Re: [v2] i2c: mediatek: Move suspend and resume handling to NOIRQ phase

[Qii Wang]

On Mon, 2020-12-14 at 22:08 +0200, Grygorii Strashko wrote:
> 
> On 14/12/2020 10:48, Qii Wang wrote:
> > On Thu, 2020-12-10 at 15:03 +0200, Grygorii Strashko wrote:
> >>
> >> On 10/12/2020 03:56, Qii Wang wrote:
> >>> On Mon, 2020-12-07 at 18:35 +0200, Grygorii Strashko wrote:
> >>>>
> >>>>>
> >>>>> On Thu, 2020-12-03 at 10:01 +0200, Grygorii Strashko wrote:
> >>>>>>
> >>>>>> On 03/12/2020 03:25, Qii Wang wrote:
> >>>>>>> On Wed, 2020-12-02 at 16:35 +0100, Wolfram Sang wrote:
> >>>>>>>> Hi,
> >>>>>>>>
> >>>>>>>>> Some i2c device driver indirectly uses I2C driver when it is now
> >>>>>>>>> being suspended. The i2c devices driver is suspended during the
> >>>>>>>>> NOIRQ phase and this cannot be changed due to other dependencies.
> >>>>>>>>> Therefore, we also need to move the suspend handling for the I2C
> >>>>>>>>> controller driver to the NOIRQ phase as well.
> >>>>>>>>>
> >>>>>>>>> Signed-off-by: Qii Wang <qii.wang@xxxxxxxxxxxx>
> >>>>>>>>
> >>>>>>>> Is this a bugfix and should go into 5.10? Or can it wait for 5.11?
> >>>>>>>>
> >>>>>>>
> >>>>>>> Yes, Can you help to apply it into 5.10? Thanks
> >>>>>>
> >>>>>> To be honest if you still do have any i2c device which accessing i2c buss after _noirq
> >>>>>> stage and your driver does not implement .master_xfer_atomic() - you definitely have a bigger problem.
> >>>>>> So adding IRQF_NO_SUSPEND sound like a hack and probably works just by luck.
> >>>>>>
> >>>>>
> >>>>> At present, it is only a problem caused by missing interrupts,
> >>>>> and .master_xfer_atomic() just a implement in polling mode. Why not set
> >>>>> the interrupt to a state that can always be triggered?
> >>>>>
> >>>>>
> >>>>
> >>>> Because you must not use any IRQ driven operations after _noirq suspend state as it might (and most probably will)
> >>>> cause unpredictable behavior later  in suspend_enter():
> >>>>
> >>>> 	arch_suspend_disable_irqs();
> >>>> 	BUG_ON(!irqs_disabled());
> >>>> ^after this point any IRQ driven I2C transfer will cause IRQ to be re-enabled
> >>>>
> >>>> if you need  turn off device from platform callbacks -  .master_xfer_atomic() has to be implemented and used.
> >>>>     
> >>> Maybe my comment is a bit disturbing.Our purpose is not to call i2c and
> >>> use interrupts after _noirq pauses.So We use
> >>> i2c_mark_adapter_suspended&i2c_mark_adapter_resumed to block these i2c
> >>> transfers， There will not have any IRQ driven I2C transfer after this
> >>> point:
> >>>           arch_suspend_disable_irqs();
> >>>           BUG_ON(!irqs_disabled());
> >>> But some device driver will do i2c transfer after
> >>> dpm_noirq_resume_devices in dpm_resume_noirq(PMSG_RESUME) when our
> >>> driver irq hasn't resume.
> >>> 	void dpm_resume_noirq(pm_message_t state)
> >>> 	{
> >>>           	dpm_noirq_resume_devices(state);
> >>
> >> Just to clarify. You have resume sequence in dpm_noirq_resume_devices
> >>    dpm_noirq_resume_devices -> resume I2C -> resume some device -> do i2c transfer after?
> >>
> > 
> > Yes.
> 
> huh. First consider IRQF_EARLY_RESUME - it's better, but still will be a hack
> 
There should be the same problem during the suspend process, So
IRQF_EARLY_RESUME should not be able to solve the problem.

> > 
> >> Is "some device" in Kernel mainline?
> >>
> > 
> > The problematic device driver is drivers/regulator/da9211-regulator.c in
> > Kernel mainline.
> 
> regulator is passive device, somebody should call it !?
> 
> And da9211-regulator IRQ handler should remain disabled till resume_device_irqs() call.
> 

Not only will i2c transfer be called in da9211-regulator IRQ handler,
but also other drivers will call da9211_buck_ops which containing i2c
transfers.

> note. regulator_class implements only
> 
> static const struct dev_pm_ops __maybe_unused regulator_pm_ops = {
> 	.suspend	= regulator_suspend,
> 	.resume		= regulator_resume,
> };
> 
> 
> > 
> >>>           	resume_device_irqs();
> >>>           	device_wakeup_disarm_wake_irqs();
> >>>           	cpuidle_resume();
> >>> 	}
> >>> .master_xfer_atomic() seems to be invalid for this question at this
> >>> time?
> >>>
> >>
> > 
>