Re: [PATCH 4/8] ARCv2: perf: Support sampling events using overflow interrupts

[Vineet Gupta]

On Tuesday 16 June 2015 02:37 PM, Peter Zijlstra wrote:

On Tue, Jun 16, 2015 at 05:37:40AM +0000, Vineet Gupta wrote:


> On Monday 15 June 2015 09:55 PM, Peter Zijlstra wrote:


> > On Tue, Jun 09, 2015 at 05:49:28PM +0530, Vineet Gupta wrote:


> >> +  if (arc_pmu->has_interrupts) {
> >> +          int irq = platform_get_irq(pdev, 0);


> > Hmm, so you're requesting a regular interrupt.
> >
> > I see your architecture has IRQ priorities, could you play games and
> > create NMIs using those?
> >
> > For example, never mask L1 (assuming that's the highest priority) and
> > treat that as an NMI.


>
> I've had this idea before, however, while ARCv2 provides hardware interrupt
> priorities, we really can't implement true NMI, because CLRI / SETI used at
> backend of loal_irq_save() / restore() impact all priorities (statsu32 register
> has a global enable interrupt bit which these wiggle). So e.g. a
> spin_lock_irqsave() will lock out even the perf interrupt.


Hmm, bugger. I (of course) only looked at the kernel source, since that
is all I have, and the current arch/arc/ frobs with those two En bits in
status32.

Yeah the new intc and handling is series #B currently floated on lkml - we are discussing #C here - perf sub-series.


So arcv2 changed all that, shame.

Not really - arcv2 intc priorities (0-15, kind of excessive for linux usecase) is a logical extension of fixed 2 levels provided by arcompact (L2 high, L1 normal). Instead of 2 bits each of enable/active we have multiple now - spread over a couple of registers (statsu32.EN, irq_priority....).
arcv2 core interrupt handling is loads better than what we had before - specially with hardware stack switching / auto reg file save etc.

Plus now hardware enforces strict priority. In arcompact, the prio enable is semi-loosely enforced. L2 (high prio) can obviously interrupt L1 (low prio). But the other way around is possible too - if E1 is set - hence the beautiful code in arch/arc/kernel/irq.c : arch_local_irq_enable()

Technically we could also do something here, not use SETI / CLRI for disabling / enabling interrupts, but play with allowed intr level etc - but I think the cost of writing to interrupt priority/enable registers that often will be prohibitive at best.





> OTOH, we can improve the perf isr path a bit - by not routing it thru regular
> interrupt return path (song and dance of CONFIG_PREEMPT_IRQ and possible
> preemption). Plus there's a bit more we can do in the isr itself - not looping
> thru 32 counters etc using ffs() etc - but I'd rather do that as separate series,
> once we have the core support in.


Yeah, borderline useful though, the reason the NMI thing is so useful is
that you can profile _inside_ IRQ-disabled regions. Now your
local_irq_enable() et al will be the hottest functions ever :-)

Regular interrupts are really only useful for userspace profiling, which
if plenty useful on its own of course. But kernel profiling is very
handy too :-)


Totally agree !
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/