Re: [PATCH net] team: grab team lock during team_change_rx_flags

[Stanislav Fomichev <stfomichev@gmail.com>]

On 05/15, Jay Vosburgh wrote:
> Stanislav Fomichev <stfomichev@gmail.com> wrote:
> 
> >On 05/15, Stanislav Fomichev wrote:
> >> On 05/15, Jakub Kicinski wrote:
> >> > On Wed, 14 May 2025 15:03:19 -0700 Stanislav Fomichev wrote:
> >> > > --- a/drivers/net/team/team_core.c
> >> > > +++ b/drivers/net/team/team_core.c
> >> > > @@ -1778,8 +1778,8 @@ static void team_change_rx_flags(struct net_device *dev, int change)
> >> > >  	struct team_port *port;
> >> > >  	int inc;
> >> > >  
> >> > > -	rcu_read_lock();
> >> > > -	list_for_each_entry_rcu(port, &team->port_list, list) {
> >> > > +	mutex_lock(&team->lock);
> >> > > +	list_for_each_entry(port, &team->port_list, list) {
> >> > 
> >> > I'm not sure if change_rx_flags is allowed to sleep.
> >> > Could you try to test it on a bond with a child without IFF_UNICAST_FLT,
> >> > add an extra unicast address to the bond and remove it?
> >> > That should flip promisc on and off IIUC.
> >> 
> >> I see, looks like you're concerned about addr_list_lock spin lock in
> >> dev_set_rx_mode? (or other callers of __dev_set_rx_mode) Let me try
> >> to reproduce with your example, but seems like it's an issue, yes
> >> and we have a lot of ndo_change_rx_flags callbacks that are sleepable :-(
> >
> >Hmm, both bond and team set IFF_UNICAST_FLT, so it seems adding/removing uc
> >address on the bonding device should not flip promisc. But still will
> >verify for real.
> 
> 	I think Jakub is saying that adding a unicast address to the
> bond would change promisc on the underlying device that's part of the
> bond (a not-IFF_UNICAST_FLT interface), not on the bond itself.  The
> question is whether that change of promisc in turn generates a sleeping
> function warning.
> 
> 	FWIW, I think an easy way to add a unicast MAC to a bond is to
> configure a VLAN above the bond, then change the MAC address of the VLAN
> interface (so it doesn't match the bond's).

This seems to work (using teaming instead of bonding, but should not matter):

  ip link add name dummy1 type dummy
  ip link add name team0 type team
  #ip link set team0 down # hit team_port_add vs team_set_rx_mode
  ip link set dev dummy1 master team0 # promisc enabled here

  ip link set dummy1 up
  ip link set team0 up # or here (if was down previously)

  ip link add link team0 name team0.100 type vlan id 100
  ip link set dev team0.100 address 00:00:00:00:00:02
  ip link set team0.100 up

But mostly because promisc is enabled via team_port_add->dev_uc_sync_multiple
(when team is up when adding a port) or via
do_setlink->netif_change_flags->ndo_set_rx_mode->team_set_rx_mode (when
upping team). The subsequent calls to __dev_set_rx_mode are noops
because the device is already in promisc. IOW, I don't see how we can
reach team_change_rx_flags from here.