1 files changed, 36 insertions, 13 deletions

diff --git a/src/mem/protocol/MOESI_CMP_directory-dir.sm b/src/mem/protocol/MOESI_CMP_directory-dir.sm
index 8d8ee7f8a..f1fb687bb 100644
--- a/src/mem/protocol/MOESI_CMP_directory-dir.sm
+++ b/src/mem/protocol/MOESI_CMP_directory-dir.sm
@@ -405,6 +405,14 @@ machine(Directory, "Directory protocol")
     }
   }
 
+  action(p_writeFwdDataToMemory, "p", desc="Write Response data to memory") {
+     peek(unblockNetwork_in, ResponseMsg) {
+      directory[in_msg.Address].DataBlk := in_msg.DataBlk;
+      DEBUG_EXPR(in_msg.Address);
+      DEBUG_EXPR(in_msg.DataBlk);
+    }
+ }
+
   action(ll_checkDataInMemory, "\ld", desc="Check PUTX/PUTO data is same as in the memory") {
     peek(unblockNetwork_in, ResponseMsg) {
       assert(in_msg.Dirty == false);
@@ -506,14 +514,31 @@ machine(Directory, "Directory protocol")
   }
 
   action(a_sendDMAAck, "\a", desc="Send DMA Ack that write completed, along with Inv Ack count") {
-    peek(memQueue_in, MemoryMsg) {
+    peek(requestQueue_in, RequestMsg) {
       enqueue(responseNetwork_out, ResponseMsg, latency="1") {
       out_msg.Address := address;
       out_msg.Sender := machineID;
       out_msg.SenderMachine := MachineType:Directory;
-        out_msg.Destination.add(in_msg.OriginalRequestorMachId);
+      out_msg.Destination.add(in_msg.Requestor);
+      out_msg.DataBlk := in_msg.DataBlk;
+      out_msg.Acks := directory[address].Sharers.count();  // for dma requests
+      out_msg.Type := CoherenceResponseType:DMA_ACK;
+      out_msg.MessageSize := MessageSizeType:Writeback_Control;
+      }
+    }
+  }
+
+  action(a_sendDMAAck2, "\aa", desc="Send DMA Ack that write completed, along with Inv Ack count") {
+    peek(unblockNetwork_in, ResponseMsg) {
+      enqueue(responseNetwork_out, ResponseMsg, latency="1") {
+      out_msg.Address := address;
+      out_msg.Sender := machineID;
+      out_msg.SenderMachine := MachineType:Directory;
+      if (TBEs.isPresent(address)) {
+        out_msg.Destination.add(TBEs[address].Requestor);
+      }
       out_msg.DataBlk := in_msg.DataBlk;
-      out_msg.Acks := in_msg.Acks;
+      out_msg.Acks := directory[address].Sharers.count();  // for dma requests
       out_msg.Type := CoherenceResponseType:DMA_ACK;
       out_msg.MessageSize := MessageSizeType:Writeback_Control;
       }
@@ -560,8 +585,9 @@ machine(Directory, "Directory protocol")
     i_popIncomingRequestQueue;
   }
 
-  transition(I, DMA_WRITE, XI_M) {
+  transition(I, DMA_WRITE, XI_U) {
     qw_queueMemoryWBRequest2;
+    a_sendDMAAck;  // ack count may be zero
     l_writeDMADataToMemory;
     i_popIncomingRequestQueue;
   }
@@ -571,11 +597,6 @@ machine(Directory, "Directory protocol")
     q_popMemQueue;
   }
 
-  transition(XI_M, Memory_Ack, XI_U) {
-    a_sendDMAAck;  // ack count may be zero
-    q_popMemQueue;
-  }
-
   transition(XI_U, Exclusive_Unblock, I) {
     cc_clearSharers;
     c_clearOwner;
@@ -594,8 +615,9 @@ machine(Directory, "Directory protocol")
     i_popIncomingRequestQueue;
   }
 
-  transition(S, DMA_WRITE, XI_M) {
+  transition(S, DMA_WRITE, XI_U) {
     qw_queueMemoryWBRequest2;
+    a_sendDMAAck;  // ack count may be zero
     l_writeDMADataToMemory;
     g_sendInvalidations;  // the DMA will collect invalidations
     i_popIncomingRequestQueue;
@@ -641,9 +663,10 @@ machine(Directory, "Directory protocol")
     i_popIncomingRequestQueue;
   }
 
-  transition(OI_D, Data, XI_M) {
+  transition(OI_D, Data, XI_U) {
     qw_queueMemoryWBRequest;
-    l_writeDataToMemory;
+    a_sendDMAAck2;  // ack count may be zero
+    p_writeFwdDataToMemory;
     l_writeDMADataToMemoryFromTBE;
     w_deallocateTBE;
     j_popIncomingUnblockQueue;
@@ -821,7 +844,7 @@ machine(Directory, "Directory protocol")
     q_popMemQueue;
   }
 
-  transition({I, S, O, M, IS, SS, OO, MO, MM, MI, MIS, OS, OSS}, Memory_Ack) {
+  transition({I, S, O, M, IS, SS, OO, MO, MM, MI, MIS, OS, OSS, XI_U, XI_M}, Memory_Ack) {
     //a_sendAck;
     q_popMemQueue;
   }