Re: [PATCH V3] exit: trigger panic when global init has exited

[Oleg Nesterov <oleg@redhat.com>]

On 03/23, qianli zhao wrote:
>
> Hi,Oleg
>
> > No, there is at least one alive init thread. If they all have exited, we have
> > the thread which calls panic() above.
>
> By current logic, setting PF_EXITING(exit_signals()) is before the
> panic(),

You certainly don't understand me :/

Please read my email you quoted below. I didn't mean the current logic.
I meant the logic after your patch which moves atomic_dec_and_test() and
panic() before exit_signals().

Oleg.

> find_alive_thread() determines the PF_EXITING of all child
> threads, the panic thread's PF_EXITING has been set before panic(),so
> find_alive_thread() thinks this thread also dead, resulting in
> find_alive_thread returning NULL.It is possible to trigger a
> zap_pid_ns_processes()->BUG() in this case.
> ========
> exit_signals(tsk);  /* sets PF_EXITING */
> ...
> group_dead = atomic_dec_and_test(&tsk->signal->live);
> if (group_dead) {
> if (unlikely(is_global_init(tsk)))
> panic("Attempted to kill init!
> exitcode=0x%08x\n",-------------------->//PF_EXITING has been set
> tsk->signal->group_exit_code ?: (int)code);
> 
> =======
> 
> > Why do you think so? It can affect _any_ code which runs under
> > "if (group_dead)". Again, I don't see anything wrong, but I didn't even
> > try to audit these code paths.
> 
> Yes,all places where checked the "signal->live" may be affected,but
> even before my changes, each program that checks "signal->live" may
> get different state(group_dead or not), depending on the timing of the
> caller,this situation will not change after my change.
> After my patch,"signal->live--" and other variable are set in a
> different order(such as signal->live and PF_EXITING),this can cause
> abnormalities in the logic associated with these two variables,that is
> my thinking.
> Of course, check all the "signal->live--" path is definitely
> necessary,it's just the case above that we need more attention.
> 
> Thanks
> 
> Oleg Nesterov <oleg@redhat.com> 于2021年3月23日周二 上午12:37写道：
> >
> > Hi,
> >
> > It seems that we don't understand each other.
> >
> > If we move atomic_dec_and_test(signal->live) and do
> >
> >         if (group_dead && is_global_init)
> >                 panic(...);
> >
> >
> > before setting PF_EXITING like your patch does, then zap_pid_ns_processes()
> > simply won't be called.
> >
> > Because:
> >
> > On 03/21, qianli zhao wrote:
> > >
> > > Hi,Oleg
> > >
> > > > How? Perhaps I missed something again, but I don't think this is possible.
> > >
> > > > zap_pid_ns_processes() simply won't be called, find_child_reaper() will
> > > > see the !PF_EXITING thread which calls panic().
> > >
> > > > So I think this should be documented somehow, at least in the changelog.
> > >
> > > This problem occurs when both two init threads enter the do_exit,
> > > One of the init thread is syscall sys_exit_group,and set SIGNAL_GROUP_EXIT
> > > The other init thread perform ret_to_user()->get_signal() and found
> > > SIGNAL_GROUP_EXIT is set,then do_group_exit()->do_exit(),since there
> > > are no alive init threads it finally goes to
> > > zap_pid_ns_processes()
> >
> > No, there is at least one alive init thread. If they all have exited, we have
> > the thread which calls panic() above.
> >
> > > and BUG().
> >
> > so we don't need the SIGNAL_GROUP_EXIT check to avoid this BUG().
> >
> > What have I missed?
> >
> > Oleg.
> >
>