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