Re: [PATCH] irqchip/armada-370-xp: Enable MSI affinity configuration

[Marc Zyngier <maz@kernel.org>]

On Thu, 21 Apr 2022 09:32:23 +0100,
Nathan Rossi <nathan@nathanrossi.com> wrote:
> 
> On Thu, 21 Apr 2022 at 16:54, Marc Zyngier <maz@kernel.org> wrote:
> >
> > Hi Nathan,
> >
> > On Thu, 21 Apr 2022 02:57:28 +0100,
> > Nathan Rossi <nathan@nathanrossi.com> wrote:
> > >
> > > From: Nathan Rossi <nathan.rossi@digi.com>
> > >
> > > With multiple devices attached via PCIe to an Armada 385 it is possible
> > > to overwhelm a single CPU with MSI interrupts. Under certain scenarios
> > > configuring the interrupts to be handled by more than one CPU would
> > > prevent the system from being overwhelmed. However the
> > > irqchip-aramada-370-xp driver is configured to only handle MSIs on the
> > > boot CPU, and provides no affinity configuration.
> > >
> > > This change adds support to the armada-370-xp driver to allow for
> > > configuring the affinity of specific MSI irqs and to generate the
> > > interrupts on secondary CPUs. This is done by enabling the private
> > > doorbell for all online CPUs and configures all CPUs to unmask MSI
> > > specific private doorbell bits. The CPU affinity selection of the
> > > interrupt is handled by the target list of the software triggered
> > > interrupt value, which is provided as the MSI message. The message has
> > > the associated CPU bit set for the target CPU. For private doorbell
> > > interrupts only one bit can be set otherwise all CPUs will receive the
> > > interrupt, so the lowest CPU in the affinity mask is used. This means
> > > that by default the first CPU will handle all the interrupts as was the
> > > case before.
> > >
> > > Signed-off-by: Nathan Rossi <nathan.rossi@digi.com>
> > > ---
> > >  drivers/irqchip/irq-armada-370-xp.c | 34 ++++++++++++++++++++++++++++++++--
> > >  1 file changed, 32 insertions(+), 2 deletions(-)
> > >
> > > diff --git a/drivers/irqchip/irq-armada-370-xp.c b/drivers/irqchip/irq-armada-370-xp.c
> > > index 5b8d571c04..42c257f576 100644
> > > --- a/drivers/irqchip/irq-armada-370-xp.c
> > > +++ b/drivers/irqchip/irq-armada-370-xp.c
> > > @@ -209,15 +209,37 @@ static struct msi_domain_info armada_370_xp_msi_domain_info = {
> > >
> > >  static void armada_370_xp_compose_msi_msg(struct irq_data *data, struct msi_msg *msg)
> > >  {
> > > +#ifdef CONFIG_SMP
> > > +     unsigned int cpu = cpumask_first(irq_data_get_effective_affinity_mask(data));
> > > +
> > > +     msg->data = (1 << (cpu + 8)) | (data->hwirq + PCI_MSI_DOORBELL_START);
> >
> > BIT(cpu + 8) | ...
> >
> > > +#else
> > > +     msg->data = 0xf00 | (data->hwirq + PCI_MSI_DOORBELL_START);
> >
> > This paints the existing code a bit differently. This seems to target
> > all 4 CPUs. Why is that? I'd expect only bit 8 to be set, and the
> > whole #ifdefery to go away.
> 
> I am not sure why this is targeting 4 CPUs, it will be masked by the
> percpu doorbell mask register and is effectively BIT(8). At least
> based on the documentation I have (only for armada 370/38x), which is
> why I left it as an #ifdef. I was also not able to find any specifics
> as to why it is targeting all 4 CPUs in git history. However this
> value was added with the initial driver implementation when only
> armada 370 was available in the kernel, so it is perhaps an
> inconsistent value that was never an issue due to the bits being
> reserved. I will remove the #ifdef in a v2 patch that addresses your
> other comments.

I guess we can get at least some testing from the platform maintainers
to check that this doesn't regress the UP systems.

> 
> >
> > > +#endif
> > >       msg->address_lo = lower_32_bits(msi_doorbell_addr);
> > >       msg->address_hi = upper_32_bits(msi_doorbell_addr);
> > > -     msg->data = 0xf00 | (data->hwirq + PCI_MSI_DOORBELL_START);
> > >  }
> > >
> > >  static int armada_370_xp_msi_set_affinity(struct irq_data *irq_data,
> > >                                         const struct cpumask *mask, bool force)
> > >  {
> > > -      return -EINVAL;
> > > +#ifdef CONFIG_SMP
> > > +     unsigned int cpu;
> > > +
> > > +     if (!force)
> > > +             cpu = cpumask_any_and(mask, cpu_online_mask);
> > > +     else
> > > +             cpu = cpumask_first(mask);
> > > +
> > > +     if (cpu >= nr_cpu_ids)
> > > +             return -EINVAL;
> > > +
> > > +     irq_data_update_effective_affinity(irq_data, cpumask_of(cpu));
> > > +
> > > +     return IRQ_SET_MASK_OK;
> > > +#else
> > > +     return -EINVAL;
> > > +#endif
> > >  }
> > >
> > >  static struct irq_chip armada_370_xp_msi_bottom_irq_chip = {
> > > @@ -482,6 +504,7 @@ static void armada_xp_mpic_smp_cpu_init(void)
> > >  static void armada_xp_mpic_reenable_percpu(void)
> > >  {
> > >       unsigned int irq;
> > > +     u32 reg;
> > >
> > >       /* Re-enable per-CPU interrupts that were enabled before suspend */
> > >       for (irq = 0; irq < ARMADA_370_XP_MAX_PER_CPU_IRQS; irq++) {
> > > @@ -501,6 +524,13 @@ static void armada_xp_mpic_reenable_percpu(void)
> > >       }
> > >
> > >       ipi_resume();
> > > +
> > > +     /* Enable MSI doorbell mask and combined cpu local interrupt */
> > > +     reg = readl(per_cpu_int_base + ARMADA_370_XP_IN_DRBEL_MSK_OFFS)
> > > +             | PCI_MSI_DOORBELL_MASK;
> > > +     writel(reg, per_cpu_int_base + ARMADA_370_XP_IN_DRBEL_MSK_OFFS);
> > > +     /* Unmask local doorbell interrupt */
> > > +     writel(1, per_cpu_int_base + ARMADA_370_XP_INT_CLEAR_MASK_OFFS);
> >
> > This is a duplicate of what is already in armada_370_xp_msi_init().
> > Please refactor it so that this doesn't happen twice on the first CPU.
> 
> It is duplicated, however armada_xp_mpic_reenable_percpu is not called
> on the boot cpu as the setup is called with cpuhp_setup_state_nocalls.

Ah, right. Make sure we can get rid of the code duplication then.

Thanks,

	M.

-- 
Without deviation from the norm, progress is not possible.