Re: [PATCH] random: do not use jump labels before they are initialized

[Ard Biesheuvel <ardb@kernel.org>]

On Tue, 7 Jun 2022 at 13:04, Jason A. Donenfeld <Jason@zx2c4.com> wrote:
>
> Hi Ard,
>
> On Tue, Jun 07, 2022 at 12:56:20PM +0200, Ard Biesheuvel wrote:
> > Could we do this to defer the static key manipulation? That way, the
> > first call to crng_reseed() that occurs after the static keys API
> > becomes available will set the static key, and patch itself away at
> > the same time.
>
> That's almost the same as the patch I just posted, except you
> pushed the logic down into crng_reseed() instead of credit_init_bits().

Sure.

> (A previous mini-project aimed to remove as much logic as possible from
> crng_reseed(), counting on those blocks in crng_init_bits() to only ever
> run once.) What this means is that the static key won't get changed
> until whenever the next reseeding is. I guess that's "fine" but I think
> I'd prefer to keep the entropy counting stuff as separate from the init
> bits stuff as possible.
>

Fair enough. What I would like is to remove the need to play around
with the placement of jump_label_init() across architectures. Jump
labels are fundamentally a performance optimization, so unless you can
explain how setting it as early as possible makes a material
difference, performance or otherwise, I really think we should pursue
a solution that does the static key manipulation at some later time.

> >> As a third, I could just defer doing anything with the bootloader seed
> >> until random_init(). This might actually be the simplest solution...
> >> I'll sketch something out. A downside, which might be sort of
> >> significant, is that a few odd things actually use randomness before
> >> random_init() is called. So these would miss out on having that seed.
> >> I'll have to look what exactly to see if we're actually getting anything
> >> real out of that.
> >>
> >
> > This is kind of the point of using a firmware provided seed, i.e.,
> > that it is available much earlier than anything else.
>
> I'll send a patch for this anyway because I'm sort of curious now. Maybe
> it'll be a dead end, for the reason you mentioned, but I think I'll
> still try to evaluate it.
>

Sure. Anything that can be deferred to an initcall() should be, as the
early arch code is much too fragile to much around with.