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     /* Declare a single capability set representing the process.
 272        Each capability will get added to this capset. */
 273     traceCapsetCreate(CAPSET_OSPROCESS_DEFAULT, CapsetTypeOsProcess);
 274
 275 #if defined(THREADED_RTS)
 276     nat i;
 277
 278 #ifndef REG_Base
 279     // We can't support multiple CPUs if BaseReg is not a register
 280     if (RtsFlags.ParFlags.nNodes > 1) {
 281         errorBelch("warning: multiple CPUs not supported in this build, reverting to 1");
 282         RtsFlags.ParFlags.nNodes = 1;
 283     }
 284 #endif
 285
 286     n_capabilities = RtsFlags.ParFlags.nNodes;
 287
 288     if (n_capabilities == 1) {
 289         capabilities = &MainCapability;
 290         // THREADED_RTS must work on builds that don't have a mutable
 291         // BaseReg (eg. unregisterised), so in this case
 292         // capabilities[0] must coincide with &MainCapability.
 293     } else {
 294         capabilities = stgMallocBytes(n_capabilities * sizeof(Capability),
 295                                       "initCapabilities");
 296     }
 297
 298     for (i = 0; i < n_capabilities; i++) {
 299         initCapability(&capabilities[i], i);
 300     }
 301
 302     debugTrace(DEBUG_sched, "allocated %d capabilities", n_capabilities);
 303
 304 #else /* !THREADED_RTS */
 305
 306     n_capabilities = 1;
 307     capabilities = &MainCapability;
 308     initCapability(&MainCapability, 0);
 309
 310 #endif
 311
 312     // There are no free capabilities to begin with.  We will start
 313     // a worker Task to each Capability, which will quickly put the
 314     // Capability on the free list when it finds nothing to do.
 315     last_free_capability = &capabilities[0];
 316 }
 317
 318 /* ----------------------------------------------------------------------------
 319  * setContextSwitches: cause all capabilities to context switch as
 320  * soon as possible.
 321  * ------------------------------------------------------------------------- */
 322
 323 void setContextSwitches(void)
 324 {
 325     nat i;
 326     for (i=0; i < n_capabilities; i++) {
 327         contextSwitchCapability(&capabilities[i]);
 328     }
 329 }
 330
 331 /* ----------------------------------------------------------------------------
 332  * Give a Capability to a Task.  The task must currently be sleeping
 333  * on its condition variable.
 334  *
 335  * Requires cap->lock (modifies cap->running_task).
 336  *
 337  * When migrating a Task, the migrater must take task->lock before
 338  * modifying task->cap, to synchronise with the waking up Task.
 339  * Additionally, the migrater should own the Capability (when
 340  * migrating the run queue), or cap->lock (when migrating
 341  * returning_workers).
 342  *
 343  * ------------------------------------------------------------------------- */
 344
 345 #if defined(THREADED_RTS)
 346 STATIC_INLINE void
 347 giveCapabilityToTask (Capability *cap USED_IF_DEBUG, Task *task)
 348 {
 349     ASSERT_LOCK_HELD(&cap->lock);
 350     ASSERT(task->cap == cap);
 351     debugTrace(DEBUG_sched, "passing capability %d to %s %p",
 352                cap->no, task->incall->tso ? "bound task" : "worker",
 353                (void *)task->id);
 354     ACQUIRE_LOCK(&task->lock);
 355     task->wakeup = rtsTrue;
 356     // the wakeup flag is needed because signalCondition() doesn't
 357     // flag the condition if the thread is already runniing, but we want
 358     // it to be sticky.
 359     signalCondition(&task->cond);
 360     RELEASE_LOCK(&task->lock);
 361 }
 362 #endif
 363
 364 /* ----------------------------------------------------------------------------
 365  * Function:  releaseCapability(Capability*)
 366  *
 367  * Purpose:   Letting go of a capability. Causes a
 368  *            'returning worker' thread or a 'waiting worker'
 369  *            to wake up, in that order.
 370  * ------------------------------------------------------------------------- */
 371
 372 #if defined(THREADED_RTS)
 373 void
 374 releaseCapability_ (Capability* cap,
 375                     rtsBool always_wakeup)
 376 {
 377     Task *task;
 378
 379     task = cap->running_task;
 380
 381     ASSERT_PARTIAL_CAPABILITY_INVARIANTS(cap,task);
 382
 383     cap->running_task = NULL;
 384
 385     // Check to see whether a worker thread can be given
 386     // the go-ahead to return the result of an external call..
 387     if (cap->returning_tasks_hd != NULL) {
 388         giveCapabilityToTask(cap,cap->returning_tasks_hd);
 389         // The Task pops itself from the queue (see waitForReturnCapability())
 390         return;
 391     }
 392
 393     if (waiting_for_gc == PENDING_GC_SEQ) {
 394       last_free_capability = cap; // needed?
 395       debugTrace(DEBUG_sched, "GC pending, set capability %d free", cap->no);
 396       return;
 397     }
 398
 399
 400     // If the next thread on the run queue is a bound thread,
 401     // give this Capability to the appropriate Task.
 402     if (!emptyRunQueue(cap) && cap->run_queue_hd->bound) {
 403         // Make sure we're not about to try to wake ourselves up
 404         // ASSERT(task != cap->run_queue_hd->bound);
 405         // assertion is false: in schedule() we force a yield after
 406         // ThreadBlocked, but the thread may be back on the run queue
 407         // by now.
 408         task = cap->run_queue_hd->bound->task;
 409         giveCapabilityToTask(cap,task);
 410         return;
 411     }
 412
 413     if (!cap->spare_workers) {
 414         // Create a worker thread if we don't have one.  If the system
 415         // is interrupted, we only create a worker task if there
 416         // are threads that need to be completed.  If the system is
 417         // shutting down, we never create a new worker.
 418         if (sched_state < SCHED_SHUTTING_DOWN || !emptyRunQueue(cap)) {
 419             debugTrace(DEBUG_sched,
 420                        "starting new worker on capability %d", cap->no);
 421             startWorkerTask(cap);
 422             return;
 423         }
 424     }
 425
 426     // If we have an unbound thread on the run queue, or if there's
 427     // anything else to do, give the Capability to a worker thread.
 428     if (always_wakeup ||
 429         !emptyRunQueue(cap) || !emptyInbox(cap) ||
 430         !emptySparkPoolCap(cap) || globalWorkToDo()) {
 431         if (cap->spare_workers) {
 432             giveCapabilityToTask(cap,cap->spare_workers);
 433             // The worker Task pops itself from the queue;
 434             return;
 435         }
 436     }
 437
 438     last_free_capability = cap;
 439     debugTrace(DEBUG_sched, "freeing capability %d", cap->no);
 440 }
 441
 442 void
 443 releaseCapability (Capability* cap USED_IF_THREADS)
 444 {
 445     ACQUIRE_LOCK(&cap->lock);
 446     releaseCapability_(cap, rtsFalse);
 447     RELEASE_LOCK(&cap->lock);
 448 }
 449
 450 void
 451 releaseAndWakeupCapability (Capability* cap USED_IF_THREADS)
 452 {
 453     ACQUIRE_LOCK(&cap->lock);
 454     releaseCapability_(cap, rtsTrue);
 455     RELEASE_LOCK(&cap->lock);
 456 }
 457
 458 static void
 459 releaseCapabilityAndQueueWorker (Capability* cap USED_IF_THREADS)
 460 {
 461     Task *task;
 462
 463     ACQUIRE_LOCK(&cap->lock);
 464
 465     task = cap->running_task;
 466
 467     // If the Task is stopped, we shouldn't be yielding, we should
 468     // be just exiting.
 469     ASSERT(!task->stopped);
 470
 471     // If the current task is a worker, save it on the spare_workers
 472     // list of this Capability.  A worker can mark itself as stopped,
 473     // in which case it is not replaced on the spare_worker queue.
 474     // This happens when the system is shutting down (see
 475     // Schedule.c:workerStart()).
 476     if (!isBoundTask(task))
 477     {
 478         if (cap->n_spare_workers < MAX_SPARE_WORKERS)
 479         {
 480             task->next = cap->spare_workers;
 481             cap->spare_workers = task;
 482             cap->n_spare_workers++;
 483         }
 484         else
 485         {
 486             debugTrace(DEBUG_sched, "%d spare workers already, exiting",
 487                        cap->n_spare_workers);
 488             releaseCapability_(cap,rtsFalse);
 489             // hold the lock until after workerTaskStop; c.f. scheduleWorker()
 490             workerTaskStop(task);
 491             RELEASE_LOCK(&cap->lock);
 492             shutdownThread();
 493         }
 494     }
 495     // Bound tasks just float around attached to their TSOs.
 496
 497     releaseCapability_(cap,rtsFalse);
 498
 499     RELEASE_LOCK(&cap->lock);
 500 }
 501 #endif
 502
 503 /* ----------------------------------------------------------------------------
 504  * waitForReturnCapability( Task *task )
 505  *
 506  * Purpose:  when an OS thread returns from an external call,
 507  * it calls waitForReturnCapability() (via Schedule.resumeThread())
 508  * to wait for permission to enter the RTS & communicate the
 509  * result of the external call back to the Haskell thread that
 510  * made it.
 511  *
 512  * ------------------------------------------------------------------------- */
 513 void
 514 waitForReturnCapability (Capability **pCap, Task *task)
 515 {
 516 #if !defined(THREADED_RTS)
 517
 518     MainCapability.running_task = task;
 519     task->cap = &MainCapability;
 520     *pCap = &MainCapability;
 521
 522 #else
 523     Capability *cap = *pCap;
 524
 525     if (cap == NULL) {
 526         // Try last_free_capability first
 527         cap = last_free_capability;
 528         if (cap->running_task) {
 529             nat i;
 530             // otherwise, search for a free capability
 531             cap = NULL;
 532             for (i = 0; i < n_capabilities; i++) {
 533                 if (!capabilities[i].running_task) {
 534                     cap = &capabilities[i];
 535                     break;
 536                 }
 537             }
 538             if (cap == NULL) {
 539                 // Can't find a free one, use last_free_capability.
 540                 cap = last_free_capability;
 541             }
 542         }
 543
 544         // record the Capability as the one this Task is now assocated with.
 545         task->cap = cap;
 546
 547     } else {
 548         ASSERT(task->cap == cap);
 549     }
 550
 551     ACQUIRE_LOCK(&cap->lock);
 552
 553     debugTrace(DEBUG_sched, "returning; I want capability %d", cap->no);
 554
 555     if (!cap->running_task) {
 556         // It's free; just grab it
 557         cap->running_task = task;
 558         RELEASE_LOCK(&cap->lock);
 559     } else {
 560         newReturningTask(cap,task);
 561         RELEASE_LOCK(&cap->lock);
 562
 563         for (;;) {
 564             ACQUIRE_LOCK(&task->lock);
 565             // task->lock held, cap->lock not held
 566             if (!task->wakeup) waitCondition(&task->cond, &task->lock);
 567             cap = task->cap;
 568             task->wakeup = rtsFalse;
 569             RELEASE_LOCK(&task->lock);
 570
 571             // now check whether we should wake up...
 572             ACQUIRE_LOCK(&cap->lock);
 573             if (cap->running_task == NULL) {
 574                 if (cap->returning_tasks_hd != task) {
 575                     giveCapabilityToTask(cap,cap->returning_tasks_hd);
 576                     RELEASE_LOCK(&cap->lock);
 577                     continue;
 578                 }
 579                 cap->running_task = task;
 580                 popReturningTask(cap);
 581                 RELEASE_LOCK(&cap->lock);
 582                 break;
 583             }
 584             RELEASE_LOCK(&cap->lock);
 585         }
 586
 587     }
 588
 589     ASSERT_FULL_CAPABILITY_INVARIANTS(cap,task);
 590
 591     debugTrace(DEBUG_sched, "resuming capability %d", cap->no);
 592
 593     *pCap = cap;
 594 #endif
 595 }
 596
 597 #if defined(THREADED_RTS)
 598 /* ----------------------------------------------------------------------------
 599  * yieldCapability
 600  * ------------------------------------------------------------------------- */
 601
 602 void
 603 yieldCapability (Capability** pCap, Task *task)
 604 {
 605     Capability *cap = *pCap;
 606
 607     if (waiting_for_gc == PENDING_GC_PAR) {
 608         traceEventGcStart(cap);
 609         gcWorkerThread(cap);
 610         traceEventGcEnd(cap);
 611         return;
 612     }
 613
 614         debugTrace(DEBUG_sched, "giving up capability %d", cap->no);
 615
 616         // We must now release the capability and wait to be woken up
 617         // again.
 618         task->wakeup = rtsFalse;
 619         releaseCapabilityAndQueueWorker(cap);
 620
 621         for (;;) {
 622             ACQUIRE_LOCK(&task->lock);
 623             // task->lock held, cap->lock not held
 624             if (!task->wakeup) waitCondition(&task->cond, &task->lock);
 625             cap = task->cap;
 626             task->wakeup = rtsFalse;
 627             RELEASE_LOCK(&task->lock);
 628
 629             debugTrace(DEBUG_sched, "woken up on capability %d", cap->no);
 630
 631             ACQUIRE_LOCK(&cap->lock);
 632             if (cap->running_task != NULL) {
 633                 debugTrace(DEBUG_sched,
 634                            "capability %d is owned by another task", cap->no);
 635                 RELEASE_LOCK(&cap->lock);
 636                 continue;
 637             }
 638
 639             if (task->incall->tso == NULL) {
 640                 ASSERT(cap->spare_workers != NULL);
 641                 // if we're not at the front of the queue, release it
 642                 // again.  This is unlikely to happen.
 643                 if (cap->spare_workers != task) {
 644                     giveCapabilityToTask(cap,cap->spare_workers);
 645                     RELEASE_LOCK(&cap->lock);
 646                     continue;
 647                 }
 648                 cap->spare_workers = task->next;
 649                 task->next = NULL;
 650                 cap->n_spare_workers--;
 651             }
 652             cap->running_task = task;
 653             RELEASE_LOCK(&cap->lock);
 654             break;
 655         }
 656
 657         debugTrace(DEBUG_sched, "resuming capability %d", cap->no);
 658         ASSERT(cap->running_task == task);
 659
 660     *pCap = cap;
 661
 662     ASSERT_FULL_CAPABILITY_INVARIANTS(cap,task);
 663
 664     return;
 665 }
 666
 667 /* ----------------------------------------------------------------------------
 668  * prodCapability
 669  *
 670  * If a Capability is currently idle, wake up a Task on it.  Used to
 671  * get every Capability into the GC.
 672  * ------------------------------------------------------------------------- */
 673
 674 void
 675 prodCapability (Capability *cap, Task *task)
 676 {
 677     ACQUIRE_LOCK(&cap->lock);
 678     if (!cap->running_task) {
 679         cap->running_task = task;
 680         releaseCapability_(cap,rtsTrue);
 681     }
 682     RELEASE_LOCK(&cap->lock);
 683 }
 684
 685 /* ----------------------------------------------------------------------------
 686  * tryGrabCapability
 687  *
 688  * Attempt to gain control of a Capability if it is free.
 689  *
 690  * ------------------------------------------------------------------------- */
 691
 692 rtsBool
 693 tryGrabCapability (Capability *cap, Task *task)
 694 {
 695     if (cap->running_task != NULL) return rtsFalse;
 696     ACQUIRE_LOCK(&cap->lock);
 697     if (cap->running_task != NULL) {
 698         RELEASE_LOCK(&cap->lock);
 699         return rtsFalse;
 700     }
 701     task->cap = cap;
 702     cap->running_task = task;
 703     RELEASE_LOCK(&cap->lock);
 704     return rtsTrue;
 705 }
 706
 707
 708 #endif /* THREADED_RTS */
 709
 710 /* ----------------------------------------------------------------------------
 711  * shutdownCapability
 712  *
 713  * At shutdown time, we want to let everything exit as cleanly as
 714  * possible.  For each capability, we let its run queue drain, and
 715  * allow the workers to stop.
 716  *
 717  * This function should be called when interrupted and
 718  * shutting_down_scheduler = rtsTrue, thus any worker that wakes up
 719  * will exit the scheduler and call taskStop(), and any bound thread
 720  * that wakes up will return to its caller.  Runnable threads are
 721  * killed.
 722  *
 723  * ------------------------------------------------------------------------- */
 724
 725 void
 726 shutdownCapability (Capability *cap,
 727                     Task *task USED_IF_THREADS,
 728                     rtsBool safe USED_IF_THREADS)
 729 {
 730 #if defined(THREADED_RTS)
 731     nat i;
 732
 733     task->cap = cap;
 734
 735     // Loop indefinitely until all the workers have exited and there
 736     // are no Haskell threads left.  We used to bail out after 50
 737     // iterations of this loop, but that occasionally left a worker
 738     // running which caused problems later (the closeMutex() below
 739     // isn't safe, for one thing).
 740
 741     for (i = 0; /* i < 50 */; i++) {
 742         ASSERT(sched_state == SCHED_SHUTTING_DOWN);
 743
 744         debugTrace(DEBUG_sched,
 745                    "shutting down capability %d, attempt %d", cap->no, i);
 746         ACQUIRE_LOCK(&cap->lock);
 747         if (cap->running_task) {
 748             RELEASE_LOCK(&cap->lock);
 749             debugTrace(DEBUG_sched, "not owner, yielding");
 750             yieldThread();
 751             continue;
 752         }
 753         cap->running_task = task;
 754
 755         if (cap->spare_workers) {
 756             // Look for workers that have died without removing
 757             // themselves from the list; this could happen if the OS
 758             // summarily killed the thread, for example.  This
 759             // actually happens on Windows when the system is
 760             // terminating the program, and the RTS is running in a
 761             // DLL.
 762             Task *t, *prev;
 763             prev = NULL;
 764             for (t = cap->spare_workers; t != NULL; t = t->next) {
 765                 if (!osThreadIsAlive(t->id)) {
 766                     debugTrace(DEBUG_sched,
 767                                "worker thread %p has died unexpectedly", (void *)t->id);
 768                     cap->n_spare_workers--;
 769                     if (!prev) {
 770                         cap->spare_workers = t->next;
 771                     } else {
 772                         prev->next = t->next;
 773                     }
 774                     prev = t;
 775                 }
 776             }
 777         }
 778
 779         if (!emptyRunQueue(cap) || cap->spare_workers) {
 780             debugTrace(DEBUG_sched,
 781                        "runnable threads or workers still alive, yielding");
 782             releaseCapability_(cap,rtsFalse); // this will wake up a worker
 783             RELEASE_LOCK(&cap->lock);
 784             yieldThread();
 785             continue;
 786         }
 787
 788         // If "safe", then busy-wait for any threads currently doing
 789         // foreign calls.  If we're about to unload this DLL, for
 790         // example, we need to be sure that there are no OS threads
 791         // that will try to return to code that has been unloaded.
 792         // We can be a bit more relaxed when this is a standalone
 793         // program that is about to terminate, and let safe=false.
 794         if (cap->suspended_ccalls && safe) {
 795             debugTrace(DEBUG_sched,
 796                        "thread(s) are involved in foreign calls, yielding");
 797             cap->running_task = NULL;
 798             RELEASE_LOCK(&cap->lock);
 799             // The IO manager thread might have been slow to start up,
 800             // so the first attempt to kill it might not have
 801             // succeeded.  Just in case, try again - the kill message
 802             // will only be sent once.
 803             //
 804             // To reproduce this deadlock: run ffi002(threaded1)
 805             // repeatedly on a loaded machine.
 806             ioManagerDie();
 807             yieldThread();
 808             continue;
 809         }
 810
 811         traceEventShutdown(cap);
 812         RELEASE_LOCK(&cap->lock);
 813         break;
 814     }
 815     // we now have the Capability, its run queue and spare workers
 816     // list are both empty.
 817
 818     // ToDo: we can't drop this mutex, because there might still be
 819     // threads performing foreign calls that will eventually try to
 820     // return via resumeThread() and attempt to grab cap->lock.
 821     // closeMutex(&cap->lock);
 822
 823 #endif /* THREADED_RTS */
 824
 825     traceCapsetRemoveCap(CAPSET_OSPROCESS_DEFAULT, cap->no);
 826 }
 827
 828 void
 829 shutdownCapabilities(Task *task, rtsBool safe)
 830 {
 831     nat i;
 832     for (i=0; i < n_capabilities; i++) {
 833         ASSERT(task->incall->tso == NULL);
 834         shutdownCapability(&capabilities[i], task, safe);
 835     }
 836     traceCapsetDelete(CAPSET_OSPROCESS_DEFAULT);
 837 }
 838
 839 static void
 840 freeCapability (Capability *cap)
 841 {
 842     stgFree(cap->mut_lists);
 843     stgFree(cap->saved_mut_lists);
 844 #if defined(THREADED_RTS)
 845     freeSparkPool(cap->sparks);
 846 #endif
 847 }
 848
 849 void
 850 freeCapabilities (void)
 851 {
 852 #if defined(THREADED_RTS)
 853     nat i;
 854     for (i=0; i < n_capabilities; i++) {
 855         freeCapability(&capabilities[i]);
 856     }
 857 #else
 858     freeCapability(&MainCapability);
 859 #endif
 860 }
 861
 862 /* ---------------------------------------------------------------------------
 863    Mark everything directly reachable from the Capabilities.  When
 864    using multiple GC threads, each GC thread marks all Capabilities
 865    for which (c `mod` n == 0), for Capability c and thread n.
 866    ------------------------------------------------------------------------ */
 867
 868 void
 869 markCapability (evac_fn evac, void *user, Capability *cap,
 870                 rtsBool no_mark_sparks USED_IF_THREADS)
 871 {
 872     InCall *incall;
 873
 874     // Each GC thread is responsible for following roots from the
 875     // Capability of the same number.  There will usually be the same
 876     // or fewer Capabilities as GC threads, but just in case there
 877     // are more, we mark every Capability whose number is the GC
 878     // thread's index plus a multiple of the number of GC threads.
 879     evac(user, (StgClosure **)(void *)&cap->run_queue_hd);
 880     evac(user, (StgClosure **)(void *)&cap->run_queue_tl);
 881 #if defined(THREADED_RTS)
 882     evac(user, (StgClosure **)(void *)&cap->inbox);
 883 #endif
 884     for (incall = cap->suspended_ccalls; incall != NULL;
 885          incall=incall->next) {
 886         evac(user, (StgClosure **)(void *)&incall->suspended_tso);
 887     }
 888
 889 #if defined(THREADED_RTS)
 890     if (!no_mark_sparks) {
 891         traverseSparkQueue (evac, user, cap);
 892     }
 893 #endif
 894
 895     // Free STM structures for this Capability
 896     stmPreGCHook(cap);
 897 }
 898
 899 void
 900 markCapabilities (evac_fn evac, void *user)
 901 {
 902     nat n;
 903     for (n = 0; n < n_capabilities; n++) {
 904         markCapability(evac, user, &capabilities[n], rtsFalse);
 905     }
 906 }