Re: [PATCH net v1] net: defer __dev_set_promiscuity() to avoid sleeping in atomic context

[Stanislav Fomichev <stfomichev@gmail.com>]

On 02/19, Jakub Kicinski wrote:
> On Thu, 19 Feb 2026 10:59:01 -0800 Stanislav Fomichev wrote:
> > On 02/18, Jakub Kicinski wrote:
> > > On Tue, 17 Feb 2026 17:10:36 -0800 Stanislav Fomichev wrote:  
> > > > > Reproducer:
> > > > > 
> > > > >   ip link add dummy0 type dummy
> > > > >   ip link add team0 type team
> > > > >   ip link set dummy0 master team0
> > > > >   ip link set team0 up
> > > > >   ip link add bridge0 type bridge vlan_filtering 1
> > > > >   ip link set bridge0 up
> > > > >   ip link set team0 master bridge0
> > > > >   ip link add macsec0 link bridge0 type macsec
> > > > >   ip link set macsec0 up   # triggers the bug    
> > > > 
> > > > Can you add it as a selftest under selftests/drivers/net/team/?  
> > > 
> > > Stan, this "fix" may work for the promisc flag but won't we have 
> > > the same problem with sync'ing the address list? Looks like team
> > > will do:
> > >  - team_set_rx_mode()
> > >    - dev_uc_sync_multiple()
> > >      - __dev_set_rx_mode(port->dev)
> > > so AFAICT we're calling ndo_set_rx_mode without holding the instance
> > > lock?  
> > 
> > Not sure I understand your trace without more details about the hierarchy.
> 
> Team on top of a ops-locked netdev 
> 
>  - team_set_rx_mode() # set_rx_mode on team
>    - dev_uc_sync_multiple()
>      - __dev_set_rx_mode(port->dev) # calls ndo_set_rx_mode on ops-locked
>                                     # netdev without holding the inst. lock
> 
> IOW this will fire:
> 
> diff --git a/drivers/net/netdevsim/netdev.c b/drivers/net/netdevsim/netdev.c
> index 6285fbefe38a..77991f62bffc 100644
> --- a/drivers/net/netdevsim/netdev.c
> +++ b/drivers/net/netdevsim/netdev.c
> @@ -184,6 +184,7 @@ static netdev_tx_t nsim_start_xmit(struct sk_buff *skb, struct net_device *dev)
>  
>  static void nsim_set_rx_mode(struct net_device *dev)
>  {
> +       netdev_assert_locked(dev);
>  }
>  
>  static int nsim_change_mtu(struct net_device *dev, int new_mtu)

Ah, you're saying in general.. Yeah, agreed, for the instance locked
devices not grabbing an instance lock for these paths looks problematic.
 
> > But you have a point, per netdevices.rst ndo_set_rx_mode is synchronized via
> > netif_addr_lock and we are breaking that with this patch.. :-(
> > (so I don't think we need an instance lock if we keep netif_addr_lock?)
> > 
> > For this particular issue, maybe we can do something similar to net_todo_list?
> > Instead of changing the promisc for !FLT under right here right now, move it
> > to the rtnl_unlock? Not sure how important the ordering is..
> 
> Not sure. Another alternative is to implement the long standing idea of
> having an async / sleeping version of ndo_set_rx_mode() orchestrated
> by the core. Because a lot of drivers need to sleep, anyway, so they
> just schedule a work from that callback.
> 
> Then we can say old ndo_set_rx_mode is under netif_addr_lock.
> ndo_set_rx_mode_async is under instance lock.

That sounds like a better plan going forward, but gonna need a bunch of
work to redo the addr lock it seems? We can start with moving promisc into
rtnl_unlock to unblock that "bridge vlan_filtering 1" and I
can look into adding an instance lock for set_rx_mode.. LMK if you prefer
me to focus on the latter and don't waste time on the former.