drivers/net/af_xdp/rte_eth_af_xdp.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2019 Intel Corporation.
   3  */
   4 #include <unistd.h>
   5 #include <errno.h>
   6 #include <stdlib.h>
   7 #include <string.h>
   8 #include <netinet/in.h>
   9 #include <net/if.h>
  10 #include <sys/socket.h>
  11 #include <sys/ioctl.h>
  12 #include <linux/if_ether.h>
  13 #include <linux/if_xdp.h>
  14 #include <linux/if_link.h>
  15 #include "af_xdp_deps.h"
  16 #include <bpf/xsk.h>
  17
  18 #include <rte_ethdev.h>
  19 #include <rte_ethdev_driver.h>
  20 #include <rte_ethdev_vdev.h>
  21 #include <rte_kvargs.h>
  22 #include <rte_bus_vdev.h>
  23 #include <rte_string_fns.h>
  24 #include <rte_branch_prediction.h>
  25 #include <rte_common.h>
  26 #include <rte_config.h>
  27 #include <rte_dev.h>
  28 #include <rte_eal.h>
  29 #include <rte_ether.h>
  30 #include <rte_lcore.h>
  31 #include <rte_log.h>
  32 #include <rte_memory.h>
  33 #include <rte_memzone.h>
  34 #include <rte_mbuf.h>
  35 #include <rte_malloc.h>
  36 #include <rte_ring.h>
  37
  38 #ifndef SOL_XDP
  39 #define SOL_XDP 283
  40 #endif
  41
  42 #ifndef AF_XDP
  43 #define AF_XDP 44
  44 #endif
  45
  46 #ifndef PF_XDP
  47 #define PF_XDP AF_XDP
  48 #endif
  49
  50 static int af_xdp_logtype;
  51
  52 #define AF_XDP_LOG(level, fmt, args...)                 \
  53         rte_log(RTE_LOG_ ## level, af_xdp_logtype,      \
  54                 "%s(): " fmt, __func__, ##args)
  55
  56 #define ETH_AF_XDP_FRAME_SIZE           XSK_UMEM__DEFAULT_FRAME_SIZE
  57 #define ETH_AF_XDP_NUM_BUFFERS          4096
  58 #define ETH_AF_XDP_DATA_HEADROOM        0
  59 #define ETH_AF_XDP_DFLT_NUM_DESCS       XSK_RING_CONS__DEFAULT_NUM_DESCS
  60 #define ETH_AF_XDP_DFLT_QUEUE_IDX       0
  61
  62 #define ETH_AF_XDP_RX_BATCH_SIZE        32
  63 #define ETH_AF_XDP_TX_BATCH_SIZE        32
  64
  65 #define ETH_AF_XDP_MAX_QUEUE_PAIRS     16
  66
  67 struct xsk_umem_info {
  68         struct xsk_ring_prod fq;
  69         struct xsk_ring_cons cq;
  70         struct xsk_umem *umem;
  71         struct rte_ring *buf_ring;
  72         const struct rte_memzone *mz;
  73 };
  74
  75 struct rx_stats {
  76         uint64_t rx_pkts;
  77         uint64_t rx_bytes;
  78         uint64_t rx_dropped;
  79 };
  80
  81 struct pkt_rx_queue {
  82         struct xsk_ring_cons rx;
  83         struct xsk_umem_info *umem;
  84         struct xsk_socket *xsk;
  85         struct rte_mempool *mb_pool;
  86
  87         struct rx_stats stats;
  88
  89         struct pkt_tx_queue *pair;
  90         uint16_t queue_idx;
  91 };
  92
  93 struct tx_stats {
  94         uint64_t tx_pkts;
  95         uint64_t err_pkts;
  96         uint64_t tx_bytes;
  97 };
  98
  99 struct pkt_tx_queue {
 100         struct xsk_ring_prod tx;
 101
 102         struct tx_stats stats;
 103
 104         struct pkt_rx_queue *pair;
 105         uint16_t queue_idx;
 106 };
 107
 108 struct pmd_internals {
 109         int if_index;
 110         char if_name[IFNAMSIZ];
 111         uint16_t queue_idx;
 112         struct ether_addr eth_addr;
 113         struct xsk_umem_info *umem;
 114         struct rte_mempool *mb_pool_share;
 115
 116         struct pkt_rx_queue rx_queues[ETH_AF_XDP_MAX_QUEUE_PAIRS];
 117         struct pkt_tx_queue tx_queues[ETH_AF_XDP_MAX_QUEUE_PAIRS];
 118 };
 119
 120 #define ETH_AF_XDP_IFACE_ARG                    "iface"
 121 #define ETH_AF_XDP_QUEUE_IDX_ARG                "queue"
 122
 123 static const char * const valid_arguments[] = {
 124         ETH_AF_XDP_IFACE_ARG,
 125         ETH_AF_XDP_QUEUE_IDX_ARG,
 126         NULL
 127 };
 128
 129 static const struct rte_eth_link pmd_link = {
 130         .link_speed = ETH_SPEED_NUM_10G,
 131         .link_duplex = ETH_LINK_FULL_DUPLEX,
 132         .link_status = ETH_LINK_DOWN,
 133         .link_autoneg = ETH_LINK_AUTONEG
 134 };
 135
 136 static inline int
 137 reserve_fill_queue(struct xsk_umem_info *umem, uint16_t reserve_size)
 138 {
 139         struct xsk_ring_prod *fq = &umem->fq;
 140         void *addrs[reserve_size];
 141         uint32_t idx;
 142         uint16_t i;
 143
 144         if (rte_ring_dequeue_bulk(umem->buf_ring, addrs, reserve_size, NULL)
 145                     != reserve_size) {
 146                 AF_XDP_LOG(DEBUG, "Failed to get enough buffers for fq.\n");
 147                 return -1;
 148         }
 149
 150         if (unlikely(!xsk_ring_prod__reserve(fq, reserve_size, &idx))) {
 151                 AF_XDP_LOG(DEBUG, "Failed to reserve enough fq descs.\n");
 152                 rte_ring_enqueue_bulk(umem->buf_ring, addrs,
 153                                 reserve_size, NULL);
 154                 return -1;
 155         }
 156
 157         for (i = 0; i < reserve_size; i++) {
 158                 __u64 *fq_addr;
 159
 160                 fq_addr = xsk_ring_prod__fill_addr(fq, idx++);
 161                 *fq_addr = (uint64_t)addrs[i];
 162         }
 163
 164         xsk_ring_prod__submit(fq, reserve_size);
 165
 166         return 0;
 167 }
 168
 169 static uint16_t
 170 eth_af_xdp_rx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 171 {
 172         struct pkt_rx_queue *rxq = queue;
 173         struct xsk_ring_cons *rx = &rxq->rx;
 174         struct xsk_umem_info *umem = rxq->umem;
 175         struct xsk_ring_prod *fq = &umem->fq;
 176         uint32_t idx_rx = 0;
 177         uint32_t free_thresh = fq->size >> 1;
 178         struct rte_mbuf *mbufs[ETH_AF_XDP_RX_BATCH_SIZE];
 179         unsigned long dropped = 0;
 180         unsigned long rx_bytes = 0;
 181         int rcvd, i;
 182
 183         nb_pkts = RTE_MIN(nb_pkts, ETH_AF_XDP_RX_BATCH_SIZE);
 184
 185         if (unlikely(rte_pktmbuf_alloc_bulk(rxq->mb_pool, mbufs, nb_pkts) != 0))
 186                 return 0;
 187
 188         rcvd = xsk_ring_cons__peek(rx, nb_pkts, &idx_rx);
 189         if (rcvd == 0)
 190                 goto out;
 191
 192         if (xsk_prod_nb_free(fq, free_thresh) >= free_thresh)
 193                 (void)reserve_fill_queue(umem, ETH_AF_XDP_RX_BATCH_SIZE);
 194
 195         for (i = 0; i < rcvd; i++) {
 196                 const struct xdp_desc *desc;
 197                 uint64_t addr;
 198                 uint32_t len;
 199                 void *pkt;
 200
 201                 desc = xsk_ring_cons__rx_desc(rx, idx_rx++);
 202                 addr = desc->addr;
 203                 len = desc->len;
 204                 pkt = xsk_umem__get_data(rxq->umem->mz->addr, addr);
 205
 206                 rte_memcpy(rte_pktmbuf_mtod(mbufs[i], void *), pkt, len);
 207                 rte_pktmbuf_pkt_len(mbufs[i]) = len;
 208                 rte_pktmbuf_data_len(mbufs[i]) = len;
 209                 rx_bytes += len;
 210                 bufs[i] = mbufs[i];
 211
 212                 rte_ring_enqueue(umem->buf_ring, (void *)addr);
 213         }
 214
 215         xsk_ring_cons__release(rx, rcvd);
 216
 217         /* statistics */
 218         rxq->stats.rx_pkts += (rcvd - dropped);
 219         rxq->stats.rx_bytes += rx_bytes;
 220
 221 out:
 222         if (rcvd != nb_pkts)
 223                 rte_mempool_put_bulk(rxq->mb_pool, (void **)&mbufs[rcvd],
 224                                      nb_pkts - rcvd);
 225
 226         return rcvd;
 227 }
 228
 229 static void
 230 pull_umem_cq(struct xsk_umem_info *umem, int size)
 231 {
 232         struct xsk_ring_cons *cq = &umem->cq;
 233         size_t i, n;
 234         uint32_t idx_cq = 0;
 235
 236         n = xsk_ring_cons__peek(cq, size, &idx_cq);
 237
 238         for (i = 0; i < n; i++) {
 239                 uint64_t addr;
 240                 addr = *xsk_ring_cons__comp_addr(cq, idx_cq++);
 241                 rte_ring_enqueue(umem->buf_ring, (void *)addr);
 242         }
 243
 244         xsk_ring_cons__release(cq, n);
 245 }
 246
 247 static void
 248 kick_tx(struct pkt_tx_queue *txq)
 249 {
 250         struct xsk_umem_info *umem = txq->pair->umem;
 251
 252         while (send(xsk_socket__fd(txq->pair->xsk), NULL,
 253                       0, MSG_DONTWAIT) < 0) {
 254                 /* some thing unexpected */
 255                 if (errno != EBUSY && errno != EAGAIN && errno != EINTR)
 256                         break;
 257
 258                 /* pull from completion queue to leave more space */
 259                 if (errno == EAGAIN)
 260                         pull_umem_cq(umem, ETH_AF_XDP_TX_BATCH_SIZE);
 261         }
 262         pull_umem_cq(umem, ETH_AF_XDP_TX_BATCH_SIZE);
 263 }
 264
 265 static uint16_t
 266 eth_af_xdp_tx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 267 {
 268         struct pkt_tx_queue *txq = queue;
 269         struct xsk_umem_info *umem = txq->pair->umem;
 270         struct rte_mbuf *mbuf;
 271         void *addrs[ETH_AF_XDP_TX_BATCH_SIZE];
 272         unsigned long tx_bytes = 0;
 273         int i;
 274         uint32_t idx_tx;
 275
 276         nb_pkts = RTE_MIN(nb_pkts, ETH_AF_XDP_TX_BATCH_SIZE);
 277
 278         pull_umem_cq(umem, nb_pkts);
 279
 280         nb_pkts = rte_ring_dequeue_bulk(umem->buf_ring, addrs,
 281                                         nb_pkts, NULL);
 282         if (nb_pkts == 0)
 283                 return 0;
 284
 285         if (xsk_ring_prod__reserve(&txq->tx, nb_pkts, &idx_tx) != nb_pkts) {
 286                 kick_tx(txq);
 287                 rte_ring_enqueue_bulk(umem->buf_ring, addrs, nb_pkts, NULL);
 288                 return 0;
 289         }
 290
 291         for (i = 0; i < nb_pkts; i++) {
 292                 struct xdp_desc *desc;
 293                 void *pkt;
 294
 295                 desc = xsk_ring_prod__tx_desc(&txq->tx, idx_tx + i);
 296                 mbuf = bufs[i];
 297
 298                 desc->addr = (uint64_t)addrs[i];
 299                 desc->len = mbuf->pkt_len;
 300                 pkt = xsk_umem__get_data(umem->mz->addr,
 301                                          desc->addr);
 302                 rte_memcpy(pkt, rte_pktmbuf_mtod(mbuf, void *),
 303                            desc->len);
 304                 tx_bytes += mbuf->pkt_len;
 305
 306                 rte_pktmbuf_free(mbuf);
 307         }
 308
 309         xsk_ring_prod__submit(&txq->tx, nb_pkts);
 310
 311         kick_tx(txq);
 312
 313         txq->stats.tx_pkts += nb_pkts;
 314         txq->stats.tx_bytes += tx_bytes;
 315
 316         return nb_pkts;
 317 }
 318
 319 static int
 320 eth_dev_start(struct rte_eth_dev *dev)
 321 {
 322         dev->data->dev_link.link_status = ETH_LINK_UP;
 323
 324         return 0;
 325 }
 326
 327 /* This function gets called when the current port gets stopped. */
 328 static void
 329 eth_dev_stop(struct rte_eth_dev *dev)
 330 {
 331         dev->data->dev_link.link_status = ETH_LINK_DOWN;
 332 }
 333
 334 static int
 335 eth_dev_configure(struct rte_eth_dev *dev)
 336 {
 337         /* rx/tx must be paired */
 338         if (dev->data->nb_rx_queues != dev->data->nb_tx_queues)
 339                 return -EINVAL;
 340
 341         return 0;
 342 }
 343
 344 static void
 345 eth_dev_info(struct rte_eth_dev *dev, struct rte_eth_dev_info *dev_info)
 346 {
 347         struct pmd_internals *internals = dev->data->dev_private;
 348
 349         dev_info->if_index = internals->if_index;
 350         dev_info->max_mac_addrs = 1;
 351         dev_info->max_rx_pktlen = ETH_FRAME_LEN;
 352         dev_info->max_rx_queues = 1;
 353         dev_info->max_tx_queues = 1;
 354
 355         dev_info->min_mtu = ETHER_MIN_MTU;
 356         dev_info->max_mtu = ETH_AF_XDP_FRAME_SIZE - ETH_AF_XDP_DATA_HEADROOM;
 357
 358         dev_info->default_rxportconf.nb_queues = 1;
 359         dev_info->default_txportconf.nb_queues = 1;
 360         dev_info->default_rxportconf.ring_size = ETH_AF_XDP_DFLT_NUM_DESCS;
 361         dev_info->default_txportconf.ring_size = ETH_AF_XDP_DFLT_NUM_DESCS;
 362 }
 363
 364 static int
 365 eth_stats_get(struct rte_eth_dev *dev, struct rte_eth_stats *stats)
 366 {
 367         struct pmd_internals *internals = dev->data->dev_private;
 368         struct xdp_statistics xdp_stats;
 369         struct pkt_rx_queue *rxq;
 370         socklen_t optlen;
 371         int i, ret;
 372
 373         for (i = 0; i < dev->data->nb_rx_queues; i++) {
 374                 optlen = sizeof(struct xdp_statistics);
 375                 rxq = &internals->rx_queues[i];
 376                 stats->q_ipackets[i] = internals->rx_queues[i].stats.rx_pkts;
 377                 stats->q_ibytes[i] = internals->rx_queues[i].stats.rx_bytes;
 378
 379                 stats->q_opackets[i] = internals->tx_queues[i].stats.tx_pkts;
 380                 stats->q_obytes[i] = internals->tx_queues[i].stats.tx_bytes;
 381
 382                 stats->ipackets += stats->q_ipackets[i];
 383                 stats->ibytes += stats->q_ibytes[i];
 384                 stats->imissed += internals->rx_queues[i].stats.rx_dropped;
 385                 ret = getsockopt(xsk_socket__fd(rxq->xsk), SOL_XDP,
 386                                 XDP_STATISTICS, &xdp_stats, &optlen);
 387                 if (ret != 0) {
 388                         AF_XDP_LOG(ERR, "getsockopt() failed for XDP_STATISTICS.\n");
 389                         return -1;
 390                 }
 391                 stats->imissed += xdp_stats.rx_dropped;
 392
 393                 stats->opackets += stats->q_opackets[i];
 394                 stats->oerrors += internals->tx_queues[i].stats.err_pkts;
 395                 stats->obytes += stats->q_obytes[i];
 396         }
 397
 398         return 0;
 399 }
 400
 401 static void
 402 eth_stats_reset(struct rte_eth_dev *dev)
 403 {
 404         struct pmd_internals *internals = dev->data->dev_private;
 405         int i;
 406
 407         for (i = 0; i < ETH_AF_XDP_MAX_QUEUE_PAIRS; i++) {
 408                 memset(&internals->rx_queues[i].stats, 0,
 409                                         sizeof(struct rx_stats));
 410                 memset(&internals->tx_queues[i].stats, 0,
 411                                         sizeof(struct tx_stats));
 412         }
 413 }
 414
 415 static void
 416 remove_xdp_program(struct pmd_internals *internals)
 417 {
 418         uint32_t curr_prog_id = 0;
 419
 420         if (bpf_get_link_xdp_id(internals->if_index, &curr_prog_id,
 421                                 XDP_FLAGS_UPDATE_IF_NOEXIST)) {
 422                 AF_XDP_LOG(ERR, "bpf_get_link_xdp_id failed\n");
 423                 return;
 424         }
 425         bpf_set_link_xdp_fd(internals->if_index, -1,
 426                         XDP_FLAGS_UPDATE_IF_NOEXIST);
 427 }
 428
 429 static void
 430 eth_dev_close(struct rte_eth_dev *dev)
 431 {
 432         struct pmd_internals *internals = dev->data->dev_private;
 433         struct pkt_rx_queue *rxq;
 434         int i;
 435
 436         AF_XDP_LOG(INFO, "Closing AF_XDP ethdev on numa socket %u\n",
 437                 rte_socket_id());
 438
 439         for (i = 0; i < ETH_AF_XDP_MAX_QUEUE_PAIRS; i++) {
 440                 rxq = &internals->rx_queues[i];
 441                 if (rxq->umem == NULL)
 442                         break;
 443                 xsk_socket__delete(rxq->xsk);
 444         }
 445
 446         (void)xsk_umem__delete(internals->umem->umem);
 447         remove_xdp_program(internals);
 448 }
 449
 450 static void
 451 eth_queue_release(void *q __rte_unused)
 452 {
 453 }
 454
 455 static int
 456 eth_link_update(struct rte_eth_dev *dev __rte_unused,
 457                 int wait_to_complete __rte_unused)
 458 {
 459         return 0;
 460 }
 461
 462 static void
 463 xdp_umem_destroy(struct xsk_umem_info *umem)
 464 {
 465         rte_memzone_free(umem->mz);
 466         umem->mz = NULL;
 467
 468         rte_ring_free(umem->buf_ring);
 469         umem->buf_ring = NULL;
 470
 471         rte_free(umem);
 472         umem = NULL;
 473 }
 474
 475 static struct
 476 xsk_umem_info *xdp_umem_configure(void)
 477 {
 478         struct xsk_umem_info *umem;
 479         const struct rte_memzone *mz;
 480         struct xsk_umem_config usr_config = {
 481                 .fill_size = ETH_AF_XDP_DFLT_NUM_DESCS,
 482                 .comp_size = ETH_AF_XDP_DFLT_NUM_DESCS,
 483                 .frame_size = ETH_AF_XDP_FRAME_SIZE,
 484                 .frame_headroom = ETH_AF_XDP_DATA_HEADROOM };
 485         int ret;
 486         uint64_t i;
 487
 488         umem = rte_zmalloc_socket("umem", sizeof(*umem), 0, rte_socket_id());
 489         if (umem == NULL) {
 490                 AF_XDP_LOG(ERR, "Failed to allocate umem info");
 491                 return NULL;
 492         }
 493
 494         umem->buf_ring = rte_ring_create("af_xdp_ring",
 495                                          ETH_AF_XDP_NUM_BUFFERS,
 496                                          rte_socket_id(),
 497                                          0x0);
 498         if (umem->buf_ring == NULL) {
 499                 AF_XDP_LOG(ERR, "Failed to create rte_ring\n");
 500                 goto err;
 501         }
 502
 503         for (i = 0; i < ETH_AF_XDP_NUM_BUFFERS; i++)
 504                 rte_ring_enqueue(umem->buf_ring,
 505                                  (void *)(i * ETH_AF_XDP_FRAME_SIZE +
 506                                           ETH_AF_XDP_DATA_HEADROOM));
 507
 508         mz = rte_memzone_reserve_aligned("af_xdp uemem",
 509                         ETH_AF_XDP_NUM_BUFFERS * ETH_AF_XDP_FRAME_SIZE,
 510                         rte_socket_id(), RTE_MEMZONE_IOVA_CONTIG,
 511                         getpagesize());
 512         if (mz == NULL) {
 513                 AF_XDP_LOG(ERR, "Failed to reserve memzone for af_xdp umem.\n");
 514                 goto err;
 515         }
 516
 517         ret = xsk_umem__create(&umem->umem, mz->addr,
 518                                ETH_AF_XDP_NUM_BUFFERS * ETH_AF_XDP_FRAME_SIZE,
 519                                &umem->fq, &umem->cq,
 520                                &usr_config);
 521
 522         if (ret) {
 523                 AF_XDP_LOG(ERR, "Failed to create umem");
 524                 goto err;
 525         }
 526         umem->mz = mz;
 527
 528         return umem;
 529
 530 err:
 531         xdp_umem_destroy(umem);
 532         return NULL;
 533 }
 534
 535 static int
 536 xsk_configure(struct pmd_internals *internals, struct pkt_rx_queue *rxq,
 537               int ring_size)
 538 {
 539         struct xsk_socket_config cfg;
 540         struct pkt_tx_queue *txq = rxq->pair;
 541         int ret = 0;
 542         int reserve_size;
 543
 544         rxq->umem = xdp_umem_configure();
 545         if (rxq->umem == NULL)
 546                 return -ENOMEM;
 547
 548         cfg.rx_size = ring_size;
 549         cfg.tx_size = ring_size;
 550         cfg.libbpf_flags = 0;
 551         cfg.xdp_flags = XDP_FLAGS_UPDATE_IF_NOEXIST;
 552         cfg.bind_flags = 0;
 553         ret = xsk_socket__create(&rxq->xsk, internals->if_name,
 554                         internals->queue_idx, rxq->umem->umem, &rxq->rx,
 555                         &txq->tx, &cfg);
 556         if (ret) {
 557                 AF_XDP_LOG(ERR, "Failed to create xsk socket.\n");
 558                 goto err;
 559         }
 560
 561         reserve_size = ETH_AF_XDP_DFLT_NUM_DESCS / 2;
 562         ret = reserve_fill_queue(rxq->umem, reserve_size);
 563         if (ret) {
 564                 xsk_socket__delete(rxq->xsk);
 565                 AF_XDP_LOG(ERR, "Failed to reserve fill queue.\n");
 566                 goto err;
 567         }
 568
 569         return 0;
 570
 571 err:
 572         xdp_umem_destroy(rxq->umem);
 573
 574         return ret;
 575 }
 576
 577 static void
 578 queue_reset(struct pmd_internals *internals, uint16_t queue_idx)
 579 {
 580         struct pkt_rx_queue *rxq = &internals->rx_queues[queue_idx];
 581         struct pkt_tx_queue *txq = rxq->pair;
 582
 583         memset(rxq, 0, sizeof(*rxq));
 584         memset(txq, 0, sizeof(*txq));
 585         rxq->pair = txq;
 586         txq->pair = rxq;
 587         rxq->queue_idx = queue_idx;
 588         txq->queue_idx = queue_idx;
 589 }
 590
 591 static int
 592 eth_rx_queue_setup(struct rte_eth_dev *dev,
 593                    uint16_t rx_queue_id,
 594                    uint16_t nb_rx_desc,
 595                    unsigned int socket_id __rte_unused,
 596                    const struct rte_eth_rxconf *rx_conf __rte_unused,
 597                    struct rte_mempool *mb_pool)
 598 {
 599         struct pmd_internals *internals = dev->data->dev_private;
 600         uint32_t buf_size, data_size;
 601         struct pkt_rx_queue *rxq;
 602         int ret;
 603
 604         rxq = &internals->rx_queues[rx_queue_id];
 605         queue_reset(internals, rx_queue_id);
 606
 607         /* Now get the space available for data in the mbuf */
 608         buf_size = rte_pktmbuf_data_room_size(mb_pool) -
 609                 RTE_PKTMBUF_HEADROOM;
 610         data_size = ETH_AF_XDP_FRAME_SIZE - ETH_AF_XDP_DATA_HEADROOM;
 611
 612         if (data_size > buf_size) {
 613                 AF_XDP_LOG(ERR, "%s: %d bytes will not fit in mbuf (%d bytes)\n",
 614                         dev->device->name, data_size, buf_size);
 615                 ret = -ENOMEM;
 616                 goto err;
 617         }
 618
 619         rxq->mb_pool = mb_pool;
 620
 621         if (xsk_configure(internals, rxq, nb_rx_desc)) {
 622                 AF_XDP_LOG(ERR, "Failed to configure xdp socket\n");
 623                 ret = -EINVAL;
 624                 goto err;
 625         }
 626
 627         internals->umem = rxq->umem;
 628
 629         dev->data->rx_queues[rx_queue_id] = rxq;
 630         return 0;
 631
 632 err:
 633         queue_reset(internals, rx_queue_id);
 634         return ret;
 635 }
 636
 637 static int
 638 eth_tx_queue_setup(struct rte_eth_dev *dev,
 639                    uint16_t tx_queue_id,
 640                    uint16_t nb_tx_desc __rte_unused,
 641                    unsigned int socket_id __rte_unused,
 642                    const struct rte_eth_txconf *tx_conf __rte_unused)
 643 {
 644         struct pmd_internals *internals = dev->data->dev_private;
 645         struct pkt_tx_queue *txq;
 646
 647         txq = &internals->tx_queues[tx_queue_id];
 648
 649         dev->data->tx_queues[tx_queue_id] = txq;
 650         return 0;
 651 }
 652
 653 static int
 654 eth_dev_mtu_set(struct rte_eth_dev *dev, uint16_t mtu)
 655 {
 656         struct pmd_internals *internals = dev->data->dev_private;
 657         struct ifreq ifr = { .ifr_mtu = mtu };
 658         int ret;
 659         int s;
 660
 661         s = socket(PF_INET, SOCK_DGRAM, 0);
 662         if (s < 0)
 663                 return -EINVAL;
 664
 665         strlcpy(ifr.ifr_name, internals->if_name, IFNAMSIZ);
 666         ret = ioctl(s, SIOCSIFMTU, &ifr);
 667         close(s);
 668
 669         return (ret < 0) ? -errno : 0;
 670 }
 671
 672 static void
 673 eth_dev_change_flags(char *if_name, uint32_t flags, uint32_t mask)
 674 {
 675         struct ifreq ifr;
 676         int s;
 677
 678         s = socket(PF_INET, SOCK_DGRAM, 0);
 679         if (s < 0)
 680                 return;
 681
 682         strlcpy(ifr.ifr_name, if_name, IFNAMSIZ);
 683         if (ioctl(s, SIOCGIFFLAGS, &ifr) < 0)
 684                 goto out;
 685         ifr.ifr_flags &= mask;
 686         ifr.ifr_flags |= flags;
 687         if (ioctl(s, SIOCSIFFLAGS, &ifr) < 0)
 688                 goto out;
 689 out:
 690         close(s);
 691 }
 692
 693 static void
 694 eth_dev_promiscuous_enable(struct rte_eth_dev *dev)
 695 {
 696         struct pmd_internals *internals = dev->data->dev_private;
 697
 698         eth_dev_change_flags(internals->if_name, IFF_PROMISC, ~0);
 699 }
 700
 701 static void
 702 eth_dev_promiscuous_disable(struct rte_eth_dev *dev)
 703 {
 704         struct pmd_internals *internals = dev->data->dev_private;
 705
 706         eth_dev_change_flags(internals->if_name, 0, ~IFF_PROMISC);
 707 }
 708
 709 static const struct eth_dev_ops ops = {
 710         .dev_start = eth_dev_start,
 711         .dev_stop = eth_dev_stop,
 712         .dev_close = eth_dev_close,
 713         .dev_configure = eth_dev_configure,
 714         .dev_infos_get = eth_dev_info,
 715         .mtu_set = eth_dev_mtu_set,
 716         .promiscuous_enable = eth_dev_promiscuous_enable,
 717         .promiscuous_disable = eth_dev_promiscuous_disable,
 718         .rx_queue_setup = eth_rx_queue_setup,
 719         .tx_queue_setup = eth_tx_queue_setup,
 720         .rx_queue_release = eth_queue_release,
 721         .tx_queue_release = eth_queue_release,
 722         .link_update = eth_link_update,
 723         .stats_get = eth_stats_get,
 724         .stats_reset = eth_stats_reset,
 725 };
 726
 727 /** parse integer from integer argument */
 728 static int
 729 parse_integer_arg(const char *key __rte_unused,
 730                   const char *value, void *extra_args)
 731 {
 732         int *i = (int *)extra_args;
 733         char *end;
 734
 735         *i = strtol(value, &end, 10);
 736         if (*i < 0) {
 737                 AF_XDP_LOG(ERR, "Argument has to be positive.\n");
 738                 return -EINVAL;
 739         }
 740
 741         return 0;
 742 }
 743
 744 /** parse name argument */
 745 static int
 746 parse_name_arg(const char *key __rte_unused,
 747                const char *value, void *extra_args)
 748 {
 749         char *name = extra_args;
 750
 751         if (strnlen(value, IFNAMSIZ) > IFNAMSIZ - 1) {
 752                 AF_XDP_LOG(ERR, "Invalid name %s, should be less than %u bytes.\n",
 753                            value, IFNAMSIZ);
 754                 return -EINVAL;
 755         }
 756
 757         strlcpy(name, value, IFNAMSIZ);
 758
 759         return 0;
 760 }
 761
 762 static int
 763 parse_parameters(struct rte_kvargs *kvlist,
 764                  char *if_name,
 765                  int *queue_idx)
 766 {
 767         int ret;
 768
 769         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_IFACE_ARG,
 770                                  &parse_name_arg, if_name);
 771         if (ret < 0)
 772                 goto free_kvlist;
 773
 774         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_QUEUE_IDX_ARG,
 775                                  &parse_integer_arg, queue_idx);
 776         if (ret < 0)
 777                 goto free_kvlist;
 778
 779 free_kvlist:
 780         rte_kvargs_free(kvlist);
 781         return ret;
 782 }
 783
 784 static int
 785 get_iface_info(const char *if_name,
 786                struct ether_addr *eth_addr,
 787                int *if_index)
 788 {
 789         struct ifreq ifr;
 790         int sock = socket(AF_INET, SOCK_DGRAM, IPPROTO_IP);
 791
 792         if (sock < 0)
 793                 return -1;
 794
 795         strlcpy(ifr.ifr_name, if_name, IFNAMSIZ);
 796         if (ioctl(sock, SIOCGIFINDEX, &ifr))
 797                 goto error;
 798
 799         *if_index = ifr.ifr_ifindex;
 800
 801         if (ioctl(sock, SIOCGIFHWADDR, &ifr))
 802                 goto error;
 803
 804         rte_memcpy(eth_addr, ifr.ifr_hwaddr.sa_data, ETHER_ADDR_LEN);
 805
 806         close(sock);
 807         return 0;
 808
 809 error:
 810         close(sock);
 811         return -1;
 812 }
 813
 814 static struct rte_eth_dev *
 815 init_internals(struct rte_vdev_device *dev,
 816                const char *if_name,
 817                int queue_idx)
 818 {
 819         const char *name = rte_vdev_device_name(dev);
 820         const unsigned int numa_node = dev->device.numa_node;
 821         struct pmd_internals *internals;
 822         struct rte_eth_dev *eth_dev;
 823         int ret;
 824         int i;
 825
 826         internals = rte_zmalloc_socket(name, sizeof(*internals), 0, numa_node);
 827         if (internals == NULL)
 828                 return NULL;
 829
 830         internals->queue_idx = queue_idx;
 831         strlcpy(internals->if_name, if_name, IFNAMSIZ);
 832
 833         for (i = 0; i < ETH_AF_XDP_MAX_QUEUE_PAIRS; i++) {
 834                 internals->tx_queues[i].pair = &internals->rx_queues[i];
 835                 internals->rx_queues[i].pair = &internals->tx_queues[i];
 836         }
 837
 838         ret = get_iface_info(if_name, &internals->eth_addr,
 839                              &internals->if_index);
 840         if (ret)
 841                 goto err;
 842
 843         eth_dev = rte_eth_vdev_allocate(dev, 0);
 844         if (eth_dev == NULL)
 845                 goto err;
 846
 847         eth_dev->data->dev_private = internals;
 848         eth_dev->data->dev_link = pmd_link;
 849         eth_dev->data->mac_addrs = &internals->eth_addr;
 850         eth_dev->dev_ops = &ops;
 851         eth_dev->rx_pkt_burst = eth_af_xdp_rx;
 852         eth_dev->tx_pkt_burst = eth_af_xdp_tx;
 853
 854         return eth_dev;
 855
 856 err:
 857         rte_free(internals);
 858         return NULL;
 859 }
 860
 861 static int
 862 rte_pmd_af_xdp_probe(struct rte_vdev_device *dev)
 863 {
 864         struct rte_kvargs *kvlist;
 865         char if_name[IFNAMSIZ] = {'\0'};
 866         int xsk_queue_idx = ETH_AF_XDP_DFLT_QUEUE_IDX;
 867         struct rte_eth_dev *eth_dev = NULL;
 868         const char *name;
 869
 870         AF_XDP_LOG(INFO, "Initializing pmd_af_xdp for %s\n",
 871                 rte_vdev_device_name(dev));
 872
 873         name = rte_vdev_device_name(dev);
 874         if (rte_eal_process_type() == RTE_PROC_SECONDARY &&
 875                 strlen(rte_vdev_device_args(dev)) == 0) {
 876                 eth_dev = rte_eth_dev_attach_secondary(name);
 877                 if (eth_dev == NULL) {
 878                         AF_XDP_LOG(ERR, "Failed to probe %s\n", name);
 879                         return -EINVAL;
 880                 }
 881                 eth_dev->dev_ops = &ops;
 882                 rte_eth_dev_probing_finish(eth_dev);
 883                 return 0;
 884         }
 885
 886         kvlist = rte_kvargs_parse(rte_vdev_device_args(dev), valid_arguments);
 887         if (kvlist == NULL) {
 888                 AF_XDP_LOG(ERR, "Invalid kvargs key\n");
 889                 return -EINVAL;
 890         }
 891
 892         if (dev->device.numa_node == SOCKET_ID_ANY)
 893                 dev->device.numa_node = rte_socket_id();
 894
 895         if (parse_parameters(kvlist, if_name, &xsk_queue_idx) < 0) {
 896                 AF_XDP_LOG(ERR, "Invalid kvargs value\n");
 897                 return -EINVAL;
 898         }
 899
 900         if (strlen(if_name) == 0) {
 901                 AF_XDP_LOG(ERR, "Network interface must be specified\n");
 902                 return -EINVAL;
 903         }
 904
 905         eth_dev = init_internals(dev, if_name, xsk_queue_idx);
 906         if (eth_dev == NULL) {
 907                 AF_XDP_LOG(ERR, "Failed to init internals\n");
 908                 return -1;
 909         }
 910
 911         rte_eth_dev_probing_finish(eth_dev);
 912
 913         return 0;
 914 }
 915
 916 static int
 917 rte_pmd_af_xdp_remove(struct rte_vdev_device *dev)
 918 {
 919         struct rte_eth_dev *eth_dev = NULL;
 920         struct pmd_internals *internals;
 921
 922         AF_XDP_LOG(INFO, "Removing AF_XDP ethdev on numa socket %u\n",
 923                 rte_socket_id());
 924
 925         if (dev == NULL)
 926                 return -1;
 927
 928         /* find the ethdev entry */
 929         eth_dev = rte_eth_dev_allocated(rte_vdev_device_name(dev));
 930         if (eth_dev == NULL)
 931                 return -1;
 932
 933         internals = eth_dev->data->dev_private;
 934
 935         rte_ring_free(internals->umem->buf_ring);
 936         rte_memzone_free(internals->umem->mz);
 937         rte_free(internals->umem);
 938
 939         rte_eth_dev_release_port(eth_dev);
 940
 941
 942         return 0;
 943 }
 944
 945 static struct rte_vdev_driver pmd_af_xdp_drv = {
 946         .probe = rte_pmd_af_xdp_probe,
 947         .remove = rte_pmd_af_xdp_remove,
 948 };
 949
 950 RTE_PMD_REGISTER_VDEV(net_af_xdp, pmd_af_xdp_drv);
 951 RTE_PMD_REGISTER_PARAM_STRING(net_af_xdp,
 952                               "iface=<string> "
 953                               "queue=<int> ");
 954
 955 RTE_INIT(af_xdp_init_log)
 956 {
 957         af_xdp_logtype = rte_log_register("pmd.net.af_xdp");
 958         if (af_xdp_logtype >= 0)
 959                 rte_log_set_level(af_xdp_logtype, RTE_LOG_NOTICE);
 960 }