Correct behavior of EXIT and ERR traps when using `set -eu`

From man bash:

• set -u
  • Treat unset variables and parameters other than the special parameters "@" and "*" as an error when performing parameter expansion. If expansion is attempted on an unset variable or parameter, the shell prints an error message, and, if not -interactive, exits with a nonzero status.

POSIX states that, in the event of an expansion error, a non-interactive shell shall exit when the expansion is associated with either a shell special builtin (which is a distinction bash regularly ignores anyway, and so maybe is irrelevant) or any other utility besides.

• Consequences of Shell Errors:
  • An expansion error is one that occurs when the shell expansions defined in Word Expansions are carried out (for example, "${x!y}", because ! is not a valid operator); an implementation may treat these as syntax errors if it is able to detect them during tokenization, rather than during expansion.
  • [A]n interactive shell shall write a diagnostic message to standard error without exiting.

Also from man bash:

• trap ... ERR
  • If a sigspec is ERR, the command arg is executed whenever a pipeline (which may consist of a single simple command), a list, or a compound command returns a non-zero exit status, subject to the following conditions:
    • The ERR trap is not executed if the failed command is part of the command list immediately following a while or until keyword...
    • ...part of the test in an if statement...
    • ...part of a command executed in a && or || list except the command following the final && or ||...
    • ...any command in a pipeline but the last...
    • ...or if the command's return value is being inverted using !.
  • These are the same conditions obeyed by the errexit -e option.

Note above that the ERR trap is all about the evaluation of some other command's return. But when an expansion error occurs, there is no command run to return anything. In your example, echo never happens - because while the shell evaluates and expands its arguments it encounters an -unset variable, which has been specified by explicit shell option to cause an immediate exit from the current, scripted shell.

And so the EXIT trap, if any, is executed, and the shell exits with a diagnostic message and exit status other than 0 - exactly as it should do.

As for the rc: 0 thing, I expect that is a version specific bug of some kind - probably to do with the two triggers for the EXIT occurring at the same time and the one getting the other's exit code (which should not occur). And anyway, with an up-to-date bash binary as installed by pacman:

bash <<\IN
    printf "shell options:\t$-\n"
    trap 'echo "EXIT (rc: $?)"' EXIT
    set -eu
    echo ${UNSET_VAR}
IN

I added the first line so you can see that the shell's conditions are those of a scripted shell - it is not interactive. The output is:

shell options:  hB
bash: line 4: UNSET_VAR: unbound variable
EXIT (rc: 1)

Here are some relevant notes from recent changelogs:

• Fixed a bug that caused asynchronous commands to not set $? correctly.
• Fixed a bug that caused error messages generated by expansion errors in for commands to have the wrong line number.
• Fixed a bug that caused SIGINT and SIGQUIT to not be trappable in asynchronous subshell commands.
• Fixed a problem with interrupt handling that caused a second and subsequent SIGINT to be ignored by interactive shells.
• The shell no longer blocks receipt of signals while running trap handlers for those signals, and allows most trap handlers to be run recursively (running trap handlers while a trap handler is executing).

I think it is either the last or the first that is most relevant - or possibly a combination of the two. A trap handler is by its very nature asynchronous because its whole job is to wait for and handle asynchronous signals. And you trigger two simultaneously with -eu and $UNSET_VAR.

And so maybe you should just update, but if you like yourself, you'll do it with a different shell altogether.

(I'm using bash 4.2.53). For part 1, the bash man page just says "An error message will be written to the standard error, and a non-interactive shell will exit". It doesn't say an ERR trap will be called, though I agree it would be useful if it did.

To be pragmatic, if what you really want is to cope more cleanly with undefined variables, a possible solution is to put most of your code inside a function, then execute that function in a sub-shell and recover the return code and stderr output. Here's an example where "cmd()" is the function:

#!/bin/bash
trap 'rc=$?; echo "ERR at line ${LINENO} (rc: $rc)"; exit $rc' ERR
trap 'rc=$?; echo "EXIT (rc: $rc)"; exit $rc' EXIT
set -u
set -E # export trap to functions

cmd(){
 echo "args=$*"
 echo ${UNSET_VAR}
 echo hello
}
oops(){
 rc=$?
 echo "$@"
 return $rc # provoke ERR trap
}

exec 3>&1 # copy stdin to use in $()
if output=$(cmd "$@" 2>&1 >&3) # collect stderr, not stdout 
then    echo ok
else    oops "fail: $output"
fi

On my bash I get

./script my stuff; echo "exit was $?"
args=my stuff
fail: ./script: line 9: UNSET_VAR: unbound variable
ERR at line 15 (rc: 1)
EXIT (rc: 1)
exit was 1