Re: [REGRESSION] 6.7.1: md: raid5 hang and unresponsive system; successfully bisected

From: Song Liu <song@kernel.org>
To: junxiao.bi@oracle.com
Cc: Yu Kuai <yukuai1@huaweicloud.com>,
	 Linux regressions mailing list <regressions@lists.linux.dev>,
	gregkh@linuxfoundation.org,  linux-kernel@vger.kernel.org,
	linux-raid@vger.kernel.org,  stable@vger.kernel.org,
	Dan Moulding <dan@danm.net>, "yukuai (C)" <yukuai3@huawei.com>
Subject: Re: [REGRESSION] 6.7.1: md: raid5 hang and unresponsive system; successfully bisected
Date: Thu, 14 Mar 2024 15:36:38 -0700	[thread overview]
Message-ID: <CAPhsuW7V_n96PDp0wBe5pFNrrkZsbiTQZgZkY+FN7fpJYUmmaw@mail.gmail.com> (raw)
In-Reply-To: <e58155bf-71c1-48c4-94c7-b0d9ecbbef99@oracle.com>

On Thu, Mar 14, 2024 at 11:20 AM <junxiao.bi@oracle.com> wrote:
>
[...]
> >>
> >> This patch eliminated the benefit of blk_plug, i think it will not be
> >> good for IO performance perspective?
> >
> > There is only one daemon thread, so IO should not be handled here as
> > much as possible. The IO should be handled by the thread that is
> > submitting the IO, and let daemon to hanldle the case that IO failed or
> > can't be submitted at that time.

raid5 can have multiple threads calling handle_stripe(). See raid5_do_work().
Only chunk_aligned_read() can be handled in raid5_make_request.

>
> I am not sure how much it will impact regarding drop the blk_plug.
>
> Song, what's your take on this?

I think we need to evaluate the impact of (removing) blk_plug. We had
some performance regressions related to blk_plug a couple years ago.

Thanks,
Song