[v1] i2c: i2c-qcom-geni: Serve transfer during early resume stage

pm_runtime_get_sync() function fails during PM early resume and returning
-EACCES because runtime PM for the device is disabled at the early stage
causing i2c transfer to fail. Make changes to serve transfer with force
resume.

1. Register interrupt with IRQF_EARLY_RESUME and IRQF_NO_SUSPEND flags
   to avoid timeout of transfer when IRQ is not enabled during early stage.
2. Do force resume if pm_runtime_get_sync() is failing after system
   suspend when runtime PM is not enabled.
3. Increment power usage count after forced resume to balance
   it against regular runtime suspend.

Co-developed-by: Viken Dadhaniya <quic_vdadhani@quicinc.com>
Signed-off-by: Viken Dadhaniya <quic_vdadhani@quicinc.com>
Signed-off-by: Mukesh Kumar Savaliya <quic_msavaliy@quicinc.com>
---
 drivers/i2c/busses/i2c-qcom-geni.c | 55 ++++++++++++++++++++++++------
 1 file changed, 45 insertions(+), 10 deletions(-)

Thanks Andi !

On 3/30/2024 3:24 AM, Andi Shyti wrote:
> Hi Mukesh,
> 
> On Thu, Mar 28, 2024 at 06:07:43PM +0530, Mukesh Kumar Savaliya wrote:
>> pm_runtime_get_sync() function fails during PM early resume and returning
>> -EACCES because runtime PM for the device is disabled at the early stage
>> causing i2c transfer to fail. Make changes to serve transfer with force
>> resume.
>>
>> 1. Register interrupt with IRQF_EARLY_RESUME and IRQF_NO_SUSPEND flags
>>     to avoid timeout of transfer when IRQ is not enabled during early stage.
>> 2. Do force resume if pm_runtime_get_sync() is failing after system
>>     suspend when runtime PM is not enabled.
>> 3. Increment power usage count after forced resume to balance
>>     it against regular runtime suspend.
>>
>> Co-developed-by: Viken Dadhaniya <quic_vdadhani@quicinc.com>
>> Signed-off-by: Viken Dadhaniya <quic_vdadhani@quicinc.com>
>> Signed-off-by: Mukesh Kumar Savaliya <quic_msavaliy@quicinc.com>
> 
> Is this a fix? O mostly an improvement?
It's an improvement to serve transfer during early resume time. Doesn't 
fix any crash issue.
> 
> ...
> 
>>   	gi2c->err = 0;
>>   	reinit_completion(&gi2c->done);
>> -	ret = pm_runtime_get_sync(gi2c->se.dev);
>> -	if (ret < 0) {
>> -		dev_err(gi2c->se.dev, "error turning SE resources:%d\n", ret);
>> -		pm_runtime_put_noidle(gi2c->se.dev);
>> -		/* Set device in suspended since resume failed */
>> -		pm_runtime_set_suspended(gi2c->se.dev);
>> -		return ret;
>> +
>> +	if (!pm_runtime_enabled(dev) && gi2c->suspended) {
>> +		dev_dbg(gi2c->se.dev, "RT_PM disabled, Do force resume, usage_count:%d\n",
> 
> dev
> 
Sure, Done.

> nit: you could leave a space after the ':'.
> 
Sure, Done.

>> +			atomic_read(&dev->power.usage_count));
>> +		ret = geni_i2c_force_resume(gi2c);
>> +		if (ret)
>> +			return ret;
>> +	} else {
>> +		ret = pm_runtime_get_sync(dev);
>> +		if (ret == -EACCES && gi2c->suspended) {
>> +			dev_dbg(gi2c->se.dev, "PM get_sync() failed-%d, force resume\n", ret);
> 
> dev
> 
Done.
>> +			ret = geni_i2c_force_resume(gi2c);
>> +			if (ret)
>> +				return ret;
>> +		}
> 
> Are we missing some cases here? Do we need a few more else's?
> 
> Andi
> 
No More else cases required.
>>   	}
>>   
>>   	qcom_geni_i2c_conf(gi2c);
>> @@ -702,8 +730,15 @@ static int geni_i2c_xfer(struct i2c_adapter *adap,
>>   	else
>>   		ret = geni_i2c_fifo_xfer(gi2c, msgs, num);
>>   
>> -	pm_runtime_mark_last_busy(gi2c->se.dev);
>> -	pm_runtime_put_autosuspend(gi2c->se.dev);
>> +	if (!pm_runtime_enabled(dev) && !gi2c->suspended) {
>> +		pm_runtime_put_noidle(dev);
>> +		pm_runtime_set_suspended(dev);
> 
> this looks like repeated code, how about making it a function?
> 
Checked it, but seems we need additional flag too to be added at one 
place where as common part is 2 lines of code. If this code continues to 
be repeated, i can think of keeping under a wrapper function. Please let 
me know if still needed to make a separate function.
>> +		gi2c->suspended = 0;
>> +	} else {
>> +		pm_runtime_mark_last_busy(gi2c->se.dev);
>> +		pm_runtime_put_autosuspend(gi2c->se.dev);
>> +	}
>> +
>>   	gi2c->cur = NULL;
>>   	gi2c->err = 0;
>>   	return ret;
>> @@ -820,7 +855,7 @@ static int geni_i2c_probe(struct platform_device *pdev)
>>   	init_completion(&gi2c->done);
>>   	spin_lock_init(&gi2c->lock);
>>   	platform_set_drvdata(pdev, gi2c);
>> -	ret = devm_request_irq(dev, gi2c->irq, geni_i2c_irq, 0,
>> +	ret = devm_request_irq(dev, gi2c->irq, geni_i2c_irq, IRQF_EARLY_RESUME | IRQF_NO_SUSPEND,
> 
> Can you provide a bit more information on how an early interrupt
> would be handled once the device is resumed?
> 
Sure, Let me go little descriptive here.
Post system suspend when system resume starts (because of PCI/touch 
device wakeup interrupt or anything else).

1.That time i2c client device handler will directly start i2c transfer, 
but we could not serve transfer because device is runtime suspended and 
runtime PM remains disabled. This happens during early_resume callback 
time, hence we have added forced_resume()

Basically system is still suspended and not reached to the point of 
enabling runtime PM of the i2c master.

2. If the IRQF_EARLY_RESUME is not registered, then we get timeout as 
the interrupt is not enabled. kernel/irq/pm.c has function which Enables 
interrupt lines having IRQF_EARLY_RESUME set during syscore instead of 
device resume time.

irq_pm_syscore_resume() => resume_irqs() => __enable_irq()

> Thanks,
> Andi
> 
>>   			       dev_name(dev), gi2c);
>>   	if (ret) {
>>   		dev_err(dev, "Request_irq failed:%d: err:%d\n",
>> -- 
>> 2.25.1
>>

Thanks Bjorn for the review.

On 4/5/2024 8:40 AM, Bjorn Andersson wrote:
> On Thu, Mar 28, 2024 at 06:07:43PM +0530, Mukesh Kumar Savaliya wrote:
>> pm_runtime_get_sync() function fails during PM early resume and returning
>> -EACCES because runtime PM for the device is disabled at the early stage
>> causing i2c transfer to fail. Make changes to serve transfer with force
>> resume.
>>
>> 1. Register interrupt with IRQF_EARLY_RESUME and IRQF_NO_SUSPEND flags
>>     to avoid timeout of transfer when IRQ is not enabled during early stage.
>> 2. Do force resume if pm_runtime_get_sync() is failing after system
>>     suspend when runtime PM is not enabled.
>> 3. Increment power usage count after forced resume to balance
>>     it against regular runtime suspend.
>>
>> Co-developed-by: Viken Dadhaniya <quic_vdadhani@quicinc.com>
>> Signed-off-by: Viken Dadhaniya <quic_vdadhani@quicinc.com>
>> Signed-off-by: Mukesh Kumar Savaliya <quic_msavaliy@quicinc.com>
>> ---
>>   drivers/i2c/busses/i2c-qcom-geni.c | 55 ++++++++++++++++++++++++------
>>   1 file changed, 45 insertions(+), 10 deletions(-)
>>
>> diff --git a/drivers/i2c/busses/i2c-qcom-geni.c b/drivers/i2c/busses/i2c-qcom-geni.c
>> index da94df466e83..ed8201983a03 100644
>> --- a/drivers/i2c/busses/i2c-qcom-geni.c
>> +++ b/drivers/i2c/busses/i2c-qcom-geni.c
>> @@ -134,6 +134,8 @@ struct geni_i2c_clk_fld {
>>   	u8	t_cycle_cnt;
>>   };
>>   
>> +static int geni_i2c_runtime_resume(struct device *dev);
>> +
>>   /*
>>    * Hardware uses the underlying formula to calculate time periods of
>>    * SCL clock cycle. Firmware uses some additional cycles excluded from the
>> @@ -677,22 +679,48 @@ static int geni_i2c_fifo_xfer(struct geni_i2c_dev *gi2c,
>>   	return num;
>>   }
>>   
>> +static int geni_i2c_force_resume(struct geni_i2c_dev *gi2c)
>> +{
>> +	struct device *dev = gi2c->se.dev;
>> +	int ret;
>> +
>> +	ret = geni_i2c_runtime_resume(dev);
>> +	if (ret) {
>> +		dev_err(gi2c->se.dev, "Error turning SE resources:%d\n", ret);
> 
> "error turning"? How about "Failed to enable SE resources: %d\n"?
Sounds good, Thanks. Made this change.
> 
>> +		pm_runtime_put_noidle(dev);
>> +		pm_runtime_set_suspended(dev);
> 
> These two where the error-handling for pm_runtime_get_sync() failing, in
> one case below you get here after pm_runtime_get_sync() failed in the
> other you have not attempted to call that function.
> 
Hi Bjorn, could you please clarify further on this question ? I am not 
getting two instances for failure of pm_runtime_get_sync().
> Do you really get through this with things balanced?
We haven't seen imbalance in the usage count with added debug logs locally.
> 
>> +		return ret;
>> +	}
>> +	pm_runtime_get_noresume(dev);
>> +	pm_runtime_set_active(dev);
>> +	return ret;
> 
> ret is 0.
> 
>> +}
>> +
>>   static int geni_i2c_xfer(struct i2c_adapter *adap,
>>   			 struct i2c_msg msgs[],
>>   			 int num)
>>   {
>>   	struct geni_i2c_dev *gi2c = i2c_get_adapdata(adap);
>> +	struct device *dev = gi2c->se.dev;
>>   	int ret;
>>   
>>   	gi2c->err = 0;
>>   	reinit_completion(&gi2c->done);
>> -	ret = pm_runtime_get_sync(gi2c->se.dev);
>> -	if (ret < 0) {
>> -		dev_err(gi2c->se.dev, "error turning SE resources:%d\n", ret);
>> -		pm_runtime_put_noidle(gi2c->se.dev);
>> -		/* Set device in suspended since resume failed */
>> -		pm_runtime_set_suspended(gi2c->se.dev);
>> -		return ret;
>> +
>> +	if (!pm_runtime_enabled(dev) && gi2c->suspended) {
>> +		dev_dbg(gi2c->se.dev, "RT_PM disabled, Do force resume, usage_count:%d\n",
> 
> "RT_PM" is not a widely used abbreviation for this...
Agree. Modified as "Runtime PM is disabled"
> 
>> +			atomic_read(&dev->power.usage_count));
>> +		ret = geni_i2c_force_resume(gi2c);
>> +		if (ret)
>> +			return ret;
>> +	} else {
>> +		ret = pm_runtime_get_sync(dev);
>> +		if (ret == -EACCES && gi2c->suspended) {
>> +			dev_dbg(gi2c->se.dev, "PM get_sync() failed-%d, force resume\n", ret);
> 
> Different abbreviation, different formatting of error value in log
> line...
yes, improved as ""pm_runtime_get_sync() failed"
> 
>> +			ret = geni_i2c_force_resume(gi2c);
>> +			if (ret)
>> +				return ret;
> 
> This chunk looks identical to the chunk above, can this somehow be
> restructured to avoid the duplication?
I see different failure condition and then conditional call to 
geni_i2c_force_resume(). Could you please highlight exact optimization ?
> 
>> +		}
>>   	}
>>   
>>   	qcom_geni_i2c_conf(gi2c);
>> @@ -702,8 +730,15 @@ static int geni_i2c_xfer(struct i2c_adapter *adap,
>>   	else
>>   		ret = geni_i2c_fifo_xfer(gi2c, msgs, num);
>>   
>> -	pm_runtime_mark_last_busy(gi2c->se.dev);
>> -	pm_runtime_put_autosuspend(gi2c->se.dev);
> 
> I'm not sure I follow this part, please add a comment here for the
> future reader.
sure, Added comment in next patch.

> 
>> +	if (!pm_runtime_enabled(dev) && !gi2c->suspended) {
>> +		pm_runtime_put_noidle(dev);
>> +		pm_runtime_set_suspended(dev);
>> +		gi2c->suspended = 0;
> 
> Why alter suspended flag in xfer()?
> 
Because force_resume() makes it active with respect to the
> Regards,
> Bjorn
> 
>> +	} else {
>> +		pm_runtime_mark_last_busy(gi2c->se.dev);
>> +		pm_runtime_put_autosuspend(gi2c->se.dev);
>> +	}
>> +
>>   	gi2c->cur = NULL;
>>   	gi2c->err = 0;
>>   	return ret;
>> @@ -820,7 +855,7 @@ static int geni_i2c_probe(struct platform_device *pdev)
>>   	init_completion(&gi2c->done);
>>   	spin_lock_init(&gi2c->lock);
>>   	platform_set_drvdata(pdev, gi2c);
>> -	ret = devm_request_irq(dev, gi2c->irq, geni_i2c_irq, 0,
>> +	ret = devm_request_irq(dev, gi2c->irq, geni_i2c_irq, IRQF_EARLY_RESUME | IRQF_NO_SUSPEND,
>>   			       dev_name(dev), gi2c);
>>   	if (ret) {
>>   		dev_err(dev, "Request_irq failed:%d: err:%d\n",
>> -- 
>> 2.25.1
>>