2018-01-30 10:43:38

by Philippe Cornu

[permalink] [raw]
Subject: [PATCH v2] drm/stm: drv: Improve data transfers

To optimize data transfers, align pitch on 128 bytes & height
on 4 bytes. This optimization is not applicable on hw without MMU.

Signed-off-by: Yannick Fertre <[email protected]>
Signed-off-by: Vincent Abriou <[email protected]>
Signed-off-by: Philippe Cornu <[email protected]>
---
Changes in v2: Rename stm_dumb_create() to stm_gem_cma_dumb_create() and
move CONFIG_MMU inside the function following comments from Benjamin
Gaignard.

drivers/gpu/drm/stm/drv.c | 20 +++++++++++++++++++-
1 file changed, 19 insertions(+), 1 deletion(-)

diff --git a/drivers/gpu/drm/stm/drv.c b/drivers/gpu/drm/stm/drv.c
index 8fe954c27fba..8bc7e8418b8d 100644
--- a/drivers/gpu/drm/stm/drv.c
+++ b/drivers/gpu/drm/stm/drv.c
@@ -31,6 +31,24 @@ static const struct drm_mode_config_funcs drv_mode_config_funcs = {
.atomic_commit = drm_atomic_helper_commit,
};

+static int stm_gem_cma_dumb_create(struct drm_file *file,
+ struct drm_device *dev,
+ struct drm_mode_create_dumb *args)
+{
+#ifdef CONFIG_MMU
+ unsigned int min_pitch = DIV_ROUND_UP(args->width * args->bpp, 8);
+
+ /*
+ * in order to optimize data transfer, pitch is aligned on
+ * 128 bytes, height is aligned on 4 bytes
+ */
+ args->pitch = roundup(min_pitch, 128);
+ args->height = roundup(args->height, 4);
+#endif
+
+ return drm_gem_cma_dumb_create_internal(file, dev, args);
+}
+
DEFINE_DRM_GEM_CMA_FOPS(drv_driver_fops);

static struct drm_driver drv_driver = {
@@ -44,7 +62,7 @@ static struct drm_driver drv_driver = {
.minor = 0,
.patchlevel = 0,
.fops = &drv_driver_fops,
- .dumb_create = drm_gem_cma_dumb_create,
+ .dumb_create = stm_gem_cma_dumb_create,
.prime_handle_to_fd = drm_gem_prime_handle_to_fd,
.prime_fd_to_handle = drm_gem_prime_fd_to_handle,
.gem_free_object_unlocked = drm_gem_cma_free_object,
--
2.15.1



2018-01-30 10:52:07

by Benjamin Gaignard

[permalink] [raw]
Subject: Re: [PATCH v2] drm/stm: drv: Improve data transfers

2018-01-30 11:42 GMT+01:00 Philippe Cornu <[email protected]>:
> To optimize data transfers, align pitch on 128 bytes & height
> on 4 bytes. This optimization is not applicable on hw without MMU.
>
> Signed-off-by: Yannick Fertre <[email protected]>
> Signed-off-by: Vincent Abriou <[email protected]>
> Signed-off-by: Philippe Cornu <[email protected]>

Applied on drm-misc-next.

Regards,
Benjamin
> ---
> Changes in v2: Rename stm_dumb_create() to stm_gem_cma_dumb_create() and
> move CONFIG_MMU inside the function following comments from Benjamin
> Gaignard.
>
> drivers/gpu/drm/stm/drv.c | 20 +++++++++++++++++++-
> 1 file changed, 19 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/gpu/drm/stm/drv.c b/drivers/gpu/drm/stm/drv.c
> index 8fe954c27fba..8bc7e8418b8d 100644
> --- a/drivers/gpu/drm/stm/drv.c
> +++ b/drivers/gpu/drm/stm/drv.c
> @@ -31,6 +31,24 @@ static const struct drm_mode_config_funcs drv_mode_config_funcs = {
> .atomic_commit = drm_atomic_helper_commit,
> };
>
> +static int stm_gem_cma_dumb_create(struct drm_file *file,
> + struct drm_device *dev,
> + struct drm_mode_create_dumb *args)
> +{
> +#ifdef CONFIG_MMU
> + unsigned int min_pitch = DIV_ROUND_UP(args->width * args->bpp, 8);
> +
> + /*
> + * in order to optimize data transfer, pitch is aligned on
> + * 128 bytes, height is aligned on 4 bytes
> + */
> + args->pitch = roundup(min_pitch, 128);
> + args->height = roundup(args->height, 4);
> +#endif
> +
> + return drm_gem_cma_dumb_create_internal(file, dev, args);
> +}
> +
> DEFINE_DRM_GEM_CMA_FOPS(drv_driver_fops);
>
> static struct drm_driver drv_driver = {
> @@ -44,7 +62,7 @@ static struct drm_driver drv_driver = {
> .minor = 0,
> .patchlevel = 0,
> .fops = &drv_driver_fops,
> - .dumb_create = drm_gem_cma_dumb_create,
> + .dumb_create = stm_gem_cma_dumb_create,
> .prime_handle_to_fd = drm_gem_prime_handle_to_fd,
> .prime_fd_to_handle = drm_gem_prime_fd_to_handle,
> .gem_free_object_unlocked = drm_gem_cma_free_object,
> --
> 2.15.1
>

2018-01-30 14:59:08

by Laurent Pinchart

[permalink] [raw]
Subject: Re: [PATCH v2] drm/stm: drv: Improve data transfers

Hi Philippe,

Thank you for the patch.

On Tuesday, 30 January 2018 12:42:00 EET Philippe Cornu wrote:
> To optimize data transfers, align pitch on 128 bytes & height
> on 4 bytes. This optimization is not applicable on hw without MMU.
>
> Signed-off-by: Yannick Fertre <[email protected]>
> Signed-off-by: Vincent Abriou <[email protected]>
> Signed-off-by: Philippe Cornu <[email protected]>
> ---
> Changes in v2: Rename stm_dumb_create() to stm_gem_cma_dumb_create() and
> move CONFIG_MMU inside the function following comments from Benjamin
> Gaignard.
>
> drivers/gpu/drm/stm/drv.c | 20 +++++++++++++++++++-
> 1 file changed, 19 insertions(+), 1 deletion(-)
>
> diff --git a/drivers/gpu/drm/stm/drv.c b/drivers/gpu/drm/stm/drv.c
> index 8fe954c27fba..8bc7e8418b8d 100644
> --- a/drivers/gpu/drm/stm/drv.c
> +++ b/drivers/gpu/drm/stm/drv.c
> @@ -31,6 +31,24 @@ static const struct drm_mode_config_funcs
> drv_mode_config_funcs = { .atomic_commit = drm_atomic_helper_commit,
> };
>
> +static int stm_gem_cma_dumb_create(struct drm_file *file,
> + struct drm_device *dev,
> + struct drm_mode_create_dumb *args)
> +{
> +#ifdef CONFIG_MMU
> + unsigned int min_pitch = DIV_ROUND_UP(args->width * args->bpp, 8);
> +
> + /*
> + * in order to optimize data transfer, pitch is aligned on
> + * 128 bytes, height is aligned on 4 bytes
> + */
> + args->pitch = roundup(min_pitch, 128);
> + args->height = roundup(args->height, 4);
> +#endif
> +
> + return drm_gem_cma_dumb_create_internal(file, dev, args);

In the !CONFIG_MMU case you now bypass the pitch and size calculations
performed by drm_gem_cma_dumb_create(), allowing userspace to allocate
arbitrarily large buffers. Is that intentional ?

> +}
> +
> DEFINE_DRM_GEM_CMA_FOPS(drv_driver_fops);
>
> static struct drm_driver drv_driver = {
> @@ -44,7 +62,7 @@ static struct drm_driver drv_driver = {
> .minor = 0,
> .patchlevel = 0,
> .fops = &drv_driver_fops,
> - .dumb_create = drm_gem_cma_dumb_create,
> + .dumb_create = stm_gem_cma_dumb_create,
> .prime_handle_to_fd = drm_gem_prime_handle_to_fd,
> .prime_fd_to_handle = drm_gem_prime_fd_to_handle,
> .gem_free_object_unlocked = drm_gem_cma_free_object,

--
Regards,

Laurent Pinchart


2018-01-30 15:02:10

by Laurent Pinchart

[permalink] [raw]
Subject: Re: [PATCH v2] drm/stm: drv: Improve data transfers

Hi Benjamin,

On Tuesday, 30 January 2018 12:51:25 EET Benjamin Gaignard wrote:
> 2018-01-30 11:42 GMT+01:00 Philippe Cornu <[email protected]>:
> > To optimize data transfers, align pitch on 128 bytes & height
> > on 4 bytes. This optimization is not applicable on hw without MMU.
> >
> > Signed-off-by: Yannick Fertre <[email protected]>
> > Signed-off-by: Vincent Abriou <[email protected]>
> > Signed-off-by: Philippe Cornu <[email protected]>
>
> Applied on drm-misc-next.

A 9 minutes review window is pretty small... I've been told before that I
should then review v1 instead of complaining about short review times for v2,
but in this case v2 introduces a bug that wasn't present in v1.

> > ---
> > Changes in v2: Rename stm_dumb_create() to stm_gem_cma_dumb_create() and
> >
> > move CONFIG_MMU inside the function following comments from Benjamin
> > Gaignard.
> >
> > drivers/gpu/drm/stm/drv.c | 20 +++++++++++++++++++-
> > 1 file changed, 19 insertions(+), 1 deletion(-)
> >
> > diff --git a/drivers/gpu/drm/stm/drv.c b/drivers/gpu/drm/stm/drv.c
> > index 8fe954c27fba..8bc7e8418b8d 100644
> > --- a/drivers/gpu/drm/stm/drv.c
> > +++ b/drivers/gpu/drm/stm/drv.c
> > @@ -31,6 +31,24 @@ static const struct drm_mode_config_funcs
> > drv_mode_config_funcs = {>
> > .atomic_commit = drm_atomic_helper_commit,
> >
> > };
> >
> > +static int stm_gem_cma_dumb_create(struct drm_file *file,
> > + struct drm_device *dev,
> > + struct drm_mode_create_dumb *args)
> > +{
> > +#ifdef CONFIG_MMU
> > + unsigned int min_pitch = DIV_ROUND_UP(args->width * args->bpp, 8);
> > +
> > + /*
> > + * in order to optimize data transfer, pitch is aligned on
> > + * 128 bytes, height is aligned on 4 bytes
> > + */
> > + args->pitch = roundup(min_pitch, 128);
> > + args->height = roundup(args->height, 4);
> > +#endif
> > +
> > + return drm_gem_cma_dumb_create_internal(file, dev, args);
> > +}
> > +
> >
> > DEFINE_DRM_GEM_CMA_FOPS(drv_driver_fops);
> >
> > static struct drm_driver drv_driver = {
> >
> > @@ -44,7 +62,7 @@ static struct drm_driver drv_driver = {
> >
> > .minor = 0,
> > .patchlevel = 0,
> > .fops = &drv_driver_fops,
> >
> > - .dumb_create = drm_gem_cma_dumb_create,
> > + .dumb_create = stm_gem_cma_dumb_create,
> >
> > .prime_handle_to_fd = drm_gem_prime_handle_to_fd,
> > .prime_fd_to_handle = drm_gem_prime_fd_to_handle,
> > .gem_free_object_unlocked = drm_gem_cma_free_object,
> >
> > --
> > 2.15.1

--
Regards,

Laurent Pinchart


2018-01-30 15:09:55

by Benjamin Gaignard

[permalink] [raw]
Subject: Re: [PATCH v2] drm/stm: drv: Improve data transfers

2018-01-30 15:58 GMT+01:00 Laurent Pinchart <[email protected]>:
> Hi Philippe,
>
> Thank you for the patch.
>
> On Tuesday, 30 January 2018 12:42:00 EET Philippe Cornu wrote:
>> To optimize data transfers, align pitch on 128 bytes & height
>> on 4 bytes. This optimization is not applicable on hw without MMU.
>>
>> Signed-off-by: Yannick Fertre <[email protected]>
>> Signed-off-by: Vincent Abriou <[email protected]>
>> Signed-off-by: Philippe Cornu <[email protected]>
>> ---
>> Changes in v2: Rename stm_dumb_create() to stm_gem_cma_dumb_create() and
>> move CONFIG_MMU inside the function following comments from Benjamin
>> Gaignard.
>>
>> drivers/gpu/drm/stm/drv.c | 20 +++++++++++++++++++-
>> 1 file changed, 19 insertions(+), 1 deletion(-)
>>
>> diff --git a/drivers/gpu/drm/stm/drv.c b/drivers/gpu/drm/stm/drv.c
>> index 8fe954c27fba..8bc7e8418b8d 100644
>> --- a/drivers/gpu/drm/stm/drv.c
>> +++ b/drivers/gpu/drm/stm/drv.c
>> @@ -31,6 +31,24 @@ static const struct drm_mode_config_funcs
>> drv_mode_config_funcs = { .atomic_commit = drm_atomic_helper_commit,
>> };
>>
>> +static int stm_gem_cma_dumb_create(struct drm_file *file,
>> + struct drm_device *dev,
>> + struct drm_mode_create_dumb *args)
>> +{
>> +#ifdef CONFIG_MMU
>> + unsigned int min_pitch = DIV_ROUND_UP(args->width * args->bpp, 8);
>> +
>> + /*
>> + * in order to optimize data transfer, pitch is aligned on
>> + * 128 bytes, height is aligned on 4 bytes
>> + */
>> + args->pitch = roundup(min_pitch, 128);
>> + args->height = roundup(args->height, 4);
>> +#endif
>> +
>> + return drm_gem_cma_dumb_create_internal(file, dev, args);
>
> In the !CONFIG_MMU case you now bypass the pitch and size calculations
> performed by drm_gem_cma_dumb_create(), allowing userspace to allocate
> arbitrarily large buffers. Is that intentional ?

My bad, I will send a fix

>
>> +}
>> +
>> DEFINE_DRM_GEM_CMA_FOPS(drv_driver_fops);
>>
>> static struct drm_driver drv_driver = {
>> @@ -44,7 +62,7 @@ static struct drm_driver drv_driver = {
>> .minor = 0,
>> .patchlevel = 0,
>> .fops = &drv_driver_fops,
>> - .dumb_create = drm_gem_cma_dumb_create,
>> + .dumb_create = stm_gem_cma_dumb_create,
>> .prime_handle_to_fd = drm_gem_prime_handle_to_fd,
>> .prime_fd_to_handle = drm_gem_prime_fd_to_handle,
>> .gem_free_object_unlocked = drm_gem_cma_free_object,
>
> --
> Regards,
>
> Laurent Pinchart
>

2018-01-30 15:12:31

by Laurent Pinchart

[permalink] [raw]
Subject: Re: [PATCH v2] drm/stm: drv: Improve data transfers

Hi Benjamin,

On Tuesday, 30 January 2018 17:08:48 EET Benjamin Gaignard wrote:
> 2018-01-30 15:58 GMT+01:00 Laurent Pinchart:
> > On Tuesday, 30 January 2018 12:42:00 EET Philippe Cornu wrote:
> >> To optimize data transfers, align pitch on 128 bytes & height
> >> on 4 bytes. This optimization is not applicable on hw without MMU.
> >>
> >> Signed-off-by: Yannick Fertre <[email protected]>
> >> Signed-off-by: Vincent Abriou <[email protected]>
> >> Signed-off-by: Philippe Cornu <[email protected]>
> >> ---
> >> Changes in v2: Rename stm_dumb_create() to stm_gem_cma_dumb_create() and
> >>
> >> move CONFIG_MMU inside the function following comments from Benjamin
> >> Gaignard.
> >>
> >> drivers/gpu/drm/stm/drv.c | 20 +++++++++++++++++++-
> >> 1 file changed, 19 insertions(+), 1 deletion(-)
> >>
> >> diff --git a/drivers/gpu/drm/stm/drv.c b/drivers/gpu/drm/stm/drv.c
> >> index 8fe954c27fba..8bc7e8418b8d 100644
> >> --- a/drivers/gpu/drm/stm/drv.c
> >> +++ b/drivers/gpu/drm/stm/drv.c
> >> @@ -31,6 +31,24 @@ static const struct drm_mode_config_funcs
> >> drv_mode_config_funcs = { .atomic_commit = drm_atomic_helper_commit,
> >>
> >> };
> >>
> >> +static int stm_gem_cma_dumb_create(struct drm_file *file,
> >> + struct drm_device *dev,
> >> + struct drm_mode_create_dumb *args)
> >> +{
> >> +#ifdef CONFIG_MMU
> >> + unsigned int min_pitch = DIV_ROUND_UP(args->width * args->bpp, 8);
> >> +
> >> + /*
> >> + * in order to optimize data transfer, pitch is aligned on
> >> + * 128 bytes, height is aligned on 4 bytes
> >> + */
> >> + args->pitch = roundup(min_pitch, 128);
> >> + args->height = roundup(args->height, 4);
> >> +#endif
> >> +
> >> + return drm_gem_cma_dumb_create_internal(file, dev, args);
> >
> > In the !CONFIG_MMU case you now bypass the pitch and size calculations
> > performed by drm_gem_cma_dumb_create(), allowing userspace to allocate
> > arbitrarily large buffers. Is that intentional ?
>
> My bad, I will send a fix

Thank you.

> >> +}
> >> +
> >>
> >> DEFINE_DRM_GEM_CMA_FOPS(drv_driver_fops);
> >>
> >> static struct drm_driver drv_driver = {
> >>
> >> @@ -44,7 +62,7 @@ static struct drm_driver drv_driver = {
> >>
> >> .minor = 0,
> >> .patchlevel = 0,
> >> .fops = &drv_driver_fops,
> >>
> >> - .dumb_create = drm_gem_cma_dumb_create,
> >> + .dumb_create = stm_gem_cma_dumb_create,
> >>
> >> .prime_handle_to_fd = drm_gem_prime_handle_to_fd,
> >> .prime_fd_to_handle = drm_gem_prime_fd_to_handle,
> >> .gem_free_object_unlocked = drm_gem_cma_free_object,

--
Regards,

Laurent Pinchart


2018-01-30 16:58:06

by Jani Nikula

[permalink] [raw]
Subject: Re: [PATCH v2] drm/stm: drv: Improve data transfers

On Tue, 30 Jan 2018, Laurent Pinchart <[email protected]> wrote:
> Hi Benjamin,
>
> On Tuesday, 30 January 2018 12:51:25 EET Benjamin Gaignard wrote:
>> 2018-01-30 11:42 GMT+01:00 Philippe Cornu <[email protected]>:
>> > To optimize data transfers, align pitch on 128 bytes & height
>> > on 4 bytes. This optimization is not applicable on hw without MMU.
>> >
>> > Signed-off-by: Yannick Fertre <[email protected]>
>> > Signed-off-by: Vincent Abriou <[email protected]>
>> > Signed-off-by: Philippe Cornu <[email protected]>
>>
>> Applied on drm-misc-next.
>
> A 9 minutes review window is pretty small... I've been told before that I
> should then review v1 instead of complaining about short review times for v2,
> but in this case v2 introduces a bug that wasn't present in v1.

IMO no patch ever needs to be applied this fast. Give more eyeballs a
chance.

BR,
Jani.


>
>> > ---
>> > Changes in v2: Rename stm_dumb_create() to stm_gem_cma_dumb_create() and
>> >
>> > move CONFIG_MMU inside the function following comments from Benjamin
>> > Gaignard.
>> >
>> > drivers/gpu/drm/stm/drv.c | 20 +++++++++++++++++++-
>> > 1 file changed, 19 insertions(+), 1 deletion(-)
>> >
>> > diff --git a/drivers/gpu/drm/stm/drv.c b/drivers/gpu/drm/stm/drv.c
>> > index 8fe954c27fba..8bc7e8418b8d 100644
>> > --- a/drivers/gpu/drm/stm/drv.c
>> > +++ b/drivers/gpu/drm/stm/drv.c
>> > @@ -31,6 +31,24 @@ static const struct drm_mode_config_funcs
>> > drv_mode_config_funcs = {>
>> > .atomic_commit = drm_atomic_helper_commit,
>> >
>> > };
>> >
>> > +static int stm_gem_cma_dumb_create(struct drm_file *file,
>> > + struct drm_device *dev,
>> > + struct drm_mode_create_dumb *args)
>> > +{
>> > +#ifdef CONFIG_MMU
>> > + unsigned int min_pitch = DIV_ROUND_UP(args->width * args->bpp, 8);
>> > +
>> > + /*
>> > + * in order to optimize data transfer, pitch is aligned on
>> > + * 128 bytes, height is aligned on 4 bytes
>> > + */
>> > + args->pitch = roundup(min_pitch, 128);
>> > + args->height = roundup(args->height, 4);
>> > +#endif
>> > +
>> > + return drm_gem_cma_dumb_create_internal(file, dev, args);
>> > +}
>> > +
>> >
>> > DEFINE_DRM_GEM_CMA_FOPS(drv_driver_fops);
>> >
>> > static struct drm_driver drv_driver = {
>> >
>> > @@ -44,7 +62,7 @@ static struct drm_driver drv_driver = {
>> >
>> > .minor = 0,
>> > .patchlevel = 0,
>> > .fops = &drv_driver_fops,
>> >
>> > - .dumb_create = drm_gem_cma_dumb_create,
>> > + .dumb_create = stm_gem_cma_dumb_create,
>> >
>> > .prime_handle_to_fd = drm_gem_prime_handle_to_fd,
>> > .prime_fd_to_handle = drm_gem_prime_fd_to_handle,
>> > .gem_free_object_unlocked = drm_gem_cma_free_object,
>> >
>> > --
>> > 2.15.1

--
Jani Nikula, Intel Open Source Technology Center