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@xxxxxxxxx> 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@xxxxxxxxxxxxx> 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@xxxxxxxxxxxxx> 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.
