X-Git-Url: http://git.megacz.com/?a=blobdiff_plain;f=rts%2FCapability.h;h=d2fcc5ed7bfdb90449773560938618a59cc18a72;hb=de75026f5a48d3d052135a973ab4dff76c5b20f5;hp=89b813f6d3f725e77bc7dabfc30886ae17cdd486;hpb=17a596022d01987d5ffe7db0742aa8db267b8cc3;p=ghc-hetmet.git

diff --git a/rts/Capability.h b/rts/Capability.h
index 89b813f..d2fcc5e 100644
--- a/rts/Capability.h
+++ b/rts/Capability.h
@@ -46,6 +46,9 @@ struct Capability_ {
     // catching unsafe call-ins.
     rtsBool in_haskell;
 
+    // true if this Capability is currently in the GC
+    rtsBool in_gc;
+
     // The run queue.  The Task owning this Capability has exclusive
     // access to its run queue, so can wake up threads without
     // taking a lock, and the common path through the scheduler is
@@ -62,10 +65,12 @@ struct Capability_ {
     Task *suspended_ccalling_tasks;
 
     // One mutable list per generation, so we don't need to take any
-    // locks when updating an old-generation thunk.  These
-    // mini-mut-lists are moved onto the respective gen->mut_list at
-    // each GC.
+    // locks when updating an old-generation thunk.  This also lets us
+    // keep track of which closures this CPU has been mutating, so we
+    // can traverse them using the right thread during GC and avoid
+    // unnecessarily moving the data from one cache to another.
     bdescr **mut_lists;
+    bdescr **saved_mut_lists; // tmp use during GC
 
     // Context switch flag. We used to have one global flag, now one 
     // per capability. Locks required  : none (conflicts are harmless)
@@ -106,8 +111,13 @@ struct Capability_ {
     StgTRecChunk *free_trec_chunks;
     StgTRecHeader *free_trec_headers;
     nat transaction_tokens;
-
-}; // typedef Capability, defined in RtsAPI.h
+} // typedef Capability is defined in RtsAPI.h
+  // Capabilities are stored in an array, so make sure that adjacent
+  // Capabilities don't share any cache-lines:
+#ifndef mingw32_HOST_OS
+  ATTRIBUTE_ALIGNED(64)
+#endif
+  ;
 
 
 #if defined(THREADED_RTS)
@@ -138,7 +148,7 @@ struct Capability_ {
 INLINE_HEADER Capability *
 regTableToCapability (StgRegTable *reg)
 {
-    return (Capability *)((void *)((unsigned char*)reg - sizeof(StgFunTable)));
+    return (Capability *)((void *)((unsigned char*)reg - STG_FIELD_OFFSET(Capability,r)));
 }
 
 // Initialise the available capabilities.
@@ -180,6 +190,8 @@ extern Capability *capabilities;
 extern Capability *last_free_capability;
 
 // GC indicator, in scope for the scheduler
+#define PENDING_GC_SEQ 1
+#define PENDING_GC_PAR 2
 extern volatile StgWord waiting_for_gc;
 
 // Acquires a capability at a return point.  If *cap is non-NULL, then
@@ -226,6 +238,7 @@ void wakeupThreadOnCapability (Capability *my_cap, Capability *other_cap,
 // need to service some global event.
 //
 void prodOneCapability (void);
+void prodCapability (Capability *cap, Task *task);
 
 // Similar to prodOneCapability(), but prods all of them.
 //
@@ -240,9 +253,13 @@ void shutdownCapability (Capability *cap, Task *task, rtsBool wait_foreign);
 //
 rtsBool tryGrabCapability (Capability *cap, Task *task);
 
-// Try to steal a spark from other Capabilities
+// Try to find a spark to run
+//
+StgClosure *findSpark (Capability *cap);
+
+// True if any capabilities have sparks
 //
-rtsBool stealWork (Capability *cap);
+rtsBool anySparks (void);
 
 INLINE_HEADER rtsBool emptySparkPoolCap (Capability *cap);
 INLINE_HEADER nat     sparkPoolSizeCap  (Capability *cap);
@@ -259,9 +276,10 @@ extern void grabCapability (Capability **pCap);
 
 // cause all capabilities to context switch as soon as possible.
 void setContextSwitches(void);
+INLINE_HEADER void contextSwitchCapability(Capability *cap);
 
-// Free a capability on exit
-void freeCapability (Capability *cap);
+// Free all capabilities
+void freeCapabilities (void);
 
 // FOr the GC:
 void markSomeCapabilities (evac_fn evac, void *user, nat i0, nat delta, 
@@ -305,4 +323,16 @@ discardSparksCap (Capability *cap)
 { return discardSparks(cap->sparks); }
 #endif
 
+INLINE_HEADER void
+contextSwitchCapability (Capability *cap)
+{
+    // setting HpLim to NULL ensures that the next heap check will
+    // fail, and the thread will return to the scheduler.
+    cap->r.rHpLim = NULL;
+    // But just in case it didn't work (the target thread might be
+    // modifying HpLim at the same time), we set the end-of-block
+    // context-switch flag too:
+    cap->context_switch = 1;
+}
+
 #endif /* CAPABILITY_H */