On 22/03/2021 18:56, Vladimir Oltean wrote:
> On Mon, Mar 22, 2021 at 06:35:10PM +0200, Nikolay Aleksandrov wrote:
>>> + hlist_for_each_entry(mp, &br->mdb_list, mdb_node) {
>>
>> You cannot walk over these lists without the multicast lock or RCU. RTNL is
>> not
>> enough because of various timers and leave messages that can alter both the
>> mdb_list
>> and the port group lists. I'd prefer RCU to avoid blocking the bridge mcast.
>
> The trouble is that I need to emulate the calling context that is
> provided to SWITCHDEV_OBJ_ID_HOST_MDB and SWITCHDEV_OBJ_ID_PORT_MDB, and
> that means blocking context.
>
> So if I hold rcu_read_lock(), I need to queue up the mdb entries, and
> notify the driver only after I leave the RCU critical section. The
> memory footprint may temporarily blow up.
>
> In fact this is what I did in v1:
> https://patchwork.kernel.org/project/netdevbpf/patch/20210224114350.2791260-15-olte...@gmail.com/
>
> I just figured I could get away with rtnl_mutex protection, but it looks
> like I can't. So I guess you prefer my v1?
>
Indeed, if you need a blocking context then you'd have to go with v1.
