Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
MIME-Version: 1.0
References: <20220620173843.1462198-1-daeho43@gmail.com> <20220620173843.1462198-2-daeho43@gmail.com>
 <YrEXmiRH+DcbW2EI@B-P7TQMD6M-0146.local>
In-Reply-To: <YrEXmiRH+DcbW2EI@B-P7TQMD6M-0146.local>
From:   Daeho Jeong <daeho43@gmail.com>
Date:   Tue, 21 Jun 2022 09:10:13 -0700
Message-ID: <CACOAw_zBi_TnOmqLZwBF9Zdd=0m7EkRP9M210BOrabh3L=7bqQ@mail.gmail.com>
Subject: Re: [f2fs-dev] [PATCH 2/2] f2fs: handle decompress only post
 processing in softirq
To:     Gao Xiang <hsiangkao@linux.alibaba.com>
Cc:     linux-kernel@vger.kernel.org,
        linux-f2fs-devel@lists.sourceforge.net, kernel-team@android.com,
        Daeho Jeong <daehojeong@google.com>
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

On Mon, Jun 20, 2022 at 5:58 PM Gao Xiang <hsiangkao@linux.alibaba.com> wrote:
>
> On Mon, Jun 20, 2022 at 10:38:43AM -0700, Daeho Jeong wrote:
> > From: Daeho Jeong <daehojeong@google.com>
> >
> > Now decompression is being handled in workqueue and it makes read I/O
> > latency non-deterministic, because of the non-deterministic scheduling
> > nature of workqueues. So, I made it handled in softirq context only if
> > possible, not in low memory devices, since this modification will
> > maintain decompresion related memory a little longer.
> >
>
> Again, I don't think this method scalable.  Since you already handle
> all decompression algorithms in this way.  Laterly, I wonder if you'd
> like to handle all:
>  - decompression algorithms;
>  - verity algorithms;
>  - decrypt algorithms;
>
> in this way, regardless of slow decompression algorithms, that would be a
> long latency and CPU overhead of softirq context.  This is my last words
> on this, I will not talk anymore.
>
> Thanks,
> Gao Xiang

I understand what you're worried about. However, verity cannot be
inlined because of its nature triggering I/Os. Except that, other twos
are almost inducing overhead comparable to memcpy. Still, decrypt part
is done in a H/W way mostly these days, though.

Thanks,