Re: [RFC] bind/unbind uevent

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

On Wed, 15 Dec 2010 08:23:16 -0800,
Greg KH <gregkh@suse.de> wrote:

> On Wed, Dec 15, 2010 at 02:21:13PM +0100, Cornelia Huck wrote:
> > On Tue, 14 Dec 2010 11:29:52 -0800,
> > Greg KH <gregkh@suse.de> wrote:
> > 
> > > On Tue, Dec 14, 2010 at 07:26:40PM +0100, Sebastian Ott wrote:
> > > > 
> > > > On Mon, 13 Dec 2010, Greg KH wrote:
> > > > > On Mon, Dec 13, 2010 at 08:27:45PM +0100, Sebastian Ott wrote:
> > 
> > > > > Don't do that, have your _driver_ register the attributes with the bus
> > > > > it is on, then when the binding happens, the attributes will
> > > > > automatically get created for the device before the notification is sent
> > > > > to userspace.  That is the proper proceedure here.
> > > > 
> > > > are you suggesting that these driver specific device attributes should be
> > > > created by the bus code which registers the device?
> > > 
> > > Yes.
> > > 
> > > > at this time it is not determinable which driver will bind to the device
> > > > (and therefore which attributes to create), also driver specific data may 
> > > > not be initialized.
> > > 
> > > {sigh}
> > > 
> > > Please go _look_ at how the driver model handles this type of thing.  It
> > > does _exactly_ what you want, as we have been doing it for _years_
> > > properly.
> > > 
> > 
> > I don't understand how this could work. All of the attribute (groups)
> > registered before the uevent are driver-ignorant. If the bus wants to
> > specify the attributes, it needs to know the driver the device will
> > bind to (regardless of whatever tables exist that show the driver <->
> > attribute relationship). But it cannot know the driver until after the
> > uevent. Or else it would need to create _all_ attributes for _all_
> > devices (surely you didn't mean that)? And what happens with drivers
> > that are loaded later on?
> > 
> > Could you please elaborate on how the attributes could be created in a
> > compatible way with today's driver core?
> 
> Come on people, just look at the code!  It does exactly this today just
> fine with no changes needed to the driver core.
> 
> How about I turn it around for you, please show me how the driver core
> does _not_ support this today?  If you can prove that this isn't working
> properly, then great, I'll gladly accept patches to resolve it.

Looking at device_add():

various setup
add kobject to tree
...
call device_add_attrs
(this adds pre-specified attribute
groups, depending on dev->type,
dev->class or the device specific
attribute groups)
call bus_add_device
(add per-bus pre-specified attributes;
the bus does not know the driver yet)
...
call kobject_uevent                       <--- first time userspace knows
                                               about device
call bus_probe_device                     <--- first time a driver may
                                               actually see the device
                                               (depending on the ->match
                                                and ->probe functions,
                                                this may be quite a bit of
                                                time afterwards)

This will not be a problem if a device driver registers a child device
(since it can specify the attributes there).

I think the basic problem is that the KOBJ_ADD uevent notifies
userspace that "a device is there", while the device will only be
really useable by userspace once a driver has bound to it. A module
load triggered by KOBJ_ADD is fine, but trying to actually use the
device after KOBJ_ADD is racy. This will not matter in the usual case,
since either the matching/probing is fast enough or userspace will wait
for something like a block device anyway, but we've seen problems on
s390. A KOBJ_BIND/UNBIND would make a proper distinction between
"device is there" and "device is usable".

(Besides, what happens on unbind/bind? Shouldn't userspace know that a
device is now bound to a different driver?)

Cornelia