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

[Michal Hocko <mhocko@suse.cz>]

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?

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?

Besides that do we need to have any memory barrier there?

Thanks!
-- 
Michal Hocko
SUSE Labs

--
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>