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
   9 #include <mlx5_glue.h>
  10 #include <mlx5_common.h>
  11 #include <mlx5_devx_cmds.h>
  12 #include <mlx5_prm.h>
  13
  14 #include "mlx5_vdpa_utils.h"
  15 #include "mlx5_vdpa.h"
  16
  17
  18 #define MLX5_VDPA_DEFAULT_FEATURES ((1ULL << VHOST_USER_F_PROTOCOL_FEATURES) | \
  19                             (1ULL << VIRTIO_F_ANY_LAYOUT) | \
  20                             (1ULL << VIRTIO_NET_F_MQ) | \
  21                             (1ULL << VIRTIO_NET_F_GUEST_ANNOUNCE) | \
  22                             (1ULL << VIRTIO_F_ORDER_PLATFORM))
  23
  24 #define MLX5_VDPA_PROTOCOL_FEATURES \
  25                             ((1ULL << VHOST_USER_PROTOCOL_F_SLAVE_REQ) | \
  26                              (1ULL << VHOST_USER_PROTOCOL_F_SLAVE_SEND_FD) | \
  27                              (1ULL << VHOST_USER_PROTOCOL_F_HOST_NOTIFIER) | \
  28                              (1ULL << VHOST_USER_PROTOCOL_F_LOG_SHMFD) | \
  29                              (1ULL << VHOST_USER_PROTOCOL_F_MQ))
  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 int
  72 mlx5_vdpa_get_vdpa_features(int did, uint64_t *features)
  73 {
  74         struct mlx5_vdpa_priv *priv = mlx5_vdpa_find_priv_resource_by_did(did);
  75
  76         if (priv == NULL) {
  77                 DRV_LOG(ERR, "Invalid device id: %d.", did);
  78                 return -1;
  79         }
  80         *features = MLX5_VDPA_DEFAULT_FEATURES;
  81         if (priv->caps.virtio_queue_type & (1 << MLX5_VIRTQ_TYPE_PACKED))
  82                 *features |= (1ULL << VIRTIO_F_RING_PACKED);
  83         if (priv->caps.tso_ipv4)
  84                 *features |= (1ULL << VIRTIO_NET_F_HOST_TSO4);
  85         if (priv->caps.tso_ipv6)
  86                 *features |= (1ULL << VIRTIO_NET_F_HOST_TSO6);
  87         if (priv->caps.tx_csum)
  88                 *features |= (1ULL << VIRTIO_NET_F_CSUM);
  89         if (priv->caps.rx_csum)
  90                 *features |= (1ULL << VIRTIO_NET_F_GUEST_CSUM);
  91         if (priv->caps.virtio_version_1_0)
  92                 *features |= (1ULL << VIRTIO_F_VERSION_1);
  93         return 0;
  94 }
  95
  96 static int
  97 mlx5_vdpa_get_protocol_features(int did, uint64_t *features)
  98 {
  99         struct mlx5_vdpa_priv *priv = mlx5_vdpa_find_priv_resource_by_did(did);
 100
 101         if (priv == NULL) {
 102                 DRV_LOG(ERR, "Invalid device id: %d.", did);
 103                 return -1;
 104         }
 105         *features = MLX5_VDPA_PROTOCOL_FEATURES;
 106         return 0;
 107 }
 108
 109 static struct rte_vdpa_dev_ops mlx5_vdpa_ops = {
 110         .get_queue_num = mlx5_vdpa_get_queue_num,
 111         .get_features = mlx5_vdpa_get_vdpa_features,
 112         .get_protocol_features = mlx5_vdpa_get_protocol_features,
 113         .dev_conf = NULL,
 114         .dev_close = NULL,
 115         .set_vring_state = NULL,
 116         .set_features = NULL,
 117         .migration_done = NULL,
 118         .get_vfio_group_fd = NULL,
 119         .get_vfio_device_fd = NULL,
 120         .get_notify_area = NULL,
 121 };
 122
 123 /**
 124  * DPDK callback to register a PCI device.
 125  *
 126  * This function spawns vdpa device out of a given PCI device.
 127  *
 128  * @param[in] pci_drv
 129  *   PCI driver structure (mlx5_vpda_driver).
 130  * @param[in] pci_dev
 131  *   PCI device information.
 132  *
 133  * @return
 134  *   0 on success, 1 to skip this driver, a negative errno value otherwise
 135  *   and rte_errno is set.
 136  */
 137 static int
 138 mlx5_vdpa_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
 139                     struct rte_pci_device *pci_dev __rte_unused)
 140 {
 141         struct ibv_device **ibv_list;
 142         struct ibv_device *ibv_match = NULL;
 143         struct mlx5_vdpa_priv *priv = NULL;
 144         struct ibv_context *ctx = NULL;
 145         struct mlx5_hca_attr attr;
 146         int ret;
 147
 148         if (mlx5_class_get(pci_dev->device.devargs) != MLX5_CLASS_VDPA) {
 149                 DRV_LOG(DEBUG, "Skip probing - should be probed by other mlx5"
 150                         " driver.");
 151                 return 1;
 152         }
 153         errno = 0;
 154         ibv_list = mlx5_glue->get_device_list(&ret);
 155         if (!ibv_list) {
 156                 rte_errno = ENOSYS;
 157                 DRV_LOG(ERR, "Failed to get device list, is ib_uverbs loaded?");
 158                 return -rte_errno;
 159         }
 160         while (ret-- > 0) {
 161                 struct rte_pci_addr pci_addr;
 162
 163                 DRV_LOG(DEBUG, "Checking device \"%s\"..", ibv_list[ret]->name);
 164                 if (mlx5_dev_to_pci_addr(ibv_list[ret]->ibdev_path, &pci_addr))
 165                         continue;
 166                 if (pci_dev->addr.domain != pci_addr.domain ||
 167                     pci_dev->addr.bus != pci_addr.bus ||
 168                     pci_dev->addr.devid != pci_addr.devid ||
 169                     pci_dev->addr.function != pci_addr.function)
 170                         continue;
 171                 DRV_LOG(INFO, "PCI information matches for device \"%s\".",
 172                         ibv_list[ret]->name);
 173                 ibv_match = ibv_list[ret];
 174                 break;
 175         }
 176         mlx5_glue->free_device_list(ibv_list);
 177         if (!ibv_match) {
 178                 DRV_LOG(ERR, "No matching IB device for PCI slot "
 179                         "%" SCNx32 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 ".",
 180                         pci_dev->addr.domain, pci_dev->addr.bus,
 181                         pci_dev->addr.devid, pci_dev->addr.function);
 182                 rte_errno = ENOENT;
 183                 return -rte_errno;
 184         }
 185         ctx = mlx5_glue->dv_open_device(ibv_match);
 186         if (!ctx) {
 187                 DRV_LOG(ERR, "Failed to open IB device \"%s\".",
 188                         ibv_match->name);
 189                 rte_errno = ENODEV;
 190                 return -rte_errno;
 191         }
 192         priv = rte_zmalloc("mlx5 vDPA device private", sizeof(*priv),
 193                            RTE_CACHE_LINE_SIZE);
 194         if (!priv) {
 195                 DRV_LOG(ERR, "Failed to allocate private memory.");
 196                 rte_errno = ENOMEM;
 197                 goto error;
 198         }
 199         ret = mlx5_devx_cmd_query_hca_attr(ctx, &attr);
 200         if (ret) {
 201                 DRV_LOG(ERR, "Unable to read HCA capabilities.");
 202                 rte_errno = ENOTSUP;
 203                 goto error;
 204         } else {
 205                 if (!attr.vdpa.valid || !attr.vdpa.max_num_virtio_queues) {
 206                         DRV_LOG(ERR, "Not enough capabilities to support vdpa,"
 207                                 " maybe old FW/OFED version?");
 208                         rte_errno = ENOTSUP;
 209                         goto error;
 210                 }
 211                 priv->caps = attr.vdpa;
 212                 priv->log_max_rqt_size = attr.log_max_rqt_size;
 213         }
 214         priv->ctx = ctx;
 215         priv->dev_addr.pci_addr = pci_dev->addr;
 216         priv->dev_addr.type = PCI_ADDR;
 217         priv->id = rte_vdpa_register_device(&priv->dev_addr, &mlx5_vdpa_ops);
 218         if (priv->id < 0) {
 219                 DRV_LOG(ERR, "Failed to register vDPA device.");
 220                 rte_errno = rte_errno ? rte_errno : EINVAL;
 221                 goto error;
 222         }
 223         SLIST_INIT(&priv->mr_list);
 224         SLIST_INIT(&priv->virtq_list);
 225         pthread_mutex_lock(&priv_list_lock);
 226         TAILQ_INSERT_TAIL(&priv_list, priv, next);
 227         pthread_mutex_unlock(&priv_list_lock);
 228         return 0;
 229
 230 error:
 231         if (priv)
 232                 rte_free(priv);
 233         if (ctx)
 234                 mlx5_glue->close_device(ctx);
 235         return -rte_errno;
 236 }
 237
 238 /**
 239  * DPDK callback to remove a PCI device.
 240  *
 241  * This function removes all vDPA devices belong to a given PCI device.
 242  *
 243  * @param[in] pci_dev
 244  *   Pointer to the PCI device.
 245  *
 246  * @return
 247  *   0 on success, the function cannot fail.
 248  */
 249 static int
 250 mlx5_vdpa_pci_remove(struct rte_pci_device *pci_dev)
 251 {
 252         struct mlx5_vdpa_priv *priv = NULL;
 253         int found = 0;
 254
 255         pthread_mutex_lock(&priv_list_lock);
 256         TAILQ_FOREACH(priv, &priv_list, next) {
 257                 if (memcmp(&priv->dev_addr.pci_addr, &pci_dev->addr,
 258                            sizeof(pci_dev->addr)) == 0) {
 259                         found = 1;
 260                         break;
 261                 }
 262         }
 263         if (found) {
 264                 TAILQ_REMOVE(&priv_list, priv, next);
 265                 mlx5_glue->close_device(priv->ctx);
 266                 rte_free(priv);
 267         }
 268         pthread_mutex_unlock(&priv_list_lock);
 269         return 0;
 270 }
 271
 272 static const struct rte_pci_id mlx5_vdpa_pci_id_map[] = {
 273         {
 274                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
 275                                PCI_DEVICE_ID_MELLANOX_CONNECTX5BF)
 276         },
 277         {
 278                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
 279                                PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF)
 280         },
 281         {
 282                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
 283                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6)
 284         },
 285         {
 286                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
 287                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6VF)
 288         },
 289         {
 290                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
 291                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6DX)
 292         },
 293         {
 294                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
 295                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6DXVF)
 296         },
 297         {
 298                 .vendor_id = 0
 299         }
 300 };
 301
 302 static struct rte_pci_driver mlx5_vdpa_driver = {
 303         .driver = {
 304                 .name = "mlx5_vdpa",
 305         },
 306         .id_table = mlx5_vdpa_pci_id_map,
 307         .probe = mlx5_vdpa_pci_probe,
 308         .remove = mlx5_vdpa_pci_remove,
 309         .drv_flags = 0,
 310 };
 311
 312 /**
 313  * Driver initialization routine.
 314  */
 315 RTE_INIT(rte_mlx5_vdpa_init)
 316 {
 317         /* Initialize common log type. */
 318         mlx5_vdpa_logtype = rte_log_register("pmd.vdpa.mlx5");
 319         if (mlx5_vdpa_logtype >= 0)
 320                 rte_log_set_level(mlx5_vdpa_logtype, RTE_LOG_NOTICE);
 321         if (mlx5_glue)
 322                 rte_pci_register(&mlx5_vdpa_driver);
 323 }
 324
 325 RTE_PMD_EXPORT_NAME(net_mlx5_vdpa, __COUNTER__);
 326 RTE_PMD_REGISTER_PCI_TABLE(net_mlx5_vdpa, mlx5_vdpa_pci_id_map);
 327 RTE_PMD_REGISTER_KMOD_DEP(net_mlx5_vdpa, "* ib_uverbs & mlx5_core & mlx5_ib");