On Fri, 2012-11-16 at 16:22 -0800, Greg Kroah-Hartman wrote: > On Fri, Nov 16, 2012 at 05:08:53PM -0700, Toshi Kani wrote: > > > > > > > > > So the question is, does the ACPI core have to do that and if so, then why? > > > > > > > > > > > > > > > > The problem is that acpi_memory_devcie_remove() can fail. However, > > > > > > > > device_release_driver() is a void function, so it cannot report its > > > > > > > > error. Here are function flows for SCI, sysfs eject and unbind. > > > > > > > > > > > > > > Then don't ever let acpi_memory_device_remove() fail. If the user wants > > > > > > > it gone, it needs to go away. Just like any other device in the system > > > > > > > that can go away at any point in time, you can't "fail" that. > > > > > > > > > > > > That would be ideal, but we cannot delete a memory device that contains > > > > > > kernel memory. I am curious, how do you deal with a USB device that is > > > > > > being mounted in this case? > > > > > > > > > > As the device is physically gone now, we deal with it and clean up > > > > > properly. > > > > > > > > > > And that's the point here, what happens if the memory really is gone? > > > > > You will still have to handle it now being removed, you can't "fail" a > > > > > physical removal of a device. > > > > > > > > > > If you remove a memory device that has kernel memory on it, well, you > > > > > better be able to somehow remap it before the kernel needs it :) > > > > > > > > :) > > > > > > > > Well, we are not trying to support surprise removal here. All three > > > > use-cases (SCI, eject, and unbind) are for graceful removal. Therefore > > > > they should fail if the removal operation cannot complete in graceful > > > > way. > > > > > > Then handle that in the ACPI bus code, it isn't anything that the driver > > > core should care about, right? > > > > Unfortunately not. Please take a look at the function flow for the > > unbind case in my first email. This request directly goes to > > driver_unbind(), which is a driver core function. > > Yes, and as the user asked for the driver to be unbound from the device, > it can not fail. > > And that is WAY different from removing the memory from the system > itself. Don't think that this is the "normal" way that memory should be > removed, that is what stuff like "eject" was created for the PCI slots. > > Don't confuse the two things here, unbinding a driver from a device > should not remove the memory from the system, it doesn't do that for any > other type of 'unbind' call for any other bus. The device is still > present, just that specific driver isn't controlling it anymore. > > In other words, you should NEVER have a normal userspace flow that is > trying to do unbind. unbind is only for radical things like > disconnecting a driver from a device if a userspace driver wants to > control it, or a hacked up way to implement revoke() for a device. > > Again, no driver core changes are needed here. Okay, we might be able to make the eject case to fail if an ACPI driver is not bound to a device. This way, the unbind case may be harmless to proceed. Let us think about this further on this (but we may come up again :). Thanks, -Toshi -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxx. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>