When using the felix driver (the only one which supports UC filtering
and MC filtering) as a DSA master for a random other DSA switch, one can
see the following stack trace when the downstream switch ports join a
VLAN-aware bridge:
=============================
WARNING: suspicious RCU usage
-----------------------------
net/8021q/vlan_core.c:238 suspicious rcu_dereference_protected() usage!
stack backtrace:
Workqueue: dsa_ordered dsa_slave_switchdev_event_work
Call trace:
lockdep_rcu_suspicious+0x170/0x210
vlan_for_each+0x8c/0x188
dsa_slave_sync_uc+0x128/0x178
__hw_addr_sync_dev+0x138/0x158
dsa_slave_set_rx_mode+0x58/0x70
__dev_set_rx_mode+0x88/0xa8
dev_uc_add+0x74/0xa0
dsa_port_bridge_host_fdb_add+0xec/0x180
dsa_slave_switchdev_event_work+0x7c/0x1c8
process_one_work+0x290/0x568
What it's saying is that vlan_for_each() expects rtnl_lock() context and
it's not getting it, when it's called from the DSA master's ndo_set_rx_mode().
The caller of that - dsa_slave_set_rx_mode() - is the slave DSA
interface's dsa_port_bridge_host_fdb_add() which comes from the deferred
dsa_slave_switchdev_event_work().
We went to great lengths to avoid the rtnl_lock() context in that call
path in commit 0faf890fc5 ("net: dsa: drop rtnl_lock from
dsa_slave_switchdev_event_work"), and calling rtnl_lock() is simply not
an option due to the possibility of deadlocking when calling
dsa_flush_workqueue() from the call paths that do hold rtnl_lock() -
basically all of them.
So, when the DSA master calls vlan_for_each() from its ndo_set_rx_mode(),
the state of the 8021q driver on this device is really not protected
from concurrent access by anything.
Looking at net/8021q/, I don't think that vlan_info->vid_list was
particularly designed with RCU traversal in mind, so introducing an RCU
read-side form of vlan_for_each() - vlan_for_each_rcu() - won't be so
easy, and it also wouldn't be exactly what we need anyway.
In general I believe that the solution isn't in net/8021q/ anyway;
vlan_for_each() is not cut out for this task. DSA doesn't need rtnl_lock()
to be held per se - since it's not a netdev state change that we're
blocking, but rather, just concurrent additions/removals to a VLAN list.
We don't even need sleepable context - the callback of vlan_for_each()
just schedules deferred work.
The proposed escape is to remove the dependency on vlan_for_each() and
to open-code a non-sleepable, rtnl-free alternative to that, based on
copies of the VLAN list modified from .ndo_vlan_rx_add_vid() and
.ndo_vlan_rx_kill_vid().
Fixes: 64fdc5f341 ("net: dsa: sync unicast and multicast addresses for VLAN filters too")
Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Link: https://lore.kernel.org/r/20230626154402.3154454-1-vladimir.oltean@nxp.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
d06f925f13