Hi Greg,
Suggest including next patch (available in linux-mainline) to 5.4-stable branch: commit ae7e86108b12 ("usb: dwc3: Stop active transfers before halting the controller"). It's also already present in 5.10 stable. Some fixes exist in 5.10-stable for that patch too.
This patch fixes panic in case of using USB2.0 Dual Role Device controller, as described below.
1. Boot in peripheral role 2. Configure RNDIS gadget, perform ping, stop ping 3. Switch to host role 4. Kernel panic occurs
Kernel panic happens because gadget->udc->driver->disconnect() (which is configfs_composite_disconnect()) is not called from usb_gadget_disconnect() function, due to timeout condition in dwc3_gadget_run_stop(), which leads to not called rndis_disable(). And although previously created endpoints are not valid anymore, eth_start_xmit() gets called and tries to use those, which leads to invalid memory access. This patch fixes timeout condition, so next call chain doesn't fail anymore, and RNDIS uninitialized properly on gadget to host role switch:
<<<<<<<<<<<<<<<<<<<< cut here >>>>>>>>>>>>>>>>>>> usb_role_switch_set_role() v dwc3_usb_role_switch_set() v dwc3_set_mode() v __dwc3_set_mode() v dwc3_gadget_exit() v usb_del_gadget_udc() v usb_gadget_remove_driver() v usb_gadget_disconnect() v // THIS IS NOT CALLED because gadget->ops->pullup() = // dwc3_gadget_pullup() returns -ETIMEDOUT (-110) gadget->udc->driver->disconnect() // = configfs_composite_disconnect() v composite_disconnect() v reset_config() v foreach (f : function) : f->disable v rndis_disable() v gether_disconnect() v usb_ep_disable(), dev->port_usb = NULL <<<<<<<<<<<<<<<<<<<< cut here >>>>>>>>>>>>>>>>>>>
Thanks!