On Sun, Apr 25, 2021 at 08:02:10PM -0700, Bart Van Assche wrote:

On 4/25/21 1:57 AM, Ming Lei wrote:

quoted

diff --git a/drivers/scsi/scsi_lib.c b/drivers/scsi/scsi_lib.c
index c289991ffaed..7cbaee282b6d 100644
--- a/drivers/scsi/scsi_lib.c
+++ b/drivers/scsi/scsi_lib.c

@@ -1568,7 +1568,11 @@ static void scsi_mq_done(struct scsi_cmnd *cmd)
 	if (unlikely(test_and_set_bit(SCMD_STATE_COMPLETE, &cmd->state)))
 		return;
 	trace_scsi_dispatch_cmd_done(cmd);
-	blk_mq_complete_request(cmd->request);
+
+	if (unlikely(host_byte(cmd->result) != DID_OK))
+		blk_mq_complete_request_locally(cmd->request);
+	else
+		blk_mq_complete_request(cmd->request);
 }

This change is so tricky that it deserves a comment.

An even better approach would be *not* to export
blk_mq_complete_request_locally() from the block layer to block drivers
and instead modify the block layer such that it completes a request on
the same CPU if request completion happens from inside the context of a
tag iteration function. That would save driver writers the trouble of
learning yet another block layer API.

Yeah, that is possible, and one request flag(eg. RQF_ITERATED) can be added.
The flag is set before calling ->fn(), and evaluated in
blk_mq_complete_request_remote().

Thanks,
Ming