lkml.org 
[lkml]   [2022]   [Jan]   [25]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
Patch in this message
/
Date
From
SubjectRe: [PATCH] irqchip/gic-v3-its: Select housekeeping CPUs preferentially for managed IRQs
On Tue, 25 Jan 2022 12:49:20 +0000,
Xiongfeng Wang <wangxiongfeng2@huawei.com> wrote:
>
> Hi Marc,
>
> On 2022/1/24 19:24, Marc Zyngier wrote:
> > + John Garry, as he was reporting issues around the same piece of code[1]
> >
> > On Mon, 24 Jan 2022 07:34:40 +0000,
> > Xiongfeng Wang <wangxiongfeng2@huawei.com> wrote:
> >>
> >> When using kernel parameter 'isolcpus=managed_irq,xxxx' to bind the
> >> managed IRQs to housekeeping CPUs, the effective_affinity sometimes
> >> still contains the non-housekeeping CPUs.
> >>
> >> irq_do_set_affinity() passes the housekeeping cpumask to
> >> chip->irq_set_affinity(), but ITS driver select CPU according to
> >> irq_common_data->affinity. While 'irq_common_data->affinity' is updated
> >> after chip->irq_set_affinity() is called in irq_do_set_affinity(). Also
> >> 'irq_common_data->affinity' may contains non-housekeeping CPUs. I found
> >> the below link explaining the reason.
> >> https://www.mail-archive.com/linux-kernel@vger.kernel.org/msg2267032.html
> >>
> >> To modify CPU selecting logic to prefer housekeeping CPUs, select CPU
> >> from the input cpumask parameter first. If none of it is online, then
> >> select CPU from 'irq_common_data->affinity'.
> >>
> >> Signed-off-by: Xiongfeng Wang <wangxiongfeng2@huawei.com>
> >> ---
> >> drivers/irqchip/irq-gic-v3-its.c | 5 ++++-
> >> 1 file changed, 4 insertions(+), 1 deletion(-)
> >>
> >> diff --git a/drivers/irqchip/irq-gic-v3-its.c b/drivers/irqchip/irq-gic-v3-its.c
> >> index d25b7a864bbb..17c15d3b2784 100644
> >> --- a/drivers/irqchip/irq-gic-v3-its.c
> >> +++ b/drivers/irqchip/irq-gic-v3-its.c
> >> @@ -1624,7 +1624,10 @@ static int its_select_cpu(struct irq_data *d,
> >>
> >> cpu = cpumask_pick_least_loaded(d, tmpmask);
> >> } else {
> >> - cpumask_and(tmpmask, irq_data_get_affinity_mask(d), cpu_online_mask);
> >> + cpumask_and(tmpmask, aff_mask, cpu_online_mask);
> >> + if (cpumask_empty(tmpmask))
> >> + cpumask_and(tmpmask, irq_data_get_affinity_mask(d),
> >> + cpu_online_mask);
> >
> > I think that the online_cpu_mask logical and is a bit wrong. A managed
> > interrupt should be able to target an offline CPU:
> >
> > diff --git a/drivers/irqchip/irq-gic-v3-its.c b/drivers/irqchip/irq-gic-v3-its.c
> > index eb0882d15366..0cea46bdaf99 100644
> > --- a/drivers/irqchip/irq-gic-v3-its.c
> > +++ b/drivers/irqchip/irq-gic-v3-its.c
> > @@ -1620,7 +1620,7 @@ static int its_select_cpu(struct irq_data *d,
> >
> > cpu = cpumask_pick_least_loaded(d, tmpmask);
> > } else {
> > - cpumask_and(tmpmask, irq_data_get_affinity_mask(d), cpu_online_mask);
> > + cpumask_copy(tmpmask, aff_mask);
> >
> > /* If we cannot cross sockets, limit the search to that node */
> > if ((its_dev->its->flags & ITS_FLAGS_WORKAROUND_CAVIUM_23144) &&
>
> I have tested the above modification with 'maxcpus=1' kernel parameter and got
> the following CallTrace.
>
> [ 14.679493][ T5] pstate: 204000c9 (nzCv daIF +PAN -UAO -TCO -DIT -SSBS
> BTYPE=--)
> [ 14.687114][ T5] pc : lpi_update_config+0xe0/0x300
> [ 14.692146][ T5] lr : lpi_update_config+0x3c/0x300

That's a problem similar to what John was seeing: the CPU isn't there,
and a lot of stuff goes very wrong in the absence of a CPU targeted by
a managed interrupt.

> > We still have an issue when the system hasn't booted with all its
> > CPUs, as the corresponding collections aren't initialised and we
> > end-up in a rather bad place.
>
> Shall we fix this 'effective CPU of managed IRQs is not housekeeping
> CPU' issue first, or we will wait until the 'maxcpus=1' issue is
> fixed.

I this we need to address this first. There is no point in only half
fixing it.

Thanks,

M.

--
Without deviation from the norm, progress is not possible.
\
 
 \ /
  Last update: 2022-01-25 14:46    [W:0.670 / U:0.584 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site