10.2. Building and Using a Debug Version of Python¶

There is a spectrum of debug builds of Python that you can create. This chapter describes how to create them.

10.2.1. Building a Standard Debug Version of Python¶

Download and unpack the Python source. Then in the source directory create a debug directory for the debug build:

mkdir debug
cd debug
../configure --with-pydebug
make
make test

10.2.2. Specifying Macros¶

They can be specified at the configure stage, this works:

../configure CFLAGS='-DPy_DEBUG -DPy_TRACE_REFS' --with-pydebug
make

However the python documentation suggests the alternative way of specifying them when invoking make:

../configure --with-pydebug
make EXTRA_CFLAGS="-DPy_REF_DEBUG"

I don’t know why one way would be regarded as better than the other.

10.2.3. The Debug Builds¶

The builds are controlled by the following macros:

Macro	Description	Must Rebuild Extensions?
`Py_DEBUG`	A standard debug build. `Py_DEBUG` sets `LLTRACE`, `Py_REF_DEBUG`, `Py_TRACE_REFS`, and `PYMALLOC_DEBUG` (if `WITH_PYMALLOC` is enabled).	Yes
`Py_REF_DEBUG`	Turn on aggregate reference counting which will be displayed in the interactive interpreter when invoked with `-X showrefcount` on the command line. If you are not keeping references to objects and the count is increasing there is probably a leak. Also adds `sys.gettotalrefcount()` to the `sys` module and this returns the total number of references.	No
`Py_TRACE_REFS`	Turns on reference tracing. Sets `Py_REF_DEBUG`.	Yes
`COUNT_ALLOCS`	Keeps track of the number of objects of each type have been allocated and how many freed. See: Python Debug build with COUNT_ALLOCS	Yes
`WITH_PYMALLOC`	Enables Pythons small memory allocator. For Valgrind this must be disabled, if using Pythons malloc debugger (using `PYMALLOC_DEBUG`) this must be enabled. See: Python’s Memory Allocator	No
`PYMALLOC_DEBUG`	Enables Python’s malloc debugger that annotates memory blocks. Requires `WITH_PYMALLOC`. See: Python’s Memory Allocator	No

Here is the description of other debug macros that are set by one of the macros above:

Macro	Description
`LLTRACE`	Low level tracing. See `Python/ceval.c`.

In the source directory:

mkdir debug
cd debug
../configure --with-pydebug
make
make test

10.2.4. Python’s Memory Allocator¶

A normal build of Python gives CPython a special memory allocator ‘PyMalloc’. When enabled this mallocs largish chunks of memory from the OS and then uses this pool for the actual PyObjects. With PyMalloc active Valgrind can not see all allocations and deallocations.

There are two Python builds of interest to help solve memory problems:

Disable PyMalloc so that Valgrind can analyse the memory usage.
Enable PyMalloc in debug mode, this creates memory blocks with special bit patterns and adds debugging information on each end of any dynamically allocated memory. This pattern is checked on every alloc/free and if found to be corrupt a diagnostic is printed and the process terminated.

To make a version of Python with its memory allocator suitable for use with Valgrind:

../configure --with-pydebug --without-pymalloc
make

See Using Valgrind for using Valgrind.

To make a version of Python with its memory allocator using Python’s malloc debugger either:

../configure CFLAGS='-DPYMALLOC_DEBUG' --with-pydebug
make

Or:

../configure --with-pydebug
make EXTRA_CFLAGS="-DPYMALLOC_DEBUG"

This builds Python with the WITH_PYMALLOC and PYMALLOC_DEBUG macros defined.

10.2.4.1. Finding Access after Free With `PYMALLOC_DEBUG`¶

Python built with PYMALLOC_DEBUG is the most effective way of detecting access after free. For example if we have this CPython code:

static PyObject *access_after_free(PyObject *pModule) {
    PyObject *pA = PyLong_FromLong(1024L);
    Py_DECREF(pA);
    PyObject_Print(pA, stdout, 0);
    Py_RETURN_NONE;
}

And we call this from the interpreter we get a diagnostic:

Python 3.4.3 (default, Sep 16 2015, 16:56:10)
[GCC 4.2.1 Compatible Apple LLVM 6.0 (clang-600.0.51)] on darwin
Type "help", "copyright", "credits" or "license" for more information.
>>> import cPyRefs
>>> cPyRefs.afterFree()
<refcnt -2604246222170760229 at 0x10a474130>
>>>

10.2.4.2. Getting Statistics on PyMalloc¶

If the environment variable PYTHONMALLOCSTATS exists when running Python built with WITH_PYMALLOC``+``PYMALLOC_DEBUG then a (detailed) report of pymalloc activity is output on stderr whenever a new ‘arena’ is allocated.

PYTHONMALLOCSTATS=1 python.exe

I have no special knowledge about the output you see when running Python this way which looks like this:

>>> cPyRefs.leakNewRefs(1000, 10000)
loose_new_reference: value=1000 count=10000
Small block threshold = 512, in 64 size classes.

class   size   num pools   blocks in use  avail blocks
-----   ----   ---------   -------------  ------------
    4     40           2             139            63
    5     48           1               2            82
  ...
   62    504           3              21             3
   63    512           3              18             3

# times object malloc called       =            2,042,125
# arenas allocated total           =                  636
# arenas reclaimed                 =                    1
# arenas highwater mark            =                  635
# arenas allocated current         =                  635
635 arenas * 262144 bytes/arena    =          166,461,440

# bytes in allocated blocks        =          162,432,624
# bytes in available blocks        =              116,824
0 unused pools * 4096 bytes        =                    0
# bytes lost to pool headers       =            1,950,720
# bytes lost to quantization       =            1,961,272
# bytes lost to arena alignment    =                    0
Total                              =          166,461,440
Small block threshold = 512, in 64 size classes.

class   size   num pools   blocks in use  avail blocks
-----   ----   ---------   -------------  ------------
    4     40           2             139            63
    5     48           1               2            82
  ...
   62    504           3              21             3
   63    512           3              18             3

# times object malloc called       =            2,045,325
# arenas allocated total           =                  637
# arenas reclaimed                 =                    1
# arenas highwater mark            =                  636
# arenas allocated current         =                  636
636 arenas * 262144 bytes/arena    =          166,723,584

# bytes in allocated blocks        =          162,688,624
# bytes in available blocks        =              116,824
0 unused pools * 4096 bytes        =                    0
# bytes lost to pool headers       =            1,953,792
# bytes lost to quantization       =            1,964,344
# bytes lost to arena alignment    =                    0
Total                              =          166,723,584
Small block threshold = 512, in 64 size classes.

class   size   num pools   blocks in use  avail blocks
-----   ----   ---------   -------------  ------------
    4     40           2             139            63
    5     48           1               2            82
  ...
   62    504           3              21             3
   63    512           3              18             3

# times object malloc called       =            2,048,525
# arenas allocated total           =                  638
# arenas reclaimed                 =                    1
# arenas highwater mark            =                  637
# arenas allocated current         =                  637
637 arenas * 262144 bytes/arena    =          166,985,728

# bytes in allocated blocks        =          162,944,624
# bytes in available blocks        =              116,824
0 unused pools * 4096 bytes        =                    0
# bytes lost to pool headers       =            1,956,864
# bytes lost to quantization       =            1,967,416
# bytes lost to arena alignment    =                    0
Total                              =          166,985,728
loose_new_reference: DONE

10.2.5. Python Debug build with `COUNT_ALLOCS`¶

A Python debug build with COUNT_ALLOCS give some additional information about each object type (not the individual objects themselves). A PyObject grows some extra fields that track the reference counts for that type. The fields are:

Field	Description
`tp_allocs`	The number of times an object of this type was allocated.
`tp_frees`	The number of times an object of this type was freed.
`tp_maxalloc`	The maximum seen value of `tp_allocs - tp_frees` so this is the maximum count of this type allocated at the same time.

The sys module also gets an extra function sys.getcounts() that returns a list of tuples: [(tp_typename, tp_allocs, tp_frees, tp_maxalloc), ...].

10.2.5.1. Building the Python Executable with `COUNT_ALLOCS`¶

Either:

../configure CFLAGS='-DCOUNT_ALLOCS' --with-pydebug
make

Or:

../configure --with-pydebug
make EXTRA_CFLAGS="-DCOUNT_ALLOCS"

Warning

When using COUNT_ALLOCS any Python extensions now need to be rebuilt with this Python executable as it fundementally changes the structure of a PyObject.

10.2.5.2. Using the Python Executable with `COUNT_ALLOCS`¶

An example of using this build is here: Observing the Reference Counts for a Particular Type

10.2.6. Identifying the Python Build Configuration from the Runtime¶

The module sysconfig allows you access to the configuration of the Python runtime. At its simplest, and most verbose, this can be used thus:

$ ./python.exe -m sysconfig
Platform: "macosx-10.9-x86_64"
Python version: "3.4"
Current installation scheme: "posix_prefix"

Paths:
    data = "/usr/local"
    ...
    stdlib = "/usr/local/lib/python3.4"

Variables:
    ABIFLAGS = "dm"
    AC_APPLE_UNIVERSAL_BUILD = "0"
    AIX_GENUINE_CPLUSPLUS = "0"
    AR = "ar"
    ...
    py_version = "3.4.3"
    py_version_nodot = "34"
    py_version_short = "3.4"

Importing sysconfig into an interpreter session gives two useful functions are get_config_var(...) which gets the setting for a particular macro and get_config_vars() which gets a dict of {macro : value, ...}. For example:

>>> import sysconfig
>>> sysconfig.get_config_var('Py_DEBUG')
1

For advanced usage you can parse any pyconfig.h into a dict by opening that file and passing it to sysconfig.parse_config_h(f) as a file object. sysconfig.get_config_h_filename() will give you the configuration file for the runtime (assuming it still exists). So:

>>> with open(sysconfig.get_config_h_filename()) as f:
      cfg = sysconfig.parse_config_h(f)