// Patrick Louis

Terminals

Terminals

(Transcript of the podcast)

Intro

We spend so much time typing at a terminal and yet the underlying mechanisms and history behind it are often overlooked.

The TTY is an integral part of Unix and we take most of its behavior for granted even though it has a huge history baggage that it carries to this day.

For instance pressing control-C or control-Z to stop or put in the background a process, or using control-A to go to the beginning of the line.

You might think that the control-A comes from the EMACS keybinds but it doesn’t, it’s the opposite, the EMACS keybinds are inspired by the TTY.

In this episode we’re going to dive in the world of terminals.

A big, rough, and unhoned overview of this part of Unix.

The terminal itself

Definition

So really, what’s a terminal?

This is a rather evolving or evolved definition, originally it was a hardware device with a single purpose, now it more or less isn’t.

Because of the changing definition many will disagree about it, even fight over the semantics and epistemology.

The name itself is about: teleprinters/teletypewriters (TeleTYpewriter, TTY).

Originally, they meant a piece of equipment through which you could interact with a computer: in the early days of unix, that meant a teleprinter-style device resembling a typewriter, sometimes called a teletypewriter, or “tty” in shorthand. The name “terminal” came from the electronic point of view, then end of a serial connection, and the name “console” from the furniture point of view. Very early in unix history, electronic keyboards and displays became the norm for terminals.

Generically, it’s an “input/output device which function is to display and input data”, but that’s overly generic.

Thus many things are regarded as terminals, they come in many forms:

It could be a “serial devices connected by a serial port such as printers/teleprinters teletypewriters, or modems supporting remote terminals via dial-up access, and directly-connected local terminals.

Or it could be a “display adapter and keyboard hardware directly incorporated into the system unit taken together to form a local “console”, which may be presented to users and to programs as a single CRT terminal or as multiple virtual terminals”.

Or it could be a “software terminal emulators, such as the xterm, Konsole, GNOME Terminal, and Terminal programs, and network servers such as the rlogin daemon and the SSH daemon, which make use of pseudoterminals”.

For the moment we haven’t talked about graphics and text but a lot of the time when we think of terminals we picture character-cell terminals or text-terminals with maybe a keyboard attached to it.

Those are devices connected to a host computer by a serial cable which transfers the textual information, in ASCII or EBCDIC form. The text on those terminals is displayed character by character from a pre-selected array of characters that the terminal has for each.

But that doesn’t mean it’s only normal text that appears on the terminal it can be represented in any way, a good example is the braille terminal that displays the braille representation of asciis.

This kind of device can have a typewriter or a video display, in this case we would call it a video display terminal or video display unit, VDU, or “glass TTY”.

But there are graphical terminals too that use vector mode or raster mode and can display images.

The cable used to hook up the text terminal to the host machine is usually a RS-232-C cable that talks the RS-232, the defacto protocol of text terminals, which is plainly said just 1 start bit, 8 data bits (the first 7 usually for ascii and the last one unspecified), no parity, and 1 stop bit.

For graphic terminals the protocols could be Tektronix for vector graphics, or ReGIS, or anything else as there are many and graphic terminals weren’t used that much.

This all make it seem like a terminal is just a thin client.

The device where you swipe/scan your credit card at a restaurant is called a terminal too.

The word terminal in its traditional meaning is a device through which you interact with a computer.

In the Unix world this all gets mixed up, a tty is a special file, a character device with additional commands beyond read and write, terminal may be synonymous with tty because the teletype were the first to be terminal ends to computers, as we’ll see.

A lot of the vocabulary is used interchangeably.

Dumb & Smart

We now have a vague idea of what a terminal is, an input/output bidirectional device that doesn’t process anything, a thin client.

However that’s not even close to what a terminal is.

There is a distinction between what some consider dumb terminals and intelligent terminals.

In fact there are many different definitions that may contradict each others.

Some say that “dumb terminals” are the ones that do no big processing and only interpret a limited number of “control codes” such as carriage return and line feed but don’t have the ability to process special escape sequence like clearing the screen. The “intelligent” terminals in that case are the ones that process the special escape sequences.

Remember here the keyword “control codes”, we’ll come back to those later on.

This also means that everything that travels the I/O streams for “dumb terminals” is dead simple and unstructured while that for intelligent terminals there could be format specifications, for instance TCP/IP is a good example of such protocol.

But then again the RS-232 or similar protocols would only fit the dumb terminal definition.

Another definition says that “dumb terminals” don’t do their own processing while “intelligent terminals” have microprocessor built in them to process inputs before sending them.

Others say that “smart terminals” are “fat clients”, that they don’t depend on a host, unlike thin clients. They do all the processing themselves, while “dumb terminals” are “thin clients”.

In that case Chromebooks and Citrix are considered dumb terminals.

But they aren’t really, this was just an extreme example to show the absurdity. Dumb terminal are thin clients but thin clients are not dumb terminals.

Others say that the terms “dumb” and “intelligent” terminals are marketing terms to sell personal computers and that in fact all terminals are dumb from the straight go, that it goes with the definition of what a terminal is.

I quite like that last approach.

Some bits of history

Those are all fine definitions but they’re not really precise, they’re just examples of what could represent the concept of a terminal.

The essence of terminals is more in their history than anything else, and it’s a changing history.

Let’s start with a part that might be surprising, but when you think about it is not so surprising, which is that the teletype evolved independently from the computer in what you could call its infancy.

The teletype derives from a device that was invented in 1869, the ticker tape or stock ticker which was an electro-mechnical machine, basically a typewriter connected through wires to the telegraph network that printed on ticker tape printer. It’s purpose was to distribute stock prices over long distances in realtime over telegraph lines. This evolved into ASCII-based teletype which were connected in a large network called a Telex, used to transfer commercial telegrams. All of this without ever being connected to a computer.

On the other side computers were gigantic monsters that executed programs through batch processing, that is you entered the program through a punch card, for example, and then you couldn’t do any manual intervention until it was finished, that’s it end of the line. Soon enough realtime interaction was possible and teletypes were used as input and output devices because they were already everywhere on the market.

Printing terminals were truly limited by the speed at which paper could be printed, and for interactive use the paper record was necessary so that you could see what you were typing.

The brand new shiny video computer displays were introduced at this point, also nicknamed “Glass TTYs” or “Visual Display Units”-VDU, which had a significant advantage and improvement over typewriters.

There were a lot of vendors but the two most popular and well known of those were the (DEC) Digital Equipment Corporation VT100 and the IBM3270.

IBM 3270 because it was used with the IBM mainframe, not with Unix, the VT100 was, and is still iconic today.

But why is the VT100 iconic, from all the vendors?

Digital Equipment Corporation which was founded in 1957 is one of the classic company that made computers a bit more practical starting with their PDP-1 that included a CRT in its operator console, then continued with today’s well known VT series of glass TTY.

At the time there were so many video terminal that the market was labelled “the tower of Babel”, each manufacturer had their own proprietary set of terminal functions, codes, escape sequences, control characters, etc..

Each of them had different characteristics used to control the physical appearance on the screen, some of the controls used when connected online or offline, as in to a computer or not. Because remember that teletype were there before being connected to computers.

It was a conventions used when the tele-typewriters were connected to printers. Which was essential because both end needed to know what characteristics the connection had. Both ends need to communicate the exact same way, the default settings didn’t cut it off.

More on that later… Again remember that keyword “control characters”, I said I’ll come back to it.

So programmers grew frustrated, there needed to be regrouping and identifying of terminal types and related control codes to use.

A compatibility layer was needed for the panoply of different teletype models.

Eventually it led to the development of the ANSI standard for device control, the X3.64 standard in 1977.

The VT100 from DEC, with its 8-bit Intel 8080 chip to interpret escape sequences, was one of the first, but not the first, to implement it. However it got really popular.

From this point on all terminals had backward compatibility with the VT100, and more precisely the X3.64 standard, so programmers favored the standard over added product features.

The output appearance escape sequences were now standardized

A last note on this topic.

Green monochrome screens, from IBM PCs for example, were not text terminals, per say, to be pedantic. Those screens didn’t contain any character generation, the image displayed is processed by the CPU and sent directly to the device to be displayed.

More on that in a previous episode about green on black.

Terminal Emulators - AKA Today

It’s good to know that there were terminal emulators for a long time. Before TCP/IP that’s what was used to remotely communicate with terminals of different types, it was the emulation layer.

Modern personal computers have a built-in keyboard and screen and so there’s not much need to own a physical terminal anymore.

Instead we have terminal emulators and virtual consoles providing multiple text-terminals on a single machine, which are provided by most of today’s operating systems, especially on Unix-like systems.

Terminal emulators are pieces of softwares which emulate and behave like real text terminal to some degree, some with better emulation than others.

We’ll dive into the architecture later on.

Those emulation softwares are compatible with all the most common terminal escape sequences and control characters, they use the ANSI standard that we talked about earlier X3.64.

It’s pretty advantageous and useful to be able to support anything on a single machine.

We’ll talk about why is everything supported later one, using termcap/terminfo.

Even though some argue that this emulation sometimes has many drabacks such as the constructed architecture to make it work and physical things such as the keyboard layouts not being the same, and the legacy aspect it has.

…my experience with “VT100” emulators suggests that many, maybe even most, lack some fundamental capability [such as] the ability to respond to an auto-identification query, the ability to perform certain screen or cursor operations, the ability to transmit appropriate “editing keypad” escape sequences, etc. Beware.

Most of today’s terminal emulators emulate the VT100 or at least are compatible with it. For example xterm emulates VT220, urxvt the VT102, the termite terminal uses the VTE library which is also used in the gnome-terminal, a sort of easy wrapper that emulates the VT220.

How close a given terminal or terminal-emulator program comes to precisely emulating a real VT100 can be measured by using the program vttest.

vttest - tool for testing VT100 compatibility of terminals https://en.wikipedia.org/wiki/Vttest

Control characters

A First Approach

I kept on repeating that we’ll explain what those control codes and escape sequences are about, so let’s do that now.

The terminal protocol only lets 8 bits, which 7 are ASCII, go over the wire and so there was a need for a way to send special commands to the terminal through ASCII, a way to control the terminal or the computer that the terminal is connected to.

The rather simplistic way implemented was to add a key that if held down, that key being the control key, and pressing another character it would clear up, grounds, zeros the voltage, of the 7th and 6th bits, so that the 8bits, ASCII character code is modified.

When you remove the 6 and 7th bit you don’t get actual printing characters but you get special ones that have a special meaning.

That’s simple enough, right?

This is apparent when you look at a physical typewriter such as the Teletype Model 33 ASR, which was originally used for telegraphy. (https://www.flickr.com/photos/osr/8699656409/lightbox) It was written on the keys what those control codes would do, on the image I’ve linked you can clearly see the CTRL-G - BELL and CTRL-I - TAB.

One interesting consequence is that adding SHIFT to the Control code doesn’t give a new bit sequence. That is because holding shift has the effect of setting to true the 6th bit which is then nullified by holding control.

As a result multiple key combinations can be equivalent, for example:

Ctrl+i, Ctrl+Shift+i, Tab
Ctrl+j, Ctrl+Shift+j, Enter
Ctrl+[, Ctrl+Shift+[, Escape
Ctrl+d, Ctrl+Shift+d, EOF
Ctrl+g, Ctrl+Shift+g, BEL

Those could be advantageous when the usual key, for instance escape, is not placed in an ergonomic manner, you can use the equivalent key combination instead.

In the early days, the control character implementation was hardware based, the changing of the ASCII code was done before being sent to the machine, but these days every key are distinguished and the interpretation is left to the software.

Learning To Walk

Before continuing with the other kinds of special sequences let’s discuss some methods, docs, and investigation tools that can be used to verify what happens when certain keys are pressed and interpreted by the terminal.

The first most important one is the man page section 7 for ascii.

It contains a table of ascii characters where it’s easy to see how control characters are triggered, they’re adjacent to their non 6&7th bit nullified version. For example the “start of heading” control character is next to the “A” character, pressing ctrl-a is equivalent to entering the “start of heading”.

Another thing you can do with that table is checkout the list of control characters.

Now a great way to actually analyze what is happening when entering characters on a terminal is get inside a mode where only the terminal interprets them via its reading library, via its readline function, or any of its low level mechanism.

You can run cat -v or tee or script and enter the key combination you want, it’ll greet you back with a visual representation of those non-printable characters we call the “caret notation”.

For instance it shows ^[ , caret & bracket, for the escape character.

This is an easy way to make sure everything you thought was happening is truly happening.

Flow Control

There are two control characters that are present for the control of transmission over the RS-232 protocol.

They are usually called “flow control characters” as they have the ability to tell the terminal to stop the transmission and to resume it.

The XOFF and XON controls, which are otherwise referred in the ASCII table as device control 1 and 3 respectively.

This is a useful feature if the computer at the end of the terminal is sending characters too fast and the terminal is not able to keep up with the printing.

As we said and will come back to later, every terminal has their own settings and features and the computer and the terminals had to agree on what base settings they are gonna communicate.

One of those settings is the baud rate, the rate of character transfer, the output speed of the terminal in bits per second.

You can imagine that the flow control is useful if a terminal is mis-configured and its baudrate is lower than what the computer presumes.

We’ll come back to the settings later.

Meta|Special|Super|Magic

What’s up with the other special keys on the keyboard?

Generically, those keys are called “modifiers” as they change the behavior of other keys when you hold it.

We already saw one subset of those modifiers when discussing control, another subset is about “bucky bit”s.

A bucky bit is a bit that gets set when you press a modifier at the same time.

You might already know one, for instance we talked about the SHIFT key which set the 6th bit, but that’s the exception to the bucky bit criteria because SHIFT isn’t considered a bucky bit changer.

One of those “bucky bit setter” originated from the MIT Lisp machine keyboard, the space-cadet keyboard, and the Sun Microsystems keyboard, which had a META and diamond keys which would set the 8th bit to true, the 8th bit which value was left unspecified because the ASCII set fit in 7 bits and not 8.

And so by only checking the 8th bit is set you can know if META/Special/Super is pressed or not.

The Meta key is used a lot these days for multiple purposes, but that’s all left to the software interpretation and not the terminal. There are now many new bucky bit setters these days but that’s not a subject for this podcast.

On the Alt, Meta, & Esc Keys

Another modifier is the ALT key, the weird kiddo in the family.

It’s original purpose was more or less one of a bucky-bit setter, but it then morphed with time.

ALT historically comes from IBM PCs where you held the ALT key and typed decimal numbers on the keypad to insert character by their keycode instead of typing them. It was used for characters that didn’t appear on the keyboard such as UTF-8 chars for other languages. The way it worked was by using a BIOS buffer that would wait for input when ALT was pressed and then map it as if it was a single keystroke.

This is a feature we today refer to as a “compose key” or “Alt codes key”.

ALT was re-purposed to serve as a META key when Unices were ported to the PC, but the technique of setting the 8th bit to 1 was not so compatible with the advent of networking, which would mess up the bits. That last bit is not inside ASCII. META-aware application checking the 8th bit was not viable anymore, it would cause troubles.

Instead they found a new way to do META-like behavior, they would first send and “ESC” before sending the character that was pressed when holding ALT.

This more or less worked because applications were already handling ESCape codes, a series of characters to manipulate at a software level the appearance of text on the screen, and so they were already waiting to process the rest of the stream.

Those escape sequences we talked about earlier.

Now meta would mean, “if you see escape whatever follows should be considered as if it was entered with META held down”.

This creates the issue that you can’t have separate key interpretation in your terminal software for ESC and ALT+char since they are the same thing.

ESC sends directly an escape characters while ALT waits for a second character to follow.

For instance when in Vim while in insert mode if you press ESC or ALT+anything it will go back to normal mode.

Or open a terminal and type some command for instance “ls -lah” then try “ALT-b” to go a word backward, then “ALT-f” to go a word forward, now replace ALT with ESC but press each key separately, see how it accomplishes the same thing.

Now you too have to add to the equation what we’ve talked about earlier about equivalence when you hold the control key and you get an idea of the mess.

Moreover, on slow and low quality connections this can cause issues if the handling of the ALT key is left to the application at the other end, this is usually corrected by increasing the time it waits for a key following the ALT.

Internal Functions Keys

One last thing to mention are the internal to the terminal debugging or functionalities/features.

There are a bunch of keys, the function keys that act on the terminal itself and are not transferred.

According to your definition and depending on the features of those function keys you could possibly call those terminal “intelligent terminals”.

Those keys were originally used to debug and have an idea of what was sent over the wire.

On other terminals where you could connect to multiple computers they are used to switch between sessions.

But I have something to admit, whatever I said about the purpose of keys and their interpretations is a bit of a lie, more or less.

Those all depends on the terminal settings and their interpretation by that terminal. It’s not anchored, so take it lightly. This is what we’re gonna discuss next.

Terminal settings

As I’ve said those special characters varies from one terminal to the other, but they also vary in their interpretation at the host level. This layer of configuration and agreement between the host and terminal, of what settings are set for what terminal is what we’re going to discuss now.

Later on we’ll deal with the architecture of the terminal emulation interface and how things fall into place.

The actual interpretation of control codes resides in the POSIX termios interface.

The termios functions describe a general terminal interface that is provided to control asynchronous communications ports.

Physical terminals

Historically, physical terminal were connected to a UART, a Universal Asynchronous Receiver and Transmitter. Which is a hardware component that controls serial communications sending bytes asynchronously from one end to the other with adjustable speed rate.

The UART is a computer hardware device for asynchronous serial communication in which the data format and transmission speeds are configurable.

So the terminal and host are connected to the UART and the OS manages this via the UART driver to control the speed of the transmission, the parity, and the control flow.

In direct relation with that, Unix systems used to manipulate and interpret the input and output streams with the ioctl() system call before that with stty() and gtty() system calls.

ioctl() reads stream from the terminal device file, as with Unix everything is a file.

This is called “line editing” or “line discipline”, which is a buffer that the OS provides for interpreting, editing, erasing, clearing lines, reprinting, echoing, etc..

Basically, a “line discipline” aka LDISC is an input/output policy for interpretation of characters that goes with the terminal requirements, it sits one layer above the serial driver and shapes its behavior. It’s the glue between kernel space and user space.

There’s a “line disciplines” for every terminal attached to the machine according to the terminal and host settings.

On Linux you can check the current line discipline for the tty in use at the special: /proc/tty/ldiscs file.

Along with that comes the session management that knows what to do with processes according to the control character received. For instance putting a process in the background or foreground.

We’ll discuss more of that later with the architecture.

Together, the UART driver, the line discipline instance, and the TTY driver, can be referred to as the TTY.

Standards

Simply explained for now, there’s a controlling terminal, a terminal your shell is connected to, a terminal with specific settings, and there’s a line discipline, or some layer that intercepts and understands control characters you enter via that terminal.

But it was not all that great in the days.

For instance there was a time when some keys were set in the drivers, for instance to represent “erase” and “kill” you were forced to use ‘#’ and ‘@’ and you couldn’t change it. Or then later you could only change it as part of the login process.

There were many different interfaces for different Unixes which led to the need for standardization, which POSIX took on.

They made every Unix use a single specification for the “line discipline”, and integrated the BSD job control (which we’ll discuss later). They dumped the raw ioctl(), which was different from platform to platform, for the more generic abstract termios that can be ported to any architecture.

The names and parameters to configure the terminals got standardized through a data structure that contains all the options.

Now processes can read and modify the configurations of the opened TTY devices via a cleaner and stable interface. And if you’re writing a program that intercepts those control characters or disables them or interact with terminal streams you can use that as the base.

That data structure in the termios, a simple C data structure, is the representation of a terminal interface.

There are also functions that let you interact with that structure, namely the tcgetattr and tcsetattr.

struct termios {
    tcflag_t c_iflag ;  // Input modes
    tcflag_t c_oflag ;  // Output modes
    tcflag_t c_cflag ;  // Control modes
    tcflag_t c_lflag ;  // Local modes
    cc_t c_cc[NCCS] ;   // Control characters
};

The data structure contains the UART parameters, the line discipline stuffs, and the control stuffs, all in one place.

It has flags for many configs that you can enable or disable, there’s a good list of them in the termios man page.

For instance you can set the baud rate, the speed of the UART, or the parity, with the c_cflag field in the structure. Or you can change the input or output modes with the c_iflag and c_oflag, or even what the control characters mean via the c_cc field, or how signals are fired via the c_lflag.

The output processing flag affect how certain characters are displayed on the screen, for instance the newline and tab characters.

The input processing c_iflag directly affect how the read() system call works on a terminal device and how the line discipline and signal-generation will work.

There are multiple modes that are possible:

There’s the character mode aka character-at-a-time or raw mode, which let’s everything you type to be sent immediately to the receiving system without even doing any line editing or interpretation of control sequences. Everything is treated as normal character input and the application receives entire character streams unaltered.

There’s the line mode aka cooked mode aka line-at-a-time mode, which provides a local line editing buffer function, and sends an entire input, when pressing RETURN, line at a time after passing through the line discipline. A so-called “line mode terminal” operates solely in this mode.

There’s the cbreak mode, which is another subset of the character-at-a-time modes jokingly referred to as a “half-cooked” mode or “rare” mode, which like the raw mode sends directly all characters to the application without line discipline performed however it handles “interrupt” and “quit” control characters, as well as modem flow control or signal generation characters.

There’s a last mode which is the block mode, aka screen-at-a-time mode, which sends whole screens to and from the terminal. The user fills many forms and then sends the data to the server. Sort of like a web REST API.

The popular terminal IBM 3270 for the IBM mainframe we talked about earlier is block oriented.

When you think of it block modes are one of the instance of “intelligent terminal” in one of the earlier definitions we gave.

POSIX doesn’t have anything for that kind of the terminals, it wasn’t built for block terminals.

Some refer to the line mode as canonical and the raw mode as non-canonical.

Command Line

The programmatic interface to control terminals might be too tedious so a much more convenient way to configure the terminal is to use the stty command directly from the shell. It is a thin wrapper around the functions of the termios API to set and get values.

The stty command acts on a terminal device and by default if you don’t pass the parameter it acts on the one you are connected to.

To get the terminal you are currently sitting at you can use the tty command, but it might not be portable.

When you call stty without arguments it prints out the differences between the current settings and what it considers “sane” defaults. When you pass the -a, or everything on BSD, it print the values for all the flags of the termios structure in a human readable format.

You can also set or unset flags and options as you wishes adding a dash(-) before it for switching off and nothing for setting it.

Overall, it’s an easier way to interact with termios.

For example, you can play around with the control characters by changing what the interrupt character is instead of the usual default of ctrl-c or change the erase character from backspace to something else or change the word erase character from ctrl-w to something else or you can disable control flow or change the baud rate, change between canonical and non-canonical mode, enable echoing back to the screen, change the number of rows and lines, etc..

There are a lot of interesting things you can do and it’s a portable way to do it.

You can add those to your shell so that your preferred settings are there at login. The issue becomes if your terminal emulator intercepts and does something with it before sending it to the line discipline, which then nullifies its effect.

Though it’s better to get used to the default “sane” settings, like ctrl-u for line erase, like that you’ll be at ease on any terminal.

Applications Expectations

Applications are written to run on many, if not all, terminals.

It doesn’t want to care if it’s a physical “vt100” or if it’s PUTTY on windows.

The software gets “told” which terminal it is currently working on and assumes and expect certain capabilities in response to that.

Usually that name is stored in an environment variable called TERM, this variable can be automatically set or manually set.

The tset or reset or getty programs initializes that value by querying the terminal trying to find a sensible state, it’s usually called inside a .profile or anything that’s there before.

That implies that before login the system can not have any idea what kind of terminal you are running. It’s a misconception that misconfigured terminfo/termcap/gettytab or $TERM environment variable affects the ability to login to a terminal.

Sometimes the expectations of the program are wrong, an example would be to think the terminal has 20 columns instead of 10 and that can mess the output but that usually happens if it’s a badly written program.

Most softwares assume the most common default terminal settings, the standard one we discussed in an earlier section, the VT100, just so that the software and terminal can agree on something.

But How Do You Know which settings goes with which terminal?

How do you map the terminal settings related to the terminal in that TERM environment variable?

That’s where terminfo and termcap appear.

They are are databases of terminal capabilities and features which mainly are comprise of the escape codes that are supported by the terminals, those are mostly for visual attributes.

But it doesn’t take car of everything, for instance it’s the tabs program that is used to control the tab behavior on terminals and not necessarily terminfo nor termcap.

The terminfo is the newest version of the termcap one, termcap can be converted to terminfo using the captoinfo command.

termcap is a text file in /etc/termcap while terminfo is a bunch of binary files in sub-directories under /usr/lib/terminfo for every different type of terminal.

Libraries such as ncurses are built on top of terminfo so that it can use the escape codes properly.

It’s better to retrieve them from the database then to have them hardcoded, and yes, there was a time where those have been hardcoded in softwares.

There’s a much simpler version of termcap for BSD that is called by getty during login called gettytab.

For example checkout the infocmp command which will output the escape characters for your current terminal or the tput which can be used to query terminfo.

For more information you can check the man pages for terminfo and termcap.

The architecture

Now let’s deal with the architecture.

How does everything fall into place, all the pieces.

First of all let’s review a bit of what we’ve seen so far.

We have the physical terminals with their history, capabilities, and keys that are connected via RS-232 to the host.

Or those could be virtual terminals, we’ll discuss how those work in a bit.

Then there’s the mapping of the terminal as a file device in /dev so that it can be interacted with.

For instance you can check the man page ttyS to find out how to manually create this file for serial terminals.

After that there’s the actual login, then the tset or getty to get & set the TERM environment variable to the appropriate value and appropriate related terminfo.

The getty utility is called by init(8) to open and initialize the tty line, read a login name, and invoke login(1).

The argument tty is the special device file in /dev to open for the ter- minal (for example, ``ttyh0’’). If there is no argument or the argument is `-‘, the tty line is assumed to be open as file descriptor 0.

After that starts the actual architectural cycle of input, send, interpret, loop.

What we’ve seen so far of that is the actual physical input on the terminal, the transfer of 8 bits for RS-232 (or RS-422 or RS-485 serial interface) , the UART to control the rate of transfer/baud-rate, and the line discipline that is attached to this current terminal with its settings.

Now let’s actually finish that cycle by discussing the architecture of virtual terminals, of virtual consoles, and of session and process control that we said we would come back to.

Generic Arch

Unix was designed with this sort of thin client, or dumb text terminal if you prefer, in mind. Much of the architectural thinking emerges from that.

That’s why it was helpful to review a bunch of things about terminal before explaining the architecture.

On most Unix there are many terminal drivers responsible for the underlying control of the I/O instructions and interrupt requests for input and output, that can be used in different situations, the console driver, the virtual console (vc) driver, the serial driver, the pseudo terminal drivers, etc..

The line discipline itself is independent from the driver.

/proc/tty/drivers
/dev/tty             /dev/tty        5       0 system:/dev/tty
/dev/console         /dev/console    5       1 system:console
/dev/ptmx            /dev/ptmx       5       2 system
/dev/vc/0            /dev/vc/0       4       0 system:vtmaster
serial               /dev/ttyS       4 64-95 serial
pty_slave            /dev/pts      136 0-1048575 pty:slave
pty_master           /dev/ptm      128 0-1048575 pty:master
unknown              /dev/tty        4 1-63 console

We’ve already seen how the physical line for physical terminal worked, and this doesn’t change much even if you add a long-distance phone line in the middle, be it a corporate terminal or console server or not (Which is outside the scope of this podcast).

Terminal <-> modem <-> physical line <-> modem <-> UART <-> ...

Now how does the Linux console work, it’s not a physical terminal, it’s directly attached to the same machine you are running Linux on.

You got on one side the display with its VGA driver. On the other side you have the keyboard with its keyboard driver.

Both are connected to a terminal emulator, one end for output - the screen, and the other for input - the keyboard.

This terminal emulator is connected to the line discipline as usual plus the TTY driver for the capabilities and the rest, and that TTY driver handles user process management.

display <- VGA driver <- terminal emulator <-> line discipline <-> TTY driver <-> user process
keyboard -> keyboard driver .-^

The only big difference is that there’s no UART, this is handled in the emulator, “video terminal”, instead.

Let’s also note that this doesn’t employ terminfo, the console has its own set of escape codes and special sequences hardcoded.

Or sometimes is has a simplified version of those info, for example the gettytab file, which is a simplified version of the termcap database accessed during the login process.

When we move up to userland things get more abstract.

It uses a concept called the pseudo terminal or pty a virtual terminal, which now requires two different driver part (under /dev again), two different end-points, one called the PTY slave or pts and another the PTY master or ptmx.

Both ends are bidirectional communication channels but do different jobs.

The processes connect to the slave end, the slave behaves exactly like a classic terminal, the slave driver responds to what the line discipline sends it.

On the other end, the master side is the one that acts as input and output. It sends the characters to the line discipline so that it can interact with the slave and receives back the output of the interaction.

It’s the master that is connected to your terminal emulator and the slave that is connected to the program you are running.

Whatever is written in one end can be read out from the other after passing through the line discipline.

Line discipline <-> TTY driver (PTY slave side) <-> user process
 `-> PTY master side <-> xterm process

Technically under Linux the ptmx is a character file with major number 5 and minor number 2 and owned by the root user. When a process opens /dev/ptmx it gets back from it a file descriptor to a slave, a pts device under /dev/pts directory, which will be owned by the user that requested it.

Every time you request from the master you get a different slave, this slave file is the reply to the tty command we talked about earlier.

Note that there’s usually a limit to the number of opened pseudo terminal, in Linux you can check it in /proc/sys/kernel/pty/max, usually 256. Another file /proc/sys/kernel/pty/nr in Linux indicates the number of already opened pseudo terminals.

Now when you want to open the pseudoterminal slave for process interaction you must pass the master’s file descriptor to be able to unlock it, so that it grants you access.

There’s a bunch of POSIX system calls that helps with all that such as posix_openpt, getpt, grantpt, unlockpt, ptsname, etc..

The BSD style of pseudo terminals provides them as precreated pairs with names of the form of /dev/ptyXY for the master and /dev/ttyXY for the slave.

For example /dev/ptyp1 and /dev/ttyp1 is a BSD pseudoterminal pair. When a process wants a pseudo terminal pair on BSD it tries to open(2) each pseudo terminal master until it succeeds.

It’s also nice to know that you can send data to the slave pseudo tty from the command line and it’ll get processed and outputed to the master end. It’s useful to send input to programs that don’t usually accept input such as su or passwd. It can also be used for chatting.

This whole pseudo terminal thing is a bit of a hack to facilitate terminal emulation in userland and not break the TTY subsystem.

Also here, you can note that there’s no UART, obviously.

Pseudo terminals have kernel buffer to sync the rate of data flowing which is the same thing as the baud-rate.

When the process can’t call write(2) to a stream because the buffer is filled it puts the process in a sleep state, and then alternates it back to running when it can.

And as with physical terminals you can stop the stream using flow control characters XON, XOFF.

This shows how the OS process states are tied to the flow.

On a terminal you usually run processes, and frequently it’s through the shell, the command prompt.

So let’s take a step back and see how all of this fits into the process model.

For example when ctrl-c is received it won’t be handed off to the application through read(2), but will instead cause a SIGINT to be delivered to the foreground job.

Why and how is that?

Signals

In kernel land, things communicate via interrupts, be them hardware interrupts or software interrupts.

Signals are a sort of software interrupt and this is what the kernel uses to communicate with processes. In our case, it’s the line discipline that lives in kernel land and that sends that to the processes.

If you want more info about signals you can go back to the episode we had about them.

One interesting thing to add to what we previously discussed is how the name and behavior of the signals originated from the actual physical terminals.

For instance SIGHUP is a signal that indicates that the UART driver hanged up, that the connection is cut.

SIGINT, is a signal sent by the TTY driver to the foreground process to say that the control character for interrupt, ctrl-c usually, has been received.

There’s also a signal sent for when the terminal size changes, the SIGWINCH.

There are quite a bunch of them.

Job Control

If you remember I’ve mentioned that POSIX implemented along with the termios the job control mechanism of BSD.

It was a feature of the C shell.

This allowed flexible manipulation of processes.

A shell gets a terminal session, that means access to a slave terminal for example.

And then there’s an array of processes under that session called the process group.

The processes within a group are called jobs.

There’s the notion of foreground process in the tty driver, the process that receives the signal that is sent to that session, and the process who may perform input and output to that terminal.

There’s only one foreground process for one session on a terminal. It’s a way to control which process may perform input output from the terminal at a given time.

There are also signal sent because processes from that group that aren’t the foreground process, so called background processes, because there can only be one foreground, are trying to access the controlling terminal.

The grouping also allows you to deliver signals to them all at the same time.

You can swap between which job within the group is the foreground or background one using those signals, putting the one in the foreground in running state and the others in the wait state.

For example when you login you get a session created by the setsid() system call on the login terminal, and all processes there are under the same group and inherit this terminal, and the foreground process is the session leader of that process group.

It’s the session leader that tells the terminal which job is the foreground one.

When a controlling terminal process terminates the session ends and it allows it to be acquired by a new session leader.

Process management is outside of the scope of this episode so we’ll keep to what is interesting to us.

This is the kind of structure inside a TTY driver, here pts:

TTY Driver (/dev/pts/0).

    Size: 45x13
    Controlling process group: (101)
Foreground process group: (103)
    UART configuration (ignored, since this is an xterm):
        Baud rate, parity, word length and much more.
        Line discipline configuration:
        cooked/raw mode, linefeed correction,
        meaning of interrupt characters etc.
        Line discipline state:
        edit buffer (currently empty),
        cursor position within buffer etc.

Overall, it’s all about juggling with signals and swapping between foreground, background.

Conclusion

One thing we forgot to add to the equation is the layer the shell has before interacting with the terminal architecture. This adds more complexity as the shell can capture and interpret characters which could be other control characters. And over all that, if you’re running an X environment and communicating over SSH, this all becomes overly complex for the simple explanation I provide here.

If terminals interests you you can get one of the old physical one and an adapter for the RS-232 protocol, or whatever the protocol that terminal uses, so that you can have fun with it.

Terminals themselves and the architecture is a bit messy but it is still very useful and widely used. You can’t really remove terminals because of the huge baggage, the existing tooling, and expectations softwares have for them. If you really wanted to replace them you’d have to support backward compatibility and it would go against the point of innovation.

This is all truly fascinating, the whole story and how it evolved with time and all the pieces joining each others.

This was a bit of a rough explanation but I hope it was fun to go through.

As usual, if you want more info you can check the show notes and the related manpages. You can even go through the POSIX standard page, it’ll be easier to read now.

Cheers!


Music: Libre De Droits/Creative Commons - Volume 1 : Experiences by Maxime Blégent (Azolan Sollicitus) https://azolanmusic.bandcamp.com/album/libre-de-droits-creative-commons-volume-1-experiences


References:

History:

Text-terminal:

Keys, Modifying Keys & Control Chars

Terminal Settings

Architecture:




If you want to have a more in depth discussion I'm always available by email or irc. We can discuss and argue about what you like and dislike, about new ideas to consider, opinions, etc..
If you don't feel like "having a discussion" or are intimidated by emails then you can simply say something small in the comment sections below and/or share it with your friends.