drivers/net/af_xdp/rte_eth_af_xdp.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2019 Intel Corporation.
   3  */
   4 #include <unistd.h>
   5 #include <errno.h>
   6 #include <stdlib.h>
   7 #include <string.h>
   8 #include <netinet/in.h>
   9 #include <net/if.h>
  10 #include <sys/socket.h>
  11 #include <sys/ioctl.h>
  12 #include <linux/if_ether.h>
  13 #include <linux/if_xdp.h>
  14 #include <linux/if_link.h>
  15 #include "af_xdp_deps.h"
  16 #include <bpf/xsk.h>
  17
  18 #include <rte_ethdev.h>
  19 #include <rte_ethdev_driver.h>
  20 #include <rte_ethdev_vdev.h>
  21 #include <rte_kvargs.h>
  22 #include <rte_bus_vdev.h>
  23 #include <rte_string_fns.h>
  24 #include <rte_branch_prediction.h>
  25 #include <rte_common.h>
  26 #include <rte_config.h>
  27 #include <rte_dev.h>
  28 #include <rte_eal.h>
  29 #include <rte_ether.h>
  30 #include <rte_lcore.h>
  31 #include <rte_log.h>
  32 #include <rte_memory.h>
  33 #include <rte_memzone.h>
  34 #include <rte_mbuf.h>
  35 #include <rte_malloc.h>
  36 #include <rte_ring.h>
  37
  38 #ifndef SOL_XDP
  39 #define SOL_XDP 283
  40 #endif
  41
  42 #ifndef AF_XDP
  43 #define AF_XDP 44
  44 #endif
  45
  46 #ifndef PF_XDP
  47 #define PF_XDP AF_XDP
  48 #endif
  49
  50 static int af_xdp_logtype;
  51
  52 #define AF_XDP_LOG(level, fmt, args...)                 \
  53         rte_log(RTE_LOG_ ## level, af_xdp_logtype,      \
  54                 "%s(): " fmt, __func__, ##args)
  55
  56 #define ETH_AF_XDP_FRAME_SIZE           XSK_UMEM__DEFAULT_FRAME_SIZE
  57 #define ETH_AF_XDP_NUM_BUFFERS          4096
  58 #define ETH_AF_XDP_DATA_HEADROOM        0
  59 #define ETH_AF_XDP_DFLT_NUM_DESCS       XSK_RING_CONS__DEFAULT_NUM_DESCS
  60 #define ETH_AF_XDP_DFLT_QUEUE_IDX       0
  61
  62 #define ETH_AF_XDP_RX_BATCH_SIZE        32
  63 #define ETH_AF_XDP_TX_BATCH_SIZE        32
  64
  65 #define ETH_AF_XDP_MAX_QUEUE_PAIRS     16
  66
  67 struct xsk_umem_info {
  68         struct xsk_ring_prod fq;
  69         struct xsk_ring_cons cq;
  70         struct xsk_umem *umem;
  71         struct rte_ring *buf_ring;
  72         const struct rte_memzone *mz;
  73         int pmd_zc;
  74 };
  75
  76 struct rx_stats {
  77         uint64_t rx_pkts;
  78         uint64_t rx_bytes;
  79         uint64_t rx_dropped;
  80 };
  81
  82 struct pkt_rx_queue {
  83         struct xsk_ring_cons rx;
  84         struct xsk_umem_info *umem;
  85         struct xsk_socket *xsk;
  86         struct rte_mempool *mb_pool;
  87
  88         struct rx_stats stats;
  89
  90         struct pkt_tx_queue *pair;
  91         uint16_t queue_idx;
  92 };
  93
  94 struct tx_stats {
  95         uint64_t tx_pkts;
  96         uint64_t err_pkts;
  97         uint64_t tx_bytes;
  98 };
  99
 100 struct pkt_tx_queue {
 101         struct xsk_ring_prod tx;
 102
 103         struct tx_stats stats;
 104
 105         struct pkt_rx_queue *pair;
 106         uint16_t queue_idx;
 107 };
 108
 109 struct pmd_internals {
 110         int if_index;
 111         char if_name[IFNAMSIZ];
 112         uint16_t queue_idx;
 113         int pmd_zc;
 114         struct rte_ether_addr eth_addr;
 115         struct rte_mempool *mb_pool_share;
 116
 117         struct pkt_rx_queue rx_queues[ETH_AF_XDP_MAX_QUEUE_PAIRS];
 118         struct pkt_tx_queue tx_queues[ETH_AF_XDP_MAX_QUEUE_PAIRS];
 119 };
 120
 121 #define ETH_AF_XDP_IFACE_ARG                    "iface"
 122 #define ETH_AF_XDP_QUEUE_IDX_ARG                "queue"
 123 #define ETH_AF_XDP_PMD_ZC_ARG                   "pmd_zero_copy"
 124
 125 static const char * const valid_arguments[] = {
 126         ETH_AF_XDP_IFACE_ARG,
 127         ETH_AF_XDP_QUEUE_IDX_ARG,
 128         ETH_AF_XDP_PMD_ZC_ARG,
 129         NULL
 130 };
 131
 132 static const struct rte_eth_link pmd_link = {
 133         .link_speed = ETH_SPEED_NUM_10G,
 134         .link_duplex = ETH_LINK_FULL_DUPLEX,
 135         .link_status = ETH_LINK_DOWN,
 136         .link_autoneg = ETH_LINK_AUTONEG
 137 };
 138
 139 static inline int
 140 reserve_fill_queue(struct xsk_umem_info *umem, uint16_t reserve_size)
 141 {
 142         struct xsk_ring_prod *fq = &umem->fq;
 143         void *addrs[reserve_size];
 144         uint32_t idx;
 145         uint16_t i;
 146
 147         if (rte_ring_dequeue_bulk(umem->buf_ring, addrs, reserve_size, NULL)
 148                     != reserve_size) {
 149                 AF_XDP_LOG(DEBUG, "Failed to get enough buffers for fq.\n");
 150                 return -1;
 151         }
 152
 153         if (unlikely(!xsk_ring_prod__reserve(fq, reserve_size, &idx))) {
 154                 AF_XDP_LOG(DEBUG, "Failed to reserve enough fq descs.\n");
 155                 rte_ring_enqueue_bulk(umem->buf_ring, addrs,
 156                                 reserve_size, NULL);
 157                 return -1;
 158         }
 159
 160         for (i = 0; i < reserve_size; i++) {
 161                 __u64 *fq_addr;
 162
 163                 fq_addr = xsk_ring_prod__fill_addr(fq, idx++);
 164                 *fq_addr = (uint64_t)addrs[i];
 165         }
 166
 167         xsk_ring_prod__submit(fq, reserve_size);
 168
 169         return 0;
 170 }
 171
 172 static void
 173 umem_buf_release_to_fq(void *addr, void *opaque)
 174 {
 175         struct xsk_umem_info *umem = (struct xsk_umem_info *)opaque;
 176         uint64_t umem_addr = (uint64_t)addr - umem->mz->addr_64;
 177
 178         rte_ring_enqueue(umem->buf_ring, (void *)umem_addr);
 179 }
 180
 181 static uint16_t
 182 eth_af_xdp_rx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 183 {
 184         struct pkt_rx_queue *rxq = queue;
 185         struct xsk_ring_cons *rx = &rxq->rx;
 186         struct xsk_umem_info *umem = rxq->umem;
 187         struct xsk_ring_prod *fq = &umem->fq;
 188         uint32_t idx_rx = 0;
 189         uint32_t free_thresh = fq->size >> 1;
 190         int pmd_zc = umem->pmd_zc;
 191         struct rte_mbuf *mbufs[ETH_AF_XDP_RX_BATCH_SIZE];
 192         unsigned long dropped = 0;
 193         unsigned long rx_bytes = 0;
 194         int rcvd, i;
 195
 196         nb_pkts = RTE_MIN(nb_pkts, ETH_AF_XDP_RX_BATCH_SIZE);
 197
 198         if (unlikely(rte_pktmbuf_alloc_bulk(rxq->mb_pool, mbufs, nb_pkts) != 0))
 199                 return 0;
 200
 201         rcvd = xsk_ring_cons__peek(rx, nb_pkts, &idx_rx);
 202         if (rcvd == 0)
 203                 goto out;
 204
 205         if (xsk_prod_nb_free(fq, free_thresh) >= free_thresh)
 206                 (void)reserve_fill_queue(umem, ETH_AF_XDP_RX_BATCH_SIZE);
 207
 208         for (i = 0; i < rcvd; i++) {
 209                 const struct xdp_desc *desc;
 210                 uint64_t addr;
 211                 uint32_t len;
 212                 void *pkt;
 213                 uint16_t buf_len = ETH_AF_XDP_FRAME_SIZE;
 214                 struct rte_mbuf_ext_shared_info *shinfo;
 215
 216                 desc = xsk_ring_cons__rx_desc(rx, idx_rx++);
 217                 addr = desc->addr;
 218                 len = desc->len;
 219                 pkt = xsk_umem__get_data(rxq->umem->mz->addr, addr);
 220
 221                 if (pmd_zc) {
 222                         shinfo = rte_pktmbuf_ext_shinfo_init_helper(pkt,
 223                                         &buf_len, umem_buf_release_to_fq, umem);
 224
 225                         rte_pktmbuf_attach_extbuf(mbufs[i], pkt, 0, buf_len,
 226                                                   shinfo);
 227                 } else {
 228                         rte_memcpy(rte_pktmbuf_mtod(mbufs[i], void *),
 229                                                         pkt, len);
 230                         rte_ring_enqueue(umem->buf_ring, (void *)addr);
 231                 }
 232                 rte_pktmbuf_pkt_len(mbufs[i]) = len;
 233                 rte_pktmbuf_data_len(mbufs[i]) = len;
 234                 rx_bytes += len;
 235                 bufs[i] = mbufs[i];
 236         }
 237
 238         xsk_ring_cons__release(rx, rcvd);
 239
 240         /* statistics */
 241         rxq->stats.rx_pkts += (rcvd - dropped);
 242         rxq->stats.rx_bytes += rx_bytes;
 243
 244 out:
 245         if (rcvd != nb_pkts)
 246                 rte_mempool_put_bulk(rxq->mb_pool, (void **)&mbufs[rcvd],
 247                                      nb_pkts - rcvd);
 248
 249         return rcvd;
 250 }
 251
 252 static void
 253 pull_umem_cq(struct xsk_umem_info *umem, int size)
 254 {
 255         struct xsk_ring_cons *cq = &umem->cq;
 256         size_t i, n;
 257         uint32_t idx_cq = 0;
 258
 259         n = xsk_ring_cons__peek(cq, size, &idx_cq);
 260
 261         for (i = 0; i < n; i++) {
 262                 uint64_t addr;
 263                 addr = *xsk_ring_cons__comp_addr(cq, idx_cq++);
 264                 rte_ring_enqueue(umem->buf_ring, (void *)addr);
 265         }
 266
 267         xsk_ring_cons__release(cq, n);
 268 }
 269
 270 static void
 271 kick_tx(struct pkt_tx_queue *txq)
 272 {
 273         struct xsk_umem_info *umem = txq->pair->umem;
 274
 275         while (send(xsk_socket__fd(txq->pair->xsk), NULL,
 276                       0, MSG_DONTWAIT) < 0) {
 277                 /* some thing unexpected */
 278                 if (errno != EBUSY && errno != EAGAIN && errno != EINTR)
 279                         break;
 280
 281                 /* pull from completion queue to leave more space */
 282                 if (errno == EAGAIN)
 283                         pull_umem_cq(umem, ETH_AF_XDP_TX_BATCH_SIZE);
 284         }
 285         pull_umem_cq(umem, ETH_AF_XDP_TX_BATCH_SIZE);
 286 }
 287
 288 static inline bool
 289 in_umem_range(struct xsk_umem_info *umem, uint64_t addr)
 290 {
 291         uint64_t mz_base_addr = umem->mz->addr_64;
 292
 293         return addr >= mz_base_addr && addr < mz_base_addr + umem->mz->len;
 294 }
 295
 296 static uint16_t
 297 eth_af_xdp_tx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 298 {
 299         struct pkt_tx_queue *txq = queue;
 300         struct xsk_umem_info *umem = txq->pair->umem;
 301         struct rte_mbuf *mbuf;
 302         int pmd_zc = umem->pmd_zc;
 303         void *addrs[ETH_AF_XDP_TX_BATCH_SIZE];
 304         unsigned long tx_bytes = 0;
 305         int i;
 306         uint32_t idx_tx;
 307
 308         nb_pkts = RTE_MIN(nb_pkts, ETH_AF_XDP_TX_BATCH_SIZE);
 309
 310         pull_umem_cq(umem, nb_pkts);
 311
 312         nb_pkts = rte_ring_dequeue_bulk(umem->buf_ring, addrs,
 313                                         nb_pkts, NULL);
 314         if (nb_pkts == 0)
 315                 return 0;
 316
 317         if (xsk_ring_prod__reserve(&txq->tx, nb_pkts, &idx_tx) != nb_pkts) {
 318                 kick_tx(txq);
 319                 rte_ring_enqueue_bulk(umem->buf_ring, addrs, nb_pkts, NULL);
 320                 return 0;
 321         }
 322
 323         for (i = 0; i < nb_pkts; i++) {
 324                 struct xdp_desc *desc;
 325                 void *pkt;
 326
 327                 desc = xsk_ring_prod__tx_desc(&txq->tx, idx_tx + i);
 328                 mbuf = bufs[i];
 329                 desc->len = mbuf->pkt_len;
 330
 331                 /*
 332                  * We need to make sure the external mbuf address is within
 333                  * current port's umem memzone range
 334                  */
 335                 if (pmd_zc && RTE_MBUF_HAS_EXTBUF(mbuf) &&
 336                                 in_umem_range(umem, (uint64_t)mbuf->buf_addr)) {
 337                         desc->addr = (uint64_t)mbuf->buf_addr -
 338                                 umem->mz->addr_64;
 339                         mbuf->buf_addr = xsk_umem__get_data(umem->mz->addr,
 340                                         (uint64_t)addrs[i]);
 341                 } else {
 342                         desc->addr = (uint64_t)addrs[i];
 343                         pkt = xsk_umem__get_data(umem->mz->addr,
 344                                         desc->addr);
 345                         rte_memcpy(pkt, rte_pktmbuf_mtod(mbuf, void *),
 346                                         desc->len);
 347                 }
 348                 tx_bytes += mbuf->pkt_len;
 349         }
 350
 351         xsk_ring_prod__submit(&txq->tx, nb_pkts);
 352
 353         kick_tx(txq);
 354
 355         txq->stats.tx_pkts += nb_pkts;
 356         txq->stats.tx_bytes += tx_bytes;
 357
 358         for (i = 0; i < nb_pkts; i++)
 359                 rte_pktmbuf_free(bufs[i]);
 360
 361         return nb_pkts;
 362 }
 363
 364 static int
 365 eth_dev_start(struct rte_eth_dev *dev)
 366 {
 367         dev->data->dev_link.link_status = ETH_LINK_UP;
 368
 369         return 0;
 370 }
 371
 372 /* This function gets called when the current port gets stopped. */
 373 static void
 374 eth_dev_stop(struct rte_eth_dev *dev)
 375 {
 376         dev->data->dev_link.link_status = ETH_LINK_DOWN;
 377 }
 378
 379 static int
 380 eth_dev_configure(struct rte_eth_dev *dev)
 381 {
 382         /* rx/tx must be paired */
 383         if (dev->data->nb_rx_queues != dev->data->nb_tx_queues)
 384                 return -EINVAL;
 385
 386         return 0;
 387 }
 388
 389 static void
 390 eth_dev_info(struct rte_eth_dev *dev, struct rte_eth_dev_info *dev_info)
 391 {
 392         struct pmd_internals *internals = dev->data->dev_private;
 393
 394         dev_info->if_index = internals->if_index;
 395         dev_info->max_mac_addrs = 1;
 396         dev_info->max_rx_pktlen = ETH_FRAME_LEN;
 397         dev_info->max_rx_queues = 1;
 398         dev_info->max_tx_queues = 1;
 399
 400         dev_info->min_mtu = RTE_ETHER_MIN_MTU;
 401         dev_info->max_mtu = ETH_AF_XDP_FRAME_SIZE - ETH_AF_XDP_DATA_HEADROOM;
 402
 403         dev_info->default_rxportconf.nb_queues = 1;
 404         dev_info->default_txportconf.nb_queues = 1;
 405         dev_info->default_rxportconf.ring_size = ETH_AF_XDP_DFLT_NUM_DESCS;
 406         dev_info->default_txportconf.ring_size = ETH_AF_XDP_DFLT_NUM_DESCS;
 407 }
 408
 409 static int
 410 eth_stats_get(struct rte_eth_dev *dev, struct rte_eth_stats *stats)
 411 {
 412         struct pmd_internals *internals = dev->data->dev_private;
 413         struct xdp_statistics xdp_stats;
 414         struct pkt_rx_queue *rxq;
 415         socklen_t optlen;
 416         int i, ret;
 417
 418         for (i = 0; i < dev->data->nb_rx_queues; i++) {
 419                 optlen = sizeof(struct xdp_statistics);
 420                 rxq = &internals->rx_queues[i];
 421                 stats->q_ipackets[i] = internals->rx_queues[i].stats.rx_pkts;
 422                 stats->q_ibytes[i] = internals->rx_queues[i].stats.rx_bytes;
 423
 424                 stats->q_opackets[i] = internals->tx_queues[i].stats.tx_pkts;
 425                 stats->q_obytes[i] = internals->tx_queues[i].stats.tx_bytes;
 426
 427                 stats->ipackets += stats->q_ipackets[i];
 428                 stats->ibytes += stats->q_ibytes[i];
 429                 stats->imissed += internals->rx_queues[i].stats.rx_dropped;
 430                 ret = getsockopt(xsk_socket__fd(rxq->xsk), SOL_XDP,
 431                                 XDP_STATISTICS, &xdp_stats, &optlen);
 432                 if (ret != 0) {
 433                         AF_XDP_LOG(ERR, "getsockopt() failed for XDP_STATISTICS.\n");
 434                         return -1;
 435                 }
 436                 stats->imissed += xdp_stats.rx_dropped;
 437
 438                 stats->opackets += stats->q_opackets[i];
 439                 stats->oerrors += internals->tx_queues[i].stats.err_pkts;
 440                 stats->obytes += stats->q_obytes[i];
 441         }
 442
 443         return 0;
 444 }
 445
 446 static void
 447 eth_stats_reset(struct rte_eth_dev *dev)
 448 {
 449         struct pmd_internals *internals = dev->data->dev_private;
 450         int i;
 451
 452         for (i = 0; i < ETH_AF_XDP_MAX_QUEUE_PAIRS; i++) {
 453                 memset(&internals->rx_queues[i].stats, 0,
 454                                         sizeof(struct rx_stats));
 455                 memset(&internals->tx_queues[i].stats, 0,
 456                                         sizeof(struct tx_stats));
 457         }
 458 }
 459
 460 static void
 461 remove_xdp_program(struct pmd_internals *internals)
 462 {
 463         uint32_t curr_prog_id = 0;
 464
 465         if (bpf_get_link_xdp_id(internals->if_index, &curr_prog_id,
 466                                 XDP_FLAGS_UPDATE_IF_NOEXIST)) {
 467                 AF_XDP_LOG(ERR, "bpf_get_link_xdp_id failed\n");
 468                 return;
 469         }
 470         bpf_set_link_xdp_fd(internals->if_index, -1,
 471                         XDP_FLAGS_UPDATE_IF_NOEXIST);
 472 }
 473
 474 static void
 475 xdp_umem_destroy(struct xsk_umem_info *umem)
 476 {
 477         rte_memzone_free(umem->mz);
 478         umem->mz = NULL;
 479
 480         rte_ring_free(umem->buf_ring);
 481         umem->buf_ring = NULL;
 482
 483         rte_free(umem);
 484         umem = NULL;
 485 }
 486
 487 static void
 488 eth_dev_close(struct rte_eth_dev *dev)
 489 {
 490         struct pmd_internals *internals = dev->data->dev_private;
 491         struct pkt_rx_queue *rxq;
 492         int i;
 493
 494         AF_XDP_LOG(INFO, "Closing AF_XDP ethdev on numa socket %u\n",
 495                 rte_socket_id());
 496
 497         for (i = 0; i < ETH_AF_XDP_MAX_QUEUE_PAIRS; i++) {
 498                 rxq = &internals->rx_queues[i];
 499                 if (rxq->umem == NULL)
 500                         break;
 501                 xsk_socket__delete(rxq->xsk);
 502                 (void)xsk_umem__delete(rxq->umem->umem);
 503                 xdp_umem_destroy(rxq->umem);
 504         }
 505
 506         /*
 507          * MAC is not allocated dynamically, setting it to NULL would prevent
 508          * from releasing it in rte_eth_dev_release_port.
 509          */
 510         dev->data->mac_addrs = NULL;
 511
 512         remove_xdp_program(internals);
 513 }
 514
 515 static void
 516 eth_queue_release(void *q __rte_unused)
 517 {
 518 }
 519
 520 static int
 521 eth_link_update(struct rte_eth_dev *dev __rte_unused,
 522                 int wait_to_complete __rte_unused)
 523 {
 524         return 0;
 525 }
 526
 527 static struct
 528 xsk_umem_info *xdp_umem_configure(struct pmd_internals *internals)
 529 {
 530         struct xsk_umem_info *umem;
 531         const struct rte_memzone *mz;
 532         struct xsk_umem_config usr_config = {
 533                 .fill_size = ETH_AF_XDP_DFLT_NUM_DESCS,
 534                 .comp_size = ETH_AF_XDP_DFLT_NUM_DESCS,
 535                 .frame_size = ETH_AF_XDP_FRAME_SIZE,
 536                 .frame_headroom = ETH_AF_XDP_DATA_HEADROOM };
 537         char ring_name[RTE_RING_NAMESIZE];
 538         char mz_name[RTE_MEMZONE_NAMESIZE];
 539         int ret;
 540         uint64_t i;
 541
 542         umem = rte_zmalloc_socket("umem", sizeof(*umem), 0, rte_socket_id());
 543         if (umem == NULL) {
 544                 AF_XDP_LOG(ERR, "Failed to allocate umem info");
 545                 return NULL;
 546         }
 547
 548         snprintf(ring_name, sizeof(ring_name), "af_xdp_ring_%s_%u",
 549                        internals->if_name, internals->queue_idx);
 550         umem->buf_ring = rte_ring_create(ring_name,
 551                                          ETH_AF_XDP_NUM_BUFFERS,
 552                                          rte_socket_id(),
 553                                          0x0);
 554         if (umem->buf_ring == NULL) {
 555                 AF_XDP_LOG(ERR, "Failed to create rte_ring\n");
 556                 goto err;
 557         }
 558
 559         for (i = 0; i < ETH_AF_XDP_NUM_BUFFERS; i++)
 560                 rte_ring_enqueue(umem->buf_ring,
 561                                  (void *)(i * ETH_AF_XDP_FRAME_SIZE +
 562                                           ETH_AF_XDP_DATA_HEADROOM));
 563
 564         snprintf(mz_name, sizeof(mz_name), "af_xdp_umem_%s_%u",
 565                        internals->if_name, internals->queue_idx);
 566         mz = rte_memzone_reserve_aligned(mz_name,
 567                         ETH_AF_XDP_NUM_BUFFERS * ETH_AF_XDP_FRAME_SIZE,
 568                         rte_socket_id(), RTE_MEMZONE_IOVA_CONTIG,
 569                         getpagesize());
 570         if (mz == NULL) {
 571                 AF_XDP_LOG(ERR, "Failed to reserve memzone for af_xdp umem.\n");
 572                 goto err;
 573         }
 574
 575         ret = xsk_umem__create(&umem->umem, mz->addr,
 576                                ETH_AF_XDP_NUM_BUFFERS * ETH_AF_XDP_FRAME_SIZE,
 577                                &umem->fq, &umem->cq,
 578                                &usr_config);
 579
 580         if (ret) {
 581                 AF_XDP_LOG(ERR, "Failed to create umem");
 582                 goto err;
 583         }
 584         umem->mz = mz;
 585
 586         return umem;
 587
 588 err:
 589         xdp_umem_destroy(umem);
 590         return NULL;
 591 }
 592
 593 static int
 594 xsk_configure(struct pmd_internals *internals, struct pkt_rx_queue *rxq,
 595               int ring_size)
 596 {
 597         struct xsk_socket_config cfg;
 598         struct pkt_tx_queue *txq = rxq->pair;
 599         int ret = 0;
 600         int reserve_size;
 601
 602         rxq->umem = xdp_umem_configure(internals);
 603         if (rxq->umem == NULL)
 604                 return -ENOMEM;
 605
 606         cfg.rx_size = ring_size;
 607         cfg.tx_size = ring_size;
 608         cfg.libbpf_flags = 0;
 609         cfg.xdp_flags = XDP_FLAGS_UPDATE_IF_NOEXIST;
 610         cfg.bind_flags = 0;
 611         ret = xsk_socket__create(&rxq->xsk, internals->if_name,
 612                         internals->queue_idx, rxq->umem->umem, &rxq->rx,
 613                         &txq->tx, &cfg);
 614         if (ret) {
 615                 AF_XDP_LOG(ERR, "Failed to create xsk socket.\n");
 616                 goto err;
 617         }
 618
 619         reserve_size = ETH_AF_XDP_DFLT_NUM_DESCS / 2;
 620         ret = reserve_fill_queue(rxq->umem, reserve_size);
 621         if (ret) {
 622                 xsk_socket__delete(rxq->xsk);
 623                 AF_XDP_LOG(ERR, "Failed to reserve fill queue.\n");
 624                 goto err;
 625         }
 626
 627         return 0;
 628
 629 err:
 630         xdp_umem_destroy(rxq->umem);
 631
 632         return ret;
 633 }
 634
 635 static void
 636 queue_reset(struct pmd_internals *internals, uint16_t queue_idx)
 637 {
 638         struct pkt_rx_queue *rxq = &internals->rx_queues[queue_idx];
 639         struct pkt_tx_queue *txq = rxq->pair;
 640
 641         memset(rxq, 0, sizeof(*rxq));
 642         memset(txq, 0, sizeof(*txq));
 643         rxq->pair = txq;
 644         txq->pair = rxq;
 645         rxq->queue_idx = queue_idx;
 646         txq->queue_idx = queue_idx;
 647 }
 648
 649 static int
 650 eth_rx_queue_setup(struct rte_eth_dev *dev,
 651                    uint16_t rx_queue_id,
 652                    uint16_t nb_rx_desc,
 653                    unsigned int socket_id __rte_unused,
 654                    const struct rte_eth_rxconf *rx_conf __rte_unused,
 655                    struct rte_mempool *mb_pool)
 656 {
 657         struct pmd_internals *internals = dev->data->dev_private;
 658         uint32_t buf_size, data_size;
 659         struct pkt_rx_queue *rxq;
 660         int ret;
 661
 662         rxq = &internals->rx_queues[rx_queue_id];
 663         queue_reset(internals, rx_queue_id);
 664
 665         /* Now get the space available for data in the mbuf */
 666         buf_size = rte_pktmbuf_data_room_size(mb_pool) -
 667                 RTE_PKTMBUF_HEADROOM;
 668         data_size = ETH_AF_XDP_FRAME_SIZE - ETH_AF_XDP_DATA_HEADROOM;
 669
 670         if (data_size > buf_size) {
 671                 AF_XDP_LOG(ERR, "%s: %d bytes will not fit in mbuf (%d bytes)\n",
 672                         dev->device->name, data_size, buf_size);
 673                 ret = -ENOMEM;
 674                 goto err;
 675         }
 676
 677         rxq->mb_pool = mb_pool;
 678
 679         if (xsk_configure(internals, rxq, nb_rx_desc)) {
 680                 AF_XDP_LOG(ERR, "Failed to configure xdp socket\n");
 681                 ret = -EINVAL;
 682                 goto err;
 683         }
 684
 685         rxq->umem->pmd_zc = internals->pmd_zc;
 686
 687         dev->data->rx_queues[rx_queue_id] = rxq;
 688         return 0;
 689
 690 err:
 691         queue_reset(internals, rx_queue_id);
 692         return ret;
 693 }
 694
 695 static int
 696 eth_tx_queue_setup(struct rte_eth_dev *dev,
 697                    uint16_t tx_queue_id,
 698                    uint16_t nb_tx_desc __rte_unused,
 699                    unsigned int socket_id __rte_unused,
 700                    const struct rte_eth_txconf *tx_conf __rte_unused)
 701 {
 702         struct pmd_internals *internals = dev->data->dev_private;
 703         struct pkt_tx_queue *txq;
 704
 705         txq = &internals->tx_queues[tx_queue_id];
 706
 707         dev->data->tx_queues[tx_queue_id] = txq;
 708         return 0;
 709 }
 710
 711 static int
 712 eth_dev_mtu_set(struct rte_eth_dev *dev, uint16_t mtu)
 713 {
 714         struct pmd_internals *internals = dev->data->dev_private;
 715         struct ifreq ifr = { .ifr_mtu = mtu };
 716         int ret;
 717         int s;
 718
 719         s = socket(PF_INET, SOCK_DGRAM, 0);
 720         if (s < 0)
 721                 return -EINVAL;
 722
 723         strlcpy(ifr.ifr_name, internals->if_name, IFNAMSIZ);
 724         ret = ioctl(s, SIOCSIFMTU, &ifr);
 725         close(s);
 726
 727         return (ret < 0) ? -errno : 0;
 728 }
 729
 730 static void
 731 eth_dev_change_flags(char *if_name, uint32_t flags, uint32_t mask)
 732 {
 733         struct ifreq ifr;
 734         int s;
 735
 736         s = socket(PF_INET, SOCK_DGRAM, 0);
 737         if (s < 0)
 738                 return;
 739
 740         strlcpy(ifr.ifr_name, if_name, IFNAMSIZ);
 741         if (ioctl(s, SIOCGIFFLAGS, &ifr) < 0)
 742                 goto out;
 743         ifr.ifr_flags &= mask;
 744         ifr.ifr_flags |= flags;
 745         if (ioctl(s, SIOCSIFFLAGS, &ifr) < 0)
 746                 goto out;
 747 out:
 748         close(s);
 749 }
 750
 751 static void
 752 eth_dev_promiscuous_enable(struct rte_eth_dev *dev)
 753 {
 754         struct pmd_internals *internals = dev->data->dev_private;
 755
 756         eth_dev_change_flags(internals->if_name, IFF_PROMISC, ~0);
 757 }
 758
 759 static void
 760 eth_dev_promiscuous_disable(struct rte_eth_dev *dev)
 761 {
 762         struct pmd_internals *internals = dev->data->dev_private;
 763
 764         eth_dev_change_flags(internals->if_name, 0, ~IFF_PROMISC);
 765 }
 766
 767 static const struct eth_dev_ops ops = {
 768         .dev_start = eth_dev_start,
 769         .dev_stop = eth_dev_stop,
 770         .dev_close = eth_dev_close,
 771         .dev_configure = eth_dev_configure,
 772         .dev_infos_get = eth_dev_info,
 773         .mtu_set = eth_dev_mtu_set,
 774         .promiscuous_enable = eth_dev_promiscuous_enable,
 775         .promiscuous_disable = eth_dev_promiscuous_disable,
 776         .rx_queue_setup = eth_rx_queue_setup,
 777         .tx_queue_setup = eth_tx_queue_setup,
 778         .rx_queue_release = eth_queue_release,
 779         .tx_queue_release = eth_queue_release,
 780         .link_update = eth_link_update,
 781         .stats_get = eth_stats_get,
 782         .stats_reset = eth_stats_reset,
 783 };
 784
 785 /** parse integer from integer argument */
 786 static int
 787 parse_integer_arg(const char *key __rte_unused,
 788                   const char *value, void *extra_args)
 789 {
 790         int *i = (int *)extra_args;
 791         char *end;
 792
 793         *i = strtol(value, &end, 10);
 794         if (*i < 0) {
 795                 AF_XDP_LOG(ERR, "Argument has to be positive.\n");
 796                 return -EINVAL;
 797         }
 798
 799         return 0;
 800 }
 801
 802 /** parse name argument */
 803 static int
 804 parse_name_arg(const char *key __rte_unused,
 805                const char *value, void *extra_args)
 806 {
 807         char *name = extra_args;
 808
 809         if (strnlen(value, IFNAMSIZ) > IFNAMSIZ - 1) {
 810                 AF_XDP_LOG(ERR, "Invalid name %s, should be less than %u bytes.\n",
 811                            value, IFNAMSIZ);
 812                 return -EINVAL;
 813         }
 814
 815         strlcpy(name, value, IFNAMSIZ);
 816
 817         return 0;
 818 }
 819
 820 static int
 821 parse_parameters(struct rte_kvargs *kvlist, char *if_name, int *queue_idx,
 822                                 int *pmd_zc)
 823 {
 824         int ret;
 825
 826         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_IFACE_ARG,
 827                                  &parse_name_arg, if_name);
 828         if (ret < 0)
 829                 goto free_kvlist;
 830
 831         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_QUEUE_IDX_ARG,
 832                                  &parse_integer_arg, queue_idx);
 833         if (ret < 0)
 834                 goto free_kvlist;
 835
 836         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_PMD_ZC_ARG,
 837                                  &parse_integer_arg, pmd_zc);
 838         if (ret < 0)
 839                 goto free_kvlist;
 840
 841 free_kvlist:
 842         rte_kvargs_free(kvlist);
 843         return ret;
 844 }
 845
 846 static int
 847 get_iface_info(const char *if_name,
 848                struct rte_ether_addr *eth_addr,
 849                int *if_index)
 850 {
 851         struct ifreq ifr;
 852         int sock = socket(AF_INET, SOCK_DGRAM, IPPROTO_IP);
 853
 854         if (sock < 0)
 855                 return -1;
 856
 857         strlcpy(ifr.ifr_name, if_name, IFNAMSIZ);
 858         if (ioctl(sock, SIOCGIFINDEX, &ifr))
 859                 goto error;
 860
 861         *if_index = ifr.ifr_ifindex;
 862
 863         if (ioctl(sock, SIOCGIFHWADDR, &ifr))
 864                 goto error;
 865
 866         rte_memcpy(eth_addr, ifr.ifr_hwaddr.sa_data, RTE_ETHER_ADDR_LEN);
 867
 868         close(sock);
 869         return 0;
 870
 871 error:
 872         close(sock);
 873         return -1;
 874 }
 875
 876 static struct rte_eth_dev *
 877 init_internals(struct rte_vdev_device *dev, const char *if_name, int queue_idx,
 878                                         int pmd_zc)
 879 {
 880         const char *name = rte_vdev_device_name(dev);
 881         const unsigned int numa_node = dev->device.numa_node;
 882         struct pmd_internals *internals;
 883         struct rte_eth_dev *eth_dev;
 884         int ret;
 885         int i;
 886
 887         internals = rte_zmalloc_socket(name, sizeof(*internals), 0, numa_node);
 888         if (internals == NULL)
 889                 return NULL;
 890
 891         internals->queue_idx = queue_idx;
 892         internals->pmd_zc = pmd_zc;
 893         strlcpy(internals->if_name, if_name, IFNAMSIZ);
 894
 895         for (i = 0; i < ETH_AF_XDP_MAX_QUEUE_PAIRS; i++) {
 896                 internals->tx_queues[i].pair = &internals->rx_queues[i];
 897                 internals->rx_queues[i].pair = &internals->tx_queues[i];
 898         }
 899
 900         ret = get_iface_info(if_name, &internals->eth_addr,
 901                              &internals->if_index);
 902         if (ret)
 903                 goto err;
 904
 905         eth_dev = rte_eth_vdev_allocate(dev, 0);
 906         if (eth_dev == NULL)
 907                 goto err;
 908
 909         eth_dev->data->dev_private = internals;
 910         eth_dev->data->dev_link = pmd_link;
 911         eth_dev->data->mac_addrs = &internals->eth_addr;
 912         eth_dev->dev_ops = &ops;
 913         eth_dev->rx_pkt_burst = eth_af_xdp_rx;
 914         eth_dev->tx_pkt_burst = eth_af_xdp_tx;
 915         /* Let rte_eth_dev_close() release the port resources. */
 916         eth_dev->data->dev_flags |= RTE_ETH_DEV_CLOSE_REMOVE;
 917
 918         if (internals->pmd_zc)
 919                 AF_XDP_LOG(INFO, "Zero copy between umem and mbuf enabled.\n");
 920
 921         return eth_dev;
 922
 923 err:
 924         rte_free(internals);
 925         return NULL;
 926 }
 927
 928 static int
 929 rte_pmd_af_xdp_probe(struct rte_vdev_device *dev)
 930 {
 931         struct rte_kvargs *kvlist;
 932         char if_name[IFNAMSIZ] = {'\0'};
 933         int xsk_queue_idx = ETH_AF_XDP_DFLT_QUEUE_IDX;
 934         struct rte_eth_dev *eth_dev = NULL;
 935         const char *name;
 936         int pmd_zc = 0;
 937
 938         AF_XDP_LOG(INFO, "Initializing pmd_af_xdp for %s\n",
 939                 rte_vdev_device_name(dev));
 940
 941         name = rte_vdev_device_name(dev);
 942         if (rte_eal_process_type() == RTE_PROC_SECONDARY &&
 943                 strlen(rte_vdev_device_args(dev)) == 0) {
 944                 eth_dev = rte_eth_dev_attach_secondary(name);
 945                 if (eth_dev == NULL) {
 946                         AF_XDP_LOG(ERR, "Failed to probe %s\n", name);
 947                         return -EINVAL;
 948                 }
 949                 eth_dev->dev_ops = &ops;
 950                 rte_eth_dev_probing_finish(eth_dev);
 951                 return 0;
 952         }
 953
 954         kvlist = rte_kvargs_parse(rte_vdev_device_args(dev), valid_arguments);
 955         if (kvlist == NULL) {
 956                 AF_XDP_LOG(ERR, "Invalid kvargs key\n");
 957                 return -EINVAL;
 958         }
 959
 960         if (dev->device.numa_node == SOCKET_ID_ANY)
 961                 dev->device.numa_node = rte_socket_id();
 962
 963         if (parse_parameters(kvlist, if_name, &xsk_queue_idx, &pmd_zc) < 0) {
 964                 AF_XDP_LOG(ERR, "Invalid kvargs value\n");
 965                 return -EINVAL;
 966         }
 967
 968         if (strlen(if_name) == 0) {
 969                 AF_XDP_LOG(ERR, "Network interface must be specified\n");
 970                 return -EINVAL;
 971         }
 972
 973         eth_dev = init_internals(dev, if_name, xsk_queue_idx, pmd_zc);
 974         if (eth_dev == NULL) {
 975                 AF_XDP_LOG(ERR, "Failed to init internals\n");
 976                 return -1;
 977         }
 978
 979         rte_eth_dev_probing_finish(eth_dev);
 980
 981         return 0;
 982 }
 983
 984 static int
 985 rte_pmd_af_xdp_remove(struct rte_vdev_device *dev)
 986 {
 987         struct rte_eth_dev *eth_dev = NULL;
 988
 989         AF_XDP_LOG(INFO, "Removing AF_XDP ethdev on numa socket %u\n",
 990                 rte_socket_id());
 991
 992         if (dev == NULL)
 993                 return -1;
 994
 995         /* find the ethdev entry */
 996         eth_dev = rte_eth_dev_allocated(rte_vdev_device_name(dev));
 997         if (eth_dev == NULL)
 998                 return 0;
 999
1000         eth_dev_close(eth_dev);
1001         rte_eth_dev_release_port(eth_dev);
1002
1003
1004         return 0;
1005 }
1006
1007 static struct rte_vdev_driver pmd_af_xdp_drv = {
1008         .probe = rte_pmd_af_xdp_probe,
1009         .remove = rte_pmd_af_xdp_remove,
1010 };
1011
1012 RTE_PMD_REGISTER_VDEV(net_af_xdp, pmd_af_xdp_drv);
1013 RTE_PMD_REGISTER_PARAM_STRING(net_af_xdp,
1014                               "iface=<string> "
1015                               "queue=<int> "
1016                               "pmd_zero_copy=<0|1>");
1017
1018 RTE_INIT(af_xdp_init_log)
1019 {
1020         af_xdp_logtype = rte_log_register("pmd.net.af_xdp");
1021         if (af_xdp_logtype >= 0)
1022                 rte_log_set_level(af_xdp_logtype, RTE_LOG_NOTICE);
1023 }