Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::3:2 as permitted sender) client-ip=2620:137:e000::3:2;
MIME-Version: 1.0
References: <20231115172344.4155593-1-nphamcs@gmail.com> <CAF8kJuN-4UE0skVHvjUzpGefavkLULMonjgkXUZSBVJrcGFXCA@mail.gmail.com>
 <CAJD7tkZ1U+YuvoBAnrXFxQDiQV2hXdbMG-gbzu64R8GLAtNAPA@mail.gmail.com>
 <CAF8kJuPTNwQM413UdeQTkMQ8HkJFyF4OWVrxJSf7uWbege0CXQ@mail.gmail.com>
 <CAKEwX=O5M-vZE5YhYQ5_CbCmXovS1XECO4ROXKWo06K880M1Mg@mail.gmail.com> <CAF8kJuOD6zq2VPcVdoZGvkzYX8iXn1akuYhNDJx-LUdS+Sx3GA@mail.gmail.com>
In-Reply-To: <CAF8kJuOD6zq2VPcVdoZGvkzYX8iXn1akuYhNDJx-LUdS+Sx3GA@mail.gmail.com>
From:   Nhat Pham <nphamcs@gmail.com>
Date:   Sun, 19 Nov 2023 11:08:40 -0800
Message-ID: <CAKEwX=NdFjemcmf27PVpgHpVHWQEo19KfApepWJBRYeyVCWvCw@mail.gmail.com>
Subject: Re: [PATCH v5] zswap: memcontrol: implement zswap writeback disabling
To:     Chris Li <chrisl@kernel.org>
Cc:     Yosry Ahmed <yosryahmed@google.com>,
        Andrew Morton <akpm@linux-foundation.org>, tj@kernel.org,
        lizefan.x@bytedance.com, Johannes Weiner <hannes@cmpxchg.org>,
        Domenico Cerasuolo <cerasuolodomenico@gmail.com>,
        Seth Jennings <sjenning@redhat.com>,
        Dan Streetman <ddstreet@ieee.org>,
        Vitaly Wool <vitaly.wool@konsulko.com>, mhocko@kernel.org,
        roman.gushchin@linux.dev, Shakeel Butt <shakeelb@google.com>,
        muchun.song@linux.dev, Hugh Dickins <hughd@google.com>,
        corbet@lwn.net, Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>,
        senozhatsky@chromium.org, rppt@kernel.org,
        linux-mm <linux-mm@kvack.org>, kernel-team@meta.com,
        LKML <linux-kernel@vger.kernel.org>, linux-doc@vger.kernel.org,
        david@ixit.cz
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable
Precedence: bulk

On Sun, Nov 19, 2023 at 1:39=E2=80=AFAM Chris Li <chrisl@kernel.org> wrote:
>
> On Sat, Nov 18, 2023 at 11:23=E2=80=AFAM Nhat Pham <nphamcs@gmail.com> wr=
ote:
> >
> > Hmm how about this - in the future, we support the following
> > options:
> >
> > 1. zswap.writeback =3D=3D 1: no limitation to zswap writeback.
> > All backing swap devices (sorted by priorities?) are fair game.
> >
> > 2. zswap.writeback =3D=3D 0: disable all forms of zswap writeback.
> >
> > 3. zswap.writeback =3D=3D <tiers description>:  attempt to write to eac=
h
> > tier, one at a time.
>
> We can merge the zswap.writeback as it is for now to unblock you.
>
> For the future. I think we should remove zswap.writeback completely.

I'm a bit weary about API changes, especially changes that affect
backward compatibility. Breaking existing userspace programs simply
with a kernel upgrade does not sound very nice to me.

(although I've heard that the eventual plan is to deprecate cgroupv1
- not sure how that is gonna proceed).

Hence my attempt at creating something that can both serve the
current use case, while still remaining (fairly) extensible for future
ideas.

>
> Instead we have:
>
> swap.tiers =3D=3D <swap_tier_list_name>
> swap.tiers =3D=3D "all" all available swap tiers. "zswap + swap file".
> This is the default.
> swap.tiers =3D=3D "zswap" zswap only, no other swap file. Internally set
> zswap.writeback =3D 0
> swap.tiers =3D=3D "foo" foo is a list of swap devices it can use. You can
> define your town custom swap tier list in
> swap.tiers =3D=3D "none" or "disabled" Not allowed to swap.

swap.tiers =3D=3D "none" or "disabled" means disallowing zswap as
well, correct?

>
> "all", "zswap", "none" are reserved keywords.
> "foo", "bar" etc are custom lists of swap tiers. User define custom
> tier list in sys/kernel/mm/swap/tiers:
> ssd:zswap,/dev/nvme01p4
> hdd:/dev/sda4,/dev/sdb4

I don't have any major argument against this. It just seems a bit
heavyweight for what we need at the moment (only disabling
swap-to-disk usage).

I'll let other people weigh in about this of course.
Johannes, how do you feel about this proposed API?

>
> That would define two custom tiers. "ssd" can use zswap then /dev/nvme01p=
4.
> The exact name of the "swap.tiers" and tiers name are open to suggestions=
.
>
> >
> > The first two are basically what we have for this patch.
> > The last one will be added in a future patch.
> >
> > This is from the userspace perspective. Internally, we can modify
> > memcg->writeback to be a pointer or a struct instead of this bool.
> > (as you suggested).
>
> Internally I would suggest memcg->swaptiers, the write back name is
> somewhat confusing. As your patch indicated. It has two situation:
> 1. shrinking from zpool to real swapfile. The write back is appropriate h=
ere.
>  2. zswap store failed (compression ratio too low, out of memory etc).
> The write back is confusing here. It is more like writing through or
> skip.
>
> >
> > This way, the API remains intact and backward compatible
> > (and FWIW, I think there are still a lot of values in having simple
> > options for the users who have simple memory hierarchies).
>
> swap.tiers can be simple. For example, you can modify your patch to
> "swap.tires =3D=3D zswap" to
> set zswap.writeback bool to 0 for now. Most of your patch is still re-usa=
ble.
>

I'm less concerned about internals - that is always up to changes.
I'm a bit more concerned with the API we're exposing to the users.

> I think we should discuss if we want to keep zswap.writeback in the
> future because that would be some code undeletable and functionally
> overlap with swap.tiers

This is a fair point.

>
> Chris