Re: [PATCH] irqchip/sifive-plic: allow many cores to handle IRQs

[Marc Zyngier]

Hi Anup,

On 2020-04-26 14:38, Anup Patel wrote:
> +Mark Z

s/k/c/ ;-)

> On Sun, Apr 26, 2020 at 6:49 PM Zong Li <zong.li@xxxxxxxxxx> wrote:
> > On Sun, Apr 26, 2020 at 8:47 PM Anup Patel <anup@xxxxxxxxxxxxxx> 
> > wrote:
> > >
> > > On Sun, Apr 26, 2020 at 4:37 PM Zong Li <zong.li@xxxxxxxxxx> wrote:
> > > >
> > > > Currently, driver forces the IRQs to be handled by only one core. This
> > > > patch provides the way to enable others cores to handle IRQs if needed,
> > > > so users could decide how many cores they wanted on default by boot
> > > > argument.
> > > >
> > > > Use 'irqaffinity' boot argument to determine affinity. If there is no
> > > > irqaffinity in dts or kernel configuration, use irq default affinity,
> > > > so all harts would try to claim IRQ.
> > > >
> > > > For example, add irqaffinity=0 in chosen node to set irq affinity to
> > > > hart 0. It also supports more than one harts to handle irq, such as set
> > > > irqaffinity=0,3,4.
> > > >
> > > > You can change IRQ affinity from user-space using procfs. For example,
> > > > you can make CPU0 and CPU2 serve IRQ together by the following command:
> > > >
> > > > echo 4 > /proc/irq/<x>/smp_affinity
> > > >
> > > > Signed-off-by: Zong Li <zong.li@xxxxxxxxxx>
> > > > ---
> > > >  drivers/irqchip/irq-sifive-plic.c | 21 +++++++--------------
> > > >  1 file changed, 7 insertions(+), 14 deletions(-)
> > > >
> > > > diff --git a/drivers/irqchip/irq-sifive-plic.c b/drivers/irqchip/irq-sifive-plic.c
> > > > index d0a71febdadc..bc1440d54185 100644
> > > > --- a/drivers/irqchip/irq-sifive-plic.c
> > > > +++ b/drivers/irqchip/irq-sifive-plic.c
> > > > @@ -111,15 +111,12 @@ static inline void plic_irq_toggle(const struct cpumask *mask,
> > > >  static void plic_irq_unmask(struct irq_data *d)
> > > >  {
> > > >         struct cpumask amask;
> > > > -       unsigned int cpu;
> > > >         struct plic_priv *priv = irq_get_chip_data(d->irq);
> > > >
> > > >         cpumask_and(&amask, &priv->lmask, cpu_online_mask);
> > > > -       cpu = cpumask_any_and(irq_data_get_affinity_mask(d),
> > > > -                                          &amask);
> > > > -       if (WARN_ON_ONCE(cpu >= nr_cpu_ids))
> > > > -               return;
> > > > -       plic_irq_toggle(cpumask_of(cpu), d, 1);
> > > > +       cpumask_and(&amask, &amask, irq_data_get_affinity_mask(d));
> > > > +
> > > > +       plic_irq_toggle(&amask, d, 1);
> > > >  }
> > > >
> > > >  static void plic_irq_mask(struct irq_data *d)
> > > > @@ -133,24 +130,20 @@ static void plic_irq_mask(struct irq_data *d)
> > > >  static int plic_set_affinity(struct irq_data *d,
> > > >                              const struct cpumask *mask_val, bool force)
> > > >  {
> > > > -       unsigned int cpu;
> > > >         struct cpumask amask;
> > > >         struct plic_priv *priv = irq_get_chip_data(d->irq);
> > > >
> > > >         cpumask_and(&amask, &priv->lmask, mask_val);
> > > >
> > > >         if (force)
> > > > -               cpu = cpumask_first(&amask);
> > > > +               cpumask_copy(&amask, mask_val);
> > > >         else
> > > > -               cpu = cpumask_any_and(&amask, cpu_online_mask);
> > > > -
> > > > -       if (cpu >= nr_cpu_ids)
> > > > -               return -EINVAL;
> > > > +               cpumask_and(&amask, &amask, cpu_online_mask);
> > > >
> > > >         plic_irq_toggle(&priv->lmask, d, 0);
> > > > -       plic_irq_toggle(cpumask_of(cpu), d, 1);
> > > > +       plic_irq_toggle(&amask, d, 1);
> > > >
> > > > -       irq_data_update_effective_affinity(d, cpumask_of(cpu));
> > > > +       irq_data_update_effective_affinity(d, &amask);
> > > >
> > > >         return IRQ_SET_MASK_OK_DONE;
> > > >  }
> > > > --
> > > > 2.26.1
> > > >
> > >
> > > I strongly oppose (NACK) this patch due to performance reasons.
> > >
> > > In PLIC, if we enable an IRQ X for N CPUs then when IRQ X occurs:
> > > 1) All N CPUs will take interrupt
> > > 2) All N CPUs will try to read PLIC CLAIM register
> > > 3) Only one of the CPUs will see IRQ X using the CLAIM register
> > > but other N - 1 CPUs will see no interrupt and return back to what
> > > they were doing. In other words, N - 1 CPUs will just waste CPU
> > > every time IRQ X occurs.
> > >
> > > Example1, one Application doing heavy network traffic will
> > > degrade performance of other applications because with every
> > > network RX/TX interrupt N-1 CPUs will waste CPU trying to
> > > process network interrupt.
> > >
> > > Example1, one Application doing heavy MMC/SD traffic will
> > > degrade performance of other applications because with every
> > > SPI read/write interrupt N-1 CPUs will waste CPU trying to
> > > process it.
> > >
> > > In fact, the current PLIC approach is actually a performance
> > > optimization. This implementation also works fine with in-kernel
> > > load-balancer and user space load balancer.
> > >
> >
> > Yes, it's exactly, I know what you pointed out. But the idea of this
> > patch is just providing a way that users could enable other cores if
> > they wanted, it could still enable only one core by this change. The
> > purpose here is thinking of flexible use, rather than limitation.
> > Maybe it would be a happy medium that we make the default case enable
> > only one core? It is a good open discussion.
>
> Making the default case as enable only one core is just a work-around.
>
> As-per my understanding, if we set affinity mask of N CPUs for IRQ X
> then it does not mean all N CPUs should receive IRQ X rather it means
> that exactly one of the N CPUs will receive IRQ X and the IRQ receiving
> CPU will be fixed (reflected by effective affinity returned by the 
> driver).
>
> If we ignore above semantics and still provide a mechanism to target
> IRQ X to N CPUs then most likely someone will try and run into
> performance issues.
>
> Please don't go this path. The performance impact in case of Guest/VM
> is even worst because PLIC is trap-n-emulated by hypervisors as MMIO
> device.

Just to add my view on this, as we get the same request on the ARM side
once every other year:

- the "broadcast" aspect of the interrupt signalling is usually a waste
  of CPU cycles on all CPUs but the one that has actually takes it.

- the fact that you have to use MMIO to interact with the interrupt
  controller potentially makes this even worse if you don't have some
  per-CPU datapath that will handle this really fast.

- not hitting the same CPU at all times for a given interrupt means that
  the caches are usually cold for this interrupt, meaning the gain in
  latency is pretty dubious.

There is a very small (but also very vocal) fringe of users that insist
that they cannot live without this broadcast aspect, because it allows
them to shave a couple of cycles off a trivial interrupt handler on 
systems
that are otherwise mostly idle.

My answer to this is that Linux is a general purpose OS, and not a
glorified interrupt handler. We aim to have balanced response times, and
give *userspace* control over where the interrupt happens in the long 
run.

HTH,

        M.
--
Jazz is not dead. It just smells funny...