Re: [PATCH] Driver core: Don't ignore bus_attach_device() retval

[Cornelia Huck]

On Wed, 11 Oct 2006 10:49:36 -0400 (EDT),
Alan Stern <stern@xxxxxxxxxxxxxxxxxxx> wrote:

> You know, I'm not so sure device registration should fail when 
> bus_attach_device() returns an error.

Hm, let's see why bus_attach_device() might fail:

* device_bind_driver() failed to create some symlinks. We may
consider not to fail in this case, since sysfs_remove_link() is fine
even for non-existing links.

* probing failed for one possible driver with something other than
-ENODEV or -ENXIO. Not sure if we really should abort in this case.
We'd just end up with an unbound device, and a driver returning (for
example) -ENOMEM for probing may just be a really dumb driver trying to
allocate an insane amount of memory (and the next driver might just be
fine).

> Furthermore there are subtle problems that can arise.  In effect, the
> device is registered for a brief time (while the driver is probed) and
> then unregistered without giving the bus subsystem a chance to prepare for
> the removal.  With USB this can lead to problems; if the driver called
> usb_set_interface() then child devices would be created below the one
> being probed -- and they would never get removed.

One way to fix this would be to make device_bind_driver() always
succeed (even without symlinks), the other to call the ->remove
function if device_bind_driver() fails (assuming that the ->remove
method should undo the stuff done in ->probe).

> In fact, we might want to separate driver probing from device_add()  
> entirely.  That is, make them available as two separate function calls.  
> That way the subsystem driver will have a chance to create attribute files
> before a uevent is generated and a driver is loaded.  (That should help
> udev to work better.)  This would require a larger change, though --
> probably requiring an alternate version of device_add().

Shouldn't subsystems that need attributes early just use dev->groups,
class->dev_attrs or bus->dev_attrs? These attribute groups are added
before the uevent is generated.

-- 
Cornelia Huck
Linux for zSeries Developer
Tel.: +49-7031-16-4837, Mail: cornelia.huck@xxxxxxxxxx
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/