From ab85674f8a061b4f29c1cd8bfc200c06b080d568 Mon Sep 17 00:00:00 2001
From: German Andryeyev <German.Andryeyev@amd.com>
Date: Thu, 12 Mar 2020 12:26:00 -0400
Subject: [PATCH] SWDEV-224471 - [19H1][Polaris10][Acer]

Test window will quit or show error message when run OpenCL API.

Add a workaround for the race condition with the first page
during pinning.

Change-Id: I9a27b4e173cf94c84aefcb94e255f11169453d94
---
 rocclr/device/gpu/gpuvirtual.cpp | 47 ++++++++++++++++++++++++++++++--
 1 file changed, 44 insertions(+), 3 deletions(-)

diff --git a/rocclr/device/gpu/gpuvirtual.cpp b/rocclr/device/gpu/gpuvirtual.cpp
index a7d4fa5bfc..f84cfadaf7 100644
--- a/rocclr/device/gpu/gpuvirtual.cpp
+++ b/rocclr/device/gpu/gpuvirtual.cpp
@@ -680,8 +680,28 @@ void VirtualGPU::submitReadMemory(amd::ReadMemoryCommand& vcmd) {
         result = blitMgr().copyBuffer(*memory, *hostMemory, origin, dstOrigin, size,
                                       vcmd.isEntireMemory());
       } else {
-        result =
-            blitMgr().readBuffer(*memory, vcmd.destination(), origin, size, vcmd.isEntireMemory());
+        // The logic below will perform 2 step copy to make sure memory pinning doesn't
+        // occur on the first unaligned page, because in Windows memory manager can
+        // have CPU access to the allocation header in another thread
+        // and a race condition is possible.
+        char* tmpHost =
+            amd::alignUp(reinterpret_cast<char*>(vcmd.destination()), PinnedMemoryAlignment);
+
+        // Find the partial size for unaligned copy
+        size_t partial = tmpHost - reinterpret_cast<char*>(vcmd.destination());
+        result = true;
+        // Check if it's staging copy, then ignore unaligned address
+        if (size[0] <= dev().settings().pinnedMinXferSize_) {
+          partial = size[0];
+        }
+        // Make first step transfer
+        if (partial > 0) {
+          result = blitMgr().readBuffer(*memory, vcmd.destination(), origin, partial);
+        }
+        // Second step transfer if something left to copy
+        if (partial < size[0]) {
+          result &= blitMgr().readBuffer(*memory, tmpHost, origin[0] + partial, size[0] - partial);
+        }
       }
       if (NULL != bufferFromImage) {
         bufferFromImage->release();
@@ -777,7 +797,28 @@ void VirtualGPU::submitWriteMemory(amd::WriteMemoryCommand& vcmd) {
         result = blitMgr().copyBuffer(*hostMemory, *memory, srcOrigin, origin, size,
                                       vcmd.isEntireMemory());
       } else {
-        result = blitMgr().writeBuffer(vcmd.source(), *memory, origin, size, vcmd.isEntireMemory());
+        // The logic below will perform 2 step copy to make sure memory pinning doesn't
+        // occur on the first unaligned page, because in Windows memory manager can
+        // have CPU access to the allocation header in another thread
+        // and a race condition is possible.
+        const char* tmpHost =
+            amd::alignUp(reinterpret_cast<const char*>(vcmd.source()), PinnedMemoryAlignment);
+
+        // Find the partial size for unaligned copy
+        size_t partial = tmpHost - reinterpret_cast<const char*>(vcmd.source());
+        result = true;
+        // Check if it's staging copy, then ignore unaligned address
+        if (size[0] <= dev().settings().pinnedMinXferSize_) {
+          partial = size[0];
+        }
+        // Make first step transfer
+        if (partial > 0) {
+          result = blitMgr().writeBuffer(vcmd.source(), *memory, origin, partial);
+        }
+        // Second step transfer if something left to copy
+        if (partial < size[0]) {
+          result &= blitMgr().writeBuffer(tmpHost, *memory, origin[0] + partial, size[0] - partial);
+        }
       }
       if (NULL != bufferFromImage) {
         bufferFromImage->release();