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 <linux/ethtool.h>
  16 #include <linux/sockios.h>
  17 #include "af_xdp_deps.h"
  18 #include <bpf/xsk.h>
  19
  20 #include <rte_ethdev.h>
  21 #include <rte_ethdev_driver.h>
  22 #include <rte_ethdev_vdev.h>
  23 #include <rte_kvargs.h>
  24 #include <rte_bus_vdev.h>
  25 #include <rte_string_fns.h>
  26 #include <rte_branch_prediction.h>
  27 #include <rte_common.h>
  28 #include <rte_config.h>
  29 #include <rte_dev.h>
  30 #include <rte_eal.h>
  31 #include <rte_ether.h>
  32 #include <rte_lcore.h>
  33 #include <rte_log.h>
  34 #include <rte_memory.h>
  35 #include <rte_memzone.h>
  36 #include <rte_mbuf.h>
  37 #include <rte_malloc.h>
  38 #include <rte_ring.h>
  39
  40 #ifndef SOL_XDP
  41 #define SOL_XDP 283
  42 #endif
  43
  44 #ifndef AF_XDP
  45 #define AF_XDP 44
  46 #endif
  47
  48 #ifndef PF_XDP
  49 #define PF_XDP AF_XDP
  50 #endif
  51
  52 static int af_xdp_logtype;
  53
  54 #define AF_XDP_LOG(level, fmt, args...)                 \
  55         rte_log(RTE_LOG_ ## level, af_xdp_logtype,      \
  56                 "%s(): " fmt, __func__, ##args)
  57
  58 #define ETH_AF_XDP_FRAME_SIZE           XSK_UMEM__DEFAULT_FRAME_SIZE
  59 #define ETH_AF_XDP_NUM_BUFFERS          4096
  60 #define ETH_AF_XDP_DATA_HEADROOM        0
  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         int pmd_zc;
  76 };
  77
  78 struct rx_stats {
  79         uint64_t rx_pkts;
  80         uint64_t rx_bytes;
  81         uint64_t rx_dropped;
  82 };
  83
  84 struct pkt_rx_queue {
  85         struct xsk_ring_cons rx;
  86         struct xsk_umem_info *umem;
  87         struct xsk_socket *xsk;
  88         struct rte_mempool *mb_pool;
  89
  90         struct rx_stats stats;
  91
  92         struct pkt_tx_queue *pair;
  93         int xsk_queue_idx;
  94 };
  95
  96 struct tx_stats {
  97         uint64_t tx_pkts;
  98         uint64_t err_pkts;
  99         uint64_t tx_bytes;
 100 };
 101
 102 struct pkt_tx_queue {
 103         struct xsk_ring_prod tx;
 104
 105         struct tx_stats stats;
 106
 107         struct pkt_rx_queue *pair;
 108         int xsk_queue_idx;
 109 };
 110
 111 struct pmd_internals {
 112         int if_index;
 113         char if_name[IFNAMSIZ];
 114         int start_queue_idx;
 115         int queue_cnt;
 116         int max_queue_cnt;
 117         int combined_queue_cnt;
 118
 119         int pmd_zc;
 120         struct rte_ether_addr eth_addr;
 121         struct rte_mempool *mb_pool_share;
 122
 123         struct pkt_rx_queue *rx_queues;
 124         struct pkt_tx_queue *tx_queues;
 125 };
 126
 127 #define ETH_AF_XDP_IFACE_ARG                    "iface"
 128 #define ETH_AF_XDP_START_QUEUE_ARG              "start_queue"
 129 #define ETH_AF_XDP_QUEUE_COUNT_ARG              "queue_count"
 130 #define ETH_AF_XDP_PMD_ZC_ARG                   "pmd_zero_copy"
 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         ETH_AF_XDP_PMD_ZC_ARG,
 137         NULL
 138 };
 139
 140 static const struct rte_eth_link pmd_link = {
 141         .link_speed = ETH_SPEED_NUM_10G,
 142         .link_duplex = ETH_LINK_FULL_DUPLEX,
 143         .link_status = ETH_LINK_DOWN,
 144         .link_autoneg = ETH_LINK_AUTONEG
 145 };
 146
 147 static inline int
 148 reserve_fill_queue(struct xsk_umem_info *umem, uint16_t reserve_size)
 149 {
 150         struct xsk_ring_prod *fq = &umem->fq;
 151         void *addrs[reserve_size];
 152         uint32_t idx;
 153         uint16_t i;
 154
 155         if (rte_ring_dequeue_bulk(umem->buf_ring, addrs, reserve_size, NULL)
 156                     != reserve_size) {
 157                 AF_XDP_LOG(DEBUG, "Failed to get enough buffers for fq.\n");
 158                 return -1;
 159         }
 160
 161         if (unlikely(!xsk_ring_prod__reserve(fq, reserve_size, &idx))) {
 162                 AF_XDP_LOG(DEBUG, "Failed to reserve enough fq descs.\n");
 163                 rte_ring_enqueue_bulk(umem->buf_ring, addrs,
 164                                 reserve_size, NULL);
 165                 return -1;
 166         }
 167
 168         for (i = 0; i < reserve_size; i++) {
 169                 __u64 *fq_addr;
 170
 171                 fq_addr = xsk_ring_prod__fill_addr(fq, idx++);
 172                 *fq_addr = (uint64_t)addrs[i];
 173         }
 174
 175         xsk_ring_prod__submit(fq, reserve_size);
 176
 177         return 0;
 178 }
 179
 180 static void
 181 umem_buf_release_to_fq(void *addr, void *opaque)
 182 {
 183         struct xsk_umem_info *umem = (struct xsk_umem_info *)opaque;
 184         uint64_t umem_addr = (uint64_t)addr - umem->mz->addr_64;
 185
 186         rte_ring_enqueue(umem->buf_ring, (void *)umem_addr);
 187 }
 188
 189 static uint16_t
 190 eth_af_xdp_rx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 191 {
 192         struct pkt_rx_queue *rxq = queue;
 193         struct xsk_ring_cons *rx = &rxq->rx;
 194         struct xsk_umem_info *umem = rxq->umem;
 195         struct xsk_ring_prod *fq = &umem->fq;
 196         uint32_t idx_rx = 0;
 197         uint32_t free_thresh = fq->size >> 1;
 198         int pmd_zc = umem->pmd_zc;
 199         struct rte_mbuf *mbufs[ETH_AF_XDP_RX_BATCH_SIZE];
 200         unsigned long dropped = 0;
 201         unsigned long rx_bytes = 0;
 202         int rcvd, i;
 203
 204         nb_pkts = RTE_MIN(nb_pkts, ETH_AF_XDP_RX_BATCH_SIZE);
 205
 206         if (unlikely(rte_pktmbuf_alloc_bulk(rxq->mb_pool, mbufs, nb_pkts) != 0))
 207                 return 0;
 208
 209         rcvd = xsk_ring_cons__peek(rx, nb_pkts, &idx_rx);
 210         if (rcvd == 0)
 211                 goto out;
 212
 213         if (xsk_prod_nb_free(fq, free_thresh) >= free_thresh)
 214                 (void)reserve_fill_queue(umem, ETH_AF_XDP_RX_BATCH_SIZE);
 215
 216         for (i = 0; i < rcvd; i++) {
 217                 const struct xdp_desc *desc;
 218                 uint64_t addr;
 219                 uint32_t len;
 220                 void *pkt;
 221                 uint16_t buf_len = ETH_AF_XDP_FRAME_SIZE;
 222                 struct rte_mbuf_ext_shared_info *shinfo;
 223
 224                 desc = xsk_ring_cons__rx_desc(rx, idx_rx++);
 225                 addr = desc->addr;
 226                 len = desc->len;
 227                 pkt = xsk_umem__get_data(rxq->umem->mz->addr, addr);
 228
 229                 if (pmd_zc) {
 230                         shinfo = rte_pktmbuf_ext_shinfo_init_helper(pkt,
 231                                         &buf_len, umem_buf_release_to_fq, umem);
 232
 233                         rte_pktmbuf_attach_extbuf(mbufs[i], pkt, 0, buf_len,
 234                                                   shinfo);
 235                 } else {
 236                         rte_memcpy(rte_pktmbuf_mtod(mbufs[i], void *),
 237                                                         pkt, len);
 238                         rte_ring_enqueue(umem->buf_ring, (void *)addr);
 239                 }
 240                 rte_pktmbuf_pkt_len(mbufs[i]) = len;
 241                 rte_pktmbuf_data_len(mbufs[i]) = len;
 242                 rx_bytes += len;
 243                 bufs[i] = mbufs[i];
 244         }
 245
 246         xsk_ring_cons__release(rx, rcvd);
 247
 248         /* statistics */
 249         rxq->stats.rx_pkts += (rcvd - dropped);
 250         rxq->stats.rx_bytes += rx_bytes;
 251
 252 out:
 253         if (rcvd != nb_pkts)
 254                 rte_mempool_put_bulk(rxq->mb_pool, (void **)&mbufs[rcvd],
 255                                      nb_pkts - rcvd);
 256
 257         return rcvd;
 258 }
 259
 260 static void
 261 pull_umem_cq(struct xsk_umem_info *umem, int size)
 262 {
 263         struct xsk_ring_cons *cq = &umem->cq;
 264         size_t i, n;
 265         uint32_t idx_cq = 0;
 266
 267         n = xsk_ring_cons__peek(cq, size, &idx_cq);
 268
 269         for (i = 0; i < n; i++) {
 270                 uint64_t addr;
 271                 addr = *xsk_ring_cons__comp_addr(cq, idx_cq++);
 272                 rte_ring_enqueue(umem->buf_ring, (void *)addr);
 273         }
 274
 275         xsk_ring_cons__release(cq, n);
 276 }
 277
 278 static void
 279 kick_tx(struct pkt_tx_queue *txq)
 280 {
 281         struct xsk_umem_info *umem = txq->pair->umem;
 282
 283         while (send(xsk_socket__fd(txq->pair->xsk), NULL,
 284                       0, MSG_DONTWAIT) < 0) {
 285                 /* some thing unexpected */
 286                 if (errno != EBUSY && errno != EAGAIN && errno != EINTR)
 287                         break;
 288
 289                 /* pull from completion queue to leave more space */
 290                 if (errno == EAGAIN)
 291                         pull_umem_cq(umem, ETH_AF_XDP_TX_BATCH_SIZE);
 292         }
 293         pull_umem_cq(umem, ETH_AF_XDP_TX_BATCH_SIZE);
 294 }
 295
 296 static inline bool
 297 in_umem_range(struct xsk_umem_info *umem, uint64_t addr)
 298 {
 299         uint64_t mz_base_addr = umem->mz->addr_64;
 300
 301         return addr >= mz_base_addr && addr < mz_base_addr + umem->mz->len;
 302 }
 303
 304 static uint16_t
 305 eth_af_xdp_tx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 306 {
 307         struct pkt_tx_queue *txq = queue;
 308         struct xsk_umem_info *umem = txq->pair->umem;
 309         struct rte_mbuf *mbuf;
 310         int pmd_zc = umem->pmd_zc;
 311         void *addrs[ETH_AF_XDP_TX_BATCH_SIZE];
 312         unsigned long tx_bytes = 0;
 313         int i;
 314         uint32_t idx_tx;
 315
 316         nb_pkts = RTE_MIN(nb_pkts, ETH_AF_XDP_TX_BATCH_SIZE);
 317
 318         pull_umem_cq(umem, nb_pkts);
 319
 320         nb_pkts = rte_ring_dequeue_bulk(umem->buf_ring, addrs,
 321                                         nb_pkts, NULL);
 322         if (nb_pkts == 0)
 323                 return 0;
 324
 325         if (xsk_ring_prod__reserve(&txq->tx, nb_pkts, &idx_tx) != nb_pkts) {
 326                 kick_tx(txq);
 327                 rte_ring_enqueue_bulk(umem->buf_ring, addrs, nb_pkts, NULL);
 328                 return 0;
 329         }
 330
 331         for (i = 0; i < nb_pkts; i++) {
 332                 struct xdp_desc *desc;
 333                 void *pkt;
 334
 335                 desc = xsk_ring_prod__tx_desc(&txq->tx, idx_tx + i);
 336                 mbuf = bufs[i];
 337                 desc->len = mbuf->pkt_len;
 338
 339                 /*
 340                  * We need to make sure the external mbuf address is within
 341                  * current port's umem memzone range
 342                  */
 343                 if (pmd_zc && RTE_MBUF_HAS_EXTBUF(mbuf) &&
 344                                 in_umem_range(umem, (uint64_t)mbuf->buf_addr)) {
 345                         desc->addr = (uint64_t)mbuf->buf_addr -
 346                                 umem->mz->addr_64;
 347                         mbuf->buf_addr = xsk_umem__get_data(umem->mz->addr,
 348                                         (uint64_t)addrs[i]);
 349                 } else {
 350                         desc->addr = (uint64_t)addrs[i];
 351                         pkt = xsk_umem__get_data(umem->mz->addr,
 352                                         desc->addr);
 353                         rte_memcpy(pkt, rte_pktmbuf_mtod(mbuf, void *),
 354                                         desc->len);
 355                 }
 356                 tx_bytes += mbuf->pkt_len;
 357         }
 358
 359         xsk_ring_prod__submit(&txq->tx, nb_pkts);
 360
 361         kick_tx(txq);
 362
 363         txq->stats.tx_pkts += nb_pkts;
 364         txq->stats.tx_bytes += tx_bytes;
 365
 366         for (i = 0; i < nb_pkts; i++)
 367                 rte_pktmbuf_free(bufs[i]);
 368
 369         return nb_pkts;
 370 }
 371
 372 static int
 373 eth_dev_start(struct rte_eth_dev *dev)
 374 {
 375         dev->data->dev_link.link_status = ETH_LINK_UP;
 376
 377         return 0;
 378 }
 379
 380 /* This function gets called when the current port gets stopped. */
 381 static void
 382 eth_dev_stop(struct rte_eth_dev *dev)
 383 {
 384         dev->data->dev_link.link_status = ETH_LINK_DOWN;
 385 }
 386
 387 static int
 388 eth_dev_configure(struct rte_eth_dev *dev)
 389 {
 390         /* rx/tx must be paired */
 391         if (dev->data->nb_rx_queues != dev->data->nb_tx_queues)
 392                 return -EINVAL;
 393
 394         return 0;
 395 }
 396
 397 static void
 398 eth_dev_info(struct rte_eth_dev *dev, struct rte_eth_dev_info *dev_info)
 399 {
 400         struct pmd_internals *internals = dev->data->dev_private;
 401
 402         dev_info->if_index = internals->if_index;
 403         dev_info->max_mac_addrs = 1;
 404         dev_info->max_rx_pktlen = ETH_FRAME_LEN;
 405         dev_info->max_rx_queues = internals->queue_cnt;
 406         dev_info->max_tx_queues = internals->queue_cnt;
 407
 408         dev_info->min_mtu = RTE_ETHER_MIN_MTU;
 409         dev_info->max_mtu = ETH_AF_XDP_FRAME_SIZE - ETH_AF_XDP_DATA_HEADROOM;
 410
 411         dev_info->default_rxportconf.nb_queues = 1;
 412         dev_info->default_txportconf.nb_queues = 1;
 413         dev_info->default_rxportconf.ring_size = ETH_AF_XDP_DFLT_NUM_DESCS;
 414         dev_info->default_txportconf.ring_size = ETH_AF_XDP_DFLT_NUM_DESCS;
 415 }
 416
 417 static int
 418 eth_stats_get(struct rte_eth_dev *dev, struct rte_eth_stats *stats)
 419 {
 420         struct pmd_internals *internals = dev->data->dev_private;
 421         struct xdp_statistics xdp_stats;
 422         struct pkt_rx_queue *rxq;
 423         struct pkt_tx_queue *txq;
 424         socklen_t optlen;
 425         int i, ret;
 426
 427         for (i = 0; i < dev->data->nb_rx_queues; i++) {
 428                 optlen = sizeof(struct xdp_statistics);
 429                 rxq = &internals->rx_queues[i];
 430                 txq = rxq->pair;
 431                 stats->q_ipackets[i] = rxq->stats.rx_pkts;
 432                 stats->q_ibytes[i] = rxq->stats.rx_bytes;
 433
 434                 stats->q_opackets[i] = txq->stats.tx_pkts;
 435                 stats->q_obytes[i] = txq->stats.tx_bytes;
 436
 437                 stats->ipackets += stats->q_ipackets[i];
 438                 stats->ibytes += stats->q_ibytes[i];
 439                 stats->imissed += rxq->stats.rx_dropped;
 440                 ret = getsockopt(xsk_socket__fd(rxq->xsk), SOL_XDP,
 441                                 XDP_STATISTICS, &xdp_stats, &optlen);
 442                 if (ret != 0) {
 443                         AF_XDP_LOG(ERR, "getsockopt() failed for XDP_STATISTICS.\n");
 444                         return -1;
 445                 }
 446                 stats->imissed += xdp_stats.rx_dropped;
 447
 448                 stats->opackets += stats->q_opackets[i];
 449                 stats->oerrors += txq->stats.err_pkts;
 450                 stats->obytes += stats->q_obytes[i];
 451         }
 452
 453         return 0;
 454 }
 455
 456 static void
 457 eth_stats_reset(struct rte_eth_dev *dev)
 458 {
 459         struct pmd_internals *internals = dev->data->dev_private;
 460         int i;
 461
 462         for (i = 0; i < internals->queue_cnt; i++) {
 463                 memset(&internals->rx_queues[i].stats, 0,
 464                                         sizeof(struct rx_stats));
 465                 memset(&internals->tx_queues[i].stats, 0,
 466                                         sizeof(struct tx_stats));
 467         }
 468 }
 469
 470 static void
 471 remove_xdp_program(struct pmd_internals *internals)
 472 {
 473         uint32_t curr_prog_id = 0;
 474
 475         if (bpf_get_link_xdp_id(internals->if_index, &curr_prog_id,
 476                                 XDP_FLAGS_UPDATE_IF_NOEXIST)) {
 477                 AF_XDP_LOG(ERR, "bpf_get_link_xdp_id failed\n");
 478                 return;
 479         }
 480         bpf_set_link_xdp_fd(internals->if_index, -1,
 481                         XDP_FLAGS_UPDATE_IF_NOEXIST);
 482 }
 483
 484 static void
 485 xdp_umem_destroy(struct xsk_umem_info *umem)
 486 {
 487         rte_memzone_free(umem->mz);
 488         umem->mz = NULL;
 489
 490         rte_ring_free(umem->buf_ring);
 491         umem->buf_ring = NULL;
 492
 493         rte_free(umem);
 494         umem = NULL;
 495 }
 496
 497 static void
 498 eth_dev_close(struct rte_eth_dev *dev)
 499 {
 500         struct pmd_internals *internals = dev->data->dev_private;
 501         struct pkt_rx_queue *rxq;
 502         int i;
 503
 504         AF_XDP_LOG(INFO, "Closing AF_XDP ethdev on numa socket %u\n",
 505                 rte_socket_id());
 506
 507         for (i = 0; i < internals->queue_cnt; i++) {
 508                 rxq = &internals->rx_queues[i];
 509                 if (rxq->umem == NULL)
 510                         break;
 511                 xsk_socket__delete(rxq->xsk);
 512                 (void)xsk_umem__delete(rxq->umem->umem);
 513                 xdp_umem_destroy(rxq->umem);
 514
 515                 /* free pkt_tx_queue */
 516                 rte_free(rxq->pair);
 517                 rte_free(rxq);
 518         }
 519
 520         /*
 521          * MAC is not allocated dynamically, setting it to NULL would prevent
 522          * from releasing it in rte_eth_dev_release_port.
 523          */
 524         dev->data->mac_addrs = NULL;
 525
 526         remove_xdp_program(internals);
 527 }
 528
 529 static void
 530 eth_queue_release(void *q __rte_unused)
 531 {
 532 }
 533
 534 static int
 535 eth_link_update(struct rte_eth_dev *dev __rte_unused,
 536                 int wait_to_complete __rte_unused)
 537 {
 538         return 0;
 539 }
 540
 541 static struct
 542 xsk_umem_info *xdp_umem_configure(struct pmd_internals *internals,
 543                                   struct pkt_rx_queue *rxq)
 544 {
 545         struct xsk_umem_info *umem;
 546         const struct rte_memzone *mz;
 547         struct xsk_umem_config usr_config = {
 548                 .fill_size = ETH_AF_XDP_DFLT_NUM_DESCS,
 549                 .comp_size = ETH_AF_XDP_DFLT_NUM_DESCS,
 550                 .frame_size = ETH_AF_XDP_FRAME_SIZE,
 551                 .frame_headroom = ETH_AF_XDP_DATA_HEADROOM };
 552         char ring_name[RTE_RING_NAMESIZE];
 553         char mz_name[RTE_MEMZONE_NAMESIZE];
 554         int ret;
 555         uint64_t i;
 556
 557         umem = rte_zmalloc_socket("umem", sizeof(*umem), 0, rte_socket_id());
 558         if (umem == NULL) {
 559                 AF_XDP_LOG(ERR, "Failed to allocate umem info");
 560                 return NULL;
 561         }
 562
 563         snprintf(ring_name, sizeof(ring_name), "af_xdp_ring_%s_%u",
 564                        internals->if_name, rxq->xsk_queue_idx);
 565         umem->buf_ring = rte_ring_create(ring_name,
 566                                          ETH_AF_XDP_NUM_BUFFERS,
 567                                          rte_socket_id(),
 568                                          0x0);
 569         if (umem->buf_ring == NULL) {
 570                 AF_XDP_LOG(ERR, "Failed to create rte_ring\n");
 571                 goto err;
 572         }
 573
 574         for (i = 0; i < ETH_AF_XDP_NUM_BUFFERS; i++)
 575                 rte_ring_enqueue(umem->buf_ring,
 576                                  (void *)(i * ETH_AF_XDP_FRAME_SIZE +
 577                                           ETH_AF_XDP_DATA_HEADROOM));
 578
 579         snprintf(mz_name, sizeof(mz_name), "af_xdp_umem_%s_%u",
 580                        internals->if_name, rxq->xsk_queue_idx);
 581         mz = rte_memzone_reserve_aligned(mz_name,
 582                         ETH_AF_XDP_NUM_BUFFERS * ETH_AF_XDP_FRAME_SIZE,
 583                         rte_socket_id(), RTE_MEMZONE_IOVA_CONTIG,
 584                         getpagesize());
 585         if (mz == NULL) {
 586                 AF_XDP_LOG(ERR, "Failed to reserve memzone for af_xdp umem.\n");
 587                 goto err;
 588         }
 589
 590         ret = xsk_umem__create(&umem->umem, mz->addr,
 591                                ETH_AF_XDP_NUM_BUFFERS * ETH_AF_XDP_FRAME_SIZE,
 592                                &umem->fq, &umem->cq,
 593                                &usr_config);
 594
 595         if (ret) {
 596                 AF_XDP_LOG(ERR, "Failed to create umem");
 597                 goto err;
 598         }
 599         umem->mz = mz;
 600
 601         return umem;
 602
 603 err:
 604         xdp_umem_destroy(umem);
 605         return NULL;
 606 }
 607
 608 static int
 609 xsk_configure(struct pmd_internals *internals, struct pkt_rx_queue *rxq,
 610               int ring_size)
 611 {
 612         struct xsk_socket_config cfg;
 613         struct pkt_tx_queue *txq = rxq->pair;
 614         int ret = 0;
 615         int reserve_size;
 616
 617         rxq->umem = xdp_umem_configure(internals, rxq);
 618         if (rxq->umem == NULL)
 619                 return -ENOMEM;
 620
 621         cfg.rx_size = ring_size;
 622         cfg.tx_size = ring_size;
 623         cfg.libbpf_flags = 0;
 624         cfg.xdp_flags = XDP_FLAGS_UPDATE_IF_NOEXIST;
 625         cfg.bind_flags = 0;
 626         ret = xsk_socket__create(&rxq->xsk, internals->if_name,
 627                         rxq->xsk_queue_idx, rxq->umem->umem, &rxq->rx,
 628                         &txq->tx, &cfg);
 629         if (ret) {
 630                 AF_XDP_LOG(ERR, "Failed to create xsk socket.\n");
 631                 goto err;
 632         }
 633
 634         reserve_size = ETH_AF_XDP_DFLT_NUM_DESCS / 2;
 635         ret = reserve_fill_queue(rxq->umem, reserve_size);
 636         if (ret) {
 637                 xsk_socket__delete(rxq->xsk);
 638                 AF_XDP_LOG(ERR, "Failed to reserve fill queue.\n");
 639                 goto err;
 640         }
 641
 642         return 0;
 643
 644 err:
 645         xdp_umem_destroy(rxq->umem);
 646
 647         return ret;
 648 }
 649
 650 static int
 651 eth_rx_queue_setup(struct rte_eth_dev *dev,
 652                    uint16_t rx_queue_id,
 653                    uint16_t nb_rx_desc,
 654                    unsigned int socket_id __rte_unused,
 655                    const struct rte_eth_rxconf *rx_conf __rte_unused,
 656                    struct rte_mempool *mb_pool)
 657 {
 658         struct pmd_internals *internals = dev->data->dev_private;
 659         uint32_t buf_size, data_size;
 660         struct pkt_rx_queue *rxq;
 661         int ret;
 662
 663         rxq = &internals->rx_queues[rx_queue_id];
 664
 665         AF_XDP_LOG(INFO, "Set up rx queue, rx queue id: %d, xsk queue id: %d\n",
 666                    rx_queue_id, rxq->xsk_queue_idx);
 667         /* Now get the space available for data in the mbuf */
 668         buf_size = rte_pktmbuf_data_room_size(mb_pool) -
 669                 RTE_PKTMBUF_HEADROOM;
 670         data_size = ETH_AF_XDP_FRAME_SIZE - ETH_AF_XDP_DATA_HEADROOM;
 671
 672         if (data_size > buf_size) {
 673                 AF_XDP_LOG(ERR, "%s: %d bytes will not fit in mbuf (%d bytes)\n",
 674                         dev->device->name, data_size, buf_size);
 675                 ret = -ENOMEM;
 676                 goto err;
 677         }
 678
 679         rxq->mb_pool = mb_pool;
 680
 681         if (xsk_configure(internals, rxq, nb_rx_desc)) {
 682                 AF_XDP_LOG(ERR, "Failed to configure xdp socket\n");
 683                 ret = -EINVAL;
 684                 goto err;
 685         }
 686
 687         rxq->umem->pmd_zc = internals->pmd_zc;
 688
 689         dev->data->rx_queues[rx_queue_id] = rxq;
 690         return 0;
 691
 692 err:
 693         return ret;
 694 }
 695
 696 static int
 697 eth_tx_queue_setup(struct rte_eth_dev *dev,
 698                    uint16_t tx_queue_id,
 699                    uint16_t nb_tx_desc __rte_unused,
 700                    unsigned int socket_id __rte_unused,
 701                    const struct rte_eth_txconf *tx_conf __rte_unused)
 702 {
 703         struct pmd_internals *internals = dev->data->dev_private;
 704         struct pkt_tx_queue *txq;
 705
 706         txq = &internals->tx_queues[tx_queue_id];
 707
 708         dev->data->tx_queues[tx_queue_id] = txq;
 709         return 0;
 710 }
 711
 712 static int
 713 eth_dev_mtu_set(struct rte_eth_dev *dev, uint16_t mtu)
 714 {
 715         struct pmd_internals *internals = dev->data->dev_private;
 716         struct ifreq ifr = { .ifr_mtu = mtu };
 717         int ret;
 718         int s;
 719
 720         s = socket(PF_INET, SOCK_DGRAM, 0);
 721         if (s < 0)
 722                 return -EINVAL;
 723
 724         strlcpy(ifr.ifr_name, internals->if_name, IFNAMSIZ);
 725         ret = ioctl(s, SIOCSIFMTU, &ifr);
 726         close(s);
 727
 728         return (ret < 0) ? -errno : 0;
 729 }
 730
 731 static void
 732 eth_dev_change_flags(char *if_name, uint32_t flags, uint32_t mask)
 733 {
 734         struct ifreq ifr;
 735         int s;
 736
 737         s = socket(PF_INET, SOCK_DGRAM, 0);
 738         if (s < 0)
 739                 return;
 740
 741         strlcpy(ifr.ifr_name, if_name, IFNAMSIZ);
 742         if (ioctl(s, SIOCGIFFLAGS, &ifr) < 0)
 743                 goto out;
 744         ifr.ifr_flags &= mask;
 745         ifr.ifr_flags |= flags;
 746         if (ioctl(s, SIOCSIFFLAGS, &ifr) < 0)
 747                 goto out;
 748 out:
 749         close(s);
 750 }
 751
 752 static void
 753 eth_dev_promiscuous_enable(struct rte_eth_dev *dev)
 754 {
 755         struct pmd_internals *internals = dev->data->dev_private;
 756
 757         eth_dev_change_flags(internals->if_name, IFF_PROMISC, ~0);
 758 }
 759
 760 static void
 761 eth_dev_promiscuous_disable(struct rte_eth_dev *dev)
 762 {
 763         struct pmd_internals *internals = dev->data->dev_private;
 764
 765         eth_dev_change_flags(internals->if_name, 0, ~IFF_PROMISC);
 766 }
 767
 768 static const struct eth_dev_ops ops = {
 769         .dev_start = eth_dev_start,
 770         .dev_stop = eth_dev_stop,
 771         .dev_close = eth_dev_close,
 772         .dev_configure = eth_dev_configure,
 773         .dev_infos_get = eth_dev_info,
 774         .mtu_set = eth_dev_mtu_set,
 775         .promiscuous_enable = eth_dev_promiscuous_enable,
 776         .promiscuous_disable = eth_dev_promiscuous_disable,
 777         .rx_queue_setup = eth_rx_queue_setup,
 778         .tx_queue_setup = eth_tx_queue_setup,
 779         .rx_queue_release = eth_queue_release,
 780         .tx_queue_release = eth_queue_release,
 781         .link_update = eth_link_update,
 782         .stats_get = eth_stats_get,
 783         .stats_reset = eth_stats_reset,
 784 };
 785
 786 /** parse integer from integer argument */
 787 static int
 788 parse_integer_arg(const char *key __rte_unused,
 789                   const char *value, void *extra_args)
 790 {
 791         int *i = (int *)extra_args;
 792         char *end;
 793
 794         *i = strtol(value, &end, 10);
 795         if (*i < 0) {
 796                 AF_XDP_LOG(ERR, "Argument has to be positive.\n");
 797                 return -EINVAL;
 798         }
 799
 800         return 0;
 801 }
 802
 803 /** parse name argument */
 804 static int
 805 parse_name_arg(const char *key __rte_unused,
 806                const char *value, void *extra_args)
 807 {
 808         char *name = extra_args;
 809
 810         if (strnlen(value, IFNAMSIZ) > IFNAMSIZ - 1) {
 811                 AF_XDP_LOG(ERR, "Invalid name %s, should be less than %u bytes.\n",
 812                            value, IFNAMSIZ);
 813                 return -EINVAL;
 814         }
 815
 816         strlcpy(name, value, IFNAMSIZ);
 817
 818         return 0;
 819 }
 820
 821 static int
 822 xdp_get_channels_info(const char *if_name, int *max_queues,
 823                                 int *combined_queues)
 824 {
 825         struct ethtool_channels channels;
 826         struct ifreq ifr;
 827         int fd, ret;
 828
 829         fd = socket(AF_INET, SOCK_DGRAM, 0);
 830         if (fd < 0)
 831                 return -1;
 832
 833         channels.cmd = ETHTOOL_GCHANNELS;
 834         ifr.ifr_data = (void *)&channels;
 835         strncpy(ifr.ifr_name, if_name, IFNAMSIZ);
 836         ret = ioctl(fd, SIOCETHTOOL, &ifr);
 837         if (ret && errno != EOPNOTSUPP) {
 838                 ret = -errno;
 839                 goto out;
 840         }
 841
 842         if (channels.max_combined == 0 || errno == EOPNOTSUPP) {
 843                 /* If the device says it has no channels, then all traffic
 844                  * is sent to a single stream, so max queues = 1.
 845                  */
 846                 *max_queues = 1;
 847                 *combined_queues = 1;
 848         } else {
 849                 *max_queues = channels.max_combined;
 850                 *combined_queues = channels.combined_count;
 851         }
 852
 853  out:
 854         close(fd);
 855         return ret;
 856 }
 857
 858 static int
 859 parse_parameters(struct rte_kvargs *kvlist, char *if_name, int *start_queue,
 860                         int *queue_cnt, int *pmd_zc)
 861 {
 862         int ret;
 863
 864         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_IFACE_ARG,
 865                                  &parse_name_arg, if_name);
 866         if (ret < 0)
 867                 goto free_kvlist;
 868
 869         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_START_QUEUE_ARG,
 870                                  &parse_integer_arg, start_queue);
 871         if (ret < 0)
 872                 goto free_kvlist;
 873
 874         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_QUEUE_COUNT_ARG,
 875                                  &parse_integer_arg, queue_cnt);
 876         if (ret < 0 || *queue_cnt <= 0) {
 877                 ret = -EINVAL;
 878                 goto free_kvlist;
 879         }
 880
 881         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_PMD_ZC_ARG,
 882                                  &parse_integer_arg, pmd_zc);
 883         if (ret < 0)
 884                 goto free_kvlist;
 885
 886 free_kvlist:
 887         rte_kvargs_free(kvlist);
 888         return ret;
 889 }
 890
 891 static int
 892 get_iface_info(const char *if_name,
 893                struct rte_ether_addr *eth_addr,
 894                int *if_index)
 895 {
 896         struct ifreq ifr;
 897         int sock = socket(AF_INET, SOCK_DGRAM, IPPROTO_IP);
 898
 899         if (sock < 0)
 900                 return -1;
 901
 902         strlcpy(ifr.ifr_name, if_name, IFNAMSIZ);
 903         if (ioctl(sock, SIOCGIFINDEX, &ifr))
 904                 goto error;
 905
 906         *if_index = ifr.ifr_ifindex;
 907
 908         if (ioctl(sock, SIOCGIFHWADDR, &ifr))
 909                 goto error;
 910
 911         rte_memcpy(eth_addr, ifr.ifr_hwaddr.sa_data, RTE_ETHER_ADDR_LEN);
 912
 913         close(sock);
 914         return 0;
 915
 916 error:
 917         close(sock);
 918         return -1;
 919 }
 920
 921 static struct rte_eth_dev *
 922 init_internals(struct rte_vdev_device *dev, const char *if_name,
 923                         int start_queue_idx, int queue_cnt, int pmd_zc)
 924 {
 925         const char *name = rte_vdev_device_name(dev);
 926         const unsigned int numa_node = dev->device.numa_node;
 927         struct pmd_internals *internals;
 928         struct rte_eth_dev *eth_dev;
 929         int ret;
 930         int i;
 931
 932         internals = rte_zmalloc_socket(name, sizeof(*internals), 0, numa_node);
 933         if (internals == NULL)
 934                 return NULL;
 935
 936         internals->start_queue_idx = start_queue_idx;
 937         internals->queue_cnt = queue_cnt;
 938         internals->pmd_zc = pmd_zc;
 939         strlcpy(internals->if_name, if_name, IFNAMSIZ);
 940
 941         if (xdp_get_channels_info(if_name, &internals->max_queue_cnt,
 942                                   &internals->combined_queue_cnt)) {
 943                 AF_XDP_LOG(ERR, "Failed to get channel info of interface: %s\n",
 944                                 if_name);
 945                 goto err_free_internals;
 946         }
 947
 948         if (queue_cnt > internals->combined_queue_cnt) {
 949                 AF_XDP_LOG(ERR, "Specified queue count %d is larger than combined queue count %d.\n",
 950                                 queue_cnt, internals->combined_queue_cnt);
 951                 goto err_free_internals;
 952         }
 953
 954         internals->rx_queues = rte_zmalloc_socket(NULL,
 955                                         sizeof(struct pkt_rx_queue) * queue_cnt,
 956                                         0, numa_node);
 957         if (internals->rx_queues == NULL) {
 958                 AF_XDP_LOG(ERR, "Failed to allocate memory for rx queues.\n");
 959                 goto err_free_internals;
 960         }
 961
 962         internals->tx_queues = rte_zmalloc_socket(NULL,
 963                                         sizeof(struct pkt_tx_queue) * queue_cnt,
 964                                         0, numa_node);
 965         if (internals->tx_queues == NULL) {
 966                 AF_XDP_LOG(ERR, "Failed to allocate memory for tx queues.\n");
 967                 goto err_free_rx;
 968         }
 969         for (i = 0; i < queue_cnt; i++) {
 970                 internals->tx_queues[i].pair = &internals->rx_queues[i];
 971                 internals->rx_queues[i].pair = &internals->tx_queues[i];
 972                 internals->rx_queues[i].xsk_queue_idx = start_queue_idx + i;
 973                 internals->tx_queues[i].xsk_queue_idx = start_queue_idx + i;
 974         }
 975
 976         ret = get_iface_info(if_name, &internals->eth_addr,
 977                              &internals->if_index);
 978         if (ret)
 979                 goto err_free_tx;
 980
 981         eth_dev = rte_eth_vdev_allocate(dev, 0);
 982         if (eth_dev == NULL)
 983                 goto err_free_tx;
 984
 985         eth_dev->data->dev_private = internals;
 986         eth_dev->data->dev_link = pmd_link;
 987         eth_dev->data->mac_addrs = &internals->eth_addr;
 988         eth_dev->dev_ops = &ops;
 989         eth_dev->rx_pkt_burst = eth_af_xdp_rx;
 990         eth_dev->tx_pkt_burst = eth_af_xdp_tx;
 991         /* Let rte_eth_dev_close() release the port resources. */
 992         eth_dev->data->dev_flags |= RTE_ETH_DEV_CLOSE_REMOVE;
 993
 994         if (internals->pmd_zc)
 995                 AF_XDP_LOG(INFO, "Zero copy between umem and mbuf enabled.\n");
 996
 997         return eth_dev;
 998
 999 err_free_tx:
1000         rte_free(internals->tx_queues);
1001 err_free_rx:
1002         rte_free(internals->rx_queues);
1003 err_free_internals:
1004         rte_free(internals);
1005         return NULL;
1006 }
1007
1008 static int
1009 rte_pmd_af_xdp_probe(struct rte_vdev_device *dev)
1010 {
1011         struct rte_kvargs *kvlist;
1012         char if_name[IFNAMSIZ] = {'\0'};
1013         int xsk_start_queue_idx = ETH_AF_XDP_DFLT_START_QUEUE_IDX;
1014         int xsk_queue_cnt = ETH_AF_XDP_DFLT_QUEUE_COUNT;
1015         struct rte_eth_dev *eth_dev = NULL;
1016         const char *name;
1017         int pmd_zc = 0;
1018
1019         AF_XDP_LOG(INFO, "Initializing pmd_af_xdp for %s\n",
1020                 rte_vdev_device_name(dev));
1021
1022         name = rte_vdev_device_name(dev);
1023         if (rte_eal_process_type() == RTE_PROC_SECONDARY &&
1024                 strlen(rte_vdev_device_args(dev)) == 0) {
1025                 eth_dev = rte_eth_dev_attach_secondary(name);
1026                 if (eth_dev == NULL) {
1027                         AF_XDP_LOG(ERR, "Failed to probe %s\n", name);
1028                         return -EINVAL;
1029                 }
1030                 eth_dev->dev_ops = &ops;
1031                 rte_eth_dev_probing_finish(eth_dev);
1032                 return 0;
1033         }
1034
1035         kvlist = rte_kvargs_parse(rte_vdev_device_args(dev), valid_arguments);
1036         if (kvlist == NULL) {
1037                 AF_XDP_LOG(ERR, "Invalid kvargs key\n");
1038                 return -EINVAL;
1039         }
1040
1041         if (dev->device.numa_node == SOCKET_ID_ANY)
1042                 dev->device.numa_node = rte_socket_id();
1043
1044         if (parse_parameters(kvlist, if_name, &xsk_start_queue_idx,
1045                              &xsk_queue_cnt, &pmd_zc) < 0) {
1046                 AF_XDP_LOG(ERR, "Invalid kvargs value\n");
1047                 return -EINVAL;
1048         }
1049
1050         if (strlen(if_name) == 0) {
1051                 AF_XDP_LOG(ERR, "Network interface must be specified\n");
1052                 return -EINVAL;
1053         }
1054
1055         eth_dev = init_internals(dev, if_name, xsk_start_queue_idx,
1056                                         xsk_queue_cnt, pmd_zc);
1057         if (eth_dev == NULL) {
1058                 AF_XDP_LOG(ERR, "Failed to init internals\n");
1059                 return -1;
1060         }
1061
1062         rte_eth_dev_probing_finish(eth_dev);
1063
1064         return 0;
1065 }
1066
1067 static int
1068 rte_pmd_af_xdp_remove(struct rte_vdev_device *dev)
1069 {
1070         struct rte_eth_dev *eth_dev = NULL;
1071
1072         AF_XDP_LOG(INFO, "Removing AF_XDP ethdev on numa socket %u\n",
1073                 rte_socket_id());
1074
1075         if (dev == NULL)
1076                 return -1;
1077
1078         /* find the ethdev entry */
1079         eth_dev = rte_eth_dev_allocated(rte_vdev_device_name(dev));
1080         if (eth_dev == NULL)
1081                 return 0;
1082
1083         eth_dev_close(eth_dev);
1084         rte_eth_dev_release_port(eth_dev);
1085
1086
1087         return 0;
1088 }
1089
1090 static struct rte_vdev_driver pmd_af_xdp_drv = {
1091         .probe = rte_pmd_af_xdp_probe,
1092         .remove = rte_pmd_af_xdp_remove,
1093 };
1094
1095 RTE_PMD_REGISTER_VDEV(net_af_xdp, pmd_af_xdp_drv);
1096 RTE_PMD_REGISTER_PARAM_STRING(net_af_xdp,
1097                               "iface=<string> "
1098                               "start_queue=<int> "
1099                               "queue_count=<int> "
1100                               "pmd_zero_copy=<0|1>");
1101
1102 RTE_INIT(af_xdp_init_log)
1103 {
1104         af_xdp_logtype = rte_log_register("pmd.net.af_xdp");
1105         if (af_xdp_logtype >= 0)
1106                 rte_log_set_level(af_xdp_logtype, RTE_LOG_NOTICE);
1107 }