Re: BUG_ON(nd->inode->i_op->follow_link);

[Al Viro <viro@ZenIV.linux.org.uk>]

On Thu, Mar 07, 2013 at 02:50:55PM -0800, Linus Torvalds wrote:

> and you get the oops, because when it follows that pseudo-symlink for
> fd3, it will hit that symlink we just looked up. And trigger the
> BUG_ON() you see.
> 
> Al, I think the BUG_ON() is simply bogus. Should we just remove it? Or
> is there some conceptual reason why we can't handle a symlink there?

BUG_ON() is bogus these days, but failing ->follow_link() is not.  The thing
is, what would you do with that sucker?  Suppose you have ->follow_link()
leaving you in a symlink.  What would you do with it?  Either you were
not going to follow it (in which case you wouldn't have called ->follow_link()
in the first place) *or* you need to know which directory it's in.  Which
we don't have in this case...

It really used to be an impossible case, until we'd introduced O_PATH.  And
IIRC, back when O_PATH had been developed, the plan was to fail attempts
to follow these suckers with -ELOOP; looks like we forgot to do that.  Mea
culpa...

And no, just removing BUG_ON() is not enough - callers of follow_link() will
do rather weird stuff (in nested symlink case it'll fail with ENOTDIR, in
path_lookupat() with LOOKUP_FOLLOW it'll give you a symlink, in case of
path_openat() with LOOKUP_FOLLOW fail with ELOOP).

See the patch I'd posted upthread.  Let's make it fail consistently...