Many operating systems that came before Unix treated each input or output device differently. Each time a new device became available, programs would have to be modified in order to access it. This is intuitive, since the devices all look different and perform different functions.

The Unix designers realized that this is actually unnecessary and a waste of programming effort, so they developed the concept of device independence. What this means is that Unix treats virtually every input and output device exactly like an ordinary file. All input and output, whether to/from a file on a disk, a keyboard, a mouse, a scanner, or a printer, is simply a stream of bytes to be input or output using the same tools.

Most I/O devices are actually accessible as a device file in /dev. For example, the primary CD-ROM might be /dev/cd0, the main disk /dev/ad0, the keyboard /dev/kbd0, and the mouse /dev/sysmouse.

A user with sufficient permissions can view input coming from these devices using the same Unix commands we use to view a file:

shell-prompt: cat /dev/kbd0
shell-prompt: more /dev/cd0
            

In fact, data are often recovered from corrupted file systems or accidentally deleted files by searching the raw disk partition as a file using standard Unix commands such as grep!

shell-prompt: grep string /dev/ad0s1f
            

A keyboard sends text data, so /dev/kbd0 is like a text file. Many other devices send binary data, so using cat to view them would output gibberish. To see the raw input from a mouse as it is being moved, we could instead use hexdump, which displays the bytes of input as numbers rather than characters:

shell-prompt: hexdump /dev/sysmouse
            

Some years ago while mentoring my son's robotics team, as part of a side project, I reverse-engineered a USB game pad so I could control a Lego robot via Bluetooth from a laptop. Thanks to device-independence, no special software was needed to figure out the game pad's communication protocol.

After plugging the game pad into my FreeBSD laptop, the system creates a new UHID (USB Human Interface Device) under /dev. The dmesg command shows the name of the new device file:

ugen1.2: <vendor 0x046d product 0xc216> at usbus1
uhid0 on uhub3
uhid0: <vendor 0x046d product 0xc216, class 0/0, rev 1.10/3.00, addr 2> on usbus1
            

One can then view the input from the game pad using hexdump:

FreeBSD manatee.acadix  bacon ~ 410: hexdump /dev/uhid0
0000000 807f 7d80 0008 fc04 807f 7b80 0008 fc04
0000010 807f 7780 0008 fc04 807f 6780 0008 fc04
0000020 807f 5080 0008 fc04 807f 3080 0008 fc04
0000030 807f 0d80 0008 fc04 807f 0080 0008 fc04
0000060 807f 005e 0008 fc04 807f 005d 0008 fc04
0000070 807f 0060 0008 fc04 807f 0063 0008 fc04
0000080 807f 006c 0008 fc04 807f 0075 0008 fc04
0000090 807f 0476 0008 fc04 807f 1978 0008 fc04
00000a0 807f 4078 0008 fc04 807f 8c7f 0008 fc04
00000b0 807f 807f 0008 fc04 807f 7f7f 0008 fc04
00000c0 807f 827f 0008 fc04 807f 847f 0008 fc04
00000d0 807f 897f 0008 fc04 807f 967f 0008 fc04
00000e0 807f a77f 0008 fc04 807f be80 0008 fc04
00000f0 807f d980 0008 fc04 807f f780 0008 fc04
0000100 807f ff80 0008 fc04 807f ff83 0008 fc04
0000110 807f ff8f 0008 fc04 807f ff93 0008 fc04
            

To understand these numbers, we need to know a little about hexadecimal, base 16. This is covered in detail in ???. In short, it works the same as decimal, but we multiply by powers of 16 rather than 10, and digits go up to 15 rather than 9. Digits for 10 through 15 are A, B, C, D, E, and F. The largest possible 4-digit number is therefore FFFF_16. 8000_16 is in the middle of the range.

0000_16 =  0 * 16^3 +  0 * 16^2 +  0 * 16^1 +  0 * 16^0 = 0_10
8000_16 =  8 * 16^3 +  0 * 16^2 +  0 * 16^1 +  0 * 16^0 = 32,678_10
FFFF_16 = 15 * 16^3 + 15 * 16^2 + 15 * 16^1 + 15 * 16^0 = 65,535_10
            

It was easy to see that moving the right joystick up resulted in lower numbers in the 3rd and 7th columns, while moving down increased the values. Center position sends a value around 8000 (hexadecimal), fully up is around 0, fully down is ffff.

It was then easy to write a small program to read the joystick position from the game pad (by simply opening /dev/uhid0 like any other file) and send commands over Bluetooth to the robot, adjusting motor speeds accordingly. The Bluetooth interface is simply treated as an output file.

Since I/O devices and files are interchangeable, Unix shells can provide a facility called redirection to easily interchange them for any process without the process even knowing it.

Redirection depends on the notion of a file stream. You can think of a file stream as a hose connecting a program to a particular file or device. Redirection simply disconnects the hose from the default file or device (such as the keyboard or terminal screen) and connects it to another file or device chosen by the user.

Every Unix process has three standard streams that are open from the moment the process is born. The standard streams for a shell process are normally connected to the terminal, as shown in Table 1.9, “Standard Streams”.

Redirection in the shell allows any or all of the three standard streams to be disconnected from the terminal and connected to a file or other I/O device. It uses special operator characters within the commands to indicate which stream(s) to redirect and where. The basic redirection operators shells are shown in Table 1.10, “Redirection Operators”.

shell-prompt: ls > listing.txt         # Overwrite with listing of .
shell-prompt: ls /etc >> listing.txt   # Append listing of /etc
            

In the examples above, the ls process sends its output to listing.txt instead of the standard output. However, the filename listing.txt is not an argument to the ls process. The ls process never even knows about this output file.

The redirection is handled by the shell and the shell removes "> listing.txt" and ">> listing.txt" from these commands before executing them. So, the first ls receives no arguments, and the second receives only /etc. Most programs have no idea whether their output is going to a file, a terminal, or some other device. They don't need to know and they don't care.

More often than not, we want to redirect both normal output and error messages to the same place. This is why C shell and its derivatives use a combined operator that redirects both at once.

shell-prompt: find /etc -name >& all-output.txt
            

The same effect can be achieved with Bourne-shell derivatives using another operator that redirects one stream to another stream. In particular, we redirect the standard output (stream 1) to a file (or device) and at the same time redirect the standard error (stream 2) to stream 1.

shell-prompt: find /etc > all-output.txt 2>&1
            

In Bourne family shells, we can separately redirect the standard output with > and the standard error with 2>:

shell-prompt: find /etc > list.txt 2> errors.txt
            

If we want to separate standard output and standard error in a C shell or T shell session, we can use a subshell under which the find command redirects only the standard output. The output from the subshell process will then only contain the standard error left over from find, which we can redirect with &>:

shell-prompt: (find /etc > list.txt) >& errors.txt
            

If a program takes input from the standard input, we can redirect input from a file as follows:

shell-prompt: command < input-file
            

For example, consider the "bc" (binary calculator) command is an arbitrary-precision calculator which inputs numerical expressions from the standard input and writes the results to the standard output. It's a good idea to use the --mathlib flag with bc for more complete functionality.

shell-prompt: bc --mathlib
3.14159265359 * 4.2 ^ 2 + sqrt(30)
60.89491440932
quit
            

In the example above, the user entered "3.14159265359 * 4.2 ^ 2 + sqrt(30)" and "quit" and the bc program output "60.89491440932". We could instead place the input shown above in a file using any text editor, such as nano or vi, or even using cat with keyboard input and output redirection as a primitive editor:

shell-prompt: cat > bc-input.txt
3.14159265359 * 4.2 ^ 2 + sqrt(30)
quit
(Type Ctrl+d to signal the end of input to the cat process)
shell-prompt: cat bc-input.txt
3.14159265359 * 4.2 ^ 2 + sqrt(30)
quit
            

Now that we have the input in a file, we can feed it to the bc process using input redirection instead of retyping it on the keyboard:

shell-prompt: bc --mathlib < bc-input.txt 
60.29203070318
            

The standard streams themselves are represented as device files on Unix systems. This allows us to redirect one stream to another without modifying a program, by appending the stream to one of the device files /dev/stdout or /dev/stderr. For example, if a program sends output to the standard output and we want to send it instead to the standard error, we could do something like the following:

shell-prompt: printf "Oops!" >> /dev/stderr
            

If we would like to simply discard output sent to the standard output or standard error, we can redirect it to /dev/null. For example, to see only error messages (standard error) from myprog, we could do the following:

shell-prompt: ./myprog > /dev/null
            

To see only normal output and not error messages, assuming Bourne shell family:

shell-prompt: ./myprog 2> /dev/null
            

In C shell family:

shell-prompt: ( find /etc > output.txt ) >& /dev/null ; cat output.txt
            

The device /dev/zero is a readable file that produces a stream of zero bytes.

The device /dev/random is a readable file that produces a stream of random integers in binary format. We can use the dd command, a bit copy program, to copy a fixed number of bytes from one file to another. We specify the input file with "if=", output with "of=", block size with "bs=", and the number of blocks with "count=". Total data copied will be block-size * count.

shell-prompt: dd if=/dev/random of=random-data bs=1000000 count=1
            

Very often, we want to use the output of one program as input to another. Such a thing could be done using redirection, as shown below:

shell-prompt: ls > listing.txt
shell-prompt: more listing.txt
            

The same task can be accomplished in one command using a pipe. A pipe redirects one of the standard streams, just as redirection does, but to or from another process instead of a file or device. In other words, we can use a pipe to send the standard output and/or standard error of one process directly to the standard input of another process.

A pipe is constructed by placing the pipe operator (|) between two commands. The whole chain of commands connected by pipes is called a pipeline.

shell-prompt: ls | more
                

Since a pipe runs multiple processes in the same shell, it is necessary to understand the concept of foreground and background processes, which are covered in detail in Section 1.18, “Process Control”.

Multiple processes can output to a terminal at the same time, although the results would obviously be chaos in most cases.

In contrast to output, only one process can receiving input from the keyboard, however. It would be a remarkable coincidence if the same input made sense to two different programs.

The foreground process running under a given shell process is defined as the process that receives the input from the standard input device (usually the keyboard). This is the only difference between a foreground process and a background process.

When running a pipeline command, the last command in the pipeline becomes the foreground process. All others run in the background, i.e. do not use the standard input device inherited from the shell process. Hence, when we run:

shell-prompt: ls | more
            

It is the more command that receives input from the keyboard. The more command has its standard input redirected from the standard output of ls, and the standard input of the ls command is effectively disabled.

This is such a common practice that Unix has defined the term filter to apply to programs that can be used in this way. A filter is any command that can receive input from the standard input and send output to the standard output. Many Unix commands are designed to accept a file name as an argument, but to use the standard input and/or standard output if no filename arguments are provided.

shell-prompt: more names.txt
shell-prompt: more < names.txt
                

shell-prompt: ls | more
shell-prompt: find . -name '*.c' | more
shell-prompt: sort names.txt | more
                

We can string any number of commands together using pipes. The only limitations are imposed by the memory requirements of the processes in the pipeline. For example, the following pipeline sorts the names in names.txt, removes duplicates, filters out all names not beginning with 'B', and shows the first 100 results one page at a time.

shell-prompt: sort names.txt | uniq | grep '^B' | head -n 100 | more
            

To see lines 101 through 200 of a file output.txt:

shell-prompt: head -n 200 output.txt | tail -n 100
            

One more useful tool worth mentioning is the tee command. The tee command is a simple program that reads from its standard input and writes to both the standard output and to one or more files whose names are provided on the command line. This allows you to view the output of a program on the screen and save it to a file at the same time.

shell-prompt: ls | tee listing.txt
            

Recall that Bourne-shell derivatives do not have combined operators for redirecting standard output and standard error at the same time. Instead, we redirect the standard output to a file or device, and redirect the standard error to the standard output using 2>&1.

We can use the same technique with a pipe, but there is one more condition: For technical reasons, the 2>&1 must come before the pipe.

shell-prompt: ls | tee listing.txt 2>&1    # Won't work
shell-prompt: ls 2>&1 | tee listing.txt    # Will work
            

The yes command (much like Jim Carrey in "Yes Man") produces a stream of y's followed by newlines. It is meant to be piped into a program that prompts for y's or n's in response to yes/no questions, so that the program will receive a yes answer to all of its prompts and run without user input.

shell-prompt: yes | ./myprog
            

The yes command can actually print any response we want, via a command line argument. To answer 'n' to every prompt, we could do the following:

shell-prompt: yes n | ./myprog
            

In cases where the response isn't always the same, we can feed a program any sequence of responses using redirection or pipes. Be sure to add a newline (\n) after each response to simulate pressing the Enter key:

shell-prompt: printf "y\nn\ny\n" | ./myprog
            

Or, to save the responses to a file for repeated use:

shell-prompt: printf "y\nn\ny\n" > responses.txt
shell-prompt: ./myprog < responses.txt
            

Users who don't fully understand Unix and processes often fall into bad habits that can potentially be costly. There are far too many such habits to cover here: One could write a separate 1,000-page volume called "Favorite Bad Habits of Unix Users". As a less painful alternative, we'll explore one common bad habit in detail and try to help you understand how to spot others. Our feature habit of the day is the use of the cat command at the head of a pipeline:

shell-prompt: cat names.txt | sort | uniq > outfile
            

So what's the alternative, what's wrong with using cat this way, what's the big deal, why do people do it, and how do we know it's a problem?