On Tue, 2023-04-04 at 10:37 +0800, Ping-Ke Shih wrote:
> On Mon, 2023-04-03 at 13:32 +0300, Kalle Valo wrote:
> > I would expect that there's polling if you are waiting something from
> > hardware, or maybe when implementing a spin lock, but not when waiting
> > for another kernel thread. This just doesn't feel right but I don't have
> > time to propose a good alternative either, sorry.
> >
>
> I have found a solution that uses an owner variable with a spin lock
> to determine which side to free completion object. Simply show two use
> cases below:
>
> Use case 1: (normal case; free completion object by work 1)
> work 1 work 2
> wait
> (spin_lock)
> complete
> check & set owner --> owner = work1
> (spin_unlock)
> wait ok
> (spin_lock)
> check & check owner --> free by work 1
> (spin_unlock)
> free completion
>
>
> Use case 2: (timeout case; free completion object by work 2)
> work 1 work 2
> wait
> wait timeout
> (spin_lock)
> check & set owner --> owner = work 2
> (spin_unlock)
> (spin_lock)
> completion
> check & set owner --> free by work 2
> (spin_unlock)
> free completion
>
> I will apply this by v5.
>
We have a better idea that use kfree_rcu() to free completion, so no
need spin_lock(). Then, the use cases become below, and I have sent
this change by v6.
Use case 1: (normal case)
work 1 work 2
(rcu_assign_pointer(wait))
wait
(rcu_read_lock)
wait = rcu_dereference();
if (wait) --> wait != NULL
complete
(rcu_read_unlock)
wait ok
(rcu_assign_pointer(NULL))
kfree_rcu(completion)
Use case 2: (timeout case)
work 1 work 2
(rcu_assign_pointer(wait))
wait
wait timeout
(rcu_assign_pointer(NULL))
kfree_rcu(completion)
(rcu_read_lock)
wait = rcu_dereference();
if (wait) --> wait == NULL
complete
(rcu_read_unlock)
Ping-Ke
