Hello!
RCU makes considerable use of swait and friends. The motivation I recall
was around offloaded callbacks, where in the old days the grace-period
kthread might do a wakeup for up to N tasks, where N is the number of
CPUs, all with interrupts disabled. This has since been reduced to
roughly sqrt(N) tasks, which might well still be too many wakeups to do
with interrupts disabled throughout.
However, the other use cases have at most one waiter to be awakened.
So I am guessing that I could usefully convert all but the rcu_node
structure's ->nocb_gp_wq field from swait to wait. Particularly the
use cases in SRCU and Tiny RCU.
Or is there some other reason why {S,}RCU needs to use swait that I
am forgetting?
Thanx, Paul
On 2020-03-04 16:35:26 [-0800], Paul E. McKenney wrote:
> Hello!
Hi Paul,
> Or is there some other reason why {S,}RCU needs to use swait that I
> am forgetting?
swait can be used in hardirq-context (on RT) that is in a section within
local_irq_save() / raw_spin_lock() and so.
wait on the hand uses spinlock_t which can not be used in this context.
So if you have no users which fall into that category then you could
move back to wait.h.
> Thanx, Paul
Sebastian
On Thu, Mar 5, 2020 at 3:11 AM Sebastian Andrzej Siewior
<[email protected]> wrote:
>
> On 2020-03-04 16:35:26 [-0800], Paul E. McKenney wrote:
> > Hello!
> Hi Paul,
>
> > Or is there some other reason why {S,}RCU needs to use swait that I
> > am forgetting?
>
> swait can be used in hardirq-context (on RT) that is in a section within
> local_irq_save() / raw_spin_lock() and so.
> wait on the hand uses spinlock_t which can not be used in this context.
> So if you have no users which fall into that category then you could
> move back to wait.h.
In RCU, there are some truly-atomic code sections containing an
swake_upXX() call, which would be considered atomic also on
PREEMPT_RT, one example is:
rcu_core() contains an local_irq_{save,restore}() section.
/* No grace period and unregistered callbacks? */
if (!rcu_gp_in_progress() &&
rcu_segcblist_is_enabled(&rdp->cblist) && !offloaded) {
local_irq_save(flags);
if (!rcu_segcblist_restempty(&rdp->cblist, RCU_NEXT_READY_TAIL))
rcu_accelerate_cbs_unlocked(rnp, rdp);
local_irq_restore(flags);
}
And rcu_accelerate_cbs_unlocked(rnp, rdp) calls rcu_gp_kthread_wake()
which does an swake_up_one(), so I think we will have to leave it as
swake_up() the way it is.
thanks,
- Joel
On 2020-03-18 20:47:18 [-0400], Joel Fernandes wrote:
> In RCU, there are some truly-atomic code sections containing an
> swake_upXX() call, which would be considered atomic also on
> PREEMPT_RT, one example is:
>
> rcu_core() contains an local_irq_{save,restore}() section.
>
> /* No grace period and unregistered callbacks? */
> if (!rcu_gp_in_progress() &&
> rcu_segcblist_is_enabled(&rdp->cblist) && !offloaded) {
> local_irq_save(flags);
> if (!rcu_segcblist_restempty(&rdp->cblist, RCU_NEXT_READY_TAIL))
> rcu_accelerate_cbs_unlocked(rnp, rdp);
> local_irq_restore(flags);
> }
>
> And rcu_accelerate_cbs_unlocked(rnp, rdp) calls rcu_gp_kthread_wake()
> which does an swake_up_one(), so I think we will have to leave it as
> swake_up() the way it is.
There is also this:
irq_exit()
rcu_irq_exit()
rcu_nmi_exit_common(true);
rcu_prepare_for_idle()
if (rcu_segcblist_pend_cbs(&rdp->cblist))
rcu_gp_kthread_wake()
swake_up_one(&rcu_state.gp_wq);
So I *think* this is another one.
> thanks,
>
> - Joel
Sebastian