btrfs: set BTRFS_ROOT_ORPHAN_CLEANUP during subvol create

In the Linux kernel, the following vulnerability has been resolved:

btrfs: set BTRFS_ROOT_ORPHAN_CLEANUP during subvol create

We have recently observed a number of subvolumes with broken dentries.
ls-ing the parent dir looks like:

drwxrwxrwt 1 root root 16 Jan 23 16:49 .
drwxr-xr-x 1 root root 24 Jan 23 16:48 ..
d????????? ? ? ? ? ? broken_subvol

and similarly stat-ing the file fails.

In this state, deleting the subvol fails with ENOENT, but attempting to
create a new file or subvol over it errors out with EEXIST and even
aborts the fs. Which leaves us a bit stuck.

dmesg contains a single notable error message reading:
"could not do orphan cleanup -2"

2 is ENOENT and the error comes from the failure handling path of
btrfs_orphan_cleanup(), with the stack leading back up to
btrfs_lookup().

btrfs_lookup
btrfs_lookup_dentry
btrfs_orphan_cleanup // prints that message and returns -ENOENT

After some detailed inspection of the internal state, it became clear
that:
- there are no orphan items for the subvol
- the subvol is otherwise healthy looking, it is not half-deleted or
anything, there is no drop progress, etc.
- the subvol was created a while ago and does the meaningful first
btrfs_orphan_cleanup() call that sets BTRFS_ROOT_ORPHAN_CLEANUP much
later.
- after btrfs_orphan_cleanup() fails, btrfs_lookup_dentry() returns -ENOENT,
which results in a negative dentry for the subvolume via
d_splice_alias(NULL, dentry), leading to the observed behavior. The
bug can be mitigated by dropping the dentry cache, at which point we
can successfully delete the subvolume if we want.

i.e.,
btrfs_lookup()
btrfs_lookup_dentry()
if (!sb_rdonly(inode->vfs_inode)->vfs_inode)
btrfs_orphan_cleanup(sub_root)
test_and_set_bit(BTRFS_ROOT_ORPHAN_CLEANUP)
btrfs_search_slot() // finds orphan item for inode N
...
prints "could not do orphan cleanup -2"
if (inode == ERR_PTR(-ENOENT))
inode = NULL;
return d_splice_alias(NULL, dentry) // NEGATIVE DENTRY for valid subvolume

btrfs_orphan_cleanup() does test_and_set_bit(BTRFS_ROOT_ORPHAN_CLEANUP)
on the root when it runs, so it cannot run more than once on a given
root, so something else must run concurrently. However, the obvious
routes to deleting an orphan when nlinks goes to 0 should not be able to
run without first doing a lookup into the subvolume, which should run
btrfs_orphan_cleanup() and set the bit.

The final important observation is that create_subvol() calls
d_instantiate_new() but does not set BTRFS_ROOT_ORPHAN_CLEANUP, so if
the dentry cache gets dropped, the next lookup into the subvolume will
make a real call into btrfs_orphan_cleanup() for the first time. This
opens up the possibility of concurrently deleting the inode/orphan items
but most typical evict() paths will be holding a reference on the parent
dentry (child dentry holds parent->d_lockref.count via dget in
d_alloc(), released in __dentry_kill()) and prevent the parent from
being removed from the dentry cache.

The one exception is delayed iputs. Ordered extent creation calls
igrab() on the inode. If the file is unlinked and closed while those
refs are held, iput() in __dentry_kill() decrements i_count but does
not trigger eviction (i_count > 0). The child dentry is freed and the
subvol dentry's d_lockref.count drops to 0, making it evictable while
the inode is still alive.

Since there are two races (the race between writeback and unlink and
the race between lookup and delayed iputs), and there are too many moving
parts, the following three diagrams show the complete picture.
(Only the second and third are races)

Phase 1:
Create Subvol in dentry cache without BTRFS_ROOT_ORPHAN_CLEANUP set

btrfs_mksubvol()
lookup_one_len()
__lookup_slow()
d_alloc_parallel()
__d_alloc() // d_lockref.count = 1
create_subvol(dentry)
// doesn't touch the bit..
d_instantiate_new(dentry, inode) // dentry in cache with d_lockref.c
---truncated---