24.4 TCP Out-of-Band Data Recap

All our examples using out-of-band data so far have been trivial. Unfortunately, out-of-band data gets messy when we consider the timing problems that may arise. The first point to consider is that the concept of out-of-band data really conveys three different pieces of information to the receiver:

The fact that the sender went into urgent mode. The receiving process can be notified of this with either the SIGURG signal or with select. This notification is transmitted immediately after the sender sends the out-of-band byte, because we saw in Figure 24.11 that TCP sends the notification even if it is stopped by flow control from sending any data to the receiver. This notification might cause the receiver to go into some special mode of processing for any subsequent data that it receives.
The position of the out-of-band byte, that is, where it was sent with regard to the rest of data from the sender: the out-of-band mark.
The actual value of the out-of-band byte. Since TCP is a byte stream protocol that does not interpret the data sent by the application, this can be any 8-bit value.

With TCP's urgent mode, we can think of the URG flag as being the notification, the urgent pointer as being the mark, and the byte of data as itself.

The problems with this concept of out-of-band data are that: (i) there is only one TCP urgent pointer per connection; (ii) there is only one out-of-band mark per connection; and (iii) there is only a single one-byte out-of-band buffer per connection (which is an issue only if the data is not being read inline). We saw with Figure 24.12 that an arriving mark overwrites any previous mark that the process has not yet encountered. If the data is being read inline, previous out-of-band bytes are not lost when new out-of-band data arrives, but the mark is lost.

One common use of out-of-band data is with rlogin, when the client interrupts the program that it is running on the server (pp. 393–394 of TCPv1). The server needs to tell the client to discard all queued output because up to one window's worth of output may be queued to send from the server to the client. The server sends a special byte to the client, telling it to flush all output, and this byte is sent as out-of-band data. When the client receives the SIGURG signal, it just reads from the socket until it encounters the mark, discarding everything up through the mark. (Pages 398–401 of TCPv1 contain an example of this use of out-of-band data, along with the corresponding tcpdump output.) In this scenario, if the server sent multiple out-of-band bytes in quick succession, it wouldn't affect the client, as the client just reads up through the final mark, discarding all the data.

In summary, the usefulness of out-of-band data depends on why it is being used by the application. If the purpose is to tell the peer to discard the normal data up through the mark, then losing an intermediate out-of-band byte and its corresponding mark is of no consequence. But if it is important that no out-of-band bytes be lost, then the data must be received inline. Furthermore, the data bytes that are sent as out-of-band data should be differentiated from normal data since intermediate marks can be overwritten when a new mark is received, effectively mixing out-of-band bytes with the normal data. telnet, for example, sends its own commands in the normal stream of data between the client and server, prefixing its commands with a byte of 255. (To send this value as data then requires two successive bytes of 255.) This lets it differentiate its commands from normal user data, but requires that the client and server process each byte of data looking for commands.

[ Team LiB ]