[PATCH 1/2] drm/etnaviv: Preallocate STLB according to CPU PAGE_SIZE

[PATCH 1/2] drm/etnaviv: Preallocate STLB according to CPU PAGE_SIZE

[Sui Jingfeng]

The dma_direct_alloc() allocate one page at minmium, which size is the CPU
PAGE_SIZE. while the etnaviv_iommuv2_ensure_stlb() only ask for 4KiB. The
rest memory space that beyond 4KiB gets wasted on bigger page size systems.
For example, on 16KiB CPU page size systems, we will waste the rest 12KiB.
On 64KiB CPU page size systems, we will waste the rest 60KiB.

Since addresses within one page are always contiguous, the rest memory can
be used to store adjacent slave TLB entries. Then, when the neighbourhoods
TLB is being hit on the next time, we don't have to ask another one page
from the system. Saving both memorys and times overhead because of that.

Signed-off-by: Sui Jingfeng <sui.jingfeng@linux.dev>
---
 drivers/gpu/drm/etnaviv/etnaviv_iommu_v2.c | 64 +++++++++++++++++++---
 1 file changed, 56 insertions(+), 8 deletions(-)

diff --git a/drivers/gpu/drm/etnaviv/etnaviv_iommu_v2.c b/drivers/gpu/drm/etnaviv/etnaviv_iommu_v2.c
index d664ae29ae20..fa6eed1ae1be 100644
--- a/drivers/gpu/drm/etnaviv/etnaviv_iommu_v2.c
+++ b/drivers/gpu/drm/etnaviv/etnaviv_iommu_v2.c
@@ -44,19 +44,66 @@ to_v2_context(struct etnaviv_iommu_context *context)
 	return container_of(context, struct etnaviv_iommuv2_context, base);
 }
 
+static int etnaviv_iommuv2_stlb_free(struct etnaviv_iommuv2_context *context)
+{
+	struct device *dev = context->base.global->dev;
+	unsigned int i;
+
+	for (i = 0; i < MMUv2_MAX_STLB_ENTRIES; ++i) {
+		u32 *vaddr = context->stlb_cpu[i];
+
+		if (!vaddr)
+			continue;
+
+		context->stlb_cpu[i] = NULL;
+
+		if (i % (PAGE_SIZE / SZ_4K))
+			continue;
+
+		dma_free_wc(dev, PAGE_SIZE, vaddr, context->stlb_dma[i]);
+	}
+
+	return 0;
+}
+
+static int
+etnaviv_iommuv2_ensure_stlb_new(struct etnaviv_iommuv2_context *context,
+				unsigned int stlb)
+{
+	struct device *dev = context->base.global->dev;
+	void *vaddr;
+	dma_addr_t daddr;
+	unsigned int i;
+
+	if (context->stlb_cpu[stlb])
+		return 0;
+
+	vaddr = dma_alloc_wc(dev, PAGE_SIZE, &daddr, GFP_KERNEL);
+	if (!vaddr)
+		return -ENOMEM;
+
+	memset32(vaddr, MMUv2_PTE_EXCEPTION, PAGE_SIZE / sizeof(u32));
+
+	stlb &= ~(PAGE_SIZE / SZ_4K - 1);
+
+	for (i = 0; i < PAGE_SIZE / SZ_4K; ++i) {
+		context->stlb_cpu[stlb + i] = vaddr;
+		context->stlb_dma[stlb + i] = daddr;
+		context->mtlb_cpu[stlb + i] = daddr | MMUv2_PTE_PRESENT;
+		vaddr += SZ_4K;
+		daddr += SZ_4K;
+	}
+
+	return 0;
+}
+
 static void etnaviv_iommuv2_free(struct etnaviv_iommu_context *context)
 {
 	struct etnaviv_iommuv2_context *v2_context = to_v2_context(context);
-	int i;
 
 	drm_mm_takedown(&context->mm);
 
-	for (i = 0; i < MMUv2_MAX_STLB_ENTRIES; i++) {
-		if (v2_context->stlb_cpu[i])
-			dma_free_wc(context->global->dev, SZ_4K,
-				    v2_context->stlb_cpu[i],
-				    v2_context->stlb_dma[i]);
-	}
+	etnaviv_iommuv2_stlb_free(v2_context);
 
 	dma_free_wc(context->global->dev, SZ_4K, v2_context->mtlb_cpu,
 		    v2_context->mtlb_dma);
@@ -65,6 +112,7 @@ static void etnaviv_iommuv2_free(struct etnaviv_iommu_context *context)
 
 	vfree(v2_context);
 }
+
 static int
 etnaviv_iommuv2_ensure_stlb(struct etnaviv_iommuv2_context *v2_context,
 			    int stlb)
@@ -109,7 +157,7 @@ static int etnaviv_iommuv2_map(struct etnaviv_iommu_context *context,
 	mtlb_entry = (iova & MMUv2_MTLB_MASK) >> MMUv2_MTLB_SHIFT;
 	stlb_entry = (iova & MMUv2_STLB_MASK) >> MMUv2_STLB_SHIFT;
 
-	ret = etnaviv_iommuv2_ensure_stlb(v2_context, mtlb_entry);
+	ret = etnaviv_iommuv2_ensure_stlb_new(v2_context, mtlb_entry);
 	if (ret)
 		return ret;
 
-- 
2.34.1

[PATCH 2/2] drm/etnaviv: Remove the etnaviv_iommuv2_ensure_stlb()

[Sui Jingfeng]

The etnaviv_iommuv2_ensure_stlb() function have been replaced by the
etnaviv_iommuv2_ensure_stlb_new(), remove it.

Signed-off-by: Sui Jingfeng <sui.jingfeng@linux.dev>
---
 drivers/gpu/drm/etnaviv/etnaviv_iommu_v2.c | 24 ----------------------
 1 file changed, 24 deletions(-)

diff --git a/drivers/gpu/drm/etnaviv/etnaviv_iommu_v2.c b/drivers/gpu/drm/etnaviv/etnaviv_iommu_v2.c
index fa6eed1ae1be..65b91eabb278 100644
--- a/drivers/gpu/drm/etnaviv/etnaviv_iommu_v2.c
+++ b/drivers/gpu/drm/etnaviv/etnaviv_iommu_v2.c
@@ -113,30 +113,6 @@ static void etnaviv_iommuv2_free(struct etnaviv_iommu_context *context)
 	vfree(v2_context);
 }
 
-static int
-etnaviv_iommuv2_ensure_stlb(struct etnaviv_iommuv2_context *v2_context,
-			    int stlb)
-{
-	if (v2_context->stlb_cpu[stlb])
-		return 0;
-
-	v2_context->stlb_cpu[stlb] =
-			dma_alloc_wc(v2_context->base.global->dev, SZ_4K,
-				     &v2_context->stlb_dma[stlb],
-				     GFP_KERNEL);
-
-	if (!v2_context->stlb_cpu[stlb])
-		return -ENOMEM;
-
-	memset32(v2_context->stlb_cpu[stlb], MMUv2_PTE_EXCEPTION,
-		 SZ_4K / sizeof(u32));
-
-	v2_context->mtlb_cpu[stlb] =
-			v2_context->stlb_dma[stlb] | MMUv2_PTE_PRESENT;
-
-	return 0;
-}
-
 static int etnaviv_iommuv2_map(struct etnaviv_iommu_context *context,
 			       unsigned long iova, phys_addr_t paddr,
 			       size_t size, int prot)
-- 
2.34.1

Re: [PATCH 1/2] drm/etnaviv: Preallocate STLB according to CPU PAGE_SIZE

[Lucas Stach]

Am Freitag, dem 08.11.2024 um 22:36 +0800 schrieb Sui Jingfeng:
> The dma_direct_alloc() allocate one page at minmium, which size is the CPU
> PAGE_SIZE. while the etnaviv_iommuv2_ensure_stlb() only ask for 4KiB. The
> rest memory space that beyond 4KiB gets wasted on bigger page size systems.
> For example, on 16KiB CPU page size systems, we will waste the rest 12KiB.
> On 64KiB CPU page size systems, we will waste the rest 60KiB.
> 
> Since addresses within one page are always contiguous, the rest memory can
> be used to store adjacent slave TLB entries. Then, when the neighbourhoods
> TLB is being hit on the next time, we don't have to ask another one page
> from the system. Saving both memorys and times overhead because of that.
> 
While this isn't adding a lot of code to etnaviv, I wonder if this
couldn't be handled by using a dma_pool for the pagetable allocations.

Regards,
Lucas

> Signed-off-by: Sui Jingfeng <sui.jingfeng@linux.dev>
> ---
>  drivers/gpu/drm/etnaviv/etnaviv_iommu_v2.c | 64 +++++++++++++++++++---
>  1 file changed, 56 insertions(+), 8 deletions(-)
> 
> diff --git a/drivers/gpu/drm/etnaviv/etnaviv_iommu_v2.c b/drivers/gpu/drm/etnaviv/etnaviv_iommu_v2.c
> index d664ae29ae20..fa6eed1ae1be 100644
> --- a/drivers/gpu/drm/etnaviv/etnaviv_iommu_v2.c
> +++ b/drivers/gpu/drm/etnaviv/etnaviv_iommu_v2.c
> @@ -44,19 +44,66 @@ to_v2_context(struct etnaviv_iommu_context *context)
>  	return container_of(context, struct etnaviv_iommuv2_context, base);
>  }
>  
> +static int etnaviv_iommuv2_stlb_free(struct etnaviv_iommuv2_context *context)
> +{
> +	struct device *dev = context->base.global->dev;
> +	unsigned int i;
> +
> +	for (i = 0; i < MMUv2_MAX_STLB_ENTRIES; ++i) {
> +		u32 *vaddr = context->stlb_cpu[i];
> +
> +		if (!vaddr)
> +			continue;
> +
> +		context->stlb_cpu[i] = NULL;
> +
> +		if (i % (PAGE_SIZE / SZ_4K))
> +			continue;
> +
> +		dma_free_wc(dev, PAGE_SIZE, vaddr, context->stlb_dma[i]);
> +	}
> +
> +	return 0;
> +}
> +
> +static int
> +etnaviv_iommuv2_ensure_stlb_new(struct etnaviv_iommuv2_context *context,
> +				unsigned int stlb)
> +{
> +	struct device *dev = context->base.global->dev;
> +	void *vaddr;
> +	dma_addr_t daddr;
> +	unsigned int i;
> +
> +	if (context->stlb_cpu[stlb])
> +		return 0;
> +
> +	vaddr = dma_alloc_wc(dev, PAGE_SIZE, &daddr, GFP_KERNEL);
> +	if (!vaddr)
> +		return -ENOMEM;
> +
> +	memset32(vaddr, MMUv2_PTE_EXCEPTION, PAGE_SIZE / sizeof(u32));
> +
> +	stlb &= ~(PAGE_SIZE / SZ_4K - 1);
> +
> +	for (i = 0; i < PAGE_SIZE / SZ_4K; ++i) {
> +		context->stlb_cpu[stlb + i] = vaddr;
> +		context->stlb_dma[stlb + i] = daddr;
> +		context->mtlb_cpu[stlb + i] = daddr | MMUv2_PTE_PRESENT;
> +		vaddr += SZ_4K;
> +		daddr += SZ_4K;
> +	}
> +
> +	return 0;
> +}
> +
>  static void etnaviv_iommuv2_free(struct etnaviv_iommu_context *context)
>  {
>  	struct etnaviv_iommuv2_context *v2_context = to_v2_context(context);
> -	int i;
>  
>  	drm_mm_takedown(&context->mm);
>  
> -	for (i = 0; i < MMUv2_MAX_STLB_ENTRIES; i++) {
> -		if (v2_context->stlb_cpu[i])
> -			dma_free_wc(context->global->dev, SZ_4K,
> -				    v2_context->stlb_cpu[i],
> -				    v2_context->stlb_dma[i]);
> -	}
> +	etnaviv_iommuv2_stlb_free(v2_context);
>  
>  	dma_free_wc(context->global->dev, SZ_4K, v2_context->mtlb_cpu,
>  		    v2_context->mtlb_dma);
> @@ -65,6 +112,7 @@ static void etnaviv_iommuv2_free(struct etnaviv_iommu_context *context)
>  
>  	vfree(v2_context);
>  }
> +
>  static int
>  etnaviv_iommuv2_ensure_stlb(struct etnaviv_iommuv2_context *v2_context,
>  			    int stlb)
> @@ -109,7 +157,7 @@ static int etnaviv_iommuv2_map(struct etnaviv_iommu_context *context,
>  	mtlb_entry = (iova & MMUv2_MTLB_MASK) >> MMUv2_MTLB_SHIFT;
>  	stlb_entry = (iova & MMUv2_STLB_MASK) >> MMUv2_STLB_SHIFT;
>  
> -	ret = etnaviv_iommuv2_ensure_stlb(v2_context, mtlb_entry);
> +	ret = etnaviv_iommuv2_ensure_stlb_new(v2_context, mtlb_entry);
>  	if (ret)
>  		return ret;
>