Merge zizzer:/z/m5/Bitkeeper/newmem

into  zazzer.eecs.umich.edu:/z/rdreslin/m5bk/newmemcleanest

--HG--
extra : convert_revision : 87f83c4edf6ea51adc767d98265d1e74c0fbb46f

src/mem/cache/base_cache.cc

@@ -104,9 +104,11 @@ BaseCache::CachePort::recvRetry()
             if (result)
                 drainList.pop_front();
         }
+        if (!result) return;
     }
     else if (!isCpuSide)
     {
+        if (!cache->doMasterRequest()) return;
         pkt = cache->getPacket();
         MSHR* mshr = (MSHR*)pkt->senderState;
         bool success = sendTiming(pkt);
@@ -178,10 +180,23 @@ BaseCache::CacheEvent::CacheEvent(CachePort *_cachePort, Packet *_pkt)
 void
 BaseCache::CacheEvent::process()
 {
+    if (!cachePort->drainList.empty()) {
+        //We have some responses to drain first
+        bool result = true;
+        while (result && !cachePort->drainList.empty()) {
+            result = cachePort->sendTiming(cachePort->drainList.front());
+            if (result)
+                cachePort->drainList.pop_front();
+        }
+        if (!result) return;
+    }
+
     if (!pkt)
     {
         if (!cachePort->isCpuSide)
         {
+            //For now, doMasterRequest somehow is still getting set
+            if (!cachePort->cache->doMasterRequest()) return;
             //MSHR
             pkt = cachePort->cache->getPacket();
             MSHR* mshr = (MSHR*) pkt->senderState;

src/mem/cache/base_cache.hh

@@ -392,11 +392,13 @@ class BaseCache : public MemObject
             blocked_causes[cause]++;
             blockedCycle = curTick;
         }
+        int old_state = blocked;
         if (!(blocked & flag)) {
             //Wasn't already blocked for this cause
             blocked |= flag;
             DPRINTF(Cache,"Blocking for cause %s\n", cause);
-            cpuSidePort->setBlocked();
+            if (!old_state)
+                cpuSidePort->setBlocked();
         }
     }
 
@@ -408,10 +410,12 @@ class BaseCache : public MemObject
     void setBlockedForSnoop(BlockedCause cause)
     {
         uint8_t flag = 1 << cause;
-        if (!(blocked & flag)) {
+        uint8_t old_state = blockedSnoop;
+        if (!(blockedSnoop & flag)) {
             //Wasn't already blocked for this cause
             blockedSnoop |= flag;
-            memSidePort->setBlocked();
+            if (!old_state)
+                memSidePort->setBlocked();
         }
     }
 

src/mem/cache/cache.hh

@@ -103,6 +103,7 @@ class Cache : public BaseCache
       * Used to append to target list, to cause an invalidation.
       */
     Packet * invalidatePkt;
+    Request *invalidateReq;
 
     /**
      * Temporarily move a block into a MSHR.

src/mem/cache/cache_impl.hh

@@ -163,10 +163,8 @@ Cache(const std::string &_name,
     prefetcher->setCache(this);
     prefetcher->setTags(tags);
     prefetcher->setBuffer(missQueue);
-#if 0
+    invalidateReq = new Request((Addr) NULL, blkSize, 0);
-    invalidatePkt = new Packet;
+    invalidatePkt = new Packet(invalidateReq, Packet::InvalidateReq, 0);
-    invalidatePkt->cmd = Packet::InvalidateReq;
-#endif
 }
 
 template<class TagStore, class Buffering, class Coherence>
@@ -267,6 +265,7 @@ template<class TagStore, class Buffering, class Coherence>
 Packet *
 Cache<TagStore,Buffering,Coherence>::getPacket()
 {
+    assert(missQueue->havePending());
     Packet * pkt = missQueue->getPacket();
     if (pkt) {
         if (!pkt->req->isUncacheable()) {
@@ -287,13 +286,27 @@ template<class TagStore, class Buffering, class Coherence>
 void
 Cache<TagStore,Buffering,Coherence>::sendResult(PacketPtr &pkt, MSHR* mshr, bool success)
 {
-    if (success) {
+    if (success && !(pkt->flags & NACKED_LINE)) {
               missQueue->markInService(pkt, mshr);
           //Temp Hack for UPGRADES
           if (pkt->cmd == Packet::UpgradeReq) {
-              handleResponse(pkt);
+              pkt->flags &= ~CACHE_LINE_FILL;
+              BlkType *blk = tags->findBlock(pkt);
+              CacheBlk::State old_state = (blk) ? blk->status : 0;
+              CacheBlk::State new_state = coherence->getNewState(pkt,old_state);
+              DPRINTF(Cache, "Block for blk addr %x moving from state %i to %i\n",
+                    pkt->getAddr() & (((ULL(1))<<48)-1), old_state, new_state);
+              //Set the state on the upgrade
+              memcpy(pkt->getPtr<uint8_t>(), blk->data, blkSize);
+              PacketList writebacks;
+              tags->handleFill(blk, mshr, new_state, writebacks, pkt);
+              assert(writebacks.empty());
+              missQueue->handleResponse(pkt, curTick + hitLatency);
           }
     } else if (pkt && !pkt->req->isUncacheable()) {
+        pkt->flags &= ~NACKED_LINE;
+        pkt->flags &= ~SATISFIED;
+        pkt->flags &= ~SNOOP_COMMIT;
         missQueue->restoreOrigCmd(pkt);
     }
 }
@@ -305,8 +318,9 @@ Cache<TagStore,Buffering,Coherence>::handleResponse(Packet * &pkt)
     BlkType *blk = NULL;
     if (pkt->senderState) {
         if (pkt->result == Packet::Nacked) {
-            pkt->reinitFromRequest();
+            //pkt->reinitFromRequest();
-            panic("Unimplemented NACK of packet\n");
+            warn("NACKs from devices not connected to the same bus not implemented\n");
+            return;
         }
         if (pkt->result == Packet::BadAddress) {
             //Make the response a Bad address and send it
@@ -397,7 +411,9 @@ Cache<TagStore,Buffering,Coherence>::snoop(Packet * &pkt)
                     assert(!(pkt->flags & SATISFIED));
                     pkt->flags |= SATISFIED;
                     pkt->flags |= NACKED_LINE;
-                    respondToSnoop(pkt, curTick + hitLatency);
+                    ///@todo NACK's from other levels
+                    //warn("NACKs from devices not connected to the same bus not implemented\n");
+                    //respondToSnoop(pkt, curTick + hitLatency);
                     return;
                 }
                 else {
@@ -410,7 +426,7 @@ Cache<TagStore,Buffering,Coherence>::snoop(Packet * &pkt)
                     //@todo Make it so that a read to a pending read can't be exclusive now.
 
                     //Set the address so find match works
-                    panic("Don't have invalidates yet\n");
+                    //panic("Don't have invalidates yet\n");
                     invalidatePkt->addrOverride(pkt->getAddr());
 
                     //Append the invalidate on
@@ -441,7 +457,7 @@ Cache<TagStore,Buffering,Coherence>::snoop(Packet * &pkt)
                         pkt->flags |= SHARED_LINE;
 
                         assert(pkt->isRead());
-                        Addr offset = pkt->getAddr() & ~(blkSize - 1);
+                        Addr offset = pkt->getAddr() & (blkSize - 1);
                         assert(offset < blkSize);
                         assert(pkt->getSize() <= blkSize);
                         assert(offset + pkt->getSize() <=blkSize);
@@ -462,16 +478,16 @@ Cache<TagStore,Buffering,Coherence>::snoop(Packet * &pkt)
     CacheBlk::State new_state;
     bool satisfy = coherence->handleBusRequest(pkt,blk,mshr, new_state);
     if (satisfy) {
-        DPRINTF(Cache, "Cache snooped a %s request and now supplying data,"
+        DPRINTF(Cache, "Cache snooped a %s request for addr %x and now supplying data,"
                 "new state is %i\n",
-                pkt->cmdString(), new_state);
+                pkt->cmdString(), blk_addr, new_state);
 
         tags->handleSnoop(blk, new_state, pkt);
         respondToSnoop(pkt, curTick + hitLatency);
         return;
     }
-    if (blk) DPRINTF(Cache, "Cache snooped a %s request, new state is %i\n",
+    if (blk) DPRINTF(Cache, "Cache snooped a %s request for addr %x, new state is %i\n",
-                     pkt->cmdString(), new_state);
+                     pkt->cmdString(), blk_addr, new_state);
     tags->handleSnoop(blk, new_state);
 }
 
@@ -689,15 +705,15 @@ Cache<TagStore,Buffering,Coherence>::snoopProbe(PacketPtr &pkt)
         CacheBlk::State new_state = 0;
         bool satisfy = coherence->handleBusRequest(pkt,blk,mshr, new_state);
         if (satisfy) {
-            DPRINTF(Cache, "Cache snooped a %s request and now supplying data,"
+            DPRINTF(Cache, "Cache snooped a %s request for addr %x and now supplying data,"
                     "new state is %i\n",
-                    pkt->cmdString(), new_state);
+                    pkt->cmdString(), blk_addr, new_state);
 
             tags->handleSnoop(blk, new_state, pkt);
             return hitLatency;
         }
-        if (blk) DPRINTF(Cache, "Cache snooped a %s request, new state is %i\n",
+        if (blk) DPRINTF(Cache, "Cache snooped a %s request for addr %x, new state is %i\n",
-                     pkt->cmdString(), new_state);
+                     pkt->cmdString(), blk_addr, new_state);
         tags->handleSnoop(blk, new_state);
         return 0;
 }

src/mem/cache/miss/miss_queue.cc

@@ -515,6 +515,14 @@ MissQueue::setBusCmd(Packet * &pkt, Packet::Command cmd)
     assert(pkt->senderState != 0);
     MSHR * mshr = (MSHR*)pkt->senderState;
     mshr->originalCmd = pkt->cmd;
+    if (cmd == Packet::UpgradeReq || cmd == Packet::InvalidateReq) {
+        pkt->flags |= NO_ALLOCATE;
+        pkt->flags &= ~CACHE_LINE_FILL;
+    }
+    else if (!pkt->req->isUncacheable() && !pkt->isNoAllocate() &&
+             (cmd & (1 << 6)/*NeedsResponse*/)) {
+        pkt->flags |= CACHE_LINE_FILL;
+    }
     if (pkt->isCacheFill() || pkt->isNoAllocate())
         pkt->cmd = cmd;
 }

src/mem/cache/miss/mshr.cc

@@ -100,6 +100,7 @@ MSHR::deallocate()
 {
     assert(targets.empty());
     assert(ntargets == 0);
+    delete pkt;
     pkt = NULL;
     inService = false;
     //allocIter = NULL;

src/mem/cache/miss/mshr_queue.cc

@@ -213,7 +213,7 @@ void
 MSHRQueue::markInService(MSHR* mshr)
 {
     //assert(mshr == pendingList.front());
-    if (!mshr->pkt->needsResponse()) {
+    if (!(mshr->pkt->needsResponse() || mshr->pkt->cmd == Packet::UpgradeReq)) {
         assert(mshr->getNumTargets() == 0);
         deallocate(mshr);
         return;

src/mem/packet.hh

@@ -58,6 +58,8 @@ typedef std::list<PacketPtr> PacketList;
 #define NO_ALLOCATE 1 << 5
 #define SNOOP_COMMIT 1 << 6
 
+//for now.  @todo fix later
+#define NUM_MEM_CMDS 1 << 11
 /**
  * A Packet is used to encapsulate a transfer between two objects in
  * the memory system (e.g., the L1 and L2 cache).  (In contrast, a
@@ -173,13 +175,10 @@ class Packet
         NeedsResponse	= 1 << 6,
         IsSWPrefetch    = 1 << 7,
         IsHWPrefetch    = 1 << 8,
-        HasData		= 1 << 9
+        IsUpgrade       = 1 << 9,
+        HasData		= 1 << 10
     };
 
-//For statistics we need max number of commands, hard code it at
-//20 for now.  @todo fix later
-#define NUM_MEM_CMDS 1 << 10
-
   public:
     /** List of all commands associated with a packet. */
     enum Command
@@ -199,7 +198,7 @@ class Packet
                                 | NeedsResponse | HasData,
         InvalidateReq   = IsInvalidate | IsRequest,
         WriteInvalidateReq = IsWrite | IsInvalidate | IsRequest | HasData,
-        UpgradeReq      = IsInvalidate | IsRequest,
+        UpgradeReq      = IsInvalidate | IsRequest | IsUpgrade,
         ReadExReq       = IsRead | IsInvalidate | IsRequest | NeedsResponse,
         ReadExResp      = IsRead | IsInvalidate | IsResponse
                                 | NeedsResponse | HasData
@@ -326,6 +325,10 @@ class Packet
         int icmd = (int)cmd;
         icmd &= ~(IsRequest);
         icmd |= IsResponse;
+        if (isRead())
+            icmd |= HasData;
+        if (isWrite())
+            icmd &= ~HasData;
         cmd = (Command)icmd;
         dest = src;
         srcValid = false;

tests/configs/memtest.py

@@ -36,7 +36,7 @@ from m5.objects import *
 class L1(BaseCache):
     latency = 1
     block_size = 64
-    mshrs = 4
+    mshrs = 12
     tgts_per_mshr = 8
     protocol = CoherenceProtocol(protocol='moesi')
 
@@ -46,14 +46,14 @@ class L1(BaseCache):
 
 class L2(BaseCache):
     block_size = 64
-    latency = 100
+    latency = 10
     mshrs = 92
     tgts_per_mshr = 16
     write_buffers = 8
 
 #MAX CORES IS 8 with the fals sharing method
 nb_cores = 8
-cpus = [ MemTest(max_loads=1e12) for i in xrange(nb_cores) ]
+cpus = [ MemTest(max_loads=1e12, percent_uncacheable=0) for i in xrange(nb_cores) ]
 
 # system simulated
 system = System(cpu = cpus, funcmem = PhysicalMemory(),
@@ -61,7 +61,7 @@ system = System(cpu = cpus, funcmem = PhysicalMemory(),
 
 # l2cache & bus
 system.toL2Bus = Bus()
-system.l2c = L2(size='4MB', assoc=8)
+system.l2c = L2(size='64kB', assoc=8)
 system.l2c.cpu_side = system.toL2Bus.port
 
 # connect l2c to membus
@@ -90,5 +90,4 @@ system.physmem.port = system.membus.port
 
 root = Root( system = system )
 root.system.mem_mode = 'timing'
-#root.trace.flags="InstExec"
+root.trace.flags="Cache"
-root.trace.flags="Bus"