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_gcd         = 0;
 231     cap->sparks_fizzled     = 0;
 232 #endif
 233
 234     cap->f.stgEagerBlackholeInfo = (W_)&__stg_EAGER_BLACKHOLE_info;
 235     cap->f.stgGCEnter1     = (StgFunPtr)__stg_gc_enter_1;
 236     cap->f.stgGCFun        = (StgFunPtr)__stg_gc_fun;
 237
 238     cap->mut_lists  = stgMallocBytes(sizeof(bdescr *) *
 239                                      RtsFlags.GcFlags.generations,
 240                                      "initCapability");
 241     cap->saved_mut_lists = stgMallocBytes(sizeof(bdescr *) *
 242                                           RtsFlags.GcFlags.generations,
 243                                           "initCapability");
 244
 245     for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
 246         cap->mut_lists[g] = NULL;
 247     }
 248
 249     cap->free_tvar_watch_queues = END_STM_WATCH_QUEUE;
 250     cap->free_invariant_check_queues = END_INVARIANT_CHECK_QUEUE;
 251     cap->free_trec_chunks = END_STM_CHUNK_LIST;
 252     cap->free_trec_headers = NO_TREC;
 253     cap->transaction_tokens = 0;
 254     cap->context_switch = 0;
 255     cap->pinned_object_block = NULL;
 256 }
 257
 258 /* ---------------------------------------------------------------------------
 259  * Function:  initCapabilities()
 260  *
 261  * Purpose:   set up the Capability handling. For the THREADED_RTS build,
 262  *            we keep a table of them, the size of which is
 263  *            controlled by the user via the RTS flag -N.
 264  *
 265  * ------------------------------------------------------------------------- */
 266 void
 267 initCapabilities( void )
 268 {
 269 #if defined(THREADED_RTS)
 270     nat i;
 271
 272 #ifndef REG_Base
 273     // We can't support multiple CPUs if BaseReg is not a register
 274     if (RtsFlags.ParFlags.nNodes > 1) {
 275         errorBelch("warning: multiple CPUs not supported in this build, reverting to 1");
 276         RtsFlags.ParFlags.nNodes = 1;
 277     }
 278 #endif
 279
 280     n_capabilities = RtsFlags.ParFlags.nNodes;
 281
 282     if (n_capabilities == 1) {
 283         capabilities = &MainCapability;
 284         // THREADED_RTS must work on builds that don't have a mutable
 285         // BaseReg (eg. unregisterised), so in this case
 286         // capabilities[0] must coincide with &MainCapability.
 287     } else {
 288         capabilities = stgMallocBytes(n_capabilities * sizeof(Capability),
 289                                       "initCapabilities");
 290     }
 291
 292     for (i = 0; i < n_capabilities; i++) {
 293         initCapability(&capabilities[i], i);
 294     }
 295
 296     debugTrace(DEBUG_sched, "allocated %d capabilities", n_capabilities);
 297
 298 #else /* !THREADED_RTS */
 299
 300     n_capabilities = 1;
 301     capabilities = &MainCapability;
 302     initCapability(&MainCapability, 0);
 303
 304 #endif
 305
 306     // There are no free capabilities to begin with.  We will start
 307     // a worker Task to each Capability, which will quickly put the
 308     // Capability on the free list when it finds nothing to do.
 309     last_free_capability = &capabilities[0];
 310 }
 311
 312 /* ----------------------------------------------------------------------------
 313  * setContextSwitches: cause all capabilities to context switch as
 314  * soon as possible.
 315  * ------------------------------------------------------------------------- */
 316
 317 void setContextSwitches(void)
 318 {
 319     nat i;
 320     for (i=0; i < n_capabilities; i++) {
 321         contextSwitchCapability(&capabilities[i]);
 322     }
 323 }
 324
 325 /* ----------------------------------------------------------------------------
 326  * Give a Capability to a Task.  The task must currently be sleeping
 327  * on its condition variable.
 328  *
 329  * Requires cap->lock (modifies cap->running_task).
 330  *
 331  * When migrating a Task, the migrater must take task->lock before
 332  * modifying task->cap, to synchronise with the waking up Task.
 333  * Additionally, the migrater should own the Capability (when
 334  * migrating the run queue), or cap->lock (when migrating
 335  * returning_workers).
 336  *
 337  * ------------------------------------------------------------------------- */
 338
 339 #if defined(THREADED_RTS)
 340 STATIC_INLINE void
 341 giveCapabilityToTask (Capability *cap USED_IF_DEBUG, Task *task)
 342 {
 343     ASSERT_LOCK_HELD(&cap->lock);
 344     ASSERT(task->cap == cap);
 345     debugTrace(DEBUG_sched, "passing capability %d to %s %p",
 346                cap->no, task->incall->tso ? "bound task" : "worker",
 347                (void *)task->id);
 348     ACQUIRE_LOCK(&task->lock);
 349     task->wakeup = rtsTrue;
 350     // the wakeup flag is needed because signalCondition() doesn't
 351     // flag the condition if the thread is already runniing, but we want
 352     // it to be sticky.
 353     signalCondition(&task->cond);
 354     RELEASE_LOCK(&task->lock);
 355 }
 356 #endif
 357
 358 /* ----------------------------------------------------------------------------
 359  * Function:  releaseCapability(Capability*)
 360  *
 361  * Purpose:   Letting go of a capability. Causes a
 362  *            'returning worker' thread or a 'waiting worker'
 363  *            to wake up, in that order.
 364  * ------------------------------------------------------------------------- */
 365
 366 #if defined(THREADED_RTS)
 367 void
 368 releaseCapability_ (Capability* cap,
 369                     rtsBool always_wakeup)
 370 {
 371     Task *task;
 372
 373     task = cap->running_task;
 374
 375     ASSERT_PARTIAL_CAPABILITY_INVARIANTS(cap,task);
 376
 377     cap->running_task = NULL;
 378
 379     // Check to see whether a worker thread can be given
 380     // the go-ahead to return the result of an external call..
 381     if (cap->returning_tasks_hd != NULL) {
 382         giveCapabilityToTask(cap,cap->returning_tasks_hd);
 383         // The Task pops itself from the queue (see waitForReturnCapability())
 384         return;
 385     }
 386
 387     if (waiting_for_gc == PENDING_GC_SEQ) {
 388       last_free_capability = cap; // needed?
 389       debugTrace(DEBUG_sched, "GC pending, set capability %d free", cap->no);
 390       return;
 391     }
 392
 393
 394     // If the next thread on the run queue is a bound thread,
 395     // give this Capability to the appropriate Task.
 396     if (!emptyRunQueue(cap) && cap->run_queue_hd->bound) {
 397         // Make sure we're not about to try to wake ourselves up
 398         // ASSERT(task != cap->run_queue_hd->bound);
 399         // assertion is false: in schedule() we force a yield after
 400         // ThreadBlocked, but the thread may be back on the run queue
 401         // by now.
 402         task = cap->run_queue_hd->bound->task;
 403         giveCapabilityToTask(cap,task);
 404         return;
 405     }
 406
 407     if (!cap->spare_workers) {
 408         // Create a worker thread if we don't have one.  If the system
 409         // is interrupted, we only create a worker task if there
 410         // are threads that need to be completed.  If the system is
 411         // shutting down, we never create a new worker.
 412         if (sched_state < SCHED_SHUTTING_DOWN || !emptyRunQueue(cap)) {
 413             debugTrace(DEBUG_sched,
 414                        "starting new worker on capability %d", cap->no);
 415             startWorkerTask(cap);
 416             return;
 417         }
 418     }
 419
 420     // If we have an unbound thread on the run queue, or if there's
 421     // anything else to do, give the Capability to a worker thread.
 422     if (always_wakeup ||
 423         !emptyRunQueue(cap) || !emptyInbox(cap) ||
 424         !emptySparkPoolCap(cap) || globalWorkToDo()) {
 425         if (cap->spare_workers) {
 426             giveCapabilityToTask(cap,cap->spare_workers);
 427             // The worker Task pops itself from the queue;
 428             return;
 429         }
 430     }
 431
 432     last_free_capability = cap;
 433     debugTrace(DEBUG_sched, "freeing capability %d", cap->no);
 434 }
 435
 436 void
 437 releaseCapability (Capability* cap USED_IF_THREADS)
 438 {
 439     ACQUIRE_LOCK(&cap->lock);
 440     releaseCapability_(cap, rtsFalse);
 441     RELEASE_LOCK(&cap->lock);
 442 }
 443
 444 void
 445 releaseAndWakeupCapability (Capability* cap USED_IF_THREADS)
 446 {
 447     ACQUIRE_LOCK(&cap->lock);
 448     releaseCapability_(cap, rtsTrue);
 449     RELEASE_LOCK(&cap->lock);
 450 }
 451
 452 static void
 453 releaseCapabilityAndQueueWorker (Capability* cap USED_IF_THREADS)
 454 {
 455     Task *task;
 456
 457     ACQUIRE_LOCK(&cap->lock);
 458
 459     task = cap->running_task;
 460
 461     // If the Task is stopped, we shouldn't be yielding, we should
 462     // be just exiting.
 463     ASSERT(!task->stopped);
 464
 465     // If the current task is a worker, save it on the spare_workers
 466     // list of this Capability.  A worker can mark itself as stopped,
 467     // in which case it is not replaced on the spare_worker queue.
 468     // This happens when the system is shutting down (see
 469     // Schedule.c:workerStart()).
 470     if (!isBoundTask(task))
 471     {
 472         if (cap->n_spare_workers < MAX_SPARE_WORKERS)
 473         {
 474             task->next = cap->spare_workers;
 475             cap->spare_workers = task;
 476             cap->n_spare_workers++;
 477         }
 478         else
 479         {
 480             debugTrace(DEBUG_sched, "%d spare workers already, exiting",
 481                        cap->n_spare_workers);
 482             releaseCapability_(cap,rtsFalse);
 483             // hold the lock until after workerTaskStop; c.f. scheduleWorker()
 484             workerTaskStop(task);
 485             RELEASE_LOCK(&cap->lock);
 486             shutdownThread();
 487         }
 488     }
 489     // Bound tasks just float around attached to their TSOs.
 490
 491     releaseCapability_(cap,rtsFalse);
 492
 493     RELEASE_LOCK(&cap->lock);
 494 }
 495 #endif
 496
 497 /* ----------------------------------------------------------------------------
 498  * waitForReturnCapability( Task *task )
 499  *
 500  * Purpose:  when an OS thread returns from an external call,
 501  * it calls waitForReturnCapability() (via Schedule.resumeThread())
 502  * to wait for permission to enter the RTS & communicate the
 503  * result of the external call back to the Haskell thread that
 504  * made it.
 505  *
 506  * ------------------------------------------------------------------------- */
 507 void
 508 waitForReturnCapability (Capability **pCap, Task *task)
 509 {
 510 #if !defined(THREADED_RTS)
 511
 512     MainCapability.running_task = task;
 513     task->cap = &MainCapability;
 514     *pCap = &MainCapability;
 515
 516 #else
 517     Capability *cap = *pCap;
 518
 519     if (cap == NULL) {
 520         // Try last_free_capability first
 521         cap = last_free_capability;
 522         if (cap->running_task) {
 523             nat i;
 524             // otherwise, search for a free capability
 525             cap = NULL;
 526             for (i = 0; i < n_capabilities; i++) {
 527                 if (!capabilities[i].running_task) {
 528                     cap = &capabilities[i];
 529                     break;
 530                 }
 531             }
 532             if (cap == NULL) {
 533                 // Can't find a free one, use last_free_capability.
 534                 cap = last_free_capability;
 535             }
 536         }
 537
 538         // record the Capability as the one this Task is now assocated with.
 539         task->cap = cap;
 540
 541     } else {
 542         ASSERT(task->cap == cap);
 543     }
 544
 545     ACQUIRE_LOCK(&cap->lock);
 546
 547     debugTrace(DEBUG_sched, "returning; I want capability %d", cap->no);
 548
 549     if (!cap->running_task) {
 550         // It's free; just grab it
 551         cap->running_task = task;
 552         RELEASE_LOCK(&cap->lock);
 553     } else {
 554         newReturningTask(cap,task);
 555         RELEASE_LOCK(&cap->lock);
 556
 557         for (;;) {
 558             ACQUIRE_LOCK(&task->lock);
 559             // task->lock held, cap->lock not held
 560             if (!task->wakeup) waitCondition(&task->cond, &task->lock);
 561             cap = task->cap;
 562             task->wakeup = rtsFalse;
 563             RELEASE_LOCK(&task->lock);
 564
 565             // now check whether we should wake up...
 566             ACQUIRE_LOCK(&cap->lock);
 567             if (cap->running_task == NULL) {
 568                 if (cap->returning_tasks_hd != task) {
 569                     giveCapabilityToTask(cap,cap->returning_tasks_hd);
 570                     RELEASE_LOCK(&cap->lock);
 571                     continue;
 572                 }
 573                 cap->running_task = task;
 574                 popReturningTask(cap);
 575                 RELEASE_LOCK(&cap->lock);
 576                 break;
 577             }
 578             RELEASE_LOCK(&cap->lock);
 579         }
 580
 581     }
 582
 583     ASSERT_FULL_CAPABILITY_INVARIANTS(cap,task);
 584
 585     debugTrace(DEBUG_sched, "resuming capability %d", cap->no);
 586
 587     *pCap = cap;
 588 #endif
 589 }
 590
 591 #if defined(THREADED_RTS)
 592 /* ----------------------------------------------------------------------------
 593  * yieldCapability
 594  * ------------------------------------------------------------------------- */
 595
 596 void
 597 yieldCapability (Capability** pCap, Task *task)
 598 {
 599     Capability *cap = *pCap;
 600
 601     if (waiting_for_gc == PENDING_GC_PAR) {
 602         traceEventGcStart(cap);
 603         gcWorkerThread(cap);
 604         traceEventGcEnd(cap);
 605         return;
 606     }
 607
 608         debugTrace(DEBUG_sched, "giving up capability %d", cap->no);
 609
 610         // We must now release the capability and wait to be woken up
 611         // again.
 612         task->wakeup = rtsFalse;
 613         releaseCapabilityAndQueueWorker(cap);
 614
 615         for (;;) {
 616             ACQUIRE_LOCK(&task->lock);
 617             // task->lock held, cap->lock not held
 618             if (!task->wakeup) waitCondition(&task->cond, &task->lock);
 619             cap = task->cap;
 620             task->wakeup = rtsFalse;
 621             RELEASE_LOCK(&task->lock);
 622
 623             debugTrace(DEBUG_sched, "woken up on capability %d", cap->no);
 624
 625             ACQUIRE_LOCK(&cap->lock);
 626             if (cap->running_task != NULL) {
 627                 debugTrace(DEBUG_sched,
 628                            "capability %d is owned by another task", cap->no);
 629                 RELEASE_LOCK(&cap->lock);
 630                 continue;
 631             }
 632
 633             if (task->incall->tso == NULL) {
 634                 ASSERT(cap->spare_workers != NULL);
 635                 // if we're not at the front of the queue, release it
 636                 // again.  This is unlikely to happen.
 637                 if (cap->spare_workers != task) {
 638                     giveCapabilityToTask(cap,cap->spare_workers);
 639                     RELEASE_LOCK(&cap->lock);
 640                     continue;
 641                 }
 642                 cap->spare_workers = task->next;
 643                 task->next = NULL;
 644                 cap->n_spare_workers--;
 645             }
 646             cap->running_task = task;
 647             RELEASE_LOCK(&cap->lock);
 648             break;
 649         }
 650
 651         debugTrace(DEBUG_sched, "resuming capability %d", cap->no);
 652         ASSERT(cap->running_task == task);
 653
 654     *pCap = cap;
 655
 656     ASSERT_FULL_CAPABILITY_INVARIANTS(cap,task);
 657
 658     return;
 659 }
 660
 661 /* ----------------------------------------------------------------------------
 662  * prodCapability
 663  *
 664  * If a Capability is currently idle, wake up a Task on it.  Used to
 665  * get every Capability into the GC.
 666  * ------------------------------------------------------------------------- */
 667
 668 void
 669 prodCapability (Capability *cap, Task *task)
 670 {
 671     ACQUIRE_LOCK(&cap->lock);
 672     if (!cap->running_task) {
 673         cap->running_task = task;
 674         releaseCapability_(cap,rtsTrue);
 675     }
 676     RELEASE_LOCK(&cap->lock);
 677 }
 678
 679 /* ----------------------------------------------------------------------------
 680  * tryGrabCapability
 681  *
 682  * Attempt to gain control of a Capability if it is free.
 683  *
 684  * ------------------------------------------------------------------------- */
 685
 686 rtsBool
 687 tryGrabCapability (Capability *cap, Task *task)
 688 {
 689     if (cap->running_task != NULL) return rtsFalse;
 690     ACQUIRE_LOCK(&cap->lock);
 691     if (cap->running_task != NULL) {
 692         RELEASE_LOCK(&cap->lock);
 693         return rtsFalse;
 694     }
 695     task->cap = cap;
 696     cap->running_task = task;
 697     RELEASE_LOCK(&cap->lock);
 698     return rtsTrue;
 699 }
 700
 701
 702 #endif /* THREADED_RTS */
 703
 704 /* ----------------------------------------------------------------------------
 705  * shutdownCapability
 706  *
 707  * At shutdown time, we want to let everything exit as cleanly as
 708  * possible.  For each capability, we let its run queue drain, and
 709  * allow the workers to stop.
 710  *
 711  * This function should be called when interrupted and
 712  * shutting_down_scheduler = rtsTrue, thus any worker that wakes up
 713  * will exit the scheduler and call taskStop(), and any bound thread
 714  * that wakes up will return to its caller.  Runnable threads are
 715  * killed.
 716  *
 717  * ------------------------------------------------------------------------- */
 718
 719 void
 720 shutdownCapability (Capability *cap USED_IF_THREADS, Task *task USED_IF_THREADS, rtsBool safe USED_IF_THREADS)
 721 {
 722 #if defined(THREADED_RTS)
 723     nat i;
 724
 725     task->cap = cap;
 726
 727     // Loop indefinitely until all the workers have exited and there
 728     // are no Haskell threads left.  We used to bail out after 50
 729     // iterations of this loop, but that occasionally left a worker
 730     // running which caused problems later (the closeMutex() below
 731     // isn't safe, for one thing).
 732
 733     for (i = 0; /* i < 50 */; i++) {
 734         ASSERT(sched_state == SCHED_SHUTTING_DOWN);
 735
 736         debugTrace(DEBUG_sched,
 737                    "shutting down capability %d, attempt %d", cap->no, i);
 738         ACQUIRE_LOCK(&cap->lock);
 739         if (cap->running_task) {
 740             RELEASE_LOCK(&cap->lock);
 741             debugTrace(DEBUG_sched, "not owner, yielding");
 742             yieldThread();
 743             continue;
 744         }
 745         cap->running_task = task;
 746
 747         if (cap->spare_workers) {
 748             // Look for workers that have died without removing
 749             // themselves from the list; this could happen if the OS
 750             // summarily killed the thread, for example.  This
 751             // actually happens on Windows when the system is
 752             // terminating the program, and the RTS is running in a
 753             // DLL.
 754             Task *t, *prev;
 755             prev = NULL;
 756             for (t = cap->spare_workers; t != NULL; t = t->next) {
 757                 if (!osThreadIsAlive(t->id)) {
 758                     debugTrace(DEBUG_sched,
 759                                "worker thread %p has died unexpectedly", (void *)t->id);
 760                     cap->n_spare_workers--;
 761                     if (!prev) {
 762                         cap->spare_workers = t->next;
 763                     } else {
 764                         prev->next = t->next;
 765                     }
 766                     prev = t;
 767                 }
 768             }
 769         }
 770
 771         if (!emptyRunQueue(cap) || cap->spare_workers) {
 772             debugTrace(DEBUG_sched,
 773                        "runnable threads or workers still alive, yielding");
 774             releaseCapability_(cap,rtsFalse); // this will wake up a worker
 775             RELEASE_LOCK(&cap->lock);
 776             yieldThread();
 777             continue;
 778         }
 779
 780         // If "safe", then busy-wait for any threads currently doing
 781         // foreign calls.  If we're about to unload this DLL, for
 782         // example, we need to be sure that there are no OS threads
 783         // that will try to return to code that has been unloaded.
 784         // We can be a bit more relaxed when this is a standalone
 785         // program that is about to terminate, and let safe=false.
 786         if (cap->suspended_ccalls && safe) {
 787             debugTrace(DEBUG_sched,
 788                        "thread(s) are involved in foreign calls, yielding");
 789             cap->running_task = NULL;
 790             RELEASE_LOCK(&cap->lock);
 791             // The IO manager thread might have been slow to start up,
 792             // so the first attempt to kill it might not have
 793             // succeeded.  Just in case, try again - the kill message
 794             // will only be sent once.
 795             //
 796             // To reproduce this deadlock: run ffi002(threaded1)
 797             // repeatedly on a loaded machine.
 798             ioManagerDie();
 799             yieldThread();
 800             continue;
 801         }
 802
 803         traceEventShutdown(cap);
 804         RELEASE_LOCK(&cap->lock);
 805         break;
 806     }
 807     // we now have the Capability, its run queue and spare workers
 808     // list are both empty.
 809
 810     // ToDo: we can't drop this mutex, because there might still be
 811     // threads performing foreign calls that will eventually try to
 812     // return via resumeThread() and attempt to grab cap->lock.
 813     // closeMutex(&cap->lock);
 814
 815 #endif /* THREADED_RTS */
 816 }
 817
 818 void
 819 shutdownCapabilities(Task *task, rtsBool safe)
 820 {
 821     nat i;
 822     for (i=0; i < n_capabilities; i++) {
 823         ASSERT(task->incall->tso == NULL);
 824         shutdownCapability(&capabilities[i], task, safe);
 825     }
 826 }
 827
 828 static void
 829 freeCapability (Capability *cap)
 830 {
 831     stgFree(cap->mut_lists);
 832     stgFree(cap->saved_mut_lists);
 833 #if defined(THREADED_RTS)
 834     freeSparkPool(cap->sparks);
 835 #endif
 836 }
 837
 838 void
 839 freeCapabilities (void)
 840 {
 841 #if defined(THREADED_RTS)
 842     nat i;
 843     for (i=0; i < n_capabilities; i++) {
 844         freeCapability(&capabilities[i]);
 845     }
 846 #else
 847     freeCapability(&MainCapability);
 848 #endif
 849 }
 850
 851 /* ---------------------------------------------------------------------------
 852    Mark everything directly reachable from the Capabilities.  When
 853    using multiple GC threads, each GC thread marks all Capabilities
 854    for which (c `mod` n == 0), for Capability c and thread n.
 855    ------------------------------------------------------------------------ */
 856
 857 void
 858 markCapability (evac_fn evac, void *user, Capability *cap,
 859                 rtsBool no_mark_sparks USED_IF_THREADS)
 860 {
 861     InCall *incall;
 862
 863     // Each GC thread is responsible for following roots from the
 864     // Capability of the same number.  There will usually be the same
 865     // or fewer Capabilities as GC threads, but just in case there
 866     // are more, we mark every Capability whose number is the GC
 867     // thread's index plus a multiple of the number of GC threads.
 868     evac(user, (StgClosure **)(void *)&cap->run_queue_hd);
 869     evac(user, (StgClosure **)(void *)&cap->run_queue_tl);
 870 #if defined(THREADED_RTS)
 871     evac(user, (StgClosure **)(void *)&cap->inbox);
 872 #endif
 873     for (incall = cap->suspended_ccalls; incall != NULL;
 874          incall=incall->next) {
 875         evac(user, (StgClosure **)(void *)&incall->suspended_tso);
 876     }
 877
 878 #if defined(THREADED_RTS)
 879     if (!no_mark_sparks) {
 880         traverseSparkQueue (evac, user, cap);
 881     }
 882 #endif
 883
 884     // Free STM structures for this Capability
 885     stmPreGCHook(cap);
 886 }
 887
 888 void
 889 markCapabilities (evac_fn evac, void *user)
 890 {
 891     nat n;
 892     for (n = 0; n < n_capabilities; n++) {
 893         markCapability(evac, user, &capabilities[n], rtsFalse);
 894     }
 895 }