Re: [RFC PATCH] PCI: Introduce new device binding path using pci_dev.driver_override

[Greg KH <gregkh@linuxfoundation.org>]

On Tue, Apr 01, 2014 at 06:52:12PM -0500, Kim Phillips wrote:
> On Tue, 01 Apr 2014 10:28:54 -0600
> Alex Williamson <alex.williamson@redhat.com> wrote:
> 
> > The driver_override field allows us to specify the driver for a device
> > rather than relying on the driver to provide a positive match of the
> > device.  This shortcuts the existing process of looking up the vendor
> > and device ID, adding them to the driver new_id, binding the device,
> > then removing the ID, but it also provides a couple advantages.
> > 
> > First, the above process allows the driver to bind to any device
> > matching the new_id for the window where it's enabled.  This is often
> > not desired, such as the case of trying to bind a single device to a
> > meta driver like pci-stub or vfio-pci.  Using driver_override we can
> > do this deterministically using:
> > 
> > echo pci-stub > /sys/bus/pci/devices/0000:03:00.0/driver_override
> > echo 0000:03:00.0 > /sys/bus/pci/devices/0000:03:00.0/driver/unbind
> > echo 0000:03:00.0 > /sys/bus/pci/drivers_probe
> > 
> > Previously we could not invoke drivers_probe after adding a device
> > to new_id for a driver as we get non-deterministic behavior whether
> > the driver we intend or the standard driver will claim the device.
> > Now it becomes a deterministic process, only the driver matching
> > driver_override will probe the device.
> > 
> > To return the device to the standard driver, we simply clear the
> > driver_override and reprobe the device, ex:
> > 
> > echo > /sys/bus/pci/devices/0000:03:00.0/preferred_driver
> > echo 0000:03:00.0 > /sys/bus/pci/devices/0000:03:00.0/driver/unbind
> > echo 0000:03:00.0 > /sys/bus/pci/drivers_probe
> > 
> > Another advantage to this approach is that we can specify a driver
> > override to force a specific binding or prevent any binding.  For
> > instance when an IOMMU group is exposed to userspace through VFIO
> > we require that all devices within that group are owned by VFIO.
> > However, devices can be hot-added into an IOMMU group, in which case
> > we want to prevent the device from binding to any driver (preferred
> > driver = "none") or perhaps have it automatically bind to vfio-pci.
> > With driver_override it's a simple matter for this field to be set
> > internally when the device is first discovered to prevent driver
> > matches.
> > 
> > Signed-off-by: Alex Williamson <alex.williamson@redhat.com>
> > ---
> > 
> > Apologies for the exceptionally long cc list, this is a follow-up to
> > Stuart's "Subject: mechanism to allow a driver to bind to any device"
> > thread.  This is effectively a v2 of the proof-of-concept patch I
> > posted in that thread.  This version changes to use a dummy id struct
> > to return on an "override" match, which removes the collateral damage
> > and greatly simplifies the patch.  This feels fairly well baked for
> > PCI and I would expect that platform drivers could do a similar
> > implementation.  From there perhaps we can discuss whether there's
> > any advantage to placing driver_override on struct device.  The logic
> > for incorporating it into the match still needs to happen per bus
> > driver, so it might only contribute to consistency of the show/store
> > sysfs attributes to move it up to struct device.  Please comment.
> 
> Sounds like Greg likes this approach more than {drv,dev}_sysfs_only.

I have made no such judgement, I only pointed out that if you
modify/add/remove a sysfs file, it needs to have documentation for it.

> The diff below is the result of duplicating and converting this patch
> for platform devices, and, indeed, binding a device to the
> vfio-platform driver succeeds with:
> 
> echo vfio-platform > /sys/bus/platform/devices/fff51000.ethernet/driver_override
> echo fff51000.ethernet > /sys/bus/platform/devices/fff51000.ethernet/driver/unbind
> echo fff51000.ethernet > /sys/bus/platform/drivers_probe
> 
> However, it's almost pure duplication modulo the bus match code.  The
> only other place I can see where to put the common bus check is
> drivers/base/base.h:driver_match_device(), which I'm guessing is
> off-limits?  So should we leave this as per-bus code, and somehow
> refactor driver_override_{show,store}?

If you can provide a way for this to be done in a bus-independant way,
like we did for new_id and the like, I'd be open to reviewing it.


greg k-h