MOESI data corruption bug fix

4 files changed, 110 insertions, 27 deletions

diff --git a/src/mem/protocol/MOESI_CMP_directory-L1cache.sm b/src/mem/protocol/MOESI_CMP_directory-L1cache.sm
index 28800b2bd..db2efd3e7 100644
--- a/src/mem/protocol/MOESI_CMP_directory-L1cache.sm
+++ b/src/mem/protocol/MOESI_CMP_directory-L1cache.sm
@@ -90,6 +90,7 @@ machine(L1Cache, "Directory protocol")
     Own_GETX,      desc="We observe our own GetX forwarded back to us";
     Fwd_GETX,      desc="A GetX from another processor";
     Fwd_GETS,      desc="A GetS from another processor";
+    Fwd_DMA,      desc="A GetS from another processor";
     Inv,           desc="Invalidations from the directory";
 
     // Responses
@@ -309,7 +310,7 @@ machine(L1Cache, "Directory protocol")
         assert(in_msg.Destination.isElement(machineID));
         DEBUG_EXPR("MRM_DEBUG: L1 received");
         DEBUG_EXPR(in_msg.Type);
-if (in_msg.Type == CoherenceRequestType:GETX || in_msg.Type == CoherenceRequestType:DMA_READ || in_msg.Type == CoherenceRequestType:DMA_WRITE) {
+if (in_msg.Type == CoherenceRequestType:GETX || in_msg.Type == CoherenceRequestType:DMA_WRITE) {
           if (in_msg.Requestor == machineID && in_msg.RequestorMachine == MachineType:L1Cache) {
             trigger(Event:Own_GETX, in_msg.Address);
           } else {
@@ -317,6 +318,8 @@ if (in_msg.Type == CoherenceRequestType:GETX || in_msg.Type == CoherenceRequestT
           }
         } else if (in_msg.Type == CoherenceRequestType:GETS) {
           trigger(Event:Fwd_GETS, in_msg.Address);
+        } else if (in_msg.Type == CoherenceRequestType:DMA_READ) {
+          trigger(Event:Fwd_DMA, in_msg.Address);
         } else if (in_msg.Type == CoherenceRequestType:WB_ACK) {
           trigger(Event:Writeback_Ack, in_msg.Address);
         } else if (in_msg.Type == CoherenceRequestType:WB_ACK_DATA) {
@@ -826,7 +829,7 @@ if (in_msg.Type == CoherenceRequestType:GETX || in_msg.Type == CoherenceRequestT
     zz_recycleMandatoryQueue;
   }
 
-  transition({M_W, MM_W}, {Fwd_GETS, Fwd_GETX, Own_GETX, Inv}) {
+  transition({M_W, MM_W}, {Fwd_GETS, Fwd_DMA, Fwd_GETX, Own_GETX, Inv}) {
     z_recycleRequestQueue;
   }
 
@@ -892,7 +895,7 @@ if (in_msg.Type == CoherenceRequestType:GETX || in_msg.Type == CoherenceRequestT
     l_popForwardQueue;
   }
 
-  transition(S, Fwd_GETS) {
+  transition(S, {Fwd_GETS, Fwd_DMA}) {
     e_sendData;
     l_popForwardQueue;
   }
@@ -921,7 +924,7 @@ if (in_msg.Type == CoherenceRequestType:GETX || in_msg.Type == CoherenceRequestT
     l_popForwardQueue;
   }
 
-  transition(O, Fwd_GETS) {
+  transition(O, {Fwd_GETS, Fwd_DMA}) {
     e_sendData;
     l_popForwardQueue;
   }
@@ -953,6 +956,12 @@ if (in_msg.Type == CoherenceRequestType:GETX || in_msg.Type == CoherenceRequestT
     l_popForwardQueue;
   }
 
+  transition(MM, Fwd_DMA, MM) {
+    //ee_sendDataExclusive;
+    e_sendData;
+    l_popForwardQueue;
+  }
+
   // Transitions from M
   transition({M, M_W}, {Load, Ifetch}) {
     h_load_hit;
@@ -986,6 +995,11 @@ if (in_msg.Type == CoherenceRequestType:GETX || in_msg.Type == CoherenceRequestT
     l_popForwardQueue;
   }
 
+  transition(M, Fwd_DMA, M) {
+    e_sendData;
+    l_popForwardQueue;
+  }
+
   // Transitions from IM
 
   transition(IM, Inv) {
@@ -1025,7 +1039,7 @@ if (in_msg.Type == CoherenceRequestType:GETX || in_msg.Type == CoherenceRequestT
     n_popResponseQueue;
   }
 
-  transition(SM, Fwd_GETS) {
+  transition(SM, {Fwd_DMA, Fwd_GETS}) {
     e_sendData;
     l_popForwardQueue;
   }
@@ -1044,7 +1058,7 @@ if (in_msg.Type == CoherenceRequestType:GETX || in_msg.Type == CoherenceRequestT
     l_popForwardQueue;
   }
 
-  transition(OM, Fwd_GETS, OM) {
+  transition(OM, {Fwd_DMA, Fwd_GETS}, OM) {
     e_sendData;
     l_popForwardQueue;
   }
@@ -1105,12 +1119,17 @@ if (in_msg.Type == CoherenceRequestType:GETX || in_msg.Type == CoherenceRequestT
     l_popForwardQueue;
   }
 
+  transition(MI, Fwd_DMA, MI) {
+    q_sendDataFromTBEToCache;
+    l_popForwardQueue;
+  }
+
   transition(MI, Fwd_GETX, II) {
     q_sendExclusiveDataFromTBEToCache;
     l_popForwardQueue;
   }
 
-  transition({SI, OI}, Fwd_GETS) {
+  transition({SI, OI}, {Fwd_DMA, Fwd_GETS}) {
     q_sendDataFromTBEToCache;
     l_popForwardQueue;
   }
diff --git a/src/mem/protocol/MOESI_CMP_directory-L2cache.sm b/src/mem/protocol/MOESI_CMP_directory-L2cache.sm
index 68d3a2cd3..9ee909199 100644
--- a/src/mem/protocol/MOESI_CMP_directory-L2cache.sm
+++ b/src/mem/protocol/MOESI_CMP_directory-L2cache.sm
@@ -138,6 +138,7 @@ machine(L2Cache, "Token protocol")
     L1_PUTS,             desc="local sharer wants to writeback";
     Fwd_GETX,      desc="A GetX from another processor";
     Fwd_GETS,      desc="A GetS from another processor";
+    Fwd_DMA,      desc="A request from DMA";
     Own_GETX,      desc="A GetX from this node";
     Inv,           desc="Invalidations from the directory";
 
@@ -584,7 +585,7 @@ machine(L2Cache, "Token protocol")
   in_port(requestNetwork_in, RequestMsg, GlobalRequestToL2Cache) {
     if (requestNetwork_in.isReady()) {
       peek(requestNetwork_in, RequestMsg) {
-        if (in_msg.Type == CoherenceRequestType:GETX || in_msg.Type == CoherenceRequestType:DMA_READ || in_msg.Type == CoherenceRequestType:DMA_WRITE) {
+        if (in_msg.Type == CoherenceRequestType:GETX || in_msg.Type == CoherenceRequestType:DMA_WRITE) {
           if (in_msg.Requestor == machineID) {
             trigger(Event:Own_GETX, in_msg.Address);
           } else {
@@ -592,6 +593,8 @@ machine(L2Cache, "Token protocol")
           }
         } else if (in_msg.Type == CoherenceRequestType:GETS) {
           trigger(Event:Fwd_GETS, in_msg.Address);
+        } else if(in_msg.Type == CoherenceRequestType:DMA_READ) {
+          trigger(Event:Fwd_DMA, in_msg.Address);
         } else if (in_msg.Type == CoherenceRequestType:INV) {
           trigger(Event:Inv, in_msg.Address);
         } else if (in_msg.Type == CoherenceRequestType:WB_ACK) {
@@ -1456,7 +1459,15 @@ machine(L2Cache, "Token protocol")
     zz_recycleResponseQueue;
   }
 
-  transition({IFGX, IFGS, ISFGS, IFGXX, IFLXO, OFGX, ILOW, ILOXW, ILOSW, ILOSXW, SLSW, OLSW, ILSW, IW, OW, SW, OXW, OLSXW, ILXW, IFLS, IFLO, IFLOX, IFLOXX, IFLOSX,OLSXS, MM, SS, OO, SLSS, OLSS, OLSF, IGMIOFS}, {Fwd_GETX, Fwd_GETS, Inv}) {
+  transition({IFGX, IFGS, ISFGS, IFGXX, IFLXO, OFGX, ILOW, ILOXW, ILOSW, ILOSXW, SLSW, OLSW, ILSW, IW, OW, SW, OXW, OLSXW, ILXW, IFLS, IFLO, IFLOX, IFLOXX, IFLOSX,OLSXS, IGS, IGM, MM, SS, OO, SLSS, OLSS, OLSF, IGMIOFS}, {Fwd_GETX, Fwd_GETS, Fwd_DMA}) {
+    zz_recycleRequestQueue;
+  }
+
+  transition({IFGX, IFGS, ISFGS, IFGXX, IFLXO, OFGX, ILOW, ILOXW, ILOSW, ILOSXW, SLSW, OLSW, ILSW, IW, OW, SW, OXW, OLSXW, ILXW, IFLS, IFLO, IFLOX, IFLOXX, IFLOSX,OLSXS, MM, SS, OO, SLSS, OLSS, OLSF, IGMIOFS}, {Inv}) {
+    zz_recycleRequestQueue;
+  }
+
+  transition({IGM, IGS}, {Own_GETX}) {
     zz_recycleRequestQueue;
   }
 
@@ -1521,6 +1532,27 @@ machine(L2Cache, "Token protocol")
     m_popRequestQueue;
   }
 
+  transition({ILOS, ILOSX}, Fwd_DMA) {
+    i_allocateTBE;
+    t_recordFwdSID;
+    j_forwardGlobalRequestToLocalOwner;
+    m_popRequestQueue;
+  }
+
+  transition({ILO, ILX, ILOX}, Fwd_DMA) {
+    i_allocateTBE;
+    t_recordFwdSID;
+    j_forwardGlobalRequestToLocalOwner;
+    m_popRequestQueue;
+  }
+ 
+  transition({ILOS, ILOSX, ILO, ILX, ILOX, ILXW}, Data) {
+    i_copyDataToTBE;
+    c_sendDataFromTBEToFwdGETS;
+    s_deallocateTBE;
+    n_popResponseQueue;
+  }
+
   transition(IFGS, Data, ILO) {
     i_copyDataToTBE;
     c_sendDataFromTBEToFwdGETS;
@@ -1598,7 +1630,7 @@ machine(L2Cache, "Token protocol")
     m_popRequestQueue;
   }
 
-  transition({O, OLS}, Fwd_GETS) {
+  transition({O, OLS}, {Fwd_GETS, Fwd_DMA}) {
     dd_sendDataToFwdGETS;
     m_popRequestQueue;
   }
@@ -1609,6 +1641,10 @@ machine(L2Cache, "Token protocol")
     m_popRequestQueue;
   }
 
+  transition(OLSX, Fwd_DMA) {
+    dd_sendDataToFwdGETS;
+    m_popRequestQueue;
+  }
 
   transition(M, Fwd_GETX, I) {
     dd_sendDataToFwdGETX;
@@ -1629,6 +1665,10 @@ machine(L2Cache, "Token protocol")
      m_popRequestQueue;
   }
 
+  transition(M, Fwd_DMA) {
+     dd_sendExclusiveDataToFwdGETS;
+     m_popRequestQueue;
+  }
 
   transition({OLS, OLSX}, Fwd_GETX, OLSF) {
     i_allocateTBE;
@@ -1837,8 +1877,6 @@ machine(L2Cache, "Token protocol")
     n_popResponseQueue;
   }
 
-
-
   // LOCAL REQUESTS THAT MUST ISSUE
 
   transition(NP, {L1_PUTS, L1_PUTX, L1_PUTO}) {
@@ -1949,7 +1987,7 @@ machine(L2Cache, "Token protocol")
     o_popL1RequestQueue;
   }
 
-  transition(OGMIO, Fwd_GETS) {
+  transition(OGMIO, {Fwd_GETS, Fwd_DMA}) {
     t_recordFwdSID;
     c_sendDataFromTBEToFwdGETS;
     m_popRequestQueue;
@@ -1984,6 +2022,12 @@ machine(L2Cache, "Token protocol")
     m_popRequestQueue;
   }
 
+  transition(IGMIO, Fwd_DMA) {
+    t_recordFwdSID;
+    j_forwardGlobalRequestToLocalOwner;
+    m_popRequestQueue;
+  }
+
   transition(IGMIOFS, Data, IGMIO) {
     i_copyDataToTBE;
     c_sendDataFromTBEToFwdGETS;
@@ -2163,7 +2207,7 @@ machine(L2Cache, "Token protocol")
 
   }
 
-  transition(IGMO, Fwd_GETS) {
+  transition(IGMO, {Fwd_GETS, Fwd_DMA}) {
     t_recordFwdSID;
     c_sendDataFromTBEToFwdGETS;
     m_popRequestQueue;
@@ -2518,13 +2562,13 @@ machine(L2Cache, "Token protocol")
     n_popTriggerQueue;
   }
 
-  transition(OLSI, Fwd_GETS) {
+  transition(OLSI, {Fwd_GETS, Fwd_DMA}) {
     t_recordFwdSID;
     c_sendDataFromTBEToFwdGETS;
     m_popRequestQueue;
   }
 
-  transition({MI, OI}, Fwd_GETS, OI) {
+  transition({MI, OI}, {Fwd_GETS, Fwd_DMA}, OI) {
     t_recordFwdSID;
     c_sendDataFromTBEToFwdGETS;
     m_popRequestQueue;
diff --git a/src/mem/protocol/MOESI_CMP_directory-dir.sm b/src/mem/protocol/MOESI_CMP_directory-dir.sm
index f1fb687bb..8e48fc9ab 100644
--- a/src/mem/protocol/MOESI_CMP_directory-dir.sm
+++ b/src/mem/protocol/MOESI_CMP_directory-dir.sm
@@ -331,6 +331,23 @@ machine(Directory, "Directory protocol")
     }
   }
 
+  action(p_fwdDataToDMA, "\d", desc="Send data to requestor") {
+    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.Requestor);
+        out_msg.DataBlk := directory[in_msg.Address].DataBlk;
+        out_msg.Dirty := false; // By definition, the block is now clean
+        out_msg.Type := CoherenceResponseType:DATA_EXCLUSIVE;
+        out_msg.MessageSize := MessageSizeType:Response_Data;
+      }
+    }
+  }
+
+
+
   action(e_ownerIsUnblocker, "e", desc="The owner is now the unblocker") {
     peek(unblockNetwork_in, ResponseMsg) {
       directory[address].Owner.clear();
@@ -592,7 +609,7 @@ machine(Directory, "Directory protocol")
     i_popIncomingRequestQueue;
   }
 
-  transition(XI_M, Memory_Data, XI_U) {
+  transition(XI_M, Memory_Data, I) {
     d_sendDataMsg;  // ack count may be zero
     q_popMemQueue;
   }
@@ -609,9 +626,10 @@ machine(Directory, "Directory protocol")
     i_popIncomingRequestQueue;
   }
 
-  transition(S, DMA_READ, XI_M) {
-    qf_queueMemoryFetchRequest;
-    g_sendInvalidations;  // the DMA will collect the invalidations then send an Unblock Exclusive
+  transition(S, DMA_READ, S) {
+    //qf_queueMemoryFetchRequest;
+    p_fwdDataToDMA;
+    //g_sendInvalidations;  // the DMA will collect the invalidations then send an Unblock Exclusive
     i_popIncomingRequestQueue;
   }
 
@@ -650,9 +668,9 @@ machine(Directory, "Directory protocol")
     i_popIncomingRequestQueue;
   }
 
-  transition(O, DMA_READ, XI_U) {
+  transition(O, DMA_READ, O) {
     f_forwardRequest;     // this will cause the data to go to DMA directly
-    g_sendInvalidations;  // this will cause acks to be sent to the DMA
+    //g_sendInvalidations;  // this will cause acks to be sent to the DMA
     i_popIncomingRequestQueue;
   }
 
@@ -684,7 +702,7 @@ machine(Directory, "Directory protocol")
   }
 
   // no exclusive unblock will show up to the directory
-  transition(M, DMA_READ, XI_U) {
+  transition(M, DMA_READ, M) {
     f_forwardRequest;     // this will cause the data to go to DMA directly
     i_popIncomingRequestQueue;
   }
diff --git a/src/mem/protocol/MOESI_CMP_directory-dma.sm b/src/mem/protocol/MOESI_CMP_directory-dma.sm
index ae86e24da..6105778bd 100644
--- a/src/mem/protocol/MOESI_CMP_directory-dma.sm
+++ b/src/mem/protocol/MOESI_CMP_directory-dma.sm
@@ -227,16 +227,18 @@ machine(DMA, "DMA Controller")
     p_popResponseQueue;
   }
 
-  transition(BUSY_RD, Data) {
+  transition(BUSY_RD, Data, READY) {
     t_updateTBEData;
-    u_updateAckCount;
-    o_checkForCompletion;
+    d_dataCallbackFromTBE;
+    w_deallocateTBE;
+    //u_updateAckCount;
+    //o_checkForCompletion;
     p_popResponseQueue;
   }
 
   transition(BUSY_RD, All_Acks, READY) {
     d_dataCallbackFromTBE;
-    u_sendExclusiveUnblockToDir;
+    //u_sendExclusiveUnblockToDir;
     w_deallocateTBE;
     p_popTriggerQueue;
   }