Re: [PATCH 6/7] cpuidle: Handle TIF_NR_POLLING on behalf of CPUIDLE_FLAG_POLLING_HARD states

From: Rafael J. Wysocki
Date: Tue Dec 12 2023 - 08:22:50 EST

On Fri, Nov 24, 2023 at 11:32 PM Frederic Weisbecker
<frederic@xxxxxxxxxx> wrote:
>
> From: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
>
> The current handling of TIF_NR_POLLING is a bit of a maze:
>
> 1) A common brief part in the generic idle loop sets TIF_NR_POLLING
> while cpuidle selects an appropriate state and the tick is evaluated
> and then stopped. Summary: One pair of set/clear
>
> 2) The state cpuidle is then called with TIF_NR_POLLING cleared but if
> the state polls on need_resched() (software or hardware), it sets
> again TIF_NR_POLLING and clears it when it completes. Summary: another
> pair of set/clear
>
> 3) goto 1)
>
> However those costly atomic operations, fully ordered RmW for some of
> them, could be avoided if the cpuidle core knew in advance if the target
> state polls on need_resched(). If so, TIF_NR_POLLING could simply be
> set once before entering the idle loop and cleared once after idle loop
> exit.
>
> Start dealing with that with handling TIF_NR_POLLING on behalf of
> CPUIDLE_FLAG_POLLING_HARD states.
>
> [fweisbec: _ Handle broadcast properly
> _ Ignore mwait_idle() as it can be used by default_idle_call()]
>
> Not-yet-signed-off-by: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
> Signed-off-by: Frederic Weisbecker <frederic@xxxxxxxxxx>
> ---
> arch/x86/include/asm/mwait.h | 3 +--
> drivers/cpuidle/cpuidle.c | 22 +++++++++++++++++++-
> include/linux/sched/idle.h | 7 ++++++-
> kernel/sched/idle.c | 40 +++++++++++++-----------------------
> 4 files changed, 42 insertions(+), 30 deletions(-)
>
> diff --git a/arch/x86/include/asm/mwait.h b/arch/x86/include/asm/mwait.h
> index 920426d691ce..3634d00e5c37 100644
> --- a/arch/x86/include/asm/mwait.h
> +++ b/arch/x86/include/asm/mwait.h
> @@ -116,7 +116,7 @@ static __always_inline void __sti_mwait(unsigned long eax, unsigned long ecx)
> */
> static __always_inline void mwait_idle_with_hints(unsigned long eax, unsigned long ecx)
> {
> - if (static_cpu_has_bug(X86_BUG_MONITOR) || !current_set_polling_and_test()) {
> + if (static_cpu_has_bug(X86_BUG_MONITOR) || !need_resched()) {
> if (static_cpu_has_bug(X86_BUG_CLFLUSH_MONITOR)) {
> mb();
> clflush((void *)&current_thread_info()->flags);
> @@ -134,7 +134,6 @@ static __always_inline void mwait_idle_with_hints(unsigned long eax, unsigned lo
> }
> }
> }
> - current_clr_polling();
> }
>
> /*
> diff --git a/drivers/cpuidle/cpuidle.c b/drivers/cpuidle/cpuidle.c
> index 737a026ef58a..49078cc83f4a 100644
> --- a/drivers/cpuidle/cpuidle.c
> +++ b/drivers/cpuidle/cpuidle.c
> @@ -213,10 +213,10 @@ noinstr int cpuidle_enter_state(struct cpuidle_device *dev,
> int index)
> {
> int entered_state;
> -
> struct cpuidle_state *target_state = &drv->states[index];
> bool broadcast = !!(target_state->flags & CPUIDLE_FLAG_TIMER_STOP);
> ktime_t time_start, time_end;
> + bool polling;
>
> instrumentation_begin();
>
> @@ -236,6 +236,23 @@ noinstr int cpuidle_enter_state(struct cpuidle_device *dev,
> broadcast = false;
> }
>
> + polling = target_state->flags & CPUIDLE_FLAG_POLLING_HARD;
> +
> + /*
> + * If the target state doesn't poll on need_resched(), this is
> + * the last check after which further TIF_NEED_RESCHED remote setting
> + * will involve an IPI.
> + */
> + if (!polling && current_clr_polling_and_test()) {
> + if (broadcast)
> + tick_broadcast_exit();
> + dev->last_residency_ns = 0;
> + local_irq_enable();
> + instrumentation_end();
> + return -EBUSY;
> + }
> +
> +
> if (target_state->flags & CPUIDLE_FLAG_TLB_FLUSHED)
> leave_mm(dev->cpu);
>
> @@ -335,6 +352,9 @@ noinstr int cpuidle_enter_state(struct cpuidle_device *dev,
> dev->states_usage[index].rejected++;
> }
>
> + if (!polling)
> + __current_set_polling();
> +
> instrumentation_end();
>
> return entered_state;
> diff --git a/include/linux/sched/idle.h b/include/linux/sched/idle.h
> index 478084f9105e..50c13531f5d8 100644
> --- a/include/linux/sched/idle.h
> +++ b/include/linux/sched/idle.h
> @@ -68,6 +68,8 @@ static __always_inline bool __must_check current_set_polling_and_test(void)
>
> static __always_inline bool __must_check current_clr_polling_and_test(void)
> {
> + bool ret;
> +
> __current_clr_polling();
>
> /*
> @@ -76,7 +78,10 @@ static __always_inline bool __must_check current_clr_polling_and_test(void)
> */
> smp_mb__after_atomic();
>
> - return unlikely(tif_need_resched());
> + ret = unlikely(tif_need_resched());
> + if (ret)
> + __current_set_polling();
> + return ret;
> }
>
> #else
> diff --git a/kernel/sched/idle.c b/kernel/sched/idle.c
> index 565f8374ddbb..4e554b4e3781 100644
> --- a/kernel/sched/idle.c
> +++ b/kernel/sched/idle.c
> @@ -94,11 +94,12 @@ void __cpuidle default_idle_call(void)
> stop_critical_timings();
>
> ct_cpuidle_enter();
> - arch_cpu_idle();
> + arch_cpu_idle(); // XXX assumes !polling
> ct_cpuidle_exit();
>
> start_critical_timings();
> trace_cpu_idle(PWR_EVENT_EXIT, smp_processor_id());
> + __current_set_polling();
> }
> local_irq_enable();
> instrumentation_end();
> @@ -107,31 +108,14 @@ void __cpuidle default_idle_call(void)
> static int call_cpuidle_s2idle(struct cpuidle_driver *drv,
> struct cpuidle_device *dev)
> {
> + int ret;
> +
> if (current_clr_polling_and_test())
> return -EBUSY;
>
> - return cpuidle_enter_s2idle(drv, dev);
> -}
> -
> -static int call_cpuidle(struct cpuidle_driver *drv, struct cpuidle_device *dev,
> - int next_state)

Since you are removing call_cpuidle(), you may as well remove
call_cpuidle_s2idle() which only has one caller anyway.

> -{
> - /*
> - * The idle task must be scheduled, it is pointless to go to idle, just
> - * update no idle residency and return.
> - */
> - if (current_clr_polling_and_test()) {
> - dev->last_residency_ns = 0;
> - local_irq_enable();
> - return -EBUSY;
> - }
> -
> - /*
> - * Enter the idle state previously returned by the governor decision.
> - * This function will block until an interrupt occurs and will take
> - * care of re-enabling the local interrupts
> - */
> - return cpuidle_enter(drv, dev, next_state);
> + ret = cpuidle_enter_s2idle(drv, dev);
> + __current_set_polling();
> + return ret;
> }
>
> /**
> @@ -198,7 +182,7 @@ static void cpuidle_idle_call(void)
> tick_nohz_idle_stop_tick();
>
> next_state = cpuidle_find_deepest_state(drv, dev, max_latency_ns);
> - call_cpuidle(drv, dev, next_state);
> + cpuidle_enter(drv, dev, next_state);
> } else {
> bool stop_tick = true;
>
> @@ -212,7 +196,12 @@ static void cpuidle_idle_call(void)
> else
> tick_nohz_idle_retain_tick();
>
> - entered_state = call_cpuidle(drv, dev, next_state);
> + /*
> + * Enter the idle state previously returned by the governor decision.
> + * This function will block until an interrupt occurs and will take
> + * care of re-enabling the local interrupts.
> + */
> + entered_state = cpuidle_enter(drv, dev, next_state);
> /*
> * Give the governor an opportunity to reflect on the outcome
> */
> @@ -220,7 +209,6 @@ static void cpuidle_idle_call(void)
> }
>
> exit_idle:
> - __current_set_polling();
>
> /*
> * It is up to the idle functions to reenable local interrupts
> --
> 2.42.1
>