Re: Possible problem with device_move()

[Cornelia Huck <cornelia.huck@de.ibm.com>]

On Sat, 28 Jul 2007 11:36:42 -0400 (EDT),
Alan Stern <stern@rowland.harvard.edu> wrote:

> Can you tell whether this will ever cause a problem?  Or is it known to
> be safe because whenever you call device_move(), the new parent was
> registered before the device being moved?

Unfortunately, we can't make any assumptions in cio whether the new
parent was registered before the moving device.

(A little background: We have each ccw_device as parent of a subchannel
as discovered during initial sense. In the case of z/VM (a hypervisor),
all subchannels are always consecutively numbered. Now the following
may happen: A dasd on subchannel 0 is detached and a new device, say a
ctc, is defined. As subchannel 0 was free from z/VMs perspective, the
ctc ccw_device now sits on subchannel 0 and the dasd ccw_device is
moved to the artificial 'defunct' subchannel. When the dasd is attached
again, z/VM will hand out the next free subchannel, which we allocate a
subchannel structure for and move the dasd ccw_device to. And here's
the case where the child is older than the parent...)

While we don't do suspend/resume yet on s390, I don't want to rule it
out for the future...