Re: [PATCH] x86: Add check for number of available vectors before CPU down [v2]

[Prarit Bhargava <prarit@redhat.com>]

On 12/20/2013 04:41 AM, rui wang wrote:
> On 12/20/13, Prarit Bhargava <prarit@redhat.com> wrote:
>>
>>
>> On 12/19/2013 01:05 PM, Tony Luck wrote:
>>> On Wed, Dec 18, 2013 at 11:50 AM, Tony Luck <tony.luck@intel.com> wrote:
>>>> Looks good to me.
>>>
>>> Though now I've been confused by an offline question about affinity.
>>
>> Heh :)  I'm pursuing it now.  Rui has asked a pretty good question that I
>> don't
>> know the answer to off the top of my head.  I'm still looking at the code.
>>
>>>
>>> Suppose we have some interrupt that has affinity to multiple cpus. E.g.
>>> (real example from one of my machines):
>>>
>>> # cat /proc/irq/94/smp_affinity_list
>>> 26,54
>>>
>>> Now If I want to take either cpu26 or cpu54 offline - I'm guessing that I
>>> don't
>>> really need to find a new home for vector 94 - because the other one of
>>> that
>>> pair already has that set up.  But your check_vectors code doesn't look
>>> like
>>> it accounts for that - if we take cpu26 offline - it would see that
>>> cpu54 doesn't
>>> have 94 free - but doesn't check that it is for the same interrupt.
>>>
>>> But I may be mixing "vectors" and "irqs" here.
>>
>> Yep.  The question really is this: is the irq mapped to a single vector or
>> multiple vectors. (I think)
>>
> 
> The vector number for an irq is programmed in the LSB of the IOAPIC
> IRTE (or MSI data register in the case of MSI/MSIx). So there can be
> only one vector number (although multiple CPUs can be specified
> through DM). An MSI-capable device can dynamically change the lower
> few bits in the LSB to signal multiple interrupts with a contiguous
> range of vectors in powers of 2,but each of these vectors is treated
> as a separate IRQ. i.e. each of them has a separate irq desc, or a
> separate line in the /proc/interrupt file. This patch shows the MSI
> irq allocation in detail:
> http://git.kernel.org/cgit/linux/kernel/git/tip/tip.git/commit/?id=51906e779f2b13b38f8153774c4c7163d412ffd9
> 
> Thanks
> Rui
> 

Gong and Rui,

After looking at this in detail I realized I made a mistake in my patch by
including the check for the smp_affinity.  Simply put, it shouldn't be there
given Rui's explanation above.

So I think the patch simply needs to do:

        this_count = 0;
        for (vector = FIRST_EXTERNAL_VECTOR; vector < NR_VECTORS; vector++) {
                irq = __this_cpu_read(vector_irq[vector]);
                if (irq >= 0) {
                        desc = irq_to_desc(irq);
                        data = irq_desc_get_irq_data(desc);
                        affinity = data->affinity;
                        if (irq_has_action(irq) && !irqd_is_per_cpu(data))
                                this_count++;
                }
        }

Can the two of you confirm the above is correct?  It would be greatly appreciated.

Tony, I apologize -- your comments made me think you were stating a fact and not
asking a question on the behavior of affinity.  I completely misunderstood what
you were suggesting.  I thought you were implying that that the affinity "tied"
IRQ behavior together; it does not.  It is simply a suggestion of what IRQs
should be assigned to a particular CPU.  There is an expectation that the system
will attempt to honour the affinity, however, it is not like each CPU is
assigned a separate IRQ.

P.