Message-ID: <9f162aae-00fb-dafd-848f-52214836789d@google.com>
Date: Sun, 26 Jan 2025 18:59:04 -0800 (PST)
From: Hugh Dickins <hughd@...gle.com>
To: Johannes Weiner <hannes@...xchg.org>
cc: Hugh Dickins <hughd@...gle.com>, Andrew Morton <akpm@...ux-foundation.org>, 
    Michal Hocko <mhocko@...e.com>, Roman Gushchin <roman.gushchin@...ux.dev>, 
    Shakeel Butt <shakeel.butt@...ux.dev>, Muchun Song <muchun.song@...ux.dev>, 
    linux-mm@...ck.org, cgroups@...r.kernel.org, linux-kernel@...r.kernel.org
Subject: Re: [PATCH] mm: memcontrol: move memsw charge callbacks to v1

On Fri, 24 Jan 2025, Johannes Weiner wrote:
> On Thu, Jan 23, 2025 at 10:53:04PM -0800, Hugh Dickins wrote:
> > On Fri, 24 Jan 2025, Johannes Weiner wrote:
> > 
> > > The interweaving of two entirely different swap accounting strategies
> > > has been one of the more confusing parts of the memcg code. Split out
> > > the v1 code to clarify the implementation and a handful of callsites,
> > > and to avoid building the v1 bits when !CONFIG_MEMCG_V1.
> > > 
> > >    text	  data	   bss	   dec	   hex	filename
> > >   39253	  6446	  4160	 49859	  c2c3	mm/memcontrol.o.old
> > >   38877	  6382	  4160	 49419	  c10b	mm/memcontrol.o
> > > 
> > > Signed-off-by: Johannes Weiner <hannes@...xchg.org>
> > 
> > I'm not really looking at this, but want to chime in that I found the
> > memcg1 swap stuff in mm/memcontrol.c, not in mm/memcontrol-v1.c, very
> > misleading when I was doing the folio_unqueue_deferred_split() business:
> > so, without looking into the details of it, strongly approve of the
> > direction you're taking here - thank you.
> 
> Thanks, I'm glad to hear that!
> 
> > But thought you could go even further, given that
> > static inline bool do_memsw_account(void)
> > {
> > 	return !cgroup_subsys_on_dfl(memory_cgrp_subsys);
> > }
> > 
> > I thought that amounted to do_memsw_account iff memcg_v1;
> > but I never did grasp cgroup_subsys_on_dfl very well,
> > so ignore me if I'm making no sense to you.
> 
> Yes, technically we should be able to move all the code guarded by
> this check to v1 proper in some form.
> 
> [ It's a runtime check for whether the memory controller is attached
>   to a cgroup1 or a cgroup2 mount. You can still mount the v1
>   controller when !CONFIG_MEMCG_V1, but in that case it won't have any
>   memory control files, so whether we update the memsw counter or not,
>   the results of it won't be visible. ]
> 
> But memcg1_swapout()/swapin() are special in that they are completely
> separate, v1-specific memcg entry points. The same is not true for the
> other occurrences:

Information overload!  Thank you for going to the trouble of explaining
those other cases, appreciated, but by "thought you could go even further",
all I had meant was that the do_memsw_account() checks in memcg1_swap*()
looked redundant to me; and possibly some other do_memsw_account() checks.
I'll say no more, I don't want to expose my memcg2 ignorance further,
and I don't deserve another reply.

Hugh

> 
> - mem_cgroup_margin():
> - mem_cgroup_get_max():
> 
> 	The v1 part is about half the function in both cases. We could
> 	split that out into a v1 subfunction, but IMO at a relatively
> 	high cost to the readability of the v1 control flow.
> 
> - drain_stock:
> - try_charge_memcg:
> - uncharge_batch:
> - mem_cgroup_replace_folio:
> - __mem_cgroup_try_charge_swap:
> - __mem_cgroup_uncharge_swap:
> - mem_cgroup_get_nr_swap_pages:
> - mem_cgroup_swap_full:
> 
> 	The majority of the code applies to v2 or both versions, and
> 	the v1 checks either cause an early return or guard the update
> 	to the memsw page_counter.
> 
> 	So not much to farm out code-wise. And the test uses a static
> 	branch, so not much overhead to be cut either.

Powered by blists - more mailing lists