Re: git fsck complains about error: refs/tags/.cyg000000000559e25517156b51cf219f51/libgcj-2.95.0: badRefName: invalid refname format?!

On Feb  3 22:14, Dan Shelton via Cygwin wrote:
> On Tue, 3 Feb 2026 at 13:20, Corinna Vinschen via Cygwin
> <[email protected]> wrote:
> >
> > On Feb  3 00:55, Dan Shelton via Cygwin wrote:
> > > On Mon, 2 Feb 2026 at 17:04, Corinna Vinschen via Cygwin
> > > <[email protected]> wrote:
> > > >
> > > > On Feb  2 14:47, Dan Shelton via Cygwin wrote:
> > > > > On Mon, 2 Feb 2026 at 14:40, Corinna Vinschen via Cygwin
> > > > > <[email protected]> wrote:
> > > > > >
> > > > > > On Feb  2 13:24, Dan Shelton via Cygwin wrote:
> > > > > > > I'm not sure whether the Cygwin code is correct. I did a peek with a
> > > > > > > kernel debugger, and I see that FILE_RENAME_INFORMATION.RootDirectory
> > > > > > > is always NULL if a file gets renamed to .cyg000000000xxxx. But if I
> > > > > > > try that with NTFS or SMB, the NtSetInformationFile() to set
> > > > > > > FileRenameInformation always fails.
> > > > > >
> > > > > > Your testcase is incorrect, unfortunately.
> > > > > >
> > > > > > >     fri->FileNameLength = (wcslen(dstfile)+1)*sizeof(wchar_t);
> > > > > >
> > > > > > For NT file paths, never count the trailing \0 to the length:
> > > > > >
> > > > > >       fri->FileNameLength = wcslen(dstfile) * sizeof (WCHAR);
> > > > > >
> > > > > > With that, your testcase works fine for me.
> > > > > >
> > > > > > On which filesystem did you see the problem?
> > > > >
> > > > > Windows NFSv3 client (the builtin one, not the newer NFSv4.1 one).
> > > >
> > > > The files in question are actually files which got renamed while
> > > > in use.  I don't know another way to implement removing in-use files
> > > > on remote file systems not supporting delete POSIX semantics.  If
> > > > somebody has a brilliant idea, https://cygwin.com/acronyms/#PTC.
> > >
> > > Did you see that these are directories, not files? How does that happen?
> >
> > Yes, I saw that.  Same as for files.  Basically something like this:
> >
> >   mkdir ("dir");
> >   fd = open ("dir", O_DIRECTORY);
> >   unlink ("dir");
> >   fd2 = openat (fd, "file", O_CREAT...);
> 
> I'm pretty sure this is a bug in Cygwin, unlink() should not work on dirs
> 1. unlink() is only defined for files, but not dirs. POSIX says ...
> The path argument shall not name a directory ...

Yeah, I was lazy.  unlink("dir") returns EISDIR.  Replace with
rmdir("dir").  Under the hood it's all a single function unlink_nt(),
though.

> 2. The /bin/unlink.exe utility says it can only work on files
> 3. Looking at the SystemV sources you see the NFS client code only
> creates .nfsXXXXX files for files, but NEVER for directories

See the code from
https://sourceware.org/cgit/newlib-cygwin/tree/winsup/cygwin/syscalls.cc#n880
to
https://sourceware.org/cgit/newlib-cygwin/tree/winsup/cygwin/syscalls.cc#n919

> 4. Testing this on Solaris 10 Update 8 says the dir is not empty

Testing what on Solaris?  In the above example the directory is created,
then unlinked (yes, rmdir'ed) while it's still empty, and only then a
file is created inside the dir via openat().  I didn't try this myself, it
was just off the top of my head to create an unlinked dir with a file in
it ¯\_(ツ)_/¯

> 5. It really breaks git and git's test suite

I tried the same locally on MSFT NFS3 (git clone gcc.git; git fsck) twice
and couldn't reproduce your problem.


Corinna

-- 
Problem reports:      https://cygwin.com/problems.html
FAQ:                  https://cygwin.com/faq/
Documentation:        https://cygwin.com/docs.html
Unsubscribe info:     https://cygwin.com/ml/#unsubscribe-simple