On 1/24/22 21:10, Dumitru Ceara wrote:
> On 1/24/22 19:40, Adrian Moreno wrote:
>>
>>
>> On 1/24/22 18:49, Jeffrey Walton wrote:
>>> On Mon, Jan 24, 2022 at 9:17 AM Dumitru Ceara <[email protected]> wrote:
>>>>
>>>> As privately reported by Aaron Conole, and by Jeffrey Walton [0]
>>>> there's currently a number of undefined behavior instances in
>>>> the OVS code base.  Running the OVS (and OVN) tests with UBSan [1]
>>>> enabled uncovers these.
>>>> ...
>>>
>>>> Note: depending on the order of reviews, if Adrian's "Fix undefined
>>>> behavior in loop macros" series [2] (or a follow up) is accepted first,
>>>> then patch 12/14 ("util: Avoid false positive UB when iterating
>>>> collections.") can be skipped.  Adrian's series seems to be the more
>>>> correct way of fixing the issue.
>>>
>>> One small nit... UBsan (and Asan) do not produce false positives. They
>>> operate on real data, and when they produce a finding it is valid.
>>> That's also why a complete set of self tests are important. The
>>> complete set of tests are important because UBsan and Asan need real
>>> data.
>>>
>>
>> I agree, it's not a false positive. Furthermore, the patch that Dumitru
>> is referring to ("util: Avoid false positive UB when iterating
>> collections") reduces UBsan's sensitivity by changing some pointer
>> arithmetics to integer arithmetics. This silences the UBsan but
>> according to the discussion with the compiler folks [1], this can still
>> yield UB.
>>
>> Therefore, I think it would be safer to keep the pointer arithmetics
>> (and UBsan's high-sensitivity), fix the actual callers (i.e: the
>> iterator macros), and run UBsan in the CI to spot all future errors
>> (which Dumitru's series does).
>>
>> [1]https://gcc.gnu.org/bugzilla/show_bug.cgi?id=103964
>>
> 
> Yes, I probably should've rephrased the commit title of patch 12/14.
> It's not a false positive.  I just kept it for now until Adrian's series
> [2] is merged.  Otherwise jobs would've failed in CI, and it's
> technically not worse than before.
> 
> But I completely agree, once Adrian's changes get accepted, the safest
> way is to keep the pointer arithmetic and rely on UBSan to complain if
> undefined behavior is detected.
> 
> [2]
> https://patchwork.ozlabs.org/project/openvswitch/list/?series=282481&state=*
> 

For now I applied 7 out of 14 patches from this series and replied with
a comments to a few of the remaining patches.
I still have a couple of ofp patches that I didn't look close enough yet.

Best regards, Ilya Maximets.
_______________________________________________
dev mailing list
[email protected]
https://mail.openvswitch.org/mailman/listinfo/ovs-dev

Reply via email to