mbox series

[0/2] intel_pstate: fix turbo not being used after a processor is rebooted

Message ID 20221221155203.11347-1-ptyadav@amazon.de
Headers show
Series intel_pstate: fix turbo not being used after a processor is rebooted | expand

Message

Pratyush Yadav Dec. 21, 2022, 3:52 p.m. UTC
When a processor is brought offline and online again, it is unable to
use Turbo mode because the _PSS table does not contain the whole turbo
frequency range, but only +1 MHz above the max non-turbo frequency. This
causes problems when ACPI processor driver tries to set frequency
constraints. See patch 2 for more details.

Pratyush Yadav (2):
  acpi: processor: allow fixing up the frequency for a performance state
  cpufreq: intel_pstate: use acpi perflib to update turbo frequency

 drivers/acpi/processor_perflib.c | 40 ++++++++++++++++++++++++++++++++
 drivers/cpufreq/intel_pstate.c   |  5 ++--
 include/acpi/processor.h         |  2 ++
 3 files changed, 45 insertions(+), 2 deletions(-)

--
2.38.1

Comments

Srinivas Pandruvada Dec. 21, 2022, 9:34 p.m. UTC | #1
On Wed, 2022-12-21 at 16:52 +0100, Pratyush Yadav wrote:
> When a processor is brought offline and online again, it is unable to
> use Turbo mode because the _PSS table does not contain the whole
> turbo
> frequency range, but only +1 MHz above the max non-turbo frequency.
> This
> causes problems when ACPI processor driver tries to set frequency
> constraints. See patch 2 for more details.
> 
Are you using some _PPC constraint to force to limit frequency?
I did a offline/online with PPC=0 with no HWP, I can get to full turbo
range.

[  121.237752] smpboot: CPU 1 is now offline
[  125.734886] x86: Booting SMP configuration:
[  125.734892] smpboot: Booting Node 0 Processor 1 APIC 0x2
[  125.741007] intel_pstate: CPU 1 going online
[  125.741692] intel_pstate: CPU1 - ACPI _PSS perf data
[  125.741698] intel_pstate:      *P0: 2301 MHz, 28000 mW, 0x2a00
[  125.741703] intel_pstate:       P1: 2300 MHz, 28000 mW, 0x1700
[  125.741705] intel_pstate:       P2: 2200 MHz, 26297 mW, 0x1600
[  125.741707] intel_pstate:       P3: 2000 MHz, 23263 mW, 0x1400
[  125.741710] intel_pstate:       P4: 1900 MHz, 21924 mW, 0x1300
[  125.741712] intel_pstate:       P5: 1800 MHz, 20612 mW, 0x1200
[  125.741714] intel_pstate:       P6: 1600 MHz, 17812 mW, 0x1000
[  125.741716] intel_pstate:       P7: 1500 MHz, 16581 mW, 0xf00
[  125.741718] intel_pstate:       P8: 1300 MHz, 13946 mW, 0xd00
[  125.741720] intel_pstate:       P9: 1200 MHz, 12796 mW, 0xc00
[  125.741722] intel_pstate:       P10: 1100 MHz, 11426 mW, 0xb00
[  125.741724] intel_pstate:       P11: 900 MHz, 9250 mW, 0x900
[  125.741726] intel_pstate:       P12: 800 MHz, 7965 mW, 0x800
[  125.741729] intel_pstate:       P13: 700 MHz, 6940 mW, 0x700
[  125.741731] intel_pstate:       P14: 500 MHz, 4738 mW, 0x500
[  125.741733] intel_pstate:       P15: 400 MHz, 3787 mW, 0x400
[  125.741735] intel_pstate: _PPC limits will be enforced
[  125.741740] intel_pstate: policy->max > max non turbo frequency
[  125.741742] intel_pstate: cpu:1 min_policy_perf:4 max_policy_perf:42
[  125.741745] intel_pstate: cpu:1 global_min:4 global_max:42
[  125.741747] intel_pstate: cpu:1 max_perf_ratio:42 min_perf_ratio:4
[  125.742243] intel_pstate: policy->max > max non turbo frequency
[  125.742247] intel_pstate: cpu:1 min_policy_perf:4 max_policy_perf:42
[  125.742251] intel_pstate: cpu:1 global_min:4 global_max:42
[  125.742255] intel_pstate: cpu:1 max_perf_ratio:42 min_perf_ratio:4


It is not clear how to get to this non turbo situation.

Thanks,
Srinivas

> Pratyush Yadav (2):
>   acpi: processor: allow fixing up the frequency for a performance
> state
>   cpufreq: intel_pstate: use acpi perflib to update turbo frequency
> 
>  drivers/acpi/processor_perflib.c | 40
> ++++++++++++++++++++++++++++++++
>  drivers/cpufreq/intel_pstate.c   |  5 ++--
>  include/acpi/processor.h         |  2 ++
>  3 files changed, 45 insertions(+), 2 deletions(-)
> 
> --
> 2.38.1
>
Pratyush Yadav Dec. 22, 2022, 10:39 a.m. UTC | #2
Hi Srinivas,

On Wed, Dec 21 2022, srinivas pandruvada wrote:
> On Wed, 2022-12-21 at 16:52 +0100, Pratyush Yadav wrote:
>> When a processor is brought offline and online again, it is unable to
>> use Turbo mode because the _PSS table does not contain the whole
>> turbo
>> frequency range, but only +1 MHz above the max non-turbo frequency.
>> This
>> causes problems when ACPI processor driver tries to set frequency
>> constraints. See patch 2 for more details.
>>
> Are you using some _PPC constraint to force to limit frequency?
> I did a offline/online with PPC=0 with no HWP, I can get to full turbo
> range.
>
> [  121.237752] smpboot: CPU 1 is now offline
> [  125.734886] x86: Booting SMP configuration:
> [  125.734892] smpboot: Booting Node 0 Processor 1 APIC 0x2
> [  125.741007] intel_pstate: CPU 1 going online
> [  125.741692] intel_pstate: CPU1 - ACPI _PSS perf data
> [  125.741698] intel_pstate:      *P0: 2301 MHz, 28000 mW, 0x2a00
> [  125.741703] intel_pstate:       P1: 2300 MHz, 28000 mW, 0x1700
> [  125.741705] intel_pstate:       P2: 2200 MHz, 26297 mW, 0x1600
> [  125.741707] intel_pstate:       P3: 2000 MHz, 23263 mW, 0x1400
> [  125.741710] intel_pstate:       P4: 1900 MHz, 21924 mW, 0x1300
> [  125.741712] intel_pstate:       P5: 1800 MHz, 20612 mW, 0x1200
> [  125.741714] intel_pstate:       P6: 1600 MHz, 17812 mW, 0x1000
> [  125.741716] intel_pstate:       P7: 1500 MHz, 16581 mW, 0xf00
> [  125.741718] intel_pstate:       P8: 1300 MHz, 13946 mW, 0xd00
> [  125.741720] intel_pstate:       P9: 1200 MHz, 12796 mW, 0xc00
> [  125.741722] intel_pstate:       P10: 1100 MHz, 11426 mW, 0xb00
> [  125.741724] intel_pstate:       P11: 900 MHz, 9250 mW, 0x900
> [  125.741726] intel_pstate:       P12: 800 MHz, 7965 mW, 0x800
> [  125.741729] intel_pstate:       P13: 700 MHz, 6940 mW, 0x700
> [  125.741731] intel_pstate:       P14: 500 MHz, 4738 mW, 0x500
> [  125.741733] intel_pstate:       P15: 400 MHz, 3787 mW, 0x400
> [  125.741735] intel_pstate: _PPC limits will be enforced
> [  125.741740] intel_pstate: policy->max > max non turbo frequency
> [  125.741742] intel_pstate: cpu:1 min_policy_perf:4 max_policy_perf:42
> [  125.741745] intel_pstate: cpu:1 global_min:4 global_max:42
> [  125.741747] intel_pstate: cpu:1 max_perf_ratio:42 min_perf_ratio:4
> [  125.742243] intel_pstate: policy->max > max non turbo frequency
> [  125.742247] intel_pstate: cpu:1 min_policy_perf:4 max_policy_perf:42
> [  125.742251] intel_pstate: cpu:1 global_min:4 global_max:42
> [  125.742255] intel_pstate: cpu:1 max_perf_ratio:42 min_perf_ratio:4
>
>
> It is not clear how to get to this non turbo situation.

Look at the scaling_max_freq before and after rebooting the CPU. Before
you do it, it should be the max turbo frequency (say 2500 MHz). After
rebooting the CPU, it should now be 2301 MHz. So the kernel will now not
ask for anything above 2301 MHz, so you will never get to 2500 MHz.

Another interesting thing I observed is that if I reboot only 1 CPU, its
scaling_max_freq goes down to 2301, but it still keeps working at 2500
MHz. This might be something to do with how turbo works, I don't
understand that very well. But if you reboot say 20 CPUs, then you see
the frequency drop.

I use the below steps to reproduce this bug on my system, which has 40
CPUs with a base frequency of 2500 MHz and turbo frequency of 3300 MHz:

$ grep 'cpu MHz' /proc/cpuinfo
cpu MHz         : 3300.000
cpu MHz         : 1199.652
cpu MHz         : 3300.000
cpu MHz         : 3300.000
cpu MHz         : 3300.000
cpu MHz         : 3300.000
cpu MHz         : 3300.000
cpu MHz         : 3300.000
cpu MHz         : 3300.000
cpu MHz         : 3300.000
[ repeat 30 times ]

$ cat /sys/devices/system/cpu/cpu*/cpufreq/scaling_max_freq | sort -n | uniq -c
     40 3300000
$ for i in `seq 1 20`; do echo 0 | sudo tee /sys/devices/system/cpu/cpu$i/online; done
[...]
$ for i in `seq 1 20`; do echo 1 | sudo tee /sys/devices/system/cpu/cpu$i/online; done
[...]
$ grep 'cpu MHz' /proc/cpuinfo
cpu MHz         : 3300.000
cpu MHz         : 2500.000
cpu MHz         : 2500.000
cpu MHz         : 2500.000
cpu MHz         : 2500.000
cpu MHz         : 2500.000
[ repeat 15 times ]
cpu MHz         : 3300.000
cpu MHz         : 3300.000
[ repeat 17 times ]
$ cat /sys/devices/system/cpu/cpu*/cpufreq/scaling_max_freq | sort -n | uniq -c
     20 2501000
     20 3300000

>
> Thanks,
> Srinivas
>
>> Pratyush Yadav (2):
>>   acpi: processor: allow fixing up the frequency for a performance
>> state
>>   cpufreq: intel_pstate: use acpi perflib to update turbo frequency
>>
>>  drivers/acpi/processor_perflib.c | 40
>> ++++++++++++++++++++++++++++++++
>>  drivers/cpufreq/intel_pstate.c   |  5 ++--
>>  include/acpi/processor.h         |  2 ++
>>  3 files changed, 45 insertions(+), 2 deletions(-)
>>
>> --
>> 2.38.1
>>
>
Pratyush Yadav Dec. 22, 2022, 10:18 p.m. UTC | #3
On Thu, Dec 22 2022, Rafael J. Wysocki wrote:
> On Wed, Dec 21, 2022 at 4:52 PM Pratyush Yadav <ptyadav@amazon.de> wrote:
>>
>> In some cases the ACPI table can have an incorrect frequency populated
>> for a performance state. For example, in Intel platforms, the Turbo
>> frequency is just listed as +1 MHz above the max non-turbo frequency.
>
> Which is a known convention based on compatibility with some older OSes.

Interesting. I did not know that.

>
>> The frequency can actually go much higher based on various factors like
>> temperature, voltage, etc.
>
> It can.
>
>> Allow drivers like intel_pstate to fix up performance state frequencies
>> with the actual maximum value.
>
> Why do you want to do that?

To be able to use my processors at the full frequency they are capable
of. See [0] for more details.

[0] https://lore.kernel.org/linux-pm/mafs0k02jd8oh.fsf_-_@dev-dsk-ptyadav-1c-37607b33.eu-west-1.amazon.com/

>
[...]
Srinivas Pandruvada Dec. 23, 2022, 6:10 p.m. UTC | #4
Hi Pratyush,

On Thu, 2022-12-22 at 11:39 +0100, Pratyush Yadav wrote:
> 
> Hi Srinivas,
> 
> On Wed, Dec 21 2022, srinivas pandruvada wrote:
> > On Wed, 2022-12-21 at 16:52 +0100, Pratyush Yadav wrote:
> > > When a processor is brought offline and online again, it is
> > > unable to
> > > use Turbo mode because the _PSS table does not contain the whole
> > > turbo
> > > frequency range, but only +1 MHz above the max non-turbo
> > > frequency.
> > > This
> > > causes problems when ACPI processor driver tries to set frequency
> > > constraints. See patch 2 for more details.
> > > 
I can reproduce on a Broadwell server platform. But not on a client
system with acpi_ppc usage.

Need to check what change broke this.

Thanks,
Srinivas

> > 
> > Thanks,
> > Srinivas
> > 
> > > Pratyush Yadav (2):
> > >   acpi: processor: allow fixing up the frequency for a
> > > performance
> > > state
> > >   cpufreq: intel_pstate: use acpi perflib to update turbo
> > > frequency
> > > 
> > >  drivers/acpi/processor_perflib.c | 40
> > > ++++++++++++++++++++++++++++++++
> > >  drivers/cpufreq/intel_pstate.c   |  5 ++--
> > >  include/acpi/processor.h         |  2 ++
> > >  3 files changed, 45 insertions(+), 2 deletions(-)
> > > 
> > > --
> > > 2.38.1
> > > 
> > 
>
Srinivas Pandruvada Dec. 25, 2022, 12:28 a.m. UTC | #5
On Fri, 2022-12-23 at 10:10 -0800, srinivas pandruvada wrote:
> Hi Pratyush,
> 
> On Thu, 2022-12-22 at 11:39 +0100, Pratyush Yadav wrote:
> > 
> > Hi Srinivas,
> > 
> > On Wed, Dec 21 2022, srinivas pandruvada wrote:
> > > On Wed, 2022-12-21 at 16:52 +0100, Pratyush Yadav wrote:
> > > > When a processor is brought offline and online again, it is
> > > > unable to
> > > > use Turbo mode because the _PSS table does not contain the whole
> > > > turbo
> > > > frequency range, but only +1 MHz above the max non-turbo
> > > > frequency.
> > > > This
> > > > causes problems when ACPI processor driver tries to set frequency
> > > > constraints. See patch 2 for more details.
> > > > 
> I can reproduce on a Broadwell server platform. But not on a client
> system with acpi_ppc usage.
> 
> Need to check what change broke this.

When PPC limits enforcement changed to PM QOS, this broke. Previously
acpi_processor_get_platform_limit() was not enforcing any limits. It
was just setting variable. So any update done after
acpi_register_performance_state() call to pr->performance-
>states[ppc].core_frequency, was effective.

We don't really need to call
	ret = freq_qos_update_request(&pr->perflib_req,
			pr->performance->states[ppc].core_frequency *
1000);

if the PPC is not changed. When PPC is changed, this gets called again,
so then we can call the above function to update cpufreq limit.

The below change fixed for me.

diff --git a/drivers/acpi/processor_perflib.c
b/drivers/acpi/processor_perflib.c
index 757a98f6d7a2..c6ced89c00dd 100644
--- a/drivers/acpi/processor_perflib.c
+++ b/drivers/acpi/processor_perflib.c
@@ -75,6 +75,11 @@ static int acpi_processor_get_platform_limit(struct
acpi_processor *pr)
        pr_debug("CPU %d: _PPC is %d - frequency %s limited\n", pr->id,
                       (int)ppc, ppc ? "" : "not");
 
+       if (ppc == pr->performance_platform_limit) {
+               pr_debug("CPU %d: _PPC is %d - frequency not
changed\n", pr->id, ppc);
+               return 0;
+       }
+
        pr->performance_platform_limit = (int)ppc;
 
        if (ppc >= pr->performance->state_count ||

Thanks,
Srinivas

> 
> Thanks,
> Srinivas
> 
> > > 
> > > Thanks,
> > > Srinivas
> > > 
> > > > Pratyush Yadav (2):
> > > >   acpi: processor: allow fixing up the frequency for a
> > > > performance
> > > > state
> > > >   cpufreq: intel_pstate: use acpi perflib to update turbo
> > > > frequency
> > > > 
> > > >  drivers/acpi/processor_perflib.c | 40
> > > > ++++++++++++++++++++++++++++++++
> > > >  drivers/cpufreq/intel_pstate.c   |  5 ++--
> > > >  include/acpi/processor.h         |  2 ++
> > > >  3 files changed, 45 insertions(+), 2 deletions(-)
> > > > 
> > > > --
> > > > 2.38.1
> > > > 
> > > 
> > 
>
Pratyush Yadav Dec. 27, 2022, 3:38 p.m. UTC | #6
Hi Srinivas,

On Sat, Dec 24 2022, srinivas pandruvada wrote:

> On Fri, 2022-12-23 at 10:10 -0800, srinivas pandruvada wrote:
>> Hi Pratyush,
>>
>> On Thu, 2022-12-22 at 11:39 +0100, Pratyush Yadav wrote:
>> >
>> > Hi Srinivas,
>> >
>> > On Wed, Dec 21 2022, srinivas pandruvada wrote:
>> > > On Wed, 2022-12-21 at 16:52 +0100, Pratyush Yadav wrote:
>> > > > When a processor is brought offline and online again, it is
>> > > > unable to
>> > > > use Turbo mode because the _PSS table does not contain the whole
>> > > > turbo
>> > > > frequency range, but only +1 MHz above the max non-turbo
>> > > > frequency.
>> > > > This
>> > > > causes problems when ACPI processor driver tries to set frequency
>> > > > constraints. See patch 2 for more details.
>> > > >
>> I can reproduce on a Broadwell server platform. But not on a client
>> system with acpi_ppc usage.
>>
>> Need to check what change broke this.
>
> When PPC limits enforcement changed to PM QOS, this broke. Previously
> acpi_processor_get_platform_limit() was not enforcing any limits. It
> was just setting variable. So any update done after
> acpi_register_performance_state() call to pr->performance-
>>states[ppc].core_frequency, was effective.
>
> We don't really need to call
>         ret = freq_qos_update_request(&pr->perflib_req,
>                         pr->performance->states[ppc].core_frequency *
> 1000);
>
> if the PPC is not changed. When PPC is changed, this gets called again,
> so then we can call the above function to update cpufreq limit.
>
> The below change fixed for me.

Right. Should I re-roll my patches with your diff below then? Or do you
think my patches should be good to merge as-is?

>
> diff --git a/drivers/acpi/processor_perflib.c
> b/drivers/acpi/processor_perflib.c
> index 757a98f6d7a2..c6ced89c00dd 100644
> --- a/drivers/acpi/processor_perflib.c
> +++ b/drivers/acpi/processor_perflib.c
> @@ -75,6 +75,11 @@ static int acpi_processor_get_platform_limit(struct
> acpi_processor *pr)
>         pr_debug("CPU %d: _PPC is %d - frequency %s limited\n", pr->id,
>                        (int)ppc, ppc ? "" : "not");
>
> +       if (ppc == pr->performance_platform_limit) {
> +               pr_debug("CPU %d: _PPC is %d - frequency not
> changed\n", pr->id, ppc);
> +               return 0;
> +       }
> +
>         pr->performance_platform_limit = (int)ppc;
>
>         if (ppc >= pr->performance->state_count ||
>
Rafael J. Wysocki Dec. 27, 2022, 3:57 p.m. UTC | #7
On Tue, Dec 27, 2022 at 4:38 PM Pratyush Yadav <ptyadav@amazon.de> wrote:
>
> Hi Srinivas,
>
> On Sat, Dec 24 2022, srinivas pandruvada wrote:
>
> > On Fri, 2022-12-23 at 10:10 -0800, srinivas pandruvada wrote:
> >> Hi Pratyush,
> >>
> >> On Thu, 2022-12-22 at 11:39 +0100, Pratyush Yadav wrote:
> >> >
> >> > Hi Srinivas,
> >> >
> >> > On Wed, Dec 21 2022, srinivas pandruvada wrote:
> >> > > On Wed, 2022-12-21 at 16:52 +0100, Pratyush Yadav wrote:
> >> > > > When a processor is brought offline and online again, it is
> >> > > > unable to
> >> > > > use Turbo mode because the _PSS table does not contain the whole
> >> > > > turbo
> >> > > > frequency range, but only +1 MHz above the max non-turbo
> >> > > > frequency.
> >> > > > This
> >> > > > causes problems when ACPI processor driver tries to set frequency
> >> > > > constraints. See patch 2 for more details.
> >> > > >
> >> I can reproduce on a Broadwell server platform. But not on a client
> >> system with acpi_ppc usage.
> >>
> >> Need to check what change broke this.
> >
> > When PPC limits enforcement changed to PM QOS, this broke. Previously
> > acpi_processor_get_platform_limit() was not enforcing any limits. It
> > was just setting variable. So any update done after
> > acpi_register_performance_state() call to pr->performance-
> >>states[ppc].core_frequency, was effective.
> >
> > We don't really need to call
> >         ret = freq_qos_update_request(&pr->perflib_req,
> >                         pr->performance->states[ppc].core_frequency *
> > 1000);
> >
> > if the PPC is not changed. When PPC is changed, this gets called again,
> > so then we can call the above function to update cpufreq limit.
> >
> > The below change fixed for me.
>
> Right. Should I re-roll my patches with your diff below then? Or do you
> think my patches should be good to merge as-is?

No, they are not good to merge.
Rafael J. Wysocki Dec. 27, 2022, 4:15 p.m. UTC | #8
On Wed, Dec 21, 2022 at 4:52 PM Pratyush Yadav <ptyadav@amazon.de> wrote:
>
> When a processor is brought offline and online again, it is unable to
> use Turbo mode because the _PSS table does not contain the whole turbo
> frequency range, but only +1 MHz above the max non-turbo frequency.

That's because of the way P-state limits in the turbo range are
handled by the given processor.

Some of them restrict the P-state even if the limit is located within
the turbo range and some of them don't (that is, requesting any
P-state in the turbo range gives the processor a license to use the
whole of it).

> This causes problems when ACPI processor driver tries to set frequency
> constraints.

The problem is that acpi_processor_get_platform_limit() sets the limit
to the frequency for all of the _PSS states including the last special
one and it should update the QoS to "no limit" in that case.
Srinivas Pandruvada Dec. 27, 2022, 4:40 p.m. UTC | #9
On Tue, 2022-12-27 at 16:38 +0100, Pratyush Yadav wrote:
> Hi Srinivas,
> 
> On Sat, Dec 24 2022, srinivas pandruvada wrote:
> 
> > On Fri, 2022-12-23 at 10:10 -0800, srinivas pandruvada wrote:
> > > Hi Pratyush,
> > > 
> > > On Thu, 2022-12-22 at 11:39 +0100, Pratyush Yadav wrote:
> > > > 
> > > > Hi Srinivas,
> > > > 
> > > > On Wed, Dec 21 2022, srinivas pandruvada wrote:
> > > > > On Wed, 2022-12-21 at 16:52 +0100, Pratyush Yadav wrote:
> > > > > > When a processor is brought offline and online again, it is
> > > > > > unable to
> > > > > > use Turbo mode because the _PSS table does not contain the
> > > > > > whole
> > > > > > turbo
> > > > > > frequency range, but only +1 MHz above the max non-turbo
> > > > > > frequency.
> > > > > > This
> > > > > > causes problems when ACPI processor driver tries to set
> > > > > > frequency
> > > > > > constraints. See patch 2 for more details.
> > > > > > 
> > > I can reproduce on a Broadwell server platform. But not on a
> > > client
> > > system with acpi_ppc usage.
> > > 
> > > Need to check what change broke this.
> > 
> > When PPC limits enforcement changed to PM QOS, this broke.
> > Previously
> > acpi_processor_get_platform_limit() was not enforcing any limits.
> > It
> > was just setting variable. So any update done after
> > acpi_register_performance_state() call to pr->performance-
> > > states[ppc].core_frequency, was effective.
> > 
> > We don't really need to call
> >         ret = freq_qos_update_request(&pr->perflib_req,
> >                         pr->performance->states[ppc].core_frequency
> > *
> > 1000);
> > 
> > if the PPC is not changed. When PPC is changed, this gets called
> > again,
> > so then we can call the above function to update cpufreq limit.
> > 
> > The below change fixed for me.
> 
> Right. 
I think, this is the only change you require to fix this. In addition
set pr->performance_platform_limit = 0 in
acpi_processor_unregister_performance().

Thanks,
Srinivas

> Should I re-roll my patches with your diff below then? Or do you
> think my patches should be good to merge as-is?
> 
> > 
> > diff --git a/drivers/acpi/processor_perflib.c
> > b/drivers/acpi/processor_perflib.c
> > index 757a98f6d7a2..c6ced89c00dd 100644
> > --- a/drivers/acpi/processor_perflib.c
> > +++ b/drivers/acpi/processor_perflib.c
> > @@ -75,6 +75,11 @@ static int
> > acpi_processor_get_platform_limit(struct
> > acpi_processor *pr)
> >         pr_debug("CPU %d: _PPC is %d - frequency %s limited\n", pr-
> > >id,
> >                        (int)ppc, ppc ? "" : "not");
> > 
> > +       if (ppc == pr->performance_platform_limit) {
> > +               pr_debug("CPU %d: _PPC is %d - frequency not
> > changed\n", pr->id, ppc);
> > +               return 0;
> > +       }
> > +
> >         pr->performance_platform_limit = (int)ppc;
> > 
> >         if (ppc >= pr->performance->state_count ||
> > 
>
Rafael J. Wysocki Dec. 27, 2022, 5:02 p.m. UTC | #10
On Tue, Dec 27, 2022 at 5:40 PM srinivas pandruvada
<srinivas.pandruvada@linux.intel.com> wrote:
>
> On Tue, 2022-12-27 at 16:38 +0100, Pratyush Yadav wrote:
> > Hi Srinivas,
> >
> > On Sat, Dec 24 2022, srinivas pandruvada wrote:
> >
> > > On Fri, 2022-12-23 at 10:10 -0800, srinivas pandruvada wrote:
> > > > Hi Pratyush,
> > > >
> > > > On Thu, 2022-12-22 at 11:39 +0100, Pratyush Yadav wrote:
> > > > >
> > > > > Hi Srinivas,
> > > > >
> > > > > On Wed, Dec 21 2022, srinivas pandruvada wrote:
> > > > > > On Wed, 2022-12-21 at 16:52 +0100, Pratyush Yadav wrote:
> > > > > > > When a processor is brought offline and online again, it is
> > > > > > > unable to
> > > > > > > use Turbo mode because the _PSS table does not contain the
> > > > > > > whole
> > > > > > > turbo
> > > > > > > frequency range, but only +1 MHz above the max non-turbo
> > > > > > > frequency.
> > > > > > > This
> > > > > > > causes problems when ACPI processor driver tries to set
> > > > > > > frequency
> > > > > > > constraints. See patch 2 for more details.
> > > > > > >
> > > > I can reproduce on a Broadwell server platform. But not on a
> > > > client
> > > > system with acpi_ppc usage.
> > > >
> > > > Need to check what change broke this.
> > >
> > > When PPC limits enforcement changed to PM QOS, this broke.
> > > Previously
> > > acpi_processor_get_platform_limit() was not enforcing any limits.
> > > It
> > > was just setting variable. So any update done after
> > > acpi_register_performance_state() call to pr->performance-
> > > > states[ppc].core_frequency, was effective.
> > >
> > > We don't really need to call
> > >         ret = freq_qos_update_request(&pr->perflib_req,
> > >                         pr->performance->states[ppc].core_frequency
> > > *
> > > 1000);
> > >
> > > if the PPC is not changed. When PPC is changed, this gets called
> > > again,
> > > so then we can call the above function to update cpufreq limit.
> > >
> > > The below change fixed for me.
> >
> > Right.
> I think, this is the only change you require to fix this. In addition
> set pr->performance_platform_limit = 0 in
> acpi_processor_unregister_performance().

Not really, because if the limit is set to a lower frequency and then
reset to _PSS[0], it needs to be set back to "no limit".

I'll send a patch for that in a while.
Rafael J. Wysocki Dec. 27, 2022, 5:13 p.m. UTC | #11
On Tuesday, December 27, 2022 6:02:50 PM CET Rafael J. Wysocki wrote:
> On Tue, Dec 27, 2022 at 5:40 PM srinivas pandruvada
> <srinivas.pandruvada@linux.intel.com> wrote:
> >
> > On Tue, 2022-12-27 at 16:38 +0100, Pratyush Yadav wrote:
> > > Hi Srinivas,
> > >
> > > On Sat, Dec 24 2022, srinivas pandruvada wrote:
> > >
> > > > On Fri, 2022-12-23 at 10:10 -0800, srinivas pandruvada wrote:
> > > > > Hi Pratyush,
> > > > >
> > > > > On Thu, 2022-12-22 at 11:39 +0100, Pratyush Yadav wrote:
> > > > > >
> > > > > > Hi Srinivas,
> > > > > >
> > > > > > On Wed, Dec 21 2022, srinivas pandruvada wrote:
> > > > > > > On Wed, 2022-12-21 at 16:52 +0100, Pratyush Yadav wrote:
> > > > > > > > When a processor is brought offline and online again, it is
> > > > > > > > unable to
> > > > > > > > use Turbo mode because the _PSS table does not contain the
> > > > > > > > whole
> > > > > > > > turbo
> > > > > > > > frequency range, but only +1 MHz above the max non-turbo
> > > > > > > > frequency.
> > > > > > > > This
> > > > > > > > causes problems when ACPI processor driver tries to set
> > > > > > > > frequency
> > > > > > > > constraints. See patch 2 for more details.
> > > > > > > >
> > > > > I can reproduce on a Broadwell server platform. But not on a
> > > > > client
> > > > > system with acpi_ppc usage.
> > > > >
> > > > > Need to check what change broke this.
> > > >
> > > > When PPC limits enforcement changed to PM QOS, this broke.
> > > > Previously
> > > > acpi_processor_get_platform_limit() was not enforcing any limits.
> > > > It
> > > > was just setting variable. So any update done after
> > > > acpi_register_performance_state() call to pr->performance-
> > > > > states[ppc].core_frequency, was effective.
> > > >
> > > > We don't really need to call
> > > >         ret = freq_qos_update_request(&pr->perflib_req,
> > > >                         pr->performance->states[ppc].core_frequency
> > > > *
> > > > 1000);
> > > >
> > > > if the PPC is not changed. When PPC is changed, this gets called
> > > > again,
> > > > so then we can call the above function to update cpufreq limit.
> > > >
> > > > The below change fixed for me.
> > >
> > > Right.
> > I think, this is the only change you require to fix this. In addition
> > set pr->performance_platform_limit = 0 in
> > acpi_processor_unregister_performance().
> 
> Not really, because if the limit is set to a lower frequency and then
> reset to _PSS[0], it needs to be set back to "no limit".
> 
> I'll send a patch for that in a while.

This has not been tested beyond compilation, so please be careful.

---
 drivers/acpi/processor_perflib.c |   27 +++++++++++++++++++++------
 1 file changed, 21 insertions(+), 6 deletions(-)

Index: linux-pm/drivers/acpi/processor_perflib.c
===================================================================
--- linux-pm.orig/drivers/acpi/processor_perflib.c
+++ linux-pm/drivers/acpi/processor_perflib.c
@@ -53,6 +53,8 @@ static int acpi_processor_get_platform_l
 {
 	acpi_status status = 0;
 	unsigned long long ppc = 0;
+	s32 qos_value;
+	int index;
 	int ret;
 
 	if (!pr)
@@ -72,17 +74,30 @@ static int acpi_processor_get_platform_l
 		}
 	}
 
+	index = ppc;
+
+	if (pr->performance_platform_limit == index ||
+	    ppc >= pr->performance->state_count)
+		return 0;
+
 	pr_debug("CPU %d: _PPC is %d - frequency %s limited\n", pr->id,
-		       (int)ppc, ppc ? "" : "not");
+		 index, index ? "is" : "is not");
 
-	pr->performance_platform_limit = (int)ppc;
+	pr->performance_platform_limit = index;
 
-	if (ppc >= pr->performance->state_count ||
-	    unlikely(!freq_qos_request_active(&pr->perflib_req)))
+	if (unlikely(!freq_qos_request_active(&pr->perflib_req)))
 		return 0;
 
-	ret = freq_qos_update_request(&pr->perflib_req,
-			pr->performance->states[ppc].core_frequency * 1000);
+	/*
+	 * If _PPC returns 0, it means that all of the available states can be
+	 * used ("no limit").
+	 */
+	if (index == 0)
+		qos_value = FREQ_QOS_MAX_DEFAULT_VALUE;
+	else
+		qos_value = pr->performance->states[index].core_frequency * 1000;
+
+	ret = freq_qos_update_request(&pr->perflib_req, qos_value);
 	if (ret < 0) {
 		pr_warn("Failed to update perflib freq constraint: CPU%d (%d)\n",
 			pr->id, ret);
Srinivas Pandruvada Dec. 27, 2022, 6:07 p.m. UTC | #12
On Tue, 2022-12-27 at 18:02 +0100, Rafael J. Wysocki wrote:
> On Tue, Dec 27, 2022 at 5:40 PM srinivas pandruvada
> <srinivas.pandruvada@linux.intel.com> wrote:
> > 
> > On Tue, 2022-12-27 at 16:38 +0100, Pratyush Yadav wrote:
> > > Hi Srinivas,
> > > 
> > > On Sat, Dec 24 2022, srinivas pandruvada wrote:
> > > 
> > > > On Fri, 2022-12-23 at 10:10 -0800, srinivas pandruvada wrote:
> > > > > Hi Pratyush,
> > > > > 
> > > > > On Thu, 2022-12-22 at 11:39 +0100, Pratyush Yadav wrote:
> > > > > > 
> > > > > > Hi Srinivas,
> > > > > > 
> > > > > > On Wed, Dec 21 2022, srinivas pandruvada wrote:
> > > > > > > On Wed, 2022-12-21 at 16:52 +0100, Pratyush Yadav wrote:
> > > > > > > > When a processor is brought offline and online again,
> > > > > > > > it is
> > > > > > > > unable to
> > > > > > > > use Turbo mode because the _PSS table does not contain
> > > > > > > > the
> > > > > > > > whole
> > > > > > > > turbo
> > > > > > > > frequency range, but only +1 MHz above the max non-
> > > > > > > > turbo
> > > > > > > > frequency.
> > > > > > > > This
> > > > > > > > causes problems when ACPI processor driver tries to set
> > > > > > > > frequency
> > > > > > > > constraints. See patch 2 for more details.
> > > > > > > > 
> > > > > I can reproduce on a Broadwell server platform. But not on a
> > > > > client
> > > > > system with acpi_ppc usage.
> > > > > 
> > > > > Need to check what change broke this.
> > > > 
> > > > When PPC limits enforcement changed to PM QOS, this broke.
> > > > Previously
> > > > acpi_processor_get_platform_limit() was not enforcing any
> > > > limits.
> > > > It
> > > > was just setting variable. So any update done after
> > > > acpi_register_performance_state() call to pr->performance-
> > > > > states[ppc].core_frequency, was effective.
> > > > 
> > > > We don't really need to call
> > > >         ret = freq_qos_update_request(&pr->perflib_req,
> > > >                         pr->performance-
> > > > >states[ppc].core_frequency
> > > > *
> > > > 1000);
> > > > 
> > > > if the PPC is not changed. When PPC is changed, this gets
> > > > called
> > > > again,
> > > > so then we can call the above function to update cpufreq limit.
> > > > 
> > > > The below change fixed for me.
> > > 
> > > Right.
> > I think, this is the only change you require to fix this. In
> > addition
> > set pr->performance_platform_limit = 0 in
> > acpi_processor_unregister_performance().
> 
> Not really, because if the limit is set to a lower frequency and then
> reset to _PSS[0], it needs to be set back to "no limit".
> 

If PPC becomes 0 again after ppc > 0 during dynamic PPC change, pr-
>performance_platform_limit will not match current PPC, so will set to
PPC 0 performance ( which is already patched by driver after return
from acpi_register_performance_state()).

But fine, you can always set freq qos to FREQ_QOS_MAX_DEFAULT_VALUE for
PPC 0 as you are doing in your patch.

Thanks,
Srinivas




> I'll send a patch for that in a while.
Rafael J. Wysocki Dec. 27, 2022, 6:47 p.m. UTC | #13
On Tue, Dec 27, 2022 at 7:07 PM srinivas pandruvada
<srinivas.pandruvada@linux.intel.com> wrote:
>
> On Tue, 2022-12-27 at 18:02 +0100, Rafael J. Wysocki wrote:
> > On Tue, Dec 27, 2022 at 5:40 PM srinivas pandruvada
> > <srinivas.pandruvada@linux.intel.com> wrote:
> > >
> > > On Tue, 2022-12-27 at 16:38 +0100, Pratyush Yadav wrote:
> > > > Hi Srinivas,
> > > >
> > > > On Sat, Dec 24 2022, srinivas pandruvada wrote:
> > > >
> > > > > On Fri, 2022-12-23 at 10:10 -0800, srinivas pandruvada wrote:
> > > > > > Hi Pratyush,
> > > > > >
> > > > > > On Thu, 2022-12-22 at 11:39 +0100, Pratyush Yadav wrote:
> > > > > > >
> > > > > > > Hi Srinivas,
> > > > > > >
> > > > > > > On Wed, Dec 21 2022, srinivas pandruvada wrote:
> > > > > > > > On Wed, 2022-12-21 at 16:52 +0100, Pratyush Yadav wrote:
> > > > > > > > > When a processor is brought offline and online again,
> > > > > > > > > it is
> > > > > > > > > unable to
> > > > > > > > > use Turbo mode because the _PSS table does not contain
> > > > > > > > > the
> > > > > > > > > whole
> > > > > > > > > turbo
> > > > > > > > > frequency range, but only +1 MHz above the max non-
> > > > > > > > > turbo
> > > > > > > > > frequency.
> > > > > > > > > This
> > > > > > > > > causes problems when ACPI processor driver tries to set
> > > > > > > > > frequency
> > > > > > > > > constraints. See patch 2 for more details.
> > > > > > > > >
> > > > > > I can reproduce on a Broadwell server platform. But not on a
> > > > > > client
> > > > > > system with acpi_ppc usage.
> > > > > >
> > > > > > Need to check what change broke this.
> > > > >
> > > > > When PPC limits enforcement changed to PM QOS, this broke.
> > > > > Previously
> > > > > acpi_processor_get_platform_limit() was not enforcing any
> > > > > limits.
> > > > > It
> > > > > was just setting variable. So any update done after
> > > > > acpi_register_performance_state() call to pr->performance-
> > > > > > states[ppc].core_frequency, was effective.
> > > > >
> > > > > We don't really need to call
> > > > >         ret = freq_qos_update_request(&pr->perflib_req,
> > > > >                         pr->performance-
> > > > > >states[ppc].core_frequency
> > > > > *
> > > > > 1000);
> > > > >
> > > > > if the PPC is not changed. When PPC is changed, this gets
> > > > > called
> > > > > again,
> > > > > so then we can call the above function to update cpufreq limit.
> > > > >
> > > > > The below change fixed for me.
> > > >
> > > > Right.
> > > I think, this is the only change you require to fix this. In
> > > addition
> > > set pr->performance_platform_limit = 0 in
> > > acpi_processor_unregister_performance().
> >
> > Not really, because if the limit is set to a lower frequency and then
> > reset to _PSS[0], it needs to be set back to "no limit".
> >
>
> If PPC becomes 0 again after ppc > 0 during dynamic PPC change, pr-
> >performance_platform_limit will not match current PPC, so will set to
> PPC 0 performance ( which is already patched by driver after return
> from acpi_register_performance_state()).

I see.

> But fine, you can always set freq qos to FREQ_QOS_MAX_DEFAULT_VALUE for
> PPC 0 as you are doing in your patch.

I think that using the "no limit" value to represent the "no limit"
condition makes sense.

Also, I'm wondering if the patching of states[0].core_frequency will
still be necessary after this change.
Srinivas Pandruvada Dec. 27, 2022, 6:49 p.m. UTC | #14
On Tue, 2022-12-27 at 19:47 +0100, Rafael J. Wysocki wrote:
> On Tue, Dec 27, 2022 at 7:07 PM srinivas pandruvada
> <srinivas.pandruvada@linux.intel.com> wrote:
> > 
> > On Tue, 2022-12-27 at 18:02 +0100, Rafael J. Wysocki wrote:
> > > On Tue, Dec 27, 2022 at 5:40 PM srinivas pandruvada
> > > <srinivas.pandruvada@linux.intel.com> wrote:
> > > > 
> > > > On Tue, 2022-12-27 at 16:38 +0100, Pratyush Yadav wrote:
> > > > > Hi Srinivas,
> > > > > 
> > > > > On Sat, Dec 24 2022, srinivas pandruvada wrote:
> > > > > 
> > > > > > On Fri, 2022-12-23 at 10:10 -0800, srinivas pandruvada
> > > > > > wrote:
> > > > > > > Hi Pratyush,
> > > > > > > 
> > > > > > > On Thu, 2022-12-22 at 11:39 +0100, Pratyush Yadav wrote:
> > > > > > > > 
> > > > > > > > Hi Srinivas,
> > > > > > > > 
> > > > > > > > On Wed, Dec 21 2022, srinivas pandruvada wrote:
> > > > > > > > > On Wed, 2022-12-21 at 16:52 +0100, Pratyush Yadav
> > > > > > > > > wrote:
> > > > > > > > > > When a processor is brought offline and online
> > > > > > > > > > again,
> > > > > > > > > > it is
> > > > > > > > > > unable to
> > > > > > > > > > use Turbo mode because the _PSS table does not
> > > > > > > > > > contain
> > > > > > > > > > the
> > > > > > > > > > whole
> > > > > > > > > > turbo
> > > > > > > > > > frequency range, but only +1 MHz above the max non-
> > > > > > > > > > turbo
> > > > > > > > > > frequency.
> > > > > > > > > > This
> > > > > > > > > > causes problems when ACPI processor driver tries to
> > > > > > > > > > set
> > > > > > > > > > frequency
> > > > > > > > > > constraints. See patch 2 for more details.
> > > > > > > > > > 
> > > > > > > I can reproduce on a Broadwell server platform. But not
> > > > > > > on a
> > > > > > > client
> > > > > > > system with acpi_ppc usage.
> > > > > > > 
> > > > > > > Need to check what change broke this.
> > > > > > 
> > > > > > When PPC limits enforcement changed to PM QOS, this broke.
> > > > > > Previously
> > > > > > acpi_processor_get_platform_limit() was not enforcing any
> > > > > > limits.
> > > > > > It
> > > > > > was just setting variable. So any update done after
> > > > > > acpi_register_performance_state() call to pr->performance-
> > > > > > > states[ppc].core_frequency, was effective.
> > > > > > 
> > > > > > We don't really need to call
> > > > > >         ret = freq_qos_update_request(&pr->perflib_req,
> > > > > >                         pr->performance-
> > > > > > > states[ppc].core_frequency
> > > > > > *
> > > > > > 1000);
> > > > > > 
> > > > > > if the PPC is not changed. When PPC is changed, this gets
> > > > > > called
> > > > > > again,
> > > > > > so then we can call the above function to update cpufreq
> > > > > > limit.
> > > > > > 
> > > > > > The below change fixed for me.
> > > > > 
> > > > > Right.
> > > > I think, this is the only change you require to fix this. In
> > > > addition
> > > > set pr->performance_platform_limit = 0 in
> > > > acpi_processor_unregister_performance().
> > > 
> > > Not really, because if the limit is set to a lower frequency and
> > > then
> > > reset to _PSS[0], it needs to be set back to "no limit".
> > > 
> > 
> > If PPC becomes 0 again after ppc > 0 during dynamic PPC change, pr-
> > > performance_platform_limit will not match current PPC, so will
> > > set to
> > PPC 0 performance ( which is already patched by driver after return
> > from acpi_register_performance_state()).
> 
> I see.
> 
> > But fine, you can always set freq qos to FREQ_QOS_MAX_DEFAULT_VALUE
> > for
> > PPC 0 as you are doing in your patch.
> 
> I think that using the "no limit" value to represent the "no limit"
> condition makes sense.
Agree.

> 
> Also, I'm wondering if the patching of states[0].core_frequency will
> still be necessary after this change.

I don't think so. We can remove the patching.
Rafael J. Wysocki Dec. 27, 2022, 6:54 p.m. UTC | #15
On Tue, Dec 27, 2022 at 7:49 PM srinivas pandruvada
<srinivas.pandruvada@linux.intel.com> wrote:
>
> On Tue, 2022-12-27 at 19:47 +0100, Rafael J. Wysocki wrote:
> > On Tue, Dec 27, 2022 at 7:07 PM srinivas pandruvada
> > <srinivas.pandruvada@linux.intel.com> wrote:
> > >
> > > On Tue, 2022-12-27 at 18:02 +0100, Rafael J. Wysocki wrote:
> > > > On Tue, Dec 27, 2022 at 5:40 PM srinivas pandruvada
> > > > <srinivas.pandruvada@linux.intel.com> wrote:
> > > > >
> > > > > On Tue, 2022-12-27 at 16:38 +0100, Pratyush Yadav wrote:
> > > > > > Hi Srinivas,
> > > > > >
> > > > > > On Sat, Dec 24 2022, srinivas pandruvada wrote:
> > > > > >
> > > > > > > On Fri, 2022-12-23 at 10:10 -0800, srinivas pandruvada
> > > > > > > wrote:
> > > > > > > > Hi Pratyush,
> > > > > > > >
> > > > > > > > On Thu, 2022-12-22 at 11:39 +0100, Pratyush Yadav wrote:
> > > > > > > > >
> > > > > > > > > Hi Srinivas,
> > > > > > > > >
> > > > > > > > > On Wed, Dec 21 2022, srinivas pandruvada wrote:
> > > > > > > > > > On Wed, 2022-12-21 at 16:52 +0100, Pratyush Yadav
> > > > > > > > > > wrote:
> > > > > > > > > > > When a processor is brought offline and online
> > > > > > > > > > > again,
> > > > > > > > > > > it is
> > > > > > > > > > > unable to
> > > > > > > > > > > use Turbo mode because the _PSS table does not
> > > > > > > > > > > contain
> > > > > > > > > > > the
> > > > > > > > > > > whole
> > > > > > > > > > > turbo
> > > > > > > > > > > frequency range, but only +1 MHz above the max non-
> > > > > > > > > > > turbo
> > > > > > > > > > > frequency.
> > > > > > > > > > > This
> > > > > > > > > > > causes problems when ACPI processor driver tries to
> > > > > > > > > > > set
> > > > > > > > > > > frequency
> > > > > > > > > > > constraints. See patch 2 for more details.
> > > > > > > > > > >
> > > > > > > > I can reproduce on a Broadwell server platform. But not
> > > > > > > > on a
> > > > > > > > client
> > > > > > > > system with acpi_ppc usage.
> > > > > > > >
> > > > > > > > Need to check what change broke this.
> > > > > > >
> > > > > > > When PPC limits enforcement changed to PM QOS, this broke.
> > > > > > > Previously
> > > > > > > acpi_processor_get_platform_limit() was not enforcing any
> > > > > > > limits.
> > > > > > > It
> > > > > > > was just setting variable. So any update done after
> > > > > > > acpi_register_performance_state() call to pr->performance-
> > > > > > > > states[ppc].core_frequency, was effective.
> > > > > > >
> > > > > > > We don't really need to call
> > > > > > >         ret = freq_qos_update_request(&pr->perflib_req,
> > > > > > >                         pr->performance-
> > > > > > > > states[ppc].core_frequency
> > > > > > > *
> > > > > > > 1000);
> > > > > > >
> > > > > > > if the PPC is not changed. When PPC is changed, this gets
> > > > > > > called
> > > > > > > again,
> > > > > > > so then we can call the above function to update cpufreq
> > > > > > > limit.
> > > > > > >
> > > > > > > The below change fixed for me.
> > > > > >
> > > > > > Right.
> > > > > I think, this is the only change you require to fix this. In
> > > > > addition
> > > > > set pr->performance_platform_limit = 0 in
> > > > > acpi_processor_unregister_performance().
> > > >
> > > > Not really, because if the limit is set to a lower frequency and
> > > > then
> > > > reset to _PSS[0], it needs to be set back to "no limit".
> > > >
> > >
> > > If PPC becomes 0 again after ppc > 0 during dynamic PPC change, pr-
> > > > performance_platform_limit will not match current PPC, so will
> > > > set to
> > > PPC 0 performance ( which is already patched by driver after return
> > > from acpi_register_performance_state()).
> >
> > I see.
> >
> > > But fine, you can always set freq qos to FREQ_QOS_MAX_DEFAULT_VALUE
> > > for
> > > PPC 0 as you are doing in your patch.
> >
> > I think that using the "no limit" value to represent the "no limit"
> > condition makes sense.
> Agree.
>
> >
> > Also, I'm wondering if the patching of states[0].core_frequency will
> > still be necessary after this change.
>
> I don't think so. We can remove the patching.

OK, let me cut the patches.