[PATCH net-next v3 0/7] dpaa_eth: add XDP support

[PATCH net-next v3 0/7] dpaa_eth: add XDP support

[Camelia Groza]

Enable XDP support for the QorIQ DPAA1 platforms.

Implement all the current actions (DROP, ABORTED, PASS, TX, REDIRECT). No
Tx batching is added at this time.

Additional XDP_PACKET_HEADROOM bytes are reserved in each frame's headroom.

After transmit, a reference to the xdp_frame is saved in the buffer for
clean-up on confirmation in a newly created structure for software
annotations.

Changes in v2:
- warn only once if extracting the timestamp from a received frame fails
  in 2/7

Changes in v3:
- drop received S/G frames when XDP is enabled in 2/7

Camelia Groza (7):
  dpaa_eth: add struct for software backpointers
  dpaa_eth: add basic XDP support
  dpaa_eth: limit the possible MTU range when XDP is enabled
  dpaa_eth: add XDP_TX support
  dpaa_eth: add XDP_REDIRECT support
  dpaa_eth: rename current skb A050385 erratum workaround
  dpaa_eth: implement the A050385 erratum workaround for XDP

 drivers/net/ethernet/freescale/dpaa/dpaa_eth.c | 457 +++++++++++++++++++++++--
 drivers/net/ethernet/freescale/dpaa/dpaa_eth.h |  13 +
 2 files changed, 440 insertions(+), 30 deletions(-)

--
1.9.1

[PATCH net-next v3 1/7] dpaa_eth: add struct for software backpointers

[Camelia Groza]

We maintain an skb backpointer in the software annotations area of Tx
frames. Introduce a structure for explicit handling.

Acked-by: Madalin Bucur <madalin.bucur@oss.nxp.com>
Signed-off-by: Camelia Groza <camelia.groza@nxp.com>
---
 drivers/net/ethernet/freescale/dpaa/dpaa_eth.c | 16 +++++++++-------
 drivers/net/ethernet/freescale/dpaa/dpaa_eth.h |  8 ++++++++
 2 files changed, 17 insertions(+), 7 deletions(-)

diff --git a/drivers/net/ethernet/freescale/dpaa/dpaa_eth.c b/drivers/net/ethernet/freescale/dpaa/dpaa_eth.c
index 8867693..88533a2 100644
--- a/drivers/net/ethernet/freescale/dpaa/dpaa_eth.c
+++ b/drivers/net/ethernet/freescale/dpaa/dpaa_eth.c
@@ -1633,6 +1633,7 @@ static struct sk_buff *dpaa_cleanup_tx_fd(const struct dpaa_priv *priv,
 	dma_addr_t addr = qm_fd_addr(fd);
 	void *vaddr = phys_to_virt(addr);
 	const struct qm_sg_entry *sgt;
+	struct dpaa_eth_swbp *swbp;
 	struct sk_buff *skb;
 	u64 ns;
 	int i;
@@ -1665,7 +1666,8 @@ static struct sk_buff *dpaa_cleanup_tx_fd(const struct dpaa_priv *priv,
 				 dma_dir);
 	}
 
-	skb = *(struct sk_buff **)vaddr;
+	swbp = (struct dpaa_eth_swbp *)vaddr;
+	skb = swbp->skb;
 
 	/* DMA unmapping is required before accessing the HW provided info */
 	if (ts && priv->tx_tstamp &&
@@ -1879,8 +1881,8 @@ static int skb_to_contig_fd(struct dpaa_priv *priv,
 {
 	struct net_device *net_dev = priv->net_dev;
 	enum dma_data_direction dma_dir;
+	struct dpaa_eth_swbp *swbp;
 	unsigned char *buff_start;
-	struct sk_buff **skbh;
 	dma_addr_t addr;
 	int err;
 
@@ -1891,8 +1893,8 @@ static int skb_to_contig_fd(struct dpaa_priv *priv,
 	buff_start = skb->data - priv->tx_headroom;
 	dma_dir = DMA_TO_DEVICE;
 
-	skbh = (struct sk_buff **)buff_start;
-	*skbh = skb;
+	swbp = (struct dpaa_eth_swbp *)buff_start;
+	swbp->skb = skb;
 
 	/* Enable L3/L4 hardware checksum computation.
 	 *
@@ -1931,8 +1933,8 @@ static int skb_to_sg_fd(struct dpaa_priv *priv,
 	const enum dma_data_direction dma_dir = DMA_TO_DEVICE;
 	const int nr_frags = skb_shinfo(skb)->nr_frags;
 	struct net_device *net_dev = priv->net_dev;
+	struct dpaa_eth_swbp *swbp;
 	struct qm_sg_entry *sgt;
-	struct sk_buff **skbh;
 	void *buff_start;
 	skb_frag_t *frag;
 	dma_addr_t addr;
@@ -2005,8 +2007,8 @@ static int skb_to_sg_fd(struct dpaa_priv *priv,
 	qm_fd_set_sg(fd, priv->tx_headroom, skb->len);
 
 	/* DMA map the SGT page */
-	skbh = (struct sk_buff **)buff_start;
-	*skbh = skb;
+	swbp = (struct dpaa_eth_swbp *)buff_start;
+	swbp->skb = skb;
 
 	addr = dma_map_page(priv->tx_dma_dev, p, 0,
 			    priv->tx_headroom + DPAA_SGT_SIZE, dma_dir);
diff --git a/drivers/net/ethernet/freescale/dpaa/dpaa_eth.h b/drivers/net/ethernet/freescale/dpaa/dpaa_eth.h
index fc2cc4c..da30e5d 100644
--- a/drivers/net/ethernet/freescale/dpaa/dpaa_eth.h
+++ b/drivers/net/ethernet/freescale/dpaa/dpaa_eth.h
@@ -144,6 +144,14 @@ struct dpaa_buffer_layout {
 	u16 priv_data_size;
 };
 
+/* Information to be used on the Tx confirmation path. Stored just
+ * before the start of the transmit buffer. Maximum size allowed
+ * is DPAA_TX_PRIV_DATA_SIZE bytes.
+ */
+struct dpaa_eth_swbp {
+	struct sk_buff *skb;
+};
+
 struct dpaa_priv {
 	struct dpaa_percpu_priv __percpu *percpu_priv;
 	struct dpaa_bp *dpaa_bp;
-- 
1.9.1

[PATCH net-next v3 2/7] dpaa_eth: add basic XDP support

[Camelia Groza]

Implement the XDP_DROP and XDP_PASS actions.

Avoid draining and reconfiguring the buffer pool at each XDP
setup/teardown by increasing the frame headroom and reserving
XDP_PACKET_HEADROOM bytes from the start. Since we always reserve an
entire page per buffer, this change only impacts Jumbo frame scenarios
where the maximum linear frame size is reduced by 256 bytes. Multi
buffer Scatter/Gather frames are now used instead in these scenarios.

Allow XDP programs to access the entire buffer.

The data in the received frame's headroom can be overwritten by the XDP
program. Extract the relevant fields from the headroom while they are
still available, before running the XDP program.

Since the headroom might be resized before the frame is passed up to the
stack, remove the check for a fixed headroom value when building an skb.

Allow the meta data to be updated and pass the information up the stack.

Scatter/Gather frames are dropped when XDP is enabled.

Acked-by: Madalin Bucur <madalin.bucur@oss.nxp.com>
Signed-off-by: Camelia Groza <camelia.groza@nxp.com>
---
Changes in v2:
- warn only once if extracting the timestamp from a received frame fails

Changes in v3:
- drop received S/G frames when XDP is enabled

 drivers/net/ethernet/freescale/dpaa/dpaa_eth.c | 158 ++++++++++++++++++++++---
 drivers/net/ethernet/freescale/dpaa/dpaa_eth.h |   2 +
 2 files changed, 144 insertions(+), 16 deletions(-)

diff --git a/drivers/net/ethernet/freescale/dpaa/dpaa_eth.c b/drivers/net/ethernet/freescale/dpaa/dpaa_eth.c
index 88533a2..102023c 100644
--- a/drivers/net/ethernet/freescale/dpaa/dpaa_eth.c
+++ b/drivers/net/ethernet/freescale/dpaa/dpaa_eth.c
@@ -53,6 +53,8 @@
 #include <linux/dma-mapping.h>
 #include <linux/sort.h>
 #include <linux/phy_fixed.h>
+#include <linux/bpf.h>
+#include <linux/bpf_trace.h>
 #include <soc/fsl/bman.h>
 #include <soc/fsl/qman.h>
 #include "fman.h"
@@ -177,7 +179,7 @@
 #define DPAA_HWA_SIZE (DPAA_PARSE_RESULTS_SIZE + DPAA_TIME_STAMP_SIZE \
 		       + DPAA_HASH_RESULTS_SIZE)
 #define DPAA_RX_PRIV_DATA_DEFAULT_SIZE (DPAA_TX_PRIV_DATA_SIZE + \
-					dpaa_rx_extra_headroom)
+					XDP_PACKET_HEADROOM - DPAA_HWA_SIZE)
 #ifdef CONFIG_DPAA_ERRATUM_A050385
 #define DPAA_RX_PRIV_DATA_A050385_SIZE (DPAA_A050385_ALIGN - DPAA_HWA_SIZE)
 #define DPAA_RX_PRIV_DATA_SIZE (fman_has_errata_a050385() ? \
@@ -1733,7 +1735,6 @@ static struct sk_buff *contig_fd_to_skb(const struct dpaa_priv *priv,
 			SKB_DATA_ALIGN(sizeof(struct skb_shared_info)));
 	if (WARN_ONCE(!skb, "Build skb failure on Rx\n"))
 		goto free_buffer;
-	WARN_ON(fd_off != priv->rx_headroom);
 	skb_reserve(skb, fd_off);
 	skb_put(skb, qm_fd_get_length(fd));

@@ -2349,12 +2350,62 @@ static enum qman_cb_dqrr_result rx_error_dqrr(struct qman_portal *portal,
 	return qman_cb_dqrr_consume;
 }

+static u32 dpaa_run_xdp(struct dpaa_priv *priv, struct qm_fd *fd, void *vaddr,
+			unsigned int *xdp_meta_len)
+{
+	ssize_t fd_off = qm_fd_get_offset(fd);
+	struct bpf_prog *xdp_prog;
+	struct xdp_buff xdp;
+	u32 xdp_act;
+
+	rcu_read_lock();
+
+	xdp_prog = READ_ONCE(priv->xdp_prog);
+	if (!xdp_prog) {
+		rcu_read_unlock();
+		return XDP_PASS;
+	}
+
+	xdp.data = vaddr + fd_off;
+	xdp.data_meta = xdp.data;
+	xdp.data_hard_start = xdp.data - XDP_PACKET_HEADROOM;
+	xdp.data_end = xdp.data + qm_fd_get_length(fd);
+	xdp.frame_sz = DPAA_BP_RAW_SIZE;
+
+	xdp_act = bpf_prog_run_xdp(xdp_prog, &xdp);
+
+	/* Update the length and the offset of the FD */
+	qm_fd_set_contig(fd, xdp.data - vaddr, xdp.data_end - xdp.data);
+
+	switch (xdp_act) {
+	case XDP_PASS:
+		*xdp_meta_len = xdp.data - xdp.data_meta;
+		break;
+	default:
+		bpf_warn_invalid_xdp_action(xdp_act);
+		fallthrough;
+	case XDP_ABORTED:
+		trace_xdp_exception(priv->net_dev, xdp_prog, xdp_act);
+		fallthrough;
+	case XDP_DROP:
+		/* Free the buffer */
+		free_pages((unsigned long)vaddr, 0);
+		break;
+	}
+
+	rcu_read_unlock();
+
+	return xdp_act;
+}
+
 static enum qman_cb_dqrr_result rx_default_dqrr(struct qman_portal *portal,
 						struct qman_fq *fq,
 						const struct qm_dqrr_entry *dq,
 						bool sched_napi)
 {
+	bool ts_valid = false, hash_valid = false;
 	struct skb_shared_hwtstamps *shhwtstamps;
+	unsigned int skb_len, xdp_meta_len = 0;
 	struct rtnl_link_stats64 *percpu_stats;
 	struct dpaa_percpu_priv *percpu_priv;
 	const struct qm_fd *fd = &dq->fd;
@@ -2362,12 +2413,14 @@ static enum qman_cb_dqrr_result rx_default_dqrr(struct qman_portal *portal,
 	enum qm_fd_format fd_format;
 	struct net_device *net_dev;
 	u32 fd_status, hash_offset;
+	struct qm_sg_entry *sgt;
 	struct dpaa_bp *dpaa_bp;
 	struct dpaa_priv *priv;
-	unsigned int skb_len;
 	struct sk_buff *skb;
 	int *count_ptr;
+	u32 xdp_act;
 	void *vaddr;
+	u32 hash;
 	u64 ns;

 	fd_status = be32_to_cpu(fd->status);
@@ -2423,35 +2476,67 @@ static enum qman_cb_dqrr_result rx_default_dqrr(struct qman_portal *portal,
 	count_ptr = this_cpu_ptr(dpaa_bp->percpu_count);
 	(*count_ptr)--;

-	if (likely(fd_format == qm_fd_contig))
+	/* Extract the timestamp stored in the headroom before running XDP */
+	if (priv->rx_tstamp) {
+		if (!fman_port_get_tstamp(priv->mac_dev->port[RX], vaddr, &ns))
+			ts_valid = true;
+		else
+			WARN_ONCE(1, "fman_port_get_tstamp failed!\n");
+	}
+
+	/* Extract the hash stored in the headroom before running XDP */
+	if (net_dev->features & NETIF_F_RXHASH && priv->keygen_in_use &&
+	    !fman_port_get_hash_result_offset(priv->mac_dev->port[RX],
+					      &hash_offset)) {
+		hash = be32_to_cpu(*(u32 *)(vaddr + hash_offset));
+		hash_valid = true;
+	}
+
+	if (likely(fd_format == qm_fd_contig)) {
+		xdp_act = dpaa_run_xdp(priv, (struct qm_fd *)fd, vaddr,
+				       &xdp_meta_len);
+		if (xdp_act != XDP_PASS) {
+			percpu_stats->rx_packets++;
+			percpu_stats->rx_bytes += qm_fd_get_length(fd);
+			return qman_cb_dqrr_consume;
+		}
 		skb = contig_fd_to_skb(priv, fd);
-	else
+	} else {
+		/* XDP doesn't support S/G frames. Return the fragments to the
+		 * buffer pool and release the SGT.
+		 */
+		if (READ_ONCE(priv->xdp_prog)) {
+			WARN_ONCE(1, "S/G frames not supported under XDP\n");
+			sgt = vaddr + qm_fd_get_offset(fd);
+			dpaa_release_sgt_members(sgt);
+			free_pages((unsigned long)vaddr, 0);
+			return qman_cb_dqrr_consume;
+		}
 		skb = sg_fd_to_skb(priv, fd);
+	}
 	if (!skb)
 		return qman_cb_dqrr_consume;

-	if (priv->rx_tstamp) {
+	if (xdp_meta_len)
+		skb_metadata_set(skb, xdp_meta_len);
+
+	/* Set the previously extracted timestamp */
+	if (ts_valid) {
 		shhwtstamps = skb_hwtstamps(skb);
 		memset(shhwtstamps, 0, sizeof(*shhwtstamps));
-
-		if (!fman_port_get_tstamp(priv->mac_dev->port[RX], vaddr, &ns))
-			shhwtstamps->hwtstamp = ns_to_ktime(ns);
-		else
-			dev_warn(net_dev->dev.parent, "fman_port_get_tstamp failed!\n");
+		shhwtstamps->hwtstamp = ns_to_ktime(ns);
 	}

 	skb->protocol = eth_type_trans(skb, net_dev);

-	if (net_dev->features & NETIF_F_RXHASH && priv->keygen_in_use &&
-	    !fman_port_get_hash_result_offset(priv->mac_dev->port[RX],
-					      &hash_offset)) {
+	/* Set the previously extracted hash */
+	if (hash_valid) {
 		enum pkt_hash_types type;

 		/* if L4 exists, it was used in the hash generation */
 		type = be32_to_cpu(fd->status) & FM_FD_STAT_L4CV ?
 			PKT_HASH_TYPE_L4 : PKT_HASH_TYPE_L3;
-		skb_set_hash(skb, be32_to_cpu(*(u32 *)(vaddr + hash_offset)),
-			     type);
+		skb_set_hash(skb, hash, type);
 	}

 	skb_len = skb->len;
@@ -2671,6 +2756,46 @@ static int dpaa_eth_stop(struct net_device *net_dev)
 	return err;
 }

+static int dpaa_setup_xdp(struct net_device *net_dev, struct bpf_prog *prog)
+{
+	struct dpaa_priv *priv = netdev_priv(net_dev);
+	struct bpf_prog *old_prog;
+	bool up;
+	int err;
+
+	/* S/G fragments are not supported in XDP-mode */
+	if (prog && (priv->dpaa_bp->raw_size <
+	    net_dev->mtu + VLAN_ETH_HLEN + ETH_FCS_LEN))
+		return -EINVAL;
+
+	up = netif_running(net_dev);
+
+	if (up)
+		dpaa_eth_stop(net_dev);
+
+	old_prog = xchg(&priv->xdp_prog, prog);
+	if (old_prog)
+		bpf_prog_put(old_prog);
+
+	if (up) {
+		err = dpaa_open(net_dev);
+		if (err)
+			return err;
+	}
+
+	return 0;
+}
+
+static int dpaa_xdp(struct net_device *net_dev, struct netdev_bpf *xdp)
+{
+	switch (xdp->command) {
+	case XDP_SETUP_PROG:
+		return dpaa_setup_xdp(net_dev, xdp->prog);
+	default:
+		return -EINVAL;
+	}
+}
+
 static int dpaa_ts_ioctl(struct net_device *dev, struct ifreq *rq, int cmd)
 {
 	struct dpaa_priv *priv = netdev_priv(dev);
@@ -2737,6 +2862,7 @@ static int dpaa_ioctl(struct net_device *net_dev, struct ifreq *rq, int cmd)
 	.ndo_set_rx_mode = dpaa_set_rx_mode,
 	.ndo_do_ioctl = dpaa_ioctl,
 	.ndo_setup_tc = dpaa_setup_tc,
+	.ndo_bpf = dpaa_xdp,
 };

 static int dpaa_napi_add(struct net_device *net_dev)
diff --git a/drivers/net/ethernet/freescale/dpaa/dpaa_eth.h b/drivers/net/ethernet/freescale/dpaa/dpaa_eth.h
index da30e5d..94e8613 100644
--- a/drivers/net/ethernet/freescale/dpaa/dpaa_eth.h
+++ b/drivers/net/ethernet/freescale/dpaa/dpaa_eth.h
@@ -196,6 +196,8 @@ struct dpaa_priv {

 	bool tx_tstamp; /* Tx timestamping enabled */
 	bool rx_tstamp; /* Rx timestamping enabled */
+
+	struct bpf_prog *xdp_prog;
 };

 /* from dpaa_ethtool.c */
--
1.9.1

Re: [PATCH net-next v3 2/7] dpaa_eth: add basic XDP support

[Maciej Fijalkowski]

On Thu, Nov 19, 2020 at 06:29:31PM +0200, Camelia Groza wrote:
> Implement the XDP_DROP and XDP_PASS actions.
> 
> Avoid draining and reconfiguring the buffer pool at each XDP
> setup/teardown by increasing the frame headroom and reserving
> XDP_PACKET_HEADROOM bytes from the start. Since we always reserve an
> entire page per buffer, this change only impacts Jumbo frame scenarios
> where the maximum linear frame size is reduced by 256 bytes. Multi
> buffer Scatter/Gather frames are now used instead in these scenarios.
> 
> Allow XDP programs to access the entire buffer.
> 
> The data in the received frame's headroom can be overwritten by the XDP
> program. Extract the relevant fields from the headroom while they are
> still available, before running the XDP program.
> 
> Since the headroom might be resized before the frame is passed up to the
> stack, remove the check for a fixed headroom value when building an skb.
> 
> Allow the meta data to be updated and pass the information up the stack.
> 
> Scatter/Gather frames are dropped when XDP is enabled.
> 
> Acked-by: Madalin Bucur <madalin.bucur@oss.nxp.com>
> Signed-off-by: Camelia Groza <camelia.groza@nxp.com>
> ---
> Changes in v2:
> - warn only once if extracting the timestamp from a received frame fails
> 
> Changes in v3:
> - drop received S/G frames when XDP is enabled
> 
>  drivers/net/ethernet/freescale/dpaa/dpaa_eth.c | 158 ++++++++++++++++++++++---
>  drivers/net/ethernet/freescale/dpaa/dpaa_eth.h |   2 +
>  2 files changed, 144 insertions(+), 16 deletions(-)
> 
> diff --git a/drivers/net/ethernet/freescale/dpaa/dpaa_eth.c b/drivers/net/ethernet/freescale/dpaa/dpaa_eth.c
> index 88533a2..102023c 100644
> --- a/drivers/net/ethernet/freescale/dpaa/dpaa_eth.c
> +++ b/drivers/net/ethernet/freescale/dpaa/dpaa_eth.c
> @@ -53,6 +53,8 @@
>  #include <linux/dma-mapping.h>
>  #include <linux/sort.h>
>  #include <linux/phy_fixed.h>
> +#include <linux/bpf.h>
> +#include <linux/bpf_trace.h>
>  #include <soc/fsl/bman.h>
>  #include <soc/fsl/qman.h>
>  #include "fman.h"
> @@ -177,7 +179,7 @@
>  #define DPAA_HWA_SIZE (DPAA_PARSE_RESULTS_SIZE + DPAA_TIME_STAMP_SIZE \
>  		       + DPAA_HASH_RESULTS_SIZE)
>  #define DPAA_RX_PRIV_DATA_DEFAULT_SIZE (DPAA_TX_PRIV_DATA_SIZE + \
> -					dpaa_rx_extra_headroom)
> +					XDP_PACKET_HEADROOM - DPAA_HWA_SIZE)
>  #ifdef CONFIG_DPAA_ERRATUM_A050385
>  #define DPAA_RX_PRIV_DATA_A050385_SIZE (DPAA_A050385_ALIGN - DPAA_HWA_SIZE)
>  #define DPAA_RX_PRIV_DATA_SIZE (fman_has_errata_a050385() ? \
> @@ -1733,7 +1735,6 @@ static struct sk_buff *contig_fd_to_skb(const struct dpaa_priv *priv,
>  			SKB_DATA_ALIGN(sizeof(struct skb_shared_info)));
>  	if (WARN_ONCE(!skb, "Build skb failure on Rx\n"))
>  		goto free_buffer;
> -	WARN_ON(fd_off != priv->rx_headroom);
>  	skb_reserve(skb, fd_off);
>  	skb_put(skb, qm_fd_get_length(fd));
> 
> @@ -2349,12 +2350,62 @@ static enum qman_cb_dqrr_result rx_error_dqrr(struct qman_portal *portal,
>  	return qman_cb_dqrr_consume;
>  }
> 
> +static u32 dpaa_run_xdp(struct dpaa_priv *priv, struct qm_fd *fd, void *vaddr,
> +			unsigned int *xdp_meta_len)
> +{
> +	ssize_t fd_off = qm_fd_get_offset(fd);
> +	struct bpf_prog *xdp_prog;
> +	struct xdp_buff xdp;
> +	u32 xdp_act;
> +
> +	rcu_read_lock();
> +
> +	xdp_prog = READ_ONCE(priv->xdp_prog);
> +	if (!xdp_prog) {
> +		rcu_read_unlock();
> +		return XDP_PASS;
> +	}
> +
> +	xdp.data = vaddr + fd_off;

I feel like a little drawing of xdp_buff layout would help me with
understanding what is going on over here :)

> +	xdp.data_meta = xdp.data;
> +	xdp.data_hard_start = xdp.data - XDP_PACKET_HEADROOM;
> +	xdp.data_end = xdp.data + qm_fd_get_length(fd);
> +	xdp.frame_sz = DPAA_BP_RAW_SIZE;

Maybe you could fill xdp_buff outside of this function so that later on
you could set xdp.rxq once per napi?

> +
> +	xdp_act = bpf_prog_run_xdp(xdp_prog, &xdp);
> +
> +	/* Update the length and the offset of the FD */
> +	qm_fd_set_contig(fd, xdp.data - vaddr, xdp.data_end - xdp.data);
> +
> +	switch (xdp_act) {
> +	case XDP_PASS:
> +		*xdp_meta_len = xdp.data - xdp.data_meta;
> +		break;
> +	default:
> +		bpf_warn_invalid_xdp_action(xdp_act);
> +		fallthrough;
> +	case XDP_ABORTED:
> +		trace_xdp_exception(priv->net_dev, xdp_prog, xdp_act);
> +		fallthrough;
> +	case XDP_DROP:
> +		/* Free the buffer */
> +		free_pages((unsigned long)vaddr, 0);
> +		break;
> +	}
> +
> +	rcu_read_unlock();
> +
> +	return xdp_act;
> +}
> +
>  static enum qman_cb_dqrr_result rx_default_dqrr(struct qman_portal *portal,
>  						struct qman_fq *fq,
>  						const struct qm_dqrr_entry *dq,
>  						bool sched_napi)
>  {
> +	bool ts_valid = false, hash_valid = false;
>  	struct skb_shared_hwtstamps *shhwtstamps;
> +	unsigned int skb_len, xdp_meta_len = 0;
>  	struct rtnl_link_stats64 *percpu_stats;
>  	struct dpaa_percpu_priv *percpu_priv;
>  	const struct qm_fd *fd = &dq->fd;
> @@ -2362,12 +2413,14 @@ static enum qman_cb_dqrr_result rx_default_dqrr(struct qman_portal *portal,
>  	enum qm_fd_format fd_format;
>  	struct net_device *net_dev;
>  	u32 fd_status, hash_offset;
> +	struct qm_sg_entry *sgt;
>  	struct dpaa_bp *dpaa_bp;
>  	struct dpaa_priv *priv;
> -	unsigned int skb_len;
>  	struct sk_buff *skb;
>  	int *count_ptr;
> +	u32 xdp_act;
>  	void *vaddr;
> +	u32 hash;
>  	u64 ns;
> 
>  	fd_status = be32_to_cpu(fd->status);
> @@ -2423,35 +2476,67 @@ static enum qman_cb_dqrr_result rx_default_dqrr(struct qman_portal *portal,
>  	count_ptr = this_cpu_ptr(dpaa_bp->percpu_count);
>  	(*count_ptr)--;
> 
> -	if (likely(fd_format == qm_fd_contig))
> +	/* Extract the timestamp stored in the headroom before running XDP */
> +	if (priv->rx_tstamp) {
> +		if (!fman_port_get_tstamp(priv->mac_dev->port[RX], vaddr, &ns))
> +			ts_valid = true;
> +		else
> +			WARN_ONCE(1, "fman_port_get_tstamp failed!\n");
> +	}
> +
> +	/* Extract the hash stored in the headroom before running XDP */
> +	if (net_dev->features & NETIF_F_RXHASH && priv->keygen_in_use &&
> +	    !fman_port_get_hash_result_offset(priv->mac_dev->port[RX],
> +					      &hash_offset)) {
> +		hash = be32_to_cpu(*(u32 *)(vaddr + hash_offset));
> +		hash_valid = true;
> +	}
> +
> +	if (likely(fd_format == qm_fd_contig)) {
> +		xdp_act = dpaa_run_xdp(priv, (struct qm_fd *)fd, vaddr,
> +				       &xdp_meta_len);
> +		if (xdp_act != XDP_PASS) {
> +			percpu_stats->rx_packets++;
> +			percpu_stats->rx_bytes += qm_fd_get_length(fd);
> +			return qman_cb_dqrr_consume;
> +		}
>  		skb = contig_fd_to_skb(priv, fd);
> -	else
> +	} else {
> +		/* XDP doesn't support S/G frames. Return the fragments to the
> +		 * buffer pool and release the SGT.
> +		 */
> +		if (READ_ONCE(priv->xdp_prog)) {
> +			WARN_ONCE(1, "S/G frames not supported under XDP\n");
> +			sgt = vaddr + qm_fd_get_offset(fd);
> +			dpaa_release_sgt_members(sgt);
> +			free_pages((unsigned long)vaddr, 0);
> +			return qman_cb_dqrr_consume;
> +		}
>  		skb = sg_fd_to_skb(priv, fd);
> +	}
>  	if (!skb)
>  		return qman_cb_dqrr_consume;
> 
> -	if (priv->rx_tstamp) {
> +	if (xdp_meta_len)
> +		skb_metadata_set(skb, xdp_meta_len);

This is working on a single buffer, right? So there's no need to clear
xdp_meta_len?

> +
> +	/* Set the previously extracted timestamp */
> +	if (ts_valid) {
>  		shhwtstamps = skb_hwtstamps(skb);
>  		memset(shhwtstamps, 0, sizeof(*shhwtstamps));
> -
> -		if (!fman_port_get_tstamp(priv->mac_dev->port[RX], vaddr, &ns))
> -			shhwtstamps->hwtstamp = ns_to_ktime(ns);
> -		else
> -			dev_warn(net_dev->dev.parent, "fman_port_get_tstamp failed!\n");
> +		shhwtstamps->hwtstamp = ns_to_ktime(ns);
>  	}
> 
>  	skb->protocol = eth_type_trans(skb, net_dev);
> 
> -	if (net_dev->features & NETIF_F_RXHASH && priv->keygen_in_use &&
> -	    !fman_port_get_hash_result_offset(priv->mac_dev->port[RX],
> -					      &hash_offset)) {
> +	/* Set the previously extracted hash */
> +	if (hash_valid) {
>  		enum pkt_hash_types type;
> 
>  		/* if L4 exists, it was used in the hash generation */
>  		type = be32_to_cpu(fd->status) & FM_FD_STAT_L4CV ?
>  			PKT_HASH_TYPE_L4 : PKT_HASH_TYPE_L3;
> -		skb_set_hash(skb, be32_to_cpu(*(u32 *)(vaddr + hash_offset)),
> -			     type);
> +		skb_set_hash(skb, hash, type);
>  	}
> 
>  	skb_len = skb->len;
> @@ -2671,6 +2756,46 @@ static int dpaa_eth_stop(struct net_device *net_dev)
>  	return err;
>  }
> 
> +static int dpaa_setup_xdp(struct net_device *net_dev, struct bpf_prog *prog)
> +{
> +	struct dpaa_priv *priv = netdev_priv(net_dev);
> +	struct bpf_prog *old_prog;
> +	bool up;
> +	int err;
> +
> +	/* S/G fragments are not supported in XDP-mode */
> +	if (prog && (priv->dpaa_bp->raw_size <
> +	    net_dev->mtu + VLAN_ETH_HLEN + ETH_FCS_LEN))
> +		return -EINVAL;

Don't you want to include extack message here?

> +
> +	up = netif_running(net_dev);
> +
> +	if (up)
> +		dpaa_eth_stop(net_dev);
> +
> +	old_prog = xchg(&priv->xdp_prog, prog);
> +	if (old_prog)
> +		bpf_prog_put(old_prog);
> +
> +	if (up) {
> +		err = dpaa_open(net_dev);
> +		if (err)
> +			return err;

Out of curiosity, what should user do for that case? Would he be aware of
the weird state of interface?

> +	}
> +
> +	return 0;
> +}
> +
> +static int dpaa_xdp(struct net_device *net_dev, struct netdev_bpf *xdp)
> +{
> +	switch (xdp->command) {
> +	case XDP_SETUP_PROG:
> +		return dpaa_setup_xdp(net_dev, xdp->prog);
> +	default:
> +		return -EINVAL;
> +	}
> +}
> +
>  static int dpaa_ts_ioctl(struct net_device *dev, struct ifreq *rq, int cmd)
>  {
>  	struct dpaa_priv *priv = netdev_priv(dev);
> @@ -2737,6 +2862,7 @@ static int dpaa_ioctl(struct net_device *net_dev, struct ifreq *rq, int cmd)
>  	.ndo_set_rx_mode = dpaa_set_rx_mode,
>  	.ndo_do_ioctl = dpaa_ioctl,
>  	.ndo_setup_tc = dpaa_setup_tc,
> +	.ndo_bpf = dpaa_xdp,
>  };
> 
>  static int dpaa_napi_add(struct net_device *net_dev)
> diff --git a/drivers/net/ethernet/freescale/dpaa/dpaa_eth.h b/drivers/net/ethernet/freescale/dpaa/dpaa_eth.h
> index da30e5d..94e8613 100644
> --- a/drivers/net/ethernet/freescale/dpaa/dpaa_eth.h
> +++ b/drivers/net/ethernet/freescale/dpaa/dpaa_eth.h
> @@ -196,6 +196,8 @@ struct dpaa_priv {
> 
>  	bool tx_tstamp; /* Tx timestamping enabled */
>  	bool rx_tstamp; /* Rx timestamping enabled */
> +
> +	struct bpf_prog *xdp_prog;
>  };
> 
>  /* from dpaa_ethtool.c */
> --
> 1.9.1
> 

RE: [PATCH net-next v3 2/7] dpaa_eth: add basic XDP support

[Camelia Alexandra Groza]

> -----Original Message-----
> From: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
> Sent: Friday, November 20, 2020 02:19
> To: Camelia Alexandra Groza <camelia.groza@nxp.com>
> Cc: kuba@kernel.org; brouer@redhat.com; saeed@kernel.org;
> davem@davemloft.net; Madalin Bucur (OSS)
> <madalin.bucur@oss.nxp.com>; Ioana Ciornei <ioana.ciornei@nxp.com>;
> netdev@vger.kernel.org
> Subject: Re: [PATCH net-next v3 2/7] dpaa_eth: add basic XDP support
> 
> On Thu, Nov 19, 2020 at 06:29:31PM +0200, Camelia Groza wrote:
> > Implement the XDP_DROP and XDP_PASS actions.
> >
> > Avoid draining and reconfiguring the buffer pool at each XDP
> > setup/teardown by increasing the frame headroom and reserving
> > XDP_PACKET_HEADROOM bytes from the start. Since we always reserve
> an
> > entire page per buffer, this change only impacts Jumbo frame scenarios
> > where the maximum linear frame size is reduced by 256 bytes. Multi
> > buffer Scatter/Gather frames are now used instead in these scenarios.
> >
> > Allow XDP programs to access the entire buffer.
> >
> > The data in the received frame's headroom can be overwritten by the XDP
> > program. Extract the relevant fields from the headroom while they are
> > still available, before running the XDP program.
> >
> > Since the headroom might be resized before the frame is passed up to the
> > stack, remove the check for a fixed headroom value when building an skb.
> >
> > Allow the meta data to be updated and pass the information up the stack.
> >
> > Scatter/Gather frames are dropped when XDP is enabled.
> >
> > Acked-by: Madalin Bucur <madalin.bucur@oss.nxp.com>
> > Signed-off-by: Camelia Groza <camelia.groza@nxp.com>
> > ---
> > Changes in v2:
> > - warn only once if extracting the timestamp from a received frame fails
> >
> > Changes in v3:
> > - drop received S/G frames when XDP is enabled
> >
> >  drivers/net/ethernet/freescale/dpaa/dpaa_eth.c | 158
> ++++++++++++++++++++++---
> >  drivers/net/ethernet/freescale/dpaa/dpaa_eth.h |   2 +
> >  2 files changed, 144 insertions(+), 16 deletions(-)
> >
> > diff --git a/drivers/net/ethernet/freescale/dpaa/dpaa_eth.c
> b/drivers/net/ethernet/freescale/dpaa/dpaa_eth.c
> > index 88533a2..102023c 100644
> > --- a/drivers/net/ethernet/freescale/dpaa/dpaa_eth.c
> > +++ b/drivers/net/ethernet/freescale/dpaa/dpaa_eth.c
> > @@ -53,6 +53,8 @@
> >  #include <linux/dma-mapping.h>
> >  #include <linux/sort.h>
> >  #include <linux/phy_fixed.h>
> > +#include <linux/bpf.h>
> > +#include <linux/bpf_trace.h>
> >  #include <soc/fsl/bman.h>
> >  #include <soc/fsl/qman.h>
> >  #include "fman.h"
> > @@ -177,7 +179,7 @@
> >  #define DPAA_HWA_SIZE (DPAA_PARSE_RESULTS_SIZE +
> DPAA_TIME_STAMP_SIZE \
> >  		       + DPAA_HASH_RESULTS_SIZE)
> >  #define DPAA_RX_PRIV_DATA_DEFAULT_SIZE
> (DPAA_TX_PRIV_DATA_SIZE + \
> > -					dpaa_rx_extra_headroom)
> > +					XDP_PACKET_HEADROOM -
> DPAA_HWA_SIZE)
> >  #ifdef CONFIG_DPAA_ERRATUM_A050385
> >  #define DPAA_RX_PRIV_DATA_A050385_SIZE (DPAA_A050385_ALIGN -
> DPAA_HWA_SIZE)
> >  #define DPAA_RX_PRIV_DATA_SIZE (fman_has_errata_a050385() ? \
> > @@ -1733,7 +1735,6 @@ static struct sk_buff *contig_fd_to_skb(const
> struct dpaa_priv *priv,
> >  			SKB_DATA_ALIGN(sizeof(struct skb_shared_info)));
> >  	if (WARN_ONCE(!skb, "Build skb failure on Rx\n"))
> >  		goto free_buffer;
> > -	WARN_ON(fd_off != priv->rx_headroom);
> >  	skb_reserve(skb, fd_off);
> >  	skb_put(skb, qm_fd_get_length(fd));
> >
> > @@ -2349,12 +2350,62 @@ static enum qman_cb_dqrr_result
> rx_error_dqrr(struct qman_portal *portal,
> >  	return qman_cb_dqrr_consume;
> >  }
> >
> > +static u32 dpaa_run_xdp(struct dpaa_priv *priv, struct qm_fd *fd, void
> *vaddr,
> > +			unsigned int *xdp_meta_len)
> > +{
> > +	ssize_t fd_off = qm_fd_get_offset(fd);
> > +	struct bpf_prog *xdp_prog;
> > +	struct xdp_buff xdp;
> > +	u32 xdp_act;
> > +
> > +	rcu_read_lock();
> > +
> > +	xdp_prog = READ_ONCE(priv->xdp_prog);
> > +	if (!xdp_prog) {
> > +		rcu_read_unlock();
> > +		return XDP_PASS;
> > +	}
> > +
> > +	xdp.data = vaddr + fd_off;
> 
> I feel like a little drawing of xdp_buff layout would help me with
> understanding what is going on over here :)

A region at the start of the buffer is reserved for storing hardware annotations and room for the XDP_PACKET_HEADROOM, before the actual data starts. So vaddr points to the start of the buffer, while fd offset provides the offset of the data inside the buffer. I don't feel that we are filling the xdp_buff in a majorly different way from other drivers, so please mention what is unclear here and I can provide more details.

> > +	xdp.data_meta = xdp.data;
> > +	xdp.data_hard_start = xdp.data - XDP_PACKET_HEADROOM;
> > +	xdp.data_end = xdp.data + qm_fd_get_length(fd);
> > +	xdp.frame_sz = DPAA_BP_RAW_SIZE;
> 
> Maybe you could fill xdp_buff outside of this function so that later on
> you could set xdp.rxq once per napi?

I admit I haven't looked into exactly how much performance we would gain from this, but I don't think it would be enough to justify the code churn. We don't have a clean loop for processing the received frames like I see the Intel and ENA drivers have.

> > +
> > +	xdp_act = bpf_prog_run_xdp(xdp_prog, &xdp);
> > +
> > +	/* Update the length and the offset of the FD */
> > +	qm_fd_set_contig(fd, xdp.data - vaddr, xdp.data_end - xdp.data);
> > +
> > +	switch (xdp_act) {
> > +	case XDP_PASS:
> > +		*xdp_meta_len = xdp.data - xdp.data_meta;
> > +		break;
> > +	default:
> > +		bpf_warn_invalid_xdp_action(xdp_act);
> > +		fallthrough;
> > +	case XDP_ABORTED:
> > +		trace_xdp_exception(priv->net_dev, xdp_prog, xdp_act);
> > +		fallthrough;
> > +	case XDP_DROP:
> > +		/* Free the buffer */
> > +		free_pages((unsigned long)vaddr, 0);
> > +		break;
> > +	}
> > +
> > +	rcu_read_unlock();
> > +
> > +	return xdp_act;
> > +}
> > +
> >  static enum qman_cb_dqrr_result rx_default_dqrr(struct qman_portal
> *portal,
> >  						struct qman_fq *fq,
> >  						const struct qm_dqrr_entry
> *dq,
> >  						bool sched_napi)
> >  {
> > +	bool ts_valid = false, hash_valid = false;
> >  	struct skb_shared_hwtstamps *shhwtstamps;
> > +	unsigned int skb_len, xdp_meta_len = 0;
> >  	struct rtnl_link_stats64 *percpu_stats;
> >  	struct dpaa_percpu_priv *percpu_priv;
> >  	const struct qm_fd *fd = &dq->fd;
> > @@ -2362,12 +2413,14 @@ static enum qman_cb_dqrr_result
> rx_default_dqrr(struct qman_portal *portal,
> >  	enum qm_fd_format fd_format;
> >  	struct net_device *net_dev;
> >  	u32 fd_status, hash_offset;
> > +	struct qm_sg_entry *sgt;
> >  	struct dpaa_bp *dpaa_bp;
> >  	struct dpaa_priv *priv;
> > -	unsigned int skb_len;
> >  	struct sk_buff *skb;
> >  	int *count_ptr;
> > +	u32 xdp_act;
> >  	void *vaddr;
> > +	u32 hash;
> >  	u64 ns;
> >
> >  	fd_status = be32_to_cpu(fd->status);
> > @@ -2423,35 +2476,67 @@ static enum qman_cb_dqrr_result
> rx_default_dqrr(struct qman_portal *portal,
> >  	count_ptr = this_cpu_ptr(dpaa_bp->percpu_count);
> >  	(*count_ptr)--;
> >
> > -	if (likely(fd_format == qm_fd_contig))
> > +	/* Extract the timestamp stored in the headroom before running
> XDP */
> > +	if (priv->rx_tstamp) {
> > +		if (!fman_port_get_tstamp(priv->mac_dev->port[RX], vaddr,
> &ns))
> > +			ts_valid = true;
> > +		else
> > +			WARN_ONCE(1, "fman_port_get_tstamp failed!\n");
> > +	}
> > +
> > +	/* Extract the hash stored in the headroom before running XDP */
> > +	if (net_dev->features & NETIF_F_RXHASH && priv->keygen_in_use
> &&
> > +	    !fman_port_get_hash_result_offset(priv->mac_dev->port[RX],
> > +					      &hash_offset)) {
> > +		hash = be32_to_cpu(*(u32 *)(vaddr + hash_offset));
> > +		hash_valid = true;
> > +	}
> > +
> > +	if (likely(fd_format == qm_fd_contig)) {
> > +		xdp_act = dpaa_run_xdp(priv, (struct qm_fd *)fd, vaddr,
> > +				       &xdp_meta_len);
> > +		if (xdp_act != XDP_PASS) {
> > +			percpu_stats->rx_packets++;
> > +			percpu_stats->rx_bytes += qm_fd_get_length(fd);
> > +			return qman_cb_dqrr_consume;
> > +		}
> >  		skb = contig_fd_to_skb(priv, fd);
> > -	else
> > +	} else {
> > +		/* XDP doesn't support S/G frames. Return the fragments to
> the
> > +		 * buffer pool and release the SGT.
> > +		 */
> > +		if (READ_ONCE(priv->xdp_prog)) {
> > +			WARN_ONCE(1, "S/G frames not supported under
> XDP\n");
> > +			sgt = vaddr + qm_fd_get_offset(fd);
> > +			dpaa_release_sgt_members(sgt);
> > +			free_pages((unsigned long)vaddr, 0);
> > +			return qman_cb_dqrr_consume;
> > +		}
> >  		skb = sg_fd_to_skb(priv, fd);
> > +	}
> >  	if (!skb)
> >  		return qman_cb_dqrr_consume;
> >
> > -	if (priv->rx_tstamp) {
> > +	if (xdp_meta_len)
> > +		skb_metadata_set(skb, xdp_meta_len);
> 
> This is working on a single buffer, right? So there's no need to clear
> xdp_meta_len?

I don't think I understand what you mean. Are you saying I shouldn't be initializing xdp_meta_len to 0? This receive path is used when XDP is disabled as well, in which case we don't propagate the metadata.

> > +
> > +	/* Set the previously extracted timestamp */
> > +	if (ts_valid) {
> >  		shhwtstamps = skb_hwtstamps(skb);
> >  		memset(shhwtstamps, 0, sizeof(*shhwtstamps));
> > -
> > -		if (!fman_port_get_tstamp(priv->mac_dev->port[RX], vaddr,
> &ns))
> > -			shhwtstamps->hwtstamp = ns_to_ktime(ns);
> > -		else
> > -			dev_warn(net_dev->dev.parent,
> "fman_port_get_tstamp failed!\n");
> > +		shhwtstamps->hwtstamp = ns_to_ktime(ns);
> >  	}
> >
> >  	skb->protocol = eth_type_trans(skb, net_dev);
> >
> > -	if (net_dev->features & NETIF_F_RXHASH && priv->keygen_in_use
> &&
> > -	    !fman_port_get_hash_result_offset(priv->mac_dev->port[RX],
> > -					      &hash_offset)) {
> > +	/* Set the previously extracted hash */
> > +	if (hash_valid) {
> >  		enum pkt_hash_types type;
> >
> >  		/* if L4 exists, it was used in the hash generation */
> >  		type = be32_to_cpu(fd->status) & FM_FD_STAT_L4CV ?
> >  			PKT_HASH_TYPE_L4 : PKT_HASH_TYPE_L3;
> > -		skb_set_hash(skb, be32_to_cpu(*(u32 *)(vaddr +
> hash_offset)),
> > -			     type);
> > +		skb_set_hash(skb, hash, type);
> >  	}
> >
> >  	skb_len = skb->len;
> > @@ -2671,6 +2756,46 @@ static int dpaa_eth_stop(struct net_device
> *net_dev)
> >  	return err;
> >  }
> >
> > +static int dpaa_setup_xdp(struct net_device *net_dev, struct bpf_prog
> *prog)
> > +{
> > +	struct dpaa_priv *priv = netdev_priv(net_dev);
> > +	struct bpf_prog *old_prog;
> > +	bool up;
> > +	int err;
> > +
> > +	/* S/G fragments are not supported in XDP-mode */
> > +	if (prog && (priv->dpaa_bp->raw_size <
> > +	    net_dev->mtu + VLAN_ETH_HLEN + ETH_FCS_LEN))
> > +		return -EINVAL;
> 
> Don't you want to include extack message here?

This code is moved in the third patch and a dev_warn() is added. I should have added the warning from the start. I'll update it in the next version.

> > +
> > +	up = netif_running(net_dev);
> > +
> > +	if (up)
> > +		dpaa_eth_stop(net_dev);
> > +
> > +	old_prog = xchg(&priv->xdp_prog, prog);
> > +	if (old_prog)
> > +		bpf_prog_put(old_prog);
> > +
> > +	if (up) {
> > +		err = dpaa_open(net_dev);
> > +		if (err)
> > +			return err;
> 
> Out of curiosity, what should user do for that case? Would he be aware of
> the weird state of interface?

This is an improbable state to reach. An error message would be useful. I'll add it in the next version. Thanks.

> > +	}
> > +
> > +	return 0;
> > +}
> > +
> > +static int dpaa_xdp(struct net_device *net_dev, struct netdev_bpf *xdp)
> > +{
> > +	switch (xdp->command) {
> > +	case XDP_SETUP_PROG:
> > +		return dpaa_setup_xdp(net_dev, xdp->prog);
> > +	default:
> > +		return -EINVAL;
> > +	}
> > +}
> > +
> >  static int dpaa_ts_ioctl(struct net_device *dev, struct ifreq *rq, int cmd)
> >  {
> >  	struct dpaa_priv *priv = netdev_priv(dev);
> > @@ -2737,6 +2862,7 @@ static int dpaa_ioctl(struct net_device *net_dev,
> struct ifreq *rq, int cmd)
> >  	.ndo_set_rx_mode = dpaa_set_rx_mode,
> >  	.ndo_do_ioctl = dpaa_ioctl,
> >  	.ndo_setup_tc = dpaa_setup_tc,
> > +	.ndo_bpf = dpaa_xdp,
> >  };
> >
> >  static int dpaa_napi_add(struct net_device *net_dev)
> > diff --git a/drivers/net/ethernet/freescale/dpaa/dpaa_eth.h
> b/drivers/net/ethernet/freescale/dpaa/dpaa_eth.h
> > index da30e5d..94e8613 100644
> > --- a/drivers/net/ethernet/freescale/dpaa/dpaa_eth.h
> > +++ b/drivers/net/ethernet/freescale/dpaa/dpaa_eth.h
> > @@ -196,6 +196,8 @@ struct dpaa_priv {
> >
> >  	bool tx_tstamp; /* Tx timestamping enabled */
> >  	bool rx_tstamp; /* Rx timestamping enabled */
> > +
> > +	struct bpf_prog *xdp_prog;
> >  };
> >
> >  /* from dpaa_ethtool.c */
> > --
> > 1.9.1
> >

Re: [PATCH net-next v3 2/7] dpaa_eth: add basic XDP support

[Maciej Fijalkowski]

On Fri, Nov 20, 2020 at 06:50:28PM +0000, Camelia Alexandra Groza wrote:
> > -----Original Message-----
> > From: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
> > Sent: Friday, November 20, 2020 02:19
> > To: Camelia Alexandra Groza <camelia.groza@nxp.com>
> > Cc: kuba@kernel.org; brouer@redhat.com; saeed@kernel.org;
> > davem@davemloft.net; Madalin Bucur (OSS)
> > <madalin.bucur@oss.nxp.com>; Ioana Ciornei <ioana.ciornei@nxp.com>;
> > netdev@vger.kernel.org
> > Subject: Re: [PATCH net-next v3 2/7] dpaa_eth: add basic XDP support
> > 
> > On Thu, Nov 19, 2020 at 06:29:31PM +0200, Camelia Groza wrote:
> > > Implement the XDP_DROP and XDP_PASS actions.
> > >
> > > Avoid draining and reconfiguring the buffer pool at each XDP
> > > setup/teardown by increasing the frame headroom and reserving
> > > XDP_PACKET_HEADROOM bytes from the start. Since we always reserve
> > an
> > > entire page per buffer, this change only impacts Jumbo frame scenarios
> > > where the maximum linear frame size is reduced by 256 bytes. Multi
> > > buffer Scatter/Gather frames are now used instead in these scenarios.
> > >
> > > Allow XDP programs to access the entire buffer.
> > >
> > > The data in the received frame's headroom can be overwritten by the XDP
> > > program. Extract the relevant fields from the headroom while they are
> > > still available, before running the XDP program.
> > >
> > > Since the headroom might be resized before the frame is passed up to the
> > > stack, remove the check for a fixed headroom value when building an skb.
> > >
> > > Allow the meta data to be updated and pass the information up the stack.
> > >
> > > Scatter/Gather frames are dropped when XDP is enabled.
> > >
> > > Acked-by: Madalin Bucur <madalin.bucur@oss.nxp.com>
> > > Signed-off-by: Camelia Groza <camelia.groza@nxp.com>
> > > ---
> > > Changes in v2:
> > > - warn only once if extracting the timestamp from a received frame fails
> > >
> > > Changes in v3:
> > > - drop received S/G frames when XDP is enabled
> > >
> > >  drivers/net/ethernet/freescale/dpaa/dpaa_eth.c | 158
> > ++++++++++++++++++++++---
> > >  drivers/net/ethernet/freescale/dpaa/dpaa_eth.h |   2 +
> > >  2 files changed, 144 insertions(+), 16 deletions(-)
> > >
> > > diff --git a/drivers/net/ethernet/freescale/dpaa/dpaa_eth.c
> > b/drivers/net/ethernet/freescale/dpaa/dpaa_eth.c
> > > index 88533a2..102023c 100644
> > > --- a/drivers/net/ethernet/freescale/dpaa/dpaa_eth.c
> > > +++ b/drivers/net/ethernet/freescale/dpaa/dpaa_eth.c
> > > @@ -53,6 +53,8 @@
> > >  #include <linux/dma-mapping.h>
> > >  #include <linux/sort.h>
> > >  #include <linux/phy_fixed.h>
> > > +#include <linux/bpf.h>
> > > +#include <linux/bpf_trace.h>
> > >  #include <soc/fsl/bman.h>
> > >  #include <soc/fsl/qman.h>
> > >  #include "fman.h"
> > > @@ -177,7 +179,7 @@
> > >  #define DPAA_HWA_SIZE (DPAA_PARSE_RESULTS_SIZE +
> > DPAA_TIME_STAMP_SIZE \
> > >  		       + DPAA_HASH_RESULTS_SIZE)
> > >  #define DPAA_RX_PRIV_DATA_DEFAULT_SIZE
> > (DPAA_TX_PRIV_DATA_SIZE + \
> > > -					dpaa_rx_extra_headroom)
> > > +					XDP_PACKET_HEADROOM -
> > DPAA_HWA_SIZE)
> > >  #ifdef CONFIG_DPAA_ERRATUM_A050385
> > >  #define DPAA_RX_PRIV_DATA_A050385_SIZE (DPAA_A050385_ALIGN -
> > DPAA_HWA_SIZE)
> > >  #define DPAA_RX_PRIV_DATA_SIZE (fman_has_errata_a050385() ? \
> > > @@ -1733,7 +1735,6 @@ static struct sk_buff *contig_fd_to_skb(const
> > struct dpaa_priv *priv,
> > >  			SKB_DATA_ALIGN(sizeof(struct skb_shared_info)));
> > >  	if (WARN_ONCE(!skb, "Build skb failure on Rx\n"))
> > >  		goto free_buffer;
> > > -	WARN_ON(fd_off != priv->rx_headroom);
> > >  	skb_reserve(skb, fd_off);
> > >  	skb_put(skb, qm_fd_get_length(fd));
> > >
> > > @@ -2349,12 +2350,62 @@ static enum qman_cb_dqrr_result
> > rx_error_dqrr(struct qman_portal *portal,
> > >  	return qman_cb_dqrr_consume;
> > >  }
> > >
> > > +static u32 dpaa_run_xdp(struct dpaa_priv *priv, struct qm_fd *fd, void
> > *vaddr,
> > > +			unsigned int *xdp_meta_len)
> > > +{
> > > +	ssize_t fd_off = qm_fd_get_offset(fd);
> > > +	struct bpf_prog *xdp_prog;
> > > +	struct xdp_buff xdp;
> > > +	u32 xdp_act;
> > > +
> > > +	rcu_read_lock();
> > > +
> > > +	xdp_prog = READ_ONCE(priv->xdp_prog);
> > > +	if (!xdp_prog) {
> > > +		rcu_read_unlock();
> > > +		return XDP_PASS;
> > > +	}
> > > +
> > > +	xdp.data = vaddr + fd_off;
> > 
> > I feel like a little drawing of xdp_buff layout would help me with
> > understanding what is going on over here :)
> 
> A region at the start of the buffer is reserved for storing hardware annotations and room for the XDP_PACKET_HEADROOM, before the actual data starts. So vaddr points to the start of the buffer, while fd offset provides the offset of the data inside the buffer. I don't feel that we are filling the xdp_buff in a majorly different way from other drivers, so please mention what is unclear here and I can provide more details.

Okay, so fd_off tells me where the frame starts, from vaddr to vaddr +
fd_off there might be some HW provided data, so you extract it and then
you are free to go with setting the data_hard_start?

> 
> > > +	xdp.data_meta = xdp.data;
> > > +	xdp.data_hard_start = xdp.data - XDP_PACKET_HEADROOM;
> > > +	xdp.data_end = xdp.data + qm_fd_get_length(fd);
> > > +	xdp.frame_sz = DPAA_BP_RAW_SIZE;
> > 
> > Maybe you could fill xdp_buff outside of this function so that later on
> > you could set xdp.rxq once per napi?
> 
> I admit I haven't looked into exactly how much performance we would gain from this, but I don't think it would be enough to justify the code churn. We don't have a clean loop for processing the received frames like I see the Intel and ENA drivers have.
> 
> > > +
> > > +	xdp_act = bpf_prog_run_xdp(xdp_prog, &xdp);
> > > +
> > > +	/* Update the length and the offset of the FD */
> > > +	qm_fd_set_contig(fd, xdp.data - vaddr, xdp.data_end - xdp.data);
> > > +
> > > +	switch (xdp_act) {
> > > +	case XDP_PASS:
> > > +		*xdp_meta_len = xdp.data - xdp.data_meta;
> > > +		break;
> > > +	default:
> > > +		bpf_warn_invalid_xdp_action(xdp_act);
> > > +		fallthrough;
> > > +	case XDP_ABORTED:
> > > +		trace_xdp_exception(priv->net_dev, xdp_prog, xdp_act);
> > > +		fallthrough;
> > > +	case XDP_DROP:
> > > +		/* Free the buffer */
> > > +		free_pages((unsigned long)vaddr, 0);
> > > +		break;
> > > +	}
> > > +
> > > +	rcu_read_unlock();
> > > +
> > > +	return xdp_act;
> > > +}
> > > +
> > >  static enum qman_cb_dqrr_result rx_default_dqrr(struct qman_portal
> > *portal,
> > >  						struct qman_fq *fq,
> > >  						const struct qm_dqrr_entry
> > *dq,
> > >  						bool sched_napi)
> > >  {
> > > +	bool ts_valid = false, hash_valid = false;
> > >  	struct skb_shared_hwtstamps *shhwtstamps;
> > > +	unsigned int skb_len, xdp_meta_len = 0;
> > >  	struct rtnl_link_stats64 *percpu_stats;
> > >  	struct dpaa_percpu_priv *percpu_priv;
> > >  	const struct qm_fd *fd = &dq->fd;
> > > @@ -2362,12 +2413,14 @@ static enum qman_cb_dqrr_result
> > rx_default_dqrr(struct qman_portal *portal,
> > >  	enum qm_fd_format fd_format;
> > >  	struct net_device *net_dev;
> > >  	u32 fd_status, hash_offset;
> > > +	struct qm_sg_entry *sgt;
> > >  	struct dpaa_bp *dpaa_bp;
> > >  	struct dpaa_priv *priv;
> > > -	unsigned int skb_len;
> > >  	struct sk_buff *skb;
> > >  	int *count_ptr;
> > > +	u32 xdp_act;
> > >  	void *vaddr;
> > > +	u32 hash;
> > >  	u64 ns;
> > >
> > >  	fd_status = be32_to_cpu(fd->status);
> > > @@ -2423,35 +2476,67 @@ static enum qman_cb_dqrr_result
> > rx_default_dqrr(struct qman_portal *portal,
> > >  	count_ptr = this_cpu_ptr(dpaa_bp->percpu_count);
> > >  	(*count_ptr)--;
> > >
> > > -	if (likely(fd_format == qm_fd_contig))
> > > +	/* Extract the timestamp stored in the headroom before running
> > XDP */
> > > +	if (priv->rx_tstamp) {
> > > +		if (!fman_port_get_tstamp(priv->mac_dev->port[RX], vaddr,
> > &ns))
> > > +			ts_valid = true;
> > > +		else
> > > +			WARN_ONCE(1, "fman_port_get_tstamp failed!\n");
> > > +	}
> > > +
> > > +	/* Extract the hash stored in the headroom before running XDP */
> > > +	if (net_dev->features & NETIF_F_RXHASH && priv->keygen_in_use
> > &&
> > > +	    !fman_port_get_hash_result_offset(priv->mac_dev->port[RX],
> > > +					      &hash_offset)) {
> > > +		hash = be32_to_cpu(*(u32 *)(vaddr + hash_offset));
> > > +		hash_valid = true;
> > > +	}
> > > +
> > > +	if (likely(fd_format == qm_fd_contig)) {
> > > +		xdp_act = dpaa_run_xdp(priv, (struct qm_fd *)fd, vaddr,
> > > +				       &xdp_meta_len);
> > > +		if (xdp_act != XDP_PASS) {
> > > +			percpu_stats->rx_packets++;
> > > +			percpu_stats->rx_bytes += qm_fd_get_length(fd);
> > > +			return qman_cb_dqrr_consume;
> > > +		}
> > >  		skb = contig_fd_to_skb(priv, fd);
> > > -	else
> > > +	} else {
> > > +		/* XDP doesn't support S/G frames. Return the fragments to
> > the
> > > +		 * buffer pool and release the SGT.
> > > +		 */
> > > +		if (READ_ONCE(priv->xdp_prog)) {
> > > +			WARN_ONCE(1, "S/G frames not supported under
> > XDP\n");
> > > +			sgt = vaddr + qm_fd_get_offset(fd);
> > > +			dpaa_release_sgt_members(sgt);
> > > +			free_pages((unsigned long)vaddr, 0);
> > > +			return qman_cb_dqrr_consume;
> > > +		}
> > >  		skb = sg_fd_to_skb(priv, fd);
> > > +	}
> > >  	if (!skb)
> > >  		return qman_cb_dqrr_consume;
> > >
> > > -	if (priv->rx_tstamp) {
> > > +	if (xdp_meta_len)
> > > +		skb_metadata_set(skb, xdp_meta_len);
> > 
> > This is working on a single buffer, right? So there's no need to clear
> > xdp_meta_len?
> 
> I don't think I understand what you mean. Are you saying I shouldn't be initializing xdp_meta_len to 0? This receive path is used when XDP is disabled as well, in which case we don't propagate the metadata.

What I meant was that if this function would operate on many buffers then
we would have to clear the xdp_meta_len, so that next buffers wouldn't get
the value from previous bufs, but I suppose that this rx_default_dqrr
callback is called once per each buffer.

> 
> > > +
> > > +	/* Set the previously extracted timestamp */
> > > +	if (ts_valid) {
> > >  		shhwtstamps = skb_hwtstamps(skb);
> > >  		memset(shhwtstamps, 0, sizeof(*shhwtstamps));
> > > -
> > > -		if (!fman_port_get_tstamp(priv->mac_dev->port[RX], vaddr,
> > &ns))
> > > -			shhwtstamps->hwtstamp = ns_to_ktime(ns);
> > > -		else
> > > -			dev_warn(net_dev->dev.parent,
> > "fman_port_get_tstamp failed!\n");
> > > +		shhwtstamps->hwtstamp = ns_to_ktime(ns);
> > >  	}
> > >
> > >  	skb->protocol = eth_type_trans(skb, net_dev);
> > >
> > > -	if (net_dev->features & NETIF_F_RXHASH && priv->keygen_in_use
> > &&
> > > -	    !fman_port_get_hash_result_offset(priv->mac_dev->port[RX],
> > > -					      &hash_offset)) {
> > > +	/* Set the previously extracted hash */
> > > +	if (hash_valid) {
> > >  		enum pkt_hash_types type;
> > >
> > >  		/* if L4 exists, it was used in the hash generation */
> > >  		type = be32_to_cpu(fd->status) & FM_FD_STAT_L4CV ?
> > >  			PKT_HASH_TYPE_L4 : PKT_HASH_TYPE_L3;
> > > -		skb_set_hash(skb, be32_to_cpu(*(u32 *)(vaddr +
> > hash_offset)),
> > > -			     type);
> > > +		skb_set_hash(skb, hash, type);
> > >  	}
> > >
> > >  	skb_len = skb->len;
> > > @@ -2671,6 +2756,46 @@ static int dpaa_eth_stop(struct net_device
> > *net_dev)
> > >  	return err;
> > >  }

RE: [PATCH net-next v3 2/7] dpaa_eth: add basic XDP support

[Camelia Alexandra Groza]

> -----Original Message-----
> From: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
> Sent: Tuesday, November 24, 2020 00:08
> To: Camelia Alexandra Groza <camelia.groza@nxp.com>
> Cc: kuba@kernel.org; brouer@redhat.com; saeed@kernel.org;
> davem@davemloft.net; Madalin Bucur (OSS)
> <madalin.bucur@oss.nxp.com>; Ioana Ciornei <ioana.ciornei@nxp.com>;
> netdev@vger.kernel.org
> Subject: Re: [PATCH net-next v3 2/7] dpaa_eth: add basic XDP support
> 
> On Fri, Nov 20, 2020 at 06:50:28PM +0000, Camelia Alexandra Groza wrote:
> > > -----Original Message-----
> > > From: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
> > > Sent: Friday, November 20, 2020 02:19
> > > To: Camelia Alexandra Groza <camelia.groza@nxp.com>
> > > Cc: kuba@kernel.org; brouer@redhat.com; saeed@kernel.org;
> > > davem@davemloft.net; Madalin Bucur (OSS)
> > > <madalin.bucur@oss.nxp.com>; Ioana Ciornei <ioana.ciornei@nxp.com>;
> > > netdev@vger.kernel.org
> > > Subject: Re: [PATCH net-next v3 2/7] dpaa_eth: add basic XDP support
> > >
> > > On Thu, Nov 19, 2020 at 06:29:31PM +0200, Camelia Groza wrote:
> > > > Implement the XDP_DROP and XDP_PASS actions.
> > > >
> > > > Avoid draining and reconfiguring the buffer pool at each XDP
> > > > setup/teardown by increasing the frame headroom and reserving
> > > > XDP_PACKET_HEADROOM bytes from the start. Since we always
> reserve
> > > an
> > > > entire page per buffer, this change only impacts Jumbo frame scenarios
> > > > where the maximum linear frame size is reduced by 256 bytes. Multi
> > > > buffer Scatter/Gather frames are now used instead in these scenarios.
> > > >
> > > > Allow XDP programs to access the entire buffer.
> > > >
> > > > The data in the received frame's headroom can be overwritten by the
> XDP
> > > > program. Extract the relevant fields from the headroom while they are
> > > > still available, before running the XDP program.
> > > >
> > > > Since the headroom might be resized before the frame is passed up to
> the
> > > > stack, remove the check for a fixed headroom value when building an
> skb.
> > > >
> > > > Allow the meta data to be updated and pass the information up the
> stack.
> > > >
> > > > Scatter/Gather frames are dropped when XDP is enabled.
> > > >
> > > > Acked-by: Madalin Bucur <madalin.bucur@oss.nxp.com>
> > > > Signed-off-by: Camelia Groza <camelia.groza@nxp.com>
> > > > ---
> > > > Changes in v2:
> > > > - warn only once if extracting the timestamp from a received frame fails
> > > >
> > > > Changes in v3:
> > > > - drop received S/G frames when XDP is enabled
> > > >
> > > >  drivers/net/ethernet/freescale/dpaa/dpaa_eth.c | 158
> > > ++++++++++++++++++++++---
> > > >  drivers/net/ethernet/freescale/dpaa/dpaa_eth.h |   2 +
> > > >  2 files changed, 144 insertions(+), 16 deletions(-)
> > > >
> > > > diff --git a/drivers/net/ethernet/freescale/dpaa/dpaa_eth.c
> > > b/drivers/net/ethernet/freescale/dpaa/dpaa_eth.c
> > > > index 88533a2..102023c 100644
> > > > --- a/drivers/net/ethernet/freescale/dpaa/dpaa_eth.c
> > > > +++ b/drivers/net/ethernet/freescale/dpaa/dpaa_eth.c
> > > > @@ -53,6 +53,8 @@
> > > >  #include <linux/dma-mapping.h>
> > > >  #include <linux/sort.h>
> > > >  #include <linux/phy_fixed.h>
> > > > +#include <linux/bpf.h>
> > > > +#include <linux/bpf_trace.h>
> > > >  #include <soc/fsl/bman.h>
> > > >  #include <soc/fsl/qman.h>
> > > >  #include "fman.h"
> > > > @@ -177,7 +179,7 @@
> > > >  #define DPAA_HWA_SIZE (DPAA_PARSE_RESULTS_SIZE +
> > > DPAA_TIME_STAMP_SIZE \
> > > >  		       + DPAA_HASH_RESULTS_SIZE)
> > > >  #define DPAA_RX_PRIV_DATA_DEFAULT_SIZE
> > > (DPAA_TX_PRIV_DATA_SIZE + \
> > > > -					dpaa_rx_extra_headroom)
> > > > +					XDP_PACKET_HEADROOM -
> > > DPAA_HWA_SIZE)
> > > >  #ifdef CONFIG_DPAA_ERRATUM_A050385
> > > >  #define DPAA_RX_PRIV_DATA_A050385_SIZE (DPAA_A050385_ALIGN
> -
> > > DPAA_HWA_SIZE)
> > > >  #define DPAA_RX_PRIV_DATA_SIZE (fman_has_errata_a050385() ? \
> > > > @@ -1733,7 +1735,6 @@ static struct sk_buff *contig_fd_to_skb(const
> > > struct dpaa_priv *priv,
> > > >  			SKB_DATA_ALIGN(sizeof(struct skb_shared_info)));
> > > >  	if (WARN_ONCE(!skb, "Build skb failure on Rx\n"))
> > > >  		goto free_buffer;
> > > > -	WARN_ON(fd_off != priv->rx_headroom);
> > > >  	skb_reserve(skb, fd_off);
> > > >  	skb_put(skb, qm_fd_get_length(fd));
> > > >
> > > > @@ -2349,12 +2350,62 @@ static enum qman_cb_dqrr_result
> > > rx_error_dqrr(struct qman_portal *portal,
> > > >  	return qman_cb_dqrr_consume;
> > > >  }
> > > >
> > > > +static u32 dpaa_run_xdp(struct dpaa_priv *priv, struct qm_fd *fd,
> void
> > > *vaddr,
> > > > +			unsigned int *xdp_meta_len)
> > > > +{
> > > > +	ssize_t fd_off = qm_fd_get_offset(fd);
> > > > +	struct bpf_prog *xdp_prog;
> > > > +	struct xdp_buff xdp;
> > > > +	u32 xdp_act;
> > > > +
> > > > +	rcu_read_lock();
> > > > +
> > > > +	xdp_prog = READ_ONCE(priv->xdp_prog);
> > > > +	if (!xdp_prog) {
> > > > +		rcu_read_unlock();
> > > > +		return XDP_PASS;
> > > > +	}
> > > > +
> > > > +	xdp.data = vaddr + fd_off;
> > >
> > > I feel like a little drawing of xdp_buff layout would help me with
> > > understanding what is going on over here :)
> >
> > A region at the start of the buffer is reserved for storing hardware
> annotations and room for the XDP_PACKET_HEADROOM, before the actual
> data starts. So vaddr points to the start of the buffer, while fd offset provides
> the offset of the data inside the buffer. I don't feel that we are filling the
> xdp_buff in a majorly different way from other drivers, so please mention
> what is unclear here and I can provide more details.
> 
> Okay, so fd_off tells me where the frame starts, from vaddr to vaddr +
> fd_off there might be some HW provided data, so you extract it and then
> you are free to go with setting the data_hard_start?

Yes, that's right.

> >
> > > > +	xdp.data_meta = xdp.data;
> > > > +	xdp.data_hard_start = xdp.data - XDP_PACKET_HEADROOM;
> > > > +	xdp.data_end = xdp.data + qm_fd_get_length(fd);
> > > > +	xdp.frame_sz = DPAA_BP_RAW_SIZE;
> > >
> > > Maybe you could fill xdp_buff outside of this function so that later on
> > > you could set xdp.rxq once per napi?
> >
> > I admit I haven't looked into exactly how much performance we would gain
> from this, but I don't think it would be enough to justify the code churn. We
> don't have a clean loop for processing the received frames like I see the Intel
> and ENA drivers have.
> >
> > > > +
> > > > +	xdp_act = bpf_prog_run_xdp(xdp_prog, &xdp);
> > > > +
> > > > +	/* Update the length and the offset of the FD */
> > > > +	qm_fd_set_contig(fd, xdp.data - vaddr, xdp.data_end - xdp.data);
> > > > +
> > > > +	switch (xdp_act) {
> > > > +	case XDP_PASS:
> > > > +		*xdp_meta_len = xdp.data - xdp.data_meta;
> > > > +		break;
> > > > +	default:
> > > > +		bpf_warn_invalid_xdp_action(xdp_act);
> > > > +		fallthrough;
> > > > +	case XDP_ABORTED:
> > > > +		trace_xdp_exception(priv->net_dev, xdp_prog, xdp_act);
> > > > +		fallthrough;
> > > > +	case XDP_DROP:
> > > > +		/* Free the buffer */
> > > > +		free_pages((unsigned long)vaddr, 0);
> > > > +		break;
> > > > +	}
> > > > +
> > > > +	rcu_read_unlock();
> > > > +
> > > > +	return xdp_act;
> > > > +}
> > > > +
> > > >  static enum qman_cb_dqrr_result rx_default_dqrr(struct qman_portal
> > > *portal,
> > > >  						struct qman_fq *fq,
> > > >  						const struct qm_dqrr_entry
> > > *dq,
> > > >  						bool sched_napi)
> > > >  {
> > > > +	bool ts_valid = false, hash_valid = false;
> > > >  	struct skb_shared_hwtstamps *shhwtstamps;
> > > > +	unsigned int skb_len, xdp_meta_len = 0;
> > > >  	struct rtnl_link_stats64 *percpu_stats;
> > > >  	struct dpaa_percpu_priv *percpu_priv;
> > > >  	const struct qm_fd *fd = &dq->fd;
> > > > @@ -2362,12 +2413,14 @@ static enum qman_cb_dqrr_result
> > > rx_default_dqrr(struct qman_portal *portal,
> > > >  	enum qm_fd_format fd_format;
> > > >  	struct net_device *net_dev;
> > > >  	u32 fd_status, hash_offset;
> > > > +	struct qm_sg_entry *sgt;
> > > >  	struct dpaa_bp *dpaa_bp;
> > > >  	struct dpaa_priv *priv;
> > > > -	unsigned int skb_len;
> > > >  	struct sk_buff *skb;
> > > >  	int *count_ptr;
> > > > +	u32 xdp_act;
> > > >  	void *vaddr;
> > > > +	u32 hash;
> > > >  	u64 ns;
> > > >
> > > >  	fd_status = be32_to_cpu(fd->status);
> > > > @@ -2423,35 +2476,67 @@ static enum qman_cb_dqrr_result
> > > rx_default_dqrr(struct qman_portal *portal,
> > > >  	count_ptr = this_cpu_ptr(dpaa_bp->percpu_count);
> > > >  	(*count_ptr)--;
> > > >
> > > > -	if (likely(fd_format == qm_fd_contig))
> > > > +	/* Extract the timestamp stored in the headroom before running
> > > XDP */
> > > > +	if (priv->rx_tstamp) {
> > > > +		if (!fman_port_get_tstamp(priv->mac_dev->port[RX], vaddr,
> > > &ns))
> > > > +			ts_valid = true;
> > > > +		else
> > > > +			WARN_ONCE(1, "fman_port_get_tstamp failed!\n");
> > > > +	}
> > > > +
> > > > +	/* Extract the hash stored in the headroom before running XDP */
> > > > +	if (net_dev->features & NETIF_F_RXHASH && priv->keygen_in_use
> > > &&
> > > > +	    !fman_port_get_hash_result_offset(priv->mac_dev->port[RX],
> > > > +					      &hash_offset)) {
> > > > +		hash = be32_to_cpu(*(u32 *)(vaddr + hash_offset));
> > > > +		hash_valid = true;
> > > > +	}
> > > > +
> > > > +	if (likely(fd_format == qm_fd_contig)) {
> > > > +		xdp_act = dpaa_run_xdp(priv, (struct qm_fd *)fd, vaddr,
> > > > +				       &xdp_meta_len);
> > > > +		if (xdp_act != XDP_PASS) {
> > > > +			percpu_stats->rx_packets++;
> > > > +			percpu_stats->rx_bytes += qm_fd_get_length(fd);
> > > > +			return qman_cb_dqrr_consume;
> > > > +		}
> > > >  		skb = contig_fd_to_skb(priv, fd);
> > > > -	else
> > > > +	} else {
> > > > +		/* XDP doesn't support S/G frames. Return the fragments to
> > > the
> > > > +		 * buffer pool and release the SGT.
> > > > +		 */
> > > > +		if (READ_ONCE(priv->xdp_prog)) {
> > > > +			WARN_ONCE(1, "S/G frames not supported under
> > > XDP\n");
> > > > +			sgt = vaddr + qm_fd_get_offset(fd);
> > > > +			dpaa_release_sgt_members(sgt);
> > > > +			free_pages((unsigned long)vaddr, 0);
> > > > +			return qman_cb_dqrr_consume;
> > > > +		}
> > > >  		skb = sg_fd_to_skb(priv, fd);
> > > > +	}
> > > >  	if (!skb)
> > > >  		return qman_cb_dqrr_consume;
> > > >
> > > > -	if (priv->rx_tstamp) {
> > > > +	if (xdp_meta_len)
> > > > +		skb_metadata_set(skb, xdp_meta_len);
> > >
> > > This is working on a single buffer, right? So there's no need to clear
> > > xdp_meta_len?
> >
> > I don't think I understand what you mean. Are you saying I shouldn't be
> initializing xdp_meta_len to 0? This receive path is used when XDP is disabled
> as well, in which case we don't propagate the metadata.
> 
> What I meant was that if this function would operate on many buffers then
> we would have to clear the xdp_meta_len, so that next buffers wouldn't get
> the value from previous bufs, but I suppose that this rx_default_dqrr
> callback is called once per each buffer.

Yes, rx_default_dqrr is processing only one buffer.

> >
> > > > +
> > > > +	/* Set the previously extracted timestamp */
> > > > +	if (ts_valid) {
> > > >  		shhwtstamps = skb_hwtstamps(skb);
> > > >  		memset(shhwtstamps, 0, sizeof(*shhwtstamps));
> > > > -
> > > > -		if (!fman_port_get_tstamp(priv->mac_dev->port[RX], vaddr,
> > > &ns))
> > > > -			shhwtstamps->hwtstamp = ns_to_ktime(ns);
> > > > -		else
> > > > -			dev_warn(net_dev->dev.parent,
> > > "fman_port_get_tstamp failed!\n");
> > > > +		shhwtstamps->hwtstamp = ns_to_ktime(ns);
> > > >  	}
> > > >
> > > >  	skb->protocol = eth_type_trans(skb, net_dev);
> > > >
> > > > -	if (net_dev->features & NETIF_F_RXHASH && priv->keygen_in_use
> > > &&
> > > > -	    !fman_port_get_hash_result_offset(priv->mac_dev->port[RX],
> > > > -					      &hash_offset)) {
> > > > +	/* Set the previously extracted hash */
> > > > +	if (hash_valid) {
> > > >  		enum pkt_hash_types type;
> > > >
> > > >  		/* if L4 exists, it was used in the hash generation */
> > > >  		type = be32_to_cpu(fd->status) & FM_FD_STAT_L4CV ?
> > > >  			PKT_HASH_TYPE_L4 : PKT_HASH_TYPE_L3;
> > > > -		skb_set_hash(skb, be32_to_cpu(*(u32 *)(vaddr +
> > > hash_offset)),
> > > > -			     type);
> > > > +		skb_set_hash(skb, hash, type);
> > > >  	}
> > > >
> > > >  	skb_len = skb->len;
> > > > @@ -2671,6 +2756,46 @@ static int dpaa_eth_stop(struct net_device
> > > *net_dev)
> > > >  	return err;
> > > >  }

[PATCH net-next v3 3/7] dpaa_eth: limit the possible MTU range when XDP is enabled

[Camelia Groza]

Implement the ndo_change_mtu callback to prevent users from setting an
MTU that would permit processing of S/G frames. The maximum MTU size
is dependent on the buffer size.

Acked-by: Madalin Bucur <madalin.bucur@oss.nxp.com>
Signed-off-by: Camelia Groza <camelia.groza@nxp.com>
---
 drivers/net/ethernet/freescale/dpaa/dpaa_eth.c | 32 ++++++++++++++++++++++++--
 1 file changed, 30 insertions(+), 2 deletions(-)

diff --git a/drivers/net/ethernet/freescale/dpaa/dpaa_eth.c b/drivers/net/ethernet/freescale/dpaa/dpaa_eth.c
index 102023c..242ed45 100644
--- a/drivers/net/ethernet/freescale/dpaa/dpaa_eth.c
+++ b/drivers/net/ethernet/freescale/dpaa/dpaa_eth.c
@@ -2756,6 +2756,34 @@ static int dpaa_eth_stop(struct net_device *net_dev)
 	return err;
 }
 
+static bool xdp_validate_mtu(struct dpaa_priv *priv, int mtu)
+{
+	int max_contig_data = priv->dpaa_bp->size - priv->rx_headroom;
+
+	/* We do not support S/G fragments when XDP is enabled.
+	 * Limit the MTU in relation to the buffer size.
+	 */
+	if (mtu + VLAN_ETH_HLEN + ETH_FCS_LEN > max_contig_data) {
+		dev_warn(priv->net_dev->dev.parent,
+			 "The maximum MTU for XDP is %d\n",
+			 max_contig_data - VLAN_ETH_HLEN - ETH_FCS_LEN);
+		return false;
+	}
+
+	return true;
+}
+
+static int dpaa_change_mtu(struct net_device *net_dev, int new_mtu)
+{
+	struct dpaa_priv *priv = netdev_priv(net_dev);
+
+	if (priv->xdp_prog && !xdp_validate_mtu(priv, new_mtu))
+		return -EINVAL;
+
+	net_dev->mtu = new_mtu;
+	return 0;
+}
+
 static int dpaa_setup_xdp(struct net_device *net_dev, struct bpf_prog *prog)
 {
 	struct dpaa_priv *priv = netdev_priv(net_dev);
@@ -2764,8 +2792,7 @@ static int dpaa_setup_xdp(struct net_device *net_dev, struct bpf_prog *prog)
 	int err;
 
 	/* S/G fragments are not supported in XDP-mode */
-	if (prog && (priv->dpaa_bp->raw_size <
-	    net_dev->mtu + VLAN_ETH_HLEN + ETH_FCS_LEN))
+	if (prog && !xdp_validate_mtu(priv, net_dev->mtu))
 		return -EINVAL;
 
 	up = netif_running(net_dev);
@@ -2862,6 +2889,7 @@ static int dpaa_ioctl(struct net_device *net_dev, struct ifreq *rq, int cmd)
 	.ndo_set_rx_mode = dpaa_set_rx_mode,
 	.ndo_do_ioctl = dpaa_ioctl,
 	.ndo_setup_tc = dpaa_setup_tc,
+	.ndo_change_mtu = dpaa_change_mtu,
 	.ndo_bpf = dpaa_xdp,
 };
 
-- 
1.9.1

[PATCH net-next v3 4/7] dpaa_eth: add XDP_TX support

[Camelia Groza]

Use an xdp_frame structure for managing the frame. Store a backpointer to
the structure at the start of the buffer before enqueueing. Use the XDP
API for freeing the buffer when it returns to the driver on the TX
confirmation path.

This approach will be reused for XDP REDIRECT.

Acked-by: Madalin Bucur <madalin.bucur@oss.nxp.com>
Signed-off-by: Camelia Groza <camelia.groza@nxp.com>
---
 drivers/net/ethernet/freescale/dpaa/dpaa_eth.c | 129 ++++++++++++++++++++++++-
 drivers/net/ethernet/freescale/dpaa/dpaa_eth.h |   2 +
 2 files changed, 126 insertions(+), 5 deletions(-)

diff --git a/drivers/net/ethernet/freescale/dpaa/dpaa_eth.c b/drivers/net/ethernet/freescale/dpaa/dpaa_eth.c
index 242ed45..cd5f4f6 100644
--- a/drivers/net/ethernet/freescale/dpaa/dpaa_eth.c
+++ b/drivers/net/ethernet/freescale/dpaa/dpaa_eth.c
@@ -1130,6 +1130,24 @@ static int dpaa_fq_init(struct dpaa_fq *dpaa_fq, bool td_enable)
 
 	dpaa_fq->fqid = qman_fq_fqid(fq);
 
+	if (dpaa_fq->fq_type == FQ_TYPE_RX_DEFAULT ||
+	    dpaa_fq->fq_type == FQ_TYPE_RX_PCD) {
+		err = xdp_rxq_info_reg(&dpaa_fq->xdp_rxq, dpaa_fq->net_dev,
+				       dpaa_fq->fqid);
+		if (err) {
+			dev_err(dev, "xdp_rxq_info_reg failed\n");
+			return err;
+		}
+
+		err = xdp_rxq_info_reg_mem_model(&dpaa_fq->xdp_rxq,
+						 MEM_TYPE_PAGE_ORDER0, NULL);
+		if (err) {
+			dev_err(dev, "xdp_rxq_info_reg_mem_model failed\n");
+			xdp_rxq_info_unreg(&dpaa_fq->xdp_rxq);
+			return err;
+		}
+	}
+
 	return 0;
 }
 
@@ -1159,6 +1177,10 @@ static int dpaa_fq_free_entry(struct device *dev, struct qman_fq *fq)
 		}
 	}
 
+	if (dpaa_fq->fq_type == FQ_TYPE_RX_DEFAULT ||
+	    dpaa_fq->fq_type == FQ_TYPE_RX_PCD)
+		xdp_rxq_info_unreg(&dpaa_fq->xdp_rxq);
+
 	qman_destroy_fq(fq);
 	list_del(&dpaa_fq->list);
 
@@ -1625,6 +1647,9 @@ static int dpaa_eth_refill_bpools(struct dpaa_priv *priv)
  *
  * Return the skb backpointer, since for S/G frames the buffer containing it
  * gets freed here.
+ *
+ * No skb backpointer is set when transmitting XDP frames. Cleanup the buffer
+ * and return NULL in this case.
  */
 static struct sk_buff *dpaa_cleanup_tx_fd(const struct dpaa_priv *priv,
 					  const struct qm_fd *fd, bool ts)
@@ -1636,6 +1661,7 @@ static struct sk_buff *dpaa_cleanup_tx_fd(const struct dpaa_priv *priv,
 	void *vaddr = phys_to_virt(addr);
 	const struct qm_sg_entry *sgt;
 	struct dpaa_eth_swbp *swbp;
+	struct xdp_frame *xdpf;
 	struct sk_buff *skb;
 	u64 ns;
 	int i;
@@ -1664,13 +1690,22 @@ static struct sk_buff *dpaa_cleanup_tx_fd(const struct dpaa_priv *priv,
 		}
 	} else {
 		dma_unmap_single(priv->tx_dma_dev, addr,
-				 priv->tx_headroom + qm_fd_get_length(fd),
+				 qm_fd_get_offset(fd) + qm_fd_get_length(fd),
 				 dma_dir);
 	}
 
 	swbp = (struct dpaa_eth_swbp *)vaddr;
 	skb = swbp->skb;
 
+	/* No skb backpointer is set when running XDP. An xdp_frame
+	 * backpointer is saved instead.
+	 */
+	if (!skb) {
+		xdpf = swbp->xdpf;
+		xdp_return_frame(xdpf);
+		return NULL;
+	}
+
 	/* DMA unmapping is required before accessing the HW provided info */
 	if (ts && priv->tx_tstamp &&
 	    skb_shinfo(skb)->tx_flags & SKBTX_HW_TSTAMP) {
@@ -2350,11 +2385,76 @@ static enum qman_cb_dqrr_result rx_error_dqrr(struct qman_portal *portal,
 	return qman_cb_dqrr_consume;
 }
 
+static int dpaa_xdp_xmit_frame(struct net_device *net_dev,
+			       struct xdp_frame *xdpf)
+{
+	struct dpaa_priv *priv = netdev_priv(net_dev);
+	struct rtnl_link_stats64 *percpu_stats;
+	struct dpaa_percpu_priv *percpu_priv;
+	struct dpaa_eth_swbp *swbp;
+	struct netdev_queue *txq;
+	void *buff_start;
+	struct qm_fd fd;
+	dma_addr_t addr;
+	int err;
+
+	percpu_priv = this_cpu_ptr(priv->percpu_priv);
+	percpu_stats = &percpu_priv->stats;
+
+	if (xdpf->headroom < DPAA_TX_PRIV_DATA_SIZE) {
+		err = -EINVAL;
+		goto out_error;
+	}
+
+	buff_start = xdpf->data - xdpf->headroom;
+
+	/* Leave empty the skb backpointer at the start of the buffer.
+	 * Save the XDP frame for easy cleanup on confirmation.
+	 */
+	swbp = (struct dpaa_eth_swbp *)buff_start;
+	swbp->skb = NULL;
+	swbp->xdpf = xdpf;
+
+	qm_fd_clear_fd(&fd);
+	fd.bpid = FSL_DPAA_BPID_INV;
+	fd.cmd |= cpu_to_be32(FM_FD_CMD_FCO);
+	qm_fd_set_contig(&fd, xdpf->headroom, xdpf->len);
+
+	addr = dma_map_single(priv->tx_dma_dev, buff_start,
+			      xdpf->headroom + xdpf->len,
+			      DMA_TO_DEVICE);
+	if (unlikely(dma_mapping_error(priv->tx_dma_dev, addr))) {
+		err = -EINVAL;
+		goto out_error;
+	}
+
+	qm_fd_addr_set64(&fd, addr);
+
+	/* Bump the trans_start */
+	txq = netdev_get_tx_queue(net_dev, smp_processor_id());
+	txq->trans_start = jiffies;
+
+	err = dpaa_xmit(priv, percpu_stats, smp_processor_id(), &fd);
+	if (err) {
+		dma_unmap_single(priv->tx_dma_dev, addr,
+				 qm_fd_get_offset(&fd) + qm_fd_get_length(&fd),
+				 DMA_TO_DEVICE);
+		goto out_error;
+	}
+
+	return 0;
+
+out_error:
+	percpu_stats->tx_errors++;
+	return err;
+}
+
 static u32 dpaa_run_xdp(struct dpaa_priv *priv, struct qm_fd *fd, void *vaddr,
-			unsigned int *xdp_meta_len)
+			struct dpaa_fq *dpaa_fq, unsigned int *xdp_meta_len)
 {
 	ssize_t fd_off = qm_fd_get_offset(fd);
 	struct bpf_prog *xdp_prog;
+	struct xdp_frame *xdpf;
 	struct xdp_buff xdp;
 	u32 xdp_act;
 
@@ -2370,7 +2470,8 @@ static u32 dpaa_run_xdp(struct dpaa_priv *priv, struct qm_fd *fd, void *vaddr,
 	xdp.data_meta = xdp.data;
 	xdp.data_hard_start = xdp.data - XDP_PACKET_HEADROOM;
 	xdp.data_end = xdp.data + qm_fd_get_length(fd);
-	xdp.frame_sz = DPAA_BP_RAW_SIZE;
+	xdp.frame_sz = DPAA_BP_RAW_SIZE - DPAA_TX_PRIV_DATA_SIZE;
+	xdp.rxq = &dpaa_fq->xdp_rxq;
 
 	xdp_act = bpf_prog_run_xdp(xdp_prog, &xdp);
 
@@ -2381,6 +2482,22 @@ static u32 dpaa_run_xdp(struct dpaa_priv *priv, struct qm_fd *fd, void *vaddr,
 	case XDP_PASS:
 		*xdp_meta_len = xdp.data - xdp.data_meta;
 		break;
+	case XDP_TX:
+		/* We can access the full headroom when sending the frame
+		 * back out
+		 */
+		xdp.data_hard_start = vaddr;
+		xdp.frame_sz = DPAA_BP_RAW_SIZE;
+		xdpf = xdp_convert_buff_to_frame(&xdp);
+		if (unlikely(!xdpf)) {
+			free_pages((unsigned long)vaddr, 0);
+			break;
+		}
+
+		if (dpaa_xdp_xmit_frame(priv->net_dev, xdpf))
+			xdp_return_frame_rx_napi(xdpf);
+
+		break;
 	default:
 		bpf_warn_invalid_xdp_action(xdp_act);
 		fallthrough;
@@ -2415,6 +2532,7 @@ static enum qman_cb_dqrr_result rx_default_dqrr(struct qman_portal *portal,
 	u32 fd_status, hash_offset;
 	struct qm_sg_entry *sgt;
 	struct dpaa_bp *dpaa_bp;
+	struct dpaa_fq *dpaa_fq;
 	struct dpaa_priv *priv;
 	struct sk_buff *skb;
 	int *count_ptr;
@@ -2423,9 +2541,10 @@ static enum qman_cb_dqrr_result rx_default_dqrr(struct qman_portal *portal,
 	u32 hash;
 	u64 ns;
 
+	dpaa_fq = container_of(fq, struct dpaa_fq, fq_base);
 	fd_status = be32_to_cpu(fd->status);
 	fd_format = qm_fd_get_format(fd);
-	net_dev = ((struct dpaa_fq *)fq)->net_dev;
+	net_dev = dpaa_fq->net_dev;
 	priv = netdev_priv(net_dev);
 	dpaa_bp = dpaa_bpid2pool(dq->fd.bpid);
 	if (!dpaa_bp)
@@ -2494,7 +2613,7 @@ static enum qman_cb_dqrr_result rx_default_dqrr(struct qman_portal *portal,
 
 	if (likely(fd_format == qm_fd_contig)) {
 		xdp_act = dpaa_run_xdp(priv, (struct qm_fd *)fd, vaddr,
-				       &xdp_meta_len);
+				       dpaa_fq, &xdp_meta_len);
 		if (xdp_act != XDP_PASS) {
 			percpu_stats->rx_packets++;
 			percpu_stats->rx_bytes += qm_fd_get_length(fd);
diff --git a/drivers/net/ethernet/freescale/dpaa/dpaa_eth.h b/drivers/net/ethernet/freescale/dpaa/dpaa_eth.h
index 94e8613..5c8d52a 100644
--- a/drivers/net/ethernet/freescale/dpaa/dpaa_eth.h
+++ b/drivers/net/ethernet/freescale/dpaa/dpaa_eth.h
@@ -68,6 +68,7 @@ struct dpaa_fq {
 	u16 channel;
 	u8 wq;
 	enum dpaa_fq_type fq_type;
+	struct xdp_rxq_info xdp_rxq;
 };
 
 struct dpaa_fq_cbs {
@@ -150,6 +151,7 @@ struct dpaa_buffer_layout {
  */
 struct dpaa_eth_swbp {
 	struct sk_buff *skb;
+	struct xdp_frame *xdpf;
 };
 
 struct dpaa_priv {
-- 
1.9.1

Re: [PATCH net-next v3 4/7] dpaa_eth: add XDP_TX support

[Maciej Fijalkowski]

On Thu, Nov 19, 2020 at 06:29:33PM +0200, Camelia Groza wrote:
> Use an xdp_frame structure for managing the frame. Store a backpointer to
> the structure at the start of the buffer before enqueueing. Use the XDP
> API for freeing the buffer when it returns to the driver on the TX
> confirmation path.

Completion path?

> 
> This approach will be reused for XDP REDIRECT.
> 
> Acked-by: Madalin Bucur <madalin.bucur@oss.nxp.com>
> Signed-off-by: Camelia Groza <camelia.groza@nxp.com>
> ---
>  drivers/net/ethernet/freescale/dpaa/dpaa_eth.c | 129 ++++++++++++++++++++++++-
>  drivers/net/ethernet/freescale/dpaa/dpaa_eth.h |   2 +
>  2 files changed, 126 insertions(+), 5 deletions(-)
> 
> diff --git a/drivers/net/ethernet/freescale/dpaa/dpaa_eth.c b/drivers/net/ethernet/freescale/dpaa/dpaa_eth.c
> index 242ed45..cd5f4f6 100644
> --- a/drivers/net/ethernet/freescale/dpaa/dpaa_eth.c
> +++ b/drivers/net/ethernet/freescale/dpaa/dpaa_eth.c
> @@ -1130,6 +1130,24 @@ static int dpaa_fq_init(struct dpaa_fq *dpaa_fq, bool td_enable)
>  
>  	dpaa_fq->fqid = qman_fq_fqid(fq);
>  
> +	if (dpaa_fq->fq_type == FQ_TYPE_RX_DEFAULT ||
> +	    dpaa_fq->fq_type == FQ_TYPE_RX_PCD) {
> +		err = xdp_rxq_info_reg(&dpaa_fq->xdp_rxq, dpaa_fq->net_dev,
> +				       dpaa_fq->fqid);
> +		if (err) {
> +			dev_err(dev, "xdp_rxq_info_reg failed\n");

Print out the err?

Also, shouldn't you call qman_destroy_fq() for these error paths?

> +			return err;
> +		}
> +
> +		err = xdp_rxq_info_reg_mem_model(&dpaa_fq->xdp_rxq,
> +						 MEM_TYPE_PAGE_ORDER0, NULL);
> +		if (err) {
> +			dev_err(dev, "xdp_rxq_info_reg_mem_model failed\n");
> +			xdp_rxq_info_unreg(&dpaa_fq->xdp_rxq);
> +			return err;
> +		}
> +	}
> +
>  	return 0;
>  }
>  
> @@ -1159,6 +1177,10 @@ static int dpaa_fq_free_entry(struct device *dev, struct qman_fq *fq)
>  		}
>  	}
>  
> +	if (dpaa_fq->fq_type == FQ_TYPE_RX_DEFAULT ||
> +	    dpaa_fq->fq_type == FQ_TYPE_RX_PCD)

You should call xdp_rxq_info_is_reg() before the unregister below.

> +		xdp_rxq_info_unreg(&dpaa_fq->xdp_rxq);
> +
>  	qman_destroy_fq(fq);
>  	list_del(&dpaa_fq->list);
>  
> @@ -1625,6 +1647,9 @@ static int dpaa_eth_refill_bpools(struct dpaa_priv *priv)
>   *
>   * Return the skb backpointer, since for S/G frames the buffer containing it
>   * gets freed here.
> + *
> + * No skb backpointer is set when transmitting XDP frames. Cleanup the buffer
> + * and return NULL in this case.
>   */
>  static struct sk_buff *dpaa_cleanup_tx_fd(const struct dpaa_priv *priv,
>  					  const struct qm_fd *fd, bool ts)
> @@ -1636,6 +1661,7 @@ static struct sk_buff *dpaa_cleanup_tx_fd(const struct dpaa_priv *priv,
>  	void *vaddr = phys_to_virt(addr);
>  	const struct qm_sg_entry *sgt;
>  	struct dpaa_eth_swbp *swbp;
> +	struct xdp_frame *xdpf;

This local variable feels a bit unnecessary.

>  	struct sk_buff *skb;
>  	u64 ns;
>  	int i;
> @@ -1664,13 +1690,22 @@ static struct sk_buff *dpaa_cleanup_tx_fd(const struct dpaa_priv *priv,
>  		}
>  	} else {
>  		dma_unmap_single(priv->tx_dma_dev, addr,
> -				 priv->tx_headroom + qm_fd_get_length(fd),
> +				 qm_fd_get_offset(fd) + qm_fd_get_length(fd),
>  				 dma_dir);
>  	}
>  
>  	swbp = (struct dpaa_eth_swbp *)vaddr;
>  	skb = swbp->skb;
>  
> +	/* No skb backpointer is set when running XDP. An xdp_frame
> +	 * backpointer is saved instead.
> +	 */
> +	if (!skb) {
> +		xdpf = swbp->xdpf;
> +		xdp_return_frame(xdpf);
> +		return NULL;
> +	}
> +
>  	/* DMA unmapping is required before accessing the HW provided info */
>  	if (ts && priv->tx_tstamp &&
>  	    skb_shinfo(skb)->tx_flags & SKBTX_HW_TSTAMP) {
> @@ -2350,11 +2385,76 @@ static enum qman_cb_dqrr_result rx_error_dqrr(struct qman_portal *portal,
>  	return qman_cb_dqrr_consume;
>  }
>  
> +static int dpaa_xdp_xmit_frame(struct net_device *net_dev,
> +			       struct xdp_frame *xdpf)
> +{
> +	struct dpaa_priv *priv = netdev_priv(net_dev);
> +	struct rtnl_link_stats64 *percpu_stats;
> +	struct dpaa_percpu_priv *percpu_priv;
> +	struct dpaa_eth_swbp *swbp;
> +	struct netdev_queue *txq;
> +	void *buff_start;
> +	struct qm_fd fd;
> +	dma_addr_t addr;
> +	int err;
> +
> +	percpu_priv = this_cpu_ptr(priv->percpu_priv);
> +	percpu_stats = &percpu_priv->stats;
> +
> +	if (xdpf->headroom < DPAA_TX_PRIV_DATA_SIZE) {

Could you shed some light on DPAA_TX_PRIV_DATA_SIZE usage?

> +		err = -EINVAL;
> +		goto out_error;
> +	}
> +
> +	buff_start = xdpf->data - xdpf->headroom;
> +
> +	/* Leave empty the skb backpointer at the start of the buffer.
> +	 * Save the XDP frame for easy cleanup on confirmation.
> +	 */
> +	swbp = (struct dpaa_eth_swbp *)buff_start;
> +	swbp->skb = NULL;
> +	swbp->xdpf = xdpf;
> +
> +	qm_fd_clear_fd(&fd);
> +	fd.bpid = FSL_DPAA_BPID_INV;
> +	fd.cmd |= cpu_to_be32(FM_FD_CMD_FCO);
> +	qm_fd_set_contig(&fd, xdpf->headroom, xdpf->len);
> +
> +	addr = dma_map_single(priv->tx_dma_dev, buff_start,
> +			      xdpf->headroom + xdpf->len,
> +			      DMA_TO_DEVICE);
> +	if (unlikely(dma_mapping_error(priv->tx_dma_dev, addr))) {
> +		err = -EINVAL;
> +		goto out_error;
> +	}
> +
> +	qm_fd_addr_set64(&fd, addr);
> +
> +	/* Bump the trans_start */
> +	txq = netdev_get_tx_queue(net_dev, smp_processor_id());
> +	txq->trans_start = jiffies;
> +
> +	err = dpaa_xmit(priv, percpu_stats, smp_processor_id(), &fd);

So it looks like you don't provide the XDP Tx resources and you share the
netstack's Tx queues with XDP, if I'm reading this right.

Please mention it/explain in the cover letter or commit message of this
patch. Furthermore, I don't see any locking happenning over here?

> +	if (err) {
> +		dma_unmap_single(priv->tx_dma_dev, addr,
> +				 qm_fd_get_offset(&fd) + qm_fd_get_length(&fd),
> +				 DMA_TO_DEVICE);
> +		goto out_error;
> +	}
> +
> +	return 0;
> +
> +out_error:
> +	percpu_stats->tx_errors++;
> +	return err;
> +}
> +
>  static u32 dpaa_run_xdp(struct dpaa_priv *priv, struct qm_fd *fd, void *vaddr,
> -			unsigned int *xdp_meta_len)
> +			struct dpaa_fq *dpaa_fq, unsigned int *xdp_meta_len)
>  {
>  	ssize_t fd_off = qm_fd_get_offset(fd);
>  	struct bpf_prog *xdp_prog;
> +	struct xdp_frame *xdpf;
>  	struct xdp_buff xdp;
>  	u32 xdp_act;
>  
> @@ -2370,7 +2470,8 @@ static u32 dpaa_run_xdp(struct dpaa_priv *priv, struct qm_fd *fd, void *vaddr,
>  	xdp.data_meta = xdp.data;
>  	xdp.data_hard_start = xdp.data - XDP_PACKET_HEADROOM;
>  	xdp.data_end = xdp.data + qm_fd_get_length(fd);
> -	xdp.frame_sz = DPAA_BP_RAW_SIZE;
> +	xdp.frame_sz = DPAA_BP_RAW_SIZE - DPAA_TX_PRIV_DATA_SIZE;
> +	xdp.rxq = &dpaa_fq->xdp_rxq;
>  
>  	xdp_act = bpf_prog_run_xdp(xdp_prog, &xdp);
>  
> @@ -2381,6 +2482,22 @@ static u32 dpaa_run_xdp(struct dpaa_priv *priv, struct qm_fd *fd, void *vaddr,
>  	case XDP_PASS:
>  		*xdp_meta_len = xdp.data - xdp.data_meta;
>  		break;
> +	case XDP_TX:
> +		/* We can access the full headroom when sending the frame
> +		 * back out

And normally why a piece of headroom is taken away? I probably should have
started from the basic XDP support patch, but if you don't mind, please
explain it a bit.

> +		 */
> +		xdp.data_hard_start = vaddr;
> +		xdp.frame_sz = DPAA_BP_RAW_SIZE;
> +		xdpf = xdp_convert_buff_to_frame(&xdp);
> +		if (unlikely(!xdpf)) {
> +			free_pages((unsigned long)vaddr, 0);
> +			break;
> +		}
> +
> +		if (dpaa_xdp_xmit_frame(priv->net_dev, xdpf))
> +			xdp_return_frame_rx_napi(xdpf);
> +
> +		break;
>  	default:
>  		bpf_warn_invalid_xdp_action(xdp_act);
>  		fallthrough;
> @@ -2415,6 +2532,7 @@ static enum qman_cb_dqrr_result rx_default_dqrr(struct qman_portal *portal,
>  	u32 fd_status, hash_offset;
>  	struct qm_sg_entry *sgt;
>  	struct dpaa_bp *dpaa_bp;
> +	struct dpaa_fq *dpaa_fq;
>  	struct dpaa_priv *priv;
>  	struct sk_buff *skb;
>  	int *count_ptr;
> @@ -2423,9 +2541,10 @@ static enum qman_cb_dqrr_result rx_default_dqrr(struct qman_portal *portal,
>  	u32 hash;
>  	u64 ns;
>  
> +	dpaa_fq = container_of(fq, struct dpaa_fq, fq_base);
>  	fd_status = be32_to_cpu(fd->status);
>  	fd_format = qm_fd_get_format(fd);
> -	net_dev = ((struct dpaa_fq *)fq)->net_dev;
> +	net_dev = dpaa_fq->net_dev;
>  	priv = netdev_priv(net_dev);
>  	dpaa_bp = dpaa_bpid2pool(dq->fd.bpid);
>  	if (!dpaa_bp)
> @@ -2494,7 +2613,7 @@ static enum qman_cb_dqrr_result rx_default_dqrr(struct qman_portal *portal,
>  
>  	if (likely(fd_format == qm_fd_contig)) {
>  		xdp_act = dpaa_run_xdp(priv, (struct qm_fd *)fd, vaddr,
> -				       &xdp_meta_len);
> +				       dpaa_fq, &xdp_meta_len);
>  		if (xdp_act != XDP_PASS) {
>  			percpu_stats->rx_packets++;
>  			percpu_stats->rx_bytes += qm_fd_get_length(fd);
> diff --git a/drivers/net/ethernet/freescale/dpaa/dpaa_eth.h b/drivers/net/ethernet/freescale/dpaa/dpaa_eth.h
> index 94e8613..5c8d52a 100644
> --- a/drivers/net/ethernet/freescale/dpaa/dpaa_eth.h
> +++ b/drivers/net/ethernet/freescale/dpaa/dpaa_eth.h
> @@ -68,6 +68,7 @@ struct dpaa_fq {
>  	u16 channel;
>  	u8 wq;
>  	enum dpaa_fq_type fq_type;
> +	struct xdp_rxq_info xdp_rxq;
>  };
>  
>  struct dpaa_fq_cbs {
> @@ -150,6 +151,7 @@ struct dpaa_buffer_layout {
>   */
>  struct dpaa_eth_swbp {
>  	struct sk_buff *skb;
> +	struct xdp_frame *xdpf;
>  };
>  
>  struct dpaa_priv {
> -- 
> 1.9.1
> 

RE: [PATCH net-next v3 4/7] dpaa_eth: add XDP_TX support

[Camelia Alexandra Groza]

> -----Original Message-----
> From: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
> Sent: Friday, November 20, 2020 01:51
> To: Camelia Alexandra Groza <camelia.groza@nxp.com>
> Cc: kuba@kernel.org; brouer@redhat.com; saeed@kernel.org;
> davem@davemloft.net; Madalin Bucur (OSS)
> <madalin.bucur@oss.nxp.com>; Ioana Ciornei <ioana.ciornei@nxp.com>;
> netdev@vger.kernel.org
> Subject: Re: [PATCH net-next v3 4/7] dpaa_eth: add XDP_TX support
> 
> On Thu, Nov 19, 2020 at 06:29:33PM +0200, Camelia Groza wrote:
> > Use an xdp_frame structure for managing the frame. Store a backpointer
> to
> > the structure at the start of the buffer before enqueueing. Use the XDP
> > API for freeing the buffer when it returns to the driver on the TX
> > confirmation path.
> 
> Completion path?

The DPAA terminology uses confirmation instead of completion, but the meaning is similar. The dpaa.rst documentation has more details.

> >
> > This approach will be reused for XDP REDIRECT.
> >
> > Acked-by: Madalin Bucur <madalin.bucur@oss.nxp.com>
> > Signed-off-by: Camelia Groza <camelia.groza@nxp.com>
> > ---
> >  drivers/net/ethernet/freescale/dpaa/dpaa_eth.c | 129
> ++++++++++++++++++++++++-
> >  drivers/net/ethernet/freescale/dpaa/dpaa_eth.h |   2 +
> >  2 files changed, 126 insertions(+), 5 deletions(-)
> >
> > diff --git a/drivers/net/ethernet/freescale/dpaa/dpaa_eth.c
> b/drivers/net/ethernet/freescale/dpaa/dpaa_eth.c
> > index 242ed45..cd5f4f6 100644
> > --- a/drivers/net/ethernet/freescale/dpaa/dpaa_eth.c
> > +++ b/drivers/net/ethernet/freescale/dpaa/dpaa_eth.c
> > @@ -1130,6 +1130,24 @@ static int dpaa_fq_init(struct dpaa_fq *dpaa_fq,
> bool td_enable)
> >
> >  	dpaa_fq->fqid = qman_fq_fqid(fq);
> >
> > +	if (dpaa_fq->fq_type == FQ_TYPE_RX_DEFAULT ||
> > +	    dpaa_fq->fq_type == FQ_TYPE_RX_PCD) {
> > +		err = xdp_rxq_info_reg(&dpaa_fq->xdp_rxq, dpaa_fq-
> >net_dev,
> > +				       dpaa_fq->fqid);
> > +		if (err) {
> > +			dev_err(dev, "xdp_rxq_info_reg failed\n");
> 
> Print out the err?

I'll add the err in the next version.

> Also, shouldn't you call qman_destroy_fq() for these error paths?

Yes, for the point of undoing all configurations on the queue. The entire probe is aborted if things fail at this point, so everything is cleaned either way. I'll fix this in the next version.

> > +			return err;
> > +		}
> > +
> > +		err = xdp_rxq_info_reg_mem_model(&dpaa_fq->xdp_rxq,
> > +						 MEM_TYPE_PAGE_ORDER0,
> NULL);
> > +		if (err) {
> > +			dev_err(dev, "xdp_rxq_info_reg_mem_model
> failed\n");
> > +			xdp_rxq_info_unreg(&dpaa_fq->xdp_rxq);
> > +			return err;
> > +		}
> > +	}
> > +
> >  	return 0;
> >  }
> >
> > @@ -1159,6 +1177,10 @@ static int dpaa_fq_free_entry(struct device
> *dev, struct qman_fq *fq)
> >  		}
> >  	}
> >
> > +	if (dpaa_fq->fq_type == FQ_TYPE_RX_DEFAULT ||
> > +	    dpaa_fq->fq_type == FQ_TYPE_RX_PCD)
> 
> You should call xdp_rxq_info_is_reg() before the unregister below.

I'll add it.

> > +		xdp_rxq_info_unreg(&dpaa_fq->xdp_rxq);
> > +
> >  	qman_destroy_fq(fq);
> >  	list_del(&dpaa_fq->list);
> >
> > @@ -1625,6 +1647,9 @@ static int dpaa_eth_refill_bpools(struct dpaa_priv
> *priv)
> >   *
> >   * Return the skb backpointer, since for S/G frames the buffer containing it
> >   * gets freed here.
> > + *
> > + * No skb backpointer is set when transmitting XDP frames. Cleanup the
> buffer
> > + * and return NULL in this case.
> >   */
> >  static struct sk_buff *dpaa_cleanup_tx_fd(const struct dpaa_priv *priv,
> >  					  const struct qm_fd *fd, bool ts)
> > @@ -1636,6 +1661,7 @@ static struct sk_buff *dpaa_cleanup_tx_fd(const
> struct dpaa_priv *priv,
> >  	void *vaddr = phys_to_virt(addr);
> >  	const struct qm_sg_entry *sgt;
> >  	struct dpaa_eth_swbp *swbp;
> > +	struct xdp_frame *xdpf;
> 
> This local variable feels a bit unnecessary.

Sure, I'll remove it.

> >  	struct sk_buff *skb;
> >  	u64 ns;
> >  	int i;
> > @@ -1664,13 +1690,22 @@ static struct sk_buff
> *dpaa_cleanup_tx_fd(const struct dpaa_priv *priv,
> >  		}
> >  	} else {
> >  		dma_unmap_single(priv->tx_dma_dev, addr,
> > -				 priv->tx_headroom +
> qm_fd_get_length(fd),
> > +				 qm_fd_get_offset(fd) +
> qm_fd_get_length(fd),
> >  				 dma_dir);
> >  	}
> >
> >  	swbp = (struct dpaa_eth_swbp *)vaddr;
> >  	skb = swbp->skb;
> >
> > +	/* No skb backpointer is set when running XDP. An xdp_frame
> > +	 * backpointer is saved instead.
> > +	 */
> > +	if (!skb) {
> > +		xdpf = swbp->xdpf;
> > +		xdp_return_frame(xdpf);
> > +		return NULL;
> > +	}
> > +
> >  	/* DMA unmapping is required before accessing the HW provided
> info */
> >  	if (ts && priv->tx_tstamp &&
> >  	    skb_shinfo(skb)->tx_flags & SKBTX_HW_TSTAMP) {
> > @@ -2350,11 +2385,76 @@ static enum qman_cb_dqrr_result
> rx_error_dqrr(struct qman_portal *portal,
> >  	return qman_cb_dqrr_consume;
> >  }
> >
> > +static int dpaa_xdp_xmit_frame(struct net_device *net_dev,
> > +			       struct xdp_frame *xdpf)
> > +{
> > +	struct dpaa_priv *priv = netdev_priv(net_dev);
> > +	struct rtnl_link_stats64 *percpu_stats;
> > +	struct dpaa_percpu_priv *percpu_priv;
> > +	struct dpaa_eth_swbp *swbp;
> > +	struct netdev_queue *txq;
> > +	void *buff_start;
> > +	struct qm_fd fd;
> > +	dma_addr_t addr;
> > +	int err;
> > +
> > +	percpu_priv = this_cpu_ptr(priv->percpu_priv);
> > +	percpu_stats = &percpu_priv->stats;
> > +
> > +	if (xdpf->headroom < DPAA_TX_PRIV_DATA_SIZE) {
> 
> Could you shed some light on DPAA_TX_PRIV_DATA_SIZE usage?

We store information at the start of egress buffers that we use on the confirmation path for cleanup. DPAA_TX_PRIV_DATA_SIZE is the size of this reserved area. We require buffers to have at least this mush space available at the start before transmitting them.

> > +		err = -EINVAL;
> > +		goto out_error;
> > +	}
> > +
> > +	buff_start = xdpf->data - xdpf->headroom;
> > +
> > +	/* Leave empty the skb backpointer at the start of the buffer.
> > +	 * Save the XDP frame for easy cleanup on confirmation.
> > +	 */
> > +	swbp = (struct dpaa_eth_swbp *)buff_start;
> > +	swbp->skb = NULL;
> > +	swbp->xdpf = xdpf;
> > +
> > +	qm_fd_clear_fd(&fd);
> > +	fd.bpid = FSL_DPAA_BPID_INV;
> > +	fd.cmd |= cpu_to_be32(FM_FD_CMD_FCO);
> > +	qm_fd_set_contig(&fd, xdpf->headroom, xdpf->len);
> > +
> > +	addr = dma_map_single(priv->tx_dma_dev, buff_start,
> > +			      xdpf->headroom + xdpf->len,
> > +			      DMA_TO_DEVICE);
> > +	if (unlikely(dma_mapping_error(priv->tx_dma_dev, addr))) {
> > +		err = -EINVAL;
> > +		goto out_error;
> > +	}
> > +
> > +	qm_fd_addr_set64(&fd, addr);
> > +
> > +	/* Bump the trans_start */
> > +	txq = netdev_get_tx_queue(net_dev, smp_processor_id());
> > +	txq->trans_start = jiffies;
> > +
> > +	err = dpaa_xmit(priv, percpu_stats, smp_processor_id(), &fd);
> 
> So it looks like you don't provide the XDP Tx resources and you share the
> netstack's Tx queues with XDP, if I'm reading this right.
> 
> Please mention it/explain in the cover letter or commit message of this
> patch. Furthermore, I don't see any locking happenning over here?

Yes, that's correct, we don't have dedicated Tx queues for XDP, we share them with the netstack.

We don't need locking at this level. We are a LLTX driver so we don't take the netif_tx_lock. I'll mention this in the cover letter.

> > +	if (err) {
> > +		dma_unmap_single(priv->tx_dma_dev, addr,
> > +				 qm_fd_get_offset(&fd) +
> qm_fd_get_length(&fd),
> > +				 DMA_TO_DEVICE);
> > +		goto out_error;
> > +	}
> > +
> > +	return 0;
> > +
> > +out_error:
> > +	percpu_stats->tx_errors++;
> > +	return err;
> > +}
> > +
> >  static u32 dpaa_run_xdp(struct dpaa_priv *priv, struct qm_fd *fd, void
> *vaddr,
> > -			unsigned int *xdp_meta_len)
> > +			struct dpaa_fq *dpaa_fq, unsigned int
> *xdp_meta_len)
> >  {
> >  	ssize_t fd_off = qm_fd_get_offset(fd);
> >  	struct bpf_prog *xdp_prog;
> > +	struct xdp_frame *xdpf;
> >  	struct xdp_buff xdp;
> >  	u32 xdp_act;
> >
> > @@ -2370,7 +2470,8 @@ static u32 dpaa_run_xdp(struct dpaa_priv *priv,
> struct qm_fd *fd, void *vaddr,
> >  	xdp.data_meta = xdp.data;
> >  	xdp.data_hard_start = xdp.data - XDP_PACKET_HEADROOM;
> >  	xdp.data_end = xdp.data + qm_fd_get_length(fd);
> > -	xdp.frame_sz = DPAA_BP_RAW_SIZE;
> > +	xdp.frame_sz = DPAA_BP_RAW_SIZE - DPAA_TX_PRIV_DATA_SIZE;
> > +	xdp.rxq = &dpaa_fq->xdp_rxq;
> >
> >  	xdp_act = bpf_prog_run_xdp(xdp_prog, &xdp);
> >
> > @@ -2381,6 +2482,22 @@ static u32 dpaa_run_xdp(struct dpaa_priv *priv,
> struct qm_fd *fd, void *vaddr,
> >  	case XDP_PASS:
> >  		*xdp_meta_len = xdp.data - xdp.data_meta;
> >  		break;
> > +	case XDP_TX:
> > +		/* We can access the full headroom when sending the frame
> > +		 * back out
> 
> And normally why a piece of headroom is taken away? I probably should
> have
> started from the basic XDP support patch, but if you don't mind, please
> explain it a bit.

I mentioned we require DPAA_TX_PRIV_DATA_SIZE bytes at the start of the buffer in order to make sure we have enough space for our private info.

When setting up the xdp_buff, this area is reserved from the frame size exposed to the XDP program.
 -	xdp.frame_sz = DPAA_BP_RAW_SIZE;
 +	xdp.frame_sz = DPAA_BP_RAW_SIZE - DPAA_TX_PRIV_DATA_SIZE;

After the XDP_TX verdict, we're sure that DPAA_TX_PRIV_DATA_SIZE bytes at the start of the buffer are free and we can use the full headroom how it suits us, hence the increase of the frame size back to DPAA_BP_RAW_SIZE.

Thanks for all your feedback.

> > +		 */
> > +		xdp.data_hard_start = vaddr;
> > +		xdp.frame_sz = DPAA_BP_RAW_SIZE;
> > +		xdpf = xdp_convert_buff_to_frame(&xdp);
> > +		if (unlikely(!xdpf)) {
> > +			free_pages((unsigned long)vaddr, 0);
> > +			break;
> > +		}
> > +
> > +		if (dpaa_xdp_xmit_frame(priv->net_dev, xdpf))
> > +			xdp_return_frame_rx_napi(xdpf);
> > +
> > +		break;
> >  	default:
> >  		bpf_warn_invalid_xdp_action(xdp_act);
> >  		fallthrough;
> > @@ -2415,6 +2532,7 @@ static enum qman_cb_dqrr_result
> rx_default_dqrr(struct qman_portal *portal,
> >  	u32 fd_status, hash_offset;
> >  	struct qm_sg_entry *sgt;
> >  	struct dpaa_bp *dpaa_bp;
> > +	struct dpaa_fq *dpaa_fq;
> >  	struct dpaa_priv *priv;
> >  	struct sk_buff *skb;
> >  	int *count_ptr;
> > @@ -2423,9 +2541,10 @@ static enum qman_cb_dqrr_result
> rx_default_dqrr(struct qman_portal *portal,
> >  	u32 hash;
> >  	u64 ns;
> >
> > +	dpaa_fq = container_of(fq, struct dpaa_fq, fq_base);
> >  	fd_status = be32_to_cpu(fd->status);
> >  	fd_format = qm_fd_get_format(fd);
> > -	net_dev = ((struct dpaa_fq *)fq)->net_dev;
> > +	net_dev = dpaa_fq->net_dev;
> >  	priv = netdev_priv(net_dev);
> >  	dpaa_bp = dpaa_bpid2pool(dq->fd.bpid);
> >  	if (!dpaa_bp)
> > @@ -2494,7 +2613,7 @@ static enum qman_cb_dqrr_result
> rx_default_dqrr(struct qman_portal *portal,
> >
> >  	if (likely(fd_format == qm_fd_contig)) {
> >  		xdp_act = dpaa_run_xdp(priv, (struct qm_fd *)fd, vaddr,
> > -				       &xdp_meta_len);
> > +				       dpaa_fq, &xdp_meta_len);
> >  		if (xdp_act != XDP_PASS) {
> >  			percpu_stats->rx_packets++;
> >  			percpu_stats->rx_bytes += qm_fd_get_length(fd);
> > diff --git a/drivers/net/ethernet/freescale/dpaa/dpaa_eth.h
> b/drivers/net/ethernet/freescale/dpaa/dpaa_eth.h
> > index 94e8613..5c8d52a 100644
> > --- a/drivers/net/ethernet/freescale/dpaa/dpaa_eth.h
> > +++ b/drivers/net/ethernet/freescale/dpaa/dpaa_eth.h
> > @@ -68,6 +68,7 @@ struct dpaa_fq {
> >  	u16 channel;
> >  	u8 wq;
> >  	enum dpaa_fq_type fq_type;
> > +	struct xdp_rxq_info xdp_rxq;
> >  };
> >
> >  struct dpaa_fq_cbs {
> > @@ -150,6 +151,7 @@ struct dpaa_buffer_layout {
> >   */
> >  struct dpaa_eth_swbp {
> >  	struct sk_buff *skb;
> > +	struct xdp_frame *xdpf;
> >  };
> >
> >  struct dpaa_priv {
> > --
> > 1.9.1
> >

Re: [PATCH net-next v3 4/7] dpaa_eth: add XDP_TX support

[Maciej Fijalkowski]

On Fri, Nov 20, 2020 at 06:54:42PM +0000, Camelia Alexandra Groza wrote:
> > -----Original Message-----
> > From: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
> > Sent: Friday, November 20, 2020 01:51
> > To: Camelia Alexandra Groza <camelia.groza@nxp.com>
> > Cc: kuba@kernel.org; brouer@redhat.com; saeed@kernel.org;
> > davem@davemloft.net; Madalin Bucur (OSS)
> > <madalin.bucur@oss.nxp.com>; Ioana Ciornei <ioana.ciornei@nxp.com>;
> > netdev@vger.kernel.org
> > Subject: Re: [PATCH net-next v3 4/7] dpaa_eth: add XDP_TX support
> > 
> > On Thu, Nov 19, 2020 at 06:29:33PM +0200, Camelia Groza wrote:
> > > Use an xdp_frame structure for managing the frame. Store a backpointer
> > to
> > > the structure at the start of the buffer before enqueueing. Use the XDP
> > > API for freeing the buffer when it returns to the driver on the TX
> > > confirmation path.
> > 

[...]

> > >  static u32 dpaa_run_xdp(struct dpaa_priv *priv, struct qm_fd *fd, void
> > *vaddr,
> > > -			unsigned int *xdp_meta_len)
> > > +			struct dpaa_fq *dpaa_fq, unsigned int
> > *xdp_meta_len)
> > >  {
> > >  	ssize_t fd_off = qm_fd_get_offset(fd);
> > >  	struct bpf_prog *xdp_prog;
> > > +	struct xdp_frame *xdpf;
> > >  	struct xdp_buff xdp;
> > >  	u32 xdp_act;
> > >
> > > @@ -2370,7 +2470,8 @@ static u32 dpaa_run_xdp(struct dpaa_priv *priv,
> > struct qm_fd *fd, void *vaddr,
> > >  	xdp.data_meta = xdp.data;
> > >  	xdp.data_hard_start = xdp.data - XDP_PACKET_HEADROOM;
> > >  	xdp.data_end = xdp.data + qm_fd_get_length(fd);
> > > -	xdp.frame_sz = DPAA_BP_RAW_SIZE;
> > > +	xdp.frame_sz = DPAA_BP_RAW_SIZE - DPAA_TX_PRIV_DATA_SIZE;
> > > +	xdp.rxq = &dpaa_fq->xdp_rxq;
> > >
> > >  	xdp_act = bpf_prog_run_xdp(xdp_prog, &xdp);
> > >
> > > @@ -2381,6 +2482,22 @@ static u32 dpaa_run_xdp(struct dpaa_priv *priv,
> > struct qm_fd *fd, void *vaddr,
> > >  	case XDP_PASS:
> > >  		*xdp_meta_len = xdp.data - xdp.data_meta;
> > >  		break;
> > > +	case XDP_TX:
> > > +		/* We can access the full headroom when sending the frame
> > > +		 * back out
> > 
> > And normally why a piece of headroom is taken away? I probably should
> > have
> > started from the basic XDP support patch, but if you don't mind, please
> > explain it a bit.
> 
> I mentioned we require DPAA_TX_PRIV_DATA_SIZE bytes at the start of the buffer in order to make sure we have enough space for our private info.

What is your private info?

> 
> When setting up the xdp_buff, this area is reserved from the frame size exposed to the XDP program.
>  -	xdp.frame_sz = DPAA_BP_RAW_SIZE;
>  +	xdp.frame_sz = DPAA_BP_RAW_SIZE - DPAA_TX_PRIV_DATA_SIZE;
> 
> After the XDP_TX verdict, we're sure that DPAA_TX_PRIV_DATA_SIZE bytes at the start of the buffer are free and we can use the full headroom how it suits us, hence the increase of the frame size back to DPAA_BP_RAW_SIZE.

Not at the *end* of the buffer?

> 
> Thanks for all your feedback.
> 
> > > +		 */
> > > +		xdp.data_hard_start = vaddr;
> > > +		xdp.frame_sz = DPAA_BP_RAW_SIZE;
> > > +		xdpf = xdp_convert_buff_to_frame(&xdp);
> > > +		if (unlikely(!xdpf)) {
> > > +			free_pages((unsigned long)vaddr, 0);
> > > +			break;
> > > +		}
> > > +
> > > +		if (dpaa_xdp_xmit_frame(priv->net_dev, xdpf))
> > > +			xdp_return_frame_rx_napi(xdpf);
> > > +
> > > +		break;
> > >  	default:
> > >  		bpf_warn_invalid_xdp_action(xdp_act);
> > >  		fallthrough;
> > > @@ -2415,6 +2532,7 @@ static enum qman_cb_dqrr_result

RE: [PATCH net-next v3 4/7] dpaa_eth: add XDP_TX support

[Camelia Alexandra Groza]

> -----Original Message-----
> From: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
> Sent: Tuesday, November 24, 2020 00:18
> To: Camelia Alexandra Groza <camelia.groza@nxp.com>
> Cc: kuba@kernel.org; brouer@redhat.com; saeed@kernel.org;
> davem@davemloft.net; Madalin Bucur (OSS)
> <madalin.bucur@oss.nxp.com>; Ioana Ciornei <ioana.ciornei@nxp.com>;
> netdev@vger.kernel.org
> Subject: Re: [PATCH net-next v3 4/7] dpaa_eth: add XDP_TX support
> 
> On Fri, Nov 20, 2020 at 06:54:42PM +0000, Camelia Alexandra Groza wrote:
> > > -----Original Message-----
> > > From: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
> > > Sent: Friday, November 20, 2020 01:51
> > > To: Camelia Alexandra Groza <camelia.groza@nxp.com>
> > > Cc: kuba@kernel.org; brouer@redhat.com; saeed@kernel.org;
> > > davem@davemloft.net; Madalin Bucur (OSS)
> > > <madalin.bucur@oss.nxp.com>; Ioana Ciornei <ioana.ciornei@nxp.com>;
> > > netdev@vger.kernel.org
> > > Subject: Re: [PATCH net-next v3 4/7] dpaa_eth: add XDP_TX support
> > >
> > > On Thu, Nov 19, 2020 at 06:29:33PM +0200, Camelia Groza wrote:
> > > > Use an xdp_frame structure for managing the frame. Store a
> backpointer
> > > to
> > > > the structure at the start of the buffer before enqueueing. Use the XDP
> > > > API for freeing the buffer when it returns to the driver on the TX
> > > > confirmation path.
> > >
> 
> [...]
> 
> > > >  static u32 dpaa_run_xdp(struct dpaa_priv *priv, struct qm_fd *fd, void
> > > *vaddr,
> > > > -			unsigned int *xdp_meta_len)
> > > > +			struct dpaa_fq *dpaa_fq, unsigned int
> > > *xdp_meta_len)
> > > >  {
> > > >  	ssize_t fd_off = qm_fd_get_offset(fd);
> > > >  	struct bpf_prog *xdp_prog;
> > > > +	struct xdp_frame *xdpf;
> > > >  	struct xdp_buff xdp;
> > > >  	u32 xdp_act;
> > > >
> > > > @@ -2370,7 +2470,8 @@ static u32 dpaa_run_xdp(struct dpaa_priv
> *priv,
> > > struct qm_fd *fd, void *vaddr,
> > > >  	xdp.data_meta = xdp.data;
> > > >  	xdp.data_hard_start = xdp.data - XDP_PACKET_HEADROOM;
> > > >  	xdp.data_end = xdp.data + qm_fd_get_length(fd);
> > > > -	xdp.frame_sz = DPAA_BP_RAW_SIZE;
> > > > +	xdp.frame_sz = DPAA_BP_RAW_SIZE - DPAA_TX_PRIV_DATA_SIZE;
> > > > +	xdp.rxq = &dpaa_fq->xdp_rxq;
> > > >
> > > >  	xdp_act = bpf_prog_run_xdp(xdp_prog, &xdp);
> > > >
> > > > @@ -2381,6 +2482,22 @@ static u32 dpaa_run_xdp(struct dpaa_priv
> *priv,
> > > struct qm_fd *fd, void *vaddr,
> > > >  	case XDP_PASS:
> > > >  		*xdp_meta_len = xdp.data - xdp.data_meta;
> > > >  		break;
> > > > +	case XDP_TX:
> > > > +		/* We can access the full headroom when sending the frame
> > > > +		 * back out
> > >
> > > And normally why a piece of headroom is taken away? I probably should
> > > have
> > > started from the basic XDP support patch, but if you don't mind, please
> > > explain it a bit.
> >
> > I mentioned we require DPAA_TX_PRIV_DATA_SIZE bytes at the start of
> the buffer in order to make sure we have enough space for our private info.
> 
> What is your private info?

The dpaa_eth_swbp struct from the first patch. It's the xdp_frame reference mentioned in the patch description, stored for cleanup on confirmation. We also store a skb reference for non-XDP use cases.

> >
> > When setting up the xdp_buff, this area is reserved from the frame size
> exposed to the XDP program.
> >  -	xdp.frame_sz = DPAA_BP_RAW_SIZE;
> >  +	xdp.frame_sz = DPAA_BP_RAW_SIZE - DPAA_TX_PRIV_DATA_SIZE;
> >
> > After the XDP_TX verdict, we're sure that DPAA_TX_PRIV_DATA_SIZE
> bytes at the start of the buffer are free and we can use the full headroom
> how it suits us, hence the increase of the frame size back to
> DPAA_BP_RAW_SIZE.
> 
> Not at the *end* of the buffer?

No, we store this information at the start of the buffer, before the XDP_PACKET_HEADROOM.
 
> >
> > Thanks for all your feedback.
> >
> > > > +		 */
> > > > +		xdp.data_hard_start = vaddr;
> > > > +		xdp.frame_sz = DPAA_BP_RAW_SIZE;
> > > > +		xdpf = xdp_convert_buff_to_frame(&xdp);
> > > > +		if (unlikely(!xdpf)) {
> > > > +			free_pages((unsigned long)vaddr, 0);
> > > > +			break;
> > > > +		}
> > > > +
> > > > +		if (dpaa_xdp_xmit_frame(priv->net_dev, xdpf))
> > > > +			xdp_return_frame_rx_napi(xdpf);
> > > > +
> > > > +		break;
> > > >  	default:
> > > >  		bpf_warn_invalid_xdp_action(xdp_act);
> > > >  		fallthrough;
> > > > @@ -2415,6 +2532,7 @@ static enum qman_cb_dqrr_result

[PATCH net-next v3 5/7] dpaa_eth: add XDP_REDIRECT support

[Camelia Groza]

After transmission, the frame is returned on confirmation queues for
cleanup. For this, store a backpointer to the xdp_frame in the private
reserved area at the start of the TX buffer.

No TX batching support is implemented at this time.

Acked-by: Madalin Bucur <madalin.bucur@oss.nxp.com>
Signed-off-by: Camelia Groza <camelia.groza@nxp.com>
---
 drivers/net/ethernet/freescale/dpaa/dpaa_eth.c | 48 +++++++++++++++++++++++++-
 drivers/net/ethernet/freescale/dpaa/dpaa_eth.h |  1 +
 2 files changed, 48 insertions(+), 1 deletion(-)

diff --git a/drivers/net/ethernet/freescale/dpaa/dpaa_eth.c b/drivers/net/ethernet/freescale/dpaa/dpaa_eth.c
index cd5f4f6..3214ea0 100644
--- a/drivers/net/ethernet/freescale/dpaa/dpaa_eth.c
+++ b/drivers/net/ethernet/freescale/dpaa/dpaa_eth.c
@@ -2305,8 +2305,11 @@ static int dpaa_eth_poll(struct napi_struct *napi, int budget)
 {
 	struct dpaa_napi_portal *np =
 			container_of(napi, struct dpaa_napi_portal, napi);
+	int cleaned;
 
-	int cleaned = qman_p_poll_dqrr(np->p, budget);
+	np->xdp_act = 0;
+
+	cleaned = qman_p_poll_dqrr(np->p, budget);
 
 	if (cleaned < budget) {
 		napi_complete_done(napi, cleaned);
@@ -2315,6 +2318,9 @@ static int dpaa_eth_poll(struct napi_struct *napi, int budget)
 		qman_p_irqsource_add(np->p, QM_PIRQ_DQRI);
 	}
 
+	if (np->xdp_act & XDP_REDIRECT)
+		xdp_do_flush();
+
 	return cleaned;
 }
 
@@ -2457,6 +2463,7 @@ static u32 dpaa_run_xdp(struct dpaa_priv *priv, struct qm_fd *fd, void *vaddr,
 	struct xdp_frame *xdpf;
 	struct xdp_buff xdp;
 	u32 xdp_act;
+	int err;
 
 	rcu_read_lock();
 
@@ -2498,6 +2505,17 @@ static u32 dpaa_run_xdp(struct dpaa_priv *priv, struct qm_fd *fd, void *vaddr,
 			xdp_return_frame_rx_napi(xdpf);
 
 		break;
+	case XDP_REDIRECT:
+		/* Allow redirect to use the full headroom */
+		xdp.data_hard_start = vaddr;
+		xdp.frame_sz = DPAA_BP_RAW_SIZE;
+
+		err = xdp_do_redirect(priv->net_dev, &xdp, xdp_prog);
+		if (err) {
+			trace_xdp_exception(priv->net_dev, xdp_prog, xdp_act);
+			free_pages((unsigned long)vaddr, 0);
+		}
+		break;
 	default:
 		bpf_warn_invalid_xdp_action(xdp_act);
 		fallthrough;
@@ -2527,6 +2545,7 @@ static enum qman_cb_dqrr_result rx_default_dqrr(struct qman_portal *portal,
 	struct dpaa_percpu_priv *percpu_priv;
 	const struct qm_fd *fd = &dq->fd;
 	dma_addr_t addr = qm_fd_addr(fd);
+	struct dpaa_napi_portal *np;
 	enum qm_fd_format fd_format;
 	struct net_device *net_dev;
 	u32 fd_status, hash_offset;
@@ -2541,6 +2560,7 @@ static enum qman_cb_dqrr_result rx_default_dqrr(struct qman_portal *portal,
 	u32 hash;
 	u64 ns;
 
+	np = container_of(&portal, struct dpaa_napi_portal, p);
 	dpaa_fq = container_of(fq, struct dpaa_fq, fq_base);
 	fd_status = be32_to_cpu(fd->status);
 	fd_format = qm_fd_get_format(fd);
@@ -2614,6 +2634,7 @@ static enum qman_cb_dqrr_result rx_default_dqrr(struct qman_portal *portal,
 	if (likely(fd_format == qm_fd_contig)) {
 		xdp_act = dpaa_run_xdp(priv, (struct qm_fd *)fd, vaddr,
 				       dpaa_fq, &xdp_meta_len);
+		np->xdp_act |= xdp_act;
 		if (xdp_act != XDP_PASS) {
 			percpu_stats->rx_packets++;
 			percpu_stats->rx_bytes += qm_fd_get_length(fd);
@@ -2942,6 +2963,30 @@ static int dpaa_xdp(struct net_device *net_dev, struct netdev_bpf *xdp)
 	}
 }
 
+static int dpaa_xdp_xmit(struct net_device *net_dev, int n,
+			 struct xdp_frame **frames, u32 flags)
+{
+	struct xdp_frame *xdpf;
+	int i, err, drops = 0;
+
+	if (unlikely(flags & ~XDP_XMIT_FLAGS_MASK))
+		return -EINVAL;
+
+	if (!netif_running(net_dev))
+		return -ENETDOWN;
+
+	for (i = 0; i < n; i++) {
+		xdpf = frames[i];
+		err = dpaa_xdp_xmit_frame(net_dev, xdpf);
+		if (err) {
+			xdp_return_frame_rx_napi(xdpf);
+			drops++;
+		}
+	}
+
+	return n - drops;
+}
+
 static int dpaa_ts_ioctl(struct net_device *dev, struct ifreq *rq, int cmd)
 {
 	struct dpaa_priv *priv = netdev_priv(dev);
@@ -3010,6 +3055,7 @@ static int dpaa_ioctl(struct net_device *net_dev, struct ifreq *rq, int cmd)
 	.ndo_setup_tc = dpaa_setup_tc,
 	.ndo_change_mtu = dpaa_change_mtu,
 	.ndo_bpf = dpaa_xdp,
+	.ndo_xdp_xmit = dpaa_xdp_xmit,
 };
 
 static int dpaa_napi_add(struct net_device *net_dev)
diff --git a/drivers/net/ethernet/freescale/dpaa/dpaa_eth.h b/drivers/net/ethernet/freescale/dpaa/dpaa_eth.h
index 5c8d52a..daf894a 100644
--- a/drivers/net/ethernet/freescale/dpaa/dpaa_eth.h
+++ b/drivers/net/ethernet/freescale/dpaa/dpaa_eth.h
@@ -127,6 +127,7 @@ struct dpaa_napi_portal {
 	struct napi_struct napi;
 	struct qman_portal *p;
 	bool down;
+	int xdp_act;
 };
 
 struct dpaa_percpu_priv {
-- 
1.9.1

[PATCH net-next v3 6/7] dpaa_eth: rename current skb A050385 erratum workaround

[Camelia Groza]

Explicitly point that the current workaround addresses skbs. This change is
in preparation for adding a workaround for XDP scenarios.

Acked-by: Madalin Bucur <madalin.bucur@oss.nxp.com>
Signed-off-by: Camelia Groza <camelia.groza@nxp.com>
---
 drivers/net/ethernet/freescale/dpaa/dpaa_eth.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/drivers/net/ethernet/freescale/dpaa/dpaa_eth.c b/drivers/net/ethernet/freescale/dpaa/dpaa_eth.c
index 3214ea0..b9d46e6 100644
--- a/drivers/net/ethernet/freescale/dpaa/dpaa_eth.c
+++ b/drivers/net/ethernet/freescale/dpaa/dpaa_eth.c
@@ -2105,7 +2105,7 @@ static inline int dpaa_xmit(struct dpaa_priv *priv,
 }
 
 #ifdef CONFIG_DPAA_ERRATUM_A050385
-static int dpaa_a050385_wa(struct net_device *net_dev, struct sk_buff **s)
+static int dpaa_a050385_wa_skb(struct net_device *net_dev, struct sk_buff **s)
 {
 	struct dpaa_priv *priv = netdev_priv(net_dev);
 	struct sk_buff *new_skb, *skb = *s;
@@ -2221,7 +2221,7 @@ static int dpaa_a050385_wa(struct net_device *net_dev, struct sk_buff **s)
 
 #ifdef CONFIG_DPAA_ERRATUM_A050385
 	if (unlikely(fman_has_errata_a050385())) {
-		if (dpaa_a050385_wa(net_dev, &skb))
+		if (dpaa_a050385_wa_skb(net_dev, &skb))
 			goto enomem;
 		nonlinear = skb_is_nonlinear(skb);
 	}
-- 
1.9.1

[PATCH net-next v3 7/7] dpaa_eth: implement the A050385 erratum workaround for XDP

[Camelia Groza]

For XDP TX, even tough we start out with correctly aligned buffers, the
XDP program might change the data's alignment. For REDIRECT, we have no
control over the alignment either.

Create a new workaround for xdp_frame structures to verify the erratum
conditions and move the data to a fresh buffer if necessary. Create a new
xdp_frame for managing the new buffer and free the old one using the XDP
API.

Due to alignment constraints, all frames have a 256 byte headroom that
is offered fully to XDP under the erratum. If the XDP program uses all
of it, the data needs to be move to make room for the xdpf backpointer.

Disable the metadata support since the information can be lost.

Acked-by: Madalin Bucur <madalin.bucur@oss.nxp.com>
Signed-off-by: Camelia Groza <camelia.groza@nxp.com>
---
 drivers/net/ethernet/freescale/dpaa/dpaa_eth.c | 82 +++++++++++++++++++++++++-
 1 file changed, 79 insertions(+), 3 deletions(-)

diff --git a/drivers/net/ethernet/freescale/dpaa/dpaa_eth.c b/drivers/net/ethernet/freescale/dpaa/dpaa_eth.c
index b9d46e6..aaf9112 100644
--- a/drivers/net/ethernet/freescale/dpaa/dpaa_eth.c
+++ b/drivers/net/ethernet/freescale/dpaa/dpaa_eth.c
@@ -2171,6 +2171,52 @@ static int dpaa_a050385_wa_skb(struct net_device *net_dev, struct sk_buff **s)
 
 	return 0;
 }
+
+static int dpaa_a050385_wa_xdpf(struct dpaa_priv *priv,
+				struct xdp_frame **init_xdpf)
+{
+	struct xdp_frame *new_xdpf, *xdpf = *init_xdpf;
+	void *new_buff;
+	struct page *p;
+
+	/* Check the data alignment and make sure the headroom is large
+	 * enough to store the xdpf backpointer. Use an aligned headroom
+	 * value.
+	 *
+	 * Due to alignment constraints, we give XDP access to the full 256
+	 * byte frame headroom. If the XDP program uses all of it, copy the
+	 * data to a new buffer and make room for storing the backpointer.
+	 */
+	if (PTR_IS_ALIGNED(xdpf->data, DPAA_A050385_ALIGN) &&
+	    xdpf->headroom >= priv->tx_headroom) {
+		xdpf->headroom = priv->tx_headroom;
+		return 0;
+	}
+
+	p = dev_alloc_pages(0);
+	if (unlikely(!p))
+		return -ENOMEM;
+
+	/* Copy the data to the new buffer at a properly aligned offset */
+	new_buff = page_address(p);
+	memcpy(new_buff + priv->tx_headroom, xdpf->data, xdpf->len);
+
+	/* Create an XDP frame around the new buffer in a similar fashion
+	 * to xdp_convert_buff_to_frame.
+	 */
+	new_xdpf = new_buff;
+	new_xdpf->data = new_buff + priv->tx_headroom;
+	new_xdpf->len = xdpf->len;
+	new_xdpf->headroom = priv->tx_headroom;
+	new_xdpf->frame_sz = DPAA_BP_RAW_SIZE;
+	new_xdpf->mem.type = MEM_TYPE_PAGE_ORDER0;
+
+	/* Release the initial buffer */
+	xdp_return_frame_rx_napi(xdpf);
+
+	*init_xdpf = new_xdpf;
+	return 0;
+}
 #endif
 
 static netdev_tx_t
@@ -2407,6 +2453,15 @@ static int dpaa_xdp_xmit_frame(struct net_device *net_dev,
 	percpu_priv = this_cpu_ptr(priv->percpu_priv);
 	percpu_stats = &percpu_priv->stats;
 
+#ifdef CONFIG_DPAA_ERRATUM_A050385
+	if (unlikely(fman_has_errata_a050385())) {
+		if (dpaa_a050385_wa_xdpf(priv, &xdpf)) {
+			err = -ENOMEM;
+			goto out_error;
+		}
+	}
+#endif
+
 	if (xdpf->headroom < DPAA_TX_PRIV_DATA_SIZE) {
 		err = -EINVAL;
 		goto out_error;
@@ -2480,6 +2535,20 @@ static u32 dpaa_run_xdp(struct dpaa_priv *priv, struct qm_fd *fd, void *vaddr,
 	xdp.frame_sz = DPAA_BP_RAW_SIZE - DPAA_TX_PRIV_DATA_SIZE;
 	xdp.rxq = &dpaa_fq->xdp_rxq;
 
+	/* We reserve a fixed headroom of 256 bytes under the erratum and we
+	 * offer it all to XDP programs to use. If no room is left for the
+	 * xdpf backpointer on TX, we will need to copy the data.
+	 * Disable metadata support since data realignments might be required
+	 * and the information can be lost.
+	 */
+#ifdef CONFIG_DPAA_ERRATUM_A050385
+	if (unlikely(fman_has_errata_a050385())) {
+		xdp_set_data_meta_invalid(&xdp);
+		xdp.data_hard_start = vaddr;
+		xdp.frame_sz = DPAA_BP_RAW_SIZE;
+	}
+#endif
+
 	xdp_act = bpf_prog_run_xdp(xdp_prog, &xdp);
 
 	/* Update the length and the offset of the FD */
@@ -2487,7 +2556,8 @@ static u32 dpaa_run_xdp(struct dpaa_priv *priv, struct qm_fd *fd, void *vaddr,
 
 	switch (xdp_act) {
 	case XDP_PASS:
-		*xdp_meta_len = xdp.data - xdp.data_meta;
+		*xdp_meta_len = xdp_data_meta_unsupported(&xdp) ? 0 :
+				xdp.data - xdp.data_meta;
 		break;
 	case XDP_TX:
 		/* We can access the full headroom when sending the frame
@@ -3187,10 +3257,16 @@ static u16 dpaa_get_headroom(struct dpaa_buffer_layout *bl,
 	 */
 	headroom = (u16)(bl[port].priv_data_size + DPAA_HWA_SIZE);
 
-	if (port == RX)
+	if (port == RX) {
+#ifdef CONFIG_DPAA_ERRATUM_A050385
+		if (unlikely(fman_has_errata_a050385()))
+			headroom = XDP_PACKET_HEADROOM;
+#endif
+
 		return ALIGN(headroom, DPAA_FD_RX_DATA_ALIGNMENT);
-	else
+	} else {
 		return ALIGN(headroom, DPAA_FD_DATA_ALIGNMENT);
+	}
 }
 
 static int dpaa_eth_probe(struct platform_device *pdev)
-- 
1.9.1