Re: [PATCH] random32: avoid attempt to late reseed if in the middle of seeding

[Daniel Borkmann]

On 03/27/2014 10:18 AM, Hannes Frederic Sowa wrote:
> On Thu, Mar 27, 2014 at 10:04:03AM +0100, Daniel Borkmann wrote:
> > On 03/27/2014 03:21 AM, Hannes Frederic Sowa wrote:
> > > On Wed, Mar 26, 2014 at 07:35:01PM -0400, Sasha Levin wrote:
> > > > On 03/26/2014 07:18 PM, Daniel Borkmann wrote:
> > > > > On 03/26/2014 06:12 PM, Sasha Levin wrote:
> > > > > > Commit 4af712e8df ("random32: add prandom_reseed_late() and call when
> > > > > > nonblocking pool becomes initialized") has added a late reseed stage
> > > > > > that happens as soon as the nonblocking pool is marked as initialized.
> > > > > >
> > > > > > This fails in the case that the nonblocking pool gets initialized
> > > > > > during __prandom_reseed()'s call to get_random_bytes(). In that case
> > > > > > we'd double back into __prandom_reseed() in an attempt to do a late
> > > > > > reseed - deadlocking on 'lock' early on in the boot process.
> > > > > >
> > > > > > Instead, just avoid even waiting to do a reseed if a reseed is already
> > > > > > occuring.
> > > > > >
> > > > > > Signed-off-by: Sasha Levin <sasha.levin@xxxxxxxxxx>
> > > > >
> > > > > Thanks for catching! (If you want Dave to pick it up, please also
> > > > > Cc netdev.)
> > > > >
> > > > > Why not via spin_trylock_irqsave() ? Thus, if we already hold the
> > > > > lock, we do not bother any longer with doing the same work twice
> > > > > and just return.
> > >
> > > I totally agree with Daniel spin_trylock_irqsave seems like the best
> > > solution.
> > >
> > > In case we really want to make sure that even early seeding doesn't
> > > race with late seed and the pool is only filled by another CPU, we would
> > > actually need per-cpu bools to get this case correct.
> >
> > But then again, we would just exit via spin_trylock_irqsave()
> > now, no? Whenever something enters this section protected under
> > irq save spinlock we would do a reseed of the entire state (s1-s4)
> > for each cpu.
>
> If early reseed races with late one, we would actually need to spin on
> maybe another cpu, so the early call can leave critical section before
> late call enters. If we don't spin we could possibly miss the late call
> when nonblocking pool is fully seeded (entropy may be added in batches
> and first cpus of the early reseeding might miss better entropy).
>
> If the early call blocks the late call, maybe even on another cpu, the late
> call should spin until the early call left the critical section. We can only
> deadlock on same cpu.
>
> I consider this just hypothetical.

I think I agree with you, it seems hypothetical. We perhaps could have
tested e.g. if the nonblocking pool is initialized and in case not,
would have a fallback to get_random_bytes_arch() for that short time
to have presumably better 'entropy'. On the other hand, if the _arch()
version would have been backdoored by the NSA, as we possibly reseed
all states later on, or just s1 and xor it with all other states, it
seems that eventually there's still no real gain for doing that. So
the spin_trylock_irqsave() version seems okay ... plus prandom() API
is not cryptographically secure anyway.
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/