RE: [RFC 06/12] genirq: Add per-cpu flow handler with conditional IRQ stats

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



From: Thomas Gleixner <tglx@xxxxxxxxxxxxx> Sent: Thursday, June 6, 2024 2:34 AM
> 
> On Thu, Jun 06 2024 at 03:14, Michael Kelley wrote:
> > From: Thomas Gleixner <tglx@xxxxxxxxxxxxx> Sent: Wednesday, June 5, 2024 7:20 AM
> >>
> >> On Wed, Jun 05 2024 at 13:45, Michael Kelley wrote:
> >> > From: Thomas Gleixner <tglx@xxxxxxxxxxxxx> Sent: Wednesday, June 5, 2024 6:20 AM
> >> >
> >> > In /proc/interrupts, the double-counting isn't a problem, and is
> >> > potentially helpful as you say. But /proc/stat, for example, shows a total
> >> > interrupt count, which will be roughly double what it was before. That
> >> > /proc/stat value then shows up in user space in vmstat, for example.
> >> > That's what I was concerned about, though it's not a huge problem in
> >> > the grand scheme of things.
> >>
> >> That's trivial to solve. We can mark interrupts to be excluded from
> >> /proc/stat accounting.
> >>
> >
> > OK.  On x86, some simple #ifdef'ery in arch_irq_stat_cpu() can filter
> > out the HYP interrupts. But what do you envision on arm64, where
> > there is no arch_irq_stat_cpu()?  On arm64, the top-level interrupt is a
> > normal Linux IRQ, and its count is included in the "kstat.irqs_sum" field
> > with no breakout by IRQ. Identifying the right IRQ and subtracting it
> > out later looks a lot uglier than the conditional stats accounting.
> 
> Sure. There are two ways to solve that:
> 
> 1) Introduce a IRQ_NO_PER_CPU_STATS flag, mark the interrupt
>    accordingly and make the stats increment conditional on it.
>    The downside is that the conditional affects every interrupt.
> 
> 2) Do something like this:
> 
> static inline
> void __handle_percpu_irq(struct irq_desc *desc, irqreturn_t (*handle)(struct irq_desc
> *))
> {
> 	struct irq_chip *chip = irq_desc_get_chip(desc);
> 
> 	if (chip->irq_ack)
> 		chip->irq_ack(&desc->irq_data);
> 
> 	handle(desc);
> 
> 	if (chip->irq_eoi)
> 		chip->irq_eoi(&desc->irq_data);
> }
> 
> void handle_percpu_irq(struct irq_desc *desc)
> {
> 	/*
> 	 * PER CPU interrupts are not serialized. Do not touch
> 	 * desc->tot_count.
> 	 */
> 	__kstat_incr_irqs_this_cpu(desc);
> 	__handle_percpu_irq(desc, handle_irq_event_percpu);
> }
> 
> void handle_percpu_irq_nostat(struct irq_desc *desc)
> {
> 	__this_cpu_inc(desc->kstat_irqs->cnt);
> 	__handle_percpu_irq(desc, __handle_irq_event_percpu);
> }
> 
> So that keeps the interrupt accounted for in /proc/interrupts. If you
> don't want that remove the __this_cpu_inc() and mark the interrupt with
> irq_set_status_flags(irq, IRQ_HIDDEN). That will exclude it from
> /proc/interrupts too.
> 

Yes, this works for not double-counting in the first place. Account for the
control message interrupts in their own Linux IRQ. Then for the top-level
interrupt, instead of adding a new handler with conditional accounting,
add a new per-CPU handler that does no accounting. I had not noticed
the IRQ_HIDDEN flag, and that solves my concern about having an
entry in /proc/interrupts that always shows zero interrupts.  And with
no double-counting, the interrupt counts in /proc/stat won't be bloated.

On x86, I'll have to separately make the "HYP" line in /proc/interrupts
go away, but that's easy.

Thanks,

Michael





[Index of Archives]     [Linux Kernel]     [Kernel Newbies]     [x86 Platform Driver]     [Netdev]     [Linux Wireless]     [Netfilter]     [Bugtraq]     [Linux Filesystems]     [Yosemite Discussion]     [MIPS Linux]     [ARM Linux]     [Linux Security]     [Linux RAID]     [Samba]     [Device Mapper]

  Powered by Linux