On Tue, Apr 18, 2017 at 11:39:27AM +0200, Johannes Berg wrote:
> On Mon, 2017-04-17 at 09:01 -0700, Paul E. McKenney wrote:
>
> > If you have not already done so, please run this with debug enabled,
> > especially CONFIG_PROVE_LOCKING=y (which implies CONFIG_PROVE_RCU=y).
> > This is important because there are configurations for which the
> > deadlocks you saw with SRCU turn into silent failure, including
> > memory corruption.
> > CONFIG_PROVE_RCU=y will catch many of those situations.
>
> Can you elaborate on that? I think we may have had CONFIG_PROVE_RCU
> enabled in the builds where we saw the problem, but I'm not sure.
CONFIG_PROVE_RCU=y will reliably catch things like this:
1. rcu_read_lock();
synchronize_rcu();
rcu_read_unlock();
With CONFIG_PROVE_RCU=n and CONFIG_PREEMPT=n, this will result in
too-short grace periods, which can free things out from under the
read-side critical section, which in turn can result in arbitrary
memory corruption. You might not even get a "scheduling while
atomic", though CONFIG_PREEMPT_COUNT=y will produce this message.
With CONFIG_PREEMPT=y, on the other hand, this should
deadlock in a manner similar to the earlier SRCU deadlocks
seen in debugfs.
2. rcu_read_lock();
schedule_timeout_interruptible(HZ);
rcu_read_unlock();
With CONFIG_PROVE_RCU=y and CONFIG_PREEMPT=y, this will just
work, more or less. Until someone runs with CONFIG_PREEMPT=n,
which will produce "scheduling while atomic". (I have a
fix for this queued for 4.13, FWIW, so that in the future
CONFIG_PROVE_RCU=y and CONFIG_PREEMPT=y will complain about
this. But for now, silent bug.)
There are more, but this should get you the flavor of the types
of bugs CONFIG_PROVE_RCU=y can locate for you.
> Can you say which configurations you're thinking of? And perhaps what
> kind of corruption you're thinking of also? I'm having a hard time
> imagining any corruption that should happen?
#1 is the silent corruption case given CONFIG_PROVE_RCU=n,
CONFIG_PREEMPT=n, and CONFIG_PREEMPT_COUNT=n.
> Nicolai probably never even ran into this problem, though it should be
> easy to reproduce.
I am just worried that the situation resulting in the earlier SRCU
deadlocks might be hiding behind CONFIG_PROVE_RCU=n, CONFIG_PREEMPT=n,
and CONFIG_PREEMPT_COUNT=n. Or some other bug hiding behind some
other set of Kconfig options.
Thanx, Paul
