On Tue, 24 Mar 2026 15:49:27 -0700 Stanislav Fomichev wrote: > > > Not sure why cancel+release, maybe you're thinking about the unregister > > > path? This is rtnl_unlock -> netdev_run_todo -> __rtnl_unlock + some > > > extras. > > > > > > And the flush is here to plumb the addresses to the real devices > > > before we return to the callers. Mostly because of the following > > > things we have in the tests: > > > > > > # TEST: team cleanup mode lacp > > > [FAIL] > > > # macvlan unicast address not found on a slave > > > > > > Can you explain a bit more on the suggestion? > > > > Oh, I thought it's here for unregister! Feels like it'd be cleaner to > > add the flush in dev_*c_add() and friends? How hard would it be to > > identify the callers in atomic context? > > Not sure we can do it in dev_xc_add because it runs under rtnl :-( > I currently do flush in netdev_run_todo because that's the place that > doesn't hold rtnl. Otherwise flush will get stuck because the work > handler grabs it...
I was thinking of something a'la linkwatch. We can "steal" / "flush" the pending work inline. I guess linkwatch is a major source of races over the years... Does the macvlan + team problem still happens with the current implementation minus the flush? We are only flushing once so only pushing the addresses thru one layer of async callbacks.
