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_config.h>
  30 #include <rte_dev.h>
  31 #include <rte_eal.h>
  32 #include <rte_ether.h>
  33 #include <rte_lcore.h>
  34 #include <rte_log.h>
  35 #include <rte_memory.h>
  36 #include <rte_memzone.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 #ifdef XDP_UMEM_UNALIGNED_CHUNK_FLAG
  62 #define ETH_AF_XDP_MBUF_OVERHEAD        128 /* sizeof(struct rte_mbuf) */
  63 #define ETH_AF_XDP_DATA_HEADROOM \
  64         (ETH_AF_XDP_MBUF_OVERHEAD + RTE_PKTMBUF_HEADROOM)
  65 #else
  66 #define ETH_AF_XDP_DATA_HEADROOM        0
  67 #endif
  68 #define ETH_AF_XDP_DFLT_NUM_DESCS       XSK_RING_CONS__DEFAULT_NUM_DESCS
  69 #define ETH_AF_XDP_DFLT_START_QUEUE_IDX 0
  70 #define ETH_AF_XDP_DFLT_QUEUE_COUNT     1
  71
  72 #define ETH_AF_XDP_RX_BATCH_SIZE        32
  73 #define ETH_AF_XDP_TX_BATCH_SIZE        32
  74
  75
  76 struct xsk_umem_info {
  77         struct xsk_ring_prod fq;
  78         struct xsk_ring_cons cq;
  79         struct xsk_umem *umem;
  80         struct rte_ring *buf_ring;
  81         const struct rte_memzone *mz;
  82         struct rte_mempool *mb_pool;
  83         void *buffer;
  84 };
  85
  86 struct rx_stats {
  87         uint64_t rx_pkts;
  88         uint64_t rx_bytes;
  89         uint64_t rx_dropped;
  90 };
  91
  92 struct pkt_rx_queue {
  93         struct xsk_ring_cons rx;
  94         struct xsk_umem_info *umem;
  95         struct xsk_socket *xsk;
  96         struct rte_mempool *mb_pool;
  97
  98         struct rx_stats stats;
  99
 100         struct pkt_tx_queue *pair;
 101         struct pollfd fds[1];
 102         int xsk_queue_idx;
 103 };
 104
 105 struct tx_stats {
 106         uint64_t tx_pkts;
 107         uint64_t tx_bytes;
 108         uint64_t tx_dropped;
 109 };
 110
 111 struct pkt_tx_queue {
 112         struct xsk_ring_prod tx;
 113         struct xsk_umem_info *umem;
 114
 115         struct tx_stats stats;
 116
 117         struct pkt_rx_queue *pair;
 118         int xsk_queue_idx;
 119 };
 120
 121 struct pmd_internals {
 122         int if_index;
 123         char if_name[IFNAMSIZ];
 124         int start_queue_idx;
 125         int queue_cnt;
 126         int max_queue_cnt;
 127         int combined_queue_cnt;
 128
 129         struct rte_ether_addr eth_addr;
 130
 131         struct pkt_rx_queue *rx_queues;
 132         struct pkt_tx_queue *tx_queues;
 133 };
 134
 135 #define ETH_AF_XDP_IFACE_ARG                    "iface"
 136 #define ETH_AF_XDP_START_QUEUE_ARG              "start_queue"
 137 #define ETH_AF_XDP_QUEUE_COUNT_ARG              "queue_count"
 138
 139 static const char * const valid_arguments[] = {
 140         ETH_AF_XDP_IFACE_ARG,
 141         ETH_AF_XDP_START_QUEUE_ARG,
 142         ETH_AF_XDP_QUEUE_COUNT_ARG,
 143         NULL
 144 };
 145
 146 static const struct rte_eth_link pmd_link = {
 147         .link_speed = ETH_SPEED_NUM_10G,
 148         .link_duplex = ETH_LINK_FULL_DUPLEX,
 149         .link_status = ETH_LINK_DOWN,
 150         .link_autoneg = ETH_LINK_AUTONEG
 151 };
 152
 153 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 154 static inline int
 155 reserve_fill_queue_zc(struct xsk_umem_info *umem, uint16_t reserve_size,
 156                       struct rte_mbuf **bufs)
 157 {
 158         struct xsk_ring_prod *fq = &umem->fq;
 159         uint32_t idx;
 160         uint16_t i;
 161
 162         if (unlikely(!xsk_ring_prod__reserve(fq, reserve_size, &idx))) {
 163                 for (i = 0; i < reserve_size; i++)
 164                         rte_pktmbuf_free(bufs[i]);
 165                 AF_XDP_LOG(DEBUG, "Failed to reserve enough fq descs.\n");
 166                 return -1;
 167         }
 168
 169         for (i = 0; i < reserve_size; i++) {
 170                 __u64 *fq_addr;
 171                 uint64_t addr;
 172
 173                 fq_addr = xsk_ring_prod__fill_addr(fq, idx++);
 174                 addr = (uint64_t)bufs[i] - (uint64_t)umem->buffer;
 175                 *fq_addr = addr;
 176         }
 177
 178         xsk_ring_prod__submit(fq, reserve_size);
 179
 180         return 0;
 181 }
 182 #else
 183 static inline int
 184 reserve_fill_queue_cp(struct xsk_umem_info *umem, uint16_t reserve_size,
 185                       struct rte_mbuf **bufs __rte_unused)
 186 {
 187         struct xsk_ring_prod *fq = &umem->fq;
 188         void *addrs[reserve_size];
 189         uint32_t idx;
 190         uint16_t i;
 191
 192         if (rte_ring_dequeue_bulk(umem->buf_ring, addrs, reserve_size, NULL)
 193                     != reserve_size) {
 194                 AF_XDP_LOG(DEBUG, "Failed to get enough buffers for fq.\n");
 195                 return -1;
 196         }
 197
 198         if (unlikely(!xsk_ring_prod__reserve(fq, reserve_size, &idx))) {
 199                 AF_XDP_LOG(DEBUG, "Failed to reserve enough fq descs.\n");
 200                 rte_ring_enqueue_bulk(umem->buf_ring, addrs,
 201                                 reserve_size, NULL);
 202                 return -1;
 203         }
 204
 205         for (i = 0; i < reserve_size; i++) {
 206                 __u64 *fq_addr;
 207
 208                 fq_addr = xsk_ring_prod__fill_addr(fq, idx++);
 209                 *fq_addr = (uint64_t)addrs[i];
 210         }
 211
 212         xsk_ring_prod__submit(fq, reserve_size);
 213
 214         return 0;
 215 }
 216 #endif
 217
 218 static inline int
 219 reserve_fill_queue(struct xsk_umem_info *umem, uint16_t reserve_size,
 220                    struct rte_mbuf **bufs)
 221 {
 222 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 223         return reserve_fill_queue_zc(umem, reserve_size, bufs);
 224 #else
 225         return reserve_fill_queue_cp(umem, reserve_size, bufs);
 226 #endif
 227 }
 228
 229 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 230 static uint16_t
 231 af_xdp_rx_zc(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 232 {
 233         struct pkt_rx_queue *rxq = queue;
 234         struct xsk_ring_cons *rx = &rxq->rx;
 235         struct xsk_umem_info *umem = rxq->umem;
 236         uint32_t idx_rx = 0;
 237         unsigned long rx_bytes = 0;
 238         int rcvd, i;
 239         struct rte_mbuf *fq_bufs[ETH_AF_XDP_RX_BATCH_SIZE];
 240
 241         /* allocate bufs for fill queue replenishment after rx */
 242         if (rte_pktmbuf_alloc_bulk(umem->mb_pool, fq_bufs, nb_pkts)) {
 243                 AF_XDP_LOG(DEBUG,
 244                         "Failed to get enough buffers for fq.\n");
 245                 return -1;
 246         }
 247
 248         rcvd = xsk_ring_cons__peek(rx, nb_pkts, &idx_rx);
 249
 250         if (rcvd == 0) {
 251 #if defined(XDP_USE_NEED_WAKEUP)
 252                 if (xsk_ring_prod__needs_wakeup(&umem->fq))
 253                         (void)poll(rxq->fds, 1, 1000);
 254 #endif
 255
 256                 goto out;
 257         }
 258
 259         for (i = 0; i < rcvd; i++) {
 260                 const struct xdp_desc *desc;
 261                 uint64_t addr;
 262                 uint32_t len;
 263                 uint64_t offset;
 264
 265                 desc = xsk_ring_cons__rx_desc(rx, idx_rx++);
 266                 addr = desc->addr;
 267                 len = desc->len;
 268
 269                 offset = xsk_umem__extract_offset(addr);
 270                 addr = xsk_umem__extract_addr(addr);
 271
 272                 bufs[i] = (struct rte_mbuf *)
 273                                 xsk_umem__get_data(umem->buffer, addr);
 274                 bufs[i]->data_off = offset - sizeof(struct rte_mbuf);
 275
 276                 rte_pktmbuf_pkt_len(bufs[i]) = len;
 277                 rte_pktmbuf_data_len(bufs[i]) = len;
 278                 rx_bytes += len;
 279         }
 280
 281         xsk_ring_cons__release(rx, rcvd);
 282
 283         (void)reserve_fill_queue(umem, rcvd, fq_bufs);
 284
 285         /* statistics */
 286         rxq->stats.rx_pkts += rcvd;
 287         rxq->stats.rx_bytes += rx_bytes;
 288
 289 out:
 290         if (rcvd != nb_pkts)
 291                 rte_mempool_put_bulk(umem->mb_pool, (void **)&fq_bufs[rcvd],
 292                                      nb_pkts - rcvd);
 293
 294         return rcvd;
 295 }
 296 #else
 297 static uint16_t
 298 af_xdp_rx_cp(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 299 {
 300         struct pkt_rx_queue *rxq = queue;
 301         struct xsk_ring_cons *rx = &rxq->rx;
 302         struct xsk_umem_info *umem = rxq->umem;
 303         struct xsk_ring_prod *fq = &umem->fq;
 304         uint32_t idx_rx = 0;
 305         unsigned long rx_bytes = 0;
 306         int rcvd, i;
 307         uint32_t free_thresh = fq->size >> 1;
 308         struct rte_mbuf *mbufs[ETH_AF_XDP_RX_BATCH_SIZE];
 309
 310         if (unlikely(rte_pktmbuf_alloc_bulk(rxq->mb_pool, mbufs, nb_pkts) != 0))
 311                 return 0;
 312
 313         rcvd = xsk_ring_cons__peek(rx, nb_pkts, &idx_rx);
 314         if (rcvd == 0) {
 315 #if defined(XDP_USE_NEED_WAKEUP)
 316                 if (xsk_ring_prod__needs_wakeup(fq))
 317                         (void)poll(rxq->fds, 1, 1000);
 318 #endif
 319
 320                 goto out;
 321         }
 322
 323         if (xsk_prod_nb_free(fq, free_thresh) >= free_thresh)
 324                 (void)reserve_fill_queue(umem, ETH_AF_XDP_RX_BATCH_SIZE, NULL);
 325
 326         for (i = 0; i < rcvd; i++) {
 327                 const struct xdp_desc *desc;
 328                 uint64_t addr;
 329                 uint32_t len;
 330                 void *pkt;
 331
 332                 desc = xsk_ring_cons__rx_desc(rx, idx_rx++);
 333                 addr = desc->addr;
 334                 len = desc->len;
 335                 pkt = xsk_umem__get_data(rxq->umem->mz->addr, addr);
 336
 337                 rte_memcpy(rte_pktmbuf_mtod(mbufs[i], void *), pkt, len);
 338                 rte_ring_enqueue(umem->buf_ring, (void *)addr);
 339                 rte_pktmbuf_pkt_len(mbufs[i]) = len;
 340                 rte_pktmbuf_data_len(mbufs[i]) = len;
 341                 rx_bytes += len;
 342                 bufs[i] = mbufs[i];
 343         }
 344
 345         xsk_ring_cons__release(rx, rcvd);
 346
 347         /* statistics */
 348         rxq->stats.rx_pkts += rcvd;
 349         rxq->stats.rx_bytes += rx_bytes;
 350
 351 out:
 352         if (rcvd != nb_pkts)
 353                 rte_mempool_put_bulk(rxq->mb_pool, (void **)&mbufs[rcvd],
 354                                      nb_pkts - rcvd);
 355
 356         return rcvd;
 357 }
 358 #endif
 359
 360 static uint16_t
 361 eth_af_xdp_rx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 362 {
 363         nb_pkts = RTE_MIN(nb_pkts, ETH_AF_XDP_RX_BATCH_SIZE);
 364
 365 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 366         return af_xdp_rx_zc(queue, bufs, nb_pkts);
 367 #else
 368         return af_xdp_rx_cp(queue, bufs, nb_pkts);
 369 #endif
 370 }
 371
 372 static void
 373 pull_umem_cq(struct xsk_umem_info *umem, int size)
 374 {
 375         struct xsk_ring_cons *cq = &umem->cq;
 376         size_t i, n;
 377         uint32_t idx_cq = 0;
 378
 379         n = xsk_ring_cons__peek(cq, size, &idx_cq);
 380
 381         for (i = 0; i < n; i++) {
 382                 uint64_t addr;
 383                 addr = *xsk_ring_cons__comp_addr(cq, idx_cq++);
 384 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 385                 addr = xsk_umem__extract_addr(addr);
 386                 rte_pktmbuf_free((struct rte_mbuf *)
 387                                         xsk_umem__get_data(umem->buffer, addr));
 388 #else
 389                 rte_ring_enqueue(umem->buf_ring, (void *)addr);
 390 #endif
 391         }
 392
 393         xsk_ring_cons__release(cq, n);
 394 }
 395
 396 static void
 397 kick_tx(struct pkt_tx_queue *txq)
 398 {
 399         struct xsk_umem_info *umem = txq->umem;
 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, ETH_AF_XDP_TX_BATCH_SIZE);
 413                 }
 414 #ifndef XDP_UMEM_UNALIGNED_CHUNK_FLAG
 415         pull_umem_cq(umem, ETH_AF_XDP_TX_BATCH_SIZE);
 416 #endif
 417 }
 418
 419 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 420 static uint16_t
 421 af_xdp_tx_zc(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 422 {
 423         struct pkt_tx_queue *txq = queue;
 424         struct xsk_umem_info *umem = txq->umem;
 425         struct rte_mbuf *mbuf;
 426         unsigned long tx_bytes = 0;
 427         int i;
 428         uint32_t idx_tx;
 429         uint16_t count = 0;
 430         struct xdp_desc *desc;
 431         uint64_t addr, offset;
 432
 433         pull_umem_cq(umem, nb_pkts);
 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                                 goto out;
 442                         }
 443                         desc = xsk_ring_prod__tx_desc(&txq->tx, idx_tx);
 444                         desc->len = mbuf->pkt_len;
 445                         addr = (uint64_t)mbuf - (uint64_t)umem->buffer;
 446                         offset = rte_pktmbuf_mtod(mbuf, uint64_t) -
 447                                         (uint64_t)mbuf;
 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                         offset = rte_pktmbuf_mtod(local_mbuf, uint64_t) -
 470                                         (uint64_t)local_mbuf;
 471                         pkt = xsk_umem__get_data(umem->buffer, addr + offset);
 472                         offset = offset << XSK_UNALIGNED_BUF_OFFSET_SHIFT;
 473                         desc->addr = addr | offset;
 474                         rte_memcpy(pkt, rte_pktmbuf_mtod(mbuf, void *),
 475                                         desc->len);
 476                         rte_pktmbuf_free(mbuf);
 477                         count++;
 478                 }
 479
 480                 tx_bytes += mbuf->pkt_len;
 481         }
 482
 483 #if defined(XDP_USE_NEED_WAKEUP)
 484         if (xsk_ring_prod__needs_wakeup(&txq->tx))
 485 #endif
 486                 kick_tx(txq);
 487
 488 out:
 489         xsk_ring_prod__submit(&txq->tx, count);
 490
 491         txq->stats.tx_pkts += count;
 492         txq->stats.tx_bytes += tx_bytes;
 493         txq->stats.tx_dropped += nb_pkts - count;
 494
 495         return count;
 496 }
 497 #else
 498 static uint16_t
 499 af_xdp_tx_cp(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 500 {
 501         struct pkt_tx_queue *txq = queue;
 502         struct xsk_umem_info *umem = txq->umem;
 503         struct rte_mbuf *mbuf;
 504         void *addrs[ETH_AF_XDP_TX_BATCH_SIZE];
 505         unsigned long tx_bytes = 0;
 506         int i;
 507         uint32_t idx_tx;
 508
 509         nb_pkts = RTE_MIN(nb_pkts, ETH_AF_XDP_TX_BATCH_SIZE);
 510
 511         pull_umem_cq(umem, nb_pkts);
 512
 513         nb_pkts = rte_ring_dequeue_bulk(umem->buf_ring, addrs,
 514                                         nb_pkts, NULL);
 515         if (nb_pkts == 0)
 516                 return 0;
 517
 518         if (xsk_ring_prod__reserve(&txq->tx, nb_pkts, &idx_tx) != nb_pkts) {
 519                 kick_tx(txq);
 520                 rte_ring_enqueue_bulk(umem->buf_ring, addrs, nb_pkts, NULL);
 521                 return 0;
 522         }
 523
 524         for (i = 0; i < nb_pkts; i++) {
 525                 struct xdp_desc *desc;
 526                 void *pkt;
 527
 528                 desc = xsk_ring_prod__tx_desc(&txq->tx, idx_tx + i);
 529                 mbuf = bufs[i];
 530                 desc->len = mbuf->pkt_len;
 531
 532                 desc->addr = (uint64_t)addrs[i];
 533                 pkt = xsk_umem__get_data(umem->mz->addr,
 534                                          desc->addr);
 535                 rte_memcpy(pkt, rte_pktmbuf_mtod(mbuf, void *), desc->len);
 536                 tx_bytes += mbuf->pkt_len;
 537                 rte_pktmbuf_free(mbuf);
 538         }
 539
 540         xsk_ring_prod__submit(&txq->tx, nb_pkts);
 541
 542         kick_tx(txq);
 543
 544         txq->stats.tx_pkts += nb_pkts;
 545         txq->stats.tx_bytes += tx_bytes;
 546
 547         return nb_pkts;
 548 }
 549 #endif
 550
 551 static uint16_t
 552 eth_af_xdp_tx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 553 {
 554 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 555         return af_xdp_tx_zc(queue, bufs, nb_pkts);
 556 #else
 557         return af_xdp_tx_cp(queue, bufs, nb_pkts);
 558 #endif
 559 }
 560
 561 static int
 562 eth_dev_start(struct rte_eth_dev *dev)
 563 {
 564         dev->data->dev_link.link_status = ETH_LINK_UP;
 565
 566         return 0;
 567 }
 568
 569 /* This function gets called when the current port gets stopped. */
 570 static void
 571 eth_dev_stop(struct rte_eth_dev *dev)
 572 {
 573         dev->data->dev_link.link_status = ETH_LINK_DOWN;
 574 }
 575
 576 static int
 577 eth_dev_configure(struct rte_eth_dev *dev)
 578 {
 579         /* rx/tx must be paired */
 580         if (dev->data->nb_rx_queues != dev->data->nb_tx_queues)
 581                 return -EINVAL;
 582
 583         return 0;
 584 }
 585
 586 static int
 587 eth_dev_info(struct rte_eth_dev *dev, struct rte_eth_dev_info *dev_info)
 588 {
 589         struct pmd_internals *internals = dev->data->dev_private;
 590
 591         dev_info->if_index = internals->if_index;
 592         dev_info->max_mac_addrs = 1;
 593         dev_info->max_rx_pktlen = ETH_FRAME_LEN;
 594         dev_info->max_rx_queues = internals->queue_cnt;
 595         dev_info->max_tx_queues = internals->queue_cnt;
 596
 597         dev_info->min_mtu = RTE_ETHER_MIN_MTU;
 598         dev_info->max_mtu = ETH_AF_XDP_FRAME_SIZE - ETH_AF_XDP_DATA_HEADROOM;
 599
 600         dev_info->default_rxportconf.nb_queues = 1;
 601         dev_info->default_txportconf.nb_queues = 1;
 602         dev_info->default_rxportconf.ring_size = ETH_AF_XDP_DFLT_NUM_DESCS;
 603         dev_info->default_txportconf.ring_size = ETH_AF_XDP_DFLT_NUM_DESCS;
 604
 605         return 0;
 606 }
 607
 608 static int
 609 eth_stats_get(struct rte_eth_dev *dev, struct rte_eth_stats *stats)
 610 {
 611         struct pmd_internals *internals = dev->data->dev_private;
 612         struct xdp_statistics xdp_stats;
 613         struct pkt_rx_queue *rxq;
 614         struct pkt_tx_queue *txq;
 615         socklen_t optlen;
 616         int i, ret;
 617
 618         for (i = 0; i < dev->data->nb_rx_queues; i++) {
 619                 optlen = sizeof(struct xdp_statistics);
 620                 rxq = &internals->rx_queues[i];
 621                 txq = rxq->pair;
 622                 stats->q_ipackets[i] = rxq->stats.rx_pkts;
 623                 stats->q_ibytes[i] = rxq->stats.rx_bytes;
 624
 625                 stats->q_opackets[i] = txq->stats.tx_pkts;
 626                 stats->q_obytes[i] = txq->stats.tx_bytes;
 627
 628                 stats->ipackets += stats->q_ipackets[i];
 629                 stats->ibytes += stats->q_ibytes[i];
 630                 stats->imissed += rxq->stats.rx_dropped;
 631                 stats->oerrors += txq->stats.tx_dropped;
 632                 ret = getsockopt(xsk_socket__fd(rxq->xsk), SOL_XDP,
 633                                 XDP_STATISTICS, &xdp_stats, &optlen);
 634                 if (ret != 0) {
 635                         AF_XDP_LOG(ERR, "getsockopt() failed for XDP_STATISTICS.\n");
 636                         return -1;
 637                 }
 638                 stats->imissed += xdp_stats.rx_dropped;
 639
 640                 stats->opackets += stats->q_opackets[i];
 641                 stats->obytes += stats->q_obytes[i];
 642         }
 643
 644         return 0;
 645 }
 646
 647 static int
 648 eth_stats_reset(struct rte_eth_dev *dev)
 649 {
 650         struct pmd_internals *internals = dev->data->dev_private;
 651         int i;
 652
 653         for (i = 0; i < internals->queue_cnt; i++) {
 654                 memset(&internals->rx_queues[i].stats, 0,
 655                                         sizeof(struct rx_stats));
 656                 memset(&internals->tx_queues[i].stats, 0,
 657                                         sizeof(struct tx_stats));
 658         }
 659
 660         return 0;
 661 }
 662
 663 static void
 664 remove_xdp_program(struct pmd_internals *internals)
 665 {
 666         uint32_t curr_prog_id = 0;
 667
 668         if (bpf_get_link_xdp_id(internals->if_index, &curr_prog_id,
 669                                 XDP_FLAGS_UPDATE_IF_NOEXIST)) {
 670                 AF_XDP_LOG(ERR, "bpf_get_link_xdp_id failed\n");
 671                 return;
 672         }
 673         bpf_set_link_xdp_fd(internals->if_index, -1,
 674                         XDP_FLAGS_UPDATE_IF_NOEXIST);
 675 }
 676
 677 static void
 678 xdp_umem_destroy(struct xsk_umem_info *umem)
 679 {
 680 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 681         rte_mempool_free(umem->mb_pool);
 682         umem->mb_pool = NULL;
 683 #else
 684         rte_memzone_free(umem->mz);
 685         umem->mz = NULL;
 686
 687         rte_ring_free(umem->buf_ring);
 688         umem->buf_ring = NULL;
 689 #endif
 690
 691         rte_free(umem);
 692         umem = NULL;
 693 }
 694
 695 static void
 696 eth_dev_close(struct rte_eth_dev *dev)
 697 {
 698         struct pmd_internals *internals = dev->data->dev_private;
 699         struct pkt_rx_queue *rxq;
 700         int i;
 701
 702         AF_XDP_LOG(INFO, "Closing AF_XDP ethdev on numa socket %u\n",
 703                 rte_socket_id());
 704
 705         for (i = 0; i < internals->queue_cnt; i++) {
 706                 rxq = &internals->rx_queues[i];
 707                 if (rxq->umem == NULL)
 708                         break;
 709                 xsk_socket__delete(rxq->xsk);
 710                 (void)xsk_umem__delete(rxq->umem->umem);
 711                 xdp_umem_destroy(rxq->umem);
 712
 713                 /* free pkt_tx_queue */
 714                 rte_free(rxq->pair);
 715                 rte_free(rxq);
 716         }
 717
 718         /*
 719          * MAC is not allocated dynamically, setting it to NULL would prevent
 720          * from releasing it in rte_eth_dev_release_port.
 721          */
 722         dev->data->mac_addrs = NULL;
 723
 724         remove_xdp_program(internals);
 725 }
 726
 727 static void
 728 eth_queue_release(void *q __rte_unused)
 729 {
 730 }
 731
 732 static int
 733 eth_link_update(struct rte_eth_dev *dev __rte_unused,
 734                 int wait_to_complete __rte_unused)
 735 {
 736         return 0;
 737 }
 738
 739 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 740 static inline uint64_t get_base_addr(struct rte_mempool *mp)
 741 {
 742         struct rte_mempool_memhdr *memhdr;
 743
 744         memhdr = STAILQ_FIRST(&mp->mem_list);
 745         return (uint64_t)memhdr->addr & ~(getpagesize() - 1);
 746 }
 747
 748 static struct
 749 xsk_umem_info *xdp_umem_configure(struct pmd_internals *internals __rte_unused,
 750                                   struct pkt_rx_queue *rxq)
 751 {
 752         struct xsk_umem_info *umem;
 753         int ret;
 754         struct xsk_umem_config usr_config = {
 755                 .fill_size = ETH_AF_XDP_DFLT_NUM_DESCS,
 756                 .comp_size = ETH_AF_XDP_DFLT_NUM_DESCS,
 757                 .flags = XDP_UMEM_UNALIGNED_CHUNK_FLAG};
 758         void *base_addr = NULL;
 759         struct rte_mempool *mb_pool = rxq->mb_pool;
 760
 761         usr_config.frame_size = rte_pktmbuf_data_room_size(mb_pool) +
 762                                         ETH_AF_XDP_MBUF_OVERHEAD +
 763                                         mb_pool->private_data_size;
 764         usr_config.frame_headroom = ETH_AF_XDP_DATA_HEADROOM +
 765                                         mb_pool->private_data_size;
 766
 767         umem = rte_zmalloc_socket("umem", sizeof(*umem), 0, rte_socket_id());
 768         if (umem == NULL) {
 769                 AF_XDP_LOG(ERR, "Failed to allocate umem info");
 770                 return NULL;
 771         }
 772
 773         umem->mb_pool = mb_pool;
 774         base_addr = (void *)get_base_addr(mb_pool);
 775
 776         ret = xsk_umem__create(&umem->umem, base_addr,
 777                                mb_pool->populated_size * usr_config.frame_size,
 778                                &umem->fq, &umem->cq,
 779                                &usr_config);
 780
 781         if (ret) {
 782                 AF_XDP_LOG(ERR, "Failed to create umem");
 783                 goto err;
 784         }
 785         umem->buffer = base_addr;
 786
 787 #else
 788 static struct
 789 xsk_umem_info *xdp_umem_configure(struct pmd_internals *internals,
 790                                   struct pkt_rx_queue *rxq)
 791 {
 792         struct xsk_umem_info *umem;
 793         const struct rte_memzone *mz;
 794         struct xsk_umem_config usr_config = {
 795                 .fill_size = ETH_AF_XDP_DFLT_NUM_DESCS,
 796                 .comp_size = ETH_AF_XDP_DFLT_NUM_DESCS,
 797                 .frame_size = ETH_AF_XDP_FRAME_SIZE,
 798                 .frame_headroom = ETH_AF_XDP_DATA_HEADROOM };
 799         char ring_name[RTE_RING_NAMESIZE];
 800         char mz_name[RTE_MEMZONE_NAMESIZE];
 801         int ret;
 802         uint64_t i;
 803
 804         umem = rte_zmalloc_socket("umem", sizeof(*umem), 0, rte_socket_id());
 805         if (umem == NULL) {
 806                 AF_XDP_LOG(ERR, "Failed to allocate umem info");
 807                 return NULL;
 808         }
 809
 810         snprintf(ring_name, sizeof(ring_name), "af_xdp_ring_%s_%u",
 811                        internals->if_name, rxq->xsk_queue_idx);
 812         umem->buf_ring = rte_ring_create(ring_name,
 813                                          ETH_AF_XDP_NUM_BUFFERS,
 814                                          rte_socket_id(),
 815                                          0x0);
 816         if (umem->buf_ring == NULL) {
 817                 AF_XDP_LOG(ERR, "Failed to create rte_ring\n");
 818                 goto err;
 819         }
 820
 821         for (i = 0; i < ETH_AF_XDP_NUM_BUFFERS; i++)
 822                 rte_ring_enqueue(umem->buf_ring,
 823                                  (void *)(i * ETH_AF_XDP_FRAME_SIZE +
 824                                           ETH_AF_XDP_DATA_HEADROOM));
 825
 826         snprintf(mz_name, sizeof(mz_name), "af_xdp_umem_%s_%u",
 827                        internals->if_name, rxq->xsk_queue_idx);
 828         mz = rte_memzone_reserve_aligned(mz_name,
 829                         ETH_AF_XDP_NUM_BUFFERS * ETH_AF_XDP_FRAME_SIZE,
 830                         rte_socket_id(), RTE_MEMZONE_IOVA_CONTIG,
 831                         getpagesize());
 832         if (mz == NULL) {
 833                 AF_XDP_LOG(ERR, "Failed to reserve memzone for af_xdp umem.\n");
 834                 goto err;
 835         }
 836
 837         ret = xsk_umem__create(&umem->umem, mz->addr,
 838                                ETH_AF_XDP_NUM_BUFFERS * ETH_AF_XDP_FRAME_SIZE,
 839                                &umem->fq, &umem->cq,
 840                                &usr_config);
 841
 842         if (ret) {
 843                 AF_XDP_LOG(ERR, "Failed to create umem");
 844                 goto err;
 845         }
 846         umem->mz = mz;
 847
 848 #endif
 849         return umem;
 850
 851 err:
 852         xdp_umem_destroy(umem);
 853         return NULL;
 854 }
 855
 856 static int
 857 xsk_configure(struct pmd_internals *internals, struct pkt_rx_queue *rxq,
 858               int ring_size)
 859 {
 860         struct xsk_socket_config cfg;
 861         struct pkt_tx_queue *txq = rxq->pair;
 862         int ret = 0;
 863         int reserve_size = ETH_AF_XDP_DFLT_NUM_DESCS / 2;
 864         struct rte_mbuf *fq_bufs[reserve_size];
 865
 866         rxq->umem = xdp_umem_configure(internals, rxq);
 867         if (rxq->umem == NULL)
 868                 return -ENOMEM;
 869         txq->umem = rxq->umem;
 870
 871         cfg.rx_size = ring_size;
 872         cfg.tx_size = ring_size;
 873         cfg.libbpf_flags = 0;
 874         cfg.xdp_flags = XDP_FLAGS_UPDATE_IF_NOEXIST;
 875         cfg.bind_flags = 0;
 876
 877 #if defined(XDP_USE_NEED_WAKEUP)
 878         cfg.bind_flags |= XDP_USE_NEED_WAKEUP;
 879 #endif
 880
 881         ret = xsk_socket__create(&rxq->xsk, internals->if_name,
 882                         rxq->xsk_queue_idx, rxq->umem->umem, &rxq->rx,
 883                         &txq->tx, &cfg);
 884         if (ret) {
 885                 AF_XDP_LOG(ERR, "Failed to create xsk socket.\n");
 886                 goto err;
 887         }
 888
 889 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 890         if (rte_pktmbuf_alloc_bulk(rxq->umem->mb_pool, fq_bufs, reserve_size)) {
 891                 AF_XDP_LOG(DEBUG, "Failed to get enough buffers for fq.\n");
 892                 goto err;
 893         }
 894 #endif
 895         ret = reserve_fill_queue(rxq->umem, reserve_size, fq_bufs);
 896         if (ret) {
 897                 xsk_socket__delete(rxq->xsk);
 898                 AF_XDP_LOG(ERR, "Failed to reserve fill queue.\n");
 899                 goto err;
 900         }
 901
 902         return 0;
 903
 904 err:
 905         xdp_umem_destroy(rxq->umem);
 906
 907         return ret;
 908 }
 909
 910 static int
 911 eth_rx_queue_setup(struct rte_eth_dev *dev,
 912                    uint16_t rx_queue_id,
 913                    uint16_t nb_rx_desc,
 914                    unsigned int socket_id __rte_unused,
 915                    const struct rte_eth_rxconf *rx_conf __rte_unused,
 916                    struct rte_mempool *mb_pool)
 917 {
 918         struct pmd_internals *internals = dev->data->dev_private;
 919         struct pkt_rx_queue *rxq;
 920         int ret;
 921
 922         rxq = &internals->rx_queues[rx_queue_id];
 923
 924         AF_XDP_LOG(INFO, "Set up rx queue, rx queue id: %d, xsk queue id: %d\n",
 925                    rx_queue_id, rxq->xsk_queue_idx);
 926
 927 #ifndef XDP_UMEM_UNALIGNED_CHUNK_FLAG
 928         uint32_t buf_size, data_size;
 929
 930         /* Now get the space available for data in the mbuf */
 931         buf_size = rte_pktmbuf_data_room_size(mb_pool) -
 932                 RTE_PKTMBUF_HEADROOM;
 933         data_size = ETH_AF_XDP_FRAME_SIZE - ETH_AF_XDP_DATA_HEADROOM;
 934
 935         if (data_size > buf_size) {
 936                 AF_XDP_LOG(ERR, "%s: %d bytes will not fit in mbuf (%d bytes)\n",
 937                         dev->device->name, data_size, buf_size);
 938                 ret = -ENOMEM;
 939                 goto err;
 940         }
 941 #endif
 942
 943         rxq->mb_pool = mb_pool;
 944
 945         if (xsk_configure(internals, rxq, nb_rx_desc)) {
 946                 AF_XDP_LOG(ERR, "Failed to configure xdp socket\n");
 947                 ret = -EINVAL;
 948                 goto err;
 949         }
 950
 951         rxq->fds[0].fd = xsk_socket__fd(rxq->xsk);
 952         rxq->fds[0].events = POLLIN;
 953
 954         dev->data->rx_queues[rx_queue_id] = rxq;
 955         return 0;
 956
 957 err:
 958         return ret;
 959 }
 960
 961 static int
 962 eth_tx_queue_setup(struct rte_eth_dev *dev,
 963                    uint16_t tx_queue_id,
 964                    uint16_t nb_tx_desc __rte_unused,
 965                    unsigned int socket_id __rte_unused,
 966                    const struct rte_eth_txconf *tx_conf __rte_unused)
 967 {
 968         struct pmd_internals *internals = dev->data->dev_private;
 969         struct pkt_tx_queue *txq;
 970
 971         txq = &internals->tx_queues[tx_queue_id];
 972
 973         dev->data->tx_queues[tx_queue_id] = txq;
 974         return 0;
 975 }
 976
 977 static int
 978 eth_dev_mtu_set(struct rte_eth_dev *dev, uint16_t mtu)
 979 {
 980         struct pmd_internals *internals = dev->data->dev_private;
 981         struct ifreq ifr = { .ifr_mtu = mtu };
 982         int ret;
 983         int s;
 984
 985         s = socket(PF_INET, SOCK_DGRAM, 0);
 986         if (s < 0)
 987                 return -EINVAL;
 988
 989         strlcpy(ifr.ifr_name, internals->if_name, IFNAMSIZ);
 990         ret = ioctl(s, SIOCSIFMTU, &ifr);
 991         close(s);
 992
 993         return (ret < 0) ? -errno : 0;
 994 }
 995
 996 static int
 997 eth_dev_change_flags(char *if_name, uint32_t flags, uint32_t mask)
 998 {
 999         struct ifreq ifr;
1000         int ret = 0;
1001         int s;
1002
1003         s = socket(PF_INET, SOCK_DGRAM, 0);
1004         if (s < 0)
1005                 return -errno;
1006
1007         strlcpy(ifr.ifr_name, if_name, IFNAMSIZ);
1008         if (ioctl(s, SIOCGIFFLAGS, &ifr) < 0) {
1009                 ret = -errno;
1010                 goto out;
1011         }
1012         ifr.ifr_flags &= mask;
1013         ifr.ifr_flags |= flags;
1014         if (ioctl(s, SIOCSIFFLAGS, &ifr) < 0) {
1015                 ret = -errno;
1016                 goto out;
1017         }
1018 out:
1019         close(s);
1020         return ret;
1021 }
1022
1023 static int
1024 eth_dev_promiscuous_enable(struct rte_eth_dev *dev)
1025 {
1026         struct pmd_internals *internals = dev->data->dev_private;
1027
1028         return eth_dev_change_flags(internals->if_name, IFF_PROMISC, ~0);
1029 }
1030
1031 static int
1032 eth_dev_promiscuous_disable(struct rte_eth_dev *dev)
1033 {
1034         struct pmd_internals *internals = dev->data->dev_private;
1035
1036         return eth_dev_change_flags(internals->if_name, 0, ~IFF_PROMISC);
1037 }
1038
1039 static const struct eth_dev_ops ops = {
1040         .dev_start = eth_dev_start,
1041         .dev_stop = eth_dev_stop,
1042         .dev_close = eth_dev_close,
1043         .dev_configure = eth_dev_configure,
1044         .dev_infos_get = eth_dev_info,
1045         .mtu_set = eth_dev_mtu_set,
1046         .promiscuous_enable = eth_dev_promiscuous_enable,
1047         .promiscuous_disable = eth_dev_promiscuous_disable,
1048         .rx_queue_setup = eth_rx_queue_setup,
1049         .tx_queue_setup = eth_tx_queue_setup,
1050         .rx_queue_release = eth_queue_release,
1051         .tx_queue_release = eth_queue_release,
1052         .link_update = eth_link_update,
1053         .stats_get = eth_stats_get,
1054         .stats_reset = eth_stats_reset,
1055 };
1056
1057 /** parse integer from integer argument */
1058 static int
1059 parse_integer_arg(const char *key __rte_unused,
1060                   const char *value, void *extra_args)
1061 {
1062         int *i = (int *)extra_args;
1063         char *end;
1064
1065         *i = strtol(value, &end, 10);
1066         if (*i < 0) {
1067                 AF_XDP_LOG(ERR, "Argument has to be positive.\n");
1068                 return -EINVAL;
1069         }
1070
1071         return 0;
1072 }
1073
1074 /** parse name argument */
1075 static int
1076 parse_name_arg(const char *key __rte_unused,
1077                const char *value, void *extra_args)
1078 {
1079         char *name = extra_args;
1080
1081         if (strnlen(value, IFNAMSIZ) > IFNAMSIZ - 1) {
1082                 AF_XDP_LOG(ERR, "Invalid name %s, should be less than %u bytes.\n",
1083                            value, IFNAMSIZ);
1084                 return -EINVAL;
1085         }
1086
1087         strlcpy(name, value, IFNAMSIZ);
1088
1089         return 0;
1090 }
1091
1092 static int
1093 xdp_get_channels_info(const char *if_name, int *max_queues,
1094                                 int *combined_queues)
1095 {
1096         struct ethtool_channels channels;
1097         struct ifreq ifr;
1098         int fd, ret;
1099
1100         fd = socket(AF_INET, SOCK_DGRAM, 0);
1101         if (fd < 0)
1102                 return -1;
1103
1104         channels.cmd = ETHTOOL_GCHANNELS;
1105         ifr.ifr_data = (void *)&channels;
1106         strncpy(ifr.ifr_name, if_name, IFNAMSIZ);
1107         ret = ioctl(fd, SIOCETHTOOL, &ifr);
1108         if (ret) {
1109                 if (errno == EOPNOTSUPP) {
1110                         ret = 0;
1111                 } else {
1112                         ret = -errno;
1113                         goto out;
1114                 }
1115         }
1116
1117         if (channels.max_combined == 0 || errno == EOPNOTSUPP) {
1118                 /* If the device says it has no channels, then all traffic
1119                  * is sent to a single stream, so max queues = 1.
1120                  */
1121                 *max_queues = 1;
1122                 *combined_queues = 1;
1123         } else {
1124                 *max_queues = channels.max_combined;
1125                 *combined_queues = channels.combined_count;
1126         }
1127
1128  out:
1129         close(fd);
1130         return ret;
1131 }
1132
1133 static int
1134 parse_parameters(struct rte_kvargs *kvlist, char *if_name, int *start_queue,
1135                         int *queue_cnt)
1136 {
1137         int ret;
1138
1139         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_IFACE_ARG,
1140                                  &parse_name_arg, if_name);
1141         if (ret < 0)
1142                 goto free_kvlist;
1143
1144         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_START_QUEUE_ARG,
1145                                  &parse_integer_arg, start_queue);
1146         if (ret < 0)
1147                 goto free_kvlist;
1148
1149         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_QUEUE_COUNT_ARG,
1150                                  &parse_integer_arg, queue_cnt);
1151         if (ret < 0 || *queue_cnt <= 0) {
1152                 ret = -EINVAL;
1153                 goto free_kvlist;
1154         }
1155
1156 free_kvlist:
1157         rte_kvargs_free(kvlist);
1158         return ret;
1159 }
1160
1161 static int
1162 get_iface_info(const char *if_name,
1163                struct rte_ether_addr *eth_addr,
1164                int *if_index)
1165 {
1166         struct ifreq ifr;
1167         int sock = socket(AF_INET, SOCK_DGRAM, IPPROTO_IP);
1168
1169         if (sock < 0)
1170                 return -1;
1171
1172         strlcpy(ifr.ifr_name, if_name, IFNAMSIZ);
1173         if (ioctl(sock, SIOCGIFINDEX, &ifr))
1174                 goto error;
1175
1176         *if_index = ifr.ifr_ifindex;
1177
1178         if (ioctl(sock, SIOCGIFHWADDR, &ifr))
1179                 goto error;
1180
1181         rte_memcpy(eth_addr, ifr.ifr_hwaddr.sa_data, RTE_ETHER_ADDR_LEN);
1182
1183         close(sock);
1184         return 0;
1185
1186 error:
1187         close(sock);
1188         return -1;
1189 }
1190
1191 static struct rte_eth_dev *
1192 init_internals(struct rte_vdev_device *dev, const char *if_name,
1193                         int start_queue_idx, int queue_cnt)
1194 {
1195         const char *name = rte_vdev_device_name(dev);
1196         const unsigned int numa_node = dev->device.numa_node;
1197         struct pmd_internals *internals;
1198         struct rte_eth_dev *eth_dev;
1199         int ret;
1200         int i;
1201
1202         internals = rte_zmalloc_socket(name, sizeof(*internals), 0, numa_node);
1203         if (internals == NULL)
1204                 return NULL;
1205
1206         internals->start_queue_idx = start_queue_idx;
1207         internals->queue_cnt = queue_cnt;
1208         strlcpy(internals->if_name, if_name, IFNAMSIZ);
1209
1210         if (xdp_get_channels_info(if_name, &internals->max_queue_cnt,
1211                                   &internals->combined_queue_cnt)) {
1212                 AF_XDP_LOG(ERR, "Failed to get channel info of interface: %s\n",
1213                                 if_name);
1214                 goto err_free_internals;
1215         }
1216
1217         if (queue_cnt > internals->combined_queue_cnt) {
1218                 AF_XDP_LOG(ERR, "Specified queue count %d is larger than combined queue count %d.\n",
1219                                 queue_cnt, internals->combined_queue_cnt);
1220                 goto err_free_internals;
1221         }
1222
1223         internals->rx_queues = rte_zmalloc_socket(NULL,
1224                                         sizeof(struct pkt_rx_queue) * queue_cnt,
1225                                         0, numa_node);
1226         if (internals->rx_queues == NULL) {
1227                 AF_XDP_LOG(ERR, "Failed to allocate memory for rx queues.\n");
1228                 goto err_free_internals;
1229         }
1230
1231         internals->tx_queues = rte_zmalloc_socket(NULL,
1232                                         sizeof(struct pkt_tx_queue) * queue_cnt,
1233                                         0, numa_node);
1234         if (internals->tx_queues == NULL) {
1235                 AF_XDP_LOG(ERR, "Failed to allocate memory for tx queues.\n");
1236                 goto err_free_rx;
1237         }
1238         for (i = 0; i < queue_cnt; i++) {
1239                 internals->tx_queues[i].pair = &internals->rx_queues[i];
1240                 internals->rx_queues[i].pair = &internals->tx_queues[i];
1241                 internals->rx_queues[i].xsk_queue_idx = start_queue_idx + i;
1242                 internals->tx_queues[i].xsk_queue_idx = start_queue_idx + i;
1243         }
1244
1245         ret = get_iface_info(if_name, &internals->eth_addr,
1246                              &internals->if_index);
1247         if (ret)
1248                 goto err_free_tx;
1249
1250         eth_dev = rte_eth_vdev_allocate(dev, 0);
1251         if (eth_dev == NULL)
1252                 goto err_free_tx;
1253
1254         eth_dev->data->dev_private = internals;
1255         eth_dev->data->dev_link = pmd_link;
1256         eth_dev->data->mac_addrs = &internals->eth_addr;
1257         eth_dev->dev_ops = &ops;
1258         eth_dev->rx_pkt_burst = eth_af_xdp_rx;
1259         eth_dev->tx_pkt_burst = eth_af_xdp_tx;
1260         /* Let rte_eth_dev_close() release the port resources. */
1261         eth_dev->data->dev_flags |= RTE_ETH_DEV_CLOSE_REMOVE;
1262
1263 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
1264         AF_XDP_LOG(INFO, "Zero copy between umem and mbuf enabled.\n");
1265 #endif
1266
1267         return eth_dev;
1268
1269 err_free_tx:
1270         rte_free(internals->tx_queues);
1271 err_free_rx:
1272         rte_free(internals->rx_queues);
1273 err_free_internals:
1274         rte_free(internals);
1275         return NULL;
1276 }
1277
1278 static int
1279 rte_pmd_af_xdp_probe(struct rte_vdev_device *dev)
1280 {
1281         struct rte_kvargs *kvlist;
1282         char if_name[IFNAMSIZ] = {'\0'};
1283         int xsk_start_queue_idx = ETH_AF_XDP_DFLT_START_QUEUE_IDX;
1284         int xsk_queue_cnt = ETH_AF_XDP_DFLT_QUEUE_COUNT;
1285         struct rte_eth_dev *eth_dev = NULL;
1286         const char *name;
1287
1288         AF_XDP_LOG(INFO, "Initializing pmd_af_xdp for %s\n",
1289                 rte_vdev_device_name(dev));
1290
1291         name = rte_vdev_device_name(dev);
1292         if (rte_eal_process_type() == RTE_PROC_SECONDARY &&
1293                 strlen(rte_vdev_device_args(dev)) == 0) {
1294                 eth_dev = rte_eth_dev_attach_secondary(name);
1295                 if (eth_dev == NULL) {
1296                         AF_XDP_LOG(ERR, "Failed to probe %s\n", name);
1297                         return -EINVAL;
1298                 }
1299                 eth_dev->dev_ops = &ops;
1300                 rte_eth_dev_probing_finish(eth_dev);
1301                 return 0;
1302         }
1303
1304         kvlist = rte_kvargs_parse(rte_vdev_device_args(dev), valid_arguments);
1305         if (kvlist == NULL) {
1306                 AF_XDP_LOG(ERR, "Invalid kvargs key\n");
1307                 return -EINVAL;
1308         }
1309
1310         if (dev->device.numa_node == SOCKET_ID_ANY)
1311                 dev->device.numa_node = rte_socket_id();
1312
1313         if (parse_parameters(kvlist, if_name, &xsk_start_queue_idx,
1314                              &xsk_queue_cnt) < 0) {
1315                 AF_XDP_LOG(ERR, "Invalid kvargs value\n");
1316                 return -EINVAL;
1317         }
1318
1319         if (strlen(if_name) == 0) {
1320                 AF_XDP_LOG(ERR, "Network interface must be specified\n");
1321                 return -EINVAL;
1322         }
1323
1324         eth_dev = init_internals(dev, if_name, xsk_start_queue_idx,
1325                                         xsk_queue_cnt);
1326         if (eth_dev == NULL) {
1327                 AF_XDP_LOG(ERR, "Failed to init internals\n");
1328                 return -1;
1329         }
1330
1331         rte_eth_dev_probing_finish(eth_dev);
1332
1333         return 0;
1334 }
1335
1336 static int
1337 rte_pmd_af_xdp_remove(struct rte_vdev_device *dev)
1338 {
1339         struct rte_eth_dev *eth_dev = NULL;
1340
1341         AF_XDP_LOG(INFO, "Removing AF_XDP ethdev on numa socket %u\n",
1342                 rte_socket_id());
1343
1344         if (dev == NULL)
1345                 return -1;
1346
1347         /* find the ethdev entry */
1348         eth_dev = rte_eth_dev_allocated(rte_vdev_device_name(dev));
1349         if (eth_dev == NULL)
1350                 return 0;
1351
1352         eth_dev_close(eth_dev);
1353         rte_eth_dev_release_port(eth_dev);
1354
1355
1356         return 0;
1357 }
1358
1359 static struct rte_vdev_driver pmd_af_xdp_drv = {
1360         .probe = rte_pmd_af_xdp_probe,
1361         .remove = rte_pmd_af_xdp_remove,
1362 };
1363
1364 RTE_PMD_REGISTER_VDEV(net_af_xdp, pmd_af_xdp_drv);
1365 RTE_PMD_REGISTER_PARAM_STRING(net_af_xdp,
1366                               "iface=<string> "
1367                               "start_queue=<int> "
1368                               "queue_count=<int> ");
1369
1370 RTE_INIT(af_xdp_init_log)
1371 {
1372         af_xdp_logtype = rte_log_register("pmd.net.af_xdp");
1373         if (af_xdp_logtype >= 0)
1374                 rte_log_set_level(af_xdp_logtype, RTE_LOG_NOTICE);
1375 }