diff mbox series

[v2,2/2] cpufreq/amd-pstate: Fix the scaling_max_freq setting on shared memory CPPC systems

Message ID 20240702081413.5688-3-Dhananjay.Ugwekar@amd.com
State Accepted
Commit 738d7d03571c7e38565bd245c0815a2c74665018
Headers show
Series AMD Pstate driver fixes | expand

Commit Message

Dhananjay Ugwekar July 2, 2024, 8:14 a.m. UTC
On shared memory CPPC systems, with amd_pstate=active mode, the change
in scaling_max_freq doesn't get written to the shared memory
region. Due to this, the writes to the scaling_max_freq sysfs file
don't take effect. Fix this by propagating the scaling_max_freq
changes to the shared memory region.

Fixes: ffa5096a7c33 ("cpufreq: amd-pstate: implement Pstate EPP support for the AMD processors")
Reported-by: David Arcari <darcari@redhat.com>
Signed-off-by: Dhananjay Ugwekar <Dhananjay.Ugwekar@amd.com>
---
 drivers/cpufreq/amd-pstate.c | 43 +++++++++++++++++++-----------------
 1 file changed, 23 insertions(+), 20 deletions(-)

Comments

Jones, Morgan Sept. 3, 2024, 5:51 p.m. UTC | #1
Hi there,

We are experiencing a ~35% performance regression on an AMD EPYC 7702 64 core machine after applying this patch. We observed Linux 6.6.44 starting to cause the issue, and performed a bisect involving rebooting the machine over 15 times. (Note that kexec didn't successfully identify the problem, since the PM memory mailbox is never reset).

It appears that we are getting a max of 2.18 GHz when this CPU can boost to 3.35 GHz, explaining the slowdown. Blacklisting amd-pstate solves the issue at the expense of the performance increase we used to get by using it.

Is it possible that the upper limits were implicitly at the max CPU power before, and setting the upper limit to something other than the boost frequency can reduce performance now?

# bad: [7213910600667c51c978e577bf5454d3f7b313b7] Linux 6.6.44
# good: [58b0425ff5df680d0b67f64ae1f3f1ebdf1c4de9] Linux 6.6.43
git bisect start '7213910600667c51c978e577bf5454d3f7b313b7' '58b0425ff5df680d0b67f64ae1f3f1ebdf1c4de9'
# good: [72ff9d26964a3a80f7650df719df139f5c1f965d] arm64: dts: qcom: sm6350: Add missing qcom,non-secure-domain property
git bisect good 72ff9d26964a3a80f7650df719df139f5c1f965d
# good: [0fffc2e1bf40a2220ef5a38f834ea063dba832d3] ARM: dts: sunxi: remove duplicated entries in makefile
git bisect good 0fffc2e1bf40a2220ef5a38f834ea063dba832d3
# bad: [8cdbe6ebfd1763a5c41a2a3058497c0a9163311c] pinctrl: renesas: r8a779g0: Fix CANFD5 suffix
git bisect bad 8cdbe6ebfd1763a5c41a2a3058497c0a9163311c
# bad: [5dbb98e7fa42bebc1325899193d8f13f0705a148] drm/mediatek: Turn off the layers with zero width or height
git bisect bad 5dbb98e7fa42bebc1325899193d8f13f0705a148
# bad: [691ec7043122c9c8c46d84f6e6cd85d13d50cd93] selftests/bpf: Null checks for links in bpf_tcp_ca
git bisect bad 691ec7043122c9c8c46d84f6e6cd85d13d50cd93
# bad: [a1359e085d75d7393a250054e66c0a7bc6c3dbfa] perf/x86: Serialize set_attr_rdpmc()
git bisect bad a1359e085d75d7393a250054e66c0a7bc6c3dbfa
# bad: [e99d9b16ff153de9540073239d24adc3b0a3a997] wifi: ath12k: change DMA direction while mapping reinjected packets
git bisect bad e99d9b16ff153de9540073239d24adc3b0a3a997
# bad: [d027ac4a08541beb2a89563d3e034da7085050ba] firmware: turris-mox-rwtm: Initialize completion before mailbox
git bisect bad d027ac4a08541beb2a89563d3e034da7085050ba
# bad: [e6c9eca327e6a41a81e7eba0d0ddc13da37f82a1] ARM: spitz: fix GPIO assignment for backlight
git bisect bad e6c9eca327e6a41a81e7eba0d0ddc13da37f82a1
# bad: [b8cdefdaa555bbfc269c2198803f8791a8923960] m68k: cmpxchg: Fix return value for default case in __arch_xchg()
git bisect bad b8cdefdaa555bbfc269c2198803f8791a8923960
# bad: [13a71384ae6a8779da809b00c6f378dcead10427] cpufreq/amd-pstate: Fix the scaling_max_freq setting on shared memory CPPC systems
git bisect bad 13a71384ae6a8779da809b00c6f378dcead10427
# first bad commit: [13a71384ae6a8779da809b00c6f378dcead10427] cpufreq/amd-pstate: Fix the scaling_max_freq setting on shared memory CPPC systems  

cpupower output:

analyzing CPU 47:
  driver: amd-pstate-epp
  CPUs which run at the same hardware frequency: 47
  CPUs which need to have their frequency coordinated by software: 47
  maximum transition latency:  Cannot determine or is not supported.
  hardware limits: 400 MHz - 2.18 GHz
  available cpufreq governors: performance powersave
  current policy: frequency should be within 400 MHz and 2.18 GHz.
                  The governor "performance" may decide which speed to use
                  within this range.
  current CPU frequency: Unable to call hardware
  current CPU frequency: 2.17 GHz (asserted by call to kernel)
  boost state support:
    Supported: yes
    Active: yes
    AMD PSTATE Highest Performance: 166. Maximum Frequency: 2.18 GHz.
    AMD PSTATE Nominal Performance: 152. Nominal Frequency: 2.00 GHz.
    AMD PSTATE Lowest Non-linear Performance: 115. Lowest Non-linear Frequency: 1.51 GHz.
    AMD PSTATE Lowest Performance: 31. Lowest Frequency: 400 MHz.

Thanks,
Morgan

-----Original Message-----
From: Mario Limonciello <mario.limonciello@amd.com> 
Sent: Tuesday, July 2, 2024 10:49 AM
To: Dhananjay Ugwekar <Dhananjay.Ugwekar@amd.com>; rafael@kernel.org; viresh.kumar@linaro.org; gautham.shenoy@amd.com; perry.yuan@amd.com; skhan@linuxfoundation.org; li.meng@amd.com; ray.huang@amd.com
Cc: linux-pm@vger.kernel.org; linux-kernel@vger.kernel.org; David Arcari <darcari@redhat.com>
Subject: Re: [PATCH v2 2/2] cpufreq/amd-pstate: Fix the scaling_max_freq setting on shared memory CPPC systems

On 7/2/2024 3:14, Dhananjay Ugwekar wrote:
> On shared memory CPPC systems, with amd_pstate=active mode, the change 
> in scaling_max_freq doesn't get written to the shared memory region. 
> Due to this, the writes to the scaling_max_freq sysfs file don't take 
> effect. Fix this by propagating the scaling_max_freq changes to the 
> shared memory region.
> 
> Fixes: ffa5096a7c33 ("cpufreq: amd-pstate: implement Pstate EPP 
> support for the AMD processors")
> Reported-by: David Arcari <darcari@redhat.com>
> Signed-off-by: Dhananjay Ugwekar <Dhananjay.Ugwekar@amd.com>
Reviewed-by: Mario Limonciello <mario.limonciello@amd.com>
> ---
>   drivers/cpufreq/amd-pstate.c | 43 +++++++++++++++++++-----------------
>   1 file changed, 23 insertions(+), 20 deletions(-)
> 
> diff --git a/drivers/cpufreq/amd-pstate.c 
> b/drivers/cpufreq/amd-pstate.c index 9ad62dbe8bfb..a092b13ffbc2 100644
> --- a/drivers/cpufreq/amd-pstate.c
> +++ b/drivers/cpufreq/amd-pstate.c
> @@ -247,6 +247,26 @@ static int amd_pstate_get_energy_pref_index(struct amd_cpudata *cpudata)
>   	return index;
>   }
>   
> +static void pstate_update_perf(struct amd_cpudata *cpudata, u32 min_perf,
> +			       u32 des_perf, u32 max_perf, bool fast_switch) {
> +	if (fast_switch)
> +		wrmsrl(MSR_AMD_CPPC_REQ, READ_ONCE(cpudata->cppc_req_cached));
> +	else
> +		wrmsrl_on_cpu(cpudata->cpu, MSR_AMD_CPPC_REQ,
> +			      READ_ONCE(cpudata->cppc_req_cached));
> +}
> +
> +DEFINE_STATIC_CALL(amd_pstate_update_perf, pstate_update_perf);
> +
> +static inline void amd_pstate_update_perf(struct amd_cpudata *cpudata,
> +					  u32 min_perf, u32 des_perf,
> +					  u32 max_perf, bool fast_switch) {
> +	static_call(amd_pstate_update_perf)(cpudata, min_perf, des_perf,
> +					    max_perf, fast_switch);
> +}
> +
>   static int amd_pstate_set_epp(struct amd_cpudata *cpudata, u32 epp)
>   {
>   	int ret;
> @@ -263,6 +283,9 @@ static int amd_pstate_set_epp(struct amd_cpudata *cpudata, u32 epp)
>   		if (!ret)
>   			cpudata->epp_cached = epp;
>   	} else {
> +		amd_pstate_update_perf(cpudata, cpudata->min_limit_perf, 0U,
> +					     cpudata->max_limit_perf, false);
> +
>   		perf_ctrls.energy_perf = epp;
>   		ret = cppc_set_epp_perf(cpudata->cpu, &perf_ctrls, 1);
>   		if (ret) {
> @@ -452,16 +475,6 @@ static inline int amd_pstate_init_perf(struct amd_cpudata *cpudata)
>   	return static_call(amd_pstate_init_perf)(cpudata);
>   }
>   
> -static void pstate_update_perf(struct amd_cpudata *cpudata, u32 min_perf,
> -			       u32 des_perf, u32 max_perf, bool fast_switch)
> -{
> -	if (fast_switch)
> -		wrmsrl(MSR_AMD_CPPC_REQ, READ_ONCE(cpudata->cppc_req_cached));
> -	else
> -		wrmsrl_on_cpu(cpudata->cpu, MSR_AMD_CPPC_REQ,
> -			      READ_ONCE(cpudata->cppc_req_cached));
> -}
> -
>   static void cppc_update_perf(struct amd_cpudata *cpudata,
>   			     u32 min_perf, u32 des_perf,
>   			     u32 max_perf, bool fast_switch) @@ -475,16 +488,6 @@ static 
> void cppc_update_perf(struct amd_cpudata *cpudata,
>   	cppc_set_perf(cpudata->cpu, &perf_ctrls);
>   }
>   
> -DEFINE_STATIC_CALL(amd_pstate_update_perf, pstate_update_perf);
> -
> -static inline void amd_pstate_update_perf(struct amd_cpudata *cpudata,
> -					  u32 min_perf, u32 des_perf,
> -					  u32 max_perf, bool fast_switch)
> -{
> -	static_call(amd_pstate_update_perf)(cpudata, min_perf, des_perf,
> -					    max_perf, fast_switch);
> -}
> -
>   static inline bool amd_pstate_sample(struct amd_cpudata *cpudata)
>   {
>   	u64 aperf, mperf, tsc;
Mario Limonciello Sept. 3, 2024, 5:54 p.m. UTC | #2
Hi Morgan,

Can you please cross reference 6.11-rc6 to see if you're still having a 
problem?  I would like to understand if we're missing a backport to 
stable or this is still an upstream problem.

Thanks,

On 9/3/2024 12:51, Jones, Morgan wrote:
> Hi there,
> 
> We are experiencing a ~35% performance regression on an AMD EPYC 7702 64 core machine after applying this patch. We observed Linux 6.6.44 starting to cause the issue, and performed a bisect involving rebooting the machine over 15 times. (Note that kexec didn't successfully identify the problem, since the PM memory mailbox is never reset).
> 
> It appears that we are getting a max of 2.18 GHz when this CPU can boost to 3.35 GHz, explaining the slowdown. Blacklisting amd-pstate solves the issue at the expense of the performance increase we used to get by using it.
> 
> Is it possible that the upper limits were implicitly at the max CPU power before, and setting the upper limit to something other than the boost frequency can reduce performance now?
> 
> # bad: [7213910600667c51c978e577bf5454d3f7b313b7] Linux 6.6.44
> # good: [58b0425ff5df680d0b67f64ae1f3f1ebdf1c4de9] Linux 6.6.43
> git bisect start '7213910600667c51c978e577bf5454d3f7b313b7' '58b0425ff5df680d0b67f64ae1f3f1ebdf1c4de9'
> # good: [72ff9d26964a3a80f7650df719df139f5c1f965d] arm64: dts: qcom: sm6350: Add missing qcom,non-secure-domain property
> git bisect good 72ff9d26964a3a80f7650df719df139f5c1f965d
> # good: [0fffc2e1bf40a2220ef5a38f834ea063dba832d3] ARM: dts: sunxi: remove duplicated entries in makefile
> git bisect good 0fffc2e1bf40a2220ef5a38f834ea063dba832d3
> # bad: [8cdbe6ebfd1763a5c41a2a3058497c0a9163311c] pinctrl: renesas: r8a779g0: Fix CANFD5 suffix
> git bisect bad 8cdbe6ebfd1763a5c41a2a3058497c0a9163311c
> # bad: [5dbb98e7fa42bebc1325899193d8f13f0705a148] drm/mediatek: Turn off the layers with zero width or height
> git bisect bad 5dbb98e7fa42bebc1325899193d8f13f0705a148
> # bad: [691ec7043122c9c8c46d84f6e6cd85d13d50cd93] selftests/bpf: Null checks for links in bpf_tcp_ca
> git bisect bad 691ec7043122c9c8c46d84f6e6cd85d13d50cd93
> # bad: [a1359e085d75d7393a250054e66c0a7bc6c3dbfa] perf/x86: Serialize set_attr_rdpmc()
> git bisect bad a1359e085d75d7393a250054e66c0a7bc6c3dbfa
> # bad: [e99d9b16ff153de9540073239d24adc3b0a3a997] wifi: ath12k: change DMA direction while mapping reinjected packets
> git bisect bad e99d9b16ff153de9540073239d24adc3b0a3a997
> # bad: [d027ac4a08541beb2a89563d3e034da7085050ba] firmware: turris-mox-rwtm: Initialize completion before mailbox
> git bisect bad d027ac4a08541beb2a89563d3e034da7085050ba
> # bad: [e6c9eca327e6a41a81e7eba0d0ddc13da37f82a1] ARM: spitz: fix GPIO assignment for backlight
> git bisect bad e6c9eca327e6a41a81e7eba0d0ddc13da37f82a1
> # bad: [b8cdefdaa555bbfc269c2198803f8791a8923960] m68k: cmpxchg: Fix return value for default case in __arch_xchg()
> git bisect bad b8cdefdaa555bbfc269c2198803f8791a8923960
> # bad: [13a71384ae6a8779da809b00c6f378dcead10427] cpufreq/amd-pstate: Fix the scaling_max_freq setting on shared memory CPPC systems
> git bisect bad 13a71384ae6a8779da809b00c6f378dcead10427
> # first bad commit: [13a71384ae6a8779da809b00c6f378dcead10427] cpufreq/amd-pstate: Fix the scaling_max_freq setting on shared memory CPPC systems
> 
> cpupower output:
> 
> analyzing CPU 47:
>    driver: amd-pstate-epp
>    CPUs which run at the same hardware frequency: 47
>    CPUs which need to have their frequency coordinated by software: 47
>    maximum transition latency:  Cannot determine or is not supported.
>    hardware limits: 400 MHz - 2.18 GHz
>    available cpufreq governors: performance powersave
>    current policy: frequency should be within 400 MHz and 2.18 GHz.
>                    The governor "performance" may decide which speed to use
>                    within this range.
>    current CPU frequency: Unable to call hardware
>    current CPU frequency: 2.17 GHz (asserted by call to kernel)
>    boost state support:
>      Supported: yes
>      Active: yes
>      AMD PSTATE Highest Performance: 166. Maximum Frequency: 2.18 GHz.
>      AMD PSTATE Nominal Performance: 152. Nominal Frequency: 2.00 GHz.
>      AMD PSTATE Lowest Non-linear Performance: 115. Lowest Non-linear Frequency: 1.51 GHz.
>      AMD PSTATE Lowest Performance: 31. Lowest Frequency: 400 MHz.
> 
> Thanks,
> Morgan
> 
> -----Original Message-----
> From: Mario Limonciello <mario.limonciello@amd.com>
> Sent: Tuesday, July 2, 2024 10:49 AM
> To: Dhananjay Ugwekar <Dhananjay.Ugwekar@amd.com>; rafael@kernel.org; viresh.kumar@linaro.org; gautham.shenoy@amd.com; perry.yuan@amd.com; skhan@linuxfoundation.org; li.meng@amd.com; ray.huang@amd.com
> Cc: linux-pm@vger.kernel.org; linux-kernel@vger.kernel.org; David Arcari <darcari@redhat.com>
> Subject: Re: [PATCH v2 2/2] cpufreq/amd-pstate: Fix the scaling_max_freq setting on shared memory CPPC systems
> 
> On 7/2/2024 3:14, Dhananjay Ugwekar wrote:
>> On shared memory CPPC systems, with amd_pstate=active mode, the change
>> in scaling_max_freq doesn't get written to the shared memory region.
>> Due to this, the writes to the scaling_max_freq sysfs file don't take
>> effect. Fix this by propagating the scaling_max_freq changes to the
>> shared memory region.
>>
>> Fixes: ffa5096a7c33 ("cpufreq: amd-pstate: implement Pstate EPP
>> support for the AMD processors")
>> Reported-by: David Arcari <darcari@redhat.com>
>> Signed-off-by: Dhananjay Ugwekar <Dhananjay.Ugwekar@amd.com>
> Reviewed-by: Mario Limonciello <mario.limonciello@amd.com>
>> ---
>>    drivers/cpufreq/amd-pstate.c | 43 +++++++++++++++++++-----------------
>>    1 file changed, 23 insertions(+), 20 deletions(-)
>>
>> diff --git a/drivers/cpufreq/amd-pstate.c
>> b/drivers/cpufreq/amd-pstate.c index 9ad62dbe8bfb..a092b13ffbc2 100644
>> --- a/drivers/cpufreq/amd-pstate.c
>> +++ b/drivers/cpufreq/amd-pstate.c
>> @@ -247,6 +247,26 @@ static int amd_pstate_get_energy_pref_index(struct amd_cpudata *cpudata)
>>    	return index;
>>    }
>>    
>> +static void pstate_update_perf(struct amd_cpudata *cpudata, u32 min_perf,
>> +			       u32 des_perf, u32 max_perf, bool fast_switch) {
>> +	if (fast_switch)
>> +		wrmsrl(MSR_AMD_CPPC_REQ, READ_ONCE(cpudata->cppc_req_cached));
>> +	else
>> +		wrmsrl_on_cpu(cpudata->cpu, MSR_AMD_CPPC_REQ,
>> +			      READ_ONCE(cpudata->cppc_req_cached));
>> +}
>> +
>> +DEFINE_STATIC_CALL(amd_pstate_update_perf, pstate_update_perf);
>> +
>> +static inline void amd_pstate_update_perf(struct amd_cpudata *cpudata,
>> +					  u32 min_perf, u32 des_perf,
>> +					  u32 max_perf, bool fast_switch) {
>> +	static_call(amd_pstate_update_perf)(cpudata, min_perf, des_perf,
>> +					    max_perf, fast_switch);
>> +}
>> +
>>    static int amd_pstate_set_epp(struct amd_cpudata *cpudata, u32 epp)
>>    {
>>    	int ret;
>> @@ -263,6 +283,9 @@ static int amd_pstate_set_epp(struct amd_cpudata *cpudata, u32 epp)
>>    		if (!ret)
>>    			cpudata->epp_cached = epp;
>>    	} else {
>> +		amd_pstate_update_perf(cpudata, cpudata->min_limit_perf, 0U,
>> +					     cpudata->max_limit_perf, false);
>> +
>>    		perf_ctrls.energy_perf = epp;
>>    		ret = cppc_set_epp_perf(cpudata->cpu, &perf_ctrls, 1);
>>    		if (ret) {
>> @@ -452,16 +475,6 @@ static inline int amd_pstate_init_perf(struct amd_cpudata *cpudata)
>>    	return static_call(amd_pstate_init_perf)(cpudata);
>>    }
>>    
>> -static void pstate_update_perf(struct amd_cpudata *cpudata, u32 min_perf,
>> -			       u32 des_perf, u32 max_perf, bool fast_switch)
>> -{
>> -	if (fast_switch)
>> -		wrmsrl(MSR_AMD_CPPC_REQ, READ_ONCE(cpudata->cppc_req_cached));
>> -	else
>> -		wrmsrl_on_cpu(cpudata->cpu, MSR_AMD_CPPC_REQ,
>> -			      READ_ONCE(cpudata->cppc_req_cached));
>> -}
>> -
>>    static void cppc_update_perf(struct amd_cpudata *cpudata,
>>    			     u32 min_perf, u32 des_perf,
>>    			     u32 max_perf, bool fast_switch) @@ -475,16 +488,6 @@ static
>> void cppc_update_perf(struct amd_cpudata *cpudata,
>>    	cppc_set_perf(cpudata->cpu, &perf_ctrls);
>>    }
>>    
>> -DEFINE_STATIC_CALL(amd_pstate_update_perf, pstate_update_perf);
>> -
>> -static inline void amd_pstate_update_perf(struct amd_cpudata *cpudata,
>> -					  u32 min_perf, u32 des_perf,
>> -					  u32 max_perf, bool fast_switch)
>> -{
>> -	static_call(amd_pstate_update_perf)(cpudata, min_perf, des_perf,
>> -					    max_perf, fast_switch);
>> -}
>> -
>>    static inline bool amd_pstate_sample(struct amd_cpudata *cpudata)
>>    {
>>    	u64 aperf, mperf, tsc;
> 
>
Jones, Morgan Sept. 3, 2024, 8:07 p.m. UTC | #3
Hey Mario,

Smoking gun here, the max frequency is incorrect on 6.6.44+ but is correct on 6.11.0-rc6.

Linux redact 6.11.0-rc6 #1-NixOS SMP PREEMPT_DYNAMIC Tue Jan  1 00:00:00 UTC 1980 x86_64 GNU/Linux

analyzing CPU 12:
  driver: amd-pstate-epp
  CPUs which run at the same hardware frequency: 12
  CPUs which need to have their frequency coordinated by software: 12
  maximum transition latency:  Cannot determine or is not supported.
  hardware limits: 400 MHz - 3.35 GHz
  available cpufreq governors: performance powersave
  current policy: frequency should be within 400 MHz and 3.35 GHz.
                  The governor "performance" may decide which speed to use
                  within this range.
  current CPU frequency: Unable to call hardware
  current CPU frequency: 3.34 GHz (asserted by call to kernel)
  boost state support:
    Supported: yes
    Active: yes
    AMD PSTATE Highest Performance: 255. Maximum Frequency: 3.35 GHz.
    AMD PSTATE Nominal Performance: 152. Nominal Frequency: 2.00 GHz.
    AMD PSTATE Lowest Non-linear Performance: 115. Lowest Non-linear Frequency: 1.51 GHz.
    AMD PSTATE Lowest Performance: 31. Lowest Frequency: 400 MHz.

We're running amd_pstate=active and amd_pstate.shared_mem=1, and our workloads are back to normal performance on 6.11.0-rc6.

Morgan

-----Original Message-----
From: Mario Limonciello <mario.limonciello@amd.com> 
Sent: Tuesday, September 3, 2024 10:55 AM
To: Jones, Morgan <Morgan.Jones@viasat.com>; Dhananjay Ugwekar <Dhananjay.Ugwekar@amd.com>; rafael@kernel.org; viresh.kumar@linaro.org; gautham.shenoy@amd.com; perry.yuan@amd.com; skhan@linuxfoundation.org; li.meng@amd.com; ray.huang@amd.com
Cc: linux-pm@vger.kernel.org; linux-kernel@vger.kernel.org; David Arcari <darcari@redhat.com>
Subject: [EXTERNAL] Re: [PATCH v2 2/2] cpufreq/amd-pstate: Fix the scaling_max_freq setting on shared memory CPPC systems

Hi Morgan,

Can you please cross reference 6.11-rc6 to see if you're still having a problem?  I would like to understand if we're missing a backport to stable or this is still an upstream problem.

Thanks,

On 9/3/2024 12:51, Jones, Morgan wrote:
> Hi there,
> 
> We are experiencing a ~35% performance regression on an AMD EPYC 7702 64 core machine after applying this patch. We observed Linux 6.6.44 starting to cause the issue, and performed a bisect involving rebooting the machine over 15 times. (Note that kexec didn't successfully identify the problem, since the PM memory mailbox is never reset).
> 
> It appears that we are getting a max of 2.18 GHz when this CPU can boost to 3.35 GHz, explaining the slowdown. Blacklisting amd-pstate solves the issue at the expense of the performance increase we used to get by using it.
> 
> Is it possible that the upper limits were implicitly at the max CPU power before, and setting the upper limit to something other than the boost frequency can reduce performance now?
> 
> # bad: [7213910600667c51c978e577bf5454d3f7b313b7] Linux 6.6.44 # good: 
> [58b0425ff5df680d0b67f64ae1f3f1ebdf1c4de9] Linux 6.6.43 git bisect 
> start '7213910600667c51c978e577bf5454d3f7b313b7' '58b0425ff5df680d0b67f64ae1f3f1ebdf1c4de9'
> # good: [72ff9d26964a3a80f7650df719df139f5c1f965d] arm64: dts: qcom: 
> sm6350: Add missing qcom,non-secure-domain property git bisect good 
> 72ff9d26964a3a80f7650df719df139f5c1f965d
> # good: [0fffc2e1bf40a2220ef5a38f834ea063dba832d3] ARM: dts: sunxi: 
> remove duplicated entries in makefile git bisect good 
> 0fffc2e1bf40a2220ef5a38f834ea063dba832d3
> # bad: [8cdbe6ebfd1763a5c41a2a3058497c0a9163311c] pinctrl: renesas: 
> r8a779g0: Fix CANFD5 suffix git bisect bad 
> 8cdbe6ebfd1763a5c41a2a3058497c0a9163311c
> # bad: [5dbb98e7fa42bebc1325899193d8f13f0705a148] drm/mediatek: Turn 
> off the layers with zero width or height git bisect bad 
> 5dbb98e7fa42bebc1325899193d8f13f0705a148
> # bad: [691ec7043122c9c8c46d84f6e6cd85d13d50cd93] selftests/bpf: Null 
> checks for links in bpf_tcp_ca git bisect bad 
> 691ec7043122c9c8c46d84f6e6cd85d13d50cd93
> # bad: [a1359e085d75d7393a250054e66c0a7bc6c3dbfa] perf/x86: Serialize 
> set_attr_rdpmc() git bisect bad 
> a1359e085d75d7393a250054e66c0a7bc6c3dbfa
> # bad: [e99d9b16ff153de9540073239d24adc3b0a3a997] wifi: ath12k: change 
> DMA direction while mapping reinjected packets git bisect bad 
> e99d9b16ff153de9540073239d24adc3b0a3a997
> # bad: [d027ac4a08541beb2a89563d3e034da7085050ba] firmware: 
> turris-mox-rwtm: Initialize completion before mailbox git bisect bad 
> d027ac4a08541beb2a89563d3e034da7085050ba
> # bad: [e6c9eca327e6a41a81e7eba0d0ddc13da37f82a1] ARM: spitz: fix GPIO 
> assignment for backlight git bisect bad 
> e6c9eca327e6a41a81e7eba0d0ddc13da37f82a1
> # bad: [b8cdefdaa555bbfc269c2198803f8791a8923960] m68k: cmpxchg: Fix 
> return value for default case in __arch_xchg() git bisect bad 
> b8cdefdaa555bbfc269c2198803f8791a8923960
> # bad: [13a71384ae6a8779da809b00c6f378dcead10427] cpufreq/amd-pstate: 
> Fix the scaling_max_freq setting on shared memory CPPC systems git 
> bisect bad 13a71384ae6a8779da809b00c6f378dcead10427
> # first bad commit: [13a71384ae6a8779da809b00c6f378dcead10427] 
> cpufreq/amd-pstate: Fix the scaling_max_freq setting on shared memory 
> CPPC systems
> 
> cpupower output:
> 
> analyzing CPU 47:
>    driver: amd-pstate-epp
>    CPUs which run at the same hardware frequency: 47
>    CPUs which need to have their frequency coordinated by software: 47
>    maximum transition latency:  Cannot determine or is not supported.
>    hardware limits: 400 MHz - 2.18 GHz
>    available cpufreq governors: performance powersave
>    current policy: frequency should be within 400 MHz and 2.18 GHz.
>                    The governor "performance" may decide which speed to use
>                    within this range.
>    current CPU frequency: Unable to call hardware
>    current CPU frequency: 2.17 GHz (asserted by call to kernel)
>    boost state support:
>      Supported: yes
>      Active: yes
>      AMD PSTATE Highest Performance: 166. Maximum Frequency: 2.18 GHz.
>      AMD PSTATE Nominal Performance: 152. Nominal Frequency: 2.00 GHz.
>      AMD PSTATE Lowest Non-linear Performance: 115. Lowest Non-linear Frequency: 1.51 GHz.
>      AMD PSTATE Lowest Performance: 31. Lowest Frequency: 400 MHz.
> 
> Thanks,
> Morgan
> 
> -----Original Message-----
> From: Mario Limonciello <mario.limonciello@amd.com>
> Sent: Tuesday, July 2, 2024 10:49 AM
> To: Dhananjay Ugwekar <Dhananjay.Ugwekar@amd.com>; rafael@kernel.org; 
> viresh.kumar@linaro.org; gautham.shenoy@amd.com; perry.yuan@amd.com; 
> skhan@linuxfoundation.org; li.meng@amd.com; ray.huang@amd.com
> Cc: linux-pm@vger.kernel.org; linux-kernel@vger.kernel.org; David 
> Arcari <darcari@redhat.com>
> Subject: Re: [PATCH v2 2/2] cpufreq/amd-pstate: Fix the 
> scaling_max_freq setting on shared memory CPPC systems
> 
> On 7/2/2024 3:14, Dhananjay Ugwekar wrote:
>> On shared memory CPPC systems, with amd_pstate=active mode, the 
>> change in scaling_max_freq doesn't get written to the shared memory region.
>> Due to this, the writes to the scaling_max_freq sysfs file don't take 
>> effect. Fix this by propagating the scaling_max_freq changes to the 
>> shared memory region.
>>
>> Fixes: ffa5096a7c33 ("cpufreq: amd-pstate: implement Pstate EPP 
>> support for the AMD processors")
>> Reported-by: David Arcari <darcari@redhat.com>
>> Signed-off-by: Dhananjay Ugwekar <Dhananjay.Ugwekar@amd.com>
> Reviewed-by: Mario Limonciello <mario.limonciello@amd.com>
>> ---
>>    drivers/cpufreq/amd-pstate.c | 43 +++++++++++++++++++-----------------
>>    1 file changed, 23 insertions(+), 20 deletions(-)
>>
>> diff --git a/drivers/cpufreq/amd-pstate.c 
>> b/drivers/cpufreq/amd-pstate.c index 9ad62dbe8bfb..a092b13ffbc2 
>> 100644
>> --- a/drivers/cpufreq/amd-pstate.c
>> +++ b/drivers/cpufreq/amd-pstate.c
>> @@ -247,6 +247,26 @@ static int amd_pstate_get_energy_pref_index(struct amd_cpudata *cpudata)
>>    	return index;
>>    }
>>    
>> +static void pstate_update_perf(struct amd_cpudata *cpudata, u32 min_perf,
>> +			       u32 des_perf, u32 max_perf, bool fast_switch) {
>> +	if (fast_switch)
>> +		wrmsrl(MSR_AMD_CPPC_REQ, READ_ONCE(cpudata->cppc_req_cached));
>> +	else
>> +		wrmsrl_on_cpu(cpudata->cpu, MSR_AMD_CPPC_REQ,
>> +			      READ_ONCE(cpudata->cppc_req_cached));
>> +}
>> +
>> +DEFINE_STATIC_CALL(amd_pstate_update_perf, pstate_update_perf);
>> +
>> +static inline void amd_pstate_update_perf(struct amd_cpudata *cpudata,
>> +					  u32 min_perf, u32 des_perf,
>> +					  u32 max_perf, bool fast_switch) {
>> +	static_call(amd_pstate_update_perf)(cpudata, min_perf, des_perf,
>> +					    max_perf, fast_switch);
>> +}
>> +
>>    static int amd_pstate_set_epp(struct amd_cpudata *cpudata, u32 epp)
>>    {
>>    	int ret;
>> @@ -263,6 +283,9 @@ static int amd_pstate_set_epp(struct amd_cpudata *cpudata, u32 epp)
>>    		if (!ret)
>>    			cpudata->epp_cached = epp;
>>    	} else {
>> +		amd_pstate_update_perf(cpudata, cpudata->min_limit_perf, 0U,
>> +					     cpudata->max_limit_perf, false);
>> +
>>    		perf_ctrls.energy_perf = epp;
>>    		ret = cppc_set_epp_perf(cpudata->cpu, &perf_ctrls, 1);
>>    		if (ret) {
>> @@ -452,16 +475,6 @@ static inline int amd_pstate_init_perf(struct amd_cpudata *cpudata)
>>    	return static_call(amd_pstate_init_perf)(cpudata);
>>    }
>>    
>> -static void pstate_update_perf(struct amd_cpudata *cpudata, u32 min_perf,
>> -			       u32 des_perf, u32 max_perf, bool fast_switch)
>> -{
>> -	if (fast_switch)
>> -		wrmsrl(MSR_AMD_CPPC_REQ, READ_ONCE(cpudata->cppc_req_cached));
>> -	else
>> -		wrmsrl_on_cpu(cpudata->cpu, MSR_AMD_CPPC_REQ,
>> -			      READ_ONCE(cpudata->cppc_req_cached));
>> -}
>> -
>>    static void cppc_update_perf(struct amd_cpudata *cpudata,
>>    			     u32 min_perf, u32 des_perf,
>>    			     u32 max_perf, bool fast_switch) @@ -475,16 +488,6 @@ 
>> static void cppc_update_perf(struct amd_cpudata *cpudata,
>>    	cppc_set_perf(cpudata->cpu, &perf_ctrls);
>>    }
>>    
>> -DEFINE_STATIC_CALL(amd_pstate_update_perf, pstate_update_perf);
>> -
>> -static inline void amd_pstate_update_perf(struct amd_cpudata *cpudata,
>> -					  u32 min_perf, u32 des_perf,
>> -					  u32 max_perf, bool fast_switch)
>> -{
>> -	static_call(amd_pstate_update_perf)(cpudata, min_perf, des_perf,
>> -					    max_perf, fast_switch);
>> -}
>> -
>>    static inline bool amd_pstate_sample(struct amd_cpudata *cpudata)
>>    {
>>    	u64 aperf, mperf, tsc;
> 
>
Mario Limonciello Sept. 3, 2024, 8:09 p.m. UTC | #4
Morgan,

OK that's great news that it's just a backport effort.  That same commit 
also backported to 6.10.3.  Can you see if 6.10.y is affected?

Ugwekar,

Any thoughts on what else needs to come back to 6.6.y off hand?

Thanks,

On 9/3/2024 15:07, Jones, Morgan wrote:
> Hey Mario,
> 
> Smoking gun here, the max frequency is incorrect on 6.6.44+ but is correct on 6.11.0-rc6.
> 
> Linux redact 6.11.0-rc6 #1-NixOS SMP PREEMPT_DYNAMIC Tue Jan  1 00:00:00 UTC 1980 x86_64 GNU/Linux
> 
> analyzing CPU 12:
>    driver: amd-pstate-epp
>    CPUs which run at the same hardware frequency: 12
>    CPUs which need to have their frequency coordinated by software: 12
>    maximum transition latency:  Cannot determine or is not supported.
>    hardware limits: 400 MHz - 3.35 GHz
>    available cpufreq governors: performance powersave
>    current policy: frequency should be within 400 MHz and 3.35 GHz.
>                    The governor "performance" may decide which speed to use
>                    within this range.
>    current CPU frequency: Unable to call hardware
>    current CPU frequency: 3.34 GHz (asserted by call to kernel)
>    boost state support:
>      Supported: yes
>      Active: yes
>      AMD PSTATE Highest Performance: 255. Maximum Frequency: 3.35 GHz.
>      AMD PSTATE Nominal Performance: 152. Nominal Frequency: 2.00 GHz.
>      AMD PSTATE Lowest Non-linear Performance: 115. Lowest Non-linear Frequency: 1.51 GHz.
>      AMD PSTATE Lowest Performance: 31. Lowest Frequency: 400 MHz.
> 
> We're running amd_pstate=active and amd_pstate.shared_mem=1, and our workloads are back to normal performance on 6.11.0-rc6.
> 
> Morgan
> 
> -----Original Message-----
> From: Mario Limonciello <mario.limonciello@amd.com>
> Sent: Tuesday, September 3, 2024 10:55 AM
> To: Jones, Morgan <Morgan.Jones@viasat.com>; Dhananjay Ugwekar <Dhananjay.Ugwekar@amd.com>; rafael@kernel.org; viresh.kumar@linaro.org; gautham.shenoy@amd.com; perry.yuan@amd.com; skhan@linuxfoundation.org; li.meng@amd.com; ray.huang@amd.com
> Cc: linux-pm@vger.kernel.org; linux-kernel@vger.kernel.org; David Arcari <darcari@redhat.com>
> Subject: [EXTERNAL] Re: [PATCH v2 2/2] cpufreq/amd-pstate: Fix the scaling_max_freq setting on shared memory CPPC systems
> 
> Hi Morgan,
> 
> Can you please cross reference 6.11-rc6 to see if you're still having a problem?  I would like to understand if we're missing a backport to stable or this is still an upstream problem.
> 
> Thanks,
> 
> On 9/3/2024 12:51, Jones, Morgan wrote:
>> Hi there,
>>
>> We are experiencing a ~35% performance regression on an AMD EPYC 7702 64 core machine after applying this patch. We observed Linux 6.6.44 starting to cause the issue, and performed a bisect involving rebooting the machine over 15 times. (Note that kexec didn't successfully identify the problem, since the PM memory mailbox is never reset).
>>
>> It appears that we are getting a max of 2.18 GHz when this CPU can boost to 3.35 GHz, explaining the slowdown. Blacklisting amd-pstate solves the issue at the expense of the performance increase we used to get by using it.
>>
>> Is it possible that the upper limits were implicitly at the max CPU power before, and setting the upper limit to something other than the boost frequency can reduce performance now?
>>
>> # bad: [7213910600667c51c978e577bf5454d3f7b313b7] Linux 6.6.44 # good:
>> [58b0425ff5df680d0b67f64ae1f3f1ebdf1c4de9] Linux 6.6.43 git bisect
>> start '7213910600667c51c978e577bf5454d3f7b313b7' '58b0425ff5df680d0b67f64ae1f3f1ebdf1c4de9'
>> # good: [72ff9d26964a3a80f7650df719df139f5c1f965d] arm64: dts: qcom:
>> sm6350: Add missing qcom,non-secure-domain property git bisect good
>> 72ff9d26964a3a80f7650df719df139f5c1f965d
>> # good: [0fffc2e1bf40a2220ef5a38f834ea063dba832d3] ARM: dts: sunxi:
>> remove duplicated entries in makefile git bisect good
>> 0fffc2e1bf40a2220ef5a38f834ea063dba832d3
>> # bad: [8cdbe6ebfd1763a5c41a2a3058497c0a9163311c] pinctrl: renesas:
>> r8a779g0: Fix CANFD5 suffix git bisect bad
>> 8cdbe6ebfd1763a5c41a2a3058497c0a9163311c
>> # bad: [5dbb98e7fa42bebc1325899193d8f13f0705a148] drm/mediatek: Turn
>> off the layers with zero width or height git bisect bad
>> 5dbb98e7fa42bebc1325899193d8f13f0705a148
>> # bad: [691ec7043122c9c8c46d84f6e6cd85d13d50cd93] selftests/bpf: Null
>> checks for links in bpf_tcp_ca git bisect bad
>> 691ec7043122c9c8c46d84f6e6cd85d13d50cd93
>> # bad: [a1359e085d75d7393a250054e66c0a7bc6c3dbfa] perf/x86: Serialize
>> set_attr_rdpmc() git bisect bad
>> a1359e085d75d7393a250054e66c0a7bc6c3dbfa
>> # bad: [e99d9b16ff153de9540073239d24adc3b0a3a997] wifi: ath12k: change
>> DMA direction while mapping reinjected packets git bisect bad
>> e99d9b16ff153de9540073239d24adc3b0a3a997
>> # bad: [d027ac4a08541beb2a89563d3e034da7085050ba] firmware:
>> turris-mox-rwtm: Initialize completion before mailbox git bisect bad
>> d027ac4a08541beb2a89563d3e034da7085050ba
>> # bad: [e6c9eca327e6a41a81e7eba0d0ddc13da37f82a1] ARM: spitz: fix GPIO
>> assignment for backlight git bisect bad
>> e6c9eca327e6a41a81e7eba0d0ddc13da37f82a1
>> # bad: [b8cdefdaa555bbfc269c2198803f8791a8923960] m68k: cmpxchg: Fix
>> return value for default case in __arch_xchg() git bisect bad
>> b8cdefdaa555bbfc269c2198803f8791a8923960
>> # bad: [13a71384ae6a8779da809b00c6f378dcead10427] cpufreq/amd-pstate:
>> Fix the scaling_max_freq setting on shared memory CPPC systems git
>> bisect bad 13a71384ae6a8779da809b00c6f378dcead10427
>> # first bad commit: [13a71384ae6a8779da809b00c6f378dcead10427]
>> cpufreq/amd-pstate: Fix the scaling_max_freq setting on shared memory
>> CPPC systems
>>
>> cpupower output:
>>
>> analyzing CPU 47:
>>     driver: amd-pstate-epp
>>     CPUs which run at the same hardware frequency: 47
>>     CPUs which need to have their frequency coordinated by software: 47
>>     maximum transition latency:  Cannot determine or is not supported.
>>     hardware limits: 400 MHz - 2.18 GHz
>>     available cpufreq governors: performance powersave
>>     current policy: frequency should be within 400 MHz and 2.18 GHz.
>>                     The governor "performance" may decide which speed to use
>>                     within this range.
>>     current CPU frequency: Unable to call hardware
>>     current CPU frequency: 2.17 GHz (asserted by call to kernel)
>>     boost state support:
>>       Supported: yes
>>       Active: yes
>>       AMD PSTATE Highest Performance: 166. Maximum Frequency: 2.18 GHz.
>>       AMD PSTATE Nominal Performance: 152. Nominal Frequency: 2.00 GHz.
>>       AMD PSTATE Lowest Non-linear Performance: 115. Lowest Non-linear Frequency: 1.51 GHz.
>>       AMD PSTATE Lowest Performance: 31. Lowest Frequency: 400 MHz.
>>
>> Thanks,
>> Morgan
>>
>> -----Original Message-----
>> From: Mario Limonciello <mario.limonciello@amd.com>
>> Sent: Tuesday, July 2, 2024 10:49 AM
>> To: Dhananjay Ugwekar <Dhananjay.Ugwekar@amd.com>; rafael@kernel.org;
>> viresh.kumar@linaro.org; gautham.shenoy@amd.com; perry.yuan@amd.com;
>> skhan@linuxfoundation.org; li.meng@amd.com; ray.huang@amd.com
>> Cc: linux-pm@vger.kernel.org; linux-kernel@vger.kernel.org; David
>> Arcari <darcari@redhat.com>
>> Subject: Re: [PATCH v2 2/2] cpufreq/amd-pstate: Fix the
>> scaling_max_freq setting on shared memory CPPC systems
>>
>> On 7/2/2024 3:14, Dhananjay Ugwekar wrote:
>>> On shared memory CPPC systems, with amd_pstate=active mode, the
>>> change in scaling_max_freq doesn't get written to the shared memory region.
>>> Due to this, the writes to the scaling_max_freq sysfs file don't take
>>> effect. Fix this by propagating the scaling_max_freq changes to the
>>> shared memory region.
>>>
>>> Fixes: ffa5096a7c33 ("cpufreq: amd-pstate: implement Pstate EPP
>>> support for the AMD processors")
>>> Reported-by: David Arcari <darcari@redhat.com>
>>> Signed-off-by: Dhananjay Ugwekar <Dhananjay.Ugwekar@amd.com>
>> Reviewed-by: Mario Limonciello <mario.limonciello@amd.com>
>>> ---
>>>     drivers/cpufreq/amd-pstate.c | 43 +++++++++++++++++++-----------------
>>>     1 file changed, 23 insertions(+), 20 deletions(-)
>>>
>>> diff --git a/drivers/cpufreq/amd-pstate.c
>>> b/drivers/cpufreq/amd-pstate.c index 9ad62dbe8bfb..a092b13ffbc2
>>> 100644
>>> --- a/drivers/cpufreq/amd-pstate.c
>>> +++ b/drivers/cpufreq/amd-pstate.c
>>> @@ -247,6 +247,26 @@ static int amd_pstate_get_energy_pref_index(struct amd_cpudata *cpudata)
>>>     	return index;
>>>     }
>>>     
>>> +static void pstate_update_perf(struct amd_cpudata *cpudata, u32 min_perf,
>>> +			       u32 des_perf, u32 max_perf, bool fast_switch) {
>>> +	if (fast_switch)
>>> +		wrmsrl(MSR_AMD_CPPC_REQ, READ_ONCE(cpudata->cppc_req_cached));
>>> +	else
>>> +		wrmsrl_on_cpu(cpudata->cpu, MSR_AMD_CPPC_REQ,
>>> +			      READ_ONCE(cpudata->cppc_req_cached));
>>> +}
>>> +
>>> +DEFINE_STATIC_CALL(amd_pstate_update_perf, pstate_update_perf);
>>> +
>>> +static inline void amd_pstate_update_perf(struct amd_cpudata *cpudata,
>>> +					  u32 min_perf, u32 des_perf,
>>> +					  u32 max_perf, bool fast_switch) {
>>> +	static_call(amd_pstate_update_perf)(cpudata, min_perf, des_perf,
>>> +					    max_perf, fast_switch);
>>> +}
>>> +
>>>     static int amd_pstate_set_epp(struct amd_cpudata *cpudata, u32 epp)
>>>     {
>>>     	int ret;
>>> @@ -263,6 +283,9 @@ static int amd_pstate_set_epp(struct amd_cpudata *cpudata, u32 epp)
>>>     		if (!ret)
>>>     			cpudata->epp_cached = epp;
>>>     	} else {
>>> +		amd_pstate_update_perf(cpudata, cpudata->min_limit_perf, 0U,
>>> +					     cpudata->max_limit_perf, false);
>>> +
>>>     		perf_ctrls.energy_perf = epp;
>>>     		ret = cppc_set_epp_perf(cpudata->cpu, &perf_ctrls, 1);
>>>     		if (ret) {
>>> @@ -452,16 +475,6 @@ static inline int amd_pstate_init_perf(struct amd_cpudata *cpudata)
>>>     	return static_call(amd_pstate_init_perf)(cpudata);
>>>     }
>>>     
>>> -static void pstate_update_perf(struct amd_cpudata *cpudata, u32 min_perf,
>>> -			       u32 des_perf, u32 max_perf, bool fast_switch)
>>> -{
>>> -	if (fast_switch)
>>> -		wrmsrl(MSR_AMD_CPPC_REQ, READ_ONCE(cpudata->cppc_req_cached));
>>> -	else
>>> -		wrmsrl_on_cpu(cpudata->cpu, MSR_AMD_CPPC_REQ,
>>> -			      READ_ONCE(cpudata->cppc_req_cached));
>>> -}
>>> -
>>>     static void cppc_update_perf(struct amd_cpudata *cpudata,
>>>     			     u32 min_perf, u32 des_perf,
>>>     			     u32 max_perf, bool fast_switch) @@ -475,16 +488,6 @@
>>> static void cppc_update_perf(struct amd_cpudata *cpudata,
>>>     	cppc_set_perf(cpudata->cpu, &perf_ctrls);
>>>     }
>>>     
>>> -DEFINE_STATIC_CALL(amd_pstate_update_perf, pstate_update_perf);
>>> -
>>> -static inline void amd_pstate_update_perf(struct amd_cpudata *cpudata,
>>> -					  u32 min_perf, u32 des_perf,
>>> -					  u32 max_perf, bool fast_switch)
>>> -{
>>> -	static_call(amd_pstate_update_perf)(cpudata, min_perf, des_perf,
>>> -					    max_perf, fast_switch);
>>> -}
>>> -
>>>     static inline bool amd_pstate_sample(struct amd_cpudata *cpudata)
>>>     {
>>>     	u64 aperf, mperf, tsc;
>>
>>
>
Mario Limonciello Sept. 3, 2024, 8:51 p.m. UTC | #5
Morgan,

This does remind me of a clamping issue that touches some of the same 
variables.  Can you please see if backporting

commit 8164f7433264 ("cpufreq: amd-pstate: adjust min/max limit perf")

to 6.6.y helps for you?

Thanks,

On 9/3/2024 15:09, Mario Limonciello wrote:
> Morgan,
> 
> OK that's great news that it's just a backport effort.  That same commit 
> also backported to 6.10.3.  Can you see if 6.10.y is affected?
> 
> Ugwekar,
> 
> Any thoughts on what else needs to come back to 6.6.y off hand?
> 
> Thanks,
> 
> On 9/3/2024 15:07, Jones, Morgan wrote:
>> Hey Mario,
>>
>> Smoking gun here, the max frequency is incorrect on 6.6.44+ but is 
>> correct on 6.11.0-rc6.
>>
>> Linux redact 6.11.0-rc6 #1-NixOS SMP PREEMPT_DYNAMIC Tue Jan  1 
>> 00:00:00 UTC 1980 x86_64 GNU/Linux
>>
>> analyzing CPU 12:
>>    driver: amd-pstate-epp
>>    CPUs which run at the same hardware frequency: 12
>>    CPUs which need to have their frequency coordinated by software: 12
>>    maximum transition latency:  Cannot determine or is not supported.
>>    hardware limits: 400 MHz - 3.35 GHz
>>    available cpufreq governors: performance powersave
>>    current policy: frequency should be within 400 MHz and 3.35 GHz.
>>                    The governor "performance" may decide which speed 
>> to use
>>                    within this range.
>>    current CPU frequency: Unable to call hardware
>>    current CPU frequency: 3.34 GHz (asserted by call to kernel)
>>    boost state support:
>>      Supported: yes
>>      Active: yes
>>      AMD PSTATE Highest Performance: 255. Maximum Frequency: 3.35 GHz.
>>      AMD PSTATE Nominal Performance: 152. Nominal Frequency: 2.00 GHz.
>>      AMD PSTATE Lowest Non-linear Performance: 115. Lowest Non-linear 
>> Frequency: 1.51 GHz.
>>      AMD PSTATE Lowest Performance: 31. Lowest Frequency: 400 MHz.
>>
>> We're running amd_pstate=active and amd_pstate.shared_mem=1, and our 
>> workloads are back to normal performance on 6.11.0-rc6.
>>
>> Morgan
>>
>> -----Original Message-----
>> From: Mario Limonciello <mario.limonciello@amd.com>
>> Sent: Tuesday, September 3, 2024 10:55 AM
>> To: Jones, Morgan <Morgan.Jones@viasat.com>; Dhananjay Ugwekar 
>> <Dhananjay.Ugwekar@amd.com>; rafael@kernel.org; 
>> viresh.kumar@linaro.org; gautham.shenoy@amd.com; perry.yuan@amd.com; 
>> skhan@linuxfoundation.org; li.meng@amd.com; ray.huang@amd.com
>> Cc: linux-pm@vger.kernel.org; linux-kernel@vger.kernel.org; David 
>> Arcari <darcari@redhat.com>
>> Subject: [EXTERNAL] Re: [PATCH v2 2/2] cpufreq/amd-pstate: Fix the 
>> scaling_max_freq setting on shared memory CPPC systems
>>
>> Hi Morgan,
>>
>> Can you please cross reference 6.11-rc6 to see if you're still having 
>> a problem?  I would like to understand if we're missing a backport to 
>> stable or this is still an upstream problem.
>>
>> Thanks,
>>
>> On 9/3/2024 12:51, Jones, Morgan wrote:
>>> Hi there,
>>>
>>> We are experiencing a ~35% performance regression on an AMD EPYC 7702 
>>> 64 core machine after applying this patch. We observed Linux 6.6.44 
>>> starting to cause the issue, and performed a bisect involving 
>>> rebooting the machine over 15 times. (Note that kexec didn't 
>>> successfully identify the problem, since the PM memory mailbox is 
>>> never reset).
>>>
>>> It appears that we are getting a max of 2.18 GHz when this CPU can 
>>> boost to 3.35 GHz, explaining the slowdown. Blacklisting amd-pstate 
>>> solves the issue at the expense of the performance increase we used 
>>> to get by using it.
>>>
>>> Is it possible that the upper limits were implicitly at the max CPU 
>>> power before, and setting the upper limit to something other than the 
>>> boost frequency can reduce performance now?
>>>
>>> # bad: [7213910600667c51c978e577bf5454d3f7b313b7] Linux 6.6.44 # good:
>>> [58b0425ff5df680d0b67f64ae1f3f1ebdf1c4de9] Linux 6.6.43 git bisect
>>> start '7213910600667c51c978e577bf5454d3f7b313b7' 
>>> '58b0425ff5df680d0b67f64ae1f3f1ebdf1c4de9'
>>> # good: [72ff9d26964a3a80f7650df719df139f5c1f965d] arm64: dts: qcom:
>>> sm6350: Add missing qcom,non-secure-domain property git bisect good
>>> 72ff9d26964a3a80f7650df719df139f5c1f965d
>>> # good: [0fffc2e1bf40a2220ef5a38f834ea063dba832d3] ARM: dts: sunxi:
>>> remove duplicated entries in makefile git bisect good
>>> 0fffc2e1bf40a2220ef5a38f834ea063dba832d3
>>> # bad: [8cdbe6ebfd1763a5c41a2a3058497c0a9163311c] pinctrl: renesas:
>>> r8a779g0: Fix CANFD5 suffix git bisect bad
>>> 8cdbe6ebfd1763a5c41a2a3058497c0a9163311c
>>> # bad: [5dbb98e7fa42bebc1325899193d8f13f0705a148] drm/mediatek: Turn
>>> off the layers with zero width or height git bisect bad
>>> 5dbb98e7fa42bebc1325899193d8f13f0705a148
>>> # bad: [691ec7043122c9c8c46d84f6e6cd85d13d50cd93] selftests/bpf: Null
>>> checks for links in bpf_tcp_ca git bisect bad
>>> 691ec7043122c9c8c46d84f6e6cd85d13d50cd93
>>> # bad: [a1359e085d75d7393a250054e66c0a7bc6c3dbfa] perf/x86: Serialize
>>> set_attr_rdpmc() git bisect bad
>>> a1359e085d75d7393a250054e66c0a7bc6c3dbfa
>>> # bad: [e99d9b16ff153de9540073239d24adc3b0a3a997] wifi: ath12k: change
>>> DMA direction while mapping reinjected packets git bisect bad
>>> e99d9b16ff153de9540073239d24adc3b0a3a997
>>> # bad: [d027ac4a08541beb2a89563d3e034da7085050ba] firmware:
>>> turris-mox-rwtm: Initialize completion before mailbox git bisect bad
>>> d027ac4a08541beb2a89563d3e034da7085050ba
>>> # bad: [e6c9eca327e6a41a81e7eba0d0ddc13da37f82a1] ARM: spitz: fix GPIO
>>> assignment for backlight git bisect bad
>>> e6c9eca327e6a41a81e7eba0d0ddc13da37f82a1
>>> # bad: [b8cdefdaa555bbfc269c2198803f8791a8923960] m68k: cmpxchg: Fix
>>> return value for default case in __arch_xchg() git bisect bad
>>> b8cdefdaa555bbfc269c2198803f8791a8923960
>>> # bad: [13a71384ae6a8779da809b00c6f378dcead10427] cpufreq/amd-pstate:
>>> Fix the scaling_max_freq setting on shared memory CPPC systems git
>>> bisect bad 13a71384ae6a8779da809b00c6f378dcead10427
>>> # first bad commit: [13a71384ae6a8779da809b00c6f378dcead10427]
>>> cpufreq/amd-pstate: Fix the scaling_max_freq setting on shared memory
>>> CPPC systems
>>>
>>> cpupower output:
>>>
>>> analyzing CPU 47:
>>>     driver: amd-pstate-epp
>>>     CPUs which run at the same hardware frequency: 47
>>>     CPUs which need to have their frequency coordinated by software: 47
>>>     maximum transition latency:  Cannot determine or is not supported.
>>>     hardware limits: 400 MHz - 2.18 GHz
>>>     available cpufreq governors: performance powersave
>>>     current policy: frequency should be within 400 MHz and 2.18 GHz.
>>>                     The governor "performance" may decide which speed 
>>> to use
>>>                     within this range.
>>>     current CPU frequency: Unable to call hardware
>>>     current CPU frequency: 2.17 GHz (asserted by call to kernel)
>>>     boost state support:
>>>       Supported: yes
>>>       Active: yes
>>>       AMD PSTATE Highest Performance: 166. Maximum Frequency: 2.18 GHz.
>>>       AMD PSTATE Nominal Performance: 152. Nominal Frequency: 2.00 GHz.
>>>       AMD PSTATE Lowest Non-linear Performance: 115. Lowest 
>>> Non-linear Frequency: 1.51 GHz.
>>>       AMD PSTATE Lowest Performance: 31. Lowest Frequency: 400 MHz.
>>>
>>> Thanks,
>>> Morgan
>>>
>>> -----Original Message-----
>>> From: Mario Limonciello <mario.limonciello@amd.com>
>>> Sent: Tuesday, July 2, 2024 10:49 AM
>>> To: Dhananjay Ugwekar <Dhananjay.Ugwekar@amd.com>; rafael@kernel.org;
>>> viresh.kumar@linaro.org; gautham.shenoy@amd.com; perry.yuan@amd.com;
>>> skhan@linuxfoundation.org; li.meng@amd.com; ray.huang@amd.com
>>> Cc: linux-pm@vger.kernel.org; linux-kernel@vger.kernel.org; David
>>> Arcari <darcari@redhat.com>
>>> Subject: Re: [PATCH v2 2/2] cpufreq/amd-pstate: Fix the
>>> scaling_max_freq setting on shared memory CPPC systems
>>>
>>> On 7/2/2024 3:14, Dhananjay Ugwekar wrote:
>>>> On shared memory CPPC systems, with amd_pstate=active mode, the
>>>> change in scaling_max_freq doesn't get written to the shared memory 
>>>> region.
>>>> Due to this, the writes to the scaling_max_freq sysfs file don't take
>>>> effect. Fix this by propagating the scaling_max_freq changes to the
>>>> shared memory region.
>>>>
>>>> Fixes: ffa5096a7c33 ("cpufreq: amd-pstate: implement Pstate EPP
>>>> support for the AMD processors")
>>>> Reported-by: David Arcari <darcari@redhat.com>
>>>> Signed-off-by: Dhananjay Ugwekar <Dhananjay.Ugwekar@amd.com>
>>> Reviewed-by: Mario Limonciello <mario.limonciello@amd.com>
>>>> ---
>>>>     drivers/cpufreq/amd-pstate.c | 43 
>>>> +++++++++++++++++++-----------------
>>>>     1 file changed, 23 insertions(+), 20 deletions(-)
>>>>
>>>> diff --git a/drivers/cpufreq/amd-pstate.c
>>>> b/drivers/cpufreq/amd-pstate.c index 9ad62dbe8bfb..a092b13ffbc2
>>>> 100644
>>>> --- a/drivers/cpufreq/amd-pstate.c
>>>> +++ b/drivers/cpufreq/amd-pstate.c
>>>> @@ -247,6 +247,26 @@ static int 
>>>> amd_pstate_get_energy_pref_index(struct amd_cpudata *cpudata)
>>>>         return index;
>>>>     }
>>>> +static void pstate_update_perf(struct amd_cpudata *cpudata, u32 
>>>> min_perf,
>>>> +                   u32 des_perf, u32 max_perf, bool fast_switch) {
>>>> +    if (fast_switch)
>>>> +        wrmsrl(MSR_AMD_CPPC_REQ, READ_ONCE(cpudata->cppc_req_cached));
>>>> +    else
>>>> +        wrmsrl_on_cpu(cpudata->cpu, MSR_AMD_CPPC_REQ,
>>>> +                  READ_ONCE(cpudata->cppc_req_cached));
>>>> +}
>>>> +
>>>> +DEFINE_STATIC_CALL(amd_pstate_update_perf, pstate_update_perf);
>>>> +
>>>> +static inline void amd_pstate_update_perf(struct amd_cpudata *cpudata,
>>>> +                      u32 min_perf, u32 des_perf,
>>>> +                      u32 max_perf, bool fast_switch) {
>>>> +    static_call(amd_pstate_update_perf)(cpudata, min_perf, des_perf,
>>>> +                        max_perf, fast_switch);
>>>> +}
>>>> +
>>>>     static int amd_pstate_set_epp(struct amd_cpudata *cpudata, u32 epp)
>>>>     {
>>>>         int ret;
>>>> @@ -263,6 +283,9 @@ static int amd_pstate_set_epp(struct amd_cpudata 
>>>> *cpudata, u32 epp)
>>>>             if (!ret)
>>>>                 cpudata->epp_cached = epp;
>>>>         } else {
>>>> +        amd_pstate_update_perf(cpudata, cpudata->min_limit_perf, 0U,
>>>> +                         cpudata->max_limit_perf, false);
>>>> +
>>>>             perf_ctrls.energy_perf = epp;
>>>>             ret = cppc_set_epp_perf(cpudata->cpu, &perf_ctrls, 1);
>>>>             if (ret) {
>>>> @@ -452,16 +475,6 @@ static inline int amd_pstate_init_perf(struct 
>>>> amd_cpudata *cpudata)
>>>>         return static_call(amd_pstate_init_perf)(cpudata);
>>>>     }
>>>> -static void pstate_update_perf(struct amd_cpudata *cpudata, u32 
>>>> min_perf,
>>>> -                   u32 des_perf, u32 max_perf, bool fast_switch)
>>>> -{
>>>> -    if (fast_switch)
>>>> -        wrmsrl(MSR_AMD_CPPC_REQ, READ_ONCE(cpudata->cppc_req_cached));
>>>> -    else
>>>> -        wrmsrl_on_cpu(cpudata->cpu, MSR_AMD_CPPC_REQ,
>>>> -                  READ_ONCE(cpudata->cppc_req_cached));
>>>> -}
>>>> -
>>>>     static void cppc_update_perf(struct amd_cpudata *cpudata,
>>>>                      u32 min_perf, u32 des_perf,
>>>>                      u32 max_perf, bool fast_switch) @@ -475,16 
>>>> +488,6 @@
>>>> static void cppc_update_perf(struct amd_cpudata *cpudata,
>>>>         cppc_set_perf(cpudata->cpu, &perf_ctrls);
>>>>     }
>>>> -DEFINE_STATIC_CALL(amd_pstate_update_perf, pstate_update_perf);
>>>> -
>>>> -static inline void amd_pstate_update_perf(struct amd_cpudata *cpudata,
>>>> -                      u32 min_perf, u32 des_perf,
>>>> -                      u32 max_perf, bool fast_switch)
>>>> -{
>>>> -    static_call(amd_pstate_update_perf)(cpudata, min_perf, des_perf,
>>>> -                        max_perf, fast_switch);
>>>> -}
>>>> -
>>>>     static inline bool amd_pstate_sample(struct amd_cpudata *cpudata)
>>>>     {
>>>>         u64 aperf, mperf, tsc;
>>>
>>>
>>
>
Jones, Morgan Sept. 3, 2024, 8:52 p.m. UTC | #6
Seems like 6.10 is OK but 6.6 is not.

Linux redact 6.10.7 #1-NixOS SMP PREEMPT_DYNAMIC Tue Jan  1 00:00:00 UTC 1980 x86_64 GNU/Linux

analyzing CPU 50:
  driver: amd-pstate-epp
  CPUs which run at the same hardware frequency: 50
  CPUs which need to have their frequency coordinated by software: 50
  maximum transition latency:  Cannot determine or is not supported.
  hardware limits: 400 MHz - 3.35 GHz
  available cpufreq governors: performance powersave
  current policy: frequency should be within 400 MHz and 3.35 GHz.
                  The governor "performance" may decide which speed to use
                  within this range.
  current CPU frequency: Unable to call hardware
  current CPU frequency: 3.32 GHz (asserted by call to kernel)
  boost state support:
    Supported: yes
    Active: yes
    AMD PSTATE Highest Performance: 255. Maximum Frequency: 3.35 GHz.
    AMD PSTATE Nominal Performance: 152. Nominal Frequency: 2.00 GHz.
    AMD PSTATE Lowest Non-linear Performance: 115. Lowest Non-linear Frequency: 1.51 GHz.
    AMD PSTATE Lowest Performance: 31. Lowest Frequency: 400 MHz.


-----Original Message-----
From: Mario Limonciello <mario.limonciello@amd.com> 
Sent: Tuesday, September 3, 2024 1:10 PM
To: Jones, Morgan <Morgan.Jones@viasat.com>; Dhananjay Ugwekar <Dhananjay.Ugwekar@amd.com>; rafael@kernel.org; viresh.kumar@linaro.org; gautham.shenoy@amd.com; perry.yuan@amd.com; skhan@linuxfoundation.org; li.meng@amd.com; ray.huang@amd.com
Cc: linux-pm@vger.kernel.org; linux-kernel@vger.kernel.org; David Arcari <darcari@redhat.com>
Subject: Re: [EXTERNAL] Re: [PATCH v2 2/2] cpufreq/amd-pstate: Fix the scaling_max_freq setting on shared memory CPPC systems

Morgan,

OK that's great news that it's just a backport effort.  That same commit also backported to 6.10.3.  Can you see if 6.10.y is affected?

Ugwekar,

Any thoughts on what else needs to come back to 6.6.y off hand?

Thanks,

On 9/3/2024 15:07, Jones, Morgan wrote:
> Hey Mario,
> 
> Smoking gun here, the max frequency is incorrect on 6.6.44+ but is correct on 6.11.0-rc6.
> 
> Linux redact 6.11.0-rc6 #1-NixOS SMP PREEMPT_DYNAMIC Tue Jan  1 
> 00:00:00 UTC 1980 x86_64 GNU/Linux
> 
> analyzing CPU 12:
>    driver: amd-pstate-epp
>    CPUs which run at the same hardware frequency: 12
>    CPUs which need to have their frequency coordinated by software: 12
>    maximum transition latency:  Cannot determine or is not supported.
>    hardware limits: 400 MHz - 3.35 GHz
>    available cpufreq governors: performance powersave
>    current policy: frequency should be within 400 MHz and 3.35 GHz.
>                    The governor "performance" may decide which speed to use
>                    within this range.
>    current CPU frequency: Unable to call hardware
>    current CPU frequency: 3.34 GHz (asserted by call to kernel)
>    boost state support:
>      Supported: yes
>      Active: yes
>      AMD PSTATE Highest Performance: 255. Maximum Frequency: 3.35 GHz.
>      AMD PSTATE Nominal Performance: 152. Nominal Frequency: 2.00 GHz.
>      AMD PSTATE Lowest Non-linear Performance: 115. Lowest Non-linear Frequency: 1.51 GHz.
>      AMD PSTATE Lowest Performance: 31. Lowest Frequency: 400 MHz.
> 
> We're running amd_pstate=active and amd_pstate.shared_mem=1, and our workloads are back to normal performance on 6.11.0-rc6.
> 
> Morgan
> 
> -----Original Message-----
> From: Mario Limonciello <mario.limonciello@amd.com>
> Sent: Tuesday, September 3, 2024 10:55 AM
> To: Jones, Morgan <Morgan.Jones@viasat.com>; Dhananjay Ugwekar 
> <Dhananjay.Ugwekar@amd.com>; rafael@kernel.org; 
> viresh.kumar@linaro.org; gautham.shenoy@amd.com; perry.yuan@amd.com; 
> skhan@linuxfoundation.org; li.meng@amd.com; ray.huang@amd.com
> Cc: linux-pm@vger.kernel.org; linux-kernel@vger.kernel.org; David 
> Arcari <darcari@redhat.com>
> Subject: [EXTERNAL] Re: [PATCH v2 2/2] cpufreq/amd-pstate: Fix the 
> scaling_max_freq setting on shared memory CPPC systems
> 
> Hi Morgan,
> 
> Can you please cross reference 6.11-rc6 to see if you're still having a problem?  I would like to understand if we're missing a backport to stable or this is still an upstream problem.
> 
> Thanks,
> 
> On 9/3/2024 12:51, Jones, Morgan wrote:
>> Hi there,
>>
>> We are experiencing a ~35% performance regression on an AMD EPYC 7702 64 core machine after applying this patch. We observed Linux 6.6.44 starting to cause the issue, and performed a bisect involving rebooting the machine over 15 times. (Note that kexec didn't successfully identify the problem, since the PM memory mailbox is never reset).
>>
>> It appears that we are getting a max of 2.18 GHz when this CPU can boost to 3.35 GHz, explaining the slowdown. Blacklisting amd-pstate solves the issue at the expense of the performance increase we used to get by using it.
>>
>> Is it possible that the upper limits were implicitly at the max CPU power before, and setting the upper limit to something other than the boost frequency can reduce performance now?
>>
>> # bad: [7213910600667c51c978e577bf5454d3f7b313b7] Linux 6.6.44 # good:
>> [58b0425ff5df680d0b67f64ae1f3f1ebdf1c4de9] Linux 6.6.43 git bisect 
>> start '7213910600667c51c978e577bf5454d3f7b313b7' '58b0425ff5df680d0b67f64ae1f3f1ebdf1c4de9'
>> # good: [72ff9d26964a3a80f7650df719df139f5c1f965d] arm64: dts: qcom:
>> sm6350: Add missing qcom,non-secure-domain property git bisect good 
>> 72ff9d26964a3a80f7650df719df139f5c1f965d
>> # good: [0fffc2e1bf40a2220ef5a38f834ea063dba832d3] ARM: dts: sunxi:
>> remove duplicated entries in makefile git bisect good
>> 0fffc2e1bf40a2220ef5a38f834ea063dba832d3
>> # bad: [8cdbe6ebfd1763a5c41a2a3058497c0a9163311c] pinctrl: renesas:
>> r8a779g0: Fix CANFD5 suffix git bisect bad 
>> 8cdbe6ebfd1763a5c41a2a3058497c0a9163311c
>> # bad: [5dbb98e7fa42bebc1325899193d8f13f0705a148] drm/mediatek: Turn 
>> off the layers with zero width or height git bisect bad
>> 5dbb98e7fa42bebc1325899193d8f13f0705a148
>> # bad: [691ec7043122c9c8c46d84f6e6cd85d13d50cd93] selftests/bpf: Null 
>> checks for links in bpf_tcp_ca git bisect bad
>> 691ec7043122c9c8c46d84f6e6cd85d13d50cd93
>> # bad: [a1359e085d75d7393a250054e66c0a7bc6c3dbfa] perf/x86: Serialize
>> set_attr_rdpmc() git bisect bad
>> a1359e085d75d7393a250054e66c0a7bc6c3dbfa
>> # bad: [e99d9b16ff153de9540073239d24adc3b0a3a997] wifi: ath12k: 
>> change DMA direction while mapping reinjected packets git bisect bad
>> e99d9b16ff153de9540073239d24adc3b0a3a997
>> # bad: [d027ac4a08541beb2a89563d3e034da7085050ba] firmware:
>> turris-mox-rwtm: Initialize completion before mailbox git bisect bad 
>> d027ac4a08541beb2a89563d3e034da7085050ba
>> # bad: [e6c9eca327e6a41a81e7eba0d0ddc13da37f82a1] ARM: spitz: fix 
>> GPIO assignment for backlight git bisect bad
>> e6c9eca327e6a41a81e7eba0d0ddc13da37f82a1
>> # bad: [b8cdefdaa555bbfc269c2198803f8791a8923960] m68k: cmpxchg: Fix 
>> return value for default case in __arch_xchg() git bisect bad
>> b8cdefdaa555bbfc269c2198803f8791a8923960
>> # bad: [13a71384ae6a8779da809b00c6f378dcead10427] cpufreq/amd-pstate:
>> Fix the scaling_max_freq setting on shared memory CPPC systems git 
>> bisect bad 13a71384ae6a8779da809b00c6f378dcead10427
>> # first bad commit: [13a71384ae6a8779da809b00c6f378dcead10427]
>> cpufreq/amd-pstate: Fix the scaling_max_freq setting on shared memory 
>> CPPC systems
>>
>> cpupower output:
>>
>> analyzing CPU 47:
>>     driver: amd-pstate-epp
>>     CPUs which run at the same hardware frequency: 47
>>     CPUs which need to have their frequency coordinated by software: 47
>>     maximum transition latency:  Cannot determine or is not supported.
>>     hardware limits: 400 MHz - 2.18 GHz
>>     available cpufreq governors: performance powersave
>>     current policy: frequency should be within 400 MHz and 2.18 GHz.
>>                     The governor "performance" may decide which speed to use
>>                     within this range.
>>     current CPU frequency: Unable to call hardware
>>     current CPU frequency: 2.17 GHz (asserted by call to kernel)
>>     boost state support:
>>       Supported: yes
>>       Active: yes
>>       AMD PSTATE Highest Performance: 166. Maximum Frequency: 2.18 GHz.
>>       AMD PSTATE Nominal Performance: 152. Nominal Frequency: 2.00 GHz.
>>       AMD PSTATE Lowest Non-linear Performance: 115. Lowest Non-linear Frequency: 1.51 GHz.
>>       AMD PSTATE Lowest Performance: 31. Lowest Frequency: 400 MHz.
>>
>> Thanks,
>> Morgan
>>
>> -----Original Message-----
>> From: Mario Limonciello <mario.limonciello@amd.com>
>> Sent: Tuesday, July 2, 2024 10:49 AM
>> To: Dhananjay Ugwekar <Dhananjay.Ugwekar@amd.com>; rafael@kernel.org; 
>> viresh.kumar@linaro.org; gautham.shenoy@amd.com; perry.yuan@amd.com; 
>> skhan@linuxfoundation.org; li.meng@amd.com; ray.huang@amd.com
>> Cc: linux-pm@vger.kernel.org; linux-kernel@vger.kernel.org; David 
>> Arcari <darcari@redhat.com>
>> Subject: Re: [PATCH v2 2/2] cpufreq/amd-pstate: Fix the 
>> scaling_max_freq setting on shared memory CPPC systems
>>
>> On 7/2/2024 3:14, Dhananjay Ugwekar wrote:
>>> On shared memory CPPC systems, with amd_pstate=active mode, the 
>>> change in scaling_max_freq doesn't get written to the shared memory region.
>>> Due to this, the writes to the scaling_max_freq sysfs file don't 
>>> take effect. Fix this by propagating the scaling_max_freq changes to 
>>> the shared memory region.
>>>
>>> Fixes: ffa5096a7c33 ("cpufreq: amd-pstate: implement Pstate EPP 
>>> support for the AMD processors")
>>> Reported-by: David Arcari <darcari@redhat.com>
>>> Signed-off-by: Dhananjay Ugwekar <Dhananjay.Ugwekar@amd.com>
>> Reviewed-by: Mario Limonciello <mario.limonciello@amd.com>
>>> ---
>>>     drivers/cpufreq/amd-pstate.c | 43 +++++++++++++++++++-----------------
>>>     1 file changed, 23 insertions(+), 20 deletions(-)
>>>
>>> diff --git a/drivers/cpufreq/amd-pstate.c 
>>> b/drivers/cpufreq/amd-pstate.c index 9ad62dbe8bfb..a092b13ffbc2
>>> 100644
>>> --- a/drivers/cpufreq/amd-pstate.c
>>> +++ b/drivers/cpufreq/amd-pstate.c
>>> @@ -247,6 +247,26 @@ static int amd_pstate_get_energy_pref_index(struct amd_cpudata *cpudata)
>>>     	return index;
>>>     }
>>>     
>>> +static void pstate_update_perf(struct amd_cpudata *cpudata, u32 min_perf,
>>> +			       u32 des_perf, u32 max_perf, bool fast_switch) {
>>> +	if (fast_switch)
>>> +		wrmsrl(MSR_AMD_CPPC_REQ, READ_ONCE(cpudata->cppc_req_cached));
>>> +	else
>>> +		wrmsrl_on_cpu(cpudata->cpu, MSR_AMD_CPPC_REQ,
>>> +			      READ_ONCE(cpudata->cppc_req_cached));
>>> +}
>>> +
>>> +DEFINE_STATIC_CALL(amd_pstate_update_perf, pstate_update_perf);
>>> +
>>> +static inline void amd_pstate_update_perf(struct amd_cpudata *cpudata,
>>> +					  u32 min_perf, u32 des_perf,
>>> +					  u32 max_perf, bool fast_switch) {
>>> +	static_call(amd_pstate_update_perf)(cpudata, min_perf, des_perf,
>>> +					    max_perf, fast_switch);
>>> +}
>>> +
>>>     static int amd_pstate_set_epp(struct amd_cpudata *cpudata, u32 epp)
>>>     {
>>>     	int ret;
>>> @@ -263,6 +283,9 @@ static int amd_pstate_set_epp(struct amd_cpudata *cpudata, u32 epp)
>>>     		if (!ret)
>>>     			cpudata->epp_cached = epp;
>>>     	} else {
>>> +		amd_pstate_update_perf(cpudata, cpudata->min_limit_perf, 0U,
>>> +					     cpudata->max_limit_perf, false);
>>> +
>>>     		perf_ctrls.energy_perf = epp;
>>>     		ret = cppc_set_epp_perf(cpudata->cpu, &perf_ctrls, 1);
>>>     		if (ret) {
>>> @@ -452,16 +475,6 @@ static inline int amd_pstate_init_perf(struct amd_cpudata *cpudata)
>>>     	return static_call(amd_pstate_init_perf)(cpudata);
>>>     }
>>>     
>>> -static void pstate_update_perf(struct amd_cpudata *cpudata, u32 min_perf,
>>> -			       u32 des_perf, u32 max_perf, bool fast_switch)
>>> -{
>>> -	if (fast_switch)
>>> -		wrmsrl(MSR_AMD_CPPC_REQ, READ_ONCE(cpudata->cppc_req_cached));
>>> -	else
>>> -		wrmsrl_on_cpu(cpudata->cpu, MSR_AMD_CPPC_REQ,
>>> -			      READ_ONCE(cpudata->cppc_req_cached));
>>> -}
>>> -
>>>     static void cppc_update_perf(struct amd_cpudata *cpudata,
>>>     			     u32 min_perf, u32 des_perf,
>>>     			     u32 max_perf, bool fast_switch) @@ -475,16 +488,6 @@ 
>>> static void cppc_update_perf(struct amd_cpudata *cpudata,
>>>     	cppc_set_perf(cpudata->cpu, &perf_ctrls);
>>>     }
>>>     
>>> -DEFINE_STATIC_CALL(amd_pstate_update_perf, pstate_update_perf);
>>> -
>>> -static inline void amd_pstate_update_perf(struct amd_cpudata *cpudata,
>>> -					  u32 min_perf, u32 des_perf,
>>> -					  u32 max_perf, bool fast_switch)
>>> -{
>>> -	static_call(amd_pstate_update_perf)(cpudata, min_perf, des_perf,
>>> -					    max_perf, fast_switch);
>>> -}
>>> -
>>>     static inline bool amd_pstate_sample(struct amd_cpudata *cpudata)
>>>     {
>>>     	u64 aperf, mperf, tsc;
>>
>>
>
Jones, Morgan Sept. 3, 2024, 8:52 p.m. UTC | #7
On it.

-----Original Message-----
From: Mario Limonciello <mario.limonciello@amd.com> 
Sent: Tuesday, September 3, 2024 1:52 PM
To: Jones, Morgan <Morgan.Jones@viasat.com>; Dhananjay Ugwekar <Dhananjay.Ugwekar@amd.com>; rafael@kernel.org; viresh.kumar@linaro.org; gautham.shenoy@amd.com; perry.yuan@amd.com; skhan@linuxfoundation.org; li.meng@amd.com; ray.huang@amd.com
Cc: linux-pm@vger.kernel.org; linux-kernel@vger.kernel.org; David Arcari <darcari@redhat.com>
Subject: Re: [EXTERNAL] Re: [PATCH v2 2/2] cpufreq/amd-pstate: Fix the scaling_max_freq setting on shared memory CPPC systems

Morgan,

This does remind me of a clamping issue that touches some of the same variables.  Can you please see if backporting

commit 8164f7433264 ("cpufreq: amd-pstate: adjust min/max limit perf")

to 6.6.y helps for you?

Thanks,

On 9/3/2024 15:09, Mario Limonciello wrote:
> Morgan,
> 
> OK that's great news that it's just a backport effort.  That same 
> commit also backported to 6.10.3.  Can you see if 6.10.y is affected?
> 
> Ugwekar,
> 
> Any thoughts on what else needs to come back to 6.6.y off hand?
> 
> Thanks,
> 
> On 9/3/2024 15:07, Jones, Morgan wrote:
>> Hey Mario,
>>
>> Smoking gun here, the max frequency is incorrect on 6.6.44+ but is 
>> correct on 6.11.0-rc6.
>>
>> Linux redact 6.11.0-rc6 #1-NixOS SMP PREEMPT_DYNAMIC Tue Jan  1
>> 00:00:00 UTC 1980 x86_64 GNU/Linux
>>
>> analyzing CPU 12:
>>    driver: amd-pstate-epp
>>    CPUs which run at the same hardware frequency: 12
>>    CPUs which need to have their frequency coordinated by software: 
>> 12
>>    maximum transition latency:  Cannot determine or is not supported.
>>    hardware limits: 400 MHz - 3.35 GHz
>>    available cpufreq governors: performance powersave
>>    current policy: frequency should be within 400 MHz and 3.35 GHz.
>>                    The governor "performance" may decide which speed 
>> to use
>>                    within this range.
>>    current CPU frequency: Unable to call hardware
>>    current CPU frequency: 3.34 GHz (asserted by call to kernel)
>>    boost state support:
>>      Supported: yes
>>      Active: yes
>>      AMD PSTATE Highest Performance: 255. Maximum Frequency: 3.35 GHz.
>>      AMD PSTATE Nominal Performance: 152. Nominal Frequency: 2.00 GHz.
>>      AMD PSTATE Lowest Non-linear Performance: 115. Lowest Non-linear
>> Frequency: 1.51 GHz.
>>      AMD PSTATE Lowest Performance: 31. Lowest Frequency: 400 MHz.
>>
>> We're running amd_pstate=active and amd_pstate.shared_mem=1, and our 
>> workloads are back to normal performance on 6.11.0-rc6.
>>
>> Morgan
>>
>> -----Original Message-----
>> From: Mario Limonciello <mario.limonciello@amd.com>
>> Sent: Tuesday, September 3, 2024 10:55 AM
>> To: Jones, Morgan <Morgan.Jones@viasat.com>; Dhananjay Ugwekar 
>> <Dhananjay.Ugwekar@amd.com>; rafael@kernel.org; 
>> viresh.kumar@linaro.org; gautham.shenoy@amd.com; perry.yuan@amd.com; 
>> skhan@linuxfoundation.org; li.meng@amd.com; ray.huang@amd.com
>> Cc: linux-pm@vger.kernel.org; linux-kernel@vger.kernel.org; David 
>> Arcari <darcari@redhat.com>
>> Subject: [EXTERNAL] Re: [PATCH v2 2/2] cpufreq/amd-pstate: Fix the 
>> scaling_max_freq setting on shared memory CPPC systems
>>
>> Hi Morgan,
>>
>> Can you please cross reference 6.11-rc6 to see if you're still having 
>> a problem?  I would like to understand if we're missing a backport to 
>> stable or this is still an upstream problem.
>>
>> Thanks,
>>
>> On 9/3/2024 12:51, Jones, Morgan wrote:
>>> Hi there,
>>>
>>> We are experiencing a ~35% performance regression on an AMD EPYC 
>>> 7702
>>> 64 core machine after applying this patch. We observed Linux 6.6.44 
>>> starting to cause the issue, and performed a bisect involving 
>>> rebooting the machine over 15 times. (Note that kexec didn't 
>>> successfully identify the problem, since the PM memory mailbox is 
>>> never reset).
>>>
>>> It appears that we are getting a max of 2.18 GHz when this CPU can 
>>> boost to 3.35 GHz, explaining the slowdown. Blacklisting amd-pstate 
>>> solves the issue at the expense of the performance increase we used 
>>> to get by using it.
>>>
>>> Is it possible that the upper limits were implicitly at the max CPU 
>>> power before, and setting the upper limit to something other than 
>>> the boost frequency can reduce performance now?
>>>
>>> # bad: [7213910600667c51c978e577bf5454d3f7b313b7] Linux 6.6.44 # good:
>>> [58b0425ff5df680d0b67f64ae1f3f1ebdf1c4de9] Linux 6.6.43 git bisect 
>>> start '7213910600667c51c978e577bf5454d3f7b313b7'
>>> '58b0425ff5df680d0b67f64ae1f3f1ebdf1c4de9'
>>> # good: [72ff9d26964a3a80f7650df719df139f5c1f965d] arm64: dts: qcom:
>>> sm6350: Add missing qcom,non-secure-domain property git bisect good 
>>> 72ff9d26964a3a80f7650df719df139f5c1f965d
>>> # good: [0fffc2e1bf40a2220ef5a38f834ea063dba832d3] ARM: dts: sunxi:
>>> remove duplicated entries in makefile git bisect good
>>> 0fffc2e1bf40a2220ef5a38f834ea063dba832d3
>>> # bad: [8cdbe6ebfd1763a5c41a2a3058497c0a9163311c] pinctrl: renesas:
>>> r8a779g0: Fix CANFD5 suffix git bisect bad 
>>> 8cdbe6ebfd1763a5c41a2a3058497c0a9163311c
>>> # bad: [5dbb98e7fa42bebc1325899193d8f13f0705a148] drm/mediatek: Turn 
>>> off the layers with zero width or height git bisect bad
>>> 5dbb98e7fa42bebc1325899193d8f13f0705a148
>>> # bad: [691ec7043122c9c8c46d84f6e6cd85d13d50cd93] selftests/bpf: 
>>> Null checks for links in bpf_tcp_ca git bisect bad
>>> 691ec7043122c9c8c46d84f6e6cd85d13d50cd93
>>> # bad: [a1359e085d75d7393a250054e66c0a7bc6c3dbfa] perf/x86: 
>>> Serialize
>>> set_attr_rdpmc() git bisect bad
>>> a1359e085d75d7393a250054e66c0a7bc6c3dbfa
>>> # bad: [e99d9b16ff153de9540073239d24adc3b0a3a997] wifi: ath12k: 
>>> change DMA direction while mapping reinjected packets git bisect bad
>>> e99d9b16ff153de9540073239d24adc3b0a3a997
>>> # bad: [d027ac4a08541beb2a89563d3e034da7085050ba] firmware:
>>> turris-mox-rwtm: Initialize completion before mailbox git bisect bad 
>>> d027ac4a08541beb2a89563d3e034da7085050ba
>>> # bad: [e6c9eca327e6a41a81e7eba0d0ddc13da37f82a1] ARM: spitz: fix 
>>> GPIO assignment for backlight git bisect bad
>>> e6c9eca327e6a41a81e7eba0d0ddc13da37f82a1
>>> # bad: [b8cdefdaa555bbfc269c2198803f8791a8923960] m68k: cmpxchg: Fix 
>>> return value for default case in __arch_xchg() git bisect bad
>>> b8cdefdaa555bbfc269c2198803f8791a8923960
>>> # bad: [13a71384ae6a8779da809b00c6f378dcead10427] cpufreq/amd-pstate:
>>> Fix the scaling_max_freq setting on shared memory CPPC systems git 
>>> bisect bad 13a71384ae6a8779da809b00c6f378dcead10427
>>> # first bad commit: [13a71384ae6a8779da809b00c6f378dcead10427]
>>> cpufreq/amd-pstate: Fix the scaling_max_freq setting on shared 
>>> memory CPPC systems
>>>
>>> cpupower output:
>>>
>>> analyzing CPU 47:
>>>     driver: amd-pstate-epp
>>>     CPUs which run at the same hardware frequency: 47
>>>     CPUs which need to have their frequency coordinated by software: 
>>> 47
>>>     maximum transition latency:  Cannot determine or is not supported.
>>>     hardware limits: 400 MHz - 2.18 GHz
>>>     available cpufreq governors: performance powersave
>>>     current policy: frequency should be within 400 MHz and 2.18 GHz.
>>>                     The governor "performance" may decide which 
>>> speed to use
>>>                     within this range.
>>>     current CPU frequency: Unable to call hardware
>>>     current CPU frequency: 2.17 GHz (asserted by call to kernel)
>>>     boost state support:
>>>       Supported: yes
>>>       Active: yes
>>>       AMD PSTATE Highest Performance: 166. Maximum Frequency: 2.18 GHz.
>>>       AMD PSTATE Nominal Performance: 152. Nominal Frequency: 2.00 GHz.
>>>       AMD PSTATE Lowest Non-linear Performance: 115. Lowest 
>>> Non-linear Frequency: 1.51 GHz.
>>>       AMD PSTATE Lowest Performance: 31. Lowest Frequency: 400 MHz.
>>>
>>> Thanks,
>>> Morgan
>>>
>>> -----Original Message-----
>>> From: Mario Limonciello <mario.limonciello@amd.com>
>>> Sent: Tuesday, July 2, 2024 10:49 AM
>>> To: Dhananjay Ugwekar <Dhananjay.Ugwekar@amd.com>; 
>>> rafael@kernel.org; viresh.kumar@linaro.org; gautham.shenoy@amd.com; 
>>> perry.yuan@amd.com; skhan@linuxfoundation.org; li.meng@amd.com; 
>>> ray.huang@amd.com
>>> Cc: linux-pm@vger.kernel.org; linux-kernel@vger.kernel.org; David 
>>> Arcari <darcari@redhat.com>
>>> Subject: Re: [PATCH v2 2/2] cpufreq/amd-pstate: Fix the 
>>> scaling_max_freq setting on shared memory CPPC systems
>>>
>>> On 7/2/2024 3:14, Dhananjay Ugwekar wrote:
>>>> On shared memory CPPC systems, with amd_pstate=active mode, the 
>>>> change in scaling_max_freq doesn't get written to the shared memory 
>>>> region.
>>>> Due to this, the writes to the scaling_max_freq sysfs file don't 
>>>> take effect. Fix this by propagating the scaling_max_freq changes 
>>>> to the shared memory region.
>>>>
>>>> Fixes: ffa5096a7c33 ("cpufreq: amd-pstate: implement Pstate EPP 
>>>> support for the AMD processors")
>>>> Reported-by: David Arcari <darcari@redhat.com>
>>>> Signed-off-by: Dhananjay Ugwekar <Dhananjay.Ugwekar@amd.com>
>>> Reviewed-by: Mario Limonciello <mario.limonciello@amd.com>
>>>> ---
>>>>     drivers/cpufreq/amd-pstate.c | 43
>>>> +++++++++++++++++++-----------------
>>>>     1 file changed, 23 insertions(+), 20 deletions(-)
>>>>
>>>> diff --git a/drivers/cpufreq/amd-pstate.c 
>>>> b/drivers/cpufreq/amd-pstate.c index 9ad62dbe8bfb..a092b13ffbc2
>>>> 100644
>>>> --- a/drivers/cpufreq/amd-pstate.c
>>>> +++ b/drivers/cpufreq/amd-pstate.c
>>>> @@ -247,6 +247,26 @@ static int
>>>> amd_pstate_get_energy_pref_index(struct amd_cpudata *cpudata)
>>>>         return index;
>>>>     }
>>>> +static void pstate_update_perf(struct amd_cpudata *cpudata, u32
>>>> min_perf,
>>>> +                   u32 des_perf, u32 max_perf, bool fast_switch) {
>>>> +    if (fast_switch)
>>>> +        wrmsrl(MSR_AMD_CPPC_REQ, 
>>>> +READ_ONCE(cpudata->cppc_req_cached));
>>>> +    else
>>>> +        wrmsrl_on_cpu(cpudata->cpu, MSR_AMD_CPPC_REQ,
>>>> +                  READ_ONCE(cpudata->cppc_req_cached));
>>>> +}
>>>> +
>>>> +DEFINE_STATIC_CALL(amd_pstate_update_perf, pstate_update_perf);
>>>> +
>>>> +static inline void amd_pstate_update_perf(struct amd_cpudata 
>>>> +*cpudata,
>>>> +                      u32 min_perf, u32 des_perf,
>>>> +                      u32 max_perf, bool fast_switch) {
>>>> +    static_call(amd_pstate_update_perf)(cpudata, min_perf, 
>>>> +des_perf,
>>>> +                        max_perf, fast_switch); }
>>>> +
>>>>     static int amd_pstate_set_epp(struct amd_cpudata *cpudata, u32 
>>>> epp)
>>>>     {
>>>>         int ret;
>>>> @@ -263,6 +283,9 @@ static int amd_pstate_set_epp(struct 
>>>> amd_cpudata *cpudata, u32 epp)
>>>>             if (!ret)
>>>>                 cpudata->epp_cached = epp;
>>>>         } else {
>>>> +        amd_pstate_update_perf(cpudata, cpudata->min_limit_perf, 
>>>> +0U,
>>>> +                         cpudata->max_limit_perf, false);
>>>> +
>>>>             perf_ctrls.energy_perf = epp;
>>>>             ret = cppc_set_epp_perf(cpudata->cpu, &perf_ctrls, 1);
>>>>             if (ret) {
>>>> @@ -452,16 +475,6 @@ static inline int amd_pstate_init_perf(struct 
>>>> amd_cpudata *cpudata)
>>>>         return static_call(amd_pstate_init_perf)(cpudata);
>>>>     }
>>>> -static void pstate_update_perf(struct amd_cpudata *cpudata, u32 
>>>> min_perf,
>>>> -                   u32 des_perf, u32 max_perf, bool fast_switch) 
>>>> -{
>>>> -    if (fast_switch)
>>>> -        wrmsrl(MSR_AMD_CPPC_REQ, 
>>>> READ_ONCE(cpudata->cppc_req_cached));
>>>> -    else
>>>> -        wrmsrl_on_cpu(cpudata->cpu, MSR_AMD_CPPC_REQ,
>>>> -                  READ_ONCE(cpudata->cppc_req_cached));
>>>> -}
>>>> -
>>>>     static void cppc_update_perf(struct amd_cpudata *cpudata,
>>>>                      u32 min_perf, u32 des_perf,
>>>>                      u32 max_perf, bool fast_switch) @@ -475,16
>>>> +488,6 @@
>>>> static void cppc_update_perf(struct amd_cpudata *cpudata,
>>>>         cppc_set_perf(cpudata->cpu, &perf_ctrls);
>>>>     }
>>>> -DEFINE_STATIC_CALL(amd_pstate_update_perf, pstate_update_perf);
>>>> -
>>>> -static inline void amd_pstate_update_perf(struct amd_cpudata 
>>>> *cpudata,
>>>> -                      u32 min_perf, u32 des_perf,
>>>> -                      u32 max_perf, bool fast_switch) -{
>>>> -    static_call(amd_pstate_update_perf)(cpudata, min_perf, 
>>>> des_perf,
>>>> -                        max_perf, fast_switch); -}
>>>> -
>>>>     static inline bool amd_pstate_sample(struct amd_cpudata 
>>>> *cpudata)
>>>>     {
>>>>         u64 aperf, mperf, tsc;
>>>
>>>
>>
>
Jones, Morgan Sept. 3, 2024, 10:21 p.m. UTC | #8
With and without the patch, I get the same:

Linux redact 6.6.48 #1-NixOS SMP PREEMPT_DYNAMIC Tue Jan  1 00:00:00 UTC 1980 x86_64 GNU/Linux

analyzing CPU 31:
  driver: amd-pstate-epp
  CPUs which run at the same hardware frequency: 31
  CPUs which need to have their frequency coordinated by software: 31
  maximum transition latency:  Cannot determine or is not supported.
  hardware limits: 400 MHz - 2.18 GHz
  available cpufreq governors: performance powersave
  current policy: frequency should be within 400 MHz and 2.18 GHz.
                  The governor "performance" may decide which speed to use
                  within this range.
  current CPU frequency: Unable to call hardware
  current CPU frequency: 2.17 GHz (asserted by call to kernel)
  boost state support:
    Supported: yes
    Active: yes
    AMD PSTATE Highest Performance: 166. Maximum Frequency: 2.18 GHz.
    AMD PSTATE Nominal Performance: 152. Nominal Frequency: 2.00 GHz.
    AMD PSTATE Lowest Non-linear Performance: 115. Lowest Non-linear Frequency: 1.51 GHz.
    AMD PSTATE Lowest Performance: 31. Lowest Frequency: 400 MHz.


-----Original Message-----
From: Mario Limonciello <mario.limonciello@amd.com> 
Sent: Tuesday, September 3, 2024 1:52 PM
To: Jones, Morgan <Morgan.Jones@viasat.com>; Dhananjay Ugwekar <Dhananjay.Ugwekar@amd.com>; rafael@kernel.org; viresh.kumar@linaro.org; gautham.shenoy@amd.com; perry.yuan@amd.com; skhan@linuxfoundation.org; li.meng@amd.com; ray.huang@amd.com
Cc: linux-pm@vger.kernel.org; linux-kernel@vger.kernel.org; David Arcari <darcari@redhat.com>
Subject: Re: [EXTERNAL] Re: [PATCH v2 2/2] cpufreq/amd-pstate: Fix the scaling_max_freq setting on shared memory CPPC systems

Morgan,

This does remind me of a clamping issue that touches some of the same variables.  Can you please see if backporting

commit 8164f7433264 ("cpufreq: amd-pstate: adjust min/max limit perf")

to 6.6.y helps for you?

Thanks,

On 9/3/2024 15:09, Mario Limonciello wrote:
> Morgan,
> 
> OK that's great news that it's just a backport effort.  That same 
> commit also backported to 6.10.3.  Can you see if 6.10.y is affected?
> 
> Ugwekar,
> 
> Any thoughts on what else needs to come back to 6.6.y off hand?
> 
> Thanks,
> 
> On 9/3/2024 15:07, Jones, Morgan wrote:
>> Hey Mario,
>>
>> Smoking gun here, the max frequency is incorrect on 6.6.44+ but is 
>> correct on 6.11.0-rc6.
>>
>> Linux redact 6.11.0-rc6 #1-NixOS SMP PREEMPT_DYNAMIC Tue Jan  1
>> 00:00:00 UTC 1980 x86_64 GNU/Linux
>>
>> analyzing CPU 12:
>>    driver: amd-pstate-epp
>>    CPUs which run at the same hardware frequency: 12
>>    CPUs which need to have their frequency coordinated by software: 
>> 12
>>    maximum transition latency:  Cannot determine or is not supported.
>>    hardware limits: 400 MHz - 3.35 GHz
>>    available cpufreq governors: performance powersave
>>    current policy: frequency should be within 400 MHz and 3.35 GHz.
>>                    The governor "performance" may decide which speed 
>> to use
>>                    within this range.
>>    current CPU frequency: Unable to call hardware
>>    current CPU frequency: 3.34 GHz (asserted by call to kernel)
>>    boost state support:
>>      Supported: yes
>>      Active: yes
>>      AMD PSTATE Highest Performance: 255. Maximum Frequency: 3.35 GHz.
>>      AMD PSTATE Nominal Performance: 152. Nominal Frequency: 2.00 GHz.
>>      AMD PSTATE Lowest Non-linear Performance: 115. Lowest Non-linear
>> Frequency: 1.51 GHz.
>>      AMD PSTATE Lowest Performance: 31. Lowest Frequency: 400 MHz.
>>
>> We're running amd_pstate=active and amd_pstate.shared_mem=1, and our 
>> workloads are back to normal performance on 6.11.0-rc6.
>>
>> Morgan
>>
>> -----Original Message-----
>> From: Mario Limonciello <mario.limonciello@amd.com>
>> Sent: Tuesday, September 3, 2024 10:55 AM
>> To: Jones, Morgan <Morgan.Jones@viasat.com>; Dhananjay Ugwekar 
>> <Dhananjay.Ugwekar@amd.com>; rafael@kernel.org; 
>> viresh.kumar@linaro.org; gautham.shenoy@amd.com; perry.yuan@amd.com; 
>> skhan@linuxfoundation.org; li.meng@amd.com; ray.huang@amd.com
>> Cc: linux-pm@vger.kernel.org; linux-kernel@vger.kernel.org; David 
>> Arcari <darcari@redhat.com>
>> Subject: [EXTERNAL] Re: [PATCH v2 2/2] cpufreq/amd-pstate: Fix the 
>> scaling_max_freq setting on shared memory CPPC systems
>>
>> Hi Morgan,
>>
>> Can you please cross reference 6.11-rc6 to see if you're still having 
>> a problem?  I would like to understand if we're missing a backport to 
>> stable or this is still an upstream problem.
>>
>> Thanks,
>>
>> On 9/3/2024 12:51, Jones, Morgan wrote:
>>> Hi there,
>>>
>>> We are experiencing a ~35% performance regression on an AMD EPYC 
>>> 7702
>>> 64 core machine after applying this patch. We observed Linux 6.6.44 
>>> starting to cause the issue, and performed a bisect involving 
>>> rebooting the machine over 15 times. (Note that kexec didn't 
>>> successfully identify the problem, since the PM memory mailbox is 
>>> never reset).
>>>
>>> It appears that we are getting a max of 2.18 GHz when this CPU can 
>>> boost to 3.35 GHz, explaining the slowdown. Blacklisting amd-pstate 
>>> solves the issue at the expense of the performance increase we used 
>>> to get by using it.
>>>
>>> Is it possible that the upper limits were implicitly at the max CPU 
>>> power before, and setting the upper limit to something other than 
>>> the boost frequency can reduce performance now?
>>>
>>> # bad: [7213910600667c51c978e577bf5454d3f7b313b7] Linux 6.6.44 # good:
>>> [58b0425ff5df680d0b67f64ae1f3f1ebdf1c4de9] Linux 6.6.43 git bisect 
>>> start '7213910600667c51c978e577bf5454d3f7b313b7'
>>> '58b0425ff5df680d0b67f64ae1f3f1ebdf1c4de9'
>>> # good: [72ff9d26964a3a80f7650df719df139f5c1f965d] arm64: dts: qcom:
>>> sm6350: Add missing qcom,non-secure-domain property git bisect good 
>>> 72ff9d26964a3a80f7650df719df139f5c1f965d
>>> # good: [0fffc2e1bf40a2220ef5a38f834ea063dba832d3] ARM: dts: sunxi:
>>> remove duplicated entries in makefile git bisect good
>>> 0fffc2e1bf40a2220ef5a38f834ea063dba832d3
>>> # bad: [8cdbe6ebfd1763a5c41a2a3058497c0a9163311c] pinctrl: renesas:
>>> r8a779g0: Fix CANFD5 suffix git bisect bad 
>>> 8cdbe6ebfd1763a5c41a2a3058497c0a9163311c
>>> # bad: [5dbb98e7fa42bebc1325899193d8f13f0705a148] drm/mediatek: Turn 
>>> off the layers with zero width or height git bisect bad
>>> 5dbb98e7fa42bebc1325899193d8f13f0705a148
>>> # bad: [691ec7043122c9c8c46d84f6e6cd85d13d50cd93] selftests/bpf: 
>>> Null checks for links in bpf_tcp_ca git bisect bad
>>> 691ec7043122c9c8c46d84f6e6cd85d13d50cd93
>>> # bad: [a1359e085d75d7393a250054e66c0a7bc6c3dbfa] perf/x86: 
>>> Serialize
>>> set_attr_rdpmc() git bisect bad
>>> a1359e085d75d7393a250054e66c0a7bc6c3dbfa
>>> # bad: [e99d9b16ff153de9540073239d24adc3b0a3a997] wifi: ath12k: 
>>> change DMA direction while mapping reinjected packets git bisect bad
>>> e99d9b16ff153de9540073239d24adc3b0a3a997
>>> # bad: [d027ac4a08541beb2a89563d3e034da7085050ba] firmware:
>>> turris-mox-rwtm: Initialize completion before mailbox git bisect bad 
>>> d027ac4a08541beb2a89563d3e034da7085050ba
>>> # bad: [e6c9eca327e6a41a81e7eba0d0ddc13da37f82a1] ARM: spitz: fix 
>>> GPIO assignment for backlight git bisect bad
>>> e6c9eca327e6a41a81e7eba0d0ddc13da37f82a1
>>> # bad: [b8cdefdaa555bbfc269c2198803f8791a8923960] m68k: cmpxchg: Fix 
>>> return value for default case in __arch_xchg() git bisect bad
>>> b8cdefdaa555bbfc269c2198803f8791a8923960
>>> # bad: [13a71384ae6a8779da809b00c6f378dcead10427] cpufreq/amd-pstate:
>>> Fix the scaling_max_freq setting on shared memory CPPC systems git 
>>> bisect bad 13a71384ae6a8779da809b00c6f378dcead10427
>>> # first bad commit: [13a71384ae6a8779da809b00c6f378dcead10427]
>>> cpufreq/amd-pstate: Fix the scaling_max_freq setting on shared 
>>> memory CPPC systems
>>>
>>> cpupower output:
>>>
>>> analyzing CPU 47:
>>>     driver: amd-pstate-epp
>>>     CPUs which run at the same hardware frequency: 47
>>>     CPUs which need to have their frequency coordinated by software: 
>>> 47
>>>     maximum transition latency:  Cannot determine or is not supported.
>>>     hardware limits: 400 MHz - 2.18 GHz
>>>     available cpufreq governors: performance powersave
>>>     current policy: frequency should be within 400 MHz and 2.18 GHz.
>>>                     The governor "performance" may decide which 
>>> speed to use
>>>                     within this range.
>>>     current CPU frequency: Unable to call hardware
>>>     current CPU frequency: 2.17 GHz (asserted by call to kernel)
>>>     boost state support:
>>>       Supported: yes
>>>       Active: yes
>>>       AMD PSTATE Highest Performance: 166. Maximum Frequency: 2.18 GHz.
>>>       AMD PSTATE Nominal Performance: 152. Nominal Frequency: 2.00 GHz.
>>>       AMD PSTATE Lowest Non-linear Performance: 115. Lowest 
>>> Non-linear Frequency: 1.51 GHz.
>>>       AMD PSTATE Lowest Performance: 31. Lowest Frequency: 400 MHz.
>>>
>>> Thanks,
>>> Morgan
>>>
>>> -----Original Message-----
>>> From: Mario Limonciello <mario.limonciello@amd.com>
>>> Sent: Tuesday, July 2, 2024 10:49 AM
>>> To: Dhananjay Ugwekar <Dhananjay.Ugwekar@amd.com>; 
>>> rafael@kernel.org; viresh.kumar@linaro.org; gautham.shenoy@amd.com; 
>>> perry.yuan@amd.com; skhan@linuxfoundation.org; li.meng@amd.com; 
>>> ray.huang@amd.com
>>> Cc: linux-pm@vger.kernel.org; linux-kernel@vger.kernel.org; David 
>>> Arcari <darcari@redhat.com>
>>> Subject: Re: [PATCH v2 2/2] cpufreq/amd-pstate: Fix the 
>>> scaling_max_freq setting on shared memory CPPC systems
>>>
>>> On 7/2/2024 3:14, Dhananjay Ugwekar wrote:
>>>> On shared memory CPPC systems, with amd_pstate=active mode, the 
>>>> change in scaling_max_freq doesn't get written to the shared memory 
>>>> region.
>>>> Due to this, the writes to the scaling_max_freq sysfs file don't 
>>>> take effect. Fix this by propagating the scaling_max_freq changes 
>>>> to the shared memory region.
>>>>
>>>> Fixes: ffa5096a7c33 ("cpufreq: amd-pstate: implement Pstate EPP 
>>>> support for the AMD processors")
>>>> Reported-by: David Arcari <darcari@redhat.com>
>>>> Signed-off-by: Dhananjay Ugwekar <Dhananjay.Ugwekar@amd.com>
>>> Reviewed-by: Mario Limonciello <mario.limonciello@amd.com>
>>>> ---
>>>>     drivers/cpufreq/amd-pstate.c | 43
>>>> +++++++++++++++++++-----------------
>>>>     1 file changed, 23 insertions(+), 20 deletions(-)
>>>>
>>>> diff --git a/drivers/cpufreq/amd-pstate.c 
>>>> b/drivers/cpufreq/amd-pstate.c index 9ad62dbe8bfb..a092b13ffbc2
>>>> 100644
>>>> --- a/drivers/cpufreq/amd-pstate.c
>>>> +++ b/drivers/cpufreq/amd-pstate.c
>>>> @@ -247,6 +247,26 @@ static int
>>>> amd_pstate_get_energy_pref_index(struct amd_cpudata *cpudata)
>>>>         return index;
>>>>     }
>>>> +static void pstate_update_perf(struct amd_cpudata *cpudata, u32
>>>> min_perf,
>>>> +                   u32 des_perf, u32 max_perf, bool fast_switch) {
>>>> +    if (fast_switch)
>>>> +        wrmsrl(MSR_AMD_CPPC_REQ, 
>>>> +READ_ONCE(cpudata->cppc_req_cached));
>>>> +    else
>>>> +        wrmsrl_on_cpu(cpudata->cpu, MSR_AMD_CPPC_REQ,
>>>> +                  READ_ONCE(cpudata->cppc_req_cached));
>>>> +}
>>>> +
>>>> +DEFINE_STATIC_CALL(amd_pstate_update_perf, pstate_update_perf);
>>>> +
>>>> +static inline void amd_pstate_update_perf(struct amd_cpudata 
>>>> +*cpudata,
>>>> +                      u32 min_perf, u32 des_perf,
>>>> +                      u32 max_perf, bool fast_switch) {
>>>> +    static_call(amd_pstate_update_perf)(cpudata, min_perf, 
>>>> +des_perf,
>>>> +                        max_perf, fast_switch); }
>>>> +
>>>>     static int amd_pstate_set_epp(struct amd_cpudata *cpudata, u32 
>>>> epp)
>>>>     {
>>>>         int ret;
>>>> @@ -263,6 +283,9 @@ static int amd_pstate_set_epp(struct 
>>>> amd_cpudata *cpudata, u32 epp)
>>>>             if (!ret)
>>>>                 cpudata->epp_cached = epp;
>>>>         } else {
>>>> +        amd_pstate_update_perf(cpudata, cpudata->min_limit_perf, 
>>>> +0U,
>>>> +                         cpudata->max_limit_perf, false);
>>>> +
>>>>             perf_ctrls.energy_perf = epp;
>>>>             ret = cppc_set_epp_perf(cpudata->cpu, &perf_ctrls, 1);
>>>>             if (ret) {
>>>> @@ -452,16 +475,6 @@ static inline int amd_pstate_init_perf(struct 
>>>> amd_cpudata *cpudata)
>>>>         return static_call(amd_pstate_init_perf)(cpudata);
>>>>     }
>>>> -static void pstate_update_perf(struct amd_cpudata *cpudata, u32 
>>>> min_perf,
>>>> -                   u32 des_perf, u32 max_perf, bool fast_switch) 
>>>> -{
>>>> -    if (fast_switch)
>>>> -        wrmsrl(MSR_AMD_CPPC_REQ, 
>>>> READ_ONCE(cpudata->cppc_req_cached));
>>>> -    else
>>>> -        wrmsrl_on_cpu(cpudata->cpu, MSR_AMD_CPPC_REQ,
>>>> -                  READ_ONCE(cpudata->cppc_req_cached));
>>>> -}
>>>> -
>>>>     static void cppc_update_perf(struct amd_cpudata *cpudata,
>>>>                      u32 min_perf, u32 des_perf,
>>>>                      u32 max_perf, bool fast_switch) @@ -475,16
>>>> +488,6 @@
>>>> static void cppc_update_perf(struct amd_cpudata *cpudata,
>>>>         cppc_set_perf(cpudata->cpu, &perf_ctrls);
>>>>     }
>>>> -DEFINE_STATIC_CALL(amd_pstate_update_perf, pstate_update_perf);
>>>> -
>>>> -static inline void amd_pstate_update_perf(struct amd_cpudata 
>>>> *cpudata,
>>>> -                      u32 min_perf, u32 des_perf,
>>>> -                      u32 max_perf, bool fast_switch) -{
>>>> -    static_call(amd_pstate_update_perf)(cpudata, min_perf, 
>>>> des_perf,
>>>> -                        max_perf, fast_switch); -}
>>>> -
>>>>     static inline bool amd_pstate_sample(struct amd_cpudata 
>>>> *cpudata)
>>>>     {
>>>>         u64 aperf, mperf, tsc;
>>>
>>>
>>
>
Jones, Morgan Sept. 3, 2024, 10:24 p.m. UTC | #9
Just to verify, this was the patch in question, correct? I had some trouble finding it on patchwork.

https://patches.linaro.org/project/linux-acpi/patch/20240227073924.3573398-1-li.meng@amd.com/

-----Original Message-----
From: Mario Limonciello <mario.limonciello@amd.com> 
Sent: Tuesday, September 3, 2024 1:52 PM
To: Jones, Morgan <Morgan.Jones@viasat.com>; Dhananjay Ugwekar <Dhananjay.Ugwekar@amd.com>; rafael@kernel.org; viresh.kumar@linaro.org; gautham.shenoy@amd.com; perry.yuan@amd.com; skhan@linuxfoundation.org; li.meng@amd.com; ray.huang@amd.com
Cc: linux-pm@vger.kernel.org; linux-kernel@vger.kernel.org; David Arcari <darcari@redhat.com>
Subject: Re: [EXTERNAL] Re: [PATCH v2 2/2] cpufreq/amd-pstate: Fix the scaling_max_freq setting on shared memory CPPC systems

Morgan,

This does remind me of a clamping issue that touches some of the same variables.  Can you please see if backporting

commit 8164f7433264 ("cpufreq: amd-pstate: adjust min/max limit perf")

to 6.6.y helps for you?

Thanks,

On 9/3/2024 15:09, Mario Limonciello wrote:
> Morgan,
> 
> OK that's great news that it's just a backport effort.  That same 
> commit also backported to 6.10.3.  Can you see if 6.10.y is affected?
> 
> Ugwekar,
> 
> Any thoughts on what else needs to come back to 6.6.y off hand?
> 
> Thanks,
> 
> On 9/3/2024 15:07, Jones, Morgan wrote:
>> Hey Mario,
>>
>> Smoking gun here, the max frequency is incorrect on 6.6.44+ but is 
>> correct on 6.11.0-rc6.
>>
>> Linux redact 6.11.0-rc6 #1-NixOS SMP PREEMPT_DYNAMIC Tue Jan  1
>> 00:00:00 UTC 1980 x86_64 GNU/Linux
>>
>> analyzing CPU 12:
>>    driver: amd-pstate-epp
>>    CPUs which run at the same hardware frequency: 12
>>    CPUs which need to have their frequency coordinated by software: 
>> 12
>>    maximum transition latency:  Cannot determine or is not supported.
>>    hardware limits: 400 MHz - 3.35 GHz
>>    available cpufreq governors: performance powersave
>>    current policy: frequency should be within 400 MHz and 3.35 GHz.
>>                    The governor "performance" may decide which speed 
>> to use
>>                    within this range.
>>    current CPU frequency: Unable to call hardware
>>    current CPU frequency: 3.34 GHz (asserted by call to kernel)
>>    boost state support:
>>      Supported: yes
>>      Active: yes
>>      AMD PSTATE Highest Performance: 255. Maximum Frequency: 3.35 GHz.
>>      AMD PSTATE Nominal Performance: 152. Nominal Frequency: 2.00 GHz.
>>      AMD PSTATE Lowest Non-linear Performance: 115. Lowest Non-linear
>> Frequency: 1.51 GHz.
>>      AMD PSTATE Lowest Performance: 31. Lowest Frequency: 400 MHz.
>>
>> We're running amd_pstate=active and amd_pstate.shared_mem=1, and our 
>> workloads are back to normal performance on 6.11.0-rc6.
>>
>> Morgan
>>
>> -----Original Message-----
>> From: Mario Limonciello <mario.limonciello@amd.com>
>> Sent: Tuesday, September 3, 2024 10:55 AM
>> To: Jones, Morgan <Morgan.Jones@viasat.com>; Dhananjay Ugwekar 
>> <Dhananjay.Ugwekar@amd.com>; rafael@kernel.org; 
>> viresh.kumar@linaro.org; gautham.shenoy@amd.com; perry.yuan@amd.com; 
>> skhan@linuxfoundation.org; li.meng@amd.com; ray.huang@amd.com
>> Cc: linux-pm@vger.kernel.org; linux-kernel@vger.kernel.org; David 
>> Arcari <darcari@redhat.com>
>> Subject: [EXTERNAL] Re: [PATCH v2 2/2] cpufreq/amd-pstate: Fix the 
>> scaling_max_freq setting on shared memory CPPC systems
>>
>> Hi Morgan,
>>
>> Can you please cross reference 6.11-rc6 to see if you're still having 
>> a problem?  I would like to understand if we're missing a backport to 
>> stable or this is still an upstream problem.
>>
>> Thanks,
>>
>> On 9/3/2024 12:51, Jones, Morgan wrote:
>>> Hi there,
>>>
>>> We are experiencing a ~35% performance regression on an AMD EPYC 
>>> 7702
>>> 64 core machine after applying this patch. We observed Linux 6.6.44 
>>> starting to cause the issue, and performed a bisect involving 
>>> rebooting the machine over 15 times. (Note that kexec didn't 
>>> successfully identify the problem, since the PM memory mailbox is 
>>> never reset).
>>>
>>> It appears that we are getting a max of 2.18 GHz when this CPU can 
>>> boost to 3.35 GHz, explaining the slowdown. Blacklisting amd-pstate 
>>> solves the issue at the expense of the performance increase we used 
>>> to get by using it.
>>>
>>> Is it possible that the upper limits were implicitly at the max CPU 
>>> power before, and setting the upper limit to something other than 
>>> the boost frequency can reduce performance now?
>>>
>>> # bad: [7213910600667c51c978e577bf5454d3f7b313b7] Linux 6.6.44 # good:
>>> [58b0425ff5df680d0b67f64ae1f3f1ebdf1c4de9] Linux 6.6.43 git bisect 
>>> start '7213910600667c51c978e577bf5454d3f7b313b7'
>>> '58b0425ff5df680d0b67f64ae1f3f1ebdf1c4de9'
>>> # good: [72ff9d26964a3a80f7650df719df139f5c1f965d] arm64: dts: qcom:
>>> sm6350: Add missing qcom,non-secure-domain property git bisect good 
>>> 72ff9d26964a3a80f7650df719df139f5c1f965d
>>> # good: [0fffc2e1bf40a2220ef5a38f834ea063dba832d3] ARM: dts: sunxi:
>>> remove duplicated entries in makefile git bisect good
>>> 0fffc2e1bf40a2220ef5a38f834ea063dba832d3
>>> # bad: [8cdbe6ebfd1763a5c41a2a3058497c0a9163311c] pinctrl: renesas:
>>> r8a779g0: Fix CANFD5 suffix git bisect bad 
>>> 8cdbe6ebfd1763a5c41a2a3058497c0a9163311c
>>> # bad: [5dbb98e7fa42bebc1325899193d8f13f0705a148] drm/mediatek: Turn 
>>> off the layers with zero width or height git bisect bad
>>> 5dbb98e7fa42bebc1325899193d8f13f0705a148
>>> # bad: [691ec7043122c9c8c46d84f6e6cd85d13d50cd93] selftests/bpf: 
>>> Null checks for links in bpf_tcp_ca git bisect bad
>>> 691ec7043122c9c8c46d84f6e6cd85d13d50cd93
>>> # bad: [a1359e085d75d7393a250054e66c0a7bc6c3dbfa] perf/x86: 
>>> Serialize
>>> set_attr_rdpmc() git bisect bad
>>> a1359e085d75d7393a250054e66c0a7bc6c3dbfa
>>> # bad: [e99d9b16ff153de9540073239d24adc3b0a3a997] wifi: ath12k: 
>>> change DMA direction while mapping reinjected packets git bisect bad
>>> e99d9b16ff153de9540073239d24adc3b0a3a997
>>> # bad: [d027ac4a08541beb2a89563d3e034da7085050ba] firmware:
>>> turris-mox-rwtm: Initialize completion before mailbox git bisect bad 
>>> d027ac4a08541beb2a89563d3e034da7085050ba
>>> # bad: [e6c9eca327e6a41a81e7eba0d0ddc13da37f82a1] ARM: spitz: fix 
>>> GPIO assignment for backlight git bisect bad
>>> e6c9eca327e6a41a81e7eba0d0ddc13da37f82a1
>>> # bad: [b8cdefdaa555bbfc269c2198803f8791a8923960] m68k: cmpxchg: Fix 
>>> return value for default case in __arch_xchg() git bisect bad
>>> b8cdefdaa555bbfc269c2198803f8791a8923960
>>> # bad: [13a71384ae6a8779da809b00c6f378dcead10427] cpufreq/amd-pstate:
>>> Fix the scaling_max_freq setting on shared memory CPPC systems git 
>>> bisect bad 13a71384ae6a8779da809b00c6f378dcead10427
>>> # first bad commit: [13a71384ae6a8779da809b00c6f378dcead10427]
>>> cpufreq/amd-pstate: Fix the scaling_max_freq setting on shared 
>>> memory CPPC systems
>>>
>>> cpupower output:
>>>
>>> analyzing CPU 47:
>>>     driver: amd-pstate-epp
>>>     CPUs which run at the same hardware frequency: 47
>>>     CPUs which need to have their frequency coordinated by software: 
>>> 47
>>>     maximum transition latency:  Cannot determine or is not supported.
>>>     hardware limits: 400 MHz - 2.18 GHz
>>>     available cpufreq governors: performance powersave
>>>     current policy: frequency should be within 400 MHz and 2.18 GHz.
>>>                     The governor "performance" may decide which 
>>> speed to use
>>>                     within this range.
>>>     current CPU frequency: Unable to call hardware
>>>     current CPU frequency: 2.17 GHz (asserted by call to kernel)
>>>     boost state support:
>>>       Supported: yes
>>>       Active: yes
>>>       AMD PSTATE Highest Performance: 166. Maximum Frequency: 2.18 GHz.
>>>       AMD PSTATE Nominal Performance: 152. Nominal Frequency: 2.00 GHz.
>>>       AMD PSTATE Lowest Non-linear Performance: 115. Lowest 
>>> Non-linear Frequency: 1.51 GHz.
>>>       AMD PSTATE Lowest Performance: 31. Lowest Frequency: 400 MHz.
>>>
>>> Thanks,
>>> Morgan
>>>
>>> -----Original Message-----
>>> From: Mario Limonciello <mario.limonciello@amd.com>
>>> Sent: Tuesday, July 2, 2024 10:49 AM
>>> To: Dhananjay Ugwekar <Dhananjay.Ugwekar@amd.com>; 
>>> rafael@kernel.org; viresh.kumar@linaro.org; gautham.shenoy@amd.com; 
>>> perry.yuan@amd.com; skhan@linuxfoundation.org; li.meng@amd.com; 
>>> ray.huang@amd.com
>>> Cc: linux-pm@vger.kernel.org; linux-kernel@vger.kernel.org; David 
>>> Arcari <darcari@redhat.com>
>>> Subject: Re: [PATCH v2 2/2] cpufreq/amd-pstate: Fix the 
>>> scaling_max_freq setting on shared memory CPPC systems
>>>
>>> On 7/2/2024 3:14, Dhananjay Ugwekar wrote:
>>>> On shared memory CPPC systems, with amd_pstate=active mode, the 
>>>> change in scaling_max_freq doesn't get written to the shared memory 
>>>> region.
>>>> Due to this, the writes to the scaling_max_freq sysfs file don't 
>>>> take effect. Fix this by propagating the scaling_max_freq changes 
>>>> to the shared memory region.
>>>>
>>>> Fixes: ffa5096a7c33 ("cpufreq: amd-pstate: implement Pstate EPP 
>>>> support for the AMD processors")
>>>> Reported-by: David Arcari <darcari@redhat.com>
>>>> Signed-off-by: Dhananjay Ugwekar <Dhananjay.Ugwekar@amd.com>
>>> Reviewed-by: Mario Limonciello <mario.limonciello@amd.com>
>>>> ---
>>>>     drivers/cpufreq/amd-pstate.c | 43
>>>> +++++++++++++++++++-----------------
>>>>     1 file changed, 23 insertions(+), 20 deletions(-)
>>>>
>>>> diff --git a/drivers/cpufreq/amd-pstate.c 
>>>> b/drivers/cpufreq/amd-pstate.c index 9ad62dbe8bfb..a092b13ffbc2
>>>> 100644
>>>> --- a/drivers/cpufreq/amd-pstate.c
>>>> +++ b/drivers/cpufreq/amd-pstate.c
>>>> @@ -247,6 +247,26 @@ static int
>>>> amd_pstate_get_energy_pref_index(struct amd_cpudata *cpudata)
>>>>         return index;
>>>>     }
>>>> +static void pstate_update_perf(struct amd_cpudata *cpudata, u32
>>>> min_perf,
>>>> +                   u32 des_perf, u32 max_perf, bool fast_switch) {
>>>> +    if (fast_switch)
>>>> +        wrmsrl(MSR_AMD_CPPC_REQ, 
>>>> +READ_ONCE(cpudata->cppc_req_cached));
>>>> +    else
>>>> +        wrmsrl_on_cpu(cpudata->cpu, MSR_AMD_CPPC_REQ,
>>>> +                  READ_ONCE(cpudata->cppc_req_cached));
>>>> +}
>>>> +
>>>> +DEFINE_STATIC_CALL(amd_pstate_update_perf, pstate_update_perf);
>>>> +
>>>> +static inline void amd_pstate_update_perf(struct amd_cpudata 
>>>> +*cpudata,
>>>> +                      u32 min_perf, u32 des_perf,
>>>> +                      u32 max_perf, bool fast_switch) {
>>>> +    static_call(amd_pstate_update_perf)(cpudata, min_perf, 
>>>> +des_perf,
>>>> +                        max_perf, fast_switch); }
>>>> +
>>>>     static int amd_pstate_set_epp(struct amd_cpudata *cpudata, u32 
>>>> epp)
>>>>     {
>>>>         int ret;
>>>> @@ -263,6 +283,9 @@ static int amd_pstate_set_epp(struct 
>>>> amd_cpudata *cpudata, u32 epp)
>>>>             if (!ret)
>>>>                 cpudata->epp_cached = epp;
>>>>         } else {
>>>> +        amd_pstate_update_perf(cpudata, cpudata->min_limit_perf, 
>>>> +0U,
>>>> +                         cpudata->max_limit_perf, false);
>>>> +
>>>>             perf_ctrls.energy_perf = epp;
>>>>             ret = cppc_set_epp_perf(cpudata->cpu, &perf_ctrls, 1);
>>>>             if (ret) {
>>>> @@ -452,16 +475,6 @@ static inline int amd_pstate_init_perf(struct 
>>>> amd_cpudata *cpudata)
>>>>         return static_call(amd_pstate_init_perf)(cpudata);
>>>>     }
>>>> -static void pstate_update_perf(struct amd_cpudata *cpudata, u32 
>>>> min_perf,
>>>> -                   u32 des_perf, u32 max_perf, bool fast_switch) 
>>>> -{
>>>> -    if (fast_switch)
>>>> -        wrmsrl(MSR_AMD_CPPC_REQ, 
>>>> READ_ONCE(cpudata->cppc_req_cached));
>>>> -    else
>>>> -        wrmsrl_on_cpu(cpudata->cpu, MSR_AMD_CPPC_REQ,
>>>> -                  READ_ONCE(cpudata->cppc_req_cached));
>>>> -}
>>>> -
>>>>     static void cppc_update_perf(struct amd_cpudata *cpudata,
>>>>                      u32 min_perf, u32 des_perf,
>>>>                      u32 max_perf, bool fast_switch) @@ -475,16
>>>> +488,6 @@
>>>> static void cppc_update_perf(struct amd_cpudata *cpudata,
>>>>         cppc_set_perf(cpudata->cpu, &perf_ctrls);
>>>>     }
>>>> -DEFINE_STATIC_CALL(amd_pstate_update_perf, pstate_update_perf);
>>>> -
>>>> -static inline void amd_pstate_update_perf(struct amd_cpudata 
>>>> *cpudata,
>>>> -                      u32 min_perf, u32 des_perf,
>>>> -                      u32 max_perf, bool fast_switch) -{
>>>> -    static_call(amd_pstate_update_perf)(cpudata, min_perf, 
>>>> des_perf,
>>>> -                        max_perf, fast_switch); -}
>>>> -
>>>>     static inline bool amd_pstate_sample(struct amd_cpudata 
>>>> *cpudata)
>>>>     {
>>>>         u64 aperf, mperf, tsc;
>>>
>>>
>>
>
Mario Limonciello Sept. 4, 2024, 1:57 p.m. UTC | #10
Morgan,

I was referring specfiically to the version that landed in Linus' tree:
https://git.kernel.org/torvalds/c/8164f7433264

But yeah it's effectively the same thing.  In any case, it's not the 
solution.

We had some internal discussion and suspect this is due to missing 
prefcore patches in 6.6 as that feature landed in 6.9.  We'll try to 
reproduce this on a Rome system and come back with our findings and 
suggestions what to do.

Thanks,

On 9/3/2024 17:24, Jones, Morgan wrote:
> Just to verify, this was the patch in question, correct? I had some trouble finding it on patchwork.
> 
> https://patches.linaro.org/project/linux-acpi/patch/20240227073924.3573398-1-li.meng@amd.com/
> 
> -----Original Message-----
> From: Mario Limonciello <mario.limonciello@amd.com>
> Sent: Tuesday, September 3, 2024 1:52 PM
> To: Jones, Morgan <Morgan.Jones@viasat.com>; Dhananjay Ugwekar <Dhananjay.Ugwekar@amd.com>; rafael@kernel.org; viresh.kumar@linaro.org; gautham.shenoy@amd.com; perry.yuan@amd.com; skhan@linuxfoundation.org; li.meng@amd.com; ray.huang@amd.com
> Cc: linux-pm@vger.kernel.org; linux-kernel@vger.kernel.org; David Arcari <darcari@redhat.com>
> Subject: Re: [EXTERNAL] Re: [PATCH v2 2/2] cpufreq/amd-pstate: Fix the scaling_max_freq setting on shared memory CPPC systems
> 
> Morgan,
> 
> This does remind me of a clamping issue that touches some of the same variables.  Can you please see if backporting
> 
> commit 8164f7433264 ("cpufreq: amd-pstate: adjust min/max limit perf")
> 
> to 6.6.y helps for you?
> 
> Thanks,
> 
> On 9/3/2024 15:09, Mario Limonciello wrote:
>> Morgan,
>>
>> OK that's great news that it's just a backport effort.  That same
>> commit also backported to 6.10.3.  Can you see if 6.10.y is affected?
>>
>> Ugwekar,
>>
>> Any thoughts on what else needs to come back to 6.6.y off hand?
>>
>> Thanks,
>>
>> On 9/3/2024 15:07, Jones, Morgan wrote:
>>> Hey Mario,
>>>
>>> Smoking gun here, the max frequency is incorrect on 6.6.44+ but is
>>> correct on 6.11.0-rc6.
>>>
>>> Linux redact 6.11.0-rc6 #1-NixOS SMP PREEMPT_DYNAMIC Tue Jan  1
>>> 00:00:00 UTC 1980 x86_64 GNU/Linux
>>>
>>> analyzing CPU 12:
>>>     driver: amd-pstate-epp
>>>     CPUs which run at the same hardware frequency: 12
>>>     CPUs which need to have their frequency coordinated by software:
>>> 12
>>>     maximum transition latency:  Cannot determine or is not supported.
>>>     hardware limits: 400 MHz - 3.35 GHz
>>>     available cpufreq governors: performance powersave
>>>     current policy: frequency should be within 400 MHz and 3.35 GHz.
>>>                     The governor "performance" may decide which speed
>>> to use
>>>                     within this range.
>>>     current CPU frequency: Unable to call hardware
>>>     current CPU frequency: 3.34 GHz (asserted by call to kernel)
>>>     boost state support:
>>>       Supported: yes
>>>       Active: yes
>>>       AMD PSTATE Highest Performance: 255. Maximum Frequency: 3.35 GHz.
>>>       AMD PSTATE Nominal Performance: 152. Nominal Frequency: 2.00 GHz.
>>>       AMD PSTATE Lowest Non-linear Performance: 115. Lowest Non-linear
>>> Frequency: 1.51 GHz.
>>>       AMD PSTATE Lowest Performance: 31. Lowest Frequency: 400 MHz.
>>>
>>> We're running amd_pstate=active and amd_pstate.shared_mem=1, and our
>>> workloads are back to normal performance on 6.11.0-rc6.
>>>
>>> Morgan
>>>
>>> -----Original Message-----
>>> From: Mario Limonciello <mario.limonciello@amd.com>
>>> Sent: Tuesday, September 3, 2024 10:55 AM
>>> To: Jones, Morgan <Morgan.Jones@viasat.com>; Dhananjay Ugwekar
>>> <Dhananjay.Ugwekar@amd.com>; rafael@kernel.org;
>>> viresh.kumar@linaro.org; gautham.shenoy@amd.com; perry.yuan@amd.com;
>>> skhan@linuxfoundation.org; li.meng@amd.com; ray.huang@amd.com
>>> Cc: linux-pm@vger.kernel.org; linux-kernel@vger.kernel.org; David
>>> Arcari <darcari@redhat.com>
>>> Subject: [EXTERNAL] Re: [PATCH v2 2/2] cpufreq/amd-pstate: Fix the
>>> scaling_max_freq setting on shared memory CPPC systems
>>>
>>> Hi Morgan,
>>>
>>> Can you please cross reference 6.11-rc6 to see if you're still having
>>> a problem?  I would like to understand if we're missing a backport to
>>> stable or this is still an upstream problem.
>>>
>>> Thanks,
>>>
>>> On 9/3/2024 12:51, Jones, Morgan wrote:
>>>> Hi there,
>>>>
>>>> We are experiencing a ~35% performance regression on an AMD EPYC
>>>> 7702
>>>> 64 core machine after applying this patch. We observed Linux 6.6.44
>>>> starting to cause the issue, and performed a bisect involving
>>>> rebooting the machine over 15 times. (Note that kexec didn't
>>>> successfully identify the problem, since the PM memory mailbox is
>>>> never reset).
>>>>
>>>> It appears that we are getting a max of 2.18 GHz when this CPU can
>>>> boost to 3.35 GHz, explaining the slowdown. Blacklisting amd-pstate
>>>> solves the issue at the expense of the performance increase we used
>>>> to get by using it.
>>>>
>>>> Is it possible that the upper limits were implicitly at the max CPU
>>>> power before, and setting the upper limit to something other than
>>>> the boost frequency can reduce performance now?
>>>>
>>>> # bad: [7213910600667c51c978e577bf5454d3f7b313b7] Linux 6.6.44 # good:
>>>> [58b0425ff5df680d0b67f64ae1f3f1ebdf1c4de9] Linux 6.6.43 git bisect
>>>> start '7213910600667c51c978e577bf5454d3f7b313b7'
>>>> '58b0425ff5df680d0b67f64ae1f3f1ebdf1c4de9'
>>>> # good: [72ff9d26964a3a80f7650df719df139f5c1f965d] arm64: dts: qcom:
>>>> sm6350: Add missing qcom,non-secure-domain property git bisect good
>>>> 72ff9d26964a3a80f7650df719df139f5c1f965d
>>>> # good: [0fffc2e1bf40a2220ef5a38f834ea063dba832d3] ARM: dts: sunxi:
>>>> remove duplicated entries in makefile git bisect good
>>>> 0fffc2e1bf40a2220ef5a38f834ea063dba832d3
>>>> # bad: [8cdbe6ebfd1763a5c41a2a3058497c0a9163311c] pinctrl: renesas:
>>>> r8a779g0: Fix CANFD5 suffix git bisect bad
>>>> 8cdbe6ebfd1763a5c41a2a3058497c0a9163311c
>>>> # bad: [5dbb98e7fa42bebc1325899193d8f13f0705a148] drm/mediatek: Turn
>>>> off the layers with zero width or height git bisect bad
>>>> 5dbb98e7fa42bebc1325899193d8f13f0705a148
>>>> # bad: [691ec7043122c9c8c46d84f6e6cd85d13d50cd93] selftests/bpf:
>>>> Null checks for links in bpf_tcp_ca git bisect bad
>>>> 691ec7043122c9c8c46d84f6e6cd85d13d50cd93
>>>> # bad: [a1359e085d75d7393a250054e66c0a7bc6c3dbfa] perf/x86:
>>>> Serialize
>>>> set_attr_rdpmc() git bisect bad
>>>> a1359e085d75d7393a250054e66c0a7bc6c3dbfa
>>>> # bad: [e99d9b16ff153de9540073239d24adc3b0a3a997] wifi: ath12k:
>>>> change DMA direction while mapping reinjected packets git bisect bad
>>>> e99d9b16ff153de9540073239d24adc3b0a3a997
>>>> # bad: [d027ac4a08541beb2a89563d3e034da7085050ba] firmware:
>>>> turris-mox-rwtm: Initialize completion before mailbox git bisect bad
>>>> d027ac4a08541beb2a89563d3e034da7085050ba
>>>> # bad: [e6c9eca327e6a41a81e7eba0d0ddc13da37f82a1] ARM: spitz: fix
>>>> GPIO assignment for backlight git bisect bad
>>>> e6c9eca327e6a41a81e7eba0d0ddc13da37f82a1
>>>> # bad: [b8cdefdaa555bbfc269c2198803f8791a8923960] m68k: cmpxchg: Fix
>>>> return value for default case in __arch_xchg() git bisect bad
>>>> b8cdefdaa555bbfc269c2198803f8791a8923960
>>>> # bad: [13a71384ae6a8779da809b00c6f378dcead10427] cpufreq/amd-pstate:
>>>> Fix the scaling_max_freq setting on shared memory CPPC systems git
>>>> bisect bad 13a71384ae6a8779da809b00c6f378dcead10427
>>>> # first bad commit: [13a71384ae6a8779da809b00c6f378dcead10427]
>>>> cpufreq/amd-pstate: Fix the scaling_max_freq setting on shared
>>>> memory CPPC systems
>>>>
>>>> cpupower output:
>>>>
>>>> analyzing CPU 47:
>>>>      driver: amd-pstate-epp
>>>>      CPUs which run at the same hardware frequency: 47
>>>>      CPUs which need to have their frequency coordinated by software:
>>>> 47
>>>>      maximum transition latency:  Cannot determine or is not supported.
>>>>      hardware limits: 400 MHz - 2.18 GHz
>>>>      available cpufreq governors: performance powersave
>>>>      current policy: frequency should be within 400 MHz and 2.18 GHz.
>>>>                      The governor "performance" may decide which
>>>> speed to use
>>>>                      within this range.
>>>>      current CPU frequency: Unable to call hardware
>>>>      current CPU frequency: 2.17 GHz (asserted by call to kernel)
>>>>      boost state support:
>>>>        Supported: yes
>>>>        Active: yes
>>>>        AMD PSTATE Highest Performance: 166. Maximum Frequency: 2.18 GHz.
>>>>        AMD PSTATE Nominal Performance: 152. Nominal Frequency: 2.00 GHz.
>>>>        AMD PSTATE Lowest Non-linear Performance: 115. Lowest
>>>> Non-linear Frequency: 1.51 GHz.
>>>>        AMD PSTATE Lowest Performance: 31. Lowest Frequency: 400 MHz.
>>>>
>>>> Thanks,
>>>> Morgan
>>>>
>>>> -----Original Message-----
>>>> From: Mario Limonciello <mario.limonciello@amd.com>
>>>> Sent: Tuesday, July 2, 2024 10:49 AM
>>>> To: Dhananjay Ugwekar <Dhananjay.Ugwekar@amd.com>;
>>>> rafael@kernel.org; viresh.kumar@linaro.org; gautham.shenoy@amd.com;
>>>> perry.yuan@amd.com; skhan@linuxfoundation.org; li.meng@amd.com;
>>>> ray.huang@amd.com
>>>> Cc: linux-pm@vger.kernel.org; linux-kernel@vger.kernel.org; David
>>>> Arcari <darcari@redhat.com>
>>>> Subject: Re: [PATCH v2 2/2] cpufreq/amd-pstate: Fix the
>>>> scaling_max_freq setting on shared memory CPPC systems
>>>>
>>>> On 7/2/2024 3:14, Dhananjay Ugwekar wrote:
>>>>> On shared memory CPPC systems, with amd_pstate=active mode, the
>>>>> change in scaling_max_freq doesn't get written to the shared memory
>>>>> region.
>>>>> Due to this, the writes to the scaling_max_freq sysfs file don't
>>>>> take effect. Fix this by propagating the scaling_max_freq changes
>>>>> to the shared memory region.
>>>>>
>>>>> Fixes: ffa5096a7c33 ("cpufreq: amd-pstate: implement Pstate EPP
>>>>> support for the AMD processors")
>>>>> Reported-by: David Arcari <darcari@redhat.com>
>>>>> Signed-off-by: Dhananjay Ugwekar <Dhananjay.Ugwekar@amd.com>
>>>> Reviewed-by: Mario Limonciello <mario.limonciello@amd.com>
>>>>> ---
>>>>>      drivers/cpufreq/amd-pstate.c | 43
>>>>> +++++++++++++++++++-----------------
>>>>>      1 file changed, 23 insertions(+), 20 deletions(-)
>>>>>
>>>>> diff --git a/drivers/cpufreq/amd-pstate.c
>>>>> b/drivers/cpufreq/amd-pstate.c index 9ad62dbe8bfb..a092b13ffbc2
>>>>> 100644
>>>>> --- a/drivers/cpufreq/amd-pstate.c
>>>>> +++ b/drivers/cpufreq/amd-pstate.c
>>>>> @@ -247,6 +247,26 @@ static int
>>>>> amd_pstate_get_energy_pref_index(struct amd_cpudata *cpudata)
>>>>>          return index;
>>>>>      }
>>>>> +static void pstate_update_perf(struct amd_cpudata *cpudata, u32
>>>>> min_perf,
>>>>> +                   u32 des_perf, u32 max_perf, bool fast_switch) {
>>>>> +    if (fast_switch)
>>>>> +        wrmsrl(MSR_AMD_CPPC_REQ,
>>>>> +READ_ONCE(cpudata->cppc_req_cached));
>>>>> +    else
>>>>> +        wrmsrl_on_cpu(cpudata->cpu, MSR_AMD_CPPC_REQ,
>>>>> +                  READ_ONCE(cpudata->cppc_req_cached));
>>>>> +}
>>>>> +
>>>>> +DEFINE_STATIC_CALL(amd_pstate_update_perf, pstate_update_perf);
>>>>> +
>>>>> +static inline void amd_pstate_update_perf(struct amd_cpudata
>>>>> +*cpudata,
>>>>> +                      u32 min_perf, u32 des_perf,
>>>>> +                      u32 max_perf, bool fast_switch) {
>>>>> +    static_call(amd_pstate_update_perf)(cpudata, min_perf,
>>>>> +des_perf,
>>>>> +                        max_perf, fast_switch); }
>>>>> +
>>>>>      static int amd_pstate_set_epp(struct amd_cpudata *cpudata, u32
>>>>> epp)
>>>>>      {
>>>>>          int ret;
>>>>> @@ -263,6 +283,9 @@ static int amd_pstate_set_epp(struct
>>>>> amd_cpudata *cpudata, u32 epp)
>>>>>              if (!ret)
>>>>>                  cpudata->epp_cached = epp;
>>>>>          } else {
>>>>> +        amd_pstate_update_perf(cpudata, cpudata->min_limit_perf,
>>>>> +0U,
>>>>> +                         cpudata->max_limit_perf, false);
>>>>> +
>>>>>              perf_ctrls.energy_perf = epp;
>>>>>              ret = cppc_set_epp_perf(cpudata->cpu, &perf_ctrls, 1);
>>>>>              if (ret) {
>>>>> @@ -452,16 +475,6 @@ static inline int amd_pstate_init_perf(struct
>>>>> amd_cpudata *cpudata)
>>>>>          return static_call(amd_pstate_init_perf)(cpudata);
>>>>>      }
>>>>> -static void pstate_update_perf(struct amd_cpudata *cpudata, u32
>>>>> min_perf,
>>>>> -                   u32 des_perf, u32 max_perf, bool fast_switch)
>>>>> -{
>>>>> -    if (fast_switch)
>>>>> -        wrmsrl(MSR_AMD_CPPC_REQ,
>>>>> READ_ONCE(cpudata->cppc_req_cached));
>>>>> -    else
>>>>> -        wrmsrl_on_cpu(cpudata->cpu, MSR_AMD_CPPC_REQ,
>>>>> -                  READ_ONCE(cpudata->cppc_req_cached));
>>>>> -}
>>>>> -
>>>>>      static void cppc_update_perf(struct amd_cpudata *cpudata,
>>>>>                       u32 min_perf, u32 des_perf,
>>>>>                       u32 max_perf, bool fast_switch) @@ -475,16
>>>>> +488,6 @@
>>>>> static void cppc_update_perf(struct amd_cpudata *cpudata,
>>>>>          cppc_set_perf(cpudata->cpu, &perf_ctrls);
>>>>>      }
>>>>> -DEFINE_STATIC_CALL(amd_pstate_update_perf, pstate_update_perf);
>>>>> -
>>>>> -static inline void amd_pstate_update_perf(struct amd_cpudata
>>>>> *cpudata,
>>>>> -                      u32 min_perf, u32 des_perf,
>>>>> -                      u32 max_perf, bool fast_switch) -{
>>>>> -    static_call(amd_pstate_update_perf)(cpudata, min_perf,
>>>>> des_perf,
>>>>> -                        max_perf, fast_switch); -}
>>>>> -
>>>>>      static inline bool amd_pstate_sample(struct amd_cpudata
>>>>> *cpudata)
>>>>>      {
>>>>>          u64 aperf, mperf, tsc;
>>>>
>>>>
>>>
>>
>
Jones, Morgan Sept. 5, 2024, 9:09 p.m. UTC | #11
Mario,

Confirmed. Thank you for the help! Slightly different refs on my end:

Remotes:

next    https://git.kernel.org/pub/scm/linux/kernel/git/next/linux-next.git (fetch)
next    https://git.kernel.org/pub/scm/linux/kernel/git/next/linux-next.git (push)
origin  git://git.kernel.org/pub/scm/linux/kernel/git/stable/linux.git (fetch)
origin  git://git.kernel.org/pub/scm/linux/kernel/git/stable/linux.git (push)
superm1 https://git.kernel.org/pub/scm/linux/kernel/git/superm1/linux.git/ (fetch)
superm1 https://git.kernel.org/pub/scm/linux/kernel/git/superm1/linux.git/ (push)
torvalds        git://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git (fetch)
torvalds        git://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git (push)

Patches:

git format-patch 12753d71e8c5^..12753d71e8c5
git format-patch f3a052391822b772b4e27f2594526cf1eb103cab^..f3a052391822b772b4e27f2594526cf1eb103cab
git format-patch bf202e654bfa57fb8cf9d93d4c6855890b70b9c4^..bf202e654bfa57fb8cf9d93d4c6855890b70b9c4

Results:

Linux redact 6.6.48 #1-NixOS SMP PREEMPT_DYNAMIC Tue Jan  1 00:00:00 UTC 1980 x86_64 GNU/Linux

analyzing CPU 56:
  driver: amd-pstate-epp
  CPUs which run at the same hardware frequency: 56
  CPUs which need to have their frequency coordinated by software: 56
  maximum transition latency:  Cannot determine or is not supported.
  hardware limits: 400 MHz - 3.35 GHz
  available cpufreq governors: performance powersave
  current policy: frequency should be within 400 MHz and 3.35 GHz.
                  The governor "performance" may decide which speed to use
                  within this range.
  current CPU frequency: Unable to call hardware
  current CPU frequency: 2.09 GHz (asserted by call to kernel)
  boost state support:
    Supported: yes
    Active: yes
    AMD PSTATE Highest Performance: 255. Maximum Frequency: 3.35 GHz.
    AMD PSTATE Nominal Performance: 152. Nominal Frequency: 2.00 GHz.
    AMD PSTATE Lowest Non-linear Performance: 115. Lowest Non-linear Frequency: 1.51 GHz.
    AMD PSTATE Lowest Performance: 31. Lowest Frequency: 400 MHz.

And our builds are back to being fast with `amd_pstate=active amd_prefcore=enable amd_pstate.shared_mem=1`.

Morgan

-----Original Message-----
From: Mario Limonciello <mario.limonciello@amd.com> 
Sent: Thursday, September 5, 2024 8:12 AM
To: Jones, Morgan <Morgan.Jones@viasat.com>
Cc: linux-pm@vger.kernel.org; linux-kernel@vger.kernel.org; David Arcari <darcari@redhat.com>; Dhananjay Ugwekar <Dhananjay.Ugwekar@amd.com>; rafael@kernel.org; viresh.kumar@linaro.org; gautham.shenoy@amd.com; perry.yuan@amd.com; skhan@linuxfoundation.org; li.meng@amd.com; ray.huang@amd.com
Subject: Re: [EXTERNAL] Re: [PATCH v2 2/2] cpufreq/amd-pstate: Fix the scaling_max_freq setting on shared memory CPPC systems

Hi Morgan,

Please apply these 3 commits:

commit 12753d71e8c5 ("ACPI: CPPC: Add helper to get the highest performance value") commit ed429c686b79 ("cpufreq: amd-pstate: Enable amd-pstate preferred core support") commit 3d291fe47fe1 ("cpufreq: amd-pstate: fix the highest frequency issue which limits performance")

The first two should help your system, the third will prevent introducing a regression on a different one.

Assuming that works we should ask @stable to pull all 3 in to fix this regression.

Thanks,

On 9/4/2024 08:57, Mario Limonciello wrote:
> Morgan,
> 
> I was referring specfiically to the version that landed in Linus' tree:
> https://urldefense.us/v3/__https://git.kernel.org/torvalds/c/8164f7433
> 264__;!!C5Asm8uRnZQmlRln!aIZEDEbIUKD7OrxN0b0KjoqKYDL2yMkwk4EK7x_oSnyHQ
> 6MEq7yt6JHjd0TD9DgEYEWDcF58OKL8c7G11bT3dSqL8eM$
> 
> But yeah it's effectively the same thing.  In any case, it's not the 
> solution.
> 
> We had some internal discussion and suspect this is due to missing 
> prefcore patches in 6.6 as that feature landed in 6.9.  We'll try to 
> reproduce this on a Rome system and come back with our findings and 
> suggestions what to do.
> 
> Thanks,
>
Mario Limonciello Sept. 5, 2024, 9:14 p.m. UTC | #12
+ stable
+ regressions
New subject

Great news.

Greg, Sasha,

Can you please pull in these 3 commits specifically to 6.6.y to fix a 
regression that was reported by Morgan in 6.6.y:

commit 12753d71e8c5 ("ACPI: CPPC: Add helper to get the highest 
performance value")
commit ed429c686b79 ("cpufreq: amd-pstate: Enable amd-pstate preferred 
core support")
commit 3d291fe47fe1 ("cpufreq: amd-pstate: fix the highest frequency 
issue which limits performance")

Further details are below.

Thanks!

On 9/5/2024 16:09, Jones, Morgan wrote:
> Mario,
> 
> Confirmed. Thank you for the help! Slightly different refs on my end:
> 
> Remotes:
> 
> next    https://git.kernel.org/pub/scm/linux/kernel/git/next/linux-next.git (fetch)
> next    https://git.kernel.org/pub/scm/linux/kernel/git/next/linux-next.git (push)
> origin  git://git.kernel.org/pub/scm/linux/kernel/git/stable/linux.git (fetch)
> origin  git://git.kernel.org/pub/scm/linux/kernel/git/stable/linux.git (push)
> superm1 https://git.kernel.org/pub/scm/linux/kernel/git/superm1/linux.git/ (fetch)
> superm1 https://git.kernel.org/pub/scm/linux/kernel/git/superm1/linux.git/ (push)
> torvalds        git://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git (fetch)
> torvalds        git://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git (push)
> 
> Patches:
> 
> git format-patch 12753d71e8c5^..12753d71e8c5
> git format-patch f3a052391822b772b4e27f2594526cf1eb103cab^..f3a052391822b772b4e27f2594526cf1eb103cab
> git format-patch bf202e654bfa57fb8cf9d93d4c6855890b70b9c4^..bf202e654bfa57fb8cf9d93d4c6855890b70b9c4
> 
> Results:
> 
> Linux redact 6.6.48 #1-NixOS SMP PREEMPT_DYNAMIC Tue Jan  1 00:00:00 UTC 1980 x86_64 GNU/Linux
> 
> analyzing CPU 56:
>    driver: amd-pstate-epp
>    CPUs which run at the same hardware frequency: 56
>    CPUs which need to have their frequency coordinated by software: 56
>    maximum transition latency:  Cannot determine or is not supported.
>    hardware limits: 400 MHz - 3.35 GHz
>    available cpufreq governors: performance powersave
>    current policy: frequency should be within 400 MHz and 3.35 GHz.
>                    The governor "performance" may decide which speed to use
>                    within this range.
>    current CPU frequency: Unable to call hardware
>    current CPU frequency: 2.09 GHz (asserted by call to kernel)
>    boost state support:
>      Supported: yes
>      Active: yes
>      AMD PSTATE Highest Performance: 255. Maximum Frequency: 3.35 GHz.
>      AMD PSTATE Nominal Performance: 152. Nominal Frequency: 2.00 GHz.
>      AMD PSTATE Lowest Non-linear Performance: 115. Lowest Non-linear Frequency: 1.51 GHz.
>      AMD PSTATE Lowest Performance: 31. Lowest Frequency: 400 MHz.
> 
> And our builds are back to being fast with `amd_pstate=active amd_prefcore=enable amd_pstate.shared_mem=1`.
> 
> Morgan
> 
> -----Original Message-----
> From: Mario Limonciello <mario.limonciello@amd.com>
> Sent: Thursday, September 5, 2024 8:12 AM
> To: Jones, Morgan <Morgan.Jones@viasat.com>
> Cc: linux-pm@vger.kernel.org; linux-kernel@vger.kernel.org; David Arcari <darcari@redhat.com>; Dhananjay Ugwekar <Dhananjay.Ugwekar@amd.com>; rafael@kernel.org; viresh.kumar@linaro.org; gautham.shenoy@amd.com; perry.yuan@amd.com; skhan@linuxfoundation.org; li.meng@amd.com; ray.huang@amd.com
> Subject: Re: [EXTERNAL] Re: [PATCH v2 2/2] cpufreq/amd-pstate: Fix the scaling_max_freq setting on shared memory CPPC systems
> 
> Hi Morgan,
> 
> Please apply these 3 commits:
> 
> commit 12753d71e8c5 ("ACPI: CPPC: Add helper to get the highest performance value") commit ed429c686b79 ("cpufreq: amd-pstate: Enable amd-pstate preferred core support") commit 3d291fe47fe1 ("cpufreq: amd-pstate: fix the highest frequency issue which limits performance")
> 
> The first two should help your system, the third will prevent introducing a regression on a different one.
> 
> Assuming that works we should ask @stable to pull all 3 in to fix this regression.
> 
> Thanks,
> 
> On 9/4/2024 08:57, Mario Limonciello wrote:
>> Morgan,
>>
>> I was referring specfiically to the version that landed in Linus' tree:
>> https://urldefense.us/v3/__https://git.kernel.org/torvalds/c/8164f7433
>> 264__;!!C5Asm8uRnZQmlRln!aIZEDEbIUKD7OrxN0b0KjoqKYDL2yMkwk4EK7x_oSnyHQ
>> 6MEq7yt6JHjd0TD9DgEYEWDcF58OKL8c7G11bT3dSqL8eM$
>>
>> But yeah it's effectively the same thing.  In any case, it's not the
>> solution.
>>
>> We had some internal discussion and suspect this is due to missing
>> prefcore patches in 6.6 as that feature landed in 6.9.  We'll try to
>> reproduce this on a Rome system and come back with our findings and
>> suggestions what to do.
>>
>> Thanks,
>>
>
Greg Kroah-Hartman Sept. 8, 2024, 2:05 p.m. UTC | #13
On Thu, Sep 05, 2024 at 04:14:26PM -0500, Mario Limonciello wrote:
> + stable
> + regressions
> New subject
> 
> Great news.
> 
> Greg, Sasha,
> 
> Can you please pull in these 3 commits specifically to 6.6.y to fix a
> regression that was reported by Morgan in 6.6.y:
> 
> commit 12753d71e8c5 ("ACPI: CPPC: Add helper to get the highest performance
> value")

This is fine, but:

> commit ed429c686b79 ("cpufreq: amd-pstate: Enable amd-pstate preferred core
> support")

This is not a valid git id in Linus's tree :(

> commit 3d291fe47fe1 ("cpufreq: amd-pstate: fix the highest frequency issue
> which limits performance")

And neither is this :(

So perhaps you got them wrong?

thanks,

greg k-h
Christian Heusel Sept. 8, 2024, 2:12 p.m. UTC | #14
Hey Greg,

On 24/09/08 04:05PM, Greg Kroah-Hartman wrote:
> On Thu, Sep 05, 2024 at 04:14:26PM -0500, Mario Limonciello wrote:
> > + stable
> > + regressions
> > New subject
> > 
> > Great news.
> > 
> > Greg, Sasha,
> > 
> > Can you please pull in these 3 commits specifically to 6.6.y to fix a
> > regression that was reported by Morgan in 6.6.y:
> > 
> > commit 12753d71e8c5 ("ACPI: CPPC: Add helper to get the highest performance
> > value")
> 
> This is fine, but:
> 
> > commit ed429c686b79 ("cpufreq: amd-pstate: Enable amd-pstate preferred core
> > support")
> 
> This is not a valid git id in Linus's tree :(

f3a052391822 ("cpufreq: amd-pstate: Enable amd-pstate preferred core support")

> 
> > commit 3d291fe47fe1 ("cpufreq: amd-pstate: fix the highest frequency issue
> > which limits performance")
> 
> And neither is this :(

bf202e654bfa ("cpufreq: amd-pstate: fix the highest frequency issue which limits performance")

> So perhaps you got them wrong?

I have added the ID's of the matching commits from Linus' tree above! :)

> thanks,
> 
> greg k-h

Cheers,
Chris
Greg Kroah-Hartman Sept. 8, 2024, 2:29 p.m. UTC | #15
On Sun, Sep 08, 2024 at 04:12:28PM +0200, Christian Heusel wrote:
> Hey Greg,
> 
> On 24/09/08 04:05PM, Greg Kroah-Hartman wrote:
> > On Thu, Sep 05, 2024 at 04:14:26PM -0500, Mario Limonciello wrote:
> > > + stable
> > > + regressions
> > > New subject
> > > 
> > > Great news.
> > > 
> > > Greg, Sasha,
> > > 
> > > Can you please pull in these 3 commits specifically to 6.6.y to fix a
> > > regression that was reported by Morgan in 6.6.y:
> > > 
> > > commit 12753d71e8c5 ("ACPI: CPPC: Add helper to get the highest performance
> > > value")
> > 
> > This is fine, but:
> > 
> > > commit ed429c686b79 ("cpufreq: amd-pstate: Enable amd-pstate preferred core
> > > support")
> > 
> > This is not a valid git id in Linus's tree :(
> 
> f3a052391822 ("cpufreq: amd-pstate: Enable amd-pstate preferred core support")
> 
> > 
> > > commit 3d291fe47fe1 ("cpufreq: amd-pstate: fix the highest frequency issue
> > > which limits performance")
> > 
> > And neither is this :(
> 
> bf202e654bfa ("cpufreq: amd-pstate: fix the highest frequency issue which limits performance")
> 
> > So perhaps you got them wrong?
> 
> I have added the ID's of the matching commits from Linus' tree above! :)
> 

Thanks, that works, all now queued up.

greg k-h
diff mbox series

Patch

diff --git a/drivers/cpufreq/amd-pstate.c b/drivers/cpufreq/amd-pstate.c
index 9ad62dbe8bfb..a092b13ffbc2 100644
--- a/drivers/cpufreq/amd-pstate.c
+++ b/drivers/cpufreq/amd-pstate.c
@@ -247,6 +247,26 @@  static int amd_pstate_get_energy_pref_index(struct amd_cpudata *cpudata)
 	return index;
 }
 
+static void pstate_update_perf(struct amd_cpudata *cpudata, u32 min_perf,
+			       u32 des_perf, u32 max_perf, bool fast_switch)
+{
+	if (fast_switch)
+		wrmsrl(MSR_AMD_CPPC_REQ, READ_ONCE(cpudata->cppc_req_cached));
+	else
+		wrmsrl_on_cpu(cpudata->cpu, MSR_AMD_CPPC_REQ,
+			      READ_ONCE(cpudata->cppc_req_cached));
+}
+
+DEFINE_STATIC_CALL(amd_pstate_update_perf, pstate_update_perf);
+
+static inline void amd_pstate_update_perf(struct amd_cpudata *cpudata,
+					  u32 min_perf, u32 des_perf,
+					  u32 max_perf, bool fast_switch)
+{
+	static_call(amd_pstate_update_perf)(cpudata, min_perf, des_perf,
+					    max_perf, fast_switch);
+}
+
 static int amd_pstate_set_epp(struct amd_cpudata *cpudata, u32 epp)
 {
 	int ret;
@@ -263,6 +283,9 @@  static int amd_pstate_set_epp(struct amd_cpudata *cpudata, u32 epp)
 		if (!ret)
 			cpudata->epp_cached = epp;
 	} else {
+		amd_pstate_update_perf(cpudata, cpudata->min_limit_perf, 0U,
+					     cpudata->max_limit_perf, false);
+
 		perf_ctrls.energy_perf = epp;
 		ret = cppc_set_epp_perf(cpudata->cpu, &perf_ctrls, 1);
 		if (ret) {
@@ -452,16 +475,6 @@  static inline int amd_pstate_init_perf(struct amd_cpudata *cpudata)
 	return static_call(amd_pstate_init_perf)(cpudata);
 }
 
-static void pstate_update_perf(struct amd_cpudata *cpudata, u32 min_perf,
-			       u32 des_perf, u32 max_perf, bool fast_switch)
-{
-	if (fast_switch)
-		wrmsrl(MSR_AMD_CPPC_REQ, READ_ONCE(cpudata->cppc_req_cached));
-	else
-		wrmsrl_on_cpu(cpudata->cpu, MSR_AMD_CPPC_REQ,
-			      READ_ONCE(cpudata->cppc_req_cached));
-}
-
 static void cppc_update_perf(struct amd_cpudata *cpudata,
 			     u32 min_perf, u32 des_perf,
 			     u32 max_perf, bool fast_switch)
@@ -475,16 +488,6 @@  static void cppc_update_perf(struct amd_cpudata *cpudata,
 	cppc_set_perf(cpudata->cpu, &perf_ctrls);
 }
 
-DEFINE_STATIC_CALL(amd_pstate_update_perf, pstate_update_perf);
-
-static inline void amd_pstate_update_perf(struct amd_cpudata *cpudata,
-					  u32 min_perf, u32 des_perf,
-					  u32 max_perf, bool fast_switch)
-{
-	static_call(amd_pstate_update_perf)(cpudata, min_perf, des_perf,
-					    max_perf, fast_switch);
-}
-
 static inline bool amd_pstate_sample(struct amd_cpudata *cpudata)
 {
 	u64 aperf, mperf, tsc;