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 <poll.h>
   9 #include <netinet/in.h>
  10 #include <net/if.h>
  11 #include <sys/socket.h>
  12 #include <sys/ioctl.h>
  13 #include <linux/if_ether.h>
  14 #include <linux/if_xdp.h>
  15 #include <linux/if_link.h>
  16 #include <linux/ethtool.h>
  17 #include <linux/sockios.h>
  18 #include "af_xdp_deps.h"
  19 #include <bpf/xsk.h>
  20
  21 #include <rte_ethdev.h>
  22 #include <rte_ethdev_driver.h>
  23 #include <rte_ethdev_vdev.h>
  24 #include <rte_kvargs.h>
  25 #include <rte_bus_vdev.h>
  26 #include <rte_string_fns.h>
  27 #include <rte_branch_prediction.h>
  28 #include <rte_common.h>
  29 #include <rte_dev.h>
  30 #include <rte_eal.h>
  31 #include <rte_ether.h>
  32 #include <rte_lcore.h>
  33 #include <rte_log.h>
  34 #include <rte_memory.h>
  35 #include <rte_memzone.h>
  36 #include <rte_mbuf.h>
  37 #include <rte_malloc.h>
  38 #include <rte_ring.h>
  39
  40 #ifndef SOL_XDP
  41 #define SOL_XDP 283
  42 #endif
  43
  44 #ifndef AF_XDP
  45 #define AF_XDP 44
  46 #endif
  47
  48 #ifndef PF_XDP
  49 #define PF_XDP AF_XDP
  50 #endif
  51
  52 static int af_xdp_logtype;
  53
  54 #define AF_XDP_LOG(level, fmt, args...)                 \
  55         rte_log(RTE_LOG_ ## level, af_xdp_logtype,      \
  56                 "%s(): " fmt, __func__, ##args)
  57
  58 #define ETH_AF_XDP_FRAME_SIZE           2048
  59 #define ETH_AF_XDP_NUM_BUFFERS          4096
  60 #ifdef XDP_UMEM_UNALIGNED_CHUNK_FLAG
  61 #define ETH_AF_XDP_MBUF_OVERHEAD        128 /* sizeof(struct rte_mbuf) */
  62 #define ETH_AF_XDP_DATA_HEADROOM \
  63         (ETH_AF_XDP_MBUF_OVERHEAD + RTE_PKTMBUF_HEADROOM)
  64 #else
  65 #define ETH_AF_XDP_DATA_HEADROOM        0
  66 #endif
  67 #define ETH_AF_XDP_DFLT_NUM_DESCS       XSK_RING_CONS__DEFAULT_NUM_DESCS
  68 #define ETH_AF_XDP_DFLT_START_QUEUE_IDX 0
  69 #define ETH_AF_XDP_DFLT_QUEUE_COUNT     1
  70
  71 #define ETH_AF_XDP_RX_BATCH_SIZE        32
  72 #define ETH_AF_XDP_TX_BATCH_SIZE        32
  73
  74
  75 struct xsk_umem_info {
  76         struct xsk_ring_prod fq;
  77         struct xsk_ring_cons cq;
  78         struct xsk_umem *umem;
  79         struct rte_ring *buf_ring;
  80         const struct rte_memzone *mz;
  81         struct rte_mempool *mb_pool;
  82         void *buffer;
  83 };
  84
  85 struct rx_stats {
  86         uint64_t rx_pkts;
  87         uint64_t rx_bytes;
  88         uint64_t rx_dropped;
  89 };
  90
  91 struct pkt_rx_queue {
  92         struct xsk_ring_cons rx;
  93         struct xsk_umem_info *umem;
  94         struct xsk_socket *xsk;
  95         struct rte_mempool *mb_pool;
  96
  97         struct rx_stats stats;
  98
  99         struct pkt_tx_queue *pair;
 100         struct pollfd fds[1];
 101         int xsk_queue_idx;
 102 };
 103
 104 struct tx_stats {
 105         uint64_t tx_pkts;
 106         uint64_t tx_bytes;
 107         uint64_t tx_dropped;
 108 };
 109
 110 struct pkt_tx_queue {
 111         struct xsk_ring_prod tx;
 112         struct xsk_umem_info *umem;
 113
 114         struct tx_stats stats;
 115
 116         struct pkt_rx_queue *pair;
 117         int xsk_queue_idx;
 118 };
 119
 120 struct pmd_internals {
 121         int if_index;
 122         char if_name[IFNAMSIZ];
 123         int start_queue_idx;
 124         int queue_cnt;
 125         int max_queue_cnt;
 126         int combined_queue_cnt;
 127
 128         struct rte_ether_addr eth_addr;
 129
 130         struct pkt_rx_queue *rx_queues;
 131         struct pkt_tx_queue *tx_queues;
 132 };
 133
 134 #define ETH_AF_XDP_IFACE_ARG                    "iface"
 135 #define ETH_AF_XDP_START_QUEUE_ARG              "start_queue"
 136 #define ETH_AF_XDP_QUEUE_COUNT_ARG              "queue_count"
 137
 138 static const char * const valid_arguments[] = {
 139         ETH_AF_XDP_IFACE_ARG,
 140         ETH_AF_XDP_START_QUEUE_ARG,
 141         ETH_AF_XDP_QUEUE_COUNT_ARG,
 142         NULL
 143 };
 144
 145 static const struct rte_eth_link pmd_link = {
 146         .link_speed = ETH_SPEED_NUM_10G,
 147         .link_duplex = ETH_LINK_FULL_DUPLEX,
 148         .link_status = ETH_LINK_DOWN,
 149         .link_autoneg = ETH_LINK_AUTONEG
 150 };
 151
 152 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 153 static inline int
 154 reserve_fill_queue_zc(struct xsk_umem_info *umem, uint16_t reserve_size,
 155                       struct rte_mbuf **bufs)
 156 {
 157         struct xsk_ring_prod *fq = &umem->fq;
 158         uint32_t idx;
 159         uint16_t i;
 160
 161         if (unlikely(!xsk_ring_prod__reserve(fq, reserve_size, &idx))) {
 162                 for (i = 0; i < reserve_size; i++)
 163                         rte_pktmbuf_free(bufs[i]);
 164                 AF_XDP_LOG(DEBUG, "Failed to reserve enough fq descs.\n");
 165                 return -1;
 166         }
 167
 168         for (i = 0; i < reserve_size; i++) {
 169                 __u64 *fq_addr;
 170                 uint64_t addr;
 171
 172                 fq_addr = xsk_ring_prod__fill_addr(fq, idx++);
 173                 addr = (uint64_t)bufs[i] - (uint64_t)umem->buffer;
 174                 *fq_addr = addr;
 175         }
 176
 177         xsk_ring_prod__submit(fq, reserve_size);
 178
 179         return 0;
 180 }
 181 #else
 182 static inline int
 183 reserve_fill_queue_cp(struct xsk_umem_info *umem, uint16_t reserve_size,
 184                       struct rte_mbuf **bufs __rte_unused)
 185 {
 186         struct xsk_ring_prod *fq = &umem->fq;
 187         void *addrs[reserve_size];
 188         uint32_t idx;
 189         uint16_t i;
 190
 191         if (rte_ring_dequeue_bulk(umem->buf_ring, addrs, reserve_size, NULL)
 192                     != reserve_size) {
 193                 AF_XDP_LOG(DEBUG, "Failed to get enough buffers for fq.\n");
 194                 return -1;
 195         }
 196
 197         if (unlikely(!xsk_ring_prod__reserve(fq, reserve_size, &idx))) {
 198                 AF_XDP_LOG(DEBUG, "Failed to reserve enough fq descs.\n");
 199                 rte_ring_enqueue_bulk(umem->buf_ring, addrs,
 200                                 reserve_size, NULL);
 201                 return -1;
 202         }
 203
 204         for (i = 0; i < reserve_size; i++) {
 205                 __u64 *fq_addr;
 206
 207                 fq_addr = xsk_ring_prod__fill_addr(fq, idx++);
 208                 *fq_addr = (uint64_t)addrs[i];
 209         }
 210
 211         xsk_ring_prod__submit(fq, reserve_size);
 212
 213         return 0;
 214 }
 215 #endif
 216
 217 static inline int
 218 reserve_fill_queue(struct xsk_umem_info *umem, uint16_t reserve_size,
 219                    struct rte_mbuf **bufs)
 220 {
 221 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 222         return reserve_fill_queue_zc(umem, reserve_size, bufs);
 223 #else
 224         return reserve_fill_queue_cp(umem, reserve_size, bufs);
 225 #endif
 226 }
 227
 228 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 229 static uint16_t
 230 af_xdp_rx_zc(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 231 {
 232         struct pkt_rx_queue *rxq = queue;
 233         struct xsk_ring_cons *rx = &rxq->rx;
 234         struct xsk_umem_info *umem = rxq->umem;
 235         uint32_t idx_rx = 0;
 236         unsigned long rx_bytes = 0;
 237         int rcvd, i;
 238         struct rte_mbuf *fq_bufs[ETH_AF_XDP_RX_BATCH_SIZE];
 239
 240         /* allocate bufs for fill queue replenishment after rx */
 241         if (rte_pktmbuf_alloc_bulk(umem->mb_pool, fq_bufs, nb_pkts)) {
 242                 AF_XDP_LOG(DEBUG,
 243                         "Failed to get enough buffers for fq.\n");
 244                 return -1;
 245         }
 246
 247         rcvd = xsk_ring_cons__peek(rx, nb_pkts, &idx_rx);
 248
 249         if (rcvd == 0) {
 250 #if defined(XDP_USE_NEED_WAKEUP)
 251                 if (xsk_ring_prod__needs_wakeup(&umem->fq))
 252                         (void)poll(rxq->fds, 1, 1000);
 253 #endif
 254
 255                 goto out;
 256         }
 257
 258         for (i = 0; i < rcvd; i++) {
 259                 const struct xdp_desc *desc;
 260                 uint64_t addr;
 261                 uint32_t len;
 262                 uint64_t offset;
 263
 264                 desc = xsk_ring_cons__rx_desc(rx, idx_rx++);
 265                 addr = desc->addr;
 266                 len = desc->len;
 267
 268                 offset = xsk_umem__extract_offset(addr);
 269                 addr = xsk_umem__extract_addr(addr);
 270
 271                 bufs[i] = (struct rte_mbuf *)
 272                                 xsk_umem__get_data(umem->buffer, addr);
 273                 bufs[i]->data_off = offset - sizeof(struct rte_mbuf);
 274
 275                 rte_pktmbuf_pkt_len(bufs[i]) = len;
 276                 rte_pktmbuf_data_len(bufs[i]) = len;
 277                 rx_bytes += len;
 278         }
 279
 280         xsk_ring_cons__release(rx, rcvd);
 281
 282         (void)reserve_fill_queue(umem, rcvd, fq_bufs);
 283
 284         /* statistics */
 285         rxq->stats.rx_pkts += rcvd;
 286         rxq->stats.rx_bytes += rx_bytes;
 287
 288 out:
 289         if (rcvd != nb_pkts)
 290                 rte_mempool_put_bulk(umem->mb_pool, (void **)&fq_bufs[rcvd],
 291                                      nb_pkts - rcvd);
 292
 293         return rcvd;
 294 }
 295 #else
 296 static uint16_t
 297 af_xdp_rx_cp(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 298 {
 299         struct pkt_rx_queue *rxq = queue;
 300         struct xsk_ring_cons *rx = &rxq->rx;
 301         struct xsk_umem_info *umem = rxq->umem;
 302         struct xsk_ring_prod *fq = &umem->fq;
 303         uint32_t idx_rx = 0;
 304         unsigned long rx_bytes = 0;
 305         int rcvd, i;
 306         uint32_t free_thresh = fq->size >> 1;
 307         struct rte_mbuf *mbufs[ETH_AF_XDP_RX_BATCH_SIZE];
 308
 309         if (unlikely(rte_pktmbuf_alloc_bulk(rxq->mb_pool, mbufs, nb_pkts) != 0))
 310                 return 0;
 311
 312         rcvd = xsk_ring_cons__peek(rx, nb_pkts, &idx_rx);
 313         if (rcvd == 0) {
 314 #if defined(XDP_USE_NEED_WAKEUP)
 315                 if (xsk_ring_prod__needs_wakeup(fq))
 316                         (void)poll(rxq->fds, 1, 1000);
 317 #endif
 318
 319                 goto out;
 320         }
 321
 322         if (xsk_prod_nb_free(fq, free_thresh) >= free_thresh)
 323                 (void)reserve_fill_queue(umem, ETH_AF_XDP_RX_BATCH_SIZE, NULL);
 324
 325         for (i = 0; i < rcvd; i++) {
 326                 const struct xdp_desc *desc;
 327                 uint64_t addr;
 328                 uint32_t len;
 329                 void *pkt;
 330
 331                 desc = xsk_ring_cons__rx_desc(rx, idx_rx++);
 332                 addr = desc->addr;
 333                 len = desc->len;
 334                 pkt = xsk_umem__get_data(rxq->umem->mz->addr, addr);
 335
 336                 rte_memcpy(rte_pktmbuf_mtod(mbufs[i], void *), pkt, len);
 337                 rte_ring_enqueue(umem->buf_ring, (void *)addr);
 338                 rte_pktmbuf_pkt_len(mbufs[i]) = len;
 339                 rte_pktmbuf_data_len(mbufs[i]) = len;
 340                 rx_bytes += len;
 341                 bufs[i] = mbufs[i];
 342         }
 343
 344         xsk_ring_cons__release(rx, rcvd);
 345
 346         /* statistics */
 347         rxq->stats.rx_pkts += rcvd;
 348         rxq->stats.rx_bytes += rx_bytes;
 349
 350 out:
 351         if (rcvd != nb_pkts)
 352                 rte_mempool_put_bulk(rxq->mb_pool, (void **)&mbufs[rcvd],
 353                                      nb_pkts - rcvd);
 354
 355         return rcvd;
 356 }
 357 #endif
 358
 359 static uint16_t
 360 eth_af_xdp_rx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 361 {
 362         nb_pkts = RTE_MIN(nb_pkts, ETH_AF_XDP_RX_BATCH_SIZE);
 363
 364 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 365         return af_xdp_rx_zc(queue, bufs, nb_pkts);
 366 #else
 367         return af_xdp_rx_cp(queue, bufs, nb_pkts);
 368 #endif
 369 }
 370
 371 static void
 372 pull_umem_cq(struct xsk_umem_info *umem, int size)
 373 {
 374         struct xsk_ring_cons *cq = &umem->cq;
 375         size_t i, n;
 376         uint32_t idx_cq = 0;
 377
 378         n = xsk_ring_cons__peek(cq, size, &idx_cq);
 379
 380         for (i = 0; i < n; i++) {
 381                 uint64_t addr;
 382                 addr = *xsk_ring_cons__comp_addr(cq, idx_cq++);
 383 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 384                 addr = xsk_umem__extract_addr(addr);
 385                 rte_pktmbuf_free((struct rte_mbuf *)
 386                                         xsk_umem__get_data(umem->buffer, addr));
 387 #else
 388                 rte_ring_enqueue(umem->buf_ring, (void *)addr);
 389 #endif
 390         }
 391
 392         xsk_ring_cons__release(cq, n);
 393 }
 394
 395 static void
 396 kick_tx(struct pkt_tx_queue *txq)
 397 {
 398         struct xsk_umem_info *umem = txq->umem;
 399
 400 #if defined(XDP_USE_NEED_WAKEUP)
 401         if (xsk_ring_prod__needs_wakeup(&txq->tx))
 402 #endif
 403                 while (send(xsk_socket__fd(txq->pair->xsk), NULL,
 404                             0, MSG_DONTWAIT) < 0) {
 405                         /* some thing unexpected */
 406                         if (errno != EBUSY && errno != EAGAIN && errno != EINTR)
 407                                 break;
 408
 409                         /* pull from completion queue to leave more space */
 410                         if (errno == EAGAIN)
 411                                 pull_umem_cq(umem, ETH_AF_XDP_TX_BATCH_SIZE);
 412                 }
 413 #ifndef XDP_UMEM_UNALIGNED_CHUNK_FLAG
 414         pull_umem_cq(umem, ETH_AF_XDP_TX_BATCH_SIZE);
 415 #endif
 416 }
 417
 418 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 419 static uint16_t
 420 af_xdp_tx_zc(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 421 {
 422         struct pkt_tx_queue *txq = queue;
 423         struct xsk_umem_info *umem = txq->umem;
 424         struct rte_mbuf *mbuf;
 425         unsigned long tx_bytes = 0;
 426         int i;
 427         uint32_t idx_tx;
 428         uint16_t count = 0;
 429         struct xdp_desc *desc;
 430         uint64_t addr, offset;
 431
 432         pull_umem_cq(umem, nb_pkts);
 433
 434         for (i = 0; i < nb_pkts; i++) {
 435                 mbuf = bufs[i];
 436
 437                 if (mbuf->pool == umem->mb_pool) {
 438                         if (!xsk_ring_prod__reserve(&txq->tx, 1, &idx_tx)) {
 439                                 kick_tx(txq);
 440                                 goto out;
 441                         }
 442                         desc = xsk_ring_prod__tx_desc(&txq->tx, idx_tx);
 443                         desc->len = mbuf->pkt_len;
 444                         addr = (uint64_t)mbuf - (uint64_t)umem->buffer;
 445                         offset = rte_pktmbuf_mtod(mbuf, uint64_t) -
 446                                         (uint64_t)mbuf;
 447                         offset = offset << XSK_UNALIGNED_BUF_OFFSET_SHIFT;
 448                         desc->addr = addr | offset;
 449                         count++;
 450                 } else {
 451                         struct rte_mbuf *local_mbuf =
 452                                         rte_pktmbuf_alloc(umem->mb_pool);
 453                         void *pkt;
 454
 455                         if (local_mbuf == NULL)
 456                                 goto out;
 457
 458                         if (!xsk_ring_prod__reserve(&txq->tx, 1, &idx_tx)) {
 459                                 rte_pktmbuf_free(local_mbuf);
 460                                 kick_tx(txq);
 461                                 goto out;
 462                         }
 463
 464                         desc = xsk_ring_prod__tx_desc(&txq->tx, idx_tx);
 465                         desc->len = mbuf->pkt_len;
 466
 467                         addr = (uint64_t)local_mbuf - (uint64_t)umem->buffer;
 468                         offset = rte_pktmbuf_mtod(local_mbuf, uint64_t) -
 469                                         (uint64_t)local_mbuf;
 470                         pkt = xsk_umem__get_data(umem->buffer, addr + offset);
 471                         offset = offset << XSK_UNALIGNED_BUF_OFFSET_SHIFT;
 472                         desc->addr = addr | offset;
 473                         rte_memcpy(pkt, rte_pktmbuf_mtod(mbuf, void *),
 474                                         desc->len);
 475                         rte_pktmbuf_free(mbuf);
 476                         count++;
 477                 }
 478
 479                 tx_bytes += mbuf->pkt_len;
 480         }
 481
 482         kick_tx(txq);
 483
 484 out:
 485         xsk_ring_prod__submit(&txq->tx, count);
 486
 487         txq->stats.tx_pkts += count;
 488         txq->stats.tx_bytes += tx_bytes;
 489         txq->stats.tx_dropped += nb_pkts - count;
 490
 491         return count;
 492 }
 493 #else
 494 static uint16_t
 495 af_xdp_tx_cp(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 496 {
 497         struct pkt_tx_queue *txq = queue;
 498         struct xsk_umem_info *umem = txq->umem;
 499         struct rte_mbuf *mbuf;
 500         void *addrs[ETH_AF_XDP_TX_BATCH_SIZE];
 501         unsigned long tx_bytes = 0;
 502         int i;
 503         uint32_t idx_tx;
 504
 505         nb_pkts = RTE_MIN(nb_pkts, ETH_AF_XDP_TX_BATCH_SIZE);
 506
 507         pull_umem_cq(umem, nb_pkts);
 508
 509         nb_pkts = rte_ring_dequeue_bulk(umem->buf_ring, addrs,
 510                                         nb_pkts, NULL);
 511         if (nb_pkts == 0)
 512                 return 0;
 513
 514         if (xsk_ring_prod__reserve(&txq->tx, nb_pkts, &idx_tx) != nb_pkts) {
 515                 kick_tx(txq);
 516                 rte_ring_enqueue_bulk(umem->buf_ring, addrs, nb_pkts, NULL);
 517                 return 0;
 518         }
 519
 520         for (i = 0; i < nb_pkts; i++) {
 521                 struct xdp_desc *desc;
 522                 void *pkt;
 523
 524                 desc = xsk_ring_prod__tx_desc(&txq->tx, idx_tx + i);
 525                 mbuf = bufs[i];
 526                 desc->len = mbuf->pkt_len;
 527
 528                 desc->addr = (uint64_t)addrs[i];
 529                 pkt = xsk_umem__get_data(umem->mz->addr,
 530                                          desc->addr);
 531                 rte_memcpy(pkt, rte_pktmbuf_mtod(mbuf, void *), desc->len);
 532                 tx_bytes += mbuf->pkt_len;
 533                 rte_pktmbuf_free(mbuf);
 534         }
 535
 536         xsk_ring_prod__submit(&txq->tx, nb_pkts);
 537
 538         kick_tx(txq);
 539
 540         txq->stats.tx_pkts += nb_pkts;
 541         txq->stats.tx_bytes += tx_bytes;
 542
 543         return nb_pkts;
 544 }
 545 #endif
 546
 547 static uint16_t
 548 eth_af_xdp_tx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 549 {
 550 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 551         return af_xdp_tx_zc(queue, bufs, nb_pkts);
 552 #else
 553         return af_xdp_tx_cp(queue, bufs, nb_pkts);
 554 #endif
 555 }
 556
 557 static int
 558 eth_dev_start(struct rte_eth_dev *dev)
 559 {
 560         dev->data->dev_link.link_status = ETH_LINK_UP;
 561
 562         return 0;
 563 }
 564
 565 /* This function gets called when the current port gets stopped. */
 566 static void
 567 eth_dev_stop(struct rte_eth_dev *dev)
 568 {
 569         dev->data->dev_link.link_status = ETH_LINK_DOWN;
 570 }
 571
 572 static int
 573 eth_dev_configure(struct rte_eth_dev *dev)
 574 {
 575         /* rx/tx must be paired */
 576         if (dev->data->nb_rx_queues != dev->data->nb_tx_queues)
 577                 return -EINVAL;
 578
 579         return 0;
 580 }
 581
 582 static int
 583 eth_dev_info(struct rte_eth_dev *dev, struct rte_eth_dev_info *dev_info)
 584 {
 585         struct pmd_internals *internals = dev->data->dev_private;
 586
 587         dev_info->if_index = internals->if_index;
 588         dev_info->max_mac_addrs = 1;
 589         dev_info->max_rx_pktlen = ETH_FRAME_LEN;
 590         dev_info->max_rx_queues = internals->queue_cnt;
 591         dev_info->max_tx_queues = internals->queue_cnt;
 592
 593         dev_info->min_mtu = RTE_ETHER_MIN_MTU;
 594         dev_info->max_mtu = ETH_AF_XDP_FRAME_SIZE - ETH_AF_XDP_DATA_HEADROOM;
 595
 596         dev_info->default_rxportconf.nb_queues = 1;
 597         dev_info->default_txportconf.nb_queues = 1;
 598         dev_info->default_rxportconf.ring_size = ETH_AF_XDP_DFLT_NUM_DESCS;
 599         dev_info->default_txportconf.ring_size = ETH_AF_XDP_DFLT_NUM_DESCS;
 600
 601         return 0;
 602 }
 603
 604 static int
 605 eth_stats_get(struct rte_eth_dev *dev, struct rte_eth_stats *stats)
 606 {
 607         struct pmd_internals *internals = dev->data->dev_private;
 608         struct xdp_statistics xdp_stats;
 609         struct pkt_rx_queue *rxq;
 610         struct pkt_tx_queue *txq;
 611         socklen_t optlen;
 612         int i, ret;
 613
 614         for (i = 0; i < dev->data->nb_rx_queues; i++) {
 615                 optlen = sizeof(struct xdp_statistics);
 616                 rxq = &internals->rx_queues[i];
 617                 txq = rxq->pair;
 618                 stats->q_ipackets[i] = rxq->stats.rx_pkts;
 619                 stats->q_ibytes[i] = rxq->stats.rx_bytes;
 620
 621                 stats->q_opackets[i] = txq->stats.tx_pkts;
 622                 stats->q_obytes[i] = txq->stats.tx_bytes;
 623
 624                 stats->ipackets += stats->q_ipackets[i];
 625                 stats->ibytes += stats->q_ibytes[i];
 626                 stats->imissed += rxq->stats.rx_dropped;
 627                 stats->oerrors += txq->stats.tx_dropped;
 628                 ret = getsockopt(xsk_socket__fd(rxq->xsk), SOL_XDP,
 629                                 XDP_STATISTICS, &xdp_stats, &optlen);
 630                 if (ret != 0) {
 631                         AF_XDP_LOG(ERR, "getsockopt() failed for XDP_STATISTICS.\n");
 632                         return -1;
 633                 }
 634                 stats->imissed += xdp_stats.rx_dropped;
 635
 636                 stats->opackets += stats->q_opackets[i];
 637                 stats->obytes += stats->q_obytes[i];
 638         }
 639
 640         return 0;
 641 }
 642
 643 static int
 644 eth_stats_reset(struct rte_eth_dev *dev)
 645 {
 646         struct pmd_internals *internals = dev->data->dev_private;
 647         int i;
 648
 649         for (i = 0; i < internals->queue_cnt; i++) {
 650                 memset(&internals->rx_queues[i].stats, 0,
 651                                         sizeof(struct rx_stats));
 652                 memset(&internals->tx_queues[i].stats, 0,
 653                                         sizeof(struct tx_stats));
 654         }
 655
 656         return 0;
 657 }
 658
 659 static void
 660 remove_xdp_program(struct pmd_internals *internals)
 661 {
 662         uint32_t curr_prog_id = 0;
 663
 664         if (bpf_get_link_xdp_id(internals->if_index, &curr_prog_id,
 665                                 XDP_FLAGS_UPDATE_IF_NOEXIST)) {
 666                 AF_XDP_LOG(ERR, "bpf_get_link_xdp_id failed\n");
 667                 return;
 668         }
 669         bpf_set_link_xdp_fd(internals->if_index, -1,
 670                         XDP_FLAGS_UPDATE_IF_NOEXIST);
 671 }
 672
 673 static void
 674 xdp_umem_destroy(struct xsk_umem_info *umem)
 675 {
 676 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 677         rte_mempool_free(umem->mb_pool);
 678         umem->mb_pool = NULL;
 679 #else
 680         rte_memzone_free(umem->mz);
 681         umem->mz = NULL;
 682
 683         rte_ring_free(umem->buf_ring);
 684         umem->buf_ring = NULL;
 685 #endif
 686
 687         rte_free(umem);
 688         umem = NULL;
 689 }
 690
 691 static void
 692 eth_dev_close(struct rte_eth_dev *dev)
 693 {
 694         struct pmd_internals *internals = dev->data->dev_private;
 695         struct pkt_rx_queue *rxq;
 696         int i;
 697
 698         AF_XDP_LOG(INFO, "Closing AF_XDP ethdev on numa socket %u\n",
 699                 rte_socket_id());
 700
 701         for (i = 0; i < internals->queue_cnt; i++) {
 702                 rxq = &internals->rx_queues[i];
 703                 if (rxq->umem == NULL)
 704                         break;
 705                 xsk_socket__delete(rxq->xsk);
 706                 (void)xsk_umem__delete(rxq->umem->umem);
 707                 xdp_umem_destroy(rxq->umem);
 708
 709                 /* free pkt_tx_queue */
 710                 rte_free(rxq->pair);
 711                 rte_free(rxq);
 712         }
 713
 714         /*
 715          * MAC is not allocated dynamically, setting it to NULL would prevent
 716          * from releasing it in rte_eth_dev_release_port.
 717          */
 718         dev->data->mac_addrs = NULL;
 719
 720         remove_xdp_program(internals);
 721 }
 722
 723 static void
 724 eth_queue_release(void *q __rte_unused)
 725 {
 726 }
 727
 728 static int
 729 eth_link_update(struct rte_eth_dev *dev __rte_unused,
 730                 int wait_to_complete __rte_unused)
 731 {
 732         return 0;
 733 }
 734
 735 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 736 static inline uint64_t get_base_addr(struct rte_mempool *mp)
 737 {
 738         struct rte_mempool_memhdr *memhdr;
 739
 740         memhdr = STAILQ_FIRST(&mp->mem_list);
 741         return (uint64_t)memhdr->addr & ~(getpagesize() - 1);
 742 }
 743
 744 static struct
 745 xsk_umem_info *xdp_umem_configure(struct pmd_internals *internals __rte_unused,
 746                                   struct pkt_rx_queue *rxq)
 747 {
 748         struct xsk_umem_info *umem;
 749         int ret;
 750         struct xsk_umem_config usr_config = {
 751                 .fill_size = ETH_AF_XDP_DFLT_NUM_DESCS,
 752                 .comp_size = ETH_AF_XDP_DFLT_NUM_DESCS,
 753                 .flags = XDP_UMEM_UNALIGNED_CHUNK_FLAG};
 754         void *base_addr = NULL;
 755         struct rte_mempool *mb_pool = rxq->mb_pool;
 756
 757         usr_config.frame_size = rte_pktmbuf_data_room_size(mb_pool) +
 758                                         ETH_AF_XDP_MBUF_OVERHEAD +
 759                                         mb_pool->private_data_size;
 760         usr_config.frame_headroom = ETH_AF_XDP_DATA_HEADROOM +
 761                                         mb_pool->private_data_size;
 762
 763         umem = rte_zmalloc_socket("umem", sizeof(*umem), 0, rte_socket_id());
 764         if (umem == NULL) {
 765                 AF_XDP_LOG(ERR, "Failed to allocate umem info");
 766                 return NULL;
 767         }
 768
 769         umem->mb_pool = mb_pool;
 770         base_addr = (void *)get_base_addr(mb_pool);
 771
 772         ret = xsk_umem__create(&umem->umem, base_addr,
 773                                mb_pool->populated_size * usr_config.frame_size,
 774                                &umem->fq, &umem->cq,
 775                                &usr_config);
 776
 777         if (ret) {
 778                 AF_XDP_LOG(ERR, "Failed to create umem");
 779                 goto err;
 780         }
 781         umem->buffer = base_addr;
 782
 783 #else
 784 static struct
 785 xsk_umem_info *xdp_umem_configure(struct pmd_internals *internals,
 786                                   struct pkt_rx_queue *rxq)
 787 {
 788         struct xsk_umem_info *umem;
 789         const struct rte_memzone *mz;
 790         struct xsk_umem_config usr_config = {
 791                 .fill_size = ETH_AF_XDP_DFLT_NUM_DESCS,
 792                 .comp_size = ETH_AF_XDP_DFLT_NUM_DESCS,
 793                 .frame_size = ETH_AF_XDP_FRAME_SIZE,
 794                 .frame_headroom = ETH_AF_XDP_DATA_HEADROOM };
 795         char ring_name[RTE_RING_NAMESIZE];
 796         char mz_name[RTE_MEMZONE_NAMESIZE];
 797         int ret;
 798         uint64_t i;
 799
 800         umem = rte_zmalloc_socket("umem", sizeof(*umem), 0, rte_socket_id());
 801         if (umem == NULL) {
 802                 AF_XDP_LOG(ERR, "Failed to allocate umem info");
 803                 return NULL;
 804         }
 805
 806         snprintf(ring_name, sizeof(ring_name), "af_xdp_ring_%s_%u",
 807                        internals->if_name, rxq->xsk_queue_idx);
 808         umem->buf_ring = rte_ring_create(ring_name,
 809                                          ETH_AF_XDP_NUM_BUFFERS,
 810                                          rte_socket_id(),
 811                                          RING_F_SP_ENQ | RING_F_SC_DEQ);
 812         if (umem->buf_ring == NULL) {
 813                 AF_XDP_LOG(ERR, "Failed to create rte_ring\n");
 814                 goto err;
 815         }
 816
 817         for (i = 0; i < ETH_AF_XDP_NUM_BUFFERS; i++)
 818                 rte_ring_enqueue(umem->buf_ring,
 819                                  (void *)(i * ETH_AF_XDP_FRAME_SIZE +
 820                                           ETH_AF_XDP_DATA_HEADROOM));
 821
 822         snprintf(mz_name, sizeof(mz_name), "af_xdp_umem_%s_%u",
 823                        internals->if_name, rxq->xsk_queue_idx);
 824         mz = rte_memzone_reserve_aligned(mz_name,
 825                         ETH_AF_XDP_NUM_BUFFERS * ETH_AF_XDP_FRAME_SIZE,
 826                         rte_socket_id(), RTE_MEMZONE_IOVA_CONTIG,
 827                         getpagesize());
 828         if (mz == NULL) {
 829                 AF_XDP_LOG(ERR, "Failed to reserve memzone for af_xdp umem.\n");
 830                 goto err;
 831         }
 832
 833         ret = xsk_umem__create(&umem->umem, mz->addr,
 834                                ETH_AF_XDP_NUM_BUFFERS * ETH_AF_XDP_FRAME_SIZE,
 835                                &umem->fq, &umem->cq,
 836                                &usr_config);
 837
 838         if (ret) {
 839                 AF_XDP_LOG(ERR, "Failed to create umem");
 840                 goto err;
 841         }
 842         umem->mz = mz;
 843
 844 #endif
 845         return umem;
 846
 847 err:
 848         xdp_umem_destroy(umem);
 849         return NULL;
 850 }
 851
 852 static int
 853 xsk_configure(struct pmd_internals *internals, struct pkt_rx_queue *rxq,
 854               int ring_size)
 855 {
 856         struct xsk_socket_config cfg;
 857         struct pkt_tx_queue *txq = rxq->pair;
 858         int ret = 0;
 859         int reserve_size = ETH_AF_XDP_DFLT_NUM_DESCS / 2;
 860         struct rte_mbuf *fq_bufs[reserve_size];
 861
 862         rxq->umem = xdp_umem_configure(internals, rxq);
 863         if (rxq->umem == NULL)
 864                 return -ENOMEM;
 865         txq->umem = rxq->umem;
 866
 867         cfg.rx_size = ring_size;
 868         cfg.tx_size = ring_size;
 869         cfg.libbpf_flags = 0;
 870         cfg.xdp_flags = XDP_FLAGS_UPDATE_IF_NOEXIST;
 871         cfg.bind_flags = 0;
 872
 873 #if defined(XDP_USE_NEED_WAKEUP)
 874         cfg.bind_flags |= XDP_USE_NEED_WAKEUP;
 875 #endif
 876
 877         ret = xsk_socket__create(&rxq->xsk, internals->if_name,
 878                         rxq->xsk_queue_idx, rxq->umem->umem, &rxq->rx,
 879                         &txq->tx, &cfg);
 880         if (ret) {
 881                 AF_XDP_LOG(ERR, "Failed to create xsk socket.\n");
 882                 goto err;
 883         }
 884
 885 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 886         if (rte_pktmbuf_alloc_bulk(rxq->umem->mb_pool, fq_bufs, reserve_size)) {
 887                 AF_XDP_LOG(DEBUG, "Failed to get enough buffers for fq.\n");
 888                 goto err;
 889         }
 890 #endif
 891         ret = reserve_fill_queue(rxq->umem, reserve_size, fq_bufs);
 892         if (ret) {
 893                 xsk_socket__delete(rxq->xsk);
 894                 AF_XDP_LOG(ERR, "Failed to reserve fill queue.\n");
 895                 goto err;
 896         }
 897
 898         return 0;
 899
 900 err:
 901         xdp_umem_destroy(rxq->umem);
 902
 903         return ret;
 904 }
 905
 906 static int
 907 eth_rx_queue_setup(struct rte_eth_dev *dev,
 908                    uint16_t rx_queue_id,
 909                    uint16_t nb_rx_desc,
 910                    unsigned int socket_id __rte_unused,
 911                    const struct rte_eth_rxconf *rx_conf __rte_unused,
 912                    struct rte_mempool *mb_pool)
 913 {
 914         struct pmd_internals *internals = dev->data->dev_private;
 915         struct pkt_rx_queue *rxq;
 916         int ret;
 917
 918         rxq = &internals->rx_queues[rx_queue_id];
 919
 920         AF_XDP_LOG(INFO, "Set up rx queue, rx queue id: %d, xsk queue id: %d\n",
 921                    rx_queue_id, rxq->xsk_queue_idx);
 922
 923 #ifndef XDP_UMEM_UNALIGNED_CHUNK_FLAG
 924         uint32_t buf_size, data_size;
 925
 926         /* Now get the space available for data in the mbuf */
 927         buf_size = rte_pktmbuf_data_room_size(mb_pool) -
 928                 RTE_PKTMBUF_HEADROOM;
 929         data_size = ETH_AF_XDP_FRAME_SIZE - ETH_AF_XDP_DATA_HEADROOM;
 930
 931         if (data_size > buf_size) {
 932                 AF_XDP_LOG(ERR, "%s: %d bytes will not fit in mbuf (%d bytes)\n",
 933                         dev->device->name, data_size, buf_size);
 934                 ret = -ENOMEM;
 935                 goto err;
 936         }
 937 #endif
 938
 939         rxq->mb_pool = mb_pool;
 940
 941         if (xsk_configure(internals, rxq, nb_rx_desc)) {
 942                 AF_XDP_LOG(ERR, "Failed to configure xdp socket\n");
 943                 ret = -EINVAL;
 944                 goto err;
 945         }
 946
 947         rxq->fds[0].fd = xsk_socket__fd(rxq->xsk);
 948         rxq->fds[0].events = POLLIN;
 949
 950         dev->data->rx_queues[rx_queue_id] = rxq;
 951         return 0;
 952
 953 err:
 954         return ret;
 955 }
 956
 957 static int
 958 eth_tx_queue_setup(struct rte_eth_dev *dev,
 959                    uint16_t tx_queue_id,
 960                    uint16_t nb_tx_desc __rte_unused,
 961                    unsigned int socket_id __rte_unused,
 962                    const struct rte_eth_txconf *tx_conf __rte_unused)
 963 {
 964         struct pmd_internals *internals = dev->data->dev_private;
 965         struct pkt_tx_queue *txq;
 966
 967         txq = &internals->tx_queues[tx_queue_id];
 968
 969         dev->data->tx_queues[tx_queue_id] = txq;
 970         return 0;
 971 }
 972
 973 static int
 974 eth_dev_mtu_set(struct rte_eth_dev *dev, uint16_t mtu)
 975 {
 976         struct pmd_internals *internals = dev->data->dev_private;
 977         struct ifreq ifr = { .ifr_mtu = mtu };
 978         int ret;
 979         int s;
 980
 981         s = socket(PF_INET, SOCK_DGRAM, 0);
 982         if (s < 0)
 983                 return -EINVAL;
 984
 985         strlcpy(ifr.ifr_name, internals->if_name, IFNAMSIZ);
 986         ret = ioctl(s, SIOCSIFMTU, &ifr);
 987         close(s);
 988
 989         return (ret < 0) ? -errno : 0;
 990 }
 991
 992 static int
 993 eth_dev_change_flags(char *if_name, uint32_t flags, uint32_t mask)
 994 {
 995         struct ifreq ifr;
 996         int ret = 0;
 997         int s;
 998
 999         s = socket(PF_INET, SOCK_DGRAM, 0);
1000         if (s < 0)
1001                 return -errno;
1002
1003         strlcpy(ifr.ifr_name, if_name, IFNAMSIZ);
1004         if (ioctl(s, SIOCGIFFLAGS, &ifr) < 0) {
1005                 ret = -errno;
1006                 goto out;
1007         }
1008         ifr.ifr_flags &= mask;
1009         ifr.ifr_flags |= flags;
1010         if (ioctl(s, SIOCSIFFLAGS, &ifr) < 0) {
1011                 ret = -errno;
1012                 goto out;
1013         }
1014 out:
1015         close(s);
1016         return ret;
1017 }
1018
1019 static int
1020 eth_dev_promiscuous_enable(struct rte_eth_dev *dev)
1021 {
1022         struct pmd_internals *internals = dev->data->dev_private;
1023
1024         return eth_dev_change_flags(internals->if_name, IFF_PROMISC, ~0);
1025 }
1026
1027 static int
1028 eth_dev_promiscuous_disable(struct rte_eth_dev *dev)
1029 {
1030         struct pmd_internals *internals = dev->data->dev_private;
1031
1032         return eth_dev_change_flags(internals->if_name, 0, ~IFF_PROMISC);
1033 }
1034
1035 static const struct eth_dev_ops ops = {
1036         .dev_start = eth_dev_start,
1037         .dev_stop = eth_dev_stop,
1038         .dev_close = eth_dev_close,
1039         .dev_configure = eth_dev_configure,
1040         .dev_infos_get = eth_dev_info,
1041         .mtu_set = eth_dev_mtu_set,
1042         .promiscuous_enable = eth_dev_promiscuous_enable,
1043         .promiscuous_disable = eth_dev_promiscuous_disable,
1044         .rx_queue_setup = eth_rx_queue_setup,
1045         .tx_queue_setup = eth_tx_queue_setup,
1046         .rx_queue_release = eth_queue_release,
1047         .tx_queue_release = eth_queue_release,
1048         .link_update = eth_link_update,
1049         .stats_get = eth_stats_get,
1050         .stats_reset = eth_stats_reset,
1051 };
1052
1053 /** parse integer from integer argument */
1054 static int
1055 parse_integer_arg(const char *key __rte_unused,
1056                   const char *value, void *extra_args)
1057 {
1058         int *i = (int *)extra_args;
1059         char *end;
1060
1061         *i = strtol(value, &end, 10);
1062         if (*i < 0) {
1063                 AF_XDP_LOG(ERR, "Argument has to be positive.\n");
1064                 return -EINVAL;
1065         }
1066
1067         return 0;
1068 }
1069
1070 /** parse name argument */
1071 static int
1072 parse_name_arg(const char *key __rte_unused,
1073                const char *value, void *extra_args)
1074 {
1075         char *name = extra_args;
1076
1077         if (strnlen(value, IFNAMSIZ) > IFNAMSIZ - 1) {
1078                 AF_XDP_LOG(ERR, "Invalid name %s, should be less than %u bytes.\n",
1079                            value, IFNAMSIZ);
1080                 return -EINVAL;
1081         }
1082
1083         strlcpy(name, value, IFNAMSIZ);
1084
1085         return 0;
1086 }
1087
1088 static int
1089 xdp_get_channels_info(const char *if_name, int *max_queues,
1090                                 int *combined_queues)
1091 {
1092         struct ethtool_channels channels;
1093         struct ifreq ifr;
1094         int fd, ret;
1095
1096         fd = socket(AF_INET, SOCK_DGRAM, 0);
1097         if (fd < 0)
1098                 return -1;
1099
1100         channels.cmd = ETHTOOL_GCHANNELS;
1101         ifr.ifr_data = (void *)&channels;
1102         strncpy(ifr.ifr_name, if_name, IFNAMSIZ);
1103         ret = ioctl(fd, SIOCETHTOOL, &ifr);
1104         if (ret) {
1105                 if (errno == EOPNOTSUPP) {
1106                         ret = 0;
1107                 } else {
1108                         ret = -errno;
1109                         goto out;
1110                 }
1111         }
1112
1113         if (channels.max_combined == 0 || errno == EOPNOTSUPP) {
1114                 /* If the device says it has no channels, then all traffic
1115                  * is sent to a single stream, so max queues = 1.
1116                  */
1117                 *max_queues = 1;
1118                 *combined_queues = 1;
1119         } else {
1120                 *max_queues = channels.max_combined;
1121                 *combined_queues = channels.combined_count;
1122         }
1123
1124  out:
1125         close(fd);
1126         return ret;
1127 }
1128
1129 static int
1130 parse_parameters(struct rte_kvargs *kvlist, char *if_name, int *start_queue,
1131                         int *queue_cnt)
1132 {
1133         int ret;
1134
1135         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_IFACE_ARG,
1136                                  &parse_name_arg, if_name);
1137         if (ret < 0)
1138                 goto free_kvlist;
1139
1140         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_START_QUEUE_ARG,
1141                                  &parse_integer_arg, start_queue);
1142         if (ret < 0)
1143                 goto free_kvlist;
1144
1145         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_QUEUE_COUNT_ARG,
1146                                  &parse_integer_arg, queue_cnt);
1147         if (ret < 0 || *queue_cnt <= 0) {
1148                 ret = -EINVAL;
1149                 goto free_kvlist;
1150         }
1151
1152 free_kvlist:
1153         rte_kvargs_free(kvlist);
1154         return ret;
1155 }
1156
1157 static int
1158 get_iface_info(const char *if_name,
1159                struct rte_ether_addr *eth_addr,
1160                int *if_index)
1161 {
1162         struct ifreq ifr;
1163         int sock = socket(AF_INET, SOCK_DGRAM, IPPROTO_IP);
1164
1165         if (sock < 0)
1166                 return -1;
1167
1168         strlcpy(ifr.ifr_name, if_name, IFNAMSIZ);
1169         if (ioctl(sock, SIOCGIFINDEX, &ifr))
1170                 goto error;
1171
1172         *if_index = ifr.ifr_ifindex;
1173
1174         if (ioctl(sock, SIOCGIFHWADDR, &ifr))
1175                 goto error;
1176
1177         rte_memcpy(eth_addr, ifr.ifr_hwaddr.sa_data, RTE_ETHER_ADDR_LEN);
1178
1179         close(sock);
1180         return 0;
1181
1182 error:
1183         close(sock);
1184         return -1;
1185 }
1186
1187 static struct rte_eth_dev *
1188 init_internals(struct rte_vdev_device *dev, const char *if_name,
1189                         int start_queue_idx, int queue_cnt)
1190 {
1191         const char *name = rte_vdev_device_name(dev);
1192         const unsigned int numa_node = dev->device.numa_node;
1193         struct pmd_internals *internals;
1194         struct rte_eth_dev *eth_dev;
1195         int ret;
1196         int i;
1197
1198         internals = rte_zmalloc_socket(name, sizeof(*internals), 0, numa_node);
1199         if (internals == NULL)
1200                 return NULL;
1201
1202         internals->start_queue_idx = start_queue_idx;
1203         internals->queue_cnt = queue_cnt;
1204         strlcpy(internals->if_name, if_name, IFNAMSIZ);
1205
1206         if (xdp_get_channels_info(if_name, &internals->max_queue_cnt,
1207                                   &internals->combined_queue_cnt)) {
1208                 AF_XDP_LOG(ERR, "Failed to get channel info of interface: %s\n",
1209                                 if_name);
1210                 goto err_free_internals;
1211         }
1212
1213         if (queue_cnt > internals->combined_queue_cnt) {
1214                 AF_XDP_LOG(ERR, "Specified queue count %d is larger than combined queue count %d.\n",
1215                                 queue_cnt, internals->combined_queue_cnt);
1216                 goto err_free_internals;
1217         }
1218
1219         internals->rx_queues = rte_zmalloc_socket(NULL,
1220                                         sizeof(struct pkt_rx_queue) * queue_cnt,
1221                                         0, numa_node);
1222         if (internals->rx_queues == NULL) {
1223                 AF_XDP_LOG(ERR, "Failed to allocate memory for rx queues.\n");
1224                 goto err_free_internals;
1225         }
1226
1227         internals->tx_queues = rte_zmalloc_socket(NULL,
1228                                         sizeof(struct pkt_tx_queue) * queue_cnt,
1229                                         0, numa_node);
1230         if (internals->tx_queues == NULL) {
1231                 AF_XDP_LOG(ERR, "Failed to allocate memory for tx queues.\n");
1232                 goto err_free_rx;
1233         }
1234         for (i = 0; i < queue_cnt; i++) {
1235                 internals->tx_queues[i].pair = &internals->rx_queues[i];
1236                 internals->rx_queues[i].pair = &internals->tx_queues[i];
1237                 internals->rx_queues[i].xsk_queue_idx = start_queue_idx + i;
1238                 internals->tx_queues[i].xsk_queue_idx = start_queue_idx + i;
1239         }
1240
1241         ret = get_iface_info(if_name, &internals->eth_addr,
1242                              &internals->if_index);
1243         if (ret)
1244                 goto err_free_tx;
1245
1246         eth_dev = rte_eth_vdev_allocate(dev, 0);
1247         if (eth_dev == NULL)
1248                 goto err_free_tx;
1249
1250         eth_dev->data->dev_private = internals;
1251         eth_dev->data->dev_link = pmd_link;
1252         eth_dev->data->mac_addrs = &internals->eth_addr;
1253         eth_dev->dev_ops = &ops;
1254         eth_dev->rx_pkt_burst = eth_af_xdp_rx;
1255         eth_dev->tx_pkt_burst = eth_af_xdp_tx;
1256         /* Let rte_eth_dev_close() release the port resources. */
1257         eth_dev->data->dev_flags |= RTE_ETH_DEV_CLOSE_REMOVE;
1258
1259 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
1260         AF_XDP_LOG(INFO, "Zero copy between umem and mbuf enabled.\n");
1261 #endif
1262
1263         return eth_dev;
1264
1265 err_free_tx:
1266         rte_free(internals->tx_queues);
1267 err_free_rx:
1268         rte_free(internals->rx_queues);
1269 err_free_internals:
1270         rte_free(internals);
1271         return NULL;
1272 }
1273
1274 static int
1275 rte_pmd_af_xdp_probe(struct rte_vdev_device *dev)
1276 {
1277         struct rte_kvargs *kvlist;
1278         char if_name[IFNAMSIZ] = {'\0'};
1279         int xsk_start_queue_idx = ETH_AF_XDP_DFLT_START_QUEUE_IDX;
1280         int xsk_queue_cnt = ETH_AF_XDP_DFLT_QUEUE_COUNT;
1281         struct rte_eth_dev *eth_dev = NULL;
1282         const char *name;
1283
1284         AF_XDP_LOG(INFO, "Initializing pmd_af_xdp for %s\n",
1285                 rte_vdev_device_name(dev));
1286
1287         name = rte_vdev_device_name(dev);
1288         if (rte_eal_process_type() == RTE_PROC_SECONDARY &&
1289                 strlen(rte_vdev_device_args(dev)) == 0) {
1290                 eth_dev = rte_eth_dev_attach_secondary(name);
1291                 if (eth_dev == NULL) {
1292                         AF_XDP_LOG(ERR, "Failed to probe %s\n", name);
1293                         return -EINVAL;
1294                 }
1295                 eth_dev->dev_ops = &ops;
1296                 rte_eth_dev_probing_finish(eth_dev);
1297                 return 0;
1298         }
1299
1300         kvlist = rte_kvargs_parse(rte_vdev_device_args(dev), valid_arguments);
1301         if (kvlist == NULL) {
1302                 AF_XDP_LOG(ERR, "Invalid kvargs key\n");
1303                 return -EINVAL;
1304         }
1305
1306         if (dev->device.numa_node == SOCKET_ID_ANY)
1307                 dev->device.numa_node = rte_socket_id();
1308
1309         if (parse_parameters(kvlist, if_name, &xsk_start_queue_idx,
1310                              &xsk_queue_cnt) < 0) {
1311                 AF_XDP_LOG(ERR, "Invalid kvargs value\n");
1312                 return -EINVAL;
1313         }
1314
1315         if (strlen(if_name) == 0) {
1316                 AF_XDP_LOG(ERR, "Network interface must be specified\n");
1317                 return -EINVAL;
1318         }
1319
1320         eth_dev = init_internals(dev, if_name, xsk_start_queue_idx,
1321                                         xsk_queue_cnt);
1322         if (eth_dev == NULL) {
1323                 AF_XDP_LOG(ERR, "Failed to init internals\n");
1324                 return -1;
1325         }
1326
1327         rte_eth_dev_probing_finish(eth_dev);
1328
1329         return 0;
1330 }
1331
1332 static int
1333 rte_pmd_af_xdp_remove(struct rte_vdev_device *dev)
1334 {
1335         struct rte_eth_dev *eth_dev = NULL;
1336
1337         AF_XDP_LOG(INFO, "Removing AF_XDP ethdev on numa socket %u\n",
1338                 rte_socket_id());
1339
1340         if (dev == NULL)
1341                 return -1;
1342
1343         /* find the ethdev entry */
1344         eth_dev = rte_eth_dev_allocated(rte_vdev_device_name(dev));
1345         if (eth_dev == NULL)
1346                 return 0;
1347
1348         eth_dev_close(eth_dev);
1349         rte_eth_dev_release_port(eth_dev);
1350
1351
1352         return 0;
1353 }
1354
1355 static struct rte_vdev_driver pmd_af_xdp_drv = {
1356         .probe = rte_pmd_af_xdp_probe,
1357         .remove = rte_pmd_af_xdp_remove,
1358 };
1359
1360 RTE_PMD_REGISTER_VDEV(net_af_xdp, pmd_af_xdp_drv);
1361 RTE_PMD_REGISTER_PARAM_STRING(net_af_xdp,
1362                               "iface=<string> "
1363                               "start_queue=<int> "
1364                               "queue_count=<int> ");
1365
1366 RTE_INIT(af_xdp_init_log)
1367 {
1368         af_xdp_logtype = rte_log_register("pmd.net.af_xdp");
1369         if (af_xdp_logtype >= 0)
1370                 rte_log_set_level(af_xdp_logtype, RTE_LOG_NOTICE);
1371 }