Doc/howto/instrumentation.rst

.. highlight:: shell-session

.. _instrumentation:

===============================================
Instrumenting CPython with DTrace and SystemTap
===============================================

:author: David Malcolm
:author: Łukasz Langa

DTrace and SystemTap are monitoring tools, each providing a way to inspect
what the processes on a computer system are doing.  They both use
domain-specific languages allowing a user to write scripts which:

  - filter which processes are to be observed
  - gather data from the processes of interest
  - generate reports on the data

As of Python 3.6, CPython can be built with embedded "markers", also
known as "probes", that can be observed by a DTrace or SystemTap script,
making it easier to monitor what the CPython processes on a system are
doing.

.. impl-detail::

   DTrace markers are implementation details of the CPython interpreter.
   No guarantees are made about probe compatibility between versions of
   CPython. DTrace scripts can stop working or work incorrectly without
   warning when changing CPython versions.


Enabling the static markers
---------------------------

macOS comes with built-in support for DTrace.  On Linux, in order to
build CPython with the embedded markers for SystemTap, the SystemTap
development tools must be installed.

On a Linux machine, this can be done via::

   $ yum install systemtap-sdt-devel

or::

   $ sudo apt-get install systemtap-sdt-dev


CPython must then be :option:`configured with the --with-dtrace option
<--with-dtrace>`:

.. code-block:: none

   checking for --with-dtrace... yes

On macOS, you can list available DTrace probes by running a Python
process in the background and listing all probes made available by the
Python provider::

   $ python3.6 -q &
   $ sudo dtrace -l -P python$!  # or: dtrace -l -m python3.6

      ID   PROVIDER            MODULE                          FUNCTION NAME
   29564 python18035        python3.6          _PyEval_EvalFrameDefault function-entry
   29565 python18035        python3.6             dtrace_function_entry function-entry
   29566 python18035        python3.6          _PyEval_EvalFrameDefault function-return
   29567 python18035        python3.6            dtrace_function_return function-return
   29568 python18035        python3.6                           collect gc-done
   29569 python18035        python3.6                           collect gc-start
   29570 python18035        python3.6          _PyEval_EvalFrameDefault line
   29571 python18035        python3.6                 maybe_dtrace_line line

On Linux, you can verify if the SystemTap static markers are present in
the built binary by seeing if it contains a ".note.stapsdt" section.

::

   $ readelf -S ./python | grep .note.stapsdt
   [30] .note.stapsdt        NOTE         0000000000000000 00308d78

If you've built Python as a shared library
(with the :option:`--enable-shared` configure option), you
need to look instead within the shared library.  For example::

   $ readelf -S libpython3.3dm.so.1.0 | grep .note.stapsdt
   [29] .note.stapsdt        NOTE         0000000000000000 00365b68

Sufficiently modern readelf can print the metadata::

    $ readelf -n ./python

    Displaying notes found at file offset 0x00000254 with length 0x00000020:
        Owner                 Data size          Description
        GNU                  0x00000010          NT_GNU_ABI_TAG (ABI version tag)
            OS: Linux, ABI: 2.6.32

    Displaying notes found at file offset 0x00000274 with length 0x00000024:
        Owner                 Data size          Description
        GNU                  0x00000014          NT_GNU_BUILD_ID (unique build ID bitstring)
            Build ID: df924a2b08a7e89f6e11251d4602022977af2670

    Displaying notes found at file offset 0x002d6c30 with length 0x00000144:
        Owner                 Data size          Description
        stapsdt              0x00000031          NT_STAPSDT (SystemTap probe descriptors)
            Provider: python
            Name: gc__start
            Location: 0x00000000004371c3, Base: 0x0000000000630ce2, Semaphore: 0x00000000008d6bf6
            Arguments: -4@%ebx
        stapsdt              0x00000030          NT_STAPSDT (SystemTap probe descriptors)
            Provider: python
            Name: gc__done
            Location: 0x00000000004374e1, Base: 0x0000000000630ce2, Semaphore: 0x00000000008d6bf8
            Arguments: -8@%rax
        stapsdt              0x00000045          NT_STAPSDT (SystemTap probe descriptors)
            Provider: python
            Name: function__entry
            Location: 0x000000000053db6c, Base: 0x0000000000630ce2, Semaphore: 0x00000000008d6be8
            Arguments: 8@%rbp 8@%r12 -4@%eax
        stapsdt              0x00000046          NT_STAPSDT (SystemTap probe descriptors)
            Provider: python
            Name: function__return
            Location: 0x000000000053dba8, Base: 0x0000000000630ce2, Semaphore: 0x00000000008d6bea
            Arguments: 8@%rbp 8@%r12 -4@%eax

The above metadata contains information for SystemTap describing how it
can patch strategically-placed machine code instructions to enable the
tracing hooks used by a SystemTap script.


Static DTrace probes
--------------------

The following example DTrace script can be used to show the call/return
hierarchy of a Python script, only tracing within the invocation of
a function called "start". In other words, import-time function
invocations are not going to be listed:

.. code-block:: none

    self int indent;

    python$target:::function-entry
    /copyinstr(arg1) == "start"/
    {
            self->trace = 1;
    }

    python$target:::function-entry
    /self->trace/
    {
            printf("%d\t%*s:", timestamp, 15, probename);
            printf("%*s", self->indent, "");
            printf("%s:%s:%d\n", basename(copyinstr(arg0)), copyinstr(arg1), arg2);
            self->indent++;
    }

    python$target:::function-return
    /self->trace/
    {
            self->indent--;
            printf("%d\t%*s:", timestamp, 15, probename);
            printf("%*s", self->indent, "");
            printf("%s:%s:%d\n", basename(copyinstr(arg0)), copyinstr(arg1), arg2);
    }

    python$target:::function-return
    /copyinstr(arg1) == "start"/
    {
            self->trace = 0;
    }

It can be invoked like this::

  $ sudo dtrace -q -s call_stack.d -c "python3.6 script.py"

The output looks like this:

.. code-block:: none

    156641360502280  function-entry:call_stack.py:start:23
    156641360518804  function-entry: call_stack.py:function_1:1
    156641360532797  function-entry:  call_stack.py:function_3:9
    156641360546807 function-return:  call_stack.py:function_3:10
    156641360563367 function-return: call_stack.py:function_1:2
    156641360578365  function-entry: call_stack.py:function_2:5
    156641360591757  function-entry:  call_stack.py:function_1:1
    156641360605556  function-entry:   call_stack.py:function_3:9
    156641360617482 function-return:   call_stack.py:function_3:10
    156641360629814 function-return:  call_stack.py:function_1:2
    156641360642285 function-return: call_stack.py:function_2:6
    156641360656770  function-entry: call_stack.py:function_3:9
    156641360669707 function-return: call_stack.py:function_3:10
    156641360687853  function-entry: call_stack.py:function_4:13
    156641360700719 function-return: call_stack.py:function_4:14
    156641360719640  function-entry: call_stack.py:function_5:18
    156641360732567 function-return: call_stack.py:function_5:21
    156641360747370 function-return:call_stack.py:start:28


Static SystemTap markers
------------------------

The low-level way to use the SystemTap integration is to use the static
markers directly.  This requires you to explicitly state the binary file
containing them.

For example, this SystemTap script can be used to show the call/return
hierarchy of a Python script:

.. code-block:: none

   probe process("python").mark("function__entry") {
        filename = user_string($arg1);
        funcname = user_string($arg2);
        lineno = $arg3;

        printf("%s => %s in %s:%d\\n",
               thread_indent(1), funcname, filename, lineno);
   }

   probe process("python").mark("function__return") {
       filename = user_string($arg1);
       funcname = user_string($arg2);
       lineno = $arg3;

       printf("%s <= %s in %s:%d\\n",
              thread_indent(-1), funcname, filename, lineno);
   }

It can be invoked like this::

   $ stap \
     show-call-hierarchy.stp \
     -c "./python test.py"

The output looks like this:

.. code-block:: none

   11408 python(8274):        => __contains__ in Lib/_abcoll.py:362
   11414 python(8274):         => __getitem__ in Lib/os.py:425
   11418 python(8274):          => encode in Lib/os.py:490
   11424 python(8274):          <= encode in Lib/os.py:493
   11428 python(8274):         <= __getitem__ in Lib/os.py:426
   11433 python(8274):        <= __contains__ in Lib/_abcoll.py:366

where the columns are:

  - time in microseconds since start of script

  - name of executable

  - PID of process

and the remainder indicates the call/return hierarchy as the script executes.

For a :option:`--enable-shared` build of CPython, the markers are contained within the
libpython shared library, and the probe's dotted path needs to reflect this. For
example, this line from the above example:

.. code-block:: none

   probe process("python").mark("function__entry") {

should instead read:

.. code-block:: none

   probe process("python").library("libpython3.6dm.so.1.0").mark("function__entry") {

(assuming a :ref:`debug build <debug-build>` of CPython 3.6)


Available static markers
------------------------

.. object:: function__entry(str filename, str funcname, int lineno)

   This marker indicates that execution of a Python function has begun.
   It is only triggered for pure-Python (bytecode) functions.

   The filename, function name, and line number are provided back to the
   tracing script as positional arguments, which must be accessed using
   ``$arg1``, ``$arg2``, ``$arg3``:

       * ``$arg1`` : ``(const char *)`` filename, accessible using ``user_string($arg1)``

       * ``$arg2`` : ``(const char *)`` function name, accessible using
         ``user_string($arg2)``

       * ``$arg3`` : ``int`` line number

.. object:: function__return(str filename, str funcname, int lineno)

   This marker is the converse of :c:func:`function__entry`, and indicates that
   execution of a Python function has ended (either via ``return``, or via an
   exception).  It is only triggered for pure-Python (bytecode) functions.

   The arguments are the same as for :c:func:`function__entry`

.. object:: line(str filename, str funcname, int lineno)

   This marker indicates a Python line is about to be executed.  It is
   the equivalent of line-by-line tracing with a Python profiler.  It is
   not triggered within C functions.

   The arguments are the same as for :c:func:`function__entry`.

.. object:: gc__start(int generation)

   Fires when the Python interpreter starts a garbage collection cycle.
   ``arg0`` is the generation to scan, like :func:`gc.collect()`.

.. object:: gc__done(long collected)

   Fires when the Python interpreter finishes a garbage collection
   cycle. ``arg0`` is the number of collected objects.

.. object:: import__find__load__start(str modulename)

   Fires before :mod:`importlib` attempts to find and load the module.
   ``arg0`` is the module name.

   .. versionadded:: 3.7

.. object:: import__find__load__done(str modulename, int found)

   Fires after :mod:`importlib`'s find_and_load function is called.
   ``arg0`` is the module name, ``arg1`` indicates if module was
   successfully loaded.

   .. versionadded:: 3.7


.. object:: audit(str event, void *tuple)

   Fires when :func:`sys.audit` or :c:func:`PySys_Audit` is called.
   ``arg0`` is the event name as C string, ``arg1`` is a :c:type:`PyObject`
   pointer to a tuple object.

   .. versionadded:: 3.8


SystemTap Tapsets
-----------------

The higher-level way to use the SystemTap integration is to use a "tapset":
SystemTap's equivalent of a library, which hides some of the lower-level
details of the static markers.

Here is a tapset file, based on a non-shared build of CPython:

.. code-block:: none

    /*
       Provide a higher-level wrapping around the function__entry and
       function__return markers:
     \*/
    probe python.function.entry = process("python").mark("function__entry")
    {
        filename = user_string($arg1);
        funcname = user_string($arg2);
        lineno = $arg3;
        frameptr = $arg4
    }
    probe python.function.return = process("python").mark("function__return")
    {
        filename = user_string($arg1);
        funcname = user_string($arg2);
        lineno = $arg3;
        frameptr = $arg4
    }

If this file is installed in SystemTap's tapset directory (e.g.
``/usr/share/systemtap/tapset``), then these additional probepoints become
available:

.. object:: python.function.entry(str filename, str funcname, int lineno, frameptr)

   This probe point indicates that execution of a Python function has begun.
   It is only triggered for pure-Python (bytecode) functions.

.. object:: python.function.return(str filename, str funcname, int lineno, frameptr)

   This probe point is the converse of ``python.function.return``, and
   indicates that execution of a Python function has ended (either via
   ``return``, or via an exception).  It is only triggered for pure-Python
   (bytecode) functions.


Examples
--------
This SystemTap script uses the tapset above to more cleanly implement the
example given above of tracing the Python function-call hierarchy, without
needing to directly name the static markers:

.. code-block:: none

    probe python.function.entry
    {
      printf("%s => %s in %s:%d\n",
             thread_indent(1), funcname, filename, lineno);
    }

    probe python.function.return
    {
      printf("%s <= %s in %s:%d\n",
             thread_indent(-1), funcname, filename, lineno);
    }


The following script uses the tapset above to provide a top-like view of all
running CPython code, showing the top 20 most frequently-entered bytecode
frames, each second, across the whole system:

.. code-block:: none

    global fn_calls;

    probe python.function.entry
    {
        fn_calls[pid(), filename, funcname, lineno] += 1;
    }

    probe timer.ms(1000) {
        printf("\033[2J\033[1;1H") /* clear screen \*/
        printf("%6s %80s %6s %30s %6s\n",
               "PID", "FILENAME", "LINE", "FUNCTION", "CALLS")
        foreach ([pid, filename, funcname, lineno] in fn_calls- limit 20) {
            printf("%6d %80s %6d %30s %6d\n",
                pid, filename, lineno, funcname,
                fn_calls[pid, filename, funcname, lineno]);
        }
        delete fn_calls;
    }