Re: [tip:efi/core] x86/mm/pat: Use _PAGE_GLOBAL bit for EFI page table mappings

From: Andy Lutomirski
Date: Wed Feb 24 2016 - 14:50:43 EST

On Wed, Feb 24, 2016 at 11:45 AM, Matt Fleming <matt@xxxxxxxxxxxxxxxxxxx> wrote:
> On Wed, 24 Feb, at 10:56:13AM, Linus Torvalds wrote:
>> On Wed, Feb 24, 2016 at 8:36 AM, Andy Lutomirski <luto@xxxxxxxxxxxxxx> wrote:
>> >>
>> >> So the EFI runtime crap should not change once it is mapped. And those
>> >> should be global. It is only natural.
>> >
>> > Why is it natural?
>> >
>> > Long-term, I'd rather see EFI runtime services use an actual mm_struct
>> > and use_mm.
>> Definitely.
>> The EFI runtime page mapping may be unchanging, but that doesn't mean
>> we should be mapping it all the time - the mapping may not change, but
>> we will change away from it.
> There is movement towards hanging the EFI memory map off of mm_struct
> for x86. ARM and arm64 already do this and there were some patches
> from Sylvain (Cc'd) to do this for the purposes of having a task
> context that could be preempted while in the middle of an EFI runtime
> call for some Intel platforms,
> Apart from the code simplification and not being required to open-code
> the %cr3 diddling, are there other benefits of mm_struct and use_mm()
> that make it appealing in the non-preemptible case?
> Not that those aren't reasons enough.

If we add PCID support, then use_mm will get the benefits (~200ns
savings for a round trip) for free.

>> So marking those pages global is very wrong.
> Ingo, Andy, how do you want to handle this patch? Maybe just drop it
> from tip/efi/core while we prod around making all the EFI mappings
> non-global? Nothing else depends on it, it can be dropped without any
> harm.

If the patch is harmless as is, I'm okay with letting it stay.


Andy Lutomirski
AMA Capital Management, LLC