rts/Capability.c

   1 /* ---------------------------------------------------------------------------
   2  *
   3  * (c) The GHC Team, 2003-2006
   4  *
   5  * Capabilities
   6  *
   7  * A Capability represent the token required to execute STG code,
   8  * and all the state an OS thread/task needs to run Haskell code:
   9  * its STG registers, a pointer to its TSO, a nursery etc. During
  10  * STG execution, a pointer to the capabilitity is kept in a
  11  * register (BaseReg; actually it is a pointer to cap->r).
  12  *
  13  * Only in an THREADED_RTS build will there be multiple capabilities,
  14  * for non-threaded builds there is only one global capability, namely
  15  * MainCapability.
  16  *
  17  * --------------------------------------------------------------------------*/
  18
  19 #include "PosixSource.h"
  20 #include "Rts.h"
  21
  22 #include "Capability.h"
  23 #include "Schedule.h"
  24 #include "Sparks.h"
  25 #include "Trace.h"
  26 #include "sm/GC.h" // for gcWorkerThread()
  27 #include "STM.h"
  28 #include "RtsUtils.h"
  29
  30 // one global capability, this is the Capability for non-threaded
  31 // builds, and for +RTS -N1
  32 Capability MainCapability;
  33
  34 nat n_capabilities = 0;
  35 Capability *capabilities = NULL;
  36
  37 // Holds the Capability which last became free.  This is used so that
  38 // an in-call has a chance of quickly finding a free Capability.
  39 // Maintaining a global free list of Capabilities would require global
  40 // locking, so we don't do that.
  41 Capability *last_free_capability = NULL;
  42
  43 /* GC indicator, in scope for the scheduler, init'ed to false */
  44 volatile StgWord waiting_for_gc = 0;
  45
  46 /* Let foreign code get the current Capability -- assuming there is one!
  47  * This is useful for unsafe foreign calls because they are called with
  48  * the current Capability held, but they are not passed it. For example,
  49  * see see the integer-gmp package which calls allocateLocal() in its
  50  * stgAllocForGMP() function (which gets called by gmp functions).
  51  * */
  52 Capability * rts_unsafeGetMyCapability (void)
  53 {
  54 #if defined(THREADED_RTS)
  55   return myTask()->cap;
  56 #else
  57   return &MainCapability;
  58 #endif
  59 }
  60
  61 #if defined(THREADED_RTS)
  62 STATIC_INLINE rtsBool
  63 globalWorkToDo (void)
  64 {
  65     return sched_state >= SCHED_INTERRUPTING
  66         || recent_activity == ACTIVITY_INACTIVE; // need to check for deadlock
  67 }
  68 #endif
  69
  70 #if defined(THREADED_RTS)
  71 StgClosure *
  72 findSpark (Capability *cap)
  73 {
  74   Capability *robbed;
  75   StgClosurePtr spark;
  76   rtsBool retry;
  77   nat i = 0;
  78
  79   if (!emptyRunQueue(cap) || cap->returning_tasks_hd != NULL) {
  80       // If there are other threads, don't try to run any new
  81       // sparks: sparks might be speculative, we don't want to take
  82       // resources away from the main computation.
  83       return 0;
  84   }
  85
  86   do {
  87       retry = rtsFalse;
  88
  89       // first try to get a spark from our own pool.
  90       // We should be using reclaimSpark(), because it works without
  91       // needing any atomic instructions:
  92       //   spark = reclaimSpark(cap->sparks);
  93       // However, measurements show that this makes at least one benchmark
  94       // slower (prsa) and doesn't affect the others.
  95       spark = tryStealSpark(cap);
  96       if (spark != NULL) {
  97           cap->sparks_converted++;
  98
  99           // Post event for running a spark from capability's own pool.
 100           traceEventRunSpark(cap, cap->r.rCurrentTSO);
 101
 102           return spark;
 103       }
 104       if (!emptySparkPoolCap(cap)) {
 105           retry = rtsTrue;
 106       }
 107
 108       if (n_capabilities == 1) { return NULL; } // makes no sense...
 109
 110       debugTrace(DEBUG_sched,
 111                  "cap %d: Trying to steal work from other capabilities",
 112                  cap->no);
 113
 114       /* visit cap.s 0..n-1 in sequence until a theft succeeds. We could
 115       start at a random place instead of 0 as well.  */
 116       for ( i=0 ; i < n_capabilities ; i++ ) {
 117           robbed = &capabilities[i];
 118           if (cap == robbed)  // ourselves...
 119               continue;
 120
 121           if (emptySparkPoolCap(robbed)) // nothing to steal here
 122               continue;
 123
 124           spark = tryStealSpark(robbed);
 125           if (spark == NULL && !emptySparkPoolCap(robbed)) {
 126               // we conflicted with another thread while trying to steal;
 127               // try again later.
 128               retry = rtsTrue;
 129           }
 130
 131           if (spark != NULL) {
 132               cap->sparks_converted++;
 133
 134               traceEventStealSpark(cap, cap->r.rCurrentTSO, robbed->no);
 135
 136               return spark;
 137           }
 138           // otherwise: no success, try next one
 139       }
 140   } while (retry);
 141
 142   debugTrace(DEBUG_sched, "No sparks stolen");
 143   return NULL;
 144 }
 145
 146 // Returns True if any spark pool is non-empty at this moment in time
 147 // The result is only valid for an instant, of course, so in a sense
 148 // is immediately invalid, and should not be relied upon for
 149 // correctness.
 150 rtsBool
 151 anySparks (void)
 152 {
 153     nat i;
 154
 155     for (i=0; i < n_capabilities; i++) {
 156         if (!emptySparkPoolCap(&capabilities[i])) {
 157             return rtsTrue;
 158         }
 159     }
 160     return rtsFalse;
 161 }
 162 #endif
 163
 164 /* -----------------------------------------------------------------------------
 165  * Manage the returning_tasks lists.
 166  *
 167  * These functions require cap->lock
 168  * -------------------------------------------------------------------------- */
 169
 170 #if defined(THREADED_RTS)
 171 STATIC_INLINE void
 172 newReturningTask (Capability *cap, Task *task)
 173 {
 174     ASSERT_LOCK_HELD(&cap->lock);
 175     ASSERT(task->next == NULL);
 176     if (cap->returning_tasks_hd) {
 177         ASSERT(cap->returning_tasks_tl->next == NULL);
 178         cap->returning_tasks_tl->next = task;
 179     } else {
 180         cap->returning_tasks_hd = task;
 181     }
 182     cap->returning_tasks_tl = task;
 183 }
 184
 185 STATIC_INLINE Task *
 186 popReturningTask (Capability *cap)
 187 {
 188     ASSERT_LOCK_HELD(&cap->lock);
 189     Task *task;
 190     task = cap->returning_tasks_hd;
 191     ASSERT(task);
 192     cap->returning_tasks_hd = task->next;
 193     if (!cap->returning_tasks_hd) {
 194         cap->returning_tasks_tl = NULL;
 195     }
 196     task->next = NULL;
 197     return task;
 198 }
 199 #endif
 200
 201 /* ----------------------------------------------------------------------------
 202  * Initialisation
 203  *
 204  * The Capability is initially marked not free.
 205  * ------------------------------------------------------------------------- */
 206
 207 static void
 208 initCapability( Capability *cap, nat i )
 209 {
 210     nat g;
 211
 212     cap->no = i;
 213     cap->in_haskell        = rtsFalse;
 214
 215     cap->run_queue_hd      = END_TSO_QUEUE;
 216     cap->run_queue_tl      = END_TSO_QUEUE;
 217
 218 #if defined(THREADED_RTS)
 219     initMutex(&cap->lock);
 220     cap->running_task      = NULL; // indicates cap is free
 221     cap->spare_workers     = NULL;
 222     cap->n_spare_workers   = 0;
 223     cap->suspended_ccalls  = NULL;
 224     cap->returning_tasks_hd = NULL;
 225     cap->returning_tasks_tl = NULL;
 226     cap->inbox              = (Message*)END_TSO_QUEUE;
 227     cap->sparks_created     = 0;
 228     cap->sparks_dud         = 0;
 229     cap->sparks_converted   = 0;
 230     cap->sparks_pruned      = 0;
 231 #endif
 232
 233     cap->f.stgEagerBlackholeInfo = (W_)&__stg_EAGER_BLACKHOLE_info;
 234     cap->f.stgGCEnter1     = (StgFunPtr)__stg_gc_enter_1;
 235     cap->f.stgGCFun        = (StgFunPtr)__stg_gc_fun;
 236
 237     cap->mut_lists  = stgMallocBytes(sizeof(bdescr *) *
 238                                      RtsFlags.GcFlags.generations,
 239                                      "initCapability");
 240     cap->saved_mut_lists = stgMallocBytes(sizeof(bdescr *) *
 241                                           RtsFlags.GcFlags.generations,
 242                                           "initCapability");
 243
 244     for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
 245         cap->mut_lists[g] = NULL;
 246     }
 247
 248     cap->free_tvar_watch_queues = END_STM_WATCH_QUEUE;
 249     cap->free_invariant_check_queues = END_INVARIANT_CHECK_QUEUE;
 250     cap->free_trec_chunks = END_STM_CHUNK_LIST;
 251     cap->free_trec_headers = NO_TREC;
 252     cap->transaction_tokens = 0;
 253     cap->context_switch = 0;
 254     cap->pinned_object_block = NULL;
 255 }
 256
 257 /* ---------------------------------------------------------------------------
 258  * Function:  initCapabilities()
 259  *
 260  * Purpose:   set up the Capability handling. For the THREADED_RTS build,
 261  *            we keep a table of them, the size of which is
 262  *            controlled by the user via the RTS flag -N.
 263  *
 264  * ------------------------------------------------------------------------- */
 265 void
 266 initCapabilities( void )
 267 {
 268 #if defined(THREADED_RTS)
 269     nat i;
 270
 271 #ifndef REG_Base
 272     // We can't support multiple CPUs if BaseReg is not a register
 273     if (RtsFlags.ParFlags.nNodes > 1) {
 274         errorBelch("warning: multiple CPUs not supported in this build, reverting to 1");
 275         RtsFlags.ParFlags.nNodes = 1;
 276     }
 277 #endif
 278
 279     n_capabilities = RtsFlags.ParFlags.nNodes;
 280
 281     if (n_capabilities == 1) {
 282         capabilities = &MainCapability;
 283         // THREADED_RTS must work on builds that don't have a mutable
 284         // BaseReg (eg. unregisterised), so in this case
 285         // capabilities[0] must coincide with &MainCapability.
 286     } else {
 287         capabilities = stgMallocBytes(n_capabilities * sizeof(Capability),
 288                                       "initCapabilities");
 289     }
 290
 291     for (i = 0; i < n_capabilities; i++) {
 292         initCapability(&capabilities[i], i);
 293     }
 294
 295     debugTrace(DEBUG_sched, "allocated %d capabilities", n_capabilities);
 296
 297 #else /* !THREADED_RTS */
 298
 299     n_capabilities = 1;
 300     capabilities = &MainCapability;
 301     initCapability(&MainCapability, 0);
 302
 303 #endif
 304
 305     // There are no free capabilities to begin with.  We will start
 306     // a worker Task to each Capability, which will quickly put the
 307     // Capability on the free list when it finds nothing to do.
 308     last_free_capability = &capabilities[0];
 309 }
 310
 311 /* ----------------------------------------------------------------------------
 312  * setContextSwitches: cause all capabilities to context switch as
 313  * soon as possible.
 314  * ------------------------------------------------------------------------- */
 315
 316 void setContextSwitches(void)
 317 {
 318     nat i;
 319     for (i=0; i < n_capabilities; i++) {
 320         contextSwitchCapability(&capabilities[i]);
 321     }
 322 }
 323
 324 /* ----------------------------------------------------------------------------
 325  * Give a Capability to a Task.  The task must currently be sleeping
 326  * on its condition variable.
 327  *
 328  * Requires cap->lock (modifies cap->running_task).
 329  *
 330  * When migrating a Task, the migrater must take task->lock before
 331  * modifying task->cap, to synchronise with the waking up Task.
 332  * Additionally, the migrater should own the Capability (when
 333  * migrating the run queue), or cap->lock (when migrating
 334  * returning_workers).
 335  *
 336  * ------------------------------------------------------------------------- */
 337
 338 #if defined(THREADED_RTS)
 339 STATIC_INLINE void
 340 giveCapabilityToTask (Capability *cap USED_IF_DEBUG, Task *task)
 341 {
 342     ASSERT_LOCK_HELD(&cap->lock);
 343     ASSERT(task->cap == cap);
 344     debugTrace(DEBUG_sched, "passing capability %d to %s %p",
 345                cap->no, task->incall->tso ? "bound task" : "worker",
 346                (void *)task->id);
 347     ACQUIRE_LOCK(&task->lock);
 348     task->wakeup = rtsTrue;
 349     // the wakeup flag is needed because signalCondition() doesn't
 350     // flag the condition if the thread is already runniing, but we want
 351     // it to be sticky.
 352     signalCondition(&task->cond);
 353     RELEASE_LOCK(&task->lock);
 354 }
 355 #endif
 356
 357 /* ----------------------------------------------------------------------------
 358  * Function:  releaseCapability(Capability*)
 359  *
 360  * Purpose:   Letting go of a capability. Causes a
 361  *            'returning worker' thread or a 'waiting worker'
 362  *            to wake up, in that order.
 363  * ------------------------------------------------------------------------- */
 364
 365 #if defined(THREADED_RTS)
 366 void
 367 releaseCapability_ (Capability* cap,
 368                     rtsBool always_wakeup)
 369 {
 370     Task *task;
 371
 372     task = cap->running_task;
 373
 374     ASSERT_PARTIAL_CAPABILITY_INVARIANTS(cap,task);
 375
 376     cap->running_task = NULL;
 377
 378     // Check to see whether a worker thread can be given
 379     // the go-ahead to return the result of an external call..
 380     if (cap->returning_tasks_hd != NULL) {
 381         giveCapabilityToTask(cap,cap->returning_tasks_hd);
 382         // The Task pops itself from the queue (see waitForReturnCapability())
 383         return;
 384     }
 385
 386     if (waiting_for_gc == PENDING_GC_SEQ) {
 387       last_free_capability = cap; // needed?
 388       debugTrace(DEBUG_sched, "GC pending, set capability %d free", cap->no);
 389       return;
 390     }
 391
 392
 393     // If the next thread on the run queue is a bound thread,
 394     // give this Capability to the appropriate Task.
 395     if (!emptyRunQueue(cap) && cap->run_queue_hd->bound) {
 396         // Make sure we're not about to try to wake ourselves up
 397         // ASSERT(task != cap->run_queue_hd->bound);
 398         // assertion is false: in schedule() we force a yield after
 399         // ThreadBlocked, but the thread may be back on the run queue
 400         // by now.
 401         task = cap->run_queue_hd->bound->task;
 402         giveCapabilityToTask(cap,task);
 403         return;
 404     }
 405
 406     if (!cap->spare_workers) {
 407         // Create a worker thread if we don't have one.  If the system
 408         // is interrupted, we only create a worker task if there
 409         // are threads that need to be completed.  If the system is
 410         // shutting down, we never create a new worker.
 411         if (sched_state < SCHED_SHUTTING_DOWN || !emptyRunQueue(cap)) {
 412             debugTrace(DEBUG_sched,
 413                        "starting new worker on capability %d", cap->no);
 414             startWorkerTask(cap);
 415             return;
 416         }
 417     }
 418
 419     // If we have an unbound thread on the run queue, or if there's
 420     // anything else to do, give the Capability to a worker thread.
 421     if (always_wakeup ||
 422         !emptyRunQueue(cap) || !emptyInbox(cap) ||
 423         !emptySparkPoolCap(cap) || globalWorkToDo()) {
 424         if (cap->spare_workers) {
 425             giveCapabilityToTask(cap,cap->spare_workers);
 426             // The worker Task pops itself from the queue;
 427             return;
 428         }
 429     }
 430
 431     last_free_capability = cap;
 432     debugTrace(DEBUG_sched, "freeing capability %d", cap->no);
 433 }
 434
 435 void
 436 releaseCapability (Capability* cap USED_IF_THREADS)
 437 {
 438     ACQUIRE_LOCK(&cap->lock);
 439     releaseCapability_(cap, rtsFalse);
 440     RELEASE_LOCK(&cap->lock);
 441 }
 442
 443 void
 444 releaseAndWakeupCapability (Capability* cap USED_IF_THREADS)
 445 {
 446     ACQUIRE_LOCK(&cap->lock);
 447     releaseCapability_(cap, rtsTrue);
 448     RELEASE_LOCK(&cap->lock);
 449 }
 450
 451 static void
 452 releaseCapabilityAndQueueWorker (Capability* cap USED_IF_THREADS)
 453 {
 454     Task *task;
 455
 456     ACQUIRE_LOCK(&cap->lock);
 457
 458     task = cap->running_task;
 459
 460     // If the Task is stopped, we shouldn't be yielding, we should
 461     // be just exiting.
 462     ASSERT(!task->stopped);
 463
 464     // If the current task is a worker, save it on the spare_workers
 465     // list of this Capability.  A worker can mark itself as stopped,
 466     // in which case it is not replaced on the spare_worker queue.
 467     // This happens when the system is shutting down (see
 468     // Schedule.c:workerStart()).
 469     if (!isBoundTask(task))
 470     {
 471         if (cap->n_spare_workers < MAX_SPARE_WORKERS)
 472         {
 473             task->next = cap->spare_workers;
 474             cap->spare_workers = task;
 475             cap->n_spare_workers++;
 476         }
 477         else
 478         {
 479             debugTrace(DEBUG_sched, "%d spare workers already, exiting",
 480                        cap->n_spare_workers);
 481             releaseCapability_(cap,rtsFalse);
 482             // hold the lock until after workerTaskStop; c.f. scheduleWorker()
 483             workerTaskStop(task);
 484             RELEASE_LOCK(&cap->lock);
 485             shutdownThread();
 486         }
 487     }
 488     // Bound tasks just float around attached to their TSOs.
 489
 490     releaseCapability_(cap,rtsFalse);
 491
 492     RELEASE_LOCK(&cap->lock);
 493 }
 494 #endif
 495
 496 /* ----------------------------------------------------------------------------
 497  * waitForReturnCapability( Task *task )
 498  *
 499  * Purpose:  when an OS thread returns from an external call,
 500  * it calls waitForReturnCapability() (via Schedule.resumeThread())
 501  * to wait for permission to enter the RTS & communicate the
 502  * result of the external call back to the Haskell thread that
 503  * made it.
 504  *
 505  * ------------------------------------------------------------------------- */
 506 void
 507 waitForReturnCapability (Capability **pCap, Task *task)
 508 {
 509 #if !defined(THREADED_RTS)
 510
 511     MainCapability.running_task = task;
 512     task->cap = &MainCapability;
 513     *pCap = &MainCapability;
 514
 515 #else
 516     Capability *cap = *pCap;
 517
 518     if (cap == NULL) {
 519         // Try last_free_capability first
 520         cap = last_free_capability;
 521         if (cap->running_task) {
 522             nat i;
 523             // otherwise, search for a free capability
 524             cap = NULL;
 525             for (i = 0; i < n_capabilities; i++) {
 526                 if (!capabilities[i].running_task) {
 527                     cap = &capabilities[i];
 528                     break;
 529                 }
 530             }
 531             if (cap == NULL) {
 532                 // Can't find a free one, use last_free_capability.
 533                 cap = last_free_capability;
 534             }
 535         }
 536
 537         // record the Capability as the one this Task is now assocated with.
 538         task->cap = cap;
 539
 540     } else {
 541         ASSERT(task->cap == cap);
 542     }
 543
 544     ACQUIRE_LOCK(&cap->lock);
 545
 546     debugTrace(DEBUG_sched, "returning; I want capability %d", cap->no);
 547
 548     if (!cap->running_task) {
 549         // It's free; just grab it
 550         cap->running_task = task;
 551         RELEASE_LOCK(&cap->lock);
 552     } else {
 553         newReturningTask(cap,task);
 554         RELEASE_LOCK(&cap->lock);
 555
 556         for (;;) {
 557             ACQUIRE_LOCK(&task->lock);
 558             // task->lock held, cap->lock not held
 559             if (!task->wakeup) waitCondition(&task->cond, &task->lock);
 560             cap = task->cap;
 561             task->wakeup = rtsFalse;
 562             RELEASE_LOCK(&task->lock);
 563
 564             // now check whether we should wake up...
 565             ACQUIRE_LOCK(&cap->lock);
 566             if (cap->running_task == NULL) {
 567                 if (cap->returning_tasks_hd != task) {
 568                     giveCapabilityToTask(cap,cap->returning_tasks_hd);
 569                     RELEASE_LOCK(&cap->lock);
 570                     continue;
 571                 }
 572                 cap->running_task = task;
 573                 popReturningTask(cap);
 574                 RELEASE_LOCK(&cap->lock);
 575                 break;
 576             }
 577             RELEASE_LOCK(&cap->lock);
 578         }
 579
 580     }
 581
 582     ASSERT_FULL_CAPABILITY_INVARIANTS(cap,task);
 583
 584     debugTrace(DEBUG_sched, "resuming capability %d", cap->no);
 585
 586     *pCap = cap;
 587 #endif
 588 }
 589
 590 #if defined(THREADED_RTS)
 591 /* ----------------------------------------------------------------------------
 592  * yieldCapability
 593  * ------------------------------------------------------------------------- */
 594
 595 void
 596 yieldCapability (Capability** pCap, Task *task)
 597 {
 598     Capability *cap = *pCap;
 599
 600     if (waiting_for_gc == PENDING_GC_PAR) {
 601         traceEventGcStart(cap);
 602         gcWorkerThread(cap);
 603         traceEventGcEnd(cap);
 604         return;
 605     }
 606
 607         debugTrace(DEBUG_sched, "giving up capability %d", cap->no);
 608
 609         // We must now release the capability and wait to be woken up
 610         // again.
 611         task->wakeup = rtsFalse;
 612         releaseCapabilityAndQueueWorker(cap);
 613
 614         for (;;) {
 615             ACQUIRE_LOCK(&task->lock);
 616             // task->lock held, cap->lock not held
 617             if (!task->wakeup) waitCondition(&task->cond, &task->lock);
 618             cap = task->cap;
 619             task->wakeup = rtsFalse;
 620             RELEASE_LOCK(&task->lock);
 621
 622             debugTrace(DEBUG_sched, "woken up on capability %d", cap->no);
 623
 624             ACQUIRE_LOCK(&cap->lock);
 625             if (cap->running_task != NULL) {
 626                 debugTrace(DEBUG_sched,
 627                            "capability %d is owned by another task", cap->no);
 628                 RELEASE_LOCK(&cap->lock);
 629                 continue;
 630             }
 631
 632             if (task->incall->tso == NULL) {
 633                 ASSERT(cap->spare_workers != NULL);
 634                 // if we're not at the front of the queue, release it
 635                 // again.  This is unlikely to happen.
 636                 if (cap->spare_workers != task) {
 637                     giveCapabilityToTask(cap,cap->spare_workers);
 638                     RELEASE_LOCK(&cap->lock);
 639                     continue;
 640                 }
 641                 cap->spare_workers = task->next;
 642                 task->next = NULL;
 643                 cap->n_spare_workers--;
 644             }
 645             cap->running_task = task;
 646             RELEASE_LOCK(&cap->lock);
 647             break;
 648         }
 649
 650         debugTrace(DEBUG_sched, "resuming capability %d", cap->no);
 651         ASSERT(cap->running_task == task);
 652
 653     *pCap = cap;
 654
 655     ASSERT_FULL_CAPABILITY_INVARIANTS(cap,task);
 656
 657     return;
 658 }
 659
 660 /* ----------------------------------------------------------------------------
 661  * prodCapability
 662  *
 663  * If a Capability is currently idle, wake up a Task on it.  Used to
 664  * get every Capability into the GC.
 665  * ------------------------------------------------------------------------- */
 666
 667 void
 668 prodCapability (Capability *cap, Task *task)
 669 {
 670     ACQUIRE_LOCK(&cap->lock);
 671     if (!cap->running_task) {
 672         cap->running_task = task;
 673         releaseCapability_(cap,rtsTrue);
 674     }
 675     RELEASE_LOCK(&cap->lock);
 676 }
 677
 678 /* ----------------------------------------------------------------------------
 679  * shutdownCapability
 680  *
 681  * At shutdown time, we want to let everything exit as cleanly as
 682  * possible.  For each capability, we let its run queue drain, and
 683  * allow the workers to stop.
 684  *
 685  * This function should be called when interrupted and
 686  * shutting_down_scheduler = rtsTrue, thus any worker that wakes up
 687  * will exit the scheduler and call taskStop(), and any bound thread
 688  * that wakes up will return to its caller.  Runnable threads are
 689  * killed.
 690  *
 691  * ------------------------------------------------------------------------- */
 692
 693 void
 694 shutdownCapability (Capability *cap, Task *task, rtsBool safe)
 695 {
 696     nat i;
 697
 698     task->cap = cap;
 699
 700     // Loop indefinitely until all the workers have exited and there
 701     // are no Haskell threads left.  We used to bail out after 50
 702     // iterations of this loop, but that occasionally left a worker
 703     // running which caused problems later (the closeMutex() below
 704     // isn't safe, for one thing).
 705
 706     for (i = 0; /* i < 50 */; i++) {
 707         ASSERT(sched_state == SCHED_SHUTTING_DOWN);
 708
 709         debugTrace(DEBUG_sched,
 710                    "shutting down capability %d, attempt %d", cap->no, i);
 711         ACQUIRE_LOCK(&cap->lock);
 712         if (cap->running_task) {
 713             RELEASE_LOCK(&cap->lock);
 714             debugTrace(DEBUG_sched, "not owner, yielding");
 715             yieldThread();
 716             continue;
 717         }
 718         cap->running_task = task;
 719
 720         if (cap->spare_workers) {
 721             // Look for workers that have died without removing
 722             // themselves from the list; this could happen if the OS
 723             // summarily killed the thread, for example.  This
 724             // actually happens on Windows when the system is
 725             // terminating the program, and the RTS is running in a
 726             // DLL.
 727             Task *t, *prev;
 728             prev = NULL;
 729             for (t = cap->spare_workers; t != NULL; t = t->next) {
 730                 if (!osThreadIsAlive(t->id)) {
 731                     debugTrace(DEBUG_sched,
 732                                "worker thread %p has died unexpectedly", (void *)t->id);
 733                     cap->n_spare_workers--;
 734                     if (!prev) {
 735                         cap->spare_workers = t->next;
 736                     } else {
 737                         prev->next = t->next;
 738                     }
 739                     prev = t;
 740                 }
 741             }
 742         }
 743
 744         if (!emptyRunQueue(cap) || cap->spare_workers) {
 745             debugTrace(DEBUG_sched,
 746                        "runnable threads or workers still alive, yielding");
 747             releaseCapability_(cap,rtsFalse); // this will wake up a worker
 748             RELEASE_LOCK(&cap->lock);
 749             yieldThread();
 750             continue;
 751         }
 752
 753         // If "safe", then busy-wait for any threads currently doing
 754         // foreign calls.  If we're about to unload this DLL, for
 755         // example, we need to be sure that there are no OS threads
 756         // that will try to return to code that has been unloaded.
 757         // We can be a bit more relaxed when this is a standalone
 758         // program that is about to terminate, and let safe=false.
 759         if (cap->suspended_ccalls && safe) {
 760             debugTrace(DEBUG_sched,
 761                        "thread(s) are involved in foreign calls, yielding");
 762             cap->running_task = NULL;
 763             RELEASE_LOCK(&cap->lock);
 764             // The IO manager thread might have been slow to start up,
 765             // so the first attempt to kill it might not have
 766             // succeeded.  Just in case, try again - the kill message
 767             // will only be sent once.
 768             //
 769             // To reproduce this deadlock: run ffi002(threaded1)
 770             // repeatedly on a loaded machine.
 771             ioManagerDie();
 772             yieldThread();
 773             continue;
 774         }
 775
 776         traceEventShutdown(cap);
 777         RELEASE_LOCK(&cap->lock);
 778         break;
 779     }
 780     // we now have the Capability, its run queue and spare workers
 781     // list are both empty.
 782
 783     // ToDo: we can't drop this mutex, because there might still be
 784     // threads performing foreign calls that will eventually try to
 785     // return via resumeThread() and attempt to grab cap->lock.
 786     // closeMutex(&cap->lock);
 787 }
 788
 789 /* ----------------------------------------------------------------------------
 790  * tryGrabCapability
 791  *
 792  * Attempt to gain control of a Capability if it is free.
 793  *
 794  * ------------------------------------------------------------------------- */
 795
 796 rtsBool
 797 tryGrabCapability (Capability *cap, Task *task)
 798 {
 799     if (cap->running_task != NULL) return rtsFalse;
 800     ACQUIRE_LOCK(&cap->lock);
 801     if (cap->running_task != NULL) {
 802         RELEASE_LOCK(&cap->lock);
 803         return rtsFalse;
 804     }
 805     task->cap = cap;
 806     cap->running_task = task;
 807     RELEASE_LOCK(&cap->lock);
 808     return rtsTrue;
 809 }
 810
 811
 812 #endif /* THREADED_RTS */
 813
 814 static void
 815 freeCapability (Capability *cap)
 816 {
 817     stgFree(cap->mut_lists);
 818     stgFree(cap->saved_mut_lists);
 819 #if defined(THREADED_RTS)
 820     freeSparkPool(cap->sparks);
 821 #endif
 822 }
 823
 824 void
 825 freeCapabilities (void)
 826 {
 827 #if defined(THREADED_RTS)
 828     nat i;
 829     for (i=0; i < n_capabilities; i++) {
 830         freeCapability(&capabilities[i]);
 831     }
 832 #else
 833     freeCapability(&MainCapability);
 834 #endif
 835 }
 836
 837 /* ---------------------------------------------------------------------------
 838    Mark everything directly reachable from the Capabilities.  When
 839    using multiple GC threads, each GC thread marks all Capabilities
 840    for which (c `mod` n == 0), for Capability c and thread n.
 841    ------------------------------------------------------------------------ */
 842
 843 void
 844 markSomeCapabilities (evac_fn evac, void *user, nat i0, nat delta,
 845                       rtsBool no_mark_sparks USED_IF_THREADS)
 846 {
 847     nat i;
 848     Capability *cap;
 849     InCall *incall;
 850
 851     // Each GC thread is responsible for following roots from the
 852     // Capability of the same number.  There will usually be the same
 853     // or fewer Capabilities as GC threads, but just in case there
 854     // are more, we mark every Capability whose number is the GC
 855     // thread's index plus a multiple of the number of GC threads.
 856     for (i = i0; i < n_capabilities; i += delta) {
 857         cap = &capabilities[i];
 858         evac(user, (StgClosure **)(void *)&cap->run_queue_hd);
 859         evac(user, (StgClosure **)(void *)&cap->run_queue_tl);
 860 #if defined(THREADED_RTS)
 861         evac(user, (StgClosure **)(void *)&cap->inbox);
 862 #endif
 863         for (incall = cap->suspended_ccalls; incall != NULL;
 864              incall=incall->next) {
 865             evac(user, (StgClosure **)(void *)&incall->suspended_tso);
 866         }
 867
 868 #if defined(THREADED_RTS)
 869         if (!no_mark_sparks) {
 870             traverseSparkQueue (evac, user, cap);
 871         }
 872 #endif
 873     }
 874
 875 #if !defined(THREADED_RTS)
 876     evac(user, (StgClosure **)(void *)&blocked_queue_hd);
 877     evac(user, (StgClosure **)(void *)&blocked_queue_tl);
 878     evac(user, (StgClosure **)(void *)&sleeping_queue);
 879 #endif
 880 }
 881
 882 void
 883 markCapabilities (evac_fn evac, void *user)
 884 {
 885     markSomeCapabilities(evac, user, 0, 1, rtsFalse);
 886 }
 887
 888 /* -----------------------------------------------------------------------------
 889    Messages
 890    -------------------------------------------------------------------------- */
 891