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