Re: [Xen-ia64-devel] vIOSAPIC and IRQs delivery

[Kenji Kaneshige <kaneshige.kenji@xxxxxxxxxxxxxx>]

Tristan Gingold wrote:
> Le Jeudi 09 Mars 2006 11:13, Kenji Kaneshige a écrit :
>
> > Hi,
> >
> > Tristan Gingold wrote:
> >
> > > Le Jeudi 09 Mars 2006 01:31, Luck, Tony a écrit :
> > >
> > > > > BTW, I think it will be *very* hard to find an ia64 platform which can
> > > > > share an IRQ line.  Maybe I am wrong but I don't know such a platform
> > > > > :-(
> > > > Hard? No.
> > > I'd like to have a vendor name and a platform name!
> > Fujitsu PRIMEQUEST can have so many (> 512) gsi sources and it
> > needs shared IRQ (vector). Note that what I'm talking is not
> > sharing IRQ lines but sharing interrupt vector between multiple
> > RTEs.
> Ok this is interesting.
> So, you don't share IRQ lines.  My question still stands!
Our server also has one IRQ line shared between multiple devices.
Here is a /proc/interrupts output of linux running on the small
partition on our server.

          CPU0       CPU1
28:          0          0          LSAPIC  cpe_poll
29:          0          0          LSAPIC  cmc_poll
30:          0          0  IO-SAPIC-level  cpe_hndlr
31:          0          0          LSAPIC  cmc_hndlr
39:          0          0  IO-SAPIC-level  acpi
48:          0          0  IO-SAPIC-level  uhci_hcd, ehci_hcd
(snip.)

> Sharing vectors is easier, since we should assign one vector to one machine 
> for optimization.  We don't want to deliver useless interrupts.
>
> > > > Expensive? Yes.
> > > >
> > > > There are only 255 irqs - minus some that the base kernel steals
> > > > for cmc, mca, timer, perfmon, etc.  So any large system with a
> > > > few hundred i/o cards _must_ share irqs.  There aren't enough
> > > > for one each.
> > > I don't agree.  You can have interrupt domains.  See "Intel Itanium
> > > Processor Family Interrupt Architecture Guide".
> > As you said, interrupt domains approach is better because
> >
> >    - Shared IRQ (vector) approach would not work for
> >      edge-triggered interrupts including MSIs.
> >
> >    - Interrupt domains approach would have much better
> >      performance
> >
> > Unfortunately, AFAIK, interrupt domains approach has not be
> > implemented for generic ia64-linux kernel yet. It would be
> > very nice if it is implemented someday. But I guess we need
> > a lot of efforts for it. To tell the truth, I've posted the
> > patches for interrupt domains to ia64-linux mailing list
> > before, though it contains many TBDs and bugs (maybe).
> Ok.
> Maybe it can be more easily done in Xen ?
Sorry, I'm new to Xen. So I don't know.

Thanks,
Kenji Kaneshige




_______________________________________________
Xen-ia64-devel mailing list
Xen-ia64-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-ia64-devel