Kernel hang with excessive resize-journal on replicated fs

[Chris Webb <chris@arachsys.com>]

The following ktest hangs in the kernel instead of producing an error from the
excessive journal resize:

  test_oversize_journal()
  {
      bcachefs format --errors=panic --replicas=2 /dev/sd[bc]
      mount -t bcachefs /dev/sdb:/dev/sdc /mnt

      # should fail, but actually hangs:
      ! bcachefs device resize-journal /dev/sdb $(blockdev --getsize64 /dev/sdb)

      umount /mnt
  }

(It would correctly return with an error on a single-device filesystem.)

The kernel is continuously looping on ret == -EAGAIN in
bch2_set_nr_journal_buckets. The disk reservation (unexpectedly) succeeds but
__bch2_set_nr_journal_buckets then keeps returning -EAGAIN because
bch2_bucket_alloc (correctly) fails.

Is the problem here that the disk reservation isn't confined to the specific
device we then want to allocate from? So although the space is available, it's
not available solely on the device where we want it?

Best wishes,

Chris.