gh-96143: Allow Linux perf profiler to see Python calls (GH-96123)

⚠️  ⚠️ Note for reviewers, hackers and fellow systems/low-level/compiler engineers ⚠️ ⚠️ 

If you have a lot of experience with this kind of shenanigans and want to improve the **first** version, **please make a PR against my branch** or **reach out by email** or **suggest code changes directly on GitHub**. 

If you have any **refinements or optimizations** please, wait until the first version is merged before starting hacking or proposing those so we can keep this PR productive.

Doc/c-api/init_config.rst

@@ -1155,6 +1155,20 @@ PyConfig
 
       Default: ``-1`` in Python mode, ``0`` in isolated mode.
 
+   .. c:member:: int perf_profiling
+
+      Enable compatibility mode with the perf profiler?
+
+      If non-zero, initialize the perf trampoline. See :ref:`perf_profiling`
+      for more information.
+
+      Set by :option:`-X perf <-X>` command line option and by the
+      :envvar:`PYTHONPERFSUPPORT` environment variable.
+
+      Default: ``-1``.
+
+      .. versionadded:: 3.12
+
    .. c:member:: int use_environment
 
       Use :ref:`environment variables <using-on-envvars>`?

Doc/howto/index.rst

@@ -30,6 +30,7 @@ Currently, the HOWTOs are:
    ipaddress.rst
    clinic.rst
    instrumentation.rst
+   perf_profiling.rst
    annotations.rst
    isolating-extensions.rst
 

Doc/howto/perf_profiling.rst

@@ -0,0 +1,200 @@
+.. highlight:: shell-session
+
+.. _perf_profiling:
+
+==============================================
+Python support for the Linux ``perf`` profiler
+==============================================
+
+:author: Pablo Galindo
+
+The Linux ``perf`` profiler is a very powerful tool that allows you to profile and
+obtain information about the performance of your application. ``perf`` also has
+a very vibrant ecosystem of tools that aid with the analysis of the data that it
+produces.
+
+The main problem with using the ``perf`` profiler with Python applications is that
+``perf`` only allows to get information about native symbols, this is, the names of
+the functions and procedures written in C. This means that the names and file names
+of the Python functions in your code will not appear in the output of the ``perf``.
+
+Since Python 3.12, the interpreter can run in a special mode that allows Python
+functions to appear in the output of the ``perf`` profiler. When this mode is
+enabled, the interpreter will interpose a small piece of code compiled on the
+fly before the execution of every Python function and it will teach ``perf`` the
+relationship between this piece of code and the associated Python function using
+`perf map files`_.
+
+.. warning::
+
+    Support for the ``perf`` profiler is only currently available for Linux on
+    selected architectures. Check the output of the configure build step or
+    check the output of ``python -m sysconfig | grep HAVE_PERF_TRAMPOLINE``
+    to see if your system is supported.
+
+For example, consider the following script:
+
+.. code-block:: python
+
+    def foo(n):
+        result = 0
+        for _ in range(n):
+            result += 1
+        return result
+
+    def bar(n):
+        foo(n)
+
+    def baz(n):
+        bar(n)
+
+    if __name__ == "__main__":
+        baz(1000000)
+
+We can run perf to sample CPU stack traces at 9999 Hertz:
+
+    $ perf record -F 9999 -g -o perf.data python my_script.py
+
+Then we can use perf report to analyze the data:
+
+.. code-block:: shell-session
+
+    $ perf report --stdio -n -g
+
+    # Children      Self       Samples  Command     Shared Object       Symbol
+    # ........  ........  ............  ..........  ..................  ..........................................
+    #
+        91.08%     0.00%             0  python.exe  python.exe          [.] _start
+                |
+                ---_start
+                |
+                    --90.71%--__libc_start_main
+                            Py_BytesMain
+                            |
+                            |--56.88%--pymain_run_python.constprop.0
+                            |          |
+                            |          |--56.13%--_PyRun_AnyFileObject
+                            |          |          _PyRun_SimpleFileObject
+                            |          |          |
+                            |          |          |--55.02%--run_mod
+                            |          |          |          |
+                            |          |          |           --54.65%--PyEval_EvalCode
+                            |          |          |                     _PyEval_EvalFrameDefault
+                            |          |          |                     PyObject_Vectorcall
+                            |          |          |                     _PyEval_Vector
+                            |          |          |                     _PyEval_EvalFrameDefault
+                            |          |          |                     PyObject_Vectorcall
+                            |          |          |                     _PyEval_Vector
+                            |          |          |                     _PyEval_EvalFrameDefault
+                            |          |          |                     PyObject_Vectorcall
+                            |          |          |                     _PyEval_Vector
+                            |          |          |                     |
+                            |          |          |                     |--51.67%--_PyEval_EvalFrameDefault
+                            |          |          |                     |          |
+                            |          |          |                     |          |--11.52%--_PyLong_Add
+                            |          |          |                     |          |          |
+                            |          |          |                     |          |          |--2.97%--_PyObject_Malloc
+    ...
+
+As you can see here, the Python functions are not shown in the output, only ``_Py_Eval_EvalFrameDefault`` appears
+(the function that evaluates the Python bytecode) shows up. Unfortunately that's not very useful because all Python
+functions use the same C function to evaluate bytecode so we cannot know which Python function corresponds to which
+bytecode-evaluating function.
+
+Instead, if we run the same experiment with perf support activated we get:
+
+.. code-block:: shell-session
+
+    $ perf report --stdio -n -g
+
+    # Children      Self       Samples  Command     Shared Object       Symbol
+    # ........  ........  ............  ..........  ..................  .....................................................................
+    #
+        90.58%     0.36%             1  python.exe  python.exe          [.] _start
+                |
+                ---_start
+                |
+                    --89.86%--__libc_start_main
+                            Py_BytesMain
+                            |
+                            |--55.43%--pymain_run_python.constprop.0
+                            |          |
+                            |          |--54.71%--_PyRun_AnyFileObject
+                            |          |          _PyRun_SimpleFileObject
+                            |          |          |
+                            |          |          |--53.62%--run_mod
+                            |          |          |          |
+                            |          |          |           --53.26%--PyEval_EvalCode
+                            |          |          |                     py::<module>:/src/script.py
+                            |          |          |                     _PyEval_EvalFrameDefault
+                            |          |          |                     PyObject_Vectorcall
+                            |          |          |                     _PyEval_Vector
+                            |          |          |                     py::baz:/src/script.py
+                            |          |          |                     _PyEval_EvalFrameDefault
+                            |          |          |                     PyObject_Vectorcall
+                            |          |          |                     _PyEval_Vector
+                            |          |          |                     py::bar:/src/script.py
+                            |          |          |                     _PyEval_EvalFrameDefault
+                            |          |          |                     PyObject_Vectorcall
+                            |          |          |                     _PyEval_Vector
+                            |          |          |                     py::foo:/src/script.py
+                            |          |          |                     |
+                            |          |          |                     |--51.81%--_PyEval_EvalFrameDefault
+                            |          |          |                     |          |
+                            |          |          |                     |          |--13.77%--_PyLong_Add
+                            |          |          |                     |          |          |
+                            |          |          |                     |          |          |--3.26%--_PyObject_Malloc
+
+
+
+Enabling perf profiling mode
+----------------------------
+
+There are two main ways to activate the perf profiling mode. If you want it to be
+active since the start of the Python interpreter, you can use the `-Xperf` option:
+
+    $ python -Xperf my_script.py
+
+There is also support for dynamically activating and deactivating the perf
+profiling mode by using the APIs in the :mod:`sys` module:
+
+.. code-block:: python
+
+    import sys
+    sys.activate_stack_trampoline("perf")
+
+    # Run some code with Perf profiling active
+
+    sys.deactivate_stack_trampoline()
+
+    # Perf profiling is not active anymore
+
+These APIs can be handy if you want to activate/deactivate profiling mode in
+response to a signal or other communication mechanism with your process.
+
+
+
+Now we can analyze the data with ``perf report``:
+
+    $ perf report -g -i perf.data
+
+
+How to obtain the best results
+-------------------------------
+
+For the best results, Python should be compiled with
+``CFLAGS="-fno-omit-frame-pointer -mno-omit-leaf-frame-pointer"`` as this allows
+profilers to unwind using only the frame pointer and not on DWARF debug
+information. This is because as the code that is interposed to allow perf
+support is dynamically generated it doesn't have any DWARF debugging information
+available.
+
+You can check if you system has been compiled with this flag by running:
+
+    $ python -m sysconfig | grep 'no-omit-frame-pointer'
+
+If you don't see any output it means that your interpreter has not been compiled with
+frame pointers and therefore it may not be able to show Python functions in the output
+of ``perf``.
+
+.. _perf map files: https://github.com/torvalds/linux/blob/0513e464f9007b70b96740271a948ca5ab6e7dd7/tools/perf/Documentation/jit-interface.txt

Doc/using/cmdline.rst

@@ -535,6 +535,12 @@ Miscellaneous options
      development (running from the source tree) then the default is "off".
      Note that the "importlib_bootstrap" and "importlib_bootstrap_external"
      frozen modules are always used, even if this flag is set to "off".
+   * ``-X perf`` to activate compatibility mode with the ``perf`` profiler.
+     When this option is activated, the Linux ``perf`` profiler will be able to
+     report Python calls. This option is only available on some platforms and
+     will do nothing if is not supported on the current system. The default value
+     is "off". See also :envvar:`PYTHONPERFSUPPORT` and :ref:`perf_profiling`
+     for more information.
 
    It also allows passing arbitrary values and retrieving them through the
    :data:`sys._xoptions` dictionary.
@@ -1025,6 +1031,13 @@ conflict.
 
    .. versionadded:: 3.11
 
+.. envvar:: PYTHONPERFSUPPORT
+
+   If this variable is set to a nonzero value, it activates compatibility mode
+   with the ``perf`` profiler so Python calls can be detected by it. See the
+   :ref:`perf_profiling` section for more information.
+
+   .. versionadded:: 3.12
 
 
 Debug-mode variables

Include/cpython/initconfig.h

@@ -142,6 +142,7 @@ typedef struct PyConfig {
     unsigned long hash_seed;
     int faulthandler;
     int tracemalloc;
+    int perf_profiling;
     int import_time;
     int code_debug_ranges;
     int show_ref_count;

Include/internal/pycore_ceval.h

@@ -65,6 +65,27 @@ extern PyObject* _PyEval_BuiltinsFromGlobals(
     PyThreadState *tstate,
     PyObject *globals);
 
+// Trampoline API
+
+typedef struct {
+    // Callback to initialize the trampoline state
+    void* (*init_state)(void);
+    // Callback to register every trampoline being created
+    void (*write_state)(void* state, const void *code_addr,
+                        unsigned int code_size, PyCodeObject* code);
+    // Callback to free the trampoline state
+    int (*free_state)(void* state);
+} _PyPerf_Callbacks;
+
+extern int _PyPerfTrampoline_SetCallbacks(_PyPerf_Callbacks *);
+extern void _PyPerfTrampoline_GetCallbacks(_PyPerf_Callbacks *);
+extern int _PyPerfTrampoline_Init(int activate);
+extern int _PyPerfTrampoline_Fini(void);
+extern int _PyIsPerfTrampolineActive(void);
+extern PyStatus _PyPerfTrampoline_AfterFork_Child(void);
+#ifdef PY_HAVE_PERF_TRAMPOLINE
+extern _PyPerf_Callbacks _Py_perfmap_callbacks;
+#endif
 
 static inline PyObject*
 _PyEval_EvalFrame(PyThreadState *tstate, struct _PyInterpreterFrame *frame, int throwflag)

Lib/test/test_embed.py

@@ -436,6 +436,7 @@ class InitConfigTests(EmbeddingTestsMixin, unittest.TestCase):
         'hash_seed': 0,
         'faulthandler': 0,
         'tracemalloc': 0,
+        'perf_profiling': 0,
         'import_time': 0,
         'code_debug_ranges': 1,
         'show_ref_count': 0,
@@ -520,6 +521,7 @@ class InitConfigTests(EmbeddingTestsMixin, unittest.TestCase):
         use_hash_seed=0,
         faulthandler=0,
         tracemalloc=0,
+        perf_profiling=0,
         pathconfig_warnings=0,
     )
     if MS_WINDOWS:
@@ -828,6 +830,7 @@ class InitConfigTests(EmbeddingTestsMixin, unittest.TestCase):
             'use_hash_seed': 1,
             'hash_seed': 123,
             'tracemalloc': 2,
+            'perf_profiling': 0,
             'import_time': 1,
             'code_debug_ranges': 0,
             'show_ref_count': 1,
@@ -890,6 +893,7 @@ class InitConfigTests(EmbeddingTestsMixin, unittest.TestCase):
             'use_hash_seed': 1,
             'hash_seed': 42,
             'tracemalloc': 2,
+            'perf_profiling': 0,
             'import_time': 1,
             'code_debug_ranges': 0,
             'malloc_stats': 1,
@@ -921,6 +925,7 @@ class InitConfigTests(EmbeddingTestsMixin, unittest.TestCase):
             'use_hash_seed': 1,
             'hash_seed': 42,
             'tracemalloc': 2,
+            'perf_profiling': 0,
             'import_time': 1,
             'code_debug_ranges': 0,
             'malloc_stats': 1,

Lib/test/test_perf_profiler.py

@@ -0,0 +1,348 @@
+import unittest
+import subprocess
+import sys
+import sysconfig
+import os
+import pathlib
+from test import support
+from test.support.script_helper import (
+    make_script,
+    assert_python_failure,
+    assert_python_ok,
+)
+from test.support.os_helper import temp_dir
+
+
+if not support.has_subprocess_support:
+    raise unittest.SkipTest("test module requires subprocess")
+
+
+def supports_trampoline_profiling():
+    perf_trampoline = sysconfig.get_config_var("PY_HAVE_PERF_TRAMPOLINE")
+    if not perf_trampoline:
+        return False
+    return int(perf_trampoline) == 1
+
+
+if not supports_trampoline_profiling():
+    raise unittest.SkipTest("perf trampoline profiling not supported")
+
+
+class TestPerfTrampoline(unittest.TestCase):
+    def setUp(self):
+        super().setUp()
+        self.perf_files = set(pathlib.Path("/tmp/").glob("perf-*.map"))
+
+    def tearDown(self) -> None:
+        super().tearDown()
+        files_to_delete = (
+            set(pathlib.Path("/tmp/").glob("perf-*.map")) - self.perf_files
+        )
+        for file in files_to_delete:
+            file.unlink()
+
+    def test_trampoline_works(self):
+        code = """if 1:
+                def foo():
+                    pass
+
+                def bar():
+                    foo()
+
+                def baz():
+                    bar()
+
+                baz()
+                """
+        with temp_dir() as script_dir:
+            script = make_script(script_dir, "perftest", code)
+            with subprocess.Popen(
+                [sys.executable, "-Xperf", script],
+                universal_newlines=True,
+                stderr=subprocess.PIPE,
+                stdout=subprocess.PIPE,
+            ) as process:
+                stdout, stderr = process.communicate()
+
+        self.assertEqual(stderr, "")
+        self.assertEqual(stdout, "")
+
+        perf_file = pathlib.Path(f"/tmp/perf-{process.pid}.map")
+        self.assertTrue(perf_file.exists())
+        perf_file_contents = perf_file.read_text()
+        self.assertIn(f"py::foo:{script}", perf_file_contents)
+        self.assertIn(f"py::bar:{script}", perf_file_contents)
+        self.assertIn(f"py::baz:{script}", perf_file_contents)
+
+    def test_trampoline_works_with_forks(self):
+        code = """if 1:
+                import os, sys
+
+                def foo_fork():
+                    pass
+
+                def bar_fork():
+                    foo_fork()
+
+                def baz_fork():
+                    bar_fork()
+
+                def foo():
+                    pid = os.fork()
+                    if pid == 0:
+                        print(os.getpid())
+                        baz_fork()
+                    else:
+                        _, status = os.waitpid(-1, 0)
+                        sys.exit(status)
+
+                def bar():
+                    foo()
+
+                def baz():
+                    bar()
+
+                baz()
+                """
+        with temp_dir() as script_dir:
+            script = make_script(script_dir, "perftest", code)
+            with subprocess.Popen(
+                [sys.executable, "-Xperf", script],
+                universal_newlines=True,
+                stderr=subprocess.PIPE,
+                stdout=subprocess.PIPE,
+            ) as process:
+                stdout, stderr = process.communicate()
+
+        self.assertEqual(process.returncode, 0)
+        self.assertEqual(stderr, "")
+        child_pid = int(stdout.strip())
+        perf_file = pathlib.Path(f"/tmp/perf-{process.pid}.map")
+        perf_child_file = pathlib.Path(f"/tmp/perf-{child_pid}.map")
+        self.assertTrue(perf_file.exists())
+        self.assertTrue(perf_child_file.exists())
+
+        perf_file_contents = perf_file.read_text()
+        self.assertIn(f"py::foo:{script}", perf_file_contents)
+        self.assertIn(f"py::bar:{script}", perf_file_contents)
+        self.assertIn(f"py::baz:{script}", perf_file_contents)
+
+        child_perf_file_contents = perf_child_file.read_text()
+        self.assertIn(f"py::foo_fork:{script}", child_perf_file_contents)
+        self.assertIn(f"py::bar_fork:{script}", child_perf_file_contents)
+        self.assertIn(f"py::baz_fork:{script}", child_perf_file_contents)
+
+    def test_sys_api(self):
+        code = """if 1:
+                import sys
+                def foo():
+                    pass
+
+                def spam():
+                    pass
+
+                def bar():
+                    sys.deactivate_stack_trampoline()
+                    foo()
+                    sys.activate_stack_trampoline("perf")
+                    spam()
+
+                def baz():
+                    bar()
+
+                sys.activate_stack_trampoline("perf")
+                baz()
+                """
+        with temp_dir() as script_dir:
+            script = make_script(script_dir, "perftest", code)
+            with subprocess.Popen(
+                [sys.executable, script],
+                universal_newlines=True,
+                stderr=subprocess.PIPE,
+                stdout=subprocess.PIPE,
+            ) as process:
+                stdout, stderr = process.communicate()
+
+        self.assertEqual(stderr, "")
+        self.assertEqual(stdout, "")
+
+        perf_file = pathlib.Path(f"/tmp/perf-{process.pid}.map")
+        self.assertTrue(perf_file.exists())
+        perf_file_contents = perf_file.read_text()
+        self.assertNotIn(f"py::foo:{script}", perf_file_contents)
+        self.assertIn(f"py::spam:{script}", perf_file_contents)
+        self.assertIn(f"py::bar:{script}", perf_file_contents)
+        self.assertIn(f"py::baz:{script}", perf_file_contents)
+
+    def test_sys_api_with_existing_trampoline(self):
+        code = """if 1:
+                import sys
+                sys.activate_stack_trampoline("perf")
+                sys.activate_stack_trampoline("perf")
+                """
+        assert_python_ok("-c", code)
+
+    def test_sys_api_with_invalid_trampoline(self):
+        code = """if 1:
+                import sys
+                sys.activate_stack_trampoline("invalid")
+                """
+        rc, out, err = assert_python_failure("-c", code)
+        self.assertIn("invalid backend: invalid", err.decode())
+
+    def test_sys_api_get_status(self):
+        code = """if 1:
+                import sys
+                sys.activate_stack_trampoline("perf")
+                assert sys.is_stack_trampoline_active() is True
+                sys.deactivate_stack_trampoline()
+                assert sys.is_stack_trampoline_active() is False
+                """
+        assert_python_ok("-c", code)
+
+
+def is_unwinding_reliable():
+    cflags = sysconfig.get_config_var("PY_CORE_CFLAGS")
+    if not cflags:
+        return False
+    return "no-omit-frame-pointer" in cflags
+
+
+def perf_command_works():
+    try:
+        cmd = ["perf", "--help"]
+        stdout = subprocess.check_output(cmd, universal_newlines=True)
+    except (subprocess.SubprocessError, OSError):
+        return False
+
+    # perf version does not return a version number on Fedora. Use presence
+    # of "perf.data" in help as indicator that it's perf from Linux tools.
+    if "perf.data" not in stdout:
+        return False
+
+    # Check that we can run a simple perf run
+    with temp_dir() as script_dir:
+        try:
+            output_file = script_dir + "/perf_output.perf"
+            cmd = (
+                "perf",
+                "record",
+                "-g",
+                "--call-graph=fp",
+                "-o",
+                output_file,
+                "--",
+                sys.executable,
+                "-c",
+                'print("hello")',
+            )
+            stdout = subprocess.check_output(
+                cmd, cwd=script_dir, universal_newlines=True, stderr=subprocess.STDOUT
+            )
+        except (subprocess.SubprocessError, OSError):
+            return False
+
+        if "hello" not in stdout:
+            return False
+
+    return True
+
+
+def run_perf(cwd, *args, **env_vars):
+    if env_vars:
+        env = os.environ.copy()
+        env.update(env_vars)
+    else:
+        env = None
+    output_file = cwd + "/perf_output.perf"
+    base_cmd = ("perf", "record", "-g", "--call-graph=fp", "-o", output_file, "--")
+    proc = subprocess.run(
+        base_cmd + args,
+        stdout=subprocess.PIPE,
+        stderr=subprocess.PIPE,
+        env=env,
+    )
+    if proc.returncode:
+        print(proc.stderr)
+        raise ValueError(f"Perf failed with return code {proc.returncode}")
+
+    base_cmd = ("perf", "script")
+    proc = subprocess.run(
+        ("perf", "script", "-i", output_file),
+        stdout=subprocess.PIPE,
+        stderr=subprocess.PIPE,
+        env=env,
+        check=True,
+    )
+    return proc.stdout.decode("utf-8", "replace"), proc.stderr.decode(
+        "utf-8", "replace"
+    )
+
+
+@unittest.skipUnless(perf_command_works(), "perf command doesn't work")
+@unittest.skipUnless(is_unwinding_reliable(), "Unwinding is unreliable")
+@support.skip_if_sanitizer(address=True, memory=True, ub=True)
+class TestPerfProfiler(unittest.TestCase):
+    def setUp(self):
+        super().setUp()
+        self.perf_files = set(pathlib.Path("/tmp/").glob("perf-*.map"))
+
+    def tearDown(self) -> None:
+        super().tearDown()
+        files_to_delete = (
+            set(pathlib.Path("/tmp/").glob("perf-*.map")) - self.perf_files
+        )
+        for file in files_to_delete:
+            file.unlink()
+
+    def test_python_calls_appear_in_the_stack_if_perf_activated(self):
+        with temp_dir() as script_dir:
+            code = """if 1:
+                def foo(n):
+                    x = 0
+                    for i in range(n):
+                        x += i
+
+                def bar(n):
+                    foo(n)
+
+                def baz(n):
+                    bar(n)
+
+                baz(10000000)
+                """
+            script = make_script(script_dir, "perftest", code)
+            stdout, stderr = run_perf(script_dir, sys.executable, "-Xperf", script)
+            self.assertEqual(stderr, "")
+
+            self.assertIn(f"py::foo:{script}", stdout)
+            self.assertIn(f"py::bar:{script}", stdout)
+            self.assertIn(f"py::baz:{script}", stdout)
+
+    def test_python_calls_do_not_appear_in_the_stack_if_perf_activated(self):
+        with temp_dir() as script_dir:
+            code = """if 1:
+                def foo(n):
+                    x = 0
+                    for i in range(n):
+                        x += i
+
+                def bar(n):
+                    foo(n)
+
+                def baz(n):
+                    bar(n)
+
+                baz(10000000)
+                """
+            script = make_script(script_dir, "perftest", code)
+            stdout, stderr = run_perf(script_dir, sys.executable, script)
+            self.assertEqual(stderr, "")
+
+            self.assertNotIn(f"py::foo:{script}", stdout)
+            self.assertNotIn(f"py::bar:{script}", stdout)
+            self.assertNotIn(f"py::baz:{script}", stdout)
+
+
+if __name__ == "__main__":
+    unittest.main()

Makefile.pre.in

@@ -478,7 +478,9 @@ OBJECT_OBJS=	\
 		Objects/unicodeobject.o \
 		Objects/unicodectype.o \
 		Objects/unionobject.o \
-		Objects/weakrefobject.o
+		Objects/weakrefobject.o \
+		Objects/perf_trampoline.o \
+		@PERF_TRAMPOLINE_OBJ@
 
 DEEPFREEZE_OBJS = Python/deepfreeze/deepfreeze.o
 
@@ -2358,6 +2360,9 @@ config.status:	$(srcdir)/configure
 
 .PRECIOUS: config.status $(BUILDPYTHON) Makefile Makefile.pre
 
+Objects/asm_trampoline.o: $(srcdir)/Objects/asm_trampoline.S
+	$(CC) -c $(PY_CORE_CFLAGS) -o $@ $<
+
 # Some make's put the object file in the current directory
 .c.o:
 	$(CC) -c $(PY_CORE_CFLAGS) -o $@ $<

Misc/NEWS.d/next/Core and Builtins/2022-08-20-18-36-40.gh-issue-96143.nh3GFM.rst

@@ -0,0 +1,7 @@
+Add a new ``-X perf`` Python command line option as well as
+:func:`sys.activate_stack_trampoline` and :func:`sys.deactivate_stack_trampoline`
+function in the :mod:`sys` module that allows to set/unset the interpreter in a
+way that the Linux ``perf`` profiler can detect Python calls. The new
+:func:`sys.is_stack_trampoline_active` function allows to query the state of the
+perf trampoline. Design by Pablo Galindo. Patch by Pablo Galindo and Christian Heimes
+with contributions from Gregory P. Smith [Google] and Mark Shannon.

Modules/posixmodule.c

@@ -606,6 +606,11 @@ PyOS_AfterFork_Child(void)
     }
     assert(_PyThreadState_GET() == tstate);
 
+    status = _PyPerfTrampoline_AfterFork_Child();
+    if (_PyStatus_EXCEPTION(status)) {
+        goto fatal_error;
+    }
+
     run_at_forkers(tstate->interp->after_forkers_child, 0);
     return;
 

Objects/asm_trampoline.S

@@ -0,0 +1,28 @@
+    .text
+    .globl	_Py_trampoline_func_start
+# The following assembly is equivalent to:
+# PyObject *
+# trampoline(PyThreadState *ts, _PyInterpreterFrame *f,
+#            int throwflag, py_evaluator evaluator)
+# {
+#     return evaluator(ts, f, throwflag);
+# }
+_Py_trampoline_func_start:
+#ifdef __x86_64__
+    sub    $8, %rsp
+    call    *%rcx
+    add    $8, %rsp
+    ret
+#endif // __x86_64__
+#if defined(__aarch64__) && defined(__AARCH64EL__) && !defined(__ILP32__)
+    // ARM64 little endian, 64bit ABI
+    // generate with aarch64-linux-gnu-gcc 12.1
+    stp     x29, x30, [sp, -16]!
+    mov     x29, sp
+    blr     x3
+    ldp     x29, x30, [sp], 16
+    ret
+#endif
+    .globl	_Py_trampoline_func_end
+_Py_trampoline_func_end:
+    .section        .note.GNU-stack,"",@progbits

Objects/perf_trampoline.c

@@ -0,0 +1,501 @@
+/*
+
+Perf trampoline instrumentation
+===============================
+
+This file contains instrumentation to allow to associate
+calls to the CPython eval loop back to the names of the Python
+functions and filename being executed.
+
+Many native performance profilers like the Linux perf tools are
+only available to 'see' the C stack when sampling from the profiled
+process. This means that if we have the following python code:
+
+    import time
+    def foo(n):
+        # Some CPU intensive code
+
+    def bar(n):
+        foo(n)
+
+    def baz(n):
+        bar(n)
+
+    baz(10000000)
+
+A performance profiler that is only able to see native frames will
+produce the following backtrace when sampling from foo():
+
+    _PyEval_EvalFrameDefault -----> Evaluation frame of foo()
+    _PyEval_Vector
+    _PyFunction_Vectorcall
+    PyObject_Vectorcall
+    call_function
+
+    _PyEval_EvalFrameDefault ------> Evaluation frame of bar()
+    _PyEval_EvalFrame
+    _PyEval_Vector
+    _PyFunction_Vectorcall
+    PyObject_Vectorcall
+    call_function
+
+    _PyEval_EvalFrameDefault -------> Evaluation frame of baz()
+    _PyEval_EvalFrame
+    _PyEval_Vector
+    _PyFunction_Vectorcall
+    PyObject_Vectorcall
+    call_function
+
+    ...
+
+    Py_RunMain
+
+Because the profiler is only able to see the native frames and the native
+function that runs the evaluation loop is the same (_PyEval_EvalFrameDefault)
+then the profiler and any reporter generated by it will not be able to
+associate the names of the Python functions and the filenames associated with
+those calls, rendering the results useless in the Python world.
+
+To fix this problem, we introduce the concept of a trampoline frame. A
+trampoline frame is a piece of code that is unique per Python code object that
+is executed before entering the CPython eval loop. This piece of code just
+calls the original Python evaluation function (_PyEval_EvalFrameDefault) and
+forwards all the arguments received. In this way, when a profiler samples
+frames from the previous example it will see;
+
+    _PyEval_EvalFrameDefault -----> Evaluation frame of foo()
+    [Jit compiled code 3]
+    _PyEval_Vector
+    _PyFunction_Vectorcall
+    PyObject_Vectorcall
+    call_function
+
+    _PyEval_EvalFrameDefault ------> Evaluation frame of bar()
+    [Jit compiled code 2]
+    _PyEval_EvalFrame
+    _PyEval_Vector
+    _PyFunction_Vectorcall
+    PyObject_Vectorcall
+    call_function
+
+    _PyEval_EvalFrameDefault -------> Evaluation frame of baz()
+    [Jit compiled code 1]
+    _PyEval_EvalFrame
+    _PyEval_Vector
+    _PyFunction_Vectorcall
+    PyObject_Vectorcall
+    call_function
+
+    ...
+
+    Py_RunMain
+
+When we generate every unique copy of the trampoline (what here we called "[Jit
+compiled code N]") we write the relationship between the compiled code and the
+Python function that is associated with it. Every profiler requires this
+information in a different format. For example, the Linux "perf" profiler
+requires a file in "/tmp/perf-PID.map" (name and location not configurable)
+with the following format:
+
+    <compiled code address> <compiled code size> <name of the compiled code>
+
+If this file is available when "perf" generates reports, it will automatically
+associate every trampoline with the Python function that it is associated with
+allowing it to generate reports that include Python information. These reports
+then can also be filtered in a way that *only* Python information appears.
+
+Notice that for this to work, there must be a unique copied of the trampoline
+per Python code object even if the code in the trampoline is the same. To
+achieve this we have a assembly template in Objects/asm_trampiline.S that is
+compiled into the Python executable/shared library. This template generates a
+symbol that maps the start of the assembly code and another that marks the end
+of the assembly code for the trampoline.  Then, every time we need a unique
+trampoline for a Python code object, we copy the assembly code into a mmaped
+area that has executable permissions and we return the start of that area as
+our trampoline function.
+
+Asking for a mmap-ed memory area for trampoline is very wasteful so we
+allocate big arenas of memory in a single mmap call, we populate the entire
+arena with copies of the trampoline (this allows us to now have to invalidate
+the icache for the instructions in the page) and then we return the next
+available chunk every time someone asks for a new trampoline. We keep a linked
+list of arenas in case the current memory arena is exhausted and another one is
+needed.
+
+For the best results, Python should be compiled with
+CFLAGS="-fno-omit-frame-pointer -mno-omit-leaf-frame-pointer" as this allows
+profilers to unwind using only the frame pointer and not on DWARF debug
+information (note that as trampilines are dynamically generated there won't be
+any DWARF information available for them).
+*/
+
+#include "Python.h"
+#include "pycore_ceval.h"
+#include "pycore_frame.h"
+#include "pycore_interp.h"
+
+typedef enum {
+    PERF_STATUS_FAILED = -1,  // Perf trampoline is in an invalid state
+    PERF_STATUS_NO_INIT = 0,  // Perf trampoline is not initialized
+    PERF_STATUS_OK = 1,       // Perf trampoline is ready to be executed
+} perf_status_t;
+
+#ifdef PY_HAVE_PERF_TRAMPOLINE
+
+#include <fcntl.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <sys/mman.h>
+#include <sys/types.h>
+#include <unistd.h>
+
+/* The function pointer is passed as last argument. The other three arguments
+ * are passed in the same order as the function requires. This results in
+ * shorter, more efficient ASM code for trampoline.
+ */
+typedef PyObject *(*py_evaluator)(PyThreadState *, _PyInterpreterFrame *,
+                                  int throwflag);
+typedef PyObject *(*py_trampoline)(PyThreadState *, _PyInterpreterFrame *, int,
+                                   py_evaluator);
+
+extern void *_Py_trampoline_func_start;  // Start of the template of the
+                                         // assembly trampoline
+extern void *
+    _Py_trampoline_func_end;  // End of the template of the assembly trampoline
+
+struct code_arena_st {
+    char *start_addr;    // Start of the memory arena
+    char *current_addr;  // Address of the current trampoline within the arena
+    size_t size;         // Size of the memory arena
+    size_t size_left;    // Remaining size of the memory arena
+    size_t code_size;    // Size of the code of every trampoline in the arena
+    struct code_arena_st
+        *prev;  // Pointer to the arena  or NULL if this is the first arena.
+};
+
+typedef struct code_arena_st code_arena_t;
+
+struct trampoline_api_st {
+    void* (*init_state)(void);
+    void (*write_state)(void* state, const void *code_addr,
+                        unsigned int code_size, PyCodeObject* code);
+    int (*free_state)(void* state);
+    void *state;
+};
+
+typedef struct trampoline_api_st trampoline_api_t;
+
+static perf_status_t perf_status = PERF_STATUS_NO_INIT;
+static Py_ssize_t extra_code_index = -1;
+static code_arena_t *code_arena;
+static trampoline_api_t trampoline_api;
+
+static FILE *perf_map_file;
+
+static void *
+perf_map_get_file(void)
+{
+    if (perf_map_file) {
+        return perf_map_file;
+    }
+    char filename[100];
+    pid_t pid = getpid();
+    // Location and file name of perf map is hard-coded in perf tool.
+    // Use exclusive create flag wit nofollow to prevent symlink attacks.
+    int flags = O_WRONLY | O_CREAT | O_EXCL | O_NOFOLLOW | O_CLOEXEC;
+    snprintf(filename, sizeof(filename) - 1, "/tmp/perf-%jd.map",
+             (intmax_t)pid);
+    int fd = open(filename, flags, 0600);
+    if (fd == -1) {
+        perf_status = PERF_STATUS_FAILED;
+        PyErr_SetFromErrnoWithFilename(PyExc_OSError, filename);
+        return NULL;
+    }
+    perf_map_file = fdopen(fd, "w");
+    if (!perf_map_file) {
+        perf_status = PERF_STATUS_FAILED;
+        PyErr_SetFromErrnoWithFilename(PyExc_OSError, filename);
+        close(fd);
+        return NULL;
+    }
+    return perf_map_file;
+}
+
+static int
+perf_map_close(void *state)
+{
+    FILE *fp = (FILE *)state;
+    int ret = 0;
+    if (fp) {
+        ret = fclose(fp);
+    }
+    perf_map_file = NULL;
+    perf_status = PERF_STATUS_NO_INIT;
+    return ret;
+}
+
+static void
+perf_map_write_entry(void *state, const void *code_addr,
+                         unsigned int code_size, PyCodeObject *co)
+{
+    assert(state != NULL);
+    FILE *method_file = (FILE *)state;
+    const char *entry = PyUnicode_AsUTF8(co->co_qualname);
+    if (entry == NULL) {
+        _PyErr_WriteUnraisableMsg("Failed to get qualname from code object",
+                                  NULL);
+        return;
+    }
+    const char *filename = PyUnicode_AsUTF8(co->co_filename);
+    if (filename == NULL) {
+        _PyErr_WriteUnraisableMsg("Failed to get filename from code object",
+                                  NULL);
+        return;
+    }
+    fprintf(method_file, "%p %x py::%s:%s\n", code_addr, code_size, entry,
+            filename);
+    fflush(method_file);
+}
+
+_PyPerf_Callbacks _Py_perfmap_callbacks = {
+    &perf_map_get_file,
+    &perf_map_write_entry,
+    &perf_map_close
+};
+
+static int
+new_code_arena(void)
+{
+    // non-trivial programs typically need 64 to 256 kiB.
+    size_t mem_size = 4096 * 16;
+    assert(mem_size % sysconf(_SC_PAGESIZE) == 0);
+    char *memory =
+        mmap(NULL,  // address
+             mem_size, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS,
+             -1,  // fd (not used here)
+             0);  // offset (not used here)
+    if (!memory) {
+        PyErr_SetFromErrno(PyExc_OSError);
+        _PyErr_WriteUnraisableMsg(
+            "Failed to create new mmap for perf trampoline", NULL);
+        perf_status = PERF_STATUS_FAILED;
+        return -1;
+    }
+    void *start = &_Py_trampoline_func_start;
+    void *end = &_Py_trampoline_func_end;
+    size_t code_size = end - start;
+
+    size_t n_copies = mem_size / code_size;
+    for (size_t i = 0; i < n_copies; i++) {
+        memcpy(memory + i * code_size, start, code_size * sizeof(char));
+    }
+    // Some systems may prevent us from creating executable code on the fly.
+    int res = mprotect(memory, mem_size, PROT_READ | PROT_EXEC);
+    if (res == -1) {
+        PyErr_SetFromErrno(PyExc_OSError);
+        munmap(memory, mem_size);
+        _PyErr_WriteUnraisableMsg(
+            "Failed to set mmap for perf trampoline to PROT_READ | PROT_EXEC",
+            NULL);
+        return -1;
+    }
+
+    code_arena_t *new_arena = PyMem_RawCalloc(1, sizeof(code_arena_t));
+    if (new_arena == NULL) {
+        PyErr_NoMemory();
+        munmap(memory, mem_size);
+        _PyErr_WriteUnraisableMsg("Failed to allocate new code arena struct",
+                                  NULL);
+        return -1;
+    }
+
+    new_arena->start_addr = memory;
+    new_arena->current_addr = memory;
+    new_arena->size = mem_size;
+    new_arena->size_left = mem_size;
+    new_arena->code_size = code_size;
+    new_arena->prev = code_arena;
+    code_arena = new_arena;
+    return 0;
+}
+
+static void
+free_code_arenas(void)
+{
+    code_arena_t *cur = code_arena;
+    code_arena_t *prev;
+    code_arena = NULL;  // invalid static pointer
+    while (cur) {
+        munmap(cur->start_addr, cur->size);
+        prev = cur->prev;
+        PyMem_RawFree(cur);
+        cur = prev;
+    }
+}
+
+static inline py_trampoline
+code_arena_new_code(code_arena_t *code_arena)
+{
+    py_trampoline trampoline = (py_trampoline)code_arena->current_addr;
+    code_arena->size_left -= code_arena->code_size;
+    code_arena->current_addr += code_arena->code_size;
+    return trampoline;
+}
+
+static inline py_trampoline
+compile_trampoline(void)
+{
+    if ((code_arena == NULL) ||
+        (code_arena->size_left <= code_arena->code_size)) {
+        if (new_code_arena() < 0) {
+            return NULL;
+        }
+    }
+    assert(code_arena->size_left <= code_arena->size);
+    return code_arena_new_code(code_arena);
+}
+
+static PyObject *
+py_trampoline_evaluator(PyThreadState *ts, _PyInterpreterFrame *frame,
+                        int throw)
+{
+    if (perf_status == PERF_STATUS_FAILED ||
+        perf_status == PERF_STATUS_NO_INIT) {
+        goto default_eval;
+    }
+    PyCodeObject *co = frame->f_code;
+    py_trampoline f = NULL;
+    assert(extra_code_index != -1);
+    int ret = _PyCode_GetExtra((PyObject *)co, extra_code_index, (void **)&f);
+    if (ret != 0 || f == NULL) {
+        // This is the first time we see this code object so we need
+        // to compile a trampoline for it.
+        py_trampoline new_trampoline = compile_trampoline();
+        if (new_trampoline == NULL) {
+            goto default_eval;
+        }
+        trampoline_api.write_state(trampoline_api.state, new_trampoline,
+                                   code_arena->code_size, co);
+        _PyCode_SetExtra((PyObject *)co, extra_code_index,
+                         (void *)new_trampoline);
+        f = new_trampoline;
+    }
+    assert(f != NULL);
+    return f(ts, frame, throw, _PyEval_EvalFrameDefault);
+default_eval:
+    // Something failed, fall back to the default evaluator.
+    return _PyEval_EvalFrameDefault(ts, frame, throw);
+}
+#endif  // PY_HAVE_PERF_TRAMPOLINE
+
+int
+_PyIsPerfTrampolineActive(void)
+{
+#ifdef PY_HAVE_PERF_TRAMPOLINE
+    PyThreadState *tstate = _PyThreadState_GET();
+    return tstate->interp->eval_frame == py_trampoline_evaluator;
+#endif
+    return 0;
+}
+
+void
+_PyPerfTrampoline_GetCallbacks(_PyPerf_Callbacks *callbacks)
+{
+    if (callbacks == NULL) {
+        return;
+    }
+#ifdef PY_HAVE_PERF_TRAMPOLINE
+    callbacks->init_state = trampoline_api.init_state;
+    callbacks->write_state = trampoline_api.write_state;
+    callbacks->free_state = trampoline_api.free_state;
+#endif
+    return;
+}
+
+int
+_PyPerfTrampoline_SetCallbacks(_PyPerf_Callbacks *callbacks)
+{
+    if (callbacks == NULL) {
+        return -1;
+    }
+#ifdef PY_HAVE_PERF_TRAMPOLINE
+    if (trampoline_api.state) {
+        _PyPerfTrampoline_Fini();
+    }
+    trampoline_api.init_state = callbacks->init_state;
+    trampoline_api.write_state = callbacks->write_state;
+    trampoline_api.free_state = callbacks->free_state;
+    trampoline_api.state = NULL;
+    perf_status = PERF_STATUS_OK;
+#endif
+    return 0;
+}
+
+int
+_PyPerfTrampoline_Init(int activate)
+{
+#ifdef PY_HAVE_PERF_TRAMPOLINE
+    PyThreadState *tstate = _PyThreadState_GET();
+    if (tstate->interp->eval_frame &&
+        tstate->interp->eval_frame != py_trampoline_evaluator) {
+        PyErr_SetString(PyExc_RuntimeError,
+                        "Trampoline cannot be initialized as a custom eval "
+                        "frame is already present");
+        return -1;
+    }
+    if (!activate) {
+        tstate->interp->eval_frame = NULL;
+    }
+    else {
+        tstate->interp->eval_frame = py_trampoline_evaluator;
+        if (new_code_arena() < 0) {
+            return -1;
+        }
+        if (trampoline_api.state == NULL) {
+            void *state = trampoline_api.init_state();
+            if (state == NULL) {
+                return -1;
+            }
+            trampoline_api.state = state;
+        }
+        extra_code_index = _PyEval_RequestCodeExtraIndex(NULL);
+        if (extra_code_index == -1) {
+            return -1;
+        }
+        perf_status = PERF_STATUS_OK;
+    }
+#endif
+    return 0;
+}
+
+int
+_PyPerfTrampoline_Fini(void)
+{
+#ifdef PY_HAVE_PERF_TRAMPOLINE
+    PyThreadState *tstate = _PyThreadState_GET();
+    if (tstate->interp->eval_frame == py_trampoline_evaluator) {
+        tstate->interp->eval_frame = NULL;
+    }
+    free_code_arenas();
+    if (trampoline_api.state != NULL) {
+        trampoline_api.free_state(trampoline_api.state);
+        trampoline_api.state = NULL;
+    }
+    extra_code_index = -1;
+#endif
+    return 0;
+}
+
+PyStatus
+_PyPerfTrampoline_AfterFork_Child(void)
+{
+#ifdef PY_HAVE_PERF_TRAMPOLINE
+    // Restart trampoline in file in child.
+    int was_active = _PyIsPerfTrampolineActive();
+    _PyPerfTrampoline_Fini();
+    if (was_active) {
+        _PyPerfTrampoline_Init(1);
+    }
+#endif
+    return PyStatus_Ok();
+}

PCbuild/_freeze_module.vcxproj

@@ -129,6 +129,7 @@
     <ClCompile Include="..\Objects\cellobject.c" />
     <ClCompile Include="..\Objects\classobject.c" />
     <ClCompile Include="..\Objects\codeobject.c" />
+    <ClCompile Include="..\Objects\perf_trampoline.c" />
     <ClCompile Include="..\Objects\complexobject.c" />
     <ClCompile Include="..\Objects\descrobject.c" />
     <ClCompile Include="..\Objects\dictobject.c" />

PCbuild/_freeze_module.vcxproj.filters

@@ -85,6 +85,9 @@
     <ClCompile Include="..\Objects\codeobject.c">
       <Filter>Source Files</Filter>
     </ClCompile>
+    <ClCompile Include="..\Objects\perf_trampoline.c">
+      <Filter>Source Files</Filter>
+    </ClCompile>
     <ClCompile Include="..\Python\compile.c">
       <Filter>Source Files</Filter>
     </ClCompile>

PCbuild/pythoncore.vcxproj

@@ -429,6 +429,7 @@
     <ClCompile Include="..\Objects\cellobject.c" />
     <ClCompile Include="..\Objects\classobject.c" />
     <ClCompile Include="..\Objects\codeobject.c" />
+    <ClCompile Include="..\Objects\perf_trampoline.c" />
     <ClCompile Include="..\Objects\complexobject.c" />
     <ClCompile Include="..\Objects\descrobject.c" />
     <ClCompile Include="..\Objects\dictobject.c" />

PCbuild/pythoncore.vcxproj.filters

@@ -923,6 +923,9 @@
     <ClCompile Include="..\Objects\codeobject.c">
       <Filter>Objects</Filter>
     </ClCompile>
+    <ClCompile Include="..\Objects\perf_trampoline.c">
+      <Filter>Objects</Filter>
+    </ClCompile>
     <ClCompile Include="..\Objects\complexobject.c">
       <Filter>Objects</Filter>
     </ClCompile>

Python/clinic/sysmodule.c.h

@@ -1151,6 +1151,79 @@ sys_getandroidapilevel(PyObject *module, PyObject *Py_UNUSED(ignored))
 
 #endif /* defined(ANDROID_API_LEVEL) */
 
+PyDoc_STRVAR(sys_activate_stack_trampoline__doc__,
+"activate_stack_trampoline($module, backend, /)\n"
+"--\n"
+"\n"
+"Activate the perf profiler trampoline.");
+
+#define SYS_ACTIVATE_STACK_TRAMPOLINE_METHODDEF    \
+    {"activate_stack_trampoline", (PyCFunction)sys_activate_stack_trampoline, METH_O, sys_activate_stack_trampoline__doc__},
+
+static PyObject *
+sys_activate_stack_trampoline_impl(PyObject *module, const char *backend);
+
+static PyObject *
+sys_activate_stack_trampoline(PyObject *module, PyObject *arg)
+{
+    PyObject *return_value = NULL;
+    const char *backend;
+
+    if (!PyUnicode_Check(arg)) {
+        _PyArg_BadArgument("activate_stack_trampoline", "argument", "str", arg);
+        goto exit;
+    }
+    Py_ssize_t backend_length;
+    backend = PyUnicode_AsUTF8AndSize(arg, &backend_length);
+    if (backend == NULL) {
+        goto exit;
+    }
+    if (strlen(backend) != (size_t)backend_length) {
+        PyErr_SetString(PyExc_ValueError, "embedded null character");
+        goto exit;
+    }
+    return_value = sys_activate_stack_trampoline_impl(module, backend);
+
+exit:
+    return return_value;
+}
+
+PyDoc_STRVAR(sys_deactivate_stack_trampoline__doc__,
+"deactivate_stack_trampoline($module, /)\n"
+"--\n"
+"\n"
+"Dectivate the perf profiler trampoline.");
+
+#define SYS_DEACTIVATE_STACK_TRAMPOLINE_METHODDEF    \
+    {"deactivate_stack_trampoline", (PyCFunction)sys_deactivate_stack_trampoline, METH_NOARGS, sys_deactivate_stack_trampoline__doc__},
+
+static PyObject *
+sys_deactivate_stack_trampoline_impl(PyObject *module);
+
+static PyObject *
+sys_deactivate_stack_trampoline(PyObject *module, PyObject *Py_UNUSED(ignored))
+{
+    return sys_deactivate_stack_trampoline_impl(module);
+}
+
+PyDoc_STRVAR(sys_is_stack_trampoline_active__doc__,
+"is_stack_trampoline_active($module, /)\n"
+"--\n"
+"\n"
+"Returns *True* if the perf profiler trampoline is active.");
+
+#define SYS_IS_STACK_TRAMPOLINE_ACTIVE_METHODDEF    \
+    {"is_stack_trampoline_active", (PyCFunction)sys_is_stack_trampoline_active, METH_NOARGS, sys_is_stack_trampoline_active__doc__},
+
+static PyObject *
+sys_is_stack_trampoline_active_impl(PyObject *module);
+
+static PyObject *
+sys_is_stack_trampoline_active(PyObject *module, PyObject *Py_UNUSED(ignored))
+{
+    return sys_is_stack_trampoline_active_impl(module);
+}
+
 #ifndef SYS_GETWINDOWSVERSION_METHODDEF
     #define SYS_GETWINDOWSVERSION_METHODDEF
 #endif /* !defined(SYS_GETWINDOWSVERSION_METHODDEF) */
@@ -1194,4 +1267,4 @@ sys_getandroidapilevel(PyObject *module, PyObject *Py_UNUSED(ignored))
 #ifndef SYS_GETANDROIDAPILEVEL_METHODDEF
     #define SYS_GETANDROIDAPILEVEL_METHODDEF
 #endif /* !defined(SYS_GETANDROIDAPILEVEL_METHODDEF) */
-/*[clinic end generated code: output=322fb0409e376ad4 input=a9049054013a1b77]*/
+/*[clinic end generated code: output=43b44240211afe95 input=a9049054013a1b77]*/

Python/initconfig.c

@@ -118,6 +118,11 @@ The following implementation-specific options are available:\n\
    files are desired as well as suppressing the extra visual location indicators \n\
    when the interpreter displays tracebacks.\n\
 \n\
+-X perf: activate support for the Linux \"perf\" profiler by activating the \"perf\"\n\
+    trampoline. When this option is activated, the Linux \"perf\" profiler will be \n\
+    able to report Python calls. This option is only available on some platforms and will \n\
+    do nothing if is not supported on the current system. The default value is \"off\".\n\
+\n\
 -X frozen_modules=[on|off]: whether or not frozen modules should be used.\n\
    The default is \"on\" (or \"off\" if you are running a local build).";
 
@@ -745,6 +750,7 @@ _PyConfig_InitCompatConfig(PyConfig *config)
     config->use_hash_seed = -1;
     config->faulthandler = -1;
     config->tracemalloc = -1;
+    config->perf_profiling = -1;
     config->module_search_paths_set = 0;
     config->parse_argv = 0;
     config->site_import = -1;
@@ -829,6 +835,7 @@ PyConfig_InitIsolatedConfig(PyConfig *config)
     config->use_hash_seed = 0;
     config->faulthandler = 0;
     config->tracemalloc = 0;
+    config->perf_profiling = 0;
     config->safe_path = 1;
     config->pathconfig_warnings = 0;
 #ifdef MS_WINDOWS
@@ -940,6 +947,7 @@ _PyConfig_Copy(PyConfig *config, const PyConfig *config2)
     COPY_ATTR(_install_importlib);
     COPY_ATTR(faulthandler);
     COPY_ATTR(tracemalloc);
+    COPY_ATTR(perf_profiling);
     COPY_ATTR(import_time);
     COPY_ATTR(code_debug_ranges);
     COPY_ATTR(show_ref_count);
@@ -1050,6 +1058,7 @@ _PyConfig_AsDict(const PyConfig *config)
     SET_ITEM_UINT(hash_seed);
     SET_ITEM_INT(faulthandler);
     SET_ITEM_INT(tracemalloc);
+    SET_ITEM_INT(perf_profiling);
     SET_ITEM_INT(import_time);
     SET_ITEM_INT(code_debug_ranges);
     SET_ITEM_INT(show_ref_count);
@@ -1331,6 +1340,7 @@ _PyConfig_FromDict(PyConfig *config, PyObject *dict)
     CHECK_VALUE("hash_seed", config->hash_seed <= MAX_HASH_SEED);
     GET_UINT(faulthandler);
     GET_UINT(tracemalloc);
+    GET_UINT(perf_profiling);
     GET_UINT(import_time);
     GET_UINT(code_debug_ranges);
     GET_UINT(show_ref_count);
@@ -1687,6 +1697,26 @@ config_read_env_vars(PyConfig *config)
     return _PyStatus_OK();
 }
 
+static PyStatus
+config_init_perf_profiling(PyConfig *config)
+{
+    int active = 0;
+    const char *env = config_get_env(config, "PYTHONPERFSUPPORT");
+    if (env) {
+        if (_Py_str_to_int(env, &active) != 0) {
+            active = 0;
+        }
+        if (active) {
+            config->perf_profiling = 1;
+        }
+    }
+    const wchar_t *xoption = config_get_xoption(config, L"perf");
+    if (xoption) {
+        config->perf_profiling = 1;
+    }
+    return _PyStatus_OK();
+
+}
 
 static PyStatus
 config_init_tracemalloc(PyConfig *config)
@@ -1788,6 +1818,12 @@ config_read_complex_options(PyConfig *config)
             return status;
         }
     }
+    if (config->perf_profiling < 0) {
+        status = config_init_perf_profiling(config);
+        if (_PyStatus_EXCEPTION(status)) {
+            return status;
+        }
+    }
 
     if (config->pycache_prefix == NULL) {
         status = config_init_pycache_prefix(config);
@@ -2104,6 +2140,9 @@ config_read(PyConfig *config, int compute_path_config)
     if (config->tracemalloc < 0) {
         config->tracemalloc = 0;
     }
+    if (config->perf_profiling < 0) {
+        config->perf_profiling = 0;
+    }
     if (config->use_hash_seed < 0) {
         config->use_hash_seed = 0;
         config->hash_seed = 0;

Python/pylifecycle.c

@@ -1149,6 +1149,16 @@ init_interp_main(PyThreadState *tstate)
         if (_PyTraceMalloc_Init(config->tracemalloc) < 0) {
             return _PyStatus_ERR("can't initialize tracemalloc");
         }
+
+
+#ifdef PY_HAVE_PERF_TRAMPOLINE
+        if (config->perf_profiling) {
+            if (_PyPerfTrampoline_SetCallbacks(&_Py_perfmap_callbacks) < 0 ||
+                    _PyPerfTrampoline_Init(config->perf_profiling) < 0) {
+                return _PyStatus_ERR("can't initialize the perf trampoline");
+            }
+        }
+#endif
     }
 
     status = init_sys_streams(tstate);
@@ -1723,6 +1733,7 @@ finalize_interp_clear(PyThreadState *tstate)
         _PyArg_Fini();
         _Py_ClearFileSystemEncoding();
         _Py_Deepfreeze_Fini();
+        _PyPerfTrampoline_Fini();
     }
 
     finalize_interp_types(tstate->interp);

Python/sysmodule.c

@@ -2053,6 +2053,80 @@ sys_getandroidapilevel_impl(PyObject *module)
 }
 #endif   /* ANDROID_API_LEVEL */
 
+/*[clinic input]
+sys.activate_stack_trampoline
+
+    backend: str
+    /
+
+Activate the perf profiler trampoline.
+[clinic start generated code]*/
+
+static PyObject *
+sys_activate_stack_trampoline_impl(PyObject *module, const char *backend)
+/*[clinic end generated code: output=5783cdeb51874b43 input=b09020e3a17c78c5]*/
+{
+#ifdef PY_HAVE_PERF_TRAMPOLINE
+    if (strcmp(backend, "perf") == 0) {
+        _PyPerf_Callbacks cur_cb;
+        _PyPerfTrampoline_GetCallbacks(&cur_cb);
+        if (cur_cb.init_state != _Py_perfmap_callbacks.init_state) {
+            if (_PyPerfTrampoline_SetCallbacks(&_Py_perfmap_callbacks) < 0 ) {
+                PyErr_SetString(PyExc_ValueError, "can't activate perf trampoline");
+                return NULL;
+            }
+        }
+    }
+    else {
+        PyErr_Format(PyExc_ValueError, "invalid backend: %s", backend);
+        return NULL;
+    }
+    if (_PyPerfTrampoline_Init(1) < 0) {
+        return NULL;
+    }
+    Py_RETURN_NONE;
+#else
+    PyErr_SetString(PyExc_ValueError, "perf trampoline not available");
+    return NULL;
+#endif
+}
+
+
+/*[clinic input]
+sys.deactivate_stack_trampoline
+
+Dectivate the perf profiler trampoline.
+[clinic start generated code]*/
+
+static PyObject *
+sys_deactivate_stack_trampoline_impl(PyObject *module)
+/*[clinic end generated code: output=b50da25465df0ef1 input=491f4fc1ed615736]*/
+{
+    if  (_PyPerfTrampoline_Init(0) < 0) {
+        return NULL;
+    }
+    Py_RETURN_NONE;
+}
+
+/*[clinic input]
+sys.is_stack_trampoline_active
+
+Returns *True* if the perf profiler trampoline is active.
+[clinic start generated code]*/
+
+static PyObject *
+sys_is_stack_trampoline_active_impl(PyObject *module)
+/*[clinic end generated code: output=ab2746de0ad9d293 input=061fa5776ac9dd59]*/
+{
+#ifdef PY_HAVE_PERF_TRAMPOLINE
+    if (_PyIsPerfTrampolineActive()) {
+        Py_RETURN_TRUE;
+    }
+#endif
+    Py_RETURN_FALSE;
+}
+
+
 static PyMethodDef sys_methods[] = {
     /* Might as well keep this in alphabetic order */
     SYS_ADDAUDITHOOK_METHODDEF
@@ -2108,6 +2182,9 @@ static PyMethodDef sys_methods[] = {
      METH_VARARGS | METH_KEYWORDS, set_asyncgen_hooks_doc},
     SYS_GET_ASYNCGEN_HOOKS_METHODDEF
     SYS_GETANDROIDAPILEVEL_METHODDEF
+    SYS_ACTIVATE_STACK_TRAMPOLINE_METHODDEF
+    SYS_DEACTIVATE_STACK_TRAMPOLINE_METHODDEF
+    SYS_IS_STACK_TRAMPOLINE_ACTIVE_METHODDEF
     SYS_UNRAISABLEHOOK_METHODDEF
 #ifdef Py_STATS
     SYS__STATS_ON_METHODDEF

configure

@@ -861,6 +861,7 @@ LIBEXPAT_CFLAGS
 TZPATH
 LIBUUID_LIBS
 LIBUUID_CFLAGS
+PERF_TRAMPOLINE_OBJ
 SHLIBS
 CFLAGSFORSHARED
 LINKFORSHARED
@@ -11498,6 +11499,35 @@ esac
 { $as_echo "$as_me:${as_lineno-$LINENO}: result: $SHLIBS" >&5
 $as_echo "$SHLIBS" >&6; }
 
+{ $as_echo "$as_me:${as_lineno-$LINENO}: checking perf trampoline" >&5
+$as_echo_n "checking perf trampoline... " >&6; }
+case $PLATFORM_TRIPLET in #(
+  x86_64-linux-gnu) :
+    perf_trampoline=yes ;; #(
+  aarch64-linux-gnu) :
+    perf_trampoline=yes ;; #(
+  *) :
+    perf_trampoline=no
+ ;;
+esac
+{ $as_echo "$as_me:${as_lineno-$LINENO}: result: $perf_trampoline" >&5
+$as_echo "$perf_trampoline" >&6; }
+
+if test "x$perf_trampoline" = xyes; then :
+
+
+$as_echo "#define PY_HAVE_PERF_TRAMPOLINE 1" >>confdefs.h
+
+  PERF_TRAMPOLINE_OBJ=Objects/asm_trampoline.o
+
+    if test "x$Py_DEBUG" = xtrue; then :
+
+    as_fn_append BASECFLAGS " -fno-omit-frame-pointer -mno-omit-leaf-frame-pointer"
+
+fi
+
+fi
+
 
 # checks for libraries
 { $as_echo "$as_me:${as_lineno-$LINENO}: checking for sendfile in -lsendfile" >&5

configure.ac

@@ -3452,6 +3452,26 @@ case "$ac_sys_system" in
 esac
 AC_MSG_RESULT($SHLIBS)
 
+dnl perf trampoline is Linux specific and requires an arch-specific
+dnl trampoline in asssembly.
+AC_MSG_CHECKING([perf trampoline])
+AS_CASE([$PLATFORM_TRIPLET],
+  [x86_64-linux-gnu], [perf_trampoline=yes],
+  [aarch64-linux-gnu], [perf_trampoline=yes],
+  [perf_trampoline=no]
+)
+AC_MSG_RESULT([$perf_trampoline])
+
+AS_VAR_IF([perf_trampoline], [yes], [
+  AC_DEFINE([PY_HAVE_PERF_TRAMPOLINE], [1], [Define to 1 if you have the perf trampoline.])
+  PERF_TRAMPOLINE_OBJ=Objects/asm_trampoline.o
+
+  dnl perf needs frame pointers for unwinding, include compiler option in debug builds
+  AS_VAR_IF([Py_DEBUG], [true], [
+    AS_VAR_APPEND([BASECFLAGS], [" -fno-omit-frame-pointer -mno-omit-leaf-frame-pointer"])
+  ])
+])
+AC_SUBST([PERF_TRAMPOLINE_OBJ])
 
 # checks for libraries
 AC_CHECK_LIB(sendfile, sendfile)

pyconfig.h.in

@@ -1568,6 +1568,9 @@
 /* Define if you want to coerce the C locale to a UTF-8 based locale */
 #undef PY_COERCE_C_LOCALE
 
+/* Define to 1 if you have the perf trampoline. */
+#undef PY_HAVE_PERF_TRAMPOLINE
+
 /* Define to 1 to build the sqlite module with loadable extensions support. */
 #undef PY_SQLITE_ENABLE_LOAD_EXTENSION