[PATCH] mmc: core: clear MQRQ_XFER_SINGLE_BLOCK for new requests

[Fexiven]

mmc_blk_mq_complete_rq() and mmc_blk_cqe_complete_rq() set
MQRQ_XFER_SINGLE_BLOCK in the request PDU when a transfer fails, so the
retry is done one block at a time. Nothing ever clears the flag again.

The PDU belongs to the blk-mq tag and is reused for later requests, while
mmc_mq_queue_rq() only resets ->retries for a new request. Every tag that
has seen a single transfer error therefore keeps issuing single-block
transfers (CMD17/CMD24) for the rest of the uptime.

On an RK3576 board (HS400ES eMMC) with occasional data CRC errors, which
are recovered correctly, sequential direct reads dropped from ~200 MB/s
after boot to 3-10 MB/s within minutes and never recovered. The host
registers showed only CMD17 with a 512 byte block size. With this change
the read throughput stays at ~200 MB/s and writes at 60-90 MB/s while the
data CRC errors continue to be retried.

Reset the flags together with the retry counter when a request is
prepared for the first time. Requeued requests keep RQF_DONTPREP, so the
single-block retry of the failing request itself is unchanged.

Fixes: d34124edffdb ("mmc: block: Use MQRQ_XFER_SINGLE_BLOCK for both read and write recovery")
Cc: stable@xxxxxxxxxxxxxxx
Signed-off-by: Fexiven <kernel@xxxxxxxxxxxx>
---
 drivers/mmc/core/queue.c | 2 ++
 1 file changed, 2 insertions(+)

--- a/drivers/mmc/core/queue.c
+++ b/drivers/mmc/core/queue.c
@@ -301,6 +301,8 @@

 	if (!(req->rq_flags & RQF_DONTPREP)) {
 		req_to_mmc_queue_req(req)->retries = 0;
+		/* the request PDU is reused: don't inherit single-block recovery */
+		req_to_mmc_queue_req(req)->flags = 0;
 		req->rq_flags |= RQF_DONTPREP;
 	}