Currently we go GFP_ATOMIC allocation for qgroup relation add, this
includes the following 3 call sites:

- btrfs_read_qgroup_config()
  This is not really needed, as at that time we're still in single
  thread mode, and no spin lock is held.

- btrfs_add_qgroup_relation()
  This one is holding a spinlock, but we're ensured to add at most one
  relation, thus we can easily do a preallocation and use the
  preallocated memory to avoid GFP_ATOMIC.

- btrfs_qgroup_inherit()
  This is a little more tricky, as we may have as many relationships as
  inherit::num_qgroups.
  Thus we have to properly allocate an array then preallocate all the
  memory.

This patch would remove the GFP_ATOMIC allocation for above involved
call sites, by doing preallocation before holding the spinlock, and let
__add_relation_rb() to handle the freeing of the structure.
Signed-off-by: Qu Wenruo <wqu@suse.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Signed-off-by: David Sterba <dsterba@suse.com>