On 06/15/2016 02:37 PM, Max Kellermann wrote: > On 2016/06/15 22:32, Shuah Khan <shuahkh@xxxxxxxxxxxxxxx> wrote: >> This change introduces memory leaks, since drivers are relying on >> media_device_unregister() to free interfaces. > > This is what I thought, too, until I checked the code paths. Who adds > entries to that list? Only media_gobj_create() does, and only when > type==MEDIA_GRAPH_INTF_DEVNODE. That is called via > media_interface_init(), via media_devnode_create(). > > In the whole kernel, there are two calls to media_devnode_create(): > one in dvbdev.c and another one in v4l2-dev.c. Both callers take care > for freeing their interface. Both would crash if somebody else would > free it for them before they get a chance to do it. Which is the very > thing my patch addresses. > > Did I miss something? > Yes media_devnode_create() creates the interfaces links and these links are deleted by media_devnode_remove(). media_device_unregister() still needs to delete the interfaces links. The reason for that is the API dynalic use-case. Drivers (other than dvb-core and v4l2-core) can create and delete media devnode interfaces during run-time, hence media_devnode_remove() has to call media_remove_intf_links(). However, driver isn't required to call media_devnode_remove() and media-core can't enforce that. So it is safe for media_device_unregister() to remove interface links if the list isn't empty. If driver does delete them, media_device_unregister() has nothing to do since the list is going to be empty. So removing kfree() from media_device_unregister() isn't the correct fix. I don't see the stack trace for the double free error you are seeing? Could it be that there is a driver problem in the order in which it is calling media_device_unregister()? thanks, -- Shuah -- To unsubscribe from this list: send the line "unsubscribe linux-media" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
