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

[Johannes Weiner]

Hello Michal,

On Thu, Feb 04, 2021 at 03:19:17PM +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.
> 
> I am still digesting details and need to look deeper into how rstat
> works but removing our own stats is definitely a good plan. Especially
> when there are existing limitations and problems that would need fixing.
> 
> Just to check that my high level understanding is correct. The
> transition is effectivelly removing a need to manually sync counters up
> the hierarchy and partially outsources that decision to rstat core. The
> controller is responsible just to tell the core how that syncing is done
> (e.g. which specific counters etc).

Yes, exactly.

rstat implements a tree of cgroups that have local changes pending,
and a flush walk on that tree. But it's all driven by the controller.

memcg needs to tell rstat 1) when stats in a local cgroup change
e.g. when we do mod_memcg_state() (cgroup_rstat_updated), 2) when to
flush, e.g. before a memory.stat read (cgroup_rstat_flush), and 3) how
to flush one cgroup's per-cpu state and propagate it upward to the
parent during rstat's flush walk (.css_rstat_flush).

> Excplicit flushes are needed when you want an exact value (e.g. when
> values are presented to the userspace). I do not see any flushes to
> be done by the core pro-actively except for clean up on a release.
> 
> Is the above correct understanding?

Yes, that's correct.