[ptrace] Problem with ptrace cleanup logic.

[ptrace] Problem with ptrace cleanup logic.

[Tetsuo Handa]

I noticed that ptrace(PTRACE_ATTACH, 1) makes the global init process remain
killable even after a process that called ptrace(PTRACE_ATTACH, 1) exits.

Steps to reproduce:

(1) Compile ptrace-init.c as ./a.out .

----- ptrace-init.c -----
#include <sys/ptrace.h>
#include <stdio.h>

int main(int argc, char *argv[])
{
        printf("PTRACE_ATTACH=%ld\n", ptrace(PTRACE_ATTACH, 1));
        return 0;
}
----- ptrace-init.c -----

(2) Run below commands as root user.

    # ./a.out
    # kill -KILL 1




Result on 2.6.26.8 to 2.6.29.6 is

    # head -n 2 /proc/1/status
    Name:   init
    State:  S (sleeping)
    # ./a.out
    PTRACE_ATTACH=0
    # head -n 2 /proc/1/status
    Name:   init
    State:  S (sleeping)
    # kill -KILL 1

Result on 2.6.30.10 to 3.3-rc4 is

    # head -n 2 /proc/1/status
    Name:   init
    State:  S (sleeping)
    # ./a.out
    PTRACE_ATTACH=0
    # head -n 2 /proc/1/status
    Name:   init
    State:  T (stopped)
    # kill -KILL 1
    Kernel panic - not syncing: Attempted to kill init!

As you can see, the global init process remains "T (stopped)" state. I think
this is a regression because it always triggers kernel panic upon sysvinit's
shutdown sequence.

  Sending all processes the TERM signal...done.
  Sending all processes the KILL signal...
  Kernel panic - not syncing: Attempted to kill init!
  Pid: 1, comm: init Not tainted 3.2.7 #1
  Call Trace:
   [<c10354c3>] panic+0x63/0x170
   [<c1038123>] find_new_reaper+0xa3/0xb0
   [<c103824c>] forget_original_parent+0x2c/0x180
   [<c11c7f24>] ? get_current_tty+0x54/0x70

If ptrace-init.c were expected to do some cleanup logic before exit, we can't
force ptrace-init.c to do it, for ptrace-init.c can be terminated unexpectedly.

Re: [ptrace] Problem with ptrace cleanup logic.

[Oleg Nesterov]

On 02/23, Tetsuo Handa wrote:
>
> I noticed that ptrace(PTRACE_ATTACH, 1) makes the global init process remain
> killable even after a process that called ptrace(PTRACE_ATTACH, 1) exits.

Yes, known problem. We should cleanup the usage of signal->flags.
This also allows to kill the new bitfields in signal_struct.

But firstly we should cleanup force_sig_info(), I sent the initial
patches.

The problem is: init loses SIGNAL_UNKILLABLE.

Oleg.