Re: [PATCH 1/4] softirq: implement IRQ flood detection mechanism
From: Ming Lei
Date: Wed Aug 28 2019 - 09:51:14 EST
On Wed, Aug 28, 2019 at 01:23:06PM +0200, Thomas Gleixner wrote:
> On Wed, 28 Aug 2019, Ming Lei wrote:
> > On Wed, Aug 28, 2019 at 01:09:44AM +0200, Thomas Gleixner wrote:
> > > > > Also how is that supposed to work when sched_clock is jiffies based?
> > > >
> > > > Good catch, looks ktime_get_ns() is needed.
> > >
> > > And what is ktime_get_ns() returning when the only available clocksource is
> > > jiffies?
> > IMO, it isn't one issue. If the only clocksource is jiffies, we needn't to
> > expect high IO performance. Then it is fine to always handle the irq in
> > interrupt context or thread context.
> > However, if it can be recognized runtime, irq_flood_detected() can
> > always return true or false.
> Right. The clocksource is determined at runtime. And if there is no high
> resolution clocksource then that function will return crap.
This patch still works even though the only clocksource is jiffies.
> > > No. Talk to Daniel. There is not going to be yet another ad hoc thingy just
> > > to make block happy.
> > I just take a first glance at the code of 'interrupt timing', and its
> > motivation is to predicate of the next occurrence of interested interrupt
> > for use cases of PM, such as predicate wakeup time.
> > And predication could be one much more difficult thing, and its implementation
> > requires to record more info: such as irq number, recent multiple irq timestamps,
> > that means its overhead is a bit high. Meantime IRQS_TIMINGS should only be set
> > on interested interrupt(s).
> Well, yes. But it's trivial enough to utilize parts of it for your
>From the code of kernel/irq/timing.c:
1) record_irq_time() only records the start time of one irq, and not
consider the time taken in interrupt handler, so we can't figure out
the real interval between two do_IRQ() on one CPU
2) irq/timing doesn't cover softirq
Daniel, could you take a look and see if irq flood detection can be
implemented easily by irq/timing.c?