On Mon, Dec 22, 2014 at 3:08 PM, Pravin Shelar <pshe...@nicira.com> wrote: > On Mon, Dec 22, 2014 at 11:56 AM, Jesse Gross <je...@nicira.com> wrote: >> On Mon, Dec 22, 2014 at 1:25 PM, Pravin Shelar <pshe...@nicira.com> wrote: >>> On Sun, Dec 21, 2014 at 9:39 PM, Jesse Gross <je...@nicira.com> wrote: >>>> On Sun, Dec 21, 2014 at 6:17 PM, Jesse Gross <je...@nicira.com> wrote: >>>>> On Fri, Dec 19, 2014 at 7:25 PM, Pravin B Shelar <pshe...@nicira.com> >>>>> wrote: >>>>>> Today vport-send has complex error handling because it involves >>>>>> freeing skb and updating stats depending on return value from >>>>>> vport send implementation. >>>>>> This can be simplified by delegating responsibility of freeing >>>>>> skb to the vport implementation for all cases. So that >>>>>> vport-send needs just update stats. >>>>>> >>>>>> Signed-off-by: Pravin B Shelar <pshe...@nicira.com> >>>>> >>>>> It's somewhat non-obvious to me that handle_offloads should free the >>>>> skb in the event of an error. It seems like this introduces more >>>>> complexity to this patch as well. >>>>> >>>>> As an example of the problem, I think that the previous patch >>>>> introduces a double free in the non-compat case. This is because it >>>>> updates the skb freeing for the compat code in handle_offloads() but >>>>> the equivalent update to the non-compat code isn't made until this >>>>> patch. >>>> >>>> I now see that the upstream iptunnel_handle_offloads() also frees the >>>> skb on error but it seems like this has introduced some errors as >>>> well. It looks like several tunnel protocols continue using the skb >>>> after calling handle_offloads without any kind of check. >>> >>> right, Thats the reason. >>> I am not sure which one still missing the check, Can you point me to >>> them? I have fixed GRE already. >> >> It looks like Geneve is the only one at this point - I thought that >> VXLAN had a similar problem but I was looking at an old version of the >> code. > > OK, you are looking at upstream Geneve code.
Yes - I was referring to the upstream tunnel implementations before. That's what I meant, it seemed like these bugs were somewhat pervasive. Since almost all of the implementations have been fixed, I guess it makes sense to just leave it as is and mirror the upstream into the compat code. Will you send a patch for Geneve upstream as part of this? _______________________________________________ dev mailing list dev@openvswitch.org http://openvswitch.org/mailman/listinfo/dev
