[project @ 2004-03-13 00:56:45 by sof]
[ghc-hetmet.git] / ghc / rts / Schedule.c
index bc091be..262e4c4 100644 (file)
@@ -1,5 +1,5 @@
 /* ---------------------------------------------------------------------------
- * $Id: Schedule.c,v 1.183 2003/12/16 13:27:32 simonmar Exp $
+ * $Id: Schedule.c,v 1.194 2004/03/13 00:56:45 sof Exp $
  *
  * (c) The GHC Team, 1998-2003
  *
@@ -224,12 +224,6 @@ static void     detectBlackHoles  ( void );
 Mutex     sched_mutex       = INIT_MUTEX_VAR;
 Mutex     term_mutex        = INIT_MUTEX_VAR;
 
-/*
- * A heavyweight solution to the problem of protecting
- * the thread_id from concurrent update.
- */
-Mutex     thread_id_mutex   = INIT_MUTEX_VAR;
-
 #endif /* RTS_SUPPORTS_THREADS */
 
 #if defined(PAR)
@@ -265,6 +259,7 @@ static void
 taskStart(void)
 {
   ACQUIRE_LOCK(&sched_mutex);
+  startingWorkerThread = rtsFalse;
   schedule(NULL,NULL);
   RELEASE_LOCK(&sched_mutex);
 }
@@ -327,7 +322,7 @@ schedule( StgMainThread *mainThread USED_WHEN_RTS_SUPPORTS_THREADS,
           Capability *initialCapability )
 {
   StgTSO *t;
-  Capability *cap = initialCapability;
+  Capability *cap;
   StgThreadReturnCode ret;
 #if defined(GRAN)
   rtsEvent *event;
@@ -345,7 +340,9 @@ schedule( StgMainThread *mainThread USED_WHEN_RTS_SUPPORTS_THREADS,
   StgTSOWhatNext prev_what_next;
   
   // Pre-condition: sched_mutex is held.
+  // We might have a capability, passed in as initialCapability.
+  cap = initialCapability;
+
 #if defined(RTS_SUPPORTS_THREADS)
   //
   // in the threaded case, the capability is either passed in via the
@@ -435,103 +432,6 @@ schedule( StgMainThread *mainThread USED_WHEN_RTS_SUPPORTS_THREADS,
 #endif
     }
 
-    //
-    // Go through the list of main threads and wake up any
-    // clients whose computations have finished.  ToDo: this
-    // should be done more efficiently without a linear scan
-    // of the main threads list, somehow...
-    //
-#if defined(RTS_SUPPORTS_THREADS)
-    { 
-       StgMainThread *m, **prev;
-       prev = &main_threads;
-       for (m = main_threads; m != NULL; prev = &m->link, m = m->link) {
-         if (m->tso->what_next == ThreadComplete
-             || m->tso->what_next == ThreadKilled)
-         {
-           if (m == mainThread)
-           {
-              if (m->tso->what_next == ThreadComplete)
-              {
-                if (m->ret)
-                {
-                  // NOTE: return val is tso->sp[1] (see StgStartup.hc)
-                  *(m->ret) = (StgClosure *)m->tso->sp[1]; 
-                }
-                m->stat = Success;
-              }
-              else
-              {
-                if (m->ret)
-                {
-                  *(m->ret) = NULL;
-                }
-                if (was_interrupted)
-                {
-                  m->stat = Interrupted;
-                }
-                else
-                {
-                  m->stat = Killed;
-                }
-              }
-              *prev = m->link;
-           
-#ifdef DEBUG
-             removeThreadLabel((StgWord)m->tso->id);
-#endif
-              releaseCapability(cap);
-              return;
-            }
-            else
-            {
-                // The current OS thread can not handle the fact that
-                // the Haskell thread "m" has ended.  "m" is bound;
-                // the scheduler loop in it's bound OS thread has to
-                // return, so let's pass our capability directly to
-                // that thread.
-               passCapability(&m->bound_thread_cond);
-            }
-          }
-       }
-    }
-    
-#else /* not threaded */
-
-# if defined(PAR)
-    /* in GUM do this only on the Main PE */
-    if (IAmMainThread)
-# endif
-    /* If our main thread has finished or been killed, return.
-     */
-    {
-      StgMainThread *m = main_threads;
-      if (m->tso->what_next == ThreadComplete
-         || m->tso->what_next == ThreadKilled) {
-#ifdef DEBUG
-       removeThreadLabel((StgWord)m->tso->id);
-#endif
-       main_threads = main_threads->link;
-       if (m->tso->what_next == ThreadComplete) {
-           // We finished successfully, fill in the return value
-           // NOTE: return val is tso->sp[1] (see StgStartup.hc)
-           if (m->ret) { *(m->ret) = (StgClosure *)m->tso->sp[1]; };
-           m->stat = Success;
-           return;
-       } else {
-         if (m->ret) { *(m->ret) = NULL; };
-         if (was_interrupted) {
-           m->stat = Interrupted;
-         } else {
-           m->stat = Killed;
-         }
-         return;
-       }
-      }
-    }
-#endif
-
-
 #if defined(RTS_USER_SIGNALS)
     // check for signals each time around the scheduler
     if (signals_pending()) {
@@ -541,19 +441,20 @@ schedule( StgMainThread *mainThread USED_WHEN_RTS_SUPPORTS_THREADS,
     }
 #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
-     * can wait indefinitely for something to happen.
-     */
-    if ( !EMPTY_QUEUE(blocked_queue_hd) || !EMPTY_QUEUE(sleeping_queue) 
+    //
+    // Check whether any waiting threads need to be woken up.  If the
+    // run queue is empty, and there are no other tasks running, we
+    // can wait indefinitely for something to happen.
+    //
+    if ( !EMPTY_QUEUE(blocked_queue_hd) || !EMPTY_QUEUE(sleeping_queue)
 #if defined(RTS_SUPPORTS_THREADS)
                || EMPTY_RUN_QUEUE()
 #endif
-        )
+       )
     {
       awaitEvent( EMPTY_RUN_QUEUE() );
     }
-    /* we can be interrupted while waiting for I/O... */
+    // we can be interrupted while waiting for I/O...
     if (interrupted) continue;
 
     /* 
@@ -619,11 +520,9 @@ schedule( StgMainThread *mainThread USED_WHEN_RTS_SUPPORTS_THREADS,
            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);
+               raiseAsync(m->tso, (StgClosure *)NonTermination_closure);
                break;
            default:
                barf("deadlock: main thread blocked in a strange way");
@@ -640,7 +539,7 @@ schedule( StgMainThread *mainThread USED_WHEN_RTS_SUPPORTS_THREADS,
 
 #if defined(RTS_SUPPORTS_THREADS)
     if ( EMPTY_RUN_QUEUE() ) {
-      continue; // nothing to do
+       continue; // nothing to do
     }
 #endif
 
@@ -929,12 +828,7 @@ schedule( StgMainThread *mainThread USED_WHEN_RTS_SUPPORTS_THREADS,
 
 #ifdef THREADED_RTS
     {
-      StgMainThread *m;
-      for(m = main_threads; m; m = m->link)
-      {
-       if(m->tso == t)
-         break;
-      }
+      StgMainThread *m = t->main;
       
       if(m)
       {
@@ -1158,7 +1052,6 @@ run_thread:
        */
       threadPaused(t);
       { 
-       StgMainThread *m;
        /* enlarge the stack */
        StgTSO *new_t = threadStackOverflow(t);
        
@@ -1166,10 +1059,8 @@ run_thread:
         * main thread stack.  It better not be on any other queues...
         * (it shouldn't be).
         */
-       for (m = main_threads; m != NULL; m = m->link) {
-         if (m->tso == t) {
-           m->tso = new_t;
-         }
+       if (t->main != NULL) {
+           t->main->tso = new_t;
        }
        threadPaused(new_t);
        PUSH_ON_RUN_QUEUE(new_t);
@@ -1332,8 +1223,74 @@ run_thread:
          !RtsFlags.ParFlags.ParStats.Suppressed) 
        DumpEndEvent(CURRENT_PROC, t, rtsFalse /* not mandatory */);
 #endif
+
+      //
+      // Check whether the thread that just completed was a main
+      // thread, and if so return with the result.  
+      //
+      // There is an assumption here that all thread completion goes
+      // through this point; we need to make sure that if a thread
+      // ends up in the ThreadKilled state, that it stays on the run
+      // queue so it can be dealt with here.
+      //
+      if (
+#if defined(RTS_SUPPORTS_THREADS)
+         mainThread != NULL
+#else
+         mainThread->tso == t
+#endif
+         )
+      {
+         // We are a bound thread: this must be our thread that just
+         // completed.
+         ASSERT(mainThread->tso == t);
+
+         if (t->what_next == ThreadComplete) {
+             if (mainThread->ret) {
+                 // NOTE: return val is tso->sp[1] (see StgStartup.hc)
+                 *(mainThread->ret) = (StgClosure *)mainThread->tso->sp[1]; 
+             }
+             mainThread->stat = Success;
+         } else {
+             if (mainThread->ret) {
+                 *(mainThread->ret) = NULL;
+             }
+             if (was_interrupted) {
+                 mainThread->stat = Interrupted;
+             } else {
+                 mainThread->stat = Killed;
+             }
+         }
+#ifdef DEBUG
+         removeThreadLabel((StgWord)mainThread->tso->id);
+#endif
+         if (mainThread->prev == NULL) {
+             main_threads = mainThread->link;
+         } else {
+             mainThread->prev->link = mainThread->link;
+         }
+         if (mainThread->link != NULL) {
+             mainThread->link->prev = NULL;
+         }
+         releaseCapability(cap);
+         return;
+      }
+
+#ifdef RTS_SUPPORTS_THREADS
+      ASSERT(t->main == NULL);
+#else
+      if (t->main != NULL) {
+         // Must be a main thread that is not the topmost one.  Leave
+         // it on the run queue until the stack has unwound to the
+         // point where we can deal with this.  Leaving it on the run
+         // queue also ensures that the garbage collector knows about
+         // this thread and its return value (it gets dropped from the
+         // all_threads list so there's no other way to find it).
+         APPEND_TO_RUN_QUEUE(t);
+      }
+#endif
       break;
-      
+
     default:
       barf("schedule: invalid thread return code %d", (int)ret);
     }
@@ -1414,12 +1371,7 @@ StgBool
 isThreadBound(StgTSO* tso USED_IN_THREADED_RTS)
 {
 #ifdef THREADED_RTS
-  StgMainThread *m;
-  for(m = main_threads; m; m = m->link)
-  {
-    if(m->tso == tso)
-      return rtsTrue;
-  }
+  return (tso->main != NULL);
 #endif
   return rtsFalse;
 }
@@ -1428,13 +1380,22 @@ isThreadBound(StgTSO* tso USED_IN_THREADED_RTS)
  * Singleton fork(). Do not copy any running threads.
  * ------------------------------------------------------------------------- */
 
+#ifndef mingw32_TARGET_OS
+#define FORKPROCESS_PRIMOP_SUPPORTED
+#endif
+
+#ifdef FORKPROCESS_PRIMOP_SUPPORTED
 static void 
 deleteThreadImmediately(StgTSO *tso);
-
+#endif
 StgInt
-forkProcess(HsStablePtr *entry)
+forkProcess(HsStablePtr *entry
+#ifndef FORKPROCESS_PRIMOP_SUPPORTED
+           STG_UNUSED
+#endif
+           )
 {
-#ifndef mingw32_TARGET_OS
+#ifdef FORKPROCESS_PRIMOP_SUPPORTED
   pid_t pid;
   StgTSO* t,*next;
   StgMainThread *m;
@@ -1468,17 +1429,17 @@ forkProcess(HsStablePtr *entry)
       // wipe the main thread list
     while((m = main_threads) != NULL) {
       main_threads = m->link;
-#ifdef THREADED_RTS
+# ifdef THREADED_RTS
       closeCondition(&m->bound_thread_cond);
-#endif
+# endif
       stgFree(m);
     }
     
-#ifdef RTS_SUPPORTS_THREADS
+# ifdef RTS_SUPPORTS_THREADS
     resetTaskManagerAfterFork();      // tell startTask() and friends that
     startingWorkerThread = rtsFalse;  // we have no worker threads any more
     resetWorkerWakeupPipeAfterFork();
-#endif
+# endif
     
     rc = rts_evalStableIO(entry, NULL);  // run the action
     rts_checkSchedStatus("forkProcess",rc);
@@ -1488,10 +1449,10 @@ forkProcess(HsStablePtr *entry)
     hs_exit();                      // clean up and exit
     stg_exit(0);
   }
-#else /* mingw32 */
-  barf("forkProcess#: primop not implemented for mingw32, sorry!\n");
+#else /* !FORKPROCESS_PRIMOP_SUPPORTED */
+  barf("forkProcess#: primop not supported, sorry!\n");
   return -1;
-#endif /* mingw32 */
+#endif
 }
 
 /* ---------------------------------------------------------------------------
@@ -1564,17 +1525,12 @@ suspendThread( StgRegTable *reg,
   cap->r.rCurrentTSO->link = suspended_ccalling_threads;
   suspended_ccalling_threads = cap->r.rCurrentTSO;
 
-#if defined(RTS_SUPPORTS_THREADS)
-  if(cap->r.rCurrentTSO->blocked_exceptions == NULL)
-  {
+  if(cap->r.rCurrentTSO->blocked_exceptions == NULL)  {
       cap->r.rCurrentTSO->why_blocked = BlockedOnCCall;
       cap->r.rCurrentTSO->blocked_exceptions = END_TSO_QUEUE;
-  }
-  else
-  {
+  } else {
       cap->r.rCurrentTSO->why_blocked = BlockedOnCCall_NoUnblockExc;
   }
-#endif
 
   /* Use the thread ID as the token; it should be unique */
   tok = cap->r.rCurrentTSO->id;
@@ -1630,13 +1586,10 @@ resumeThread( StgInt tok,
   }
   tso->link = END_TSO_QUEUE;
   
-#if defined(RTS_SUPPORTS_THREADS)
-  if(tso->why_blocked == BlockedOnCCall)
-  {
+  if(tso->why_blocked == BlockedOnCCall) {
       awakenBlockedQueueNoLock(tso->blocked_exceptions);
       tso->blocked_exceptions = NULL;
   }
-#endif
   
   /* Reset blocking status */
   tso->why_blocked  = NotBlocked;
@@ -1760,18 +1713,12 @@ createThread(nat size)
   // Always start with the compiled code evaluator
   tso->what_next = ThreadRunGHC;
 
-  /* tso->id needs to be unique.  For now we use a heavyweight mutex to
-   * protect the increment operation on next_thread_id.
-   * In future, we could use an atomic increment instead.
-   */
-  ACQUIRE_LOCK(&thread_id_mutex);
   tso->id = next_thread_id++; 
-  RELEASE_LOCK(&thread_id_mutex);
-
   tso->why_blocked  = NotBlocked;
   tso->blocked_exceptions = NULL;
 
   tso->saved_errno = 0;
+  tso->main = NULL;
   
   tso->stack_size   = stack_size;
   tso->max_stack_size = round_to_mblocks(RtsFlags.GcFlags.maxStkSize) 
@@ -1964,27 +1911,24 @@ void
 scheduleThread_(StgTSO *tso)
 {
   // Precondition: sched_mutex must be held.
-
-  /* Put the new thread on the head of the runnable queue.  The caller
-   * better push an appropriate closure on this thread's stack
-   * beforehand.  In the SMP case, the thread may start running as
-   * soon as we release the scheduler lock below.
-   */
   PUSH_ON_RUN_QUEUE(tso);
   THREAD_RUNNABLE();
-
-#if 0
-  IF_DEBUG(scheduler,printTSO(tso));
-#endif
 }
 
-void scheduleThread(StgTSO* tso)
+void
+scheduleThread(StgTSO* tso)
 {
   ACQUIRE_LOCK(&sched_mutex);
   scheduleThread_(tso);
   RELEASE_LOCK(&sched_mutex);
 }
 
+#if defined(RTS_SUPPORTS_THREADS)
+static Condition bound_cond_cache;
+static int bound_cond_cache_full = 0;
+#endif
+
+
 SchedulerStatus
 scheduleWaitThread(StgTSO* tso, /*[out]*/HaskellObj* ret,
                   Capability *initialCapability)
@@ -1994,10 +1938,26 @@ scheduleWaitThread(StgTSO* tso, /*[out]*/HaskellObj* ret,
 
     m = stgMallocBytes(sizeof(StgMainThread), "waitThread");
     m->tso = tso;
+    tso->main = m;
     m->ret = ret;
     m->stat = NoStatus;
+    m->link = main_threads;
+    m->prev = NULL;
+    if (main_threads != NULL) {
+       main_threads->prev = m;
+    }
+    main_threads = m;
+
 #if defined(RTS_SUPPORTS_THREADS)
-    initCondition(&m->bound_thread_cond);
+    // Allocating a new condition for each thread is expensive, so we
+    // cache one.  This is a pretty feeble hack, but it helps speed up
+    // consecutive call-ins quite a bit.
+    if (bound_cond_cache_full) {
+       m->bound_thread_cond = bound_cond_cache;
+       bound_cond_cache_full = 0;
+    } else {
+       initCondition(&m->bound_thread_cond);
+    }
 #endif
 
     /* Put the thread on the main-threads list prior to scheduling the TSO.
@@ -2011,10 +1971,10 @@ scheduleWaitThread(StgTSO* tso, /*[out]*/HaskellObj* ret,
     */
     IF_DEBUG(scheduler, sched_belch("waiting for thread (%d)", tso->id));
     
-    m->link = main_threads;
-    main_threads = m;
-    
-    scheduleThread_(tso);
+    PUSH_ON_RUN_QUEUE(tso);
+    // NB. Don't call THREAD_RUNNABLE() here, because the thread is
+    // bound and only runnable by *this* OS thread, so waking up other
+    // workers will just slow things down.
 
     return waitThread_(m, initialCapability);
 }
@@ -2066,14 +2026,9 @@ initScheduler(void)
    * the scheduler. */
   initMutex(&sched_mutex);
   initMutex(&term_mutex);
-  initMutex(&thread_id_mutex);
-
-  initCondition(&thread_ready_cond);
 #endif
   
-#if defined(RTS_SUPPORTS_THREADS)
   ACQUIRE_LOCK(&sched_mutex);
-#endif
 
   /* A capability holds the state a native thread needs in
    * order to execute STG code. At least one capability is
@@ -2091,7 +2046,6 @@ initScheduler(void)
 #endif
 
   RELEASE_LOCK(&sched_mutex);
-
 }
 
 void
@@ -2135,7 +2089,13 @@ waitThread_(StgMainThread* m, Capability *initialCapability)
   stat = m->stat;
 
 #if defined(RTS_SUPPORTS_THREADS)
-  closeCondition(&m->bound_thread_cond);
+  // Free the condition variable, returning it to the cache if possible.
+  if (!bound_cond_cache_full) {
+      bound_cond_cache = m->bound_thread_cond;
+      bound_cond_cache_full = 1;
+  } else {
+      closeCondition(&m->bound_thread_cond);
+  }
 #endif
 
   IF_DEBUG(scheduler, sched_belch("main thread (%d) finished", m->tso->id));
@@ -2214,25 +2174,6 @@ GetRoots( evac_fn evac )
   // mark the signal handlers (signals should be already blocked)
   markSignalHandlers(evac);
 #endif
-
-  // main threads which have completed need to be retained until they
-  // are dealt with in the main scheduler loop.  They won't be
-  // retained any other way: the GC will drop them from the
-  // all_threads list, so we have to be careful to treat them as roots
-  // here.
-  { 
-      StgMainThread *m;
-      for (m = main_threads; m != NULL; m = m->link) {
-         switch (m->tso->what_next) {
-         case ThreadComplete:
-         case ThreadKilled:
-             evac((StgClosure **)&m->tso);
-             break;
-         default:
-             break;
-         }
-      }
-  }
 }
 
 /* -----------------------------------------------------------------------------
@@ -2648,7 +2589,6 @@ awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node)
 
 #else   /* !GRAN && !PAR */
 
-#ifdef RTS_SUPPORTS_THREADS
 void
 awakenBlockedQueueNoLock(StgTSO *tso)
 {
@@ -2656,7 +2596,6 @@ awakenBlockedQueueNoLock(StgTSO *tso)
     tso = unblockOneLocked(tso);
   }
 }
-#endif
 
 void
 awakenBlockedQueue(StgTSO *tso)
@@ -3001,6 +2940,7 @@ deleteThread(StgTSO *tso)
   raiseAsync(tso,NULL);
 }
 
+#ifdef FORKPROCESS_PRIMOP_SUPPORTED
 static void 
 deleteThreadImmediately(StgTSO *tso)
 { // for forkProcess only:
@@ -3009,13 +2949,15 @@ deleteThreadImmediately(StgTSO *tso)
   if (tso->what_next == ThreadComplete || tso->what_next == ThreadKilled) {
       return;
   }
-#if defined(RTS_SUPPORTS_THREADS)
-  if (tso->why_blocked != BlockedOnCCall
-      && tso->why_blocked != BlockedOnCCall_NoUnblockExc)
-#endif
+
+  if (tso->why_blocked != BlockedOnCCall &&
+      tso->why_blocked != BlockedOnCCall_NoUnblockExc) {
     unblockThread(tso);
+  }
+
   tso->what_next = ThreadKilled;
 }
+#endif
 
 void
 raiseAsyncWithLock(StgTSO *tso, StgClosure *exception)
@@ -3347,14 +3289,12 @@ 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;
   case BlockedOnCCall_NoUnblockExc:
     fprintf(stderr,"is blocked on an external call (exceptions were already blocked)");
     break;
-#endif
   default:
     barf("printThreadBlockage: strange tso->why_blocked: %d for TSO %d (%d)",
         tso->why_blocked, tso->id, tso);