Linux Perf tool does not work as expected in Linux ARM runners

[visheshruparelia]

Description

In Linux ARM based runners, the perf record command does not collect any samples.
I have set the kernel.perf_event_paranoid=1 and kernel.kptr_restrict=0. Upon running the command: sudo perf record -g -e cycles -a -o perf.data sleep 5 it collects no samples in case of Linux ARM runners.
The same command works just fine in Linux x64 runners.

This can be consistently reproduced and reproducer can be found here: https://github.com/visheshruparelia/gh-linux-arm-perf-tool-issue

Platforms affected

• Azure DevOps
• GitHub Actions - Standard Runners
• GitHub Actions - Larger Runners

Runner images affected

• Ubuntu 20.04
• Ubuntu 22.04
• Ubuntu 24.04
• macOS 13
• macOS 13 Arm64
• macOS 14
• macOS 14 Arm64
• macOS 15
• macOS 15 Arm64
• Windows Server 2019
• Windows Server 2022
• Windows Server 2025

Image version and build link

Please check the reproducer build here: https://github.com/visheshruparelia/gh-linux-arm-perf-tool-issue/actions/runs/13587925274

When you compare the output of the step named View perf output on both the different runners, you can clearly see that it works as expected in x64 but does not collect any data in ARM.

Is it regression?

Yes - this is not seen any Linux x64 machines.

Expected behavior

The Linux ARM machine should also produce output similar to x64 machine in the last step of the job (name: View perf output).

Actual behavior

The ARM machine does not collect any samples and thus the output is empty.

Repro steps

Run the following steps on Linux ARM runner and the last step will give an empty output.

    steps:
      - name: Install perf
        run: sudo apt-get install linux-tools-generic
      - name: Enable perf event paranoid flag
        run: sudo sysctl kernel.perf_event_paranoid=1
      - name: Set Kernel KPTR  Restriction
        run: sudo sysctl kernel.kptr_restrict=0
      - name: Run perf command with -a
        run:  sudo perf record -g -e cycles -a -o perf.data sleep 5
      - name: Run ls
        run: ls -la
      - name: View perf output
        run: sudo perf script -i perf.data

I have created a simple reproducer for your convenience: https://github.com/visheshruparelia/gh-linux-arm-perf-tool-issue/blob/main/.github/workflows/perf-tool-issue.yml

[RaviAkshintala]

Hi @visheshruparelia Thank you for bringing this issue to our attention. We are looking into this issue and will update you on this issue after investigating.

[RaviAkshintala]

Hi @visheshruparelia Could you please try with the below workflow for the Linux ARM runner and It is displaying the output.Thanks.

https://github.com/RaviAkshintala/gh-linux-arm-perf-tool-issue-test/actions/runs/13673462959/workflow#L9

FYI:

https://github.com/RaviAkshintala/gh-linux-arm-perf-tool-issue-test/actions/runs/13673462959/job/38228630946

[visheshruparelia]

Thank you @RaviAkshintala for looking into this. May I ask why cpu-clock and not cycles? I am specifically interested in the cycles event.

[RaviAkshintala]

@visheshruparelia Some ARM architectures might not support the same cycles event name or might require a different events like instr or cpu-clock. Thanks.

[visheshruparelia]

@RaviAkshintala when I run sudo perf list on the ARM runner, I do see cycles as the listed event. Sample run: https://github.com/visheshruparelia/async-profiler/actions/runs/13548483543/job/37865980407

[apangin]

@RaviAkshintala If the hardware did not support the given event, perf would fail with an error. The problem is, perf runs silently as if everything works fine, and it write an output file, but no samples are written. The same happens with all PERF_TYPE_HARDWARE events, such as instructions, cache-misses etc.

cpu-clock works well, but it is a software event (PERF_TYPE_SOFTWARE). We are interested specifically in running tests with the hardware performance counters which are normally supported on AArch64 machines, including those available in public clouds. Thanks.