On Sun, Jul 7, 2019 at 8:11 PM Yufen Yu <yuyufen@xxxxxxxxxx> wrote:
>
> When raid1d() submits a lot of IO, block plug can improve IO merge
> and increase IO throughput. But, after commit 18022a1bd37 ("md/raid1/10:
> add missed blk plug"), flush_pending_writes have added start/finish
> plug by itself.
>
> Except flush_pending_writes(), other processing in raid1d() would
> not submit a lot of IO. They may issue some IO, but these IOs are
> synchronous, i.e. submit bio and wait it finish. Plug can increase
> io latency, which is not expected.
>
> Thus, we may need to remove the unneceessary start/finish plug.
Could you please add some data showing the difference in performance?
Thanks,
Song
