On Fri, Sep 30, 2022 at 08:27:09PM +0200, Ard Biesheuvel wrote:
> On Fri, 30 Sept 2022 at 19:12, Demi Marie Obenour
> <demi@xxxxxxxxxxxxxxxxxxxxxx> wrote:
> >
> > On Fri, Sep 30, 2022 at 06:30:57PM +0200, Ard Biesheuvel wrote:
> > > On Fri, 30 Sept 2022 at 08:44, Jan Beulich <jbeulich@xxxxxxxx> wrote:
> > > >
> > > > On 30.09.2022 01:02, Demi Marie Obenour wrote:
> > > > > Memory of type EFI_CONVENTIONAL_MEMORY, EFI_LOADER_CODE, EFI_LOADER_DATA,
> > > > > EFI_BOOT_SERVICES_CODE, and EFI_BOOT_SERVICES_DATA may be clobbered by
> > > > > Xen before Linux gets to start using it. Therefore, Linux under Xen
> > > > > must not use EFI tables from such memory. Most of the remaining EFI
> > > > > memory types are not suitable for EFI tables, leaving only
> > > > > EFI_ACPI_RECLAIM_MEMORY, EFI_RUNTIME_SERVICES_DATA, and
> > > > > EFI_RUNTIME_SERVICES_CODE. When running under Xen, Linux should only
> > > > > use tables that are located in one of these types of memory.
> > > > >
> > > > > This patch ensures this, and also adds a function
> > > > > (xen_config_table_memory_region_max()) that will be used later to
> > > > > replace the usage of the EFI memory map in esrt.c when running under
> > > > > Xen. This function can also be used in mokvar-table.c and efi-bgrt.c,
> > > > > but I have not implemented this.
> > > > >
> > > > > Signed-off-by: Demi Marie Obenour <demi@xxxxxxxxxxxxxxxxxxxxxx>
> > > >
> > > > In Xen we don't clobber EfiBootServices{Code,Data} when xen.efi was passed
> > > > "-mapbs". Should we perhaps extend the interface such that Dom0 can then
> > > > also use tables located in such regions, perhaps by faking
> > > > EFI_MEMORY_RUNTIME in the attributes returned by XEN_FW_EFI_MEM_INFO?
> > > >
> > >
> > > I know this ship has sailed for x86, but for the sake of other
> > > architectures, I'd strongly recommend leaving the EFI_MEMORY_RUNTIME
> > > bits alone, for the same reasons I gave earlier. (Runtime mappings for
> > > the firmware code itself, page table fragmentation etc etc)
> >
> > Why do you say that it has sailed for x86?
> >
>
> The x86 EFI code in Linux makes changes to the EFI memory map in many
> different places in the code. On other architectures, we have managed
> to avoid this, so that the EFI memory map is always identical to the
> one provided by the firmware at boot.
>
> > > I know very little about Xen, but based on the context you provided in
> > > this thread, I'd say that the best approach from the Xen side is to
> > > convert all EfiBootServicesData regions that have configuration tables
> > > pointing into them into EfiAcpiReclaimMemory.
> >
> > Should Xen convert the entire region, or should it try to reserve only
> > the memory it needs? The latter would require it to parse the
> > configuration tables. Is there a list of configuration tables that can
> > legitimately be in EfiBootServicesData regions?
> >
>
> Not really, no.
Is there a list of tables that can be in EfiBootServicesData and which
Linux cares about?
> So you would have to convert the entire region
> /unless/ Xen knows the GUID, and therefore knows how to derive the
> size of the table, allowing it to reserve memory more conservatively.
My worry is that this will wind up being equivalent to mapping all (or
most) of EfiBootServicesData memory.
> However, I doubt whether this is worth it: splitting entries implies
> rewriting the memory map, which is a thing I'd rather avoid if I were
> in your shoes.
Xen actually uses a different approach: instead of rewriting the memory
map, it uses the EFI pool allocator to allocate memory of the desired
type, copies the table to the newly allocated memory, and installs the
new table in place of the old one. That only works for tables Xen
understands, though.
--
Sincerely,
Demi Marie Obenour (she/her/hers)
Invisible Things Lab
Attachment:
signature.asc
Description: PGP signature
