Re: [PATCH RFC] time: drop do_sys_times spinlock

[Peter Zijlstra <peterz@infradead.org>]

[-- Attachment #1: Type: text/plain, Size: 1369 bytes --]

On Wed, Aug 13, 2014 at 07:22:30PM +0200, Oleg Nesterov wrote:
> On 08/12, Rik van Riel wrote:
> >
> > Any other ideas?
> 
> To simplify, lets suppose that we only need sum_exec_runtime.
> 
> Perhaps we can do something like this
> 
> 	u64 thread_group_sched_runtime(void)
> 	{
> 		struct task_struct *tsk = current;
> 		spinlock_t *siglock = &tsk->sighand->siglock; /* stable */
> 		struct task_struct *t;
> 		u64 x1, x2;
> 
> 	retry:
> 		x1 = tsk->signal->sum_sched_runtime;
> 		rmb();
> 		spin_unlock_wait(siglock);
> 		rmb();
> 
> 		x2 = 0;
> 		rcu_read_lock();
> 		for_each_thread(tsk, t)
> 			x2 += task_sched_runtime(t);
> 		rcu_read_unlock();
> 
> 		rmb();
> 		spin_unlock_wait(siglock);
> 		rmb();
> 
> 		if (x1 != tsk->signal->sum_sched_runtime)
> 			goto retry;
> 
> 		return x1 + x2;
> 	}
> 
> ?
> 
> We do not care if for_each_thread() misses the new thread, we can pretend
> thread_group_sched_runtime() was called before clone.
> 
> We do not care if a thread with sum_sched_runtime == 0 exits, obviously.
> 
> Otherwise "x1 != tsk->signal->sum_sched_runtime" should tell us that we
> raced with __exit_signal().

So the problem with the above is the lack of fwd progress; if there's
enough clone()/exit() happening in the thread group (and the more CPUs
the more possible), we'll keep repeating.




[-- Attachment #2: Type: application/pgp-signature, Size: 836 bytes --]