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