ruby: Allow multiple outstanding DMA requests

DMA sequencers and protocols can currently only issue one DMA access at
a time. This patch implements the necessary functionality to support
multiple outstanding DMA requests in Ruby.

src/mem/protocol/MESI_Two_Level-dma.sm

@@ -50,15 +50,38 @@ machine(MachineType:DMA, "DMA Controller")
     Ack,          desc="DMA write to memory completed";
   }
 
-  State cur_state;
+  structure(TBE, desc="...") {
-  Tick clockEdge();
+    State TBEState,    desc="Transient state";
-
+    DataBlock DataBlk, desc="Data";
-  State getState(Addr addr) {
-    return cur_state;
   }
 
-  void setState(Addr addr, State state) {
+  structure(TBETable, external = "yes") {
-    cur_state := state;
+    TBE lookup(Addr);
+    void allocate(Addr);
+    void deallocate(Addr);
+    bool isPresent(Addr);
+  }
+
+  void set_tbe(TBE b);
+  void unset_tbe();
+  void wakeUpAllBuffers();
+
+  TBETable TBEs, template="<DMA_TBE>", constructor="m_number_of_TBEs";
+
+  Tick clockEdge();
+
+  State getState(TBE tbe, Addr addr) {
+    if (is_valid(tbe)) {
+        return tbe.TBEState;
+    } else {
+        return State:READY;
+    }
+  }
+
+  void setState(TBE tbe, Addr addr, State state) {
+    if (is_valid(tbe)) {
+        tbe.TBEState := state;
+    }
   }
 
   AccessPermission getAccessPermission(Addr addr) {
@@ -82,9 +105,9 @@ machine(MachineType:DMA, "DMA Controller")
     if (dmaRequestQueue_in.isReady(clockEdge())) {
       peek(dmaRequestQueue_in, SequencerMsg) {
         if (in_msg.Type == SequencerRequestType:LD ) {
-          trigger(Event:ReadRequest, in_msg.LineAddress);
+          trigger(Event:ReadRequest, in_msg.LineAddress, TBEs[in_msg.LineAddress]);
         } else if (in_msg.Type == SequencerRequestType:ST) {
-          trigger(Event:WriteRequest, in_msg.LineAddress);
+          trigger(Event:WriteRequest, in_msg.LineAddress, TBEs[in_msg.LineAddress]);
         } else {
           error("Invalid request type");
         }
@@ -96,9 +119,11 @@ machine(MachineType:DMA, "DMA Controller")
     if (dmaResponseQueue_in.isReady(clockEdge())) {
       peek( dmaResponseQueue_in, ResponseMsg) {
         if (in_msg.Type == CoherenceResponseType:ACK) {
-          trigger(Event:Ack, makeLineAddress(in_msg.addr));
+          trigger(Event:Ack, makeLineAddress(in_msg.addr),
+                  TBEs[makeLineAddress(in_msg.addr)]);
         } else if (in_msg.Type == CoherenceResponseType:DATA) {
-          trigger(Event:Data, makeLineAddress(in_msg.addr));
+          trigger(Event:Data, makeLineAddress(in_msg.addr),
+                  TBEs[makeLineAddress(in_msg.addr)]);
         } else {
           error("Invalid response type");
         }
@@ -133,13 +158,28 @@ machine(MachineType:DMA, "DMA Controller")
   }
 
   action(a_ackCallback, "a", desc="Notify dma controller that write request completed") {
-    dma_sequencer.ackCallback();
+    dma_sequencer.ackCallback(address);
   }
 
   action(d_dataCallback, "d", desc="Write data to dma sequencer") {
-    peek (dmaResponseQueue_in, ResponseMsg) {
+    dma_sequencer.dataCallback(tbe.DataBlk, address);
-      dma_sequencer.dataCallback(in_msg.DataBlk);
   }
+
+  action(t_updateTBEData, "t", desc="Update TBE Data") {
+    assert(is_valid(tbe));
+    peek( dmaResponseQueue_in, ResponseMsg) {
+        tbe.DataBlk := in_msg.DataBlk;
+    }
+  }
+
+  action(v_allocateTBE, "v", desc="Allocate TBE entry") {
+    TBEs.allocate(address);
+    set_tbe(TBEs[address]);
+  }
+
+  action(w_deallocateTBE, "w", desc="Deallocate TBE entry") {
+    TBEs.deallocate(address);
+    unset_tbe();
   }
 
   action(p_popRequestQueue, "p", desc="Pop request queue") {
@@ -150,23 +190,43 @@ machine(MachineType:DMA, "DMA Controller")
     dmaResponseQueue_in.dequeue(clockEdge());
   }
 
+  action(zz_stallAndWaitRequestQueue, "zz", desc="...") {
+    stall_and_wait(dmaRequestQueue_in, address);
+  }
+
+  action(wkad_wakeUpAllDependents, "wkad", desc="wake-up all dependents") {
+    wakeUpAllBuffers();
+  }
+
   transition(READY, ReadRequest, BUSY_RD) {
+    v_allocateTBE;
     s_sendReadRequest;
     p_popRequestQueue;
   }
 
   transition(READY, WriteRequest, BUSY_WR) {
+    v_allocateTBE;
     s_sendWriteRequest;
     p_popRequestQueue;
   }
 
   transition(BUSY_RD, Data, READY) {
+    t_updateTBEData;
     d_dataCallback;
+    w_deallocateTBE;
     p_popResponseQueue;
+    wkad_wakeUpAllDependents;
   }
 
   transition(BUSY_WR, Ack, READY) {
     a_ackCallback;
+    w_deallocateTBE;
     p_popResponseQueue;
+    wkad_wakeUpAllDependents;
   }
+
+  transition({BUSY_RD,BUSY_WR}, {ReadRequest,WriteRequest}) {
+     zz_stallAndWaitRequestQueue;
+  }
+
 }

src/mem/protocol/MI_example-dma.sm

@@ -50,17 +50,38 @@ machine(MachineType:DMA, "DMA Controller")
     Ack,          desc="DMA write to memory completed";
   }
 
-  State cur_state;
+  structure(TBE, desc="...") {
-
+    State TBEState,    desc="Transient state";
-  Tick clockEdge();
+    DataBlock DataBlk, desc="Data";
-  Cycles ticksToCycles(Tick t);
-
-  State getState(Addr addr) {
-    return cur_state;
   }
 
-  void setState(Addr addr, State state) {
+  structure(TBETable, external = "yes") {
-    cur_state := state;
+    TBE lookup(Addr);
+    void allocate(Addr);
+    void deallocate(Addr);
+    bool isPresent(Addr);
+  }
+
+  void set_tbe(TBE b);
+  void unset_tbe();
+  void wakeUpAllBuffers();
+
+  TBETable TBEs, template="<DMA_TBE>", constructor="m_number_of_TBEs";
+
+  Tick clockEdge();
+
+  State getState(TBE tbe, Addr addr) {
+    if (is_valid(tbe)) {
+        return tbe.TBEState;
+    } else {
+        return State:READY;
+    }
+  }
+
+  void setState(TBE tbe, Addr addr, State state) {
+    if (is_valid(tbe)) {
+        tbe.TBEState := state;
+    }
   }
 
   AccessPermission getAccessPermission(Addr addr) {
@@ -84,9 +105,9 @@ machine(MachineType:DMA, "DMA Controller")
     if (dmaRequestQueue_in.isReady(clockEdge())) {
       peek(dmaRequestQueue_in, SequencerMsg) {
         if (in_msg.Type == SequencerRequestType:LD ) {
-          trigger(Event:ReadRequest, in_msg.LineAddress);
+          trigger(Event:ReadRequest, in_msg.LineAddress, TBEs[in_msg.LineAddress]);
         } else if (in_msg.Type == SequencerRequestType:ST) {
-          trigger(Event:WriteRequest, in_msg.LineAddress);
+          trigger(Event:WriteRequest, in_msg.LineAddress, TBEs[in_msg.LineAddress]);
         } else {
           error("Invalid request type");
         }
@@ -98,9 +119,9 @@ machine(MachineType:DMA, "DMA Controller")
     if (dmaResponseQueue_in.isReady(clockEdge())) {
       peek( dmaResponseQueue_in, DMAResponseMsg) {
         if (in_msg.Type == DMAResponseType:ACK) {
-          trigger(Event:Ack, in_msg.LineAddress);
+          trigger(Event:Ack, in_msg.LineAddress, TBEs[in_msg.LineAddress]);
         } else if (in_msg.Type == DMAResponseType:DATA) {
-          trigger(Event:Data, in_msg.LineAddress);
+          trigger(Event:Data, in_msg.LineAddress, TBEs[in_msg.LineAddress]);
         } else {
           error("Invalid response type");
         }
@@ -139,15 +160,28 @@ machine(MachineType:DMA, "DMA Controller")
   }
 
   action(a_ackCallback, "a", desc="Notify dma controller that write request completed") {
-    peek (dmaResponseQueue_in, DMAResponseMsg) {
+    dma_sequencer.ackCallback(address);
-      dma_sequencer.ackCallback();
-    }
   }
 
   action(d_dataCallback, "d", desc="Write data to dma sequencer") {
-    peek (dmaResponseQueue_in, DMAResponseMsg) {
+    dma_sequencer.dataCallback(tbe.DataBlk, address);
-      dma_sequencer.dataCallback(in_msg.DataBlk);
   }
+
+  action(t_updateTBEData, "t", desc="Update TBE Data") {
+    assert(is_valid(tbe));
+    peek( dmaResponseQueue_in, DMAResponseMsg) {
+        tbe.DataBlk := in_msg.DataBlk;
+    }
+  }
+
+  action(v_allocateTBE, "v", desc="Allocate TBE entry") {
+    TBEs.allocate(address);
+    set_tbe(TBEs[address]);
+  }
+
+  action(w_deallocateTBE, "w", desc="Deallocate TBE entry") {
+    TBEs.deallocate(address);
+    unset_tbe();
   }
 
   action(p_popRequestQueue, "p", desc="Pop request queue") {
@@ -158,23 +192,43 @@ machine(MachineType:DMA, "DMA Controller")
     dmaResponseQueue_in.dequeue(clockEdge());
   }
 
+  action(zz_stallAndWaitRequestQueue, "zz", desc="...") {
+    stall_and_wait(dmaRequestQueue_in, address);
+  }
+
+  action(wkad_wakeUpAllDependents, "wkad", desc="wake-up all dependents") {
+    wakeUpAllBuffers();
+  }
+
   transition(READY, ReadRequest, BUSY_RD) {
+    v_allocateTBE;
     s_sendReadRequest;
     p_popRequestQueue;
   }
 
   transition(READY, WriteRequest, BUSY_WR) {
+    v_allocateTBE;
     s_sendWriteRequest;
     p_popRequestQueue;
   }
 
   transition(BUSY_RD, Data, READY) {
+    t_updateTBEData;
     d_dataCallback;
+    w_deallocateTBE;
     p_popResponseQueue;
+    wkad_wakeUpAllDependents;
   }
 
   transition(BUSY_WR, Ack, READY) {
     a_ackCallback;
+    w_deallocateTBE;
     p_popResponseQueue;
+    wkad_wakeUpAllDependents;
   }
+
+  transition({BUSY_RD,BUSY_WR}, {ReadRequest,WriteRequest}) {
+     zz_stallAndWaitRequestQueue;
+  }
+
 }

src/mem/protocol/MOESI_CMP_directory-dma.sm

@@ -184,7 +184,7 @@ machine(MachineType:DMA, "DMA Controller")
   }
 
   action(a_ackCallback, "a", desc="Notify dma controller that write request completed") {
-      dma_sequencer.ackCallback();
+      dma_sequencer.ackCallback(address);
   }
 
   action(o_checkForCompletion, "o", desc="Check if we have received all the messages required for completion") {
@@ -236,7 +236,7 @@ machine(MachineType:DMA, "DMA Controller")
 
   action(d_dataCallbackFromTBE, "/d", desc="data callback with data from TBE") {
     assert(is_valid(tbe));
-    dma_sequencer.dataCallback(tbe.DataBlk);
+    dma_sequencer.dataCallback(tbe.DataBlk, address);
   }
 
   action(v_allocateTBE, "v", desc="Allocate TBE entry") {

src/mem/protocol/MOESI_CMP_token-dma.sm

@@ -52,16 +52,38 @@ machine(MachineType:DMA, "DMA Controller")
     Ack,          desc="DMA write to memory completed";
   }
 
-  State cur_state;
+  structure(TBE, desc="...") {
+    State TBEState,    desc="Transient state";
+    DataBlock DataBlk, desc="Data";
+  }
+
+  structure(TBETable, external = "yes") {
+    TBE lookup(Addr);
+    void allocate(Addr);
+    void deallocate(Addr);
+    bool isPresent(Addr);
+  }
+
+  void set_tbe(TBE b);
+  void unset_tbe();
+  void wakeUpAllBuffers();
+
+  TBETable TBEs, template="<DMA_TBE>", constructor="m_number_of_TBEs";
 
   Tick clockEdge();
 
-  State getState(Addr addr) {
+  State getState(TBE tbe, Addr addr) {
-    return cur_state;
+    if (is_valid(tbe)) {
+        return tbe.TBEState;
+    } else {
+        return State:READY;
+    }
   }
 
-  void setState(Addr addr, State state) {
+  void setState(TBE tbe, Addr addr, State state) {
-    cur_state := state;
+    if (is_valid(tbe)) {
+        tbe.TBEState := state;
+    }
   }
 
   AccessPermission getAccessPermission(Addr addr) {
@@ -85,9 +107,9 @@ machine(MachineType:DMA, "DMA Controller")
     if (dmaRequestQueue_in.isReady(clockEdge())) {
       peek(dmaRequestQueue_in, SequencerMsg) {
         if (in_msg.Type == SequencerRequestType:LD ) {
-          trigger(Event:ReadRequest, in_msg.LineAddress);
+          trigger(Event:ReadRequest, in_msg.LineAddress, TBEs[in_msg.LineAddress]);
         } else if (in_msg.Type == SequencerRequestType:ST) {
-          trigger(Event:WriteRequest, in_msg.LineAddress);
+          trigger(Event:WriteRequest, in_msg.LineAddress, TBEs[in_msg.LineAddress]);
         } else {
           error("Invalid request type");
         }
@@ -99,9 +121,9 @@ machine(MachineType:DMA, "DMA Controller")
     if (dmaResponseQueue_in.isReady(clockEdge())) {
       peek( dmaResponseQueue_in, DMAResponseMsg) {
         if (in_msg.Type == DMAResponseType:ACK) {
-          trigger(Event:Ack, in_msg.LineAddress);
+          trigger(Event:Ack, in_msg.LineAddress, TBEs[in_msg.LineAddress]);
         } else if (in_msg.Type == DMAResponseType:DATA) {
-          trigger(Event:Data, in_msg.LineAddress);
+          trigger(Event:Data, in_msg.LineAddress, TBEs[in_msg.LineAddress]);
         } else {
           error("Invalid response type");
         }
@@ -140,15 +162,28 @@ machine(MachineType:DMA, "DMA Controller")
   }
 
   action(a_ackCallback, "a", desc="Notify dma controller that write request completed") {
-    peek (dmaResponseQueue_in, DMAResponseMsg) {
+    dma_sequencer.ackCallback(address);
-      dma_sequencer.ackCallback();
-    }
   }
 
   action(d_dataCallback, "d", desc="Write data to dma sequencer") {
-    peek (dmaResponseQueue_in, DMAResponseMsg) {
+    dma_sequencer.dataCallback(tbe.DataBlk, address);
-      dma_sequencer.dataCallback(in_msg.DataBlk);
   }
+
+  action(t_updateTBEData, "t", desc="Update TBE Data") {
+    assert(is_valid(tbe));
+    peek(dmaResponseQueue_in, DMAResponseMsg) {
+        tbe.DataBlk := in_msg.DataBlk;
+    }
+  }
+
+  action(v_allocateTBE, "v", desc="Allocate TBE entry") {
+    TBEs.allocate(address);
+    set_tbe(TBEs[address]);
+  }
+
+  action(w_deallocateTBE, "w", desc="Deallocate TBE entry") {
+    TBEs.deallocate(address);
+    unset_tbe();
   }
 
   action(p_popRequestQueue, "p", desc="Pop request queue") {
@@ -159,23 +194,42 @@ machine(MachineType:DMA, "DMA Controller")
     dmaResponseQueue_in.dequeue(clockEdge());
   }
 
+  action(zz_stallAndWaitRequestQueue, "zz", desc="...") {
+    stall_and_wait(dmaRequestQueue_in, address);
+  }
+
+  action(wkad_wakeUpAllDependents, "wkad", desc="wake-up all dependents") {
+    wakeUpAllBuffers();
+  }
+
   transition(READY, ReadRequest, BUSY_RD) {
+    v_allocateTBE;
     s_sendReadRequest;
     p_popRequestQueue;
   }
 
   transition(READY, WriteRequest, BUSY_WR) {
+    v_allocateTBE;
     s_sendWriteRequest;
     p_popRequestQueue;
   }
 
   transition(BUSY_RD, Data, READY) {
+    t_updateTBEData;
     d_dataCallback;
+    w_deallocateTBE;
     p_popResponseQueue;
+    wkad_wakeUpAllDependents;
   }
 
   transition(BUSY_WR, Ack, READY) {
     a_ackCallback;
+    w_deallocateTBE;
     p_popResponseQueue;
+    wkad_wakeUpAllDependents;
+  }
+
+  transition({BUSY_RD,BUSY_WR}, {ReadRequest,WriteRequest}) {
+     zz_stallAndWaitRequestQueue;
   }
 }

src/mem/protocol/MOESI_hammer-dma.sm

@@ -50,15 +50,38 @@ machine(MachineType:DMA, "DMA Controller")
     Ack,          desc="DMA write to memory completed";
   }
 
-  State cur_state;
+  structure(TBE, desc="...") {
+    State TBEState,    desc="Transient state";
+    DataBlock DataBlk, desc="Data";
+  }
+
+  structure(TBETable, external = "yes") {
+    TBE lookup(Addr);
+    void allocate(Addr);
+    void deallocate(Addr);
+    bool isPresent(Addr);
+  }
+
+  void set_tbe(TBE b);
+  void unset_tbe();
+  void wakeUpAllBuffers();
+
+  TBETable TBEs, template="<DMA_TBE>", constructor="m_number_of_TBEs";
 
   Tick clockEdge();
 
-  State getState(Addr addr) {
+  State getState(TBE tbe, Addr addr) {
-    return cur_state;
+    if (is_valid(tbe)) {
+        return tbe.TBEState;
+    } else {
+        return State:READY;
+    }
+  }
+
+  void setState(TBE tbe, Addr addr, State state) {
+    if (is_valid(tbe)) {
+        tbe.TBEState := state;
     }
-  void setState(Addr addr, State state) {
-    cur_state := state;
   }
 
   AccessPermission getAccessPermission(Addr addr) {
@@ -82,9 +105,9 @@ machine(MachineType:DMA, "DMA Controller")
     if (dmaRequestQueue_in.isReady(clockEdge())) {
       peek(dmaRequestQueue_in, SequencerMsg) {
         if (in_msg.Type == SequencerRequestType:LD ) {
-          trigger(Event:ReadRequest, in_msg.LineAddress);
+          trigger(Event:ReadRequest, in_msg.LineAddress, TBEs[in_msg.LineAddress]);
         } else if (in_msg.Type == SequencerRequestType:ST) {
-          trigger(Event:WriteRequest, in_msg.LineAddress);
+          trigger(Event:WriteRequest, in_msg.LineAddress, TBEs[in_msg.LineAddress]);
         } else {
           error("Invalid request type");
         }
@@ -96,9 +119,9 @@ machine(MachineType:DMA, "DMA Controller")
     if (dmaResponseQueue_in.isReady(clockEdge())) {
       peek( dmaResponseQueue_in, DMAResponseMsg) {
         if (in_msg.Type == DMAResponseType:ACK) {
-          trigger(Event:Ack, in_msg.LineAddress);
+          trigger(Event:Ack, in_msg.LineAddress, TBEs[in_msg.LineAddress]);
         } else if (in_msg.Type == DMAResponseType:DATA) {
-          trigger(Event:Data, in_msg.LineAddress);
+          trigger(Event:Data, in_msg.LineAddress, TBEs[in_msg.LineAddress]);
         } else {
           error("Invalid response type");
         }
@@ -137,15 +160,28 @@ machine(MachineType:DMA, "DMA Controller")
   }
 
   action(a_ackCallback, "a", desc="Notify dma controller that write request completed") {
-    peek (dmaResponseQueue_in, DMAResponseMsg) {
+    dma_sequencer.ackCallback(address);
-      dma_sequencer.ackCallback();
-    }
   }
 
   action(d_dataCallback, "d", desc="Write data to dma sequencer") {
-    peek (dmaResponseQueue_in, DMAResponseMsg) {
+    dma_sequencer.dataCallback(tbe.DataBlk, address);
-      dma_sequencer.dataCallback(in_msg.DataBlk);
   }
+
+  action(t_updateTBEData, "t", desc="Update TBE Data") {
+    assert(is_valid(tbe));
+    peek( dmaResponseQueue_in, DMAResponseMsg) {
+        tbe.DataBlk := in_msg.DataBlk;
+    }
+  }
+
+  action(v_allocateTBE, "v", desc="Allocate TBE entry") {
+    TBEs.allocate(address);
+    set_tbe(TBEs[address]);
+  }
+
+  action(w_deallocateTBE, "w", desc="Deallocate TBE entry") {
+    TBEs.deallocate(address);
+    unset_tbe();
   }
 
   action(p_popRequestQueue, "p", desc="Pop request queue") {
@@ -156,23 +192,43 @@ machine(MachineType:DMA, "DMA Controller")
     dmaResponseQueue_in.dequeue(clockEdge());
   }
 
+  action(zz_stallAndWaitRequestQueue, "zz", desc="...") {
+    stall_and_wait(dmaRequestQueue_in, address);
+  }
+
+  action(wkad_wakeUpAllDependents, "wkad", desc="wake-up all dependents") {
+    wakeUpAllBuffers();
+  }
+
   transition(READY, ReadRequest, BUSY_RD) {
+    v_allocateTBE;
     s_sendReadRequest;
     p_popRequestQueue;
   }
 
   transition(READY, WriteRequest, BUSY_WR) {
+    v_allocateTBE;
     s_sendWriteRequest;
     p_popRequestQueue;
   }
 
   transition(BUSY_RD, Data, READY) {
+    t_updateTBEData;
     d_dataCallback;
+    w_deallocateTBE;
     p_popResponseQueue;
+    wkad_wakeUpAllDependents;
   }
 
   transition(BUSY_WR, Ack, READY) {
     a_ackCallback;
+    w_deallocateTBE;
     p_popResponseQueue;
+    wkad_wakeUpAllDependents;
   }
+
+  transition({BUSY_RD,BUSY_WR}, {ReadRequest,WriteRequest}) {
+     zz_stallAndWaitRequestQueue;
+  }
+
 }

src/mem/protocol/RubySlicc_Types.sm

@@ -220,8 +220,8 @@ structure (WireBuffer, inport="yes", outport="yes", external = "yes") {
 }
 
 structure (DMASequencer, external = "yes") {
-  void ackCallback();
+  void ackCallback(Addr);
-  void dataCallback(DataBlock);
+  void dataCallback(DataBlock,Addr);
   void recordRequestType(CacheRequestType);
 }
 

src/mem/ruby/system/DMASequencer.cc

@@ -35,8 +35,18 @@
 #include "mem/ruby/system/DMASequencer.hh"
 #include "mem/ruby/system/RubySystem.hh"
 
+DMARequest::DMARequest(uint64_t start_paddr, int len, bool write,
+                       int bytes_completed, int bytes_issued, uint8_t *data,
+                       PacketPtr pkt)
+    : start_paddr(start_paddr), len(len), write(write),
+      bytes_completed(bytes_completed), bytes_issued(bytes_issued), data(data),
+      pkt(pkt)
+{
+}
+
 DMASequencer::DMASequencer(const Params *p)
-    : RubyPort(p)
+    : RubyPort(p), m_outstanding_count(0),
+      m_max_outstanding_requests(p->max_outstanding_requests)
 {
 }
 
@@ -44,7 +54,6 @@ void
 DMASequencer::init()
 {
     RubyPort::init();
-    m_is_busy = false;
     m_data_block_mask = mask(RubySystem::getBlockSizeBits());
 
     for (const auto &s_port : slave_ports)
@@ -54,7 +63,7 @@ DMASequencer::init()
 RequestStatus
 DMASequencer::makeRequest(PacketPtr pkt)
 {
-    if (m_is_busy) {
+    if (m_outstanding_count == m_max_outstanding_requests) {
         return RequestStatus_BufferFull;
     }
 
@@ -63,21 +72,29 @@ DMASequencer::makeRequest(PacketPtr pkt)
     int len = pkt->getSize();
     bool write = pkt->isWrite();
 
-    assert(!m_is_busy);  // only support one outstanding DMA request
+    assert(m_outstanding_count < m_max_outstanding_requests);
-    m_is_busy = true;
+    Addr line_addr = makeLineAddress(paddr);
+    auto emplace_pair =
+        m_RequestTable.emplace(std::piecewise_construct,
+                               std::forward_as_tuple(line_addr),
+                               std::forward_as_tuple(paddr, len, write, 0,
+                                                     0, data, pkt));
+    DMARequest& active_request = emplace_pair.first->second;
 
-    active_request.start_paddr = paddr;
+    // This is pretty conservative.  A regular Sequencer with a  more beefy
-    active_request.write = write;
+    // request table that can track multiple requests for a cache line should
-    active_request.data = data;
+    // be used if a more aggressive policy is needed.
-    active_request.len = len;
+    if (!emplace_pair.second) {
-    active_request.bytes_completed = 0;
+            DPRINTF(RubyDma, "DMA aliased: addr %p, len %d\n", line_addr, len);
-    active_request.bytes_issued = 0;
+            return RequestStatus_Aliased;
-    active_request.pkt = pkt;
+    }
+
+    DPRINTF(RubyDma, "DMA req created: addr %p, len %d\n", line_addr, len);
 
     std::shared_ptr<SequencerMsg> msg =
         std::make_shared<SequencerMsg>(clockEdge());
     msg->getPhysicalAddress() = paddr;
-    msg->getLineAddress() = makeLineAddress(msg->getPhysicalAddress());
+    msg->getLineAddress() = line_addr;
     msg->getType() = write ? SequencerRequestType_ST : SequencerRequestType_LD;
     int offset = paddr & m_data_block_mask;
 
@@ -90,6 +107,8 @@ DMASequencer::makeRequest(PacketPtr pkt)
         }
     }
 
+    m_outstanding_count++;
+
     assert(m_mandatory_q_ptr != NULL);
     m_mandatory_q_ptr->enqueue(msg, clockEdge(), cyclesToTicks(Cycles(1)));
     active_request.bytes_issued += msg->getLen();
@@ -98,18 +117,22 @@ DMASequencer::makeRequest(PacketPtr pkt)
 }
 
 void
-DMASequencer::issueNext()
+DMASequencer::issueNext(const Addr& address)
 {
-    assert(m_is_busy);
+    RequestTable::iterator i = m_RequestTable.find(address);
+    assert(i != m_RequestTable.end());
+
+    DMARequest &active_request = i->second;
+
+    assert(m_outstanding_count <= m_max_outstanding_requests);
     active_request.bytes_completed = active_request.bytes_issued;
     if (active_request.len == active_request.bytes_completed) {
-        //
+        DPRINTF(RubyDma, "DMA request completed: addr %p, size %d\n",
-        // Must unset the busy flag before calling back the dma port because
+                address, active_request.len);
-        // the callback may cause a previously nacked request to be reissued
+        m_outstanding_count--;
-        //
+        PacketPtr pkt = active_request.pkt;
-        DPRINTF(RubyDma, "DMA request completed\n");
+        m_RequestTable.erase(i);
-        m_is_busy = false;
+        ruby_hit_callback(pkt);
-        ruby_hit_callback(active_request.pkt);
         return;
     }
 
@@ -146,9 +169,13 @@ DMASequencer::issueNext()
 }
 
 void
-DMASequencer::dataCallback(const DataBlock & dblk)
+DMASequencer::dataCallback(const DataBlock & dblk, const Addr& address)
 {
-    assert(m_is_busy);
+
+    RequestTable::iterator i = m_RequestTable.find(address);
+    assert(i != m_RequestTable.end());
+
+    DMARequest &active_request = i->second;
     int len = active_request.bytes_issued - active_request.bytes_completed;
     int offset = 0;
     if (active_request.bytes_completed == 0)
@@ -158,13 +185,16 @@ DMASequencer::dataCallback(const DataBlock & dblk)
         memcpy(&active_request.data[active_request.bytes_completed],
                dblk.getData(offset, len), len);
     }
-    issueNext();
+    issueNext(address);
 }
 
 void
-DMASequencer::ackCallback()
+DMASequencer::ackCallback(const Addr& address)
 {
-    issueNext();
+    RequestTable::iterator i = m_RequestTable.find(address);
+    assert(i != m_RequestTable.end());
+
+    issueNext(address);
 }
 
 void

src/mem/ruby/system/DMASequencer.hh

@@ -31,14 +31,19 @@
 
 #include <memory>
 #include <ostream>
+#include <unordered_map>
 
 #include "mem/protocol/DMASequencerRequestType.hh"
+#include "mem/ruby/common/Address.hh"
 #include "mem/ruby/common/DataBlock.hh"
 #include "mem/ruby/system/RubyPort.hh"
 #include "params/DMASequencer.hh"
 
 struct DMARequest
 {
+    DMARequest(uint64_t start_paddr, int len, bool write, int bytes_completed,
+               int bytes_issued, uint8_t *data, PacketPtr pkt);
+
     uint64_t start_paddr;
     int len;
     bool write;
@@ -57,23 +62,27 @@ class DMASequencer : public RubyPort
 
     /* external interface */
     RequestStatus makeRequest(PacketPtr pkt) override;
-    bool busy() { return m_is_busy;}
+    bool busy() { return m_outstanding_count > 0; }
-    int outstandingCount() const override { return (m_is_busy ? 1 : 0); }
+    int outstandingCount() const override { return m_outstanding_count; }
     bool isDeadlockEventScheduled() const override { return false; }
     void descheduleDeadlockEvent() override {}
 
     /* SLICC callback */
-    void dataCallback(const DataBlock & dblk);
+    void dataCallback(const DataBlock &dblk, const Addr &addr);
-    void ackCallback();
+    void ackCallback(const Addr &addr);
 
     void recordRequestType(DMASequencerRequestType requestType);
 
   private:
-    void issueNext();
+    void issueNext(const Addr &addr);
 
-    bool m_is_busy;
     uint64_t m_data_block_mask;
-    DMARequest active_request;
+
+    typedef std::unordered_map<Addr, DMARequest> RequestTable;
+    RequestTable m_RequestTable;
+
+    int m_outstanding_count;
+    int m_max_outstanding_requests;
 };
 
 #endif // __MEM_RUBY_SYSTEM_DMASEQUENCER_HH__

src/mem/ruby/system/Sequencer.py

@@ -81,3 +81,4 @@ class RubySequencer(RubyPort):
 class DMASequencer(RubyPort):
    type = 'DMASequencer'
    cxx_header = "mem/ruby/system/DMASequencer.hh"
+   max_outstanding_requests = Param.Int(64, "max outstanding requests")