[MODERATED] Re: Additional sampling fun

On Fri, Feb 28, 2020 at 05:34:47PM +0100, speck for Greg KH wrote:
> On Fri, Feb 28, 2020 at 05:21:40PM +0100, speck for Borislav Petkov wrote:
> > On Thu, Feb 20, 2020 at 09:14:20AM +0100, speck for Greg KH wrote:
> > > Then we need to stop using RDRAND internally for our "give me a random
> > > number api" which has spread to more and more parts of the kernel.
> > > 
> > > Here's a patch that does so:
> > > 	https://lore.kernel.org/lkml/[email protected]/
> > > which I'm going to advise get merged now and backported to the stable
> > > branches.
> 
> Note, the above patch (well the v2 version) is now merged and should
> show up in the next -rc1 release.
> 
> > So one of our guys - Nicolai Stange - was looking at this
> > wrt backporting it to trees and there's another problem in
> > add_interrupt_randomness() which could potentially turn out
> > to be nasty.
> > 
> > We asked him to write it up for speck@ (he's not subscribed) so that we
> > can discuss it here first. Here is the deal in his own words:
> > 
> > "In the context of the get_random_long() patch posted at [1], I noticed
> > that there's also a RDSEED insn issued from the interrupt path, which
> > perhaps might have undesired effects performance-wise.
> > 
> > More specifically, add_interrupt_randomness() would issue one RDSEED
> > either once a second or every 64 interrupts, whichever comes first:
> > 
> >   void add_interrupt_randomness(int irq, int irq_flags)
> >   {
> >     fast_mix(fast_pool); /* increments fast_pool->count */
> >     ...
> >     if ((fast_pool->count < 64) &&
> >          !time_after(now, fast_pool->last + HZ))
> >       return;
> >     ...
> >     fast_pool->last = now;
> >     if (arch_get_random_seed_long(&seed)) {}
> >     ...
> >   }
> > 
> > So while this certainly won't matter much on average, I'm still
> > wondering whether or not this RDSEED could potentially cause IRQ
> > latency spikes relevant e.g. to -RT and/or under high IRQ load?
> > 
> > FWIW, the commit introducing the arch_get_random_seed_long() invocation
> > to add_interrupt_randomness() was commit 83664a6928a4 ("random: Use
> > arch_get_random_seed*() at init time and once a second").
> > 
> > I can only guess, but I think the motivation for mixing
> > arch_get_random_seed_long() from the interrupt path was probably to
> > sync that with the IRQ rate. That is, to make sure that the entropy
> > mixed from RDSEED doesn't dominate the interrupt entropy source.
> > 
> > [1] https://lkml.kernel.org/r/[email protected]
> > "
> 
> Ugh.  I think we need to drag Jason into this as well, but really,
> talking about that can be done on the mailing list as there's nothing
> wrong with trying to get that slow code out of the irq path today,
> right?
> 
> thanks,
> 
> greg k-h
FWIW unless someone is abusing rdrand/rdseed I don't think the impact of the
mitigation will be measurable.  Running multiple instances of spanking rdrand
in a loop will show nonlinear impacts due to bus lock contention but, I don't
think there is any contention issues with once/64IRS's or once a second.  you
are looking at approximately O(100cycles) vrs O(1000cycles) every second or
every 64th interrupt.  I don't think you'll be able to measure the impact of
that.  (unless you force lock contention on the HW bus lock)

--mark