a static analysis tool for CPython extension code — gcc-python-plugin 0.16 documentation (original) (raw)

Docs »
Usage example: a static analysis tool for CPython extension code
Edit on GitHub

Note

This code is under heavy development, and still contains bugs. It is not unusual to see Python tracebacks when running the checker. You should verify what the checker reports before acting on it: it could be wrong.

An example of using the plugin is a static analysis tool I’m working on which checks the C source of CPython extension modules for common coding errors.

This was one of my main motivations for writing the GCC plugin, and I often need to extend the plugin to support this use case.

For this reason, the checker is embedded within the gcc-python source tree itself for now:

gcc-with-cpychecker is a harness script, which invokes GCC, adding the arguments necessary to use the Python plugin, using thelibcpychecker Python code

the libcpychecker subdirectory contains the code that does the actual work

various test cases (in the source tree, below tests/cpychecker)

gcc-with-cpychecker¶

gcc-with-cpychecker is a harness script, which invokes GCC, adding the arguments necessary to use the Python plugin, using thelibcpychecker Python code

You should be able to use the checker on arbitrary CPython extension code by replacing “gcc” with “gcc-with-cpychecker” in your build with something like:

make CC=/path/to/built/plugin/gcc-with-cpychecker

to override the Makefile variable CC.

You may need to supply an absolute path, especially if the “make” recursively invokes “make” within subdirectories (thus having a different working directory).

Similarly, for projects that use distutils, the code is typically built with an invocation like this:

This respects the environment variable CC, so typically you can replace the above with something like this in order to add the additional checks:

CC=/path/to/built/plugin/gcc-with-cpychecker python setup.py build

Additional arguments for gcc-with-cpychecker¶

--maxtrans <int>¶

Set the maximum number of transitions to consider within each function before pruning the analysis tree. You may need to increase this limit for complicated functions.

--dump-json ``¶

Dump a JSON representation of any problems. For example, given a functionfoo.c, if any warnings or errors are found in function bar, a filefoo.c.bar.json will be written out in JSON form.

Reference-count checking¶

The checker attempts to analyze all possible paths through each function, tracking the various PyObject* objects encountered.

For each path through the function and PyObject*, it determines what the reference count ought to be at the end of the function, issuing warnings for any that are incorrect.

The warnings are in two forms: the classic textual output to GCC’s standard error stream, together with an HTML report indicating the flow through the function, in graphical form.

For example, given this buggy C code:

PyObject * test(PyObject *self, PyObject *args) { PyObject *list; PyObject item; list = PyList_New(1); if (!list) return NULL; item = PyLong_FromLong(42); / This error handling is incorrect: it's missing an invocation of Py_DECREF(list): / if (!item) return NULL; / This steals a reference to item; item is not leaked when we get here: */ PyList_SetItem(list, 0, item); return list; }

the checker emits these messages to stderr:

input.c: In function 'test': input.c:38:1: warning: ob_refcnt of '*list' is 1 too high [enabled by default] input.c:38:1: note: was expecting final ob_refcnt to be N + 0 (for some unknown N) input.c:38:1: note: but final ob_refcnt is N + 1 input.c:27:10: note: PyListObject allocated at: list = PyList_New(1); input.c:27:10: note: when PyList_New() succeeds at: list = PyList_New(1); input.c:27:10: note: ob_refcnt is now refs: 1 + N where N >= 0 input.c:28:8: note: taking False path at: if (!list) input.c:30:10: note: reaching: item = PyLong_FromLong(42); input.c:30:10: note: when PyLong_FromLong() fails at: item = PyLong_FromLong(42); input.c:33:8: note: taking True path at: if (!item) input.c:34:9: note: reaching: return NULL; input.c:38:1: note: returning input.c:24:1: note: graphical error report for function 'test' written out to 'input.c.test-refcount-errors.html'

along with this HTML report (as referred to by the final line on stderr):

The HTML report is intended to be relatively self-contained, and thus easy to attach to bug tracking systems (it embeds its own CSS inline, and references the JavaScript it uses via URLs to the web).

Note

The arrow graphics in the HTML form of the report are added by using the JSPlumb JavaScript library to generate HTML 5 elements. You may need a relatively modern browser to see them.

Note

The checker tracks reference counts in an abstract way, in two parts: a part of the reference count that it knows about within the context of the function, along with a second part: all of the other references held by the rest of the program.

For example, in a call to PyInt_FromLong(0), it is assumed that if the call succeeds, the object has a reference count of 1 + N, where N is some unknown amount of other references held by the rest of the program. The checker knows that N >= 0.

If the object is then stored in an opaque container which is known to increment the reference count, the checker can say that the reference count is then 1 + (N+1).

If the function then decrements the reference count (to finish transferring the reference to the opaque container), the checker now treats the object as having a reference count of 0 + (N+1): it no longer owns any references on the object, but the reference count is actually unchanged relative to the original 1 + N amount. It also knows, given that N >= 0 that the actual reference count is >= 1, and thus the object won’t (yet) be deallocated.

Assumptions and configuration¶

For any function returning a PyObject*, it assumes that the PyObject*should be either a new reference to an object, or NULL (with an exception set) - the function’s caller should “own” a reference to that object. For all other PyObject*, it assumes that there should be no references owned by the function when the function terminates.

It will assume this behavior for any function (or call through a function pointer) that returns a PyObject*.

It is possible to override this behavior using custom compiler attributes as follows:

Marking functions that return borrowed references¶

The checker provides a custom GCC attribute:

attribute((cpychecker_returns_borrowed_ref))

which can be used to mark function declarations:

/* The checker automatically defines this preprocessor name when creating the custom attribute: */ #if defined(WITH_CPYCHECKER_RETURNS_BORROWED_REF_ATTRIBUTE) #define CPYCHECKER_RETURNS_BORROWED_REF
attribute((cpychecker_returns_borrowed_ref)) #else #define CPYCHECKER_RETURNS_BORROWED_REF #endif

PyObject *foo(void) CPYCHECKER_RETURNS_BORROWED_REF;

Given the above, the checker will assume that invocations of foo() are returning a borrowed reference (or NULL), rather than a new reference. It will also check that this is that case when verifying the implementation offoo() itself.

Marking functions that steal references to their arguments¶

The checker provides a custom GCC attribute:

attribute((cpychecker_steals_reference_to_arg(n)))

which can be used to mark function declarations:

/* The checker automatically defines this preprocessor name when creating the custom attribute: */ #if defined(WITH_CPYCHECKER_STEALS_REFERENCE_TO_ARG_ATTRIBUTE) #define CPYCHECKER_STEALS_REFERENCE_TO_ARG(n)
attribute((cpychecker_steals_reference_to_arg(n))) #else #define CPYCHECKER_STEALS_REFERENCE_TO_ARG(n) #endif

extern void foo(PyObject *obj) CPYCHECKER_STEALS_REFERENCE_TO_ARG(1);

Given the above, the checker will assume that invocations of foo() steal a reference to the first argument (obj). It will also verify that this is the case when analyzing the implementation of foo() itself.

More then one argument can be marked:

extern void bar(int i, PyObject *obj, int j, PyObject *other) CPYCHECKER_STEALS_REFERENCE_TO_ARG(2) CPYCHECKER_STEALS_REFERENCE_TO_ARG(4);

The argument indices are 1-based (the above example is thus referring toobj and to other).

All such arguments to the attribute should be PyObject* (or a pointer to a derived structure type).

It is assumed that such references are stolen for all possible outcomes of the function - if a function can either succeed or fail, the reference is stolen in both possible worlds.

Error-handling checking¶

The checker has knowledge of much of the CPython C API, and will generate a trace tree containing many of the possible error paths. It will issue warnings for code that appears to not gracefully handle an error.

(TODO: show example)

As noted above, it assumes that any function that returns a PyObject* can return can either NULL (setting an exception), or a new reference. It knows about much of the other parts of the CPython C API, including many other functions that can fail.

The checker will emit warnings for various events:

if it detects a dereferencing of a NULL value

if a NULL value is erroneously passed to various CPython API entrypoints which are known to implicitly dereference those arguments (which would lead to a segmentation fault if that code path were executed):
input.c: In function 'test':
input.c:38:33: warning: calling PyString_AsString with NULL (gcc.VarDecl('repr_args')) as argument 1 at input.c:38
input.c:31:15: note: when PyObject_Repr() fails at: repr_args = PyObject_Repr(args);
input.c:38:33: note: PyString_AsString() invokes Py_TYPE() on the pointer via the PyString_Check() macro, thus accessing (NULL)->ob_type
input.c:27:1: note: graphical error report for function 'test' written out to 'input.c.test-refcount-errors.html'

if it detects that an uninitialized local variable has been used

if it detects access to an object that has been deallocated, or such an object being returned:
input.c: In function 'test':
input.c:43:1: warning: returning pointer to deallocated memory
input.c:29:15: note: when PyLong_FromLong() succeeds at: PyObject *tmp = PyLong_FromLong(0x1000);
input.c:31:8: note: taking False path at: if (!tmp) {
input.c:39:5: note: reaching: Py_DECREF(tmp);
input.c:39:5: note: when taking False path at: Py_DECREF(tmp);
input.c:39:5: note: reaching: Py_DECREF(tmp);
input.c:39:5: note: calling tp_dealloc on PyLongObject allocated at input.c:29 at: Py_DECREF(tmp);
input.c:42:5: note: reaching: return tmp;
input.c:43:1: note: returning
input.c:39:5: note: memory deallocated here
input.c:27:1: note: graphical error report for function 'returning_dead_object' written out to 'input.c.test.html'

Errors in exception-handling¶

The checker keeps track of the per-thread exception state. It will issue a warning about any paths through functions returning a PyObject* that return NULL for which the per-thread exception state has not been set:

input.c: In function 'test': input.c:32:5: warning: returning (PyObject*)NULL without setting an exception

The checker does not emit the warning for cases where it is known that such behavior is acceptable. Currently this covers functions used as tp_iternext callbacks of aPyTypeObject.

If you have a helper function that always sets an exception, you can mark this property using a custom GCC attribute:

attribute((cpychecker_sets_exception))

which can be used to mark function declarations.

/* The checker automatically defines this preprocessor name when creating the custom attribute: */ #if defined(WITH_CPYCHECKER_SETS_EXCEPTION_ATTRIBUTE) #define CPYCHECKER_SETS_EXCEPTION
attribute((cpychecker_sets_exception)) #else #define CPYCHECKER_SETS_EXCEPTION #endif

extern void raise_error(const char *msg) CPYCHECKER_SETS_EXCEPTION;

Given the above, the checker will know that an exception is set whenever a call to raise_error() occurs. It will also verify that raise_error()actually behaves this way when compiling the implementation of raise_error.

There is an analogous attribute for the case where a function returns a negative value to signify an error, where the exception state is set whenever a negative value is returned:

attribute((cpychecker_negative_result_sets_exception))

which can be used to mark function declarations.

/* The checker automatically defines this preprocessor name when creating the custom attribute: */ #if defined(WITH_CPYCHECKER_NEGATIVE_RESULT_SETS_EXCEPTION_ATTRIBUTE) #define CPYCHECKER_NEGATIVE_RESULT_SETS_EXCEPTION
attribute((cpychecker_negative_result_sets_exception)) #else #define CPYCHECKER_NEGATIVE_RESULT_SETS_EXCEPTION #endif

extern int foo(void) CPYCHECKER_NEGATIVE_RESULT_SETS_EXCEPTION;

Given the above, the checker will know that an exception is raised whenever a call to foo returns a negative value. It will also verify that fooactually behaves this way when compiling the implementation of foo.

The checker already knows about many of the functions within the CPython API which behave this way.

Format string checking¶

The checker will analyze some Python APIs that take format strings and detect mismatches between the number and types of arguments that are passed in, as compared with those described by the format string.

It currently verifies the arguments to the following API entrypoints:

PyArg_ParseTuple

PyArg_ParseTupleAndKeywords

PyArg_Parse

Py_BuildValue

PyObject_CallFunction

PyObject_CallMethod

along with the variants that occur if you define PY_SSIZE_T_CLEAN before#include <Python.h>.

For example, type mismatches between int vs long can lead to flaws when the code is compiled on big-endian 64-bit architectures, wheresizeof(int) != sizeof(long) and the in-memory layout of those types differs from what you might expect.

The checker will also issue a warning if the list of keyword arguments in a call to PyArg_ParseTupleAndKeywords is not NULL-terminated.

Note

All of the various “#” codes in these format strings are affected by the presence of the macro PY_SSIZE_T_CLEAN. If the macro was defined before including Python.h, the various lengths for these format codes are of C type Py_ssize_t rather than int.

This behavior was clarified in the Python 3 version of the C API documentation, though the Python 2 version of the API docs leave the matter of which codes are affected somewhat ambiguous.

Nevertheless, the API does work this way in Python 2: all format codes with a “#” do work this way.

Internally, the C preprocessor converts such function calls into invocations of:

_PyArg_ParseTuple_SizeT

_PyArg_ParseTupleAndKeywords_SizeT

The checker handles this behavior correctly, by checking “#” codes in the regular functions against int and those in the modified functions againstPy_ssize_t.

Associating PyTypeObject instances with compile-time types¶

The “O!” format code to PyArg_ParseTuple takes a PyTypeObject followed by the address of an object. This second argument can point to aPyObject*, but it can also point to a pointer to a derived class.

For example, CPython’s own implementation contains code like this:

static PyObject * unicodedata_decomposition(PyObject *self, PyObject *args) { PyUnicodeObject *v;

/* ...snip... */

if (!PyArg_ParseTuple(args, "O!:decomposition",
                      &PyUnicode_Type, &v))

/* ...etc... */

in which the input argument is written out into the PyUnicodeObject*, provided that it is indeed a unicode instance.

When the cpychecker verifies the types in this format string it verifies that the run-time type of the PyTypeObject matches the compile-time type (PyUnicodeObject *). It is able to do this since it contains hard-coded associations between these worlds for all of Python’s built-in types: for the above case, it “knows” that PyUnicode_Type is associated withPyUnicodeObject.

If you need to provide a similar association for an extension type, the checker provides a custom GCC attribute:

attribute((cpychecker_type_object_for_typedef(typename)))

which can be used to mark PyTypeObject instance, giving the name of the typedef that PyObject instances of that type can be safely cast to.

/* The checker automatically defines this preprocessor name when creating the custom attribute: / #if defined(WITH_CPYCHECKER_TYPE_OBJECT_FOR_TYPEDEF_ATTRIBUTE) #define CPYCHECKER_TYPE_OBJECT_FOR_TYPEDEF(typename)
attribute((cpychecker_type_object_for_typedef(typename))) #else / This handles the case where we're compiling with a "vanilla" compiler that doesn't supply this attribute: */ #define CPYCHECKER_TYPE_OBJECT_FOR_TYPEDEF(typename) #endif

/* Define some PyObject subclass, as both a struct and a typedef / struct OurObjectStruct { PyObject_HEAD / other fields */ }; typedef struct OurObjectStruct OurExtensionObject;

/* Declare the PyTypeObject, using the custom attribute to associate it with the typedef above: */ extern PyTypeObject UserDefinedExtension_Type CPYCHECKER_TYPE_OBJECT_FOR_TYPEDEF("OurExtensionObject");

Given the above, the checker will associate the given PyTypeObject with the given typedef.

Verification of PyMethodDef tables¶

The checker will verify the types within tables of PyMethodDef initializers: the callbacks are typically cast to PyCFunction, but the exact type needs to correspond to the flags given. For example (METH_VARARGS | METH_KEYWORDS)implies a different function signature to the default, which the vanilla C compiler has no way of verifying.

/* BUG: there's a mismatch between the signature of the callback and that implied by ml_flags below. */ static PyObject *widget_display(PyObject *self, PyObject *args);

static PyMethodDef widget_methods[] = { {"display", (PyCFunction)widget_display, (METH_VARARGS | METH_KEYWORDS), /* ml_flags */ NULL},

{NULL, NULL, 0, NULL} /* terminator */

};

Given the above, the checker will emit an error like this:

input.c:59:6: warning: flags do not match callback signature for 'widget_display' within PyMethodDef table input.c:59:6: note: expected ml_meth callback of type "PyObject (fn)(someobject *, PyObject *args, PyObject *kwargs)" due to METH_KEYWORDS flag (3 arguments) input.c:59:6: note: actual type of underlying callback: struct PyObject * (struct PyObject *, struct PyObject *) (2 arguments) input.c:59:6: note: see http://docs.python.org/c-api/structures.html#PyMethodDef

It will also warn about tables of PyMethodDef initializers that are lacking a NULL sentinel value to terminate the iteration:

static PyMethodDef widget_methods[] = { {"display", (PyCFunction)widget_display, 0, /* ml_flags */ NULL},

/* BUG: this array is missing a NULL value to terminate
   the list of methods, leading to a possible segfault
   at run-time */

};

Given the above, the checker will emit this warning:

input.c:39:6: warning: missing NULL sentinel value at end of PyMethodDef table

Additional tests¶

the checker will verify the argument lists of invocations ofPyObject_CallFunctionObjArgs andPyObject_CallMethodObjArgs, checking that all of the arguments are of the correct type (PyObject* or subclasses), and that the list is NULL-terminated:
input.c: In function 'test':
input.c:33:5: warning: argument 2 had type char[12] * but was expecting a PyObject* (or subclass)
input.c:33:5: warning: arguments to PyObject_CallFunctionObjArgs were not NULL-terminated

Limitations and caveats¶

Compiling with the checker is significantly slower than with “vanilla” gcc. I have been focussing on correctness and features, rather than optimization. I hope that it will be possible to greatly speed up the checker via ahead-of-time compilation of the Python code (e.g. using Cython).

The checker does not yet fully implement all of C: expect to see Python tracebacks when it encounters less common parts of the language. (We’ll fix those bugs as we come to them)

The checker has a rather simplistic way of tracking the flow through a function: it builds a tree of all possible traces of execution through a function. This brings with it some shortcomings:

In order to guarantee that the analysis terminates, the checker will only track the first time through any loop, and stop analysing that trace for subsequent iterations. This appears to be good enough for detecting many kinds of reference leaks, especially in simple wrapper code, but is clearly suboptimal.

In order to avoid combinatorial explosion, the checker will stop analyzing a function once the trace tree gets sufficiently large. When it reaches this cutoff, a warning is issued:
input.c: In function 'add_module_objects':
input.c:31:1: note: this function is too complicated for the reference-count checker to analyze
To increase this limit, see the --maxtrans option.

The checker doesn’t yet match up similar traces, and so a single bug that affects multiple traces in the trace tree can lead to duplicate error reports.

Only a subset of the CPython API has been modelled so far. The functions known to the checker are:

PyArg_Parse and _PyArg_Parse_SizeT,PyArg_ParseTuple and _PyArg_ParseTuple_SizeT,PyArg_ParseTupleAndKeywords and _PyArg_ParseTupleAndKeywords_SizeT,PyArg_UnpackTuple,Py_AtExit,PyBool_FromLong,Py_BuildValue and _Py_BuildValue_SizeT,PyCallable_Check,PyCapsule_GetPointer,PyCObject_AsVoidPtr,PyCObject_FromVoidPtr,PyCObject_FromVoidPtrAndDesc,PyCode_New,PyDict_GetItem,PyDict_GetItemString,PyDict_New,PyDict_SetItem,PyDict_SetItemString,PyDict_Size,PyErr_Format,PyErr_NewException,PyErr_NoMemory,PyErr_Occurred,PyErr_Print,PyErr_PrintEx,PyErr_SetFromErrno,PyErr_SetFromErrnoWithFilename,PyErr_SetNone,PyErr_SetObject,PyErr_SetString,PyErr_WarnEx,PyEval_CallMethod,PyEval_CallObjectWithKeywords,PyEval_InitThreads,PyEval_RestoreThread,PyEval_SaveThread,Py_FatalError,PyFile_SoftSpace,PyFile_WriteObject,PyFile_WriteString,Py_Finalize,PyFrame_New,Py_GetVersion,PyGILState_Ensure,PyGILState_Release,PyImport_AddModule,PyImport_AppendInittab,PyImport_ImportModule,Py_Initialize,Py_InitModule4_64,PyInt_AsLong,PyInt_FromLong,PyList_Append,PyList_GetItem,PyList_New,PyList_SetItem,PyList_Size,PyLong_FromLong,PyLong_FromLongLong,PyLong_FromString,PyLong_FromVoidPtr,PyMapping_Size,PyMem_Free,PyMem_Malloc,PyModule_AddIntConstant,PyModule_AddObject,PyModule_AddStringConstant,_,PyModule_GetDict,PyNumber_Int,PyNumber_Remainer,PyObject_AsFileDescriptor,PyObject_Call,PyObject_CallFunction and _PyObject_CallFunction_SizeT,PyObject_CallFunctionObjArgs,PyObject_CallMethod and _PyObject_CallMethod_SizeT,PyObject_CallMethodObjArgs,PyObject_CallObject,PyObject_GetAttr,PyObject_GetAttrString,PyObject_GetItem,PyObject_GenericGetAttr,PyObject_GenericSetAttr,PyObject_HasAttrString,PyObject_IsTrue,_PyObject_New,PyObject_Repr,PyObject_SetAttr,PyObject_SetAttrString,PyObject_Str,PyOS_snprintf,PyRun_SimpleFileExFlags,PyRun_SimpleStringFlags,PySequence_Concat,PySequence_DelItem,PySequence_GetItem,PySequence_GetSlice,PySequence_SetItem,PySequence_Size,PyString_AsString,PyString_Concat,PyString_ConcatAndDel,PyString_FromFormat,PyString_FromString,PyString_FromStringAndSize,PyString_InternFromString,PyString_Size,PyStructSequence_InitType,PyStructSequence_New,PySys_GetObject,PySys_SetObject,PyTraceBack_Here,PyTuple_GetItem,PyTuple_New,PyTuple_Pack,PyTuple_SetItem,PyTuple_Size,PyType_IsSubtype,PyType_Ready,PyUnicodeUCS4_AsUTF8String,PyUnicodeUCS4_DecodeUTF8,PyWeakref_GetObject

The checker also has some knowledge about these SWIG-generated functions:SWIG_Python_ErrorType,SWIG_Python_SetErrorMsg

and of this Cython-generated function:__Pyx_GetStdout

Ideas for future tests¶

Here’s a list of some other C coding bugs I intend for the tool to detect:

tp_traverse errors (which can mess up the garbage collector); missing it altogether, or omitting fields

errors in GIL-handling

lock/release mismatches

missed opportunities to release the GIL (e.g. compute-intensive functions; functions that wait on IO/syscalls)

Ideas for other tests are most welcome (patches even more so!)

We will probably need various fallbacks and suppression modes for turning off individual tests (perhaps pragmas, perhaps compile-line flags, etc)

Reusing this code for other projects¶

It may be possible to reuse the analysis engine from cpychecker for other kinds of analysis - hopefully the python-specific parts are relatively self-contained. Email the gcc-python-plugin’s mailing list if you’re interested in adding verifiers for other kinds of code.

Common mistakes¶

Here are some common mistakes made using the CPython extension API, along with the fixes.

Missing Py_INCREF() on Py_None¶

The following is typically incorrect: a method implementation is required to return a new reference, but this code isn’t incrementing the reference count on Py_None.

PyObject* some_method(PyObject *self, PyObject *args) { [...snip...]

/* BUG: loses a reference to Py_None */
return Py_None;

}

If called enough, this could cause Py_None to be deallocated, crashing the interpreter:

Fatal error: deallocating None

The Py_RETURN_NONEmacro takes care of incrementing the reference count for you:

PyObject* some_method(PyObject *self, PyObject *args) { [...snip...]

/* Fixed version of the above: */
Py_RETURN_NONE;

}

Reference leak in Py_BuildValue¶

Py_BuildValue with “O” adds a new reference on the object for use by the new tuple, hence the following code leaks the reference already owned on the object:

/* BUG: reference leak: */ return Py_BuildValue("O", some_object_we_own_a_ref_on);

Py_BuildValue with “N” steals the reference (and copes with it being NULL by propagating the exception):

/* Fixed version of the above: */ return Py_BuildValue("N", some_object_we_own_a_ref_on);