Re: [PATCH 1/2] signal: Always notice exiting tasks

[Eric W. Biederman]

ebiederm@xxxxxxxxxxxx (Eric W. Biederman) writes:

> Oleg Nesterov <oleg@xxxxxxxxxx> writes:
>
>> sorry again for delay...
>>
>> On 02/07, Eric W. Biederman wrote:
>>>
>>> --- a/kernel/signal.c
>>> +++ b/kernel/signal.c
>>> @@ -2393,6 +2393,11 @@ bool get_signal(struct ksignal *ksig)
>>>  		goto relock;
>>>  	}
>>>  
>>> +	/* Has this task already been marked for death? */
>>> +	ksig->info.si_signo = signr = SIGKILL;
>>> +	if (signal_group_exit(signal))
>>> +		goto fatal;
>>> +
>>>  	for (;;) {
>>>  		struct k_sigaction *ka;
>>>  
>>> @@ -2488,6 +2493,7 @@ bool get_signal(struct ksignal *ksig)
>>>  			continue;
>>>  		}
>>>  
>>> +	fatal:
>>>  		spin_unlock_irq(&sighand->siglock);
>>
>> Eric, but this is wrong. At least this is the serious user-visible
>> change.
>>
>> Afaics, with this patch the tracee will never stop in PTRACE_EVENT_EXIT in case
>> of group_exit/exec, because schedule() in TASK_TRACED state won't block due to
>> __fatal_signal_pending().
>>
>> Yes, yes, as I said many times the semantics of PTRACE_EVENT_EXIT was never really
>> defined, it depends on /dev/random, but still I don't think we should break it even
>> more.
>
> Well it changes PTRACE_EVENT_EXIT I grant that.  It looks like that
> changes makes PTRACE_EVENT_EXIT is less than useful.
>
> The only way to perfectly preserve the previous semantics is probably to
> do something like my JOBCTL_TASK_EXIT proposal.
>
> That said I don't think even adding a JOBCTL_TASK_EXIT is enough to have
> a reliable stop of ptrace_event_exit after a process has exited.  As any
> other pending signal can cause problems there as well.
>
> I have received a report that strace -f in some cases is not noticing
> children before they die and it looks like a stop in PTRACE_EVENT_EXIT
> would fix that strace behavior.
>
> Sigh.
>
> Here I was trying for the simple minimal change and I hit this landmine.
> Which leaves me with the question of what should be semantics of signal
> handling after exit.
>
> I think from dim memory of previous conversations the desired semantics
> look like:
> a) Ignore all signal state except for SIGKILL.
> b) Letting SIGKILL wake up the process should be sufficient.
>
> I will see if I can reproduce the strace failure and see if I can cook
> up something minimal that addresses just that.  If you have suggestions
> I would love to hear them.
>
> As this was a minimal fix for SIGKILL being broken I have already sent
> the fix to Linus.  So we are looking at an incremental fix at this
> point.

In my testing I found something that concerns me.  Because we wind up
with SIGKILL in shard_pending we can not kill a process in do_exit that
has stopped at PTRACE_EVENT_EXIT.  That bug seems to go back a long ways.

Other than that, it looks like we can do the following to fix the
regression I introduced.

Oleg any ideas on how to make PTRACE_EVENT_EXIT reliably killable?

diff --git a/kernel/signal.c b/kernel/signal.c
index 99fa8ff06fd9..a1f154dca73c 100644
--- a/kernel/signal.c
+++ b/kernel/signal.c
@@ -2544,6 +2544,9 @@ bool get_signal(struct ksignal *ksig)
                }
 
        fatal:
+               /* No more signals can be pending past this point */
+               sigdelset(&current->pending.signal, SIGKILL);
+               clear_tsk_thread_flag(current, TIF_SIGPENDING);
                spin_unlock_irq(&sighand->siglock);
 
                /*

Eric