The EXIT trap, when set, runs its code when the function or script within which it was set exits. Here’s a simple example:

function func {
    trap 'print "exiting from the function"' EXIT
    print 'start of the function'
}

trap 'print "exiting from the script"' EXIT
print 'start of the script'
func

If you run this script, you see this output:

start of the script
start of the function
exiting from the function
exiting from the script

In other words, the script starts by setting the trap for its own exit. Then it prints a message and finally calls the function. The function does the same — sets a trap for its exit and prints a message. (Remember that function-style functions can have their own local traps that supersede any traps set by the surrounding script, while POSIX functions share traps with the main script.)

The function then exits, which causes the shell to send it the fake signal EXIT, which in turn runs the code print "exiting from the function". Then the script exits, and its own EXIT trap code is run. Note also that traps “stack;” the EXIT fake signal is sent to each running function in turn as each more recently called function exits.

An EXIT trap occurs no matter how the script or function exits, whether normally (by finishing the last statement), by an explicit exit or return statement, or by receiving a “real” signal such as INT or TERM. Consider the following inane number-guessing program:

trap 'print "Thank you for playing!"' EXIT

magicnum=$(($RANDOM%10+1))
print 'Guess a number between 1 and 10:'
while read guess'?number> '; do
    sleep 10
    if (( $guess == $magicnum )); then
        print 'Right!'
        exit
    fi
    print 'Wrong!'
done

This program picks a number between 1 and 10 by getting a random number (via the built-in variable RANDOM, see Appendix B), extracting the last digit (the remainder when divided by 10), and adding 1. Then it prompts you for a guess, and after 10 seconds, it tells you if you guessed right.

If you did, the program exits with the message, “Thank you for playing!”, i.e., it runs the EXIT trap code. If you were wrong, it prompts you again and repeats the process until you get it right. If you get bored with this little game and hit CTRL-C while waiting for it to tell you whether you were right, you also see the message.

The fake signal ERR enables you to run code whenever a command in the surrounding script or function exits with non-zero status. Trap code for ERR can take advantage of the built-in variable ?, which holds the exit status of the previous command. It survives the trap and is accessible at the beginning of the trap-handling code.

A simple but effective use of this is to put the following code into a script you want to debug:

function errtrap {
    typeset es=$?
    print "ERROR: Command exited with status $es."
}

trap errtrap ERR

The first line saves the nonzero exit status in the local variable es.

For example, if the shell can’t find a command, it returns status 1. If you put the code in a script with a line of gibberish (like “lskdjfafd”), the shell responds with:

                     scriptname: line N: lskdjfafd:  not found
ERROR: command exited with status 1.

N is the number of the line in the script that contains the bad command. In this case, the shell prints the line number as part of its own error-reporting mechanism, since the error was a command that the shell could not find. But if the nonzero exit status comes from another program, the shell doesn’t report the line number. For example:

function errtrap {
    typeset es=$?
    print "ERROR: Command exited with status $es."
}

trap errtrap ERR

function bad {
    return 17
}

bad

This only prints ERROR: Command exited with status 17.

It would obviously be an improvement to include the line number in this error message. The built-in variable LINENO exists, but if you use it inside a function, it evaluates to the line number in the function, not in the overall file. In other words, if you used $LINENO in the print statement in the errtrap routine, it would always evaluate to 2.

To get around this problem, we simply pass $LINENO as an argument to the trap handler, surrounding it in single quotes so that it doesn’t get evaluated until the fake signal actually comes in:

function errtrap {
    typeset es=$?
    print "ERROR line $1: Command exited with status $es."
}
trap 'errtrap $LINENO' ERR
...

If you use this with the above example, the result is the message, ERROR line 12: Command exited with status 17. This is much more useful. We’ll see a variation on this technique shortly.

This simple code is actually not a bad all-purpose debugging mechanism. It takes into account that a nonzero exit status does not necessarily indicate an undesirable condition or event: remember that every control construct with a conditional (if, while, etc.) uses a nonzero exit status to mean “false.” Accordingly, the shell doesn’t generate ERR traps when statements or expressions in the “condition” parts of control structures produce nonzero exit statuses.

But a disadvantage is that exit statuses are not as uniform (or even as meaningful) as they should be, as we explained in Chapter 5. A particular exit status need not say anything about the nature of the error or even that there was an error.

The final debugging-related fake signal, DEBUG, causes the trap code to be run before every statement in the surrounding function or script.^[116] This has two possible uses. First is the use for humans, as a sort of a “brute force” method of tracking a certain element of a program’s state that you notice is going awry.

For example, you notice that the value of a particular variable is running amok. The naive approach would be to put in lots of print statements to check the variable’s value at several points. The DEBUG trap makes this easier:

function dbgtrap {
    print "badvar is $badvar"
}

trap dbgtrap DEBUG

... Section of code in which problem occurs ...

trap - DEBUG            # turn off DEBUG trap

This code prints the value of the wayward variable before every statement between the two traps.

The second and far more important use of the DEBUG trap is as a primitive for implementing Korn shell debuggers. In fact, it would be fair to say that the DEBUG trap reduces the task of implementing a useful shell debugger from a large-scale software development project to a manageable exercise. We will get to this shortly.

It is possible for multiple signals to arrive simultaneously (or close to it). In that case, the shell runs the trap commands in the following order:

Therefore the code has two parts: the part that implements the debugger’s functionality, and the part that installs that functionality into the script being debugged. The second part, which we’ll see first, is the script called kshdb. It’s very simple:

# kshdb -- Korn Shell debugger
# Main driver: constructs full script (with preamble) and runs it

print "Korn Shell Debugger version 2.0 for ksh '${.sh.version}'" >&2
_guineapig=$1
if [[ ! -r $1 ]]; then      # file not found or readable
    print "Cannot read $_guineapig." >&2
    exit 1
fi
shift

_tmpdir=/tmp
_libdir=.                   # set to real directory upon installation
_dbgfile=$_tmpdir/kshdb$$   # temp file for script being debugged (copy)
cat $_libdir/kshdb.pre $_guineapig > $_dbgfile
exec ksh $_dbgfile $_guineapig $_tmpdir $_libdir "$@"

kshdb takes as argument the name of the script being debugged, which, for the sake of brevity, we’ll call the guinea pig. Any additional arguments are passed to the guinea pig as its positional parameters. Notice that ${.sh.version} indicates the version of the Korn shell for the startup message.

If the argument is invalid (the file isn’t readable), kshdb exits with an error status. Otherwise, after an introductory message, it constructs a temporary filename like we saw in Chapter 8. If you don’t have (or don’t have access to) /tmp on your system, you can substitute a different directory for _tmpdir.^[117] Also, make sure that _libdir is set to the directory where the kshdb.pre and kshdb.fns files (which we’ll see soon) reside. /usr/share/lib is a good choice if you have access to it.

The cat statement builds the temp file: it consists of a file that we’ll see soon called kshdb.pre, which contains the actual debugger code, followed immediately by a copy of the guinea pig. Therefore the temp file contains a shell script that has been turned into a debugger for itself.

The last line runs this script with exec, a statement that we haven’t seen yet. We’ve chosen to wait until now to introduce it because — as we think you’ll agree — it can be dangerous. exec takes its arguments as a command line and runs the command in place of the current program, in the same process. In other words, the shell running the above script will terminate immediately and be replaced by exec’s arguments. The situations in which you would want to use exec are few, far between, and quite arcane — though this is one of them.

In this case, exec just runs the newly constructed shell script, i.e., the guinea pig with its debugger, in another Korn shell. It passes the new script three arguments — the names of the original guinea pig ($_guineapig), the temp directory ($_tmpdir), and the directory where kshdb.pre and kshdb.fns are kept — followed by the user’s positional parameters, if any.

exec can also be used with just an I/O redirector; this causes the redirector to take effect for the remainder of the script or login session. For example, the line exec 2>errlog at the top of a script directs the shell’s own standard error to the file errlog for the entire script. This can also be used to move the input or output of a coprocess to a regular numbered file descriptor. For example, exec 5<&p moves the coprocess’s output (which is input to the shell) to file descriptor 5. Similarly, exec 6>&p moves the coprocess’s input (which is output from the shell) to file descriptor 6. The predefined alias redirect='command exec' is more mnemonic.

We’ll explain shortly how _steptrap determines these things; now we’ll look at _cmdloop. It’s a typical command loop, resembling a combination of the case statements we saw in Chapter 5 and the calculator loop we saw in Chapter 8.

# Debugger command loop.
# Here at start of debugger session, when breakpoint reached,
# after single-step.  Optionally here inside watchpoint.
function _cmdloop {
    typeset cmd args

    while read -s cmd"?kshdb> " args; do
        case $cmd in
        #bp ) _setbp $args ;;       # set breakpoint at line num or string.
        #bc ) _setbc $args ;;       # set break condition.
        #cb ) _clearbp ;;           # clear all breakpoints.
        #g  ) return ;;             # start/resume execution
        #s  ) let _steps=${args:-1} # single-step N times (default 1)
               return ;;
        #wp ) _setwp $args ;;       # set a watchpoint
        #cw ) _clearwp $args ;;     # clear one or more watchpoints

        #x  ) _xtrace ;;            # toggle execution trace
        #? | #h ) _menu ;;        # print command menu
        #q  ) exit ;;               # quit
        #*  ) _msg "Invalid command: $cmd" ;;
        *  ) eval $cmd $args ;;      # otherwise, run shell command
        esac
    done

At each iteration, _cmdloop prints a prompt, reads a command, and processes it. We use read -s so that the user can take advantage of command-line editing within kshdb. All kshdb commands start with # to prevent confusion with shell commands. Anything that isn’t a kshdb command (and doesn’t start with #) is passed off to the shell for execution. Using # as the command character prevents a mistyped command from having any ill effect when the last case catches it and runs it through eval. Table 9-5 summarizes the debugger commands.

Before we look at the individual commands, it is important that you understand how control passes through _steptrap, the command loop, and the guinea pig.

_steptrap runs before every statement in the guinea pig as a result of the trap ... DEBUG statement in the preamble. If a breakpoint has been reached or the user previously typed in a step command (#s), _steptrap calls the command loop. In doing so, it effectively interrupts the shell that is running the guinea pig to hand control over to the user.^[118]

The user can invoke debugger commands as well as shell commands that run in the same shell as the guinea pig. This means that you can use shell commands to check values of variables, signal traps, and any other information local to the script being debugged.

The command loop runs, and the user stays in control, until the user types #g, #s, or #q. Let’s look in detail at what happens in each of these cases.

#g has the effect of running the guinea pig uninterrupted until it finishes or hits a breakpoint. But actually, it simply exits the command loop and returns to _steptrap, which exits as well. The shell takes control back; it runs the next statement in the guinea pig script and calls _steptrap again. Assuming that there is no breakpoint, this time _steptrap just exits again, and the process repeats until there is a breakpoint or the guinea pig is done.

When the user types #s, the command loop code sets the variable _steps to the number of steps the user wants to execute, i.e., to the argument given. Assume at first that the user omits the argument, meaning that _steps is set to 1. Then the command loop exits and returns control to _steptrap, which (as above) exits and hands control back to the shell. The shell runs the next statement and returns to _steptrap, which sees that _steps is 1 and decrements it to 0. Then the third elif conditional sees that _steps is 0, so it prints a “stopped” message and calls the command loop.

Now assume that the user supplies an argument to #s, say 3. _steps is set to 3. Then the following happens:

The overall effect is that three steps run and then the debugger takes over again.

Finally, the #q command exits. The EXIT trap then calls the function _cleanup, which just erases the temp file and exits the entire program.

All other debugger commands (#bp, #bc, #cb, #wp, #cw, #x, and shell commands) cause the shell to stay in the command loop, meaning that the user prolongs the interruption of the shell.

Now we’ll examine the breakpoint-related commands and the breakpoint mechanism in general. The #bp command calls the function _setbp, which can set two kinds of breakpoints, depending on the type of argument given. If it is a number, it’s treated as a line number; otherwise, it’s interpreted as a string that the breakpoint line should contain.

For example, the command #bp 15 sets a breakpoint at line 15, and #bp grep sets a breakpoint at the next line that contains the string grep — whatever number that turns out to be. Although you can always look at a numbered listing of a file,^[119] string arguments to #bp can make that unnecessary.

Here is the code for _setbp:

# Set breakpoint(s) at given line numbers or strings
# by appending patterns to breakpoint variables
function _setbp {
    if [[ -z $1 ]]; then
        _listbp
    elif [[ $1 == +([0-9]) ]]; then  # number, set bp at that line
        _linebp="${_linebp}$1|"
        _msg "Breakpoint at line " $1
    else                             # string, set bp at next line w/string
        _stringbp="${_stringbp}$@|"
        _msg "Breakpoint at next line containing '$@'."
    fi
}

_setbp sets the breakpoints by storing them in the variables _linebp (line number breakpoints) and _stringbp (string breakpoints). Both have breakpoints separated by pipe character delimiters, for reasons that will become clear shortly. This implies that breakpoints are cumulative; setting new breakpoints does not erase the old ones.

The only way to remove breakpoints is with the command #cb, which (in function _clearbp) clears all of them at once by simply resetting the two variables to null. If you don’t remember what breakpoints you have set, the command #bp without arguments lists them.

The functions _at_linenumbp and _at_stringbp are called by _steptrap after every statement; they check whether the shell has arrived at a line number or string breakpoint, respectively.

Here is _at_linenumbp:

# See if next line no. is a breakpoint.
function _at_linenumbp {
    [[ $_curline == @(${_linebp%|}) ]]
}

_at_linenumbp takes advantage of the pipe character as the separator between line numbers: it constructs a regular expression of the form @( N1 | N2 |...) by taking the list of line numbers _linebp, removing the trailing |, and surrounding it with @( and ). For example, if $_linebp is 3|15|19|, the resulting expression is @(3|15|19).

If the current line is any of these numbers, the conditional becomes true, and _at_linenumbp also returns a “true” (0) exit status.

The check for a string breakpoint works on the same principle, but it’s slightly more complicated; here is _at_stringbp:

# Search string breakpoints to see if next line in script matches.
function _at_stringbp {
    [[ -n $_stringbp && ${_lines[$_curline]} == *@(${_stringbp%|})* ]]
}

The conditional first checks if $_stringbp is non-null (meaning that string breakpoints have been defined). If not, the conditional evaluates to false, but if so, its value depends on the pattern match after the && — which tests the current line to see if it contains any of the breakpoint strings.

The expression on the right side of the double equal sign is similar to the one in _at_linenumbp above, except that it has * before and after it. This gives expressions of the form *@( S1 | S2 |...)*, where the Ss are the string breakpoints. This expression matches any line that contains any one of the possibilities in the parentheses.

The left side of the double equal sign is the text of the current line in the guinea pig. So, if this text matches the regular expression, we’ve reached a string breakpoint; accordingly, the conditional expression and _at_stringbp return exit status 0.

_steptrap tests each condition separately, so that it can tell you which kind of breakpoint stopped execution. In both cases, it calls the main command loop.

kshdb has another feature related to breakpoints: the break condition. This is a string that the user can specify that is evaluated as a command; if it is true (i.e., returns exit status 0), the debugger enters the command loop. Since the break condition can be any line of shell code, there’s lots of flexibility in what can be tested. For example, you can break when a variable reaches a certain value (e.g., (( $x < 0 ))) or when a particular piece of text has been written to a file (grep string file). You will probably think of all kinds of uses for this feature.^[120] To set a break condition, type #bc string. To remove it, type #bc without arguments — this installs the null string, which is ignored. _steptrap evaluates the break condition $_brcond only if it’s non-null. If the break condition evaluates to 0, the if clause is true and, once again, _steptrap calls the command loop.

The next feature is execution tracing, available through the #x command. This feature is meant to overcome the fact that a kshdb user can’t use set -o xtrace while debugging (by entering it as a shell command), because its scope is limited to the _cmdloop function.^[121]

The function _xtrace toggles execution tracing by simply assigning to the variable _trace the logical “not” of its current value, so that it alternates between 0 (off) and 1 (on). The preamble initializes it to 0.

kshdb takes advantage of the shell’s discipline functions to provide watchpoints. You can set a watchpoint on any variable when the variable’s value is retrieved or changed, or when the variable is unset. Optionally, the watchpoint can be set up to drop into the command loop as well. You do this with the #wp command, which in turn calls _setwp:

# Set a watchpoint on a variable
# usage: _setwp [-c] var discipline
# $1 = variable
# $2 = get|set|unset
typeset -A _watchpoints
function _setwp {
    typeset funcdef do_cmdloop=0
    if [[ $1 == -c ]]; then
        do_cmdloop=1
        shift
    fi

    funcdef="function $1.$2 { "

    case $2 in
    get)    funcdef+="_msg $1 ($$1) retrieved, line $_curline"
            ;;
    set)    funcdef+="_msg $1 set to "'${.sh.value}'", line $_curline"
            ;;
    unset)  funcdef+="_msg $1 cleared at line $_curline"
            funcdef+=$'
unset '"$1"
            ;;
    *)      _msg invalid watchpoint function $2
            return 1
            ;;
    esac

    if ((do_cmdloop)); then
        funcdef+=$'
_cmdloop'
    fi
    funcdef+=$'
}'

    eval "$funcdef"

    _watchpoints[$1.$2]=1
}

This function illustrates several interesting techniques. The first thing it does is declare some local variables and check if it was invoked with the -c option. This indicates that the watchpoint should enter the command loop.

The general idea is to build up the text of the appropriate discipline function in the variable funcdef. The initial value is the function keyword, the discipline function name, and the opening left curly brace. The space following the brace is important, so that the shell will correctly recognize it as a keyword.

Then, for each kind of discipline function, the case construct appends the appropriate function body to the funcdef string. The code uses judiciously placed backslashes to get the correct mixture of immediate and delayed shell variable evaluation. Consider the get case: for the (, the backslash stays intact for use as a quoting character inside the body of the discipline function. For $$1, the quoting happens as follows: the $ becomes a $ inside the function, while the $1 is evaluated immediately inside the double quoted string.

In the case that the -c option was supplied, it uses the $'...' notation to append a newline and a call to _cmdloop to the function body, and then at the end appends another newline and closing right brace. Finally, by using eval, it installs the newly created function.

For example, if -c was used, the text of the generated get function for the variable count ends up looking like this:

function count.get { _msg count ($count) retrieved, line $_curline
_cmdloop
}

At the end of _setwp, _watchpoints[$1.$2] is set to 1. This creates an entry in the associative array _watchpoints indexed by discipline function name. This conveniently stores the names of all watchpoints for when we want to clear them.

Watchpoints are cleared with the #cw command, which in turn runs the _clearwp function. Here it is:

# Clear watchpoints:
# no args: clear all
# two args: same as for setting: var get|set|unset
function _clearwp {
    if [ $# = 0 ]; then
        typeset _i
        for _i in ${!_watchpoints[*]}; do
            unset -f $_i
            unset _watchpoints[$_i]
        done
    elif [ $# = 2 ]; then
        case $2 in
        get | set | unset)
            unset -f $1.$2
            unset _watchpoints[$1.$2]
            ;;
        *)  _msg $2: invalid watchpoint
            ;;
        esac
    fi
}

When invoked with no arguments, _clearwp clears all the watchpoints, by looping over all the subscripts in the _watchpoints associative array. Otherwise, if invoked with two arguments, the variable name and discipline function, it unsets the function using unset -f. In either case, the entry in _watchpoints is also unset.

kshdb was not designed to push the state of the debugger art forward or to have an overabundance of features. It has the most useful basic features; its implementation is compact and (we hope) comprehensible. But it does have some important limitations. The ones we know of are described in the list that follows:

Many of these are not insurmountable; see the exercises.