Le Mon, Apr 15, 2024 at 06:36:31PM +0200, Valentin Schneider a écrit :
> On 10/04/24 12:19, Paul E. McKenney wrote:
> > On Wed, Apr 10, 2024 at 12:30:55PM +0200, Frederic Weisbecker wrote:
> >> Le Tue, Apr 09, 2024 at 12:53:03PM -0700, Paul E. McKenney a écrit :
> >> > I am having a hard time getting too excited about the name. I could
> >> > suggest CT_RCU_WATCHING_IDX, but that isn't exactly the shortest
> >> > possible name.
> >>
> >> I really like CT_RCU_WATCHING. It says everything. The _IDX isn't even
> >> needed after all. What do you think?
> >
> > Works for me!
> >
>
> Sounds good to me too, thanks for the suggestion :)
>
> Now, what about ct_dynticks() & friends? I was about to do:
>
> -static __always_inline int ct_dynticks(void)
> +static __always_inline int ct_rcu_watching(void)
> {
> - return atomic_read(this_cpu_ptr(&context_tracking.state)) & CT_DYNTICKS_MASK;
> + return atomic_read(this_cpu_ptr(&context_tracking.state)) & CT_RCU_WATCHING_MASK;
> }
Yup!
>
> ... but then realised that there's more siblings to the rcu_dynticks*()
> family;
Ouch right, sorry I forgot there is so much of this namespace. But in case you're
willing to clean that up:
>
> AFAICT dynticks_nesting could also get the rcu_watching prefix treatment,
> `rcu_dynticks_task_exit() -> rcu_watching_task_exit` doesn't sound as
rcu_tasks_exit() ?
But Paul, is there a reason why check_holdout_task() doesn't check
ct_dynticks_cpu(task_cpu(t)) instead of maintaining this separate counter?
> obvious though. The rcu_dyntick event probably can't be renamed either.
I think we can rename trace_rcu_dyntick() to trace_rcu_watching()
>
> I'm not sure how far to take the renaming; seeing things like:
>
> notrace bool rcu_is_watching(void)
> {
> bool ret;
>
> preempt_disable_notrace();
> ret = !rcu_dynticks_curr_cpu_in_eqs();
> preempt_enable_notrace();
> return ret;
> }
> EXPORT_SYMBOL_GPL(rcu_is_watching);
>
> makes me think most of the rcu_*dynticks / rcu_*eqs stuff could get an
> rcu_watching facelift?
The eqs part can stay as-is. But the *dynticks* needs an update.
>
> Here are my current considerations for identifiers used in context_tracking
> in decreasing order of confidence:
>
> | Old | New |
> |---------------------------------------+---------------------------------------------------------------|
> | RCU_DYNTICKS_IDX | CT_RCU_WATCHING |
> | RCU_DYNTICKS_MASK | CT_RCU_WATCHING_MASK |
> | context_tracking.dynticks_nesting | context_tracking.rcu_watching_nesting |
This can be context_tracking.nesting (and yes one day we might need to lock up
context_tracking.nesting and context_tracking.recursion together in a room and see
who wins after a day or two).
> | context_tracking.dynticks_nmi_nesting | context_tracking.rcu_watching_nmi_nesting [bit of a mouthful] |
context_tracking.nmi_nesting
> | rcu_dynticks_curr_cpu_in_eqs() | rcu_watching_curr_cpu() [with an added negation] |
Nice!
> |---------------------------------------+---------------------------------------------------------------|
> | TRACE_EVENT_RCU(rcu_dyntick, | [Can't change?] |
It can change. Officially trace events aren't ABI. Unoficially I wouldn't dare
changing the sched switch trace event but this one is fine.
> |---------------------------------------+---------------------------------------------------------------|
> | rcu_dynticks_task_enter() | rcu_watching_task_enter()> | |
rcu_tasks_enter() ?
> | rcu_dynticks_task_exit() | rcu_watching_task_exit() |
rcu_tasks_exit() ?
> | rcu_dynticks_task_trace_enter() | rcu_watching_task_trace_enter() |
rcu_tasks_trace_enter()?
> | rcu_dynticks_task_trace_exit() | rcu_watching_task_trace_exit() |
rcu_tasks_trace_exit() ?
Thanks.
>
> Thoughts?
>
