On Tue, 2009-02-10 at 22:53 -0700, Matthew Wilcox wrote:
> I've been working on support for systems with large numbers of MSI-X
> vectors. Here's the architecture interface I'm currently working with:
>
> extern int arch_msi_check_device(struct pci_dev *, int nvec, int type);
> extern int arch_setup_msi_irqs(struct pci_dev *, int nvec);
> extern int arch_setup_msix_irq(struct pci_dev *, int irq);
> extern void arch_teardown_msi_irqs(struct pci_dev *);
>
> and the MSI-X code looks like this:
>
> for (i = 0; i < nvec; i++) {
> unsigned irq, table_entry = entries[i].entry;
> struct msi_desc desc;
> ret = arch_setup_msix_irq(dev, table_entry);
>
> I now perceive this would throw a bit of a spanner in the works for ppc.
If that was the generic code then yes. We can't (sanely) allocate them
one at a time.
> How about the following API?
>
> /*
> * Set up nvec MSI interrupts. Returns 0 on success, a positive number if
> * it could have set up fewer interrupts than requested, or -Errno for other
> * errors
> */
> int arch_msi_setup_irqs(struct pci_dev *, int nvec);
> void arch_msi_free_irqs(struct pci_dev *);
>
> /*
> * Allocate nvec MSI-X interrupts. Returns 0 on success, a positive number if
> * it could have set up fewer interrupts than requested, or -Errno for other
> * errors
> */
> int arch_msix_alloc_irqs(struct pci_dev *, int nvec);
>
> /*
> * Program the @table_entry MSI-X interrupt on the card. Returns the Linux
> * interrupt number on success or -Errno if an error occurs.
> */
> int arch_msix_setup_irq(struct pci_dev *, int table_entry);
> void arch_msix_free_irqs(struct pci_dev *);
>
> There's a lot more I can say about the heavily patched msi.c file I have
> here ... but perhaps I'll save that for another day.
I'm not quite sure I follow. Your MSI-X routine is still setting up just
one irq, that won't work for us.
I don't understand why you'd want a different interface for MSI vs
MSI-X, rather than just passing the type as we currently do.
I say that because in general the arch code doesn't care whether it's
MSI or MSI-X, it's just an irq. And it's simpler to have one interface
with a type argument for the rare case where it matters. Looking at the
current implementations, type is used exactly once.
arch/x86/kernel/io_apic.c
arch/powerpc/platforms/cell/axon_msi.c
arch/powerpc/sysdev/fsl_msi.c
arch/powerpc/sysdev/mpic_pasemi_msi.c
arch/powerpc/platforms/pseries/msi.c <- uses type
And that's only because we have a different argument to the firmware
call for MSI-X. The rest of the code is the same.
I take it you're planning on replacing the list of msi_descs with
something more compact? It seems to me we could do that fairly easily
without changing the arch interface much at all - just replace the list
with some per-pci_dev attributes and then an array of per-MSI info.
But I haven't thought about lots of MSI-X much, so perhaps there are
some requirements I'm missing.
cheers
--
Michael Ellerman
OzLabs, IBM Australia Development Lab
wwweb: http://michael.ellerman.id.au
phone: +61 2 6212 1183 (tie line 70 21183)
We do not inherit the earth from our ancestors,
we borrow it from our children. - S.M.A.R.T Person
Attachment:
signature.asc
Description: This is a digitally signed message part
