[PATCH] drm/sti: Use 64-bit timestamps

[PATCH] drm/sti: Use 64-bit timestamps

[Tina Ruchandani]

'struct timespec' uses a 32-bit field for seconds, which
will overflow in year 2038 and beyond. This patch is part
of a larger attempt to remove instances of timeval, timespec
and time_t, all of which suffer from the y2038 issue, from the
kernel.

Signed-off-by: Tina Ruchandani <ruchandani.tina@gmail.com>
---
 drivers/gpu/drm/sti/sti_plane.c | 16 +++-------------
 drivers/gpu/drm/sti/sti_plane.h |  2 +-
 2 files changed, 4 insertions(+), 14 deletions(-)

diff --git a/drivers/gpu/drm/sti/sti_plane.c b/drivers/gpu/drm/sti/sti_plane.c
index f10c98d..3b46899 100644
--- a/drivers/gpu/drm/sti/sti_plane.c
+++ b/drivers/gpu/drm/sti/sti_plane.c
@@ -45,25 +45,15 @@ const char *sti_plane_to_str(struct sti_plane *plane)

 #define STI_FPS_INTERVAL_MS     3000

-static int sti_plane_timespec_ms_diff(struct timespec lhs, struct timespec rhs)
-{
-	struct timespec tmp_ts = timespec_sub(lhs, rhs);
-	u64 tmp_ns = (u64)timespec_to_ns(&tmp_ts);
-
-	do_div(tmp_ns, NSEC_PER_MSEC);
-
-	return (u32)tmp_ns;
-}
-
 void sti_plane_update_fps(struct sti_plane *plane,
 			  bool new_frame,
 			  bool new_field)
 {
-	struct timespec now;
+	ktime_t now;
 	struct sti_fps_info *fps;
 	int fpks, fipks, ms_since_last, num_frames, num_fields;

-	getrawmonotonic(&now);
+	now = ktime_get();

 	/* Compute number of frame updates */
 	fps = &plane->fps_info;
@@ -76,7 +66,7 @@ void sti_plane_update_fps(struct sti_plane *plane,
 		return;

 	fps->curr_frame_counter++;
-	ms_since_last = sti_plane_timespec_ms_diff(now, fps->last_timestamp);
+	ms_since_last = ktime_to_ms(ktime_sub(now, fps->last_timestamp));
 	num_frames = fps->curr_frame_counter - fps->last_frame_counter;

 	if (num_frames <= 0  || ms_since_last < STI_FPS_INTERVAL_MS)
diff --git a/drivers/gpu/drm/sti/sti_plane.h b/drivers/gpu/drm/sti/sti_plane.h
index c50a3b9..0a64eb0 100644
--- a/drivers/gpu/drm/sti/sti_plane.h
+++ b/drivers/gpu/drm/sti/sti_plane.h
@@ -57,7 +57,7 @@ struct sti_fps_info {
 	unsigned int last_frame_counter;
 	unsigned int curr_field_counter;
 	unsigned int last_field_counter;
-	struct timespec last_timestamp;
+	ktime_t	     last_timestamp;
 	char fps_str[FPS_LENGTH];
 	char fips_str[FPS_LENGTH];
 };
--
2.8.0.rc3.226.g39d4020

Re: [Y2038] [PATCH] drm/sti: Use 64-bit timestamps

[Arnd Bergmann]

On Wednesday 13 April 2016 02:28:02 Tina Ruchandani wrote:
> 'struct timespec' uses a 32-bit field for seconds, which
> will overflow in year 2038 and beyond. This patch is part
> of a larger attempt to remove instances of timeval, timespec
> and time_t, all of which suffer from the y2038 issue, from the
> kernel.
> 
> Signed-off-by: Tina Ruchandani <ruchandani.tina@gmail.com>

Looks good in principle. Two small points:

>  void sti_plane_update_fps(struct sti_plane *plane,
>  			  bool new_frame,
>  			  bool new_field)
>  {
> -	struct timespec now;
> +	ktime_t now;
>  	struct sti_fps_info *fps;
>  	int fpks, fipks, ms_since_last, num_frames, num_fields;
> 
> -	getrawmonotonic(&now);
> +	now = ktime_get();

It's unclear why the driver was using getrawmonotonic() here rather
than ktime_get_ts(). The code is fairly new, so Vincent can
probably explain this.

If it was intentional, we should use ktime_get_raw() instead of
ktime_get().

> @@ -76,7 +66,7 @@ void sti_plane_update_fps(struct sti_plane *plane,
>  		return;
> 
>  	fps->curr_frame_counter++;
> -	ms_since_last = sti_plane_timespec_ms_diff(now, fps->last_timestamp);
> +	ms_since_last = ktime_to_ms(ktime_sub(now, fps->last_timestamp));
>  	num_frames = fps->curr_frame_counter - fps->last_frame_counter;

This could be expressed in a more compact way using ktime_ms_delta().

	Arnd

Re: [Y2038] [PATCH] drm/sti: Use 64-bit timestamps

[Vincent ABRIOU]

On 04/17/2016 01:39 AM, Arnd Bergmann wrote:
> On Wednesday 13 April 2016 02:28:02 Tina Ruchandani wrote:
>> 'struct timespec' uses a 32-bit field for seconds, which
>> will overflow in year 2038 and beyond. This patch is part
>> of a larger attempt to remove instances of timeval, timespec
>> and time_t, all of which suffer from the y2038 issue, from the
>> kernel.
>>
>> Signed-off-by: Tina Ruchandani <ruchandani.tina@gmail.com>
>
> Looks good in principle. Two small points:
>
>>   void sti_plane_update_fps(struct sti_plane *plane,
>>   			  bool new_frame,
>>   			  bool new_field)
>>   {
>> -	struct timespec now;
>> +	ktime_t now;
>>   	struct sti_fps_info *fps;
>>   	int fpks, fipks, ms_since_last, num_frames, num_fields;
>>
>> -	getrawmonotonic(&now);
>> +	now = ktime_get();
>
> It's unclear why the driver was using getrawmonotonic() here rather
> than ktime_get_ts(). The code is fairly new, so Vincent can
> probably explain this.
>
> If it was intentional, we should use ktime_get_raw() instead of
> ktime_get().
>

getrawmonotonic comes from a legacy code so the use is not intentional.
Honestly, it is not clear to me the difference between monotonic and 
rawmonotonic. But in the debug context in which it is used, ktime_get 
and ktime_get_raw will deliver the same level of information we need. So 
implementation done by Tina is fine for me.

Vincent

>> @@ -76,7 +66,7 @@ void sti_plane_update_fps(struct sti_plane *plane,
>>   		return;
>>
>>   	fps->curr_frame_counter++;
>> -	ms_since_last = sti_plane_timespec_ms_diff(now, fps->last_timestamp);
>> +	ms_since_last = ktime_to_ms(ktime_sub(now, fps->last_timestamp));
>>   	num_frames = fps->curr_frame_counter - fps->last_frame_counter;
>
> This could be expressed in a more compact way using ktime_ms_delta().
>
> 	Arnd
>

Re: [Y2038] [PATCH] drm/sti: Use 64-bit timestamps

[Arnd Bergmann]

On Monday 18 April 2016 12:02:35 Vincent ABRIOU wrote:
> 
> getrawmonotonic comes from a legacy code so the use is not intentional.
> Honestly, it is not clear to me the difference between monotonic and 
> rawmonotonic. But in the debug context in which it is used, ktime_get 
> and ktime_get_raw will deliver the same level of information we need. So 
> implementation done by Tina is fine for me.
> 

Ok, cool, thanks for confirming!

FWIW, the best way I can see for illustrating the difference is that
rawmonotonic time is for things that should be synchronized with the
machines clock generators (e.g. A/V sync), while monotonic time is
for the case where you want to synchronize with another machine (or
the internet) that may have a slightly different clock generator but
uses NTP to correct for that.

In most cases the difference is irrelevant and we tend to use monotonic
time by default.

	Arnd