Re: + taskstats-add-all-thread-stats-for-taskstats_cmd_attr_tgid.patch added to -mm tree

Re: + taskstats-add-all-thread-stats-for-taskstats_cmd_attr_tgid.patch added to -mm tree

[Guillaume Chazarain]

Le Thu, 13 Sep 2007 16:41:41 +0400,
Oleg Nesterov <oleg@tv-sign.ru> a écrit :

> (sorry, have no time to read the patch carefully, just a couple of
> minor random "unless I misread this patch" nits)

Thank you Oleg, hopefully I addressed all your comments in v4.

> > +	do
> > +		if (!tsk->exit_state)
> > +			add_tsk(stats, tsk);
> > +	while_each_thread(first, tsk);
> 
> This reminds me. The "!tsk->exit_state" is not good and racy, it
> can't prevent the double accounting (of course, I don't blame this
> patch).

I documented this in v4, but I was wondering if always setting
tsk->exit_state under the protection of lock_task_sighand() would fix
the race.

> > +	if (thread_group_leader(task)) {
> > +		stats->ac_exitcode = task->exit_code;
> > +		if (task->flags & PF_FORKNOEXEC)
> >  			stats->ac_flag |= AFORK;
> 
> Actually, I can't understand this "thread_group_leader()" check, but

Seems like this code was copied from kernel/acct.c:acct_collect():
	if (thread_group_leader(current)) {
		pacct->ac_exitcode = exitcode;
		if (current->flags & PF_FORKNOEXEC)
			pacct->ac_flag |= AFORK;
	}

> I don't know what ->ac_exitcode and AFORK means to the user-space.

>From include/linux/acct.h: executed fork, but did not exec

Except that it talks about fork(2), not do_fork() hence the
thread_group_leader() test to avoid flagging threads.

> But it looks a bit suspicious. Perhaps we need
> ->signal->group_exit_code, not task->exit_code?

I put ->signal->group_exit_code if not null, otherwise
leader->exit_code.

Thank again to you and Andrew for the review.

-- 
Guillaume

Re: + taskstats-add-all-thread-stats-for-taskstats_cmd_attr_tgid.patch added to -mm tree

[Oleg Nesterov]

On 09/15, Guillaume Chazarain wrote:
>
> Le Thu, 13 Sep 2007 16:41:41 +0400,
> Oleg Nesterov <oleg@tv-sign.ru> a ??crit :
> 
> > > +	do
> > > +		if (!tsk->exit_state)
> > > +			add_tsk(stats, tsk);
> > > +	while_each_thread(first, tsk);
> > 
> > This reminds me. The "!tsk->exit_state" is not good and racy, it
> > can't prevent the double accounting (of course, I don't blame this
> > patch).
> 
> I documented this in v4, but I was wondering if always setting
> tsk->exit_state under the protection of lock_task_sighand() would fix
> the race.

No, this is not enough. There is a window between do_exit()->taskstats_exit()
and do_exit()->exit_notify() which sets ->exit_state. Probably, we can
shift taskstats_exit() into exit_notify() but this is not good.

Oleg.