RDMA/mlx5: Add modify_op_stat() support

Add support for ib callback modify_op_stat() to add or remove an optional
counter. When adding, a steering flow table is created with a rule that
catches and counts all the matching packets. When removing, the table and
flow counter are destroyed.

Link: https://lore.kernel.org/r/20211008122439.166063-13-markzhang@nvidia.com
Signed-off-by: Aharon Landau <aharonl@nvidia.com>
Signed-off-by: Leon Romanovsky <leonro@nvidia.com>
Signed-off-by: Mark Zhang <markzhang@nvidia.com>
Signed-off-by: Jason Gunthorpe <jgg@nvidia.com>

drivers/infiniband/hw/mlx5/counters.c

@@ -12,6 +12,7 @@
 struct mlx5_ib_counter {
 	const char *name;
 	size_t offset;
+	u32 type;
 };
 
 #define INIT_Q_COUNTER(_name)		\
@@ -75,19 +76,19 @@ static const struct mlx5_ib_counter ext_ppcnt_cnts[] = {
 	INIT_EXT_PPCNT_COUNTER(rx_icrc_encapsulated),
 };
 
-#define INIT_OP_COUNTER(_name)                                          \
-	{ .name = #_name }
+#define INIT_OP_COUNTER(_name, _type)		\
+	{ .name = #_name, .type = MLX5_IB_OPCOUNTER_##_type}
 
 static const struct mlx5_ib_counter basic_op_cnts[] = {
-	INIT_OP_COUNTER(cc_rx_ce_pkts),
+	INIT_OP_COUNTER(cc_rx_ce_pkts, CC_RX_CE_PKTS),
 };
 
 static const struct mlx5_ib_counter rdmarx_cnp_op_cnts[] = {
-	INIT_OP_COUNTER(cc_rx_cnp_pkts),
+	INIT_OP_COUNTER(cc_rx_cnp_pkts, CC_RX_CNP_PKTS),
 };
 
 static const struct mlx5_ib_counter rdmatx_cnp_op_cnts[] = {
-	INIT_OP_COUNTER(cc_tx_cnp_pkts),
+	INIT_OP_COUNTER(cc_tx_cnp_pkts, CC_TX_CNP_PKTS),
 };
 
 static int mlx5_ib_read_counters(struct ib_counters *counters,
@@ -451,6 +452,7 @@ static void mlx5_ib_fill_counters(struct mlx5_ib_dev *dev,
 	for (i = 0; i < ARRAY_SIZE(basic_op_cnts); i++, j++) {
 		descs[j].name = basic_op_cnts[i].name;
 		descs[j].flags |= IB_STAT_FLAG_OPTIONAL;
+		descs[j].priv = &basic_op_cnts[i].type;
 	}
 
 	if (MLX5_CAP_FLOWTABLE(dev->mdev,
@@ -458,6 +460,7 @@ static void mlx5_ib_fill_counters(struct mlx5_ib_dev *dev,
 		for (i = 0; i < ARRAY_SIZE(rdmarx_cnp_op_cnts); i++, j++) {
 			descs[j].name = rdmarx_cnp_op_cnts[i].name;
 			descs[j].flags |= IB_STAT_FLAG_OPTIONAL;
+			descs[j].priv = &rdmarx_cnp_op_cnts[i].type;
 		}
 	}
 
@@ -466,6 +469,7 @@ static void mlx5_ib_fill_counters(struct mlx5_ib_dev *dev,
 		for (i = 0; i < ARRAY_SIZE(rdmatx_cnp_op_cnts); i++, j++) {
 			descs[j].name = rdmatx_cnp_op_cnts[i].name;
 			descs[j].flags |= IB_STAT_FLAG_OPTIONAL;
+			descs[j].priv = &rdmatx_cnp_op_cnts[i].type;
 		}
 	}
 }
@@ -535,7 +539,7 @@ static void mlx5_ib_dealloc_counters(struct mlx5_ib_dev *dev)
 {
 	u32 in[MLX5_ST_SZ_DW(dealloc_q_counter_in)] = {};
 	int num_cnt_ports;
-	int i;
+	int i, j;
 
 	num_cnt_ports = is_mdev_switchdev_mode(dev->mdev) ? 1 : dev->num_ports;
 
@@ -550,6 +554,18 @@ static void mlx5_ib_dealloc_counters(struct mlx5_ib_dev *dev)
 		}
 		kfree(dev->port[i].cnts.descs);
 		kfree(dev->port[i].cnts.offsets);
+
+		for (j = 0; j < MLX5_IB_OPCOUNTER_MAX; j++) {
+			if (!dev->port[i].cnts.opfcs[j].fc)
+				continue;
+
+			mlx5_ib_fs_remove_op_fc(dev,
+						&dev->port[i].cnts.opfcs[j],
+						j);
+			mlx5_fc_destroy(dev->mdev,
+					dev->port[i].cnts.opfcs[j].fc);
+			dev->port[i].cnts.opfcs[j].fc = NULL;
+		}
 	}
 }
 
@@ -729,6 +745,56 @@ void mlx5_ib_counters_clear_description(struct ib_counters *counters)
 	mutex_unlock(&mcounters->mcntrs_mutex);
 }
 
+static int mlx5_ib_modify_stat(struct ib_device *device, u32 port,
+			       unsigned int index, bool enable)
+{
+	struct mlx5_ib_dev *dev = to_mdev(device);
+	struct mlx5_ib_counters *cnts;
+	struct mlx5_ib_op_fc *opfc;
+	u32 num_hw_counters, type;
+	int ret;
+
+	cnts = &dev->port[port - 1].cnts;
+	num_hw_counters = cnts->num_q_counters + cnts->num_cong_counters +
+		cnts->num_ext_ppcnt_counters;
+	if (index < num_hw_counters ||
+	    index >= (num_hw_counters + cnts->num_op_counters))
+		return -EINVAL;
+
+	if (!(cnts->descs[index].flags & IB_STAT_FLAG_OPTIONAL))
+		return -EINVAL;
+
+	type = *(u32 *)cnts->descs[index].priv;
+	if (type >= MLX5_IB_OPCOUNTER_MAX)
+		return -EINVAL;
+
+	opfc = &cnts->opfcs[type];
+
+	if (enable) {
+		if (opfc->fc)
+			return -EEXIST;
+
+		opfc->fc = mlx5_fc_create(dev->mdev, false);
+		if (IS_ERR(opfc->fc))
+			return PTR_ERR(opfc->fc);
+
+		ret = mlx5_ib_fs_add_op_fc(dev, port, opfc, type);
+		if (ret) {
+			mlx5_fc_destroy(dev->mdev, opfc->fc);
+			opfc->fc = NULL;
+		}
+		return ret;
+	}
+
+	if (!opfc->fc)
+		return -EINVAL;
+
+	mlx5_ib_fs_remove_op_fc(dev, opfc, type);
+	mlx5_fc_destroy(dev->mdev, opfc->fc);
+	opfc->fc = NULL;
+	return 0;
+}
+
 static const struct ib_device_ops hw_stats_ops = {
 	.alloc_hw_port_stats = mlx5_ib_alloc_hw_port_stats,
 	.get_hw_stats = mlx5_ib_get_hw_stats,
@@ -737,6 +803,7 @@ static const struct ib_device_ops hw_stats_ops = {
 	.counter_dealloc = mlx5_ib_counter_dealloc,
 	.counter_alloc_stats = mlx5_ib_counter_alloc_stats,
 	.counter_update_stats = mlx5_ib_counter_update_stats,
+	.modify_hw_stat = mlx5_ib_modify_stat,
 };
 
 static const struct ib_device_ops hw_switchdev_stats_ops = {

drivers/infiniband/hw/mlx5/mlx5_ib.h

@@ -821,6 +821,7 @@ struct mlx5_ib_counters {
 	u32 num_ext_ppcnt_counters;
 	u32 num_op_counters;
 	u16 set_id;
+	struct mlx5_ib_op_fc opfcs[MLX5_IB_OPCOUNTER_MAX];
 };
 
 int mlx5_ib_fs_add_op_fc(struct mlx5_ib_dev *dev, u32 port_num,

include/rdma/ib_verbs.h

@@ -553,10 +553,12 @@ enum ib_stat_flag {
  * struct rdma_stat_desc
  * @name - The name of the counter
  * @flags - Flags of the counter; For example, IB_STAT_FLAG_OPTIONAL
+ * @priv - Driver private information; Core code should not use
  */
 struct rdma_stat_desc {
 	const char *name;
 	unsigned int flags;
+	const void *priv;
 };
 
 /**