[Bug 219640] New: Preferred cores working incorrectly for Zen3 CPU (regression)

[bugzilla-daemon@kernel.org]

https://bugzilla.kernel.org/show_bug.cgi?id=219640

            Bug ID: 219640
           Summary: Preferred cores working incorrectly for Zen3 CPU
                    (regression)
           Product: Power Management
           Version: 2.5
          Hardware: AMD
                OS: Linux
            Status: NEW
          Severity: normal
          Priority: P3
         Component: cpufreq
          Assignee: linux-pm@vger.kernel.org
          Reporter: sobrus@gmail.com
        Regression: No

Created attachment 307415
  --> https://bugzilla.kernel.org/attachment.cgi?id=307415&action=edit
Outputs comparing affected and unaffected kernels

Hi Kernel Team,

I've noticed a preferred cores regression while using most recent CachyOS
kernels on Zen3 platform with amd_pstate_epp driver (5950x, amd_pstate=active).
Regression is clearly visible in ryzen_monitor when running low thread
workloads - instead of sticking to best cores, it chooses the cores randomly
(?) for each subsequent workload run.

CachyOS team developer has bissected the issue to commit
39311a230e04eab2fe7e257ad79922040bfdaf1c

Original CachyOS forum thread:
https://discuss.cachyos.org/t/preferred-cores-stopped-working-for-zen3-cpu/5108/6

Steps to reproduce:
- use ryzen_monitor or some kind of system monitor to see which cores are being
utilized
- run “stress - c 2” or other software causing low-thread workload.
- observe core utilization, frequency and consumed power

Below I'm attaching detailed commit info, output of cpupower frequency-info,
output of amd-pstate-triage.py and output of ryzen_monitor while running
"stress -c 4" for both affected and unaffected kernels.

The differences I've noticed are:
- affected kernel has ITMT:     Y instead of 1
- affected kernel highest performance reported by cpupower frequency-info is
166 instead of 196.

Hardware is Ryzen 5950x, undervolted using per-core PBO and voltage offset,
limited to 4750Mhz and 80C, AGESA 1.2.0.7, chipset B550.

Commit info:

commit 39311a230e04eab2fe7e257ad79922040bfdaf1c
Author: Mario Limonciello <mario.limonciello@amd.com>
Date:   Mon Dec 9 12:52:34 2024 -0600

    cpufreq/amd-pstate: Store the boost numerator as highest perf again

    commit ad4caad58d91d ("cpufreq: amd-pstate: Merge
    amd_pstate_highest_perf_set() into amd_get_boost_ratio_numerator()")
    changed the semantics for highest perf and commit 18d9b52271213
    ("cpufreq/amd-pstate: Use nominal perf for limits when boost is disabled")
    worked around those semantic changes.

    This however is a confusing result and furthermore makes it awkward to
    change frequency limits and boost due to the scaling differences. Restore
    the boost numerator to highest perf again.

    Suggested-by: Dhananjay Ugwekar <Dhananjay.Ugwekar@amd.com>
    Reviewed-by: Gautham R. Shenoy <gautham.shenoy@amd.com>
    Fixes: ad4caad58d91 ("cpufreq: amd-pstate: Merge
amd_pstate_highest_perf_set() into amd_get_boost_ratio_numerator()")
    Link:
https://lore.kernel.org/r/20241209185248.16301-2-mario.limonciello@amd.com
    Signed-off-by: Mario Limonciello <mario.limonciello@amd.com>

 Documentation/admin-guide/pm/amd-pstate.rst |  4 +---
 drivers/cpufreq/amd-pstate.c                | 25 +++++++++++++++----------
 2 files changed, 16 insertions(+), 13 deletions(-)

-- 
You may reply to this email to add a comment.

You are receiving this mail because:
You are the assignee for the bug.