[PATCH v5 net-nex 0/5] xdp: introduce bulking for page_pool tx return path

[PATCH v5 net-nex 0/5] xdp: introduce bulking for page_pool tx return path

[Lorenzo Bianconi]

XDP bulk APIs introduce a defer/flush mechanism to return
pages belonging to the same xdp_mem_allocator object
(identified via the mem.id field) in bulk to optimize
I-cache and D-cache since xdp_return_frame is usually run
inside the driver NAPI tx completion loop.
Convert mvneta, mvpp2 and mlx5 drivers to xdp_return_frame_bulk APIs.

Changes since v4:
- fix comments
- introduce xdp_frame_bulk_init utility routine
- compiler annotations for I-cache code layout
- move rcu_read_lock outside fast-path
- mlx5 xdp bulking code optimization

Changes since v3:
- align DEV_MAP_BULK_SIZE to XDP_BULK_QUEUE_SIZE
- refactor page_pool_put_page_bulk to avoid code duplication

Changes since v2:
- move mvneta changes in a dedicated patch

Changes since v1:
- improve comments
- rework xdp_return_frame_bulk routine logic
- move count and xa fields at the beginning of xdp_frame_bulk struct
- invert logic in page_pool_put_page_bulk for loop

Lorenzo Bianconi (5):
  net: xdp: introduce bulking for xdp tx return path
  net: page_pool: add bulk support for ptr_ring
  net: mvneta: add xdp tx return bulking support
  net: mvpp2: add xdp tx return bulking support
  net: mlx5: add xdp tx return bulking support

 drivers/net/ethernet/marvell/mvneta.c         | 10 ++-
 .../net/ethernet/marvell/mvpp2/mvpp2_main.c   | 10 ++-
 .../net/ethernet/mellanox/mlx5/core/en/xdp.c  | 22 ++++--
 include/net/page_pool.h                       | 26 +++++++
 include/net/xdp.h                             | 17 ++++-
 net/core/page_pool.c                          | 69 ++++++++++++++++---
 net/core/xdp.c                                | 54 +++++++++++++++
 7 files changed, 191 insertions(+), 17 deletions(-)

-- 
2.26.2

[PATCH v5 net-nex 1/5] net: xdp: introduce bulking for xdp tx return path

[Lorenzo Bianconi]

XDP bulk APIs introduce a defer/flush mechanism to return
pages belonging to the same xdp_mem_allocator object
(identified via the mem.id field) in bulk to optimize
I-cache and D-cache since xdp_return_frame is usually run
inside the driver NAPI tx completion loop.
The bulk queue size is set to 16 to be aligned to how
XDP_REDIRECT bulking works. The bulk is flushed when
it is full or when mem.id changes.
xdp_frame_bulk is usually stored/allocated on the function
call-stack to avoid locking penalties.
Current implementation considers only page_pool memory model.

Suggested-by: Jesper Dangaard Brouer <brouer@redhat.com>
Co-developed-by: Jesper Dangaard Brouer <brouer@redhat.com>
Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
Signed-off-by: Lorenzo Bianconi <lorenzo@kernel.org>
---
 include/net/xdp.h | 17 +++++++++++++-
 net/core/xdp.c    | 59 +++++++++++++++++++++++++++++++++++++++++++++++
 2 files changed, 75 insertions(+), 1 deletion(-)

diff --git a/include/net/xdp.h b/include/net/xdp.h
index 3814fb631d52..7d48b2ae217a 100644
--- a/include/net/xdp.h
+++ b/include/net/xdp.h
@@ -104,6 +104,18 @@ struct xdp_frame {
 	struct net_device *dev_rx; /* used by cpumap */
 };
 
+#define XDP_BULK_QUEUE_SIZE	16
+struct xdp_frame_bulk {
+	int count;
+	void *xa;
+	void *q[XDP_BULK_QUEUE_SIZE];
+};
+
+static __always_inline void xdp_frame_bulk_init(struct xdp_frame_bulk *bq)
+{
+	/* bq->count will be zero'ed when bq->xa gets updated */
+	bq->xa = NULL;
+}
 
 static inline struct skb_shared_info *
 xdp_get_shared_info_from_frame(struct xdp_frame *frame)
@@ -194,6 +206,9 @@ struct xdp_frame *xdp_convert_buff_to_frame(struct xdp_buff *xdp)
 void xdp_return_frame(struct xdp_frame *xdpf);
 void xdp_return_frame_rx_napi(struct xdp_frame *xdpf);
 void xdp_return_buff(struct xdp_buff *xdp);
+void xdp_flush_frame_bulk(struct xdp_frame_bulk *bq);
+void xdp_return_frame_bulk(struct xdp_frame *xdpf,
+			   struct xdp_frame_bulk *bq);
 
 /* When sending xdp_frame into the network stack, then there is no
  * return point callback, which is needed to release e.g. DMA-mapping
@@ -245,6 +260,6 @@ bool xdp_attachment_flags_ok(struct xdp_attachment_info *info,
 void xdp_attachment_setup(struct xdp_attachment_info *info,
 			  struct netdev_bpf *bpf);
 
-#define DEV_MAP_BULK_SIZE 16
+#define DEV_MAP_BULK_SIZE XDP_BULK_QUEUE_SIZE
 
 #endif /* __LINUX_NET_XDP_H__ */
diff --git a/net/core/xdp.c b/net/core/xdp.c
index 48aba933a5a8..bbaee7fdd44f 100644
--- a/net/core/xdp.c
+++ b/net/core/xdp.c
@@ -380,6 +380,65 @@ void xdp_return_frame_rx_napi(struct xdp_frame *xdpf)
 }
 EXPORT_SYMBOL_GPL(xdp_return_frame_rx_napi);
 
+/* XDP bulk APIs introduce a defer/flush mechanism to return
+ * pages belonging to the same xdp_mem_allocator object
+ * (identified via the mem.id field) in bulk to optimize
+ * I-cache and D-cache.
+ * The bulk queue size is set to 16 to be aligned to how
+ * XDP_REDIRECT bulking works. The bulk is flushed when
+ * it is full or when mem.id changes.
+ * xdp_frame_bulk is usually stored/allocated on the function
+ * call-stack to avoid locking penalties.
+ */
+void xdp_flush_frame_bulk(struct xdp_frame_bulk *bq)
+{
+	struct xdp_mem_allocator *xa = bq->xa;
+	int i;
+
+	if (unlikely(!xa))
+		return;
+
+	for (i = 0; i < bq->count; i++) {
+		struct page *page = virt_to_head_page(bq->q[i]);
+
+		page_pool_put_full_page(xa->page_pool, page, false);
+	}
+	/* bq->xa is not cleared to save lookup, if mem.id same in next bulk */
+	bq->count = 0;
+}
+EXPORT_SYMBOL_GPL(xdp_flush_frame_bulk);
+
+/* Must be called with rcu_read_lock held */
+void xdp_return_frame_bulk(struct xdp_frame *xdpf,
+			   struct xdp_frame_bulk *bq)
+{
+	struct xdp_mem_info *mem = &xdpf->mem;
+	struct xdp_mem_allocator *xa;
+
+	if (mem->type != MEM_TYPE_PAGE_POOL) {
+		__xdp_return(xdpf->data, &xdpf->mem, false);
+		return;
+	}
+
+	xa = bq->xa;
+	if (unlikely(!xa)) {
+		xa = rhashtable_lookup(mem_id_ht, &mem->id, mem_id_rht_params);
+		bq->count = 0;
+		bq->xa = xa;
+	}
+
+	if (bq->count == XDP_BULK_QUEUE_SIZE)
+		xdp_flush_frame_bulk(bq);
+
+	if (unlikely(mem->id != xa->mem.id)) {
+		xdp_flush_frame_bulk(bq);
+		bq->xa = rhashtable_lookup(mem_id_ht, &mem->id, mem_id_rht_params);
+	}
+
+	bq->q[bq->count++] = xdpf->data;
+}
+EXPORT_SYMBOL_GPL(xdp_return_frame_bulk);
+
 void xdp_return_buff(struct xdp_buff *xdp)
 {
 	__xdp_return(xdp->data, &xdp->rxq->mem, true);
-- 
2.26.2

[PATCH v5 net-nex 2/5] net: page_pool: add bulk support for ptr_ring

[Lorenzo Bianconi]

Introduce the capability to batch page_pool ptr_ring refill since it is
usually run inside the driver NAPI tx completion loop.

Suggested-by: Jesper Dangaard Brouer <brouer@redhat.com>
Co-developed-by: Jesper Dangaard Brouer <brouer@redhat.com>
Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
Signed-off-by: Lorenzo Bianconi <lorenzo@kernel.org>
---
 include/net/page_pool.h | 26 ++++++++++++++++
 net/core/page_pool.c    | 69 +++++++++++++++++++++++++++++++++++------
 net/core/xdp.c          |  9 ++----
 3 files changed, 87 insertions(+), 17 deletions(-)

diff --git a/include/net/page_pool.h b/include/net/page_pool.h
index 81d7773f96cd..b5b195305346 100644
--- a/include/net/page_pool.h
+++ b/include/net/page_pool.h
@@ -152,6 +152,8 @@ struct page_pool *page_pool_create(const struct page_pool_params *params);
 void page_pool_destroy(struct page_pool *pool);
 void page_pool_use_xdp_mem(struct page_pool *pool, void (*disconnect)(void *));
 void page_pool_release_page(struct page_pool *pool, struct page *page);
+void page_pool_put_page_bulk(struct page_pool *pool, void **data,
+			     int count);
 #else
 static inline void page_pool_destroy(struct page_pool *pool)
 {
@@ -165,6 +167,11 @@ static inline void page_pool_release_page(struct page_pool *pool,
 					  struct page *page)
 {
 }
+
+static inline void page_pool_put_page_bulk(struct page_pool *pool, void **data,
+					   int count)
+{
+}
 #endif
 
 void page_pool_put_page(struct page_pool *pool, struct page *page,
@@ -215,4 +222,23 @@ static inline void page_pool_nid_changed(struct page_pool *pool, int new_nid)
 	if (unlikely(pool->p.nid != new_nid))
 		page_pool_update_nid(pool, new_nid);
 }
+
+static inline void page_pool_ring_lock(struct page_pool *pool)
+	__acquires(&pool->ring.producer_lock)
+{
+	if (in_serving_softirq())
+		spin_lock(&pool->ring.producer_lock);
+	else
+		spin_lock_bh(&pool->ring.producer_lock);
+}
+
+static inline void page_pool_ring_unlock(struct page_pool *pool)
+	__releases(&pool->ring.producer_lock)
+{
+	if (in_serving_softirq())
+		spin_unlock(&pool->ring.producer_lock);
+	else
+		spin_unlock_bh(&pool->ring.producer_lock);
+}
+
 #endif /* _NET_PAGE_POOL_H */
diff --git a/net/core/page_pool.c b/net/core/page_pool.c
index ef98372facf6..a3e6051ac978 100644
--- a/net/core/page_pool.c
+++ b/net/core/page_pool.c
@@ -11,6 +11,8 @@
 #include <linux/device.h>
 
 #include <net/page_pool.h>
+#include <net/xdp.h>
+
 #include <linux/dma-direction.h>
 #include <linux/dma-mapping.h>
 #include <linux/page-flags.h>
@@ -362,8 +364,9 @@ static bool pool_page_reusable(struct page_pool *pool, struct page *page)
  * If the page refcnt != 1, then the page will be returned to memory
  * subsystem.
  */
-void page_pool_put_page(struct page_pool *pool, struct page *page,
-			unsigned int dma_sync_size, bool allow_direct)
+static __always_inline struct page *
+__page_pool_put_page(struct page_pool *pool, struct page *page,
+		     unsigned int dma_sync_size, bool allow_direct)
 {
 	/* This allocator is optimized for the XDP mode that uses
 	 * one-frame-per-page, but have fallbacks that act like the
@@ -379,15 +382,12 @@ void page_pool_put_page(struct page_pool *pool, struct page *page,
 			page_pool_dma_sync_for_device(pool, page,
 						      dma_sync_size);
 
-		if (allow_direct && in_serving_softirq())
-			if (page_pool_recycle_in_cache(page, pool))
-				return;
+		if (allow_direct && in_serving_softirq() &&
+		    page_pool_recycle_in_cache(page, pool))
+			return NULL;
 
-		if (!page_pool_recycle_in_ring(pool, page)) {
-			/* Cache full, fallback to free pages */
-			page_pool_return_page(pool, page);
-		}
-		return;
+		/* Page found as candidate for recycling */
+		return page;
 	}
 	/* Fallback/non-XDP mode: API user have elevated refcnt.
 	 *
@@ -405,9 +405,58 @@ void page_pool_put_page(struct page_pool *pool, struct page *page,
 	/* Do not replace this with page_pool_return_page() */
 	page_pool_release_page(pool, page);
 	put_page(page);
+
+	return NULL;
+}
+
+void page_pool_put_page(struct page_pool *pool, struct page *page,
+			unsigned int dma_sync_size, bool allow_direct)
+{
+	page = __page_pool_put_page(pool, page, dma_sync_size, allow_direct);
+	if (page && !page_pool_recycle_in_ring(pool, page)) {
+		/* Cache full, fallback to free pages */
+		page_pool_return_page(pool, page);
+	}
 }
 EXPORT_SYMBOL(page_pool_put_page);
 
+/* Caller must not use data area after call, as this function overwrites it */
+void page_pool_put_page_bulk(struct page_pool *pool, void **data,
+			     int count)
+{
+	int i, bulk_len = 0, pa_len = 0;
+
+	for (i = 0; i < count; i++) {
+		struct page *page = virt_to_head_page(data[i]);
+
+		page = __page_pool_put_page(pool, page, -1, false);
+		/* Approved for bulk recycling in ptr_ring cache */
+		if (page)
+			data[bulk_len++] = page;
+	}
+
+	if (unlikely(!bulk_len))
+		return;
+
+	/* Bulk producer into ptr_ring page_pool cache */
+	page_pool_ring_lock(pool);
+	for (i = 0; i < bulk_len; i++) {
+		if (__ptr_ring_produce(&pool->ring, data[i]))
+			data[pa_len++] = data[i];
+	}
+	page_pool_ring_unlock(pool);
+
+	if (likely(!pa_len))
+		return;
+
+	/* ptr_ring cache full, free pages outside producer lock since
+	 * put_page() with refcnt == 1 can be an expensive operation
+	 */
+	for (i = 0; i < pa_len; i++)
+		page_pool_return_page(pool, data[i]);
+}
+EXPORT_SYMBOL(page_pool_put_page_bulk);
+
 static void page_pool_empty_ring(struct page_pool *pool)
 {
 	struct page *page;
diff --git a/net/core/xdp.c b/net/core/xdp.c
index bbaee7fdd44f..3d330ebda893 100644
--- a/net/core/xdp.c
+++ b/net/core/xdp.c
@@ -393,16 +393,11 @@ EXPORT_SYMBOL_GPL(xdp_return_frame_rx_napi);
 void xdp_flush_frame_bulk(struct xdp_frame_bulk *bq)
 {
 	struct xdp_mem_allocator *xa = bq->xa;
-	int i;
 
-	if (unlikely(!xa))
+	if (unlikely(!xa || !bq->count))
 		return;
 
-	for (i = 0; i < bq->count; i++) {
-		struct page *page = virt_to_head_page(bq->q[i]);
-
-		page_pool_put_full_page(xa->page_pool, page, false);
-	}
+	page_pool_put_page_bulk(xa->page_pool, bq->q, bq->count);
 	/* bq->xa is not cleared to save lookup, if mem.id same in next bulk */
 	bq->count = 0;
 }
-- 
2.26.2

[PATCH v5 net-nex 3/5] net: mvneta: add xdp tx return bulking support

[Lorenzo Bianconi]

Convert mvneta driver to xdp_return_frame_bulk APIs.

XDP_REDIRECT (upstream codepath): 275Kpps
XDP_REDIRECT (upstream codepath + bulking APIs): 284Kpps

Co-developed-by: Jesper Dangaard Brouer <brouer@redhat.com>
Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
Signed-off-by: Lorenzo Bianconi <lorenzo@kernel.org>
---
 drivers/net/ethernet/marvell/mvneta.c | 10 +++++++++-
 1 file changed, 9 insertions(+), 1 deletion(-)

diff --git a/drivers/net/ethernet/marvell/mvneta.c b/drivers/net/ethernet/marvell/mvneta.c
index 54b0bf574c05..183530ed4d1d 100644
--- a/drivers/net/ethernet/marvell/mvneta.c
+++ b/drivers/net/ethernet/marvell/mvneta.c
@@ -1834,8 +1834,13 @@ static void mvneta_txq_bufs_free(struct mvneta_port *pp,
 				 struct netdev_queue *nq, bool napi)
 {
 	unsigned int bytes_compl = 0, pkts_compl = 0;
+	struct xdp_frame_bulk bq;
 	int i;
 
+	xdp_frame_bulk_init(&bq);
+
+	rcu_read_lock(); /* need for xdp_return_frame_bulk */
+
 	for (i = 0; i < num; i++) {
 		struct mvneta_tx_buf *buf = &txq->buf[txq->txq_get_index];
 		struct mvneta_tx_desc *tx_desc = txq->descs +
@@ -1857,9 +1862,12 @@ static void mvneta_txq_bufs_free(struct mvneta_port *pp,
 			if (napi && buf->type == MVNETA_TYPE_XDP_TX)
 				xdp_return_frame_rx_napi(buf->xdpf);
 			else
-				xdp_return_frame(buf->xdpf);
+				xdp_return_frame_bulk(buf->xdpf, &bq);
 		}
 	}
+	xdp_flush_frame_bulk(&bq);
+
+	rcu_read_unlock();
 
 	netdev_tx_completed_queue(nq, pkts_compl, bytes_compl);
 }
-- 
2.26.2

[PATCH v5 net-nex 4/5] net: mvpp2: add xdp tx return bulking support

[Lorenzo Bianconi]

Convert mvpp2 driver to xdp_return_frame_bulk APIs.

XDP_REDIRECT (upstream codepath): 1.79Mpps
XDP_REDIRECT (upstream codepath + bulking APIs): 1.93Mpps

Tested-by: Matteo Croce <mcroce@microsoft.com>
Co-developed-by: Jesper Dangaard Brouer <brouer@redhat.com>
Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
Signed-off-by: Lorenzo Bianconi <lorenzo@kernel.org>
---
 drivers/net/ethernet/marvell/mvpp2/mvpp2_main.c | 10 +++++++++-
 1 file changed, 9 insertions(+), 1 deletion(-)

diff --git a/drivers/net/ethernet/marvell/mvpp2/mvpp2_main.c b/drivers/net/ethernet/marvell/mvpp2/mvpp2_main.c
index f6616c8933ca..3069e192d773 100644
--- a/drivers/net/ethernet/marvell/mvpp2/mvpp2_main.c
+++ b/drivers/net/ethernet/marvell/mvpp2/mvpp2_main.c
@@ -2440,8 +2440,13 @@ static void mvpp2_txq_bufs_free(struct mvpp2_port *port,
 				struct mvpp2_tx_queue *txq,
 				struct mvpp2_txq_pcpu *txq_pcpu, int num)
 {
+	struct xdp_frame_bulk bq;
 	int i;
 
+	xdp_frame_bulk_init(&bq);
+
+	rcu_read_lock(); /* need for xdp_return_frame_bulk */
+
 	for (i = 0; i < num; i++) {
 		struct mvpp2_txq_pcpu_buf *tx_buf =
 			txq_pcpu->buffs + txq_pcpu->txq_get_index;
@@ -2454,10 +2459,13 @@ static void mvpp2_txq_bufs_free(struct mvpp2_port *port,
 			dev_kfree_skb_any(tx_buf->skb);
 		else if (tx_buf->type == MVPP2_TYPE_XDP_TX ||
 			 tx_buf->type == MVPP2_TYPE_XDP_NDO)
-			xdp_return_frame(tx_buf->xdpf);
+			xdp_return_frame_bulk(tx_buf->xdpf, &bq);
 
 		mvpp2_txq_inc_get(txq_pcpu);
 	}
+	xdp_flush_frame_bulk(&bq);
+
+	rcu_read_unlock();
 }
 
 static inline struct mvpp2_rx_queue *mvpp2_get_rx_queue(struct mvpp2_port *port,
-- 
2.26.2

[PATCH v5 net-nex 5/5] net: mlx5: add xdp tx return bulking support

[Lorenzo Bianconi]

Convert mlx5 driver to xdp_return_frame_bulk APIs.

XDP_REDIRECT (upstream codepath): 8.9Mpps
XDP_REDIRECT (upstream codepath + bulking APIs): 10.2Mpps

Co-developed-by: Jesper Dangaard Brouer <brouer@redhat.com>
Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
Signed-off-by: Lorenzo Bianconi <lorenzo@kernel.org>
---
 .../net/ethernet/mellanox/mlx5/core/en/xdp.c  | 22 +++++++++++++++----
 1 file changed, 18 insertions(+), 4 deletions(-)

diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en/xdp.c b/drivers/net/ethernet/mellanox/mlx5/core/en/xdp.c
index ae90d533a350..2e3e78b0f333 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/en/xdp.c
+++ b/drivers/net/ethernet/mellanox/mlx5/core/en/xdp.c
@@ -366,7 +366,8 @@ mlx5e_xmit_xdp_frame(struct mlx5e_xdpsq *sq, struct mlx5e_xmit_data *xdptxd,
 static void mlx5e_free_xdpsq_desc(struct mlx5e_xdpsq *sq,
 				  struct mlx5e_xdp_wqe_info *wi,
 				  u32 *xsk_frames,
-				  bool recycle)
+				  bool recycle,
+				  struct xdp_frame_bulk *bq)
 {
 	struct mlx5e_xdp_info_fifo *xdpi_fifo = &sq->db.xdpi_fifo;
 	u16 i;
@@ -379,7 +380,7 @@ static void mlx5e_free_xdpsq_desc(struct mlx5e_xdpsq *sq,
 			/* XDP_TX from the XSK RQ and XDP_REDIRECT */
 			dma_unmap_single(sq->pdev, xdpi.frame.dma_addr,
 					 xdpi.frame.xdpf->len, DMA_TO_DEVICE);
-			xdp_return_frame(xdpi.frame.xdpf);
+			xdp_return_frame_bulk(xdpi.frame.xdpf, bq);
 			break;
 		case MLX5E_XDP_XMIT_MODE_PAGE:
 			/* XDP_TX from the regular RQ */
@@ -397,12 +398,15 @@ static void mlx5e_free_xdpsq_desc(struct mlx5e_xdpsq *sq,
 
 bool mlx5e_poll_xdpsq_cq(struct mlx5e_cq *cq)
 {
+	struct xdp_frame_bulk bq;
 	struct mlx5e_xdpsq *sq;
 	struct mlx5_cqe64 *cqe;
 	u32 xsk_frames = 0;
 	u16 sqcc;
 	int i;
 
+	xdp_frame_bulk_init(&bq);
+
 	sq = container_of(cq, struct mlx5e_xdpsq, cq);
 
 	if (unlikely(!test_bit(MLX5E_SQ_STATE_ENABLED, &sq->state)))
@@ -434,7 +438,7 @@ bool mlx5e_poll_xdpsq_cq(struct mlx5e_cq *cq)
 
 			sqcc += wi->num_wqebbs;
 
-			mlx5e_free_xdpsq_desc(sq, wi, &xsk_frames, true);
+			mlx5e_free_xdpsq_desc(sq, wi, &xsk_frames, true, &bq);
 		} while (!last_wqe);
 
 		if (unlikely(get_cqe_opcode(cqe) != MLX5_CQE_REQ)) {
@@ -447,6 +451,8 @@ bool mlx5e_poll_xdpsq_cq(struct mlx5e_cq *cq)
 		}
 	} while ((++i < MLX5E_TX_CQ_POLL_BUDGET) && (cqe = mlx5_cqwq_get_cqe(&cq->wq)));
 
+	xdp_flush_frame_bulk(&bq);
+
 	if (xsk_frames)
 		xsk_tx_completed(sq->xsk_pool, xsk_frames);
 
@@ -463,8 +469,13 @@ bool mlx5e_poll_xdpsq_cq(struct mlx5e_cq *cq)
 
 void mlx5e_free_xdpsq_descs(struct mlx5e_xdpsq *sq)
 {
+	struct xdp_frame_bulk bq;
 	u32 xsk_frames = 0;
 
+	xdp_frame_bulk_init(&bq);
+
+	rcu_read_lock(); /* need for xdp_return_frame_bulk */
+
 	while (sq->cc != sq->pc) {
 		struct mlx5e_xdp_wqe_info *wi;
 		u16 ci;
@@ -474,9 +485,12 @@ void mlx5e_free_xdpsq_descs(struct mlx5e_xdpsq *sq)
 
 		sq->cc += wi->num_wqebbs;
 
-		mlx5e_free_xdpsq_desc(sq, wi, &xsk_frames, false);
+		mlx5e_free_xdpsq_desc(sq, wi, &xsk_frames, false, &bq);
 	}
 
+	xdp_flush_frame_bulk(&bq);
+	rcu_read_unlock();
+
 	if (xsk_frames)
 		xsk_tx_completed(sq->xsk_pool, xsk_frames);
 }
-- 
2.26.2

Re: [PATCH v5 net-nex 0/5] xdp: introduce bulking for page_pool tx return path

[Jesper Dangaard Brouer]

On Tue, 10 Nov 2020 16:37:55 +0100
Lorenzo Bianconi <lorenzo@kernel.org> wrote:

> XDP bulk APIs introduce a defer/flush mechanism to return
> pages belonging to the same xdp_mem_allocator object
> (identified via the mem.id field) in bulk to optimize
> I-cache and D-cache since xdp_return_frame is usually run
> inside the driver NAPI tx completion loop.
> Convert mvneta, mvpp2 and mlx5 drivers to xdp_return_frame_bulk APIs.

Series

Acked-by: Jesper Dangaard Brouer <brouer@redhat.com>

> Changes since v4:
> - fix comments
> - introduce xdp_frame_bulk_init utility routine
> - compiler annotations for I-cache code layout
> - move rcu_read_lock outside fast-path
> - mlx5 xdp bulking code optimization

I've done a lot of these changes, and benchmarked them on mlx5, details in[1].

[1] https://github.com/xdp-project/xdp-project/blob/master/areas/mem/xdp_bulk_return01.org

> Changes since v3:
> - align DEV_MAP_BULK_SIZE to XDP_BULK_QUEUE_SIZE
> - refactor page_pool_put_page_bulk to avoid code duplication
> 
> Changes since v2:
> - move mvneta changes in a dedicated patch
> 
> Changes since v1:
> - improve comments
> - rework xdp_return_frame_bulk routine logic
> - move count and xa fields at the beginning of xdp_frame_bulk struct
> - invert logic in page_pool_put_page_bulk for loop
> 
> Lorenzo Bianconi (5):
>   net: xdp: introduce bulking for xdp tx return path
>   net: page_pool: add bulk support for ptr_ring
>   net: mvneta: add xdp tx return bulking support
>   net: mvpp2: add xdp tx return bulking support
>   net: mlx5: add xdp tx return bulking support
> 
>  drivers/net/ethernet/marvell/mvneta.c         | 10 ++-
>  .../net/ethernet/marvell/mvpp2/mvpp2_main.c   | 10 ++-
>  .../net/ethernet/mellanox/mlx5/core/en/xdp.c  | 22 ++++--
>  include/net/page_pool.h                       | 26 +++++++
>  include/net/xdp.h                             | 17 ++++-
>  net/core/page_pool.c                          | 69 ++++++++++++++++---
>  net/core/xdp.c                                | 54 +++++++++++++++
>  7 files changed, 191 insertions(+), 17 deletions(-)
> 



-- 
Best regards,
  Jesper Dangaard Brouer
  MSc.CS, Principal Kernel Engineer at Red Hat
  LinkedIn: http://www.linkedin.com/in/brouer

Re: [PATCH v5 net-nex 0/5] xdp: introduce bulking for page_pool tx return path

[Lorenzo Bianconi]

[-- Attachment #1: Type: text/plain, Size: 2490 bytes --]

> On Tue, 10 Nov 2020 16:37:55 +0100
> Lorenzo Bianconi <lorenzo@kernel.org> wrote:
> 
> > XDP bulk APIs introduce a defer/flush mechanism to return
> > pages belonging to the same xdp_mem_allocator object
> > (identified via the mem.id field) in bulk to optimize
> > I-cache and D-cache since xdp_return_frame is usually run
> > inside the driver NAPI tx completion loop.
> > Convert mvneta, mvpp2 and mlx5 drivers to xdp_return_frame_bulk APIs.
> 
> Series
> 
> Acked-by: Jesper Dangaard Brouer <brouer@redhat.com>
> 
> > Changes since v4:
> > - fix comments
> > - introduce xdp_frame_bulk_init utility routine
> > - compiler annotations for I-cache code layout
> > - move rcu_read_lock outside fast-path
> > - mlx5 xdp bulking code optimization
> 
> I've done a lot of these changes, and benchmarked them on mlx5, details in[1].
> 
> [1] https://github.com/xdp-project/xdp-project/blob/master/areas/mem/xdp_bulk_return01.org

ops sorry, I forgot to add it to the cover-letter.

Regards,
Lorenzo

> 
> > Changes since v3:
> > - align DEV_MAP_BULK_SIZE to XDP_BULK_QUEUE_SIZE
> > - refactor page_pool_put_page_bulk to avoid code duplication
> > 
> > Changes since v2:
> > - move mvneta changes in a dedicated patch
> > 
> > Changes since v1:
> > - improve comments
> > - rework xdp_return_frame_bulk routine logic
> > - move count and xa fields at the beginning of xdp_frame_bulk struct
> > - invert logic in page_pool_put_page_bulk for loop
> > 
> > Lorenzo Bianconi (5):
> >   net: xdp: introduce bulking for xdp tx return path
> >   net: page_pool: add bulk support for ptr_ring
> >   net: mvneta: add xdp tx return bulking support
> >   net: mvpp2: add xdp tx return bulking support
> >   net: mlx5: add xdp tx return bulking support
> > 
> >  drivers/net/ethernet/marvell/mvneta.c         | 10 ++-
> >  .../net/ethernet/marvell/mvpp2/mvpp2_main.c   | 10 ++-
> >  .../net/ethernet/mellanox/mlx5/core/en/xdp.c  | 22 ++++--
> >  include/net/page_pool.h                       | 26 +++++++
> >  include/net/xdp.h                             | 17 ++++-
> >  net/core/page_pool.c                          | 69 ++++++++++++++++---
> >  net/core/xdp.c                                | 54 +++++++++++++++
> >  7 files changed, 191 insertions(+), 17 deletions(-)
> > 
> 
> 
> 
> -- 
> Best regards,
>   Jesper Dangaard Brouer
>   MSc.CS, Principal Kernel Engineer at Red Hat
>   LinkedIn: http://www.linkedin.com/in/brouer
> 

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 228 bytes --]

Re: [PATCH v5 net-nex 5/5] net: mlx5: add xdp tx return bulking support

[Jesper Dangaard Brouer]

On Tue, 10 Nov 2020 16:38:00 +0100
Lorenzo Bianconi <lorenzo@kernel.org> wrote:

> Convert mlx5 driver to xdp_return_frame_bulk APIs.
> 
> XDP_REDIRECT (upstream codepath): 8.9Mpps
> XDP_REDIRECT (upstream codepath + bulking APIs): 10.2Mpps
> 
> Co-developed-by: Jesper Dangaard Brouer <brouer@redhat.com>
> Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
> Signed-off-by: Lorenzo Bianconi <lorenzo@kernel.org>

I did most of my testing on this driver:

Tested-by: Jesper Dangaard Brouer <brouer@redhat.com>

-- 
Best regards,
  Jesper Dangaard Brouer
  MSc.CS, Principal Kernel Engineer at Red Hat
  LinkedIn: http://www.linkedin.com/in/brouer

RE: [PATCH v5 net-nex 2/5] net: page_pool: add bulk support for ptr_ring

[John Fastabend]

Lorenzo Bianconi wrote:
> Introduce the capability to batch page_pool ptr_ring refill since it is
> usually run inside the driver NAPI tx completion loop.
> 
> Suggested-by: Jesper Dangaard Brouer <brouer@redhat.com>
> Co-developed-by: Jesper Dangaard Brouer <brouer@redhat.com>
> Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
> Signed-off-by: Lorenzo Bianconi <lorenzo@kernel.org>
> ---
>  include/net/page_pool.h | 26 ++++++++++++++++
>  net/core/page_pool.c    | 69 +++++++++++++++++++++++++++++++++++------
>  net/core/xdp.c          |  9 ++----
>  3 files changed, 87 insertions(+), 17 deletions(-)

[...]

> +/* Caller must not use data area after call, as this function overwrites it */
> +void page_pool_put_page_bulk(struct page_pool *pool, void **data,
> +			     int count)
> +{
> +	int i, bulk_len = 0, pa_len = 0;
> +
> +	for (i = 0; i < count; i++) {
> +		struct page *page = virt_to_head_page(data[i]);
> +
> +		page = __page_pool_put_page(pool, page, -1, false);
> +		/* Approved for bulk recycling in ptr_ring cache */
> +		if (page)
> +			data[bulk_len++] = page;
> +	}
> +
> +	if (unlikely(!bulk_len))
> +		return;
> +
> +	/* Bulk producer into ptr_ring page_pool cache */
> +	page_pool_ring_lock(pool);
> +	for (i = 0; i < bulk_len; i++) {
> +		if (__ptr_ring_produce(&pool->ring, data[i]))
> +			data[pa_len++] = data[i];

How about bailing out on the first error? bulk_len should be less than
16 right, so should we really keep retying hoping ring->size changes?

> +	}
> +	page_pool_ring_unlock(pool);
> +
> +	if (likely(!pa_len))
> +		return;
> +
> +	/* ptr_ring cache full, free pages outside producer lock since
> +	 * put_page() with refcnt == 1 can be an expensive operation
> +	 */
> +	for (i = 0; i < pa_len; i++)
> +		page_pool_return_page(pool, data[i]);
> +}
> +EXPORT_SYMBOL(page_pool_put_page_bulk);
> +

Otherwise LGTM.

Re: [PATCH v5 net-nex 2/5] net: page_pool: add bulk support for ptr_ring

[Lorenzo Bianconi]

[-- Attachment #1: Type: text/plain, Size: 2407 bytes --]

> Lorenzo Bianconi wrote:
> > Introduce the capability to batch page_pool ptr_ring refill since it is
> > usually run inside the driver NAPI tx completion loop.
> > 
> > Suggested-by: Jesper Dangaard Brouer <brouer@redhat.com>
> > Co-developed-by: Jesper Dangaard Brouer <brouer@redhat.com>
> > Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
> > Signed-off-by: Lorenzo Bianconi <lorenzo@kernel.org>
> > ---
> >  include/net/page_pool.h | 26 ++++++++++++++++
> >  net/core/page_pool.c    | 69 +++++++++++++++++++++++++++++++++++------
> >  net/core/xdp.c          |  9 ++----
> >  3 files changed, 87 insertions(+), 17 deletions(-)
> 
> [...]
> 
> > +/* Caller must not use data area after call, as this function overwrites it */
> > +void page_pool_put_page_bulk(struct page_pool *pool, void **data,
> > +			     int count)
> > +{
> > +	int i, bulk_len = 0, pa_len = 0;
> > +
> > +	for (i = 0; i < count; i++) {
> > +		struct page *page = virt_to_head_page(data[i]);
> > +
> > +		page = __page_pool_put_page(pool, page, -1, false);
> > +		/* Approved for bulk recycling in ptr_ring cache */
> > +		if (page)
> > +			data[bulk_len++] = page;
> > +	}
> > +
> > +	if (unlikely(!bulk_len))
> > +		return;
> > +
> > +	/* Bulk producer into ptr_ring page_pool cache */
> > +	page_pool_ring_lock(pool);
> > +	for (i = 0; i < bulk_len; i++) {
> > +		if (__ptr_ring_produce(&pool->ring, data[i]))
> > +			data[pa_len++] = data[i];
> 
> How about bailing out on the first error? bulk_len should be less than
> 16 right, so should we really keep retying hoping ring->size changes?

do you mean doing something like:

	page_pool_ring_lock(pool);
	if (__ptr_ring_full(&pool->ring)) {
		pa_len = bulk_len;
		page_pool_ring_unlock(pool);
		goto out;
	}
	...
out:
	for (i = 0; i < pa_len; i++) {
		...
	}

I do not know if it is better or not since the consumer can run in parallel.
@Jesper/Ilias: any idea?

Regards,
Lorenzo

> 
> > +	}
> > +	page_pool_ring_unlock(pool);
> > +
> > +	if (likely(!pa_len))
> > +		return;
> > +
> > +	/* ptr_ring cache full, free pages outside producer lock since
> > +	 * put_page() with refcnt == 1 can be an expensive operation
> > +	 */
> > +	for (i = 0; i < pa_len; i++)
> > +		page_pool_return_page(pool, data[i]);
> > +}
> > +EXPORT_SYMBOL(page_pool_put_page_bulk);
> > +
> 
> Otherwise LGTM.

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 228 bytes --]

Re: [PATCH v5 net-nex 2/5] net: page_pool: add bulk support for ptr_ring

[Jesper Dangaard Brouer]

On Wed, 11 Nov 2020 11:43:31 +0100
Lorenzo Bianconi <lorenzo@kernel.org> wrote:

> > Lorenzo Bianconi wrote:  
> > > Introduce the capability to batch page_pool ptr_ring refill since it is
> > > usually run inside the driver NAPI tx completion loop.
> > > 
> > > Suggested-by: Jesper Dangaard Brouer <brouer@redhat.com>
> > > Co-developed-by: Jesper Dangaard Brouer <brouer@redhat.com>
> > > Signed-off-by: Jesper Dangaard Brouer <brouer@redhat.com>
> > > Signed-off-by: Lorenzo Bianconi <lorenzo@kernel.org>
> > > ---
> > >  include/net/page_pool.h | 26 ++++++++++++++++
> > >  net/core/page_pool.c    | 69 +++++++++++++++++++++++++++++++++++------
> > >  net/core/xdp.c          |  9 ++----
> > >  3 files changed, 87 insertions(+), 17 deletions(-)  
> > 
> > [...]
> >   
> > > +/* Caller must not use data area after call, as this function overwrites it */
> > > +void page_pool_put_page_bulk(struct page_pool *pool, void **data,
> > > +			     int count)
> > > +{
> > > +	int i, bulk_len = 0, pa_len = 0;
> > > +
> > > +	for (i = 0; i < count; i++) {
> > > +		struct page *page = virt_to_head_page(data[i]);
> > > +
> > > +		page = __page_pool_put_page(pool, page, -1, false);
> > > +		/* Approved for bulk recycling in ptr_ring cache */
> > > +		if (page)
> > > +			data[bulk_len++] = page;
> > > +	}
> > > +
> > > +	if (unlikely(!bulk_len))
> > > +		return;
> > > +
> > > +	/* Bulk producer into ptr_ring page_pool cache */
> > > +	page_pool_ring_lock(pool);
> > > +	for (i = 0; i < bulk_len; i++) {
> > > +		if (__ptr_ring_produce(&pool->ring, data[i]))
> > > +			data[pa_len++] = data[i];  
> > 
> > How about bailing out on the first error? bulk_len should be less than
> > 16 right, so should we really keep retying hoping ring->size changes?  
> 
> do you mean doing something like:
> 
> 	page_pool_ring_lock(pool);
> 	if (__ptr_ring_full(&pool->ring)) {
> 		pa_len = bulk_len;
> 		page_pool_ring_unlock(pool);
> 		goto out;
> 	}
> 	...
> out:
> 	for (i = 0; i < pa_len; i++) {
> 		...
> 	}

I think this is the change John is looking for:

diff --git a/net/core/page_pool.c b/net/core/page_pool.c
index a06606f07df0..3093fe4e1cd7 100644
--- a/net/core/page_pool.c
+++ b/net/core/page_pool.c
@@ -424,7 +424,7 @@ EXPORT_SYMBOL(page_pool_put_page);
 void page_pool_put_page_bulk(struct page_pool *pool, void **data,
                             int count)
 {
-       int i, bulk_len = 0, pa_len = 0;
+       int i, bulk_len = 0;
        bool order0 = (pool->p.order == 0);
 
        for (i = 0; i < count; i++) {
@@ -448,17 +448,18 @@ void page_pool_put_page_bulk(struct page_pool *pool, void **data,
        page_pool_ring_lock(pool);
        for (i = 0; i < bulk_len; i++) {
                if (__ptr_ring_produce(&pool->ring, data[i]))
-                       data[pa_len++] = data[i];
+                       break; /* ring_full */
        }
        page_pool_ring_unlock(pool);
 
-       if (likely(!pa_len))
+       /* Hopefully all pages was return into ptr_ring */
+       if (likely(i == bulk_len))
                return;
 
-       /* ptr_ring cache full, free pages outside producer lock since
-        * put_page() with refcnt == 1 can be an expensive operation
+       /* ptr_ring cache full, free remaining pages outside producer lock
+        * since put_page() with refcnt == 1 can be an expensive operation
         */
-       for (i = 0; i < pa_len; i++)
+       for (; i < bulk_len; i++)
                page_pool_return_page(pool, data[i]);
 }
 EXPORT_SYMBOL(page_pool_put_page_bulk);


> I do not know if it is better or not since the consumer can run in
> parallel. @Jesper/Ilias: any idea?

Currently it is not very likely that the consumer runs in parallel, but
is it possible. (As you experienced on your testlab with mlx5, the
DMA-TX completion did run on another CPU, but I asked you to
reconfigure this to have it run on same CPU, as it is suboptimal).
When we (finally) support this memory type for SKBs it will be more
normal to happen.

But, John is right, for ptr_ring we should exit as soon the first
"produce" fails.  This is because I know how ptr_ring works internally.
The "consumer" will free slots in chunks of 16 slots, so it is not very
likely that a slot opens up.

> >   
> > > +	}
> > > +	page_pool_ring_unlock(pool);
> > > +
> > > +	if (likely(!pa_len))
> > > +		return;
> > > +
> > > +	/* ptr_ring cache full, free pages outside producer lock since
> > > +	 * put_page() with refcnt == 1 can be an expensive operation
> > > +	 */
> > > +	for (i = 0; i < pa_len; i++)
> > > +		page_pool_return_page(pool, data[i]);
> > > +}
> > > +EXPORT_SYMBOL(page_pool_put_page_bulk);
> > > +  
> > 
> > Otherwise LGTM.  



-- 
Best regards,
  Jesper Dangaard Brouer
  MSc.CS, Principal Kernel Engineer at Red Hat
  LinkedIn: http://www.linkedin.com/in/brouer