[PATCH 3/3] KVM: PPC: Book3S: Add support for hwrng found on some powernv systems

[Paul Mackerras]

On Thu, Oct 03, 2013 at 08:48:03AM +0300, Gleb Natapov wrote:
> On Thu, Oct 03, 2013 at 08:45:42AM +1000, Paul Mackerras wrote:
> > On Wed, Oct 02, 2013 at 04:36:05PM +0200, Alexander Graf wrote:
> > > 
> > > On 02.10.2013, at 16:33, Paolo Bonzini wrote:
> > > 
> > > > Il 02/10/2013 16:08, Alexander Graf ha scritto:
> > > >>> The hwrng is accessible by host userspace via /dev/mem.
> > > >> 
> > > >> A guest should live on the same permission level as a user space
> > > >> application. If you run QEMU as UID 1000 without access to /dev/mem, why
> > > >> should the guest suddenly be able to directly access a memory location
> > > >> (MMIO) it couldn't access directly through a normal user space interface.
> > > >> 
> > > >> It's basically a layering violation.
> > > > 
> > > > With Michael's earlier patch in this series, the hwrng is accessible by
> > > > host userspace via /dev/hwrng, no?
> > > 
> > > Yes, but there's not token from user space that gets passed into the kernel to check whether access is ok or not. So while QEMU may not have permission to open /dev/hwrng it could spawn a guest that opens it, drains all entropy out of it and thus stall other processes which try to fetch entropy, no?
> > 
> > Even if you drain all entropy out of it, wait 64 microseconds and it
> > will be full again. :)  Basically it produces 64 bits every
> > microsecond and puts that in a 64 entry x 64-bit FIFO buffer, which is
> > what is read by the MMIO.  So there is no danger of stalling other
> > processes for any significant amount of time.
> > 
> Even if user crates 100s guests each one of which reads hwrng in a loop?

Well, you can't actually have more guests running than there are cores
in a system.  POWER7+ has one RNG per chip and 8 cores per chip, each
of which can run 4 threads (which have to be in the same guest).

Michael's code uses the RNG on the same chip.  Worst case therefore is
32 threads accessing the same RNG, so a given thread might have to
wait up to 32 microseconds for its data.

Paul.