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 <sys/mman.h>
   9 #include <fcntl.h>
  10 #include <netinet/in.h>
  11
  12 #include <rte_malloc.h>
  13 #include <rte_log.h>
  14 #include <rte_errno.h>
  15 #include <rte_string_fns.h>
  16 #include <rte_bus_pci.h>
  17
  18 #include <mlx5_glue.h>
  19 #include <mlx5_common.h>
  20 #include <mlx5_common_defs.h>
  21 #include <mlx5_devx_cmds.h>
  22 #include <mlx5_prm.h>
  23 #include <mlx5_nl.h>
  24
  25 #include "mlx5_vdpa_utils.h"
  26 #include "mlx5_vdpa.h"
  27
  28 #define MLX5_VDPA_DRIVER_NAME vdpa_mlx5
  29
  30 #define MLX5_VDPA_DEFAULT_FEATURES ((1ULL << VHOST_USER_F_PROTOCOL_FEATURES) | \
  31                             (1ULL << VIRTIO_F_ANY_LAYOUT) | \
  32                             (1ULL << VIRTIO_NET_F_MQ) | \
  33                             (1ULL << VIRTIO_NET_F_GUEST_ANNOUNCE) | \
  34                             (1ULL << VIRTIO_F_ORDER_PLATFORM) | \
  35                             (1ULL << VHOST_F_LOG_ALL) | \
  36                             (1ULL << VIRTIO_NET_F_MTU))
  37
  38 #define MLX5_VDPA_PROTOCOL_FEATURES \
  39                             ((1ULL << VHOST_USER_PROTOCOL_F_SLAVE_REQ) | \
  40                              (1ULL << VHOST_USER_PROTOCOL_F_SLAVE_SEND_FD) | \
  41                              (1ULL << VHOST_USER_PROTOCOL_F_HOST_NOTIFIER) | \
  42                              (1ULL << VHOST_USER_PROTOCOL_F_LOG_SHMFD) | \
  43                              (1ULL << VHOST_USER_PROTOCOL_F_MQ) | \
  44                              (1ULL << VHOST_USER_PROTOCOL_F_NET_MTU) | \
  45                              (1ULL << VHOST_USER_PROTOCOL_F_STATUS))
  46
  47 #define MLX5_VDPA_DEFAULT_NO_TRAFFIC_MAX 16LLU
  48
  49 TAILQ_HEAD(mlx5_vdpa_privs, mlx5_vdpa_priv) priv_list =
  50                                               TAILQ_HEAD_INITIALIZER(priv_list);
  51 static pthread_mutex_t priv_list_lock = PTHREAD_MUTEX_INITIALIZER;
  52
  53 static void mlx5_vdpa_dev_release(struct mlx5_vdpa_priv *priv);
  54
  55 static struct mlx5_vdpa_priv *
  56 mlx5_vdpa_find_priv_resource_by_vdev(struct rte_vdpa_device *vdev)
  57 {
  58         struct mlx5_vdpa_priv *priv;
  59         int found = 0;
  60
  61         pthread_mutex_lock(&priv_list_lock);
  62         TAILQ_FOREACH(priv, &priv_list, next) {
  63                 if (vdev == priv->vdev) {
  64                         found = 1;
  65                         break;
  66                 }
  67         }
  68         pthread_mutex_unlock(&priv_list_lock);
  69         if (!found) {
  70                 DRV_LOG(ERR, "Invalid vDPA device: %s.", vdev->device->name);
  71                 rte_errno = EINVAL;
  72                 return NULL;
  73         }
  74         return priv;
  75 }
  76
  77 static int
  78 mlx5_vdpa_get_queue_num(struct rte_vdpa_device *vdev, uint32_t *queue_num)
  79 {
  80         struct mlx5_vdpa_priv *priv =
  81                 mlx5_vdpa_find_priv_resource_by_vdev(vdev);
  82
  83         if (priv == NULL) {
  84                 DRV_LOG(ERR, "Invalid vDPA device: %s.", vdev->device->name);
  85                 return -1;
  86         }
  87         *queue_num = priv->caps.max_num_virtio_queues;
  88         return 0;
  89 }
  90
  91 static int
  92 mlx5_vdpa_get_vdpa_features(struct rte_vdpa_device *vdev, uint64_t *features)
  93 {
  94         struct mlx5_vdpa_priv *priv =
  95                 mlx5_vdpa_find_priv_resource_by_vdev(vdev);
  96
  97         if (priv == NULL) {
  98                 DRV_LOG(ERR, "Invalid vDPA device: %s.", vdev->device->name);
  99                 return -1;
 100         }
 101         *features = MLX5_VDPA_DEFAULT_FEATURES;
 102         if (priv->caps.virtio_queue_type & (1 << MLX5_VIRTQ_TYPE_PACKED))
 103                 *features |= (1ULL << VIRTIO_F_RING_PACKED);
 104         if (priv->caps.tso_ipv4)
 105                 *features |= (1ULL << VIRTIO_NET_F_HOST_TSO4);
 106         if (priv->caps.tso_ipv6)
 107                 *features |= (1ULL << VIRTIO_NET_F_HOST_TSO6);
 108         if (priv->caps.tx_csum)
 109                 *features |= (1ULL << VIRTIO_NET_F_CSUM);
 110         if (priv->caps.rx_csum)
 111                 *features |= (1ULL << VIRTIO_NET_F_GUEST_CSUM);
 112         if (priv->caps.virtio_version_1_0)
 113                 *features |= (1ULL << VIRTIO_F_VERSION_1);
 114         return 0;
 115 }
 116
 117 static int
 118 mlx5_vdpa_get_protocol_features(struct rte_vdpa_device *vdev,
 119                 uint64_t *features)
 120 {
 121         struct mlx5_vdpa_priv *priv =
 122                 mlx5_vdpa_find_priv_resource_by_vdev(vdev);
 123
 124         if (priv == NULL) {
 125                 DRV_LOG(ERR, "Invalid vDPA device: %s.", vdev->device->name);
 126                 return -1;
 127         }
 128         *features = MLX5_VDPA_PROTOCOL_FEATURES;
 129         return 0;
 130 }
 131
 132 static int
 133 mlx5_vdpa_set_vring_state(int vid, int vring, int state)
 134 {
 135         struct rte_vdpa_device *vdev = rte_vhost_get_vdpa_device(vid);
 136         struct mlx5_vdpa_priv *priv =
 137                 mlx5_vdpa_find_priv_resource_by_vdev(vdev);
 138         int ret;
 139
 140         if (priv == NULL) {
 141                 DRV_LOG(ERR, "Invalid vDPA device: %s.", vdev->device->name);
 142                 return -EINVAL;
 143         }
 144         if (vring >= (int)priv->caps.max_num_virtio_queues * 2) {
 145                 DRV_LOG(ERR, "Too big vring id: %d.", vring);
 146                 return -E2BIG;
 147         }
 148         pthread_mutex_lock(&priv->vq_config_lock);
 149         ret = mlx5_vdpa_virtq_enable(priv, vring, state);
 150         pthread_mutex_unlock(&priv->vq_config_lock);
 151         return ret;
 152 }
 153
 154 static int
 155 mlx5_vdpa_features_set(int vid)
 156 {
 157         struct rte_vdpa_device *vdev = rte_vhost_get_vdpa_device(vid);
 158         struct mlx5_vdpa_priv *priv =
 159                 mlx5_vdpa_find_priv_resource_by_vdev(vdev);
 160         uint64_t log_base, log_size;
 161         uint64_t features;
 162         int ret;
 163
 164         if (priv == NULL) {
 165                 DRV_LOG(ERR, "Invalid vDPA device: %s.", vdev->device->name);
 166                 return -EINVAL;
 167         }
 168         ret = rte_vhost_get_negotiated_features(vid, &features);
 169         if (ret) {
 170                 DRV_LOG(ERR, "Failed to get negotiated features.");
 171                 return ret;
 172         }
 173         if (RTE_VHOST_NEED_LOG(features)) {
 174                 ret = rte_vhost_get_log_base(vid, &log_base, &log_size);
 175                 if (ret) {
 176                         DRV_LOG(ERR, "Failed to get log base.");
 177                         return ret;
 178                 }
 179                 ret = mlx5_vdpa_dirty_bitmap_set(priv, log_base, log_size);
 180                 if (ret) {
 181                         DRV_LOG(ERR, "Failed to set dirty bitmap.");
 182                         return ret;
 183                 }
 184                 DRV_LOG(INFO, "mlx5 vdpa: enabling dirty logging...");
 185                 ret = mlx5_vdpa_logging_enable(priv, 1);
 186                 if (ret) {
 187                         DRV_LOG(ERR, "Failed t enable dirty logging.");
 188                         return ret;
 189                 }
 190         }
 191         return 0;
 192 }
 193
 194 static int
 195 mlx5_vdpa_mtu_set(struct mlx5_vdpa_priv *priv)
 196 {
 197         struct ifreq request;
 198         uint16_t vhost_mtu = 0;
 199         uint16_t kern_mtu = 0;
 200         int ret = rte_vhost_get_mtu(priv->vid, &vhost_mtu);
 201         int sock;
 202         int retries = MLX5_VDPA_MAX_RETRIES;
 203
 204         if (ret) {
 205                 DRV_LOG(DEBUG, "Cannot get vhost MTU - %d.", ret);
 206                 return ret;
 207         }
 208         if (!vhost_mtu) {
 209                 DRV_LOG(DEBUG, "Vhost MTU is 0.");
 210                 return ret;
 211         }
 212         ret = mlx5_get_ifname_sysfs
 213                                 (mlx5_os_get_ctx_device_name(priv->cdev->ctx),
 214                                  request.ifr_name);
 215         if (ret) {
 216                 DRV_LOG(DEBUG, "Cannot get kernel IF name - %d.", ret);
 217                 return ret;
 218         }
 219         sock = socket(PF_INET, SOCK_DGRAM, IPPROTO_IP);
 220         if (sock == -1) {
 221                 DRV_LOG(DEBUG, "Cannot open IF socket.");
 222                 return sock;
 223         }
 224         while (retries--) {
 225                 ret = ioctl(sock, SIOCGIFMTU, &request);
 226                 if (ret == -1)
 227                         break;
 228                 kern_mtu = request.ifr_mtu;
 229                 DRV_LOG(DEBUG, "MTU: current %d requested %d.", (int)kern_mtu,
 230                         (int)vhost_mtu);
 231                 if (kern_mtu == vhost_mtu)
 232                         break;
 233                 request.ifr_mtu = vhost_mtu;
 234                 ret = ioctl(sock, SIOCSIFMTU, &request);
 235                 if (ret == -1)
 236                         break;
 237                 request.ifr_mtu = 0;
 238                 usleep(MLX5_VDPA_USEC);
 239         }
 240         close(sock);
 241         return kern_mtu == vhost_mtu ? 0 : -1;
 242 }
 243
 244 static void
 245 mlx5_vdpa_dev_cache_clean(struct mlx5_vdpa_priv *priv)
 246 {
 247         mlx5_vdpa_virtqs_cleanup(priv);
 248         mlx5_vdpa_mem_dereg(priv);
 249 }
 250
 251 static int
 252 mlx5_vdpa_dev_close(int vid)
 253 {
 254         struct rte_vdpa_device *vdev = rte_vhost_get_vdpa_device(vid);
 255         struct mlx5_vdpa_priv *priv =
 256                 mlx5_vdpa_find_priv_resource_by_vdev(vdev);
 257         int ret = 0;
 258
 259         if (priv == NULL) {
 260                 DRV_LOG(ERR, "Invalid vDPA device: %s.", vdev->device->name);
 261                 return -1;
 262         }
 263         mlx5_vdpa_cqe_event_unset(priv);
 264         if (priv->state == MLX5_VDPA_STATE_CONFIGURED) {
 265                 ret |= mlx5_vdpa_lm_log(priv);
 266                 priv->state = MLX5_VDPA_STATE_IN_PROGRESS;
 267         }
 268         mlx5_vdpa_steer_unset(priv);
 269         mlx5_vdpa_virtqs_release(priv);
 270         if (priv->lm_mr.addr)
 271                 mlx5_os_wrapped_mkey_destroy(&priv->lm_mr);
 272         priv->state = MLX5_VDPA_STATE_PROBED;
 273         if (!priv->connected)
 274                 mlx5_vdpa_dev_cache_clean(priv);
 275         priv->vid = 0;
 276         /* The mutex may stay locked after event thread cancel - initiate it. */
 277         pthread_mutex_init(&priv->vq_config_lock, NULL);
 278         DRV_LOG(INFO, "vDPA device %d was closed.", vid);
 279         return ret;
 280 }
 281
 282 static int
 283 mlx5_vdpa_dev_config(int vid)
 284 {
 285         struct rte_vdpa_device *vdev = rte_vhost_get_vdpa_device(vid);
 286         struct mlx5_vdpa_priv *priv =
 287                 mlx5_vdpa_find_priv_resource_by_vdev(vdev);
 288
 289         if (priv == NULL) {
 290                 DRV_LOG(ERR, "Invalid vDPA device: %s.", vdev->device->name);
 291                 return -EINVAL;
 292         }
 293         if (priv->state == MLX5_VDPA_STATE_CONFIGURED &&
 294             mlx5_vdpa_dev_close(vid)) {
 295                 DRV_LOG(ERR, "Failed to reconfigure vid %d.", vid);
 296                 return -1;
 297         }
 298         priv->vid = vid;
 299         priv->connected = true;
 300         if (mlx5_vdpa_mtu_set(priv))
 301                 DRV_LOG(WARNING, "MTU cannot be set on device %s.",
 302                                 vdev->device->name);
 303         if (mlx5_vdpa_mem_register(priv) ||
 304             mlx5_vdpa_virtqs_prepare(priv) || mlx5_vdpa_steer_setup(priv) ||
 305             mlx5_vdpa_cqe_event_setup(priv)) {
 306                 mlx5_vdpa_dev_close(vid);
 307                 return -1;
 308         }
 309         priv->state = MLX5_VDPA_STATE_CONFIGURED;
 310         DRV_LOG(INFO, "vDPA device %d was configured.", vid);
 311         return 0;
 312 }
 313
 314 static int
 315 mlx5_vdpa_get_device_fd(int vid)
 316 {
 317         struct rte_vdpa_device *vdev = rte_vhost_get_vdpa_device(vid);
 318         struct mlx5_vdpa_priv *priv =
 319                 mlx5_vdpa_find_priv_resource_by_vdev(vdev);
 320
 321         if (priv == NULL) {
 322                 DRV_LOG(ERR, "Invalid vDPA device: %s.", vdev->device->name);
 323                 return -EINVAL;
 324         }
 325         return ((struct ibv_context *)priv->cdev->ctx)->cmd_fd;
 326 }
 327
 328 static int
 329 mlx5_vdpa_get_notify_area(int vid, int qid, uint64_t *offset, uint64_t *size)
 330 {
 331         struct rte_vdpa_device *vdev = rte_vhost_get_vdpa_device(vid);
 332         struct mlx5_vdpa_priv *priv =
 333                 mlx5_vdpa_find_priv_resource_by_vdev(vdev);
 334
 335         RTE_SET_USED(qid);
 336         if (priv == NULL) {
 337                 DRV_LOG(ERR, "Invalid vDPA device: %s.", vdev->device->name);
 338                 return -EINVAL;
 339         }
 340         if (!priv->var) {
 341                 DRV_LOG(ERR, "VAR was not created for device %s, is the device"
 342                         " configured?.", vdev->device->name);
 343                 return -EINVAL;
 344         }
 345         *offset = priv->var->mmap_off;
 346         *size = priv->var->length;
 347         return 0;
 348 }
 349
 350 static int
 351 mlx5_vdpa_get_stats_names(struct rte_vdpa_device *vdev,
 352                 struct rte_vdpa_stat_name *stats_names,
 353                 unsigned int size)
 354 {
 355         static const char *mlx5_vdpa_stats_names[MLX5_VDPA_STATS_MAX] = {
 356                 "received_descriptors",
 357                 "completed_descriptors",
 358                 "bad descriptor errors",
 359                 "exceed max chain",
 360                 "invalid buffer",
 361                 "completion errors",
 362         };
 363         struct mlx5_vdpa_priv *priv =
 364                 mlx5_vdpa_find_priv_resource_by_vdev(vdev);
 365         unsigned int i;
 366
 367         if (priv == NULL) {
 368                 DRV_LOG(ERR, "Invalid device: %s.", vdev->device->name);
 369                 return -ENODEV;
 370         }
 371         if (!stats_names)
 372                 return MLX5_VDPA_STATS_MAX;
 373         size = RTE_MIN(size, (unsigned int)MLX5_VDPA_STATS_MAX);
 374         for (i = 0; i < size; ++i)
 375                 strlcpy(stats_names[i].name, mlx5_vdpa_stats_names[i],
 376                         RTE_VDPA_STATS_NAME_SIZE);
 377         return size;
 378 }
 379
 380 static int
 381 mlx5_vdpa_get_stats(struct rte_vdpa_device *vdev, int qid,
 382                 struct rte_vdpa_stat *stats, unsigned int n)
 383 {
 384         struct mlx5_vdpa_priv *priv =
 385                 mlx5_vdpa_find_priv_resource_by_vdev(vdev);
 386
 387         if (priv == NULL) {
 388                 DRV_LOG(ERR, "Invalid device: %s.", vdev->device->name);
 389                 return -ENODEV;
 390         }
 391         if (qid >= (int)priv->caps.max_num_virtio_queues * 2) {
 392                 DRV_LOG(ERR, "Too big vring id: %d for device %s.", qid,
 393                                 vdev->device->name);
 394                 return -E2BIG;
 395         }
 396         if (!priv->caps.queue_counters_valid) {
 397                 DRV_LOG(ERR, "Virtq statistics is not supported for device %s.",
 398                         vdev->device->name);
 399                 return -ENOTSUP;
 400         }
 401         return mlx5_vdpa_virtq_stats_get(priv, qid, stats, n);
 402 }
 403
 404 static int
 405 mlx5_vdpa_reset_stats(struct rte_vdpa_device *vdev, int qid)
 406 {
 407         struct mlx5_vdpa_priv *priv =
 408                 mlx5_vdpa_find_priv_resource_by_vdev(vdev);
 409
 410         if (priv == NULL) {
 411                 DRV_LOG(ERR, "Invalid device: %s.", vdev->device->name);
 412                 return -ENODEV;
 413         }
 414         if (qid >= (int)priv->caps.max_num_virtio_queues * 2) {
 415                 DRV_LOG(ERR, "Too big vring id: %d for device %s.", qid,
 416                                 vdev->device->name);
 417                 return -E2BIG;
 418         }
 419         if (!priv->caps.queue_counters_valid) {
 420                 DRV_LOG(ERR, "Virtq statistics is not supported for device %s.",
 421                         vdev->device->name);
 422                 return -ENOTSUP;
 423         }
 424         return mlx5_vdpa_virtq_stats_reset(priv, qid);
 425 }
 426
 427 static int
 428 mlx5_vdpa_dev_cleanup(int vid)
 429 {
 430         struct rte_vdpa_device *vdev = rte_vhost_get_vdpa_device(vid);
 431         struct mlx5_vdpa_priv *priv;
 432
 433         if (vdev == NULL)
 434                 return -1;
 435         priv = mlx5_vdpa_find_priv_resource_by_vdev(vdev);
 436         if (priv == NULL) {
 437                 DRV_LOG(ERR, "Invalid vDPA device: %s.", vdev->device->name);
 438                 return -1;
 439         }
 440         if (priv->state == MLX5_VDPA_STATE_PROBED)
 441                 mlx5_vdpa_dev_cache_clean(priv);
 442         priv->connected = false;
 443         return 0;
 444 }
 445
 446 static struct rte_vdpa_dev_ops mlx5_vdpa_ops = {
 447         .get_queue_num = mlx5_vdpa_get_queue_num,
 448         .get_features = mlx5_vdpa_get_vdpa_features,
 449         .get_protocol_features = mlx5_vdpa_get_protocol_features,
 450         .dev_conf = mlx5_vdpa_dev_config,
 451         .dev_close = mlx5_vdpa_dev_close,
 452         .dev_cleanup = mlx5_vdpa_dev_cleanup,
 453         .set_vring_state = mlx5_vdpa_set_vring_state,
 454         .set_features = mlx5_vdpa_features_set,
 455         .migration_done = NULL,
 456         .get_vfio_group_fd = NULL,
 457         .get_vfio_device_fd = mlx5_vdpa_get_device_fd,
 458         .get_notify_area = mlx5_vdpa_get_notify_area,
 459         .get_stats_names = mlx5_vdpa_get_stats_names,
 460         .get_stats = mlx5_vdpa_get_stats,
 461         .reset_stats = mlx5_vdpa_reset_stats,
 462 };
 463
 464 static int
 465 mlx5_vdpa_args_check_handler(const char *key, const char *val, void *opaque)
 466 {
 467         struct mlx5_vdpa_priv *priv = opaque;
 468         unsigned long tmp;
 469         int n_cores = sysconf(_SC_NPROCESSORS_ONLN);
 470
 471         errno = 0;
 472         tmp = strtoul(val, NULL, 0);
 473         if (errno) {
 474                 DRV_LOG(WARNING, "%s: \"%s\" is an invalid integer.", key, val);
 475                 return -errno;
 476         }
 477         if (strcmp(key, "event_mode") == 0) {
 478                 if (tmp <= MLX5_VDPA_EVENT_MODE_ONLY_INTERRUPT)
 479                         priv->event_mode = (int)tmp;
 480                 else
 481                         DRV_LOG(WARNING, "Invalid event_mode %s.", val);
 482         } else if (strcmp(key, "event_us") == 0) {
 483                 priv->event_us = (uint32_t)tmp;
 484         } else if (strcmp(key, "no_traffic_time") == 0) {
 485                 priv->no_traffic_max = (uint32_t)tmp;
 486         } else if (strcmp(key, "event_core") == 0) {
 487                 if (tmp >= (unsigned long)n_cores)
 488                         DRV_LOG(WARNING, "Invalid event_core %s.", val);
 489                 else
 490                         priv->event_core = tmp;
 491         } else if (strcmp(key, "hw_latency_mode") == 0) {
 492                 priv->hw_latency_mode = (uint32_t)tmp;
 493         } else if (strcmp(key, "hw_max_latency_us") == 0) {
 494                 priv->hw_max_latency_us = (uint32_t)tmp;
 495         } else if (strcmp(key, "hw_max_pending_comp") == 0) {
 496                 priv->hw_max_pending_comp = (uint32_t)tmp;
 497         }
 498         return 0;
 499 }
 500
 501 static void
 502 mlx5_vdpa_config_get(struct mlx5_kvargs_ctrl *mkvlist,
 503                      struct mlx5_vdpa_priv *priv)
 504 {
 505         const char **params = (const char *[]){
 506                 "event_core",
 507                 "event_mode",
 508                 "event_us",
 509                 "hw_latency_mode",
 510                 "hw_max_latency_us",
 511                 "hw_max_pending_comp",
 512                 "no_traffic_time",
 513                 NULL,
 514         };
 515
 516         priv->event_mode = MLX5_VDPA_EVENT_MODE_FIXED_TIMER;
 517         priv->event_us = 0;
 518         priv->event_core = -1;
 519         priv->no_traffic_max = MLX5_VDPA_DEFAULT_NO_TRAFFIC_MAX;
 520         if (mkvlist == NULL)
 521                 return;
 522         mlx5_kvargs_process(mkvlist, params, mlx5_vdpa_args_check_handler,
 523                             priv);
 524         if (!priv->event_us &&
 525             priv->event_mode == MLX5_VDPA_EVENT_MODE_DYNAMIC_TIMER)
 526                 priv->event_us = MLX5_VDPA_DEFAULT_TIMER_STEP_US;
 527         DRV_LOG(DEBUG, "event mode is %d.", priv->event_mode);
 528         DRV_LOG(DEBUG, "event_us is %u us.", priv->event_us);
 529         DRV_LOG(DEBUG, "no traffic max is %u.", priv->no_traffic_max);
 530 }
 531
 532 static int
 533 mlx5_vdpa_create_dev_resources(struct mlx5_vdpa_priv *priv)
 534 {
 535         struct mlx5_devx_tis_attr tis_attr = {0};
 536         struct ibv_context *ctx = priv->cdev->ctx;
 537         uint32_t i;
 538         int retry;
 539
 540         for (retry = 0; retry < 7; retry++) {
 541                 priv->var = mlx5_glue->dv_alloc_var(ctx, 0);
 542                 if (priv->var != NULL)
 543                         break;
 544                 DRV_LOG(WARNING, "Failed to allocate VAR, retry %d.", retry);
 545                 /* Wait Qemu release VAR during vdpa restart, 0.1 sec based. */
 546                 usleep(100000U << retry);
 547         }
 548         if (!priv->var) {
 549                 DRV_LOG(ERR, "Failed to allocate VAR %u.", errno);
 550                 rte_errno = ENOMEM;
 551                 return -rte_errno;
 552         }
 553         /* Always map the entire page. */
 554         priv->virtq_db_addr = mmap(NULL, priv->var->length, PROT_READ |
 555                                    PROT_WRITE, MAP_SHARED, ctx->cmd_fd,
 556                                    priv->var->mmap_off);
 557         if (priv->virtq_db_addr == MAP_FAILED) {
 558                 DRV_LOG(ERR, "Failed to map doorbell page %u.", errno);
 559                 priv->virtq_db_addr = NULL;
 560                 rte_errno = errno;
 561                 return -rte_errno;
 562         }
 563         DRV_LOG(DEBUG, "VAR address of doorbell mapping is %p.",
 564                 priv->virtq_db_addr);
 565         priv->td = mlx5_devx_cmd_create_td(ctx);
 566         if (!priv->td) {
 567                 DRV_LOG(ERR, "Failed to create transport domain.");
 568                 rte_errno = errno;
 569                 return -rte_errno;
 570         }
 571         tis_attr.transport_domain = priv->td->id;
 572         for (i = 0; i < priv->num_lag_ports; i++) {
 573                 /* 0 is auto affinity, non-zero value to propose port. */
 574                 tis_attr.lag_tx_port_affinity = i + 1;
 575                 priv->tiss[i] = mlx5_devx_cmd_create_tis(ctx, &tis_attr);
 576                 if (!priv->tiss[i]) {
 577                         DRV_LOG(ERR, "Failed to create TIS %u.", i);
 578                         return -rte_errno;
 579                 }
 580         }
 581         priv->null_mr = mlx5_glue->alloc_null_mr(priv->cdev->pd);
 582         if (!priv->null_mr) {
 583                 DRV_LOG(ERR, "Failed to allocate null MR.");
 584                 rte_errno = errno;
 585                 return -rte_errno;
 586         }
 587         DRV_LOG(DEBUG, "Dump fill Mkey = %u.", priv->null_mr->lkey);
 588 #ifdef HAVE_MLX5DV_DR
 589         priv->steer.domain = mlx5_glue->dr_create_domain(ctx,
 590                                         MLX5DV_DR_DOMAIN_TYPE_NIC_RX);
 591         if (!priv->steer.domain) {
 592                 DRV_LOG(ERR, "Failed to create Rx domain.");
 593                 rte_errno = errno;
 594                 return -rte_errno;
 595         }
 596 #endif
 597         priv->steer.tbl = mlx5_glue->dr_create_flow_tbl(priv->steer.domain, 0);
 598         if (!priv->steer.tbl) {
 599                 DRV_LOG(ERR, "Failed to create table 0 with Rx domain.");
 600                 rte_errno = errno;
 601                 return -rte_errno;
 602         }
 603         if (mlx5_vdpa_err_event_setup(priv) != 0)
 604                 return -rte_errno;
 605         if (mlx5_vdpa_event_qp_global_prepare(priv))
 606                 return -rte_errno;
 607         return 0;
 608 }
 609
 610 static int
 611 mlx5_vdpa_dev_probe(struct mlx5_common_device *cdev,
 612                     struct mlx5_kvargs_ctrl *mkvlist)
 613 {
 614         struct mlx5_vdpa_priv *priv = NULL;
 615         struct mlx5_hca_attr *attr = &cdev->config.hca_attr;
 616
 617         if (!attr->vdpa.valid || !attr->vdpa.max_num_virtio_queues) {
 618                 DRV_LOG(ERR, "Not enough capabilities to support vdpa, maybe "
 619                         "old FW/OFED version?");
 620                 rte_errno = ENOTSUP;
 621                 return -rte_errno;
 622         }
 623         if (!attr->vdpa.queue_counters_valid)
 624                 DRV_LOG(DEBUG, "No capability to support virtq statistics.");
 625         priv = rte_zmalloc("mlx5 vDPA device private", sizeof(*priv) +
 626                            sizeof(struct mlx5_vdpa_virtq) *
 627                            attr->vdpa.max_num_virtio_queues * 2,
 628                            RTE_CACHE_LINE_SIZE);
 629         if (!priv) {
 630                 DRV_LOG(ERR, "Failed to allocate private memory.");
 631                 rte_errno = ENOMEM;
 632                 return -rte_errno;
 633         }
 634         priv->caps = attr->vdpa;
 635         priv->log_max_rqt_size = attr->log_max_rqt_size;
 636         priv->num_lag_ports = attr->num_lag_ports;
 637         if (attr->num_lag_ports == 0)
 638                 priv->num_lag_ports = 1;
 639         pthread_mutex_init(&priv->vq_config_lock, NULL);
 640         priv->cdev = cdev;
 641         if (mlx5_vdpa_create_dev_resources(priv))
 642                 goto error;
 643         priv->vdev = rte_vdpa_register_device(cdev->dev, &mlx5_vdpa_ops);
 644         if (priv->vdev == NULL) {
 645                 DRV_LOG(ERR, "Failed to register vDPA device.");
 646                 rte_errno = rte_errno ? rte_errno : EINVAL;
 647                 goto error;
 648         }
 649         mlx5_vdpa_config_get(mkvlist, priv);
 650         SLIST_INIT(&priv->mr_list);
 651         pthread_mutex_lock(&priv_list_lock);
 652         TAILQ_INSERT_TAIL(&priv_list, priv, next);
 653         pthread_mutex_unlock(&priv_list_lock);
 654         return 0;
 655 error:
 656         if (priv)
 657                 mlx5_vdpa_dev_release(priv);
 658         return -rte_errno;
 659 }
 660
 661 static int
 662 mlx5_vdpa_dev_remove(struct mlx5_common_device *cdev)
 663 {
 664         struct mlx5_vdpa_priv *priv = NULL;
 665         int found = 0;
 666
 667         pthread_mutex_lock(&priv_list_lock);
 668         TAILQ_FOREACH(priv, &priv_list, next) {
 669                 if (priv->vdev->device == cdev->dev) {
 670                         found = 1;
 671                         break;
 672                 }
 673         }
 674         if (found)
 675                 TAILQ_REMOVE(&priv_list, priv, next);
 676         pthread_mutex_unlock(&priv_list_lock);
 677         if (found)
 678                 mlx5_vdpa_dev_release(priv);
 679         return 0;
 680 }
 681
 682 static void
 683 mlx5_vdpa_release_dev_resources(struct mlx5_vdpa_priv *priv)
 684 {
 685         uint32_t i;
 686
 687         mlx5_vdpa_dev_cache_clean(priv);
 688         for (i = 0; i < priv->caps.max_num_virtio_queues * 2; i++) {
 689                 if (!priv->virtqs[i].counters)
 690                         continue;
 691                 claim_zero(mlx5_devx_cmd_destroy(priv->virtqs[i].counters));
 692         }
 693         mlx5_vdpa_event_qp_global_release(priv);
 694         mlx5_vdpa_err_event_unset(priv);
 695         if (priv->steer.tbl)
 696                 claim_zero(mlx5_glue->dr_destroy_flow_tbl(priv->steer.tbl));
 697         if (priv->steer.domain)
 698                 claim_zero(mlx5_glue->dr_destroy_domain(priv->steer.domain));
 699         if (priv->null_mr)
 700                 claim_zero(mlx5_glue->dereg_mr(priv->null_mr));
 701         for (i = 0; i < priv->num_lag_ports; i++) {
 702                 if (priv->tiss[i])
 703                         claim_zero(mlx5_devx_cmd_destroy(priv->tiss[i]));
 704         }
 705         if (priv->td)
 706                 claim_zero(mlx5_devx_cmd_destroy(priv->td));
 707         if (priv->virtq_db_addr)
 708                 claim_zero(munmap(priv->virtq_db_addr, priv->var->length));
 709         if (priv->var)
 710                 mlx5_glue->dv_free_var(priv->var);
 711 }
 712
 713 static void
 714 mlx5_vdpa_dev_release(struct mlx5_vdpa_priv *priv)
 715 {
 716         if (priv->state == MLX5_VDPA_STATE_CONFIGURED)
 717                 mlx5_vdpa_dev_close(priv->vid);
 718         mlx5_vdpa_release_dev_resources(priv);
 719         if (priv->vdev)
 720                 rte_vdpa_unregister_device(priv->vdev);
 721         pthread_mutex_destroy(&priv->vq_config_lock);
 722         rte_free(priv);
 723 }
 724
 725 static const struct rte_pci_id mlx5_vdpa_pci_id_map[] = {
 726         {
 727                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
 728                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6)
 729         },
 730         {
 731                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
 732                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6VF)
 733         },
 734         {
 735                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
 736                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6DX)
 737         },
 738         {
 739                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
 740                                 PCI_DEVICE_ID_MELLANOX_CONNECTXVF)
 741         },
 742         {
 743                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
 744                                 PCI_DEVICE_ID_MELLANOX_CONNECTX6DXBF)
 745         },
 746         {
 747                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
 748                                 PCI_DEVICE_ID_MELLANOX_CONNECTX7)
 749         },
 750         {
 751                 RTE_PCI_DEVICE(PCI_VENDOR_ID_MELLANOX,
 752                                 PCI_DEVICE_ID_MELLANOX_CONNECTX7BF)
 753         },
 754         {
 755                 .vendor_id = 0
 756         }
 757 };
 758
 759 static struct mlx5_class_driver mlx5_vdpa_driver = {
 760         .drv_class = MLX5_CLASS_VDPA,
 761         .name = RTE_STR(MLX5_VDPA_DRIVER_NAME),
 762         .id_table = mlx5_vdpa_pci_id_map,
 763         .probe = mlx5_vdpa_dev_probe,
 764         .remove = mlx5_vdpa_dev_remove,
 765 };
 766
 767 RTE_LOG_REGISTER_DEFAULT(mlx5_vdpa_logtype, NOTICE)
 768
 769 /**
 770  * Driver initialization routine.
 771  */
 772 RTE_INIT(rte_mlx5_vdpa_init)
 773 {
 774         mlx5_common_init();
 775         if (mlx5_glue)
 776                 mlx5_class_driver_register(&mlx5_vdpa_driver);
 777 }
 778
 779 RTE_PMD_EXPORT_NAME(MLX5_VDPA_DRIVER_NAME, __COUNTER__);
 780 RTE_PMD_REGISTER_PCI_TABLE(MLX5_VDPA_DRIVER_NAME, mlx5_vdpa_pci_id_map);
 781 RTE_PMD_REGISTER_KMOD_DEP(MLX5_VDPA_DRIVER_NAME, "* ib_uverbs & mlx5_core & mlx5_ib");