Re: BUG: KCSAN: data-race in add_device_randomness+0x20d/0x290

[Jann Horn]

On Mon, Feb 7, 2022 at 10:50 PM Jann Horn <jannh@xxxxxxxxxx> wrote:
>
> On Mon, Feb 7, 2022 at 10:45 PM Jann Horn <jannh@xxxxxxxxxx> wrote:
> > On Mon, Feb 7, 2022 at 7:42 PM Jason A. Donenfeld <Jason@xxxxxxxxx> wrote:
> > > Thanks for the report. I assume that this is actually an old bug. Do
> > > you have a vmlinux or a random.o from this kernel you could send me to
> > > double check? Without that, my best guess, which I'd say I have
> > > relatively high confidence about, is that the "1 byte read" is
> > > actually a `movzx eax, cs:lfsr` referring to the `static u8 lfsr`
> > > here, which gets inlined into add_device_randomness:
> > >
> > > static int crng_slow_load(const u8 *cp, size_t len)
> > > {
> > >     unsigned long flags;
> > >     static u8 lfsr = 1;
> > >
> > > This was added in 2008 with dc12baacb95f ("random: use a different
> > > mixing algorithm for add_device_randomness()"). My understanding is
> > > that the race here isn't super problematic as we're in kind of a half
> > > assed "low effort" phase anyway. But I'll give it some thought. I'm
> > > CCing Jann as well who reported the original issue that motivated that
> > > change.
> >
> > But the "lfsr" variable is never accessed outside the part of this
> > method that holds a global spinlock. So that can't really be it,
> > right?
>
> There is a data race in crng_ready(), it just loads from "crng_init"
> without READ_ONCE()... maybe that's what KCSAN is noticing?

(Actually, maybe it should technically even be a smp_load_acquire()?
Or do we have anything else that ensures that we can't get ordering
issues here?)