drivers/vdpa/mlx5/mlx5_vdpa.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright 2019 Mellanox Technologies, Ltd
   3  */
   4 #include <rte_malloc.h>
   5 #include <rte_log.h>
   6 #include <rte_errno.h>
   7 #include <rte_bus_pci.h>
   8 #ifdef PEDANTIC
   9 #pragma GCC diagnostic ignored "-Wpedantic"
  10 #endif
  11 #include <rte_vdpa.h>
  12 #ifdef PEDANTIC
  13 #pragma GCC diagnostic error "-Wpedantic"
  14 #endif
  15
  16 #include <mlx5_glue.h>
  17 #include <mlx5_common.h>
  18 #include <mlx5_devx_cmds.h>
  19
  20 #include "mlx5_vdpa_utils.h"
  21
  22
  23 struct mlx5_vdpa_priv {
  24         TAILQ_ENTRY(mlx5_vdpa_priv) next;
  25         int id; /* vDPA device id. */
  26         struct ibv_context *ctx; /* Device context. */
  27         struct rte_vdpa_dev_addr dev_addr;
  28         struct mlx5_hca_vdpa_attr caps;
  29 };
  30
  31 TAILQ_HEAD(mlx5_vdpa_privs, mlx5_vdpa_priv) priv_list =
  32                                               TAILQ_HEAD_INITIALIZER(priv_list);
  33 static pthread_mutex_t priv_list_lock = PTHREAD_MUTEX_INITIALIZER;
  34 int mlx5_vdpa_logtype;
  35
  36 static struct mlx5_vdpa_priv *
  37 mlx5_vdpa_find_priv_resource_by_did(int did)
  38 {
  39         struct mlx5_vdpa_priv *priv;
  40         int found = 0;
  41
  42         pthread_mutex_lock(&priv_list_lock);
  43         TAILQ_FOREACH(priv, &priv_list, next) {
  44                 if (did == priv->id) {
  45                         found = 1;
  46                         break;
  47                 }
  48         }
  49         pthread_mutex_unlock(&priv_list_lock);
  50         if (!found) {
  51                 DRV_LOG(ERR, "Invalid device id: %d.", did);
  52                 rte_errno = EINVAL;
  53                 return NULL;
  54         }
  55         return priv;
  56 }
  57
  58 static int
  59 mlx5_vdpa_get_queue_num(int did, uint32_t *queue_num)
  60 {
  61         struct mlx5_vdpa_priv *priv = mlx5_vdpa_find_priv_resource_by_did(did);
  62
  63         if (priv == NULL) {
  64                 DRV_LOG(ERR, "Invalid device id: %d.", did);
  65                 return -1;
  66         }
  67         *queue_num = priv->caps.max_num_virtio_queues;
  68         return 0;
  69 }
  70
  71 static struct rte_vdpa_dev_ops mlx5_vdpa_ops = {
  72         .get_queue_num = mlx5_vdpa_get_queue_num,
  73         .get_features = NULL,
  74         .get_protocol_features = NULL,
  75         .dev_conf = NULL,
  76         .dev_close = NULL,
  77         .set_vring_state = NULL,
  78         .set_features = NULL,
  79         .migration_done = NULL,
  80         .get_vfio_group_fd = NULL,
  81         .get_vfio_device_fd = NULL,
  82         .get_notify_area = NULL,
  83 };
  84
  85 /**
  86  * DPDK callback to register a PCI device.
  87  *
  88  * This function spawns vdpa device out of a given PCI device.
  89  *
  90  * @param[in] pci_drv
  91  *   PCI driver structure (mlx5_vpda_driver).
  92  * @param[in] pci_dev
  93  *   PCI device information.
  94  *
  95  * @return
  96  *   0 on success, 1 to skip this driver, a negative errno value otherwise
  97  *   and rte_errno is set.
  98  */
  99 static int
 100 mlx5_vdpa_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
 101                     struct rte_pci_device *pci_dev __rte_unused)
 102 {
 103         struct ibv_device **ibv_list;
 104         struct ibv_device *ibv_match = NULL;
 105         struct mlx5_vdpa_priv *priv = NULL;
 106         struct ibv_context *ctx = NULL;
 107         struct mlx5_hca_attr attr;
 108         int ret;
 109
 110         if (mlx5_class_get(pci_dev->device.devargs) != MLX5_CLASS_VDPA) {
 111                 DRV_LOG(DEBUG, "Skip probing - should be probed by other mlx5"
 112                         " driver.");
 113                 return 1;
 114         }
 115         errno = 0;
 116         ibv_list = mlx5_glue->get_device_list(&ret);
 117         if (!ibv_list) {
 118                 rte_errno = ENOSYS;
 119                 DRV_LOG(ERR, "Failed to get device list, is ib_uverbs loaded?");
 120                 return -rte_errno;
 121         }
 122         while (ret-- > 0) {
 123                 struct rte_pci_addr pci_addr;
 124
 125                 DRV_LOG(DEBUG, "Checking device \"%s\"..", ibv_list[ret]->name);
 126                 if (mlx5_dev_to_pci_addr(ibv_list[ret]->ibdev_path, &pci_addr))
 127                         continue;
 128                 if (pci_dev->addr.domain != pci_addr.domain ||
 129                     pci_dev->addr.bus != pci_addr.bus ||
 130                     pci_dev->addr.devid != pci_addr.devid ||
 131                     pci_dev->addr.function != pci_addr.function)
 132                         continue;
 133                 DRV_LOG(INFO, "PCI information matches for device \"%s\".",
 134                         ibv_list[ret]->name);
 135                 ibv_match = ibv_list[ret];
 136                 break;
 137         }
 138         mlx5_glue->free_device_list(ibv_list);
 139         if (!ibv_match) {
 140                 DRV_LOG(ERR, "No matching IB device for PCI slot "
 141                         "%" SCNx32 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 ".",
 142                         pci_dev->addr.domain, pci_dev->addr.bus,
 143                         pci_dev->addr.devid, pci_dev->addr.function);
 144                 rte_errno = ENOENT;
 145                 return -rte_errno;
 146         }
 147         ctx = mlx5_glue->dv_open_device(ibv_match);
 148         if (!ctx) {
 149                 DRV_LOG(ERR, "Failed to open IB device \"%s\".",
 150                         ibv_match->name);
 151                 rte_errno = ENODEV;
 152                 return -rte_errno;
 153         }
 154         priv = rte_zmalloc("mlx5 vDPA device private", sizeof(*priv),
 155                            RTE_CACHE_LINE_SIZE);
 156         if (!priv) {
 157                 DRV_LOG(ERR, "Failed to allocate private memory.");
 158                 rte_errno = ENOMEM;
 159                 goto error;
 160         }
 161         ret = mlx5_devx_cmd_query_hca_attr(ctx, &attr);
 162         if (ret) {
 163                 DRV_LOG(ERR, "Unable to read HCA capabilities.");
 164                 rte_errno = ENOTSUP;
 165                 goto error;
 166         } else {
 167                 if (!attr.vdpa.valid || !attr.vdpa.max_num_virtio_queues) {
 168                         DRV_LOG(ERR, "Not enough capabilities to support vdpa,"
 169                                 " maybe old FW/OFED version?");
 170                         rte_errno = ENOTSUP;
 171                         goto error;
 172                 }
 173                 priv->caps = attr.vdpa;
 174         }
 175         priv->ctx = ctx;
 176         priv->dev_addr.pci_addr = pci_dev->addr;
 177         priv->dev_addr.type = PCI_ADDR;
 178         priv->id = rte_vdpa_register_device(&priv->dev_addr, &mlx5_vdpa_ops);
 179         if (priv->id < 0) {
 180                 DRV_LOG(ERR, "Failed to register vDPA device.");
 181                 rte_errno = rte_errno ? rte_errno : EINVAL;
 182                 goto error;
 183         }
 184         pthread_mutex_lock(&priv_list_lock);
 185         TAILQ_INSERT_TAIL(&priv_list, priv, next);
 186         pthread_mutex_unlock(&priv_list_lock);
 187         return 0;
 188
 189 error:
 190         if (priv)
 191                 rte_free(priv);
 192         if (ctx)
 193                 mlx5_glue->close_device(ctx);
 194         return -rte_errno;
 195 }
 196
 197 /**
 198  * DPDK callback to remove a PCI device.
 199  *
 200  * This function removes all vDPA devices belong to a given PCI device.
 201  *
 202  * @param[in] pci_dev
 203  *   Pointer to the PCI device.
 204  *
 205  * @return
 206  *   0 on success, the function cannot fail.
 207  */
 208 static int
 209 mlx5_vdpa_pci_remove(struct rte_pci_device *pci_dev)
 210 {
 211         struct mlx5_vdpa_priv *priv = NULL;
 212         int found = 0;
 213
 214         pthread_mutex_lock(&priv_list_lock);
 215         TAILQ_FOREACH(priv, &priv_list, next) {
 216                 if (memcmp(&priv->dev_addr.pci_addr, &pci_dev->addr,
 217                            sizeof(pci_dev->addr)) == 0) {
 218                         found = 1;
 219                         break;
 220                 }
 221         }
 222         if (found) {
 223                 TAILQ_REMOVE(&priv_list, priv, next);
 224                 mlx5_glue->close_device(priv->ctx);
 225                 rte_free(priv);
 226         }
 227         pthread_mutex_unlock(&priv_list_lock);
 228         return 0;
 229 }
 230
 231 static const struct rte_pci_id mlx5_vdpa_pci_id_map[] = {
 232         {
 233                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
 234                                PCI_DEVICE_ID_MELLANOX_CONNECTX5BF)
 235         },
 236         {
 237                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
 238                                PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF)
 239         },
 240         {
 241                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
 242                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6)
 243         },
 244         {
 245                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
 246                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6VF)
 247         },
 248         {
 249                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
 250                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6DX)
 251         },
 252         {
 253                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
 254                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6DXVF)
 255         },
 256         {
 257                 .vendor_id = 0
 258         }
 259 };
 260
 261 static struct rte_pci_driver mlx5_vdpa_driver = {
 262         .driver = {
 263                 .name = "mlx5_vdpa",
 264         },
 265         .id_table = mlx5_vdpa_pci_id_map,
 266         .probe = mlx5_vdpa_pci_probe,
 267         .remove = mlx5_vdpa_pci_remove,
 268         .drv_flags = 0,
 269 };
 270
 271 /**
 272  * Driver initialization routine.
 273  */
 274 RTE_INIT(rte_mlx5_vdpa_init)
 275 {
 276         /* Initialize common log type. */
 277         mlx5_vdpa_logtype = rte_log_register("pmd.vdpa.mlx5");
 278         if (mlx5_vdpa_logtype >= 0)
 279                 rte_log_set_level(mlx5_vdpa_logtype, RTE_LOG_NOTICE);
 280         if (mlx5_glue)
 281                 rte_pci_register(&mlx5_vdpa_driver);
 282 }
 283
 284 RTE_PMD_EXPORT_NAME(net_mlx5_vdpa, __COUNTER__);
 285 RTE_PMD_REGISTER_PCI_TABLE(net_mlx5_vdpa, mlx5_vdpa_pci_id_map);
 286 RTE_PMD_REGISTER_KMOD_DEP(net_mlx5_vdpa, "* ib_uverbs & mlx5_core & mlx5_ib");