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: <20220621125658.64935-1-songmuchun@bytedance.com>
 <CAJD7tkbeTtkMcVri9vzKHATjxAVZfC_vPNfApJiyRxY53c8pZQ@mail.gmail.com>
 <YrlYG0ecJyzFTVVz@FVFYT0MHHV2J.usts.net> <CAJD7tkbubDhpJV81qkXqRE5-yvvcjKrtPnZnECenafz+to=3_Q@mail.gmail.com>
 <YrmC3FbUAD74lKq/@FVFYT0MHHV2J.usts.net> <YrpYPl1K/k1W30nh@castle>
 <CAJD7tkaOJoZ-8S5RaefDKtD_+aCGVa-RJ3hR23EHOZqT_wf+ew@mail.gmail.com> <Yrpbdk2xVa+B44rK@castle>
In-Reply-To: <Yrpbdk2xVa+B44rK@castle>
From:   Yosry Ahmed <yosryahmed@google.com>
Date:   Mon, 27 Jun 2022 18:45:18 -0700
Message-ID: <CAJD7tkZzXzt5pAjmzawF0dnYLN1AhXPnz5+wWHY+N0qyE3SdQw@mail.gmail.com>
Subject: Re: [PATCH v6 00/11] Use obj_cgroup APIs to charge the LRU pages
To:     Roman Gushchin <roman.gushchin@linux.dev>
Cc:     Muchun Song <songmuchun@bytedance.com>,
        Andrew Morton <akpm@linux-foundation.org>,
        Johannes Weiner <hannes@cmpxchg.org>, longman@redhat.com,
        Michal Hocko <mhocko@kernel.org>,
        Shakeel Butt <shakeelb@google.com>,
        Cgroups <cgroups@vger.kernel.org>, duanxiongchun@bytedance.com,
        Linux Kernel Mailing List <linux-kernel@vger.kernel.org>,
        Linux-MM <linux-mm@kvack.org>
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

On Mon, Jun 27, 2022 at 6:38 PM Roman Gushchin <roman.gushchin@linux.dev> wrote:
>
> On Mon, Jun 27, 2022 at 06:31:14PM -0700, Yosry Ahmed wrote:
> > On Mon, Jun 27, 2022 at 6:24 PM Roman Gushchin <roman.gushchin@linux.dev> wrote:
> > >
> > > On Mon, Jun 27, 2022 at 06:13:48PM +0800, Muchun Song wrote:
> > > > On Mon, Jun 27, 2022 at 01:05:06AM -0700, Yosry Ahmed wrote:
> > > > > On Mon, Jun 27, 2022 at 12:11 AM Muchun Song <songmuchun@bytedance.com> wrote:
> > > > > >
> > > > > > On Sun, Jun 26, 2022 at 03:32:02AM -0700, Yosry Ahmed wrote:
> > > > > > > On Tue, Jun 21, 2022 at 5:57 AM Muchun Song <songmuchun@bytedance.com> wrote:
> > > > > > > >
> > > > > > > > This version is rebased on mm-unstable. Hopefully, Andrew can get this series
> > > > > > > > into mm-unstable which will help to determine whether there is a problem or
> > > > > > > > degradation. I am also doing some benchmark tests in parallel.
> > > > > > > >
> > > > > > > > Since the following patchsets applied. All the kernel memory are charged
> > > > > > > > with the new APIs of obj_cgroup.
> > > > > > > >
> > > > > > > >         commit f2fe7b09a52b ("mm: memcg/slab: charge individual slab objects instead of pages")
> > > > > > > >         commit b4e0b68fbd9d ("mm: memcontrol: use obj_cgroup APIs to charge kmem pages")
> > > > > > > >
> > > > > > > > But user memory allocations (LRU pages) pinning memcgs for a long time -
> > > > > > > > it exists at a larger scale and is causing recurring problems in the real
> > > > > > > > world: page cache doesn't get reclaimed for a long time, or is used by the
> > > > > > > > second, third, fourth, ... instance of the same job that was restarted into
> > > > > > > > a new cgroup every time. Unreclaimable dying cgroups pile up, waste memory,
> > > > > > > > and make page reclaim very inefficient.
> > > > > > > >
> > > > > > > > We can convert LRU pages and most other raw memcg pins to the objcg direction
> > > > > > > > to fix this problem, and then the LRU pages will not pin the memcgs.
> > > > > > > >
> > > > > > > > This patchset aims to make the LRU pages to drop the reference to memory
> > > > > > > > cgroup by using the APIs of obj_cgroup. Finally, we can see that the number
> > > > > > > > of the dying cgroups will not increase if we run the following test script.
> > > > > > >
> > > > > > > This is amazing work!
> > > > > > >
> > > > > > > Sorry if I came late, I didn't follow the threads of previous versions
> > > > > > > so this might be redundant, I just have a couple of questions.
> > > > > > >
> > > > > > > a) If LRU pages keep getting parented until they reach root_mem_cgroup
> > > > > > > (assuming they can), aren't these pages effectively unaccounted at
> > > > > > > this point or leaked? Is there protection against this?
> > > > > > >
> > > > > >
> > > > > > In this case, those pages are accounted in root memcg level. Unfortunately,
> > > > > > there is no mechanism now to transfer a page's memcg from one to another.
> > > > > >
> > > > > > > b) Since moving charged pages between memcgs is now becoming easier by
> > > > > > > using the APIs of obj_cgroup, I wonder if this opens the door for
> > > > > > > future work to transfer charges to memcgs that are actually using
> > > > > > > reparented resources. For example, let's say cgroup A reads a few
> > > > > > > pages into page cache, and then they are no longer used by cgroup A.
> > > > > > > cgroup B, however, is using the same pages that are currently charged
> > > > > > > to cgroup A, so it keeps taxing cgroup A for its use. When cgroup A
> > > > > > > dies, and these pages are reparented to A's parent, can we possibly
> > > > > > > mark these reparented pages (maybe in the page tables somewhere) so
> > > > > > > that next time they get accessed we recharge them to B instead
> > > > > > > (possibly asynchronously)?
> > > > > > > I don't have much experience about page tables but I am pretty sure
> > > > > > > they are loaded so maybe there is no room in PTEs for something like
> > > > > > > this, but I have always wondered about what we can do for this case
> > > > > > > where a cgroup is consistently using memory charged to another cgroup.
> > > > > > > Maybe when this memory is reparented is a good point in time to decide
> > > > > > > to recharge appropriately. It would also fix the reparenty leak to
> > > > > > > root problem (if it even exists).
> > > > > > >
> > > > > >
> > > > > > From my point of view, this is going to be an improvement to the memcg
> > > > > > subsystem in the future.  IIUC, most reparented pages are page cache
> > > > > > pages without be mapped to users. So page tables are not a suitable
> > > > > > place to record this information. However, we already have this information
> > > > > > in struct obj_cgroup and struct mem_cgroup. If a page's obj_cgroup is not
> > > > > > equal to the page's obj_cgroup->memcg->objcg, it means this page have
> > > > > > been reparented. I am thinking if a place where a page is mapped (probably
> > > > > > page fault patch) or page (cache) is written (usually vfs write path)
> > > > > > is suitable to transfer page's memcg from one to another. But need more
> > > > >
> > > > > Very good point about unmapped pages, I missed this. Page tables will
> > > > > do us no good here. Such a change would indeed require careful thought
> > > > > because (like you mentioned) there are multiple points in time where
> > > > > it might be suitable to consider recharging the page (e.g. when the
> > > > > page is mapped). This could be an incremental change though. Right now
> > > > > we have no recharging at all, so maybe we can gradually add recharging
> > > > > to suitable paths.
> > > > >
> > > >
> > > > Agree.
> > > >
> > > > > > thinking, e.g. How to decide if a reparented page needs to be transferred?
> > > > >
> > > > > Maybe if (page's obj_cgroup->memcg == root_mem_cgroup) OR (memcg of
> > > >
> > > > This is a good start.
> > > >
> > > > > current is not a descendant of page's obj_cgroup->memcg) is a good
> > > >
> > > > I am not sure this one since a page could be shared between different
> > > > memcg.
> > >
> > > No way :)
> >
> > No way in terms of charging or usage? AFAIU a page is only charged to
> > one memcg, but can be used by multiple memcgs if it exists in the page
> > cache for example. Am I missing something here?
>
> Charging of course. I mean we can't realistically precisely account for
> shared use of a page between multiple cgroups, at least not at 4k granularity.
>
> >
> > >
> > > >
> > > >     root
> > > >    /   \
> > > >   A     B
> > > >  / \     \
> > > > C   E     D
> > > >
> > > > e.g. a page (originally, it belongs to memcg E and E is dying) is reparented
> > > > to memcg A, and it is shared between C and D now. Then we need to consider
> > > > whether it should be recharged. Yep, we need more thinging about recharging.
> > >
> > > This is why I wasn't sure that objcg-based reparenting is the best approach.
> > > Instead (or maybe even _with_ the reparenting) we can recharge pages on, say,
> > > page activation and/or rotation (inactive->inactive). Pagefaults/reads are
> > > probably to hot to do it there. But the reclaim path should be more accessible
> > > in terms of the performance overhead. Just some ideas.
> >
> > Thanks for chipping in, Roman! I am honestly not sure on what paths
> > the recharge should occur, but I know that we will probably need a
> > recharge mechanism at some point. We can start adding recharging
> > gradually to paths that don't affect performance, reclaim is a very
> > good place. Maybe we sort LRUs such that reparented pages are scanned
> > first, and possibly recharged under memcg pressure.
>
> I think the activation path is a good place to start because we know for sure
> that a page is actively used and we know who is using it.

I agree. What I am suggesting is to additionally scan reparented pages
first under memory pressure. These pages were used by a dead
descendant, so there is a big chance they aren't being used anymore or
they are used by a different memcg, in this case recharging these
pages (if possible) might put the memcg back below its limit. If a
memcg reaches its limit and undergoes reclaim because of reparented
pages it isn't using, this is bad. If during reclaim we keep those
pages and reclaim other pages that are actually being used by the
memcg (even if colder), is arguably worse. WDYT?