Re: [PATCH] Use delayed disable mode of ioapic edge triggered interrupts

[ebiederm@xmission.com (Eric W. Biederman)]

Linus Torvalds <torvalds@osdl.org> writes:

> On Tue, 14 Nov 2006, Eric W. Biederman wrote:
>
>> The truth is in practice I don't think it matters because I don't
>> think anyone actually disables MSI or hypertransport interrupts.
>
> Fair enough, at least for a 2.6.19 kind of release timeframe (and that is 
> what I worry about most, at least right now).
>
>> At this point I have two questions.
>> - What is the easiest path to get us to a stable 2.6.19 where
>>   everything works?
>
> If people don't expect HT and MSI interrupts to be masked (and I can well 
> imagine that), then I think your two-liner patch is good to go. Komuro 
> seems to have acked it already, and in many ways that's the "minimal 
> change" for 2.6.19 right now.

Well I just doubled checked this assertion.  The one driver that uses
the hypertransport irqs doesn't call disable_irq.  On the msi side
at least the forcedeth driver does call disable_irq when in msi mode.

I just doubled checked the historical behavior of the msi code and
it has never done the delayed disable thing.  So not doing it there
is not a regression.

The MSI case is different.  MSI is fundamentally about non-shared
interrupts, and interrupts that don't race with your DMAs.  So with
MSI you don't need a status register read to process the interrupt.

In the context of Ingo's patch I don't like the idea of saddling MSI
interrupts down with the best in class work arounds for a completely
different hardware interrupt model.  Although I don't doubt MSI will
get it's own set of work arounds as we come to know it better.

> I do like Ingo's patch because it seems "safe" (even if I think it might 
> be a bit _overly_ safe), but it changes semantics enough that I don't like 
> it for 2.6.19. Even his second version definitely changes semantics for 
> level-triggered PCI interrupts, even though he fixed ExtInt/i8259 ones.
>
> So I think I'll go with your patch for now, and we can re-visit Ingo's 
> thing after 2.6.19.

Sounds like a plan.

Eric