Re: [PATCH v2] ACPI: Replace msleep() with usleep_range() in acpi_os_sleep().

[Hans de Goede <hdegoede@redhat.com>]

Hi Rafael,

On 18-Nov-24 1:02 PM, Rafael J. Wysocki wrote:
> Hi Hans,
> 
> On Mon, Nov 18, 2024 at 12:38 PM Hans de Goede <hdegoede@redhat.com> wrote:
>>
>> Hi Rafael, Len,
>>
>> On 18-Nov-24 12:03 PM, Rafael J. Wysocki wrote:
>>> On Sat, Nov 16, 2024 at 12:11 AM Len Brown <lenb@kernel.org> wrote:
>>>>
>>>> From: Len Brown <len.brown@intel.com>
>>>>
>>>> Replace msleep() with usleep_range() in acpi_os_sleep().
>>>>
>>>> This has a significant user-visible performance benefit
>>>> on some ACPI flows on some systems.  eg. Kernel resume
>>>> time of a Dell XPS-13-9300 drops from 1943ms to 1127ms (42%).
>>>
>>> Sure.
>>>
>>> And the argument seems to be that it is better to always use more
>>> resources in a given path (ACPI sleep in this particular case) than to
>>> be somewhat inaccurate which is visible in some cases.
>>>
>>> This would mean that hrtimers should always be used everywhere, but they aren't.
>>>
>>> While I have nothing against addressing the short sleeps issue where
>>> the msleep() inaccuracy is too large, I don't see why this requires
>>> using a hrtimer with no slack in all cases.
>>>
>>> The argument seems to be that the short sleeps case is hard to
>>> distinguish from the other cases, but I'm not sure about this.
>>>
>>> Also, something like this might work, but for some reason you don't
>>> want to do it:
>>>
>>> if (ms >= 12 * MSEC_PER_SEC / HZ) {
>>>         msleep(ms);
>>> } else {
>>>        u64 us = ms * USEC_PER_MSEC;
>>>
>>>       usleep_range(us, us / 8);
> 
> Should be
> 
>       usleep_range(us, us + us / 8);
> 
> (I notoriously confuse this API).

I see.

>>> }
>>
>> FWIW I was thinking the same thing, that it would be good to still
>> use msleep when the sleep is > (MSEC_PER_SEC / HZ), not sure
>> why you added the 12 there ? Surely something like a sleep longer
>> then 3 timerticks (I know we have NOHZ but still) would already be
>> long enough to not worry about msleep slack ?
> 
> The typical msleep() overhead in 6.12 appears to be 1.5 jiffy which is
> 1.5 * MSEC_PER_SEC / HZ and I want the usleep() delta to be less than
> this, so
> 
> delta = ms / 8 <= 1.5 * MSEC_PER_SEC / HZ

Ok, that makes sense. But this probably requires a comment explaining
this so that when someone looks at this in the future they understand
where the 12 comes from.

Where as the / 8 is just a choice right? I think it is decent choice,
but still this is just a value you picked which should work nicely,
right ?

>> OTOH it is not like we will hit these ACPI acpi_os_sleep()
>> calls multiple times per second all the time. On a normal idle
>> system I expect there to not be that many calls (could still
>> be a few from ACPI managed devices going into + out of
>> runtime-pm regularly). And if don't hit acpi_os_sleep() calls
>> multiple times per second then the chances of time coalescing
>> are not that big anyways.
>>
>> Still I think that finding something middle ground between always
>> sleeping the exact min time and the old msleep() call, as Rafael
>> is proposing, would be good IMHO.
> 
> Thanks for the feedback!

You're welcome.

Len any chance you can give Rafael's proposal a test run on the
same Dell XPS 13 9300 and see what this means for the resume time ?

If this gets close enough to your patch I think we should go with
what Rafael is proposing.

Regards,

Hans




>>>> usleep_range(min, min) is used because there is scant
>>>> opportunity for timer coalescing during ACPI flows
>>>> related to system suspend, resume (or initialization).
>>>>
>>>> ie. During these flows usleep_range(min, max) is observed to
>>>> be effectvely be the same as usleep_range(max, max).
>>>>
>>>> Similarly, msleep() for long sleeps is not considered because
>>>> these flows almost never have opportunities to coalesce
>>>> with other activity on jiffie boundaries, leaving no
>>>> measurably benefit to rounding up to jiffie boundaries.
>>>>
>>>> Background:
>>>>
>>>> acpi_os_sleep() supports the ACPI AML Sleep(msec) operator,
>>>> and it must not return before the requested number of msec.
>>>>
>>>> Until Linux-3.13, this contract was sometimes violated by using
>>>> schedule_timeout_interruptible(j), which could return early.
>>>>
>>>> Since Linux-3.13, acpi_os_sleep() uses msleep(),
>>>> which doesn't return early, but is still subject
>>>> to long delays due to the low resolution of the jiffie clock.
>>>>
>>>> Linux-6.12 removed a stray jiffie from msleep: commit 4381b895f544
>>>> ("timers: Remove historical extra jiffie for timeout in msleep()")
>>>> The 4ms savings is material for some durations,
>>>> but msleep is still generally too course. eg msleep(5)
>>>> on a 250HZ system still takes 11.9ms.
>>>>
>>>> System resume performance of a Dell XPS 13 9300:
>>>>
>>>> Linux-6.11:
>>>> msleep HZ 250   2460 ms
>>>>
>>>> Linux-6.12:
>>>> msleep HZ 250   1943 ms
>>>> msleep HZ 1000  1233 ms
>>>> usleep HZ 250   1127 ms
>>>> usleep HZ 1000  1130 ms
>>>>
>>>> Closes: https://bugzilla.kernel.org/show_bug.cgi?id=216263
>>>> Signed-off-by: Len Brown <len.brown@intel.com>
>>>> Suggested-by: Arjan van de Ven <arjan@linux.intel.com>
>>>> Tested-by: Todd Brandt <todd.e.brandt@intel.com>
>>>> ---
>>>>  drivers/acpi/osl.c | 4 +++-
>>>>  1 file changed, 3 insertions(+), 1 deletion(-)
>>>>
>>>> diff --git a/drivers/acpi/osl.c b/drivers/acpi/osl.c
>>>> index 70af3fbbebe5..daf87e33b8ea 100644
>>>> --- a/drivers/acpi/osl.c
>>>> +++ b/drivers/acpi/osl.c
>>>> @@ -607,7 +607,9 @@ acpi_status acpi_os_remove_interrupt_handler(u32 gsi, acpi_osd_handler handler)
>>>>
>>>>  void acpi_os_sleep(u64 ms)
>>>>  {
>>>> -       msleep(ms);
>>>> +       u64 us = ms * USEC_PER_MSEC;
>>>> +
>>>> +       usleep_range(us, us);
>>>>  }
>>>>
>>>>  void acpi_os_stall(u32 us)
>>>> --
>>>> 2.43.0
>>>>
>>>
>>
>