2005-01-31 12:24:14

by Oleg Nesterov

Subject: [PATCH] rc2-mm2: unneeded cli/sti in fix-preemption-race patch

Hello.

Thomas Gleixner wrote:
>
> The patch prevents this by leaving interrupts disabled and calling
> a seperate function preempt_schedule_irq().

sched-fix-preemption-race-core-i386.patch:
>
> +++ 25-akpm/arch/i386/kernel/entry.S
> @@ -189,6 +189,7 @@ ENTRY(resume_userspace)
>
> #ifdef CONFIG_PREEMPT
> ENTRY(resume_kernel)
> + cli

I think this 'cli' is unneeded. resume_kernel: is used on
return path from do_IRQ() which leaves interrupts disabled.
irq_desc->handler->end() should not enable interrupts, yes?

And we have preempt_stop==cli in ret_from_exception: case,
before ret_from_intr.

> +++ 25-akpm/kernel/sched.c
> @@ -2872,6 +2872,48 @@ need_resched:
> ...
> +asmlinkage void __sched preempt_schedule_irq(void)
> +{
> ...
> + local_irq_enable();
> + schedule();

It is ok to enter schedule() with interrupts disabled, so this
'sti' seems to be unneeded too.

Am I missed something?

Signed-off-by: Oleg Nesterov <[email protected]>

--- 2.6.11-rc2-mm2/arch/i386/kernel/entry.S~ Mon Jan 31 14:09:37 2005
+++ 2.6.11-rc2-mm2/arch/i386/kernel/entry.S Mon Jan 31 14:55:25 2005
@@ -189,7 +189,6 @@ ENTRY(resume_userspace)

#ifdef CONFIG_PREEMPT
ENTRY(resume_kernel)
- cli
cmpl $0,TI_preempt_count(%ebp) # non-zero preempt_count ?
jnz restore_all
need_resched:
--- 2.6.11-rc2-mm2/kernel/sched.c~ Mon Jan 31 14:09:54 2005
+++ 2.6.11-rc2-mm2/kernel/sched.c Mon Jan 31 14:57:44 2005
@@ -2971,7 +2971,6 @@ need_resched:
saved_lock_depth = task->lock_depth;
task->lock_depth = -1;
#endif
- local_irq_enable();
schedule();
local_irq_disable();
#ifdef CONFIG_PREEMPT_BKL

2005-01-31 13:23:56

by Thomas Gleixner

Subject: Re: [PATCH] rc2-mm2: unneeded cli/sti in fix-preemption-race patch

On Mon, 2005-01-31 at 16:24 +0300, Oleg Nesterov wrote:
> Hello.
>
> Thomas Gleixner wrote:
> >
> > The patch prevents this by leaving interrupts disabled and calling
> > a seperate function preempt_schedule_irq().
>
> sched-fix-preemption-race-core-i386.patch:
> >
> > +++ 25-akpm/arch/i386/kernel/entry.S
> > @@ -189,6 +189,7 @@ ENTRY(resume_userspace)
> >
> > #ifdef CONFIG_PREEMPT
> > ENTRY(resume_kernel)
> > + cli
>
> I think this 'cli' is unneeded. resume_kernel: is used on
> return path from do_IRQ() which leaves interrupts disabled.
> irq_desc->handler->end() should not enable interrupts, yes?
>
> And we have preempt_stop==cli in ret_from_exception: case,
> before ret_from_intr.

Right. I missed the preempt_stop == cli.

> > +++ 25-akpm/kernel/sched.c
> > @@ -2872,6 +2872,48 @@ need_resched:
> > ...
> > +asmlinkage void __sched preempt_schedule_irq(void)
> > +{
> > ...
> > + local_irq_enable();
> > + schedule();
>
> It is ok to enter schedule() with interrupts disabled, so this
> 'sti' seems to be unneeded too.

Makes sense.

tglx

> Signed-off-by: Oleg Nesterov <[email protected]>
>
> --- 2.6.11-rc2-mm2/arch/i386/kernel/entry.S~ Mon Jan 31 14:09:37 2005
> +++ 2.6.11-rc2-mm2/arch/i386/kernel/entry.S Mon Jan 31 14:55:25 2005
> @@ -189,7 +189,6 @@ ENTRY(resume_userspace)
>
> #ifdef CONFIG_PREEMPT
> ENTRY(resume_kernel)
> - cli
> cmpl $0,TI_preempt_count(%ebp) # non-zero preempt_count ?
> jnz restore_all
> need_resched:
> --- 2.6.11-rc2-mm2/kernel/sched.c~ Mon Jan 31 14:09:54 2005
> +++ 2.6.11-rc2-mm2/kernel/sched.c Mon Jan 31 14:57:44 2005
> @@ -2971,7 +2971,6 @@ need_resched:
> saved_lock_depth = task->lock_depth;
> task->lock_depth = -1;
> #endif
> - local_irq_enable();
> schedule();
> local_irq_disable();
> #ifdef CONFIG_PREEMPT_BKL