Support Software and Hardware Flow Control

PCCOM - Solution Provider for RS232/422/485 Serial Communication
Multi Port Serial Card
Serial - TCP/IP Communication
RS232/422/485 Converter and Repeater
Data Acquisition and Instrument Control
Real - Time Data Collection form any Serial Device
Introduction to UARTS

Flow Control

Flow control means the ability to slow down the flow of bytes in a wire. For serial ports this means the ability to stop and then restart the flow without any loss of bytes. Flow control is needed for modems and other hardware to allow a jump in instantaneous flow rates.

Example of Flow Control

For example, consider the case where you connect a 33.6k external modem via a short cable to your serial port. The modem sends and receives bytes over the phone line at 33.6k bits per second (bps). Assume it's not doing any data compression or error correction. You have set the serial port speed to 115,200 bits/sec (bps), and you are sending data from your computer to the phone line. Then the flow from the your computer to your modem over the short cable is at 115.2k bps. However the flow from your modem out the phone line is only 33.6k bps. Since a faster flow (115.2k) is going into your modem than is coming out of it, the modem is storing the excess flow (115.2k -33.6k = 81.6k bps) in one of its buffers. This buffer would soon overrun (run out of free storage space) unless the high 115.2k flow is stopped.

But now flow control comes to the rescue. When the modem's buffer is almost full, the modem sends a stop signal to the serial port. The serial port passes on the stop signal on to the device driver and the 115.2k bps flow is halted. Then the modem continues to send out data at 33.6k bps drawing on the data it previous accumulated in its buffer. Since nothing is coming into this buffer, the number of bytes in it starts to drop. When almost no bytes are left in the buffer, the modem sends a start signal to the serial port and the 115.2k flows from the computer to the modem resumes. In effect, flow control creates an average flow rate in the short cable (in this case 33.6k), which is significantly less than the "on" flow rate of 115.2k bps. This is "start-stop" flow control.

In the above simple example it was assumed that the modem did no data compression. This could happen when the modem is sending a file, which is already compressed and can't be compressed further. Now let's consider the opposite extreme where the modem is compressing the data with a high compression ratio. In such a case the modem might need an input flow rate of say 115.2k bps to provide an output (to the phone line) of 33.6k bps (compressed data). This compression ratio is 3.43 (115.2/33.6). In this case the modem is able to compress the 115.2 bps PC-to-modem flow and send the same data (in compressed form) out the phone line at 33.6bps. There's no need for flow control here so long as the compression ratio remains higher that 3.43. But the compression ratio varies from second to second and if it should drop below 3.43, flow control will be needed.

In the above example, the modem was an external modem. But the same situation exists (as of early 2003) for most internal modems. There is still a speed limit on the PC-to-modem speed even though this flow doesn't take place over an external cable. This makes the internal modems compatible with the external modems.

In the above example of flow control, the flow was from the computer to a modem. But there is also flow control, which is used for the opposite direction of flow: from a modem (or other device) to a computer. Each direction of flow involves 3 buffers: 1. in the modem 2. In the UART chip (called FIFOs) and 3. in main memory managed by the serial driver. Flow control protects all buffers (except the FIFOs) from overflowing.

Under Linux, the small UART FIFO buffers are not protected by flow control but instead rely on a fast response to the interrupts they issue. Some UART chips can be set to do hardware flow control to protect their FIFOs but Linux (as of early 2003) doesn't seem to support it. FIFO stands for "First In, First Out" which is the way it handles bytes in a queue. All the 3 buffers use the FIFO rule but only the one in the UART is named "FIFO". This is the essence of flow control but there are still some more details.

Symptoms of No Flow Control

Understanding flow-control theory can be of practical use. The symptom of no flow control is that chunks of data missing from files sent without the benefit of flow control. When overflow happens, often hundreds or even thousands of bytes get lost, and all in contiguous chunks.

Hardware vs. Software Flow Control

If feasible, it's best to use "hardware" flow control that uses two dedicated "modem control" wires to send the "stop" and "start" signals. Hardware flow control at the serial port works like this: The two pins, RTS (Request to send) and CTS (Clear to send) are used. When the computer is ready to receive date it asserts RTS by putting a positive voltage on the RTS pin (meaning "Request To Send to me"). When the computer is not able to receive any more bytes, it negates RTS by putting a negative voltage on the pin saying: "stop sending to me". The RTS pin is connected by the serial cable to another pin on the modem, printer, terminal, etc. This other pin's only function is to receive the signal.

For the case of a modem this "other" pin will be the modem's RTS pin. But for a printer, another PC, or a non-modem device, it's usually a CTS pin so a "crossover" or "null modem" cable is required. This cable connects the CTS pin at one end with the RTS pin at the other end (two wires since each end of the cable has a CTS pin). For a modem, a straight-thru cable is used.

For the opposite direction of flow a similar scheme is used. For a modem, the CTS pin is used to send the flow control signal to the CTS pin on the PC. For a non-modem, the RTS pin sends the signal. Thus modems and non-modems have the roles of their RTS and CTS pins interchanged. Some non-modems such as dumb terminals may use other pins for flow control such as the DTR pin instead of RTS.

Software flow control uses the main receive and transmit wires to send the start and stop signals. It uses the ASCII control characters DC1 (start) and DC3 (stop) for this purpose. They are just inserted into the regular stream of data. Software flow control is not only slower in reacting but also does not allow the sending of binary data unless special precautions are taken. Since binary data will likely contain DC1 and DC3, special means must be taken to distinguish between a DC3 that means a flow control stop and a DC3 that is part of the binary code.

Data Flow Path; Buffers

It's been mention that there are 3 buffers for each direction of flow (3 pairs altogether): 16-byte FIFO buffers (in the UART), a pair of larger buffers inside a device connected to the serial port (such as a modem), and a pair of buffers (say 8k) in main memory. When an application program sends bytes to the serial port they first get stashed in transmit serial port buffer in main memory. The other member of the pair consists of a receive buffer for the opposite direction of byte-flow. Here's an example diagram for the case of browsing the Internet with a browser. Transmit data flow is left to right while receive flow is right to left. There is a separate buffer for each direction of flow.

Application	8k-byte	16-byte	1k-byte
BROWSER	MEMORY	FIFO	MODEM	Telephone line
Program	Buffer	Buffer	Buffer

For the transmit case, the serial device driver takes out say 16 bytes from this transmit buffer (in main memory), one byte at a time and puts them into the 16-byte transmit buffer in the serial UART for transmission. Once in that transmit buffer, there is no way to stop them from being transmitted. They are then transmitted to the modem or (other device connected to the serial port) which also has a fair sized (say 1k) buffer. When the device driver (on orders from flow control) stops the flow of outgoing bytes from the computer, what it actually stops is the flow of outgoing bytes from the large transmit buffer in main memory. Even after this has happened and the flow to the modem has stopped, an application program may keep sending bytes to the 8k transmit buffer until it becomes fill. At the same time, the bytes stored in the FIFO and MODEM continues to be sent out until these buffers empty.

When the memory buffer gets fill, the application program can't send any more bytes to it (a "write" statement in a C program blocks) and the application program temporarily stops running and waits until some buffer space becomes available. Thus a flow control "stop" is ultimately able to stop the program that is sending the bytes. Even though this program stops, the computer does not necessarily stop computing since it may switch to running other processes while it's waiting at a flow control stop.

The above was a little oversimplified in three ways. First, some UARTs can do automatic hardware flow control, which can stop the transmission out of the FIFO buffers if needed (not yet supported by Linux). Second, while an application process is waiting to write to the transmit buffer, it could possibly perform other tasks. Third, the serial driver (located between the memory buffer and the FIFO) has it's own buffer (in main memory) used to process characters.

Complex Flow Control Example

For many situations, there is a transmit path involving several links, each with its own flow control. For example, I type at a text-terminal connected to a PC with a modem to access a BBS (Bulletin Board System). For this I use the application program "minicom" which deals with 2 serial ports: one connected to a modem and another connected to the text-terminal. What I type at the text terminal goes into the first serial port to minicom, then from minicom out the second serial port to the modem, and then onto the telephone line to the BBS. The text-terminal has a limit to the speed at which bytes can be displayed on its screen and issues a flow control "stop" from time to time to slow down the flow. What happens when such a "stop" is issued? Let's consider a case where the "stop" is long enough to get thru to the BBS and stop the program at the BBS which is sending out the bytes.

Let's trace out the flow of this "stop" (which may be "hardware" on some links and "software" on others). First, suppose I'm "capturing" a long file from the BBS which is being sent simultaneously to both my text-terminal and a to file on my hard disk. The bytes are coming in faster than the terminal can handle them so it sends a "stop" out its serial port to the first serial port on my PC. The device driver detects it and stops sending bytes from the 8 k serial buffer (in main memory) to the terminal. Now minicom still keeps sending out bytes for the terminal into this 8k buffer.

When this 8k transmit buffer (on the first serial port) is full, minicom must stop writing to it. Minicom stops and waits. But this also causes minicom to stop reading from the 8k receive buffer on the 2nd serial port connected to the modem. Flow from the modem continues until this 8k buffer too fills up and sends a different "stop" to the modem. Now the modem's buffer ceases to send to the serial port and also fills up. The modem (assuming error correction is enabled) sends a "stop signal" to the other modem at the BBS. This modem stops sending bytes out of its buffer and when its buffer gets fill, another stop signal is sent to the serial port of the BBS. At the BBS, the 8-k (or whatever) buffer fills up and the program at the BBS can't write to it anymore and thus temporarily halts.

Thus a stop signal from a text terminal has halted a program on a BBS computer. What a complex sequence of events! Note that the stop signal passed thru 4 serial ports, 2 modems, and one application program (minicom). Each serial port has 2 buffers (in one direction of flow): the 8k one and the hardware 16-byte one. The application program may have a buffer in its C code. This adds up to 11 different buffers the data is passing thru. Note that the small serial hardware buffers do not participate directly in flow control.

If the terminal speed limitation is the bottleneck in the flow from the BBS to the terminal, then its flow control "stop" is actually stopping the program that is sending from the BBS as explained above. But you may ask: How can a "stop" last so long that 11 buffers (some of them large) all get filled up? It can actually happen this way if all the buffers were near their upper limits when the terminal sent out the "stop".

But if you were to run a simulation on it you would discover that it's usually more complicated than this. At an instant of time some links are flowing and others are stopped (due to flow control). A "stop" from the terminal seldom propagates back to the BBS neatly as described above. It may take a few "stops" from the terminal to result in one "stop" at the BBS. To understand what is going on you really need to observe a simulation, which can be done for a simple case with coins on a table. Use only a few buffers and set the upper level for each buffer at only a few coins.

Does one really need to understand all this? Well, understanding this explained to me why capturing text from a BBS was loosing text. The situation was exactly the above example but modem-to-modem flow control was disabled. Chunks of captured text that were supposed to also get to my hard-disk never got there because of an overflow at my modem buffer due to flow control "stops" from the terminal. Even though the BBS had a flow path to the hard disk without bottlenecks, the overflow due to the terminal happened on this path and chunks of text were lost and never even made it to the hard disk. Note that the flow to the hard disk passed thru my modem and since the overflow happened there, bytes intended for the hard disk were lost.