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.
As I understand things, the code in question is the iterator code for
lists and it was taken from the Linux kernel. The kernel code is not
correct. C language rules are not "opt-in" or opt-out". The kernel has
been getting lucky, and it could break as soon as the compiler's
analysis improves and the compiler starts removing the undefined
behavior. (I've had it happen to me in userland using Intel's ICC
compiler. ICC is ruthless, and it removes all the UB it can find in an
effort to make the code run faster).
Jeff
_______________________________________________
dev mailing list
[email protected]
https://mail.openvswitch.org/mailman/listinfo/ovs-dev
