On Fri, 19 Jan 2007 04:02:42 +0100 Simon 'corecode' Schubert wrote:

quoted text
> Bart Trojanowski wrote:
> >  	/* If input_fd is a file, we should have reached its end now. */
> > -	if (fstat(input_fd, &st))
> > -		die("cannot fstat packfile: %s", strerror(errno));
> > -	if (S_ISREG(st.st_mode) && st.st_size != consumed_bytes)
> > -		die("pack has junk at the end");
> > +        if (input_fd) {
> > +                if (fstat(input_fd, &st))
> > +                        die("cannot fstat packfile: %s", strerror(errno));
> > +                if (S_ISREG(st.st_mode) && st.st_size != consumed_bytes)
> > +                        die("pack has junk at the end");
> > +        }
>
> This is clearly the wrong fix.  input_fd being 0 doesn't mean that
> it is *not* a regular file.  Only doing a fstat can tell.  You are
> simply hiding your real issue there, which is that you can't fstat
> on a pipe or whatever input_fd is.
>
> The problem here is that your 64bit kernel can't fit the data into
> your struct stat provided by your 32bit libc.  Not a problem of git.
> However, it would be interesting to know what exactly produces the
> EOVERFLOW.

Most likely it is the st_ino field - the kernel assigns unique inode
numbers for pipes from an "unsigned long" counter, which is 64-bit in
this case, and *stat() calls must fail with EOVERFLOW if the inode
number does not fit into ino_t, which is 32-bit here.  This problem is
known for some time, and there is even a kernel patch proposed as a
workaround (which makes the counter 32-bit):

http://permalink.gmane.org/gmane.linux.file-systems/12526

AFAIK, that patch is not upstream yet - so upgrading the kernel did
not really fix the issue, it will appear again once the system will
use more than 4G of pipe inodes.

Compiling git with -D_FILE_OFFSET_BITS=64 will make ino_t 64-bit and
therefore will fix the problem (however, I'm not sure whether the git
code is ready for this).