Message-ID: <CAJD7tkZF_EW9GC22G-6=j8wVYYS2hTRGRPjwUJfTRr9Ah14Oqg@mail.gmail.com>
Date:   Sat, 9 Jul 2022 02:23:51 -0700
From:   Yosry Ahmed <yosryahmed@...gle.com>
To:     Muchun Song <songmuchun@...edance.com>
Cc:     Andrew Morton <akpm@...ux-foundation.org>,
        Johannes Weiner <hannes@...xchg.org>, longman@...hat.com,
        Michal Hocko <mhocko@...nel.org>,
        Roman Gushchin <roman.gushchin@...ux.dev>,
        Shakeel Butt <shakeelb@...gle.com>,
        Cgroups <cgroups@...r.kernel.org>, duanxiongchun@...edance.com,
        Linux Kernel Mailing List <linux-kernel@...r.kernel.org>,
        Linux-MM <linux-mm@...ck.org>
Subject: Re: [PATCH v6 00/11] Use obj_cgroup APIs to charge the LRU pages

On Fri, Jul 8, 2022 at 10:51 PM Muchun Song <songmuchun@...edance.com> wrote:
>
> On Fri, Jul 08, 2022 at 02:26:08AM -0700, Yosry Ahmed wrote:
> > On Thu, Jul 7, 2022 at 11:52 PM Muchun Song <songmuchun@...edance.com> wrote:
> > >
> > > On Thu, Jul 07, 2022 at 03:14:26PM -0700, Yosry Ahmed wrote:
> > > > On Mon, Jun 27, 2022 at 12:11 AM Muchun Song <songmuchun@...edance.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@...edance.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.
> > > > >
> > > >
> > > > Hey Muchun,
> > > >
> > > > Quick question regarding the behavior of this change on cgroup v1 (I
> > > > know .. I know .. sorry):
> > > >
> > > > When a memcg dies, its LRU pages are reparented, but what happens to
> > > > the charge? IIUC we don't do anything because the pages are already
> > > > hierarchically charged to the parent. Is this correct?
> > > >
> > >
> > > Correct.
> > >
> > > > In cgroup v1, we have non-hierarchical stats as well, so I am trying
> > > > to understand if the reparented memory will appear in the
> > > > non-hierarchical stats of the parent (my understanding is that the
> > > > will not). I am also particularly interested in the charging behavior
> > > > of pages that get reparented to root_mem_cgroup.
> > > >
> > >
> > > I didn't change any memory stats when reparenting.
> > >
> > > > The main reason I am asking is that (hierarchical_usage -
> > > > non-hierarchical_usage - children_hierarchical_usage) is *roughly*
> > > > something that we use, especially at the root level, to estimate
> > > > zombie memory usage. I am trying to see if this change will break such
> > > > calculations. Thanks!
> > > >
> > >
> > > So I think your calculations will still be correct. If you have
> > > any unexpected result, please let me know. Thanks.
> >
> > I have been looking at the code and the patchset and I think there
> > might be a problem with the stats, at least for cgroup v1. Lets say we
> > have a parent memcg P, which has a child memcg C. When processes in
> > memcg C allocate memory the stats (e.g. NR_ANON_MAPPED) are updated
> > for C (non-hierarchical per-cpu counters, memcg->vmstats_percpu), and
> > for P (aggregated stats, memcg->vmstats).
> >
> > When memcg C is offlined, its pages are reparented to memcg P, so far
> > P->vmstats (hierarchical) still have those pages, and
> > P->vmstats_percpu (non-hierarchical) don't. So far so good.
> >
> > Now those reparented pages get uncharged, but their memcg is P now, so
> > they get subtracted from P's *non-hierarchical* stats (and eventually
> > hierarchical stats as well). So now P->vmstats (hierarchical)
> > decreases, which is correct, but P->vmstats_percpu (non-hierarchical)
> > also decreases, which is wrong, as those stats were never added to
> > P->vmstats_percpu to begin with.
> >
> > From a cgroup v2 perspective *maybe* everything continues to work, but
> > this breaks cgroup v1 non-hierarchical stats. In fact, if the
> > reparented memory exceeds the original non-hierarchical memory in P,
> > we can underflow those stats  because we are subtracting stats that
> > were never added in the first place.
> >
> > Please let me know if I am misunderstanding something and there is
> > actually no problem with the non-hierarchical stats (you can stop
> > reading here if this is all in my head and there's actually no
> > problem).
> >
>
> Thanks for patient explanation. Now I got your point.
>
> > Off the top of my mind we can handle stats modifications of reparented
> > memory separately. We should not updated local per-cpu counters, maybe
> > we should rather update memcg->vmstat.state_pending directly so that
> > the changes appear as if they come from a child memcg. Two problems
> > come with such an approach:
> >
>
> Instead of avoiding updating local per-cpu counters for reparented pages,
> after reparenting, how about propagating the child memcg's local per-cpu
> counters to its parent after LRU pages reparenting? And we do not need to
> propagate all vmstats, just some vmstats exposed to cgroup v1 users (like
> memcg1_stats, memcg1_events and lru list pages). I think a reparented page
> is just a little bit of difference compared to other non-reparented pages,
> propagating local per-cpu counters may be acceptable. What do you think?
>

I think this introduces another problem. Now the non-hierarchical
stats of a parent memcg (P in the above example) would include
reparented memory. This hides zombie memory usage. As I elaborated
earlier, parent_hierarchical_usage - parent_non_hierarchical_usage -
SUM(children_hierarchical_usage) should give an estimate of the zombie
memory under parent. If we propagate reparented memory stats (aka
zombies) to the parent's non-hierarchical stats, then we have no way
of finding out how much zombie memory lives in a memcg. This problem
becomes more significant when we are reparenting to root, where zombie
memory is part of unaccounted system overhead.

Actually there is a different problem even in cgroup v2. At root level
there will be no way of finding out whether unaccounted system
overhead (root_usage - SUM(top_level_memcgs_usage)) comes from zombie
memcgs or not, because zombie memcgs will no longer exist and
reparented/zombie memory can be indistinguishable from memory that has
always lived in root. This makes debugging high system overhead even
harder, but that's a problem with the reparenting approach in general,
unrelated to the non-hierarchical stats problem.

> > 1) memcg->vmstat.state_pending is shared between cpus, and so far is
> > only modified by mem_cgroup_css_rstat_flush() in locked context. A
> > solution would be to add reparented state to
> > memcg->vmstat.state_percpu instead and treat it like
> > memcg->vmstat.state_pending in mem_cgroup_css_rstat_flush(). Keep in
> > mind that this adds a tiny bit of memory overhead (roughly 8
> > bytes*num_cpus for each memcg).
> >
> > 2) Identifying that we are updating stats of reparented memory. This
> > should be easy if we have a pointer to the page to compare page->objcg
> > with page->objcg->memcg->objcg, but AFAICT the memcg stats are updated
> > in __mod_memcg_state() and __mod_memcg_lruvec_state(), and we have no
> > idea in each of these what page(s) is the stats update associated
> > with. They are called from many different places, it would be
> > troublesome to pass such information down from all call sites. I have
> > nothing off the top of my head to fix this problem except passing the
> > necessary info through all code paths to __mod_memcg_state() and
> > __mod_memcg_lruvec_state(), which is far from ideal.
> >
> > Again, I am sorry if these discussions are late, I didn't have time to
> > look at previous versions of this patchset.
> >
>
> Not late, thanks for your feedback.
>
> > >
> > > > > > 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
> > > > > thinking, e.g. How to decide if a reparented page needs to be transferred?
> > > > > If we need more information to make this decision, where to store those
> > > > > information? This is my primary thoughts on this question.
> > > > >
> > > > > Thanks.
> > > > >
> > > > > > Thanks again for this work and please excuse my ignorance if any part
> > > > > > of what I said doesn't make sense :)
> > > > > >
> > > > > > >
> > > > > > > ```bash
> > > > > > > #!/bin/bash
> > > > > > >
> > > > > > > dd if=/dev/zero of=temp bs=4096 count=1
> > > > > > > cat /proc/cgroups | grep memory
> > > > > > >
> > > > > > > for i in {0..2000}
> > > > > > > do
> > > > > > >         mkdir /sys/fs/cgroup/memory/test$i
> > > > > > >         echo $$ > /sys/fs/cgroup/memory/test$i/cgroup.procs
> > > > > > >         cat temp >> log
> > > > > > >         echo $$ > /sys/fs/cgroup/memory/cgroup.procs
> > > > > > >         rmdir /sys/fs/cgroup/memory/test$i
> > > > > > > done
> > > > > > >
> > > > > > > cat /proc/cgroups | grep memory
> > > > > > >
> > > > > > > rm -f temp log
> > > > > > > ```
> > > > > > >
> > > > > > > v5: https://lore.kernel.org/all/20220530074919.46352-1-songmuchun@bytedance.com/
> > > > > > > v4: https://lore.kernel.org/all/20220524060551.80037-1-songmuchun@bytedance.com/
> > > > > > > v3: https://lore.kernel.org/all/20220216115132.52602-1-songmuchun@bytedance.com/
> > > > > > > v2: https://lore.kernel.org/all/20210916134748.67712-1-songmuchun@bytedance.com/
> > > > > > > v1: https://lore.kernel.org/all/20210814052519.86679-1-songmuchun@bytedance.com/
> > > > > > > RFC v4: https://lore.kernel.org/all/20210527093336.14895-1-songmuchun@bytedance.com/
> > > > > > > RFC v3: https://lore.kernel.org/all/20210421070059.69361-1-songmuchun@bytedance.com/
> > > > > > > RFC v2: https://lore.kernel.org/all/20210409122959.82264-1-songmuchun@bytedance.com/
> > > > > > > RFC v1: https://lore.kernel.org/all/20210330101531.82752-1-songmuchun@bytedance.com/
> > > > > > >
> > > > > > > v6:
> > > > > > >  - Collect Acked-by and Reviewed-by from Roman and Michal Koutný. Thanks.
> > > > > > >  - Rebase to mm-unstable.
> > > > > > >
> > > > > > > v5:
> > > > > > >  - Lots of improvements from Johannes, Roman and Waiman.
> > > > > > >  - Fix lockdep warning reported by kernel test robot.
> > > > > > >  - Add two new patches to do code cleanup.
> > > > > > >  - Collect Acked-by and Reviewed-by from Johannes and Roman.
> > > > > > >  - I didn't replace local_irq_disable/enable() to local_lock/unlock_irq() since
> > > > > > >    local_lock/unlock_irq() takes an parameter, it needs more thinking to transform
> > > > > > >    it to local_lock.  It could be an improvement in the future.
> > > > > > >
> > > > > > > v4:
> > > > > > >  - Resend and rebased on v5.18.
> > > > > > >
> > > > > > > v3:
> > > > > > >  - Removed the Acked-by tags from Roman since this version is based on
> > > > > > >    the folio relevant.
> > > > > > >
> > > > > > > v2:
> > > > > > >  - Rename obj_cgroup_release_kmem() to obj_cgroup_release_bytes() and the
> > > > > > >    dependencies of CONFIG_MEMCG_KMEM (suggested by Roman, Thanks).
> > > > > > >  - Rebase to linux 5.15-rc1.
> > > > > > >  - Add a new pacth to cleanup mem_cgroup_kmem_disabled().
> > > > > > >
> > > > > > > v1:
> > > > > > >  - Drop RFC tag.
> > > > > > >  - Rebase to linux next-20210811.
> > > > > > >
> > > > > > > RFC v4:
> > > > > > >  - Collect Acked-by from Roman.
> > > > > > >  - Rebase to linux next-20210525.
> > > > > > >  - Rename obj_cgroup_release_uncharge() to obj_cgroup_release_kmem().
> > > > > > >  - Change the patch 1 title to "prepare objcg API for non-kmem usage".
> > > > > > >  - Convert reparent_ops_head to an array in patch 8.
> > > > > > >
> > > > > > > Thanks for Roman's review and suggestions.
> > > > > > >
> > > > > > > RFC v3:
> > > > > > >  - Drop the code cleanup and simplification patches. Gather those patches
> > > > > > >    into a separate series[1].
> > > > > > >  - Rework patch #1 suggested by Johannes.
> > > > > > >
> > > > > > > RFC v2:
> > > > > > >  - Collect Acked-by tags by Johannes. Thanks.
> > > > > > >  - Rework lruvec_holds_page_lru_lock() suggested by Johannes. Thanks.
> > > > > > >  - Fix move_pages_to_lru().
> > > > > > >
> > > > > > > Muchun Song (11):
> > > > > > >   mm: memcontrol: remove dead code and comments
> > > > > > >   mm: rename unlock_page_lruvec{_irq, _irqrestore} to
> > > > > > >     lruvec_unlock{_irq, _irqrestore}
> > > > > > >   mm: memcontrol: prepare objcg API for non-kmem usage
> > > > > > >   mm: memcontrol: make lruvec lock safe when LRU pages are reparented
> > > > > > >   mm: vmscan: rework move_pages_to_lru()
> > > > > > >   mm: thp: make split queue lock safe when LRU pages are reparented
> > > > > > >   mm: memcontrol: make all the callers of {folio,page}_memcg() safe
> > > > > > >   mm: memcontrol: introduce memcg_reparent_ops
> > > > > > >   mm: memcontrol: use obj_cgroup APIs to charge the LRU pages
> > > > > > >   mm: lru: add VM_WARN_ON_ONCE_FOLIO to lru maintenance function
> > > > > > >   mm: lru: use lruvec lock to serialize memcg changes
> > > > > > >
> > > > > > >  fs/buffer.c                      |   4 +-
> > > > > > >  fs/fs-writeback.c                |  23 +-
> > > > > > >  include/linux/memcontrol.h       | 218 +++++++++------
> > > > > > >  include/linux/mm_inline.h        |   6 +
> > > > > > >  include/trace/events/writeback.h |   5 +
> > > > > > >  mm/compaction.c                  |  39 ++-
> > > > > > >  mm/huge_memory.c                 | 153 ++++++++--
> > > > > > >  mm/memcontrol.c                  | 584 +++++++++++++++++++++++++++------------
> > > > > > >  mm/migrate.c                     |   4 +
> > > > > > >  mm/mlock.c                       |   2 +-
> > > > > > >  mm/page_io.c                     |   5 +-
> > > > > > >  mm/swap.c                        |  49 ++--
> > > > > > >  mm/vmscan.c                      |  66 ++---
> > > > > > >  13 files changed, 776 insertions(+), 382 deletions(-)
> > > > > > >
> > > > > > >
> > > > > > > base-commit: 882be1ed6b1b5073fc88552181b99bd2b9c0031f
> > > > > > > --
> > > > > > > 2.11.0
> > > > > > >
> > > > > > >
> > > > > >
> > > >
> >

Powered by blists - more mailing lists