Re: RFC: (re-)binding the VFIO platform driver to a platform device

[Kim Phillips]

On Wed, 2 Oct 2013 11:43:30 -0700
Christoffer Dall <christoffer.dall@xxxxxxxxxx> wrote:

> On Wed, Oct 02, 2013 at 01:32:38PM -0500, Scott Wood wrote:
> > On Wed, 2013-10-02 at 13:25 -0500, Yoder Stuart-B08248 wrote:
> > > 
> > > > -----Original Message-----
> > > > From: Christoffer Dall [mailto:christoffer.dall@xxxxxxxxxx]
> > > > Sent: Wednesday, October 02, 2013 10:14 AM
> > > > To: Alex Williamson
> > > > Cc: Kim Phillips; gregkh@xxxxxxxxxxxxxxxxxxx; linux-
> > > > kernel@xxxxxxxxxxxxxxx; a.motakis@xxxxxxxxxxxxxxxxxxxxxx; agraf@xxxxxxx;
> > > > Yoder Stuart-B08248; Wood Scott-B07421; Sethi Varun-B16395; Bhushan
> > > > Bharat-R65777; peter.maydell@xxxxxxxxxx; santosh.shukla@xxxxxxxxxx;
> > > > kvm@xxxxxxxxxxxxxxx
> > > > Subject: Re: RFC: (re-)binding the VFIO platform driver to a platform
> > > > device
> > > > 
> > > > Wouldn't a sysfs file to add compatibility strings to the vfio-platform
> > > > driver make driver_match_device return true and make everyone happy?
> > > 
> > > I had a similar thought.  Why can't we do something like:
> > > 
> > >   echo "fsl,i2c" > /sys/bus/platform/drivers/vfio-platform/new_compatible
> > >   echo 12ce0000.i2c > /sys/bus/platform/drivers/vfio-platform/bind
> > > 
> > > The first steps tell vfio-platform to register itself to handle
> > > "fsl,i2c" compatible devices.  The second step does the bind.
> > 
> > Needing to specify the compatible is hacky (we already know what device
> > we want to bind; why do we need to scrounge up more information than
> > that, and add a new sysfs interface for extending compatible matches,
> > and a more flexible data structure to back that up?), and is racy on
> > buses that can hotplug (which driver gets the new device?).
> 
> Why hacky?  It seems quite reasonable to me that the user has to tell a
> subsystem that from a certain point it should be capable of handling
> some device.

I think what Scott is saying is that the first echo above is somewhat
redundant with the second: they're both talking to the vfio-platform
driver about an i2c device.

> As for the data structure, isn't this a simple linked list?
> 
> The problem with the race seems to be a common problem that hasn't even
> been solved for PCI yet, so I'm wondering if this is not an orthogonal
> issue with a separate solution, such as a priority or something like
> that.

I agree; adding 'direct'/'atomic' functionality to the existing bind
sysfs file, i.e., bind_store() logic to perform device_release_driver()
logic if dev->driver != NULL, all under the same device_lock() is an
independent problem from binding the VFIO platform driver to a platform
device.

> Yes, once you've added the new_compatible to the vfio-platform driver,
> it's up for grabs from both the new and the old driver, but that could
> be solved by always making sure that the vfio-platform driver is checked
> first.

not sure I understand the priority problem here - haven't looked at PCI
yet - but, given the above 'atomic bind' functionality described above,
the new and old driver wouldn't be requesting to bind to the same
device simultaneously, no?

> (I'm not familiar with these data structures, but I would imagine
> something like re-inserting the vfio-platform driver in the
> list/tree/... whenever adding a new_compatible value might possibly be
> one solution).
> 
> > What's wrong with a non-vfio-specific flag that a driver can set, that
> > indicates that the driver is willing to try to bind to any device on the
> > bus if explicitly requested via the existing sysfs bind mechanism?
> > 
> It sounds more hackish to me to invent some 'generic' flag to solve a
> very specific case.  What you're suggesting would let users specify that
> a serial driver should handle a NIC hardware, no?  That sounds much much
> worse to me.

I thought that was the nature of VFIO drivers...it's a 'meta-' driver,
used for enabling userspace drivers at large.

Kim
--
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/