On Tue, Feb 05, 2019 at 07:41:18PM +0000, Ben Hutchings wrote:
> > > Peter Oskolkov checked an earlier version of this backport, but I have
> > > since rebased and added another 3 commits to it. I tested with the
> > > ip_defrag.sh self-test that he added upstream, and it passed. I have
> > > included the fix that is currently queued for the 4.9, 4.14 and 4.19
> > > branches.
> >
> > That's a lot of patches, some of which I have already queued up in the
> > next 4.4 release which will happen in a day or so. Are they all still
> > needed after the changes there are merged?
>
> Ah, yes, a lot of the fragment-handling changes are already in your
> queue and I'm not certain that all of mine are needed. However I don't
> think the changes in your queue are complete and correct. When I run
> the ip_defrag.sh self-test:
>
> 1. The ipv4 non-overlap case fails after a few seconds, with recv()
> returning an EAGAIN error. If I modify the script to continue after an
> error, the other cases do pass, however. This is not a regression from
> 4.4.172, but with my changes all cases pass.
>
> 2. There is a reference leak which prevents the new network namespaces
> being cleaned up ("unregister_netdevice: waiting for lo to become free.
> Usage count = 61"). With 4.4.172 or with my changes applied, the
> warnings appear, but only for about a minute with the number gradually
> decreasing. So this is a regression.
>
> 3. If I run the test again, it hangs. Shutting down the VM also hangs.
> I think this is related to the previous issue. Again, this is a
> regression.
Ok, I dropped those patches from the 4.4 queue before releasing it. Let
me go add them back for the moment and then I'll dig through all of this
over the next few days and see what it looks like...
thanks,
greg k-h
