !_sx_sunlock_try(sx, &x)))
1278                 _sx_sunlock_hard(sx, x LOCK_FILE_LINE_ARG);
1279         else
1280                 lock_profile_release_lock(&sx->lock_object);
1281
1282         TD_LOCKS_DEC(curthread);
1283 }
1284
1285 void
1286 _sx_sunlock(struct sx *sx, const char *file, int line)
1287 {
1288
1289         _sx_sunlock_int(sx LOCK_FILE_LINE_ARG);
1290 }
1291
1292 #ifdef INVARIANT_SUPPORT
1293 #ifndef INVARIANTS
1294 #undef  _sx_assert
1295 #endif
1296
1297 /*
1298  * In the non-WITNESS case, sx_assert() can only detect that at least
1299  * *some* thread owns an slock, but it cannot guarantee that *this*
1300  * thread owns an slock.
1301  */
1302 void
1303 _sx_assert(const struct sx *sx, int what, const char *file, int line)
1304 {
1305 #ifndef WITNESS
1306         int slocked = 0;
1307 #endif
1308
1309         if (panicstr != NULL)
1310                 return;
1311         switch (what) {
1312         case SA_SLOCKED:
1313         case SA_SLOCKED | SA_NOTRECURSED:
1314         case SA_SLOCKED | SA_RECURSED:
1315 #ifndef WITNESS
1316                 slocked = 1;
1317                 /* FALLTHROUGH */
1318 #endif
1319         case SA_LOCKED:
1320         case SA_LOCKED | SA_NOTRECURSED:
1321         case SA_LOCKED | SA_RECURSED:
1322 #ifdef WITNESS
1323                 witness_assert(&sx->lock_object, what, file, line);
1324 #else
1325                 /*
1326                  * If some other thread has an exclusive lock or we
1327                  * have one and are asserting a shared lock, fail.
1328                  * Also, if no one has a lock at all, fail.
1329                  */
1330                 if (sx->sx_lock == SX_LOCK_UNLOCKED ||
1331                     (!(sx->sx_lock & SX_LOCK_SHARED) && (slocked ||
1332                     sx_xholder(sx) != curthread)))
1333                         panic("Lock %s not %slocked @ %s:%d\n",
1334                             sx->lock_object.lo_name, slocked ? "share " : "",
1335                             file, line);
1336
1337                 if (!(sx->sx_lock & SX_LOCK_SHARED)) {
1338                         if (sx_recursed(sx)) {
1339                                 if (what & SA_NOTRECURSED)
1340                                         panic("Lock %s recursed @ %s:%d\n",
1341                                             sx->lock_object.lo_name, file,
1342                                             line);
1343                         } else if (what & SA_RECURSED)
1344                                 panic("Lock %s not recursed @ %s:%d\n",
1345                                     sx->lock_object.lo_name, file, line);
1346                 }
1347 #endif
1348                 break;
1349         case SA_XLOCKED:
1350         case SA_XLOCKED | SA_NOTRECURSED:
1351         case SA_XLOCKED | SA_RECURSED:
1352                 if (sx_xholder(sx) != curthread)
1353                         panic("Lock %s not exclusively locked @ %s:%d\n",
1354                             sx->lock_object.lo_name, file, line);
1355                 if (sx_recursed(sx)) {
1356                         if (what & SA_NOTRECURSED)
1357                                 panic("Lock %s recursed @ %s:%d\n",
1358                                     sx->lock_object.lo_name, file, line);
1359                 } else if (what & SA_RECURSED)
1360                         panic("Lock %s not recursed @ %s:%d\n",
1361                             sx->lock_object.lo_name, file, line);
1362                 break;
1363         case SA_UNLOCKED:
1364 #ifdef WITNESS
1365                 witness_assert(&sx->lock_object, what, file, line);
1366 #else
1367                 /*
1368                  * If we hold an exclusve lock fail.  We can't
1369                  * reliably check to see if we hold a shared lock or
1370                  * not.
1371                  */
1372                 if (sx_xholder(sx) == curthread)
1373                         panic("Lock %s exclusively locked @ %s:%d\n",
1374                             sx->lock_object.lo_name, file, line);
1375 #endif
1376                 break;
1377         default:
1378                 panic("Unknown sx lock assertion: %d @ %s:%d", what, file,
1379                     line);
1380         }
1381 }
1382 #endif  /* INVARIANT_SUPPORT */
1383
1384 #ifdef DDB
1385 static void
1386 db_show_sx(const struct lock_object *lock)
1387 {
1388         struct thread *td;
1389         const struct sx *sx;
1390
1391         sx = (const struct sx *)lock;
1392
1393         db_printf(" state: ");
1394         if (sx->sx_lock == SX_LOCK_UNLOCKED)
1395                 db_printf("UNLOCKED\n");
1396         else if (sx->sx_lock == SX_LOCK_DESTROYED) {
1397                 db_printf("DESTROYED\n");
1398                 return;
1399         } else if (sx->sx_lock & SX_LOCK_SHARED)
1400                 db_printf("SLOCK: %ju\n", (uintmax_t)SX_SHARERS(sx->sx_lock));
1401         else {
1402                 td = sx_xholder(sx);
1403                 db_printf("XLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1404                     td->td_tid, td->td_proc->p_pid, td->td_name);
1405                 if (sx_recursed(sx))
1406                         db_printf(" recursed: %d\n", sx->sx_recurse);
1407         }
1408
1409         db_printf(" waiters: ");
1410         switch(sx->sx_lock &
1411             (SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS)) {
1412         case SX_LOCK_SHARED_WAITERS:
1413                 db_printf("shared\n");
1414                 break;
1415         case SX_LOCK_EXCLUSIVE_WAITERS:
1416                 db_printf("exclusive\n");
1417                 break;
1418         case SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS:
1419                 db_printf("exclusive and shared\n");
1420                 break;
1421         default:
1422                 db_printf("none\n");
1423         }
1424 }
1425
1426 /*
1427  * Check to see if a thread that is blocked on a sleep queue is actually
1428  * blocked on an sx lock.  If so, output some details and return true.
1429  * If the lock has an exclusive owner, return that in *ownerp.
1430  */
1431 int
1432 sx_chain(struct thread *td, struct thread **ownerp)
1433 {
1434         struct sx *sx;
1435
1436         /*
1437          * Check to see if this thread is blocked on an sx lock.
1438          * First, we check the lock class.  If that is ok, then we
1439          * compare the lock name against the wait message.
1440          */
1441         sx = td->td_wchan;
1442         if (LOCK_CLASS(&sx->lock_object) != &lock_class_sx ||
1443             sx->lock_object.lo_name != td->td_wmesg)
1444                 return (0);
1445
1446         /* We think we have an sx lock, so output some details. */
1447         db_printf("blocked on sx \"%s\" ", td->td_wmesg);
1448         *ownerp = sx_xholder(sx);
1449         if (sx->sx_lock & SX_LOCK_SHARED)
1450                 db_printf("SLOCK (count %ju)\n",
1451                     (uintmax_t)SX_SHARERS(sx->sx_lock));
1452         else
1453                 db_printf("XLOCK\n");
1454         return (1);
1455 }
1456 #endif