Merge branch 'mlx5e-use-tls-tx-pool-to-improve-connection-rate'

Tariq Toukan says:

====================
mlx5e use TLS TX pool to improve connection rate

To offload encryption operations, the mlx5 device maintains state and
keeps track of every kTLS device-offloaded connection.  Two HW objects
are used per TX context of a kTLS offloaded connection: a. Transport
interface send (TIS) object, to reach the HW context.  b. Data Encryption
Key (DEK) to perform the crypto operations.

These two objects are created and destroyed per TLS TX context, via FW
commands.  In total, 4 FW commands are issued per TLS TX context, which
seriously limits the connection rate.

In this series, we aim to save creation and destroy of TIS objects by
recycling them.  Upon recycling of a TIS, the HW still needs to be
notified for the re-mapping between a TIS and a context. This is done by
posting WQEs via an SQ, significantly faster API than the FW command
interface.

A pool is used for recycling. The pool dynamically interacts to the load
and connection rate, growing and shrinking accordingly.

Saving the TIS FW commands per context increases connection rate by ~42%,
from 11.6K to 16.5K connections per sec.

Connection rate is still limited by FW bottleneck due to the remaining
per context FW commands (DEK create/destroy). This will soon be addressed
in a followup series.  By combining the two series, the FW bottleneck
will be released, and a significantly higher (about 100K connections per
sec) kTLS TX device-offloaded connection rate is reached.
====================

Link: https://lore.kernel.org/r/20220727094346.10540-1-tariqt@nvidia.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>

7 files changed, 527 insertions, 102 deletions

diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en_accel/en_accel.h b/drivers/net/ethernet/mellanox/mlx5/core/en_accel/en_accel.h
index 04c0a5e1c89a..1839f1ab1ddd 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/en_accel/en_accel.h
+++ b/drivers/net/ethernet/mellanox/mlx5/core/en_accel/en_accel.h
@@ -194,4 +194,14 @@ static inline void mlx5e_accel_cleanup_rx(struct mlx5e_priv *priv)
 {
 	mlx5e_ktls_cleanup_rx(priv);
 }
+
+static inline int mlx5e_accel_init_tx(struct mlx5e_priv *priv)
+{
+	return mlx5e_ktls_init_tx(priv);
+}
+
+static inline void mlx5e_accel_cleanup_tx(struct mlx5e_priv *priv)
+{
+	mlx5e_ktls_cleanup_tx(priv);
+}
 #endif /* __MLX5E_EN_ACCEL_H__ */
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en_accel/ktls.h b/drivers/net/ethernet/mellanox/mlx5/core/en_accel/ktls.h
index d016624fbc9d..948400dee525 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/en_accel/ktls.h
+++ b/drivers/net/ethernet/mellanox/mlx5/core/en_accel/ktls.h
@@ -42,6 +42,8 @@ static inline bool mlx5e_ktls_type_check(struct mlx5_core_dev *mdev,
 }
 
 void mlx5e_ktls_build_netdev(struct mlx5e_priv *priv);
+int mlx5e_ktls_init_tx(struct mlx5e_priv *priv);
+void mlx5e_ktls_cleanup_tx(struct mlx5e_priv *priv);
 int mlx5e_ktls_init_rx(struct mlx5e_priv *priv);
 void mlx5e_ktls_cleanup_rx(struct mlx5e_priv *priv);
 int mlx5e_ktls_set_feature_rx(struct net_device *netdev, bool enable);
@@ -62,6 +64,8 @@ static inline bool mlx5e_is_ktls_rx(struct mlx5_core_dev *mdev)
 struct mlx5e_tls_sw_stats {
 	atomic64_t tx_tls_ctx;
 	atomic64_t tx_tls_del;
+	atomic64_t tx_tls_pool_alloc;
+	atomic64_t tx_tls_pool_free;
 	atomic64_t rx_tls_ctx;
 	atomic64_t rx_tls_del;
 };
@@ -69,6 +73,7 @@ struct mlx5e_tls_sw_stats {
 struct mlx5e_tls {
 	struct mlx5e_tls_sw_stats sw_stats;
 	struct workqueue_struct *rx_wq;
+	struct mlx5e_tls_tx_pool *tx_pool;
 };
 
 int mlx5e_ktls_init(struct mlx5e_priv *priv);
@@ -83,6 +88,15 @@ static inline void mlx5e_ktls_build_netdev(struct mlx5e_priv *priv)
 {
 }
 
+static inline int mlx5e_ktls_init_tx(struct mlx5e_priv *priv)
+{
+	return 0;
+}
+
+static inline void mlx5e_ktls_cleanup_tx(struct mlx5e_priv *priv)
+{
+}
+
 static inline int mlx5e_ktls_init_rx(struct mlx5e_priv *priv)
 {
 	return 0;
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en_accel/ktls_stats.c b/drivers/net/ethernet/mellanox/mlx5/core/en_accel/ktls_stats.c
index 2ab46c4247ff..7c1c0eb16787 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/en_accel/ktls_stats.c
+++ b/drivers/net/ethernet/mellanox/mlx5/core/en_accel/ktls_stats.c
@@ -41,6 +41,8 @@
 static const struct counter_desc mlx5e_ktls_sw_stats_desc[] = {
 	{ MLX5E_DECLARE_STAT(struct mlx5e_tls_sw_stats, tx_tls_ctx) },
 	{ MLX5E_DECLARE_STAT(struct mlx5e_tls_sw_stats, tx_tls_del) },
+	{ MLX5E_DECLARE_STAT(struct mlx5e_tls_sw_stats, tx_tls_pool_alloc) },
+	{ MLX5E_DECLARE_STAT(struct mlx5e_tls_sw_stats, tx_tls_pool_free) },
 	{ MLX5E_DECLARE_STAT(struct mlx5e_tls_sw_stats, rx_tls_ctx) },
 	{ MLX5E_DECLARE_STAT(struct mlx5e_tls_sw_stats, rx_tls_del) },
 };
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en_accel/ktls_tx.c b/drivers/net/ethernet/mellanox/mlx5/core/en_accel/ktls_tx.c
index fba21edf88d8..6b6c7044b64a 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/en_accel/ktls_tx.c
+++ b/drivers/net/ethernet/mellanox/mlx5/core/en_accel/ktls_tx.c
@@ -35,30 +35,70 @@ u16 mlx5e_ktls_get_stop_room(struct mlx5_core_dev *mdev, struct mlx5e_params *pa
 	stop_room += mlx5e_stop_room_for_wqe(mdev, MLX5E_TLS_SET_STATIC_PARAMS_WQEBBS);
 	stop_room += mlx5e_stop_room_for_wqe(mdev, MLX5E_TLS_SET_PROGRESS_PARAMS_WQEBBS);
 	stop_room += num_dumps * mlx5e_stop_room_for_wqe(mdev, MLX5E_KTLS_DUMP_WQEBBS);
+	stop_room += 1; /* fence nop */
 
 	return stop_room;
 }
 
+static void mlx5e_ktls_set_tisc(struct mlx5_core_dev *mdev, void *tisc)
+{
+	MLX5_SET(tisc, tisc, tls_en, 1);
+	MLX5_SET(tisc, tisc, pd, mdev->mlx5e_res.hw_objs.pdn);
+	MLX5_SET(tisc, tisc, transport_domain, mdev->mlx5e_res.hw_objs.td.tdn);
+}
+
 static int mlx5e_ktls_create_tis(struct mlx5_core_dev *mdev, u32 *tisn)
 {
 	u32 in[MLX5_ST_SZ_DW(create_tis_in)] = {};
-	void *tisc;
 
-	tisc = MLX5_ADDR_OF(create_tis_in, in, ctx);
+	mlx5e_ktls_set_tisc(mdev, MLX5_ADDR_OF(create_tis_in, in, ctx));
 
-	MLX5_SET(tisc, tisc, tls_en, 1);
+	return mlx5_core_create_tis(mdev, in, tisn);
+}
+
+static int mlx5e_ktls_create_tis_cb(struct mlx5_core_dev *mdev,
+				    struct mlx5_async_ctx *async_ctx,
+				    u32 *out, int outlen,
+				    mlx5_async_cbk_t callback,
+				    struct mlx5_async_work *context)
+{
+	u32 in[MLX5_ST_SZ_DW(create_tis_in)] = {};
+
+	mlx5e_ktls_set_tisc(mdev, MLX5_ADDR_OF(create_tis_in, in, ctx));
+	MLX5_SET(create_tis_in, in, opcode, MLX5_CMD_OP_CREATE_TIS);
+
+	return mlx5_cmd_exec_cb(async_ctx, in, sizeof(in),
+				out, outlen, callback, context);
+}
+
+static int mlx5e_ktls_destroy_tis_cb(struct mlx5_core_dev *mdev, u32 tisn,
+				     struct mlx5_async_ctx *async_ctx,
+				     u32 *out, int outlen,
+				     mlx5_async_cbk_t callback,
+				     struct mlx5_async_work *context)
+{
+	u32 in[MLX5_ST_SZ_DW(destroy_tis_in)] = {};
+
+	MLX5_SET(destroy_tis_in, in, opcode, MLX5_CMD_OP_DESTROY_TIS);
+	MLX5_SET(destroy_tis_in, in, tisn, tisn);
 
-	return mlx5e_create_tis(mdev, in, tisn);
+	return mlx5_cmd_exec_cb(async_ctx, in, sizeof(in),
+				out, outlen, callback, context);
 }
 
 struct mlx5e_ktls_offload_context_tx {
-	struct tls_offload_context_tx *tx_ctx;
-	struct tls12_crypto_info_aes_gcm_128 crypto_info;
-	struct mlx5e_tls_sw_stats *sw_stats;
+	/* fast path */
 	u32 expected_seq;
 	u32 tisn;
-	u32 key_id;
 	bool ctx_post_pending;
+	/* control / resync */
+	struct list_head list_node; /* member of the pool */
+	struct tls12_crypto_info_aes_gcm_128 crypto_info;
+	struct tls_offload_context_tx *tx_ctx;
+	struct mlx5_core_dev *mdev;
+	struct mlx5e_tls_sw_stats *sw_stats;
+	u32 key_id;
+	u8 create_err : 1;
 };
 
 static void
@@ -82,28 +122,368 @@ mlx5e_get_ktls_tx_priv_ctx(struct tls_context *tls_ctx)
 	return *ctx;
 }
 
+/* struct for callback API management */
+struct mlx5e_async_ctx {
+	struct mlx5_async_work context;
+	struct mlx5_async_ctx async_ctx;
+	struct work_struct work;
+	struct mlx5e_ktls_offload_context_tx *priv_tx;
+	struct completion complete;
+	int err;
+	union {
+		u32 out_create[MLX5_ST_SZ_DW(create_tis_out)];
+		u32 out_destroy[MLX5_ST_SZ_DW(destroy_tis_out)];
+	};
+};
+
+static struct mlx5e_async_ctx *mlx5e_bulk_async_init(struct mlx5_core_dev *mdev, int n)
+{
+	struct mlx5e_async_ctx *bulk_async;
+	int i;
+
+	bulk_async = kvcalloc(n, sizeof(struct mlx5e_async_ctx), GFP_KERNEL);
+	if (!bulk_async)
+		return NULL;
+
+	for (i = 0; i < n; i++) {
+		struct mlx5e_async_ctx *async = &bulk_async[i];
+
+		mlx5_cmd_init_async_ctx(mdev, &async->async_ctx);
+		init_completion(&async->complete);
+	}
+
+	return bulk_async;
+}
+
+static void mlx5e_bulk_async_cleanup(struct mlx5e_async_ctx *bulk_async, int n)
+{
+	int i;
+
+	for (i = 0; i < n; i++) {
+		struct mlx5e_async_ctx *async = &bulk_async[i];
+
+		mlx5_cmd_cleanup_async_ctx(&async->async_ctx);
+	}
+	kvfree(bulk_async);
+}
+
+static void create_tis_callback(int status, struct mlx5_async_work *context)
+{
+	struct mlx5e_async_ctx *async =
+		container_of(context, struct mlx5e_async_ctx, context);
+	struct mlx5e_ktls_offload_context_tx *priv_tx = async->priv_tx;
+
+	if (status) {
+		async->err = status;
+		priv_tx->create_err = 1;
+		goto out;
+	}
+
+	priv_tx->tisn = MLX5_GET(create_tis_out, async->out_create, tisn);
+out:
+	complete(&async->complete);
+}
+
+static void destroy_tis_callback(int status, struct mlx5_async_work *context)
+{
+	struct mlx5e_async_ctx *async =
+		container_of(context, struct mlx5e_async_ctx, context);
+	struct mlx5e_ktls_offload_context_tx *priv_tx = async->priv_tx;
+
+	complete(&async->complete);
+	kfree(priv_tx);
+}
+
+static struct mlx5e_ktls_offload_context_tx *
+mlx5e_tls_priv_tx_init(struct mlx5_core_dev *mdev, struct mlx5e_tls_sw_stats *sw_stats,
+		       struct mlx5e_async_ctx *async)
+{
+	struct mlx5e_ktls_offload_context_tx *priv_tx;
+	int err;
+
+	priv_tx = kzalloc(sizeof(*priv_tx), GFP_KERNEL);
+	if (!priv_tx)
+		return ERR_PTR(-ENOMEM);
+
+	priv_tx->mdev = mdev;
+	priv_tx->sw_stats = sw_stats;
+
+	if (!async) {
+		err = mlx5e_ktls_create_tis(mdev, &priv_tx->tisn);
+		if (err)
+			goto err_out;
+	} else {
+		async->priv_tx = priv_tx;
+		err = mlx5e_ktls_create_tis_cb(mdev, &async->async_ctx,
+					       async->out_create, sizeof(async->out_create),
+					       create_tis_callback, &async->context);
+		if (err)
+			goto err_out;
+	}
+
+	return priv_tx;
+
+err_out:
+	kfree(priv_tx);
+	return ERR_PTR(err);
+}
+
+static void mlx5e_tls_priv_tx_cleanup(struct mlx5e_ktls_offload_context_tx *priv_tx,
+				      struct mlx5e_async_ctx *async)
+{
+	if (priv_tx->create_err) {
+		complete(&async->complete);
+		kfree(priv_tx);
+		return;
+	}
+	async->priv_tx = priv_tx;
+	mlx5e_ktls_destroy_tis_cb(priv_tx->mdev, priv_tx->tisn,
+				  &async->async_ctx,
+				  async->out_destroy, sizeof(async->out_destroy),
+				  destroy_tis_callback, &async->context);
+}
+
+static void mlx5e_tls_priv_tx_list_cleanup(struct mlx5_core_dev *mdev,
+					   struct list_head *list, int size)
+{
+	struct mlx5e_ktls_offload_context_tx *obj;
+	struct mlx5e_async_ctx *bulk_async;
+	int i;
+
+	bulk_async = mlx5e_bulk_async_init(mdev, size);
+	if (!bulk_async)
+		return;
+
+	i = 0;
+	list_for_each_entry(obj, list, list_node) {
+		mlx5e_tls_priv_tx_cleanup(obj, &bulk_async[i]);
+		i++;
+	}
+
+	for (i = 0; i < size; i++) {
+		struct mlx5e_async_ctx *async = &bulk_async[i];
+
+		wait_for_completion(&async->complete);
+	}
+	mlx5e_bulk_async_cleanup(bulk_async, size);
+}
+
+/* Recycling pool API */
+
+#define MLX5E_TLS_TX_POOL_BULK (16)
+#define MLX5E_TLS_TX_POOL_HIGH (4 * 1024)
+#define MLX5E_TLS_TX_POOL_LOW (MLX5E_TLS_TX_POOL_HIGH / 4)
+
+struct mlx5e_tls_tx_pool {
+	struct mlx5_core_dev *mdev;
+	struct mlx5e_tls_sw_stats *sw_stats;
+	struct mutex lock; /* Protects access to the pool */
+	struct list_head list;
+	size_t size;
+
+	struct workqueue_struct *wq;
+	struct work_struct create_work;
+	struct work_struct destroy_work;
+};
+
+static void create_work(struct work_struct *work)
+{
+	struct mlx5e_tls_tx_pool *pool =
+		container_of(work, struct mlx5e_tls_tx_pool, create_work);
+	struct mlx5e_ktls_offload_context_tx *obj;
+	struct mlx5e_async_ctx *bulk_async;
+	LIST_HEAD(local_list);
+	int i, j, err = 0;
+
+	bulk_async = mlx5e_bulk_async_init(pool->mdev, MLX5E_TLS_TX_POOL_BULK);
+	if (!bulk_async)
+		return;
+
+	for (i = 0; i < MLX5E_TLS_TX_POOL_BULK; i++) {
+		obj = mlx5e_tls_priv_tx_init(pool->mdev, pool->sw_stats, &bulk_async[i]);
+		if (IS_ERR(obj)) {
+			err = PTR_ERR(obj);
+			break;
+		}
+		list_add(&obj->list_node, &local_list);
+	}
+
+	for (j = 0; j < i; j++) {
+		struct mlx5e_async_ctx *async = &bulk_async[j];
+
+		wait_for_completion(&async->complete);
+		if (!err && async->err)
+			err = async->err;
+	}
+	atomic64_add(i, &pool->sw_stats->tx_tls_pool_alloc);
+	mlx5e_bulk_async_cleanup(bulk_async, MLX5E_TLS_TX_POOL_BULK);
+	if (err)
+		goto err_out;
+
+	mutex_lock(&pool->lock);
+	if (pool->size + MLX5E_TLS_TX_POOL_BULK >= MLX5E_TLS_TX_POOL_HIGH) {
+		mutex_unlock(&pool->lock);
+		goto err_out;
+	}
+	list_splice(&local_list, &pool->list);
+	pool->size += MLX5E_TLS_TX_POOL_BULK;
+	if (pool->size <= MLX5E_TLS_TX_POOL_LOW)
+		queue_work(pool->wq, work);
+	mutex_unlock(&pool->lock);
+	return;
+
+err_out:
+	mlx5e_tls_priv_tx_list_cleanup(pool->mdev, &local_list, i);
+	atomic64_add(i, &pool->sw_stats->tx_tls_pool_free);
+}
+
+static void destroy_work(struct work_struct *work)
+{
+	struct mlx5e_tls_tx_pool *pool =
+		container_of(work, struct mlx5e_tls_tx_pool, destroy_work);
+	struct mlx5e_ktls_offload_context_tx *obj;
+	LIST_HEAD(local_list);
+	int i = 0;
+
+	mutex_lock(&pool->lock);
+	if (pool->size < MLX5E_TLS_TX_POOL_HIGH) {
+		mutex_unlock(&pool->lock);
+		return;
+	}
+
+	list_for_each_entry(obj, &pool->list, list_node)
+		if (++i == MLX5E_TLS_TX_POOL_BULK)
+			break;
+
+	list_cut_position(&local_list, &pool->list, &obj->list_node);
+	pool->size -= MLX5E_TLS_TX_POOL_BULK;
+	if (pool->size >= MLX5E_TLS_TX_POOL_HIGH)
+		queue_work(pool->wq, work);
+	mutex_unlock(&pool->lock);
+
+	mlx5e_tls_priv_tx_list_cleanup(pool->mdev, &local_list, MLX5E_TLS_TX_POOL_BULK);
+	atomic64_add(MLX5E_TLS_TX_POOL_BULK, &pool->sw_stats->tx_tls_pool_free);
+}
+
+static struct mlx5e_tls_tx_pool *mlx5e_tls_tx_pool_init(struct mlx5_core_dev *mdev,
+							struct mlx5e_tls_sw_stats *sw_stats)
+{
+	struct mlx5e_tls_tx_pool *pool;
+
+	BUILD_BUG_ON(MLX5E_TLS_TX_POOL_LOW + MLX5E_TLS_TX_POOL_BULK >= MLX5E_TLS_TX_POOL_HIGH);
+
+	pool = kvzalloc(sizeof(*pool), GFP_KERNEL);
+	if (!pool)
+		return NULL;
+
+	pool->wq = create_singlethread_workqueue("mlx5e_tls_tx_pool");
+	if (!pool->wq)
+		goto err_free;
+
+	INIT_LIST_HEAD(&pool->list);
+	mutex_init(&pool->lock);
+
+	INIT_WORK(&pool->create_work, create_work);
+	INIT_WORK(&pool->destroy_work, destroy_work);
+
+	pool->mdev = mdev;
+	pool->sw_stats = sw_stats;
+
+	return pool;
+
+err_free:
+	kvfree(pool);
+	return NULL;
+}
+
+static void mlx5e_tls_tx_pool_list_cleanup(struct mlx5e_tls_tx_pool *pool)
+{
+	while (pool->size > MLX5E_TLS_TX_POOL_BULK) {
+		struct mlx5e_ktls_offload_context_tx *obj;
+		LIST_HEAD(local_list);
+		int i = 0;
+
+		list_for_each_entry(obj, &pool->list, list_node)
+			if (++i == MLX5E_TLS_TX_POOL_BULK)
+				break;
+
+		list_cut_position(&local_list, &pool->list, &obj->list_node);
+		mlx5e_tls_priv_tx_list_cleanup(pool->mdev, &local_list, MLX5E_TLS_TX_POOL_BULK);
+		atomic64_add(MLX5E_TLS_TX_POOL_BULK, &pool->sw_stats->tx_tls_pool_free);
+		pool->size -= MLX5E_TLS_TX_POOL_BULK;
+	}
+	if (pool->size) {
+		mlx5e_tls_priv_tx_list_cleanup(pool->mdev, &pool->list, pool->size);
+		atomic64_add(pool->size, &pool->sw_stats->tx_tls_pool_free);
+	}
+}
+
+static void mlx5e_tls_tx_pool_cleanup(struct mlx5e_tls_tx_pool *pool)
+{
+	mlx5e_tls_tx_pool_list_cleanup(pool);
+	destroy_workqueue(pool->wq);
+	kvfree(pool);
+}
+
+static void pool_push(struct mlx5e_tls_tx_pool *pool, struct mlx5e_ktls_offload_context_tx *obj)
+{
+	mutex_lock(&pool->lock);
+	list_add(&obj->list_node, &pool->list);
+	if (++pool->size == MLX5E_TLS_TX_POOL_HIGH)
+		queue_work(pool->wq, &pool->destroy_work);
+	mutex_unlock(&pool->lock);
+}
+
+static struct mlx5e_ktls_offload_context_tx *pool_pop(struct mlx5e_tls_tx_pool *pool)
+{
+	struct mlx5e_ktls_offload_context_tx *obj;
+
+	mutex_lock(&pool->lock);
+	if (unlikely(pool->size == 0)) {
+		/* pool is empty:
+		 * - trigger the populating work, and
+		 * - serve the current context via the regular blocking api.
+		 */
+		queue_work(pool->wq, &pool->create_work);
+		mutex_unlock(&pool->lock);
+		obj = mlx5e_tls_priv_tx_init(pool->mdev, pool->sw_stats, NULL);
+		if (!IS_ERR(obj))
+			atomic64_inc(&pool->sw_stats->tx_tls_pool_alloc);
+		return obj;
+	}
+
+	obj = list_first_entry(&pool->list, struct mlx5e_ktls_offload_context_tx,
+			       list_node);
+	list_del(&obj->list_node);
+	if (--pool->size == MLX5E_TLS_TX_POOL_LOW)
+		queue_work(pool->wq, &pool->create_work);
+	mutex_unlock(&pool->lock);
+	return obj;
+}
+
+/* End of pool API */
+
 int mlx5e_ktls_add_tx(struct net_device *netdev, struct sock *sk,
 		      struct tls_crypto_info *crypto_info, u32 start_offload_tcp_sn)
 {
 	struct mlx5e_ktls_offload_context_tx *priv_tx;
+	struct mlx5e_tls_tx_pool *pool;
 	struct tls_context *tls_ctx;
-	struct mlx5_core_dev *mdev;
 	struct mlx5e_priv *priv;
 	int err;
 
 	tls_ctx = tls_get_ctx(sk);
 	priv = netdev_priv(netdev);
-	mdev = priv->mdev;
+	pool = priv->tls->tx_pool;
 
-	priv_tx = kzalloc(sizeof(*priv_tx), GFP_KERNEL);
-	if (!priv_tx)
-		return -ENOMEM;
+	priv_tx = pool_pop(pool);
+	if (IS_ERR(priv_tx))
+		return PTR_ERR(priv_tx);
 
-	err = mlx5_ktls_create_key(mdev, crypto_info, &priv_tx->key_id);
+	err = mlx5_ktls_create_key(pool->mdev, crypto_info, &priv_tx->key_id);
 	if (err)
 		goto err_create_key;
 
-	priv_tx->sw_stats = &priv->tls->sw_stats;
 	priv_tx->expected_seq = start_offload_tcp_sn;
 	priv_tx->crypto_info  =
 		*(struct tls12_crypto_info_aes_gcm_128 *)crypto_info;
@@ -111,36 +491,29 @@ int mlx5e_ktls_add_tx(struct net_device *netdev, struct sock *sk,
 
 	mlx5e_set_ktls_tx_priv_ctx(tls_ctx, priv_tx);
 
-	err = mlx5e_ktls_create_tis(mdev, &priv_tx->tisn);
-	if (err)
-		goto err_create_tis;
-
 	priv_tx->ctx_post_pending = true;
 	atomic64_inc(&priv_tx->sw_stats->tx_tls_ctx);
 
 	return 0;
 
-err_create_tis:
-	mlx5_ktls_destroy_key(mdev, priv_tx->key_id);
 err_create_key:
-	kfree(priv_tx);
+	pool_push(pool, priv_tx);
 	return err;
 }
 
 void mlx5e_ktls_del_tx(struct net_device *netdev, struct tls_context *tls_ctx)
 {
 	struct mlx5e_ktls_offload_context_tx *priv_tx;
-	struct mlx5_core_dev *mdev;
+	struct mlx5e_tls_tx_pool *pool;
 	struct mlx5e_priv *priv;
 
 	priv_tx = mlx5e_get_ktls_tx_priv_ctx(tls_ctx);
 	priv = netdev_priv(netdev);
-	mdev = priv->mdev;
+	pool = priv->tls->tx_pool;
 
 	atomic64_inc(&priv_tx->sw_stats->tx_tls_del);
-	mlx5e_destroy_tis(mdev, priv_tx->tisn);
-	mlx5_ktls_destroy_key(mdev, priv_tx->key_id);
-	kfree(priv_tx);
+	mlx5_ktls_destroy_key(priv_tx->mdev, priv_tx->key_id);
+	pool_push(pool, priv_tx);
 }
 
 static void tx_fill_wi(struct mlx5e_txqsq *sq,
@@ -201,6 +574,16 @@ post_progress_params(struct mlx5e_txqsq *sq,
 	sq->pc += num_wqebbs;
 }
 
+static void tx_post_fence_nop(struct mlx5e_txqsq *sq)
+{
+	struct mlx5_wq_cyc *wq = &sq->wq;
+	u16 pi = mlx5_wq_cyc_ctr2ix(wq, sq->pc);
+
+	tx_fill_wi(sq, pi, 1, 0, NULL);
+
+	mlx5e_post_nop_fence(wq, sq->sqn, &sq->pc);
+}
+
 static void
 mlx5e_ktls_tx_post_param_wqes(struct mlx5e_txqsq *sq,
 			      struct mlx5e_ktls_offload_context_tx *priv_tx,
@@ -212,6 +595,7 @@ mlx5e_ktls_tx_post_param_wqes(struct mlx5e_txqsq *sq,
 		post_static_params(sq, priv_tx, fence_first_post);
 
 	post_progress_params(sq, priv_tx, progress_fence);
+	tx_post_fence_nop(sq);
 }
 
 struct tx_sync_info {
@@ -304,7 +688,7 @@ tx_post_resync_params(struct mlx5e_txqsq *sq,
 }
 
 static int
-tx_post_resync_dump(struct mlx5e_txqsq *sq, skb_frag_t *frag, u32 tisn, bool first)
+tx_post_resync_dump(struct mlx5e_txqsq *sq, skb_frag_t *frag, u32 tisn)
 {
 	struct mlx5_wqe_ctrl_seg *cseg;
 	struct mlx5_wqe_data_seg *dseg;
@@ -326,7 +710,6 @@ tx_post_resync_dump(struct mlx5e_txqsq *sq, skb_frag_t *frag, u32 tisn, bool fir
 	cseg->opmod_idx_opcode = cpu_to_be32((sq->pc << 8)  | MLX5_OPCODE_DUMP);
 	cseg->qpn_ds           = cpu_to_be32((sq->sqn << 8) | ds_cnt);
 	cseg->tis_tir_num      = cpu_to_be32(tisn << 8);
-	cseg->fm_ce_se         = first ? MLX5_FENCE_MODE_INITIATOR_SMALL : 0;
 
 	fsz = skb_frag_size(frag);
 	dma_addr = skb_frag_dma_map(sq->pdev, frag, 0, fsz,
@@ -361,67 +744,39 @@ void mlx5e_ktls_tx_handle_resync_dump_comp(struct mlx5e_txqsq *sq,
 	stats->tls_dump_bytes += wi->num_bytes;
 }
 
-static void tx_post_fence_nop(struct mlx5e_txqsq *sq)
-{
-	struct mlx5_wq_cyc *wq = &sq->wq;
-	u16 pi = mlx5_wq_cyc_ctr2ix(wq, sq->pc);
-
-	tx_fill_wi(sq, pi, 1, 0, NULL);
-
-	mlx5e_post_nop_fence(wq, sq->sqn, &sq->pc);
-}
-
 static enum mlx5e_ktls_sync_retval
 mlx5e_ktls_tx_handle_ooo(struct mlx5e_ktls_offload_context_tx *priv_tx,
 			 struct mlx5e_txqsq *sq,
 			 int datalen,
 			 u32 seq)
 {
-	struct mlx5e_sq_stats *stats = sq->stats;
 	enum mlx5e_ktls_sync_retval ret;
 	struct tx_sync_info info = {};
-	int i = 0;
+	int i;
 
 	ret = tx_sync_info_get(priv_tx, seq, datalen, &info);
-	if (unlikely(ret != MLX5E_KTLS_SYNC_DONE)) {
-		if (ret == MLX5E_KTLS_SYNC_SKIP_NO_DATA) {
-			stats->tls_skip_no_sync_data++;
-			return MLX5E_KTLS_SYNC_SKIP_NO_DATA;
-		}
-		/* We might get here if a retransmission reaches the driver
-		 * after the relevant record is acked.
+	if (unlikely(ret != MLX5E_KTLS_SYNC_DONE))
+		/* We might get here with ret == FAIL if a retransmission
+		 * reaches the driver after the relevant record is acked.
 		 * It should be safe to drop the packet in this case
 		 */
-		stats->tls_drop_no_sync_data++;
-		goto err_out;
-	}
-
-	stats->tls_ooo++;
+		return ret;
 
 	tx_post_resync_params(sq, priv_tx, info.rcd_sn);
 
-	/* If no dump WQE was sent, we need to have a fence NOP WQE before the
-	 * actual data xmit.
-	 */
-	if (!info.nr_frags) {
-		tx_post_fence_nop(sq);
-		return MLX5E_KTLS_SYNC_DONE;
-	}
-
-	for (; i < info.nr_frags; i++) {
+	for (i = 0; i < info.nr_frags; i++) {
 		unsigned int orig_fsz, frag_offset = 0, n = 0;
 		skb_frag_t *f = &info.frags[i];
 
 		orig_fsz = skb_frag_size(f);
 
 		do {
-			bool fence = !(i || frag_offset);
 			unsigned int fsz;
 
 			n++;
 			fsz = min_t(unsigned int, sq->hw_mtu, orig_fsz - frag_offset);
 			skb_frag_size_set(f, fsz);
-			if (tx_post_resync_dump(sq, f, priv_tx->tisn, fence)) {
+			if (tx_post_resync_dump(sq, f, priv_tx->tisn)) {
 				page_ref_add(skb_frag_page(f), n - 1);
 				goto err_out;
 			}
@@ -469,24 +824,27 @@ bool mlx5e_ktls_handle_tx_skb(struct net_device *netdev, struct mlx5e_txqsq *sq,
 
 	priv_tx = mlx5e_get_ktls_tx_priv_ctx(tls_ctx);
 
-	if (unlikely(mlx5e_ktls_tx_offload_test_and_clear_pending(priv_tx))) {
+	if (unlikely(mlx5e_ktls_tx_offload_test_and_clear_pending(priv_tx)))
 		mlx5e_ktls_tx_post_param_wqes(sq, priv_tx, false, false);
-	}
 
 	seq = ntohl(tcp_hdr(skb)->seq);
 	if (unlikely(priv_tx->expected_seq != seq)) {
 		enum mlx5e_ktls_sync_retval ret =
 			mlx5e_ktls_tx_handle_ooo(priv_tx, sq, datalen, seq);
 
+		stats->tls_ooo++;
+
 		switch (ret) {
 		case MLX5E_KTLS_SYNC_DONE:
 			break;
 		case MLX5E_KTLS_SYNC_SKIP_NO_DATA:
+			stats->tls_skip_no_sync_data++;
 			if (likely(!skb->decrypted))
 				goto out;
 			WARN_ON_ONCE(1);
-			fallthrough;
+			goto err_out;
 		case MLX5E_KTLS_SYNC_FAIL:
+			stats->tls_drop_no_sync_data++;
 			goto err_out;
 		}
 	}
@@ -505,3 +863,24 @@ err_out:
 	dev_kfree_skb_any(skb);
 	return false;
 }
+
+int mlx5e_ktls_init_tx(struct mlx5e_priv *priv)
+{
+	if (!mlx5e_is_ktls_tx(priv->mdev))
+		return 0;
+
+	priv->tls->tx_pool = mlx5e_tls_tx_pool_init(priv->mdev, &priv->tls->sw_stats);
+	if (!priv->tls->tx_pool)
+		return -ENOMEM;
+
+	return 0;
+}
+
+void mlx5e_ktls_cleanup_tx(struct mlx5e_priv *priv)
+{
+	if (!mlx5e_is_ktls_tx(priv->mdev))
+		return;
+
+	mlx5e_tls_tx_pool_cleanup(priv->tls->tx_pool);
+	priv->tls->tx_pool = NULL;
+}
diff --git a/drivers/net/ethernet/mellanox/mlx5/core/en_main.c b/drivers/net/ethernet/mellanox/mlx5/core/en_main.c
index 180b2f418339..24ddd438c066 100644
--- a/drivers/net/ethernet/mellanox/mlx5/core/en_main.c
+++ b/drivers/net/ethernet/mellanox/mlx5/core/en_main.c
@@ -3144,6 +3144,7 @@ static void mlx5e_cleanup_nic_tx(struct mlx5e_priv *priv)
 		mlx5e_mqprio_rl_free(priv->mqprio_rl);
 		priv->mqprio_rl = NULL;
 	}
+	mlx5e_accel_cleanup_tx(priv);
 	mlx5e_destroy_tises(priv);
 }
 
@@ -5147,9 +5148,17 @@ static int mlx5e_init_nic_tx(struct mlx5e_priv *priv)
 		return err;
 	}
 
+	err = mlx5e_accel_init_tx(priv);
+	if (err)
+		goto err_destroy_tises;
+
 	mlx5e_set_mqprio_rl(priv);
 	mlx5e_dcbnl_initialize(priv);
 	return 0;
+
+err_destroy_tises:
+	mlx5e_destroy_tises(priv);
+	return err;
 }
 
 static void mlx5e_nic_enable(struct mlx5e_priv *priv)
diff --git a/include/net/tls.h b/include/net/tls.h
index abb050b0df83..b75b5727abdb 100644
--- a/include/net/tls.h
+++ b/include/net/tls.h
@@ -161,6 +161,8 @@ struct tls_offload_context_tx {
 
 	struct scatterlist sg_tx_data[MAX_SKB_FRAGS];
 	void (*sk_destruct)(struct sock *sk);
+	struct work_struct destruct_work;
+	struct tls_context *ctx;
 	u8 driver_state[] __aligned(8);
 	/* The TLS layer reserves room for driver specific state
 	 * Currently the belief is that there is not enough
diff --git a/net/tls/tls_device.c b/net/tls/tls_device.c
index b3e2a30041c6..18c7e5c6d228 100644
--- a/net/tls/tls_device.c
+++ b/net/tls/tls_device.c
@@ -46,10 +46,8 @@
  */
 static DECLARE_RWSEM(device_offload_lock);
 
-static void tls_device_gc_task(struct work_struct *work);
+static struct workqueue_struct *destruct_wq __read_mostly;
 
-static DECLARE_WORK(tls_device_gc_work, tls_device_gc_task);
-static LIST_HEAD(tls_device_gc_list);
 static LIST_HEAD(tls_device_list);
 static LIST_HEAD(tls_device_down_list);
 static DEFINE_SPINLOCK(tls_device_lock);
@@ -68,47 +66,44 @@ static void tls_device_free_ctx(struct tls_context *ctx)
 	tls_ctx_free(NULL, ctx);
 }
 
-static void tls_device_gc_task(struct work_struct *work)
+static void tls_device_tx_del_task(struct work_struct *work)
 {
-	struct tls_context *ctx, *tmp;
-	unsigned long flags;
-	LIST_HEAD(gc_list);
-
-	spin_lock_irqsave(&tls_device_lock, flags);
-	list_splice_init(&tls_device_gc_list, &gc_list);
-	spin_unlock_irqrestore(&tls_device_lock, flags);
-
-	list_for_each_entry_safe(ctx, tmp, &gc_list, list) {
-		struct net_device *netdev = ctx->netdev;
-
-		if (netdev && ctx->tx_conf == TLS_HW) {
-			netdev->tlsdev_ops->tls_dev_del(netdev, ctx,
-							TLS_OFFLOAD_CTX_DIR_TX);
-			dev_put(netdev);
-			ctx->netdev = NULL;
-		}
+	struct tls_offload_context_tx *offload_ctx =
+		container_of(work, struct tls_offload_context_tx, destruct_work);
+	struct tls_context *ctx = offload_ctx->ctx;
+	struct net_device *netdev = ctx->netdev;
 
-		list_del(&ctx->list);
-		tls_device_free_ctx(ctx);
-	}
+	netdev->tlsdev_ops->tls_dev_del(netdev, ctx, TLS_OFFLOAD_CTX_DIR_TX);
+	dev_put(netdev);
+	ctx->netdev = NULL;
+	tls_device_free_ctx(ctx);
 }
 
 static void tls_device_queue_ctx_destruction(struct tls_context *ctx)
 {
 	unsigned long flags;
+	bool async_cleanup;
 
 	spin_lock_irqsave(&tls_device_lock, flags);
-	if (unlikely(!refcount_dec_and_test(&ctx->refcount)))
-		goto unlock;
+	if (unlikely(!refcount_dec_and_test(&ctx->refcount))) {
+		spin_unlock_irqrestore(&tls_device_lock, flags);
+		return;
+	}
 
-	list_move_tail(&ctx->list, &tls_device_gc_list);
+	list_del(&ctx->list); /* Remove from tls_device_list / tls_device_down_list */
+	async_cleanup = ctx->netdev && ctx->tx_conf == TLS_HW;
+	if (async_cleanup) {
+		struct tls_offload_context_tx *offload_ctx = tls_offload_ctx_tx(ctx);
 
-	/* schedule_work inside the spinlock
-	 * to make sure tls_device_down waits for that work.
-	 */
-	schedule_work(&tls_device_gc_work);
-unlock:
+		/* queue_work inside the spinlock
+		 * to make sure tls_device_down waits for that work.
+		 */
+		queue_work(destruct_wq, &offload_ctx->destruct_work);
+	}
 	spin_unlock_irqrestore(&tls_device_lock, flags);
+
+	if (!async_cleanup)
+		tls_device_free_ctx(ctx);
 }
 
 /* We assume that the socket is already connected */
@@ -1150,6 +1145,9 @@ int tls_set_device_offload(struct sock *sk, struct tls_context *ctx)
 	start_marker_record->len = 0;
 	start_marker_record->num_frags = 0;
 
+	INIT_WORK(&offload_ctx->destruct_work, tls_device_tx_del_task);
+	offload_ctx->ctx = ctx;
+
 	INIT_LIST_HEAD(&offload_ctx->records_list);
 	list_add_tail(&start_marker_record->list, &offload_ctx->records_list);
 	spin_lock_init(&offload_ctx->lock);
@@ -1394,7 +1392,7 @@ static int tls_device_down(struct net_device *netdev)
 
 	up_write(&device_offload_lock);
 
-	flush_work(&tls_device_gc_work);
+	flush_workqueue(destruct_wq);
 
 	return NOTIFY_DONE;
 }
@@ -1435,12 +1433,23 @@ static struct notifier_block tls_dev_notifier = {
 
 int __init tls_device_init(void)
 {
-	return register_netdevice_notifier(&tls_dev_notifier);
+	int err;
+
+	destruct_wq = alloc_workqueue("ktls_device_destruct", 0, 0);
+	if (!destruct_wq)
+		return -ENOMEM;
+
+	err = register_netdevice_notifier(&tls_dev_notifier);
+	if (err)
+		destroy_workqueue(destruct_wq);
+
+	return err;
 }
 
 void __exit tls_device_cleanup(void)
 {
 	unregister_netdevice_notifier(&tls_dev_notifier);
-	flush_work(&tls_device_gc_work);
+	flush_workqueue(destruct_wq);
+	destroy_workqueue(destruct_wq);
 	clean_acked_data_flush();
 }