+ * Historical RSS hash key.
+ *
+ * This used to be the default for mlx4 in Linux before v3.19 switched to
+ * generating random hash keys through netdev_rss_key_fill().
+ *
+ * It is used in this PMD for consistency with past DPDK releases but can
+ * now be overridden through user configuration.
+ *
+ * Note: this is not const to work around API quirks.
+ */
+uint8_t
+mlx4_rss_hash_key_default[MLX4_RSS_HASH_KEY_SIZE] = {
+ 0x2c, 0xc6, 0x81, 0xd1,
+ 0x5b, 0xdb, 0xf4, 0xf7,
+ 0xfc, 0xa2, 0x83, 0x19,
+ 0xdb, 0x1a, 0x3e, 0x94,
+ 0x6b, 0x9e, 0x38, 0xd9,
+ 0x2c, 0x9c, 0x03, 0xd1,
+ 0xad, 0x99, 0x44, 0xa7,
+ 0xd9, 0x56, 0x3d, 0x59,
+ 0x06, 0x3c, 0x25, 0xf3,
+ 0xfc, 0x1f, 0xdc, 0x2a,
+};
+
+/**
+ * Obtain a RSS context with specified properties.
+ *
+ * Used when creating a flow rule targeting one or several Rx queues.
+ *
+ * If a matching RSS context already exists, it is returned with its
+ * reference count incremented.
+ *
+ * @param priv
+ * Pointer to private structure.
+ * @param fields
+ * Fields for RSS processing (Verbs format).
+ * @param[in] key
+ * Hash key to use (whose size is exactly MLX4_RSS_HASH_KEY_SIZE).
+ * @param queues
+ * Number of target queues.
+ * @param[in] queue_id
+ * Target queues.
+ *
+ * @return
+ * Pointer to RSS context on success, NULL otherwise and rte_errno is set.
+ */
+struct mlx4_rss *
+mlx4_rss_get(struct mlx4_priv *priv, uint64_t fields,
+ const uint8_t key[MLX4_RSS_HASH_KEY_SIZE],
+ uint16_t queues, const uint16_t queue_id[])
+{
+ struct mlx4_rss *rss;
+ size_t queue_id_size = sizeof(queue_id[0]) * queues;
+
+ LIST_FOREACH(rss, &priv->rss, next)
+ if (fields == rss->fields &&
+ queues == rss->queues &&
+ !memcmp(key, rss->key, MLX4_RSS_HASH_KEY_SIZE) &&
+ !memcmp(queue_id, rss->queue_id, queue_id_size)) {
+ ++rss->refcnt;
+ return rss;
+ }
+ rss = rte_malloc(__func__, offsetof(struct mlx4_rss, queue_id) +
+ queue_id_size, 0);
+ if (!rss)
+ goto error;
+ *rss = (struct mlx4_rss){
+ .priv = priv,
+ .refcnt = 1,
+ .usecnt = 0,
+ .qp = NULL,
+ .ind = NULL,
+ .fields = fields,
+ .queues = queues,
+ };
+ memcpy(rss->key, key, MLX4_RSS_HASH_KEY_SIZE);
+ memcpy(rss->queue_id, queue_id, queue_id_size);
+ LIST_INSERT_HEAD(&priv->rss, rss, next);
+ return rss;
+error:
+ rte_errno = ENOMEM;
+ return NULL;
+}
+
+/**
+ * Release a RSS context instance.
+ *
+ * Used when destroying a flow rule targeting one or several Rx queues.
+ *
+ * This function decrements the reference count of the context and destroys
+ * it after reaching 0. The context must have no users at this point; all
+ * prior calls to mlx4_rss_attach() must have been followed by matching
+ * calls to mlx4_rss_detach().
+ *
+ * @param rss
+ * RSS context to release.
+ */
+void
+mlx4_rss_put(struct mlx4_rss *rss)
+{
+ MLX4_ASSERT(rss->refcnt);
+ if (--rss->refcnt)
+ return;
+ MLX4_ASSERT(!rss->usecnt);
+ MLX4_ASSERT(!rss->qp);
+ MLX4_ASSERT(!rss->ind);
+ LIST_REMOVE(rss, next);
+ rte_free(rss);
+}
+
+/**
+ * Attach a user to a RSS context instance.
+ *
+ * Used when the RSS QP and indirection table objects must be instantiated,
+ * that is, when a flow rule must be enabled.
+ *
+ * This function increments the usage count of the context.
+ *
+ * @param rss
+ * RSS context to attach to.
+ *
+ * @return
+ * 0 on success, a negative errno value otherwise and rte_errno is set.
+ */
+int
+mlx4_rss_attach(struct mlx4_rss *rss)
+{
+ MLX4_ASSERT(rss->refcnt);
+ if (rss->usecnt++) {
+ MLX4_ASSERT(rss->qp);
+ MLX4_ASSERT(rss->ind);
+ return 0;
+ }
+
+ struct ibv_wq *ind_tbl[rss->queues];
+ struct mlx4_priv *priv = rss->priv;
+ struct rte_eth_dev *dev = ETH_DEV(priv);
+ const char *msg;
+ unsigned int i = 0;
+ int ret;
+
+ if (!rte_is_power_of_2(RTE_DIM(ind_tbl))) {
+ ret = EINVAL;
+ msg = "number of RSS queues must be a power of two";
+ goto error;
+ }
+ for (i = 0; i != RTE_DIM(ind_tbl); ++i) {
+ uint16_t id = rss->queue_id[i];
+ struct rxq *rxq = NULL;
+
+ if (id < dev->data->nb_rx_queues)
+ rxq = dev->data->rx_queues[id];
+ if (!rxq) {
+ ret = EINVAL;
+ msg = "RSS target queue is not configured";
+ goto error;
+ }
+ ret = mlx4_rxq_attach(rxq);
+ if (ret) {
+ ret = -ret;
+ msg = "unable to attach RSS target queue";
+ goto error;
+ }
+ ind_tbl[i] = rxq->wq;
+ }
+ rss->ind = mlx4_glue->create_rwq_ind_table
+ (priv->ctx,
+ &(struct ibv_rwq_ind_table_init_attr){
+ .log_ind_tbl_size = rte_log2_u32(RTE_DIM(ind_tbl)),
+ .ind_tbl = ind_tbl,
+ .comp_mask = 0,
+ });
+ if (!rss->ind) {
+ ret = errno ? errno : EINVAL;
+ msg = "RSS indirection table creation failure";
+ goto error;
+ }
+ rss->qp = mlx4_glue->create_qp_ex
+ (priv->ctx,
+ &(struct ibv_qp_init_attr_ex){
+ .comp_mask = (IBV_QP_INIT_ATTR_PD |
+ IBV_QP_INIT_ATTR_RX_HASH |
+ IBV_QP_INIT_ATTR_IND_TABLE),
+ .qp_type = IBV_QPT_RAW_PACKET,
+ .pd = priv->pd,
+ .rwq_ind_tbl = rss->ind,
+ .rx_hash_conf = {
+ .rx_hash_function = IBV_RX_HASH_FUNC_TOEPLITZ,
+ .rx_hash_key_len = MLX4_RSS_HASH_KEY_SIZE,
+ .rx_hash_key = rss->key,
+ .rx_hash_fields_mask = rss->fields,
+ },
+ });
+ if (!rss->qp) {
+ ret = errno ? errno : EINVAL;
+ msg = "RSS hash QP creation failure";
+ goto error;
+ }
+ ret = mlx4_glue->modify_qp
+ (rss->qp,
+ &(struct ibv_qp_attr){
+ .qp_state = IBV_QPS_INIT,
+ .port_num = priv->port,
+ },
+ IBV_QP_STATE | IBV_QP_PORT);
+ if (ret) {
+ msg = "failed to switch RSS hash QP to INIT state";
+ goto error;
+ }
+ ret = mlx4_glue->modify_qp
+ (rss->qp,
+ &(struct ibv_qp_attr){
+ .qp_state = IBV_QPS_RTR,
+ },
+ IBV_QP_STATE);
+ if (ret) {
+ msg = "failed to switch RSS hash QP to RTR state";
+ goto error;
+ }
+ return 0;
+error:
+ if (rss->qp) {
+ claim_zero(mlx4_glue->destroy_qp(rss->qp));
+ rss->qp = NULL;
+ }
+ if (rss->ind) {
+ claim_zero(mlx4_glue->destroy_rwq_ind_table(rss->ind));
+ rss->ind = NULL;
+ }
+ while (i--)
+ mlx4_rxq_detach(dev->data->rx_queues[rss->queue_id[i]]);
+ ERROR("mlx4: %s", msg);
+ --rss->usecnt;
+ rte_errno = ret;
+ return -ret;
+}
+
+/**
+ * Detach a user from a RSS context instance.
+ *
+ * Used when disabling (not destroying) a flow rule.
+ *
+ * This function decrements the usage count of the context and destroys
+ * usage resources after reaching 0.
+ *
+ * @param rss
+ * RSS context to detach from.
+ */
+void
+mlx4_rss_detach(struct mlx4_rss *rss)
+{
+ struct mlx4_priv *priv = rss->priv;
+ struct rte_eth_dev *dev = ETH_DEV(priv);
+ unsigned int i;
+
+ MLX4_ASSERT(rss->refcnt);
+ MLX4_ASSERT(rss->qp);
+ MLX4_ASSERT(rss->ind);
+ if (--rss->usecnt)
+ return;
+ claim_zero(mlx4_glue->destroy_qp(rss->qp));
+ rss->qp = NULL;
+ claim_zero(mlx4_glue->destroy_rwq_ind_table(rss->ind));
+ rss->ind = NULL;
+ for (i = 0; i != rss->queues; ++i)
+ mlx4_rxq_detach(dev->data->rx_queues[rss->queue_id[i]]);
+}
+
+/**
+ * Initialize common RSS context resources.
+ *
+ * Because ConnectX-3 hardware limitations require a fixed order in the
+ * indirection table, WQs must be allocated sequentially to be part of a
+ * common RSS context.
+ *
+ * Since a newly created WQ cannot be moved to a different context, this
+ * function allocates them all at once, one for each configured Rx queue,
+ * as well as all related resources (CQs and mbufs).
+ *
+ * This must therefore be done before creating any Rx flow rules relying on
+ * indirection tables.
+ *
+ * @param priv
+ * Pointer to private structure.
+ *
+ * @return
+ * 0 on success, a negative errno value otherwise and rte_errno is set.
+ */
+int
+mlx4_rss_init(struct mlx4_priv *priv)
+{
+ struct rte_eth_dev *dev = ETH_DEV(priv);
+ uint8_t log2_range = rte_log2_u32(dev->data->nb_rx_queues);
+ uint32_t wq_num_prev = 0;
+ const char *msg;
+ unsigned int i;
+ int ret;
+
+ if (priv->rss_init)
+ return 0;
+ if (ETH_DEV(priv)->data->nb_rx_queues > priv->hw_rss_max_qps) {
+ ERROR("RSS does not support more than %d queues",
+ priv->hw_rss_max_qps);
+ rte_errno = EINVAL;
+ return -rte_errno;
+ }
+ /* Prepare range for RSS contexts before creating the first WQ. */
+ ret = mlx4_glue->dv_set_context_attr
+ (priv->ctx,
+ MLX4DV_SET_CTX_ATTR_LOG_WQS_RANGE_SZ,
+ &log2_range);
+ if (ret) {
+ ERROR("cannot set up range size for RSS context to %u"
+ " (for %u Rx queues), error: %s",
+ 1 << log2_range, dev->data->nb_rx_queues, strerror(ret));
+ rte_errno = ret;
+ return -ret;
+ }
+ for (i = 0; i != ETH_DEV(priv)->data->nb_rx_queues; ++i) {
+ struct rxq *rxq = ETH_DEV(priv)->data->rx_queues[i];
+ struct ibv_cq *cq;
+ struct ibv_wq *wq;
+ uint32_t wq_num;
+
+ /* Attach the configured Rx queues. */
+ if (rxq) {
+ MLX4_ASSERT(!rxq->usecnt);
+ ret = mlx4_rxq_attach(rxq);
+ if (!ret) {
+ wq_num = rxq->wq->wq_num;
+ goto wq_num_check;
+ }
+ ret = -ret;
+ msg = "unable to create Rx queue resources";
+ goto error;
+ }
+ /*
+ * WQs are temporarily allocated for unconfigured Rx queues
+ * to maintain proper index alignment in indirection table
+ * by skipping unused WQ numbers.
+ *
+ * The reason this works at all even though these WQs are
+ * immediately destroyed is that WQNs are allocated
+ * sequentially and are guaranteed to never be reused in the
+ * same context by the underlying implementation.
+ */
+ cq = mlx4_glue->create_cq(priv->ctx, 1, NULL, NULL, 0);
+ if (!cq) {
+ ret = ENOMEM;
+ msg = "placeholder CQ creation failure";
+ goto error;
+ }
+ wq = mlx4_glue->create_wq
+ (priv->ctx,
+ &(struct ibv_wq_init_attr){
+ .wq_type = IBV_WQT_RQ,
+ .max_wr = 1,
+ .max_sge = 1,
+ .pd = priv->pd,
+ .cq = cq,
+ });
+ if (wq) {
+ wq_num = wq->wq_num;
+ claim_zero(mlx4_glue->destroy_wq(wq));
+ } else {
+ wq_num = 0; /* Shut up GCC 4.8 warnings. */
+ }
+ claim_zero(mlx4_glue->destroy_cq(cq));
+ if (!wq) {
+ ret = ENOMEM;
+ msg = "placeholder WQ creation failure";
+ goto error;
+ }
+wq_num_check:
+ /*
+ * While guaranteed by the implementation, make sure WQ
+ * numbers are really sequential (as the saying goes,
+ * trust, but verify).
+ */
+ if (i && wq_num - wq_num_prev != 1) {
+ if (rxq)
+ mlx4_rxq_detach(rxq);
+ ret = ERANGE;
+ msg = "WQ numbers are not sequential";
+ goto error;
+ }
+ wq_num_prev = wq_num;
+ }
+ priv->rss_init = 1;
+ return 0;
+error:
+ ERROR("cannot initialize common RSS resources (queue %u): %s: %s",
+ i, msg, strerror(ret));
+ while (i--) {
+ struct rxq *rxq = ETH_DEV(priv)->data->rx_queues[i];
+
+ if (rxq)
+ mlx4_rxq_detach(rxq);
+ }
+ rte_errno = ret;
+ return -ret;
+}
+
+/**
+ * Release common RSS context resources.
+ *
+ * As the reverse of mlx4_rss_init(), this must be done after removing all
+ * flow rules relying on indirection tables.
+ *
+ * @param priv
+ * Pointer to private structure.
+ */
+void
+mlx4_rss_deinit(struct mlx4_priv *priv)
+{
+ unsigned int i;
+
+ if (!priv->rss_init)
+ return;
+ for (i = 0; i != ETH_DEV(priv)->data->nb_rx_queues; ++i) {
+ struct rxq *rxq = ETH_DEV(priv)->data->rx_queues[i];
+
+ if (rxq) {
+ MLX4_ASSERT(rxq->usecnt == 1);
+ mlx4_rxq_detach(rxq);
+ }
+ }
+ priv->rss_init = 0;
+}
+
+/**
+ * Attach a user to a Rx queue.
+ *
+ * Used when the resources of an Rx queue must be instantiated for it to
+ * become in a usable state.
+ *
+ * This function increments the usage count of the Rx queue.