2024-01-29 11:56:53

by Mihai Carabas

Subject: [PATCH v3 7/7] cpuidle/poll_state: replace cpu_relax with smp_cond_load_relaxed

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 <[email protected]>
Signed-off-by: Mihai Carabas <[email protected]>
---
drivers/cpuidle/poll_state.c | 14 +++++++++-----
1 file changed, 9 insertions(+), 5 deletions(-)

diff --git a/drivers/cpuidle/poll_state.c b/drivers/cpuidle/poll_state.c
index 9b6d90a72601..440cd713e39a 100644
--- a/drivers/cpuidle/poll_state.c
+++ b/drivers/cpuidle/poll_state.c
@@ -26,12 +26,16 @@ 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;
+
+ smp_cond_load_relaxed(&current_thread_info()->flags,
+ (VAL & _TIF_NEED_RESCHED) ||
+ (loop_count++ >= POLL_IDLE_RELAX_COUNT));
+
+ if (loop_count < POLL_IDLE_RELAX_COUNT)
+ break;
+
if (local_clock_noinstr() - time_start > limit) {
dev->poll_time_limit = true;
break;
--
1.8.3.1

2024-01-29 14:52:47

by Rafael J. Wysocki

Subject: Re: [PATCH v3 7/7] cpuidle/poll_state: replace cpu_relax with smp_cond_load_relaxed

On Mon, Jan 29, 2024 at 12:56 PM Mihai Carabas <[email protected]> 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 <[email protected]>
> Signed-off-by: Mihai Carabas <[email protected]>
> ---
> drivers/cpuidle/poll_state.c | 14 +++++++++-----
> 1 file changed, 9 insertions(+), 5 deletions(-)
>
> diff --git a/drivers/cpuidle/poll_state.c b/drivers/cpuidle/poll_state.c
> index 9b6d90a72601..440cd713e39a 100644
> --- a/drivers/cpuidle/poll_state.c
> +++ b/drivers/cpuidle/poll_state.c
> @@ -26,12 +26,16 @@ 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;
> +
> + smp_cond_load_relaxed(&current_thread_info()->flags,
> + (VAL & _TIF_NEED_RESCHED) ||
> + (loop_count++ >= POLL_IDLE_RELAX_COUNT));

The inner parens are not necessary AFAICS.

Also, doesn't this return a value which can be used for checking if
_TIF_NEED_RESCHED is set instead of the condition below?

> +
> + if (loop_count < POLL_IDLE_RELAX_COUNT)
> + break;
> +
> if (local_clock_noinstr() - time_start > limit) {
> dev->poll_time_limit = true;
> break;
> --

2024-01-29 19:22:13

by Mihai Carabas

Subject: Re: [PATCH v3 7/7] cpuidle/poll_state: replace cpu_relax with smp_cond_load_relaxed

La 29.01.2024 16:52, Rafael J. Wysocki a scris:
> On Mon, Jan 29, 2024 at 12:56 PM Mihai Carabas <[email protected]> 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 <[email protected]>
>> Signed-off-by: Mihai Carabas <[email protected]>
>> ---
>> drivers/cpuidle/poll_state.c | 14 +++++++++-----
>> 1 file changed, 9 insertions(+), 5 deletions(-)
>>
>> diff --git a/drivers/cpuidle/poll_state.c b/drivers/cpuidle/poll_state.c
>> index 9b6d90a72601..440cd713e39a 100644
>> --- a/drivers/cpuidle/poll_state.c
>> +++ b/drivers/cpuidle/poll_state.c
>> @@ -26,12 +26,16 @@ 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;
>> +
>> + smp_cond_load_relaxed(&current_thread_info()->flags,
>> + (VAL & _TIF_NEED_RESCHED) ||
>> + (loop_count++ >= POLL_IDLE_RELAX_COUNT));
> The inner parens are not necessary AFAICS.

Provides better reading. Do you want to remove these?

> Also, doesn't this return a value which can be used for checking if
> _TIF_NEED_RESCHED is set instead of the condition below?

Yes, indeed - should I modify this check? (somehow I wanted to preserve
the original check)

Thank you,
Mihai

>> +
>> + if (loop_count < POLL_IDLE_RELAX_COUNT)
>> + break;
>> +
>> if (local_clock_noinstr() - time_start > limit) {
>> dev->poll_time_limit = true;
>> break;
>> --

2024-01-29 19:47:47

by Rafael J. Wysocki

Subject: Re: [PATCH v3 7/7] cpuidle/poll_state: replace cpu_relax with smp_cond_load_relaxed

On Mon, Jan 29, 2024 at 8:21 PM Mihai Carabas <mihai.carabas@oraclecom> wrote:
>
> La 29.01.2024 16:52, Rafael J. Wysocki a scris:
> > On Mon, Jan 29, 2024 at 12:56 PM Mihai Carabas <[email protected]> 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 <[email protected]>
> >> Signed-off-by: Mihai Carabas <[email protected]>
> >> ---
> >> drivers/cpuidle/poll_state.c | 14 +++++++++-----
> >> 1 file changed, 9 insertions(+), 5 deletions(-)
> >>
> >> diff --git a/drivers/cpuidle/poll_state.c b/drivers/cpuidle/poll_statec
> >> index 9b6d90a72601..440cd713e39a 100644
> >> --- a/drivers/cpuidle/poll_state.c
> >> +++ b/drivers/cpuidle/poll_state.c
> >> @@ -26,12 +26,16 @@ 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;
> >> +
> >> + smp_cond_load_relaxed(&current_thread_info()->flags,
> >> + (VAL & _TIF_NEED_RESCHED) ||
> >> + (loop_count++ >= POLL_IDLE_RELAX_COUNT));
> > The inner parens are not necessary AFAICS.
>
> Provides better reading. Do you want to remove these?

Whether or not it provides better reading is in the eye of the reader.
We seem to disagree here, because IMO redundant characters don't help
clarity.

> > Also, doesn't this return a value which can be used for checking if
> > _TIF_NEED_RESCHED is set instead of the condition below?
>
> Yes, indeed - should I modify this check? (somehow I wanted to preserve
> the original check)

But you haven't - it goes the other way around now.

In theory, _TIF_NEED_RESCHED may be set in the last iteration, in
which the check below will miss it, won't it?

> >> +
> >> + if (loop_count < POLL_IDLE_RELAX_COUNT)
> >> + break;
> >> +
> >> if (local_clock_noinstr() - time_start > limit) {
> >> dev->poll_time_limit = true;
> >> break;
> >> --