lkml.org 
[lkml]   [2014]   [Sep]   [25]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [patch] mm: memcontrol: lockless page counters
On Wed 24-09-14 13:00:17, Johannes Weiner wrote:
> On Wed, Sep 24, 2014 at 04:16:33PM +0200, Michal Hocko wrote:
> > On Tue 23-09-14 13:05:25, Johannes Weiner wrote:
> > [...]
> > > #include <trace/events/vmscan.h>
> > >
> > > -int page_counter_sub(struct page_counter *counter, unsigned long nr_pages)
> > > +/**
> > > + * page_counter_cancel - take pages out of the local counter
> > > + * @counter: counter
> > > + * @nr_pages: number of pages to cancel
> > > + *
> > > + * Returns whether there are remaining pages in the counter.
> > > + */
> > > +int page_counter_cancel(struct page_counter *counter, unsigned long nr_pages)
> > > {
> > > long new;
> > >
> > > new = atomic_long_sub_return(nr_pages, &counter->count);
> > >
> > > - if (WARN_ON(unlikely(new < 0)))
> > > - atomic_long_set(&counter->count, 0);
> > > + if (WARN_ON_ONCE(unlikely(new < 0)))
> > > + atomic_long_add(nr_pages, &counter->count);
> > >
> > > return new > 0;
> > > }
> >
> > I am not sure I understand this correctly.
> >
> > The original res_counter code has protection against < 0 because it used
> > unsigned longs and wanted to protect from really disturbing effects of
> > underflow I guess (this wasn't documented anywhere). But you are using
> > long so even underflow shouldn't be a big problem so why do we need a
> > fixup?
>
> Immediate issues might be bogus numbers showing up in userspace or
> endless looping during reparenting. Negative values are just not
> defined for that counter, so I want to mitigate exposing them.
>
> It's not completely leak-free, as you can see, but I don't think it'd
> be worth weighing down the hot path any more than this just to
> mitigate the unlikely consequences of kernel bug.
>
> > The only way how we can end up < 0 would be a cancel without pairing
> > charge AFAICS. A charge should always appear before uncharge
> > because both of them are using atomics which imply memory barriers
> > (atomic_*_return). So do I understand correctly that your motivation
> > is to fix up those cancel-without-charge automatically? This would
> > definitely ask for a fat comment. Or am I missing something?
>
> This function is also used by the uncharge path, so any imbalance in
> accounting, not just from spurious cancels, is caught that way.
>
> As you said, these are all atomics, so it has nothing to do with
> memory ordering. It's simply catching logical underflows.

OK, I think we should document this in the changelog and/or in the
comment. These things are easy to forget...

Thanks!

--
Michal Hocko
SUSE Labs


\
 
 \ /
  Last update: 2014-09-25 15:21    [W:0.111 / U:3.512 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site