Le 01/05/18 à 17:09, Dan Williams a écrit :
> Quoting Mark's original RFC:
>
> "Recently, Google Project Zero discovered several classes of attack
> against speculative execution. One of these, known as variant-1, allows
> explicit bounds checks to be bypassed under speculation, providing an
> arbitrary read gadget. Further details can be found on the GPZ blog [1]
> and the Documentation patch in this series."
>
> This series incorporates Mark Rutland's latest api and adds the x86
> specific implementation of nospec_barrier. The
> nospec_{array_ptr,ptr,barrier} helpers are then combined with a kernel
> wide analysis performed by Elena Reshetova to address static analysis
> reports where speculative execution on a userspace controlled value
> could bypass a bounds check. The patches address a precondition for the
> attack discussed in the Spectre paper [2].
>
> A consideration worth noting for reviewing these patches is to weigh the
> dramatic cost of being wrong about whether a given report is exploitable
> vs the overhead nospec_{array_ptr,ptr} may introduce. In other words,
> lets make the bar for applying these patches be "can you prove that the
> bounds check bypass is *not* exploitable". Consider that the Spectre
> paper reports one example of a speculation window being ~180 cycles.
>
> Note that there is also a proposal from Linus, array_access [3], that
> attempts to quash speculative execution past a bounds check without
> introducing an lfence instruction. That may be a future optimization
> possibility that is compatible with this api, but it would appear to
> need guarantees from the compiler that it is not clear the kernel can
> rely on at this point. It is also not clear that it would be a
> significant performance win vs lfence.
>
> These patches also will also be available via the 'nospec' git branch
> here:
>
> git://git.kernel.org/pub/scm/linux/kernel/git/djbw/linux nospec
Although I suppose -stable and distribution maintainers will keep a
close eye on these patches, is there a particular reason why they don't
include the relevant CVE number in their commit messages?
It sounds like Coverity was used to produce these patches? If so, is
there a plan to have smatch (hey Dan) or other open source static
analysis tool be possibly enhanced to do a similar type of work?
Thanks!
--
Florian
