drivers/vdpa/mlx5/mlx5_vdpa.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright 2019 Mellanox Technologies, Ltd
   3  */
   4 #include <unistd.h>
   5 #include <net/if.h>
   6 #include <sys/socket.h>
   7 #include <sys/ioctl.h>
   8 #include <fcntl.h>
   9 #include <netinet/in.h>
  10
  11 #include <rte_malloc.h>
  12 #include <rte_log.h>
  13 #include <rte_errno.h>
  14 #include <rte_pci.h>
  15 #include <rte_string_fns.h>
  16
  17 #include <mlx5_glue.h>
  18 #include <mlx5_common.h>
  19 #include <mlx5_common_pci.h>
  20 #include <mlx5_devx_cmds.h>
  21 #include <mlx5_prm.h>
  22 #include <mlx5_nl.h>
  23
  24 #include "mlx5_vdpa_utils.h"
  25 #include "mlx5_vdpa.h"
  26
  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                             (1ULL << VHOST_F_LOG_ALL) | \
  34                             (1ULL << VIRTIO_NET_F_MTU))
  35
  36 #define MLX5_VDPA_PROTOCOL_FEATURES \
  37                             ((1ULL << VHOST_USER_PROTOCOL_F_SLAVE_REQ) | \
  38                              (1ULL << VHOST_USER_PROTOCOL_F_SLAVE_SEND_FD) | \
  39                              (1ULL << VHOST_USER_PROTOCOL_F_HOST_NOTIFIER) | \
  40                              (1ULL << VHOST_USER_PROTOCOL_F_LOG_SHMFD) | \
  41                              (1ULL << VHOST_USER_PROTOCOL_F_MQ) | \
  42                              (1ULL << VHOST_USER_PROTOCOL_F_NET_MTU) | \
  43                              (1ULL << VHOST_USER_PROTOCOL_F_STATUS))
  44
  45 #define MLX5_VDPA_MAX_RETRIES 20
  46 #define MLX5_VDPA_USEC 1000
  47 #define MLX5_VDPA_DEFAULT_NO_TRAFFIC_TIME_S 2LLU
  48
  49 TAILQ_HEAD(mlx5_vdpa_privs, mlx5_vdpa_priv) priv_list =
  50                                               TAILQ_HEAD_INITIALIZER(priv_list);
  51 static pthread_mutex_t priv_list_lock = PTHREAD_MUTEX_INITIALIZER;
  52
  53 static struct mlx5_vdpa_priv *
  54 mlx5_vdpa_find_priv_resource_by_vdev(struct rte_vdpa_device *vdev)
  55 {
  56         struct mlx5_vdpa_priv *priv;
  57         int found = 0;
  58
  59         pthread_mutex_lock(&priv_list_lock);
  60         TAILQ_FOREACH(priv, &priv_list, next) {
  61                 if (vdev == priv->vdev) {
  62                         found = 1;
  63                         break;
  64                 }
  65         }
  66         pthread_mutex_unlock(&priv_list_lock);
  67         if (!found) {
  68                 DRV_LOG(ERR, "Invalid vDPA device: %s.", vdev->device->name);
  69                 rte_errno = EINVAL;
  70                 return NULL;
  71         }
  72         return priv;
  73 }
  74
  75 static int
  76 mlx5_vdpa_get_queue_num(struct rte_vdpa_device *vdev, uint32_t *queue_num)
  77 {
  78         struct mlx5_vdpa_priv *priv =
  79                 mlx5_vdpa_find_priv_resource_by_vdev(vdev);
  80
  81         if (priv == NULL) {
  82                 DRV_LOG(ERR, "Invalid vDPA device: %s.", vdev->device->name);
  83                 return -1;
  84         }
  85         *queue_num = priv->caps.max_num_virtio_queues;
  86         return 0;
  87 }
  88
  89 static int
  90 mlx5_vdpa_get_vdpa_features(struct rte_vdpa_device *vdev, uint64_t *features)
  91 {
  92         struct mlx5_vdpa_priv *priv =
  93                 mlx5_vdpa_find_priv_resource_by_vdev(vdev);
  94
  95         if (priv == NULL) {
  96                 DRV_LOG(ERR, "Invalid vDPA device: %s.", vdev->device->name);
  97                 return -1;
  98         }
  99         *features = MLX5_VDPA_DEFAULT_FEATURES;
 100         if (priv->caps.virtio_queue_type & (1 << MLX5_VIRTQ_TYPE_PACKED))
 101                 *features |= (1ULL << VIRTIO_F_RING_PACKED);
 102         if (priv->caps.tso_ipv4)
 103                 *features |= (1ULL << VIRTIO_NET_F_HOST_TSO4);
 104         if (priv->caps.tso_ipv6)
 105                 *features |= (1ULL << VIRTIO_NET_F_HOST_TSO6);
 106         if (priv->caps.tx_csum)
 107                 *features |= (1ULL << VIRTIO_NET_F_CSUM);
 108         if (priv->caps.rx_csum)
 109                 *features |= (1ULL << VIRTIO_NET_F_GUEST_CSUM);
 110         if (priv->caps.virtio_version_1_0)
 111                 *features |= (1ULL << VIRTIO_F_VERSION_1);
 112         return 0;
 113 }
 114
 115 static int
 116 mlx5_vdpa_get_protocol_features(struct rte_vdpa_device *vdev,
 117                 uint64_t *features)
 118 {
 119         struct mlx5_vdpa_priv *priv =
 120                 mlx5_vdpa_find_priv_resource_by_vdev(vdev);
 121
 122         if (priv == NULL) {
 123                 DRV_LOG(ERR, "Invalid vDPA device: %s.", vdev->device->name);
 124                 return -1;
 125         }
 126         *features = MLX5_VDPA_PROTOCOL_FEATURES;
 127         return 0;
 128 }
 129
 130 static int
 131 mlx5_vdpa_set_vring_state(int vid, int vring, int state)
 132 {
 133         struct rte_vdpa_device *vdev = rte_vhost_get_vdpa_device(vid);
 134         struct mlx5_vdpa_priv *priv =
 135                 mlx5_vdpa_find_priv_resource_by_vdev(vdev);
 136         int ret;
 137
 138         if (priv == NULL) {
 139                 DRV_LOG(ERR, "Invalid vDPA device: %s.", vdev->device->name);
 140                 return -EINVAL;
 141         }
 142         if (vring >= (int)priv->caps.max_num_virtio_queues * 2) {
 143                 DRV_LOG(ERR, "Too big vring id: %d.", vring);
 144                 return -E2BIG;
 145         }
 146         pthread_mutex_lock(&priv->vq_config_lock);
 147         ret = mlx5_vdpa_virtq_enable(priv, vring, state);
 148         pthread_mutex_unlock(&priv->vq_config_lock);
 149         return ret;
 150 }
 151
 152 static int
 153 mlx5_vdpa_features_set(int vid)
 154 {
 155         struct rte_vdpa_device *vdev = rte_vhost_get_vdpa_device(vid);
 156         struct mlx5_vdpa_priv *priv =
 157                 mlx5_vdpa_find_priv_resource_by_vdev(vdev);
 158         uint64_t log_base, log_size;
 159         uint64_t features;
 160         int ret;
 161
 162         if (priv == NULL) {
 163                 DRV_LOG(ERR, "Invalid vDPA device: %s.", vdev->device->name);
 164                 return -EINVAL;
 165         }
 166         ret = rte_vhost_get_negotiated_features(vid, &features);
 167         if (ret) {
 168                 DRV_LOG(ERR, "Failed to get negotiated features.");
 169                 return ret;
 170         }
 171         if (RTE_VHOST_NEED_LOG(features)) {
 172                 ret = rte_vhost_get_log_base(vid, &log_base, &log_size);
 173                 if (ret) {
 174                         DRV_LOG(ERR, "Failed to get log base.");
 175                         return ret;
 176                 }
 177                 ret = mlx5_vdpa_dirty_bitmap_set(priv, log_base, log_size);
 178                 if (ret) {
 179                         DRV_LOG(ERR, "Failed to set dirty bitmap.");
 180                         return ret;
 181                 }
 182                 DRV_LOG(INFO, "mlx5 vdpa: enabling dirty logging...");
 183                 ret = mlx5_vdpa_logging_enable(priv, 1);
 184                 if (ret) {
 185                         DRV_LOG(ERR, "Failed t enable dirty logging.");
 186                         return ret;
 187                 }
 188         }
 189         return 0;
 190 }
 191
 192 static int
 193 mlx5_vdpa_pd_create(struct mlx5_vdpa_priv *priv)
 194 {
 195 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
 196         priv->pd = mlx5_glue->alloc_pd(priv->ctx);
 197         if (priv->pd == NULL) {
 198                 DRV_LOG(ERR, "Failed to allocate PD.");
 199                 return errno ? -errno : -ENOMEM;
 200         }
 201         struct mlx5dv_obj obj;
 202         struct mlx5dv_pd pd_info;
 203         int ret = 0;
 204
 205         obj.pd.in = priv->pd;
 206         obj.pd.out = &pd_info;
 207         ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_PD);
 208         if (ret) {
 209                 DRV_LOG(ERR, "Fail to get PD object info.");
 210                 mlx5_glue->dealloc_pd(priv->pd);
 211                 priv->pd = NULL;
 212                 return -errno;
 213         }
 214         priv->pdn = pd_info.pdn;
 215         return 0;
 216 #else
 217         (void)priv;
 218         DRV_LOG(ERR, "Cannot get pdn - no DV support.");
 219         return -ENOTSUP;
 220 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
 221 }
 222
 223 static int
 224 mlx5_vdpa_mtu_set(struct mlx5_vdpa_priv *priv)
 225 {
 226         struct ifreq request;
 227         uint16_t vhost_mtu = 0;
 228         uint16_t kern_mtu = 0;
 229         int ret = rte_vhost_get_mtu(priv->vid, &vhost_mtu);
 230         int sock;
 231         int retries = MLX5_VDPA_MAX_RETRIES;
 232
 233         if (ret) {
 234                 DRV_LOG(DEBUG, "Cannot get vhost MTU - %d.", ret);
 235                 return ret;
 236         }
 237         if (!vhost_mtu) {
 238                 DRV_LOG(DEBUG, "Vhost MTU is 0.");
 239                 return ret;
 240         }
 241         ret = mlx5_get_ifname_sysfs(priv->ctx->device->ibdev_path,
 242                                     request.ifr_name);
 243         if (ret) {
 244                 DRV_LOG(DEBUG, "Cannot get kernel IF name - %d.", ret);
 245                 return ret;
 246         }
 247         sock = socket(PF_INET, SOCK_DGRAM, IPPROTO_IP);
 248         if (sock == -1) {
 249                 DRV_LOG(DEBUG, "Cannot open IF socket.");
 250                 return sock;
 251         }
 252         while (retries--) {
 253                 ret = ioctl(sock, SIOCGIFMTU, &request);
 254                 if (ret == -1)
 255                         break;
 256                 kern_mtu = request.ifr_mtu;
 257                 DRV_LOG(DEBUG, "MTU: current %d requested %d.", (int)kern_mtu,
 258                         (int)vhost_mtu);
 259                 if (kern_mtu == vhost_mtu)
 260                         break;
 261                 request.ifr_mtu = vhost_mtu;
 262                 ret = ioctl(sock, SIOCSIFMTU, &request);
 263                 if (ret == -1)
 264                         break;
 265                 request.ifr_mtu = 0;
 266                 usleep(MLX5_VDPA_USEC);
 267         }
 268         close(sock);
 269         return kern_mtu == vhost_mtu ? 0 : -1;
 270 }
 271
 272 static int
 273 mlx5_vdpa_dev_close(int vid)
 274 {
 275         struct rte_vdpa_device *vdev = rte_vhost_get_vdpa_device(vid);
 276         struct mlx5_vdpa_priv *priv =
 277                 mlx5_vdpa_find_priv_resource_by_vdev(vdev);
 278         int ret = 0;
 279
 280         if (priv == NULL) {
 281                 DRV_LOG(ERR, "Invalid vDPA device: %s.", vdev->device->name);
 282                 return -1;
 283         }
 284         if (priv->configured)
 285                 ret |= mlx5_vdpa_lm_log(priv);
 286         mlx5_vdpa_err_event_unset(priv);
 287         mlx5_vdpa_cqe_event_unset(priv);
 288         mlx5_vdpa_steer_unset(priv);
 289         mlx5_vdpa_virtqs_release(priv);
 290         mlx5_vdpa_event_qp_global_release(priv);
 291         mlx5_vdpa_mem_dereg(priv);
 292         if (priv->pd) {
 293                 claim_zero(mlx5_glue->dealloc_pd(priv->pd));
 294                 priv->pd = NULL;
 295         }
 296         priv->configured = 0;
 297         priv->vid = 0;
 298         DRV_LOG(INFO, "vDPA device %d was closed.", vid);
 299         return ret;
 300 }
 301
 302 static int
 303 mlx5_vdpa_dev_config(int vid)
 304 {
 305         struct rte_vdpa_device *vdev = rte_vhost_get_vdpa_device(vid);
 306         struct mlx5_vdpa_priv *priv =
 307                 mlx5_vdpa_find_priv_resource_by_vdev(vdev);
 308
 309         if (priv == NULL) {
 310                 DRV_LOG(ERR, "Invalid vDPA device: %s.", vdev->device->name);
 311                 return -EINVAL;
 312         }
 313         if (priv->configured && mlx5_vdpa_dev_close(vid)) {
 314                 DRV_LOG(ERR, "Failed to reconfigure vid %d.", vid);
 315                 return -1;
 316         }
 317         priv->vid = vid;
 318         if (mlx5_vdpa_mtu_set(priv))
 319                 DRV_LOG(WARNING, "MTU cannot be set on device %s.",
 320                                 vdev->device->name);
 321         if (mlx5_vdpa_pd_create(priv) || mlx5_vdpa_mem_register(priv) ||
 322             mlx5_vdpa_err_event_setup(priv) ||
 323             mlx5_vdpa_virtqs_prepare(priv) || mlx5_vdpa_steer_setup(priv) ||
 324             mlx5_vdpa_cqe_event_setup(priv)) {
 325                 mlx5_vdpa_dev_close(vid);
 326                 return -1;
 327         }
 328         priv->configured = 1;
 329         DRV_LOG(INFO, "vDPA device %d was configured.", vid);
 330         return 0;
 331 }
 332
 333 static int
 334 mlx5_vdpa_get_device_fd(int vid)
 335 {
 336         struct rte_vdpa_device *vdev = rte_vhost_get_vdpa_device(vid);
 337         struct mlx5_vdpa_priv *priv =
 338                 mlx5_vdpa_find_priv_resource_by_vdev(vdev);
 339
 340         if (priv == NULL) {
 341                 DRV_LOG(ERR, "Invalid vDPA device: %s.", vdev->device->name);
 342                 return -EINVAL;
 343         }
 344         return priv->ctx->cmd_fd;
 345 }
 346
 347 static int
 348 mlx5_vdpa_get_notify_area(int vid, int qid, uint64_t *offset, uint64_t *size)
 349 {
 350         struct rte_vdpa_device *vdev = rte_vhost_get_vdpa_device(vid);
 351         struct mlx5_vdpa_priv *priv =
 352                 mlx5_vdpa_find_priv_resource_by_vdev(vdev);
 353
 354         RTE_SET_USED(qid);
 355         if (priv == NULL) {
 356                 DRV_LOG(ERR, "Invalid vDPA device: %s.", vdev->device->name);
 357                 return -EINVAL;
 358         }
 359         if (!priv->var) {
 360                 DRV_LOG(ERR, "VAR was not created for device %s, is the device"
 361                         " configured?.", vdev->device->name);
 362                 return -EINVAL;
 363         }
 364         *offset = priv->var->mmap_off;
 365         *size = priv->var->length;
 366         return 0;
 367 }
 368
 369 static int
 370 mlx5_vdpa_get_stats_names(struct rte_vdpa_device *vdev,
 371                 struct rte_vdpa_stat_name *stats_names,
 372                 unsigned int size)
 373 {
 374         static const char *mlx5_vdpa_stats_names[MLX5_VDPA_STATS_MAX] = {
 375                 "received_descriptors",
 376                 "completed_descriptors",
 377                 "bad descriptor errors",
 378                 "exceed max chain",
 379                 "invalid buffer",
 380                 "completion errors",
 381         };
 382         struct mlx5_vdpa_priv *priv =
 383                 mlx5_vdpa_find_priv_resource_by_vdev(vdev);
 384         unsigned int i;
 385
 386         if (priv == NULL) {
 387                 DRV_LOG(ERR, "Invalid device: %s.", vdev->device->name);
 388                 return -ENODEV;
 389         }
 390         if (!stats_names)
 391                 return MLX5_VDPA_STATS_MAX;
 392         size = RTE_MIN(size, (unsigned int)MLX5_VDPA_STATS_MAX);
 393         for (i = 0; i < size; ++i)
 394                 strlcpy(stats_names[i].name, mlx5_vdpa_stats_names[i],
 395                         RTE_VDPA_STATS_NAME_SIZE);
 396         return size;
 397 }
 398
 399 static int
 400 mlx5_vdpa_get_stats(struct rte_vdpa_device *vdev, int qid,
 401                 struct rte_vdpa_stat *stats, unsigned int n)
 402 {
 403         struct mlx5_vdpa_priv *priv =
 404                 mlx5_vdpa_find_priv_resource_by_vdev(vdev);
 405
 406         if (priv == NULL) {
 407                 DRV_LOG(ERR, "Invalid device: %s.", vdev->device->name);
 408                 return -ENODEV;
 409         }
 410         if (!priv->configured) {
 411                 DRV_LOG(ERR, "Device %s was not configured.",
 412                                 vdev->device->name);
 413                 return -ENODATA;
 414         }
 415         if (qid >= (int)priv->nr_virtqs) {
 416                 DRV_LOG(ERR, "Too big vring id: %d for device %s.", qid,
 417                                 vdev->device->name);
 418                 return -E2BIG;
 419         }
 420         if (!priv->caps.queue_counters_valid) {
 421                 DRV_LOG(ERR, "Virtq statistics is not supported for device %s.",
 422                         vdev->device->name);
 423                 return -ENOTSUP;
 424         }
 425         return mlx5_vdpa_virtq_stats_get(priv, qid, stats, n);
 426 }
 427
 428 static int
 429 mlx5_vdpa_reset_stats(struct rte_vdpa_device *vdev, int qid)
 430 {
 431         struct mlx5_vdpa_priv *priv =
 432                 mlx5_vdpa_find_priv_resource_by_vdev(vdev);
 433
 434         if (priv == NULL) {
 435                 DRV_LOG(ERR, "Invalid device: %s.", vdev->device->name);
 436                 return -ENODEV;
 437         }
 438         if (!priv->configured) {
 439                 DRV_LOG(ERR, "Device %s was not configured.",
 440                                 vdev->device->name);
 441                 return -ENODATA;
 442         }
 443         if (qid >= (int)priv->nr_virtqs) {
 444                 DRV_LOG(ERR, "Too big vring id: %d for device %s.", qid,
 445                                 vdev->device->name);
 446                 return -E2BIG;
 447         }
 448         if (!priv->caps.queue_counters_valid) {
 449                 DRV_LOG(ERR, "Virtq statistics is not supported for device %s.",
 450                         vdev->device->name);
 451                 return -ENOTSUP;
 452         }
 453         return mlx5_vdpa_virtq_stats_reset(priv, qid);
 454 }
 455
 456 static struct rte_vdpa_dev_ops mlx5_vdpa_ops = {
 457         .get_queue_num = mlx5_vdpa_get_queue_num,
 458         .get_features = mlx5_vdpa_get_vdpa_features,
 459         .get_protocol_features = mlx5_vdpa_get_protocol_features,
 460         .dev_conf = mlx5_vdpa_dev_config,
 461         .dev_close = mlx5_vdpa_dev_close,
 462         .set_vring_state = mlx5_vdpa_set_vring_state,
 463         .set_features = mlx5_vdpa_features_set,
 464         .migration_done = NULL,
 465         .get_vfio_group_fd = NULL,
 466         .get_vfio_device_fd = mlx5_vdpa_get_device_fd,
 467         .get_notify_area = mlx5_vdpa_get_notify_area,
 468         .get_stats_names = mlx5_vdpa_get_stats_names,
 469         .get_stats = mlx5_vdpa_get_stats,
 470         .reset_stats = mlx5_vdpa_reset_stats,
 471 };
 472
 473 static struct ibv_device *
 474 mlx5_vdpa_get_ib_device_match(struct rte_pci_addr *addr)
 475 {
 476         int n;
 477         struct ibv_device **ibv_list = mlx5_glue->get_device_list(&n);
 478         struct ibv_device *ibv_match = NULL;
 479
 480         if (!ibv_list) {
 481                 rte_errno = ENOSYS;
 482                 return NULL;
 483         }
 484         while (n-- > 0) {
 485                 struct rte_pci_addr pci_addr;
 486
 487                 DRV_LOG(DEBUG, "Checking device \"%s\"..", ibv_list[n]->name);
 488                 if (mlx5_dev_to_pci_addr(ibv_list[n]->ibdev_path, &pci_addr))
 489                         continue;
 490                 if (rte_pci_addr_cmp(addr, &pci_addr))
 491                         continue;
 492                 ibv_match = ibv_list[n];
 493                 break;
 494         }
 495         if (!ibv_match)
 496                 rte_errno = ENOENT;
 497         mlx5_glue->free_device_list(ibv_list);
 498         return ibv_match;
 499 }
 500
 501 /* Try to disable ROCE by Netlink\Devlink. */
 502 static int
 503 mlx5_vdpa_nl_roce_disable(const char *addr)
 504 {
 505         int nlsk_fd = mlx5_nl_init(NETLINK_GENERIC);
 506         int devlink_id;
 507         int enable;
 508         int ret;
 509
 510         if (nlsk_fd < 0)
 511                 return nlsk_fd;
 512         devlink_id = mlx5_nl_devlink_family_id_get(nlsk_fd);
 513         if (devlink_id < 0) {
 514                 ret = devlink_id;
 515                 DRV_LOG(DEBUG, "Failed to get devlink id for ROCE operations by"
 516                         " Netlink.");
 517                 goto close;
 518         }
 519         ret = mlx5_nl_enable_roce_get(nlsk_fd, devlink_id, addr, &enable);
 520         if (ret) {
 521                 DRV_LOG(DEBUG, "Failed to get ROCE enable by Netlink: %d.",
 522                         ret);
 523                 goto close;
 524         } else if (!enable) {
 525                 DRV_LOG(INFO, "ROCE has already disabled(Netlink).");
 526                 goto close;
 527         }
 528         ret = mlx5_nl_enable_roce_set(nlsk_fd, devlink_id, addr, 0);
 529         if (ret)
 530                 DRV_LOG(DEBUG, "Failed to disable ROCE by Netlink: %d.", ret);
 531         else
 532                 DRV_LOG(INFO, "ROCE is disabled by Netlink successfully.");
 533 close:
 534         close(nlsk_fd);
 535         return ret;
 536 }
 537
 538 /* Try to disable ROCE by sysfs. */
 539 static int
 540 mlx5_vdpa_sys_roce_disable(const char *addr)
 541 {
 542         FILE *file_o;
 543         int enable;
 544         int ret;
 545
 546         MKSTR(file_p, "/sys/bus/pci/devices/%s/roce_enable", addr);
 547         file_o = fopen(file_p, "rb");
 548         if (!file_o) {
 549                 rte_errno = ENOTSUP;
 550                 return -ENOTSUP;
 551         }
 552         ret = fscanf(file_o, "%d", &enable);
 553         if (ret != 1) {
 554                 rte_errno = EINVAL;
 555                 ret = EINVAL;
 556                 goto close;
 557         } else if (!enable) {
 558                 ret = 0;
 559                 DRV_LOG(INFO, "ROCE has already disabled(sysfs).");
 560                 goto close;
 561         }
 562         fclose(file_o);
 563         file_o = fopen(file_p, "wb");
 564         if (!file_o) {
 565                 rte_errno = ENOTSUP;
 566                 return -ENOTSUP;
 567         }
 568         fprintf(file_o, "0\n");
 569         ret = 0;
 570 close:
 571         if (ret)
 572                 DRV_LOG(DEBUG, "Failed to disable ROCE by sysfs: %d.", ret);
 573         else
 574                 DRV_LOG(INFO, "ROCE is disabled by sysfs successfully.");
 575         fclose(file_o);
 576         return ret;
 577 }
 578
 579 static int
 580 mlx5_vdpa_roce_disable(struct rte_pci_addr *addr, struct ibv_device **ibv)
 581 {
 582         char addr_name[64] = {0};
 583
 584         rte_pci_device_name(addr, addr_name, sizeof(addr_name));
 585         /* Firstly try to disable ROCE by Netlink and fallback to sysfs. */
 586         if (mlx5_vdpa_nl_roce_disable(addr_name) == 0 ||
 587             mlx5_vdpa_sys_roce_disable(addr_name) == 0) {
 588                 /*
 589                  * Succeed to disable ROCE, wait for the IB device to appear
 590                  * again after reload.
 591                  */
 592                 int r;
 593                 struct ibv_device *ibv_new;
 594
 595                 for (r = MLX5_VDPA_MAX_RETRIES; r; r--) {
 596                         ibv_new = mlx5_vdpa_get_ib_device_match(addr);
 597                         if (ibv_new) {
 598                                 *ibv = ibv_new;
 599                                 return 0;
 600                         }
 601                         usleep(MLX5_VDPA_USEC);
 602                 }
 603                 DRV_LOG(ERR, "Cannot much device %s after ROCE disable, "
 604                         "retries exceed %d", addr_name, MLX5_VDPA_MAX_RETRIES);
 605                 rte_errno = EAGAIN;
 606         }
 607         return -rte_errno;
 608 }
 609
 610 static int
 611 mlx5_vdpa_args_check_handler(const char *key, const char *val, void *opaque)
 612 {
 613         struct mlx5_vdpa_priv *priv = opaque;
 614         unsigned long tmp;
 615
 616         if (strcmp(key, "class") == 0)
 617                 return 0;
 618         errno = 0;
 619         tmp = strtoul(val, NULL, 0);
 620         if (errno) {
 621                 DRV_LOG(WARNING, "%s: \"%s\" is an invalid integer.", key, val);
 622                 return -errno;
 623         }
 624         if (strcmp(key, "event_mode") == 0) {
 625                 if (tmp <= MLX5_VDPA_EVENT_MODE_ONLY_INTERRUPT)
 626                         priv->event_mode = (int)tmp;
 627                 else
 628                         DRV_LOG(WARNING, "Invalid event_mode %s.", val);
 629         } else if (strcmp(key, "event_us") == 0) {
 630                 priv->event_us = (uint32_t)tmp;
 631         } else if (strcmp(key, "no_traffic_time") == 0) {
 632                 priv->no_traffic_time_s = (uint32_t)tmp;
 633         } else {
 634                 DRV_LOG(WARNING, "Invalid key %s.", key);
 635         }
 636         return 0;
 637 }
 638
 639 static void
 640 mlx5_vdpa_config_get(struct rte_devargs *devargs, struct mlx5_vdpa_priv *priv)
 641 {
 642         struct rte_kvargs *kvlist;
 643
 644         priv->event_mode = MLX5_VDPA_EVENT_MODE_DYNAMIC_TIMER;
 645         priv->event_us = 0;
 646         priv->no_traffic_time_s = MLX5_VDPA_DEFAULT_NO_TRAFFIC_TIME_S;
 647         if (devargs == NULL)
 648                 return;
 649         kvlist = rte_kvargs_parse(devargs->args, NULL);
 650         if (kvlist == NULL)
 651                 return;
 652         rte_kvargs_process(kvlist, NULL, mlx5_vdpa_args_check_handler, priv);
 653         rte_kvargs_free(kvlist);
 654         if (!priv->event_us &&
 655             priv->event_mode == MLX5_VDPA_EVENT_MODE_DYNAMIC_TIMER)
 656                 priv->event_us = MLX5_VDPA_DEFAULT_TIMER_STEP_US;
 657         DRV_LOG(DEBUG, "event mode is %d.", priv->event_mode);
 658         DRV_LOG(DEBUG, "event_us is %u us.", priv->event_us);
 659         DRV_LOG(DEBUG, "no traffic time is %u s.", priv->no_traffic_time_s);
 660 }
 661
 662 /**
 663  * DPDK callback to register a mlx5 PCI device.
 664  *
 665  * This function spawns vdpa device out of a given PCI device.
 666  *
 667  * @param[in] pci_drv
 668  *   PCI driver structure (mlx5_vpda_driver).
 669  * @param[in] pci_dev
 670  *   PCI device information.
 671  *
 672  * @return
 673  *   0 on success, 1 to skip this driver, a negative errno value otherwise
 674  *   and rte_errno is set.
 675  */
 676 static int
 677 mlx5_vdpa_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
 678                     struct rte_pci_device *pci_dev __rte_unused)
 679 {
 680         struct ibv_device *ibv;
 681         struct mlx5_vdpa_priv *priv = NULL;
 682         struct ibv_context *ctx = NULL;
 683         struct mlx5_hca_attr attr;
 684         int ret;
 685
 686         ibv = mlx5_vdpa_get_ib_device_match(&pci_dev->addr);
 687         if (!ibv) {
 688                 DRV_LOG(ERR, "No matching IB device for PCI slot "
 689                         PCI_PRI_FMT ".", pci_dev->addr.domain,
 690                         pci_dev->addr.bus, pci_dev->addr.devid,
 691                         pci_dev->addr.function);
 692                 return -rte_errno;
 693         } else {
 694                 DRV_LOG(INFO, "PCI information matches for device \"%s\".",
 695                         ibv->name);
 696         }
 697         if (mlx5_vdpa_roce_disable(&pci_dev->addr, &ibv) != 0) {
 698                 DRV_LOG(WARNING, "Failed to disable ROCE for \"%s\".",
 699                         ibv->name);
 700                 return -rte_errno;
 701         }
 702         ctx = mlx5_glue->dv_open_device(ibv);
 703         if (!ctx) {
 704                 DRV_LOG(ERR, "Failed to open IB device \"%s\".", ibv->name);
 705                 rte_errno = ENODEV;
 706                 return -rte_errno;
 707         }
 708         ret = mlx5_devx_cmd_query_hca_attr(ctx, &attr);
 709         if (ret) {
 710                 DRV_LOG(ERR, "Unable to read HCA capabilities.");
 711                 rte_errno = ENOTSUP;
 712                 goto error;
 713         } else if (!attr.vdpa.valid || !attr.vdpa.max_num_virtio_queues) {
 714                 DRV_LOG(ERR, "Not enough capabilities to support vdpa, maybe "
 715                         "old FW/OFED version?");
 716                 rte_errno = ENOTSUP;
 717                 goto error;
 718         }
 719         if (!attr.vdpa.queue_counters_valid)
 720                 DRV_LOG(DEBUG, "No capability to support virtq statistics.");
 721         priv = rte_zmalloc("mlx5 vDPA device private", sizeof(*priv) +
 722                            sizeof(struct mlx5_vdpa_virtq) *
 723                            attr.vdpa.max_num_virtio_queues * 2,
 724                            RTE_CACHE_LINE_SIZE);
 725         if (!priv) {
 726                 DRV_LOG(ERR, "Failed to allocate private memory.");
 727                 rte_errno = ENOMEM;
 728                 goto error;
 729         }
 730         priv->caps = attr.vdpa;
 731         priv->log_max_rqt_size = attr.log_max_rqt_size;
 732         priv->num_lag_ports = attr.num_lag_ports;
 733         if (attr.num_lag_ports == 0)
 734                 priv->num_lag_ports = 1;
 735         priv->ctx = ctx;
 736         priv->pci_dev = pci_dev;
 737         priv->var = mlx5_glue->dv_alloc_var(ctx, 0);
 738         if (!priv->var) {
 739                 DRV_LOG(ERR, "Failed to allocate VAR %u.\n", errno);
 740                 goto error;
 741         }
 742         priv->vdev = rte_vdpa_register_device(&pci_dev->device,
 743                         &mlx5_vdpa_ops);
 744         if (priv->vdev == NULL) {
 745                 DRV_LOG(ERR, "Failed to register vDPA device.");
 746                 rte_errno = rte_errno ? rte_errno : EINVAL;
 747                 goto error;
 748         }
 749         mlx5_vdpa_config_get(pci_dev->device.devargs, priv);
 750         SLIST_INIT(&priv->mr_list);
 751         pthread_mutex_init(&priv->vq_config_lock, NULL);
 752         pthread_mutex_lock(&priv_list_lock);
 753         TAILQ_INSERT_TAIL(&priv_list, priv, next);
 754         pthread_mutex_unlock(&priv_list_lock);
 755         return 0;
 756
 757 error:
 758         if (priv) {
 759                 if (priv->var)
 760                         mlx5_glue->dv_free_var(priv->var);
 761                 rte_free(priv);
 762         }
 763         if (ctx)
 764                 mlx5_glue->close_device(ctx);
 765         return -rte_errno;
 766 }
 767
 768 /**
 769  * DPDK callback to remove a PCI device.
 770  *
 771  * This function removes all vDPA devices belong to a given PCI device.
 772  *
 773  * @param[in] pci_dev
 774  *   Pointer to the PCI device.
 775  *
 776  * @return
 777  *   0 on success, the function cannot fail.
 778  */
 779 static int
 780 mlx5_vdpa_pci_remove(struct rte_pci_device *pci_dev)
 781 {
 782         struct mlx5_vdpa_priv *priv = NULL;
 783         int found = 0;
 784
 785         pthread_mutex_lock(&priv_list_lock);
 786         TAILQ_FOREACH(priv, &priv_list, next) {
 787                 if (!rte_pci_addr_cmp(&priv->pci_dev->addr, &pci_dev->addr)) {
 788                         found = 1;
 789                         break;
 790                 }
 791         }
 792         if (found)
 793                 TAILQ_REMOVE(&priv_list, priv, next);
 794         pthread_mutex_unlock(&priv_list_lock);
 795         if (found) {
 796                 if (priv->configured)
 797                         mlx5_vdpa_dev_close(priv->vid);
 798                 if (priv->var) {
 799                         mlx5_glue->dv_free_var(priv->var);
 800                         priv->var = NULL;
 801                 }
 802                 mlx5_glue->close_device(priv->ctx);
 803                 pthread_mutex_destroy(&priv->vq_config_lock);
 804                 rte_free(priv);
 805         }
 806         return 0;
 807 }
 808
 809 static const struct rte_pci_id mlx5_vdpa_pci_id_map[] = {
 810         {
 811                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
 812                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6)
 813         },
 814         {
 815                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
 816                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6VF)
 817         },
 818         {
 819                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
 820                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6DX)
 821         },
 822         {
 823                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
 824                                 PCI_DEVICE_ID_MELLANOX_CONNECTXVF)
 825         },
 826         {
 827                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
 828                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6DXBF)
 829         },
 830         {
 831                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
 832                                 PCI_DEVICE_ID_MELLANOX_CONNECTX7)
 833         },
 834         {
 835                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
 836                                 PCI_DEVICE_ID_MELLANOX_CONNECTX7BF)
 837         },
 838         {
 839                 .vendor_id = 0
 840         }
 841 };
 842
 843 static struct mlx5_pci_driver mlx5_vdpa_driver = {
 844         .driver_class = MLX5_CLASS_VDPA,
 845         .pci_driver = {
 846                 .driver = {
 847                         .name = "mlx5_vdpa",
 848                 },
 849                 .id_table = mlx5_vdpa_pci_id_map,
 850                 .probe = mlx5_vdpa_pci_probe,
 851                 .remove = mlx5_vdpa_pci_remove,
 852                 .drv_flags = 0,
 853         },
 854 };
 855
 856 RTE_LOG_REGISTER(mlx5_vdpa_logtype, pmd.vdpa.mlx5, NOTICE)
 857
 858 /**
 859  * Driver initialization routine.
 860  */
 861 RTE_INIT(rte_mlx5_vdpa_init)
 862 {
 863         mlx5_common_init();
 864         if (mlx5_glue)
 865                 mlx5_pci_driver_register(&mlx5_vdpa_driver);
 866 }
 867
 868 RTE_PMD_EXPORT_NAME(net_mlx5_vdpa, __COUNTER__);
 869 RTE_PMD_REGISTER_PCI_TABLE(net_mlx5_vdpa, mlx5_vdpa_pci_id_map);
 870 RTE_PMD_REGISTER_KMOD_DEP(net_mlx5_vdpa, "* ib_uverbs & mlx5_core & mlx5_ib");