Improve backwards compatibility.

While writing the protocol document, we decided that a server should
always send a payload, even when an error occurred. Only it turned out
that this was not in line with what existing implementations did.

Since
- It seems to make more sense to not send invalid data to a client when
  we know it to be invalid, and
- this is what current implementations do at any rate,

change the spec to say that servers MUST NOT send payload on read error,
and change the implementation to match.

doc/proto.md

@@ -887,12 +887,9 @@ The following request types exist:
     of data, read from *offset* bytes into the file, unless an error
     condition has occurred.
 
-    If an error occurs, the server SHOULD set the appropriate error
-    code in the error field. The server MUST then either initiate
-    a hard disconnect, or send *length* bytes of data (these bytes MAY be
-    invalid, in which case they SHOULD be zero); this is true even if
-    the error is `EINVAL` for bad flags detected before even
-    attempting to read.
+    If an error occurs, the server SHOULD set the appropriate error code
+    in the error field. The server MAY then initiate a hard disconnect.
+    If it chooses not to, it MUST NOT send any payload for this request.
 
     If an error occurs while reading after the server has already sent
     out the reply header with an error field set to zero (i.e.,

nbd-server.c

@@ -1466,7 +1466,9 @@ static void handle_read(CLIENT* client, struct nbd_request* req) {
 	}
 	pthread_mutex_lock(&(client->lock));
 	writeit(client->net, &rep, sizeof rep);
-	writeit(client->net, buf, req->len);
+	if(!rep.error) {
+		writeit(client->net, buf, req->len);
+	}
 	pthread_mutex_unlock(&(client->lock));
 	free(buf);
 }