On Wed, Dec 09, 2015 at 02:08:35PM +0100, Linus Walleij wrote: > Because we want to have a proper userspace ABI for GPIO chips, > which involves using a character device that the user opens > and closes. While the character device is open, the underlying > kernel objects must not go away. Okay, so you stop the gpio_chip struct from going away. What about the code which is called via gpio_chip - say, if userspace keep shte chardev open, and someone rmmod's the driver providing the GPIO driver. I'm not sure that splitting up objects in this way really solves anything at all. Yes, it divorses the driver's private data from the subsystem data, but is that really an advantage? Network drivers have a similar issue, and the way this problem is solved there is that alloc_netdev() is always used to allocate the subsystem data structure and any driver private data structure as one allocation, and the lifetime of both objects remains under the control of the subsystem. The allocated memory is only freed when the last user goes away, and net has protection to prevent an unregistered driver from being called (via locks on every path into the layer.) Things get a little more complex with gpio, because there's the issue that some methods are spinlocked while others can take semaphores, but it should be possible to come up with a solution to that - maybe an atomic_t which is incremented whenever we're in some operation provided it's >= 0 (otherwise it fails), and decremented when the operation completes. We can then control in the unregistration path further GPIO accesses, and also prevent new accesses occuring by setting the atomic_t to -1. This shouldn't require any additional locking in any path. It does mean that the unregistration path needs careful thought to ensure that when we set it to -1, we wait for it to be dropped by the appropriate amount. -- RMK's Patch system: http://www.arm.linux.org.uk/developer/patches/ FTTC broadband for 0.8mile line: currently at 9.6Mbps down 400kbps up according to speedtest.net.
