[patch 14/19] softirq: Make softirq control and processing RT aware

[Frederic Weisbecker]

On Tue, Nov 24, 2020 at 01:06:15AM +0100, Thomas Gleixner wrote:
> On Tue, Nov 24 2020 at 00:58, Frederic Weisbecker wrote:
> > On Mon, Nov 23, 2020 at 08:27:33PM +0100, Thomas Gleixner wrote:
> >> On Mon, Nov 23 2020 at 14:44, Frederic Weisbecker wrote:
> >> > On Fri, Nov 13, 2020 at 03:02:21PM +0100, Thomas Gleixner wrote:
> >> >> +	/*
> >> >> +	 * Adjust softirq count to SOFTIRQ_OFFSET which makes
> >> >> +	 * in_serving_softirq() become true.
> >> >> +	 */
> >> >> +	cnt = SOFTIRQ_OFFSET;
> >> >> +	__local_bh_enable(cnt, false);
> >> >
> >> > But then you enter __do_softirq() with softirq_count() == SOFTIRQ_OFFSET.
> >> > __do_softirq() calls softirq_handle_begin() which then sets it back to
> >> > SOFTIRQ_DISABLE_OFFSET...
> >> 
> >> The RT variant of it added in this very same patch
> >> > +static inline void softirq_handle_begin(void) { }
> >> > +static inline void softirq_handle_end(void) { }
> >
> > Ah but then account_irq_enter_time() is called with SOFTIRQ_OFFSET (it's
> > currently called with softirq_count == 0 at this point) and that may mess
> > up irqtime accounting which relies on it. It could spuriously account all
> > the time between the last (soft-)IRQ exit until now as softirq time.
> 
> Good point. Haven't thought about that. Let me have a look again.

But I'm cooking a patchset which moves account_irq_enter_time() after
HARDIRQ_OFFSET or SOFTIRQ_OFFSET is incremented. This will allow us to move
tick_irq_enter() under this layout:

		 preempt_count_add(HARDIRQ_OFFSET)
		 lockdep_hardirq_enter()
		 tick_irq_enter()
		 account_irq_enter_time()

This way tick_irq_enter() can be correctly handled by lockdep and we can remove
the nasty hack which temporarily disables softirqs around it.

And as a side effect it should also fix your issue.

I should have that ready soonish.

Thanks.