+/* ----------------------------------------------------------------------------
+ * Setting up the scheduler loop
+ * ------------------------------------------------------------------------- */
+
+static void
+schedulePreLoop(void)
+{
+#if defined(GRAN)
+ /* set up first event to get things going */
+ /* ToDo: assign costs for system setup and init MainTSO ! */
+ new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
+ ContinueThread,
+ CurrentTSO, (StgClosure*)NULL, (rtsSpark*)NULL);
+
+ IF_DEBUG(gran,
+ debugBelch("GRAN: Init CurrentTSO (in schedule) = %p\n",
+ CurrentTSO);
+ G_TSO(CurrentTSO, 5));
+
+ if (RtsFlags.GranFlags.Light) {
+ /* Save current time; GranSim Light only */
+ CurrentTSO->gran.clock = CurrentTime[CurrentProc];
+ }
+#endif
+}
+
+/* ----------------------------------------------------------------------------
+ * Start any pending signal handlers
+ * ------------------------------------------------------------------------- */
+
+static void
+scheduleStartSignalHandlers(void)
+{
+#if defined(RTS_USER_SIGNALS) && !defined(THREADED_RTS)
+ if (signals_pending()) { // safe outside the lock
+ startSignalHandlers();
+ }
+#endif
+}
+
+/* ----------------------------------------------------------------------------
+ * Check for blocked threads that can be woken up.
+ * ------------------------------------------------------------------------- */
+
+static void
+scheduleCheckBlockedThreads(Capability *cap USED_WHEN_NON_THREADED_RTS)
+{
+#if !defined(THREADED_RTS)
+ //
+ // Check whether any waiting threads need to be woken up. If the
+ // run queue is empty, and there are no other tasks running, we
+ // can wait indefinitely for something to happen.
+ //
+ if ( !emptyQueue(blocked_queue_hd) || !emptyQueue(sleeping_queue) )
+ {
+ awaitEvent( emptyRunQueue(cap) && !blackholes_need_checking );
+ }
+#endif
+}
+
+
+/* ----------------------------------------------------------------------------
+ * Check for threads blocked on BLACKHOLEs that can be woken up
+ * ------------------------------------------------------------------------- */
+static void
+scheduleCheckBlackHoles (Capability *cap)
+{
+ if ( blackholes_need_checking ) // check without the lock first
+ {
+ ACQUIRE_LOCK(&sched_mutex);
+ if ( blackholes_need_checking ) {
+ checkBlackHoles(cap);
+ blackholes_need_checking = rtsFalse;
+ }
+ RELEASE_LOCK(&sched_mutex);
+ }
+}
+
+/* ----------------------------------------------------------------------------
+ * Detect deadlock conditions and attempt to resolve them.
+ * ------------------------------------------------------------------------- */
+
+static void
+scheduleDetectDeadlock (Capability *cap, Task *task)
+{
+
+#if defined(PARALLEL_HASKELL)
+ // ToDo: add deadlock detection in GUM (similar to SMP) -- HWL
+ return;
+#endif
+
+ /*
+ * Detect deadlock: when we have no threads to run, there are no
+ * threads blocked, waiting for I/O, or sleeping, and all the
+ * other tasks are waiting for work, we must have a deadlock of
+ * some description.
+ */
+ if ( emptyThreadQueues(cap) )
+ {
+#if defined(THREADED_RTS)
+ /*
+ * In the threaded RTS, we only check for deadlock if there
+ * has been no activity in a complete timeslice. This means
+ * we won't eagerly start a full GC just because we don't have
+ * any threads to run currently.
+ */
+ if (recent_activity != ACTIVITY_INACTIVE) return;
+#endif
+
+ IF_DEBUG(scheduler, sched_belch("deadlocked, forcing major GC..."));
+
+ // Garbage collection can release some new threads due to
+ // either (a) finalizers or (b) threads resurrected because
+ // they are unreachable and will therefore be sent an
+ // exception. Any threads thus released will be immediately
+ // runnable.
+ scheduleDoGC( cap, task, rtsTrue/*force major GC*/ );
+ recent_activity = ACTIVITY_DONE_GC;
+
+ if ( !emptyRunQueue(cap) ) return;
+
+#if defined(RTS_USER_SIGNALS) && !defined(THREADED_RTS)
+ /* If we have user-installed signal handlers, then wait
+ * for signals to arrive rather then bombing out with a
+ * deadlock.
+ */
+ if ( anyUserHandlers() ) {
+ IF_DEBUG(scheduler,
+ sched_belch("still deadlocked, waiting for signals..."));
+
+ awaitUserSignals();
+
+ if (signals_pending()) {
+ startSignalHandlers();
+ }
+
+ // either we have threads to run, or we were interrupted:
+ ASSERT(!emptyRunQueue(cap) || interrupted);
+ }
+#endif
+
+#if !defined(THREADED_RTS)
+ /* Probably a real deadlock. Send the current main thread the
+ * Deadlock exception.
+ */
+ if (task->tso) {
+ switch (task->tso->why_blocked) {
+ case BlockedOnSTM:
+ case BlockedOnBlackHole:
+ case BlockedOnException:
+ case BlockedOnMVar:
+ raiseAsync(cap, task->tso, (StgClosure *)NonTermination_closure);
+ return;
+ default:
+ barf("deadlock: main thread blocked in a strange way");
+ }
+ }
+ return;
+#endif
+ }
+}
+
+/* ----------------------------------------------------------------------------
+ * Process an event (GRAN only)
+ * ------------------------------------------------------------------------- */
+
+#if defined(GRAN)
+static StgTSO *
+scheduleProcessEvent(rtsEvent *event)
+{
+ StgTSO *t;
+
+ if (RtsFlags.GranFlags.Light)
+ GranSimLight_enter_system(event, &ActiveTSO); // adjust ActiveTSO etc
+
+ /* adjust time based on time-stamp */
+ if (event->time > CurrentTime[CurrentProc] &&
+ event->evttype != ContinueThread)
+ CurrentTime[CurrentProc] = event->time;
+
+ /* Deal with the idle PEs (may issue FindWork or MoveSpark events) */
+ if (!RtsFlags.GranFlags.Light)
+ handleIdlePEs();
+
+ IF_DEBUG(gran, debugBelch("GRAN: switch by event-type\n"));
+
+ /* main event dispatcher in GranSim */
+ switch (event->evttype) {
+ /* Should just be continuing execution */
+ case ContinueThread:
+ IF_DEBUG(gran, debugBelch("GRAN: doing ContinueThread\n"));
+ /* ToDo: check assertion
+ ASSERT(run_queue_hd != (StgTSO*)NULL &&
+ run_queue_hd != END_TSO_QUEUE);
+ */
+ /* Ignore ContinueThreads for fetching threads (if synchr comm) */
+ if (!RtsFlags.GranFlags.DoAsyncFetch &&
+ procStatus[CurrentProc]==Fetching) {
+ debugBelch("ghuH: Spurious ContinueThread while Fetching ignored; TSO %d (%p) [PE %d]\n",
+ CurrentTSO->id, CurrentTSO, CurrentProc);
+ goto next_thread;
+ }
+ /* Ignore ContinueThreads for completed threads */
+ if (CurrentTSO->what_next == ThreadComplete) {
+ debugBelch("ghuH: found a ContinueThread event for completed thread %d (%p) [PE %d] (ignoring ContinueThread)\n",
+ CurrentTSO->id, CurrentTSO, CurrentProc);
+ goto next_thread;
+ }
+ /* Ignore ContinueThreads for threads that are being migrated */
+ if (PROCS(CurrentTSO)==Nowhere) {
+ debugBelch("ghuH: trying to run the migrating TSO %d (%p) [PE %d] (ignoring ContinueThread)\n",
+ CurrentTSO->id, CurrentTSO, CurrentProc);
+ goto next_thread;
+ }
+ /* The thread should be at the beginning of the run queue */
+ if (CurrentTSO!=run_queue_hds[CurrentProc]) {
+ debugBelch("ghuH: TSO %d (%p) [PE %d] is not at the start of the run_queue when doing a ContinueThread\n",
+ CurrentTSO->id, CurrentTSO, CurrentProc);
+ break; // run the thread anyway
+ }
+ /*
+ new_event(proc, proc, CurrentTime[proc],
+ FindWork,
+ (StgTSO*)NULL, (StgClosure*)NULL, (rtsSpark*)NULL);
+ goto next_thread;
+ */ /* Catches superfluous CONTINUEs -- should be unnecessary */
+ break; // now actually run the thread; DaH Qu'vam yImuHbej
+
+ case FetchNode:
+ do_the_fetchnode(event);
+ goto next_thread; /* handle next event in event queue */
+
+ case GlobalBlock:
+ do_the_globalblock(event);
+ goto next_thread; /* handle next event in event queue */
+
+ case FetchReply:
+ do_the_fetchreply(event);
+ goto next_thread; /* handle next event in event queue */
+
+ case UnblockThread: /* Move from the blocked queue to the tail of */
+ do_the_unblock(event);
+ goto next_thread; /* handle next event in event queue */
+
+ case ResumeThread: /* Move from the blocked queue to the tail of */
+ /* the runnable queue ( i.e. Qu' SImqa'lu') */
+ event->tso->gran.blocktime +=
+ CurrentTime[CurrentProc] - event->tso->gran.blockedat;
+ do_the_startthread(event);
+ goto next_thread; /* handle next event in event queue */
+
+ case StartThread:
+ do_the_startthread(event);
+ goto next_thread; /* handle next event in event queue */
+
+ case MoveThread:
+ do_the_movethread(event);
+ goto next_thread; /* handle next event in event queue */
+
+ case MoveSpark:
+ do_the_movespark(event);
+ goto next_thread; /* handle next event in event queue */
+
+ case FindWork:
+ do_the_findwork(event);
+ goto next_thread; /* handle next event in event queue */
+
+ default:
+ barf("Illegal event type %u\n", event->evttype);
+ } /* switch */
+
+ /* This point was scheduler_loop in the old RTS */
+
+ IF_DEBUG(gran, debugBelch("GRAN: after main switch\n"));
+
+ TimeOfLastEvent = CurrentTime[CurrentProc];
+ TimeOfNextEvent = get_time_of_next_event();
+ IgnoreEvents=(TimeOfNextEvent==0); // HWL HACK
+ // CurrentTSO = ThreadQueueHd;
+
+ IF_DEBUG(gran, debugBelch("GRAN: time of next event is: %ld\n",
+ TimeOfNextEvent));
+
+ if (RtsFlags.GranFlags.Light)
+ GranSimLight_leave_system(event, &ActiveTSO);
+
+ EndOfTimeSlice = CurrentTime[CurrentProc]+RtsFlags.GranFlags.time_slice;
+
+ IF_DEBUG(gran,
+ debugBelch("GRAN: end of time-slice is %#lx\n", EndOfTimeSlice));
+
+ /* in a GranSim setup the TSO stays on the run queue */
+ t = CurrentTSO;
+ /* Take a thread from the run queue. */
+ POP_RUN_QUEUE(t); // take_off_run_queue(t);
+
+ IF_DEBUG(gran,
+ debugBelch("GRAN: About to run current thread, which is\n");
+ G_TSO(t,5));
+
+ context_switch = 0; // turned on via GranYield, checking events and time slice
+
+ IF_DEBUG(gran,
+ DumpGranEvent(GR_SCHEDULE, t));
+
+ procStatus[CurrentProc] = Busy;
+}
+#endif // GRAN
+
+/* ----------------------------------------------------------------------------
+ * Send pending messages (PARALLEL_HASKELL only)
+ * ------------------------------------------------------------------------- */
+
+#if defined(PARALLEL_HASKELL)
+static StgTSO *
+scheduleSendPendingMessages(void)
+{
+ StgSparkPool *pool;
+ rtsSpark spark;
+ StgTSO *t;
+
+# if defined(PAR) // global Mem.Mgmt., omit for now
+ if (PendingFetches != END_BF_QUEUE) {
+ processFetches();
+ }
+# endif
+
+ if (RtsFlags.ParFlags.BufferTime) {
+ // if we use message buffering, we must send away all message
+ // packets which have become too old...
+ sendOldBuffers();
+ }
+}
+#endif
+
+/* ----------------------------------------------------------------------------
+ * Activate spark threads (PARALLEL_HASKELL only)
+ * ------------------------------------------------------------------------- */
+
+#if defined(PARALLEL_HASKELL)
+static void
+scheduleActivateSpark(void)
+{
+#if defined(SPARKS)
+ ASSERT(emptyRunQueue());
+/* We get here if the run queue is empty and want some work.
+ We try to turn a spark into a thread, and add it to the run queue,
+ from where it will be picked up in the next iteration of the scheduler
+ loop.
+*/
+
+ /* :-[ no local threads => look out for local sparks */
+ /* the spark pool for the current PE */
+ pool = &(cap.r.rSparks); // JB: cap = (old) MainCap
+ if (advisory_thread_count < RtsFlags.ParFlags.maxThreads &&
+ pool->hd < pool->tl) {
+ /*
+ * ToDo: add GC code check that we really have enough heap afterwards!!
+ * Old comment:
+ * If we're here (no runnable threads) and we have pending
+ * sparks, we must have a space problem. Get enough space
+ * to turn one of those pending sparks into a
+ * thread...
+ */
+
+ spark = findSpark(rtsFalse); /* get a spark */
+ if (spark != (rtsSpark) NULL) {
+ tso = createThreadFromSpark(spark); /* turn the spark into a thread */
+ IF_PAR_DEBUG(fish, // schedule,
+ debugBelch("==== schedule: Created TSO %d (%p); %d threads active\n",
+ tso->id, tso, advisory_thread_count));
+
+ if (tso==END_TSO_QUEUE) { /* failed to activate spark->back to loop */
+ IF_PAR_DEBUG(fish, // schedule,
+ debugBelch("==^^ failed to create thread from spark @ %lx\n",
+ spark));
+ return rtsFalse; /* failed to generate a thread */
+ } /* otherwise fall through & pick-up new tso */
+ } else {
+ IF_PAR_DEBUG(fish, // schedule,
+ debugBelch("==^^ no local sparks (spark pool contains only NFs: %d)\n",
+ spark_queue_len(pool)));
+ return rtsFalse; /* failed to generate a thread */
+ }
+ return rtsTrue; /* success in generating a thread */
+ } else { /* no more threads permitted or pool empty */
+ return rtsFalse; /* failed to generateThread */
+ }
+#else
+ tso = NULL; // avoid compiler warning only
+ return rtsFalse; /* dummy in non-PAR setup */
+#endif // SPARKS
+}
+#endif // PARALLEL_HASKELL
+
+/* ----------------------------------------------------------------------------
+ * Get work from a remote node (PARALLEL_HASKELL only)
+ * ------------------------------------------------------------------------- */
+
+#if defined(PARALLEL_HASKELL)
+static rtsBool
+scheduleGetRemoteWork(rtsBool *receivedFinish)
+{
+ ASSERT(emptyRunQueue());
+
+ if (RtsFlags.ParFlags.BufferTime) {
+ IF_PAR_DEBUG(verbose,
+ debugBelch("...send all pending data,"));
+ {
+ nat i;
+ for (i=1; i<=nPEs; i++)
+ sendImmediately(i); // send all messages away immediately
+ }
+ }
+# ifndef SPARKS
+ //++EDEN++ idle() , i.e. send all buffers, wait for work
+ // suppress fishing in EDEN... just look for incoming messages
+ // (blocking receive)
+ IF_PAR_DEBUG(verbose,
+ debugBelch("...wait for incoming messages...\n"));
+ *receivedFinish = processMessages(); // blocking receive...
+
+ // and reenter scheduling loop after having received something
+ // (return rtsFalse below)
+
+# else /* activate SPARKS machinery */
+/* We get here, if we have no work, tried to activate a local spark, but still
+ have no work. We try to get a remote spark, by sending a FISH message.
+ Thread migration should be added here, and triggered when a sequence of
+ fishes returns without work. */
+ delay = (RtsFlags.ParFlags.fishDelay!=0ll ? RtsFlags.ParFlags.fishDelay : 0ll);
+
+ /* =8-[ no local sparks => look for work on other PEs */
+ /*
+ * We really have absolutely no work. Send out a fish
+ * (there may be some out there already), and wait for
+ * something to arrive. We clearly can't run any threads
+ * until a SCHEDULE or RESUME arrives, and so that's what
+ * we're hoping to see. (Of course, we still have to
+ * respond to other types of messages.)
+ */
+ rtsTime now = msTime() /*CURRENT_TIME*/;
+ IF_PAR_DEBUG(verbose,
+ debugBelch("-- now=%ld\n", now));
+ IF_PAR_DEBUG(fish, // verbose,
+ if (outstandingFishes < RtsFlags.ParFlags.maxFishes &&
+ (last_fish_arrived_at!=0 &&
+ last_fish_arrived_at+delay > now)) {
+ debugBelch("--$$ <%llu> delaying FISH until %llu (last fish %llu, delay %llu)\n",
+ now, last_fish_arrived_at+delay,
+ last_fish_arrived_at,
+ delay);
+ });
+
+ if (outstandingFishes < RtsFlags.ParFlags.maxFishes &&
+ advisory_thread_count < RtsFlags.ParFlags.maxThreads) { // send a FISH, but when?
+ if (last_fish_arrived_at==0 ||
+ (last_fish_arrived_at+delay <= now)) { // send FISH now!
+ /* outstandingFishes is set in sendFish, processFish;
+ avoid flooding system with fishes via delay */
+ next_fish_to_send_at = 0;
+ } else {
+ /* ToDo: this should be done in the main scheduling loop to avoid the
+ busy wait here; not so bad if fish delay is very small */
+ int iq = 0; // DEBUGGING -- HWL
+ next_fish_to_send_at = last_fish_arrived_at+delay; // remember when to send
+ /* send a fish when ready, but process messages that arrive in the meantime */
+ do {
+ if (PacketsWaiting()) {
+ iq++; // DEBUGGING
+ *receivedFinish = processMessages();
+ }
+ now = msTime();
+ } while (!*receivedFinish || now<next_fish_to_send_at);
+ // JB: This means the fish could become obsolete, if we receive
+ // work. Better check for work again?
+ // last line: while (!receivedFinish || !haveWork || now<...)
+ // next line: if (receivedFinish || haveWork )
+
+ if (*receivedFinish) // no need to send a FISH if we are finishing anyway
+ return rtsFalse; // NB: this will leave scheduler loop
+ // immediately after return!
+
+ IF_PAR_DEBUG(fish, // verbose,
+ debugBelch("--$$ <%llu> sent delayed fish (%d processMessages); active/total threads=%d/%d\n",now,iq,run_queue_len(),advisory_thread_count));
+
+ }
+
+ // JB: IMHO, this should all be hidden inside sendFish(...)
+ /* pe = choosePE();
+ sendFish(pe, thisPE, NEW_FISH_AGE, NEW_FISH_HISTORY,
+ NEW_FISH_HUNGER);
+
+ // Global statistics: count no. of fishes
+ if (RtsFlags.ParFlags.ParStats.Global &&
+ RtsFlags.GcFlags.giveStats > NO_GC_STATS) {
+ globalParStats.tot_fish_mess++;
+ }
+ */
+
+ /* delayed fishes must have been sent by now! */
+ next_fish_to_send_at = 0;
+ }
+
+ *receivedFinish = processMessages();
+# endif /* SPARKS */
+
+ return rtsFalse;
+ /* NB: this function always returns rtsFalse, meaning the scheduler
+ loop continues with the next iteration;
+ rationale:
+ return code means success in finding work; we enter this function
+ if there is no local work, thus have to send a fish which takes
+ time until it arrives with work; in the meantime we should process
+ messages in the main loop;
+ */
+}
+#endif // PARALLEL_HASKELL
+
+/* ----------------------------------------------------------------------------
+ * PAR/GRAN: Report stats & debugging info(?)
+ * ------------------------------------------------------------------------- */
+
+#if defined(PAR) || defined(GRAN)
+static void
+scheduleGranParReport(void)
+{
+ ASSERT(run_queue_hd != END_TSO_QUEUE);
+
+ /* Take a thread from the run queue, if we have work */
+ POP_RUN_QUEUE(t); // take_off_run_queue(END_TSO_QUEUE);
+
+ /* If this TSO has got its outport closed in the meantime,
+ * it mustn't be run. Instead, we have to clean it up as if it was finished.
+ * It has to be marked as TH_DEAD for this purpose.
+ * If it is TH_TERM instead, it is supposed to have finished in the normal way.
+
+JB: TODO: investigate wether state change field could be nuked
+ entirely and replaced by the normal tso state (whatnext
+ field). All we want to do is to kill tsos from outside.
+ */
+
+ /* ToDo: write something to the log-file
+ if (RTSflags.ParFlags.granSimStats && !sameThread)
+ DumpGranEvent(GR_SCHEDULE, RunnableThreadsHd);
+
+ CurrentTSO = t;
+ */
+ /* the spark pool for the current PE */
+ pool = &(cap.r.rSparks); // cap = (old) MainCap
+
+ IF_DEBUG(scheduler,
+ debugBelch("--=^ %d threads, %d sparks on [%#x]\n",
+ run_queue_len(), spark_queue_len(pool), CURRENT_PROC));
+
+ IF_PAR_DEBUG(fish,
+ debugBelch("--=^ %d threads, %d sparks on [%#x]\n",
+ run_queue_len(), spark_queue_len(pool), CURRENT_PROC));
+
+ if (RtsFlags.ParFlags.ParStats.Full &&
+ (t->par.sparkname != (StgInt)0) && // only log spark generated threads
+ (emitSchedule || // forced emit
+ (t && LastTSO && t->id != LastTSO->id))) {
+ /*
+ we are running a different TSO, so write a schedule event to log file
+ NB: If we use fair scheduling we also have to write a deschedule
+ event for LastTSO; with unfair scheduling we know that the
+ previous tso has blocked whenever we switch to another tso, so
+ we don't need it in GUM for now
+ */
+ IF_PAR_DEBUG(fish, // schedule,
+ debugBelch("____ scheduling spark generated thread %d (%lx) (%lx) via a forced emit\n",t->id,t,t->par.sparkname));
+
+ DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
+ GR_SCHEDULE, t, (StgClosure *)NULL, 0, 0);
+ emitSchedule = rtsFalse;
+ }
+}
+#endif
+
+/* ----------------------------------------------------------------------------
+ * After running a thread...
+ * ------------------------------------------------------------------------- */
+
+static void
+schedulePostRunThread(void)
+{
+#if defined(PAR)
+ /* HACK 675: if the last thread didn't yield, make sure to print a
+ SCHEDULE event to the log file when StgRunning the next thread, even
+ if it is the same one as before */
+ LastTSO = t;
+ TimeOfLastYield = CURRENT_TIME;
+#endif
+
+ /* some statistics gathering in the parallel case */
+
+#if defined(GRAN) || defined(PAR) || defined(EDEN)
+ switch (ret) {
+ case HeapOverflow:
+# if defined(GRAN)
+ IF_DEBUG(gran, DumpGranEvent(GR_DESCHEDULE, t));
+ globalGranStats.tot_heapover++;
+# elif defined(PAR)
+ globalParStats.tot_heapover++;
+# endif
+ break;
+
+ case StackOverflow:
+# if defined(GRAN)
+ IF_DEBUG(gran,
+ DumpGranEvent(GR_DESCHEDULE, t));
+ globalGranStats.tot_stackover++;
+# elif defined(PAR)
+ // IF_DEBUG(par,
+ // DumpGranEvent(GR_DESCHEDULE, t);
+ globalParStats.tot_stackover++;
+# endif
+ break;
+
+ case ThreadYielding:
+# if defined(GRAN)
+ IF_DEBUG(gran,
+ DumpGranEvent(GR_DESCHEDULE, t));
+ globalGranStats.tot_yields++;
+# elif defined(PAR)
+ // IF_DEBUG(par,
+ // DumpGranEvent(GR_DESCHEDULE, t);
+ globalParStats.tot_yields++;
+# endif
+ break;
+
+ case ThreadBlocked:
+# if defined(GRAN)
+ IF_DEBUG(scheduler,
+ debugBelch("--<< thread %ld (%p; %s) stopped, blocking on node %p [PE %d] with BQ: ",
+ t->id, t, whatNext_strs[t->what_next], t->block_info.closure,
+ (t->block_info.closure==(StgClosure*)NULL ? 99 : where_is(t->block_info.closure)));
+ if (t->block_info.closure!=(StgClosure*)NULL)
+ print_bq(t->block_info.closure);
+ debugBelch("\n"));
+
+ // ??? needed; should emit block before
+ IF_DEBUG(gran,
+ DumpGranEvent(GR_DESCHEDULE, t));
+ prune_eventq(t, (StgClosure *)NULL); // prune ContinueThreads for t
+ /*
+ ngoq Dogh!
+ ASSERT(procStatus[CurrentProc]==Busy ||
+ ((procStatus[CurrentProc]==Fetching) &&
+ (t->block_info.closure!=(StgClosure*)NULL)));
+ if (run_queue_hds[CurrentProc] == END_TSO_QUEUE &&
+ !(!RtsFlags.GranFlags.DoAsyncFetch &&
+ procStatus[CurrentProc]==Fetching))
+ procStatus[CurrentProc] = Idle;
+ */
+# elif defined(PAR)
+//++PAR++ blockThread() writes the event (change?)
+# endif
+ break;
+
+ case ThreadFinished:
+ break;
+
+ default:
+ barf("parGlobalStats: unknown return code");
+ break;
+ }
+#endif
+}
+
+/* -----------------------------------------------------------------------------
+ * Handle a thread that returned to the scheduler with ThreadHeepOverflow
+ * -------------------------------------------------------------------------- */
+
+static rtsBool
+scheduleHandleHeapOverflow( Capability *cap, StgTSO *t )
+{
+ // did the task ask for a large block?
+ if (cap->r.rHpAlloc > BLOCK_SIZE) {
+ // if so, get one and push it on the front of the nursery.
+ bdescr *bd;
+ lnat blocks;
+
+ blocks = (lnat)BLOCK_ROUND_UP(cap->r.rHpAlloc) / BLOCK_SIZE;
+
+ IF_DEBUG(scheduler,
+ debugBelch("--<< thread %ld (%s) stopped: requesting a large block (size %ld)\n",
+ (long)t->id, whatNext_strs[t->what_next], blocks));
+
+ // don't do this if the nursery is (nearly) full, we'll GC first.
+ if (cap->r.rCurrentNursery->link != NULL ||
+ cap->r.rNursery->n_blocks == 1) { // paranoia to prevent infinite loop
+ // if the nursery has only one block.
+
+ ACQUIRE_SM_LOCK
+ bd = allocGroup( blocks );
+ RELEASE_SM_LOCK
+ cap->r.rNursery->n_blocks += blocks;
+
+ // link the new group into the list
+ bd->link = cap->r.rCurrentNursery;
+ bd->u.back = cap->r.rCurrentNursery->u.back;
+ if (cap->r.rCurrentNursery->u.back != NULL) {
+ cap->r.rCurrentNursery->u.back->link = bd;
+ } else {
+#if !defined(SMP)
+ ASSERT(g0s0->blocks == cap->r.rCurrentNursery &&
+ g0s0 == cap->r.rNursery);
+#endif
+ cap->r.rNursery->blocks = bd;
+ }
+ cap->r.rCurrentNursery->u.back = bd;
+
+ // initialise it as a nursery block. We initialise the
+ // step, gen_no, and flags field of *every* sub-block in
+ // this large block, because this is easier than making
+ // sure that we always find the block head of a large
+ // block whenever we call Bdescr() (eg. evacuate() and
+ // isAlive() in the GC would both have to do this, at
+ // least).
+ {
+ bdescr *x;
+ for (x = bd; x < bd + blocks; x++) {
+ x->step = cap->r.rNursery;
+ x->gen_no = 0;
+ x->flags = 0;
+ }
+ }
+
+ // This assert can be a killer if the app is doing lots
+ // of large block allocations.
+ IF_DEBUG(sanity, checkNurserySanity(cap->r.rNursery));
+
+ // now update the nursery to point to the new block
+ cap->r.rCurrentNursery = bd;
+
+ // we might be unlucky and have another thread get on the
+ // run queue before us and steal the large block, but in that
+ // case the thread will just end up requesting another large
+ // block.
+ pushOnRunQueue(cap,t);
+ return rtsFalse; /* not actually GC'ing */
+ }
+ }
+
+ IF_DEBUG(scheduler,
+ debugBelch("--<< thread %ld (%s) stopped: HeapOverflow\n",
+ (long)t->id, whatNext_strs[t->what_next]));
+#if defined(GRAN)
+ ASSERT(!is_on_queue(t,CurrentProc));
+#elif defined(PARALLEL_HASKELL)
+ /* Currently we emit a DESCHEDULE event before GC in GUM.
+ ToDo: either add separate event to distinguish SYSTEM time from rest
+ or just nuke this DESCHEDULE (and the following SCHEDULE) */
+ if (0 && RtsFlags.ParFlags.ParStats.Full) {
+ DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
+ GR_DESCHEDULE, t, (StgClosure *)NULL, 0, 0);
+ emitSchedule = rtsTrue;
+ }
+#endif
+
+ pushOnRunQueue(cap,t);
+ return rtsTrue;
+ /* actual GC is done at the end of the while loop in schedule() */
+}
+
+/* -----------------------------------------------------------------------------
+ * Handle a thread that returned to the scheduler with ThreadStackOverflow
+ * -------------------------------------------------------------------------- */
+
+static void
+scheduleHandleStackOverflow (Capability *cap, Task *task, StgTSO *t)
+{
+ IF_DEBUG(scheduler,debugBelch("--<< thread %ld (%s) stopped, StackOverflow\n",
+ (long)t->id, whatNext_strs[t->what_next]));
+ /* just adjust the stack for this thread, then pop it back
+ * on the run queue.
+ */
+ {
+ /* enlarge the stack */
+ StgTSO *new_t = threadStackOverflow(cap, t);
+
+ /* This TSO has moved, so update any pointers to it from the
+ * main thread stack. It better not be on any other queues...
+ * (it shouldn't be).
+ */
+ if (task->tso != NULL) {
+ task->tso = new_t;
+ }
+ pushOnRunQueue(cap,new_t);
+ }
+}
+
+/* -----------------------------------------------------------------------------
+ * Handle a thread that returned to the scheduler with ThreadYielding
+ * -------------------------------------------------------------------------- */
+
+static rtsBool
+scheduleHandleYield( Capability *cap, StgTSO *t, nat prev_what_next )
+{
+ // Reset the context switch flag. We don't do this just before
+ // running the thread, because that would mean we would lose ticks
+ // during GC, which can lead to unfair scheduling (a thread hogs
+ // the CPU because the tick always arrives during GC). This way
+ // penalises threads that do a lot of allocation, but that seems
+ // better than the alternative.
+ context_switch = 0;
+
+ /* put the thread back on the run queue. Then, if we're ready to
+ * GC, check whether this is the last task to stop. If so, wake
+ * up the GC thread. getThread will block during a GC until the
+ * GC is finished.
+ */
+ IF_DEBUG(scheduler,
+ if (t->what_next != prev_what_next) {
+ debugBelch("--<< thread %ld (%s) stopped to switch evaluators\n",
+ (long)t->id, whatNext_strs[t->what_next]);
+ } else {
+ debugBelch("--<< thread %ld (%s) stopped, yielding\n",
+ (long)t->id, whatNext_strs[t->what_next]);
+ }
+ );
+
+ IF_DEBUG(sanity,
+ //debugBelch("&& Doing sanity check on yielding TSO %ld.", t->id);
+ checkTSO(t));
+ ASSERT(t->link == END_TSO_QUEUE);
+
+ // Shortcut if we're just switching evaluators: don't bother
+ // doing stack squeezing (which can be expensive), just run the
+ // thread.
+ if (t->what_next != prev_what_next) {
+ return rtsTrue;
+ }
+
+#if defined(GRAN)
+ ASSERT(!is_on_queue(t,CurrentProc));
+
+ IF_DEBUG(sanity,
+ //debugBelch("&& Doing sanity check on all ThreadQueues (and their TSOs).");
+ checkThreadQsSanity(rtsTrue));
+
+#endif
+
+ addToRunQueue(cap,t);
+
+#if defined(GRAN)
+ /* add a ContinueThread event to actually process the thread */
+ new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
+ ContinueThread,
+ t, (StgClosure*)NULL, (rtsSpark*)NULL);
+ IF_GRAN_DEBUG(bq,
+ debugBelch("GRAN: eventq and runnableq after adding yielded thread to queue again:\n");
+ G_EVENTQ(0);
+ G_CURR_THREADQ(0));
+#endif
+ return rtsFalse;
+}
+
+/* -----------------------------------------------------------------------------
+ * Handle a thread that returned to the scheduler with ThreadBlocked
+ * -------------------------------------------------------------------------- */
+
+static void
+scheduleHandleThreadBlocked( StgTSO *t
+#if !defined(GRAN) && !defined(DEBUG)
+ STG_UNUSED
+#endif
+ )
+{
+#if defined(GRAN)
+ IF_DEBUG(scheduler,
+ debugBelch("--<< thread %ld (%p; %s) stopped, blocking on node %p [PE %d] with BQ: \n",
+ t->id, t, whatNext_strs[t->what_next], t->block_info.closure, (t->block_info.closure==(StgClosure*)NULL ? 99 : where_is(t->block_info.closure)));
+ if (t->block_info.closure!=(StgClosure*)NULL) print_bq(t->block_info.closure));
+
+ // ??? needed; should emit block before
+ IF_DEBUG(gran,
+ DumpGranEvent(GR_DESCHEDULE, t));
+ prune_eventq(t, (StgClosure *)NULL); // prune ContinueThreads for t
+ /*
+ ngoq Dogh!
+ ASSERT(procStatus[CurrentProc]==Busy ||
+ ((procStatus[CurrentProc]==Fetching) &&
+ (t->block_info.closure!=(StgClosure*)NULL)));
+ if (run_queue_hds[CurrentProc] == END_TSO_QUEUE &&
+ !(!RtsFlags.GranFlags.DoAsyncFetch &&
+ procStatus[CurrentProc]==Fetching))
+ procStatus[CurrentProc] = Idle;
+ */
+#elif defined(PAR)
+ IF_DEBUG(scheduler,
+ debugBelch("--<< thread %ld (%p; %s) stopped, blocking on node %p with BQ: \n",
+ t->id, t, whatNext_strs[t->what_next], t->block_info.closure));
+ IF_PAR_DEBUG(bq,
+
+ if (t->block_info.closure!=(StgClosure*)NULL)
+ print_bq(t->block_info.closure));
+
+ /* Send a fetch (if BlockedOnGA) and dump event to log file */
+ blockThread(t);
+
+ /* whatever we schedule next, we must log that schedule */
+ emitSchedule = rtsTrue;
+
+#else /* !GRAN */
+
+ // We don't need to do anything. The thread is blocked, and it
+ // has tidied up its stack and placed itself on whatever queue
+ // it needs to be on.
+
+#if !defined(SMP)
+ ASSERT(t->why_blocked != NotBlocked);
+ // This might not be true under SMP: we don't have
+ // exclusive access to this TSO, so someone might have
+ // woken it up by now. This actually happens: try
+ // conc023 +RTS -N2.
+#endif
+
+ IF_DEBUG(scheduler,
+ debugBelch("--<< thread %d (%s) stopped: ",
+ t->id, whatNext_strs[t->what_next]);
+ printThreadBlockage(t);
+ debugBelch("\n"));
+
+ /* Only for dumping event to log file
+ ToDo: do I need this in GranSim, too?
+ blockThread(t);
+ */
+#endif
+}
+
+/* -----------------------------------------------------------------------------
+ * Handle a thread that returned to the scheduler with ThreadFinished
+ * -------------------------------------------------------------------------- */
+
+static rtsBool
+scheduleHandleThreadFinished (Capability *cap STG_UNUSED, Task *task, StgTSO *t)
+{
+ /* Need to check whether this was a main thread, and if so,
+ * return with the return value.
+ *
+ * We also end up here if the thread kills itself with an
+ * uncaught exception, see Exception.cmm.
+ */
+ IF_DEBUG(scheduler,debugBelch("--++ thread %d (%s) finished\n",
+ t->id, whatNext_strs[t->what_next]));
+
+#if defined(GRAN)
+ endThread(t, CurrentProc); // clean-up the thread
+#elif defined(PARALLEL_HASKELL)
+ /* For now all are advisory -- HWL */
+ //if(t->priority==AdvisoryPriority) ??
+ advisory_thread_count--; // JB: Caution with this counter, buggy!
+
+# if defined(DIST)
+ if(t->dist.priority==RevalPriority)
+ FinishReval(t);
+# endif
+
+# if defined(EDENOLD)
+ // the thread could still have an outport... (BUG)
+ if (t->eden.outport != -1) {
+ // delete the outport for the tso which has finished...
+ IF_PAR_DEBUG(eden_ports,
+ debugBelch("WARNING: Scheduler removes outport %d for TSO %d.\n",
+ t->eden.outport, t->id));
+ deleteOPT(t);
+ }
+ // thread still in the process (HEAVY BUG! since outport has just been closed...)
+ if (t->eden.epid != -1) {
+ IF_PAR_DEBUG(eden_ports,
+ debugBelch("WARNING: Scheduler removes TSO %d from process %d .\n",
+ t->id, t->eden.epid));
+ removeTSOfromProcess(t);
+ }
+# endif
+
+# if defined(PAR)
+ if (RtsFlags.ParFlags.ParStats.Full &&
+ !RtsFlags.ParFlags.ParStats.Suppressed)
+ DumpEndEvent(CURRENT_PROC, t, rtsFalse /* not mandatory */);
+
+ // t->par only contains statistics: left out for now...
+ IF_PAR_DEBUG(fish,
+ debugBelch("**** end thread: ended sparked thread %d (%lx); sparkname: %lx\n",
+ t->id,t,t->par.sparkname));
+# endif
+#endif // PARALLEL_HASKELL
+
+ //
+ // Check whether the thread that just completed was a bound
+ // thread, and if so return with the result.
+ //
+ // There is an assumption here that all thread completion goes
+ // through this point; we need to make sure that if a thread
+ // ends up in the ThreadKilled state, that it stays on the run
+ // queue so it can be dealt with here.
+ //
+
+ if (t->bound) {
+
+ if (t->bound != task) {
+#if !defined(THREADED_RTS)
+ // Must be a bound thread that is not the topmost one. Leave
+ // it on the run queue until the stack has unwound to the
+ // point where we can deal with this. Leaving it on the run
+ // queue also ensures that the garbage collector knows about
+ // this thread and its return value (it gets dropped from the
+ // all_threads list so there's no other way to find it).
+ appendToRunQueue(cap,t);
+ return rtsFalse;
+#else
+ // this cannot happen in the threaded RTS, because a
+ // bound thread can only be run by the appropriate Task.
+ barf("finished bound thread that isn't mine");
+#endif
+ }
+
+ ASSERT(task->tso == t);
+
+ if (t->what_next == ThreadComplete) {
+ if (task->ret) {
+ // NOTE: return val is tso->sp[1] (see StgStartup.hc)
+ *(task->ret) = (StgClosure *)task->tso->sp[1];
+ }
+ task->stat = Success;
+ } else {
+ if (task->ret) {
+ *(task->ret) = NULL;
+ }
+ if (interrupted) {
+ task->stat = Interrupted;
+ } else {
+ task->stat = Killed;
+ }
+ }
+#ifdef DEBUG
+ removeThreadLabel((StgWord)task->tso->id);
+#endif
+ return rtsTrue; // tells schedule() to return
+ }
+
+ return rtsFalse;
+}
+
+/* -----------------------------------------------------------------------------
+ * Perform a heap census, if PROFILING
+ * -------------------------------------------------------------------------- */
+
+static rtsBool
+scheduleDoHeapProfile( rtsBool ready_to_gc STG_UNUSED )
+{
+#if defined(PROFILING)
+ // When we have +RTS -i0 and we're heap profiling, do a census at
+ // every GC. This lets us get repeatable runs for debugging.
+ if (performHeapProfile ||
+ (RtsFlags.ProfFlags.profileInterval==0 &&
+ RtsFlags.ProfFlags.doHeapProfile && ready_to_gc)) {
+ GarbageCollect(GetRoots, rtsTrue);
+ heapCensus();
+ performHeapProfile = rtsFalse;
+ return rtsTrue; // true <=> we already GC'd
+ }
+#endif
+ return rtsFalse;
+}
+
+/* -----------------------------------------------------------------------------
+ * Perform a garbage collection if necessary
+ * -------------------------------------------------------------------------- */
+
+static void
+scheduleDoGC( Capability *cap, Task *task USED_WHEN_SMP, rtsBool force_major )
+{
+ StgTSO *t;
+#ifdef SMP
+ static volatile StgWord waiting_for_gc;
+ rtsBool was_waiting;
+ nat i;
+#endif
+
+#ifdef SMP
+ // In order to GC, there must be no threads running Haskell code.
+ // Therefore, the GC thread needs to hold *all* the capabilities,
+ // and release them after the GC has completed.
+ //
+ // This seems to be the simplest way: previous attempts involved
+ // making all the threads with capabilities give up their
+ // capabilities and sleep except for the *last* one, which
+ // actually did the GC. But it's quite hard to arrange for all
+ // the other tasks to sleep and stay asleep.
+ //
+
+ was_waiting = cas(&waiting_for_gc, 0, 1);
+ if (was_waiting) return;
+
+ for (i=0; i < n_capabilities; i++) {
+ IF_DEBUG(scheduler, sched_belch("ready_to_gc, grabbing all the capabilies (%d/%d)", i, n_capabilities));
+ if (cap != &capabilities[i]) {
+ Capability *pcap = &capabilities[i];
+ // we better hope this task doesn't get migrated to
+ // another Capability while we're waiting for this one.
+ // It won't, because load balancing happens while we have
+ // all the Capabilities, but even so it's a slightly
+ // unsavoury invariant.
+ task->cap = pcap;
+ waitForReturnCapability(&pcap, task);
+ if (pcap != &capabilities[i]) {
+ barf("scheduleDoGC: got the wrong capability");
+ }
+ }
+ }
+
+ waiting_for_gc = rtsFalse;
+#endif
+
+ /* Kick any transactions which are invalid back to their
+ * atomically frames. When next scheduled they will try to
+ * commit, this commit will fail and they will retry.
+ */
+ {
+ StgTSO *next;
+
+ for (t = all_threads; t != END_TSO_QUEUE; t = next) {
+ if (t->what_next == ThreadRelocated) {
+ next = t->link;
+ } else {
+ next = t->global_link;
+ if (t -> trec != NO_TREC && t -> why_blocked == NotBlocked) {
+ if (!stmValidateNestOfTransactions (t -> trec)) {
+ IF_DEBUG(stm, sched_belch("trec %p found wasting its time", t));
+
+ // strip the stack back to the
+ // ATOMICALLY_FRAME, aborting the (nested)
+ // transaction, and saving the stack of any
+ // partially-evaluated thunks on the heap.
+ raiseAsync_(cap, t, NULL, rtsTrue);
+
+#ifdef REG_R1
+ ASSERT(get_itbl((StgClosure *)t->sp)->type == ATOMICALLY_FRAME);
+#endif
+ }
+ }
+ }
+ }
+ }
+
+ // so this happens periodically:
+ scheduleCheckBlackHoles(cap);
+
+ IF_DEBUG(scheduler, printAllThreads());
+
+ /* everybody back, start the GC.
+ * Could do it in this thread, or signal a condition var
+ * to do it in another thread. Either way, we need to
+ * broadcast on gc_pending_cond afterward.
+ */
+#if defined(THREADED_RTS)
+ IF_DEBUG(scheduler,sched_belch("doing GC"));
+#endif
+ GarbageCollect(GetRoots, force_major);
+
+#if defined(SMP)
+ // release our stash of capabilities.
+ for (i = 0; i < n_capabilities; i++) {
+ if (cap != &capabilities[i]) {
+ task->cap = &capabilities[i];
+ releaseCapability(&capabilities[i]);
+ }
+ }
+ task->cap = cap;
+#endif
+
+#if defined(GRAN)
+ /* add a ContinueThread event to continue execution of current thread */
+ new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
+ ContinueThread,
+ t, (StgClosure*)NULL, (rtsSpark*)NULL);
+ IF_GRAN_DEBUG(bq,
+ debugBelch("GRAN: eventq and runnableq after Garbage collection:\n\n");
+ G_EVENTQ(0);
+ G_CURR_THREADQ(0));
+#endif /* GRAN */
+}
+
+/* ---------------------------------------------------------------------------
+ * rtsSupportsBoundThreads(): is the RTS built to support bound threads?
+ * used by Control.Concurrent for error checking.
+ * ------------------------------------------------------------------------- */
+
+StgBool
+rtsSupportsBoundThreads(void)
+{
+#if defined(THREADED_RTS)
+ return rtsTrue;
+#else
+ return rtsFalse;
+#endif
+}
+
+/* ---------------------------------------------------------------------------
+ * isThreadBound(tso): check whether tso is bound to an OS thread.
+ * ------------------------------------------------------------------------- */
+
+StgBool
+isThreadBound(StgTSO* tso USED_WHEN_THREADED_RTS)
+{
+#if defined(THREADED_RTS)
+ return (tso->bound != NULL);
+#endif
+ return rtsFalse;
+}
+
+/* ---------------------------------------------------------------------------
+ * Singleton fork(). Do not copy any running threads.
+ * ------------------------------------------------------------------------- */
+
+#if !defined(mingw32_HOST_OS) && !defined(SMP)
+#define FORKPROCESS_PRIMOP_SUPPORTED
+#endif
+
+#ifdef FORKPROCESS_PRIMOP_SUPPORTED
+static void
+deleteThreadImmediately(Capability *cap, StgTSO *tso);
+#endif
+StgInt
+forkProcess(HsStablePtr *entry
+#ifndef FORKPROCESS_PRIMOP_SUPPORTED
+ STG_UNUSED
+#endif
+ )
+{
+#ifdef FORKPROCESS_PRIMOP_SUPPORTED
+ pid_t pid;
+ StgTSO* t,*next;
+ Task *task;
+ Capability *cap;
+
+ IF_DEBUG(scheduler,sched_belch("forking!"));
+
+ // ToDo: for SMP, we should probably acquire *all* the capabilities
+ cap = rts_lock();
+
+ pid = fork();
+
+ if (pid) { // parent
+
+ // just return the pid
+ return pid;
+
+ } else { // child
+
+ // delete all threads
+ cap->run_queue_hd = END_TSO_QUEUE;
+ cap->run_queue_tl = END_TSO_QUEUE;
+
+ for (t = all_threads; t != END_TSO_QUEUE; t = next) {
+ next = t->link;
+
+ // don't allow threads to catch the ThreadKilled exception
+ deleteThreadImmediately(cap,t);
+ }
+
+ // wipe the main thread list
+ while ((task = all_tasks) != NULL) {
+ all_tasks = task->all_link;
+ discardTask(task);
+ }
+
+ cap = rts_evalStableIO(cap, entry, NULL); // run the action
+ rts_checkSchedStatus("forkProcess",cap);
+
+ rts_unlock(cap);
+ hs_exit(); // clean up and exit
+ stg_exit(EXIT_SUCCESS);
+ }
+#else /* !FORKPROCESS_PRIMOP_SUPPORTED */
+ barf("forkProcess#: primop not supported on this platform, sorry!\n");
+ return -1;
+#endif
+}
+
+/* ---------------------------------------------------------------------------
+ * Delete the threads on the run queue of the current capability.
+ * ------------------------------------------------------------------------- */
+
+static void
+deleteRunQueue (Capability *cap)
+{
+ StgTSO *t, *next;
+ for (t = cap->run_queue_hd; t != END_TSO_QUEUE; t = next) {
+ ASSERT(t->what_next != ThreadRelocated);
+ next = t->link;
+ deleteThread(cap, t);
+ }
+}
+
+/* startThread and insertThread are now in GranSim.c -- HWL */
+
+
+/* -----------------------------------------------------------------------------
+ Managing the suspended_ccalling_tasks list.
+ Locks required: sched_mutex
+ -------------------------------------------------------------------------- */
+
+STATIC_INLINE void
+suspendTask (Capability *cap, Task *task)
+{
+ ASSERT(task->next == NULL && task->prev == NULL);
+ task->next = cap->suspended_ccalling_tasks;
+ task->prev = NULL;
+ if (cap->suspended_ccalling_tasks) {
+ cap->suspended_ccalling_tasks->prev = task;
+ }
+ cap->suspended_ccalling_tasks = task;
+}
+
+STATIC_INLINE void
+recoverSuspendedTask (Capability *cap, Task *task)
+{
+ if (task->prev) {
+ task->prev->next = task->next;
+ } else {
+ ASSERT(cap->suspended_ccalling_tasks == task);
+ cap->suspended_ccalling_tasks = task->next;
+ }
+ if (task->next) {
+ task->next->prev = task->prev;
+ }
+ task->next = task->prev = NULL;
+}
+
+/* ---------------------------------------------------------------------------
+ * Suspending & resuming Haskell threads.
+ *
+ * When making a "safe" call to C (aka _ccall_GC), the task gives back
+ * its capability before calling the C function. This allows another
+ * task to pick up the capability and carry on running Haskell
+ * threads. It also means that if the C call blocks, it won't lock
+ * the whole system.
+ *
+ * The Haskell thread making the C call is put to sleep for the
+ * duration of the call, on the susepended_ccalling_threads queue. We
+ * give out a token to the task, which it can use to resume the thread
+ * on return from the C function.
+ * ------------------------------------------------------------------------- */
+
+void *
+suspendThread (StgRegTable *reg)
+{
+ Capability *cap;
+ int saved_errno = errno;
+ StgTSO *tso;
+ Task *task;
+
+ /* assume that *reg is a pointer to the StgRegTable part of a Capability.
+ */
+ cap = regTableToCapability(reg);
+
+ task = cap->running_task;
+ tso = cap->r.rCurrentTSO;
+
+ IF_DEBUG(scheduler,
+ sched_belch("thread %d did a safe foreign call", cap->r.rCurrentTSO->id));
+
+ // XXX this might not be necessary --SDM
+ tso->what_next = ThreadRunGHC;
+
+ threadPaused(tso);
+
+ if(tso->blocked_exceptions == NULL) {
+ tso->why_blocked = BlockedOnCCall;
+ tso->blocked_exceptions = END_TSO_QUEUE;
+ } else {
+ tso->why_blocked = BlockedOnCCall_NoUnblockExc;
+ }
+
+ // Hand back capability
+ task->suspended_tso = tso;
+
+ ACQUIRE_LOCK(&cap->lock);
+
+ suspendTask(cap,task);
+ cap->in_haskell = rtsFalse;
+ releaseCapability_(cap);
+
+ RELEASE_LOCK(&cap->lock);
+
+#if defined(THREADED_RTS)
+ /* Preparing to leave the RTS, so ensure there's a native thread/task
+ waiting to take over.
+ */
+ IF_DEBUG(scheduler, sched_belch("thread %d: leaving RTS", tso->id));
+#endif
+
+ errno = saved_errno;
+ return task;
+}
+
+StgRegTable *
+resumeThread (void *task_)
+{
+ StgTSO *tso;
+ Capability *cap;
+ int saved_errno = errno;
+ Task *task = task_;
+
+ cap = task->cap;
+ // Wait for permission to re-enter the RTS with the result.
+ waitForReturnCapability(&cap,task);
+ // we might be on a different capability now... but if so, our
+ // entry on the suspended_ccalling_tasks list will also have been
+ // migrated.
+
+ // Remove the thread from the suspended list
+ recoverSuspendedTask(cap,task);
+
+ tso = task->suspended_tso;
+ task->suspended_tso = NULL;
+ tso->link = END_TSO_QUEUE;
+ IF_DEBUG(scheduler, sched_belch("thread %d: re-entering RTS", tso->id));
+
+ if (tso->why_blocked == BlockedOnCCall) {
+ awakenBlockedQueue(cap,tso->blocked_exceptions);
+ tso->blocked_exceptions = NULL;
+ }
+
+ /* Reset blocking status */
+ tso->why_blocked = NotBlocked;
+
+ cap->r.rCurrentTSO = tso;
+ cap->in_haskell = rtsTrue;
+ errno = saved_errno;
+
+ return &cap->r;
+}
+
+/* ---------------------------------------------------------------------------
+ * Comparing Thread ids.
+ *
+ * This is used from STG land in the implementation of the
+ * instances of Eq/Ord for ThreadIds.
+ * ------------------------------------------------------------------------ */
+
+int
+cmp_thread(StgPtr tso1, StgPtr tso2)
+{
+ StgThreadID id1 = ((StgTSO *)tso1)->id;
+ StgThreadID id2 = ((StgTSO *)tso2)->id;
+
+ if (id1 < id2) return (-1);
+ if (id1 > id2) return 1;
+ return 0;
+}
+
+/* ---------------------------------------------------------------------------
+ * Fetching the ThreadID from an StgTSO.
+ *
+ * This is used in the implementation of Show for ThreadIds.
+ * ------------------------------------------------------------------------ */
+int
+rts_getThreadId(StgPtr tso)
+{
+ return ((StgTSO *)tso)->id;
+}
+
+#ifdef DEBUG
+void
+labelThread(StgPtr tso, char *label)
+{
+ int len;
+ void *buf;
+
+ /* Caveat: Once set, you can only set the thread name to "" */
+ len = strlen(label)+1;
+ buf = stgMallocBytes(len * sizeof(char), "Schedule.c:labelThread()");
+ strncpy(buf,label,len);
+ /* Update will free the old memory for us */
+ updateThreadLabel(((StgTSO *)tso)->id,buf);
+}
+#endif /* DEBUG */
+
+/* ---------------------------------------------------------------------------
+ Create a new thread.
+
+ The new thread starts with the given stack size. Before the
+ scheduler can run, however, this thread needs to have a closure
+ (and possibly some arguments) pushed on its stack. See
+ pushClosure() in Schedule.h.
+
+ createGenThread() and createIOThread() (in SchedAPI.h) are
+ convenient packaged versions of this function.
+
+ currently pri (priority) is only used in a GRAN setup -- HWL
+ ------------------------------------------------------------------------ */
+#if defined(GRAN)
+/* currently pri (priority) is only used in a GRAN setup -- HWL */
+StgTSO *
+createThread(nat size, StgInt pri)
+#else
+StgTSO *
+createThread(Capability *cap, nat size)
+#endif
+{
+ StgTSO *tso;
+ nat stack_size;
+
+ /* sched_mutex is *not* required */
+
+ /* First check whether we should create a thread at all */
+#if defined(PARALLEL_HASKELL)
+ /* check that no more than RtsFlags.ParFlags.maxThreads threads are created */
+ if (advisory_thread_count >= RtsFlags.ParFlags.maxThreads) {
+ threadsIgnored++;
+ debugBelch("{createThread}Daq ghuH: refusing to create another thread; no more than %d threads allowed (currently %d)\n",
+ RtsFlags.ParFlags.maxThreads, advisory_thread_count);
+ return END_TSO_QUEUE;
+ }
+ threadsCreated++;
+#endif
+
+#if defined(GRAN)
+ ASSERT(!RtsFlags.GranFlags.Light || CurrentProc==0);
+#endif
+
+ // ToDo: check whether size = stack_size - TSO_STRUCT_SIZEW
+
+ /* catch ridiculously small stack sizes */
+ if (size < MIN_STACK_WORDS + TSO_STRUCT_SIZEW) {
+ size = MIN_STACK_WORDS + TSO_STRUCT_SIZEW;
+ }
+
+ stack_size = size - TSO_STRUCT_SIZEW;
+
+ tso = (StgTSO *)allocateLocal(cap, size);
+ TICK_ALLOC_TSO(stack_size, 0);
+
+ SET_HDR(tso, &stg_TSO_info, CCS_SYSTEM);
+#if defined(GRAN)
+ SET_GRAN_HDR(tso, ThisPE);
+#endif
+
+ // Always start with the compiled code evaluator
+ tso->what_next = ThreadRunGHC;
+
+ tso->why_blocked = NotBlocked;
+ tso->blocked_exceptions = NULL;
+
+ tso->saved_errno = 0;
+ tso->bound = NULL;
+
+ tso->stack_size = stack_size;
+ tso->max_stack_size = round_to_mblocks(RtsFlags.GcFlags.maxStkSize)
+ - TSO_STRUCT_SIZEW;
+ tso->sp = (P_)&(tso->stack) + stack_size;
+
+ tso->trec = NO_TREC;
+
+#ifdef PROFILING
+ tso->prof.CCCS = CCS_MAIN;
+#endif
+
+ /* put a stop frame on the stack */
+ tso->sp -= sizeofW(StgStopFrame);
+ SET_HDR((StgClosure*)tso->sp,(StgInfoTable *)&stg_stop_thread_info,CCS_SYSTEM);
+ tso->link = END_TSO_QUEUE;
+
+ // ToDo: check this
+#if defined(GRAN)
+ /* uses more flexible routine in GranSim */
+ insertThread(tso, CurrentProc);
+#else
+ /* In a non-GranSim setup the pushing of a TSO onto the runq is separated
+ * from its creation
+ */
+#endif
+
+#if defined(GRAN)
+ if (RtsFlags.GranFlags.GranSimStats.Full)
+ DumpGranEvent(GR_START,tso);
+#elif defined(PARALLEL_HASKELL)
+ if (RtsFlags.ParFlags.ParStats.Full)
+ DumpGranEvent(GR_STARTQ,tso);
+ /* HACk to avoid SCHEDULE
+ LastTSO = tso; */
+#endif
+
+ /* Link the new thread on the global thread list.
+ */
+ ACQUIRE_LOCK(&sched_mutex);
+ tso->id = next_thread_id++; // while we have the mutex
+ tso->global_link = all_threads;
+ all_threads = tso;
+ RELEASE_LOCK(&sched_mutex);
+
+#if defined(DIST)
+ tso->dist.priority = MandatoryPriority; //by default that is...
+#endif
+
+#if defined(GRAN)
+ tso->gran.pri = pri;
+# if defined(DEBUG)
+ tso->gran.magic = TSO_MAGIC; // debugging only
+# endif
+ tso->gran.sparkname = 0;
+ tso->gran.startedat = CURRENT_TIME;
+ tso->gran.exported = 0;
+ tso->gran.basicblocks = 0;
+ tso->gran.allocs = 0;
+ tso->gran.exectime = 0;
+ tso->gran.fetchtime = 0;
+ tso->gran.fetchcount = 0;
+ tso->gran.blocktime = 0;
+ tso->gran.blockcount = 0;
+ tso->gran.blockedat = 0;
+ tso->gran.globalsparks = 0;
+ tso->gran.localsparks = 0;
+ if (RtsFlags.GranFlags.Light)
+ tso->gran.clock = Now; /* local clock */
+ else
+ tso->gran.clock = 0;
+
+ IF_DEBUG(gran,printTSO(tso));
+#elif defined(PARALLEL_HASKELL)
+# if defined(DEBUG)
+ tso->par.magic = TSO_MAGIC; // debugging only
+# endif
+ tso->par.sparkname = 0;
+ tso->par.startedat = CURRENT_TIME;
+ tso->par.exported = 0;
+ tso->par.basicblocks = 0;
+ tso->par.allocs = 0;
+ tso->par.exectime = 0;
+ tso->par.fetchtime = 0;
+ tso->par.fetchcount = 0;
+ tso->par.blocktime = 0;
+ tso->par.blockcount = 0;
+ tso->par.blockedat = 0;
+ tso->par.globalsparks = 0;
+ tso->par.localsparks = 0;
+#endif
+
+#if defined(GRAN)
+ globalGranStats.tot_threads_created++;
+ globalGranStats.threads_created_on_PE[CurrentProc]++;
+ globalGranStats.tot_sq_len += spark_queue_len(CurrentProc);
+ globalGranStats.tot_sq_probes++;
+#elif defined(PARALLEL_HASKELL)
+ // collect parallel global statistics (currently done together with GC stats)
+ if (RtsFlags.ParFlags.ParStats.Global &&
+ RtsFlags.GcFlags.giveStats > NO_GC_STATS) {
+ //debugBelch("Creating thread %d @ %11.2f\n", tso->id, usertime());
+ globalParStats.tot_threads_created++;
+ }
+#endif
+
+#if defined(GRAN)
+ IF_GRAN_DEBUG(pri,
+ sched_belch("==__ schedule: Created TSO %d (%p);",
+ CurrentProc, tso, tso->id));
+#elif defined(PARALLEL_HASKELL)
+ IF_PAR_DEBUG(verbose,
+ sched_belch("==__ schedule: Created TSO %d (%p); %d threads active",
+ (long)tso->id, tso, advisory_thread_count));
+#else
+ IF_DEBUG(scheduler,sched_belch("created thread %ld, stack size = %lx words",
+ (long)tso->id, (long)tso->stack_size));
+#endif
+ return tso;
+}
+
+#if defined(PAR)
+/* RFP:
+ all parallel thread creation calls should fall through the following routine.
+*/
+StgTSO *
+createThreadFromSpark(rtsSpark spark)
+{ StgTSO *tso;
+ ASSERT(spark != (rtsSpark)NULL);
+// JB: TAKE CARE OF THIS COUNTER! BUGGY
+ if (advisory_thread_count >= RtsFlags.ParFlags.maxThreads)
+ { threadsIgnored++;
+ barf("{createSparkThread}Daq ghuH: refusing to create another thread; no more than %d threads allowed (currently %d)",
+ RtsFlags.ParFlags.maxThreads, advisory_thread_count);
+ return END_TSO_QUEUE;
+ }
+ else
+ { threadsCreated++;
+ tso = createThread(RtsFlags.GcFlags.initialStkSize);
+ if (tso==END_TSO_QUEUE)
+ barf("createSparkThread: Cannot create TSO");
+#if defined(DIST)
+ tso->priority = AdvisoryPriority;
+#endif
+ pushClosure(tso,spark);
+ addToRunQueue(tso);
+ advisory_thread_count++; // JB: TAKE CARE OF THIS COUNTER! BUGGY
+ }
+ return tso;
+}
+#endif
+
+/*
+ Turn a spark into a thread.
+ ToDo: fix for SMP (needs to acquire SCHED_MUTEX!)
+*/
+#if 0
+StgTSO *
+activateSpark (rtsSpark spark)
+{
+ StgTSO *tso;
+
+ tso = createSparkThread(spark);
+ if (RtsFlags.ParFlags.ParStats.Full) {
+ //ASSERT(run_queue_hd == END_TSO_QUEUE); // I think ...
+ IF_PAR_DEBUG(verbose,
+ debugBelch("==^^ activateSpark: turning spark of closure %p (%s) into a thread\n",
+ (StgClosure *)spark, info_type((StgClosure *)spark)));
+ }
+ // ToDo: fwd info on local/global spark to thread -- HWL
+ // tso->gran.exported = spark->exported;
+ // tso->gran.locked = !spark->global;
+ // tso->gran.sparkname = spark->name;
+
+ return tso;
+}
+#endif
+
+/* ---------------------------------------------------------------------------
+ * scheduleThread()
+ *
+ * scheduleThread puts a thread on the end of the runnable queue.
+ * This will usually be done immediately after a thread is created.
+ * The caller of scheduleThread must create the thread using e.g.
+ * createThread and push an appropriate closure
+ * on this thread's stack before the scheduler is invoked.
+ * ------------------------------------------------------------------------ */
+
+void
+scheduleThread(Capability *cap, StgTSO *tso)
+{
+ // The thread goes at the *end* of the run-queue, to avoid possible
+ // starvation of any threads already on the queue.
+ appendToRunQueue(cap,tso);
+}
+
+Capability *
+scheduleWaitThread (StgTSO* tso, /*[out]*/HaskellObj* ret, Capability *cap)
+{
+ Task *task;
+
+ // We already created/initialised the Task
+ task = cap->running_task;
+
+ // This TSO is now a bound thread; make the Task and TSO
+ // point to each other.
+ tso->bound = task;
+
+ task->tso = tso;
+ task->ret = ret;
+ task->stat = NoStatus;
+
+ appendToRunQueue(cap,tso);
+
+ IF_DEBUG(scheduler, sched_belch("new bound thread (%d)", tso->id));
+
+#if defined(GRAN)
+ /* GranSim specific init */
+ CurrentTSO = m->tso; // the TSO to run
+ procStatus[MainProc] = Busy; // status of main PE
+ CurrentProc = MainProc; // PE to run it on
+#endif
+
+ cap = schedule(cap,task);
+
+ ASSERT(task->stat != NoStatus);
+
+ IF_DEBUG(scheduler, sched_belch("bound thread (%d) finished", task->tso->id));
+ return cap;
+}
+
+/* ----------------------------------------------------------------------------
+ * Starting Tasks
+ * ------------------------------------------------------------------------- */
+
+#if defined(THREADED_RTS)
+void
+workerStart(Task *task)
+{
+ Capability *cap;
+
+ // See startWorkerTask().
+ ACQUIRE_LOCK(&task->lock);
+ cap = task->cap;
+ RELEASE_LOCK(&task->lock);
+
+ // set the thread-local pointer to the Task:
+ taskEnter(task);
+
+ // schedule() runs without a lock.
+ cap = schedule(cap,task);
+
+ // On exit from schedule(), we have a Capability.
+ releaseCapability(cap);
+ taskStop(task);
+}
+#endif
+
+/* ---------------------------------------------------------------------------
+ * initScheduler()
+ *
+ * Initialise the scheduler. This resets all the queues - if the
+ * queues contained any threads, they'll be garbage collected at the
+ * next pass.
+ *
+ * ------------------------------------------------------------------------ */
+
+void
+initScheduler(void)
+{
+#if defined(GRAN)
+ nat i;
+ for (i=0; i<=MAX_PROC; i++) {
+ run_queue_hds[i] = END_TSO_QUEUE;
+ run_queue_tls[i] = END_TSO_QUEUE;
+ blocked_queue_hds[i] = END_TSO_QUEUE;
+ blocked_queue_tls[i] = END_TSO_QUEUE;
+ ccalling_threadss[i] = END_TSO_QUEUE;
+ blackhole_queue[i] = END_TSO_QUEUE;
+ sleeping_queue = END_TSO_QUEUE;
+ }
+#elif !defined(THREADED_RTS)
+ blocked_queue_hd = END_TSO_QUEUE;
+ blocked_queue_tl = END_TSO_QUEUE;
+ sleeping_queue = END_TSO_QUEUE;
+#endif
+
+ blackhole_queue = END_TSO_QUEUE;
+ all_threads = END_TSO_QUEUE;
+
+ context_switch = 0;
+ interrupted = 0;
+
+ RtsFlags.ConcFlags.ctxtSwitchTicks =
+ RtsFlags.ConcFlags.ctxtSwitchTime / TICK_MILLISECS;
+
+#if defined(THREADED_RTS)
+ /* Initialise the mutex and condition variables used by
+ * the scheduler. */
+ initMutex(&sched_mutex);
+#endif
+
+ ACQUIRE_LOCK(&sched_mutex);
+
+ /* A capability holds the state a native thread needs in
+ * order to execute STG code. At least one capability is
+ * floating around (only SMP builds have more than one).
+ */
+ initCapabilities();
+
+ initTaskManager();
+
+#if defined(SMP)
+ /*
+ * Eagerly start one worker to run each Capability, except for
+ * Capability 0. The idea is that we're probably going to start a
+ * bound thread on Capability 0 pretty soon, so we don't want a
+ * worker task hogging it.
+ */
+ {
+ nat i;
+ Capability *cap;
+ for (i = 1; i < n_capabilities; i++) {
+ cap = &capabilities[i];
+ ACQUIRE_LOCK(&cap->lock);
+ startWorkerTask(cap, workerStart);
+ RELEASE_LOCK(&cap->lock);
+ }
+ }
+#endif
+
+#if /* defined(SMP) ||*/ defined(PARALLEL_HASKELL)
+ initSparkPools();
+#endif
+
+ RELEASE_LOCK(&sched_mutex);
+}
+
+void
+exitScheduler( void )
+{
+ interrupted = rtsTrue;
+ shutting_down_scheduler = rtsTrue;
+
+#if defined(THREADED_RTS)
+ {
+ Task *task;
+ nat i;
+
+ ACQUIRE_LOCK(&sched_mutex);
+ task = newBoundTask();
+ RELEASE_LOCK(&sched_mutex);
+
+ for (i = 0; i < n_capabilities; i++) {
+ shutdownCapability(&capabilities[i], task);
+ }
+ boundTaskExiting(task);
+ stopTaskManager();
+ }
+#endif
+}
+
+/* ---------------------------------------------------------------------------