Hi,
On 3/19/2020 2:44 AM, Stephen Boyd wrote:
> Quoting Maulik Shah (2020-03-16 23:47:21)
>> Hi,
>>
>> On 3/17/2020 7:34 AM, Stephen Boyd wrote:
>>> Quoting Maulik Shah (2020-03-12 06:22:59)
>>>> Change the way interrupts get enabled at wakeup capable PDC irq chip.
>>>>
>>>> Introduce irq_set_wake call which lets interrupts enabled at PDC with
>>>> enable_irq_wake and disabled with disable_irq_wake with certain
>>>> conditions.
>>>>
>>>> Interrupt will get enabled in HW at PDC and its parent GIC if they are
>>>> either enabled is SW or marked as wake up capable.
>>> Shouldn't we only enable in PDC and GIC if it's marked wakeup capable
>>> and we're entering suspend? Otherwise we should let the hardware enable
>>> state follow the software irq enable state?
>> Not only during "sleep" but PDC (and GIC) have a role during "active" time as well.
>> so we can not just enabled at PDC and GIC when entering to suspend, interrupt need
>> to keep interrupt enabled at PDC and GIC HW when out of suspend as well.
> Yes, but if an interrupt is only marked for wakeup and not actually
> enabled we shouldn't deliver it to the GIC. That's what I'm asking
> about.
please see below details.
>>>> interrupt will get disabled in HW at PDC and its parent GIC only if its
>>>> disabled in SW and also marked as non-wake up capable.
>>>>
>>>> Signed-off-by: Maulik Shah <mkshah@xxxxxxxxxxxxxx>
>>>> ---
>>>> drivers/irqchip/qcom-pdc.c | 124 ++++++++++++++++++++++++++++++++++++++++++---
>>>> 1 file changed, 117 insertions(+), 7 deletions(-)
>>>>
>>>> diff --git a/drivers/irqchip/qcom-pdc.c b/drivers/irqchip/qcom-pdc.c
>>>> index 6ae9e1f..d698cec 100644
>>>> --- a/drivers/irqchip/qcom-pdc.c
>>>> +++ b/drivers/irqchip/qcom-pdc.c
>>>> @@ -1,6 +1,6 @@
> [...]
>>>> +
>>>> if (d->hwirq == GPIO_NO_WAKE_IRQ)
>>>> return;
>>>>
>>>> - pdc_enable_intr(d, false);
>>>> - irq_chip_disable_parent(d);
>>>> + raw_spin_lock(&pdc_lock);
>>>> +
>>>> + clear_bit(d->hwirq, pdc_enabled_irqs);
>>> clear_bit() is atomic, so why inside the lock?
>> I will move it out of lock.
>>>> + wake_status = test_bit(d->hwirq, pdc_wake_irqs);
>>>> +
>>>> + /* Disable at PDC HW if wake_status also says same */
>>>> + if (!wake_status)
>>> Should read as "if not wakeup_enabled".
>> I will update comment.
> Hopefully the comment isn't useful and can just be removed if the code
> reads properly.
Sure i will remove.
>>>> + pdc_enable_intr(d, false);
>>>> +
>>>> + raw_spin_unlock(&pdc_lock);
>>>> +
>>>> + /* Disable at GIC HW if wake_status also says same */
>>>> + if (!wake_status)
>>> This happens outside the lock, so I'm confused why any locking is needed
>>> in this function.
>> Okay, since test_bit() is also atomic so i will keep locking inside pc_enable_intr() as it is.
>>>> + irq_chip_disable_parent(d);
>>>> }
>>>>
>>>> static void qcom_pdc_gic_enable(struct irq_data *d)
>>>> @@ -101,7 +116,16 @@ static void qcom_pdc_gic_enable(struct irq_data *d)
>>>> if (d->hwirq == GPIO_NO_WAKE_IRQ)
>>>> return;
>>>>
>>>> + raw_spin_lock(&pdc_lock);
>>>> +
>>>> + set_bit(d->hwirq, pdc_enabled_irqs);
>>>> +
>>>> + /* We can blindly enable at PDC HW as we are already in enable path */
>>>> pdc_enable_intr(d, true);
>>>> +
>>>> + raw_spin_unlock(&pdc_lock);
>>>> +
>>>> + /* We can blindly enable at GIC HW as we are already in enable path */
>>>> irq_chip_enable_parent(d);
>>>> }
>>>>
> [...]
>>>> + */
>>>> +
>>>> +static int qcom_pdc_gic_set_wake(struct irq_data *d, unsigned int on)
>>>> +{
>>>> + bool enabled_status;
>>>> +
>>>> + if (d->hwirq == GPIO_NO_WAKE_IRQ)
>>>> + return 0;
>>>> +
>>>> + raw_spin_lock(&pdc_lock);
>>>> + enabled_status = test_bit(d->hwirq, pdc_enabled_irqs);
>>>> + if (on) {
>>>> + set_bit(d->hwirq, pdc_wake_irqs);
>>>> + pdc_enable_intr(d, true);
>>>> + } else {
>>>> + clear_bit(d->hwirq, pdc_wake_irqs);
>>>> + pdc_enable_intr(d, enabled_status);
>>>> + }
>>>> +
>>>> + raw_spin_unlock(&pdc_lock);
>>>> +
>>>> + /* Either "wake" or "enabled" need same status at parent as well */
>>>> + if (on || enabled_status)
>>>> + irq_chip_enable_parent(d);
>>>> + else
>>>> + irq_chip_disable_parent(d);
>>> What happens if irq is "disabled" in software, because this is the first
>>> function called on the irq, and we aren't in suspend yet. Then we get
>>> the irq. Won't we be interrupting the CPU because we've enabled in PDC
>>> and GIC hardware? Why doesn't this function update the wake bit and then
>>> leave the force on irq logic to suspend entry? Will we miss an interrupt
>>> while entering suspend because of that?
>> As PDC (and GIC) have a role during "active" time as well, interrupt should be
>> enabled in PDC and GIC HW.
> Sure. When the irq is enabled we want to enable at the GIC, but if it
> isn't enabled and we're not in suspend I would think we don't want the
> irq enabled at the GIC. But this code is doing that. Why?
Since we want to wake up in idle path LPM as well, when IRQ is marked as wake up capable and even though its disabled.
> I'd think we
> would want to make enable in the PDC driver enable the parent and then
> make the set_wake path just update some bitmap tracking wakeup enabled
> irqs.
>
> Then when we enter suspend we will enable any pdc interrupts only in the
> PDC so that we can wakeup from suspend if that interrupt comes in. When
> we wakeup we'll resend the edge interrupts to the GIC on the resume path
> and level interrupts will "just work" because they'll stay asserted
> throughout resume.
>
> The bigger problem would look to be suspend entry, but in that case we
> leave any interrupts enabled at the GIC on the path to suspend (see
> suspend_device_irq() and how it bails out early if it's marked for
> wakeup)
No it doesn't happen this way in suspend_device_irq(), not for this disabled IRQ.
Agree, it’s a bigger problem to set IRQ enabled at GIC HW which is already disabled in HW and SW but marked for wake up.
However suspend_device_irq() is of little or no-use here (for that particular IRQ). Let me step by step give details.
This will benefit everyone to understand this problem. Correct me if something below in not happening as I listed.
Step-1
Driver invokes enable_irq_wake() to mark their IRQ wake up capable.
Step-2
After enable_irq_wake(), drivers invokes disable_irq().
Let’s break it down how this disable_irq() will traverse (in current code, without this RFC PATCH)
Step-2.1
In kernel/irq/manage.c
disable_irq() => __disable_irq_nosync() => __disable_irq() => irq_disable()
Step-2.2
This will jump to kernel/irq/chip.c
irq_disable() => __irq_disable() => which then calls chip’s .irq_disable via desc->irq_data.chip->irq_disable(&desc->irq_data);
Note that this is a GPIO IRQ, gpiolib set’s this .irq_disable for every gpio chip during registration
(see below in drivers/gpio/gpiolib.c)
irqchip->irq_disable = gpiochip_irq_disable;
So it doesn’t take lazy path to disable at HW, its always unlazy (at-least for gpio IRQs)
Step-2.3
Call Jumps to gpiochip_irq_disable()
This then invokes irq_chip’s .irq_disable via chip->irq.irq_disable(d)
which finally arrives at msmgpio irq_chip’s msm_gpio_irq_disable() call from here.
it finds that IRQ controller is in hierarchy, so it invokes irq_chip_disable_parent().
Step-2.4
This invokes PDC irq_chip’s qcom_pdc_gic_disable()
At this point,
This will go ahead and “disabled in PDC HW”
This also invokes irq_chip_disable_parent() since PDC is also in hierarchy with GIC.
Step-2.5
This invokes GIC’s gic_mask_irq() since GIC doesn’t have .irq_disable implemented, it instead invokes mask.
This will go ahead and “disables at GIC HW”.
Final status at the end of step 2:
(a) IRQ is marked as wake up capable in SW
(b) IRQ is disabled in both SW and HW at GIC and PDC
Step-3
Device enters “suspend to RAM” which invokes suspend_device_irq()
Pasting the interested part here…
if (irqd_is_wakeup_set(&desc->irq_data)) {
irqd_set(&desc->irq_data, IRQD_WAKEUP_ARMED);
…
return true;
}
..
__disable_irq(desc);
Note that it bails out with above if condition here for IRQs that are marked wake up capable, as you have already pointed out.
For the rest of IRQs it goes ahead and disables them at HW via invoking __disable_irq() (be it a GIC HW or PDC HW)
so when device is in suspend only wake up capable IRQs are finally left enabled in HW.
Now, If any driver that has only done step-1, then it make sense that it will bail out here and leave IRQ enabled in HW (to be precise in its
original state at HW, without disabling it)
But some drivers did step-1 and step-2 both.
Yes even for this case, it will still bail out here, since its marked as wake up capable but it defeats the purpose of bail out
(bail out is to NOT go ahead and disable at HW), but by doing step-2 driver already disabled the IRQ at HW well before
suspend_device_irq() is invoked.
In other words, IRQ status stays same as what was at the end of step-2 (disabled in HW)
So this whole function is of no use (for that particular IRQ part). Here driver is disabling IRQ on their own via step-2 and then asks
“hey remember I marked it as wake up capable in step-1, so IRQ should resume system from suspend when it occurs” completely ignoring
the fact that its already disabled at HW in step-2.
IMO, drivers should not even do step-2 here. They should just mark irq as wake up capable and then let suspend framework decide whether
to keep it enabled or disabled in HW when entering to suspend (Read may only do step-1, then everything works fine with current code)
Hope that above details makes it clear on why I way asking earlier to remove unnecessary disable_irq() call from driver, I don’t understand its usage
(at least till now its not clear to me). May be there are some reasons like seeing spurious IRQ when entering to suspend so driver may want to disable it in HW.
But then responsibility is falling on either suspend framework to re-enable such wake up capable interrupts in HW.
This may be done by again invoking enable_irq() before bailing out for wakeup capable IRQ in suspend_device_irq(). I don’t know if this is acceptable.
Someone from To/Cc may clarify if above can be done. Note that it may create problem for other drivers which does only step-1, by calling enable_irq()
during suspend, it will update desc->depth, so this agin need to be undo when resuming.
Otherwise, it is currently falling onto the individual irq_chip’s to avoid disabling in HW in the first place.
this RFC patch tries to do same and address this problem in PDC irq_chip when control reaches at Step 2.4, to NOT disable at HW when the IRQ
is already marked wake up capable, it bails out at Step 2.4 in PDC irq_chip so that interrupt is left enabled in HW at both PDC and GIC HW level.
> so we should be able to have some suspend entry hook in pdc that
> enables the irq in the PDC if it's in the wakeup bitmap. Then on the
> path to suspend the GIC can lose power at any point after we enable the
> wakeup path in PDC and then the system should resume and get the
> interrupt through the resend mechanism.
I thought of this to introduce suspend hook in PDC which decides to keep wake up marked irq enabled at PDC.
But then someone need to keep it enabled at GIC as well.
PDC does not directly forward IRQ to CPU. PDC brings SoC out of low power mode where GIC does
not have power cut, it replays the interrupt at GIC in HW and that leads to forwarding interrupt
to CPU and resume from low power mode.
So PDC and GIC HW status should need to be in sync.
Thanks,
Maulik
--
QUALCOMM INDIA, on behalf of Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, hosted by The Linux Foundation
