drivers/net/af_xdp/rte_eth_af_xdp.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2019-2020 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/bpf.h>
  19 #include <bpf/xsk.h>
  20
  21 #include <rte_ethdev.h>
  22 #include <ethdev_driver.h>
  23 #include <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_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_mempool.h>
  37 #include <rte_mbuf.h>
  38 #include <rte_malloc.h>
  39 #include <rte_ring.h>
  40 #include <rte_spinlock.h>
  41 #include <rte_power_intrinsics.h>
  42
  43 #include "compat.h"
  44
  45 #ifndef SO_PREFER_BUSY_POLL
  46 #define SO_PREFER_BUSY_POLL 69
  47 #endif
  48 #ifndef SO_BUSY_POLL_BUDGET
  49 #define SO_BUSY_POLL_BUDGET 70
  50 #endif
  51
  52
  53 #ifndef SOL_XDP
  54 #define SOL_XDP 283
  55 #endif
  56
  57 #ifndef AF_XDP
  58 #define AF_XDP 44
  59 #endif
  60
  61 #ifndef PF_XDP
  62 #define PF_XDP AF_XDP
  63 #endif
  64
  65 RTE_LOG_REGISTER_DEFAULT(af_xdp_logtype, NOTICE);
  66
  67 #define AF_XDP_LOG(level, fmt, args...)                 \
  68         rte_log(RTE_LOG_ ## level, af_xdp_logtype,      \
  69                 "%s(): " fmt, __func__, ##args)
  70
  71 #define ETH_AF_XDP_FRAME_SIZE           2048
  72 #define ETH_AF_XDP_NUM_BUFFERS          4096
  73 #define ETH_AF_XDP_DFLT_NUM_DESCS       XSK_RING_CONS__DEFAULT_NUM_DESCS
  74 #define ETH_AF_XDP_DFLT_START_QUEUE_IDX 0
  75 #define ETH_AF_XDP_DFLT_QUEUE_COUNT     1
  76 #define ETH_AF_XDP_DFLT_BUSY_BUDGET     64
  77 #define ETH_AF_XDP_DFLT_BUSY_TIMEOUT    20
  78
  79 #define ETH_AF_XDP_RX_BATCH_SIZE        XSK_RING_CONS__DEFAULT_NUM_DESCS
  80 #define ETH_AF_XDP_TX_BATCH_SIZE        XSK_RING_CONS__DEFAULT_NUM_DESCS
  81
  82 #define ETH_AF_XDP_ETH_OVERHEAD         (RTE_ETHER_HDR_LEN + RTE_ETHER_CRC_LEN)
  83
  84 struct xsk_umem_info {
  85         struct xsk_umem *umem;
  86         struct rte_ring *buf_ring;
  87         const struct rte_memzone *mz;
  88         struct rte_mempool *mb_pool;
  89         void *buffer;
  90         uint8_t refcnt;
  91         uint32_t max_xsks;
  92 };
  93
  94 struct rx_stats {
  95         uint64_t rx_pkts;
  96         uint64_t rx_bytes;
  97         uint64_t rx_dropped;
  98 };
  99
 100 struct pkt_rx_queue {
 101         struct xsk_ring_cons rx;
 102         struct xsk_umem_info *umem;
 103         struct xsk_socket *xsk;
 104         struct rte_mempool *mb_pool;
 105
 106         struct rx_stats stats;
 107
 108         struct xsk_ring_prod fq;
 109         struct xsk_ring_cons cq;
 110
 111         struct pkt_tx_queue *pair;
 112         struct pollfd fds[1];
 113         int xsk_queue_idx;
 114         int busy_budget;
 115 };
 116
 117 struct tx_stats {
 118         uint64_t tx_pkts;
 119         uint64_t tx_bytes;
 120         uint64_t tx_dropped;
 121 };
 122
 123 struct pkt_tx_queue {
 124         struct xsk_ring_prod tx;
 125         struct xsk_umem_info *umem;
 126
 127         struct tx_stats stats;
 128
 129         struct pkt_rx_queue *pair;
 130         int xsk_queue_idx;
 131 };
 132
 133 struct pmd_internals {
 134         int if_index;
 135         char if_name[IFNAMSIZ];
 136         int start_queue_idx;
 137         int queue_cnt;
 138         int max_queue_cnt;
 139         int combined_queue_cnt;
 140         bool shared_umem;
 141         char prog_path[PATH_MAX];
 142         bool custom_prog_configured;
 143         struct bpf_map *map;
 144
 145         struct rte_ether_addr eth_addr;
 146
 147         struct pkt_rx_queue *rx_queues;
 148         struct pkt_tx_queue *tx_queues;
 149 };
 150
 151 #define ETH_AF_XDP_IFACE_ARG                    "iface"
 152 #define ETH_AF_XDP_START_QUEUE_ARG              "start_queue"
 153 #define ETH_AF_XDP_QUEUE_COUNT_ARG              "queue_count"
 154 #define ETH_AF_XDP_SHARED_UMEM_ARG              "shared_umem"
 155 #define ETH_AF_XDP_PROG_ARG                     "xdp_prog"
 156 #define ETH_AF_XDP_BUDGET_ARG                   "busy_budget"
 157
 158 static const char * const valid_arguments[] = {
 159         ETH_AF_XDP_IFACE_ARG,
 160         ETH_AF_XDP_START_QUEUE_ARG,
 161         ETH_AF_XDP_QUEUE_COUNT_ARG,
 162         ETH_AF_XDP_SHARED_UMEM_ARG,
 163         ETH_AF_XDP_PROG_ARG,
 164         ETH_AF_XDP_BUDGET_ARG,
 165         NULL
 166 };
 167
 168 static const struct rte_eth_link pmd_link = {
 169         .link_speed = RTE_ETH_SPEED_NUM_10G,
 170         .link_duplex = RTE_ETH_LINK_FULL_DUPLEX,
 171         .link_status = RTE_ETH_LINK_DOWN,
 172         .link_autoneg = RTE_ETH_LINK_AUTONEG
 173 };
 174
 175 /* List which tracks PMDs to facilitate sharing UMEMs across them. */
 176 struct internal_list {
 177         TAILQ_ENTRY(internal_list) next;
 178         struct rte_eth_dev *eth_dev;
 179 };
 180
 181 TAILQ_HEAD(internal_list_head, internal_list);
 182 static struct internal_list_head internal_list =
 183         TAILQ_HEAD_INITIALIZER(internal_list);
 184
 185 static pthread_mutex_t internal_list_lock = PTHREAD_MUTEX_INITIALIZER;
 186
 187 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 188 static inline int
 189 reserve_fill_queue_zc(struct xsk_umem_info *umem, uint16_t reserve_size,
 190                       struct rte_mbuf **bufs, struct xsk_ring_prod *fq)
 191 {
 192         uint32_t idx;
 193         uint16_t i;
 194
 195         if (unlikely(!xsk_ring_prod__reserve(fq, reserve_size, &idx))) {
 196                 for (i = 0; i < reserve_size; i++)
 197                         rte_pktmbuf_free(bufs[i]);
 198                 AF_XDP_LOG(DEBUG, "Failed to reserve enough fq descs.\n");
 199                 return -1;
 200         }
 201
 202         for (i = 0; i < reserve_size; i++) {
 203                 __u64 *fq_addr;
 204                 uint64_t addr;
 205
 206                 fq_addr = xsk_ring_prod__fill_addr(fq, idx++);
 207                 addr = (uint64_t)bufs[i] - (uint64_t)umem->buffer -
 208                                 umem->mb_pool->header_size;
 209                 *fq_addr = addr;
 210         }
 211
 212         xsk_ring_prod__submit(fq, reserve_size);
 213
 214         return 0;
 215 }
 216 #else
 217 static inline int
 218 reserve_fill_queue_cp(struct xsk_umem_info *umem, uint16_t reserve_size,
 219                       struct rte_mbuf **bufs __rte_unused,
 220                       struct xsk_ring_prod *fq)
 221 {
 222         void *addrs[reserve_size];
 223         uint32_t idx;
 224         uint16_t i;
 225
 226         if (rte_ring_dequeue_bulk(umem->buf_ring, addrs, reserve_size, NULL)
 227                     != reserve_size) {
 228                 AF_XDP_LOG(DEBUG, "Failed to get enough buffers for fq.\n");
 229                 return -1;
 230         }
 231
 232         if (unlikely(!xsk_ring_prod__reserve(fq, reserve_size, &idx))) {
 233                 AF_XDP_LOG(DEBUG, "Failed to reserve enough fq descs.\n");
 234                 rte_ring_enqueue_bulk(umem->buf_ring, addrs,
 235                                 reserve_size, NULL);
 236                 return -1;
 237         }
 238
 239         for (i = 0; i < reserve_size; i++) {
 240                 __u64 *fq_addr;
 241
 242                 fq_addr = xsk_ring_prod__fill_addr(fq, idx++);
 243                 *fq_addr = (uint64_t)addrs[i];
 244         }
 245
 246         xsk_ring_prod__submit(fq, reserve_size);
 247
 248         return 0;
 249 }
 250 #endif
 251
 252 static inline int
 253 reserve_fill_queue(struct xsk_umem_info *umem, uint16_t reserve_size,
 254                    struct rte_mbuf **bufs, struct xsk_ring_prod *fq)
 255 {
 256 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 257         return reserve_fill_queue_zc(umem, reserve_size, bufs, fq);
 258 #else
 259         return reserve_fill_queue_cp(umem, reserve_size, bufs, fq);
 260 #endif
 261 }
 262
 263 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 264 static uint16_t
 265 af_xdp_rx_zc(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 266 {
 267         struct pkt_rx_queue *rxq = queue;
 268         struct xsk_ring_cons *rx = &rxq->rx;
 269         struct xsk_ring_prod *fq = &rxq->fq;
 270         struct xsk_umem_info *umem = rxq->umem;
 271         uint32_t idx_rx = 0;
 272         unsigned long rx_bytes = 0;
 273         int i;
 274         struct rte_mbuf *fq_bufs[ETH_AF_XDP_RX_BATCH_SIZE];
 275
 276         nb_pkts = xsk_ring_cons__peek(rx, nb_pkts, &idx_rx);
 277
 278         if (nb_pkts == 0) {
 279                 /* we can assume a kernel >= 5.11 is in use if busy polling is
 280                  * enabled and thus we can safely use the recvfrom() syscall
 281                  * which is only supported for AF_XDP sockets in kernels >=
 282                  * 5.11.
 283                  */
 284                 if (rxq->busy_budget) {
 285                         (void)recvfrom(xsk_socket__fd(rxq->xsk), NULL, 0,
 286                                        MSG_DONTWAIT, NULL, NULL);
 287                 } else if (xsk_ring_prod__needs_wakeup(fq)) {
 288                         (void)poll(&rxq->fds[0], 1, 1000);
 289                 }
 290
 291                 return 0;
 292         }
 293
 294         /* allocate bufs for fill queue replenishment after rx */
 295         if (rte_pktmbuf_alloc_bulk(umem->mb_pool, fq_bufs, nb_pkts)) {
 296                 AF_XDP_LOG(DEBUG,
 297                         "Failed to get enough buffers for fq.\n");
 298                 /* rollback cached_cons which is added by
 299                  * xsk_ring_cons__peek
 300                  */
 301                 rx->cached_cons -= nb_pkts;
 302                 return 0;
 303         }
 304
 305         for (i = 0; i < nb_pkts; i++) {
 306                 const struct xdp_desc *desc;
 307                 uint64_t addr;
 308                 uint32_t len;
 309                 uint64_t offset;
 310
 311                 desc = xsk_ring_cons__rx_desc(rx, idx_rx++);
 312                 addr = desc->addr;
 313                 len = desc->len;
 314
 315                 offset = xsk_umem__extract_offset(addr);
 316                 addr = xsk_umem__extract_addr(addr);
 317
 318                 bufs[i] = (struct rte_mbuf *)
 319                                 xsk_umem__get_data(umem->buffer, addr +
 320                                         umem->mb_pool->header_size);
 321                 bufs[i]->data_off = offset - sizeof(struct rte_mbuf) -
 322                         rte_pktmbuf_priv_size(umem->mb_pool) -
 323                         umem->mb_pool->header_size;
 324
 325                 rte_pktmbuf_pkt_len(bufs[i]) = len;
 326                 rte_pktmbuf_data_len(bufs[i]) = len;
 327                 rx_bytes += len;
 328         }
 329
 330         xsk_ring_cons__release(rx, nb_pkts);
 331         (void)reserve_fill_queue(umem, nb_pkts, fq_bufs, fq);
 332
 333         /* statistics */
 334         rxq->stats.rx_pkts += nb_pkts;
 335         rxq->stats.rx_bytes += rx_bytes;
 336
 337         return nb_pkts;
 338 }
 339 #else
 340 static uint16_t
 341 af_xdp_rx_cp(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 342 {
 343         struct pkt_rx_queue *rxq = queue;
 344         struct xsk_ring_cons *rx = &rxq->rx;
 345         struct xsk_umem_info *umem = rxq->umem;
 346         struct xsk_ring_prod *fq = &rxq->fq;
 347         uint32_t idx_rx = 0;
 348         unsigned long rx_bytes = 0;
 349         int i;
 350         uint32_t free_thresh = fq->size >> 1;
 351         struct rte_mbuf *mbufs[ETH_AF_XDP_RX_BATCH_SIZE];
 352
 353         if (xsk_prod_nb_free(fq, free_thresh) >= free_thresh)
 354                 (void)reserve_fill_queue(umem, nb_pkts, NULL, fq);
 355
 356         nb_pkts = xsk_ring_cons__peek(rx, nb_pkts, &idx_rx);
 357         if (nb_pkts == 0) {
 358 #if defined(XDP_USE_NEED_WAKEUP)
 359                 if (xsk_ring_prod__needs_wakeup(fq))
 360                         (void)poll(rxq->fds, 1, 1000);
 361 #endif
 362                 return 0;
 363         }
 364
 365         if (unlikely(rte_pktmbuf_alloc_bulk(rxq->mb_pool, mbufs, nb_pkts))) {
 366                 /* rollback cached_cons which is added by
 367                  * xsk_ring_cons__peek
 368                  */
 369                 rx->cached_cons -= nb_pkts;
 370                 return 0;
 371         }
 372
 373         for (i = 0; i < nb_pkts; i++) {
 374                 const struct xdp_desc *desc;
 375                 uint64_t addr;
 376                 uint32_t len;
 377                 void *pkt;
 378
 379                 desc = xsk_ring_cons__rx_desc(rx, idx_rx++);
 380                 addr = desc->addr;
 381                 len = desc->len;
 382                 pkt = xsk_umem__get_data(rxq->umem->mz->addr, addr);
 383
 384                 rte_memcpy(rte_pktmbuf_mtod(mbufs[i], void *), pkt, len);
 385                 rte_ring_enqueue(umem->buf_ring, (void *)addr);
 386                 rte_pktmbuf_pkt_len(mbufs[i]) = len;
 387                 rte_pktmbuf_data_len(mbufs[i]) = len;
 388                 rx_bytes += len;
 389                 bufs[i] = mbufs[i];
 390         }
 391
 392         xsk_ring_cons__release(rx, nb_pkts);
 393
 394         /* statistics */
 395         rxq->stats.rx_pkts += nb_pkts;
 396         rxq->stats.rx_bytes += rx_bytes;
 397
 398         return nb_pkts;
 399 }
 400 #endif
 401
 402 static uint16_t
 403 af_xdp_rx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 404 {
 405 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 406         return af_xdp_rx_zc(queue, bufs, nb_pkts);
 407 #else
 408         return af_xdp_rx_cp(queue, bufs, nb_pkts);
 409 #endif
 410 }
 411
 412 static uint16_t
 413 eth_af_xdp_rx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 414 {
 415         uint16_t nb_rx;
 416
 417         if (likely(nb_pkts <= ETH_AF_XDP_RX_BATCH_SIZE))
 418                 return af_xdp_rx(queue, bufs, nb_pkts);
 419
 420         /* Split larger batch into smaller batches of size
 421          * ETH_AF_XDP_RX_BATCH_SIZE or less.
 422          */
 423         nb_rx = 0;
 424         while (nb_pkts) {
 425                 uint16_t ret, n;
 426
 427                 n = (uint16_t)RTE_MIN(nb_pkts, ETH_AF_XDP_RX_BATCH_SIZE);
 428                 ret = af_xdp_rx(queue, &bufs[nb_rx], n);
 429                 nb_rx = (uint16_t)(nb_rx + ret);
 430                 nb_pkts = (uint16_t)(nb_pkts - ret);
 431                 if (ret < n)
 432                         break;
 433         }
 434
 435         return nb_rx;
 436 }
 437
 438 static void
 439 pull_umem_cq(struct xsk_umem_info *umem, int size, struct xsk_ring_cons *cq)
 440 {
 441         size_t i, n;
 442         uint32_t idx_cq = 0;
 443
 444         n = xsk_ring_cons__peek(cq, size, &idx_cq);
 445
 446         for (i = 0; i < n; i++) {
 447                 uint64_t addr;
 448                 addr = *xsk_ring_cons__comp_addr(cq, idx_cq++);
 449 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 450                 addr = xsk_umem__extract_addr(addr);
 451                 rte_pktmbuf_free((struct rte_mbuf *)
 452                                         xsk_umem__get_data(umem->buffer,
 453                                         addr + umem->mb_pool->header_size));
 454 #else
 455                 rte_ring_enqueue(umem->buf_ring, (void *)addr);
 456 #endif
 457         }
 458
 459         xsk_ring_cons__release(cq, n);
 460 }
 461
 462 static void
 463 kick_tx(struct pkt_tx_queue *txq, struct xsk_ring_cons *cq)
 464 {
 465         struct xsk_umem_info *umem = txq->umem;
 466
 467         pull_umem_cq(umem, XSK_RING_CONS__DEFAULT_NUM_DESCS, cq);
 468
 469         if (tx_syscall_needed(&txq->tx))
 470                 while (send(xsk_socket__fd(txq->pair->xsk), NULL,
 471                             0, MSG_DONTWAIT) < 0) {
 472                         /* some thing unexpected */
 473                         if (errno != EBUSY && errno != EAGAIN && errno != EINTR)
 474                                 break;
 475
 476                         /* pull from completion queue to leave more space */
 477                         if (errno == EAGAIN)
 478                                 pull_umem_cq(umem,
 479                                              XSK_RING_CONS__DEFAULT_NUM_DESCS,
 480                                              cq);
 481                 }
 482 }
 483
 484 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 485 static uint16_t
 486 af_xdp_tx_zc(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 487 {
 488         struct pkt_tx_queue *txq = queue;
 489         struct xsk_umem_info *umem = txq->umem;
 490         struct rte_mbuf *mbuf;
 491         unsigned long tx_bytes = 0;
 492         int i;
 493         uint32_t idx_tx;
 494         uint16_t count = 0;
 495         struct xdp_desc *desc;
 496         uint64_t addr, offset;
 497         struct xsk_ring_cons *cq = &txq->pair->cq;
 498         uint32_t free_thresh = cq->size >> 1;
 499
 500         if (xsk_cons_nb_avail(cq, free_thresh) >= free_thresh)
 501                 pull_umem_cq(umem, XSK_RING_CONS__DEFAULT_NUM_DESCS, cq);
 502
 503         for (i = 0; i < nb_pkts; i++) {
 504                 mbuf = bufs[i];
 505
 506                 if (mbuf->pool == umem->mb_pool) {
 507                         if (!xsk_ring_prod__reserve(&txq->tx, 1, &idx_tx)) {
 508                                 kick_tx(txq, cq);
 509                                 if (!xsk_ring_prod__reserve(&txq->tx, 1,
 510                                                             &idx_tx))
 511                                         goto out;
 512                         }
 513                         desc = xsk_ring_prod__tx_desc(&txq->tx, idx_tx);
 514                         desc->len = mbuf->pkt_len;
 515                         addr = (uint64_t)mbuf - (uint64_t)umem->buffer -
 516                                         umem->mb_pool->header_size;
 517                         offset = rte_pktmbuf_mtod(mbuf, uint64_t) -
 518                                         (uint64_t)mbuf +
 519                                         umem->mb_pool->header_size;
 520                         offset = offset << XSK_UNALIGNED_BUF_OFFSET_SHIFT;
 521                         desc->addr = addr | offset;
 522                         count++;
 523                 } else {
 524                         struct rte_mbuf *local_mbuf =
 525                                         rte_pktmbuf_alloc(umem->mb_pool);
 526                         void *pkt;
 527
 528                         if (local_mbuf == NULL)
 529                                 goto out;
 530
 531                         if (!xsk_ring_prod__reserve(&txq->tx, 1, &idx_tx)) {
 532                                 rte_pktmbuf_free(local_mbuf);
 533                                 goto out;
 534                         }
 535
 536                         desc = xsk_ring_prod__tx_desc(&txq->tx, idx_tx);
 537                         desc->len = mbuf->pkt_len;
 538
 539                         addr = (uint64_t)local_mbuf - (uint64_t)umem->buffer -
 540                                         umem->mb_pool->header_size;
 541                         offset = rte_pktmbuf_mtod(local_mbuf, uint64_t) -
 542                                         (uint64_t)local_mbuf +
 543                                         umem->mb_pool->header_size;
 544                         pkt = xsk_umem__get_data(umem->buffer, addr + offset);
 545                         offset = offset << XSK_UNALIGNED_BUF_OFFSET_SHIFT;
 546                         desc->addr = addr | offset;
 547                         rte_memcpy(pkt, rte_pktmbuf_mtod(mbuf, void *),
 548                                         desc->len);
 549                         rte_pktmbuf_free(mbuf);
 550                         count++;
 551                 }
 552
 553                 tx_bytes += mbuf->pkt_len;
 554         }
 555
 556 out:
 557         xsk_ring_prod__submit(&txq->tx, count);
 558         kick_tx(txq, cq);
 559
 560         txq->stats.tx_pkts += count;
 561         txq->stats.tx_bytes += tx_bytes;
 562         txq->stats.tx_dropped += nb_pkts - count;
 563
 564         return count;
 565 }
 566 #else
 567 static uint16_t
 568 af_xdp_tx_cp(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 569 {
 570         struct pkt_tx_queue *txq = queue;
 571         struct xsk_umem_info *umem = txq->umem;
 572         struct rte_mbuf *mbuf;
 573         void *addrs[ETH_AF_XDP_TX_BATCH_SIZE];
 574         unsigned long tx_bytes = 0;
 575         int i;
 576         uint32_t idx_tx;
 577         struct xsk_ring_cons *cq = &txq->pair->cq;
 578
 579         pull_umem_cq(umem, nb_pkts, cq);
 580
 581         nb_pkts = rte_ring_dequeue_bulk(umem->buf_ring, addrs,
 582                                         nb_pkts, NULL);
 583         if (nb_pkts == 0)
 584                 return 0;
 585
 586         if (xsk_ring_prod__reserve(&txq->tx, nb_pkts, &idx_tx) != nb_pkts) {
 587                 kick_tx(txq, cq);
 588                 rte_ring_enqueue_bulk(umem->buf_ring, addrs, nb_pkts, NULL);
 589                 return 0;
 590         }
 591
 592         for (i = 0; i < nb_pkts; i++) {
 593                 struct xdp_desc *desc;
 594                 void *pkt;
 595
 596                 desc = xsk_ring_prod__tx_desc(&txq->tx, idx_tx + i);
 597                 mbuf = bufs[i];
 598                 desc->len = mbuf->pkt_len;
 599
 600                 desc->addr = (uint64_t)addrs[i];
 601                 pkt = xsk_umem__get_data(umem->mz->addr,
 602                                          desc->addr);
 603                 rte_memcpy(pkt, rte_pktmbuf_mtod(mbuf, void *), desc->len);
 604                 tx_bytes += mbuf->pkt_len;
 605                 rte_pktmbuf_free(mbuf);
 606         }
 607
 608         xsk_ring_prod__submit(&txq->tx, nb_pkts);
 609
 610         kick_tx(txq, cq);
 611
 612         txq->stats.tx_pkts += nb_pkts;
 613         txq->stats.tx_bytes += tx_bytes;
 614
 615         return nb_pkts;
 616 }
 617
 618 static uint16_t
 619 af_xdp_tx_cp_batch(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 620 {
 621         uint16_t nb_tx;
 622
 623         if (likely(nb_pkts <= ETH_AF_XDP_TX_BATCH_SIZE))
 624                 return af_xdp_tx_cp(queue, bufs, nb_pkts);
 625
 626         nb_tx = 0;
 627         while (nb_pkts) {
 628                 uint16_t ret, n;
 629
 630                 /* Split larger batch into smaller batches of size
 631                  * ETH_AF_XDP_TX_BATCH_SIZE or less.
 632                  */
 633                 n = (uint16_t)RTE_MIN(nb_pkts, ETH_AF_XDP_TX_BATCH_SIZE);
 634                 ret = af_xdp_tx_cp(queue, &bufs[nb_tx], n);
 635                 nb_tx = (uint16_t)(nb_tx + ret);
 636                 nb_pkts = (uint16_t)(nb_pkts - ret);
 637                 if (ret < n)
 638                         break;
 639         }
 640
 641         return nb_tx;
 642 }
 643 #endif
 644
 645 static uint16_t
 646 eth_af_xdp_tx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 647 {
 648 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 649         return af_xdp_tx_zc(queue, bufs, nb_pkts);
 650 #else
 651         return af_xdp_tx_cp_batch(queue, bufs, nb_pkts);
 652 #endif
 653 }
 654
 655 static int
 656 eth_dev_start(struct rte_eth_dev *dev)
 657 {
 658         dev->data->dev_link.link_status = RTE_ETH_LINK_UP;
 659
 660         return 0;
 661 }
 662
 663 /* This function gets called when the current port gets stopped. */
 664 static int
 665 eth_dev_stop(struct rte_eth_dev *dev)
 666 {
 667         dev->data->dev_link.link_status = RTE_ETH_LINK_DOWN;
 668         return 0;
 669 }
 670
 671 /* Find ethdev in list */
 672 static inline struct internal_list *
 673 find_internal_resource(struct pmd_internals *port_int)
 674 {
 675         int found = 0;
 676         struct internal_list *list = NULL;
 677
 678         if (port_int == NULL)
 679                 return NULL;
 680
 681         pthread_mutex_lock(&internal_list_lock);
 682
 683         TAILQ_FOREACH(list, &internal_list, next) {
 684                 struct pmd_internals *list_int =
 685                                 list->eth_dev->data->dev_private;
 686                 if (list_int == port_int) {
 687                         found = 1;
 688                         break;
 689                 }
 690         }
 691
 692         pthread_mutex_unlock(&internal_list_lock);
 693
 694         if (!found)
 695                 return NULL;
 696
 697         return list;
 698 }
 699
 700 /* Check if the netdev,qid context already exists */
 701 static inline bool
 702 ctx_exists(struct pkt_rx_queue *rxq, const char *ifname,
 703                 struct pkt_rx_queue *list_rxq, const char *list_ifname)
 704 {
 705         bool exists = false;
 706
 707         if (rxq->xsk_queue_idx == list_rxq->xsk_queue_idx &&
 708                         !strncmp(ifname, list_ifname, IFNAMSIZ)) {
 709                 AF_XDP_LOG(ERR, "ctx %s,%i already exists, cannot share umem\n",
 710                                         ifname, rxq->xsk_queue_idx);
 711                 exists = true;
 712         }
 713
 714         return exists;
 715 }
 716
 717 /* Get a pointer to an existing UMEM which overlays the rxq's mb_pool */
 718 static inline int
 719 get_shared_umem(struct pkt_rx_queue *rxq, const char *ifname,
 720                         struct xsk_umem_info **umem)
 721 {
 722         struct internal_list *list;
 723         struct pmd_internals *internals;
 724         int i = 0, ret = 0;
 725         struct rte_mempool *mb_pool = rxq->mb_pool;
 726
 727         if (mb_pool == NULL)
 728                 return ret;
 729
 730         pthread_mutex_lock(&internal_list_lock);
 731
 732         TAILQ_FOREACH(list, &internal_list, next) {
 733                 internals = list->eth_dev->data->dev_private;
 734                 for (i = 0; i < internals->queue_cnt; i++) {
 735                         struct pkt_rx_queue *list_rxq =
 736                                                 &internals->rx_queues[i];
 737                         if (rxq == list_rxq)
 738                                 continue;
 739                         if (mb_pool == internals->rx_queues[i].mb_pool) {
 740                                 if (ctx_exists(rxq, ifname, list_rxq,
 741                                                 internals->if_name)) {
 742                                         ret = -1;
 743                                         goto out;
 744                                 }
 745                                 if (__atomic_load_n(
 746                                         &internals->rx_queues[i].umem->refcnt,
 747                                                         __ATOMIC_ACQUIRE)) {
 748                                         *umem = internals->rx_queues[i].umem;
 749                                         goto out;
 750                                 }
 751                         }
 752                 }
 753         }
 754
 755 out:
 756         pthread_mutex_unlock(&internal_list_lock);
 757
 758         return ret;
 759 }
 760
 761 static int
 762 eth_dev_configure(struct rte_eth_dev *dev)
 763 {
 764         struct pmd_internals *internal = dev->data->dev_private;
 765
 766         /* rx/tx must be paired */
 767         if (dev->data->nb_rx_queues != dev->data->nb_tx_queues)
 768                 return -EINVAL;
 769
 770         if (internal->shared_umem) {
 771                 struct internal_list *list = NULL;
 772                 const char *name = dev->device->name;
 773
 774                 /* Ensure PMD is not already inserted into the list */
 775                 list = find_internal_resource(internal);
 776                 if (list)
 777                         return 0;
 778
 779                 list = rte_zmalloc_socket(name, sizeof(*list), 0,
 780                                         dev->device->numa_node);
 781                 if (list == NULL)
 782                         return -1;
 783
 784                 list->eth_dev = dev;
 785                 pthread_mutex_lock(&internal_list_lock);
 786                 TAILQ_INSERT_TAIL(&internal_list, list, next);
 787                 pthread_mutex_unlock(&internal_list_lock);
 788         }
 789
 790         return 0;
 791 }
 792
 793 #define CLB_VAL_IDX 0
 794 static int
 795 eth_monitor_callback(const uint64_t value,
 796                 const uint64_t opaque[RTE_POWER_MONITOR_OPAQUE_SZ])
 797 {
 798         const uint64_t v = opaque[CLB_VAL_IDX];
 799         const uint64_t m = (uint32_t)~0;
 800
 801         /* if the value has changed, abort entering power optimized state */
 802         return (value & m) == v ? 0 : -1;
 803 }
 804
 805 static int
 806 eth_get_monitor_addr(void *rx_queue, struct rte_power_monitor_cond *pmc)
 807 {
 808         struct pkt_rx_queue *rxq = rx_queue;
 809         unsigned int *prod = rxq->rx.producer;
 810         const uint32_t cur_val = rxq->rx.cached_prod; /* use cached value */
 811
 812         /* watch for changes in producer ring */
 813         pmc->addr = (void *)prod;
 814
 815         /* store current value */
 816         pmc->opaque[CLB_VAL_IDX] = cur_val;
 817         pmc->fn = eth_monitor_callback;
 818
 819         /* AF_XDP producer ring index is 32-bit */
 820         pmc->size = sizeof(uint32_t);
 821
 822         return 0;
 823 }
 824
 825 static int
 826 eth_dev_info(struct rte_eth_dev *dev, struct rte_eth_dev_info *dev_info)
 827 {
 828         struct pmd_internals *internals = dev->data->dev_private;
 829
 830         dev_info->if_index = internals->if_index;
 831         dev_info->max_mac_addrs = 1;
 832         dev_info->max_rx_queues = internals->queue_cnt;
 833         dev_info->max_tx_queues = internals->queue_cnt;
 834
 835         dev_info->min_mtu = RTE_ETHER_MIN_MTU;
 836 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 837         dev_info->max_rx_pktlen = getpagesize() -
 838                                   sizeof(struct rte_mempool_objhdr) -
 839                                   sizeof(struct rte_mbuf) -
 840                                   RTE_PKTMBUF_HEADROOM - XDP_PACKET_HEADROOM;
 841 #else
 842         dev_info->max_rx_pktlen = ETH_AF_XDP_FRAME_SIZE - XDP_PACKET_HEADROOM;
 843 #endif
 844         dev_info->max_mtu = dev_info->max_rx_pktlen - ETH_AF_XDP_ETH_OVERHEAD;
 845
 846         dev_info->default_rxportconf.burst_size = ETH_AF_XDP_DFLT_BUSY_BUDGET;
 847         dev_info->default_txportconf.burst_size = ETH_AF_XDP_DFLT_BUSY_BUDGET;
 848         dev_info->default_rxportconf.nb_queues = 1;
 849         dev_info->default_txportconf.nb_queues = 1;
 850         dev_info->default_rxportconf.ring_size = ETH_AF_XDP_DFLT_NUM_DESCS;
 851         dev_info->default_txportconf.ring_size = ETH_AF_XDP_DFLT_NUM_DESCS;
 852
 853         return 0;
 854 }
 855
 856 static int
 857 eth_stats_get(struct rte_eth_dev *dev, struct rte_eth_stats *stats)
 858 {
 859         struct pmd_internals *internals = dev->data->dev_private;
 860         struct xdp_statistics xdp_stats;
 861         struct pkt_rx_queue *rxq;
 862         struct pkt_tx_queue *txq;
 863         socklen_t optlen;
 864         int i, ret;
 865
 866         for (i = 0; i < dev->data->nb_rx_queues; i++) {
 867                 optlen = sizeof(struct xdp_statistics);
 868                 rxq = &internals->rx_queues[i];
 869                 txq = rxq->pair;
 870                 stats->q_ipackets[i] = rxq->stats.rx_pkts;
 871                 stats->q_ibytes[i] = rxq->stats.rx_bytes;
 872
 873                 stats->q_opackets[i] = txq->stats.tx_pkts;
 874                 stats->q_obytes[i] = txq->stats.tx_bytes;
 875
 876                 stats->ipackets += stats->q_ipackets[i];
 877                 stats->ibytes += stats->q_ibytes[i];
 878                 stats->imissed += rxq->stats.rx_dropped;
 879                 stats->oerrors += txq->stats.tx_dropped;
 880                 ret = getsockopt(xsk_socket__fd(rxq->xsk), SOL_XDP,
 881                                 XDP_STATISTICS, &xdp_stats, &optlen);
 882                 if (ret != 0) {
 883                         AF_XDP_LOG(ERR, "getsockopt() failed for XDP_STATISTICS.\n");
 884                         return -1;
 885                 }
 886                 stats->imissed += xdp_stats.rx_dropped;
 887
 888                 stats->opackets += stats->q_opackets[i];
 889                 stats->obytes += stats->q_obytes[i];
 890         }
 891
 892         return 0;
 893 }
 894
 895 static int
 896 eth_stats_reset(struct rte_eth_dev *dev)
 897 {
 898         struct pmd_internals *internals = dev->data->dev_private;
 899         int i;
 900
 901         for (i = 0; i < internals->queue_cnt; i++) {
 902                 memset(&internals->rx_queues[i].stats, 0,
 903                                         sizeof(struct rx_stats));
 904                 memset(&internals->tx_queues[i].stats, 0,
 905                                         sizeof(struct tx_stats));
 906         }
 907
 908         return 0;
 909 }
 910
 911 static void
 912 remove_xdp_program(struct pmd_internals *internals)
 913 {
 914         uint32_t curr_prog_id = 0;
 915
 916         if (bpf_get_link_xdp_id(internals->if_index, &curr_prog_id,
 917                                 XDP_FLAGS_UPDATE_IF_NOEXIST)) {
 918                 AF_XDP_LOG(ERR, "bpf_get_link_xdp_id failed\n");
 919                 return;
 920         }
 921         bpf_set_link_xdp_fd(internals->if_index, -1,
 922                         XDP_FLAGS_UPDATE_IF_NOEXIST);
 923 }
 924
 925 static void
 926 xdp_umem_destroy(struct xsk_umem_info *umem)
 927 {
 928 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 929         umem->mb_pool = NULL;
 930 #else
 931         rte_memzone_free(umem->mz);
 932         umem->mz = NULL;
 933
 934         rte_ring_free(umem->buf_ring);
 935         umem->buf_ring = NULL;
 936 #endif
 937
 938         rte_free(umem);
 939 }
 940
 941 static int
 942 eth_dev_close(struct rte_eth_dev *dev)
 943 {
 944         struct pmd_internals *internals = dev->data->dev_private;
 945         struct pkt_rx_queue *rxq;
 946         int i;
 947
 948         if (rte_eal_process_type() != RTE_PROC_PRIMARY)
 949                 return 0;
 950
 951         AF_XDP_LOG(INFO, "Closing AF_XDP ethdev on numa socket %u\n",
 952                 rte_socket_id());
 953
 954         for (i = 0; i < internals->queue_cnt; i++) {
 955                 rxq = &internals->rx_queues[i];
 956                 if (rxq->umem == NULL)
 957                         break;
 958                 xsk_socket__delete(rxq->xsk);
 959
 960                 if (__atomic_sub_fetch(&rxq->umem->refcnt, 1, __ATOMIC_ACQUIRE)
 961                                 == 0) {
 962                         (void)xsk_umem__delete(rxq->umem->umem);
 963                         xdp_umem_destroy(rxq->umem);
 964                 }
 965
 966                 /* free pkt_tx_queue */
 967                 rte_free(rxq->pair);
 968                 rte_free(rxq);
 969         }
 970
 971         /*
 972          * MAC is not allocated dynamically, setting it to NULL would prevent
 973          * from releasing it in rte_eth_dev_release_port.
 974          */
 975         dev->data->mac_addrs = NULL;
 976
 977         remove_xdp_program(internals);
 978
 979         if (internals->shared_umem) {
 980                 struct internal_list *list;
 981
 982                 /* Remove ethdev from list used to track and share UMEMs */
 983                 list = find_internal_resource(internals);
 984                 if (list) {
 985                         pthread_mutex_lock(&internal_list_lock);
 986                         TAILQ_REMOVE(&internal_list, list, next);
 987                         pthread_mutex_unlock(&internal_list_lock);
 988                         rte_free(list);
 989                 }
 990         }
 991
 992         return 0;
 993 }
 994
 995 static int
 996 eth_link_update(struct rte_eth_dev *dev __rte_unused,
 997                 int wait_to_complete __rte_unused)
 998 {
 999         return 0;
1000 }
1001
1002 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
1003 static inline uintptr_t get_base_addr(struct rte_mempool *mp, uint64_t *align)
1004 {
1005         struct rte_mempool_memhdr *memhdr;
1006         uintptr_t memhdr_addr, aligned_addr;
1007
1008         memhdr = STAILQ_FIRST(&mp->mem_list);
1009         memhdr_addr = (uintptr_t)memhdr->addr;
1010         aligned_addr = memhdr_addr & ~(getpagesize() - 1);
1011         *align = memhdr_addr - aligned_addr;
1012
1013         return aligned_addr;
1014 }
1015
1016 static struct
1017 xsk_umem_info *xdp_umem_configure(struct pmd_internals *internals,
1018                                   struct pkt_rx_queue *rxq)
1019 {
1020         struct xsk_umem_info *umem = NULL;
1021         int ret;
1022         struct xsk_umem_config usr_config = {
1023                 .fill_size = ETH_AF_XDP_DFLT_NUM_DESCS * 2,
1024                 .comp_size = ETH_AF_XDP_DFLT_NUM_DESCS,
1025                 .flags = XDP_UMEM_UNALIGNED_CHUNK_FLAG};
1026         void *base_addr = NULL;
1027         struct rte_mempool *mb_pool = rxq->mb_pool;
1028         uint64_t umem_size, align = 0;
1029
1030         if (internals->shared_umem) {
1031                 if (get_shared_umem(rxq, internals->if_name, &umem) < 0)
1032                         return NULL;
1033
1034                 if (umem != NULL &&
1035                         __atomic_load_n(&umem->refcnt, __ATOMIC_ACQUIRE) <
1036                                         umem->max_xsks) {
1037                         AF_XDP_LOG(INFO, "%s,qid%i sharing UMEM\n",
1038                                         internals->if_name, rxq->xsk_queue_idx);
1039                         __atomic_fetch_add(&umem->refcnt, 1, __ATOMIC_ACQUIRE);
1040                 }
1041         }
1042
1043         if (umem == NULL) {
1044                 usr_config.frame_size =
1045                         rte_mempool_calc_obj_size(mb_pool->elt_size,
1046                                                   mb_pool->flags, NULL);
1047                 usr_config.frame_headroom = mb_pool->header_size +
1048                                                 sizeof(struct rte_mbuf) +
1049                                                 rte_pktmbuf_priv_size(mb_pool) +
1050                                                 RTE_PKTMBUF_HEADROOM;
1051
1052                 umem = rte_zmalloc_socket("umem", sizeof(*umem), 0,
1053                                           rte_socket_id());
1054                 if (umem == NULL) {
1055                         AF_XDP_LOG(ERR, "Failed to allocate umem info");
1056                         return NULL;
1057                 }
1058
1059                 umem->mb_pool = mb_pool;
1060                 base_addr = (void *)get_base_addr(mb_pool, &align);
1061                 umem_size = (uint64_t)mb_pool->populated_size *
1062                                 (uint64_t)usr_config.frame_size +
1063                                 align;
1064
1065                 ret = xsk_umem__create(&umem->umem, base_addr, umem_size,
1066                                 &rxq->fq, &rxq->cq, &usr_config);
1067                 if (ret) {
1068                         AF_XDP_LOG(ERR, "Failed to create umem");
1069                         goto err;
1070                 }
1071                 umem->buffer = base_addr;
1072
1073                 if (internals->shared_umem) {
1074                         umem->max_xsks = mb_pool->populated_size /
1075                                                 ETH_AF_XDP_NUM_BUFFERS;
1076                         AF_XDP_LOG(INFO, "Max xsks for UMEM %s: %u\n",
1077                                                 mb_pool->name, umem->max_xsks);
1078                 }
1079
1080                 __atomic_store_n(&umem->refcnt, 1, __ATOMIC_RELEASE);
1081         }
1082
1083 #else
1084 static struct
1085 xsk_umem_info *xdp_umem_configure(struct pmd_internals *internals,
1086                                   struct pkt_rx_queue *rxq)
1087 {
1088         struct xsk_umem_info *umem;
1089         const struct rte_memzone *mz;
1090         struct xsk_umem_config usr_config = {
1091                 .fill_size = ETH_AF_XDP_DFLT_NUM_DESCS,
1092                 .comp_size = ETH_AF_XDP_DFLT_NUM_DESCS,
1093                 .frame_size = ETH_AF_XDP_FRAME_SIZE,
1094                 .frame_headroom = 0 };
1095         char ring_name[RTE_RING_NAMESIZE];
1096         char mz_name[RTE_MEMZONE_NAMESIZE];
1097         int ret;
1098         uint64_t i;
1099
1100         umem = rte_zmalloc_socket("umem", sizeof(*umem), 0, rte_socket_id());
1101         if (umem == NULL) {
1102                 AF_XDP_LOG(ERR, "Failed to allocate umem info");
1103                 return NULL;
1104         }
1105
1106         snprintf(ring_name, sizeof(ring_name), "af_xdp_ring_%s_%u",
1107                        internals->if_name, rxq->xsk_queue_idx);
1108         umem->buf_ring = rte_ring_create(ring_name,
1109                                          ETH_AF_XDP_NUM_BUFFERS,
1110                                          rte_socket_id(),
1111                                          0x0);
1112         if (umem->buf_ring == NULL) {
1113                 AF_XDP_LOG(ERR, "Failed to create rte_ring\n");
1114                 goto err;
1115         }
1116
1117         for (i = 0; i < ETH_AF_XDP_NUM_BUFFERS; i++)
1118                 rte_ring_enqueue(umem->buf_ring,
1119                                  (void *)(i * ETH_AF_XDP_FRAME_SIZE));
1120
1121         snprintf(mz_name, sizeof(mz_name), "af_xdp_umem_%s_%u",
1122                        internals->if_name, rxq->xsk_queue_idx);
1123         mz = rte_memzone_reserve_aligned(mz_name,
1124                         ETH_AF_XDP_NUM_BUFFERS * ETH_AF_XDP_FRAME_SIZE,
1125                         rte_socket_id(), RTE_MEMZONE_IOVA_CONTIG,
1126                         getpagesize());
1127         if (mz == NULL) {
1128                 AF_XDP_LOG(ERR, "Failed to reserve memzone for af_xdp umem.\n");
1129                 goto err;
1130         }
1131
1132         ret = xsk_umem__create(&umem->umem, mz->addr,
1133                                ETH_AF_XDP_NUM_BUFFERS * ETH_AF_XDP_FRAME_SIZE,
1134                                &rxq->fq, &rxq->cq,
1135                                &usr_config);
1136
1137         if (ret) {
1138                 AF_XDP_LOG(ERR, "Failed to create umem");
1139                 goto err;
1140         }
1141         umem->mz = mz;
1142
1143 #endif
1144         return umem;
1145
1146 err:
1147         xdp_umem_destroy(umem);
1148         return NULL;
1149 }
1150
1151 static int
1152 load_custom_xdp_prog(const char *prog_path, int if_index, struct bpf_map **map)
1153 {
1154         int ret, prog_fd = -1;
1155         struct bpf_object *obj;
1156
1157         ret = bpf_prog_load(prog_path, BPF_PROG_TYPE_XDP, &obj, &prog_fd);
1158         if (ret) {
1159                 AF_XDP_LOG(ERR, "Failed to load program %s\n", prog_path);
1160                 return ret;
1161         }
1162
1163         /*
1164          * The loaded program must provision for a map of xsks, such that some
1165          * traffic can be redirected to userspace.
1166          */
1167         *map = bpf_object__find_map_by_name(obj, "xsks_map");
1168         if (!*map) {
1169                 AF_XDP_LOG(ERR, "Failed to find xsks_map in %s\n", prog_path);
1170                 return -1;
1171         }
1172
1173         /* Link the program with the given network device */
1174         ret = bpf_set_link_xdp_fd(if_index, prog_fd,
1175                                         XDP_FLAGS_UPDATE_IF_NOEXIST);
1176         if (ret) {
1177                 AF_XDP_LOG(ERR, "Failed to set prog fd %d on interface\n",
1178                                 prog_fd);
1179                 return -1;
1180         }
1181
1182         AF_XDP_LOG(INFO, "Successfully loaded XDP program %s with fd %d\n",
1183                                 prog_path, prog_fd);
1184
1185         return 0;
1186 }
1187
1188 /* Detect support for busy polling through setsockopt(). */
1189 static int
1190 configure_preferred_busy_poll(struct pkt_rx_queue *rxq)
1191 {
1192         int sock_opt = 1;
1193         int fd = xsk_socket__fd(rxq->xsk);
1194         int ret = 0;
1195
1196         ret = setsockopt(fd, SOL_SOCKET, SO_PREFER_BUSY_POLL,
1197                         (void *)&sock_opt, sizeof(sock_opt));
1198         if (ret < 0) {
1199                 AF_XDP_LOG(DEBUG, "Failed to set SO_PREFER_BUSY_POLL\n");
1200                 goto err_prefer;
1201         }
1202
1203         sock_opt = ETH_AF_XDP_DFLT_BUSY_TIMEOUT;
1204         ret = setsockopt(fd, SOL_SOCKET, SO_BUSY_POLL, (void *)&sock_opt,
1205                         sizeof(sock_opt));
1206         if (ret < 0) {
1207                 AF_XDP_LOG(DEBUG, "Failed to set SO_BUSY_POLL\n");
1208                 goto err_timeout;
1209         }
1210
1211         sock_opt = rxq->busy_budget;
1212         ret = setsockopt(fd, SOL_SOCKET, SO_BUSY_POLL_BUDGET,
1213                         (void *)&sock_opt, sizeof(sock_opt));
1214         if (ret < 0) {
1215                 AF_XDP_LOG(DEBUG, "Failed to set SO_BUSY_POLL_BUDGET\n");
1216         } else {
1217                 AF_XDP_LOG(INFO, "Busy polling budget set to: %u\n",
1218                                         rxq->busy_budget);
1219                 return 0;
1220         }
1221
1222         /* setsockopt failure - attempt to restore xsk to default state and
1223          * proceed without busy polling support.
1224          */
1225         sock_opt = 0;
1226         ret = setsockopt(fd, SOL_SOCKET, SO_BUSY_POLL, (void *)&sock_opt,
1227                         sizeof(sock_opt));
1228         if (ret < 0) {
1229                 AF_XDP_LOG(ERR, "Failed to unset SO_BUSY_POLL\n");
1230                 return -1;
1231         }
1232
1233 err_timeout:
1234         sock_opt = 0;
1235         ret = setsockopt(fd, SOL_SOCKET, SO_PREFER_BUSY_POLL,
1236                         (void *)&sock_opt, sizeof(sock_opt));
1237         if (ret < 0) {
1238                 AF_XDP_LOG(ERR, "Failed to unset SO_PREFER_BUSY_POLL\n");
1239                 return -1;
1240         }
1241
1242 err_prefer:
1243         rxq->busy_budget = 0;
1244         return 0;
1245 }
1246
1247 static int
1248 xsk_configure(struct pmd_internals *internals, struct pkt_rx_queue *rxq,
1249               int ring_size)
1250 {
1251         struct xsk_socket_config cfg;
1252         struct pkt_tx_queue *txq = rxq->pair;
1253         int ret = 0;
1254         int reserve_size = ETH_AF_XDP_DFLT_NUM_DESCS;
1255         struct rte_mbuf *fq_bufs[reserve_size];
1256
1257         rxq->umem = xdp_umem_configure(internals, rxq);
1258         if (rxq->umem == NULL)
1259                 return -ENOMEM;
1260         txq->umem = rxq->umem;
1261
1262         cfg.rx_size = ring_size;
1263         cfg.tx_size = ring_size;
1264         cfg.libbpf_flags = 0;
1265         cfg.xdp_flags = XDP_FLAGS_UPDATE_IF_NOEXIST;
1266         cfg.bind_flags = 0;
1267
1268 #if defined(XDP_USE_NEED_WAKEUP)
1269         cfg.bind_flags |= XDP_USE_NEED_WAKEUP;
1270 #endif
1271
1272         if (strnlen(internals->prog_path, PATH_MAX) &&
1273                                 !internals->custom_prog_configured) {
1274                 ret = load_custom_xdp_prog(internals->prog_path,
1275                                            internals->if_index,
1276                                            &internals->map);
1277                 if (ret) {
1278                         AF_XDP_LOG(ERR, "Failed to load custom XDP program %s\n",
1279                                         internals->prog_path);
1280                         goto err;
1281                 }
1282                 internals->custom_prog_configured = 1;
1283                 cfg.libbpf_flags = XSK_LIBBPF_FLAGS__INHIBIT_PROG_LOAD;
1284         }
1285
1286         if (internals->shared_umem)
1287                 ret = create_shared_socket(&rxq->xsk, internals->if_name,
1288                                 rxq->xsk_queue_idx, rxq->umem->umem, &rxq->rx,
1289                                 &txq->tx, &rxq->fq, &rxq->cq, &cfg);
1290         else
1291                 ret = xsk_socket__create(&rxq->xsk, internals->if_name,
1292                                 rxq->xsk_queue_idx, rxq->umem->umem, &rxq->rx,
1293                                 &txq->tx, &cfg);
1294
1295         if (ret) {
1296                 AF_XDP_LOG(ERR, "Failed to create xsk socket.\n");
1297                 goto err;
1298         }
1299
1300         /* insert the xsk into the xsks_map */
1301         if (internals->custom_prog_configured) {
1302                 int err, fd;
1303
1304                 fd = xsk_socket__fd(rxq->xsk);
1305                 err = bpf_map_update_elem(bpf_map__fd(internals->map),
1306                                           &rxq->xsk_queue_idx, &fd, 0);
1307                 if (err) {
1308                         AF_XDP_LOG(ERR, "Failed to insert xsk in map.\n");
1309                         goto err;
1310                 }
1311         }
1312
1313 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
1314         ret = rte_pktmbuf_alloc_bulk(rxq->umem->mb_pool, fq_bufs, reserve_size);
1315         if (ret) {
1316                 AF_XDP_LOG(DEBUG, "Failed to get enough buffers for fq.\n");
1317                 goto err;
1318         }
1319 #endif
1320
1321         if (rxq->busy_budget) {
1322                 ret = configure_preferred_busy_poll(rxq);
1323                 if (ret) {
1324                         AF_XDP_LOG(ERR, "Failed configure busy polling.\n");
1325                         goto err;
1326                 }
1327         }
1328
1329         ret = reserve_fill_queue(rxq->umem, reserve_size, fq_bufs, &rxq->fq);
1330         if (ret) {
1331                 xsk_socket__delete(rxq->xsk);
1332                 AF_XDP_LOG(ERR, "Failed to reserve fill queue.\n");
1333                 goto err;
1334         }
1335
1336         return 0;
1337
1338 err:
1339         if (__atomic_sub_fetch(&rxq->umem->refcnt, 1, __ATOMIC_ACQUIRE) == 0)
1340                 xdp_umem_destroy(rxq->umem);
1341
1342         return ret;
1343 }
1344
1345 static int
1346 eth_rx_queue_setup(struct rte_eth_dev *dev,
1347                    uint16_t rx_queue_id,
1348                    uint16_t nb_rx_desc,
1349                    unsigned int socket_id __rte_unused,
1350                    const struct rte_eth_rxconf *rx_conf __rte_unused,
1351                    struct rte_mempool *mb_pool)
1352 {
1353         struct pmd_internals *internals = dev->data->dev_private;
1354         struct pkt_rx_queue *rxq;
1355         int ret;
1356
1357         rxq = &internals->rx_queues[rx_queue_id];
1358
1359         AF_XDP_LOG(INFO, "Set up rx queue, rx queue id: %d, xsk queue id: %d\n",
1360                    rx_queue_id, rxq->xsk_queue_idx);
1361
1362 #ifndef XDP_UMEM_UNALIGNED_CHUNK_FLAG
1363         uint32_t buf_size, data_size;
1364
1365         /* Now get the space available for data in the mbuf */
1366         buf_size = rte_pktmbuf_data_room_size(mb_pool) -
1367                 RTE_PKTMBUF_HEADROOM;
1368         data_size = ETH_AF_XDP_FRAME_SIZE;
1369
1370         if (data_size > buf_size) {
1371                 AF_XDP_LOG(ERR, "%s: %d bytes will not fit in mbuf (%d bytes)\n",
1372                         dev->device->name, data_size, buf_size);
1373                 ret = -ENOMEM;
1374                 goto err;
1375         }
1376 #endif
1377
1378         rxq->mb_pool = mb_pool;
1379
1380         if (xsk_configure(internals, rxq, nb_rx_desc)) {
1381                 AF_XDP_LOG(ERR, "Failed to configure xdp socket\n");
1382                 ret = -EINVAL;
1383                 goto err;
1384         }
1385
1386         if (!rxq->busy_budget)
1387                 AF_XDP_LOG(DEBUG, "Preferred busy polling not enabled\n");
1388
1389         rxq->fds[0].fd = xsk_socket__fd(rxq->xsk);
1390         rxq->fds[0].events = POLLIN;
1391
1392         dev->data->rx_queues[rx_queue_id] = rxq;
1393         return 0;
1394
1395 err:
1396         return ret;
1397 }
1398
1399 static int
1400 eth_tx_queue_setup(struct rte_eth_dev *dev,
1401                    uint16_t tx_queue_id,
1402                    uint16_t nb_tx_desc __rte_unused,
1403                    unsigned int socket_id __rte_unused,
1404                    const struct rte_eth_txconf *tx_conf __rte_unused)
1405 {
1406         struct pmd_internals *internals = dev->data->dev_private;
1407         struct pkt_tx_queue *txq;
1408
1409         txq = &internals->tx_queues[tx_queue_id];
1410
1411         dev->data->tx_queues[tx_queue_id] = txq;
1412         return 0;
1413 }
1414
1415 static int
1416 eth_dev_mtu_set(struct rte_eth_dev *dev, uint16_t mtu)
1417 {
1418         struct pmd_internals *internals = dev->data->dev_private;
1419         struct ifreq ifr = { .ifr_mtu = mtu };
1420         int ret;
1421         int s;
1422
1423         s = socket(PF_INET, SOCK_DGRAM, 0);
1424         if (s < 0)
1425                 return -EINVAL;
1426
1427         strlcpy(ifr.ifr_name, internals->if_name, IFNAMSIZ);
1428         ret = ioctl(s, SIOCSIFMTU, &ifr);
1429         close(s);
1430
1431         return (ret < 0) ? -errno : 0;
1432 }
1433
1434 static int
1435 eth_dev_change_flags(char *if_name, uint32_t flags, uint32_t mask)
1436 {
1437         struct ifreq ifr;
1438         int ret = 0;
1439         int s;
1440
1441         s = socket(PF_INET, SOCK_DGRAM, 0);
1442         if (s < 0)
1443                 return -errno;
1444
1445         strlcpy(ifr.ifr_name, if_name, IFNAMSIZ);
1446         if (ioctl(s, SIOCGIFFLAGS, &ifr) < 0) {
1447                 ret = -errno;
1448                 goto out;
1449         }
1450         ifr.ifr_flags &= mask;
1451         ifr.ifr_flags |= flags;
1452         if (ioctl(s, SIOCSIFFLAGS, &ifr) < 0) {
1453                 ret = -errno;
1454                 goto out;
1455         }
1456 out:
1457         close(s);
1458         return ret;
1459 }
1460
1461 static int
1462 eth_dev_promiscuous_enable(struct rte_eth_dev *dev)
1463 {
1464         struct pmd_internals *internals = dev->data->dev_private;
1465
1466         return eth_dev_change_flags(internals->if_name, IFF_PROMISC, ~0);
1467 }
1468
1469 static int
1470 eth_dev_promiscuous_disable(struct rte_eth_dev *dev)
1471 {
1472         struct pmd_internals *internals = dev->data->dev_private;
1473
1474         return eth_dev_change_flags(internals->if_name, 0, ~IFF_PROMISC);
1475 }
1476
1477 static const struct eth_dev_ops ops = {
1478         .dev_start = eth_dev_start,
1479         .dev_stop = eth_dev_stop,
1480         .dev_close = eth_dev_close,
1481         .dev_configure = eth_dev_configure,
1482         .dev_infos_get = eth_dev_info,
1483         .mtu_set = eth_dev_mtu_set,
1484         .promiscuous_enable = eth_dev_promiscuous_enable,
1485         .promiscuous_disable = eth_dev_promiscuous_disable,
1486         .rx_queue_setup = eth_rx_queue_setup,
1487         .tx_queue_setup = eth_tx_queue_setup,
1488         .link_update = eth_link_update,
1489         .stats_get = eth_stats_get,
1490         .stats_reset = eth_stats_reset,
1491         .get_monitor_addr = eth_get_monitor_addr,
1492 };
1493
1494 /** parse busy_budget argument */
1495 static int
1496 parse_budget_arg(const char *key __rte_unused,
1497                   const char *value, void *extra_args)
1498 {
1499         int *i = (int *)extra_args;
1500         char *end;
1501
1502         *i = strtol(value, &end, 10);
1503         if (*i < 0 || *i > UINT16_MAX) {
1504                 AF_XDP_LOG(ERR, "Invalid busy_budget %i, must be >= 0 and <= %u\n",
1505                                 *i, UINT16_MAX);
1506                 return -EINVAL;
1507         }
1508
1509         return 0;
1510 }
1511
1512 /** parse integer from integer argument */
1513 static int
1514 parse_integer_arg(const char *key __rte_unused,
1515                   const char *value, void *extra_args)
1516 {
1517         int *i = (int *)extra_args;
1518         char *end;
1519
1520         *i = strtol(value, &end, 10);
1521         if (*i < 0) {
1522                 AF_XDP_LOG(ERR, "Argument has to be positive.\n");
1523                 return -EINVAL;
1524         }
1525
1526         return 0;
1527 }
1528
1529 /** parse name argument */
1530 static int
1531 parse_name_arg(const char *key __rte_unused,
1532                const char *value, void *extra_args)
1533 {
1534         char *name = extra_args;
1535
1536         if (strnlen(value, IFNAMSIZ) > IFNAMSIZ - 1) {
1537                 AF_XDP_LOG(ERR, "Invalid name %s, should be less than %u bytes.\n",
1538                            value, IFNAMSIZ);
1539                 return -EINVAL;
1540         }
1541
1542         strlcpy(name, value, IFNAMSIZ);
1543
1544         return 0;
1545 }
1546
1547 /** parse xdp prog argument */
1548 static int
1549 parse_prog_arg(const char *key __rte_unused,
1550                const char *value, void *extra_args)
1551 {
1552         char *path = extra_args;
1553
1554         if (strnlen(value, PATH_MAX) == PATH_MAX) {
1555                 AF_XDP_LOG(ERR, "Invalid path %s, should be less than %u bytes.\n",
1556                            value, PATH_MAX);
1557                 return -EINVAL;
1558         }
1559
1560         if (access(value, F_OK) != 0) {
1561                 AF_XDP_LOG(ERR, "Error accessing %s: %s\n",
1562                            value, strerror(errno));
1563                 return -EINVAL;
1564         }
1565
1566         strlcpy(path, value, PATH_MAX);
1567
1568         return 0;
1569 }
1570
1571 static int
1572 xdp_get_channels_info(const char *if_name, int *max_queues,
1573                                 int *combined_queues)
1574 {
1575         struct ethtool_channels channels;
1576         struct ifreq ifr;
1577         int fd, ret;
1578
1579         fd = socket(AF_INET, SOCK_DGRAM, 0);
1580         if (fd < 0)
1581                 return -1;
1582
1583         channels.cmd = ETHTOOL_GCHANNELS;
1584         ifr.ifr_data = (void *)&channels;
1585         strlcpy(ifr.ifr_name, if_name, IFNAMSIZ);
1586         ret = ioctl(fd, SIOCETHTOOL, &ifr);
1587         if (ret) {
1588                 if (errno == EOPNOTSUPP) {
1589                         ret = 0;
1590                 } else {
1591                         ret = -errno;
1592                         goto out;
1593                 }
1594         }
1595
1596         if (channels.max_combined == 0 || errno == EOPNOTSUPP) {
1597                 /* If the device says it has no channels, then all traffic
1598                  * is sent to a single stream, so max queues = 1.
1599                  */
1600                 *max_queues = 1;
1601                 *combined_queues = 1;
1602         } else {
1603                 *max_queues = channels.max_combined;
1604                 *combined_queues = channels.combined_count;
1605         }
1606
1607  out:
1608         close(fd);
1609         return ret;
1610 }
1611
1612 static int
1613 parse_parameters(struct rte_kvargs *kvlist, char *if_name, int *start_queue,
1614                         int *queue_cnt, int *shared_umem, char *prog_path,
1615                         int *busy_budget)
1616 {
1617         int ret;
1618
1619         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_IFACE_ARG,
1620                                  &parse_name_arg, if_name);
1621         if (ret < 0)
1622                 goto free_kvlist;
1623
1624         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_START_QUEUE_ARG,
1625                                  &parse_integer_arg, start_queue);
1626         if (ret < 0)
1627                 goto free_kvlist;
1628
1629         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_QUEUE_COUNT_ARG,
1630                                  &parse_integer_arg, queue_cnt);
1631         if (ret < 0 || *queue_cnt <= 0) {
1632                 ret = -EINVAL;
1633                 goto free_kvlist;
1634         }
1635
1636         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_SHARED_UMEM_ARG,
1637                                 &parse_integer_arg, shared_umem);
1638         if (ret < 0)
1639                 goto free_kvlist;
1640
1641         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_PROG_ARG,
1642                                  &parse_prog_arg, prog_path);
1643         if (ret < 0)
1644                 goto free_kvlist;
1645
1646         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_BUDGET_ARG,
1647                                 &parse_budget_arg, busy_budget);
1648         if (ret < 0)
1649                 goto free_kvlist;
1650
1651 free_kvlist:
1652         rte_kvargs_free(kvlist);
1653         return ret;
1654 }
1655
1656 static int
1657 get_iface_info(const char *if_name,
1658                struct rte_ether_addr *eth_addr,
1659                int *if_index)
1660 {
1661         struct ifreq ifr;
1662         int sock = socket(AF_INET, SOCK_DGRAM, IPPROTO_IP);
1663
1664         if (sock < 0)
1665                 return -1;
1666
1667         strlcpy(ifr.ifr_name, if_name, IFNAMSIZ);
1668         if (ioctl(sock, SIOCGIFINDEX, &ifr))
1669                 goto error;
1670
1671         *if_index = ifr.ifr_ifindex;
1672
1673         if (ioctl(sock, SIOCGIFHWADDR, &ifr))
1674                 goto error;
1675
1676         rte_memcpy(eth_addr, ifr.ifr_hwaddr.sa_data, RTE_ETHER_ADDR_LEN);
1677
1678         close(sock);
1679         return 0;
1680
1681 error:
1682         close(sock);
1683         return -1;
1684 }
1685
1686 static struct rte_eth_dev *
1687 init_internals(struct rte_vdev_device *dev, const char *if_name,
1688                 int start_queue_idx, int queue_cnt, int shared_umem,
1689                 const char *prog_path, int busy_budget)
1690 {
1691         const char *name = rte_vdev_device_name(dev);
1692         const unsigned int numa_node = dev->device.numa_node;
1693         struct pmd_internals *internals;
1694         struct rte_eth_dev *eth_dev;
1695         int ret;
1696         int i;
1697
1698         internals = rte_zmalloc_socket(name, sizeof(*internals), 0, numa_node);
1699         if (internals == NULL)
1700                 return NULL;
1701
1702         internals->start_queue_idx = start_queue_idx;
1703         internals->queue_cnt = queue_cnt;
1704         strlcpy(internals->if_name, if_name, IFNAMSIZ);
1705         strlcpy(internals->prog_path, prog_path, PATH_MAX);
1706         internals->custom_prog_configured = 0;
1707
1708 #ifndef ETH_AF_XDP_SHARED_UMEM
1709         if (shared_umem) {
1710                 AF_XDP_LOG(ERR, "Shared UMEM feature not available. "
1711                                 "Check kernel and libbpf version\n");
1712                 goto err_free_internals;
1713         }
1714 #endif
1715         internals->shared_umem = shared_umem;
1716
1717         if (xdp_get_channels_info(if_name, &internals->max_queue_cnt,
1718                                   &internals->combined_queue_cnt)) {
1719                 AF_XDP_LOG(ERR, "Failed to get channel info of interface: %s\n",
1720                                 if_name);
1721                 goto err_free_internals;
1722         }
1723
1724         if (queue_cnt > internals->combined_queue_cnt) {
1725                 AF_XDP_LOG(ERR, "Specified queue count %d is larger than combined queue count %d.\n",
1726                                 queue_cnt, internals->combined_queue_cnt);
1727                 goto err_free_internals;
1728         }
1729
1730         internals->rx_queues = rte_zmalloc_socket(NULL,
1731                                         sizeof(struct pkt_rx_queue) * queue_cnt,
1732                                         0, numa_node);
1733         if (internals->rx_queues == NULL) {
1734                 AF_XDP_LOG(ERR, "Failed to allocate memory for rx queues.\n");
1735                 goto err_free_internals;
1736         }
1737
1738         internals->tx_queues = rte_zmalloc_socket(NULL,
1739                                         sizeof(struct pkt_tx_queue) * queue_cnt,
1740                                         0, numa_node);
1741         if (internals->tx_queues == NULL) {
1742                 AF_XDP_LOG(ERR, "Failed to allocate memory for tx queues.\n");
1743                 goto err_free_rx;
1744         }
1745         for (i = 0; i < queue_cnt; i++) {
1746                 internals->tx_queues[i].pair = &internals->rx_queues[i];
1747                 internals->rx_queues[i].pair = &internals->tx_queues[i];
1748                 internals->rx_queues[i].xsk_queue_idx = start_queue_idx + i;
1749                 internals->tx_queues[i].xsk_queue_idx = start_queue_idx + i;
1750                 internals->rx_queues[i].busy_budget = busy_budget;
1751         }
1752
1753         ret = get_iface_info(if_name, &internals->eth_addr,
1754                              &internals->if_index);
1755         if (ret)
1756                 goto err_free_tx;
1757
1758         eth_dev = rte_eth_vdev_allocate(dev, 0);
1759         if (eth_dev == NULL)
1760                 goto err_free_tx;
1761
1762         eth_dev->data->dev_private = internals;
1763         eth_dev->data->dev_link = pmd_link;
1764         eth_dev->data->mac_addrs = &internals->eth_addr;
1765         eth_dev->data->dev_flags |= RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS;
1766         eth_dev->dev_ops = &ops;
1767         eth_dev->rx_pkt_burst = eth_af_xdp_rx;
1768         eth_dev->tx_pkt_burst = eth_af_xdp_tx;
1769
1770 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
1771         AF_XDP_LOG(INFO, "Zero copy between umem and mbuf enabled.\n");
1772 #endif
1773
1774         return eth_dev;
1775
1776 err_free_tx:
1777         rte_free(internals->tx_queues);
1778 err_free_rx:
1779         rte_free(internals->rx_queues);
1780 err_free_internals:
1781         rte_free(internals);
1782         return NULL;
1783 }
1784
1785 static int
1786 rte_pmd_af_xdp_probe(struct rte_vdev_device *dev)
1787 {
1788         struct rte_kvargs *kvlist;
1789         char if_name[IFNAMSIZ] = {'\0'};
1790         int xsk_start_queue_idx = ETH_AF_XDP_DFLT_START_QUEUE_IDX;
1791         int xsk_queue_cnt = ETH_AF_XDP_DFLT_QUEUE_COUNT;
1792         int shared_umem = 0;
1793         char prog_path[PATH_MAX] = {'\0'};
1794         int busy_budget = -1;
1795         struct rte_eth_dev *eth_dev = NULL;
1796         const char *name;
1797
1798         AF_XDP_LOG(INFO, "Initializing pmd_af_xdp for %s\n",
1799                 rte_vdev_device_name(dev));
1800
1801         name = rte_vdev_device_name(dev);
1802         if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1803                 AF_XDP_LOG(ERR, "Failed to probe %s. "
1804                                 "AF_XDP PMD does not support secondary processes.\n",
1805                                 name);
1806                 return -ENOTSUP;
1807         }
1808
1809         kvlist = rte_kvargs_parse(rte_vdev_device_args(dev), valid_arguments);
1810         if (kvlist == NULL) {
1811                 AF_XDP_LOG(ERR, "Invalid kvargs key\n");
1812                 return -EINVAL;
1813         }
1814
1815         if (dev->device.numa_node == SOCKET_ID_ANY)
1816                 dev->device.numa_node = rte_socket_id();
1817
1818         if (parse_parameters(kvlist, if_name, &xsk_start_queue_idx,
1819                              &xsk_queue_cnt, &shared_umem, prog_path,
1820                              &busy_budget) < 0) {
1821                 AF_XDP_LOG(ERR, "Invalid kvargs value\n");
1822                 return -EINVAL;
1823         }
1824
1825         if (strlen(if_name) == 0) {
1826                 AF_XDP_LOG(ERR, "Network interface must be specified\n");
1827                 return -EINVAL;
1828         }
1829
1830         busy_budget = busy_budget == -1 ? ETH_AF_XDP_DFLT_BUSY_BUDGET :
1831                                         busy_budget;
1832
1833         eth_dev = init_internals(dev, if_name, xsk_start_queue_idx,
1834                                         xsk_queue_cnt, shared_umem, prog_path,
1835                                         busy_budget);
1836         if (eth_dev == NULL) {
1837                 AF_XDP_LOG(ERR, "Failed to init internals\n");
1838                 return -1;
1839         }
1840
1841         rte_eth_dev_probing_finish(eth_dev);
1842
1843         return 0;
1844 }
1845
1846 static int
1847 rte_pmd_af_xdp_remove(struct rte_vdev_device *dev)
1848 {
1849         struct rte_eth_dev *eth_dev = NULL;
1850
1851         AF_XDP_LOG(INFO, "Removing AF_XDP ethdev on numa socket %u\n",
1852                 rte_socket_id());
1853
1854         if (dev == NULL)
1855                 return -1;
1856
1857         /* find the ethdev entry */
1858         eth_dev = rte_eth_dev_allocated(rte_vdev_device_name(dev));
1859         if (eth_dev == NULL)
1860                 return 0;
1861
1862         eth_dev_close(eth_dev);
1863         rte_eth_dev_release_port(eth_dev);
1864
1865
1866         return 0;
1867 }
1868
1869 static struct rte_vdev_driver pmd_af_xdp_drv = {
1870         .probe = rte_pmd_af_xdp_probe,
1871         .remove = rte_pmd_af_xdp_remove,
1872 };
1873
1874 RTE_PMD_REGISTER_VDEV(net_af_xdp, pmd_af_xdp_drv);
1875 RTE_PMD_REGISTER_PARAM_STRING(net_af_xdp,
1876                               "iface=<string> "
1877                               "start_queue=<int> "
1878                               "queue_count=<int> "
1879                               "shared_umem=<int> "
1880                               "xdp_prog=<string> "
1881                               "busy_budget=<int>");