On Tue, Feb 13, 2018 at 08:23:14PM +0200, Eyal Birger wrote: > Hi Xin Long, > > On Tue, 13 Feb 2018 23:18:14 +0800 > Xin Long <[email protected]> wrote: > > > On Tue, Feb 13, 2018 at 6:54 PM, Eyal Birger <[email protected]> > > wrote: > > > In setups like the following: > > > > > > Host A -- Host B > > > tun0 -- ipsec -- eth0 -- eth0 -- ipsec -- tun0 > > > > > > where tun0 are tunnel devices using dst_cache (ipip, ipip6, etc...). > > > > > > Unregistration of an underlying eth0 device leads to the following > > > log messages: > > > > > > unregister_netdevice: waiting for eth0 to become free. Usage count > > > = 2 > > > > > > This is since xfrm dsts device references are not released upon > > > device unregistration when the xfrm dst is cached in a dst_cache. > > > > > > This issue was first introduced in commit 52df157f17e5 > > > ("xfrm: take refcnt of dst when creating struct xfrm_dst bundle") > > > as part of an effort to remove routing garbage collection. > > > > > > Several approaches for fixing this were discussed in [1]; this > > > commit keeps track of allocated xdsts and releases their device > > > references on a netdev unregister/down events. > > > > > > Signed-off-by: Eyal Birger <[email protected]> > > > Fixes: 52df157f17e5 ("xfrm: take refcnt of dst when creating struct > > > xfrm_dst bundle") > > I had another fix [1] for this issue in my local tree. > > That reuses uncached_list to track xdsts, and fewer job to do for > > this. As naturally uncached_list is already the list that is supposed > > to do this job. We don't have to introduce a new list and members. > > > > [1] https://paste.fedoraproject.org/paste/X8FQzEXjGXQBTKAVLnDn4w > > I have tested your patch and it resolves the issue described. > I like not having to introduce a new list and lock.
Me too. This looks like the best of all proposed solutions. > > Small note: I think you need to increase net->ipv6.rt6_stats->fib_rt_uncache > when adding to the IPv6 uncached list from the xfrm code. Xing, plaese incorporate the review and submit the patch. Thanks!
