Re: [PATCH v5 4/4] x86: re-enable rng seeding via SetupData

[Jason A. Donenfeld]

Hi,

I'm currently stumped at the moment, so adding linux-mm@ and x86@. Still
working on it though. Details of where I'm at are below the quote below.

On Sat, Dec 24, 2022 at 05:21:46AM +0100, Jason A. Donenfeld wrote:
> On Sat, Dec 24, 2022 at 04:09:08AM +0100, Jason A. Donenfeld wrote:
> > Hi Eric,
> > 
> > Replying to you from my telephone, and I'm traveling the next two days,
> > but I thought I should mention some preliminary results right away from
> > doing some termux compiles:
> > 
> > On Fri, Dec 23, 2022 at 04:14:00PM -0800, Eric Biggers wrote:
> > > Hi Jason,
> > > 
> > > On Wed, Sep 21, 2022 at 11:31:34AM +0200, Jason A. Donenfeld wrote:
> > > > This reverts 3824e25db1 ("x86: disable rng seeding via setup_data"), but
> > > > for 7.2 rather than 7.1, now that modifying setup_data is safe to do.
> > > > 
> > > > Cc: Laurent Vivier <laurent@vivier.eu>
> > > > Cc: Michael S. Tsirkin <mst@redhat.com>
> > > > Cc: Paolo Bonzini <pbonzini@redhat.com>
> > > > Cc: Peter Maydell <peter.maydell@linaro.org>
> > > > Cc: Philippe Mathieu-Daudé <f4bug@amsat.org>
> > > > Cc: Richard Henderson <richard.henderson@linaro.org>
> > > > Cc: Ard Biesheuvel <ardb@kernel.org>
> > > > Acked-by: Gerd Hoffmann <kraxel@redhat.com>
> > > > Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
> > > > ---
> > > >  hw/i386/microvm.c | 2 +-
> > > >  hw/i386/pc_piix.c | 3 ++-
> > > >  hw/i386/pc_q35.c  | 3 ++-
> > > >  3 files changed, 5 insertions(+), 3 deletions(-)
> > > > 
> > > 
> > > After upgrading to QEMU 7.2, Linux 6.1 no longer boots with some configs. 
> > >  There
> > > is no output at all.  I bisected it to this commit, and I verified that 
> > > the
> > > following change to QEMU's master branch makes the problem go away:
> > > 
> > > diff --git a/hw/i386/pc_piix.c b/hw/i386/pc_piix.c
> > > index b48047f50c..42f5b07d2f 100644
> > > --- a/hw/i386/pc_piix.c
> > > +++ b/hw/i386/pc_piix.c
> > > @@ -441,6 +441,7 @@ static void 
> > > pc_i440fx_8_0_machine_options(MachineClass *m)
> > >      pc_i440fx_machine_options(m);
> > >      m->alias = "pc";
> > >      m->is_default = true;
> > > +    PC_MACHINE_CLASS(m)->legacy_no_rng_seed = true;
> > >  }
> > > 
> > > I've attached the kernel config I am seeing the problem on.
> > > 
> > > For some reason, the problem also goes away if I disable CONFIG_KASAN.
> > > 
> > > Any idea what is causing this?
> > 
> > - Commenting out the call to parse_setup_data() doesn't fix the issue.
> >   So there's no KASAN issue with the actual parser.
> > 
> > - Using KASAN_OUTLINE rather than INLINE does fix the issue!
> > 
> > That makes me suspect that it's file size related, and QEMU or the BIOS
> > is placing setup data at an overlapping offset by accident, or something
> > similar.
> 
> I removed the file systems from your config to bring the kernel size
> back down, and voila, it works, even with KASAN_INLINE. So perhaps I'm
> on the right track here...

QEMU sticks setup_data after the kernel image, the same as kexec-tools
and everything else. Apparently, when the kernel image is large, the
call to early_memremap(boot_params.hdr.setup_data, ...) returns a value
that points some place bogus, and the system crashes or does something
weird. I haven't yet determined what this limit is, but in my current
test kernel, a value of 0x0000000001327650 is enough to make it point to
rubbish.

Is this expected? What's going on here?

Jason