vdpa/mlx5: add hardware queue moderation
[dpdk.git] / drivers / vdpa / mlx5 / mlx5_vdpa.c
index b4b6bf4..0f22a86 100644 (file)
 #include <rte_malloc.h>
 #include <rte_log.h>
 #include <rte_errno.h>
-#include <rte_bus_pci.h>
 #include <rte_pci.h>
 #include <rte_string_fns.h>
 
 #include <mlx5_glue.h>
 #include <mlx5_common.h>
+#include <mlx5_common_pci.h>
 #include <mlx5_devx_cmds.h>
 #include <mlx5_prm.h>
 #include <mlx5_nl.h>
@@ -39,7 +39,8 @@
                             (1ULL << VHOST_USER_PROTOCOL_F_HOST_NOTIFIER) | \
                             (1ULL << VHOST_USER_PROTOCOL_F_LOG_SHMFD) | \
                             (1ULL << VHOST_USER_PROTOCOL_F_MQ) | \
-                            (1ULL << VHOST_USER_PROTOCOL_F_NET_MTU))
+                            (1ULL << VHOST_USER_PROTOCOL_F_NET_MTU) | \
+                            (1ULL << VHOST_USER_PROTOCOL_F_STATUS))
 
 #define MLX5_VDPA_MAX_RETRIES 20
 #define MLX5_VDPA_USEC 1000
@@ -48,7 +49,6 @@
 TAILQ_HEAD(mlx5_vdpa_privs, mlx5_vdpa_priv) priv_list =
                                              TAILQ_HEAD_INITIALIZER(priv_list);
 static pthread_mutex_t priv_list_lock = PTHREAD_MUTEX_INITIALIZER;
-int mlx5_vdpa_logtype;
 
 static struct mlx5_vdpa_priv *
 mlx5_vdpa_find_priv_resource_by_vdev(struct rte_vdpa_device *vdev)
@@ -133,6 +133,7 @@ mlx5_vdpa_set_vring_state(int vid, int vring, int state)
        struct rte_vdpa_device *vdev = rte_vhost_get_vdpa_device(vid);
        struct mlx5_vdpa_priv *priv =
                mlx5_vdpa_find_priv_resource_by_vdev(vdev);
+       int ret;
 
        if (priv == NULL) {
                DRV_LOG(ERR, "Invalid vDPA device: %s.", vdev->device->name);
@@ -142,32 +143,10 @@ mlx5_vdpa_set_vring_state(int vid, int vring, int state)
                DRV_LOG(ERR, "Too big vring id: %d.", vring);
                return -E2BIG;
        }
-       return mlx5_vdpa_virtq_enable(priv, vring, state);
-}
-
-static int
-mlx5_vdpa_direct_db_prepare(struct mlx5_vdpa_priv *priv)
-{
-       int ret;
-
-       if (priv->direct_notifier) {
-               ret = rte_vhost_host_notifier_ctrl(priv->vid,
-                                                  RTE_VHOST_QUEUE_ALL, false);
-               if (ret != 0) {
-                       DRV_LOG(INFO, "Direct HW notifier FD cannot be "
-                               "destroyed for device %d: %d.", priv->vid, ret);
-                       return -1;
-               }
-               priv->direct_notifier = 0;
-       }
-       ret = rte_vhost_host_notifier_ctrl(priv->vid, RTE_VHOST_QUEUE_ALL,
-                                          true);
-       if (ret != 0)
-               DRV_LOG(INFO, "Direct HW notifier FD cannot be configured for"
-                       " device %d: %d.", priv->vid, ret);
-       else
-               priv->direct_notifier = 1;
-       return 0;
+       pthread_mutex_lock(&priv->vq_config_lock);
+       ret = mlx5_vdpa_virtq_enable(priv, vring, state);
+       pthread_mutex_unlock(&priv->vq_config_lock);
+       return ret;
 }
 
 static int
@@ -304,6 +283,7 @@ mlx5_vdpa_dev_close(int vid)
        }
        if (priv->configured)
                ret |= mlx5_vdpa_lm_log(priv);
+       mlx5_vdpa_err_event_unset(priv);
        mlx5_vdpa_cqe_event_unset(priv);
        mlx5_vdpa_steer_unset(priv);
        mlx5_vdpa_virtqs_release(priv);
@@ -339,7 +319,7 @@ mlx5_vdpa_dev_config(int vid)
                DRV_LOG(WARNING, "MTU cannot be set on device %s.",
                                vdev->device->name);
        if (mlx5_vdpa_pd_create(priv) || mlx5_vdpa_mem_register(priv) ||
-           mlx5_vdpa_direct_db_prepare(priv) ||
+           mlx5_vdpa_err_event_setup(priv) ||
            mlx5_vdpa_virtqs_prepare(priv) || mlx5_vdpa_steer_setup(priv) ||
            mlx5_vdpa_cqe_event_setup(priv)) {
                mlx5_vdpa_dev_close(vid);
@@ -632,6 +612,7 @@ mlx5_vdpa_args_check_handler(const char *key, const char *val, void *opaque)
 {
        struct mlx5_vdpa_priv *priv = opaque;
        unsigned long tmp;
+       int n_cores = sysconf(_SC_NPROCESSORS_ONLN);
 
        if (strcmp(key, "class") == 0)
                return 0;
@@ -650,6 +631,17 @@ mlx5_vdpa_args_check_handler(const char *key, const char *val, void *opaque)
                priv->event_us = (uint32_t)tmp;
        } else if (strcmp(key, "no_traffic_time") == 0) {
                priv->no_traffic_time_s = (uint32_t)tmp;
+       } else if (strcmp(key, "event_core") == 0) {
+               if (tmp >= (unsigned long)n_cores)
+                       DRV_LOG(WARNING, "Invalid event_core %s.", val);
+               else
+                       priv->event_core = tmp;
+       } else if (strcmp(key, "hw_latency_mode") == 0) {
+               priv->hw_latency_mode = (uint32_t)tmp;
+       } else if (strcmp(key, "hw_max_latency_us") == 0) {
+               priv->hw_max_latency_us = (uint32_t)tmp;
+       } else if (strcmp(key, "hw_max_pending_comp") == 0) {
+               priv->hw_max_pending_comp = (uint32_t)tmp;
        } else {
                DRV_LOG(WARNING, "Invalid key %s.", key);
        }
@@ -661,8 +653,9 @@ mlx5_vdpa_config_get(struct rte_devargs *devargs, struct mlx5_vdpa_priv *priv)
 {
        struct rte_kvargs *kvlist;
 
-       priv->event_mode = MLX5_VDPA_EVENT_MODE_DYNAMIC_TIMER;
+       priv->event_mode = MLX5_VDPA_EVENT_MODE_FIXED_TIMER;
        priv->event_us = 0;
+       priv->event_core = -1;
        priv->no_traffic_time_s = MLX5_VDPA_DEFAULT_NO_TRAFFIC_TIME_S;
        if (devargs == NULL)
                return;
@@ -671,19 +664,16 @@ mlx5_vdpa_config_get(struct rte_devargs *devargs, struct mlx5_vdpa_priv *priv)
                return;
        rte_kvargs_process(kvlist, NULL, mlx5_vdpa_args_check_handler, priv);
        rte_kvargs_free(kvlist);
-       if (!priv->event_us) {
-               if (priv->event_mode == MLX5_VDPA_EVENT_MODE_DYNAMIC_TIMER)
-                       priv->event_us = MLX5_VDPA_DEFAULT_TIMER_STEP_US;
-               else if (priv->event_mode == MLX5_VDPA_EVENT_MODE_FIXED_TIMER)
-                       priv->event_us = MLX5_VDPA_DEFAULT_TIMER_DELAY_US;
-       }
+       if (!priv->event_us &&
+           priv->event_mode == MLX5_VDPA_EVENT_MODE_DYNAMIC_TIMER)
+               priv->event_us = MLX5_VDPA_DEFAULT_TIMER_STEP_US;
        DRV_LOG(DEBUG, "event mode is %d.", priv->event_mode);
        DRV_LOG(DEBUG, "event_us is %u us.", priv->event_us);
        DRV_LOG(DEBUG, "no traffic time is %u s.", priv->no_traffic_time_s);
 }
 
 /**
- * DPDK callback to register a PCI device.
+ * DPDK callback to register a mlx5 PCI device.
  *
  * This function spawns vdpa device out of a given PCI device.
  *
@@ -706,11 +696,6 @@ mlx5_vdpa_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
        struct mlx5_hca_attr attr;
        int ret;
 
-       if (mlx5_class_get(pci_dev->device.devargs) != MLX5_CLASS_VDPA) {
-               DRV_LOG(DEBUG, "Skip probing - should be probed by other mlx5"
-                       " driver.");
-               return 1;
-       }
        ibv = mlx5_vdpa_get_ib_device_match(&pci_dev->addr);
        if (!ibv) {
                DRV_LOG(ERR, "No matching IB device for PCI slot "
@@ -757,6 +742,9 @@ mlx5_vdpa_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
        }
        priv->caps = attr.vdpa;
        priv->log_max_rqt_size = attr.log_max_rqt_size;
+       priv->num_lag_ports = attr.num_lag_ports;
+       if (attr.num_lag_ports == 0)
+               priv->num_lag_ports = 1;
        priv->ctx = ctx;
        priv->pci_dev = pci_dev;
        priv->var = mlx5_glue->dv_alloc_var(ctx, 0);
@@ -773,6 +761,7 @@ mlx5_vdpa_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
        }
        mlx5_vdpa_config_get(pci_dev->device.devargs, priv);
        SLIST_INIT(&priv->mr_list);
+       pthread_mutex_init(&priv->vq_config_lock, NULL);
        pthread_mutex_lock(&priv_list_lock);
        TAILQ_INSERT_TAIL(&priv_list, priv, next);
        pthread_mutex_unlock(&priv_list_lock);
@@ -824,6 +813,7 @@ mlx5_vdpa_pci_remove(struct rte_pci_device *pci_dev)
                        priv->var = NULL;
                }
                mlx5_glue->close_device(priv->ctx);
+               pthread_mutex_destroy(&priv->vq_config_lock);
                rte_free(priv);
        }
        return 0;
@@ -844,38 +834,48 @@ static const struct rte_pci_id mlx5_vdpa_pci_id_map[] = {
        },
        {
                RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
-                               PCI_DEVICE_ID_MELLANOX_CONNECTX6DXVF)
+                               PCI_DEVICE_ID_MELLANOX_CONNECTXVF)
        },
        {
                RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
                                PCI_DEVICE_ID_MELLANOX_CONNECTX6DXBF)
        },
+       {
+               RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
+                               PCI_DEVICE_ID_MELLANOX_CONNECTX7)
+       },
+       {
+               RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
+                               PCI_DEVICE_ID_MELLANOX_CONNECTX7BF)
+       },
        {
                .vendor_id = 0
        }
 };
 
-static struct rte_pci_driver mlx5_vdpa_driver = {
-       .driver = {
-               .name = "mlx5_vdpa",
+static struct mlx5_pci_driver mlx5_vdpa_driver = {
+       .driver_class = MLX5_CLASS_VDPA,
+       .pci_driver = {
+               .driver = {
+                       .name = "mlx5_vdpa",
+               },
+               .id_table = mlx5_vdpa_pci_id_map,
+               .probe = mlx5_vdpa_pci_probe,
+               .remove = mlx5_vdpa_pci_remove,
+               .drv_flags = 0,
        },
-       .id_table = mlx5_vdpa_pci_id_map,
-       .probe = mlx5_vdpa_pci_probe,
-       .remove = mlx5_vdpa_pci_remove,
-       .drv_flags = 0,
 };
 
+RTE_LOG_REGISTER(mlx5_vdpa_logtype, pmd.vdpa.mlx5, NOTICE)
+
 /**
  * Driver initialization routine.
  */
 RTE_INIT(rte_mlx5_vdpa_init)
 {
-       /* Initialize common log type. */
-       mlx5_vdpa_logtype = rte_log_register("pmd.vdpa.mlx5");
-       if (mlx5_vdpa_logtype >= 0)
-               rte_log_set_level(mlx5_vdpa_logtype, RTE_LOG_NOTICE);
+       mlx5_common_init();
        if (mlx5_glue)
-               rte_pci_register(&mlx5_vdpa_driver);
+               mlx5_pci_driver_register(&mlx5_vdpa_driver);
 }
 
 RTE_PMD_EXPORT_NAME(net_mlx5_vdpa, __COUNTER__);