drivers/net/tap/rte_eth_tap.c

   1 /*-
   2  *   BSD LICENSE
   3  *
   4  *   Copyright(c) 2016 Intel Corporation. All rights reserved.
   5  *   All rights reserved.
   6  *
   7  *   Redistribution and use in source and binary forms, with or without
   8  *   modification, are permitted provided that the following conditions
   9  *   are met:
  10  *
  11  *     * Redistributions of source code must retain the above copyright
  12  *       notice, this list of conditions and the following disclaimer.
  13  *     * Redistributions in binary form must reproduce the above copyright
  14  *       notice, this list of conditions and the following disclaimer in
  15  *       the documentation and/or other materials provided with the
  16  *       distribution.
  17  *     * Neither the name of Intel Corporation nor the names of its
  18  *       contributors may be used to endorse or promote products derived
  19  *       from this software without specific prior written permission.
  20  *
  21  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  24  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  25  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  26  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  27  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  28  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  29  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  30  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  31  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  32  */
  33
  34 #include <rte_mbuf.h>
  35 #include <rte_ethdev.h>
  36 #include <rte_malloc.h>
  37 #include <rte_vdev.h>
  38 #include <rte_kvargs.h>
  39
  40 #include <sys/types.h>
  41 #include <sys/stat.h>
  42 #include <sys/socket.h>
  43 #include <sys/ioctl.h>
  44 #include <sys/mman.h>
  45 #include <unistd.h>
  46 #include <poll.h>
  47 #include <arpa/inet.h>
  48 #include <linux/if.h>
  49 #include <linux/if_tun.h>
  50 #include <linux/if_ether.h>
  51 #include <fcntl.h>
  52
  53 /* Linux based path to the TUN device */
  54 #define TUN_TAP_DEV_PATH        "/dev/net/tun"
  55 #define DEFAULT_TAP_NAME        "dtap"
  56
  57 #define ETH_TAP_IFACE_ARG       "iface"
  58 #define ETH_TAP_SPEED_ARG       "speed"
  59
  60 #ifdef IFF_MULTI_QUEUE
  61 #define RTE_PMD_TAP_MAX_QUEUES  16
  62 #else
  63 #define RTE_PMD_TAP_MAX_QUEUES  1
  64 #endif
  65
  66 static struct rte_vdev_driver pmd_tap_drv;
  67
  68 static const char *valid_arguments[] = {
  69         ETH_TAP_IFACE_ARG,
  70         ETH_TAP_SPEED_ARG,
  71         NULL
  72 };
  73
  74 static int tap_unit;
  75
  76 static struct rte_eth_link pmd_link = {
  77         .link_speed = ETH_SPEED_NUM_10G,
  78         .link_duplex = ETH_LINK_FULL_DUPLEX,
  79         .link_status = ETH_LINK_DOWN,
  80         .link_autoneg = ETH_LINK_SPEED_AUTONEG
  81 };
  82
  83 struct pkt_stats {
  84         uint64_t opackets;              /* Number of output packets */
  85         uint64_t ipackets;              /* Number of input packets */
  86         uint64_t obytes;                /* Number of bytes on output */
  87         uint64_t ibytes;                /* Number of bytes on input */
  88         uint64_t errs;                  /* Number of error packets */
  89 };
  90
  91 struct rx_queue {
  92         struct rte_mempool *mp;         /* Mempool for RX packets */
  93         uint16_t in_port;               /* Port ID */
  94         int fd;
  95
  96         struct pkt_stats stats;         /* Stats for this RX queue */
  97 };
  98
  99 struct tx_queue {
 100         int fd;
 101         struct pkt_stats stats;         /* Stats for this TX queue */
 102 };
 103
 104 struct pmd_internals {
 105         char name[RTE_ETH_NAME_MAX_LEN];        /* Internal Tap device name */
 106         uint16_t nb_queues;             /* Number of queues supported */
 107         struct ether_addr eth_addr;     /* Mac address of the device port */
 108
 109         int if_index;                   /* IF_INDEX for the port */
 110
 111         struct rx_queue rxq[RTE_PMD_TAP_MAX_QUEUES];    /* List of RX queues */
 112         struct tx_queue txq[RTE_PMD_TAP_MAX_QUEUES];    /* List of TX queues */
 113 };
 114
 115 /* Tun/Tap allocation routine
 116  *
 117  * name is the number of the interface to use, unless NULL to take the host
 118  * supplied name.
 119  */
 120 static int
 121 tun_alloc(struct pmd_internals *pmd, uint16_t qid)
 122 {
 123         struct ifreq ifr;
 124 #ifdef IFF_MULTI_QUEUE
 125         unsigned int features;
 126 #endif
 127         int fd;
 128
 129         memset(&ifr, 0, sizeof(struct ifreq));
 130
 131         ifr.ifr_flags = IFF_TAP | IFF_NO_PI;
 132         strncpy(ifr.ifr_name, pmd->name, IFNAMSIZ);
 133
 134         RTE_LOG(DEBUG, PMD, "ifr_name '%s'\n", ifr.ifr_name);
 135
 136         fd = open(TUN_TAP_DEV_PATH, O_RDWR);
 137         if (fd < 0) {
 138                 RTE_LOG(ERR, PMD, "Unable to create TAP interface");
 139                 goto error;
 140         }
 141
 142 #ifdef IFF_MULTI_QUEUE
 143         /* Grab the TUN features to verify we can work multi-queue */
 144         if (ioctl(fd, TUNGETFEATURES, &features) < 0) {
 145                 RTE_LOG(ERR, PMD, "TAP unable to get TUN/TAP features\n");
 146                 goto error;
 147         }
 148         RTE_LOG(DEBUG, PMD, "  TAP Features %08x\n", features);
 149
 150         if (features & IFF_MULTI_QUEUE) {
 151                 RTE_LOG(DEBUG, PMD, "  Multi-queue support for %d queues\n",
 152                         RTE_PMD_TAP_MAX_QUEUES);
 153                 ifr.ifr_flags |= IFF_MULTI_QUEUE;
 154         } else
 155 #endif
 156         {
 157                 ifr.ifr_flags |= IFF_ONE_QUEUE;
 158                 RTE_LOG(DEBUG, PMD, "Single queue only support\n");
 159         }
 160
 161         /* Set the TUN/TAP configuration and set the name if needed */
 162         if (ioctl(fd, TUNSETIFF, (void *)&ifr) < 0) {
 163                 RTE_LOG(ERR, PMD, "Unable to set TUNSETIFF for %s\n",
 164                         ifr.ifr_name);
 165                 perror("TUNSETIFF");
 166                 goto error;
 167         }
 168
 169         /* Always set the file descriptor to non-blocking */
 170         if (fcntl(fd, F_SETFL, O_NONBLOCK) < 0) {
 171                 RTE_LOG(ERR, PMD, "Unable to set to nonblocking\n");
 172                 perror("F_SETFL, NONBLOCK");
 173                 goto error;
 174         }
 175
 176         if (qid == 0) {
 177                 if (ioctl(fd, SIOCGIFHWADDR, &ifr) == -1) {
 178                         RTE_LOG(ERR, PMD, "ioctl failed (SIOCGIFHWADDR) (%s)\n",
 179                                 ifr.ifr_name);
 180                         goto error;
 181                 }
 182
 183                 rte_memcpy(&pmd->eth_addr, ifr.ifr_hwaddr.sa_data, ETH_ALEN);
 184         }
 185
 186         return fd;
 187
 188 error:
 189         if (fd > 0)
 190                 close(fd);
 191         return -1;
 192 }
 193
 194 /* Callback to handle the rx burst of packets to the correct interface and
 195  * file descriptor(s) in a multi-queue setup.
 196  */
 197 static uint16_t
 198 pmd_rx_burst(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 199 {
 200         int len;
 201         struct rte_mbuf *mbuf;
 202         struct rx_queue *rxq = queue;
 203         uint16_t num_rx;
 204         unsigned long num_rx_bytes = 0;
 205
 206         for (num_rx = 0; num_rx < nb_pkts; ) {
 207                 /* allocate the next mbuf */
 208                 mbuf = rte_pktmbuf_alloc(rxq->mp);
 209                 if (unlikely(!mbuf)) {
 210                         RTE_LOG(WARNING, PMD, "Unable to allocate mbuf\n");
 211                         break;
 212                 }
 213
 214                 len = read(rxq->fd, rte_pktmbuf_mtod(mbuf, char *),
 215                            rte_pktmbuf_tailroom(mbuf));
 216                 if (len <= 0) {
 217                         rte_pktmbuf_free(mbuf);
 218                         break;
 219                 }
 220
 221                 mbuf->data_len = len;
 222                 mbuf->pkt_len = len;
 223                 mbuf->port = rxq->in_port;
 224
 225                 /* account for the receive frame */
 226                 bufs[num_rx++] = mbuf;
 227                 num_rx_bytes += mbuf->pkt_len;
 228         }
 229         rxq->stats.ipackets += num_rx;
 230         rxq->stats.ibytes += num_rx_bytes;
 231
 232         return num_rx;
 233 }
 234
 235 /* Callback to handle sending packets from the tap interface
 236  */
 237 static uint16_t
 238 pmd_tx_burst(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 239 {
 240         struct rte_mbuf *mbuf;
 241         struct tx_queue *txq = queue;
 242         struct pollfd pfd;
 243         uint16_t num_tx = 0;
 244         unsigned long num_tx_bytes = 0;
 245         int i, n;
 246
 247         if (unlikely(nb_pkts == 0))
 248                 return 0;
 249
 250         pfd.events = POLLOUT;
 251         pfd.fd = txq->fd;
 252         for (i = 0; i < nb_pkts; i++) {
 253                 n = poll(&pfd, 1, 0);
 254
 255                 if (n <= 0)
 256                         break;
 257
 258                 if (pfd.revents & POLLOUT) {
 259                         /* copy the tx frame data */
 260                         mbuf = bufs[num_tx];
 261                         n = write(pfd.fd, rte_pktmbuf_mtod(mbuf, void*),
 262                                   rte_pktmbuf_pkt_len(mbuf));
 263                         if (n <= 0)
 264                                 break;
 265
 266                         num_tx++;
 267                         num_tx_bytes += mbuf->pkt_len;
 268                         rte_pktmbuf_free(mbuf);
 269                 }
 270         }
 271
 272         txq->stats.opackets += num_tx;
 273         txq->stats.errs += nb_pkts - num_tx;
 274         txq->stats.obytes += num_tx_bytes;
 275
 276         return num_tx;
 277 }
 278
 279 static int tap_link_set_flags(struct pmd_internals *pmd, short flags, int add)
 280 {
 281         struct ifreq ifr;
 282         int err, s;
 283
 284         /*
 285          * An AF_INET/DGRAM socket is needed for
 286          * SIOCGIFFLAGS/SIOCSIFFLAGS, using fd won't work.
 287          */
 288         s = socket(AF_INET, SOCK_DGRAM, 0);
 289         if (s < 0) {
 290                 RTE_LOG(ERR, PMD,
 291                         "Unable to get a socket to set flags: %s\n",
 292                         strerror(errno));
 293                 return -1;
 294         }
 295         memset(&ifr, 0, sizeof(ifr));
 296         strncpy(ifr.ifr_name, pmd->name, IFNAMSIZ);
 297         err = ioctl(s, SIOCGIFFLAGS, &ifr);
 298         if (err < 0) {
 299                 RTE_LOG(ERR, PMD, "Unable to get tap netdevice flags: %s\n",
 300                         strerror(errno));
 301                 close(s);
 302                 return -1;
 303         }
 304         if (add)
 305                 ifr.ifr_flags |= flags;
 306         else
 307                 ifr.ifr_flags &= ~flags;
 308         err = ioctl(s, SIOCSIFFLAGS, &ifr);
 309         if (err < 0) {
 310                 RTE_LOG(ERR, PMD, "Unable to %s flags 0x%x: %s\n",
 311                         add ? "set" : "unset", flags, strerror(errno));
 312                 close(s);
 313                 return -1;
 314         }
 315         close(s);
 316
 317         return 0;
 318 }
 319
 320 static int
 321 tap_link_set_down(struct rte_eth_dev *dev)
 322 {
 323         struct pmd_internals *pmd = dev->data->dev_private;
 324
 325         dev->data->dev_link.link_status = ETH_LINK_DOWN;
 326         return tap_link_set_flags(pmd, IFF_UP | IFF_NOARP, 0);
 327 }
 328
 329 static int
 330 tap_link_set_up(struct rte_eth_dev *dev)
 331 {
 332         struct pmd_internals *pmd = dev->data->dev_private;
 333
 334         dev->data->dev_link.link_status = ETH_LINK_UP;
 335         return tap_link_set_flags(pmd, IFF_UP | IFF_NOARP, 1);
 336 }
 337
 338 static int
 339 tap_dev_start(struct rte_eth_dev *dev)
 340 {
 341         return tap_link_set_up(dev);
 342 }
 343
 344 /* This function gets called when the current port gets stopped.
 345  */
 346 static void
 347 tap_dev_stop(struct rte_eth_dev *dev)
 348 {
 349         int i;
 350         struct pmd_internals *internals = dev->data->dev_private;
 351
 352         for (i = 0; i < internals->nb_queues; i++)
 353                 if (internals->rxq[i].fd != -1)
 354                         close(internals->rxq[i].fd);
 355         tap_link_set_down(dev);
 356 }
 357
 358 static int
 359 tap_dev_configure(struct rte_eth_dev *dev __rte_unused)
 360 {
 361         return 0;
 362 }
 363
 364 static void
 365 tap_dev_info(struct rte_eth_dev *dev, struct rte_eth_dev_info *dev_info)
 366 {
 367         struct pmd_internals *internals = dev->data->dev_private;
 368
 369         dev_info->if_index = internals->if_index;
 370         dev_info->max_mac_addrs = 1;
 371         dev_info->max_rx_pktlen = (uint32_t)ETHER_MAX_VLAN_FRAME_LEN;
 372         dev_info->max_rx_queues = internals->nb_queues;
 373         dev_info->max_tx_queues = internals->nb_queues;
 374         dev_info->min_rx_bufsize = 0;
 375         dev_info->pci_dev = NULL;
 376 }
 377
 378 static void
 379 tap_stats_get(struct rte_eth_dev *dev, struct rte_eth_stats *tap_stats)
 380 {
 381         unsigned int i, imax;
 382         unsigned long rx_total = 0, tx_total = 0, tx_err_total = 0;
 383         unsigned long rx_bytes_total = 0, tx_bytes_total = 0;
 384         const struct pmd_internals *pmd = dev->data->dev_private;
 385
 386         imax = (pmd->nb_queues < RTE_ETHDEV_QUEUE_STAT_CNTRS) ?
 387                 pmd->nb_queues : RTE_ETHDEV_QUEUE_STAT_CNTRS;
 388
 389         for (i = 0; i < imax; i++) {
 390                 tap_stats->q_ipackets[i] = pmd->rxq[i].stats.ipackets;
 391                 tap_stats->q_ibytes[i] = pmd->rxq[i].stats.ibytes;
 392                 rx_total += tap_stats->q_ipackets[i];
 393                 rx_bytes_total += tap_stats->q_ibytes[i];
 394
 395                 tap_stats->q_opackets[i] = pmd->txq[i].stats.opackets;
 396                 tap_stats->q_errors[i] = pmd->txq[i].stats.errs;
 397                 tap_stats->q_obytes[i] = pmd->txq[i].stats.obytes;
 398                 tx_total += tap_stats->q_opackets[i];
 399                 tx_err_total += tap_stats->q_errors[i];
 400                 tx_bytes_total += tap_stats->q_obytes[i];
 401         }
 402
 403         tap_stats->ipackets = rx_total;
 404         tap_stats->ibytes = rx_bytes_total;
 405         tap_stats->opackets = tx_total;
 406         tap_stats->oerrors = tx_err_total;
 407         tap_stats->obytes = tx_bytes_total;
 408 }
 409
 410 static void
 411 tap_stats_reset(struct rte_eth_dev *dev)
 412 {
 413         int i;
 414         struct pmd_internals *pmd = dev->data->dev_private;
 415
 416         for (i = 0; i < pmd->nb_queues; i++) {
 417                 pmd->rxq[i].stats.ipackets = 0;
 418                 pmd->rxq[i].stats.ibytes = 0;
 419
 420                 pmd->txq[i].stats.opackets = 0;
 421                 pmd->txq[i].stats.errs = 0;
 422                 pmd->txq[i].stats.obytes = 0;
 423         }
 424 }
 425
 426 static void
 427 tap_dev_close(struct rte_eth_dev *dev __rte_unused)
 428 {
 429 }
 430
 431 static void
 432 tap_rx_queue_release(void *queue)
 433 {
 434         struct rx_queue *rxq = queue;
 435
 436         if (rxq && (rxq->fd > 0)) {
 437                 close(rxq->fd);
 438                 rxq->fd = -1;
 439         }
 440 }
 441
 442 static void
 443 tap_tx_queue_release(void *queue)
 444 {
 445         struct tx_queue *txq = queue;
 446
 447         if (txq && (txq->fd > 0)) {
 448                 close(txq->fd);
 449                 txq->fd = -1;
 450         }
 451 }
 452
 453 static int
 454 tap_link_update(struct rte_eth_dev *dev __rte_unused,
 455                 int wait_to_complete __rte_unused)
 456 {
 457         return 0;
 458 }
 459
 460 static void
 461 tap_promisc_enable(struct rte_eth_dev *dev)
 462 {
 463         struct pmd_internals *pmd = dev->data->dev_private;
 464
 465         dev->data->promiscuous = 1;
 466         tap_link_set_flags(pmd, IFF_PROMISC, 1);
 467 }
 468
 469 static void
 470 tap_promisc_disable(struct rte_eth_dev *dev)
 471 {
 472         struct pmd_internals *pmd = dev->data->dev_private;
 473
 474         dev->data->promiscuous = 0;
 475         tap_link_set_flags(pmd, IFF_PROMISC, 0);
 476 }
 477
 478 static void
 479 tap_allmulti_enable(struct rte_eth_dev *dev)
 480 {
 481         struct pmd_internals *pmd = dev->data->dev_private;
 482
 483         dev->data->all_multicast = 1;
 484         tap_link_set_flags(pmd, IFF_ALLMULTI, 1);
 485 }
 486
 487 static void
 488 tap_allmulti_disable(struct rte_eth_dev *dev)
 489 {
 490         struct pmd_internals *pmd = dev->data->dev_private;
 491
 492         dev->data->all_multicast = 0;
 493         tap_link_set_flags(pmd, IFF_ALLMULTI, 0);
 494 }
 495
 496 static int
 497 tap_setup_queue(struct rte_eth_dev *dev,
 498                 struct pmd_internals *internals,
 499                 uint16_t qid)
 500 {
 501         struct pmd_internals *pmd = dev->data->dev_private;
 502         struct rx_queue *rx = &internals->rxq[qid];
 503         struct tx_queue *tx = &internals->txq[qid];
 504         int fd;
 505
 506         fd = rx->fd;
 507         if (fd < 0) {
 508                 fd = tx->fd;
 509                 if (fd < 0) {
 510                         RTE_LOG(INFO, PMD, "Add queue to TAP %s for qid %d\n",
 511                                 pmd->name, qid);
 512                         fd = tun_alloc(pmd, qid);
 513                         if (fd < 0) {
 514                                 RTE_LOG(ERR, PMD, "tun_alloc(%s) failed\n",
 515                                         pmd->name);
 516                                 return -1;
 517                         }
 518                 }
 519         }
 520
 521         rx->fd = fd;
 522         tx->fd = fd;
 523
 524         return fd;
 525 }
 526
 527 static int
 528 rx_setup_queue(struct rte_eth_dev *dev,
 529                 struct pmd_internals *internals,
 530                 uint16_t qid)
 531 {
 532         dev->data->rx_queues[qid] = &internals->rxq[qid];
 533
 534         return tap_setup_queue(dev, internals, qid);
 535 }
 536
 537 static int
 538 tx_setup_queue(struct rte_eth_dev *dev,
 539                 struct pmd_internals *internals,
 540                 uint16_t qid)
 541 {
 542         dev->data->tx_queues[qid] = &internals->txq[qid];
 543
 544         return tap_setup_queue(dev, internals, qid);
 545 }
 546
 547 static int
 548 tap_rx_queue_setup(struct rte_eth_dev *dev,
 549                    uint16_t rx_queue_id,
 550                    uint16_t nb_rx_desc __rte_unused,
 551                    unsigned int socket_id __rte_unused,
 552                    const struct rte_eth_rxconf *rx_conf __rte_unused,
 553                    struct rte_mempool *mp)
 554 {
 555         struct pmd_internals *internals = dev->data->dev_private;
 556         uint16_t buf_size;
 557         int fd;
 558
 559         if ((rx_queue_id >= internals->nb_queues) || !mp) {
 560                 RTE_LOG(ERR, PMD, "nb_queues %d mp %p\n",
 561                         internals->nb_queues, mp);
 562                 return -1;
 563         }
 564
 565         internals->rxq[rx_queue_id].mp = mp;
 566         internals->rxq[rx_queue_id].in_port = dev->data->port_id;
 567
 568         /* Now get the space available for data in the mbuf */
 569         buf_size = (uint16_t)(rte_pktmbuf_data_room_size(mp) -
 570                                 RTE_PKTMBUF_HEADROOM);
 571
 572         if (buf_size < ETH_FRAME_LEN) {
 573                 RTE_LOG(ERR, PMD,
 574                         "%s: %d bytes will not fit in mbuf (%d bytes)\n",
 575                         dev->data->name, ETH_FRAME_LEN, buf_size);
 576                 return -ENOMEM;
 577         }
 578
 579         fd = rx_setup_queue(dev, internals, rx_queue_id);
 580         if (fd == -1)
 581                 return -1;
 582
 583         RTE_LOG(INFO, PMD, "RX TAP device name %s, qid %d on fd %d\n",
 584                 internals->name, rx_queue_id, internals->rxq[rx_queue_id].fd);
 585
 586         return 0;
 587 }
 588
 589 static int
 590 tap_tx_queue_setup(struct rte_eth_dev *dev,
 591                    uint16_t tx_queue_id,
 592                    uint16_t nb_tx_desc __rte_unused,
 593                    unsigned int socket_id __rte_unused,
 594                    const struct rte_eth_txconf *tx_conf __rte_unused)
 595 {
 596         struct pmd_internals *internals = dev->data->dev_private;
 597         int ret;
 598
 599         if (tx_queue_id >= internals->nb_queues)
 600                 return -1;
 601
 602         ret = tx_setup_queue(dev, internals, tx_queue_id);
 603         if (ret == -1)
 604                 return -1;
 605
 606         RTE_LOG(INFO, PMD, "TX TAP device name %s, qid %d on fd %d\n",
 607                 internals->name, tx_queue_id, internals->txq[tx_queue_id].fd);
 608
 609         return 0;
 610 }
 611
 612 static const struct eth_dev_ops ops = {
 613         .dev_start              = tap_dev_start,
 614         .dev_stop               = tap_dev_stop,
 615         .dev_close              = tap_dev_close,
 616         .dev_configure          = tap_dev_configure,
 617         .dev_infos_get          = tap_dev_info,
 618         .rx_queue_setup         = tap_rx_queue_setup,
 619         .tx_queue_setup         = tap_tx_queue_setup,
 620         .rx_queue_release       = tap_rx_queue_release,
 621         .tx_queue_release       = tap_tx_queue_release,
 622         .link_update            = tap_link_update,
 623         .dev_set_link_up        = tap_link_set_up,
 624         .dev_set_link_down      = tap_link_set_down,
 625         .promiscuous_enable     = tap_promisc_enable,
 626         .promiscuous_disable    = tap_promisc_disable,
 627         .allmulticast_enable    = tap_allmulti_enable,
 628         .allmulticast_disable   = tap_allmulti_disable,
 629         .stats_get              = tap_stats_get,
 630         .stats_reset            = tap_stats_reset,
 631 };
 632
 633 static int
 634 eth_dev_tap_create(const char *name, char *tap_name)
 635 {
 636         int numa_node = rte_socket_id();
 637         struct rte_eth_dev *dev = NULL;
 638         struct pmd_internals *pmd = NULL;
 639         struct rte_eth_dev_data *data = NULL;
 640         int i;
 641
 642         RTE_LOG(INFO, PMD,
 643                 "%s: Create TAP Ethernet device with %d queues on numa %u\n",
 644                  name, RTE_PMD_TAP_MAX_QUEUES, rte_socket_id());
 645
 646         data = rte_zmalloc_socket(tap_name, sizeof(*data), 0, numa_node);
 647         if (!data) {
 648                 RTE_LOG(ERR, PMD, "Failed to allocate data\n");
 649                 goto error_exit;
 650         }
 651
 652         pmd = rte_zmalloc_socket(tap_name, sizeof(*pmd), 0, numa_node);
 653         if (!pmd) {
 654                 RTE_LOG(ERR, PMD, "Unable to allocate internal struct\n");
 655                 goto error_exit;
 656         }
 657
 658         dev = rte_eth_dev_allocate(tap_name);
 659         if (!dev) {
 660                 RTE_LOG(ERR, PMD, "Unable to allocate device struct\n");
 661                 goto error_exit;
 662         }
 663
 664         snprintf(pmd->name, sizeof(pmd->name), "%s", tap_name);
 665
 666         pmd->nb_queues = RTE_PMD_TAP_MAX_QUEUES;
 667
 668         /* Setup some default values */
 669         data->dev_private = pmd;
 670         data->port_id = dev->data->port_id;
 671         data->dev_flags = RTE_ETH_DEV_DETACHABLE;
 672         data->kdrv = RTE_KDRV_NONE;
 673         data->drv_name = pmd_tap_drv.driver.name;
 674         data->numa_node = numa_node;
 675
 676         data->dev_link = pmd_link;
 677         data->mac_addrs = &pmd->eth_addr;
 678         data->nb_rx_queues = pmd->nb_queues;
 679         data->nb_tx_queues = pmd->nb_queues;
 680
 681         dev->data = data;
 682         dev->dev_ops = &ops;
 683         dev->driver = NULL;
 684         dev->rx_pkt_burst = pmd_rx_burst;
 685         dev->tx_pkt_burst = pmd_tx_burst;
 686         snprintf(dev->data->name, sizeof(dev->data->name), "%s", name);
 687
 688         /* Presetup the fds to -1 as being not valid */
 689         for (i = 0; i < RTE_PMD_TAP_MAX_QUEUES; i++) {
 690                 pmd->rxq[i].fd = -1;
 691                 pmd->txq[i].fd = -1;
 692         }
 693
 694         return 0;
 695
 696 error_exit:
 697         RTE_PMD_DEBUG_TRACE("Unable to initialize %s\n", name);
 698
 699         rte_free(data);
 700         rte_free(pmd);
 701
 702         rte_eth_dev_release_port(dev);
 703
 704         return -EINVAL;
 705 }
 706
 707 static int
 708 set_interface_name(const char *key __rte_unused,
 709                    const char *value,
 710                    void *extra_args)
 711 {
 712         char *name = (char *)extra_args;
 713
 714         if (value)
 715                 snprintf(name, RTE_ETH_NAME_MAX_LEN - 1, "%s", value);
 716         else
 717                 snprintf(name, RTE_ETH_NAME_MAX_LEN - 1, "%s%d",
 718                          DEFAULT_TAP_NAME, (tap_unit - 1));
 719
 720         return 0;
 721 }
 722
 723 static int
 724 set_interface_speed(const char *key __rte_unused,
 725                     const char *value,
 726                     void *extra_args)
 727 {
 728         *(int *)extra_args = (value) ? atoi(value) : ETH_SPEED_NUM_10G;
 729
 730         return 0;
 731 }
 732
 733 /* Open a TAP interface device.
 734  */
 735 static int
 736 rte_pmd_tap_probe(const char *name, const char *params)
 737 {
 738         int ret;
 739         struct rte_kvargs *kvlist = NULL;
 740         int speed;
 741         char tap_name[RTE_ETH_NAME_MAX_LEN];
 742
 743         speed = ETH_SPEED_NUM_10G;
 744         snprintf(tap_name, sizeof(tap_name), "%s%d",
 745                  DEFAULT_TAP_NAME, tap_unit++);
 746
 747         if (params && (params[0] != '\0')) {
 748                 RTE_LOG(INFO, PMD, "paramaters (%s)\n", params);
 749
 750                 kvlist = rte_kvargs_parse(params, valid_arguments);
 751                 if (kvlist) {
 752                         if (rte_kvargs_count(kvlist, ETH_TAP_SPEED_ARG) == 1) {
 753                                 ret = rte_kvargs_process(kvlist,
 754                                                          ETH_TAP_SPEED_ARG,
 755                                                          &set_interface_speed,
 756                                                          &speed);
 757                                 if (ret == -1)
 758                                         goto leave;
 759                         }
 760
 761                         if (rte_kvargs_count(kvlist, ETH_TAP_IFACE_ARG) == 1) {
 762                                 ret = rte_kvargs_process(kvlist,
 763                                                          ETH_TAP_IFACE_ARG,
 764                                                          &set_interface_name,
 765                                                          tap_name);
 766                                 if (ret == -1)
 767                                         goto leave;
 768                         }
 769                 }
 770         }
 771         pmd_link.link_speed = speed;
 772
 773         RTE_LOG(INFO, PMD, "Initializing pmd_tap for %s as %s\n",
 774                 name, tap_name);
 775
 776         ret = eth_dev_tap_create(name, tap_name);
 777
 778 leave:
 779         if (ret == -1) {
 780                 RTE_LOG(INFO, PMD, "Failed to create pmd for %s as %s\n",
 781                         name, tap_name);
 782                 tap_unit--;             /* Restore the unit number */
 783         }
 784         rte_kvargs_free(kvlist);
 785
 786         return ret;
 787 }
 788
 789 /* detach a TAP device.
 790  */
 791 static int
 792 rte_pmd_tap_remove(const char *name)
 793 {
 794         struct rte_eth_dev *eth_dev = NULL;
 795         struct pmd_internals *internals;
 796         int i;
 797
 798         RTE_LOG(INFO, PMD, "Closing TUN/TAP Ethernet device on numa %u\n",
 799                 rte_socket_id());
 800
 801         /* find the ethdev entry */
 802         eth_dev = rte_eth_dev_allocated(name);
 803         if (!eth_dev)
 804                 return 0;
 805
 806         internals = eth_dev->data->dev_private;
 807         for (i = 0; i < internals->nb_queues; i++)
 808                 if (internals->rxq[i].fd != -1)
 809                         close(internals->rxq[i].fd);
 810
 811         rte_free(eth_dev->data->dev_private);
 812         rte_free(eth_dev->data);
 813
 814         rte_eth_dev_release_port(eth_dev);
 815
 816         return 0;
 817 }
 818
 819 static struct rte_vdev_driver pmd_tap_drv = {
 820         .probe = rte_pmd_tap_probe,
 821         .remove = rte_pmd_tap_remove,
 822 };
 823 RTE_PMD_REGISTER_VDEV(net_tap, pmd_tap_drv);
 824 RTE_PMD_REGISTER_ALIAS(net_tap, eth_tap);
 825 RTE_PMD_REGISTER_PARAM_STRING(net_tap, "iface=<string>,speed=N");