Neal Norwitz | c3cd9df | 2004-06-06 19:58:40 +0000 | [diff] [blame] | 1 | This document describes some caveats about the use of Valgrind with |
| 2 | Python. Valgrind is used periodically by Python developers to try |
| 3 | to ensure there are no memory leaks or invalid memory reads/writes. |
| 4 | |
| 5 | If you don't want to read about the details of using Valgrind, there |
| 6 | are still two things you must do to suppress the warnings. First, |
| 7 | you must use a suppressions file. One is supplied in |
| 8 | Misc/valgrind-python.supp. Second, you must do one of the following: |
| 9 | |
| 10 | * Uncomment Py_USING_MEMORY_DEBUGGER in Objects/obmalloc.c, |
| 11 | then rebuild Python |
| 12 | * Uncomment the lines in Misc/valgrind-python.supp that |
| 13 | suppress the warnings for PyObject_Free and PyObject_Realloc |
| 14 | |
| 15 | Details: |
| 16 | -------- |
| 17 | Python uses its own allocation scheme on top of malloc called PyMalloc. |
| 18 | Valgrind my show some unexpected results when PyMalloc is used. |
| 19 | Starting with Python 2.3, PyMalloc is used by default. You can disable |
| 20 | PyMalloc when configuring python by adding the --without-pymalloc option. |
| 21 | If you disable PyMalloc, most of the information in this document and |
| 22 | the supplied suppressions file will not be useful. |
| 23 | |
| 24 | If you use valgrind on a default build of Python, you will see |
| 25 | many errors like: |
| 26 | |
| 27 | ==6399== Use of uninitialised value of size 4 |
| 28 | ==6399== at 0x4A9BDE7E: PyObject_Free (obmalloc.c:711) |
| 29 | ==6399== by 0x4A9B8198: dictresize (dictobject.c:477) |
| 30 | |
| 31 | These are expected and not a problem. Tim Peters explains |
| 32 | the situation: |
| 33 | |
| 34 | PyMalloc needs to know whether an arbitrary address is one |
| 35 | that's managed by it, or is managed by the system malloc. |
| 36 | The current scheme allows this to be determined in constant |
| 37 | time, regardless of how many memory areas are under pymalloc's |
| 38 | control. |
| 39 | |
| 40 | The memory pymalloc manages itself is in one or more "arenas", |
| 41 | each a large contiguous memory area obtained from malloc. |
| 42 | The base address of each arena is saved by pymalloc |
| 43 | in a vector, and a field at the start of each arena contains |
| 44 | the index of that arena's base address in that vector. |
| 45 | |
| 46 | Given an arbitrary address, pymalloc computes the arena base |
| 47 | address corresponding to it, then looks at "the index" stored |
| 48 | near there. If the index read up is out of bounds for the |
| 49 | vector of arena base addresses pymalloc maintains, then |
| 50 | pymalloc knows for certain that this address is not under |
| 51 | pymalloc's control. Otherwise the index is in bounds, and |
| 52 | pymalloc compares |
| 53 | |
| 54 | the arena base address stored at that index in the vector |
| 55 | |
| 56 | to |
| 57 | |
| 58 | the computed arena address |
| 59 | |
| 60 | pymalloc controls this arena if and only if they're equal. |
| 61 | |
| 62 | It doesn't matter whether the memory pymalloc reads up ("the |
| 63 | index") is initialized. If it's not initialized, then |
| 64 | whatever trash gets read up will lead pymalloc to conclude |
| 65 | (correctly) that the address isn't controlled by it. |
| 66 | |
| 67 | This determination has to be made on every call to one of |
| 68 | pymalloc's free/realloc entry points, so its speed is critical |
| 69 | (Python allocates and frees dynamic memory at a ferocious rate |
| 70 | -- everything in Python, from integers to "stack frames", |
| 71 | lives in the heap). |