Re: [PATCH 14/18] 2.6.17.9 perfmon2 patch for review: new i386 files

[Stephane Eranian <eranian@hpl.hp.com>]

Andi,

On Tue, Sep 12, 2006 at 07:04:53AM -0700, Stephane Eranian wrote:
> Andi,
> 
> On Fri, Aug 25, 2006 at 03:13:58PM +0200, Andi Kleen wrote:
> 
> > > Are we already running with cr4.pce set today?
> > 
> > Not yet. But soon.
> >  
> > > The cr4.pce allows all PMC (counter) to be read at user level, not just perfctr0.
> > > When enabled all counters are readable at user level from any process. A process
> > > can see the value accumulated by another process (assuming monitoring in per-thread
> > > mode).
> > 
> > Yes, we'll have to live with that.
> > 
> > > Some people may see this as a security risk.
> > 
> > Maybe some paranoiacs, but we normally don't design software for these people's
> > obsessions.
> > 
> > > On the other hand all you see  
> > > is counts.
> > 
> > Exactly. And you always have RDTSC anyways.
> > 
> > 
> > > So as long as the i386/x86_64 PMU only collect counts, this could be 
> > > fine. The day they can capture addresses, this becomes more problematic, I think.
> > 
> > We can worry about it when it happens. Whenever anyone adds that capability
> > to the hardware they will hopefully add new separate ring 3 control bits.
> > 
> Just a follow-up on this. It already exists.
> 
> On the P4, I am planning on exporting the Last Brnch Record Stack (LBR Stack) to users of perfmon.
> This provides some branch buffer similar to what we have on Itanium. Obviously, the MSRs do contains
> addresses (source/target of branches).
> 
I did some more checking on this.
I think we are fine, because you need to use RDPMC to access from userland when cr4.pce=1 and it
gives access ONLY to MSR that contain counters, not all performance monitoring related MSRs.

-- 
-Stephane