net/mlx: enhance Rx scatter mode detection
[dpdk.git] / drivers / net / mlx5 / mlx5_rxq.c
index 8e02440..d5825fd 100644 (file)
@@ -36,6 +36,7 @@
 #include <errno.h>
 #include <string.h>
 #include <stdint.h>
+#include <fcntl.h>
 
 /* Verbs header. */
 /* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
@@ -57,6 +58,7 @@
 #include <rte_malloc.h>
 #include <rte_ethdev.h>
 #include <rte_common.h>
+#include <rte_interrupts.h>
 #ifdef PEDANTIC
 #pragma GCC diagnostic error "-Wpedantic"
 #endif
@@ -723,7 +725,7 @@ rxq_free_elts(struct rxq_ctrl *rxq_ctrl)
        if (rxq_ctrl->rxq.elts == NULL)
                return;
 
-       for (i = 0; (i != rxq_ctrl->rxq.elts_n); ++i) {
+       for (i = 0; (i != (1u << rxq_ctrl->rxq.elts_n)); ++i) {
                if ((*rxq_ctrl->rxq.elts)[i] != NULL)
                        rte_pktmbuf_free_seg((*rxq_ctrl->rxq.elts)[i]);
                (*rxq_ctrl->rxq.elts)[i] = NULL;
@@ -773,6 +775,8 @@ rxq_cleanup(struct rxq_ctrl *rxq_ctrl)
                claim_zero(ibv_exp_destroy_wq(rxq_ctrl->wq));
        if (rxq_ctrl->cq != NULL)
                claim_zero(ibv_destroy_cq(rxq_ctrl->cq));
+       if (rxq_ctrl->channel != NULL)
+               claim_zero(ibv_destroy_comp_channel(rxq_ctrl->channel));
        if (rxq_ctrl->rd != NULL) {
                struct ibv_exp_destroy_res_domain_attr attr = {
                        .comp_mask = 0,
@@ -807,7 +811,7 @@ rxq_cleanup(struct rxq_ctrl *rxq_ctrl)
 int
 rxq_rehash(struct rte_eth_dev *dev, struct rxq_ctrl *rxq_ctrl)
 {
-       unsigned int elts_n = rxq_ctrl->rxq.elts_n;
+       unsigned int elts_n = 1 << rxq_ctrl->rxq.elts_n;
        unsigned int i;
        struct ibv_exp_wq_attr mod;
        int err;
@@ -870,7 +874,7 @@ rxq_setup(struct rxq_ctrl *tmpl)
        struct ibv_cq *ibcq = tmpl->cq;
        struct mlx5_cq *cq = to_mxxx(cq, cq);
        struct mlx5_rwq *rwq = container_of(tmpl->wq, struct mlx5_rwq, wq);
-       struct rte_mbuf *(*elts)[tmpl->rxq.elts_n] =
+       struct rte_mbuf *(*elts)[1 << tmpl->rxq.elts_n] =
                rte_calloc_socket("RXQ", 1, sizeof(*elts), 0, tmpl->socket);
 
        if (cq->cqe_sz != RTE_CACHE_LINE_SIZE) {
@@ -881,7 +885,7 @@ rxq_setup(struct rxq_ctrl *tmpl)
        if (elts == NULL)
                return ENOMEM;
        tmpl->rxq.rq_db = rwq->rq.db;
-       tmpl->rxq.cqe_n = ibcq->cqe + 1;
+       tmpl->rxq.cqe_n = log2above(ibcq->cqe);
        tmpl->rxq.cq_ci = 0;
        tmpl->rxq.rq_ci = 0;
        tmpl->rxq.cq_db = cq->dbrec;
@@ -924,8 +928,9 @@ rxq_ctrl_setup(struct rte_eth_dev *dev, struct rxq_ctrl *rxq_ctrl,
                .priv = priv,
                .socket = socket,
                .rxq = {
-                       .elts_n = desc,
+                       .elts_n = log2above(desc),
                        .mp = mp,
+                       .rss_hash = priv->rxqs_n > 1,
                },
        };
        struct ibv_exp_wq_attr mod;
@@ -945,9 +950,10 @@ rxq_ctrl_setup(struct rte_eth_dev *dev, struct rxq_ctrl *rxq_ctrl,
        (void)conf; /* Thresholds configuration (ignored). */
        /* Enable scattered packets support for this queue if necessary. */
        assert(mb_len >= RTE_PKTMBUF_HEADROOM);
-       if ((dev->data->dev_conf.rxmode.jumbo_frame) &&
-           (dev->data->dev_conf.rxmode.max_rx_pkt_len >
-            (mb_len - RTE_PKTMBUF_HEADROOM))) {
+       if (dev->data->dev_conf.rxmode.max_rx_pkt_len <=
+           (mb_len - RTE_PKTMBUF_HEADROOM)) {
+               tmpl.rxq.sges_n = 0;
+       } else if (dev->data->dev_conf.rxmode.enable_scatter) {
                unsigned int size =
                        RTE_PKTMBUF_HEADROOM +
                        dev->data->dev_conf.rxmode.max_rx_pkt_len;
@@ -970,6 +976,13 @@ rxq_ctrl_setup(struct rte_eth_dev *dev, struct rxq_ctrl *rxq_ctrl,
                              dev->data->dev_conf.rxmode.max_rx_pkt_len);
                        return EOVERFLOW;
                }
+       } else {
+               WARN("%p: the requested maximum Rx packet size (%u) is"
+                    " larger than a single mbuf (%u) and scattered"
+                    " mode has not been requested",
+                    (void *)dev,
+                    dev->data->dev_conf.rxmode.max_rx_pkt_len,
+                    mb_len - RTE_PKTMBUF_HEADROOM);
        }
        DEBUG("%p: maximum number of segments per packet: %u",
              (void *)dev, 1 << tmpl.rxq.sges_n);
@@ -1008,6 +1021,16 @@ rxq_ctrl_setup(struct rte_eth_dev *dev, struct rxq_ctrl *rxq_ctrl,
                      (void *)dev, strerror(ret));
                goto error;
        }
+       if (dev->data->dev_conf.intr_conf.rxq) {
+               tmpl.channel = ibv_create_comp_channel(priv->ctx);
+               if (tmpl.channel == NULL) {
+                       dev->data->dev_conf.intr_conf.rxq = 0;
+                       ret = ENOMEM;
+                       ERROR("%p: Comp Channel creation failure: %s",
+                       (void *)dev, strerror(ret));
+                       goto error;
+               }
+       }
        attr.cq = (struct ibv_exp_cq_init_attr){
                .comp_mask = IBV_EXP_CQ_INIT_ATTR_RES_DOMAIN,
                .res_domain = tmpl.rd,
@@ -1017,7 +1040,7 @@ rxq_ctrl_setup(struct rte_eth_dev *dev, struct rxq_ctrl *rxq_ctrl,
                attr.cq.flags |= IBV_EXP_CQ_COMPRESSED_CQE;
                cqe_n = (desc * 2) - 1; /* Double the number of CQEs. */
        }
-       tmpl.cq = ibv_exp_create_cq(priv->ctx, cqe_n, NULL, NULL, 0,
+       tmpl.cq = ibv_exp_create_cq(priv->ctx, cqe_n, NULL, tmpl.channel, 0,
                                    &attr.cq);
        if (tmpl.cq == NULL) {
                ret = ENOMEM;
@@ -1148,7 +1171,7 @@ rxq_ctrl_setup(struct rte_eth_dev *dev, struct rxq_ctrl *rxq_ctrl,
        }
        /* Reuse buffers from original queue if possible. */
        if (rxq_ctrl->rxq.elts_n) {
-               assert(rxq_ctrl->rxq.elts_n == desc);
+               assert(1 << rxq_ctrl->rxq.elts_n == desc);
                assert(rxq_ctrl->rxq.elts != tmpl.rxq.elts);
                ret = rxq_alloc_elts(&tmpl, desc, rxq_ctrl->rxq.elts);
        } else
@@ -1241,6 +1264,19 @@ mlx5_rx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
                }
                (*priv->rxqs)[idx] = NULL;
                rxq_cleanup(rxq_ctrl);
+               /* Resize if rxq size is changed. */
+               if (rxq_ctrl->rxq.elts_n != log2above(desc)) {
+                       rxq_ctrl = rte_realloc(rxq_ctrl,
+                                              sizeof(*rxq_ctrl) +
+                                              desc * sizeof(struct rte_mbuf *),
+                                              RTE_CACHE_LINE_SIZE);
+                       if (!rxq_ctrl) {
+                               ERROR("%p: unable to reallocate queue index %u",
+                                       (void *)dev, idx);
+                               priv_unlock(priv);
+                               return -ENOMEM;
+                       }
+               }
        } else {
                rxq_ctrl = rte_calloc_socket("RXQ", 1, sizeof(*rxq_ctrl) +
                                             desc * sizeof(struct rte_mbuf *),
@@ -1261,7 +1297,7 @@ mlx5_rx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t desc,
                      (void *)dev, (void *)rxq_ctrl);
                (*priv->rxqs)[idx] = &rxq_ctrl->rxq;
                /* Update receive callback. */
-               dev->rx_pkt_burst = mlx5_rx_burst;
+               priv_select_rx_function(priv);
        }
        priv_unlock(priv);
        return -ret;
@@ -1341,3 +1377,113 @@ mlx5_rx_burst_secondary_setup(void *dpdk_rxq, struct rte_mbuf **pkts,
        rxq = (*priv->rxqs)[index];
        return priv->dev->rx_pkt_burst(rxq, pkts, pkts_n);
 }
+
+/**
+ * Fill epoll fd list for rxq interrupts.
+ *
+ * @param priv
+ *   Private structure.
+ *
+ * @return
+ *   0 on success, negative on failure.
+ */
+int
+priv_intr_efd_enable(struct priv *priv)
+{
+       unsigned int i;
+       unsigned int rxqs_n = priv->rxqs_n;
+       unsigned int n = RTE_MIN(rxqs_n, (uint32_t)RTE_MAX_RXTX_INTR_VEC_ID);
+       struct rte_intr_handle *intr_handle = priv->dev->intr_handle;
+
+       if (n == 0)
+               return 0;
+       if (n < rxqs_n) {
+               WARN("rxqs num is larger than EAL max interrupt vector "
+                    "%u > %u unable to supprt rxq interrupts",
+                    rxqs_n, (uint32_t)RTE_MAX_RXTX_INTR_VEC_ID);
+               return -EINVAL;
+       }
+       intr_handle->type = RTE_INTR_HANDLE_EXT;
+       for (i = 0; i != n; ++i) {
+               struct rxq *rxq = (*priv->rxqs)[i];
+               struct rxq_ctrl *rxq_ctrl =
+                       container_of(rxq, struct rxq_ctrl, rxq);
+               int fd = rxq_ctrl->channel->fd;
+               int flags;
+               int rc;
+
+               flags = fcntl(fd, F_GETFL);
+               rc = fcntl(fd, F_SETFL, flags | O_NONBLOCK);
+               if (rc < 0) {
+                       WARN("failed to change rxq interrupt file "
+                            "descriptor %d for queue index %d", fd, i);
+                       return -1;
+               }
+               intr_handle->efds[i] = fd;
+       }
+       intr_handle->nb_efd = n;
+       return 0;
+}
+
+/**
+ * Clean epoll fd list for rxq interrupts.
+ *
+ * @param priv
+ *   Private structure.
+ */
+void
+priv_intr_efd_disable(struct priv *priv)
+{
+       struct rte_intr_handle *intr_handle = priv->dev->intr_handle;
+
+       rte_intr_free_epoll_fd(intr_handle);
+}
+
+/**
+ * Create and init interrupt vector array.
+ *
+ * @param priv
+ *   Private structure.
+ *
+ * @return
+ *   0 on success, negative on failure.
+ */
+int
+priv_create_intr_vec(struct priv *priv)
+{
+       unsigned int rxqs_n = priv->rxqs_n;
+       unsigned int i;
+       struct rte_intr_handle *intr_handle = priv->dev->intr_handle;
+
+       if (rxqs_n == 0)
+               return 0;
+       intr_handle->intr_vec = (int *)
+               rte_malloc("intr_vec", rxqs_n * sizeof(int), 0);
+       if (intr_handle->intr_vec == NULL) {
+               WARN("Failed to allocate memory for intr_vec "
+                    "rxq interrupt will not be supported");
+               return -ENOMEM;
+       }
+       for (i = 0; i != rxqs_n; ++i) {
+               /* 1:1 mapping between rxq and interrupt. */
+               intr_handle->intr_vec[i] = RTE_INTR_VEC_RXTX_OFFSET + i;
+       }
+       return 0;
+}
+
+/**
+ * Destroy init interrupt vector array.
+ *
+ * @param priv
+ *   Private structure.
+ *
+ * @return
+ *   0 on success, negative on failure.
+ */
+void
+priv_destroy_intr_vec(struct priv *priv)
+{
+       struct rte_intr_handle *intr_handle = priv->dev->intr_handle;
+
+       rte_free(intr_handle->intr_vec);
+}