[dpdk-dev] [PATCH v2 18/19] net/mlx5: split Rx flows to provide metadata copy
Viacheslav Ovsiienko
viacheslavo at mellanox.com
Wed Nov 6 18:37:52 CET 2019
Values set by MARK and SET_META actions should be carried over
to the VF representor in case of flow miss on Tx path. However,
as not all metadata registers are preserved across the different
domains (NIC Rx/Tx and E-Switch FDB), as a workaround, those
values should be carried by reg_c's which are preserved across
domains and copied to STE flow_tag (MARK) and reg_b (META) fields
in the last stage of flow steering, in order to scatter those
values to flow_tag and flow_table_metadata of CQE.
While reg_c[meta] can be copied to reg_b simply by modify-header
action (it is supported by hardware), it is not possible to copy
reg_c[mark] to the STE flow_tag as flow_tag is not a metadata
register and this is not supported by hardware. Instead, it should
be manually set by a flow per MARK ID. For this purpose, there
should be a dedicated flow table - RX_CP_TBL and all the Rx flow
should pass by the table to properly copy values.
As the last action of Rx flow steering must be a terminal action
such as QUEUE, RSS or DROP, if a user flow has Q/RSS action, the
flow must be split in order to pass by the RX_CP_TBL. And the
remained Q/RSS action will be performed by another dedicated
action table - RX_ACT_TBL.
For example, for an ingress flow:
pattern,
actions_having_QRSS
it must be split into two flows. The first one is,
pattern,
actions_except_QRSS / copy (reg_c[2] := flow_id) / jump to RX_CP_TBL
and the second one in RX_ACT_TBL.
(if reg_c[2] == flow_id),
action_QRSS
where flow_id is uniquely allocated and managed identifier.
This patch implements the Rx flow splitting and build the RX_ACT_TBL.
Also, per each egress flow on NIC Tx, a copy action (reg_c[]= reg_a)
should be added in order to transfer metadata from WQE.
Signed-off-by: Yongseok Koh <yskoh at mellanox.com>
Signed-off-by: Viacheslav Ovsiienko <viacheslavo at mellanox.com>
Acked-by: Matan Azrad <matan at mellanox.com>
---
drivers/net/mlx5/mlx5.c | 8 +
drivers/net/mlx5/mlx5.h | 1 +
drivers/net/mlx5/mlx5_flow.c | 428 ++++++++++++++++++++++++++++++++++++++++++-
drivers/net/mlx5/mlx5_flow.h | 1 +
4 files changed, 436 insertions(+), 2 deletions(-)
diff --git a/drivers/net/mlx5/mlx5.c b/drivers/net/mlx5/mlx5.c
index fb7b94b..6359bc9 100644
--- a/drivers/net/mlx5/mlx5.c
+++ b/drivers/net/mlx5/mlx5.c
@@ -2411,6 +2411,12 @@ struct mlx5_flow_id_pool *
err = mlx5_alloc_shared_dr(priv);
if (err)
goto error;
+ priv->qrss_id_pool = mlx5_flow_id_pool_alloc();
+ if (!priv->qrss_id_pool) {
+ DRV_LOG(ERR, "can't create flow id pool");
+ err = ENOMEM;
+ goto error;
+ }
}
/* Supported Verbs flow priority number detection. */
err = mlx5_flow_discover_priorities(eth_dev);
@@ -2463,6 +2469,8 @@ struct mlx5_flow_id_pool *
close(priv->nl_socket_rdma);
if (priv->vmwa_context)
mlx5_vlan_vmwa_exit(priv->vmwa_context);
+ if (priv->qrss_id_pool)
+ mlx5_flow_id_pool_release(priv->qrss_id_pool);
if (own_domain_id)
claim_zero(rte_eth_switch_domain_free(priv->domain_id));
rte_free(priv);
diff --git a/drivers/net/mlx5/mlx5.h b/drivers/net/mlx5/mlx5.h
index 92d445a..9c1a88a 100644
--- a/drivers/net/mlx5/mlx5.h
+++ b/drivers/net/mlx5/mlx5.h
@@ -733,6 +733,7 @@ struct mlx5_priv {
uint32_t nl_sn; /* Netlink message sequence number. */
LIST_HEAD(dbrpage, mlx5_devx_dbr_page) dbrpgs; /* Door-bell pages. */
struct mlx5_vlan_vmwa_context *vmwa_context; /* VLAN WA context. */
+ struct mlx5_flow_id_pool *qrss_id_pool;
#ifndef RTE_ARCH_64
rte_spinlock_t uar_lock_cq; /* CQs share a common distinct UAR */
rte_spinlock_t uar_lock[MLX5_UAR_PAGE_NUM_MAX];
diff --git a/drivers/net/mlx5/mlx5_flow.c b/drivers/net/mlx5/mlx5_flow.c
index d97a0b2..2f6ace0 100644
--- a/drivers/net/mlx5/mlx5_flow.c
+++ b/drivers/net/mlx5/mlx5_flow.c
@@ -2222,6 +2222,49 @@ uint32_t mlx5_flow_adjust_priority(struct rte_eth_dev *dev, int32_t priority,
return 0;
}
+/* Allocate unique ID for the split Q/RSS subflows. */
+static uint32_t
+flow_qrss_get_id(struct rte_eth_dev *dev)
+{
+ struct mlx5_priv *priv = dev->data->dev_private;
+ uint32_t qrss_id, ret;
+
+ ret = mlx5_flow_id_get(priv->qrss_id_pool, &qrss_id);
+ if (ret)
+ return 0;
+ assert(qrss_id);
+ return qrss_id;
+}
+
+/* Free unique ID for the split Q/RSS subflows. */
+static void
+flow_qrss_free_id(struct rte_eth_dev *dev, uint32_t qrss_id)
+{
+ struct mlx5_priv *priv = dev->data->dev_private;
+
+ if (qrss_id)
+ mlx5_flow_id_release(priv->qrss_id_pool, qrss_id);
+}
+
+/**
+ * Release resource related QUEUE/RSS action split.
+ *
+ * @param dev
+ * Pointer to Ethernet device.
+ * @param flow
+ * Flow to release id's from.
+ */
+static void
+flow_mreg_split_qrss_release(struct rte_eth_dev *dev,
+ struct rte_flow *flow)
+{
+ struct mlx5_flow *dev_flow;
+
+ LIST_FOREACH(dev_flow, &flow->dev_flows, next)
+ if (dev_flow->qrss_id)
+ flow_qrss_free_id(dev, dev_flow->qrss_id);
+}
+
static int
flow_null_validate(struct rte_eth_dev *dev __rte_unused,
const struct rte_flow_attr *attr __rte_unused,
@@ -2511,6 +2554,7 @@ uint32_t mlx5_flow_adjust_priority(struct rte_eth_dev *dev, int32_t priority,
const struct mlx5_flow_driver_ops *fops;
enum mlx5_flow_drv_type type = flow->drv_type;
+ flow_mreg_split_qrss_release(dev, flow);
assert(type > MLX5_FLOW_TYPE_MIN && type < MLX5_FLOW_TYPE_MAX);
fops = flow_get_drv_ops(type);
fops->destroy(dev, flow);
@@ -2581,6 +2625,41 @@ uint32_t mlx5_flow_adjust_priority(struct rte_eth_dev *dev, int32_t priority,
}
/**
+ * Get QUEUE/RSS action from the action list.
+ *
+ * @param[in] actions
+ * Pointer to the list of actions.
+ * @param[out] qrss
+ * Pointer to the return pointer.
+ * @param[out] qrss_type
+ * Pointer to the action type to return. RTE_FLOW_ACTION_TYPE_END is returned
+ * if no QUEUE/RSS is found.
+ *
+ * @return
+ * Total number of actions.
+ */
+static int
+flow_parse_qrss_action(const struct rte_flow_action actions[],
+ const struct rte_flow_action **qrss)
+{
+ int actions_n = 0;
+
+ for (; actions->type != RTE_FLOW_ACTION_TYPE_END; actions++) {
+ switch (actions->type) {
+ case RTE_FLOW_ACTION_TYPE_QUEUE:
+ case RTE_FLOW_ACTION_TYPE_RSS:
+ *qrss = actions;
+ break;
+ default:
+ break;
+ }
+ actions_n++;
+ }
+ /* Count RTE_FLOW_ACTION_TYPE_END. */
+ return actions_n + 1;
+}
+
+/**
* Check if the flow should be splited due to hairpin.
* The reason for the split is that in current HW we can't
* support encap on Rx, so if a flow have encap we move it
@@ -2832,6 +2911,351 @@ uint32_t mlx5_flow_adjust_priority(struct rte_eth_dev *dev, int32_t priority,
}
/**
+ * Split action list having QUEUE/RSS for metadata register copy.
+ *
+ * Once Q/RSS action is detected in user's action list, the flow action
+ * should be split in order to copy metadata registers, which will happen in
+ * RX_CP_TBL like,
+ * - CQE->flow_tag := reg_c[1] (MARK)
+ * - CQE->flow_table_metadata (reg_b) := reg_c[0] (META)
+ * The Q/RSS action will be performed on RX_ACT_TBL after passing by RX_CP_TBL.
+ * This is because the last action of each flow must be a terminal action
+ * (QUEUE, RSS or DROP).
+ *
+ * Flow ID must be allocated to identify actions in the RX_ACT_TBL and it is
+ * stored and kept in the mlx5_flow structure per each sub_flow.
+ *
+ * The Q/RSS action is replaced with,
+ * - SET_TAG, setting the allocated flow ID to reg_c[2].
+ * And the following JUMP action is added at the end,
+ * - JUMP, to RX_CP_TBL.
+ *
+ * A flow to perform remained Q/RSS action will be created in RX_ACT_TBL by
+ * flow_create_split_metadata() routine. The flow will look like,
+ * - If flow ID matches (reg_c[2]), perform Q/RSS.
+ *
+ * @param dev
+ * Pointer to Ethernet device.
+ * @param[out] split_actions
+ * Pointer to store split actions to jump to CP_TBL.
+ * @param[in] actions
+ * Pointer to the list of original flow actions.
+ * @param[in] qrss
+ * Pointer to the Q/RSS action.
+ * @param[in] actions_n
+ * Number of original actions.
+ * @param[out] error
+ * Perform verbose error reporting if not NULL.
+ *
+ * @return
+ * non-zero unique flow_id on success, otherwise 0 and
+ * error/rte_error are set.
+ */
+static uint32_t
+flow_mreg_split_qrss_prep(struct rte_eth_dev *dev,
+ struct rte_flow_action *split_actions,
+ const struct rte_flow_action *actions,
+ const struct rte_flow_action *qrss,
+ int actions_n, struct rte_flow_error *error)
+{
+ struct mlx5_rte_flow_action_set_tag *set_tag;
+ struct rte_flow_action_jump *jump;
+ const int qrss_idx = qrss - actions;
+ uint32_t flow_id;
+ int ret = 0;
+
+ /*
+ * Given actions will be split
+ * - Replace QUEUE/RSS action with SET_TAG to set flow ID.
+ * - Add jump to mreg CP_TBL.
+ * As a result, there will be one more action.
+ */
+ ++actions_n;
+ /*
+ * Allocate the new subflow ID. This one is unique within
+ * device and not shared with representors. Otherwise,
+ * we would have to resolve multi-thread access synch
+ * issue. Each flow on the shared device is appended
+ * with source vport identifier, so the resulting
+ * flows will be unique in the shared (by master and
+ * representors) domain even if they have coinciding
+ * IDs.
+ */
+ flow_id = flow_qrss_get_id(dev);
+ if (!flow_id)
+ return rte_flow_error_set(error, ENOMEM,
+ RTE_FLOW_ERROR_TYPE_ACTION,
+ NULL, "can't allocate id "
+ "for split Q/RSS subflow");
+ /* Internal SET_TAG action to set flow ID. */
+ set_tag = (void *)(split_actions + actions_n);
+ *set_tag = (struct mlx5_rte_flow_action_set_tag){
+ .data = flow_id,
+ };
+ ret = mlx5_flow_get_reg_id(dev, MLX5_COPY_MARK, 0, error);
+ if (ret < 0)
+ return ret;
+ set_tag->id = ret;
+ /* JUMP action to jump to mreg copy table (CP_TBL). */
+ jump = (void *)(set_tag + 1);
+ *jump = (struct rte_flow_action_jump){
+ .group = MLX5_FLOW_MREG_CP_TABLE_GROUP,
+ };
+ /* Construct new actions array. */
+ memcpy(split_actions, actions, sizeof(*split_actions) * actions_n);
+ /* Replace QUEUE/RSS action. */
+ split_actions[qrss_idx] = (struct rte_flow_action){
+ .type = MLX5_RTE_FLOW_ACTION_TYPE_TAG,
+ .conf = set_tag,
+ };
+ split_actions[actions_n - 2] = (struct rte_flow_action){
+ .type = RTE_FLOW_ACTION_TYPE_JUMP,
+ .conf = jump,
+ };
+ split_actions[actions_n - 1] = (struct rte_flow_action){
+ .type = RTE_FLOW_ACTION_TYPE_END,
+ };
+ return flow_id;
+}
+
+/**
+ * Extend the given action list for Tx metadata copy.
+ *
+ * Copy the given action list to the ext_actions and add flow metadata register
+ * copy action in order to copy reg_a set by WQE to reg_c[0].
+ *
+ * @param[out] ext_actions
+ * Pointer to the extended action list.
+ * @param[in] actions
+ * Pointer to the list of actions.
+ * @param[in] actions_n
+ * Number of actions in the list.
+ * @param[out] error
+ * Perform verbose error reporting if not NULL.
+ *
+ * @return
+ * 0 on success, negative value otherwise
+ */
+static int
+flow_mreg_tx_copy_prep(struct rte_eth_dev *dev,
+ struct rte_flow_action *ext_actions,
+ const struct rte_flow_action *actions,
+ int actions_n, struct rte_flow_error *error)
+{
+ struct mlx5_flow_action_copy_mreg *cp_mreg =
+ (struct mlx5_flow_action_copy_mreg *)
+ (ext_actions + actions_n + 1);
+ int ret;
+
+ ret = mlx5_flow_get_reg_id(dev, MLX5_METADATA_RX, 0, error);
+ if (ret < 0)
+ return ret;
+ cp_mreg->dst = ret;
+ ret = mlx5_flow_get_reg_id(dev, MLX5_METADATA_TX, 0, error);
+ if (ret < 0)
+ return ret;
+ cp_mreg->src = ret;
+ memcpy(ext_actions, actions,
+ sizeof(*ext_actions) * actions_n);
+ ext_actions[actions_n - 1] = (struct rte_flow_action){
+ .type = MLX5_RTE_FLOW_ACTION_TYPE_COPY_MREG,
+ .conf = cp_mreg,
+ };
+ ext_actions[actions_n] = (struct rte_flow_action){
+ .type = RTE_FLOW_ACTION_TYPE_END,
+ };
+ return 0;
+}
+
+/**
+ * The splitting for metadata feature.
+ *
+ * - Q/RSS action on NIC Rx should be split in order to pass by
+ * the mreg copy table (RX_CP_TBL) and then it jumps to the
+ * action table (RX_ACT_TBL) which has the split Q/RSS action.
+ *
+ * - All the actions on NIC Tx should have a mreg copy action to
+ * copy reg_a from WQE to reg_c[0].
+ *
+ * @param dev
+ * Pointer to Ethernet device.
+ * @param[in] flow
+ * Parent flow structure pointer.
+ * @param[in] attr
+ * Flow rule attributes.
+ * @param[in] items
+ * Pattern specification (list terminated by the END pattern item).
+ * @param[in] actions
+ * Associated actions (list terminated by the END action).
+ * @param[in] external
+ * This flow rule is created by request external to PMD.
+ * @param[out] error
+ * Perform verbose error reporting if not NULL.
+ * @return
+ * 0 on success, negative value otherwise
+ */
+static int
+flow_create_split_metadata(struct rte_eth_dev *dev,
+ struct rte_flow *flow,
+ const struct rte_flow_attr *attr,
+ const struct rte_flow_item items[],
+ const struct rte_flow_action actions[],
+ bool external, struct rte_flow_error *error)
+{
+ struct mlx5_priv *priv = dev->data->dev_private;
+ struct mlx5_dev_config *config = &priv->config;
+ const struct rte_flow_action *qrss = NULL;
+ struct rte_flow_action *ext_actions = NULL;
+ struct mlx5_flow *dev_flow = NULL;
+ uint32_t qrss_id = 0;
+ size_t act_size;
+ int actions_n;
+ int ret;
+
+ /* Check whether extensive metadata feature is engaged. */
+ if (!config->dv_flow_en ||
+ config->dv_xmeta_en == MLX5_XMETA_MODE_LEGACY ||
+ !mlx5_flow_ext_mreg_supported(dev))
+ return flow_create_split_inner(dev, flow, NULL, attr, items,
+ actions, external, error);
+ actions_n = flow_parse_qrss_action(actions, &qrss);
+ if (qrss) {
+ /* Exclude hairpin flows from splitting. */
+ if (qrss->type == RTE_FLOW_ACTION_TYPE_QUEUE) {
+ const struct rte_flow_action_queue *queue;
+
+ queue = qrss->conf;
+ if (mlx5_rxq_get_type(dev, queue->index) ==
+ MLX5_RXQ_TYPE_HAIRPIN)
+ qrss = NULL;
+ } else if (qrss->type == RTE_FLOW_ACTION_TYPE_RSS) {
+ const struct rte_flow_action_rss *rss;
+
+ rss = qrss->conf;
+ if (mlx5_rxq_get_type(dev, rss->queue[0]) ==
+ MLX5_RXQ_TYPE_HAIRPIN)
+ qrss = NULL;
+ }
+ }
+ if (qrss) {
+ /*
+ * Q/RSS action on NIC Rx should be split in order to pass by
+ * the mreg copy table (RX_CP_TBL) and then it jumps to the
+ * action table (RX_ACT_TBL) which has the split Q/RSS action.
+ */
+ act_size = sizeof(struct rte_flow_action) * (actions_n + 1) +
+ sizeof(struct rte_flow_action_set_tag) +
+ sizeof(struct rte_flow_action_jump);
+ ext_actions = rte_zmalloc(__func__, act_size, 0);
+ if (!ext_actions)
+ return rte_flow_error_set(error, ENOMEM,
+ RTE_FLOW_ERROR_TYPE_ACTION,
+ NULL, "no memory to split "
+ "metadata flow");
+ /*
+ * Create the new actions list with removed Q/RSS action
+ * and appended set tag and jump to register copy table
+ * (RX_CP_TBL). We should preallocate unique tag ID here
+ * in advance, because it is needed for set tag action.
+ */
+ qrss_id = flow_mreg_split_qrss_prep(dev, ext_actions, actions,
+ qrss, actions_n, error);
+ if (!qrss_id) {
+ ret = -rte_errno;
+ goto exit;
+ }
+ } else if (attr->egress && !attr->transfer) {
+ /*
+ * All the actions on NIC Tx should have a metadata register
+ * copy action to copy reg_a from WQE to reg_c[meta]
+ */
+ act_size = sizeof(struct rte_flow_action) * (actions_n + 1) +
+ sizeof(struct mlx5_flow_action_copy_mreg);
+ ext_actions = rte_zmalloc(__func__, act_size, 0);
+ if (!ext_actions)
+ return rte_flow_error_set(error, ENOMEM,
+ RTE_FLOW_ERROR_TYPE_ACTION,
+ NULL, "no memory to split "
+ "metadata flow");
+ /* Create the action list appended with copy register. */
+ ret = flow_mreg_tx_copy_prep(dev, ext_actions, actions,
+ actions_n, error);
+ if (ret < 0)
+ goto exit;
+ }
+ /* Add the unmodified original or prefix subflow. */
+ ret = flow_create_split_inner(dev, flow, &dev_flow, attr, items,
+ ext_actions ? ext_actions : actions,
+ external, error);
+ if (ret < 0)
+ goto exit;
+ assert(dev_flow);
+ if (qrss_id) {
+ const struct rte_flow_attr q_attr = {
+ .group = MLX5_FLOW_MREG_ACT_TABLE_GROUP,
+ .ingress = 1,
+ };
+ /* Internal PMD action to set register. */
+ struct mlx5_rte_flow_item_tag q_tag_spec = {
+ .data = qrss_id,
+ .id = 0,
+ };
+ struct rte_flow_item q_items[] = {
+ {
+ .type = MLX5_RTE_FLOW_ITEM_TYPE_TAG,
+ .spec = &q_tag_spec,
+ .last = NULL,
+ .mask = NULL,
+ },
+ {
+ .type = RTE_FLOW_ITEM_TYPE_END,
+ },
+ };
+ struct rte_flow_action q_actions[] = {
+ {
+ .type = qrss->type,
+ .conf = qrss->conf,
+ },
+ {
+ .type = RTE_FLOW_ACTION_TYPE_END,
+ },
+ };
+ uint64_t hash_fields = dev_flow->hash_fields;
+ /*
+ * Put unique id in prefix flow due to it is destroyed after
+ * prefix flow and id will be freed after there is no actual
+ * flows with this id and identifier reallocation becomes
+ * possible (for example, for other flows in other threads).
+ */
+ dev_flow->qrss_id = qrss_id;
+ qrss_id = 0;
+ dev_flow = NULL;
+ ret = mlx5_flow_get_reg_id(dev, MLX5_COPY_MARK, 0, error);
+ if (ret < 0)
+ goto exit;
+ q_tag_spec.id = ret;
+ /* Add suffix subflow to execute Q/RSS. */
+ ret = flow_create_split_inner(dev, flow, &dev_flow,
+ &q_attr, q_items, q_actions,
+ external, error);
+ if (ret < 0)
+ goto exit;
+ assert(dev_flow);
+ dev_flow->hash_fields = hash_fields;
+ }
+
+exit:
+ /*
+ * We do not destroy the partially created sub_flows in case of error.
+ * These ones are included into parent flow list and will be destroyed
+ * by flow_drv_destroy.
+ */
+ flow_qrss_free_id(dev, qrss_id);
+ rte_free(ext_actions);
+ return ret;
+}
+
+/**
* Split the flow to subflow set. The splitters might be linked
* in the chain, like this:
* flow_create_split_outer() calls:
@@ -2876,8 +3300,8 @@ uint32_t mlx5_flow_adjust_priority(struct rte_eth_dev *dev, int32_t priority,
{
int ret;
- ret = flow_create_split_inner(dev, flow, NULL, attr, items,
- actions, external, error);
+ ret = flow_create_split_metadata(dev, flow, attr, items,
+ actions, external, error);
assert(ret <= 0);
return ret;
}
diff --git a/drivers/net/mlx5/mlx5_flow.h b/drivers/net/mlx5/mlx5_flow.h
index ef16aef..c71938b 100644
--- a/drivers/net/mlx5/mlx5_flow.h
+++ b/drivers/net/mlx5/mlx5_flow.h
@@ -500,6 +500,7 @@ struct mlx5_flow {
#endif
struct mlx5_flow_verbs verbs;
};
+ uint32_t qrss_id; /**< Uniqie Q/RSS suffix subflow tag. */
bool external; /**< true if the flow is created external to PMD. */
};
--
1.8.3.1
More information about the dev
mailing list