test_is_python_process_non_python_subprocess always times out when sleep is a multicall binary (Alpine/BusyBox, RHEL UBI9) #160881

Bug report

Bug description:

The fix for gh-159534 (GH-160028, backported to 3.15 in GH-160080) makes test_is_python_process_non_python_subprocess fail every time on systems where sleep is provided by a multicall binary. That includes the Alpine buildbots, where the original gh-159534 failure was found:

FAIL: test_is_python_process_non_python_subprocess (test.test_profiling.test_sampling_profiler.test_children.TestIsPythonProcess.test_is_python_process_non_python_subprocess)
Test that a non-Python subprocess is not detected as Python.
----------------------------------------------------------------------
Traceback (most recent call last):
  File ".../Lib/test/test_profiling/test_sampling_profiler/test_children.py", line 628, in test_is_python_process_non_python_subprocess
    _wait_for_process_exec(proc, "sleep", SHORT_TIMEOUT)
  File ".../Lib/test/test_profiling/test_sampling_profiler/test_children.py", line 91, in _wait_for_process_exec
    for _ in sleeping_retry(timeout, max_delay=0.1):
  File ".../Lib/test/support/__init__.py", line 2779, in sleeping_retry
    for _ in busy_retry(timeout, err_msg, error=error):
  File ".../Lib/test/support/__init__.py", line 2749, in busy_retry
    raise AssertionError(msg)
AssertionError: timeout (45.1 seconds)

Buildbots

AMD64 Alpine Linux NoGIL 3.x:

• build 3444 (before gh-159534: Fix race condition in test_sampling_profiler #160028): the original intermittent failure, True is not false.
• build 3445 (first build with gh-159534: Fix race condition in test_sampling_profiler #160028): timeout (45.1 seconds). Every build since fails the same way, e.g. build 3523.

AMD64 Alpine Linux 3.15:

• build 800 (before [3.15] gh-159534: Fix race condition in test_sampling_profiler (GH-160028) #160080): only test_lazy_import fails.
• build 815 (first build with [3.15] gh-159534: Fix race condition in test_sampling_profiler (GH-160028) #160080) and every build since, including build 851 just before v3.15.2 was tagged: this test fails too.

This failure is separate from the test_lazy_import crash on Alpine (gh-159535). The Alpine builders will stay red after that crash is fixed.

Cause

The new _wait_for_process_exec() helper waits until os.path.basename(os.readlink(f"/proc/{pid}/exe")) == "sleep". When sleep is a multicall binary, /proc/{pid}/exe points to the multicall binary instead:

The condition never becomes true, so the test always times out.

Reproducer, tested in python:3-alpine and registry.access.redhat.com/ubi9/ubi:

import os, subprocess, time
proc = subprocess.Popen(["sleep", "10"])
time.sleep(0.5)
print(os.readlink(f"/proc/{proc.pid}/exe"))
proc.kill()

Suggested fix

The helper only needs to know that the child has called exec(). The program name doesn't matter. Before exec(), the forked child's /proc/{pid}/exe is the parent's Python executable, so wait until it differs from the parent's:

-def _wait_for_process_exec(proc, name, timeout):
+def _wait_for_process_exec(proc, timeout):
     # Wait until a child process calls exec() to replace the forked process
-    # executable. Raise AssertionError on timeout.
+    # executable. Don't check the executable name: it can be a multicall
+    # binary such as /bin/busybox. Raise AssertionError on timeout.
+    parent_exe = os.readlink("/proc/self/exe")
     for _ in sleeping_retry(timeout, max_delay=0.1):
         exe = os.readlink(f"/proc/{proc.pid}/exe")
-        if os.path.basename(exe) == name:
+        if exe != parent_exe:
             break

-            _wait_for_process_exec(proc, "sleep", SHORT_TIMEOUT)
+            _wait_for_process_exec(proc, SHORT_TIMEOUT)

I tested this check with the reproducer on Alpine (BusyBox) and on UBI9 with both coreutils-single and coreutils. It detects the exec() in all three cases, while the current name check fails on Alpine and on UBI9 with coreutils-single.

CPython versions tested on:

3.15.2, CPython main branch

Operating systems tested on:

Linux (Alpine 3.24 buildbots, RHEL UBI9 container)

CPython versions tested on:

3.15

Operating systems tested on:

Linux

Linked PRs

• gh-160881: Fix test_is_python_process_non_python_subprocess with multicall sleep #160901