On 11/10/17 19:57, Willem de Bruijn wrote:
> On Wed, Oct 11, 2017 at 2:39 PM, Anton Ivanov
> <anton.iva...@kot-begemot.co.uk> wrote:
>> The check as now insists that the actual driver supports GSO_ROBUST, because
>> we have marked the skb dodgy.
>>
>> The specific bit which does this check is in net_gso_ok()
>>
>> Now, lets's see how many Ethernet drivers set GSO_ROBUST.
>>
>> find drivers/net/ethernet -type f -name "*.[c,h]" -exec grep -H GSO_ROBUST
>> {} \;
>>
>> That returns nothing in 4.x
>>
>> IMHO - af_packet allocates the skb, does all checks (and extra may be added)
>> on the gso, why is this set dodgy in the first place?
> It is set when the header has to be validated.
>
> The segmentation logic will validate and fixup gso_segs. See for
> instance tcp_gso_segment:
>
> if (skb_gso_ok(skb, features | NETIF_F_GSO_ROBUST)) {
> /* Packet is from an untrusted source, reset gso_segs. */
>
> skb_shinfo(skb)->gso_segs = DIV_ROUND_UP(skb->len, mss);
>
> segs = NULL;
> goto out;
> }
>
> If the device would have the robust bit set and otherwise supports the
> required features, fix up gso_segs and pass the large packet to the
> device.
>
> Else it continues to the software gso path.
>
> Large packets generated with psock_txring_vnet.c pass this test. I
That test is indeed a different path - this goes via the tpacket_snd
which allocs via sock_alloc_send_skb. That results in a non-fragged skb
as it calls pskb after that with data_len = 0 asking for a contiguous one.
My stuff is using sendmmsg which ends up via packet_snd which allocs
via sock_alloc_send_pskb which is invoked in a way which always creates
2 segments - one for the linear section and one for the rest (and more
if needed). It is faster than tpacket by the way (several times).
As a comparison tap and other virtual drivers use sock_alloc_send_pskb
with non-zero data length which results in multiple frags. The code in
packet_snd is in fact identical with tap (+/- some cosmetic differences).
That is the difference between the tests and that is why your test works
and mine fails.
Now, alloc-ing a 64k contiguous skb every time IMHO is wrong.
So the logic in the xmit check at present works only because it is given
only a very corner case for a GSO frame and tested versus it. It should
work with the generic case which is what comes out of
sock_alloc_send_pskb (same as in tap).
A.
> suspect that there is a subtle difference in the virtio_net_hdr fields
> that that generates vs. your program.
>
