Re: [PATCH 03/30] thread_info: tif_need_resched() now takes resched_t as param

From: Mark Rutland
Date: Mon Feb 19 2024 - 07:30:52 EST


On Wed, Feb 14, 2024 at 08:08:30PM -0800, Ankur Arora wrote:
>
> Mark Rutland <mark.rutland@xxxxxxx> writes:
>
> > On Mon, Feb 12, 2024 at 09:55:27PM -0800, Ankur Arora wrote:
> >> tif_need_resched() now takes a resched_t parameter to decide the
> >> immediacy of the need-resched.
> >
> > I see at the end of the series, most callers pass a constant:
> >
> > [mark@lakrids:~/src/linux]% git grep -w tif_need_resched
> > arch/s390/include/asm/preempt.h: return !--S390_lowcore.preempt_count && tif_need_resched(NR_now);
> > arch/s390/include/asm/preempt.h: tif_need_resched(NR_now));
> > include/asm-generic/preempt.h: return !--*preempt_count_ptr() && tif_need_resched(NR_now);
> > include/asm-generic/preempt.h: tif_need_resched(NR_now));
> > include/linux/preempt.h: if (tif_need_resched(NR_now)) \
> > include/linux/sched.h: return unlikely(tif_need_resched(NR_now));
> > include/linux/sched.h: unlikely(tif_need_resched(NR_lazy));
> > include/linux/thread_info.h:static __always_inline bool tif_need_resched(resched_t rs)
> > include/linux/thread_info.h: * With !PREEMPT_AUTO tif_need_resched(NR_lazy) is defined
> > kernel/entry/common.c: if (tif_need_resched(NR_now))
> > kernel/sched/debug.c: nr = tif_need_resched(NR_now) ? "need_resched" : "need_resched_lazy";
> > kernel/trace/trace.c: if (tif_need_resched(NR_now))
> > kernel/trace/trace.c: if (tif_need_resched(NR_lazy))
> >
> > I think it'd be clearer if we had tif_need_resched_now() and
> > tif_need_resched_lazy() wrappers rather than taking a parameter. I think that
> > if we did similar elsewhere (e.g. {set,test}_tsk_need_resched_{now,lazy}()),
> > it'd be a bit cleaner overall, since we can special-case the lazy behaviour
> > more easily/clearly.
>
> So, we have three need-resched interfaces:
>
> 1. need_resched(), need_resched_lazy()
> These are used all over non-core (and idle) code, and I don't
> see a case where the user would find it useful to dynamically
> choose one or the other.
> So, here two separate interfaces, need_resched()/need_resched_lazy()
> make sense.
>
> 2. tif_need_resched()
> This is mostly used from preempt.h or scheduler adjacent code to drive
> preemption and at least in current uses, the resched_t param is a
> compile time constant.
>
> I think the scheduler might find it useful in some cases to parametrize
> it (ex. maybe the scheduler knows how long which bit has been set for
> over long and wants to pass that on to resched_latency_warn().)
>
> But that's a contrived example. I think this one would be fine
> either way. Will try it out and see which (tif_need_resched(rs),
> or tif_need_resched_now()/tif_need_resched_lazy()) seems cleaner.
>
> 3. *_tsk_need_resched()
> This is is used almost entirely from the scheduler and RCU.
>
> One place where I found the ability to parametrize quite useful
> was __resched_curr(). So this I would like to keep.
>
> All of that said, and I wonder if we need these new interfaces at all.
> Most of the code only uses the NR_now interface. Only the scheduler and
> the entry code need to distinguish between lazy and eager.
> (Plus, this way lazy and eager becomes an implementation detail which
> doesn't need to be known outside the scheduler. Which is also kind of
> the point of PREEMPT_AUTO.)
>
> Say something like the patch below (and similar for tif_need_resched(),
> need_resched() etc.)
>
> What do you think?
>
> Ankur
>
> ---------
> diff --git a/include/linux/sched.h b/include/linux/sched.h
> index 58e6ea7572a0..b836b238b117 100644
> --- a/include/linux/sched.h
> +++ b/include/linux/sched.h
> @@ -1953,7 +1953,7 @@ static inline bool test_tsk_thread_flag(struct task_struct *tsk, int flag)
> * tif_resched(NR_now). Add a check in the helpers below to ensure
> * we don't touch the tif_reshed(NR_now) bit unnecessarily.
> */
> -static inline void set_tsk_need_resched(struct task_struct *tsk, resched_t rs)
> +static inline void __set_tsk_need_resched(struct task_struct *tsk, resched_t rs)
> {
> if (IS_ENABLED(CONFIG_PREEMPT_AUTO) || rs == NR_now)
> set_tsk_thread_flag(tsk, tif_resched(rs));
> @@ -1964,6 +1964,11 @@ static inline void set_tsk_need_resched(struct task_struct *tsk, resched_t rs)
> BUG();
> }
>
> +static inline void set_tsk_need_resched(struct task_struct *tsk, resched_t rs)
> +{
> + __set_tsk_need_resched(tsk, NR_now);
> +}

I assume for this (and test_tsk_need_resched() below), you mean to drop the
resched_t argument, i.e. this should be:

static inline void set_tsk_need_resched(struct task_struct *tsk)
{
__set_tsk_need_resched(tsk, NR_now);
}

Assuming so, this looks good to me!

Mark.

> +
> static inline void clear_tsk_need_resched(struct task_struct *tsk)
> {
> clear_tsk_thread_flag(tsk, tif_resched(NR_now));
> @@ -1972,7 +1977,7 @@ static inline void clear_tsk_need_resched(struct task_struct *tsk)
> clear_tsk_thread_flag(tsk, tif_resched(NR_lazy));
> }
>
> -static inline bool test_tsk_need_resched(struct task_struct *tsk, resched_t rs)
> +static inline bool __test_tsk_need_resched(struct task_struct *tsk, resched_t rs)
> {
> if (IS_ENABLED(CONFIG_PREEMPT_AUTO) || rs == NR_now)
> return unlikely(test_tsk_thread_flag(tsk, tif_resched(rs)));
> @@ -1980,6 +1985,11 @@ static inline bool test_tsk_need_resched(struct task_struct *tsk, resched_t rs)
> return false;
> }
>
> +static inline bool test_tsk_need_resched(struct task_struct *tsk, resched_t rs)
> +{
> + return __test_tsk_need_resched(tsk, NR_now);
> +}
> +
> /*
> * cond_resched() and cond_resched_lock(): latency reduction via
> * explicit rescheduling in places that are safe. The return