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 assumes
 100  * curthread currently has an exclusive lock.
 101  */
 102 #define sx_recursed(sx)         ((sx)->sx_recurse != 0)
 103
 104 #ifdef DDB
 105 static void     db_show_sx(struct lock_object *lock);
 106 #endif
 107 static void     lock_sx(struct lock_object *lock, int how);
 108 static int      unlock_sx(struct lock_object *lock);
 109
 110 struct lock_class lock_class_sx = {
 111         .lc_name = "sx",
 112         .lc_flags = LC_SLEEPLOCK | LC_SLEEPABLE | LC_RECURSABLE | LC_UPGRADABLE,
 113 #ifdef DDB
 114         .lc_ddb_show = db_show_sx,
 115 #endif
 116         .lc_lock = lock_sx,
 117         .lc_unlock = unlock_sx,
 118 };
 119
 120 #ifndef INVARIANTS
 121 #define _sx_assert(sx, what, file, line)
 122 #endif
 123
 124 void
 125 lock_sx(struct lock_object *lock, int how)
 126 {
 127         struct sx *sx;
 128
 129         sx = (struct sx *)lock;
 130         if (how)
 131                 sx_xlock(sx);
 132         else
 133                 sx_slock(sx);
 134 }
 135
 136 int
 137 unlock_sx(struct lock_object *lock)
 138 {
 139         struct sx *sx;
 140
 141         sx = (struct sx *)lock;
 142         sx_assert(sx, SX_LOCKED | SX_NOTRECURSED);
 143         if (sx_xlocked(sx)) {
 144                 sx_xunlock(sx);
 145                 return (1);
 146         } else {
 147                 sx_sunlock(sx);
 148                 return (0);
 149         }
 150 }
 151
 152 void
 153 sx_sysinit(void *arg)
 154 {
 155         struct sx_args *sargs = arg;
 156
 157         sx_init(sargs->sa_sx, sargs->sa_desc);
 158 }
 159
 160 void
 161 sx_init_flags(struct sx *sx, const char *description, int opts)
 162 {
 163         int flags;
 164
 165         MPASS((opts & ~(SX_QUIET | SX_RECURSE | SX_NOWITNESS | SX_DUPOK |
 166             SX_NOPROFILE | SX_ADAPTIVESPIN)) == 0);
 167
 168         flags = LO_SLEEPABLE | LO_UPGRADABLE | LO_RECURSABLE;
 169         if (opts & SX_DUPOK)
 170                 flags |= LO_DUPOK;
 171         if (opts & SX_NOPROFILE)
 172                 flags |= LO_NOPROFILE;
 173         if (!(opts & SX_NOWITNESS))
 174                 flags |= LO_WITNESS;
 175         if (opts & SX_QUIET)
 176                 flags |= LO_QUIET;
 177
 178         flags |= opts & (SX_ADAPTIVESPIN | SX_RECURSE);
 179         sx->sx_lock = SX_LOCK_UNLOCKED;
 180         sx->sx_recurse = 0;
 181         lock_init(&sx->lock_object, &lock_class_sx, description, NULL, flags);
 182 }
 183
 184 void
 185 sx_destroy(struct sx *sx)
 186 {
 187
 188         KASSERT(sx->sx_lock == SX_LOCK_UNLOCKED, ("sx lock still held"));
 189         KASSERT(sx->sx_recurse == 0, ("sx lock still recursed"));
 190         sx->sx_lock = SX_LOCK_DESTROYED;
 191         lock_destroy(&sx->lock_object);
 192 }
 193
 194 void
 195 _sx_slock(struct sx *sx, const char *file, int line)
 196 {
 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         __sx_slock(sx, file, line);
 203         LOCK_LOG_LOCK("SLOCK", &sx->lock_object, 0, 0, file, line);
 204         WITNESS_LOCK(&sx->lock_object, 0, file, line);
 205         curthread->td_locks++;
 206 }
 207
 208 int
 209 _sx_try_slock(struct sx *sx, const char *file, int line)
 210 {
 211         uintptr_t x;
 212
 213         x = sx->sx_lock;
 214         KASSERT(x != SX_LOCK_DESTROYED,
 215             ("sx_try_slock() of destroyed sx @ %s:%d", file, line));
 216         if ((x & SX_LOCK_SHARED) && atomic_cmpset_acq_ptr(&sx->sx_lock, x,
 217             x + SX_ONE_SHARER)) {
 218                 LOCK_LOG_TRY("SLOCK", &sx->lock_object, 0, 1, file, line);
 219                 WITNESS_LOCK(&sx->lock_object, LOP_TRYLOCK, file, line);
 220                 curthread->td_locks++;
 221                 return (1);
 222         }
 223
 224         LOCK_LOG_TRY("SLOCK", &sx->lock_object, 0, 0, file, line);
 225         return (0);
 226 }
 227
 228 void
 229 _sx_xlock(struct sx *sx, const char *file, int line)
 230 {
 231
 232         MPASS(curthread != NULL);
 233         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 234             ("sx_xlock() of destroyed sx @ %s:%d", file, line));
 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         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 250             ("sx_try_xlock() of destroyed sx @ %s:%d", file, line));
 251
 252         if (sx_xlocked(sx) && (sx->lock_object.lo_flags & SX_RECURSE) != 0) {
 253                 sx->sx_recurse++;
 254                 atomic_set_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
 255                 rval = 1;
 256         } else
 257                 rval = atomic_cmpset_acq_ptr(&sx->sx_lock, SX_LOCK_UNLOCKED,
 258                     (uintptr_t)curthread);
 259         LOCK_LOG_TRY("XLOCK", &sx->lock_object, 0, rval, file, line);
 260         if (rval) {
 261                 WITNESS_LOCK(&sx->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
 262                     file, line);
 263                 curthread->td_locks++;
 264         }
 265
 266         return (rval);
 267 }
 268
 269 void
 270 _sx_sunlock(struct sx *sx, const char *file, int line)
 271 {
 272
 273         MPASS(curthread != NULL);
 274         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 275             ("sx_sunlock() of destroyed sx @ %s:%d", file, line));
 276         _sx_assert(sx, SX_SLOCKED, file, line);
 277         curthread->td_locks--;
 278         WITNESS_UNLOCK(&sx->lock_object, 0, file, line);
 279         LOCK_LOG_LOCK("SUNLOCK", &sx->lock_object, 0, 0, file, line);
 280         if (SX_SHARERS(sx->sx_lock) == 0)
 281                 lock_profile_release_lock(&sx->lock_object);
 282         __sx_sunlock(sx, file, line);
 283 }
 284
 285 void
 286 _sx_xunlock(struct sx *sx, const char *file, int line)
 287 {
 288
 289         MPASS(curthread != NULL);
 290         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 291             ("sx_xunlock() of destroyed sx @ %s:%d", file, line));
 292         _sx_assert(sx, SX_XLOCKED, file, line);
 293         curthread->td_locks--;
 294         WITNESS_UNLOCK(&sx->lock_object, LOP_EXCLUSIVE, file, line);
 295         LOCK_LOG_LOCK("XUNLOCK", &sx->lock_object, 0, sx->sx_recurse, file,
 296             line);
 297         if (!sx_recursed(sx))
 298                 lock_profile_release_lock(&sx->lock_object);
 299         __sx_xunlock(sx, curthread, file, line);
 300 }
 301
 302 /*
 303  * Try to do a non-blocking upgrade from a shared lock to an exclusive lock.
 304  * This will only succeed if this thread holds a single shared lock.
 305  * Return 1 if if the upgrade succeed, 0 otherwise.
 306  */
 307 int
 308 _sx_try_upgrade(struct sx *sx, const char *file, int line)
 309 {
 310         uintptr_t x;
 311         int success;
 312
 313         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 314             ("sx_try_upgrade() of destroyed sx @ %s:%d", file, line));
 315         _sx_assert(sx, SX_SLOCKED, file, line);
 316
 317         /*
 318          * Try to switch from one shared lock to an exclusive lock.  We need
 319          * to maintain the SX_LOCK_EXCLUSIVE_WAITERS flag if set so that
 320          * we will wake up the exclusive waiters when we drop the lock.
 321          */
 322         x = sx->sx_lock & SX_LOCK_EXCLUSIVE_WAITERS;
 323         success = atomic_cmpset_ptr(&sx->sx_lock, SX_SHARERS_LOCK(1) | x,
 324             (uintptr_t)curthread | x);
 325         LOCK_LOG_TRY("XUPGRADE", &sx->lock_object, 0, success, file, line);
 326         if (success)
 327                 WITNESS_UPGRADE(&sx->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
 328                     file, line);
 329         return (success);
 330 }
 331
 332 /*
 333  * Downgrade an unrecursed exclusive lock into a single shared lock.
 334  */
 335 void
 336 _sx_downgrade(struct sx *sx, const char *file, int line)
 337 {
 338         uintptr_t x;
 339
 340         KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
 341             ("sx_downgrade() of destroyed sx @ %s:%d", file, line));
 342         _sx_assert(sx, SX_XLOCKED | SX_NOTRECURSED, file, line);
 343 #ifndef INVARIANTS
 344         if (sx_recursed(sx))
 345                 panic("downgrade of a recursed lock");
 346 #endif
 347
 348         WITNESS_DOWNGRADE(&sx->lock_object, 0, file, line);
 349
 350         /*
 351          * Try to switch from an exclusive lock with no shared waiters
 352          * to one sharer with no shared waiters.  If there are
 353          * exclusive waiters, we don't need to lock the sleep queue so
 354          * long as we preserve the flag.  We do one quick try and if
 355          * that fails we grab the sleepq lock to keep the flags from
 356          * changing and do it the slow way.
 357          *
 358          * We have to lock the sleep queue if there are shared waiters
 359          * so we can wake them up.
 360          */
 361         x = sx->sx_lock;
 362         if (!(x & SX_LOCK_SHARED_WAITERS) &&
 363             atomic_cmpset_rel_ptr(&sx->sx_lock, x, SX_SHARERS_LOCK(1) |
 364             (x & SX_LOCK_EXCLUSIVE_WAITERS))) {
 365                 LOCK_LOG_LOCK("XDOWNGRADE", &sx->lock_object, 0, 0, file, line);
 366                 return;
 367         }
 368
 369         /*
 370          * Lock the sleep queue so we can read the waiters bits
 371          * without any races and wakeup any shared waiters.
 372          */
 373         sleepq_lock(&sx->lock_object);
 374
 375         /*
 376          * Preserve SX_LOCK_EXCLUSIVE_WAITERS while downgraded to a single
 377          * shared lock.  If there are any shared waiters, wake them up.
 378          */
 379         x = sx->sx_lock;
 380         atomic_store_rel_ptr(&sx->sx_lock, SX_SHARERS_LOCK(1) |
 381             (x & SX_LOCK_EXCLUSIVE_WAITERS));
 382         if (x & SX_LOCK_SHARED_WAITERS)
 383                 sleepq_broadcast(&sx->lock_object, SLEEPQ_SX, -1,
 384                     SQ_SHARED_QUEUE);
 385         else
 386                 sleepq_release(&sx->lock_object);
 387
 388         LOCK_LOG_LOCK("XDOWNGRADE", &sx->lock_object, 0, 0, file, line);
 389 }
 390
 391 /*
 392  * This function represents the so-called 'hard case' for sx_xlock
 393  * operation.  All 'easy case' failures are redirected to this.  Note
 394  * that ideally this would be a static function, but it needs to be
 395  * accessible from at least sx.h.
 396  */
 397 void
 398 _sx_xlock_hard(struct sx *sx, uintptr_t tid, const char *file, int line)
 399 {
 400         GIANT_DECLARE;
 401 #ifdef ADAPTIVE_SX
 402         volatile struct thread *owner;
 403 #endif
 404         uintptr_t x;
 405         int contested = 0;
 406         uint64_t waitstart = 0;
 407
 408         /* If we already hold an exclusive lock, then recurse. */
 409         if (sx_xlocked(sx)) {
 410                 KASSERT((sx->lock_object.lo_flags & SX_RECURSE) != 0,
 411             ("_sx_xlock_hard: recursed on non-recursive sx %s @ %s:%d\n",
 412                     sx->lock_object.lo_name, file, line));
 413                 sx->sx_recurse++;
 414                 atomic_set_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
 415                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 416                         CTR2(KTR_LOCK, "%s: %p recursing", __func__, sx);
 417                 return;
 418         }
 419         lock_profile_obtain_lock_failed(&(sx)->lock_object,
 420             &contested, &waitstart);
 421
 422         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 423                 CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__,
 424                     sx->lock_object.lo_name, (void *)sx->sx_lock, file, line);
 425
 426         while (!atomic_cmpset_acq_ptr(&sx->sx_lock, SX_LOCK_UNLOCKED, tid)) {
 427 #ifdef ADAPTIVE_SX
 428                 /*
 429                  * If the lock is write locked and the owner is
 430                  * running on another CPU, spin until the owner stops
 431                  * running or the state of the lock changes.
 432                  */
 433                 x = sx->sx_lock;
 434                 if (!(x & SX_LOCK_SHARED) &&
 435                     (sx->lock_object.lo_flags & SX_ADAPTIVESPIN)) {
 436                         x = SX_OWNER(x);
 437                         owner = (struct thread *)x;
 438                         if (TD_IS_RUNNING(owner)) {
 439                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 440                                         CTR3(KTR_LOCK,
 441                                             "%s: spinning on %p held by %p",
 442                                             __func__, sx, owner);
 443                                 GIANT_SAVE();
 444                                 while (SX_OWNER(sx->sx_lock) == x &&
 445                                     TD_IS_RUNNING(owner))
 446                                         cpu_spinwait();
 447                                 continue;
 448                         }
 449                 }
 450 #endif
 451
 452                 sleepq_lock(&sx->lock_object);
 453                 x = sx->sx_lock;
 454
 455                 /*
 456                  * If the lock was released while spinning on the
 457                  * sleep queue chain lock, try again.
 458                  */
 459                 if (x == SX_LOCK_UNLOCKED) {
 460                         sleepq_release(&sx->lock_object);
 461                         continue;
 462                 }
 463
 464 #ifdef ADAPTIVE_SX
 465                 /*
 466                  * The current lock owner might have started executing
 467                  * on another CPU (or the lock could have changed
 468                  * owners) while we were waiting on the sleep queue
 469                  * chain lock.  If so, drop the sleep queue lock and try
 470                  * again.
 471                  */
 472                 if (!(x & SX_LOCK_SHARED) &&
 473                     (sx->lock_object.lo_flags & SX_ADAPTIVESPIN)) {
 474                         owner = (struct thread *)SX_OWNER(x);
 475                         if (TD_IS_RUNNING(owner)) {
 476                                 sleepq_release(&sx->lock_object);
 477                                 continue;
 478                         }
 479                 }
 480 #endif
 481
 482                 /*
 483                  * If an exclusive lock was released with both shared
 484                  * and exclusive waiters and a shared waiter hasn't
 485                  * woken up and acquired the lock yet, sx_lock will be
 486                  * set to SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS.
 487                  * If we see that value, try to acquire it once.  Note
 488                  * that we have to preserve SX_LOCK_EXCLUSIVE_WAITERS
 489                  * as there are other exclusive waiters still.  If we
 490                  * fail, restart the loop.
 491                  */
 492                 if (x == (SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS)) {
 493                         if (atomic_cmpset_acq_ptr(&sx->sx_lock,
 494                             SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS,
 495                             tid | SX_LOCK_EXCLUSIVE_WAITERS)) {
 496                                 sleepq_release(&sx->lock_object);
 497                                 CTR2(KTR_LOCK, "%s: %p claimed by new writer",
 498                                     __func__, sx);
 499                                 break;
 500                         }
 501                         sleepq_release(&sx->lock_object);
 502                         continue;
 503                 }
 504
 505                 /*
 506                  * Try to set the SX_LOCK_EXCLUSIVE_WAITERS.  If we fail,
 507                  * than loop back and retry.
 508                  */
 509                 if (!(x & SX_LOCK_EXCLUSIVE_WAITERS)) {
 510                         if (!atomic_cmpset_ptr(&sx->sx_lock, x,
 511                             x | SX_LOCK_EXCLUSIVE_WAITERS)) {
 512                                 sleepq_release(&sx->lock_object);
 513                                 continue;
 514                         }
 515                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 516                                 CTR2(KTR_LOCK, "%s: %p set excl waiters flag",
 517                                     __func__, sx);
 518                 }
 519
 520                 /*
 521                  * Since we have been unable to acquire the exclusive
 522                  * lock and the exclusive waiters flag is set, we have
 523                  * to sleep.
 524                  */
 525                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 526                         CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
 527                             __func__, sx);
 528
 529                 GIANT_SAVE();
 530                 sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
 531                     SLEEPQ_SX, SQ_EXCLUSIVE_QUEUE);
 532                 sleepq_wait(&sx->lock_object);
 533
 534                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 535                         CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
 536                             __func__, sx);
 537         }
 538
 539         GIANT_RESTORE();
 540         lock_profile_obtain_lock_success(&(sx)->lock_object, contested,
 541             waitstart, file, line);
 542 }
 543
 544 /*
 545  * This function represents the so-called 'hard case' for sx_xunlock
 546  * operation.  All 'easy case' failures are redirected to this.  Note
 547  * that ideally this would be a static function, but it needs to be
 548  * accessible from at least sx.h.
 549  */
 550 void
 551 _sx_xunlock_hard(struct sx *sx, uintptr_t tid, const char *file, int line)
 552 {
 553         uintptr_t x;
 554         int queue;
 555
 556         MPASS(!(sx->sx_lock & SX_LOCK_SHARED));
 557
 558         /* If the lock is recursed, then unrecurse one level. */
 559         if (sx_xlocked(sx) && sx_recursed(sx)) {
 560                 if ((--sx->sx_recurse) == 0)
 561                         atomic_clear_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
 562                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 563                         CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, sx);
 564                 return;
 565         }
 566         MPASS(sx->sx_lock & (SX_LOCK_SHARED_WAITERS |
 567             SX_LOCK_EXCLUSIVE_WAITERS));
 568         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 569                 CTR2(KTR_LOCK, "%s: %p contested", __func__, sx);
 570
 571         sleepq_lock(&sx->lock_object);
 572         x = SX_LOCK_UNLOCKED;
 573
 574         /*
 575          * The wake up algorithm here is quite simple and probably not
 576          * ideal.  It gives precedence to shared waiters if they are
 577          * present.  For this condition, we have to preserve the
 578          * state of the exclusive waiters flag.
 579          */
 580         if (sx->sx_lock & SX_LOCK_SHARED_WAITERS) {
 581                 queue = SQ_SHARED_QUEUE;
 582                 x |= (sx->sx_lock & SX_LOCK_EXCLUSIVE_WAITERS);
 583         } else
 584                 queue = SQ_EXCLUSIVE_QUEUE;
 585
 586         /* Wake up all the waiters for the specific queue. */
 587         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 588                 CTR3(KTR_LOCK, "%s: %p waking up all threads on %s queue",
 589                     __func__, sx, queue == SQ_SHARED_QUEUE ? "shared" :
 590                     "exclusive");
 591         atomic_store_rel_ptr(&sx->sx_lock, x);
 592         sleepq_broadcast(&sx->lock_object, SLEEPQ_SX, -1, queue);
 593 }
 594
 595 /*
 596  * This function represents the so-called 'hard case' for sx_slock
 597  * operation.  All 'easy case' failures are redirected to this.  Note
 598  * that ideally this would be a static function, but it needs to be
 599  * accessible from at least sx.h.
 600  */
 601 void
 602 _sx_slock_hard(struct sx *sx, const char *file, int line)
 603 {
 604         GIANT_DECLARE;
 605 #ifdef ADAPTIVE_SX
 606         volatile struct thread *owner;
 607 #endif
 608         uintptr_t x;
 609         uint64_t waitstart = 0;
 610         int contested = 0;
 611         /*
 612          * As with rwlocks, we don't make any attempt to try to block
 613          * shared locks once there is an exclusive waiter.
 614          */
 615
 616         for (;;) {
 617                 x = sx->sx_lock;
 618
 619                 /*
 620                  * If no other thread has an exclusive lock then try to bump up
 621                  * the count of sharers.  Since we have to preserve the state
 622                  * of SX_LOCK_EXCLUSIVE_WAITERS, if we fail to acquire the
 623                  * shared lock loop back and retry.
 624                  */
 625                 if (x & SX_LOCK_SHARED) {
 626                         MPASS(!(x & SX_LOCK_SHARED_WAITERS));
 627                         if (atomic_cmpset_acq_ptr(&sx->sx_lock, x,
 628                             x + SX_ONE_SHARER)) {
 629                                 if (SX_SHARERS(x) == 0)
 630                                         lock_profile_obtain_lock_success(
 631                                             &sx->lock_object, contested,
 632                                             waitstart, file, line);
 633                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 634                                         CTR4(KTR_LOCK,
 635                                             "%s: %p succeed %p -> %p", __func__,
 636                                             sx, (void *)x,
 637                                             (void *)(x + SX_ONE_SHARER));
 638                                 break;
 639                         }
 640                         lock_profile_obtain_lock_failed(&sx->lock_object, &contested,
 641                             &waitstart);
 642
 643                         continue;
 644                 }
 645
 646 #ifdef ADAPTIVE_SX
 647                 /*
 648                  * If the owner is running on another CPU, spin until
 649                  * the owner stops running or the state of the lock
 650                  * changes.
 651                  */
 652                 else if (sx->lock_object.lo_flags & SX_ADAPTIVESPIN) {
 653                         x = SX_OWNER(x);
 654                         owner = (struct thread *)x;
 655                         if (TD_IS_RUNNING(owner)) {
 656                                 lock_profile_obtain_lock_failed(&sx->lock_object, &contested,
 657                                     &waitstart);
 658                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 659                                         CTR3(KTR_LOCK,
 660                                             "%s: spinning on %p held by %p",
 661                                             __func__, sx, owner);
 662                                 GIANT_SAVE();
 663                                 while (SX_OWNER(sx->sx_lock) == x &&
 664                                     TD_IS_RUNNING(owner))
 665                                         cpu_spinwait();
 666                                 continue;
 667                         }
 668                 }
 669 #endif
 670                 else
 671                         lock_profile_obtain_lock_failed(&sx->lock_object, &contested,
 672                             &waitstart);
 673
 674                 /*
 675                  * Some other thread already has an exclusive lock, so
 676                  * start the process of blocking.
 677                  */
 678                 sleepq_lock(&sx->lock_object);
 679                 x = sx->sx_lock;
 680
 681                 /*
 682                  * The lock could have been released while we spun.
 683                  * In this case loop back and retry.
 684                  */
 685                 if (x & SX_LOCK_SHARED) {
 686                         sleepq_release(&sx->lock_object);
 687                         continue;
 688                 }
 689
 690 #ifdef ADAPTIVE_SX
 691                 /*
 692                  * If the owner is running on another CPU, spin until
 693                  * the owner stops running or the state of the lock
 694                  * changes.
 695                  */
 696                 if (!(x & SX_LOCK_SHARED) &&
 697                     (sx->lock_object.lo_flags & SX_ADAPTIVESPIN)) {
 698                         owner = (struct thread *)SX_OWNER(x);
 699                         if (TD_IS_RUNNING(owner)) {
 700                                 sleepq_release(&sx->lock_object);
 701                                 continue;
 702                         }
 703                 }
 704 #endif
 705
 706                 /*
 707                  * Try to set the SX_LOCK_SHARED_WAITERS flag.  If we
 708                  * fail to set it drop the sleep queue lock and loop
 709                  * back.
 710                  */
 711                 if (!(x & SX_LOCK_SHARED_WAITERS)) {
 712                         if (!atomic_cmpset_ptr(&sx->sx_lock, x,
 713                             x | SX_LOCK_SHARED_WAITERS)) {
 714                                 sleepq_release(&sx->lock_object);
 715                                 continue;
 716                         }
 717                         if (LOCK_LOG_TEST(&sx->lock_object, 0))
 718                                 CTR2(KTR_LOCK, "%s: %p set shared waiters flag",
 719                                     __func__, sx);
 720                 }
 721
 722                 /*
 723                  * Since we have been unable to acquire the shared lock,
 724                  * we have to sleep.
 725                  */
 726                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 727                         CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
 728                             __func__, sx);
 729
 730                 GIANT_SAVE();
 731                 sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
 732                     SLEEPQ_SX, SQ_SHARED_QUEUE);
 733                 sleepq_wait(&sx->lock_object);
 734
 735                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 736                         CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
 737                             __func__, sx);
 738         }
 739
 740         GIANT_RESTORE();
 741 }
 742
 743 /*
 744  * This function represents the so-called 'hard case' for sx_sunlock
 745  * operation.  All 'easy case' failures are redirected to this.  Note
 746  * that ideally this would be a static function, but it needs to be
 747  * accessible from at least sx.h.
 748  */
 749 void
 750 _sx_sunlock_hard(struct sx *sx, const char *file, int line)
 751 {
 752         uintptr_t x;
 753
 754         for (;;) {
 755                 x = sx->sx_lock;
 756
 757                 /*
 758                  * We should never have sharers while at least one thread
 759                  * holds a shared lock.
 760                  */
 761                 KASSERT(!(x & SX_LOCK_SHARED_WAITERS),
 762                     ("%s: waiting sharers", __func__));
 763
 764                 /*
 765                  * See if there is more than one shared lock held.  If
 766                  * so, just drop one and return.
 767                  */
 768                 if (SX_SHARERS(x) > 1) {
 769                         if (atomic_cmpset_ptr(&sx->sx_lock, x,
 770                             x - SX_ONE_SHARER)) {
 771                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 772                                         CTR4(KTR_LOCK,
 773                                             "%s: %p succeeded %p -> %p",
 774                                             __func__, sx, (void *)x,
 775                                             (void *)(x - SX_ONE_SHARER));
 776                                 break;
 777                         }
 778                         continue;
 779                 }
 780
 781                 /*
 782                  * If there aren't any waiters for an exclusive lock,
 783                  * then try to drop it quickly.
 784                  */
 785                 if (!(x & SX_LOCK_EXCLUSIVE_WAITERS)) {
 786                         MPASS(x == SX_SHARERS_LOCK(1));
 787                         if (atomic_cmpset_ptr(&sx->sx_lock, SX_SHARERS_LOCK(1),
 788                             SX_LOCK_UNLOCKED)) {
 789                                 lock_profile_release_lock(&sx->lock_object);
 790                                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 791                                         CTR2(KTR_LOCK, "%s: %p last succeeded",
 792                                             __func__, sx);
 793                                 break;
 794                         }
 795                         continue;
 796                 }
 797
 798                 /*
 799                  * At this point, there should just be one sharer with
 800                  * exclusive waiters.
 801                  */
 802                 MPASS(x == (SX_SHARERS_LOCK(1) | SX_LOCK_EXCLUSIVE_WAITERS));
 803
 804                 lock_profile_release_lock(&sx->lock_object);
 805                 sleepq_lock(&sx->lock_object);
 806
 807                 /*
 808                  * Wake up semantic here is quite simple:
 809                  * Just wake up all the exclusive waiters.
 810                  * Note that the state of the lock could have changed,
 811                  * so if it fails loop back and retry.
 812                  */
 813                 if (!atomic_cmpset_ptr(&sx->sx_lock,
 814                     SX_SHARERS_LOCK(1) | SX_LOCK_EXCLUSIVE_WAITERS,
 815                     SX_LOCK_UNLOCKED)) {
 816                         sleepq_release(&sx->lock_object);
 817                         continue;
 818                 }
 819                 if (LOCK_LOG_TEST(&sx->lock_object, 0))
 820                         CTR2(KTR_LOCK, "%s: %p waking up all thread on"
 821                             "exclusive queue", __func__, sx);
 822                 sleepq_broadcast(&sx->lock_object, SLEEPQ_SX, -1,
 823                     SQ_EXCLUSIVE_QUEUE);
 824                 break;
 825         }
 826 }
 827
 828 #ifdef INVARIANT_SUPPORT
 829 #ifndef INVARIANTS
 830 #undef  _sx_assert
 831 #endif
 832
 833 /*
 834  * In the non-WITNESS case, sx_assert() can only detect that at least
 835  * *some* thread owns an slock, but it cannot guarantee that *this*
 836  * thread owns an slock.
 837  */
 838 void
 839 _sx_assert(struct sx *sx, int what, const char *file, int line)
 840 {
 841 #ifndef WITNESS
 842         int slocked = 0;
 843 #endif
 844
 845         if (panicstr != NULL)
 846                 return;
 847         switch (what) {
 848         case SX_SLOCKED:
 849         case SX_SLOCKED | SX_NOTRECURSED:
 850         case SX_SLOCKED | SX_RECURSED:
 851 #ifndef WITNESS
 852                 slocked = 1;
 853                 /* FALLTHROUGH */
 854 #endif
 855         case SX_LOCKED:
 856         case SX_LOCKED | SX_NOTRECURSED:
 857         case SX_LOCKED | SX_RECURSED:
 858 #ifdef WITNESS
 859                 witness_assert(&sx->lock_object, what, file, line);
 860 #else
 861                 /*
 862                  * If some other thread has an exclusive lock or we
 863                  * have one and are asserting a shared lock, fail.
 864                  * Also, if no one has a lock at all, fail.
 865                  */
 866                 if (sx->sx_lock == SX_LOCK_UNLOCKED ||
 867                     (!(sx->sx_lock & SX_LOCK_SHARED) && (slocked ||
 868                     sx_xholder(sx) != curthread)))
 869                         panic("Lock %s not %slocked @ %s:%d\n",
 870                             sx->lock_object.lo_name, slocked ? "share " : "",
 871                             file, line);
 872
 873                 if (!(sx->sx_lock & SX_LOCK_SHARED)) {
 874                         if (sx_recursed(sx)) {
 875                                 if (what & SX_NOTRECURSED)
 876                                         panic("Lock %s recursed @ %s:%d\n",
 877                                             sx->lock_object.lo_name, file,
 878                                             line);
 879                         } else if (what & SX_RECURSED)
 880                                 panic("Lock %s not recursed @ %s:%d\n",
 881                                     sx->lock_object.lo_name, file, line);
 882                 }
 883 #endif
 884                 break;
 885         case SX_XLOCKED:
 886         case SX_XLOCKED | SX_NOTRECURSED:
 887         case SX_XLOCKED | SX_RECURSED:
 888                 if (sx_xholder(sx) != curthread)
 889                         panic("Lock %s not exclusively locked @ %s:%d\n",
 890                             sx->lock_object.lo_name, file, line);
 891                 if (sx_recursed(sx)) {
 892                         if (what & SX_NOTRECURSED)
 893                                 panic("Lock %s recursed @ %s:%d\n",
 894                                     sx->lock_object.lo_name, file, line);
 895                 } else if (what & SX_RECURSED)
 896                         panic("Lock %s not recursed @ %s:%d\n",
 897                             sx->lock_object.lo_name, file, line);
 898                 break;
 899         case SX_UNLOCKED:
 900 #ifdef WITNESS
 901                 witness_assert(&sx->lock_object, what, file, line);
 902 #else
 903                 /*
 904                  * If we hold an exclusve lock fail.  We can't
 905                  * reliably check to see if we hold a shared lock or
 906                  * not.
 907                  */
 908                 if (sx_xholder(sx) == curthread)
 909                         panic("Lock %s exclusively locked @ %s:%d\n",
 910                             sx->lock_object.lo_name, file, line);
 911 #endif
 912                 break;
 913         default:
 914                 panic("Unknown sx lock assertion: %d @ %s:%d", what, file,
 915                     line);
 916         }
 917 }
 918 #endif  /* INVARIANT_SUPPORT */
 919
 920 #ifdef DDB
 921 static void
 922 db_show_sx(struct lock_object *lock)
 923 {
 924         struct thread *td;
 925         struct sx *sx;
 926
 927         sx = (struct sx *)lock;
 928
 929         db_printf(" state: ");
 930         if (sx->sx_lock == SX_LOCK_UNLOCKED)
 931                 db_printf("UNLOCKED\n");
 932         else if (sx->sx_lock == SX_LOCK_DESTROYED) {
 933                 db_printf("DESTROYED\n");
 934                 return;
 935         } else if (sx->sx_lock & SX_LOCK_SHARED)
 936                 db_printf("SLOCK: %ju\n", (uintmax_t)SX_SHARERS(sx->sx_lock));
 937         else {
 938                 td = sx_xholder(sx);
 939                 db_printf("XLOCK: %p (tid %d, pid %d, \"%s\")\n", td,
 940                     td->td_tid, td->td_proc->p_pid, td->td_proc->p_comm);
 941                 if (sx_recursed(sx))
 942                         db_printf(" recursed: %d\n", sx->sx_recurse);
 943         }
 944
 945         db_printf(" waiters: ");
 946         switch(sx->sx_lock &
 947             (SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS)) {
 948         case SX_LOCK_SHARED_WAITERS:
 949                 db_printf("shared\n");
 950                 break;
 951         case SX_LOCK_EXCLUSIVE_WAITERS:
 952                 db_printf("exclusive\n");
 953                 break;
 954         case SX_LOCK_SHARED_WAITERS | SX_LOCK_EXCLUSIVE_WAITERS:
 955                 db_printf("exclusive and shared\n");
 956                 break;
 957         default:
 958                 db_printf("none\n");
 959         }
 960 }
 961
 962 /*
 963  * Check to see if a thread that is blocked on a sleep queue is actually
 964  * blocked on an sx lock.  If so, output some details and return true.
 965  * If the lock has an exclusive owner, return that in *ownerp.
 966  */
 967 int
 968 sx_chain(struct thread *td, struct thread **ownerp)
 969 {
 970         struct sx *sx;
 971
 972         /*
 973          * Check to see if this thread is blocked on an sx lock.
 974          * First, we check the lock class.  If that is ok, then we
 975          * compare the lock name against the wait message.
 976          */
 977         sx = td->td_wchan;
 978         if (LOCK_CLASS(&sx->lock_object) != &lock_class_sx ||
 979             sx->lock_object.lo_name != td->td_wmesg)
 980                 return (0);
 981
 982         /* We think we have an sx lock, so output some details. */
 983         db_printf("blocked on sx \"%s\" ", td->td_wmesg);
 984         *ownerp = sx_xholder(sx);
 985         if (sx->sx_lock & SX_LOCK_SHARED)
 986                 db_printf("SLOCK (count %ju)\n",
 987                     (uintmax_t)SX_SHARERS(sx->sx_lock));
 988         else
 989                 db_printf("XLOCK\n");
 990         return (1);
 991 }
 992 #endif