[project @ 2002-01-31 11:18:06 by sof]
[ghc-hetmet.git] / ghc / rts / Schedule.c
index fa26c03..026ffeb 100644 (file)
@@ -1,5 +1,5 @@
 /* ---------------------------------------------------------------------------
- * $Id: Schedule.c,v 1.99 2001/07/30 09:47:13 simonmar Exp $
+ * $Id: Schedule.c,v 1.114 2002/01/31 11:18:07 sof Exp $
  *
  * (c) The GHC Team, 1998-2000
  *
  *
  * 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)
+ * 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, , ,
@@ -74,6 +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 "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 <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.
  *
@@ -129,8 +137,8 @@ typedef struct StgMainThread_ {
   StgTSO *         tso;
   SchedulerStatus  stat;
   StgClosure **    ret;
-#ifdef SMP
-  pthread_cond_t wakeup;
+#if defined(RTS_SUPPORTS_THREADS)
+  CondVar          wakeup;
 #endif
   struct StgMainThread_ *link;
 } StgMainThread;
@@ -180,7 +188,6 @@ StgTSO *all_threads;
  */
 static StgTSO *suspended_ccalling_threads;
 
-static void GetRoots(evac_fn);
 static StgTSO *threadStackOverflow(StgTSO *tso);
 
 /* KH: The following two flags are shared memory locations.  There is no need
@@ -220,17 +227,9 @@ 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 */
+
+#if !defined(SMP)
+Capability MainCapability;     /* for non-SMP, we have one global capability */
 #endif
 
 #if defined(GRAN)
@@ -247,7 +246,7 @@ rtsBool ready_to_gc;
 
 /* All our current task ids, saved in case we need to kill them later.
  */
-#ifdef SMP
+#if defined(SMP)
 //@cindex task_ids
 task_info *task_ids;
 #endif
@@ -268,15 +267,14 @@ 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.
+ */
+MutexVar  sched_mutex       = INIT_MUTEX_VAR;
+MutexVar  term_mutex        = INIT_MUTEX_VAR;
+CondVar   thread_ready_cond = INIT_COND_VAR;
+CondVar   gc_pending_cond   = INIT_COND_VAR;
 
 nat await_death;
 #endif
@@ -305,7 +303,7 @@ char *threadReturnCode_strs[] = {
 };
 #endif
 
-#ifdef PAR
+#if defined(PAR)
 StgTSO * createSparkThread(rtsSpark spark);
 StgTSO * activateSpark (rtsSpark spark);  
 #endif
@@ -429,7 +427,7 @@ 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;
@@ -441,16 +439,17 @@ schedule( void )
          }
          *prev = m->link;
          m->stat = Success;
-         pthread_cond_broadcast(&m->wakeup);
+         broadcastCondVar(&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);
+         broadcastCondVar(&m->wakeup);
          break;
        default:
          break;
@@ -458,7 +457,8 @@ schedule( void )
       }
     }
 
-#else
+#else /* not threaded */
+
 # if defined(PAR)
     /* in GUM do this only on the Main PE */
     if (IAmMainThread)
@@ -476,6 +476,7 @@ schedule( void )
          m->stat = Success;
          return;
        } else {
+         if (m->ret) { *(m->ret) = NULL; };
          if (was_interrupted) {
            m->stat = Interrupted;
          } else {
@@ -490,10 +491,13 @@ 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 */
@@ -520,11 +524,18 @@ schedule( void )
       /* 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) {
+         signalCondVar ( &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
@@ -535,21 +546,14 @@ schedule( void )
     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 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
@@ -565,8 +569,8 @@ schedule( void )
     if (blocked_queue_hd == END_TSO_QUEUE
        && run_queue_hd == END_TSO_QUEUE
        && sleeping_queue == END_TSO_QUEUE
-#ifdef SMP
-       && (n_free_capabilities == RtsFlags.ParFlags.nNodes)
+#if defined(SMP)
+       && allFreeCapabilities()
 #endif
        )
     {
@@ -575,45 +579,76 @@ schedule( void )
        if (blocked_queue_hd == END_TSO_QUEUE
            && run_queue_hd == END_TSO_QUEUE
            && sleeping_queue == END_TSO_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 (run_queue_hd == END_TSO_QUEUE) {
-               StgMainThread *m = main_threads;
-#ifdef SMP
-               for (; m != NULL; m = m->link) {
-                   m->ret = NULL;
-                   m->stat = Deadlock;
-                   pthread_cond_broadcast(&m->wakeup);
+               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");
+                   }
                }
-               main_threads = NULL;
 #else
-               m->ret = NULL;
-               m->stat = Deadlock;
-               main_threads = m->link;
-               return;
+               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
            }
+#if !defined(RTS_SUPPORTS_THREADS)
+           ASSERT( run_queue_hd != END_TSO_QUEUE );
+#endif
        }
     }
 #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);
+      waitCondVar ( &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) {
+    while ( run_queue_hd == END_TSO_QUEUE 
+#if defined(SMP)
+           || noFreeCapabilities() 
+#endif
+           ) {
       IF_DEBUG(scheduler, sched_belch("waiting for work"));
-      pthread_cond_wait(&thread_ready_cond, &sched_mutex);
+      waitCondVar ( &thread_ready_cond, &sched_mutex );
       IF_DEBUG(scheduler, sched_belch("work now available"));
     }
 #endif
@@ -897,30 +932,33 @@ schedule( void )
      */
     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--;
+    grabCapability(&cap);
 #else
-    cap = &MainRegTable;
+    cap = &MainCapability;
 #endif
 
-    cap->rCurrentTSO = t;
+    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;
@@ -930,20 +968,24 @@ 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;
     case ThreadEnterGHC:
-       ret = StgRun((StgFunPtr) stg_enterStackTop, cap);
+       ret = StgRun((StgFunPtr) stg_enterStackTop, &cap->r);
        break;
     case ThreadRunGHC:
-       ret = StgRun((StgFunPtr) stg_returnToStackTop, cap);
+       ret = StgRun((StgFunPtr) stg_returnToStackTop, &cap->r);
        break;
     case ThreadEnterInterp:
        ret = interpretBCO(cap);
@@ -955,17 +997,18 @@ 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 
@@ -978,14 +1021,65 @@ schedule( void )
     switch (ret) {
     case HeapOverflow:
 #if defined(GRAN)
-      IF_DEBUG(gran, 
-              DumpGranEvent(GR_DESCHEDULE, t));
+      IF_DEBUG(gran, DumpGranEvent(GR_DESCHEDULE, t));
       globalGranStats.tot_heapover++;
 #elif defined(PAR)
-      // IF_DEBUG(par, 
-      //DumpGranEvent(GR_DESCHEDULE, t);
       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.
@@ -1200,13 +1294,20 @@ schedule( void )
     }
     
 #ifdef SMP
-    cap->link = free_capabilities;
-    free_capabilities = cap;
-    n_free_capabilities++;
+    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
 
 #ifdef SMP
-    if (ready_to_gc && n_free_capabilities == RtsFlags.ParFlags.nNodes) 
+    if (ready_to_gc && allFreeCapabilities() )
 #else
     if (ready_to_gc) 
 #endif
@@ -1216,13 +1317,13 @@ schedule( void )
        * 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);
+      broadcastCondVar(&gc_pending_cond);
 #endif
 #if defined(GRAN)
       /* add a ContinueThread event to continue execution of current thread */
@@ -1235,24 +1336,20 @@ schedule( void )
               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"));
 }
@@ -1303,33 +1400,36 @@ 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->rCurrentTSO);
-  cap->rCurrentTSO->link = suspended_ccalling_threads;
-  suspended_ccalling_threads = cap->rCurrentTSO;
+  threadPaused(cap->r.rCurrentTSO);
+  cap->r.rCurrentTSO->link = suspended_ccalling_threads;
+  suspended_ccalling_threads = cap->r.rCurrentTSO;
 
   /* 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);
 #endif
 
   RELEASE_LOCK(&sched_mutex);
   return tok; 
 }
 
-Capability *
+StgRegTable *
 resumeThread( StgInt tok )
 {
   StgTSO *tso, **prev;
@@ -1352,22 +1452,20 @@ resumeThread( StgInt tok )
   tso->link = END_TSO_QUEUE;
 
 #ifdef SMP
-  while (free_capabilities == NULL) {
+  while ( noFreeCapabilities() ) {
     IF_DEBUG(scheduler, sched_belch("waiting to resume"));
-    pthread_cond_wait(&thread_ready_cond, &sched_mutex);
+    waitCondVar(&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--;
+  grabCapability(&cap);
 #else  
-  cap = &MainRegTable;
+  cap = &MainCapability;
 #endif
 
-  cap->rCurrentTSO = tso;
+  cap->r.rCurrentTSO = tso;
 
   RELEASE_LOCK(&sched_mutex);
-  return cap;
+  return &cap->r;
 }
 
 
@@ -1394,6 +1492,16 @@ 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
@@ -1459,7 +1567,7 @@ 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)
@@ -1664,11 +1772,6 @@ activateSpark (rtsSpark spark)
 void
 scheduleThread(StgTSO *tso)
 {
-  if (tso==END_TSO_QUEUE){    
-    schedule();
-    return;
-  }
-
   ACQUIRE_LOCK(&sched_mutex);
 
   /* Put the new thread on the head of the runnable queue.  The caller
@@ -1697,7 +1800,7 @@ scheduleThread(StgTSO *tso)
 void
 taskStart(void) /*  ( void *arg STG_UNUSED)  */
 {
-  scheduleThread(END_TSO_QUEUE);
+  schedule();
 }
 #endif
 
@@ -1719,11 +1822,10 @@ 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)
 {
@@ -1773,24 +1875,12 @@ initScheduler(void)
 
 #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););
+  initCapabilities(RtsFlags.ParFlags.nNodes);
+#else
+  initCapability(&MainCapability);
 #endif
 
-#if defined(SMP) || defined(PAR)
+#if /* defined(SMP) ||*/ defined(PAR)
   initSparkPools();
 #endif
 }
@@ -1801,7 +1891,7 @@ startTasks( void )
 {
   nat i;
   int r;
-  pthread_t tid;
+  OSThreadId tid;
   
   /* make some space for saving all the thread ids */
   task_ids = stgMallocBytes(RtsFlags.ParFlags.nNodes * sizeof(task_info),
@@ -1809,7 +1899,7 @@ startTasks( void )
   
   /* and create all the threads */
   for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) {
-    r = pthread_create(&tid,NULL,taskStart,NULL);
+    r = createOSThread(&tid,taskStart);
     if (r != 0) {
       barf("startTasks: Can't create new Posix thread");
     }
@@ -1818,7 +1908,7 @@ startTasks( void )
     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();
+    task_ids[i].elapsedtimestart = stat_getElapsedTime();
     IF_DEBUG(scheduler,fprintf(stderr,"scheduler: Started task: %ld\n",tid););
   }
 }
@@ -1931,8 +2021,8 @@ 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)
+  initCondVar(&m->wakeup);
 #endif
 
   m->link = main_threads;
@@ -1943,7 +2033,7 @@ waitThread(StgTSO *tso, /*out*/StgClosure **ret)
 
 #ifdef SMP
   do {
-    pthread_cond_wait(&m->wakeup, &sched_mutex);
+    waitCondVar(&m->wakeup, &sched_mutex);
   } while (m->stat == NoStatus);
 #elif defined(GRAN)
   /* GranSim specific init */
@@ -1959,8 +2049,8 @@ waitThread(StgTSO *tso, /*out*/StgClosure **ret)
 
   stat = m->stat;
 
-#ifdef SMP
-  pthread_cond_destroy(&m->wakeup);
+#if defined(RTS_SUPPORTS_THREADS)
+  closeCondVar(&m->wakeup);
 #endif
 
   IF_DEBUG(scheduler, fprintf(stderr, "== scheduler: main thread (%d) finished\n", 
@@ -2092,7 +2182,7 @@ take_off_run_queue(StgTSO *tso) {
        KH @ 25/10/99
 */
 
-static void
+void
 GetRoots(evac_fn evac)
 {
   StgMainThread *m;
@@ -2142,7 +2232,7 @@ GetRoots(evac_fn evac)
       evac((StgClosure **)&suspended_ccalling_threads);
   }
 
-#if defined(SMP) || defined(PAR) || defined(GRAN)
+#if defined(PAR) || defined(GRAN)
   markSparkQueue(evac);
 #endif
 }
@@ -2232,7 +2322,7 @@ threadStackOverflow(StgTSO *tso)
   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);
@@ -3465,7 +3555,7 @@ 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
@@ -3489,11 +3579,9 @@ sched_belch(char *s, ...)
 //* 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