[cc += Heiner Kallweit, Andrew Lunn]

On Thu, Mar 10, 2022 at 12:38:20PM +0100, Oleksij Rempel wrote:

On Thu, Mar 10, 2022 at 12:25:08PM +0100, Oliver Neukum wrote:

quoted

I got bug reports that 2c9d6c2b871d ("usbnet: run unbind() before
unregister_netdev()")
is causing regressions.

I would like to see this reverted as well.  For obvious reasons,
the order in usbnet_disconnect() should be the inverse of
usbnet_probe().  Since 2c9d6c2b871d, that's no longer the case.

quoted

Rather than simply reverting it,
it seems to me that the call needs to be split. One in the old place
and one in the place you moved it to.

I disagree.  The commit message claims that the change is necessary
because phy_disconnect() fails if called with phydev->attached_dev == NULL.

I've just gone through the code to check that and the only thing that
caught my eye is an unconditional call to netif_testing_off(dev) in
phy_stop().  It seems to me that making that call conditional on
"if (dev)" would solve the issue.

That's a bug introduced by
   4a459bdc7472 ("net: phy: Put interface into oper testing during cable
                  test")
or 472115d9834c ("net: phy: stop PHY if needed when entering
                  phy_disconnect")
depending on how you look at it.

Thanks,

Lukas

quoted

Could you tell me which drivers you moved the call for?

I moved it for asix_devices.c:ax88772_unbind()

Probably smsc95xx.c:smsc95xx_unbind() has same issue.