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/xsk.h>
  19
  20 #include <rte_ethdev.h>
  21 #include <ethdev_driver.h>
  22 #include <ethdev_vdev.h>
  23 #include <rte_kvargs.h>
  24 #include <rte_bus_vdev.h>
  25 #include <rte_string_fns.h>
  26 #include <rte_branch_prediction.h>
  27 #include <rte_common.h>
  28 #include <rte_dev.h>
  29 #include <rte_eal.h>
  30 #include <rte_ether.h>
  31 #include <rte_lcore.h>
  32 #include <rte_log.h>
  33 #include <rte_memory.h>
  34 #include <rte_memzone.h>
  35 #include <rte_mempool.h>
  36 #include <rte_mbuf.h>
  37 #include <rte_malloc.h>
  38 #include <rte_ring.h>
  39 #include <rte_spinlock.h>
  40 #include <rte_power_intrinsics.h>
  41
  42 #include "compat.h"
  43
  44 #ifndef SO_PREFER_BUSY_POLL
  45 #define SO_PREFER_BUSY_POLL 69
  46 #endif
  47 #ifndef SO_BUSY_POLL_BUDGET
  48 #define SO_BUSY_POLL_BUDGET 70
  49 #endif
  50
  51
  52 #ifndef SOL_XDP
  53 #define SOL_XDP 283
  54 #endif
  55
  56 #ifndef AF_XDP
  57 #define AF_XDP 44
  58 #endif
  59
  60 #ifndef PF_XDP
  61 #define PF_XDP AF_XDP
  62 #endif
  63
  64 RTE_LOG_REGISTER_DEFAULT(af_xdp_logtype, NOTICE);
  65
  66 #define AF_XDP_LOG(level, fmt, args...)                 \
  67         rte_log(RTE_LOG_ ## level, af_xdp_logtype,      \
  68                 "%s(): " fmt, __func__, ##args)
  69
  70 #define ETH_AF_XDP_FRAME_SIZE           2048
  71 #define ETH_AF_XDP_NUM_BUFFERS          4096
  72 #define ETH_AF_XDP_DFLT_NUM_DESCS       XSK_RING_CONS__DEFAULT_NUM_DESCS
  73 #define ETH_AF_XDP_DFLT_START_QUEUE_IDX 0
  74 #define ETH_AF_XDP_DFLT_QUEUE_COUNT     1
  75 #define ETH_AF_XDP_DFLT_BUSY_BUDGET     64
  76 #define ETH_AF_XDP_DFLT_BUSY_TIMEOUT    20
  77
  78 #define ETH_AF_XDP_RX_BATCH_SIZE        XSK_RING_CONS__DEFAULT_NUM_DESCS
  79 #define ETH_AF_XDP_TX_BATCH_SIZE        XSK_RING_CONS__DEFAULT_NUM_DESCS
  80
  81
  82 struct xsk_umem_info {
  83         struct xsk_umem *umem;
  84         struct rte_ring *buf_ring;
  85         const struct rte_memzone *mz;
  86         struct rte_mempool *mb_pool;
  87         void *buffer;
  88         uint8_t refcnt;
  89         uint32_t max_xsks;
  90 };
  91
  92 struct rx_stats {
  93         uint64_t rx_pkts;
  94         uint64_t rx_bytes;
  95         uint64_t rx_dropped;
  96 };
  97
  98 struct pkt_rx_queue {
  99         struct xsk_ring_cons rx;
 100         struct xsk_umem_info *umem;
 101         struct xsk_socket *xsk;
 102         struct rte_mempool *mb_pool;
 103
 104         struct rx_stats stats;
 105
 106         struct xsk_ring_prod fq;
 107         struct xsk_ring_cons cq;
 108
 109         struct pkt_tx_queue *pair;
 110         struct pollfd fds[1];
 111         int xsk_queue_idx;
 112         int busy_budget;
 113 };
 114
 115 struct tx_stats {
 116         uint64_t tx_pkts;
 117         uint64_t tx_bytes;
 118         uint64_t tx_dropped;
 119 };
 120
 121 struct pkt_tx_queue {
 122         struct xsk_ring_prod tx;
 123         struct xsk_umem_info *umem;
 124
 125         struct tx_stats stats;
 126
 127         struct pkt_rx_queue *pair;
 128         int xsk_queue_idx;
 129 };
 130
 131 struct pmd_internals {
 132         int if_index;
 133         char if_name[IFNAMSIZ];
 134         int start_queue_idx;
 135         int queue_cnt;
 136         int max_queue_cnt;
 137         int combined_queue_cnt;
 138         bool shared_umem;
 139         char prog_path[PATH_MAX];
 140         bool custom_prog_configured;
 141
 142         struct rte_ether_addr eth_addr;
 143
 144         struct pkt_rx_queue *rx_queues;
 145         struct pkt_tx_queue *tx_queues;
 146 };
 147
 148 #define ETH_AF_XDP_IFACE_ARG                    "iface"
 149 #define ETH_AF_XDP_START_QUEUE_ARG              "start_queue"
 150 #define ETH_AF_XDP_QUEUE_COUNT_ARG              "queue_count"
 151 #define ETH_AF_XDP_SHARED_UMEM_ARG              "shared_umem"
 152 #define ETH_AF_XDP_PROG_ARG                     "xdp_prog"
 153 #define ETH_AF_XDP_BUDGET_ARG                   "busy_budget"
 154
 155 static const char * const valid_arguments[] = {
 156         ETH_AF_XDP_IFACE_ARG,
 157         ETH_AF_XDP_START_QUEUE_ARG,
 158         ETH_AF_XDP_QUEUE_COUNT_ARG,
 159         ETH_AF_XDP_SHARED_UMEM_ARG,
 160         ETH_AF_XDP_PROG_ARG,
 161         ETH_AF_XDP_BUDGET_ARG,
 162         NULL
 163 };
 164
 165 static const struct rte_eth_link pmd_link = {
 166         .link_speed = ETH_SPEED_NUM_10G,
 167         .link_duplex = ETH_LINK_FULL_DUPLEX,
 168         .link_status = ETH_LINK_DOWN,
 169         .link_autoneg = ETH_LINK_AUTONEG
 170 };
 171
 172 /* List which tracks PMDs to facilitate sharing UMEMs across them. */
 173 struct internal_list {
 174         TAILQ_ENTRY(internal_list) next;
 175         struct rte_eth_dev *eth_dev;
 176 };
 177
 178 TAILQ_HEAD(internal_list_head, internal_list);
 179 static struct internal_list_head internal_list =
 180         TAILQ_HEAD_INITIALIZER(internal_list);
 181
 182 static pthread_mutex_t internal_list_lock = PTHREAD_MUTEX_INITIALIZER;
 183
 184 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 185 static inline int
 186 reserve_fill_queue_zc(struct xsk_umem_info *umem, uint16_t reserve_size,
 187                       struct rte_mbuf **bufs, struct xsk_ring_prod *fq)
 188 {
 189         uint32_t idx;
 190         uint16_t i;
 191
 192         if (unlikely(!xsk_ring_prod__reserve(fq, reserve_size, &idx))) {
 193                 for (i = 0; i < reserve_size; i++)
 194                         rte_pktmbuf_free(bufs[i]);
 195                 AF_XDP_LOG(DEBUG, "Failed to reserve enough fq descs.\n");
 196                 return -1;
 197         }
 198
 199         for (i = 0; i < reserve_size; i++) {
 200                 __u64 *fq_addr;
 201                 uint64_t addr;
 202
 203                 fq_addr = xsk_ring_prod__fill_addr(fq, idx++);
 204                 addr = (uint64_t)bufs[i] - (uint64_t)umem->buffer -
 205                                 umem->mb_pool->header_size;
 206                 *fq_addr = addr;
 207         }
 208
 209         xsk_ring_prod__submit(fq, reserve_size);
 210
 211         return 0;
 212 }
 213 #else
 214 static inline int
 215 reserve_fill_queue_cp(struct xsk_umem_info *umem, uint16_t reserve_size,
 216                       struct rte_mbuf **bufs __rte_unused,
 217                       struct xsk_ring_prod *fq)
 218 {
 219         void *addrs[reserve_size];
 220         uint32_t idx;
 221         uint16_t i;
 222
 223         if (rte_ring_dequeue_bulk(umem->buf_ring, addrs, reserve_size, NULL)
 224                     != reserve_size) {
 225                 AF_XDP_LOG(DEBUG, "Failed to get enough buffers for fq.\n");
 226                 return -1;
 227         }
 228
 229         if (unlikely(!xsk_ring_prod__reserve(fq, reserve_size, &idx))) {
 230                 AF_XDP_LOG(DEBUG, "Failed to reserve enough fq descs.\n");
 231                 rte_ring_enqueue_bulk(umem->buf_ring, addrs,
 232                                 reserve_size, NULL);
 233                 return -1;
 234         }
 235
 236         for (i = 0; i < reserve_size; i++) {
 237                 __u64 *fq_addr;
 238
 239                 fq_addr = xsk_ring_prod__fill_addr(fq, idx++);
 240                 *fq_addr = (uint64_t)addrs[i];
 241         }
 242
 243         xsk_ring_prod__submit(fq, reserve_size);
 244
 245         return 0;
 246 }
 247 #endif
 248
 249 static inline int
 250 reserve_fill_queue(struct xsk_umem_info *umem, uint16_t reserve_size,
 251                    struct rte_mbuf **bufs, struct xsk_ring_prod *fq)
 252 {
 253 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 254         return reserve_fill_queue_zc(umem, reserve_size, bufs, fq);
 255 #else
 256         return reserve_fill_queue_cp(umem, reserve_size, bufs, fq);
 257 #endif
 258 }
 259
 260 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 261 static uint16_t
 262 af_xdp_rx_zc(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 263 {
 264         struct pkt_rx_queue *rxq = queue;
 265         struct xsk_ring_cons *rx = &rxq->rx;
 266         struct xsk_ring_prod *fq = &rxq->fq;
 267         struct xsk_umem_info *umem = rxq->umem;
 268         uint32_t idx_rx = 0;
 269         unsigned long rx_bytes = 0;
 270         int i;
 271         struct rte_mbuf *fq_bufs[ETH_AF_XDP_RX_BATCH_SIZE];
 272
 273         nb_pkts = xsk_ring_cons__peek(rx, nb_pkts, &idx_rx);
 274
 275         if (nb_pkts == 0) {
 276                 /* we can assume a kernel >= 5.11 is in use if busy polling is
 277                  * enabled and thus we can safely use the recvfrom() syscall
 278                  * which is only supported for AF_XDP sockets in kernels >=
 279                  * 5.11.
 280                  */
 281                 if (rxq->busy_budget) {
 282                         (void)recvfrom(xsk_socket__fd(rxq->xsk), NULL, 0,
 283                                        MSG_DONTWAIT, NULL, NULL);
 284                 } else if (xsk_ring_prod__needs_wakeup(fq)) {
 285                         (void)poll(&rxq->fds[0], 1, 1000);
 286                 }
 287
 288                 return 0;
 289         }
 290
 291         /* allocate bufs for fill queue replenishment after rx */
 292         if (rte_pktmbuf_alloc_bulk(umem->mb_pool, fq_bufs, nb_pkts)) {
 293                 AF_XDP_LOG(DEBUG,
 294                         "Failed to get enough buffers for fq.\n");
 295                 /* rollback cached_cons which is added by
 296                  * xsk_ring_cons__peek
 297                  */
 298                 rx->cached_cons -= nb_pkts;
 299                 return 0;
 300         }
 301
 302         for (i = 0; i < nb_pkts; i++) {
 303                 const struct xdp_desc *desc;
 304                 uint64_t addr;
 305                 uint32_t len;
 306                 uint64_t offset;
 307
 308                 desc = xsk_ring_cons__rx_desc(rx, idx_rx++);
 309                 addr = desc->addr;
 310                 len = desc->len;
 311
 312                 offset = xsk_umem__extract_offset(addr);
 313                 addr = xsk_umem__extract_addr(addr);
 314
 315                 bufs[i] = (struct rte_mbuf *)
 316                                 xsk_umem__get_data(umem->buffer, addr +
 317                                         umem->mb_pool->header_size);
 318                 bufs[i]->data_off = offset - sizeof(struct rte_mbuf) -
 319                         rte_pktmbuf_priv_size(umem->mb_pool) -
 320                         umem->mb_pool->header_size;
 321
 322                 rte_pktmbuf_pkt_len(bufs[i]) = len;
 323                 rte_pktmbuf_data_len(bufs[i]) = len;
 324                 rx_bytes += len;
 325         }
 326
 327         xsk_ring_cons__release(rx, nb_pkts);
 328         (void)reserve_fill_queue(umem, nb_pkts, fq_bufs, fq);
 329
 330         /* statistics */
 331         rxq->stats.rx_pkts += nb_pkts;
 332         rxq->stats.rx_bytes += rx_bytes;
 333
 334         return nb_pkts;
 335 }
 336 #else
 337 static uint16_t
 338 af_xdp_rx_cp(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 339 {
 340         struct pkt_rx_queue *rxq = queue;
 341         struct xsk_ring_cons *rx = &rxq->rx;
 342         struct xsk_umem_info *umem = rxq->umem;
 343         struct xsk_ring_prod *fq = &rxq->fq;
 344         uint32_t idx_rx = 0;
 345         unsigned long rx_bytes = 0;
 346         int i;
 347         uint32_t free_thresh = fq->size >> 1;
 348         struct rte_mbuf *mbufs[ETH_AF_XDP_RX_BATCH_SIZE];
 349
 350         if (xsk_prod_nb_free(fq, free_thresh) >= free_thresh)
 351                 (void)reserve_fill_queue(umem, nb_pkts, NULL, fq);
 352
 353         nb_pkts = xsk_ring_cons__peek(rx, nb_pkts, &idx_rx);
 354         if (nb_pkts == 0) {
 355 #if defined(XDP_USE_NEED_WAKEUP)
 356                 if (xsk_ring_prod__needs_wakeup(fq))
 357                         (void)poll(rxq->fds, 1, 1000);
 358 #endif
 359                 return 0;
 360         }
 361
 362         if (unlikely(rte_pktmbuf_alloc_bulk(rxq->mb_pool, mbufs, nb_pkts))) {
 363                 /* rollback cached_cons which is added by
 364                  * xsk_ring_cons__peek
 365                  */
 366                 rx->cached_cons -= nb_pkts;
 367                 return 0;
 368         }
 369
 370         for (i = 0; i < nb_pkts; i++) {
 371                 const struct xdp_desc *desc;
 372                 uint64_t addr;
 373                 uint32_t len;
 374                 void *pkt;
 375
 376                 desc = xsk_ring_cons__rx_desc(rx, idx_rx++);
 377                 addr = desc->addr;
 378                 len = desc->len;
 379                 pkt = xsk_umem__get_data(rxq->umem->mz->addr, addr);
 380
 381                 rte_memcpy(rte_pktmbuf_mtod(mbufs[i], void *), pkt, len);
 382                 rte_ring_enqueue(umem->buf_ring, (void *)addr);
 383                 rte_pktmbuf_pkt_len(mbufs[i]) = len;
 384                 rte_pktmbuf_data_len(mbufs[i]) = len;
 385                 rx_bytes += len;
 386                 bufs[i] = mbufs[i];
 387         }
 388
 389         xsk_ring_cons__release(rx, nb_pkts);
 390
 391         /* statistics */
 392         rxq->stats.rx_pkts += nb_pkts;
 393         rxq->stats.rx_bytes += rx_bytes;
 394
 395         return nb_pkts;
 396 }
 397 #endif
 398
 399 static uint16_t
 400 af_xdp_rx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 401 {
 402 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 403         return af_xdp_rx_zc(queue, bufs, nb_pkts);
 404 #else
 405         return af_xdp_rx_cp(queue, bufs, nb_pkts);
 406 #endif
 407 }
 408
 409 static uint16_t
 410 eth_af_xdp_rx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 411 {
 412         uint16_t nb_rx;
 413
 414         if (likely(nb_pkts <= ETH_AF_XDP_RX_BATCH_SIZE))
 415                 return af_xdp_rx(queue, bufs, nb_pkts);
 416
 417         /* Split larger batch into smaller batches of size
 418          * ETH_AF_XDP_RX_BATCH_SIZE or less.
 419          */
 420         nb_rx = 0;
 421         while (nb_pkts) {
 422                 uint16_t ret, n;
 423
 424                 n = (uint16_t)RTE_MIN(nb_pkts, ETH_AF_XDP_RX_BATCH_SIZE);
 425                 ret = af_xdp_rx(queue, &bufs[nb_rx], n);
 426                 nb_rx = (uint16_t)(nb_rx + ret);
 427                 nb_pkts = (uint16_t)(nb_pkts - ret);
 428                 if (ret < n)
 429                         break;
 430         }
 431
 432         return nb_rx;
 433 }
 434
 435 static void
 436 pull_umem_cq(struct xsk_umem_info *umem, int size, struct xsk_ring_cons *cq)
 437 {
 438         size_t i, n;
 439         uint32_t idx_cq = 0;
 440
 441         n = xsk_ring_cons__peek(cq, size, &idx_cq);
 442
 443         for (i = 0; i < n; i++) {
 444                 uint64_t addr;
 445                 addr = *xsk_ring_cons__comp_addr(cq, idx_cq++);
 446 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 447                 addr = xsk_umem__extract_addr(addr);
 448                 rte_pktmbuf_free((struct rte_mbuf *)
 449                                         xsk_umem__get_data(umem->buffer,
 450                                         addr + umem->mb_pool->header_size));
 451 #else
 452                 rte_ring_enqueue(umem->buf_ring, (void *)addr);
 453 #endif
 454         }
 455
 456         xsk_ring_cons__release(cq, n);
 457 }
 458
 459 static void
 460 kick_tx(struct pkt_tx_queue *txq, struct xsk_ring_cons *cq)
 461 {
 462         struct xsk_umem_info *umem = txq->umem;
 463
 464         pull_umem_cq(umem, XSK_RING_CONS__DEFAULT_NUM_DESCS, cq);
 465
 466         if (tx_syscall_needed(&txq->tx))
 467                 while (send(xsk_socket__fd(txq->pair->xsk), NULL,
 468                             0, MSG_DONTWAIT) < 0) {
 469                         /* some thing unexpected */
 470                         if (errno != EBUSY && errno != EAGAIN && errno != EINTR)
 471                                 break;
 472
 473                         /* pull from completion queue to leave more space */
 474                         if (errno == EAGAIN)
 475                                 pull_umem_cq(umem,
 476                                              XSK_RING_CONS__DEFAULT_NUM_DESCS,
 477                                              cq);
 478                 }
 479 }
 480
 481 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 482 static uint16_t
 483 af_xdp_tx_zc(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 484 {
 485         struct pkt_tx_queue *txq = queue;
 486         struct xsk_umem_info *umem = txq->umem;
 487         struct rte_mbuf *mbuf;
 488         unsigned long tx_bytes = 0;
 489         int i;
 490         uint32_t idx_tx;
 491         uint16_t count = 0;
 492         struct xdp_desc *desc;
 493         uint64_t addr, offset;
 494         struct xsk_ring_cons *cq = &txq->pair->cq;
 495         uint32_t free_thresh = cq->size >> 1;
 496
 497         if (xsk_cons_nb_avail(cq, free_thresh) >= free_thresh)
 498                 pull_umem_cq(umem, XSK_RING_CONS__DEFAULT_NUM_DESCS, cq);
 499
 500         for (i = 0; i < nb_pkts; i++) {
 501                 mbuf = bufs[i];
 502
 503                 if (mbuf->pool == umem->mb_pool) {
 504                         if (!xsk_ring_prod__reserve(&txq->tx, 1, &idx_tx)) {
 505                                 kick_tx(txq, cq);
 506                                 if (!xsk_ring_prod__reserve(&txq->tx, 1,
 507                                                             &idx_tx))
 508                                         goto out;
 509                         }
 510                         desc = xsk_ring_prod__tx_desc(&txq->tx, idx_tx);
 511                         desc->len = mbuf->pkt_len;
 512                         addr = (uint64_t)mbuf - (uint64_t)umem->buffer -
 513                                         umem->mb_pool->header_size;
 514                         offset = rte_pktmbuf_mtod(mbuf, uint64_t) -
 515                                         (uint64_t)mbuf +
 516                                         umem->mb_pool->header_size;
 517                         offset = offset << XSK_UNALIGNED_BUF_OFFSET_SHIFT;
 518                         desc->addr = addr | offset;
 519                         count++;
 520                 } else {
 521                         struct rte_mbuf *local_mbuf =
 522                                         rte_pktmbuf_alloc(umem->mb_pool);
 523                         void *pkt;
 524
 525                         if (local_mbuf == NULL)
 526                                 goto out;
 527
 528                         if (!xsk_ring_prod__reserve(&txq->tx, 1, &idx_tx)) {
 529                                 rte_pktmbuf_free(local_mbuf);
 530                                 goto out;
 531                         }
 532
 533                         desc = xsk_ring_prod__tx_desc(&txq->tx, idx_tx);
 534                         desc->len = mbuf->pkt_len;
 535
 536                         addr = (uint64_t)local_mbuf - (uint64_t)umem->buffer -
 537                                         umem->mb_pool->header_size;
 538                         offset = rte_pktmbuf_mtod(local_mbuf, uint64_t) -
 539                                         (uint64_t)local_mbuf +
 540                                         umem->mb_pool->header_size;
 541                         pkt = xsk_umem__get_data(umem->buffer, addr + offset);
 542                         offset = offset << XSK_UNALIGNED_BUF_OFFSET_SHIFT;
 543                         desc->addr = addr | offset;
 544                         rte_memcpy(pkt, rte_pktmbuf_mtod(mbuf, void *),
 545                                         desc->len);
 546                         rte_pktmbuf_free(mbuf);
 547                         count++;
 548                 }
 549
 550                 tx_bytes += mbuf->pkt_len;
 551         }
 552
 553 out:
 554         xsk_ring_prod__submit(&txq->tx, count);
 555         kick_tx(txq, cq);
 556
 557         txq->stats.tx_pkts += count;
 558         txq->stats.tx_bytes += tx_bytes;
 559         txq->stats.tx_dropped += nb_pkts - count;
 560
 561         return count;
 562 }
 563 #else
 564 static uint16_t
 565 af_xdp_tx_cp(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 566 {
 567         struct pkt_tx_queue *txq = queue;
 568         struct xsk_umem_info *umem = txq->umem;
 569         struct rte_mbuf *mbuf;
 570         void *addrs[ETH_AF_XDP_TX_BATCH_SIZE];
 571         unsigned long tx_bytes = 0;
 572         int i;
 573         uint32_t idx_tx;
 574         struct xsk_ring_cons *cq = &txq->pair->cq;
 575
 576         pull_umem_cq(umem, nb_pkts, cq);
 577
 578         nb_pkts = rte_ring_dequeue_bulk(umem->buf_ring, addrs,
 579                                         nb_pkts, NULL);
 580         if (nb_pkts == 0)
 581                 return 0;
 582
 583         if (xsk_ring_prod__reserve(&txq->tx, nb_pkts, &idx_tx) != nb_pkts) {
 584                 kick_tx(txq, cq);
 585                 rte_ring_enqueue_bulk(umem->buf_ring, addrs, nb_pkts, NULL);
 586                 return 0;
 587         }
 588
 589         for (i = 0; i < nb_pkts; i++) {
 590                 struct xdp_desc *desc;
 591                 void *pkt;
 592
 593                 desc = xsk_ring_prod__tx_desc(&txq->tx, idx_tx + i);
 594                 mbuf = bufs[i];
 595                 desc->len = mbuf->pkt_len;
 596
 597                 desc->addr = (uint64_t)addrs[i];
 598                 pkt = xsk_umem__get_data(umem->mz->addr,
 599                                          desc->addr);
 600                 rte_memcpy(pkt, rte_pktmbuf_mtod(mbuf, void *), desc->len);
 601                 tx_bytes += mbuf->pkt_len;
 602                 rte_pktmbuf_free(mbuf);
 603         }
 604
 605         xsk_ring_prod__submit(&txq->tx, nb_pkts);
 606
 607         kick_tx(txq, cq);
 608
 609         txq->stats.tx_pkts += nb_pkts;
 610         txq->stats.tx_bytes += tx_bytes;
 611
 612         return nb_pkts;
 613 }
 614
 615 static uint16_t
 616 af_xdp_tx_cp_batch(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 617 {
 618         uint16_t nb_tx;
 619
 620         if (likely(nb_pkts <= ETH_AF_XDP_TX_BATCH_SIZE))
 621                 return af_xdp_tx_cp(queue, bufs, nb_pkts);
 622
 623         nb_tx = 0;
 624         while (nb_pkts) {
 625                 uint16_t ret, n;
 626
 627                 /* Split larger batch into smaller batches of size
 628                  * ETH_AF_XDP_TX_BATCH_SIZE or less.
 629                  */
 630                 n = (uint16_t)RTE_MIN(nb_pkts, ETH_AF_XDP_TX_BATCH_SIZE);
 631                 ret = af_xdp_tx_cp(queue, &bufs[nb_tx], n);
 632                 nb_tx = (uint16_t)(nb_tx + ret);
 633                 nb_pkts = (uint16_t)(nb_pkts - ret);
 634                 if (ret < n)
 635                         break;
 636         }
 637
 638         return nb_tx;
 639 }
 640 #endif
 641
 642 static uint16_t
 643 eth_af_xdp_tx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 644 {
 645 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 646         return af_xdp_tx_zc(queue, bufs, nb_pkts);
 647 #else
 648         return af_xdp_tx_cp_batch(queue, bufs, nb_pkts);
 649 #endif
 650 }
 651
 652 static int
 653 eth_dev_start(struct rte_eth_dev *dev)
 654 {
 655         dev->data->dev_link.link_status = ETH_LINK_UP;
 656
 657         return 0;
 658 }
 659
 660 /* This function gets called when the current port gets stopped. */
 661 static int
 662 eth_dev_stop(struct rte_eth_dev *dev)
 663 {
 664         dev->data->dev_link.link_status = ETH_LINK_DOWN;
 665         return 0;
 666 }
 667
 668 /* Find ethdev in list */
 669 static inline struct internal_list *
 670 find_internal_resource(struct pmd_internals *port_int)
 671 {
 672         int found = 0;
 673         struct internal_list *list = NULL;
 674
 675         if (port_int == NULL)
 676                 return NULL;
 677
 678         pthread_mutex_lock(&internal_list_lock);
 679
 680         TAILQ_FOREACH(list, &internal_list, next) {
 681                 struct pmd_internals *list_int =
 682                                 list->eth_dev->data->dev_private;
 683                 if (list_int == port_int) {
 684                         found = 1;
 685                         break;
 686                 }
 687         }
 688
 689         pthread_mutex_unlock(&internal_list_lock);
 690
 691         if (!found)
 692                 return NULL;
 693
 694         return list;
 695 }
 696
 697 /* Check if the netdev,qid context already exists */
 698 static inline bool
 699 ctx_exists(struct pkt_rx_queue *rxq, const char *ifname,
 700                 struct pkt_rx_queue *list_rxq, const char *list_ifname)
 701 {
 702         bool exists = false;
 703
 704         if (rxq->xsk_queue_idx == list_rxq->xsk_queue_idx &&
 705                         !strncmp(ifname, list_ifname, IFNAMSIZ)) {
 706                 AF_XDP_LOG(ERR, "ctx %s,%i already exists, cannot share umem\n",
 707                                         ifname, rxq->xsk_queue_idx);
 708                 exists = true;
 709         }
 710
 711         return exists;
 712 }
 713
 714 /* Get a pointer to an existing UMEM which overlays the rxq's mb_pool */
 715 static inline int
 716 get_shared_umem(struct pkt_rx_queue *rxq, const char *ifname,
 717                         struct xsk_umem_info **umem)
 718 {
 719         struct internal_list *list;
 720         struct pmd_internals *internals;
 721         int i = 0, ret = 0;
 722         struct rte_mempool *mb_pool = rxq->mb_pool;
 723
 724         if (mb_pool == NULL)
 725                 return ret;
 726
 727         pthread_mutex_lock(&internal_list_lock);
 728
 729         TAILQ_FOREACH(list, &internal_list, next) {
 730                 internals = list->eth_dev->data->dev_private;
 731                 for (i = 0; i < internals->queue_cnt; i++) {
 732                         struct pkt_rx_queue *list_rxq =
 733                                                 &internals->rx_queues[i];
 734                         if (rxq == list_rxq)
 735                                 continue;
 736                         if (mb_pool == internals->rx_queues[i].mb_pool) {
 737                                 if (ctx_exists(rxq, ifname, list_rxq,
 738                                                 internals->if_name)) {
 739                                         ret = -1;
 740                                         goto out;
 741                                 }
 742                                 if (__atomic_load_n(
 743                                         &internals->rx_queues[i].umem->refcnt,
 744                                                         __ATOMIC_ACQUIRE)) {
 745                                         *umem = internals->rx_queues[i].umem;
 746                                         goto out;
 747                                 }
 748                         }
 749                 }
 750         }
 751
 752 out:
 753         pthread_mutex_unlock(&internal_list_lock);
 754
 755         return ret;
 756 }
 757
 758 static int
 759 eth_dev_configure(struct rte_eth_dev *dev)
 760 {
 761         struct pmd_internals *internal = dev->data->dev_private;
 762
 763         /* rx/tx must be paired */
 764         if (dev->data->nb_rx_queues != dev->data->nb_tx_queues)
 765                 return -EINVAL;
 766
 767         if (internal->shared_umem) {
 768                 struct internal_list *list = NULL;
 769                 const char *name = dev->device->name;
 770
 771                 /* Ensure PMD is not already inserted into the list */
 772                 list = find_internal_resource(internal);
 773                 if (list)
 774                         return 0;
 775
 776                 list = rte_zmalloc_socket(name, sizeof(*list), 0,
 777                                         dev->device->numa_node);
 778                 if (list == NULL)
 779                         return -1;
 780
 781                 list->eth_dev = dev;
 782                 pthread_mutex_lock(&internal_list_lock);
 783                 TAILQ_INSERT_TAIL(&internal_list, list, next);
 784                 pthread_mutex_unlock(&internal_list_lock);
 785         }
 786
 787         return 0;
 788 }
 789
 790 #define CLB_VAL_IDX 0
 791 static int
 792 eth_monitor_callback(const uint64_t value,
 793                 const uint64_t opaque[RTE_POWER_MONITOR_OPAQUE_SZ])
 794 {
 795         const uint64_t v = opaque[CLB_VAL_IDX];
 796         const uint64_t m = (uint32_t)~0;
 797
 798         /* if the value has changed, abort entering power optimized state */
 799         return (value & m) == v ? 0 : -1;
 800 }
 801
 802 static int
 803 eth_get_monitor_addr(void *rx_queue, struct rte_power_monitor_cond *pmc)
 804 {
 805         struct pkt_rx_queue *rxq = rx_queue;
 806         unsigned int *prod = rxq->rx.producer;
 807         const uint32_t cur_val = rxq->rx.cached_prod; /* use cached value */
 808
 809         /* watch for changes in producer ring */
 810         pmc->addr = (void *)prod;
 811
 812         /* store current value */
 813         pmc->opaque[CLB_VAL_IDX] = cur_val;
 814         pmc->fn = eth_monitor_callback;
 815
 816         /* AF_XDP producer ring index is 32-bit */
 817         pmc->size = sizeof(uint32_t);
 818
 819         return 0;
 820 }
 821
 822 static int
 823 eth_dev_info(struct rte_eth_dev *dev, struct rte_eth_dev_info *dev_info)
 824 {
 825         struct pmd_internals *internals = dev->data->dev_private;
 826
 827         dev_info->if_index = internals->if_index;
 828         dev_info->max_mac_addrs = 1;
 829         dev_info->max_rx_pktlen = ETH_FRAME_LEN;
 830         dev_info->max_rx_queues = internals->queue_cnt;
 831         dev_info->max_tx_queues = internals->queue_cnt;
 832
 833         dev_info->min_mtu = RTE_ETHER_MIN_MTU;
 834 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 835         dev_info->max_mtu = getpagesize() -
 836                                 sizeof(struct rte_mempool_objhdr) -
 837                                 sizeof(struct rte_mbuf) -
 838                                 RTE_PKTMBUF_HEADROOM - XDP_PACKET_HEADROOM;
 839 #else
 840         dev_info->max_mtu = ETH_AF_XDP_FRAME_SIZE - XDP_PACKET_HEADROOM;
 841 #endif
 842
 843         dev_info->default_rxportconf.burst_size = ETH_AF_XDP_DFLT_BUSY_BUDGET;
 844         dev_info->default_txportconf.burst_size = ETH_AF_XDP_DFLT_BUSY_BUDGET;
 845         dev_info->default_rxportconf.nb_queues = 1;
 846         dev_info->default_txportconf.nb_queues = 1;
 847         dev_info->default_rxportconf.ring_size = ETH_AF_XDP_DFLT_NUM_DESCS;
 848         dev_info->default_txportconf.ring_size = ETH_AF_XDP_DFLT_NUM_DESCS;
 849
 850         return 0;
 851 }
 852
 853 static int
 854 eth_stats_get(struct rte_eth_dev *dev, struct rte_eth_stats *stats)
 855 {
 856         struct pmd_internals *internals = dev->data->dev_private;
 857         struct xdp_statistics xdp_stats;
 858         struct pkt_rx_queue *rxq;
 859         struct pkt_tx_queue *txq;
 860         socklen_t optlen;
 861         int i, ret;
 862
 863         for (i = 0; i < dev->data->nb_rx_queues; i++) {
 864                 optlen = sizeof(struct xdp_statistics);
 865                 rxq = &internals->rx_queues[i];
 866                 txq = rxq->pair;
 867                 stats->q_ipackets[i] = rxq->stats.rx_pkts;
 868                 stats->q_ibytes[i] = rxq->stats.rx_bytes;
 869
 870                 stats->q_opackets[i] = txq->stats.tx_pkts;
 871                 stats->q_obytes[i] = txq->stats.tx_bytes;
 872
 873                 stats->ipackets += stats->q_ipackets[i];
 874                 stats->ibytes += stats->q_ibytes[i];
 875                 stats->imissed += rxq->stats.rx_dropped;
 876                 stats->oerrors += txq->stats.tx_dropped;
 877                 ret = getsockopt(xsk_socket__fd(rxq->xsk), SOL_XDP,
 878                                 XDP_STATISTICS, &xdp_stats, &optlen);
 879                 if (ret != 0) {
 880                         AF_XDP_LOG(ERR, "getsockopt() failed for XDP_STATISTICS.\n");
 881                         return -1;
 882                 }
 883                 stats->imissed += xdp_stats.rx_dropped;
 884
 885                 stats->opackets += stats->q_opackets[i];
 886                 stats->obytes += stats->q_obytes[i];
 887         }
 888
 889         return 0;
 890 }
 891
 892 static int
 893 eth_stats_reset(struct rte_eth_dev *dev)
 894 {
 895         struct pmd_internals *internals = dev->data->dev_private;
 896         int i;
 897
 898         for (i = 0; i < internals->queue_cnt; i++) {
 899                 memset(&internals->rx_queues[i].stats, 0,
 900                                         sizeof(struct rx_stats));
 901                 memset(&internals->tx_queues[i].stats, 0,
 902                                         sizeof(struct tx_stats));
 903         }
 904
 905         return 0;
 906 }
 907
 908 static void
 909 remove_xdp_program(struct pmd_internals *internals)
 910 {
 911         uint32_t curr_prog_id = 0;
 912
 913         if (bpf_get_link_xdp_id(internals->if_index, &curr_prog_id,
 914                                 XDP_FLAGS_UPDATE_IF_NOEXIST)) {
 915                 AF_XDP_LOG(ERR, "bpf_get_link_xdp_id failed\n");
 916                 return;
 917         }
 918         bpf_set_link_xdp_fd(internals->if_index, -1,
 919                         XDP_FLAGS_UPDATE_IF_NOEXIST);
 920 }
 921
 922 static void
 923 xdp_umem_destroy(struct xsk_umem_info *umem)
 924 {
 925 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 926         umem->mb_pool = NULL;
 927 #else
 928         rte_memzone_free(umem->mz);
 929         umem->mz = NULL;
 930
 931         rte_ring_free(umem->buf_ring);
 932         umem->buf_ring = NULL;
 933 #endif
 934
 935         rte_free(umem);
 936 }
 937
 938 static int
 939 eth_dev_close(struct rte_eth_dev *dev)
 940 {
 941         struct pmd_internals *internals = dev->data->dev_private;
 942         struct pkt_rx_queue *rxq;
 943         int i;
 944
 945         if (rte_eal_process_type() != RTE_PROC_PRIMARY)
 946                 return 0;
 947
 948         AF_XDP_LOG(INFO, "Closing AF_XDP ethdev on numa socket %u\n",
 949                 rte_socket_id());
 950
 951         for (i = 0; i < internals->queue_cnt; i++) {
 952                 rxq = &internals->rx_queues[i];
 953                 if (rxq->umem == NULL)
 954                         break;
 955                 xsk_socket__delete(rxq->xsk);
 956
 957                 if (__atomic_sub_fetch(&rxq->umem->refcnt, 1, __ATOMIC_ACQUIRE)
 958                                 == 0) {
 959                         (void)xsk_umem__delete(rxq->umem->umem);
 960                         xdp_umem_destroy(rxq->umem);
 961                 }
 962
 963                 /* free pkt_tx_queue */
 964                 rte_free(rxq->pair);
 965                 rte_free(rxq);
 966         }
 967
 968         /*
 969          * MAC is not allocated dynamically, setting it to NULL would prevent
 970          * from releasing it in rte_eth_dev_release_port.
 971          */
 972         dev->data->mac_addrs = NULL;
 973
 974         remove_xdp_program(internals);
 975
 976         if (internals->shared_umem) {
 977                 struct internal_list *list;
 978
 979                 /* Remove ethdev from list used to track and share UMEMs */
 980                 list = find_internal_resource(internals);
 981                 if (list) {
 982                         pthread_mutex_lock(&internal_list_lock);
 983                         TAILQ_REMOVE(&internal_list, list, next);
 984                         pthread_mutex_unlock(&internal_list_lock);
 985                         rte_free(list);
 986                 }
 987         }
 988
 989         return 0;
 990 }
 991
 992 static void
 993 eth_queue_release(void *q __rte_unused)
 994 {
 995 }
 996
 997 static int
 998 eth_link_update(struct rte_eth_dev *dev __rte_unused,
 999                 int wait_to_complete __rte_unused)
1000 {
1001         return 0;
1002 }
1003
1004 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
1005 static inline uintptr_t get_base_addr(struct rte_mempool *mp, uint64_t *align)
1006 {
1007         struct rte_mempool_memhdr *memhdr;
1008         uintptr_t memhdr_addr, aligned_addr;
1009
1010         memhdr = STAILQ_FIRST(&mp->mem_list);
1011         memhdr_addr = (uintptr_t)memhdr->addr;
1012         aligned_addr = memhdr_addr & ~(getpagesize() - 1);
1013         *align = memhdr_addr - aligned_addr;
1014
1015         return aligned_addr;
1016 }
1017
1018 static struct
1019 xsk_umem_info *xdp_umem_configure(struct pmd_internals *internals,
1020                                   struct pkt_rx_queue *rxq)
1021 {
1022         struct xsk_umem_info *umem = NULL;
1023         int ret;
1024         struct xsk_umem_config usr_config = {
1025                 .fill_size = ETH_AF_XDP_DFLT_NUM_DESCS * 2,
1026                 .comp_size = ETH_AF_XDP_DFLT_NUM_DESCS,
1027                 .flags = XDP_UMEM_UNALIGNED_CHUNK_FLAG};
1028         void *base_addr = NULL;
1029         struct rte_mempool *mb_pool = rxq->mb_pool;
1030         uint64_t umem_size, align = 0;
1031
1032         if (internals->shared_umem) {
1033                 if (get_shared_umem(rxq, internals->if_name, &umem) < 0)
1034                         return NULL;
1035
1036                 if (umem != NULL &&
1037                         __atomic_load_n(&umem->refcnt, __ATOMIC_ACQUIRE) <
1038                                         umem->max_xsks) {
1039                         AF_XDP_LOG(INFO, "%s,qid%i sharing UMEM\n",
1040                                         internals->if_name, rxq->xsk_queue_idx);
1041                         __atomic_fetch_add(&umem->refcnt, 1, __ATOMIC_ACQUIRE);
1042                 }
1043         }
1044
1045         if (umem == NULL) {
1046                 usr_config.frame_size =
1047                         rte_mempool_calc_obj_size(mb_pool->elt_size,
1048                                                   mb_pool->flags, NULL);
1049                 usr_config.frame_headroom = mb_pool->header_size +
1050                                                 sizeof(struct rte_mbuf) +
1051                                                 rte_pktmbuf_priv_size(mb_pool) +
1052                                                 RTE_PKTMBUF_HEADROOM;
1053
1054                 umem = rte_zmalloc_socket("umem", sizeof(*umem), 0,
1055                                           rte_socket_id());
1056                 if (umem == NULL) {
1057                         AF_XDP_LOG(ERR, "Failed to allocate umem info");
1058                         return NULL;
1059                 }
1060
1061                 umem->mb_pool = mb_pool;
1062                 base_addr = (void *)get_base_addr(mb_pool, &align);
1063                 umem_size = (uint64_t)mb_pool->populated_size *
1064                                 (uint64_t)usr_config.frame_size +
1065                                 align;
1066
1067                 ret = xsk_umem__create(&umem->umem, base_addr, umem_size,
1068                                 &rxq->fq, &rxq->cq, &usr_config);
1069                 if (ret) {
1070                         AF_XDP_LOG(ERR, "Failed to create umem");
1071                         goto err;
1072                 }
1073                 umem->buffer = base_addr;
1074
1075                 if (internals->shared_umem) {
1076                         umem->max_xsks = mb_pool->populated_size /
1077                                                 ETH_AF_XDP_NUM_BUFFERS;
1078                         AF_XDP_LOG(INFO, "Max xsks for UMEM %s: %u\n",
1079                                                 mb_pool->name, umem->max_xsks);
1080                 }
1081
1082                 __atomic_store_n(&umem->refcnt, 1, __ATOMIC_RELEASE);
1083         }
1084
1085 #else
1086 static struct
1087 xsk_umem_info *xdp_umem_configure(struct pmd_internals *internals,
1088                                   struct pkt_rx_queue *rxq)
1089 {
1090         struct xsk_umem_info *umem;
1091         const struct rte_memzone *mz;
1092         struct xsk_umem_config usr_config = {
1093                 .fill_size = ETH_AF_XDP_DFLT_NUM_DESCS,
1094                 .comp_size = ETH_AF_XDP_DFLT_NUM_DESCS,
1095                 .frame_size = ETH_AF_XDP_FRAME_SIZE,
1096                 .frame_headroom = 0 };
1097         char ring_name[RTE_RING_NAMESIZE];
1098         char mz_name[RTE_MEMZONE_NAMESIZE];
1099         int ret;
1100         uint64_t i;
1101
1102         umem = rte_zmalloc_socket("umem", sizeof(*umem), 0, rte_socket_id());
1103         if (umem == NULL) {
1104                 AF_XDP_LOG(ERR, "Failed to allocate umem info");
1105                 return NULL;
1106         }
1107
1108         snprintf(ring_name, sizeof(ring_name), "af_xdp_ring_%s_%u",
1109                        internals->if_name, rxq->xsk_queue_idx);
1110         umem->buf_ring = rte_ring_create(ring_name,
1111                                          ETH_AF_XDP_NUM_BUFFERS,
1112                                          rte_socket_id(),
1113                                          0x0);
1114         if (umem->buf_ring == NULL) {
1115                 AF_XDP_LOG(ERR, "Failed to create rte_ring\n");
1116                 goto err;
1117         }
1118
1119         for (i = 0; i < ETH_AF_XDP_NUM_BUFFERS; i++)
1120                 rte_ring_enqueue(umem->buf_ring,
1121                                  (void *)(i * ETH_AF_XDP_FRAME_SIZE));
1122
1123         snprintf(mz_name, sizeof(mz_name), "af_xdp_umem_%s_%u",
1124                        internals->if_name, rxq->xsk_queue_idx);
1125         mz = rte_memzone_reserve_aligned(mz_name,
1126                         ETH_AF_XDP_NUM_BUFFERS * ETH_AF_XDP_FRAME_SIZE,
1127                         rte_socket_id(), RTE_MEMZONE_IOVA_CONTIG,
1128                         getpagesize());
1129         if (mz == NULL) {
1130                 AF_XDP_LOG(ERR, "Failed to reserve memzone for af_xdp umem.\n");
1131                 goto err;
1132         }
1133
1134         ret = xsk_umem__create(&umem->umem, mz->addr,
1135                                ETH_AF_XDP_NUM_BUFFERS * ETH_AF_XDP_FRAME_SIZE,
1136                                &rxq->fq, &rxq->cq,
1137                                &usr_config);
1138
1139         if (ret) {
1140                 AF_XDP_LOG(ERR, "Failed to create umem");
1141                 goto err;
1142         }
1143         umem->mz = mz;
1144
1145 #endif
1146         return umem;
1147
1148 err:
1149         xdp_umem_destroy(umem);
1150         return NULL;
1151 }
1152
1153 static int
1154 load_custom_xdp_prog(const char *prog_path, int if_index)
1155 {
1156         int ret, prog_fd = -1;
1157         struct bpf_object *obj;
1158         struct bpf_map *map;
1159
1160         ret = bpf_prog_load(prog_path, BPF_PROG_TYPE_XDP, &obj, &prog_fd);
1161         if (ret) {
1162                 AF_XDP_LOG(ERR, "Failed to load program %s\n", prog_path);
1163                 return ret;
1164         }
1165
1166         /*
1167          * The loaded program must provision for a map of xsks, such that some
1168          * traffic can be redirected to userspace. When the xsk is created,
1169          * libbpf inserts it into the map.
1170          */
1171         map = bpf_object__find_map_by_name(obj, "xsks_map");
1172         if (!map) {
1173                 AF_XDP_LOG(ERR, "Failed to find xsks_map in %s\n", prog_path);
1174                 return -1;
1175         }
1176
1177         /* Link the program with the given network device */
1178         ret = bpf_set_link_xdp_fd(if_index, prog_fd,
1179                                         XDP_FLAGS_UPDATE_IF_NOEXIST);
1180         if (ret) {
1181                 AF_XDP_LOG(ERR, "Failed to set prog fd %d on interface\n",
1182                                 prog_fd);
1183                 return -1;
1184         }
1185
1186         AF_XDP_LOG(INFO, "Successfully loaded XDP program %s with fd %d\n",
1187                                 prog_path, prog_fd);
1188
1189         return 0;
1190 }
1191
1192 /* Detect support for busy polling through setsockopt(). */
1193 static int
1194 configure_preferred_busy_poll(struct pkt_rx_queue *rxq)
1195 {
1196         int sock_opt = 1;
1197         int fd = xsk_socket__fd(rxq->xsk);
1198         int ret = 0;
1199
1200         ret = setsockopt(fd, SOL_SOCKET, SO_PREFER_BUSY_POLL,
1201                         (void *)&sock_opt, sizeof(sock_opt));
1202         if (ret < 0) {
1203                 AF_XDP_LOG(DEBUG, "Failed to set SO_PREFER_BUSY_POLL\n");
1204                 goto err_prefer;
1205         }
1206
1207         sock_opt = ETH_AF_XDP_DFLT_BUSY_TIMEOUT;
1208         ret = setsockopt(fd, SOL_SOCKET, SO_BUSY_POLL, (void *)&sock_opt,
1209                         sizeof(sock_opt));
1210         if (ret < 0) {
1211                 AF_XDP_LOG(DEBUG, "Failed to set SO_BUSY_POLL\n");
1212                 goto err_timeout;
1213         }
1214
1215         sock_opt = rxq->busy_budget;
1216         ret = setsockopt(fd, SOL_SOCKET, SO_BUSY_POLL_BUDGET,
1217                         (void *)&sock_opt, sizeof(sock_opt));
1218         if (ret < 0) {
1219                 AF_XDP_LOG(DEBUG, "Failed to set SO_BUSY_POLL_BUDGET\n");
1220         } else {
1221                 AF_XDP_LOG(INFO, "Busy polling budget set to: %u\n",
1222                                         rxq->busy_budget);
1223                 return 0;
1224         }
1225
1226         /* setsockopt failure - attempt to restore xsk to default state and
1227          * proceed without busy polling support.
1228          */
1229         sock_opt = 0;
1230         ret = setsockopt(fd, SOL_SOCKET, SO_BUSY_POLL, (void *)&sock_opt,
1231                         sizeof(sock_opt));
1232         if (ret < 0) {
1233                 AF_XDP_LOG(ERR, "Failed to unset SO_BUSY_POLL\n");
1234                 return -1;
1235         }
1236
1237 err_timeout:
1238         sock_opt = 0;
1239         ret = setsockopt(fd, SOL_SOCKET, SO_PREFER_BUSY_POLL,
1240                         (void *)&sock_opt, sizeof(sock_opt));
1241         if (ret < 0) {
1242                 AF_XDP_LOG(ERR, "Failed to unset SO_PREFER_BUSY_POLL\n");
1243                 return -1;
1244         }
1245
1246 err_prefer:
1247         rxq->busy_budget = 0;
1248         return 0;
1249 }
1250
1251 static int
1252 xsk_configure(struct pmd_internals *internals, struct pkt_rx_queue *rxq,
1253               int ring_size)
1254 {
1255         struct xsk_socket_config cfg;
1256         struct pkt_tx_queue *txq = rxq->pair;
1257         int ret = 0;
1258         int reserve_size = ETH_AF_XDP_DFLT_NUM_DESCS;
1259         struct rte_mbuf *fq_bufs[reserve_size];
1260
1261         rxq->umem = xdp_umem_configure(internals, rxq);
1262         if (rxq->umem == NULL)
1263                 return -ENOMEM;
1264         txq->umem = rxq->umem;
1265
1266         cfg.rx_size = ring_size;
1267         cfg.tx_size = ring_size;
1268         cfg.libbpf_flags = 0;
1269         cfg.xdp_flags = XDP_FLAGS_UPDATE_IF_NOEXIST;
1270         cfg.bind_flags = 0;
1271
1272 #if defined(XDP_USE_NEED_WAKEUP)
1273         cfg.bind_flags |= XDP_USE_NEED_WAKEUP;
1274 #endif
1275
1276         if (strnlen(internals->prog_path, PATH_MAX) &&
1277                                 !internals->custom_prog_configured) {
1278                 ret = load_custom_xdp_prog(internals->prog_path,
1279                                            internals->if_index);
1280                 if (ret) {
1281                         AF_XDP_LOG(ERR, "Failed to load custom XDP program %s\n",
1282                                         internals->prog_path);
1283                         goto err;
1284                 }
1285                 internals->custom_prog_configured = 1;
1286         }
1287
1288         if (internals->shared_umem)
1289                 ret = create_shared_socket(&rxq->xsk, internals->if_name,
1290                                 rxq->xsk_queue_idx, rxq->umem->umem, &rxq->rx,
1291                                 &txq->tx, &rxq->fq, &rxq->cq, &cfg);
1292         else
1293                 ret = xsk_socket__create(&rxq->xsk, internals->if_name,
1294                                 rxq->xsk_queue_idx, rxq->umem->umem, &rxq->rx,
1295                                 &txq->tx, &cfg);
1296
1297         if (ret) {
1298                 AF_XDP_LOG(ERR, "Failed to create xsk socket.\n");
1299                 goto err;
1300         }
1301
1302 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
1303         ret = rte_pktmbuf_alloc_bulk(rxq->umem->mb_pool, fq_bufs, reserve_size);
1304         if (ret) {
1305                 AF_XDP_LOG(DEBUG, "Failed to get enough buffers for fq.\n");
1306                 goto err;
1307         }
1308 #endif
1309
1310         if (rxq->busy_budget) {
1311                 ret = configure_preferred_busy_poll(rxq);
1312                 if (ret) {
1313                         AF_XDP_LOG(ERR, "Failed configure busy polling.\n");
1314                         goto err;
1315                 }
1316         }
1317
1318         ret = reserve_fill_queue(rxq->umem, reserve_size, fq_bufs, &rxq->fq);
1319         if (ret) {
1320                 xsk_socket__delete(rxq->xsk);
1321                 AF_XDP_LOG(ERR, "Failed to reserve fill queue.\n");
1322                 goto err;
1323         }
1324
1325         return 0;
1326
1327 err:
1328         if (__atomic_sub_fetch(&rxq->umem->refcnt, 1, __ATOMIC_ACQUIRE) == 0)
1329                 xdp_umem_destroy(rxq->umem);
1330
1331         return ret;
1332 }
1333
1334 static int
1335 eth_rx_queue_setup(struct rte_eth_dev *dev,
1336                    uint16_t rx_queue_id,
1337                    uint16_t nb_rx_desc,
1338                    unsigned int socket_id __rte_unused,
1339                    const struct rte_eth_rxconf *rx_conf __rte_unused,
1340                    struct rte_mempool *mb_pool)
1341 {
1342         struct pmd_internals *internals = dev->data->dev_private;
1343         struct pkt_rx_queue *rxq;
1344         int ret;
1345
1346         rxq = &internals->rx_queues[rx_queue_id];
1347
1348         AF_XDP_LOG(INFO, "Set up rx queue, rx queue id: %d, xsk queue id: %d\n",
1349                    rx_queue_id, rxq->xsk_queue_idx);
1350
1351 #ifndef XDP_UMEM_UNALIGNED_CHUNK_FLAG
1352         uint32_t buf_size, data_size;
1353
1354         /* Now get the space available for data in the mbuf */
1355         buf_size = rte_pktmbuf_data_room_size(mb_pool) -
1356                 RTE_PKTMBUF_HEADROOM;
1357         data_size = ETH_AF_XDP_FRAME_SIZE;
1358
1359         if (data_size > buf_size) {
1360                 AF_XDP_LOG(ERR, "%s: %d bytes will not fit in mbuf (%d bytes)\n",
1361                         dev->device->name, data_size, buf_size);
1362                 ret = -ENOMEM;
1363                 goto err;
1364         }
1365 #endif
1366
1367         rxq->mb_pool = mb_pool;
1368
1369         if (xsk_configure(internals, rxq, nb_rx_desc)) {
1370                 AF_XDP_LOG(ERR, "Failed to configure xdp socket\n");
1371                 ret = -EINVAL;
1372                 goto err;
1373         }
1374
1375         if (!rxq->busy_budget)
1376                 AF_XDP_LOG(DEBUG, "Preferred busy polling not enabled\n");
1377
1378         rxq->fds[0].fd = xsk_socket__fd(rxq->xsk);
1379         rxq->fds[0].events = POLLIN;
1380
1381         dev->data->rx_queues[rx_queue_id] = rxq;
1382         return 0;
1383
1384 err:
1385         return ret;
1386 }
1387
1388 static int
1389 eth_tx_queue_setup(struct rte_eth_dev *dev,
1390                    uint16_t tx_queue_id,
1391                    uint16_t nb_tx_desc __rte_unused,
1392                    unsigned int socket_id __rte_unused,
1393                    const struct rte_eth_txconf *tx_conf __rte_unused)
1394 {
1395         struct pmd_internals *internals = dev->data->dev_private;
1396         struct pkt_tx_queue *txq;
1397
1398         txq = &internals->tx_queues[tx_queue_id];
1399
1400         dev->data->tx_queues[tx_queue_id] = txq;
1401         return 0;
1402 }
1403
1404 static int
1405 eth_dev_mtu_set(struct rte_eth_dev *dev, uint16_t mtu)
1406 {
1407         struct pmd_internals *internals = dev->data->dev_private;
1408         struct ifreq ifr = { .ifr_mtu = mtu };
1409         int ret;
1410         int s;
1411
1412         s = socket(PF_INET, SOCK_DGRAM, 0);
1413         if (s < 0)
1414                 return -EINVAL;
1415
1416         strlcpy(ifr.ifr_name, internals->if_name, IFNAMSIZ);
1417         ret = ioctl(s, SIOCSIFMTU, &ifr);
1418         close(s);
1419
1420         return (ret < 0) ? -errno : 0;
1421 }
1422
1423 static int
1424 eth_dev_change_flags(char *if_name, uint32_t flags, uint32_t mask)
1425 {
1426         struct ifreq ifr;
1427         int ret = 0;
1428         int s;
1429
1430         s = socket(PF_INET, SOCK_DGRAM, 0);
1431         if (s < 0)
1432                 return -errno;
1433
1434         strlcpy(ifr.ifr_name, if_name, IFNAMSIZ);
1435         if (ioctl(s, SIOCGIFFLAGS, &ifr) < 0) {
1436                 ret = -errno;
1437                 goto out;
1438         }
1439         ifr.ifr_flags &= mask;
1440         ifr.ifr_flags |= flags;
1441         if (ioctl(s, SIOCSIFFLAGS, &ifr) < 0) {
1442                 ret = -errno;
1443                 goto out;
1444         }
1445 out:
1446         close(s);
1447         return ret;
1448 }
1449
1450 static int
1451 eth_dev_promiscuous_enable(struct rte_eth_dev *dev)
1452 {
1453         struct pmd_internals *internals = dev->data->dev_private;
1454
1455         return eth_dev_change_flags(internals->if_name, IFF_PROMISC, ~0);
1456 }
1457
1458 static int
1459 eth_dev_promiscuous_disable(struct rte_eth_dev *dev)
1460 {
1461         struct pmd_internals *internals = dev->data->dev_private;
1462
1463         return eth_dev_change_flags(internals->if_name, 0, ~IFF_PROMISC);
1464 }
1465
1466 static const struct eth_dev_ops ops = {
1467         .dev_start = eth_dev_start,
1468         .dev_stop = eth_dev_stop,
1469         .dev_close = eth_dev_close,
1470         .dev_configure = eth_dev_configure,
1471         .dev_infos_get = eth_dev_info,
1472         .mtu_set = eth_dev_mtu_set,
1473         .promiscuous_enable = eth_dev_promiscuous_enable,
1474         .promiscuous_disable = eth_dev_promiscuous_disable,
1475         .rx_queue_setup = eth_rx_queue_setup,
1476         .tx_queue_setup = eth_tx_queue_setup,
1477         .rx_queue_release = eth_queue_release,
1478         .tx_queue_release = eth_queue_release,
1479         .link_update = eth_link_update,
1480         .stats_get = eth_stats_get,
1481         .stats_reset = eth_stats_reset,
1482         .get_monitor_addr = eth_get_monitor_addr,
1483 };
1484
1485 /** parse busy_budget argument */
1486 static int
1487 parse_budget_arg(const char *key __rte_unused,
1488                   const char *value, void *extra_args)
1489 {
1490         int *i = (int *)extra_args;
1491         char *end;
1492
1493         *i = strtol(value, &end, 10);
1494         if (*i < 0 || *i > UINT16_MAX) {
1495                 AF_XDP_LOG(ERR, "Invalid busy_budget %i, must be >= 0 and <= %u\n",
1496                                 *i, UINT16_MAX);
1497                 return -EINVAL;
1498         }
1499
1500         return 0;
1501 }
1502
1503 /** parse integer from integer argument */
1504 static int
1505 parse_integer_arg(const char *key __rte_unused,
1506                   const char *value, void *extra_args)
1507 {
1508         int *i = (int *)extra_args;
1509         char *end;
1510
1511         *i = strtol(value, &end, 10);
1512         if (*i < 0) {
1513                 AF_XDP_LOG(ERR, "Argument has to be positive.\n");
1514                 return -EINVAL;
1515         }
1516
1517         return 0;
1518 }
1519
1520 /** parse name argument */
1521 static int
1522 parse_name_arg(const char *key __rte_unused,
1523                const char *value, void *extra_args)
1524 {
1525         char *name = extra_args;
1526
1527         if (strnlen(value, IFNAMSIZ) > IFNAMSIZ - 1) {
1528                 AF_XDP_LOG(ERR, "Invalid name %s, should be less than %u bytes.\n",
1529                            value, IFNAMSIZ);
1530                 return -EINVAL;
1531         }
1532
1533         strlcpy(name, value, IFNAMSIZ);
1534
1535         return 0;
1536 }
1537
1538 /** parse xdp prog argument */
1539 static int
1540 parse_prog_arg(const char *key __rte_unused,
1541                const char *value, void *extra_args)
1542 {
1543         char *path = extra_args;
1544
1545         if (strnlen(value, PATH_MAX) == PATH_MAX) {
1546                 AF_XDP_LOG(ERR, "Invalid path %s, should be less than %u bytes.\n",
1547                            value, PATH_MAX);
1548                 return -EINVAL;
1549         }
1550
1551         if (access(value, F_OK) != 0) {
1552                 AF_XDP_LOG(ERR, "Error accessing %s: %s\n",
1553                            value, strerror(errno));
1554                 return -EINVAL;
1555         }
1556
1557         strlcpy(path, value, PATH_MAX);
1558
1559         return 0;
1560 }
1561
1562 static int
1563 xdp_get_channels_info(const char *if_name, int *max_queues,
1564                                 int *combined_queues)
1565 {
1566         struct ethtool_channels channels;
1567         struct ifreq ifr;
1568         int fd, ret;
1569
1570         fd = socket(AF_INET, SOCK_DGRAM, 0);
1571         if (fd < 0)
1572                 return -1;
1573
1574         channels.cmd = ETHTOOL_GCHANNELS;
1575         ifr.ifr_data = (void *)&channels;
1576         strlcpy(ifr.ifr_name, if_name, IFNAMSIZ);
1577         ret = ioctl(fd, SIOCETHTOOL, &ifr);
1578         if (ret) {
1579                 if (errno == EOPNOTSUPP) {
1580                         ret = 0;
1581                 } else {
1582                         ret = -errno;
1583                         goto out;
1584                 }
1585         }
1586
1587         if (channels.max_combined == 0 || errno == EOPNOTSUPP) {
1588                 /* If the device says it has no channels, then all traffic
1589                  * is sent to a single stream, so max queues = 1.
1590                  */
1591                 *max_queues = 1;
1592                 *combined_queues = 1;
1593         } else {
1594                 *max_queues = channels.max_combined;
1595                 *combined_queues = channels.combined_count;
1596         }
1597
1598  out:
1599         close(fd);
1600         return ret;
1601 }
1602
1603 static int
1604 parse_parameters(struct rte_kvargs *kvlist, char *if_name, int *start_queue,
1605                         int *queue_cnt, int *shared_umem, char *prog_path,
1606                         int *busy_budget)
1607 {
1608         int ret;
1609
1610         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_IFACE_ARG,
1611                                  &parse_name_arg, if_name);
1612         if (ret < 0)
1613                 goto free_kvlist;
1614
1615         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_START_QUEUE_ARG,
1616                                  &parse_integer_arg, start_queue);
1617         if (ret < 0)
1618                 goto free_kvlist;
1619
1620         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_QUEUE_COUNT_ARG,
1621                                  &parse_integer_arg, queue_cnt);
1622         if (ret < 0 || *queue_cnt <= 0) {
1623                 ret = -EINVAL;
1624                 goto free_kvlist;
1625         }
1626
1627         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_SHARED_UMEM_ARG,
1628                                 &parse_integer_arg, shared_umem);
1629         if (ret < 0)
1630                 goto free_kvlist;
1631
1632         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_PROG_ARG,
1633                                  &parse_prog_arg, prog_path);
1634         if (ret < 0)
1635                 goto free_kvlist;
1636
1637         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_BUDGET_ARG,
1638                                 &parse_budget_arg, busy_budget);
1639         if (ret < 0)
1640                 goto free_kvlist;
1641
1642 free_kvlist:
1643         rte_kvargs_free(kvlist);
1644         return ret;
1645 }
1646
1647 static int
1648 get_iface_info(const char *if_name,
1649                struct rte_ether_addr *eth_addr,
1650                int *if_index)
1651 {
1652         struct ifreq ifr;
1653         int sock = socket(AF_INET, SOCK_DGRAM, IPPROTO_IP);
1654
1655         if (sock < 0)
1656                 return -1;
1657
1658         strlcpy(ifr.ifr_name, if_name, IFNAMSIZ);
1659         if (ioctl(sock, SIOCGIFINDEX, &ifr))
1660                 goto error;
1661
1662         *if_index = ifr.ifr_ifindex;
1663
1664         if (ioctl(sock, SIOCGIFHWADDR, &ifr))
1665                 goto error;
1666
1667         rte_memcpy(eth_addr, ifr.ifr_hwaddr.sa_data, RTE_ETHER_ADDR_LEN);
1668
1669         close(sock);
1670         return 0;
1671
1672 error:
1673         close(sock);
1674         return -1;
1675 }
1676
1677 static struct rte_eth_dev *
1678 init_internals(struct rte_vdev_device *dev, const char *if_name,
1679                 int start_queue_idx, int queue_cnt, int shared_umem,
1680                 const char *prog_path, int busy_budget)
1681 {
1682         const char *name = rte_vdev_device_name(dev);
1683         const unsigned int numa_node = dev->device.numa_node;
1684         struct pmd_internals *internals;
1685         struct rte_eth_dev *eth_dev;
1686         int ret;
1687         int i;
1688
1689         internals = rte_zmalloc_socket(name, sizeof(*internals), 0, numa_node);
1690         if (internals == NULL)
1691                 return NULL;
1692
1693         internals->start_queue_idx = start_queue_idx;
1694         internals->queue_cnt = queue_cnt;
1695         strlcpy(internals->if_name, if_name, IFNAMSIZ);
1696         strlcpy(internals->prog_path, prog_path, PATH_MAX);
1697         internals->custom_prog_configured = 0;
1698
1699 #ifndef ETH_AF_XDP_SHARED_UMEM
1700         if (shared_umem) {
1701                 AF_XDP_LOG(ERR, "Shared UMEM feature not available. "
1702                                 "Check kernel and libbpf version\n");
1703                 goto err_free_internals;
1704         }
1705 #endif
1706         internals->shared_umem = shared_umem;
1707
1708         if (xdp_get_channels_info(if_name, &internals->max_queue_cnt,
1709                                   &internals->combined_queue_cnt)) {
1710                 AF_XDP_LOG(ERR, "Failed to get channel info of interface: %s\n",
1711                                 if_name);
1712                 goto err_free_internals;
1713         }
1714
1715         if (queue_cnt > internals->combined_queue_cnt) {
1716                 AF_XDP_LOG(ERR, "Specified queue count %d is larger than combined queue count %d.\n",
1717                                 queue_cnt, internals->combined_queue_cnt);
1718                 goto err_free_internals;
1719         }
1720
1721         internals->rx_queues = rte_zmalloc_socket(NULL,
1722                                         sizeof(struct pkt_rx_queue) * queue_cnt,
1723                                         0, numa_node);
1724         if (internals->rx_queues == NULL) {
1725                 AF_XDP_LOG(ERR, "Failed to allocate memory for rx queues.\n");
1726                 goto err_free_internals;
1727         }
1728
1729         internals->tx_queues = rte_zmalloc_socket(NULL,
1730                                         sizeof(struct pkt_tx_queue) * queue_cnt,
1731                                         0, numa_node);
1732         if (internals->tx_queues == NULL) {
1733                 AF_XDP_LOG(ERR, "Failed to allocate memory for tx queues.\n");
1734                 goto err_free_rx;
1735         }
1736         for (i = 0; i < queue_cnt; i++) {
1737                 internals->tx_queues[i].pair = &internals->rx_queues[i];
1738                 internals->rx_queues[i].pair = &internals->tx_queues[i];
1739                 internals->rx_queues[i].xsk_queue_idx = start_queue_idx + i;
1740                 internals->tx_queues[i].xsk_queue_idx = start_queue_idx + i;
1741                 internals->rx_queues[i].busy_budget = busy_budget;
1742         }
1743
1744         ret = get_iface_info(if_name, &internals->eth_addr,
1745                              &internals->if_index);
1746         if (ret)
1747                 goto err_free_tx;
1748
1749         eth_dev = rte_eth_vdev_allocate(dev, 0);
1750         if (eth_dev == NULL)
1751                 goto err_free_tx;
1752
1753         eth_dev->data->dev_private = internals;
1754         eth_dev->data->dev_link = pmd_link;
1755         eth_dev->data->mac_addrs = &internals->eth_addr;
1756         eth_dev->data->dev_flags |= RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS;
1757         eth_dev->dev_ops = &ops;
1758         eth_dev->rx_pkt_burst = eth_af_xdp_rx;
1759         eth_dev->tx_pkt_burst = eth_af_xdp_tx;
1760
1761 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
1762         AF_XDP_LOG(INFO, "Zero copy between umem and mbuf enabled.\n");
1763 #endif
1764
1765         return eth_dev;
1766
1767 err_free_tx:
1768         rte_free(internals->tx_queues);
1769 err_free_rx:
1770         rte_free(internals->rx_queues);
1771 err_free_internals:
1772         rte_free(internals);
1773         return NULL;
1774 }
1775
1776 static int
1777 rte_pmd_af_xdp_probe(struct rte_vdev_device *dev)
1778 {
1779         struct rte_kvargs *kvlist;
1780         char if_name[IFNAMSIZ] = {'\0'};
1781         int xsk_start_queue_idx = ETH_AF_XDP_DFLT_START_QUEUE_IDX;
1782         int xsk_queue_cnt = ETH_AF_XDP_DFLT_QUEUE_COUNT;
1783         int shared_umem = 0;
1784         char prog_path[PATH_MAX] = {'\0'};
1785         int busy_budget = -1;
1786         struct rte_eth_dev *eth_dev = NULL;
1787         const char *name;
1788
1789         AF_XDP_LOG(INFO, "Initializing pmd_af_xdp for %s\n",
1790                 rte_vdev_device_name(dev));
1791
1792         name = rte_vdev_device_name(dev);
1793         if (rte_eal_process_type() == RTE_PROC_SECONDARY &&
1794                 strlen(rte_vdev_device_args(dev)) == 0) {
1795                 eth_dev = rte_eth_dev_attach_secondary(name);
1796                 if (eth_dev == NULL) {
1797                         AF_XDP_LOG(ERR, "Failed to probe %s\n", name);
1798                         return -EINVAL;
1799                 }
1800                 eth_dev->dev_ops = &ops;
1801                 rte_eth_dev_probing_finish(eth_dev);
1802                 return 0;
1803         }
1804
1805         kvlist = rte_kvargs_parse(rte_vdev_device_args(dev), valid_arguments);
1806         if (kvlist == NULL) {
1807                 AF_XDP_LOG(ERR, "Invalid kvargs key\n");
1808                 return -EINVAL;
1809         }
1810
1811         if (dev->device.numa_node == SOCKET_ID_ANY)
1812                 dev->device.numa_node = rte_socket_id();
1813
1814         if (parse_parameters(kvlist, if_name, &xsk_start_queue_idx,
1815                              &xsk_queue_cnt, &shared_umem, prog_path,
1816                              &busy_budget) < 0) {
1817                 AF_XDP_LOG(ERR, "Invalid kvargs value\n");
1818                 return -EINVAL;
1819         }
1820
1821         if (strlen(if_name) == 0) {
1822                 AF_XDP_LOG(ERR, "Network interface must be specified\n");
1823                 return -EINVAL;
1824         }
1825
1826         busy_budget = busy_budget == -1 ? ETH_AF_XDP_DFLT_BUSY_BUDGET :
1827                                         busy_budget;
1828
1829         eth_dev = init_internals(dev, if_name, xsk_start_queue_idx,
1830                                         xsk_queue_cnt, shared_umem, prog_path,
1831                                         busy_budget);
1832         if (eth_dev == NULL) {
1833                 AF_XDP_LOG(ERR, "Failed to init internals\n");
1834                 return -1;
1835         }
1836
1837         rte_eth_dev_probing_finish(eth_dev);
1838
1839         return 0;
1840 }
1841
1842 static int
1843 rte_pmd_af_xdp_remove(struct rte_vdev_device *dev)
1844 {
1845         struct rte_eth_dev *eth_dev = NULL;
1846
1847         AF_XDP_LOG(INFO, "Removing AF_XDP ethdev on numa socket %u\n",
1848                 rte_socket_id());
1849
1850         if (dev == NULL)
1851                 return -1;
1852
1853         /* find the ethdev entry */
1854         eth_dev = rte_eth_dev_allocated(rte_vdev_device_name(dev));
1855         if (eth_dev == NULL)
1856                 return 0;
1857
1858         eth_dev_close(eth_dev);
1859         rte_eth_dev_release_port(eth_dev);
1860
1861
1862         return 0;
1863 }
1864
1865 static struct rte_vdev_driver pmd_af_xdp_drv = {
1866         .probe = rte_pmd_af_xdp_probe,
1867         .remove = rte_pmd_af_xdp_remove,
1868 };
1869
1870 RTE_PMD_REGISTER_VDEV(net_af_xdp, pmd_af_xdp_drv);
1871 RTE_PMD_REGISTER_PARAM_STRING(net_af_xdp,
1872                               "iface=<string> "
1873                               "start_queue=<int> "
1874                               "queue_count=<int> "
1875                               "shared_umem=<int> "
1876                               "xdp_prog=<string> "
1877                               "busy_budget=<int>");