On Wed, Jun 13, 2012 at 03:29:36PM -0400, Rik van Riel wrote:
For one, there are separate kernel boot arguments to control whether
32 and 64 bit processes need to have their addresses aligned for
page colouring.
Do we really need that?
Yes.
Mind you, this is only enabled on AMD F15h - all other x86 simply can't
tweak it without code change.
Would it be a problem if I discarded that code, in order to get to one
common cache colouring implementation?
Sorry, but, we'd like to keep it in.
Secondly, MAP_FIXED never checks for page colouring alignment. I
assume the cache aliasing on AMD Bulldozer is merely a performance
issue, and we can simply ignore page colouring for MAP_FIXED?
Right, AFAICR, MAP_FIXED is not generally used for shared libs (correct
me if I'm wrong here, my memory is very fuzzy about it) and since we see
the perf issue with shared libs, this was fine.
That will be easy to get right in an architecture-independent
implementation.
A third issue is this:
if (!(current->flags& PF_RANDOMIZE))
return addr;
Do we really want to skip page colouring merely because the
application does not have PF_RANDOMIZE set? What is this
conditional supposed to do?
Linus said that without this we are probably breaking old userspace
which can't stomach ASLR so we had to respect such userspace which
clears that flag.