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
  11 #include <mlx5_glue.h>
  12 #include <mlx5_common.h>
  13 #include <mlx5_devx_cmds.h>
  14 #include <mlx5_prm.h>
  15
  16 #include "mlx5_vdpa_utils.h"
  17 #include "mlx5_vdpa.h"
  18
  19
  20 #ifndef VIRTIO_F_ORDER_PLATFORM
  21 #define VIRTIO_F_ORDER_PLATFORM 36
  22 #endif
  23
  24 #ifndef VIRTIO_F_RING_PACKED
  25 #define VIRTIO_F_RING_PACKED 34
  26 #endif
  27
  28 #define MLX5_VDPA_DEFAULT_FEATURES ((1ULL << VHOST_USER_F_PROTOCOL_FEATURES) | \
  29                             (1ULL << VIRTIO_F_ANY_LAYOUT) | \
  30                             (1ULL << VIRTIO_NET_F_MQ) | \
  31                             (1ULL << VIRTIO_NET_F_GUEST_ANNOUNCE) | \
  32                             (1ULL << VIRTIO_F_ORDER_PLATFORM))
  33
  34 #define MLX5_VDPA_PROTOCOL_FEATURES \
  35                             ((1ULL << VHOST_USER_PROTOCOL_F_SLAVE_REQ) | \
  36                              (1ULL << VHOST_USER_PROTOCOL_F_SLAVE_SEND_FD) | \
  37                              (1ULL << VHOST_USER_PROTOCOL_F_HOST_NOTIFIER) | \
  38                              (1ULL << VHOST_USER_PROTOCOL_F_LOG_SHMFD) | \
  39                              (1ULL << VHOST_USER_PROTOCOL_F_MQ))
  40
  41 TAILQ_HEAD(mlx5_vdpa_privs, mlx5_vdpa_priv) priv_list =
  42                                               TAILQ_HEAD_INITIALIZER(priv_list);
  43 static pthread_mutex_t priv_list_lock = PTHREAD_MUTEX_INITIALIZER;
  44 int mlx5_vdpa_logtype;
  45
  46 static struct mlx5_vdpa_priv *
  47 mlx5_vdpa_find_priv_resource_by_did(int did)
  48 {
  49         struct mlx5_vdpa_priv *priv;
  50         int found = 0;
  51
  52         pthread_mutex_lock(&priv_list_lock);
  53         TAILQ_FOREACH(priv, &priv_list, next) {
  54                 if (did == priv->id) {
  55                         found = 1;
  56                         break;
  57                 }
  58         }
  59         pthread_mutex_unlock(&priv_list_lock);
  60         if (!found) {
  61                 DRV_LOG(ERR, "Invalid device id: %d.", did);
  62                 rte_errno = EINVAL;
  63                 return NULL;
  64         }
  65         return priv;
  66 }
  67
  68 static int
  69 mlx5_vdpa_get_queue_num(int did, uint32_t *queue_num)
  70 {
  71         struct mlx5_vdpa_priv *priv = mlx5_vdpa_find_priv_resource_by_did(did);
  72
  73         if (priv == NULL) {
  74                 DRV_LOG(ERR, "Invalid device id: %d.", did);
  75                 return -1;
  76         }
  77         *queue_num = priv->caps.max_num_virtio_queues;
  78         return 0;
  79 }
  80
  81 static int
  82 mlx5_vdpa_get_vdpa_features(int did, uint64_t *features)
  83 {
  84         struct mlx5_vdpa_priv *priv = mlx5_vdpa_find_priv_resource_by_did(did);
  85
  86         if (priv == NULL) {
  87                 DRV_LOG(ERR, "Invalid device id: %d.", did);
  88                 return -1;
  89         }
  90         *features = MLX5_VDPA_DEFAULT_FEATURES;
  91         if (priv->caps.virtio_queue_type & (1 << MLX5_VIRTQ_TYPE_PACKED))
  92                 *features |= (1ULL << VIRTIO_F_RING_PACKED);
  93         if (priv->caps.tso_ipv4)
  94                 *features |= (1ULL << VIRTIO_NET_F_HOST_TSO4);
  95         if (priv->caps.tso_ipv6)
  96                 *features |= (1ULL << VIRTIO_NET_F_HOST_TSO6);
  97         if (priv->caps.tx_csum)
  98                 *features |= (1ULL << VIRTIO_NET_F_CSUM);
  99         if (priv->caps.rx_csum)
 100                 *features |= (1ULL << VIRTIO_NET_F_GUEST_CSUM);
 101         if (priv->caps.virtio_version_1_0)
 102                 *features |= (1ULL << VIRTIO_F_VERSION_1);
 103         return 0;
 104 }
 105
 106 static int
 107 mlx5_vdpa_get_protocol_features(int did, uint64_t *features)
 108 {
 109         struct mlx5_vdpa_priv *priv = mlx5_vdpa_find_priv_resource_by_did(did);
 110
 111         if (priv == NULL) {
 112                 DRV_LOG(ERR, "Invalid device id: %d.", did);
 113                 return -1;
 114         }
 115         *features = MLX5_VDPA_PROTOCOL_FEATURES;
 116         return 0;
 117 }
 118
 119 static struct rte_vdpa_dev_ops mlx5_vdpa_ops = {
 120         .get_queue_num = mlx5_vdpa_get_queue_num,
 121         .get_features = mlx5_vdpa_get_vdpa_features,
 122         .get_protocol_features = mlx5_vdpa_get_protocol_features,
 123         .dev_conf = NULL,
 124         .dev_close = NULL,
 125         .set_vring_state = NULL,
 126         .set_features = NULL,
 127         .migration_done = NULL,
 128         .get_vfio_group_fd = NULL,
 129         .get_vfio_device_fd = NULL,
 130         .get_notify_area = NULL,
 131 };
 132
 133 /**
 134  * DPDK callback to register a PCI device.
 135  *
 136  * This function spawns vdpa device out of a given PCI device.
 137  *
 138  * @param[in] pci_drv
 139  *   PCI driver structure (mlx5_vpda_driver).
 140  * @param[in] pci_dev
 141  *   PCI device information.
 142  *
 143  * @return
 144  *   0 on success, 1 to skip this driver, a negative errno value otherwise
 145  *   and rte_errno is set.
 146  */
 147 static int
 148 mlx5_vdpa_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
 149                     struct rte_pci_device *pci_dev __rte_unused)
 150 {
 151         struct ibv_device **ibv_list;
 152         struct ibv_device *ibv_match = NULL;
 153         struct mlx5_vdpa_priv *priv = NULL;
 154         struct ibv_context *ctx = NULL;
 155         struct mlx5_hca_attr attr;
 156         int ret;
 157
 158         if (mlx5_class_get(pci_dev->device.devargs) != MLX5_CLASS_VDPA) {
 159                 DRV_LOG(DEBUG, "Skip probing - should be probed by other mlx5"
 160                         " driver.");
 161                 return 1;
 162         }
 163         errno = 0;
 164         ibv_list = mlx5_glue->get_device_list(&ret);
 165         if (!ibv_list) {
 166                 rte_errno = ENOSYS;
 167                 DRV_LOG(ERR, "Failed to get device list, is ib_uverbs loaded?");
 168                 return -rte_errno;
 169         }
 170         while (ret-- > 0) {
 171                 struct rte_pci_addr pci_addr;
 172
 173                 DRV_LOG(DEBUG, "Checking device \"%s\"..", ibv_list[ret]->name);
 174                 if (mlx5_dev_to_pci_addr(ibv_list[ret]->ibdev_path, &pci_addr))
 175                         continue;
 176                 if (pci_dev->addr.domain != pci_addr.domain ||
 177                     pci_dev->addr.bus != pci_addr.bus ||
 178                     pci_dev->addr.devid != pci_addr.devid ||
 179                     pci_dev->addr.function != pci_addr.function)
 180                         continue;
 181                 DRV_LOG(INFO, "PCI information matches for device \"%s\".",
 182                         ibv_list[ret]->name);
 183                 ibv_match = ibv_list[ret];
 184                 break;
 185         }
 186         mlx5_glue->free_device_list(ibv_list);
 187         if (!ibv_match) {
 188                 DRV_LOG(ERR, "No matching IB device for PCI slot "
 189                         "%" SCNx32 ":%" SCNx8 ":%" SCNx8 ".%" SCNx8 ".",
 190                         pci_dev->addr.domain, pci_dev->addr.bus,
 191                         pci_dev->addr.devid, pci_dev->addr.function);
 192                 rte_errno = ENOENT;
 193                 return -rte_errno;
 194         }
 195         ctx = mlx5_glue->dv_open_device(ibv_match);
 196         if (!ctx) {
 197                 DRV_LOG(ERR, "Failed to open IB device \"%s\".",
 198                         ibv_match->name);
 199                 rte_errno = ENODEV;
 200                 return -rte_errno;
 201         }
 202         priv = rte_zmalloc("mlx5 vDPA device private", sizeof(*priv),
 203                            RTE_CACHE_LINE_SIZE);
 204         if (!priv) {
 205                 DRV_LOG(ERR, "Failed to allocate private memory.");
 206                 rte_errno = ENOMEM;
 207                 goto error;
 208         }
 209         ret = mlx5_devx_cmd_query_hca_attr(ctx, &attr);
 210         if (ret) {
 211                 DRV_LOG(ERR, "Unable to read HCA capabilities.");
 212                 rte_errno = ENOTSUP;
 213                 goto error;
 214         } else {
 215                 if (!attr.vdpa.valid || !attr.vdpa.max_num_virtio_queues) {
 216                         DRV_LOG(ERR, "Not enough capabilities to support vdpa,"
 217                                 " maybe old FW/OFED version?");
 218                         rte_errno = ENOTSUP;
 219                         goto error;
 220                 }
 221                 priv->caps = attr.vdpa;
 222         }
 223         priv->ctx = ctx;
 224         priv->dev_addr.pci_addr = pci_dev->addr;
 225         priv->dev_addr.type = PCI_ADDR;
 226         priv->id = rte_vdpa_register_device(&priv->dev_addr, &mlx5_vdpa_ops);
 227         if (priv->id < 0) {
 228                 DRV_LOG(ERR, "Failed to register vDPA device.");
 229                 rte_errno = rte_errno ? rte_errno : EINVAL;
 230                 goto error;
 231         }
 232         SLIST_INIT(&priv->mr_list);
 233         SLIST_INIT(&priv->virtq_list);
 234         pthread_mutex_lock(&priv_list_lock);
 235         TAILQ_INSERT_TAIL(&priv_list, priv, next);
 236         pthread_mutex_unlock(&priv_list_lock);
 237         return 0;
 238
 239 error:
 240         if (priv)
 241                 rte_free(priv);
 242         if (ctx)
 243                 mlx5_glue->close_device(ctx);
 244         return -rte_errno;
 245 }
 246
 247 /**
 248  * DPDK callback to remove a PCI device.
 249  *
 250  * This function removes all vDPA devices belong to a given PCI device.
 251  *
 252  * @param[in] pci_dev
 253  *   Pointer to the PCI device.
 254  *
 255  * @return
 256  *   0 on success, the function cannot fail.
 257  */
 258 static int
 259 mlx5_vdpa_pci_remove(struct rte_pci_device *pci_dev)
 260 {
 261         struct mlx5_vdpa_priv *priv = NULL;
 262         int found = 0;
 263
 264         pthread_mutex_lock(&priv_list_lock);
 265         TAILQ_FOREACH(priv, &priv_list, next) {
 266                 if (memcmp(&priv->dev_addr.pci_addr, &pci_dev->addr,
 267                            sizeof(pci_dev->addr)) == 0) {
 268                         found = 1;
 269                         break;
 270                 }
 271         }
 272         if (found) {
 273                 TAILQ_REMOVE(&priv_list, priv, next);
 274                 mlx5_glue->close_device(priv->ctx);
 275                 rte_free(priv);
 276         }
 277         pthread_mutex_unlock(&priv_list_lock);
 278         return 0;
 279 }
 280
 281 static const struct rte_pci_id mlx5_vdpa_pci_id_map[] = {
 282         {
 283                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
 284                                PCI_DEVICE_ID_MELLANOX_CONNECTX5BF)
 285         },
 286         {
 287                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
 288                                PCI_DEVICE_ID_MELLANOX_CONNECTX5BFVF)
 289         },
 290         {
 291                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
 292                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6)
 293         },
 294         {
 295                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
 296                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6VF)
 297         },
 298         {
 299                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
 300                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6DX)
 301         },
 302         {
 303                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
 304                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6DXVF)
 305         },
 306         {
 307                 .vendor_id = 0
 308         }
 309 };
 310
 311 static struct rte_pci_driver mlx5_vdpa_driver = {
 312         .driver = {
 313                 .name = "mlx5_vdpa",
 314         },
 315         .id_table = mlx5_vdpa_pci_id_map,
 316         .probe = mlx5_vdpa_pci_probe,
 317         .remove = mlx5_vdpa_pci_remove,
 318         .drv_flags = 0,
 319 };
 320
 321 /**
 322  * Driver initialization routine.
 323  */
 324 RTE_INIT(rte_mlx5_vdpa_init)
 325 {
 326         /* Initialize common log type. */
 327         mlx5_vdpa_logtype = rte_log_register("pmd.vdpa.mlx5");
 328         if (mlx5_vdpa_logtype >= 0)
 329                 rte_log_set_level(mlx5_vdpa_logtype, RTE_LOG_NOTICE);
 330         if (mlx5_glue)
 331                 rte_pci_register(&mlx5_vdpa_driver);
 332 }
 333
 334 RTE_PMD_EXPORT_NAME(net_mlx5_vdpa, __COUNTER__);
 335 RTE_PMD_REGISTER_PCI_TABLE(net_mlx5_vdpa, mlx5_vdpa_pci_id_map);
 336 RTE_PMD_REGISTER_KMOD_DEP(net_mlx5_vdpa, "* ib_uverbs & mlx5_core & mlx5_ib");