Re: [PATCH v4 3/3] iio: magnetometer: Add driver support for PNI RM3100

[Song Qiang <songqiang1304521@gmail.com>]

On 2018/11/2 下午5:24, Jonathan Cameron wrote:
> On Fri, 2 Nov 2018 15:55:27 +0800
> Song Qiang <songqiang1304521@gmail.com> wrote:
>
>> On 2018/10/21 下午10:14, Jonathan Cameron wrote:
>>> On Thu, 18 Oct 2018 16:24:15 +0800
>>> Song Qiang <songqiang1304521@gmail.com> wrote:
>>>
>>> ...
>>>>>> +static irqreturn_t rm3100_trigger_handler(int irq, void *p)
>>>>>> +{
>>>>>> +	struct iio_poll_func *pf = p;
>>>>>> +	struct iio_dev *indio_dev = pf->indio_dev;
>>>>>> +	unsigned long scan_mask = *indio_dev->active_scan_mask;
>>>>>> +	unsigned int mask_len = indio_dev->masklength;
>>>>>> +	struct rm3100_data *data = iio_priv(indio_dev);
>>>>>> +	struct regmap *regmap = data->regmap;
>>>>>> +	int ret, i, bit;
>>>>>> +
>>>>>> +	mutex_lock(&data->lock);
>>>>>> +	switch (scan_mask) {
>>>>>> +	case BIT(0) | BIT(1) | BIT(2):
>>>>>> +		ret = regmap_bulk_read(regmap, RM3100_REG_MX2, data->buffer, 9);
>>>>>> +		mutex_unlock(&data->lock);
>>>>>> +		if (ret < 0)
>>>>>> +			goto done;
>>>>>> +	break;
>>>>>> +	case BIT(0) | BIT(1):
>>>>>> +		ret = regmap_bulk_read(regmap, RM3100_REG_MX2, data->buffer, 6);
>>>>>> +		mutex_unlock(&data->lock);
>>>>>> +		if (ret < 0)
>>>>>> +			goto done;
>>>>>> +	break;
>>>>>> +	case BIT(1) | BIT(2):
>>>>>> +		ret = regmap_bulk_read(regmap, RM3100_REG_MY2, data->buffer, 6);
>>>>>> +		mutex_unlock(&data->lock);
>>>>>> +		if (ret < 0)
>>>>>> +			goto done;
>>>>>> +	break;
>>>>> What about BIT(0) | BIT(2)?
>>>>>
>>>>> Now you can do it like you have here and on that one corner case let the iio core
>>>>> demux code take care of it, but then you will need to provide available_scan_masks
>>>>> so the core knows it needs to handle this case.
>>>>>      
>>>> This confused me a little. The available_scan_masks I was using is {BIT(0) |
>>>> BIT(1) | BIT(2), 0x0}. Apparently in this version of patch I would like it to
>>>> handle every circumstances like BIT(0), BIT(0) | BIT(2), BIT(1) | BIT(2), etc.
>>>> Since Phil mentioned he would like this to reduce bus usage as much as we can
>>>> and I want it, too, I think these three circumstances can be read consecutively
>>>> while others can be read one axis at a time. So I plan to let  BIT(0) | BIT(2)
>>>> fall into the 'default' section, which reads axis one by one.
>>>>
>>>> My question is, since this handles every possible combination, do I still need
>>>> to list every available scan masks in available_scan_masks?
>>> Ah. I see, I'd missed that the default was picking up that case as well as the
>>> single axes.   It would be interesting to sanity check if it is quicker on
>>> a 'typical' platform to do the all axis read for the BIT(0) | BIT(2) case
>>> and drop the middle value (which would be done using available scan_masks)
>>> or to just do two independent reads.
>>>
>>> (I would guess it is worth reading the 'dead' axis).
>>>   
>>>> All other problems will be fixed in the next patch.
>>>>
>>>> yours,
>>>>
>>>> Song Qiang
>>>>
>>>>
>>>> ...
>>> Thanks,
>>>
>>> Jonathan
>> I tested this two ways of getting data with the following code snippet:
>>
>>
>>       u8 buffer[9];
>>       struct timeval timebefore, timeafter;
>>
>>       do_gettimeofday(&timebefore);
>>       ret = regmap_bulk_read(regmap, RM3100_REG_MX2, buffer, 9);
>>       if (ret < 0)
>>           goto unlock_return;
>>       do_gettimeofday(&timeafter);
>>       printk(KERN_INFO "read with dead axis time: %ld",
>>              timeafter.tv_sec * 1000000 + timeafter.tv_usec -
>>              timebefore.tv_sec * 1000000 - timebefore.tv_usec);
>>       do_gettimeofday(&timebefore);
>>
>>       ret = regmap_bulk_read(regmap, RM3100_REG_MX2, buffer, 3);
>>       if (ret < 0)
>>           goto unlock_return;
>>       ret = regmap_bulk_read(regmap, RM3100_REG_MZ2, buffer + 6, 3);
>>       if (ret < 0)
>>           goto unlock_return;
>>       do_gettimeofday(&timeafter);
>>       printk(KERN_INFO "read two single axis time: %ld",
>>              timeafter.tv_sec * 1000000 + timeafter.tv_usec -
>>              timebefore.tv_sec * 1000000 - timebefore.tv_usec);
>>
>>
>> And get this result:
>>
>> [  161.264777] read with dead axis time: 883
>> [  161.270621] read two single axis time: 1359
>> [  161.575134] read with dead axis time: 852
>> [  161.580973] read two single axis time: 1356
>> [  161.895704] read with dead axis time: 854
>> [  161.903744] read two single axis time: 3540
>> [  162.223600] read with dead axis time: 853
>> [  162.229451] read two single axis time: 1363
>> [  162.591227] read with dead axis time: 850
>> [  162.597630] read two single axis time: 1555
>> [  162.920102] read with dead axis time: 852
>> [  162.926467] read two single axis time: 1534
>> [  163.303121] read with dead axis time: 881
>> [  163.308997] read two single axis time: 1390
>> [  163.711004] read with dead axis time: 861
>>
>>
>> It seems like you're right! Reading consecutively 9 bytes does save a lot time
>> compared to read 3 bytes twice.
>>
> I've done this stuff before ;)  We had this on the adis16365 parts back
> in the early days of IIO.  A worse case as it has a lot more channels,
> but otherwise similar from what I recall.
>
> It would be an interesting exercise to trace those paths and find out the
> balance between real hardware stuff we can't change and potential software
> overheads.
>
> Chances are this is mostly 'real' stuff though but would be great to
> confirm this. It's been on my list of things to do for years (not on
> this driver obviously but in general)...
>
> Jonathan
>
I think I can try to use ftrace to trace it's flow path on my platform. I don't 
familiar with it, may need some time.

yours,

Song Qiang