On Thu, 11 Jan 2018, Steven Sistare wrote:
> On 1/11/2018 5:30 PM, Thomas Gleixner wrote:
> > On Thu, 11 Jan 2018, Thomas Gleixner wrote:
> >> On Thu, 11 Jan 2018, Linus Torvalds wrote:
> >>
> >>> On Thu, Jan 11, 2018 at 12:37 PM, Thomas Gleixner <tglx@xxxxxxxxxxxxx> wrote:
> >>>>
> >>>> 67a9108ed431 ("x86/efi: Build our own page table structures")
> >>>>
> >>>> got rid of EFI depending on real_mode_header->trampoline_pgd
> >>>
> >>> So I think it only got rid of by default - the codepath is still
> >>> there, the allocation is still there, it's just that it's not actually
> >>> used unless somebody does that "efi=old_mmap" thing.
> >>
> >> Yes, the trampoline_pgd is still around, but I can't figure out how it
> >> would be used after boot. Confused, digging more.
> >
> > So coming back to the same commit. From the changelog:
> >
> > This is caused by mapping EFI regions with RWX permissions.
> > There isn't much we can do to restrict the permissions for these
> > regions due to the way the firmware toolchains mix code and
> > data, but we can at least isolate these mappings so that they do
> > not appear in the regular kernel page tables.
> >
> > In commit d2f7cbe7b26a ("x86/efi: Runtime services virtual
> > mapping") we started using 'trampoline_pgd' to map the EFI
> > regions because there was an existing identity mapping there
> > which we use during the SetVirtualAddressMap() call and for
> > broken firmware that accesses those addresses.
> >
> > So this very commit gets rid of the (ab)use of trampoline_pgd and allocates
> > efi_pgd, which we made use the proper size.
> >
> > trampoline_pgd is since then only used to get into long mode in
> > realmode/rm/trampoline_64.S and for reboot in machine_real_restart().
> >
> > The runtime services stuff does not use it in kernel versions >= 4.6
> >
> > Thanks,
> >
> > tglx
>
> Yes, and addressing Linus' concern about EFI_OLD_MEMMAP, those paths are
> independent of it. When EFI_OLD_MMAP is enabled, the efi pgd is not
> used, and the bug will not bite.
We have a fix queued in tip/x86/pti which addresses a missing NX clear, but
that's a different story.
Thanks,
tglx