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

[Prarit Bhargava <prarit@redhat.com>]

On 12/19/2013 02:19 AM, rui wang wrote:
> On 12/19/13, Prarit Bhargava <prarit@redhat.com> wrote:
>>
>>
>> On 12/03/2013 09:48 PM, rui wang wrote:
>>> On 11/20/13, Prarit Bhargava <prarit@redhat.com> wrote:
>>> Have you considered the case when an IRQ is destined to more than one CPU?
>>> e.g.:
>>>
>>> bash# cat /proc/irq/89/smp_affinity_list
>>> 30,62
>>> bash#
>>>
>>> In this case offlining CPU30 does not seem to require an empty vector
>>> slot. It seems that we only need to change the affinity mask of irq89.
>>> Your check_vectors() assumed that each irq on the offlining cpu
>>> requires a new vector slot.
>>>
>>
>> Rui,
>>
>> The smp_affinity_list only indicates a preferred destination of the IRQ, not
>> the
>> *actual* location of the CPU.  So the IRQ is on one of cpu 30 or 62 but not
>> both
>> simultaneously.
>>
> 
> It depends on how IOAPIC (or MSI/MSIx) is configured. An IRQ can be
> simultaneously broadcast to all destination CPUs (Fixed Mode) or
> delivered to the CPU with the lowest priority task (Lowest Priority
> Mode). It's programmed in the Delivery Mode bits of the IOAPIC's IO
> Redirection table registers, or the Message Data Register in the case
> of MSI/MSIx

Hmm ... I didn't realize that this was a possibility.  I'll go back and rework
the patch.

Thanks for the info Rui!

P.

> 
>> If the case is that 62 is being brought down, then the smp_affinity mask
>> will be
>> updated to reflect only cpu 30 (and vice versa).
>>
> 
> Yes the affinity mask should be updated. But if it was destined to
> more than one CPU, your "this_counter" does not seem to count the
> right numbers. Are you saying that smp_affinity mask is broken on
> Linux so that there's no way to configure an IRQ to target more than
> one CPU?
> 
> Thanks
> Rui
> 
>> P.
>>