Re: [PATCH 04/21] Generic percpu refcounting

[Tejun Heo <tj@kernel.org>]

Hey,

On Wed, May 15, 2013 at 02:07:42AM -0700, Kent Overstreet wrote:
> > > +		__this_cpu_dec(*pcpu_count);
> > > +	else
> > > +		ret = atomic_dec_and_test(&ref->count);
> > > +
> > > +	preempt_enable();
> > > +
> > > +	return ret;
> > 
> > With likely() added, I think the compiler should be able to recognize
> > that the branch on pcpu_count should exclude later branch in the
> > caller to test for the final put in most cases but I'm a bit worried
> > whether that would always be the case and wonder whether ->release
> > based interface would be better.  Another concern is that the above
> > interface is likely to encourage its users to put the release
> > implementation in the same function.  e.g.
> 
> I... don't follow what you mean hear at all - what exactly would the
> compiler do differently? and how would passing a release function
> matter?

So, on the fast path, there should be one branch on the percpu
pointer; however, given the above code, especially without likely(),
the compiler may well choose to emit two branches which are shared by
both hot and cold paths - the first one on the percpu pointer, the
second on whether ref->count reached zero.  It just isn't clear to the
compiler whether duplicated preempt_enable() or an extra branch would
be cheaper.

> > 	void my_put(my_obj)
> > 	{
> > 		if (!percpu_ref_put(&my_obj->ref))
> > 			return;
> > 		destroy my_obj;
> > 		free my_obj;
> > 	}
> > 
> > Which in turn is likely to nudge the developer or compiler towards not
> > inlining the fast path.
> 
> I'm kind of skeptical partial inlining would be worth it for just an
> atomic_dec_and_test()...

Ooh, you can do the slow path inline too but I *suspect* we probably
need a bit more logic in the slowpath anyway if we wanna take care of
the bias overflow and maybe the release callback, and it really
doesn't matter a bit whether you have a call for slowpath, so...

> > So, while I do like the simplicity of put() returning %true on the
> > final put, I suspect it's more likely to slowing down fast paths due
> > to its interface compared to having separate ->release function
> > combined with void put().  Any ideas?
> 
> Oh, you mean having one branch instead of two when we're in percpu mode.
> Yeah, that is a good point.

Yeap, heh, I should have read to the end before repling. :)

> I bet with the likely() added the compiler is going to generate the same
> code either way, but I suppose I can have a look at what gcc actually
> does...

Yeah, with likely(), I *think* gcc should get it right most of the
time.  There might be some edge cases tho.

Thanks.

-- 
tejun