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_bus_pci.h>
  15 #include <rte_pci.h>
  16 #include <rte_string_fns.h>
  17
  18 #include <mlx5_glue.h>
  19 #include <mlx5_common.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
  44 #define MLX5_VDPA_MAX_RETRIES 20
  45 #define MLX5_VDPA_USEC 1000
  46
  47 TAILQ_HEAD(mlx5_vdpa_privs, mlx5_vdpa_priv) priv_list =
  48                                               TAILQ_HEAD_INITIALIZER(priv_list);
  49 static pthread_mutex_t priv_list_lock = PTHREAD_MUTEX_INITIALIZER;
  50 int mlx5_vdpa_logtype;
  51
  52 static struct mlx5_vdpa_priv *
  53 mlx5_vdpa_find_priv_resource_by_vdev(struct rte_vdpa_device *vdev)
  54 {
  55         struct mlx5_vdpa_priv *priv;
  56         int found = 0;
  57
  58         pthread_mutex_lock(&priv_list_lock);
  59         TAILQ_FOREACH(priv, &priv_list, next) {
  60                 if (vdev == priv->vdev) {
  61                         found = 1;
  62                         break;
  63                 }
  64         }
  65         pthread_mutex_unlock(&priv_list_lock);
  66         if (!found) {
  67                 DRV_LOG(ERR, "Invalid vDPA device: %s.", vdev->device->name);
  68                 rte_errno = EINVAL;
  69                 return NULL;
  70         }
  71         return priv;
  72 }
  73
  74 static int
  75 mlx5_vdpa_get_queue_num(struct rte_vdpa_device *vdev, uint32_t *queue_num)
  76 {
  77         struct mlx5_vdpa_priv *priv =
  78                 mlx5_vdpa_find_priv_resource_by_vdev(vdev);
  79
  80         if (priv == NULL) {
  81                 DRV_LOG(ERR, "Invalid vDPA device: %s.", vdev->device->name);
  82                 return -1;
  83         }
  84         *queue_num = priv->caps.max_num_virtio_queues;
  85         return 0;
  86 }
  87
  88 static int
  89 mlx5_vdpa_get_vdpa_features(struct rte_vdpa_device *vdev, uint64_t *features)
  90 {
  91         struct mlx5_vdpa_priv *priv =
  92                 mlx5_vdpa_find_priv_resource_by_vdev(vdev);
  93
  94         if (priv == NULL) {
  95                 DRV_LOG(ERR, "Invalid vDPA device: %s.", vdev->device->name);
  96                 return -1;
  97         }
  98         *features = MLX5_VDPA_DEFAULT_FEATURES;
  99         if (priv->caps.virtio_queue_type & (1 << MLX5_VIRTQ_TYPE_PACKED))
 100                 *features |= (1ULL << VIRTIO_F_RING_PACKED);
 101         if (priv->caps.tso_ipv4)
 102                 *features |= (1ULL << VIRTIO_NET_F_HOST_TSO4);
 103         if (priv->caps.tso_ipv6)
 104                 *features |= (1ULL << VIRTIO_NET_F_HOST_TSO6);
 105         if (priv->caps.tx_csum)
 106                 *features |= (1ULL << VIRTIO_NET_F_CSUM);
 107         if (priv->caps.rx_csum)
 108                 *features |= (1ULL << VIRTIO_NET_F_GUEST_CSUM);
 109         if (priv->caps.virtio_version_1_0)
 110                 *features |= (1ULL << VIRTIO_F_VERSION_1);
 111         return 0;
 112 }
 113
 114 static int
 115 mlx5_vdpa_get_protocol_features(struct rte_vdpa_device *vdev,
 116                 uint64_t *features)
 117 {
 118         struct mlx5_vdpa_priv *priv =
 119                 mlx5_vdpa_find_priv_resource_by_vdev(vdev);
 120
 121         if (priv == NULL) {
 122                 DRV_LOG(ERR, "Invalid vDPA device: %s.", vdev->device->name);
 123                 return -1;
 124         }
 125         *features = MLX5_VDPA_PROTOCOL_FEATURES;
 126         return 0;
 127 }
 128
 129 static int
 130 mlx5_vdpa_set_vring_state(int vid, int vring, int state)
 131 {
 132         struct rte_vdpa_device *vdev = rte_vdpa_get_device(
 133                         rte_vhost_get_vdpa_device_id(vid));
 134         struct mlx5_vdpa_priv *priv =
 135                 mlx5_vdpa_find_priv_resource_by_vdev(vdev);
 136
 137         if (priv == NULL) {
 138                 DRV_LOG(ERR, "Invalid vDPA device: %s.", vdev->device->name);
 139                 return -EINVAL;
 140         }
 141         if (vring >= (int)priv->caps.max_num_virtio_queues * 2) {
 142                 DRV_LOG(ERR, "Too big vring id: %d.", vring);
 143                 return -E2BIG;
 144         }
 145         return mlx5_vdpa_virtq_enable(priv, vring, state);
 146 }
 147
 148 static int
 149 mlx5_vdpa_direct_db_prepare(struct mlx5_vdpa_priv *priv)
 150 {
 151         int ret;
 152
 153         if (priv->direct_notifier) {
 154                 ret = rte_vhost_host_notifier_ctrl(priv->vid, false);
 155                 if (ret != 0) {
 156                         DRV_LOG(INFO, "Direct HW notifier FD cannot be "
 157                                 "destroyed for device %d: %d.", priv->vid, ret);
 158                         return -1;
 159                 }
 160                 priv->direct_notifier = 0;
 161         }
 162         ret = rte_vhost_host_notifier_ctrl(priv->vid, true);
 163         if (ret != 0)
 164                 DRV_LOG(INFO, "Direct HW notifier FD cannot be configured for"
 165                         " device %d: %d.", priv->vid, ret);
 166         else
 167                 priv->direct_notifier = 1;
 168         return 0;
 169 }
 170
 171 static int
 172 mlx5_vdpa_features_set(int vid)
 173 {
 174         struct rte_vdpa_device *vdev = rte_vdpa_get_device(
 175                         rte_vhost_get_vdpa_device_id(vid));
 176         struct mlx5_vdpa_priv *priv =
 177                 mlx5_vdpa_find_priv_resource_by_vdev(vdev);
 178         uint64_t log_base, log_size;
 179         uint64_t features;
 180         int ret;
 181
 182         if (priv == NULL) {
 183                 DRV_LOG(ERR, "Invalid vDPA device: %s.", vdev->device->name);
 184                 return -EINVAL;
 185         }
 186         ret = rte_vhost_get_negotiated_features(vid, &features);
 187         if (ret) {
 188                 DRV_LOG(ERR, "Failed to get negotiated features.");
 189                 return ret;
 190         }
 191         if (RTE_VHOST_NEED_LOG(features)) {
 192                 ret = rte_vhost_get_log_base(vid, &log_base, &log_size);
 193                 if (ret) {
 194                         DRV_LOG(ERR, "Failed to get log base.");
 195                         return ret;
 196                 }
 197                 ret = mlx5_vdpa_dirty_bitmap_set(priv, log_base, log_size);
 198                 if (ret) {
 199                         DRV_LOG(ERR, "Failed to set dirty bitmap.");
 200                         return ret;
 201                 }
 202                 DRV_LOG(INFO, "mlx5 vdpa: enabling dirty logging...");
 203                 ret = mlx5_vdpa_logging_enable(priv, 1);
 204                 if (ret) {
 205                         DRV_LOG(ERR, "Failed t enable dirty logging.");
 206                         return ret;
 207                 }
 208         }
 209         return 0;
 210 }
 211
 212 static int
 213 mlx5_vdpa_pd_create(struct mlx5_vdpa_priv *priv)
 214 {
 215 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
 216         priv->pd = mlx5_glue->alloc_pd(priv->ctx);
 217         if (priv->pd == NULL) {
 218                 DRV_LOG(ERR, "Failed to allocate PD.");
 219                 return errno ? -errno : -ENOMEM;
 220         }
 221         struct mlx5dv_obj obj;
 222         struct mlx5dv_pd pd_info;
 223         int ret = 0;
 224
 225         obj.pd.in = priv->pd;
 226         obj.pd.out = &pd_info;
 227         ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_PD);
 228         if (ret) {
 229                 DRV_LOG(ERR, "Fail to get PD object info.");
 230                 mlx5_glue->dealloc_pd(priv->pd);
 231                 priv->pd = NULL;
 232                 return -errno;
 233         }
 234         priv->pdn = pd_info.pdn;
 235         return 0;
 236 #else
 237         (void)priv;
 238         DRV_LOG(ERR, "Cannot get pdn - no DV support.");
 239         return -ENOTSUP;
 240 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
 241 }
 242
 243 static int
 244 mlx5_vdpa_mtu_set(struct mlx5_vdpa_priv *priv)
 245 {
 246         struct ifreq request;
 247         uint16_t vhost_mtu = 0;
 248         uint16_t kern_mtu = 0;
 249         int ret = rte_vhost_get_mtu(priv->vid, &vhost_mtu);
 250         int sock;
 251         int retries = MLX5_VDPA_MAX_RETRIES;
 252
 253         if (ret) {
 254                 DRV_LOG(DEBUG, "Cannot get vhost MTU - %d.", ret);
 255                 return ret;
 256         }
 257         if (!vhost_mtu) {
 258                 DRV_LOG(DEBUG, "Vhost MTU is 0.");
 259                 return ret;
 260         }
 261         ret = mlx5_get_ifname_sysfs(priv->ctx->device->ibdev_path,
 262                                     request.ifr_name);
 263         if (ret) {
 264                 DRV_LOG(DEBUG, "Cannot get kernel IF name - %d.", ret);
 265                 return ret;
 266         }
 267         sock = socket(PF_INET, SOCK_DGRAM, IPPROTO_IP);
 268         if (sock == -1) {
 269                 DRV_LOG(DEBUG, "Cannot open IF socket.");
 270                 return sock;
 271         }
 272         while (retries--) {
 273                 ret = ioctl(sock, SIOCGIFMTU, &request);
 274                 if (ret == -1)
 275                         break;
 276                 kern_mtu = request.ifr_mtu;
 277                 DRV_LOG(DEBUG, "MTU: current %d requested %d.", (int)kern_mtu,
 278                         (int)vhost_mtu);
 279                 if (kern_mtu == vhost_mtu)
 280                         break;
 281                 request.ifr_mtu = vhost_mtu;
 282                 ret = ioctl(sock, SIOCSIFMTU, &request);
 283                 if (ret == -1)
 284                         break;
 285                 request.ifr_mtu = 0;
 286                 usleep(MLX5_VDPA_USEC);
 287         }
 288         close(sock);
 289         return kern_mtu == vhost_mtu ? 0 : -1;
 290 }
 291
 292 static int
 293 mlx5_vdpa_dev_close(int vid)
 294 {
 295         struct rte_vdpa_device *vdev = rte_vdpa_get_device(
 296                         rte_vhost_get_vdpa_device_id(vid));
 297         struct mlx5_vdpa_priv *priv =
 298                 mlx5_vdpa_find_priv_resource_by_vdev(vdev);
 299         int ret = 0;
 300
 301         if (priv == NULL) {
 302                 DRV_LOG(ERR, "Invalid vDPA device: %s.", vdev->device->name);
 303                 return -1;
 304         }
 305         if (priv->configured)
 306                 ret |= mlx5_vdpa_lm_log(priv);
 307         mlx5_vdpa_cqe_event_unset(priv);
 308         mlx5_vdpa_steer_unset(priv);
 309         mlx5_vdpa_virtqs_release(priv);
 310         mlx5_vdpa_event_qp_global_release(priv);
 311         mlx5_vdpa_mem_dereg(priv);
 312         if (priv->pd) {
 313                 claim_zero(mlx5_glue->dealloc_pd(priv->pd));
 314                 priv->pd = NULL;
 315         }
 316         priv->configured = 0;
 317         priv->vid = 0;
 318         DRV_LOG(INFO, "vDPA device %d was closed.", vid);
 319         return ret;
 320 }
 321
 322 static int
 323 mlx5_vdpa_dev_config(int vid)
 324 {
 325         struct rte_vdpa_device *vdev = rte_vdpa_get_device(
 326                         rte_vhost_get_vdpa_device_id(vid));
 327         struct mlx5_vdpa_priv *priv =
 328                 mlx5_vdpa_find_priv_resource_by_vdev(vdev);
 329
 330         if (priv == NULL) {
 331                 DRV_LOG(ERR, "Invalid vDPA device: %s.", vdev->device->name);
 332                 return -EINVAL;
 333         }
 334         if (priv->configured && mlx5_vdpa_dev_close(vid)) {
 335                 DRV_LOG(ERR, "Failed to reconfigure vid %d.", vid);
 336                 return -1;
 337         }
 338         priv->vid = vid;
 339         if (mlx5_vdpa_mtu_set(priv))
 340                 DRV_LOG(WARNING, "MTU cannot be set on device %s.",
 341                                 vdev->device->name);
 342         if (mlx5_vdpa_pd_create(priv) || mlx5_vdpa_mem_register(priv) ||
 343             mlx5_vdpa_direct_db_prepare(priv) ||
 344             mlx5_vdpa_virtqs_prepare(priv) || mlx5_vdpa_steer_setup(priv) ||
 345             mlx5_vdpa_cqe_event_setup(priv)) {
 346                 mlx5_vdpa_dev_close(vid);
 347                 return -1;
 348         }
 349         priv->configured = 1;
 350         DRV_LOG(INFO, "vDPA device %d was configured.", vid);
 351         return 0;
 352 }
 353
 354 static int
 355 mlx5_vdpa_get_device_fd(int vid)
 356 {
 357         struct rte_vdpa_device *vdev = rte_vdpa_get_device(
 358                         rte_vhost_get_vdpa_device_id(vid));
 359         struct mlx5_vdpa_priv *priv =
 360                 mlx5_vdpa_find_priv_resource_by_vdev(vdev);
 361
 362         if (priv == NULL) {
 363                 DRV_LOG(ERR, "Invalid vDPA device: %s.", vdev->device->name);
 364                 return -EINVAL;
 365         }
 366         return priv->ctx->cmd_fd;
 367 }
 368
 369 static int
 370 mlx5_vdpa_get_notify_area(int vid, int qid, uint64_t *offset, uint64_t *size)
 371 {
 372         struct rte_vdpa_device *vdev = rte_vdpa_get_device(
 373                         rte_vhost_get_vdpa_device_id(vid));
 374         struct mlx5_vdpa_priv *priv =
 375                 mlx5_vdpa_find_priv_resource_by_vdev(vdev);
 376
 377         RTE_SET_USED(qid);
 378         if (priv == NULL) {
 379                 DRV_LOG(ERR, "Invalid vDPA device: %s.", vdev->device->name);
 380                 return -EINVAL;
 381         }
 382         if (!priv->var) {
 383                 DRV_LOG(ERR, "VAR was not created for device %s, is the device"
 384                         " configured?.", vdev->device->name);
 385                 return -EINVAL;
 386         }
 387         *offset = priv->var->mmap_off;
 388         *size = priv->var->length;
 389         return 0;
 390 }
 391
 392 static int
 393 mlx5_vdpa_get_stats_names(struct rte_vdpa_device *vdev,
 394                 struct rte_vdpa_stat_name *stats_names,
 395                 unsigned int size)
 396 {
 397         static const char *mlx5_vdpa_stats_names[MLX5_VDPA_STATS_MAX] = {
 398                 "received_descriptors",
 399                 "completed_descriptors",
 400                 "bad descriptor errors",
 401                 "exceed max chain",
 402                 "invalid buffer",
 403                 "completion errors",
 404         };
 405         struct mlx5_vdpa_priv *priv =
 406                 mlx5_vdpa_find_priv_resource_by_vdev(vdev);
 407         unsigned int i;
 408
 409         if (priv == NULL) {
 410                 DRV_LOG(ERR, "Invalid device: %s.", vdev->device->name);
 411                 return -ENODEV;
 412         }
 413         if (!stats_names)
 414                 return MLX5_VDPA_STATS_MAX;
 415         size = RTE_MIN(size, (unsigned int)MLX5_VDPA_STATS_MAX);
 416         for (i = 0; i < size; ++i)
 417                 strlcpy(stats_names[i].name, mlx5_vdpa_stats_names[i],
 418                         RTE_VDPA_STATS_NAME_SIZE);
 419         return size;
 420 }
 421
 422 static int
 423 mlx5_vdpa_get_stats(struct rte_vdpa_device *vdev, int qid,
 424                 struct rte_vdpa_stat *stats, unsigned int n)
 425 {
 426         struct mlx5_vdpa_priv *priv =
 427                 mlx5_vdpa_find_priv_resource_by_vdev(vdev);
 428
 429         if (priv == NULL) {
 430                 DRV_LOG(ERR, "Invalid device: %s.", vdev->device->name);
 431                 return -ENODEV;
 432         }
 433         if (!priv->configured) {
 434                 DRV_LOG(ERR, "Device %s was not configured.",
 435                                 vdev->device->name);
 436                 return -ENODATA;
 437         }
 438         if (qid >= (int)priv->nr_virtqs) {
 439                 DRV_LOG(ERR, "Too big vring id: %d for device %s.", qid,
 440                                 vdev->device->name);
 441                 return -E2BIG;
 442         }
 443         if (!priv->caps.queue_counters_valid) {
 444                 DRV_LOG(ERR, "Virtq statistics is not supported for device %s.",
 445                         vdev->device->name);
 446                 return -ENOTSUP;
 447         }
 448         return mlx5_vdpa_virtq_stats_get(priv, qid, stats, n);
 449 }
 450
 451 static int
 452 mlx5_vdpa_reset_stats(struct rte_vdpa_device *vdev, int qid)
 453 {
 454         struct mlx5_vdpa_priv *priv =
 455                 mlx5_vdpa_find_priv_resource_by_vdev(vdev);
 456
 457         if (priv == NULL) {
 458                 DRV_LOG(ERR, "Invalid device: %s.", vdev->device->name);
 459                 return -ENODEV;
 460         }
 461         if (!priv->configured) {
 462                 DRV_LOG(ERR, "Device %s was not configured.",
 463                                 vdev->device->name);
 464                 return -ENODATA;
 465         }
 466         if (qid >= (int)priv->nr_virtqs) {
 467                 DRV_LOG(ERR, "Too big vring id: %d for device %s.", qid,
 468                                 vdev->device->name);
 469                 return -E2BIG;
 470         }
 471         if (!priv->caps.queue_counters_valid) {
 472                 DRV_LOG(ERR, "Virtq statistics is not supported for device %s.",
 473                         vdev->device->name);
 474                 return -ENOTSUP;
 475         }
 476         return mlx5_vdpa_virtq_stats_reset(priv, qid);
 477 }
 478
 479 static struct rte_vdpa_dev_ops mlx5_vdpa_ops = {
 480         .get_queue_num = mlx5_vdpa_get_queue_num,
 481         .get_features = mlx5_vdpa_get_vdpa_features,
 482         .get_protocol_features = mlx5_vdpa_get_protocol_features,
 483         .dev_conf = mlx5_vdpa_dev_config,
 484         .dev_close = mlx5_vdpa_dev_close,
 485         .set_vring_state = mlx5_vdpa_set_vring_state,
 486         .set_features = mlx5_vdpa_features_set,
 487         .migration_done = NULL,
 488         .get_vfio_group_fd = NULL,
 489         .get_vfio_device_fd = mlx5_vdpa_get_device_fd,
 490         .get_notify_area = mlx5_vdpa_get_notify_area,
 491         .get_stats_names = mlx5_vdpa_get_stats_names,
 492         .get_stats = mlx5_vdpa_get_stats,
 493         .reset_stats = mlx5_vdpa_reset_stats,
 494 };
 495
 496 static struct ibv_device *
 497 mlx5_vdpa_get_ib_device_match(struct rte_pci_addr *addr)
 498 {
 499         int n;
 500         struct ibv_device **ibv_list = mlx5_glue->get_device_list(&n);
 501         struct ibv_device *ibv_match = NULL;
 502
 503         if (!ibv_list) {
 504                 rte_errno = ENOSYS;
 505                 return NULL;
 506         }
 507         while (n-- > 0) {
 508                 struct rte_pci_addr pci_addr;
 509
 510                 DRV_LOG(DEBUG, "Checking device \"%s\"..", ibv_list[n]->name);
 511                 if (mlx5_dev_to_pci_addr(ibv_list[n]->ibdev_path, &pci_addr))
 512                         continue;
 513                 if (rte_pci_addr_cmp(addr, &pci_addr))
 514                         continue;
 515                 ibv_match = ibv_list[n];
 516                 break;
 517         }
 518         if (!ibv_match)
 519                 rte_errno = ENOENT;
 520         mlx5_glue->free_device_list(ibv_list);
 521         return ibv_match;
 522 }
 523
 524 /* Try to disable ROCE by Netlink\Devlink. */
 525 static int
 526 mlx5_vdpa_nl_roce_disable(const char *addr)
 527 {
 528         int nlsk_fd = mlx5_nl_init(NETLINK_GENERIC);
 529         int devlink_id;
 530         int enable;
 531         int ret;
 532
 533         if (nlsk_fd < 0)
 534                 return nlsk_fd;
 535         devlink_id = mlx5_nl_devlink_family_id_get(nlsk_fd);
 536         if (devlink_id < 0) {
 537                 ret = devlink_id;
 538                 DRV_LOG(DEBUG, "Failed to get devlink id for ROCE operations by"
 539                         " Netlink.");
 540                 goto close;
 541         }
 542         ret = mlx5_nl_enable_roce_get(nlsk_fd, devlink_id, addr, &enable);
 543         if (ret) {
 544                 DRV_LOG(DEBUG, "Failed to get ROCE enable by Netlink: %d.",
 545                         ret);
 546                 goto close;
 547         } else if (!enable) {
 548                 DRV_LOG(INFO, "ROCE has already disabled(Netlink).");
 549                 goto close;
 550         }
 551         ret = mlx5_nl_enable_roce_set(nlsk_fd, devlink_id, addr, 0);
 552         if (ret)
 553                 DRV_LOG(DEBUG, "Failed to disable ROCE by Netlink: %d.", ret);
 554         else
 555                 DRV_LOG(INFO, "ROCE is disabled by Netlink successfully.");
 556 close:
 557         close(nlsk_fd);
 558         return ret;
 559 }
 560
 561 /* Try to disable ROCE by sysfs. */
 562 static int
 563 mlx5_vdpa_sys_roce_disable(const char *addr)
 564 {
 565         FILE *file_o;
 566         int enable;
 567         int ret;
 568
 569         MKSTR(file_p, "/sys/bus/pci/devices/%s/roce_enable", addr);
 570         file_o = fopen(file_p, "rb");
 571         if (!file_o) {
 572                 rte_errno = ENOTSUP;
 573                 return -ENOTSUP;
 574         }
 575         ret = fscanf(file_o, "%d", &enable);
 576         if (ret != 1) {
 577                 rte_errno = EINVAL;
 578                 ret = EINVAL;
 579                 goto close;
 580         } else if (!enable) {
 581                 ret = 0;
 582                 DRV_LOG(INFO, "ROCE has already disabled(sysfs).");
 583                 goto close;
 584         }
 585         fclose(file_o);
 586         file_o = fopen(file_p, "wb");
 587         if (!file_o) {
 588                 rte_errno = ENOTSUP;
 589                 return -ENOTSUP;
 590         }
 591         fprintf(file_o, "0\n");
 592         ret = 0;
 593 close:
 594         if (ret)
 595                 DRV_LOG(DEBUG, "Failed to disable ROCE by sysfs: %d.", ret);
 596         else
 597                 DRV_LOG(INFO, "ROCE is disabled by sysfs successfully.");
 598         fclose(file_o);
 599         return ret;
 600 }
 601
 602 static int
 603 mlx5_vdpa_roce_disable(struct rte_pci_addr *addr, struct ibv_device **ibv)
 604 {
 605         char addr_name[64] = {0};
 606
 607         rte_pci_device_name(addr, addr_name, sizeof(addr_name));
 608         /* Firstly try to disable ROCE by Netlink and fallback to sysfs. */
 609         if (mlx5_vdpa_nl_roce_disable(addr_name) == 0 ||
 610             mlx5_vdpa_sys_roce_disable(addr_name) == 0) {
 611                 /*
 612                  * Succeed to disable ROCE, wait for the IB device to appear
 613                  * again after reload.
 614                  */
 615                 int r;
 616                 struct ibv_device *ibv_new;
 617
 618                 for (r = MLX5_VDPA_MAX_RETRIES; r; r--) {
 619                         ibv_new = mlx5_vdpa_get_ib_device_match(addr);
 620                         if (ibv_new) {
 621                                 *ibv = ibv_new;
 622                                 return 0;
 623                         }
 624                         usleep(MLX5_VDPA_USEC);
 625                 }
 626                 DRV_LOG(ERR, "Cannot much device %s after ROCE disable, "
 627                         "retries exceed %d", addr_name, MLX5_VDPA_MAX_RETRIES);
 628                 rte_errno = EAGAIN;
 629         }
 630         return -rte_errno;
 631 }
 632
 633 /**
 634  * DPDK callback to register a PCI device.
 635  *
 636  * This function spawns vdpa device out of a given PCI device.
 637  *
 638  * @param[in] pci_drv
 639  *   PCI driver structure (mlx5_vpda_driver).
 640  * @param[in] pci_dev
 641  *   PCI device information.
 642  *
 643  * @return
 644  *   0 on success, 1 to skip this driver, a negative errno value otherwise
 645  *   and rte_errno is set.
 646  */
 647 static int
 648 mlx5_vdpa_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
 649                     struct rte_pci_device *pci_dev __rte_unused)
 650 {
 651         struct ibv_device *ibv;
 652         struct mlx5_vdpa_priv *priv = NULL;
 653         struct ibv_context *ctx = NULL;
 654         struct mlx5_hca_attr attr;
 655         int ret;
 656
 657         if (mlx5_class_get(pci_dev->device.devargs) != MLX5_CLASS_VDPA) {
 658                 DRV_LOG(DEBUG, "Skip probing - should be probed by other mlx5"
 659                         " driver.");
 660                 return 1;
 661         }
 662         ibv = mlx5_vdpa_get_ib_device_match(&pci_dev->addr);
 663         if (!ibv) {
 664                 DRV_LOG(ERR, "No matching IB device for PCI slot "
 665                         PCI_PRI_FMT ".", pci_dev->addr.domain,
 666                         pci_dev->addr.bus, pci_dev->addr.devid,
 667                         pci_dev->addr.function);
 668                 return -rte_errno;
 669         } else {
 670                 DRV_LOG(INFO, "PCI information matches for device \"%s\".",
 671                         ibv->name);
 672         }
 673         if (mlx5_vdpa_roce_disable(&pci_dev->addr, &ibv) != 0) {
 674                 DRV_LOG(WARNING, "Failed to disable ROCE for \"%s\".",
 675                         ibv->name);
 676                 return -rte_errno;
 677         }
 678         ctx = mlx5_glue->dv_open_device(ibv);
 679         if (!ctx) {
 680                 DRV_LOG(ERR, "Failed to open IB device \"%s\".", ibv->name);
 681                 rte_errno = ENODEV;
 682                 return -rte_errno;
 683         }
 684         ret = mlx5_devx_cmd_query_hca_attr(ctx, &attr);
 685         if (ret) {
 686                 DRV_LOG(ERR, "Unable to read HCA capabilities.");
 687                 rte_errno = ENOTSUP;
 688                 goto error;
 689         } else if (!attr.vdpa.valid || !attr.vdpa.max_num_virtio_queues) {
 690                 DRV_LOG(ERR, "Not enough capabilities to support vdpa, maybe "
 691                         "old FW/OFED version?");
 692                 rte_errno = ENOTSUP;
 693                 goto error;
 694         }
 695         if (!attr.vdpa.queue_counters_valid)
 696                 DRV_LOG(DEBUG, "No capability to support virtq statistics.");
 697         priv = rte_zmalloc("mlx5 vDPA device private", sizeof(*priv) +
 698                            sizeof(struct mlx5_vdpa_virtq) *
 699                            attr.vdpa.max_num_virtio_queues * 2,
 700                            RTE_CACHE_LINE_SIZE);
 701         if (!priv) {
 702                 DRV_LOG(ERR, "Failed to allocate private memory.");
 703                 rte_errno = ENOMEM;
 704                 goto error;
 705         }
 706         priv->caps = attr.vdpa;
 707         priv->log_max_rqt_size = attr.log_max_rqt_size;
 708         priv->ctx = ctx;
 709         priv->pci_dev = pci_dev;
 710         priv->var = mlx5_glue->dv_alloc_var(ctx, 0);
 711         if (!priv->var) {
 712                 DRV_LOG(ERR, "Failed to allocate VAR %u.\n", errno);
 713                 goto error;
 714         }
 715         priv->vdev = rte_vdpa_register_device(&pci_dev->device,
 716                         &mlx5_vdpa_ops);
 717         if (priv->vdev == NULL) {
 718                 DRV_LOG(ERR, "Failed to register vDPA device.");
 719                 rte_errno = rte_errno ? rte_errno : EINVAL;
 720                 goto error;
 721         }
 722         SLIST_INIT(&priv->mr_list);
 723         pthread_mutex_lock(&priv_list_lock);
 724         TAILQ_INSERT_TAIL(&priv_list, priv, next);
 725         pthread_mutex_unlock(&priv_list_lock);
 726         return 0;
 727
 728 error:
 729         if (priv) {
 730                 if (priv->var)
 731                         mlx5_glue->dv_free_var(priv->var);
 732                 rte_free(priv);
 733         }
 734         if (ctx)
 735                 mlx5_glue->close_device(ctx);
 736         return -rte_errno;
 737 }
 738
 739 /**
 740  * DPDK callback to remove a PCI device.
 741  *
 742  * This function removes all vDPA devices belong to a given PCI device.
 743  *
 744  * @param[in] pci_dev
 745  *   Pointer to the PCI device.
 746  *
 747  * @return
 748  *   0 on success, the function cannot fail.
 749  */
 750 static int
 751 mlx5_vdpa_pci_remove(struct rte_pci_device *pci_dev)
 752 {
 753         struct mlx5_vdpa_priv *priv = NULL;
 754         int found = 0;
 755
 756         pthread_mutex_lock(&priv_list_lock);
 757         TAILQ_FOREACH(priv, &priv_list, next) {
 758                 if (!rte_pci_addr_cmp(&priv->pci_dev->addr, &pci_dev->addr)) {
 759                         found = 1;
 760                         break;
 761                 }
 762         }
 763         if (found)
 764                 TAILQ_REMOVE(&priv_list, priv, next);
 765         pthread_mutex_unlock(&priv_list_lock);
 766         if (found) {
 767                 if (priv->configured)
 768                         mlx5_vdpa_dev_close(priv->vid);
 769                 if (priv->var) {
 770                         mlx5_glue->dv_free_var(priv->var);
 771                         priv->var = NULL;
 772                 }
 773                 mlx5_glue->close_device(priv->ctx);
 774                 rte_free(priv);
 775         }
 776         return 0;
 777 }
 778
 779 static const struct rte_pci_id mlx5_vdpa_pci_id_map[] = {
 780         {
 781                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
 782                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6)
 783         },
 784         {
 785                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
 786                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6VF)
 787         },
 788         {
 789                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
 790                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6DX)
 791         },
 792         {
 793                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
 794                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6DXVF)
 795         },
 796         {
 797                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
 798                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6DXBF)
 799         },
 800         {
 801                 .vendor_id = 0
 802         }
 803 };
 804
 805 static struct rte_pci_driver mlx5_vdpa_driver = {
 806         .driver = {
 807                 .name = "mlx5_vdpa",
 808         },
 809         .id_table = mlx5_vdpa_pci_id_map,
 810         .probe = mlx5_vdpa_pci_probe,
 811         .remove = mlx5_vdpa_pci_remove,
 812         .drv_flags = 0,
 813 };
 814
 815 /**
 816  * Driver initialization routine.
 817  */
 818 RTE_INIT(rte_mlx5_vdpa_init)
 819 {
 820         /* Initialize common log type. */
 821         mlx5_vdpa_logtype = rte_log_register("pmd.vdpa.mlx5");
 822         if (mlx5_vdpa_logtype >= 0)
 823                 rte_log_set_level(mlx5_vdpa_logtype, RTE_LOG_NOTICE);
 824         if (mlx5_glue)
 825                 rte_pci_register(&mlx5_vdpa_driver);
 826 }
 827
 828 RTE_PMD_EXPORT_NAME(net_mlx5_vdpa, __COUNTER__);
 829 RTE_PMD_REGISTER_PCI_TABLE(net_mlx5_vdpa, mlx5_vdpa_pci_id_map);
 830 RTE_PMD_REGISTER_KMOD_DEP(net_mlx5_vdpa, "* ib_uverbs & mlx5_core & mlx5_ib");