Re: [PATCH 01/14] VFS: Add additional RESOLVE_* flags [ver #18]

[Linus Torvalds <torvalds@linux-foundation.org>]

On Thu, Mar 12, 2020 at 9:56 AM David Howells <dhowells@redhat.com> wrote:
>
> Any thoughts on which set of flags should override the other?

Do we need to care? I don't think we actually have conflicts, because
the semantics aren't the same, and they are about independent issues.

>                  If we're making
> RESOLVE_* flags the new definitive interface, then I feel they should probably
> override the AT_* flags where there's a conflict, ie. RESOLVE_NO_SYMLINKS
> should override AT_SYMLINK_FOLLOW for example.

That's just for a linkat2() system call? I think the natural semantic
is the one that falls out directly: RESOLVE_NO_SYMLINKS will cause it
to fail with -ELOOP if it is a symlink.

NOTE! This isn't really a "conflict". It's actually two different and
independent things:

 - without AT_SYMLINK_FOLLOW, a linkat() simply won't even try to
follow the symlink, and will link to the symlink itself instead.

 - RESOLVE_NO_SYMLINKS says "never follow symlinks".

Note how one does *NOT* override the other, quite the reverse. They
are about different things. One is about the _behavior_ when the last
component is a symlink, and the other is about permission to follow
any symlinks.

So all combinations make sense:

 - no AT_SYMLINK_FOLLOW, no RESOLVE_NO_SYMLINKS: just link to the
target, whether it's a symlink or not

   This is obviously our historical link() behavior.

 - no AT_SYMLINK_FOLLOW, yes RESOLVE_NO_SYMLINKS: just link to the
target, whether it's a symlink or not, but if there's a symlink in the
middle, return -ELOOP

   Note how this case doesn't follow the last one, so
RESOLVE_NO_SYMLINKS isn't an issue for the last component, but _is_ an
issue for the components in the middle.

 - AT_SYMLINK_FOLLOW, no RESOLVE_NO_SYMLINKS: just link to the target,
following the symlink if it exists

   This is obviously the historical AT_SYMLINK_FOLLOW behavior

 - AT_SYMLINK_FOLLOW | RESOLVE_NO_SYMLINKS: just link to the target,
return -ELOOP if it's a symlink (of if there's a symlink on the way).

   This is the natural behavior for "refuse to follow any symlinks anywhere".

note how they are all completely sane versions, and in no case does
one flag really override the other.

If anything, we actually miss a third flag: the "don't allow linking
to a final symlink at all" (but allow intermediate symlinks). We've
never had that behavior, although I think POSIX makes that case
undefined (ie you're not guaranteed to be able to link to a symlink in
the first place in POSIX).

I guess that third case could be emulated with open(O_PATH) + fstat to
check it's not a symlink + linkat(fd,AT_EMPTY_PATH) if it turns
somebody would want something like that (and we decided that
AT_EMPTY_PATH is ok for linkat()).

I doubt anybody cares.

                 Linus