Re: [RFC][PATCH] do d_instantiate/unlock_new_inode combinations safely

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

On Fri, May 11, 2018 at 11:32:08AM +1000, Dave Chinner wrote:

> i.e. we already have code in xfs_setup_inode() that sets the xfs
> inode ILOCK rwsem dir/non-dir lockdep class before the new inode is
> unlocked - we could just do the i_rwsem lockdep setup there, too.

... which would suffice -

        if (S_ISDIR(inode->i_mode)) {
                struct file_system_type *type = inode->i_sb->s_type;

                /* Set new key only if filesystem hasn't already changed it */
                if (lockdep_match_class(&inode->i_rwsem, &type->i_mutex_key)) {

in lockdep_annotate_inode_mutex_key() would make sure that ->i_rwsem will be
left alone by unlock_new_inode().

> Then, if we were to factor unlock_new_inode() as Andreas suggested,
> we could call __unlock_new_inode() from xfs_finish_inode_setup().

No need - if you set the class in xfs_setup_inode(), you are fine.

Said that, hash insertion is also potentially delicate - another ext2/nfsd
race from the same pile back in 2008 had been
	* ext2_new_inode() chooses inumber
	* open-by-fhandle guesses the inumber and hits ext2_iget(), which
inserts a locked in-core inode into icache and proceeds to block reading
it from disk.
	* ext2_new_inode() inserts *its* in-core inode into icache (with
the same inumber) and sets the things up, both in-core and on disk
	* open-by-fhandle is back and sees a good live on-disk inode.
It finishes setting the in-core one up and we'd got *TWO* in-core inodes
with the same inumber, both hashed, both with dentries, both used by
syscalls to do IO.  Good times all around - fs corruption is fun.

That was fixed by using insert_inode_locked() in ext2_new_inode(), and doing
that before the on-disk inode would start looking good.  If it came during
ext2_iget(), it would've found an in-core inode with that inumber (locked,
doomed to be rejected), waited for it to come unlocked, see it unhashed
(since ext2_iget() said it was no good) and inserted its in-core inode
into hash (after having rechecked that nobody had an in-core inode with
the same inumber in there, that is).

I'm not familiar enough with XFS icache replacment to tell if anything
of that sort is a problem there; might be a non-issue for any number
of reasons.