On Fri, Feb 10, 2023 at 03:49:41PM +0100, Sascha Hauer wrote: > On Thu, Feb 09, 2023 at 10:05:35AM -0500, Alan Stern wrote: > > Sascha: > > > > When you first posted your original patch, I wondered if it was really > > the right thing to do. Making the net device not be a child of the UDC > > device means you can (in theory) have strange behavior such as the > > kernel suspending the USB device controller while expecting the network > > interface to keep on working. > > > > Is there a different way of solving the original problem? > > I don't know which. One thing would be to couple the lifetime of the > ethernet device to the lifetime of the UDC, but the result would look > different to userspace, so wouldn't be ideal either. > > Note the original reason doing this change was that we saw backtraces > when doing a 'reboot -f', the 'rmmod dummy_hcd' was just an easy > reproducer for the problem. > > One other possibility might be to take a reference to the UDC while > it is in use so that the module can't be rmmoded. Not sure if that fixes > my original problem though. Not being familiar with the networking code, I don't really understand the original problem. Does the use-after-free error occur when you try to dereference a dev->parent pointer in the ethernet device? If that's so, then taking a reference (i.e. get_device()) on the parent device should fix the problem. If not, maybe you can give a more detailed guide as to what's going wrong. Alan Stern