Re: [PATCH] random: fix crash on multiple early calls to add_bootloader_randomness()

[Ard Biesheuvel]

On Wed, 3 Nov 2021 at 08:17, Dominik Brodowski
<linux@xxxxxxxxxxxxxxxxxxxx> wrote:
>
> Am Sun, Oct 31, 2021 at 01:33:34PM +0100 schrieb Ard Biesheuvel:
> > On Sun, 31 Oct 2021 at 07:31, Dominik Brodowski
> > <linux@xxxxxxxxxxxxxxxxxxxx> wrote:
> > >
> > > If add_bootloader_randomness() or add_hwgenerator_randomness() is
> > > called for the first time during early boot, crng_init equals 0. Then,
> > > crng_fast_load() gets called -- which is safe to do even if the input
> > > pool is not yet properly set up.
> > >
> > > If the added entropy suffices to increase crng_init to 1, future calls
> > > to add_bootloader_randomness() or add_hwgenerator_randomness() used to
> > > progress to credit_entropy_bits(). However, if the input pool is not yet
> > > properly set up, the cmpxchg call within that function can lead to an
> > > infinite recursion. This is not only a hypothetical problem, as qemu
> > > on x86 may provide bootloader entropy via EFI and via devicetree.
> > >
> >
> > arm64 not x86
>
> Thanks, fixed in v2
>
> > > As crng_global_init_time is set to != 0 once the input pool is properly
> > > set up, check (also) for this condition to determine which branch to take.
> > >
> > > Calls to crng_fast_load() do not modify the input pool; therefore, the
> > > entropy_count for the input pool must not be modified at that early
> > > stage.
> > >
> > > Reported-and-tested-by: Ivan T. Ivanov <iivanov@xxxxxxx>
> >
> > Nit: fancy tags like this are more difficult to grep for
> >
> > Better to use separate Reported-by and Tested-by tags
>
> Well, it's used not all that rarely, but I don't care that much, so updated for v2.
>
> > Please don't drop the diffstat. Are you using git format-patch?
>
> For singular patches no; but fixed for v2.
>
> > > diff --git a/drivers/char/random.c b/drivers/char/random.c
> > > index 605969ed0f96..4211ff3092f9 100644
> > > --- a/drivers/char/random.c
> > > +++ b/drivers/char/random.c
> > > @@ -1763,8 +1763,8 @@ static void __init init_std_data(struct entropy_store *r)
> > >  }
> > >
> > >  /*
> > > - * Note that setup_arch() may call add_device_randomness()
> > > - * long before we get here. This allows seeding of the pools
> > > + * add_device_randomness() or add_bootloader_randomness() may be
> > > + * called long before we get here. This allows seeding of the pools
> > >   * with some platform dependent data very early in the boot
> > >   * process. But it limits our options here. We must use
> > >   * statically allocated structures that already have all
> > > @@ -2274,7 +2274,12 @@ void add_hwgenerator_randomness(const char *buffer, size_t count,
> > >  {
> > >         struct entropy_store *poolp = &input_pool;
> > >
> > > -       if (unlikely(crng_init == 0)) {
> > > +       /* We cannot do much with the input pool until it is set up in
> > > +        * rand_initalize(); therefore just mix into the crng state.
> > > +        * As this does not affect the input pool, we cannot credit
> > > +        * entropy for this.
> > > +        */
> > > +       if (unlikely(crng_init == 0) || unlikely(crng_global_init_time == 0)) {
> >
> > Can we just drop the unlikely()s here?
>
> As that would be a different change to the one necessary to resolve the bug,
> I'd like to defer that decision to the maintainer of random.c.
>

In that case, can we at least using a single unlikely() for the whole condition?