Alexei,
On 12/22/18 10:12 PM, Alexei Starovoitov wrote:
On Sat, Dec 22, 2018 at 09:37:02PM -0600, Gustavo A. R. Silva wrote:
Can't we have the case in which the code can be "trained" to read
perfectly valid values for prog->len for quite a while, making the
microcode come into place and speculate about:
1013 if (flen == 0 || flen > BPF_MAXINSNS)
1014 return false;
and then make flen to be greater than BPF_MAXINSNS?
Yes. The user space can train line 1013 to mispredict by passing
smaller flen N times and then passing large flen.
Why do you think it's exploitable?
Without the patch in the mispredicted path the cpu will do
if (0 < flen) condition and since flen is hot in the cache
it will happily execute the filter[0] load...
and about 12-20 u-ops later (depending on u-arch of cpu) when
branch predictor realizes that it's a miss, the cpu will ignore
the values computed in the shadow cpu registers used by speculative execution
and go back to the 'return false' execution path.
The side effect of bringing filter[0] value in L1 cache is still there.
The cpu is incapable to undo that cache load. That's what spectre1 is about.
Do you see how filter[0] value in cpu L1 cache is exploitable?
In this regard, the policy has been to kill the speculation on that
first load (as I mentioned in my last email. It is also mentioned in
the commit log for every patch).
I took another look at the following patches:
"net: core: Fix Spectre v1 vulnerability"
"nfc: af_nfc: Fix Spectre v1 vulnerability"
"can: af_can: Fix Spectre v1 vulnerability"
and I have to say that none of them are necessary.
I'm not sure whether there were other patches that pretend to fix spectre1.
It's not about pretending to fix it. It's about trying to prevent the
conditions that can actually trigger the exploitation of a potential
vulnerability. The code is not always the same, it changes, it evolves,
and we are currently trying to catch that first load (that's what smatch
does in all these cases) that could eventually lead to an actual vuln.
Thanks
--
Gustavo
