Re: [PATCH] memory-hotplug: fix zone stat mismatch

[Minchan Kim]

On Thu, Sep 20, 2012 at 11:12:40AM +0800, Wen Congyang wrote:
> At 09/20/2012 10:30 AM, Minchan Kim Wrote:
> > On Thu, Sep 20, 2012 at 09:17:43AM +0800, Wen Congyang wrote:
> >> At 09/20/2012 04:17 AM, Minchan Kim Wrote:
> >>> Hi KOSAKI,
> >>>
> >>> On Wed, Sep 19, 2012 at 02:05:20PM -0400, KOSAKI Motohiro wrote:
> >>>> On Wed, Sep 19, 2012 at 3:29 AM, Minchan Kim <minchan@xxxxxxxxxx> wrote:
> >>>>> During memory-hotplug stress test, I found NR_ISOLATED_[ANON|FILE]
> >>>>> are increasing so that kernel are hang out.
> >>>>>
> >>>>> The cause is that when we do memory-hotadd after memory-remove,
> >>>>> __zone_pcp_update clear out zone's ZONE_STAT_ITEMS in setup_pageset
> >>>>> without draining vm_stat_diff of all CPU.
> >>>>>
> >>>>> This patch fixes it.
> >>>>
> >>>> zone_pcp_update() is called from online pages path. but IMHO,
> >>>> the statistics should be drained offline path. isn't it?
> >>>
> >>> It isn't necessary because statistics is right until we reset it to zero
> >>> in online path.
> >>> Do you have something on your mind that we have to drain it in offline path?
> >>
> >> When a node is offlined and onlined again. We create node_data[i] in the
> > 
> > I would like to clarify your word.
> > Create or recreate?
> > Why I have a question is as I look over the source code, hotadd_new_pgdat
> > seem to be called if we do hotadd *new* memory. It's not the case for
> > offline and online again you mentioned. If you're right, I should find 
> > arch_free_nodedata to free pgdat when node is disappear but I can't find it.
> > Do I miss something?
> 
> Hmm, when a memory is removed, we don't do cleanup now. We(Fujitsu) posted
> a patchset to do this:
> https://lkml.org/lkml/2012/9/5/201
> 
> We don't free pgdat in this patchset now. We have two choice:
> 1. free pgdat
> 2. don't free it, and reuse it when it is onlined again
> 
> I'm not sure which choice is better.

I have no idea because I don't know how you guys uses.
If there is use case that sometime you ues many node burstly but
ues a few node in most time, 1) would be good POV memory efficiency
although it makes code rather complicated.

Anyway, it's another story with this patch because it's not merged yet.

> 
> > 
> > 
> >> function hotadd_new_pgdat(), and we will lost the statistics stored in
> >> zone->pageset. So we should drain it in offline path.
> > 
> > Even we drain in offline patch, it still has a problem.
> > 
> > 1. offline
> > 2. drain -> OKAY 
> > 3. schedule
> > 4. Process A increase zone stat
> > 5. Process B increase zone stat
> > 6. online
> > 7. reset it -> we ends up lost zone stat counter which is modified between 2-6
> > 
> 
> I understand why you drain it in online path now. But it still should drain it
> in offline path because if all pages in this zone are offlined, we will call
> zone_pcp_reset() to reset zone's pcp. We should also drop it in the function
> zone_pcp_reset().

Good point.
How about this?