On Wed, 28 Mar 2012 18:41:03 -0700, David Daney <david.daney@xxxxxxxxxx> wrote:
> On 03/28/2012 03:22 PM, Grant Likely wrote:
> > On Tue, 27 Mar 2012 11:24:51 -0700, David Daney<david.daney@xxxxxxxxxx> wrote:
> >> On 03/26/2012 06:56 PM, Rob Herring wrote:
> >>> On 03/26/2012 02:31 PM, David Daney wrote:
> >>>> From: David Daney<david.daney@xxxxxxxxxx>
> >> [...]
> >>>> +static int octeon_irq_ciu_map(struct irq_domain *d,
> >>>> + unsigned int virq, irq_hw_number_t hw)
> >>>> +{
> >>>> + unsigned int line = hw>> 6;
> >>>> + unsigned int bit = hw& 63;
> >>>> +
> >>>> + if (virq>= 256)
> >>>> + return -EINVAL;
> >>>
> >>> Drop this. You should not care what the virq numbers are.
> >>
> >>
> >> I care that they don't overflow the width of octeon_irq_ciu_to_irq (a u8).
> >>
> >> So really I want to say:
> >>
> >> if (virq>= (1<< sizeof (octeon_irq_ciu_to_irq[0][0]))) {
> >> WARN(...);
> >> return -EINVAL;
> >> }
> >>
> >>
> >> I need a map external to any one irq_domain. The irq handling code
> >> handles sources that come from two separate irq_domains, as well as irqs
> >> that are not part of any domain.
> >
> > You can get past this limitation by using the struct irq_data .hwirq and
> > .domain members for the irq ==> hwirq translation, and for hwirq ==>
> > irq the code should already have the context to know which user it is.
> >
> > For the irqs that are not covered by an irq_domain, the driver is free
> > to set the .hwirq value directly. Ultimately however, it will
> > probably be best to add an irq domain for those users also.
> >
> > ...
> >
> > Howver, I don't understand where the risk is in overflowing
> > octeon_irq_ciu_to_irq[][]. From what I can see, the virq value isn't
> > used at all to calculate the array dereference. line and bit are
> > calculated from the hwirq value. What am I missing?
> >
>
> We do the opposite. We extract the hwirq value from the interrupt
> controller and then look up virq in the table. If the range of virq
> overflows the width of u8, we would end up calling do_IRQ() with a bad
> value. Also this dispatch code is not aware of the various irq_domains
> and non irq_domain irqs, it is a single function that handles them all
> calling do_IRQ() with whatever it looks up in the table.
>
> We could use a wider type for this lookup array, but that would increase
> the cache footprint of the irq dispatcher...
Ah, I missed that octeon_irq_ciu_to_irq was a u8. You're using Linux
though; your cache footprint is already trashed. :-) Please just use
unsigned int for all irq storage since that is the type used by all
core interrupt handling code. Anything else smells like premature
optimization. :-)
Besides, now that we have it you should plan to switch to the common
mechanism of irq_domain for hwirq->irq reverse mapping anyway. It
doesn't make any sense for each platform to reinvent it's own reverse
mapping scheme.
g.
