On 11/09, Jakub Kicinski wrote:
netdev netlink iterates over devices when dumping queues/NAPIs/qstats
etc. and takes rtnl_lock or instance lock depending on the underlying
device. Currently even if all the "real" devices in the system are ops
locked we still have to take rtnl_lock for lo (only to find out
that it doesn't even have queues or NAPIs to report).
Opt loopback into having control path under the ops lock.
This is a pretty obvious thing to do, I've held back this
patch because we used to only apply ops locking on physical devices.
netkit queue leasing made a precedent for (far more complex)
SW devices enabling ops locking. Now adding it to lo should
not create much new bug surface.
I considered an alternative of adding a "predicate" to the iteration
primitive so that we can skip the devices which obviously don't
support given API (eg qstat) without any locking. But it's more
LoC and real_num_.x_queues is not currently WRITE_ONCE()ed so
it doesn't work too well for queues.
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
---
v3:
- pre-req commit b6f74dff6d26 ("net: use two lockdep classes for
the netdev instance lock") was merged
- no change
Acked-by: Stanislav Fomichev <sdf@fomichev.me>