drivers/vdpa/mlx5/mlx5_vdpa.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright 2019 Mellanox Technologies, Ltd
   3  */
   4 #include <linux/virtio_net.h>
   5
   6 #include <rte_malloc.h>
   7 #include <rte_log.h>
   8 #include <rte_errno.h>
   9 #include <rte_bus_pci.h>
  10 #ifdef PEDANTIC
  11 #pragma GCC diagnostic ignored "-Wpedantic"
  12 #endif
  13 #include <rte_vdpa.h>
  14 #ifdef PEDANTIC
  15 #pragma GCC diagnostic error "-Wpedantic"
  16 #endif
  17
  18 #include <mlx5_glue.h>
  19 #include <mlx5_common.h>
  20 #include <mlx5_devx_cmds.h>
  21 #include <mlx5_prm.h>
  22
  23 #include "mlx5_vdpa_utils.h"
  24
  25
  26 struct mlx5_vdpa_priv {
  27         TAILQ_ENTRY(mlx5_vdpa_priv) next;
  28         int id; /* vDPA device id. */
  29         struct ibv_context *ctx; /* Device context. */
  30         struct rte_vdpa_dev_addr dev_addr;
  31         struct mlx5_hca_vdpa_attr caps;
  32 };
  33
  34 #ifndef VIRTIO_F_ORDER_PLATFORM
  35 #define VIRTIO_F_ORDER_PLATFORM 36
  36 #endif
  37
  38 #ifndef VIRTIO_F_RING_PACKED
  39 #define VIRTIO_F_RING_PACKED 34
  40 #endif
  41
  42 #define MLX5_VDPA_DEFAULT_FEATURES ((1ULL << VHOST_USER_F_PROTOCOL_FEATURES) | \
  43                             (1ULL << VIRTIO_F_ANY_LAYOUT) | \
  44                             (1ULL << VIRTIO_NET_F_MQ) | \
  45                             (1ULL << VIRTIO_NET_F_GUEST_ANNOUNCE) | \
  46                             (1ULL << VIRTIO_F_ORDER_PLATFORM))
  47
  48 #define MLX5_VDPA_PROTOCOL_FEATURES \
  49                             ((1ULL << VHOST_USER_PROTOCOL_F_SLAVE_REQ) | \
  50                              (1ULL << VHOST_USER_PROTOCOL_F_SLAVE_SEND_FD) | \
  51                              (1ULL << VHOST_USER_PROTOCOL_F_HOST_NOTIFIER) | \
  52                              (1ULL << VHOST_USER_PROTOCOL_F_LOG_SHMFD) | \
  53                              (1ULL << VHOST_USER_PROTOCOL_F_MQ))
  54
  55 TAILQ_HEAD(mlx5_vdpa_privs, mlx5_vdpa_priv) priv_list =
  56                                               TAILQ_HEAD_INITIALIZER(priv_list);
  57 static pthread_mutex_t priv_list_lock = PTHREAD_MUTEX_INITIALIZER;
  58 int mlx5_vdpa_logtype;
  59
  60 static struct mlx5_vdpa_priv *
  61 mlx5_vdpa_find_priv_resource_by_did(int did)
  62 {
  63         struct mlx5_vdpa_priv *priv;
  64         int found = 0;
  65
  66         pthread_mutex_lock(&priv_list_lock);
  67         TAILQ_FOREACH(priv, &priv_list, next) {
  68                 if (did == priv->id) {
  69                         found = 1;
  70                         break;
  71                 }
  72         }
  73         pthread_mutex_unlock(&priv_list_lock);
  74         if (!found) {
  75                 DRV_LOG(ERR, "Invalid device id: %d.", did);
  76                 rte_errno = EINVAL;
  77                 return NULL;
  78         }
  79         return priv;
  80 }
  81
  82 static int
  83 mlx5_vdpa_get_queue_num(int did, uint32_t *queue_num)
  84 {
  85         struct mlx5_vdpa_priv *priv = mlx5_vdpa_find_priv_resource_by_did(did);
  86
  87         if (priv == NULL) {
  88                 DRV_LOG(ERR, "Invalid device id: %d.", did);
  89                 return -1;
  90         }
  91         *queue_num = priv->caps.max_num_virtio_queues;
  92         return 0;
  93 }
  94
  95 static int
  96 mlx5_vdpa_get_vdpa_features(int did, uint64_t *features)
  97 {
  98         struct mlx5_vdpa_priv *priv = mlx5_vdpa_find_priv_resource_by_did(did);
  99
 100         if (priv == NULL) {
 101                 DRV_LOG(ERR, "Invalid device id: %d.", did);
 102                 return -1;
 103         }
 104         *features = MLX5_VDPA_DEFAULT_FEATURES;
 105         if (priv->caps.virtio_queue_type & (1 << MLX5_VIRTQ_TYPE_PACKED))
 106                 *features |= (1ULL << VIRTIO_F_RING_PACKED);
 107         if (priv->caps.tso_ipv4)
 108                 *features |= (1ULL << VIRTIO_NET_F_HOST_TSO4);
 109         if (priv->caps.tso_ipv6)
 110                 *features |= (1ULL << VIRTIO_NET_F_HOST_TSO6);
 111         if (priv->caps.tx_csum)
 112                 *features |= (1ULL << VIRTIO_NET_F_CSUM);
 113         if (priv->caps.rx_csum)
 114                 *features |= (1ULL << VIRTIO_NET_F_GUEST_CSUM);
 115         if (priv->caps.virtio_version_1_0)
 116                 *features |= (1ULL << VIRTIO_F_VERSION_1);
 117         return 0;
 118 }
 119
 120 static int
 121 mlx5_vdpa_get_protocol_features(int did, uint64_t *features)
 122 {
 123         struct mlx5_vdpa_priv *priv = mlx5_vdpa_find_priv_resource_by_did(did);
 124
 125         if (priv == NULL) {
 126                 DRV_LOG(ERR, "Invalid device id: %d.", did);
 127                 return -1;
 128         }
 129         *features = MLX5_VDPA_PROTOCOL_FEATURES;
 130         return 0;
 131 }
 132
 133 static struct rte_vdpa_dev_ops mlx5_vdpa_ops = {
 134         .get_queue_num = mlx5_vdpa_get_queue_num,
 135         .get_features = mlx5_vdpa_get_vdpa_features,
 136         .get_protocol_features = mlx5_vdpa_get_protocol_features,
 137         .dev_conf = NULL,
 138         .dev_close = NULL,
 139         .set_vring_state = NULL,
 140         .set_features = NULL,
 141         .migration_done = NULL,
 142         .get_vfio_group_fd = NULL,
 143         .get_vfio_device_fd = NULL,
 144         .get_notify_area = NULL,
 145 };
 146
 147 /**
 148  * DPDK callback to register a PCI device.
 149  *
 150  * This function spawns vdpa device out of a given PCI device.
 151  *
 152  * @param[in] pci_drv
 153  *   PCI driver structure (mlx5_vpda_driver).
 154  * @param[in] pci_dev
 155  *   PCI device information.
 156  *
 157  * @return
 158  *   0 on success, 1 to skip this driver, a negative errno value otherwise
 159  *   and rte_errno is set.
 160  */
 161 static int
 162 mlx5_vdpa_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
 163                     struct rte_pci_device *pci_dev __rte_unused)
 164 {
 165         struct ibv_device **ibv_list;
 166         struct ibv_device *ibv_match = NULL;
 167         struct mlx5_vdpa_priv *priv = NULL;
 168         struct ibv_context *ctx = NULL;
 169         struct mlx5_hca_attr attr;
 170         int ret;
 171
 172         if (mlx5_class_get(pci_dev->device.devargs) != MLX5_CLASS_VDPA) {
 173                 DRV_LOG(DEBUG, "Skip probing - should be probed by other mlx5"
 174                         " driver.");
 175                 return 1;
 176         }
 177         errno = 0;
 178         ibv_list = mlx5_glue->get_device_list(&ret);
 179         if (!ibv_list) {
 180                 rte_errno = ENOSYS;
 181                 DRV_LOG(ERR, "Failed to get device list, is ib_uverbs loaded?");
 182                 return -rte_errno;
 183         }
 184         while (ret-- > 0) {
 185                 struct rte_pci_addr pci_addr;
 186
 187                 DRV_LOG(DEBUG, "Checking device \"%s\"..", ibv_list[ret]->name);
 188                 if (mlx5_dev_to_pci_addr(ibv_list[ret]->ibdev_path, &pci_addr))
 189                         continue;
 190                 if (pci_dev->addr.domain != pci_addr.domain ||
 191                     pci_dev->addr.bus != pci_addr.bus ||
 192                     pci_dev->addr.devid != pci_addr.devid ||
 193                     pci_dev->addr.function != pci_addr.function)
 194                         continue;
 195                 DRV_LOG(INFO, "PCI information matches for device \"%s\".",
 196                         ibv_list[ret]->name);
 197                 ibv_match = ibv_list[ret];
 198                 break;
 199         }
 200         mlx5_glue->free_device_list(ibv_list);
 201         if (!ibv_match) {
 202                 DRV_LOG(ERR, "No matching IB device for PCI slot "
 203                         "%" SCNx32 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 ".",
 204                         pci_dev->addr.domain, pci_dev->addr.bus,
 205                         pci_dev->addr.devid, pci_dev->addr.function);
 206                 rte_errno = ENOENT;
 207                 return -rte_errno;
 208         }
 209         ctx = mlx5_glue->dv_open_device(ibv_match);
 210         if (!ctx) {
 211                 DRV_LOG(ERR, "Failed to open IB device \"%s\".",
 212                         ibv_match->name);
 213                 rte_errno = ENODEV;
 214                 return -rte_errno;
 215         }
 216         priv = rte_zmalloc("mlx5 vDPA device private", sizeof(*priv),
 217                            RTE_CACHE_LINE_SIZE);
 218         if (!priv) {
 219                 DRV_LOG(ERR, "Failed to allocate private memory.");
 220                 rte_errno = ENOMEM;
 221                 goto error;
 222         }
 223         ret = mlx5_devx_cmd_query_hca_attr(ctx, &attr);
 224         if (ret) {
 225                 DRV_LOG(ERR, "Unable to read HCA capabilities.");
 226                 rte_errno = ENOTSUP;
 227                 goto error;
 228         } else {
 229                 if (!attr.vdpa.valid || !attr.vdpa.max_num_virtio_queues) {
 230                         DRV_LOG(ERR, "Not enough capabilities to support vdpa,"
 231                                 " maybe old FW/OFED version?");
 232                         rte_errno = ENOTSUP;
 233                         goto error;
 234                 }
 235                 priv->caps = attr.vdpa;
 236         }
 237         priv->ctx = ctx;
 238         priv->dev_addr.pci_addr = pci_dev->addr;
 239         priv->dev_addr.type = PCI_ADDR;
 240         priv->id = rte_vdpa_register_device(&priv->dev_addr, &mlx5_vdpa_ops);
 241         if (priv->id < 0) {
 242                 DRV_LOG(ERR, "Failed to register vDPA device.");
 243                 rte_errno = rte_errno ? rte_errno : EINVAL;
 244                 goto error;
 245         }
 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");