Merge vm1.(none):/home/stever/bk/newmem-head

into  vm1.(none):/home/stever/bk/newmem-cache2

--HG--
extra : convert_revision : 8a501917daf81021212d136b4ebbfa059b452a13

7 files changed, 115 insertions, 28 deletions

diff --git a/configs/boot/devtime.rcS b/configs/boot/devtime.rcS
index 22a5469b8..4d1ca9407 100644
--- a/configs/boot/devtime.rcS
+++ b/configs/boot/devtime.rcS
@@ -1,7 +1,4 @@
-echo "switching cpus"
-m5 switchcpu
-echo "done"
-insmod /modules/devtime.ko dataAddr=0x9000004 count=100
+insmod /modules/devtime.ko dataAddr=0x9000008 count=100
 rmmod devtime
 insmod /modules/devtime.ko dataAddr=0x1a0000300 count=100
 rmmod devtime
diff --git a/src/dev/i8254xGBe.cc b/src/dev/i8254xGBe.cc
index e0272c655..baf13c49a 100644
--- a/src/dev/i8254xGBe.cc
+++ b/src/dev/i8254xGBe.cc
@@ -656,7 +656,7 @@ IGbE::RxDescCache::writePacket(EthPacketPtr packet)
         return false;
 
     pktPtr = packet;
-
+    pktDone = false;
     igbe->dmaWrite(igbe->platform->pciToDma(unusedCache.front()->buf),
             packet->length, &pktEvent, packet->data);
     return true;
@@ -683,8 +683,12 @@ IGbE::RxDescCache::pktComplete()
 
     uint8_t status = RXDS_DD | RXDS_EOP;
     uint8_t err = 0;
+
     IpPtr ip(pktPtr);
+
     if (ip) {
+        DPRINTF(EthernetDesc, "Proccesing Ip packet with Id=%d\n", ip->id());
+
         if (igbe->regs.rxcsum.ipofld()) {
             DPRINTF(EthernetDesc, "Checking IP checksum\n");
             status |= RXDS_IPCS;
@@ -715,7 +719,10 @@ IGbE::RxDescCache::pktComplete()
                 err |= RXDE_TCPE;
             }
         }
-    } // if ip
+    } else { // if ip
+        DPRINTF(EthernetSM, "Proccesing Non-Ip packet\n");
+    }
+
 
     desc->status = htole(status);
     desc->errors = htole(err);
@@ -912,10 +919,20 @@ IGbE::TxDescCache::pktComplete()
 
     DPRINTF(EthernetDesc, "TxDescriptor data d1: %#llx d2: %#llx\n", desc->d1, desc->d2);
 
+    if (DTRACE(EthernetDesc)) {
+        IpPtr ip(pktPtr);
+        if (ip)
+            DPRINTF(EthernetDesc, "Proccesing Ip packet with Id=%d\n",
+                    ip->id());
+        else
+            DPRINTF(EthernetSM, "Proccesing Non-Ip packet\n");
+    }
+
     // Checksums are only ofloaded for new descriptor types
     if (TxdOp::isData(desc) && ( TxdOp::ixsm(desc) || TxdOp::txsm(desc)) ) {
         DPRINTF(EthernetDesc, "Calculating checksums for packet\n");
         IpPtr ip(pktPtr);
+
         if (TxdOp::ixsm(desc)) {
             ip->sum(0);
             ip->sum(cksum(ip));
@@ -1192,6 +1209,7 @@ IGbE::rxStateMachine()
 
     // If the packet is done check for interrupts/descriptors/etc
     if (rxDescCache.packetDone()) {
+        rxDmaPacket = false;
         DPRINTF(EthernetSM, "RXS: Packet completed DMA to memory\n");
         int descLeft = rxDescCache.descLeft();
         switch (regs.rctl.rdmts()) {
@@ -1236,6 +1254,12 @@ IGbE::rxStateMachine()
         return;
     }
 
+    if (rxDmaPacket) {
+        DPRINTF(EthernetSM, "RXS: stopping ticking until packet DMA completes\n");
+        rxTick = false;
+        return;
+    }
+
     if (!rxDescCache.descUnused()) {
         DPRINTF(EthernetSM, "RXS: No descriptors available in cache, stopping ticking\n");
         rxTick = false;
@@ -1262,6 +1286,7 @@ IGbE::rxStateMachine()
     rxFifo.pop();
     DPRINTF(EthernetSM, "RXS: stopping ticking until packet DMA completes\n");
     rxTick = false;
+    rxDmaPacket = true;
 }
 
 void
diff --git a/src/dev/i8254xGBe.hh b/src/dev/i8254xGBe.hh
index 2dec3b08c..b6da53b09 100644
--- a/src/dev/i8254xGBe.hh
+++ b/src/dev/i8254xGBe.hh
@@ -80,6 +80,8 @@ class IGbE : public PciDev
     bool txTick;
     bool txFifoTick;
 
+    bool rxDmaPacket;
+
     // Event and function to deal with RDTR timer expiring
     void rdtrProcess() {
         rxDescCache.writeback(0);
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.cc b/src/mem/bridge.cc
index e89473be3..f525ccb48 100644
--- a/src/mem/bridge.cc
+++ b/src/mem/bridge.cc
@@ -119,7 +119,14 @@ Bridge::BridgePort::recvTiming(PacketPtr pkt)
     DPRINTF(BusBridge, "recvTiming: src %d dest %d addr 0x%x\n",
                 pkt->getSrc(), pkt->getDest(), pkt->getAddr());
 
-    if (pkt->isRequest() && otherPort->reqQueueFull()) {
+    DPRINTF(BusBridge, "Local queue size: %d outreq: %d outresp: %d\n",
+                    sendQueue.size(), queuedRequests, outstandingResponses);
+    DPRINTF(BusBridge, "Remove queue size: %d outreq: %d outresp: %d\n",
+                    otherPort->sendQueue.size(), otherPort->queuedRequests,
+                    otherPort->outstandingResponses);
+
+    if (pkt->isRequest() && otherPort->reqQueueFull() && pkt->result !=
+            Packet::Nacked) {
         DPRINTF(BusBridge, "Remote queue full, nacking\n");
         nackRequest(pkt);
         return true;
@@ -191,7 +198,7 @@ Bridge::BridgePort::nackRequest(PacketPtr pkt)
 void
 Bridge::BridgePort::queueForSendTiming(PacketPtr pkt)
 {
-   if (pkt->isResponse() || pkt->result == Packet::Nacked) {
+    if (pkt->isResponse() || pkt->result == Packet::Nacked) {
         // This is a response for a request we forwarded earlier.  The
         // corresponding PacketBuffer should be stored in the packet's
         // senderState field.
@@ -201,9 +208,9 @@ Bridge::BridgePort::queueForSendTiming(PacketPtr pkt)
         // from original request
         buf->fixResponse(pkt);
 
-        // Check if this packet was expecting a response (this is either it or
-        // its a nacked packet and we won't be seeing that response)
-        if (buf->expectResponse)
+        // Check if this packet was expecting a response and it's a nacked
+        // packet, in which case we will never being seeing it
+        if (buf->expectResponse && pkt->result == Packet::Nacked)
             --outstandingResponses;
 
 
@@ -213,6 +220,13 @@ Bridge::BridgePort::queueForSendTiming(PacketPtr pkt)
         delete buf;
     }
 
+
+    if (pkt->isRequest() && pkt->result != Packet::Nacked) {
+        ++queuedRequests;
+    }
+
+
+
     Tick readyTime = curTick + delay;
     PacketBuffer *buf = new PacketBuffer(pkt, readyTime);
     DPRINTF(BusBridge, "old sender state: %#X, new sender state: %#X\n",
@@ -225,7 +239,6 @@ Bridge::BridgePort::queueForSendTiming(PacketPtr pkt)
     if (sendQueue.empty()) {
         sendEvent.schedule(readyTime);
     }
-    ++queuedRequests;
     sendQueue.push_back(buf);
 }
 
@@ -254,6 +267,8 @@ Bridge::BridgePort::trySend()
     DPRINTF(BusBridge, "trySend: origSrc %d dest %d addr 0x%x\n",
             buf->origSrc, pkt->getDest(), pkt->getAddr());
 
+    bool wasReq = pkt->isRequest();
+    bool wasNacked = pkt->result == Packet::Nacked;
 
     if (sendTiming(pkt)) {
         // send successful
@@ -270,8 +285,12 @@ Bridge::BridgePort::trySend()
             delete buf;
         }
 
-        if (!buf->nacked)
+        if (!wasNacked) {
+            if (wasReq)
                 --queuedRequests;
+            else
+                --outstandingResponses;
+        }
 
         // If there are more packets to send, schedule event to try again.
         if (!sendQueue.empty()) {
@@ -305,7 +324,32 @@ Bridge::BridgePort::recvRetry()
 Tick
 Bridge::BridgePort::recvAtomic(PacketPtr pkt)
 {
-    return otherPort->sendAtomic(pkt) + delay;
+    int pbs = otherPort->peerBlockSize();
+    Tick atomic_delay;
+    // fix partial atomic writes... similar to the timing code that does the
+    // same
+    if (pkt->cmd == MemCmd::WriteInvalidateReq && fixPartialWrite &&
+             pkt->getOffset(pbs) && pkt->getSize() != pbs) {
+        PacketDataPtr data;
+        data = new uint8_t[pbs];
+        PacketPtr funcPkt = new Packet(pkt->req, MemCmd::ReadReq,
+                         Packet::Broadcast, pbs);
+
+        funcPkt->dataStatic(data);
+        otherPort->sendFunctional(funcPkt);
+        assert(funcPkt->result == Packet::Success);
+        delete funcPkt;
+        memcpy(data + pkt->getOffset(pbs), pkt->getPtr<uint8_t>(),
+                         pkt->getSize());
+        PacketPtr newPkt = new Packet(pkt->req, MemCmd::WriteInvalidateReq,
+                Packet::Broadcast, pbs);
+        pkt->dataDynamicArray(data);
+        atomic_delay = otherPort->sendAtomic(newPkt);
+        delete newPkt;
+    } else {
+        atomic_delay = otherPort->sendAtomic(pkt);
+    }
+    return atomic_delay + delay;
 }
 
 /** Function called by the port when the bus is receiving a Functional
diff --git a/src/mem/bridge.hh b/src/mem/bridge.hh
index cb5a6baed..5951eeb98 100644
--- a/src/mem/bridge.hh
+++ b/src/mem/bridge.hh
@@ -82,16 +82,15 @@ class Bridge : public MemObject
 
             bool partialWriteFixed;
             PacketPtr oldPkt;
-            bool nacked;
 
             PacketBuffer(PacketPtr _pkt, Tick t, bool nack = false)
                 : ready(t), pkt(_pkt),
                   origSenderState(_pkt->senderState), origSrc(_pkt->getSrc()),
                   expectResponse(_pkt->needsResponse() && !nack),
-                  partialWriteFixed(false), nacked(nack)
+                  partialWriteFixed(false)
 
             {
-                if (!pkt->isResponse() && !nack)
+                if (!pkt->isResponse() && !nack && pkt->result != Packet::Nacked)
                     pkt->senderState = this;
             }
 
@@ -109,18 +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 f78909b8f..9b094c1e3 100644
--- a/src/mem/cache/cache_impl.hh
+++ b/src/mem/cache/cache_impl.hh
@@ -1282,9 +1282,9 @@ 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);
 }