lkml.org 
[lkml]   [2020]   [Apr]   [27]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH] irqchip/sifive-plic: allow many cores to handle IRQs
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@sifive.com> wrote:
>>
>> On Sun, Apr 26, 2020 at 8:47 PM Anup Patel <anup@brainfault.org>
>> wrote:
>> >
>> > On Sun, Apr 26, 2020 at 4:37 PM Zong Li <zong.li@sifive.com> 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@sifive.com>
>> > > ---
>> > > 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...

\
 
 \ /
  Last update: 2020-04-27 10:28    [W:0.054 / U:2.296 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site