Am 23.05.2017 um 23:01 schrieb Lino Sanfilippo:

On 23.05.2017 21:38, Stefan Wahren wrote:

quoted

quoted

Lino Sanfilippo [off-list ref] hat am 23. Mai 2017 um 20:16 geschrieben:

I suggest to avoid this possible race by first unregistering the netdevice and then 
calling cancel_work_sync().

What makes you sure that's safe to unregister the netdev while the tx work queue is possibly active?

unregister_netdevice() calls netdev_close() if the interface is still up. netdev_close() calls flush_work()
so the unregistration is delayed until the tx work function is finished. Furthermore both close() and
tx work are synchronized by means of the qca->lock which also guarantees that unregister_netdevice() wont
be finished until the tx work is done.

Thanks for the explanation. I suspect there could be the same race
between serdev_device_close() and the tx work queue.

So i would propose a variant of your original suggestion:

unregister_netdev(qca->net_dev);

/* Flush any pending characters in the driver. */
serdev_device_close(serdev);
cancel_work_sync(&qca->tx_work);

Since we have the same pattern in the error path of the probe function,
the same applies there.

Stefan