[PATCH] efi_loader: Get rid of kaslr-seed

[Ard Biesheuvel]

On Thu, 16 Dec 2021 at 17:56, Mark Kettenis <mark.kettenis at xs4all.nl> wrote:
>
> > From: Ard Biesheuvel <ardb at kernel.org>
> > Date: Thu, 16 Dec 2021 15:54:55 +0100
>
> Hi Ard, Ilias,
>
> > On Thu, 16 Dec 2021 at 15:52, Ilias Apalodimas
> > <ilias.apalodimas at linaro.org> wrote:
> > >
> > > Right now we unconditionally pass a 'kaslr-seed' property to the kernel
> > > if the DTB we ended up in EFI includes the entry.  However the kernel
> > > EFI stub completely ignores it and only relies on EFI_RNG_PROTOCOL.
> > > So let's get rid of it unconditionally since it would mess up the
> > > (upcoming) DTB TPM measuring as well.
> > >
> > > Signed-off-by: Ilias Apalodimas <ilias.apalodimas at linaro.org>
> >
> > Acked-by: Ard Biesheuvel <ardb at kernel.org>
> >
> > Note that the EFI stub itself does not consume the DTB /chosen entry
> > for its own randomness needs (i.e., the randomization of the physical
> > placement of the kernel, which also affects low order virtual
> > placement [bits 16-20]), and will blindly overwrite the seed with
> > whatever the EFI_RNG_PROTOCOL returns.
>
> But it will only do that if EFI_RNG_PROTOCOL is implemented and
> sucessfully returns some random data.  Otherwise "kaslr-seed" will
> survive as-is.  At least that is how I read the code in
> drivers/firmware/efi/libstub/fdt.c:update_fdt().
>
> And this is good.  On Apple M1 systems, the Apple bootloader actually
> provides a block of entropy the the kernel in their version of the
> device tree.  The m1n1 bootloader uses this entropy to populate the
> kaslr-seed property in the device tree it passes on.  And U-Boot is
> used to provide an EFI implementation such that we can boot a wide
> variety of OSes.  At this point we don't know yet whether the SoC
> includes an RNG that we can use to implement EFI_RNG_PROTOCOL in
> U-Boot.
>

Wouldn't it be better to use this block of entropy to seed a DRBG and
subsequently use that as a source of random numbers?

> So the effect of tis change is that a Linux kernel on this platform
> will run without KASLR.  That doesn't seem acceptable to me.
>

I agree that this kind of regression should be avoided. But the
reality is that /chosen/kaslr-seed is Linux internal ABI (EFI
stub<->kernel) that somehow got promoted to boot protocol, in a way
that doesn't even work correctly with the EFI stub itself, since
nobody ever proposed a way to *consume* this kaslr-seed in a way that
permits the EFI stub itself to perform load address randomization when
EFI_RNG_PROTOCOL is absent. Note that randomization of the physical
placement is important not only for physical KASLR but also for
virtual KASLR in Linux. (The virtual placement modulo 2 MiB is decided
by the physical placement directly)

So in my opinion, this is a blatant layering violation, and EFI boot
should be fixed, by implementing EFI_RNG_PROTOCOL in all cases where
u-boot apparently has a source of entropy, as it is able to populate
the kaslr-seed property.

For non-EFI boot, the situation is obviously different, and it's
perfectly fine to use /chosen/kaslr-seed directly for the the parts of
KASLR that can still be used in this case.



> > > ---
> > >  cmd/bootefi.c                 |  2 ++
> > >  include/efi_loader.h          |  2 ++
> > >  lib/efi_loader/efi_dt_fixup.c | 22 ++++++++++++++++++++++
> > >  3 files changed, 26 insertions(+)
> > >
> > > diff --git a/cmd/bootefi.c b/cmd/bootefi.c
> > > index d77d3b6e943d..25f9bfce9b84 100644
> > > --- a/cmd/bootefi.c
> > > +++ b/cmd/bootefi.c
> > > @@ -310,6 +310,8 @@ efi_status_t efi_install_fdt(void *fdt)
> > >         /* Create memory reservations as indicated by the device tree */
> > >         efi_carve_out_dt_rsv(fdt);
> > >
> > > +       efi_purge_kaslr_seed(fdt);
> > > +
> > >         /* Install device tree as UEFI table */
> > >         ret = efi_install_configuration_table(&efi_guid_fdt, fdt);
> > >         if (ret != EFI_SUCCESS) {
> > > diff --git a/include/efi_loader.h b/include/efi_loader.h
> > > index 9dd6c2033634..e560401ac54f 100644
> > > --- a/include/efi_loader.h
> > > +++ b/include/efi_loader.h
> > > @@ -519,6 +519,8 @@ efi_status_t EFIAPI efi_convert_pointer(efi_uintn_t debug_disposition,
> > >                                         void **address);
> > >  /* Carve out DT reserved memory ranges */
> > >  void efi_carve_out_dt_rsv(void *fdt);
> > > +/* Purge unused kaslr-seed */
> > > +void efi_purge_kaslr_seed(void *fdt);
> > >  /* Called by bootefi to make console interface available */
> > >  efi_status_t efi_console_register(void);
> > >  /* Called by bootefi to make all disk storage accessible as EFI objects */
> > > diff --git a/lib/efi_loader/efi_dt_fixup.c b/lib/efi_loader/efi_dt_fixup.c
> > > index b6fe5d2e5a34..02f7de73872e 100644
> > > --- a/lib/efi_loader/efi_dt_fixup.c
> > > +++ b/lib/efi_loader/efi_dt_fixup.c
> > > @@ -40,6 +40,28 @@ static void efi_reserve_memory(u64 addr, u64 size, bool nomap)
> > >                         addr, size);
> > >  }
> > >
> > > +/**
> > > + * efi_remove_kaslr_seed() - Removed unused kaslr-seed
> > > + *
> > > + * Kernel's EFI STUB only relies on EFI_RNG_PROTOCOL for randomization
> > > + * and completely ignores the kaslr-seed.  Weed it out from the DTB we
> > > + * hand over, which would mess up our DTB TPM measurements as well.
> > > + *
> > > + * @fdt: Pointer to device tree
> > > + */
> > > +void efi_purge_kaslr_seed(void *fdt)
> > > +{
> > > +       int nodeoff = fdt_path_offset(fdt, "/chosen");
> > > +       int err = 0;
> > > +
> > > +       if (nodeoff < 0)
> > > +               return;
> > > +
> > > +       err = fdt_delprop(fdt, nodeoff, "kaslr-seed");
> > > +       if (err < 0)
> > > +               log_err("Error deleting kaslr-seed\n");
> > > +}
> > > +
> > >  /**
> > >   * efi_carve_out_dt_rsv() - Carve out DT reserved memory ranges
> > >   *
> > > --
> > > 2.30.2
> > >
> >