Re: [PATCH v3 1/3] namei: implement O_BENEATH-style AT_* flags

On 2018-10-13, Al Viro [off-list ref] wrote:

First of all, dirfd_path_init() part should be in a separate commit.  And I'm
really not happy with the logics in there.  dirfd_path_init() itself is
kinda-sorta reasonable.

Sure, I can do that.

It is equivalent to setting the starting point for
relative pathnames + setting ->root for LOOKUP_BENEATH, right?

Right.

But the part in path_init() is too bloody convoluted for its own good.  Let me
try to translate:

quoted

+	if (unlikely(flags & LOOKUP_XDEV)) {
+		error = dirfd_path_init(nd);
+		if (unlikely(error))
+			return ERR_PTR(error);
+	}

* if LOOKUP_XDEV is set, set the starting point as if it was a relative
  pathname.  If LOOKUP_BENEATH was set as well, set ->root to the same
  point.

Right. This is for two reasons (though if you disagree with these
semantics we can change this as well):

1. It's not clear to me whether openat(somefd->"/", "/tmp", O_XDEV)
   should return an -EXDEV or completely ignore the starting point. Same
   argument with AT_FDCWD. I opted to make it so that the starting point
   has to be on the same mountpoint, but I totally understand if you
   feel this is insane -- and I'd be happy to change it. The real
   problem comes from (2).

2. AT_THIS_ROOT chroot-scope absolute paths, and so in the second patch
   LOOKUP_CHROOT also triggers this codepath. The main argument for this
   semantic is somewhat elaborated in the cover letter -- but
   the short version is because AT_THIS_ROOT has to chroot-scope
   absolute symlinks it would be somewhat strange if it didn't scope
   absolute paths you give it -- otherwise it could either be a footgun
   or would require always returning -EXDEV here.

   Though, as above, if you feel that the current semantics (absolute
   paths override whatever dirfd you give), then -EXDEV is the
   alternative I would pitch.

* if it's an absolute pathname, 

quoted

 	if (*s == '/') {

... and we hadn't come here with LOOKUP_XDEV + LOOKUP_BENEATH, set ->root.

quoted

+		if (likely(!nd->root.mnt))
+			set_root(nd);

* if it's an absolute pathname, set the starting point to ->root.  Note that
if we came here with LOOKUP_XDEV, we'll discard the starting point we'd
calculated.

We wouldn't discard it -- nd_jump_root() will check whether a mount
crossing was implied here (otherwise an absolute symlink could cause you
to cross a mountpoint).

But as above, if you'd prefer that absolute paths disable all dirfd
handling (as is the case now), I can remove this semantic.

quoted

+		error = nd_jump_root(nd);
+		if (unlikely(error))
+			s = ERR_PTR(error);
 		return s;
 	}
+	if (likely(!nd->path.mnt)) {

* if we didn't have LOOKUP_XDEV, set the starting point as if it was a relative
pathname (which it is) and, if LOOKUP_BENEATH is also there, set ->root there
as well.

quoted

+		error = dirfd_path_init(nd);
+		if (unlikely(error))
+			return ERR_PTR(error);
+	}
+	return s;
 }

Pardon me, but... huh?  The reason for your two calls of dirfd_path_init() is,
AFAICS, the combination of absolute pathname with both LOOKUP_XDEV and
LOOKUP_BENEATH at the same time.  That combination is treated as if the pathname
had been relative.  Note that LOOKUP_BENEATH alone is ignored for absolute ones
(and with a good reason - it's a no-op on path_init() level in that case).

What the hell?  It complicates your code and doesn't seem to provide any benefits
whatsoever

The reasoning for this is because of how AT_THIS_ROOT uses both of these
codepaths (it causes dirfd_path_init() to be called before the absolute
check, and also causes ->root to be set).

I wrote the features in parallel and then split out the code for
AT_THIS_ROOT so it could be discussed separately (and so removing it if
it was rejected would be simpler). But unfortunately this does result in
the dirfd_path_init() code looking completely superfluous without seeing
the second patch.

-- you could bloody well have passed the relative pathname to start with.

(I think you mean always doing dirfd_path_init() first here?)

Right, but I didn't want to discard nd->path unnecessarily -- if we do
all of the code to grab AT_FDCWD and then it is completely unused (not
even in the AT_XDEV sense of "unused") it seems like a waste.

Did I misunderstand your suggestion? Were you referring to userspace
just being able to "[pass] the relative pathname to start with"?

IDGI...  Without that kludge it becomes simply "do as we currently do for absolute
pathnames, call dirfd_path_init() for relative ones".  And I would argue that
taking LOOKUP_BENEATH handling out of dirfd_path_init() into path_init() (relative)
case would be a good idea.

Right, I could definitely do that -- though for AT_THIS_ROOT we'd
duplicate the ->root setting in both places.

As it is, the logics is very hard to follow.

Sorry about that. Would you prefer if the two patches (AT_BENEATH family
and AT_THIS_ROOT) were sent as a single patch -- with the
dirfd_path_init() code split out? Or that the second patch do all of the
structural changes to refactor dirfd_path_init() usage?

-- 
Aleksa Sarai
Senior Software Engineer (Containers)
SUSE Linux GmbH
<https://www.cyphar.com/>