[PATCH] net/mlx5: allocate dma pool structs on the pool's NUMA node

[PATCH] net/mlx5: allocate dma pool structs on the pool's NUMA node

[Seongjun Hong]

The mlx5 DMA pool structs (mlx5_dma_pool, mlx5_dma_pool_page and its
block bitmap, and mlx5_frag_buf_node_pools) are allocated without a
node hint, so they land on the node of whichever CPU happens to create
or fill the pool, whiile the DMA pages they describe are allocated on
the pool's node.

The pool, page and bitmap are dereferenced on every block allocation
and free. Allocate them on the pool's NUMA node as well, so that all
of a pool's state lives on one node.

Signed-off-by: Seongjun Hong <hsj0512@snu.ac.kr>
---
 drivers/net/ethernet/mellanox/mlx5/core/alloc.c | 11 ++++++-----
 1 file changed, 6 insertions(+), 5 deletions(-)

diff --git a/drivers/net/ethernet/mellanox/mlx5/core/alloc.c b/drivers/net/ethernet/mellanox/mlx5/core/alloc.c
index a92cf545bdaf..dcd281c4691d 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/alloc.c
+++ b/drivers/net/ethernet/mellanox/mlx5/core/alloc.c
@@ -110,7 +110,7 @@ static struct mlx5_dma_pool *mlx5_dma_pool_create(struct mlx5_core_dev *dev,
 {
 	struct mlx5_dma_pool *pool;
 
-	pool = kzalloc_obj(*pool);
+	pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, node);
 	if (!pool)
 		return NULL;
 
@@ -127,19 +127,20 @@ mlx5_dma_pool_page_alloc(struct mlx5_dma_pool *pool)
 {
 	int blocks_per_page = BIT(PAGE_SHIFT - pool->block_shift);
 	struct mlx5_dma_pool_page *page;
+	int node = pool->node;
 
-	page = kzalloc_obj(*page);
+	page = kzalloc_node(sizeof(*page), GFP_KERNEL, node);
 	if (!page)
 		goto err_out;
 
 	page->pool = pool;
-	page->bitmap = bitmap_zalloc(blocks_per_page, GFP_KERNEL);
+	page->bitmap = bitmap_zalloc_node(blocks_per_page, GFP_KERNEL, node);
 	if (!page->bitmap)
 		goto err_free_page;
 
 	bitmap_fill(page->bitmap, blocks_per_page);
 	page->buf = mlx5_dma_zalloc_coherent_node(pool->dev, PAGE_SIZE,
-						  &page->dma, pool->node);
+						  &page->dma, node);
 	if (!page->buf)
 		goto err_free_bitmap;
 
@@ -278,7 +279,7 @@ mlx5_frag_buf_node_pools_create(struct mlx5_core_dev *dev, int node)
 {
 	struct mlx5_frag_buf_node_pools *node_pools;
 
-	node_pools = kzalloc_obj(*node_pools);
+	node_pools = kzalloc_node(sizeof(*node_pools), GFP_KERNEL, node);
 	if (!node_pools)
 		return NULL;
 

---
base-commit: 8b4e7209c842d8cb9516f1f5ef0a88aa2d8831a6
change-id: 20261005-net-mlx5-numa-allocate-pool-page-2ee20a04147e

Best regards,
-- 
Seongjun Hong

Re: [PATCH] net/mlx5: allocate dma pool structs on the pool's NUMA node

[Tariq Toukan]

On 05/10/2026 16:46, Seongjun Hong wrote:
> The mlx5 DMA pool structs (mlx5_dma_pool, mlx5_dma_pool_page and its
> block bitmap, and mlx5_frag_buf_node_pools) are allocated without a
> node hint, so they land on the node of whichever CPU happens to create
> or fill the pool, whiile 

nit: typo

> the DMA pages they describe are allocated on
> the pool's node.
> 
> The pool, page and bitmap are dereferenced on every block allocation
> and free. Allocate them on the pool's NUMA node as well, so that all
> of a pool's state lives on one node.
> 

Thanks for the patch.

That is true, but these structs are only used from the control path.
That is once per WQ/CQ/EQ or doorbell record at open/close, not per packet.

I doubt we can measure any benefit, so it's just a matter of consistency 
with the other fields access in data-path.

The API kzalloc_node falls back if node is offline, so it's safe.

No strong motivation, but we can accept it.
Please fix the typo and emphasize that these are control-path objects.

> Signed-off-by: Seongjun Hong <hsj0512@snu.ac.kr>
> ---
>   drivers/net/ethernet/mellanox/mlx5/core/alloc.c | 11 ++++++-----
>   1 file changed, 6 insertions(+), 5 deletions(-)
> 
> diff --git a/drivers/net/ethernet/mellanox/mlx5/core/alloc.c b/drivers/net/ethernet/mellanox/mlx5/core/alloc.c
> index a92cf545bdaf..dcd281c4691d 100644
> --- a/drivers/net/ethernet/mellanox/mlx5/core/alloc.c
> +++ b/drivers/net/ethernet/mellanox/mlx5/core/alloc.c
> @@ -110,7 +110,7 @@ static struct mlx5_dma_pool *mlx5_dma_pool_create(struct mlx5_core_dev *dev,
>   {
>   	struct mlx5_dma_pool *pool;
>   
> -	pool = kzalloc_obj(*pool);
> +	pool = kzalloc_node(sizeof(*pool), GFP_KERNEL, node);
>   	if (!pool)
>   		return NULL;
>   
> @@ -127,19 +127,20 @@ mlx5_dma_pool_page_alloc(struct mlx5_dma_pool *pool)
>   {
>   	int blocks_per_page = BIT(PAGE_SHIFT - pool->block_shift);
>   	struct mlx5_dma_pool_page *page;
> +	int node = pool->node;
>   
> -	page = kzalloc_obj(*page);
> +	page = kzalloc_node(sizeof(*page), GFP_KERNEL, node);
>   	if (!page)
>   		goto err_out;
>   
>   	page->pool = pool;
> -	page->bitmap = bitmap_zalloc(blocks_per_page, GFP_KERNEL);
> +	page->bitmap = bitmap_zalloc_node(blocks_per_page, GFP_KERNEL, node);
>   	if (!page->bitmap)
>   		goto err_free_page;
>   
>   	bitmap_fill(page->bitmap, blocks_per_page);
>   	page->buf = mlx5_dma_zalloc_coherent_node(pool->dev, PAGE_SIZE,
> -						  &page->dma, pool->node);
> +						  &page->dma, node);
>   	if (!page->buf)
>   		goto err_free_bitmap;
>   
> @@ -278,7 +279,7 @@ mlx5_frag_buf_node_pools_create(struct mlx5_core_dev *dev, int node)
>   {
>   	struct mlx5_frag_buf_node_pools *node_pools;
>   
> -	node_pools = kzalloc_obj(*node_pools);
> +	node_pools = kzalloc_node(sizeof(*node_pools), GFP_KERNEL, node);
>   	if (!node_pools)
>   		return NULL;
>   
> 
> ---
> base-commit: 8b4e7209c842d8cb9516f1f5ef0a88aa2d8831a6
> change-id: 20261005-net-mlx5-numa-allocate-pool-page-2ee20a04147e
> 
> Best regards,

Re: [PATCH] net/mlx5: allocate dma pool structs on the pool's NUMA node

[Seongjun Hong]

On Tue, Oct 06, 2026 at 10:10:06AM +0300, Tariq Toukan wrote:
> 
> 
> On 05/10/2026 16:46, Seongjun Hong wrote:
> > The mlx5 DMA pool structs (mlx5_dma_pool, mlx5_dma_pool_page and its
> > block bitmap, and mlx5_frag_buf_node_pools) are allocated without a
> > node hint, so they land on the node of whichever CPU happens to create
> > or fill the pool, whiile
> 
> nit: typo
> 
> > the DMA pages they describe are allocated on
> > the pool's node.
> > 
> > The pool, page and bitmap are dereferenced on every block allocation
> > and free. Allocate them on the pool's NUMA node as well, so that all
> > of a pool's state lives on one node.
> > 
> 
> Thanks for the patch.
> 
> That is true, but these structs are only used from the control path.
> That is once per WQ/CQ/EQ or doorbell record at open/close, not per packet.
> 
> I doubt we can measure any benefit, so it's just a matter of consistency
> with the other fields access in data-path.
> 
> The API kzalloc_node falls back if node is offline, so it's safe.
> 
> No strong motivation, but we can accept it.
> Please fix the typo and emphasize that these are control-path objects.
> 

Will do, thanks! :)

Best regards,
-- 
Seongjun Hong