From mboxrd@z Thu Jan 1 00:00:00 1970 From: Mark Maule Date: Tue, 03 Jan 2006 03:22:49 +0000 Subject: Re: [PATCH 0/3] msi abstractions and support for altix Message-Id: <20060103032249.GA4957@sgi.com> List-Id: References: <20051222201651.2019.37913.96422@lnx-maule.americas.sgi.com> <20051222202259.GA4959@suse.de> <20051222202627.GI17552@sgi.com> <20051222203415.GA28240@suse.de> <20051222203824.GJ17552@sgi.com> <20051222205023.GK2361@parisc-linux.org> In-Reply-To: <20051222205023.GK2361@parisc-linux.org> MIME-Version: 1.0 Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: 7bit To: Matthew Wilcox Cc: Greg KH , linuxppc64-dev@ozlabs.org, linux-pci@atrey.karlin.mff.cuni.cz, linux-ia64@vger.kernel.org, linux-kernel@vger.kernel.org, Tony Luck 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