Re: [path][rfc] add PR_DETACH prctl command

[Stas Sergeev <stsp@aknet.ru>]

24.02.2011 16:29, Oleg Nesterov wrote:
>> And I really think there is no other
>> way to daemonize the process with threads, than to use something
>> like this patch, or is there?
> Depends on what "daemonize" means. Even with this patch, setsid()
> after PR_DETACH can fail because we do not change the pids and
> the caller can still be pgrp leader.
Yes, I am using TIOCNOTTY ioctl instead.
It doesn't detach the entire group from tty though, but the plan
is to implement also TIOCNOTTY_GRP, in case PR_DETACH is done.

> And. What if the parent of PR_DETACH caller blocks or ignores
> SIGCHLD or simply doesn't call do_wait()? The caller can run with
> PR_DETACH set without any effect "forever".
I am currently rewriting the patch to solve this all.
What I am trying to do now, is to reparent directly in prctl(),
but delay the list_move_tail(&p->sibling, &p->real_parent->children);
to the wait() call. If this is a feasible solution, I'll post the new patch.

> So, to be honest, I do not think this idea will be accepted, and I don't
> really understand your motivation. But once again, I never argue with the
> "we need this feature" requests, no need to convince me.
Lets see if the clean implementation is possible first. :)

> The problem is that ptrace uses this ->exit_code member as well.
> Suppose that the (ptraced) task calls PR_DETACH and, say, recieves
> a signal after that. See ptrace_signal().
Also do_signal_stop() seems to alter it.
Do you mean right now it can't happen that multiple events
alter the exit_code, and the parent notices only the last one?
In this case I need to add a separate variable.

> I understand why you added PF_EXITING. And, once again, this is not
> right afaics. The current condition is more or less "random" and mostly
> historical, although correct. If we want to take PF_EXITING into account,
> we should just add BUG_ON(!(tsk->flags&  PF_EXITING)). IOW, it is just
> wrong to call this function unless this tsk exits.
OK, I'll address also this.

Thanks for your time, I am hoping to post the patch that addresses
the pointed problems.