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 int
 993 eth_link_update(struct rte_eth_dev *dev __rte_unused,
 994                 int wait_to_complete __rte_unused)
 995 {
 996         return 0;
 997 }
 998
 999 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
1000 static inline uintptr_t get_base_addr(struct rte_mempool *mp, uint64_t *align)
1001 {
1002         struct rte_mempool_memhdr *memhdr;
1003         uintptr_t memhdr_addr, aligned_addr;
1004
1005         memhdr = STAILQ_FIRST(&mp->mem_list);
1006         memhdr_addr = (uintptr_t)memhdr->addr;
1007         aligned_addr = memhdr_addr & ~(getpagesize() - 1);
1008         *align = memhdr_addr - aligned_addr;
1009
1010         return aligned_addr;
1011 }
1012
1013 static struct
1014 xsk_umem_info *xdp_umem_configure(struct pmd_internals *internals,
1015                                   struct pkt_rx_queue *rxq)
1016 {
1017         struct xsk_umem_info *umem = NULL;
1018         int ret;
1019         struct xsk_umem_config usr_config = {
1020                 .fill_size = ETH_AF_XDP_DFLT_NUM_DESCS * 2,
1021                 .comp_size = ETH_AF_XDP_DFLT_NUM_DESCS,
1022                 .flags = XDP_UMEM_UNALIGNED_CHUNK_FLAG};
1023         void *base_addr = NULL;
1024         struct rte_mempool *mb_pool = rxq->mb_pool;
1025         uint64_t umem_size, align = 0;
1026
1027         if (internals->shared_umem) {
1028                 if (get_shared_umem(rxq, internals->if_name, &umem) < 0)
1029                         return NULL;
1030
1031                 if (umem != NULL &&
1032                         __atomic_load_n(&umem->refcnt, __ATOMIC_ACQUIRE) <
1033                                         umem->max_xsks) {
1034                         AF_XDP_LOG(INFO, "%s,qid%i sharing UMEM\n",
1035                                         internals->if_name, rxq->xsk_queue_idx);
1036                         __atomic_fetch_add(&umem->refcnt, 1, __ATOMIC_ACQUIRE);
1037                 }
1038         }
1039
1040         if (umem == NULL) {
1041                 usr_config.frame_size =
1042                         rte_mempool_calc_obj_size(mb_pool->elt_size,
1043                                                   mb_pool->flags, NULL);
1044                 usr_config.frame_headroom = mb_pool->header_size +
1045                                                 sizeof(struct rte_mbuf) +
1046                                                 rte_pktmbuf_priv_size(mb_pool) +
1047                                                 RTE_PKTMBUF_HEADROOM;
1048
1049                 umem = rte_zmalloc_socket("umem", sizeof(*umem), 0,
1050                                           rte_socket_id());
1051                 if (umem == NULL) {
1052                         AF_XDP_LOG(ERR, "Failed to allocate umem info");
1053                         return NULL;
1054                 }
1055
1056                 umem->mb_pool = mb_pool;
1057                 base_addr = (void *)get_base_addr(mb_pool, &align);
1058                 umem_size = (uint64_t)mb_pool->populated_size *
1059                                 (uint64_t)usr_config.frame_size +
1060                                 align;
1061
1062                 ret = xsk_umem__create(&umem->umem, base_addr, umem_size,
1063                                 &rxq->fq, &rxq->cq, &usr_config);
1064                 if (ret) {
1065                         AF_XDP_LOG(ERR, "Failed to create umem");
1066                         goto err;
1067                 }
1068                 umem->buffer = base_addr;
1069
1070                 if (internals->shared_umem) {
1071                         umem->max_xsks = mb_pool->populated_size /
1072                                                 ETH_AF_XDP_NUM_BUFFERS;
1073                         AF_XDP_LOG(INFO, "Max xsks for UMEM %s: %u\n",
1074                                                 mb_pool->name, umem->max_xsks);
1075                 }
1076
1077                 __atomic_store_n(&umem->refcnt, 1, __ATOMIC_RELEASE);
1078         }
1079
1080 #else
1081 static struct
1082 xsk_umem_info *xdp_umem_configure(struct pmd_internals *internals,
1083                                   struct pkt_rx_queue *rxq)
1084 {
1085         struct xsk_umem_info *umem;
1086         const struct rte_memzone *mz;
1087         struct xsk_umem_config usr_config = {
1088                 .fill_size = ETH_AF_XDP_DFLT_NUM_DESCS,
1089                 .comp_size = ETH_AF_XDP_DFLT_NUM_DESCS,
1090                 .frame_size = ETH_AF_XDP_FRAME_SIZE,
1091                 .frame_headroom = 0 };
1092         char ring_name[RTE_RING_NAMESIZE];
1093         char mz_name[RTE_MEMZONE_NAMESIZE];
1094         int ret;
1095         uint64_t i;
1096
1097         umem = rte_zmalloc_socket("umem", sizeof(*umem), 0, rte_socket_id());
1098         if (umem == NULL) {
1099                 AF_XDP_LOG(ERR, "Failed to allocate umem info");
1100                 return NULL;
1101         }
1102
1103         snprintf(ring_name, sizeof(ring_name), "af_xdp_ring_%s_%u",
1104                        internals->if_name, rxq->xsk_queue_idx);
1105         umem->buf_ring = rte_ring_create(ring_name,
1106                                          ETH_AF_XDP_NUM_BUFFERS,
1107                                          rte_socket_id(),
1108                                          0x0);
1109         if (umem->buf_ring == NULL) {
1110                 AF_XDP_LOG(ERR, "Failed to create rte_ring\n");
1111                 goto err;
1112         }
1113
1114         for (i = 0; i < ETH_AF_XDP_NUM_BUFFERS; i++)
1115                 rte_ring_enqueue(umem->buf_ring,
1116                                  (void *)(i * ETH_AF_XDP_FRAME_SIZE));
1117
1118         snprintf(mz_name, sizeof(mz_name), "af_xdp_umem_%s_%u",
1119                        internals->if_name, rxq->xsk_queue_idx);
1120         mz = rte_memzone_reserve_aligned(mz_name,
1121                         ETH_AF_XDP_NUM_BUFFERS * ETH_AF_XDP_FRAME_SIZE,
1122                         rte_socket_id(), RTE_MEMZONE_IOVA_CONTIG,
1123                         getpagesize());
1124         if (mz == NULL) {
1125                 AF_XDP_LOG(ERR, "Failed to reserve memzone for af_xdp umem.\n");
1126                 goto err;
1127         }
1128
1129         ret = xsk_umem__create(&umem->umem, mz->addr,
1130                                ETH_AF_XDP_NUM_BUFFERS * ETH_AF_XDP_FRAME_SIZE,
1131                                &rxq->fq, &rxq->cq,
1132                                &usr_config);
1133
1134         if (ret) {
1135                 AF_XDP_LOG(ERR, "Failed to create umem");
1136                 goto err;
1137         }
1138         umem->mz = mz;
1139
1140 #endif
1141         return umem;
1142
1143 err:
1144         xdp_umem_destroy(umem);
1145         return NULL;
1146 }
1147
1148 static int
1149 load_custom_xdp_prog(const char *prog_path, int if_index)
1150 {
1151         int ret, prog_fd = -1;
1152         struct bpf_object *obj;
1153         struct bpf_map *map;
1154
1155         ret = bpf_prog_load(prog_path, BPF_PROG_TYPE_XDP, &obj, &prog_fd);
1156         if (ret) {
1157                 AF_XDP_LOG(ERR, "Failed to load program %s\n", prog_path);
1158                 return ret;
1159         }
1160
1161         /*
1162          * The loaded program must provision for a map of xsks, such that some
1163          * traffic can be redirected to userspace. When the xsk is created,
1164          * libbpf inserts it into the map.
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                 if (ret) {
1276                         AF_XDP_LOG(ERR, "Failed to load custom XDP program %s\n",
1277                                         internals->prog_path);
1278                         goto err;
1279                 }
1280                 internals->custom_prog_configured = 1;
1281         }
1282
1283         if (internals->shared_umem)
1284                 ret = create_shared_socket(&rxq->xsk, internals->if_name,
1285                                 rxq->xsk_queue_idx, rxq->umem->umem, &rxq->rx,
1286                                 &txq->tx, &rxq->fq, &rxq->cq, &cfg);
1287         else
1288                 ret = xsk_socket__create(&rxq->xsk, internals->if_name,
1289                                 rxq->xsk_queue_idx, rxq->umem->umem, &rxq->rx,
1290                                 &txq->tx, &cfg);
1291
1292         if (ret) {
1293                 AF_XDP_LOG(ERR, "Failed to create xsk socket.\n");
1294                 goto err;
1295         }
1296
1297 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
1298         ret = rte_pktmbuf_alloc_bulk(rxq->umem->mb_pool, fq_bufs, reserve_size);
1299         if (ret) {
1300                 AF_XDP_LOG(DEBUG, "Failed to get enough buffers for fq.\n");
1301                 goto err;
1302         }
1303 #endif
1304
1305         if (rxq->busy_budget) {
1306                 ret = configure_preferred_busy_poll(rxq);
1307                 if (ret) {
1308                         AF_XDP_LOG(ERR, "Failed configure busy polling.\n");
1309                         goto err;
1310                 }
1311         }
1312
1313         ret = reserve_fill_queue(rxq->umem, reserve_size, fq_bufs, &rxq->fq);
1314         if (ret) {
1315                 xsk_socket__delete(rxq->xsk);
1316                 AF_XDP_LOG(ERR, "Failed to reserve fill queue.\n");
1317                 goto err;
1318         }
1319
1320         return 0;
1321
1322 err:
1323         if (__atomic_sub_fetch(&rxq->umem->refcnt, 1, __ATOMIC_ACQUIRE) == 0)
1324                 xdp_umem_destroy(rxq->umem);
1325
1326         return ret;
1327 }
1328
1329 static int
1330 eth_rx_queue_setup(struct rte_eth_dev *dev,
1331                    uint16_t rx_queue_id,
1332                    uint16_t nb_rx_desc,
1333                    unsigned int socket_id __rte_unused,
1334                    const struct rte_eth_rxconf *rx_conf __rte_unused,
1335                    struct rte_mempool *mb_pool)
1336 {
1337         struct pmd_internals *internals = dev->data->dev_private;
1338         struct pkt_rx_queue *rxq;
1339         int ret;
1340
1341         rxq = &internals->rx_queues[rx_queue_id];
1342
1343         AF_XDP_LOG(INFO, "Set up rx queue, rx queue id: %d, xsk queue id: %d\n",
1344                    rx_queue_id, rxq->xsk_queue_idx);
1345
1346 #ifndef XDP_UMEM_UNALIGNED_CHUNK_FLAG
1347         uint32_t buf_size, data_size;
1348
1349         /* Now get the space available for data in the mbuf */
1350         buf_size = rte_pktmbuf_data_room_size(mb_pool) -
1351                 RTE_PKTMBUF_HEADROOM;
1352         data_size = ETH_AF_XDP_FRAME_SIZE;
1353
1354         if (data_size > buf_size) {
1355                 AF_XDP_LOG(ERR, "%s: %d bytes will not fit in mbuf (%d bytes)\n",
1356                         dev->device->name, data_size, buf_size);
1357                 ret = -ENOMEM;
1358                 goto err;
1359         }
1360 #endif
1361
1362         rxq->mb_pool = mb_pool;
1363
1364         if (xsk_configure(internals, rxq, nb_rx_desc)) {
1365                 AF_XDP_LOG(ERR, "Failed to configure xdp socket\n");
1366                 ret = -EINVAL;
1367                 goto err;
1368         }
1369
1370         if (!rxq->busy_budget)
1371                 AF_XDP_LOG(DEBUG, "Preferred busy polling not enabled\n");
1372
1373         rxq->fds[0].fd = xsk_socket__fd(rxq->xsk);
1374         rxq->fds[0].events = POLLIN;
1375
1376         dev->data->rx_queues[rx_queue_id] = rxq;
1377         return 0;
1378
1379 err:
1380         return ret;
1381 }
1382
1383 static int
1384 eth_tx_queue_setup(struct rte_eth_dev *dev,
1385                    uint16_t tx_queue_id,
1386                    uint16_t nb_tx_desc __rte_unused,
1387                    unsigned int socket_id __rte_unused,
1388                    const struct rte_eth_txconf *tx_conf __rte_unused)
1389 {
1390         struct pmd_internals *internals = dev->data->dev_private;
1391         struct pkt_tx_queue *txq;
1392
1393         txq = &internals->tx_queues[tx_queue_id];
1394
1395         dev->data->tx_queues[tx_queue_id] = txq;
1396         return 0;
1397 }
1398
1399 static int
1400 eth_dev_mtu_set(struct rte_eth_dev *dev, uint16_t mtu)
1401 {
1402         struct pmd_internals *internals = dev->data->dev_private;
1403         struct ifreq ifr = { .ifr_mtu = mtu };
1404         int ret;
1405         int s;
1406
1407         s = socket(PF_INET, SOCK_DGRAM, 0);
1408         if (s < 0)
1409                 return -EINVAL;
1410
1411         strlcpy(ifr.ifr_name, internals->if_name, IFNAMSIZ);
1412         ret = ioctl(s, SIOCSIFMTU, &ifr);
1413         close(s);
1414
1415         return (ret < 0) ? -errno : 0;
1416 }
1417
1418 static int
1419 eth_dev_change_flags(char *if_name, uint32_t flags, uint32_t mask)
1420 {
1421         struct ifreq ifr;
1422         int ret = 0;
1423         int s;
1424
1425         s = socket(PF_INET, SOCK_DGRAM, 0);
1426         if (s < 0)
1427                 return -errno;
1428
1429         strlcpy(ifr.ifr_name, if_name, IFNAMSIZ);
1430         if (ioctl(s, SIOCGIFFLAGS, &ifr) < 0) {
1431                 ret = -errno;
1432                 goto out;
1433         }
1434         ifr.ifr_flags &= mask;
1435         ifr.ifr_flags |= flags;
1436         if (ioctl(s, SIOCSIFFLAGS, &ifr) < 0) {
1437                 ret = -errno;
1438                 goto out;
1439         }
1440 out:
1441         close(s);
1442         return ret;
1443 }
1444
1445 static int
1446 eth_dev_promiscuous_enable(struct rte_eth_dev *dev)
1447 {
1448         struct pmd_internals *internals = dev->data->dev_private;
1449
1450         return eth_dev_change_flags(internals->if_name, IFF_PROMISC, ~0);
1451 }
1452
1453 static int
1454 eth_dev_promiscuous_disable(struct rte_eth_dev *dev)
1455 {
1456         struct pmd_internals *internals = dev->data->dev_private;
1457
1458         return eth_dev_change_flags(internals->if_name, 0, ~IFF_PROMISC);
1459 }
1460
1461 static const struct eth_dev_ops ops = {
1462         .dev_start = eth_dev_start,
1463         .dev_stop = eth_dev_stop,
1464         .dev_close = eth_dev_close,
1465         .dev_configure = eth_dev_configure,
1466         .dev_infos_get = eth_dev_info,
1467         .mtu_set = eth_dev_mtu_set,
1468         .promiscuous_enable = eth_dev_promiscuous_enable,
1469         .promiscuous_disable = eth_dev_promiscuous_disable,
1470         .rx_queue_setup = eth_rx_queue_setup,
1471         .tx_queue_setup = eth_tx_queue_setup,
1472         .link_update = eth_link_update,
1473         .stats_get = eth_stats_get,
1474         .stats_reset = eth_stats_reset,
1475         .get_monitor_addr = eth_get_monitor_addr,
1476 };
1477
1478 /** parse busy_budget argument */
1479 static int
1480 parse_budget_arg(const char *key __rte_unused,
1481                   const char *value, void *extra_args)
1482 {
1483         int *i = (int *)extra_args;
1484         char *end;
1485
1486         *i = strtol(value, &end, 10);
1487         if (*i < 0 || *i > UINT16_MAX) {
1488                 AF_XDP_LOG(ERR, "Invalid busy_budget %i, must be >= 0 and <= %u\n",
1489                                 *i, UINT16_MAX);
1490                 return -EINVAL;
1491         }
1492
1493         return 0;
1494 }
1495
1496 /** parse integer from integer argument */
1497 static int
1498 parse_integer_arg(const char *key __rte_unused,
1499                   const char *value, void *extra_args)
1500 {
1501         int *i = (int *)extra_args;
1502         char *end;
1503
1504         *i = strtol(value, &end, 10);
1505         if (*i < 0) {
1506                 AF_XDP_LOG(ERR, "Argument has to be positive.\n");
1507                 return -EINVAL;
1508         }
1509
1510         return 0;
1511 }
1512
1513 /** parse name argument */
1514 static int
1515 parse_name_arg(const char *key __rte_unused,
1516                const char *value, void *extra_args)
1517 {
1518         char *name = extra_args;
1519
1520         if (strnlen(value, IFNAMSIZ) > IFNAMSIZ - 1) {
1521                 AF_XDP_LOG(ERR, "Invalid name %s, should be less than %u bytes.\n",
1522                            value, IFNAMSIZ);
1523                 return -EINVAL;
1524         }
1525
1526         strlcpy(name, value, IFNAMSIZ);
1527
1528         return 0;
1529 }
1530
1531 /** parse xdp prog argument */
1532 static int
1533 parse_prog_arg(const char *key __rte_unused,
1534                const char *value, void *extra_args)
1535 {
1536         char *path = extra_args;
1537
1538         if (strnlen(value, PATH_MAX) == PATH_MAX) {
1539                 AF_XDP_LOG(ERR, "Invalid path %s, should be less than %u bytes.\n",
1540                            value, PATH_MAX);
1541                 return -EINVAL;
1542         }
1543
1544         if (access(value, F_OK) != 0) {
1545                 AF_XDP_LOG(ERR, "Error accessing %s: %s\n",
1546                            value, strerror(errno));
1547                 return -EINVAL;
1548         }
1549
1550         strlcpy(path, value, PATH_MAX);
1551
1552         return 0;
1553 }
1554
1555 static int
1556 xdp_get_channels_info(const char *if_name, int *max_queues,
1557                                 int *combined_queues)
1558 {
1559         struct ethtool_channels channels;
1560         struct ifreq ifr;
1561         int fd, ret;
1562
1563         fd = socket(AF_INET, SOCK_DGRAM, 0);
1564         if (fd < 0)
1565                 return -1;
1566
1567         channels.cmd = ETHTOOL_GCHANNELS;
1568         ifr.ifr_data = (void *)&channels;
1569         strlcpy(ifr.ifr_name, if_name, IFNAMSIZ);
1570         ret = ioctl(fd, SIOCETHTOOL, &ifr);
1571         if (ret) {
1572                 if (errno == EOPNOTSUPP) {
1573                         ret = 0;
1574                 } else {
1575                         ret = -errno;
1576                         goto out;
1577                 }
1578         }
1579
1580         if (channels.max_combined == 0 || errno == EOPNOTSUPP) {
1581                 /* If the device says it has no channels, then all traffic
1582                  * is sent to a single stream, so max queues = 1.
1583                  */
1584                 *max_queues = 1;
1585                 *combined_queues = 1;
1586         } else {
1587                 *max_queues = channels.max_combined;
1588                 *combined_queues = channels.combined_count;
1589         }
1590
1591  out:
1592         close(fd);
1593         return ret;
1594 }
1595
1596 static int
1597 parse_parameters(struct rte_kvargs *kvlist, char *if_name, int *start_queue,
1598                         int *queue_cnt, int *shared_umem, char *prog_path,
1599                         int *busy_budget)
1600 {
1601         int ret;
1602
1603         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_IFACE_ARG,
1604                                  &parse_name_arg, if_name);
1605         if (ret < 0)
1606                 goto free_kvlist;
1607
1608         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_START_QUEUE_ARG,
1609                                  &parse_integer_arg, start_queue);
1610         if (ret < 0)
1611                 goto free_kvlist;
1612
1613         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_QUEUE_COUNT_ARG,
1614                                  &parse_integer_arg, queue_cnt);
1615         if (ret < 0 || *queue_cnt <= 0) {
1616                 ret = -EINVAL;
1617                 goto free_kvlist;
1618         }
1619
1620         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_SHARED_UMEM_ARG,
1621                                 &parse_integer_arg, shared_umem);
1622         if (ret < 0)
1623                 goto free_kvlist;
1624
1625         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_PROG_ARG,
1626                                  &parse_prog_arg, prog_path);
1627         if (ret < 0)
1628                 goto free_kvlist;
1629
1630         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_BUDGET_ARG,
1631                                 &parse_budget_arg, busy_budget);
1632         if (ret < 0)
1633                 goto free_kvlist;
1634
1635 free_kvlist:
1636         rte_kvargs_free(kvlist);
1637         return ret;
1638 }
1639
1640 static int
1641 get_iface_info(const char *if_name,
1642                struct rte_ether_addr *eth_addr,
1643                int *if_index)
1644 {
1645         struct ifreq ifr;
1646         int sock = socket(AF_INET, SOCK_DGRAM, IPPROTO_IP);
1647
1648         if (sock < 0)
1649                 return -1;
1650
1651         strlcpy(ifr.ifr_name, if_name, IFNAMSIZ);
1652         if (ioctl(sock, SIOCGIFINDEX, &ifr))
1653                 goto error;
1654
1655         *if_index = ifr.ifr_ifindex;
1656
1657         if (ioctl(sock, SIOCGIFHWADDR, &ifr))
1658                 goto error;
1659
1660         rte_memcpy(eth_addr, ifr.ifr_hwaddr.sa_data, RTE_ETHER_ADDR_LEN);
1661
1662         close(sock);
1663         return 0;
1664
1665 error:
1666         close(sock);
1667         return -1;
1668 }
1669
1670 static struct rte_eth_dev *
1671 init_internals(struct rte_vdev_device *dev, const char *if_name,
1672                 int start_queue_idx, int queue_cnt, int shared_umem,
1673                 const char *prog_path, int busy_budget)
1674 {
1675         const char *name = rte_vdev_device_name(dev);
1676         const unsigned int numa_node = dev->device.numa_node;
1677         struct pmd_internals *internals;
1678         struct rte_eth_dev *eth_dev;
1679         int ret;
1680         int i;
1681
1682         internals = rte_zmalloc_socket(name, sizeof(*internals), 0, numa_node);
1683         if (internals == NULL)
1684                 return NULL;
1685
1686         internals->start_queue_idx = start_queue_idx;
1687         internals->queue_cnt = queue_cnt;
1688         strlcpy(internals->if_name, if_name, IFNAMSIZ);
1689         strlcpy(internals->prog_path, prog_path, PATH_MAX);
1690         internals->custom_prog_configured = 0;
1691
1692 #ifndef ETH_AF_XDP_SHARED_UMEM
1693         if (shared_umem) {
1694                 AF_XDP_LOG(ERR, "Shared UMEM feature not available. "
1695                                 "Check kernel and libbpf version\n");
1696                 goto err_free_internals;
1697         }
1698 #endif
1699         internals->shared_umem = shared_umem;
1700
1701         if (xdp_get_channels_info(if_name, &internals->max_queue_cnt,
1702                                   &internals->combined_queue_cnt)) {
1703                 AF_XDP_LOG(ERR, "Failed to get channel info of interface: %s\n",
1704                                 if_name);
1705                 goto err_free_internals;
1706         }
1707
1708         if (queue_cnt > internals->combined_queue_cnt) {
1709                 AF_XDP_LOG(ERR, "Specified queue count %d is larger than combined queue count %d.\n",
1710                                 queue_cnt, internals->combined_queue_cnt);
1711                 goto err_free_internals;
1712         }
1713
1714         internals->rx_queues = rte_zmalloc_socket(NULL,
1715                                         sizeof(struct pkt_rx_queue) * queue_cnt,
1716                                         0, numa_node);
1717         if (internals->rx_queues == NULL) {
1718                 AF_XDP_LOG(ERR, "Failed to allocate memory for rx queues.\n");
1719                 goto err_free_internals;
1720         }
1721
1722         internals->tx_queues = rte_zmalloc_socket(NULL,
1723                                         sizeof(struct pkt_tx_queue) * queue_cnt,
1724                                         0, numa_node);
1725         if (internals->tx_queues == NULL) {
1726                 AF_XDP_LOG(ERR, "Failed to allocate memory for tx queues.\n");
1727                 goto err_free_rx;
1728         }
1729         for (i = 0; i < queue_cnt; i++) {
1730                 internals->tx_queues[i].pair = &internals->rx_queues[i];
1731                 internals->rx_queues[i].pair = &internals->tx_queues[i];
1732                 internals->rx_queues[i].xsk_queue_idx = start_queue_idx + i;
1733                 internals->tx_queues[i].xsk_queue_idx = start_queue_idx + i;
1734                 internals->rx_queues[i].busy_budget = busy_budget;
1735         }
1736
1737         ret = get_iface_info(if_name, &internals->eth_addr,
1738                              &internals->if_index);
1739         if (ret)
1740                 goto err_free_tx;
1741
1742         eth_dev = rte_eth_vdev_allocate(dev, 0);
1743         if (eth_dev == NULL)
1744                 goto err_free_tx;
1745
1746         eth_dev->data->dev_private = internals;
1747         eth_dev->data->dev_link = pmd_link;
1748         eth_dev->data->mac_addrs = &internals->eth_addr;
1749         eth_dev->data->dev_flags |= RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS;
1750         eth_dev->dev_ops = &ops;
1751         eth_dev->rx_pkt_burst = eth_af_xdp_rx;
1752         eth_dev->tx_pkt_burst = eth_af_xdp_tx;
1753
1754 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
1755         AF_XDP_LOG(INFO, "Zero copy between umem and mbuf enabled.\n");
1756 #endif
1757
1758         return eth_dev;
1759
1760 err_free_tx:
1761         rte_free(internals->tx_queues);
1762 err_free_rx:
1763         rte_free(internals->rx_queues);
1764 err_free_internals:
1765         rte_free(internals);
1766         return NULL;
1767 }
1768
1769 static int
1770 rte_pmd_af_xdp_probe(struct rte_vdev_device *dev)
1771 {
1772         struct rte_kvargs *kvlist;
1773         char if_name[IFNAMSIZ] = {'\0'};
1774         int xsk_start_queue_idx = ETH_AF_XDP_DFLT_START_QUEUE_IDX;
1775         int xsk_queue_cnt = ETH_AF_XDP_DFLT_QUEUE_COUNT;
1776         int shared_umem = 0;
1777         char prog_path[PATH_MAX] = {'\0'};
1778         int busy_budget = -1;
1779         struct rte_eth_dev *eth_dev = NULL;
1780         const char *name;
1781
1782         AF_XDP_LOG(INFO, "Initializing pmd_af_xdp for %s\n",
1783                 rte_vdev_device_name(dev));
1784
1785         name = rte_vdev_device_name(dev);
1786         if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1787                 AF_XDP_LOG(ERR, "Failed to probe %s. "
1788                                 "AF_XDP PMD does not support secondary processes.\n",
1789                                 name);
1790                 return -ENOTSUP;
1791         }
1792
1793         kvlist = rte_kvargs_parse(rte_vdev_device_args(dev), valid_arguments);
1794         if (kvlist == NULL) {
1795                 AF_XDP_LOG(ERR, "Invalid kvargs key\n");
1796                 return -EINVAL;
1797         }
1798
1799         if (dev->device.numa_node == SOCKET_ID_ANY)
1800                 dev->device.numa_node = rte_socket_id();
1801
1802         if (parse_parameters(kvlist, if_name, &xsk_start_queue_idx,
1803                              &xsk_queue_cnt, &shared_umem, prog_path,
1804                              &busy_budget) < 0) {
1805                 AF_XDP_LOG(ERR, "Invalid kvargs value\n");
1806                 return -EINVAL;
1807         }
1808
1809         if (strlen(if_name) == 0) {
1810                 AF_XDP_LOG(ERR, "Network interface must be specified\n");
1811                 return -EINVAL;
1812         }
1813
1814         busy_budget = busy_budget == -1 ? ETH_AF_XDP_DFLT_BUSY_BUDGET :
1815                                         busy_budget;
1816
1817         eth_dev = init_internals(dev, if_name, xsk_start_queue_idx,
1818                                         xsk_queue_cnt, shared_umem, prog_path,
1819                                         busy_budget);
1820         if (eth_dev == NULL) {
1821                 AF_XDP_LOG(ERR, "Failed to init internals\n");
1822                 return -1;
1823         }
1824
1825         rte_eth_dev_probing_finish(eth_dev);
1826
1827         return 0;
1828 }
1829
1830 static int
1831 rte_pmd_af_xdp_remove(struct rte_vdev_device *dev)
1832 {
1833         struct rte_eth_dev *eth_dev = NULL;
1834
1835         AF_XDP_LOG(INFO, "Removing AF_XDP ethdev on numa socket %u\n",
1836                 rte_socket_id());
1837
1838         if (dev == NULL)
1839                 return -1;
1840
1841         /* find the ethdev entry */
1842         eth_dev = rte_eth_dev_allocated(rte_vdev_device_name(dev));
1843         if (eth_dev == NULL)
1844                 return 0;
1845
1846         eth_dev_close(eth_dev);
1847         rte_eth_dev_release_port(eth_dev);
1848
1849
1850         return 0;
1851 }
1852
1853 static struct rte_vdev_driver pmd_af_xdp_drv = {
1854         .probe = rte_pmd_af_xdp_probe,
1855         .remove = rte_pmd_af_xdp_remove,
1856 };
1857
1858 RTE_PMD_REGISTER_VDEV(net_af_xdp, pmd_af_xdp_drv);
1859 RTE_PMD_REGISTER_PARAM_STRING(net_af_xdp,
1860                               "iface=<string> "
1861                               "start_queue=<int> "
1862                               "queue_count=<int> "
1863                               "shared_umem=<int> "
1864                               "xdp_prog=<string> "
1865                               "busy_budget=<int>");