RFC 30: Unicode Filenames
Authors: Frank Warmerdam
This document describes steps to generally handle filenames as UTF-8 strings in GDAL/OGR. In brief it will be assumed that filenames passed into and returned by GDAL/OGR interfaces are UTF-8. On some operating systems, notably Windows, this will require use of "wide character" interfaces in the low level VSI*L API.
All filenames in the VSI*L API will be treated as UTF-8, which means the cpl_vsil_win32.cpp implementation will need substantial updates to use wide character interfaces.
Old (small file) VSI API
The old VSIFOpen() function will be adapted to use _wfopen() on windows instead of fopen() so that utf-8 filenames will be supported.
Because the path/extension delimiter characters '.', '', '/' and ':' will never appear in the non-ascii portion of utf-8 strings we can safely leave the existing path parsing functions working as they do now. They do not need to be aware of the real character boundaries for exotic characters in utf-8 paths. The following will be left unchanged.
These will all also need to treat filenames as utf-8.
Currently Windows's cpl_vsil_win32.cpp module uses CreateFile() with ascii filenames. It needs to be converted to use CreateFileW() and other wide character functions for stat(), rename, mkdir, etc. Prototype implementation already developed (r20620).
Linux / Unix / MacOS X
On modern linux, unix and MacOS operating systems the fopen(), stat(), readdir() functions already support UTF-8 strings. It is not currently anticipated that any work will be needed on Linux/Unix/MacOS X though there is some question about this. It is considered permissible under the definition of this RFC for old, and substandard operating systems (WinCE?) to support only ASCII, not UTF-8 filenames.
There are a variety of places where general text may contain filenames. One obvious case is the subdataset filenames returned from the SUBDATASET domain. Previously these were just exposed as plain text and interpretation of the character set was undefined. As part of this RFC we state that such filenames should be considered to be in utf-8 format.
I observe with Python 2.6 that functions like gdal.Open() do not accept unicode strings, but they do accept utf-8 string objects. One possible solution is to update the bindings in selective places to identify unicode strings passed in, and transform them to utf-8 strings.
filename = u'xx\u4E2D\u6587.\u4E2D\u6587' if type(filename) == type(u'a'): filename = filename.encode('utf-8')
I'm not sure what the easiest way is to accomplish this in the bindings. The key entries are:
Similarly all interfaces (ie. gdal.ReadDir()) that return filenames will hereafter return unicode objects rather than string objects.
Also note that in Python 3.x strings are always unicode.
Tamas notes that in C# we normally convert the unicode C# strings into C string with the PtrToStringAnsi marshaller. Presumably we will need to use a utf-8 converter for all interface strings considered to be filenames. I would note this should also apploy to OGR string attribute values which are also intended to be treated as utf-8.
(It is unclear who will take care of this aspect since the primary author (FrankW) is not C#-binding-competent.
The general rule in Perl is that all strings should be decoded before giving them to Perl and encoded when they are output. In practice things usually just work. To be sure, I (Ari) have added an explicit decode from utf8 to FindFile and ReadDir (#20800).
No changes are needed for Java. Java strings are unicode, and they are already converted to utf-8 in the java swig bindings. That is, the java bindings already assumed passing and receiving utf-8 strings to/from GDAL/OGR.
On windows argv as passed into main() will not generally be able to represent exotic filenames that can't be represented in the locale charset. It is possible to fetch the commandline and parse it as wide characters using GetCommandLineW() and CommandLinetoArgvW() to capture ucs-16 filenames (easily converted to utf-8); however, this interferes with the use of setargv.obj to expand wildcards on windows.
I have not been able to come up with a good solution, so for now I am not intending to make any changes to the GDAL/OGR commandline utilities to allow passing exotic filenames. So this RFC is mainly aimed at ensuring that other applications using GDAL/OGR can utilize exotic filenames.
The proposed implementation really only addresses file format drivers that use VSIFOpenL(), VSIFOpen() and related functions. Some drivers dependent on external libraries (ie. netcdf) do not have a way to hook the file IO API and may not support utf-8 filenames. It might be nice to be able to distinguish these.
At the very least any driver marked with GDAL_DCAP_VIRTUALIO as "YES" will support UTF-8. Perhaps this opportunity ought to be used to more uniformly apply this driver metadata (done).
We will need to introduce some test suite tests with multibyte utf-8 filenames. In support of that aspects of the VSI*L API - particularly the rename, mkdir, rmdir, functions and VSIFOpenL itself have been exposed in python.
Appropriate API entry points will be documented as taking and return UTF-8 strings.
Implementation is underway and being tracked in ticket #3766.