Re: [GIT PULL v2] Early SLAB fixes for 2.6.31

[Pekka Enberg <penberg@cs.helsinki.fi>]

On Mon, 2009-06-15 at 10:55 -0400, Christoph Lameter wrote:
> On Sun, 14 Jun 2009, Pekka J Enberg wrote:
> 
> > How about something like this? There should be no extra code in fastpaths
> > for production configs with this one.
> 
> Yes something like that would be good. More comments below.
> 
> > index 4d6004c..5e8cea1 100644
> > --- a/mm/slub.c
> > +++ b/mm/slub.c
> > @@ -1613,6 +1613,8 @@ another_slab:
> >  	deactivate_slab(s, c);
> >
> >  new_slab:
> > +	gfpflags &= slab_gfp_mask;
> > +
> 
> Move the processing of GFP_RECLAIM_MASK etc up to here from new_slab? Then
> the flow is also more logical. The flags handling is concentrated in one
> spot in the allocator and its more obvious how we handle gfp flags.

Sure. Will fix.

> > @@ -1668,13 +1670,14 @@ static __always_inline void *slab_alloc(struct kmem_cache *s,
> >  	struct kmem_cache_cpu *c;
> >  	unsigned long flags;
> >  	unsigned int objsize;
> > +	gfp_t real_gfp;
> >
> > -	gfpflags &= slab_gfp_mask;
> > +	real_gfp = gfpflags & slab_gfp_mask;
> >
> > -	lockdep_trace_alloc(gfpflags);
> > -	might_sleep_if(gfpflags & __GFP_WAIT);
> > +	lockdep_trace_alloc(real_gfp);
> > +	might_sleep_if(real_gfp & __GFP_WAIT);
> >
> > -	if (should_failslab(s->objsize, gfpflags))
> > +	if (should_failslab(s->objsize, real_gfp))
> >  		return NULL;
> >
> >  	local_irq_save(flags);
> 
> Dont do it there. Only modify the slow path.
> 
> Look at __might_sleep(). It already has an exception for system_state !=
> RUNNING. If it still triggers then add to the condition there.

But does this matter? When the debugging options are turned off, there
are no users for "real_gfp" and thus GCC optimizes everything away. For
debugging configs, the extra cacheline load doesn't matter, does it?