Re: [PATCH v3 2/2] random: defer fast pool mixing to worker

[Jason A. Donenfeld]

On Wed, Feb 9, 2022 at 1:36 AM Jason A. Donenfeld <Jason@xxxxxxxxx> wrote:
> On Wed, Feb 9, 2022 at 1:12 AM Eric Biggers <ebiggers@xxxxxxxxxx> wrote:
> > So, add_interrupt_randomness() can execute on the same CPU re-entrantly at any
> > time this is executing?  That could result in some pretty weird behavior, where
> > the pool gets changed half-way through being used, so what is used is neither
> > the old nor the new state of the pool.  Is there a reason why this is okay?
>
> Yes, right, that's the "idea" of this patch, if you could call it such
> a thing. The argument is that we set fast_pool->count to zero *after*
> mixing in the existing bytes + whatever partial bytes might be mixed
> in on an interrupt halfway through the execution of mix_pool_bytes.
> Since we set the count to zero after, it means we do not give any
> credit to those partial bytes for the following set of 64 interrupts.
> What winds up being mixed in will contain at least as much as what was
> there had it not been interrupted. And what gets mixed in the next
> time will only have more mixed in than it otherwise would have, not
> less.

I can actually make it even better by memcpy()ing the fast pool first.
This way any races only affect the fast_mix side -- harmless as
described above -- without affecting blake2s. The latter was _already_
buffering it, but memcpy()ing makes it more clear and doesn't rely on
that behavior. It also means that we get to set count to zero a bit
sooner.