Re: [PATCH v2 1/2] leds: core: Introduce generic pattern interface

[Baolin Wang]

On 28 June 2018 at 18:24, Andy Shevchenko <andy.shevchenko@xxxxxxxxx> wrote:
> On Thu, Jun 28, 2018 at 12:18 PM, Baolin Wang <baolin.wang@xxxxxxxxxx> wrote:
>> On 28 June 2018 at 16:31, Andy Shevchenko <andy.shevchenko@xxxxxxxxx> wrote:
>>> On Thu, Jun 28, 2018 at 8:16 AM, Baolin Wang <baolin.wang@xxxxxxxxxx> wrote:
>
>>>> +What: /sys/class/leds/<led>/pattern
>>>> +Date: June 2018
>>>> +KernelVersion: 4.18
>>>
>>> 4.19 ?
>>
>> I think this will be merged in 4.18.
>
> Is it bug fix? I don't see how it would make v4.18.

OK. 4.19 is fine.

>>>> +static ssize_t pattern_show(struct device *dev,
>>>> +                           struct device_attribute *attr, char *buf)
>>>> +{
>>>> +       struct led_classdev *led_cdev = dev_get_drvdata(dev);
>>>> +       struct led_pattern *pattern;
>>>> +       size_t offset = 0;
>>>> +       int count, n, i;
>>>
>>>> +       if (!led_cdev->pattern_get)
>>>> +               return -EOPNOTSUPP;
>>>> +
>>>
>>> Perhaps just hide an attribute completely?
>>
>> Driver need implement the pattern_get() interface, otherwise we can
>> not get any available pattern values to show.
>
> It doesn't contradict with what I said. I proposed to just hide an
> attribute from sysfs completely if there is no such callbacks
> available.

Ah, I got your points now. But the pattern is RW and we can not hide
it if only pattern_get() is not supplied without considering
pattern_set().

>
>>>> +       pattern = led_cdev->pattern_get(led_cdev, &count);
>>>> +       if (IS_ERR_OR_NULL(pattern))
>>>> +               return PTR_ERR(pattern);
>>>
>>> Hmm.. Here you shadow NULL case by returning 0.
>>> Even if it's correct behaviour IS_ERR_OR_NULL is a beast to hide such
>>> subtle detail.
>>>
>>> It also would be good idea to check for count == 0 and bail out
>>> immediately. Otherwise you will print an extra blank line.
>>
>> We can not check count, since count can be not initialized if failed
>> to call pattern_get(). So maybe force user to return error pointer,
>> and we just check like:
>> if (IS_ERR(pattern))
>>         return PTR_ERR(pattern);
>
> My question is can be counter 0 by some reason?

I am not sure, but I still think checking the return error pointer is
the common way to handle.

>>>> +       for (i = 0; i < count; i++) {
>>>> +               n = snprintf(buf + offset, PAGE_SIZE - offset, "%d %d",
>>>> +                            pattern[i].brightness, pattern[i].delta_t);
>>>> +
>>>
>>>> +               if (offset + n >= PAGE_SIZE)
>>>> +                       goto err_nospc;
>>>
>>>> +
>>>> +               offset += n;
>>>> +
>>>
>>>> +               if (i < count - 1)
>>>> +                       buf[offset++] = ' ';
>>>
>>> You might add this to the end of above format string and remove this
>>> conditional completely...
>>
>> Hmmm, I do not think we need add one extra ' ' to the end of format string.
>
> Why not? You do it anyway for all except last, but...

Previous whitespace is used to break up the brightness value and
delta_t value, but it is useless to add one extra ' ' to the end of
format string.

>
>>>> +       buf[offset++] = '\n';
>>>
>>> ...and use here something like
>>>
>>> buf[offset - 1] = '\n';
>>
>> I don't think so. We need increase the offset value at the same time.
>
> Nope, you don't need it. here we replace trailing space by '\n'.

But why we need add one extra useless space?

>
>>> (we have such patterns in the kernel)
>
> ...look at existing patterns in the kernel.
>
>>>> +       /* Trim trailing newline */
>>>> +       s[strcspn(s, "\n")] = '\0';
>>>
>>> It's usually done via strstrip().
>>>
>>> sbegin = kstrndup();
>>> ...
>>>
>>> s = strstrip(sbegin);
>>
>> Good idea, will change.
>
> Replying to your another message. And who cares about leading spaces?
> Is it wrong to have them? Why?

OK, will use strstrip() to remove leading and trailing space firstly.

>
>>>> +       /* Parse out the brightness & delta_t touples */
>>>> +       while ((elem = strsep(&s, " ")) != NULL) {
>>>> +               ret = kstrtoul(elem, 10, &val);
>>>> +               if (ret)
>>>> +                       goto out;
>>>> +
>>>
>>>> +               if (odd) {
>>>
>>> This is effectivelly if (len % 2 == 0)
>>
>> It is incorrect, we can not use len to decide the value is brightness
>> or delta. Here logical is to make sure we must keep <brightness
>> delta>, <brightness, delta> ......
>
> Right, the problem is the format itself I suppose.
> It's too error prone for one attribute.

The format was decided by previous discussion, so I still like to keep
the original approach here.

>
> What about to change it like
>
> "brightness delta[, brightness delta[, ...]]"
>
> and then you do
>
> ...elem = strsep(&s, ",")...
>
> sscanf("%d %d") == 2
>
> ?
>
>>>> +                       pattern[len].brightness = val;
>>>> +               } else {
>>>> +                       pattern[len].delta_t = val;
>>>> +                       len++;
>>>> +               }
>>>> +
>>>> +               odd = !odd;
>>>> +       }
>>>> +
>>>> +       /*
>>>> +        * Fail if we didn't find any data points or last data point was partial
>>>> +        */
>>>> +       if (!len || !odd) {
>>>> +               ret = -EINVAL;
>>>> +               goto out;
>>>> +       }
>>>
>>> For partial data can we return different error code?
>>> Does it make sense?
>>
>> Sorry I did not get you here. If user set incorrect pattern values, I
>> think '-EINVAL' is suitable.
>
> OK.
>
> --
> With Best Regards,
> Andy Shevchenko



-- 
Baolin.wang
Best Regards