Re: [IRQ] IRQ affinity not working properly?

[Thomas Gleixner <tglx@linutronix.de>]

On Wed, Apr 21 2021 at 09:31, Nitesh Narayan Lal wrote:
> On 3/28/21 2:45 PM, Thomas Gleixner wrote:
>> On Fri, Jan 29 2021 at 13:17, Chris Friesen wrote:
>>> I have a CentOS 7 linux system with 48 logical CPUs and a number of
>
> <snip>
>
>>> IR-PCI-MSI-edge      iavf-net1-TxRx-3
>>> 961:          0          0          0          0 28492         0 
>>> IR-PCI-MSI-edge iavf-0000:b5:02.7:mbx
>>> 962:          0          0          0          0 435608         0 
>>> IR-PCI-MSI-edge      iavf-net1-TxRx-0
>>> 963:          0          0          0          0 394832         0 
>>> IR-PCI-MSI-edge      iavf-net1-TxRx-1
>>> 964:          0          0          0          0 398414         0 
>>> IR-PCI-MSI-edge      iavf-net1-TxRx-2
>>> 965:          0          0          0          0 192847         0 
>>> IR-PCI-MSI-edge      iavf-net1-TxRx-3
>>>
>>> There were IRQs coming in on the "iavf-0000:b5:02.7:mbx" interrupt at 
>>> roughly 1 per second without any traffic, while the interrupt rate on 
>>> the "iavf-net1-TxRx-<X>" seemed to be related to traffic.
>>>
>>> Is this expected?  It seems like the IRQ subsystem is not respecting the 
>>> configured SMP affinity for the interrupt in question.  I've also seen 
>>> the same behaviour with igb interrupts.
>> No it's not expected. Do you see the same behaviour with a recent
>> mainline kernel, i.e. 5.10 or 5.11?
>>
>>
> Jesse pointed me to this thread and apologies that it took a while for me
> to respond here.
>
> I agree it will be interesting to see with which kernel version Chris is
> reproducing the issue.

And the output of

 /proc/irq/$NUMBER/smp_affinity_list
 /proc/irq/$NUMBER/effective_affinity_list

> Initially, I thought that this issue is the same as the one that we have
> been discussing in another thread [1].
>
> However, in that case, the smp affinity mask itself is incorrect and doesn't
> follow the default smp affinity mask (with irqbalance disabled).

That's the question...

Thanks,

        tglx