* Ard Biesheuvel <ard.biesheuvel@xxxxxxxxxx> wrote:
> On Fri, 30 Nov 2018 at 08:29, Ingo Molnar <mingo@xxxxxxxxxx> wrote:
> >
> >
> > * Ard Biesheuvel <ard.biesheuvel@xxxxxxxxxx> wrote:
> >
> > > From: Eric Snowberg <eric.snowberg@xxxxxxxxxx>
> > >
> > > Commit d64934019f6c ("x86/efi: Use efi_exit_boot_services()")
> > > introduced a regression on systems with large memory maps
> > > causing them to hang on boot. The first "goto get_map" that was removed
> > > from exit_boot insured there was enough room for the memory map when
> > > efi_call_early(exit_boot_services) was called. This happens when
> > > (nr_desc > ARRAY_SIZE(params->e820_table).
> > >
> > > Chain of events:
> > > exit_boot()
> > > efi_exit_boot_services()
> > > efi_get_memory_map <- at this point the mm can't grow over 8 desc
> > > priv_func()
> > > exit_boot_func()
> > > allocate_e820ext() <- new mm grows over 8 desc from e820 alloc
> > > efi_call_early(exit_boot_services) <- mm key doesn't match so retry
> > > efi_call_early(get_memory_map) <- not enough room for new mm
> > > system hangs
> > >
> > > This patch allocates the e820 buffer before calling efi_exit_boot_services
> > > and fixes the regression.
> > >
> > > Signed-off-by: Eric Snowberg <eric.snowberg@xxxxxxxxxx>
> > > Signed-off-by: Ard Biesheuvel <ard.biesheuvel@xxxxxxxxxx>
> > > ---
> > > arch/x86/boot/compressed/eboot.c | 65 ++++++++++++++++++++------------
> > > 1 file changed, 41 insertions(+), 24 deletions(-)
> >
> > Any objections against marking this for -stable and filing it in
> > efi/urgent? Boot hangs are show-stopper bugs, so distributions would want
> > to backport this fix anyway.
> >
>
> No objections per se, but this is the kind of patch that might go the
> other way as well, so I would prefer to give it some wider coverage at
> first, given how quickly patches are taken into -stable.
>
> I can make a note of it and send it to Greg halfway into the next -rc cycle.
So there should be at least one week of testing because I just sent the
EFI fixes to Linus, plus -stable gets at least a week of testing as well.
Also, in practice, -next and early -rc cycles get only the fraction of
testing that later -rc's or -stable gets. So if we want this in -stable
we might as well do it now - or the real testing gets delayed by ~3
months in practice. That's also the pattern encouraged by Linus: if it's
a fix that matters then it should be upstreamed with the usual regression
fixes.
Anyway, your call!
Thanks,
Ingo
