[PATCH v2 net-next 0/5] dpaa2-eth: add support for xdp bulk enqueue

[PATCH v2 net-next 0/5] dpaa2-eth: add support for xdp bulk enqueue

[Ioana Ciornei]

The first patch moves the DEV_MAP_BULK_SIZE macro into the xdp.h header
file so that drivers can take advantage of it and use it.

The following 3 patches are there to setup the scene for using the bulk
enqueue feature.  First of all, the prototype of the enqueue function is
changed so that it returns the number of enqueued frames. Second, the
bulk enqueue interface is used but without any functional changes, still
one frame at a time is enqueued.  Third, the .ndo_xdp_xmit callback is
split into two stages, create all FDs for the xdp_frames received and
then enqueue them.

The last patch of the series builds on top of the others and instead of
issuing an enqueue operation for each FD it issues a bulk enqueue call
for as many frames as possible. This is repeated until all frames are
enqueued or the maximum number of retries is hit. We do not use the
XDP_XMIT_FLUSH flag since the architecture is not capable to store all
frames dequeued in a NAPI cycle, instead we send out right away all
frames received in a .ndo_xdp_xmit call.


Changes in v2:
 - statically allocate an array of dpaa2_fd by frame queue
 - use the DEV_MAP_BULK_SIZE as the maximum number of xdp_frames
   received in .ndo_xdp_xmit()

Ioana Ciornei (5):
  xdp: export the DEV_MAP_BULK_SIZE macro
  dpaa2-eth: return num_enqueued frames from enqueue callback
  dpaa2-eth: use the bulk ring mode enqueue interface
  dpaa2-eth: split the .ndo_xdp_xmit callback into two stages
  dpaa2-eth: use bulk enqueue in .ndo_xdp_xmit

 .../net/ethernet/freescale/dpaa2/dpaa2-eth.c  | 136 ++++++++++--------
 .../net/ethernet/freescale/dpaa2/dpaa2-eth.h  |   8 +-
 include/net/xdp.h                             |   2 +
 kernel/bpf/devmap.c                           |   1 -
 4 files changed, 87 insertions(+), 60 deletions(-)

-- 
2.17.1

[PATCH v2 net-next 1/5] xdp: export the DEV_MAP_BULK_SIZE macro

[Ioana Ciornei]

Export the DEV_MAP_BULK_SIZE macro to the header file so that drivers
can directly use it as the maximum number of xdp_frames received in the
.ndo_xdp_xmit() callback.

Signed-off-by: Ioana Ciornei <ioana.ciornei@nxp.com>
---
Changes in v2:
 - patch added

 include/net/xdp.h   | 2 ++
 kernel/bpf/devmap.c | 1 -
 2 files changed, 2 insertions(+), 1 deletion(-)

diff --git a/include/net/xdp.h b/include/net/xdp.h
index 40c6d3398458..3cc6d5d84aa4 100644
--- a/include/net/xdp.h
+++ b/include/net/xdp.h
@@ -181,4 +181,6 @@ 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
+
 #endif /* __LINUX_NET_XDP_H__ */
diff --git a/kernel/bpf/devmap.c b/kernel/bpf/devmap.c
index 58bdca5d978a..a51d9fb7a359 100644
--- a/kernel/bpf/devmap.c
+++ b/kernel/bpf/devmap.c
@@ -52,7 +52,6 @@
 #define DEV_CREATE_FLAG_MASK \
 	(BPF_F_NUMA_NODE | BPF_F_RDONLY | BPF_F_WRONLY)
 
-#define DEV_MAP_BULK_SIZE 16
 struct xdp_dev_bulk_queue {
 	struct xdp_frame *q[DEV_MAP_BULK_SIZE];
 	struct list_head flush_node;
-- 
2.17.1

[PATCH v2 net-next 2/5] dpaa2-eth: return num_enqueued frames from enqueue callback

[Ioana Ciornei]

The enqueue dpaa2-eth callback now returns the number of successfully
enqueued frames. This is a preliminary patch necessary for adding
support for bulk ring mode enqueue.

Signed-off-by: Ioana Ciornei <ioana.ciornei@nxp.com>
---
Changes in v2:
 - none

 .../net/ethernet/freescale/dpaa2/dpaa2-eth.c  | 34 +++++++++++++------
 .../net/ethernet/freescale/dpaa2/dpaa2-eth.h  |  5 +--
 2 files changed, 26 insertions(+), 13 deletions(-)

diff --git a/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.c b/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.c
index b6c46639aa4c..7b41ece8f160 100644
--- a/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.c
+++ b/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.c
@@ -1,6 +1,6 @@
 // SPDX-License-Identifier: (GPL-2.0+ OR BSD-3-Clause)
 /* Copyright 2014-2016 Freescale Semiconductor Inc.
- * Copyright 2016-2019 NXP
+ * Copyright 2016-2020 NXP
  */
 #include <linux/init.h>
 #include <linux/module.h>
@@ -268,7 +268,7 @@ static int xdp_enqueue(struct dpaa2_eth_priv *priv, struct dpaa2_fd *fd,
 
 	fq = &priv->fq[queue_id];
 	for (i = 0; i < DPAA2_ETH_ENQUEUE_RETRIES; i++) {
-		err = priv->enqueue(priv, fq, fd, 0);
+		err = priv->enqueue(priv, fq, fd, 0, NULL);
 		if (err != -EBUSY)
 			break;
 	}
@@ -847,7 +847,7 @@ static netdev_tx_t dpaa2_eth_tx(struct sk_buff *skb, struct net_device *net_dev)
 	 * the Tx confirmation callback for this frame
 	 */
 	for (i = 0; i < DPAA2_ETH_ENQUEUE_RETRIES; i++) {
-		err = priv->enqueue(priv, fq, &fd, prio);
+		err = priv->enqueue(priv, fq, &fd, prio, NULL);
 		if (err != -EBUSY)
 			break;
 	}
@@ -1937,7 +1937,7 @@ static int dpaa2_eth_xdp_xmit_frame(struct net_device *net_dev,
 
 	fq = &priv->fq[smp_processor_id() % dpaa2_eth_queue_count(priv)];
 	for (i = 0; i < DPAA2_ETH_ENQUEUE_RETRIES; i++) {
-		err = priv->enqueue(priv, fq, &fd, 0);
+		err = priv->enqueue(priv, fq, &fd, 0, NULL);
 		if (err != -EBUSY)
 			break;
 	}
@@ -2523,19 +2523,31 @@ static int set_buffer_layout(struct dpaa2_eth_priv *priv)
 
 static inline int dpaa2_eth_enqueue_qd(struct dpaa2_eth_priv *priv,
 				       struct dpaa2_eth_fq *fq,
-				       struct dpaa2_fd *fd, u8 prio)
+				       struct dpaa2_fd *fd, u8 prio,
+				       int *frames_enqueued)
 {
-	return dpaa2_io_service_enqueue_qd(fq->channel->dpio,
-					   priv->tx_qdid, prio,
-					   fq->tx_qdbin, fd);
+	int err;
+
+	err = dpaa2_io_service_enqueue_qd(fq->channel->dpio,
+					  priv->tx_qdid, prio,
+					  fq->tx_qdbin, fd);
+	if (!err && frames_enqueued)
+		*frames_enqueued = 1;
+	return err;
 }
 
 static inline int dpaa2_eth_enqueue_fq(struct dpaa2_eth_priv *priv,
 				       struct dpaa2_eth_fq *fq,
-				       struct dpaa2_fd *fd, u8 prio)
+				       struct dpaa2_fd *fd, u8 prio,
+				       int *frames_enqueued)
 {
-	return dpaa2_io_service_enqueue_fq(fq->channel->dpio,
-					   fq->tx_fqid[prio], fd);
+	int err;
+
+	err = dpaa2_io_service_enqueue_fq(fq->channel->dpio,
+					  fq->tx_fqid[prio], fd);
+	if (!err && frames_enqueued)
+		*frames_enqueued = 1;
+	return err;
 }
 
 static void set_enqueue_mode(struct dpaa2_eth_priv *priv)
diff --git a/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.h b/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.h
index 7635db3ef903..085ff750e4b5 100644
--- a/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.h
+++ b/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.h
@@ -1,6 +1,6 @@
 /* SPDX-License-Identifier: (GPL-2.0+ OR BSD-3-Clause) */
 /* Copyright 2014-2016 Freescale Semiconductor Inc.
- * Copyright 2016 NXP
+ * Copyright 2016-2020 NXP
  */
 
 #ifndef __DPAA2_ETH_H
@@ -371,7 +371,8 @@ struct dpaa2_eth_priv {
 	struct dpaa2_eth_fq fq[DPAA2_ETH_MAX_QUEUES];
 	int (*enqueue)(struct dpaa2_eth_priv *priv,
 		       struct dpaa2_eth_fq *fq,
-		       struct dpaa2_fd *fd, u8 prio);
+		       struct dpaa2_fd *fd, u8 prio,
+		       int *frames_enqueued);
 
 	u8 num_channels;
 	struct dpaa2_eth_channel *channel[DPAA2_ETH_MAX_DPCONS];
-- 
2.17.1

[PATCH v2 net-next 3/5] dpaa2-eth: use the bulk ring mode enqueue interface

[Ioana Ciornei]

Update the dpaa2-eth driver to use the bulk enqueue function introduced
with the change to QBMAN ring mode. At the moment, no functional changes
are made but rather the driver just transitions to the new interface
while still enqueuing just one frame at a time.

Signed-off-by: Ioana Ciornei <ioana.ciornei@nxp.com>
---
Changes in v2:
 - none

 .../net/ethernet/freescale/dpaa2/dpaa2-eth.c  | 35 +++++++++++--------
 .../net/ethernet/freescale/dpaa2/dpaa2-eth.h  |  1 +
 2 files changed, 22 insertions(+), 14 deletions(-)

diff --git a/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.c b/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.c
index 7b41ece8f160..26c2868435d5 100644
--- a/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.c
+++ b/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.c
@@ -268,7 +268,7 @@ static int xdp_enqueue(struct dpaa2_eth_priv *priv, struct dpaa2_fd *fd,
 
 	fq = &priv->fq[queue_id];
 	for (i = 0; i < DPAA2_ETH_ENQUEUE_RETRIES; i++) {
-		err = priv->enqueue(priv, fq, fd, 0, NULL);
+		err = priv->enqueue(priv, fq, fd, 0, 1, NULL);
 		if (err != -EBUSY)
 			break;
 	}
@@ -847,7 +847,7 @@ static netdev_tx_t dpaa2_eth_tx(struct sk_buff *skb, struct net_device *net_dev)
 	 * the Tx confirmation callback for this frame
 	 */
 	for (i = 0; i < DPAA2_ETH_ENQUEUE_RETRIES; i++) {
-		err = priv->enqueue(priv, fq, &fd, prio, NULL);
+		err = priv->enqueue(priv, fq, &fd, prio, 1, NULL);
 		if (err != -EBUSY)
 			break;
 	}
@@ -1937,7 +1937,7 @@ static int dpaa2_eth_xdp_xmit_frame(struct net_device *net_dev,
 
 	fq = &priv->fq[smp_processor_id() % dpaa2_eth_queue_count(priv)];
 	for (i = 0; i < DPAA2_ETH_ENQUEUE_RETRIES; i++) {
-		err = priv->enqueue(priv, fq, &fd, 0, NULL);
+		err = priv->enqueue(priv, fq, &fd, 0, 1, NULL);
 		if (err != -EBUSY)
 			break;
 	}
@@ -2524,6 +2524,7 @@ static int set_buffer_layout(struct dpaa2_eth_priv *priv)
 static inline int dpaa2_eth_enqueue_qd(struct dpaa2_eth_priv *priv,
 				       struct dpaa2_eth_fq *fq,
 				       struct dpaa2_fd *fd, u8 prio,
+				       u32 num_frames __always_unused,
 				       int *frames_enqueued)
 {
 	int err;
@@ -2536,18 +2537,24 @@ static inline int dpaa2_eth_enqueue_qd(struct dpaa2_eth_priv *priv,
 	return err;
 }
 
-static inline int dpaa2_eth_enqueue_fq(struct dpaa2_eth_priv *priv,
-				       struct dpaa2_eth_fq *fq,
-				       struct dpaa2_fd *fd, u8 prio,
-				       int *frames_enqueued)
+static inline int dpaa2_eth_enqueue_fq_multiple(struct dpaa2_eth_priv *priv,
+						struct dpaa2_eth_fq *fq,
+						struct dpaa2_fd *fd,
+						u8 prio, u32 num_frames,
+						int *frames_enqueued)
 {
 	int err;
 
-	err = dpaa2_io_service_enqueue_fq(fq->channel->dpio,
-					  fq->tx_fqid[prio], fd);
-	if (!err && frames_enqueued)
-		*frames_enqueued = 1;
-	return err;
+	err = dpaa2_io_service_enqueue_multiple_fq(fq->channel->dpio,
+						   fq->tx_fqid[prio],
+						   fd, num_frames);
+
+	if (err == 0)
+		return -EBUSY;
+
+	if (frames_enqueued)
+		*frames_enqueued = err;
+	return 0;
 }
 
 static void set_enqueue_mode(struct dpaa2_eth_priv *priv)
@@ -2556,7 +2563,7 @@ static void set_enqueue_mode(struct dpaa2_eth_priv *priv)
 				   DPNI_ENQUEUE_FQID_VER_MINOR) < 0)
 		priv->enqueue = dpaa2_eth_enqueue_qd;
 	else
-		priv->enqueue = dpaa2_eth_enqueue_fq;
+		priv->enqueue = dpaa2_eth_enqueue_fq_multiple;
 }
 
 static int set_pause(struct dpaa2_eth_priv *priv)
@@ -2617,7 +2624,7 @@ static void update_tx_fqids(struct dpaa2_eth_priv *priv)
 		}
 	}
 
-	priv->enqueue = dpaa2_eth_enqueue_fq;
+	priv->enqueue = dpaa2_eth_enqueue_fq_multiple;
 
 	return;
 
diff --git a/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.h b/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.h
index 085ff750e4b5..2440ba6b21ef 100644
--- a/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.h
+++ b/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.h
@@ -372,6 +372,7 @@ struct dpaa2_eth_priv {
 	int (*enqueue)(struct dpaa2_eth_priv *priv,
 		       struct dpaa2_eth_fq *fq,
 		       struct dpaa2_fd *fd, u8 prio,
+		       u32 num_frames,
 		       int *frames_enqueued);
 
 	u8 num_channels;
-- 
2.17.1

[PATCH v2 net-next 4/5] dpaa2-eth: split the .ndo_xdp_xmit callback into two stages

[Ioana Ciornei]

Instead of having a function that both creates a frame descriptor from
an xdp_frame and enqueues it, split this into two stages.
Add the dpaa2_eth_xdp_create_fd that just transforms an xdp_frame into a
FD while the actual enqueue callback is called directly from the ndo for
each frame.
This is particulary useful in conjunction with bulk enqueue.

Signed-off-by: Ioana Ciornei <ioana.ciornei@nxp.com>
---
Changes in v2:
 - none

 .../net/ethernet/freescale/dpaa2/dpaa2-eth.c  | 76 ++++++++++---------
 1 file changed, 40 insertions(+), 36 deletions(-)

diff --git a/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.c b/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.c
index 26c2868435d5..9a0432cd893c 100644
--- a/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.c
+++ b/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.c
@@ -1880,20 +1880,16 @@ static int dpaa2_eth_xdp(struct net_device *dev, struct netdev_bpf *xdp)
 	return 0;
 }
 
-static int dpaa2_eth_xdp_xmit_frame(struct net_device *net_dev,
-				    struct xdp_frame *xdpf)
+static int dpaa2_eth_xdp_create_fd(struct net_device *net_dev,
+				   struct xdp_frame *xdpf,
+				   struct dpaa2_fd *fd)
 {
 	struct dpaa2_eth_priv *priv = netdev_priv(net_dev);
 	struct device *dev = net_dev->dev.parent;
-	struct rtnl_link_stats64 *percpu_stats;
-	struct dpaa2_eth_drv_stats *percpu_extras;
 	unsigned int needed_headroom;
 	struct dpaa2_eth_swa *swa;
-	struct dpaa2_eth_fq *fq;
-	struct dpaa2_fd fd;
 	void *buffer_start, *aligned_start;
 	dma_addr_t addr;
-	int err, i;
 
 	/* We require a minimum headroom to be able to transmit the frame.
 	 * Otherwise return an error and let the original net_device handle it
@@ -1902,11 +1898,8 @@ static int dpaa2_eth_xdp_xmit_frame(struct net_device *net_dev,
 	if (xdpf->headroom < needed_headroom)
 		return -EINVAL;
 
-	percpu_stats = this_cpu_ptr(priv->percpu_stats);
-	percpu_extras = this_cpu_ptr(priv->percpu_extras);
-
 	/* Setup the FD fields */
-	memset(&fd, 0, sizeof(fd));
+	memset(fd, 0, sizeof(*fd));
 
 	/* Align FD address, if possible */
 	buffer_start = xdpf->data - needed_headroom;
@@ -1924,32 +1917,14 @@ static int dpaa2_eth_xdp_xmit_frame(struct net_device *net_dev,
 	addr = dma_map_single(dev, buffer_start,
 			      swa->xdp.dma_size,
 			      DMA_BIDIRECTIONAL);
-	if (unlikely(dma_mapping_error(dev, addr))) {
-		percpu_stats->tx_dropped++;
+	if (unlikely(dma_mapping_error(dev, addr)))
 		return -ENOMEM;
-	}
-
-	dpaa2_fd_set_addr(&fd, addr);
-	dpaa2_fd_set_offset(&fd, xdpf->data - buffer_start);
-	dpaa2_fd_set_len(&fd, xdpf->len);
-	dpaa2_fd_set_format(&fd, dpaa2_fd_single);
-	dpaa2_fd_set_ctrl(&fd, FD_CTRL_PTA);
-
-	fq = &priv->fq[smp_processor_id() % dpaa2_eth_queue_count(priv)];
-	for (i = 0; i < DPAA2_ETH_ENQUEUE_RETRIES; i++) {
-		err = priv->enqueue(priv, fq, &fd, 0, 1, NULL);
-		if (err != -EBUSY)
-			break;
-	}
-	percpu_extras->tx_portal_busy += i;
-	if (unlikely(err < 0)) {
-		percpu_stats->tx_errors++;
-		/* let the Rx device handle the cleanup */
-		return err;
-	}
 
-	percpu_stats->tx_packets++;
-	percpu_stats->tx_bytes += dpaa2_fd_get_len(&fd);
+	dpaa2_fd_set_addr(fd, addr);
+	dpaa2_fd_set_offset(fd, xdpf->data - buffer_start);
+	dpaa2_fd_set_len(fd, xdpf->len);
+	dpaa2_fd_set_format(fd, dpaa2_fd_single);
+	dpaa2_fd_set_ctrl(fd, FD_CTRL_PTA);
 
 	return 0;
 }
@@ -1957,6 +1932,11 @@ static int dpaa2_eth_xdp_xmit_frame(struct net_device *net_dev,
 static int dpaa2_eth_xdp_xmit(struct net_device *net_dev, int n,
 			      struct xdp_frame **frames, u32 flags)
 {
+	struct dpaa2_eth_priv *priv = netdev_priv(net_dev);
+	struct dpaa2_eth_drv_stats *percpu_extras;
+	struct rtnl_link_stats64 *percpu_stats;
+	struct dpaa2_eth_fq *fq;
+	struct dpaa2_fd fd;
 	int drops = 0;
 	int i, err;
 
@@ -1966,14 +1946,38 @@ static int dpaa2_eth_xdp_xmit(struct net_device *net_dev, int n,
 	if (!netif_running(net_dev))
 		return -ENETDOWN;
 
+	percpu_stats = this_cpu_ptr(priv->percpu_stats);
+	percpu_extras = this_cpu_ptr(priv->percpu_extras);
+
 	for (i = 0; i < n; i++) {
 		struct xdp_frame *xdpf = frames[i];
 
-		err = dpaa2_eth_xdp_xmit_frame(net_dev, xdpf);
+		/* create the FD from the xdp_frame */
+		err = dpaa2_eth_xdp_create_fd(net_dev, xdpf, &fd);
 		if (err) {
+			percpu_stats->tx_dropped++;
 			xdp_return_frame_rx_napi(xdpf);
 			drops++;
+			continue;
+		}
+
+		/* enqueue the newly created FD */
+		fq = &priv->fq[smp_processor_id() % dpaa2_eth_queue_count(priv)];
+		for (i = 0; i < DPAA2_ETH_ENQUEUE_RETRIES; i++) {
+			err = priv->enqueue(priv, fq, &fd, 0, 1);
+			if (err != -EBUSY)
+				break;
 		}
+
+		percpu_extras->tx_portal_busy += i;
+		if (unlikely(err < 0)) {
+			percpu_stats->tx_errors++;
+			xdp_return_frame_rx_napi(xdpf);
+			continue;
+		}
+
+		percpu_stats->tx_packets++;
+		percpu_stats->tx_bytes += dpaa2_fd_get_len(&fd);
 	}
 
 	return n - drops;
-- 
2.17.1

[PATCH v2 net-next 5/5] dpaa2-eth: use bulk enqueue in .ndo_xdp_xmit

[Ioana Ciornei]

Take advantage of the bulk enqueue feature in .ndo_xdp_xmit.
We cannot use the XDP_XMIT_FLUSH since the architecture is not capable
to store all the frames dequeued in a NAPI cycle so we instead are
enqueueing all the frames received in a ndo_xdp_xmit call right away.

After setting up all FDs for the xdp_frames received, enqueue multiple
frames at a time until all are sent or the maximum number of retries is
hit.

Signed-off-by: Ioana Ciornei <ioana.ciornei@nxp.com>
---
Changes in v2:
 - use a statically allocated array of dpaa2_fd for each fq
 - use DEV_MAP_BULK_SIZE as the max number of xdp_frames received

 .../net/ethernet/freescale/dpaa2/dpaa2-eth.c  | 57 +++++++++----------
 .../net/ethernet/freescale/dpaa2/dpaa2-eth.h  |  2 +
 2 files changed, 30 insertions(+), 29 deletions(-)

diff --git a/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.c b/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.c
index 9a0432cd893c..9d4061bba0b8 100644
--- a/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.c
+++ b/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.c
@@ -1933,12 +1933,12 @@ static int dpaa2_eth_xdp_xmit(struct net_device *net_dev, int n,
 			      struct xdp_frame **frames, u32 flags)
 {
 	struct dpaa2_eth_priv *priv = netdev_priv(net_dev);
+	int total_enqueued = 0, retries = 0, enqueued;
 	struct dpaa2_eth_drv_stats *percpu_extras;
 	struct rtnl_link_stats64 *percpu_stats;
+	int num_fds, i, err, max_retries;
 	struct dpaa2_eth_fq *fq;
-	struct dpaa2_fd fd;
-	int drops = 0;
-	int i, err;
+	struct dpaa2_fd *fds;
 
 	if (unlikely(flags & ~XDP_XMIT_FLAGS_MASK))
 		return -EINVAL;
@@ -1946,41 +1946,40 @@ static int dpaa2_eth_xdp_xmit(struct net_device *net_dev, int n,
 	if (!netif_running(net_dev))
 		return -ENETDOWN;
 
+	fq = &priv->fq[smp_processor_id()];
+	fds = fq->xdp_fds;
+
 	percpu_stats = this_cpu_ptr(priv->percpu_stats);
 	percpu_extras = this_cpu_ptr(priv->percpu_extras);
 
+	/* create a FD for each xdp_frame in the list received */
 	for (i = 0; i < n; i++) {
-		struct xdp_frame *xdpf = frames[i];
-
-		/* create the FD from the xdp_frame */
-		err = dpaa2_eth_xdp_create_fd(net_dev, xdpf, &fd);
-		if (err) {
-			percpu_stats->tx_dropped++;
-			xdp_return_frame_rx_napi(xdpf);
-			drops++;
-			continue;
-		}
-
-		/* enqueue the newly created FD */
-		fq = &priv->fq[smp_processor_id() % dpaa2_eth_queue_count(priv)];
-		for (i = 0; i < DPAA2_ETH_ENQUEUE_RETRIES; i++) {
-			err = priv->enqueue(priv, fq, &fd, 0, 1);
-			if (err != -EBUSY)
-				break;
-		}
+		err = dpaa2_eth_xdp_create_fd(net_dev, frames[i], &fds[i]);
+		if (err)
+			break;
+	}
+	num_fds = i;
 
-		percpu_extras->tx_portal_busy += i;
-		if (unlikely(err < 0)) {
-			percpu_stats->tx_errors++;
-			xdp_return_frame_rx_napi(xdpf);
+	/* try to enqueue all the FDs until the max number of retries is hit */
+	max_retries = num_fds * DPAA2_ETH_ENQUEUE_RETRIES;
+	while (total_enqueued < num_fds && retries < max_retries) {
+		err = priv->enqueue(priv, fq, &fds[total_enqueued],
+				    0, num_fds - total_enqueued, &enqueued);
+		if (err == -EBUSY) {
+			percpu_extras->tx_portal_busy += ++retries;
 			continue;
 		}
-
-		percpu_stats->tx_packets++;
-		percpu_stats->tx_bytes += dpaa2_fd_get_len(&fd);
+		total_enqueued += enqueued;
 	}
 
-	return n - drops;
+	/* update statistics */
+	percpu_stats->tx_packets += total_enqueued;
+	for (i = 0; i < total_enqueued; i++)
+		percpu_stats->tx_bytes += dpaa2_fd_get_len(&fds[i]);
+	for (i = total_enqueued; i < n; i++)
+		xdp_return_frame_rx_napi(frames[i]);
+
+	return total_enqueued;
 }
 
 static int update_xps(struct dpaa2_eth_priv *priv)
diff --git a/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.h b/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.h
index 2440ba6b21ef..289053099974 100644
--- a/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.h
+++ b/drivers/net/ethernet/freescale/dpaa2/dpaa2-eth.h
@@ -325,6 +325,8 @@ struct dpaa2_eth_fq {
 			const struct dpaa2_fd *fd,
 			struct dpaa2_eth_fq *fq);
 	struct dpaa2_eth_fq_stats stats;
+
+	struct dpaa2_fd xdp_fds[DEV_MAP_BULK_SIZE];
 };
 
 struct dpaa2_eth_ch_xdp {
-- 
2.17.1

Re: [PATCH v2 net-next 1/5] xdp: export the DEV_MAP_BULK_SIZE macro

[Jesper Dangaard Brouer]

On Wed, 22 Apr 2020 15:05:09 +0300
Ioana Ciornei <ioana.ciornei@nxp.com> wrote:

> Export the DEV_MAP_BULK_SIZE macro to the header file so that drivers
> can directly use it as the maximum number of xdp_frames received in the
> .ndo_xdp_xmit() callback.
> 
> Signed-off-by: Ioana Ciornei <ioana.ciornei@nxp.com>

Acked-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 v2 net-next 5/5] dpaa2-eth: use bulk enqueue in .ndo_xdp_xmit

[Jesper Dangaard Brouer]

On Wed, 22 Apr 2020 15:05:13 +0300
Ioana Ciornei <ioana.ciornei@nxp.com> wrote:

> Take advantage of the bulk enqueue feature in .ndo_xdp_xmit.
> We cannot use the XDP_XMIT_FLUSH since the architecture is not capable
> to store all the frames dequeued in a NAPI cycle so we instead are
> enqueueing all the frames received in a ndo_xdp_xmit call right away.
> 
> After setting up all FDs for the xdp_frames received, enqueue multiple
> frames at a time until all are sent or the maximum number of retries is
> hit.
> 
> Signed-off-by: Ioana Ciornei <ioana.ciornei@nxp.com>
> ---
> Changes in v2:
>  - use a statically allocated array of dpaa2_fd for each fq
>  - use DEV_MAP_BULK_SIZE as the max number of xdp_frames received

Thanks for doing this.

Acked-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 v2 net-next 4/5] dpaa2-eth: split the .ndo_xdp_xmit callback into two stages

[Jesper Dangaard Brouer]

On Wed, 22 Apr 2020 15:05:12 +0300
Ioana Ciornei <ioana.ciornei@nxp.com> wrote:

> Instead of having a function that both creates a frame descriptor from
> an xdp_frame and enqueues it, split this into two stages.
> Add the dpaa2_eth_xdp_create_fd that just transforms an xdp_frame into a
> FD while the actual enqueue callback is called directly from the ndo for
> each frame.
> This is particulary useful in conjunction with bulk enqueue.
> 
> Signed-off-by: Ioana Ciornei <ioana.ciornei@nxp.com>

LGTM - and I very recently had to do a deep dive into this driver code.

Acked-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 v2 net-next 0/5] dpaa2-eth: add support for xdp bulk enqueue

[David Miller]

From: Ioana Ciornei <ioana.ciornei@nxp.com>
Date: Wed, 22 Apr 2020 15:05:08 +0300

> The first patch moves the DEV_MAP_BULK_SIZE macro into the xdp.h header
> file so that drivers can take advantage of it and use it.
> 
> The following 3 patches are there to setup the scene for using the bulk
> enqueue feature.  First of all, the prototype of the enqueue function is
> changed so that it returns the number of enqueued frames. Second, the
> bulk enqueue interface is used but without any functional changes, still
> one frame at a time is enqueued.  Third, the .ndo_xdp_xmit callback is
> split into two stages, create all FDs for the xdp_frames received and
> then enqueue them.
> 
> The last patch of the series builds on top of the others and instead of
> issuing an enqueue operation for each FD it issues a bulk enqueue call
> for as many frames as possible. This is repeated until all frames are
> enqueued or the maximum number of retries is hit. We do not use the
> XDP_XMIT_FLUSH flag since the architecture is not capable to store all
> frames dequeued in a NAPI cycle, instead we send out right away all
> frames received in a .ndo_xdp_xmit call.
> 
> Changes in v2:
>  - statically allocate an array of dpaa2_fd by frame queue
>  - use the DEV_MAP_BULK_SIZE as the maximum number of xdp_frames
>    received in .ndo_xdp_xmit()

Series applied, thanks.