Can I get the python call stack with the linux perf?

Question

For example,

    def test():
        print "test"

I used perf record -g -p $pid, but the result was just all about PyEval_EvalFrameEx. How can I get the real name "test" or if can not by using perf?

My python program will call some c++ extensions, so I use perf to check the call stack for c++ extensions, also I want to get the python. — wuwl, Nov 13 '14 at 07:49

maxschlepzig · Answer 1 · 2021-08-22T16:24:12.107

As of 2018, perf simply doesn't have support for reading the Python stack frames (cf. a 2014 Python mailinglist discussion).

Python 3.6 has some support for Dtrace and Systemtap.

An alternative to this is Pyflame, a stochastic profiler for Python that samples python call stacks via ptrace(). In contrast to Dtrace/Systemtap you don't need extra permissions and it also works with Python versions that are compiled without instrumentalization support.

When you use the --threads option with Pyflame you see Python lines that call into C/C++ extensions, although the stack-trace stops at the last Python frame. But perhaps this is sufficient for your use case.

Edit: Pyflame was abandoned in the end of 2019 or so. A hacker news thread mentions the following alternatives:

score 4 · Answer 2 · answered Dec 19 '16 at 22:50

You won't be able to do this with perf, that's specifically built to interface the Linux process model, decode those stack frames, etc. It's doing what it's supposed to by telling you that it was executing the function PyEval_EvalFrameEx. It would have to be extended with python-specific information to be able to actually decode Python's frame information, which isn't going to happen. Unfortunately I haven't found a really good way to debug both Python and C/C++ modules easily. It's generally pdb for one and gdb for the other.

score 2 · Answer 3 · answered Mar 14 '23 at 18:06

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.

https://docs.python.org/3.12/howto/perf_profiling.html