sys/kern/kern_sx.c

   1 /*-
   2  * Copyright (c) 2007 Attilio Rao <attilio@freebsd.org>
   3  * Copyright (c) 2001 Jason Evans <jasone@freebsd.org>
   4  * All rights reserved.
   5  *
   6  * Redistribution and use in source and binary forms, with or without
   7  * modification, are permitted provided that the following conditions
   8  * are met:
   9  * 1. Redistributions of source code must retain the above copyright
  10  *    notice(s), this list of conditions and the following disclaimer as
  11  *    the first lines of this file unmodified other than the possible
  12  *    addition of one or more copyright notices.
  13  * 2. Redistributions in binary form must reproduce the above copyright
  14  *    notice(s), this list of conditions and the following disclaimer in the
  15  *    documentation and/or other materials provided with the distribution.
  16  *
  17  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDER(S) ``AS IS'' AND ANY
  18  * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
  19  * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
  20  * DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT HOLDER(S) BE LIABLE FOR ANY
  21  * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
  22  * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
  23  * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
  24  * CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH
  27  * DAMAGE.
  28  */
  29
  30 /*
  31  * Shared/exclusive locks.  This implementation attempts to ensure
  32  * deterministic lock granting behavior, so that slocks and xlocks are
  33  * interleaved.
  34  *
  35  * Priority propagation will not generally raise the priority of lock holders,
  36  * so should not be relied upon in combination with sx locks.
  37  */
  38
  39 #include "opt_adaptive_sx.h"
  40 #include "opt_ddb.h"
  41
  42 #include <sys/cdefs.h>
  43 __FBSDID("$FreeBSD$");
  44
  45 #include <sys/param.h>
  46 #include <sys/ktr.h>
  47 #include <sys/lock.h>
  48 #include <sys/mutex.h>
  49 #include <sys/proc.h>
  50 #include <sys/sleepqueue.h>
  51 #include <sys/sx.h>
  52 #include <sys/systm.h>
  53
  54 #ifdef ADAPTIVE_SX
  55 #include <machine/cpu.h>
  56 #endif
  57
  58 #ifdef DDB
  59 #include <ddb/ddb.h>
  60 #endif
  61
  62 #if !defined(SMP) && defined(ADAPTIVE_SX)
  63 #error "You must have SMP to enable the ADAPTIVE_SX option"
  64 #endif
  65
  66 CTASSERT(((SX_ADAPTIVESPIN | SX_RECURSE) & LO_CLASSFLAGS) ==
  67     (SX_ADAPTIVESPIN | SX_RECURSE));
  68
  69 /* Handy macros for sleep queues. */
  70 #define SQ_EXCLUSIVE_QUEUE      0
  71 #define SQ_SHARED_QUEUE         1
  72
  73 /*
  74  * Variations on DROP_GIANT()/PICKUP_GIANT() for use in this file.  We
  75  * drop Giant anytime we have to sleep or if we adaptively spin.
  76  */
  77 #define GIANT_DECLARE                                                   \
  78         int _giantcnt = 0;                                              \
  79         WITNESS_SAVE_DECL(Giant)                                        \
  80
  81 #define GIANT_SAVE() do {                                               \
  82         if (mtx_owned(&Giant)) {                                        \
  83                 WITNESS_SAVE(&Giant.lock_object, Giant);                \
  84                 while (mtx_owned(&Giant)) {                             \
  85                         _giantcnt++;                                    \
  86                         mtx_unlock(&Giant);                             \
  87                 }                                                       \
  88         }                                                               \
  89 } while (0)
  90
  91 #define GIANT_RESTORE() do {                                            \
  92         if (_giantcnt > 0) {                                            \
  93                 mtx_assert(&Giant, MA_NOTOWNED);                        \
  94                 while (_giantcnt--)                                     \
  95                         mtx_lock(&Giant);                               \
  96                 WITNESS_RESTORE(&Giant.lock_object, Giant);             \
  97         }                                                               \
  98 } while (0)
  99
 100 /*
 101  * Returns true if an exclusive lock is recursed.  It assumes
 102  * curthread currently has an exclusive lock.
 103  */
 104 #define sx_recursed(sx)         ((sx)->sx_recurse != 0)
 105
 106 static void     assert_sx(struct lock_object *lock, int what);
 107 #ifdef DDB
 108 static void     db_show_sx(struct lock_object *lock);
 109 #endif
 110 static void     lock_sx(struct lock_object *lock, int how);
 111 static int      unlock_sx(struct lock_object *lock);
 112
 113 struct lock_class lock_class_sx = {
 114         .lc_name = "sx",
 115         .lc_flags = LC_SLEEPLOCK | LC_SLEEPABLE | LC_RECURSABLE | LC_UPGRADABLE,
 116         .lc_assert = assert_sx,
 117 #ifdef DDB
 118         .lc_ddb_show = db_show_sx,
 119 #endif
 120         .lc_lock = lock_sx,
 121         .lc_unlock = unlock_sx,
 122 };
 123
 124 #ifndef INVARIANTS
 125 #define _sx_assert(sx, what, file, line)
 126 #endif
 127
 128 void
 129 assert_sx(struct lock_object *lock, int what)
 130 {
 131
 132         sx_assert((struct sx *)lock, what);
 133 }
 134
 135 void
 136 lock_sx(struct lock_object *lock, int how)
 137 {
 138         struct sx *sx;
 139
 140         sx = (struct sx *)lock;
 141         if (how)
 142                 sx_xlock(sx);
 143         else
 144                 sx_slock(sx);
 145 }
 146
 147 int
 148 unlock_sx(struct lock_object *lock)
 149 {
 150         struct sx *sx;
 151
 152         sx = (struct sx *)lock;
 153         sx_assert(sx, SA_LOCKED | SA_NOTRECURSED);
 154         if (sx_xlocked(sx)) {
 155                 sx_xunlock(sx);
 156                 return (1);
 157         } else {
 158                 sx_sunlock(sx);
 159                 return (0);
 160         }
 161 }
 162
 163 void
 164 sx_sysinit(void *arg)
 165 {
 166         struct sx_args *sargs = arg;
 167
 168         sx_init(sargs->sa_sx, sargs->sa_desc);
 169 }
 170
 171 void
 172 sx_init_flags(struct sx *sx, const char *description, int opts)
 173 {
 174         int flags;
 175
 176         MPASS((opts & ~(SX_QUIET | SX_RECURSE | SX_NOWITNESS | SX_DUPOK |
 177             SX_NOPROFILE | SX_ADAPTIVESPIN)) == 0);
 178
 179         flags = LO_RECURSABLE | LO_SLEEPABLE | LO_UPGRADABLE;
 180         if (opts & SX_DUPOK)
 181                 flags |= LO_DUPOK;
 182         if (opts & SX_NOPROFILE)
 183                 flags |= LO_NOPROFILE;
 184         if (!(opts & SX_NOWITNESS))
 185                 flags |= LO_WITNESS;
 186         if (opts & SX_QUIET)
 187                 flags |= LO_QUIET;
 188
 189         flags |= opts & (SX_ADAPTIVESPIN | SX_RECURSE);
 190         sx->sx_lock = SX_LOCK_UNLOCKED;
 191         sx->sx_recurse = 0;
 192         lock_init(&sx->lock_object, &lock_class_sx, description, NULL, flags);
 193 }
 194
 195 void
 196 sx_destroy(struct sx *sx)
 197 {
 198
 199         KASSERT(sx->sx_lock == SX_LOCK_UNLOCKED, ("sx lock still held"));
 200         KASSERT(sx->sx_recurse == 0, ("sx lock still recursed"));
 201         sx->sx_lock = SX_LOCK_DESTROYED;
 202         lock_destroy(&sx->lock_object);
 203 }
 204
 205 int
 206 _sx_slock(struct sx *sx, int opts, const char *file, int line)
 207 {
 208         int error = 0;
 209
 210         MPASS(curthread != NULL);
 211         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 212             ("sx_slock() of destroyed sx @ %s:%d", file, line));
 213         WITNESS_CHECKORDER(&sx->lock_object, LOP_NEWORDER, file, line);
 214         error = __sx_slock(sx, opts, file, line);
 215         if (!error) {
 216                 LOCK_LOG_LOCK("SLOCK", &sx->lock_object, 0, 0, file, line);
 217                 WITNESS_LOCK(&sx->lock_object, 0, file, line);
 218                 curthread->td_locks++;
 219         }
 220
 221         return (error);
 222 }
 223
 224 int
 225 _sx_try_slock(struct sx *sx, const char *file, int line)
 226 {
 227         uintptr_t x;
 228
 229         for (;;) {
 230                 x = sx->sx_lock;
 231                 KASSERT(x != SX_LOCK_DESTROYED,
 232                     ("sx_try_slock() of destroyed sx @ %s:%d", file, line));
 233                 if (!(x & SX_LOCK_SHARED))
 234                         break;
 235                 if (atomic_cmpset_acq_ptr(&sx->sx_lock, x, x + SX_ONE_SHARER)) {
 236                         LOCK_LOG_TRY("SLOCK", &sx->lock_object, 0, 1, file, line);
 237                         WITNESS_LOCK(&sx->lock_object, LOP_TRYLOCK, file, line);
 238                         curthread->td_locks++;
 239                         return (1);
 240                 }
 241         }
 242
 243         LOCK_LOG_TRY("SLOCK", &sx->lock_object, 0, 0, file, line);
 244         return (0);
 245 }
 246
 247 int
 248 _sx_xlock(struct sx *sx, int opts, const char *file, int line)
 249 {
 250         int error = 0;
 251
 252         MPASS(curthread != NULL);
 253         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 254             ("sx_xlock() of destroyed sx @ %s:%d", file, line));
 255         WITNESS_CHECKORDER(&sx->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE, file,
 256             line);
 257         error = __sx_xlock(sx, curthread, opts, file, line);
 258         if (!error) {
 259                 LOCK_LOG_LOCK("XLOCK", &sx->lock_object, 0, sx->sx_recurse,
 260                     file, line);
 261                 WITNESS_LOCK(&sx->lock_object, LOP_EXCLUSIVE, file, line);
 262                 curthread->td_locks++;
 263         }
 264
 265         return (error);
 266 }
 267
 268 int
 269 _sx_try_xlock(struct sx *sx, const char *file, int line)
 270 {
 271         int rval;
 272
 273         MPASS(curthread != NULL);
 274         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 275             ("sx_try_xlock() of destroyed sx @ %s:%d", file, line));
 276
 277         if (sx_xlocked(sx) && (sx->lock_object.lo_flags & SX_RECURSE) != 0) {
 278                 sx->sx_recurse++;
 279                 atomic_set_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
 280                 rval = 1;
 281         } else
 282                 rval = atomic_cmpset_acq_ptr(&sx->sx_lock, SX_LOCK_UNLOCKED,
 283                     (uintptr_t)curthread);
 284         LOCK_LOG_TRY("XLOCK", &sx->lock_object, 0, rval, file, line);
 285         if (rval) {
 286                 WITNESS_LOCK(&sx->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
 287                     file, line);
 288                 curthread->td_locks++;
 289         }
 290
 291         return (rval);
 292 }
 293
 294 void
 295 _sx_sunlock(struct sx *sx, const char *file, int line)
 296 {
 297
 298         MPASS(curthread != NULL);
 299         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 300             ("sx_sunlock() of destroyed sx @ %s:%d", file, line));
 301         _sx_assert(sx, SA_SLOCKED, file, line);
 302         curthread->td_locks--;
 303         WITNESS_UNLOCK(&sx->lock_object, 0, file, line);
 304         LOCK_LOG_LOCK("SUNLOCK", &sx->lock_object, 0, 0, file, line);
 305         __sx_sunlock(sx, file, line);
 306         lock_profile_release_lock(&sx->lock_object);
 307 }
 308
 309 void
 310 _sx_xunlock(struct sx *sx, const char *file, int line)
 311 {
 312
 313         MPASS(curthread != NULL);
 314         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 315             ("sx_xunlock() of destroyed sx @ %s:%d", file, line));
 316         _sx_assert(sx, SA_XLOCKED, file, line);
 317         curthread->td_locks--;
 318         WITNESS_UNLOCK(&sx->lock_object, LOP_EXCLUSIVE, file, line);
 319         LOCK_LOG_LOCK("XUNLOCK", &sx->lock_object, 0, sx->sx_recurse, file,
 320             line);
 321         if (!sx_recursed(sx))
 322                 lock_profile_release_lock(&sx->lock_object);
 323         __sx_xunlock(sx, curthread, file, line);
 324 }
 325
 326 /*
 327  * Try to do a non-blocking upgrade from a shared lock to an exclusive lock.
 328  * This will only succeed if this thread holds a single shared lock.
 329  * Return 1 if if the upgrade succeed, 0 otherwise.
 330  */
 331 int
 332 _sx_try_upgrade(struct sx *sx, const char *file, int line)
 333 {
 334         uintptr_t x;
 335         int success;
 336
 337         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 338             ("sx_try_upgrade() of destroyed sx @ %s:%d", file, line));
 339         _sx_assert(sx, SA_SLOCKED, file, line);
 340
 341         /*
 342          * Try to switch from one shared lock to an exclusive lock.  We need
 343          * to maintain the SX_LOCK_EXCLUSIVE_WAITERS flag if set so that
 344          * we will wake up the exclusive waiters when we drop the lock.
 345          */
 346         x = sx->sx_lock & SX_LOCK_EXCLUSIVE_WAITERS;
 347         success = atomic_cmpset_ptr(&sx->sx_lock, SX_SHARERS_LOCK(1) | x,
 348             (uintptr_t)curthread | x);
 349         LOCK_LOG_TRY("XUPGRADE", &sx->lock_object, 0, success, file, line);
 350         if (success)
 351                 WITNESS_UPGRADE(&sx->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
 352                     file, line);
 353         return (success);
 354 }
 355
 356 /*
 357  * Downgrade an unrecursed exclusive lock into a single shared lock.
 358  */
 359 void
 360 _sx_downgrade(struct sx *sx, const char *file, int line)
 361 {
 362         uintptr_t x;
 363
 364         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 365             ("sx_downgrade() of destroyed sx @ %s:%d", file, line));
 366         _sx_assert(sx, SA_XLOCKED | SA_NOTRECURSED, file, line);
 367 #ifndef INVARIANTS
 368         if (sx_recursed(sx))
 369                 panic("downgrade of a recursed lock");
 370 #endif
 371
 372         WITNESS_DOWNGRADE(&sx->lock_object, 0, file, line);
 373
 374         /*
 375          * Try to switch from an exclusive lock with no shared waiters
 376          * to one sharer with no shared waiters.  If there are
 377          * exclusive waiters, we don't need to lock the sleep queue so
 378          * long as we preserve the flag.  We do one quick try and if
 379          * that fails we grab the sleepq lock to keep the flags from
 380          * changing and do it the slow way.
 381          *
 382          * We have to lock the sleep queue if there are shared waiters
 383          * so we can wake them up.
 384          */
 385         x = sx->sx_lock;
 386         if (!(x & SX_LOCK_SHARED_WAITERS) &&
 387             atomic_cmpset_rel_ptr(&sx->sx_lock, x, SX_SHARERS_LOCK(1) |
 388             (x & SX_LOCK_EXCLUSIVE_WAITERS))) {
 389                 LOCK_LOG_LOCK("XDOWNGRADE", &sx->lock_object, 0, 0, file, line);
 390                 return;
 391         }
 392
 393         /*
 394          * Lock the sleep queue so we can read the waiters bits
 395          * without any races and wakeup any shared waiters.
 396          */
 397         sleepq_lock(&sx->lock_object);
 398
 399         /*
 400          * Preserve SX_LOCK_EXCLUSIVE_WAITERS while downgraded to a single
 401          * shared lock.  If there are any shared waiters, wake them up.
 402          */
 403         x = sx->sx_lock;
 404         atomic_store_rel_ptr(&sx->sx_lock, SX_SHARERS_LOCK(1) |
 405             (x & SX_LOCK_EXCLUSIVE_WAITERS));
 406         if (x & SX_LOCK_SHARED_WAITERS)
 407                 sleepq_broadcast(&sx->lock_object, SLEEPQ_SX, -1,
 408                     SQ_SHARED_QUEUE);
 409         else
 410                 sleepq_release(&sx->lock_object);
 411
 412         LOCK_LOG_LOCK("XDOWNGRADE", &sx->lock_object, 0, 0, file, line);
 413 }
 414
 415 /*
 416  * This function represents the so-called 'hard case' for sx_xlock
 417  * operation.  All 'easy case' failures are redirected to this.  Note
 418  * that ideally this would be a static function, but it needs to be
 419  * accessible from at least sx.h.
 420  */
 421 int
 422 _sx_xlock_hard(struct sx *sx, uintptr_t tid, int opts, const char *file,
 423     int line)
 424 {
 425         GIANT_DECLARE;
 426 #ifdef ADAPTIVE_SX
 427         volatile struct thread *owner;
 428 #endif
 429         uint64_t waittime = 0;
 430         uintptr_t x;
 431         int contested = 0, error = 0;
 432
 433         /* If we already hold an exclusive lock, then recurse. */
 434         if (sx_xlocked(sx)) {
 435                 KASSERT((sx->lock_object.lo_flags & SX_RECURSE) != 0,
 436             ("_sx_xlock_hard: recursed on non-recursive sx %s @ %s:%d\n",
 437                     sx->lock_object.lo_name, file, line));
 438                 sx->sx_recurse++;
 439                 atomic_set_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
 440                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 441                         CTR2(KTR_LOCK, "%s: %p recursing", __func__, sx);
 442                 return (0);
 443         }
 444
 445         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 446                 CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__,
 447                     sx->lock_object.lo_name, (void *)sx->sx_lock, file, line);
 448
 449         while (!atomic_cmpset_acq_ptr(&sx->sx_lock, SX_LOCK_UNLOCKED, tid)) {
 450                 lock_profile_obtain_lock_failed(&sx->lock_object, &contested,
 451                     &waittime);
 452 #ifdef ADAPTIVE_SX
 453                 /*
 454                  * If the lock is write locked and the owner is
 455                  * running on another CPU, spin until the owner stops
 456                  * running or the state of the lock changes.
 457                  */
 458                 x = sx->sx_lock;
 459                 if (!(x & SX_LOCK_SHARED) &&
 460                     (sx->lock_object.lo_flags & SX_ADAPTIVESPIN)) {
 461                         x = SX_OWNER(x);
 462                         owner = (struct thread *)x;
 463                         if (TD_IS_RUNNING(owner)) {
 464                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 465                                         CTR3(KTR_LOCK,
 466                                             "%s: spinning on %p held by %p",
 467                                             __func__, sx, owner);
 468                                 GIANT_SAVE();
 469                                 while (SX_OWNER(sx->sx_lock) == x &&
 470                                     TD_IS_RUNNING(owner))
 471                                         cpu_spinwait();
 472                                 continue;
 473                         }
 474                 }
 475 #endif
 476
 477                 sleepq_lock(&sx->lock_object);
 478                 x = sx->sx_lock;
 479
 480                 /*
 481                  * If the lock was released while spinning on the
 482                  * sleep queue chain lock, try again.
 483                  */
 484                 if (x == SX_LOCK_UNLOCKED) {
 485                         sleepq_release(&sx->lock_object);
 486                         continue;
 487                 }
 488
 489 #ifdef ADAPTIVE_SX
 490                 /*
 491                  * The current lock owner might have started executing
 492                  * on another CPU (or the lock could have changed
 493                  * owners) while we were waiting on the sleep queue
 494                  * chain lock.  If so, drop the sleep queue lock and try
 495                  * again.
 496                  */
 497                 if (!(x & SX_LOCK_SHARED) &&
 498                     (sx->lock_object.lo_flags & SX_ADAPTIVESPIN)) {
 499                         owner = (struct thread *)SX_OWNER(x);
 500                         if (TD_IS_RUNNING(owner)) {
 501                                 sleepq_release(&sx->lock_object);
 502                                 continue;
 503                         }
 504                 }
 505 #endif
 506
 507                 /*
 508                  * If an exclusive lock was released with both shared
 509                  * and exclusive waiters and a shared waiter hasn't
 510                  * woken up and acquired the lock yet, sx_lock will be
 511                  * set to SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS.
 512                  * If we see that value, try to acquire it once.  Note
 513                  * that we have to preserve SX_LOCK_EXCLUSIVE_WAITERS
 514                  * as there are other exclusive waiters still.  If we
 515                  * fail, restart the loop.
 516                  */
 517                 if (x == (SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS)) {
 518                         if (atomic_cmpset_acq_ptr(&sx->sx_lock,
 519                             SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS,
 520                             tid | SX_LOCK_EXCLUSIVE_WAITERS)) {
 521                                 sleepq_release(&sx->lock_object);
 522                                 CTR2(KTR_LOCK, "%s: %p claimed by new writer",
 523                                     __func__, sx);
 524                                 break;
 525                         }
 526                         sleepq_release(&sx->lock_object);
 527                         continue;
 528                 }
 529
 530                 /*
 531                  * Try to set the SX_LOCK_EXCLUSIVE_WAITERS.  If we fail,
 532                  * than loop back and retry.
 533                  */
 534                 if (!(x & SX_LOCK_EXCLUSIVE_WAITERS)) {
 535                         if (!atomic_cmpset_ptr(&sx->sx_lock, x,
 536                             x | SX_LOCK_EXCLUSIVE_WAITERS)) {
 537                                 sleepq_release(&sx->lock_object);
 538                                 continue;
 539                         }
 540                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 541                                 CTR2(KTR_LOCK, "%s: %p set excl waiters flag",
 542                                     __func__, sx);
 543                 }
 544
 545                 /*
 546                  * Since we have been unable to acquire the exclusive
 547                  * lock and the exclusive waiters flag is set, we have
 548                  * to sleep.
 549                  */
 550                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 551                         CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
 552                             __func__, sx);
 553
 554                 GIANT_SAVE();
 555                 sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
 556                     SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
 557                     SLEEPQ_INTERRUPTIBLE : 0), SQ_EXCLUSIVE_QUEUE);
 558                 if (!(opts & SX_INTERRUPTIBLE))
 559                         sleepq_wait(&sx->lock_object);
 560                 else
 561                         error = sleepq_wait_sig(&sx->lock_object);
 562
 563                 if (error) {
 564                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 565                                 CTR2(KTR_LOCK,
 566                         "%s: interruptible sleep by %p suspended by signal",
 567                                     __func__, sx);
 568                         break;
 569                 }
 570                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 571                         CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
 572                             __func__, sx);
 573         }
 574
 575         GIANT_RESTORE();
 576         if (!error)
 577                 lock_profile_obtain_lock_success(&sx->lock_object, contested,
 578                     waittime, file, line);
 579         return (error);
 580 }
 581
 582 /*
 583  * This function represents the so-called 'hard case' for sx_xunlock
 584  * operation.  All 'easy case' failures are redirected to this.  Note
 585  * that ideally this would be a static function, but it needs to be
 586  * accessible from at least sx.h.
 587  */
 588 void
 589 _sx_xunlock_hard(struct sx *sx, uintptr_t tid, const char *file, int line)
 590 {
 591         uintptr_t x;
 592         int queue;
 593
 594         MPASS(!(sx->sx_lock & SX_LOCK_SHARED));
 595
 596         /* If the lock is recursed, then unrecurse one level. */
 597         if (sx_xlocked(sx) && sx_recursed(sx)) {
 598                 if ((--sx->sx_recurse) == 0)
 599                         atomic_clear_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
 600                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 601                         CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, sx);
 602                 return;
 603         }
 604         MPASS(sx->sx_lock & (SX_LOCK_SHARED_WAITERS |
 605             SX_LOCK_EXCLUSIVE_WAITERS));
 606         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 607                 CTR2(KTR_LOCK, "%s: %p contested", __func__, sx);
 608
 609         sleepq_lock(&sx->lock_object);
 610         x = SX_LOCK_UNLOCKED;
 611
 612         /*
 613          * The wake up algorithm here is quite simple and probably not
 614          * ideal.  It gives precedence to shared waiters if they are
 615          * present.  For this condition, we have to preserve the
 616          * state of the exclusive waiters flag.
 617          */
 618         if (sx->sx_lock & SX_LOCK_SHARED_WAITERS) {
 619                 queue = SQ_SHARED_QUEUE;
 620                 x |= (sx->sx_lock & SX_LOCK_EXCLUSIVE_WAITERS);
 621         } else
 622                 queue = SQ_EXCLUSIVE_QUEUE;
 623
 624         /* Wake up all the waiters for the specific queue. */
 625         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 626                 CTR3(KTR_LOCK, "%s: %p waking up all threads on %s queue",
 627                     __func__, sx, queue == SQ_SHARED_QUEUE ? "shared" :
 628                     "exclusive");
 629         atomic_store_rel_ptr(&sx->sx_lock, x);
 630         sleepq_broadcast(&sx->lock_object, SLEEPQ_SX, -1, queue);
 631 }
 632
 633 /*
 634  * This function represents the so-called 'hard case' for sx_slock
 635  * operation.  All 'easy case' failures are redirected to this.  Note
 636  * that ideally this would be a static function, but it needs to be
 637  * accessible from at least sx.h.
 638  */
 639 int
 640 _sx_slock_hard(struct sx *sx, int opts, const char *file, int line)
 641 {
 642         GIANT_DECLARE;
 643 #ifdef ADAPTIVE_SX
 644         volatile struct thread *owner;
 645 #endif
 646         uint64_t waittime = 0;
 647         int contested = 0;
 648         uintptr_t x;
 649         int error = 0;
 650
 651         /*
 652          * As with rwlocks, we don't make any attempt to try to block
 653          * shared locks once there is an exclusive waiter.
 654          */
 655         for (;;) {
 656                 x = sx->sx_lock;
 657
 658                 /*
 659                  * If no other thread has an exclusive lock then try to bump up
 660                  * the count of sharers.  Since we have to preserve the state
 661                  * of SX_LOCK_EXCLUSIVE_WAITERS, if we fail to acquire the
 662                  * shared lock loop back and retry.
 663                  */
 664                 if (x & SX_LOCK_SHARED) {
 665                         MPASS(!(x & SX_LOCK_SHARED_WAITERS));
 666                         if (atomic_cmpset_acq_ptr(&sx->sx_lock, x,
 667                             x + SX_ONE_SHARER)) {
 668                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 669                                         CTR4(KTR_LOCK,
 670                                             "%s: %p succeed %p -> %p", __func__,
 671                                             sx, (void *)x,
 672                                             (void *)(x + SX_ONE_SHARER));
 673                                 break;
 674                         }
 675                         continue;
 676                 }
 677                 lock_profile_obtain_lock_failed(&sx->lock_object, &contested,
 678                     &waittime);
 679
 680 #ifdef ADAPTIVE_SX
 681                 /*
 682                  * If the owner is running on another CPU, spin until
 683                  * the owner stops running or the state of the lock
 684                  * changes.
 685                  */
 686                 if (sx->lock_object.lo_flags & SX_ADAPTIVESPIN) {
 687                         x = SX_OWNER(x);
 688                         owner = (struct thread *)x;
 689                         if (TD_IS_RUNNING(owner)) {
 690                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 691                                         CTR3(KTR_LOCK,
 692                                             "%s: spinning on %p held by %p",
 693                                             __func__, sx, owner);
 694                                 GIANT_SAVE();
 695                                 while (SX_OWNER(sx->sx_lock) == x &&
 696                                     TD_IS_RUNNING(owner))
 697                                         cpu_spinwait();
 698                                 continue;
 699                         }
 700                 }
 701 #endif
 702
 703                 /*
 704                  * Some other thread already has an exclusive lock, so
 705                  * start the process of blocking.
 706                  */
 707                 sleepq_lock(&sx->lock_object);
 708                 x = sx->sx_lock;
 709
 710                 /*
 711                  * The lock could have been released while we spun.
 712                  * In this case loop back and retry.
 713                  */
 714                 if (x & SX_LOCK_SHARED) {
 715                         sleepq_release(&sx->lock_object);
 716                         continue;
 717                 }
 718
 719 #ifdef ADAPTIVE_SX
 720                 /*
 721                  * If the owner is running on another CPU, spin until
 722                  * the owner stops running or the state of the lock
 723                  * changes.
 724                  */
 725                 if (!(x & SX_LOCK_SHARED) &&
 726                     (sx->lock_object.lo_flags & SX_ADAPTIVESPIN)) {
 727                         owner = (struct thread *)SX_OWNER(x);
 728                         if (TD_IS_RUNNING(owner)) {
 729                                 sleepq_release(&sx->lock_object);
 730                                 continue;
 731                         }
 732                 }
 733 #endif
 734
 735                 /*
 736                  * Try to set the SX_LOCK_SHARED_WAITERS flag.  If we
 737                  * fail to set it drop the sleep queue lock and loop
 738                  * back.
 739                  */
 740                 if (!(x & SX_LOCK_SHARED_WAITERS)) {
 741                         if (!atomic_cmpset_ptr(&sx->sx_lock, x,
 742                             x | SX_LOCK_SHARED_WAITERS)) {
 743                                 sleepq_release(&sx->lock_object);
 744                                 continue;
 745                         }
 746                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 747                                 CTR2(KTR_LOCK, "%s: %p set shared waiters flag",
 748                                     __func__, sx);
 749                 }
 750
 751                 /*
 752                  * Since we have been unable to acquire the shared lock,
 753                  * we have to sleep.
 754                  */
 755                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 756                         CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
 757                             __func__, sx);
 758
 759                 GIANT_SAVE();
 760                 sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
 761                     SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
 762                     SLEEPQ_INTERRUPTIBLE : 0), SQ_SHARED_QUEUE);
 763                 if (!(opts & SX_INTERRUPTIBLE))
 764                         sleepq_wait(&sx->lock_object);
 765                 else
 766                         error = sleepq_wait_sig(&sx->lock_object);
 767
 768                 if (error) {
 769                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 770                                 CTR2(KTR_LOCK,
 771                         "%s: interruptible sleep by %p suspended by signal",
 772                                     __func__, sx);
 773                         break;
 774                 }
 775                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 776                         CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
 777                             __func__, sx);
 778         }
 779         if (error == 0)
 780                 lock_profile_obtain_lock_success(&sx->lock_object, contested,
 781                     waittime, file, line);
 782
 783         GIANT_RESTORE();
 784         return (error);
 785 }
 786
 787 /*
 788  * This function represents the so-called 'hard case' for sx_sunlock
 789  * operation.  All 'easy case' failures are redirected to this.  Note
 790  * that ideally this would be a static function, but it needs to be
 791  * accessible from at least sx.h.
 792  */
 793 void
 794 _sx_sunlock_hard(struct sx *sx, const char *file, int line)
 795 {
 796         uintptr_t x;
 797
 798         for (;;) {
 799                 x = sx->sx_lock;
 800
 801                 /*
 802                  * We should never have sharers while at least one thread
 803                  * holds a shared lock.
 804                  */
 805                 KASSERT(!(x & SX_LOCK_SHARED_WAITERS),
 806                     ("%s: waiting sharers", __func__));
 807
 808                 /*
 809                  * See if there is more than one shared lock held.  If
 810                  * so, just drop one and return.
 811                  */
 812                 if (SX_SHARERS(x) > 1) {
 813                         if (atomic_cmpset_ptr(&sx->sx_lock, x,
 814                             x - SX_ONE_SHARER)) {
 815                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 816                                         CTR4(KTR_LOCK,
 817                                             "%s: %p succeeded %p -> %p",
 818                                             __func__, sx, (void *)x,
 819                                             (void *)(x - SX_ONE_SHARER));
 820                                 break;
 821                         }
 822                         continue;
 823                 }
 824
 825                 /*
 826                  * If there aren't any waiters for an exclusive lock,
 827                  * then try to drop it quickly.
 828                  */
 829                 if (!(x & SX_LOCK_EXCLUSIVE_WAITERS)) {
 830                         MPASS(x == SX_SHARERS_LOCK(1));
 831                         if (atomic_cmpset_ptr(&sx->sx_lock, SX_SHARERS_LOCK(1),
 832                             SX_LOCK_UNLOCKED)) {
 833                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 834                                         CTR2(KTR_LOCK, "%s: %p last succeeded",
 835                                             __func__, sx);
 836                                 break;
 837                         }
 838                         continue;
 839                 }
 840
 841                 /*
 842                  * At this point, there should just be one sharer with
 843                  * exclusive waiters.
 844                  */
 845                 MPASS(x == (SX_SHARERS_LOCK(1) | SX_LOCK_EXCLUSIVE_WAITERS));
 846
 847                 sleepq_lock(&sx->lock_object);
 848
 849                 /*
 850                  * Wake up semantic here is quite simple:
 851                  * Just wake up all the exclusive waiters.
 852                  * Note that the state of the lock could have changed,
 853                  * so if it fails loop back and retry.
 854                  */
 855                 if (!atomic_cmpset_ptr(&sx->sx_lock,
 856                     SX_SHARERS_LOCK(1) | SX_LOCK_EXCLUSIVE_WAITERS,
 857                     SX_LOCK_UNLOCKED)) {
 858                         sleepq_release(&sx->lock_object);
 859                         continue;
 860                 }
 861                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 862                         CTR2(KTR_LOCK, "%s: %p waking up all thread on"
 863                             "exclusive queue", __func__, sx);
 864                 sleepq_broadcast(&sx->lock_object, SLEEPQ_SX, -1,
 865                     SQ_EXCLUSIVE_QUEUE);
 866                 break;
 867         }
 868 }
 869
 870 #ifdef INVARIANT_SUPPORT
 871 #ifndef INVARIANTS
 872 #undef  _sx_assert
 873 #endif
 874
 875 /*
 876  * In the non-WITNESS case, sx_assert() can only detect that at least
 877  * *some* thread owns an slock, but it cannot guarantee that *this*
 878  * thread owns an slock.
 879  */
 880 void
 881 _sx_assert(struct sx *sx, int what, const char *file, int line)
 882 {
 883 #ifndef WITNESS
 884         int slocked = 0;
 885 #endif
 886
 887         if (panicstr != NULL)
 888                 return;
 889         switch (what) {
 890         case SA_SLOCKED:
 891         case SA_SLOCKED | SA_NOTRECURSED:
 892         case SA_SLOCKED | SA_RECURSED:
 893 #ifndef WITNESS
 894                 slocked = 1;
 895                 /* FALLTHROUGH */
 896 #endif
 897         case SA_LOCKED:
 898         case SA_LOCKED | SA_NOTRECURSED:
 899         case SA_LOCKED | SA_RECURSED:
 900 #ifdef WITNESS
 901                 witness_assert(&sx->lock_object, what, file, line);
 902 #else
 903                 /*
 904                  * If some other thread has an exclusive lock or we
 905                  * have one and are asserting a shared lock, fail.
 906                  * Also, if no one has a lock at all, fail.
 907                  */
 908                 if (sx->sx_lock == SX_LOCK_UNLOCKED ||
 909                     (!(sx->sx_lock & SX_LOCK_SHARED) && (slocked ||
 910                     sx_xholder(sx) != curthread)))
 911                         panic("Lock %s not %slocked @ %s:%d\n",
 912                             sx->lock_object.lo_name, slocked ? "share " : "",
 913                             file, line);
 914
 915                 if (!(sx->sx_lock & SX_LOCK_SHARED)) {
 916                         if (sx_recursed(sx)) {
 917                                 if (what & SA_NOTRECURSED)
 918                                         panic("Lock %s recursed @ %s:%d\n",
 919                                             sx->lock_object.lo_name, file,
 920                                             line);
 921                         } else if (what & SA_RECURSED)
 922                                 panic("Lock %s not recursed @ %s:%d\n",
 923                                     sx->lock_object.lo_name, file, line);
 924                 }
 925 #endif
 926                 break;
 927         case SA_XLOCKED:
 928         case SA_XLOCKED | SA_NOTRECURSED:
 929         case SA_XLOCKED | SA_RECURSED:
 930                 if (sx_xholder(sx) != curthread)
 931                         panic("Lock %s not exclusively locked @ %s:%d\n",
 932                             sx->lock_object.lo_name, file, line);
 933                 if (sx_recursed(sx)) {
 934                         if (what & SA_NOTRECURSED)
 935                                 panic("Lock %s recursed @ %s:%d\n",
 936                                     sx->lock_object.lo_name, file, line);
 937                 } else if (what & SA_RECURSED)
 938                         panic("Lock %s not recursed @ %s:%d\n",
 939                             sx->lock_object.lo_name, file, line);
 940                 break;
 941         case SA_UNLOCKED:
 942 #ifdef WITNESS
 943                 witness_assert(&sx->lock_object, what, file, line);
 944 #else
 945                 /*
 946                  * If we hold an exclusve lock fail.  We can't
 947                  * reliably check to see if we hold a shared lock or
 948                  * not.
 949                  */
 950                 if (sx_xholder(sx) == curthread)
 951                         panic("Lock %s exclusively locked @ %s:%d\n",
 952                             sx->lock_object.lo_name, file, line);
 953 #endif
 954                 break;
 955         default:
 956                 panic("Unknown sx lock assertion: %d @ %s:%d", what, file,
 957                     line);
 958         }
 959 }
 960 #endif  /* INVARIANT_SUPPORT */
 961
 962 #ifdef DDB
 963 static void
 964 db_show_sx(struct lock_object *lock)
 965 {
 966         struct thread *td;
 967         struct sx *sx;
 968
 969         sx = (struct sx *)lock;
 970
 971         db_printf(" state: ");
 972         if (sx->sx_lock == SX_LOCK_UNLOCKED)
 973                 db_printf("UNLOCKED\n");
 974         else if (sx->sx_lock == SX_LOCK_DESTROYED) {
 975                 db_printf("DESTROYED\n");
 976                 return;
 977         } else if (sx->sx_lock & SX_LOCK_SHARED)
 978                 db_printf("SLOCK: %ju\n", (uintmax_t)SX_SHARERS(sx->sx_lock));
 979         else {
 980                 td = sx_xholder(sx);
 981                 db_printf("XLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
 982                     td->td_tid, td->td_proc->p_pid, td->td_name);
 983                 if (sx_recursed(sx))
 984                         db_printf(" recursed: %d\n", sx->sx_recurse);
 985         }
 986
 987         db_printf(" waiters: ");
 988         switch(sx->sx_lock &
 989             (SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS)) {
 990         case SX_LOCK_SHARED_WAITERS:
 991                 db_printf("shared\n");
 992                 break;
 993         case SX_LOCK_EXCLUSIVE_WAITERS:
 994                 db_printf("exclusive\n");
 995                 break;
 996         case SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS:
 997                 db_printf("exclusive and shared\n");
 998                 break;
 999         default:
1000                 db_printf("none\n");
1001         }
1002 }
1003
1004 /*
1005  * Check to see if a thread that is blocked on a sleep queue is actually
1006  * blocked on an sx lock.  If so, output some details and return true.
1007  * If the lock has an exclusive owner, return that in *ownerp.
1008  */
1009 int
1010 sx_chain(struct thread *td, struct thread **ownerp)
1011 {
1012         struct sx *sx;
1013
1014         /*
1015          * Check to see if this thread is blocked on an sx lock.
1016          * First, we check the lock class.  If that is ok, then we
1017          * compare the lock name against the wait message.
1018          */
1019         sx = td->td_wchan;
1020         if (LOCK_CLASS(&sx->lock_object) != &lock_class_sx ||
1021             sx->lock_object.lo_name != td->td_wmesg)
1022                 return (0);
1023
1024         /* We think we have an sx lock, so output some details. */
1025         db_printf("blocked on sx \"%s\" ", td->td_wmesg);
1026         *ownerp = sx_xholder(sx);
1027         if (sx->sx_lock & SX_LOCK_SHARED)
1028                 db_printf("SLOCK (count %ju)\n",
1029                     (uintmax_t)SX_SHARERS(sx->sx_lock));
1030         else
1031                 db_printf("XLOCK\n");
1032         return (1);
1033 }
1034 #endif