Re: [RFC 08/11] gpiolib: cdev: Add hardware timestamp clock type

[Dipen Patel <dipenp@nvidia.com>]

On 7/1/21 7:24 AM, Kent Gibson wrote:
> On Fri, Jun 25, 2021 at 04:55:29PM -0700, Dipen Patel wrote:
>> This patch adds new clock type for the GPIO controller which can
>> timestamp gpio lines using hardware means. To expose such
>> functionalities to the userspace, code has been added in this patch
>> where during line create call, it checks for new clock type and if
>> requested, calls hardware timestamp related API from gpiolib.c.
>> During line change event, it retrieves timestamp in nano seconds by
>> calling gpiod_get_hw_timestamp API from gpiolib.c. At the line release,
>> it disables this functionality by calling gpiod_hw_timestamp_control.
>>
>> Signed-off-by: Dipen Patel <dipenp@nvidia.com>
>> ---
>>  drivers/gpio/gpiolib-cdev.c | 65 +++++++++++++++++++++++++++++++++++--
>>  include/uapi/linux/gpio.h   |  1 +
>>  2 files changed, 64 insertions(+), 2 deletions(-)
>>
>> diff --git a/drivers/gpio/gpiolib-cdev.c b/drivers/gpio/gpiolib-cdev.c
>> index 1631727bf0da..9f98c727e937 100644
>> --- a/drivers/gpio/gpiolib-cdev.c
>> +++ b/drivers/gpio/gpiolib-cdev.c
>> @@ -518,6 +518,7 @@ struct linereq {
>>  	 GPIO_V2_LINE_DRIVE_FLAGS | \
>>  	 GPIO_V2_LINE_EDGE_FLAGS | \
>>  	 GPIO_V2_LINE_FLAG_EVENT_CLOCK_REALTIME | \
>> +	 GPIO_V2_LINE_FLAG_EVENT_CLOCK_HARDWARE | \
>>  	 GPIO_V2_LINE_BIAS_FLAGS)
>>  
>>  static void linereq_put_event(struct linereq *lr,
>> @@ -540,9 +541,20 @@ static void linereq_put_event(struct linereq *lr,
>>  
>>  static u64 line_event_timestamp(struct line *line)
>>  {
>> +	bool block;
>> +
>>  	if (test_bit(FLAG_EVENT_CLOCK_REALTIME, &line->desc->flags))
>>  		return ktime_get_real_ns();
>>  
>> +	if (test_bit(FLAG_EVENT_CLOCK_HARDWARE, &line->desc->flags)) {
>> +		if (irq_count())
>> +			block = false;
>> +		else
>> +			block = true;
>> +
>> +		return gpiod_get_hw_timestamp(line->desc, block);
>> +	}
>> +
> Use in_task() instead of block?
yes, will change to in_task.
>
>>  	return ktime_get_ns();
>>  }
>>  
>> @@ -828,6 +840,7 @@ static int edge_detector_setup(struct line *line,
>>  		return ret;
>>  
>>  	line->irq = irq;
>> +
>>  	return 0;
>>  }
>>  
> Remove gratuitous whitespace changes.
> If you dislike the formatting then suggest it in a separate patch.
I will remove this space.
>
>> @@ -891,7 +904,6 @@ static int gpio_v2_line_flags_validate(u64 flags)
>>  	/* Return an error if an unknown flag is set */
>>  	if (flags & ~GPIO_V2_LINE_VALID_FLAGS)
>>  		return -EINVAL;
>> -
>>  	/*
>>  	 * Do not allow both INPUT and OUTPUT flags to be set as they are
>>  	 * contradictory.
>> @@ -900,6 +912,14 @@ static int gpio_v2_line_flags_validate(u64 flags)
>>  	    (flags & GPIO_V2_LINE_FLAG_OUTPUT))
>>  		return -EINVAL;
>>  
> Same here.
>
>> +	/*
>> +	 * Do not mix with any other clocks if hardware assisted timestamp is
>> +	 * asked.
>> +	 */
>> +	if ((flags & GPIO_V2_LINE_FLAG_EVENT_CLOCK_REALTIME) &&
>> +	    (flags & GPIO_V2_LINE_FLAG_EVENT_CLOCK_HARDWARE))
>> +		return -EINVAL;
>> +
> The comment is very hw timestamp centric. It should just be something
> along the lines of "only allow one event clock source".
Sure, will change it.
>
>>  	/* Edge detection requires explicit input. */
>>  	if ((flags & GPIO_V2_LINE_EDGE_FLAGS) &&
>>  	    !(flags & GPIO_V2_LINE_FLAG_INPUT))
>> @@ -992,6 +1012,8 @@ static void gpio_v2_line_config_flags_to_desc_flags(u64 flags,
>>  
>>  	assign_bit(FLAG_EVENT_CLOCK_REALTIME, flagsp,
>>  		   flags & GPIO_V2_LINE_FLAG_EVENT_CLOCK_REALTIME);
>> +	assign_bit(FLAG_EVENT_CLOCK_HARDWARE, flagsp,
>> +		   flags & GPIO_V2_LINE_FLAG_EVENT_CLOCK_HARDWARE);
>>  }
>>  
>>  static long linereq_get_values(struct linereq *lr, void __user *ip)
>> @@ -1139,6 +1161,18 @@ static long linereq_set_config_unlocked(struct linereq *lr,
>>  			int val = gpio_v2_line_config_output_value(lc, i);
>>  
>>  			edge_detector_stop(&lr->lines[i]);
>> +
>> +			/*
>> +			 * Assuming line was input before and hardware
>> +			 * assisted timestamp only timestamps the input
>> +			 * lines.
>> +			 */
>> +			if (gpiod_is_hw_timestamp_enabled(desc)) {
>> +				ret = gpiod_hw_timestamp_control(desc, false);
>> +				if (ret)
>> +					return ret;
>> +			}
>> +
> So if you fail to disable the hw timestamp then you fail the set_config?
> Does that make sense?
> It should be impossible to fail, as per the preceding edge_detector_stop(),
> or any failure in this context is irrelevant and so can be ignored.

I am planning to remove is_hw_timestamp* API as it is not needed.

I will also remove ret check from timestamp_control API as it is not needed.

>
>>  			ret = gpiod_direction_output(desc, val);
>>  			if (ret)
>>  				return ret;
>> @@ -1152,6 +1186,13 @@ static long linereq_set_config_unlocked(struct linereq *lr,
>>  					polarity_change);
>>  			if (ret)
>>  				return ret;
>> +
>> +			/* Check if new config sets hardware assisted clock */
>> +			if (flags & GPIO_V2_LINE_FLAG_EVENT_CLOCK_HARDWARE) {
>> +				ret = gpiod_hw_timestamp_control(desc, true);
>> +				if (ret)
>> +					return ret;
>> +			}
>>  		}
>>  
> The error code here can come from the pinctrl timestamp_control(), so it
> should be sanitised before being returned to userspace.

I do not understand what do you mean by sanitise. I just followed what

gpiod_direction_output did just above which also returns ret from gpio

driver code similar to timestamp_control API.

>
>>  		blocking_notifier_call_chain(&desc->gdev->notifier,
>> @@ -1281,8 +1322,12 @@ static void linereq_free(struct linereq *lr)
>>  
>>  	for (i = 0; i < lr->num_lines; i++) {
>>  		edge_detector_stop(&lr->lines[i]);
>> -		if (lr->lines[i].desc)
>> +		if (lr->lines[i].desc) {
>> +			if (gpiod_is_hw_timestamp_enabled(lr->lines[i].desc))
>> +				gpiod_hw_timestamp_control(lr->lines[i].desc,
>> +							   false);
>>  			gpiod_free(lr->lines[i].desc);
>> +		}
> Potential race on gpiod_is_hw_timestamp_enabled() and the call to
> gpiod_hw_timestamp_control()?
> Why not put the gpiod_is_hw_timestamp_enabled() check inside
> gpiod_hw_timestamp_control()?
>
> And the gpiod_hw_timestamp_control() call should be moved inside
> gpiod_free(), or more correctly gpiod_free_commit().
> i.e. whenever you free the gpio you release any associated hw timestamp.

I am planning to remove is_hw_timestamp* API, that should take care

of race condition. For gpiod_free comment, I had thought about it before

but then ruled out as it would mean that for all the clients who did not

register with HTE, during their gpiod_free call, it has to make unncessary

call into HTE, however HTE release_ts has necessary checks which will return

without doing anything. Let me know if you still think to move it in gpiod_free.

>
>>  	}
>>  	kfifo_free(&lr->events);
>>  	kfree(lr->label);
>> @@ -1409,6 +1454,15 @@ static int linereq_create(struct gpio_device *gdev, void __user *ip)
>>  					flags & GPIO_V2_LINE_EDGE_FLAGS);
>>  			if (ret)
>>  				goto out_free_linereq;
>> +
>> +			/*
>> +			 * Check if hardware assisted timestamp is requested
>> +			 */
>> +			if (flags & GPIO_V2_LINE_FLAG_EVENT_CLOCK_HARDWARE) {
>> +				ret = gpiod_hw_timestamp_control(desc, true);
>> +				if (ret)
>> +					goto out_free_linereq;
>> +			}
>>  		}
>>  
> Comment can fit on one line, and probably isn't even necessary - the
> code is clear enough.
I will remove comment.
>
>>  		blocking_notifier_call_chain(&desc->gdev->notifier,
>> @@ -1956,8 +2010,15 @@ static void gpio_desc_to_lineinfo(struct gpio_desc *desc,
>>  	if (test_bit(FLAG_EDGE_FALLING, &desc->flags))
>>  		info->flags |= GPIO_V2_LINE_FLAG_EDGE_FALLING;
>>  
>> +	/*
>> +	 * Practically it is possible that user will want both the real time
>> +	 * and hardware timestamps on GPIO events, for now however lets just
>> +	 * work with either clocks
>> +	 */
>>  	if (test_bit(FLAG_EVENT_CLOCK_REALTIME, &desc->flags))
>>  		info->flags |= GPIO_V2_LINE_FLAG_EVENT_CLOCK_REALTIME;
>> +	else if (test_bit(FLAG_EVENT_CLOCK_HARDWARE, &desc->flags))
>> +		info->flags |= GPIO_V2_LINE_FLAG_EVENT_CLOCK_HARDWARE;
>>
> If there is any need or intent to support multiple clock sources then
> avoid creeping API changes and add it now.
> Either way, drop the comment.
I will remove comment in next RFC.
>
>>  	debounce_period_us = READ_ONCE(desc->debounce_period_us);
>>  	if (debounce_period_us) {
>> diff --git a/include/uapi/linux/gpio.h b/include/uapi/linux/gpio.h
>> index eaaea3d8e6b4..d360545b4c21 100644
>> --- a/include/uapi/linux/gpio.h
>> +++ b/include/uapi/linux/gpio.h
>> @@ -80,6 +80,7 @@ enum gpio_v2_line_flag {
>>  	GPIO_V2_LINE_FLAG_BIAS_PULL_DOWN	= _BITULL(9),
>>  	GPIO_V2_LINE_FLAG_BIAS_DISABLED		= _BITULL(10),
>>  	GPIO_V2_LINE_FLAG_EVENT_CLOCK_REALTIME	= _BITULL(11),
>> +	GPIO_V2_LINE_FLAG_EVENT_CLOCK_HARDWARE	= _BITULL(12),
>>  };
>>  
>>  /**
>> -- 
>> 2.17.1
>>
> Cheers,
> Kent.