Re: [patch v2 0/5] percpu_counter: bug fix and enhancement

[Shaohua Li <shaohua.li@intel.com>]

On Fri, 2011-05-13 at 14:34 +0800, Eric Dumazet wrote:
> Le vendredi 13 mai 2011 à 13:28 +0800, Shaohua Li a écrit :
> > Hi,
> > On Fri, May 13, 2011 at 01:20:06PM +0800, Eric Dumazet wrote:
> > > Le vendredi 13 mai 2011 à 12:37 +0800, Shaohua Li a écrit :
> > > > On Thu, 2011-05-12 at 17:05 +0800, Tejun Heo wrote:
> > > > > Hello,
> > > > > 
> > > > > On Thu, May 12, 2011 at 11:02:15AM +0200, Eric Dumazet wrote:
> > > > > > > I don't think @maxfuzzy is necessary there.  I wrote this before but
> > > > > > > why can't we track the actual deviation instead of the number of
> > > > > > > deviation events?
> > > > > > 
> > > > > > Thats roughly same thing (BATCH multiplicator factor apart)
> > > > > > 
> > > > > > Most percpu_counter users for a given percpu_counter object use a given
> > > > > > BATCH, dont they ?
> > > > > 
> > > > > Well, @maxfuzzy is much harder than @batch.  It's way less intuitive.
> > > > > Although I haven't really thought about it that much, I think it might
> > > > > be possible to eliminate it.  Maybe I'm confused.  I'll take another
> > > > > look later but if someone can think of something, please jump right
> > > > > in.
> > > > Hmm, looks Eric's approach doesn't work. because we want to remove lock
> > > > in _add, checking seq in _sum still races with _add.
> > > > 
> > > 
> > > Why ?
> > > 
> > > I'll code a patch, I believe it should work.
> > I thought your proposal is:
> > in _add
> > {
> > 	if (count >= batch || count <= -batch) {
> > 		fbc->seq_count++;
> >                atomic64_add(count, &fbc->count);
> > -------->
> >                 __this_cpu_write(*fbc->counters, 0);
> > 	}
> > }
> > 
> > in _sum
> > {
> > restart:
> > 	oldseq = fbc->seqcount;
> > 	smp_rmb();
> > 	do_sum();
> > 	smp_rmb()
> > 	newseq = fbc->seqcount;
> > 	if (newseq - oldseq >= maxfuzzy)
> > 		goto restart;
> > 	return ret;
> > } 
> > if _sum run between above line marked in _add, then the seqcount check
> > doesn't work, we still have deviation Tejun pointed out.
> > 
> 
> I see the point thanks, I'll think a bit more about it.
> 
> We currently serializes both _sum() and _add() with a spinlock.
> 
> My idea was OK if we still kept spinlock in _add(), but this obviously
> is not the need.
> 
> Your goal is letting _add() run without spinlock, but can we agree
> _sum() can run with a spinlock() like today [no more than one instance
> of _sum() running per percpu_counter] ?
locking _sum should be fine