Re: [RESEND PATCH 1/2 -mm] mm: account lazy free pages separately

[Yang Shi]

On 8/14/19 4:08 AM, Michal Hocko wrote:
> On Mon 12-08-19 10:00:17, Yang Shi wrote:
> > On 8/12/19 2:34 AM, Michal Hocko wrote:
> > > On Fri 09-08-19 16:54:43, Yang Shi wrote:
> > > > On 8/9/19 11:26 AM, Yang Shi wrote:
> > > > > On 8/9/19 11:02 AM, Michal Hocko wrote:
> > > [...]
> > > > > > I have to study the code some more but is there any reason why those
> > > > > > pages are not accounted as proper THPs anymore? Sure they are partially
> > > > > > unmaped but they are still THPs so why cannot we keep them accounted
> > > > > > like that. Having a new counter to reflect that sounds like papering
> > > > > > over the problem to me. But as I've said I might be missing something
> > > > > > important here.
> > > > > I think we could keep those pages accounted for NR_ANON_THPS since they
> > > > > are still THP although they are unmapped as you mentioned if we just
> > > > > want to fix the improper accounting.
> > > > By double checking what NR_ANON_THPS really means,
> > > > Documentation/filesystems/proc.txt says "Non-file backed huge pages mapped
> > > > into userspace page tables". Then it makes some sense to dec NR_ANON_THPS
> > > > when removing rmap even though they are still THPs.
> > > >
> > > > I don't think we would like to change the definition, if so a new counter
> > > > may make more sense.
> > > Yes, changing NR_ANON_THPS semantic sounds like a bad idea. Let
> > > me try whether I understand the problem. So we have some THP in
> > > limbo waiting for them to be split and unmapped parts to be freed,
> > > right? I can see that page_remove_anon_compound_rmap does correctly
> > > decrement NR_ANON_MAPPED for sub pages that are no longer mapped by
> > > anybody. LRU pages seem to be accounted properly as well.  As you've
> > > said NR_ANON_THPS reflects the number of THPs mapped and that should be
> > > reflecting the reality already IIUC.
> > >
> > > So the only problem seems to be that deferred THP might aggregate a lot
> > > of immediately freeable memory (if none of the subpages are mapped) and
> > > that can confuse MemAvailable because it doesn't know about the fact.
> > > Has an skewed counter resulted in a user observable behavior/failures?
> > No. But the skewed counter may make big difference for a big scale cluster.
> > The MemAvailable is an important factor for cluster scheduler to determine
> > the capacity.
> But MemAvailable is a very rough estimation. Is relying on it really a
> good measure? I mean there is a lot of reclaimable memory that is not
> reflected there (some fs. internal data structures, networking buffers
> etc.)

Yes, I agree there are other freeable objects not accounted into 
MemAvailable. Their size depends on the workload. But, deferred split 
THPs seems more common with the common workloads. A simple run with 
MariaDB test of mmtest shows it could generate over fifteen thousand 
deferred split THPs (accumulated around 30G in one hour run, 75% of 40G 
memory for my VM). So, it may be worth accounting deferred split THPs in 
MemAvailable.

> [...]
>
> > > accounting the full THP correct? What if subpages are still mapped?
> > "Deferred split" definitely doesn't mean they are free. When memory pressure
> > is hit, they would be split, then the unmapped normal pages would be freed.
> > So, when calculating MemAvailable, they are not accounted 100%, but like
> > "available += lazyfree - min(lazyfree / 2, wmark_low)", just like how page
> > cache is accounted.
> Then this is even more dubious IMHO.
>
> > We could get more accurate account, i.e. checking each sub page's mapcount
> > when accounting, but it may change before shrinker start scanning. So, just
> > use the ballpark estimation to trade off the complexity for accurate
> > accounting.
> I do not see much point in fixing up one particular counter when there
> is a whole lot that is even not considered. I would rather live with the
> fact that MemAvailable is only very rough estimate then whack a mole on
> any memory consumer that is freeable directly or indirectly via memory
> reclaim. Because this is likely to be always subtly broken and only
> visible under very specific workloads so there is no way to test for it.

I saw Vlastimil suggested KReclaimable, it seems a good fit. If so we 
don't need create a new counter anymore.