On Tue, Jan 13, 2026 at 03:39:53PM +0000, Al Viro wrote:
> On Tue, Jan 13, 2026 at 03:31:14PM +0000, Mark Brown wrote:
>
> > I'm seeing a regression in -next in the execveat kselftest which bisects
> > to 2a0db5f7653b ("struct filename: saner handling of long names"). The
> > test triggers two new failures with very long filenames for tests that
> > previously succeeded:
> >
> > # # Failed to open length 4094 filename, errno=36 (File name too long)
> > # # Invoke exec via root_dfd and relative filename
> > # # child execveat() failed, rc=-1 errno=36 (File name too long)
> > # # child 9501 exited with 36 neither 99 nor 99
> > # not ok 48 Check success of execveat(8,
> > 'opt/kselftest/exec/x...yyyyyyyyyyyyyyyyyyyy', 0)...
> > # # Failed to open length 4094 filename, errno=36 (File name too long)
> > # # Invoke script via root_dfd and relative filename
> > # # child execveat() failed, rc=-1 errno=36 (File name too long)
> > # # child 9502 exited with 36 neither 127 nor 126
> > # not ok 49 Check success of execveat(8,
> > 'opt/kselftest/exec/x...yyyyyyyyyyyyyyyyyyyy', 0)...
>
> Could you check if replacing (in include/linux/fs.h)
>
> #define EMBEDDED_NAME_MAX 192 - sizeof(struct __filename_head)
>
> with
>
> #define EMBEDDED_NAME_MAX (192 - sizeof(struct __filename_head))
>
> is sufficient for fixing that reproducer?
FWIW, an unpleasant surprise is that LTP apparently doesn't test that
case anywhere - the effect of that braino is to lower the cutoff for
name length by 48 characters and there's not a single test in there
that would test that ;-/
chdir04 does check that name component is not too long, but that's
a different codepath - it's individual filesystem's ->lookup() rejecting
a component.
Oh, well - not hard to add (to the same chdir04, for example)...
