drivers/net/af_xdp/rte_eth_af_xdp.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2019 Intel Corporation.
   3  */
   4 #include <unistd.h>
   5 #include <errno.h>
   6 #include <stdlib.h>
   7 #include <string.h>
   8 #include <poll.h>
   9 #include <netinet/in.h>
  10 #include <net/if.h>
  11 #include <sys/socket.h>
  12 #include <sys/ioctl.h>
  13 #include <linux/if_ether.h>
  14 #include <linux/if_xdp.h>
  15 #include <linux/if_link.h>
  16 #include <linux/ethtool.h>
  17 #include <linux/sockios.h>
  18 #include "af_xdp_deps.h"
  19 #include <bpf/xsk.h>
  20
  21 #include <rte_ethdev.h>
  22 #include <rte_ethdev_driver.h>
  23 #include <rte_ethdev_vdev.h>
  24 #include <rte_kvargs.h>
  25 #include <rte_bus_vdev.h>
  26 #include <rte_string_fns.h>
  27 #include <rte_branch_prediction.h>
  28 #include <rte_common.h>
  29 #include <rte_config.h>
  30 #include <rte_dev.h>
  31 #include <rte_eal.h>
  32 #include <rte_ether.h>
  33 #include <rte_lcore.h>
  34 #include <rte_log.h>
  35 #include <rte_memory.h>
  36 #include <rte_memzone.h>
  37 #include <rte_mbuf.h>
  38 #include <rte_malloc.h>
  39 #include <rte_ring.h>
  40
  41 #ifndef SOL_XDP
  42 #define SOL_XDP 283
  43 #endif
  44
  45 #ifndef AF_XDP
  46 #define AF_XDP 44
  47 #endif
  48
  49 #ifndef PF_XDP
  50 #define PF_XDP AF_XDP
  51 #endif
  52
  53 static int af_xdp_logtype;
  54
  55 #define AF_XDP_LOG(level, fmt, args...)                 \
  56         rte_log(RTE_LOG_ ## level, af_xdp_logtype,      \
  57                 "%s(): " fmt, __func__, ##args)
  58
  59 #define ETH_AF_XDP_FRAME_SIZE           2048
  60 #define ETH_AF_XDP_NUM_BUFFERS          4096
  61 #define ETH_AF_XDP_DATA_HEADROOM        0
  62 #define ETH_AF_XDP_DFLT_NUM_DESCS       XSK_RING_CONS__DEFAULT_NUM_DESCS
  63 #define ETH_AF_XDP_DFLT_START_QUEUE_IDX 0
  64 #define ETH_AF_XDP_DFLT_QUEUE_COUNT     1
  65
  66 #define ETH_AF_XDP_RX_BATCH_SIZE        32
  67 #define ETH_AF_XDP_TX_BATCH_SIZE        32
  68
  69
  70 struct xsk_umem_info {
  71         struct xsk_ring_prod fq;
  72         struct xsk_ring_cons cq;
  73         struct xsk_umem *umem;
  74         struct rte_ring *buf_ring;
  75         const struct rte_memzone *mz;
  76         int pmd_zc;
  77 };
  78
  79 struct rx_stats {
  80         uint64_t rx_pkts;
  81         uint64_t rx_bytes;
  82         uint64_t rx_dropped;
  83 };
  84
  85 struct pkt_rx_queue {
  86         struct xsk_ring_cons rx;
  87         struct xsk_umem_info *umem;
  88         struct xsk_socket *xsk;
  89         struct rte_mempool *mb_pool;
  90
  91         struct rx_stats stats;
  92
  93         struct pkt_tx_queue *pair;
  94         struct pollfd fds[1];
  95         int xsk_queue_idx;
  96 };
  97
  98 struct tx_stats {
  99         uint64_t tx_pkts;
 100         uint64_t tx_bytes;
 101 };
 102
 103 struct pkt_tx_queue {
 104         struct xsk_ring_prod tx;
 105
 106         struct tx_stats stats;
 107
 108         struct pkt_rx_queue *pair;
 109         int xsk_queue_idx;
 110 };
 111
 112 struct pmd_internals {
 113         int if_index;
 114         char if_name[IFNAMSIZ];
 115         int start_queue_idx;
 116         int queue_cnt;
 117         int max_queue_cnt;
 118         int combined_queue_cnt;
 119
 120         int pmd_zc;
 121         struct rte_ether_addr eth_addr;
 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 #if defined(XDP_USE_NEED_WAKEUP)
 212                 if (xsk_ring_prod__needs_wakeup(fq))
 213                         (void)poll(rxq->fds, 1, 1000);
 214 #endif
 215
 216                 goto out;
 217         }
 218
 219         if (xsk_prod_nb_free(fq, free_thresh) >= free_thresh)
 220                 (void)reserve_fill_queue(umem, ETH_AF_XDP_RX_BATCH_SIZE);
 221
 222         for (i = 0; i < rcvd; i++) {
 223                 const struct xdp_desc *desc;
 224                 uint64_t addr;
 225                 uint32_t len;
 226                 void *pkt;
 227                 uint16_t buf_len = ETH_AF_XDP_FRAME_SIZE;
 228                 struct rte_mbuf_ext_shared_info *shinfo;
 229
 230                 desc = xsk_ring_cons__rx_desc(rx, idx_rx++);
 231                 addr = desc->addr;
 232                 len = desc->len;
 233                 pkt = xsk_umem__get_data(rxq->umem->mz->addr, addr);
 234
 235                 if (pmd_zc) {
 236                         shinfo = rte_pktmbuf_ext_shinfo_init_helper(pkt,
 237                                         &buf_len, umem_buf_release_to_fq, umem);
 238
 239                         rte_pktmbuf_attach_extbuf(mbufs[i], pkt, 0, buf_len,
 240                                                   shinfo);
 241                 } else {
 242                         rte_memcpy(rte_pktmbuf_mtod(mbufs[i], void *),
 243                                                         pkt, len);
 244                         rte_ring_enqueue(umem->buf_ring, (void *)addr);
 245                 }
 246                 rte_pktmbuf_pkt_len(mbufs[i]) = len;
 247                 rte_pktmbuf_data_len(mbufs[i]) = len;
 248                 rx_bytes += len;
 249                 bufs[i] = mbufs[i];
 250         }
 251
 252         xsk_ring_cons__release(rx, rcvd);
 253
 254         /* statistics */
 255         rxq->stats.rx_pkts += (rcvd - dropped);
 256         rxq->stats.rx_bytes += rx_bytes;
 257
 258 out:
 259         if (rcvd != nb_pkts)
 260                 rte_mempool_put_bulk(rxq->mb_pool, (void **)&mbufs[rcvd],
 261                                      nb_pkts - rcvd);
 262
 263         return rcvd;
 264 }
 265
 266 static void
 267 pull_umem_cq(struct xsk_umem_info *umem, int size)
 268 {
 269         struct xsk_ring_cons *cq = &umem->cq;
 270         size_t i, n;
 271         uint32_t idx_cq = 0;
 272
 273         n = xsk_ring_cons__peek(cq, size, &idx_cq);
 274
 275         for (i = 0; i < n; i++) {
 276                 uint64_t addr;
 277                 addr = *xsk_ring_cons__comp_addr(cq, idx_cq++);
 278                 rte_ring_enqueue(umem->buf_ring, (void *)addr);
 279         }
 280
 281         xsk_ring_cons__release(cq, n);
 282 }
 283
 284 static void
 285 kick_tx(struct pkt_tx_queue *txq)
 286 {
 287         struct xsk_umem_info *umem = txq->pair->umem;
 288
 289 #if defined(XDP_USE_NEED_WAKEUP)
 290         if (xsk_ring_prod__needs_wakeup(&txq->tx))
 291 #endif
 292                 while (send(xsk_socket__fd(txq->pair->xsk), NULL,
 293                             0, MSG_DONTWAIT) < 0) {
 294                         /* some thing unexpected */
 295                         if (errno != EBUSY && errno != EAGAIN && errno != EINTR)
 296                                 break;
 297
 298                         /* pull from completion queue to leave more space */
 299                         if (errno == EAGAIN)
 300                                 pull_umem_cq(umem, ETH_AF_XDP_TX_BATCH_SIZE);
 301                 }
 302         pull_umem_cq(umem, ETH_AF_XDP_TX_BATCH_SIZE);
 303 }
 304
 305 static inline bool
 306 in_umem_range(struct xsk_umem_info *umem, uint64_t addr)
 307 {
 308         uint64_t mz_base_addr = umem->mz->addr_64;
 309
 310         return addr >= mz_base_addr && addr < mz_base_addr + umem->mz->len;
 311 }
 312
 313 static uint16_t
 314 eth_af_xdp_tx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 315 {
 316         struct pkt_tx_queue *txq = queue;
 317         struct xsk_umem_info *umem = txq->pair->umem;
 318         struct rte_mbuf *mbuf;
 319         int pmd_zc = umem->pmd_zc;
 320         void *addrs[ETH_AF_XDP_TX_BATCH_SIZE];
 321         unsigned long tx_bytes = 0;
 322         int i;
 323         uint32_t idx_tx;
 324
 325         nb_pkts = RTE_MIN(nb_pkts, ETH_AF_XDP_TX_BATCH_SIZE);
 326
 327         pull_umem_cq(umem, nb_pkts);
 328
 329         nb_pkts = rte_ring_dequeue_bulk(umem->buf_ring, addrs,
 330                                         nb_pkts, NULL);
 331         if (nb_pkts == 0)
 332                 return 0;
 333
 334         if (xsk_ring_prod__reserve(&txq->tx, nb_pkts, &idx_tx) != nb_pkts) {
 335                 kick_tx(txq);
 336                 rte_ring_enqueue_bulk(umem->buf_ring, addrs, nb_pkts, NULL);
 337                 return 0;
 338         }
 339
 340         for (i = 0; i < nb_pkts; i++) {
 341                 struct xdp_desc *desc;
 342                 void *pkt;
 343
 344                 desc = xsk_ring_prod__tx_desc(&txq->tx, idx_tx + i);
 345                 mbuf = bufs[i];
 346                 desc->len = mbuf->pkt_len;
 347
 348                 /*
 349                  * We need to make sure the external mbuf address is within
 350                  * current port's umem memzone range
 351                  */
 352                 if (pmd_zc && RTE_MBUF_HAS_EXTBUF(mbuf) &&
 353                                 in_umem_range(umem, (uint64_t)mbuf->buf_addr)) {
 354                         desc->addr = (uint64_t)mbuf->buf_addr -
 355                                 umem->mz->addr_64;
 356                         mbuf->buf_addr = xsk_umem__get_data(umem->mz->addr,
 357                                         (uint64_t)addrs[i]);
 358                 } else {
 359                         desc->addr = (uint64_t)addrs[i];
 360                         pkt = xsk_umem__get_data(umem->mz->addr,
 361                                         desc->addr);
 362                         rte_memcpy(pkt, rte_pktmbuf_mtod(mbuf, void *),
 363                                         desc->len);
 364                 }
 365                 tx_bytes += mbuf->pkt_len;
 366         }
 367
 368         xsk_ring_prod__submit(&txq->tx, nb_pkts);
 369
 370         kick_tx(txq);
 371
 372         txq->stats.tx_pkts += nb_pkts;
 373         txq->stats.tx_bytes += tx_bytes;
 374
 375         for (i = 0; i < nb_pkts; i++)
 376                 rte_pktmbuf_free(bufs[i]);
 377
 378         return nb_pkts;
 379 }
 380
 381 static int
 382 eth_dev_start(struct rte_eth_dev *dev)
 383 {
 384         dev->data->dev_link.link_status = ETH_LINK_UP;
 385
 386         return 0;
 387 }
 388
 389 /* This function gets called when the current port gets stopped. */
 390 static void
 391 eth_dev_stop(struct rte_eth_dev *dev)
 392 {
 393         dev->data->dev_link.link_status = ETH_LINK_DOWN;
 394 }
 395
 396 static int
 397 eth_dev_configure(struct rte_eth_dev *dev)
 398 {
 399         /* rx/tx must be paired */
 400         if (dev->data->nb_rx_queues != dev->data->nb_tx_queues)
 401                 return -EINVAL;
 402
 403         return 0;
 404 }
 405
 406 static int
 407 eth_dev_info(struct rte_eth_dev *dev, struct rte_eth_dev_info *dev_info)
 408 {
 409         struct pmd_internals *internals = dev->data->dev_private;
 410
 411         dev_info->if_index = internals->if_index;
 412         dev_info->max_mac_addrs = 1;
 413         dev_info->max_rx_pktlen = ETH_FRAME_LEN;
 414         dev_info->max_rx_queues = internals->queue_cnt;
 415         dev_info->max_tx_queues = internals->queue_cnt;
 416
 417         dev_info->min_mtu = RTE_ETHER_MIN_MTU;
 418         dev_info->max_mtu = ETH_AF_XDP_FRAME_SIZE - ETH_AF_XDP_DATA_HEADROOM;
 419
 420         dev_info->default_rxportconf.nb_queues = 1;
 421         dev_info->default_txportconf.nb_queues = 1;
 422         dev_info->default_rxportconf.ring_size = ETH_AF_XDP_DFLT_NUM_DESCS;
 423         dev_info->default_txportconf.ring_size = ETH_AF_XDP_DFLT_NUM_DESCS;
 424
 425         return 0;
 426 }
 427
 428 static int
 429 eth_stats_get(struct rte_eth_dev *dev, struct rte_eth_stats *stats)
 430 {
 431         struct pmd_internals *internals = dev->data->dev_private;
 432         struct xdp_statistics xdp_stats;
 433         struct pkt_rx_queue *rxq;
 434         struct pkt_tx_queue *txq;
 435         socklen_t optlen;
 436         int i, ret;
 437
 438         for (i = 0; i < dev->data->nb_rx_queues; i++) {
 439                 optlen = sizeof(struct xdp_statistics);
 440                 rxq = &internals->rx_queues[i];
 441                 txq = rxq->pair;
 442                 stats->q_ipackets[i] = rxq->stats.rx_pkts;
 443                 stats->q_ibytes[i] = rxq->stats.rx_bytes;
 444
 445                 stats->q_opackets[i] = txq->stats.tx_pkts;
 446                 stats->q_obytes[i] = txq->stats.tx_bytes;
 447
 448                 stats->ipackets += stats->q_ipackets[i];
 449                 stats->ibytes += stats->q_ibytes[i];
 450                 stats->imissed += rxq->stats.rx_dropped;
 451                 ret = getsockopt(xsk_socket__fd(rxq->xsk), SOL_XDP,
 452                                 XDP_STATISTICS, &xdp_stats, &optlen);
 453                 if (ret != 0) {
 454                         AF_XDP_LOG(ERR, "getsockopt() failed for XDP_STATISTICS.\n");
 455                         return -1;
 456                 }
 457                 stats->imissed += xdp_stats.rx_dropped;
 458
 459                 stats->opackets += stats->q_opackets[i];
 460                 stats->obytes += stats->q_obytes[i];
 461         }
 462
 463         return 0;
 464 }
 465
 466 static void
 467 eth_stats_reset(struct rte_eth_dev *dev)
 468 {
 469         struct pmd_internals *internals = dev->data->dev_private;
 470         int i;
 471
 472         for (i = 0; i < internals->queue_cnt; i++) {
 473                 memset(&internals->rx_queues[i].stats, 0,
 474                                         sizeof(struct rx_stats));
 475                 memset(&internals->tx_queues[i].stats, 0,
 476                                         sizeof(struct tx_stats));
 477         }
 478 }
 479
 480 static void
 481 remove_xdp_program(struct pmd_internals *internals)
 482 {
 483         uint32_t curr_prog_id = 0;
 484
 485         if (bpf_get_link_xdp_id(internals->if_index, &curr_prog_id,
 486                                 XDP_FLAGS_UPDATE_IF_NOEXIST)) {
 487                 AF_XDP_LOG(ERR, "bpf_get_link_xdp_id failed\n");
 488                 return;
 489         }
 490         bpf_set_link_xdp_fd(internals->if_index, -1,
 491                         XDP_FLAGS_UPDATE_IF_NOEXIST);
 492 }
 493
 494 static void
 495 xdp_umem_destroy(struct xsk_umem_info *umem)
 496 {
 497         rte_memzone_free(umem->mz);
 498         umem->mz = NULL;
 499
 500         rte_ring_free(umem->buf_ring);
 501         umem->buf_ring = NULL;
 502
 503         rte_free(umem);
 504         umem = NULL;
 505 }
 506
 507 static void
 508 eth_dev_close(struct rte_eth_dev *dev)
 509 {
 510         struct pmd_internals *internals = dev->data->dev_private;
 511         struct pkt_rx_queue *rxq;
 512         int i;
 513
 514         AF_XDP_LOG(INFO, "Closing AF_XDP ethdev on numa socket %u\n",
 515                 rte_socket_id());
 516
 517         for (i = 0; i < internals->queue_cnt; i++) {
 518                 rxq = &internals->rx_queues[i];
 519                 if (rxq->umem == NULL)
 520                         break;
 521                 xsk_socket__delete(rxq->xsk);
 522                 (void)xsk_umem__delete(rxq->umem->umem);
 523                 xdp_umem_destroy(rxq->umem);
 524
 525                 /* free pkt_tx_queue */
 526                 rte_free(rxq->pair);
 527                 rte_free(rxq);
 528         }
 529
 530         /*
 531          * MAC is not allocated dynamically, setting it to NULL would prevent
 532          * from releasing it in rte_eth_dev_release_port.
 533          */
 534         dev->data->mac_addrs = NULL;
 535
 536         remove_xdp_program(internals);
 537 }
 538
 539 static void
 540 eth_queue_release(void *q __rte_unused)
 541 {
 542 }
 543
 544 static int
 545 eth_link_update(struct rte_eth_dev *dev __rte_unused,
 546                 int wait_to_complete __rte_unused)
 547 {
 548         return 0;
 549 }
 550
 551 static struct
 552 xsk_umem_info *xdp_umem_configure(struct pmd_internals *internals,
 553                                   struct pkt_rx_queue *rxq)
 554 {
 555         struct xsk_umem_info *umem;
 556         const struct rte_memzone *mz;
 557         struct xsk_umem_config usr_config = {
 558                 .fill_size = ETH_AF_XDP_DFLT_NUM_DESCS,
 559                 .comp_size = ETH_AF_XDP_DFLT_NUM_DESCS,
 560                 .frame_size = ETH_AF_XDP_FRAME_SIZE,
 561                 .frame_headroom = ETH_AF_XDP_DATA_HEADROOM };
 562         char ring_name[RTE_RING_NAMESIZE];
 563         char mz_name[RTE_MEMZONE_NAMESIZE];
 564         int ret;
 565         uint64_t i;
 566
 567         umem = rte_zmalloc_socket("umem", sizeof(*umem), 0, rte_socket_id());
 568         if (umem == NULL) {
 569                 AF_XDP_LOG(ERR, "Failed to allocate umem info");
 570                 return NULL;
 571         }
 572
 573         snprintf(ring_name, sizeof(ring_name), "af_xdp_ring_%s_%u",
 574                        internals->if_name, rxq->xsk_queue_idx);
 575         umem->buf_ring = rte_ring_create(ring_name,
 576                                          ETH_AF_XDP_NUM_BUFFERS,
 577                                          rte_socket_id(),
 578                                          0x0);
 579         if (umem->buf_ring == NULL) {
 580                 AF_XDP_LOG(ERR, "Failed to create rte_ring\n");
 581                 goto err;
 582         }
 583
 584         for (i = 0; i < ETH_AF_XDP_NUM_BUFFERS; i++)
 585                 rte_ring_enqueue(umem->buf_ring,
 586                                  (void *)(i * ETH_AF_XDP_FRAME_SIZE +
 587                                           ETH_AF_XDP_DATA_HEADROOM));
 588
 589         snprintf(mz_name, sizeof(mz_name), "af_xdp_umem_%s_%u",
 590                        internals->if_name, rxq->xsk_queue_idx);
 591         mz = rte_memzone_reserve_aligned(mz_name,
 592                         ETH_AF_XDP_NUM_BUFFERS * ETH_AF_XDP_FRAME_SIZE,
 593                         rte_socket_id(), RTE_MEMZONE_IOVA_CONTIG,
 594                         getpagesize());
 595         if (mz == NULL) {
 596                 AF_XDP_LOG(ERR, "Failed to reserve memzone for af_xdp umem.\n");
 597                 goto err;
 598         }
 599
 600         ret = xsk_umem__create(&umem->umem, mz->addr,
 601                                ETH_AF_XDP_NUM_BUFFERS * ETH_AF_XDP_FRAME_SIZE,
 602                                &umem->fq, &umem->cq,
 603                                &usr_config);
 604
 605         if (ret) {
 606                 AF_XDP_LOG(ERR, "Failed to create umem");
 607                 goto err;
 608         }
 609         umem->mz = mz;
 610
 611         return umem;
 612
 613 err:
 614         xdp_umem_destroy(umem);
 615         return NULL;
 616 }
 617
 618 static int
 619 xsk_configure(struct pmd_internals *internals, struct pkt_rx_queue *rxq,
 620               int ring_size)
 621 {
 622         struct xsk_socket_config cfg;
 623         struct pkt_tx_queue *txq = rxq->pair;
 624         int ret = 0;
 625         int reserve_size;
 626
 627         rxq->umem = xdp_umem_configure(internals, rxq);
 628         if (rxq->umem == NULL)
 629                 return -ENOMEM;
 630
 631         cfg.rx_size = ring_size;
 632         cfg.tx_size = ring_size;
 633         cfg.libbpf_flags = 0;
 634         cfg.xdp_flags = XDP_FLAGS_UPDATE_IF_NOEXIST;
 635         cfg.bind_flags = 0;
 636
 637 #if defined(XDP_USE_NEED_WAKEUP)
 638         cfg.bind_flags |= XDP_USE_NEED_WAKEUP;
 639 #endif
 640
 641         ret = xsk_socket__create(&rxq->xsk, internals->if_name,
 642                         rxq->xsk_queue_idx, rxq->umem->umem, &rxq->rx,
 643                         &txq->tx, &cfg);
 644         if (ret) {
 645                 AF_XDP_LOG(ERR, "Failed to create xsk socket.\n");
 646                 goto err;
 647         }
 648
 649         reserve_size = ETH_AF_XDP_DFLT_NUM_DESCS / 2;
 650         ret = reserve_fill_queue(rxq->umem, reserve_size);
 651         if (ret) {
 652                 xsk_socket__delete(rxq->xsk);
 653                 AF_XDP_LOG(ERR, "Failed to reserve fill queue.\n");
 654                 goto err;
 655         }
 656
 657         return 0;
 658
 659 err:
 660         xdp_umem_destroy(rxq->umem);
 661
 662         return ret;
 663 }
 664
 665 static int
 666 eth_rx_queue_setup(struct rte_eth_dev *dev,
 667                    uint16_t rx_queue_id,
 668                    uint16_t nb_rx_desc,
 669                    unsigned int socket_id __rte_unused,
 670                    const struct rte_eth_rxconf *rx_conf __rte_unused,
 671                    struct rte_mempool *mb_pool)
 672 {
 673         struct pmd_internals *internals = dev->data->dev_private;
 674         uint32_t buf_size, data_size;
 675         struct pkt_rx_queue *rxq;
 676         int ret;
 677
 678         rxq = &internals->rx_queues[rx_queue_id];
 679
 680         AF_XDP_LOG(INFO, "Set up rx queue, rx queue id: %d, xsk queue id: %d\n",
 681                    rx_queue_id, rxq->xsk_queue_idx);
 682         /* Now get the space available for data in the mbuf */
 683         buf_size = rte_pktmbuf_data_room_size(mb_pool) -
 684                 RTE_PKTMBUF_HEADROOM;
 685         data_size = ETH_AF_XDP_FRAME_SIZE - ETH_AF_XDP_DATA_HEADROOM;
 686
 687         if (data_size > buf_size) {
 688                 AF_XDP_LOG(ERR, "%s: %d bytes will not fit in mbuf (%d bytes)\n",
 689                         dev->device->name, data_size, buf_size);
 690                 ret = -ENOMEM;
 691                 goto err;
 692         }
 693
 694         rxq->mb_pool = mb_pool;
 695
 696         if (xsk_configure(internals, rxq, nb_rx_desc)) {
 697                 AF_XDP_LOG(ERR, "Failed to configure xdp socket\n");
 698                 ret = -EINVAL;
 699                 goto err;
 700         }
 701
 702         rxq->fds[0].fd = xsk_socket__fd(rxq->xsk);
 703         rxq->fds[0].events = POLLIN;
 704
 705         rxq->umem->pmd_zc = internals->pmd_zc;
 706
 707         dev->data->rx_queues[rx_queue_id] = rxq;
 708         return 0;
 709
 710 err:
 711         return ret;
 712 }
 713
 714 static int
 715 eth_tx_queue_setup(struct rte_eth_dev *dev,
 716                    uint16_t tx_queue_id,
 717                    uint16_t nb_tx_desc __rte_unused,
 718                    unsigned int socket_id __rte_unused,
 719                    const struct rte_eth_txconf *tx_conf __rte_unused)
 720 {
 721         struct pmd_internals *internals = dev->data->dev_private;
 722         struct pkt_tx_queue *txq;
 723
 724         txq = &internals->tx_queues[tx_queue_id];
 725
 726         dev->data->tx_queues[tx_queue_id] = txq;
 727         return 0;
 728 }
 729
 730 static int
 731 eth_dev_mtu_set(struct rte_eth_dev *dev, uint16_t mtu)
 732 {
 733         struct pmd_internals *internals = dev->data->dev_private;
 734         struct ifreq ifr = { .ifr_mtu = mtu };
 735         int ret;
 736         int s;
 737
 738         s = socket(PF_INET, SOCK_DGRAM, 0);
 739         if (s < 0)
 740                 return -EINVAL;
 741
 742         strlcpy(ifr.ifr_name, internals->if_name, IFNAMSIZ);
 743         ret = ioctl(s, SIOCSIFMTU, &ifr);
 744         close(s);
 745
 746         return (ret < 0) ? -errno : 0;
 747 }
 748
 749 static void
 750 eth_dev_change_flags(char *if_name, uint32_t flags, uint32_t mask)
 751 {
 752         struct ifreq ifr;
 753         int s;
 754
 755         s = socket(PF_INET, SOCK_DGRAM, 0);
 756         if (s < 0)
 757                 return;
 758
 759         strlcpy(ifr.ifr_name, if_name, IFNAMSIZ);
 760         if (ioctl(s, SIOCGIFFLAGS, &ifr) < 0)
 761                 goto out;
 762         ifr.ifr_flags &= mask;
 763         ifr.ifr_flags |= flags;
 764         if (ioctl(s, SIOCSIFFLAGS, &ifr) < 0)
 765                 goto out;
 766 out:
 767         close(s);
 768 }
 769
 770 static void
 771 eth_dev_promiscuous_enable(struct rte_eth_dev *dev)
 772 {
 773         struct pmd_internals *internals = dev->data->dev_private;
 774
 775         eth_dev_change_flags(internals->if_name, IFF_PROMISC, ~0);
 776 }
 777
 778 static void
 779 eth_dev_promiscuous_disable(struct rte_eth_dev *dev)
 780 {
 781         struct pmd_internals *internals = dev->data->dev_private;
 782
 783         eth_dev_change_flags(internals->if_name, 0, ~IFF_PROMISC);
 784 }
 785
 786 static const struct eth_dev_ops ops = {
 787         .dev_start = eth_dev_start,
 788         .dev_stop = eth_dev_stop,
 789         .dev_close = eth_dev_close,
 790         .dev_configure = eth_dev_configure,
 791         .dev_infos_get = eth_dev_info,
 792         .mtu_set = eth_dev_mtu_set,
 793         .promiscuous_enable = eth_dev_promiscuous_enable,
 794         .promiscuous_disable = eth_dev_promiscuous_disable,
 795         .rx_queue_setup = eth_rx_queue_setup,
 796         .tx_queue_setup = eth_tx_queue_setup,
 797         .rx_queue_release = eth_queue_release,
 798         .tx_queue_release = eth_queue_release,
 799         .link_update = eth_link_update,
 800         .stats_get = eth_stats_get,
 801         .stats_reset = eth_stats_reset,
 802 };
 803
 804 /** parse integer from integer argument */
 805 static int
 806 parse_integer_arg(const char *key __rte_unused,
 807                   const char *value, void *extra_args)
 808 {
 809         int *i = (int *)extra_args;
 810         char *end;
 811
 812         *i = strtol(value, &end, 10);
 813         if (*i < 0) {
 814                 AF_XDP_LOG(ERR, "Argument has to be positive.\n");
 815                 return -EINVAL;
 816         }
 817
 818         return 0;
 819 }
 820
 821 /** parse name argument */
 822 static int
 823 parse_name_arg(const char *key __rte_unused,
 824                const char *value, void *extra_args)
 825 {
 826         char *name = extra_args;
 827
 828         if (strnlen(value, IFNAMSIZ) > IFNAMSIZ - 1) {
 829                 AF_XDP_LOG(ERR, "Invalid name %s, should be less than %u bytes.\n",
 830                            value, IFNAMSIZ);
 831                 return -EINVAL;
 832         }
 833
 834         strlcpy(name, value, IFNAMSIZ);
 835
 836         return 0;
 837 }
 838
 839 static int
 840 xdp_get_channels_info(const char *if_name, int *max_queues,
 841                                 int *combined_queues)
 842 {
 843         struct ethtool_channels channels;
 844         struct ifreq ifr;
 845         int fd, ret;
 846
 847         fd = socket(AF_INET, SOCK_DGRAM, 0);
 848         if (fd < 0)
 849                 return -1;
 850
 851         channels.cmd = ETHTOOL_GCHANNELS;
 852         ifr.ifr_data = (void *)&channels;
 853         strncpy(ifr.ifr_name, if_name, IFNAMSIZ);
 854         ret = ioctl(fd, SIOCETHTOOL, &ifr);
 855         if (ret) {
 856                 if (errno == EOPNOTSUPP) {
 857                         ret = 0;
 858                 } else {
 859                         ret = -errno;
 860                         goto out;
 861                 }
 862         }
 863
 864         if (channels.max_combined == 0 || errno == EOPNOTSUPP) {
 865                 /* If the device says it has no channels, then all traffic
 866                  * is sent to a single stream, so max queues = 1.
 867                  */
 868                 *max_queues = 1;
 869                 *combined_queues = 1;
 870         } else {
 871                 *max_queues = channels.max_combined;
 872                 *combined_queues = channels.combined_count;
 873         }
 874
 875  out:
 876         close(fd);
 877         return ret;
 878 }
 879
 880 static int
 881 parse_parameters(struct rte_kvargs *kvlist, char *if_name, int *start_queue,
 882                         int *queue_cnt, int *pmd_zc)
 883 {
 884         int ret;
 885
 886         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_IFACE_ARG,
 887                                  &parse_name_arg, if_name);
 888         if (ret < 0)
 889                 goto free_kvlist;
 890
 891         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_START_QUEUE_ARG,
 892                                  &parse_integer_arg, start_queue);
 893         if (ret < 0)
 894                 goto free_kvlist;
 895
 896         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_QUEUE_COUNT_ARG,
 897                                  &parse_integer_arg, queue_cnt);
 898         if (ret < 0 || *queue_cnt <= 0) {
 899                 ret = -EINVAL;
 900                 goto free_kvlist;
 901         }
 902
 903         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_PMD_ZC_ARG,
 904                                  &parse_integer_arg, pmd_zc);
 905         if (ret < 0)
 906                 goto free_kvlist;
 907
 908 free_kvlist:
 909         rte_kvargs_free(kvlist);
 910         return ret;
 911 }
 912
 913 static int
 914 get_iface_info(const char *if_name,
 915                struct rte_ether_addr *eth_addr,
 916                int *if_index)
 917 {
 918         struct ifreq ifr;
 919         int sock = socket(AF_INET, SOCK_DGRAM, IPPROTO_IP);
 920
 921         if (sock < 0)
 922                 return -1;
 923
 924         strlcpy(ifr.ifr_name, if_name, IFNAMSIZ);
 925         if (ioctl(sock, SIOCGIFINDEX, &ifr))
 926                 goto error;
 927
 928         *if_index = ifr.ifr_ifindex;
 929
 930         if (ioctl(sock, SIOCGIFHWADDR, &ifr))
 931                 goto error;
 932
 933         rte_memcpy(eth_addr, ifr.ifr_hwaddr.sa_data, RTE_ETHER_ADDR_LEN);
 934
 935         close(sock);
 936         return 0;
 937
 938 error:
 939         close(sock);
 940         return -1;
 941 }
 942
 943 static struct rte_eth_dev *
 944 init_internals(struct rte_vdev_device *dev, const char *if_name,
 945                         int start_queue_idx, int queue_cnt, int pmd_zc)
 946 {
 947         const char *name = rte_vdev_device_name(dev);
 948         const unsigned int numa_node = dev->device.numa_node;
 949         struct pmd_internals *internals;
 950         struct rte_eth_dev *eth_dev;
 951         int ret;
 952         int i;
 953
 954         internals = rte_zmalloc_socket(name, sizeof(*internals), 0, numa_node);
 955         if (internals == NULL)
 956                 return NULL;
 957
 958         internals->start_queue_idx = start_queue_idx;
 959         internals->queue_cnt = queue_cnt;
 960         internals->pmd_zc = pmd_zc;
 961         strlcpy(internals->if_name, if_name, IFNAMSIZ);
 962
 963         if (xdp_get_channels_info(if_name, &internals->max_queue_cnt,
 964                                   &internals->combined_queue_cnt)) {
 965                 AF_XDP_LOG(ERR, "Failed to get channel info of interface: %s\n",
 966                                 if_name);
 967                 goto err_free_internals;
 968         }
 969
 970         if (queue_cnt > internals->combined_queue_cnt) {
 971                 AF_XDP_LOG(ERR, "Specified queue count %d is larger than combined queue count %d.\n",
 972                                 queue_cnt, internals->combined_queue_cnt);
 973                 goto err_free_internals;
 974         }
 975
 976         internals->rx_queues = rte_zmalloc_socket(NULL,
 977                                         sizeof(struct pkt_rx_queue) * queue_cnt,
 978                                         0, numa_node);
 979         if (internals->rx_queues == NULL) {
 980                 AF_XDP_LOG(ERR, "Failed to allocate memory for rx queues.\n");
 981                 goto err_free_internals;
 982         }
 983
 984         internals->tx_queues = rte_zmalloc_socket(NULL,
 985                                         sizeof(struct pkt_tx_queue) * queue_cnt,
 986                                         0, numa_node);
 987         if (internals->tx_queues == NULL) {
 988                 AF_XDP_LOG(ERR, "Failed to allocate memory for tx queues.\n");
 989                 goto err_free_rx;
 990         }
 991         for (i = 0; i < queue_cnt; i++) {
 992                 internals->tx_queues[i].pair = &internals->rx_queues[i];
 993                 internals->rx_queues[i].pair = &internals->tx_queues[i];
 994                 internals->rx_queues[i].xsk_queue_idx = start_queue_idx + i;
 995                 internals->tx_queues[i].xsk_queue_idx = start_queue_idx + i;
 996         }
 997
 998         ret = get_iface_info(if_name, &internals->eth_addr,
 999                              &internals->if_index);
1000         if (ret)
1001                 goto err_free_tx;
1002
1003         eth_dev = rte_eth_vdev_allocate(dev, 0);
1004         if (eth_dev == NULL)
1005                 goto err_free_tx;
1006
1007         eth_dev->data->dev_private = internals;
1008         eth_dev->data->dev_link = pmd_link;
1009         eth_dev->data->mac_addrs = &internals->eth_addr;
1010         eth_dev->dev_ops = &ops;
1011         eth_dev->rx_pkt_burst = eth_af_xdp_rx;
1012         eth_dev->tx_pkt_burst = eth_af_xdp_tx;
1013         /* Let rte_eth_dev_close() release the port resources. */
1014         eth_dev->data->dev_flags |= RTE_ETH_DEV_CLOSE_REMOVE;
1015
1016         if (internals->pmd_zc)
1017                 AF_XDP_LOG(INFO, "Zero copy between umem and mbuf enabled.\n");
1018
1019         return eth_dev;
1020
1021 err_free_tx:
1022         rte_free(internals->tx_queues);
1023 err_free_rx:
1024         rte_free(internals->rx_queues);
1025 err_free_internals:
1026         rte_free(internals);
1027         return NULL;
1028 }
1029
1030 static int
1031 rte_pmd_af_xdp_probe(struct rte_vdev_device *dev)
1032 {
1033         struct rte_kvargs *kvlist;
1034         char if_name[IFNAMSIZ] = {'\0'};
1035         int xsk_start_queue_idx = ETH_AF_XDP_DFLT_START_QUEUE_IDX;
1036         int xsk_queue_cnt = ETH_AF_XDP_DFLT_QUEUE_COUNT;
1037         struct rte_eth_dev *eth_dev = NULL;
1038         const char *name;
1039         int pmd_zc = 0;
1040
1041         AF_XDP_LOG(INFO, "Initializing pmd_af_xdp for %s\n",
1042                 rte_vdev_device_name(dev));
1043
1044         name = rte_vdev_device_name(dev);
1045         if (rte_eal_process_type() == RTE_PROC_SECONDARY &&
1046                 strlen(rte_vdev_device_args(dev)) == 0) {
1047                 eth_dev = rte_eth_dev_attach_secondary(name);
1048                 if (eth_dev == NULL) {
1049                         AF_XDP_LOG(ERR, "Failed to probe %s\n", name);
1050                         return -EINVAL;
1051                 }
1052                 eth_dev->dev_ops = &ops;
1053                 rte_eth_dev_probing_finish(eth_dev);
1054                 return 0;
1055         }
1056
1057         kvlist = rte_kvargs_parse(rte_vdev_device_args(dev), valid_arguments);
1058         if (kvlist == NULL) {
1059                 AF_XDP_LOG(ERR, "Invalid kvargs key\n");
1060                 return -EINVAL;
1061         }
1062
1063         if (dev->device.numa_node == SOCKET_ID_ANY)
1064                 dev->device.numa_node = rte_socket_id();
1065
1066         if (parse_parameters(kvlist, if_name, &xsk_start_queue_idx,
1067                              &xsk_queue_cnt, &pmd_zc) < 0) {
1068                 AF_XDP_LOG(ERR, "Invalid kvargs value\n");
1069                 return -EINVAL;
1070         }
1071
1072         if (strlen(if_name) == 0) {
1073                 AF_XDP_LOG(ERR, "Network interface must be specified\n");
1074                 return -EINVAL;
1075         }
1076
1077         eth_dev = init_internals(dev, if_name, xsk_start_queue_idx,
1078                                         xsk_queue_cnt, pmd_zc);
1079         if (eth_dev == NULL) {
1080                 AF_XDP_LOG(ERR, "Failed to init internals\n");
1081                 return -1;
1082         }
1083
1084         rte_eth_dev_probing_finish(eth_dev);
1085
1086         return 0;
1087 }
1088
1089 static int
1090 rte_pmd_af_xdp_remove(struct rte_vdev_device *dev)
1091 {
1092         struct rte_eth_dev *eth_dev = NULL;
1093
1094         AF_XDP_LOG(INFO, "Removing AF_XDP ethdev on numa socket %u\n",
1095                 rte_socket_id());
1096
1097         if (dev == NULL)
1098                 return -1;
1099
1100         /* find the ethdev entry */
1101         eth_dev = rte_eth_dev_allocated(rte_vdev_device_name(dev));
1102         if (eth_dev == NULL)
1103                 return 0;
1104
1105         eth_dev_close(eth_dev);
1106         rte_eth_dev_release_port(eth_dev);
1107
1108
1109         return 0;
1110 }
1111
1112 static struct rte_vdev_driver pmd_af_xdp_drv = {
1113         .probe = rte_pmd_af_xdp_probe,
1114         .remove = rte_pmd_af_xdp_remove,
1115 };
1116
1117 RTE_PMD_REGISTER_VDEV(net_af_xdp, pmd_af_xdp_drv);
1118 RTE_PMD_REGISTER_PARAM_STRING(net_af_xdp,
1119                               "iface=<string> "
1120                               "start_queue=<int> "
1121                               "queue_count=<int> "
1122                               "pmd_zero_copy=<0|1>");
1123
1124 RTE_INIT(af_xdp_init_log)
1125 {
1126         af_xdp_logtype = rte_log_register("pmd.net.af_xdp");
1127         if (af_xdp_logtype >= 0)
1128                 rte_log_set_level(af_xdp_logtype, RTE_LOG_NOTICE);
1129 }