Date: Thu, 23 May 2024 12:55:46 -0700
From: Roman Gushchin <roman.gushchin@...ux.dev>
To: Kairui Song <ryncsn@...il.com>
Cc: Shakeel Butt <shakeel.butt@...ux.dev>,
	Andrew Morton <akpm@...ux-foundation.org>,
	Muchun Song <muchun.song@...ux.dev>,
	Johannes Weiner <hannes@...xchg.org>,
	Michal Hocko <mhocko@...nel.org>,
	Matthew Wilcox <willy@...radead.org>, linux-mm@...ck.org,
	linux-kernel@...r.kernel.org, gthelen@...gle.coma,
	rientjes@...gle.com, Chris Li <chrisl@...nel.org>
Subject: Re: [PATCH rfc 0/9] mm: memcg: separate legacy cgroup v1 code and
 put under config option

On Thu, May 23, 2024 at 01:58:49AM +0800, Kairui Song wrote:
> On Thu, May 9, 2024 at 2:33 PM Shakeel Butt <shakeel.butt@...ux.dev> wrote:
> >
> > On Wed, May 08, 2024 at 08:41:29PM -0700, Roman Gushchin wrote:
> > > Cgroups v2 have been around for a while and many users have fully adopted them,
> > > so they never use cgroups v1 features and functionality. Yet they have to "pay"
> > > for the cgroup v1 support anyway:
> > > 1) the kernel binary contains useless cgroup v1 code,
> > > 2) some common structures like task_struct and mem_cgroup have never used
> > >    cgroup v1-specific members,
> > > 3) some code paths have additional checks which are not needed.
> > >
> > > Cgroup v1's memory controller has a number of features that are not supported
> > > by cgroup v2 and their implementation is pretty much self contained.
> > > Most notably, these features are: soft limit reclaim, oom handling in userspace,
> > > complicated event notification system, charge migration.
> > >
> > > Cgroup v1-specific code in memcontrol.c is close to 4k lines in size and it's
> > > intervened with generic and cgroup v2-specific code. It's a burden on
> > > developers and maintainers.
> > >
> > > This patchset aims to solve these problems by:
> > > 1) moving cgroup v1-specific memcg code to the new mm/memcontrol-v1.c file,
> > > 2) putting definitions shared by memcontrol.c and memcontrol-v1.c into the
> > >    mm/internal.h header
> > > 3) introducing the CONFIG_MEMCG_V1 config option, turned on by default
> > > 4) making memcontrol-v1.c to compile only if CONFIG_MEMCG_V1 is set
> > > 5) putting unused struct memory_cgroup and task_struct members under
> > >    CONFIG_MEMCG_V1 as well.
> > >
> > > This is an RFC version, which is not 100% polished yet, so but it would be great
> > > to discuss and agree on the overall approach.
> > >
> > > Some open questions, opinions are appreciated:
> > > 1) I consider renaming non-static functions in memcontrol-v1.c to have
> > >    mem_cgroup_v1_ prefix. Is this a good idea?
> > > 2) Do we want to extend it beyond the memory controller? Should
> > > 3) Is it better to use a new include/linux/memcontrol-v1.h instead of
> > >    mm/internal.h? Or mm/memcontrol-v1.h.
> > >
> >
> > Hi Roman,
> >
> > A very timely and important topic and we should definitely talk about it
> > during LSFMM as well. I have been thinking about this problem for quite
> > sometime and I am getting more and more convinced that we should aim to
> > completely deprecate memcg-v1.
> >
> > More specifically:
> >
> > 1. What are the memcg-v1 features which have no alternative in memcg-v2
> > and are blocker for memcg-v1 users? (setting aside the cgroup v2
> > structual restrictions)
> >
> > 2. What are unused memcg-v1 features which we should start deprecating?
> >
> > IMO we should systematically start deprecating memcg-v1 features and
> > start unblocking the users stuck on memcg-v1.
> >
> > Now regarding the proposal in this series, I think it can be a first
> > step but should not give an impression that we are done. The only
> > concern I have is the potential of "out of sight, out of mind" situation
> > with this change but if we keep the momentum of deprecation of memcg-v1
> > it should be fine.
> >
> > I have CCed Greg and David from Google to get their opinion on what
> > memcg-v1 features are blocker for their memcg-v2 migration and if they
> > have concern in deprecation of memcg-v1 features.
> >
> > Anyone else still on memcg-v1, please do provide your input.
> >
> 
> Hi,
> 
> Sorry for joining the discussion late, but I'd like to add some info
> here: We are using the "memsw" feature a lot. It's a very useful knob
> for container memory overcommitting: It's a great abstraction of the
> "expected total memory usage" of a container, so containers can't
> allocate too much memory using SWAP, but still be able to SWAP out.
> 
> For a simple example, with memsw.limit == memory.limit, containers
> can't exceed their original memory limit, even with SWAP enabled, they
> get OOM killed as how they used to, but the host is now able to
> offload cold pages.
> 
> Similar ability seems absent with V2: With memory.swap.max == 0, the
> host can't use SWAP to reclaim container memory at all. But with a
> value larger than that, containers are able to overuse memory, causing
> delayed OOM kill, thrashing, CPU/Memory usage ratio could be heavily
> out of balance, especially with compress SWAP backends.
> 
> Cgroup accounting of ZSWAP/ZRAM doesn't really help, we want to
> account for the total raw usage, not the compressed usage. One example
> is that if a container uses tons of duplicated pages, then it can
> allocate much more memory than it is limited, that could cause
> trouble.

So you don't need separate swap knobs, only combined, right?

> I saw Chris also mentioned Google has a workaround internally for it
> for Cgroup V2. This will be a blocker for us and a similar workaround
> might be needed. It will be great so see an upstream support for this.

I think that _at least_ we should refactor the code so that it would
be a minimal patch (e.g. one #define) to switch to the old mode.

I don't think it's reasonable to add a new interface, but having a
patch/config option or even a mount option which changes the semantics
of memory.swap.max to the v1-like behavior should be ok.

I'll try to do the first part (refactoring this code), and we can have
a discussion from there.

Thanks!

Powered by blists - more mailing lists