Encoding issue on filename with Java 7 on OSX with jnlp/webstart

Question

I have this problem that has been dropped on me, and have been a couple of days of unsuccessful searches and workaround attempts.

I have now an internal java swing program distributed by jnlp/webstart, on osx and windows computers, that, among other things, downloads some files from WebDav.

Recently, on a test machine with OSX 10.8 and Java 7, filenames and directory names with accented characters started having those replaced by question marks.

No problem on OSX with versions of Java before 7.

example :

XXXYYY_è_ABCD/

becomes

XXXYYY_?_ABCD/

using java.text.Normalizer (NFD, NFC, NFKD, NFKC) on the original string, the result is different but still wrong :

XXXYYY_e?_ABCD/

or

XXXYYY_e_ABCD/

I know, from correspondence between [andrew.brygin at oracle.com] and [mik3hall at gmail.com] that

Yes, file.encoding is set based on the locale that the jvm is running on, and if you run your java vm in xxxx.UTF-8 locale, the file.encoding should be UTF-8, set to MacRoman will be problematic. So I believe Oracle/OpenJDK7 behaves correctly. That said, as Andrew Thompson pointed out, if all previous Apple JDK releases use MacRoman as the file.encoding for english/UTF-8 locale, there is a "compatibility" concern here, it might worth putting something in the release note to give Oracle/OpenJDK MacOS user a heads up.

original mail

from Joni Salonen blog (java-and-file-names-with-invalid-characters) i know that :

You probably know that Java uses a “default character encoding” to convert binary data to Strings. To read or write text using another encoding you can use an InputStreamReader or OutputStreamWriter. But for data-to-text conversions deep in the API you have no choice but to change the default encoding.

and

What about file.encoding?

The file.encoding system property can also be used to set the default character encoding that Java uses for I/O. Unfortunately it seems to have no effect on how file names are decoded into Strings.

executing locale from inside the jnlp invariabily prints

LANG=
LC_COLLATE="C"
LC_CTYPE="C"
LC_MESSAGES="C"
LC_MONETARY="C"
LC_NUMERIC="C"
LC_TIME="C"
LC_ALL=

the most similar problem on stackoverflow with a solution is this : encoding-issues-on-java-7-file-names-in-os-x

but the solution is wrapping the execution of the java program in a script with

#!/bin/bash
export LC_CTYPE="UTF-8" # Try other options if this doesn't work
exec java your.program.Here

but I don't think this option is available to me because of the webstart, and I haven't found any way to set the LC_CTYPE environment variable from within the program.

Any solutions or workarounds?

P.S. :

If we run the program directly from shell, it writes the file/directory correctly even on OSX 10+Java 7. The problem appears only with the combination of JNLP+OSX+Java7

Answer 1

I take it it's acceptable to have maximal ASCII representation of the file name, which works in virtually any encoding.

First, you want to use specifically NFKD, so that maximum information is retained in the ASCII form. For example, "2⁵" becomes "25"rather than just "2", "fi" becomes "fi" rather than "" etc once the non-ascii and non-control characters are filtered out.

String str = "XXXYYY_è_ABCD/";
str = Normalizer.normalize(str, Normalizer.Form.NFKD);
str = str.replaceAll( "[^\\x20-\\x7E]", "");
//The file name will be XXXYYY_e_ABCD no matter what system encoding

You would then always pass filenames through this filter to get their filesystem name. You only lose is some uniqueness, I.E file asdé.txt is the same as asde.txt and in this system they cannot be differentiated.

Answer 2

EDIT: After experimenting with OS X some more I realized my answer was totally wrong, so I'm redoing it.

If your JVM supports -Dfile.encoding=UTF-8 on the JVM command line, that might fix the issue. I believe that is a standard property but I'm not certain about that.

HFS Plus, like other POSIX-compliant file systems, stores filenames as bytes. But unlike Linux's ext3 filesystem, it forces filenames to be valid decomposed UTF-8. This can be seen here with the Python interpreter on my OS X system, starting in an empty directory.

$ python
Python 2.7.1 (r271:86832, Jul 31 2011, 19:30:53) 
>>> import os
>>> os.mkdir('\xc3\xa8')
>>> os.mkdir('e\xcc\x80')
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
OSError: [Errno 17] File exists: 'e\xcc\x80'
>>> os.mkdir('\x8f')
>>> os.listdir('.')
['%8F', 'e\xcc\x80']
>>> ^D
$ ls
%8F è

This proves that the directory name on your filesystem cannot be Mac-Roman encoded (i.e. with byte value 8F where the è is seen), as long as it's an HFS Plus filesystem. But of course, the JVM is not assured of an HFS Plus filesystem, and SMB and NFS do not have the same encoding guarantees, so the JVM should not assume this scheme.

Therefore, you have to convince the JVM to interpret file and directory names with UTF-8 encoding, in order to read the names as java.lang.String objects correctly.

Answer 3

Shot in the dark: File Encoding does not influence the way how the file names are created, just how the content gets written into the file - check this guy here: http://jonisalonen.com/2012/java-and-file-names-with-invalid-characters/

Here is a short entry from Apple: http://developer.apple.com/library/mac/#qa/qa1173/_index.html

Comparing this to http://docs.oracle.com/javase/tutorial/i18n/text/normalizerapi.html I would assume you want to use

normalized_string = Normalizer.normalize(target_chars, Normalizer.Form.NFD);

to normalize the file names before you pass them to the File constructor. Does this help?