nbd: Permit simple error to NBD_CMD_BLOCK_STATUS
commitebd82cd872726549d0a55d329d22c731e2e660ff
authorEric Blake <eblake@redhat.com>
Mon, 25 Mar 2019 19:01:04 +0000 (25 14:01 -0500)
committerEric Blake <eblake@redhat.com>
Sat, 30 Mar 2019 15:06:08 +0000 (30 10:06 -0500)
treea34633e4afbca0ee6785a1cbdeff51473c1374c4
parentb29f3a3d2a5fab40dbb4a65fa2f91821ebffae51
nbd: Permit simple error to NBD_CMD_BLOCK_STATUS

The NBD spec is clear that when structured replies are active, a
simple error reply is acceptable to any command except for
NBD_CMD_READ.  However, we were mistakenly requiring structured errors
for NBD_CMD_BLOCK_STATUS, and hanging up on a server that gave a
simple error (since qemu does not behave as such a server, we didn't
notice the problem until now).  Broken since its introduction in
commit 78a33ab5 (v2.12).

Noticed while debugging a separate failure reported by nbdkit while
working out its initial implementation of BLOCK_STATUS, although it
turns out that nbdkit also chose to send structured error replies for
BLOCK_STATUS, so I had to manually provoke the situation by hacking
qemu's server to send a simple error reply:

| diff --git i/nbd/server.c w/nbd/server.c
| index fd013a2817a..833288d7c45 100644
| 00--- i/nbd/server.c
| +++ w/nbd/server.c
| @@ -2269,6 +2269,8 @@ static coroutine_fn int nbd_handle_request(NBDClient *client,
|                                        "discard failed", errp);
|
|      case NBD_CMD_BLOCK_STATUS:
| +        return nbd_co_send_simple_reply(client, request->handle, ENOMEM,
| +                                        NULL, 0, errp);
|          if (!request->len) {
|              return nbd_send_generic_reply(client, request->handle, -EINVAL,
|                                            "need non-zero length", errp);
|

Signed-off-by: Eric Blake <eblake@redhat.com>
Acked-by: Richard W.M. Jones <rjones@redhat.com>
Message-Id: <20190325190104.30213-3-eblake@redhat.com>
Reviewed-by: Vladimir Sementsov-Ogievskiy <vsementsov@virtuozzo.com>
block/nbd-client.c