diff mbox series

[v3,8/8] drm/omap: add OMAP_BO flags to affect buffer allocation

Message ID 20191007112555.25278-9-jjhiblot@ti.com
State Superseded
Headers show
Series None | expand

Commit Message

Jean-Jacques Hiblot Oct. 7, 2019, 11:25 a.m. UTC
From: Tomi Valkeinen <tomi.valkeinen@ti.com>

On SoCs with DMM/TILER, we have two ways to allocate buffers: normal
dma_alloc or via DMM (which basically functions as an IOMMU). DMM can
map 128MB at a time, and we only map the DMM buffers when they are used
(i.e. not at alloc time). If DMM is present, omapdrm always uses DMM.

There are use cases that require lots of big buffers that are being used
at the same time by different IPs. At the moment the userspace has a
hard maximum of 128MB.

This patch adds three new flags that can be used by the userspace to
solve the situation:

OMAP_BO_MEM_CONTIG: The driver will use dma_alloc to get the memory.
This can be used to avoid DMM if the userspace knows it needs more than
128M of memory at the same time.

OMAP_BO_MEM_DMM: The driver will use DMM to get the memory. There's not
much use for this flag at the moment, as on platforms with DMM it is
used by default, but it's here for completeness.

OMAP_BO_MEM_PIN: The driver will pin the memory at alloc time, and keep
it pinned. This can be used to 1) get an error at alloc time if DMM
space is full, and 2) get rid of the constant pin/unpin operations which
may have some effect on performance.

If none of the flags are given, the behavior is the same as currently.

Signed-off-by: Tomi Valkeinen <tomi.valkeinen@ti.com>
---
 drivers/gpu/drm/omapdrm/omap_gem.c | 54 ++++++++++++++++++++++++++++--
 include/uapi/drm/omap_drm.h        |  9 +++++
 2 files changed, 61 insertions(+), 2 deletions(-)

Comments

Jean-Jacques Hiblot Oct. 8, 2019, 4:01 p.m. UTC | #1
On 07/10/2019 13:25, Jean-Jacques Hiblot wrote:
> From: Tomi Valkeinen <tomi.valkeinen@ti.com>
>
> On SoCs with DMM/TILER, we have two ways to allocate buffers: normal
> dma_alloc or via DMM (which basically functions as an IOMMU). DMM can
> map 128MB at a time, and we only map the DMM buffers when they are used
> (i.e. not at alloc time). If DMM is present, omapdrm always uses DMM.
>
> There are use cases that require lots of big buffers that are being used
> at the same time by different IPs. At the moment the userspace has a
> hard maximum of 128MB.
>
> This patch adds three new flags that can be used by the userspace to
> solve the situation:
>
> OMAP_BO_MEM_CONTIG: The driver will use dma_alloc to get the memory.
> This can be used to avoid DMM if the userspace knows it needs more than
> 128M of memory at the same time.
>
> OMAP_BO_MEM_DMM: The driver will use DMM to get the memory. There's not
> much use for this flag at the moment, as on platforms with DMM it is
> used by default, but it's here for completeness.
>
> OMAP_BO_MEM_PIN: The driver will pin the memory at alloc time, and keep
> it pinned. This can be used to 1) get an error at alloc time if DMM
> space is full, and 2) get rid of the constant pin/unpin operations which
> may have some effect on performance.
>
> If none of the flags are given, the behavior is the same as currently.
>
> Signed-off-by: Tomi Valkeinen <tomi.valkeinen@ti.com>
> ---
>   drivers/gpu/drm/omapdrm/omap_gem.c | 54 ++++++++++++++++++++++++++++--
>   include/uapi/drm/omap_drm.h        |  9 +++++
>   2 files changed, 61 insertions(+), 2 deletions(-)
>
> diff --git a/drivers/gpu/drm/omapdrm/omap_gem.c b/drivers/gpu/drm/omapdrm/omap_gem.c
> index e518d93ca6df..bf18dfe2b689 100644
> --- a/drivers/gpu/drm/omapdrm/omap_gem.c
> +++ b/drivers/gpu/drm/omapdrm/omap_gem.c
> @@ -1097,6 +1097,9 @@ void omap_gem_free_object(struct drm_gem_object *obj)
>   	list_del(&omap_obj->mm_list);
>   	mutex_unlock(&priv->list_lock);
>   
> +	if (omap_obj->flags & OMAP_BO_MEM_PIN)
> +		omap_gem_unpin_locked(obj);
> +
>   	/*
>   	 * We own the sole reference to the object at this point, but to keep
>   	 * lockdep happy, we must still take the omap_obj_lock to call
> @@ -1147,10 +1150,19 @@ static bool omap_gem_validate_flags(struct drm_device *dev, u32 flags)
>   		return false;
>   	}
>   
> +	if ((flags & OMAP_BO_MEM_CONTIG) && (flags & OMAP_BO_MEM_DMM))
> +		return false;
> +
> +	if ((flags & OMAP_BO_MEM_DMM) && !priv->usergart)
> +		return false;
> +
>   	if (flags & OMAP_BO_TILED_MASK) {
>   		if (!priv->usergart)
>   			return false;
>   
> +		if (flags & OMAP_BO_MEM_CONTIG)
> +			return false;
> +
>   		switch (flags & OMAP_BO_TILED_MASK) {
>   		case OMAP_BO_TILED_8:
>   		case OMAP_BO_TILED_16:
> @@ -1165,7 +1177,34 @@ static bool omap_gem_validate_flags(struct drm_device *dev, u32 flags)
>   	return true;
>   }
>   
> -/* GEM buffer object constructor */
> +/**
> + * omap_gem_new() - Create a new GEM buffer
> + * @dev: The DRM device
> + * @gsize: The requested size for the GEM buffer. If the buffer is tiled
> + *         (2D buffer), the size is a pair of values: height and width
> + *         expressed in pixels. If the buffers is not tiled, it is expressed
> + *         in bytes.
> + * @flags: Flags give additionnal information about the allocation:
> + *         OMAP_BO_TILED_x: use the TILER (2D buffers). The TILER container
> + *              unit can be 8, 16 or 32 bits. Cache is always disabled for
> + *              tiled buffers.
> + *         OMAP_BO_SCANOUT: Scannout buffer, consummable by the DSS
> + *         OMAP_BO_CACHED: Buffer CPU caching mode: cached
> + *         OMAP_BO_WC: Buffer CPU caching mode: write-combined
> + *         OMAP_BO_UNCACHED: Buffer CPU caching mode: uncached
> + *         OMAP_BO_MEM_CONTIG: The driver will use dma_alloc to get the memory.
> + *              This can be used to avoid DMM if the userspace knows it needs
> + *              more than 128M of memory at the same time.
> + *         OMAP_BO_MEM_DMM: The driver will use DMM to get the memory. There's
> + *              not much use for this flag at the moment, as on platforms with
> + *              DMM it is used by default, but it's here for completeness.
> + *         OMAP_BO_MEM_PIN: The driver will pin the memory at alloc time, and
> + *              keep it pinned. This can be used to 1) get an error at alloc
> + *              time if DMM space is full, and 2) get rid of the constant
> + *              pin/unpin operations which may have some effect on performance.
> + *
> + * Return: The GEM buffer or NULL if the allocation failed
> + */
>   struct drm_gem_object *omap_gem_new(struct drm_device *dev,
>   		union omap_gem_size gsize, u32 flags)
>   {
> @@ -1193,7 +1232,8 @@ struct drm_gem_object *omap_gem_new(struct drm_device *dev,
>   		 */
>   		flags &= ~(OMAP_BO_CACHED|OMAP_BO_WC|OMAP_BO_UNCACHED);
>   		flags |= tiler_get_cpu_cache_flags();
> -	} else if ((flags & OMAP_BO_SCANOUT) && !priv->has_dmm) {
> +	} else if ((flags & OMAP_BO_MEM_CONTIG) ||
> +		((flags & OMAP_BO_SCANOUT) && !priv->has_dmm)) {
>   		/*
>   		 * If we don't have DMM, we must allocate scanout buffers
>   		 * from contiguous DMA memory.
> @@ -1253,12 +1293,22 @@ struct drm_gem_object *omap_gem_new(struct drm_device *dev,
>   			goto err_release;
>   	}
>   
> +	if (flags & OMAP_BO_MEM_PIN) {
> +		ret = omap_gem_pin(obj, NULL);
> +		if (ret)
> +			goto err_free_dma;
> +	}
> +
>   	mutex_lock(&priv->list_lock);
>   	list_add(&omap_obj->mm_list, &priv->obj_list);
>   	mutex_unlock(&priv->list_lock);
>   
>   	return obj;
>   
> +err_free_dma:
> +	if (flags & OMAP_BO_MEM_DMA_API)
> +		dma_free_wc(dev->dev, size, omap_obj->vaddr,
> +			    omap_obj->dma_addr);
>   err_release:
>   	drm_gem_object_release(obj);
>   err_free:
> diff --git a/include/uapi/drm/omap_drm.h b/include/uapi/drm/omap_drm.h
> index 5a142fad473c..842d3180a442 100644
> --- a/include/uapi/drm/omap_drm.h
> +++ b/include/uapi/drm/omap_drm.h
> @@ -47,6 +47,15 @@ struct drm_omap_param {
>   #define OMAP_BO_UNCACHED	0x00000004
>   #define OMAP_BO_CACHE_MASK	0x00000006
>   
> +/* Force allocation from contiguous DMA memory */
> +#define OMAP_BO_MEM_CONTIG	0x00000008
> +
> +/* Force allocation via DMM */
> +#define OMAP_BO_MEM_DMM		0x00000010
> +
> +/* Pin the buffer when allocating and keep pinned */
> +#define OMAP_BO_MEM_PIN		0x00000020
> +
>   /* Use TILER for the buffer. The TILER container unit can be 8, 16 or 32 bits. */
>   #define OMAP_BO_TILED_8		0x00000100
>   #define OMAP_BO_TILED_16	0x00000200

I modified the patch to add comment following the kernel-doc style, so 
my Reviewed-by actually applies only to the code modifications, not the 
comment wording.

Reviewed-by: Jean-Jacques Hiblot <jjhiblot@ti.com>
diff mbox series

Patch

diff --git a/drivers/gpu/drm/omapdrm/omap_gem.c b/drivers/gpu/drm/omapdrm/omap_gem.c
index e518d93ca6df..bf18dfe2b689 100644
--- a/drivers/gpu/drm/omapdrm/omap_gem.c
+++ b/drivers/gpu/drm/omapdrm/omap_gem.c
@@ -1097,6 +1097,9 @@  void omap_gem_free_object(struct drm_gem_object *obj)
 	list_del(&omap_obj->mm_list);
 	mutex_unlock(&priv->list_lock);
 
+	if (omap_obj->flags & OMAP_BO_MEM_PIN)
+		omap_gem_unpin_locked(obj);
+
 	/*
 	 * We own the sole reference to the object at this point, but to keep
 	 * lockdep happy, we must still take the omap_obj_lock to call
@@ -1147,10 +1150,19 @@  static bool omap_gem_validate_flags(struct drm_device *dev, u32 flags)
 		return false;
 	}
 
+	if ((flags & OMAP_BO_MEM_CONTIG) && (flags & OMAP_BO_MEM_DMM))
+		return false;
+
+	if ((flags & OMAP_BO_MEM_DMM) && !priv->usergart)
+		return false;
+
 	if (flags & OMAP_BO_TILED_MASK) {
 		if (!priv->usergart)
 			return false;
 
+		if (flags & OMAP_BO_MEM_CONTIG)
+			return false;
+
 		switch (flags & OMAP_BO_TILED_MASK) {
 		case OMAP_BO_TILED_8:
 		case OMAP_BO_TILED_16:
@@ -1165,7 +1177,34 @@  static bool omap_gem_validate_flags(struct drm_device *dev, u32 flags)
 	return true;
 }
 
-/* GEM buffer object constructor */
+/**
+ * omap_gem_new() - Create a new GEM buffer
+ * @dev: The DRM device
+ * @gsize: The requested size for the GEM buffer. If the buffer is tiled
+ *         (2D buffer), the size is a pair of values: height and width
+ *         expressed in pixels. If the buffers is not tiled, it is expressed
+ *         in bytes.
+ * @flags: Flags give additionnal information about the allocation:
+ *         OMAP_BO_TILED_x: use the TILER (2D buffers). The TILER container
+ *              unit can be 8, 16 or 32 bits. Cache is always disabled for
+ *              tiled buffers.
+ *         OMAP_BO_SCANOUT: Scannout buffer, consummable by the DSS
+ *         OMAP_BO_CACHED: Buffer CPU caching mode: cached
+ *         OMAP_BO_WC: Buffer CPU caching mode: write-combined
+ *         OMAP_BO_UNCACHED: Buffer CPU caching mode: uncached
+ *         OMAP_BO_MEM_CONTIG: The driver will use dma_alloc to get the memory.
+ *              This can be used to avoid DMM if the userspace knows it needs
+ *              more than 128M of memory at the same time.
+ *         OMAP_BO_MEM_DMM: The driver will use DMM to get the memory. There's
+ *              not much use for this flag at the moment, as on platforms with
+ *              DMM it is used by default, but it's here for completeness.
+ *         OMAP_BO_MEM_PIN: The driver will pin the memory at alloc time, and
+ *              keep it pinned. This can be used to 1) get an error at alloc
+ *              time if DMM space is full, and 2) get rid of the constant
+ *              pin/unpin operations which may have some effect on performance.
+ *
+ * Return: The GEM buffer or NULL if the allocation failed
+ */
 struct drm_gem_object *omap_gem_new(struct drm_device *dev,
 		union omap_gem_size gsize, u32 flags)
 {
@@ -1193,7 +1232,8 @@  struct drm_gem_object *omap_gem_new(struct drm_device *dev,
 		 */
 		flags &= ~(OMAP_BO_CACHED|OMAP_BO_WC|OMAP_BO_UNCACHED);
 		flags |= tiler_get_cpu_cache_flags();
-	} else if ((flags & OMAP_BO_SCANOUT) && !priv->has_dmm) {
+	} else if ((flags & OMAP_BO_MEM_CONTIG) ||
+		((flags & OMAP_BO_SCANOUT) && !priv->has_dmm)) {
 		/*
 		 * If we don't have DMM, we must allocate scanout buffers
 		 * from contiguous DMA memory.
@@ -1253,12 +1293,22 @@  struct drm_gem_object *omap_gem_new(struct drm_device *dev,
 			goto err_release;
 	}
 
+	if (flags & OMAP_BO_MEM_PIN) {
+		ret = omap_gem_pin(obj, NULL);
+		if (ret)
+			goto err_free_dma;
+	}
+
 	mutex_lock(&priv->list_lock);
 	list_add(&omap_obj->mm_list, &priv->obj_list);
 	mutex_unlock(&priv->list_lock);
 
 	return obj;
 
+err_free_dma:
+	if (flags & OMAP_BO_MEM_DMA_API)
+		dma_free_wc(dev->dev, size, omap_obj->vaddr,
+			    omap_obj->dma_addr);
 err_release:
 	drm_gem_object_release(obj);
 err_free:
diff --git a/include/uapi/drm/omap_drm.h b/include/uapi/drm/omap_drm.h
index 5a142fad473c..842d3180a442 100644
--- a/include/uapi/drm/omap_drm.h
+++ b/include/uapi/drm/omap_drm.h
@@ -47,6 +47,15 @@  struct drm_omap_param {
 #define OMAP_BO_UNCACHED	0x00000004
 #define OMAP_BO_CACHE_MASK	0x00000006
 
+/* Force allocation from contiguous DMA memory */
+#define OMAP_BO_MEM_CONTIG	0x00000008
+
+/* Force allocation via DMM */
+#define OMAP_BO_MEM_DMM		0x00000010
+
+/* Pin the buffer when allocating and keep pinned */
+#define OMAP_BO_MEM_PIN		0x00000020
+
 /* Use TILER for the buffer. The TILER container unit can be 8, 16 or 32 bits. */
 #define OMAP_BO_TILED_8		0x00000100
 #define OMAP_BO_TILED_16	0x00000200