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 RTE_LOG_REGISTER(af_xdp_logtype, pmd.net.af_xdp, NOTICE);
  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         pull_umem_cq(umem, XSK_RING_CONS__DEFAULT_NUM_DESCS);
 400
 401 #if defined(XDP_USE_NEED_WAKEUP)
 402         if (xsk_ring_prod__needs_wakeup(&txq->tx))
 403 #endif
 404                 while (send(xsk_socket__fd(txq->pair->xsk), NULL,
 405                             0, MSG_DONTWAIT) < 0) {
 406                         /* some thing unexpected */
 407                         if (errno != EBUSY && errno != EAGAIN && errno != EINTR)
 408                                 break;
 409
 410                         /* pull from completion queue to leave more space */
 411                         if (errno == EAGAIN)
 412                                 pull_umem_cq(umem,
 413                                              XSK_RING_CONS__DEFAULT_NUM_DESCS);
 414                 }
 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         uint32_t free_thresh = umem->cq.size >> 1;
 431
 432         if (xsk_cons_nb_avail(&umem->cq, free_thresh) >= free_thresh)
 433                 pull_umem_cq(umem, XSK_RING_CONS__DEFAULT_NUM_DESCS);
 434
 435         for (i = 0; i < nb_pkts; i++) {
 436                 mbuf = bufs[i];
 437
 438                 if (mbuf->pool == umem->mb_pool) {
 439                         if (!xsk_ring_prod__reserve(&txq->tx, 1, &idx_tx)) {
 440                                 kick_tx(txq);
 441                                 if (!xsk_ring_prod__reserve(&txq->tx, 1,
 442                                                             &idx_tx))
 443                                         goto out;
 444                         }
 445                         desc = xsk_ring_prod__tx_desc(&txq->tx, idx_tx);
 446                         desc->len = mbuf->pkt_len;
 447                         addr = (uint64_t)mbuf - (uint64_t)umem->buffer -
 448                                         umem->mb_pool->header_size;
 449                         offset = rte_pktmbuf_mtod(mbuf, uint64_t) -
 450                                         (uint64_t)mbuf +
 451                                         umem->mb_pool->header_size;
 452                         offset = offset << XSK_UNALIGNED_BUF_OFFSET_SHIFT;
 453                         desc->addr = addr | offset;
 454                         count++;
 455                 } else {
 456                         struct rte_mbuf *local_mbuf =
 457                                         rte_pktmbuf_alloc(umem->mb_pool);
 458                         void *pkt;
 459
 460                         if (local_mbuf == NULL)
 461                                 goto out;
 462
 463                         if (!xsk_ring_prod__reserve(&txq->tx, 1, &idx_tx)) {
 464                                 rte_pktmbuf_free(local_mbuf);
 465                                 kick_tx(txq);
 466                                 goto out;
 467                         }
 468
 469                         desc = xsk_ring_prod__tx_desc(&txq->tx, idx_tx);
 470                         desc->len = mbuf->pkt_len;
 471
 472                         addr = (uint64_t)local_mbuf - (uint64_t)umem->buffer -
 473                                         umem->mb_pool->header_size;
 474                         offset = rte_pktmbuf_mtod(local_mbuf, uint64_t) -
 475                                         (uint64_t)local_mbuf +
 476                                         umem->mb_pool->header_size;
 477                         pkt = xsk_umem__get_data(umem->buffer, addr + offset);
 478                         offset = offset << XSK_UNALIGNED_BUF_OFFSET_SHIFT;
 479                         desc->addr = addr | offset;
 480                         rte_memcpy(pkt, rte_pktmbuf_mtod(mbuf, void *),
 481                                         desc->len);
 482                         rte_pktmbuf_free(mbuf);
 483                         count++;
 484                 }
 485
 486                 tx_bytes += mbuf->pkt_len;
 487         }
 488
 489         kick_tx(txq);
 490
 491 out:
 492         xsk_ring_prod__submit(&txq->tx, count);
 493
 494         txq->stats.tx_pkts += count;
 495         txq->stats.tx_bytes += tx_bytes;
 496         txq->stats.tx_dropped += nb_pkts - count;
 497
 498         return count;
 499 }
 500 #else
 501 static uint16_t
 502 af_xdp_tx_cp(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 503 {
 504         struct pkt_tx_queue *txq = queue;
 505         struct xsk_umem_info *umem = txq->umem;
 506         struct rte_mbuf *mbuf;
 507         void *addrs[ETH_AF_XDP_TX_BATCH_SIZE];
 508         unsigned long tx_bytes = 0;
 509         int i;
 510         uint32_t idx_tx;
 511
 512         nb_pkts = RTE_MIN(nb_pkts, ETH_AF_XDP_TX_BATCH_SIZE);
 513
 514         pull_umem_cq(umem, nb_pkts);
 515
 516         nb_pkts = rte_ring_dequeue_bulk(umem->buf_ring, addrs,
 517                                         nb_pkts, NULL);
 518         if (nb_pkts == 0)
 519                 return 0;
 520
 521         if (xsk_ring_prod__reserve(&txq->tx, nb_pkts, &idx_tx) != nb_pkts) {
 522                 kick_tx(txq);
 523                 rte_ring_enqueue_bulk(umem->buf_ring, addrs, nb_pkts, NULL);
 524                 return 0;
 525         }
 526
 527         for (i = 0; i < nb_pkts; i++) {
 528                 struct xdp_desc *desc;
 529                 void *pkt;
 530
 531                 desc = xsk_ring_prod__tx_desc(&txq->tx, idx_tx + i);
 532                 mbuf = bufs[i];
 533                 desc->len = mbuf->pkt_len;
 534
 535                 desc->addr = (uint64_t)addrs[i];
 536                 pkt = xsk_umem__get_data(umem->mz->addr,
 537                                          desc->addr);
 538                 rte_memcpy(pkt, rte_pktmbuf_mtod(mbuf, void *), desc->len);
 539                 tx_bytes += mbuf->pkt_len;
 540                 rte_pktmbuf_free(mbuf);
 541         }
 542
 543         xsk_ring_prod__submit(&txq->tx, nb_pkts);
 544
 545         kick_tx(txq);
 546
 547         txq->stats.tx_pkts += nb_pkts;
 548         txq->stats.tx_bytes += tx_bytes;
 549
 550         return nb_pkts;
 551 }
 552 #endif
 553
 554 static uint16_t
 555 eth_af_xdp_tx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 556 {
 557 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 558         return af_xdp_tx_zc(queue, bufs, nb_pkts);
 559 #else
 560         return af_xdp_tx_cp(queue, bufs, nb_pkts);
 561 #endif
 562 }
 563
 564 static int
 565 eth_dev_start(struct rte_eth_dev *dev)
 566 {
 567         dev->data->dev_link.link_status = ETH_LINK_UP;
 568
 569         return 0;
 570 }
 571
 572 /* This function gets called when the current port gets stopped. */
 573 static void
 574 eth_dev_stop(struct rte_eth_dev *dev)
 575 {
 576         dev->data->dev_link.link_status = ETH_LINK_DOWN;
 577 }
 578
 579 static int
 580 eth_dev_configure(struct rte_eth_dev *dev)
 581 {
 582         /* rx/tx must be paired */
 583         if (dev->data->nb_rx_queues != dev->data->nb_tx_queues)
 584                 return -EINVAL;
 585
 586         return 0;
 587 }
 588
 589 static int
 590 eth_dev_info(struct rte_eth_dev *dev, struct rte_eth_dev_info *dev_info)
 591 {
 592         struct pmd_internals *internals = dev->data->dev_private;
 593
 594         dev_info->if_index = internals->if_index;
 595         dev_info->max_mac_addrs = 1;
 596         dev_info->max_rx_pktlen = ETH_FRAME_LEN;
 597         dev_info->max_rx_queues = internals->queue_cnt;
 598         dev_info->max_tx_queues = internals->queue_cnt;
 599
 600         dev_info->min_mtu = RTE_ETHER_MIN_MTU;
 601 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 602         dev_info->max_mtu = getpagesize() -
 603                                 sizeof(struct rte_mempool_objhdr) -
 604                                 sizeof(struct rte_mbuf) -
 605                                 RTE_PKTMBUF_HEADROOM - XDP_PACKET_HEADROOM;
 606 #else
 607         dev_info->max_mtu = ETH_AF_XDP_FRAME_SIZE - XDP_PACKET_HEADROOM;
 608 #endif
 609
 610         dev_info->default_rxportconf.nb_queues = 1;
 611         dev_info->default_txportconf.nb_queues = 1;
 612         dev_info->default_rxportconf.ring_size = ETH_AF_XDP_DFLT_NUM_DESCS;
 613         dev_info->default_txportconf.ring_size = ETH_AF_XDP_DFLT_NUM_DESCS;
 614
 615         return 0;
 616 }
 617
 618 static int
 619 eth_stats_get(struct rte_eth_dev *dev, struct rte_eth_stats *stats)
 620 {
 621         struct pmd_internals *internals = dev->data->dev_private;
 622         struct xdp_statistics xdp_stats;
 623         struct pkt_rx_queue *rxq;
 624         struct pkt_tx_queue *txq;
 625         socklen_t optlen;
 626         int i, ret;
 627
 628         for (i = 0; i < dev->data->nb_rx_queues; i++) {
 629                 optlen = sizeof(struct xdp_statistics);
 630                 rxq = &internals->rx_queues[i];
 631                 txq = rxq->pair;
 632                 stats->q_ipackets[i] = rxq->stats.rx_pkts;
 633                 stats->q_ibytes[i] = rxq->stats.rx_bytes;
 634
 635                 stats->q_opackets[i] = txq->stats.tx_pkts;
 636                 stats->q_obytes[i] = txq->stats.tx_bytes;
 637
 638                 stats->ipackets += stats->q_ipackets[i];
 639                 stats->ibytes += stats->q_ibytes[i];
 640                 stats->imissed += rxq->stats.rx_dropped;
 641                 stats->oerrors += txq->stats.tx_dropped;
 642                 ret = getsockopt(xsk_socket__fd(rxq->xsk), SOL_XDP,
 643                                 XDP_STATISTICS, &xdp_stats, &optlen);
 644                 if (ret != 0) {
 645                         AF_XDP_LOG(ERR, "getsockopt() failed for XDP_STATISTICS.\n");
 646                         return -1;
 647                 }
 648                 stats->imissed += xdp_stats.rx_dropped;
 649
 650                 stats->opackets += stats->q_opackets[i];
 651                 stats->obytes += stats->q_obytes[i];
 652         }
 653
 654         return 0;
 655 }
 656
 657 static int
 658 eth_stats_reset(struct rte_eth_dev *dev)
 659 {
 660         struct pmd_internals *internals = dev->data->dev_private;
 661         int i;
 662
 663         for (i = 0; i < internals->queue_cnt; i++) {
 664                 memset(&internals->rx_queues[i].stats, 0,
 665                                         sizeof(struct rx_stats));
 666                 memset(&internals->tx_queues[i].stats, 0,
 667                                         sizeof(struct tx_stats));
 668         }
 669
 670         return 0;
 671 }
 672
 673 static void
 674 remove_xdp_program(struct pmd_internals *internals)
 675 {
 676         uint32_t curr_prog_id = 0;
 677
 678         if (bpf_get_link_xdp_id(internals->if_index, &curr_prog_id,
 679                                 XDP_FLAGS_UPDATE_IF_NOEXIST)) {
 680                 AF_XDP_LOG(ERR, "bpf_get_link_xdp_id failed\n");
 681                 return;
 682         }
 683         bpf_set_link_xdp_fd(internals->if_index, -1,
 684                         XDP_FLAGS_UPDATE_IF_NOEXIST);
 685 }
 686
 687 static void
 688 xdp_umem_destroy(struct xsk_umem_info *umem)
 689 {
 690 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 691         umem->mb_pool = NULL;
 692 #else
 693         rte_memzone_free(umem->mz);
 694         umem->mz = NULL;
 695
 696         rte_ring_free(umem->buf_ring);
 697         umem->buf_ring = NULL;
 698 #endif
 699
 700         rte_free(umem);
 701         umem = NULL;
 702 }
 703
 704 static void
 705 eth_dev_close(struct rte_eth_dev *dev)
 706 {
 707         struct pmd_internals *internals = dev->data->dev_private;
 708         struct pkt_rx_queue *rxq;
 709         int i;
 710
 711         AF_XDP_LOG(INFO, "Closing AF_XDP ethdev on numa socket %u\n",
 712                 rte_socket_id());
 713
 714         for (i = 0; i < internals->queue_cnt; i++) {
 715                 rxq = &internals->rx_queues[i];
 716                 if (rxq->umem == NULL)
 717                         break;
 718                 xsk_socket__delete(rxq->xsk);
 719                 (void)xsk_umem__delete(rxq->umem->umem);
 720                 xdp_umem_destroy(rxq->umem);
 721
 722                 /* free pkt_tx_queue */
 723                 rte_free(rxq->pair);
 724                 rte_free(rxq);
 725         }
 726
 727         /*
 728          * MAC is not allocated dynamically, setting it to NULL would prevent
 729          * from releasing it in rte_eth_dev_release_port.
 730          */
 731         dev->data->mac_addrs = NULL;
 732
 733         remove_xdp_program(internals);
 734 }
 735
 736 static void
 737 eth_queue_release(void *q __rte_unused)
 738 {
 739 }
 740
 741 static int
 742 eth_link_update(struct rte_eth_dev *dev __rte_unused,
 743                 int wait_to_complete __rte_unused)
 744 {
 745         return 0;
 746 }
 747
 748 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 749 static inline uint64_t get_base_addr(struct rte_mempool *mp)
 750 {
 751         struct rte_mempool_memhdr *memhdr;
 752
 753         memhdr = STAILQ_FIRST(&mp->mem_list);
 754         return (uint64_t)memhdr->addr & ~(getpagesize() - 1);
 755 }
 756
 757 static struct
 758 xsk_umem_info *xdp_umem_configure(struct pmd_internals *internals __rte_unused,
 759                                   struct pkt_rx_queue *rxq)
 760 {
 761         struct xsk_umem_info *umem;
 762         int ret;
 763         struct xsk_umem_config usr_config = {
 764                 .fill_size = ETH_AF_XDP_DFLT_NUM_DESCS * 2,
 765                 .comp_size = ETH_AF_XDP_DFLT_NUM_DESCS,
 766                 .flags = XDP_UMEM_UNALIGNED_CHUNK_FLAG};
 767         void *base_addr = NULL;
 768         struct rte_mempool *mb_pool = rxq->mb_pool;
 769
 770         usr_config.frame_size = rte_mempool_calc_obj_size(mb_pool->elt_size,
 771                                                                 mb_pool->flags,
 772                                                                 NULL);
 773         usr_config.frame_headroom = mb_pool->header_size +
 774                                         sizeof(struct rte_mbuf) +
 775                                         rte_pktmbuf_priv_size(mb_pool) +
 776                                         RTE_PKTMBUF_HEADROOM;
 777
 778         umem = rte_zmalloc_socket("umem", sizeof(*umem), 0, rte_socket_id());
 779         if (umem == NULL) {
 780                 AF_XDP_LOG(ERR, "Failed to allocate umem info");
 781                 return NULL;
 782         }
 783
 784         umem->mb_pool = mb_pool;
 785         base_addr = (void *)get_base_addr(mb_pool);
 786
 787         ret = xsk_umem__create(&umem->umem, base_addr,
 788                                mb_pool->populated_size * usr_config.frame_size,
 789                                &umem->fq, &umem->cq,
 790                                &usr_config);
 791
 792         if (ret) {
 793                 AF_XDP_LOG(ERR, "Failed to create umem");
 794                 goto err;
 795         }
 796         umem->buffer = base_addr;
 797
 798 #else
 799 static struct
 800 xsk_umem_info *xdp_umem_configure(struct pmd_internals *internals,
 801                                   struct pkt_rx_queue *rxq)
 802 {
 803         struct xsk_umem_info *umem;
 804         const struct rte_memzone *mz;
 805         struct xsk_umem_config usr_config = {
 806                 .fill_size = ETH_AF_XDP_DFLT_NUM_DESCS,
 807                 .comp_size = ETH_AF_XDP_DFLT_NUM_DESCS,
 808                 .frame_size = ETH_AF_XDP_FRAME_SIZE,
 809                 .frame_headroom = 0 };
 810         char ring_name[RTE_RING_NAMESIZE];
 811         char mz_name[RTE_MEMZONE_NAMESIZE];
 812         int ret;
 813         uint64_t i;
 814
 815         umem = rte_zmalloc_socket("umem", sizeof(*umem), 0, rte_socket_id());
 816         if (umem == NULL) {
 817                 AF_XDP_LOG(ERR, "Failed to allocate umem info");
 818                 return NULL;
 819         }
 820
 821         snprintf(ring_name, sizeof(ring_name), "af_xdp_ring_%s_%u",
 822                        internals->if_name, rxq->xsk_queue_idx);
 823         umem->buf_ring = rte_ring_create(ring_name,
 824                                          ETH_AF_XDP_NUM_BUFFERS,
 825                                          rte_socket_id(),
 826                                          0x0);
 827         if (umem->buf_ring == NULL) {
 828                 AF_XDP_LOG(ERR, "Failed to create rte_ring\n");
 829                 goto err;
 830         }
 831
 832         for (i = 0; i < ETH_AF_XDP_NUM_BUFFERS; i++)
 833                 rte_ring_enqueue(umem->buf_ring,
 834                                  (void *)(i * ETH_AF_XDP_FRAME_SIZE));
 835
 836         snprintf(mz_name, sizeof(mz_name), "af_xdp_umem_%s_%u",
 837                        internals->if_name, rxq->xsk_queue_idx);
 838         mz = rte_memzone_reserve_aligned(mz_name,
 839                         ETH_AF_XDP_NUM_BUFFERS * ETH_AF_XDP_FRAME_SIZE,
 840                         rte_socket_id(), RTE_MEMZONE_IOVA_CONTIG,
 841                         getpagesize());
 842         if (mz == NULL) {
 843                 AF_XDP_LOG(ERR, "Failed to reserve memzone for af_xdp umem.\n");
 844                 goto err;
 845         }
 846
 847         ret = xsk_umem__create(&umem->umem, mz->addr,
 848                                ETH_AF_XDP_NUM_BUFFERS * ETH_AF_XDP_FRAME_SIZE,
 849                                &umem->fq, &umem->cq,
 850                                &usr_config);
 851
 852         if (ret) {
 853                 AF_XDP_LOG(ERR, "Failed to create umem");
 854                 goto err;
 855         }
 856         umem->mz = mz;
 857
 858 #endif
 859         return umem;
 860
 861 err:
 862         xdp_umem_destroy(umem);
 863         return NULL;
 864 }
 865
 866 static int
 867 xsk_configure(struct pmd_internals *internals, struct pkt_rx_queue *rxq,
 868               int ring_size)
 869 {
 870         struct xsk_socket_config cfg;
 871         struct pkt_tx_queue *txq = rxq->pair;
 872         int ret = 0;
 873         int reserve_size = ETH_AF_XDP_DFLT_NUM_DESCS;
 874         struct rte_mbuf *fq_bufs[reserve_size];
 875
 876         rxq->umem = xdp_umem_configure(internals, rxq);
 877         if (rxq->umem == NULL)
 878                 return -ENOMEM;
 879         txq->umem = rxq->umem;
 880
 881         cfg.rx_size = ring_size;
 882         cfg.tx_size = ring_size;
 883         cfg.libbpf_flags = 0;
 884         cfg.xdp_flags = XDP_FLAGS_UPDATE_IF_NOEXIST;
 885         cfg.bind_flags = 0;
 886
 887 #if defined(XDP_USE_NEED_WAKEUP)
 888         cfg.bind_flags |= XDP_USE_NEED_WAKEUP;
 889 #endif
 890
 891         ret = xsk_socket__create(&rxq->xsk, internals->if_name,
 892                         rxq->xsk_queue_idx, rxq->umem->umem, &rxq->rx,
 893                         &txq->tx, &cfg);
 894         if (ret) {
 895                 AF_XDP_LOG(ERR, "Failed to create xsk socket.\n");
 896                 goto err;
 897         }
 898
 899 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 900         if (rte_pktmbuf_alloc_bulk(rxq->umem->mb_pool, fq_bufs, reserve_size)) {
 901                 AF_XDP_LOG(DEBUG, "Failed to get enough buffers for fq.\n");
 902                 goto err;
 903         }
 904 #endif
 905         ret = reserve_fill_queue(rxq->umem, reserve_size, fq_bufs);
 906         if (ret) {
 907                 xsk_socket__delete(rxq->xsk);
 908                 AF_XDP_LOG(ERR, "Failed to reserve fill queue.\n");
 909                 goto err;
 910         }
 911
 912         return 0;
 913
 914 err:
 915         xdp_umem_destroy(rxq->umem);
 916
 917         return ret;
 918 }
 919
 920 static int
 921 eth_rx_queue_setup(struct rte_eth_dev *dev,
 922                    uint16_t rx_queue_id,
 923                    uint16_t nb_rx_desc,
 924                    unsigned int socket_id __rte_unused,
 925                    const struct rte_eth_rxconf *rx_conf __rte_unused,
 926                    struct rte_mempool *mb_pool)
 927 {
 928         struct pmd_internals *internals = dev->data->dev_private;
 929         struct pkt_rx_queue *rxq;
 930         int ret;
 931
 932         rxq = &internals->rx_queues[rx_queue_id];
 933
 934         AF_XDP_LOG(INFO, "Set up rx queue, rx queue id: %d, xsk queue id: %d\n",
 935                    rx_queue_id, rxq->xsk_queue_idx);
 936
 937 #ifndef XDP_UMEM_UNALIGNED_CHUNK_FLAG
 938         uint32_t buf_size, data_size;
 939
 940         /* Now get the space available for data in the mbuf */
 941         buf_size = rte_pktmbuf_data_room_size(mb_pool) -
 942                 RTE_PKTMBUF_HEADROOM;
 943         data_size = ETH_AF_XDP_FRAME_SIZE;
 944
 945         if (data_size > buf_size) {
 946                 AF_XDP_LOG(ERR, "%s: %d bytes will not fit in mbuf (%d bytes)\n",
 947                         dev->device->name, data_size, buf_size);
 948                 ret = -ENOMEM;
 949                 goto err;
 950         }
 951 #endif
 952
 953         rxq->mb_pool = mb_pool;
 954
 955         if (xsk_configure(internals, rxq, nb_rx_desc)) {
 956                 AF_XDP_LOG(ERR, "Failed to configure xdp socket\n");
 957                 ret = -EINVAL;
 958                 goto err;
 959         }
 960
 961         rxq->fds[0].fd = xsk_socket__fd(rxq->xsk);
 962         rxq->fds[0].events = POLLIN;
 963
 964         dev->data->rx_queues[rx_queue_id] = rxq;
 965         return 0;
 966
 967 err:
 968         return ret;
 969 }
 970
 971 static int
 972 eth_tx_queue_setup(struct rte_eth_dev *dev,
 973                    uint16_t tx_queue_id,
 974                    uint16_t nb_tx_desc __rte_unused,
 975                    unsigned int socket_id __rte_unused,
 976                    const struct rte_eth_txconf *tx_conf __rte_unused)
 977 {
 978         struct pmd_internals *internals = dev->data->dev_private;
 979         struct pkt_tx_queue *txq;
 980
 981         txq = &internals->tx_queues[tx_queue_id];
 982
 983         dev->data->tx_queues[tx_queue_id] = txq;
 984         return 0;
 985 }
 986
 987 static int
 988 eth_dev_mtu_set(struct rte_eth_dev *dev, uint16_t mtu)
 989 {
 990         struct pmd_internals *internals = dev->data->dev_private;
 991         struct ifreq ifr = { .ifr_mtu = mtu };
 992         int ret;
 993         int s;
 994
 995         s = socket(PF_INET, SOCK_DGRAM, 0);
 996         if (s < 0)
 997                 return -EINVAL;
 998
 999         strlcpy(ifr.ifr_name, internals->if_name, IFNAMSIZ);
1000         ret = ioctl(s, SIOCSIFMTU, &ifr);
1001         close(s);
1002
1003         return (ret < 0) ? -errno : 0;
1004 }
1005
1006 static int
1007 eth_dev_change_flags(char *if_name, uint32_t flags, uint32_t mask)
1008 {
1009         struct ifreq ifr;
1010         int ret = 0;
1011         int s;
1012
1013         s = socket(PF_INET, SOCK_DGRAM, 0);
1014         if (s < 0)
1015                 return -errno;
1016
1017         strlcpy(ifr.ifr_name, if_name, IFNAMSIZ);
1018         if (ioctl(s, SIOCGIFFLAGS, &ifr) < 0) {
1019                 ret = -errno;
1020                 goto out;
1021         }
1022         ifr.ifr_flags &= mask;
1023         ifr.ifr_flags |= flags;
1024         if (ioctl(s, SIOCSIFFLAGS, &ifr) < 0) {
1025                 ret = -errno;
1026                 goto out;
1027         }
1028 out:
1029         close(s);
1030         return ret;
1031 }
1032
1033 static int
1034 eth_dev_promiscuous_enable(struct rte_eth_dev *dev)
1035 {
1036         struct pmd_internals *internals = dev->data->dev_private;
1037
1038         return eth_dev_change_flags(internals->if_name, IFF_PROMISC, ~0);
1039 }
1040
1041 static int
1042 eth_dev_promiscuous_disable(struct rte_eth_dev *dev)
1043 {
1044         struct pmd_internals *internals = dev->data->dev_private;
1045
1046         return eth_dev_change_flags(internals->if_name, 0, ~IFF_PROMISC);
1047 }
1048
1049 static const struct eth_dev_ops ops = {
1050         .dev_start = eth_dev_start,
1051         .dev_stop = eth_dev_stop,
1052         .dev_close = eth_dev_close,
1053         .dev_configure = eth_dev_configure,
1054         .dev_infos_get = eth_dev_info,
1055         .mtu_set = eth_dev_mtu_set,
1056         .promiscuous_enable = eth_dev_promiscuous_enable,
1057         .promiscuous_disable = eth_dev_promiscuous_disable,
1058         .rx_queue_setup = eth_rx_queue_setup,
1059         .tx_queue_setup = eth_tx_queue_setup,
1060         .rx_queue_release = eth_queue_release,
1061         .tx_queue_release = eth_queue_release,
1062         .link_update = eth_link_update,
1063         .stats_get = eth_stats_get,
1064         .stats_reset = eth_stats_reset,
1065 };
1066
1067 /** parse integer from integer argument */
1068 static int
1069 parse_integer_arg(const char *key __rte_unused,
1070                   const char *value, void *extra_args)
1071 {
1072         int *i = (int *)extra_args;
1073         char *end;
1074
1075         *i = strtol(value, &end, 10);
1076         if (*i < 0) {
1077                 AF_XDP_LOG(ERR, "Argument has to be positive.\n");
1078                 return -EINVAL;
1079         }
1080
1081         return 0;
1082 }
1083
1084 /** parse name argument */
1085 static int
1086 parse_name_arg(const char *key __rte_unused,
1087                const char *value, void *extra_args)
1088 {
1089         char *name = extra_args;
1090
1091         if (strnlen(value, IFNAMSIZ) > IFNAMSIZ - 1) {
1092                 AF_XDP_LOG(ERR, "Invalid name %s, should be less than %u bytes.\n",
1093                            value, IFNAMSIZ);
1094                 return -EINVAL;
1095         }
1096
1097         strlcpy(name, value, IFNAMSIZ);
1098
1099         return 0;
1100 }
1101
1102 static int
1103 xdp_get_channels_info(const char *if_name, int *max_queues,
1104                                 int *combined_queues)
1105 {
1106         struct ethtool_channels channels;
1107         struct ifreq ifr;
1108         int fd, ret;
1109
1110         fd = socket(AF_INET, SOCK_DGRAM, 0);
1111         if (fd < 0)
1112                 return -1;
1113
1114         channels.cmd = ETHTOOL_GCHANNELS;
1115         ifr.ifr_data = (void *)&channels;
1116         strncpy(ifr.ifr_name, if_name, IFNAMSIZ);
1117         ret = ioctl(fd, SIOCETHTOOL, &ifr);
1118         if (ret) {
1119                 if (errno == EOPNOTSUPP) {
1120                         ret = 0;
1121                 } else {
1122                         ret = -errno;
1123                         goto out;
1124                 }
1125         }
1126
1127         if (channels.max_combined == 0 || errno == EOPNOTSUPP) {
1128                 /* If the device says it has no channels, then all traffic
1129                  * is sent to a single stream, so max queues = 1.
1130                  */
1131                 *max_queues = 1;
1132                 *combined_queues = 1;
1133         } else {
1134                 *max_queues = channels.max_combined;
1135                 *combined_queues = channels.combined_count;
1136         }
1137
1138  out:
1139         close(fd);
1140         return ret;
1141 }
1142
1143 static int
1144 parse_parameters(struct rte_kvargs *kvlist, char *if_name, int *start_queue,
1145                         int *queue_cnt)
1146 {
1147         int ret;
1148
1149         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_IFACE_ARG,
1150                                  &parse_name_arg, if_name);
1151         if (ret < 0)
1152                 goto free_kvlist;
1153
1154         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_START_QUEUE_ARG,
1155                                  &parse_integer_arg, start_queue);
1156         if (ret < 0)
1157                 goto free_kvlist;
1158
1159         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_QUEUE_COUNT_ARG,
1160                                  &parse_integer_arg, queue_cnt);
1161         if (ret < 0 || *queue_cnt <= 0) {
1162                 ret = -EINVAL;
1163                 goto free_kvlist;
1164         }
1165
1166 free_kvlist:
1167         rte_kvargs_free(kvlist);
1168         return ret;
1169 }
1170
1171 static int
1172 get_iface_info(const char *if_name,
1173                struct rte_ether_addr *eth_addr,
1174                int *if_index)
1175 {
1176         struct ifreq ifr;
1177         int sock = socket(AF_INET, SOCK_DGRAM, IPPROTO_IP);
1178
1179         if (sock < 0)
1180                 return -1;
1181
1182         strlcpy(ifr.ifr_name, if_name, IFNAMSIZ);
1183         if (ioctl(sock, SIOCGIFINDEX, &ifr))
1184                 goto error;
1185
1186         *if_index = ifr.ifr_ifindex;
1187
1188         if (ioctl(sock, SIOCGIFHWADDR, &ifr))
1189                 goto error;
1190
1191         rte_memcpy(eth_addr, ifr.ifr_hwaddr.sa_data, RTE_ETHER_ADDR_LEN);
1192
1193         close(sock);
1194         return 0;
1195
1196 error:
1197         close(sock);
1198         return -1;
1199 }
1200
1201 static struct rte_eth_dev *
1202 init_internals(struct rte_vdev_device *dev, const char *if_name,
1203                         int start_queue_idx, int queue_cnt)
1204 {
1205         const char *name = rte_vdev_device_name(dev);
1206         const unsigned int numa_node = dev->device.numa_node;
1207         struct pmd_internals *internals;
1208         struct rte_eth_dev *eth_dev;
1209         int ret;
1210         int i;
1211
1212         internals = rte_zmalloc_socket(name, sizeof(*internals), 0, numa_node);
1213         if (internals == NULL)
1214                 return NULL;
1215
1216         internals->start_queue_idx = start_queue_idx;
1217         internals->queue_cnt = queue_cnt;
1218         strlcpy(internals->if_name, if_name, IFNAMSIZ);
1219
1220         if (xdp_get_channels_info(if_name, &internals->max_queue_cnt,
1221                                   &internals->combined_queue_cnt)) {
1222                 AF_XDP_LOG(ERR, "Failed to get channel info of interface: %s\n",
1223                                 if_name);
1224                 goto err_free_internals;
1225         }
1226
1227         if (queue_cnt > internals->combined_queue_cnt) {
1228                 AF_XDP_LOG(ERR, "Specified queue count %d is larger than combined queue count %d.\n",
1229                                 queue_cnt, internals->combined_queue_cnt);
1230                 goto err_free_internals;
1231         }
1232
1233         internals->rx_queues = rte_zmalloc_socket(NULL,
1234                                         sizeof(struct pkt_rx_queue) * queue_cnt,
1235                                         0, numa_node);
1236         if (internals->rx_queues == NULL) {
1237                 AF_XDP_LOG(ERR, "Failed to allocate memory for rx queues.\n");
1238                 goto err_free_internals;
1239         }
1240
1241         internals->tx_queues = rte_zmalloc_socket(NULL,
1242                                         sizeof(struct pkt_tx_queue) * queue_cnt,
1243                                         0, numa_node);
1244         if (internals->tx_queues == NULL) {
1245                 AF_XDP_LOG(ERR, "Failed to allocate memory for tx queues.\n");
1246                 goto err_free_rx;
1247         }
1248         for (i = 0; i < queue_cnt; i++) {
1249                 internals->tx_queues[i].pair = &internals->rx_queues[i];
1250                 internals->rx_queues[i].pair = &internals->tx_queues[i];
1251                 internals->rx_queues[i].xsk_queue_idx = start_queue_idx + i;
1252                 internals->tx_queues[i].xsk_queue_idx = start_queue_idx + i;
1253         }
1254
1255         ret = get_iface_info(if_name, &internals->eth_addr,
1256                              &internals->if_index);
1257         if (ret)
1258                 goto err_free_tx;
1259
1260         eth_dev = rte_eth_vdev_allocate(dev, 0);
1261         if (eth_dev == NULL)
1262                 goto err_free_tx;
1263
1264         eth_dev->data->dev_private = internals;
1265         eth_dev->data->dev_link = pmd_link;
1266         eth_dev->data->mac_addrs = &internals->eth_addr;
1267         eth_dev->dev_ops = &ops;
1268         eth_dev->rx_pkt_burst = eth_af_xdp_rx;
1269         eth_dev->tx_pkt_burst = eth_af_xdp_tx;
1270         /* Let rte_eth_dev_close() release the port resources. */
1271         eth_dev->data->dev_flags |= RTE_ETH_DEV_CLOSE_REMOVE;
1272
1273 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
1274         AF_XDP_LOG(INFO, "Zero copy between umem and mbuf enabled.\n");
1275 #endif
1276
1277         return eth_dev;
1278
1279 err_free_tx:
1280         rte_free(internals->tx_queues);
1281 err_free_rx:
1282         rte_free(internals->rx_queues);
1283 err_free_internals:
1284         rte_free(internals);
1285         return NULL;
1286 }
1287
1288 static int
1289 rte_pmd_af_xdp_probe(struct rte_vdev_device *dev)
1290 {
1291         struct rte_kvargs *kvlist;
1292         char if_name[IFNAMSIZ] = {'\0'};
1293         int xsk_start_queue_idx = ETH_AF_XDP_DFLT_START_QUEUE_IDX;
1294         int xsk_queue_cnt = ETH_AF_XDP_DFLT_QUEUE_COUNT;
1295         struct rte_eth_dev *eth_dev = NULL;
1296         const char *name;
1297
1298         AF_XDP_LOG(INFO, "Initializing pmd_af_xdp for %s\n",
1299                 rte_vdev_device_name(dev));
1300
1301         name = rte_vdev_device_name(dev);
1302         if (rte_eal_process_type() == RTE_PROC_SECONDARY &&
1303                 strlen(rte_vdev_device_args(dev)) == 0) {
1304                 eth_dev = rte_eth_dev_attach_secondary(name);
1305                 if (eth_dev == NULL) {
1306                         AF_XDP_LOG(ERR, "Failed to probe %s\n", name);
1307                         return -EINVAL;
1308                 }
1309                 eth_dev->dev_ops = &ops;
1310                 rte_eth_dev_probing_finish(eth_dev);
1311                 return 0;
1312         }
1313
1314         kvlist = rte_kvargs_parse(rte_vdev_device_args(dev), valid_arguments);
1315         if (kvlist == NULL) {
1316                 AF_XDP_LOG(ERR, "Invalid kvargs key\n");
1317                 return -EINVAL;
1318         }
1319
1320         if (dev->device.numa_node == SOCKET_ID_ANY)
1321                 dev->device.numa_node = rte_socket_id();
1322
1323         if (parse_parameters(kvlist, if_name, &xsk_start_queue_idx,
1324                              &xsk_queue_cnt) < 0) {
1325                 AF_XDP_LOG(ERR, "Invalid kvargs value\n");
1326                 return -EINVAL;
1327         }
1328
1329         if (strlen(if_name) == 0) {
1330                 AF_XDP_LOG(ERR, "Network interface must be specified\n");
1331                 return -EINVAL;
1332         }
1333
1334         eth_dev = init_internals(dev, if_name, xsk_start_queue_idx,
1335                                         xsk_queue_cnt);
1336         if (eth_dev == NULL) {
1337                 AF_XDP_LOG(ERR, "Failed to init internals\n");
1338                 return -1;
1339         }
1340
1341         rte_eth_dev_probing_finish(eth_dev);
1342
1343         return 0;
1344 }
1345
1346 static int
1347 rte_pmd_af_xdp_remove(struct rte_vdev_device *dev)
1348 {
1349         struct rte_eth_dev *eth_dev = NULL;
1350
1351         AF_XDP_LOG(INFO, "Removing AF_XDP ethdev on numa socket %u\n",
1352                 rte_socket_id());
1353
1354         if (dev == NULL)
1355                 return -1;
1356
1357         /* find the ethdev entry */
1358         eth_dev = rte_eth_dev_allocated(rte_vdev_device_name(dev));
1359         if (eth_dev == NULL)
1360                 return 0;
1361
1362         eth_dev_close(eth_dev);
1363         rte_eth_dev_release_port(eth_dev);
1364
1365
1366         return 0;
1367 }
1368
1369 static struct rte_vdev_driver pmd_af_xdp_drv = {
1370         .probe = rte_pmd_af_xdp_probe,
1371         .remove = rte_pmd_af_xdp_remove,
1372 };
1373
1374 RTE_PMD_REGISTER_VDEV(net_af_xdp, pmd_af_xdp_drv);
1375 RTE_PMD_REGISTER_PARAM_STRING(net_af_xdp,
1376                               "iface=<string> "
1377                               "start_queue=<int> "
1378                               "queue_count=<int> ");