fix proc_fill_cache() in case of d_alloc_parallel() failure

If d_alloc_parallel() returns ERR_PTR(...), we don't want to dput()
that.  Small reorganization allows to have all error-in-lookup
cases rejoin the main codepath after dput(child), avoiding the
entire problem.
Spotted-by: Tetsuo Handa <penguin-kernel@i-love.sakura.ne.jp>
Fixes: 0168b9e3 "procfs: switch instantiate_t to d_splice_alias()"
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>

fs/proc/base.c

@@ -1891,19 +1891,19 @@ bool proc_fill_cache(struct file *file, struct dir_context *ctx,
 			struct dentry *res;
 			res = instantiate(child, task, ptr);
 			d_lookup_done(child);
-			if (IS_ERR(res))
-				goto end_instantiate;
 			if (unlikely(res)) {
 				dput(child);
 				child = res;
+				if (IS_ERR(child))
+					goto end_instantiate;
 			}
 		}
 	}
 	inode = d_inode(child);
 	ino = inode->i_ino;
 	type = inode->i_mode >> 12;
-end_instantiate:
 	dput(child);
+end_instantiate:
 	return dir_emit(ctx, name, len, ino, type);
 }