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

[Ankur Arora <ankur.a.arora@oracle.com>]

Okanovic, Haris <harisokn@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@infradead.org>
>> Signed-off-by: Mihai Carabas <mihai.carabas@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.

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.

Will address this in the next version.

Thanks for pointing this out.

--
ankur