Re: [patch 2/2] x86, irq: use 0x20 for the IRQ_MOVE_CLEANUP_VECTORinstead of 0x1f

[Suresh Siddha]

On Sat, 2010-01-30 at 23:19 -0800, Maciej W. Rozycki wrote:
> On Sun, 24 Jan 2010, H. Peter Anvin wrote:
> 
> > > > So change the IRQ_MOVE_CLEANUP_VECTOR to 0x20 and  allow 0x21-0x2f to be
> > > > used
> > > > for device interrupts. 0x30-0x3f will be used for ISA interrupts (these
> > > > also can be migrated in the context of IOAPIC and hence need to be at a
> > > > higher
> > > > priority level than IRQ_MOVE_CLEANUP_VECTOR).
> > > 
> > >   I have troubles understanding what exactly this change is needed for
> > > (i.e. what's the difference between using vectors 0x20-0x2f and 0x30-0x3f
> > > as ExtINT interrupts, what's the gain from relocating them? -- they are
> > > transparent to the APIC, so the exact priority level used does not matter
> > > at all), but since I've been cc-ed, I have one question -- have you
> > > verified that with the new arrangement the mixed interrupt mode (where
> > > some interrupts come via the APIC and some via the 8259A PICs) still
> > > works?
> > > 
> > 
> > The difference is relevant when they are *not* invoked as ExtInt interrupts,
> > but when used as IOAPIC interrupts it matters.
> 
>  Hmm, I/O APIC interrupts coming from ISA devices used not to differ from 
> ones from PCI devices and their vectors were evenly distributed across the 
> whole device range (one reason for this was the (in)famous Pentium APIC 
> limitation WRT multiple outstanding requests at the same priority level).  
> Now what you've written suggests this has changed and now ISA devices only 
> get vectors within a single priority level -- am I getting this right?

Even before the current changes (2.6.32 for example), for irq0 to irq15,
irrespective of whether those irq's were handled by PIC or an IO-APIC,
we were using 16 vectors with in a single priority level for these
irq's.

Looking at the changelog, it looks this issue happened when we merged
io_apic.c for 32-bit and 64-bit. 2.6.28 and beyond uses the same
priority level for irq0..irq15

32bit kernels 2.6.27 and before has the behavior of evenly distributed
vectors for all the io-apic irq's (legacy or non-legacy) but starting
from 2.6.28, only for non-legacy irq's (16 and above) we try to spread
the vectors uniformly across priority levels.

>  If so, then to push my original question further: how are these vectors 
> allocated -- are they identity mapped with the corresponding i8259A 
> vectors?  And how does it play with the Pentium APIC limitation (that may 
> actually apply to all the local APIC cores that use serial bus delivery; 
> I'm not sure) I mentioned above?

As we are using the code from 2.6.28 and no one noticed/complained about
this issue for more than 1.5 years, probably the pentium APIC issue is
not wide-spread.

If we care about this, I can post a fix (which is needed irrespective of
the current changes). 

thanks,
suresh

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/