diff --git a/src/dev/io_device.cc b/src/dev/io_device.cc
index d430ace72..6528bd81c 100644
--- a/src/dev/io_device.cc
+++ b/src/dev/io_device.cc
@@ -218,6 +218,9 @@ DmaPort::dmaAction(Packet::Command cmd, Addr addr, int size, Event *event,
 
     DmaReqState *reqState = new DmaReqState(event, this, size);
 
+
+    DPRINTF(DMA, "Starting DMA for addr: %#x size: %d sched: %d\n", addr, size,
+            event->scheduled());
     for (ChunkGenerator gen(addr, size, peerBlockSize());
          !gen.done(); gen.next()) {
             Request *req = new Request(gen.addr(), gen.size(), 0);
@@ -231,6 +234,8 @@ DmaPort::dmaAction(Packet::Command cmd, Addr addr, int size, Event *event,
 
             assert(pendingCount >= 0);
             pendingCount++;
+            DPRINTF(DMA, "--Queuing DMA for addr: %#x size: %d\n", gen.addr(),
+                    gen.size());
             queueDma(pkt);
     }
 
@@ -281,19 +286,28 @@ DmaPort::sendDma()
 
         if (transmitList.size() && backoffTime && !inRetry &&
                 !backoffEvent.scheduled()) {
+            DPRINTF(DMA, "-- Scheduling backoff timer for %d\n",
+                    backoffTime+curTick);
             backoffEvent.schedule(backoffTime+curTick);
         }
     } else if (state == System::Atomic) {
         transmitList.pop_front();
 
         Tick lat;
+        DPRINTF(DMA, "--Sending  DMA for addr: %#x size: %d\n",
+                pkt->req->getPaddr(), pkt->req->getSize());
         lat = sendAtomic(pkt);
         assert(pkt->senderState);
         DmaReqState *state = dynamic_cast<DmaReqState*>(pkt->senderState);
         assert(state);
-
         state->numBytes += pkt->req->getSize();
+
+        DPRINTF(DMA, "--Received response for  DMA for addr: %#x size: %d nb: %d, tot: %d sched %d\n",
+                pkt->req->getPaddr(), pkt->req->getSize(), state->numBytes,
+                state->totBytes, state->completionEvent->scheduled());
+
         if (state->totBytes == state->numBytes) {
+            assert(!state->completionEvent->scheduled());
             state->completionEvent->schedule(curTick + lat);
             delete state;
             delete pkt->req;
diff --git a/src/mem/bridge.hh b/src/mem/bridge.hh
index 7df3c767f..5951eeb98 100644
--- a/src/mem/bridge.hh
+++ b/src/mem/bridge.hh
@@ -108,18 +108,24 @@ class Bridge : public MemObject
                 assert(!partialWriteFixed);
                 assert(expectResponse);
 
-                int pbs = port->peerBlockSize();
+                Addr pbs = port->peerBlockSize();
+                Addr blockAddr = pkt->getAddr() & ~(pbs-1);
                 partialWriteFixed = true;
                 PacketDataPtr data;
 
                 data = new uint8_t[pbs];
-                PacketPtr funcPkt = new Packet(pkt->req, MemCmd::ReadReq,
-                        Packet::Broadcast, pbs);
-
-                funcPkt->dataStatic(data);
-                port->sendFunctional(funcPkt);
-                assert(funcPkt->result == Packet::Success);
+                RequestPtr funcReq = new Request(blockAddr, 4, 0);
+                PacketPtr funcPkt = new Packet(funcReq, MemCmd::ReadReq,
+                            Packet::Broadcast);
+                for (int x = 0; x < pbs; x+=4) {
+                    funcReq->setPhys(blockAddr + x, 4, 0);
+                    funcPkt->reinitFromRequest();
+                    funcPkt->dataStatic(data + x);
+                    port->sendFunctional(funcPkt);
+                    assert(funcPkt->result == Packet::Success);
+                }
                 delete funcPkt;
+                delete funcReq;
 
                 oldPkt = pkt;
                 memcpy(data + oldPkt->getOffset(pbs), pkt->getPtr<uint8_t>(),
diff --git a/src/mem/cache/cache_impl.hh b/src/mem/cache/cache_impl.hh
index c70f10151..db488d33d 100644
--- a/src/mem/cache/cache_impl.hh
+++ b/src/mem/cache/cache_impl.hh
@@ -1290,9 +1290,9 @@ template<class TagStore, class Coherence>
 void
 Cache<TagStore,Coherence>::MemSidePort::recvFunctional(PacketPtr pkt)
 {
-    if (checkFunctional(pkt)) {
-        myCache()->probe(pkt, false, cache->cpuSidePort);
-    }
+    myCache()->probe(pkt, false, cache->cpuSidePort);
+    if (pkt->result != Packet::Success)
+        checkFunctional(pkt);
 }