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? |
---|---|---|
|
A standard debug build. |
Yes |
|
Turn on aggregate reference counting which will be
displayed in the interactive interpreter when
invoked with |
No |
|
Turns on reference tracing.
Sets |
Yes |
|
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 |
|
Enables Pythons small memory allocator. For Valgrind
this must be disabled, if using Pythons malloc
debugger (using |
No |
|
Enables Python’s malloc debugger that annotates
memory blocks. Requires |
No |
Here is the description of other debug macros that are set by one of the macros above:
Macro |
Description |
---|---|
|
Low level tracing. See |
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:
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 | >>> 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 |
---|---|
|
The number of times an object of this type was allocated. |
|
The number of times an object of this type was freed. |
|
The maximum seen value of |
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:
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 | $ ./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)