2007-08-13 23:27:20

by Joe Korty

Subject: [PATCH] hres_timers_resume must block interrupts

Retrigger_next_event() must be called with interrupts disabled.

All internal (to hrtimer.c) uses of retrigger_next_event() are correct.
But the version exported to other files, hres_timers_resume(), does not
do the IRQ blocking, nor does the (single) external caller of it.

Rather than require that users of hres_timers_resume() do the IRQ blocking,
this patch makes the blocking part of the hres_timers_resume() functionality.

(Also remove the meaningless WARN_ON_ONCE() call in hres_timers_resume)

Signed-off-by: Joe Korty ([email protected])

Index: 2.6.23-rc3/kernel/hrtimer.c
===================================================================
--- 2.6.23-rc3.orig/kernel/hrtimer.c 2007-08-13 18:30:09.000000000 -0400
+++ 2.6.23-rc3/kernel/hrtimer.c 2007-08-13 18:38:48.000000000 -0400
@@ -463,10 +463,11 @@
*/
void hres_timers_resume(void)
{
- WARN_ON_ONCE(num_online_cpus() > 1);
+ unsigned long flags;

- /* Retrigger the CPU local events: */
+ local_irq_save(flags);
retrigger_next_event(NULL);
+ local_irq_restore(flags);
}

/*

2007-08-14 05:37:04

by Thomas Gleixner

Subject: Re: [PATCH] hres_timers_resume must block interrupts

On Mon, 2007-08-13 at 19:26 -0400, Joe Korty wrote:
> Retrigger_next_event() must be called with interrupts disabled.
>
> All internal (to hrtimer.c) uses of retrigger_next_event() are correct.
> But the version exported to other files, hres_timers_resume(), does not
> do the IRQ blocking, nor does the (single) external caller of it.
>
> Rather than require that users of hres_timers_resume() do the IRQ blocking,
> this patch makes the blocking part of the hres_timers_resume() functionality.

Ack.

> (Also remove the meaningless WARN_ON_ONCE() call in hres_timers_resume)

It's not that meaningless. It catched a resume order problem some time
ago.

tglx

> Signed-off-by: Joe Korty ([email protected])
>
> Index: 2.6.23-rc3/kernel/hrtimer.c
> ===================================================================
> --- 2.6.23-rc3.orig/kernel/hrtimer.c 2007-08-13 18:30:09.000000000 -0400
> +++ 2.6.23-rc3/kernel/hrtimer.c 2007-08-13 18:38:48.000000000 -0400
> @@ -463,10 +463,11 @@
> */
> void hres_timers_resume(void)
> {
> - WARN_ON_ONCE(num_online_cpus() > 1);
> + unsigned long flags;
>
> - /* Retrigger the CPU local events: */
> + local_irq_save(flags);
> retrigger_next_event(NULL);
> + local_irq_restore(flags);
> }
>
> /*
>

2007-08-14 06:28:00

by Thomas Gleixner

Subject: Re: [PATCH] hres_timers_resume must block interrupts

On Tue, 2007-08-14 at 07:36 +0200, Thomas Gleixner wrote:
> On Mon, 2007-08-13 at 19:26 -0400, Joe Korty wrote:
> > Retrigger_next_event() must be called with interrupts disabled.
> >
> > All internal (to hrtimer.c) uses of retrigger_next_event() are correct.
> > But the version exported to other files, hres_timers_resume(), does not
> > do the IRQ blocking, nor does the (single) external caller of it.
> >
> > Rather than require that users of hres_timers_resume() do the IRQ blocking,
> > this patch makes the blocking part of the hres_timers_resume() functionality.
>
> Ack.

-ENOTENOUGHCOFFEE.

The resume functions are called with interrupts disabled. Nothing to fix
here.

tglx