While fixing bpo-8391, I realized that subprocess doesn't support bytes program name if it's not an absolute path:
-------

$ ./python
>>> import subprocess
>>> subprocess.call([b'echo'])
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/SHARE/SVN/py3k/Lib/subprocess.py", line 449, in call
    return Popen(*popenargs, **kwargs).wait()
  File "/home/SHARE/SVN/py3k/Lib/subprocess.py", line 681, in __init__
    restore_signals, start_new_session)
  File "/home/SHARE/SVN/py3k/Lib/subprocess.py", line 1116, in _execute_child
    for exe in executable_list)
  File "/home/SHARE/SVN/py3k/Lib/subprocess.py", line 1115, in <genexpr>
    executable_list = tuple(fs_encode(exe)
  File "/home/SHARE/SVN/py3k/Lib/subprocess.py", line 1114, in <genexpr>
    for dir in path_list)
  File "/home/SHARE/SVN/py3k/Lib/posixpath.py", line 75, in join
    if b.startswith(sep):
TypeError: expected an object with the buffer interface
[62826 refs]

I'm working on a patch.

Attached patch (draft version) fixes this issue.

I proposed the creation of fs_encode() and fs_decode() functions in os.path: see bpo-8514. There functions would simplify my patch, but also make it correct on all OS (Windows, Mac, Linux).

My patch changes:

• os._execvpe(): support bytes type for the file argument (program name)
• os.get_exec_path(): support bytes type for the PATH environment variable
• Popen._execute_child(): decode the executable name before encoding the arguments (if the program name is not an absolute path)

Bytes program name problem should be splitted in two parts:

(a) subprocess.call([b'env']) and subprocess.call([b'env'], env={'PATH': '/usr/bin'}): bytes program and unicode environ
(b) bytes program and bytes environ

Part (a)
--------

(a) would benefit from the creation of os(.path).fsencode() function (issue bpo-8514).

Part (b)
--------

If the creation of os.environb is accepted (bpo-8603), I think that subprocess should also be modified to support pure bytes environ. Mix bytes and unicode variables in env would lead to mojibake and headaches.

So I propose the creation of an "envb" (environment bytes) argument to Popen constructor. (b) would be written as:

   subprocess.call([b'env'], envb={b'PATH': b'/usr/bin'})

or

   envb = os.environb.copy()
   envb[b'PATH'] = b'/usr/bin'
   subprocess.call([b'env'], envb=envb)

(and it will not be possible to define env and envb at the same time)

In this case, we will need a pure bytes version of os.get_exec_path(), and os._execvpe() should have a "bytes environment" mode (add an optional argument).

--

I have a patch fixing both problems, parts (a) and (b), but the patch depends on os.environb. I prefer to wait until os.environb is accepted to submit my patch.

New patch (issue8513_partA.patch):

• don't *decode*, only encode (str->bytes)
• only patch os._execvpe() for POSIX

If the creation of os.environb is accepted (bpo-8603), I think that
subprocess should also be modified to support pure bytes environ.

I fixed bpo-8603 and opened bpo-8640 for subprocess and envb.

I think your partA patch makes sense.

It would benefit from fsencode/fsdecode functions rather than manually doing the 'surrogateescape' thing everywhere.

Also, could you add a unittest for os._execvpe to test its behavior?

I think your partA patch makes sense.

I can fix part A and B in two commits.

It would benefit from fsencode/fsdecode functions rather
than manually doing the 'surrogateescape' thing everywhere.

I choosed to drop the idea of fsdecode() (patch for part A doesn't decode bytes anymore, it only encodes str). bpo-8514 has now a short and simple patch. I'm waiting for the final decision on bpo-8514 to commit the part A.

Also, could you add a unittest for os._execvpe to test its behavior?

os._execvpe() is a protected function. issue8513_partA.patch includes a test for subprocess. test_subprocess in two twices: with _posixsubprocess (C module) and with the pure Python implementation. The pure Python implementation calls os._execvpe(), that's why I patched also this function in my patch ;-)

Update the patch to use os.fsencode().

Build on the os._execvpe unittest I added in py3k r81001. Protected functions are perfectly fine things to unittest.

Build on the os._execvpe unittest I added in py3k r81001.

The test fails on Windows.

======================================================================
FAIL: test_internal_execvpe (test.test_os.ExecTests)
----------------------------------------------------------------------

Traceback (most recent call last):
  File "D:\cygwin\home\db3l\buildarea\3.x.bolen-windows\build\lib\test\test_os.py", line 683, in test_internal_execvpe
    exec_stubbed.calls)
AssertionError: Lists differ: [('execve', '/p/f', (['-a'], {... != [('execve', '/p\\f', (['-a'], ...

First differing element 0:
('execve', '/p/f', (['-a'], {'spam': 'beans'}))
('execve', '/p\\f', (['-a'], {'spam': 'beans'}))

• [('execve', '/p/f', (['-a'], {'spam': 'beans'})),
  ? ^

+ [('execve', '/p\\f', (['-a'], {'spam': 'beans'})),
? ^^

• ('execve', '/pp/f', (['-a'], {'spam': 'beans'}))]
  ? ^

+ ('execve', '/pp\\f', (['-a'], {'spam': 'beans'}))]
? ^^

my bad. hopefully r81019 fixes that.

New patch fixing this issue:

• os.get_exec_path() type now depends on the OS: str on Windows, bytes on Unix
• os.get_exec_path(None) uses os.environ on Windows, os.environb on Unix
• os.get_exec_path(env) uses 'PATH' or b'PATH' key, but raise a ValueError if both keys exist
• add os.supports_bytes_environ flag (boolean)
• os._execvpe() and subprocess._execute_child() canonicalize the program to bytes
• test "not path.supports_unicode_filenames" to check if fsencode() should be defined and used (instead of testing name != "nt")

I'm not proud of the change on os.get_exec_path() result type, I'm not sure that it's the right thing to do. But at least, the patch works :-)

Oops, I forgot to add the new patch: subprocess_bytes_program-3.patch.

I asked on #python-dev about os.get_exec_path() result type. As expected, the answer was "It's a really bad idea".

So here is a new version of my patch. Summary of the patch version 4:

• subprocess.Popen() and os._execvpe() support bytes program name
• os.get_exec_path() now supports b'PATH' key and bytes value
• add os.supports_bytes_environ flag: "True if the native OS type of the environment is bytes (eg. False on Windows)"

Changes since the version 3:

• document the new os.supports_bytes_environ flag
• os.get_exec_path() result type is always str (decode bytes to str using sys.getfilesystemencoding() + surrogateescape)
• path.supports_unicode_filenames is False on Windows 9x but Windows 9x native type is unicode and so fsencode() should not be defined on this OS: revert the fsencode() test (if not path.supports_unicode_filenames => if name != 'nt')

I forgot to update test_os: patch version 5.

Ok, everything is ready for this issue: os.environb and os.fsencode() are commited, and test_os is prepared to test os._execvpe() change. I'm just waiting for a review.

Execpt the change on os.get_exec_path(), the patch is now simple.

Hum, os.get_exec_path() has no test for the new features (support b'PATH' key and bytes value).

Fixed by r81291 + r81292 (py3k).

The final commit contains much more tests ;-) I will watch the buildbot next hours and block the commit in 3.1.

Oh, I forget subprocess.call(b'ls'): command as a byte string => fixed in Python 3.3 (r88720).