[PATCH v4 7/8] cpuidle/poll_state: replace cpu_relax with smp_cond_load_relaxed

[Okanovic, Haris]

On Mon, 2024-04-08 at 11:46 -0700, Ankur Arora wrote:
> CAUTION: This email originated from outside of the organization. Do not click links or open attachments unless you can confirm the sender and know the content is safe.
> 
> 
> 
> Okanovic, Haris <harisokn at amazon.com> writes:
> 
> > On Fri, 2024-04-05 at 16:14 -0700, Ankur Arora wrote:
> > > CAUTION: This email originated from outside of the organization. Do not click links or open attachments unless you can confirm the sender and know the content is safe.
> > > 
> > > 
> > > 
> > > Okanovic, Haris <harisokn at amazon.com> writes:
> > > 
> > > > On Thu, 2024-02-15 at 09:41 +0200, Mihai Carabas wrote:
> > > > > cpu_relax on ARM64 does a simple "yield". Thus we replace it with
> > > > > smp_cond_load_relaxed which basically does a "wfe".
> > > > > 
> > > > > Suggested-by: Peter Zijlstra <peterz at infradead.org>
> > > > > Signed-off-by: Mihai Carabas <mihai.carabas at oracle.com>
> > > > > ---
> > > > >  drivers/cpuidle/poll_state.c | 15 ++++++++++-----
> > > > >  1 file changed, 10 insertions(+), 5 deletions(-)
> > > > > 
> > > > > diff --git a/drivers/cpuidle/poll_state.c b/drivers/cpuidle/poll_state.c
> > > > > index 9b6d90a72601..1e45be906e72 100644
> > > > > --- a/drivers/cpuidle/poll_state.c
> > > > > +++ b/drivers/cpuidle/poll_state.c
> > > > > @@ -13,6 +13,7 @@
> > > > >  static int __cpuidle poll_idle(struct cpuidle_device *dev,
> > > > >                             struct cpuidle_driver *drv, int index)
> > > > >  {
> > > > > +    unsigned long ret;
> > > > >      u64 time_start;
> > > > > 
> > > > >      time_start = local_clock_noinstr();
> > > > > @@ -26,12 +27,16 @@ static int __cpuidle poll_idle(struct cpuidle_device *dev,
> > > > > 
> > > > >              limit = cpuidle_poll_time(drv, dev);
> > > > > 
> > > > > -            while (!need_resched()) {
> > > > > -                    cpu_relax();
> > > > > -                    if (loop_count++ < POLL_IDLE_RELAX_COUNT)
> > > > > -                            continue;
> > > > > -
> > > > > +            for (;;) {
> > > > >                      loop_count = 0;
> > > > > +
> > > > > +                    ret = smp_cond_load_relaxed(&current_thread_info()->flags,
> > > > > +                                                VAL & _TIF_NEED_RESCHED ||
> > > > > +                                                loop_count++ >= POLL_IDLE_RELAX_COUNT);
> > > > 
> > > > Is it necessary to repeat this 200 times with a wfe poll?
> > > 
> > > The POLL_IDLE_RELAX_COUNT is there because on x86 each cpu_relax()
> > > iteration is much shorter.
> > > 
> > > With WFE, it makes less sense.
> > > 
> > > > Does kvm not implement a timeout period?
> > > 
> > > Not yet, but it does become more useful after a WFE haltpoll is
> > > available on ARM64.
> > 
> > Note that kvm conditionally traps WFE and WFI based on number of host
> > CPU tasks. VMs will sometimes see hardware behavior - potentially
> > polling for a long time before entering WFI.
> > 
> > https://elixir.bootlin.com/linux/latest/source/arch/arm64/kvm/arm.c#L459
> 
> Yeah. There was a discussion on this
> https://lore.kernel.org/lkml/871qc6qufy.fsf@oracle.com/.
> 
> > > Haltpoll does have a timeout, which you should be able to tune via
> > > /sys/module/haltpoll/parameters/ but that, of course, won't help here.
> > > 
> > > > Could you make it configurable? This patch improves certain workloads
> > > > on AWS Graviton instances as well, but blocks up to 6ms in 200 * 30us
> > > > increments before going to wfi, which is a bit excessive.
> > > 
> > > Yeah, this looks like a problem. We could solve it by making it an
> > > architectural parameter. Though I worry about ARM platforms with
> > > much smaller default timeouts.
> > > The other possibility is using WFET in the primitive, but then we
> > > have that dependency and that's a bigger change.
> > 
> > See arm64's delay() for inspiration:
> > 
> > https://elixir.bootlin.com/linux/v6.9-rc2/source/arch/arm64/lib/delay.c#L26
> 
> Sure, that part is straight-forward enough. However, this will need a fallback
> the case when WFET is not available. And, because this path is used on x86,
> so we need a cross platform smp_cond*timeout(). Though given that the x86
> version is based on cpu_relax() then that could just fold the sched_clock()
> check in.

I was trying to point out how delay() handles different configurations:
It prefers WFET when available, falls back to WFE when event stream is
available, and finally falls back to cpu_relax() as last resort. Same
logic can apply here. The x86 case can always use cpu_relax() fallback,
for same behavior as smp_cond_load_relaxed().

Re your concern about "ARM platforms with much smaller default
timeouts": You could do something different when arch_timer_get_rate()
is too small. Although I'm not sure this is a huge concern, given that
delay() doesn't seem to care in the WFE case.

-- Haris Okanovic

> 
> Maybe another place to do this would be by KVM forcing a WFE timeout. Arguably
> that is needed regardless of whether we use a smp_cond*timeout() or not.
> 
> --
> ankur