Re: [PATCH v2 02/13] vfio/mdev: Allow the mdev_parent_ops to specify the device driver to bind

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On Wed, Apr 28, 2021 at 12:58:29PM -0700, Dan Williams wrote:
> On Wed, Apr 28, 2021 at 7:00 AM Jason Gunthorpe <jgg@xxxxxxxxxx> wrote:
> >
> > On Wed, Apr 28, 2021 at 02:41:53PM +0200, Christoph Hellwig wrote:
> > > On Wed, Apr 28, 2021 at 12:56:21AM -0700, Dan Williams wrote:
> > > > > I still think this going the wrong way.  Why can't we enhance the core
> > > > > driver code with a version of device_bind_driver() that does call into
> > > > > ->probe?  That probably seems like a better model for those existing
> > > > > direct users of device_bind_driver or device_attach with a pre-set
> > > > > ->drv anyway.
> > > >
> > > > Wouldn't that just be "export device_driver_attach()" so that drivers
> > > > can implement their own custom bind implementation?
> > >
> > > That looks like it might be all that is needed.
> >
> > I thought about doing it like that, it is generally a good idea,
> > however, if I add new API surface to the driver core I really want to
> > get rid of device_bind_driver(), or at least most of its users.
> 
> I might be missing where you are going with this comment, but
> device_driver_attach() isn't a drop-in replacement for
> device_bind_driver().

Many of the places calling device_bind_driver() are wonky things
like this:

        dev->dev.driver = &drv->link.driver;
        if (pnp_bus_type.probe(&dev->dev))
                goto err_out;
        if (device_bind_driver(&dev->dev))
                goto err_out;

So device_driver_attach() does replace that - with some differences.

Notable is that bind_driver requires the driver_lock but driver_attach
gets it internally. However, as far as I can tell, none of the
bind_driver callers do get it, so huh.

Aside from the driver_lock there are lots of small subtle differences
that are probably not important unless they are for some very complex
reason. :\

Of the callers:
  drivers/input/serio/serio.c
    This definitely doesn't have the device_lock
    It uses connect instead of probe and for some reason uses its own
    mutex instead of the device_lock. Murky.

  drivers/input/gameport/gameport.c
    This looks alot like serio, same comments

  drivers/net/phy/phy_device.c
    device_driver_attach() is better, looks unlikely that
    device_lock is properly held here. Little unclear on what
    the bus is and if bus->probe will be OK

  drivers/net/wireless/mac80211_hwsim.c
    Definitely does not hold the driver lock, the class and the driver
    have NULL probes so this could be changed

  drivers/pnp/card.c
    device_driver_attach() is better, very unlikely that a random
    device pulled from a linked list has the driver_lock held

  drivers/usb/core/driver.c
    This comment says the caller must have the device lock, but it
    doesn't call probe, and when I look at cdc_ether.c I wonder
    where the device_lock is hidden? Murky.

Basically, there is some mess here, and eliminating
device_bind_driver() for device_driver_attach() is quite a reasonable
cleanup. But hard, complex enough it needs testing each patch.

The other driver self bind scenario is to directly assign driver
before device_add, but I have a hard time finding those cases in the
tree with grep.

> If this export prevented a new device_bind_driver() user, I think
> that's a net positive, because device_bind_driver() seems an odd way
> to implement bus code to me.

Yes, I looked into why it is like this and concluded it is just very
very old.
 
> I have an ulterior motive / additional use case in mind here which is
> the work-in-progress cleanup of the DSA driver. It uses the driver
> model to assign an engine to different use cases via driver binding.
> However, it currently has a custom bind implementation that does not
> operate like a typical /sys/bus/$bus/drivers interface. If
> device_driver_attach() was exported then some DSA compat code could
> model the current way while also allowing a transition path to the
> right way. As is I was telling Dave that the compat code would need to
> be built-in because I don't think fixing a DSA device-model problem is
> enough justification on its own to ask for a device_driver_attach()
> export.

Can you make and test a DSA patch? If we have two concrete things and
I can sketch two more out of the above that should meet Greg's "need 4
things" general thinking for driver core API changes.

But I still would like to keep this going while we wait for acks, you
know how long that can take...

Jason



[Index of Archives]     [KVM ARM]     [KVM ia64]     [KVM ppc]     [Virtualization Tools]     [Spice Development]     [Libvirt]     [Libvirt Users]     [Linux USB Devel]     [Linux Audio Users]     [Yosemite Questions]     [Linux Kernel]     [Linux SCSI]     [XFree86]

  Powered by Linux