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_mempool.h>
  37 #include <rte_mbuf.h>
  38 #include <rte_malloc.h>
  39 #include <rte_ring.h>
  40
  41 #ifndef SOL_XDP
  42 #define SOL_XDP 283
  43 #endif
  44
  45 #ifndef AF_XDP
  46 #define AF_XDP 44
  47 #endif
  48
  49 #ifndef PF_XDP
  50 #define PF_XDP AF_XDP
  51 #endif
  52
  53 static int af_xdp_logtype;
  54
  55 #define AF_XDP_LOG(level, fmt, args...)                 \
  56         rte_log(RTE_LOG_ ## level, af_xdp_logtype,      \
  57                 "%s(): " fmt, __func__, ##args)
  58
  59 #define ETH_AF_XDP_FRAME_SIZE           2048
  60 #define ETH_AF_XDP_NUM_BUFFERS          4096
  61 #define ETH_AF_XDP_DFLT_NUM_DESCS       XSK_RING_CONS__DEFAULT_NUM_DESCS
  62 #define ETH_AF_XDP_DFLT_START_QUEUE_IDX 0
  63 #define ETH_AF_XDP_DFLT_QUEUE_COUNT     1
  64
  65 #define ETH_AF_XDP_RX_BATCH_SIZE        32
  66 #define ETH_AF_XDP_TX_BATCH_SIZE        32
  67
  68
  69 struct xsk_umem_info {
  70         struct xsk_ring_prod fq;
  71         struct xsk_ring_cons cq;
  72         struct xsk_umem *umem;
  73         struct rte_ring *buf_ring;
  74         const struct rte_memzone *mz;
  75         struct rte_mempool *mb_pool;
  76         void *buffer;
  77 };
  78
  79 struct rx_stats {
  80         uint64_t rx_pkts;
  81         uint64_t rx_bytes;
  82         uint64_t rx_dropped;
  83 };
  84
  85 struct pkt_rx_queue {
  86         struct xsk_ring_cons rx;
  87         struct xsk_umem_info *umem;
  88         struct xsk_socket *xsk;
  89         struct rte_mempool *mb_pool;
  90
  91         struct rx_stats stats;
  92
  93         struct pkt_tx_queue *pair;
  94         struct pollfd fds[1];
  95         int xsk_queue_idx;
  96 };
  97
  98 struct tx_stats {
  99         uint64_t tx_pkts;
 100         uint64_t tx_bytes;
 101         uint64_t tx_dropped;
 102 };
 103
 104 struct pkt_tx_queue {
 105         struct xsk_ring_prod tx;
 106         struct xsk_umem_info *umem;
 107
 108         struct tx_stats stats;
 109
 110         struct pkt_rx_queue *pair;
 111         int xsk_queue_idx;
 112 };
 113
 114 struct pmd_internals {
 115         int if_index;
 116         char if_name[IFNAMSIZ];
 117         int start_queue_idx;
 118         int queue_cnt;
 119         int max_queue_cnt;
 120         int combined_queue_cnt;
 121
 122         struct rte_ether_addr eth_addr;
 123
 124         struct pkt_rx_queue *rx_queues;
 125         struct pkt_tx_queue *tx_queues;
 126 };
 127
 128 #define ETH_AF_XDP_IFACE_ARG                    "iface"
 129 #define ETH_AF_XDP_START_QUEUE_ARG              "start_queue"
 130 #define ETH_AF_XDP_QUEUE_COUNT_ARG              "queue_count"
 131
 132 static const char * const valid_arguments[] = {
 133         ETH_AF_XDP_IFACE_ARG,
 134         ETH_AF_XDP_START_QUEUE_ARG,
 135         ETH_AF_XDP_QUEUE_COUNT_ARG,
 136         NULL
 137 };
 138
 139 static const struct rte_eth_link pmd_link = {
 140         .link_speed = ETH_SPEED_NUM_10G,
 141         .link_duplex = ETH_LINK_FULL_DUPLEX,
 142         .link_status = ETH_LINK_DOWN,
 143         .link_autoneg = ETH_LINK_AUTONEG
 144 };
 145
 146 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 147 static inline int
 148 reserve_fill_queue_zc(struct xsk_umem_info *umem, uint16_t reserve_size,
 149                       struct rte_mbuf **bufs)
 150 {
 151         struct xsk_ring_prod *fq = &umem->fq;
 152         uint32_t idx;
 153         uint16_t i;
 154
 155         if (unlikely(!xsk_ring_prod__reserve(fq, reserve_size, &idx))) {
 156                 for (i = 0; i < reserve_size; i++)
 157                         rte_pktmbuf_free(bufs[i]);
 158                 AF_XDP_LOG(DEBUG, "Failed to reserve enough fq descs.\n");
 159                 return -1;
 160         }
 161
 162         for (i = 0; i < reserve_size; i++) {
 163                 __u64 *fq_addr;
 164                 uint64_t addr;
 165
 166                 fq_addr = xsk_ring_prod__fill_addr(fq, idx++);
 167                 addr = (uint64_t)bufs[i] - (uint64_t)umem->buffer -
 168                                 umem->mb_pool->header_size;
 169                 *fq_addr = addr;
 170         }
 171
 172         xsk_ring_prod__submit(fq, reserve_size);
 173
 174         return 0;
 175 }
 176 #else
 177 static inline int
 178 reserve_fill_queue_cp(struct xsk_umem_info *umem, uint16_t reserve_size,
 179                       struct rte_mbuf **bufs __rte_unused)
 180 {
 181         struct xsk_ring_prod *fq = &umem->fq;
 182         void *addrs[reserve_size];
 183         uint32_t idx;
 184         uint16_t i;
 185
 186         if (rte_ring_dequeue_bulk(umem->buf_ring, addrs, reserve_size, NULL)
 187                     != reserve_size) {
 188                 AF_XDP_LOG(DEBUG, "Failed to get enough buffers for fq.\n");
 189                 return -1;
 190         }
 191
 192         if (unlikely(!xsk_ring_prod__reserve(fq, reserve_size, &idx))) {
 193                 AF_XDP_LOG(DEBUG, "Failed to reserve enough fq descs.\n");
 194                 rte_ring_enqueue_bulk(umem->buf_ring, addrs,
 195                                 reserve_size, NULL);
 196                 return -1;
 197         }
 198
 199         for (i = 0; i < reserve_size; i++) {
 200                 __u64 *fq_addr;
 201
 202                 fq_addr = xsk_ring_prod__fill_addr(fq, idx++);
 203                 *fq_addr = (uint64_t)addrs[i];
 204         }
 205
 206         xsk_ring_prod__submit(fq, reserve_size);
 207
 208         return 0;
 209 }
 210 #endif
 211
 212 static inline int
 213 reserve_fill_queue(struct xsk_umem_info *umem, uint16_t reserve_size,
 214                    struct rte_mbuf **bufs)
 215 {
 216 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 217         return reserve_fill_queue_zc(umem, reserve_size, bufs);
 218 #else
 219         return reserve_fill_queue_cp(umem, reserve_size, bufs);
 220 #endif
 221 }
 222
 223 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 224 static uint16_t
 225 af_xdp_rx_zc(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 226 {
 227         struct pkt_rx_queue *rxq = queue;
 228         struct xsk_ring_cons *rx = &rxq->rx;
 229         struct xsk_umem_info *umem = rxq->umem;
 230         uint32_t idx_rx = 0;
 231         unsigned long rx_bytes = 0;
 232         int rcvd, i;
 233         struct rte_mbuf *fq_bufs[ETH_AF_XDP_RX_BATCH_SIZE];
 234
 235         /* allocate bufs for fill queue replenishment after rx */
 236         if (rte_pktmbuf_alloc_bulk(umem->mb_pool, fq_bufs, nb_pkts)) {
 237                 AF_XDP_LOG(DEBUG,
 238                         "Failed to get enough buffers for fq.\n");
 239                 return -1;
 240         }
 241
 242         rcvd = xsk_ring_cons__peek(rx, nb_pkts, &idx_rx);
 243
 244         if (rcvd == 0) {
 245 #if defined(XDP_USE_NEED_WAKEUP)
 246                 if (xsk_ring_prod__needs_wakeup(&umem->fq))
 247                         (void)poll(rxq->fds, 1, 1000);
 248 #endif
 249
 250                 goto out;
 251         }
 252
 253         for (i = 0; i < rcvd; i++) {
 254                 const struct xdp_desc *desc;
 255                 uint64_t addr;
 256                 uint32_t len;
 257                 uint64_t offset;
 258
 259                 desc = xsk_ring_cons__rx_desc(rx, idx_rx++);
 260                 addr = desc->addr;
 261                 len = desc->len;
 262
 263                 offset = xsk_umem__extract_offset(addr);
 264                 addr = xsk_umem__extract_addr(addr);
 265
 266                 bufs[i] = (struct rte_mbuf *)
 267                                 xsk_umem__get_data(umem->buffer, addr +
 268                                         umem->mb_pool->header_size);
 269                 bufs[i]->data_off = offset - sizeof(struct rte_mbuf) -
 270                         rte_pktmbuf_priv_size(umem->mb_pool) -
 271                         umem->mb_pool->header_size;
 272
 273                 rte_pktmbuf_pkt_len(bufs[i]) = len;
 274                 rte_pktmbuf_data_len(bufs[i]) = len;
 275                 rx_bytes += len;
 276         }
 277
 278         xsk_ring_cons__release(rx, rcvd);
 279
 280         (void)reserve_fill_queue(umem, rcvd, fq_bufs);
 281
 282         /* statistics */
 283         rxq->stats.rx_pkts += rcvd;
 284         rxq->stats.rx_bytes += rx_bytes;
 285
 286 out:
 287         if (rcvd != nb_pkts)
 288                 rte_mempool_put_bulk(umem->mb_pool, (void **)&fq_bufs[rcvd],
 289                                      nb_pkts - rcvd);
 290
 291         return rcvd;
 292 }
 293 #else
 294 static uint16_t
 295 af_xdp_rx_cp(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 296 {
 297         struct pkt_rx_queue *rxq = queue;
 298         struct xsk_ring_cons *rx = &rxq->rx;
 299         struct xsk_umem_info *umem = rxq->umem;
 300         struct xsk_ring_prod *fq = &umem->fq;
 301         uint32_t idx_rx = 0;
 302         unsigned long rx_bytes = 0;
 303         int rcvd, i;
 304         uint32_t free_thresh = fq->size >> 1;
 305         struct rte_mbuf *mbufs[ETH_AF_XDP_RX_BATCH_SIZE];
 306
 307         if (unlikely(rte_pktmbuf_alloc_bulk(rxq->mb_pool, mbufs, nb_pkts) != 0))
 308                 return 0;
 309
 310         rcvd = xsk_ring_cons__peek(rx, nb_pkts, &idx_rx);
 311         if (rcvd == 0) {
 312 #if defined(XDP_USE_NEED_WAKEUP)
 313                 if (xsk_ring_prod__needs_wakeup(fq))
 314                         (void)poll(rxq->fds, 1, 1000);
 315 #endif
 316
 317                 goto out;
 318         }
 319
 320         if (xsk_prod_nb_free(fq, free_thresh) >= free_thresh)
 321                 (void)reserve_fill_queue(umem, ETH_AF_XDP_RX_BATCH_SIZE, NULL);
 322
 323         for (i = 0; i < rcvd; i++) {
 324                 const struct xdp_desc *desc;
 325                 uint64_t addr;
 326                 uint32_t len;
 327                 void *pkt;
 328
 329                 desc = xsk_ring_cons__rx_desc(rx, idx_rx++);
 330                 addr = desc->addr;
 331                 len = desc->len;
 332                 pkt = xsk_umem__get_data(rxq->umem->mz->addr, addr);
 333
 334                 rte_memcpy(rte_pktmbuf_mtod(mbufs[i], void *), pkt, len);
 335                 rte_ring_enqueue(umem->buf_ring, (void *)addr);
 336                 rte_pktmbuf_pkt_len(mbufs[i]) = len;
 337                 rte_pktmbuf_data_len(mbufs[i]) = len;
 338                 rx_bytes += len;
 339                 bufs[i] = mbufs[i];
 340         }
 341
 342         xsk_ring_cons__release(rx, rcvd);
 343
 344         /* statistics */
 345         rxq->stats.rx_pkts += rcvd;
 346         rxq->stats.rx_bytes += rx_bytes;
 347
 348 out:
 349         if (rcvd != nb_pkts)
 350                 rte_mempool_put_bulk(rxq->mb_pool, (void **)&mbufs[rcvd],
 351                                      nb_pkts - rcvd);
 352
 353         return rcvd;
 354 }
 355 #endif
 356
 357 static uint16_t
 358 eth_af_xdp_rx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 359 {
 360         nb_pkts = RTE_MIN(nb_pkts, ETH_AF_XDP_RX_BATCH_SIZE);
 361
 362 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 363         return af_xdp_rx_zc(queue, bufs, nb_pkts);
 364 #else
 365         return af_xdp_rx_cp(queue, bufs, nb_pkts);
 366 #endif
 367 }
 368
 369 static void
 370 pull_umem_cq(struct xsk_umem_info *umem, int size)
 371 {
 372         struct xsk_ring_cons *cq = &umem->cq;
 373         size_t i, n;
 374         uint32_t idx_cq = 0;
 375
 376         n = xsk_ring_cons__peek(cq, size, &idx_cq);
 377
 378         for (i = 0; i < n; i++) {
 379                 uint64_t addr;
 380                 addr = *xsk_ring_cons__comp_addr(cq, idx_cq++);
 381 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 382                 addr = xsk_umem__extract_addr(addr);
 383                 rte_pktmbuf_free((struct rte_mbuf *)
 384                                         xsk_umem__get_data(umem->buffer,
 385                                         addr + umem->mb_pool->header_size));
 386 #else
 387                 rte_ring_enqueue(umem->buf_ring, (void *)addr);
 388 #endif
 389         }
 390
 391         xsk_ring_cons__release(cq, n);
 392 }
 393
 394 static void
 395 kick_tx(struct pkt_tx_queue *txq)
 396 {
 397         struct xsk_umem_info *umem = txq->umem;
 398
 399 #if defined(XDP_USE_NEED_WAKEUP)
 400         if (xsk_ring_prod__needs_wakeup(&txq->tx))
 401 #endif
 402                 while (send(xsk_socket__fd(txq->pair->xsk), NULL,
 403                             0, MSG_DONTWAIT) < 0) {
 404                         /* some thing unexpected */
 405                         if (errno != EBUSY && errno != EAGAIN && errno != EINTR)
 406                                 break;
 407
 408                         /* pull from completion queue to leave more space */
 409                         if (errno == EAGAIN)
 410                                 pull_umem_cq(umem, ETH_AF_XDP_TX_BATCH_SIZE);
 411                 }
 412 #ifndef XDP_UMEM_UNALIGNED_CHUNK_FLAG
 413         pull_umem_cq(umem, ETH_AF_XDP_TX_BATCH_SIZE);
 414 #endif
 415 }
 416
 417 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 418 static uint16_t
 419 af_xdp_tx_zc(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 420 {
 421         struct pkt_tx_queue *txq = queue;
 422         struct xsk_umem_info *umem = txq->umem;
 423         struct rte_mbuf *mbuf;
 424         unsigned long tx_bytes = 0;
 425         int i;
 426         uint32_t idx_tx;
 427         uint16_t count = 0;
 428         struct xdp_desc *desc;
 429         uint64_t addr, offset;
 430
 431         pull_umem_cq(umem, nb_pkts);
 432
 433         for (i = 0; i < nb_pkts; i++) {
 434                 mbuf = bufs[i];
 435
 436                 if (mbuf->pool == umem->mb_pool) {
 437                         if (!xsk_ring_prod__reserve(&txq->tx, 1, &idx_tx)) {
 438                                 kick_tx(txq);
 439                                 goto out;
 440                         }
 441                         desc = xsk_ring_prod__tx_desc(&txq->tx, idx_tx);
 442                         desc->len = mbuf->pkt_len;
 443                         addr = (uint64_t)mbuf - (uint64_t)umem->buffer -
 444                                         umem->mb_pool->header_size;
 445                         offset = rte_pktmbuf_mtod(mbuf, uint64_t) -
 446                                         (uint64_t)mbuf +
 447                                         umem->mb_pool->header_size;
 448                         offset = offset << XSK_UNALIGNED_BUF_OFFSET_SHIFT;
 449                         desc->addr = addr | offset;
 450                         count++;
 451                 } else {
 452                         struct rte_mbuf *local_mbuf =
 453                                         rte_pktmbuf_alloc(umem->mb_pool);
 454                         void *pkt;
 455
 456                         if (local_mbuf == NULL)
 457                                 goto out;
 458
 459                         if (!xsk_ring_prod__reserve(&txq->tx, 1, &idx_tx)) {
 460                                 rte_pktmbuf_free(local_mbuf);
 461                                 kick_tx(txq);
 462                                 goto out;
 463                         }
 464
 465                         desc = xsk_ring_prod__tx_desc(&txq->tx, idx_tx);
 466                         desc->len = mbuf->pkt_len;
 467
 468                         addr = (uint64_t)local_mbuf - (uint64_t)umem->buffer -
 469                                         umem->mb_pool->header_size;
 470                         offset = rte_pktmbuf_mtod(local_mbuf, uint64_t) -
 471                                         (uint64_t)local_mbuf +
 472                                         umem->mb_pool->header_size;
 473                         pkt = xsk_umem__get_data(umem->buffer, addr + offset);
 474                         offset = offset << XSK_UNALIGNED_BUF_OFFSET_SHIFT;
 475                         desc->addr = addr | offset;
 476                         rte_memcpy(pkt, rte_pktmbuf_mtod(mbuf, void *),
 477                                         desc->len);
 478                         rte_pktmbuf_free(mbuf);
 479                         count++;
 480                 }
 481
 482                 tx_bytes += mbuf->pkt_len;
 483         }
 484
 485         kick_tx(txq);
 486
 487 out:
 488         xsk_ring_prod__submit(&txq->tx, count);
 489
 490         txq->stats.tx_pkts += count;
 491         txq->stats.tx_bytes += tx_bytes;
 492         txq->stats.tx_dropped += nb_pkts - count;
 493
 494         return count;
 495 }
 496 #else
 497 static uint16_t
 498 af_xdp_tx_cp(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 499 {
 500         struct pkt_tx_queue *txq = queue;
 501         struct xsk_umem_info *umem = txq->umem;
 502         struct rte_mbuf *mbuf;
 503         void *addrs[ETH_AF_XDP_TX_BATCH_SIZE];
 504         unsigned long tx_bytes = 0;
 505         int i;
 506         uint32_t idx_tx;
 507
 508         nb_pkts = RTE_MIN(nb_pkts, ETH_AF_XDP_TX_BATCH_SIZE);
 509
 510         pull_umem_cq(umem, nb_pkts);
 511
 512         nb_pkts = rte_ring_dequeue_bulk(umem->buf_ring, addrs,
 513                                         nb_pkts, NULL);
 514         if (nb_pkts == 0)
 515                 return 0;
 516
 517         if (xsk_ring_prod__reserve(&txq->tx, nb_pkts, &idx_tx) != nb_pkts) {
 518                 kick_tx(txq);
 519                 rte_ring_enqueue_bulk(umem->buf_ring, addrs, nb_pkts, NULL);
 520                 return 0;
 521         }
 522
 523         for (i = 0; i < nb_pkts; i++) {
 524                 struct xdp_desc *desc;
 525                 void *pkt;
 526
 527                 desc = xsk_ring_prod__tx_desc(&txq->tx, idx_tx + i);
 528                 mbuf = bufs[i];
 529                 desc->len = mbuf->pkt_len;
 530
 531                 desc->addr = (uint64_t)addrs[i];
 532                 pkt = xsk_umem__get_data(umem->mz->addr,
 533                                          desc->addr);
 534                 rte_memcpy(pkt, rte_pktmbuf_mtod(mbuf, void *), desc->len);
 535                 tx_bytes += mbuf->pkt_len;
 536                 rte_pktmbuf_free(mbuf);
 537         }
 538
 539         xsk_ring_prod__submit(&txq->tx, nb_pkts);
 540
 541         kick_tx(txq);
 542
 543         txq->stats.tx_pkts += nb_pkts;
 544         txq->stats.tx_bytes += tx_bytes;
 545
 546         return nb_pkts;
 547 }
 548 #endif
 549
 550 static uint16_t
 551 eth_af_xdp_tx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 552 {
 553 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 554         return af_xdp_tx_zc(queue, bufs, nb_pkts);
 555 #else
 556         return af_xdp_tx_cp(queue, bufs, nb_pkts);
 557 #endif
 558 }
 559
 560 static int
 561 eth_dev_start(struct rte_eth_dev *dev)
 562 {
 563         dev->data->dev_link.link_status = ETH_LINK_UP;
 564
 565         return 0;
 566 }
 567
 568 /* This function gets called when the current port gets stopped. */
 569 static void
 570 eth_dev_stop(struct rte_eth_dev *dev)
 571 {
 572         dev->data->dev_link.link_status = ETH_LINK_DOWN;
 573 }
 574
 575 static int
 576 eth_dev_configure(struct rte_eth_dev *dev)
 577 {
 578         /* rx/tx must be paired */
 579         if (dev->data->nb_rx_queues != dev->data->nb_tx_queues)
 580                 return -EINVAL;
 581
 582         return 0;
 583 }
 584
 585 static int
 586 eth_dev_info(struct rte_eth_dev *dev, struct rte_eth_dev_info *dev_info)
 587 {
 588         struct pmd_internals *internals = dev->data->dev_private;
 589
 590         dev_info->if_index = internals->if_index;
 591         dev_info->max_mac_addrs = 1;
 592         dev_info->max_rx_pktlen = ETH_FRAME_LEN;
 593         dev_info->max_rx_queues = internals->queue_cnt;
 594         dev_info->max_tx_queues = internals->queue_cnt;
 595
 596         dev_info->min_mtu = RTE_ETHER_MIN_MTU;
 597 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 598         dev_info->max_mtu = getpagesize() -
 599                                 sizeof(struct rte_mempool_objhdr) -
 600                                 sizeof(struct rte_mbuf) -
 601                                 RTE_PKTMBUF_HEADROOM - XDP_PACKET_HEADROOM;
 602 #else
 603         dev_info->max_mtu = ETH_AF_XDP_FRAME_SIZE - XDP_PACKET_HEADROOM;
 604 #endif
 605
 606         dev_info->default_rxportconf.nb_queues = 1;
 607         dev_info->default_txportconf.nb_queues = 1;
 608         dev_info->default_rxportconf.ring_size = ETH_AF_XDP_DFLT_NUM_DESCS;
 609         dev_info->default_txportconf.ring_size = ETH_AF_XDP_DFLT_NUM_DESCS;
 610
 611         return 0;
 612 }
 613
 614 static int
 615 eth_stats_get(struct rte_eth_dev *dev, struct rte_eth_stats *stats)
 616 {
 617         struct pmd_internals *internals = dev->data->dev_private;
 618         struct xdp_statistics xdp_stats;
 619         struct pkt_rx_queue *rxq;
 620         struct pkt_tx_queue *txq;
 621         socklen_t optlen;
 622         int i, ret;
 623
 624         for (i = 0; i < dev->data->nb_rx_queues; i++) {
 625                 optlen = sizeof(struct xdp_statistics);
 626                 rxq = &internals->rx_queues[i];
 627                 txq = rxq->pair;
 628                 stats->q_ipackets[i] = rxq->stats.rx_pkts;
 629                 stats->q_ibytes[i] = rxq->stats.rx_bytes;
 630
 631                 stats->q_opackets[i] = txq->stats.tx_pkts;
 632                 stats->q_obytes[i] = txq->stats.tx_bytes;
 633
 634                 stats->ipackets += stats->q_ipackets[i];
 635                 stats->ibytes += stats->q_ibytes[i];
 636                 stats->imissed += rxq->stats.rx_dropped;
 637                 stats->oerrors += txq->stats.tx_dropped;
 638                 ret = getsockopt(xsk_socket__fd(rxq->xsk), SOL_XDP,
 639                                 XDP_STATISTICS, &xdp_stats, &optlen);
 640                 if (ret != 0) {
 641                         AF_XDP_LOG(ERR, "getsockopt() failed for XDP_STATISTICS.\n");
 642                         return -1;
 643                 }
 644                 stats->imissed += xdp_stats.rx_dropped;
 645
 646                 stats->opackets += stats->q_opackets[i];
 647                 stats->obytes += stats->q_obytes[i];
 648         }
 649
 650         return 0;
 651 }
 652
 653 static int
 654 eth_stats_reset(struct rte_eth_dev *dev)
 655 {
 656         struct pmd_internals *internals = dev->data->dev_private;
 657         int i;
 658
 659         for (i = 0; i < internals->queue_cnt; i++) {
 660                 memset(&internals->rx_queues[i].stats, 0,
 661                                         sizeof(struct rx_stats));
 662                 memset(&internals->tx_queues[i].stats, 0,
 663                                         sizeof(struct tx_stats));
 664         }
 665
 666         return 0;
 667 }
 668
 669 static void
 670 remove_xdp_program(struct pmd_internals *internals)
 671 {
 672         uint32_t curr_prog_id = 0;
 673
 674         if (bpf_get_link_xdp_id(internals->if_index, &curr_prog_id,
 675                                 XDP_FLAGS_UPDATE_IF_NOEXIST)) {
 676                 AF_XDP_LOG(ERR, "bpf_get_link_xdp_id failed\n");
 677                 return;
 678         }
 679         bpf_set_link_xdp_fd(internals->if_index, -1,
 680                         XDP_FLAGS_UPDATE_IF_NOEXIST);
 681 }
 682
 683 static void
 684 xdp_umem_destroy(struct xsk_umem_info *umem)
 685 {
 686 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 687         rte_mempool_free(umem->mb_pool);
 688         umem->mb_pool = NULL;
 689 #else
 690         rte_memzone_free(umem->mz);
 691         umem->mz = NULL;
 692
 693         rte_ring_free(umem->buf_ring);
 694         umem->buf_ring = NULL;
 695 #endif
 696
 697         rte_free(umem);
 698         umem = NULL;
 699 }
 700
 701 static void
 702 eth_dev_close(struct rte_eth_dev *dev)
 703 {
 704         struct pmd_internals *internals = dev->data->dev_private;
 705         struct pkt_rx_queue *rxq;
 706         int i;
 707
 708         AF_XDP_LOG(INFO, "Closing AF_XDP ethdev on numa socket %u\n",
 709                 rte_socket_id());
 710
 711         for (i = 0; i < internals->queue_cnt; i++) {
 712                 rxq = &internals->rx_queues[i];
 713                 if (rxq->umem == NULL)
 714                         break;
 715                 xsk_socket__delete(rxq->xsk);
 716                 (void)xsk_umem__delete(rxq->umem->umem);
 717                 xdp_umem_destroy(rxq->umem);
 718
 719                 /* free pkt_tx_queue */
 720                 rte_free(rxq->pair);
 721                 rte_free(rxq);
 722         }
 723
 724         /*
 725          * MAC is not allocated dynamically, setting it to NULL would prevent
 726          * from releasing it in rte_eth_dev_release_port.
 727          */
 728         dev->data->mac_addrs = NULL;
 729
 730         remove_xdp_program(internals);
 731 }
 732
 733 static void
 734 eth_queue_release(void *q __rte_unused)
 735 {
 736 }
 737
 738 static int
 739 eth_link_update(struct rte_eth_dev *dev __rte_unused,
 740                 int wait_to_complete __rte_unused)
 741 {
 742         return 0;
 743 }
 744
 745 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 746 static inline uint64_t get_base_addr(struct rte_mempool *mp)
 747 {
 748         struct rte_mempool_memhdr *memhdr;
 749
 750         memhdr = STAILQ_FIRST(&mp->mem_list);
 751         return (uint64_t)memhdr->addr & ~(getpagesize() - 1);
 752 }
 753
 754 static struct
 755 xsk_umem_info *xdp_umem_configure(struct pmd_internals *internals __rte_unused,
 756                                   struct pkt_rx_queue *rxq)
 757 {
 758         struct xsk_umem_info *umem;
 759         int ret;
 760         struct xsk_umem_config usr_config = {
 761                 .fill_size = ETH_AF_XDP_DFLT_NUM_DESCS,
 762                 .comp_size = ETH_AF_XDP_DFLT_NUM_DESCS,
 763                 .flags = XDP_UMEM_UNALIGNED_CHUNK_FLAG};
 764         void *base_addr = NULL;
 765         struct rte_mempool *mb_pool = rxq->mb_pool;
 766
 767         usr_config.frame_size = rte_mempool_calc_obj_size(mb_pool->elt_size,
 768                                                                 mb_pool->flags,
 769                                                                 NULL);
 770         usr_config.frame_headroom = mb_pool->header_size +
 771                                         sizeof(struct rte_mbuf) +
 772                                         rte_pktmbuf_priv_size(mb_pool) +
 773                                         RTE_PKTMBUF_HEADROOM;
 774
 775         umem = rte_zmalloc_socket("umem", sizeof(*umem), 0, rte_socket_id());
 776         if (umem == NULL) {
 777                 AF_XDP_LOG(ERR, "Failed to allocate umem info");
 778                 return NULL;
 779         }
 780
 781         umem->mb_pool = mb_pool;
 782         base_addr = (void *)get_base_addr(mb_pool);
 783
 784         ret = xsk_umem__create(&umem->umem, base_addr,
 785                                mb_pool->populated_size * usr_config.frame_size,
 786                                &umem->fq, &umem->cq,
 787                                &usr_config);
 788
 789         if (ret) {
 790                 AF_XDP_LOG(ERR, "Failed to create umem");
 791                 goto err;
 792         }
 793         umem->buffer = base_addr;
 794
 795 #else
 796 static struct
 797 xsk_umem_info *xdp_umem_configure(struct pmd_internals *internals,
 798                                   struct pkt_rx_queue *rxq)
 799 {
 800         struct xsk_umem_info *umem;
 801         const struct rte_memzone *mz;
 802         struct xsk_umem_config usr_config = {
 803                 .fill_size = ETH_AF_XDP_DFLT_NUM_DESCS,
 804                 .comp_size = ETH_AF_XDP_DFLT_NUM_DESCS,
 805                 .frame_size = ETH_AF_XDP_FRAME_SIZE,
 806                 .frame_headroom = 0 };
 807         char ring_name[RTE_RING_NAMESIZE];
 808         char mz_name[RTE_MEMZONE_NAMESIZE];
 809         int ret;
 810         uint64_t i;
 811
 812         umem = rte_zmalloc_socket("umem", sizeof(*umem), 0, rte_socket_id());
 813         if (umem == NULL) {
 814                 AF_XDP_LOG(ERR, "Failed to allocate umem info");
 815                 return NULL;
 816         }
 817
 818         snprintf(ring_name, sizeof(ring_name), "af_xdp_ring_%s_%u",
 819                        internals->if_name, rxq->xsk_queue_idx);
 820         umem->buf_ring = rte_ring_create(ring_name,
 821                                          ETH_AF_XDP_NUM_BUFFERS,
 822                                          rte_socket_id(),
 823                                          RING_F_SP_ENQ | RING_F_SC_DEQ);
 824         if (umem->buf_ring == NULL) {
 825                 AF_XDP_LOG(ERR, "Failed to create rte_ring\n");
 826                 goto err;
 827         }
 828
 829         for (i = 0; i < ETH_AF_XDP_NUM_BUFFERS; i++)
 830                 rte_ring_enqueue(umem->buf_ring,
 831                                  (void *)(i * ETH_AF_XDP_FRAME_SIZE));
 832
 833         snprintf(mz_name, sizeof(mz_name), "af_xdp_umem_%s_%u",
 834                        internals->if_name, rxq->xsk_queue_idx);
 835         mz = rte_memzone_reserve_aligned(mz_name,
 836                         ETH_AF_XDP_NUM_BUFFERS * ETH_AF_XDP_FRAME_SIZE,
 837                         rte_socket_id(), RTE_MEMZONE_IOVA_CONTIG,
 838                         getpagesize());
 839         if (mz == NULL) {
 840                 AF_XDP_LOG(ERR, "Failed to reserve memzone for af_xdp umem.\n");
 841                 goto err;
 842         }
 843
 844         ret = xsk_umem__create(&umem->umem, mz->addr,
 845                                ETH_AF_XDP_NUM_BUFFERS * ETH_AF_XDP_FRAME_SIZE,
 846                                &umem->fq, &umem->cq,
 847                                &usr_config);
 848
 849         if (ret) {
 850                 AF_XDP_LOG(ERR, "Failed to create umem");
 851                 goto err;
 852         }
 853         umem->mz = mz;
 854
 855 #endif
 856         return umem;
 857
 858 err:
 859         xdp_umem_destroy(umem);
 860         return NULL;
 861 }
 862
 863 static int
 864 xsk_configure(struct pmd_internals *internals, struct pkt_rx_queue *rxq,
 865               int ring_size)
 866 {
 867         struct xsk_socket_config cfg;
 868         struct pkt_tx_queue *txq = rxq->pair;
 869         int ret = 0;
 870         int reserve_size = ETH_AF_XDP_DFLT_NUM_DESCS / 2;
 871         struct rte_mbuf *fq_bufs[reserve_size];
 872
 873         rxq->umem = xdp_umem_configure(internals, rxq);
 874         if (rxq->umem == NULL)
 875                 return -ENOMEM;
 876         txq->umem = rxq->umem;
 877
 878         cfg.rx_size = ring_size;
 879         cfg.tx_size = ring_size;
 880         cfg.libbpf_flags = 0;
 881         cfg.xdp_flags = XDP_FLAGS_UPDATE_IF_NOEXIST;
 882         cfg.bind_flags = 0;
 883
 884 #if defined(XDP_USE_NEED_WAKEUP)
 885         cfg.bind_flags |= XDP_USE_NEED_WAKEUP;
 886 #endif
 887
 888         ret = xsk_socket__create(&rxq->xsk, internals->if_name,
 889                         rxq->xsk_queue_idx, rxq->umem->umem, &rxq->rx,
 890                         &txq->tx, &cfg);
 891         if (ret) {
 892                 AF_XDP_LOG(ERR, "Failed to create xsk socket.\n");
 893                 goto err;
 894         }
 895
 896 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 897         if (rte_pktmbuf_alloc_bulk(rxq->umem->mb_pool, fq_bufs, reserve_size)) {
 898                 AF_XDP_LOG(DEBUG, "Failed to get enough buffers for fq.\n");
 899                 goto err;
 900         }
 901 #endif
 902         ret = reserve_fill_queue(rxq->umem, reserve_size, fq_bufs);
 903         if (ret) {
 904                 xsk_socket__delete(rxq->xsk);
 905                 AF_XDP_LOG(ERR, "Failed to reserve fill queue.\n");
 906                 goto err;
 907         }
 908
 909         return 0;
 910
 911 err:
 912         xdp_umem_destroy(rxq->umem);
 913
 914         return ret;
 915 }
 916
 917 static int
 918 eth_rx_queue_setup(struct rte_eth_dev *dev,
 919                    uint16_t rx_queue_id,
 920                    uint16_t nb_rx_desc,
 921                    unsigned int socket_id __rte_unused,
 922                    const struct rte_eth_rxconf *rx_conf __rte_unused,
 923                    struct rte_mempool *mb_pool)
 924 {
 925         struct pmd_internals *internals = dev->data->dev_private;
 926         struct pkt_rx_queue *rxq;
 927         int ret;
 928
 929         rxq = &internals->rx_queues[rx_queue_id];
 930
 931         AF_XDP_LOG(INFO, "Set up rx queue, rx queue id: %d, xsk queue id: %d\n",
 932                    rx_queue_id, rxq->xsk_queue_idx);
 933
 934 #ifndef XDP_UMEM_UNALIGNED_CHUNK_FLAG
 935         uint32_t buf_size, data_size;
 936
 937         /* Now get the space available for data in the mbuf */
 938         buf_size = rte_pktmbuf_data_room_size(mb_pool) -
 939                 RTE_PKTMBUF_HEADROOM;
 940         data_size = ETH_AF_XDP_FRAME_SIZE;
 941
 942         if (data_size > buf_size) {
 943                 AF_XDP_LOG(ERR, "%s: %d bytes will not fit in mbuf (%d bytes)\n",
 944                         dev->device->name, data_size, buf_size);
 945                 ret = -ENOMEM;
 946                 goto err;
 947         }
 948 #endif
 949
 950         rxq->mb_pool = mb_pool;
 951
 952         if (xsk_configure(internals, rxq, nb_rx_desc)) {
 953                 AF_XDP_LOG(ERR, "Failed to configure xdp socket\n");
 954                 ret = -EINVAL;
 955                 goto err;
 956         }
 957
 958         rxq->fds[0].fd = xsk_socket__fd(rxq->xsk);
 959         rxq->fds[0].events = POLLIN;
 960
 961         dev->data->rx_queues[rx_queue_id] = rxq;
 962         return 0;
 963
 964 err:
 965         return ret;
 966 }
 967
 968 static int
 969 eth_tx_queue_setup(struct rte_eth_dev *dev,
 970                    uint16_t tx_queue_id,
 971                    uint16_t nb_tx_desc __rte_unused,
 972                    unsigned int socket_id __rte_unused,
 973                    const struct rte_eth_txconf *tx_conf __rte_unused)
 974 {
 975         struct pmd_internals *internals = dev->data->dev_private;
 976         struct pkt_tx_queue *txq;
 977
 978         txq = &internals->tx_queues[tx_queue_id];
 979
 980         dev->data->tx_queues[tx_queue_id] = txq;
 981         return 0;
 982 }
 983
 984 static int
 985 eth_dev_mtu_set(struct rte_eth_dev *dev, uint16_t mtu)
 986 {
 987         struct pmd_internals *internals = dev->data->dev_private;
 988         struct ifreq ifr = { .ifr_mtu = mtu };
 989         int ret;
 990         int s;
 991
 992         s = socket(PF_INET, SOCK_DGRAM, 0);
 993         if (s < 0)
 994                 return -EINVAL;
 995
 996         strlcpy(ifr.ifr_name, internals->if_name, IFNAMSIZ);
 997         ret = ioctl(s, SIOCSIFMTU, &ifr);
 998         close(s);
 999
1000         return (ret < 0) ? -errno : 0;
1001 }
1002
1003 static int
1004 eth_dev_change_flags(char *if_name, uint32_t flags, uint32_t mask)
1005 {
1006         struct ifreq ifr;
1007         int ret = 0;
1008         int s;
1009
1010         s = socket(PF_INET, SOCK_DGRAM, 0);
1011         if (s < 0)
1012                 return -errno;
1013
1014         strlcpy(ifr.ifr_name, if_name, IFNAMSIZ);
1015         if (ioctl(s, SIOCGIFFLAGS, &ifr) < 0) {
1016                 ret = -errno;
1017                 goto out;
1018         }
1019         ifr.ifr_flags &= mask;
1020         ifr.ifr_flags |= flags;
1021         if (ioctl(s, SIOCSIFFLAGS, &ifr) < 0) {
1022                 ret = -errno;
1023                 goto out;
1024         }
1025 out:
1026         close(s);
1027         return ret;
1028 }
1029
1030 static int
1031 eth_dev_promiscuous_enable(struct rte_eth_dev *dev)
1032 {
1033         struct pmd_internals *internals = dev->data->dev_private;
1034
1035         return eth_dev_change_flags(internals->if_name, IFF_PROMISC, ~0);
1036 }
1037
1038 static int
1039 eth_dev_promiscuous_disable(struct rte_eth_dev *dev)
1040 {
1041         struct pmd_internals *internals = dev->data->dev_private;
1042
1043         return eth_dev_change_flags(internals->if_name, 0, ~IFF_PROMISC);
1044 }
1045
1046 static const struct eth_dev_ops ops = {
1047         .dev_start = eth_dev_start,
1048         .dev_stop = eth_dev_stop,
1049         .dev_close = eth_dev_close,
1050         .dev_configure = eth_dev_configure,
1051         .dev_infos_get = eth_dev_info,
1052         .mtu_set = eth_dev_mtu_set,
1053         .promiscuous_enable = eth_dev_promiscuous_enable,
1054         .promiscuous_disable = eth_dev_promiscuous_disable,
1055         .rx_queue_setup = eth_rx_queue_setup,
1056         .tx_queue_setup = eth_tx_queue_setup,
1057         .rx_queue_release = eth_queue_release,
1058         .tx_queue_release = eth_queue_release,
1059         .link_update = eth_link_update,
1060         .stats_get = eth_stats_get,
1061         .stats_reset = eth_stats_reset,
1062 };
1063
1064 /** parse integer from integer argument */
1065 static int
1066 parse_integer_arg(const char *key __rte_unused,
1067                   const char *value, void *extra_args)
1068 {
1069         int *i = (int *)extra_args;
1070         char *end;
1071
1072         *i = strtol(value, &end, 10);
1073         if (*i < 0) {
1074                 AF_XDP_LOG(ERR, "Argument has to be positive.\n");
1075                 return -EINVAL;
1076         }
1077
1078         return 0;
1079 }
1080
1081 /** parse name argument */
1082 static int
1083 parse_name_arg(const char *key __rte_unused,
1084                const char *value, void *extra_args)
1085 {
1086         char *name = extra_args;
1087
1088         if (strnlen(value, IFNAMSIZ) > IFNAMSIZ - 1) {
1089                 AF_XDP_LOG(ERR, "Invalid name %s, should be less than %u bytes.\n",
1090                            value, IFNAMSIZ);
1091                 return -EINVAL;
1092         }
1093
1094         strlcpy(name, value, IFNAMSIZ);
1095
1096         return 0;
1097 }
1098
1099 static int
1100 xdp_get_channels_info(const char *if_name, int *max_queues,
1101                                 int *combined_queues)
1102 {
1103         struct ethtool_channels channels;
1104         struct ifreq ifr;
1105         int fd, ret;
1106
1107         fd = socket(AF_INET, SOCK_DGRAM, 0);
1108         if (fd < 0)
1109                 return -1;
1110
1111         channels.cmd = ETHTOOL_GCHANNELS;
1112         ifr.ifr_data = (void *)&channels;
1113         strncpy(ifr.ifr_name, if_name, IFNAMSIZ);
1114         ret = ioctl(fd, SIOCETHTOOL, &ifr);
1115         if (ret) {
1116                 if (errno == EOPNOTSUPP) {
1117                         ret = 0;
1118                 } else {
1119                         ret = -errno;
1120                         goto out;
1121                 }
1122         }
1123
1124         if (channels.max_combined == 0 || errno == EOPNOTSUPP) {
1125                 /* If the device says it has no channels, then all traffic
1126                  * is sent to a single stream, so max queues = 1.
1127                  */
1128                 *max_queues = 1;
1129                 *combined_queues = 1;
1130         } else {
1131                 *max_queues = channels.max_combined;
1132                 *combined_queues = channels.combined_count;
1133         }
1134
1135  out:
1136         close(fd);
1137         return ret;
1138 }
1139
1140 static int
1141 parse_parameters(struct rte_kvargs *kvlist, char *if_name, int *start_queue,
1142                         int *queue_cnt)
1143 {
1144         int ret;
1145
1146         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_IFACE_ARG,
1147                                  &parse_name_arg, if_name);
1148         if (ret < 0)
1149                 goto free_kvlist;
1150
1151         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_START_QUEUE_ARG,
1152                                  &parse_integer_arg, start_queue);
1153         if (ret < 0)
1154                 goto free_kvlist;
1155
1156         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_QUEUE_COUNT_ARG,
1157                                  &parse_integer_arg, queue_cnt);
1158         if (ret < 0 || *queue_cnt <= 0) {
1159                 ret = -EINVAL;
1160                 goto free_kvlist;
1161         }
1162
1163 free_kvlist:
1164         rte_kvargs_free(kvlist);
1165         return ret;
1166 }
1167
1168 static int
1169 get_iface_info(const char *if_name,
1170                struct rte_ether_addr *eth_addr,
1171                int *if_index)
1172 {
1173         struct ifreq ifr;
1174         int sock = socket(AF_INET, SOCK_DGRAM, IPPROTO_IP);
1175
1176         if (sock < 0)
1177                 return -1;
1178
1179         strlcpy(ifr.ifr_name, if_name, IFNAMSIZ);
1180         if (ioctl(sock, SIOCGIFINDEX, &ifr))
1181                 goto error;
1182
1183         *if_index = ifr.ifr_ifindex;
1184
1185         if (ioctl(sock, SIOCGIFHWADDR, &ifr))
1186                 goto error;
1187
1188         rte_memcpy(eth_addr, ifr.ifr_hwaddr.sa_data, RTE_ETHER_ADDR_LEN);
1189
1190         close(sock);
1191         return 0;
1192
1193 error:
1194         close(sock);
1195         return -1;
1196 }
1197
1198 static struct rte_eth_dev *
1199 init_internals(struct rte_vdev_device *dev, const char *if_name,
1200                         int start_queue_idx, int queue_cnt)
1201 {
1202         const char *name = rte_vdev_device_name(dev);
1203         const unsigned int numa_node = dev->device.numa_node;
1204         struct pmd_internals *internals;
1205         struct rte_eth_dev *eth_dev;
1206         int ret;
1207         int i;
1208
1209         internals = rte_zmalloc_socket(name, sizeof(*internals), 0, numa_node);
1210         if (internals == NULL)
1211                 return NULL;
1212
1213         internals->start_queue_idx = start_queue_idx;
1214         internals->queue_cnt = queue_cnt;
1215         strlcpy(internals->if_name, if_name, IFNAMSIZ);
1216
1217         if (xdp_get_channels_info(if_name, &internals->max_queue_cnt,
1218                                   &internals->combined_queue_cnt)) {
1219                 AF_XDP_LOG(ERR, "Failed to get channel info of interface: %s\n",
1220                                 if_name);
1221                 goto err_free_internals;
1222         }
1223
1224         if (queue_cnt > internals->combined_queue_cnt) {
1225                 AF_XDP_LOG(ERR, "Specified queue count %d is larger than combined queue count %d.\n",
1226                                 queue_cnt, internals->combined_queue_cnt);
1227                 goto err_free_internals;
1228         }
1229
1230         internals->rx_queues = rte_zmalloc_socket(NULL,
1231                                         sizeof(struct pkt_rx_queue) * queue_cnt,
1232                                         0, numa_node);
1233         if (internals->rx_queues == NULL) {
1234                 AF_XDP_LOG(ERR, "Failed to allocate memory for rx queues.\n");
1235                 goto err_free_internals;
1236         }
1237
1238         internals->tx_queues = rte_zmalloc_socket(NULL,
1239                                         sizeof(struct pkt_tx_queue) * queue_cnt,
1240                                         0, numa_node);
1241         if (internals->tx_queues == NULL) {
1242                 AF_XDP_LOG(ERR, "Failed to allocate memory for tx queues.\n");
1243                 goto err_free_rx;
1244         }
1245         for (i = 0; i < queue_cnt; i++) {
1246                 internals->tx_queues[i].pair = &internals->rx_queues[i];
1247                 internals->rx_queues[i].pair = &internals->tx_queues[i];
1248                 internals->rx_queues[i].xsk_queue_idx = start_queue_idx + i;
1249                 internals->tx_queues[i].xsk_queue_idx = start_queue_idx + i;
1250         }
1251
1252         ret = get_iface_info(if_name, &internals->eth_addr,
1253                              &internals->if_index);
1254         if (ret)
1255                 goto err_free_tx;
1256
1257         eth_dev = rte_eth_vdev_allocate(dev, 0);
1258         if (eth_dev == NULL)
1259                 goto err_free_tx;
1260
1261         eth_dev->data->dev_private = internals;
1262         eth_dev->data->dev_link = pmd_link;
1263         eth_dev->data->mac_addrs = &internals->eth_addr;
1264         eth_dev->dev_ops = &ops;
1265         eth_dev->rx_pkt_burst = eth_af_xdp_rx;
1266         eth_dev->tx_pkt_burst = eth_af_xdp_tx;
1267         /* Let rte_eth_dev_close() release the port resources. */
1268         eth_dev->data->dev_flags |= RTE_ETH_DEV_CLOSE_REMOVE;
1269
1270 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
1271         AF_XDP_LOG(INFO, "Zero copy between umem and mbuf enabled.\n");
1272 #endif
1273
1274         return eth_dev;
1275
1276 err_free_tx:
1277         rte_free(internals->tx_queues);
1278 err_free_rx:
1279         rte_free(internals->rx_queues);
1280 err_free_internals:
1281         rte_free(internals);
1282         return NULL;
1283 }
1284
1285 static int
1286 rte_pmd_af_xdp_probe(struct rte_vdev_device *dev)
1287 {
1288         struct rte_kvargs *kvlist;
1289         char if_name[IFNAMSIZ] = {'\0'};
1290         int xsk_start_queue_idx = ETH_AF_XDP_DFLT_START_QUEUE_IDX;
1291         int xsk_queue_cnt = ETH_AF_XDP_DFLT_QUEUE_COUNT;
1292         struct rte_eth_dev *eth_dev = NULL;
1293         const char *name;
1294
1295         AF_XDP_LOG(INFO, "Initializing pmd_af_xdp for %s\n",
1296                 rte_vdev_device_name(dev));
1297
1298         name = rte_vdev_device_name(dev);
1299         if (rte_eal_process_type() == RTE_PROC_SECONDARY &&
1300                 strlen(rte_vdev_device_args(dev)) == 0) {
1301                 eth_dev = rte_eth_dev_attach_secondary(name);
1302                 if (eth_dev == NULL) {
1303                         AF_XDP_LOG(ERR, "Failed to probe %s\n", name);
1304                         return -EINVAL;
1305                 }
1306                 eth_dev->dev_ops = &ops;
1307                 rte_eth_dev_probing_finish(eth_dev);
1308                 return 0;
1309         }
1310
1311         kvlist = rte_kvargs_parse(rte_vdev_device_args(dev), valid_arguments);
1312         if (kvlist == NULL) {
1313                 AF_XDP_LOG(ERR, "Invalid kvargs key\n");
1314                 return -EINVAL;
1315         }
1316
1317         if (dev->device.numa_node == SOCKET_ID_ANY)
1318                 dev->device.numa_node = rte_socket_id();
1319
1320         if (parse_parameters(kvlist, if_name, &xsk_start_queue_idx,
1321                              &xsk_queue_cnt) < 0) {
1322                 AF_XDP_LOG(ERR, "Invalid kvargs value\n");
1323                 return -EINVAL;
1324         }
1325
1326         if (strlen(if_name) == 0) {
1327                 AF_XDP_LOG(ERR, "Network interface must be specified\n");
1328                 return -EINVAL;
1329         }
1330
1331         eth_dev = init_internals(dev, if_name, xsk_start_queue_idx,
1332                                         xsk_queue_cnt);
1333         if (eth_dev == NULL) {
1334                 AF_XDP_LOG(ERR, "Failed to init internals\n");
1335                 return -1;
1336         }
1337
1338         rte_eth_dev_probing_finish(eth_dev);
1339
1340         return 0;
1341 }
1342
1343 static int
1344 rte_pmd_af_xdp_remove(struct rte_vdev_device *dev)
1345 {
1346         struct rte_eth_dev *eth_dev = NULL;
1347
1348         AF_XDP_LOG(INFO, "Removing AF_XDP ethdev on numa socket %u\n",
1349                 rte_socket_id());
1350
1351         if (dev == NULL)
1352                 return -1;
1353
1354         /* find the ethdev entry */
1355         eth_dev = rte_eth_dev_allocated(rte_vdev_device_name(dev));
1356         if (eth_dev == NULL)
1357                 return 0;
1358
1359         eth_dev_close(eth_dev);
1360         rte_eth_dev_release_port(eth_dev);
1361
1362
1363         return 0;
1364 }
1365
1366 static struct rte_vdev_driver pmd_af_xdp_drv = {
1367         .probe = rte_pmd_af_xdp_probe,
1368         .remove = rte_pmd_af_xdp_remove,
1369 };
1370
1371 RTE_PMD_REGISTER_VDEV(net_af_xdp, pmd_af_xdp_drv);
1372 RTE_PMD_REGISTER_PARAM_STRING(net_af_xdp,
1373                               "iface=<string> "
1374                               "start_queue=<int> "
1375                               "queue_count=<int> ");
1376
1377 RTE_INIT(af_xdp_init_log)
1378 {
1379         af_xdp_logtype = rte_log_register("pmd.net.af_xdp");
1380         if (af_xdp_logtype >= 0)
1381                 rte_log_set_level(af_xdp_logtype, RTE_LOG_NOTICE);
1382 }