Re: [PATCH v1] [mm-unstable] mm: Fix memcg reclaim on memory tiered systems

From: Mina Almasry
Date: Mon Dec 05 2022 - 19:04:24 EST


On Sun, Dec 4, 2022 at 6:39 PM Huang, Ying <ying.huang@xxxxxxxxx> wrote:
>
> Mina Almasry <almasrymina@xxxxxxxxxx> writes:
>
> > commit 3f1509c57b1b ("Revert "mm/vmscan: never demote for memcg
> > reclaim"") enabled demotion in memcg reclaim, which is the right thing
> > to do, however, I suspect it introduced a regression in the behavior of
> > try_to_free_mem_cgroup_pages().
> >
> > The callers of try_to_free_mem_cgroup_pages() expect it to attempt to
> > reclaim - not demote - nr_pages from the cgroup. I.e. the memory usage
> > of the cgroup should reduce by nr_pages. The callers expect
> > try_to_free_mem_cgroup_pages() to also return the number of pages
> > reclaimed, not demoted.
> >
> > However, what try_to_free_mem_cgroup_pages() actually does is it
> > unconditionally counts demoted pages as reclaimed pages. So in practice
> > when it is called it will often demote nr_pages and return the number of
> > demoted pages to the caller. Demoted pages don't lower the memcg usage,
> > and so I think try_to_free_mem_cgroup_pages() is not actually doing what
> > the callers want it to do.
> >
> > I suspect various things work suboptimally on memory systems or don't
> > work at all due to this:
> >
> > - memory.high enforcement likely doesn't work (it just demotes nr_pages
> > instead of lowering the memcg usage by nr_pages).
> > - try_charge_memcg() will keep retrying the charge while
> > try_to_free_mem_cgroup_pages() is just demoting pages and not actually
> > making any room for the charge.
> > - memory.reclaim has a wonky interface. It advertises to the user it
> > reclaims the provided amount but it will actually demote that amount.
> >
> > There may be more effects to this issue.
> >
> > To fix these issues I propose shrink_folio_list() to only count pages
> > demoted from inside of sc->nodemask to outside of sc->nodemask as
> > 'reclaimed'.
> >
> > For callers such as reclaim_high() or try_charge_memcg() that set
> > sc->nodemask to NULL, try_to_free_mem_cgroup_pages() will try to
> > actually reclaim nr_pages and return the number of pages reclaimed. No
> > demoted pages would count towards the nr_pages requirement.
> >
> > For callers such as memory_reclaim() that set sc->nodemask,
> > try_to_free_mem_cgroup_pages() will free nr_pages from that nodemask
> > with either reclaim or demotion.
>
> Have you checked all callers? For example, IIUC, in
> reclaim_clean_pages_from_list(), although sc.nodemask == NULL, the
> demoted pages should be counted as reclaimed.

I checked all call stacks leading to shrink_folio_list() now (at least
I hope). Here is what I think they do and how I propose to handle
them:

- reclaim_clean_pages_from_list() & __node_reclaim() & balance_pgdat()
These try to free memory from a specific node, and both demotion and
reclaim from that node should be counted. I propose these calls set
sc>nodemask = pgdat.node_id to signal to shrink_folio_list() that both
demotion and reclaim from this node should be counted.

- try_to_free_pages()
Tries to free pages from a specific nodemask. It sets sc->nodemask to
ac->nodemask. In this case pages demoted within the nodemask should
not count. Pages demoted outside of the nodemask should count, which
this patch already tries to do.

- mem_cgroup_shrink_node()
This is memcg soft limit reclaim. AFAIU only reclaim should be
counted. It already sets sc->nodemask=NULL to indicate that it
requires reclaim from all nodes and that only reclaimed memory should
be counted, which this patch already tries to do.

- try_to_free_mem_cgroup_pages()
This is covered in the commit message. Many callers set nodemask=NULL
indicating they want reclaim and demotion should not count.
memory.reclaim sets nodemask depending on the 'nodes=' arg and wants
demotion and reclaim from that nodemask.

- reclaim_folio_list()
Sets no_demotion = 1. No ambiguity here, only reclaims and counts
reclaimed pages.

If agreeable I can fix reclaim_clean_pages_from_list() &
__node_reclaim() & balance_pgdat() call sites in v3.

> How about count both
> "demoted" and "reclaimed" in struct scan_control, and let callers to
> determine how to use the number?
>

I don't think this is by itself enough. Pages demoted between 2 nodes
that are both in sc->nodemask should not count, I think. So 'demoted'
needs to be specifically pages demoted outside of the nodemask. We can
do 2 things:

1. Only allow the kernel to demote outside the nodemask (which you
don't prefer).
2. Allow the kernel to demote inside the nodemask but not count them.

I will see if I can implement #2.

> > Tested this change using memory.reclaim interface. With this change,
> >
> > echo "1m" > memory.reclaim
> >
> > Will cause freeing of 1m of memory from the cgroup regardless of the
> > demotions happening inside.
> >
> > echo "1m nodes=0" > memory.reclaim
>
> Have you tested these tests in the original kernel? If so, whether does
> the issue you suspected above occurs during testing?
>

Yes. I set up a test case where I allocate 500m in a cgroup, and then do:

echo "50m" > memory.reclaim

Without my fix, my kernel demotes 70mb and reclaims 4 mb.

With my v1 fix, my kernel demotes all memory possible and reclaims 60mb.

I will add this to the commit message in the next version.


> Best Regards,
> Huang, Ying
>
> > Will cause freeing of 1m of node 0 by demotion if a demotion target is
> > available, and by reclaim if no demotion target is available.
> >
> > Signed-off-by: Mina Almasry <almasrymina@xxxxxxxxxx>
> >
> > ---
> >
> > This is developed on top of mm-unstable largely because I need the
> > memory.reclaim nodes= arg to test it properly.
> > ---
> > mm/vmscan.c | 13 ++++++++++++-
> > 1 file changed, 12 insertions(+), 1 deletion(-)
> >
> > diff --git a/mm/vmscan.c b/mm/vmscan.c
> > index 2b42ac9ad755..8f6e993b870d 100644
> > --- a/mm/vmscan.c
> > +++ b/mm/vmscan.c
> > @@ -1653,6 +1653,7 @@ static unsigned int shrink_folio_list(struct list_head *folio_list,
> > LIST_HEAD(free_folios);
> > LIST_HEAD(demote_folios);
> > unsigned int nr_reclaimed = 0;
> > + unsigned int nr_demoted = 0;
> > unsigned int pgactivate = 0;
> > bool do_demote_pass;
> > struct swap_iocb *plug = NULL;
> > @@ -2085,7 +2086,17 @@ static unsigned int shrink_folio_list(struct list_head *folio_list,
> > /* 'folio_list' is always empty here */
> >
> > /* Migrate folios selected for demotion */
> > - nr_reclaimed += demote_folio_list(&demote_folios, pgdat);
> > + nr_demoted = demote_folio_list(&demote_folios, pgdat);
> > +
> > + /*
> > + * Only count demoted folios as reclaimed if we demoted them from
> > + * inside of the nodemask to outside of the nodemask, hence reclaiming
> > + * pages in the nodemask.
> > + */
> > + if (sc->nodemask && node_isset(pgdat->node_id, *sc->nodemask) &&
> > + !node_isset(next_demotion_node(pgdat->node_id), *sc->nodemask))
> > + nr_reclaimed += nr_demoted;
> > +
> > /* Folios that could not be demoted are still in @demote_folios */
> > if (!list_empty(&demote_folios)) {
> > /* Folios which weren't demoted go back on @folio_list */
> > --
> > 2.39.0.rc0.267.gcb52ba06e7-goog
>