Call for testers: Apple ATA DMA

[John Baldwin]

On Thursday 11 September 2008 06:13:47 pm Nathan Whitehorn wrote:
> John Baldwin wrote:
> > On Thursday 11 September 2008 02:32:10 pm Peter Grehan wrote:
> >   
> >> Hi John,
> >>
> >>     
> >>> So when I did the MSI stuff I had assumed (apparently incorrectly), that 
> >>>       
> > PCI 
> >   
> >>> functions would only every have 1 non-MSI interrupt (since there is only 
a 
> >>> single INTLINE config register).  Is the extra interrupt coming from OF?  
> >>>       
> > If 
> >   
> >>> so, does OF support MSI at all?  You could always change the OF PCI bus 
> >>> driver to not do MSI and use rid 1 IRQ for the OF indicated IRQ for a 
PCI 
> >>> device by having custom alloc_resource/setup_intr/teardown_intr methods.
> >>>       
> >>   Int lines on the Mac go directly into the OpenPIC, allowing as many 
> >> int sources as desired. The intline config register isn't really used, 
> >> though there is code that attempts to read the OFW interrupt properties 
> >> and then program that register to avoid messing with the PCI common 
> >> code. Unfortunately, some Mac devices ignore writes to that register :( 
> >> The G5 does support MSI.
> >>
> >>   I had sent a possible solution to Nathan (Nathan: check your junk :) 
> >> that in pci_setup_intr did something like:
> >>
> >>    if (dinfo->cfg.msi.msi_addr > 0) {
> >>      ...
> >>    } else if (dinfo->cfg.msi.msix_alloc > 0) {
> >>      ...
> >>    } else {
> >> #ifndef __powerpc__
> >>      KASSERT("No MSI or MSI-X interrupts allocated")
> >> #endif
> >>    }
> >>
> >>   There's probably a bunch of other places that need fixing but this was 
> >> an obvious one.
> >>     
> >
> > OFW should already have its own PCI bus driver, so I'd rather you give it 
its 
> > own bus_setup_intr() method that DTRT for these interrupt resources (rid > 
0 
> > and !MSI) and then calls pci_setup_intr() for the rest.  Then you don't 
have 
> > to add MD hacks to the generic PCI bus driver.
> >   
> It doesn't on PowerPC. There are a bunch of hacks done at attach-time to 
> compensate for this (see ofw_pci_fixup()). It might be nice to import 
> sparc64's PCI OFW bus code, though.

Ah, yeah, it certainly might be.  Having separate PCI bus drivers for ACPI vs. 
not (and all the various PCI bridge drivers) helped detangle a bunch of mess 
in PCI and made ACPI support a lot easier.

-- 
John Baldwin