This critical bug is one of the reasons that non-English speaking communities doesn't adopt Python as broadly as it happens in English world compared to other technologies (PHP etc.).

# -- coding: utf-8 --

import os

os.mkdir(u'Русское имя')
os.mkdir(u'English name')

for r, dirs, files in os.walk('.'):
  print dirs

This gives:
['English name']
[]

Windows Vista.

dir /b
English name
test.py
Русское имя

Is it reproduced on 3.x?

No.

Oops, clicked submit too soon.

It isn't likely to get fixed in 2.7, because 2.7's unicode support problems is the major reason python3 was developed.

For that matter, it isn't reproduced in python2.7, either:

>>> for r, dirs, files in os.walk(u'.'):
...   print dirs
... 
[u'\u0420\u0443\u0441\u0441\u043a\u043e\u0435 \u0438\u043c\u044f']
[]

The problem exhibited is not coming from the os.walk() implementation, but from the use of a byte-string as the argument to it.

The directories are created with unicode literals and therefore the argument must also be a unicode literal (u'.') for them to be shown. See the note in the listdir() documentation.

As it stands, I suggest that this is closed as invalid, or at the minimum that it could be a documentation bug for walk() not also having a similar note as listdir().

Works for me without the u'.', too, though less usefully:

>>> for r, dirs, files in os.walk('.'):
...   print dirs
... 
['\xd0\xa0\xd1\x83\xd1\x81\xd1\x81\xd0\xba\xd0\xbe\xd0\xb5 \xd0\xb8\xd0\xbc\xd1\x8f']

Maybe that doesn't work on Windows, though. I am, of course, assuming that python3 does the right thing on Windows, but I can't imagine Victor would have overlooked that.

In Python 3 it fails with UnicodeEncodeError in "C:\Python33\lib\encodings\cp437.py", while Vista's 'dir' command shows everything correctly in the same console, so somebody definitely overlooked that aspect.

This bug is clearly an issue for developers who write products for international markets. It is neither out of date, nor it is invalid. Note in documentation in red is a must have, also a warning should be issued in warning mode when os.walk() ignores international dirs. I doubt there are many people who aware of this racist behavior and want it be default.

• Do you have a full traceback of the failing os.walk() in Python3.3?
• What's the result of os.listdir(u'.') ?

What are the results of os.listdir(b'.') and os.listdir(u'.') on Python 2.7 and Python 3.3+?

What are the results of os.stat(b'Русское имя') and os.stat(b'Русское имя') on Python 2.7 and Python 3.3+?

What are the results of sys.getdefaultencoding(), sys.getfilesystemencoding(), locale.getpreferredencoding(False) and locale.getpreferredencoding(True) on Python 2.7 and Python 3.3+?

If any of those calls fail, please provide a full traceback.

My guess is that your unicode issue is bpo-1602, which is non-trivial to solve.

My guess is that your unicode issue is bpo-1602, which is non-trivial to solve.

In such case the output will be something like:

['English name', '']
[]
[]

• Do you have a full traceback of the failing os.walk() in Python3.3?

Traceback (most recent call last):
  File "test.py", line 9, in <module>
    print(dirs)
  File "C:\Python33\lib\encodings\cp437.py", line 19, in encode
    return codecs.charmap_encode(input,self.errors,encoding_map)[0]
UnicodeEncodeError: 'charmap' codec can't encode characters in position
18-24: character maps to <undefined>

• What's the result of os.listdir(u'.') ?

>python3 -c "import os; print(os.listdir(u'.'))"
Traceback (most recent call last):
  File "<string>", line 1, in <module>
  File "C:\Python33\lib\encodings\cp437.py", line 19, in encode
    return codecs.charmap_encode(input,self.errors,encoding_map)[0]
UnicodeEncodeError: 'charmap' codec can't encode characters in position
41-47: character maps to <undefined>

python2 -c "import os; print(os.listdir(u'.'))"
[u'English name', u'test.py', u'test2.py',
u'\u0420\u0443\u0441\u0441\u043a\u043e\u0435 \u0438\u043c\u044f']

python2 -c "import os; print(os.listdir('.'))"
['English name', 'test.py', 'test2.py', '??????? ???']

I attach tests.py file used to run the tests. Results are in python2.out.txt and python3.out.txt also attached.

What are the results of os.stat(b'Русское имя') and os.stat(b'Русское имя') on Python 2.7 and Python 3.3+?

b'Русское имя' is not a valid syntax construct in Python 3 even though I have correct 'coding: utf-8' header and expect characters to be utf-8 bytes. Therefore I skipped this test for Python 3.

python test.py
File "tests.py", line 23
print(os.stat(b'\u0420\u0443\u0441\u0441\u043a\u043e\u0435 \u0438\u043c\u044f'))
^
SyntaxError: bytes can only contain ASCII literal characters.

Thank you, Anatoly, for report. I'll try to investigate this issue.

So, it seems that os.walk() and os.listdir() work correctly with Python3.3, but print(u'Русское имя') fails because the terminal encoding is cp437.

See bpo-1602 for the print issue.
As a quick workaround, try to set PYTHONIOENCODING=cp437:backslashreplace as suggested in http://wiki.python.org/moin/PrintFails

If nothing is wrong with os.walk() and os.listdir(), this issue should be closed.

Anatoly, can you please run the attached test?

Based on the pasted results I'm pretty sure there's nothing wrong with walk and listdir. But it sounds like Serhiy will check to make sure, so we'll wait for his report.

The byte string vs the coding cookie is an interesting observation, but is a separate issue and should probably be raised on python-ideas, since I'm guessing it the current behavior was a conscious design choice.

Anatoly

b'Русское имя' is not a valid syntax construct in Python 3 even though I have
correct 'coding: utf-8' header and expect characters to be utf-8 bytes.

David

The byte string vs the coding cookie is an interesting observation, but is a separate
issue and should probably be raised on python-ideas, since I'm guessing it the
current behavior was a conscious design choice.

Yes, it works as designed: the coding cookie is used to decode bytes to characters in unicode literals (e.g. if I have u'Éric' in my source file, not a \u escape); bytes literals are independent of the coding cookie and should always contain only bytes, not characters (including \u escapes), e.g. '\xc3\x89ric' for UTF-8 bytes.

There is one more problem - when I redirect the output with:

py test_unicode_fname.py > test.log 2>&1

In Python 2.7 the traceback is at the end of file, in Python 3.3 it is at the beginning. Therefore I just copied data from the screen, where it appears in correct order.

(current mood: Python debugging is a mess)

Anatoly, please file another issue for the 2>&1 mess.

Thanks, Anatoly. I see an actual bug. FindFirstFile and FindNextFile return broken name if file unicode name can't be represented in current codepage.

I don't know what is perfect solution for this issue.

On 2.7 we can decode listdir() argument to unicode and then encode result names to str with sys.getfilesystemencoding() only if it is possible. Therefore listdir() with str argument will return unicode for non-encodable names. This should not make many new problems in addition to those which 2.7 already have with Unicode.

But on 3.x listdir() with bytes argument can returns only bytes objects. I don't know what to do with non-encodable names in such case. Perhaps an exception should be raised. Fortunately listdir() with bytes argument is rarely used on 3.x.

That's what surrogateescape is for, on linux. I thought Victor dealt with this a different way in Windows. Maybe by deprecating the bytes interface :)

Surrogateescape is for non-decodable names. Here we have a problem with non-encodable names.

I know that naive approach with using only Unicode API inside is not work because Windows use complex logic for filename encoding (for example dropping diacritics). Perhaps Martin have more to say.

Ah, I misunderstood your comment.

So, listdir is returning the "correct" the filename, it's just that we can't encode it to the console encoding. So, it is working as expected within the current windows console limitations, if not in a particularly useful fashion.

(That is, listdir/os.walk are *not* ignoring the international dirs.)

Ah, I misunderstood your comment.

Ah, you misunderstood my comment right now.

So, listdir is returning the "correct" the filename, it's just that we can't encode it to the console encoding.

listdir() returns already irremediably broken filename (all Cyrillic
letters replaced with '?'). My test script outputs only ascii data, you
see literally what you get, there is no output encoding issues.

Oh, I remember Victor complaining about that behavior of Windows. I'm pretty sure it is the windows API and not python that is doing that mangling. But Victor would know for sure.

I'm a little confused. FindFirstFile is an ANSI API, so we get a narrow string back. We call PyBytes_FromString(), which expects a narrow string and returns a bytes object. Who's trying (and failing) to encode the filename?

Who's trying (and failing) to encode the filename?

Windows. File created using Unicode API and stored UTF-16 encoded in
NTFS. Windows fails to represent this filename using ANSI API.

Here is a patch against 2.7 which always uses Unicode API in listdir()
and tries to encode filenames to str if str argument used.

I can't test (and even compile) the patch as I don't have a Windows,
please test it for me.

On Windows with Python 2, unencodable characters are replaced with "?". It is the default behaviour of WideCharToMultiByte() and so all ANSI functions have this behaviour. Python doesn't try to behave differently, it just exposes system function as Python functions.

So for example, os.listdir(bytes) returns filename with "?" if some characters are not encodable to the ANSI codepage. It's a choice in the design of Windows.

This critical bug is one of the reasons that non-English speaking
communities doesn't adopt Python as broadly as it happens in
English world compared to other technologies (PHP etc.).

I don't understand this point.

PHP doesn't have a Unicode type, I'm quite sure that PHP have exactly the same issue. And this issue is only solved in Python 3... except if you explicitly uses a bytes filename (for os.listdir/os.walk), but the bytes filename API has been deprecated in Python 3.3.

In Python 2, you can use Unicode filenames to workaround this issue. But it doesn't work as well as Python 3: on UNIX, you will get a similar issue with undecodable filenames (which is the opposite of unencodable filenames).

Read my book for more information: https://github.com/haypo/unicode_book/wiki

--

About listdir_unicode-2.7.patch: Python chose to work as Windows with unencodable characters. If you want to change the behaviour, you must change *all* calls to the Windows ANSI API (which is not trivial). Anyway, as I wrote, the bytes API is deprecated for filenames in Python 3.3. I prefer to not change anything in Python 2, because it may break existing applications. For example, os.listdir(bytes) doesn't fail in Python 2.7 with unencodable names, whereas it fails with your patch.

Nothing interesting in this issue, I'm closing it. If your consider the redirection issue important, please open a new issue.

And this issue is only solved in Python 3...

Ooops, I mean: this issue is *already* solved in Python 3

For example, os.listdir(bytes) doesn't fail in Python 2.7 with unencodable names, whereas it fails with your patch.

No. The purpose of this patch is that it doesn't fail and should return
a usable result.

haypo didn't understood the main user story for this ticket and closed it, so I reopen it with simplified user story.

"""As a developer, I want Python os.listdir('.') function return all directories in current directory on Windows, not matter how international they are, in the same way as other Windows applications return them.

I want returned name to be normal string, but reuseable in subsequent functions, so that I can query stats for this dir or CD to it. It doesn't matter how this name is binary represented. It can be quoted, in UTF-8 byte code or whatever - I don't care as long as I can access the same name from the same session."""

Anatoly,

• in Python2.7, try this:
  print repr(os.listdir(u'.'))
• in Python3, try this:
  print(ascii(os.listdir('.')))
  Do the commands above work correctly?

Looks as this is a duplicate of bpo-13247. And Victor submitted a patch with approach similar to me (except that my patch does not raises an exception, but returns an unicode for unencodable names).

It looks as longliving design bug and perhaps really be closed as "won't fix".

Indeed, os.listdir() should always be called with a unicode argument under Windows (which is the natural thing to do in 3.x, anyway). The other issue (with print() unable to display some symbols depending on the codepage) is unrelated.