I've attached my first pass at a patch to change the file system encoding on Windows to UTF-8 and remove use of the *A APIs.

It would be trivial to change the encoding from UTF-8 back to CP_ACP and change the error mode if that's what we decide is better, but my vote is strongly for an encoding that never drops characters when converted from UTF-16.

Discussion is still ongoing on python-ideas, so let's argue about yes/no and utf-8/mbcs there and just discuss the patch here.

I personally hate ansi myself so +1 to UTF-8/UTF-16.

Would it be acceptable for you to add a new option to switch to UTF-8 in Python 3.6, and discuss later if it's ok to enable it by default?

In the python-ideas threed, you wrote that Windows allow surrogate characters in filenames, but not the UTF-8/strict Python codec. Would it make sense to use UTF-8/surrogatepass codec to avoid any unicode error?

Steve Dower: Please don't use git format for diff, or the bug tracker is unable to create reviews. I regenerated the patch.

By the way, you introduced a bug in posix_do_stat(): you added a new "else" in the !MS_WINDOWS path which leads to a compilation error. I fixed it.

Thanks for the regen. I don't think git format is the problem as most of my patches are fine, it's probably because it was in a patch queue and so the parent isn't actually a known commit. I haven't tested whether this works without my other console patches but I think it should.

Is there a surrogatepass option? If so, I'll definitely use that, as that'll fix the one remaining edge case.

I suspect we'll have to go to Guido to get a ruling on the default, but I'll add an environment variable to switch.

Is there a surrogatepass option?

I'm talking about error handlers of Python codecs: text.encode('utf8',
'surrogatepass')

I suspect we'll have to go to Guido to get a ruling on the default, but I'll add an environment variable to switch.

If you go in this direction, I would like to follow you for the
UNIX/BSD side to make the switch portable. I was thinking about "-X
utf8" which avoids to change the command line parser.

If we agree on a plan, I would like to write it down as a PEP since I
expect a lot of complains and questions which I would prefer to only
answer once (see for example the length of your thread on python-ideas
where each people repeated the same things multiple times ;-))

By portable, do you mean not using an environment variable?

Command line parsing is potentially affected by this on Windows - I'd have to look deeper - as command lines are provided as UTF-16. But we may not ever expose them as bytes.

I don't even know that this matters on the UNIX/BSD side as the file system encoding provided there is correct, no? It's just Windows where the file system encoding used for bytes doesn't match what the file system actually uses.

I was afraid a PEP would be necessary out of this, but I want to see how the python-dev discussion goes first.

Steve Dower added the comment:

By portable, do you mean not using an environment variable?

I mean that "python3 -X utf8" should force sys.getfilesystemencoding()
to UTF-8 on UNIX/BSD, it would ignore the current locale setting.

Ah I see, if we end up sticking with MBCS and offering a switch to enable UTF-8. In that case, we'll definitely ensure the flag is the same (but I'm hopeful we will just make the reliable behavior on Windows the default, so it won't matter).

I belatedly remembered I've had this new test case hanging around for a while, and never got around to getting it into shape for inclusion in the standard library.

With the prospect of reasonable cross-platform consistency in this area, it could be a good thing to add as part of this PEP.

Also see PEP-529 for the latest updates there.

This is likely to be accepted as experimental for 3.6.0b1-3, and we'll commit to either the new default or a compatible default for b4.

PEP-529 has been accepted, so this really needs a review now. But since it's experimental and all the tests pass, I'll be committing it shortly anyway and will be tidying up issues during beta.

One minor change - I removed the unused definition of Py_FileSystemDefaultDecodeErrors.

Thanks for that review, Eryk, but I'm going to defer those to other issues (specifically bpo-27998 for scandir and we should file a new issue for the symlink concerns).

I've got some more doc updates to do though, and then I'll check in if there are no other concerns.

New changeset e20c7d8a8187 by Steve Dower in branch 'default':
Issue bpo-27781: Change file system encoding on Windows to UTF-8 (PEP-529)
https://hg.python.org/cpython/rev/e20c7d8a8187

This is pushed now - let the bug fixing begin :)

New changeset faca0730270b by Steve Dower in branch 'default':
Fixes tests broken by issue bpo-27781.
https://hg.python.org/cpython/rev/faca0730270b

It looks as though this change might have broken the compile on OS X. On my OS X 10.9 machine, building from a clean Git checkout of the master branch fails; the tail of the failed build looks like this:

./python.exe -E -S -m sysconfig --generate-posix-vars ;\
	if test $? -ne 0 ; then \
		echo "generate-posix-vars failed" ; \
		rm -f ./pybuilddir.txt ; \
		exit 1 ; \
	fi
Fatal Python error: Py_Initialize: unable to load the file system codec
Traceback (most recent call last):
  File "<frozen importlib._bootstrap>", line 962, in _find_and_load
  File "<frozen importlib._bootstrap>", line 951, in _find_and_load_unlocked
  File "<frozen importlib._bootstrap>", line 656, in _load_unlocked
  File "<frozen importlib._bootstrap_external>", line 668, in exec_module
  File "<frozen importlib._bootstrap_external>", line 782, in get_code
  File "<frozen importlib._bootstrap_external>", line 842, in _cache_bytecode
  File "<frozen importlib._bootstrap_external>", line 867, in set_data
  File "<frozen importlib._bootstrap_external>", line 117, in _write_atomic
ValueError: negative file descriptor
/bin/sh: line 1: 35829 Abort trap: 6           ./python.exe -E -S -m sysconfig --generate-posix-vars
generate-posix-vars failed
make: *** [pybuilddir.txt] Error 1

Full build output attached.

It looks as though this change in posixmodule.c is the cause:

 #ifdef MS_WINDOWS
-        if (path->wide)
-            fd = _wopen(path->wide, flags, mode);
-        else
+        fd = _wopen(path->wide, flags, mode);
 #endif
 #ifdef HAVE_OPENAT
         if (dir_fd != DEFAULT_DIR_FD)
             fd = openat(dir_fd, path->narrow, flags, mode);
         else
-#endif
             fd = open(path->narrow, flags, mode);
+#endif

The move of the final #endif means that fd is not defined on OS X. If I move the #endif back again, the compile succeeds.

New changeset 801634d3c105 by Steve Dower in branch 'default':
Issue bpo-27781: Fixes uninitialized fd when !MS_WINDOWS and !HAVE_OPENAT
https://hg.python.org/cpython/rev/801634d3c105

That seems to have done the trick. Thanks!

Before 3.6.0 beta 4 I need to make this change permanent. From memory, it's just an exception message that needs changing (and PEP-529 becomes final), but I'll review the changeset to be sure.

New changeset b26c8104e54f by Steve Dower in branch '3.6':
Closes bpo-27781: Removes special cases for the experimental aspect of PEP-529
https://hg.python.org/cpython/rev/b26c8104e54f

New changeset b8233c779ff7 by Steve Dower in branch 'default':
Closes bpo-27781: Removes special cases for the experimental aspect of PEP-529
https://hg.python.org/cpython/rev/b8233c779ff7