vdpa/mlx5: pre-create virtq at probing time
[dpdk.git] / drivers / vdpa / mlx5 / mlx5_vdpa_event.c
index 88f6a42..b43dca9 100644 (file)
 void
 mlx5_vdpa_event_qp_global_release(struct mlx5_vdpa_priv *priv)
 {
-       if (priv->uar) {
-               mlx5_glue->devx_free_uar(priv->uar);
-               priv->uar = NULL;
-       }
+       mlx5_devx_uar_release(&priv->uar);
 #ifdef HAVE_IBV_DEVX_EVENT
        if (priv->eventc) {
                mlx5_os_devx_destroy_event_channel(priv->eventc);
@@ -43,12 +40,10 @@ mlx5_vdpa_event_qp_global_release(struct mlx5_vdpa_priv *priv)
 }
 
 /* Prepare all the global resources for all the event objects.*/
-static int
+int
 mlx5_vdpa_event_qp_global_prepare(struct mlx5_vdpa_priv *priv)
 {
-       if (priv->eventc)
-               return 0;
-       priv->eventc = mlx5_os_devx_create_event_channel(priv->ctx,
+       priv->eventc = mlx5_os_devx_create_event_channel(priv->cdev->ctx,
                           MLX5DV_DEVX_CREATE_EVENT_CHANNEL_FLAGS_OMIT_EV_DATA);
        if (!priv->eventc) {
                rte_errno = errno;
@@ -56,14 +51,7 @@ mlx5_vdpa_event_qp_global_prepare(struct mlx5_vdpa_priv *priv)
                        rte_errno);
                goto error;
        }
-       /*
-        * This PMD always claims the write memory barrier on UAR
-        * registers writings, it is safe to allocate UAR with any
-        * memory mapping type.
-        */
-       priv->uar = mlx5_devx_alloc_uar(priv->ctx, -1);
-       if (!priv->uar) {
-               rte_errno = errno;
+       if (mlx5_devx_uar_prepare(priv->cdev, &priv->uar) != 0) {
                DRV_LOG(ERR, "Failed to allocate UAR.");
                goto error;
        }
@@ -88,18 +76,9 @@ mlx5_vdpa_cq_arm(struct mlx5_vdpa_priv *priv, struct mlx5_vdpa_cq *cq)
        uint32_t doorbell_hi = arm_sn | MLX5_CQ_DBR_CMD_ALL | cq_ci;
        uint64_t doorbell = ((uint64_t)doorbell_hi << 32) | cq->cq_obj.cq->id;
        uint64_t db_be = rte_cpu_to_be_64(doorbell);
-       uint32_t *addr = RTE_PTR_ADD(priv->uar->base_addr, MLX5_CQ_DOORBELL);
-
-       rte_io_wmb();
-       cq->cq_obj.db_rec[MLX5_CQ_ARM_DB] = rte_cpu_to_be_32(doorbell_hi);
-       rte_wmb();
-#ifdef RTE_ARCH_64
-       *(uint64_t *)addr = db_be;
-#else
-       *(uint32_t *)addr = db_be;
-       rte_io_wmb();
-       *((uint32_t *)addr + 1) = db_be >> 32;
-#endif
+
+       mlx5_doorbell_ring(&priv->uar.cq_db, db_be, doorbell_hi,
+                          &cq->cq_obj.db_rec[MLX5_CQ_ARM_DB], 0);
        cq->arm_sn++;
        cq->armed = 1;
 }
@@ -110,13 +89,13 @@ mlx5_vdpa_cq_create(struct mlx5_vdpa_priv *priv, uint16_t log_desc_n,
 {
        struct mlx5_devx_cq_attr attr = {
                .use_first_only = 1,
-               .uar_page_id = priv->uar->page_id,
+               .uar_page_id = mlx5_os_get_devx_uar_page_id(priv->uar.obj),
        };
        uint16_t event_nums[1] = {0};
        int ret;
 
-       ret = mlx5_devx_cq_create(priv->ctx, &cq->cq_obj, log_desc_n, &attr,
-                                 SOCKET_ID_ANY);
+       ret = mlx5_devx_cq_create(priv->cdev->ctx, &cq->cq_obj, log_desc_n,
+                                 &attr, SOCKET_ID_ANY);
        if (ret)
                goto error;
        cq->cq_ci = 0;
@@ -158,7 +137,7 @@ mlx5_vdpa_cq_poll(struct mlx5_vdpa_cq *cq)
                };
                uint32_t word;
        } last_word;
-       uint16_t next_wqe_counter = cq->cq_ci;
+       uint16_t next_wqe_counter = eqp->qp_pi;
        uint16_t cur_wqe_counter;
        uint16_t comp;
 
@@ -177,9 +156,10 @@ mlx5_vdpa_cq_poll(struct mlx5_vdpa_cq *cq)
                rte_io_wmb();
                /* Ring CQ doorbell record. */
                cq->cq_obj.db_rec[0] = rte_cpu_to_be_32(cq->cq_ci);
+               eqp->qp_pi += comp;
                rte_io_wmb();
                /* Ring SW QP doorbell record. */
-               eqp->db_rec[0] = rte_cpu_to_be_32(cq->cq_ci + cq_size);
+               eqp->sw_qp.db_rec[0] = rte_cpu_to_be_32(eqp->qp_pi + cq_size);
        }
        return comp;
 }
@@ -253,6 +233,25 @@ mlx5_vdpa_queues_complete(struct mlx5_vdpa_priv *priv)
        return max;
 }
 
+void
+mlx5_vdpa_drain_cq(struct mlx5_vdpa_priv *priv)
+{
+       unsigned int i;
+
+       for (i = 0; i < priv->caps.max_num_virtio_queues * 2; i++) {
+               struct mlx5_vdpa_cq *cq = &priv->virtqs[i].eqp.cq;
+
+               mlx5_vdpa_queue_complete(cq);
+               if (cq->cq_obj.cq) {
+                       cq->cq_obj.cqes[0].wqe_counter =
+                               rte_cpu_to_be_16(UINT16_MAX);
+                       priv->virtqs[i].eqp.qp_pi = 0;
+                       if (!cq->armed)
+                               mlx5_vdpa_cq_arm(priv, cq);
+               }
+       }
+}
+
 /* Wait on all CQs channel for completion event. */
 static struct mlx5_vdpa_cq *
 mlx5_vdpa_event_wait(struct mlx5_vdpa_priv *priv __rte_unused)
@@ -397,7 +396,8 @@ mlx5_vdpa_err_event_setup(struct mlx5_vdpa_priv *priv)
        int flags;
 
        /* Setup device event channel. */
-       priv->err_chnl = mlx5_glue->devx_create_event_channel(priv->ctx, 0);
+       priv->err_chnl = mlx5_glue->devx_create_event_channel(priv->cdev->ctx,
+                                                             0);
        if (!priv->err_chnl) {
                rte_errno = errno;
                DRV_LOG(ERR, "Failed to create device event channel %d.",
@@ -407,17 +407,30 @@ mlx5_vdpa_err_event_setup(struct mlx5_vdpa_priv *priv)
        flags = fcntl(priv->err_chnl->fd, F_GETFL);
        ret = fcntl(priv->err_chnl->fd, F_SETFL, flags | O_NONBLOCK);
        if (ret) {
+               rte_errno = errno;
                DRV_LOG(ERR, "Failed to change device event channel FD.");
                goto error;
        }
-       priv->err_intr_handle.fd = priv->err_chnl->fd;
-       priv->err_intr_handle.type = RTE_INTR_HANDLE_EXT;
-       if (rte_intr_callback_register(&priv->err_intr_handle,
-                                      mlx5_vdpa_err_interrupt_handler,
-                                      priv)) {
-               priv->err_intr_handle.fd = 0;
+       priv->err_intr_handle =
+               rte_intr_instance_alloc(RTE_INTR_INSTANCE_F_SHARED);
+       if (priv->err_intr_handle == NULL) {
+               DRV_LOG(ERR, "Fail to allocate intr_handle");
+               goto error;
+       }
+       if (rte_intr_fd_set(priv->err_intr_handle, priv->err_chnl->fd))
+               goto error;
+
+       if (rte_intr_type_set(priv->err_intr_handle, RTE_INTR_HANDLE_EXT))
+               goto error;
+
+       ret = rte_intr_callback_register(priv->err_intr_handle,
+                                        mlx5_vdpa_err_interrupt_handler,
+                                        priv);
+       if (ret != 0) {
+               rte_intr_fd_set(priv->err_intr_handle, 0);
                DRV_LOG(ERR, "Failed to register error interrupt for device %d.",
                        priv->vid);
+               rte_errno = -ret;
                goto error;
        } else {
                DRV_LOG(DEBUG, "Registered error interrupt for device%d.",
@@ -435,20 +448,20 @@ mlx5_vdpa_err_event_unset(struct mlx5_vdpa_priv *priv)
        int retries = MLX5_VDPA_INTR_RETRIES;
        int ret = -EAGAIN;
 
-       if (!priv->err_intr_handle.fd)
+       if (!rte_intr_fd_get(priv->err_intr_handle))
                return;
        while (retries-- && ret == -EAGAIN) {
-               ret = rte_intr_callback_unregister(&priv->err_intr_handle,
+               ret = rte_intr_callback_unregister(priv->err_intr_handle,
                                            mlx5_vdpa_err_interrupt_handler,
                                            priv);
                if (ret == -EAGAIN) {
                        DRV_LOG(DEBUG, "Try again to unregister fd %d "
                                "of error interrupt, retries = %d.",
-                               priv->err_intr_handle.fd, retries);
+                               rte_intr_fd_get(priv->err_intr_handle),
+                               retries);
                        rte_pause();
                }
        }
-       memset(&priv->err_intr_handle, 0, sizeof(priv->err_intr_handle));
        if (priv->err_chnl) {
 #ifdef HAVE_IBV_DEVX_EVENT
                union {
@@ -466,6 +479,7 @@ mlx5_vdpa_err_event_unset(struct mlx5_vdpa_priv *priv)
                mlx5_glue->devx_destroy_event_channel(priv->err_chnl);
                priv->err_chnl = NULL;
        }
+       rte_intr_instance_free(priv->err_intr_handle);
 }
 
 int
@@ -531,12 +545,7 @@ mlx5_vdpa_cqe_event_unset(struct mlx5_vdpa_priv *priv)
 void
 mlx5_vdpa_event_qp_destroy(struct mlx5_vdpa_event_qp *eqp)
 {
-       if (eqp->sw_qp)
-               claim_zero(mlx5_devx_cmd_destroy(eqp->sw_qp));
-       if (eqp->umem_obj)
-               claim_zero(mlx5_glue->devx_umem_dereg(eqp->umem_obj));
-       if (eqp->umem_buf)
-               rte_free(eqp->umem_buf);
+       mlx5_devx_qp_destroy(&eqp->sw_qp);
        if (eqp->fw_qp)
                claim_zero(mlx5_devx_cmd_destroy(eqp->fw_qp));
        mlx5_vdpa_cq_destroy(&eqp->cq);
@@ -547,36 +556,36 @@ static int
 mlx5_vdpa_qps2rts(struct mlx5_vdpa_event_qp *eqp)
 {
        if (mlx5_devx_cmd_modify_qp_state(eqp->fw_qp, MLX5_CMD_OP_RST2INIT_QP,
-                                         eqp->sw_qp->id)) {
+                                         eqp->sw_qp.qp->id)) {
                DRV_LOG(ERR, "Failed to modify FW QP to INIT state(%u).",
                        rte_errno);
                return -1;
        }
-       if (mlx5_devx_cmd_modify_qp_state(eqp->sw_qp, MLX5_CMD_OP_RST2INIT_QP,
-                                         eqp->fw_qp->id)) {
+       if (mlx5_devx_cmd_modify_qp_state(eqp->sw_qp.qp,
+                       MLX5_CMD_OP_RST2INIT_QP, eqp->fw_qp->id)) {
                DRV_LOG(ERR, "Failed to modify SW QP to INIT state(%u).",
                        rte_errno);
                return -1;
        }
        if (mlx5_devx_cmd_modify_qp_state(eqp->fw_qp, MLX5_CMD_OP_INIT2RTR_QP,
-                                         eqp->sw_qp->id)) {
+                                         eqp->sw_qp.qp->id)) {
                DRV_LOG(ERR, "Failed to modify FW QP to RTR state(%u).",
                        rte_errno);
                return -1;
        }
-       if (mlx5_devx_cmd_modify_qp_state(eqp->sw_qp, MLX5_CMD_OP_INIT2RTR_QP,
-                                         eqp->fw_qp->id)) {
+       if (mlx5_devx_cmd_modify_qp_state(eqp->sw_qp.qp,
+                       MLX5_CMD_OP_INIT2RTR_QP, eqp->fw_qp->id)) {
                DRV_LOG(ERR, "Failed to modify SW QP to RTR state(%u).",
                        rte_errno);
                return -1;
        }
        if (mlx5_devx_cmd_modify_qp_state(eqp->fw_qp, MLX5_CMD_OP_RTR2RTS_QP,
-                                         eqp->sw_qp->id)) {
+                                         eqp->sw_qp.qp->id)) {
                DRV_LOG(ERR, "Failed to modify FW QP to RTS state(%u).",
                        rte_errno);
                return -1;
        }
-       if (mlx5_devx_cmd_modify_qp_state(eqp->sw_qp, MLX5_CMD_OP_RTR2RTS_QP,
+       if (mlx5_devx_cmd_modify_qp_state(eqp->sw_qp.qp, MLX5_CMD_OP_RTR2RTS_QP,
                                          eqp->fw_qp->id)) {
                DRV_LOG(ERR, "Failed to modify SW QP to RTS state(%u).",
                        rte_errno);
@@ -585,62 +594,75 @@ mlx5_vdpa_qps2rts(struct mlx5_vdpa_event_qp *eqp)
        return 0;
 }
 
+static int
+mlx5_vdpa_qps2rst2rts(struct mlx5_vdpa_event_qp *eqp)
+{
+       if (mlx5_devx_cmd_modify_qp_state(eqp->fw_qp, MLX5_CMD_OP_QP_2RST,
+                                         eqp->sw_qp.qp->id)) {
+               DRV_LOG(ERR, "Failed to modify FW QP to RST state(%u).",
+                       rte_errno);
+               return -1;
+       }
+       if (mlx5_devx_cmd_modify_qp_state(eqp->sw_qp.qp,
+                       MLX5_CMD_OP_QP_2RST, eqp->fw_qp->id)) {
+               DRV_LOG(ERR, "Failed to modify SW QP to RST state(%u).",
+                       rte_errno);
+               return -1;
+       }
+       return mlx5_vdpa_qps2rts(eqp);
+}
+
 int
-mlx5_vdpa_event_qp_create(struct mlx5_vdpa_priv *priv, uint16_t desc_n,
+mlx5_vdpa_event_qp_prepare(struct mlx5_vdpa_priv *priv, uint16_t desc_n,
                          int callfd, struct mlx5_vdpa_event_qp *eqp)
 {
        struct mlx5_devx_qp_attr attr = {0};
        uint16_t log_desc_n = rte_log2_u32(desc_n);
-       uint32_t umem_size = (1 << log_desc_n) * MLX5_WSEG_SIZE +
-                                                      sizeof(*eqp->db_rec) * 2;
-
-       if (mlx5_vdpa_event_qp_global_prepare(priv))
-               return -1;
+       uint32_t ret;
+
+       if (eqp->cq.cq_obj.cq != NULL && log_desc_n == eqp->cq.log_desc_n) {
+               /* Reuse existing resources. */
+               eqp->cq.callfd = callfd;
+               /* FW will set event qp to error state in q destroy. */
+               if (!mlx5_vdpa_qps2rst2rts(eqp)) {
+                       rte_write32(rte_cpu_to_be_32(RTE_BIT32(log_desc_n)),
+                                       &eqp->sw_qp.db_rec[0]);
+                       return 0;
+               }
+       }
+       if (eqp->fw_qp)
+               mlx5_vdpa_event_qp_destroy(eqp);
        if (mlx5_vdpa_cq_create(priv, log_desc_n, callfd, &eqp->cq))
                return -1;
-       attr.pd = priv->pdn;
-       attr.ts_format = mlx5_ts_format_conv(priv->qp_ts_format);
-       eqp->fw_qp = mlx5_devx_cmd_create_qp(priv->ctx, &attr);
+       attr.pd = priv->cdev->pdn;
+       attr.ts_format =
+               mlx5_ts_format_conv(priv->cdev->config.hca_attr.qp_ts_format);
+       eqp->fw_qp = mlx5_devx_cmd_create_qp(priv->cdev->ctx, &attr);
        if (!eqp->fw_qp) {
                DRV_LOG(ERR, "Failed to create FW QP(%u).", rte_errno);
                goto error;
        }
-       eqp->umem_buf = rte_zmalloc(__func__, umem_size, 4096);
-       if (!eqp->umem_buf) {
-               DRV_LOG(ERR, "Failed to allocate memory for SW QP.");
-               rte_errno = ENOMEM;
-               goto error;
-       }
-       eqp->umem_obj = mlx5_glue->devx_umem_reg(priv->ctx,
-                                              (void *)(uintptr_t)eqp->umem_buf,
-                                              umem_size,
-                                              IBV_ACCESS_LOCAL_WRITE);
-       if (!eqp->umem_obj) {
-               DRV_LOG(ERR, "Failed to register umem for SW QP.");
-               goto error;
-       }
-       attr.uar_index = priv->uar->page_id;
+       attr.uar_index = mlx5_os_get_devx_uar_page_id(priv->uar.obj);
        attr.cqn = eqp->cq.cq_obj.cq->id;
-       attr.log_page_size = rte_log2_u32(sysconf(_SC_PAGESIZE));
-       attr.rq_size = 1 << log_desc_n;
+       attr.num_of_receive_wqes = RTE_BIT32(log_desc_n);
        attr.log_rq_stride = rte_log2_u32(MLX5_WSEG_SIZE);
-       attr.sq_size = 0; /* No need SQ. */
-       attr.dbr_umem_valid = 1;
-       attr.wq_umem_id = eqp->umem_obj->umem_id;
-       attr.wq_umem_offset = 0;
-       attr.dbr_umem_id = eqp->umem_obj->umem_id;
-       attr.dbr_address = (1 << log_desc_n) * MLX5_WSEG_SIZE;
-       attr.ts_format = mlx5_ts_format_conv(priv->qp_ts_format);
-       eqp->sw_qp = mlx5_devx_cmd_create_qp(priv->ctx, &attr);
-       if (!eqp->sw_qp) {
+       attr.num_of_send_wqbbs = 0; /* No need SQ. */
+       attr.ts_format =
+               mlx5_ts_format_conv(priv->cdev->config.hca_attr.qp_ts_format);
+       ret = mlx5_devx_qp_create(priv->cdev->ctx, &(eqp->sw_qp),
+                                       attr.num_of_receive_wqes *
+                                       MLX5_WSEG_SIZE, &attr, SOCKET_ID_ANY);
+       if (ret) {
                DRV_LOG(ERR, "Failed to create SW QP(%u).", rte_errno);
                goto error;
        }
-       eqp->db_rec = RTE_PTR_ADD(eqp->umem_buf, (uintptr_t)attr.dbr_address);
        if (mlx5_vdpa_qps2rts(eqp))
                goto error;
+       eqp->qp_pi = 0;
        /* First ringing. */
-       rte_write32(rte_cpu_to_be_32(1 << log_desc_n), &eqp->db_rec[0]);
+       if (eqp->sw_qp.db_rec)
+               rte_write32(rte_cpu_to_be_32(RTE_BIT32(log_desc_n)),
+                       &eqp->sw_qp.db_rec[0]);
        return 0;
 error:
        mlx5_vdpa_event_qp_destroy(eqp);