[ofa-general] which CPU will ib_mthca interrupt next?

[Rick Jones]

Roland Dreier wrote:
>  > I've no idea which CPU ib_mthca will interrupt next.  ISTR (but could
>  > be wrong) that as I repeated a test with the 1.1 bits that the same
>  > CPU would be interrupted, but with 1.2 it seems that the
>  > card/firmware/whatever is deciding to migrate interrupts around.
> 
> I don't think this is an OFED change but rather a kernel change.
> 
> Anyway, first make sure you don't have a userspace irq balancer
> running.  (irqbalanced or something like that).

Grrr - indeed that is what was happening, the blessed irqbalancer was 
running.  I run into that from time to time, then go run to/in an 
environment blissfully free from it and forget about its evil ways :(

It seems to have been entirely too aggressive here - changing the 
interrupt assignements between successive netperf runs.  I have decided 
to terminate it with extreme predjudice.

> 
> Then you can set IRQ affinity through 
> 
>     /proc/irq/77/smp_affinity
> 
> The file takes a bitmap of allowed CPUs.
> (where 77 is your real IRQ number of course).

Yep - once the wicked-irq-witch is dead does a:

03:00.0 InfiniBand: Mellanox Technologies MT25208 InfiniHost III Ex 
(Tavor compatibility mode) (rev 20)

naturally want to interrupt more than one CPU at a time?

thanks,

rick jones