Re: [PATCH 0/3] msi abstractions and support for altix

[Mark Maule <maule@sgi.com>]

On Thu, Dec 22, 2005 at 01:50:23PM -0700, Matthew Wilcox wrote:
> On Thu, Dec 22, 2005 at 02:38:24PM -0600, Mark Maule wrote:
> > Because on ia64 IA64_FIRST_DEVICE_VECTOR and IA64_LAST_DEVICE_VECTOR
> > (from which MSI FIRST_DEVICE_VECTOR/LAST_DEVICE_VECTOR are derived) are not
> > constants.  The are now global variables (see change to asm-ia64/hw_irq.h)
> > to allow the platform to override them.  Altix uses a reduced range of
> > vectors for devices, and this change was necessary to make assign_irq_vector()
> > to work on altix.
> 
> To be honest, I think this is just adding a third layer of paper over
> the crack in the wall.  The original code assumed x86; the ia64 port
> added enough emulation to make it look like x86 and now altix fixes a
> couple of assumptions.  I say: bleh.
> 
> What we actually need is an interface provided by the architecture that
> allocates a new irq.  I have a hankering to implement MSI on PA-RISC but
> haven't found the time ... 

Matt, Greg, et. al:

Did you guys have something in mind for a vector allocation interface?  It
seems to me that assign_irq_vector() more or less does what we want,
but what is missing is a way for the platform to prime which vectors
are available to choose from.

One possibly better solution would be to call something in the init_IRQ path
that would set up the vector pool available to assign_irq_vector().

Any opinions on this?  I would maintain that this effort should be done
independently of this patchset.

thanks
Mark