[project @ 2002-02-15 17:49:23 by sof]
[ghc-hetmet.git] / ghc / rts / Schedule.c
index f0c6019..1f8d06b 100644 (file)
@@ -1,20 +1,27 @@
 /* ---------------------------------------------------------------------------
- * $Id: Schedule.c,v 1.83 2000/12/04 12:31:21 simonmar Exp $
+ * $Id: Schedule.c,v 1.127 2002/02/15 17:49:23 sof Exp $
  *
  * (c) The GHC Team, 1998-2000
  *
  * Scheduler
  *
- * The main scheduling code in GranSim is quite different from that in std
- * (concurrent) Haskell: while concurrent Haskell just iterates over the
- * threads in the runnable queue, GranSim is event driven, i.e. it iterates
- * over the events in the global event queue.  -- HWL
+ * Different GHC ways use this scheduler quite differently (see comments below)
+ * Here is the global picture:
+ *
+ * WAY  Name     CPP flag  What's it for
+ * --------------------------------------
+ * mp   GUM      PAR          Parallel execution on a distributed memory machine
+ * s    SMP      SMP          Parallel execution on a shared memory machine
+ * mg   GranSim  GRAN         Simulation of parallel execution
+ * md   GUM/GdH  DIST         Distributed execution (based on GUM)
+ *
  * --------------------------------------------------------------------------*/
 
 //@node Main scheduling code, , ,
 //@section Main scheduling code
 
-/* Version with scheduler monitor support for SMPs.
+/* 
+ * Version with scheduler monitor support for SMPs (WAY=s):
 
    This design provides a high-level API to create and schedule threads etc.
    as documented in the SMP design document.
    In a non-SMP build, there is one global capability, namely MainRegTable.
 
    SDM & KH, 10/99
+
+ * Version with support for distributed memory parallelism aka GUM (WAY=mp):
+
+   The main scheduling loop in GUM iterates until a finish message is received.
+   In that case a global flag @receivedFinish@ is set and this instance of
+   the RTS shuts down. See ghc/rts/parallel/HLComms.c:processMessages()
+   for the handling of incoming messages, such as PP_FINISH.
+   Note that in the parallel case we have a system manager that coordinates
+   different PEs, each of which are running one instance of the RTS.
+   See ghc/rts/parallel/SysMan.c for the main routine of the parallel program.
+   From this routine processes executing ghc/rts/Main.c are spawned. -- HWL
+
+ * Version with support for simulating parallel execution aka GranSim (WAY=mg):
+
+   The main scheduling code in GranSim is quite different from that in std
+   (concurrent) Haskell: while concurrent Haskell just iterates over the
+   threads in the runnable queue, GranSim is event driven, i.e. it iterates
+   over the events in the global event queue.  -- HWL
 */
 
 //@menu
@@ -50,6 +75,7 @@
 //@node Includes, Variables and Data structures, Main scheduling code, Main scheduling code
 //@subsection Includes
 
+#include "PosixSource.h"
 #include "Rts.h"
 #include "SchedAPI.h"
 #include "RtsUtils.h"
 #include "Storage.h"
 #include "StgRun.h"
 #include "StgStartup.h"
-#include "GC.h"
 #include "Hooks.h"
 #include "Schedule.h"
 #include "StgMiscClosures.h"
 #include "Storage.h"
-#include "Evaluator.h"
+#include "Interpreter.h"
 #include "Exception.h"
 #include "Printer.h"
 #include "Main.h"
 #include "Stats.h"
 #include "Itimer.h"
 #include "Prelude.h"
+#ifdef PROFILING
+#include "Proftimer.h"
+#include "ProfHeap.h"
+#endif
 #if defined(GRAN) || defined(PAR)
 # include "GranSimRts.h"
 # include "GranSim.h"
 # include "HLC.h"
 #endif
 #include "Sparks.h"
+#include "Capability.h"
+#include "OSThreads.h"
+#include  "Task.h"
 
 #include <stdarg.h>
 
  *
  * These are the threads which clients have requested that we run.  
  *
- * In an SMP build, we might have several concurrent clients all
+ * In a 'threaded' build, we might have several concurrent clients all
  * waiting for results, and each one will wait on a condition variable
  * until the result is available.
  *
@@ -106,8 +138,8 @@ typedef struct StgMainThread_ {
   StgTSO *         tso;
   SchedulerStatus  stat;
   StgClosure **    ret;
-#ifdef SMP
-  pthread_cond_t wakeup;
+#if defined(RTS_SUPPORTS_THREADS)
+  Condition        wakeup;
 #endif
   struct StgMainThread_ *link;
 } StgMainThread;
@@ -126,7 +158,7 @@ StgTSO* ActiveTSO = NULL; /* for assigning system costs; GranSim-Light only */
 /* rtsTime TimeOfNextEvent, EndOfTimeSlice;            now in GranSim.c */
 
 /* 
-   In GranSim we have a runable and a blocked queue for each processor.
+   In GranSim we have a runnable and a blocked queue for each processor.
    In order to minimise code changes new arrays run_queue_hds/tls
    are created. run_queue_hd is then a short cut (macro) for
    run_queue_hds[CurrentProc] (see GranSim.h).
@@ -153,11 +185,12 @@ StgTSO *sleeping_queue;           /* perhaps replace with a hash table? */
  */
 StgTSO *all_threads;
 
-/* Threads suspended in _ccall_GC.
+/* When a thread performs a safe C call (_ccall_GC, using old
+ * terminology), it gets put on the suspended_ccalling_threads
+ * list. Used by the garbage collector.
  */
 static StgTSO *suspended_ccalling_threads;
 
-static void GetRoots(void);
 static StgTSO *threadStackOverflow(StgTSO *tso);
 
 /* KH: The following two flags are shared memory locations.  There is no need
@@ -197,18 +230,6 @@ StgThreadID next_thread_id = 1;
 
 #define MIN_STACK_WORDS (RESERVED_STACK_WORDS + sizeofW(StgStopFrame) + 2)
 
-/* Free capability list.
- * Locks required: sched_mutex.
- */
-#ifdef SMP
-//@cindex free_capabilities
-//@cindex n_free_capabilities
-Capability *free_capabilities; /* Available capabilities for running threads */
-nat n_free_capabilities;       /* total number of available capabilities */
-#else
-//@cindex MainRegTable
-Capability MainRegTable;       /* for non-SMP, we have one global capability */
-#endif
 
 #if defined(GRAN)
 StgTSO *CurrentTSO;
@@ -222,13 +243,6 @@ StgTSO dummy_tso;
 
 rtsBool ready_to_gc;
 
-/* All our current task ids, saved in case we need to kill them later.
- */
-#ifdef SMP
-//@cindex task_ids
-task_info *task_ids;
-#endif
-
 void            addToBlockedQueue ( StgTSO *tso );
 
 static void     schedule          ( void );
@@ -245,29 +259,31 @@ static void     detectBlackHoles  ( void );
 static void sched_belch(char *s, ...);
 #endif
 
-#ifdef SMP
-//@cindex sched_mutex
-//@cindex term_mutex
-//@cindex thread_ready_cond
-//@cindex gc_pending_cond
-pthread_mutex_t sched_mutex       = PTHREAD_MUTEX_INITIALIZER;
-pthread_mutex_t term_mutex        = PTHREAD_MUTEX_INITIALIZER;
-pthread_cond_t  thread_ready_cond = PTHREAD_COND_INITIALIZER;
-pthread_cond_t  gc_pending_cond   = PTHREAD_COND_INITIALIZER;
+#if defined(RTS_SUPPORTS_THREADS)
+/* ToDo: carefully document the invariants that go together
+ *       with these synchronisation objects.
+ */
+Mutex     sched_mutex       = INIT_MUTEX_VAR;
+Mutex     term_mutex        = INIT_MUTEX_VAR;
 
+# if defined(SMP)
+static Condition gc_pending_cond = INIT_COND_VAR;
 nat await_death;
-#endif
+# endif
+
+#endif /* RTS_SUPPORTS_THREADS */
 
 #if defined(PAR)
 StgTSO *LastTSO;
 rtsTime TimeOfLastYield;
+rtsBool emitSchedule = rtsTrue;
 #endif
 
 #if DEBUG
 char *whatNext_strs[] = {
   "ThreadEnterGHC",
   "ThreadRunGHC",
-  "ThreadEnterHugs",
+  "ThreadEnterInterp",
   "ThreadKilled",
   "ThreadComplete"
 };
@@ -281,12 +297,29 @@ char *threadReturnCode_strs[] = {
 };
 #endif
 
+#if defined(PAR)
+StgTSO * createSparkThread(rtsSpark spark);
+StgTSO * activateSpark (rtsSpark spark);  
+#endif
+
 /*
  * The thread state for the main thread.
 // ToDo: check whether not needed any more
 StgTSO   *MainTSO;
  */
 
+#if defined(PAR) || defined(RTS_SUPPORTS_THREADS)
+static void taskStart(void);
+static void
+taskStart(void)
+{
+  schedule();
+}
+#endif
+
+
+
+
 //@node Main scheduling loop, Suspend and Resume, Prototypes, Main scheduling code
 //@subsection Main scheduling loop
 
@@ -339,13 +372,25 @@ schedule( void )
   rtsSpark spark;
   StgTSO *tso;
   GlobalTaskId pe;
+  rtsBool receivedFinish = rtsFalse;
+# if defined(DEBUG)
+  nat tp_size, sp_size; // stats only
+# endif
 #endif
   rtsBool was_interrupted = rtsFalse;
   
   ACQUIRE_LOCK(&sched_mutex);
+#if defined(RTS_SUPPORTS_THREADS)
+  /* Check to see whether there are any worker threads
+     waiting to deposit external call results. If so,
+     yield our capability */
+  yieldToReturningWorker(&sched_mutex, cap);
 
-#if defined(GRAN)
+  waitForWorkCapability(&sched_mutex, &cap, rtsFalse);
+#endif
 
+#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],
@@ -370,8 +415,8 @@ schedule( void )
 
 #elif defined(PAR)
 
-  while (!GlobalStopPending) {          /* GlobalStopPending set in par_exit */
-
+  while (!receivedFinish) {    /* set by processMessages */
+                               /* when receiving PP_FINISH message         */ 
 #else
 
   while (1) {
@@ -396,7 +441,7 @@ schedule( void )
      * should be done more efficiently without a linear scan
      * of the main threads list, somehow...
      */
-#ifdef SMP
+#if defined(RTS_SUPPORTS_THREADS)
     { 
       StgMainThread *m, **prev;
       prev = &main_threads;
@@ -408,16 +453,17 @@ schedule( void )
          }
          *prev = m->link;
          m->stat = Success;
-         pthread_cond_broadcast(&m->wakeup);
+         broadcastCondition(&m->wakeup);
          break;
        case ThreadKilled:
+         if (m->ret) *(m->ret) = NULL;
          *prev = m->link;
          if (was_interrupted) {
            m->stat = Interrupted;
          } else {
            m->stat = Killed;
          }
-         pthread_cond_broadcast(&m->wakeup);
+         broadcastCondition(&m->wakeup);
          break;
        default:
          break;
@@ -425,7 +471,8 @@ schedule( void )
       }
     }
 
-#else
+#else /* not threaded */
+
 # if defined(PAR)
     /* in GUM do this only on the Main PE */
     if (IAmMainThread)
@@ -443,6 +490,7 @@ schedule( void )
          m->stat = Success;
          return;
        } else {
+         if (m->ret) { *(m->ret) = NULL; };
          if (was_interrupted) {
            m->stat = Interrupted;
          } else {
@@ -457,10 +505,13 @@ schedule( void )
     /* Top up the run queue from our spark pool.  We try to make the
      * number of threads in the run queue equal to the number of
      * free capabilities.
+     *
+     * Disable spark support in SMP for now, non-essential & requires
+     * a little bit of work to make it compile cleanly. -- sof 1/02.
      */
-#if defined(SMP)
+#if 0 /* defined(SMP) */
     {
-      nat n = n_free_capabilities;
+      nat n = getFreeCapabilities();
       StgTSO *tso = run_queue_hd;
 
       /* Count the run queue */
@@ -471,34 +522,34 @@ schedule( void )
 
       for (; n > 0; n--) {
        StgClosure *spark;
-       spark = findSpark();
+       spark = findSpark(rtsFalse);
        if (spark == NULL) {
          break; /* no more sparks in the pool */
        } else {
          /* I'd prefer this to be done in activateSpark -- HWL */
          /* tricky - it needs to hold the scheduler lock and
           * not try to re-acquire it -- SDM */
-         StgTSO *tso;
-         tso = createThread_(RtsFlags.GcFlags.initialStkSize, rtsTrue);
-         pushClosure(tso,spark);
-         PUSH_ON_RUN_QUEUE(tso);
-#ifdef PAR
-         advisory_thread_count++;
-#endif
-         
+         createSparkThread(spark);       
          IF_DEBUG(scheduler,
-                  sched_belch("turning spark of closure %p into a thread",
+                  sched_belch("==^^ turning spark of closure %p into a thread",
                               (StgClosure *)spark));
        }
       }
       /* We need to wake up the other tasks if we just created some
        * work for them.
        */
-      if (n_free_capabilities - n > 1) {
-         pthread_cond_signal(&thread_ready_cond);
+      if (getFreeCapabilities() - n > 1) {
+         signalCondition( &thread_ready_cond );
       }
     }
-#endif /* SMP */
+#endif // SMP
+
+    /* check for signals each time around the scheduler */
+#ifndef mingw32_TARGET_OS
+    if (signals_pending()) {
+      startSignalHandlers();
+    }
+#endif
 
     /* Check whether any waiting threads need to be woken up.  If the
      * run queue is empty, and there are no other tasks running, we
@@ -506,24 +557,16 @@ schedule( void )
      * ToDo: what if another client comes along & requests another
      * main thread?
      */
-    if (blocked_queue_hd != END_TSO_QUEUE || sleeping_queue != END_TSO_QUEUE) {
-      awaitEvent(
-          (run_queue_hd == END_TSO_QUEUE)
-#ifdef SMP
-       && (n_free_capabilities == RtsFlags.ParFlags.nNodes)
+    if ( !EMPTY_QUEUE(blocked_queue_hd) || !EMPTY_QUEUE(sleeping_queue) ) {
+      awaitEvent( EMPTY_RUN_QUEUE()
+#if defined(SMP)
+       && allFreeCapabilities()
 #endif
        );
     }
     /* we can be interrupted while waiting for I/O... */
     if (interrupted) continue;
 
-    /* check for signals each time around the scheduler */
-#ifndef mingw32_TARGET_OS
-    if (signals_pending()) {
-      start_signal_handlers();
-    }
-#endif
-
     /* 
      * Detect deadlock: when we have no threads to run, there are no
      * threads waiting on I/O or sleeping, and all the other tasks are
@@ -535,62 +578,115 @@ schedule( void )
      * If no threads are black holed, we have a deadlock situation, so
      * inform all the main threads.
      */
+#ifndef PAR
+    if (   EMPTY_RUN_QUEUE()
+       && EMPTY_QUEUE(blocked_queue_hd)
+       && EMPTY_QUEUE(sleeping_queue)
+#if defined(RTS_SUPPORTS_THREADS)
+       && EMPTY_QUEUE(suspended_ccalling_threads)
+#endif
 #ifdef SMP
-    if (blocked_queue_hd == END_TSO_QUEUE
-       && run_queue_hd == END_TSO_QUEUE
-       && sleeping_queue == END_TSO_QUEUE
-       && (n_free_capabilities == RtsFlags.ParFlags.nNodes))
+       && allFreeCapabilities()
+#endif
+       )
     {
-       IF_DEBUG(scheduler, sched_belch("deadlocked, checking for black holes..."));
-       detectBlackHoles();
-       if (run_queue_hd == END_TSO_QUEUE) {
-           StgMainThread *m;
-           for (m = main_threads; m != NULL; m = m->link) {
-               m->ret = NULL;
-               m->stat = Deadlock;
-               pthread_cond_broadcast(&m->wakeup);
+       IF_DEBUG(scheduler, sched_belch("deadlocked, forcing major GC..."));
+#if defined(THREADED_RTS)
+       /* and SMP mode ..? */
+       releaseCapability(cap);
+#endif
+       RELEASE_LOCK(&sched_mutex);
+       GarbageCollect(GetRoots,rtsTrue);
+       ACQUIRE_LOCK(&sched_mutex);
+       if (   EMPTY_QUEUE(blocked_queue_hd)
+           && EMPTY_RUN_QUEUE()
+           && EMPTY_QUEUE(sleeping_queue) ) {
+
+           IF_DEBUG(scheduler, sched_belch("still deadlocked, checking for black holes..."));
+           detectBlackHoles();
+
+           /* No black holes, so 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 progress).
+            */
+           if ( EMPTY_RUN_QUEUE() ) {
+               StgMainThread *m;
+#if defined(RTS_SUPPORTS_THREADS)
+               for (m = main_threads; m != NULL; m = m->link) {
+                   switch (m->tso->why_blocked) {
+                   case BlockedOnBlackHole:
+                       raiseAsync(m->tso, (StgClosure *)NonTermination_closure);
+                       break;
+                   case BlockedOnException:
+                   case BlockedOnMVar:
+                       raiseAsync(m->tso, (StgClosure *)Deadlock_closure);
+                       break;
+                   default:
+                       barf("deadlock: main thread blocked in a strange way");
+                   }
+               }
+#else
+               m = main_threads;
+               switch (m->tso->why_blocked) {
+               case BlockedOnBlackHole:
+                   raiseAsync(m->tso, (StgClosure *)NonTermination_closure);
+                   break;
+               case BlockedOnException:
+               case BlockedOnMVar:
+                   raiseAsync(m->tso, (StgClosure *)Deadlock_closure);
+                   break;
+               default:
+                   barf("deadlock: main thread blocked in a strange way");
+               }
+#endif
            }
-           main_threads = NULL;
-       }
-    }
-#else /* ! SMP */
-    if (blocked_queue_hd == END_TSO_QUEUE
-       && run_queue_hd == END_TSO_QUEUE
-       && sleeping_queue == END_TSO_QUEUE)
-    {
-       IF_DEBUG(scheduler, sched_belch("deadlocked, checking for black holes..."));
-       detectBlackHoles();
-       if (run_queue_hd == END_TSO_QUEUE) {
-           StgMainThread *m = main_threads;
-           m->ret = NULL;
-           m->stat = Deadlock;
-           main_threads = m->link;
-           return;
+#if defined(RTS_SUPPORTS_THREADS)
+           /* ToDo: revisit conditions (and mechanism) for shutting
+              down a multi-threaded world  */
+           if ( EMPTY_RUN_QUEUE() ) {
+             IF_DEBUG(scheduler, sched_belch("all done, i think...shutting down."));
+             shutdownHaskellAndExit(0);
+           }
+#endif
+           ASSERT( !EMPTY_RUN_QUEUE() );
        }
     }
+#elif defined(PAR)
+    /* ToDo: add deadlock detection in GUM (similar to SMP) -- HWL */
 #endif
 
-#ifdef SMP
+#if defined(SMP)
     /* If there's a GC pending, don't do anything until it has
      * completed.
      */
     if (ready_to_gc) {
       IF_DEBUG(scheduler,sched_belch("waiting for GC"));
-      pthread_cond_wait(&gc_pending_cond, &sched_mutex);
+      waitCondition( &gc_pending_cond, &sched_mutex );
     }
-    
+#endif    
+
+#if defined(RTS_SUPPORTS_THREADS)
     /* block until we've got a thread on the run queue and a free
      * capability.
+     *
      */
-    while (run_queue_hd == END_TSO_QUEUE || free_capabilities == NULL) {
-      IF_DEBUG(scheduler, sched_belch("waiting for work"));
-      pthread_cond_wait(&thread_ready_cond, &sched_mutex);
-      IF_DEBUG(scheduler, sched_belch("work now available"));
+    if ( EMPTY_RUN_QUEUE() ) {
+      /* Give up our capability */
+      releaseCapability(cap);
+      IF_DEBUG(scheduler, sched_belch("thread %d: waiting for work", osThreadId()));
+      waitForWorkCapability(&sched_mutex, &cap, rtsTrue);
+      IF_DEBUG(scheduler, sched_belch("thread %d: work now available", osThreadId()));
+#if 0
+      while ( EMPTY_RUN_QUEUE() ) {
+       waitForWorkCapability(&sched_mutex, &cap);
+       IF_DEBUG(scheduler, sched_belch("thread %d: work now available", osThreadId()));
+      }
+#endif
     }
 #endif
 
 #if defined(GRAN)
-
     if (RtsFlags.GranFlags.Light)
       GranSimLight_enter_system(event, &ActiveTSO); // adjust ActiveTSO etc
 
@@ -603,7 +699,7 @@ schedule( void )
     if (!RtsFlags.GranFlags.Light)
       handleIdlePEs();
 
-    IF_DEBUG(gran, fprintf(stderr, "GRAN: switch by event-type\n"))
+    IF_DEBUG(gran, fprintf(stderr, "GRAN: switch by event-type\n"));
 
     /* main event dispatcher in GranSim */
     switch (event->evttype) {
@@ -717,7 +813,7 @@ schedule( void )
 
     IF_DEBUG(gran, 
             fprintf(stderr, "GRAN: About to run current thread, which is\n");
-            G_TSO(t,5))
+            G_TSO(t,5));
 
     context_switch = 0; // turned on via GranYield, checking events and time slice
 
@@ -727,14 +823,13 @@ schedule( void )
     procStatus[CurrentProc] = Busy;
 
 #elif defined(PAR)
-
     if (PendingFetches != END_BF_QUEUE) {
         processFetches();
     }
 
     /* ToDo: phps merge with spark activation above */
     /* check whether we have local work and send requests if we have none */
-    if (run_queue_hd == END_TSO_QUEUE) {  /* no runnable threads */
+    if (EMPTY_RUN_QUEUE()) {  /* no runnable threads */
       /* :-[  no local threads => look out for local sparks */
       /* the spark pool for the current PE */
       pool = &(MainRegTable.rSparks); // generalise to cap = &MainRegTable
@@ -748,8 +843,8 @@ schedule( void )
         * to turn one of those pending sparks into a
         * thread... 
         */
-       
-       spark = findSpark();                /* get a spark */
+
+       spark = findSpark(rtsFalse);                /* get a spark */
        if (spark != (rtsSpark) NULL) {
          tso = activateSpark(spark);       /* turn the spark into a thread */
          IF_PAR_DEBUG(schedule,
@@ -766,9 +861,13 @@ schedule( void )
                             spark_queue_len(pool)));
          goto next_thread;
        }
-      } else  
+      }
+
+      /* If we still have no work we need to send a FISH to get a spark
+        from another PE 
+      */
+      if (EMPTY_RUN_QUEUE()) {
       /* =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
@@ -777,28 +876,48 @@ schedule( void )
         * we're hoping to see.  (Of course, we still have to
         * respond to other types of messages.)
         */
-       if (//!fishing &&  
-           outstandingFishes < RtsFlags.ParFlags.maxFishes ) { // &&
-         // (last_fish_arrived_at+FISH_DELAY < CURRENT_TIME)) {
-         /* fishing set in sendFish, processFish;
+       TIME now = msTime() /*CURRENT_TIME*/;
+       IF_PAR_DEBUG(verbose, 
+                    belch("--  now=%ld", now));
+       IF_PAR_DEBUG(verbose,
+                    if (outstandingFishes < RtsFlags.ParFlags.maxFishes &&
+                        (last_fish_arrived_at!=0 &&
+                         last_fish_arrived_at+RtsFlags.ParFlags.fishDelay > now)) {
+                      belch("--$$ delaying FISH until %ld (last fish %ld, delay %ld, now %ld)",
+                            last_fish_arrived_at+RtsFlags.ParFlags.fishDelay,
+                            last_fish_arrived_at,
+                            RtsFlags.ParFlags.fishDelay, now);
+                    });
+       
+       if (outstandingFishes < RtsFlags.ParFlags.maxFishes &&
+           (last_fish_arrived_at==0 ||
+            (last_fish_arrived_at+RtsFlags.ParFlags.fishDelay <= now))) {
+         /* outstandingFishes is set in sendFish, processFish;
             avoid flooding system with fishes via delay */
          pe = choosePE();
          sendFish(pe, mytid, 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++;
+         }
        }
-       
-       processMessages();
+      
+       receivedFinish = processMessages();
        goto next_thread;
-       // ReSchedule(0);
       }
     } else if (PacketsWaiting()) {  /* Look for incoming messages */
-      processMessages();
+      receivedFinish = processMessages();
     }
 
     /* Now we are sure that we have some work available */
     ASSERT(run_queue_hd != END_TSO_QUEUE);
+
     /* Take a thread from the run queue, if we have work */
     t = POP_RUN_QUEUE();  // take_off_run_queue(END_TSO_QUEUE);
+    IF_DEBUG(sanity,checkTSO(t));
 
     /* ToDo: write something to the log-file
     if (RTSflags.ParFlags.granSimStats && !sameThread)
@@ -809,17 +928,23 @@ schedule( void )
     /* the spark pool for the current PE */
     pool = &(MainRegTable.rSparks); // generalise to cap = &MainRegTable
 
-    IF_DEBUG(scheduler, belch("--^^ %d sparks on [%#x] (hd=%x; tl=%x; base=%x, lim=%x)", 
-                             spark_queue_len(pool), 
-                             CURRENT_PROC,
-                             pool->hd, pool->tl, pool->base, pool->lim));
-
-    IF_DEBUG(scheduler, belch("--== %d threads on [%#x] (hd=%x; tl=%x)", 
-                             run_queue_len(), CURRENT_PROC,
-                             run_queue_hd, run_queue_tl));
+    IF_DEBUG(scheduler, 
+            belch("--=^ %d threads, %d sparks on [%#x]", 
+                  run_queue_len(), spark_queue_len(pool), CURRENT_PROC));
+
+# if 1
+    if (0 && RtsFlags.ParFlags.ParStats.Full && 
+       t && LastTSO && t->id != LastTSO->id && 
+       LastTSO->why_blocked == NotBlocked && 
+       LastTSO->what_next != ThreadComplete) {
+      // if previously scheduled TSO not blocked we have to record the context switch
+      DumpVeryRawGranEvent(TimeOfLastYield, CURRENT_PROC, CURRENT_PROC,
+                          GR_DESCHEDULE, LastTSO, (StgClosure *)NULL, 0, 0);
+    }
 
-#if 0
-    if (t != LastTSO) {
+    if (RtsFlags.ParFlags.ParStats.Full && 
+       (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 
@@ -829,39 +954,35 @@ schedule( void )
       */
       DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC,
                       GR_SCHEDULE, t, (StgClosure *)NULL, 0, 0);
-      
+      emitSchedule = rtsFalse;
     }
-#endif
+     
+# endif
 #else /* !GRAN && !PAR */
   
-    /* grab a thread from the run queue
-     */
+    /* grab a thread from the run queue */
     ASSERT(run_queue_hd != END_TSO_QUEUE);
     t = POP_RUN_QUEUE();
+    // Sanity check the thread we're about to run.  This can be
+    // expensive if there is lots of thread switching going on...
     IF_DEBUG(sanity,checkTSO(t));
-
 #endif
     
-    /* grab a capability
-     */
-#ifdef SMP
-    cap = free_capabilities;
-    free_capabilities = cap->link;
-    n_free_capabilities--;
-#else
-    cap = &MainRegTable;
-#endif
-    
-    cap->rCurrentTSO = t;
+    grabCapability(&cap);
+    cap->r.rCurrentTSO = t;
     
     /* context switches are now initiated by the timer signal, unless
      * the user specified "context switch as often as possible", with
      * +RTS -C0
      */
-    if (RtsFlags.ConcFlags.ctxtSwitchTicks == 0
-       && (run_queue_hd != END_TSO_QUEUE
-           || blocked_queue_hd != END_TSO_QUEUE
-           || sleeping_queue != END_TSO_QUEUE))
+    if (
+#ifdef PROFILING
+       RtsFlags.ProfFlags.profileInterval == 0 ||
+#endif
+       (RtsFlags.ConcFlags.ctxtSwitchTicks == 0
+        && (run_queue_hd != END_TSO_QUEUE
+            || blocked_queue_hd != END_TSO_QUEUE
+            || sleeping_queue != END_TSO_QUEUE)))
        context_switch = 1;
     else
        context_switch = 0;
@@ -871,34 +992,28 @@ schedule( void )
     IF_DEBUG(scheduler, sched_belch("-->> Running TSO %ld (%p) %s ...", 
                              t->id, t, whatNext_strs[t->what_next]));
 
+#ifdef PROFILING
+    startHeapProfTimer();
+#endif
+
     /* +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ */
     /* Run the current thread 
      */
-    switch (cap->rCurrentTSO->what_next) {
+    switch (cap->r.rCurrentTSO->what_next) {
     case ThreadKilled:
     case ThreadComplete:
-      /* Thread already finished, return to scheduler. */
-      ret = ThreadFinished;
-      break;
+       /* Thread already finished, return to scheduler. */
+       ret = ThreadFinished;
+       break;
     case ThreadEnterGHC:
-      ret = StgRun((StgFunPtr) stg_enterStackTop, cap);
-      break;
+       ret = StgRun((StgFunPtr) stg_enterStackTop, &cap->r);
+       break;
     case ThreadRunGHC:
-      ret = StgRun((StgFunPtr) stg_returnToStackTop, cap);
-      break;
-    case ThreadEnterHugs:
-#ifdef INTERPRETER
-      {
-         StgClosure* c;
-        IF_DEBUG(scheduler,sched_belch("entering Hugs"));
-        c = (StgClosure *)(cap->rCurrentTSO->sp[0]);
-        cap->rCurrentTSO->sp += 1;
-        ret = enter(cap,c);
-         break;
-      }
-#else
-      barf("Panic: entered a BCO but no bytecode interpreter in this build");
-#endif
+       ret = StgRun((StgFunPtr) stg_returnToStackTop, &cap->r);
+       break;
+    case ThreadEnterInterp:
+       ret = interpretBCO(cap);
+       break;
     default:
       barf("schedule: invalid what_next field");
     }
@@ -906,28 +1021,89 @@ schedule( void )
     
     /* Costs for the scheduler are assigned to CCS_SYSTEM */
 #ifdef PROFILING
+    stopHeapProfTimer();
     CCCS = CCS_SYSTEM;
 #endif
     
     ACQUIRE_LOCK(&sched_mutex);
 
 #ifdef SMP
-    IF_DEBUG(scheduler,fprintf(stderr,"scheduler (task %ld): ", pthread_self()););
+    IF_DEBUG(scheduler,fprintf(stderr,"scheduler (task %ld): ", osThreadId()););
 #elif !defined(GRAN) && !defined(PAR)
     IF_DEBUG(scheduler,fprintf(stderr,"scheduler: "););
 #endif
-    t = cap->rCurrentTSO;
+    t = cap->r.rCurrentTSO;
     
 #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; //(ret == ThreadBlocked) ? END_TSO_QUEUE : t; 
+    LastTSO = t; 
     TimeOfLastYield = CURRENT_TIME;
 #endif
 
     switch (ret) {
     case HeapOverflow:
+#if defined(GRAN)
+      IF_DEBUG(gran, DumpGranEvent(GR_DESCHEDULE, t));
+      globalGranStats.tot_heapover++;
+#elif defined(PAR)
+      globalParStats.tot_heapover++;
+#endif
+
+      // did the task ask for a large block?
+      if (cap->r.rHpAlloc > BLOCK_SIZE_W) {
+         // if so, get one and push it on the front of the nursery.
+         bdescr *bd;
+         nat blocks;
+         
+         blocks = (nat)BLOCK_ROUND_UP(cap->r.rHpAlloc * sizeof(W_)) / BLOCK_SIZE;
+
+         IF_DEBUG(scheduler,belch("--<< thread %ld (%p; %s) stopped: requesting a large block (size %d)", 
+                                  t->id, t,
+                                  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) {
+
+             alloc_blocks += blocks;
+             bd = allocGroup( 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 {
+                 ASSERT(g0s0->blocks == cap->r.rCurrentNursery &&
+                        g0s0->blocks == cap->r.rNursery);
+                 cap->r.rNursery = g0s0->blocks = bd;
+             }           
+             cap->r.rCurrentNursery->u.back = bd;
+
+             // initialise it as a nursery block
+             bd->step = g0s0;
+             bd->gen_no = 0;
+             bd->flags = 0;
+             bd->free = bd->start;
+
+             // don't forget to update the block count in g0s0.
+             g0s0->n_blocks += blocks;
+             ASSERT(countBlocks(g0s0->blocks) == g0s0->n_blocks);
+
+             // 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.
+             PUSH_ON_RUN_QUEUE(t);
+             break;
+         }
+      }
+
       /* 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.
@@ -937,6 +1113,15 @@ schedule( void )
       threadPaused(t);
 #if defined(GRAN)
       ASSERT(!is_on_queue(t,CurrentProc));
+#elif defined(PAR)
+      /* 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
       
       ready_to_gc = rtsTrue;
@@ -946,6 +1131,15 @@ schedule( void )
       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
       IF_DEBUG(scheduler,belch("--<< thread %ld (%p; %s) stopped, StackOverflow", 
                               t->id, t, whatNext_strs[t->what_next]));
       /* just adjust the stack for this thread, then pop it back
@@ -977,8 +1171,9 @@ schedule( void )
               DumpGranEvent(GR_DESCHEDULE, t));
       globalGranStats.tot_yields++;
 #elif defined(PAR)
-      IF_DEBUG(par, 
-              DumpGranEvent(GR_DESCHEDULE, t));
+      // IF_DEBUG(par, 
+      // DumpGranEvent(GR_DESCHEDULE, t);
+      globalParStats.tot_yields++;
 #endif
       /* 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
@@ -986,7 +1181,7 @@ schedule( void )
        * GC is finished.
        */
       IF_DEBUG(scheduler,
-               if (t->what_next == ThreadEnterHugs) {
+               if (t->what_next == ThreadEnterInterp) {
                   /* ToDo: or maybe a timer expired when we were in Hugs?
                    * or maybe someone hit ctrl-C
                     */
@@ -1011,7 +1206,18 @@ schedule( void )
               //belch("&& Doing sanity check on all ThreadQueues (and their TSOs).");
               checkThreadQsSanity(rtsTrue));
 #endif
+#if defined(PAR)
+      if (RtsFlags.ParFlags.doFairScheduling) { 
+       /* this does round-robin scheduling; good for concurrency */
+       APPEND_TO_RUN_QUEUE(t);
+      } else {
+       /* this does unfair scheduling; good for parallelism */
+       PUSH_ON_RUN_QUEUE(t);
+      }
+#else
+      /* this does round-robin scheduling; good for concurrency */
       APPEND_TO_RUN_QUEUE(t);
+#endif
 #if defined(GRAN)
       /* add a ContinueThread event to actually process the thread */
       new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc],
@@ -1020,7 +1226,7 @@ schedule( void )
       IF_GRAN_DEBUG(bq, 
               belch("GRAN: eventq and runnableq after adding yielded thread to queue again:");
               G_EVENTQ(0);
-              G_CURR_THREADQ(0))
+              G_CURR_THREADQ(0));
 #endif /* GRAN */
       break;
       
@@ -1046,16 +1252,19 @@ schedule( void )
        procStatus[CurrentProc] = Idle;
       */
 #elif defined(PAR)
-      IF_DEBUG(par, 
-              DumpGranEvent(GR_DESCHEDULE, t)); 
+      IF_DEBUG(scheduler,
+              belch("--<< thread %ld (%p; %s) stopped, blocking on node %p with BQ: ", 
+                    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);
 
-      IF_DEBUG(scheduler,
-              belch("--<< thread %ld (%p; %s) stopped, blocking on node %p with BQ: ", 
-                              t->id, t, whatNext_strs[t->what_next], t->block_info.closure);
-              if (t->block_info.closure!=(StgClosure*)NULL) print_bq(t->block_info.closure));
+      /* whatever we schedule next, we must log that schedule */
+      emitSchedule = rtsTrue;
 
 #else /* !GRAN */
       /* don't need to do anything.  Either the thread is blocked on
@@ -1089,8 +1298,17 @@ schedule( void )
 #if defined(GRAN)
       endThread(t, CurrentProc); // clean-up the thread
 #elif defined(PAR)
+      /* For now all are advisory -- HWL */
+      //if(t->priority==AdvisoryPriority) ??
       advisory_thread_count--;
-      if (RtsFlags.ParFlags.ParStats.Full) 
+      
+# ifdef DIST
+      if(t->dist.priority==RevalPriority)
+       FinishReval(t);
+# endif
+      
+      if (RtsFlags.ParFlags.ParStats.Full &&
+         !RtsFlags.ParFlags.ParStats.Suppressed) 
        DumpEndEvent(CURRENT_PROC, t, rtsFalse /* not mandatory */);
 #endif
       break;
@@ -1099,30 +1317,37 @@ schedule( void )
       barf("schedule: invalid thread return code %d", (int)ret);
     }
     
-#ifdef SMP
-    cap->link = free_capabilities;
-    free_capabilities = cap;
-    n_free_capabilities++;
+#if defined(RTS_SUPPORTS_THREADS)
+    /* I don't understand what this re-grab is doing -- sof */
+    grabCapability(&cap);
 #endif
 
+#ifdef PROFILING
+    if (RtsFlags.ProfFlags.profileInterval==0 || performHeapProfile) {
+       GarbageCollect(GetRoots, rtsTrue);
+       heapCensus();
+       performHeapProfile = rtsFalse;
+       ready_to_gc = rtsFalse; // we already GC'd
+    }
+#endif
+
+    if (ready_to_gc 
 #ifdef SMP
-    if (ready_to_gc && n_free_capabilities == RtsFlags.ParFlags.nNodes) 
-#else
-    if (ready_to_gc) 
+       && allFreeCapabilities() 
 #endif
-      {
+       ) {
       /* 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.
        */
-#ifdef SMP
+#if defined(RTS_SUPPORTS_THREADS)
       IF_DEBUG(scheduler,sched_belch("doing GC"));
 #endif
       GarbageCollect(GetRoots,rtsFalse);
       ready_to_gc = rtsFalse;
 #ifdef SMP
-      pthread_cond_broadcast(&gc_pending_cond);
+      broadcastCondition(&gc_pending_cond);
 #endif
 #if defined(GRAN)
       /* add a ContinueThread event to continue execution of current thread */
@@ -1132,27 +1357,25 @@ schedule( void )
       IF_GRAN_DEBUG(bq, 
               fprintf(stderr, "GRAN: eventq and runnableq after Garbage collection:\n");
               G_EVENTQ(0);
-              G_CURR_THREADQ(0))
+              G_CURR_THREADQ(0));
 #endif /* GRAN */
     }
+
 #if defined(GRAN)
   next_thread:
     IF_GRAN_DEBUG(unused,
                  print_eventq(EventHd));
 
     event = get_next_event();
-
 #elif defined(PAR)
   next_thread:
     /* ToDo: wait for next message to arrive rather than busy wait */
-
-#else /* GRAN */
-  /* not any more
-  next_thread:
-    t = take_off_run_queue(END_TSO_QUEUE);
-  */
 #endif /* GRAN */
+
   } /* end of while(1) */
+
+  IF_PAR_DEBUG(verbose,
+              belch("== Leaving schedule() after having received Finish"));
 }
 
 /* ---------------------------------------------------------------------------
@@ -1164,15 +1387,18 @@ schedule( void )
    
 void deleteAllThreads ( void )
 {
-  StgTSO* t;
+  StgTSO* t, *next;
   IF_DEBUG(scheduler,sched_belch("deleting all threads"));
-  for (t = run_queue_hd; t != END_TSO_QUEUE; t = t->link) {
+  for (t = run_queue_hd; t != END_TSO_QUEUE; t = next) {
+      next = t->link;
       deleteThread(t);
   }
-  for (t = blocked_queue_hd; t != END_TSO_QUEUE; t = t->link) {
+  for (t = blocked_queue_hd; t != END_TSO_QUEUE; t = next) {
+      next = t->link;
       deleteThread(t);
   }
-  for (t = sleeping_queue; t != END_TSO_QUEUE; t = t->link) {
+  for (t = sleeping_queue; t != END_TSO_QUEUE; t = next) {
+      next = t->link;
       deleteThread(t);
   }
   run_queue_hd = run_queue_tl = END_TSO_QUEUE;
@@ -1182,6 +1408,7 @@ void deleteAllThreads ( void )
 
 /* startThread and  insertThread are now in GranSim.c -- HWL */
 
+
 //@node Suspend and Resume, Run queue code, Main scheduling loop, Main scheduling code
 //@subsection Suspend and Resume
 
@@ -1201,40 +1428,66 @@ void deleteAllThreads ( void )
  * ------------------------------------------------------------------------- */
    
 StgInt
-suspendThread( Capability *cap )
+suspendThread( StgRegTable *reg )
 {
   nat tok;
+  Capability *cap;
+
+  /* assume that *reg is a pointer to the StgRegTable part
+   * of a Capability.
+   */
+  cap = (Capability *)((void *)reg - sizeof(StgFunTable));
 
   ACQUIRE_LOCK(&sched_mutex);
 
   IF_DEBUG(scheduler,
-          sched_belch("thread %d did a _ccall_gc", cap->rCurrentTSO->id));
+          sched_belch("thread %d did a _ccall_gc", cap->r.rCurrentTSO->id));
+
+  threadPaused(cap->r.rCurrentTSO);
+  cap->r.rCurrentTSO->link = suspended_ccalling_threads;
+  suspended_ccalling_threads = cap->r.rCurrentTSO;
 
-  threadPaused(cap->rCurrentTSO);
-  cap->rCurrentTSO->link = suspended_ccalling_threads;
-  suspended_ccalling_threads = cap->rCurrentTSO;
+#if defined(RTS_SUPPORTS_THREADS)
+  cap->r.rCurrentTSO->why_blocked  = BlockedOnCCall;
+#endif
 
   /* Use the thread ID as the token; it should be unique */
-  tok = cap->rCurrentTSO->id;
+  tok = cap->r.rCurrentTSO->id;
 
-#ifdef SMP
-  cap->link = free_capabilities;
-  free_capabilities = cap;
-  n_free_capabilities++;
+  /* Hand back capability */
+  releaseCapability(cap);
+  
+#if defined(RTS_SUPPORTS_THREADS) && !defined(SMP)
+  /* Preparing to leave the RTS, so ensure there's a native thread/task
+     waiting to take over.
+     
+     ToDo: optimise this and only create a new task if there's a need
+     for one (i.e., if there's only one Concurrent Haskell thread alive,
+     there's no need to create a new task).
+  */
+  IF_DEBUG(scheduler, sched_belch("worker thread (%d): leaving RTS", tok));
+  startTask(taskStart);
 #endif
 
+  THREAD_RUNNABLE();
   RELEASE_LOCK(&sched_mutex);
   return tok; 
 }
 
-Capability *
+StgRegTable *
 resumeThread( StgInt tok )
 {
   StgTSO *tso, **prev;
   Capability *cap;
 
-  ACQUIRE_LOCK(&sched_mutex);
+#if defined(RTS_SUPPORTS_THREADS)
+  /* Wait for permission to re-enter the RTS with the result. */
+  grabReturnCapability(&sched_mutex, &cap);
+#else
+  grabCapability(&cap);
+#endif
 
+  /* Remove the thread off of the suspended list */
   prev = &suspended_ccalling_threads;
   for (tso = suspended_ccalling_threads; 
        tso != END_TSO_QUEUE; 
@@ -1248,24 +1501,13 @@ resumeThread( StgInt tok )
     barf("resumeThread: thread not found");
   }
   tso->link = END_TSO_QUEUE;
-
-#ifdef SMP
-  while (free_capabilities == NULL) {
-    IF_DEBUG(scheduler, sched_belch("waiting to resume"));
-    pthread_cond_wait(&thread_ready_cond, &sched_mutex);
-    IF_DEBUG(scheduler, sched_belch("resuming thread %d", tso->id));
-  }
-  cap = free_capabilities;
-  free_capabilities = cap->link;
-  n_free_capabilities--;
-#else  
-  cap = &MainRegTable;
-#endif
-
-  cap->rCurrentTSO = tso;
+  /* Reset blocking status */
+  tso->why_blocked  = NotBlocked;
 
   RELEASE_LOCK(&sched_mutex);
-  return cap;
+
+  cap->r.rCurrentTSO = tso;
+  return &cap->r;
 }
 
 
@@ -1292,6 +1534,16 @@ int cmp_thread(const StgTSO *tso1, const StgTSO *tso2)
 }
 
 /* ---------------------------------------------------------------------------
+ * Fetching the ThreadID from an StgTSO.
+ *
+ * This is used in the implementation of Show for ThreadIds.
+ * ------------------------------------------------------------------------ */
+int rts_getThreadId(const StgTSO *tso) 
+{
+  return tso->id;
+}
+
+/* ---------------------------------------------------------------------------
    Create a new thread.
 
    The new thread starts with the given stack size.  Before the
@@ -1357,7 +1609,7 @@ createThread_(nat size, rtsBool have_lock)
   stack_size = size - TSO_STRUCT_SIZEW;
 
   tso = (StgTSO *)allocate(size);
-  TICK_ALLOC_TSO(size-TSO_STRUCT_SIZEW, 0);
+  TICK_ALLOC_TSO(stack_size, 0);
 
   SET_HDR(tso, &stg_TSO_info, CCS_SYSTEM);
 #if defined(GRAN)
@@ -1401,8 +1653,14 @@ createThread_(nat size, rtsBool have_lock)
    */
 #endif
 
-#if defined(GRAN) || defined(PAR)
-  DumpGranEvent(GR_START,tso);
+#if defined(GRAN) 
+  if (RtsFlags.GranFlags.GranSimStats.Full) 
+    DumpGranEvent(GR_START,tso);
+#elif defined(PAR)
+  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.
@@ -1410,6 +1668,10 @@ createThread_(nat size, rtsBool have_lock)
   tso->global_link = all_threads;
   all_threads = tso;
 
+#if defined(DIST)
+  tso->dist.priority = MandatoryPriority; //by default that is...
+#endif
+
 #if defined(GRAN)
   tso->gran.pri = pri;
 # if defined(DEBUG)
@@ -1458,6 +1720,13 @@ createThread_(nat size, rtsBool have_lock)
   globalGranStats.threads_created_on_PE[CurrentProc]++;
   globalGranStats.tot_sq_len += spark_queue_len(CurrentProc);
   globalGranStats.tot_sq_probes++;
+#elif defined(PAR)
+  // collect parallel global statistics (currently done together with GC stats)
+  if (RtsFlags.ParFlags.ParStats.Global &&
+      RtsFlags.GcFlags.giveStats > NO_GC_STATS) {
+    //fprintf(stderr, "Creating thread %d @ %11.2f\n", tso->id, usertime()); 
+    globalParStats.tot_threads_created++;
+  }
 #endif 
 
 #if defined(GRAN)
@@ -1475,6 +1744,36 @@ createThread_(nat size, rtsBool have_lock)
   return tso;
 }
 
+#if defined(PAR)
+/* RFP:
+   all parallel thread creation calls should fall through the following routine.
+*/
+StgTSO *
+createSparkThread(rtsSpark spark) 
+{ StgTSO *tso;
+  ASSERT(spark != (rtsSpark)NULL);
+  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, rtsTrue);
+    if (tso==END_TSO_QUEUE)    
+      barf("createSparkThread: Cannot create TSO");
+#if defined(DIST)
+    tso->priority = AdvisoryPriority;
+#endif
+    pushClosure(tso,spark);
+    PUSH_ON_RUN_QUEUE(tso);
+    advisory_thread_count++;    
+  }
+  return tso;
+}
+#endif
+
 /*
   Turn a spark into a thread.
   ToDo: fix for SMP (needs to acquire SCHED_MUTEX!)
@@ -1485,22 +1784,13 @@ StgTSO *
 activateSpark (rtsSpark spark) 
 {
   StgTSO *tso;
-  
-  ASSERT(spark != (rtsSpark)NULL);
-  tso = createThread_(RtsFlags.GcFlags.initialStkSize, rtsTrue);
-  if (tso!=END_TSO_QUEUE) {
-    pushClosure(tso,spark);
-    PUSH_ON_RUN_QUEUE(tso);
-    advisory_thread_count++;
 
-    if (RtsFlags.ParFlags.ParStats.Full) {
-      //ASSERT(run_queue_hd == END_TSO_QUEUE); // I think ...
-      IF_PAR_DEBUG(verbose,
-                  belch("==^^ activateSpark: turning spark of closure %p (%s) into a thread",
-                        (StgClosure *)spark, info_type((StgClosure *)spark)));
-    }
-  } else {
-    barf("activateSpark: Cannot create TSO");
+  tso = createSparkThread(spark);
+  if (RtsFlags.ParFlags.ParStats.Full) {   
+    //ASSERT(run_queue_hd == END_TSO_QUEUE); // I think ...
+    IF_PAR_DEBUG(verbose,
+                belch("==^^ activateSpark: turning spark of closure %p (%s) into a thread",
+                      (StgClosure *)spark, info_type((StgClosure *)spark)));
   }
   // ToDo: fwd info on local/global spark to thread -- HWL
   // tso->gran.exported =  spark->exported;
@@ -1521,14 +1811,16 @@ activateSpark (rtsSpark spark)
  * on this thread's stack before the scheduler is invoked.
  * ------------------------------------------------------------------------ */
 
+static void scheduleThread_ (StgTSO* tso, rtsBool createTask);
+
 void
-scheduleThread(StgTSO *tso)
+scheduleThread_(StgTSO *tso
+              , rtsBool createTask
+#if !defined(THREADED_RTS)
+                STG_UNUSED
+#endif
+             )
 {
-  if (tso==END_TSO_QUEUE){    
-    schedule();
-    return;
-  }
-
   ACQUIRE_LOCK(&sched_mutex);
 
   /* Put the new thread on the head of the runnable queue.  The caller
@@ -1537,6 +1829,14 @@ scheduleThread(StgTSO *tso)
    * soon as we release the scheduler lock below.
    */
   PUSH_ON_RUN_QUEUE(tso);
+#if defined(THREADED_RTS)
+  /* If main() is scheduling a thread, don't bother creating a 
+   * new task.
+   */
+  if ( createTask ) {
+    startTask(taskStart);
+  }
+#endif
   THREAD_RUNNABLE();
 
 #if 0
@@ -1545,21 +1845,15 @@ scheduleThread(StgTSO *tso)
   RELEASE_LOCK(&sched_mutex);
 }
 
-/* ---------------------------------------------------------------------------
- * startTasks()
- *
- * Start up Posix threads to run each of the scheduler tasks.
- * I believe the task ids are not needed in the system as defined.
- *  KH @ 25/10/99
- * ------------------------------------------------------------------------ */
+void scheduleThread(StgTSO* tso)
+{
+  return scheduleThread_(tso, rtsFalse);
+}
 
-#if defined(PAR) || defined(SMP)
-void *
-taskStart( void *arg STG_UNUSED )
+void scheduleExtThread(StgTSO* tso)
 {
-  rts_evalNothing(NULL);
+  return scheduleThread_(tso, rtsTrue);
 }
-#endif
 
 /* ---------------------------------------------------------------------------
  * initScheduler()
@@ -1568,7 +1862,6 @@ taskStart( void *arg STG_UNUSED )
  * queues contained any threads, they'll be garbage collected at the
  * next pass.
  *
- * This now calls startTasks(), so should only be called once!  KH @ 25/10/99
  * ------------------------------------------------------------------------ */
 
 #ifdef SMP
@@ -1579,11 +1872,10 @@ term_handler(int sig STG_UNUSED)
   ACQUIRE_LOCK(&term_mutex);
   await_death--;
   RELEASE_LOCK(&term_mutex);
-  pthread_exit(NULL);
+  shutdownThread();
 }
 #endif
 
-//@cindex initScheduler
 void 
 initScheduler(void)
 {
@@ -1616,14 +1908,26 @@ initScheduler(void)
 
   RtsFlags.ConcFlags.ctxtSwitchTicks =
       RtsFlags.ConcFlags.ctxtSwitchTime / TICK_MILLISECS;
+      
+#if defined(RTS_SUPPORTS_THREADS)
+  /* Initialise the mutex and condition variables used by
+   * the scheduler. */
+  initMutex(&sched_mutex);
+  initMutex(&term_mutex);
+
+  initCondition(&thread_ready_cond);
+#endif
+  
+#if defined(SMP)
+  initCondition(&gc_pending_cond);
+#endif
 
-#ifdef INTERPRETER
-  ecafList = END_ECAF_LIST;
-  clearECafTable();
+#if defined(RTS_SUPPORTS_THREADS)
+  ACQUIRE_LOCK(&sched_mutex);
 #endif
 
   /* Install the SIGHUP handler */
-#ifdef SMP
+#if defined(SMP)
   {
     struct sigaction action,oact;
 
@@ -1636,92 +1940,36 @@ initScheduler(void)
   }
 #endif
 
-#ifdef SMP
-  /* Allocate N Capabilities */
-  {
-    nat i;
-    Capability *cap, *prev;
-    cap  = NULL;
-    prev = NULL;
-    for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
-      cap = stgMallocBytes(sizeof(Capability), "initScheduler:capabilities");
-      cap->link = prev;
-      prev = cap;
-    }
-    free_capabilities = cap;
-    n_free_capabilities = RtsFlags.ParFlags.nNodes;
-  }
-  IF_DEBUG(scheduler,fprintf(stderr,"scheduler: Allocated %d capabilities\n",
-                            n_free_capabilities););
+  /* 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();
+  
+#if defined(RTS_SUPPORTS_THREADS)
+    /* start our haskell execution tasks */
+# if defined(SMP)
+    startTaskManager(RtsFlags.ParFlags.nNodes, taskStart);
+# else
+    startTaskManager(0,taskStart);
+# endif
 #endif
 
-#if defined(SMP) || defined(PAR)
+#if /* defined(SMP) ||*/ defined(PAR)
   initSparkPools();
 #endif
-}
 
-#ifdef SMP
-void
-startTasks( void )
-{
-  nat i;
-  int r;
-  pthread_t tid;
-  
-  /* make some space for saving all the thread ids */
-  task_ids = stgMallocBytes(RtsFlags.ParFlags.nNodes * sizeof(task_info),
-                           "initScheduler:task_ids");
-  
-  /* and create all the threads */
-  for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
-    r = pthread_create(&tid,NULL,taskStart,NULL);
-    if (r != 0) {
-      barf("startTasks: Can't create new Posix thread");
-    }
-    task_ids[i].id = tid;
-    task_ids[i].mut_time = 0.0;
-    task_ids[i].mut_etime = 0.0;
-    task_ids[i].gc_time = 0.0;
-    task_ids[i].gc_etime = 0.0;
-    task_ids[i].elapsedtimestart = elapsedtime();
-    IF_DEBUG(scheduler,fprintf(stderr,"scheduler: Started task: %ld\n",tid););
-  }
-}
+#if defined(RTS_SUPPORTS_THREADS)
+  RELEASE_LOCK(&sched_mutex);
 #endif
 
+}
+
 void
 exitScheduler( void )
 {
-#ifdef SMP
-  nat i;
-
-  /* Don't want to use pthread_cancel, since we'd have to install
-   * these silly exception handlers (pthread_cleanup_{push,pop}) around
-   * all our locks.
-   */
-#if 0
-  /* Cancel all our tasks */
-  for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
-    pthread_cancel(task_ids[i].id);
-  }
-  
-  /* Wait for all the tasks to terminate */
-  for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
-    IF_DEBUG(scheduler,fprintf(stderr,"scheduler: waiting for task %ld\n", 
-                              task_ids[i].id));
-    pthread_join(task_ids[i].id, NULL);
-  }
-#endif
-
-  /* Send 'em all a SIGHUP.  That should shut 'em up.
-   */
-  await_death = RtsFlags.ParFlags.nNodes;
-  for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
-    pthread_kill(task_ids[i].id,SIGTERM);
-  }
-  while (await_death > 0) {
-    sched_yield();
-  }
+#if defined(RTS_SUPPORTS_THREADS)
+  stopTaskManager();
 #endif
 }
 
@@ -1769,13 +2017,13 @@ finishAllThreads ( void )
 {
    do {
       while (run_queue_hd != END_TSO_QUEUE) {
-         waitThread ( run_queue_hd, NULL );
+         waitThread ( run_queue_hd, NULL);
       }
       while (blocked_queue_hd != END_TSO_QUEUE) {
-         waitThread ( blocked_queue_hd, NULL );
+         waitThread ( blocked_queue_hd, NULL);
       }
       while (sleeping_queue != END_TSO_QUEUE) {
-         waitThread ( blocked_queue_hd, NULL );
+         waitThread ( blocked_queue_hd, NULL);
       }
    } while 
       (blocked_queue_hd != END_TSO_QUEUE || 
@@ -1785,6 +2033,21 @@ finishAllThreads ( void )
 
 SchedulerStatus
 waitThread(StgTSO *tso, /*out*/StgClosure **ret)
+{ 
+#if defined(THREADED_RTS)
+  return waitThread_(tso,ret, rtsFalse);
+#else
+  return waitThread_(tso,ret);
+#endif
+}
+
+SchedulerStatus
+waitThread_(StgTSO *tso,
+           /*out*/StgClosure **ret
+#if defined(THREADED_RTS)
+           , rtsBool blockWaiting
+#endif
+          )
 {
   StgMainThread *m;
   SchedulerStatus stat;
@@ -1796,20 +2059,34 @@ waitThread(StgTSO *tso, /*out*/StgClosure **ret)
   m->tso = tso;
   m->ret = ret;
   m->stat = NoStatus;
-#ifdef SMP
-  pthread_cond_init(&m->wakeup, NULL);
+#if defined(RTS_SUPPORTS_THREADS)
+  initCondition(&m->wakeup);
 #endif
 
   m->link = main_threads;
   main_threads = m;
 
-  IF_DEBUG(scheduler, fprintf(stderr, "scheduler: new main thread (%d)\n", 
-                             m->tso->id));
+  IF_DEBUG(scheduler, sched_belch("== scheduler: new main thread (%d)\n", m->tso->id));
 
-#ifdef SMP
-  do {
-    pthread_cond_wait(&m->wakeup, &sched_mutex);
-  } while (m->stat == NoStatus);
+#if defined(RTS_SUPPORTS_THREADS)
+
+# if defined(THREADED_RTS)
+  if (!blockWaiting) {
+    /* In the threaded case, the OS thread that called main()
+     * gets to enter the RTS directly without going via another
+     * task/thread.
+     */
+    RELEASE_LOCK(&sched_mutex);
+    schedule();
+    ASSERT(m->stat != NoStatus);
+  } else 
+# endif
+  {
+    IF_DEBUG(scheduler, sched_belch("sfoo"));
+    do {
+      waitCondition(&m->wakeup, &sched_mutex);
+    } while (m->stat == NoStatus);
+  }
 #elif defined(GRAN)
   /* GranSim specific init */
   CurrentTSO = m->tso;                // the TSO to run
@@ -1818,21 +2095,25 @@ waitThread(StgTSO *tso, /*out*/StgClosure **ret)
 
   schedule();
 #else
+  RELEASE_LOCK(&sched_mutex);
   schedule();
   ASSERT(m->stat != NoStatus);
 #endif
 
   stat = m->stat;
 
-#ifdef SMP
-  pthread_cond_destroy(&m->wakeup);
+#if defined(RTS_SUPPORTS_THREADS)
+  closeCondition(&m->wakeup);
 #endif
 
-  IF_DEBUG(scheduler, fprintf(stderr, "scheduler: main thread (%d) finished\n", 
+  IF_DEBUG(scheduler, fprintf(stderr, "== scheduler: main thread (%d) finished\n", 
                              m->tso->id));
   free(m);
 
-  RELEASE_LOCK(&sched_mutex);
+#if defined(THREADED_RTS)
+  if (blockWaiting) 
+#endif
+    RELEASE_LOCK(&sched_mutex);
 
   return stat;
 }
@@ -1957,7 +2238,8 @@ take_off_run_queue(StgTSO *tso) {
        KH @ 25/10/99
 */
 
-static void GetRoots(void)
+void
+GetRoots(evac_fn evac)
 {
   StgMainThread *m;
 
@@ -1966,16 +2248,16 @@ static void GetRoots(void)
     nat i;
     for (i=0; i<=RtsFlags.GranFlags.proc; i++) {
       if ((run_queue_hds[i] != END_TSO_QUEUE) && ((run_queue_hds[i] != NULL)))
-       run_queue_hds[i]    = (StgTSO *)MarkRoot((StgClosure *)run_queue_hds[i]);
+         evac((StgClosure **)&run_queue_hds[i]);
       if ((run_queue_tls[i] != END_TSO_QUEUE) && ((run_queue_tls[i] != NULL)))
-       run_queue_tls[i]    = (StgTSO *)MarkRoot((StgClosure *)run_queue_tls[i]);
+         evac((StgClosure **)&run_queue_tls[i]);
       
       if ((blocked_queue_hds[i] != END_TSO_QUEUE) && ((blocked_queue_hds[i] != NULL)))
-       blocked_queue_hds[i] = (StgTSO *)MarkRoot((StgClosure *)blocked_queue_hds[i]);
+         evac((StgClosure **)&blocked_queue_hds[i]);
       if ((blocked_queue_tls[i] != END_TSO_QUEUE) && ((blocked_queue_tls[i] != NULL)))
-       blocked_queue_tls[i] = (StgTSO *)MarkRoot((StgClosure *)blocked_queue_tls[i]);
+         evac((StgClosure **)&blocked_queue_tls[i]);
       if ((ccalling_threadss[i] != END_TSO_QUEUE) && ((ccalling_threadss[i] != NULL)))
-       ccalling_threadss[i] = (StgTSO *)MarkRoot((StgClosure *)ccalling_threadss[i]);
+         evac((StgClosure **)&ccalling_threads[i]);
     }
   }
 
@@ -1983,31 +2265,31 @@ static void GetRoots(void)
 
 #else /* !GRAN */
   if (run_queue_hd != END_TSO_QUEUE) {
-    ASSERT(run_queue_tl != END_TSO_QUEUE);
-    run_queue_hd      = (StgTSO *)MarkRoot((StgClosure *)run_queue_hd);
-    run_queue_tl      = (StgTSO *)MarkRoot((StgClosure *)run_queue_tl);
+      ASSERT(run_queue_tl != END_TSO_QUEUE);
+      evac((StgClosure **)&run_queue_hd);
+      evac((StgClosure **)&run_queue_tl);
   }
-
+  
   if (blocked_queue_hd != END_TSO_QUEUE) {
-    ASSERT(blocked_queue_tl != END_TSO_QUEUE);
-    blocked_queue_hd  = (StgTSO *)MarkRoot((StgClosure *)blocked_queue_hd);
-    blocked_queue_tl  = (StgTSO *)MarkRoot((StgClosure *)blocked_queue_tl);
+      ASSERT(blocked_queue_tl != END_TSO_QUEUE);
+      evac((StgClosure **)&blocked_queue_hd);
+      evac((StgClosure **)&blocked_queue_tl);
   }
-
+  
   if (sleeping_queue != END_TSO_QUEUE) {
-    sleeping_queue  = (StgTSO *)MarkRoot((StgClosure *)sleeping_queue);
+      evac((StgClosure **)&sleeping_queue);
   }
 #endif 
 
   for (m = main_threads; m != NULL; m = m->link) {
-    m->tso = (StgTSO *)MarkRoot((StgClosure *)m->tso);
+      evac((StgClosure **)&m->tso);
+  }
+  if (suspended_ccalling_threads != END_TSO_QUEUE) {
+      evac((StgClosure **)&suspended_ccalling_threads);
   }
-  if (suspended_ccalling_threads != END_TSO_QUEUE)
-    suspended_ccalling_threads = 
-      (StgTSO *)MarkRoot((StgClosure *)suspended_ccalling_threads);
 
-#if defined(SMP) || defined(PAR) || defined(GRAN)
-  markSparkQueue();
+#if defined(PAR) || defined(GRAN)
+  markSparkQueue(evac);
 #endif
 }
 
@@ -2024,7 +2306,7 @@ static void GetRoots(void)
    This needs to be protected by the GC condition variable above.  KH.
    -------------------------------------------------------------------------- */
 
-void (*extra_roots)(void);
+void (*extra_roots)(evac_fn);
 
 void
 performGC(void)
@@ -2039,17 +2321,16 @@ performMajorGC(void)
 }
 
 static void
-AllRoots(void)
+AllRoots(evac_fn evac)
 {
-  GetRoots();                  /* the scheduler's roots */
-  extra_roots();               /* the user's roots */
+    GetRoots(evac);            // the scheduler's roots
+    extra_roots(evac);         // the user's roots
 }
 
 void
-performGCWithRoots(void (*get_roots)(void))
+performGCWithRoots(void (*get_roots)(evac_fn))
 {
   extra_roots = get_roots;
-
   GarbageCollect(AllRoots,rtsFalse);
 }
 
@@ -2079,13 +2360,8 @@ threadStackOverflow(StgTSO *tso)
             printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size, 
                                              tso->sp+64)));
 
-#ifdef INTERPRETER
-    fprintf(stderr, "fatal: stack overflow in Hugs; aborting\n" );
-    exit(1);
-#else
     /* Send this thread the StackOverflow exception */
     raiseAsync(tso, (StgClosure *)stackOverflow_closure);
-#endif
     return tso;
   }
 
@@ -2099,10 +2375,10 @@ threadStackOverflow(StgTSO *tso)
   new_tso_size = round_to_mblocks(new_tso_size);  /* Be MBLOCK-friendly */
   new_stack_size = new_tso_size - TSO_STRUCT_SIZEW;
 
-  IF_DEBUG(scheduler, fprintf(stderr,"scheduler: increasing stack size from %d words to %d.\n", tso->stack_size, new_stack_size));
+  IF_DEBUG(scheduler, fprintf(stderr,"== scheduler: increasing stack size from %d words to %d.\n", tso->stack_size, new_stack_size));
 
   dest = (StgTSO *)allocate(new_tso_size);
-  TICK_ALLOC_TSO(new_tso_size-sizeofW(StgTSO),0);
+  TICK_ALLOC_TSO(new_stack_size,0);
 
   /* copy the TSO block and the old stack into the new area */
   memcpy(dest,tso,TSO_STRUCT_SIZE);
@@ -2117,7 +2393,7 @@ threadStackOverflow(StgTSO *tso)
   dest->stack_size = new_stack_size;
        
   /* and relocate the update frame list */
-  relocate_TSO(tso, dest);
+  relocate_stack(dest, diff);
 
   /* Mark the old TSO as relocated.  We have to check for relocated
    * TSOs in the garbage collector and any primops that deal with TSOs.
@@ -2168,8 +2444,10 @@ unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
      update blocked and fetch time (depending on type of the orig closure) */
   if (RtsFlags.ParFlags.ParStats.Full) {
     DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC, 
-                    GR_RESUME, ((StgTSO *)bqe), ((StgTSO *)bqe)->block_info.closure,
+                    GR_RESUMEQ, ((StgTSO *)bqe), ((StgTSO *)bqe)->block_info.closure,
                     0, 0 /* spark_queue_len(ADVISORY_POOL) */);
+    if (EMPTY_RUN_QUEUE())
+      emitSchedule = rtsTrue;
 
     switch (get_itbl(node)->type) {
        case FETCH_ME_BQ:
@@ -2180,6 +2458,10 @@ unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node )
        case BLACKHOLE_BQ:
          ((StgTSO *)bqe)->par.blocktime += CURRENT_TIME-((StgTSO *)bqe)->par.blockedat;
          break;
+#ifdef DIST
+        case MVAR:
+          break;
+#endif   
        default:
          barf("{unblockOneLocked}Daq Qagh: unexpected closure in blocking queue");
        }
@@ -2249,8 +2531,8 @@ unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
     case BLOCKED_FETCH:
       /* if it's a BLOCKED_FETCH put it on the PendingFetches list */
       next = bqe->link;
-      bqe->link = PendingFetches;
-      PendingFetches = bqe;
+      bqe->link = (StgBlockingQueueElement *)PendingFetches;
+      PendingFetches = (StgBlockedFetch *)bqe;
       break;
 
 # if defined(DEBUG)
@@ -2269,7 +2551,7 @@ unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node)
           (StgClosure *)bqe);
 # endif
     }
-  // IF_DEBUG(scheduler,sched_belch("waking up thread %ld", tso->id));
+  IF_PAR_DEBUG(bq, fprintf(stderr, ", %p (%s)", bqe, info_type((StgClosure*)bqe)));
   return next;
 }
 
@@ -2319,13 +2601,14 @@ awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
   nat len = 0; 
 
   IF_GRAN_DEBUG(bq, 
-               belch("## AwBQ for node %p on PE %d @ %ld by TSO %d (%p): ", \
+               belch("##-_ AwBQ for node %p on PE %d @ %ld by TSO %d (%p): ", \
                      node, CurrentProc, CurrentTime[CurrentProc], 
                      CurrentTSO->id, CurrentTSO));
 
   node_loc = where_is(node);
 
-  ASSERT(get_itbl(q)->type == TSO ||   // q is either a TSO or an RBHSave
+  ASSERT(q == END_BQ_QUEUE ||
+        get_itbl(q)->type == TSO ||   // q is either a TSO or an RBHSave
         get_itbl(q)->type == CONSTR); // closure (type constructor)
   ASSERT(is_unique(node));
 
@@ -2395,15 +2678,23 @@ awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
 void 
 awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
 {
-  StgBlockingQueueElement *bqe, *next;
+  StgBlockingQueueElement *bqe;
 
   ACQUIRE_LOCK(&sched_mutex);
 
   IF_PAR_DEBUG(verbose, 
-              belch("## AwBQ for node %p on [%x]: ",
+              belch("##-_ AwBQ for node %p on [%x]: ",
                     node, mytid));
-
-  ASSERT(get_itbl(q)->type == TSO ||           
+#ifdef DIST  
+  //RFP
+  if(get_itbl(q)->type == CONSTR || q==END_BQ_QUEUE) {
+    IF_PAR_DEBUG(verbose, belch("## ... nothing to unblock so lets just return. RFP (BUG?)"));
+    return;
+  }
+#endif
+  
+  ASSERT(q == END_BQ_QUEUE ||
+        get_itbl(q)->type == TSO ||           
         get_itbl(q)->type == BLOCKED_FETCH || 
         get_itbl(q)->type == CONSTR); 
 
@@ -2510,6 +2801,12 @@ unblockThread(StgTSO *tso)
       StgTSO *target  = tso->block_info.tso;
 
       ASSERT(get_itbl(target)->type == TSO);
+
+      if (target->what_next == ThreadRelocated) {
+         target = target->link;
+         ASSERT(get_itbl(target)->type == TSO);
+      }
+
       ASSERT(target->blocked_exceptions != NULL);
 
       last = (StgBlockingQueueElement **)&target->blocked_exceptions;
@@ -2528,6 +2825,7 @@ unblockThread(StgTSO *tso)
   case BlockedOnRead:
   case BlockedOnWrite:
     {
+      /* take TSO off blocked_queue */
       StgBlockingQueueElement *prev = NULL;
       for (t = (StgBlockingQueueElement *)blocked_queue_hd; t != END_BQ_QUEUE; 
           prev = t, t = t->link) {
@@ -2551,6 +2849,7 @@ unblockThread(StgTSO *tso)
 
   case BlockedOnDelay:
     {
+      /* take TSO off sleeping_queue */
       StgBlockingQueueElement *prev = NULL;
       for (t = (StgBlockingQueueElement *)sleeping_queue; t != END_BQ_QUEUE; 
           prev = t, t = t->link) {
@@ -2630,6 +2929,12 @@ unblockThread(StgTSO *tso)
       StgTSO *target  = tso->block_info.tso;
 
       ASSERT(get_itbl(target)->type == TSO);
+
+      while (target->what_next == ThreadRelocated) {
+         target = target->link;
+         ASSERT(get_itbl(target)->type == TSO);
+      }
+      
       ASSERT(target->blocked_exceptions != NULL);
 
       last = &target->blocked_exceptions;
@@ -2762,56 +3067,44 @@ raiseAsync(StgTSO *tso, StgClosure *exception)
   }
 
   while (1) {
-    int words = ((P_)su - (P_)sp) - 1;
+    nat words = ((P_)su - (P_)sp) - 1;
     nat i;
     StgAP_UPD * ap;
 
     /* If we find a CATCH_FRAME, and we've got an exception to raise,
-     * then build PAP(handler,exception,realworld#), and leave it on
-     * top of the stack ready to enter.
+     * then build the THUNK raise(exception), and leave it on
+     * top of the CATCH_FRAME ready to enter.
      */
     if (get_itbl(su)->type == CATCH_FRAME && exception != NULL) {
       StgCatchFrame *cf = (StgCatchFrame *)su;
+      StgClosure *raise;
+
       /* we've got an exception to raise, so let's pass it to the
        * handler in this frame.
        */
-      ap = (StgAP_UPD *)allocate(sizeofW(StgPAP) + 2);
-      TICK_ALLOC_UPD_PAP(3,0);
-      SET_HDR(ap,&stg_PAP_info,cf->header.prof.ccs);
-             
-      ap->n_args = 2;
-      ap->fun = cf->handler;   /* :: Exception -> IO a */
-      ap->payload[0] = exception;
-      ap->payload[1] = ARG_TAG(0); /* realworld token */
-
-      /* throw away the stack from Sp up to and including the
-       * CATCH_FRAME.
-       */
-      sp = (P_)su + sizeofW(StgCatchFrame) - 1; 
-      tso->su = cf->link;
-
-      /* Restore the blocked/unblocked state for asynchronous exceptions
-       * at the CATCH_FRAME.  
-       *
-       * If exceptions were unblocked at the catch, arrange that they
-       * are unblocked again after executing the handler by pushing an
-       * unblockAsyncExceptions_ret stack frame.
+      raise = (StgClosure *)allocate(sizeofW(StgClosure)+1);
+      TICK_ALLOC_SE_THK(1,0);
+      SET_HDR(raise,&stg_raise_info,cf->header.prof.ccs);
+      raise->payload[0] = exception;
+
+      /* throw away the stack from Sp up to the CATCH_FRAME.
        */
-      if (!cf->exceptions_blocked) {
-       *(sp--) = (W_)&stg_unblockAsyncExceptionszh_ret_info;
-      }
-      
-      /* Ensure that async exceptions are blocked when running the handler.
+      sp = (P_)su - 1;
+
+      /* Ensure that async excpetions are blocked now, so we don't get
+       * a surprise exception before we get around to executing the
+       * handler.
        */
       if (tso->blocked_exceptions == NULL) {
-       tso->blocked_exceptions = END_TSO_QUEUE;
+         tso->blocked_exceptions = END_TSO_QUEUE;
       }
-      
-      /* Put the newly-built PAP on top of the stack, ready to execute
+
+      /* Put the newly-built THUNK on top of the stack, ready to execute
        * when the thread restarts.
        */
-      sp[0] = (W_)ap;
+      sp[0] = (W_)raise;
       tso->sp = sp;
+      tso->su = su;
       tso->what_next = ThreadEnterGHC;
       IF_DEBUG(sanity, checkTSO(tso));
       return;
@@ -2849,14 +3142,24 @@ raiseAsync(StgTSO *tso, StgClosure *exception)
        /* Replace the updatee with an indirection - happily
         * this will also wake up any threads currently
         * waiting on the result.
+        *
+        * Warning: if we're in a loop, more than one update frame on
+        * the stack may point to the same object.  Be careful not to
+        * overwrite an IND_OLDGEN in this case, because we'll screw
+        * up the mutable lists.  To be on the safe side, don't
+        * overwrite any kind of indirection at all.  See also
+        * threadSqueezeStack in GC.c, where we have to make a similar
+        * check.
         */
-       UPD_IND_NOLOCK(su->updatee,ap);  /* revert the black hole */
+       if (!closure_IND(su->updatee)) {
+           UPD_IND_NOLOCK(su->updatee,ap);  /* revert the black hole */
+       }
        su = su->link;
        sp += sizeofW(StgUpdateFrame) -1;
        sp[0] = (W_)ap; /* push onto stack */
        break;
       }
-      
+
     case CATCH_FRAME:
       {
        StgCatchFrame *cf = (StgCatchFrame *)su;
@@ -2986,6 +3289,11 @@ detectBlackHoles( void )
 
     for (t = all_threads; t != END_TSO_QUEUE; t = t->global_link) {
 
+       while (t->what_next == ThreadRelocated) {
+           t = t->link;
+           ASSERT(get_itbl(t)->type == TSO);
+       }
+      
        if (t->why_blocked != BlockedOnBlackHole) {
            continue;
        }
@@ -3019,7 +3327,7 @@ detectBlackHoles( void )
            break;
        }
 
-    done:
+    done: ;
     }   
 }
 
@@ -3068,6 +3376,11 @@ printThreadBlockage(StgTSO *tso)
            tso->block_info.closure, info_type(tso->block_info.closure));
     break;
 #endif
+#if defined(RTS_SUPPORTS_THREADS)
+  case BlockedOnCCall:
+    fprintf(stderr,"is blocked on an external call");
+    break;
+#endif
   default:
     barf("printThreadBlockage: strange tso->why_blocked: %d for TSO %d (%d)",
         tso->why_blocked, tso->id, tso);
@@ -3094,7 +3407,22 @@ printAllThreads(void)
 {
   StgTSO *t;
 
+# if defined(GRAN)
+  char time_string[TIME_STR_LEN], node_str[NODE_STR_LEN];
+  ullong_format_string(TIME_ON_PROC(CurrentProc), 
+                      time_string, rtsFalse/*no commas!*/);
+
+  sched_belch("all threads at [%s]:", time_string);
+# elif defined(PAR)
+  char time_string[TIME_STR_LEN], node_str[NODE_STR_LEN];
+  ullong_format_string(CURRENT_TIME,
+                      time_string, rtsFalse/*no commas!*/);
+
+  sched_belch("all threads at [%s]:", time_string);
+# else
   sched_belch("all threads:");
+# endif
+
   for (t = all_threads; t != END_TSO_QUEUE; t = t->global_link) {
     fprintf(stderr, "\tthread %d ", t->id);
     printThreadStatus(t);
@@ -3120,27 +3448,41 @@ print_bq (StgClosure *node)
   /* should cover all closures that may have a blocking queue */
   ASSERT(get_itbl(node)->type == BLACKHOLE_BQ ||
         get_itbl(node)->type == FETCH_ME_BQ ||
-        get_itbl(node)->type == RBH);
+        get_itbl(node)->type == RBH ||
+        get_itbl(node)->type == MVAR);
     
   ASSERT(node!=(StgClosure*)NULL);         // sanity check
+
+  print_bqe(((StgBlockingQueue*)node)->blocking_queue);
+}
+
+/* 
+   Print a whole blocking queue starting with the element bqe.
+*/
+void 
+print_bqe (StgBlockingQueueElement *bqe)
+{
+  rtsBool end;
+
   /* 
      NB: In a parallel setup a BQ of an RBH must end with an RBH_Save closure;
   */
-  for (bqe = ((StgBlockingQueue*)node)->blocking_queue, end = (bqe==END_BQ_QUEUE);
+  for (end = (bqe==END_BQ_QUEUE);
        !end; // iterate until bqe points to a CONSTR
-       end = (get_itbl(bqe)->type == CONSTR) || (bqe->link==END_BQ_QUEUE), bqe = end ? END_BQ_QUEUE : bqe->link) {
-    ASSERT(bqe != END_BQ_QUEUE);             // sanity check
-    ASSERT(bqe != (StgTSO*)NULL);            // sanity check
+       end = (get_itbl(bqe)->type == CONSTR) || (bqe->link==END_BQ_QUEUE), 
+       bqe = end ? END_BQ_QUEUE : bqe->link) {
+    ASSERT(bqe != END_BQ_QUEUE);                               // sanity check
+    ASSERT(bqe != (StgBlockingQueueElement *)NULL);            // sanity check
     /* types of closures that may appear in a blocking queue */
     ASSERT(get_itbl(bqe)->type == TSO ||           
           get_itbl(bqe)->type == BLOCKED_FETCH || 
           get_itbl(bqe)->type == CONSTR); 
     /* only BQs of an RBH end with an RBH_Save closure */
-    ASSERT(get_itbl(bqe)->type != CONSTR || get_itbl(node)->type == RBH);
+    //ASSERT(get_itbl(bqe)->type != CONSTR || get_itbl(node)->type == RBH);
 
     switch (get_itbl(bqe)->type) {
     case TSO:
-      fprintf(stderr," TSO %d (%x),",
+      fprintf(stderr," TSO %u (%x),",
              ((StgTSO *)bqe)->id, ((StgTSO *)bqe));
       break;
     case BLOCKED_FETCH:
@@ -3158,8 +3500,8 @@ print_bq (StgClosure *node)
               "RBH_Save_?"), get_itbl(bqe));
       break;
     default:
-      barf("Unexpected closure type %s in blocking queue of %p (%s)",
-          info_type(bqe), node, info_type(node));
+      barf("Unexpected closure type %s in blocking queue", // of %p (%s)",
+          info_type((StgClosure *)bqe)); // , node, info_type(node));
       break;
     }
   } /* for */
@@ -3262,7 +3604,9 @@ sched_belch(char *s, ...)
   va_list ap;
   va_start(ap,s);
 #ifdef SMP
-  fprintf(stderr, "scheduler (task %ld): ", pthread_self());
+  fprintf(stderr, "scheduler (task %ld): ", osThreadId());
+#elif defined(PAR)
+  fprintf(stderr, "== ");
 #else
   fprintf(stderr, "scheduler: ");
 #endif
@@ -3277,18 +3621,15 @@ sched_belch(char *s, ...)
 //@subsection Index
 
 //@index
-//* MainRegTable::  @cindex\s-+MainRegTable
 //* StgMainThread::  @cindex\s-+StgMainThread
 //* awaken_blocked_queue::  @cindex\s-+awaken_blocked_queue
 //* blocked_queue_hd::  @cindex\s-+blocked_queue_hd
 //* blocked_queue_tl::  @cindex\s-+blocked_queue_tl
 //* context_switch::  @cindex\s-+context_switch
 //* createThread::  @cindex\s-+createThread
-//* free_capabilities::  @cindex\s-+free_capabilities
 //* gc_pending_cond::  @cindex\s-+gc_pending_cond
 //* initScheduler::  @cindex\s-+initScheduler
 //* interrupted::  @cindex\s-+interrupted
-//* n_free_capabilities::  @cindex\s-+n_free_capabilities
 //* next_thread_id::  @cindex\s-+next_thread_id
 //* print_bq::  @cindex\s-+print_bq
 //* run_queue_hd::  @cindex\s-+run_queue_hd
@@ -3296,7 +3637,5 @@ sched_belch(char *s, ...)
 //* sched_mutex::  @cindex\s-+sched_mutex
 //* schedule::  @cindex\s-+schedule
 //* take_off_run_queue::  @cindex\s-+take_off_run_queue
-//* task_ids::  @cindex\s-+task_ids
 //* term_mutex::  @cindex\s-+term_mutex
-//* thread_ready_cond::  @cindex\s-+thread_ready_cond
 //@end index