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