Re: [PATCH net-next 1/1] net_sched: Introduce skbmod action

[Daniel Borkmann <daniel@iogearbox.net>]

On 07/19/2016 08:04 PM, Cong Wang wrote:
> On Tue, Jul 19, 2016 at 8:03 AM, Daniel Borkmann <daniel@iogearbox.net> wrote:
>> On 07/19/2016 03:56 PM, Jamal Hadi Salim wrote:
>> [...]
>>>>
>>>> But apart from this,
>>>> neither pedit nor tcf_skbmod_run() here handle checksum complete, so
>>>> you'll
>>>> potentially get false positives wrt csum corruption and drops as a result
>>>> when using either of the two.
>>>
>>> pedit maybe tricky. Any suggestions?
>>> On tcf_skbmod_run, mostly ignorance: while doing only ethernet updates;
>>> is it still needed to do the checksum complete?
>>
>> Well, what Cong recently fixed with mirred was related to mac header ...
>>
>> You probably need skb_postpull_rcsum(), skb_postpush_rcsum() pair.
>>
>> Also, what about skb_try_make_writable()?
>
> I don't think so. 1) checksum is supposed to be done by csum action
> rather than pedit (or skbmod if it matters), 2) csum action currently
> already handles that correctly for both egress and ingress: 2a)
> CHECKSUM_COMPLETE is meaningless on egress; 2b) it forces
> CHECKSUM_COMPLETE to be CHECKSUM_NONE on ingress
> and it is correctly respected.

Ahh, right, so it redoes entire csums when worst-case changing one byte
only, since it cannot know what other actions like pedit did previously.
But it also means for your l2 mac addr changes to force a CHECKSUM_COMPLETE
into CHECKSUM_NONE with csum action that you need to call into one of the
l3/l4 helpers as far as I see.