Re: [9fans] off topic - a good Git reference

[Steffen Nurpmeso <sdaoden@yandex.com>]

erik quanstrom <quanstro@quanstro.net> wrote:
 |On Wed Oct  7 14:25:58 PDT 2015, elbingmiss@gmail.com wrote:
 |in the case of system crash, fsync doesn't provide strong \
 |guarantees that the
 |write will not be lost, or the fs not corrupted even in li\
 |nux.  some versions of
 |the linux kernel do nothing on fsync, fsync doesn't actual\

Linux fsync(2) says that the file itself is fine after it, but for
the filesystem metadata you need an additional fsync(2) on the
directory the file resides in.
The final wording of the related POSIX issue[1] states that the
latter is only necessary if directory entries have been created or
removed (also regarding rename(2) which needs two fsync(2) calls;
luckily we're binary only), and noone has opposed to this wording
in more than two years.
(Mac OS X Snow Leopard had a F_FULLFSYNC fcntl(2) which provided
the fully-synced guarantee, as in
        _os_fsync = os.fsync
        def __fullfsync(fd):
            if not SAFE_FSYNC:
                return _os_fsync(fd)
            try:
                fcntl.fcntl(fd, fcntl.F_FULLFSYNC)
            except IOError as e:
                if e.errno == errno.ENOTTY:
                    return _os_fsync(fd)
                raise
        os.fsync = __fullfsync
which i thought of as a good idea, from a programmers point of
view who wants to get it done.)

  [1] http://austingroupbugs.net/view.php?id=672#c1618

--steffen