net/bnxt: handle multi queue mode properly
authorAjit Khaparde <ajit.khaparde@broadcom.com>
Thu, 28 Sep 2017 21:43:26 +0000 (16:43 -0500)
committerFerruh Yigit <ferruh.yigit@intel.com>
Fri, 6 Oct 2017 00:49:49 +0000 (02:49 +0200)
We are currently not handling multi queue RX/RSS modes correctly.
If RSS is not requested, create one VNIC per RXQ.

Fixes: 6133f207970c ("net/bnxt: add Rx queue create/destroy")
Cc: stable@dpdk.org
Signed-off-by: Ajit Khaparde <ajit.khaparde@broadcom.com>
drivers/net/bnxt/bnxt_ethdev.c
drivers/net/bnxt/bnxt_rxq.c

index c9d1122..6e5cb88 100644 (file)
@@ -360,6 +360,7 @@ static void bnxt_dev_info_get_op(struct rte_eth_dev *eth_dev,
 {
        struct bnxt *bp = (struct bnxt *)eth_dev->data->dev_private;
        uint16_t max_vnics, i, j, vpool, vrxq;
+       unsigned int max_rx_rings;
 
        dev_info->pci_dev = RTE_ETH_DEV_TO_PCI(eth_dev);
 
@@ -370,8 +371,12 @@ static void bnxt_dev_info_get_op(struct rte_eth_dev *eth_dev,
        /* PF/VF specifics */
        if (BNXT_PF(bp))
                dev_info->max_vfs = bp->pdev->max_vfs;
-       dev_info->max_rx_queues = bp->max_rx_rings;
-       dev_info->max_tx_queues = bp->max_tx_rings;
+       max_rx_rings = RTE_MIN(bp->max_vnics, RTE_MIN(bp->max_l2_ctx,
+                                               RTE_MIN(bp->max_rsscos_ctx,
+                                               bp->max_stat_ctx)));
+       /* For the sake of symmetry, max_rx_queues = max_tx_queues */
+       dev_info->max_rx_queues = max_rx_rings;
+       dev_info->max_tx_queues = max_rx_rings;
        dev_info->reta_size = bp->max_rsscos_ctx;
        max_vnics = bp->max_vnics;
 
@@ -827,7 +832,7 @@ static int bnxt_rss_hash_update_op(struct rte_eth_dev *eth_dev,
         */
        if (dev_conf->rxmode.mq_mode & ETH_MQ_RX_RSS_FLAG) {
                if (!rss_conf->rss_hf)
-                       return -EINVAL;
+                       RTE_LOG(ERR, PMD, "Hash type NONE\n");
        } else {
                if (rss_conf->rss_hf & BNXT_ETH_RSS_SUPPORT)
                        return -EINVAL;
index 441e543..8459fcc 100644 (file)
@@ -60,11 +60,13 @@ void bnxt_free_rxq_stats(struct bnxt_rx_queue *rxq)
 int bnxt_mq_rx_configure(struct bnxt *bp)
 {
        struct rte_eth_conf *dev_conf = &bp->eth_dev->data->dev_conf;
-       unsigned int i, j, nb_q_per_grp, ring_idx;
-       int start_grp_id, end_grp_id, rc = 0;
+       unsigned int i, j, nb_q_per_grp = 1, ring_idx = 0;
+       int start_grp_id, end_grp_id = 1, rc = 0;
        struct bnxt_vnic_info *vnic;
        struct bnxt_filter_info *filter;
+       enum rte_eth_nb_pools pools = bp->rx_cp_nr_rings, max_pools = 0;
        struct bnxt_rx_queue *rxq;
+       bool rss_dflt_cr = false;
 
        bp->nr_vnics = 0;
 
@@ -98,116 +100,123 @@ int bnxt_mq_rx_configure(struct bnxt *bp)
        }
 
        /* Multi-queue mode */
-       if (dev_conf->rxmode.mq_mode & ETH_MQ_RX_VMDQ_FLAG) {
+       if (dev_conf->rxmode.mq_mode & ETH_MQ_RX_VMDQ_DCB_RSS) {
                /* VMDq ONLY, VMDq+RSS, VMDq+DCB, VMDq+DCB+RSS */
-               enum rte_eth_nb_pools pools;
+               const struct rte_eth_vmdq_rx_conf *conf =
+                   &dev_conf->rx_adv_conf.vmdq_rx_conf;
+
 
                switch (dev_conf->rxmode.mq_mode) {
                case ETH_MQ_RX_VMDQ_RSS:
                case ETH_MQ_RX_VMDQ_ONLY:
-                       {
-                               const struct rte_eth_vmdq_rx_conf *conf =
-                                   &dev_conf->rx_adv_conf.vmdq_rx_conf;
-
-                               /* ETH_8/64_POOLs */
-                               pools = conf->nb_queue_pools;
-                               break;
-                       }
+                       /* ETH_8/64_POOLs */
+                       pools = conf->nb_queue_pools;
+                       /* For each pool, allocate MACVLAN CFA rule & VNIC */
+                       max_pools = RTE_MIN(bp->max_vnics,
+                                           RTE_MIN(bp->max_l2_ctx,
+                                           RTE_MIN(bp->max_rsscos_ctx,
+                                                   ETH_64_POOLS)));
+                       if (pools > max_pools)
+                               pools = max_pools;
+                       break;
+               case ETH_MQ_RX_RSS:
+                       pools = 1;
+                       break;
                default:
                        RTE_LOG(ERR, PMD, "Unsupported mq_mod %d\n",
                                dev_conf->rxmode.mq_mode);
                        rc = -EINVAL;
                        goto err_out;
                }
-               /* For each pool, allocate MACVLAN CFA rule & VNIC */
-               if (!pools) {
-                       pools = RTE_MIN(bp->max_vnics,
-                           RTE_MIN(bp->max_l2_ctx,
-                            RTE_MIN(bp->max_rsscos_ctx, ETH_64_POOLS)));
-                       RTE_LOG(ERR, PMD,
-                               "VMDq pool not set, defaulted to %d\n", pools);
+       }
+       /*
+        * If MQ RX w/o RSS no need for per VNIC filter.
+        */
+       if ((dev_conf->rxmode.mq_mode & ETH_MQ_RX_VMDQ_DCB) ||
+           (bp->rx_cp_nr_rings &&
+            !(dev_conf->rxmode.mq_mode & ETH_MQ_RX_RSS)))
+               rss_dflt_cr = true;
+
+       nb_q_per_grp = bp->rx_cp_nr_rings / pools;
+       start_grp_id = 0;
+       end_grp_id = nb_q_per_grp;
+
+       for (i = 0; i < pools; i++) {
+               vnic = bnxt_alloc_vnic(bp);
+               if (!vnic) {
+                       RTE_LOG(ERR, PMD, "VNIC alloc failed\n");
+                       rc = -ENOMEM;
+                       goto err_out;
                }
-               nb_q_per_grp = bp->rx_cp_nr_rings / pools;
-               start_grp_id = 0;
-               end_grp_id = nb_q_per_grp;
-
-               ring_idx = 0;
-               for (i = 0; i < pools; i++) {
-                       vnic = bnxt_alloc_vnic(bp);
-                       if (!vnic) {
-                               RTE_LOG(ERR, PMD,
-                                       "VNIC alloc failed\n");
-                               rc = -ENOMEM;
-                               goto err_out;
-                       }
-                       vnic->flags |= BNXT_VNIC_INFO_BCAST;
-                       STAILQ_INSERT_TAIL(&bp->ff_pool[i], vnic, next);
-                       bp->nr_vnics++;
-
-                       for (j = 0; j < nb_q_per_grp; j++, ring_idx++) {
-                               rxq = bp->eth_dev->data->rx_queues[ring_idx];
-                               rxq->vnic = vnic;
-                       }
-                       if (i == 0)
-                               vnic->func_default = true;
-                       vnic->ff_pool_idx = i;
-                       vnic->start_grp_id = start_grp_id;
-                       vnic->end_grp_id = end_grp_id;
-
-                       filter = bnxt_alloc_filter(bp);
-                       if (!filter) {
-                               RTE_LOG(ERR, PMD,
-                                       "L2 filter alloc failed\n");
-                               rc = -ENOMEM;
-                               goto err_out;
-                       }
-                       /*
-                        * TODO: Configure & associate CFA rule for
-                        * each VNIC for each VMDq with MACVLAN, MACVLAN+TC
-                        */
-                       STAILQ_INSERT_TAIL(&vnic->filter, filter, next);
+               vnic->flags |= BNXT_VNIC_INFO_BCAST;
+               STAILQ_INSERT_TAIL(&bp->ff_pool[i], vnic, next);
+               bp->nr_vnics++;
 
-                       start_grp_id = end_grp_id;
-                       end_grp_id += nb_q_per_grp;
+               for (j = 0, ring_idx = 0; j < nb_q_per_grp; j++, ring_idx++) {
+                       rxq = bp->eth_dev->data->rx_queues[ring_idx];
+                       rxq->vnic = vnic;
                }
-               goto out;
-       }
+               if (i == 0)
+                       vnic->func_default = true;
+               vnic->ff_pool_idx = i;
+               vnic->start_grp_id = start_grp_id;
+               vnic->end_grp_id = end_grp_id;
+
+               if (rss_dflt_cr && i) {
+                       vnic->rss_dflt_cr = true;
+                       goto skip_filter_allocation;
+               }
+               filter = bnxt_alloc_filter(bp);
+               if (!filter) {
+                       RTE_LOG(ERR, PMD, "L2 filter alloc failed\n");
+                       rc = -ENOMEM;
+                       goto err_out;
+               }
+               /*
+                * TODO: Configure & associate CFA rule for
+                * each VNIC for each VMDq with MACVLAN, MACVLAN+TC
+                */
+               STAILQ_INSERT_TAIL(&vnic->filter, filter, next);
 
-       /* Non-VMDq mode - RSS, DCB, RSS+DCB */
-       /* Init default VNIC for RSS or DCB only */
-       vnic = bnxt_alloc_vnic(bp);
-       if (!vnic) {
-               RTE_LOG(ERR, PMD, "VNIC alloc failed\n");
-               rc = -ENOMEM;
-               goto err_out;
-       }
-       vnic->flags |= BNXT_VNIC_INFO_BCAST;
-       /* Partition the rx queues for the single pool */
-       for (i = 0; i < bp->rx_cp_nr_rings; i++) {
-               rxq = bp->eth_dev->data->rx_queues[i];
-               rxq->vnic = vnic;
+skip_filter_allocation:
+               start_grp_id = end_grp_id;
+               end_grp_id += nb_q_per_grp;
        }
-       STAILQ_INSERT_TAIL(&bp->ff_pool[0], vnic, next);
-       bp->nr_vnics++;
-
-       vnic->func_default = true;
-       vnic->ff_pool_idx = 0;
-       vnic->start_grp_id = 0;
-       vnic->end_grp_id = bp->rx_cp_nr_rings;
-       filter = bnxt_alloc_filter(bp);
-       if (!filter) {
-               RTE_LOG(ERR, PMD, "L2 filter alloc failed\n");
-               rc = -ENOMEM;
-               goto err_out;
-       }
-       STAILQ_INSERT_TAIL(&vnic->filter, filter, next);
-
-       if (dev_conf->rxmode.mq_mode & ETH_MQ_RX_RSS_FLAG)
-               vnic->hash_type =
-                       HWRM_VNIC_RSS_CFG_INPUT_HASH_TYPE_IPV4 |
-                       HWRM_VNIC_RSS_CFG_INPUT_HASH_TYPE_IPV6;
 
 out:
+       if (dev_conf->rxmode.mq_mode & ETH_MQ_RX_RSS_FLAG) {
+               struct rte_eth_rss_conf *rss = &dev_conf->rx_adv_conf.rss_conf;
+               uint16_t hash_type = 0;
+
+               if (rss->rss_hf & ETH_RSS_IPV4)
+                       hash_type |= HWRM_VNIC_RSS_CFG_INPUT_HASH_TYPE_IPV4;
+               if (rss->rss_hf & ETH_RSS_NONFRAG_IPV4_TCP)
+                       hash_type |= HWRM_VNIC_RSS_CFG_INPUT_HASH_TYPE_TCP_IPV4;
+               if (rss->rss_hf & ETH_RSS_NONFRAG_IPV4_UDP)
+                       hash_type |= HWRM_VNIC_RSS_CFG_INPUT_HASH_TYPE_UDP_IPV4;
+               if (rss->rss_hf & ETH_RSS_IPV6)
+                       hash_type |= HWRM_VNIC_RSS_CFG_INPUT_HASH_TYPE_IPV6;
+               if (rss->rss_hf & ETH_RSS_NONFRAG_IPV6_TCP)
+                       hash_type |= HWRM_VNIC_RSS_CFG_INPUT_HASH_TYPE_TCP_IPV6;
+               if (rss->rss_hf & ETH_RSS_NONFRAG_IPV6_UDP)
+                       hash_type |= HWRM_VNIC_RSS_CFG_INPUT_HASH_TYPE_UDP_IPV6;
+
+               for (i = 0; i < bp->nr_vnics; i++) {
+                       STAILQ_FOREACH(vnic, &bp->ff_pool[i], next) {
+                       vnic->hash_type |= hash_type;
+
+                       /*
+                        * Use the supplied key if the key length is
+                        * acceptable and the rss_key is not NULL
+                        */
+                       if (rss->rss_key &&
+                           rss->rss_key_len <= HW_HASH_KEY_SIZE)
+                               memcpy(vnic->rss_hash_key,
+                                      rss->rss_key, rss->rss_key_len);
+                       }
+               }
+       }
+
        return rc;
 
 err_out: