[PATCH v2 1/2] tee: amdtee: use page_alloc_exact() for memory allocations

[PATCH v2 1/2] tee: amdtee: use page_alloc_exact() for memory allocations

[Devaraj Rangasamy]

Use page_alloc_exact() to get buffers, instead of
get_free_pages(), so as to avoid wastage of memory.
Currently get_free_pages() is allocating at next order,
while page_alloc_exact() will free the unused pages.

Signed-off-by: Devaraj Rangasamy <Devaraj.Rangasamy@amd.com>
---
 v2:
 * Replaced __get_free_pages() with alloc_pages_exact().

 drivers/tee/amdtee/shm_pool.c | 18 ++++++++----------
 1 file changed, 8 insertions(+), 10 deletions(-)

diff --git a/drivers/tee/amdtee/shm_pool.c b/drivers/tee/amdtee/shm_pool.c
index f0303126f199..156e8a6f631f 100644
--- a/drivers/tee/amdtee/shm_pool.c
+++ b/drivers/tee/amdtee/shm_pool.c
@@ -4,6 +4,7 @@
  */
 
 #include <linux/slab.h>
+#include <linux/mm.h>
 #include <linux/tee_drv.h>
 #include <linux/psp.h>
 #include "amdtee_private.h"
@@ -11,26 +12,23 @@
 static int pool_op_alloc(struct tee_shm_pool *pool, struct tee_shm *shm,
 			 size_t size, size_t align)
 {
-	unsigned int order = get_order(size);
-	unsigned long va;
+	void *va;
 	int rc;
 
-	/*
-	 * Ignore alignment since this is already going to be page aligned
-	 * and there's no need for any larger alignment.
-	 */
-	va = __get_free_pages(GFP_KERNEL | __GFP_ZERO, order);
+	size = PAGE_ALIGN(size);
+
+	va = alloc_pages_exact(size, GFP_KERNEL | __GFP_ZERO);
 	if (!va)
 		return -ENOMEM;
 
 	shm->kaddr = (void *)va;
 	shm->paddr = __psp_pa((void *)va);
-	shm->size = PAGE_SIZE << order;
+	shm->size = size;
 
 	/* Map the allocated memory in to TEE */
 	rc = amdtee_map_shmem(shm);
 	if (rc) {
-		free_pages(va, order);
+		free_pages_exact(va, size);
 		shm->kaddr = NULL;
 		return rc;
 	}
@@ -42,7 +40,7 @@ static void pool_op_free(struct tee_shm_pool *pool, struct tee_shm *shm)
 {
 	/* Unmap the shared memory from TEE */
 	amdtee_unmap_shmem(shm);
-	free_pages((unsigned long)shm->kaddr, get_order(shm->size));
+	free_pages_exact(shm->kaddr, shm->size);
 	shm->kaddr = NULL;
 }
 
-- 
2.25.1

[PATCH v2 2/2] tee: amdtee: add support for CMA buffer allocations

[Devaraj Rangasamy]

amdtee driver shall use CMA region for contiguous
buffer allocation, if CMA is available.

since CMA and DMA contiguous APIs are not exported,
this support is enabled only when amdtee is built
as a builtin driver.

Signed-off-by: Devaraj Rangasamy <Devaraj.Rangasamy@amd.com>
Signed-off-by: SivaSangeetha SK <SivaSangeetha.SK@amd.com>
Reviewed-by: Rijo Thomas <Rijo-john.Thomas@amd.com>
---
v2:
* Switched to generic CMA pool.
* Replaced __get_free_pages() with alloc_pages_exact().
* Compacted patch with simplified changes within single file.

 drivers/tee/amdtee/shm_pool.c | 51 +++++++++++++++++++++++++++++++++--
 1 file changed, 49 insertions(+), 2 deletions(-)

diff --git a/drivers/tee/amdtee/shm_pool.c b/drivers/tee/amdtee/shm_pool.c
index 156e8a6f631f..d504d9749114 100644
--- a/drivers/tee/amdtee/shm_pool.c
+++ b/drivers/tee/amdtee/shm_pool.c
@@ -5,10 +5,50 @@
 
 #include <linux/slab.h>
 #include <linux/mm.h>
+#include <linux/dma-map-ops.h>
 #include <linux/tee_drv.h>
 #include <linux/psp.h>
 #include "amdtee_private.h"
 
+#if IS_BUILTIN(CONFIG_AMDTEE) && IS_ENABLED(CONFIG_DMA_CMA)
+static void *alloc_from_cma(size_t size)
+{
+
+	int nr_pages = size >> PAGE_SHIFT;
+	struct page *page;
+
+	page = dma_alloc_from_contiguous(NULL, nr_pages, 0, false);
+	if (page)
+		return page_to_virt(page);
+
+	return NULL;
+}
+
+static bool free_from_cma(struct tee_shm *shm)
+{
+
+	int nr_pages;
+	struct page *page;
+
+	if (!dev_get_cma_area(NULL))
+		return false;
+
+	nr_pages = shm->size >> PAGE_SHIFT;
+	page = virt_to_page(shm->kaddr);
+	return dma_release_from_contiguous(NULL, page, nr_pages);
+}
+#else
+static void *alloc_from_cma(size_t size)
+{
+	return NULL;
+}
+
+static bool free_from_cma(struct tee_shm *shm)
+{
+	return false;
+}
+#endif
+
 static int pool_op_alloc(struct tee_shm_pool *pool, struct tee_shm *shm,
 			 size_t size, size_t align)
 {
@@ -17,7 +57,11 @@ static int pool_op_alloc(struct tee_shm_pool *pool, struct tee_shm *shm,
 
 	size = PAGE_ALIGN(size);
 
-	va = alloc_pages_exact(size, GFP_KERNEL | __GFP_ZERO);
+	va = alloc_from_cma(size);
+
+	if (!va)
+		va = alloc_pages_exact(size, GFP_KERNEL | __GFP_ZERO);
+
 	if (!va)
 		return -ENOMEM;
 
@@ -40,7 +84,10 @@ static void pool_op_free(struct tee_shm_pool *pool, struct tee_shm *shm)
 {
 	/* Unmap the shared memory from TEE */
 	amdtee_unmap_shmem(shm);
-	free_pages_exact(shm->kaddr, shm->size);
+
+	if (!free_from_cma(shm))
+		free_pages_exact(shm->kaddr, shm->size);
+
 	shm->kaddr = NULL;
 }
 
-- 
2.25.1

Re: [PATCH v2 1/2] tee: amdtee: use page_alloc_exact() for memory allocations

[Rijo Thomas]

On 8/30/2023 1:49 AM, Devaraj Rangasamy wrote:
> Use page_alloc_exact() to get buffers, instead of
> get_free_pages(), so as to avoid wastage of memory.
> Currently get_free_pages() is allocating at next order,
> while page_alloc_exact() will free the unused pages.
> 
> Signed-off-by: Devaraj Rangasamy <Devaraj.Rangasamy@amd.com>
> ---
>  v2:
>  * Replaced __get_free_pages() with alloc_pages_exact().
> 

Reviewed-by: Rijo Thomas <Rijo-john.Thomas@amd.com>

Thanks,
Rijo

>  drivers/tee/amdtee/shm_pool.c | 18 ++++++++----------
>  1 file changed, 8 insertions(+), 10 deletions(-)
> 
> diff --git a/drivers/tee/amdtee/shm_pool.c b/drivers/tee/amdtee/shm_pool.c
> index f0303126f199..156e8a6f631f 100644
> --- a/drivers/tee/amdtee/shm_pool.c
> +++ b/drivers/tee/amdtee/shm_pool.c
> @@ -4,6 +4,7 @@
>   */
>  
>  #include <linux/slab.h>
> +#include <linux/mm.h>
>  #include <linux/tee_drv.h>
>  #include <linux/psp.h>
>  #include "amdtee_private.h"
> @@ -11,26 +12,23 @@
>  static int pool_op_alloc(struct tee_shm_pool *pool, struct tee_shm *shm,
>  			 size_t size, size_t align)
>  {
> -	unsigned int order = get_order(size);
> -	unsigned long va;
> +	void *va;
>  	int rc;
>  
> -	/*
> -	 * Ignore alignment since this is already going to be page aligned
> -	 * and there's no need for any larger alignment.
> -	 */
> -	va = __get_free_pages(GFP_KERNEL | __GFP_ZERO, order);
> +	size = PAGE_ALIGN(size);
> +
> +	va = alloc_pages_exact(size, GFP_KERNEL | __GFP_ZERO);
>  	if (!va)
>  		return -ENOMEM;
>  
>  	shm->kaddr = (void *)va;
>  	shm->paddr = __psp_pa((void *)va);
> -	shm->size = PAGE_SIZE << order;
> +	shm->size = size;
>  
>  	/* Map the allocated memory in to TEE */
>  	rc = amdtee_map_shmem(shm);
>  	if (rc) {
> -		free_pages(va, order);
> +		free_pages_exact(va, size);
>  		shm->kaddr = NULL;
>  		return rc;
>  	}
> @@ -42,7 +40,7 @@ static void pool_op_free(struct tee_shm_pool *pool, struct tee_shm *shm)
>  {
>  	/* Unmap the shared memory from TEE */
>  	amdtee_unmap_shmem(shm);
> -	free_pages((unsigned long)shm->kaddr, get_order(shm->size));
> +	free_pages_exact(shm->kaddr, shm->size);
>  	shm->kaddr = NULL;
>  }
>  

Re: [PATCH v2 1/2] tee: amdtee: use page_alloc_exact() for memory allocations

[Jeff Johnson]

On 8/29/2023 1:19 PM, Devaraj Rangasamy wrote:> Use page_alloc_exact() 
to get buffers, instead of
> get_free_pages(), so as to avoid wastage of memory.
> Currently get_free_pages() is allocating at next order,
> while page_alloc_exact() will free the unused pages.

s/page_alloc_exact()/alloc_pages_exact()/ everywhere including subject 
to match the actual code change?

Re: [PATCH v2 1/2] tee: amdtee: use page_alloc_exact() for memory allocations

[Devaraj Rangasamy]

On 10/5/2023 12:39 AM, Jeff Johnson wrote:
> On 8/29/2023 1:19 PM, Devaraj Rangasamy wrote:> Use page_alloc_exact() 
> to get buffers, instead of
>> get_free_pages(), so as to avoid wastage of memory.
>> Currently get_free_pages() is allocating at next order,
>> while page_alloc_exact() will free the unused pages.
>
> s/page_alloc_exact()/alloc_pages_exact()/ everywhere including subject 
> to match the actual code change?

Ack. Thanks for the catching this.

While I could send an updated patch taking care of this, we are working 
on a rearchitected patch, due to a new requirement.

Kindly abandon this patch series, and we will be posting new support 
soon to the community.