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