Re: [RFC 08/12] lockdep: Apply crossrelease to PG_locked lock

[Peter Zijlstra]

On Fri, Jul 01, 2016 at 08:21:21AM +0900, Byungchul Park wrote:
> On Thu, Jun 30, 2016 at 03:04:58PM +0200, Peter Zijlstra wrote:
> > On Mon, Jun 20, 2016 at 01:55:23PM +0900, Byungchul Park wrote:
> > > @@ -215,6 +219,11 @@ struct page {
> > >  #ifdef LAST_CPUPID_NOT_IN_PAGE_FLAGS
> > >  	int _last_cpupid;
> > >  #endif
> > > +
> > > +#ifdef CONFIG_LOCKDEP_PAGELOCK
> > > +	struct lockdep_map map;
> > > +	struct cross_lock xlock;
> > > +#endif
> > >  }
> > 
> > So that's 32+64=96 bytes (CONFIG_LOCK_STAT=n) added to struct page,
> > really!?
> 
> Yes... I concerned it at first, but I thought it would be ok since
> CONFIG_LOCKDEP_PAGE is a debug feature.

Right, but still, that's 0.75 GB of memory on my desktop (32GB total)
just for a debug feature. It grows struct page from 1.5% to 3.9% of
total memory, that is immense.

We've avoided doing this for ptl; which was doable because typically
only a small number of pages ends up being a pagetable.

In any case, I feel PG_locked is special enough to fudge. After all, the
content of all these lockdep_map thingies would basically be the same,
which is a massive waste of space.

I still need to bend my brain around this xlock stuff, that just didn't
want to parse when I looked at it last night.