Re: Physical and dynamic event channels

Re: Physical and dynamic event channels

[Keir Fraser]

On 11/8/06 12:41 am, "Jeremy Fitzhardinge" <jeremy@goop.org> wrote:

> In the current Xen patches, it reserves a range of 256 irqs for 1:1
> mapping with hardware irqs, and another 256 for dynamically allocated
> event channels.
> 
> Is this really necessary?  256 irqs for hardware interrupts is excessive
> in itself, because there can only be 224.  But aside from that, is it
> necessary to have a 1:1 mapping?  Couldn't they be dynamic as well?  Do
> some subset of "historic" irqs need to be 1:1 mapped, but everything
> else is up for grabs?
> 
> How many event channels do guests end up using anyway?  Would 256 irqs
> be enough general use?

256 would be enough for anyone I think. domain0 is obviously the main user
of event channels but most of those are demuxed to userspace rather than the
kernel IRQ space.

Adding a level of indirection for non-domain-0 wouldn't be too hard. We have
control over which IRQ a PCI device driver thinks it is binding to because
we control the PCI config space via pciback. We need more care for older
devices, but could arrange that by maintaining a 1:1 mapping for legacy PIC
irq range (0 to 15).

Doing the same for domain 0 is probably more 'exciting'. If we could force
the 'vector space' IRQ allocation strategy of PCI_MSI then I think this is
more plausible -- since that effectively gives a level of indirection from
the GSI space. The vectors that Xen currently hands out to domain0 are real
vector numbers but, of course, it would be trivial to add a level of
indirection there, or in the caller.

This all raises an obvious question: what do you think the scope of work for
upstream merging right now should be? Previously it was non-driver domUs
only (and a simplified form at that). Are you thinking about upstreaming
everything, or is this just review and preparation for that effort in the
future?

 -- Keir

Re: Physical and dynamic event channels

[Jeremy Fitzhardinge]

Keir Fraser wrote:
> 256 would be enough for anyone I think. domain0 is obviously the main user
> of event channels but most of those are demuxed to userspace rather than the
> kernel IRQ space.
>
> Adding a level of indirection for non-domain-0 wouldn't be too hard. We have
> control over which IRQ a PCI device driver thinks it is binding to because
> we control the PCI config space via pciback. We need more care for older
> devices, but could arrange that by maintaining a 1:1 mapping for legacy PIC
> irq range (0 to 15).
>
> Doing the same for domain 0 is probably more 'exciting'. If we could force
> the 'vector space' IRQ allocation strategy of PCI_MSI then I think this is
> more plausible -- since that effectively gives a level of indirection from
> the GSI space. The vectors that Xen currently hands out to domain0 are real
> vector numbers but, of course, it would be trivial to add a level of
> indirection there, or in the caller.
>
> This all raises an obvious question: what do you think the scope of work for
> upstream merging right now should be? Previously it was non-driver domUs
> only (and a simplified form at that). Are you thinking about upstreaming
> everything, or is this just review and preparation for that effort in the
> future?
>   

My plan is:

   1. simple domU (UP only, shadow page tables, no pciback)
   2. domU SMP (to really exercise the paravirt interface)
   3. add more to domU (explicit MMU ops, optimising, etc)
   4. port to x86-64
   5. dom0

Once the paravirt version of simple domU is working, it will be time to 
consider how to sync all this with the xen-unstable tree.  Also after 
step 1, I don't have any strong feelings about the order.  I haven't 
really looked into what the dom0 work really requires, so I'm not sure 
whether it can start early, or if it would be best done once domU is 
relatively sophisticated.  Similarly I'm not sure how much work would be 
required for 64-bit (though an obvious prerequisite is that we implement 
paravirtops for x86-64).

As far as the event channels stuff goes, I think I'll keep it as simple 
as possible for now, and just support as much as needed to get domU 
going, with a view to making it more complex later.

    J

Re: Physical and dynamic event channels

[Keir Fraser]

On 11/8/06 4:04 pm, "Jeremy Fitzhardinge" <jeremy@goop.org> wrote:

> As far as the event channels stuff goes, I think I'll keep it as simple
> as possible for now, and just support as much as needed to get domU
> going, with a view to making it more complex later.

For domU you can get rid of the 1:1 section entirely and start the dynamic
space from 0.

 -- Keir

Re: Physical and dynamic event channels

[Jeremy Fitzhardinge]

Keir Fraser wrote:
> For domU you can get rid of the 1:1 section entirely and start the dynamic
> space from 0.
>   

Yup, that's what I was planning.

    J