ACPI vs PCI: configuration space

ACPI vs PCI: configuration space

[Matthew Wilcox]

OK, I've been bashing my head against this for two days,
it's time to get more eyes on the problem.  The problem is in
acpi_os_read_pci_configuration() and acpi_os_write_pci_configuration().
These functions can be called for busses which have not yet been scanned
(and therefore do not have a corresponding pci_bus).  The code in 2.5
doesn't work for ia64 beause we need a valid ->sysdata pointer to handle
PCI domains.

The current patch in the ia64 tree does this:

        struct pci_bus          bus;
+#ifdef CONFIG_IA64
+       struct pci_controller   ctrl;
+#endif
 
...
        bus.number = pci_id->bus;
+#ifdef CONFIG_IA64
+       ctrl.segment = pci_id->segment;
+       bus.sysdata = &ctrl;
+#endif
        result = pci_root_ops->read(&bus, PCI_DEVFN(pci_id->device,

I think we can all agree that's ugly.  But there's no _clear_ way to
improve this.

Try 1: Ask the architecture code to provide a sysdata for us.  That's bad;
it needs to allocate with GFP_ATOMIC (since this code can be called from
interrupt context).  So it can fail on low mem conditions.

Try 2: Define a `struct pci_controller' on architectures that don't have it.
And a macro pci_set_domain() so x86 can have a zero-length pci_controller.
Not the prettiest idea, but best of this batch.

Try 3: Move the segment/domain into the pci_bus.  This already got NAKed
by a few people.

Try 4: redefine the pci_ops again.  Haha, very funny.

It's kind of annoying to invent some structures and put some values into
them only to pull them out again.  This leads to try 5 ...

struct acpi_pci_ops {
	int (*read)(int domain, int bus, int devfn, int where, int size, u32 *val);
	int (*write)(int domain, int bus, int devfn, int where, int size, u32 val);
}

It reduces stack consumption, which is a clear win ... it's also _incredibly_
easy to implement since all the existing pci_ops call functions which take
exactly this form.

Go on, approve Try 5.  You know you want to ;-)

-- 
"It's not Hollywood.  War is real, war is primarily not about defeat or
victory, it is about death.  I've seen thousands and thousands of dead bodies.
Do you think I want to have an academic debate on this subject?" -- Robert Fisk

Re: ACPI vs PCI: configuration space

[Greg KH]

On Wed, Jun 18, 2003 at 11:17:52PM +0100, Matthew Wilcox wrote:
> 
> It's kind of annoying to invent some structures and put some values into
> them only to pull them out again.  This leads to try 5 ...
> 
> struct acpi_pci_ops {
> 	int (*read)(int domain, int bus, int devfn, int where, int size, u32 *val);
> 	int (*write)(int domain, int bus, int devfn, int where, int size, u32 val);
> }
> 
> It reduces stack consumption, which is a clear win ... it's also _incredibly_
> easy to implement since all the existing pci_ops call functions which take
> exactly this form.
> 
> Go on, approve Try 5.  You know you want to ;-)

So for i386, what would domain be?

Anyway, yeah, I agree with try 5, that seems the most sane.

thanks,

greg k-h

Re: ACPI vs PCI: configuration space

[Matthew Wilcox]

On Wed, Jun 18, 2003 at 03:30:03PM -0700, Greg KH wrote:
> So for i386, what would domain be?

The `seg' argument currently in place to all these macros ;-)

> Anyway, yeah, I agree with try 5, that seems the most sane.

Yay.  you'll have a patch tomorrow.

-- 
"It's not Hollywood.  War is real, war is primarily not about defeat or
victory, it is about death.  I've seen thousands and thousands of dead bodies.
Do you think I want to have an academic debate on this subject?" -- Robert Fisk