On Mon, Apr 29, 2024 at 06:07:25PM -0400, Joseph Huang wrote:
Something like this (some layers omitted for brevity)?
+br_iterator
| for each mdb
| _br_switchdev_mdb_notify
rtnl_lock | without F_DEFER flag
| | |
+switchdev_port_attr_set_deferred | +switchdev_port_obj_notify
| | |
+dsa_port_mrouter | +dsa_user_port_obj_a/d
| | |
+mv88e6xxx_port_mrouter----------+ +mv88e6xxx_port_obj_a/d
|
+--------------------------------------+
|
rtnl_unlock
At a _very_ superficial glance, I don't think you are properly
accounting for the fact that even with rtnl_lock() held, there are still
SWITCHDEV_OBJ_ID_PORT_MDB events which may be pending on the switchdev
chain. Without a switchdev_deferred_process() flush call, you won't be
getting rid of them, so when you rtnl_unlock(), they will still run.
Even worse, holding rtnl_lock() will not stop the bridge multicast layer
from modifying its br->mdb_list; only br->multicast_lock will.
So you may be better off also acquiring br->multicast_lock, and
notifying the MDB entries to the switchdev chain _with_the F_DEFER flag.
Note that on the system I tested, each register read/write takes about 100us
to complete. For 100's of mdb groups, this would mean that we will be
holding rtnl lock for 10's of ms. I don't know if it's considered too long.
Not sure how this is going to be any better if the iteration over MDB
entries is done 100% in the driver, though - since its hook,
dsa_port_mrouter(), runs entirely under rtnl_lock(). >
Anyway, with the SWITCHDEV_F_DEFER flag, maybe the mdb object
notifications can be made to run by switchdev only a few at a time, to
give the network stack time to do other things as well.