On Thu, Mar 26, 2026 at 08:28:46AM -0600, Keith Busch wrote:

On Thu, Mar 26, 2026 at 03:51:52PM +0900, 전민식 wrote:

quoted

 {
 	struct nvme_ctrl *ctrl = nvme_req(req)->ctrl;
 
-	trace_nvme_complete_rq(req);
+	/*
+	 * The idea for these trace events was to match up commands
+	 * dispatched to hardware with the hardware's posted response.
+	 * So skip tracing for undispatched commands.
+	 */
+	if (nvme_req(req)->status != NVME_SC_HOST_PATH_ERROR)
+		trace_nvme_complete_rq(req);
+

Well, how do we know a controller doesnn't actually return that status
code? I was just suggesting to skip the trace for the condition we never
dispatched the command. An added bonus is we don't need a mostly
unnecessary 'if' check on every IO.

The 7?h error values were added for host use.  The description of
the section in the spec suggests this, but isn't actually as clear
as I would like it.  I wonder if we need to verify that controller
don't incorrectly return it, as that could cause some problems?

Independent of that your patch below looks sane to me.