Merge zizzer:/bk/newmem

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

--HG--
extra : convert_revision : 498304c24435437f8f1942bb8aeafe69ba22a089

8 files changed, 297 insertions, 63 deletions

diff --git a/src/mem/cache/base_cache.cc b/src/mem/cache/base_cache.cc
index c26d7782b..c16cb6945 100644
--- a/src/mem/cache/base_cache.cc
+++ b/src/mem/cache/base_cache.cc
@@ -102,21 +102,56 @@ BaseCache::CachePort::recvAtomic(PacketPtr pkt)
     return cache->doAtomicAccess(pkt, isCpuSide);
 }
 
-void
-BaseCache::CachePort::recvFunctional(PacketPtr pkt)
+bool
+BaseCache::CachePort::checkFunctional(PacketPtr pkt)
 {
     //Check storage here first
     list<PacketPtr>::iterator i = drainList.begin();
-    list<PacketPtr>::iterator end = drainList.end();
-    for (; i != end; ++i) {
+    list<PacketPtr>::iterator iend = drainList.end();
+    bool notDone = true;
+    while (i != iend && notDone) {
         PacketPtr target = *i;
         // If the target contains data, and it overlaps the
         // probed request, need to update data
         if (target->intersect(pkt)) {
-            fixPacket(pkt, target);
+            DPRINTF(Cache, "Functional %s access to blk_addr %x intersects a drain\n",
+                    pkt->cmdString(), pkt->getAddr() & ~(cache->getBlockSize() - 1));
+            notDone = fixPacket(pkt, target);
         }
+        i++;
     }
-    cache->doFunctionalAccess(pkt, isCpuSide);
+    //Also check the response not yet ready to be on the list
+    std::list<std::pair<Tick,PacketPtr> >::iterator j = transmitList.begin();
+    std::list<std::pair<Tick,PacketPtr> >::iterator jend = transmitList.end();
+
+    while (j != jend && notDone) {
+        PacketPtr target = j->second;
+        // If the target contains data, and it overlaps the
+        // probed request, need to update data
+        if (target->intersect(pkt)) {
+            DPRINTF(Cache, "Functional %s access to blk_addr %x intersects a response\n",
+                    pkt->cmdString(), pkt->getAddr() & ~(cache->getBlockSize() - 1));
+            notDone = fixDelayedResponsePacket(pkt, target);
+        }
+        j++;
+    }
+    return notDone;
+}
+
+void
+BaseCache::CachePort::recvFunctional(PacketPtr pkt)
+{
+    bool notDone = checkFunctional(pkt);
+    if (notDone)
+        cache->doFunctionalAccess(pkt, isCpuSide);
+}
+
+void
+BaseCache::CachePort::checkAndSendFunctional(PacketPtr pkt)
+{
+    bool notDone = checkFunctional(pkt);
+    if (notDone)
+        sendFunctional(pkt);
 }
 
 void
@@ -135,7 +170,7 @@ BaseCache::CachePort::recvRetry()
                 isCpuSide && cache->doSlaveRequest()) {
 
                 DPRINTF(CachePort, "%s has more responses/requests\n", name());
-                BaseCache::CacheEvent * reqCpu = new BaseCache::CacheEvent(this);
+                BaseCache::CacheEvent * reqCpu = new BaseCache::CacheEvent(this, false);
                 reqCpu->schedule(curTick + 1);
             }
             waitingOnRetry = false;
@@ -176,7 +211,7 @@ BaseCache::CachePort::recvRetry()
         {
             DPRINTF(CachePort, "%s has more requests\n", name());
             //Still more to issue, rerequest in 1 cycle
-            BaseCache::CacheEvent * reqCpu = new BaseCache::CacheEvent(this);
+            BaseCache::CacheEvent * reqCpu = new BaseCache::CacheEvent(this, false);
             reqCpu->schedule(curTick + 1);
         }
     }
@@ -194,7 +229,7 @@ BaseCache::CachePort::recvRetry()
         {
             DPRINTF(CachePort, "%s has more requests\n", name());
             //Still more to issue, rerequest in 1 cycle
-            BaseCache::CacheEvent * reqCpu = new BaseCache::CacheEvent(this);
+            BaseCache::CacheEvent * reqCpu = new BaseCache::CacheEvent(this, false);
             reqCpu->schedule(curTick + 1);
         }
     }
@@ -226,23 +261,19 @@ BaseCache::CachePort::clearBlocked()
     }
 }
 
-BaseCache::CacheEvent::CacheEvent(CachePort *_cachePort)
-    : Event(&mainEventQueue, CPU_Tick_Pri), cachePort(_cachePort)
+BaseCache::CacheEvent::CacheEvent(CachePort *_cachePort, bool _newResponse)
+    : Event(&mainEventQueue, CPU_Tick_Pri), cachePort(_cachePort),
+      newResponse(_newResponse)
 {
-    this->setFlags(AutoDelete);
+    if (!newResponse)
+        this->setFlags(AutoDelete);
     pkt = NULL;
 }
 
-BaseCache::CacheEvent::CacheEvent(CachePort *_cachePort, PacketPtr _pkt)
-    : Event(&mainEventQueue, CPU_Tick_Pri), cachePort(_cachePort), pkt(_pkt)
-{
-    this->setFlags(AutoDelete);
-}
-
 void
 BaseCache::CacheEvent::process()
 {
-    if (!pkt)
+    if (!newResponse)
     {
         if (cachePort->waitingOnRetry) return;
        //We have some responses to drain first
@@ -322,8 +353,16 @@ BaseCache::CacheEvent::process()
         }
         return;
     }
-    //Response
-    //Know the packet to send
+    //Else it's a response
+    assert(cachePort->transmitList.size());
+    assert(cachePort->transmitList.front().first <= curTick);
+    pkt = cachePort->transmitList.front().second;
+    cachePort->transmitList.pop_front();
+    if (!cachePort->transmitList.empty()) {
+        Tick time = cachePort->transmitList.front().first;
+        schedule(time <= curTick ? curTick+1 : time);
+    }
+
     if (pkt->flags & NACKED_LINE)
         pkt->result = Packet::Nacked;
     else
@@ -343,7 +382,7 @@ BaseCache::CacheEvent::process()
     }
 
     // Check if we're done draining once this list is empty
-    if (cachePort->drainList.empty())
+    if (cachePort->drainList.empty() && cachePort->transmitList.empty())
         cachePort->cache->checkDrain();
 }
 
@@ -358,8 +397,10 @@ BaseCache::getPort(const std::string &if_name, int idx)
 {
     if (if_name == "")
     {
-        if(cpuSidePort == NULL)
+        if(cpuSidePort == NULL) {
             cpuSidePort = new CachePort(name() + "-cpu_side_port", this, true);
+            sendEvent = new CacheEvent(cpuSidePort, true);
+        }
         return cpuSidePort;
     }
     else if (if_name == "functional")
@@ -368,8 +409,10 @@ BaseCache::getPort(const std::string &if_name, int idx)
     }
     else if (if_name == "cpu_side")
     {
-        if(cpuSidePort == NULL)
+        if(cpuSidePort == NULL) {
             cpuSidePort = new CachePort(name() + "-cpu_side_port", this, true);
+            sendEvent = new CacheEvent(cpuSidePort, true);
+        }
         return cpuSidePort;
     }
     else if (if_name == "mem_side")
@@ -377,6 +420,7 @@ BaseCache::getPort(const std::string &if_name, int idx)
         if (memSidePort != NULL)
             panic("Already have a mem side for this cache\n");
         memSidePort = new CachePort(name() + "-mem_side_port", this, false);
+        memSendEvent = new CacheEvent(memSidePort, true);
         return memSidePort;
     }
     else panic("Port name %s unrecognized\n", if_name);
diff --git a/src/mem/cache/base_cache.hh b/src/mem/cache/base_cache.hh
index ea7544fbb..584c2d5df 100644
--- a/src/mem/cache/base_cache.hh
+++ b/src/mem/cache/base_cache.hh
@@ -105,7 +105,11 @@ class BaseCache : public MemObject
 
         void clearBlocked();
 
-        bool canDrain() { return drainList.empty(); }
+        bool checkFunctional(PacketPtr pkt);
+
+        void checkAndSendFunctional(PacketPtr pkt);
+
+        bool canDrain() { return drainList.empty() && transmitList.empty(); }
 
         bool blocked;
 
@@ -117,15 +121,16 @@ class BaseCache : public MemObject
 
         std::list<PacketPtr> drainList;
 
+        std::list<std::pair<Tick,PacketPtr> > transmitList;
     };
 
     struct CacheEvent : public Event
     {
         CachePort *cachePort;
         PacketPtr pkt;
+        bool newResponse;
 
-        CacheEvent(CachePort *_cachePort);
-        CacheEvent(CachePort *_cachePort, PacketPtr _pkt);
+        CacheEvent(CachePort *_cachePort, bool response);
         void process();
         const char *description();
     };
@@ -133,6 +138,9 @@ class BaseCache : public MemObject
   public: //Made public so coherence can get at it.
     CachePort *cpuSidePort;
 
+    CacheEvent *sendEvent;
+    CacheEvent *memSendEvent;
+
   protected:
     CachePort *memSidePort;
 
@@ -353,6 +361,12 @@ class BaseCache : public MemObject
         snoopRangesSent = false;
     }
 
+    ~BaseCache()
+    {
+        delete sendEvent;
+        delete memSendEvent;
+    }
+
     virtual void init();
 
     /**
@@ -467,7 +481,8 @@ class BaseCache : public MemObject
     {
         if (!doMasterRequest() && !memSidePort->waitingOnRetry)
         {
-            BaseCache::CacheEvent * reqCpu = new BaseCache::CacheEvent(memSidePort);
+            BaseCache::CacheEvent * reqCpu =
+                new BaseCache::CacheEvent(memSidePort, false);
             reqCpu->schedule(time);
         }
         uint8_t flag = 1<<cause;
@@ -503,7 +518,8 @@ class BaseCache : public MemObject
     {
         if (!doSlaveRequest() && !cpuSidePort->waitingOnRetry)
         {
-            BaseCache::CacheEvent * reqCpu = new BaseCache::CacheEvent(cpuSidePort);
+            BaseCache::CacheEvent * reqCpu =
+                new BaseCache::CacheEvent(cpuSidePort, false);
             reqCpu->schedule(time);
         }
         uint8_t flag = 1<<cause;
@@ -528,9 +544,44 @@ class BaseCache : public MemObject
      */
     void respond(PacketPtr pkt, Tick time)
     {
+        assert(time >= curTick);
         if (pkt->needsResponse()) {
-            CacheEvent *reqCpu = new CacheEvent(cpuSidePort, pkt);
+/*            CacheEvent *reqCpu = new CacheEvent(cpuSidePort, pkt);
             reqCpu->schedule(time);
+*/
+            if (cpuSidePort->transmitList.empty()) {
+                assert(!sendEvent->scheduled());
+                sendEvent->schedule(time);
+                cpuSidePort->transmitList.push_back(std::pair<Tick,PacketPtr>
+                                                    (time,pkt));
+                return;
+            }
+
+            // something is on the list and this belongs at the end
+            if (time >= cpuSidePort->transmitList.back().first) {
+                cpuSidePort->transmitList.push_back(std::pair<Tick,PacketPtr>
+                                                    (time,pkt));
+                return;
+            }
+            // Something is on the list and this belongs somewhere else
+            std::list<std::pair<Tick,PacketPtr> >::iterator i =
+                cpuSidePort->transmitList.begin();
+            std::list<std::pair<Tick,PacketPtr> >::iterator end =
+                cpuSidePort->transmitList.end();
+            bool done = false;
+
+            while (i != end && !done) {
+                if (time < i->first) {
+                    if (i == cpuSidePort->transmitList.begin()) {
+                        //Inserting at begining, reschedule
+                        sendEvent->reschedule(time);
+                    }
+                    cpuSidePort->transmitList.insert(i,std::pair<Tick,PacketPtr>
+                                                     (time,pkt));
+                    done = true;
+                }
+                i++;
+            }
         }
         else {
             if (pkt->cmd != Packet::UpgradeReq)
@@ -548,12 +599,48 @@ class BaseCache : public MemObject
      */
     void respondToMiss(PacketPtr pkt, Tick time)
     {
+        assert(time >= curTick);
         if (!pkt->req->isUncacheable()) {
-            missLatency[pkt->cmdToIndex()][0/*pkt->req->getThreadNum()*/] += time - pkt->time;
+            missLatency[pkt->cmdToIndex()][0/*pkt->req->getThreadNum()*/] +=
+                time - pkt->time;
         }
         if (pkt->needsResponse()) {
-            CacheEvent *reqCpu = new CacheEvent(cpuSidePort, pkt);
+/*            CacheEvent *reqCpu = new CacheEvent(cpuSidePort, pkt);
             reqCpu->schedule(time);
+*/
+            if (cpuSidePort->transmitList.empty()) {
+                assert(!sendEvent->scheduled());
+                sendEvent->schedule(time);
+                cpuSidePort->transmitList.push_back(std::pair<Tick,PacketPtr>
+                                                    (time,pkt));
+                return;
+            }
+
+            // something is on the list and this belongs at the end
+            if (time >= cpuSidePort->transmitList.back().first) {
+                cpuSidePort->transmitList.push_back(std::pair<Tick,PacketPtr>
+                                                    (time,pkt));
+                return;
+            }
+            // Something is on the list and this belongs somewhere else
+            std::list<std::pair<Tick,PacketPtr> >::iterator i =
+                cpuSidePort->transmitList.begin();
+            std::list<std::pair<Tick,PacketPtr> >::iterator end =
+                cpuSidePort->transmitList.end();
+            bool done = false;
+
+            while (i != end && !done) {
+                if (time < i->first) {
+                    if (i == cpuSidePort->transmitList.begin()) {
+                        //Inserting at begining, reschedule
+                        sendEvent->reschedule(time);
+                    }
+                    cpuSidePort->transmitList.insert(i,std::pair<Tick,PacketPtr>
+                                                     (time,pkt));
+                    done = true;
+                }
+                i++;
+            }
         }
         else {
             if (pkt->cmd != Packet::UpgradeReq)
@@ -570,9 +657,43 @@ class BaseCache : public MemObject
      */
     void respondToSnoop(PacketPtr pkt, Tick time)
     {
+        assert(time >= curTick);
         assert (pkt->needsResponse());
-        CacheEvent *reqMem = new CacheEvent(memSidePort, pkt);
+/*        CacheEvent *reqMem = new CacheEvent(memSidePort, pkt);
         reqMem->schedule(time);
+*/
+        if (memSidePort->transmitList.empty()) {
+            assert(!memSendEvent->scheduled());
+            memSendEvent->schedule(time);
+            memSidePort->transmitList.push_back(std::pair<Tick,PacketPtr>
+                                                (time,pkt));
+            return;
+        }
+
+        // something is on the list and this belongs at the end
+        if (time >= memSidePort->transmitList.back().first) {
+            memSidePort->transmitList.push_back(std::pair<Tick,PacketPtr>
+                                                (time,pkt));
+            return;
+        }
+        // Something is on the list and this belongs somewhere else
+        std::list<std::pair<Tick,PacketPtr> >::iterator i =
+            memSidePort->transmitList.begin();
+        std::list<std::pair<Tick,PacketPtr> >::iterator end =
+            memSidePort->transmitList.end();
+        bool done = false;
+
+        while (i != end && !done) {
+            if (time < i->first) {
+                if (i == memSidePort->transmitList.begin()) {
+                    //Inserting at begining, reschedule
+                    memSendEvent->reschedule(time);
+                }
+                memSidePort->transmitList.insert(i,std::pair<Tick,PacketPtr>(time,pkt));
+                done = true;
+            }
+            i++;
+        }
     }
 
     /**
diff --git a/src/mem/cache/cache_impl.hh b/src/mem/cache/cache_impl.hh
index 9bb72e85c..df59b0a4f 100644
--- a/src/mem/cache/cache_impl.hh
+++ b/src/mem/cache/cache_impl.hh
@@ -53,6 +53,8 @@
 
 #include "sim/sim_exit.hh" // for SimExitEvent
 
+bool SIGNAL_NACK_HACK;
+
 template<class TagStore, class Buffering, class Coherence>
 bool
 Cache<TagStore,Buffering,Coherence>::
@@ -242,6 +244,11 @@ Cache<TagStore,Buffering,Coherence>::access(PacketPtr &pkt)
         missQueue->handleMiss(pkt, size, curTick + hitLatency);
     }
 
+    if (pkt->cmd == Packet::Writeback) {
+        //Need to clean up the packet on a writeback miss, but leave the request
+        delete pkt;
+    }
+
     return true;
 }
 
@@ -265,6 +272,7 @@ Cache<TagStore,Buffering,Coherence>::getPacket()
 
     assert(!doMasterRequest() || missQueue->havePending());
     assert(!pkt || pkt->time <= curTick);
+    SIGNAL_NACK_HACK = false;
     return pkt;
 }
 
@@ -273,16 +281,15 @@ void
 Cache<TagStore,Buffering,Coherence>::sendResult(PacketPtr &pkt, MSHR* mshr,
                                                 bool success)
 {
-    if (success && !(pkt && (pkt->flags & NACKED_LINE))) {
-        if (!mshr->pkt->needsResponse()
-            && !(mshr->pkt->cmd == Packet::UpgradeReq)
-            && (pkt && (pkt->flags & SATISFIED))) {
-            //Writeback, clean up the non copy version of the packet
-            delete pkt;
-        }
+    if (success && !(SIGNAL_NACK_HACK)) {
+        //Remember if it was an upgrade because writeback MSHR's are removed
+        //in Mark in Service
+        bool upgrade = (mshr->pkt && mshr->pkt->cmd == Packet::UpgradeReq);
+
         missQueue->markInService(mshr->pkt, mshr);
+
         //Temp Hack for UPGRADES
-        if (mshr->pkt && mshr->pkt->cmd == Packet::UpgradeReq) {
+        if (upgrade) {
             assert(pkt);  //Upgrades need to be fixed
             pkt->flags &= ~CACHE_LINE_FILL;
             BlkType *blk = tags->findBlock(pkt);
@@ -300,6 +307,7 @@ Cache<TagStore,Buffering,Coherence>::sendResult(PacketPtr &pkt, MSHR* mshr,
         }
     } else if (pkt && !pkt->req->isUncacheable()) {
         pkt->flags &= ~NACKED_LINE;
+        SIGNAL_NACK_HACK = false;
         pkt->flags &= ~SATISFIED;
         pkt->flags &= ~SNOOP_COMMIT;
 
@@ -333,6 +341,8 @@ Cache<TagStore,Buffering,Coherence>::handleResponse(PacketPtr &pkt)
         DPRINTF(Cache, "Handling reponse to %x\n", pkt->getAddr());
 
         if (pkt->isCacheFill() && !pkt->isNoAllocate()) {
+            DPRINTF(Cache, "Block for addr %x being updated in Cache\n",
+                    pkt->getAddr());
             blk = tags->findBlock(pkt);
             CacheBlk::State old_state = (blk) ? blk->status : 0;
             PacketList writebacks;
@@ -402,6 +412,7 @@ Cache<TagStore,Buffering,Coherence>::snoop(PacketPtr &pkt)
                     assert(!(pkt->flags & SATISFIED));
                     pkt->flags |= SATISFIED;
                     pkt->flags |= NACKED_LINE;
+                    SIGNAL_NACK_HACK = true;
                     ///@todo NACK's from other levels
                     //warn("NACKs from devices not connected to the same bus "
                     //"not implemented\n");
@@ -474,6 +485,13 @@ Cache<TagStore,Buffering,Coherence>::snoop(PacketPtr &pkt)
     }
     CacheBlk::State new_state;
     bool satisfy = coherence->handleBusRequest(pkt,blk,mshr, new_state);
+
+    if (blk && mshr && !mshr->inService && new_state == 0) {
+            //There was a outstanding write to a shared block, not need ReadEx
+            //not update, so change No Allocate param in MSHR
+            mshr->pkt->flags &= ~NO_ALLOCATE;
+    }
+
     if (satisfy) {
         DPRINTF(Cache, "Cache snooped a %s request for addr %x and "
                 "now supplying data, new state is %i\n",
@@ -486,6 +504,7 @@ Cache<TagStore,Buffering,Coherence>::snoop(PacketPtr &pkt)
     if (blk)
         DPRINTF(Cache, "Cache snooped a %s request for addr %x, "
                 "new state is %i\n", pkt->cmdString(), blk_addr, new_state);
+
     tags->handleSnoop(blk, new_state);
 }
 
@@ -534,9 +553,9 @@ Cache<TagStore,Buffering,Coherence>::probe(PacketPtr &pkt, bool update,
         }
     }
 
-    if (!update && (pkt->isWrite() || (otherSidePort == cpuSidePort))) {
+    if (!update && (otherSidePort == cpuSidePort)) {
         // Still need to change data in all locations.
-        otherSidePort->sendFunctional(pkt);
+        otherSidePort->checkAndSendFunctional(pkt);
         if (pkt->isRead() && pkt->result == Packet::Success)
             return 0;
     }
@@ -560,30 +579,33 @@ Cache<TagStore,Buffering,Coherence>::probe(PacketPtr &pkt, bool update,
     missQueue->findWrites(blk_addr, writes);
 
     if (!update) {
+        bool notDone = !(pkt->flags & SATISFIED); //Hit in cache (was a block)
         // Check for data in MSHR and writebuffer.
         if (mshr) {
             MSHR::TargetList *targets = mshr->getTargetList();
             MSHR::TargetList::iterator i = targets->begin();
             MSHR::TargetList::iterator end = targets->end();
-            for (; i != end; ++i) {
+            for (; i != end && notDone; ++i) {
                 PacketPtr target = *i;
                 // If the target contains data, and it overlaps the
                 // probed request, need to update data
                 if (target->intersect(pkt)) {
-                    fixPacket(pkt, target);
+                    DPRINTF(Cache, "Functional %s access to blk_addr %x intersects a MSHR\n",
+                            pkt->cmdString(), blk_addr);
+                    notDone = fixPacket(pkt, target);
                 }
             }
         }
-        for (int i = 0; i < writes.size(); ++i) {
+        for (int i = 0; i < writes.size() && notDone; ++i) {
             PacketPtr write = writes[i]->pkt;
             if (write->intersect(pkt)) {
-                fixPacket(pkt, write);
+                DPRINTF(Cache, "Functional %s access to blk_addr %x intersects a writeback\n",
+                        pkt->cmdString(), blk_addr);
+                notDone = fixPacket(pkt, write);
             }
         }
-        if (pkt->isRead()
-            && pkt->result != Packet::Success
-            && otherSidePort == memSidePort) {
-            otherSidePort->sendFunctional(pkt);
+        if (notDone && otherSidePort == memSidePort) {
+            otherSidePort->checkAndSendFunctional(pkt);
             assert(pkt->result == Packet::Success);
         }
         return 0;
diff --git a/src/mem/cache/miss/mshr_queue.cc b/src/mem/cache/miss/mshr_queue.cc
index d3a7a7933..6cb62429d 100644
--- a/src/mem/cache/miss/mshr_queue.cc
+++ b/src/mem/cache/miss/mshr_queue.cc
@@ -198,11 +198,6 @@ MSHRQueue::markInService(MSHR* mshr)
     //assert(mshr == pendingList.front());
     if (!mshr->pkt->needsResponse() && !(mshr->pkt->cmd == Packet::UpgradeReq)) {
         assert(mshr->getNumTargets() == 0);
-        if ((mshr->pkt->flags & SATISFIED) && (mshr->pkt->cmd == Packet::Writeback)) {
-            //Writeback hit, so delete it
-            //otherwise the consumer will delete it
-            delete mshr->pkt->req;
-        }
         deallocate(mshr);
         return;
     }
diff --git a/src/mem/packet.cc b/src/mem/packet.cc
index a342af634..e2faf4527 100644
--- a/src/mem/packet.cc
+++ b/src/mem/packet.cc
@@ -144,6 +144,24 @@ Packet::intersect(PacketPtr p)
 }
 
 bool
+fixDelayedResponsePacket(PacketPtr func, PacketPtr timing)
+{
+    bool result;
+
+    if (timing->isRead() || timing->isWrite()) {
+        timing->toggleData();
+        result = fixPacket(func, timing);
+        timing->toggleData();
+    }
+    else {
+        //Don't toggle if it isn't a read/write response
+        result = fixPacket(func, timing);
+    }
+
+    return result;
+}
+
+bool
 fixPacket(PacketPtr func, PacketPtr timing)
 {
     Addr funcStart      = func->getAddr();
@@ -168,6 +186,7 @@ fixPacket(PacketPtr func, PacketPtr timing)
             memcpy(func->getPtr<uint8_t>(), timing->getPtr<uint8_t>() +
                     funcStart - timingStart, func->getSize());
             func->result = Packet::Success;
+            func->flags |= SATISFIED;
             return false;
         } else {
             // In this case the timing packet only partially satisfies the
@@ -182,11 +201,11 @@ fixPacket(PacketPtr func, PacketPtr timing)
         if (funcStart >= timingStart) {
             memcpy(timing->getPtr<uint8_t>() + (funcStart - timingStart),
                    func->getPtr<uint8_t>(),
-                   std::min(funcEnd, timingEnd) - funcStart);
+                   (std::min(funcEnd, timingEnd) - funcStart) + 1);
         } else { // timingStart > funcStart
             memcpy(timing->getPtr<uint8_t>(),
                    func->getPtr<uint8_t>() + (timingStart - funcStart),
-                   std::min(funcEnd, timingEnd) - timingStart);
+                   (std::min(funcEnd, timingEnd) - timingStart) + 1);
         }
         // we always want to keep going with a write
         return true;
diff --git a/src/mem/packet.hh b/src/mem/packet.hh
index cb97dd036..2bc51bf12 100644
--- a/src/mem/packet.hh
+++ b/src/mem/packet.hh
@@ -344,6 +344,13 @@ class Packet
         srcValid = false;
     }
 
+
+    void toggleData() {
+        int icmd = (int)cmd;
+        icmd ^= HasData;
+        cmd = (Command)icmd;
+    }
+
     /**
      * Take a request packet and modify it in place to be suitable for
      * returning as a response to that request.
@@ -448,7 +455,6 @@ class Packet
     bool intersect(PacketPtr p);
 };
 
-
 /** This function given a functional packet and a timing packet either satisfies
  * the timing packet, or updates the timing packet to reflect the updated state
  * in the timing packet. It returns if the functional packet should continue to
@@ -456,6 +462,12 @@ class Packet
  */
 bool fixPacket(PacketPtr func, PacketPtr timing);
 
+/** This function is a wrapper for the fixPacket field that toggles the hasData bit
+ * it is used when a response is waiting in the caches, but hasn't been marked as a
+ * response yet (so the fixPacket needs to get the correct value for the hasData)
+ */
+bool fixDelayedResponsePacket(PacketPtr func, PacketPtr timing);
+
 std::ostream & operator<<(std::ostream &o, const Packet &p);
 
 #endif //__MEM_PACKET_HH
diff --git a/src/mem/physical.cc b/src/mem/physical.cc
index 39eb63108..94f60ad80 100644
--- a/src/mem/physical.cc
+++ b/src/mem/physical.cc
@@ -288,6 +288,21 @@ PhysicalMemory::MemoryPort::recvAtomic(PacketPtr pkt)
 void
 PhysicalMemory::MemoryPort::recvFunctional(PacketPtr pkt)
 {
+    //Since we are overriding the function, make sure to have the impl of the
+    //check or functional accesses here.
+    std::list<std::pair<Tick,PacketPtr> >::iterator i = transmitList.begin();
+    std::list<std::pair<Tick,PacketPtr> >::iterator end = transmitList.end();
+    bool notDone = true;
+
+    while (i != end && notDone) {
+        PacketPtr target = i->second;
+        // If the target contains data, and it overlaps the
+        // probed request, need to update data
+        if (target->intersect(pkt))
+            notDone = fixPacket(pkt, target);
+        i++;
+    }
+
     // Default implementation of SimpleTimingPort::recvFunctional()
     // calls recvAtomic() and throws away the latency; we can save a
     // little here by just not calculating the latency.
diff --git a/src/mem/tport.cc b/src/mem/tport.cc
index 086d91279..c43c9aac0 100644
--- a/src/mem/tport.cc
+++ b/src/mem/tport.cc
@@ -35,14 +35,14 @@ SimpleTimingPort::recvFunctional(PacketPtr pkt)
 {
     std::list<std::pair<Tick,PacketPtr> >::iterator i = transmitList.begin();
     std::list<std::pair<Tick,PacketPtr> >::iterator end = transmitList.end();
-    bool done = false;
+    bool notDone = true;
 
-    while (i != end && !done) {
+    while (i != end && notDone) {
         PacketPtr target = i->second;
         // If the target contains data, and it overlaps the
         // probed request, need to update data
         if (target->intersect(pkt))
-            done = fixPacket(pkt, target);
+            notDone = fixPacket(pkt, target);
 
         i++;
     }
@@ -118,8 +118,14 @@ SimpleTimingPort::sendTiming(PacketPtr pkt, Tick time)
     bool done = false;
 
     while (i != end && !done) {
-        if (time+curTick < i->first)
+        if (time+curTick < i->first) {
+            if (i == transmitList.begin()) {
+                //Inserting at begining, reschedule
+                sendEvent.reschedule(time+curTick);
+            }
             transmitList.insert(i,std::pair<Tick,PacketPtr>(time+curTick,pkt));
+            done = true;
+        }
         i++;
     }
 }