Re: Unkillable processes due to PTRACE_TRACEME

[Oleg Nesterov <oleg@redhat.com>]

On 10/19, Dmitry Vyukov wrote:
>
> The following program hangs in some interesting state and is not
> killable (started by a normal user, not root):

Thanks.

> #include <pthread.h>
> #include <unistd.h>
> #include <sys/ptrace.h>
> #include <stdio.h>
> #include <signal.h>
>
> void *thr(void *arg) {
>         ptrace(PTRACE_TRACEME, 0, 0, 0);
>         sleep(3);
>         kill(getpid(), SIGCHLD);
>         return 0;
> }
>
> int main() {
>         if (fork() == 0) {
>                 sleep(1);
>                 pthread_t th;
>                 pthread_create(&th, 0, thr, 0);
>                 sleep(1);
>         }
>         return 0;
> }
>
>
> The child process attaches as tracee to init process

Yes, although in a racy manner, the parent can exit after
PTRACE_TRACEME in this case the kernel will untrace the task
before reparenting. Not that this matters.

> and then hangs in
> a state that I don't understand. When I did a similar thing but
> attached it to a normal parent process (shell), I still was able to
> get rid of it by killing parent (shell).

See above.

So I bet the problem is that your /sbin/init doesn't use __WALL,
so wait() doesn't reap the traced zombie sub-thread, and thus it
can't release the non-empty thread group.

Could you please verify? Just do "strace -p1" and send SIGCHLD to
init.

perhaps eligible_child() should assume WALL if ptrace && ZOMBIE...

Oleg.