>>> This implements subset of function for codes which was removed by the
>>> commit
>>> I mentioned below.
>>
>> No, as I explain above, it performs a different check.
>>
>>>>>
>
> [...]
Clearly I was wrong, sorry. Thanks for pointing out that commit and
576a30eb6453 ("[NET]: Added GSO header verification").
>>>>> For performance reason. I think we should delay the check or
>>>>> segmentation
>>>>> as
>>>>> much as possible until it was really needed.
>>>>
>>>> Going through segmentation is probably as expensive as flow dissector,
>>>> if not more so because of the indirect branches.
>>>
>>> I think we don't even need to care about this consider the evil packet
>>> should be rare.
>>
>> How does frequency matter when a single packet can crash a host?
>
>
> I mean consider we had fix the crash, we don't care how expensive do we spot
> this.
>
>>
>>> And what you propose here is just a very small subset of the
>>> necessary checking, more comes at gso header checking. So even if we care
>>> performance, it only help for some specific case.
>>
>> It also fixed the bug that Eric sent a separate patch for, as that did
>> not dissect as a valid TCP packet, either.
>
>
> I may miss something but how did this patch protects an evil thoff?
Actually, it blocked that specific reproducer because the ip protocol
did not match.
I think that __skb_flow_dissect_tcp should return a boolean, causing
dissection return FLOW_DISSECT_RET_OUT_BAD if the tcph is bad.
That would be needed to really catch it with flow dissection at the source.
