On Wed, 21 Aug 2019 at 08:57, Hsin-Yi Wang <hsinyi@xxxxxxxxxxxx> wrote: > > Then we'd still use add_device_randomness() in case that bootloader > provides weak entropy. > (please don't top post) Whether to trust the firmware provided entropy is a policy decision, and typically, we try to avoid dictating policy in the kernel, and instead, we try to provide a sane default but give the user control over it. So in this case, we should probably introduce add_firmware_randomness() with a Kconfig/cmdline option pair to decide whether it should be trusted or not (or reuse the one we have for trusting RDRAND etc) > On Tue, Aug 20, 2019 at 7:14 PM Ard Biesheuvel > <ard.biesheuvel@xxxxxxxxxx> wrote: > > > > On Tue, 20 Aug 2019 at 10:43, Hsin-Yi Wang <hsinyi@xxxxxxxxxxxx> wrote: > > > > > > Hi Ted, > > > > > > Thanks for raising this question. > > > > > > For UEFI based system, they have a config table that carries rng seed > > > and can be passed to device randomness. However, they also use > > > add_device_randomness (not sure if it's the same reason that they > > > can't guarantee _all_ bootloader can be trusted) > > > > The config table is actually a Linux invention: it is populated by the > > EFI stub code (which is part of the kernel) based on the output of a > > call into the EFI_RNG_PROTOCOL, which is defined in the UEFI spec, but > > optional and not widely available. > > > > I have opted for add_device_randomness() since there is no way to > > establish the quality level of the output of EFI_RNG_PROTOCOL, and so > > it is currently only used to prevent the bootup state of the entropy > > pool to be too predictable, and the output does not contribute to the > > entropy estimate kept by the RNG core. > > > > > > > This patch is to let DT based system also have similar features, which > > > can make initial random number stronger. (We only care initial > > > situation here, since more entropy would be added to kernel as time > > > goes on ) > > > > > > Conservatively, we can use add_device_randomness() as well, which > > > would pass buffer to crng_slow_load() instead of crng_fast_load(). > > > But I think we should trust bootloader here. Whoever wants to use this > > > feature should make sure their bootloader can pass valid (random > > > enough) seeds. If they are not sure, they can just don't add the > > > property to DT. > > > > It is the firmware that adds the property to the DT, not the user.
