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 static int
 701 eth_dev_configure(struct rte_eth_dev *dev)
 702 {
 703         struct pmd_internals *internal = dev->data->dev_private;
 704
 705         /* rx/tx must be paired */
 706         if (dev->data->nb_rx_queues != dev->data->nb_tx_queues)
 707                 return -EINVAL;
 708
 709         if (internal->shared_umem) {
 710                 struct internal_list *list = NULL;
 711                 const char *name = dev->device->name;
 712
 713                 /* Ensure PMD is not already inserted into the list */
 714                 list = find_internal_resource(internal);
 715                 if (list)
 716                         return 0;
 717
 718                 list = rte_zmalloc_socket(name, sizeof(*list), 0,
 719                                         dev->device->numa_node);
 720                 if (list == NULL)
 721                         return -1;
 722
 723                 list->eth_dev = dev;
 724                 pthread_mutex_lock(&internal_list_lock);
 725                 TAILQ_INSERT_TAIL(&internal_list, list, next);
 726                 pthread_mutex_unlock(&internal_list_lock);
 727         }
 728
 729         return 0;
 730 }
 731
 732 #define CLB_VAL_IDX 0
 733 static int
 734 eth_monitor_callback(const uint64_t value,
 735                 const uint64_t opaque[RTE_POWER_MONITOR_OPAQUE_SZ])
 736 {
 737         const uint64_t v = opaque[CLB_VAL_IDX];
 738         const uint64_t m = (uint32_t)~0;
 739
 740         /* if the value has changed, abort entering power optimized state */
 741         return (value & m) == v ? 0 : -1;
 742 }
 743
 744 static int
 745 eth_get_monitor_addr(void *rx_queue, struct rte_power_monitor_cond *pmc)
 746 {
 747         struct pkt_rx_queue *rxq = rx_queue;
 748         unsigned int *prod = rxq->rx.producer;
 749         const uint32_t cur_val = rxq->rx.cached_prod; /* use cached value */
 750
 751         /* watch for changes in producer ring */
 752         pmc->addr = (void *)prod;
 753
 754         /* store current value */
 755         pmc->opaque[CLB_VAL_IDX] = cur_val;
 756         pmc->fn = eth_monitor_callback;
 757
 758         /* AF_XDP producer ring index is 32-bit */
 759         pmc->size = sizeof(uint32_t);
 760
 761         return 0;
 762 }
 763
 764 static int
 765 eth_dev_info(struct rte_eth_dev *dev, struct rte_eth_dev_info *dev_info)
 766 {
 767         struct pmd_internals *internals = dev->data->dev_private;
 768
 769         dev_info->if_index = internals->if_index;
 770         dev_info->max_mac_addrs = 1;
 771         dev_info->max_rx_queues = internals->queue_cnt;
 772         dev_info->max_tx_queues = internals->queue_cnt;
 773
 774         dev_info->min_mtu = RTE_ETHER_MIN_MTU;
 775 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 776         dev_info->max_rx_pktlen = getpagesize() -
 777                                   sizeof(struct rte_mempool_objhdr) -
 778                                   sizeof(struct rte_mbuf) -
 779                                   RTE_PKTMBUF_HEADROOM - XDP_PACKET_HEADROOM;
 780 #else
 781         dev_info->max_rx_pktlen = ETH_AF_XDP_FRAME_SIZE - XDP_PACKET_HEADROOM;
 782 #endif
 783         dev_info->max_mtu = dev_info->max_rx_pktlen - ETH_AF_XDP_ETH_OVERHEAD;
 784
 785         dev_info->default_rxportconf.burst_size = ETH_AF_XDP_DFLT_BUSY_BUDGET;
 786         dev_info->default_txportconf.burst_size = ETH_AF_XDP_DFLT_BUSY_BUDGET;
 787         dev_info->default_rxportconf.nb_queues = 1;
 788         dev_info->default_txportconf.nb_queues = 1;
 789         dev_info->default_rxportconf.ring_size = ETH_AF_XDP_DFLT_NUM_DESCS;
 790         dev_info->default_txportconf.ring_size = ETH_AF_XDP_DFLT_NUM_DESCS;
 791
 792         return 0;
 793 }
 794
 795 static int
 796 eth_stats_get(struct rte_eth_dev *dev, struct rte_eth_stats *stats)
 797 {
 798         struct pmd_internals *internals = dev->data->dev_private;
 799         struct xdp_statistics xdp_stats;
 800         struct pkt_rx_queue *rxq;
 801         struct pkt_tx_queue *txq;
 802         socklen_t optlen;
 803         int i, ret;
 804
 805         for (i = 0; i < dev->data->nb_rx_queues; i++) {
 806                 optlen = sizeof(struct xdp_statistics);
 807                 rxq = &internals->rx_queues[i];
 808                 txq = rxq->pair;
 809                 stats->q_ipackets[i] = rxq->stats.rx_pkts;
 810                 stats->q_ibytes[i] = rxq->stats.rx_bytes;
 811
 812                 stats->q_opackets[i] = txq->stats.tx_pkts;
 813                 stats->q_obytes[i] = txq->stats.tx_bytes;
 814
 815                 stats->ipackets += stats->q_ipackets[i];
 816                 stats->ibytes += stats->q_ibytes[i];
 817                 stats->imissed += rxq->stats.rx_dropped;
 818                 stats->oerrors += txq->stats.tx_dropped;
 819                 ret = getsockopt(xsk_socket__fd(rxq->xsk), SOL_XDP,
 820                                 XDP_STATISTICS, &xdp_stats, &optlen);
 821                 if (ret != 0) {
 822                         AF_XDP_LOG(ERR, "getsockopt() failed for XDP_STATISTICS.\n");
 823                         return -1;
 824                 }
 825                 stats->imissed += xdp_stats.rx_dropped;
 826
 827                 stats->opackets += stats->q_opackets[i];
 828                 stats->obytes += stats->q_obytes[i];
 829         }
 830
 831         return 0;
 832 }
 833
 834 static int
 835 eth_stats_reset(struct rte_eth_dev *dev)
 836 {
 837         struct pmd_internals *internals = dev->data->dev_private;
 838         int i;
 839
 840         for (i = 0; i < internals->queue_cnt; i++) {
 841                 memset(&internals->rx_queues[i].stats, 0,
 842                                         sizeof(struct rx_stats));
 843                 memset(&internals->tx_queues[i].stats, 0,
 844                                         sizeof(struct tx_stats));
 845         }
 846
 847         return 0;
 848 }
 849
 850 static void
 851 remove_xdp_program(struct pmd_internals *internals)
 852 {
 853         uint32_t curr_prog_id = 0;
 854
 855         if (bpf_get_link_xdp_id(internals->if_index, &curr_prog_id,
 856                                 XDP_FLAGS_UPDATE_IF_NOEXIST)) {
 857                 AF_XDP_LOG(ERR, "bpf_get_link_xdp_id failed\n");
 858                 return;
 859         }
 860         bpf_set_link_xdp_fd(internals->if_index, -1,
 861                         XDP_FLAGS_UPDATE_IF_NOEXIST);
 862 }
 863
 864 static void
 865 xdp_umem_destroy(struct xsk_umem_info *umem)
 866 {
 867 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 868         umem->mb_pool = NULL;
 869 #else
 870         rte_memzone_free(umem->mz);
 871         umem->mz = NULL;
 872
 873         rte_ring_free(umem->buf_ring);
 874         umem->buf_ring = NULL;
 875 #endif
 876
 877         rte_free(umem);
 878 }
 879
 880 static int
 881 eth_dev_close(struct rte_eth_dev *dev)
 882 {
 883         struct pmd_internals *internals = dev->data->dev_private;
 884         struct pkt_rx_queue *rxq;
 885         int i;
 886
 887         if (rte_eal_process_type() != RTE_PROC_PRIMARY)
 888                 return 0;
 889
 890         AF_XDP_LOG(INFO, "Closing AF_XDP ethdev on numa socket %u\n",
 891                 rte_socket_id());
 892
 893         for (i = 0; i < internals->queue_cnt; i++) {
 894                 rxq = &internals->rx_queues[i];
 895                 if (rxq->umem == NULL)
 896                         break;
 897                 xsk_socket__delete(rxq->xsk);
 898
 899                 if (__atomic_sub_fetch(&rxq->umem->refcnt, 1, __ATOMIC_ACQUIRE)
 900                                 == 0) {
 901                         (void)xsk_umem__delete(rxq->umem->umem);
 902                         xdp_umem_destroy(rxq->umem);
 903                 }
 904
 905                 /* free pkt_tx_queue */
 906                 rte_free(rxq->pair);
 907                 rte_free(rxq);
 908         }
 909
 910         /*
 911          * MAC is not allocated dynamically, setting it to NULL would prevent
 912          * from releasing it in rte_eth_dev_release_port.
 913          */
 914         dev->data->mac_addrs = NULL;
 915
 916         remove_xdp_program(internals);
 917
 918         if (internals->shared_umem) {
 919                 struct internal_list *list;
 920
 921                 /* Remove ethdev from list used to track and share UMEMs */
 922                 list = find_internal_resource(internals);
 923                 if (list) {
 924                         pthread_mutex_lock(&internal_list_lock);
 925                         TAILQ_REMOVE(&internal_list, list, next);
 926                         pthread_mutex_unlock(&internal_list_lock);
 927                         rte_free(list);
 928                 }
 929         }
 930
 931         return 0;
 932 }
 933
 934 static int
 935 eth_link_update(struct rte_eth_dev *dev __rte_unused,
 936                 int wait_to_complete __rte_unused)
 937 {
 938         return 0;
 939 }
 940
 941 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 942 static inline uintptr_t get_base_addr(struct rte_mempool *mp, uint64_t *align)
 943 {
 944         struct rte_mempool_memhdr *memhdr;
 945         uintptr_t memhdr_addr, aligned_addr;
 946
 947         memhdr = STAILQ_FIRST(&mp->mem_list);
 948         memhdr_addr = (uintptr_t)memhdr->addr;
 949         aligned_addr = memhdr_addr & ~(getpagesize() - 1);
 950         *align = memhdr_addr - aligned_addr;
 951
 952         return aligned_addr;
 953 }
 954
 955 /* Check if the netdev,qid context already exists */
 956 static inline bool
 957 ctx_exists(struct pkt_rx_queue *rxq, const char *ifname,
 958                 struct pkt_rx_queue *list_rxq, const char *list_ifname)
 959 {
 960         bool exists = false;
 961
 962         if (rxq->xsk_queue_idx == list_rxq->xsk_queue_idx &&
 963                         !strncmp(ifname, list_ifname, IFNAMSIZ)) {
 964                 AF_XDP_LOG(ERR, "ctx %s,%i already exists, cannot share umem\n",
 965                                         ifname, rxq->xsk_queue_idx);
 966                 exists = true;
 967         }
 968
 969         return exists;
 970 }
 971
 972 /* Get a pointer to an existing UMEM which overlays the rxq's mb_pool */
 973 static inline int
 974 get_shared_umem(struct pkt_rx_queue *rxq, const char *ifname,
 975                         struct xsk_umem_info **umem)
 976 {
 977         struct internal_list *list;
 978         struct pmd_internals *internals;
 979         int i = 0, ret = 0;
 980         struct rte_mempool *mb_pool = rxq->mb_pool;
 981
 982         if (mb_pool == NULL)
 983                 return ret;
 984
 985         pthread_mutex_lock(&internal_list_lock);
 986
 987         TAILQ_FOREACH(list, &internal_list, next) {
 988                 internals = list->eth_dev->data->dev_private;
 989                 for (i = 0; i < internals->queue_cnt; i++) {
 990                         struct pkt_rx_queue *list_rxq =
 991                                                 &internals->rx_queues[i];
 992                         if (rxq == list_rxq)
 993                                 continue;
 994                         if (mb_pool == internals->rx_queues[i].mb_pool) {
 995                                 if (ctx_exists(rxq, ifname, list_rxq,
 996                                                 internals->if_name)) {
 997                                         ret = -1;
 998                                         goto out;
 999                                 }
1000                                 if (__atomic_load_n(&internals->rx_queues[i].umem->refcnt,
1001                                                     __ATOMIC_ACQUIRE)) {
1002                                         *umem = internals->rx_queues[i].umem;
1003                                         goto out;
1004                                 }
1005                         }
1006                 }
1007         }
1008
1009 out:
1010         pthread_mutex_unlock(&internal_list_lock);
1011
1012         return ret;
1013 }
1014
1015 static struct
1016 xsk_umem_info *xdp_umem_configure(struct pmd_internals *internals,
1017                                   struct pkt_rx_queue *rxq)
1018 {
1019         struct xsk_umem_info *umem = NULL;
1020         int ret;
1021         struct xsk_umem_config usr_config = {
1022                 .fill_size = ETH_AF_XDP_DFLT_NUM_DESCS * 2,
1023                 .comp_size = ETH_AF_XDP_DFLT_NUM_DESCS,
1024                 .flags = XDP_UMEM_UNALIGNED_CHUNK_FLAG};
1025         void *base_addr = NULL;
1026         struct rte_mempool *mb_pool = rxq->mb_pool;
1027         uint64_t umem_size, align = 0;
1028
1029         if (internals->shared_umem) {
1030                 if (get_shared_umem(rxq, internals->if_name, &umem) < 0)
1031                         return NULL;
1032
1033                 if (umem != NULL &&
1034                         __atomic_load_n(&umem->refcnt, __ATOMIC_ACQUIRE) <
1035                                         umem->max_xsks) {
1036                         AF_XDP_LOG(INFO, "%s,qid%i sharing UMEM\n",
1037                                         internals->if_name, rxq->xsk_queue_idx);
1038                         __atomic_fetch_add(&umem->refcnt, 1, __ATOMIC_ACQUIRE);
1039                 }
1040         }
1041
1042         if (umem == NULL) {
1043                 usr_config.frame_size =
1044                         rte_mempool_calc_obj_size(mb_pool->elt_size,
1045                                                   mb_pool->flags, NULL);
1046                 usr_config.frame_headroom = mb_pool->header_size +
1047                                                 sizeof(struct rte_mbuf) +
1048                                                 rte_pktmbuf_priv_size(mb_pool) +
1049                                                 RTE_PKTMBUF_HEADROOM;
1050
1051                 umem = rte_zmalloc_socket("umem", sizeof(*umem), 0,
1052                                           rte_socket_id());
1053                 if (umem == NULL) {
1054                         AF_XDP_LOG(ERR, "Failed to allocate umem info");
1055                         return NULL;
1056                 }
1057
1058                 umem->mb_pool = mb_pool;
1059                 base_addr = (void *)get_base_addr(mb_pool, &align);
1060                 umem_size = (uint64_t)mb_pool->populated_size *
1061                                 (uint64_t)usr_config.frame_size +
1062                                 align;
1063
1064                 ret = xsk_umem__create(&umem->umem, base_addr, umem_size,
1065                                 &rxq->fq, &rxq->cq, &usr_config);
1066                 if (ret) {
1067                         AF_XDP_LOG(ERR, "Failed to create umem");
1068                         goto err;
1069                 }
1070                 umem->buffer = base_addr;
1071
1072                 if (internals->shared_umem) {
1073                         umem->max_xsks = mb_pool->populated_size /
1074                                                 ETH_AF_XDP_NUM_BUFFERS;
1075                         AF_XDP_LOG(INFO, "Max xsks for UMEM %s: %u\n",
1076                                                 mb_pool->name, umem->max_xsks);
1077                 }
1078
1079                 __atomic_store_n(&umem->refcnt, 1, __ATOMIC_RELEASE);
1080         }
1081
1082 #else
1083 static struct
1084 xsk_umem_info *xdp_umem_configure(struct pmd_internals *internals,
1085                                   struct pkt_rx_queue *rxq)
1086 {
1087         struct xsk_umem_info *umem;
1088         const struct rte_memzone *mz;
1089         struct xsk_umem_config usr_config = {
1090                 .fill_size = ETH_AF_XDP_DFLT_NUM_DESCS,
1091                 .comp_size = ETH_AF_XDP_DFLT_NUM_DESCS,
1092                 .frame_size = ETH_AF_XDP_FRAME_SIZE,
1093                 .frame_headroom = 0 };
1094         char ring_name[RTE_RING_NAMESIZE];
1095         char mz_name[RTE_MEMZONE_NAMESIZE];
1096         int ret;
1097         uint64_t i;
1098
1099         umem = rte_zmalloc_socket("umem", sizeof(*umem), 0, rte_socket_id());
1100         if (umem == NULL) {
1101                 AF_XDP_LOG(ERR, "Failed to allocate umem info");
1102                 return NULL;
1103         }
1104
1105         snprintf(ring_name, sizeof(ring_name), "af_xdp_ring_%s_%u",
1106                        internals->if_name, rxq->xsk_queue_idx);
1107         umem->buf_ring = rte_ring_create(ring_name,
1108                                          ETH_AF_XDP_NUM_BUFFERS,
1109                                          rte_socket_id(),
1110                                          0x0);
1111         if (umem->buf_ring == NULL) {
1112                 AF_XDP_LOG(ERR, "Failed to create rte_ring\n");
1113                 goto err;
1114         }
1115
1116         for (i = 0; i < ETH_AF_XDP_NUM_BUFFERS; i++)
1117                 rte_ring_enqueue(umem->buf_ring,
1118                                  (void *)(i * ETH_AF_XDP_FRAME_SIZE));
1119
1120         snprintf(mz_name, sizeof(mz_name), "af_xdp_umem_%s_%u",
1121                        internals->if_name, rxq->xsk_queue_idx);
1122         mz = rte_memzone_reserve_aligned(mz_name,
1123                         ETH_AF_XDP_NUM_BUFFERS * ETH_AF_XDP_FRAME_SIZE,
1124                         rte_socket_id(), RTE_MEMZONE_IOVA_CONTIG,
1125                         getpagesize());
1126         if (mz == NULL) {
1127                 AF_XDP_LOG(ERR, "Failed to reserve memzone for af_xdp umem.\n");
1128                 goto err;
1129         }
1130
1131         ret = xsk_umem__create(&umem->umem, mz->addr,
1132                                ETH_AF_XDP_NUM_BUFFERS * ETH_AF_XDP_FRAME_SIZE,
1133                                &rxq->fq, &rxq->cq,
1134                                &usr_config);
1135
1136         if (ret) {
1137                 AF_XDP_LOG(ERR, "Failed to create umem");
1138                 goto err;
1139         }
1140         umem->mz = mz;
1141
1142 #endif
1143         return umem;
1144
1145 err:
1146         xdp_umem_destroy(umem);
1147         return NULL;
1148 }
1149
1150 static int
1151 load_custom_xdp_prog(const char *prog_path, int if_index, struct bpf_map **map)
1152 {
1153         int ret, prog_fd = -1;
1154         struct bpf_object *obj;
1155
1156         ret = bpf_prog_load(prog_path, BPF_PROG_TYPE_XDP, &obj, &prog_fd);
1157         if (ret) {
1158                 AF_XDP_LOG(ERR, "Failed to load program %s\n", prog_path);
1159                 return ret;
1160         }
1161
1162         /*
1163          * The loaded program must provision for a map of xsks, such that some
1164          * traffic can be redirected to userspace.
1165          */
1166         *map = bpf_object__find_map_by_name(obj, "xsks_map");
1167         if (!*map) {
1168                 AF_XDP_LOG(ERR, "Failed to find xsks_map in %s\n", prog_path);
1169                 return -1;
1170         }
1171
1172         /* Link the program with the given network device */
1173         ret = bpf_set_link_xdp_fd(if_index, prog_fd,
1174                                         XDP_FLAGS_UPDATE_IF_NOEXIST);
1175         if (ret) {
1176                 AF_XDP_LOG(ERR, "Failed to set prog fd %d on interface\n",
1177                                 prog_fd);
1178                 return -1;
1179         }
1180
1181         AF_XDP_LOG(INFO, "Successfully loaded XDP program %s with fd %d\n",
1182                                 prog_path, prog_fd);
1183
1184         return 0;
1185 }
1186
1187 /* Detect support for busy polling through setsockopt(). */
1188 static int
1189 configure_preferred_busy_poll(struct pkt_rx_queue *rxq)
1190 {
1191         int sock_opt = 1;
1192         int fd = xsk_socket__fd(rxq->xsk);
1193         int ret = 0;
1194
1195         ret = setsockopt(fd, SOL_SOCKET, SO_PREFER_BUSY_POLL,
1196                         (void *)&sock_opt, sizeof(sock_opt));
1197         if (ret < 0) {
1198                 AF_XDP_LOG(DEBUG, "Failed to set SO_PREFER_BUSY_POLL\n");
1199                 goto err_prefer;
1200         }
1201
1202         sock_opt = ETH_AF_XDP_DFLT_BUSY_TIMEOUT;
1203         ret = setsockopt(fd, SOL_SOCKET, SO_BUSY_POLL, (void *)&sock_opt,
1204                         sizeof(sock_opt));
1205         if (ret < 0) {
1206                 AF_XDP_LOG(DEBUG, "Failed to set SO_BUSY_POLL\n");
1207                 goto err_timeout;
1208         }
1209
1210         sock_opt = rxq->busy_budget;
1211         ret = setsockopt(fd, SOL_SOCKET, SO_BUSY_POLL_BUDGET,
1212                         (void *)&sock_opt, sizeof(sock_opt));
1213         if (ret < 0) {
1214                 AF_XDP_LOG(DEBUG, "Failed to set SO_BUSY_POLL_BUDGET\n");
1215         } else {
1216                 AF_XDP_LOG(INFO, "Busy polling budget set to: %u\n",
1217                                         rxq->busy_budget);
1218                 return 0;
1219         }
1220
1221         /* setsockopt failure - attempt to restore xsk to default state and
1222          * proceed without busy polling support.
1223          */
1224         sock_opt = 0;
1225         ret = setsockopt(fd, SOL_SOCKET, SO_BUSY_POLL, (void *)&sock_opt,
1226                         sizeof(sock_opt));
1227         if (ret < 0) {
1228                 AF_XDP_LOG(ERR, "Failed to unset SO_BUSY_POLL\n");
1229                 return -1;
1230         }
1231
1232 err_timeout:
1233         sock_opt = 0;
1234         ret = setsockopt(fd, SOL_SOCKET, SO_PREFER_BUSY_POLL,
1235                         (void *)&sock_opt, sizeof(sock_opt));
1236         if (ret < 0) {
1237                 AF_XDP_LOG(ERR, "Failed to unset SO_PREFER_BUSY_POLL\n");
1238                 return -1;
1239         }
1240
1241 err_prefer:
1242         rxq->busy_budget = 0;
1243         return 0;
1244 }
1245
1246 static int
1247 xsk_configure(struct pmd_internals *internals, struct pkt_rx_queue *rxq,
1248               int ring_size)
1249 {
1250         struct xsk_socket_config cfg;
1251         struct pkt_tx_queue *txq = rxq->pair;
1252         int ret = 0;
1253         int reserve_size = ETH_AF_XDP_DFLT_NUM_DESCS;
1254         struct rte_mbuf *fq_bufs[reserve_size];
1255
1256         rxq->umem = xdp_umem_configure(internals, rxq);
1257         if (rxq->umem == NULL)
1258                 return -ENOMEM;
1259         txq->umem = rxq->umem;
1260
1261         cfg.rx_size = ring_size;
1262         cfg.tx_size = ring_size;
1263         cfg.libbpf_flags = 0;
1264         cfg.xdp_flags = XDP_FLAGS_UPDATE_IF_NOEXIST;
1265         cfg.bind_flags = 0;
1266
1267 #if defined(XDP_USE_NEED_WAKEUP)
1268         cfg.bind_flags |= XDP_USE_NEED_WAKEUP;
1269 #endif
1270
1271         if (strnlen(internals->prog_path, PATH_MAX) &&
1272                                 !internals->custom_prog_configured) {
1273                 ret = load_custom_xdp_prog(internals->prog_path,
1274                                            internals->if_index,
1275                                            &internals->map);
1276                 if (ret) {
1277                         AF_XDP_LOG(ERR, "Failed to load custom XDP program %s\n",
1278                                         internals->prog_path);
1279                         goto err;
1280                 }
1281                 internals->custom_prog_configured = 1;
1282                 cfg.libbpf_flags = XSK_LIBBPF_FLAGS__INHIBIT_PROG_LOAD;
1283         }
1284
1285         if (internals->shared_umem)
1286                 ret = create_shared_socket(&rxq->xsk, internals->if_name,
1287                                 rxq->xsk_queue_idx, rxq->umem->umem, &rxq->rx,
1288                                 &txq->tx, &rxq->fq, &rxq->cq, &cfg);
1289         else
1290                 ret = xsk_socket__create(&rxq->xsk, internals->if_name,
1291                                 rxq->xsk_queue_idx, rxq->umem->umem, &rxq->rx,
1292                                 &txq->tx, &cfg);
1293
1294         if (ret) {
1295                 AF_XDP_LOG(ERR, "Failed to create xsk socket.\n");
1296                 goto err;
1297         }
1298
1299         /* insert the xsk into the xsks_map */
1300         if (internals->custom_prog_configured) {
1301                 int err, fd;
1302
1303                 fd = xsk_socket__fd(rxq->xsk);
1304                 err = bpf_map_update_elem(bpf_map__fd(internals->map),
1305                                           &rxq->xsk_queue_idx, &fd, 0);
1306                 if (err) {
1307                         AF_XDP_LOG(ERR, "Failed to insert xsk in map.\n");
1308                         goto err;
1309                 }
1310         }
1311
1312 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
1313         ret = rte_pktmbuf_alloc_bulk(rxq->umem->mb_pool, fq_bufs, reserve_size);
1314         if (ret) {
1315                 AF_XDP_LOG(DEBUG, "Failed to get enough buffers for fq.\n");
1316                 goto err;
1317         }
1318 #endif
1319
1320         if (rxq->busy_budget) {
1321                 ret = configure_preferred_busy_poll(rxq);
1322                 if (ret) {
1323                         AF_XDP_LOG(ERR, "Failed configure busy polling.\n");
1324                         goto err;
1325                 }
1326         }
1327
1328         ret = reserve_fill_queue(rxq->umem, reserve_size, fq_bufs, &rxq->fq);
1329         if (ret) {
1330                 xsk_socket__delete(rxq->xsk);
1331                 AF_XDP_LOG(ERR, "Failed to reserve fill queue.\n");
1332                 goto err;
1333         }
1334
1335         return 0;
1336
1337 err:
1338         if (__atomic_sub_fetch(&rxq->umem->refcnt, 1, __ATOMIC_ACQUIRE) == 0)
1339                 xdp_umem_destroy(rxq->umem);
1340
1341         return ret;
1342 }
1343
1344 static int
1345 eth_rx_queue_setup(struct rte_eth_dev *dev,
1346                    uint16_t rx_queue_id,
1347                    uint16_t nb_rx_desc,
1348                    unsigned int socket_id __rte_unused,
1349                    const struct rte_eth_rxconf *rx_conf __rte_unused,
1350                    struct rte_mempool *mb_pool)
1351 {
1352         struct pmd_internals *internals = dev->data->dev_private;
1353         struct pkt_rx_queue *rxq;
1354         int ret;
1355
1356         rxq = &internals->rx_queues[rx_queue_id];
1357
1358         AF_XDP_LOG(INFO, "Set up rx queue, rx queue id: %d, xsk queue id: %d\n",
1359                    rx_queue_id, rxq->xsk_queue_idx);
1360
1361 #ifndef XDP_UMEM_UNALIGNED_CHUNK_FLAG
1362         uint32_t buf_size, data_size;
1363
1364         /* Now get the space available for data in the mbuf */
1365         buf_size = rte_pktmbuf_data_room_size(mb_pool) -
1366                 RTE_PKTMBUF_HEADROOM;
1367         data_size = ETH_AF_XDP_FRAME_SIZE;
1368
1369         if (data_size > buf_size) {
1370                 AF_XDP_LOG(ERR, "%s: %d bytes will not fit in mbuf (%d bytes)\n",
1371                         dev->device->name, data_size, buf_size);
1372                 ret = -ENOMEM;
1373                 goto err;
1374         }
1375 #endif
1376
1377         rxq->mb_pool = mb_pool;
1378
1379         if (xsk_configure(internals, rxq, nb_rx_desc)) {
1380                 AF_XDP_LOG(ERR, "Failed to configure xdp socket\n");
1381                 ret = -EINVAL;
1382                 goto err;
1383         }
1384
1385         if (!rxq->busy_budget)
1386                 AF_XDP_LOG(DEBUG, "Preferred busy polling not enabled\n");
1387
1388         rxq->fds[0].fd = xsk_socket__fd(rxq->xsk);
1389         rxq->fds[0].events = POLLIN;
1390
1391         dev->data->rx_queues[rx_queue_id] = rxq;
1392         return 0;
1393
1394 err:
1395         return ret;
1396 }
1397
1398 static int
1399 eth_tx_queue_setup(struct rte_eth_dev *dev,
1400                    uint16_t tx_queue_id,
1401                    uint16_t nb_tx_desc __rte_unused,
1402                    unsigned int socket_id __rte_unused,
1403                    const struct rte_eth_txconf *tx_conf __rte_unused)
1404 {
1405         struct pmd_internals *internals = dev->data->dev_private;
1406         struct pkt_tx_queue *txq;
1407
1408         txq = &internals->tx_queues[tx_queue_id];
1409
1410         dev->data->tx_queues[tx_queue_id] = txq;
1411         return 0;
1412 }
1413
1414 static int
1415 eth_dev_mtu_set(struct rte_eth_dev *dev, uint16_t mtu)
1416 {
1417         struct pmd_internals *internals = dev->data->dev_private;
1418         struct ifreq ifr = { .ifr_mtu = mtu };
1419         int ret;
1420         int s;
1421
1422         s = socket(PF_INET, SOCK_DGRAM, 0);
1423         if (s < 0)
1424                 return -EINVAL;
1425
1426         strlcpy(ifr.ifr_name, internals->if_name, IFNAMSIZ);
1427         ret = ioctl(s, SIOCSIFMTU, &ifr);
1428         close(s);
1429
1430         return (ret < 0) ? -errno : 0;
1431 }
1432
1433 static int
1434 eth_dev_change_flags(char *if_name, uint32_t flags, uint32_t mask)
1435 {
1436         struct ifreq ifr;
1437         int ret = 0;
1438         int s;
1439
1440         s = socket(PF_INET, SOCK_DGRAM, 0);
1441         if (s < 0)
1442                 return -errno;
1443
1444         strlcpy(ifr.ifr_name, if_name, IFNAMSIZ);
1445         if (ioctl(s, SIOCGIFFLAGS, &ifr) < 0) {
1446                 ret = -errno;
1447                 goto out;
1448         }
1449         ifr.ifr_flags &= mask;
1450         ifr.ifr_flags |= flags;
1451         if (ioctl(s, SIOCSIFFLAGS, &ifr) < 0) {
1452                 ret = -errno;
1453                 goto out;
1454         }
1455 out:
1456         close(s);
1457         return ret;
1458 }
1459
1460 static int
1461 eth_dev_promiscuous_enable(struct rte_eth_dev *dev)
1462 {
1463         struct pmd_internals *internals = dev->data->dev_private;
1464
1465         return eth_dev_change_flags(internals->if_name, IFF_PROMISC, ~0);
1466 }
1467
1468 static int
1469 eth_dev_promiscuous_disable(struct rte_eth_dev *dev)
1470 {
1471         struct pmd_internals *internals = dev->data->dev_private;
1472
1473         return eth_dev_change_flags(internals->if_name, 0, ~IFF_PROMISC);
1474 }
1475
1476 static const struct eth_dev_ops ops = {
1477         .dev_start = eth_dev_start,
1478         .dev_stop = eth_dev_stop,
1479         .dev_close = eth_dev_close,
1480         .dev_configure = eth_dev_configure,
1481         .dev_infos_get = eth_dev_info,
1482         .mtu_set = eth_dev_mtu_set,
1483         .promiscuous_enable = eth_dev_promiscuous_enable,
1484         .promiscuous_disable = eth_dev_promiscuous_disable,
1485         .rx_queue_setup = eth_rx_queue_setup,
1486         .tx_queue_setup = eth_tx_queue_setup,
1487         .link_update = eth_link_update,
1488         .stats_get = eth_stats_get,
1489         .stats_reset = eth_stats_reset,
1490         .get_monitor_addr = eth_get_monitor_addr,
1491 };
1492
1493 /** parse busy_budget argument */
1494 static int
1495 parse_budget_arg(const char *key __rte_unused,
1496                   const char *value, void *extra_args)
1497 {
1498         int *i = (int *)extra_args;
1499         char *end;
1500
1501         *i = strtol(value, &end, 10);
1502         if (*i < 0 || *i > UINT16_MAX) {
1503                 AF_XDP_LOG(ERR, "Invalid busy_budget %i, must be >= 0 and <= %u\n",
1504                                 *i, UINT16_MAX);
1505                 return -EINVAL;
1506         }
1507
1508         return 0;
1509 }
1510
1511 /** parse integer from integer argument */
1512 static int
1513 parse_integer_arg(const char *key __rte_unused,
1514                   const char *value, void *extra_args)
1515 {
1516         int *i = (int *)extra_args;
1517         char *end;
1518
1519         *i = strtol(value, &end, 10);
1520         if (*i < 0) {
1521                 AF_XDP_LOG(ERR, "Argument has to be positive.\n");
1522                 return -EINVAL;
1523         }
1524
1525         return 0;
1526 }
1527
1528 /** parse name argument */
1529 static int
1530 parse_name_arg(const char *key __rte_unused,
1531                const char *value, void *extra_args)
1532 {
1533         char *name = extra_args;
1534
1535         if (strnlen(value, IFNAMSIZ) > IFNAMSIZ - 1) {
1536                 AF_XDP_LOG(ERR, "Invalid name %s, should be less than %u bytes.\n",
1537                            value, IFNAMSIZ);
1538                 return -EINVAL;
1539         }
1540
1541         strlcpy(name, value, IFNAMSIZ);
1542
1543         return 0;
1544 }
1545
1546 /** parse xdp prog argument */
1547 static int
1548 parse_prog_arg(const char *key __rte_unused,
1549                const char *value, void *extra_args)
1550 {
1551         char *path = extra_args;
1552
1553         if (strnlen(value, PATH_MAX) == PATH_MAX) {
1554                 AF_XDP_LOG(ERR, "Invalid path %s, should be less than %u bytes.\n",
1555                            value, PATH_MAX);
1556                 return -EINVAL;
1557         }
1558
1559         if (access(value, F_OK) != 0) {
1560                 AF_XDP_LOG(ERR, "Error accessing %s: %s\n",
1561                            value, strerror(errno));
1562                 return -EINVAL;
1563         }
1564
1565         strlcpy(path, value, PATH_MAX);
1566
1567         return 0;
1568 }
1569
1570 static int
1571 xdp_get_channels_info(const char *if_name, int *max_queues,
1572                                 int *combined_queues)
1573 {
1574         struct ethtool_channels channels;
1575         struct ifreq ifr;
1576         int fd, ret;
1577
1578         fd = socket(AF_INET, SOCK_DGRAM, 0);
1579         if (fd < 0)
1580                 return -1;
1581
1582         channels.cmd = ETHTOOL_GCHANNELS;
1583         ifr.ifr_data = (void *)&channels;
1584         strlcpy(ifr.ifr_name, if_name, IFNAMSIZ);
1585         ret = ioctl(fd, SIOCETHTOOL, &ifr);
1586         if (ret) {
1587                 if (errno == EOPNOTSUPP) {
1588                         ret = 0;
1589                 } else {
1590                         ret = -errno;
1591                         goto out;
1592                 }
1593         }
1594
1595         if (channels.max_combined == 0 || errno == EOPNOTSUPP) {
1596                 /* If the device says it has no channels, then all traffic
1597                  * is sent to a single stream, so max queues = 1.
1598                  */
1599                 *max_queues = 1;
1600                 *combined_queues = 1;
1601         } else {
1602                 *max_queues = channels.max_combined;
1603                 *combined_queues = channels.combined_count;
1604         }
1605
1606  out:
1607         close(fd);
1608         return ret;
1609 }
1610
1611 static int
1612 parse_parameters(struct rte_kvargs *kvlist, char *if_name, int *start_queue,
1613                         int *queue_cnt, int *shared_umem, char *prog_path,
1614                         int *busy_budget)
1615 {
1616         int ret;
1617
1618         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_IFACE_ARG,
1619                                  &parse_name_arg, if_name);
1620         if (ret < 0)
1621                 goto free_kvlist;
1622
1623         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_START_QUEUE_ARG,
1624                                  &parse_integer_arg, start_queue);
1625         if (ret < 0)
1626                 goto free_kvlist;
1627
1628         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_QUEUE_COUNT_ARG,
1629                                  &parse_integer_arg, queue_cnt);
1630         if (ret < 0 || *queue_cnt <= 0) {
1631                 ret = -EINVAL;
1632                 goto free_kvlist;
1633         }
1634
1635         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_SHARED_UMEM_ARG,
1636                                 &parse_integer_arg, shared_umem);
1637         if (ret < 0)
1638                 goto free_kvlist;
1639
1640         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_PROG_ARG,
1641                                  &parse_prog_arg, prog_path);
1642         if (ret < 0)
1643                 goto free_kvlist;
1644
1645         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_BUDGET_ARG,
1646                                 &parse_budget_arg, busy_budget);
1647         if (ret < 0)
1648                 goto free_kvlist;
1649
1650 free_kvlist:
1651         rte_kvargs_free(kvlist);
1652         return ret;
1653 }
1654
1655 static int
1656 get_iface_info(const char *if_name,
1657                struct rte_ether_addr *eth_addr,
1658                int *if_index)
1659 {
1660         struct ifreq ifr;
1661         int sock = socket(AF_INET, SOCK_DGRAM, IPPROTO_IP);
1662
1663         if (sock < 0)
1664                 return -1;
1665
1666         strlcpy(ifr.ifr_name, if_name, IFNAMSIZ);
1667         if (ioctl(sock, SIOCGIFINDEX, &ifr))
1668                 goto error;
1669
1670         *if_index = ifr.ifr_ifindex;
1671
1672         if (ioctl(sock, SIOCGIFHWADDR, &ifr))
1673                 goto error;
1674
1675         rte_memcpy(eth_addr, ifr.ifr_hwaddr.sa_data, RTE_ETHER_ADDR_LEN);
1676
1677         close(sock);
1678         return 0;
1679
1680 error:
1681         close(sock);
1682         return -1;
1683 }
1684
1685 static struct rte_eth_dev *
1686 init_internals(struct rte_vdev_device *dev, const char *if_name,
1687                 int start_queue_idx, int queue_cnt, int shared_umem,
1688                 const char *prog_path, int busy_budget)
1689 {
1690         const char *name = rte_vdev_device_name(dev);
1691         const unsigned int numa_node = dev->device.numa_node;
1692         struct pmd_internals *internals;
1693         struct rte_eth_dev *eth_dev;
1694         int ret;
1695         int i;
1696
1697         internals = rte_zmalloc_socket(name, sizeof(*internals), 0, numa_node);
1698         if (internals == NULL)
1699                 return NULL;
1700
1701         internals->start_queue_idx = start_queue_idx;
1702         internals->queue_cnt = queue_cnt;
1703         strlcpy(internals->if_name, if_name, IFNAMSIZ);
1704         strlcpy(internals->prog_path, prog_path, PATH_MAX);
1705         internals->custom_prog_configured = 0;
1706
1707 #ifndef ETH_AF_XDP_SHARED_UMEM
1708         if (shared_umem) {
1709                 AF_XDP_LOG(ERR, "Shared UMEM feature not available. "
1710                                 "Check kernel and libbpf version\n");
1711                 goto err_free_internals;
1712         }
1713 #endif
1714         internals->shared_umem = shared_umem;
1715
1716         if (xdp_get_channels_info(if_name, &internals->max_queue_cnt,
1717                                   &internals->combined_queue_cnt)) {
1718                 AF_XDP_LOG(ERR, "Failed to get channel info of interface: %s\n",
1719                                 if_name);
1720                 goto err_free_internals;
1721         }
1722
1723         if (queue_cnt > internals->combined_queue_cnt) {
1724                 AF_XDP_LOG(ERR, "Specified queue count %d is larger than combined queue count %d.\n",
1725                                 queue_cnt, internals->combined_queue_cnt);
1726                 goto err_free_internals;
1727         }
1728
1729         internals->rx_queues = rte_zmalloc_socket(NULL,
1730                                         sizeof(struct pkt_rx_queue) * queue_cnt,
1731                                         0, numa_node);
1732         if (internals->rx_queues == NULL) {
1733                 AF_XDP_LOG(ERR, "Failed to allocate memory for rx queues.\n");
1734                 goto err_free_internals;
1735         }
1736
1737         internals->tx_queues = rte_zmalloc_socket(NULL,
1738                                         sizeof(struct pkt_tx_queue) * queue_cnt,
1739                                         0, numa_node);
1740         if (internals->tx_queues == NULL) {
1741                 AF_XDP_LOG(ERR, "Failed to allocate memory for tx queues.\n");
1742                 goto err_free_rx;
1743         }
1744         for (i = 0; i < queue_cnt; i++) {
1745                 internals->tx_queues[i].pair = &internals->rx_queues[i];
1746                 internals->rx_queues[i].pair = &internals->tx_queues[i];
1747                 internals->rx_queues[i].xsk_queue_idx = start_queue_idx + i;
1748                 internals->tx_queues[i].xsk_queue_idx = start_queue_idx + i;
1749                 internals->rx_queues[i].busy_budget = busy_budget;
1750         }
1751
1752         ret = get_iface_info(if_name, &internals->eth_addr,
1753                              &internals->if_index);
1754         if (ret)
1755                 goto err_free_tx;
1756
1757         eth_dev = rte_eth_vdev_allocate(dev, 0);
1758         if (eth_dev == NULL)
1759                 goto err_free_tx;
1760
1761         eth_dev->data->dev_private = internals;
1762         eth_dev->data->dev_link = pmd_link;
1763         eth_dev->data->mac_addrs = &internals->eth_addr;
1764         eth_dev->data->dev_flags |= RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS;
1765         eth_dev->dev_ops = &ops;
1766         eth_dev->rx_pkt_burst = eth_af_xdp_rx;
1767         eth_dev->tx_pkt_burst = eth_af_xdp_tx;
1768
1769 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
1770         AF_XDP_LOG(INFO, "Zero copy between umem and mbuf enabled.\n");
1771 #endif
1772
1773         return eth_dev;
1774
1775 err_free_tx:
1776         rte_free(internals->tx_queues);
1777 err_free_rx:
1778         rte_free(internals->rx_queues);
1779 err_free_internals:
1780         rte_free(internals);
1781         return NULL;
1782 }
1783
1784 static int
1785 rte_pmd_af_xdp_probe(struct rte_vdev_device *dev)
1786 {
1787         struct rte_kvargs *kvlist;
1788         char if_name[IFNAMSIZ] = {'\0'};
1789         int xsk_start_queue_idx = ETH_AF_XDP_DFLT_START_QUEUE_IDX;
1790         int xsk_queue_cnt = ETH_AF_XDP_DFLT_QUEUE_COUNT;
1791         int shared_umem = 0;
1792         char prog_path[PATH_MAX] = {'\0'};
1793         int busy_budget = -1;
1794         struct rte_eth_dev *eth_dev = NULL;
1795         const char *name;
1796
1797         AF_XDP_LOG(INFO, "Initializing pmd_af_xdp for %s\n",
1798                 rte_vdev_device_name(dev));
1799
1800         name = rte_vdev_device_name(dev);
1801         if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1802                 AF_XDP_LOG(ERR, "Failed to probe %s. "
1803                                 "AF_XDP PMD does not support secondary processes.\n",
1804                                 name);
1805                 return -ENOTSUP;
1806         }
1807
1808         kvlist = rte_kvargs_parse(rte_vdev_device_args(dev), valid_arguments);
1809         if (kvlist == NULL) {
1810                 AF_XDP_LOG(ERR, "Invalid kvargs key\n");
1811                 return -EINVAL;
1812         }
1813
1814         if (dev->device.numa_node == SOCKET_ID_ANY)
1815                 dev->device.numa_node = rte_socket_id();
1816
1817         if (parse_parameters(kvlist, if_name, &xsk_start_queue_idx,
1818                              &xsk_queue_cnt, &shared_umem, prog_path,
1819                              &busy_budget) < 0) {
1820                 AF_XDP_LOG(ERR, "Invalid kvargs value\n");
1821                 return -EINVAL;
1822         }
1823
1824         if (strlen(if_name) == 0) {
1825                 AF_XDP_LOG(ERR, "Network interface must be specified\n");
1826                 return -EINVAL;
1827         }
1828
1829         busy_budget = busy_budget == -1 ? ETH_AF_XDP_DFLT_BUSY_BUDGET :
1830                                         busy_budget;
1831
1832         eth_dev = init_internals(dev, if_name, xsk_start_queue_idx,
1833                                         xsk_queue_cnt, shared_umem, prog_path,
1834                                         busy_budget);
1835         if (eth_dev == NULL) {
1836                 AF_XDP_LOG(ERR, "Failed to init internals\n");
1837                 return -1;
1838         }
1839
1840         rte_eth_dev_probing_finish(eth_dev);
1841
1842         return 0;
1843 }
1844
1845 static int
1846 rte_pmd_af_xdp_remove(struct rte_vdev_device *dev)
1847 {
1848         struct rte_eth_dev *eth_dev = NULL;
1849
1850         AF_XDP_LOG(INFO, "Removing AF_XDP ethdev on numa socket %u\n",
1851                 rte_socket_id());
1852
1853         if (dev == NULL)
1854                 return -1;
1855
1856         /* find the ethdev entry */
1857         eth_dev = rte_eth_dev_allocated(rte_vdev_device_name(dev));
1858         if (eth_dev == NULL)
1859                 return 0;
1860
1861         eth_dev_close(eth_dev);
1862         rte_eth_dev_release_port(eth_dev);
1863
1864
1865         return 0;
1866 }
1867
1868 static struct rte_vdev_driver pmd_af_xdp_drv = {
1869         .probe = rte_pmd_af_xdp_probe,
1870         .remove = rte_pmd_af_xdp_remove,
1871 };
1872
1873 RTE_PMD_REGISTER_VDEV(net_af_xdp, pmd_af_xdp_drv);
1874 RTE_PMD_REGISTER_PARAM_STRING(net_af_xdp,
1875                               "iface=<string> "
1876                               "start_queue=<int> "
1877                               "queue_count=<int> "
1878                               "shared_umem=<int> "
1879                               "xdp_prog=<string> "
1880                               "busy_budget=<int>");