On Wed, 2023-07-19 at 20:30 -0700, Jakub Kicinski wrote: > On Thu, 20 Jul 2023 01:12:07 +0200 Florian Westphal wrote: > > I don't see any netfilter involvement here. > > > > The repro just creates a massive amount of team devices. > > > > At the time it hits the LOCKDEP limits on my test vm it has > > created ~2k team devices, system load is at +14 because udev > > is also busy spawing hotplug scripts for the new devices. > > > > After reboot and suspending the running reproducer after about 1500 > > devices (before hitting lockdep limits), followed by 'ip link del' for > > the team devices gets the lockdep entries down to ~8k (from 40k), > > which is in the range that it has on this VM after a fresh boot. > > > > So as far as I can see this workload is just pushing lockdep > > past what it can handle with the configured settings and is > > not triggering any actual bug. > > The lockdep splat because of netdevice stacking is one of our top > reports from syzbot. Is anyone else feeling like we should add > an artificial but very high limit on netdev stacking? :( We already have a similar limit for xmit: XMIT_RECURSION_LIMIT. I guess stacking more then such devices will be quite useless/non functional. We could use such value to limit the device stacking, too. Cheers, Paolo
