preempt_schedule_irq() loop question

preempt_schedule_irq() loop question

[Valentin Schneider]

Hi,

I've been wandering around preempt_schedule_irq() in sched/core.c, and
got curious regarding how the arch code calls it.

The main part of preempt_schedule_irq() is:

    do {
	    preempt_disable();
	    local_irq_enable();
	    __schedule(true);
	    local_irq_disable();
	    sched_preempt_enable_no_resched();
    } while (need_resched());

Yet all the arch entry.S I looked at (I stopped after arm64, arm, x86_32,
MIPS, powerpc) wrap the call to preempt_schedule_irq() in another

    do { ... } while (need_resched())

For instance, this is what's done in arm64:

    1:	bl	preempt_schedule_irq		// irq en/disable is done inside
	ldr	x0, [tsk, #TSK_TI_FLAGS]	// get new tasks TI_FLAGS
	tbnz	x0, #TIF_NEED_RESCHED, 1b	// needs rescheduling?


I naively thought this could be attributed to something like
preempt_schedule_irq() historically not having an inner loop, but it seems
to have been there since the beginning of time (or at least up to the point
where the git history stops).

I don't see why we need to have these nested loops - AFAICT the one in
preempt_schedule_irq() would suffice. What am I missing?

Re: preempt_schedule_irq() loop question

[Peter Zijlstra]

On Wed, Jan 16, 2019 at 12:50:42PM +0000, Valentin Schneider wrote:
> Hi,
> 
> I've been wandering around preempt_schedule_irq() in sched/core.c, and
> got curious regarding how the arch code calls it.
> 
> The main part of preempt_schedule_irq() is:
> 
>     do {
> 	    preempt_disable();
> 	    local_irq_enable();
> 	    __schedule(true);
> 	    local_irq_disable();
> 	    sched_preempt_enable_no_resched();
>     } while (need_resched());
> 
> Yet all the arch entry.S I looked at (I stopped after arm64, arm, x86_32,
> MIPS, powerpc) wrap the call to preempt_schedule_irq() in another
> 
>     do { ... } while (need_resched())
> 
> For instance, this is what's done in arm64:
> 
>     1:	bl	preempt_schedule_irq		// irq en/disable is done inside
> 	ldr	x0, [tsk, #TSK_TI_FLAGS]	// get new tasks TI_FLAGS
> 	tbnz	x0, #TIF_NEED_RESCHED, 1b	// needs rescheduling?
> 
> 
> I naively thought this could be attributed to something like
> preempt_schedule_irq() historically not having an inner loop, but it seems
> to have been there since the beginning of time (or at least up to the point
> where the git history stops).
> 
> I don't see why we need to have these nested loops - AFAICT the one in
> preempt_schedule_irq() would suffice. What am I missing?

I think you're quite right; but I wasn't doing kernel work back when rml
added the preemptible bits. Ingo, do you have any recollections that far
back?

Re: preempt_schedule_irq() loop question

[Thomas Gleixner]

On Wed, 16 Jan 2019, Peter Zijlstra wrote:
> On Wed, Jan 16, 2019 at 12:50:42PM +0000, Valentin Schneider wrote:
> > Hi,
> > 
> > I've been wandering around preempt_schedule_irq() in sched/core.c, and
> > got curious regarding how the arch code calls it.
> > 
> > The main part of preempt_schedule_irq() is:
> > 
> >     do {
> > 	    preempt_disable();
> > 	    local_irq_enable();
> > 	    __schedule(true);
> > 	    local_irq_disable();
> > 	    sched_preempt_enable_no_resched();
> >     } while (need_resched());
> > 
> > Yet all the arch entry.S I looked at (I stopped after arm64, arm, x86_32,
> > MIPS, powerpc) wrap the call to preempt_schedule_irq() in another
> > 
> >     do { ... } while (need_resched())
> > 
> > For instance, this is what's done in arm64:
> > 
> >     1:	bl	preempt_schedule_irq		// irq en/disable is done inside
> > 	ldr	x0, [tsk, #TSK_TI_FLAGS]	// get new tasks TI_FLAGS
> > 	tbnz	x0, #TIF_NEED_RESCHED, 1b	// needs rescheduling?
> > 
> > 
> > I naively thought this could be attributed to something like
> > preempt_schedule_irq() historically not having an inner loop, but it seems
> > to have been there since the beginning of time (or at least up to the point
> > where the git history stops).
> > 
> > I don't see why we need to have these nested loops - AFAICT the one in
> > preempt_schedule_irq() would suffice. What am I missing?
> 
> I think you're quite right; but I wasn't doing kernel work back when rml
> added the preemptible bits. Ingo, do you have any recollections that far
> back?

I just went back in the history tree and had to figure out that it's my
fault :)

preempt_schedule_irq() was introduced to plug a stupid race, but I did not
notice (and obviously nobody else) that this made the extra loop in the
entry code pointless. So it's just histerical raisins without any value.

Thanks,

	tglx

Re: preempt_schedule_irq() loop question

[Valentin Schneider]

On 16/01/2019 14:57, Thomas Gleixner wrote:
[...]
> 
> I just went back in the history tree and had to figure out that it's my
> fault :)
> 
> preempt_schedule_irq() was introduced to plug a stupid race, but I did not
> notice (and obviously nobody else) that this made the extra loop in the
> entry code pointless. So it's just histerical raisins without any value.
> 

I see, thanks for clearing that up! Credit goes to Julien & Will for
unearthing the oddity.

> Thanks,
> 
> 	tglx
>