net/mlx5e: Save memory by using dynamic allocation in netdev priv

Many arrays in priv are statically allocated with a pre-defined maximum
(for num channels, num TCs, etc...), that is in some cases significantly
larger than the actual maximum. Examples:
- The more VFs are supported, the less MSIX vectors each of them could
  have. This limits the max_nch for each.
- Systems with limited number of cores or MSIX (< 64).
- Netdev profiles that do not support: QoS (DCB / HTB), PTP TX port
  timestamping.

Here we save some amount of memory by moving several structures
and arrays to follow the actual maximum instead.
This patch also prepares the code for even more savings to follow.

For example, on a system where the maximum num of channel is 8,
the channels stats structs alone go down from 3648*64 = 228 KB to
3648*8 = 28.5 KB per interface.

This is important for environments with high number of VFs/SFs or
limited memory.
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Reviewed-by: Moshe Shemesh <moshe@nvidia.com>
Reviewed-by: Parav Pandit <parav@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>

drivers/net/ethernet/mellanox/mlx5/core/en.h

@@ -145,7 +145,6 @@ struct page_pool;
 
 #define MLX5E_MIN_NUM_CHANNELS         0x1
 #define MLX5E_MAX_NUM_CHANNELS         (MLX5E_INDIR_RQT_SIZE / 2)
-#define MLX5E_MAX_NUM_SQS              (MLX5E_MAX_NUM_CHANNELS * MLX5E_MAX_NUM_TC)
 #define MLX5E_TX_CQ_POLL_BUDGET        128
 #define MLX5E_TX_XSK_POLL_BUDGET       64
 #define MLX5E_SQ_RECOVER_MIN_INTERVAL  500 /* msecs */
@@ -875,10 +874,8 @@ struct mlx5e_trap;
 
 struct mlx5e_priv {
 	/* priv data path fields - start */
-	/* +1 for port ptp ts */
-	struct mlx5e_txqsq *txq2sq[(MLX5E_MAX_NUM_CHANNELS + 1) * MLX5E_MAX_NUM_TC +
-				   MLX5E_QOS_MAX_LEAF_NODES];
-	int channel_tc2realtxq[MLX5E_MAX_NUM_CHANNELS][MLX5E_MAX_NUM_TC];
+	struct mlx5e_txqsq **txq2sq;
+	int **channel_tc2realtxq;
 	int port_ptp_tc2realtxq[MLX5E_MAX_NUM_TC];
 #ifdef CONFIG_MLX5_CORE_EN_DCB
 	struct mlx5e_dcbx_dp       dcbx_dp;
@@ -893,7 +890,7 @@ struct mlx5e_priv {
 	struct mlx5e_channels      channels;
 	u32                        tisn[MLX5_MAX_PORTS][MLX5E_MAX_NUM_TC];
 	struct mlx5e_rx_res       *rx_res;
-	u32                        tx_rates[MLX5E_MAX_NUM_SQS];
+	u32                       *tx_rates;
 
 	struct mlx5e_flow_steering fs;
 
@@ -909,7 +906,7 @@ struct mlx5e_priv {
 	struct net_device         *netdev;
 	struct mlx5e_trap         *en_trap;
 	struct mlx5e_stats         stats;
-	struct mlx5e_channel_stats channel_stats[MLX5E_MAX_NUM_CHANNELS];
+	struct mlx5e_channel_stats *channel_stats;
 	struct mlx5e_channel_stats trap_stats;
 	struct mlx5e_ptp_stats     ptp_stats;
 	u16                        stats_nch;

drivers/net/ethernet/mellanox/mlx5/core/en_main.c

@@ -5130,12 +5130,18 @@ int mlx5e_priv_init(struct mlx5e_priv *priv,
 		    struct net_device *netdev,
 		    struct mlx5_core_dev *mdev)
 {
+	int nch, num_txqs, node, i;
+
+	num_txqs = netdev->num_tx_queues;
+	nch = mlx5e_calc_max_nch(mdev, netdev, profile);
+	node = dev_to_node(mlx5_core_dma_dev(mdev));
+
 	/* priv init */
 	priv->mdev        = mdev;
 	priv->netdev      = netdev;
 	priv->msglevel    = MLX5E_MSG_LEVEL;
-	priv->max_nch     = mlx5e_calc_max_nch(mdev, netdev, profile);
-	priv->stats_nch   = priv->max_nch;
+	priv->max_nch     = nch;
+	priv->stats_nch   = nch;
 	priv->max_opened_tc = 1;
 
 	if (!alloc_cpumask_var(&priv->scratchpad.cpumask, GFP_KERNEL))
@@ -5152,11 +5158,46 @@ int mlx5e_priv_init(struct mlx5e_priv *priv,
 	if (!priv->wq)
 		goto err_free_cpumask;
 
+	priv->txq2sq = kcalloc_node(num_txqs, sizeof(*priv->txq2sq), GFP_KERNEL, node);
+	if (!priv->txq2sq)
+		goto err_destroy_workqueue;
+
+	priv->tx_rates = kcalloc_node(num_txqs, sizeof(*priv->tx_rates), GFP_KERNEL, node);
+	if (!priv->tx_rates)
+		goto err_free_txq2sq;
+
+	priv->channel_tc2realtxq =
+		kcalloc_node(nch, sizeof(*priv->channel_tc2realtxq), GFP_KERNEL, node);
+	if (!priv->channel_tc2realtxq)
+		goto err_free_tx_rates;
+
+	for (i = 0; i < nch; i++) {
+		priv->channel_tc2realtxq[i] =
+			kcalloc_node(profile->max_tc, sizeof(**priv->channel_tc2realtxq),
+				     GFP_KERNEL, node);
+		if (!priv->channel_tc2realtxq[i])
+			goto err_free_channel_tc2realtxq;
+	}
+
+	priv->channel_stats =
+		kcalloc_node(nch, sizeof(*priv->channel_stats), GFP_KERNEL, node);
+	if (!priv->channel_stats)
+		goto err_free_channel_tc2realtxq;
+
 	return 0;
 
+err_free_channel_tc2realtxq:
+	while (--i >= 0)
+		kfree(priv->channel_tc2realtxq[i]);
+	kfree(priv->channel_tc2realtxq);
+err_free_tx_rates:
+	kfree(priv->tx_rates);
+err_free_txq2sq:
+	kfree(priv->txq2sq);
+err_destroy_workqueue:
+	destroy_workqueue(priv->wq);
 err_free_cpumask:
 	free_cpumask_var(priv->scratchpad.cpumask);
-
 	return -ENOMEM;
 }
 
@@ -5168,6 +5209,12 @@ void mlx5e_priv_cleanup(struct mlx5e_priv *priv)
 	if (!priv->mdev)
 		return;
 
+	kfree(priv->channel_stats);
+	for (i = 0; i < priv->max_nch; i++)
+		kfree(priv->channel_tc2realtxq[i]);
+	kfree(priv->channel_tc2realtxq);
+	kfree(priv->tx_rates);
+	kfree(priv->txq2sq);
 	destroy_workqueue(priv->wq);
 	free_cpumask_var(priv->scratchpad.cpumask);