[snip]
> The test can be run both with and without ring:
>
> psock_txring_vnet -l 8000 -s $src_ip -d $dst_ip -v
> psock_txring_vnet -l 8000 -s $src_ip -d $dst_ip -v -N
>
> both with and without qdisc bypass ('-q').
Thanks, apologies, I was being inpatient. Started reading the source,
saw the tpacket bits and stopped there.
>
>> - 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.
> but attached the ring slot as fragments in tpacket_fill_skb.
>
>> 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.
> All the above test cases work for me, including those that build skbs
> with fragments. Could you try those.
Tried it, works on all of the adapters and hosts where mine fails. I
will step by step hack-in the differences so it behaves same as mine
until I find the culprit.
This will be tomorrow though, it is late here.
The only obvious difference I can see at this point is that I am using
iovs and sending the vnet header as iov[0] and the data in pieces after
that while your code is doing a send() for the whole frame. This should
not make any difference though - it all ends up as an iov internally in
the kernel.
A.
>
