net/mlx5e: Add XSK zero-copy support

This commit adds support for AF_XDP zero-copy RX and TX.

We create a dedicated XSK RQ inside the channel, it means that two
RQs are running simultaneously: one for non-XSK traffic and the other
for XSK traffic. The regular and XSK RQs use a single ID namespace split
into two halves: the lower half is regular RQs, and the upper half is
XSK RQs. When any zero-copy AF_XDP socket is active, changing the number
of channels is not allowed, because it would break to mapping between
XSK RQ IDs and channels.

XSK requires different page allocation and release routines. Such
functions as mlx5e_{alloc,free}_rx_mpwqe and mlx5e_{get,put}_rx_frag are
generic enough to be used for both regular and XSK RQs, and they use the
mlx5e_page_{alloc,release} wrappers around the real allocation
functions. Function pointers are not used to avoid losing the
performance with retpolines. Wherever it's certain that the regular
(non-XSK) page release function should be used, it's called directly.

Only the stats that could be meaningful for XSK are exposed to the
userspace. Those that don't take part in the XSK flow are not
considered.

Note that we don't wait for WQEs on the XSK RQ (unlike the regular RQ),
because the newer xdpsock sample doesn't provide any Fill Ring entries
at the setup stage.

We create a dedicated XSK SQ in the channel. This separation has its
advantages:

1. When the UMEM is closed, the XSK SQ can also be closed and stop
receiving completions. If an existing SQ was used for XSK, it would
continue receiving completions for the packets of the closed socket. If
a new UMEM was opened at that point, it would start getting completions
that don't belong to it.

2. Calculating statistics separately.

When the userspace kicks the TX, the driver triggers a hardware
interrupt by posting a NOP to a dedicated XSK ICO (internal control
operations) SQ, in order to trigger NAPI on the right CPU core. This XSK
ICO SQ is protected by a spinlock, as the userspace application may kick
the TX from any core.

Store the pointers to the UMEMs in the net device private context,
independently from the kernel. This way the driver can distinguish
between the zero-copy and non-zero-copy UMEMs. The kernel function
xdp_get_umem_from_qid does not care about this difference, but the
driver is only interested in zero-copy UMEMs, particularly, on the
cleanup it determines whether to close the XSK RQ and SQ or not by
looking at the presence of the UMEM. Use state_lock to protect the
access to this area of UMEM pointers.

LRO isn't compatible with XDP, but there may be active UMEMs while
XDP is off. If this is the case, don't allow LRO to ensure XDP can
be reenabled at any time.

The validation of XSK parameters typically happens when XSK queues
open. However, when the interface is down or the XDP program isn't
set, it's still possible to have active AF_XDP sockets and even to
open new, but the XSK queues will be closed. To cover these cases,
perform the validation also in these flows:

1. A new UMEM is registered, but the XSK queues aren't going to be
created due to missing XDP program or interface being down.

2. MTU changes while there are UMEMs registered.

Having this early check prevents mlx5e_open_channels from failing
at a later stage, where recovery is impossible and the application
has no chance to handle the error, because it got the successful
return value for an MTU change or XSK open operation.

The performance testing was performed on a machine with the following
configuration:

- 24 cores of Intel Xeon E5-2620 v3 @ 2.40 GHz
- Mellanox ConnectX-5 Ex with 100 Gbit/s link

The results with retpoline disabled, single stream:

txonly: 33.3 Mpps (21.5 Mpps with queue and app pinned to the same CPU)
rxdrop: 12.2 Mpps
l2fwd: 9.4 Mpps

The results with retpoline enabled, single stream:

txonly: 21.3 Mpps (14.1 Mpps with queue and app pinned to the same CPU)
rxdrop: 9.9 Mpps
l2fwd: 6.8 Mpps

Signed-off-by: Maxim Mikityanskiy <maximmi@mellanox.com>
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Acked-by: Saeed Mahameed <saeedm@mellanox.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>

13 files changed, 1091 insertions, 52 deletions

diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en/params.c b/drivers/net/ethernet/mellanox/mlx5/core/en/params.c
index 0de908b12fcc..79301d116667 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/en/params.c
+++ b/drivers/net/ethernet/mellanox/mlx5/core/en/params.c
@@ -49,48 +49,56 @@ u32 mlx5e_rx_get_linear_frag_sz(struct mlx5e_params *params,
 	return frag_sz;
 }
 
-u8 mlx5e_mpwqe_log_pkts_per_wqe(struct mlx5e_params *params)
+u8 mlx5e_mpwqe_log_pkts_per_wqe(struct mlx5e_params *params,
+				struct mlx5e_xsk_param *xsk)
 {
-	u32 linear_frag_sz = mlx5e_rx_get_linear_frag_sz(params, NULL);
+	u32 linear_frag_sz = mlx5e_rx_get_linear_frag_sz(params, xsk);
 
 	return MLX5_MPWRQ_LOG_WQE_SZ - order_base_2(linear_frag_sz);
 }
 
-bool mlx5e_rx_is_linear_skb(struct mlx5e_params *params)
+bool mlx5e_rx_is_linear_skb(struct mlx5e_params *params,
+			    struct mlx5e_xsk_param *xsk)
 {
-	u32 frag_sz = mlx5e_rx_get_linear_frag_sz(params, NULL);
+	/* AF_XDP allocates SKBs on XDP_PASS - ensure they don't occupy more
+	 * than one page. For this, check both with and without xsk.
+	 */
+	u32 linear_frag_sz = max(mlx5e_rx_get_linear_frag_sz(params, xsk),
+				 mlx5e_rx_get_linear_frag_sz(params, NULL));
 
-	return !params->lro_en && frag_sz <= PAGE_SIZE;
+	return !params->lro_en && linear_frag_sz <= PAGE_SIZE;
 }
 
 #define MLX5_MAX_MPWQE_LOG_WQE_STRIDE_SZ ((BIT(__mlx5_bit_sz(wq, log_wqe_stride_size)) - 1) + \
 					  MLX5_MPWQE_LOG_STRIDE_SZ_BASE)
 bool mlx5e_rx_mpwqe_is_linear_skb(struct mlx5_core_dev *mdev,
-				  struct mlx5e_params *params)
+				  struct mlx5e_params *params,
+				  struct mlx5e_xsk_param *xsk)
 {
-	u32 frag_sz = mlx5e_rx_get_linear_frag_sz(params, NULL);
+	u32 linear_frag_sz = mlx5e_rx_get_linear_frag_sz(params, xsk);
 	s8 signed_log_num_strides_param;
 	u8 log_num_strides;
 
-	if (!mlx5e_rx_is_linear_skb(params))
+	if (!mlx5e_rx_is_linear_skb(params, xsk))
 		return false;
 
-	if (order_base_2(frag_sz) > MLX5_MAX_MPWQE_LOG_WQE_STRIDE_SZ)
+	if (order_base_2(linear_frag_sz) > MLX5_MAX_MPWQE_LOG_WQE_STRIDE_SZ)
 		return false;
 
 	if (MLX5_CAP_GEN(mdev, ext_stride_num_range))
 		return true;
 
-	log_num_strides = MLX5_MPWRQ_LOG_WQE_SZ - order_base_2(frag_sz);
+	log_num_strides = MLX5_MPWRQ_LOG_WQE_SZ - order_base_2(linear_frag_sz);
 	signed_log_num_strides_param =
 		(s8)log_num_strides - MLX5_MPWQE_LOG_NUM_STRIDES_BASE;
 
 	return signed_log_num_strides_param >= 0;
 }
 
-u8 mlx5e_mpwqe_get_log_rq_size(struct mlx5e_params *params)
+u8 mlx5e_mpwqe_get_log_rq_size(struct mlx5e_params *params,
+			       struct mlx5e_xsk_param *xsk)
 {
-	u8 log_pkts_per_wqe = mlx5e_mpwqe_log_pkts_per_wqe(params);
+	u8 log_pkts_per_wqe = mlx5e_mpwqe_log_pkts_per_wqe(params, xsk);
 
 	/* Numbers are unsigned, don't subtract to avoid underflow. */
 	if (params->log_rq_mtu_frames <
@@ -101,27 +109,30 @@ u8 mlx5e_mpwqe_get_log_rq_size(struct mlx5e_params *params)
 }
 
 u8 mlx5e_mpwqe_get_log_stride_size(struct mlx5_core_dev *mdev,
-				   struct mlx5e_params *params)
+				   struct mlx5e_params *params,
+				   struct mlx5e_xsk_param *xsk)
 {
-	if (mlx5e_rx_mpwqe_is_linear_skb(mdev, params))
-		return order_base_2(mlx5e_rx_get_linear_frag_sz(params, NULL));
+	if (mlx5e_rx_mpwqe_is_linear_skb(mdev, params, xsk))
+		return order_base_2(mlx5e_rx_get_linear_frag_sz(params, xsk));
 
 	return MLX5_MPWRQ_DEF_LOG_STRIDE_SZ(mdev);
 }
 
 u8 mlx5e_mpwqe_get_log_num_strides(struct mlx5_core_dev *mdev,
-				   struct mlx5e_params *params)
+				   struct mlx5e_params *params,
+				   struct mlx5e_xsk_param *xsk)
 {
 	return MLX5_MPWRQ_LOG_WQE_SZ -
-		mlx5e_mpwqe_get_log_stride_size(mdev, params);
+		mlx5e_mpwqe_get_log_stride_size(mdev, params, xsk);
 }
 
 u16 mlx5e_get_rq_headroom(struct mlx5_core_dev *mdev,
-			  struct mlx5e_params *params)
+			  struct mlx5e_params *params,
+			  struct mlx5e_xsk_param *xsk)
 {
 	bool is_linear_skb = (params->rq_wq_type == MLX5_WQ_TYPE_CYCLIC) ?
-		mlx5e_rx_is_linear_skb(params) :
-		mlx5e_rx_mpwqe_is_linear_skb(mdev, params);
+		mlx5e_rx_is_linear_skb(params, xsk) :
+		mlx5e_rx_mpwqe_is_linear_skb(mdev, params, xsk);
 
-	return is_linear_skb ? mlx5e_get_linear_rq_headroom(params, NULL) : 0;
+	return is_linear_skb ? mlx5e_get_linear_rq_headroom(params, xsk) : 0;
 }
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en/params.h b/drivers/net/ethernet/mellanox/mlx5/core/en/params.h
index f83417b822bf..bd882b5ee9a7 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/en/params.h
+++ b/drivers/net/ethernet/mellanox/mlx5/core/en/params.h
@@ -40,22 +40,85 @@ struct mlx5e_channel_param {
 	struct mlx5e_cq_param      icosq_cq;
 };
 
+static inline bool mlx5e_qid_get_ch_if_in_group(struct mlx5e_params *params,
+						u16 qid,
+						enum mlx5e_rq_group group,
+						u16 *ix)
+{
+	int nch = params->num_channels;
+	int ch = qid - nch * group;
+
+	if (ch < 0 || ch >= nch)
+		return false;
+
+	*ix = ch;
+	return true;
+}
+
+static inline void mlx5e_qid_get_ch_and_group(struct mlx5e_params *params,
+					      u16 qid,
+					      u16 *ix,
+					      enum mlx5e_rq_group *group)
+{
+	u16 nch = params->num_channels;
+
+	*ix = qid % nch;
+	*group = qid / nch;
+}
+
+static inline bool mlx5e_qid_validate(struct mlx5e_params *params, u64 qid)
+{
+	return qid < params->num_channels * MLX5E_NUM_RQ_GROUPS;
+}
+
 /* Parameter calculations */
 
 u16 mlx5e_get_linear_rq_headroom(struct mlx5e_params *params,
 				 struct mlx5e_xsk_param *xsk);
 u32 mlx5e_rx_get_linear_frag_sz(struct mlx5e_params *params,
 				struct mlx5e_xsk_param *xsk);
-u8 mlx5e_mpwqe_log_pkts_per_wqe(struct mlx5e_params *params);
-bool mlx5e_rx_is_linear_skb(struct mlx5e_params *params);
+u8 mlx5e_mpwqe_log_pkts_per_wqe(struct mlx5e_params *params,
+				struct mlx5e_xsk_param *xsk);
+bool mlx5e_rx_is_linear_skb(struct mlx5e_params *params,
+			    struct mlx5e_xsk_param *xsk);
 bool mlx5e_rx_mpwqe_is_linear_skb(struct mlx5_core_dev *mdev,
-				  struct mlx5e_params *params);
-u8 mlx5e_mpwqe_get_log_rq_size(struct mlx5e_params *params);
+				  struct mlx5e_params *params,
+				  struct mlx5e_xsk_param *xsk);
+u8 mlx5e_mpwqe_get_log_rq_size(struct mlx5e_params *params,
+			       struct mlx5e_xsk_param *xsk);
 u8 mlx5e_mpwqe_get_log_stride_size(struct mlx5_core_dev *mdev,
-				   struct mlx5e_params *params);
+				   struct mlx5e_params *params,
+				   struct mlx5e_xsk_param *xsk);
 u8 mlx5e_mpwqe_get_log_num_strides(struct mlx5_core_dev *mdev,
-				   struct mlx5e_params *params);
+				   struct mlx5e_params *params,
+				   struct mlx5e_xsk_param *xsk);
 u16 mlx5e_get_rq_headroom(struct mlx5_core_dev *mdev,
-			  struct mlx5e_params *params);
+			  struct mlx5e_params *params,
+			  struct mlx5e_xsk_param *xsk);
+
+/* Build queue parameters */
+
+void mlx5e_build_rq_param(struct mlx5e_priv *priv,
+			  struct mlx5e_params *params,
+			  struct mlx5e_xsk_param *xsk,
+			  struct mlx5e_rq_param *param);
+void mlx5e_build_sq_param_common(struct mlx5e_priv *priv,
+				 struct mlx5e_sq_param *param);
+void mlx5e_build_rx_cq_param(struct mlx5e_priv *priv,
+			     struct mlx5e_params *params,
+			     struct mlx5e_xsk_param *xsk,
+			     struct mlx5e_cq_param *param);
+void mlx5e_build_tx_cq_param(struct mlx5e_priv *priv,
+			     struct mlx5e_params *params,
+			     struct mlx5e_cq_param *param);
+void mlx5e_build_ico_cq_param(struct mlx5e_priv *priv,
+			      u8 log_wq_size,
+			      struct mlx5e_cq_param *param);
+void mlx5e_build_icosq_param(struct mlx5e_priv *priv,
+			     u8 log_wq_size,
+			     struct mlx5e_sq_param *param);
+void mlx5e_build_xdpsq_param(struct mlx5e_priv *priv,
+			     struct mlx5e_params *params,
+			     struct mlx5e_sq_param *param);
 
 #endif /* __MLX5_EN_PARAMS_H__ */
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en/xdp.c b/drivers/net/ethernet/mellanox/mlx5/core/en/xdp.c
index ee99efde9143..b0b982cf69bb 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/en/xdp.c
+++ b/drivers/net/ethernet/mellanox/mlx5/core/en/xdp.c
@@ -31,6 +31,7 @@
  */
 
 #include <linux/bpf_trace.h>
+#include <net/xdp_sock.h>
 #include "en/xdp.h"
 #include "en/params.h"
 
@@ -113,12 +114,12 @@ mlx5e_xmit_xdp_buff(struct mlx5e_xdpsq *sq, struct mlx5e_rq *rq,
 		xdpi.page.di    = *di;
 	}
 
-	return sq->xmit_xdp_frame(sq, &xdptxd, &xdpi);
+	return sq->xmit_xdp_frame(sq, &xdptxd, &xdpi, 0);
 }
 
 /* returns true if packet was consumed by xdp */
 bool mlx5e_xdp_handle(struct mlx5e_rq *rq, struct mlx5e_dma_info *di,
-		      void *va, u16 *rx_headroom, u32 *len)
+		      void *va, u16 *rx_headroom, u32 *len, bool xsk)
 {
 	struct bpf_prog *prog = READ_ONCE(rq->xdp_prog);
 	struct xdp_buff xdp;
@@ -132,9 +133,13 @@ bool mlx5e_xdp_handle(struct mlx5e_rq *rq, struct mlx5e_dma_info *di,
 	xdp_set_data_meta_invalid(&xdp);
 	xdp.data_end = xdp.data + *len;
 	xdp.data_hard_start = va;
+	if (xsk)
+		xdp.handle = di->xsk.handle;
 	xdp.rxq = &rq->xdp_rxq;
 
 	act = bpf_prog_run_xdp(prog, &xdp);
+	if (xsk)
+		xdp.handle += xdp.data - xdp.data_hard_start;
 	switch (act) {
 	case XDP_PASS:
 		*rx_headroom = xdp.data - xdp.data_hard_start;
@@ -152,7 +157,8 @@ bool mlx5e_xdp_handle(struct mlx5e_rq *rq, struct mlx5e_dma_info *di,
 			goto xdp_abort;
 		__set_bit(MLX5E_RQ_FLAG_XDP_XMIT, rq->flags);
 		__set_bit(MLX5E_RQ_FLAG_XDP_REDIRECT, rq->flags);
-		mlx5e_page_dma_unmap(rq, di);
+		if (!xsk)
+			mlx5e_page_dma_unmap(rq, di);
 		rq->stats->xdp_redirect++;
 		return true;
 	default:
@@ -206,7 +212,7 @@ static void mlx5e_xdp_mpwqe_session_start(struct mlx5e_xdpsq *sq)
 	stats->mpwqe++;
 }
 
-static void mlx5e_xdp_mpwqe_complete(struct mlx5e_xdpsq *sq)
+void mlx5e_xdp_mpwqe_complete(struct mlx5e_xdpsq *sq)
 {
 	struct mlx5_wq_cyc       *wq    = &sq->wq;
 	struct mlx5e_xdp_mpwqe *session = &sq->mpwqe;
@@ -229,9 +235,32 @@ static void mlx5e_xdp_mpwqe_complete(struct mlx5e_xdpsq *sq)
 	session->wqe = NULL; /* Close session */
 }
 
+enum {
+	MLX5E_XDP_CHECK_OK = 1,
+	MLX5E_XDP_CHECK_START_MPWQE = 2,
+};
+
+static int mlx5e_xmit_xdp_frame_check_mpwqe(struct mlx5e_xdpsq *sq)
+{
+	if (unlikely(!sq->mpwqe.wqe)) {
+		if (unlikely(!mlx5e_wqc_has_room_for(&sq->wq, sq->cc, sq->pc,
+						     MLX5_SEND_WQE_MAX_WQEBBS))) {
+			/* SQ is full, ring doorbell */
+			mlx5e_xmit_xdp_doorbell(sq);
+			sq->stats->full++;
+			return -EBUSY;
+		}
+
+		return MLX5E_XDP_CHECK_START_MPWQE;
+	}
+
+	return MLX5E_XDP_CHECK_OK;
+}
+
 static bool mlx5e_xmit_xdp_frame_mpwqe(struct mlx5e_xdpsq *sq,
 				       struct mlx5e_xdp_xmit_data *xdptxd,
-				       struct mlx5e_xdp_info *xdpi)
+				       struct mlx5e_xdp_info *xdpi,
+				       int check_result)
 {
 	struct mlx5e_xdp_mpwqe *session = &sq->mpwqe;
 	struct mlx5e_xdpsq_stats *stats = sq->stats;
@@ -241,15 +270,16 @@ static bool mlx5e_xmit_xdp_frame_mpwqe(struct mlx5e_xdpsq *sq,
 		return false;
 	}
 
-	if (unlikely(!session->wqe)) {
-		if (unlikely(!mlx5e_wqc_has_room_for(&sq->wq, sq->cc, sq->pc,
-						     MLX5_SEND_WQE_MAX_WQEBBS))) {
-			/* SQ is full, ring doorbell */
-			mlx5e_xmit_xdp_doorbell(sq);
-			stats->full++;
-			return false;
-		}
+	if (!check_result)
+		check_result = mlx5e_xmit_xdp_frame_check_mpwqe(sq);
+	if (unlikely(check_result < 0))
+		return false;
 
+	if (check_result == MLX5E_XDP_CHECK_START_MPWQE) {
+		/* Start the session when nothing can fail, so it's guaranteed
+		 * that if there is an active session, it has at least one dseg,
+		 * and it's safe to complete it at any time.
+		 */
 		mlx5e_xdp_mpwqe_session_start(sq);
 	}
 
@@ -264,9 +294,22 @@ static bool mlx5e_xmit_xdp_frame_mpwqe(struct mlx5e_xdpsq *sq,
 	return true;
 }
 
+static int mlx5e_xmit_xdp_frame_check(struct mlx5e_xdpsq *sq)
+{
+	if (unlikely(!mlx5e_wqc_has_room_for(&sq->wq, sq->cc, sq->pc, 1))) {
+		/* SQ is full, ring doorbell */
+		mlx5e_xmit_xdp_doorbell(sq);
+		sq->stats->full++;
+		return -EBUSY;
+	}
+
+	return MLX5E_XDP_CHECK_OK;
+}
+
 static bool mlx5e_xmit_xdp_frame(struct mlx5e_xdpsq *sq,
 				 struct mlx5e_xdp_xmit_data *xdptxd,
-				 struct mlx5e_xdp_info *xdpi)
+				 struct mlx5e_xdp_info *xdpi,
+				 int check_result)
 {
 	struct mlx5_wq_cyc       *wq   = &sq->wq;
 	u16                       pi   = mlx5_wq_cyc_ctr2ix(wq, sq->pc);
@@ -288,12 +331,10 @@ static bool mlx5e_xmit_xdp_frame(struct mlx5e_xdpsq *sq,
 		return false;
 	}
 
-	if (unlikely(!mlx5e_wqc_has_room_for(wq, sq->cc, sq->pc, 1))) {
-		/* SQ is full, ring doorbell */
-		mlx5e_xmit_xdp_doorbell(sq);
-		stats->full++;
+	if (!check_result)
+		check_result = mlx5e_xmit_xdp_frame_check(sq);
+	if (unlikely(check_result < 0))
 		return false;
-	}
 
 	cseg->fm_ce_se = 0;
 
@@ -323,6 +364,7 @@ static bool mlx5e_xmit_xdp_frame(struct mlx5e_xdpsq *sq,
 
 static void mlx5e_free_xdpsq_desc(struct mlx5e_xdpsq *sq,
 				  struct mlx5e_xdp_wqe_info *wi,
+				  u32 *xsk_frames,
 				  bool recycle)
 {
 	struct mlx5e_xdp_info_fifo *xdpi_fifo = &sq->db.xdpi_fifo;
@@ -340,7 +382,11 @@ static void mlx5e_free_xdpsq_desc(struct mlx5e_xdpsq *sq,
 			break;
 		case MLX5E_XDP_XMIT_MODE_PAGE:
 			/* XDP_TX from the regular RQ */
-			mlx5e_page_release(xdpi.page.rq, &xdpi.page.di, recycle);
+			mlx5e_page_release_dynamic(xdpi.page.rq, &xdpi.page.di, recycle);
+			break;
+		case MLX5E_XDP_XMIT_MODE_XSK:
+			/* AF_XDP send */
+			(*xsk_frames)++;
 			break;
 		default:
 			WARN_ON_ONCE(true);
@@ -352,6 +398,7 @@ bool mlx5e_poll_xdpsq_cq(struct mlx5e_cq *cq)
 {
 	struct mlx5e_xdpsq *sq;
 	struct mlx5_cqe64 *cqe;
+	u32 xsk_frames = 0;
 	u16 sqcc;
 	int i;
 
@@ -393,10 +440,13 @@ bool mlx5e_poll_xdpsq_cq(struct mlx5e_cq *cq)
 
 			sqcc += wi->num_wqebbs;
 
-			mlx5e_free_xdpsq_desc(sq, wi, true);
+			mlx5e_free_xdpsq_desc(sq, wi, &xsk_frames, true);
 		} while (!last_wqe);
 	} while ((++i < MLX5E_TX_CQ_POLL_BUDGET) && (cqe = mlx5_cqwq_get_cqe(&cq->wq)));
 
+	if (xsk_frames)
+		xsk_umem_complete_tx(sq->umem, xsk_frames);
+
 	sq->stats->cqes += i;
 
 	mlx5_cqwq_update_db_record(&cq->wq);
@@ -410,6 +460,8 @@ bool mlx5e_poll_xdpsq_cq(struct mlx5e_cq *cq)
 
 void mlx5e_free_xdpsq_descs(struct mlx5e_xdpsq *sq)
 {
+	u32 xsk_frames = 0;
+
 	while (sq->cc != sq->pc) {
 		struct mlx5e_xdp_wqe_info *wi;
 		u16 ci;
@@ -419,8 +471,11 @@ void mlx5e_free_xdpsq_descs(struct mlx5e_xdpsq *sq)
 
 		sq->cc += wi->num_wqebbs;
 
-		mlx5e_free_xdpsq_desc(sq, wi, false);
+		mlx5e_free_xdpsq_desc(sq, wi, &xsk_frames, false);
 	}
+
+	if (xsk_frames)
+		xsk_umem_complete_tx(sq->umem, xsk_frames);
 }
 
 int mlx5e_xdp_xmit(struct net_device *dev, int n, struct xdp_frame **frames,
@@ -466,7 +521,7 @@ int mlx5e_xdp_xmit(struct net_device *dev, int n, struct xdp_frame **frames,
 		xdpi.frame.xdpf     = xdpf;
 		xdpi.frame.dma_addr = xdptxd.dma_addr;
 
-		if (unlikely(!sq->xmit_xdp_frame(sq, &xdptxd, &xdpi))) {
+		if (unlikely(!sq->xmit_xdp_frame(sq, &xdptxd, &xdpi, 0))) {
 			dma_unmap_single(sq->pdev, xdptxd.dma_addr,
 					 xdptxd.len, DMA_TO_DEVICE);
 			xdp_return_frame_rx_napi(xdpf);
@@ -500,6 +555,8 @@ void mlx5e_xdp_rx_poll_complete(struct mlx5e_rq *rq)
 
 void mlx5e_set_xmit_fp(struct mlx5e_xdpsq *sq, bool is_mpw)
 {
+	sq->xmit_xdp_frame_check = is_mpw ?
+		mlx5e_xmit_xdp_frame_check_mpwqe : mlx5e_xmit_xdp_frame_check;
 	sq->xmit_xdp_frame = is_mpw ?
 		mlx5e_xmit_xdp_frame_mpwqe : mlx5e_xmit_xdp_frame;
 }
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en/xdp.h b/drivers/net/ethernet/mellanox/mlx5/core/en/xdp.h
index 9200cb9f499b..2d934c8d3807 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/en/xdp.h
+++ b/drivers/net/ethernet/mellanox/mlx5/core/en/xdp.h
@@ -42,7 +42,8 @@
 struct mlx5e_xsk_param;
 int mlx5e_xdp_max_mtu(struct mlx5e_params *params, struct mlx5e_xsk_param *xsk);
 bool mlx5e_xdp_handle(struct mlx5e_rq *rq, struct mlx5e_dma_info *di,
-		      void *va, u16 *rx_headroom, u32 *len);
+		      void *va, u16 *rx_headroom, u32 *len, bool xsk);
+void mlx5e_xdp_mpwqe_complete(struct mlx5e_xdpsq *sq);
 bool mlx5e_poll_xdpsq_cq(struct mlx5e_cq *cq);
 void mlx5e_free_xdpsq_descs(struct mlx5e_xdpsq *sq);
 void mlx5e_set_xmit_fp(struct mlx5e_xdpsq *sq, bool is_mpw);
@@ -67,6 +68,21 @@ static inline bool mlx5e_xdp_tx_is_enabled(struct mlx5e_priv *priv)
 	return test_bit(MLX5E_STATE_XDP_TX_ENABLED, &priv->state);
 }
 
+static inline void mlx5e_xdp_set_open(struct mlx5e_priv *priv)
+{
+	set_bit(MLX5E_STATE_XDP_OPEN, &priv->state);
+}
+
+static inline void mlx5e_xdp_set_closed(struct mlx5e_priv *priv)
+{
+	clear_bit(MLX5E_STATE_XDP_OPEN, &priv->state);
+}
+
+static inline bool mlx5e_xdp_is_open(struct mlx5e_priv *priv)
+{
+	return test_bit(MLX5E_STATE_XDP_OPEN, &priv->state);
+}
+
 static inline void mlx5e_xmit_xdp_doorbell(struct mlx5e_xdpsq *sq)
 {
 	if (sq->doorbell_cseg) {
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en/xsk/Makefile b/drivers/net/ethernet/mellanox/mlx5/core/en/xsk/Makefile
new file mode 100644
index 000000000000..5ee42991900a
--- /dev/null
+++ b/drivers/net/ethernet/mellanox/mlx5/core/en/xsk/Makefile
@@ -0,0 +1 @@
+subdir-ccflags-y += -I$(src)/../..
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en/xsk/rx.c b/drivers/net/ethernet/mellanox/mlx5/core/en/xsk/rx.c
new file mode 100644
index 000000000000..6a55573ec8f2
--- /dev/null
+++ b/drivers/net/ethernet/mellanox/mlx5/core/en/xsk/rx.c
@@ -0,0 +1,192 @@
+// SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
+/* Copyright (c) 2019 Mellanox Technologies. */
+
+#include "rx.h"
+#include "en/xdp.h"
+#include <net/xdp_sock.h>
+
+/* RX data path */
+
+bool mlx5e_xsk_pages_enough_umem(struct mlx5e_rq *rq, int count)
+{
+	/* Check in advance that we have enough frames, instead of allocating
+	 * one-by-one, failing and moving frames to the Reuse Ring.
+	 */
+	return xsk_umem_has_addrs_rq(rq->umem, count);
+}
+
+int mlx5e_xsk_page_alloc_umem(struct mlx5e_rq *rq,
+			      struct mlx5e_dma_info *dma_info)
+{
+	struct xdp_umem *umem = rq->umem;
+	u64 handle;
+
+	if (!xsk_umem_peek_addr_rq(umem, &handle))
+		return -ENOMEM;
+
+	dma_info->xsk.handle = handle + rq->buff.umem_headroom;
+	dma_info->xsk.data = xdp_umem_get_data(umem, dma_info->xsk.handle);
+
+	/* No need to add headroom to the DMA address. In striding RQ case, we
+	 * just provide pages for UMR, and headroom is counted at the setup
+	 * stage when creating a WQE. In non-striding RQ case, headroom is
+	 * accounted in mlx5e_alloc_rx_wqe.
+	 */
+	dma_info->addr = xdp_umem_get_dma(umem, handle);
+
+	xsk_umem_discard_addr_rq(umem);
+
+	dma_sync_single_for_device(rq->pdev, dma_info->addr, PAGE_SIZE,
+				   DMA_BIDIRECTIONAL);
+
+	return 0;
+}
+
+static inline void mlx5e_xsk_recycle_frame(struct mlx5e_rq *rq, u64 handle)
+{
+	xsk_umem_fq_reuse(rq->umem, handle & rq->umem->chunk_mask);
+}
+
+/* XSKRQ uses pages from UMEM, they must not be released. They are returned to
+ * the userspace if possible, and if not, this function is called to reuse them
+ * in the driver.
+ */
+void mlx5e_xsk_page_release(struct mlx5e_rq *rq,
+			    struct mlx5e_dma_info *dma_info)
+{
+	mlx5e_xsk_recycle_frame(rq, dma_info->xsk.handle);
+}
+
+/* Return a frame back to the hardware to fill in again. It is used by XDP when
+ * the XDP program returns XDP_TX or XDP_REDIRECT not to an XSKMAP.
+ */
+void mlx5e_xsk_zca_free(struct zero_copy_allocator *zca, unsigned long handle)
+{
+	struct mlx5e_rq *rq = container_of(zca, struct mlx5e_rq, zca);
+
+	mlx5e_xsk_recycle_frame(rq, handle);
+}
+
+static struct sk_buff *mlx5e_xsk_construct_skb(struct mlx5e_rq *rq, void *data,
+					       u32 cqe_bcnt)
+{
+	struct sk_buff *skb;
+
+	skb = napi_alloc_skb(rq->cq.napi, cqe_bcnt);
+	if (unlikely(!skb)) {
+		rq->stats->buff_alloc_err++;
+		return NULL;
+	}
+
+	skb_put_data(skb, data, cqe_bcnt);
+
+	return skb;
+}
+
+struct sk_buff *mlx5e_xsk_skb_from_cqe_mpwrq_linear(struct mlx5e_rq *rq,
+						    struct mlx5e_mpw_info *wi,
+						    u16 cqe_bcnt,
+						    u32 head_offset,
+						    u32 page_idx)
+{
+	struct mlx5e_dma_info *di = &wi->umr.dma_info[page_idx];
+	u16 rx_headroom = rq->buff.headroom - rq->buff.umem_headroom;
+	u32 cqe_bcnt32 = cqe_bcnt;
+	void *va, *data;
+	u32 frag_size;
+	bool consumed;
+
+	/* Check packet size. Note LRO doesn't use linear SKB */
+	if (unlikely(cqe_bcnt > rq->hw_mtu)) {
+		rq->stats->oversize_pkts_sw_drop++;
+		return NULL;
+	}
+
+	/* head_offset is not used in this function, because di->xsk.data and
+	 * di->addr point directly to the necessary place. Furthermore, in the
+	 * current implementation, one page = one packet = one frame, so
+	 * head_offset should always be 0.
+	 */
+	WARN_ON_ONCE(head_offset);
+
+	va             = di->xsk.data;
+	data           = va + rx_headroom;
+	frag_size      = rq->buff.headroom + cqe_bcnt32;
+
+	dma_sync_single_for_cpu(rq->pdev, di->addr, frag_size, DMA_BIDIRECTIONAL);
+	prefetch(data);
+
+	rcu_read_lock();
+	consumed = mlx5e_xdp_handle(rq, di, va, &rx_headroom, &cqe_bcnt32, true);
+	rcu_read_unlock();
+
+	/* Possible flows:
+	 * - XDP_REDIRECT to XSKMAP:
+	 *   The page is owned by the userspace from now.
+	 * - XDP_TX and other XDP_REDIRECTs:
+	 *   The page was returned by ZCA and recycled.
+	 * - XDP_DROP:
+	 *   Recycle the page.
+	 * - XDP_PASS:
+	 *   Allocate an SKB, copy the data and recycle the page.
+	 *
+	 * Pages to be recycled go to the Reuse Ring on MPWQE deallocation. Its
+	 * size is the same as the Driver RX Ring's size, and pages for WQEs are
+	 * allocated first from the Reuse Ring, so it has enough space.
+	 */
+
+	if (likely(consumed)) {
+		if (likely(__test_and_clear_bit(MLX5E_RQ_FLAG_XDP_XMIT, rq->flags)))
+			__set_bit(page_idx, wi->xdp_xmit_bitmap); /* non-atomic */
+		return NULL; /* page/packet was consumed by XDP */
+	}
+
+	/* XDP_PASS: copy the data from the UMEM to a new SKB and reuse the
+	 * frame. On SKB allocation failure, NULL is returned.
+	 */
+	return mlx5e_xsk_construct_skb(rq, data, cqe_bcnt32);
+}
+
+struct sk_buff *mlx5e_xsk_skb_from_cqe_linear(struct mlx5e_rq *rq,
+					      struct mlx5_cqe64 *cqe,
+					      struct mlx5e_wqe_frag_info *wi,
+					      u32 cqe_bcnt)
+{
+	struct mlx5e_dma_info *di = wi->di;
+	u16 rx_headroom = rq->buff.headroom - rq->buff.umem_headroom;
+	void *va, *data;
+	bool consumed;
+	u32 frag_size;
+
+	/* wi->offset is not used in this function, because di->xsk.data and
+	 * di->addr point directly to the necessary place. Furthermore, in the
+	 * current implementation, one page = one packet = one frame, so
+	 * wi->offset should always be 0.
+	 */
+	WARN_ON_ONCE(wi->offset);
+
+	va             = di->xsk.data;
+	data           = va + rx_headroom;
+	frag_size      = rq->buff.headroom + cqe_bcnt;
+
+	dma_sync_single_for_cpu(rq->pdev, di->addr, frag_size, DMA_BIDIRECTIONAL);
+	prefetch(data);
+
+	if (unlikely(get_cqe_opcode(cqe) != MLX5_CQE_RESP_SEND)) {
+		rq->stats->wqe_err++;
+		return NULL;
+	}
+
+	rcu_read_lock();
+	consumed = mlx5e_xdp_handle(rq, di, va, &rx_headroom, &cqe_bcnt, true);
+	rcu_read_unlock();
+
+	if (likely(consumed))
+		return NULL; /* page/packet was consumed by XDP */
+
+	/* XDP_PASS: copy the data from the UMEM to a new SKB. The frame reuse
+	 * will be handled by mlx5e_put_rx_frag.
+	 * On SKB allocation failure, NULL is returned.
+	 */
+	return mlx5e_xsk_construct_skb(rq, data, cqe_bcnt);
+}
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en/xsk/rx.h b/drivers/net/ethernet/mellanox/mlx5/core/en/xsk/rx.h
new file mode 100644
index 000000000000..307b923a1361
--- /dev/null
+++ b/drivers/net/ethernet/mellanox/mlx5/core/en/xsk/rx.h
@@ -0,0 +1,27 @@
+/* SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB */
+/* Copyright (c) 2019 Mellanox Technologies. */
+
+#ifndef __MLX5_EN_XSK_RX_H__
+#define __MLX5_EN_XSK_RX_H__
+
+#include "en.h"
+
+/* RX data path */
+
+bool mlx5e_xsk_pages_enough_umem(struct mlx5e_rq *rq, int count);
+int mlx5e_xsk_page_alloc_umem(struct mlx5e_rq *rq,
+			      struct mlx5e_dma_info *dma_info);
+void mlx5e_xsk_page_release(struct mlx5e_rq *rq,
+			    struct mlx5e_dma_info *dma_info);
+void mlx5e_xsk_zca_free(struct zero_copy_allocator *zca, unsigned long handle);
+struct sk_buff *mlx5e_xsk_skb_from_cqe_mpwrq_linear(struct mlx5e_rq *rq,
+						    struct mlx5e_mpw_info *wi,
+						    u16 cqe_bcnt,
+						    u32 head_offset,
+						    u32 page_idx);
+struct sk_buff *mlx5e_xsk_skb_from_cqe_linear(struct mlx5e_rq *rq,
+					      struct mlx5_cqe64 *cqe,
+					      struct mlx5e_wqe_frag_info *wi,
+					      u32 cqe_bcnt);
+
+#endif /* __MLX5_EN_XSK_RX_H__ */
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en/xsk/setup.c b/drivers/net/ethernet/mellanox/mlx5/core/en/xsk/setup.c
new file mode 100644
index 000000000000..9b4d47c47c92
--- /dev/null
+++ b/drivers/net/ethernet/mellanox/mlx5/core/en/xsk/setup.c
@@ -0,0 +1,223 @@
+// SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
+/* Copyright (c) 2019 Mellanox Technologies. */
+
+#include "setup.h"
+#include "en/params.h"
+
+bool mlx5e_validate_xsk_param(struct mlx5e_params *params,
+			      struct mlx5e_xsk_param *xsk,
+			      struct mlx5_core_dev *mdev)
+{
+	/* AF_XDP doesn't support frames larger than PAGE_SIZE, and the current
+	 * mlx5e XDP implementation doesn't support multiple packets per page.
+	 */
+	if (xsk->chunk_size != PAGE_SIZE)
+		return false;
+
+	/* Current MTU and XSK headroom don't allow packets to fit the frames. */
+	if (mlx5e_rx_get_linear_frag_sz(params, xsk) > xsk->chunk_size)
+		return false;
+
+	/* frag_sz is different for regular and XSK RQs, so ensure that linear
+	 * SKB mode is possible.
+	 */
+	switch (params->rq_wq_type) {
+	case MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ:
+		return mlx5e_rx_mpwqe_is_linear_skb(mdev, params, xsk);
+	default: /* MLX5_WQ_TYPE_CYCLIC */
+		return mlx5e_rx_is_linear_skb(params, xsk);
+	}
+}
+
+static void mlx5e_build_xskicosq_param(struct mlx5e_priv *priv,
+				       u8 log_wq_size,
+				       struct mlx5e_sq_param *param)
+{
+	void *sqc = param->sqc;
+	void *wq = MLX5_ADDR_OF(sqc, sqc, wq);
+
+	mlx5e_build_sq_param_common(priv, param);
+
+	MLX5_SET(wq, wq, log_wq_sz, log_wq_size);
+}
+
+static void mlx5e_build_xsk_cparam(struct mlx5e_priv *priv,
+				   struct mlx5e_params *params,
+				   struct mlx5e_xsk_param *xsk,
+				   struct mlx5e_channel_param *cparam)
+{
+	const u8 xskicosq_size = MLX5E_PARAMS_MINIMUM_LOG_SQ_SIZE;
+
+	mlx5e_build_rq_param(priv, params, xsk, &cparam->rq);
+	mlx5e_build_xdpsq_param(priv, params, &cparam->xdp_sq);
+	mlx5e_build_xskicosq_param(priv, xskicosq_size, &cparam->icosq);
+	mlx5e_build_rx_cq_param(priv, params, xsk, &cparam->rx_cq);
+	mlx5e_build_tx_cq_param(priv, params, &cparam->tx_cq);
+	mlx5e_build_ico_cq_param(priv, xskicosq_size, &cparam->icosq_cq);
+}
+
+int mlx5e_open_xsk(struct mlx5e_priv *priv, struct mlx5e_params *params,
+		   struct mlx5e_xsk_param *xsk, struct xdp_umem *umem,
+		   struct mlx5e_channel *c)
+{
+	struct mlx5e_channel_param cparam = {};
+	struct net_dim_cq_moder icocq_moder = {};
+	int err;
+
+	if (!mlx5e_validate_xsk_param(params, xsk, priv->mdev))
+		return -EINVAL;
+
+	mlx5e_build_xsk_cparam(priv, params, xsk, &cparam);
+
+	err = mlx5e_open_cq(c, params->rx_cq_moderation, &cparam.rx_cq, &c->xskrq.cq);
+	if (unlikely(err))
+		return err;
+
+	err = mlx5e_open_rq(c, params, &cparam.rq, xsk, umem, &c->xskrq);
+	if (unlikely(err))
+		goto err_close_rx_cq;
+
+	err = mlx5e_open_cq(c, params->tx_cq_moderation, &cparam.tx_cq, &c->xsksq.cq);
+	if (unlikely(err))
+		goto err_close_rq;
+
+	/* Create a separate SQ, so that when the UMEM is disabled, we could
+	 * close this SQ safely and stop receiving CQEs. In other case, e.g., if
+	 * the XDPSQ was used instead, we might run into trouble when the UMEM
+	 * is disabled and then reenabled, but the SQ continues receiving CQEs
+	 * from the old UMEM.
+	 */
+	err = mlx5e_open_xdpsq(c, params, &cparam.xdp_sq, umem, &c->xsksq, true);
+	if (unlikely(err))
+		goto err_close_tx_cq;
+
+	err = mlx5e_open_cq(c, icocq_moder, &cparam.icosq_cq, &c->xskicosq.cq);
+	if (unlikely(err))
+		goto err_close_sq;
+
+	/* Create a dedicated SQ for posting NOPs whenever we need an IRQ to be
+	 * triggered and NAPI to be called on the correct CPU.
+	 */
+	err = mlx5e_open_icosq(c, params, &cparam.icosq, &c->xskicosq);
+	if (unlikely(err))
+		goto err_close_icocq;
+
+	spin_lock_init(&c->xskicosq_lock);
+
+	set_bit(MLX5E_CHANNEL_STATE_XSK, c->state);
+
+	return 0;
+
+err_close_icocq:
+	mlx5e_close_cq(&c->xskicosq.cq);
+
+err_close_sq:
+	mlx5e_close_xdpsq(&c->xsksq);
+
+err_close_tx_cq:
+	mlx5e_close_cq(&c->xsksq.cq);
+
+err_close_rq:
+	mlx5e_close_rq(&c->xskrq);
+
+err_close_rx_cq:
+	mlx5e_close_cq(&c->xskrq.cq);
+
+	return err;
+}
+
+void mlx5e_close_xsk(struct mlx5e_channel *c)
+{
+	clear_bit(MLX5E_CHANNEL_STATE_XSK, c->state);
+	napi_synchronize(&c->napi);
+
+	mlx5e_close_rq(&c->xskrq);
+	mlx5e_close_cq(&c->xskrq.cq);
+	mlx5e_close_icosq(&c->xskicosq);
+	mlx5e_close_cq(&c->xskicosq.cq);
+	mlx5e_close_xdpsq(&c->xsksq);
+	mlx5e_close_cq(&c->xsksq.cq);
+}
+
+void mlx5e_activate_xsk(struct mlx5e_channel *c)
+{
+	set_bit(MLX5E_RQ_STATE_ENABLED, &c->xskrq.state);
+	/* TX queue is created active. */
+	mlx5e_trigger_irq(&c->xskicosq);
+}
+
+void mlx5e_deactivate_xsk(struct mlx5e_channel *c)
+{
+	mlx5e_deactivate_rq(&c->xskrq);
+	/* TX queue is disabled on close. */
+}
+
+static int mlx5e_redirect_xsk_rqt(struct mlx5e_priv *priv, u16 ix, u32 rqn)
+{
+	struct mlx5e_redirect_rqt_param direct_rrp = {
+		.is_rss = false,
+		{
+			.rqn = rqn,
+		},
+	};
+
+	u32 rqtn = priv->xsk_tir[ix].rqt.rqtn;
+
+	return mlx5e_redirect_rqt(priv, rqtn, 1, direct_rrp);
+}
+
+int mlx5e_xsk_redirect_rqt_to_channel(struct mlx5e_priv *priv, struct mlx5e_channel *c)
+{
+	return mlx5e_redirect_xsk_rqt(priv, c->ix, c->xskrq.rqn);
+}
+
+int mlx5e_xsk_redirect_rqt_to_drop(struct mlx5e_priv *priv, u16 ix)
+{
+	return mlx5e_redirect_xsk_rqt(priv, ix, priv->drop_rq.rqn);
+}
+
+int mlx5e_xsk_redirect_rqts_to_channels(struct mlx5e_priv *priv, struct mlx5e_channels *chs)
+{
+	int err, i;
+
+	if (!priv->xsk.refcnt)
+		return 0;
+
+	for (i = 0; i < chs->num; i++) {
+		struct mlx5e_channel *c = chs->c[i];
+
+		if (!test_bit(MLX5E_CHANNEL_STATE_XSK, c->state))
+			continue;
+
+		err = mlx5e_xsk_redirect_rqt_to_channel(priv, c);
+		if (unlikely(err))
+			goto err_stop;
+	}
+
+	return 0;
+
+err_stop:
+	for (i--; i >= 0; i--) {
+		if (!test_bit(MLX5E_CHANNEL_STATE_XSK, chs->c[i]->state))
+			continue;
+
+		mlx5e_xsk_redirect_rqt_to_drop(priv, i);
+	}
+
+	return err;
+}
+
+void mlx5e_xsk_redirect_rqts_to_drop(struct mlx5e_priv *priv, struct mlx5e_channels *chs)
+{
+	int i;
+
+	if (!priv->xsk.refcnt)
+		return;
+
+	for (i = 0; i < chs->num; i++) {
+		if (!test_bit(MLX5E_CHANNEL_STATE_XSK, chs->c[i]->state))
+			continue;
+
+		mlx5e_xsk_redirect_rqt_to_drop(priv, i);
+	}
+}
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en/xsk/setup.h b/drivers/net/ethernet/mellanox/mlx5/core/en/xsk/setup.h
new file mode 100644
index 000000000000..0dd11b81c046
--- /dev/null
+++ b/drivers/net/ethernet/mellanox/mlx5/core/en/xsk/setup.h
@@ -0,0 +1,25 @@
+/* SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB */
+/* Copyright (c) 2019 Mellanox Technologies. */
+
+#ifndef __MLX5_EN_XSK_SETUP_H__
+#define __MLX5_EN_XSK_SETUP_H__
+
+#include "en.h"
+
+struct mlx5e_xsk_param;
+
+bool mlx5e_validate_xsk_param(struct mlx5e_params *params,
+			      struct mlx5e_xsk_param *xsk,
+			      struct mlx5_core_dev *mdev);
+int mlx5e_open_xsk(struct mlx5e_priv *priv, struct mlx5e_params *params,
+		   struct mlx5e_xsk_param *xsk, struct xdp_umem *umem,
+		   struct mlx5e_channel *c);
+void mlx5e_close_xsk(struct mlx5e_channel *c);
+void mlx5e_activate_xsk(struct mlx5e_channel *c);
+void mlx5e_deactivate_xsk(struct mlx5e_channel *c);
+int mlx5e_xsk_redirect_rqt_to_channel(struct mlx5e_priv *priv, struct mlx5e_channel *c);
+int mlx5e_xsk_redirect_rqt_to_drop(struct mlx5e_priv *priv, u16 ix);
+int mlx5e_xsk_redirect_rqts_to_channels(struct mlx5e_priv *priv, struct mlx5e_channels *chs);
+void mlx5e_xsk_redirect_rqts_to_drop(struct mlx5e_priv *priv, struct mlx5e_channels *chs);
+
+#endif /* __MLX5_EN_XSK_SETUP_H__ */
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en/xsk/tx.c b/drivers/net/ethernet/mellanox/mlx5/core/en/xsk/tx.c
new file mode 100644
index 000000000000..35e188cf4ea4
--- /dev/null
+++ b/drivers/net/ethernet/mellanox/mlx5/core/en/xsk/tx.c
@@ -0,0 +1,111 @@
+// SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
+/* Copyright (c) 2019 Mellanox Technologies. */
+
+#include "tx.h"
+#include "umem.h"
+#include "en/xdp.h"
+#include "en/params.h"
+#include <net/xdp_sock.h>
+
+int mlx5e_xsk_async_xmit(struct net_device *dev, u32 qid)
+{
+	struct mlx5e_priv *priv = netdev_priv(dev);
+	struct mlx5e_params *params = &priv->channels.params;
+	struct mlx5e_channel *c;
+	u16 ix;
+
+	if (unlikely(!mlx5e_xdp_is_open(priv)))
+		return -ENETDOWN;
+
+	if (unlikely(!mlx5e_qid_get_ch_if_in_group(params, qid, MLX5E_RQ_GROUP_XSK, &ix)))
+		return -EINVAL;
+
+	c = priv->channels.c[ix];
+
+	if (unlikely(!test_bit(MLX5E_CHANNEL_STATE_XSK, c->state)))
+		return -ENXIO;
+
+	if (!napi_if_scheduled_mark_missed(&c->napi)) {
+		spin_lock(&c->xskicosq_lock);
+		mlx5e_trigger_irq(&c->xskicosq);
+		spin_unlock(&c->xskicosq_lock);
+	}
+
+	return 0;
+}
+
+/* When TX fails (because of the size of the packet), we need to get completions
+ * in order, so post a NOP to get a CQE. Since AF_XDP doesn't distinguish
+ * between successful TX and errors, handling in mlx5e_poll_xdpsq_cq is the
+ * same.
+ */
+static void mlx5e_xsk_tx_post_err(struct mlx5e_xdpsq *sq,
+				  struct mlx5e_xdp_info *xdpi)
+{
+	u16 pi = mlx5_wq_cyc_ctr2ix(&sq->wq, sq->pc);
+	struct mlx5e_xdp_wqe_info *wi = &sq->db.wqe_info[pi];
+	struct mlx5e_tx_wqe *nopwqe;
+
+	wi->num_wqebbs = 1;
+	wi->num_pkts = 1;
+
+	nopwqe = mlx5e_post_nop(&sq->wq, sq->sqn, &sq->pc);
+	mlx5e_xdpi_fifo_push(&sq->db.xdpi_fifo, xdpi);
+	sq->doorbell_cseg = &nopwqe->ctrl;
+}
+
+bool mlx5e_xsk_tx(struct mlx5e_xdpsq *sq, unsigned int budget)
+{
+	struct xdp_umem *umem = sq->umem;
+	struct mlx5e_xdp_info xdpi;
+	struct mlx5e_xdp_xmit_data xdptxd;
+	bool work_done = true;
+	bool flush = false;
+
+	xdpi.mode = MLX5E_XDP_XMIT_MODE_XSK;
+
+	for (; budget; budget--) {
+		int check_result = sq->xmit_xdp_frame_check(sq);
+		struct xdp_desc desc;
+
+		if (unlikely(check_result < 0)) {
+			work_done = false;
+			break;
+		}
+
+		if (!xsk_umem_consume_tx(umem, &desc)) {
+			/* TX will get stuck until something wakes it up by
+			 * triggering NAPI. Currently it's expected that the
+			 * application calls sendto() if there are consumed, but
+			 * not completed frames.
+			 */
+			break;
+		}
+
+		xdptxd.dma_addr = xdp_umem_get_dma(umem, desc.addr);
+		xdptxd.data = xdp_umem_get_data(umem, desc.addr);
+		xdptxd.len = desc.len;
+
+		dma_sync_single_for_device(sq->pdev, xdptxd.dma_addr,
+					   xdptxd.len, DMA_BIDIRECTIONAL);
+
+		if (unlikely(!sq->xmit_xdp_frame(sq, &xdptxd, &xdpi, check_result))) {
+			if (sq->mpwqe.wqe)
+				mlx5e_xdp_mpwqe_complete(sq);
+
+			mlx5e_xsk_tx_post_err(sq, &xdpi);
+		}
+
+		flush = true;
+	}
+
+	if (flush) {
+		if (sq->mpwqe.wqe)
+			mlx5e_xdp_mpwqe_complete(sq);
+		mlx5e_xmit_xdp_doorbell(sq);
+
+		xsk_umem_consume_tx_done(umem);
+	}
+
+	return !(budget && work_done);
+}
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en/xsk/tx.h b/drivers/net/ethernet/mellanox/mlx5/core/en/xsk/tx.h
new file mode 100644
index 000000000000..7add18bf78d8
--- /dev/null
+++ b/drivers/net/ethernet/mellanox/mlx5/core/en/xsk/tx.h
@@ -0,0 +1,15 @@
+/* SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB */
+/* Copyright (c) 2019 Mellanox Technologies. */
+
+#ifndef __MLX5_EN_XSK_TX_H__
+#define __MLX5_EN_XSK_TX_H__
+
+#include "en.h"
+
+/* TX data path */
+
+int mlx5e_xsk_async_xmit(struct net_device *dev, u32 qid);
+
+bool mlx5e_xsk_tx(struct mlx5e_xdpsq *sq, unsigned int budget);
+
+#endif /* __MLX5_EN_XSK_TX_H__ */
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en/xsk/umem.c b/drivers/net/ethernet/mellanox/mlx5/core/en/xsk/umem.c
new file mode 100644
index 000000000000..4baaa5788320
--- /dev/null
+++ b/drivers/net/ethernet/mellanox/mlx5/core/en/xsk/umem.c
@@ -0,0 +1,267 @@
+// SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
+/* Copyright (c) 2019 Mellanox Technologies. */
+
+#include <net/xdp_sock.h>
+#include "umem.h"
+#include "setup.h"
+#include "en/params.h"
+
+static int mlx5e_xsk_map_umem(struct mlx5e_priv *priv,
+			      struct xdp_umem *umem)
+{
+	struct device *dev = priv->mdev->device;
+	u32 i;
+
+	for (i = 0; i < umem->npgs; i++) {
+		dma_addr_t dma = dma_map_page(dev, umem->pgs[i], 0, PAGE_SIZE,
+					      DMA_BIDIRECTIONAL);
+
+		if (unlikely(dma_mapping_error(dev, dma)))
+			goto err_unmap;
+		umem->pages[i].dma = dma;
+	}
+
+	return 0;
+
+err_unmap:
+	while (i--) {
+		dma_unmap_page(dev, umem->pages[i].dma, PAGE_SIZE,
+			       DMA_BIDIRECTIONAL);
+		umem->pages[i].dma = 0;
+	}
+
+	return -ENOMEM;
+}
+
+static void mlx5e_xsk_unmap_umem(struct mlx5e_priv *priv,
+				 struct xdp_umem *umem)
+{
+	struct device *dev = priv->mdev->device;
+	u32 i;
+
+	for (i = 0; i < umem->npgs; i++) {
+		dma_unmap_page(dev, umem->pages[i].dma, PAGE_SIZE,
+			       DMA_BIDIRECTIONAL);
+		umem->pages[i].dma = 0;
+	}
+}
+
+static int mlx5e_xsk_get_umems(struct mlx5e_xsk *xsk)
+{
+	if (!xsk->umems) {
+		xsk->umems = kcalloc(MLX5E_MAX_NUM_CHANNELS,
+				     sizeof(*xsk->umems), GFP_KERNEL);
+		if (unlikely(!xsk->umems))
+			return -ENOMEM;
+	}
+
+	xsk->refcnt++;
+	xsk->ever_used = true;
+
+	return 0;
+}
+
+static void mlx5e_xsk_put_umems(struct mlx5e_xsk *xsk)
+{
+	if (!--xsk->refcnt) {
+		kfree(xsk->umems);
+		xsk->umems = NULL;
+	}
+}
+
+static int mlx5e_xsk_add_umem(struct mlx5e_xsk *xsk, struct xdp_umem *umem, u16 ix)
+{
+	int err;
+
+	err = mlx5e_xsk_get_umems(xsk);
+	if (unlikely(err))
+		return err;
+
+	xsk->umems[ix] = umem;
+	return 0;
+}
+
+static void mlx5e_xsk_remove_umem(struct mlx5e_xsk *xsk, u16 ix)
+{
+	xsk->umems[ix] = NULL;
+
+	mlx5e_xsk_put_umems(xsk);
+}
+
+static bool mlx5e_xsk_is_umem_sane(struct xdp_umem *umem)
+{
+	return umem->headroom <= 0xffff && umem->chunk_size_nohr <= 0xffff;
+}
+
+void mlx5e_build_xsk_param(struct xdp_umem *umem, struct mlx5e_xsk_param *xsk)
+{
+	xsk->headroom = umem->headroom;
+	xsk->chunk_size = umem->chunk_size_nohr + umem->headroom;
+}
+
+static int mlx5e_xsk_enable_locked(struct mlx5e_priv *priv,
+				   struct xdp_umem *umem, u16 ix)
+{
+	struct mlx5e_params *params = &priv->channels.params;
+	struct mlx5e_xsk_param xsk;
+	struct mlx5e_channel *c;
+	int err;
+
+	if (unlikely(mlx5e_xsk_get_umem(&priv->channels.params, &priv->xsk, ix)))
+		return -EBUSY;
+
+	if (unlikely(!mlx5e_xsk_is_umem_sane(umem)))
+		return -EINVAL;
+
+	err = mlx5e_xsk_map_umem(priv, umem);
+	if (unlikely(err))
+		return err;
+
+	err = mlx5e_xsk_add_umem(&priv->xsk, umem, ix);
+	if (unlikely(err))
+		goto err_unmap_umem;
+
+	mlx5e_build_xsk_param(umem, &xsk);
+
+	if (!test_bit(MLX5E_STATE_OPENED, &priv->state)) {
+		/* XSK objects will be created on open. */
+		goto validate_closed;
+	}
+
+	if (!params->xdp_prog) {
+		/* XSK objects will be created when an XDP program is set,
+		 * and the channels are reopened.
+		 */
+		goto validate_closed;
+	}
+
+	c = priv->channels.c[ix];
+
+	err = mlx5e_open_xsk(priv, params, &xsk, umem, c);
+	if (unlikely(err))
+		goto err_remove_umem;
+
+	mlx5e_activate_xsk(c);
+
+	/* Don't wait for WQEs, because the newer xdpsock sample doesn't provide
+	 * any Fill Ring entries at the setup stage.
+	 */
+
+	err = mlx5e_xsk_redirect_rqt_to_channel(priv, priv->channels.c[ix]);
+	if (unlikely(err))
+		goto err_deactivate;
+
+	return 0;
+
+err_deactivate:
+	mlx5e_deactivate_xsk(c);
+	mlx5e_close_xsk(c);
+
+err_remove_umem:
+	mlx5e_xsk_remove_umem(&priv->xsk, ix);
+
+err_unmap_umem:
+	mlx5e_xsk_unmap_umem(priv, umem);
+
+	return err;
+
+validate_closed:
+	/* Check the configuration in advance, rather than fail at a later stage
+	 * (in mlx5e_xdp_set or on open) and end up with no channels.
+	 */
+	if (!mlx5e_validate_xsk_param(params, &xsk, priv->mdev)) {
+		err = -EINVAL;
+		goto err_remove_umem;
+	}
+
+	return 0;
+}
+
+static int mlx5e_xsk_disable_locked(struct mlx5e_priv *priv, u16 ix)
+{
+	struct xdp_umem *umem = mlx5e_xsk_get_umem(&priv->channels.params,
+						   &priv->xsk, ix);
+	struct mlx5e_channel *c;
+
+	if (unlikely(!umem))
+		return -EINVAL;
+
+	if (!test_bit(MLX5E_STATE_OPENED, &priv->state))
+		goto remove_umem;
+
+	/* XSK RQ and SQ are only created if XDP program is set. */
+	if (!priv->channels.params.xdp_prog)
+		goto remove_umem;
+
+	c = priv->channels.c[ix];
+	mlx5e_xsk_redirect_rqt_to_drop(priv, ix);
+	mlx5e_deactivate_xsk(c);
+	mlx5e_close_xsk(c);
+
+remove_umem:
+	mlx5e_xsk_remove_umem(&priv->xsk, ix);
+	mlx5e_xsk_unmap_umem(priv, umem);
+
+	return 0;
+}
+
+static int mlx5e_xsk_enable_umem(struct mlx5e_priv *priv, struct xdp_umem *umem,
+				 u16 ix)
+{
+	int err;
+
+	mutex_lock(&priv->state_lock);
+	err = mlx5e_xsk_enable_locked(priv, umem, ix);
+	mutex_unlock(&priv->state_lock);
+
+	return err;
+}
+
+static int mlx5e_xsk_disable_umem(struct mlx5e_priv *priv, u16 ix)
+{
+	int err;
+
+	mutex_lock(&priv->state_lock);
+	err = mlx5e_xsk_disable_locked(priv, ix);
+	mutex_unlock(&priv->state_lock);
+
+	return err;
+}
+
+int mlx5e_xsk_setup_umem(struct net_device *dev, struct xdp_umem *umem, u16 qid)
+{
+	struct mlx5e_priv *priv = netdev_priv(dev);
+	struct mlx5e_params *params = &priv->channels.params;
+	u16 ix;
+
+	if (unlikely(!mlx5e_qid_get_ch_if_in_group(params, qid, MLX5E_RQ_GROUP_XSK, &ix)))
+		return -EINVAL;
+
+	return umem ? mlx5e_xsk_enable_umem(priv, umem, ix) :
+		      mlx5e_xsk_disable_umem(priv, ix);
+}
+
+int mlx5e_xsk_resize_reuseq(struct xdp_umem *umem, u32 nentries)
+{
+	struct xdp_umem_fq_reuse *reuseq;
+
+	reuseq = xsk_reuseq_prepare(nentries);
+	if (unlikely(!reuseq))
+		return -ENOMEM;
+	xsk_reuseq_free(xsk_reuseq_swap(umem, reuseq));
+
+	return 0;
+}
+
+u16 mlx5e_xsk_first_unused_channel(struct mlx5e_params *params, struct mlx5e_xsk *xsk)
+{
+	u16 res = xsk->refcnt ? params->num_channels : 0;
+
+	while (res) {
+		if (mlx5e_xsk_get_umem(params, xsk, res - 1))
+			break;
+		--res;
+	}
+
+	return res;
+}
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en/xsk/umem.h b/drivers/net/ethernet/mellanox/mlx5/core/en/xsk/umem.h
new file mode 100644
index 000000000000..25b4cbe58b54
--- /dev/null
+++ b/drivers/net/ethernet/mellanox/mlx5/core/en/xsk/umem.h
@@ -0,0 +1,31 @@
+/* SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB */
+/* Copyright (c) 2019 Mellanox Technologies. */
+
+#ifndef __MLX5_EN_XSK_UMEM_H__
+#define __MLX5_EN_XSK_UMEM_H__
+
+#include "en.h"
+
+static inline struct xdp_umem *mlx5e_xsk_get_umem(struct mlx5e_params *params,
+						  struct mlx5e_xsk *xsk, u16 ix)
+{
+	if (!xsk || !xsk->umems)
+		return NULL;
+
+	if (unlikely(ix >= params->num_channels))
+		return NULL;
+
+	return xsk->umems[ix];
+}
+
+struct mlx5e_xsk_param;
+void mlx5e_build_xsk_param(struct xdp_umem *umem, struct mlx5e_xsk_param *xsk);
+
+/* .ndo_bpf callback. */
+int mlx5e_xsk_setup_umem(struct net_device *dev, struct xdp_umem *umem, u16 qid);
+
+int mlx5e_xsk_resize_reuseq(struct xdp_umem *umem, u32 nentries);
+
+u16 mlx5e_xsk_first_unused_channel(struct mlx5e_params *params, struct mlx5e_xsk *xsk);
+
+#endif /* __MLX5_EN_XSK_UMEM_H__ */