Re: [PATCH] xirc2ps_cs irq return fix

[Andrew Morton]

Alan Cox <[e-mail]> wrote:
>
> On Maw, 2003-05-27 at 11:45, Andrew Morton wrote:
> > The below patch has been in -mm for some time.  It was supposed to kill the
> > IRQ if 750 of the previous 1000 IRQs weren't handled.
> > 
> > I disabled the killing code because it was triggering on someone's
> > works-just-fine setup.
> > 
> > There will be pain involved in getting all this to work right.  Do you
> > really think there's much value in it?
> 
> Being able to at least turn it on at run time is valuable when you are debugging
> a box operated by someone who doesnt habitually rebuild kernels. The 750 of 1000
> thing doesnt work because it can happen to be timing triggered by blocks of IRQ's
> from a chip being folded together. The "million in a row" should be a stuck IRQ,
> maybe 50,000 in a row even but just "zillions in a row"
> 

The problem with zillions-in-a-row is that the babbling device could be
sharing the IRQ with a non-babbling device.  So the legitimate interrupts
from the non-babbler will cause the detection state machine to reset
itself.  It will never go off, and the box remains locked up.

Maybe the 750/1000 should be 100/100,000.  Who knows...