Re: [PATCH 0/2] mm/page_alloc: Remote per-cpu lists drain support

[Mel Gorman <mgorman@suse.de>]

On Mon, Mar 28, 2022 at 03:51:43PM +0200, Nicolas Saenz Julienne wrote:
> > Now we don't explicitly have this pattern because there isn't an
> > obvious this_cpu_read() for example but it can accidentally happen for
> > counting. __count_zid_vm_events -> __count_vm_events -> raw_cpu_add is
> > an example although a harmless one.
> > 
> > Any of the mod_page_state ones are more problematic though because we
> > lock one PCP but potentially update the per-cpu pcp stats of another CPU
> > of a different PCP that we have not locked and those counters must be
> > accurate.
> 
> But IIUC vmstats don't track pcplist usage (i.e. adding a page into the local
> pcplist doesn't affect the count at all). It is only when interacting with the
> buddy allocator that they get updated. It makes sense for the CPU that
> adds/removes pages from the allocator to do the stat update, regardless of the
> page's journey.
> 

It probably doesn't, I didn't audit it. As I said, it's subtle which is
why I'm wary of relying on accidental safety of getting a per-cpu pointer
that may not be stable. Even if it was ok *now*, I would worry that it
would break in the future. There already has been cases where patches
tried to move vmstats outside the appropriate locking accidentally.

> > It *might* still be safe but it's subtle, it could be easily accidentally
> > broken in the future and it would be hard to detect because it would be
> > very slow corruption of VM counters like NR_FREE_PAGES that must be
> > accurate.
> 
> What does accurate mean here? vmstat consumers don't get accurate data, only
> snapshots.

They are accurate in that they have "Eventual Consistency".
zone_page_state_snapshot exists to get a more accurate count but there is
always some drift but it still is accurate eventually. There is a clear
distinction between VM counters which can be inaccurate they are just to
assist debugging and vmstats like NR_FREE_PAGES that the kernel uses to
make decisions. It potentially gets very problematic if a per-cpu pointer
acquired from one zone gets migrated to another zone and the wrong vmstat
is updated. It *might* still be ok, I haven't audited it but if there is a
possible that two CPUs can be doing a RMW on one per-cpu vmstat structure,
it will corrupt and it'll be difficult to detect.

> And as I comment above you can't infer information about pcplist
> usage from these stats. So, I see no real need for CPU locality when updating
> them (which we're still retaining nonetheless, as per my comment above), the
> only thing that is really needed is atomicity, achieved by disabling IRQs (and
> preemption on RT). And this, even with your solution, is achieved through the
> struct zone's spin_lock (plus a preempt_disable() in RT).
> 

Yes, but under the series I had, I was using local_lock to stabilise what
CPU is being used before acquiring the per-cpu pointer. Strictly speaking,
it doesn't need a local_lock but the local_lock is clearer in terms of
what is being protected and it works with PROVE_LOCKING which already
caught a problematic softirq interaction for me when developing the series.

> All in all, my point is that none of the stats are affected by the change, nor
> have a dependency with the pcplists handling. And if we ever have the need to
> pin vmstat updates to pcplist usage they should share the same pcp structure.
> That said, I'm happy with either solution as long as we get remote pcplist
> draining. So if still unconvinced, let me know how can I help. I have access to
> all sorts of machines to validate perf results, time to review, or even to move
> the series forward.
> 

I also want the remote draining for PREEMPT_RT to avoid interference
of isolated CPUs due to workqueue activity but whatever the solution, I
would be happier if the per-cpu lock is acquired with the CPU stablised
and covers the scope of any vmstat delta updates stored in the per-cpu
structure.  The earliest I will be rebasing my series is 5.18-rc1 as I
see limited value in basing it on 5.17 aiming for a 5.19 merge window.

-- 
Mel Gorman
SUSE Labs