On Wednesday, December 12, 2012 11:50:06 PM Jiang Liu wrote:
> On 12/10/2012 07:00 AM, Rafael J. Wysocki wrote:
> > From: Rafael J. Wysocki <rafael.j.wysocki@xxxxxxxxx>
> >
> > Currently, as soon as an ACPI device node object (struct acpi_device)
> > is created, the driver core attempts to probe ACPI drivers against
> > it. That leads to some unpleasant side effects, like the fact that
> > the boot code path for ACPI namespace scanning is different from the
> > analogous hot-plug code path (during boot ACPI drivers are not
> > present when ACPI device node objects are registered, so they are
> > guaranteed not to be probed, which is not the case during hot-plug).
> > That, in turn, leads to unnecessary complications in the PCI
> > enumeration algorithm.
> >
> > Reduce the differences between the boot and hot-plug cases by
> > splitting the ACPI namespace scanning for devices into two passes,
> > such that struct acpi_device objects are registerd in the first
> > patch without probing ACPI drivers and the drivers are probed
> > against them directly in the second pass. This way ACPI drivers
> > can assume that all of the ACPI device node objects in the given
> > scope will be registered when their .add() routines run and the
> > hot-plug case becomes the same as the boot case from their
> > perspective.
> >
> > Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@xxxxxxxxx>
> > ---
> > drivers/acpi/scan.c | 96 +++++++++++++++++++++++++++++++++---------------
> > include/acpi/acpi_bus.h | 1
> > 2 files changed, 68 insertions(+), 29 deletions(-)
> >
> > Index: linux/include/acpi/acpi_bus.h
> > ===================================================================
> > --- linux.orig/include/acpi/acpi_bus.h
> > +++ linux/include/acpi/acpi_bus.h
> > @@ -98,6 +98,7 @@ typedef void (*acpi_op_notify) (struct a
> > struct acpi_bus_ops {
> > u32 acpi_op_add:1;
> > u32 acpi_op_start:1;
> > + u32 acpi_op_match:1;
> > };
> >
> > struct acpi_device_ops {
> > Index: linux/drivers/acpi/scan.c
> > ===================================================================
> > --- linux.orig/drivers/acpi/scan.c
> > +++ linux/drivers/acpi/scan.c
> > @@ -494,7 +494,8 @@ static int acpi_bus_match(struct device
> > struct acpi_device *acpi_dev = to_acpi_device(dev);
> > struct acpi_driver *acpi_drv = to_acpi_driver(drv);
> >
> > - return !acpi_match_device_ids(acpi_dev, acpi_drv->ids);
> > + return acpi_dev->bus_ops.acpi_op_match
> > + && !acpi_match_device_ids(acpi_dev, acpi_drv->ids);
> > }
> Hi Rafael,
> PCI host bridge hotplug has the same requirement to separate device
> enumeration from device driver binding. And VFIO has a similar requirement too.
> Yinghai and I have implemented two different solutions for PCI host bridge
> hotplug but all rejected by Greg. So it would be great if we could promote
> a common mechanism to the device core to temporarily disable binding drivers
> to devices, which could used to support ACPI hotplug, PCI hotplug and VFIO.
OK, but let's first have a good common use case, I think. I mean, let's
implement it in each of these subsystems separately and then show that it
leads to simpler code if we move it up to the driver core.
Thanks,
Rafael
--
I speak only for myself.
Rafael J. Wysocki, Intel Open Source Technology Center.
--
To unsubscribe from this list: send the line "unsubscribe linux-pci" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
