dev-amdgpu: Fix pending PCI RLC doorbell (#1157)

SDMA RLC queues do not currently remove their doorbell mapping. This can
cause issues re-registering the queue and prevents the pending doorbells
feature from working. In addition the data value of the doorbell (the
ring buffer rptr) is not saved, leading to UB when this workaround is
used.

This commit removes the doorbell mapping from the gpu device when the
SDMA engine unmaps an RLC queue and copies the next doorbell value to
the pending packet as was originally intended.

Change-Id: Ifd551450f439c065579afcf916f8ff192e7598ab

src/dev/amdgpu/amdgpu_device.cc

@@ -517,10 +517,13 @@ AMDGPUDevice::writeDoorbell(PacketPtr pkt, Addr offset)
              offset);
 
         // We have to ACK the PCI packet immediately, so create a copy of the
-        // packet here to send again.
+        // packet here to send again. The packet data contains the value of
+        // the doorbell to write so we need to copy that as the original
+        // packet gets deleted after the PCI write() method returns.
         RequestPtr pending_req(pkt->req);
         PacketPtr pending_pkt = Packet::createWrite(pending_req);
         uint8_t *pending_data = new uint8_t[pkt->getSize()];
+        memcpy(pending_data, pkt->getPtr<uint8_t>(), pkt->getSize());
         pending_pkt->dataDynamic(pending_data);
 
         pendingDoorbellPkts.emplace(offset, pending_pkt);
@@ -709,6 +712,12 @@ AMDGPUDevice::setDoorbellType(uint32_t offset, QueueType qt, int ip_id)
     doorbells[offset].ip_id = ip_id;
 }
 
+void
+AMDGPUDevice::unsetDoorbell(uint32_t offset)
+{
+    doorbells.erase(offset);
+}
+
 void
 AMDGPUDevice::setSDMAEngine(Addr offset, SDMAEngine *eng)
 {

src/dev/amdgpu/amdgpu_device.hh

@@ -196,6 +196,7 @@ class AMDGPUDevice : public PciDevice
      * Set handles to GPU blocks.
      */
     void setDoorbellType(uint32_t offset, QueueType qt, int ip_id = 0);
+    void unsetDoorbell(uint32_t offset);
     void processPendingDoorbells(uint32_t offset);
     void setSDMAEngine(Addr offset, SDMAEngine *eng);
 

src/dev/amdgpu/sdma_engine.cc

@@ -269,6 +269,7 @@ SDMAEngine::deallocateRLCQueues()
     for (auto doorbell: rlcInfo) {
         if (doorbell) {
             unregisterRLCQueue(doorbell);
+            gpuDevice->unsetDoorbell(doorbell);
         }
     }
 }