On Mon, 04 Mar 2019 23:28:14 +0100, Toke Høiland-Jørgensen wrote: > Jakub Kicinski <jakub.kicin...@netronome.com> writes: > > On Mon, 04 Mar 2019 20:05:30 +0100, Toke Høiland-Jørgensen wrote: > >> > Hm. I think you'll still need a lock (mutex?) on the alloc path, but > >> > the free path should be fine as long as you load the map pointer before > >> > looking at the refcnt (atomic op ensuring the barrier there). > >> > >> Yeah, for the per-namespace refcnt it's pretty straight forward, the > >> trouble is the global count that needs to iterate over all namespaces; > >> probably need to put that all behind a (non-spin)lock, right? > > > > Because net iteration is under RCU? You can switch to taking net_rwsem > > for that one, no? I'm probably confused again ;) > > Because there's a single refcount that needs to trigger > creation/deletion of *all* the default maps. I.e. > > if (atomic_dec_return(&global_refcnt)) > for_each_namespace(net) > destroy_default_map(net); > > which needs to not step on the toes of a subsequent > > if (atomic_inc_return(&global_refcnt) == 1) > for_each_namespace(net) > create_default_map(net); > > (or vice versa, of course). > > Not sure there's a way to do that without wrapping both of those > constructs (including the refcnt inc/dec) in a mutex?
Seems like it :(
