freebsd 12 FUSE xattr issues

[ThomasWaldmann]

The xattrs in the native freebsd filesystem are set like this, then backed up with borg:

   xattr.setxattr(os.path.join(self.input_path, 'fusexattr'), 'user.foo', b'bar')
   xattr.setxattr(os.path.join(self.input_path, 'fusexattr'), 'user.empty', b'')

Then the archive is fuse mounted and test_fuse executed:

freebsd 12 amd64 VM (see Vagrantfile), current master branch, no fakeroot:

freebsd64: >           assert sorted(no_selinux(xattr.listxattr(out_fn))) == ['user.empty', 'user.foo', ]
freebsd64: E           AssertionError: assert ['empty', 'foo'] == ['user.empty', 'user.foo']
freebsd64: E             At index 0 diff: 'empty' != 'user.empty'

borg 1.2.0.dev569+g23bd2345.d20180704

---

freebsd 12 amd64 VM (see Vagrantfile), current platform-xattr branch, no fakeroot:

borg 1.2.0.dev560+gadb298d1

behaves the same after setting borg.xattr.XATTR_FAKEROOT = False so that same test code is executed.

---

So it seems that something is stripping away the "user." from the xattr keys.

[ThomasWaldmann]

freebsd 12 amd64 VM (see Vagrantfile), current platform-xattr branch, no fakeroot:

When having XATTR_FAKEROOT = True, this happens:

                if not xattr.XATTR_FAKEROOT and xattr.is_enabled(self.input_path):
                    assert sorted(no_selinux(xattr.listxattr(out_fn))) == ['user.empty', 'user.foo', ]
                    assert xattr.getxattr(out_fn, 'user.foo') == b'bar'
                    # Special case: getxattr returns None (not b'') when reading an empty xattr.
                    assert xattr.getxattr(out_fn, 'user.empty') is None
                else:
>                   assert xattr.listxattr(out_fn) == []

src/borg/testsuite/archiver.py:2192: 
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
src/borg/platform/freebsd.pyx:62: in borg.platform.freebsd.listxattr
    n, buf = _listxattr_inner(func, path)
src/borg/platform/xattr.py:72: in _listxattr_inner
    n = _check(func(path, buf, size), path, detect_buffer_too_small=True)
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _

rv = -1, path = b'/tmp/tmpyrlor7jv/mountpoint/input/fusexattr', detect_buffer_too_small = True

    def _check(rv, path=None, detect_buffer_too_small=False):
        if rv < 0:
            e = get_errno()
            if detect_buffer_too_small and e == errno.ERANGE:
                # listxattr and getxattr signal with ERANGE that they need a bigger result buffer.
                # setxattr signals this way that e.g. a xattr key name is too long / inacceptable.
                raise BufferTooSmallError
            else:
                try:
                    msg = os.strerror(e)
                except ValueError:
                    msg = ''
                if isinstance(path, int):
                    path = '<FD %d>' % path
>               raise OSError(e, msg, path)
E               OSError: [Errno 22] Invalid argument: b'/tmp/tmpyrlor7jv/mountpoint/input/fusexattr'

src/borg/platform/xattr.py:55: OSError

So listxattr on a file that is visible via the fuse mount and has not xattrs set gives errno 22.

[ThomasWaldmann]

A quick test in the native freebsd filesystem with the platform-xattr code showed nothing special, works as expected.

[ThomasWaldmann]

So, looks like some FreeBSD 12 fuse bugs.

[ThomasWaldmann]

Still seeing the xattr name issue on a fresh freebsd 12 release install with all updates.

[asomers]

This sounds to me like a bug in borgbackup, not FreeBSD. But I can't be sure unless you can post a reproduction case that uses lsextattr(1) or extattr_list_file(2). I say that I don't think it's a bug in FreeBSD, because extattr_list_file is supposed to return just the attribute names, not prefixed with their namespaces.

[ThomasWaldmann]

Well, we are running this test on different platforms (fbsd, macOS, linux) and IIRC also used to run it on FreeBSD < 12.

And IIRC it only fails on FBSD 12, this is why I filed this ticket.

[asomers]

If you think it's a bug in FreeBSD, could you please try to reproduce it independently of borg-backup so a FreeBSD developer like myself can debug it?

[ThomasWaldmann]

I can try, but I usually don't write C code for such stuff and also usually do not work on fbsd, I just use it for platform testing.

[ThomasWaldmann]

Also, as this is about borg's fuse code working together with fuse code of freebsd, guess it will need borg for that part.

[asomers]

C code isn't necessary. sh code will work just fine using the command-line utilities lsextattr, getextattr, rmextattr, and setextattr.

[ThomasWaldmann]

script

mkdir input
touch input/file-with-xattr
setextattr user user.foo bar input/file-with-xattr
setextattr user user.empty '' input/file-with-xattr
setextattr user xxx.yyy zzz input/file-with-xattr

echo "os level xattrs (NS 'user'):"
lsextattr user input/file-with-xattr

borg init -e none repo
borg create repo::archive input

mkdir mnt
borg mount repo::archive mnt
echo "fuse level xattrs (NS 'user'):"
lsextattr user mnt/input/file-with-xattr
umount mnt

rm -rf mnt
borg delete repo
rm -rf input

testing

using borg 1.2.0 ~ alpha5 (alpha4 should also work). i also checked borg 1.1.7 from freebsd, showed same behaviour (although our xattr code had a major refactoring in between).

output freebsd 12

os level xattrs (NS 'user'):
input/file-with-xattr	user.foo	user.empty	xxx.yyy

fuse level xattrs (NS 'user'):
mnt/input/file-with-xattr	empty	foo

output freebsd 10.3

os level xattrs (NS 'user'):
input/file-with-xattr	user.foo	user.empty	xxx.yyy

fuse level xattrs (NS 'user'):
lsextattr: mnt/input/file-with-xattr: failed: Operation not supported

notes

10.3: hmm, maybe the failure for 10.3 somehow triggered a different path in our test suite (like "fuse does not support xattrs")

12: see how it reports key names with "user." on OS level, but not on fuse level. also, the "xxx.yyy" key is completely missing.

I admit the key names are a bit weird because they put a key named "user.foo" into the "user" namespace, but I guess at least the FUSE level should have same result as OS level.

The weirdness comes from linux not having an extra param for the namespace and we have the same tests for all platforms.