net/mlx5: fix Rx interrupts management
[dpdk.git] / drivers / net / mlx5 / mlx5_rxq.c
index d5825fd..22be999 100644 (file)
@@ -59,6 +59,7 @@
 #include <rte_ethdev.h>
 #include <rte_common.h>
 #include <rte_interrupts.h>
+#include <rte_debug.h>
 #ifdef PEDANTIC
 #pragma GCC diagnostic error "-Wpedantic"
 #endif
@@ -362,6 +363,8 @@ priv_create_hash_rxqs(struct priv *priv)
        assert(priv->hash_rxqs_n == 0);
        assert(priv->pd != NULL);
        assert(priv->ctx != NULL);
+       if (priv->isolated)
+               return 0;
        if (priv->rxqs_n == 0)
                return EINVAL;
        assert(priv->rxqs != NULL);
@@ -743,51 +746,16 @@ rxq_free_elts(struct rxq_ctrl *rxq_ctrl)
 void
 rxq_cleanup(struct rxq_ctrl *rxq_ctrl)
 {
-       struct ibv_exp_release_intf_params params;
-
        DEBUG("cleaning up %p", (void *)rxq_ctrl);
        rxq_free_elts(rxq_ctrl);
        if (rxq_ctrl->fdir_queue != NULL)
                priv_fdir_queue_destroy(rxq_ctrl->priv, rxq_ctrl->fdir_queue);
-       if (rxq_ctrl->if_wq != NULL) {
-               assert(rxq_ctrl->priv != NULL);
-               assert(rxq_ctrl->priv->ctx != NULL);
-               assert(rxq_ctrl->wq != NULL);
-               params = (struct ibv_exp_release_intf_params){
-                       .comp_mask = 0,
-               };
-               claim_zero(ibv_exp_release_intf(rxq_ctrl->priv->ctx,
-                                               rxq_ctrl->if_wq,
-                                               &params));
-       }
-       if (rxq_ctrl->if_cq != NULL) {
-               assert(rxq_ctrl->priv != NULL);
-               assert(rxq_ctrl->priv->ctx != NULL);
-               assert(rxq_ctrl->cq != NULL);
-               params = (struct ibv_exp_release_intf_params){
-                       .comp_mask = 0,
-               };
-               claim_zero(ibv_exp_release_intf(rxq_ctrl->priv->ctx,
-                                               rxq_ctrl->if_cq,
-                                               &params));
-       }
        if (rxq_ctrl->wq != NULL)
                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,
-               };
-
-               assert(rxq_ctrl->priv != NULL);
-               assert(rxq_ctrl->priv->ctx != NULL);
-               claim_zero(ibv_exp_destroy_res_domain(rxq_ctrl->priv->ctx,
-                                                     rxq_ctrl->rd,
-                                                     &attr));
-       }
        if (rxq_ctrl->mr != NULL)
                claim_zero(ibv_dereg_mr(rxq_ctrl->mr));
        memset(rxq_ctrl, 0, sizeof(*rxq_ctrl));
@@ -872,12 +840,16 @@ static inline int
 rxq_setup(struct rxq_ctrl *tmpl)
 {
        struct ibv_cq *ibcq = tmpl->cq;
-       struct mlx5_cq *cq = to_mxxx(cq, cq);
+       struct ibv_mlx5_cq_info cq_info;
        struct mlx5_rwq *rwq = container_of(tmpl->wq, struct mlx5_rwq, wq);
        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) {
+       if (ibv_mlx5_exp_get_cq_info(ibcq, &cq_info)) {
+               ERROR("Unable to query CQ info. check your OFED.");
+               return ENOTSUP;
+       }
+       if (cq_info.cqe_size != RTE_CACHE_LINE_SIZE) {
                ERROR("Wrong MLX5_CQE_SIZE environment variable value: "
                      "it should be set to %u", RTE_CACHE_LINE_SIZE);
                return EINVAL;
@@ -885,16 +857,16 @@ rxq_setup(struct rxq_ctrl *tmpl)
        if (elts == NULL)
                return ENOMEM;
        tmpl->rxq.rq_db = rwq->rq.db;
-       tmpl->rxq.cqe_n = log2above(ibcq->cqe);
+       tmpl->rxq.cqe_n = log2above(cq_info.cqe_cnt);
        tmpl->rxq.cq_ci = 0;
        tmpl->rxq.rq_ci = 0;
-       tmpl->rxq.cq_db = cq->dbrec;
+       tmpl->rxq.cq_db = cq_info.dbrec;
        tmpl->rxq.wqes =
                (volatile struct mlx5_wqe_data_seg (*)[])
                (uintptr_t)rwq->rq.buff;
        tmpl->rxq.cqes =
                (volatile struct mlx5_cqe (*)[])
-               (uintptr_t)cq->active_buf->buf;
+               (uintptr_t)cq_info.buf;
        tmpl->rxq.elts = elts;
        return 0;
 }
@@ -935,13 +907,10 @@ rxq_ctrl_setup(struct rte_eth_dev *dev, struct rxq_ctrl *rxq_ctrl,
        };
        struct ibv_exp_wq_attr mod;
        union {
-               struct ibv_exp_query_intf_params params;
                struct ibv_exp_cq_init_attr cq;
-               struct ibv_exp_res_domain_init_attr rd;
                struct ibv_exp_wq_init_attr wq;
                struct ibv_exp_cq_attr cq_attr;
        } attr;
-       enum ibv_exp_query_intf_status status;
        unsigned int mb_len = rte_pktmbuf_data_room_size(mp);
        unsigned int cqe_n = desc - 1;
        struct rte_mbuf *(*elts)[desc] = NULL;
@@ -1008,32 +977,18 @@ rxq_ctrl_setup(struct rte_eth_dev *dev, struct rxq_ctrl *rxq_ctrl,
                      (void *)dev, strerror(ret));
                goto error;
        }
-       attr.rd = (struct ibv_exp_res_domain_init_attr){
-               .comp_mask = (IBV_EXP_RES_DOMAIN_THREAD_MODEL |
-                             IBV_EXP_RES_DOMAIN_MSG_MODEL),
-               .thread_model = IBV_EXP_THREAD_SINGLE,
-               .msg_model = IBV_EXP_MSG_HIGH_BW,
-       };
-       tmpl.rd = ibv_exp_create_res_domain(priv->ctx, &attr.rd);
-       if (tmpl.rd == NULL) {
-               ret = ENOMEM;
-               ERROR("%p: RD creation failure: %s",
-                     (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));
+                       ERROR("%p: Rx interrupt completion 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,
+               .comp_mask = 0,
        };
        if (priv->cqe_comp) {
                attr.cq.comp_mask |= IBV_EXP_CQ_INIT_ATTR_FLAGS;
@@ -1065,10 +1020,8 @@ rxq_ctrl_setup(struct rte_eth_dev *dev, struct rxq_ctrl *rxq_ctrl,
                .pd = priv->pd,
                .cq = tmpl.cq,
                .comp_mask =
-                       IBV_EXP_CREATE_WQ_RES_DOMAIN |
                        IBV_EXP_CREATE_WQ_VLAN_OFFLOADS |
                        0,
-               .res_domain = tmpl.rd,
                .vlan_offloads = (tmpl.rxq.vlan_strip ?
                                  IBV_EXP_RECEIVE_WQ_CVLAN_STRIP :
                                  0),
@@ -1129,29 +1082,6 @@ rxq_ctrl_setup(struct rte_eth_dev *dev, struct rxq_ctrl *rxq_ctrl,
        /* Save port ID. */
        tmpl.rxq.port_id = dev->data->port_id;
        DEBUG("%p: RTE port ID: %u", (void *)rxq_ctrl, tmpl.rxq.port_id);
-       attr.params = (struct ibv_exp_query_intf_params){
-               .intf_scope = IBV_EXP_INTF_GLOBAL,
-               .intf_version = 1,
-               .intf = IBV_EXP_INTF_CQ,
-               .obj = tmpl.cq,
-       };
-       tmpl.if_cq = ibv_exp_query_intf(priv->ctx, &attr.params, &status);
-       if (tmpl.if_cq == NULL) {
-               ERROR("%p: CQ interface family query failed with status %d",
-                     (void *)dev, status);
-               goto error;
-       }
-       attr.params = (struct ibv_exp_query_intf_params){
-               .intf_scope = IBV_EXP_INTF_GLOBAL,
-               .intf = IBV_EXP_INTF_WQ,
-               .obj = tmpl.wq,
-       };
-       tmpl.if_wq = ibv_exp_query_intf(priv->ctx, &attr.params, &status);
-       if (tmpl.if_wq == NULL) {
-               ERROR("%p: WQ interface family query failed with status %d",
-                     (void *)dev, status);
-               goto error;
-       }
        /* Change queue state to ready. */
        mod = (struct ibv_exp_wq_attr){
                .attr_mask = IBV_EXP_WQ_ATTR_STATE,
@@ -1325,6 +1255,9 @@ mlx5_rx_queue_release(void *dpdk_rxq)
        rxq_ctrl = container_of(rxq, struct rxq_ctrl, rxq);
        priv = rxq_ctrl->priv;
        priv_lock(priv);
+       if (priv_flow_rxq_in_use(priv, rxq))
+               rte_panic("Rx queue %p is still used by a flow and cannot be"
+                         " removed\n", (void *)rxq_ctrl);
        for (i = 0; (i != priv->rxqs_n); ++i)
                if ((*priv->rxqs)[i] == rxq) {
                        DEBUG("%p: removing RX queue %p from list",
@@ -1379,111 +1312,159 @@ mlx5_rx_burst_secondary_setup(void *dpdk_rxq, struct rte_mbuf **pkts,
 }
 
 /**
- * Fill epoll fd list for rxq interrupts.
+ * Allocate queue vector and fill epoll fd list for Rx interrupts.
  *
  * @param priv
- *   Private structure.
+ *   Pointer to private structure.
  *
  * @return
  *   0 on success, negative on failure.
  */
 int
-priv_intr_efd_enable(struct priv *priv)
+priv_rx_intr_vec_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);
+       unsigned int count = 0;
        struct rte_intr_handle *intr_handle = priv->dev->intr_handle;
 
-       if (n == 0)
+       if (!priv->dev->data->dev_conf.intr_conf.rxq)
                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;
+       priv_rx_intr_vec_disable(priv);
+       intr_handle->intr_vec = malloc(sizeof(intr_handle->intr_vec[rxqs_n]));
+       if (intr_handle->intr_vec == NULL) {
+               ERROR("failed to allocate memory for interrupt vector,"
+                     " Rx interrupts will not be supported");
+               return -ENOMEM;
        }
        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 fd;
                int flags;
                int rc;
 
+               /* Skip queues that cannot request interrupts. */
+               if (!rxq || !rxq_ctrl->channel) {
+                       /* Use invalid intr_vec[] index to disable entry. */
+                       intr_handle->intr_vec[i] =
+                               RTE_INTR_VEC_RXTX_OFFSET +
+                               RTE_MAX_RXTX_INTR_VEC_ID;
+                       continue;
+               }
+               if (count >= RTE_MAX_RXTX_INTR_VEC_ID) {
+                       ERROR("too many Rx queues for interrupt vector size"
+                             " (%d), Rx interrupts cannot be enabled",
+                             RTE_MAX_RXTX_INTR_VEC_ID);
+                       priv_rx_intr_vec_disable(priv);
+                       return -1;
+               }
+               fd = rxq_ctrl->channel->fd;
                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);
+                       ERROR("failed to make Rx interrupt file descriptor"
+                             " %d non-blocking for queue index %d", fd, i);
+                       priv_rx_intr_vec_disable(priv);
                        return -1;
                }
-               intr_handle->efds[i] = fd;
+               intr_handle->intr_vec[i] = RTE_INTR_VEC_RXTX_OFFSET + count;
+               intr_handle->efds[count] = fd;
+               count++;
        }
-       intr_handle->nb_efd = n;
+       if (!count)
+               priv_rx_intr_vec_disable(priv);
+       else
+               intr_handle->nb_efd = count;
        return 0;
 }
 
 /**
- * Clean epoll fd list for rxq interrupts.
+ * Clean up Rx interrupts handler.
  *
  * @param priv
- *   Private structure.
+ *   Pointer to private structure.
  */
 void
-priv_intr_efd_disable(struct priv *priv)
+priv_rx_intr_vec_disable(struct priv *priv)
 {
        struct rte_intr_handle *intr_handle = priv->dev->intr_handle;
 
        rte_intr_free_epoll_fd(intr_handle);
+       free(intr_handle->intr_vec);
+       intr_handle->nb_efd = 0;
+       intr_handle->intr_vec = NULL;
 }
 
+#ifdef HAVE_UPDATE_CQ_CI
+
 /**
- * Create and init interrupt vector array.
+ * DPDK callback for Rx queue interrupt enable.
  *
- * @param priv
- *   Private structure.
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ * @param rx_queue_id
+ *   Rx queue number.
  *
  * @return
  *   0 on success, negative on failure.
  */
 int
-priv_create_intr_vec(struct priv *priv)
+mlx5_rx_intr_enable(struct rte_eth_dev *dev, uint16_t rx_queue_id)
 {
-       unsigned int rxqs_n = priv->rxqs_n;
-       unsigned int i;
-       struct rte_intr_handle *intr_handle = priv->dev->intr_handle;
+       struct priv *priv = mlx5_get_priv(dev);
+       struct rxq *rxq = (*priv->rxqs)[rx_queue_id];
+       struct rxq_ctrl *rxq_ctrl = container_of(rxq, struct rxq_ctrl, rxq);
+       int ret;
 
-       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;
+       if (!rxq || !rxq_ctrl->channel) {
+               ret = EINVAL;
+       } else {
+               ibv_mlx5_exp_update_cq_ci(rxq_ctrl->cq, rxq->cq_ci);
+               ret = ibv_req_notify_cq(rxq_ctrl->cq, 0);
        }
-       return 0;
+       if (ret)
+               WARN("unable to arm interrupt on rx queue %d", rx_queue_id);
+       return -ret;
 }
 
 /**
- * Destroy init interrupt vector array.
+ * DPDK callback for Rx queue interrupt disable.
  *
- * @param priv
- *   Private structure.
+ * @param dev
+ *   Pointer to Ethernet device structure.
+ * @param rx_queue_id
+ *   Rx queue number.
  *
  * @return
  *   0 on success, negative on failure.
  */
-void
-priv_destroy_intr_vec(struct priv *priv)
+int
+mlx5_rx_intr_disable(struct rte_eth_dev *dev, uint16_t rx_queue_id)
 {
-       struct rte_intr_handle *intr_handle = priv->dev->intr_handle;
+       struct priv *priv = mlx5_get_priv(dev);
+       struct rxq *rxq = (*priv->rxqs)[rx_queue_id];
+       struct rxq_ctrl *rxq_ctrl = container_of(rxq, struct rxq_ctrl, rxq);
+       struct ibv_cq *ev_cq;
+       void *ev_ctx;
+       int ret;
 
-       rte_free(intr_handle->intr_vec);
+       if (!rxq || !rxq_ctrl->channel) {
+               ret = EINVAL;
+       } else {
+               ret = ibv_get_cq_event(rxq_ctrl->cq->channel, &ev_cq, &ev_ctx);
+               if (ret || ev_cq != rxq_ctrl->cq)
+                       ret = EINVAL;
+       }
+       if (ret)
+               WARN("unable to disable interrupt on rx queue %d",
+                    rx_queue_id);
+       else
+               ibv_ack_cq_events(rxq_ctrl->cq, 1);
+       return -ret;
 }
+
+#endif /* HAVE_UPDATE_CQ_CI */