arm64: Avoid pointless schedule_preempt_irq() invocations
diff mbox series

Message ID alpine.DEB.2.21.1907171036490.1767@nanos.tec.linutronix.de
State New
Headers show
Series
  • arm64: Avoid pointless schedule_preempt_irq() invocations
Related show

Commit Message

Thomas Gleixner July 17, 2019, 8:43 a.m. UTC
When preempt_count is zero on return from interrupt then
schedule_preempt_irq() is invoked even if TIF_NEED_RESCHED is not set.

That does not make sense because schedule_preempt_irq() has to go through a
full __schedule() for nothing in that case.

Check TIF_NEED_RESCHED and invoke schedule_preempt_irq() only if set.

Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
---
Found while staring at some RT wrecakge in that area.
---
 arch/arm64/kernel/entry.S |    4 ++++
 1 file changed, 4 insertions(+)

Comments

Valentin Schneider July 17, 2019, 9:54 a.m. UTC | #1
On 17/07/2019 09:43, Thomas Gleixner wrote:
> When preempt_count is zero on return from interrupt then
> schedule_preempt_irq() is invoked even if TIF_NEED_RESCHED is not set.
> 
> That does not make sense because schedule_preempt_irq() has to go through a
> full __schedule() for nothing in that case.
> 
> Check TIF_NEED_RESCHED and invoke schedule_preempt_irq() only if set.
> 
> Signed-off-by: Thomas Gleixner <tglx@linutronix.de>

Don't we have NEED_RESCHED squashed into preempt count?

  396244692232 ("arm64: preempt: Provide our own implementation of asm/preempt.h")

So the existing check should cover that, unless I'm missing something?

> ---
> Found while staring at some RT wrecakge in that area.
> ---
>  arch/arm64/kernel/entry.S |    4 ++++
>  1 file changed, 4 insertions(+)
> 
> --- a/arch/arm64/kernel/entry.S
> +++ b/arch/arm64/kernel/entry.S
> @@ -680,6 +680,10 @@ alternative_if ARM64_HAS_IRQ_PRIO_MASKIN
>  	orr	x24, x24, x0
>  alternative_else_nop_endif
>  	cbnz	x24, 1f				// preempt count != 0 || NMI return path
> +
> +	ldr	x0, [tsk, #TSK_TI_FLAGS]        // get flags
> +	tbz	x0, #TIF_NEED_RESCHED, 1f      	// needs rescheduling?
> +
>  	bl	preempt_schedule_irq		// irq en/disable is done inside
>  1:
>  #endif
>
Thomas Gleixner July 17, 2019, 10:04 a.m. UTC | #2
On Wed, 17 Jul 2019, Valentin Schneider wrote:

> On 17/07/2019 09:43, Thomas Gleixner wrote:
> > When preempt_count is zero on return from interrupt then
> > schedule_preempt_irq() is invoked even if TIF_NEED_RESCHED is not set.
> > 
> > That does not make sense because schedule_preempt_irq() has to go through a
> > full __schedule() for nothing in that case.
> > 
> > Check TIF_NEED_RESCHED and invoke schedule_preempt_irq() only if set.
> > 
> > Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
> 
> Don't we have NEED_RESCHED squashed into preempt count?
> 
>   396244692232 ("arm64: preempt: Provide our own implementation of asm/preempt.h")
> 
> So the existing check should cover that, unless I'm missing something?

Right. Ignore me.

Patch
diff mbox series

--- a/arch/arm64/kernel/entry.S
+++ b/arch/arm64/kernel/entry.S
@@ -680,6 +680,10 @@  alternative_if ARM64_HAS_IRQ_PRIO_MASKIN
 	orr	x24, x24, x0
 alternative_else_nop_endif
 	cbnz	x24, 1f				// preempt count != 0 || NMI return path
+
+	ldr	x0, [tsk, #TSK_TI_FLAGS]        // get flags
+	tbz	x0, #TIF_NEED_RESCHED, 1f      	// needs rescheduling?
+
 	bl	preempt_schedule_irq		// irq en/disable is done inside
 1:
 #endif