[PATCH v2 1/2] board: sifive: unmatched: use zero copy for initrd

[David Abdurachmanov]

On Tue, Nov 9, 2021 at 7:11 PM Tom Rini <trini at konsulko.com> wrote:
>
> On Tue, Nov 09, 2021 at 05:27:52PM +0100, Heinrich Schuchardt wrote:
> > On 11/9/21 17:07, Tom Rini wrote:
> > > On Tue, Nov 09, 2021 at 04:50:27PM +0100, Heinrich Schuchardt wrote:
> > > > On 11/9/21 16:46, Tom Rini wrote:
> > > > > On Tue, Nov 09, 2021 at 03:46:00PM +0100, Heinrich Schuchardt wrote:
> > > > >
> > > > > > Booting Ubuntu Impish showed the following output:
> > > > > >
> > > > > >       relocaddr   = 0x00000000fff60000
> > > > > >
> > > > > >       Loading Ramdisk to fa118000, end fffff19d ...
> > > > > >
> > > > > > The initrd is overwriting the U-Boot binary. Booting fails.
> > > > > >
> > > > > > There is no need to copy the initrd from $ramdisk_addr_r.
> > > > > > Set init_high = ~0UL to use zero copy. Do the same for the device tree.
> > > > > >
> > > > > > Same for the devicetree.
> > > > > >
> > > > > > Signed-off-by: Heinrich Schuchardt <heinrich.schuchardt at canonical.com>
> > > > > > ---
> > > > > > v2:
> > > > > >       Don't copy fdt either.
> > > > > > ---
> > > > > >    include/configs/sifive-unmatched.h | 2 ++
> > > > > >    1 file changed, 2 insertions(+)
> > > > > >
> > > > > > diff --git a/include/configs/sifive-unmatched.h b/include/configs/sifive-unmatched.h
> > > > > > index f68d7d7676..69a4eb2f2a 100644
> > > > > > --- a/include/configs/sifive-unmatched.h
> > > > > > +++ b/include/configs/sifive-unmatched.h
> > > > > > @@ -59,6 +59,8 @@
> > > > > >       "name=system,size=-,bootable,type=${type_guid_gpt_system};"
> > > > > >    #define CONFIG_EXTRA_ENV_SETTINGS \
> > > > > > +     "fdt_high=0xffffffffffffffff\0" \
> > > > > > +     "initrd_high=0xffffffffffffffff\0" \
> > > > > >       "kernel_addr_r=0x84000000\0" \
> > > > > >       "fdt_addr_r=0x88000000\0" \
> > > > > >       "scriptaddr=0x88100000\0" \
> > > > >
> > > > > I know I'm doing this out of order, but I'm going to repeat myself here
> > > > > too.  You cannot disable device tree relocation.  I cannot count the
> > > > > number of hours that have been wasted because of this mis-feature due to
> > > > > misalignment of the device tree or overwriting of the device tree and
> > > > > then U-Boot not fixing it because it was told not to, and then people
> > > > > and projects wasting countless hours over it.  It's why checkpatch.pl
> > > > > throws out an ERROR on this now.  I didn't yell even more loudly
> > > > > previously at riscv because as it was missing the arch_lmb portion to
> > > > > avoid overwriting U-Boot at run-time, it still was a problem.  But
> > > > > that's been fixed.  So, no.  NAK.
> > > >
> > > > Why should the devicetree relocated?
> > > > This should never have been enabled on RISC-V.
> > >
> > > To repeat myself, because RISC-V has been broken until very recently and
> > > lacked the parts of lmb to avoid overwriting U-Boot while running, is
> > > why any platforms have been allowed in with fdt/initrd_high set to
> > > disable relocation.  As that problem has now been fixed, fdt relocation
> > > must be re-enabled on the currently wrong platforms, and will not be
> > > allowed on new platforms.
> > >
> > > There are specific deployment cases where the developer can choose to
> > > disable relocation because they know that there will never be a way for
> > > things to be done in an overlapping manner because the system is locked
> > > down.  That is very rarely the case for mainline and absolutely not the
> > > case for a general purpose board like the unmatched.
> >
> > __lmb_alloc_base() seems not be integrated with the UEFI sub-system. So UEFI
> > might hand out memory marked as reserved in the LMB sub-system.

Heinrich, do you plan to work on this?

It would be great if we could finally solve this situation with the
2022.01 release.

Right now Unmatched is probably the only board that doesn't turn off
the relocation, and thus some people hit the issues.

Looking at the git history, the LMB issue is fixed in v2022.01-rc1
only right now. Minus UEFI part, which is/will be important for
distributions.

david

> >
> > I guess this is still a topic to be addressed.
>
> If UEFI can still end up getting U-Boot overwritten, yes, that needs to
> be addressed.  Only slightly surprised one of the capture-the-flag or
> similar events hasn't come to us yet with some CVEs related to that,
> too.
>
> --
> Tom