Re: [Qemu-devel] [PATCH] pc: Clean up PIC-to-APIC IRQ path

[Anthony Liguori]

On 09/04/2011 08:42 AM, Jan Kiszka wrote:
> On 2011-09-04 15:38, Anthony Liguori wrote:
> > > With current kvm device assignment it's mandatory as it only support
> > > kernel/kernel IRQ delivery. Only vfio's eventfds will make it optional
> > > (but still highly desirable).
> >
> > It's not mandatory.  All you need to be able to do is calculate the APIC
> > IRQ for a given PCI device interrupt.
>
> ...and establish notifies for changes along this line. And allow to
> update intermediate states on access.

Again, this is all in a single layer.

> >   That doesn't mean we need to be
> > able to do arbitrary interrupt resolution in generic code.
>
> We will likely have to solve the same problem on none x86 as well.

But we're already seeing the that device assignment problem is very 
different on other platforms.

Given multiple well known architectures, yes, it makes sense to write 
generic code to handle both.

But we shouldn't just assume that just because we have one use-case that 
it generalizes to arbitrarily complex use-cases.

An IRQ routing table sounds good in theory but in practice it's just not 
that simple.  Image a simple device with an output pin that is high when 
there is data to read.

It looks and smells like an IRQ but there's nothing to say that it's 
ever going to be routed to the CPU as a level interrupt.  It may be 
handled entirely within another device for a certain platform.

But in another platform, it may actually be routed as a level triggered 
interrupt to the CPU.  In fact, as we do more pin level modelling, we'll 
run into a lot of scenarios like this.

Over generalizing is just going to box us into this narrow view of 
anything that's a single bit of information being directly routable to a 
CPU.

> > There is potentially tremendous complexity here because you'll have to
> > bake all interrupt rerouting logic into a declarative API and/or call
> > into generic code to update routing tables.  Given the fact that we
> > can't even generically refer to a device reliably today, this is would
> > be a daunting task.
> >
> > We're making this all more complicated than it needs to be.
>
> We can't discuss the problem away, sorry.

"We may need at some point in the future" is not a sufficient 
justification either.

If we want to talk about concrete scenarios, fine.  I'm all for doing 
what we need to do to support things.  But unless there's an absolutely 
compelling need to introduce a global interrupt routing table, I think 
we're much better off handling it as special cases.

Regards,

Anthony Liguori

> Jan