Re: [patch] mm: memcontrol: lockless page counters

[Johannes Weiner <hannes@cmpxchg.org>]

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.

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>