On Sun, Nov 20, 2022 at 05:22:19PM +0000, John Keeping wrote: > On Fri, Nov 18, 2022 at 04:07:24PM -0500, Alan Stern wrote: > > On Fri, Nov 18, 2022 at 04:37:32PM +0000, John Keeping wrote: > > > I don't think it's at all simple to fix this - I posted a series > > > addressing the lifetime issues here a few years ago but didn't chase it > > > up and there was no feedback: > > > > > > https://lore.kernel.org/linux-usb/20191028114228.3679219-1-john@xxxxxxxxxxxx/ > > > > > > That includes a patch to remove the embedded struct cdev and manage its > > > lifetime separately, which I think is needed as there are two different > > > struct device objects here and we cannot tie their lifetimes together. > > > > I still don't have a clear picture of what the real problem is. Lee's > > original patch description just said "external references are presently > > not tracked", with no details about what those external references are. > > Why not add just proper cdev_get() and cdev_put() calls to whatever code > > handles those external references, so that they _are_ tracked? > > > > What are the two different struct device objects? Why do their > > lifetimes need to be tied together? If you do need to tie their > > lifetimes somehow, why not simply make one of them (the one which is > > logically allowed to be shorter-lived) hold a reference to the other? > > The problem is that we have a struct cdev embedded in f_hidg but the > lifetime of f_hidg is not tied to any kobject so we can't solve this in > the right way by setting the parent kobject of the cdev. > > While refcounting struct f_hidg is necessary, it's not sufficient > because the only way to keep it alive long enough for the final > kobject_put() on the embedded cdev is to tie the lifetime to a kobject > of its own and there is no suitable object as this is not the model > followed by gadget function instances. I see. The solution is simple: Embed a struct device in struct f_hidg, and call cdev_device_add() to add the device and the cdev. This will automatically make the device the parent of the cdev, so the device's refcount won't go to 0 until the cdev's refcount does. Then you can tie the f_hidg's lifetime to the device's, so the device's release routine can safely deallocate the entire f_hidg structure. The parent of the new struct device should be set to &gadget->dev. If you can't think of a better name for the device, you could simply append ":I" to the parent's name, where I is the interface number, or even append ":C.I" where C is the config number (like we do on the host side). Alan Stern
