[dpdk-dev] [PATCH 4/6] net/ena/base: align IO cq allocation to a 4K

[Michal Krawczyk]

Latest generation HW requires IO completion queue descriptors to be
aligned to a 4K in order to achieve the best performance.

Because of that, the new allocation macros were added, which allows
driver to allocate the memory with specified alignemnt.

The previous allocation macros are now wrappers around the macros
doing the alignment, with the alignment value equal to cacheline size.

Fixes: b68309be44c0 ("net/ena/base: update communication layer for the ENAv2")
Cc: stable at dpdk.org

Signed-off-by: Ido Segev <idose at amazon.com>
Signed-off-by: Michal Krawczyk <mk at semihalf.com>
Reviewed-by: Igor Chauskin <igorch at amazon.com>
Reviewed-by: Amit Bernstein <amitbern at amazon.com>
---
 drivers/net/ena/base/ena_com.c       | 26 ++++++++++----------
 drivers/net/ena/base/ena_com.h       |  2 ++
 drivers/net/ena/base/ena_plat_dpdk.h | 36 +++++++++++++++++++++-------
 3 files changed, 44 insertions(+), 20 deletions(-)

diff --git a/drivers/net/ena/base/ena_com.c b/drivers/net/ena/base/ena_com.c
index 3686ae05c4..aae68721fb 100644
--- a/drivers/net/ena/base/ena_com.c
+++ b/drivers/net/ena/base/ena_com.c
@@ -413,19 +413,21 @@ static int ena_com_init_io_cq(struct ena_com_dev *ena_dev,
 	size = io_cq->cdesc_entry_size_in_bytes * io_cq->q_depth;
 	io_cq->bus = ena_dev->bus;
 
-	ENA_MEM_ALLOC_COHERENT_NODE(ena_dev->dmadev,
-			size,
-			io_cq->cdesc_addr.virt_addr,
-			io_cq->cdesc_addr.phys_addr,
-			io_cq->cdesc_addr.mem_handle,
-			ctx->numa_node,
-			prev_node);
+	ENA_MEM_ALLOC_COHERENT_NODE_ALIGNED(ena_dev->dmadev,
+					    size,
+					    io_cq->cdesc_addr.virt_addr,
+					    io_cq->cdesc_addr.phys_addr,
+					    io_cq->cdesc_addr.mem_handle,
+					    ctx->numa_node,
+					    prev_node,
+					    ENA_CDESC_RING_SIZE_ALIGNMENT);
 	if (!io_cq->cdesc_addr.virt_addr) {
-		ENA_MEM_ALLOC_COHERENT(ena_dev->dmadev,
-				       size,
-				       io_cq->cdesc_addr.virt_addr,
-				       io_cq->cdesc_addr.phys_addr,
-				       io_cq->cdesc_addr.mem_handle);
+		ENA_MEM_ALLOC_COHERENT_ALIGNED(ena_dev->dmadev,
+					       size,
+					       io_cq->cdesc_addr.virt_addr,
+					       io_cq->cdesc_addr.phys_addr,
+					       io_cq->cdesc_addr.mem_handle,
+					       ENA_CDESC_RING_SIZE_ALIGNMENT);
 	}
 
 	if (!io_cq->cdesc_addr.virt_addr) {
diff --git a/drivers/net/ena/base/ena_com.h b/drivers/net/ena/base/ena_com.h
index 8eacaeab0e..64d8f247cb 100644
--- a/drivers/net/ena/base/ena_com.h
+++ b/drivers/net/ena/base/ena_com.h
@@ -23,6 +23,8 @@
 #define ADMIN_CQ_SIZE(depth)	((depth) * sizeof(struct ena_admin_acq_entry))
 #define ADMIN_AENQ_SIZE(depth)	((depth) * sizeof(struct ena_admin_aenq_entry))
 
+#define ENA_CDESC_RING_SIZE_ALIGNMENT	(1 << 12) /* 4K */
+
 /*****************************************************************************/
 /*****************************************************************************/
 /* ENA adaptive interrupt moderation settings */
diff --git a/drivers/net/ena/base/ena_plat_dpdk.h b/drivers/net/ena/base/ena_plat_dpdk.h
index a6782f3732..48c77f0c19 100644
--- a/drivers/net/ena/base/ena_plat_dpdk.h
+++ b/drivers/net/ena/base/ena_plat_dpdk.h
@@ -172,7 +172,8 @@ do {                                                                   \
  */
 extern rte_atomic32_t ena_alloc_cnt;
 
-#define ENA_MEM_ALLOC_COHERENT(dmadev, size, virt, phys, handle)	\
+#define ENA_MEM_ALLOC_COHERENT_ALIGNED(					\
+	dmadev, size, virt, phys, handle, alignment)			\
 	do {								\
 		const struct rte_memzone *mz = NULL;			\
 		ENA_TOUCH(dmadev); ENA_TOUCH(handle);			\
@@ -181,9 +182,10 @@ extern rte_atomic32_t ena_alloc_cnt;
 			snprintf(z_name, sizeof(z_name),		\
 			 "ena_alloc_%d",				\
 			 rte_atomic32_add_return(&ena_alloc_cnt, 1));	\
-			mz = rte_memzone_reserve(z_name, size,		\
+			mz = rte_memzone_reserve_aligned(z_name, size,	\
 					SOCKET_ID_ANY,			\
-					RTE_MEMZONE_IOVA_CONTIG);	\
+					RTE_MEMZONE_IOVA_CONTIG,	\
+					alignment);			\
 			handle = mz;					\
 		}							\
 		if (mz == NULL) {					\
@@ -195,13 +197,21 @@ extern rte_atomic32_t ena_alloc_cnt;
 			phys = mz->iova;				\
 		}							\
 	} while (0)
+#define ENA_MEM_ALLOC_COHERENT(dmadev, size, virt, phys, handle)	\
+		ENA_MEM_ALLOC_COHERENT_ALIGNED(				\
+			dmadev,						\
+			size,						\
+			virt,						\
+			phys,						\
+			handle,						\
+			RTE_CACHE_LINE_SIZE)
 #define ENA_MEM_FREE_COHERENT(dmadev, size, virt, phys, handle) 	\
 		({ ENA_TOUCH(size); ENA_TOUCH(phys);			\
 		   ENA_TOUCH(dmadev);					\
 		   rte_memzone_free(handle); })
 
-#define ENA_MEM_ALLOC_COHERENT_NODE(					\
-	dmadev, size, virt, phys, mem_handle, node, dev_node)		\
+#define ENA_MEM_ALLOC_COHERENT_NODE_ALIGNED(				\
+	dmadev, size, virt, phys, mem_handle, node, dev_node, alignment) \
 	do {								\
 		const struct rte_memzone *mz = NULL;			\
 		ENA_TOUCH(dmadev); ENA_TOUCH(dev_node);			\
@@ -210,8 +220,8 @@ extern rte_atomic32_t ena_alloc_cnt;
 			snprintf(z_name, sizeof(z_name),		\
 			 "ena_alloc_%d",				\
 			 rte_atomic32_add_return(&ena_alloc_cnt, 1));   \
-			mz = rte_memzone_reserve(z_name, size, node,	\
-				RTE_MEMZONE_IOVA_CONTIG);		\
+			mz = rte_memzone_reserve_aligned(z_name, size, node, \
+				RTE_MEMZONE_IOVA_CONTIG, alignment);	\
 			mem_handle = mz;				\
 		}							\
 		if (mz == NULL) {					\
@@ -223,7 +233,17 @@ extern rte_atomic32_t ena_alloc_cnt;
 			phys = mz->iova;				\
 		}							\
 	} while (0)
-
+#define ENA_MEM_ALLOC_COHERENT_NODE(					\
+	dmadev, size, virt, phys, mem_handle, node, dev_node)		\
+		ENA_MEM_ALLOC_COHERENT_NODE_ALIGNED(			\
+			dmadev,						\
+			size,						\
+			virt,						\
+			phys,						\
+			mem_handle,					\
+			node,						\
+			dev_node,					\
+			RTE_CACHE_LINE_SIZE)
 #define ENA_MEM_ALLOC_NODE(dmadev, size, virt, node, dev_node) \
 	do {								\
 		ENA_TOUCH(dmadev); ENA_TOUCH(dev_node);			\
-- 
2.25.1