[project @ 2005-05-21 16:09:18 by panne]
[ghc-hetmet.git] / ghc / rts / Schedule.c
index 6cac6c1..8e1a43e 100644 (file)
@@ -174,14 +174,12 @@ static StgTSO *suspended_ccalling_threads;
 /* flag set by signal handler to precipitate a context switch */
 int context_switch = 0;
 
+/* flag that tracks whether we have done any execution in this time slice. */
+nat recent_activity = ACTIVITY_YES;
+
 /* if this flag is set as well, give up execution */
 rtsBool interrupted = rtsFalse;
 
-/* If this flag is set, we are running Haskell code.  Used to detect
- * uses of 'foreign import unsafe' that should be 'safe'.
- */
-static rtsBool in_haskell = rtsFalse;
-
 /* Next thread ID to allocate.
  * Locks required: thread_id_mutex
  */
@@ -217,12 +215,6 @@ StgTSO *CurrentTSO;
  */
 StgTSO dummy_tso;
 
-# if defined(SMP)
-static Condition gc_pending_cond = INIT_COND_VAR;
-# endif
-
-static rtsBool ready_to_gc;
-
 /*
  * Set to TRUE when entering a shutdown state (via shutdownHaskellAndExit()) --
  * in an MT setting, needed to signal that a worker thread shouldn't hang around
@@ -295,8 +287,8 @@ static rtsBool scheduleHandleYield( StgTSO *t, nat prev_what_next );
 static void scheduleHandleThreadBlocked( StgTSO *t );
 static rtsBool scheduleHandleThreadFinished( StgMainThread *mainThread, 
                                             Capability *cap, StgTSO *t );
-static void scheduleDoHeapProfile(void);
-static void scheduleDoGC(void);
+static rtsBool scheduleDoHeapProfile(rtsBool ready_to_gc);
+static void scheduleDoGC(Capability *cap);
 
 static void unblockThread(StgTSO *tso);
 static rtsBool checkBlackHoles(void);
@@ -324,13 +316,13 @@ StgTSO * activateSpark (rtsSpark spark);
  * ------------------------------------------------------------------------- */
 
 #if defined(RTS_SUPPORTS_THREADS)
-static rtsBool startingWorkerThread = rtsFalse;
+static nat startingWorkerThread = 0;
 
 static void
 taskStart(void)
 {
   ACQUIRE_LOCK(&sched_mutex);
-  startingWorkerThread = rtsFalse;
+  startingWorkerThread--;
   schedule(NULL,NULL);
   taskStop();
   RELEASE_LOCK(&sched_mutex);
@@ -341,14 +333,14 @@ startSchedulerTaskIfNecessary(void)
 {
     if ( !EMPTY_RUN_QUEUE()
         && !shutting_down_scheduler // not if we're shutting down
-        && !startingWorkerThread )
+        && startingWorkerThread==0)
     {
        // we don't want to start another worker thread
        // just because the last one hasn't yet reached the
        // "waiting for capability" state
-       startingWorkerThread = rtsTrue;
+       startingWorkerThread++;
        if (!maybeStartNewWorker(taskStart)) {
-           startingWorkerThread = rtsFalse;
+           startingWorkerThread--;
        }
     }
 }
@@ -429,6 +421,7 @@ schedule( StgMainThread *mainThread USED_WHEN_RTS_SUPPORTS_THREADS,
 # endif
 #endif
   nat prev_what_next;
+  rtsBool ready_to_gc;
   
   // Pre-condition: sched_mutex is held.
   // We might have a capability, passed in as initialCapability.
@@ -465,21 +458,6 @@ schedule( StgMainThread *mainThread USED_WHEN_RTS_SUPPORTS_THREADS,
       CurrentTSO = event->tso;
 #endif
 
-      IF_DEBUG(scheduler, printAllThreads());
-
-#if defined(SMP)
-      // 
-      // Wait until GC has completed, if necessary.
-      //
-      if (ready_to_gc) {
-         if (cap != NULL) {
-             releaseCapability(cap);
-             IF_DEBUG(scheduler,sched_belch("waiting for GC"));
-             waitCondition( &gc_pending_cond, &sched_mutex );
-         }
-      }
-#endif
-
 #if defined(RTS_SUPPORTS_THREADS)
       // Yield the capability to higher-priority tasks if necessary.
       //
@@ -496,11 +474,20 @@ schedule( StgMainThread *mainThread USED_WHEN_RTS_SUPPORTS_THREADS,
 
       // We now have a capability...
 #endif
+      
+#if 0 /* extra sanity checking */
+      { 
+         StgMainThread *m;
+         for (m = main_threads; m != NULL; m = m->link) {
+             ASSERT(get_itbl(m->tso)->type == TSO);
+         }
+      }
+#endif
 
     // Check whether we have re-entered the RTS from Haskell without
     // going via suspendThread()/resumeThread (i.e. a 'safe' foreign
     // call).
-    if (in_haskell) {
+    if (cap->r.rInHaskell) {
          errorBelch("schedule: re-entered unsafely.\n"
                     "   Perhaps a 'foreign import unsafe' should be 'safe'?");
          stg_exit(1);
@@ -689,7 +676,9 @@ run_thread:
     prev_what_next = t->what_next;
 
     errno = t->saved_errno;
-    in_haskell = rtsTrue;
+    cap->r.rInHaskell = rtsTrue;
+
+    recent_activity = ACTIVITY_YES;
 
     switch (prev_what_next) {
 
@@ -711,13 +700,20 @@ run_thread:
       barf("schedule: invalid what_next field");
     }
 
+#if defined(SMP)
+    // in SMP mode, we might return with a different capability than
+    // we started with, if the Haskell thread made a foreign call.  So
+    // let's find out what our current Capability is:
+    cap = myCapability();
+#endif
+
     // We have run some Haskell code: there might be blackhole-blocked
     // threads to wake up now.
     if ( blackhole_queue != END_TSO_QUEUE ) {
        blackholes_need_checking = rtsTrue;
     }
 
-    in_haskell = rtsFalse;
+    cap->r.rInHaskell = rtsFalse;
 
     // The TSO might have moved, eg. if it re-entered the RTS and a GC
     // happened.  So find the new location:
@@ -744,6 +740,8 @@ run_thread:
     
     schedulePostRunThread();
 
+    ready_to_gc = rtsFalse;
+
     switch (ret) {
     case HeapOverflow:
        ready_to_gc = scheduleHandleHeapOverflow(cap,t);
@@ -762,7 +760,6 @@ run_thread:
 
     case ThreadBlocked:
        scheduleHandleThreadBlocked(t);
-       threadPaused(t);
        break;
 
     case ThreadFinished:
@@ -773,8 +770,8 @@ run_thread:
       barf("schedule: invalid thread return code %d", (int)ret);
     }
 
-    scheduleDoHeapProfile();
-    scheduleDoGC();
+    if (scheduleDoHeapProfile(ready_to_gc)) { ready_to_gc = rtsFalse; }
+    if (ready_to_gc) { scheduleDoGC(cap); }
   } /* end of while() */
 
   IF_PAR_DEBUG(verbose,
@@ -816,7 +813,7 @@ schedulePreLoop(void)
 static void
 scheduleStartSignalHandlers(void)
 {
-#if defined(RTS_USER_SIGNALS)
+#if defined(RTS_USER_SIGNALS) && !defined(RTS_SUPPORTS_THREADS)
     if (signals_pending()) {
       RELEASE_LOCK(&sched_mutex); /* ToDo: kill */
       startSignalHandlers();
@@ -871,6 +868,12 @@ scheduleCheckBlackHoles( void )
 static void
 scheduleDetectDeadlock(void)
 {
+
+#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
@@ -879,7 +882,16 @@ scheduleDetectDeadlock(void)
      */
     if ( EMPTY_THREAD_QUEUES() )
     {
-#if !defined(PARALLEL_HASKELL) && !defined(RTS_SUPPORTS_THREADS)
+#if defined(RTS_SUPPORTS_THREADS)
+       /* 
+        * 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
@@ -887,10 +899,12 @@ scheduleDetectDeadlock(void)
        // they are unreachable and will therefore be sent an
        // exception.  Any threads thus released will be immediately
        // runnable.
+
        GarbageCollect(GetRoots,rtsTrue);
+       recent_activity = ACTIVITY_DONE_GC;
        if ( !EMPTY_RUN_QUEUE() ) return;
 
-#if defined(RTS_USER_SIGNALS)
+#if defined(RTS_USER_SIGNALS) && !defined(RTS_SUPPORTS_THREADS)
        /* If we have user-installed signal handlers, then wait
         * for signals to arrive rather then bombing out with a
         * deadlock.
@@ -912,6 +926,7 @@ scheduleDetectDeadlock(void)
        }
 #endif
 
+#if !defined(RTS_SUPPORTS_THREADS)
        /* Probably a real deadlock.  Send the current main thread the
         * Deadlock exception (or in the SMP build, send *all* main
         * threads the deadlock exception, since none of them can make
@@ -921,6 +936,7 @@ scheduleDetectDeadlock(void)
            StgMainThread *m;
            m = main_threads;
            switch (m->tso->why_blocked) {
+           case BlockedOnSTM:
            case BlockedOnBlackHole:
            case BlockedOnException:
            case BlockedOnMVar:
@@ -930,11 +946,6 @@ scheduleDetectDeadlock(void)
                barf("deadlock: main thread blocked in a strange way");
            }
        }
-
-#elif defined(RTS_SUPPORTS_THREADS)
-    // ToDo: add deadlock detection in threaded RTS
-#elif defined(PARALLEL_HASKELL)
-    // ToDo: add deadlock detection in GUM (similar to SMP) -- HWL
 #endif
     }
 }
@@ -1469,12 +1480,13 @@ scheduleHandleHeapOverflow( Capability *cap, StgTSO *t )
                 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 it would push us over the
-       // alloc_blocks_lim limit; we'll GC first.
-       if (alloc_blocks + blocks < alloc_blocks_lim) {
+       // 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.
            
-           alloc_blocks += blocks;
            bd = allocGroup( blocks );
+           cap->r.rNursery->n_blocks += blocks;
            
            // link the new group into the list
            bd->link = cap->r.rCurrentNursery;
@@ -1482,9 +1494,11 @@ scheduleHandleHeapOverflow( Capability *cap, StgTSO *t )
            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->blocks == cap->r.rNursery);
-               cap->r.rNursery = g0s0->blocks = bd;
+                      g0s0 == cap->r.rNursery);
+#endif
+               cap->r.rNursery->blocks = bd;
            }             
            cap->r.rCurrentNursery->u.back = bd;
            
@@ -1498,17 +1512,15 @@ scheduleHandleHeapOverflow( Capability *cap, StgTSO *t )
            { 
                bdescr *x;
                for (x = bd; x < bd + blocks; x++) {
-                   x->step = g0s0;
+                   x->step = cap->r.rNursery;
                    x->gen_no = 0;
                    x->flags = 0;
                }
            }
            
-           // don't forget to update the block count in g0s0.
-           g0s0->n_blocks += blocks;
            // This assert can be a killer if the app is doing lots
            // of large block allocations.
-           ASSERT(countBlocks(g0s0->blocks) == g0s0->n_blocks);
+           IF_DEBUG(sanity, checkNurserySanity(cap->r.rNursery));
            
            // now update the nursery to point to the new block
            cap->r.rCurrentNursery = bd;
@@ -1522,14 +1534,9 @@ scheduleHandleHeapOverflow( Capability *cap, StgTSO *t )
        }
     }
     
-    /* make all the running tasks block on a condition variable,
-     * maybe set context_switch and wait till they all pile in,
-     * then have them wait on a GC condition variable.
-     */
     IF_DEBUG(scheduler,
             debugBelch("--<< thread %ld (%s) stopped: HeapOverflow\n", 
                        (long)t->id, whatNext_strs[t->what_next]));
-    threadPaused(t);
 #if defined(GRAN)
     ASSERT(!is_on_queue(t,CurrentProc));
 #elif defined(PARALLEL_HASKELL)
@@ -1561,7 +1568,6 @@ scheduleHandleStackOverflow( StgTSO *t)
     /* just adjust the stack for this thread, then pop it back
      * on the run queue.
      */
-    threadPaused(t);
     { 
        /* enlarge the stack */
        StgTSO *new_t = threadStackOverflow(t);
@@ -1620,8 +1626,6 @@ scheduleHandleYield( StgTSO *t, nat prev_what_next )
        return rtsTrue;
     }
     
-    threadPaused(t);
-    
 #if defined(GRAN)
     ASSERT(!is_on_queue(t,CurrentProc));
       
@@ -1694,12 +1698,19 @@ scheduleHandleThreadBlocked( StgTSO *t
     emitSchedule = rtsTrue;
     
 #else /* !GRAN */
-      /* don't need to do anything.  Either the thread is blocked on
-       * I/O, in which case we'll have called addToBlockedQueue
-       * previously, or it's blocked on an MVar or Blackhole, in which
-       * case it'll be on the relevant queue already.
-       */
+
+      // 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]);
@@ -1816,12 +1827,13 @@ scheduleHandleThreadFinished( StgMainThread *mainThread
          removeThreadLabel((StgWord)mainThread->tso->id);
 #endif
          if (mainThread->prev == NULL) {
+             ASSERT(mainThread == main_threads);
              main_threads = mainThread->link;
          } else {
              mainThread->prev->link = mainThread->link;
          }
          if (mainThread->link != NULL) {
-             mainThread->link->prev = NULL;
+             mainThread->link->prev = mainThread->prev;
          }
          releaseCapability(cap);
          return rtsTrue; // tells schedule() to return
@@ -1847,10 +1859,10 @@ scheduleHandleThreadFinished( StgMainThread *mainThread
  * Perform a heap census, if PROFILING
  * -------------------------------------------------------------------------- */
 
-static void
-scheduleDoHeapProfile(void)
+static rtsBool
+scheduleDoHeapProfile( rtsBool ready_to_gc STG_UNUSED )
 {
-#ifdef PROFILING
+#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 ||
@@ -1859,9 +1871,10 @@ scheduleDoHeapProfile(void)
        GarbageCollect(GetRoots, rtsTrue);
        heapCensus();
        performHeapProfile = rtsFalse;
-       ready_to_gc = rtsFalse; // we already GC'd
+       return rtsTrue;  // true <=> we already GC'd
     }
 #endif
+    return rtsFalse;
 }
 
 /* -----------------------------------------------------------------------------
@@ -1870,66 +1883,99 @@ scheduleDoHeapProfile(void)
  * -------------------------------------------------------------------------- */
 
 static void
-scheduleDoGC(void)
+scheduleDoGC( Capability *cap STG_UNUSED )
 {
     StgTSO *t;
+#ifdef SMP
+    static rtsBool waiting_for_gc;
+    int n_capabilities = RtsFlags.ParFlags.nNodes - 1; 
+           // subtract one because we're already holding one.
+    Capability *caps[n_capabilities];
+#endif
 
 #ifdef SMP
-    // The last task to stop actually gets to do the GC.  The rest
-    // of the tasks release their capabilities and wait gc_pending_cond.
-    if (ready_to_gc && allFreeCapabilities())
-#else
-    if (ready_to_gc)
+    // 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.
+    //
+       
+    // Someone else is already trying to GC
+    if (waiting_for_gc) return;
+    waiting_for_gc = rtsTrue;
+
+    caps[n_capabilities] = cap;
+    while (n_capabilities > 0) {
+       IF_DEBUG(scheduler, sched_belch("ready_to_gc, grabbing all the capabilies (%d left)", n_capabilities));
+       waitForReturnCapability(&sched_mutex, &cap);
+       n_capabilities--;
+       caps[n_capabilities] = cap;
+    }
+
+    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.
-        */
-       for (t = all_threads; t != END_TSO_QUEUE; t = t -> link) {
-           if (t -> what_next != ThreadRelocated && t -> trec != NO_TREC && t -> why_blocked == NotBlocked) {
-               if (!stmValidateTransaction (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_(t, NULL, rtsTrue);
-                   
+
+    /* 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.
+     */
+    for (t = all_threads; t != END_TSO_QUEUE; t = t -> link) {
+       if (t -> what_next != ThreadRelocated && t -> trec != NO_TREC && t -> why_blocked == NotBlocked) {
+           if (!stmValidateTransaction (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_(t, NULL, rtsTrue);
+               
 #ifdef REG_R1
-                   ASSERT(get_itbl((StgClosure *)t->sp)->type == ATOMICALLY_FRAME);
+               ASSERT(get_itbl((StgClosure *)t->sp)->type == ATOMICALLY_FRAME);
 #endif
-               }
            }
        }
+    }
+    
+    // so this happens periodically:
+    scheduleCheckBlackHoles();
+    
+    IF_DEBUG(scheduler, printAllThreads());
 
-       // so this happens periodically:
-       scheduleCheckBlackHoles();
-
-       /* 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.
-        */
+    /* 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(RTS_SUPPORTS_THREADS)
-       IF_DEBUG(scheduler,sched_belch("doing GC"));
+    IF_DEBUG(scheduler,sched_belch("doing GC"));
 #endif
-       GarbageCollect(GetRoots,rtsFalse);
-       ready_to_gc = rtsFalse;
+    GarbageCollect(GetRoots,rtsFalse);
+    
 #if defined(SMP)
-       broadcastCondition(&gc_pending_cond);
+    {
+       // release our stash of capabilities.
+       nat i;
+       for (i = 0; i < RtsFlags.ParFlags.nNodes-1; i++) {
+           releaseCapability(caps[i]);
+       }
+    }
 #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));
+    /* 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 */
-    }
 }
 
 /* ---------------------------------------------------------------------------
@@ -1940,7 +1986,7 @@ scheduleDoGC(void)
 StgBool
 rtsSupportsBoundThreads(void)
 {
-#ifdef THREADED_RTS
+#if defined(RTS_SUPPORTS_THREADS)
   return rtsTrue;
 #else
   return rtsFalse;
@@ -1954,7 +2000,7 @@ rtsSupportsBoundThreads(void)
 StgBool
 isThreadBound(StgTSO* tso USED_IN_THREADED_RTS)
 {
-#ifdef THREADED_RTS
+#if defined(RTS_SUPPORTS_THREADS)
   return (tso->main != NULL);
 #endif
   return rtsFalse;
@@ -2048,8 +2094,12 @@ deleteAllThreads ( void )
   StgTSO* t, *next;
   IF_DEBUG(scheduler,sched_belch("deleting all threads"));
   for (t = all_threads; t != END_TSO_QUEUE; t = next) {
-      next = t->global_link;
-      deleteThread(t);
+      if (t->what_next == ThreadRelocated) {
+         next = t->link;
+      } else {
+         next = t->global_link;
+         deleteThread(t);
+      }
   }      
 
   // The run queue now contains a bunch of ThreadKilled threads.  We
@@ -2116,6 +2166,7 @@ suspendThread( StgRegTable *reg )
   tok = cap->r.rCurrentTSO->id;
 
   /* Hand back capability */
+  cap->r.rInHaskell = rtsFalse;
   releaseCapability(cap);
   
 #if defined(RTS_SUPPORTS_THREADS)
@@ -2125,7 +2176,6 @@ suspendThread( StgRegTable *reg )
   IF_DEBUG(scheduler, sched_belch("worker (token %d): leaving RTS", tok));
 #endif
 
-  in_haskell = rtsFalse;
   RELEASE_LOCK(&sched_mutex);
   
   errno = saved_errno;
@@ -2173,7 +2223,7 @@ resumeThread( StgInt tok )
   tso->why_blocked  = NotBlocked;
 
   cap->r.rCurrentTSO = tso;
-  in_haskell = rtsTrue;
+  cap->r.rInHaskell = rtsTrue;
   RELEASE_LOCK(&sched_mutex);
   errno = saved_errno;
   return &cap->r;
@@ -2615,6 +2665,7 @@ initScheduler(void)
 
 #if defined(SMP)
   /* eagerly start some extra workers */
+  startingWorkerThread = RtsFlags.ParFlags.nNodes;
   startTasks(RtsFlags.ParFlags.nNodes, taskStart);
 #endif
 
@@ -2933,7 +2984,7 @@ unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
 #endif
 
 #if defined(GRAN)
-static StgBlockingQueueElement *
+StgBlockingQueueElement *
 unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
 {
     StgTSO *tso;
@@ -2973,7 +3024,7 @@ unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
                             tso->id, tso));
 }
 #elif defined(PARALLEL_HASKELL)
-static StgBlockingQueueElement *
+StgBlockingQueueElement *
 unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
 {
     StgBlockingQueueElement *next;
@@ -3019,7 +3070,7 @@ unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
 }
 
 #else /* !GRAN && !PARALLEL_HASKELL */
-static StgTSO *
+StgTSO *
 unblockOneLocked(StgTSO *tso)
 {
   StgTSO *next;
@@ -3201,6 +3252,11 @@ interruptStgRts(void)
 {
     interrupted    = 1;
     context_switch = 1;
+    threadRunnable();
+    /* ToDo: if invoked from a signal handler, this threadRunnable
+     * only works if there's another thread (not this one) waiting to
+     * be woken up.
+     */
 }
 
 /* -----------------------------------------------------------------------------
@@ -3324,6 +3380,12 @@ unblockThread(StgTSO *tso)
              blocked_queue_tl = (StgTSO *)prev;
            }
          }
+#if defined(mingw32_HOST_OS)
+         /* (Cooperatively) signal that the worker thread should abort
+          * the request.
+          */
+         abandonWorkRequest(tso->block_info.async_result->reqID);
+#endif
          goto done;
        }
       }
@@ -3458,6 +3520,12 @@ unblockThread(StgTSO *tso)
              blocked_queue_tl = prev;
            }
          }
+#if defined(mingw32_HOST_OS)
+         /* (Cooperatively) signal that the worker thread should abort
+          * the request.
+          */
+         abandonWorkRequest(tso->block_info.async_result->reqID);
+#endif
          goto done;
        }
       }
@@ -3720,12 +3788,12 @@ raiseAsync_(StgTSO *tso, StgClosure *exception, rtsBool stop_at_atomically)
 #ifdef PROFILING
            StgCatchFrame *cf = (StgCatchFrame *)frame;
 #endif
-           StgClosure *raise;
+           StgThunk *raise;
            
            // we've got an exception to raise, so let's pass it to the
            // handler in this frame.
            //
-           raise = (StgClosure *)allocate(sizeofW(StgClosure)+1);
+           raise = (StgThunk *)allocate(sizeofW(StgThunk)+1);
            TICK_ALLOC_SE_THK(1,0);
            SET_HDR(raise,&stg_raise_info,cf->header.prof.ccs);
            raise->payload[0] = exception;
@@ -3763,7 +3831,7 @@ raiseAsync_(StgTSO *tso, StgClosure *exception, rtsBool stop_at_atomically)
            // fun field.
            //
            words = frame - sp - 1;
-           ap = (StgAP_STACK *)allocate(PAP_sizeW(words));
+           ap = (StgAP_STACK *)allocate(AP_STACK_sizeW(words));
            
            ap->size = words;
            ap->fun  = (StgClosure *)sp[0];
@@ -3830,7 +3898,7 @@ raiseAsync_(StgTSO *tso, StgClosure *exception, rtsBool stop_at_atomically)
 StgWord
 raiseExceptionHelper (StgTSO *tso, StgClosure *exception)
 {
-    StgClosure *raise_closure = NULL;
+    StgThunk *raise_closure = NULL;
     StgPtr p, next;
     StgRetInfoTable *info;
     //
@@ -3867,11 +3935,11 @@ raiseExceptionHelper (StgTSO *tso, StgClosure *exception)
            // Only create raise_closure if we need to.
            if (raise_closure == NULL) {
                raise_closure = 
-                   (StgClosure *)allocate(sizeofW(StgClosure)+MIN_UPD_SIZE);
+                   (StgThunk *)allocate(sizeofW(StgThunk)+MIN_UPD_SIZE);
                SET_HDR(raise_closure, &stg_raise_info, CCCS);
                raise_closure->payload[0] = exception;
            }
-           UPD_IND(((StgUpdateFrame *)p)->updatee,raise_closure);
+           UPD_IND(((StgUpdateFrame *)p)->updatee,(StgClosure *)raise_closure);
            p = next;
            continue;
 
@@ -4006,10 +4074,10 @@ printThreadBlockage(StgTSO *tso)
 {
   switch (tso->why_blocked) {
   case BlockedOnRead:
-    debugBelch("is blocked on read from fd %ld", tso->block_info.fd);
+    debugBelch("is blocked on read from fd %d", (int)(tso->block_info.fd));
     break;
   case BlockedOnWrite:
-    debugBelch("is blocked on write to fd %ld", tso->block_info.fd);
+    debugBelch("is blocked on write to fd %d", (int)(tso->block_info.fd));
     break;
 #if defined(mingw32_HOST_OS)
     case BlockedOnDoProc:
@@ -4017,7 +4085,7 @@ printThreadBlockage(StgTSO *tso)
     break;
 #endif
   case BlockedOnDelay:
-    debugBelch("is blocked until %ld", tso->block_info.target);
+    debugBelch("is blocked until %ld", (long)(tso->block_info.target));
     break;
   case BlockedOnMVar:
     debugBelch("is blocked on an MVar");
@@ -4093,7 +4161,7 @@ printAllThreads(void)
   debugBelch("all threads:\n");
 # endif
 
-  for (t = all_threads; t != END_TSO_QUEUE; t = t->global_link) {
+  for (t = all_threads; t != END_TSO_QUEUE; ) {
     debugBelch("\tthread %d @ %p ", t->id, (void *)t);
 #if defined(DEBUG)
     {
@@ -4101,8 +4169,14 @@ printAllThreads(void)
       if (label) debugBelch("[\"%s\"] ",(char *)label);
     }
 #endif
-    printThreadStatus(t);
-    debugBelch("\n");
+    if (t->what_next == ThreadRelocated) {
+       debugBelch("has been relocated...\n");
+       t = t->link;
+    } else {
+       printThreadStatus(t);
+       debugBelch("\n");
+       t = t->global_link;
+    }
   }
 }