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