On Tue, Jul 26, 2011 at 01:58:43AM -0400, Ray Lee wrote:
> > > Are only 1% of mappings 'bad'? 5%? 10%? 50%? We have no idea and the
> > > actual number matters a lot.
> >
> > Actually, it's hard to say what is a 'good' and 'bad' allocation. If
> > they don't alias, they're always good :).
> 
> Count the number of aliases per total mappings, and report that as a
> failure percentage? It doesn't have to be a perfect metric, it just
> needs to be a meaningful and repeatable one.

Ok, I went and installed a Debian wheezy with gnome and started almost
everything I could find in the startup menu. Kernel is 3.0 with Linus'
simpler fix.

The attached file shows all libraries which would create aliases (i.e. 2
or more unique values for bits [14:12] along with the respective count)
in the total of:

Total r-xp mappings: 831, aliasing: 240 (0.289%)

For example, libc gets mapped into all possible slots for [14:12]

Library                                 [14:12] count
=======                                 ======= =====
/lib/libc-2.11.2.so
                                        0       12
                                        1       12
                                        2       11
                                        3       12
                                        4       13
                                        5       16
                                        6       13
                                        7       9


and so on.

So we're close to 30% and could get higher with more apps. I guess I'll
start working on the review comments to the more involved fix.

Thanks.

-- 
Regards/Gruss,
Boris.

Advanced Micro Devices GmbH
Einsteinring 24, 85609 Dornach
GM: Alberto Bozzo
Reg: Dornach, Landkreis Muenchen
HRB Nr. 43632 WEEE Registernr: 129 19551