Re: [RFC PATCH] mm: memcontrol: don't account swap failures not due to cgroup limits

[Roman Gushchin]

On Tue, Feb 07, 2023 at 11:21:15AM -0800, Yosry Ahmed wrote:
> On Tue, Feb 7, 2023 at 11:09 AM Johannes Weiner <hannes@xxxxxxxxxxx> wrote:
> >
> > On Thu, Feb 02, 2023 at 10:30:40AM -0800, Yosry Ahmed wrote:
> > > On Thu, Feb 2, 2023 at 7:56 AM Johannes Weiner <hannes@xxxxxxxxxxx> wrote:
> > > >
> > > > Christian reports the following situation in a cgroup that doesn't
> > > > have memory.swap.max configured:
> > > >
> > > >   $ cat memory.swap.events
> > > >   high 0
> > > >   max 0
> > > >   fail 6218
> > > >
> > > > Upon closer examination, this is an ARM64 machine that doesn't support
> > > > swapping out THPs. In that case, the first get_swap_page() fails, and
> > > > the kernel falls back to splitting the THP and swapping the 4k
> > > > constituents one by one. /proc/vmstat confirms this with a high rate
> > > > of thp_swpout_fallback events.
> > > >
> > > > While the behavior can ultimately be explained, it's unexpected and
> > > > confusing. I see three choices how to address this:
> > > >
> > > > a) Specifically exlude THP fallbacks from being counted, as the
> > > >    failure is transient and the memory is ultimately swapped.
> > > >
> > > >    Arguably, though, the user would like to know if their cgroup's
> > > >    swap limit is causing high rates of THP splitting during swapout.
> > >
> > > We have the option to add THP_SWPOUT_FALLBACK (and THP_SWPOUT for
> > > completeness) to memcg events for this if/when a use case arises,
> > > right?
> >
> > Yes, we can add that to memory.stat.
> >
> > > > b) Only count cgroup swap events when they are actually due to a
> > > >    cgroup's own limit. Exclude failures that are due to physical swap
> > > >    shortage or other system-level conditions (like !THP_SWAP). Also
> > > >    count them at the level where the limit is configured, which may be
> > > >    above the local cgroup that holds the page-to-be-swapped.
> > > >
> > > >    This is in line with how memory.swap.high, memory.high and
> > > >    memory.max events are counted.
> > > >
> > > >    However, it's a change in documented behavior.
> > >
> > > This option makes sense to me, but I can't speak to the change of
> > > documented behavior. However, looking at the code, it seems like if we do this
> > > the "max" & "fail" counters become effectively the same. "fail" would
> > > not provide much value then.
> >
> > Right.
> >
> > > I wonder if it makes sense to have both, and clarify that "fail" -
> > > "max" would be non-limit based failures (e.g. ran out of swap space),
> > > or would this cause confusion as to whether those non-limit failures
> > > were transient (THP fallback) or eventual?
> >
> > If we add the fallback events, the user can calculate it. I wouldn't
> > split the fail counter itself. There are too many reasons why swap can
> > fail, half of them implementation-defined (as in the ARM example).
> >
> > So I think I'll send patches either way to:
> >
> > 1. Fix the hierarchical accounting of the events to make it consistent
> >    with other counters.
> >
> > 2. Add THP swap/fallback counts to memory.stat
> 
> Sounds good, thanks!
> 
> >
> > We could consider excluding THP fallbacks from the fail count. But it
> > seems really wrong for the cgroup controller to start classifying
> > individual types of failures in the swap layer and make decisions on
> > how to report them to the user. Cgroups really shouldn't be in the
> > business of making up its own MM events. I should provide per-cgroup
> > accounting of native MM events. And nobody has felt the need to add
> > native swap failure counts yet.
> >
> > So I'd argue we should either remove the swap fail count altogether
> > for all the reasons mentioned, or just leave it as is and as a
> > documented interface that is unfortunately out the door.
> >
> > Thoughts?
> 
> Agreed. We should either report all failures or failures specific to
> cgroup limits (which the max count already tracks).
> 
> About removing the fail count completely as-is or completely removing
> it, I don't feel strongly either way. If we add THP swap fallbacks to
> memory.stat, then the fail counter becomes more understandable as you
> can break it down into (max, THP swap fallback, others), with others
> being *probably* swap is full. Arguably, it seems like the interesting
> components (or the cgroup-relevant) components are already there
> regardless. The counter might be useful from a monitoring perspective,
> if a memcg OOMs for example a high fail count can show us that it
> tried to swap multiple times but failed, we can then look at the max
> count and THP fallbacks to understand where the failure is coming
> from.
> 
> I would say we can leave it as-is, but whatever you prefer.

+1

On one hand, the less counters the better.
On the other, removing things which is an API break should have a really good reason.
So probably let's leave it as it is.