Re: [PATCH net-next v5 02/14] net: introduce ndo_set_rx_mode_async and netdev_rx_mode_work (61d75e67dcd2)

[sdf.kernel@gmail.com]

> This is an AI-generated review of your patch. The human sending this
> email has considered the AI review valid, or at least plausible.
> ---
> Subject: net: introduce ndo_set_rx_mode_async and netdev_rx_mode_work
> 
> This commit introduces an async variant of ndo_set_rx_mode that runs from a
> workqueue, allowing drivers to sleep during RX mode changes. It also adds
> netif_rx_mode_sync() to opportunistically flush pending updates before
> returning to userspace.
> 
> > --- a/net/core/dev.c
> > +++ b/net/core/dev.c
> [ ... ]
> > @@ -12217,6 +12178,9 @@ void free_netdev(struct net_device *dev)
> >  
> >  	kfree(rcu_dereference_protected(dev->ingress_queue, 1));
> >  
> > +	if (netif_rx_mode_clean(dev))
> > +		dev_put(dev);
> > +
> >  	/* Flush device addresses */
> >  	dev_addr_flush(dev);
>
> Is this cleanup reachable during normal shutdown?
>
> When a device is queued to rx_mode_list, netif_rx_mode_queue() elevates its
> reference count via dev_hold(). During unregistration, netdev_wait_allrefs()
> blocks the system until all references are dropped.
>
> Because of this wait, the system shouldn't be able to reach free_netdev()
> until the worker thread has already dequeued the device and called
> dev_put(). It seems the device is guaranteed to not be on the list by the
> time this code executes.

I guess, yeah, we can drop this. netdev_wait_allrefs waits for all refs
to go away, so this is not needed.

> 
> > --- a/net/core/dev_addr_lists.c
> > +++ b/net/core/dev_addr_lists.c
> [ ... ]
> > +static void netdev_rx_mode_work(struct work_struct *work)
> > +{
> > +	struct net_device *dev;
> > +
> > +	rtnl_lock();
> > +
> > +	while (true) {
> > +		spin_lock_bh(&rx_mode_lock);
> > +		if (list_empty(&rx_mode_list)) {
> > +			spin_unlock_bh(&rx_mode_lock);
> > +			break;
> > +		}
> > +		dev = list_first_entry(&rx_mode_list, struct net_device,
> > +				       rx_mode_node);
> > +		list_del_init(&dev->rx_mode_node);
> > +		spin_unlock_bh(&rx_mode_lock);
> > +
> > +		netdev_lock_ops(dev);
> 
> Can this while loop starve the system of the rtnl lock?
>
> Since userspace can trigger RX mode updates via ioctls like SIOCADDMULTI
> while holding only netdev_lock_ops() and not the rtnl lock, a continuous
> stream of these operations could repeatedly add devices to rx_mode_list.
>
> Because the loop releases rx_mode_lock while processing but retains the rtnl
> lock, could it run indefinitely if the list is continually replenished?

Idk, this seems overly paranoid, the same work would happen under rtnl
on the main thread if we don't do a async callback.

> Additionally, does the order of list_del_init() and netdev_lock_ops() create
> a race condition with netif_rx_mode_sync()?
>
> If the worker thread dequeues the device and drops the spinlock, it might
> then block waiting for netdev_lock_ops(). Meanwhile, an ioctl thread
> holding netdev_lock_ops() might call netif_rx_mode_sync().
>
> The ioctl thread would call netif_rx_mode_clean(), which checks
> list_empty(). Since the worker already removed the device from the list,
> clean returns false. This causes netif_rx_mode_sync() to act as a no-op,
> allowing the ioctl to return to userspace before the hardware update
> actually completes.

The worst thing that can happen is we re-queue empty work. I don't see
how we can return to userspace without having a sync update.

> 
> > --- a/net/core/dev_api.c
> > +++ b/net/core/dev_api.c
> [ ... ]
> > @@ -311,6 +313,7 @@ int dev_set_allmulti(struct net_device *dev, int inc)
> >  
> >  	netdev_lock_ops(dev);
> >  	ret = netif_set_allmulti(dev, inc, true);
> > +	netif_rx_mode_sync(dev);
> >  	netdev_unlock_ops(dev);
> >  
> >  	return ret;
> 
> Does this correctly handle cascaded updates for stacked devices?
>
> When operating on a stacked device like a VLAN, the RX mode update cascades
> to the underlying physical device, queuing the physical device to the
> rx_mode_list via __dev_set_rx_mode().
>
> Since netif_rx_mode_sync() is only called on the stacked device here, and
> the stacked device itself is not on the rx_mode_list, will the physical
> device's update be left on the workqueue? This seems to bypass the
> synchronous update guarantee.

Yes, this assessment is correct in general. Not sure we want some new
netif_rx_mode_deep_sync or (probably better?) add some netif_rx_mode_sync
calls where appropriate. For now, leaving netif_rx_mode_sync in a few
places and planning to add more netif_rx_mode_sync if/when issues with
deep hierarchy syncing arise.