[Patches] [ python-Patches-683592 ] unicode support for os.listdir()

SourceForge.net noreply@sourceforge.net
Tue, 04 Mar 2003 07:50:18 -0800


Patches item #683592, was opened at 2003-02-09 22:43
You can respond by visiting: 
https://sourceforge.net/tracker/?func=detail&atid=305470&aid=683592&group_id=5470

Category: Library (Lib)
Group: None
Status: Closed
Resolution: Accepted
Priority: 5
Submitted By: Just van Rossum (jvr)
Assigned to: Martin v. Löwis (loewis)
Summary: unicode support for os.listdir()

Initial Comment:
The attached patch makes os.listdir() return unicode strings, on plaforms that have Py_FileSystemDefaultEncoding defined as non-NULL.

I'm by no means sure this is the right thing to do; it does seem right on OSX where Py_FileSystemDefaultEncoding is (or rather: will be real soon, I'm waiting for Jack's approval) utf-8. I'd be happy to add the code in an OSX-specific switch.

A more subtle variant could perhaps only return unicode strings if the file name is not ASCII.

----------------------------------------------------------------------

>Comment By: Just van Rossum (jvr)
Date: 2003-03-04 16:50

Message:
Logged In: YES 
user_id=92689

Here's a note about file system encodings on OSX, including a few words about NFS: http://developer.apple.com/qa/qa2001/qa1173.html.

I propose to fall back to a byte string if conversion to unicode fails.

----------------------------------------------------------------------

Comment By: Jack Jansen (jackjansen)
Date: 2003-03-04 16:44

Message:
Logged In: YES 
user_id=45365

I just did a test (created 254 files with all bytes except / and null in their names on a linux server, mounted the partition over NFS on MacOSX) and indeed MacOSX tries to interpret the bytes as UTF-8 and fails.

I know that conversion works for HFS and HFS+ volumes (which carry a filename encoding with them, or you have to specify it when mounting). I assume it works for AFP and SMB (which also carries encoding info, IIRC) but I can't test this. I haven't a clue about webdav and such.

Something to keep in mind is that we are really trying to solve someone else's problem: the inability of NFS and most unixen to handle file system encodings. If I'm on a latin-1 machine and I nfs-mount your latin-2 partition I will see garbage filenames.

----------------------------------------------------------------------

Comment By: Martin v. Löwis (loewis)
Date: 2003-03-04 16:15

Message:
Logged In: YES 
user_id=21627

Setting the file system encoding on startup should be fine,
except that we need another setlocale/query/restore locale
sequence. This is, in principle, bad, as there is no
guarantee that the restore locale operation really produces
the original state, and may cause problems if other threads
are already running. In practice, it appears to work out
just fine, as we use such sequences already (e.g. to undo
the readline initialization).

----------------------------------------------------------------------

Comment By: Martin v. Löwis (loewis)
Date: 2003-03-04 16:11

Message:
Logged In: YES 
user_id=21627

I disagree with the last assertion: In *particular* if the
file system encoding is UTF-8, there is a good chance that
decoding will fail (unlike if it is latin-1; decoding will
then never fail - it may just produce mojibake). 

OS X seems to make a guarantee to always return UTF-8 from
its low-level API, but I distrust this guarantee until I see
it with my own eyes :-) E.g. what happens if you mount an
NFS tree, and the NFS server gives file names in some other
encoding?

I see the following options:

- only enable the code for OS X. I dislike this option, as
it essentially freezes the Unix status to non-Unicode (we
won't get further insights, the de jure status won't change,
de facto, all files will be encoded in the locale's encoding).

- leave the code as-is, documenting the possibility of
exceptions.

- add byte strings instead of Unicode strings into the
result for non-decodable strings. This gives a mixed-type
result, which is fine if you only pass the resulting file
names to stat() or open(), and will likely break the
application if it tries to display the file names somehow.

----------------------------------------------------------------------

Comment By: Just van Rossum (jvr)
Date: 2003-03-04 16:07

Message:
Logged In: YES 
user_id=92689

I think it would be better to simply return byte strings if the file system encoding isn't know. (This btw. was what my original patch did.)

----------------------------------------------------------------------

Comment By: Guido van Rossum (gvanrossum)
Date: 2003-03-04 16:03

Message:
Logged In: YES 
user_id=6380

Maybe the filesystem default encoding should be set to
Latin-1 by default (when nothing better is known about it)?
Then it's hard to imagine how the conversion could fail,
since every Latin-1 byte maps 1-1 to the corresponding
Unicode code point.

----------------------------------------------------------------------

Comment By: Guido van Rossum (gvanrossum)
Date: 2003-03-04 15:54

Message:
Logged In: YES 
user_id=6380

The setlocale call indeed works.

I think I'd be happier if this was set by default, but I
don't know what other consequences there would be.

----------------------------------------------------------------------

Comment By: Just van Rossum (jvr)
Date: 2003-03-04 15:51

Message:
Logged In: YES 
user_id=92689

It would seem that even with a user's locale there's a chance os.listdir() fails when passed a unicode argument. I'm not sure it's reasonable for os.listdir() to fail at all (if the directory to be listed exists and we the right permissions).

If it's all too difficult to get right, I'm happy to put the listdir unicode support in a MacOSX switch. I know nothing about locales so I'm really not in a position to straighten this out. All I know is that if Py_FileSystemDefaultEncoding is known to be utf-8, it's just dumb _not_ to return unicode. You guys figure out the rest.

----------------------------------------------------------------------

Comment By: Martin v. Löwis (loewis)
Date: 2003-03-04 15:40

Message:
Logged In: YES 
user_id=21627

Guido's scenario was precisely the reason why Unix was left
out from consideration for PEP 277.

However, it is better than it sounds: There is a good chance
that invoking locale.setlocale(locale.LC_CTYPE, "") prior to
invoking listdir will overcome the problem, as the setlocale
call will set the file system encoding to the user's
preference. If \xff is a valid file name in the user's
preferred encoding, then listdir will succeed in converting
this file name to a Unicode string.

It might be useful to set the file system encoding on Unix
to the user's preferred encoding unconditionally (i.e. not
as a side effect of invoking setlocale). It might also be
useful to expose the file system encoding read-only for
inspection.

----------------------------------------------------------------------

Comment By: Just van Rossum (jvr)
Date: 2003-03-04 15:31

Message:
Logged In: YES 
user_id=92689

Would you prefer the error be silenced and a byte string be used instead? If so, should there be a warning?

----------------------------------------------------------------------

Comment By: Guido van Rossum (gvanrossum)
Date: 2003-03-04 15:01

Message:
Logged In: YES 
user_id=6380

I haven't seen the code, but I have a complaint.

On Linux, when I have a file named '\xff' (i.e. its name is
the single byte with value 255), os.listdir(u'.') gives me a
UnicodeDecodeError.

Is that really progress?

----------------------------------------------------------------------

Comment By: Martin v. Löwis (loewis)
Date: 2003-03-04 07:49

Message:
Logged In: YES 
user_id=21627

The current code looks fine to me. Closing this patch.

----------------------------------------------------------------------

Comment By: Just van Rossum (jvr)
Date: 2003-03-03 18:56

Message:
Logged In: YES 
user_id=92689

Martin, assigning this item to you. Please close it if you deem the changes in CVS correct.

----------------------------------------------------------------------

Comment By: Just van Rossum (jvr)
Date: 2003-03-03 18:45

Message:
Logged In: YES 
user_id=92689

Applied to CVS as:
  Modules/posixmodule.c: 2.288
  Doc/lib/libos.tex: 1.115
  Misc/NEWS: 1.687

Unicode errors are propagated as in the original version of the patch, libos.tex mentions Win NT/2k/XP and Unix.

----------------------------------------------------------------------

Comment By: Martin v. Löwis (loewis)
Date: 2003-03-03 17:39

Message:
Logged In: YES 
user_id=21627

Clearing the error is bad, I agree. I see two options:
reraise the exception, deleting the result obtained so far
(i.e. as the code did that the latest patch removes), OR add
a byte string instead of the Unicode string into the result.
Even though I have proposed the latter in the past, I could
also accept the former; applications that anticipate that
exception then just need to re-invoke listdir with a byte
string, and deal with the result themselves.

With these changes, the patch is fine with me.

----------------------------------------------------------------------

Comment By: Just van Rossum (jvr)
Date: 2003-03-03 17:08

Message:
Logged In: YES 
user_id=92689

I think this could be achieved by removing the "Py_FileSystemDefaultEncoding != NULL" part of the condition on line 1805, as indeed passing NULL as the encoding to PyUnicode_FromEncodedObject causes the default encoding to be used. Shall I check it in like that?

I'm not quite happy with the fact that exceptions are silently dropped: should a warning be issued instead? Especially when using the default encoding, exceptions are not unlikely I suppose.

----------------------------------------------------------------------

Comment By: Martin v. Löwis (loewis)
Date: 2003-03-03 16:48

Message:
Logged In: YES 
user_id=21627

I see. The right thing, IMO, is to always return Unicode
objects for Unicode arguments, just the same way the "et"
parser works: if the file system encoding is NULL, fall back
to the system default encoding. Then, you can generalize the
docs to [NT and Unix] (with OS X being a flavour of Unix),
or drop the OS reference completely (in which case the other
os modules are effectively buggy).

There might be a function already to fall back to the system
default encoding; perhaps just passing NULL works.

There should be a documentation section on Unicode file
names; I volunteer to write it (Summary: NT+ uses Unicode
natively, W9x uses "mbcs", OS X uses UTF-8, which equates to
"Unicode natively", Unices with nl_langinfo(CODEPAGE) use
that, all others use the system default encoding).

----------------------------------------------------------------------

Comment By: Just van Rossum (jvr)
Date: 2003-03-03 15:32

Message:
Logged In: YES 
user_id=92689

Ok, done, including a minor patch to Doc/lib/libos.tex. I also adapted the Misc/NEWS items. I'm not sure how to change the os.listdir() doco to better reflect the actual situation without mentioning Py_FileSystemDefaultEncoding...

----------------------------------------------------------------------

Comment By: Martin v. Löwis (loewis)
Date: 2003-03-03 14:11

Message:
Logged In: YES 
user_id=21627

Looks good, but incomplete: If the argument is Unicode,
*all* results should be Unicode. There should also be
documentation changes.

----------------------------------------------------------------------

Comment By: Just van Rossum (jvr)
Date: 2003-03-03 14:02

Message:
Logged In: YES 
user_id=92689

I've attached a patch that fixes the bug as well as addresses the unicode arg vs. return value inconsistency that Martin noted. The exception behavior has not yet been changed.

----------------------------------------------------------------------

Comment By: Just van Rossum (jvr)
Date: 2003-03-03 13:22

Message:
Logged In: YES 
user_id=92689

Jack, as noted on #bug 696261, the bug is that os.listdir() doesn't do the right thing with a Unicode string argument (it should use Py_FileSystemDefaultEncoding but it doesn't; I'm working on it.

Martin: I now see that PEP 277 says "Under this proposal, [os.listdir] will return a list of Unicode strings when its path argument is Unicode". I don't like this much (I really think we should push Unicode a little harder onto the users), but I'll look into changing the unix end of os.listdir() to do the same. I'll also review your exception comment.

----------------------------------------------------------------------

Comment By: Martin v. Löwis (loewis)
Date: 2003-03-03 12:36

Message:
Logged In: YES 
user_id=21627

I dislike this change, as it introduces inconsistency across
platforms. On Win32, as a result of PEP 277, Unicode file
names are only returned for Unicode directory names. There
was an explicit discussion about this aspect of PEP 277, and
this interface was accepted as The Right Thing. So I think
Unix should follow here: return byte string file names for
byte string directory names, and Unicode file names for
Unicode directory names. Support for Unicode directory names
should also invoke the file system encoding for the
directory name.

I'm also unsure about the exception handling. If there is a
file name that doesn't decode according to the file system
encoding, it raises the Unicode error. This means that all
other file names are lost. This might be acceptable if the
Unicode-in-Unicode-out strategy is used; in its current
form, the change can and will break existing applications
(which find all kinds of funny byte sequences on disk that
don't work with the user's file system encoding).

----------------------------------------------------------------------

Comment By: Jack Jansen (jackjansen)
Date: 2003-03-03 12:23

Message:
Logged In: YES 
user_id=45365

I think this patch does more bad than good.

A practical problem is that os.path.walk doesn't work anymore if there are 
non-ascii directories in the directory tree (os.listdir will return these as unicode names, but doesn't accept unicode on input). See bug #696261. An additional problem is that various other methods in posix don't do the unicode conversion, so for instance os.getcwd() will return 8-bit strings in Py_FileSystemDefaultEncoding which are incompatible with the unicode returned by listdir.

My preferred solution would be to do the unicode trick everywhere. Second best would be to retract the whole thing and think about it a bit more for Python 2.4.

----------------------------------------------------------------------

Comment By: Just van Rossum (jvr)
Date: 2003-02-25 22:52

Message:
Logged In: YES 
user_id=92689

Checked in as rev. 2.287 of Modules/posixmodule.c. Leaving this item open for now, in case MvL has comments when he gets back.

----------------------------------------------------------------------

Comment By: Guido van Rossum (gvanrossum)
Date: 2003-02-25 18:22

Message:
Logged In: YES 
user_id=6380

OK, check it in, just be prepared for contingencies. I
really cannot judge whether this is right on all platforms.

----------------------------------------------------------------------

Comment By: Just van Rossum (jvr)
Date: 2003-02-25 16:55

Message:
Logged In: YES 
user_id=92689

Having missed 2.3a2, I'd like to get this in way ahead of 2.3b1. Any objections?

----------------------------------------------------------------------

Comment By: Just van Rossum (jvr)
Date: 2003-02-10 19:17

Message:
Logged In: YES 
user_id=92689

I'm pretty sure os.path deals just fine with unicode strings (it's all pure string manipulations, isn't it?)

Worries: well, apparently on Windows os.listdir() has been returning unicode for some time, so it's not like we're breaking completely new grounds here.

If anything breaks it's probably good this happens, as it gives an opportunity to fix things... I just found several example of potential breakage: _bsddb.c parses a filename arg with the "z" format specifier. gdbmmodule.c uses "s". bsddbmodule.c and dbmmodule.c as well.

I'm not sure the above modules work on Windows with non-ascii filenames at all, but it doesn't look like it. Besides Windows (for which my patch is not relevant), only OSX sets Py_FileSystemDefaultEncoding, so any new breakage won't reach a mass market right away <wink>.

----------------------------------------------------------------------

Comment By: M.-A. Lemburg (lemburg)
Date: 2003-02-10 18:46

Message:
Logged In: YES 
user_id=38388

Ok, let's look at it from a different
angle: things that you get from os.listdir() should be
compatible 
to (at least) all the os.path tools and os itself.
Converting to 
Unicode has the advantage that slicing and indexing into the
path names will not break the paths (unlike UTF-8 encoded 8-bit
strings which tend to break when you slice them).

That said, I think you're right about the ASCII approach
provided
that the os, os.path tools can actually properly cope with
Unicode.

What I worry about is that if os.listdir() gives back
Unicode for
e.g. Latin-1 filenames and the application then passes the
Unicode
names to a C API using "s", prefectly working code will break...
then again the C code should really use "es" for decoding to
the Py_FileSystemDefaultEncoding as is done in e.g.
fileobject.c.

I really don't know what to do here...

----------------------------------------------------------------------

Comment By: Just van Rossum (jvr)
Date: 2003-02-10 17:24

Message:
Logged In: YES 
user_id=92689

Here's an argument for ASCII and against the default encoding: if the default encoding is different from Py_FileSystemDefaultEncoding, things go wrong: an 8-bit string passed to file() will be interpreted as Py_FileSystemDefaultEncoding (more precisely: will not be interpreted at all), not the default encoding...

----------------------------------------------------------------------

Comment By: M.-A. Lemburg (lemburg)
Date: 2003-02-10 12:24

Message:
Logged In: YES 
user_id=38388

Right, except that injecting Unicode into Unicode-unaware code
can be dangerous (e.g. some code might require a string object
to work on).

E.g. if someone sets the default encoding to Latin-1 he wouldn't
expect os.listdir() to suddenly return Unicode for him.

This may be a problem in general for the change to os.listdir().
We'll just have to see what happens during the alpha and beta
phases.

----------------------------------------------------------------------

Comment By: Just van Rossum (jvr)
Date: 2003-02-10 12:08

Message:
Logged In: YES 
user_id=92689

On the other hand, if it's not ASCII, wouldn't a unicode string be more appropriate to begin with? If it's encodable with the default encoding, this will happen as soon as the string is used in a piece of unicode-unaware code, right?

----------------------------------------------------------------------

Comment By: M.-A. Lemburg (lemburg)
Date: 2003-02-10 11:55

Message:
Logged In: YES 
user_id=38388

Good question. The default encoding would better fit 
into the concept, I guess.

Instead of PyUnicode_AsASCIIString(v) you'd
have to use PyUnicode_AsEncodedString(v, NULL, "strict").


----------------------------------------------------------------------

Comment By: Just van Rossum (jvr)
Date: 2003-02-10 11:49

Message:
Logged In: YES 
user_id=92689

Ok, I went for your original suggestion: always convert to unicode and then try to convert to ascii. See new patch. Or should this use the default encoding? Hm.

----------------------------------------------------------------------

Comment By: M.-A. Lemburg (lemburg)
Date: 2003-02-10 11:17

Message:
Logged In: YES 
user_id=38388

The file system does not need to support embedded \0 chars
even if it supports UTF-16. It only happens that your test
assumes
that you have one byte per characters encodings which may not
always be true. With UTF-16 your test will see lots of \0 bytes
but not necessarily ones which are ord(x)>=128.

I'm not sure whether other variable length encodings can result
in \0 bytes, e.g. the Asian ones. 

There's also the possibility of the
encoding mapping the ASCII range to other non-ASCII characters,
e.g. ShiftJIS does this for the Yen sign.

If you absolutely want to use the simple test, I'd at least
restrict
the test to an ASCII isalnum(x) test and then try the
encode/decode 
method I described if this test fails.

Note that isalnum() can be locale dependent on some
platforms, so
you have to hard-code it.


----------------------------------------------------------------------

Comment By: Just van Rossum (jvr)
Date: 2003-02-10 10:51

Message:
Logged In: YES 
user_id=92689

I don't see hot UTF-16 could be a valid value for Py_FileSystemDefaultEncoding, as for most platforms the file name can't contain null bytes. My looking at the NAMELEN() spaghetti, it seems platforms without HAVE_DIRENT_H might still support embedded null bytes. Any wisdom on this?

----------------------------------------------------------------------

Comment By: M.-A. Lemburg (lemburg)
Date: 2003-02-10 10:24

Message:
Logged In: YES 
user_id=38388

Your test will probably catch most cases, but it could fail
for e.g. UTF-16.

The only true test would be to first convert to Unicode and then
try to convert back to ASCII. If you get an error you can be
sure that
the text is not ASCII compatible. Given that .listdir()
involves lots of
IO I think the added performance hit wouldn't be noticable.

----------------------------------------------------------------------

Comment By: Just van Rossum (jvr)
Date: 2003-02-10 10:12

Message:
Logged In: YES 
user_id=92689

Applied both suggestions.

However, I'm not sure if my ASCII test does the right thing, or at least I don't think it does if Py_FileSystemDefaultEncoding is not a superset of ASCII.

----------------------------------------------------------------------

Comment By: Neal Norwitz (nnorwitz)
Date: 2003-02-10 04:07

Message:
Logged In: YES 
user_id=33168

The code which uses unicode APIs should probably be wrapped 
with:

#ifdef Py_USING_UNICODE
 /* code */
#endif


----------------------------------------------------------------------

Comment By: Guido van Rossum (gvanrossum)
Date: 2003-02-10 02:16

Message:
Logged In: YES 
user_id=6380

At the very least, I'd like it to return Unicode only when
the original string isn't just ASCII.

----------------------------------------------------------------------

You can respond by visiting: 
https://sourceforge.net/tracker/?func=detail&atid=305470&aid=683592&group_id=5470