On Sat, Aug 01, 2020 at 06:53:40AM +0000, Peter Chen wrote: > > So, the basic routine should like below. I thought the usb_gadget should be > > deallocated before the UDC driver remove itself (UDC device is the parent of > > usb_gadget device), I may not need to wrong about it, it is just a memory region, it > > could release later. > > > > xxx_udc_release(struct device *gadget_dev) { > > struct usb_gadget *gadget = container_of(gadget_dev, struct > > usb_gadget, dev); > > kfree(gadget); > > } > > > > > > xxx_udc_probe(pdev) > > { > > udc_priv_data = kzalloc(sizeof(*udc_priv_data), GFP_KERNEL); > > gadget = kzalloc(sizeof(struct usb_gadget), GFP_KERNEL); > > udc_priv_data->gadget = gadget; > > ... > > usb_add_gadget_udc_release(&pdev->dev, gadget, xxx_udc_release); > > > > } > > > > At xxx_udc_remove(pdev) > > { > > usb_del_gadget_udc(udc_priv_data->gadget); > > /* need to never reference udc_priv_data->gadget any more */ > > udc_priv_data other deinit; > > kfree(udc_priv_data); > > } That would work. It doesn't have to be done exactly this way. Depending on the driver's needs, you could do: xxx_udc_release(struct device *dev) { udc_priv_data = dev_get_drvdata(dev); kfree(udc_priv_data); } xxx_udc_probe(pdev) { udc_priv_data = kzalloc(sizeof(*udc_priv_data), GFP_KERNEL); dev_set_drvdata(&udc_priv_data->gadget.dev, udc_priv_data); platform_set_drvdata(pdev, udc_priv_data); ... usb_add_gadget_udc_release(&pdev->dev, &udc_priv_data->gadget, xxx_udc_release); } xxx_udc_remove(pdev) { udc_priv_data = platform_get_drvdata(pdev); usb_del_gadget_udc(&udc_priv_data->gadget); } In other words, embed the struct gadget inside the udc_priv_data structure. The difference is whether you want to keep the udc_priv_data structure hanging around even while the controller is in host mode; if you do then your approach (a separate struct gadget) is better. For a peripheral-only controller, my approach would be better. > > Since all structures xxx_udc_release uses are common one, it could replace > > usb_udc_nop_release at udc/core.c. Yes, it could. But first all the UDC drivers would have to be modified. > Since gadget structure is allocated at UDC drivers, I think it should be freed at > the same place. Current common release function at udc/core.c may not a > good idea per our discussion. Agreed. Alan Stern