Re: pci_enable_msi() for everyone?

[Greg KH <gregkh@suse.de>]

On Fri, Jun 03, 2005 at 07:16:04PM -0400, Jeff Garzik wrote:
> Greg KH wrote:
> >In talking with a few people about the MSI kernel code, they asked why
> >we can't just do the pci_enable_msi() call for every pci device in the
> >system (at somewhere like pci_enable_device() time or so).  That would
> >let all drivers and devices get the MSI functionality without changing
> >their code, and probably make the api a whole lot simpler.
> 
> Agreed.
> 
> And now is a good time to make changes, before bunches of drivers start 
> using pci_enable_msi().  I am preparing such a change for the AHCI 
> driver (see attached), which will be the standard SATA interface on most 
> new motherboards.

Yes, roll that pci_intx() into the core and have it do the
pci_disable_msi() also if needed would be what I am thinking of.

> >Now I know the e1000 driver would have to specifically disable MSI for
> >some of their broken versions, and possibly some other drivers might
> >need this, but the downside seems quite small.
> 
> tg3 needs to deal with some broken system chipsets as well.  See
> tg3_test_msi(), which I would eventually prefer to eliminate in favor of 
> PCI quirks and such.

Ok, that's fine.

> An API note of warning though...   IMO eventually different drivers are 
> going to want different behavior from pci_enable_device().  IDE already 
> hacks around this, as Alan was required to do a while ago (IDE has a 
> weird PCI BAR setup sometimes, requiring care during enabling).
> 
> Longer term, I think we will need a
> 
> 	pci_enable(info on what to enable)
> 
> so that drivers can specify ahead of time "don't enable PIO, only MMIO", 
> "don't enable MMIO, only PIO", "don't use MSI", etc.  and add a 
> pci_disable() to undo all of that.

Yes, I agree, but let's start with baby steps :)

> The more we add singleton functions like pci_enable_msi(), 
> pci_set_master(), etc. the more I wish for a single function that 
> handled all those details at one atomic point.  There is a lot of 
> standard patterns that are hand-coded into every PCI driver's probe 
> functions.

Also agreed.

thanks,

greg k-h