Merge branch 'support-tunnel-mode-in-mlx5-ipsec-packet-offload'

Leon Romanovsky says:

====================
Support tunnel mode in mlx5 IPsec packet offload

This series extends mlx5 to support tunnel mode in its IPsec packet
offload implementation.

v0: https://lore.kernel.org/all/cover.1681106636.git.leonro@nvidia.com
====================

Link: https://lore.kernel.org/r/cover.1681388425.git.leonro@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>

7 files changed, 481 insertions, 47 deletions

diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en_accel/ipsec.c b/drivers/net/ethernet/mellanox/mlx5/core/en_accel/ipsec.c
index def01bfde610..5fd609d1120e 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/en_accel/ipsec.c
+++ b/drivers/net/ethernet/mellanox/mlx5/core/en_accel/ipsec.c
@@ -35,12 +35,14 @@
 #include <crypto/aead.h>
 #include <linux/inetdevice.h>
 #include <linux/netdevice.h>
+#include <net/netevent.h>
 
 #include "en.h"
 #include "ipsec.h"
 #include "ipsec_rxtx.h"
 
 #define MLX5_IPSEC_RESCHED msecs_to_jiffies(1000)
+#define MLX5E_IPSEC_TUNNEL_SA XA_MARK_1
 
 static struct mlx5e_ipsec_sa_entry *to_ipsec_sa_entry(struct xfrm_state *x)
 {
@@ -242,6 +244,57 @@ static void mlx5e_ipsec_init_limits(struct mlx5e_ipsec_sa_entry *sa_entry,
 	attrs->lft.numb_rounds_soft = (u64)n;
 }
 
+static void mlx5e_ipsec_init_macs(struct mlx5e_ipsec_sa_entry *sa_entry,
+				  struct mlx5_accel_esp_xfrm_attrs *attrs)
+{
+	struct mlx5_core_dev *mdev = mlx5e_ipsec_sa2dev(sa_entry);
+	struct xfrm_state *x = sa_entry->x;
+	struct net_device *netdev;
+	struct neighbour *n;
+	u8 addr[ETH_ALEN];
+
+	if (attrs->mode != XFRM_MODE_TUNNEL ||
+	    attrs->type != XFRM_DEV_OFFLOAD_PACKET)
+		return;
+
+	netdev = x->xso.real_dev;
+
+	mlx5_query_mac_address(mdev, addr);
+	switch (attrs->dir) {
+	case XFRM_DEV_OFFLOAD_IN:
+		ether_addr_copy(attrs->dmac, addr);
+		n = neigh_lookup(&arp_tbl, &attrs->saddr.a4, netdev);
+		if (!n) {
+			n = neigh_create(&arp_tbl, &attrs->saddr.a4, netdev);
+			if (IS_ERR(n))
+				return;
+			neigh_event_send(n, NULL);
+			attrs->drop = true;
+			break;
+		}
+		neigh_ha_snapshot(addr, n, netdev);
+		ether_addr_copy(attrs->smac, addr);
+		break;
+	case XFRM_DEV_OFFLOAD_OUT:
+		ether_addr_copy(attrs->smac, addr);
+		n = neigh_lookup(&arp_tbl, &attrs->daddr.a4, netdev);
+		if (!n) {
+			n = neigh_create(&arp_tbl, &attrs->daddr.a4, netdev);
+			if (IS_ERR(n))
+				return;
+			neigh_event_send(n, NULL);
+			attrs->drop = true;
+			break;
+		}
+		neigh_ha_snapshot(addr, n, netdev);
+		ether_addr_copy(attrs->dmac, addr);
+		break;
+	default:
+		return;
+	}
+	neigh_release(n);
+}
+
 void mlx5e_ipsec_build_accel_xfrm_attrs(struct mlx5e_ipsec_sa_entry *sa_entry,
 					struct mlx5_accel_esp_xfrm_attrs *attrs)
 {
@@ -297,8 +350,10 @@ void mlx5e_ipsec_build_accel_xfrm_attrs(struct mlx5e_ipsec_sa_entry *sa_entry,
 	attrs->upspec.sport = ntohs(x->sel.sport);
 	attrs->upspec.sport_mask = ntohs(x->sel.sport_mask);
 	attrs->upspec.proto = x->sel.proto;
+	attrs->mode = x->props.mode;
 
 	mlx5e_ipsec_init_limits(sa_entry, attrs);
+	mlx5e_ipsec_init_macs(sa_entry, attrs);
 }
 
 static int mlx5e_xfrm_validate_state(struct mlx5_core_dev *mdev,
@@ -367,6 +422,11 @@ static int mlx5e_xfrm_validate_state(struct mlx5_core_dev *mdev,
 		return -EINVAL;
 	}
 
+	if (x->props.mode != XFRM_MODE_TRANSPORT && x->props.mode != XFRM_MODE_TUNNEL) {
+		NL_SET_ERR_MSG_MOD(extack, "Only transport and tunnel xfrm states may be offloaded");
+		return -EINVAL;
+	}
+
 	switch (x->xso.type) {
 	case XFRM_DEV_OFFLOAD_CRYPTO:
 		if (!(mlx5_ipsec_device_caps(mdev) & MLX5_IPSEC_CAP_CRYPTO)) {
@@ -374,11 +434,6 @@ static int mlx5e_xfrm_validate_state(struct mlx5_core_dev *mdev,
 			return -EINVAL;
 		}
 
-		if (x->props.mode != XFRM_MODE_TRANSPORT &&
-		    x->props.mode != XFRM_MODE_TUNNEL) {
-			NL_SET_ERR_MSG_MOD(extack, "Only transport and tunnel xfrm states may be offloaded");
-			return -EINVAL;
-		}
 		break;
 	case XFRM_DEV_OFFLOAD_PACKET:
 		if (!(mlx5_ipsec_device_caps(mdev) &
@@ -387,8 +442,9 @@ static int mlx5e_xfrm_validate_state(struct mlx5_core_dev *mdev,
 			return -EINVAL;
 		}
 
-		if (x->props.mode != XFRM_MODE_TRANSPORT) {
-			NL_SET_ERR_MSG_MOD(extack, "Only transport xfrm states may be offloaded in packet mode");
+		if (x->props.mode == XFRM_MODE_TUNNEL &&
+		    !(mlx5_ipsec_device_caps(mdev) & MLX5_IPSEC_CAP_TUNNEL)) {
+			NL_SET_ERR_MSG_MOD(extack, "Packet offload is not supported for tunnel mode");
 			return -EINVAL;
 		}
 
@@ -458,34 +514,81 @@ static void mlx5e_ipsec_set_esn_ops(struct mlx5e_ipsec_sa_entry *sa_entry)
 	sa_entry->set_iv_op = mlx5e_ipsec_set_iv;
 }
 
+static void mlx5e_ipsec_handle_netdev_event(struct work_struct *_work)
+{
+	struct mlx5e_ipsec_work *work =
+		container_of(_work, struct mlx5e_ipsec_work, work);
+	struct mlx5e_ipsec_sa_entry *sa_entry = work->sa_entry;
+	struct mlx5e_ipsec_netevent_data *data = work->data;
+	struct mlx5_accel_esp_xfrm_attrs *attrs;
+
+	attrs = &sa_entry->attrs;
+
+	switch (attrs->dir) {
+	case XFRM_DEV_OFFLOAD_IN:
+		ether_addr_copy(attrs->smac, data->addr);
+		break;
+	case XFRM_DEV_OFFLOAD_OUT:
+		ether_addr_copy(attrs->dmac, data->addr);
+		break;
+	default:
+		WARN_ON_ONCE(true);
+	}
+	attrs->drop = false;
+	mlx5e_accel_ipsec_fs_modify(sa_entry);
+}
+
 static int mlx5_ipsec_create_work(struct mlx5e_ipsec_sa_entry *sa_entry)
 {
 	struct xfrm_state *x = sa_entry->x;
 	struct mlx5e_ipsec_work *work;
+	void *data = NULL;
 
 	switch (x->xso.type) {
 	case XFRM_DEV_OFFLOAD_CRYPTO:
 		if (!(x->props.flags & XFRM_STATE_ESN))
 			return 0;
 		break;
+	case XFRM_DEV_OFFLOAD_PACKET:
+		if (x->props.mode != XFRM_MODE_TUNNEL)
+			return 0;
+		break;
 	default:
-		return 0;
+		break;
 	}
 
 	work = kzalloc(sizeof(*work), GFP_KERNEL);
 	if (!work)
 		return -ENOMEM;
 
-	work->data = kzalloc(sizeof(*sa_entry), GFP_KERNEL);
-	if (!work->data) {
-		kfree(work);
-		return -ENOMEM;
+	switch (x->xso.type) {
+	case XFRM_DEV_OFFLOAD_CRYPTO:
+		data = kzalloc(sizeof(*sa_entry), GFP_KERNEL);
+		if (!data)
+			goto free_work;
+
+		INIT_WORK(&work->work, mlx5e_ipsec_modify_state);
+		break;
+	case XFRM_DEV_OFFLOAD_PACKET:
+		data = kzalloc(sizeof(struct mlx5e_ipsec_netevent_data),
+			       GFP_KERNEL);
+		if (!data)
+			goto free_work;
+
+		INIT_WORK(&work->work, mlx5e_ipsec_handle_netdev_event);
+		break;
+	default:
+		break;
 	}
 
-	INIT_WORK(&work->work, mlx5e_ipsec_modify_state);
+	work->data = data;
 	work->sa_entry = sa_entry;
 	sa_entry->work = work;
 	return 0;
+
+free_work:
+	kfree(work);
+	return -ENOMEM;
 }
 
 static int mlx5e_ipsec_create_dwork(struct mlx5e_ipsec_sa_entry *sa_entry)
@@ -566,6 +669,14 @@ static int mlx5e_xfrm_add_state(struct xfrm_state *x,
 	if (err)
 		goto err_hw_ctx;
 
+	if (x->props.mode == XFRM_MODE_TUNNEL &&
+	    x->xso.type == XFRM_DEV_OFFLOAD_PACKET &&
+	    !mlx5e_ipsec_fs_tunnel_enabled(sa_entry)) {
+		NL_SET_ERR_MSG_MOD(extack, "Packet offload tunnel mode is disabled due to encap settings");
+		err = -EINVAL;
+		goto err_add_rule;
+	}
+
 	/* We use *_bh() variant because xfrm_timer_handler(), which runs
 	 * in softirq context, can reach our state delete logic and we need
 	 * xa_erase_bh() there.
@@ -580,6 +691,12 @@ static int mlx5e_xfrm_add_state(struct xfrm_state *x,
 	if (sa_entry->dwork)
 		queue_delayed_work(ipsec->wq, &sa_entry->dwork->dwork,
 				   MLX5_IPSEC_RESCHED);
+
+	if (x->xso.type == XFRM_DEV_OFFLOAD_PACKET &&
+	    x->props.mode == XFRM_MODE_TUNNEL)
+		xa_set_mark(&ipsec->sadb, sa_entry->ipsec_obj_id,
+			    MLX5E_IPSEC_TUNNEL_SA);
+
 out:
 	x->xso.offload_handle = (unsigned long)sa_entry;
 	return 0;
@@ -602,6 +719,7 @@ err_xfrm:
 static void mlx5e_xfrm_del_state(struct xfrm_state *x)
 {
 	struct mlx5e_ipsec_sa_entry *sa_entry = to_ipsec_sa_entry(x);
+	struct mlx5_accel_esp_xfrm_attrs *attrs = &sa_entry->attrs;
 	struct mlx5e_ipsec *ipsec = sa_entry->ipsec;
 	struct mlx5e_ipsec_sa_entry *old;
 
@@ -610,6 +728,12 @@ static void mlx5e_xfrm_del_state(struct xfrm_state *x)
 
 	old = xa_erase_bh(&ipsec->sadb, sa_entry->ipsec_obj_id);
 	WARN_ON(old != sa_entry);
+
+	if (attrs->mode == XFRM_MODE_TUNNEL &&
+	    attrs->type == XFRM_DEV_OFFLOAD_PACKET)
+		/* Make sure that no ARP requests are running in parallel */
+		flush_workqueue(ipsec->wq);
+
 }
 
 static void mlx5e_xfrm_free_state(struct xfrm_state *x)
@@ -634,6 +758,46 @@ sa_entry_free:
 	kfree(sa_entry);
 }
 
+static int mlx5e_ipsec_netevent_event(struct notifier_block *nb,
+				      unsigned long event, void *ptr)
+{
+	struct mlx5_accel_esp_xfrm_attrs *attrs;
+	struct mlx5e_ipsec_netevent_data *data;
+	struct mlx5e_ipsec_sa_entry *sa_entry;
+	struct mlx5e_ipsec *ipsec;
+	struct neighbour *n = ptr;
+	struct net_device *netdev;
+	struct xfrm_state *x;
+	unsigned long idx;
+
+	if (event != NETEVENT_NEIGH_UPDATE || !(n->nud_state & NUD_VALID))
+		return NOTIFY_DONE;
+
+	ipsec = container_of(nb, struct mlx5e_ipsec, netevent_nb);
+	xa_for_each_marked(&ipsec->sadb, idx, sa_entry, MLX5E_IPSEC_TUNNEL_SA) {
+		attrs = &sa_entry->attrs;
+
+		if (attrs->family == AF_INET) {
+			if (!neigh_key_eq32(n, &attrs->saddr.a4) &&
+			    !neigh_key_eq32(n, &attrs->daddr.a4))
+				continue;
+		} else {
+			if (!neigh_key_eq128(n, &attrs->saddr.a4) &&
+			    !neigh_key_eq128(n, &attrs->daddr.a4))
+				continue;
+		}
+
+		x = sa_entry->x;
+		netdev = x->xso.real_dev;
+		data = sa_entry->work->data;
+
+		neigh_ha_snapshot(data->addr, n, netdev);
+		queue_work(ipsec->wq, &sa_entry->work->work);
+	}
+
+	return NOTIFY_DONE;
+}
+
 void mlx5e_ipsec_init(struct mlx5e_priv *priv)
 {
 	struct mlx5e_ipsec *ipsec;
@@ -662,6 +826,13 @@ void mlx5e_ipsec_init(struct mlx5e_priv *priv)
 			goto err_aso;
 	}
 
+	if (mlx5_ipsec_device_caps(priv->mdev) & MLX5_IPSEC_CAP_TUNNEL) {
+		ipsec->netevent_nb.notifier_call = mlx5e_ipsec_netevent_event;
+		ret = register_netevent_notifier(&ipsec->netevent_nb);
+		if (ret)
+			goto clear_aso;
+	}
+
 	ret = mlx5e_accel_ipsec_fs_init(ipsec);
 	if (ret)
 		goto err_fs_init;
@@ -672,6 +843,9 @@ void mlx5e_ipsec_init(struct mlx5e_priv *priv)
 	return;
 
 err_fs_init:
+	if (mlx5_ipsec_device_caps(priv->mdev) & MLX5_IPSEC_CAP_TUNNEL)
+		unregister_netevent_notifier(&ipsec->netevent_nb);
+clear_aso:
 	if (mlx5_ipsec_device_caps(priv->mdev) & MLX5_IPSEC_CAP_PACKET_OFFLOAD)
 		mlx5e_ipsec_aso_cleanup(ipsec);
 err_aso:
@@ -690,6 +864,8 @@ void mlx5e_ipsec_cleanup(struct mlx5e_priv *priv)
 		return;
 
 	mlx5e_accel_ipsec_fs_cleanup(ipsec);
+	if (mlx5_ipsec_device_caps(priv->mdev) & MLX5_IPSEC_CAP_TUNNEL)
+		unregister_netevent_notifier(&ipsec->netevent_nb);
 	if (mlx5_ipsec_device_caps(priv->mdev) & MLX5_IPSEC_CAP_PACKET_OFFLOAD)
 		mlx5e_ipsec_aso_cleanup(ipsec);
 	destroy_workqueue(ipsec->wq);
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en_accel/ipsec.h b/drivers/net/ethernet/mellanox/mlx5/core/en_accel/ipsec.h
index 52890d7dce6b..f7f7c09d2b32 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/en_accel/ipsec.h
+++ b/drivers/net/ethernet/mellanox/mlx5/core/en_accel/ipsec.h
@@ -77,7 +77,7 @@ struct mlx5_replay_esn {
 
 struct mlx5_accel_esp_xfrm_attrs {
 	u32   spi;
-	u32   flags;
+	u32   mode;
 	struct aes_gcm_keymat aes_gcm;
 
 	union {
@@ -99,6 +99,8 @@ struct mlx5_accel_esp_xfrm_attrs {
 	u32 authsize;
 	u32 reqid;
 	struct mlx5_ipsec_lft lft;
+	u8 smac[ETH_ALEN];
+	u8 dmac[ETH_ALEN];
 };
 
 enum mlx5_ipsec_cap {
@@ -107,6 +109,7 @@ enum mlx5_ipsec_cap {
 	MLX5_IPSEC_CAP_PACKET_OFFLOAD	= 1 << 2,
 	MLX5_IPSEC_CAP_ROCE             = 1 << 3,
 	MLX5_IPSEC_CAP_PRIO             = 1 << 4,
+	MLX5_IPSEC_CAP_TUNNEL           = 1 << 5,
 };
 
 struct mlx5e_priv;
@@ -141,6 +144,10 @@ struct mlx5e_ipsec_work {
 	void *data;
 };
 
+struct mlx5e_ipsec_netevent_data {
+	u8 addr[ETH_ALEN];
+};
+
 struct mlx5e_ipsec_dwork {
 	struct delayed_work dwork;
 	struct mlx5e_ipsec_sa_entry *sa_entry;
@@ -166,6 +173,7 @@ struct mlx5e_ipsec {
 	struct mlx5e_ipsec_tx *tx;
 	struct mlx5e_ipsec_aso *aso;
 	struct notifier_block nb;
+	struct notifier_block netevent_nb;
 	struct mlx5_ipsec_fs *roce;
 };
 
@@ -243,6 +251,7 @@ void mlx5e_accel_ipsec_fs_del_rule(struct mlx5e_ipsec_sa_entry *sa_entry);
 int mlx5e_accel_ipsec_fs_add_pol(struct mlx5e_ipsec_pol_entry *pol_entry);
 void mlx5e_accel_ipsec_fs_del_pol(struct mlx5e_ipsec_pol_entry *pol_entry);
 void mlx5e_accel_ipsec_fs_modify(struct mlx5e_ipsec_sa_entry *sa_entry);
+bool mlx5e_ipsec_fs_tunnel_enabled(struct mlx5e_ipsec_sa_entry *sa_entry);
 
 int mlx5_ipsec_create_sa_ctx(struct mlx5e_ipsec_sa_entry *sa_entry);
 void mlx5_ipsec_free_sa_ctx(struct mlx5e_ipsec_sa_entry *sa_entry);
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en_accel/ipsec_fs.c b/drivers/net/ethernet/mellanox/mlx5/core/en_accel/ipsec_fs.c
index b47794d4146e..5a8fcd30fcb1 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/en_accel/ipsec_fs.c
+++ b/drivers/net/ethernet/mellanox/mlx5/core/en_accel/ipsec_fs.c
@@ -4,12 +4,15 @@
 #include <linux/netdevice.h>
 #include "en.h"
 #include "en/fs.h"
+#include "eswitch.h"
 #include "ipsec.h"
 #include "fs_core.h"
 #include "lib/ipsec_fs_roce.h"
 #include "lib/fs_chains.h"
 
 #define NUM_IPSEC_FTE BIT(15)
+#define MLX5_REFORMAT_TYPE_ADD_ESP_TRANSPORT_SIZE 16
+#define IPSEC_TUNNEL_DEFAULT_TTL 0x40
 
 struct mlx5e_ipsec_fc {
 	struct mlx5_fc *cnt;
@@ -36,6 +39,7 @@ struct mlx5e_ipsec_rx {
 	struct mlx5e_ipsec_rule status;
 	struct mlx5e_ipsec_fc *fc;
 	struct mlx5_fs_chains *chains;
+	u8 allow_tunnel_mode : 1;
 };
 
 struct mlx5e_ipsec_tx {
@@ -45,6 +49,7 @@ struct mlx5e_ipsec_tx {
 	struct mlx5_flow_namespace *ns;
 	struct mlx5e_ipsec_fc *fc;
 	struct mlx5_fs_chains *chains;
+	u8 allow_tunnel_mode : 1;
 };
 
 /* IPsec RX flow steering */
@@ -118,7 +123,7 @@ static void ipsec_chains_put_table(struct mlx5_fs_chains *chains, u32 prio)
 
 static struct mlx5_flow_table *ipsec_ft_create(struct mlx5_flow_namespace *ns,
 					       int level, int prio,
-					       int max_num_groups)
+					       int max_num_groups, u32 flags)
 {
 	struct mlx5_flow_table_attr ft_attr = {};
 
@@ -127,6 +132,7 @@ static struct mlx5_flow_table *ipsec_ft_create(struct mlx5_flow_namespace *ns,
 	ft_attr.max_fte = NUM_IPSEC_FTE;
 	ft_attr.level = level;
 	ft_attr.prio = prio;
+	ft_attr.flags = flags;
 
 	return mlx5_create_auto_grouped_flow_table(ns, &ft_attr);
 }
@@ -251,7 +257,8 @@ static void rx_destroy(struct mlx5_core_dev *mdev, struct mlx5e_ipsec *ipsec,
 	mlx5_del_flow_rules(rx->sa.rule);
 	mlx5_destroy_flow_group(rx->sa.group);
 	mlx5_destroy_flow_table(rx->ft.sa);
-
+	if (rx->allow_tunnel_mode)
+		mlx5_eswitch_unblock_encap(mdev);
 	mlx5_del_flow_rules(rx->status.rule);
 	mlx5_modify_header_dealloc(mdev, rx->status.modify_hdr);
 	mlx5_destroy_flow_table(rx->ft.status);
@@ -267,6 +274,7 @@ static int rx_create(struct mlx5_core_dev *mdev, struct mlx5e_ipsec *ipsec,
 	struct mlx5_flow_destination default_dest;
 	struct mlx5_flow_destination dest[2];
 	struct mlx5_flow_table *ft;
+	u32 flags = 0;
 	int err;
 
 	default_dest = mlx5_ttc_get_default_dest(ttc, family2tt(family));
@@ -277,7 +285,7 @@ static int rx_create(struct mlx5_core_dev *mdev, struct mlx5e_ipsec *ipsec,
 		return err;
 
 	ft = ipsec_ft_create(ns, MLX5E_ACCEL_FS_ESP_FT_ERR_LEVEL,
-			     MLX5E_NIC_PRIO, 1);
+			     MLX5E_NIC_PRIO, 1, 0);
 	if (IS_ERR(ft)) {
 		err = PTR_ERR(ft);
 		goto err_fs_ft_status;
@@ -300,8 +308,12 @@ static int rx_create(struct mlx5_core_dev *mdev, struct mlx5e_ipsec *ipsec,
 		goto err_add;
 
 	/* Create FT */
-	ft = ipsec_ft_create(ns, MLX5E_ACCEL_FS_ESP_FT_LEVEL, MLX5E_NIC_PRIO,
-			     2);
+	if (mlx5_ipsec_device_caps(mdev) & MLX5_IPSEC_CAP_TUNNEL)
+		rx->allow_tunnel_mode = mlx5_eswitch_block_encap(mdev);
+	if (rx->allow_tunnel_mode)
+		flags = MLX5_FLOW_TABLE_TUNNEL_EN_REFORMAT;
+	ft = ipsec_ft_create(ns, MLX5E_ACCEL_FS_ESP_FT_LEVEL, MLX5E_NIC_PRIO, 2,
+			     flags);
 	if (IS_ERR(ft)) {
 		err = PTR_ERR(ft);
 		goto err_fs_ft;
@@ -327,7 +339,7 @@ static int rx_create(struct mlx5_core_dev *mdev, struct mlx5e_ipsec *ipsec,
 	}
 
 	ft = ipsec_ft_create(ns, MLX5E_ACCEL_FS_POL_FT_LEVEL, MLX5E_NIC_PRIO,
-			     2);
+			     2, 0);
 	if (IS_ERR(ft)) {
 		err = PTR_ERR(ft);
 		goto err_pol_ft;
@@ -356,6 +368,8 @@ err_pol_ft:
 err_fs:
 	mlx5_destroy_flow_table(rx->ft.sa);
 err_fs_ft:
+	if (rx->allow_tunnel_mode)
+		mlx5_eswitch_unblock_encap(mdev);
 	mlx5_del_flow_rules(rx->status.rule);
 	mlx5_modify_header_dealloc(mdev, rx->status.modify_hdr);
 err_add:
@@ -490,7 +504,8 @@ err_rule:
 }
 
 /* IPsec TX flow steering */
-static void tx_destroy(struct mlx5e_ipsec_tx *tx, struct mlx5_ipsec_fs *roce)
+static void tx_destroy(struct mlx5_core_dev *mdev, struct mlx5e_ipsec_tx *tx,
+		       struct mlx5_ipsec_fs *roce)
 {
 	mlx5_ipsec_fs_roce_tx_destroy(roce);
 	if (tx->chains) {
@@ -502,6 +517,8 @@ static void tx_destroy(struct mlx5e_ipsec_tx *tx, struct mlx5_ipsec_fs *roce)
 	}
 
 	mlx5_destroy_flow_table(tx->ft.sa);
+	if (tx->allow_tunnel_mode)
+		mlx5_eswitch_unblock_encap(mdev);
 	mlx5_del_flow_rules(tx->status.rule);
 	mlx5_destroy_flow_table(tx->ft.status);
 }
@@ -511,9 +528,10 @@ static int tx_create(struct mlx5_core_dev *mdev, struct mlx5e_ipsec_tx *tx,
 {
 	struct mlx5_flow_destination dest = {};
 	struct mlx5_flow_table *ft;
+	u32 flags = 0;
 	int err;
 
-	ft = ipsec_ft_create(tx->ns, 2, 0, 1);
+	ft = ipsec_ft_create(tx->ns, 2, 0, 1, 0);
 	if (IS_ERR(ft))
 		return PTR_ERR(ft);
 	tx->ft.status = ft;
@@ -522,7 +540,11 @@ static int tx_create(struct mlx5_core_dev *mdev, struct mlx5e_ipsec_tx *tx,
 	if (err)
 		goto err_status_rule;
 
-	ft = ipsec_ft_create(tx->ns, 1, 0, 4);
+	if (mlx5_ipsec_device_caps(mdev) & MLX5_IPSEC_CAP_TUNNEL)
+		tx->allow_tunnel_mode = mlx5_eswitch_block_encap(mdev);
+	if (tx->allow_tunnel_mode)
+		flags = MLX5_FLOW_TABLE_TUNNEL_EN_REFORMAT;
+	ft = ipsec_ft_create(tx->ns, 1, 0, 4, flags);
 	if (IS_ERR(ft)) {
 		err = PTR_ERR(ft);
 		goto err_sa_ft;
@@ -541,7 +563,7 @@ static int tx_create(struct mlx5_core_dev *mdev, struct mlx5e_ipsec_tx *tx,
 		goto connect_roce;
 	}
 
-	ft = ipsec_ft_create(tx->ns, 0, 0, 2);
+	ft = ipsec_ft_create(tx->ns, 0, 0, 2, 0);
 	if (IS_ERR(ft)) {
 		err = PTR_ERR(ft);
 		goto err_pol_ft;
@@ -572,6 +594,8 @@ err_roce:
 err_pol_ft:
 	mlx5_destroy_flow_table(tx->ft.sa);
 err_sa_ft:
+	if (tx->allow_tunnel_mode)
+		mlx5_eswitch_unblock_encap(mdev);
 	mlx5_del_flow_rules(tx->status.rule);
 err_status_rule:
 	mlx5_destroy_flow_table(tx->ft.status);
@@ -600,7 +624,7 @@ static void tx_put(struct mlx5e_ipsec *ipsec, struct mlx5e_ipsec_tx *tx)
 	if (--tx->ft.refcnt)
 		return;
 
-	tx_destroy(tx, ipsec->roce);
+	tx_destroy(ipsec->mdev, tx, ipsec->roce);
 }
 
 static struct mlx5_flow_table *tx_ft_get_policy(struct mlx5_core_dev *mdev,
@@ -829,40 +853,181 @@ static int setup_modify_header(struct mlx5_core_dev *mdev, u32 val, u8 dir,
 	return 0;
 }
 
+static int
+setup_pkt_tunnel_reformat(struct mlx5_core_dev *mdev,
+			  struct mlx5_accel_esp_xfrm_attrs *attrs,
+			  struct mlx5_pkt_reformat_params *reformat_params)
+{
+	struct ip_esp_hdr *esp_hdr;
+	struct ipv6hdr *ipv6hdr;
+	struct ethhdr *eth_hdr;
+	struct iphdr *iphdr;
+	char *reformatbf;
+	size_t bfflen;
+	void *hdr;
+
+	bfflen = sizeof(*eth_hdr);
+
+	if (attrs->dir == XFRM_DEV_OFFLOAD_OUT) {
+		bfflen += sizeof(*esp_hdr) + 8;
+
+		switch (attrs->family) {
+		case AF_INET:
+			bfflen += sizeof(*iphdr);
+			break;
+		case AF_INET6:
+			bfflen += sizeof(*ipv6hdr);
+			break;
+		default:
+			return -EINVAL;
+		}
+	}
+
+	reformatbf = kzalloc(bfflen, GFP_KERNEL);
+	if (!reformatbf)
+		return -ENOMEM;
+
+	eth_hdr = (struct ethhdr *)reformatbf;
+	switch (attrs->family) {
+	case AF_INET:
+		eth_hdr->h_proto = htons(ETH_P_IP);
+		break;
+	case AF_INET6:
+		eth_hdr->h_proto = htons(ETH_P_IPV6);
+		break;
+	default:
+		goto free_reformatbf;
+	}
+
+	ether_addr_copy(eth_hdr->h_dest, attrs->dmac);
+	ether_addr_copy(eth_hdr->h_source, attrs->smac);
+
+	switch (attrs->dir) {
+	case XFRM_DEV_OFFLOAD_IN:
+		reformat_params->type = MLX5_REFORMAT_TYPE_L3_ESP_TUNNEL_TO_L2;
+		break;
+	case XFRM_DEV_OFFLOAD_OUT:
+		reformat_params->type = MLX5_REFORMAT_TYPE_L2_TO_L3_ESP_TUNNEL;
+		reformat_params->param_0 = attrs->authsize;
+
+		hdr = reformatbf + sizeof(*eth_hdr);
+		switch (attrs->family) {
+		case AF_INET:
+			iphdr = (struct iphdr *)hdr;
+			memcpy(&iphdr->saddr, &attrs->saddr.a4, 4);
+			memcpy(&iphdr->daddr, &attrs->daddr.a4, 4);
+			iphdr->version = 4;
+			iphdr->ihl = 5;
+			iphdr->ttl = IPSEC_TUNNEL_DEFAULT_TTL;
+			iphdr->protocol = IPPROTO_ESP;
+			hdr += sizeof(*iphdr);
+			break;
+		case AF_INET6:
+			ipv6hdr = (struct ipv6hdr *)hdr;
+			memcpy(&ipv6hdr->saddr, &attrs->saddr.a6, 16);
+			memcpy(&ipv6hdr->daddr, &attrs->daddr.a6, 16);
+			ipv6hdr->nexthdr = IPPROTO_ESP;
+			ipv6hdr->version = 6;
+			ipv6hdr->hop_limit = IPSEC_TUNNEL_DEFAULT_TTL;
+			hdr += sizeof(*ipv6hdr);
+			break;
+		default:
+			goto free_reformatbf;
+		}
+
+		esp_hdr = (struct ip_esp_hdr *)hdr;
+		esp_hdr->spi = htonl(attrs->spi);
+		break;
+	default:
+		goto free_reformatbf;
+	}
+
+	reformat_params->size = bfflen;
+	reformat_params->data = reformatbf;
+	return 0;
+
+free_reformatbf:
+	kfree(reformatbf);
+	return -EINVAL;
+}
+
+static int
+setup_pkt_transport_reformat(struct mlx5_accel_esp_xfrm_attrs *attrs,
+			     struct mlx5_pkt_reformat_params *reformat_params)
+{
+	u8 *reformatbf;
+	__be32 spi;
+
+	switch (attrs->dir) {
+	case XFRM_DEV_OFFLOAD_IN:
+		reformat_params->type = MLX5_REFORMAT_TYPE_DEL_ESP_TRANSPORT;
+		break;
+	case XFRM_DEV_OFFLOAD_OUT:
+		if (attrs->family == AF_INET)
+			reformat_params->type =
+				MLX5_REFORMAT_TYPE_ADD_ESP_TRANSPORT_OVER_IPV4;
+		else
+			reformat_params->type =
+				MLX5_REFORMAT_TYPE_ADD_ESP_TRANSPORT_OVER_IPV6;
+
+		reformatbf = kzalloc(MLX5_REFORMAT_TYPE_ADD_ESP_TRANSPORT_SIZE,
+				     GFP_KERNEL);
+		if (!reformatbf)
+			return -ENOMEM;
+
+		/* convert to network format */
+		spi = htonl(attrs->spi);
+		memcpy(reformatbf, &spi, sizeof(spi));
+
+		reformat_params->param_0 = attrs->authsize;
+		reformat_params->size =
+			MLX5_REFORMAT_TYPE_ADD_ESP_TRANSPORT_SIZE;
+		reformat_params->data = reformatbf;
+		break;
+	default:
+		return -EINVAL;
+	}
+
+	return 0;
+}
+
 static int setup_pkt_reformat(struct mlx5_core_dev *mdev,
 			      struct mlx5_accel_esp_xfrm_attrs *attrs,
 			      struct mlx5_flow_act *flow_act)
 {
-	enum mlx5_flow_namespace_type ns_type = MLX5_FLOW_NAMESPACE_EGRESS;
 	struct mlx5_pkt_reformat_params reformat_params = {};
 	struct mlx5_pkt_reformat *pkt_reformat;
-	u8 reformatbf[16] = {};
-	__be32 spi;
+	enum mlx5_flow_namespace_type ns_type;
+	int ret;
 
-	if (attrs->dir == XFRM_DEV_OFFLOAD_IN) {
-		reformat_params.type = MLX5_REFORMAT_TYPE_DEL_ESP_TRANSPORT;
+	switch (attrs->dir) {
+	case XFRM_DEV_OFFLOAD_IN:
 		ns_type = MLX5_FLOW_NAMESPACE_KERNEL;
-		goto cmd;
+		break;
+	case XFRM_DEV_OFFLOAD_OUT:
+		ns_type = MLX5_FLOW_NAMESPACE_EGRESS;
+		break;
+	default:
+		return -EINVAL;
 	}
 
-	if (attrs->family == AF_INET)
-		reformat_params.type =
-			MLX5_REFORMAT_TYPE_ADD_ESP_TRANSPORT_OVER_IPV4;
-	else
-		reformat_params.type =
-			MLX5_REFORMAT_TYPE_ADD_ESP_TRANSPORT_OVER_IPV6;
-
-	/* convert to network format */
-	spi = htonl(attrs->spi);
-	memcpy(reformatbf, &spi, 4);
+	switch (attrs->mode) {
+	case XFRM_MODE_TRANSPORT:
+		ret = setup_pkt_transport_reformat(attrs, &reformat_params);
+		break;
+	case XFRM_MODE_TUNNEL:
+		ret = setup_pkt_tunnel_reformat(mdev, attrs, &reformat_params);
+		break;
+	default:
+		ret = -EINVAL;
+	}
 
-	reformat_params.param_0 = attrs->authsize;
-	reformat_params.size = sizeof(reformatbf);
-	reformat_params.data = &reformatbf;
+	if (ret)
+		return ret;
 
-cmd:
 	pkt_reformat =
 		mlx5_packet_reformat_alloc(mdev, &reformat_params, ns_type);
+	kfree(reformat_params.data);
 	if (IS_ERR(pkt_reformat))
 		return PTR_ERR(pkt_reformat);
 
@@ -1453,3 +1618,15 @@ void mlx5e_accel_ipsec_fs_modify(struct mlx5e_ipsec_sa_entry *sa_entry)
 	mlx5e_accel_ipsec_fs_del_rule(sa_entry);
 	memcpy(sa_entry, &sa_entry_shadow, sizeof(*sa_entry));
 }
+
+bool mlx5e_ipsec_fs_tunnel_enabled(struct mlx5e_ipsec_sa_entry *sa_entry)
+{
+	struct mlx5e_ipsec_rx *rx =
+		ipsec_rx(sa_entry->ipsec, sa_entry->attrs.family);
+	struct mlx5e_ipsec_tx *tx = sa_entry->ipsec->tx;
+
+	if (sa_entry->attrs.dir == XFRM_DEV_OFFLOAD_OUT)
+		return tx->allow_tunnel_mode;
+
+	return rx->allow_tunnel_mode;
+}
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en_accel/ipsec_offload.c b/drivers/net/ethernet/mellanox/mlx5/core/en_accel/ipsec_offload.c
index 5fddb86bb35e..df90e19066bc 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/en_accel/ipsec_offload.c
+++ b/drivers/net/ethernet/mellanox/mlx5/core/en_accel/ipsec_offload.c
@@ -48,6 +48,12 @@ u32 mlx5_ipsec_device_caps(struct mlx5_core_dev *mdev)
 		if (MLX5_CAP_FLOWTABLE_NIC_TX(mdev, ignore_flow_level) &&
 		    MLX5_CAP_FLOWTABLE_NIC_RX(mdev, ignore_flow_level))
 			caps |= MLX5_IPSEC_CAP_PRIO;
+
+		if (MLX5_CAP_FLOWTABLE_NIC_TX(mdev,
+					      reformat_l2_to_l3_esp_tunnel) &&
+		    MLX5_CAP_FLOWTABLE_NIC_RX(mdev,
+					      reformat_l3_esp_tunnel_to_l2))
+			caps |= MLX5_IPSEC_CAP_TUNNEL;
 	}
 
 	if (mlx5_get_roce_state(mdev) &&
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/eswitch.h b/drivers/net/ethernet/mellanox/mlx5/core/eswitch.h
index 19e9a77c4633..e9d68fdf68f5 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/eswitch.h
+++ b/drivers/net/ethernet/mellanox/mlx5/core/eswitch.h
@@ -263,6 +263,7 @@ struct mlx5_esw_offload {
 	const struct mlx5_eswitch_rep_ops *rep_ops[NUM_REP_TYPES];
 	u8 inline_mode;
 	atomic64_t num_flows;
+	u64 num_block_encap;
 	enum devlink_eswitch_encap_mode encap;
 	struct ida vport_metadata_ida;
 	unsigned int host_number; /* ECPF supports one external host */
@@ -748,6 +749,9 @@ void mlx5_eswitch_offloads_destroy_single_fdb(struct mlx5_eswitch *master_esw,
 					      struct mlx5_eswitch *slave_esw);
 int mlx5_eswitch_reload_reps(struct mlx5_eswitch *esw);
 
+bool mlx5_eswitch_block_encap(struct mlx5_core_dev *dev);
+void mlx5_eswitch_unblock_encap(struct mlx5_core_dev *dev);
+
 static inline int mlx5_eswitch_num_vfs(struct mlx5_eswitch *esw)
 {
 	if (mlx5_esw_allowed(esw))
@@ -761,6 +765,7 @@ mlx5_eswitch_get_slow_fdb(struct mlx5_eswitch *esw)
 {
 	return esw->fdb_table.offloads.slow_fdb;
 }
+
 #else  /* CONFIG_MLX5_ESWITCH */
 /* eswitch API stubs */
 static inline int  mlx5_eswitch_init(struct mlx5_core_dev *dev) { return 0; }
@@ -805,6 +810,15 @@ mlx5_eswitch_reload_reps(struct mlx5_eswitch *esw)
 {
 	return 0;
 }
+
+static inline bool mlx5_eswitch_block_encap(struct mlx5_core_dev *dev)
+{
+	return true;
+}
+
+static inline void mlx5_eswitch_unblock_encap(struct mlx5_core_dev *dev)
+{
+}
 #endif /* CONFIG_MLX5_ESWITCH */
 
 #endif /* __MLX5_ESWITCH_H__ */
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/eswitch_offloads.c b/drivers/net/ethernet/mellanox/mlx5/core/eswitch_offloads.c
index 48036dfddd5e..b6e2709c1371 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/eswitch_offloads.c
+++ b/drivers/net/ethernet/mellanox/mlx5/core/eswitch_offloads.c
@@ -3586,6 +3586,47 @@ int mlx5_devlink_eswitch_inline_mode_get(struct devlink *devlink, u8 *mode)
 	return err;
 }
 
+bool mlx5_eswitch_block_encap(struct mlx5_core_dev *dev)
+{
+	struct devlink *devlink = priv_to_devlink(dev);
+	struct mlx5_eswitch *esw;
+
+	devl_lock(devlink);
+	esw = mlx5_devlink_eswitch_get(devlink);
+	if (IS_ERR(esw)) {
+		devl_unlock(devlink);
+		/* Failure means no eswitch => not possible to change encap */
+		return true;
+	}
+
+	down_write(&esw->mode_lock);
+	if (esw->mode != MLX5_ESWITCH_LEGACY &&
+	    esw->offloads.encap != DEVLINK_ESWITCH_ENCAP_MODE_NONE) {
+		up_write(&esw->mode_lock);
+		devl_unlock(devlink);
+		return false;
+	}
+
+	esw->offloads.num_block_encap++;
+	up_write(&esw->mode_lock);
+	devl_unlock(devlink);
+	return true;
+}
+
+void mlx5_eswitch_unblock_encap(struct mlx5_core_dev *dev)
+{
+	struct devlink *devlink = priv_to_devlink(dev);
+	struct mlx5_eswitch *esw;
+
+	esw = mlx5_devlink_eswitch_get(devlink);
+	if (IS_ERR(esw))
+		return;
+
+	down_write(&esw->mode_lock);
+	esw->offloads.num_block_encap--;
+	up_write(&esw->mode_lock);
+}
+
 int mlx5_devlink_eswitch_encap_mode_set(struct devlink *devlink,
 					enum devlink_eswitch_encap_mode encap,
 					struct netlink_ext_ack *extack)
@@ -3627,6 +3668,13 @@ int mlx5_devlink_eswitch_encap_mode_set(struct devlink *devlink,
 		goto unlock;
 	}
 
+	if (esw->offloads.num_block_encap) {
+		NL_SET_ERR_MSG_MOD(extack,
+				   "Can't set encapsulation when IPsec SA and/or policies are configured");
+		err = -EOPNOTSUPP;
+		goto unlock;
+	}
+
 	esw_destroy_offloads_fdb_tables(esw);
 
 	esw->offloads.encap = encap;
diff --git a/include/linux/mlx5/mlx5_ifc.h b/include/linux/mlx5/mlx5_ifc.h
index 6c84bf6eec85..20d00e09b168 100644
--- a/include/linux/mlx5/mlx5_ifc.h
+++ b/include/linux/mlx5/mlx5_ifc.h
@@ -463,9 +463,11 @@ struct mlx5_ifc_flow_table_prop_layout_bits {
 	u8         max_ft_level[0x8];
 
 	u8         reformat_add_esp_trasport[0x1];
-	u8         reserved_at_41[0x2];
+	u8         reformat_l2_to_l3_esp_tunnel[0x1];
+	u8         reserved_at_42[0x1];
 	u8         reformat_del_esp_trasport[0x1];
-	u8         reserved_at_44[0x2];
+	u8         reformat_l3_esp_tunnel_to_l2[0x1];
+	u8         reserved_at_45[0x1];
 	u8         execute_aso[0x1];
 	u8         reserved_at_47[0x19];
 
@@ -6630,7 +6632,9 @@ enum mlx5_reformat_ctx_type {
 	MLX5_REFORMAT_TYPE_L3_TUNNEL_TO_L2 = 0x3,
 	MLX5_REFORMAT_TYPE_L2_TO_L3_TUNNEL = 0x4,
 	MLX5_REFORMAT_TYPE_ADD_ESP_TRANSPORT_OVER_IPV4 = 0x5,
+	MLX5_REFORMAT_TYPE_L2_TO_L3_ESP_TUNNEL = 0x6,
 	MLX5_REFORMAT_TYPE_DEL_ESP_TRANSPORT = 0x8,
+	MLX5_REFORMAT_TYPE_L3_ESP_TUNNEL_TO_L2 = 0x9,
 	MLX5_REFORMAT_TYPE_ADD_ESP_TRANSPORT_OVER_IPV6 = 0xb,
 	MLX5_REFORMAT_TYPE_INSERT_HDR = 0xf,
 	MLX5_REFORMAT_TYPE_REMOVE_HDR = 0x10,