irq cpu binding

[Adrian Chadd]

You also have to do the taskqueue threads.

(I mean, pmc could also be broken...)


-a


On 31 March 2015 at 15:59, Slawa Olhovchenkov <slw at zxy.spb.ru> wrote:
> On Sat, Mar 28, 2015 at 12:33:52PM -0700, Adrian Chadd wrote:
>
>> That's done deferred by the bus interrupt wiring. That's something
>> John's been looking into as part of the general NUMA work (and I'm
>> trying to debug right now, on dual-socket boxes with ixgbe. :-)
>>
>> Look at bus_bind_intr() and the twisty path to intr_event_bind(), then
>> x86/x86/intr_machdep.c:intr_assign_cpu(), then intr_shuffle_cpus() at
>> boot, versus what happens via calls to pic_assign_cpu to setup the
>> wiring.
>
> I am do simple, ugle hack ixgbe driver for let start cpu binding.
> I am still see ixgbe in pmc output.
>
> What may be wrong?
> What may be miss?
>
> =====
> static int ixgbe_start_cpu = 0;
> TUNABLE_INT("hw.ix.start_cpu", &ixgbe_start_cpu);
> SYSCTL_INT(_hw_ix, OID_AUTO, start_cpu, CTLFLAG_RDTUN, &ixgbe_start_cpu, 0,
>     "Start CPU for next IRQ binding");
>
> [...]
>                 if (adapter->num_queues > 1)
>                         bus_bind_intr(dev, que->res, i+ixgbe_start_cpu);
>
> #ifndef IXGBE_LEGACY_TX
>                 TASK_INIT(&txr->txq_task, 0, ixgbe_deferred_mq_start, txr);
> #endif
>                 TASK_INIT(&que->que_task, 0, ixgbe_handle_que, que);
>                 que->tq = taskqueue_create_fast("ixgbe_que", M_NOWAIT,
>                     taskqueue_thread_enqueue, &que->tq);
>                 taskqueue_start_threads(&que->tq, 1, PI_NET, "%s que",
>                     device_get_nameunit(adapter->dev));
>         }
>         ixgbe_start_cpu += adapter->num_queues;
>