On Sat, Aug 10, 2019 at 10:26:58PM -0400, Joel Fernandes wrote:
> On Sat, Aug 10, 2019 at 11:24:46AM -0700, Paul E. McKenney wrote:
> > On Sat, Aug 10, 2019 at 12:20:37AM -0400, Joel Fernandes wrote:
> > > On Fri, Aug 09, 2019 at 08:38:14PM -0700, Paul E. McKenney wrote:
> > > > On Fri, Aug 09, 2019 at 10:42:32PM -0400, Joel Fernandes wrote:
> > > > > On Wed, Aug 07, 2019 at 10:52:15AM -0700, Paul E. McKenney wrote:
> > > > > [snip]
> > > > > > > > > @@ -3459,6 +3645,8 @@ void __init rcu_init(void)
> > > > > > > > > {
> > > > > > > > > int cpu;
> > > > > > > > >
> > > > > > > > > + kfree_rcu_batch_init();
> > > > > > > >
> > > > > > > > What happens if someone does a kfree_rcu() before this point? It looks
> > > > > > > > like it should work, but have you tested it?
> > > > > > > >
> > > > > > > > > rcu_early_boot_tests();
> > > > > > > >
> > > > > > > > For example, by testing it in rcu_early_boot_tests() and moving the
> > > > > > > > call to kfree_rcu_batch_init() here.
> > > > > > >
> > > > > > > I have not tried to do the kfree_rcu() this early. I will try it out.
> > > > > >
> > > > > > Yeah, well, call_rcu() this early came as a surprise to me back in the
> > > > > > day, so... ;-)
> > > > >
> > > > > I actually did get surprised as well!
> > > > >
> > > > > It appears the timers are not fully initialized so the really early
> > > > > kfree_rcu() call from rcu_init() does cause a splat about an initialized
> > > > > timer spinlock (even though future kfree_rcu()s and the system are working
> > > > > fine all the way into the torture tests).
> > > > >
> > > > > I think to resolve this, we can just not do batching until early_initcall,
> > > > > during which I have an initialization function which switches batching on.
> > > > > >From that point it is safe.
> > > >
> > > > Just go ahead and batch, but don't bother with the timer until
> > > > after single-threaded boot is done. For example, you could check
> > > > rcu_scheduler_active similar to how sync_rcu_exp_select_cpus() does.
> > > > (See kernel/rcu/tree_exp.h.)
> > >
> > > Cool, that works nicely and I tested it. Actually I made it such that we
> > > don't need to batch even, before the scheduler is up. I don't see any benefit
> > > of that unless we can see a kfree_rcu() flood happening that early at boot
> > > which seems highly doubtful as a real world case.
> >
> > The benefit is removing the kfree_rcu() special cases from the innards
> > of RCU, for example, in rcu_do_batch(). Another benefit is removing the
> > current restriction on the position of the rcu_head structure within the
> > enclosing data structure.
> >
> > So it would be good to avoid the current kfree_rcu() special casing within
> > RCU itself.
> >
> > Or are you using some trick that avoids both the batching and the current
> > kfree_rcu() special casing?
>
> Oh. I see what you mean. Would it be Ok with you to have that be a follow up
> patch? I am not getting rid (yet) of the special casing in rcu_do_batch in
> this patch, but can do that in another patch.
I am OK having that in another patch, and I will be looking over yours
and Byungchul's two patches tomorrow. If they look OK, I will queue them.
However, I won't send them upstream without a follow-on patch that gets
rid of the kfree_rcu() special casing within rcu_do_batch() and perhaps
elsewhere. This follow-on patch would of course also need to change rcuperf
appropriately.
> For now I am just doing something like the following in kfree_call_rcu(). I
> was almost about to hit send on the v1 and I have been testing this a lot so
> I'll post it anyway; and we can discuss more about this point on that.
>
> +void kfree_call_rcu(struct rcu_head *head, rcu_callback_t func)
> +{
> + unsigned long flags;
> + struct kfree_rcu_cpu *krcp;
> + bool monitor_todo;
> +
> + /* kfree_call_rcu() batching requires timers to be up. If the scheduler
> + * is not yet up, just skip batching and do non-batched kfree_call_rcu().
> + */
> + if (rcu_scheduler_active != RCU_SCHEDULER_RUNNING)
> + return kfree_call_rcu_nobatch(head, func);
> +
As a stopgap until the follow-on patch, this looks fine.
Thanx, Paul
