Re: [PATCH 5/5] cpuidle-haltpoll: fix up the branch check

[Zhenzhong Duan]

On 2019/11/4 23:01, Marcelo Tosatti wrote:
> On Mon, Nov 04, 2019 at 11:10:25AM +0800, Zhenzhong Duan wrote:
> > On 2019/11/2 5:26, Marcelo Tosatti wrote:
> > > On Sat, Oct 26, 2019 at 11:23:59AM +0800, Zhenzhong Duan wrote:
> > > > Ensure pool time is longer than block_ns, so there is a margin to
> > > > avoid vCPU get into block state unnecessorily.
> > > >
> > > > Signed-off-by: Zhenzhong Duan <zhenzhong.duan@xxxxxxxxxx>
> > > > ---
> > > >   drivers/cpuidle/governors/haltpoll.c | 6 +++---
> > > >   1 file changed, 3 insertions(+), 3 deletions(-)
> > > >
> > > > diff --git a/drivers/cpuidle/governors/haltpoll.c b/drivers/cpuidle/governors/haltpoll.c
> > > > index 4b00d7a..59eadaf 100644
> > > > --- a/drivers/cpuidle/governors/haltpoll.c
> > > > +++ b/drivers/cpuidle/governors/haltpoll.c
> > > > @@ -81,9 +81,9 @@ static void adjust_poll_limit(struct cpuidle_device *dev, unsigned int block_us)
> > > >   	u64 block_ns = block_us*NSEC_PER_USEC;
> > > >   	/* Grow cpu_halt_poll_us if
> > > > -	 * cpu_halt_poll_us < block_ns < guest_halt_poll_us
> > > > +	 * cpu_halt_poll_us <= block_ns < guest_halt_poll_us
> > > >   	 */
> > > > -	if (block_ns > dev->poll_limit_ns && block_ns <= guest_halt_poll_ns) {
> > > > +	if (block_ns >= dev->poll_limit_ns && block_ns < guest_halt_poll_ns) {
> > > 					      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
> > >
> > > If block_ns == guest_halt_poll_ns, you won't allow dev->poll_limit_ns to
> > > grow. Why is that?
> > Maybe I'm too strict here. My understanding is: if block_ns = guest_halt_poll_ns,
> > dev->poll_limit_ns will grow to guest_halt_poll_ns,
> OK.
>
> > then block_ns = dev->poll_limit_ns,
> block_ns = dev->poll_limit_ns = guest_halt_poll_ns. OK.
>
> > there is not a margin to ensure poll time is enough to cover the equal block time.
> > In this case, shrinking may be a better choice?
> Ok, so you are considering _on the next_ halt instance, if block_ns =
> guest_halt_poll_ns again?
Yes, I realized it's rarely to happen in nanosecond granularity.
> Then without the suggested modification: we don't shrink, poll for
> guest_halt_poll_ns again.
>
> With your modification: we shrink, because block_ns ==
> guest_halt_poll_ns.
>
> IMO what really clarifies things here is either the real sleep pattern
> or a synthetic sleep pattern similar to the real thing.
Agree
> Do you have a scenario where the current algorithm is maintaining
> a low dev->poll_limit_ns and performance is hurt?
>
> If you could come up with examples, such as the client/server pair at
> https://lore.kernel.org/lkml/20190514135022.GD4392@xxxxxxxx/T/
>
> or just a sequence of delays:
> block_ns, block_ns, block_ns-1,...
>
> It would be easier to visualize this.

Looks hard to generate a sequence of delays of same value in nanoseconds 
which is also CPU cycle granularity.

I think this patch doesn't help much for a real scenario, so pls ignore it.

Thanks

Zhenzhong