From 009cec56e0e7a082ed684e98c5600babc2d2283e Mon Sep 17 00:00:00 2001
From: Matthew Poremba <matthew.poremba@amd.com>
Date: Wed, 7 Feb 2024 13:29:44 -0600
Subject: [PATCH] dev-amdgpu: Check for SDMA copies to GART range

The SDMA engine can potentially be used to write to the GART address
range. Since gem5 has a shadow copy of the GART table to avoid sending
functional reads to device memory, the GART table must be updated when
copying to the GART range.

This changeset adds a check in the VM for GART range and implements the
SDMA copy packet writing to the GART range. A fatal is added to write
and ptePde, which are the only other two ways to write to memory, as
using these packets to update the GART table has not been observed.

Change-Id: I1e62dfd9179cc9e987659e68414209fd77bba2bd
---
 src/dev/amdgpu/amdgpu_vm.hh   |  6 ++++++
 src/dev/amdgpu/sdma_engine.cc | 37 ++++++++++++++++++++++++++++-------
 2 files changed, 36 insertions(+), 7 deletions(-)

diff --git a/src/dev/amdgpu/amdgpu_vm.hh b/src/dev/amdgpu/amdgpu_vm.hh
index f35a735111..5af666f379 100644
--- a/src/dev/amdgpu/amdgpu_vm.hh
+++ b/src/dev/amdgpu/amdgpu_vm.hh
@@ -172,6 +172,12 @@ class AMDGPUVM : public Serializable
      */
     Addr gartSize();
 
+    bool
+    inGARTRange(Addr paddr)
+    {
+        return ((paddr >= gartBase()) && (paddr <= (gartBase() + gartSize())));
+    }
+
     /**
      * Copy of GART table. Typically resides in device memory, however we use
      * a copy in gem5 to simplify the interface.
diff --git a/src/dev/amdgpu/sdma_engine.cc b/src/dev/amdgpu/sdma_engine.cc
index 34ad027234..94bcdf9cb9 100644
--- a/src/dev/amdgpu/sdma_engine.cc
+++ b/src/dev/amdgpu/sdma_engine.cc
@@ -627,10 +627,14 @@ SDMAEngine::writeReadData(SDMAQueue *q, sdmaWrite *pkt, uint32_t *dmaBuffer)
 
     // lastly we write read data to the destination address
     if (gpuDevice->getVM().inMMHUB(pkt->dest)) {
-        Addr mmhubAddr = pkt->dest - gpuDevice->getVM().getMMHUBBase();
+        Addr mmhub_addr = pkt->dest - gpuDevice->getVM().getMMHUBBase();
+
+        fatal_if(gpuDevice->getVM().inGARTRange(mmhub_addr),
+                "SDMA write to GART not implemented");
+
         auto cb = new EventFunctionWrapper(
             [ = ]{ writeDone(q, pkt, dmaBuffer); }, name());
-        gpuDevice->getMemMgr()->writeRequest(mmhubAddr, (uint8_t *)dmaBuffer,
+        gpuDevice->getMemMgr()->writeRequest(mmhub_addr, (uint8_t *)dmaBuffer,
                                            bufferSize, 0, cb);
     } else {
         if (q->priv()) {
@@ -663,9 +667,11 @@ SDMAEngine::copy(SDMAQueue *q, sdmaCopy *pkt)
     // count represents the number of bytes - 1 to be copied
     pkt->count++;
     if (q->priv()) {
-        DPRINTF(SDMAEngine, "Getting GART addr for %lx\n", pkt->source);
-        pkt->source = getGARTAddr(pkt->source);
-        DPRINTF(SDMAEngine, "GART addr %lx\n", pkt->source);
+        if (!gpuDevice->getVM().inMMHUB(pkt->source)) {
+            DPRINTF(SDMAEngine, "Getting GART addr for %lx\n", pkt->source);
+            pkt->source = getGARTAddr(pkt->source);
+            DPRINTF(SDMAEngine, "GART addr %lx\n", pkt->source);
+        }
     }
 
     // Read data from the source first, then call the copyReadData method
@@ -742,6 +748,19 @@ SDMAEngine::copyReadData(SDMAQueue *q, sdmaCopy *pkt, uint8_t *dmaBuffer)
             [ = ] (const uint64_t &) { copyDone(q, pkt, dmaBuffer); });
         dmaWriteVirt(pkt->dest, pkt->count, cb, (void *)dmaBuffer);
     }
+
+    // For destinations in the GART table, gem5 uses a mapping tables instead
+    // of functionally going to device memory, so we need to update that copy.
+    if (gpuDevice->getVM().inGARTRange(device_addr)) {
+        // GART entries are always 8 bytes.
+        assert((pkt->count % 8) == 0);
+        for (int i = 0; i < pkt->count/8; ++i) {
+            Addr gart_addr = device_addr + i*8 - gpuDevice->getVM().gartBase();
+            DPRINTF(SDMAEngine, "Shadow copying to GART table %lx -> %lx\n",
+                    gart_addr, dmaBuffer64[i]);
+            gpuDevice->getVM().gartTable[gart_addr] = dmaBuffer64[i];
+        }
+    }
 }
 
 /* Completion of a copy packet. */
@@ -971,10 +990,14 @@ SDMAEngine::ptePde(SDMAQueue *q, sdmaPtePde *pkt)
 
     // Writing generated data to the destination address.
     if (gpuDevice->getVM().inMMHUB(pkt->dest)) {
-        Addr mmhubAddr = pkt->dest - gpuDevice->getVM().getMMHUBBase();
+        Addr mmhub_addr = pkt->dest - gpuDevice->getVM().getMMHUBBase();
+
+        fatal_if(gpuDevice->getVM().inGARTRange(mmhub_addr),
+                "SDMA write to GART not implemented");
+
         auto cb = new EventFunctionWrapper(
             [ = ]{ ptePdeDone(q, pkt, dmaBuffer); }, name());
-        gpuDevice->getMemMgr()->writeRequest(mmhubAddr, (uint8_t *)dmaBuffer,
+        gpuDevice->getMemMgr()->writeRequest(mmhub_addr, (uint8_t *)dmaBuffer,
                                              sizeof(uint64_t) * pkt->count, 0,
                                              cb);
     } else {