Re: [PATCH] mmc: core: Add packed command for eMMC4.5 device

[me...@codeaurora.org]

> + if (reqs >= 2) {
> + mmc_blk_packed_hdr_wrq_prep(mq->mqrq_cur, card, mq, reqs);
> + if (rq_data_dir(rqc) == READ) {
> + areq = &mq->mqrq_cur->mmc_active;
> + mmc_wait_for_req(card->host, areq->mrq);
Packing read requests requires preparation of two requests. After sending
the header we wait for its completion before sending the next request
(mmc_wait_for_req is used). Therefore, if we try to pack 2 read requests
we might end up with worse performance in comparison to sending each
request by itself (which allows the preparation of one request while the
other is sent).
I suggest to check the size of the packed commands list and in case it is
less than 3 send the requests one by one. If you move mmc_blk_chk_packable
to queue.c after the first fetch this change should be very easy and can
be done by removing the requests from the packed_list and calling issue_fn
for each one of them.

Thanks,
Maya Erez
Consultant for Qualcomm Innovation Center, Inc.
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum






--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majo...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

[Seungwon Jeon]

Maya Erez wrote:
>
> > + if (reqs >= 2) {
> > + mmc_blk_packed_hdr_wrq_prep(mq->mqrq_cur, card, mq, reqs);
> > + if (rq_data_dir(rqc) == READ) {
> > + areq = &mq->mqrq_cur->mmc_active;
> > + mmc_wait_for_req(card->host, areq->mrq);
> Packing read requests requires preparation of two requests. After sending
> the header we wait for its completion before sending the next request
> (mmc_wait_for_req is used). Therefore, if we try to pack 2 read requests
> we might end up with worse performance in comparison to sending each
> request by itself (which allows the preparation of one request while the
> other is sent).
> I suggest to check the size of the packed commands list and in case it is
> less than 3 send the requests one by one. If you move mmc_blk_chk_packable
> to queue.c after the first fetch this change should be very easy and can
> be done by removing the requests from the packed_list and calling issue_fn
> for each one of them.

Sending header for packed read which doesn't require nand program
unlike normal data, so it may not spend long time.
Which point you think is the overhead of packed two-requests
in comparison to individual request?

>
> Thanks,
> Maya Erez
> Consultant for Qualcomm Innovation Center, Inc.
> Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum
>
>
>
>
>
>
> --

> To unsubscribe from this list: send the line "unsubscribe linux-mmc" in

[me...@codeaurora.org]

When you send the packed commands you have to send an additional sector.
When you have only 2 or even 3 requests to send this might not be
negligible.
I think it's worth checking how often we send 2-3 requests in the packed
commands and if it's better to send it separately or packed. According to
the results you can decide if it's worth handling such cases.