On Monday 12 September 2011, Stefano Lattarini wrote:
> Now some updates: the "default" Korn Shell on Debian GNU/Linux (package
> `ksh', version `93u-1') seems to exhibit the same issue:
>
> $ ksh -c "perl -e 'kill 2, \$\$'; :"; echo $?
> 130
>
> And if I'm not reading the strace output wrong, this is due to the fact
> that the Debian korn shell is apparently killing itself (yikes!) with the
> same signal that killed the child process:
>
> $ strace ksh -c "perl -e 'kill 2, \$\$'; :"
> ...
> getpid() = 15108
> ...
> lstat64("/usr/bin/perl", {st_mode=S_IFREG|0755, st_size=1437872, ...}) = 0
> rt_sigprocmask(SIG_BLOCK, [HUP INT QUIT PIPE], [], 8) = 0
> vfork() = 15109
> --- SIGCHLD (Child exited) @ 0 (0) ---
> sigreturn() = ? (mask now [HUP INT QUIT PIPE])
> rt_sigprocmask(SIG_SETMASK, [], NULL, 8) = 0
> waitpid(-1, [{WIFSIGNALED(s) && WTERMSIG(s) == SIGINT}],
> WNOHANG|WSTOPPED|WCONTINUED) = 15109
> waitpid(-1, 0xbfa0270c, WNOHANG|WSTOPPED|WCONTINUED) = -1 ECHILD (No child
> processes)
> rt_sigaction(SIGCHLD, {0x806e390, [], SA_INTERRUPT}, {0x806e390, [],
> SA_INTERRUPT}, 8) = 0
> rt_sigaction(SIGINT, {0x80603b0, [], SA_INTERRUPT}, {SIG_DFL, [], 0}, 8) = 0
> ioctl(2, TIOCGPGRP, [15107]) = 0
> rt_sigaction(SIGINT, {0x80603b0, [], SA_INTERRUPT}, {0x80603b0, [],
> SA_INTERRUPT}, 8) = 0
> getrlimit(RLIMIT_CORE, {rlim_cur=0, rlim_max=RLIM_INFINITY}) = 0
> setrlimit(RLIMIT_CORE, {rlim_cur=0, rlim_max=RLIM_INFINITY}) = 0
> rt_sigaction(SIGINT, {SIG_DFL, [], SA_INTERRUPT}, {0x80603b0, [],
> SA_INTERRUPT}, 8) = 0
> rt_sigprocmask(SIG_UNBLOCK, [INT], NULL, 8) = 0
> rt_sigprocmask(SIG_UNBLOCK, [INT], NULL, 8) = 0
> kill(15108, SIGINT) = 0
> --- SIGINT (Interrupt) @ 0 (0) ---
> +++ killed by SIGINT +++
>
> The truss output on Solaris is less clear (for me at least, I'm not used
> to it), and I'm not at all sure that the issue there is due to the same
> dynamics; more digging needed.
>
No, it's better on Solaris: its /bin/ksh (and /usr/xpg4/bin/sh) "only" exit
with status 130 [*] if a child is terminated by a SIGINT; so the signal is
not "recursively propagated".
[*] Well, it still executes the trap associated with SIGINT, but probably
does so through special-cased code, without passing through real `kill'
system calls. I guess.
Examples:
$ cat bar1.sh
trap 'echo SIGINT; exit 99;' 2
/bin/ksh -c "perl -e 'kill 2, \$\$'; :"; echo $?
echo alive
$ /bin/ksh bar1.sh
130
alive
$ cat bar2.sh
/bin/ksh -c "/bin/ksh -c 'kill -2 \$\$'; echo \$?"; echo $?
echo alive
$ /bin/ksh bar2.sh
130
0
alive
$ cat bar3.sh
trap 'echo SIGINT; exit 99' 2
/bin/ksh -c "
trap 'echo SIGINT; exit 99;' 2
perl -e 'kill 2, \$\$'
:"
echo $?
echo alive
$ /bin/ksh bar3.sh
SIGINT
99
alive
OTOH, the first and second scripts would behave much worse on Debian:
$ /bin/ksh bar1.sh; echo $?
SIGINT
99
$ /bin/ksh bar2.sh; echo $?
130
And with this, I think the issue has been explained. Now all it remains
to do is document it in the automanual ;-)
Regards,
Stefano
