[i386, x86-64] ioapic_register_intr() and assign_irq_vector() questions

[i386, x86-64] ioapic_register_intr() and assign_irq_vector() questions

[Jan Beulich]

Since ioapic_register_intr() ties, for the PCI case, the interrupt gate for vector to interrupt[vector], doesn't this,
since do_IRQ() derives the IRQ from the value loaded in the handler at that address (which is the value of vector), mean
that here irq == vector in all cases? If so, why does this function need an if/else (the 'else' case would then be good
for both cases)?

Looking at the call paths assign_irq_vector() can get called from, I would think this function, namely as it's using
static variables, lacks synchronization - is there any (hidden) reason this is not needed here?

Thanks, Jan

Re: [i386, x86-64] ioapic_register_intr() and assign_irq_vector() questions

[Andi Kleen]

"Jan Beulich" <jbeulich@novell.com> writes:

> Since ioapic_register_intr() ties, for the PCI case, the interrupt
> gate for vector to interrupt[vector], doesn't this, since do_IRQ()
> derives the IRQ from the value loaded in the handler at that address
> (which is the value of vector), mean that here irq == vector in all
> cases? If so, why does this function need an if/else (the 'else'
> case would then be good for both cases)?

Yes agreed. irq should be always equal vector. IIRC this 
stems from the MSI work from Tom Nguyen. Maybe he knows
why he coded it like this. I suppose it could be cleaned up.

> 
> Looking at the call paths assign_irq_vector() can get called from, I
> would think this function, namely as it's using static variables,
> lacks synchronization - is there any (hidden) reason this is not
> needed here?

It is only called during system initialization which is single threaded. 
If someone added ioapic hotplug they would need to do something about 
this.

-Andi

Re: [i386, x86-64] ioapic_register_intr() and assign_irq_vector() questions

[Jan Beulich]

>> Looking at the call paths assign_irq_vector() can get called from, I
>> would think this function, namely as it's using static variables,
>> lacks synchronization - is there any (hidden) reason this is not
>> needed here?

>It is only called during system initialization which is single threaded. 
>If someone added ioapic hotplug they would need to do something about 
>this.

Hmm, as I looked through this I expected this to be possibly called also later, as it seems to be on paths reachable
from exported functions (which clearly can be called only after the single-threaded phase is over.

Jan