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(WARNING, PMD,
 164                         "Unable to set TUNSETIFF for %s\n",
 165                         ifr.ifr_name);
 166                 perror("TUNSETIFF");
 167                 goto error;
 168         }
 169
 170         /* Always set the file descriptor to non-blocking */
 171         if (fcntl(fd, F_SETFL, O_NONBLOCK) < 0) {
 172                 RTE_LOG(WARNING, PMD,
 173                         "Unable to set %s to nonblocking\n",
 174                         ifr.ifr_name);
 175                 perror("F_SETFL, NONBLOCK");
 176                 goto error;
 177         }
 178
 179         if (qid == 0) {
 180                 if (ioctl(fd, SIOCGIFHWADDR, &ifr) == -1) {
 181                         RTE_LOG(ERR, PMD, "ioctl failed (SIOCGIFHWADDR) (%s)\n",
 182                                 ifr.ifr_name);
 183                         goto error;
 184                 }
 185
 186                 rte_memcpy(&pmd->eth_addr, ifr.ifr_hwaddr.sa_data, ETH_ALEN);
 187         }
 188
 189         return fd;
 190
 191 error:
 192         if (fd > 0)
 193                 close(fd);
 194         return -1;
 195 }
 196
 197 /* Callback to handle the rx burst of packets to the correct interface and
 198  * file descriptor(s) in a multi-queue setup.
 199  */
 200 static uint16_t
 201 pmd_rx_burst(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 202 {
 203         int len;
 204         struct rte_mbuf *mbuf;
 205         struct rx_queue *rxq = queue;
 206         uint16_t num_rx;
 207         unsigned long num_rx_bytes = 0;
 208
 209         for (num_rx = 0; num_rx < nb_pkts; ) {
 210                 /* allocate the next mbuf */
 211                 mbuf = rte_pktmbuf_alloc(rxq->mp);
 212                 if (unlikely(!mbuf)) {
 213                         RTE_LOG(WARNING, PMD, "TAP unable to allocate mbuf\n");
 214                         break;
 215                 }
 216
 217                 len = read(rxq->fd, rte_pktmbuf_mtod(mbuf, char *),
 218                            rte_pktmbuf_tailroom(mbuf));
 219                 if (len <= 0) {
 220                         rte_pktmbuf_free(mbuf);
 221                         break;
 222                 }
 223
 224                 mbuf->data_len = len;
 225                 mbuf->pkt_len = len;
 226                 mbuf->port = rxq->in_port;
 227
 228                 /* account for the receive frame */
 229                 bufs[num_rx++] = mbuf;
 230                 num_rx_bytes += mbuf->pkt_len;
 231         }
 232         rxq->stats.ipackets += num_rx;
 233         rxq->stats.ibytes += num_rx_bytes;
 234
 235         return num_rx;
 236 }
 237
 238 /* Callback to handle sending packets from the tap interface
 239  */
 240 static uint16_t
 241 pmd_tx_burst(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 242 {
 243         struct rte_mbuf *mbuf;
 244         struct tx_queue *txq = queue;
 245         struct pollfd pfd;
 246         uint16_t num_tx = 0;
 247         unsigned long num_tx_bytes = 0;
 248         int i, n;
 249
 250         if (unlikely(nb_pkts == 0))
 251                 return 0;
 252
 253         pfd.events = POLLOUT;
 254         pfd.fd = txq->fd;
 255         for (i = 0; i < nb_pkts; i++) {
 256                 n = poll(&pfd, 1, 0);
 257
 258                 if (n <= 0)
 259                         break;
 260
 261                 if (pfd.revents & POLLOUT) {
 262                         /* copy the tx frame data */
 263                         mbuf = bufs[num_tx];
 264                         n = write(pfd.fd, rte_pktmbuf_mtod(mbuf, void*),
 265                                   rte_pktmbuf_pkt_len(mbuf));
 266                         if (n <= 0)
 267                                 break;
 268
 269                         num_tx++;
 270                         num_tx_bytes += mbuf->pkt_len;
 271                         rte_pktmbuf_free(mbuf);
 272                 }
 273         }
 274
 275         txq->stats.opackets += num_tx;
 276         txq->stats.errs += nb_pkts - num_tx;
 277         txq->stats.obytes += num_tx_bytes;
 278
 279         return num_tx;
 280 }
 281
 282 static int
 283 tap_link_set_flags(struct pmd_internals *pmd, short flags, int add)
 284 {
 285         struct ifreq ifr;
 286         int err, s;
 287
 288         /*
 289          * An AF_INET/DGRAM socket is needed for
 290          * SIOCGIFFLAGS/SIOCSIFFLAGS, using fd won't work.
 291          */
 292         s = socket(AF_INET, SOCK_DGRAM, 0);
 293         if (s < 0) {
 294                 RTE_LOG(ERR, PMD,
 295                         "Unable to get a socket to set flags: %s\n",
 296                         strerror(errno));
 297                 return -1;
 298         }
 299         memset(&ifr, 0, sizeof(ifr));
 300         strncpy(ifr.ifr_name, pmd->name, IFNAMSIZ);
 301         err = ioctl(s, SIOCGIFFLAGS, &ifr);
 302         if (err < 0) {
 303                 RTE_LOG(WARNING, PMD, "Unable to get %s device flags: %s\n",
 304                         pmd->name, strerror(errno));
 305                 close(s);
 306                 return -1;
 307         }
 308         if (add)
 309                 ifr.ifr_flags |= flags;
 310         else
 311                 ifr.ifr_flags &= ~flags;
 312         err = ioctl(s, SIOCSIFFLAGS, &ifr);
 313         if (err < 0) {
 314                 RTE_LOG(WARNING, PMD, "Unable to %s flags 0x%x: %s\n",
 315                         add ? "set" : "unset", flags, strerror(errno));
 316                 close(s);
 317                 return -1;
 318         }
 319         close(s);
 320
 321         return 0;
 322 }
 323
 324 static int
 325 tap_link_set_down(struct rte_eth_dev *dev)
 326 {
 327         struct pmd_internals *pmd = dev->data->dev_private;
 328
 329         dev->data->dev_link.link_status = ETH_LINK_DOWN;
 330         return tap_link_set_flags(pmd, IFF_UP | IFF_NOARP, 0);
 331 }
 332
 333 static int
 334 tap_link_set_up(struct rte_eth_dev *dev)
 335 {
 336         struct pmd_internals *pmd = dev->data->dev_private;
 337
 338         dev->data->dev_link.link_status = ETH_LINK_UP;
 339         return tap_link_set_flags(pmd, IFF_UP | IFF_NOARP, 1);
 340 }
 341
 342 static int
 343 tap_dev_start(struct rte_eth_dev *dev)
 344 {
 345         return tap_link_set_up(dev);
 346 }
 347
 348 /* This function gets called when the current port gets stopped.
 349  */
 350 static void
 351 tap_dev_stop(struct rte_eth_dev *dev)
 352 {
 353         int i;
 354         struct pmd_internals *internals = dev->data->dev_private;
 355
 356         for (i = 0; i < internals->nb_queues; i++)
 357                 if (internals->rxq[i].fd != -1)
 358                         close(internals->rxq[i].fd);
 359         tap_link_set_down(dev);
 360 }
 361
 362 static int
 363 tap_dev_configure(struct rte_eth_dev *dev __rte_unused)
 364 {
 365         return 0;
 366 }
 367
 368 static void
 369 tap_dev_info(struct rte_eth_dev *dev, struct rte_eth_dev_info *dev_info)
 370 {
 371         struct pmd_internals *internals = dev->data->dev_private;
 372
 373         dev_info->if_index = internals->if_index;
 374         dev_info->max_mac_addrs = 1;
 375         dev_info->max_rx_pktlen = (uint32_t)ETHER_MAX_VLAN_FRAME_LEN;
 376         dev_info->max_rx_queues = internals->nb_queues;
 377         dev_info->max_tx_queues = internals->nb_queues;
 378         dev_info->min_rx_bufsize = 0;
 379         dev_info->pci_dev = NULL;
 380 }
 381
 382 static void
 383 tap_stats_get(struct rte_eth_dev *dev, struct rte_eth_stats *tap_stats)
 384 {
 385         unsigned int i, imax;
 386         unsigned long rx_total = 0, tx_total = 0, tx_err_total = 0;
 387         unsigned long rx_bytes_total = 0, tx_bytes_total = 0;
 388         const struct pmd_internals *pmd = dev->data->dev_private;
 389
 390         imax = (pmd->nb_queues < RTE_ETHDEV_QUEUE_STAT_CNTRS) ?
 391                 pmd->nb_queues : RTE_ETHDEV_QUEUE_STAT_CNTRS;
 392
 393         for (i = 0; i < imax; i++) {
 394                 tap_stats->q_ipackets[i] = pmd->rxq[i].stats.ipackets;
 395                 tap_stats->q_ibytes[i] = pmd->rxq[i].stats.ibytes;
 396                 rx_total += tap_stats->q_ipackets[i];
 397                 rx_bytes_total += tap_stats->q_ibytes[i];
 398
 399                 tap_stats->q_opackets[i] = pmd->txq[i].stats.opackets;
 400                 tap_stats->q_errors[i] = pmd->txq[i].stats.errs;
 401                 tap_stats->q_obytes[i] = pmd->txq[i].stats.obytes;
 402                 tx_total += tap_stats->q_opackets[i];
 403                 tx_err_total += tap_stats->q_errors[i];
 404                 tx_bytes_total += tap_stats->q_obytes[i];
 405         }
 406
 407         tap_stats->ipackets = rx_total;
 408         tap_stats->ibytes = rx_bytes_total;
 409         tap_stats->opackets = tx_total;
 410         tap_stats->oerrors = tx_err_total;
 411         tap_stats->obytes = tx_bytes_total;
 412 }
 413
 414 static void
 415 tap_stats_reset(struct rte_eth_dev *dev)
 416 {
 417         int i;
 418         struct pmd_internals *pmd = dev->data->dev_private;
 419
 420         for (i = 0; i < pmd->nb_queues; i++) {
 421                 pmd->rxq[i].stats.ipackets = 0;
 422                 pmd->rxq[i].stats.ibytes = 0;
 423
 424                 pmd->txq[i].stats.opackets = 0;
 425                 pmd->txq[i].stats.errs = 0;
 426                 pmd->txq[i].stats.obytes = 0;
 427         }
 428 }
 429
 430 static void
 431 tap_dev_close(struct rte_eth_dev *dev __rte_unused)
 432 {
 433 }
 434
 435 static void
 436 tap_rx_queue_release(void *queue)
 437 {
 438         struct rx_queue *rxq = queue;
 439
 440         if (rxq && (rxq->fd > 0)) {
 441                 close(rxq->fd);
 442                 rxq->fd = -1;
 443         }
 444 }
 445
 446 static void
 447 tap_tx_queue_release(void *queue)
 448 {
 449         struct tx_queue *txq = queue;
 450
 451         if (txq && (txq->fd > 0)) {
 452                 close(txq->fd);
 453                 txq->fd = -1;
 454         }
 455 }
 456
 457 static int
 458 tap_link_update(struct rte_eth_dev *dev __rte_unused,
 459                 int wait_to_complete __rte_unused)
 460 {
 461         return 0;
 462 }
 463
 464 static void
 465 tap_promisc_enable(struct rte_eth_dev *dev)
 466 {
 467         struct pmd_internals *pmd = dev->data->dev_private;
 468
 469         dev->data->promiscuous = 1;
 470         tap_link_set_flags(pmd, IFF_PROMISC, 1);
 471 }
 472
 473 static void
 474 tap_promisc_disable(struct rte_eth_dev *dev)
 475 {
 476         struct pmd_internals *pmd = dev->data->dev_private;
 477
 478         dev->data->promiscuous = 0;
 479         tap_link_set_flags(pmd, IFF_PROMISC, 0);
 480 }
 481
 482 static void
 483 tap_allmulti_enable(struct rte_eth_dev *dev)
 484 {
 485         struct pmd_internals *pmd = dev->data->dev_private;
 486
 487         dev->data->all_multicast = 1;
 488         tap_link_set_flags(pmd, IFF_ALLMULTI, 1);
 489 }
 490
 491 static void
 492 tap_allmulti_disable(struct rte_eth_dev *dev)
 493 {
 494         struct pmd_internals *pmd = dev->data->dev_private;
 495
 496         dev->data->all_multicast = 0;
 497         tap_link_set_flags(pmd, IFF_ALLMULTI, 0);
 498 }
 499
 500 static int
 501 tap_setup_queue(struct rte_eth_dev *dev,
 502                 struct pmd_internals *internals,
 503                 uint16_t qid)
 504 {
 505         struct pmd_internals *pmd = dev->data->dev_private;
 506         struct rx_queue *rx = &internals->rxq[qid];
 507         struct tx_queue *tx = &internals->txq[qid];
 508         int fd;
 509
 510         fd = rx->fd;
 511         if (fd < 0) {
 512                 fd = tx->fd;
 513                 if (fd < 0) {
 514                         RTE_LOG(INFO, PMD, "Add queue to TAP %s for qid %d\n",
 515                                 pmd->name, qid);
 516                         fd = tun_alloc(pmd, qid);
 517                         if (fd < 0) {
 518                                 RTE_LOG(ERR, PMD, "tun_alloc(%s, %d) failed\n",
 519                                         pmd->name, qid);
 520                                 return -1;
 521                         }
 522                 }
 523         }
 524
 525         rx->fd = fd;
 526         tx->fd = fd;
 527
 528         return fd;
 529 }
 530
 531 static int
 532 rx_setup_queue(struct rte_eth_dev *dev,
 533                 struct pmd_internals *internals,
 534                 uint16_t qid)
 535 {
 536         dev->data->rx_queues[qid] = &internals->rxq[qid];
 537
 538         return tap_setup_queue(dev, internals, qid);
 539 }
 540
 541 static int
 542 tx_setup_queue(struct rte_eth_dev *dev,
 543                 struct pmd_internals *internals,
 544                 uint16_t qid)
 545 {
 546         dev->data->tx_queues[qid] = &internals->txq[qid];
 547
 548         return tap_setup_queue(dev, internals, qid);
 549 }
 550
 551 static int
 552 tap_rx_queue_setup(struct rte_eth_dev *dev,
 553                    uint16_t rx_queue_id,
 554                    uint16_t nb_rx_desc __rte_unused,
 555                    unsigned int socket_id __rte_unused,
 556                    const struct rte_eth_rxconf *rx_conf __rte_unused,
 557                    struct rte_mempool *mp)
 558 {
 559         struct pmd_internals *internals = dev->data->dev_private;
 560         uint16_t buf_size;
 561         int fd;
 562
 563         if ((rx_queue_id >= internals->nb_queues) || !mp) {
 564                 RTE_LOG(WARNING, PMD,
 565                         "nb_queues %d too small or mempool NULL\n",
 566                         internals->nb_queues);
 567                 return -1;
 568         }
 569
 570         internals->rxq[rx_queue_id].mp = mp;
 571         internals->rxq[rx_queue_id].in_port = dev->data->port_id;
 572
 573         /* Now get the space available for data in the mbuf */
 574         buf_size = (uint16_t)(rte_pktmbuf_data_room_size(mp) -
 575                                 RTE_PKTMBUF_HEADROOM);
 576
 577         if (buf_size < ETH_FRAME_LEN) {
 578                 RTE_LOG(WARNING, PMD,
 579                         "%s: %d bytes will not fit in mbuf (%d bytes)\n",
 580                         dev->data->name, ETH_FRAME_LEN, buf_size);
 581                 return -ENOMEM;
 582         }
 583
 584         fd = rx_setup_queue(dev, internals, rx_queue_id);
 585         if (fd == -1)
 586                 return -1;
 587
 588         RTE_LOG(DEBUG, PMD, "  RX TAP device name %s, qid %d on fd %d\n",
 589                 internals->name, rx_queue_id, internals->rxq[rx_queue_id].fd);
 590
 591         return 0;
 592 }
 593
 594 static int
 595 tap_tx_queue_setup(struct rte_eth_dev *dev,
 596                    uint16_t tx_queue_id,
 597                    uint16_t nb_tx_desc __rte_unused,
 598                    unsigned int socket_id __rte_unused,
 599                    const struct rte_eth_txconf *tx_conf __rte_unused)
 600 {
 601         struct pmd_internals *internals = dev->data->dev_private;
 602         int ret;
 603
 604         if (tx_queue_id >= internals->nb_queues)
 605                 return -1;
 606
 607         ret = tx_setup_queue(dev, internals, tx_queue_id);
 608         if (ret == -1)
 609                 return -1;
 610
 611         RTE_LOG(DEBUG, PMD, "  TX TAP device name %s, qid %d on fd %d\n",
 612                 internals->name, tx_queue_id, internals->txq[tx_queue_id].fd);
 613
 614         return 0;
 615 }
 616
 617 static const struct eth_dev_ops ops = {
 618         .dev_start              = tap_dev_start,
 619         .dev_stop               = tap_dev_stop,
 620         .dev_close              = tap_dev_close,
 621         .dev_configure          = tap_dev_configure,
 622         .dev_infos_get          = tap_dev_info,
 623         .rx_queue_setup         = tap_rx_queue_setup,
 624         .tx_queue_setup         = tap_tx_queue_setup,
 625         .rx_queue_release       = tap_rx_queue_release,
 626         .tx_queue_release       = tap_tx_queue_release,
 627         .link_update            = tap_link_update,
 628         .dev_set_link_up        = tap_link_set_up,
 629         .dev_set_link_down      = tap_link_set_down,
 630         .promiscuous_enable     = tap_promisc_enable,
 631         .promiscuous_disable    = tap_promisc_disable,
 632         .allmulticast_enable    = tap_allmulti_enable,
 633         .allmulticast_disable   = tap_allmulti_disable,
 634         .stats_get              = tap_stats_get,
 635         .stats_reset            = tap_stats_reset,
 636 };
 637
 638 static int
 639 eth_dev_tap_create(const char *name, char *tap_name)
 640 {
 641         int numa_node = rte_socket_id();
 642         struct rte_eth_dev *dev = NULL;
 643         struct pmd_internals *pmd = NULL;
 644         struct rte_eth_dev_data *data = NULL;
 645         int i;
 646
 647         RTE_LOG(DEBUG, PMD, "  TAP device on numa %u\n", rte_socket_id());
 648
 649         data = rte_zmalloc_socket(tap_name, sizeof(*data), 0, numa_node);
 650         if (!data) {
 651                 RTE_LOG(ERR, PMD, "TAP Failed to allocate data\n");
 652                 goto error_exit;
 653         }
 654
 655         pmd = rte_zmalloc_socket(tap_name, sizeof(*pmd), 0, numa_node);
 656         if (!pmd) {
 657                 RTE_LOG(ERR, PMD, "TAP Unable to allocate internal struct\n");
 658                 goto error_exit;
 659         }
 660
 661         dev = rte_eth_dev_allocate(tap_name);
 662         if (!dev) {
 663                 RTE_LOG(ERR, PMD, "TAP Unable to allocate device struct\n");
 664                 goto error_exit;
 665         }
 666
 667         snprintf(pmd->name, sizeof(pmd->name), "%s", tap_name);
 668
 669         pmd->nb_queues = RTE_PMD_TAP_MAX_QUEUES;
 670
 671         /* Setup some default values */
 672         data->dev_private = pmd;
 673         data->port_id = dev->data->port_id;
 674         data->dev_flags = RTE_ETH_DEV_DETACHABLE;
 675         data->kdrv = RTE_KDRV_NONE;
 676         data->drv_name = pmd_tap_drv.driver.name;
 677         data->numa_node = numa_node;
 678
 679         data->dev_link = pmd_link;
 680         data->mac_addrs = &pmd->eth_addr;
 681         data->nb_rx_queues = pmd->nb_queues;
 682         data->nb_tx_queues = pmd->nb_queues;
 683
 684         dev->data = data;
 685         dev->dev_ops = &ops;
 686         dev->driver = NULL;
 687         dev->rx_pkt_burst = pmd_rx_burst;
 688         dev->tx_pkt_burst = pmd_tx_burst;
 689         snprintf(dev->data->name, sizeof(dev->data->name), "%s", name);
 690
 691         /* Presetup the fds to -1 as being not valid */
 692         for (i = 0; i < RTE_PMD_TAP_MAX_QUEUES; i++) {
 693                 pmd->rxq[i].fd = -1;
 694                 pmd->txq[i].fd = -1;
 695         }
 696
 697         return 0;
 698
 699 error_exit:
 700         RTE_LOG(DEBUG, PMD, "TAP Unable to initialize %s\n", name);
 701
 702         rte_free(data);
 703         rte_free(pmd);
 704
 705         rte_eth_dev_release_port(dev);
 706
 707         return -EINVAL;
 708 }
 709
 710 static int
 711 set_interface_name(const char *key __rte_unused,
 712                    const char *value,
 713                    void *extra_args)
 714 {
 715         char *name = (char *)extra_args;
 716
 717         if (value)
 718                 snprintf(name, RTE_ETH_NAME_MAX_LEN - 1, "%s", value);
 719         else
 720                 snprintf(name, RTE_ETH_NAME_MAX_LEN - 1, "%s%d",
 721                          DEFAULT_TAP_NAME, (tap_unit - 1));
 722
 723         return 0;
 724 }
 725
 726 static int
 727 set_interface_speed(const char *key __rte_unused,
 728                     const char *value,
 729                     void *extra_args)
 730 {
 731         *(int *)extra_args = (value) ? atoi(value) : ETH_SPEED_NUM_10G;
 732
 733         return 0;
 734 }
 735
 736 /* Open a TAP interface device.
 737  */
 738 static int
 739 rte_pmd_tap_probe(const char *name, const char *params)
 740 {
 741         int ret;
 742         struct rte_kvargs *kvlist = NULL;
 743         int speed;
 744         char tap_name[RTE_ETH_NAME_MAX_LEN];
 745
 746         speed = ETH_SPEED_NUM_10G;
 747         snprintf(tap_name, sizeof(tap_name), "%s%d",
 748                  DEFAULT_TAP_NAME, tap_unit++);
 749
 750         if (params && (params[0] != '\0')) {
 751                 RTE_LOG(DEBUG, PMD, "paramaters (%s)\n", params);
 752
 753                 kvlist = rte_kvargs_parse(params, valid_arguments);
 754                 if (kvlist) {
 755                         if (rte_kvargs_count(kvlist, ETH_TAP_SPEED_ARG) == 1) {
 756                                 ret = rte_kvargs_process(kvlist,
 757                                                          ETH_TAP_SPEED_ARG,
 758                                                          &set_interface_speed,
 759                                                          &speed);
 760                                 if (ret == -1)
 761                                         goto leave;
 762                         }
 763
 764                         if (rte_kvargs_count(kvlist, ETH_TAP_IFACE_ARG) == 1) {
 765                                 ret = rte_kvargs_process(kvlist,
 766                                                          ETH_TAP_IFACE_ARG,
 767                                                          &set_interface_name,
 768                                                          tap_name);
 769                                 if (ret == -1)
 770                                         goto leave;
 771                         }
 772                 }
 773         }
 774         pmd_link.link_speed = speed;
 775
 776         RTE_LOG(NOTICE, PMD, "Initializing pmd_tap for %s as %s\n",
 777                 name, tap_name);
 778
 779         ret = eth_dev_tap_create(name, tap_name);
 780
 781 leave:
 782         if (ret == -1) {
 783                 RTE_LOG(ERR, PMD, "Failed to create pmd for %s as %s\n",
 784                         name, tap_name);
 785                 tap_unit--;             /* Restore the unit number */
 786         }
 787         rte_kvargs_free(kvlist);
 788
 789         return ret;
 790 }
 791
 792 /* detach a TAP device.
 793  */
 794 static int
 795 rte_pmd_tap_remove(const char *name)
 796 {
 797         struct rte_eth_dev *eth_dev = NULL;
 798         struct pmd_internals *internals;
 799         int i;
 800
 801         RTE_LOG(DEBUG, PMD, "Closing TUN/TAP Ethernet device on numa %u\n",
 802                 rte_socket_id());
 803
 804         /* find the ethdev entry */
 805         eth_dev = rte_eth_dev_allocated(name);
 806         if (!eth_dev)
 807                 return 0;
 808
 809         internals = eth_dev->data->dev_private;
 810         for (i = 0; i < internals->nb_queues; i++)
 811                 if (internals->rxq[i].fd != -1)
 812                         close(internals->rxq[i].fd);
 813
 814         rte_free(eth_dev->data->dev_private);
 815         rte_free(eth_dev->data);
 816
 817         rte_eth_dev_release_port(eth_dev);
 818
 819         return 0;
 820 }
 821
 822 static struct rte_vdev_driver pmd_tap_drv = {
 823         .probe = rte_pmd_tap_probe,
 824         .remove = rte_pmd_tap_remove,
 825 };
 826 RTE_PMD_REGISTER_VDEV(net_tap, pmd_tap_drv);
 827 RTE_PMD_REGISTER_ALIAS(net_tap, eth_tap);
 828 RTE_PMD_REGISTER_PARAM_STRING(net_tap, "iface=<string>,speed=N");