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