Re: [PATCH] 2.6.1-mm2: Get irq_vector size right for generic subarch UP installer kernels

["Mika Penttilä" <mika.penttila@kolumbus.fi>]

James Cleverdon wrote:

>On Wednesday 14 January 2004 8:36 pm, Zwane Mwaikambo wrote:
>  
>
>>On Wed, 14 Jan 2004, Nakajima, Jun wrote:
>>    
>>
>>>I tend to agree. I think the confusing part is the range of the IRQs on
>>>that machine. Assuming that irq_vector[NR_IRQ_VECTORS = 1024] requires
>>>more entries, then the IRQs should take that range, because
>>>IO_APCI_VECTOR(irq) is just irq_vector[irq], for example. If NR_IRQS is
>>>still 224, how can do_IRQ() can get the correct IRQ (i.e. >= 224) ? So
>>>in that case, the IRQ should be smaller than 224, then irq_vector[]
>>>should be smaller.
>>>      
>>>
>>In my opinion we should be breaking after we've exceeded the maximum
>>external vectors we can install. This will of course mean less than
>>the number of RTEs. James have you actually managed to use the devices
>>connected to the high (over ~224) RTEs?
>>    
>>
>
>No, I haven't exceeded the available vectors, but wli has on a large NUMA-Q 
>box.
>
>The x440 and x445's problems are pre-reserving lots of bus numbers in the 
>BIOS, more than one per PCI slot.  They must be anticipating PCI cards with 
>bridge chips on them.
>
>I believe that the reason for irq_vector being so large is to allow IRQ (and 
>eventually vector) sharing.  The array is to map from RTE to vector.
>  
>
Any attempt to  setup an irq >= NR_IRQS will crash, because for instance 
entry.c interrupt stubs are an array of NR_IRQS entries...NR_IRQ_VECTORS 
 > NR_IRQS really doesn't make sense as is.

We do support irq sharing among devices, but not vector sharing among 
irqs. For that the handler should loop through irq_vector[] to find 
every index, index != irq, irq_vector[index] == irq_vector[irq].

--Mika