sys/kern/kern_sx.c

   1 /*-
   2  * SPDX-License-Identifier: BSD-2-Clause
   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 #include <sys/param.h>
  47 #include <sys/systm.h>
  48 #include <sys/kdb.h>
  49 #include <sys/kernel.h>
  50 #include <sys/ktr.h>
  51 #include <sys/lock.h>
  52 #include <sys/mutex.h>
  53 #include <sys/proc.h>
  54 #include <sys/sched.h>
  55 #include <sys/sleepqueue.h>
  56 #include <sys/sx.h>
  57 #include <sys/smp.h>
  58 #include <sys/sysctl.h>
  59
  60 #if defined(SMP) && !defined(NO_ADAPTIVE_SX)
  61 #include <machine/cpu.h>
  62 #endif
  63
  64 #ifdef DDB
  65 #include <ddb/ddb.h>
  66 #endif
  67
  68 #if defined(SMP) && !defined(NO_ADAPTIVE_SX)
  69 #define ADAPTIVE_SX
  70 #endif
  71
  72 #ifdef HWPMC_HOOKS
  73 #include <sys/pmckern.h>
  74 PMC_SOFT_DECLARE( , , lock, failed);
  75 #endif
  76
  77 /* Handy macros for sleep queues. */
  78 #define SQ_EXCLUSIVE_QUEUE      0
  79 #define SQ_SHARED_QUEUE         1
  80
  81 /*
  82  * Variations on DROP_GIANT()/PICKUP_GIANT() for use in this file.  We
  83  * drop Giant anytime we have to sleep or if we adaptively spin.
  84  */
  85 #define GIANT_DECLARE                                                   \
  86         int _giantcnt = 0;                                              \
  87         WITNESS_SAVE_DECL(Giant)                                        \
  88
  89 #define GIANT_SAVE(work) do {                                           \
  90         if (__predict_false(mtx_owned(&Giant))) {                       \
  91                 work++;                                                 \
  92                 WITNESS_SAVE(&Giant.lock_object, Giant);                \
  93                 while (mtx_owned(&Giant)) {                             \
  94                         _giantcnt++;                                    \
  95                         mtx_unlock(&Giant);                             \
  96                 }                                                       \
  97         }                                                               \
  98 } while (0)
  99
 100 #define GIANT_RESTORE() do {                                            \
 101         if (_giantcnt > 0) {                                            \
 102                 mtx_assert(&Giant, MA_NOTOWNED);                        \
 103                 while (_giantcnt--)                                     \
 104                         mtx_lock(&Giant);                               \
 105                 WITNESS_RESTORE(&Giant.lock_object, Giant);             \
 106         }                                                               \
 107 } while (0)
 108
 109 /*
 110  * Returns true if an exclusive lock is recursed.  It assumes
 111  * curthread currently has an exclusive lock.
 112  */
 113 #define sx_recursed(sx)         ((sx)->sx_recurse != 0)
 114
 115 static void     assert_sx(const struct lock_object *lock, int what);
 116 #ifdef DDB
 117 static void     db_show_sx(const struct lock_object *lock);
 118 #endif
 119 static void     lock_sx(struct lock_object *lock, uintptr_t how);
 120 #ifdef KDTRACE_HOOKS
 121 static int      owner_sx(const struct lock_object *lock, struct thread **owner);
 122 #endif
 123 static uintptr_t unlock_sx(struct lock_object *lock);
 124
 125 struct lock_class lock_class_sx = {
 126         .lc_name = "sx",
 127         .lc_flags = LC_SLEEPLOCK | LC_SLEEPABLE | LC_RECURSABLE | LC_UPGRADABLE,
 128         .lc_assert = assert_sx,
 129 #ifdef DDB
 130         .lc_ddb_show = db_show_sx,
 131 #endif
 132         .lc_lock = lock_sx,
 133         .lc_unlock = unlock_sx,
 134 #ifdef KDTRACE_HOOKS
 135         .lc_owner = owner_sx,
 136 #endif
 137 };
 138
 139 #ifndef INVARIANTS
 140 #define _sx_assert(sx, what, file, line)
 141 #endif
 142
 143 #ifdef ADAPTIVE_SX
 144 #ifdef SX_CUSTOM_BACKOFF
 145 static u_short __read_frequently asx_retries;
 146 static u_short __read_frequently asx_loops;
 147 static SYSCTL_NODE(_debug, OID_AUTO, sx, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL,
 148     "sxlock debugging");
 149 SYSCTL_U16(_debug_sx, OID_AUTO, retries, CTLFLAG_RW, &asx_retries, 0, "");
 150 SYSCTL_U16(_debug_sx, OID_AUTO, loops, CTLFLAG_RW, &asx_loops, 0, "");
 151
 152 static struct lock_delay_config __read_frequently sx_delay;
 153
 154 SYSCTL_U16(_debug_sx, OID_AUTO, delay_base, CTLFLAG_RW, &sx_delay.base,
 155     0, "");
 156 SYSCTL_U16(_debug_sx, OID_AUTO, delay_max, CTLFLAG_RW, &sx_delay.max,
 157     0, "");
 158
 159 static void
 160 sx_lock_delay_init(void *arg __unused)
 161 {
 162
 163         lock_delay_default_init(&sx_delay);
 164         asx_retries = 10;
 165         asx_loops = max(10000, sx_delay.max);
 166 }
 167 LOCK_DELAY_SYSINIT(sx_lock_delay_init);
 168 #else
 169 #define sx_delay        locks_delay
 170 #define asx_retries     locks_delay_retries
 171 #define asx_loops       locks_delay_loops
 172 #endif
 173 #endif
 174
 175 void
 176 assert_sx(const struct lock_object *lock, int what)
 177 {
 178
 179         sx_assert((const struct sx *)lock, what);
 180 }
 181
 182 void
 183 lock_sx(struct lock_object *lock, uintptr_t how)
 184 {
 185         struct sx *sx;
 186
 187         sx = (struct sx *)lock;
 188         if (how)
 189                 sx_slock(sx);
 190         else
 191                 sx_xlock(sx);
 192 }
 193
 194 uintptr_t
 195 unlock_sx(struct lock_object *lock)
 196 {
 197         struct sx *sx;
 198
 199         sx = (struct sx *)lock;
 200         sx_assert(sx, SA_LOCKED | SA_NOTRECURSED);
 201         if (sx_xlocked(sx)) {
 202                 sx_xunlock(sx);
 203                 return (0);
 204         } else {
 205                 sx_sunlock(sx);
 206                 return (1);
 207         }
 208 }
 209
 210 #ifdef KDTRACE_HOOKS
 211 int
 212 owner_sx(const struct lock_object *lock, struct thread **owner)
 213 {
 214         const struct sx *sx;
 215         uintptr_t x;
 216
 217         sx = (const struct sx *)lock;
 218         x = sx->sx_lock;
 219         *owner = NULL;
 220         return ((x & SX_LOCK_SHARED) != 0 ? (SX_SHARERS(x) != 0) :
 221             ((*owner = (struct thread *)SX_OWNER(x)) != NULL));
 222 }
 223 #endif
 224
 225 void
 226 sx_sysinit(void *arg)
 227 {
 228         struct sx_args *sargs = arg;
 229
 230         sx_init_flags(sargs->sa_sx, sargs->sa_desc, sargs->sa_flags);
 231 }
 232
 233 void
 234 sx_init_flags(struct sx *sx, const char *description, int opts)
 235 {
 236         int flags;
 237
 238         MPASS((opts & ~(SX_QUIET | SX_RECURSE | SX_NOWITNESS | SX_DUPOK |
 239             SX_NOPROFILE | SX_NEW)) == 0);
 240         ASSERT_ATOMIC_LOAD_PTR(sx->sx_lock,
 241             ("%s: sx_lock not aligned for %s: %p", __func__, description,
 242             &sx->sx_lock));
 243
 244         flags = LO_SLEEPABLE | LO_UPGRADABLE;
 245         if (opts & SX_DUPOK)
 246                 flags |= LO_DUPOK;
 247         if (opts & SX_NOPROFILE)
 248                 flags |= LO_NOPROFILE;
 249         if (!(opts & SX_NOWITNESS))
 250                 flags |= LO_WITNESS;
 251         if (opts & SX_RECURSE)
 252                 flags |= LO_RECURSABLE;
 253         if (opts & SX_QUIET)
 254                 flags |= LO_QUIET;
 255         if (opts & SX_NEW)
 256                 flags |= LO_NEW;
 257
 258         lock_init(&sx->lock_object, &lock_class_sx, description, NULL, flags);
 259         sx->sx_lock = SX_LOCK_UNLOCKED;
 260         sx->sx_recurse = 0;
 261 }
 262
 263 void
 264 sx_destroy(struct sx *sx)
 265 {
 266
 267         KASSERT(sx->sx_lock == SX_LOCK_UNLOCKED, ("sx lock still held"));
 268         KASSERT(sx->sx_recurse == 0, ("sx lock still recursed"));
 269         sx->sx_lock = SX_LOCK_DESTROYED;
 270         lock_destroy(&sx->lock_object);
 271 }
 272
 273 int
 274 sx_try_slock_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
 275 {
 276         uintptr_t x;
 277
 278         if (SCHEDULER_STOPPED())
 279                 return (1);
 280
 281         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
 282             ("sx_try_slock() by idle thread %p on sx %s @ %s:%d",
 283             curthread, sx->lock_object.lo_name, file, line));
 284
 285         x = sx->sx_lock;
 286         for (;;) {
 287                 KASSERT(x != SX_LOCK_DESTROYED,
 288                     ("sx_try_slock() of destroyed sx @ %s:%d", file, line));
 289                 if (!(x & SX_LOCK_SHARED))
 290                         break;
 291                 if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, x + SX_ONE_SHARER)) {
 292                         LOCK_LOG_TRY("SLOCK", &sx->lock_object, 0, 1, file, line);
 293                         WITNESS_LOCK(&sx->lock_object, LOP_TRYLOCK, file, line);
 294                         LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire,
 295                             sx, 0, 0, file, line, LOCKSTAT_READER);
 296                         TD_LOCKS_INC(curthread);
 297                         curthread->td_sx_slocks++;
 298                         return (1);
 299                 }
 300         }
 301
 302         LOCK_LOG_TRY("SLOCK", &sx->lock_object, 0, 0, file, line);
 303         return (0);
 304 }
 305
 306 int
 307 sx_try_slock_(struct sx *sx, const char *file, int line)
 308 {
 309
 310         return (sx_try_slock_int(sx LOCK_FILE_LINE_ARG));
 311 }
 312
 313 int
 314 _sx_xlock(struct sx *sx, int opts, const char *file, int line)
 315 {
 316         uintptr_t tid, x;
 317         int error = 0;
 318
 319         KASSERT(kdb_active != 0 || SCHEDULER_STOPPED() ||
 320             !TD_IS_IDLETHREAD(curthread),
 321             ("sx_xlock() by idle thread %p on sx %s @ %s:%d",
 322             curthread, sx->lock_object.lo_name, file, line));
 323         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 324             ("sx_xlock() of destroyed sx @ %s:%d", file, line));
 325         WITNESS_CHECKORDER(&sx->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE, file,
 326             line, NULL);
 327         tid = (uintptr_t)curthread;
 328         x = SX_LOCK_UNLOCKED;
 329         if (!atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid))
 330                 error = _sx_xlock_hard(sx, x, opts LOCK_FILE_LINE_ARG);
 331         else
 332                 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
 333                     0, 0, file, line, LOCKSTAT_WRITER);
 334         if (!error) {
 335                 LOCK_LOG_LOCK("XLOCK", &sx->lock_object, 0, sx->sx_recurse,
 336                     file, line);
 337                 WITNESS_LOCK(&sx->lock_object, LOP_EXCLUSIVE, file, line);
 338                 TD_LOCKS_INC(curthread);
 339         }
 340
 341         return (error);
 342 }
 343
 344 int
 345 sx_try_xlock_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
 346 {
 347         struct thread *td;
 348         uintptr_t tid, x;
 349         int rval;
 350         bool recursed;
 351
 352         td = curthread;
 353         tid = (uintptr_t)td;
 354         if (SCHEDULER_STOPPED_TD(td))
 355                 return (1);
 356
 357         KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(td),
 358             ("sx_try_xlock() by idle thread %p on sx %s @ %s:%d",
 359             curthread, sx->lock_object.lo_name, file, line));
 360         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 361             ("sx_try_xlock() of destroyed sx @ %s:%d", file, line));
 362
 363         rval = 1;
 364         recursed = false;
 365         x = SX_LOCK_UNLOCKED;
 366         for (;;) {
 367                 if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid))
 368                         break;
 369                 if (x == SX_LOCK_UNLOCKED)
 370                         continue;
 371                 if (x == tid && (sx->lock_object.lo_flags & LO_RECURSABLE)) {
 372                         sx->sx_recurse++;
 373                         atomic_set_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
 374                         break;
 375                 }
 376                 rval = 0;
 377                 break;
 378         }
 379
 380         LOCK_LOG_TRY("XLOCK", &sx->lock_object, 0, rval, file, line);
 381         if (rval) {
 382                 WITNESS_LOCK(&sx->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
 383                     file, line);
 384                 if (!recursed)
 385                         LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire,
 386                             sx, 0, 0, file, line, LOCKSTAT_WRITER);
 387                 TD_LOCKS_INC(curthread);
 388         }
 389
 390         return (rval);
 391 }
 392
 393 int
 394 sx_try_xlock_(struct sx *sx, const char *file, int line)
 395 {
 396
 397         return (sx_try_xlock_int(sx LOCK_FILE_LINE_ARG));
 398 }
 399
 400 void
 401 _sx_xunlock(struct sx *sx, const char *file, int line)
 402 {
 403
 404         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 405             ("sx_xunlock() of destroyed sx @ %s:%d", file, line));
 406         _sx_assert(sx, SA_XLOCKED, file, line);
 407         WITNESS_UNLOCK(&sx->lock_object, LOP_EXCLUSIVE, file, line);
 408         LOCK_LOG_LOCK("XUNLOCK", &sx->lock_object, 0, sx->sx_recurse, file,
 409             line);
 410 #if LOCK_DEBUG > 0
 411         _sx_xunlock_hard(sx, (uintptr_t)curthread, file, line);
 412 #else
 413         __sx_xunlock(sx, curthread, file, line);
 414 #endif
 415         TD_LOCKS_DEC(curthread);
 416 }
 417
 418 /*
 419  * Try to do a non-blocking upgrade from a shared lock to an exclusive lock.
 420  * This will only succeed if this thread holds a single shared lock.
 421  * Return 1 if if the upgrade succeed, 0 otherwise.
 422  */
 423 int
 424 sx_try_upgrade_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
 425 {
 426         uintptr_t x;
 427         uintptr_t waiters;
 428         int success;
 429
 430         if (SCHEDULER_STOPPED())
 431                 return (1);
 432
 433         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 434             ("sx_try_upgrade() of destroyed sx @ %s:%d", file, line));
 435         _sx_assert(sx, SA_SLOCKED, file, line);
 436
 437         /*
 438          * Try to switch from one shared lock to an exclusive lock.  We need
 439          * to maintain the SX_LOCK_EXCLUSIVE_WAITERS flag if set so that
 440          * we will wake up the exclusive waiters when we drop the lock.
 441          */
 442         success = 0;
 443         x = SX_READ_VALUE(sx);
 444         for (;;) {
 445                 if (SX_SHARERS(x) > 1)
 446                         break;
 447                 waiters = (x & SX_LOCK_WAITERS);
 448                 if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x,
 449                     (uintptr_t)curthread | waiters)) {
 450                         success = 1;
 451                         break;
 452                 }
 453         }
 454         LOCK_LOG_TRY("XUPGRADE", &sx->lock_object, 0, success, file, line);
 455         if (success) {
 456                 curthread->td_sx_slocks--;
 457                 WITNESS_UPGRADE(&sx->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
 458                     file, line);
 459                 LOCKSTAT_RECORD0(sx__upgrade, sx);
 460         }
 461         return (success);
 462 }
 463
 464 int
 465 sx_try_upgrade_(struct sx *sx, const char *file, int line)
 466 {
 467
 468         return (sx_try_upgrade_int(sx LOCK_FILE_LINE_ARG));
 469 }
 470
 471 /*
 472  * Downgrade an unrecursed exclusive lock into a single shared lock.
 473  */
 474 void
 475 sx_downgrade_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
 476 {
 477         uintptr_t x;
 478         int wakeup_swapper;
 479
 480         if (SCHEDULER_STOPPED())
 481                 return;
 482
 483         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 484             ("sx_downgrade() of destroyed sx @ %s:%d", file, line));
 485         _sx_assert(sx, SA_XLOCKED | SA_NOTRECURSED, file, line);
 486 #ifndef INVARIANTS
 487         if (sx_recursed(sx))
 488                 panic("downgrade of a recursed lock");
 489 #endif
 490
 491         WITNESS_DOWNGRADE(&sx->lock_object, 0, file, line);
 492
 493         /*
 494          * Try to switch from an exclusive lock with no shared waiters
 495          * to one sharer with no shared waiters.  If there are
 496          * exclusive waiters, we don't need to lock the sleep queue so
 497          * long as we preserve the flag.  We do one quick try and if
 498          * that fails we grab the sleepq lock to keep the flags from
 499          * changing and do it the slow way.
 500          *
 501          * We have to lock the sleep queue if there are shared waiters
 502          * so we can wake them up.
 503          */
 504         x = sx->sx_lock;
 505         if (!(x & SX_LOCK_SHARED_WAITERS) &&
 506             atomic_cmpset_rel_ptr(&sx->sx_lock, x, SX_SHARERS_LOCK(1) |
 507             (x & SX_LOCK_EXCLUSIVE_WAITERS)))
 508                 goto out;
 509
 510         /*
 511          * Lock the sleep queue so we can read the waiters bits
 512          * without any races and wakeup any shared waiters.
 513          */
 514         sleepq_lock(&sx->lock_object);
 515
 516         /*
 517          * Preserve SX_LOCK_EXCLUSIVE_WAITERS while downgraded to a single
 518          * shared lock.  If there are any shared waiters, wake them up.
 519          */
 520         wakeup_swapper = 0;
 521         x = sx->sx_lock;
 522         atomic_store_rel_ptr(&sx->sx_lock, SX_SHARERS_LOCK(1) |
 523             (x & SX_LOCK_EXCLUSIVE_WAITERS));
 524         if (x & SX_LOCK_SHARED_WAITERS)
 525                 wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX,
 526                     0, SQ_SHARED_QUEUE);
 527         sleepq_release(&sx->lock_object);
 528
 529         if (wakeup_swapper)
 530                 kick_proc0();
 531
 532 out:
 533         curthread->td_sx_slocks++;
 534         LOCK_LOG_LOCK("XDOWNGRADE", &sx->lock_object, 0, 0, file, line);
 535         LOCKSTAT_RECORD0(sx__downgrade, sx);
 536 }
 537
 538 void
 539 sx_downgrade_(struct sx *sx, const char *file, int line)
 540 {
 541
 542         sx_downgrade_int(sx LOCK_FILE_LINE_ARG);
 543 }
 544
 545 #ifdef  ADAPTIVE_SX
 546 static inline void
 547 sx_drop_critical(uintptr_t x, bool *in_critical, int *extra_work)
 548 {
 549
 550         if (x & SX_LOCK_WRITE_SPINNER)
 551                 return;
 552         if (*in_critical) {
 553                 critical_exit();
 554                 *in_critical = false;
 555                 (*extra_work)--;
 556         }
 557 }
 558 #else
 559 #define sx_drop_critical(x, in_critical, extra_work) do { } while (0)
 560 #endif
 561
 562 /*
 563  * This function represents the so-called 'hard case' for sx_xlock
 564  * operation.  All 'easy case' failures are redirected to this.  Note
 565  * that ideally this would be a static function, but it needs to be
 566  * accessible from at least sx.h.
 567  */
 568 int
 569 _sx_xlock_hard(struct sx *sx, uintptr_t x, int opts LOCK_FILE_LINE_ARG_DEF)
 570 {
 571         GIANT_DECLARE;
 572         uintptr_t tid, setx;
 573 #ifdef ADAPTIVE_SX
 574         struct thread *owner;
 575         u_int i, n, spintries = 0;
 576         enum { READERS, WRITER } sleep_reason = READERS;
 577         bool in_critical = false;
 578 #endif
 579 #ifdef LOCK_PROFILING
 580         uint64_t waittime = 0;
 581         int contested = 0;
 582 #endif
 583         int error = 0;
 584 #if defined(ADAPTIVE_SX) || defined(KDTRACE_HOOKS)
 585         struct lock_delay_arg lda;
 586 #endif
 587 #ifdef  KDTRACE_HOOKS
 588         u_int sleep_cnt = 0;
 589         int64_t sleep_time = 0;
 590         int64_t all_time = 0;
 591 #endif
 592 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
 593         uintptr_t state = 0;
 594         int doing_lockprof = 0;
 595 #endif
 596         int extra_work = 0;
 597
 598         tid = (uintptr_t)curthread;
 599
 600 #ifdef KDTRACE_HOOKS
 601         if (LOCKSTAT_PROFILE_ENABLED(sx__acquire)) {
 602                 while (x == SX_LOCK_UNLOCKED) {
 603                         if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid))
 604                                 goto out_lockstat;
 605                 }
 606                 extra_work = 1;
 607                 doing_lockprof = 1;
 608                 all_time -= lockstat_nsecs(&sx->lock_object);
 609                 state = x;
 610         }
 611 #endif
 612 #ifdef LOCK_PROFILING
 613         extra_work = 1;
 614         doing_lockprof = 1;
 615         state = x;
 616 #endif
 617
 618         if (SCHEDULER_STOPPED())
 619                 return (0);
 620
 621         if (__predict_false(x == SX_LOCK_UNLOCKED))
 622                 x = SX_READ_VALUE(sx);
 623
 624         /* If we already hold an exclusive lock, then recurse. */
 625         if (__predict_false(lv_sx_owner(x) == (struct thread *)tid)) {
 626                 KASSERT((sx->lock_object.lo_flags & LO_RECURSABLE) != 0,
 627             ("_sx_xlock_hard: recursed on non-recursive sx %s @ %s:%d\n",
 628                     sx->lock_object.lo_name, file, line));
 629                 sx->sx_recurse++;
 630                 atomic_set_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
 631                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 632                         CTR2(KTR_LOCK, "%s: %p recursing", __func__, sx);
 633                 return (0);
 634         }
 635
 636         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 637                 CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__,
 638                     sx->lock_object.lo_name, (void *)sx->sx_lock, file, line);
 639
 640 #if defined(ADAPTIVE_SX)
 641         lock_delay_arg_init(&lda, &sx_delay);
 642 #elif defined(KDTRACE_HOOKS)
 643         lock_delay_arg_init_noadapt(&lda);
 644 #endif
 645
 646 #ifdef HWPMC_HOOKS
 647         PMC_SOFT_CALL( , , lock, failed);
 648 #endif
 649         lock_profile_obtain_lock_failed(&sx->lock_object, false, &contested,
 650             &waittime);
 651
 652 #ifndef INVARIANTS
 653         GIANT_SAVE(extra_work);
 654 #endif
 655
 656         THREAD_CONTENDS_ON_LOCK(&sx->lock_object);
 657
 658         for (;;) {
 659                 if (x == SX_LOCK_UNLOCKED) {
 660                         if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid))
 661                                 break;
 662                         continue;
 663                 }
 664 #ifdef INVARIANTS
 665                 GIANT_SAVE(extra_work);
 666 #endif
 667 #ifdef KDTRACE_HOOKS
 668                 lda.spin_cnt++;
 669 #endif
 670 #ifdef ADAPTIVE_SX
 671                 if (x == (SX_LOCK_SHARED | SX_LOCK_WRITE_SPINNER)) {
 672                         if (atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid))
 673                                 break;
 674                         continue;
 675                 }
 676
 677                 /*
 678                  * If the lock is write locked and the owner is
 679                  * running on another CPU, spin until the owner stops
 680                  * running or the state of the lock changes.
 681                  */
 682                 if ((x & SX_LOCK_SHARED) == 0) {
 683                         sx_drop_critical(x, &in_critical, &extra_work);
 684                         sleep_reason = WRITER;
 685                         owner = lv_sx_owner(x);
 686                         if (!TD_IS_RUNNING(owner))
 687                                 goto sleepq;
 688                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 689                                 CTR3(KTR_LOCK, "%s: spinning on %p held by %p",
 690                                     __func__, sx, owner);
 691                         KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
 692                             "spinning", "lockname:\"%s\"",
 693                             sx->lock_object.lo_name);
 694                         do {
 695                                 lock_delay(&lda);
 696                                 x = SX_READ_VALUE(sx);
 697                                 owner = lv_sx_owner(x);
 698                         } while (owner != NULL && TD_IS_RUNNING(owner));
 699                         KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
 700                             "running");
 701                         continue;
 702                 } else if (SX_SHARERS(x) > 0) {
 703                         sleep_reason = READERS;
 704                         if (spintries == asx_retries)
 705                                 goto sleepq;
 706                         if (!(x & SX_LOCK_WRITE_SPINNER)) {
 707                                 if (!in_critical) {
 708                                         critical_enter();
 709                                         in_critical = true;
 710                                         extra_work++;
 711                                 }
 712                                 if (!atomic_fcmpset_ptr(&sx->sx_lock, &x,
 713                                     x | SX_LOCK_WRITE_SPINNER)) {
 714                                         critical_exit();
 715                                         in_critical = false;
 716                                         extra_work--;
 717                                         continue;
 718                                 }
 719                         }
 720                         spintries++;
 721                         KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
 722                             "spinning", "lockname:\"%s\"",
 723                             sx->lock_object.lo_name);
 724                         n = SX_SHARERS(x);
 725                         for (i = 0; i < asx_loops; i += n) {
 726                                 lock_delay_spin(n);
 727                                 x = SX_READ_VALUE(sx);
 728                                 if (!(x & SX_LOCK_WRITE_SPINNER))
 729                                         break;
 730                                 if (!(x & SX_LOCK_SHARED))
 731                                         break;
 732                                 n = SX_SHARERS(x);
 733                                 if (n == 0)
 734                                         break;
 735                         }
 736 #ifdef KDTRACE_HOOKS
 737                         lda.spin_cnt += i;
 738 #endif
 739                         KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
 740                             "running");
 741                         if (i < asx_loops)
 742                                 continue;
 743                 }
 744 sleepq:
 745 #endif
 746                 sleepq_lock(&sx->lock_object);
 747                 x = SX_READ_VALUE(sx);
 748 retry_sleepq:
 749
 750                 /*
 751                  * If the lock was released while spinning on the
 752                  * sleep queue chain lock, try again.
 753                  */
 754                 if (x == SX_LOCK_UNLOCKED) {
 755                         sleepq_release(&sx->lock_object);
 756                         sx_drop_critical(x, &in_critical, &extra_work);
 757                         continue;
 758                 }
 759
 760 #ifdef ADAPTIVE_SX
 761                 /*
 762                  * The current lock owner might have started executing
 763                  * on another CPU (or the lock could have changed
 764                  * owners) while we were waiting on the sleep queue
 765                  * chain lock.  If so, drop the sleep queue lock and try
 766                  * again.
 767                  */
 768                 if (!(x & SX_LOCK_SHARED)) {
 769                         owner = (struct thread *)SX_OWNER(x);
 770                         if (TD_IS_RUNNING(owner)) {
 771                                 sleepq_release(&sx->lock_object);
 772                                 sx_drop_critical(x, &in_critical,
 773                                     &extra_work);
 774                                 continue;
 775                         }
 776                 } else if (SX_SHARERS(x) > 0 && sleep_reason == WRITER) {
 777                         sleepq_release(&sx->lock_object);
 778                         sx_drop_critical(x, &in_critical, &extra_work);
 779                         continue;
 780                 }
 781 #endif
 782
 783                 /*
 784                  * If an exclusive lock was released with both shared
 785                  * and exclusive waiters and a shared waiter hasn't
 786                  * woken up and acquired the lock yet, sx_lock will be
 787                  * set to SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS.
 788                  * If we see that value, try to acquire it once.  Note
 789                  * that we have to preserve SX_LOCK_EXCLUSIVE_WAITERS
 790                  * as there are other exclusive waiters still.  If we
 791                  * fail, restart the loop.
 792                  */
 793                 setx = x & (SX_LOCK_WAITERS | SX_LOCK_WRITE_SPINNER);
 794                 if ((x & ~setx) == SX_LOCK_SHARED) {
 795                         setx &= ~SX_LOCK_WRITE_SPINNER;
 796                         if (!atomic_fcmpset_acq_ptr(&sx->sx_lock, &x, tid | setx))
 797                                 goto retry_sleepq;
 798                         sleepq_release(&sx->lock_object);
 799                         CTR2(KTR_LOCK, "%s: %p claimed by new writer",
 800                             __func__, sx);
 801                         break;
 802                 }
 803
 804 #ifdef ADAPTIVE_SX
 805                 /*
 806                  * It is possible we set the SX_LOCK_WRITE_SPINNER bit.
 807                  * It is an invariant that when the bit is set, there is
 808                  * a writer ready to grab the lock. Thus clear the bit since
 809                  * we are going to sleep.
 810                  */
 811                 if (in_critical) {
 812                         if ((x & SX_LOCK_WRITE_SPINNER) ||
 813                             !((x & SX_LOCK_EXCLUSIVE_WAITERS))) {
 814                                 setx = x & ~SX_LOCK_WRITE_SPINNER;
 815                                 setx |= SX_LOCK_EXCLUSIVE_WAITERS;
 816                                 if (!atomic_fcmpset_ptr(&sx->sx_lock, &x,
 817                                     setx)) {
 818                                         goto retry_sleepq;
 819                                 }
 820                         }
 821                         critical_exit();
 822                         in_critical = false;
 823                 } else {
 824 #endif
 825                         /*
 826                          * Try to set the SX_LOCK_EXCLUSIVE_WAITERS.  If we fail,
 827                          * than loop back and retry.
 828                          */
 829                         if (!(x & SX_LOCK_EXCLUSIVE_WAITERS)) {
 830                                 if (!atomic_fcmpset_ptr(&sx->sx_lock, &x,
 831                                     x | SX_LOCK_EXCLUSIVE_WAITERS)) {
 832                                         goto retry_sleepq;
 833                                 }
 834                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 835                                         CTR2(KTR_LOCK, "%s: %p set excl waiters flag",
 836                                             __func__, sx);
 837                         }
 838 #ifdef ADAPTIVE_SX
 839                 }
 840 #endif
 841
 842                 /*
 843                  * Since we have been unable to acquire the exclusive
 844                  * lock and the exclusive waiters flag is set, we have
 845                  * to sleep.
 846                  */
 847                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 848                         CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
 849                             __func__, sx);
 850
 851 #ifdef KDTRACE_HOOKS
 852                 sleep_time -= lockstat_nsecs(&sx->lock_object);
 853 #endif
 854                 sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
 855                     SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
 856                     SLEEPQ_INTERRUPTIBLE : 0), SQ_EXCLUSIVE_QUEUE);
 857                 /*
 858                  * Hack: this can land in thread_suspend_check which will
 859                  * conditionally take a mutex, tripping over an assert if a
 860                  * lock we are waiting for is set.
 861                  */
 862                 THREAD_CONTENTION_DONE(&sx->lock_object);
 863                 if (!(opts & SX_INTERRUPTIBLE))
 864                         sleepq_wait(&sx->lock_object, 0);
 865                 else
 866                         error = sleepq_wait_sig(&sx->lock_object, 0);
 867                 THREAD_CONTENDS_ON_LOCK(&sx->lock_object);
 868 #ifdef KDTRACE_HOOKS
 869                 sleep_time += lockstat_nsecs(&sx->lock_object);
 870                 sleep_cnt++;
 871 #endif
 872                 if (error) {
 873                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 874                                 CTR2(KTR_LOCK,
 875                         "%s: interruptible sleep by %p suspended by signal",
 876                                     __func__, sx);
 877                         break;
 878                 }
 879                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 880                         CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
 881                             __func__, sx);
 882                 x = SX_READ_VALUE(sx);
 883         }
 884         THREAD_CONTENTION_DONE(&sx->lock_object);
 885         if (__predict_true(!extra_work))
 886                 return (error);
 887 #ifdef ADAPTIVE_SX
 888         if (in_critical)
 889                 critical_exit();
 890 #endif
 891         GIANT_RESTORE();
 892 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
 893         if (__predict_true(!doing_lockprof))
 894                 return (error);
 895 #endif
 896 #ifdef KDTRACE_HOOKS
 897         all_time += lockstat_nsecs(&sx->lock_object);
 898         if (sleep_time)
 899                 LOCKSTAT_RECORD4(sx__block, sx, sleep_time,
 900                     LOCKSTAT_WRITER, (state & SX_LOCK_SHARED) == 0,
 901                     (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
 902         if (lda.spin_cnt > sleep_cnt)
 903                 LOCKSTAT_RECORD4(sx__spin, sx, all_time - sleep_time,
 904                     LOCKSTAT_WRITER, (state & SX_LOCK_SHARED) == 0,
 905                     (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
 906 out_lockstat:
 907 #endif
 908         if (!error)
 909                 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
 910                     contested, waittime, file, line, LOCKSTAT_WRITER);
 911         return (error);
 912 }
 913
 914 /*
 915  * This function represents the so-called 'hard case' for sx_xunlock
 916  * operation.  All 'easy case' failures are redirected to this.  Note
 917  * that ideally this would be a static function, but it needs to be
 918  * accessible from at least sx.h.
 919  */
 920 void
 921 _sx_xunlock_hard(struct sx *sx, uintptr_t x LOCK_FILE_LINE_ARG_DEF)
 922 {
 923         uintptr_t tid, setx;
 924         int queue, wakeup_swapper;
 925
 926         if (SCHEDULER_STOPPED())
 927                 return;
 928
 929         tid = (uintptr_t)curthread;
 930
 931         if (__predict_false(x == tid))
 932                 x = SX_READ_VALUE(sx);
 933
 934         MPASS(!(x & SX_LOCK_SHARED));
 935
 936         if (__predict_false(x & SX_LOCK_RECURSED)) {
 937                 /* The lock is recursed, unrecurse one level. */
 938                 if ((--sx->sx_recurse) == 0)
 939                         atomic_clear_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
 940                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 941                         CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, sx);
 942                 return;
 943         }
 944
 945         LOCKSTAT_PROFILE_RELEASE_RWLOCK(sx__release, sx, LOCKSTAT_WRITER);
 946         if (x == tid &&
 947             atomic_cmpset_rel_ptr(&sx->sx_lock, tid, SX_LOCK_UNLOCKED))
 948                 return;
 949
 950         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 951                 CTR2(KTR_LOCK, "%s: %p contested", __func__, sx);
 952
 953         sleepq_lock(&sx->lock_object);
 954         x = SX_READ_VALUE(sx);
 955         MPASS(x & (SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS));
 956
 957         /*
 958          * The wake up algorithm here is quite simple and probably not
 959          * ideal.  It gives precedence to shared waiters if they are
 960          * present.  For this condition, we have to preserve the
 961          * state of the exclusive waiters flag.
 962          * If interruptible sleeps left the shared queue empty avoid a
 963          * starvation for the threads sleeping on the exclusive queue by giving
 964          * them precedence and cleaning up the shared waiters bit anyway.
 965          */
 966         setx = SX_LOCK_UNLOCKED;
 967         queue = SQ_SHARED_QUEUE;
 968         if ((x & SX_LOCK_EXCLUSIVE_WAITERS) != 0 &&
 969             sleepq_sleepcnt(&sx->lock_object, SQ_EXCLUSIVE_QUEUE) != 0) {
 970                 queue = SQ_EXCLUSIVE_QUEUE;
 971                 setx |= (x & SX_LOCK_SHARED_WAITERS);
 972         }
 973         atomic_store_rel_ptr(&sx->sx_lock, setx);
 974
 975         /* Wake up all the waiters for the specific queue. */
 976         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 977                 CTR3(KTR_LOCK, "%s: %p waking up all threads on %s queue",
 978                     __func__, sx, queue == SQ_SHARED_QUEUE ? "shared" :
 979                     "exclusive");
 980
 981         wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX, 0,
 982             queue);
 983         sleepq_release(&sx->lock_object);
 984         if (wakeup_swapper)
 985                 kick_proc0();
 986 }
 987
 988 static bool __always_inline
 989 __sx_can_read(struct thread *td, uintptr_t x, bool fp)
 990 {
 991
 992         if ((x & (SX_LOCK_SHARED | SX_LOCK_EXCLUSIVE_WAITERS | SX_LOCK_WRITE_SPINNER))
 993                         == SX_LOCK_SHARED)
 994                 return (true);
 995         if (!fp && td->td_sx_slocks && (x & SX_LOCK_SHARED))
 996                 return (true);
 997         return (false);
 998 }
 999
1000 static bool __always_inline
1001 __sx_slock_try(struct sx *sx, struct thread *td, uintptr_t *xp, bool fp
1002     LOCK_FILE_LINE_ARG_DEF)
1003 {
1004
1005         /*
1006          * If no other thread has an exclusive lock then try to bump up
1007          * the count of sharers.  Since we have to preserve the state
1008          * of SX_LOCK_EXCLUSIVE_WAITERS, if we fail to acquire the
1009          * shared lock loop back and retry.
1010          */
1011         while (__sx_can_read(td, *xp, fp)) {
1012                 if (atomic_fcmpset_acq_ptr(&sx->sx_lock, xp,
1013                     *xp + SX_ONE_SHARER)) {
1014                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
1015                                 CTR4(KTR_LOCK, "%s: %p succeed %p -> %p",
1016                                     __func__, sx, (void *)*xp,
1017                                     (void *)(*xp + SX_ONE_SHARER));
1018                         td->td_sx_slocks++;
1019                         return (true);
1020                 }
1021         }
1022         return (false);
1023 }
1024
1025 static int __noinline
1026 _sx_slock_hard(struct sx *sx, int opts, uintptr_t x LOCK_FILE_LINE_ARG_DEF)
1027 {
1028         GIANT_DECLARE;
1029         struct thread *td;
1030 #ifdef ADAPTIVE_SX
1031         struct thread *owner;
1032         u_int i, n, spintries = 0;
1033 #endif
1034 #ifdef LOCK_PROFILING
1035         uint64_t waittime = 0;
1036         int contested = 0;
1037 #endif
1038         int error = 0;
1039 #if defined(ADAPTIVE_SX) || defined(KDTRACE_HOOKS)
1040         struct lock_delay_arg lda;
1041 #endif
1042 #ifdef KDTRACE_HOOKS
1043         u_int sleep_cnt = 0;
1044         int64_t sleep_time = 0;
1045         int64_t all_time = 0;
1046 #endif
1047 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
1048         uintptr_t state = 0;
1049 #endif
1050         int extra_work __sdt_used = 0;
1051
1052         td = curthread;
1053
1054 #ifdef KDTRACE_HOOKS
1055         if (LOCKSTAT_PROFILE_ENABLED(sx__acquire)) {
1056                 if (__sx_slock_try(sx, td, &x, false LOCK_FILE_LINE_ARG))
1057                         goto out_lockstat;
1058                 extra_work = 1;
1059                 all_time -= lockstat_nsecs(&sx->lock_object);
1060                 state = x;
1061         }
1062 #endif
1063 #ifdef LOCK_PROFILING
1064         extra_work = 1;
1065         state = x;
1066 #endif
1067
1068         if (SCHEDULER_STOPPED())
1069                 return (0);
1070
1071 #if defined(ADAPTIVE_SX)
1072         lock_delay_arg_init(&lda, &sx_delay);
1073 #elif defined(KDTRACE_HOOKS)
1074         lock_delay_arg_init_noadapt(&lda);
1075 #endif
1076
1077 #ifdef HWPMC_HOOKS
1078         PMC_SOFT_CALL( , , lock, failed);
1079 #endif
1080         lock_profile_obtain_lock_failed(&sx->lock_object, false, &contested,
1081             &waittime);
1082
1083 #ifndef INVARIANTS
1084         GIANT_SAVE(extra_work);
1085 #endif
1086
1087         THREAD_CONTENDS_ON_LOCK(&sx->lock_object);
1088
1089         /*
1090          * As with rwlocks, we don't make any attempt to try to block
1091          * shared locks once there is an exclusive waiter.
1092          */
1093         for (;;) {
1094                 if (__sx_slock_try(sx, td, &x, false LOCK_FILE_LINE_ARG))
1095                         break;
1096 #ifdef INVARIANTS
1097                 GIANT_SAVE(extra_work);
1098 #endif
1099 #ifdef KDTRACE_HOOKS
1100                 lda.spin_cnt++;
1101 #endif
1102
1103 #ifdef ADAPTIVE_SX
1104                 /*
1105                  * If the owner is running on another CPU, spin until
1106                  * the owner stops running or the state of the lock
1107                  * changes.
1108                  */
1109                 if ((x & SX_LOCK_SHARED) == 0) {
1110                         owner = lv_sx_owner(x);
1111                         if (TD_IS_RUNNING(owner)) {
1112                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1113                                         CTR3(KTR_LOCK,
1114                                             "%s: spinning on %p held by %p",
1115                                             __func__, sx, owner);
1116                                 KTR_STATE1(KTR_SCHED, "thread",
1117                                     sched_tdname(curthread), "spinning",
1118                                     "lockname:\"%s\"", sx->lock_object.lo_name);
1119                                 do {
1120                                         lock_delay(&lda);
1121                                         x = SX_READ_VALUE(sx);
1122                                         owner = lv_sx_owner(x);
1123                                 } while (owner != NULL && TD_IS_RUNNING(owner));
1124                                 KTR_STATE0(KTR_SCHED, "thread",
1125                                     sched_tdname(curthread), "running");
1126                                 continue;
1127                         }
1128                 } else {
1129                         if ((x & SX_LOCK_WRITE_SPINNER) && SX_SHARERS(x) == 0) {
1130                                 MPASS(!__sx_can_read(td, x, false));
1131                                 lock_delay_spin(2);
1132                                 x = SX_READ_VALUE(sx);
1133                                 continue;
1134                         }
1135                         if (spintries < asx_retries) {
1136                                 KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
1137                                     "spinning", "lockname:\"%s\"",
1138                                     sx->lock_object.lo_name);
1139                                 n = SX_SHARERS(x);
1140                                 for (i = 0; i < asx_loops; i += n) {
1141                                         lock_delay_spin(n);
1142                                         x = SX_READ_VALUE(sx);
1143                                         if (!(x & SX_LOCK_SHARED))
1144                                                 break;
1145                                         n = SX_SHARERS(x);
1146                                         if (n == 0)
1147                                                 break;
1148                                         if (__sx_can_read(td, x, false))
1149                                                 break;
1150                                 }
1151 #ifdef KDTRACE_HOOKS
1152                                 lda.spin_cnt += i;
1153 #endif
1154                                 KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
1155                                     "running");
1156                                 if (i < asx_loops)
1157                                         continue;
1158                         }
1159                 }
1160 #endif
1161
1162                 /*
1163                  * Some other thread already has an exclusive lock, so
1164                  * start the process of blocking.
1165                  */
1166                 sleepq_lock(&sx->lock_object);
1167                 x = SX_READ_VALUE(sx);
1168 retry_sleepq:
1169                 if (((x & SX_LOCK_WRITE_SPINNER) && SX_SHARERS(x) == 0) ||
1170                     __sx_can_read(td, x, false)) {
1171                         sleepq_release(&sx->lock_object);
1172                         continue;
1173                 }
1174
1175 #ifdef ADAPTIVE_SX
1176                 /*
1177                  * If the owner is running on another CPU, spin until
1178                  * the owner stops running or the state of the lock
1179                  * changes.
1180                  */
1181                 if (!(x & SX_LOCK_SHARED)) {
1182                         owner = (struct thread *)SX_OWNER(x);
1183                         if (TD_IS_RUNNING(owner)) {
1184                                 sleepq_release(&sx->lock_object);
1185                                 x = SX_READ_VALUE(sx);
1186                                 continue;
1187                         }
1188                 }
1189 #endif
1190
1191                 /*
1192                  * Try to set the SX_LOCK_SHARED_WAITERS flag.  If we
1193                  * fail to set it drop the sleep queue lock and loop
1194                  * back.
1195                  */
1196                 if (!(x & SX_LOCK_SHARED_WAITERS)) {
1197                         if (!atomic_fcmpset_ptr(&sx->sx_lock, &x,
1198                             x | SX_LOCK_SHARED_WAITERS))
1199                                 goto retry_sleepq;
1200                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
1201                                 CTR2(KTR_LOCK, "%s: %p set shared waiters flag",
1202                                     __func__, sx);
1203                 }
1204
1205                 /*
1206                  * Since we have been unable to acquire the shared lock,
1207                  * we have to sleep.
1208                  */
1209                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1210                         CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
1211                             __func__, sx);
1212
1213 #ifdef KDTRACE_HOOKS
1214                 sleep_time -= lockstat_nsecs(&sx->lock_object);
1215 #endif
1216                 sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
1217                     SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
1218                     SLEEPQ_INTERRUPTIBLE : 0), SQ_SHARED_QUEUE);
1219                 /*
1220                  * Hack: this can land in thread_suspend_check which will
1221                  * conditionally take a mutex, tripping over an assert if a
1222                  * lock we are waiting for is set.
1223                  */
1224                 THREAD_CONTENTION_DONE(&sx->lock_object);
1225                 if (!(opts & SX_INTERRUPTIBLE))
1226                         sleepq_wait(&sx->lock_object, 0);
1227                 else
1228                         error = sleepq_wait_sig(&sx->lock_object, 0);
1229                 THREAD_CONTENDS_ON_LOCK(&sx->lock_object);
1230 #ifdef KDTRACE_HOOKS
1231                 sleep_time += lockstat_nsecs(&sx->lock_object);
1232                 sleep_cnt++;
1233 #endif
1234                 if (error) {
1235                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
1236                                 CTR2(KTR_LOCK,
1237                         "%s: interruptible sleep by %p suspended by signal",
1238                                     __func__, sx);
1239                         break;
1240                 }
1241                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1242                         CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
1243                             __func__, sx);
1244                 x = SX_READ_VALUE(sx);
1245         }
1246         THREAD_CONTENTION_DONE(&sx->lock_object);
1247 #if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
1248         if (__predict_true(!extra_work))
1249                 return (error);
1250 #endif
1251 #ifdef KDTRACE_HOOKS
1252         all_time += lockstat_nsecs(&sx->lock_object);
1253         if (sleep_time)
1254                 LOCKSTAT_RECORD4(sx__block, sx, sleep_time,
1255                     LOCKSTAT_READER, (state & SX_LOCK_SHARED) == 0,
1256                     (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
1257         if (lda.spin_cnt > sleep_cnt)
1258                 LOCKSTAT_RECORD4(sx__spin, sx, all_time - sleep_time,
1259                     LOCKSTAT_READER, (state & SX_LOCK_SHARED) == 0,
1260                     (state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
1261 out_lockstat:
1262 #endif
1263         if (error == 0) {
1264                 LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
1265                     contested, waittime, file, line, LOCKSTAT_READER);
1266         }
1267         GIANT_RESTORE();
1268         return (error);
1269 }
1270
1271 int
1272 _sx_slock_int(struct sx *sx, int opts LOCK_FILE_LINE_ARG_DEF)
1273 {
1274         struct thread *td;
1275         uintptr_t x;
1276         int error;
1277
1278         KASSERT(kdb_active != 0 || SCHEDULER_STOPPED() ||
1279             !TD_IS_IDLETHREAD(curthread),
1280             ("sx_slock() by idle thread %p on sx %s @ %s:%d",
1281             curthread, sx->lock_object.lo_name, file, line));
1282         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
1283             ("sx_slock() of destroyed sx @ %s:%d", file, line));
1284         WITNESS_CHECKORDER(&sx->lock_object, LOP_NEWORDER, file, line, NULL);
1285
1286         error = 0;
1287         td = curthread;
1288         x = SX_READ_VALUE(sx);
1289         if (__predict_false(LOCKSTAT_PROFILE_ENABLED(sx__acquire) ||
1290             !__sx_slock_try(sx, td, &x, true LOCK_FILE_LINE_ARG)))
1291                 error = _sx_slock_hard(sx, opts, x LOCK_FILE_LINE_ARG);
1292         else
1293                 lock_profile_obtain_lock_success(&sx->lock_object, false, 0, 0,
1294                     file, line);
1295         if (error == 0) {
1296                 LOCK_LOG_LOCK("SLOCK", &sx->lock_object, 0, 0, file, line);
1297                 WITNESS_LOCK(&sx->lock_object, 0, file, line);
1298                 TD_LOCKS_INC(curthread);
1299         }
1300         return (error);
1301 }
1302
1303 int
1304 _sx_slock(struct sx *sx, int opts, const char *file, int line)
1305 {
1306
1307         return (_sx_slock_int(sx, opts LOCK_FILE_LINE_ARG));
1308 }
1309
1310 static bool __always_inline
1311 _sx_sunlock_try(struct sx *sx, struct thread *td, uintptr_t *xp)
1312 {
1313
1314         for (;;) {
1315                 if (SX_SHARERS(*xp) > 1 || !(*xp & SX_LOCK_WAITERS)) {
1316                         if (atomic_fcmpset_rel_ptr(&sx->sx_lock, xp,
1317                             *xp - SX_ONE_SHARER)) {
1318                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1319                                         CTR4(KTR_LOCK,
1320                                             "%s: %p succeeded %p -> %p",
1321                                             __func__, sx, (void *)*xp,
1322                                             (void *)(*xp - SX_ONE_SHARER));
1323                                 td->td_sx_slocks--;
1324                                 return (true);
1325                         }
1326                         continue;
1327                 }
1328                 break;
1329         }
1330         return (false);
1331 }
1332
1333 static void __noinline
1334 _sx_sunlock_hard(struct sx *sx, struct thread *td, uintptr_t x
1335     LOCK_FILE_LINE_ARG_DEF)
1336 {
1337         int wakeup_swapper = 0;
1338         uintptr_t setx, queue;
1339
1340         if (SCHEDULER_STOPPED())
1341                 return;
1342
1343         if (_sx_sunlock_try(sx, td, &x))
1344                 goto out_lockstat;
1345
1346         sleepq_lock(&sx->lock_object);
1347         x = SX_READ_VALUE(sx);
1348         for (;;) {
1349                 if (_sx_sunlock_try(sx, td, &x))
1350                         break;
1351
1352                 /*
1353                  * Wake up semantic here is quite simple:
1354                  * Just wake up all the exclusive waiters.
1355                  * Note that the state of the lock could have changed,
1356                  * so if it fails loop back and retry.
1357                  */
1358                 setx = SX_LOCK_UNLOCKED;
1359                 queue = SQ_SHARED_QUEUE;
1360                 if (x & SX_LOCK_EXCLUSIVE_WAITERS) {
1361                         setx |= (x & SX_LOCK_SHARED_WAITERS);
1362                         queue = SQ_EXCLUSIVE_QUEUE;
1363                 }
1364                 setx |= (x & SX_LOCK_WRITE_SPINNER);
1365                 if (!atomic_fcmpset_rel_ptr(&sx->sx_lock, &x, setx))
1366                         continue;
1367                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
1368                         CTR2(KTR_LOCK, "%s: %p waking up all thread on"
1369                             "exclusive queue", __func__, sx);
1370                 wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX,
1371                     0, queue);
1372                 td->td_sx_slocks--;
1373                 break;
1374         }
1375         sleepq_release(&sx->lock_object);
1376         if (wakeup_swapper)
1377                 kick_proc0();
1378 out_lockstat:
1379         LOCKSTAT_PROFILE_RELEASE_RWLOCK(sx__release, sx, LOCKSTAT_READER);
1380 }
1381
1382 void
1383 _sx_sunlock_int(struct sx *sx LOCK_FILE_LINE_ARG_DEF)
1384 {
1385         struct thread *td;
1386         uintptr_t x;
1387
1388         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
1389             ("sx_sunlock() of destroyed sx @ %s:%d", file, line));
1390         _sx_assert(sx, SA_SLOCKED, file, line);
1391         WITNESS_UNLOCK(&sx->lock_object, 0, file, line);
1392         LOCK_LOG_LOCK("SUNLOCK", &sx->lock_object, 0, 0, file, line);
1393
1394         td = curthread;
1395         x = SX_READ_VALUE(sx);
1396         if (__predict_false(LOCKSTAT_PROFILE_ENABLED(sx__release) ||
1397             !_sx_sunlock_try(sx, td, &x)))
1398                 _sx_sunlock_hard(sx, td, x LOCK_FILE_LINE_ARG);
1399         else
1400                 lock_profile_release_lock(&sx->lock_object, false);
1401
1402         TD_LOCKS_DEC(curthread);
1403 }
1404
1405 void
1406 _sx_sunlock(struct sx *sx, const char *file, int line)
1407 {
1408
1409         _sx_sunlock_int(sx LOCK_FILE_LINE_ARG);
1410 }
1411
1412 #ifdef INVARIANT_SUPPORT
1413 #ifndef INVARIANTS
1414 #undef  _sx_assert
1415 #endif
1416
1417 /*
1418  * In the non-WITNESS case, sx_assert() can only detect that at least
1419  * *some* thread owns an slock, but it cannot guarantee that *this*
1420  * thread owns an slock.
1421  */
1422 void
1423 _sx_assert(const struct sx *sx, int what, const char *file, int line)
1424 {
1425 #ifndef WITNESS
1426         int slocked = 0;
1427 #endif
1428
1429         if (SCHEDULER_STOPPED())
1430                 return;
1431         switch (what) {
1432         case SA_SLOCKED:
1433         case SA_SLOCKED | SA_NOTRECURSED:
1434         case SA_SLOCKED | SA_RECURSED:
1435 #ifndef WITNESS
1436                 slocked = 1;
1437                 /* FALLTHROUGH */
1438 #endif
1439         case SA_LOCKED:
1440         case SA_LOCKED | SA_NOTRECURSED:
1441         case SA_LOCKED | SA_RECURSED:
1442 #ifdef WITNESS
1443                 witness_assert(&sx->lock_object, what, file, line);
1444 #else
1445                 /*
1446                  * If some other thread has an exclusive lock or we
1447                  * have one and are asserting a shared lock, fail.
1448                  * Also, if no one has a lock at all, fail.
1449                  */
1450                 if (sx->sx_lock == SX_LOCK_UNLOCKED ||
1451                     (!(sx->sx_lock & SX_LOCK_SHARED) && (slocked ||
1452                     sx_xholder(sx) != curthread)))
1453                         panic("Lock %s not %slocked @ %s:%d\n",
1454                             sx->lock_object.lo_name, slocked ? "share " : "",
1455                             file, line);
1456
1457                 if (!(sx->sx_lock & SX_LOCK_SHARED)) {
1458                         if (sx_recursed(sx)) {
1459                                 if (what & SA_NOTRECURSED)
1460                                         panic("Lock %s recursed @ %s:%d\n",
1461                                             sx->lock_object.lo_name, file,
1462                                             line);
1463                         } else if (what & SA_RECURSED)
1464                                 panic("Lock %s not recursed @ %s:%d\n",
1465                                     sx->lock_object.lo_name, file, line);
1466                 }
1467 #endif
1468                 break;
1469         case SA_XLOCKED:
1470         case SA_XLOCKED | SA_NOTRECURSED:
1471         case SA_XLOCKED | SA_RECURSED:
1472                 if (sx_xholder(sx) != curthread)
1473                         panic("Lock %s not exclusively locked @ %s:%d\n",
1474                             sx->lock_object.lo_name, file, line);
1475                 if (sx_recursed(sx)) {
1476                         if (what & SA_NOTRECURSED)
1477                                 panic("Lock %s recursed @ %s:%d\n",
1478                                     sx->lock_object.lo_name, file, line);
1479                 } else if (what & SA_RECURSED)
1480                         panic("Lock %s not recursed @ %s:%d\n",
1481                             sx->lock_object.lo_name, file, line);
1482                 break;
1483         case SA_UNLOCKED:
1484 #ifdef WITNESS
1485                 witness_assert(&sx->lock_object, what, file, line);
1486 #else
1487                 /*
1488                  * If we hold an exclusve lock fail.  We can't
1489                  * reliably check to see if we hold a shared lock or
1490                  * not.
1491                  */
1492                 if (sx_xholder(sx) == curthread)
1493                         panic("Lock %s exclusively locked @ %s:%d\n",
1494                             sx->lock_object.lo_name, file, line);
1495 #endif
1496                 break;
1497         default:
1498                 panic("Unknown sx lock assertion: %d @ %s:%d", what, file,
1499                     line);
1500         }
1501 }
1502 #endif  /* INVARIANT_SUPPORT */
1503
1504 #ifdef DDB
1505 static void
1506 db_show_sx(const struct lock_object *lock)
1507 {
1508         struct thread *td;
1509         const struct sx *sx;
1510
1511         sx = (const struct sx *)lock;
1512
1513         db_printf(" state: ");
1514         if (sx->sx_lock == SX_LOCK_UNLOCKED)
1515                 db_printf("UNLOCKED\n");
1516         else if (sx->sx_lock == SX_LOCK_DESTROYED) {
1517                 db_printf("DESTROYED\n");
1518                 return;
1519         } else if (sx->sx_lock & SX_LOCK_SHARED)
1520                 db_printf("SLOCK: %ju\n", (uintmax_t)SX_SHARERS(sx->sx_lock));
1521         else {
1522                 td = sx_xholder(sx);
1523                 db_printf("XLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
1524                     td->td_tid, td->td_proc->p_pid, td->td_name);
1525                 if (sx_recursed(sx))
1526                         db_printf(" recursed: %d\n", sx->sx_recurse);
1527         }
1528
1529         db_printf(" waiters: ");
1530         switch(sx->sx_lock &
1531             (SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS)) {
1532         case SX_LOCK_SHARED_WAITERS:
1533                 db_printf("shared\n");
1534                 break;
1535         case SX_LOCK_EXCLUSIVE_WAITERS:
1536                 db_printf("exclusive\n");
1537                 break;
1538         case SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS:
1539                 db_printf("exclusive and shared\n");
1540                 break;
1541         default:
1542                 db_printf("none\n");
1543         }
1544 }
1545
1546 /*
1547  * Check to see if a thread that is blocked on a sleep queue is actually
1548  * blocked on an sx lock.  If so, output some details and return true.
1549  * If the lock has an exclusive owner, return that in *ownerp.
1550  */
1551 int
1552 sx_chain(struct thread *td, struct thread **ownerp)
1553 {
1554         const struct sx *sx;
1555
1556         /*
1557          * Check to see if this thread is blocked on an sx lock.
1558          * First, we check the lock class.  If that is ok, then we
1559          * compare the lock name against the wait message.
1560          */
1561         sx = td->td_wchan;
1562         if (LOCK_CLASS(&sx->lock_object) != &lock_class_sx ||
1563             sx->lock_object.lo_name != td->td_wmesg)
1564                 return (0);
1565
1566         /* We think we have an sx lock, so output some details. */
1567         db_printf("blocked on sx \"%s\" ", td->td_wmesg);
1568         *ownerp = sx_xholder(sx);
1569         if (sx->sx_lock & SX_LOCK_SHARED)
1570                 db_printf("SLOCK (count %ju)\n",
1571                     (uintmax_t)SX_SHARERS(sx->sx_lock));
1572         else
1573                 db_printf("XLOCK\n");
1574         return (1);
1575 }
1576 #endif