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 <rte_string_fns.h>
  17 #include <rte_eal_memconfig.h>
  18
  19 #include "vhost.h"
  20 #include "virtio_user_dev.h"
  21 #include "../virtio_ethdev.h"
  22
  23 #define VIRTIO_USER_MEM_EVENT_CLB_NAME "virtio_user_mem_event_clb"
  24
  25 const char * const virtio_user_backend_strings[] = {
  26         [VIRTIO_USER_BACKEND_UNKNOWN] = "VIRTIO_USER_BACKEND_UNKNOWN",
  27         [VIRTIO_USER_BACKEND_VHOST_USER] = "VHOST_USER",
  28         [VIRTIO_USER_BACKEND_VHOST_KERNEL] = "VHOST_NET",
  29         [VIRTIO_USER_BACKEND_VHOST_VDPA] = "VHOST_VDPA",
  30 };
  31
  32 static int
  33 virtio_user_create_queue(struct virtio_user_dev *dev, uint32_t queue_sel)
  34 {
  35         /* Of all per virtqueue MSGs, make sure VHOST_SET_VRING_CALL come
  36          * firstly because vhost depends on this msg to allocate virtqueue
  37          * pair.
  38          */
  39         struct vhost_vring_file file;
  40         int ret;
  41
  42         file.index = queue_sel;
  43         file.fd = dev->callfds[queue_sel];
  44         ret = dev->ops->set_vring_call(dev, &file);
  45         if (ret < 0) {
  46                 PMD_INIT_LOG(ERR, "(%s) Failed to create queue %u\n", dev->path, queue_sel);
  47                 return -1;
  48         }
  49
  50         return 0;
  51 }
  52
  53 static int
  54 virtio_user_kick_queue(struct virtio_user_dev *dev, uint32_t queue_sel)
  55 {
  56         int ret;
  57         struct vhost_vring_file file;
  58         struct vhost_vring_state state;
  59         struct vring *vring = &dev->vrings[queue_sel];
  60         struct vring_packed *pq_vring = &dev->packed_vrings[queue_sel];
  61         struct vhost_vring_addr addr = {
  62                 .index = queue_sel,
  63                 .log_guest_addr = 0,
  64                 .flags = 0, /* disable log */
  65         };
  66
  67         if (dev->features & (1ULL << VIRTIO_F_RING_PACKED)) {
  68                 addr.desc_user_addr =
  69                         (uint64_t)(uintptr_t)pq_vring->desc;
  70                 addr.avail_user_addr =
  71                         (uint64_t)(uintptr_t)pq_vring->driver;
  72                 addr.used_user_addr =
  73                         (uint64_t)(uintptr_t)pq_vring->device;
  74         } else {
  75                 addr.desc_user_addr = (uint64_t)(uintptr_t)vring->desc;
  76                 addr.avail_user_addr = (uint64_t)(uintptr_t)vring->avail;
  77                 addr.used_user_addr = (uint64_t)(uintptr_t)vring->used;
  78         }
  79
  80         state.index = queue_sel;
  81         state.num = vring->num;
  82         ret = dev->ops->set_vring_num(dev, &state);
  83         if (ret < 0)
  84                 goto err;
  85
  86         state.index = queue_sel;
  87         state.num = 0; /* no reservation */
  88         if (dev->features & (1ULL << VIRTIO_F_RING_PACKED))
  89                 state.num |= (1 << 15);
  90         ret = dev->ops->set_vring_base(dev, &state);
  91         if (ret < 0)
  92                 goto err;
  93
  94         ret = dev->ops->set_vring_addr(dev, &addr);
  95         if (ret < 0)
  96                 goto err;
  97
  98         /* Of all per virtqueue MSGs, make sure VHOST_USER_SET_VRING_KICK comes
  99          * lastly because vhost depends on this msg to judge if
 100          * virtio is ready.
 101          */
 102         file.index = queue_sel;
 103         file.fd = dev->kickfds[queue_sel];
 104         ret = dev->ops->set_vring_kick(dev, &file);
 105         if (ret < 0)
 106                 goto err;
 107
 108         return 0;
 109 err:
 110         PMD_INIT_LOG(ERR, "(%s) Failed to kick queue %u\n", dev->path, queue_sel);
 111
 112         return -1;
 113 }
 114
 115 static int
 116 virtio_user_queue_setup(struct virtio_user_dev *dev,
 117                         int (*fn)(struct virtio_user_dev *, uint32_t))
 118 {
 119         uint32_t i, queue_sel;
 120
 121         for (i = 0; i < dev->max_queue_pairs; ++i) {
 122                 queue_sel = 2 * i + VTNET_SQ_RQ_QUEUE_IDX;
 123                 if (fn(dev, queue_sel) < 0) {
 124                         PMD_DRV_LOG(ERR, "(%s) setup rx vq %u failed", dev->path, i);
 125                         return -1;
 126                 }
 127         }
 128         for (i = 0; i < dev->max_queue_pairs; ++i) {
 129                 queue_sel = 2 * i + VTNET_SQ_TQ_QUEUE_IDX;
 130                 if (fn(dev, queue_sel) < 0) {
 131                         PMD_DRV_LOG(INFO, "(%s) setup tx vq %u failed", dev->path, i);
 132                         return -1;
 133                 }
 134         }
 135
 136         return 0;
 137 }
 138
 139 int
 140 virtio_user_dev_set_features(struct virtio_user_dev *dev)
 141 {
 142         uint64_t features;
 143         int ret = -1;
 144
 145         pthread_mutex_lock(&dev->mutex);
 146
 147         /* Step 0: tell vhost to create queues */
 148         if (virtio_user_queue_setup(dev, virtio_user_create_queue) < 0)
 149                 goto error;
 150
 151         features = dev->features;
 152
 153         /* Strip VIRTIO_NET_F_MAC, as MAC address is handled in vdev init */
 154         features &= ~(1ull << VIRTIO_NET_F_MAC);
 155         /* Strip VIRTIO_NET_F_CTRL_VQ, as devices do not really need to know */
 156         features &= ~(1ull << VIRTIO_NET_F_CTRL_VQ);
 157         features &= ~(1ull << VIRTIO_NET_F_STATUS);
 158         ret = dev->ops->set_features(dev, features);
 159         if (ret < 0)
 160                 goto error;
 161         PMD_DRV_LOG(INFO, "(%s) set features: 0x%" PRIx64, dev->path, features);
 162 error:
 163         pthread_mutex_unlock(&dev->mutex);
 164
 165         return ret;
 166 }
 167
 168 int
 169 virtio_user_start_device(struct virtio_user_dev *dev)
 170 {
 171         int ret;
 172
 173         /*
 174          * XXX workaround!
 175          *
 176          * We need to make sure that the locks will be
 177          * taken in the correct order to avoid deadlocks.
 178          *
 179          * Before releasing this lock, this thread should
 180          * not trigger any memory hotplug events.
 181          *
 182          * This is a temporary workaround, and should be
 183          * replaced when we get proper supports from the
 184          * memory subsystem in the future.
 185          */
 186         rte_mcfg_mem_read_lock();
 187         pthread_mutex_lock(&dev->mutex);
 188
 189         /* Step 2: share memory regions */
 190         ret = dev->ops->set_memory_table(dev);
 191         if (ret < 0)
 192                 goto error;
 193
 194         /* Step 3: kick queues */
 195         ret = virtio_user_queue_setup(dev, virtio_user_kick_queue);
 196         if (ret < 0)
 197                 goto error;
 198
 199         /* Step 4: enable queues
 200          * we enable the 1st queue pair by default.
 201          */
 202         ret = dev->ops->enable_qp(dev, 0, 1);
 203         if (ret < 0)
 204                 goto error;
 205
 206         dev->started = true;
 207
 208         pthread_mutex_unlock(&dev->mutex);
 209         rte_mcfg_mem_read_unlock();
 210
 211         return 0;
 212 error:
 213         pthread_mutex_unlock(&dev->mutex);
 214         rte_mcfg_mem_read_unlock();
 215
 216         PMD_INIT_LOG(ERR, "(%s) Failed to start device\n", dev->path);
 217
 218         /* TODO: free resource here or caller to check */
 219         return -1;
 220 }
 221
 222 int virtio_user_stop_device(struct virtio_user_dev *dev)
 223 {
 224         struct vhost_vring_state state;
 225         uint32_t i;
 226         int ret;
 227
 228         pthread_mutex_lock(&dev->mutex);
 229         if (!dev->started)
 230                 goto out;
 231
 232         for (i = 0; i < dev->max_queue_pairs; ++i) {
 233                 ret = dev->ops->enable_qp(dev, i, 0);
 234                 if (ret < 0)
 235                         goto err;
 236         }
 237
 238         /* Stop the backend. */
 239         for (i = 0; i < dev->max_queue_pairs * 2; ++i) {
 240                 state.index = i;
 241                 ret = dev->ops->get_vring_base(dev, &state);
 242                 if (ret < 0) {
 243                         PMD_DRV_LOG(ERR, "(%s) get_vring_base failed, index=%u", dev->path, i);
 244                         goto err;
 245                 }
 246         }
 247
 248         dev->started = false;
 249
 250 out:
 251         pthread_mutex_unlock(&dev->mutex);
 252
 253         return 0;
 254 err:
 255         pthread_mutex_unlock(&dev->mutex);
 256
 257         PMD_INIT_LOG(ERR, "(%s) Failed to stop device\n", dev->path);
 258
 259         return -1;
 260 }
 261
 262 static inline void
 263 parse_mac(struct virtio_user_dev *dev, const char *mac)
 264 {
 265         struct rte_ether_addr tmp;
 266
 267         if (!mac)
 268                 return;
 269
 270         if (rte_ether_unformat_addr(mac, &tmp) == 0) {
 271                 memcpy(dev->mac_addr, &tmp, RTE_ETHER_ADDR_LEN);
 272                 dev->mac_specified = 1;
 273         } else {
 274                 /* ignore the wrong mac, use random mac */
 275                 PMD_DRV_LOG(ERR, "wrong format of mac: %s", mac);
 276         }
 277 }
 278
 279 static int
 280 virtio_user_dev_init_notify(struct virtio_user_dev *dev)
 281 {
 282         uint32_t i, j;
 283         int callfd;
 284         int kickfd;
 285
 286         for (i = 0; i < VIRTIO_MAX_VIRTQUEUES; ++i) {
 287                 if (i >= dev->max_queue_pairs * 2) {
 288                         dev->kickfds[i] = -1;
 289                         dev->callfds[i] = -1;
 290                         continue;
 291                 }
 292
 293                 /* May use invalid flag, but some backend uses kickfd and
 294                  * callfd as criteria to judge if dev is alive. so finally we
 295                  * use real event_fd.
 296                  */
 297                 callfd = eventfd(0, EFD_CLOEXEC | EFD_NONBLOCK);
 298                 if (callfd < 0) {
 299                         PMD_DRV_LOG(ERR, "(%s) callfd error, %s", dev->path, strerror(errno));
 300                         break;
 301                 }
 302                 kickfd = eventfd(0, EFD_CLOEXEC | EFD_NONBLOCK);
 303                 if (kickfd < 0) {
 304                         close(callfd);
 305                         PMD_DRV_LOG(ERR, "(%s) kickfd error, %s", dev->path, strerror(errno));
 306                         break;
 307                 }
 308                 dev->callfds[i] = callfd;
 309                 dev->kickfds[i] = kickfd;
 310         }
 311
 312         if (i < VIRTIO_MAX_VIRTQUEUES) {
 313                 for (j = 0; j < i; ++j) {
 314                         close(dev->callfds[j]);
 315                         close(dev->kickfds[j]);
 316                 }
 317
 318                 return -1;
 319         }
 320
 321         return 0;
 322 }
 323
 324 static int
 325 virtio_user_fill_intr_handle(struct virtio_user_dev *dev)
 326 {
 327         uint32_t i;
 328         struct rte_eth_dev *eth_dev = &rte_eth_devices[dev->port_id];
 329
 330         if (!eth_dev->intr_handle) {
 331                 eth_dev->intr_handle = malloc(sizeof(*eth_dev->intr_handle));
 332                 if (!eth_dev->intr_handle) {
 333                         PMD_DRV_LOG(ERR, "(%s) failed to allocate intr_handle", dev->path);
 334                         return -1;
 335                 }
 336                 memset(eth_dev->intr_handle, 0, sizeof(*eth_dev->intr_handle));
 337         }
 338
 339         for (i = 0; i < dev->max_queue_pairs; ++i)
 340                 eth_dev->intr_handle->efds[i] = dev->callfds[i];
 341         eth_dev->intr_handle->nb_efd = dev->max_queue_pairs;
 342         eth_dev->intr_handle->max_intr = dev->max_queue_pairs + 1;
 343         eth_dev->intr_handle->type = RTE_INTR_HANDLE_VDEV;
 344         /* For virtio vdev, no need to read counter for clean */
 345         eth_dev->intr_handle->efd_counter_size = 0;
 346         eth_dev->intr_handle->fd = dev->ops->get_intr_fd(dev);
 347
 348         return 0;
 349 }
 350
 351 static void
 352 virtio_user_mem_event_cb(enum rte_mem_event type __rte_unused,
 353                          const void *addr,
 354                          size_t len __rte_unused,
 355                          void *arg)
 356 {
 357         struct virtio_user_dev *dev = arg;
 358         struct rte_memseg_list *msl;
 359         uint16_t i;
 360         int ret = 0;
 361
 362         /* ignore externally allocated memory */
 363         msl = rte_mem_virt2memseg_list(addr);
 364         if (msl->external)
 365                 return;
 366
 367         pthread_mutex_lock(&dev->mutex);
 368
 369         if (dev->started == false)
 370                 goto exit;
 371
 372         /* Step 1: pause the active queues */
 373         for (i = 0; i < dev->queue_pairs; i++) {
 374                 ret = dev->ops->enable_qp(dev, i, 0);
 375                 if (ret < 0)
 376                         goto exit;
 377         }
 378
 379         /* Step 2: update memory regions */
 380         ret = dev->ops->set_memory_table(dev);
 381         if (ret < 0)
 382                 goto exit;
 383
 384         /* Step 3: resume the active queues */
 385         for (i = 0; i < dev->queue_pairs; i++) {
 386                 ret = dev->ops->enable_qp(dev, i, 1);
 387                 if (ret < 0)
 388                         goto exit;
 389         }
 390
 391 exit:
 392         pthread_mutex_unlock(&dev->mutex);
 393
 394         if (ret < 0)
 395                 PMD_DRV_LOG(ERR, "(%s) Failed to update memory table\n", dev->path);
 396 }
 397
 398 static int
 399 virtio_user_dev_setup(struct virtio_user_dev *dev)
 400 {
 401         uint32_t q;
 402
 403         dev->vhostfds = NULL;
 404         dev->tapfds = NULL;
 405
 406         if (dev->is_server) {
 407                 if (dev->backend_type != VIRTIO_USER_BACKEND_VHOST_USER) {
 408                         PMD_DRV_LOG(ERR, "Server mode only supports vhost-user!");
 409                         return -1;
 410                 }
 411         }
 412
 413         if (dev->backend_type == VIRTIO_USER_BACKEND_VHOST_USER) {
 414                 dev->ops = &virtio_ops_user;
 415         } else if (dev->backend_type ==
 416                         VIRTIO_USER_BACKEND_VHOST_KERNEL) {
 417                 dev->ops = &virtio_ops_kernel;
 418
 419                 dev->vhostfds = malloc(dev->max_queue_pairs *
 420                                 sizeof(int));
 421                 dev->tapfds = malloc(dev->max_queue_pairs *
 422                                 sizeof(int));
 423                 if (!dev->vhostfds || !dev->tapfds) {
 424                         PMD_INIT_LOG(ERR, "(%s) Failed to allocate FDs", dev->path);
 425                         return -1;
 426                 }
 427
 428                 for (q = 0; q < dev->max_queue_pairs; ++q) {
 429                         dev->vhostfds[q] = -1;
 430                         dev->tapfds[q] = -1;
 431                 }
 432         } else if (dev->backend_type ==
 433                         VIRTIO_USER_BACKEND_VHOST_VDPA) {
 434                 dev->ops = &virtio_ops_vdpa;
 435         } else {
 436                 PMD_DRV_LOG(ERR, "(%s) Unknown backend type", dev->path);
 437                 return -1;
 438         }
 439
 440
 441         if (dev->ops->setup(dev) < 0) {
 442                 PMD_INIT_LOG(ERR, "(%s) Failed to setup backend\n", dev->path);
 443                 return -1;
 444         }
 445
 446         if (virtio_user_dev_init_notify(dev) < 0) {
 447                 PMD_INIT_LOG(ERR, "(%s) Failed to init notifiers\n", dev->path);
 448                 return -1;
 449         }
 450
 451         if (virtio_user_fill_intr_handle(dev) < 0) {
 452                 PMD_INIT_LOG(ERR, "(%s) Failed to init interrupt handler\n", dev->path);
 453                 return -1;
 454         }
 455
 456         return 0;
 457 }
 458
 459 /* Use below macro to filter features from vhost backend */
 460 #define VIRTIO_USER_SUPPORTED_FEATURES                  \
 461         (1ULL << VIRTIO_NET_F_MAC               |       \
 462          1ULL << VIRTIO_NET_F_STATUS            |       \
 463          1ULL << VIRTIO_NET_F_MQ                |       \
 464          1ULL << VIRTIO_NET_F_CTRL_MAC_ADDR     |       \
 465          1ULL << VIRTIO_NET_F_CTRL_VQ           |       \
 466          1ULL << VIRTIO_NET_F_CTRL_RX           |       \
 467          1ULL << VIRTIO_NET_F_CTRL_VLAN         |       \
 468          1ULL << VIRTIO_NET_F_CSUM              |       \
 469          1ULL << VIRTIO_NET_F_HOST_TSO4         |       \
 470          1ULL << VIRTIO_NET_F_HOST_TSO6         |       \
 471          1ULL << VIRTIO_NET_F_MRG_RXBUF         |       \
 472          1ULL << VIRTIO_RING_F_INDIRECT_DESC    |       \
 473          1ULL << VIRTIO_NET_F_GUEST_CSUM        |       \
 474          1ULL << VIRTIO_NET_F_GUEST_TSO4        |       \
 475          1ULL << VIRTIO_NET_F_GUEST_TSO6        |       \
 476          1ULL << VIRTIO_F_IN_ORDER              |       \
 477          1ULL << VIRTIO_F_VERSION_1             |       \
 478          1ULL << VIRTIO_F_RING_PACKED)
 479
 480 int
 481 virtio_user_dev_init(struct virtio_user_dev *dev, char *path, int queues,
 482                      int cq, int queue_size, const char *mac, char **ifname,
 483                      int server, int mrg_rxbuf, int in_order, int packed_vq,
 484                      enum virtio_user_backend_type backend_type)
 485 {
 486         uint64_t backend_features;
 487
 488         pthread_mutex_init(&dev->mutex, NULL);
 489         strlcpy(dev->path, path, PATH_MAX);
 490         dev->started = 0;
 491         dev->max_queue_pairs = queues;
 492         dev->queue_pairs = 1; /* mq disabled by default */
 493         dev->queue_size = queue_size;
 494         dev->is_server = server;
 495         dev->mac_specified = 0;
 496         dev->frontend_features = 0;
 497         dev->unsupported_features = 0;
 498         dev->backend_type = backend_type;
 499
 500         parse_mac(dev, mac);
 501
 502         if (*ifname) {
 503                 dev->ifname = *ifname;
 504                 *ifname = NULL;
 505         }
 506
 507         if (virtio_user_dev_setup(dev) < 0) {
 508                 PMD_INIT_LOG(ERR, "(%s) backend set up fails", dev->path);
 509                 return -1;
 510         }
 511
 512         if (dev->ops->set_owner(dev) < 0) {
 513                 PMD_INIT_LOG(ERR, "(%s) Failed to set backend owner", dev->path);
 514                 return -1;
 515         }
 516
 517         if (dev->ops->get_backend_features(&backend_features) < 0) {
 518                 PMD_INIT_LOG(ERR, "(%s) Failed to get backend features", dev->path);
 519                 return -1;
 520         }
 521
 522         dev->unsupported_features = ~(VIRTIO_USER_SUPPORTED_FEATURES | backend_features);
 523
 524         if (dev->ops->get_features(dev, &dev->device_features) < 0) {
 525                 PMD_INIT_LOG(ERR, "(%s) Failed to get device features", dev->path);
 526                 return -1;
 527         }
 528
 529         if (!mrg_rxbuf)
 530                 dev->unsupported_features |= (1ull << VIRTIO_NET_F_MRG_RXBUF);
 531
 532         if (!in_order)
 533                 dev->unsupported_features |= (1ull << VIRTIO_F_IN_ORDER);
 534
 535         if (!packed_vq)
 536                 dev->unsupported_features |= (1ull << VIRTIO_F_RING_PACKED);
 537
 538         if (dev->mac_specified)
 539                 dev->frontend_features |= (1ull << VIRTIO_NET_F_MAC);
 540         else
 541                 dev->unsupported_features |= (1ull << VIRTIO_NET_F_MAC);
 542
 543         if (cq) {
 544                 /* device does not really need to know anything about CQ,
 545                  * so if necessary, we just claim to support CQ
 546                  */
 547                 dev->frontend_features |= (1ull << VIRTIO_NET_F_CTRL_VQ);
 548         } else {
 549                 dev->unsupported_features |= (1ull << VIRTIO_NET_F_CTRL_VQ);
 550                 /* Also disable features that depend on VIRTIO_NET_F_CTRL_VQ */
 551                 dev->unsupported_features |= (1ull << VIRTIO_NET_F_CTRL_RX);
 552                 dev->unsupported_features |= (1ull << VIRTIO_NET_F_CTRL_VLAN);
 553                 dev->unsupported_features |=
 554                         (1ull << VIRTIO_NET_F_GUEST_ANNOUNCE);
 555                 dev->unsupported_features |= (1ull << VIRTIO_NET_F_MQ);
 556                 dev->unsupported_features |=
 557                         (1ull << VIRTIO_NET_F_CTRL_MAC_ADDR);
 558         }
 559
 560         /* The backend will not report this feature, we add it explicitly */
 561         if (dev->backend_type == VIRTIO_USER_BACKEND_VHOST_USER)
 562                 dev->frontend_features |= (1ull << VIRTIO_NET_F_STATUS);
 563
 564         /*
 565          * Device features =
 566          *     (frontend_features | backend_features) & ~unsupported_features;
 567          */
 568         dev->device_features |= dev->frontend_features;
 569         dev->device_features &= ~dev->unsupported_features;
 570
 571         if (rte_mem_event_callback_register(VIRTIO_USER_MEM_EVENT_CLB_NAME,
 572                                 virtio_user_mem_event_cb, dev)) {
 573                 if (rte_errno != ENOTSUP) {
 574                         PMD_INIT_LOG(ERR, "(%s) Failed to register mem event callback\n",
 575                                         dev->path);
 576                         return -1;
 577                 }
 578         }
 579
 580         return 0;
 581 }
 582
 583 void
 584 virtio_user_dev_uninit(struct virtio_user_dev *dev)
 585 {
 586         uint32_t i;
 587
 588         virtio_user_stop_device(dev);
 589
 590         rte_mem_event_callback_unregister(VIRTIO_USER_MEM_EVENT_CLB_NAME, dev);
 591
 592         for (i = 0; i < dev->max_queue_pairs * 2; ++i) {
 593                 close(dev->callfds[i]);
 594                 close(dev->kickfds[i]);
 595         }
 596         if (dev->vhostfds) {
 597                 for (i = 0; i < dev->max_queue_pairs; ++i) {
 598                         close(dev->vhostfds[i]);
 599                         if (dev->tapfds[i] >= 0)
 600                                 close(dev->tapfds[i]);
 601                 }
 602                 free(dev->vhostfds);
 603                 free(dev->tapfds);
 604         }
 605
 606         free(dev->ifname);
 607
 608         if (dev->is_server)
 609                 unlink(dev->path);
 610
 611         dev->ops->destroy(dev);
 612 }
 613
 614 uint8_t
 615 virtio_user_handle_mq(struct virtio_user_dev *dev, uint16_t q_pairs)
 616 {
 617         uint16_t i;
 618         uint8_t ret = 0;
 619
 620         if (q_pairs > dev->max_queue_pairs) {
 621                 PMD_INIT_LOG(ERR, "(%s) multi-q config %u, but only %u supported",
 622                              dev->path, q_pairs, dev->max_queue_pairs);
 623                 return -1;
 624         }
 625
 626         for (i = 0; i < q_pairs; ++i)
 627                 ret |= dev->ops->enable_qp(dev, i, 1);
 628         for (i = q_pairs; i < dev->max_queue_pairs; ++i)
 629                 ret |= dev->ops->enable_qp(dev, i, 0);
 630
 631         dev->queue_pairs = q_pairs;
 632
 633         return ret;
 634 }
 635
 636 static uint32_t
 637 virtio_user_handle_ctrl_msg(struct virtio_user_dev *dev, struct vring *vring,
 638                             uint16_t idx_hdr)
 639 {
 640         struct virtio_net_ctrl_hdr *hdr;
 641         virtio_net_ctrl_ack status = ~0;
 642         uint16_t i, idx_data, idx_status;
 643         uint32_t n_descs = 0;
 644
 645         /* locate desc for header, data, and status */
 646         idx_data = vring->desc[idx_hdr].next;
 647         n_descs++;
 648
 649         i = idx_data;
 650         while (vring->desc[i].flags == VRING_DESC_F_NEXT) {
 651                 i = vring->desc[i].next;
 652                 n_descs++;
 653         }
 654
 655         /* locate desc for status */
 656         idx_status = i;
 657         n_descs++;
 658
 659         hdr = (void *)(uintptr_t)vring->desc[idx_hdr].addr;
 660         if (hdr->class == VIRTIO_NET_CTRL_MQ &&
 661             hdr->cmd == VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET) {
 662                 uint16_t queues;
 663
 664                 queues = *(uint16_t *)(uintptr_t)vring->desc[idx_data].addr;
 665                 status = virtio_user_handle_mq(dev, queues);
 666         } else if (hdr->class == VIRTIO_NET_CTRL_RX  ||
 667                    hdr->class == VIRTIO_NET_CTRL_MAC ||
 668                    hdr->class == VIRTIO_NET_CTRL_VLAN) {
 669                 status = 0;
 670         }
 671
 672         /* Update status */
 673         *(virtio_net_ctrl_ack *)(uintptr_t)vring->desc[idx_status].addr = status;
 674
 675         return n_descs;
 676 }
 677
 678 static inline int
 679 desc_is_avail(struct vring_packed_desc *desc, bool wrap_counter)
 680 {
 681         uint16_t flags = __atomic_load_n(&desc->flags, __ATOMIC_ACQUIRE);
 682
 683         return wrap_counter == !!(flags & VRING_PACKED_DESC_F_AVAIL) &&
 684                 wrap_counter != !!(flags & VRING_PACKED_DESC_F_USED);
 685 }
 686
 687 static uint32_t
 688 virtio_user_handle_ctrl_msg_packed(struct virtio_user_dev *dev,
 689                                    struct vring_packed *vring,
 690                                    uint16_t idx_hdr)
 691 {
 692         struct virtio_net_ctrl_hdr *hdr;
 693         virtio_net_ctrl_ack status = ~0;
 694         uint16_t idx_data, idx_status;
 695         /* initialize to one, header is first */
 696         uint32_t n_descs = 1;
 697
 698         /* locate desc for header, data, and status */
 699         idx_data = idx_hdr + 1;
 700         if (idx_data >= dev->queue_size)
 701                 idx_data -= dev->queue_size;
 702
 703         n_descs++;
 704
 705         idx_status = idx_data;
 706         while (vring->desc[idx_status].flags & VRING_DESC_F_NEXT) {
 707                 idx_status++;
 708                 if (idx_status >= dev->queue_size)
 709                         idx_status -= dev->queue_size;
 710                 n_descs++;
 711         }
 712
 713         hdr = (void *)(uintptr_t)vring->desc[idx_hdr].addr;
 714         if (hdr->class == VIRTIO_NET_CTRL_MQ &&
 715             hdr->cmd == VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET) {
 716                 uint16_t queues;
 717
 718                 queues = *(uint16_t *)(uintptr_t)
 719                                 vring->desc[idx_data].addr;
 720                 status = virtio_user_handle_mq(dev, queues);
 721         } else if (hdr->class == VIRTIO_NET_CTRL_RX  ||
 722                    hdr->class == VIRTIO_NET_CTRL_MAC ||
 723                    hdr->class == VIRTIO_NET_CTRL_VLAN) {
 724                 status = 0;
 725         }
 726
 727         /* Update status */
 728         *(virtio_net_ctrl_ack *)(uintptr_t)
 729                 vring->desc[idx_status].addr = status;
 730
 731         /* Update used descriptor */
 732         vring->desc[idx_hdr].id = vring->desc[idx_status].id;
 733         vring->desc[idx_hdr].len = sizeof(status);
 734
 735         return n_descs;
 736 }
 737
 738 void
 739 virtio_user_handle_cq_packed(struct virtio_user_dev *dev, uint16_t queue_idx)
 740 {
 741         struct virtio_user_queue *vq = &dev->packed_queues[queue_idx];
 742         struct vring_packed *vring = &dev->packed_vrings[queue_idx];
 743         uint16_t n_descs, flags;
 744
 745         /* Perform a load-acquire barrier in desc_is_avail to
 746          * enforce the ordering between desc flags and desc
 747          * content.
 748          */
 749         while (desc_is_avail(&vring->desc[vq->used_idx],
 750                              vq->used_wrap_counter)) {
 751
 752                 n_descs = virtio_user_handle_ctrl_msg_packed(dev, vring,
 753                                 vq->used_idx);
 754
 755                 flags = VRING_DESC_F_WRITE;
 756                 if (vq->used_wrap_counter)
 757                         flags |= VRING_PACKED_DESC_F_AVAIL_USED;
 758
 759                 __atomic_store_n(&vring->desc[vq->used_idx].flags, flags,
 760                                  __ATOMIC_RELEASE);
 761
 762                 vq->used_idx += n_descs;
 763                 if (vq->used_idx >= dev->queue_size) {
 764                         vq->used_idx -= dev->queue_size;
 765                         vq->used_wrap_counter ^= 1;
 766                 }
 767         }
 768 }
 769
 770 void
 771 virtio_user_handle_cq(struct virtio_user_dev *dev, uint16_t queue_idx)
 772 {
 773         uint16_t avail_idx, desc_idx;
 774         struct vring_used_elem *uep;
 775         uint32_t n_descs;
 776         struct vring *vring = &dev->vrings[queue_idx];
 777
 778         /* Consume avail ring, using used ring idx as first one */
 779         while (__atomic_load_n(&vring->used->idx, __ATOMIC_RELAXED)
 780                != vring->avail->idx) {
 781                 avail_idx = __atomic_load_n(&vring->used->idx, __ATOMIC_RELAXED)
 782                             & (vring->num - 1);
 783                 desc_idx = vring->avail->ring[avail_idx];
 784
 785                 n_descs = virtio_user_handle_ctrl_msg(dev, vring, desc_idx);
 786
 787                 /* Update used ring */
 788                 uep = &vring->used->ring[avail_idx];
 789                 uep->id = desc_idx;
 790                 uep->len = n_descs;
 791
 792                 __atomic_add_fetch(&vring->used->idx, 1, __ATOMIC_RELAXED);
 793         }
 794 }
 795
 796 int
 797 virtio_user_dev_set_status(struct virtio_user_dev *dev, uint8_t status)
 798 {
 799         int ret;
 800
 801         pthread_mutex_lock(&dev->mutex);
 802         dev->status = status;
 803         ret = dev->ops->set_status(dev, status);
 804         if (ret && ret != -ENOTSUP)
 805                 PMD_INIT_LOG(ERR, "(%s) Failed to set backend status\n", dev->path);
 806
 807         pthread_mutex_unlock(&dev->mutex);
 808         return ret;
 809 }
 810
 811 int
 812 virtio_user_dev_update_status(struct virtio_user_dev *dev)
 813 {
 814         int ret;
 815         uint8_t status;
 816
 817         pthread_mutex_lock(&dev->mutex);
 818
 819         ret = dev->ops->get_status(dev, &status);
 820         if (!ret) {
 821                 dev->status = status;
 822                 PMD_INIT_LOG(DEBUG, "Updated Device Status(0x%08x):\n"
 823                         "\t-RESET: %u\n"
 824                         "\t-ACKNOWLEDGE: %u\n"
 825                         "\t-DRIVER: %u\n"
 826                         "\t-DRIVER_OK: %u\n"
 827                         "\t-FEATURES_OK: %u\n"
 828                         "\t-DEVICE_NEED_RESET: %u\n"
 829                         "\t-FAILED: %u\n",
 830                         dev->status,
 831                         (dev->status == VIRTIO_CONFIG_STATUS_RESET),
 832                         !!(dev->status & VIRTIO_CONFIG_STATUS_ACK),
 833                         !!(dev->status & VIRTIO_CONFIG_STATUS_DRIVER),
 834                         !!(dev->status & VIRTIO_CONFIG_STATUS_DRIVER_OK),
 835                         !!(dev->status & VIRTIO_CONFIG_STATUS_FEATURES_OK),
 836                         !!(dev->status & VIRTIO_CONFIG_STATUS_DEV_NEED_RESET),
 837                         !!(dev->status & VIRTIO_CONFIG_STATUS_FAILED));
 838         } else if (ret != -ENOTSUP) {
 839                 PMD_INIT_LOG(ERR, "(%s) Failed to get backend status\n", dev->path);
 840         }
 841
 842         pthread_mutex_unlock(&dev->mutex);
 843         return ret;
 844 }
 845
 846 int
 847 virtio_user_dev_update_link_state(struct virtio_user_dev *dev)
 848 {
 849         if (dev->ops->update_link_state)
 850                 return dev->ops->update_link_state(dev);
 851
 852         return 0;
 853 }
 854
 855 static void
 856 virtio_user_dev_reset_queues_packed(struct rte_eth_dev *eth_dev)
 857 {
 858         struct virtio_user_dev *dev = eth_dev->data->dev_private;
 859         struct virtio_hw *hw = &dev->hw;
 860         struct virtnet_rx *rxvq;
 861         struct virtnet_tx *txvq;
 862         uint16_t i;
 863
 864         /* Add lock to avoid queue contention. */
 865         rte_spinlock_lock(&hw->state_lock);
 866         hw->started = 0;
 867
 868         /*
 869          * Waiting for datapath to complete before resetting queues.
 870          * 1 ms should be enough for the ongoing Tx/Rx function to finish.
 871          */
 872         rte_delay_ms(1);
 873
 874         /* Vring reset for each Tx queue and Rx queue. */
 875         for (i = 0; i < eth_dev->data->nb_rx_queues; i++) {
 876                 rxvq = eth_dev->data->rx_queues[i];
 877                 virtqueue_rxvq_reset_packed(rxvq->vq);
 878                 virtio_dev_rx_queue_setup_finish(eth_dev, i);
 879         }
 880
 881         for (i = 0; i < eth_dev->data->nb_tx_queues; i++) {
 882                 txvq = eth_dev->data->tx_queues[i];
 883                 virtqueue_txvq_reset_packed(txvq->vq);
 884         }
 885
 886         hw->started = 1;
 887         rte_spinlock_unlock(&hw->state_lock);
 888 }
 889
 890 void
 891 virtio_user_dev_delayed_handler(void *param)
 892 {
 893         struct virtio_user_dev *dev = param;
 894         struct rte_eth_dev *eth_dev = &rte_eth_devices[dev->port_id];
 895
 896         if (rte_intr_disable(eth_dev->intr_handle) < 0) {
 897                 PMD_DRV_LOG(ERR, "interrupt disable failed");
 898                 return;
 899         }
 900         rte_intr_callback_unregister(eth_dev->intr_handle,
 901                                      virtio_interrupt_handler, eth_dev);
 902         if (dev->is_server) {
 903                 if (dev->ops->server_disconnect)
 904                         dev->ops->server_disconnect(dev);
 905                 eth_dev->intr_handle->fd = dev->ops->get_intr_fd(dev);
 906                 rte_intr_callback_register(eth_dev->intr_handle,
 907                                            virtio_interrupt_handler, eth_dev);
 908                 if (rte_intr_enable(eth_dev->intr_handle) < 0) {
 909                         PMD_DRV_LOG(ERR, "interrupt enable failed");
 910                         return;
 911                 }
 912         }
 913 }
 914
 915 int
 916 virtio_user_dev_server_reconnect(struct virtio_user_dev *dev)
 917 {
 918         int ret, old_status;
 919         struct rte_eth_dev *eth_dev = &rte_eth_devices[dev->port_id];
 920         struct virtio_hw *hw = &dev->hw;
 921
 922         if (!dev->ops->server_reconnect) {
 923                 PMD_DRV_LOG(ERR, "(%s) Missing server reconnect callback", dev->path);
 924                 return -1;
 925         }
 926
 927         if (dev->ops->server_reconnect(dev)) {
 928                 PMD_DRV_LOG(ERR, "(%s) Reconnect callback call failed", dev->path);
 929                 return -1;
 930         }
 931
 932         old_status = dev->status;
 933
 934         virtio_reset(hw);
 935
 936         virtio_set_status(hw, VIRTIO_CONFIG_STATUS_ACK);
 937
 938         virtio_set_status(hw, VIRTIO_CONFIG_STATUS_DRIVER);
 939
 940         if (dev->ops->get_features(dev, &dev->device_features) < 0) {
 941                 PMD_INIT_LOG(ERR, "get_features failed: %s",
 942                              strerror(errno));
 943                 return -1;
 944         }
 945
 946         dev->device_features |= dev->frontend_features;
 947
 948         /* unmask vhost-user unsupported features */
 949         dev->device_features &= ~(dev->unsupported_features);
 950
 951         dev->features &= dev->device_features;
 952
 953         /* For packed ring, resetting queues is required in reconnection. */
 954         if (virtio_with_packed_queue(hw) &&
 955            (old_status & VIRTIO_CONFIG_STATUS_DRIVER_OK)) {
 956                 PMD_INIT_LOG(NOTICE, "Packets on the fly will be dropped"
 957                                 " when packed ring reconnecting.");
 958                 virtio_user_dev_reset_queues_packed(eth_dev);
 959         }
 960
 961         virtio_set_status(hw, VIRTIO_CONFIG_STATUS_FEATURES_OK);
 962
 963         /* Start the device */
 964         virtio_set_status(hw, VIRTIO_CONFIG_STATUS_DRIVER_OK);
 965         if (!dev->started)
 966                 return -1;
 967
 968         if (dev->queue_pairs > 1) {
 969                 ret = virtio_user_handle_mq(dev, dev->queue_pairs);
 970                 if (ret != 0) {
 971                         PMD_INIT_LOG(ERR, "Fails to enable multi-queue pairs!");
 972                         return -1;
 973                 }
 974         }
 975         if (eth_dev->data->dev_flags & RTE_ETH_DEV_INTR_LSC) {
 976                 if (rte_intr_disable(eth_dev->intr_handle) < 0) {
 977                         PMD_DRV_LOG(ERR, "interrupt disable failed");
 978                         return -1;
 979                 }
 980                 rte_intr_callback_unregister(eth_dev->intr_handle,
 981                                              virtio_interrupt_handler,
 982                                              eth_dev);
 983
 984                 eth_dev->intr_handle->fd = dev->ops->get_intr_fd(dev);
 985                 rte_intr_callback_register(eth_dev->intr_handle,
 986                                            virtio_interrupt_handler, eth_dev);
 987
 988                 if (rte_intr_enable(eth_dev->intr_handle) < 0) {
 989                         PMD_DRV_LOG(ERR, "interrupt enable failed");
 990                         return -1;
 991                 }
 992         }
 993         PMD_INIT_LOG(NOTICE, "server mode virtio-user reconnection succeeds!");
 994         return 0;
 995 }