On Mon, Sep 16, 2019 at 07:25:48AM -0700, Dave Hansen wrote:
> On 9/16/19 2:00 AM, Kirill A. Shutemov wrote:
> >>> I think we also need to make it clear that this is workaround for a broken
> >>> hardware: speculative execution must not trigger a halt.
> >> I think the word broken is a bit loaded here. According to the UEFI
> >> spec (version 2.8, page 167), "Regions that are backed by the physical
> >> hardware, but are not supposed to be accessed by the OS, must be
> >> returned as EfiReservedMemoryType." Our interpretation is that
> >> includes speculative accesses.
> > +Dave.
> >
> > I don't think it is. Speculative access is done by hardware, not OS.
> >
> > BTW, isn't it a BIOS issue?
> >
> > I believe it should have a way to hide a range of physical address space
> > from OS or force a caching mode on to exclude it from speculative
> > execution. Like setup MTRRs or something.
>
> Ugh. I bet that was a fun one to chase down. Have the hardware
> engineers learned a lesson or are they hiding behind the EFI spec in an
> act of pure cowardice? ;)
Yes, it was fun. My main BIOS contact has explained to me how they are
stuck between a rock and a hard place on any other options for this.
> The patch is small and fixes a real problem. The changelog is OK,
> although I'd prefer some differentiation between "occupied by the
> kernel" and the kernel *image*.
OK, is the phrase "kernel image" generally understood to cover
everything from _text to _end, including the bss? As long as that's
true, I will adopt this phrase.
> The code is also gloriously free of any comments about what it's
> doing or why.
I'm intending to add something like this in the next version:
/*
* Only the region occupied by the kernel image has so far been checked against
* the table of usable memory regions provided by the firmware, so
* invalidate pages outside that region. A page table entry that maps to
* a reserved area of memory would allow processor speculation into that
* area, and on some hardware (particularly the UV platform) speculation
* into reserved areas can cause a system halt.
*/
> But, I'm left with lots of questions:
>
> Why do PMD-level changes fix this? Is it because we 2MB pad the kernel
> image? Why can't we still get within 2MB of the memory address in
> question?
This fix works for our hardware because the problematic reserved
regions are 64M aligned, and going up to the next 2MB boundary from
_end is not going to cross the next 64M boundary.
One could argue the next step would be going into
boot/compressed/{kaslr.c, misc.c} and rounding the size of the kernel
up to the next 2MB boundary to ensure the chosen random location is
covered by usable RAM up to the next PMD-level boundary. I did not go
there because for us it is not necessary.
> Is it in the lower 1MB, by chance?
No, this is a reserved range at the top physical addresses for each
NUMA node in a collection of them.
> If this is all about avoiding EFI reserved ranges, why doesn't the
> patch *LOOK* At EFI reserved ranges?
Because the range the kernel image is located in is already checked
against them in boot/compressed/kaslr.c. This will now be explained
in the comment I mention above, which you had not yet seen.
--> Steve Wahl
--
Steve Wahl, Hewlett Packard Enterprise
