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 #define MLX5_VDPA_DEFAULT_NO_TRAFFIC_TIME_S 2LLU
  47
  48 TAILQ_HEAD(mlx5_vdpa_privs, mlx5_vdpa_priv) priv_list =
  49                                               TAILQ_HEAD_INITIALIZER(priv_list);
  50 static pthread_mutex_t priv_list_lock = PTHREAD_MUTEX_INITIALIZER;
  51 int mlx5_vdpa_logtype;
  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
 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_vhost_get_vdpa_device(vid);
 175         struct mlx5_vdpa_priv *priv =
 176                 mlx5_vdpa_find_priv_resource_by_vdev(vdev);
 177         uint64_t log_base, log_size;
 178         uint64_t features;
 179         int ret;
 180
 181         if (priv == NULL) {
 182                 DRV_LOG(ERR, "Invalid vDPA device: %s.", vdev->device->name);
 183                 return -EINVAL;
 184         }
 185         ret = rte_vhost_get_negotiated_features(vid, &features);
 186         if (ret) {
 187                 DRV_LOG(ERR, "Failed to get negotiated features.");
 188                 return ret;
 189         }
 190         if (RTE_VHOST_NEED_LOG(features)) {
 191                 ret = rte_vhost_get_log_base(vid, &log_base, &log_size);
 192                 if (ret) {
 193                         DRV_LOG(ERR, "Failed to get log base.");
 194                         return ret;
 195                 }
 196                 ret = mlx5_vdpa_dirty_bitmap_set(priv, log_base, log_size);
 197                 if (ret) {
 198                         DRV_LOG(ERR, "Failed to set dirty bitmap.");
 199                         return ret;
 200                 }
 201                 DRV_LOG(INFO, "mlx5 vdpa: enabling dirty logging...");
 202                 ret = mlx5_vdpa_logging_enable(priv, 1);
 203                 if (ret) {
 204                         DRV_LOG(ERR, "Failed t enable dirty logging.");
 205                         return ret;
 206                 }
 207         }
 208         return 0;
 209 }
 210
 211 static int
 212 mlx5_vdpa_pd_create(struct mlx5_vdpa_priv *priv)
 213 {
 214 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
 215         priv->pd = mlx5_glue->alloc_pd(priv->ctx);
 216         if (priv->pd == NULL) {
 217                 DRV_LOG(ERR, "Failed to allocate PD.");
 218                 return errno ? -errno : -ENOMEM;
 219         }
 220         struct mlx5dv_obj obj;
 221         struct mlx5dv_pd pd_info;
 222         int ret = 0;
 223
 224         obj.pd.in = priv->pd;
 225         obj.pd.out = &pd_info;
 226         ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_PD);
 227         if (ret) {
 228                 DRV_LOG(ERR, "Fail to get PD object info.");
 229                 mlx5_glue->dealloc_pd(priv->pd);
 230                 priv->pd = NULL;
 231                 return -errno;
 232         }
 233         priv->pdn = pd_info.pdn;
 234         return 0;
 235 #else
 236         (void)priv;
 237         DRV_LOG(ERR, "Cannot get pdn - no DV support.");
 238         return -ENOTSUP;
 239 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
 240 }
 241
 242 static int
 243 mlx5_vdpa_mtu_set(struct mlx5_vdpa_priv *priv)
 244 {
 245         struct ifreq request;
 246         uint16_t vhost_mtu = 0;
 247         uint16_t kern_mtu = 0;
 248         int ret = rte_vhost_get_mtu(priv->vid, &vhost_mtu);
 249         int sock;
 250         int retries = MLX5_VDPA_MAX_RETRIES;
 251
 252         if (ret) {
 253                 DRV_LOG(DEBUG, "Cannot get vhost MTU - %d.", ret);
 254                 return ret;
 255         }
 256         if (!vhost_mtu) {
 257                 DRV_LOG(DEBUG, "Vhost MTU is 0.");
 258                 return ret;
 259         }
 260         ret = mlx5_get_ifname_sysfs(priv->ctx->device->ibdev_path,
 261                                     request.ifr_name);
 262         if (ret) {
 263                 DRV_LOG(DEBUG, "Cannot get kernel IF name - %d.", ret);
 264                 return ret;
 265         }
 266         sock = socket(PF_INET, SOCK_DGRAM, IPPROTO_IP);
 267         if (sock == -1) {
 268                 DRV_LOG(DEBUG, "Cannot open IF socket.");
 269                 return sock;
 270         }
 271         while (retries--) {
 272                 ret = ioctl(sock, SIOCGIFMTU, &request);
 273                 if (ret == -1)
 274                         break;
 275                 kern_mtu = request.ifr_mtu;
 276                 DRV_LOG(DEBUG, "MTU: current %d requested %d.", (int)kern_mtu,
 277                         (int)vhost_mtu);
 278                 if (kern_mtu == vhost_mtu)
 279                         break;
 280                 request.ifr_mtu = vhost_mtu;
 281                 ret = ioctl(sock, SIOCSIFMTU, &request);
 282                 if (ret == -1)
 283                         break;
 284                 request.ifr_mtu = 0;
 285                 usleep(MLX5_VDPA_USEC);
 286         }
 287         close(sock);
 288         return kern_mtu == vhost_mtu ? 0 : -1;
 289 }
 290
 291 static int
 292 mlx5_vdpa_dev_close(int vid)
 293 {
 294         struct rte_vdpa_device *vdev = rte_vhost_get_vdpa_device(vid);
 295         struct mlx5_vdpa_priv *priv =
 296                 mlx5_vdpa_find_priv_resource_by_vdev(vdev);
 297         int ret = 0;
 298
 299         if (priv == NULL) {
 300                 DRV_LOG(ERR, "Invalid vDPA device: %s.", vdev->device->name);
 301                 return -1;
 302         }
 303         if (priv->configured)
 304                 ret |= mlx5_vdpa_lm_log(priv);
 305         mlx5_vdpa_cqe_event_unset(priv);
 306         mlx5_vdpa_steer_unset(priv);
 307         mlx5_vdpa_virtqs_release(priv);
 308         mlx5_vdpa_event_qp_global_release(priv);
 309         mlx5_vdpa_mem_dereg(priv);
 310         if (priv->pd) {
 311                 claim_zero(mlx5_glue->dealloc_pd(priv->pd));
 312                 priv->pd = NULL;
 313         }
 314         priv->configured = 0;
 315         priv->vid = 0;
 316         DRV_LOG(INFO, "vDPA device %d was closed.", vid);
 317         return ret;
 318 }
 319
 320 static int
 321 mlx5_vdpa_dev_config(int vid)
 322 {
 323         struct rte_vdpa_device *vdev = rte_vhost_get_vdpa_device(vid);
 324         struct mlx5_vdpa_priv *priv =
 325                 mlx5_vdpa_find_priv_resource_by_vdev(vdev);
 326
 327         if (priv == NULL) {
 328                 DRV_LOG(ERR, "Invalid vDPA device: %s.", vdev->device->name);
 329                 return -EINVAL;
 330         }
 331         if (priv->configured && mlx5_vdpa_dev_close(vid)) {
 332                 DRV_LOG(ERR, "Failed to reconfigure vid %d.", vid);
 333                 return -1;
 334         }
 335         priv->vid = vid;
 336         if (mlx5_vdpa_mtu_set(priv))
 337                 DRV_LOG(WARNING, "MTU cannot be set on device %s.",
 338                                 vdev->device->name);
 339         if (mlx5_vdpa_pd_create(priv) || mlx5_vdpa_mem_register(priv) ||
 340             mlx5_vdpa_direct_db_prepare(priv) ||
 341             mlx5_vdpa_virtqs_prepare(priv) || mlx5_vdpa_steer_setup(priv) ||
 342             mlx5_vdpa_cqe_event_setup(priv)) {
 343                 mlx5_vdpa_dev_close(vid);
 344                 return -1;
 345         }
 346         priv->configured = 1;
 347         DRV_LOG(INFO, "vDPA device %d was configured.", vid);
 348         return 0;
 349 }
 350
 351 static int
 352 mlx5_vdpa_get_device_fd(int vid)
 353 {
 354         struct rte_vdpa_device *vdev = rte_vhost_get_vdpa_device(vid);
 355         struct mlx5_vdpa_priv *priv =
 356                 mlx5_vdpa_find_priv_resource_by_vdev(vdev);
 357
 358         if (priv == NULL) {
 359                 DRV_LOG(ERR, "Invalid vDPA device: %s.", vdev->device->name);
 360                 return -EINVAL;
 361         }
 362         return priv->ctx->cmd_fd;
 363 }
 364
 365 static int
 366 mlx5_vdpa_get_notify_area(int vid, int qid, uint64_t *offset, uint64_t *size)
 367 {
 368         struct rte_vdpa_device *vdev = rte_vhost_get_vdpa_device(vid);
 369         struct mlx5_vdpa_priv *priv =
 370                 mlx5_vdpa_find_priv_resource_by_vdev(vdev);
 371
 372         RTE_SET_USED(qid);
 373         if (priv == NULL) {
 374                 DRV_LOG(ERR, "Invalid vDPA device: %s.", vdev->device->name);
 375                 return -EINVAL;
 376         }
 377         if (!priv->var) {
 378                 DRV_LOG(ERR, "VAR was not created for device %s, is the device"
 379                         " configured?.", vdev->device->name);
 380                 return -EINVAL;
 381         }
 382         *offset = priv->var->mmap_off;
 383         *size = priv->var->length;
 384         return 0;
 385 }
 386
 387 static int
 388 mlx5_vdpa_get_stats_names(struct rte_vdpa_device *vdev,
 389                 struct rte_vdpa_stat_name *stats_names,
 390                 unsigned int size)
 391 {
 392         static const char *mlx5_vdpa_stats_names[MLX5_VDPA_STATS_MAX] = {
 393                 "received_descriptors",
 394                 "completed_descriptors",
 395                 "bad descriptor errors",
 396                 "exceed max chain",
 397                 "invalid buffer",
 398                 "completion errors",
 399         };
 400         struct mlx5_vdpa_priv *priv =
 401                 mlx5_vdpa_find_priv_resource_by_vdev(vdev);
 402         unsigned int i;
 403
 404         if (priv == NULL) {
 405                 DRV_LOG(ERR, "Invalid device: %s.", vdev->device->name);
 406                 return -ENODEV;
 407         }
 408         if (!stats_names)
 409                 return MLX5_VDPA_STATS_MAX;
 410         size = RTE_MIN(size, (unsigned int)MLX5_VDPA_STATS_MAX);
 411         for (i = 0; i < size; ++i)
 412                 strlcpy(stats_names[i].name, mlx5_vdpa_stats_names[i],
 413                         RTE_VDPA_STATS_NAME_SIZE);
 414         return size;
 415 }
 416
 417 static int
 418 mlx5_vdpa_get_stats(struct rte_vdpa_device *vdev, int qid,
 419                 struct rte_vdpa_stat *stats, unsigned int n)
 420 {
 421         struct mlx5_vdpa_priv *priv =
 422                 mlx5_vdpa_find_priv_resource_by_vdev(vdev);
 423
 424         if (priv == NULL) {
 425                 DRV_LOG(ERR, "Invalid device: %s.", vdev->device->name);
 426                 return -ENODEV;
 427         }
 428         if (!priv->configured) {
 429                 DRV_LOG(ERR, "Device %s was not configured.",
 430                                 vdev->device->name);
 431                 return -ENODATA;
 432         }
 433         if (qid >= (int)priv->nr_virtqs) {
 434                 DRV_LOG(ERR, "Too big vring id: %d for device %s.", qid,
 435                                 vdev->device->name);
 436                 return -E2BIG;
 437         }
 438         if (!priv->caps.queue_counters_valid) {
 439                 DRV_LOG(ERR, "Virtq statistics is not supported for device %s.",
 440                         vdev->device->name);
 441                 return -ENOTSUP;
 442         }
 443         return mlx5_vdpa_virtq_stats_get(priv, qid, stats, n);
 444 }
 445
 446 static int
 447 mlx5_vdpa_reset_stats(struct rte_vdpa_device *vdev, int qid)
 448 {
 449         struct mlx5_vdpa_priv *priv =
 450                 mlx5_vdpa_find_priv_resource_by_vdev(vdev);
 451
 452         if (priv == NULL) {
 453                 DRV_LOG(ERR, "Invalid device: %s.", vdev->device->name);
 454                 return -ENODEV;
 455         }
 456         if (!priv->configured) {
 457                 DRV_LOG(ERR, "Device %s was not configured.",
 458                                 vdev->device->name);
 459                 return -ENODATA;
 460         }
 461         if (qid >= (int)priv->nr_virtqs) {
 462                 DRV_LOG(ERR, "Too big vring id: %d for device %s.", qid,
 463                                 vdev->device->name);
 464                 return -E2BIG;
 465         }
 466         if (!priv->caps.queue_counters_valid) {
 467                 DRV_LOG(ERR, "Virtq statistics is not supported for device %s.",
 468                         vdev->device->name);
 469                 return -ENOTSUP;
 470         }
 471         return mlx5_vdpa_virtq_stats_reset(priv, qid);
 472 }
 473
 474 static struct rte_vdpa_dev_ops mlx5_vdpa_ops = {
 475         .get_queue_num = mlx5_vdpa_get_queue_num,
 476         .get_features = mlx5_vdpa_get_vdpa_features,
 477         .get_protocol_features = mlx5_vdpa_get_protocol_features,
 478         .dev_conf = mlx5_vdpa_dev_config,
 479         .dev_close = mlx5_vdpa_dev_close,
 480         .set_vring_state = mlx5_vdpa_set_vring_state,
 481         .set_features = mlx5_vdpa_features_set,
 482         .migration_done = NULL,
 483         .get_vfio_group_fd = NULL,
 484         .get_vfio_device_fd = mlx5_vdpa_get_device_fd,
 485         .get_notify_area = mlx5_vdpa_get_notify_area,
 486         .get_stats_names = mlx5_vdpa_get_stats_names,
 487         .get_stats = mlx5_vdpa_get_stats,
 488         .reset_stats = mlx5_vdpa_reset_stats,
 489 };
 490
 491 static struct ibv_device *
 492 mlx5_vdpa_get_ib_device_match(struct rte_pci_addr *addr)
 493 {
 494         int n;
 495         struct ibv_device **ibv_list = mlx5_glue->get_device_list(&n);
 496         struct ibv_device *ibv_match = NULL;
 497
 498         if (!ibv_list) {
 499                 rte_errno = ENOSYS;
 500                 return NULL;
 501         }
 502         while (n-- > 0) {
 503                 struct rte_pci_addr pci_addr;
 504
 505                 DRV_LOG(DEBUG, "Checking device \"%s\"..", ibv_list[n]->name);
 506                 if (mlx5_dev_to_pci_addr(ibv_list[n]->ibdev_path, &pci_addr))
 507                         continue;
 508                 if (rte_pci_addr_cmp(addr, &pci_addr))
 509                         continue;
 510                 ibv_match = ibv_list[n];
 511                 break;
 512         }
 513         if (!ibv_match)
 514                 rte_errno = ENOENT;
 515         mlx5_glue->free_device_list(ibv_list);
 516         return ibv_match;
 517 }
 518
 519 /* Try to disable ROCE by Netlink\Devlink. */
 520 static int
 521 mlx5_vdpa_nl_roce_disable(const char *addr)
 522 {
 523         int nlsk_fd = mlx5_nl_init(NETLINK_GENERIC);
 524         int devlink_id;
 525         int enable;
 526         int ret;
 527
 528         if (nlsk_fd < 0)
 529                 return nlsk_fd;
 530         devlink_id = mlx5_nl_devlink_family_id_get(nlsk_fd);
 531         if (devlink_id < 0) {
 532                 ret = devlink_id;
 533                 DRV_LOG(DEBUG, "Failed to get devlink id for ROCE operations by"
 534                         " Netlink.");
 535                 goto close;
 536         }
 537         ret = mlx5_nl_enable_roce_get(nlsk_fd, devlink_id, addr, &enable);
 538         if (ret) {
 539                 DRV_LOG(DEBUG, "Failed to get ROCE enable by Netlink: %d.",
 540                         ret);
 541                 goto close;
 542         } else if (!enable) {
 543                 DRV_LOG(INFO, "ROCE has already disabled(Netlink).");
 544                 goto close;
 545         }
 546         ret = mlx5_nl_enable_roce_set(nlsk_fd, devlink_id, addr, 0);
 547         if (ret)
 548                 DRV_LOG(DEBUG, "Failed to disable ROCE by Netlink: %d.", ret);
 549         else
 550                 DRV_LOG(INFO, "ROCE is disabled by Netlink successfully.");
 551 close:
 552         close(nlsk_fd);
 553         return ret;
 554 }
 555
 556 /* Try to disable ROCE by sysfs. */
 557 static int
 558 mlx5_vdpa_sys_roce_disable(const char *addr)
 559 {
 560         FILE *file_o;
 561         int enable;
 562         int ret;
 563
 564         MKSTR(file_p, "/sys/bus/pci/devices/%s/roce_enable", addr);
 565         file_o = fopen(file_p, "rb");
 566         if (!file_o) {
 567                 rte_errno = ENOTSUP;
 568                 return -ENOTSUP;
 569         }
 570         ret = fscanf(file_o, "%d", &enable);
 571         if (ret != 1) {
 572                 rte_errno = EINVAL;
 573                 ret = EINVAL;
 574                 goto close;
 575         } else if (!enable) {
 576                 ret = 0;
 577                 DRV_LOG(INFO, "ROCE has already disabled(sysfs).");
 578                 goto close;
 579         }
 580         fclose(file_o);
 581         file_o = fopen(file_p, "wb");
 582         if (!file_o) {
 583                 rte_errno = ENOTSUP;
 584                 return -ENOTSUP;
 585         }
 586         fprintf(file_o, "0\n");
 587         ret = 0;
 588 close:
 589         if (ret)
 590                 DRV_LOG(DEBUG, "Failed to disable ROCE by sysfs: %d.", ret);
 591         else
 592                 DRV_LOG(INFO, "ROCE is disabled by sysfs successfully.");
 593         fclose(file_o);
 594         return ret;
 595 }
 596
 597 static int
 598 mlx5_vdpa_roce_disable(struct rte_pci_addr *addr, struct ibv_device **ibv)
 599 {
 600         char addr_name[64] = {0};
 601
 602         rte_pci_device_name(addr, addr_name, sizeof(addr_name));
 603         /* Firstly try to disable ROCE by Netlink and fallback to sysfs. */
 604         if (mlx5_vdpa_nl_roce_disable(addr_name) == 0 ||
 605             mlx5_vdpa_sys_roce_disable(addr_name) == 0) {
 606                 /*
 607                  * Succeed to disable ROCE, wait for the IB device to appear
 608                  * again after reload.
 609                  */
 610                 int r;
 611                 struct ibv_device *ibv_new;
 612
 613                 for (r = MLX5_VDPA_MAX_RETRIES; r; r--) {
 614                         ibv_new = mlx5_vdpa_get_ib_device_match(addr);
 615                         if (ibv_new) {
 616                                 *ibv = ibv_new;
 617                                 return 0;
 618                         }
 619                         usleep(MLX5_VDPA_USEC);
 620                 }
 621                 DRV_LOG(ERR, "Cannot much device %s after ROCE disable, "
 622                         "retries exceed %d", addr_name, MLX5_VDPA_MAX_RETRIES);
 623                 rte_errno = EAGAIN;
 624         }
 625         return -rte_errno;
 626 }
 627
 628 static int
 629 mlx5_vdpa_args_check_handler(const char *key, const char *val, void *opaque)
 630 {
 631         struct mlx5_vdpa_priv *priv = opaque;
 632         unsigned long tmp;
 633
 634         if (strcmp(key, "class") == 0)
 635                 return 0;
 636         errno = 0;
 637         tmp = strtoul(val, NULL, 0);
 638         if (errno) {
 639                 DRV_LOG(WARNING, "%s: \"%s\" is an invalid integer.", key, val);
 640                 return -errno;
 641         }
 642         if (strcmp(key, "event_mode") == 0) {
 643                 if (tmp <= MLX5_VDPA_EVENT_MODE_ONLY_INTERRUPT)
 644                         priv->event_mode = (int)tmp;
 645                 else
 646                         DRV_LOG(WARNING, "Invalid event_mode %s.", val);
 647         } else if (strcmp(key, "event_us") == 0) {
 648                 priv->event_us = (uint32_t)tmp;
 649         } else if (strcmp(key, "no_traffic_time") == 0) {
 650                 priv->no_traffic_time_s = (uint32_t)tmp;
 651         } else {
 652                 DRV_LOG(WARNING, "Invalid key %s.", key);
 653         }
 654         return 0;
 655 }
 656
 657 static void
 658 mlx5_vdpa_config_get(struct rte_devargs *devargs, struct mlx5_vdpa_priv *priv)
 659 {
 660         struct rte_kvargs *kvlist;
 661
 662         priv->event_mode = MLX5_VDPA_EVENT_MODE_DYNAMIC_TIMER;
 663         priv->event_us = 0;
 664         priv->no_traffic_time_s = MLX5_VDPA_DEFAULT_NO_TRAFFIC_TIME_S;
 665         if (devargs == NULL)
 666                 return;
 667         kvlist = rte_kvargs_parse(devargs->args, NULL);
 668         if (kvlist == NULL)
 669                 return;
 670         rte_kvargs_process(kvlist, NULL, mlx5_vdpa_args_check_handler, priv);
 671         rte_kvargs_free(kvlist);
 672         if (!priv->event_us) {
 673                 if (priv->event_mode == MLX5_VDPA_EVENT_MODE_DYNAMIC_TIMER)
 674                         priv->event_us = MLX5_VDPA_DEFAULT_TIMER_STEP_US;
 675                 else if (priv->event_mode == MLX5_VDPA_EVENT_MODE_FIXED_TIMER)
 676                         priv->event_us = MLX5_VDPA_DEFAULT_TIMER_DELAY_US;
 677         }
 678         DRV_LOG(DEBUG, "event mode is %d.", priv->event_mode);
 679         DRV_LOG(DEBUG, "event_us is %u us.", priv->event_us);
 680         DRV_LOG(DEBUG, "no traffic time is %u s.", priv->no_traffic_time_s);
 681 }
 682
 683 /**
 684  * DPDK callback to register a PCI device.
 685  *
 686  * This function spawns vdpa device out of a given PCI device.
 687  *
 688  * @param[in] pci_drv
 689  *   PCI driver structure (mlx5_vpda_driver).
 690  * @param[in] pci_dev
 691  *   PCI device information.
 692  *
 693  * @return
 694  *   0 on success, 1 to skip this driver, a negative errno value otherwise
 695  *   and rte_errno is set.
 696  */
 697 static int
 698 mlx5_vdpa_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
 699                     struct rte_pci_device *pci_dev __rte_unused)
 700 {
 701         struct ibv_device *ibv;
 702         struct mlx5_vdpa_priv *priv = NULL;
 703         struct ibv_context *ctx = NULL;
 704         struct mlx5_hca_attr attr;
 705         int ret;
 706
 707         if (mlx5_class_get(pci_dev->device.devargs) != MLX5_CLASS_VDPA) {
 708                 DRV_LOG(DEBUG, "Skip probing - should be probed by other mlx5"
 709                         " driver.");
 710                 return 1;
 711         }
 712         ibv = mlx5_vdpa_get_ib_device_match(&pci_dev->addr);
 713         if (!ibv) {
 714                 DRV_LOG(ERR, "No matching IB device for PCI slot "
 715                         PCI_PRI_FMT ".", pci_dev->addr.domain,
 716                         pci_dev->addr.bus, pci_dev->addr.devid,
 717                         pci_dev->addr.function);
 718                 return -rte_errno;
 719         } else {
 720                 DRV_LOG(INFO, "PCI information matches for device \"%s\".",
 721                         ibv->name);
 722         }
 723         if (mlx5_vdpa_roce_disable(&pci_dev->addr, &ibv) != 0) {
 724                 DRV_LOG(WARNING, "Failed to disable ROCE for \"%s\".",
 725                         ibv->name);
 726                 return -rte_errno;
 727         }
 728         ctx = mlx5_glue->dv_open_device(ibv);
 729         if (!ctx) {
 730                 DRV_LOG(ERR, "Failed to open IB device \"%s\".", ibv->name);
 731                 rte_errno = ENODEV;
 732                 return -rte_errno;
 733         }
 734         ret = mlx5_devx_cmd_query_hca_attr(ctx, &attr);
 735         if (ret) {
 736                 DRV_LOG(ERR, "Unable to read HCA capabilities.");
 737                 rte_errno = ENOTSUP;
 738                 goto error;
 739         } else if (!attr.vdpa.valid || !attr.vdpa.max_num_virtio_queues) {
 740                 DRV_LOG(ERR, "Not enough capabilities to support vdpa, maybe "
 741                         "old FW/OFED version?");
 742                 rte_errno = ENOTSUP;
 743                 goto error;
 744         }
 745         if (!attr.vdpa.queue_counters_valid)
 746                 DRV_LOG(DEBUG, "No capability to support virtq statistics.");
 747         priv = rte_zmalloc("mlx5 vDPA device private", sizeof(*priv) +
 748                            sizeof(struct mlx5_vdpa_virtq) *
 749                            attr.vdpa.max_num_virtio_queues * 2,
 750                            RTE_CACHE_LINE_SIZE);
 751         if (!priv) {
 752                 DRV_LOG(ERR, "Failed to allocate private memory.");
 753                 rte_errno = ENOMEM;
 754                 goto error;
 755         }
 756         priv->caps = attr.vdpa;
 757         priv->log_max_rqt_size = attr.log_max_rqt_size;
 758         priv->ctx = ctx;
 759         priv->pci_dev = pci_dev;
 760         priv->var = mlx5_glue->dv_alloc_var(ctx, 0);
 761         if (!priv->var) {
 762                 DRV_LOG(ERR, "Failed to allocate VAR %u.\n", errno);
 763                 goto error;
 764         }
 765         priv->vdev = rte_vdpa_register_device(&pci_dev->device,
 766                         &mlx5_vdpa_ops);
 767         if (priv->vdev == NULL) {
 768                 DRV_LOG(ERR, "Failed to register vDPA device.");
 769                 rte_errno = rte_errno ? rte_errno : EINVAL;
 770                 goto error;
 771         }
 772         mlx5_vdpa_config_get(pci_dev->device.devargs, priv);
 773         SLIST_INIT(&priv->mr_list);
 774         pthread_mutex_lock(&priv_list_lock);
 775         TAILQ_INSERT_TAIL(&priv_list, priv, next);
 776         pthread_mutex_unlock(&priv_list_lock);
 777         return 0;
 778
 779 error:
 780         if (priv) {
 781                 if (priv->var)
 782                         mlx5_glue->dv_free_var(priv->var);
 783                 rte_free(priv);
 784         }
 785         if (ctx)
 786                 mlx5_glue->close_device(ctx);
 787         return -rte_errno;
 788 }
 789
 790 /**
 791  * DPDK callback to remove a PCI device.
 792  *
 793  * This function removes all vDPA devices belong to a given PCI device.
 794  *
 795  * @param[in] pci_dev
 796  *   Pointer to the PCI device.
 797  *
 798  * @return
 799  *   0 on success, the function cannot fail.
 800  */
 801 static int
 802 mlx5_vdpa_pci_remove(struct rte_pci_device *pci_dev)
 803 {
 804         struct mlx5_vdpa_priv *priv = NULL;
 805         int found = 0;
 806
 807         pthread_mutex_lock(&priv_list_lock);
 808         TAILQ_FOREACH(priv, &priv_list, next) {
 809                 if (!rte_pci_addr_cmp(&priv->pci_dev->addr, &pci_dev->addr)) {
 810                         found = 1;
 811                         break;
 812                 }
 813         }
 814         if (found)
 815                 TAILQ_REMOVE(&priv_list, priv, next);
 816         pthread_mutex_unlock(&priv_list_lock);
 817         if (found) {
 818                 if (priv->configured)
 819                         mlx5_vdpa_dev_close(priv->vid);
 820                 if (priv->var) {
 821                         mlx5_glue->dv_free_var(priv->var);
 822                         priv->var = NULL;
 823                 }
 824                 mlx5_glue->close_device(priv->ctx);
 825                 rte_free(priv);
 826         }
 827         return 0;
 828 }
 829
 830 static const struct rte_pci_id mlx5_vdpa_pci_id_map[] = {
 831         {
 832                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
 833                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6)
 834         },
 835         {
 836                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
 837                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6VF)
 838         },
 839         {
 840                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
 841                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6DX)
 842         },
 843         {
 844                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
 845                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6DXVF)
 846         },
 847         {
 848                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
 849                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6DXBF)
 850         },
 851         {
 852                 .vendor_id = 0
 853         }
 854 };
 855
 856 static struct rte_pci_driver mlx5_vdpa_driver = {
 857         .driver = {
 858                 .name = "mlx5_vdpa",
 859         },
 860         .id_table = mlx5_vdpa_pci_id_map,
 861         .probe = mlx5_vdpa_pci_probe,
 862         .remove = mlx5_vdpa_pci_remove,
 863         .drv_flags = 0,
 864 };
 865
 866 /**
 867  * Driver initialization routine.
 868  */
 869 RTE_INIT(rte_mlx5_vdpa_init)
 870 {
 871         /* Initialize common log type. */
 872         mlx5_vdpa_logtype = rte_log_register("pmd.vdpa.mlx5");
 873         if (mlx5_vdpa_logtype >= 0)
 874                 rte_log_set_level(mlx5_vdpa_logtype, RTE_LOG_NOTICE);
 875         if (mlx5_glue)
 876                 rte_pci_register(&mlx5_vdpa_driver);
 877 }
 878
 879 RTE_PMD_EXPORT_NAME(net_mlx5_vdpa, __COUNTER__);
 880 RTE_PMD_REGISTER_PCI_TABLE(net_mlx5_vdpa, mlx5_vdpa_pci_id_map);
 881 RTE_PMD_REGISTER_KMOD_DEP(net_mlx5_vdpa, "* ib_uverbs & mlx5_core & mlx5_ib");