Re: [PATCH net-next 1/4] flow_dissector: Get vlan info from skb->vlan_tci instead of skb->data

[Jiri Pirko <jiri@resnulli.us>]

Mon, Aug 15, 2016 at 05:51:38PM CEST, hadarh@dev.mellanox.co.il wrote:
>On Mon, Aug 15, 2016 at 5:38 AM, Toshiaki Makita
><toshiaki.makita1@gmail.com> wrote:
>> On 16/08/14 (日) 23:58, Hadar Hen Zion wrote:
>>>
>>> On Fri, Aug 12, 2016 at 9:36 AM, Toshiaki Makita
>>> <makita.toshiaki@lab.ntt.co.jp> wrote:
>>>>
>>>> On 2016/08/10 22:32, Hadar Hen Zion wrote:
>>>>>
>>>>> Early in the datapath skb_vlan_untag function is called, stripped
>>>>> the vlan from the skb and set skb->vlan_tci and skb->vlan_proto fields.
>>>>>
>>>>> The current dissection doesn't handle vlan packets correctly.  Vlan
>>>>> doesn't exist in skb->data anymore when applying flow dissection on the
>>>>> skb, fix that.
>>>>
>>>>
>>>> RPS (and flow-dissector called in RPS) is performed before vlan-strip in
>>>> __netif_receive_skb_core().
>>>
>>>
>>> right, I'll fix it to v2.
>>>
>>>> Also, in cases skb is tagged with multiple vlan headers (typical when
>>>> using 802.1ad), the second level vlan tag is in skb->data.
>>>
>>>
>>> Currently, flow_dissector doesn't support multiple vlan headers, only
>>> one vlan_id field is present.
>>> There aren't any flow_dissector "customers" yet for multiple vlan support.
>>
>>
>> Sure, no need to store second level vlan tag information for now.
>> The point is that current flow-dissector correctly skips any number of vlan
>> tags and get hash value from IP/TCP/UDP headers, so RPS works for multiple
>> vlan tagged packets.
>>
>> Thanks,
>> Toshiaki Makita
>
>ok, so we are on the same page.
>The flow dissector will correctly skip any number of vlans regardless
>if the first vlan is stripped or not.


On RX the first vlan is always stripped either by hw or by skb_vlan_untag.
On TX the first vlan is also stripped as validate_xmit_skb_list which
pushes vlan header is called just before dev_hard_start_xmit.

So I believe you can safely work just with skb->vlan_*