From c52d4abeceffb406de043c5470ce320870b9a84a Mon Sep 17 00:00:00 2001
From: Tim Besard <tim.besard@gmail.com>
Date: Sat, 17 Oct 2026 10:21:51 +0200
Subject: [PATCH] fix: emit uc tag update into epilogue stream when present

When an idle immediate command list dispatches a regular command list's
command buffer as the primary batch buffer, flushTaskHeapful receives
that buffer as an exactly full task stream, with the immediate command
stream as optionalEpilogueCmdStream. A pending UC tag update was still
written into the task stream, so LinearStream::getSpace aborted.

Write it into the epilogue stream instead, as the barrier with post-sync
in flushTaskHeapful and the UC tag update in flushBcsTask already do.

Related-To: https://github.com/intel/compute-runtime/issues/1027
Signed-off-by: Tim Besard <tim.besard@gmail.com>
---
 .../command_stream_receiver_hw_base.inl       |  8 +++-
 .../command_stream_receiver_tests.cpp         | 45 +++++++++++++++++++
 2 files changed, 51 insertions(+), 2 deletions(-)

diff --git a/shared/source/command_stream/command_stream_receiver_hw_base.inl b/shared/source/command_stream/command_stream_receiver_hw_base.inl
index f75aa22995530..ba863c358d2c1 100644
--- a/shared/source/command_stream/command_stream_receiver_hw_base.inl
+++ b/shared/source/command_stream/command_stream_receiver_hw_base.inl
@@ -466,8 +466,13 @@ CompletionStamp CommandStreamReceiverHw<GfxFamily>::flushTaskHeapful(
     if (detectInitProgrammingFlagsRequired(dispatchFlags)) {
         initProgrammingFlags();
     }
+
+    // a regular command list dispatched as the primary batch buffer has no space left in
+    // the task stream, so commands following the task go into the epilogue stream
+    LinearStream &epilogueCommandStream = dispatchFlags.optionalEpilogueCmdStream != nullptr ? *dispatchFlags.optionalEpilogueCmdStream : commandStreamTask;
+
     if (this->ucResourceRequiresTagUpdate) {
-        this->emitTagUpdateWithoutDCFlush(commandStreamTask);
+        this->emitTagUpdateWithoutDCFlush(epilogueCommandStream);
     }
     const auto &hwInfo = peekHwInfo();
 
@@ -478,7 +483,6 @@ CompletionStamp CommandStreamReceiverHw<GfxFamily>::flushTaskHeapful(
                                      dispatchFlags.textureCacheFlush ||
                                      this->heapStorageRequiresRecyclingTag;
     if (barrierWithPostSyncNeeded) {
-        LinearStream &epilogueCommandStream = dispatchFlags.optionalEpilogueCmdStream != nullptr ? *dispatchFlags.optionalEpilogueCmdStream : commandStreamTask;
         processBarrierWithPostSync(epilogueCommandStream, dispatchFlags, levelClosed, currentPipeControlForNooping,
                                    epiloguePipeControlLocation, hasStallingCmdsOnTaskStream, args);
     }
diff --git a/shared/test/unit_test/command_stream/command_stream_receiver_tests.cpp b/shared/test/unit_test/command_stream/command_stream_receiver_tests.cpp
index 3b4c0714545e6..d1288715c7b56 100644
--- a/shared/test/unit_test/command_stream/command_stream_receiver_tests.cpp
+++ b/shared/test/unit_test/command_stream/command_stream_receiver_tests.cpp
@@ -6473,6 +6473,51 @@ HWTEST_F(CommandStreamReceiverHwTest, givenEpilogueStreamAvailableWhenFlushTaskC
     commandStreamReceiver.getMemoryManager()->freeGraphicsMemoryImpl(commandBuffer);
 }
 
+HWTEST_F(CommandStreamReceiverHwTest, givenUcResourceRequiresTagUpdateAndFullTaskStreamWithEpilogueStreamWhenFlushTaskCalledThenTagUpdateIsDispatchedIntoEpilogueStream) {
+    using PIPE_CONTROL = typename FamilyType::PIPE_CONTROL;
+
+    auto &commandStreamReceiver = pDevice->getUltCommandStreamReceiver<FamilyType>();
+    if (commandStreamReceiver.heaplessModeEnabled) {
+        GTEST_SKIP();
+    }
+
+    GraphicsAllocation *commandBuffer = commandStreamReceiver.getMemoryManager()->allocateGraphicsMemoryWithProperties(MockAllocationProperties{commandStreamReceiver.getRootDeviceIndex(), MemoryConstants::pageSize});
+    ASSERT_NE(nullptr, commandBuffer);
+    LinearStream epilogueStream(commandBuffer);
+
+    flushTaskFlags.optionalEpilogueCmdStream = &epilogueStream;
+    commandStreamReceiver.ucResourceRequiresTagUpdate = true;
+
+    // a regular command list dispatched as the primary batch buffer fills the task stream
+    commandStream.getSpace(commandStream.getAvailableSpace());
+
+    EXPECT_NO_THROW(commandStreamReceiver.flushTask(commandStream,
+                                                    0,
+                                                    &dsh,
+                                                    &ioh,
+                                                    nullptr,
+                                                    taskLevel,
+                                                    flushTaskFlags,
+                                                    *pDevice));
+    EXPECT_EQ(commandStream.getMaxAvailableSpace(), commandStream.getUsed());
+    EXPECT_FALSE(commandStreamReceiver.ucResourceRequiresTagUpdate);
+
+    HardwareParse hwParser;
+    hwParser.parseCommands<FamilyType>(epilogueStream, 0);
+    auto ucTagAddress = commandStreamReceiver.getUcTagGPUAddress();
+    bool ucTagUpdateFound = false;
+    for (auto &cmd : findAll<PIPE_CONTROL *>(hwParser.cmdList.begin(), hwParser.cmdList.end())) {
+        auto pipeControl = genCmdCast<PIPE_CONTROL *>(*cmd);
+        uint64_t address = pipeControl->getAddressHigh();
+        address <<= 32;
+        address |= pipeControl->getAddress();
+        ucTagUpdateFound |= (address == ucTagAddress);
+    }
+    EXPECT_TRUE(ucTagUpdateFound);
+
+    commandStreamReceiver.getMemoryManager()->freeGraphicsMemoryImpl(commandBuffer);
+}
+
 HWTEST2_F(CommandStreamReceiverHwTest, givenSpecialPipelineSelectModeChangedWhenGetCmdSizeForPielineSelectIsCalledThenCorrectSizeIsReturned, IsAtMostXeCore) {
     using PIPELINE_SELECT = typename FamilyType::PIPELINE_SELECT;
     using PIPE_CONTROL = typename FamilyType::PIPE_CONTROL;