Merge tag 'ipsec-esn-replay' of https://git.kernel.org/pub/scm/linux/kernel/git/mellanox/linux

Leon Romanovsky says:

====================
Improve IPsec limits, ESN and replay window

This series overcomes existing hardware limitations in Mellanox ConnectX
devices around handling IPsec soft and hard limits.

In addition, the ESN logic is tied and added an interface to configure
replay window sequence numbers through existing iproute2 interface.

  ip xfrm state ... [ replay-seq SEQ ] [ replay-oseq SEQ ]
                    [ replay-seq-hi SEQ ] [ replay-oseq-hi SEQ ]

Link: https://lore.kernel.org/all/cover.1680162300.git.leonro@nvidia.com
Signed-off-by: Leon Romanovsky <leon@kernel.org>

* tag 'ipsec-esn-replay' of https://git.kernel.org/pub/scm/linux/kernel/git/mellanox/linux:
  net/mlx5e: Simulate missing IPsec TX limits hardware functionality
  net/mlx5e: Generalize IPsec work structs
  net/mlx5e: Reduce contention in IPsec workqueue
  net/mlx5e: Set IPsec replay sequence numbers
  net/mlx5e: Remove ESN callbacks if it is not supported
  xfrm: don't require advance ESN callback for packet offload
  net/mlx5e: Overcome slow response for first IPsec ASO WQE
  net/mlx5e: Add SW implementation to support IPsec 64 bit soft and hard limits
  net/mlx5e: Prevent zero IPsec soft/hard limits
  net/mlx5e: Factor out IPsec ASO update function
====================

Link: https://lore.kernel.org/r/20230406071902.712388-1-leon@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>

4 files changed, 497 insertions, 110 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 91fa0a366316..def01bfde610 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/en_accel/ipsec.c
+++ b/drivers/net/ethernet/mellanox/mlx5/core/en_accel/ipsec.c
@@ -40,6 +40,8 @@
 #include "ipsec.h"
 #include "ipsec_rxtx.h"
 
+#define MLX5_IPSEC_RESCHED msecs_to_jiffies(1000)
+
 static struct mlx5e_ipsec_sa_entry *to_ipsec_sa_entry(struct xfrm_state *x)
 {
 	return (struct mlx5e_ipsec_sa_entry *)x->xso.offload_handle;
@@ -50,27 +52,71 @@ static struct mlx5e_ipsec_pol_entry *to_ipsec_pol_entry(struct xfrm_policy *x)
 	return (struct mlx5e_ipsec_pol_entry *)x->xdo.offload_handle;
 }
 
+static void mlx5e_ipsec_handle_tx_limit(struct work_struct *_work)
+{
+	struct mlx5e_ipsec_dwork *dwork =
+		container_of(_work, struct mlx5e_ipsec_dwork, dwork.work);
+	struct mlx5e_ipsec_sa_entry *sa_entry = dwork->sa_entry;
+	struct xfrm_state *x = sa_entry->x;
+
+	spin_lock(&x->lock);
+	xfrm_state_check_expire(x);
+	if (x->km.state == XFRM_STATE_EXPIRED) {
+		sa_entry->attrs.drop = true;
+		mlx5e_accel_ipsec_fs_modify(sa_entry);
+	}
+	spin_unlock(&x->lock);
+
+	if (sa_entry->attrs.drop)
+		return;
+
+	queue_delayed_work(sa_entry->ipsec->wq, &dwork->dwork,
+			   MLX5_IPSEC_RESCHED);
+}
+
 static bool mlx5e_ipsec_update_esn_state(struct mlx5e_ipsec_sa_entry *sa_entry)
 {
-	struct xfrm_replay_state_esn *replay_esn;
+	struct xfrm_state *x = sa_entry->x;
 	u32 seq_bottom = 0;
+	u32 esn, esn_msb;
 	u8 overlap;
 
-	if (!(sa_entry->x->props.flags & XFRM_STATE_ESN)) {
-		sa_entry->esn_state.trigger = 0;
+	switch (x->xso.type) {
+	case XFRM_DEV_OFFLOAD_PACKET:
+		switch (x->xso.dir) {
+		case XFRM_DEV_OFFLOAD_IN:
+			esn = x->replay_esn->seq;
+			esn_msb = x->replay_esn->seq_hi;
+			break;
+		case XFRM_DEV_OFFLOAD_OUT:
+			esn = x->replay_esn->oseq;
+			esn_msb = x->replay_esn->oseq_hi;
+			break;
+		default:
+			WARN_ON(true);
+			return false;
+		}
+		break;
+	case XFRM_DEV_OFFLOAD_CRYPTO:
+		/* Already parsed by XFRM core */
+		esn = x->replay_esn->seq;
+		break;
+	default:
+		WARN_ON(true);
 		return false;
 	}
 
-	replay_esn = sa_entry->x->replay_esn;
-	if (replay_esn->seq >= replay_esn->replay_window)
-		seq_bottom = replay_esn->seq - replay_esn->replay_window + 1;
-
 	overlap = sa_entry->esn_state.overlap;
 
-	sa_entry->esn_state.esn = xfrm_replay_seqhi(sa_entry->x,
-						    htonl(seq_bottom));
+	if (esn >= x->replay_esn->replay_window)
+		seq_bottom = esn - x->replay_esn->replay_window + 1;
+
+	if (x->xso.type == XFRM_DEV_OFFLOAD_CRYPTO)
+		esn_msb = xfrm_replay_seqhi(x, htonl(seq_bottom));
+
+	sa_entry->esn_state.esn = esn;
+	sa_entry->esn_state.esn_msb = esn_msb;
 
-	sa_entry->esn_state.trigger = 1;
 	if (unlikely(overlap && seq_bottom < MLX5E_IPSEC_ESN_SCOPE_MID)) {
 		sa_entry->esn_state.overlap = 0;
 		return true;
@@ -87,25 +133,113 @@ static void mlx5e_ipsec_init_limits(struct mlx5e_ipsec_sa_entry *sa_entry,
 				    struct mlx5_accel_esp_xfrm_attrs *attrs)
 {
 	struct xfrm_state *x = sa_entry->x;
+	s64 start_value, n;
 
-	attrs->hard_packet_limit = x->lft.hard_packet_limit;
+	attrs->lft.hard_packet_limit = x->lft.hard_packet_limit;
+	attrs->lft.soft_packet_limit = x->lft.soft_packet_limit;
 	if (x->lft.soft_packet_limit == XFRM_INF)
 		return;
 
-	/* Hardware decrements hard_packet_limit counter through
-	 * the operation. While fires an event when soft_packet_limit
-	 * is reached. It emans that we need substitute the numbers
-	 * in order to properly count soft limit.
+	/* Compute hard limit initial value and number of rounds.
+	 *
+	 * The counting pattern of hardware counter goes:
+	 *                value  -> 2^31-1
+	 *      2^31  | (2^31-1) -> 2^31-1
+	 *      2^31  | (2^31-1) -> 2^31-1
+	 *      [..]
+	 *      2^31  | (2^31-1) -> 0
+	 *
+	 * The pattern is created by using an ASO operation to atomically set
+	 * bit 31 after the down counter clears bit 31. This is effectively an
+	 * atomic addition of 2**31 to the counter.
 	 *
-	 * As an example:
-	 * XFRM user sets soft limit is 2 and hard limit is 9 and
-	 * expects to see soft event after 2 packets and hard event
-	 * after 9 packets. In our case, the hard limit will be set
-	 * to 9 and soft limit is comparator to 7 so user gets the
-	 * soft event after 2 packeta
+	 * We wish to configure the counter, within the above pattern, so that
+	 * when it reaches 0, it has hit the hard limit. This is defined by this
+	 * system of equations:
+	 *
+	 *      hard_limit == start_value + n * 2^31
+	 *      n >= 0
+	 *      start_value < 2^32, start_value >= 0
+	 *
+	 * These equations are not single-solution, there are often two choices:
+	 *      hard_limit == start_value + n * 2^31
+	 *      hard_limit == (start_value+2^31) + (n-1) * 2^31
+	 *
+	 * The algorithm selects the solution that keeps the counter value
+	 * above 2^31 until the final iteration.
+	 */
+
+	/* Start by estimating n and compute start_value */
+	n = attrs->lft.hard_packet_limit / BIT_ULL(31);
+	start_value = attrs->lft.hard_packet_limit - n * BIT_ULL(31);
+
+	/* Choose the best of the two solutions: */
+	if (n >= 1)
+		n -= 1;
+
+	/* Computed values solve the system of equations: */
+	start_value = attrs->lft.hard_packet_limit - n * BIT_ULL(31);
+
+	/* The best solution means: when there are multiple iterations we must
+	 * start above 2^31 and count down to 2**31 to get the interrupt.
 	 */
-	attrs->soft_packet_limit =
-		x->lft.hard_packet_limit - x->lft.soft_packet_limit;
+	attrs->lft.hard_packet_limit = lower_32_bits(start_value);
+	attrs->lft.numb_rounds_hard = (u64)n;
+
+	/* Compute soft limit initial value and number of rounds.
+	 *
+	 * The soft_limit is achieved by adjusting the counter's
+	 * interrupt_value. This is embedded in the counting pattern created by
+	 * hard packet calculations above.
+	 *
+	 * We wish to compute the interrupt_value for the soft_limit. This is
+	 * defined by this system of equations:
+	 *
+	 *      soft_limit == start_value - soft_value + n * 2^31
+	 *      n >= 0
+	 *      soft_value < 2^32, soft_value >= 0
+	 *      for n == 0 start_value > soft_value
+	 *
+	 * As with compute_hard_n_value() the equations are not single-solution.
+	 * The algorithm selects the solution that has:
+	 *      2^30 <= soft_limit < 2^31 + 2^30
+	 * for the interior iterations, which guarantees a large guard band
+	 * around the counter hard limit and next interrupt.
+	 */
+
+	/* Start by estimating n and compute soft_value */
+	n = (x->lft.soft_packet_limit - attrs->lft.hard_packet_limit) / BIT_ULL(31);
+	start_value = attrs->lft.hard_packet_limit + n * BIT_ULL(31) -
+		      x->lft.soft_packet_limit;
+
+	/* Compare against constraints and adjust n */
+	if (n < 0)
+		n = 0;
+	else if (start_value >= BIT_ULL(32))
+		n -= 1;
+	else if (start_value < 0)
+		n += 1;
+
+	/* Choose the best of the two solutions: */
+	start_value = attrs->lft.hard_packet_limit + n * BIT_ULL(31) - start_value;
+	if (n != attrs->lft.numb_rounds_hard && start_value < BIT_ULL(30))
+		n += 1;
+
+	/* Note that the upper limit of soft_value happens naturally because we
+	 * always select the lowest soft_value.
+	 */
+
+	/* Computed values solve the system of equations: */
+	start_value = attrs->lft.hard_packet_limit + n * BIT_ULL(31) - start_value;
+
+	/* The best solution means: when there are multiple iterations we must
+	 * not fall below 2^30 as that would get too close to the false
+	 * hard_limit and when we reach an interior iteration for soft_limit it
+	 * has to be far away from 2**32-1 which is the counter reset point
+	 * after the +2^31 to accommodate latency.
+	 */
+	attrs->lft.soft_packet_limit = lower_32_bits(start_value);
+	attrs->lft.numb_rounds_soft = (u64)n;
 }
 
 void mlx5e_ipsec_build_accel_xfrm_attrs(struct mlx5e_ipsec_sa_entry *sa_entry,
@@ -141,11 +275,11 @@ void mlx5e_ipsec_build_accel_xfrm_attrs(struct mlx5e_ipsec_sa_entry *sa_entry,
 	aes_gcm->icv_len = x->aead->alg_icv_len;
 
 	/* esn */
-	if (sa_entry->esn_state.trigger) {
-		attrs->esn_trigger = true;
-		attrs->esn = sa_entry->esn_state.esn;
-		attrs->esn_overlap = sa_entry->esn_state.overlap;
-		attrs->replay_window = x->replay_esn->replay_window;
+	if (x->props.flags & XFRM_STATE_ESN) {
+		attrs->replay_esn.trigger = true;
+		attrs->replay_esn.esn = sa_entry->esn_state.esn;
+		attrs->replay_esn.esn_msb = sa_entry->esn_state.esn_msb;
+		attrs->replay_esn.overlap = sa_entry->esn_state.overlap;
 	}
 
 	attrs->dir = x->xso.dir;
@@ -283,6 +417,11 @@ static int mlx5e_xfrm_validate_state(struct mlx5_core_dev *mdev,
 			NL_SET_ERR_MSG_MOD(extack, "Hard packet limit must be greater than soft one");
 			return -EINVAL;
 		}
+
+		if (!x->lft.soft_packet_limit || !x->lft.hard_packet_limit) {
+			NL_SET_ERR_MSG_MOD(extack, "Soft/hard packet limits can't be 0");
+			return -EINVAL;
+		}
 		break;
 	default:
 		NL_SET_ERR_MSG_MOD(extack, "Unsupported xfrm offload type");
@@ -291,14 +430,87 @@ static int mlx5e_xfrm_validate_state(struct mlx5_core_dev *mdev,
 	return 0;
 }
 
-static void _update_xfrm_state(struct work_struct *work)
+static void mlx5e_ipsec_modify_state(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 mlx5_accel_esp_xfrm_attrs *attrs;
+
+	attrs = &((struct mlx5e_ipsec_sa_entry *)work->data)->attrs;
+
+	mlx5_accel_esp_modify_xfrm(sa_entry, attrs);
+}
+
+static void mlx5e_ipsec_set_esn_ops(struct mlx5e_ipsec_sa_entry *sa_entry)
+{
+	struct xfrm_state *x = sa_entry->x;
+
+	if (x->xso.type != XFRM_DEV_OFFLOAD_CRYPTO ||
+	    x->xso.dir != XFRM_DEV_OFFLOAD_OUT)
+		return;
+
+	if (x->props.flags & XFRM_STATE_ESN) {
+		sa_entry->set_iv_op = mlx5e_ipsec_set_iv_esn;
+		return;
+	}
+
+	sa_entry->set_iv_op = mlx5e_ipsec_set_iv;
+}
+
+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;
+
+	switch (x->xso.type) {
+	case XFRM_DEV_OFFLOAD_CRYPTO:
+		if (!(x->props.flags & XFRM_STATE_ESN))
+			return 0;
+		break;
+	default:
+		return 0;
+	}
+
+	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;
+	}
+
+	INIT_WORK(&work->work, mlx5e_ipsec_modify_state);
+	work->sa_entry = sa_entry;
+	sa_entry->work = work;
+	return 0;
+}
+
+static int mlx5e_ipsec_create_dwork(struct mlx5e_ipsec_sa_entry *sa_entry)
 {
-	struct mlx5e_ipsec_modify_state_work *modify_work =
-		container_of(work, struct mlx5e_ipsec_modify_state_work, work);
-	struct mlx5e_ipsec_sa_entry *sa_entry = container_of(
-		modify_work, struct mlx5e_ipsec_sa_entry, modify_work);
+	struct xfrm_state *x = sa_entry->x;
+	struct mlx5e_ipsec_dwork *dwork;
+
+	if (x->xso.type != XFRM_DEV_OFFLOAD_PACKET)
+		return 0;
 
-	mlx5_accel_esp_modify_xfrm(sa_entry, &modify_work->attrs);
+	if (x->xso.dir != XFRM_DEV_OFFLOAD_OUT)
+		return 0;
+
+	if (x->lft.soft_packet_limit == XFRM_INF &&
+	    x->lft.hard_packet_limit == XFRM_INF)
+		return 0;
+
+	dwork = kzalloc(sizeof(*dwork), GFP_KERNEL);
+	if (!dwork)
+		return -ENOMEM;
+
+	dwork->sa_entry = sa_entry;
+	INIT_DELAYED_WORK(&dwork->dwork, mlx5e_ipsec_handle_tx_limit);
+	sa_entry->dwork = dwork;
+	return 0;
 }
 
 static int mlx5e_xfrm_add_state(struct xfrm_state *x,
@@ -332,13 +544,23 @@ static int mlx5e_xfrm_add_state(struct xfrm_state *x,
 		goto err_xfrm;
 
 	/* check esn */
-	mlx5e_ipsec_update_esn_state(sa_entry);
+	if (x->props.flags & XFRM_STATE_ESN)
+		mlx5e_ipsec_update_esn_state(sa_entry);
 
 	mlx5e_ipsec_build_accel_xfrm_attrs(sa_entry, &sa_entry->attrs);
+
+	err = mlx5_ipsec_create_work(sa_entry);
+	if (err)
+		goto err_xfrm;
+
+	err = mlx5e_ipsec_create_dwork(sa_entry);
+	if (err)
+		goto release_work;
+
 	/* create hw context */
 	err = mlx5_ipsec_create_sa_ctx(sa_entry);
 	if (err)
-		goto err_xfrm;
+		goto release_dwork;
 
 	err = mlx5e_accel_ipsec_fs_add_rule(sa_entry);
 	if (err)
@@ -353,11 +575,11 @@ static int mlx5e_xfrm_add_state(struct xfrm_state *x,
 	if (err)
 		goto err_add_rule;
 
-	if (x->xso.dir == XFRM_DEV_OFFLOAD_OUT)
-		sa_entry->set_iv_op = (x->props.flags & XFRM_STATE_ESN) ?
-				mlx5e_ipsec_set_iv_esn : mlx5e_ipsec_set_iv;
+	mlx5e_ipsec_set_esn_ops(sa_entry);
 
-	INIT_WORK(&sa_entry->modify_work.work, _update_xfrm_state);
+	if (sa_entry->dwork)
+		queue_delayed_work(ipsec->wq, &sa_entry->dwork->dwork,
+				   MLX5_IPSEC_RESCHED);
 out:
 	x->xso.offload_handle = (unsigned long)sa_entry;
 	return 0;
@@ -366,6 +588,11 @@ err_add_rule:
 	mlx5e_accel_ipsec_fs_del_rule(sa_entry);
 err_hw_ctx:
 	mlx5_ipsec_free_sa_ctx(sa_entry);
+release_dwork:
+	kfree(sa_entry->dwork);
+release_work:
+	kfree(sa_entry->work->data);
+	kfree(sa_entry->work);
 err_xfrm:
 	kfree(sa_entry);
 	NL_SET_ERR_MSG_MOD(extack, "Device failed to offload this policy");
@@ -392,9 +619,17 @@ static void mlx5e_xfrm_free_state(struct xfrm_state *x)
 	if (x->xso.flags & XFRM_DEV_OFFLOAD_FLAG_ACQ)
 		goto sa_entry_free;
 
-	cancel_work_sync(&sa_entry->modify_work.work);
+	if (sa_entry->work)
+		cancel_work_sync(&sa_entry->work->work);
+
+	if (sa_entry->dwork)
+		cancel_delayed_work_sync(&sa_entry->dwork->dwork);
+
 	mlx5e_accel_ipsec_fs_del_rule(sa_entry);
 	mlx5_ipsec_free_sa_ctx(sa_entry);
+	kfree(sa_entry->dwork);
+	kfree(sa_entry->work->data);
+	kfree(sa_entry->work);
 sa_entry_free:
 	kfree(sa_entry);
 }
@@ -415,8 +650,8 @@ void mlx5e_ipsec_init(struct mlx5e_priv *priv)
 
 	xa_init_flags(&ipsec->sadb, XA_FLAGS_ALLOC);
 	ipsec->mdev = priv->mdev;
-	ipsec->wq = alloc_ordered_workqueue("mlx5e_ipsec: %s", 0,
-					    priv->netdev->name);
+	ipsec->wq = alloc_workqueue("mlx5e_ipsec: %s", WQ_UNBOUND, 0,
+				    priv->netdev->name);
 	if (!ipsec->wq)
 		goto err_wq;
 
@@ -480,16 +715,18 @@ static bool mlx5e_ipsec_offload_ok(struct sk_buff *skb, struct xfrm_state *x)
 static void mlx5e_xfrm_advance_esn_state(struct xfrm_state *x)
 {
 	struct mlx5e_ipsec_sa_entry *sa_entry = to_ipsec_sa_entry(x);
-	struct mlx5e_ipsec_modify_state_work *modify_work =
-		&sa_entry->modify_work;
+	struct mlx5e_ipsec_work *work = sa_entry->work;
+	struct mlx5e_ipsec_sa_entry *sa_entry_shadow;
 	bool need_update;
 
 	need_update = mlx5e_ipsec_update_esn_state(sa_entry);
 	if (!need_update)
 		return;
 
-	mlx5e_ipsec_build_accel_xfrm_attrs(sa_entry, &modify_work->attrs);
-	queue_work(sa_entry->ipsec->wq, &modify_work->work);
+	sa_entry_shadow = work->data;
+	memset(sa_entry_shadow, 0x00, sizeof(*sa_entry_shadow));
+	mlx5e_ipsec_build_accel_xfrm_attrs(sa_entry, &sa_entry_shadow->attrs);
+	queue_work(sa_entry->ipsec->wq, &work->work);
 }
 
 static void mlx5e_xfrm_update_curlft(struct xfrm_state *x)
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 68ae5230eb75..52890d7dce6b 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/en_accel/ipsec.h
+++ b/drivers/net/ethernet/mellanox/mlx5/core/en_accel/ipsec.h
@@ -60,8 +60,22 @@ struct upspec {
 	u8 proto;
 };
 
+struct mlx5_ipsec_lft {
+	u64 hard_packet_limit;
+	u64 soft_packet_limit;
+	u64 numb_rounds_hard;
+	u64 numb_rounds_soft;
+};
+
+struct mlx5_replay_esn {
+	u32 replay_window;
+	u32 esn;
+	u32 esn_msb;
+	u8 overlap : 1;
+	u8 trigger : 1;
+};
+
 struct mlx5_accel_esp_xfrm_attrs {
-	u32   esn;
 	u32   spi;
 	u32   flags;
 	struct aes_gcm_keymat aes_gcm;
@@ -78,15 +92,13 @@ struct mlx5_accel_esp_xfrm_attrs {
 
 	struct upspec upspec;
 	u8 dir : 2;
-	u8 esn_overlap : 1;
-	u8 esn_trigger : 1;
 	u8 type : 2;
+	u8 drop : 1;
 	u8 family;
-	u32 replay_window;
+	struct mlx5_replay_esn replay_esn;
 	u32 authsize;
 	u32 reqid;
-	u64 hard_packet_limit;
-	u64 soft_packet_limit;
+	struct mlx5_ipsec_lft lft;
 };
 
 enum mlx5_ipsec_cap {
@@ -125,8 +137,13 @@ struct mlx5e_ipsec_tx;
 
 struct mlx5e_ipsec_work {
 	struct work_struct work;
-	struct mlx5e_ipsec *ipsec;
-	u32 id;
+	struct mlx5e_ipsec_sa_entry *sa_entry;
+	void *data;
+};
+
+struct mlx5e_ipsec_dwork {
+	struct delayed_work dwork;
+	struct mlx5e_ipsec_sa_entry *sa_entry;
 };
 
 struct mlx5e_ipsec_aso {
@@ -154,7 +171,7 @@ struct mlx5e_ipsec {
 
 struct mlx5e_ipsec_esn_state {
 	u32 esn;
-	u8 trigger: 1;
+	u32 esn_msb;
 	u8 overlap: 1;
 };
 
@@ -165,9 +182,10 @@ struct mlx5e_ipsec_rule {
 	struct mlx5_fc *fc;
 };
 
-struct mlx5e_ipsec_modify_state_work {
-	struct work_struct		work;
-	struct mlx5_accel_esp_xfrm_attrs attrs;
+struct mlx5e_ipsec_limits {
+	u64 round;
+	u8 soft_limit_hit : 1;
+	u8 fix_limit : 1;
 };
 
 struct mlx5e_ipsec_sa_entry {
@@ -180,7 +198,9 @@ struct mlx5e_ipsec_sa_entry {
 	u32 ipsec_obj_id;
 	u32 enc_key_id;
 	struct mlx5e_ipsec_rule ipsec_rule;
-	struct mlx5e_ipsec_modify_state_work modify_work;
+	struct mlx5e_ipsec_work *work;
+	struct mlx5e_ipsec_dwork *dwork;
+	struct mlx5e_ipsec_limits limits;
 };
 
 struct mlx5_accel_pol_xfrm_attrs {
@@ -222,6 +242,7 @@ int mlx5e_accel_ipsec_fs_add_rule(struct mlx5e_ipsec_sa_entry *sa_entry);
 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);
 
 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 0539640a4d88..b47794d4146e 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
@@ -926,9 +926,12 @@ static int rx_add_rule(struct mlx5e_ipsec_sa_entry *sa_entry)
 	flow_act.crypto.type = MLX5_FLOW_CONTEXT_ENCRYPT_DECRYPT_TYPE_IPSEC;
 	flow_act.crypto.obj_id = sa_entry->ipsec_obj_id;
 	flow_act.flags |= FLOW_ACT_NO_APPEND;
-	flow_act.action |= MLX5_FLOW_CONTEXT_ACTION_FWD_DEST |
-			   MLX5_FLOW_CONTEXT_ACTION_CRYPTO_DECRYPT |
+	flow_act.action |= MLX5_FLOW_CONTEXT_ACTION_CRYPTO_DECRYPT |
 			   MLX5_FLOW_CONTEXT_ACTION_COUNT;
+	if (attrs->drop)
+		flow_act.action |= MLX5_FLOW_CONTEXT_ACTION_DROP;
+	else
+		flow_act.action |= MLX5_FLOW_CONTEXT_ACTION_FWD_DEST;
 	dest[0].type = MLX5_FLOW_DESTINATION_TYPE_FLOW_TABLE;
 	dest[0].ft = rx->ft.status;
 	dest[1].type = MLX5_FLOW_DESTINATION_TYPE_COUNTER;
@@ -1018,9 +1021,13 @@ static int tx_add_rule(struct mlx5e_ipsec_sa_entry *sa_entry)
 	flow_act.crypto.type = MLX5_FLOW_CONTEXT_ENCRYPT_DECRYPT_TYPE_IPSEC;
 	flow_act.crypto.obj_id = sa_entry->ipsec_obj_id;
 	flow_act.flags |= FLOW_ACT_NO_APPEND;
-	flow_act.action |= MLX5_FLOW_CONTEXT_ACTION_FWD_DEST |
-			   MLX5_FLOW_CONTEXT_ACTION_CRYPTO_ENCRYPT |
+	flow_act.action |= MLX5_FLOW_CONTEXT_ACTION_CRYPTO_ENCRYPT |
 			   MLX5_FLOW_CONTEXT_ACTION_COUNT;
+	if (attrs->drop)
+		flow_act.action |= MLX5_FLOW_CONTEXT_ACTION_DROP;
+	else
+		flow_act.action |= MLX5_FLOW_CONTEXT_ACTION_FWD_DEST;
+
 	dest[0].ft = tx->ft.status;
 	dest[0].type = MLX5_FLOW_DESTINATION_TYPE_FLOW_TABLE;
 	dest[1].type = MLX5_FLOW_DESTINATION_TYPE_COUNTER;
@@ -1430,3 +1437,19 @@ err_rx_ipv4:
 	kfree(ipsec->tx);
 	return err;
 }
+
+void mlx5e_accel_ipsec_fs_modify(struct mlx5e_ipsec_sa_entry *sa_entry)
+{
+	struct mlx5e_ipsec_sa_entry sa_entry_shadow = {};
+	int err;
+
+	memcpy(&sa_entry_shadow, sa_entry, sizeof(*sa_entry));
+	memset(&sa_entry_shadow.ipsec_rule, 0x00, sizeof(sa_entry->ipsec_rule));
+
+	err = mlx5e_accel_ipsec_fs_add_rule(&sa_entry_shadow);
+	if (err)
+		return;
+
+	mlx5e_accel_ipsec_fs_del_rule(sa_entry);
+	memcpy(sa_entry, &sa_entry_shadow, sizeof(*sa_entry));
+}
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 5342b0b07681..5fddb86bb35e 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
@@ -8,6 +8,7 @@
 
 enum {
 	MLX5_IPSEC_ASO_REMOVE_FLOW_PKT_CNT_OFFSET,
+	MLX5_IPSEC_ASO_REMOVE_FLOW_SOFT_LFT_OFFSET,
 };
 
 u32 mlx5_ipsec_device_caps(struct mlx5_core_dev *mdev)
@@ -75,15 +76,17 @@ static void mlx5e_ipsec_packet_setup(void *obj, u32 pdn,
 	void *aso_ctx;
 
 	aso_ctx = MLX5_ADDR_OF(ipsec_obj, obj, ipsec_aso);
-	if (attrs->esn_trigger) {
+	if (attrs->replay_esn.trigger) {
 		MLX5_SET(ipsec_aso, aso_ctx, esn_event_arm, 1);
 
 		if (attrs->dir == XFRM_DEV_OFFLOAD_IN) {
 			MLX5_SET(ipsec_aso, aso_ctx, window_sz,
-				 attrs->replay_window / 64);
+				 attrs->replay_esn.replay_window / 64);
 			MLX5_SET(ipsec_aso, aso_ctx, mode,
 				 MLX5_IPSEC_ASO_REPLAY_PROTECTION);
-			}
+		}
+		MLX5_SET(ipsec_aso, aso_ctx, mode_parameter,
+			 attrs->replay_esn.esn);
 	}
 
 	/* ASO context */
@@ -100,15 +103,15 @@ static void mlx5e_ipsec_packet_setup(void *obj, u32 pdn,
 	if (attrs->dir == XFRM_DEV_OFFLOAD_OUT)
 		MLX5_SET(ipsec_aso, aso_ctx, mode, MLX5_IPSEC_ASO_INC_SN);
 
-	if (attrs->hard_packet_limit != XFRM_INF) {
+	if (attrs->lft.hard_packet_limit != XFRM_INF) {
 		MLX5_SET(ipsec_aso, aso_ctx, remove_flow_pkt_cnt,
-			 lower_32_bits(attrs->hard_packet_limit));
+			 attrs->lft.hard_packet_limit);
 		MLX5_SET(ipsec_aso, aso_ctx, hard_lft_arm, 1);
 	}
 
-	if (attrs->soft_packet_limit != XFRM_INF) {
+	if (attrs->lft.soft_packet_limit != XFRM_INF) {
 		MLX5_SET(ipsec_aso, aso_ctx, remove_flow_soft_lft,
-			 lower_32_bits(attrs->soft_packet_limit));
+			 attrs->lft.soft_packet_limit);
 
 		MLX5_SET(ipsec_aso, aso_ctx, soft_lft_arm, 1);
 	}
@@ -135,10 +138,10 @@ static int mlx5_create_ipsec_obj(struct mlx5e_ipsec_sa_entry *sa_entry)
 	salt_iv_p = MLX5_ADDR_OF(ipsec_obj, obj, implicit_iv);
 	memcpy(salt_iv_p, &aes_gcm->seq_iv, sizeof(aes_gcm->seq_iv));
 	/* esn */
-	if (attrs->esn_trigger) {
+	if (attrs->replay_esn.trigger) {
 		MLX5_SET(ipsec_obj, obj, esn_en, 1);
-		MLX5_SET(ipsec_obj, obj, esn_msb, attrs->esn);
-		MLX5_SET(ipsec_obj, obj, esn_overlap, attrs->esn_overlap);
+		MLX5_SET(ipsec_obj, obj, esn_msb, attrs->replay_esn.esn_msb);
+		MLX5_SET(ipsec_obj, obj, esn_overlap, attrs->replay_esn.overlap);
 	}
 
 	MLX5_SET(ipsec_obj, obj, dekn, sa_entry->enc_key_id);
@@ -224,9 +227,6 @@ static int mlx5_modify_ipsec_obj(struct mlx5e_ipsec_sa_entry *sa_entry,
 	void *obj;
 	int err;
 
-	if (!attrs->esn_trigger)
-		return 0;
-
 	general_obj_types = MLX5_CAP_GEN_64(mdev, general_obj_types);
 	if (!(general_obj_types & MLX5_HCA_CAP_GENERAL_OBJECT_TYPES_IPSEC))
 		return -EINVAL;
@@ -254,8 +254,8 @@ static int mlx5_modify_ipsec_obj(struct mlx5e_ipsec_sa_entry *sa_entry,
 	MLX5_SET64(ipsec_obj, obj, modify_field_select,
 		   MLX5_MODIFY_IPSEC_BITMASK_ESN_OVERLAP |
 			   MLX5_MODIFY_IPSEC_BITMASK_ESN_MSB);
-	MLX5_SET(ipsec_obj, obj, esn_msb, attrs->esn);
-	MLX5_SET(ipsec_obj, obj, esn_overlap, attrs->esn_overlap);
+	MLX5_SET(ipsec_obj, obj, esn_msb, attrs->replay_esn.esn_msb);
+	MLX5_SET(ipsec_obj, obj, esn_overlap, attrs->replay_esn.overlap);
 
 	/* general object fields set */
 	MLX5_SET(general_obj_in_cmd_hdr, in, opcode, MLX5_CMD_OP_MODIFY_GENERAL_OBJECT);
@@ -275,29 +275,24 @@ void mlx5_accel_esp_modify_xfrm(struct mlx5e_ipsec_sa_entry *sa_entry,
 	memcpy(&sa_entry->attrs, attrs, sizeof(sa_entry->attrs));
 }
 
-static void
-mlx5e_ipsec_aso_update_esn(struct mlx5e_ipsec_sa_entry *sa_entry,
-			   const struct mlx5_accel_esp_xfrm_attrs *attrs)
+static void mlx5e_ipsec_aso_update(struct mlx5e_ipsec_sa_entry *sa_entry,
+				   struct mlx5_wqe_aso_ctrl_seg *data)
 {
-	struct mlx5_wqe_aso_ctrl_seg data = {};
+	data->data_mask_mode = MLX5_ASO_DATA_MASK_MODE_BITWISE_64BIT << 6;
+	data->condition_1_0_operand = MLX5_ASO_ALWAYS_TRUE |
+				      MLX5_ASO_ALWAYS_TRUE << 4;
 
-	data.data_mask_mode = MLX5_ASO_DATA_MASK_MODE_BITWISE_64BIT << 6;
-	data.condition_1_0_operand = MLX5_ASO_ALWAYS_TRUE | MLX5_ASO_ALWAYS_TRUE
-								    << 4;
-	data.data_offset_condition_operand = MLX5_IPSEC_ASO_REMOVE_FLOW_PKT_CNT_OFFSET;
-	data.bitwise_data = cpu_to_be64(BIT_ULL(54));
-	data.data_mask = data.bitwise_data;
-
-	mlx5e_ipsec_aso_query(sa_entry, &data);
+	mlx5e_ipsec_aso_query(sa_entry, data);
 }
 
 static void mlx5e_ipsec_update_esn_state(struct mlx5e_ipsec_sa_entry *sa_entry,
 					 u32 mode_param)
 {
 	struct mlx5_accel_esp_xfrm_attrs attrs = {};
+	struct mlx5_wqe_aso_ctrl_seg data = {};
 
 	if (mode_param < MLX5E_IPSEC_ESN_SCOPE_MID) {
-		sa_entry->esn_state.esn++;
+		sa_entry->esn_state.esn_msb++;
 		sa_entry->esn_state.overlap = 0;
 	} else {
 		sa_entry->esn_state.overlap = 1;
@@ -305,25 +300,129 @@ static void mlx5e_ipsec_update_esn_state(struct mlx5e_ipsec_sa_entry *sa_entry,
 
 	mlx5e_ipsec_build_accel_xfrm_attrs(sa_entry, &attrs);
 	mlx5_accel_esp_modify_xfrm(sa_entry, &attrs);
-	mlx5e_ipsec_aso_update_esn(sa_entry, &attrs);
+
+	data.data_offset_condition_operand =
+		MLX5_IPSEC_ASO_REMOVE_FLOW_PKT_CNT_OFFSET;
+	data.bitwise_data = cpu_to_be64(BIT_ULL(54));
+	data.data_mask = data.bitwise_data;
+
+	mlx5e_ipsec_aso_update(sa_entry, &data);
+}
+
+static void mlx5e_ipsec_aso_update_hard(struct mlx5e_ipsec_sa_entry *sa_entry)
+{
+	struct mlx5_wqe_aso_ctrl_seg data = {};
+
+	data.data_offset_condition_operand =
+		MLX5_IPSEC_ASO_REMOVE_FLOW_PKT_CNT_OFFSET;
+	data.bitwise_data = cpu_to_be64(BIT_ULL(57) + BIT_ULL(31));
+	data.data_mask = data.bitwise_data;
+	mlx5e_ipsec_aso_update(sa_entry, &data);
+}
+
+static void mlx5e_ipsec_aso_update_soft(struct mlx5e_ipsec_sa_entry *sa_entry,
+					u32 val)
+{
+	struct mlx5_wqe_aso_ctrl_seg data = {};
+
+	data.data_offset_condition_operand =
+		MLX5_IPSEC_ASO_REMOVE_FLOW_SOFT_LFT_OFFSET;
+	data.bitwise_data = cpu_to_be64(val);
+	data.data_mask = cpu_to_be64(U32_MAX);
+	mlx5e_ipsec_aso_update(sa_entry, &data);
+}
+
+static void mlx5e_ipsec_handle_limits(struct mlx5e_ipsec_sa_entry *sa_entry)
+{
+	struct mlx5_accel_esp_xfrm_attrs *attrs = &sa_entry->attrs;
+	struct mlx5e_ipsec *ipsec = sa_entry->ipsec;
+	struct mlx5e_ipsec_aso *aso = ipsec->aso;
+	bool soft_arm, hard_arm;
+	u64 hard_cnt;
+
+	lockdep_assert_held(&sa_entry->x->lock);
+
+	soft_arm = !MLX5_GET(ipsec_aso, aso->ctx, soft_lft_arm);
+	hard_arm = !MLX5_GET(ipsec_aso, aso->ctx, hard_lft_arm);
+	if (!soft_arm && !hard_arm)
+		/* It is not lifetime event */
+		return;
+
+	hard_cnt = MLX5_GET(ipsec_aso, aso->ctx, remove_flow_pkt_cnt);
+	if (!hard_cnt || hard_arm) {
+		/* It is possible to see packet counter equal to zero without
+		 * hard limit event armed. Such situation can be if packet
+		 * decreased, while we handled soft limit event.
+		 *
+		 * However it will be HW/FW bug if hard limit event is raised
+		 * and packet counter is not zero.
+		 */
+		WARN_ON_ONCE(hard_arm && hard_cnt);
+
+		/* Notify about hard limit */
+		xfrm_state_check_expire(sa_entry->x);
+		return;
+	}
+
+	/* We are in soft limit event. */
+	if (!sa_entry->limits.soft_limit_hit &&
+	    sa_entry->limits.round == attrs->lft.numb_rounds_soft) {
+		sa_entry->limits.soft_limit_hit = true;
+		/* Notify about soft limit */
+		xfrm_state_check_expire(sa_entry->x);
+
+		if (sa_entry->limits.round == attrs->lft.numb_rounds_hard)
+			goto hard;
+
+		if (attrs->lft.soft_packet_limit > BIT_ULL(31)) {
+			/* We cannot avoid a soft_value that might have the high
+			 * bit set. For instance soft_value=2^31+1 cannot be
+			 * adjusted to the low bit clear version of soft_value=1
+			 * because it is too close to 0.
+			 *
+			 * Thus we have this corner case where we can hit the
+			 * soft_limit with the high bit set, but cannot adjust
+			 * the counter. Thus we set a temporary interrupt_value
+			 * at least 2^30 away from here and do the adjustment
+			 * then.
+			 */
+			mlx5e_ipsec_aso_update_soft(sa_entry,
+						    BIT_ULL(31) - BIT_ULL(30));
+			sa_entry->limits.fix_limit = true;
+			return;
+		}
+
+		sa_entry->limits.fix_limit = true;
+	}
+
+hard:
+	if (sa_entry->limits.round == attrs->lft.numb_rounds_hard) {
+		mlx5e_ipsec_aso_update_soft(sa_entry, 0);
+		attrs->lft.soft_packet_limit = XFRM_INF;
+		return;
+	}
+
+	mlx5e_ipsec_aso_update_hard(sa_entry);
+	sa_entry->limits.round++;
+	if (sa_entry->limits.round == attrs->lft.numb_rounds_soft)
+		mlx5e_ipsec_aso_update_soft(sa_entry,
+					    attrs->lft.soft_packet_limit);
+	if (sa_entry->limits.fix_limit) {
+		sa_entry->limits.fix_limit = false;
+		mlx5e_ipsec_aso_update_soft(sa_entry, BIT_ULL(31) - 1);
+	}
 }
 
 static void mlx5e_ipsec_handle_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->data;
 	struct mlx5_accel_esp_xfrm_attrs *attrs;
-	struct mlx5e_ipsec_sa_entry *sa_entry;
 	struct mlx5e_ipsec_aso *aso;
-	struct mlx5e_ipsec *ipsec;
 	int ret;
 
-	sa_entry = xa_load(&work->ipsec->sadb, work->id);
-	if (!sa_entry)
-		goto out;
-
-	ipsec = sa_entry->ipsec;
-	aso = ipsec->aso;
+	aso = sa_entry->ipsec->aso;
 	attrs = &sa_entry->attrs;
 
 	spin_lock(&sa_entry->x->lock);
@@ -331,21 +430,18 @@ static void mlx5e_ipsec_handle_event(struct work_struct *_work)
 	if (ret)
 		goto unlock;
 
-	if (attrs->esn_trigger &&
+	if (attrs->replay_esn.trigger &&
 	    !MLX5_GET(ipsec_aso, aso->ctx, esn_event_arm)) {
 		u32 mode_param = MLX5_GET(ipsec_aso, aso->ctx, mode_parameter);
 
 		mlx5e_ipsec_update_esn_state(sa_entry, mode_param);
 	}
 
-	if (attrs->soft_packet_limit != XFRM_INF)
-		if (!MLX5_GET(ipsec_aso, aso->ctx, soft_lft_arm) ||
-		    !MLX5_GET(ipsec_aso, aso->ctx, hard_lft_arm))
-			xfrm_state_check_expire(sa_entry->x);
+	if (attrs->lft.soft_packet_limit != XFRM_INF)
+		mlx5e_ipsec_handle_limits(sa_entry);
 
 unlock:
 	spin_unlock(&sa_entry->x->lock);
-out:
 	kfree(work);
 }
 
@@ -353,6 +449,7 @@ static int mlx5e_ipsec_event(struct notifier_block *nb, unsigned long event,
 			     void *data)
 {
 	struct mlx5e_ipsec *ipsec = container_of(nb, struct mlx5e_ipsec, nb);
+	struct mlx5e_ipsec_sa_entry *sa_entry;
 	struct mlx5_eqe_obj_change *object;
 	struct mlx5e_ipsec_work *work;
 	struct mlx5_eqe *eqe = data;
@@ -367,13 +464,16 @@ static int mlx5e_ipsec_event(struct notifier_block *nb, unsigned long event,
 	if (type != MLX5_GENERAL_OBJECT_TYPES_IPSEC)
 		return NOTIFY_DONE;
 
+	sa_entry = xa_load(&ipsec->sadb, be32_to_cpu(object->obj_id));
+	if (!sa_entry)
+		return NOTIFY_DONE;
+
 	work = kmalloc(sizeof(*work), GFP_ATOMIC);
 	if (!work)
 		return NOTIFY_DONE;
 
 	INIT_WORK(&work->work, mlx5e_ipsec_handle_event);
-	work->ipsec = ipsec;
-	work->id = be32_to_cpu(object->obj_id);
+	work->data = sa_entry;
 
 	queue_work(ipsec->wq, &work->work);
 	return NOTIFY_OK;
@@ -464,6 +564,7 @@ int mlx5e_ipsec_aso_query(struct mlx5e_ipsec_sa_entry *sa_entry,
 	struct mlx5_wqe_aso_ctrl_seg *ctrl;
 	struct mlx5e_hw_objs *res;
 	struct mlx5_aso_wqe *wqe;
+	unsigned long expires;
 	u8 ds_cnt;
 	int ret;
 
@@ -485,7 +586,12 @@ int mlx5e_ipsec_aso_query(struct mlx5e_ipsec_sa_entry *sa_entry,
 	mlx5e_ipsec_aso_copy(ctrl, data);
 
 	mlx5_aso_post_wqe(aso->aso, false, &wqe->ctrl);
-	ret = mlx5_aso_poll_cq(aso->aso, false);
+	expires = jiffies + msecs_to_jiffies(10);
+	do {
+		ret = mlx5_aso_poll_cq(aso->aso, false);
+		if (ret)
+			usleep_range(2, 10);
+	} while (ret && time_is_after_jiffies(expires));
 	spin_unlock_bh(&aso->lock);
 	return ret;
 }