An Introduction to the UNIX Shell

S. R. Bourne

ABSTRACT

The shell is a command programming language that provides an interface to the UNIXoperating system. Its features include control-flow primitives, parameter passing, variables and string substitution. Constructs such as while, if then else, case and for are available. Two-way communication is possible between the shell and commands. String-valued parameters, typically file names or flags, may be passed to a command. A return code is set by commands that may be used to determine control-flow, and the standard output from a command may be used as shell input.

The shell can modify the environment in which commands run. Input and output can be redirected to files, and processes that communicate through `pipes' can be invoked. Commands are found by searching directories in the file system in a sequence that can be defined by the user. Commands can be read either from the terminal or from a file, which allows command procedures to be stored for later use.


1.0 Introduction

The shell is both a command language and a programming language that provides an interface to the UNIX operating system. This memorandum describes, with examples, the UNIX shell. The first section covers most of the everyday requirements of terminal users. Some familiarity with UNIX is an advantage when reading this section; see, for example, "UNIX for beginners". unix beginn kernigh 1978 Section 2 describes those features of the shell primarily intended for use within shell procedures. These include the control-flow primitives and string-valued variables provided by the shell. A knowledge of a programming language would be a help when reading this section. The last section describes the more advanced features of the shell. References of the form "see pipe (2)" are to a section of the UNIX manual. seventh 1978 ritchie thompson

1.1 Simple commands

Simple commands consist of one or more words separated by blanks. The first word is the name of the command to be executed; any remaining words are passed as arguments to the command. For example,

who
is a command that prints the names of users logged in. The command
ls -l
prints a list of files in the current directory. The argument -l tells ls to print status information, size and the creation date for each file.

1.2 Background commands

To execute a command the shell normally creates a new process and waits for it to finish. A command may be run without waiting for it to finish. For example,

cc pgm.c &
calls the C compiler to compile the file pgm.c. The trailing & is an operator that instructs the shell not to wait for the command to finish. To help keep track of such a process the shell reports its process number following its creation. A list of currently active processes may be obtained using the ps command.

1.3 Input output redirection

Most commands produce output on the standard output that is initially connected to the terminal. This output may be sent to a file by writing, for example,

ls -l >file
The notation >file is interpreted by the shell and is not passed as an argument to ls. If file does not exist then the shell creates it; otherwise the original contents of file are replaced with the output from ls. Output may be appended to a file using the notation
ls -l >>file
In this case file is also created if it does not already exist.

The standard input of a command may be taken from a file instead of the terminal by writing, for example,

wc <file
The command wc reads its standard input (in this case redirected from file) and prints the number of characters, words and lines found. If only the number of lines is required then
wc -l <file
could be used.

1.4 Pipelines and filters

The standard output of one command may be connected to the standard input of another by writing the `pipe' operator, indicated by |, as in,

ls -l | wc
Two commands connected in this way constitute a pipeline and the overall effect is the same as
ls -l >file; wc <file
except that no file is used. Instead the two processes are connected by a pipe (see pipe (2)) and are run in parallel. Pipes are unidirectional and synchronization is achieved by halting wc when there is nothing to read and halting ls when the pipe is full.

A filter is a command that reads its standard input, transforms it in some way, and prints the result as output. One such filter, grep, selects from its input those lines that contain some specified string. For example,

ls | grep old
prints those lines, if any, of the output from ls that contain the string old. Another useful filter is sort. For example,
who | sort
will print an alphabetically sorted list of logged in users.

A pipeline may consist of more than two commands, for example,

ls | grep old | wc -l
prints the number of file names in the current directory containing the string old.

1.5 File name generation

Many commands accept arguments which are file names. For example,

ls -l main.c
prints information relating to the file main.c.

The shell provides a mechanism for generating a list of file names that match a pattern. For example,

ls -l *.c
generates, as arguments to ls, all file names in the current directory that end in .c. The character * is a pattern that will match any string including the null string. In general patterns are specified as follows.
*
Matches any string of characters including the null string.
?
Matches any single character.
[...]
Matches any one of the characters enclosed. A pair of characters separated by a minus will match any character lexically between the pair.

For example,

[a-z]*
matches all names in the current directory beginning with one of the letters a through z.
/usr/fred/test/?
matches all names in the directory /usr/fred/test that consist of a single character. If no file name is found that matches the pattern then the pattern is passed, unchanged, as an argument.

This mechanism is useful both to save typing and to select names according to some pattern. It may also be used to find files. For example,

echo /usr/fred/*/core
finds and prints the names of all core files in sub-directories of /usr/fred. (echo is a standard UNIX command that prints its arguments, separated by blanks.) This last feature can be expensive, requiring a scan of all sub-directories of /usr/fred.

There is one exception to the general rules given for patterns. The character `.' at the start of a file name must be explicitly matched.

echo *
will therefore echo all file names in the current directory not beginning with `.'.
echo .*
will echo all those file names that begin with `.'. This avoids inadvertent matching of the names `.' and `..' which mean `the current directory' and `the parent directory' respectively. (Notice that ls suppresses information for the files `.' and `..'.)

1.6 Quoting

Characters that have a special meaning to the shell, such as < > * ? | &, are called metacharacters. A complete list of metacharacters is given in appendix B. Any character preceded by a \ is quoted and loses its special meaning, if any. The \ is elided so that

echo \\?
will echo a single ?, and
echo \\\\
will echo a single \. To allow long strings to be continued over more than one line the sequence \newline is ignored.

\ is convenient for quoting single characters. When more than one character needs quoting the above mechanism is clumsy and error prone. A string of characters may be quoted by enclosing the string between single quotes. For example,

echo xx'****'xx
will echo
xx****xx
The quoted string may not contain a single quote but may contain newlines, which are preserved. This quoting mechanism is the most simple and is recommended for casual use.

A third quoting mechanism using double quotes is also available that prevents interpretation of some but not all metacharacters. Discussion of the details is deferred to section 3.4.

1.7 Prompting

When the shell is used from a terminal it will issue a prompt before reading a command. By default this prompt is `$ '. It may be changed by saying, for example,

PS1=yesdear
that sets the prompt to be the string yesdear. If a newline is typed and further input is needed then the shell will issue the prompt `> '. Sometimes this can be caused by mistyping a quote mark. If it is unexpected then an interrupt (DEL) will return the shell to read another command. This prompt may be changed by saying, for example,
PS2=more

1.8 The shell and login

Following login (1) the shell is called to read and execute commands typed at the terminal. If the user's login directory contains the file .profile then it is assumed to contain commands and is read by the shell before reading any commands from the terminal.

1.9 Summary


2.0 Shell procedures

The shell may be used to read and execute commands contained in a file. For example,

sh file [ args ... ]
calls the shell to read commands from file. Such a file is called a command procedure or shell procedure. Arguments may be supplied with the call and are referred to in file using the positional parameters $1, $2, .... For example, if the file wg contains
who | grep $1
then
sh wg fred
is equivalent to
who | grep fred

UNIX files have three independent attributes, read, write and execute. The UNIX command chmod (1) may be used to make a file executable. For example,

chmod +x wg
will ensure that the file wg has execute status. Following this, the command
wg fred
is equivalent to
sh wg fred
This allows shell procedures and programs to be used interchangeably. In either case a new process is created to run the command.

As well as providing names for the positional parameters, the number of positional parameters in the call is available as $#. The name of the file being executed is available as $0.

A special shell parameter $* is used to substitute for all positional parameters except $0. A typical use of this is to provide some default arguments, as in,

nroff -T450 -ms $*
which simply prepends some arguments to those already given.

2.1 Control flow - for

A frequent use of shell procedures is to loop through the arguments ($1, $2, ...) executing commands once for each argument. An example of such a procedure is tel that searches the file /usr/lib/telnos that contains lines of the form

...
fred mh0123
bert mh0789
...
The text of tel is
for i
do grep $i /usr/lib/telnos; done
The command
tel fred
prints those lines in /usr/lib/telnos that contain the string fred.
tel fred bert
prints those lines containing fred followed by those for bert.

The for loop notation is recognized by the shell and has the general form

for name in w1 w2 ...
do command-list
done
A command-list is a sequence of one or more simple commands separated or terminated by a newline or semicolon. Furthermore, reserved words like do and done are only recognized following a newline or semicolon. name is a shell variable that is set to the words w1 w2 ... in turn each time the command-list following do is executed. If in w1 w2 ... is omitted then the loop is executed once for each positional parameter; that is, in $* is assumed.

Another example of the use of the for loop is the create command whose text is

for i do >$i; done
The command
create alpha beta
ensures that two empty files alpha and beta exist and are empty. The notation >file may be used on its own to create or clear the contents of a file. Notice also that a semicolon (or newline) is required before done.

2.2 Control flow - case

A multiple way branch is provided for by the case notation. For example,

case $# in
1) cat >>$1 ;;
2) cat >>$2 <$1 ;;
*) echo \'usage: append [ from ] to\' ;;
esac
is an append command. When called with one argument as
append file
$# is the string 1 and the standard input is copied onto the end of file using the cat command.
append file1 file2
appends the contents of file1 onto file2. If the number of arguments supplied to append is other than 1 or 2 then a message is printed indicating proper usage.

The general form of the case command is

case word in
pattern) command-list;;
...
esac
The shell attempts to match word with each pattern, in the order in which the patterns appear. If a match is found the associated command-list is executed and execution of the case is complete. Since * is the pattern that matches any string it can be used for the default case.

A word of caution: no check is made to ensure that only one pattern matches the case argument. The first match found defines the set of commands to be executed. In the example below the commands following the second * will never be executed.

case $# in
*) ... ;;
*) ... ;;
esac

Another example of the use of the case construction is to distinguish between different forms of an argument. The following example is a fragment of a cc command.

for i
do case $i in
-[ocs]) ... ;;
-*) echo \'unknown flag $i\' ;;
*.c) /lib/c0 $i ... ;;
*) echo \'unexpected argument $i\' ;;
esac
done

To allow the same commands to be associated with more than one pattern the case command provides for alternative patterns separated by a |. For example,

case $i in
-x|-y) ...
esac
is equivalent to
case $i in
-[xy]) ...
esac

The usual quoting conventions apply so that

case $i in
\\?) ...
will match the character ?.

2.3 Here documents

The shell procedure tel in section 2.1 uses the file /usr/lib/telnos to supply the data for grep. An alternative is to include this data within the shell procedure as a here document, as in,

for i
do grep $i <<!
...
fred mh0123
bert mh0789
...
!
done
In this example the shell takes the lines between <<! and ! as the standard input for grep. The string ! is arbitrary, the document being terminated by a line that consists of the string following <<.

Parameters are substituted in the document before it is made available to grep as illustrated by the following procedure called edg.

ed $3 <<%
g/$1/s//$2/g
w
%
The call
edg string1 string2 file
is then equivalent to the command
ed file <<%
g/string1/s//string2/g
w
%
and changes all occurrences of string1 in file to string2. Substitution can be prevented using \ to quote the special character $ as in
ed $3 <<+
1,\\$s/$1/$2/g
w
+
(This version of edg is equivalent to the first except that ed will print a ? if there are no occurrences of the string $1.) Substitution within a here document may be prevented entirely by quoting the terminating string, for example,
grep $i <<\\#
...
#
The document is presented without modification to grep. If parameter substitution is not required in a here document this latter form is more efficient.

2.4 Shell variables

The shell provides string-valued variables. Variable names begin with a letter and consist of letters, digits and underscores. Variables may be given values by writing, for example,

user=fred box=m000 acct=mh0000
which assigns values to the variables user, box and acct. A variable may be set to the null string by saying, for example,
null=
The value of a variable is substituted by preceding its name with $; for example,
echo $user
will echo fred.

Variables may be used interactively to provide abbreviations for frequently used strings. For example,

b=/usr/fred/bin
mv pgm $b
will move the file pgm from the current directory to the directory /usr/fred/bin. A more general notation is available for parameter (or variable) substitution, as in,
echo ${user}
which is equivalent to
echo $user
and is used when the parameter name is followed by a letter or digit. For example,
tmp=/tmp/ps
ps a >${tmp}a
will direct the output of ps to the file /tmp/psa, whereas,
ps a >$tmpa
would cause the value of the variable tmpa to be substituted.

Except for $? the following are set initially by the shell. $? is set after executing each command.

$?
The exit status (return code) of the last command executed as a decimal string. Most commands return a zero exit status if they complete successfully, otherwise a non-zero exit status is returned. Testing the value of return codes is dealt with later under if and while commands.
$#
The number of positional parameters (in decimal). Used, for example, in the append command to check the number of parameters.
$$
The process number of this shell (in decimal). Since process numbers are unique among all existing processes, this string is frequently used to generate unique temporary file names. For example,
ps a >/tmp/ps$$
...
rm /tmp/ps$$
$!
The process number of the last process run in the background (in decimal).
$-
The current shell flags, such as -x and -v.

Some variables have a special meaning to the shell and should be avoided for general use.

$MAIL
When used interactively the shell looks at the file specified by this variable before it issues a prompt. If the specified file has been modified since it was last looked at the shell prints the message you have mail before prompting for the next command. This variable is typically set in the file .profile, in the user's login directory. For example,
MAIL=/usr/mail/fred
$HOME
The default argument for the cd command. The current directory is used to resolve file name references that do not begin with a /, and is changed using the cd command. For example,
cd /usr/fred/bin
makes the current directory /usr/fred/bin.
cat wn
will print on the terminal the file wn in this directory. The command cd with no argument is equivalent to
cd $HOME
This variable is also typically set in the the user's login profile.
$PATH
A list of directories that contain commands (the search path). Each time a command is executed by the shell a list of directories is searched for an executable file. If $PATH is not set then the current directory, /bin, and /usr/bin are searched by default. Otherwise $PATH consists of directory names separated by :. For example,
PATH=:/usr/fred/bin:/bin:/usr/bin
specifies that the current directory (the null string before the first :), /usr/fred/bin, /bin and /usr/bin are to be searched in that order. In this way individual users can have their own `private' commands that are accessible independently of the current directory. If the command name contains a / then this directory search is not used; a single attempt is made to execute the command.
$PS1
The primary shell prompt string, by default, `$ '.
$PS2
The shell prompt when further input is needed, by default, `> '.
$IFS
The set of characters used by blank interpretation (see section 3.4).

2.5 The test command

The test command, although not part of the shell, is intended for use by shell programs. For example,

test -f file
returns zero exit status if file exists and non-zero exit status otherwise. In general test evaluates a predicate and returns the result as its exit status. Some of the more frequently used test arguments are given here, see test (1) for a complete specification.
test s true if the argument s is not the null string
test -f file true if file exists
test -r file true if file is readable
test -w file true if file is writable
test -d file true if file is a directory

2.6 Control flow - while

The actions of the for loop and the case branch are determined by data available to the shell. A while or until loop and an if then else branch are also provided whose actions are determined by the exit status returned by commands. A while loop has the general form

while command-list1
do command-list2
done

The value tested by the while command is the exit status of the last simple command following while. Each time round the loop command-list1 is executed; if a zero exit status is returned then command-list2 is executed; otherwise, the loop terminates. For example,

while test $1
do ...
shift
done
is equivalent to
for i
do ...
done
shift is a shell command that renames the positional parameters $2, $3, ... as $1, $2, ... and loses $1.

Another kind of use for the while/until loop is to wait until some external event occurs and then run some commands. In an until loop the termination condition is reversed. For example,

until test -f file
do sleep 300; done
commands
will loop until file exists. Each time round the loop it waits for 5 minutes before trying again. (Presumably another process will eventually create the file.)

2.7 Control flow - if

Also available is a general conditional branch of the form,

if command-list
then command-list
else command-list
fi
that tests the value returned by the last simple command following if.

The if command may be used in conjunction with the test command to test for the existence of a file as in

if test -f file
then process file
else do something else
fi

An example of the use of if, case and for constructions is given in section 2.10.

A multiple test if command of the form

if ...
then ...
else if ...
then ...
else if ...
...
fi
fi
fi
may be written using an extension of the if notation as,
if ...
then ...
elif ...
then ...
elif ...
...
fi

The following example is the touch command which changes the `last modified' time for a list of files. The command may be used in conjunction with make (1) to force recompilation of a list of files.

flag=
for i
do case $i in
-c) flag=N ;;
*) if test -f $i
then ln $i junk$$; rm junk$$
elif test $flag
then echo file \\'$i\\' does not exist
else >$i
fi
esac
done
The -c flag is used in this command to force subsequent files to be created if they do not already exist. Otherwise, if the file does not exist, an error message is printed. The shell variable flag is set to some non-null string if the -c argument is encountered. The commands
ln ...; rm ...
make a link to the file and then remove it thus causing the last modified date to be updated.

The sequence

if command1
then command2
fi
may be written
command1 && command2
Conversely,
command1 || command2
executes command2 only if command1 fails. In each case the value returned is that of the last simple command executed.

2.8 Command grouping

Commands may be grouped in two ways,

{ command-list ; }
and
( command-list )

In the first command-list is simply executed. The second form executes command-list as a separate process. For example,

(cd x; rm junk )
executes rm junk in the directory x without changing the current directory of the invoking shell.

The commands

cd x; rm junk
have the same effect but leave the invoking shell in the directory x.

2.9 Debugging shell procedures

The shell provides two tracing mechanisms to help when debugging shell procedures. The first is invoked within the procedure as

set -v
(v for verbose) and causes lines of the procedure to be printed as they are read. It is useful to help isolate syntax errors. It may be invoked without modifying the procedure by saying
sh -v proc ...
where proc is the name of the shell procedure. This flag may be used in conjunction with the -n flag which prevents execution of subsequent commands. (Note that saying set -n at a terminal will render the terminal useless until an end-of-file is typed.)

The command

set -x
will produce an execution trace. Following parameter substitution each command is printed as it is executed. (Try these at the terminal to see what effect they have.) Both flags may be turned off by saying
set -
and the current setting of the shell flags is available as $-.

2.10 The man command

The following is the man command which is used to print sections of the UNIX manual. It is called, for example, as

man sh
man -t ed
man 2 fork
In the first the manual section for sh is printed. Since no section is specified, section 1 is used. The second example will typeset (-t option) the manual section for ed. The last prints the fork manual page from section 2.

cd /usr/man

: 'colon is the comment command'
: 'default is nroff ($N), section 1 ($s)'
N=n s=1

for i
do case $i in
[1-9]*) s=$i ;;
-t) N=t ;;
-n) N=n ;;
-*) echo unknown flag \\'$i\\' ;;
*) if test -f man$s/$i.$s
then ${N}roff man0/${N}aa man$s/$i.$s
else : 'look through all manual sections'
found=no
for j in 1 2 3 4 5 6 7 8 9
do if test -f man$j/$i.$j
then man $j $i
found=yes
fi
done
case $found in
no) echo \'$i: manual page not found\'
esac
fi
esac
done
Figure 1. A version of the man command

3.0 Keyword parameters

Shell variables may be given values by assignment or when a shell procedure is invoked. An argument to a shell procedure of the form name=value that precedes the command name causes value to be assigned to name before execution of the procedure begins. The value of name in the invoking shell is not affected. For example,

user=fred command
will execute command with user set to fred. The -k flag causes arguments of the form name=value to be interpreted in this way anywhere in the argument list. Such names are sometimes called keyword parameters. If any arguments remain they are available as positional parameters $1, $2, ....

The set command may also be used to set positional parameters from within a procedure. For example,

set - *
will set $1 to the first file name in the current directory, $2 to the next, and so on. Note that the first argument, -, ensures correct treatment when the first file name begins with a -.

3.1 Parameter transmission

When a shell procedure is invoked both positional and keyword parameters may be supplied with the call. Keyword parameters are also made available implicitly to a shell procedure by specifying in advance that such parameters are to be exported. For example,

export user box
marks the variables user and box for export. When a shell procedure is invoked copies are made of all exportable variables for use within the invoked procedure. Modification of such variables within the procedure does not affect the values in the invoking shell. It is generally true of a shell procedure that it may not modify the state of its caller without explicit request on the part of the caller. (Shared file descriptors are an exception to this rule.)

Names whose value is intended to remain constant may be declared readonly. The form of this command is the same as that of the export command,

readonly name ...
Subsequent attempts to set readonly variables are illegal.

3.2 Parameter substitution

If a shell parameter is not set then the null string is substituted for it. For example, if the variable d is not set

echo $d
or
echo ${d}
will echo nothing. A default string may be given as in
echo ${d-.}
which will echo the value of the variable d if it is set and `.' otherwise. The default string is evaluated using the usual quoting conventions so that
echo ${d-'*'}
will echo * if the variable d is not set. Similarly
echo ${d-$1}
will echo the value of d if it is set and the value (if any) of $1 otherwise. A variable may be assigned a default value using the notation
echo ${d=.}
which substitutes the same string as
echo ${d-.}
and if d were not previously set then it will be set to the string `.'. (The notation ${...=...} is not available for positional parameters.)

If there is no sensible default then the notation

echo ${d?message}
will echo the value of the variable d if it has one, otherwise message is printed by the shell and execution of the shell procedure is abandoned. If message is absent then a standard message is printed. A shell procedure that requires some parameters to be set might start as follows.
: ${user?} ${acct?} ${bin?}
...
Colon (:) is a command that is built in to the shell and does nothing once its arguments have been evaluated. If any of the variables user, acct or bin are not set then the shell will abandon execution of the procedure.

3.3 Command substitution

The standard output from a command can be substituted in a similar way to parameters. The command pwd prints on its standard output the name of the current directory. For example, if the current directory is /usr/fred/bin then the command

d=`pwd`
is equivalent to
d=/usr/fred/bin

The entire string between grave accents (`...`) is taken as the command to be executed and is replaced with the output from the command. The command is written using the usual quoting conventions except that a ` must be escaped using a \. For example,

ls `echo "$1"`
is equivalent to
ls $1
Command substitution occurs in all contexts where parameter substitution occurs (including here documents) and the treatment of the resulting text is the same in both cases. This mechanism allows string processing commands to be used within shell procedures. An example of such a command is basename which removes a specified suffix from a string. For example,
basename main.c .c
will print the string main. Its use is illustrated by the following fragment from a cc command.
case $A in
...
*.c) B=`basename $A .c`
...
esac
that sets B to the part of $A with the suffix .c stripped.

Here are some composite examples.

  • for i in `ls -t`; do ...
    The variable i is set to the names of files in time order, most recent first.
  • set `date`; echo $6 $2 $3, $4
    will print, e.g., 1977 Nov 1, 23:59:59

3.4 Evaluation and quoting

The shell is a macro processor that provides parameter substitution, command substitution and file name generation for the arguments to commands. This section discusses the order in which these evaluations occur and the effects of the various quoting mechanisms.

Commands are parsed initially according to the grammar given in appendix A. Before a command is executed the following substitutions occur.

  • parameter substitution, e.g. $user
  • command substitution, e.g. `pwd`

Only one evaluation occurs so that if, for example, the value of the variable X is the string $y then

echo $X
will echo $y.
  • blank interpretation

Following the above substitutions the resulting characters are broken into non-blank words (blank interpretation). For this purpose `blanks' are the characters of the string $IFS. By default, this string consists of blank, tab and newline. The null string is not regarded as a word unless it is quoted. For example,

echo ''
will pass on the null string as the first argument to echo, whereas
echo $null
will call echo with no arguments if the variable null is not set or set to the null string.
  • file name generation

Each word is then scanned for the file pattern characters *, ? and [...] and an alphabetical list of file names is generated to replace the word. Each such file name is a separate argument.

The evaluations just described also occur in the list of words associated with a for loop. Only substitution occurs in the word used for a case branch.

As well as the quoting mechanisms described earlier using \ and '...' a third quoting mechanism is provided using double quotes. Within double quotes parameter and command substitution occurs but file name generation and the interpretation of blanks does not. The following characters have a special meaning within double quotes and may be quoted using \.

$ parameter substitution
` command substitution
" ends the quoted string
\ quotes the special characters $ ` " \
For example,
echo "$x"
will pass the value of the variable x as a single argument to echo. Similarly,
echo "$*"
will pass the positional parameters as a single argument and is equivalent to
echo "$1 $2 ..."
The notation $@ is the same as $* except when it is quoted.
echo "$@"
will pass the positional parameters, unevaluated, to echo and is equivalent to
echo "$1" "$2" ...

The following table gives, for each quoting mechanism, the shell metacharacters that are evaluated.

metacharacter \ $ * ` " ' ' n n n n n t ` y n n t n n " y y n y t n t terminator y interpreted n not interpreted
Figure 2. Quoting mechanisms

In cases where more than one evaluation of a string is required the built-in command eval may be used. For example, if the variable X has the value $y, and if y has the value pqr then

eval echo $X
will echo the string pqr.

In general the eval command evaluates its arguments (as do all commands) and treats the result as input to the shell. The input is read and the resulting command(s) executed. For example,

wg=\'eval who|grep\'
$wg fred
is equivalent to
who|grep fred
In this example, eval is required since there is no interpretation of metacharacters, such as |, following substitution.

3.5 Error handling

The treatment of errors detected by the shell depends on the type of error and on whether the shell is being used interactively. An interactive shell is one whose input and output are connected to a terminal (as determined by gtty (2)). A shell invoked with the -i flag is also interactive.

Execution of a command (see also 3.7) may fail for any of the following reasons.

In all of these cases the shell will go on to execute the next command. Except for the last case an error message will be printed by the shell. All remaining errors cause the shell to exit from a command procedure. An interactive shell will return to read another command from the terminal. Such errors include the following.

The shell flag -e causes the shell to terminate if any error is detected.

1 hangup
2 interrupt
3* quit
4* illegal instruction
5* trace trap
6* IOT instruction
7* EMT instruction
8* floating point exception
9 kill (cannot be caught or ignored)
10* bus error
11* segmentation violation
12* bad argument to system call
13 write on a pipe with no one to read it
14 alarm clock
15 software termination (from kill (1))

Figure 3. UNIX signals
Those signals marked with an asterisk produce a core dump if not caught. However, the shell itself ignores quit which is the only external signal that can cause a dump. The signals in this list of potential interest to shell programs are 1, 2, 3, 14 and 15.

3.6 Fault handling

Shell procedures normally terminate when an interrupt is received from the terminal. The trap command is used if some cleaning up is required, such as removing temporary files. For example,

trap 'rm /tmp/ps$$; exit' 2
sets a trap for signal 2 (terminal interrupt), and if this signal is received will execute the commands
rm /tmp/ps$$; exit
exit is another built-in command that terminates execution of a shell procedure. The exit is required; otherwise, after the trap has been taken, the shell will resume executing the procedure at the place where it was interrupted.

UNIX signals can be handled in one of three ways. They can be ignored, in which case the signal is never sent to the process. They can be caught, in which case the process must decide what action to take when the signal is received. Lastly, they can be left to cause termination of the process without it having to take any further action. If a signal is being ignored on entry to the shell procedure, for example, by invoking it in the background (see 3.7) then trap commands (and the signal) are ignored.

The use of trap is illustrated by this modified version of the touch command (Figure 4). The cleanup action is to remove the file junk$$.

flag=
trap 'rm -f junk$$; exit' 1 2 3 15
for i
do case $i in
-c) flag=N ;;
*) if test -f $i
then ln $i junk$$; rm junk$$
elif test $flag
then echo file \\'$i\\' does not exist
else >$i
fi
esac
done

Figure 4. The touch command

The trap command appears before the creation of the temporary file; otherwise it would be possible for the process to die without removing the file.

Since there is no signal 0 in UNIX it is used by the shell to indicate the commands to be executed on exit from the shell procedure.

A procedure may, itself, elect to ignore signals by specifying the null string as the argument to trap. The following fragment is taken from the nohup command.

trap '' 1 2 3 15
which causes hangup, interrupt, quit and kill to be ignored both by the procedure and by invoked commands.

Traps may be reset by saying

trap 2 3
which resets the traps for signals 2 and 3 to their default values. A list of the current values of traps may be obtained by writing
trap

The procedure scan (Figure 5) is an example of the use of trap where there is no exit in the trap command. scan takes each directory in the current directory, prompts with its name, and then executes commands typed at the terminal until an end of file or an interrupt is received. Interrupts are ignored while executing the requested commands but cause termination when scan is waiting for input.

d=`pwd`
for i in *
do if test -d $d/$i
then cd $d/$i
while echo "$i:"
trap exit 2
read x
do trap : 2; eval $x; done
fi
done

Figure 5. The scan command

read x is a built-in command that reads one line from the standard input and places the result in the variable x. It returns a non-zero exit status if either an end-of-file is read or an interrupt is received.

3.7 Command execution

To run a command (other than a built-in) the shell first creates a new process using the system call fork. The execution environment for the command includes input, output and the states of signals, and is established in the child process before the command is executed. The built-in command exec is used in the rare cases when no fork is required and simply replaces the shell with a new command. For example, a simple version of the nohup command looks like

trap \'\' 1 2 3 15
exec $*
The trap turns off the signals specified so that they are ignored by subsequently created commands and exec replaces the shell by the command specified.

Most forms of input output redirection have already been described. In the following word is only subject to parameter and command substitution. No file name generation or blank interpretation takes place so that, for example,

echo ... >*.c
will write its output into a file whose name is *.c. Input output specifications are evaluated left to right as they appear in the command.
> word
The standard output (file descriptor 1) is sent to the file word which is created if it does not already exist.
>> word
The standard output is sent to file word. If the file exists then output is appended (by seeking to the end); otherwise the file is created.
< word
The standard input (file descriptor 0) is taken from the file word.
<< word
The standard input is taken from the lines of shell input that follow up to but not including a line consisting only of word. If word is quoted then no interpretation of the document occurs. If word is not quoted then parameter and command substitution occur and \ is used to quote the characters \ $ ` and the first character of word. In the latter case \newline is ignored (c.f. quoted strings).
>& digit
The file descriptor digit is duplicated using the system call dup (2) and the result is used as the standard output.
<& digit
The standard input is duplicated from file descriptor digit.
<&-
The standard input is closed.
>&-
The standard output is closed.

Any of the above may be preceded by a digit in which case the file descriptor created is that specified by the digit instead of the default 0 or 1. For example,

... 2>file
runs a command with message output (file descriptor 2) directed to file.
... 2>&1
runs a command with its standard output and message output merged. (Strictly speaking file descriptor 2 is created by duplicating file descriptor 1 but the effect is usually to merge the two streams.)

The environment for a command run in the background such as

list *.c | lpr &
is modified in two ways. Firstly, the default standard input for such a command is the empty file /dev/null. This prevents two processes (the shell and the command), which are running in parallel, from trying to read the same input. Chaos would ensue if this were not the case. For example,
ed file &
would allow both the editor and the shell to read from the same input at the same time.

The other modification to the environment of a background command is to turn off the QUIT and INTERRUPT signals so that they are ignored by the command. This allows these signals to be used at the terminal without causing background commands to terminate. For this reason the UNIX convention for a signal is that if it is set to 1 (ignored) then it is never changed even for a short time. Note that the shell command trap has no effect for an ignored signal.

3.8 Invoking the shell

The following flags are interpreted by the shell when it is invoked. If the first character of argument zero is a minus, then commands are read from the file .profile.

-c string

If the -c flag is present then commands are read from string.
-s
If the -s flag is present or if no arguments remain then commands are read from the standard input. Shell output is written to file descriptor 2.
-i
If the -i flag is present or if the shell input and output are attached to a terminal (as told by gtty) then this shell is interactive. In this case TERMINATE is ignored (so that kill 0 does not kill an interactive shell) and INTERRUPT is caught and ignored (so that wait is interruptable). In all cases QUIT is ignored by the shell.

Acknowledgements

The design of the shell is based in part on the original UNIX shell unix command language thompson and the PWB/UNIX shell, pwb shell mashey unix some features having been taken from both. Similarities also exist with the command interpreters of the Cambridge Multiple Access System cambridge multiple access system hartley and of CTSS. ctss

I would like to thank Dennis Ritchie and John Mashey for many discussions during the design of the shell. I am also grateful to the members of the Computing Science Research Center and to Joe Maranzano for their comments on drafts of this document.

$LIST$

Appendix A - Grammar

item: word
input-output
name = value

simple-command: item
simple-command item

command: simple-command
( command-list )
{ command-list }
for name do command-list done
for name in word ... do command-list done
while command-list do command-list done
until command-list do command-list done
case word in case-part ... esac
if command-list then command-list else-part fi

pipeline: command
pipeline
| command

andor: pipeline
andor
&& pipeline
andor
|| pipeline

command-list: andor
command-list
;
command-list
&
command-list
; andor
command-list
& andor

input-output:
> file
< file
>> word
<< word

file: word
& digit
& -

case-part: pattern
) command-list ;;

pattern: word
pattern
| word

else-part: elif command-list then command-list else-part
else command-list
empty

empty:

word:
a sequence of non-blank characters

name:
a sequence of letters, digits or underscores starting with a letter

digit:
0 1 2 3 4 5 6 7 8 9


Appendix B - Meta-characters and Reserved Words

a) syntactic

|
pipe symbol
&&
`andf' symbol
||
`orf' symbol
;
command separator
;;
case delimiter
&
background commands
( )
command grouping
<
input redirection
<<
input from a here document
>
output creation
>>
output append

b) patterns

*
match any character(s) including none
?
match any single character
[...]
match any of the enclosed characters

c) substitution

${...}
substitute shell variable
`...`
substitute command output

d) quoting

\
quote the next character
'...'
quote the enclosed characters except for '
"..."
quote the enclosed characters except for $ ` \ "

e) reserved words

if then else elif fi
case in esac
for while until do done
{ }