Re: [EXTERNAL] Re: [PATCH v2 2/2] cpufreq/amd-pstate: Fix the scaling_max_freq setting on shared memory CPPC systems
From: Mario Limonciello
Date: Wed Sep 04 2024 - 09:58:49 EST
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@xxxxxxx/
-----Original Message-----
From: Mario Limonciello <mario.limonciello@xxxxxxx>
Sent: Tuesday, September 3, 2024 1:52 PM
To: Jones, Morgan <Morgan.Jones@xxxxxxxxxx>; Dhananjay Ugwekar <Dhananjay.Ugwekar@xxxxxxx>; rafael@xxxxxxxxxx; viresh.kumar@xxxxxxxxxx; gautham.shenoy@xxxxxxx; perry.yuan@xxxxxxx; skhan@xxxxxxxxxxxxxxxxxxx; li.meng@xxxxxxx; ray.huang@xxxxxxx
Cc: linux-pm@xxxxxxxxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx; David Arcari <darcari@xxxxxxxxxx>
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@xxxxxxx>
Sent: Tuesday, September 3, 2024 10:55 AM
To: Jones, Morgan <Morgan.Jones@xxxxxxxxxx>; Dhananjay Ugwekar
<Dhananjay.Ugwekar@xxxxxxx>; rafael@xxxxxxxxxx;
viresh.kumar@xxxxxxxxxx; gautham.shenoy@xxxxxxx; perry.yuan@xxxxxxx;
skhan@xxxxxxxxxxxxxxxxxxx; li.meng@xxxxxxx; ray.huang@xxxxxxx
Cc: linux-pm@xxxxxxxxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx; David
Arcari <darcari@xxxxxxxxxx>
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@xxxxxxx>
Sent: Tuesday, July 2, 2024 10:49 AM
To: Dhananjay Ugwekar <Dhananjay.Ugwekar@xxxxxxx>;
rafael@xxxxxxxxxx; viresh.kumar@xxxxxxxxxx; gautham.shenoy@xxxxxxx;
perry.yuan@xxxxxxx; skhan@xxxxxxxxxxxxxxxxxxx; li.meng@xxxxxxx;
ray.huang@xxxxxxx
Cc: linux-pm@xxxxxxxxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx; David
Arcari <darcari@xxxxxxxxxx>
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@xxxxxxxxxx>
Signed-off-by: Dhananjay Ugwekar <Dhananjay.Ugwekar@xxxxxxx>
Reviewed-by: Mario Limonciello <mario.limonciello@xxxxxxx>
---
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;