On 14.04.22 17:01, Andy Shevchenko wrote:
>
> Good question. Isn't it the commit 2c9d6c2b871d ("usbnet: run unbind()
> before unregister_netdev()") which changed the ordering of the
> interface shutdown and basically makes this race happen? I don't see
> how we can guarantee that IOCTL won't be called until we quiescence
> the network device — my understanding that on device surprise removal
True. The best we could do is introduce a mutex for ioctl() and
disconnect(). That seems the least preferable solution to me.
> we have to first shutdown what it created and then unbind the device.
> If I understand the original issue correctly then the problem is in
> usbnet->unbind and it should actually be split to two hooks, otherwise
> it seems every possible IOCTL callback must have some kind of
> reference counting and keep an eye on the surprise removal.
>
> Johan, can you correct me if my understanding is wrong?
>
It seems to me that fundamentally the order of actions to handle
a hotunplug must mirror the order in a hotplug. We can add more hooks
if that turns out to be necessary for some drivers, but the basic
reverse mirrored order must be supported and I very much favor
restoring it as default.

So I am afraid I have to ask again, whether anybody sees a fundamental
issue with the attached patch, as opposed to it not being an elegant
solution?
It looks to me that we are in a fundamental disagreement on the correct
order in this question and there is no productive way forward other than
offering both ways.

    Regards
        Oliver