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_did(int did)
  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 (did == priv->id) {
  61                         found = 1;
  62                         break;
  63                 }
  64         }
  65         pthread_mutex_unlock(&priv_list_lock);
  66         if (!found) {
  67                 DRV_LOG(ERR, "Invalid device id: %d.", did);
  68                 rte_errno = EINVAL;
  69                 return NULL;
  70         }
  71         return priv;
  72 }
  73
  74 static int
  75 mlx5_vdpa_get_queue_num(int did, uint32_t *queue_num)
  76 {
  77         struct mlx5_vdpa_priv *priv = mlx5_vdpa_find_priv_resource_by_did(did);
  78
  79         if (priv == NULL) {
  80                 DRV_LOG(ERR, "Invalid device id: %d.", did);
  81                 return -1;
  82         }
  83         *queue_num = priv->caps.max_num_virtio_queues;
  84         return 0;
  85 }
  86
  87 static int
  88 mlx5_vdpa_get_vdpa_features(int did, uint64_t *features)
  89 {
  90         struct mlx5_vdpa_priv *priv = mlx5_vdpa_find_priv_resource_by_did(did);
  91
  92         if (priv == NULL) {
  93                 DRV_LOG(ERR, "Invalid device id: %d.", did);
  94                 return -1;
  95         }
  96         *features = MLX5_VDPA_DEFAULT_FEATURES;
  97         if (priv->caps.virtio_queue_type & (1 << MLX5_VIRTQ_TYPE_PACKED))
  98                 *features |= (1ULL << VIRTIO_F_RING_PACKED);
  99         if (priv->caps.tso_ipv4)
 100                 *features |= (1ULL << VIRTIO_NET_F_HOST_TSO4);
 101         if (priv->caps.tso_ipv6)
 102                 *features |= (1ULL << VIRTIO_NET_F_HOST_TSO6);
 103         if (priv->caps.tx_csum)
 104                 *features |= (1ULL << VIRTIO_NET_F_CSUM);
 105         if (priv->caps.rx_csum)
 106                 *features |= (1ULL << VIRTIO_NET_F_GUEST_CSUM);
 107         if (priv->caps.virtio_version_1_0)
 108                 *features |= (1ULL << VIRTIO_F_VERSION_1);
 109         return 0;
 110 }
 111
 112 static int
 113 mlx5_vdpa_get_protocol_features(int did, uint64_t *features)
 114 {
 115         struct mlx5_vdpa_priv *priv = mlx5_vdpa_find_priv_resource_by_did(did);
 116
 117         if (priv == NULL) {
 118                 DRV_LOG(ERR, "Invalid device id: %d.", did);
 119                 return -1;
 120         }
 121         *features = MLX5_VDPA_PROTOCOL_FEATURES;
 122         return 0;
 123 }
 124
 125 static int
 126 mlx5_vdpa_set_vring_state(int vid, int vring, int state)
 127 {
 128         int did = rte_vhost_get_vdpa_device_id(vid);
 129         struct mlx5_vdpa_priv *priv = mlx5_vdpa_find_priv_resource_by_did(did);
 130
 131         if (priv == NULL) {
 132                 DRV_LOG(ERR, "Invalid device id: %d.", did);
 133                 return -EINVAL;
 134         }
 135         if (vring >= (int)priv->caps.max_num_virtio_queues * 2) {
 136                 DRV_LOG(ERR, "Too big vring id: %d.", vring);
 137                 return -E2BIG;
 138         }
 139         return mlx5_vdpa_virtq_enable(priv, vring, state);
 140 }
 141
 142 static int
 143 mlx5_vdpa_direct_db_prepare(struct mlx5_vdpa_priv *priv)
 144 {
 145         int ret;
 146
 147         if (priv->direct_notifier) {
 148                 ret = rte_vhost_host_notifier_ctrl(priv->vid, false);
 149                 if (ret != 0) {
 150                         DRV_LOG(INFO, "Direct HW notifier FD cannot be "
 151                                 "destroyed for device %d: %d.", priv->vid, ret);
 152                         return -1;
 153                 }
 154                 priv->direct_notifier = 0;
 155         }
 156         ret = rte_vhost_host_notifier_ctrl(priv->vid, true);
 157         if (ret != 0)
 158                 DRV_LOG(INFO, "Direct HW notifier FD cannot be configured for"
 159                         " device %d: %d.", priv->vid, ret);
 160         else
 161                 priv->direct_notifier = 1;
 162         return 0;
 163 }
 164
 165 static int
 166 mlx5_vdpa_features_set(int vid)
 167 {
 168         int did = rte_vhost_get_vdpa_device_id(vid);
 169         struct mlx5_vdpa_priv *priv = mlx5_vdpa_find_priv_resource_by_did(did);
 170         uint64_t log_base, log_size;
 171         uint64_t features;
 172         int ret;
 173
 174         if (priv == NULL) {
 175                 DRV_LOG(ERR, "Invalid device id: %d.", did);
 176                 return -EINVAL;
 177         }
 178         ret = rte_vhost_get_negotiated_features(vid, &features);
 179         if (ret) {
 180                 DRV_LOG(ERR, "Failed to get negotiated features.");
 181                 return ret;
 182         }
 183         if (RTE_VHOST_NEED_LOG(features)) {
 184                 ret = rte_vhost_get_log_base(vid, &log_base, &log_size);
 185                 if (ret) {
 186                         DRV_LOG(ERR, "Failed to get log base.");
 187                         return ret;
 188                 }
 189                 ret = mlx5_vdpa_dirty_bitmap_set(priv, log_base, log_size);
 190                 if (ret) {
 191                         DRV_LOG(ERR, "Failed to set dirty bitmap.");
 192                         return ret;
 193                 }
 194                 DRV_LOG(INFO, "mlx5 vdpa: enabling dirty logging...");
 195                 ret = mlx5_vdpa_logging_enable(priv, 1);
 196                 if (ret) {
 197                         DRV_LOG(ERR, "Failed t enable dirty logging.");
 198                         return ret;
 199                 }
 200         }
 201         return 0;
 202 }
 203
 204 static int
 205 mlx5_vdpa_pd_create(struct mlx5_vdpa_priv *priv)
 206 {
 207 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
 208         priv->pd = mlx5_glue->alloc_pd(priv->ctx);
 209         if (priv->pd == NULL) {
 210                 DRV_LOG(ERR, "Failed to allocate PD.");
 211                 return errno ? -errno : -ENOMEM;
 212         }
 213         struct mlx5dv_obj obj;
 214         struct mlx5dv_pd pd_info;
 215         int ret = 0;
 216
 217         obj.pd.in = priv->pd;
 218         obj.pd.out = &pd_info;
 219         ret = mlx5_glue->dv_init_obj(&obj, MLX5DV_OBJ_PD);
 220         if (ret) {
 221                 DRV_LOG(ERR, "Fail to get PD object info.");
 222                 mlx5_glue->dealloc_pd(priv->pd);
 223                 priv->pd = NULL;
 224                 return -errno;
 225         }
 226         priv->pdn = pd_info.pdn;
 227         return 0;
 228 #else
 229         (void)priv;
 230         DRV_LOG(ERR, "Cannot get pdn - no DV support.");
 231         return -ENOTSUP;
 232 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */
 233 }
 234
 235 static int
 236 mlx5_vdpa_mtu_set(struct mlx5_vdpa_priv *priv)
 237 {
 238         struct ifreq request;
 239         uint16_t vhost_mtu = 0;
 240         uint16_t kern_mtu = 0;
 241         int ret = rte_vhost_get_mtu(priv->vid, &vhost_mtu);
 242         int sock;
 243         int retries = MLX5_VDPA_MAX_RETRIES;
 244
 245         if (ret) {
 246                 DRV_LOG(DEBUG, "Cannot get vhost MTU - %d.", ret);
 247                 return ret;
 248         }
 249         if (!vhost_mtu) {
 250                 DRV_LOG(DEBUG, "Vhost MTU is 0.");
 251                 return ret;
 252         }
 253         ret = mlx5_get_ifname_sysfs(priv->ctx->device->ibdev_path,
 254                                     request.ifr_name);
 255         if (ret) {
 256                 DRV_LOG(DEBUG, "Cannot get kernel IF name - %d.", ret);
 257                 return ret;
 258         }
 259         sock = socket(PF_INET, SOCK_DGRAM, IPPROTO_IP);
 260         if (sock == -1) {
 261                 DRV_LOG(DEBUG, "Cannot open IF socket.");
 262                 return sock;
 263         }
 264         while (retries--) {
 265                 ret = ioctl(sock, SIOCGIFMTU, &request);
 266                 if (ret == -1)
 267                         break;
 268                 kern_mtu = request.ifr_mtu;
 269                 DRV_LOG(DEBUG, "MTU: current %d requested %d.", (int)kern_mtu,
 270                         (int)vhost_mtu);
 271                 if (kern_mtu == vhost_mtu)
 272                         break;
 273                 request.ifr_mtu = vhost_mtu;
 274                 ret = ioctl(sock, SIOCSIFMTU, &request);
 275                 if (ret == -1)
 276                         break;
 277                 request.ifr_mtu = 0;
 278                 usleep(MLX5_VDPA_USEC);
 279         }
 280         close(sock);
 281         return kern_mtu == vhost_mtu ? 0 : -1;
 282 }
 283
 284 static int
 285 mlx5_vdpa_dev_close(int vid)
 286 {
 287         int did = rte_vhost_get_vdpa_device_id(vid);
 288         struct mlx5_vdpa_priv *priv = mlx5_vdpa_find_priv_resource_by_did(did);
 289         int ret = 0;
 290
 291         if (priv == NULL) {
 292                 DRV_LOG(ERR, "Invalid device id: %d.", did);
 293                 return -1;
 294         }
 295         if (priv->configured)
 296                 ret |= mlx5_vdpa_lm_log(priv);
 297         mlx5_vdpa_cqe_event_unset(priv);
 298         mlx5_vdpa_steer_unset(priv);
 299         mlx5_vdpa_virtqs_release(priv);
 300         mlx5_vdpa_event_qp_global_release(priv);
 301         mlx5_vdpa_mem_dereg(priv);
 302         if (priv->pd) {
 303                 claim_zero(mlx5_glue->dealloc_pd(priv->pd));
 304                 priv->pd = NULL;
 305         }
 306         priv->configured = 0;
 307         priv->vid = 0;
 308         DRV_LOG(INFO, "vDPA device %d was closed.", vid);
 309         return ret;
 310 }
 311
 312 static int
 313 mlx5_vdpa_dev_config(int vid)
 314 {
 315         int did = rte_vhost_get_vdpa_device_id(vid);
 316         struct mlx5_vdpa_priv *priv = mlx5_vdpa_find_priv_resource_by_did(did);
 317
 318         if (priv == NULL) {
 319                 DRV_LOG(ERR, "Invalid device id: %d.", did);
 320                 return -EINVAL;
 321         }
 322         if (priv->configured && mlx5_vdpa_dev_close(vid)) {
 323                 DRV_LOG(ERR, "Failed to reconfigure vid %d.", vid);
 324                 return -1;
 325         }
 326         priv->vid = vid;
 327         if (mlx5_vdpa_mtu_set(priv))
 328                 DRV_LOG(WARNING, "MTU cannot be set on device %d.", did);
 329         if (mlx5_vdpa_pd_create(priv) || mlx5_vdpa_mem_register(priv) ||
 330             mlx5_vdpa_direct_db_prepare(priv) ||
 331             mlx5_vdpa_virtqs_prepare(priv) || mlx5_vdpa_steer_setup(priv) ||
 332             mlx5_vdpa_cqe_event_setup(priv)) {
 333                 mlx5_vdpa_dev_close(vid);
 334                 return -1;
 335         }
 336         priv->configured = 1;
 337         DRV_LOG(INFO, "vDPA device %d was configured.", vid);
 338         return 0;
 339 }
 340
 341 static int
 342 mlx5_vdpa_get_device_fd(int vid)
 343 {
 344         int did = rte_vhost_get_vdpa_device_id(vid);
 345         struct mlx5_vdpa_priv *priv = mlx5_vdpa_find_priv_resource_by_did(did);
 346
 347         if (priv == NULL) {
 348                 DRV_LOG(ERR, "Invalid device id: %d.", did);
 349                 return -EINVAL;
 350         }
 351         return priv->ctx->cmd_fd;
 352 }
 353
 354 static int
 355 mlx5_vdpa_get_notify_area(int vid, int qid, uint64_t *offset, uint64_t *size)
 356 {
 357         int did = rte_vhost_get_vdpa_device_id(vid);
 358         struct mlx5_vdpa_priv *priv = mlx5_vdpa_find_priv_resource_by_did(did);
 359
 360         RTE_SET_USED(qid);
 361         if (priv == NULL) {
 362                 DRV_LOG(ERR, "Invalid device id: %d.", did);
 363                 return -EINVAL;
 364         }
 365         if (!priv->var) {
 366                 DRV_LOG(ERR, "VAR was not created for device %d, is the device"
 367                         " configured?.", did);
 368                 return -EINVAL;
 369         }
 370         *offset = priv->var->mmap_off;
 371         *size = priv->var->length;
 372         return 0;
 373 }
 374
 375 static int
 376 mlx5_vdpa_get_stats_names(int did, struct rte_vdpa_stat_name *stats_names,
 377                           unsigned int size)
 378 {
 379         static const char *mlx5_vdpa_stats_names[MLX5_VDPA_STATS_MAX] = {
 380                 "received_descriptors",
 381                 "completed_descriptors",
 382                 "bad descriptor errors",
 383                 "exceed max chain",
 384                 "invalid buffer",
 385                 "completion errors",
 386         };
 387         struct mlx5_vdpa_priv *priv = mlx5_vdpa_find_priv_resource_by_did(did);
 388         unsigned int i;
 389
 390         if (priv == NULL) {
 391                 DRV_LOG(ERR, "Invalid device id: %d.", did);
 392                 return -ENODEV;
 393         }
 394         if (!stats_names)
 395                 return MLX5_VDPA_STATS_MAX;
 396         size = RTE_MIN(size, (unsigned int)MLX5_VDPA_STATS_MAX);
 397         for (i = 0; i < size; ++i)
 398                 strlcpy(stats_names[i].name, mlx5_vdpa_stats_names[i],
 399                         RTE_VDPA_STATS_NAME_SIZE);
 400         return size;
 401 }
 402
 403 static int
 404 mlx5_vdpa_get_stats(int did, int qid, struct rte_vdpa_stat *stats,
 405                     unsigned int n)
 406 {
 407         struct mlx5_vdpa_priv *priv = mlx5_vdpa_find_priv_resource_by_did(did);
 408
 409         if (priv == NULL) {
 410                 DRV_LOG(ERR, "Invalid device id: %d.", did);
 411                 return -ENODEV;
 412         }
 413         if (!priv->configured) {
 414                 DRV_LOG(ERR, "Device %d was not configured.", did);
 415                 return -ENODATA;
 416         }
 417         if (qid >= (int)priv->nr_virtqs) {
 418                 DRV_LOG(ERR, "Too big vring id: %d.", qid);
 419                 return -E2BIG;
 420         }
 421         if (!priv->caps.queue_counters_valid) {
 422                 DRV_LOG(ERR, "Virtq statistics is not supported for device %d.",
 423                         did);
 424                 return -ENOTSUP;
 425         }
 426         return mlx5_vdpa_virtq_stats_get(priv, qid, stats, n);
 427 }
 428
 429 static int
 430 mlx5_vdpa_reset_stats(int did, int qid)
 431 {
 432         struct mlx5_vdpa_priv *priv = mlx5_vdpa_find_priv_resource_by_did(did);
 433
 434         if (priv == NULL) {
 435                 DRV_LOG(ERR, "Invalid device id: %d.", did);
 436                 return -ENODEV;
 437         }
 438         if (!priv->configured) {
 439                 DRV_LOG(ERR, "Device %d was not configured.", did);
 440                 return -ENODATA;
 441         }
 442         if (qid >= (int)priv->nr_virtqs) {
 443                 DRV_LOG(ERR, "Too big vring id: %d.", qid);
 444                 return -E2BIG;
 445         }
 446         if (!priv->caps.queue_counters_valid) {
 447                 DRV_LOG(ERR, "Virtq statistics is not supported for device %d.",
 448                         did);
 449                 return -ENOTSUP;
 450         }
 451         return mlx5_vdpa_virtq_stats_reset(priv, qid);
 452 }
 453
 454 static struct rte_vdpa_dev_ops mlx5_vdpa_ops = {
 455         .get_queue_num = mlx5_vdpa_get_queue_num,
 456         .get_features = mlx5_vdpa_get_vdpa_features,
 457         .get_protocol_features = mlx5_vdpa_get_protocol_features,
 458         .dev_conf = mlx5_vdpa_dev_config,
 459         .dev_close = mlx5_vdpa_dev_close,
 460         .set_vring_state = mlx5_vdpa_set_vring_state,
 461         .set_features = mlx5_vdpa_features_set,
 462         .migration_done = NULL,
 463         .get_vfio_group_fd = NULL,
 464         .get_vfio_device_fd = mlx5_vdpa_get_device_fd,
 465         .get_notify_area = mlx5_vdpa_get_notify_area,
 466         .get_stats_names = mlx5_vdpa_get_stats_names,
 467         .get_stats = mlx5_vdpa_get_stats,
 468         .reset_stats = mlx5_vdpa_reset_stats,
 469 };
 470
 471 static struct ibv_device *
 472 mlx5_vdpa_get_ib_device_match(struct rte_pci_addr *addr)
 473 {
 474         int n;
 475         struct ibv_device **ibv_list = mlx5_glue->get_device_list(&n);
 476         struct ibv_device *ibv_match = NULL;
 477
 478         if (!ibv_list) {
 479                 rte_errno = ENOSYS;
 480                 return NULL;
 481         }
 482         while (n-- > 0) {
 483                 struct rte_pci_addr pci_addr;
 484
 485                 DRV_LOG(DEBUG, "Checking device \"%s\"..", ibv_list[n]->name);
 486                 if (mlx5_dev_to_pci_addr(ibv_list[n]->ibdev_path, &pci_addr))
 487                         continue;
 488                 if (rte_pci_addr_cmp(addr, &pci_addr))
 489                         continue;
 490                 ibv_match = ibv_list[n];
 491                 break;
 492         }
 493         if (!ibv_match)
 494                 rte_errno = ENOENT;
 495         mlx5_glue->free_device_list(ibv_list);
 496         return ibv_match;
 497 }
 498
 499 /* Try to disable ROCE by Netlink\Devlink. */
 500 static int
 501 mlx5_vdpa_nl_roce_disable(const char *addr)
 502 {
 503         int nlsk_fd = mlx5_nl_init(NETLINK_GENERIC);
 504         int devlink_id;
 505         int enable;
 506         int ret;
 507
 508         if (nlsk_fd < 0)
 509                 return nlsk_fd;
 510         devlink_id = mlx5_nl_devlink_family_id_get(nlsk_fd);
 511         if (devlink_id < 0) {
 512                 ret = devlink_id;
 513                 DRV_LOG(DEBUG, "Failed to get devlink id for ROCE operations by"
 514                         " Netlink.");
 515                 goto close;
 516         }
 517         ret = mlx5_nl_enable_roce_get(nlsk_fd, devlink_id, addr, &enable);
 518         if (ret) {
 519                 DRV_LOG(DEBUG, "Failed to get ROCE enable by Netlink: %d.",
 520                         ret);
 521                 goto close;
 522         } else if (!enable) {
 523                 DRV_LOG(INFO, "ROCE has already disabled(Netlink).");
 524                 goto close;
 525         }
 526         ret = mlx5_nl_enable_roce_set(nlsk_fd, devlink_id, addr, 0);
 527         if (ret)
 528                 DRV_LOG(DEBUG, "Failed to disable ROCE by Netlink: %d.", ret);
 529         else
 530                 DRV_LOG(INFO, "ROCE is disabled by Netlink successfully.");
 531 close:
 532         close(nlsk_fd);
 533         return ret;
 534 }
 535
 536 /* Try to disable ROCE by sysfs. */
 537 static int
 538 mlx5_vdpa_sys_roce_disable(const char *addr)
 539 {
 540         FILE *file_o;
 541         int enable;
 542         int ret;
 543
 544         MKSTR(file_p, "/sys/bus/pci/devices/%s/roce_enable", addr);
 545         file_o = fopen(file_p, "rb");
 546         if (!file_o) {
 547                 rte_errno = ENOTSUP;
 548                 return -ENOTSUP;
 549         }
 550         ret = fscanf(file_o, "%d", &enable);
 551         if (ret != 1) {
 552                 rte_errno = EINVAL;
 553                 ret = EINVAL;
 554                 goto close;
 555         } else if (!enable) {
 556                 ret = 0;
 557                 DRV_LOG(INFO, "ROCE has already disabled(sysfs).");
 558                 goto close;
 559         }
 560         fclose(file_o);
 561         file_o = fopen(file_p, "wb");
 562         if (!file_o) {
 563                 rte_errno = ENOTSUP;
 564                 return -ENOTSUP;
 565         }
 566         fprintf(file_o, "0\n");
 567         ret = 0;
 568 close:
 569         if (ret)
 570                 DRV_LOG(DEBUG, "Failed to disable ROCE by sysfs: %d.", ret);
 571         else
 572                 DRV_LOG(INFO, "ROCE is disabled by sysfs successfully.");
 573         fclose(file_o);
 574         return ret;
 575 }
 576
 577 static int
 578 mlx5_vdpa_roce_disable(struct rte_pci_addr *addr, struct ibv_device **ibv)
 579 {
 580         char addr_name[64] = {0};
 581
 582         rte_pci_device_name(addr, addr_name, sizeof(addr_name));
 583         /* Firstly try to disable ROCE by Netlink and fallback to sysfs. */
 584         if (mlx5_vdpa_nl_roce_disable(addr_name) == 0 ||
 585             mlx5_vdpa_sys_roce_disable(addr_name) == 0) {
 586                 /*
 587                  * Succeed to disable ROCE, wait for the IB device to appear
 588                  * again after reload.
 589                  */
 590                 int r;
 591                 struct ibv_device *ibv_new;
 592
 593                 for (r = MLX5_VDPA_MAX_RETRIES; r; r--) {
 594                         ibv_new = mlx5_vdpa_get_ib_device_match(addr);
 595                         if (ibv_new) {
 596                                 *ibv = ibv_new;
 597                                 return 0;
 598                         }
 599                         usleep(MLX5_VDPA_USEC);
 600                 }
 601                 DRV_LOG(ERR, "Cannot much device %s after ROCE disable, "
 602                         "retries exceed %d", addr_name, MLX5_VDPA_MAX_RETRIES);
 603                 rte_errno = EAGAIN;
 604         }
 605         return -rte_errno;
 606 }
 607
 608 /**
 609  * DPDK callback to register a PCI device.
 610  *
 611  * This function spawns vdpa device out of a given PCI device.
 612  *
 613  * @param[in] pci_drv
 614  *   PCI driver structure (mlx5_vpda_driver).
 615  * @param[in] pci_dev
 616  *   PCI device information.
 617  *
 618  * @return
 619  *   0 on success, 1 to skip this driver, a negative errno value otherwise
 620  *   and rte_errno is set.
 621  */
 622 static int
 623 mlx5_vdpa_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
 624                     struct rte_pci_device *pci_dev __rte_unused)
 625 {
 626         struct ibv_device *ibv;
 627         struct mlx5_vdpa_priv *priv = NULL;
 628         struct ibv_context *ctx = NULL;
 629         struct mlx5_hca_attr attr;
 630         int ret;
 631
 632         if (mlx5_class_get(pci_dev->device.devargs) != MLX5_CLASS_VDPA) {
 633                 DRV_LOG(DEBUG, "Skip probing - should be probed by other mlx5"
 634                         " driver.");
 635                 return 1;
 636         }
 637         ibv = mlx5_vdpa_get_ib_device_match(&pci_dev->addr);
 638         if (!ibv) {
 639                 DRV_LOG(ERR, "No matching IB device for PCI slot "
 640                         PCI_PRI_FMT ".", pci_dev->addr.domain,
 641                         pci_dev->addr.bus, pci_dev->addr.devid,
 642                         pci_dev->addr.function);
 643                 return -rte_errno;
 644         } else {
 645                 DRV_LOG(INFO, "PCI information matches for device \"%s\".",
 646                         ibv->name);
 647         }
 648         if (mlx5_vdpa_roce_disable(&pci_dev->addr, &ibv) != 0) {
 649                 DRV_LOG(WARNING, "Failed to disable ROCE for \"%s\".",
 650                         ibv->name);
 651                 return -rte_errno;
 652         }
 653         ctx = mlx5_glue->dv_open_device(ibv);
 654         if (!ctx) {
 655                 DRV_LOG(ERR, "Failed to open IB device \"%s\".", ibv->name);
 656                 rte_errno = ENODEV;
 657                 return -rte_errno;
 658         }
 659         ret = mlx5_devx_cmd_query_hca_attr(ctx, &attr);
 660         if (ret) {
 661                 DRV_LOG(ERR, "Unable to read HCA capabilities.");
 662                 rte_errno = ENOTSUP;
 663                 goto error;
 664         } else if (!attr.vdpa.valid || !attr.vdpa.max_num_virtio_queues) {
 665                 DRV_LOG(ERR, "Not enough capabilities to support vdpa, maybe "
 666                         "old FW/OFED version?");
 667                 rte_errno = ENOTSUP;
 668                 goto error;
 669         }
 670         if (!attr.vdpa.queue_counters_valid)
 671                 DRV_LOG(DEBUG, "No capability to support virtq statistics.");
 672         priv = rte_zmalloc("mlx5 vDPA device private", sizeof(*priv) +
 673                            sizeof(struct mlx5_vdpa_virtq) *
 674                            attr.vdpa.max_num_virtio_queues * 2,
 675                            RTE_CACHE_LINE_SIZE);
 676         if (!priv) {
 677                 DRV_LOG(ERR, "Failed to allocate private memory.");
 678                 rte_errno = ENOMEM;
 679                 goto error;
 680         }
 681         priv->caps = attr.vdpa;
 682         priv->log_max_rqt_size = attr.log_max_rqt_size;
 683         priv->ctx = ctx;
 684         priv->pci_dev = pci_dev;
 685         priv->var = mlx5_glue->dv_alloc_var(ctx, 0);
 686         if (!priv->var) {
 687                 DRV_LOG(ERR, "Failed to allocate VAR %u.\n", errno);
 688                 goto error;
 689         }
 690         priv->id = rte_vdpa_register_device(&pci_dev->device, &mlx5_vdpa_ops);
 691         if (priv->id < 0) {
 692                 DRV_LOG(ERR, "Failed to register vDPA device.");
 693                 rte_errno = rte_errno ? rte_errno : EINVAL;
 694                 goto error;
 695         }
 696         SLIST_INIT(&priv->mr_list);
 697         pthread_mutex_lock(&priv_list_lock);
 698         TAILQ_INSERT_TAIL(&priv_list, priv, next);
 699         pthread_mutex_unlock(&priv_list_lock);
 700         return 0;
 701
 702 error:
 703         if (priv) {
 704                 if (priv->var)
 705                         mlx5_glue->dv_free_var(priv->var);
 706                 rte_free(priv);
 707         }
 708         if (ctx)
 709                 mlx5_glue->close_device(ctx);
 710         return -rte_errno;
 711 }
 712
 713 /**
 714  * DPDK callback to remove a PCI device.
 715  *
 716  * This function removes all vDPA devices belong to a given PCI device.
 717  *
 718  * @param[in] pci_dev
 719  *   Pointer to the PCI device.
 720  *
 721  * @return
 722  *   0 on success, the function cannot fail.
 723  */
 724 static int
 725 mlx5_vdpa_pci_remove(struct rte_pci_device *pci_dev)
 726 {
 727         struct mlx5_vdpa_priv *priv = NULL;
 728         int found = 0;
 729
 730         pthread_mutex_lock(&priv_list_lock);
 731         TAILQ_FOREACH(priv, &priv_list, next) {
 732                 if (!rte_pci_addr_cmp(&priv->pci_dev->addr, &pci_dev->addr)) {
 733                         found = 1;
 734                         break;
 735                 }
 736         }
 737         if (found)
 738                 TAILQ_REMOVE(&priv_list, priv, next);
 739         pthread_mutex_unlock(&priv_list_lock);
 740         if (found) {
 741                 if (priv->configured)
 742                         mlx5_vdpa_dev_close(priv->vid);
 743                 if (priv->var) {
 744                         mlx5_glue->dv_free_var(priv->var);
 745                         priv->var = NULL;
 746                 }
 747                 mlx5_glue->close_device(priv->ctx);
 748                 rte_free(priv);
 749         }
 750         return 0;
 751 }
 752
 753 static const struct rte_pci_id mlx5_vdpa_pci_id_map[] = {
 754         {
 755                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
 756                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6)
 757         },
 758         {
 759                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
 760                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6VF)
 761         },
 762         {
 763                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
 764                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6DX)
 765         },
 766         {
 767                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
 768                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6DXVF)
 769         },
 770         {
 771                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
 772                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6DXBF)
 773         },
 774         {
 775                 .vendor_id = 0
 776         }
 777 };
 778
 779 static struct rte_pci_driver mlx5_vdpa_driver = {
 780         .driver = {
 781                 .name = "mlx5_vdpa",
 782         },
 783         .id_table = mlx5_vdpa_pci_id_map,
 784         .probe = mlx5_vdpa_pci_probe,
 785         .remove = mlx5_vdpa_pci_remove,
 786         .drv_flags = 0,
 787 };
 788
 789 /**
 790  * Driver initialization routine.
 791  */
 792 RTE_INIT(rte_mlx5_vdpa_init)
 793 {
 794         /* Initialize common log type. */
 795         mlx5_vdpa_logtype = rte_log_register("pmd.vdpa.mlx5");
 796         if (mlx5_vdpa_logtype >= 0)
 797                 rte_log_set_level(mlx5_vdpa_logtype, RTE_LOG_NOTICE);
 798         if (mlx5_glue)
 799                 rte_pci_register(&mlx5_vdpa_driver);
 800 }
 801
 802 RTE_PMD_EXPORT_NAME(net_mlx5_vdpa, __COUNTER__);
 803 RTE_PMD_REGISTER_PCI_TABLE(net_mlx5_vdpa, mlx5_vdpa_pci_id_map);
 804 RTE_PMD_REGISTER_KMOD_DEP(net_mlx5_vdpa, "* ib_uverbs & mlx5_core & mlx5_ib");