On Mon, 2014-03-10 at 19:53 -0700, Julius Werner wrote: > I think usbnet_stop() raced with the dev->bh tasklet, which by itself > might not be a problem (usbnet_stop() later kills the tasklet itself, > so it should expect that it can be running before that). The issue is > that it calls usbnet_terminate_urbs() before that, which temporarily > installs a waitqueue in dev->wait in order to be able to wait on the > tasklet to run and finish up some queues. The waiting itself looks > okay, but the access to 'dev->wait' is totally unprotected and can > race arbitrarily. I think in this case usbnet_bh() managed to succeed > it's dev->wait check just before usbnet_terminate_urbs() sets it back > to NULL. The latter then finishes and the waitqueue_t structure on its > stack gets overwritten by other functions halfway through the > wake_up() call in usbnet_bh(). > > I think the best solution would be to just make dev->wait a directly > embedded structure inside struct usbnet instead of a pointer to > something stack-allocated. usbnet_bh() could just call wake_up() > unconditionally (if empty it will be a noop), and then one other check > for !dev->wait could be replaced with a call to waitqueue_active(). > Then the waitqueue-internal locks should be enough to protect all > accesses. The diagnosis seems spot on. The fix is not quite so simple. dev->wait is abused as a flag in resume(). The easiest fix is just to make sure resume() is never called while stopping. Regards Oliver -- To unsubscribe from this list: send the line "unsubscribe linux-usb" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
