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 int
 110 mlx5_vdpa_set_vring_state(int vid, int vring, int state)
 111 {
 112         int did = rte_vhost_get_vdpa_device_id(vid);
 113         struct mlx5_vdpa_priv *priv = mlx5_vdpa_find_priv_resource_by_did(did);
 114         struct mlx5_vdpa_virtq *virtq = NULL;
 115
 116         if (priv == NULL) {
 117                 DRV_LOG(ERR, "Invalid device id: %d.", did);
 118                 return -EINVAL;
 119         }
 120         SLIST_FOREACH(virtq, &priv->virtq_list, next)
 121                 if (virtq->index == vring)
 122                         break;
 123         if (!virtq) {
 124                 DRV_LOG(ERR, "Invalid or unconfigured vring id: %d.", vring);
 125                 return -EINVAL;
 126         }
 127         return mlx5_vdpa_virtq_enable(virtq, state);
 128 }
 129
 130 static struct rte_vdpa_dev_ops mlx5_vdpa_ops = {
 131         .get_queue_num = mlx5_vdpa_get_queue_num,
 132         .get_features = mlx5_vdpa_get_vdpa_features,
 133         .get_protocol_features = mlx5_vdpa_get_protocol_features,
 134         .dev_conf = NULL,
 135         .dev_close = NULL,
 136         .set_vring_state = mlx5_vdpa_set_vring_state,
 137         .set_features = NULL,
 138         .migration_done = NULL,
 139         .get_vfio_group_fd = NULL,
 140         .get_vfio_device_fd = NULL,
 141         .get_notify_area = NULL,
 142 };
 143
 144 /**
 145  * DPDK callback to register a PCI device.
 146  *
 147  * This function spawns vdpa device out of a given PCI device.
 148  *
 149  * @param[in] pci_drv
 150  *   PCI driver structure (mlx5_vpda_driver).
 151  * @param[in] pci_dev
 152  *   PCI device information.
 153  *
 154  * @return
 155  *   0 on success, 1 to skip this driver, a negative errno value otherwise
 156  *   and rte_errno is set.
 157  */
 158 static int
 159 mlx5_vdpa_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
 160                     struct rte_pci_device *pci_dev __rte_unused)
 161 {
 162         struct ibv_device **ibv_list;
 163         struct ibv_device *ibv_match = NULL;
 164         struct mlx5_vdpa_priv *priv = NULL;
 165         struct ibv_context *ctx = NULL;
 166         struct mlx5_hca_attr attr;
 167         int ret;
 168
 169         if (mlx5_class_get(pci_dev->device.devargs) != MLX5_CLASS_VDPA) {
 170                 DRV_LOG(DEBUG, "Skip probing - should be probed by other mlx5"
 171                         " driver.");
 172                 return 1;
 173         }
 174         errno = 0;
 175         ibv_list = mlx5_glue->get_device_list(&ret);
 176         if (!ibv_list) {
 177                 rte_errno = ENOSYS;
 178                 DRV_LOG(ERR, "Failed to get device list, is ib_uverbs loaded?");
 179                 return -rte_errno;
 180         }
 181         while (ret-- > 0) {
 182                 struct rte_pci_addr pci_addr;
 183
 184                 DRV_LOG(DEBUG, "Checking device \"%s\"..", ibv_list[ret]->name);
 185                 if (mlx5_dev_to_pci_addr(ibv_list[ret]->ibdev_path, &pci_addr))
 186                         continue;
 187                 if (pci_dev->addr.domain != pci_addr.domain ||
 188                     pci_dev->addr.bus != pci_addr.bus ||
 189                     pci_dev->addr.devid != pci_addr.devid ||
 190                     pci_dev->addr.function != pci_addr.function)
 191                         continue;
 192                 DRV_LOG(INFO, "PCI information matches for device \"%s\".",
 193                         ibv_list[ret]->name);
 194                 ibv_match = ibv_list[ret];
 195                 break;
 196         }
 197         mlx5_glue->free_device_list(ibv_list);
 198         if (!ibv_match) {
 199                 DRV_LOG(ERR, "No matching IB device for PCI slot "
 200                         "%" SCNx32 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 ".",
 201                         pci_dev->addr.domain, pci_dev->addr.bus,
 202                         pci_dev->addr.devid, pci_dev->addr.function);
 203                 rte_errno = ENOENT;
 204                 return -rte_errno;
 205         }
 206         ctx = mlx5_glue->dv_open_device(ibv_match);
 207         if (!ctx) {
 208                 DRV_LOG(ERR, "Failed to open IB device \"%s\".",
 209                         ibv_match->name);
 210                 rte_errno = ENODEV;
 211                 return -rte_errno;
 212         }
 213         priv = rte_zmalloc("mlx5 vDPA device private", sizeof(*priv),
 214                            RTE_CACHE_LINE_SIZE);
 215         if (!priv) {
 216                 DRV_LOG(ERR, "Failed to allocate private memory.");
 217                 rte_errno = ENOMEM;
 218                 goto error;
 219         }
 220         ret = mlx5_devx_cmd_query_hca_attr(ctx, &attr);
 221         if (ret) {
 222                 DRV_LOG(ERR, "Unable to read HCA capabilities.");
 223                 rte_errno = ENOTSUP;
 224                 goto error;
 225         } else {
 226                 if (!attr.vdpa.valid || !attr.vdpa.max_num_virtio_queues) {
 227                         DRV_LOG(ERR, "Not enough capabilities to support vdpa,"
 228                                 " maybe old FW/OFED version?");
 229                         rte_errno = ENOTSUP;
 230                         goto error;
 231                 }
 232                 priv->caps = attr.vdpa;
 233                 priv->log_max_rqt_size = attr.log_max_rqt_size;
 234         }
 235         priv->ctx = ctx;
 236         priv->dev_addr.pci_addr = pci_dev->addr;
 237         priv->dev_addr.type = PCI_ADDR;
 238         priv->id = rte_vdpa_register_device(&priv->dev_addr, &mlx5_vdpa_ops);
 239         if (priv->id < 0) {
 240                 DRV_LOG(ERR, "Failed to register vDPA device.");
 241                 rte_errno = rte_errno ? rte_errno : EINVAL;
 242                 goto error;
 243         }
 244         SLIST_INIT(&priv->mr_list);
 245         SLIST_INIT(&priv->virtq_list);
 246         pthread_mutex_lock(&priv_list_lock);
 247         TAILQ_INSERT_TAIL(&priv_list, priv, next);
 248         pthread_mutex_unlock(&priv_list_lock);
 249         return 0;
 250
 251 error:
 252         if (priv)
 253                 rte_free(priv);
 254         if (ctx)
 255                 mlx5_glue->close_device(ctx);
 256         return -rte_errno;
 257 }
 258
 259 /**
 260  * DPDK callback to remove a PCI device.
 261  *
 262  * This function removes all vDPA devices belong to a given PCI device.
 263  *
 264  * @param[in] pci_dev
 265  *   Pointer to the PCI device.
 266  *
 267  * @return
 268  *   0 on success, the function cannot fail.
 269  */
 270 static int
 271 mlx5_vdpa_pci_remove(struct rte_pci_device *pci_dev)
 272 {
 273         struct mlx5_vdpa_priv *priv = NULL;
 274         int found = 0;
 275
 276         pthread_mutex_lock(&priv_list_lock);
 277         TAILQ_FOREACH(priv, &priv_list, next) {
 278                 if (memcmp(&priv->dev_addr.pci_addr, &pci_dev->addr,
 279                            sizeof(pci_dev->addr)) == 0) {
 280                         found = 1;
 281                         break;
 282                 }
 283         }
 284         if (found) {
 285                 TAILQ_REMOVE(&priv_list, priv, next);
 286                 mlx5_glue->close_device(priv->ctx);
 287                 rte_free(priv);
 288         }
 289         pthread_mutex_unlock(&priv_list_lock);
 290         return 0;
 291 }
 292
 293 static const struct rte_pci_id mlx5_vdpa_pci_id_map[] = {
 294         {
 295                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
 296                                PCI_DEVICE_ID_MELLANOX_CONNECTX5BF)
 297         },
 298         {
 299                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
 300                                PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF)
 301         },
 302         {
 303                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
 304                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6)
 305         },
 306         {
 307                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
 308                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6VF)
 309         },
 310         {
 311                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
 312                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6DX)
 313         },
 314         {
 315                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
 316                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6DXVF)
 317         },
 318         {
 319                 .vendor_id = 0
 320         }
 321 };
 322
 323 static struct rte_pci_driver mlx5_vdpa_driver = {
 324         .driver = {
 325                 .name = "mlx5_vdpa",
 326         },
 327         .id_table = mlx5_vdpa_pci_id_map,
 328         .probe = mlx5_vdpa_pci_probe,
 329         .remove = mlx5_vdpa_pci_remove,
 330         .drv_flags = 0,
 331 };
 332
 333 /**
 334  * Driver initialization routine.
 335  */
 336 RTE_INIT(rte_mlx5_vdpa_init)
 337 {
 338         /* Initialize common log type. */
 339         mlx5_vdpa_logtype = rte_log_register("pmd.vdpa.mlx5");
 340         if (mlx5_vdpa_logtype >= 0)
 341                 rte_log_set_level(mlx5_vdpa_logtype, RTE_LOG_NOTICE);
 342         if (mlx5_glue)
 343                 rte_pci_register(&mlx5_vdpa_driver);
 344 }
 345
 346 RTE_PMD_EXPORT_NAME(net_mlx5_vdpa, __COUNTER__);
 347 RTE_PMD_REGISTER_PCI_TABLE(net_mlx5_vdpa, mlx5_vdpa_pci_id_map);
 348 RTE_PMD_REGISTER_KMOD_DEP(net_mlx5_vdpa, "* ib_uverbs & mlx5_core & mlx5_ib");