On 03/04/2016 01:16 AM, Cong Wang wrote:
On Thu, Mar 3, 2016 at 12:58 AM, Daniel Borkmann <dan...@iogearbox.net> wrote:
On 03/03/2016 02:21 AM, Cong Wang wrote:

Why not set it in tun_rx_dst() where it is allocated?

Nope, current convention is to only fill options_len when an actual
option was detected on RX, f.e. see ip_tunnel_info_opts_set() in
geneve. Consumers like ovs_flow_key_extract() check for options_len
and not TUNNEL_OPTIONS_PRESENT to copy it via ip_tunnel_info_opts_get()
from there.

But the APIs suck...

You expect to use ip_tunnel_info_opts_{get,set}() to read or write
the tun_info, but actually this is not the case here for vxlan.

Yep, since depending on the working mode either skb->mark is populated
or the tunnel opts buffer.

Also, ip_tunnel_info_opts_set() could write out of range if the len is
bigger than the allocated length. I know existing callers are fine, but this API
is problematic.

Current call sites are good agree, API could probably be better, yeah.

I think this is why we had this bug.

Reply via email to