OpenWrt 21.02 status

[Alexander E. Patrakov]

Hauke Mehrtens <hauke at hauke-m.de>:

>
> On 8/31/21 1:52 PM, Alexander E. Patrakov wrote:
> > Hauke Mehrtens wrote:
> >
> >> We did the 21.02-rc4, but there is still a problem with flow offloading
> >> as this was not fixed. The other problems should be fixed now.
> >
> >> ...
> >
> >> Some more information can be found here:
> >> https://forum.openwrt.org/t/software-flow-offloading-and-conntrack-timeouts/74588
> >> https://bugs.openwrt.org/index.php?do=details&task_id=3373
> >> https://bugs.openwrt.org/index.php?do=details&task_id=3759
> >
> >> It could be that this change causes the problems:
> >> https://patchwork.ozlabs.org/project/netdev/patch/20180720130906.27687-3-pablo@netfilter.org/
> >
> >> I do not know how much time and interest I have in debugging and fixing
> >> this problem. If someone wants to have a closer look into this problem
> >> it would be really nice. even when you can make it easier to reproduce
> >> it in a test environment it would be nice.
> >
> > There is no need to debug this problem. Just revert the patch, because
> > the upstream kernel history looks like this:
> >
> > Jun 15, 2018, e97d940: flow offload pickup timeouts were changed to
> > hard-coded values, 120 seconds for TCP and 30 seconds for UDP, because
> > the author had a concern with the "established" timeout to be too huge
> > for this purpose.
> > Jun 3, 2021, 1d91d2e: sysctls are introduced so that the timeouts for
> > picking up offloaded flows are configurable, because they were too low
> > for some use cases.
> > Aug 4, 2021, 4592ee7: the timeouts were bumped again to their old
> > "established" values, sysctls are removed, the motivation is,
> > essentially, that the pickup timeouts were effectively acting like the
> > new version of established timeouts.
> >
> > I.e., the patch was effectively reverted upstream.
> >
>
> Thanks for looking into this problem.
>
> I would prefer to not delay the release any further and will take care
> of this after the release. Did you test this and see improvements in
> your scenarios?
>
> Hauke

Not yet, and can't until I return home. No objections to making this a
.1 material.

--
Alexander E. Patrakov
CV: http://u.pc.cd/wT8otalK