Re: __sk_buff.data_end

[Daniel Borkmann <daniel@iogearbox.net>]

On 04/20/2017 02:12 AM, Alexei Starovoitov wrote:
> On Thu, Apr 20, 2017 at 02:01:49AM +0200, Daniel Borkmann wrote:
>> On 04/20/2017 12:20 AM, Johannes Berg wrote:
>>> On Wed, 2017-04-19 at 23:31 +0200, Johannes Berg wrote:
>>>> Hi Alexei, Daniel,
>>>>
>>>> I'm looking at adding the __wifi_sk_buff I talked about, and I notice
>>>> that it uses CB space to store data_end. Unfortunately, in a lot of
>>>> cases, we don't have any CB space to spare in wifi.
>>>
>>> I guess I can work around this, would this seem reasonable?
>>>
>>>   struct bpf_skb_data_end {
>>>          struct qdisc_skb_cb qdisc_cb;
>>> -       void *data_end;
>>> +       /*
>>> +        * The alignment here is for mac80211, since that doesn't use
>>> +        * a pointer but a u64 value and needs to save/restore that
>>> +        * across running its BPF programs.
>>> +        */
>>> +       void *data_end __aligned(sizeof(u64));
>>>   };
>>
>> Yeah, should work as well for the 32 bit archs, on 64 bit we
>> have this effectively already:
>>
>> struct bpf_skb_data_end {
>>          struct qdisc_skb_cb        qdisc_cb;             /*     0    28 */
>>
>>          /* XXX 4 bytes hole, try to pack */
>>
>>          void *                     data_end;             /*    32     8 */
>>
>>          /* size: 40, cachelines: 1, members: 2 */
>>          /* sum members: 36, holes: 1, sum holes: 4 */
>>          /* last cacheline: 40 bytes */
>> };
>>
>> Can you elaborate on why this works for mac80211? It uses cb
>> only up to that point from where you invoke the prog?
>
> +1
>
> also didn't we discuss that wifi has crazy non-linear skb?
> this data/data_end is used by cls_bpf with headlen only
> for direct packet access where performance matters.

bpf_skb_pull_data() helper can be used as an option to pull
in more, though, f.e. up to bpf_skb_pull_data(skb, skb->len)
in the worst case, which then results in a fully linearized
skb where data/data_end has complete access. That much may
not be needed, though, but f.e. cls_bpf can certainly expand
the available headlen for direct packet access.

> Since wifi skbs have only eth in headlen, there is not much
> pointing adding support for data/data_end to wifi.
> Just use ld_abs/ld_ind instructions and load_bytes() helper.

Afaik, the ld_abs/ld_ind are not an option due to the data
on the wire being in little endian, but the bpf_skb_load_bytes()
might be the way to go initially, agree.