Re: [PATCH] timers: Fix get_next_timer_interrupt() with no timers pending

[Frederic Weisbecker]

On Fri, Jul 16, 2021 at 06:38:37PM +0200, Nicolas Saenz Julienne wrote:
> On Sat, 2021-07-10 at 02:52 +0200, Frederic Weisbecker wrote:
> > I guess later we can turn this .timers_pending into
> > .timers_count and that would spare us the costly call to
> > __next_timer_interrupt() up to the last level after the last
> > timer is dequeued.
> 
> I've been looking into this. AFAIU there is no limit to the number of timers
> one might enqueue, so there is no fool proof way of selecting .timers_count's
> size. That said, 'struct timer_list' size is 40 bytes (as per pahole), so in
> order to overflow an u32 .timers_count you'd need to allocate ~160GB in 'struct
> timer_list' which I think is safe to assume will never happen.
> 
> Also, I measured the costy call to __next_timer_interrupt() it's slightly less
> than 1us on my test machine. Not a that big in the grand scheme of things, but
> it's in the irq exit code path, so I think it's worth the extra complexity in
> the timer code.

And also each time we iterate the idle loop. In fact __next_timer_interrupt()
won't always have the same cost: the worst case is when the wheel is entirely
empty after the last removal and we need to walk through all 9 levels. It's
a pretty common case because it happens when the last timer expires.

And that's the only one case to measure because it's the only one covered
by the counter.

> 
> Any thoughs?
> 
> -- 
> Nicolás Sáenz
>