gh-152769: Enable perf trampoline on Alpine

[weixlu]

Proposal of change

As mentioned in #152769, perf can‘t collect Python function-level hotspot info on Alpine Linux because perf trampoline support is not enabled there.

The trampoline is a CPU-architecture-specific feature and has little to do with OS or libc. Therefore, we can simply modify configure to reuse the same trampoline implementation as glibc.

Test results

I ran the tests on the following three platforms:

• Alpine 3.22.4, musl 1.2.5, x86_64 (matching the buildbot env)
• Alpine edge, musl 1.2.6, x86_64 (latest version)
• Alpine 3.22.4, musl 1.2.5, aarch64 (specifically ARMv9 Neoverse N2)

The three profiling-related tests now pass on all of these platforms:

test_perf_profiler
test_perfmaps
test_samply_profiler

Before this PR, these tests were skipped.

I also manually collected profiling data with perf and confirmed that Python-level functions appear in the profiling output.

How to reproduce

1. When running Alpine in a container, note that the container needs to be started with the additional --privileged option.

2. Apart from build dependencies, install perf and samply. samply needs to be built with Rust:

apk add perf rust cargo
export PATH="$HOME/.cargo/bin:$PATH"
cargo install --locked samply

3. Run the tests:

./python -m test test_perf_profiler test_perfmaps test_samply_profiler

4. To manually check profiling data, refer to the scripts described in Lack of perf support on Alpine Linux #152769.

• Issue: Lack of perf support on Alpine Linux #152769

[weixlu]

@vstinner Hi, this is the PR fixing perf trampolines on musl. Could you please take a look?

[thesamesam]

See also #122370 (and #121608).

[weixlu]

Thanks for pointing out, @thesamesam — you're right. I'll close #152769 as a duplicate of #121608 so the discussion stays in one place.

Hi, @corona10, since you authored the original #122370, I'd like to defer to you on how best to handle this.

• Back in 2024 when you opened the PR, the profiling tests were skipped because the Alpine perf was too old.
• Now with the latest version of Alpine, that's no longer an issue — I verified on three platforms and test_perf_profiler / test_perfmaps / test_samply_profiler now pass.
• The configure logic has also changed a bit in the meantime, so the original diff in gh-121608: Add supporting perf trampoline on -linux-musl #122370 doesn't apply cleanly to main anymore. This PR is the same change on the current code.

Whatever works best for you — if you'd prefer to update and land #122370 yourself, I'm happy to close mine; or feel free to use this PR if that's easier. Thanks!

[vstinner]

I have issues to try samply in my Alpine Linux (podman) container running on Fedora 44 :-(

First, it complained because of perf_event_paranoid:

# samply record --save-only --output profile.json.gz /root/cpython/python -c 'print("hello")' 

'/proc/sys/kernel/perf_event_paranoid' is currently set to 2.
In order for samply to work with a non-root user, this level needs
to be set to 1 or lower.
You can execute the following command and then try again:
    echo '1' | sudo tee /proc/sys/kernel/perf_event_paranoid

I ran echo '1' | sudo tee /proc/sys/kernel/perf_event_paranoid, but then it fails with:

# samply record --save-only --output profile.json.gz /root/cpython/python -c 'print("hello")' |cat
Failed to start profiling: Operation not permitted (os error 1)

That's maybe because I'm running Alpine Linux in a podman container which is running as root?

Second, test_samply_profiler hangs randomly on samply_command_works(). I can reproduce the issue in the command line, without Python:

# samply record --save-only --output profile.json.gz /root/cpython/python -c 'print("hello")' |cat
Failed to start profiling: Operation not permitted (os error 1)

The command hangs. If I run samply in strace, it no longer hangs, so I cannot debug the issue :-(

# strace -o trace samply record --save-only --output profile.json.gz /root/cpython/python -c 'print("hello")' |cat
Failed to start profiling: Operation not permitted (os error 1)

By the way, test_samply_profiler runs samply_command_works() 3 times, which is inefficient! I wrote #153424 to only run samply_command_works() once.

[vstinner]

I also had permission issues with running Linux perf in my Alpine Linux container:

# perf record pwd
perf_event_open(..., PERF_FLAG_FD_CLOEXEC) failed with unexpected error 1 (Operation not permitted)
perf_event_open(..., 0) failed unexpectedly with error 1 (Operation not permitted)
Missing support for build id in kernel mmap events.
Disable this warning with --no-buildid-mmap
Error:
Failure to open event 'cpu/cycles/P' on PMU 'cpu' which will be removed.
No fallback found for 'cpu/cycles/P' for error 1
Error:
Failure to open any events for recording.

Note: I tried setting /proc/sys/kernel/perf_event_paranoid to -1, but it didn't help.

[vstinner]

@corona10: Are you able to run the 3 tests on Alpine Linux successfully?

[weixlu]

@vstinner Hi, looks like it's related to permission issues. When you start your podman, did you add the flag --privileged? like podman run -it --privileged alpine:edge sh

[vstinner]

Ah, I forgot to mention that I tried running podman run --privileged, but it didn't solve the permission issue.

[weixlu]

Could you let me know which version of the alpine image you're using? I'll see if I can reproduce it.

[weixlu]

And perf_event_paranoid is 1 here on my side. I'm also running the container as root

~ # cat /proc/sys/kernel/perf_event_paranoid
1
~ # perf record pwd
/root
[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.017 MB perf.data ]
~ # perf version
perf version 6.15.0