On Wednesday, December 12, 2012 06:00:19 PM Bjorn Helgaas wrote:
> On Sun, Dec 9, 2012 at 4:00 PM, Rafael J. Wysocki <rjw@xxxxxxx> wrote:
> > From: Rafael J. Wysocki <rafael.j.wysocki@xxxxxxxxx>
> >
> > Devices created by acpi_create_platform_device() sometimes may need
> > to be added to the device hierarchy as children of PCI bridges.
>
> An example of this hierarchy would help to understand it.
Well, that's only hypothetical, but it goes like this.
There's a PCIe root complex with root ports. Some devices are connected to
those root ports, but for whatever (the heck) the reason they are not visible
to the kernel as PCI devices (ie. their config spaces are unavailable by any
usual means). They are, however, listed in the ACPI namespace and their
MMIO ranges can be extracted through _CRS. Still, the root ports themselves
are visible to the kernel as normal PCIe devices (I know, that's silly).
So, in order to maintain (for example) the correct runtime suspend ordering,
we need to represent those "hidden" devices as device nodes being the children
of the PCIe root ports' device nodes.
>
> > For this purpose, however, the struct pci_dev objects representing those
> > bridges need to exist before the platform devices in question are
> > added, but this is only possible if the PCI root bridge driver is
> > registered before the initial scanning of the ACPI namespace
> > (that driver's .add() routine creates the required struct pci_dev
> > objects).
>
> The previous patch (1/6) registers all acpi_device objects in the
> first pass, then calls driver .add() methods in the second pass. And
> here you're saying the .add() method has to run before platform
> devices are added. So I guess the acpi_device objects are added
> first, then the .add() methods called, then the platform devices
> added?
No, the platform devices are added in the same pass in which the .add()
methods are called, because we know that we won't call an .add() method
for any struct acpi_device that we'll be creating a platform device for.
These things are mutually exclusive.
Of course, we could add a separate pass for adding the platform devices,
but then we'd need to teach ACPI PNP that it shouldn't create PNP device
objects for the ACPI devices we'll be creating platform devices for.
> I think the call sequence looks like this:
>
> acpi_bus_scan
> acpi_walk_namespace
> acpi_bus_check_add
> acpi_add_single_object
> device = kzalloc(sizeof(struct acpi_device, ...) # (1)
> acpi_devices created here
> acpi_walk_namespace
> acpi_bus_match_device
> if (acpi_match_device_ids(device, acpi_platform_device_ids))
> acpi_create_platform_device # (3) platform device added here
> else
> device_attach # (2) driver .add() called here
> acpi_hot_add_bind
>
> (1) happens first because it's in the first acpi_walk_namespace().
Yes.
> (2) happens before (3) because acpi_walk_namespace() calls
> acpi_bus_match_device() in preorder (node visited before its children)
Yes.
> It always seems like a bit of a hack when we have to call out a driver
> specifically like this. Are these special platform devices unique to
> PCI?
I believe so.
> What would happen with these platform devices that are children
> of PCI bridges if we booted a kernel without a PCI host bridge driver?
> You would hope that the PCI host bridge and everything under it would
> just be ignored, and I assume that in that case, these platform
> devices should be ignored, too.
That's a good question and I believe the answer depends on what's there
in the ACPI namespace. Namely, if the ACPI namespace lists the PCI parents
of those platform devices, we should at least enable them to decode resources
for the children (which we should be doing in general for any parents
listed in the ACPI namespace).
> I know we can't build ACPI without PCI today, but AFAIK that's mostly
> to reduce the configuration/testing matrix, not a design restriction.
> So I guess I'm trying to figure out whether the ACPI core should be
> made smart enough to deal with these PCI-related platform devices (as
> you're doing in these patches), or whether there should be something
> in the PCI host bridge driver that deals with them.
Well, I'm actually not making the ACPI core any smarter. :-)
What I'm really doing is just to say "let's register the PCI root bridge
driver before walking the namespace, because we may want it to kick in
before anything else". And if it doesn't kick in at all, so be it.
That also makes the "boot" case be more similar to the "hotplug" case in which
the root bridge driver is always present, so I think this is worth doing
anyway.
And moreover, I think we should register _all_ ACPI drivers (except for modular
ones, but those are not really suitable for hotplug anyway) before walking the
namespace to make the "boot" and "hotplug" cases look the same. This patch
just does this for the root bridge on the "one item at a time" basis, but I
don't see a reason why not to do that for other ACPI drivers.
Thanks,
Rafael
> > For this reason, call acpi_pci_root_init() from acpi_scan_init()
> > before scanning the ACPI namespace for the first time instead of
> > running it from a separate subsys initcall. Since the previous patch
> > has changed the ACPI namespace scanning algorithm, this change does
> > not affect the PCI root bridge driver's functionality during boot.
> > It also makes the situation during boot more similar to the situation
> > during hot-plug (in which the PCI root bridge driver is always
> > present) and so it helps to reduce arbitary differences between
> > the hot-plug and boot PCI root bridge code.
> >
> > Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@xxxxxxxxx>
> > ---
> > drivers/acpi/internal.h | 1 +
> > drivers/acpi/pci_root.c | 4 +---
> > drivers/acpi/scan.c | 1 +
> > 3 files changed, 3 insertions(+), 3 deletions(-)
> >
> > Index: linux/drivers/acpi/internal.h
> > ===================================================================
> > --- linux.orig/drivers/acpi/internal.h
> > +++ linux/drivers/acpi/internal.h
> > @@ -67,6 +67,7 @@ struct acpi_ec {
> >
> > extern struct acpi_ec *first_ec;
> >
> > +int acpi_pci_root_init(void);
> > int acpi_ec_init(void);
> > int acpi_ec_ecdt_probe(void);
> > int acpi_boot_ec_enable(void);
> > Index: linux/drivers/acpi/pci_root.c
> > ===================================================================
> > --- linux.orig/drivers/acpi/pci_root.c
> > +++ linux/drivers/acpi/pci_root.c
> > @@ -674,7 +674,7 @@ static int acpi_pci_root_remove(struct a
> > return 0;
> > }
> >
> > -static int __init acpi_pci_root_init(void)
> > +int __init acpi_pci_root_init(void)
> > {
> > acpi_hest_init();
> >
> > @@ -687,5 +687,3 @@ static int __init acpi_pci_root_init(voi
> >
> > return 0;
> > }
> > -
> > -subsys_initcall(acpi_pci_root_init);
> > Index: linux/drivers/acpi/scan.c
> > ===================================================================
> > --- linux.orig/drivers/acpi/scan.c
> > +++ linux/drivers/acpi/scan.c
> > @@ -1828,6 +1828,7 @@ int __init acpi_scan_init(void)
> > }
> >
> > acpi_power_init();
> > + acpi_pci_root_init();
> >
> > /*
> > * Enumerate devices in the ACPI namespace.
> >
--
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
