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