On 10/14, Lorenz Bauer wrote:
On Thu, 13 Oct 2022, at 11:44, shaozhengchao wrote:
> Sorry, I haven't fully understood your intentions yet.
> Can you explain it more detail?
I'll try! Roughly, we do the following:
1. Create a BPF_PROG_TYPE_SOCKET_FILTER program that just returns 0
2. Load the program into the kernel
3. Call BPF_PROG_RUN with data_size_in == 14
After your bugfix, it seems like step 3 is rejected due to data_size_in
== 14. We had to increase data_size_in to 15 to
avoid this, see [0].
This breaks user space, so it would be great if you could fix this in a
way that doesn't refuse BPF_PROG_RUN with
[..]
data_size_in == 14. Since I don't understand the original problem very
well I can't tell you what the best fix is however.
The problem was that we were able to generate skb with len=0 via
BPF_PROG_RUN. Prohibiting those cases breaks backwards compatibility, so
we either have to:
a) (preferred?) accept inputs with <14, but maybe internally pad to 14
bytes to make the core stack happy
b) revert the patch and instead have length checks at runtime; doesn't seem
to
be worth the penalty in the forwarding path because of some corner cases
like these ?
0:
https://github.com/cilium/ebpf/commit/a38fb6b5a46ab3b5639ea4d421232a10013596c0
Thanks
Lorenz
