The CPython interpreter gets SIGSEGV or SIGABRT during the run. The script attempts to decode binary file using UTF-16-LE encoding and custom error handler. The error handler is poorly built, and doesn't respect the unicode standard with wrong calculation of the new position for decoder to continue. This somehow interfere with internal C code doing memory allocation. The result is invalid writes outside of allocated block.

Here is how it looks like with Python 3.7.0a4+ (heads/master:44a70e9, Jan 17 2018, 12:18:45) run under Valgrind 3.11.0. Please see the full Valgrind output in attached valgrind.log.

==24836== Invalid write of size 4
==24836==    at 0x4C6B17: ucs4lib_utf16_decode (codecs.h:540)
==24836==    by 0x4C6B17: PyUnicode_DecodeUTF16Stateful (unicodeobject.c:5600)
==24836==    by 0x55AAD3: _codecs_utf_16_le_decode_impl (_codecsmodule.c:363)
==24836==    by 0x55AB6C: _codecs_utf_16_le_decode (_codecsmodule.c.h:371)
==24836==    by 0x4315D6: _PyMethodDef_RawFastCallKeywords (call.c:651)
==24836==    by 0x431840: _PyCFunction_FastCallKeywords (call.c:730)
==24836==    by 0x4ED159: call_function (ceval.c:4580)
==24836==    by 0x4ED159: _PyEval_EvalFrameDefault (ceval.c:3134)
==24836==    by 0x4E302D: PyEval_EvalFrameEx (ceval.c:545)
==24836==    by 0x4E3A42: _PyEval_EvalCodeWithName (ceval.c:3971)
==24836==    by 0x430EDD: _PyFunction_FastCallDict (call.c:376)
==24836==    by 0x4336B0: PyObject_Call (call.c:226)
==24836==    by 0x433839: PyEval_CallObjectWithKeywords (call.c:826)
==24836==    by 0x4FEAA6: _PyCodec_DecodeInternal (codecs.c:471)
==24836==  Address 0x6cf4bf8 is 0 bytes after a block of size 339,112 alloc'd
==24836==    at 0x4C2DB8F: malloc (in /usr/lib/valgrind/vgpreload_memcheck-amd64-linux.so)
==24836==    by 0x467635: _PyMem_RawMalloc (obmalloc.c:75)
==24836==    by 0x467B7D: _PyMem_DebugRawAlloc (obmalloc.c:2033)
==24836==    by 0x467C1F: _PyMem_DebugRawMalloc (obmalloc.c:2062)
==24836==    by 0x467C40: _PyMem_DebugMalloc (obmalloc.c:2202)
==24836==    by 0x468BFF: PyObject_Malloc (obmalloc.c:616)
==24836==    by 0x493902: PyUnicode_New (unicodeobject.c:1293)
==24836==    by 0x4BEA4F: _PyUnicodeWriter_PrepareInternal (unicodeobject.c:13456)
==24836==    by 0x4C6D39: _PyUnicodeWriter_WriteCharInline (unicodeobject.c:13494)
==24836==    by 0x4C6D39: PyUnicode_DecodeUTF16Stateful (unicodeobject.c:5637)
==24836==    by 0x55AAD3: _codecs_utf_16_le_decode_impl (_codecsmodule.c:363)
==24836==    by 0x55AB6C: _codecs_utf_16_le_decode (_codecsmodule.c.h:371)
==24836==    by 0x4315D6: _PyMethodDef_RawFastCallKeywords (call.c:651)

As written, decode_crash.py crashes on Windows also.  Passing 'replace' instead of 'w3lib_replace' results in no crash and lots of boxes and blanks.

The problem is utf16 decoder almost always assumes that two bytes decodes to one unicode character, so when allocating memory, it assumes (bytes_number+1)/2 unicode slots is enough, there is even a comment in the code. And in unicode_decode_call_errorhandler_writer, it only allocates more memory when the error handler returns a unicode longer than 1, but doesn't take care pace by one, in which case one byte to one unicode character. So it's possible for the decoder to write out of bound.

This example could steadily crash on my Mac with debug version, it writes across the bound of the internal unicode buffer:

>>> import codecs
>>> def pace_by_one(exc):
...     return ('\ufffd', exc.start+1)
...
>>> codecs.register_error('pace_by_one', pace_by_one)
>>> b'\xd8\xd8\xd8\xd8\xd8\xd8\x00\x00\x00'.decode('utf-16-le', 'pace_by_one')
Debug memory block at address p=0x10210c260: API 'o'
    100 bytes originally requested
    The 7 pad bytes at p-7 are FORBIDDENBYTE, as expected.
    The 8 pad bytes at tail=0x10210c2c4 are not all FORBIDDENBYTE (0xfb):
        at tail+0: 0x00 *** OUCH
        at tail+1: 0x00 *** OUCH
        at tail+2: 0xfb
        at tail+3: 0xfb
        at tail+4: 0xfb
        at tail+5: 0xfb
        at tail+6: 0xfb
        at tail+7: 0xfb
    The block was made by call #30672 to debug malloc/realloc.
    Data at p: 00 00 00 00 00 00 00 00 ... fd ff fd ff fd ff d8 00

Fatal Python error: bad trailing pad byte

Current thread 0x00007fffab9b4340 (most recent call first):
  File "/Users/angwer/Repositories/cpython/Lib/encodings/utf_16_le.py", line 16 in decode
  File "<stdin>", line 1 in <module>
[1]    63997 abort      ~/Repositories/cpython/python.exe

I'll try to make a fix tomorrow.

Another way to crash:

>>> import codecs
>>> def replace_with_longer(exc):
...     exc.object = b'\xa0\x00' * 100
...     return ('\ufffd', exc.end)
...
>>> codecs.register
codecs.register(       codecs.register_error(
>>> codecs.register_error('replace_with_longer', rep
replace_with_longer( repr(
>>> codecs.register_error('replace_with_longer', replace_with_longer)
>>> b'\xd8\xd8'.decode('utf-16-le', 'replace_with_longer')
Debug memory block at address p=0x10b3b8c40: API 'o'
    92 bytes originally requested
    The 7 pad bytes at p-7 are FORBIDDENBYTE, as expected.
    The 8 pad bytes at tail=0x10b3b8c9c are not all FORBIDDENBYTE (0xfb):
        at tail+0: 0xa0 *** OUCH
        at tail+1: 0x00 *** OUCH
        at tail+2: 0xa0 *** OUCH
        at tail+3: 0x00 *** OUCH
        at tail+4: 0xa0 *** OUCH
        at tail+5: 0x00 *** OUCH
        at tail+6: 0xa0 *** OUCH
        at tail+7: 0x00 *** OUCH
    The block was made by call #11529390970613309440 to debug malloc/realloc.
    Data at p: 00 00 00 00 00 00 00 00 ... 00 00 00 00 fd ff a0 00

Fatal Python error: bad trailing pad byte

Current thread 0x00007fffab9b4340 (most recent call first):
  File "/Users/angwer/Repositories/cpython/Lib/encodings/utf_16_le.py", line 16 in decode
  File "<stdin>", line 1 in <module>
[1]    64081 abort      ~/Repositories/cpython/python.exe

I write a draft patch, without tests yet. I'll add them later. Reviews are appreciated. I also check the Windows codepage equivalent and encoders, look to me they don't suffer the problem.

New changeset 2c7fd46e11333ef5e5cce34212f7d087694f3658 by Xiang Zhang in branch 'master':
bpo-32583: Fix possible crashing in builtin Unicode decoders (#5325)
https://github.com/python/cpython/commit/2c7fd46e11333ef5e5cce34212f7d087694f3658

New changeset ea94fce6960d90fffeeda131e31024617912d231 by Xiang Zhang in branch '3.6':
[3.6] bpo-32583: Fix possible crashing in builtin Unicode decoders (GH-5325) (#5459)
https://github.com/python/cpython/commit/ea94fce6960d90fffeeda131e31024617912d231

New changeset 86fdad093b863db7ef6a3a00c9cff724c09442e7 by Ned Deily (Xiang Zhang) in branch '3.7':
bpo-32583: Fix possible crashing in builtin Unicode decoders (#5325)
https://github.com/python/cpython/commit/86fdad093b863db7ef6a3a00c9cff724c09442e7