Am Freitag, 12. Dezember 2014, 16:15:40 schrieb Eric Dumazet:

On Sat, 2014-12-13 at 00:47 +0100, Wolfgang Walter wrote:

quoted

I can't disable it as the driver will not allow it:
# ethtool -K eth0 tx off
Cannot change tx-checksumming
Could not change any device features

Sounds a bug in itself :(

Yes, that is what I meant in my email that there seems also to be bug in the 
GSO path.


I don't understand how tcp_sendmsg() marks gso/non-gso skbs.

tcp_send_mss() calls tcp_xmit_size_goal() which calls tcp_xmit_size_goal() and 
there is a check for GSO ( sk_can_gso(sk) ). If sk_can_gso() returns false 
then tcp_xmit_size_goal() returns mss_now. gso_segs seems to be set to zero. 
In in tcp_output.c gso_segs == 0 seems to allow GSO.

What if mss later shrinks? Couldn't it be that then tcp_set_skb_tso_segs() 
gets called and then there the gso branch is taken?

Regards,
-- 
Wolfgang Walter
Studentenwerk München
Anstalt des öffentlichen Rechts