Re: [PATCH 6/7] mm: memcontrol: switch to rstat

[Johannes Weiner]

On Fri, Feb 05, 2021 at 04:05:20PM +0100, Michal Hocko wrote:
> On Tue 02-02-21 13:47:45, Johannes Weiner wrote:
> > Replace the memory controller's custom hierarchical stats code with
> > the generic rstat infrastructure provided by the cgroup core.
> > 
> > The current implementation does batched upward propagation from the
> > write side (i.e. as stats change). The per-cpu batches introduce an
> > error, which is multiplied by the number of subgroups in a tree. In
> > systems with many CPUs and sizable cgroup trees, the error can be
> > large enough to confuse users (e.g. 32 batch pages * 32 CPUs * 32
> > subgroups results in an error of up to 128M per stat item). This can
> > entirely swallow allocation bursts inside a workload that the user is
> > expecting to see reflected in the statistics.
> > 
> > In the past, we've done read-side aggregation, where a memory.stat
> > read would have to walk the entire subtree and add up per-cpu
> > counts. This became problematic with lazily-freed cgroups: we could
> > have large subtrees where most cgroups were entirely idle. Hence the
> > switch to change-driven upward propagation. Unfortunately, it needed
> > to trade accuracy for speed due to the write side being so hot.
> > 
> > Rstat combines the best of both worlds: from the write side, it
> > cheaply maintains a queue of cgroups that have pending changes, so
> > that the read side can do selective tree aggregation. This way the
> > reported stats will always be precise and recent as can be, while the
> > aggregation can skip over potentially large numbers of idle cgroups.
> > 
> > This adds a second vmstats to struct mem_cgroup (MEMCG_NR_STAT +
> > NR_VM_EVENT_ITEMS) to track pending subtree deltas during upward
> > aggregation. It removes 3 words from the per-cpu data. It eliminates
> > memcg_exact_page_state(), since memcg_page_state() is now exact.
> 
> The above confused me a bit. I can see the pcp data size increased by
> adding _prev.  The resulting memory footprint should be increased by
> sizeof(long) * (MEMCG_NR_STAT + NR_VM_EVENT_ITEMS) * (CPUS + 1)
> which is roughly 1kB per CPU per memcg unless I have made any
> mistake. This is a quite a lot and it should be mentioned in the
> changelog.

Not quite, you missed a hunk further below in the patch.

Yes, the _prev arrays are added to the percpu struct. HOWEVER, we used
to have TWO percpu structs in a memcg: one for local data, one for
hierarchical data. In the rstat format, one is enough to capture both:

-       /* Legacy local VM stats and events */
-       struct memcg_vmstats_percpu __percpu *vmstats_local;
-
-       /* Subtree VM stats and events (batched updates) */
        struct memcg_vmstats_percpu __percpu *vmstats_percpu;

This eliminates dead duplicates of the nr_page_events and
targets[MEM_CGROUP_NTARGETS(2)] we used to carry, which means we have
a net reduction of 3 longs in the percpu data with this series.

> > Signed-off-by: Johannes Weiner <hannes@xxxxxxxxxxx>
> 
> Although the memory overhead is quite large and it scales both with
> memcg count and CPUs so it can grow quite a bit I do not think this is
> prohibitive. Although it would be really nice if this could be optimized
> in the future.
> 
> All that being said, the code looks more manageable now.
> Acked-by: Michal Hocko <mhocko@xxxxxxxx>

Thanks