drivers/net/virtio/virtio_user/virtio_user_dev.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2010-2016 Intel Corporation
   3  */
   4
   5 #include <stdint.h>
   6 #include <stdio.h>
   7 #include <fcntl.h>
   8 #include <string.h>
   9 #include <errno.h>
  10 #include <sys/mman.h>
  11 #include <unistd.h>
  12 #include <sys/eventfd.h>
  13 #include <sys/types.h>
  14 #include <sys/stat.h>
  15
  16 #include "vhost.h"
  17 #include "virtio_user_dev.h"
  18 #include "../virtio_ethdev.h"
  19
  20 #define VIRTIO_USER_MEM_EVENT_CLB_NAME "virtio_user_mem_event_clb"
  21
  22 static int
  23 virtio_user_create_queue(struct virtio_user_dev *dev, uint32_t queue_sel)
  24 {
  25         /* Of all per virtqueue MSGs, make sure VHOST_SET_VRING_CALL come
  26          * firstly because vhost depends on this msg to allocate virtqueue
  27          * pair.
  28          */
  29         struct vhost_vring_file file;
  30
  31         file.index = queue_sel;
  32         file.fd = dev->callfds[queue_sel];
  33         dev->ops->send_request(dev, VHOST_USER_SET_VRING_CALL, &file);
  34
  35         return 0;
  36 }
  37
  38 static int
  39 virtio_user_kick_queue(struct virtio_user_dev *dev, uint32_t queue_sel)
  40 {
  41         struct vhost_vring_file file;
  42         struct vhost_vring_state state;
  43         struct vring *vring = &dev->vrings[queue_sel];
  44         struct vhost_vring_addr addr = {
  45                 .index = queue_sel,
  46                 .desc_user_addr = (uint64_t)(uintptr_t)vring->desc,
  47                 .avail_user_addr = (uint64_t)(uintptr_t)vring->avail,
  48                 .used_user_addr = (uint64_t)(uintptr_t)vring->used,
  49                 .log_guest_addr = 0,
  50                 .flags = 0, /* disable log */
  51         };
  52
  53         state.index = queue_sel;
  54         state.num = vring->num;
  55         dev->ops->send_request(dev, VHOST_USER_SET_VRING_NUM, &state);
  56
  57         state.index = queue_sel;
  58         state.num = 0; /* no reservation */
  59         dev->ops->send_request(dev, VHOST_USER_SET_VRING_BASE, &state);
  60
  61         dev->ops->send_request(dev, VHOST_USER_SET_VRING_ADDR, &addr);
  62
  63         /* Of all per virtqueue MSGs, make sure VHOST_USER_SET_VRING_KICK comes
  64          * lastly because vhost depends on this msg to judge if
  65          * virtio is ready.
  66          */
  67         file.index = queue_sel;
  68         file.fd = dev->kickfds[queue_sel];
  69         dev->ops->send_request(dev, VHOST_USER_SET_VRING_KICK, &file);
  70
  71         return 0;
  72 }
  73
  74 static int
  75 virtio_user_queue_setup(struct virtio_user_dev *dev,
  76                         int (*fn)(struct virtio_user_dev *, uint32_t))
  77 {
  78         uint32_t i, queue_sel;
  79
  80         for (i = 0; i < dev->max_queue_pairs; ++i) {
  81                 queue_sel = 2 * i + VTNET_SQ_RQ_QUEUE_IDX;
  82                 if (fn(dev, queue_sel) < 0) {
  83                         PMD_DRV_LOG(INFO, "setup rx vq fails: %u", i);
  84                         return -1;
  85                 }
  86         }
  87         for (i = 0; i < dev->max_queue_pairs; ++i) {
  88                 queue_sel = 2 * i + VTNET_SQ_TQ_QUEUE_IDX;
  89                 if (fn(dev, queue_sel) < 0) {
  90                         PMD_DRV_LOG(INFO, "setup tx vq fails: %u", i);
  91                         return -1;
  92                 }
  93         }
  94
  95         return 0;
  96 }
  97
  98 int
  99 is_vhost_user_by_type(const char *path)
 100 {
 101         struct stat sb;
 102
 103         if (stat(path, &sb) == -1)
 104                 return 0;
 105
 106         return S_ISSOCK(sb.st_mode);
 107 }
 108
 109 int
 110 virtio_user_start_device(struct virtio_user_dev *dev)
 111 {
 112         uint64_t features;
 113         int ret;
 114
 115         pthread_mutex_lock(&dev->mutex);
 116
 117         if (is_vhost_user_by_type(dev->path) && dev->vhostfd < 0)
 118                 goto error;
 119
 120         /* Do not check return as already done in init, or reset in stop */
 121         dev->ops->send_request(dev, VHOST_USER_SET_OWNER, NULL);
 122
 123         /* Step 0: tell vhost to create queues */
 124         if (virtio_user_queue_setup(dev, virtio_user_create_queue) < 0)
 125                 goto error;
 126
 127         /* Step 1: set features */
 128         features = dev->features;
 129         /* Strip VIRTIO_NET_F_MAC, as MAC address is handled in vdev init */
 130         features &= ~(1ull << VIRTIO_NET_F_MAC);
 131         /* Strip VIRTIO_NET_F_CTRL_VQ, as devices do not really need to know */
 132         features &= ~(1ull << VIRTIO_NET_F_CTRL_VQ);
 133         features &= ~(1ull << VIRTIO_NET_F_STATUS);
 134         ret = dev->ops->send_request(dev, VHOST_USER_SET_FEATURES, &features);
 135         if (ret < 0)
 136                 goto error;
 137         PMD_DRV_LOG(INFO, "set features: %" PRIx64, features);
 138
 139         /* Step 2: share memory regions */
 140         ret = dev->ops->send_request(dev, VHOST_USER_SET_MEM_TABLE, NULL);
 141         if (ret < 0)
 142                 goto error;
 143
 144         /* Step 3: kick queues */
 145         if (virtio_user_queue_setup(dev, virtio_user_kick_queue) < 0)
 146                 goto error;
 147
 148         /* Step 4: enable queues
 149          * we enable the 1st queue pair by default.
 150          */
 151         dev->ops->enable_qp(dev, 0, 1);
 152
 153         dev->started = true;
 154         pthread_mutex_unlock(&dev->mutex);
 155
 156         return 0;
 157 error:
 158         pthread_mutex_unlock(&dev->mutex);
 159         /* TODO: free resource here or caller to check */
 160         return -1;
 161 }
 162
 163 int virtio_user_stop_device(struct virtio_user_dev *dev)
 164 {
 165         uint32_t i;
 166
 167         pthread_mutex_lock(&dev->mutex);
 168         for (i = 0; i < dev->max_queue_pairs; ++i)
 169                 dev->ops->enable_qp(dev, i, 0);
 170
 171         if (dev->ops->send_request(dev, VHOST_USER_RESET_OWNER, NULL) < 0) {
 172                 PMD_DRV_LOG(INFO, "Failed to reset the device\n");
 173                 pthread_mutex_unlock(&dev->mutex);
 174                 return -1;
 175         }
 176         dev->started = false;
 177         pthread_mutex_unlock(&dev->mutex);
 178
 179         return 0;
 180 }
 181
 182 static inline void
 183 parse_mac(struct virtio_user_dev *dev, const char *mac)
 184 {
 185         int i, r;
 186         uint32_t tmp[ETHER_ADDR_LEN];
 187
 188         if (!mac)
 189                 return;
 190
 191         r = sscanf(mac, "%x:%x:%x:%x:%x:%x", &tmp[0],
 192                         &tmp[1], &tmp[2], &tmp[3], &tmp[4], &tmp[5]);
 193         if (r == ETHER_ADDR_LEN) {
 194                 for (i = 0; i < ETHER_ADDR_LEN; ++i)
 195                         dev->mac_addr[i] = (uint8_t)tmp[i];
 196                 dev->mac_specified = 1;
 197         } else {
 198                 /* ignore the wrong mac, use random mac */
 199                 PMD_DRV_LOG(ERR, "wrong format of mac: %s", mac);
 200         }
 201 }
 202
 203 static int
 204 virtio_user_dev_init_notify(struct virtio_user_dev *dev)
 205 {
 206         uint32_t i, j;
 207         int callfd;
 208         int kickfd;
 209
 210         for (i = 0; i < VIRTIO_MAX_VIRTQUEUES; ++i) {
 211                 if (i >= dev->max_queue_pairs * 2) {
 212                         dev->kickfds[i] = -1;
 213                         dev->callfds[i] = -1;
 214                         continue;
 215                 }
 216
 217                 /* May use invalid flag, but some backend uses kickfd and
 218                  * callfd as criteria to judge if dev is alive. so finally we
 219                  * use real event_fd.
 220                  */
 221                 callfd = eventfd(0, EFD_CLOEXEC | EFD_NONBLOCK);
 222                 if (callfd < 0) {
 223                         PMD_DRV_LOG(ERR, "callfd error, %s", strerror(errno));
 224                         break;
 225                 }
 226                 kickfd = eventfd(0, EFD_CLOEXEC | EFD_NONBLOCK);
 227                 if (kickfd < 0) {
 228                         PMD_DRV_LOG(ERR, "kickfd error, %s", strerror(errno));
 229                         break;
 230                 }
 231                 dev->callfds[i] = callfd;
 232                 dev->kickfds[i] = kickfd;
 233         }
 234
 235         if (i < VIRTIO_MAX_VIRTQUEUES) {
 236                 for (j = 0; j <= i; ++j) {
 237                         close(dev->callfds[j]);
 238                         close(dev->kickfds[j]);
 239                 }
 240
 241                 return -1;
 242         }
 243
 244         return 0;
 245 }
 246
 247 static int
 248 virtio_user_fill_intr_handle(struct virtio_user_dev *dev)
 249 {
 250         uint32_t i;
 251         struct rte_eth_dev *eth_dev = &rte_eth_devices[dev->port_id];
 252
 253         if (!eth_dev->intr_handle) {
 254                 eth_dev->intr_handle = malloc(sizeof(*eth_dev->intr_handle));
 255                 if (!eth_dev->intr_handle) {
 256                         PMD_DRV_LOG(ERR, "fail to allocate intr_handle");
 257                         return -1;
 258                 }
 259                 memset(eth_dev->intr_handle, 0, sizeof(*eth_dev->intr_handle));
 260         }
 261
 262         for (i = 0; i < dev->max_queue_pairs; ++i)
 263                 eth_dev->intr_handle->efds[i] = dev->callfds[i];
 264         eth_dev->intr_handle->nb_efd = dev->max_queue_pairs;
 265         eth_dev->intr_handle->max_intr = dev->max_queue_pairs + 1;
 266         eth_dev->intr_handle->type = RTE_INTR_HANDLE_VDEV;
 267         /* For virtio vdev, no need to read counter for clean */
 268         eth_dev->intr_handle->efd_counter_size = 0;
 269         eth_dev->intr_handle->fd = -1;
 270         if (dev->vhostfd >= 0)
 271                 eth_dev->intr_handle->fd = dev->vhostfd;
 272         else if (dev->is_server)
 273                 eth_dev->intr_handle->fd = dev->listenfd;
 274
 275         return 0;
 276 }
 277
 278 static void
 279 virtio_user_mem_event_cb(enum rte_mem_event type __rte_unused,
 280                                                  const void *addr __rte_unused,
 281                                                  size_t len __rte_unused,
 282                                                  void *arg)
 283 {
 284         struct virtio_user_dev *dev = arg;
 285         uint16_t i;
 286
 287         pthread_mutex_lock(&dev->mutex);
 288
 289         if (dev->started == false)
 290                 goto exit;
 291
 292         /* Step 1: pause the active queues */
 293         for (i = 0; i < dev->queue_pairs; i++)
 294                 dev->ops->enable_qp(dev, i, 0);
 295
 296         /* Step 2: update memory regions */
 297         dev->ops->send_request(dev, VHOST_USER_SET_MEM_TABLE, NULL);
 298
 299         /* Step 3: resume the active queues */
 300         for (i = 0; i < dev->queue_pairs; i++)
 301                 dev->ops->enable_qp(dev, i, 1);
 302
 303 exit:
 304         pthread_mutex_unlock(&dev->mutex);
 305 }
 306
 307 static int
 308 virtio_user_dev_setup(struct virtio_user_dev *dev)
 309 {
 310         uint32_t q;
 311
 312         dev->vhostfd = -1;
 313         dev->vhostfds = NULL;
 314         dev->tapfds = NULL;
 315
 316         if (dev->is_server) {
 317                 if (access(dev->path, F_OK) == 0 &&
 318                     !is_vhost_user_by_type(dev->path)) {
 319                         PMD_DRV_LOG(ERR, "Server mode doesn't support vhost-kernel!");
 320                         return -1;
 321                 }
 322                 dev->ops = &ops_user;
 323         } else {
 324                 if (is_vhost_user_by_type(dev->path)) {
 325                         dev->ops = &ops_user;
 326                 } else {
 327                         dev->ops = &ops_kernel;
 328
 329                         dev->vhostfds = malloc(dev->max_queue_pairs *
 330                                                sizeof(int));
 331                         dev->tapfds = malloc(dev->max_queue_pairs *
 332                                              sizeof(int));
 333                         if (!dev->vhostfds || !dev->tapfds) {
 334                                 PMD_INIT_LOG(ERR, "Failed to malloc");
 335                                 return -1;
 336                         }
 337
 338                         for (q = 0; q < dev->max_queue_pairs; ++q) {
 339                                 dev->vhostfds[q] = -1;
 340                                 dev->tapfds[q] = -1;
 341                         }
 342                 }
 343         }
 344
 345         if (dev->ops->setup(dev) < 0)
 346                 return -1;
 347
 348         if (virtio_user_dev_init_notify(dev) < 0)
 349                 return -1;
 350
 351         if (virtio_user_fill_intr_handle(dev) < 0)
 352                 return -1;
 353
 354         return 0;
 355 }
 356
 357 /* Use below macro to filter features from vhost backend */
 358 #define VIRTIO_USER_SUPPORTED_FEATURES                  \
 359         (1ULL << VIRTIO_NET_F_MAC               |       \
 360          1ULL << VIRTIO_NET_F_STATUS            |       \
 361          1ULL << VIRTIO_NET_F_MQ                |       \
 362          1ULL << VIRTIO_NET_F_CTRL_MAC_ADDR     |       \
 363          1ULL << VIRTIO_NET_F_CTRL_VQ           |       \
 364          1ULL << VIRTIO_NET_F_CTRL_RX           |       \
 365          1ULL << VIRTIO_NET_F_CTRL_VLAN         |       \
 366          1ULL << VIRTIO_NET_F_CSUM              |       \
 367          1ULL << VIRTIO_NET_F_HOST_TSO4         |       \
 368          1ULL << VIRTIO_NET_F_HOST_TSO6         |       \
 369          1ULL << VIRTIO_NET_F_MRG_RXBUF         |       \
 370          1ULL << VIRTIO_RING_F_INDIRECT_DESC    |       \
 371          1ULL << VIRTIO_NET_F_GUEST_CSUM        |       \
 372          1ULL << VIRTIO_NET_F_GUEST_TSO4        |       \
 373          1ULL << VIRTIO_NET_F_GUEST_TSO6        |       \
 374          1ULL << VIRTIO_F_VERSION_1)
 375
 376 int
 377 virtio_user_dev_init(struct virtio_user_dev *dev, char *path, int queues,
 378                      int cq, int queue_size, const char *mac, char **ifname)
 379 {
 380         pthread_mutex_init(&dev->mutex, NULL);
 381         snprintf(dev->path, PATH_MAX, "%s", path);
 382         dev->started = 0;
 383         dev->max_queue_pairs = queues;
 384         dev->queue_pairs = 1; /* mq disabled by default */
 385         dev->queue_size = queue_size;
 386         dev->mac_specified = 0;
 387         parse_mac(dev, mac);
 388
 389         if (*ifname) {
 390                 dev->ifname = *ifname;
 391                 *ifname = NULL;
 392         }
 393
 394         if (virtio_user_dev_setup(dev) < 0) {
 395                 PMD_INIT_LOG(ERR, "backend set up fails");
 396                 return -1;
 397         }
 398
 399         if (dev->vhostfd >= 0) {
 400                 if (dev->ops->send_request(dev, VHOST_USER_SET_OWNER,
 401                                            NULL) < 0) {
 402                         PMD_INIT_LOG(ERR, "set_owner fails: %s",
 403                                      strerror(errno));
 404                         return -1;
 405                 }
 406
 407                 if (dev->ops->send_request(dev, VHOST_USER_GET_FEATURES,
 408                                            &dev->device_features) < 0) {
 409                         PMD_INIT_LOG(ERR, "get_features failed: %s",
 410                                      strerror(errno));
 411                         return -1;
 412                 }
 413         } else {
 414                 /* We just pretend vhost-user can support all these features.
 415                  * Note that this could be problematic that if some feature is
 416                  * negotiated but not supported by the vhost-user which comes
 417                  * later.
 418                  */
 419                 dev->device_features = VIRTIO_USER_SUPPORTED_FEATURES;
 420         }
 421
 422         if (dev->mac_specified)
 423                 dev->device_features |= (1ull << VIRTIO_NET_F_MAC);
 424
 425         if (cq) {
 426                 /* device does not really need to know anything about CQ,
 427                  * so if necessary, we just claim to support CQ
 428                  */
 429                 dev->device_features |= (1ull << VIRTIO_NET_F_CTRL_VQ);
 430         } else {
 431                 dev->device_features &= ~(1ull << VIRTIO_NET_F_CTRL_VQ);
 432                 /* Also disable features depends on VIRTIO_NET_F_CTRL_VQ */
 433                 dev->device_features &= ~(1ull << VIRTIO_NET_F_CTRL_RX);
 434                 dev->device_features &= ~(1ull << VIRTIO_NET_F_CTRL_VLAN);
 435                 dev->device_features &= ~(1ull << VIRTIO_NET_F_GUEST_ANNOUNCE);
 436                 dev->device_features &= ~(1ull << VIRTIO_NET_F_MQ);
 437                 dev->device_features &= ~(1ull << VIRTIO_NET_F_CTRL_MAC_ADDR);
 438         }
 439
 440         /* The backend will not report this feature, we add it explicitly */
 441         if (is_vhost_user_by_type(dev->path))
 442                 dev->device_features |= (1ull << VIRTIO_NET_F_STATUS);
 443
 444         dev->device_features &= VIRTIO_USER_SUPPORTED_FEATURES;
 445
 446         if (rte_mem_event_callback_register(VIRTIO_USER_MEM_EVENT_CLB_NAME,
 447                                 virtio_user_mem_event_cb, dev)) {
 448                 PMD_INIT_LOG(ERR, "Failed to register mem event callback\n");
 449                 return -1;
 450         }
 451
 452         return 0;
 453 }
 454
 455 void
 456 virtio_user_dev_uninit(struct virtio_user_dev *dev)
 457 {
 458         uint32_t i;
 459
 460         virtio_user_stop_device(dev);
 461
 462         rte_mem_event_callback_unregister(VIRTIO_USER_MEM_EVENT_CLB_NAME, dev);
 463
 464         for (i = 0; i < dev->max_queue_pairs * 2; ++i) {
 465                 close(dev->callfds[i]);
 466                 close(dev->kickfds[i]);
 467         }
 468
 469         close(dev->vhostfd);
 470
 471         if (dev->is_server && dev->listenfd >= 0) {
 472                 close(dev->listenfd);
 473                 dev->listenfd = -1;
 474         }
 475
 476         if (dev->vhostfds) {
 477                 for (i = 0; i < dev->max_queue_pairs; ++i)
 478                         close(dev->vhostfds[i]);
 479                 free(dev->vhostfds);
 480                 free(dev->tapfds);
 481         }
 482
 483         free(dev->ifname);
 484
 485         if (dev->is_server)
 486                 unlink(dev->path);
 487 }
 488
 489 uint8_t
 490 virtio_user_handle_mq(struct virtio_user_dev *dev, uint16_t q_pairs)
 491 {
 492         uint16_t i;
 493         uint8_t ret = 0;
 494
 495         if (q_pairs > dev->max_queue_pairs) {
 496                 PMD_INIT_LOG(ERR, "multi-q config %u, but only %u supported",
 497                              q_pairs, dev->max_queue_pairs);
 498                 return -1;
 499         }
 500
 501         /* Server mode can't enable queue pairs if vhostfd is invalid,
 502          * always return 0 in this case.
 503          */
 504         if (dev->vhostfd >= 0) {
 505                 for (i = 0; i < q_pairs; ++i)
 506                         ret |= dev->ops->enable_qp(dev, i, 1);
 507                 for (i = q_pairs; i < dev->max_queue_pairs; ++i)
 508                         ret |= dev->ops->enable_qp(dev, i, 0);
 509         } else if (!dev->is_server) {
 510                 ret = ~0;
 511         }
 512         dev->queue_pairs = q_pairs;
 513
 514         return ret;
 515 }
 516
 517 static uint32_t
 518 virtio_user_handle_ctrl_msg(struct virtio_user_dev *dev, struct vring *vring,
 519                             uint16_t idx_hdr)
 520 {
 521         struct virtio_net_ctrl_hdr *hdr;
 522         virtio_net_ctrl_ack status = ~0;
 523         uint16_t i, idx_data, idx_status;
 524         uint32_t n_descs = 0;
 525
 526         /* locate desc for header, data, and status */
 527         idx_data = vring->desc[idx_hdr].next;
 528         n_descs++;
 529
 530         i = idx_data;
 531         while (vring->desc[i].flags == VRING_DESC_F_NEXT) {
 532                 i = vring->desc[i].next;
 533                 n_descs++;
 534         }
 535
 536         /* locate desc for status */
 537         idx_status = i;
 538         n_descs++;
 539
 540         hdr = (void *)(uintptr_t)vring->desc[idx_hdr].addr;
 541         if (hdr->class == VIRTIO_NET_CTRL_MQ &&
 542             hdr->cmd == VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET) {
 543                 uint16_t queues;
 544
 545                 queues = *(uint16_t *)(uintptr_t)vring->desc[idx_data].addr;
 546                 status = virtio_user_handle_mq(dev, queues);
 547         }
 548
 549         /* Update status */
 550         *(virtio_net_ctrl_ack *)(uintptr_t)vring->desc[idx_status].addr = status;
 551
 552         return n_descs;
 553 }
 554
 555 void
 556 virtio_user_handle_cq(struct virtio_user_dev *dev, uint16_t queue_idx)
 557 {
 558         uint16_t avail_idx, desc_idx;
 559         struct vring_used_elem *uep;
 560         uint32_t n_descs;
 561         struct vring *vring = &dev->vrings[queue_idx];
 562
 563         /* Consume avail ring, using used ring idx as first one */
 564         while (vring->used->idx != vring->avail->idx) {
 565                 avail_idx = (vring->used->idx) & (vring->num - 1);
 566                 desc_idx = vring->avail->ring[avail_idx];
 567
 568                 n_descs = virtio_user_handle_ctrl_msg(dev, vring, desc_idx);
 569
 570                 /* Update used ring */
 571                 uep = &vring->used->ring[avail_idx];
 572                 uep->id = avail_idx;
 573                 uep->len = n_descs;
 574
 575                 vring->used->idx++;
 576         }
 577 }