How to terminate script's process tree in Cygwin bash from bash script

Question

I have a Cygwin bash script that I need to watch and terminate under certain conditions - specifically, after a certain file has been created. I'm having difficulty figuring out how exactly to terminate the script with the same level of completeness that Ctrl+C does, however.

Here's a simple script (called test1) that does little more than wait around to be terminated.

#!/bin/bash

test -f kill_me && rm kill_me

touch kill_me
tail -f kill_me

If this script is run in the foreground, Ctrl+C will terminate both the tail and the script itself. If the script is run in the background, a kill %1 (assuming it is job 1) will also terminate both tail and the script.

However, when I try to do the same thing from a script, I'm finding that only the bash process running the script is terminated, while tail hangs around disconnected from its parent. Here's one way I tried (test2):

#!/bin/bash

test -f kill_me && rm kill_me

(
    touch kill_me
    tail -f kill_me
) &

while true; do
    sleep 1
    test -f kill_me && {
        kill %1
        exit
    }
done

If this is run, the bash subshell running in the background is terminated OK, but tail still hangs around.

If I use an explicitly separate script, like this, it still doesn't work (test3):

#!/bin/bash

test -f kill_me && rm kill_me

# assuming test1 above is included in the same directory
./test1 &

while true; do
    sleep 1
    test -f kill_me && {
        kill %1
        exit
    }
done

tail is still hanging around after this script is run.

In my actual case, the process creating files is not particularly instrumentable, so I can't get it to terminate of its own accord; by finding out when it has created a particular file, however, I can at that point know that it's OK to terminate it. Unfortunately, I can't use a simple killall or equivalent, as there may be multiple instances running, and I only want to kill the specific instance.

Answer 1

/bin/kill (the program, not the bash builtin) interprets a negative PID as “kill the process group” which will get all the children too.

Changing

kill %1

to

/bin/kill -- -$$

works for me.

Answer 2

Adam's link put me in a direction that will solve the problem, albeit not without some minor caveats.

The script doesn't work unmodified under Cygwin, so I rewrote it, and with a couple more options. Here's my version:

#!/bin/bash

function usage
{
    echo "usage: $(basename $0) [-c] [-<sigspec>] <pid>..."
    echo "Recursively kill the process tree(s) rooted by <pid>."
    echo "Options:"
    echo "  -c        Only kill children; don't kill root"
    echo "  <sigspec> Arbitrary argument to pass to kill, expected to be signal specification"
    exit 1
}

kill_parent=1
sig_spec=-9

function do_kill # <pid>...
{
    kill "$sig_spec" "$@"
}

function kill_children # pid
{
    local target=$1
    local pid=
    local ppid=
    local i
    # Returns alternating ids: first is pid, second is parent
    for i in $(ps -f | tail +2 | cut -b 10-24); do
        if [ ! -n "$pid" ]; then
            # first in pair
            pid=$i
        else
            # second in pair
            ppid=$i
            (( ppid == target && pid != $$ )) && {
                kill_children $pid
                do_kill $pid
            }
            # reset pid for next pair
            pid=
        fi
    done

}

test -n "$1" || usage

while [ -n "$1" ]; do
    case "$1" in
        -c)
            kill_parent=0
            ;;

        -*)
            sig_spec="$1"
            ;;

        *)
            kill_children $1
            (( kill_parent )) && do_kill $1
            ;;
    esac
    shift
done

The only real downside is the somewhat ugly message that bash prints out when it receives a fatal signal, namely "Terminated", "Killed" or "Interrupted" (depending on what you send). However, I can live with that in batch scripts.