threads / discuss / 21749

jgit problems for file paths with non-ASCII characters

Subject: jgit problems for file paths with non-ASCII characters

## tl;dr

10 messages between Nov 25, 2009 and Nov 26, 2009.

replies: 9people: 5as markdown or json

Marc Strapetz· Nov 25, 2009, 13:47 UTC · lore

I have noticed that jgit converts file paths to UTF-8 when querying the repository. Especially, org.eclipse.jgit.treewalk.filter.PathFilter#PathFilter performs this conversion:

  private PathFilter(final String s) {
    pathStr = s;
    pathRaw = Constants.encode(pathStr);
  }

Because of this conversion, a TreeWalk fails to identify a file with German umlauts. When using platform encoding to convert the file path to bytes:

  private PathFilter(final String s) {
    pathStr = s;
    pathRaw = s.getBytes();
  }

the TreeWalk works as expected. Actually, the file path seems to be stored with platform encoding in the repository.

Is this a bug or a misconfiguration of my repository? I'm using jgit (commit e16af839e8a0cc01c52d3648d2d28e4cb915f80f) on Windows.

Thanks!

-- Best regards, Marc Strapetz ============= syntevo GmbH http://www.syntevo.com http://blog.syntevo.com

Robin Rosenberg· Nov 25, 2009, 21:11 UTC · re: Marc Strapetz · lore

Re: jgit problems for file paths with non-ASCII characters

onsdag 25 november 2009 14:47:25 skrev  Marc Strapetz:

Show 24 quoted lines

> I have noticed that jgit converts file paths to UTF-8 when querying the
> repository. Especially,
> org.eclipse.jgit.treewalk.filter.PathFilter#PathFilter performs this
> conversion:
>
>   private PathFilter(final String s) {
>     pathStr = s;
>     pathRaw = Constants.encode(pathStr);
>   }
>
> Because of this conversion, a TreeWalk fails to identify a file with
> German umlauts. When using platform encoding to convert the file path to
> bytes:
>
>   private PathFilter(final String s) {
>     pathStr = s;
>     pathRaw = s.getBytes();e pr
>   }
>
> the TreeWalk works as expected. Actually, the file path seems to be
> stored with platform encoding in the repository.
>
> Is this a bug or a misconfiguration of my repository? I'm using jgit
> (commit e16af839e8a0cc01c52d3648d2d28e4cb915f80f) on Windows.

A bug.

The problem here is that we need to allow multiple encodings since there is no reliable encoding specified anywhere. The approach I advocate is the one we use for handling encoding in general. I.e. if it looks like UTF-8, treat it like that else fallback. This is expensive however and then we have all the other issues with case insensitive name and the funny property that unicode has when it allows characters to be encoding using multiple sequences of code points as empoloyed by Apple.

-- robin

-- robin

Shawn O. Pearce· Nov 26, 2009, 00:54 UTC · re: Robin Rosenberg · lore

Re: [egit-dev] Re: jgit problems for file paths with non-ASCII characters

Robin Rosenberg <robin.rosenberg@dewire.com> wrote:

> onsdag 25 november 2009 14:47:25 skrev  Marc Strapetz:
> > I have noticed that jgit converts file paths to UTF-8 when querying the
> > repository.

...

Show 7 quoted lines

> > Is this a bug or a misconfiguration of my repository? I'm using jgit
> > (commit e16af839e8a0cc01c52d3648d2d28e4cb915f80f) on Windows.
> 
> A bug. 
> 
> The problem here is that we need to allow multiple encodings since there
> is no reliable encoding specified anywhere.

This is a design fault of both Linux and git. git gets a byte sequence from readdir and stores that as-is into the repository. We have no way of knowing what that encoding is. So now everyone touching a Git repository is screwed.

> The approach I advocate is
> the one we use for handling encoding in general. I.e. if it looks like UTF-8,
> treat it like that else fallback. This is expensive however

We should try to work harder with the git-core folks to get character set encoding for file names worked out. We might be able to use a configuration setting in the repository to tell us what the proper encoding should be, and if not set, assume UTF-8.

> and then we have
> all the other issues with case insensitive name and the funny property that
> unicode has when it allows characters to be encoding using multiple sequences
> of code points as empoloyed by Apple.

But as you said, this still doesn't make the Apple normal form any easier. Though if we know we are on such a strange filesystem we might be able to assume the paths in the repository are equally damaged. Or not.

-- 
Shawn.

Thomas Singer· Nov 26, 2009, 13:09 UTC · re: Shawn O. Pearce · lore

Re: [egit-dev] Re: jgit problems for file paths with non-ASCII characters

> But as you said, this still doesn't make the Apple normal form
> any easier.  Though if we know we are on such a strange filesystem
> we might be able to assume the paths in the repository are equally
> damaged.  Or not.

Well, if the git-core folks could standardize on, e.g., composed UTF-8 (rather then just UTF-8), for storing file names in the repository, then everything should be clear, isn't it?

-- Best regards, Thomas Singer ============= syntevo GmbH http://www.syntevo.com http://blog.syntevo.com

Shawn O. Pearce wrote:

Show 36 quoted lines

> Robin Rosenberg <robin.rosenberg@dewire.com> wrote:
>> onsdag 25 november 2009 14:47:25 skrev  Marc Strapetz:
>>> I have noticed that jgit converts file paths to UTF-8 when querying the
>>> repository.
> ...
>>> Is this a bug or a misconfiguration of my repository? I'm using jgit
>>> (commit e16af839e8a0cc01c52d3648d2d28e4cb915f80f) on Windows.
>> A bug. 
>>
>> The problem here is that we need to allow multiple encodings since there
>> is no reliable encoding specified anywhere.
> 
> This is a design fault of both Linux and git.  git gets a byte
> sequence from readdir and stores that as-is into the repository.
> We have no way of knowing what that encoding is.  So now everyone
> touching a Git repository is screwed.
> 
>> The approach I advocate is
>> the one we use for handling encoding in general. I.e. if it looks like UTF-8,
>> treat it like that else fallback. This is expensive however
> 
> We should try to work harder with the git-core folks to get character
> set encoding for file names worked out.  We might be able to use a
> configuration setting in the repository to tell us what the proper
> encoding should be, and if not set, assume UTF-8.
> 
>> and then we have
>> all the other issues with case insensitive name and the funny property that
>> unicode has when it allows characters to be encoding using multiple sequences
>> of code points as empoloyed by Apple.
> 
> But as you said, this still doesn't make the Apple normal form
> any easier.  Though if we know we are on such a strange filesystem
> we might be able to assume the paths in the repository are equally
> damaged.  Or not.
>

Johannes Schindelin· Nov 26, 2009, 14:47 UTC · re: Thomas Singer · lore

Re: [egit-dev] Re: jgit problems for file paths with non-ASCII characters

Hi,

On Thu, 26 Nov 2009, Thomas Singer wrote:

Show 10 quoted lines

> [someone said, Thomas did not say who]
>
> > But as you said, this still doesn't make the Apple normal form any 
> > easier.  Though if we know we are on such a strange filesystem we 
> > might be able to assume the paths in the repository are equally 
> > damaged.  Or not.
> 
> Well, if the git-core folks could standardize on, e.g., composed UTF-8 
> (rather then just UTF-8), for storing file names in the repository, then 
> everything should be clear, isn't it?

You mean we should do the same thing as Apple with HFS?  Are you serious?

Ciao, Dscho

Thomas Singer· Nov 26, 2009, 15:31 UTC · re: Johannes Schindelin · lore

Re: [egit-dev] Re: jgit problems for file paths with non-ASCII characters

> You mean we should do the same thing as Apple with HFS?  Are you serious?

Yes, I'm serious. IMHO there should be a defined clear encoding used for files names in the repository. Otherwise you don't know what you can expect by reading it - it could mean anything. File names are in fact strings which are based on characters. To convert characters to bytes (or visa versa) you need to know the encoding.

-- Best regards, Thomas Singer ============= syntevo GmbH http://www.syntevo.com http://blog.syntevo.com

Johannes Schindelin wrote:

Show 18 quoted lines

> Hi,
> 
> On Thu, 26 Nov 2009, Thomas Singer wrote:
> 
>> [someone said, Thomas did not say who]
>>
>>> But as you said, this still doesn't make the Apple normal form any 
>>> easier.  Though if we know we are on such a strange filesystem we 
>>> might be able to assume the paths in the repository are equally 
>>> damaged.  Or not.
>> Well, if the git-core folks could standardize on, e.g., composed UTF-8 
>> (rather then just UTF-8), for storing file names in the repository, then 
>> everything should be clear, isn't it?
> 
> You mean we should do the same thing as Apple with HFS?  Are you serious?
> 
> Ciao,
> Dscho

Shawn O. Pearce· Nov 26, 2009, 19:57 UTC · re: Thomas Singer · lore

Re: [egit-dev] Re: jgit problems for file paths with non-ASCII characters

Thomas Singer <thomas.singer@syntevo.com> wrote:

Show 7 quoted lines

> > You mean we should do the same thing as Apple with HFS?  Are you serious?
> 
> Yes, I'm serious. IMHO there should be a defined clear encoding used for
> files names in the repository. Otherwise you don't know what you can expect
> by reading it - it could mean anything. File names are in fact strings which
> are based on characters. To convert characters to bytes (or visa versa) you
> need to know the encoding.

That's likely not going to fly. HFS+ has changed their decomposition rules at least once, which means the byte sequence for the same character sequence would differ, and a tree or commit hash would come out different depending upon which rules you were following. See [1] for details on what HFS+ does.

Also, Linus has previously stated HFS+ chose the worst possible way to encode the names. Getting Linus to admit he was wrong is impossible, getting Linus to accept the HFS+ encoding rules as the standard format used in a Git repository is not likely to happen. Fortunately Linus carries a slightly smaller stick in Git than he used to, but he is quite vocal and people tend to listen.

[1] http://developer.apple.com/mac/library/technotes/tn/tn1150.html#UnicodeSubtleties

-- 
Shawn.

Robin Rosenberg· Nov 26, 2009, 16:44 UTC · re: Thomas Singer · lore

Re: [egit-dev] Re: jgit problems for file paths with non-ASCII characters

torsdag 26 november 2009 14:09:09 skrev  Thomas Singer:

Show 8 quoted lines

> > But as you said, this still doesn't make the Apple normal form
> > any easier.  Though if we know we are on such a strange filesystem
> > we might be able to assume the paths in the repository are equally
> > damaged.  Or not.
>
> Well, if the git-core folks could standardize on, e.g., composed UTF-8
> (rather then just UTF-8), for storing file names in the repository, then
> everything should be clear, isn't it?

Hey, we're trying to enforce composed characters...

-- robin

Marc Strapetz· Nov 26, 2009, 14:25 UTC · re: Shawn O. Pearce · lore

Re: [egit-dev] Re: jgit problems for file paths with non-ASCII characters

> We should try to work harder with the git-core folks to get character
> set encoding for file names worked out.  We might be able to use a
> configuration setting in the repository to tell us what the proper
> encoding should be, and if not set, assume UTF-8.

I agree that this should be the ultimate goal, though the default should better be "system encoding" for compatibility with current git repositories and instead have newer git versions always set encoding to UTF-8. Thus, for our jgit clone I've introduced a system property to configure Constants.PATH_ENCODING set to system encoding. It's used by PathFilter and this resolves my original problem.

I have tried to switch more usages from Constants.CHARACTER_ENCODING to Constants.PATH_ENCODING, but ended up in confusion due to my lack of understanding: primarily because I couldn't tell anymore whether encoded strings were file names or not. Does it make sense to explicitly distinguish encoding usages in that way? We could try to contribute here (and hopefully cause less review effort to jgit developers than the changes itself are worth ;-)

-- Best regards, Marc Strapetz ============= syntevo GmbH http://www.syntevo.com http://blog.syntevo.com

Shawn O. Pearce wrote:

Show 36 quoted lines

> Robin Rosenberg <robin.rosenberg@dewire.com> wrote:
>> onsdag 25 november 2009 14:47:25 skrev  Marc Strapetz:
>>> I have noticed that jgit converts file paths to UTF-8 when querying the
>>> repository.
> ...
>>> Is this a bug or a misconfiguration of my repository? I'm using jgit
>>> (commit e16af839e8a0cc01c52d3648d2d28e4cb915f80f) on Windows.
>> A bug. 
>>
>> The problem here is that we need to allow multiple encodings since there
>> is no reliable encoding specified anywhere.
> 
> This is a design fault of both Linux and git.  git gets a byte
> sequence from readdir and stores that as-is into the repository.
> We have no way of knowing what that encoding is.  So now everyone
> touching a Git repository is screwed.
> 
>> The approach I advocate is
>> the one we use for handling encoding in general. I.e. if it looks like UTF-8,
>> treat it like that else fallback. This is expensive however
> 
> We should try to work harder with the git-core folks to get character
> set encoding for file names worked out.  We might be able to use a
> configuration setting in the repository to tell us what the proper
> encoding should be, and if not set, assume UTF-8.
> 
>> and then we have
>> all the other issues with case insensitive name and the funny property that
>> unicode has when it allows characters to be encoding using multiple sequences
>> of code points as empoloyed by Apple.
> 
> But as you said, this still doesn't make the Apple normal form
> any easier.  Though if we know we are on such a strange filesystem
> we might be able to assume the paths in the repository are equally
> damaged.  Or not.
>

Shawn O. Pearce· Nov 26, 2009, 20:03 UTC · re: Marc Strapetz · lore

Re: [egit-dev] Re: jgit problems for file paths with non-ASCII characters

Marc Strapetz <marc.strapetz@syntevo.com> wrote:

Show 11 quoted lines

> > We should try to work harder with the git-core folks to get character
> > set encoding for file names worked out.  We might be able to use a
> > configuration setting in the repository to tell us what the proper
> > encoding should be, and if not set, assume UTF-8.
> 
> I agree that this should be the ultimate goal, though the default should
> better be "system encoding" for compatibility with current git
> repositories and instead have newer git versions always set encoding to
> UTF-8. Thus, for our jgit clone I've introduced a system property to
> configure Constants.PATH_ENCODING set to system encoding. It's used by
> PathFilter and this resolves my original problem.

That's probably a good point, using the system encoding on a repository may produce the file names in a more compatible way with git-core. But we probably don't want the encoding to be a single encoding constant in this JVM, we probably need to support a per-repository configuration of the encoding for path names so that we can eventually move to a non-platform specific encoding.

> I have tried to switch more usages from Constants.CHARACTER_ENCODING to
> Constants.PATH_ENCODING, but ended up in confusion due to my lack of
> understanding: primarily because I couldn't tell anymore whether encoded
> strings were file names or not.

Heh. Yea. There are a number of file name encoding sites. I think everything in the treewalk package, as well as the GitIndex, Tree and DirCache* classes. Also the Patch class and its FileHeader friend.

> Does it make sense to explicitly
> distinguish encoding usages in that way? We could try to contribute here
> (and hopefully cause less review effort to jgit developers than the
> changes itself are worth ;-)

Yes, it does. Because we eventually need to support encodings other than the current UTF-8 we assume for file names, especially if a repository is using the local filesystem encoding and that isn't UTF-8.

-- 
Shawn.