ruby: MOESI hammer support for DMA reads and writes
authorBrad Beckmann <Brad.Beckmann@amd.com>
Thu, 19 Nov 2009 00:34:32 +0000 (16:34 -0800)
committerBrad Beckmann <Brad.Beckmann@amd.com>
Thu, 19 Nov 2009 00:34:32 +0000 (16:34 -0800)
src/mem/protocol/MOESI_hammer-dir.sm
src/mem/protocol/MOESI_hammer-dma.sm [new file with mode: 0644]
src/mem/protocol/MOESI_hammer-msg.sm
src/mem/protocol/MOESI_hammer.slicc
src/mem/ruby/config/defaults.rb

index 49efaffb684db508966ac0e9e0fd39858201e48a..b9b001e401d8a97251aaa2b07b26d3f8eb7d3036 100644 (file)
@@ -39,11 +39,17 @@ machine(Directory, "AMD Hammer-like protocol")
 
   MessageBuffer forwardFromDir, network="To", virtual_network="2", ordered="false";
   MessageBuffer responseFromDir, network="To", virtual_network="1", ordered="false";
-  //MessageBuffer dmaRequestFromDir, network="To", virtual_network="4", ordered="true";
+  //
+  // For a finite buffered network, note that the DMA response network only 
+  // works at this relatively higher numbered (lower priority) virtual network
+  // because the trigger queue decouples cache responses from DMA responses.
+  //
+  MessageBuffer dmaResponseFromDir, network="To", virtual_network="4", ordered="true";
 
-  MessageBuffer requestToDir, network="From", virtual_network="3", ordered="false";
   MessageBuffer unblockToDir, network="From", virtual_network="0", ordered="false";
-  //MessageBuffer dmaRequestToDir, network="From", virtual_network="5", ordered="true";
+  MessageBuffer responseToDir, network="From", virtual_network="1", ordered="false";
+  MessageBuffer requestToDir, network="From", virtual_network="3", ordered="false";
+  MessageBuffer dmaRequestToDir, network="From", virtual_network="5", ordered="true";
 
   // STATES
   enumeration(State, desc="Directory states", default="Directory_State_E") {
@@ -57,6 +63,13 @@ machine(Directory, "AMD Hammer-like protocol")
     O_B_W,          desc="Owner, Blocked, waiting for Dram";
     NO_W,           desc="Not Owner, waiting for Dram";
     O_W,            desc="Owner, waiting for Dram";
+    NO_DW_B_W,      desc="Not Owner, Dma Write waiting for Dram and cache responses";
+    NO_DR_B_W,      desc="Not Owner, Dma Read waiting for Dram and cache responses";
+    NO_DR_B_D,      desc="Not Owner, Dma Read waiting for cache responses including dirty data";
+    NO_DR_B,        desc="Not Owner, Dma Read waiting for cache responses";
+    NO_DW_W,        desc="Not Owner, Dma Write waiting for Dram";
+    O_DR_B_W,       desc="Owner, Dma Read waiting for Dram and cache responses";
+    O_DR_B,         desc="Owner, Dma Read waiting for cache responses";
     WB,             desc="Blocked on a writeback";
     WB_O_W,         desc="Blocked on memory write, will go to O";
     WB_E_W,         desc="Blocked on memory write, will go to E";
@@ -73,9 +86,23 @@ machine(Directory, "AMD Hammer-like protocol")
     Writeback_Exclusive_Clean, desc="The final part of a PutX (no data, exclusive)";
     Writeback_Exclusive_Dirty, desc="The final part of a PutX (data, exclusive)";
 
+    // DMA requests
+    DMA_READ, desc="A DMA Read memory request";
+    DMA_WRITE, desc="A DMA Write memory request";
+
     // Memory Controller
     Memory_Data, desc="Fetched data from memory arrives";
     Memory_Ack, desc="Writeback Ack from memory arrives";
+
+    // Cache responses required to handle DMA
+    Ack,             desc="Received an ack message";
+    Shared_Ack,      desc="Received an ack message, responder has a shared copy";
+    Shared_Data,     desc="Received a data message, responder has a shared copy";
+    Exclusive_Data,  desc="Received a data message, responder had an exclusive copy, they gave it to us";
+
+    // Triggers
+    All_acks_and_data,            desc="Received all required data and message acks";
+    All_acks_and_data_no_sharers, desc="Received all acks and no other processor has a shared copy";
   }
 
   // TYPES
@@ -100,9 +127,13 @@ machine(Directory, "AMD Hammer-like protocol")
     Address PhysicalAddress, desc="physical address";
     State TBEState,        desc="Transient State";
     CoherenceResponseType ResponseType, desc="The type for the subsequent response message";
-    DataBlock DataBlk,     desc="Data to be written (DMA write only)";
+    DataBlock DmaDataBlk,  desc="DMA Data to be written.  Partial blocks need to merged with system memory";
+    DataBlock DataBlk,     desc="The current view of system memory";
     int Len,               desc="...";
     MachineID DmaRequestor, desc="DMA requestor";
+    int NumPendingMsgs,    desc="Number of pending acks/messages";
+    bool CacheDirty,       desc="Indicates whether a cache has responded with dirty data";
+    bool Sharers,          desc="Indicates whether a cache has indicated it is currently a sharer";
   }
 
   external_type(TBETable) {
@@ -135,10 +166,14 @@ machine(Directory, "AMD Hammer-like protocol")
     directory[addr].DirectoryState := state;
   }
   
+  MessageBuffer triggerQueue, ordered="true";
+
   // ** OUT_PORTS **
+  out_port(requestQueue_out, ResponseMsg, requestToDir); // For recycling requests
   out_port(forwardNetwork_out, RequestMsg, forwardFromDir);
   out_port(responseNetwork_out, ResponseMsg, responseFromDir);
-  out_port(requestQueue_out, ResponseMsg, requestToDir); // For recycling requests
+  out_port(dmaResponseNetwork_out, DMAResponseMsg, dmaResponseFromDir);
+  out_port(triggerQueue_out, TriggerMsg, triggerQueue);
   
   //
   // Memory buffer for memory controller to DIMM communication
@@ -147,6 +182,21 @@ machine(Directory, "AMD Hammer-like protocol")
 
   // ** IN_PORTS **
   
+  // Trigger Queue
+  in_port(triggerQueue_in, TriggerMsg, triggerQueue) {
+    if (triggerQueue_in.isReady()) {
+      peek(triggerQueue_in, TriggerMsg) {
+        if (in_msg.Type == TriggerType:ALL_ACKS) {
+          trigger(Event:All_acks_and_data, in_msg.Address);
+        } else if (in_msg.Type == TriggerType:ALL_ACKS_NO_SHARERS) {
+          trigger(Event:All_acks_and_data_no_sharers, in_msg.Address);
+        } else {
+          error("Unexpected message");
+        }
+      }
+    }
+  }
+
   in_port(unblockNetwork_in, ResponseMsg, unblockToDir) {
     if (unblockNetwork_in.isReady()) {
       peek(unblockNetwork_in, ResponseMsg) {
@@ -167,6 +217,39 @@ machine(Directory, "AMD Hammer-like protocol")
     }
   }
 
+  // Response Network
+  in_port(responseToDir_in, ResponseMsg, responseToDir) {
+    if (responseToDir_in.isReady()) {
+      peek(responseToDir_in, ResponseMsg) {
+        if (in_msg.Type == CoherenceResponseType:ACK) {
+          trigger(Event:Ack, in_msg.Address);
+        } else if (in_msg.Type == CoherenceResponseType:ACK_SHARED) {
+          trigger(Event:Shared_Ack, in_msg.Address);
+        } else if (in_msg.Type == CoherenceResponseType:DATA_SHARED) {
+          trigger(Event:Shared_Data, in_msg.Address);
+        } else if (in_msg.Type == CoherenceResponseType:DATA_EXCLUSIVE) {
+          trigger(Event:Exclusive_Data, in_msg.Address);
+        } else {
+          error("Unexpected message");
+        }
+      }
+    }
+  }
+
+  in_port(dmaRequestQueue_in, DMARequestMsg, dmaRequestToDir) {
+    if (dmaRequestQueue_in.isReady()) {
+      peek(dmaRequestQueue_in, DMARequestMsg) {
+        if (in_msg.Type == DMARequestType:READ) {
+          trigger(Event:DMA_READ, in_msg.LineAddress);
+        } else if (in_msg.Type == DMARequestType:WRITE) {
+          trigger(Event:DMA_WRITE, in_msg.LineAddress);
+        } else {
+          error("Invalid message");
+        }
+      }
+    }
+  }
+
   in_port(requestQueue_in, RequestMsg, requestToDir) {
     if (requestQueue_in.isReady()) {
       peek(requestQueue_in, RequestMsg) {
@@ -233,10 +316,61 @@ machine(Directory, "AMD Hammer-like protocol")
     }
   }
 
+  action(vd_allocateDmaRequestInTBE, "vd", desc="Record Data in TBE") {
+    peek(dmaRequestQueue_in, DMARequestMsg) {
+      TBEs.allocate(address);
+      TBEs[address].DmaDataBlk := in_msg.DataBlk;
+      TBEs[address].PhysicalAddress := in_msg.PhysicalAddress;
+      TBEs[address].Len := in_msg.Len;
+      TBEs[address].DmaRequestor := in_msg.Requestor;
+      TBEs[address].ResponseType := CoherenceResponseType:DATA_EXCLUSIVE;
+      //
+      // One ack for each last-level cache
+      //
+      TBEs[address].NumPendingMsgs := getNumberOfLastLevelCaches();
+      //
+      // Assume initially that the caches store a clean copy and that memory
+      // will provide the data
+      //
+      TBEs[address].CacheDirty := false;
+    }
+  }
+
   action(w_deallocateTBE, "w", desc="Deallocate TBE") {
     TBEs.deallocate(address);
   }
 
+  action(m_decrementNumberOfMessages, "m", desc="Decrement the number of messages for which we're waiting") {
+    peek(responseToDir_in, ResponseMsg) {
+      assert(in_msg.Acks > 0);
+      DEBUG_EXPR(TBEs[address].NumPendingMsgs);
+      //
+      // Note that cache data responses will have an ack count of 2.  However, 
+      // directory DMA requests must wait for acks from all LLC caches, so 
+      // only decrement by 1.
+      //
+      TBEs[address].NumPendingMsgs := TBEs[address].NumPendingMsgs - 1;
+      DEBUG_EXPR(TBEs[address].NumPendingMsgs);
+    }
+  }
+
+  action(n_popResponseQueue, "n", desc="Pop response queue") {
+    responseToDir_in.dequeue();
+  }
+
+  action(o_checkForCompletion, "o", desc="Check if we have received all the messages required for completion") {
+    if (TBEs[address].NumPendingMsgs == 0) {
+      enqueue(triggerQueue_out, TriggerMsg) {
+        out_msg.Address := address;
+        if (TBEs[address].Sharers) {
+          out_msg.Type := TriggerType:ALL_ACKS;
+        } else {
+          out_msg.Type := TriggerType:ALL_ACKS_NO_SHARERS;
+        }
+      }
+    }
+  }
+
   action(d_sendData, "d", desc="Send data to requestor") {
     peek(memQueue_in, MemoryMsg) {
       enqueue(responseNetwork_out, ResponseMsg, latency="1") {
@@ -252,18 +386,66 @@ machine(Directory, "AMD Hammer-like protocol")
     }
   }
 
+  action(dr_sendDmaData, "dr", desc="Send Data to DMA controller from memory") {
+    peek(memQueue_in, MemoryMsg) {
+      enqueue(dmaResponseNetwork_out, DMAResponseMsg, latency="1") {
+        out_msg.PhysicalAddress := address;
+        out_msg.LineAddress := address;
+        out_msg.Type := DMAResponseType:DATA;
+        //
+        // we send the entire data block and rely on the dma controller to 
+        // split it up if need be
+        //
+        out_msg.DataBlk := in_msg.DataBlk;
+        out_msg.Destination.add(TBEs[address].DmaRequestor);
+        out_msg.MessageSize := MessageSizeType:Response_Data;
+      }
+    }
+  }
+
+  action(dt_sendDmaDataFromTbe, "dt", desc="Send Data to DMA controller from tbe") {
+    peek(triggerQueue_in, TriggerMsg) {
+      enqueue(dmaResponseNetwork_out, DMAResponseMsg, latency="1") {
+        out_msg.PhysicalAddress := address;
+        out_msg.LineAddress := address;
+        out_msg.Type := DMAResponseType:DATA;
+        //
+        // we send the entire data block and rely on the dma controller to 
+        // split it up if need be
+        //
+        out_msg.DataBlk := TBEs[address].DataBlk;
+        out_msg.Destination.add(TBEs[address].DmaRequestor);
+        out_msg.MessageSize := MessageSizeType:Response_Data;
+      }
+    }
+  }
+
+  action(da_sendDmaAck, "da", desc="Send Ack to DMA controller") {
+    enqueue(dmaResponseNetwork_out, DMAResponseMsg, latency="1") {
+      out_msg.PhysicalAddress := address;
+      out_msg.LineAddress := address;
+      out_msg.Type := DMAResponseType:ACK;
+      out_msg.Destination.add(TBEs[address].DmaRequestor); 
+      out_msg.MessageSize := MessageSizeType:Writeback_Control;
+    }
+  }
+
   action(rx_recordExclusiveInTBE, "rx", desc="Record Exclusive in TBE") {
     peek(requestQueue_in, RequestMsg) {
       TBEs[address].ResponseType := CoherenceResponseType:DATA_EXCLUSIVE;
     }
   }
 
-  action(r_recordDataInTBE, "r", desc="Record Data in TBE") {
+  action(r_recordDataInTBE, "rt", desc="Record Data in TBE") {
     peek(requestQueue_in, RequestMsg) {
       TBEs[address].ResponseType := CoherenceResponseType:DATA;
     }
   }
 
+  action(r_setSharerBit, "r", desc="We saw other sharers") {
+    TBEs[address].Sharers := true;
+  }
+
   action(qf_queueMemoryFetchRequest, "qf", desc="Queue off-chip fetch request") {
     peek(requestQueue_in, RequestMsg) {
       enqueue(memQueue_out, MemoryMsg, latency="1") {
@@ -272,56 +454,25 @@ machine(Directory, "AMD Hammer-like protocol")
         out_msg.Sender := machineID;
         out_msg.OriginalRequestorMachId := in_msg.Requestor;
         out_msg.MessageSize := in_msg.MessageSize;
-        out_msg.DataBlk := directory[in_msg.Address].DataBlk;
+        out_msg.DataBlk := directory[address].DataBlk;
         DEBUG_EXPR(out_msg);
       }
     }
   }
 
-//  action(qx_queueMemoryFetchExclusiveRequest, "xf", desc="Queue off-chip fetch request") {
-//    peek(requestQueue_in, RequestMsg) {
-//      enqueue(memQueue_out, MemoryMsg, latency=memory_request_latency) {
-//        out_msg.Address := address;
-//        out_msg.Type := MemoryRequestType:MEMORY_READ;
-//        out_msg.ResponseType := CoherenceResponseType:DATA_EXCLUSIVE;
-//        out_msg.Sender := machineID;
-//        out_msg.OriginalRequestorMachId := in_msg.Requestor;
-//        out_msg.MessageSize := in_msg.MessageSize;
-//        out_msg.DataBlk := directory[in_msg.Address].DataBlk;
-//        DEBUG_EXPR(out_msg);
-//      }
-//    }
-//  }
-
-//  action(d_sendData, "d", desc="Send data to requestor") {
-//    peek(requestQueue_in, RequestMsg) {
-//     enqueue(responseNetwork_out, ResponseMsg, latency=memory_latency) {
-//        out_msg.Address := address;
-//        out_msg.Type := CoherenceResponseType:DATA;
-//        out_msg.Sender := machineID;
-//        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.Acks := 1;
-//        out_msg.MessageSize := MessageSizeType:Response_Data;
-//      }
-//    }
-//  }
-
-//  action(dd_sendExclusiveData, "\d", desc="Send exclusive data to requestor") {
-//    peek(requestQueue_in, RequestMsg) {
-//      enqueue(responseNetwork_out, ResponseMsg, latency=memory_latency) {
-//        out_msg.Address := address;
-//        out_msg.Type := CoherenceResponseType:DATA_EXCLUSIVE;
-//        out_msg.Sender := machineID;
-//        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.Acks := 1;
-//        out_msg.MessageSize := MessageSizeType:Response_Data;
-//      }
-//    }
-//  }
+  action(qd_queueMemoryRequestFromDmaRead, "qd", desc="Queue off-chip fetch request") {
+    peek(dmaRequestQueue_in, DMARequestMsg) {
+      enqueue(memQueue_out, MemoryMsg, latency="1") {
+        out_msg.Address := address;
+        out_msg.Type := MemoryRequestType:MEMORY_READ;
+        out_msg.Sender := machineID;
+        out_msg.OriginalRequestorMachId := in_msg.Requestor;
+        out_msg.MessageSize := in_msg.MessageSize;
+        out_msg.DataBlk := directory[address].DataBlk;
+        DEBUG_EXPR(out_msg);
+      }
+    }
+  }
 
   action(f_forwardRequest, "f", desc="Forward requests") {
     if (getNumberOfLastLevelCaches() > 1) {
@@ -338,6 +489,38 @@ machine(Directory, "AMD Hammer-like protocol")
     }
   }
 
+  action(f_forwardWriteFromDma, "fw", desc="Forward requests") {
+    peek(dmaRequestQueue_in, DMARequestMsg) {
+      enqueue(forwardNetwork_out, RequestMsg, latency=memory_controller_latency) {
+        out_msg.Address := address;
+        out_msg.Type := CoherenceRequestType:GETX;
+        //
+        // Send to all L1 caches, since the requestor is the memory controller
+        // itself
+        //
+        out_msg.Requestor := machineID;
+        out_msg.Destination.broadcast(MachineType:L1Cache); 
+        out_msg.MessageSize := MessageSizeType:Forwarded_Control;
+      }
+    }
+  }
+
+  action(f_forwardReadFromDma, "fr", desc="Forward requests") {
+    peek(dmaRequestQueue_in, DMARequestMsg) {
+      enqueue(forwardNetwork_out, RequestMsg, latency=memory_controller_latency) {
+        out_msg.Address := address;
+        out_msg.Type := CoherenceRequestType:GETS;
+        //
+        // Send to all L1 caches, since the requestor is the memory controller
+        // itself
+        //
+        out_msg.Requestor := machineID;
+        out_msg.Destination.broadcast(MachineType:L1Cache); 
+        out_msg.MessageSize := MessageSizeType:Forwarded_Control;
+      }
+    }
+  }
+
   action(i_popIncomingRequestQueue, "i", desc="Pop incoming request queue") {
     requestQueue_in.dequeue();
   }
@@ -350,16 +533,52 @@ machine(Directory, "AMD Hammer-like protocol")
     memQueue_in.dequeue();
   }
 
+  action(g_popTriggerQueue, "g", desc="Pop trigger queue") {
+    triggerQueue_in.dequeue();
+  }
+
+  action(p_popDmaRequestQueue, "pd", desc="pop dma request queue") {
+    dmaRequestQueue_in.dequeue();
+  }
+
+  action(y_recycleDmaRequestQueue, "y", desc="recycle dma request queue") {
+    dmaRequestQueue_in.recycle();
+  }
+
+  action(r_recordMemoryData, "rd", desc="record data from memory to TBE") {
+    peek(memQueue_in, MemoryMsg) {
+      if (TBEs[address].CacheDirty == false) {
+        TBEs[address].DataBlk := in_msg.DataBlk;
+      }
+    }
+  }
+
+  action(r_recordCacheData, "rc", desc="record data from cache response to TBE") {
+    peek(responseToDir_in, ResponseMsg) {
+      TBEs[address].CacheDirty := true;
+      TBEs[address].DataBlk := in_msg.DataBlk;
+    }
+  }
+
   action(l_writeDataToMemory, "l", desc="Write PUTX/PUTO data to memory") {
     peek(unblockNetwork_in, ResponseMsg) {
       assert(in_msg.Dirty);
       assert(in_msg.MessageSize == MessageSizeType:Writeback_Data);
-      directory[in_msg.Address].DataBlk := in_msg.DataBlk;
+      directory[address].DataBlk := in_msg.DataBlk;
       DEBUG_EXPR(in_msg.Address);
       DEBUG_EXPR(in_msg.DataBlk);
     }
   }
 
+  action(dwt_writeDmaDataFromTBE, "dwt", desc="DMA Write data to memory from TBE") {
+    directory[address].DataBlk := TBEs[address].DataBlk;
+    directory[address].DataBlk.copyPartial(TBEs[address].DmaDataBlk, addressOffset(TBEs[address].PhysicalAddress), TBEs[address].Len);
+  }
+
+  action(a_assertCacheData, "ac", desc="Assert that a cache provided the data") {
+    assert(TBEs[address].CacheDirty);
+  }
+
   action(l_queueMemoryWBRequest, "lq", desc="Write PUTX data to memory") {
     peek(unblockNetwork_in, ResponseMsg) {
       enqueue(memQueue_out, MemoryMsg, latency="1") {
@@ -370,6 +589,18 @@ machine(Directory, "AMD Hammer-like protocol")
     }
   }
 
+  action(ld_queueMemoryDmaWrite, "ld", desc="Write DMA data to memory") {
+    enqueue(memQueue_out, MemoryMsg, latency="1") {
+      out_msg.Address := address;
+      out_msg.Type := MemoryRequestType:MEMORY_WB;
+      // first, initialize the data blk to the current version of system memory
+      out_msg.DataBlk := TBEs[address].DataBlk;
+      // then add the dma write data
+      out_msg.DataBlk.copyPartial(TBEs[address].DmaDataBlk, addressOffset(TBEs[address].PhysicalAddress), TBEs[address].Len);
+      DEBUG_EXPR(out_msg);
+    }
+  }
+
   action(ll_checkIncomingWriteback, "\l", desc="Check PUTX/PUTO response message") {
     peek(unblockNetwork_in, ResponseMsg) {
       assert(in_msg.Dirty == false);
@@ -379,20 +610,17 @@ machine(Directory, "AMD Hammer-like protocol")
       // implementation.  We include the data in the "dataless"
       // message so we can assert the clean data matches the datablock
       // in memory
-      assert(directory[in_msg.Address].DataBlk == in_msg.DataBlk);
+      assert(directory[address].DataBlk == in_msg.DataBlk);
     }
   }
 
-  //  action(z_stall, "z", desc="Cannot be handled right now.") {
-    // Special name recognized as do nothing case
-  //  }
-
   action(zz_recycleRequest, "\z", desc="Recycle the request queue") {
     requestQueue_in.recycle();
   }
 
   // TRANSITIONS
 
+  // Transitions out of E state
   transition(E, GETX, NO_B_W) {
     v_allocateTBE;
     rx_recordExclusiveInTBE;
@@ -409,7 +637,14 @@ machine(Directory, "AMD Hammer-like protocol")
     i_popIncomingRequestQueue;
   }
 
-  // 
+  transition(E, DMA_READ, NO_DR_B_W) {
+    vd_allocateDmaRequestInTBE;
+    qd_queueMemoryRequestFromDmaRead;
+    f_forwardReadFromDma;
+    p_popDmaRequestQueue;
+  }
+
+  // Transitions out of O state
   transition(O, GETX, NO_B_W) {
     v_allocateTBE;
     r_recordDataInTBE;
@@ -426,7 +661,20 @@ machine(Directory, "AMD Hammer-like protocol")
     i_popIncomingRequestQueue;
   }
 
-  //
+  transition(O, DMA_READ, O_DR_B_W) {
+    vd_allocateDmaRequestInTBE;
+    qd_queueMemoryRequestFromDmaRead;
+    f_forwardReadFromDma;
+    p_popDmaRequestQueue;
+  }
+
+  transition({E, O, NO}, DMA_WRITE, NO_DW_B_W) {
+    vd_allocateDmaRequestInTBE;
+    f_forwardWriteFromDma;
+    p_popDmaRequestQueue;
+  }
+
+  // Transitions out of NO state
   transition(NO, GETX, NO_B) {
     f_forwardRequest;
     i_popIncomingRequestQueue;
@@ -442,16 +690,33 @@ machine(Directory, "AMD Hammer-like protocol")
     i_popIncomingRequestQueue;
   }
 
+  transition(NO, DMA_READ, NO_DR_B_D) {
+    vd_allocateDmaRequestInTBE;
+    f_forwardReadFromDma;
+    p_popDmaRequestQueue;
+  }
+
+  // Nack PUT requests when races cause us to believe we own the data
   transition({O, E}, PUT) {
     b_sendWriteBackNack;
     i_popIncomingRequestQueue;
   }
 
-  // Blocked states
-  transition({NO_B, O_B, NO_B_W, O_B_W, NO_W, O_W, WB, WB_E_W, WB_O_W}, {GETS, GETX, PUT}) {
+  // Blocked transient states
+  transition({NO_B, O_B, NO_DR_B_W, NO_DW_B_W, NO_B_W, NO_DR_B_D, 
+              NO_DR_B, O_DR_B, O_B_W, O_DR_B_W, NO_DW_W, 
+              NO_W, O_W, WB, WB_E_W, WB_O_W}, 
+             {GETS, GETX, PUT}) {
     zz_recycleRequest;
   }
 
+  transition({NO_B, O_B, NO_DR_B_W, NO_DW_B_W, NO_B_W, NO_DR_B_D, 
+              NO_DR_B, O_DR_B, O_B_W, O_DR_B_W, NO_DW_W, 
+              NO_W, O_W, WB, WB_E_W, WB_O_W}, 
+             {DMA_READ, DMA_WRITE}) {
+    y_recycleDmaRequestQueue;
+  }
+
   transition(NO_B, Unblock, NO) {
     j_popIncomingUnblockQueue;
   }
@@ -466,6 +731,134 @@ machine(Directory, "AMD Hammer-like protocol")
     l_popMemQueue;
   }
 
+  transition(NO_DR_B_W, Memory_Data, NO_DR_B) {
+    r_recordMemoryData;
+    o_checkForCompletion;
+    l_popMemQueue;
+  }
+
+  transition(O_DR_B_W, Memory_Data, O_DR_B) {
+    r_recordMemoryData;
+    dr_sendDmaData;
+    o_checkForCompletion;
+    l_popMemQueue;
+  }
+
+  transition({NO_DR_B, O_DR_B, NO_DR_B_D, NO_DW_B_W}, Ack) {
+    m_decrementNumberOfMessages;
+    o_checkForCompletion;
+    n_popResponseQueue;
+  }
+
+  transition(NO_DR_B_W, Ack) {
+    m_decrementNumberOfMessages;
+    n_popResponseQueue;
+  }
+
+  transition(NO_DR_B_W, Shared_Ack) {
+    m_decrementNumberOfMessages;
+    r_setSharerBit;
+    n_popResponseQueue;
+  }
+
+  transition({NO_DR_B, NO_DR_B_D}, Shared_Ack) {
+    m_decrementNumberOfMessages;
+    r_setSharerBit;
+    o_checkForCompletion;
+    n_popResponseQueue;
+  }
+
+  transition(NO_DR_B_W, Shared_Data) {
+    r_recordCacheData;
+    m_decrementNumberOfMessages;
+    r_setSharerBit;
+    o_checkForCompletion;
+    n_popResponseQueue;
+  }
+
+  transition({NO_DR_B, NO_DR_B_D}, Shared_Data) {
+    r_recordCacheData;
+    m_decrementNumberOfMessages;
+    r_setSharerBit;
+    o_checkForCompletion;
+    n_popResponseQueue;
+  }
+
+  transition(NO_DR_B_W, Exclusive_Data) {
+    r_recordCacheData;
+    m_decrementNumberOfMessages;
+    n_popResponseQueue;
+  }
+
+  transition({NO_DR_B, NO_DR_B_D, NO_DW_B_W}, Exclusive_Data) {
+    r_recordCacheData;
+    m_decrementNumberOfMessages;
+    o_checkForCompletion;
+    n_popResponseQueue;
+  }
+
+  transition(NO_DR_B, All_acks_and_data, O) {
+    //
+    // Note that the DMA consistency model allows us to send the DMA device
+    // a response as soon as we receive valid data and prior to receiving
+    // all acks.  However, to simplify the protocol we wait for all acks.
+    //
+    dt_sendDmaDataFromTbe;
+    w_deallocateTBE;
+    g_popTriggerQueue;
+  }
+
+  transition(NO_DR_B_D, All_acks_and_data, O) {
+    //
+    // Note that the DMA consistency model allows us to send the DMA device
+    // a response as soon as we receive valid data and prior to receiving
+    // all acks.  However, to simplify the protocol we wait for all acks.
+    //
+    dt_sendDmaDataFromTbe;
+    w_deallocateTBE;
+    g_popTriggerQueue;
+  }
+
+  transition(O_DR_B, All_acks_and_data_no_sharers, O) {
+    w_deallocateTBE;
+    g_popTriggerQueue;
+  }
+
+  transition(NO_DR_B, All_acks_and_data_no_sharers, E) {
+    //
+    // Note that the DMA consistency model allows us to send the DMA device
+    // a response as soon as we receive valid data and prior to receiving
+    // all acks.  However, to simplify the protocol we wait for all acks.
+    //
+    dt_sendDmaDataFromTbe;
+    w_deallocateTBE;
+    g_popTriggerQueue;
+  }
+
+  transition(NO_DR_B_D, All_acks_and_data_no_sharers, E) {
+    a_assertCacheData;
+    //
+    // Note that the DMA consistency model allows us to send the DMA device
+    // a response as soon as we receive valid data and prior to receiving
+    // all acks.  However, to simplify the protocol we wait for all acks.
+    //
+    dt_sendDmaDataFromTbe;
+    w_deallocateTBE;
+    g_popTriggerQueue;
+  }
+
+  transition(NO_DW_B_W, All_acks_and_data_no_sharers, NO_DW_W) {
+    dwt_writeDmaDataFromTBE;
+    ld_queueMemoryDmaWrite;
+    g_popTriggerQueue;
+  }
+
+  transition(NO_DW_W, Memory_Ack, E) {
+    da_sendDmaAck;
+    w_deallocateTBE;
+    l_popMemQueue;
+  }
+
   transition(O_B_W, Memory_Data, O_B) {
     d_sendData;
     w_deallocateTBE;
@@ -490,7 +883,7 @@ machine(Directory, "AMD Hammer-like protocol")
     l_popMemQueue;
   }
 
-  // WB
+  // WB State Transistions
   transition(WB, Writeback_Dirty, WB_E_W) {
     l_writeDataToMemory;
     l_queueMemoryWBRequest;
diff --git a/src/mem/protocol/MOESI_hammer-dma.sm b/src/mem/protocol/MOESI_hammer-dma.sm
new file mode 100644 (file)
index 0000000..b217923
--- /dev/null
@@ -0,0 +1,165 @@
+/*
+ * Copyright (c) 1999-2008 Mark D. Hill and David A. Wood
+ * All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions are
+ * met: redistributions of source code must retain the above copyright
+ * notice, this list of conditions and the following disclaimer;
+ * redistributions in binary form must reproduce the above copyright
+ * notice, this list of conditions and the following disclaimer in the
+ * documentation and/or other materials provided with the distribution;
+ * neither the name of the copyright holders nor the names of its
+ * contributors may be used to endorse or promote products derived from
+ * this software without specific prior written permission.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+
+machine(DMA, "DMA Controller") 
+: int request_latency
+{
+
+  MessageBuffer responseFromDir, network="From", virtual_network="4", ordered="true", no_vector="true";
+  MessageBuffer reqToDirectory, network="To", virtual_network="5", ordered="false", no_vector="true";
+
+  enumeration(State, desc="DMA states", default="DMA_State_READY") {
+    READY, desc="Ready to accept a new request";
+    BUSY_RD,  desc="Busy: currently processing a request";
+    BUSY_WR,  desc="Busy: currently processing a request";
+  }
+
+  enumeration(Event, desc="DMA events") {
+    ReadRequest,  desc="A new read request";
+    WriteRequest, desc="A new write request";
+    Data,         desc="Data from a DMA memory read";
+    Ack,          desc="DMA write to memory completed";
+  }
+
+  external_type(DMASequencer) {
+    void ackCallback();
+    void dataCallback(DataBlock);
+  }
+
+  MessageBuffer mandatoryQueue, ordered="false", no_vector="true";
+  DMASequencer dma_sequencer, factory='RubySystem::getDMASequencer(m_cfg["dma_sequencer"])', no_vector="true";
+  State cur_state, no_vector="true";
+
+  State getState(Address addr) {
+    return cur_state;
+  }
+  void setState(Address addr, State state) {
+  cur_state := state;
+  }
+
+  out_port(reqToDirectory_out, DMARequestMsg, reqToDirectory, desc="...");
+
+  in_port(dmaRequestQueue_in, SequencerMsg, mandatoryQueue, desc="...") {
+    if (dmaRequestQueue_in.isReady()) {
+      peek(dmaRequestQueue_in, SequencerMsg) {
+        if (in_msg.Type == SequencerRequestType:LD ) {
+          trigger(Event:ReadRequest, in_msg.LineAddress);
+        } else if (in_msg.Type == SequencerRequestType:ST) {
+          trigger(Event:WriteRequest, in_msg.LineAddress);
+        } else {
+          error("Invalid request type");
+        }
+      }
+    }
+  }
+
+  in_port(dmaResponseQueue_in, DMAResponseMsg, responseFromDir, desc="...") {
+    if (dmaResponseQueue_in.isReady()) {
+      peek( dmaResponseQueue_in, DMAResponseMsg) {
+        if (in_msg.Type == DMAResponseType:ACK) {
+          trigger(Event:Ack, in_msg.LineAddress);
+        } else if (in_msg.Type == DMAResponseType:DATA) {
+          trigger(Event:Data, in_msg.LineAddress);
+        } else {
+          error("Invalid response type");
+        }
+      }
+    }
+  }
+
+  action(s_sendReadRequest, "s", desc="Send a DMA read request to memory") {
+    peek(dmaRequestQueue_in, SequencerMsg) {
+      enqueue(reqToDirectory_out, DMARequestMsg, latency=request_latency) {
+        out_msg.PhysicalAddress := in_msg.PhysicalAddress;
+        out_msg.LineAddress := in_msg.LineAddress; 
+        out_msg.Type := DMARequestType:READ;
+        out_msg.Requestor := machineID;
+        out_msg.DataBlk := in_msg.DataBlk;
+        out_msg.Len := in_msg.Len;
+        out_msg.Destination.add(map_Address_to_Directory(address));
+        out_msg.MessageSize := MessageSizeType:Writeback_Control;
+      }
+    }
+  }
+
+  action(s_sendWriteRequest, "\s", desc="Send a DMA write request to memory") {
+    peek(dmaRequestQueue_in, SequencerMsg) {
+      enqueue(reqToDirectory_out, DMARequestMsg, latency=request_latency) {
+          out_msg.PhysicalAddress := in_msg.PhysicalAddress;
+          out_msg.LineAddress := in_msg.LineAddress; 
+          out_msg.Type := DMARequestType:WRITE;
+          out_msg.Requestor := machineID;
+          out_msg.DataBlk := in_msg.DataBlk;
+          out_msg.Len := in_msg.Len;
+          out_msg.Destination.add(map_Address_to_Directory(address));
+          out_msg.MessageSize := MessageSizeType:Writeback_Control;
+        }
+      }
+  }
+
+  action(a_ackCallback, "a", desc="Notify dma controller that write request completed") {
+    peek (dmaResponseQueue_in, DMAResponseMsg) {
+      dma_sequencer.ackCallback();
+    }
+  }
+
+  action(d_dataCallback, "d", desc="Write data to dma sequencer") {
+    peek (dmaResponseQueue_in, DMAResponseMsg) {
+      dma_sequencer.dataCallback(in_msg.DataBlk);
+    }
+  }
+
+  action(p_popRequestQueue, "p", desc="Pop request queue") {
+    dmaRequestQueue_in.dequeue();
+  }
+
+  action(p_popResponseQueue, "\p", desc="Pop request queue") {
+    dmaResponseQueue_in.dequeue();
+  }
+
+  transition(READY, ReadRequest, BUSY_RD) {
+    s_sendReadRequest;
+    p_popRequestQueue;
+  }
+
+  transition(READY, WriteRequest, BUSY_WR) {
+    s_sendWriteRequest;
+    p_popRequestQueue;
+  }
+
+  transition(BUSY_RD, Data, READY) {
+    d_dataCallback;
+    p_popResponseQueue;
+  }
+
+  transition(BUSY_WR, Ack, READY) {
+    a_ackCallback;
+    p_popResponseQueue;
+  }
+}
index c9f14681983b28dbf7d70196cd97bb22dcb8f44d..5d8226eb6b384a32c319bff3c04b423fc37ff2b6 100644 (file)
@@ -85,3 +85,35 @@ structure(ResponseMsg, desc="...", interface="NetworkMessage") {
   int Acks,                    desc="How many messages this counts as";
   MessageSizeType MessageSize, desc="size category of the message";
 }
+
+enumeration(DMARequestType, desc="...", default="DMARequestType_NULL") {
+  READ,          desc="Memory Read";
+  WRITE,         desc="Memory Write";
+  NULL,          desc="Invalid";
+}
+
+enumeration(DMAResponseType, desc="...", default="DMAResponseType_NULL") {
+  DATA,          desc="DATA read";
+  ACK,           desc="ACK write";
+  NULL,          desc="Invalid";
+}
+
+structure(DMARequestMsg, desc="...", interface="NetworkMessage") {
+  DMARequestType Type,       desc="Request type (read/write)";
+  Address PhysicalAddress,   desc="Physical address for this request";
+  Address LineAddress,       desc="Line address for this request";
+  MachineID Requestor,            desc="Node who initiated the request";
+  NetDest Destination,       desc="Destination";
+  DataBlock DataBlk,         desc="DataBlk attached to this request";
+  int Len,                   desc="The length of the request";
+  MessageSizeType MessageSize, desc="size category of the message";
+}
+
+structure(DMAResponseMsg, desc="...", interface="NetworkMessage") {
+  DMAResponseType Type,      desc="Response type (DATA/ACK)";
+  Address PhysicalAddress,   desc="Physical address for this request";
+  Address LineAddress,       desc="Line address for this request";
+  NetDest Destination,       desc="Destination";
+  DataBlock DataBlk,         desc="DataBlk attached to this request";
+  MessageSizeType MessageSize, desc="size category of the message";
+}
index d49350e3235c3755a2658f3a42bc9fb19dcd9df0..31ad47c2efc9632c14ffbb484a5a7d2512b3bf77 100644 (file)
@@ -1,4 +1,5 @@
 MOESI_hammer-msg.sm
 MOESI_hammer-cache.sm
 MOESI_hammer-dir.sm
+MOESI_hammer-dma.sm
 standard_1level_CMP-protocol.sm
index da7fa17c741adbe356a8aeca0ed9224ea8caff05..bb054ec4eb0b13628cedfd7a8884f337863d6228 100644 (file)
@@ -178,6 +178,10 @@ class MOESI_hammer_DirectoryController < DirectoryController
   default_param :memory_controller_latency, Integer, 12
 end
 
+class MOESI_hammer_DMAController < DMAController
+  default_param :request_latency, Integer, 6
+end
+
 class RubySystem
 
   # Random seed used by the simulation. If set to "rand", the seed