In fact things don't work this way. The list above stops short after
"add usb host controller devices"; the probe routines for host
controllers do not scan for USB hubs or other USB devices. Instead
they are detected by a completely separate thread (khubd).
I gather the idea is to convert dev->sem to a mutex. This idea had
occurred to me a long time ago but I didn't pursue it because of the
sheer number of places where dev->sem gets used, not to mention the
lockdep problems.
You can't possibly solve the lockdep problems here with a simple-minded
approach like your DRIVER_NORMAL, DRIVER_PARENT, etc. The device tree
is arbitrarily deep & wide, and there is at least one routine that
acquires the semaphores for _all_ the devices in the tree. This fact
alone seems to preclude using lockdep for device locks. (If there was
a form of mutex_lock() that bypassed the lockdep checks, you could use
it and avoid these issues.)
Deadlock is a serious consideration. For the most part, routines
locking devices do so along a single path in the tree. For this simple
case the rule is: Never acquire a parent's lock while holding the
child's lock.
The routine that locks all the devices acquires the locks in order of
device registration. The idea here is that children are always
registered _after_ their parents, so this should be compatible with the
previous rule. But there is a potential problem: device_move() can
move an older child under a younger parent!
Right now we have no way to deal with this. There has been some
discussion of reordering the dpm_active list when a device is moved,
but so far nothing has been done about it.
Alan Stern
-
