By not taking references to parent directories when walking, the dentry
can become negative under us, if the target file is moved then the parent
quickly deleted.  This is problematic because it means that we lose access
to any landlock rules attached to that parent, and thus we won't be able
to grant the correct access even when we're allowed to pretend the move
hasn't happened yet.

This commit tests a slightly more complicated scenario, where after moving
the file's parent directory away, the next two intermediate directories
are quickly removed.  This demonstrates that in this situation the only
choice for Landlock is to restart the walk.  Without doing that, even if
we were to re-check d_parent, if we were "cut off" when we've already
walked away from the original leaf, we would still not be able to recover.

As an illustration:

mkdir /d1 /d1/d2 /b
create landlock rule on /d1
create landlock rule on /b
touch /d1/d2/file

thread 1                              thread 2
                                      cd /d1/d2
                                      cat file
                                        landlock walks to /d1/d2 without ref,
                                        checked rule on /d1/d2 (nothing),
                                        about to walk up to /d1
mv /d1/d2/file /b
rmdir /d1/d2 /d1
(/d1/d2 and /d1 both becomes negative)
                                        notices stuff changed
                                        at this point, we're looking at /d1/d2
                                        and trying to walk to its parent.
                                        however, both /d1/d2 and /d1 are negative
                                        now.
                                        Our only choice is to restart the walk
                                        altogether from the original file's dentry,
                                        which will now have d_parent /b.

The test is probablistic as it tests for a race condition, but I found
that in my environment, with the previous patch, it pretty much always
reliably fails within 10 seconds.  I've set the timeout to 30 seconds, and
the test will pass if no permission errors (or other errors) detected.  In
those 30 seconds it will keep recreating the above directory structure
(except with a lot more sibling directories so it can run for some time
before it "exhausts" all the directories and has to recreate the whole
thing).

Signed-off-by: Tingmao Wang <m@maowtm.org>

This fixes the issue mentioned in the previous patch, by essentially
having two "modes" for the pathwalk code - in the pathwalk_ref == false
case we don't take references and just inspect `d_parent` (unless we have
to `follow_up`).  In the pathwalk_ref == true case, this is the same as
before.

When we detect any renames during a pathwalk_ref == false walk, we restart
with pathwalk_ref == true, re-initializing the layer masks.  I'm not sure
if this is completely correct in regards to is_dom_check - but seems to
work for now.  I can revisit this later.

Signed-off-by: Tingmao Wang <m@maowtm.org>