Re: [PATCH net] net: dsa: Keep a reference count on ethernet_dev

[Florian Fainelli]

On 01/24/2017 09:39 AM, David Miller wrote:
> From: Florian Fainelli <f.fainelli@xxxxxxxxx>
> Date: Sat, 21 Jan 2017 09:40:54 -0800
> 
>> of_find_net_device_by_node() just returns a reference to a net_device but does
>> not increment its reference count, which means that the master network device
>> can just vanish under our feet.
>>
>> Fixes: 83c0afaec7b7 ("net: dsa: Add new binding implementation")
>> Signed-off-by: Florian Fainelli <f.fainelli@xxxxxxxxx>
> 
> This is fine, except now this netdev is completely locked into place with
> no way to dynamically unload it.
> 
> If someone tries to modunload the driver for this ethernet device,
> their screen will fill up with warning messages indicating that the
> reference taken here in the DSA code is not going away.
> 
> You need to implement a netdev notifier that tears down this DSA
> instance during an unregister event and releases the ethernet_dev.
> Similar to how we handle protocol addresses bound to a netdev, etc.

I have been thinking about this a bit more, and this is what is going on:

- upon master network device unregister we can look up the
dsa_switch_tree in dev->dsa_ptr and call dsa_dst_unapply() that is
actually exactly what we want since it detaches the dsa_switch_tree from
the master network device

- upon master network device register, we can look up whether this
master netdev is the one of interest and re-attach the dangling switch
tree, but here we have two cases:

	- if we have an OF enabled system, doing a reverse look up of
net_device to device to device_node, and then looking it up in the
Device Tree is possible and reasonably simple, this works

	- if we have a platform data enabled system, doing a reverse lookup is
possible, but won't necessarily yield the expected results, because
platform data will have a device reference to the original master
netdev, and this one could be totally different the second time we probe
the master network device due to to unregister/register

Thoughts?
-- 
Florian