bcachefs: Fix inode early destruction path

discard_new_inode() is the wrong interface to use when we need to free
an inode that was never inserted into the inode hash table; we can
bypass the whole iput() -> evict() path and replace it with
__destroy_inode(); kmem_cache_free() - this fixes a WARN_ON() about
I_NEW.

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>

Author: Kent Overstreet

fs/bcachefs/fs.c

@@ -188,7 +188,8 @@ static struct bch_inode_info *bch2_inode_insert(struct bch_fs *c, struct bch_ino
 	BUG_ON(!old);
 
 	if (unlikely(old != inode)) {
-		discard_new_inode(&inode->v);
+		__destroy_inode(&inode->v);
+		kmem_cache_free(bch2_inode_cache, inode);
 		inode = old;
 	} else {
 		mutex_lock(&c->vfs_inodes_lock);
@@ -225,8 +226,10 @@ static struct bch_inode_info *bch2_new_inode(struct btree_trans *trans)
 
 	if (unlikely(!inode)) {
 		int ret = drop_locks_do(trans, (inode = to_bch_ei(new_inode(c->vfs_sb))) ? 0 : -ENOMEM);
-		if (ret && inode)
-			discard_new_inode(&inode->v);
+		if (ret && inode) {
+			__destroy_inode(&inode->v);
+			kmem_cache_free(bch2_inode_cache, inode);
+		}
 		if (ret)
 			return ERR_PTR(ret);
 	}