Re: [PATCH 10/14] PCI: tegra: Move PCIe driver to drivers/pci/host

[Thierry Reding <thierry.reding-RM9K5IK7kjKj5M59NBduVrNAH6kLmebB@public.gmane.org>]

[-- Attachment #1: Type: text/plain, Size: 3171 bytes --]

On Mon, Jan 14, 2013 at 09:57:07AM +0000, Andrew Murray wrote:
> On Sun, Jan 13, 2013 at 09:58:06AM +0000, Thierry Reding wrote:
> > On Sat, Jan 12, 2013 at 09:12:25PM +0000, Arnd Bergmann wrote:
> > > On Saturday 12 January 2013, Thierry Reding wrote:
> > > > > I already hinted at that in one of the other subthreads. Having such a
> > > > > multiplex would also allow the driver to be built as a module. I had
> > > > > already thought about this when I was working on an earlier version of
> > > > > these patches. Basically these would be two ops attached to the host
> > > > > bridge, and the generic arch_setup_msi_irq() could then look that up
> > > > > given the struct pci_dev that is passed to it and call this new per-
> > > > > host bridge .setup_msi_irq().
> > > > 
> > > > struct pci_ops looks like a good place to put these. They'll be
> > > > available from each struct pci_bus, so should be easy to call from
> > > > arch_setup_msi_irq().
> > > > 
> > > > Any objections?
> > > > 
> > > 
> > > struct pci_ops has a long history of being specifically about
> > > config space read/write operations, so on the one hand it does
> > > not feel like the right place to put interrupt specific operations,
> > > but on the other hand, the name sounds appropriate and I cannot
> > > think of any other place to put this, so it's fine with me.
> > > 
> > > The only alternative I can think of is to introduce a new
> > > structure next to it in struct pci_bus, but that feels a bit
> > > pointless. Maybe Bjorn has a preference one way or the other.
> > 
> > The name pci_ops is certainly generic enough. Also the comment above the
> > structure declaration says "Low-level architecture-dependent routines",
> > which applies to the MSI functions as well.
> 
> I've previously looked into this. It seems that architectures handle this
> in different ways, some use vector tables, others use a multiplex and others
> just let the end user implement the callback directly.
> 
> I've made an attempt to find a more common way. Though my implementation, which
> I will try to share later today for reference provides a registration function
> in drivers/pci/msi.c to provide implementations of the
> (setup|teardown)_msi_irq(s) ops. This seems slightly better than the current
> approach and doesn't break existing users - but is still ugly.
> 
> At present the PCI and MSI frameworks are largely uncoupled from each other and
> so I was keen to not pollute PCI structures (e.g. pci_ops) with MSI ops. Just
> because most PCI host bridges also provide MSI support I don't think there is a
> reason why they should always come as a pair or be provided by the same chip.
> 
> Perhaps the solution is to support MSI controller drivers and a means to
> associate them with PCI host controller drivers?

I'm not sure I follow you're reasoning here. Is it possible to use MSIs
without PCI? If not then I think there's little sense in keeping the
implementations separate.

Furthermore, if MSI controller and PCI host bridge are separate entities
how do you look up the MSI controller given a PCI device?

Thierry

[-- Attachment #2: Type: application/pgp-signature, Size: 836 bytes --]