[kernel-hardening] [PATCH 0/7] arm64: Privileged Access Never using TTBR0_EL1 switching
will.deacon at arm.com
Mon Aug 15 04:00:49 PDT 2016
On Mon, Aug 15, 2016 at 12:43:31PM +0200, Ard Biesheuvel wrote:
> On 15 August 2016 at 12:37, Will Deacon <will.deacon at arm.com> wrote:
> > On Mon, Aug 15, 2016 at 12:31:29PM +0200, Ard Biesheuvel wrote:
> >> On 15 August 2016 at 12:30, Will Deacon <will.deacon at arm.com> wrote:
> >> > On Mon, Aug 15, 2016 at 12:21:00PM +0200, Ard Biesheuvel wrote:
> >> >> As to Will's point, I suppose there is a window where a speculative
> >> >> TLB fill could occur, so I suppose that means updating TTBR0_EL1.ASID
> >> >> first, then TCR_EL1.EPD0, and finally perform the TLBI ASIDE1 on the
> >> >> reserved ASID.
> >> >
> >> > But then what do you gain from the reserved ASID?
> >> >
> >> To prevent TLB hits against the ASID of the current (disabled)
> >> userland translation
> > Right, but if the sequence you described ensures that, then why not just
> > set TCR_EL1.EPD0 and do TLBI ASIDE1 on the current ASID?
> ... because then you wipe all the cached translations for current
> userland, which I suppose is best avoided. Wiping the reserved ASID
> only discards TLB entries that should not exist in the first place.
True, I guess we'd need to measure that vs. the extra cost of switching
to/from the reserved ASID.
> > I don't see the difference between a TLB entry formed from a speculative
> > fill using the reserved ASID and one formed using a non-reserved ASID --
> > the page table is the same.
> No, but EPD0 does not disable translations, it disable translation
> table walks on TLB misses, so we need to switch ASIDs to prevent user
> space accesses via TLB hits.
Right, only if you don't do the invalidation for the current ASID.
> But, how about we store the reserved ASID in TTBR1_EL1 instead, and
> switch TCR_EL1.A1 and TCR_EL1.EPD0 in a single write? That way, we can
> switch ASIDs and disable table walks atomically (I hope), and we
> wouldn't need to change TTBR0_EL1 at all.
I doubt that would work, unfortunately. Whilst the writes are atomic
from the point-of-view of the TCR, if the fields can be cached in the
TLB logic, then we can't guarantee the atomicity of changes to the
cached state. Nice idea, though!
More information about the linux-arm-kernel