Re: IRQ_ONESHOT expectations vs mask/unmask

[Benjamin Herrenschmidt]

On Mon, 2017-07-31 at 09:09 +0200, Thomas Gleixner wrote:
> Ben,
> 
> On Mon, 31 Jul 2017, Benjamin Herrenschmidt wrote:
> > I noticed a problem with some powerpc systems with threaded IRQs. I
> > hadn't looked at threaded irq in the past so there was a bit of
> > discovery involved. That lead to a few questions:
> > 
> >  - Threaded IRQs rely on IRQF_ONESHOT. Now, is that a flag/feature that
> > is generally exposed to drivers even in the non-threaded case, and if
> > yes, what core callback are drivers expected to use to "unmask" the
> > oneshot interrupt ?
> 
> The oneshot flag should not have any effects when the interrupt is
> non-threaded. Emphasis on should. I'm not sure whether that's true, will
> have a look.
> 
> >  - I don't see any provisions for dealing with interrupts lost while
> > masked. So on some PICs on powerpc, while we do use "fast EOI", we also
> > have a chance of edge interrupts (MSIs) being lost while masked. This
> > wasn't a problem until now because we used lazy disabling. However, it
> > looks like IRQF_ONESHOT (and thus threaded irqs) rely on masked
> > interrupts being latched in HW, otherwise an interrupt might be lost
> > between the threaded handler completing and the interrupt being
> > unmasked, or am I missing something ?
> > 
> >  - I noticed that other flow handlers (edge, edge_eoi, ...) don't have
> > any provision for IRQF_ONESHOT. Isn't that a problem ? Or will the core
> > silently swallow subsequent interrupts until the thread has completed
> > anyway ? (I might be missing something here).
> 
> The only case where IRQF_ONESHOT should have an effect is with level type
> interrupts. That's required, because otherwise the hardware interrupt would
> fire for ever. Level type interrupts should not need any hardware latching,
> we assume that they fire once they are unmasked.
> 
> For edge type interrupts we do not mask the interrupt in order not to lose
> an interrupt. If the interrupt fires while the thread handler is running,
> we mark the thread and once it the handler returns it gets called again.
> 
> > Generally, how do you suggest I fix the threaded irq problem for XICS ?
> 
> You asked a lot of questions, but you failed to explain the problem for
> XICS.

I did but maybe it wasn't clear :-)

"So on some PICs on powerpc, while we do use "fast EOI", we also> >
have a chance of edge interrupts (MSIs) being lost while masked."

Basically, a masked interrupt might get dropped rather than "latched",
so if we use the existing fasteoi handler with IRQF_ONESHOT, we'll
lose them if they occur at the wrong time.

I could use something like the edge_eoi flow handler instead I suppose
but that will *never* lazy disable which is somewhat unfortunate, very
fast paced MSIs benefit from being HW masked if we already recorded
that they occurred.

So what I would need is something along the line of ONESHOT as done in
fasteoi but only for level interrupts.

Cheers,
Ben.

> Thanks,
> 
> 	tglx