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: <20220613155612.402297-1-daeho43@gmail.com> <Yqge0XS7jbSnNWvq@sol.localdomain>
 <YqhRBZMYPp/kyxoe@B-P7TQMD6M-0146.local>
In-Reply-To: <YqhRBZMYPp/kyxoe@B-P7TQMD6M-0146.local>
From:   Daeho Jeong <daeho43@gmail.com>
Date:   Tue, 14 Jun 2022 09:46:50 -0700
Message-ID: <CACOAw_wjCyTmwusY6S4+NgMuLOZm9fwGfrvCT272GJ01-RP6PQ@mail.gmail.com>
Subject: Re: [f2fs-dev] [PATCH] f2fs: handle decompress only post processing
 in softirq
To:     Gao Xiang <hsiangkao@linux.alibaba.com>
Cc:     Eric Biggers <ebiggers@kernel.org>,
        Daeho Jeong <daehojeong@google.com>,
        Nathan Huckleberry <nhuck@google.com>, kernel-team@android.com,
        linux-kernel@vger.kernel.org,
        linux-f2fs-devel@lists.sourceforge.net
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

>
> Some my own previous thoughts about this strategy:
>
>  - If we allocate all memory and map these before I/Os, all inflight I/Os
>    will keep such temporary pages all the time until decompression is
>    finished. In contrast, if we allocate or reuse such pages just before
>    decompression, it would minimize the memory footprints.
>
>    I think it will impact the memory numbers at least on the very
>    low-ended devices with bslow storage. (I've seen f2fs has some big
>    mempool already)
>
>  - Many compression algorithms are not suitable in the softirq contexts,
>    also I vaguely remembered if softirq context lasts for > 2ms, it will
>    push into ksoftirqd instead so it's actually another process context.
>    And it may delay other important interrupt handling.
>
>  - Go back to the non-deterministic scheduling of workqueues. I guess it
>    may be just due to scheduling punishment due to a lot of CPU consuming
>    due to decompression before so the priority becomes low, but that is
>    just a pure guess. May be we need to use RT scheduling policy instead.
>
>    At least with WQ_HIGHPRI for dm-verity at least, but I don't find
>    WQ_HIGHPRI mark for dm-verity.
>
> Thanks,
> Gao Xiang

I totally understand what you are worried about. However, in the real
world, non-determinism from workqueues is more harsh than we expected.
As you know, reading I/Os in the system are critical paths most of the
time and now I/O variations with workqueue are too bad.

I also think it's better that we have RT scheduling like things here.
We could think about it more.

Thanks,