Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Reading Unicode filenames fails on Windows #3771

Closed
vicuna opened this issue Aug 25, 2005 · 37 comments

Comments

@vicuna
Copy link

commented Aug 25, 2005

Original bug ID: 3771
Reporter: administrator
Assigned to: @dra27
Status: resolved (set by @xavierleroy on 2017-09-20T14:05:31Z)
Resolution: fixed
Priority: normal
Severity: major
Target version: 4.06.0 +dev/beta1/beta2/rc1
Fixed in version: 4.06.0 +dev/beta1/beta2/rc1
Category: platform support (windows, cross-compilation, etc)
Tags: patch
Related to: #6695
Parent of: #3786 #3789
Monitored by: mdelahaye @whitequark @ygrek jmeber @dra27 daweil @dbuenzli slipstream

Bug description

Full_Name: spiralvoice
Version: 3.08.4
OS: Windows/MinGW
Submission from: p5481eb87.dip.t-dialin.net (84.129.235.135)

Hi,

in otherlibs\win32unix\windir.c the functions win_findfirst and win_findnext
use WIN32_FIND_DATA which is not Unicode aware:

http://msdn.microsoft.com/library/default.asp?url=/library/en-us/fileio/fs/win32_find_data_str.asp

File attachments

@vicuna

This comment has been minimized.

Copy link
Author

commented Jul 30, 2012

Comment author: @ygrek

Here is an old update to mldonkey's patch which does not rely on iconv and works(ed) with msvc : http://ygrek.org.ua/p/ocaml_unicode.html
It worked ok in 2009, but YMMV today.

@vicuna

This comment has been minimized.

Copy link
Author

commented Sep 11, 2012

Comment author: @damiendoligez

Uploaded the patch referenced by ygrek.

@vicuna

This comment has been minimized.

Copy link
Author

commented Jul 2, 2013

Comment author: @damiendoligez

This is such a huge patch :-(

@vicuna

This comment has been minimized.

Copy link
Author

commented Jul 17, 2013

Comment author: @ygrek

It is unavoidable, because it is exactly that all runtime functions have to be tweaked.. Do you have any recommendations to enhance the patch and/or to make it more acceptable?

@vicuna

This comment has been minimized.

Copy link
Author

commented Jul 16, 2014

Comment author: @ygrek

see also https://github.com/mfp/win32unixw

@vicuna

This comment has been minimized.

Copy link
Author

commented Dec 3, 2014

Comment author: @johnwhitington

Just a quick note to say I ran into this problem with a customer today, with our command line tools. I had to ask them to write a script to rename the file, process it, and then rename it back, which isn't ideal.

What are the barriers to having this patch included? I would be happy to test it, though I've never built OCaml from scratch on Windows before, so it might take some time.

@vicuna

This comment has been minimized.

Copy link
Author

commented Dec 4, 2014

Comment author: @alainfrisch

Can someone explain what the patch does? Does it have the effect of interpreting all filenames (in stdlib/unix) as utf8-encoded strings? Can this break existing code? There seems to be some fallback into interpreting paths differently when not valid utf; what's the rationale? etc...

@vicuna

This comment has been minimized.

Copy link
Author

commented Dec 7, 2014

Comment author: @ygrek

Windows api has two variants of every function that deals with filenames, e.g. RenameFileA and RenameFileW, the first one treats filename as encoded in 1-byte current "locale", the second accepts UTF16. The patch switches (maybe not all places) OCaml runtime to use the second family of functions (through MS specific (IIUC) crt wrappers). The fallback is there to specifically preserve backwards compatibility - if the string is not utf8 - it is treated as 1-byte encoded, otherwise as utf8.

Probably we could remove this fallback and enable it unconditionally based on env variable, or vice-versa - leave the current mode a default and switch to utf8 only on request (function in Sys?). As for backwards compatibility - the breakage may occur if the code assumes 1-byte encoded filenames and builds filenames from other sources other than the readdir (which in utf8 mode would return utf8 strings). For simple programs - which read the filenames from filesystem and then use those filenames - there would be no breakage when switching to utf8.

I personally would vote for switching default to utf8 and leaving a knob for old code. Let me know what you think, I am not developing on windows anymore, but I am willing to spend time on this patch to see it integrated.

@vicuna

This comment has been minimized.

Copy link
Author

commented Dec 7, 2014

Comment author: @ygrek

BTW the first version of this patch (iconv-dependent) is used in mldonkey, the second one was used in commercial windows application. I wonder how unison people deal with this problem..

@vicuna

This comment has been minimized.

Copy link
Author

commented Dec 7, 2014

Comment author: @alainfrisch

Thanks ygrek for the useful background information. The patch would easily break existing code. For instance, in LexiFi application, we store filenames (currently encoded in the current "locale") in databases, we retrieve them from file pickers (using .Net dialogs and mapping .Net strings into OCaml Latin-1 strings, failing on non Latin-1 ones), etc. It would make sense to switch all filename-related strings to utf8 internally, but this requires some non-trivial migration path. I assume most existing deployed OCaml Windows applications would suffer from the same problem.

The fallback is a little bit dangerous. Interpreting OCaml strings in the current locales when not valid utf8 is probably harmless in practice; but there would still the problem that strings returned by Sys/Unix module (e.g. from readdir) would now be encoded in utf8, and application need to be adapated to deal with that (e.g. if they put that in a GUI).

A global mode switch might be preferable (perhaps in addition to the fallback when utf8 mode is enabled).

It would indeed be interesting to hear about how e.g. Unison deal currently with the problem.

@vicuna

This comment has been minimized.

Copy link
Author

commented Dec 9, 2014

Comment author: @ygrek

Unison has separate windows specific wrapper that uses Unicode-family of winapi, see https://github.com/nitrous-io/unison/blob/master/system/system_win_stubs.c (not official repo)

@vicuna

This comment has been minimized.

Copy link
Author

commented Dec 9, 2014

Comment author: @alainfrisch

I'm convinced about the need for such a change.

One general question: which of the 6 Windows ports should be impacted? Clearly msvc32/64. Also mingw32/64, I guess. What about Cygwin32/64?

About the patch itself:

  • There are two copies of u8tou16 (one for the core runtime system, one for win32unix). Couldn't they be merged?

  • I'd propose to define shared helpers so that code such as:

char * temp=String_val(name);
WCHAR * wtemp;
int retcode;
if(is_valid_utf8(temp))
wtemp = utf8_to_utf16(temp);
else
wtemp = ansi_to_utf16(temp);

can be written as:

WCHAR *wtemp = to_u16(String_val(name));

and:

tempo = utf16_to_utf8(fileinfo.cFileName);
valname = copy_string(tempo);
free(tempo);

becomes:

valname = caml_string_u8_of_u16(fileinfo.cFileName);

These helpers will also help for checking e.g. a global flag that control the desired behavior.

  • One should use functions such as caml_stat_free, etc.

  • The patch conflicts with changes related to the global runtime lock (e.g. in caml_sys_rename). Actually, the new code looks like:

char * p_old;
char * p_new;
int ret;
p_old = caml_strdup(String_val(oldname));
p_new = caml_strdup(String_val(newname));
caml_enter_blocking_section();
ret = rename(p_old, p_new);
caml_leave_blocking_section();
caml_stat_free(p_new);
caml_stat_free(p_old);

which, in essence, is closer to what needs to be done on Windows (i.e. copy the string, and finally free the copy). This should give an opportunity for more sharing between the Windows and non-Windows versions (defining a macro/typedef that expand either to "char*" or "WCHAR*" depending on the system, and functions to map between OCaml string values and that type).

  • Is there a point to have a specific macro UTF16? Shouldn't it be equivalent to e.g. _WIN32? The code under UTF16 uses the Windows API, it couldn't be used under non-Windows systems, and I don't think that we should give the option to disable the UTF16 stuff for Windows port at configure time.

  • Is it relevant to check e.g. for #ifdef HAS_GETCWD in the Windows/UTF16 code path?

@vicuna

This comment has been minimized.

Copy link
Author

commented Dec 9, 2014

Comment author: @ygrek

I believe cygwin doesn't need this change - it should take care of that as it pretends to be unix.

Agreed on the helpers.
Will need to check whether mingw port has _WIN32 defined and cygwin doesn't.
It will take me some time to setup windows environment, hold on.

@vicuna

This comment has been minimized.

Copy link
Author

commented Dec 9, 2014

Comment author: @alainfrisch

I believe cygwin doesn't need this change - it should take care of that as it pretends to be unix.

That was my guess, but as far as I understand, Unix doesn't say anything about the encoding of filenames (they are just sequences of bytes with some disallowed bytes), so I wonder how they define the mapping.

[edit] Found some information here: https://cygwin.com/cygwin-ug-net/using-specialnames.html It seems that the mapping depends on the current locale. This seems ugly, but this is a quite independent problem from the current proposal.

It will take me some time to setup windows environment, hold on.

No problem, and thanks for your effort!

@vicuna

This comment has been minimized.

Copy link
Author

commented Dec 9, 2014

Comment author: @dbuenzli

I didn't have a close look but the UTF bits look dodgy, e.g. UTF8_LENGTH tells us that an UTF-8 sequence can be 5 to 6 bytes long which is wrong, 4 bytes is the max. Also UNICODE_VALID seems to exclude non-characters but I don't think it should (usually you want to deal with the full range of Unicode scalar values).

@vicuna

This comment has been minimized.

Copy link
Author

commented Dec 9, 2014

Comment author: @alainfrisch

Couldn't we use MultiByteToWideChar and WideCharToMultiByte to convert between WCHAR and UTF-8 ?

@vicuna

This comment has been minimized.

Copy link
Author

commented Dec 9, 2014

Comment author: @ygrek

that's what the patch does actually

@vicuna

This comment has been minimized.

Copy link
Author

commented Dec 10, 2014

Comment author: @alainfrisch

Ah yes, sorry, I was confused by the beginning of u8tou16 (and dbuenzli's comment); this code is only used to validate UTF-8. Cannot this be achieved with MultiByteToWideChar (with dwflags = MB_ERR_INVALID_CHARS)? The doc says: << Starting with Windows Vista, this function fully conforms with the Unicode 4.1 specification for UTF-8 and UTF-16. >> so I assume that, except for older Windows, the function would perform a full validity check (I don't know exactly what it does for XP).

@vicuna

This comment has been minimized.

Copy link
Author

commented Dec 11, 2014

Comment author: @alainfrisch

What about environment variables? It's quite common to pass filenames/paths in environment variables. Shouldn't we switch to GetEnvironmentVariableW and co?

@vicuna

This comment has been minimized.

Copy link
Author

commented Dec 11, 2014

Comment author: @whitequark

Same concern applies for WinMainW.

@vicuna

This comment has been minimized.

Copy link
Author

commented Mar 10, 2015

Comment author: @ygrek

#153 addresses Alain's comments. TODO remaining : exposing runtime switch and use winapi for utf8 validation

@vicuna

This comment has been minimized.

Copy link
Author

commented Feb 17, 2017

Comment author: @xavierleroy

Tentatively shooting for release 4.06. #2516 on Github is a first cut but needs some work. Any help welcome.

@vicuna

This comment has been minimized.

Copy link
Author

commented May 18, 2017

Comment author: cfranchini

Hi,
after some work. I rebase and rewrite the patch to be applied on current repository. I attached to that ticket a pack.tag.gz which contain 3 patches.

It's the same patch but applied to three different commits.

  • 4.04.patch
    patch for branch 4.04
    apply on commit 34400b1

  • 4.05.patch
    patch for branch 4.05
    apply on commit dafb259

  • trunk.patch
    patch for branch trunk
    apply on commit 472ee2c

I also add my test folder utf16-test which contains different tests for those commits. I hope it would help.

@vicuna

This comment has been minimized.

Copy link
Author

commented Jun 8, 2017

Comment author: @alainfrisch

Thanks for putting some energy into it. (I think that you can safely focus on working with trunk. Such a large change is unlikely to be integrated on a release branch.)

I did not look at the updated patch closely, but it seems to me that several functions still need to be switched to the Unicode-aware variants. In particular, all accesses to the environment (getenv/putenv/environ), and functions to create process (execv and friend, and CreateProcess). Also, we probably need to do something for the command-line arguments of the current process. Can you confirm?

@vicuna

This comment has been minimized.

Copy link
Author

commented Jun 12, 2017

Comment author: @nojb

I opened a new PR at #1200 with cfranchini's trunk patch with the intention of pushing this through and getting it to a mergeable state.

So far:

  • Fixed several issues arising from the rebase/rewrite operation to the latest trunk
  • Switched to using WinAPI for UTF8 validation and also fixed a couple of bugs in the existing use of WinAPI for UTF-16->UTF-8 conversion.
  • Integrated cfranchini's tests into the OCaml testsuite.
  • Added wrappers for Sys.getenv and Sys.command

Apart from completing the patch (wrapping environment, process creation functions), one key point to think about next: how to minimize/cleanup the use of #ifdefs and make the UTF8 conversion a runtime switch.

@vicuna

This comment has been minimized.

Copy link
Author

commented Jun 12, 2017

Comment author: @dra27

Regarding the runtime switch - especially with things implemented as #define's at the moment, this is going to be a fair amount of work resulting in quite a lot of code duplication (at least some functions will need to be compiled twice) and we'd be doing this to maintain a totally broken way of handling filenames from before.

I'm toying with the proposal of making switching to UTF-8 a blocking part on Windows of upgrading to this version of OCaml, so not providing compatibility with the old behaviour (obviously, we could provide documentation and guidance on what you'd need to do!). So, rather than having a runtime switch, could we consider having some kind of new compile time action which is required - i.e. some way of declaring to the compiler that you are aware of the implications of the UTF-8 change and which triggers a warning (or possibly even error?) if you use any of the affected functions? That could be a function you need to call (e.g. a UW IMAP-style Sys.i_accept_the_risk ()), or a module/linker flags which must also be included (e.g. -cclib -lwin32utf8 or win32utf8.cmo), or an actual flag to ocamlc/ocamlopt (e.g. -win32-utf8). We could then in a few versions time relax this requirement (it's like the reverse of a deprecation warning).

Windows programs which want to compile both with this and with old versions will certainly have to worry about the two different modes, so I'm not necessarily seeing the downside of embracing this in the compiler itself? Programs which don't use filename-related functions at all would be unaffected.

@vicuna

This comment has been minimized.

Copy link
Author

commented Jun 12, 2017

Comment author: @nojb

I agree that there isn't an obvious way to do the runtime switch without overly complicating or duplicating the code.

It would be nice to reach a consensus on whether a compile-time or compiler switch is a good solution before refactoring this part of the code.

@vicuna

This comment has been minimized.

Copy link
Author

commented Jun 12, 2017

Comment author: @dra27

Alain - thoughts?

@vicuna

This comment has been minimized.

Copy link
Author

commented Jun 12, 2017

Comment author: @dbuenzli

What behaviour does exactly change with the fix ? Isn't it the case that before you could simply not deal with filenames beyond ASCII ? If that is the case I'm not sure you need a compat story on that.

@vicuna

This comment has been minimized.

Copy link
Author

commented Jun 12, 2017

Comment author: @nojb

Currently ("before" this patch) OCaml strings given to Unix/Sys functions are passed verbatim to the ANSI Windows API, which means they are interpreted in the current Windows codepage (typically Latin-1 in Western Europe).

This means that you cannot deal with filenames outside the Latin-1 (not ASCII) range.

This patch fixes this problem by translating on the OCaml side between UTF-8 and UTF-16 and using the UTF-16 Windows API.

The problem is that this change can break code that uses Latin-1 strings (non ASCII Latin-1 strings are not valid UTF-8) if it is not adapted to re-encode those strings to UTF-8 before passing them to OCaml library functions.

@vicuna

This comment has been minimized.

Copy link
Author

commented Jun 13, 2017

Comment author: @dbuenzli

I see.

It seems to me that the sanest way is to simply provide the functionality under an ocaml configure option rather than try to fiddle with this at the ocaml runtime level or worse to add more flags to ocamlc. In general it seems better if we can avoid to put the burden on build systems for this.

It guess it should be easy to ifdef the function that you interpose between the syscalls to choose between treating the result of String_val as Latin-1 or UTF-8.

The default could be kept to Latin-1 for one or two versions major versions and then switched to UTF-8. Meanwhile the UTF-8 variant could be published under another opam switch name say +win-utf-8.

@vicuna

This comment has been minimized.

Copy link
Author

commented Jun 14, 2017

Comment author: @nojb

A little update on this patch. After discussion with Alain, we settled on the following approach.

  • Switch the runtime system, under Windows, to use Unicode API. OCaml strings passed to the stdlib will be interpreted, either as 1) UTF-8 (in "modern" mode), or 2) current codepage (in "legacy" mode). In both cases they are re-encoded into wide strings using Windows API and then passed to the Unicode API.

  • Same thing going in the other direction: UTF-16 strings returned by the Unicode API are re-encoded as 1) UTF-8 in "modern" mode or 2) current codepage in "legacy" mode.

  • The switch between "modern" and "legacy" can be modified at runtime by calling Sys.enable_windows_unicode_runtime.

  • In fact, there will be a third mode, "transitional" which is like "modern" but falls back to re-encoding from current codepage if the OCaml string if not valid UTF-8. This mode comes with a runtime warning when passing invalid UTF-8.

  • For those bits of code that are compiled under both Linux and Windows a very small compatibility header is necessary.

@vicuna

This comment has been minimized.

Copy link
Author

commented Jun 14, 2017

Comment author: @dbuenzli

TBH I'm not sure I see the advantage of being able to do this at runtime.

I think it's going to give us a longer compatibility story and more erratic behaviour for everyone.

If we do it statically packages could gradually assert they need the ocaml compiler with UTF-8 and refuse to install in Latin-1.

With the runtime switch everyone out there is basically left guessing, from end user to program and package devs. And if you want to be robust you need to check which variant you are actually running which puts the work on the program code.

@vicuna

This comment has been minimized.

Copy link
Author

commented Jun 15, 2017

Comment author: @nojb

To be clear, if we agree on the basic design of switching unconditionally to the Unicode API and doing the UTF-8 or CODEPAGE conversion on the OCaml side (the first two points in my previous note), the (runtime) switch comes for free. We can discuss how to expose it; it may very well be the case that a static switch is the best for opam-delivered compilers.

Having said that, I think one argument for having a runtime switch was that it can help incremental migration for large codebases where it is not feasible to do it all at once due to the need to preserve backwards compatibility (this is the case at LexiFi), as well as making it easier to test (for example, the current testsuite in the PR is testing both the "legacy" and "modern" codepaths).

@vicuna

This comment has been minimized.

Copy link
Author

commented Jun 15, 2017

Comment author: @dbuenzli

I still don't see any compelling argument to expose that to programs at the API level via Sys.enable_windows_unicode_runtime. If you need easier testing you could do this via an env var and/or compiling against a -runtime-variant.

@vicuna

This comment has been minimized.

Copy link
Author

commented Jun 15, 2017

Comment author: @nojb

Yes, I agree. Sys.enable_windows_unicode_runtime is just a proof of concept.

@vicuna

This comment has been minimized.

Copy link
Author

commented Sep 20, 2017

Comment author: @xavierleroy

Unicode filename support was added to trunk and will be in release 4.06. Feel free to test and reopen this MPR if the issue is still there.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Projects
None yet
2 participants
You can’t perform that action at this time.