On 19 Nov 2012 at 11:37, Maxim Kammerer wrote:

> On Mon, Nov 19, 2012 at 2:25 AM, Matthew Thode
> <[email protected]> wrote:
> > Originally virtualization was slow on grsec/pax with either uderef or
> > kernexec enabled.
> 
> My impression was that UDEREF/KERNEXEC were slow in guest. Is it
> wrong, or did these settings affect host as well?

there was a bug in the per-cpu pgd feature (that those two features rely on
on amd64) that, when enabled on the host, would cause a big guest slowdown
(regardless of what the  guest was).

that these two features have a performance impact on their own is a separate
issue and something i can't help without proper hw support (think SMEP/SMAP).

> > Pipacs overcame this limitation in 3.5.4-r1 and
> > overcame a memory commit issue kvm was having in 3.5.4-r2.  He overcame
> > it using nested page tables on newer CPUs, which means older CPUs will
> > likely still be slow.
> 
> So one needs at least 3.5.4-r2 in both hardened guest and host, and
> nested page tables support in CPU?

for this bug only the host matters and use more like 3.6 please since we no
longer support 3.5 (and in a few weeks that'll become 3.7 ;) or our 2.6.32/3.2
stable series.

nested page tables help with the inherent performance impact of per-cpu pgd
(that is, if you enable it in your guest kernels as well), independently of
the performance bug i fixed some months ago.


Reply via email to