On Thu, 16 Jul 2015 18:32:28 +0100
Marc Zyngier <marc.zyngier@xxxxxxx> wrote:
> On 16/07/15 18:14, David Daney wrote:
> > On 07/16/2015 10:09 AM, Marc Zyngier wrote:
> >> On 16/07/15 17:50, David Daney wrote:
> > [...]
> >>>> Patch 5 has established that you're using "virtual wire" SPIs, so we
> >>>> need to work on exposing that with the normal kernel abstraction, and
> >>>> not by messing with the internals of the GIC.
> >>>>
> >>>
> >>> Agreed.
> >>>
> >>> The MSI system has pci_enable_msix()/pci_disable_msix().
> >>>
> >>> I would propose something like:
> >>>
> >>> struct gic_spi_entry {
> >>> int spi /* SPI number */
> >>> int irq; /* kernel irq number mapped to the spi*/
> >>> u32 msg; /* message to be written */
> >>> u64 assert_addr;
> >>> u64 deassert_addr;
> >>> };
> >>>
> >>> /* Fill in the SPI processing information */
> >>> int gic_map_spi(int spi, struct gic_spi_entry *data);
> >>
> >> Neither.
> >>
> >> The way to do it is to make this a *separate* IRQ domain stacked onto
> >> the SPI domain. No funky hook on the side. If it doesn't go through the
> >> normal kernel API, it doesn't reach the GIC.
> >
> > Yes, the irqdomain does handle mapping SPI -> irq, and the message can
> > be derived from the SPI. However, the irqdomain infrastructure cannot
> > supply values for either assert_addr or deassert_addr.
>
> This is why I suggested earlier (in my reply to patch 5) that you have a
> look at the series I posted a couple of days ago to implement non-PCI
> MSI support. This would allow you to compose the domains as such:
>
> platform-MSI -> message-SPI -> GIC
>
> You'd end up with a msi_msg containing the GICD_SETSPI_NSR doorbell, and
> the SPI as a payload.
>
> > Those are needed in order to use SPI. How would you suggest that they
> > be obtained?
>
> Two possibilities: either you derive GICD_CLRSPI_NSR by adding 8 to the
> doorbell you got from the msi_msg structure (ugly, but limited to your
> own code), or you extend msi_msg to cater for this case.
A simpler alternative to the above would be to move all this to
firmware, and only expose a *wired* SPI to the kernel. This would have
the advantage to use the existing infrastructure for both DT and ACPI.
As a side note, that was the intended use of these registers - to
provide a "virtual wire" interface that remains mostly invisible to
software.
Thanks,
M.
--
Jazz is not dead. It just smells funny.
--
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
