Re: [RFC][PATCH 01/13] sched/deadline: Impose global limits on sched_attr::sched_period

[Peter Zijlstra]

On Sat, Aug 31, 2019 at 03:41:17PM +0100, Alessio Balsini wrote:
> Right!
> 
> Verified that sysctl_sched_dl_period_max and sysctl_sched_dl_period_min values
> are now always consistent.
> 
> I spent some time in trying to figure out if not having any mutex in
> __checkparam_dl() is safe. There can surely happen that "max < min", e.g.:
> 
>           |              |               periods
> User1     | User2        | checkparam_dl()  | sysctl_sched_dl_*
> ----------|--------------|------------------|-------------------
>           |              |                  | [x, x]
> p_min = 5 |              |                  |
>           |              |                  | [5, x]
> p_max = 5 |              |                  |
>           |              |                  | [5, 5]
>           | setattr(p=8) |                  |
>           |              | p = 8            |
>           |              | [x, 5]           |
> p_max = 9 |              |                  |
>           |              |                  | [5, 9]
> p_min = 6 |              |                  |
>           |              |                  | [6, 9]
>           |              | [6, 5]           |
> ----------|--------------|------------------|-------------------
> 
> Sharing my thoughts, a "BUG_ON(max < min)" in __checkparam_dl() is then a
> guaranteed source of explosions, but the good news is that "if (period < min ||
> period > max" in __checkparam_dl() surely fails if "max < min".  Also the fact
> that, when we are writing the new sysctl_sched_dl_* values, only one is
> actually changed at a time, that surely helps to preserve the consistency.
> 
> But is that enough?

Strictly speaking, no, I suppose it is not. We can have two changes in
between the two READ_ONCE()s and then we'd be able to observe a
violation.

The easy way to fix that is do something like:

+	synchronize_rcu();
	mutex_unlock(&mutex);

in sched_dl_period_handler(). And do:

+	preempt_disable();
	max = (u64)READ_ONCE(sysctl_sched_dl_period_max) * NSEC_PER_USEC;
	min = (u64)READ_ONCE(sysctl_sched_dl_period_min) * NSEC_PER_USEC;
+	preempt_enable();

in __checkparam_dl().

That would prohibit we see two changes, and seeing only the single
change is safe.