Re: [PATCH 30 of 38] xen: implement io_apic_ops

[Ingo Molnar <mingo@elte.hu>]

* Jeremy Fitzhardinge <jeremy@goop.org> wrote:

> Ingo Molnar wrote:
>> * Jeremy Fitzhardinge <jeremy@goop.org> wrote:
>>
>>   
>>> Writes to the IO APIC are paravirtualized via hypercalls, so implement
>>> the appropriate operations.
>>>
>>> Signed-off-by: Jeremy Fitzhardinge <jeremy.fitzhardinge@citrix.com>
>>> ---
>>>  arch/x86/xen/Makefile    |    3 +-
>>>  arch/x86/xen/apic.c      |   66 ++++++++++++++++++++++++++++++++++++++++++++++
>>>  arch/x86/xen/enlighten.c |    2 +
>>>  arch/x86/xen/xen-ops.h   |    2 +
>>>  4 files changed, 72 insertions(+), 1 deletion(-)
>>>     
>>
>> hm, why is the ioapic used as the API here, and not an irqchip?
>>   
>
> In essence, the purpose of the series is to break the 1:1 
> relationship between Linux irqs and hardware GSIs.  This allows me 
> to have my own irq allocator, which in turn allows me to intermix 
> "physical" irqs (ie, a Linux irq number bound to a real hardware 
> interrupt source) with the various software/virtual irqs the Xen 
> system needs.
>
> Once a physical irq has been mapped onto a gsi interrupt source, the 
> mechanisms for handing the ioapic side of things are more or less 
> the same.  There's the same procedure of finding the ioapic/pin for 
> a gsi and programming the appropriate vector.
>
> (Presumably once I implement MSI support, all references to "gsi" 
> will become "gsi/msi/etc".)
>
> So, there's an awkward tradeoff.  I could just completely duplicate 
> the whole irq/vector/ioapic management code and hide it under my own 
> irqchip, but it would end up duplicating a lot of the existing code.  
> My alternative was to try to open out the existing code into 
> something like a thin ioapic library, which I can call into as 
> needed.  The only low-level difference is that the Xen ioapics need 
> to be programmed via a hypercall rather than register writes.
>
> If the x86 interrupt layer in general decouples irqs from GSIs, then 
> I can probably make use of that to clean things up.  A general irq 
> allocator along with some way of attaching interrupt-source-specific 
> information to each irq would get me a long way, I think.  I'd still 
> need hooks to paravirtualize the actual ioapic writes, but at least 
> I wouldn't need to have quite so much delicate hooking.

it certainly looks thin enough to me although i'm really not sure we 
want to virtualize at the IO-APIC level. Peter, what's your 
opinion/preference?

	Ingo