The exynos DRM driver uses real-time 'struct timeval' values
for exporting its timestamps to user space. This has multiple
problems:
1. signed seconds overflow in y2038
2. the 'struct timeval' definition is deprecated in the kernel
3. time may jump or go backwards after a 'settimeofday()' syscall
4. other DRM timestamps are in CLOCK_MONOTONIC domain, so they
can't be compared
5. exporting microseconds requires a division by 1000, which may
be slow on some architectures.
The code existed in two places before, but the IPP portion was
removed in 8ded59413ccc ("drm/exynos: ipp: Remove Exynos DRM
IPP subsystem"), so we no longer need to worry about it.
Ideally timestamps should just use 64-bit nanoseconds instead, but
of course we can't change that now. Instead, this tries to address
the first four points above by using monotonic 'timespec' values.
According to Tobias Jakobi, user space doesn't care about the
timestamp at the moment, so we can change the format. Even if
there is something looking at them, it will work just fine with
monotonic times as long as the application only looks at the
relative values between two events.
Link: https://patchwork.kernel.org/patch/10038593/
Cc: Tobias Jakobi <[email protected]>
Signed-off-by: Arnd Bergmann <[email protected]>
---
v2: rebased to what will be in 4.15, now that ipp is gone,
updated changelog text based on input from Tobias.
---
drivers/gpu/drm/exynos/exynos_drm_g2d.c | 6 +++---
1 file changed, 3 insertions(+), 3 deletions(-)
diff --git a/drivers/gpu/drm/exynos/exynos_drm_g2d.c b/drivers/gpu/drm/exynos/exynos_drm_g2d.c
index 2b8bf2dd6387..9effe40f5fa5 100644
--- a/drivers/gpu/drm/exynos/exynos_drm_g2d.c
+++ b/drivers/gpu/drm/exynos/exynos_drm_g2d.c
@@ -926,7 +926,7 @@ static void g2d_finish_event(struct g2d_data *g2d, u32 cmdlist_no)
struct drm_device *drm_dev = g2d->subdrv.drm_dev;
struct g2d_runqueue_node *runqueue_node = g2d->runqueue_node;
struct drm_exynos_pending_g2d_event *e;
- struct timeval now;
+ struct timespec64 now;
if (list_empty(&runqueue_node->event_list))
return;
@@ -934,9 +934,9 @@ static void g2d_finish_event(struct g2d_data *g2d, u32 cmdlist_no)
e = list_first_entry(&runqueue_node->event_list,
struct drm_exynos_pending_g2d_event, base.link);
- do_gettimeofday(&now);
+ ktime_get_ts64(&now);
e->event.tv_sec = now.tv_sec;
- e->event.tv_usec = now.tv_usec;
+ e->event.tv_usec = now.tv_nsec / NSEC_PER_USEC;
e->event.cmdlist_no = cmdlist_no;
drm_send_event(drm_dev, &e->base);
--
2.9.0
Hey Arnd,
looks good to me!
Reviewed-by: Tobias Jakobi <[email protected]>
- Tobias
Arnd Bergmann wrote:
> The exynos DRM driver uses real-time 'struct timeval' values
> for exporting its timestamps to user space. This has multiple
> problems:
>
> 1. signed seconds overflow in y2038
> 2. the 'struct timeval' definition is deprecated in the kernel
> 3. time may jump or go backwards after a 'settimeofday()' syscall
> 4. other DRM timestamps are in CLOCK_MONOTONIC domain, so they
> can't be compared
> 5. exporting microseconds requires a division by 1000, which may
> be slow on some architectures.
>
> The code existed in two places before, but the IPP portion was
> removed in 8ded59413ccc ("drm/exynos: ipp: Remove Exynos DRM
> IPP subsystem"), so we no longer need to worry about it.
>
> Ideally timestamps should just use 64-bit nanoseconds instead, but
> of course we can't change that now. Instead, this tries to address
> the first four points above by using monotonic 'timespec' values.
>
> According to Tobias Jakobi, user space doesn't care about the
> timestamp at the moment, so we can change the format. Even if
> there is something looking at them, it will work just fine with
> monotonic times as long as the application only looks at the
> relative values between two events.
>
> Link: https://patchwork.kernel.org/patch/10038593/
> Cc: Tobias Jakobi <[email protected]>
> Signed-off-by: Arnd Bergmann <[email protected]>
> ---
> v2: rebased to what will be in 4.15, now that ipp is gone,
> updated changelog text based on input from Tobias.
> ---
> drivers/gpu/drm/exynos/exynos_drm_g2d.c | 6 +++---
> 1 file changed, 3 insertions(+), 3 deletions(-)
>
> diff --git a/drivers/gpu/drm/exynos/exynos_drm_g2d.c b/drivers/gpu/drm/exynos/exynos_drm_g2d.c
> index 2b8bf2dd6387..9effe40f5fa5 100644
> --- a/drivers/gpu/drm/exynos/exynos_drm_g2d.c
> +++ b/drivers/gpu/drm/exynos/exynos_drm_g2d.c
> @@ -926,7 +926,7 @@ static void g2d_finish_event(struct g2d_data *g2d, u32 cmdlist_no)
> struct drm_device *drm_dev = g2d->subdrv.drm_dev;
> struct g2d_runqueue_node *runqueue_node = g2d->runqueue_node;
> struct drm_exynos_pending_g2d_event *e;
> - struct timeval now;
> + struct timespec64 now;
>
> if (list_empty(&runqueue_node->event_list))
> return;
> @@ -934,9 +934,9 @@ static void g2d_finish_event(struct g2d_data *g2d, u32 cmdlist_no)
> e = list_first_entry(&runqueue_node->event_list,
> struct drm_exynos_pending_g2d_event, base.link);
>
> - do_gettimeofday(&now);
> + ktime_get_ts64(&now);
> e->event.tv_sec = now.tv_sec;
> - e->event.tv_usec = now.tv_usec;
> + e->event.tv_usec = now.tv_nsec / NSEC_PER_USEC;
> e->event.cmdlist_no = cmdlist_no;
>
> drm_send_event(drm_dev, &e->base);
>
2018년 01월 18일 02:01에 Arnd Bergmann 이(가) 쓴 글:
> The exynos DRM driver uses real-time 'struct timeval' values
> for exporting its timestamps to user space. This has multiple
> problems:
>
> 1. signed seconds overflow in y2038
> 2. the 'struct timeval' definition is deprecated in the kernel
> 3. time may jump or go backwards after a 'settimeofday()' syscall
> 4. other DRM timestamps are in CLOCK_MONOTONIC domain, so they
> can't be compared
> 5. exporting microseconds requires a division by 1000, which may
> be slow on some architectures.
>
> The code existed in two places before, but the IPP portion was
> removed in 8ded59413ccc ("drm/exynos: ipp: Remove Exynos DRM
> IPP subsystem"), so we no longer need to worry about it.
>
> Ideally timestamps should just use 64-bit nanoseconds instead, but
> of course we can't change that now. Instead, this tries to address
> the first four points above by using monotonic 'timespec' values.
>
> According to Tobias Jakobi, user space doesn't care about the
> timestamp at the moment, so we can change the format. Even if
> there is something looking at them, it will work just fine with
> monotonic times as long as the application only looks at the
> relative values between two events.
>
> Link: https://patchwork.kernel.org/patch/10038593/
> Cc: Tobias Jakobi <[email protected]>
> Signed-off-by: Arnd Bergmann <[email protected]>
Picked it up.
Thanks,
Inki Dae
> ---
> v2: rebased to what will be in 4.15, now that ipp is gone,
> updated changelog text based on input from Tobias.
> ---
> drivers/gpu/drm/exynos/exynos_drm_g2d.c | 6 +++---
> 1 file changed, 3 insertions(+), 3 deletions(-)
>
> diff --git a/drivers/gpu/drm/exynos/exynos_drm_g2d.c b/drivers/gpu/drm/exynos/exynos_drm_g2d.c
> index 2b8bf2dd6387..9effe40f5fa5 100644
> --- a/drivers/gpu/drm/exynos/exynos_drm_g2d.c
> +++ b/drivers/gpu/drm/exynos/exynos_drm_g2d.c
> @@ -926,7 +926,7 @@ static void g2d_finish_event(struct g2d_data *g2d, u32 cmdlist_no)
> struct drm_device *drm_dev = g2d->subdrv.drm_dev;
> struct g2d_runqueue_node *runqueue_node = g2d->runqueue_node;
> struct drm_exynos_pending_g2d_event *e;
> - struct timeval now;
> + struct timespec64 now;
>
> if (list_empty(&runqueue_node->event_list))
> return;
> @@ -934,9 +934,9 @@ static void g2d_finish_event(struct g2d_data *g2d, u32 cmdlist_no)
> e = list_first_entry(&runqueue_node->event_list,
> struct drm_exynos_pending_g2d_event, base.link);
>
> - do_gettimeofday(&now);
> + ktime_get_ts64(&now);
> e->event.tv_sec = now.tv_sec;
> - e->event.tv_usec = now.tv_usec;
> + e->event.tv_usec = now.tv_nsec / NSEC_PER_USEC;
> e->event.cmdlist_no = cmdlist_no;
>
> drm_send_event(drm_dev, &e->base);
>