Re: [RFC] should we allow BPF to transmit empty skbs

[Eric Dumazet <eric.dumazet@gmail.com>]

On 10/29/21 8:41 AM, Alexei Starovoitov wrote:
> On Fri, Oct 29, 2021 at 8:39 AM Daniel Borkmann <daniel@iogearbox.net> wrote:
>>
>> On 10/29/21 5:22 PM, Eric Dumazet wrote:
>>> Some layers in tx path do not expect skb being empty (skb->len == 0)
>>>
>>> syzbot reported a crash [1] in fq_codel.
>>>
>>> But I expect many drivers would also crash later.
>>>
>>> Sure the immediate fq_codel crash could be 'fixed', but I would rather
>>> add some sanity checks in net/core/filter.c
>>
>> Makes sense, we shouldn't have to add this to fq_codel fast path, but rather
>> a sanity check for bpf_clone_redirect().
>>
>> I wonder if it's only related to bpf_prog_test_run() infra or if it could also
>> have been generated via stack?
> 
> probably bpf_prog_test_run_skb only.
> I would only add size !=0 check there.
> 

We have a C repro, I will release the syzbot bug so that it can be shared with you.

Thanks.