diff --git a/src/amd/vulkan/radv_cmd_buffer.c b/src/amd/vulkan/radv_cmd_buffer.c
index cae08e1ca85..2bdbdec81e0 100644
--- a/src/amd/vulkan/radv_cmd_buffer.c
+++ b/src/amd/vulkan/radv_cmd_buffer.c
@@ -15333,10 +15333,19 @@ radv_CmdBeginTransformFeedbackEXT(VkCommandBuffer commandBuffer, uint32_t firstC
 
    assert(firstCounterBuffer + counterBufferCount <= MAX_SO_BUFFERS);
 
-   if (pdev->info.gfx_level >= GFX12)
+   if (pdev->info.gfx_level >= GFX12) {
       radv_init_streamout_state(cmd_buffer);
-   else if (!pdev->use_ngg_streamout)
+
+      /* Invalidate L2 in case the buffer filled size needs to be saved because COPY_DATA isn't
+       * coherent with L2.
+       */
+      if (pdev->info.cp_sdma_ge_use_system_memory_scope) {
+         cmd_buffer->state.flush_bits |= RADV_CMD_FLAG_INV_L2;
+         radv_emit_cache_flush(cmd_buffer);
+      }
+   } else if (!pdev->use_ngg_streamout) {
       radv_flush_vgt_streamout(cmd_buffer);
+   }
 
    ASSERTED unsigned cdw_max = radeon_check_space(device->ws, cs->b, MAX_SO_BUFFERS * 10);
 
@@ -15534,8 +15543,8 @@ radv_emit_strmout_buffer(struct radv_cmd_buffer *cmd_buffer, const struct radv_d
    radeon_end();
 
    if (gfx_level >= GFX10) {
-      /* Emitting a COPY_DATA packet should be enough because RADV doesn't support preemption
-       * (shadow memory) but for unknown reasons, it can lead to GPU hangs on GFX10+.
+      /* Make sure that PFP waits for ME to avoid a race condition because the data is written by
+       * STRMOUT_BUFFER_UPDATE in ME, but LOAD_CONTEXT_REG_INDEX is in PFP.
        */
       ac_emit_cp_pfp_sync_me(cs->b, false);