drivers/net/af_xdp/rte_eth_af_xdp.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2019-2020 Intel Corporation.
   3  */
   4 #include <unistd.h>
   5 #include <errno.h>
   6 #include <stdlib.h>
   7 #include <string.h>
   8 #include <netinet/in.h>
   9 #include <net/if.h>
  10 #include <sys/socket.h>
  11 #include <sys/ioctl.h>
  12 #include <linux/if_ether.h>
  13 #include <linux/if_xdp.h>
  14 #include <linux/if_link.h>
  15 #include <linux/ethtool.h>
  16 #include <linux/sockios.h>
  17 #include "af_xdp_deps.h"
  18 #include <bpf/bpf.h>
  19
  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         struct bpf_map *map;
 143
 144         struct rte_ether_addr eth_addr;
 145
 146         struct pkt_rx_queue *rx_queues;
 147         struct pkt_tx_queue *tx_queues;
 148 };
 149
 150 #define ETH_AF_XDP_IFACE_ARG                    "iface"
 151 #define ETH_AF_XDP_START_QUEUE_ARG              "start_queue"
 152 #define ETH_AF_XDP_QUEUE_COUNT_ARG              "queue_count"
 153 #define ETH_AF_XDP_SHARED_UMEM_ARG              "shared_umem"
 154 #define ETH_AF_XDP_PROG_ARG                     "xdp_prog"
 155 #define ETH_AF_XDP_BUDGET_ARG                   "busy_budget"
 156
 157 static const char * const valid_arguments[] = {
 158         ETH_AF_XDP_IFACE_ARG,
 159         ETH_AF_XDP_START_QUEUE_ARG,
 160         ETH_AF_XDP_QUEUE_COUNT_ARG,
 161         ETH_AF_XDP_SHARED_UMEM_ARG,
 162         ETH_AF_XDP_PROG_ARG,
 163         ETH_AF_XDP_BUDGET_ARG,
 164         NULL
 165 };
 166
 167 static const struct rte_eth_link pmd_link = {
 168         .link_speed = RTE_ETH_SPEED_NUM_10G,
 169         .link_duplex = RTE_ETH_LINK_FULL_DUPLEX,
 170         .link_status = RTE_ETH_LINK_DOWN,
 171         .link_autoneg = RTE_ETH_LINK_AUTONEG
 172 };
 173
 174 /* List which tracks PMDs to facilitate sharing UMEMs across them. */
 175 struct internal_list {
 176         TAILQ_ENTRY(internal_list) next;
 177         struct rte_eth_dev *eth_dev;
 178 };
 179
 180 TAILQ_HEAD(internal_list_head, internal_list);
 181 static struct internal_list_head internal_list =
 182         TAILQ_HEAD_INITIALIZER(internal_list);
 183
 184 static pthread_mutex_t internal_list_lock = PTHREAD_MUTEX_INITIALIZER;
 185
 186 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 187 static inline int
 188 reserve_fill_queue_zc(struct xsk_umem_info *umem, uint16_t reserve_size,
 189                       struct rte_mbuf **bufs, struct xsk_ring_prod *fq)
 190 {
 191         uint32_t idx;
 192         uint16_t i;
 193
 194         if (unlikely(!xsk_ring_prod__reserve(fq, reserve_size, &idx))) {
 195                 for (i = 0; i < reserve_size; i++)
 196                         rte_pktmbuf_free(bufs[i]);
 197                 AF_XDP_LOG(DEBUG, "Failed to reserve enough fq descs.\n");
 198                 return -1;
 199         }
 200
 201         for (i = 0; i < reserve_size; i++) {
 202                 __u64 *fq_addr;
 203                 uint64_t addr;
 204
 205                 fq_addr = xsk_ring_prod__fill_addr(fq, idx++);
 206                 addr = (uint64_t)bufs[i] - (uint64_t)umem->buffer -
 207                                 umem->mb_pool->header_size;
 208                 *fq_addr = addr;
 209         }
 210
 211         xsk_ring_prod__submit(fq, reserve_size);
 212
 213         return 0;
 214 }
 215 #else
 216 static inline int
 217 reserve_fill_queue_cp(struct xsk_umem_info *umem, uint16_t reserve_size,
 218                       struct rte_mbuf **bufs __rte_unused,
 219                       struct xsk_ring_prod *fq)
 220 {
 221         void *addrs[reserve_size];
 222         uint32_t idx;
 223         uint16_t i;
 224
 225         if (rte_ring_dequeue_bulk(umem->buf_ring, addrs, reserve_size, NULL)
 226                     != reserve_size) {
 227                 AF_XDP_LOG(DEBUG, "Failed to get enough buffers for fq.\n");
 228                 return -1;
 229         }
 230
 231         if (unlikely(!xsk_ring_prod__reserve(fq, reserve_size, &idx))) {
 232                 AF_XDP_LOG(DEBUG, "Failed to reserve enough fq descs.\n");
 233                 rte_ring_enqueue_bulk(umem->buf_ring, addrs,
 234                                 reserve_size, NULL);
 235                 return -1;
 236         }
 237
 238         for (i = 0; i < reserve_size; i++) {
 239                 __u64 *fq_addr;
 240
 241                 fq_addr = xsk_ring_prod__fill_addr(fq, idx++);
 242                 *fq_addr = (uint64_t)addrs[i];
 243         }
 244
 245         xsk_ring_prod__submit(fq, reserve_size);
 246
 247         return 0;
 248 }
 249 #endif
 250
 251 static inline int
 252 reserve_fill_queue(struct xsk_umem_info *umem, uint16_t reserve_size,
 253                    struct rte_mbuf **bufs, struct xsk_ring_prod *fq)
 254 {
 255 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 256         return reserve_fill_queue_zc(umem, reserve_size, bufs, fq);
 257 #else
 258         return reserve_fill_queue_cp(umem, reserve_size, bufs, fq);
 259 #endif
 260 }
 261
 262 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 263 static uint16_t
 264 af_xdp_rx_zc(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 265 {
 266         struct pkt_rx_queue *rxq = queue;
 267         struct xsk_ring_cons *rx = &rxq->rx;
 268         struct xsk_ring_prod *fq = &rxq->fq;
 269         struct xsk_umem_info *umem = rxq->umem;
 270         uint32_t idx_rx = 0;
 271         unsigned long rx_bytes = 0;
 272         int i;
 273         struct rte_mbuf *fq_bufs[ETH_AF_XDP_RX_BATCH_SIZE];
 274
 275         nb_pkts = xsk_ring_cons__peek(rx, nb_pkts, &idx_rx);
 276
 277         if (nb_pkts == 0) {
 278                 /* we can assume a kernel >= 5.11 is in use if busy polling is
 279                  * enabled and thus we can safely use the recvfrom() syscall
 280                  * which is only supported for AF_XDP sockets in kernels >=
 281                  * 5.11.
 282                  */
 283                 if (rxq->busy_budget) {
 284                         (void)recvfrom(xsk_socket__fd(rxq->xsk), NULL, 0,
 285                                        MSG_DONTWAIT, NULL, NULL);
 286                 } else if (xsk_ring_prod__needs_wakeup(fq)) {
 287                         (void)poll(&rxq->fds[0], 1, 1000);
 288                 }
 289
 290                 return 0;
 291         }
 292
 293         /* allocate bufs for fill queue replenishment after rx */
 294         if (rte_pktmbuf_alloc_bulk(umem->mb_pool, fq_bufs, nb_pkts)) {
 295                 AF_XDP_LOG(DEBUG,
 296                         "Failed to get enough buffers for fq.\n");
 297                 /* rollback cached_cons which is added by
 298                  * xsk_ring_cons__peek
 299                  */
 300                 rx->cached_cons -= nb_pkts;
 301                 return 0;
 302         }
 303
 304         for (i = 0; i < nb_pkts; i++) {
 305                 const struct xdp_desc *desc;
 306                 uint64_t addr;
 307                 uint32_t len;
 308                 uint64_t offset;
 309
 310                 desc = xsk_ring_cons__rx_desc(rx, idx_rx++);
 311                 addr = desc->addr;
 312                 len = desc->len;
 313
 314                 offset = xsk_umem__extract_offset(addr);
 315                 addr = xsk_umem__extract_addr(addr);
 316
 317                 bufs[i] = (struct rte_mbuf *)
 318                                 xsk_umem__get_data(umem->buffer, addr +
 319                                         umem->mb_pool->header_size);
 320                 bufs[i]->data_off = offset - sizeof(struct rte_mbuf) -
 321                         rte_pktmbuf_priv_size(umem->mb_pool) -
 322                         umem->mb_pool->header_size;
 323
 324                 rte_pktmbuf_pkt_len(bufs[i]) = len;
 325                 rte_pktmbuf_data_len(bufs[i]) = len;
 326                 rx_bytes += len;
 327         }
 328
 329         xsk_ring_cons__release(rx, nb_pkts);
 330         (void)reserve_fill_queue(umem, nb_pkts, fq_bufs, fq);
 331
 332         /* statistics */
 333         rxq->stats.rx_pkts += nb_pkts;
 334         rxq->stats.rx_bytes += rx_bytes;
 335
 336         return nb_pkts;
 337 }
 338 #else
 339 static uint16_t
 340 af_xdp_rx_cp(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 341 {
 342         struct pkt_rx_queue *rxq = queue;
 343         struct xsk_ring_cons *rx = &rxq->rx;
 344         struct xsk_umem_info *umem = rxq->umem;
 345         struct xsk_ring_prod *fq = &rxq->fq;
 346         uint32_t idx_rx = 0;
 347         unsigned long rx_bytes = 0;
 348         int i;
 349         uint32_t free_thresh = fq->size >> 1;
 350         struct rte_mbuf *mbufs[ETH_AF_XDP_RX_BATCH_SIZE];
 351
 352         if (xsk_prod_nb_free(fq, free_thresh) >= free_thresh)
 353                 (void)reserve_fill_queue(umem, nb_pkts, NULL, fq);
 354
 355         nb_pkts = xsk_ring_cons__peek(rx, nb_pkts, &idx_rx);
 356         if (nb_pkts == 0) {
 357 #if defined(XDP_USE_NEED_WAKEUP)
 358                 if (xsk_ring_prod__needs_wakeup(fq))
 359                         (void)poll(rxq->fds, 1, 1000);
 360 #endif
 361                 return 0;
 362         }
 363
 364         if (unlikely(rte_pktmbuf_alloc_bulk(rxq->mb_pool, mbufs, nb_pkts))) {
 365                 /* rollback cached_cons which is added by
 366                  * xsk_ring_cons__peek
 367                  */
 368                 rx->cached_cons -= nb_pkts;
 369                 return 0;
 370         }
 371
 372         for (i = 0; i < nb_pkts; i++) {
 373                 const struct xdp_desc *desc;
 374                 uint64_t addr;
 375                 uint32_t len;
 376                 void *pkt;
 377
 378                 desc = xsk_ring_cons__rx_desc(rx, idx_rx++);
 379                 addr = desc->addr;
 380                 len = desc->len;
 381                 pkt = xsk_umem__get_data(rxq->umem->mz->addr, addr);
 382
 383                 rte_memcpy(rte_pktmbuf_mtod(mbufs[i], void *), pkt, len);
 384                 rte_ring_enqueue(umem->buf_ring, (void *)addr);
 385                 rte_pktmbuf_pkt_len(mbufs[i]) = len;
 386                 rte_pktmbuf_data_len(mbufs[i]) = len;
 387                 rx_bytes += len;
 388                 bufs[i] = mbufs[i];
 389         }
 390
 391         xsk_ring_cons__release(rx, nb_pkts);
 392
 393         /* statistics */
 394         rxq->stats.rx_pkts += nb_pkts;
 395         rxq->stats.rx_bytes += rx_bytes;
 396
 397         return nb_pkts;
 398 }
 399 #endif
 400
 401 static uint16_t
 402 af_xdp_rx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 403 {
 404 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 405         return af_xdp_rx_zc(queue, bufs, nb_pkts);
 406 #else
 407         return af_xdp_rx_cp(queue, bufs, nb_pkts);
 408 #endif
 409 }
 410
 411 static uint16_t
 412 eth_af_xdp_rx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 413 {
 414         uint16_t nb_rx;
 415
 416         if (likely(nb_pkts <= ETH_AF_XDP_RX_BATCH_SIZE))
 417                 return af_xdp_rx(queue, bufs, nb_pkts);
 418
 419         /* Split larger batch into smaller batches of size
 420          * ETH_AF_XDP_RX_BATCH_SIZE or less.
 421          */
 422         nb_rx = 0;
 423         while (nb_pkts) {
 424                 uint16_t ret, n;
 425
 426                 n = (uint16_t)RTE_MIN(nb_pkts, ETH_AF_XDP_RX_BATCH_SIZE);
 427                 ret = af_xdp_rx(queue, &bufs[nb_rx], n);
 428                 nb_rx = (uint16_t)(nb_rx + ret);
 429                 nb_pkts = (uint16_t)(nb_pkts - ret);
 430                 if (ret < n)
 431                         break;
 432         }
 433
 434         return nb_rx;
 435 }
 436
 437 static void
 438 pull_umem_cq(struct xsk_umem_info *umem, int size, struct xsk_ring_cons *cq)
 439 {
 440         size_t i, n;
 441         uint32_t idx_cq = 0;
 442
 443         n = xsk_ring_cons__peek(cq, size, &idx_cq);
 444
 445         for (i = 0; i < n; i++) {
 446                 uint64_t addr;
 447                 addr = *xsk_ring_cons__comp_addr(cq, idx_cq++);
 448 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 449                 addr = xsk_umem__extract_addr(addr);
 450                 rte_pktmbuf_free((struct rte_mbuf *)
 451                                         xsk_umem__get_data(umem->buffer,
 452                                         addr + umem->mb_pool->header_size));
 453 #else
 454                 rte_ring_enqueue(umem->buf_ring, (void *)addr);
 455 #endif
 456         }
 457
 458         xsk_ring_cons__release(cq, n);
 459 }
 460
 461 static void
 462 kick_tx(struct pkt_tx_queue *txq, struct xsk_ring_cons *cq)
 463 {
 464         struct xsk_umem_info *umem = txq->umem;
 465
 466         pull_umem_cq(umem, XSK_RING_CONS__DEFAULT_NUM_DESCS, cq);
 467
 468         if (tx_syscall_needed(&txq->tx))
 469                 while (send(xsk_socket__fd(txq->pair->xsk), NULL,
 470                             0, MSG_DONTWAIT) < 0) {
 471                         /* some thing unexpected */
 472                         if (errno != EBUSY && errno != EAGAIN && errno != EINTR)
 473                                 break;
 474
 475                         /* pull from completion queue to leave more space */
 476                         if (errno == EAGAIN)
 477                                 pull_umem_cq(umem,
 478                                              XSK_RING_CONS__DEFAULT_NUM_DESCS,
 479                                              cq);
 480                 }
 481 }
 482
 483 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 484 static uint16_t
 485 af_xdp_tx_zc(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 486 {
 487         struct pkt_tx_queue *txq = queue;
 488         struct xsk_umem_info *umem = txq->umem;
 489         struct rte_mbuf *mbuf;
 490         unsigned long tx_bytes = 0;
 491         int i;
 492         uint32_t idx_tx;
 493         uint16_t count = 0;
 494         struct xdp_desc *desc;
 495         uint64_t addr, offset;
 496         struct xsk_ring_cons *cq = &txq->pair->cq;
 497         uint32_t free_thresh = cq->size >> 1;
 498
 499         if (xsk_cons_nb_avail(cq, free_thresh) >= free_thresh)
 500                 pull_umem_cq(umem, XSK_RING_CONS__DEFAULT_NUM_DESCS, cq);
 501
 502         for (i = 0; i < nb_pkts; i++) {
 503                 mbuf = bufs[i];
 504
 505                 if (mbuf->pool == umem->mb_pool) {
 506                         if (!xsk_ring_prod__reserve(&txq->tx, 1, &idx_tx)) {
 507                                 kick_tx(txq, cq);
 508                                 if (!xsk_ring_prod__reserve(&txq->tx, 1,
 509                                                             &idx_tx))
 510                                         goto out;
 511                         }
 512                         desc = xsk_ring_prod__tx_desc(&txq->tx, idx_tx);
 513                         desc->len = mbuf->pkt_len;
 514                         addr = (uint64_t)mbuf - (uint64_t)umem->buffer -
 515                                         umem->mb_pool->header_size;
 516                         offset = rte_pktmbuf_mtod(mbuf, uint64_t) -
 517                                         (uint64_t)mbuf +
 518                                         umem->mb_pool->header_size;
 519                         offset = offset << XSK_UNALIGNED_BUF_OFFSET_SHIFT;
 520                         desc->addr = addr | offset;
 521                         count++;
 522                 } else {
 523                         struct rte_mbuf *local_mbuf =
 524                                         rte_pktmbuf_alloc(umem->mb_pool);
 525                         void *pkt;
 526
 527                         if (local_mbuf == NULL)
 528                                 goto out;
 529
 530                         if (!xsk_ring_prod__reserve(&txq->tx, 1, &idx_tx)) {
 531                                 rte_pktmbuf_free(local_mbuf);
 532                                 goto out;
 533                         }
 534
 535                         desc = xsk_ring_prod__tx_desc(&txq->tx, idx_tx);
 536                         desc->len = mbuf->pkt_len;
 537
 538                         addr = (uint64_t)local_mbuf - (uint64_t)umem->buffer -
 539                                         umem->mb_pool->header_size;
 540                         offset = rte_pktmbuf_mtod(local_mbuf, uint64_t) -
 541                                         (uint64_t)local_mbuf +
 542                                         umem->mb_pool->header_size;
 543                         pkt = xsk_umem__get_data(umem->buffer, addr + offset);
 544                         offset = offset << XSK_UNALIGNED_BUF_OFFSET_SHIFT;
 545                         desc->addr = addr | offset;
 546                         rte_memcpy(pkt, rte_pktmbuf_mtod(mbuf, void *),
 547                                         desc->len);
 548                         rte_pktmbuf_free(mbuf);
 549                         count++;
 550                 }
 551
 552                 tx_bytes += mbuf->pkt_len;
 553         }
 554
 555 out:
 556         xsk_ring_prod__submit(&txq->tx, count);
 557         kick_tx(txq, cq);
 558
 559         txq->stats.tx_pkts += count;
 560         txq->stats.tx_bytes += tx_bytes;
 561         txq->stats.tx_dropped += nb_pkts - count;
 562
 563         return count;
 564 }
 565 #else
 566 static uint16_t
 567 af_xdp_tx_cp(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 568 {
 569         struct pkt_tx_queue *txq = queue;
 570         struct xsk_umem_info *umem = txq->umem;
 571         struct rte_mbuf *mbuf;
 572         void *addrs[ETH_AF_XDP_TX_BATCH_SIZE];
 573         unsigned long tx_bytes = 0;
 574         int i;
 575         uint32_t idx_tx;
 576         struct xsk_ring_cons *cq = &txq->pair->cq;
 577
 578         pull_umem_cq(umem, nb_pkts, cq);
 579
 580         nb_pkts = rte_ring_dequeue_bulk(umem->buf_ring, addrs,
 581                                         nb_pkts, NULL);
 582         if (nb_pkts == 0)
 583                 return 0;
 584
 585         if (xsk_ring_prod__reserve(&txq->tx, nb_pkts, &idx_tx) != nb_pkts) {
 586                 kick_tx(txq, cq);
 587                 rte_ring_enqueue_bulk(umem->buf_ring, addrs, nb_pkts, NULL);
 588                 return 0;
 589         }
 590
 591         for (i = 0; i < nb_pkts; i++) {
 592                 struct xdp_desc *desc;
 593                 void *pkt;
 594
 595                 desc = xsk_ring_prod__tx_desc(&txq->tx, idx_tx + i);
 596                 mbuf = bufs[i];
 597                 desc->len = mbuf->pkt_len;
 598
 599                 desc->addr = (uint64_t)addrs[i];
 600                 pkt = xsk_umem__get_data(umem->mz->addr,
 601                                          desc->addr);
 602                 rte_memcpy(pkt, rte_pktmbuf_mtod(mbuf, void *), desc->len);
 603                 tx_bytes += mbuf->pkt_len;
 604                 rte_pktmbuf_free(mbuf);
 605         }
 606
 607         xsk_ring_prod__submit(&txq->tx, nb_pkts);
 608
 609         kick_tx(txq, cq);
 610
 611         txq->stats.tx_pkts += nb_pkts;
 612         txq->stats.tx_bytes += tx_bytes;
 613
 614         return nb_pkts;
 615 }
 616
 617 static uint16_t
 618 af_xdp_tx_cp_batch(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 619 {
 620         uint16_t nb_tx;
 621
 622         if (likely(nb_pkts <= ETH_AF_XDP_TX_BATCH_SIZE))
 623                 return af_xdp_tx_cp(queue, bufs, nb_pkts);
 624
 625         nb_tx = 0;
 626         while (nb_pkts) {
 627                 uint16_t ret, n;
 628
 629                 /* Split larger batch into smaller batches of size
 630                  * ETH_AF_XDP_TX_BATCH_SIZE or less.
 631                  */
 632                 n = (uint16_t)RTE_MIN(nb_pkts, ETH_AF_XDP_TX_BATCH_SIZE);
 633                 ret = af_xdp_tx_cp(queue, &bufs[nb_tx], n);
 634                 nb_tx = (uint16_t)(nb_tx + ret);
 635                 nb_pkts = (uint16_t)(nb_pkts - ret);
 636                 if (ret < n)
 637                         break;
 638         }
 639
 640         return nb_tx;
 641 }
 642 #endif
 643
 644 static uint16_t
 645 eth_af_xdp_tx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 646 {
 647 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 648         return af_xdp_tx_zc(queue, bufs, nb_pkts);
 649 #else
 650         return af_xdp_tx_cp_batch(queue, bufs, nb_pkts);
 651 #endif
 652 }
 653
 654 static int
 655 eth_dev_start(struct rte_eth_dev *dev)
 656 {
 657         dev->data->dev_link.link_status = RTE_ETH_LINK_UP;
 658
 659         return 0;
 660 }
 661
 662 /* This function gets called when the current port gets stopped. */
 663 static int
 664 eth_dev_stop(struct rte_eth_dev *dev)
 665 {
 666         dev->data->dev_link.link_status = RTE_ETH_LINK_DOWN;
 667         return 0;
 668 }
 669
 670 /* Find ethdev in list */
 671 static inline struct internal_list *
 672 find_internal_resource(struct pmd_internals *port_int)
 673 {
 674         int found = 0;
 675         struct internal_list *list = NULL;
 676
 677         if (port_int == NULL)
 678                 return NULL;
 679
 680         pthread_mutex_lock(&internal_list_lock);
 681
 682         TAILQ_FOREACH(list, &internal_list, next) {
 683                 struct pmd_internals *list_int =
 684                                 list->eth_dev->data->dev_private;
 685                 if (list_int == port_int) {
 686                         found = 1;
 687                         break;
 688                 }
 689         }
 690
 691         pthread_mutex_unlock(&internal_list_lock);
 692
 693         if (!found)
 694                 return NULL;
 695
 696         return list;
 697 }
 698
 699 static int
 700 eth_dev_configure(struct rte_eth_dev *dev)
 701 {
 702         struct pmd_internals *internal = dev->data->dev_private;
 703
 704         /* rx/tx must be paired */
 705         if (dev->data->nb_rx_queues != dev->data->nb_tx_queues)
 706                 return -EINVAL;
 707
 708         if (internal->shared_umem) {
 709                 struct internal_list *list = NULL;
 710                 const char *name = dev->device->name;
 711
 712                 /* Ensure PMD is not already inserted into the list */
 713                 list = find_internal_resource(internal);
 714                 if (list)
 715                         return 0;
 716
 717                 list = rte_zmalloc_socket(name, sizeof(*list), 0,
 718                                         dev->device->numa_node);
 719                 if (list == NULL)
 720                         return -1;
 721
 722                 list->eth_dev = dev;
 723                 pthread_mutex_lock(&internal_list_lock);
 724                 TAILQ_INSERT_TAIL(&internal_list, list, next);
 725                 pthread_mutex_unlock(&internal_list_lock);
 726         }
 727
 728         return 0;
 729 }
 730
 731 #define CLB_VAL_IDX 0
 732 static int
 733 eth_monitor_callback(const uint64_t value,
 734                 const uint64_t opaque[RTE_POWER_MONITOR_OPAQUE_SZ])
 735 {
 736         const uint64_t v = opaque[CLB_VAL_IDX];
 737         const uint64_t m = (uint32_t)~0;
 738
 739         /* if the value has changed, abort entering power optimized state */
 740         return (value & m) == v ? 0 : -1;
 741 }
 742
 743 static int
 744 eth_get_monitor_addr(void *rx_queue, struct rte_power_monitor_cond *pmc)
 745 {
 746         struct pkt_rx_queue *rxq = rx_queue;
 747         unsigned int *prod = rxq->rx.producer;
 748         const uint32_t cur_val = rxq->rx.cached_prod; /* use cached value */
 749
 750         /* watch for changes in producer ring */
 751         pmc->addr = (void *)prod;
 752
 753         /* store current value */
 754         pmc->opaque[CLB_VAL_IDX] = cur_val;
 755         pmc->fn = eth_monitor_callback;
 756
 757         /* AF_XDP producer ring index is 32-bit */
 758         pmc->size = sizeof(uint32_t);
 759
 760         return 0;
 761 }
 762
 763 static int
 764 eth_dev_info(struct rte_eth_dev *dev, struct rte_eth_dev_info *dev_info)
 765 {
 766         struct pmd_internals *internals = dev->data->dev_private;
 767
 768         dev_info->if_index = internals->if_index;
 769         dev_info->max_mac_addrs = 1;
 770         dev_info->max_rx_queues = internals->queue_cnt;
 771         dev_info->max_tx_queues = internals->queue_cnt;
 772
 773         dev_info->min_mtu = RTE_ETHER_MIN_MTU;
 774 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 775         dev_info->max_rx_pktlen = getpagesize() -
 776                                   sizeof(struct rte_mempool_objhdr) -
 777                                   sizeof(struct rte_mbuf) -
 778                                   RTE_PKTMBUF_HEADROOM - XDP_PACKET_HEADROOM;
 779 #else
 780         dev_info->max_rx_pktlen = ETH_AF_XDP_FRAME_SIZE - XDP_PACKET_HEADROOM;
 781 #endif
 782         dev_info->max_mtu = dev_info->max_rx_pktlen - ETH_AF_XDP_ETH_OVERHEAD;
 783
 784         dev_info->default_rxportconf.burst_size = ETH_AF_XDP_DFLT_BUSY_BUDGET;
 785         dev_info->default_txportconf.burst_size = ETH_AF_XDP_DFLT_BUSY_BUDGET;
 786         dev_info->default_rxportconf.nb_queues = 1;
 787         dev_info->default_txportconf.nb_queues = 1;
 788         dev_info->default_rxportconf.ring_size = ETH_AF_XDP_DFLT_NUM_DESCS;
 789         dev_info->default_txportconf.ring_size = ETH_AF_XDP_DFLT_NUM_DESCS;
 790
 791         return 0;
 792 }
 793
 794 static int
 795 eth_stats_get(struct rte_eth_dev *dev, struct rte_eth_stats *stats)
 796 {
 797         struct pmd_internals *internals = dev->data->dev_private;
 798         struct xdp_statistics xdp_stats;
 799         struct pkt_rx_queue *rxq;
 800         struct pkt_tx_queue *txq;
 801         socklen_t optlen;
 802         int i, ret;
 803
 804         for (i = 0; i < dev->data->nb_rx_queues; i++) {
 805                 optlen = sizeof(struct xdp_statistics);
 806                 rxq = &internals->rx_queues[i];
 807                 txq = rxq->pair;
 808                 stats->q_ipackets[i] = rxq->stats.rx_pkts;
 809                 stats->q_ibytes[i] = rxq->stats.rx_bytes;
 810
 811                 stats->q_opackets[i] = txq->stats.tx_pkts;
 812                 stats->q_obytes[i] = txq->stats.tx_bytes;
 813
 814                 stats->ipackets += stats->q_ipackets[i];
 815                 stats->ibytes += stats->q_ibytes[i];
 816                 stats->imissed += rxq->stats.rx_dropped;
 817                 stats->oerrors += txq->stats.tx_dropped;
 818                 ret = getsockopt(xsk_socket__fd(rxq->xsk), SOL_XDP,
 819                                 XDP_STATISTICS, &xdp_stats, &optlen);
 820                 if (ret != 0) {
 821                         AF_XDP_LOG(ERR, "getsockopt() failed for XDP_STATISTICS.\n");
 822                         return -1;
 823                 }
 824                 stats->imissed += xdp_stats.rx_dropped;
 825
 826                 stats->opackets += stats->q_opackets[i];
 827                 stats->obytes += stats->q_obytes[i];
 828         }
 829
 830         return 0;
 831 }
 832
 833 static int
 834 eth_stats_reset(struct rte_eth_dev *dev)
 835 {
 836         struct pmd_internals *internals = dev->data->dev_private;
 837         int i;
 838
 839         for (i = 0; i < internals->queue_cnt; i++) {
 840                 memset(&internals->rx_queues[i].stats, 0,
 841                                         sizeof(struct rx_stats));
 842                 memset(&internals->tx_queues[i].stats, 0,
 843                                         sizeof(struct tx_stats));
 844         }
 845
 846         return 0;
 847 }
 848
 849 static void
 850 remove_xdp_program(struct pmd_internals *internals)
 851 {
 852         uint32_t curr_prog_id = 0;
 853
 854         if (bpf_get_link_xdp_id(internals->if_index, &curr_prog_id,
 855                                 XDP_FLAGS_UPDATE_IF_NOEXIST)) {
 856                 AF_XDP_LOG(ERR, "bpf_get_link_xdp_id failed\n");
 857                 return;
 858         }
 859         bpf_set_link_xdp_fd(internals->if_index, -1,
 860                         XDP_FLAGS_UPDATE_IF_NOEXIST);
 861 }
 862
 863 static void
 864 xdp_umem_destroy(struct xsk_umem_info *umem)
 865 {
 866 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 867         umem->mb_pool = NULL;
 868 #else
 869         rte_memzone_free(umem->mz);
 870         umem->mz = NULL;
 871
 872         rte_ring_free(umem->buf_ring);
 873         umem->buf_ring = NULL;
 874 #endif
 875
 876         rte_free(umem);
 877 }
 878
 879 static int
 880 eth_dev_close(struct rte_eth_dev *dev)
 881 {
 882         struct pmd_internals *internals = dev->data->dev_private;
 883         struct pkt_rx_queue *rxq;
 884         int i;
 885
 886         if (rte_eal_process_type() != RTE_PROC_PRIMARY)
 887                 return 0;
 888
 889         AF_XDP_LOG(INFO, "Closing AF_XDP ethdev on numa socket %u\n",
 890                 rte_socket_id());
 891
 892         for (i = 0; i < internals->queue_cnt; i++) {
 893                 rxq = &internals->rx_queues[i];
 894                 if (rxq->umem == NULL)
 895                         break;
 896                 xsk_socket__delete(rxq->xsk);
 897
 898                 if (__atomic_sub_fetch(&rxq->umem->refcnt, 1, __ATOMIC_ACQUIRE)
 899                                 == 0) {
 900                         (void)xsk_umem__delete(rxq->umem->umem);
 901                         xdp_umem_destroy(rxq->umem);
 902                 }
 903
 904                 /* free pkt_tx_queue */
 905                 rte_free(rxq->pair);
 906                 rte_free(rxq);
 907         }
 908
 909         /*
 910          * MAC is not allocated dynamically, setting it to NULL would prevent
 911          * from releasing it in rte_eth_dev_release_port.
 912          */
 913         dev->data->mac_addrs = NULL;
 914
 915         remove_xdp_program(internals);
 916
 917         if (internals->shared_umem) {
 918                 struct internal_list *list;
 919
 920                 /* Remove ethdev from list used to track and share UMEMs */
 921                 list = find_internal_resource(internals);
 922                 if (list) {
 923                         pthread_mutex_lock(&internal_list_lock);
 924                         TAILQ_REMOVE(&internal_list, list, next);
 925                         pthread_mutex_unlock(&internal_list_lock);
 926                         rte_free(list);
 927                 }
 928         }
 929
 930         return 0;
 931 }
 932
 933 static int
 934 eth_link_update(struct rte_eth_dev *dev __rte_unused,
 935                 int wait_to_complete __rte_unused)
 936 {
 937         return 0;
 938 }
 939
 940 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 941 static inline uintptr_t get_base_addr(struct rte_mempool *mp, uint64_t *align)
 942 {
 943         struct rte_mempool_memhdr *memhdr;
 944         uintptr_t memhdr_addr, aligned_addr;
 945
 946         memhdr = STAILQ_FIRST(&mp->mem_list);
 947         memhdr_addr = (uintptr_t)memhdr->addr;
 948         aligned_addr = memhdr_addr & ~(getpagesize() - 1);
 949         *align = memhdr_addr - aligned_addr;
 950
 951         return aligned_addr;
 952 }
 953
 954 /* Check if the netdev,qid context already exists */
 955 static inline bool
 956 ctx_exists(struct pkt_rx_queue *rxq, const char *ifname,
 957                 struct pkt_rx_queue *list_rxq, const char *list_ifname)
 958 {
 959         bool exists = false;
 960
 961         if (rxq->xsk_queue_idx == list_rxq->xsk_queue_idx &&
 962                         !strncmp(ifname, list_ifname, IFNAMSIZ)) {
 963                 AF_XDP_LOG(ERR, "ctx %s,%i already exists, cannot share umem\n",
 964                                         ifname, rxq->xsk_queue_idx);
 965                 exists = true;
 966         }
 967
 968         return exists;
 969 }
 970
 971 /* Get a pointer to an existing UMEM which overlays the rxq's mb_pool */
 972 static inline int
 973 get_shared_umem(struct pkt_rx_queue *rxq, const char *ifname,
 974                         struct xsk_umem_info **umem)
 975 {
 976         struct internal_list *list;
 977         struct pmd_internals *internals;
 978         int i = 0, ret = 0;
 979         struct rte_mempool *mb_pool = rxq->mb_pool;
 980
 981         if (mb_pool == NULL)
 982                 return ret;
 983
 984         pthread_mutex_lock(&internal_list_lock);
 985
 986         TAILQ_FOREACH(list, &internal_list, next) {
 987                 internals = list->eth_dev->data->dev_private;
 988                 for (i = 0; i < internals->queue_cnt; i++) {
 989                         struct pkt_rx_queue *list_rxq =
 990                                                 &internals->rx_queues[i];
 991                         if (rxq == list_rxq)
 992                                 continue;
 993                         if (mb_pool == internals->rx_queues[i].mb_pool) {
 994                                 if (ctx_exists(rxq, ifname, list_rxq,
 995                                                 internals->if_name)) {
 996                                         ret = -1;
 997                                         goto out;
 998                                 }
 999                                 if (__atomic_load_n(&internals->rx_queues[i].umem->refcnt,
1000                                                     __ATOMIC_ACQUIRE)) {
1001                                         *umem = internals->rx_queues[i].umem;
1002                                         goto out;
1003                                 }
1004                         }
1005                 }
1006         }
1007
1008 out:
1009         pthread_mutex_unlock(&internal_list_lock);
1010
1011         return ret;
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         return umem;
1082
1083 err:
1084         xdp_umem_destroy(umem);
1085         return NULL;
1086 }
1087 #else
1088 static struct
1089 xsk_umem_info *xdp_umem_configure(struct pmd_internals *internals,
1090                                   struct pkt_rx_queue *rxq)
1091 {
1092         struct xsk_umem_info *umem;
1093         const struct rte_memzone *mz;
1094         struct xsk_umem_config usr_config = {
1095                 .fill_size = ETH_AF_XDP_DFLT_NUM_DESCS,
1096                 .comp_size = ETH_AF_XDP_DFLT_NUM_DESCS,
1097                 .frame_size = ETH_AF_XDP_FRAME_SIZE,
1098                 .frame_headroom = 0 };
1099         char ring_name[RTE_RING_NAMESIZE];
1100         char mz_name[RTE_MEMZONE_NAMESIZE];
1101         int ret;
1102         uint64_t i;
1103
1104         umem = rte_zmalloc_socket("umem", sizeof(*umem), 0, rte_socket_id());
1105         if (umem == NULL) {
1106                 AF_XDP_LOG(ERR, "Failed to allocate umem info");
1107                 return NULL;
1108         }
1109
1110         snprintf(ring_name, sizeof(ring_name), "af_xdp_ring_%s_%u",
1111                        internals->if_name, rxq->xsk_queue_idx);
1112         umem->buf_ring = rte_ring_create(ring_name,
1113                                          ETH_AF_XDP_NUM_BUFFERS,
1114                                          rte_socket_id(),
1115                                          0x0);
1116         if (umem->buf_ring == NULL) {
1117                 AF_XDP_LOG(ERR, "Failed to create rte_ring\n");
1118                 goto err;
1119         }
1120
1121         for (i = 0; i < ETH_AF_XDP_NUM_BUFFERS; i++)
1122                 rte_ring_enqueue(umem->buf_ring,
1123                                  (void *)(i * ETH_AF_XDP_FRAME_SIZE));
1124
1125         snprintf(mz_name, sizeof(mz_name), "af_xdp_umem_%s_%u",
1126                        internals->if_name, rxq->xsk_queue_idx);
1127         mz = rte_memzone_reserve_aligned(mz_name,
1128                         ETH_AF_XDP_NUM_BUFFERS * ETH_AF_XDP_FRAME_SIZE,
1129                         rte_socket_id(), RTE_MEMZONE_IOVA_CONTIG,
1130                         getpagesize());
1131         if (mz == NULL) {
1132                 AF_XDP_LOG(ERR, "Failed to reserve memzone for af_xdp umem.\n");
1133                 goto err;
1134         }
1135
1136         ret = xsk_umem__create(&umem->umem, mz->addr,
1137                                ETH_AF_XDP_NUM_BUFFERS * ETH_AF_XDP_FRAME_SIZE,
1138                                &rxq->fq, &rxq->cq,
1139                                &usr_config);
1140
1141         if (ret) {
1142                 AF_XDP_LOG(ERR, "Failed to create umem");
1143                 goto err;
1144         }
1145         umem->mz = mz;
1146
1147         return umem;
1148
1149 err:
1150         xdp_umem_destroy(umem);
1151         return NULL;
1152 }
1153 #endif
1154
1155 static int
1156 load_custom_xdp_prog(const char *prog_path, int if_index, struct bpf_map **map)
1157 {
1158         int ret, prog_fd = -1;
1159         struct bpf_object *obj;
1160
1161         ret = bpf_prog_load(prog_path, BPF_PROG_TYPE_XDP, &obj, &prog_fd);
1162         if (ret) {
1163                 AF_XDP_LOG(ERR, "Failed to load program %s\n", prog_path);
1164                 return ret;
1165         }
1166
1167         /*
1168          * The loaded program must provision for a map of xsks, such that some
1169          * traffic can be redirected to userspace.
1170          */
1171         *map = bpf_object__find_map_by_name(obj, "xsks_map");
1172         if (!*map) {
1173                 AF_XDP_LOG(ERR, "Failed to find xsks_map in %s\n", prog_path);
1174                 return -1;
1175         }
1176
1177         /* Link the program with the given network device */
1178         ret = bpf_set_link_xdp_fd(if_index, prog_fd,
1179                                         XDP_FLAGS_UPDATE_IF_NOEXIST);
1180         if (ret) {
1181                 AF_XDP_LOG(ERR, "Failed to set prog fd %d on interface\n",
1182                                 prog_fd);
1183                 return -1;
1184         }
1185
1186         AF_XDP_LOG(INFO, "Successfully loaded XDP program %s with fd %d\n",
1187                                 prog_path, prog_fd);
1188
1189         return 0;
1190 }
1191
1192 /* Detect support for busy polling through setsockopt(). */
1193 static int
1194 configure_preferred_busy_poll(struct pkt_rx_queue *rxq)
1195 {
1196         int sock_opt = 1;
1197         int fd = xsk_socket__fd(rxq->xsk);
1198         int ret = 0;
1199
1200         ret = setsockopt(fd, SOL_SOCKET, SO_PREFER_BUSY_POLL,
1201                         (void *)&sock_opt, sizeof(sock_opt));
1202         if (ret < 0) {
1203                 AF_XDP_LOG(DEBUG, "Failed to set SO_PREFER_BUSY_POLL\n");
1204                 goto err_prefer;
1205         }
1206
1207         sock_opt = ETH_AF_XDP_DFLT_BUSY_TIMEOUT;
1208         ret = setsockopt(fd, SOL_SOCKET, SO_BUSY_POLL, (void *)&sock_opt,
1209                         sizeof(sock_opt));
1210         if (ret < 0) {
1211                 AF_XDP_LOG(DEBUG, "Failed to set SO_BUSY_POLL\n");
1212                 goto err_timeout;
1213         }
1214
1215         sock_opt = rxq->busy_budget;
1216         ret = setsockopt(fd, SOL_SOCKET, SO_BUSY_POLL_BUDGET,
1217                         (void *)&sock_opt, sizeof(sock_opt));
1218         if (ret < 0) {
1219                 AF_XDP_LOG(DEBUG, "Failed to set SO_BUSY_POLL_BUDGET\n");
1220         } else {
1221                 AF_XDP_LOG(INFO, "Busy polling budget set to: %u\n",
1222                                         rxq->busy_budget);
1223                 return 0;
1224         }
1225
1226         /* setsockopt failure - attempt to restore xsk to default state and
1227          * proceed without busy polling support.
1228          */
1229         sock_opt = 0;
1230         ret = setsockopt(fd, SOL_SOCKET, SO_BUSY_POLL, (void *)&sock_opt,
1231                         sizeof(sock_opt));
1232         if (ret < 0) {
1233                 AF_XDP_LOG(ERR, "Failed to unset SO_BUSY_POLL\n");
1234                 return -1;
1235         }
1236
1237 err_timeout:
1238         sock_opt = 0;
1239         ret = setsockopt(fd, SOL_SOCKET, SO_PREFER_BUSY_POLL,
1240                         (void *)&sock_opt, sizeof(sock_opt));
1241         if (ret < 0) {
1242                 AF_XDP_LOG(ERR, "Failed to unset SO_PREFER_BUSY_POLL\n");
1243                 return -1;
1244         }
1245
1246 err_prefer:
1247         rxq->busy_budget = 0;
1248         return 0;
1249 }
1250
1251 static int
1252 xsk_configure(struct pmd_internals *internals, struct pkt_rx_queue *rxq,
1253               int ring_size)
1254 {
1255         struct xsk_socket_config cfg;
1256         struct pkt_tx_queue *txq = rxq->pair;
1257         int ret = 0;
1258         int reserve_size = ETH_AF_XDP_DFLT_NUM_DESCS;
1259         struct rte_mbuf *fq_bufs[reserve_size];
1260
1261         rxq->umem = xdp_umem_configure(internals, rxq);
1262         if (rxq->umem == NULL)
1263                 return -ENOMEM;
1264         txq->umem = rxq->umem;
1265
1266         cfg.rx_size = ring_size;
1267         cfg.tx_size = ring_size;
1268         cfg.libbpf_flags = 0;
1269         cfg.xdp_flags = XDP_FLAGS_UPDATE_IF_NOEXIST;
1270         cfg.bind_flags = 0;
1271
1272 #if defined(XDP_USE_NEED_WAKEUP)
1273         cfg.bind_flags |= XDP_USE_NEED_WAKEUP;
1274 #endif
1275
1276         if (strnlen(internals->prog_path, PATH_MAX) &&
1277                                 !internals->custom_prog_configured) {
1278                 ret = load_custom_xdp_prog(internals->prog_path,
1279                                            internals->if_index,
1280                                            &internals->map);
1281                 if (ret) {
1282                         AF_XDP_LOG(ERR, "Failed to load custom XDP program %s\n",
1283                                         internals->prog_path);
1284                         goto err;
1285                 }
1286                 internals->custom_prog_configured = 1;
1287                 cfg.libbpf_flags = XSK_LIBBPF_FLAGS__INHIBIT_PROG_LOAD;
1288         }
1289
1290         if (internals->shared_umem)
1291                 ret = create_shared_socket(&rxq->xsk, internals->if_name,
1292                                 rxq->xsk_queue_idx, rxq->umem->umem, &rxq->rx,
1293                                 &txq->tx, &rxq->fq, &rxq->cq, &cfg);
1294         else
1295                 ret = xsk_socket__create(&rxq->xsk, internals->if_name,
1296                                 rxq->xsk_queue_idx, rxq->umem->umem, &rxq->rx,
1297                                 &txq->tx, &cfg);
1298
1299         if (ret) {
1300                 AF_XDP_LOG(ERR, "Failed to create xsk socket.\n");
1301                 goto err;
1302         }
1303
1304         /* insert the xsk into the xsks_map */
1305         if (internals->custom_prog_configured) {
1306                 int err, fd;
1307
1308                 fd = xsk_socket__fd(rxq->xsk);
1309                 err = bpf_map_update_elem(bpf_map__fd(internals->map),
1310                                           &rxq->xsk_queue_idx, &fd, 0);
1311                 if (err) {
1312                         AF_XDP_LOG(ERR, "Failed to insert xsk in map.\n");
1313                         goto err;
1314                 }
1315         }
1316
1317 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
1318         ret = rte_pktmbuf_alloc_bulk(rxq->umem->mb_pool, fq_bufs, reserve_size);
1319         if (ret) {
1320                 AF_XDP_LOG(DEBUG, "Failed to get enough buffers for fq.\n");
1321                 goto err;
1322         }
1323 #endif
1324
1325         if (rxq->busy_budget) {
1326                 ret = configure_preferred_busy_poll(rxq);
1327                 if (ret) {
1328                         AF_XDP_LOG(ERR, "Failed configure busy polling.\n");
1329                         goto err;
1330                 }
1331         }
1332
1333         ret = reserve_fill_queue(rxq->umem, reserve_size, fq_bufs, &rxq->fq);
1334         if (ret) {
1335                 xsk_socket__delete(rxq->xsk);
1336                 AF_XDP_LOG(ERR, "Failed to reserve fill queue.\n");
1337                 goto err;
1338         }
1339
1340         return 0;
1341
1342 err:
1343         if (__atomic_sub_fetch(&rxq->umem->refcnt, 1, __ATOMIC_ACQUIRE) == 0)
1344                 xdp_umem_destroy(rxq->umem);
1345
1346         return ret;
1347 }
1348
1349 static int
1350 eth_rx_queue_setup(struct rte_eth_dev *dev,
1351                    uint16_t rx_queue_id,
1352                    uint16_t nb_rx_desc,
1353                    unsigned int socket_id __rte_unused,
1354                    const struct rte_eth_rxconf *rx_conf __rte_unused,
1355                    struct rte_mempool *mb_pool)
1356 {
1357         struct pmd_internals *internals = dev->data->dev_private;
1358         struct pkt_rx_queue *rxq;
1359         int ret;
1360
1361         rxq = &internals->rx_queues[rx_queue_id];
1362
1363         AF_XDP_LOG(INFO, "Set up rx queue, rx queue id: %d, xsk queue id: %d\n",
1364                    rx_queue_id, rxq->xsk_queue_idx);
1365
1366 #ifndef XDP_UMEM_UNALIGNED_CHUNK_FLAG
1367         uint32_t buf_size, data_size;
1368
1369         /* Now get the space available for data in the mbuf */
1370         buf_size = rte_pktmbuf_data_room_size(mb_pool) -
1371                 RTE_PKTMBUF_HEADROOM;
1372         data_size = ETH_AF_XDP_FRAME_SIZE;
1373
1374         if (data_size > buf_size) {
1375                 AF_XDP_LOG(ERR, "%s: %d bytes will not fit in mbuf (%d bytes)\n",
1376                         dev->device->name, data_size, buf_size);
1377                 ret = -ENOMEM;
1378                 goto err;
1379         }
1380 #endif
1381
1382         rxq->mb_pool = mb_pool;
1383
1384         if (xsk_configure(internals, rxq, nb_rx_desc)) {
1385                 AF_XDP_LOG(ERR, "Failed to configure xdp socket\n");
1386                 ret = -EINVAL;
1387                 goto err;
1388         }
1389
1390         if (!rxq->busy_budget)
1391                 AF_XDP_LOG(DEBUG, "Preferred busy polling not enabled\n");
1392
1393         rxq->fds[0].fd = xsk_socket__fd(rxq->xsk);
1394         rxq->fds[0].events = POLLIN;
1395
1396         dev->data->rx_queues[rx_queue_id] = rxq;
1397         return 0;
1398
1399 err:
1400         return ret;
1401 }
1402
1403 static int
1404 eth_tx_queue_setup(struct rte_eth_dev *dev,
1405                    uint16_t tx_queue_id,
1406                    uint16_t nb_tx_desc __rte_unused,
1407                    unsigned int socket_id __rte_unused,
1408                    const struct rte_eth_txconf *tx_conf __rte_unused)
1409 {
1410         struct pmd_internals *internals = dev->data->dev_private;
1411         struct pkt_tx_queue *txq;
1412
1413         txq = &internals->tx_queues[tx_queue_id];
1414
1415         dev->data->tx_queues[tx_queue_id] = txq;
1416         return 0;
1417 }
1418
1419 static int
1420 eth_dev_mtu_set(struct rte_eth_dev *dev, uint16_t mtu)
1421 {
1422         struct pmd_internals *internals = dev->data->dev_private;
1423         struct ifreq ifr = { .ifr_mtu = mtu };
1424         int ret;
1425         int s;
1426
1427         s = socket(PF_INET, SOCK_DGRAM, 0);
1428         if (s < 0)
1429                 return -EINVAL;
1430
1431         strlcpy(ifr.ifr_name, internals->if_name, IFNAMSIZ);
1432         ret = ioctl(s, SIOCSIFMTU, &ifr);
1433         close(s);
1434
1435         return (ret < 0) ? -errno : 0;
1436 }
1437
1438 static int
1439 eth_dev_change_flags(char *if_name, uint32_t flags, uint32_t mask)
1440 {
1441         struct ifreq ifr;
1442         int ret = 0;
1443         int s;
1444
1445         s = socket(PF_INET, SOCK_DGRAM, 0);
1446         if (s < 0)
1447                 return -errno;
1448
1449         strlcpy(ifr.ifr_name, if_name, IFNAMSIZ);
1450         if (ioctl(s, SIOCGIFFLAGS, &ifr) < 0) {
1451                 ret = -errno;
1452                 goto out;
1453         }
1454         ifr.ifr_flags &= mask;
1455         ifr.ifr_flags |= flags;
1456         if (ioctl(s, SIOCSIFFLAGS, &ifr) < 0) {
1457                 ret = -errno;
1458                 goto out;
1459         }
1460 out:
1461         close(s);
1462         return ret;
1463 }
1464
1465 static int
1466 eth_dev_promiscuous_enable(struct rte_eth_dev *dev)
1467 {
1468         struct pmd_internals *internals = dev->data->dev_private;
1469
1470         return eth_dev_change_flags(internals->if_name, IFF_PROMISC, ~0);
1471 }
1472
1473 static int
1474 eth_dev_promiscuous_disable(struct rte_eth_dev *dev)
1475 {
1476         struct pmd_internals *internals = dev->data->dev_private;
1477
1478         return eth_dev_change_flags(internals->if_name, 0, ~IFF_PROMISC);
1479 }
1480
1481 static const struct eth_dev_ops ops = {
1482         .dev_start = eth_dev_start,
1483         .dev_stop = eth_dev_stop,
1484         .dev_close = eth_dev_close,
1485         .dev_configure = eth_dev_configure,
1486         .dev_infos_get = eth_dev_info,
1487         .mtu_set = eth_dev_mtu_set,
1488         .promiscuous_enable = eth_dev_promiscuous_enable,
1489         .promiscuous_disable = eth_dev_promiscuous_disable,
1490         .rx_queue_setup = eth_rx_queue_setup,
1491         .tx_queue_setup = eth_tx_queue_setup,
1492         .link_update = eth_link_update,
1493         .stats_get = eth_stats_get,
1494         .stats_reset = eth_stats_reset,
1495         .get_monitor_addr = eth_get_monitor_addr,
1496 };
1497
1498 /** parse busy_budget argument */
1499 static int
1500 parse_budget_arg(const char *key __rte_unused,
1501                   const char *value, void *extra_args)
1502 {
1503         int *i = (int *)extra_args;
1504         char *end;
1505
1506         *i = strtol(value, &end, 10);
1507         if (*i < 0 || *i > UINT16_MAX) {
1508                 AF_XDP_LOG(ERR, "Invalid busy_budget %i, must be >= 0 and <= %u\n",
1509                                 *i, UINT16_MAX);
1510                 return -EINVAL;
1511         }
1512
1513         return 0;
1514 }
1515
1516 /** parse integer from integer argument */
1517 static int
1518 parse_integer_arg(const char *key __rte_unused,
1519                   const char *value, void *extra_args)
1520 {
1521         int *i = (int *)extra_args;
1522         char *end;
1523
1524         *i = strtol(value, &end, 10);
1525         if (*i < 0) {
1526                 AF_XDP_LOG(ERR, "Argument has to be positive.\n");
1527                 return -EINVAL;
1528         }
1529
1530         return 0;
1531 }
1532
1533 /** parse name argument */
1534 static int
1535 parse_name_arg(const char *key __rte_unused,
1536                const char *value, void *extra_args)
1537 {
1538         char *name = extra_args;
1539
1540         if (strnlen(value, IFNAMSIZ) > IFNAMSIZ - 1) {
1541                 AF_XDP_LOG(ERR, "Invalid name %s, should be less than %u bytes.\n",
1542                            value, IFNAMSIZ);
1543                 return -EINVAL;
1544         }
1545
1546         strlcpy(name, value, IFNAMSIZ);
1547
1548         return 0;
1549 }
1550
1551 /** parse xdp prog argument */
1552 static int
1553 parse_prog_arg(const char *key __rte_unused,
1554                const char *value, void *extra_args)
1555 {
1556         char *path = extra_args;
1557
1558         if (strnlen(value, PATH_MAX) == PATH_MAX) {
1559                 AF_XDP_LOG(ERR, "Invalid path %s, should be less than %u bytes.\n",
1560                            value, PATH_MAX);
1561                 return -EINVAL;
1562         }
1563
1564         if (access(value, F_OK) != 0) {
1565                 AF_XDP_LOG(ERR, "Error accessing %s: %s\n",
1566                            value, strerror(errno));
1567                 return -EINVAL;
1568         }
1569
1570         strlcpy(path, value, PATH_MAX);
1571
1572         return 0;
1573 }
1574
1575 static int
1576 xdp_get_channels_info(const char *if_name, int *max_queues,
1577                                 int *combined_queues)
1578 {
1579         struct ethtool_channels channels;
1580         struct ifreq ifr;
1581         int fd, ret;
1582
1583         fd = socket(AF_INET, SOCK_DGRAM, 0);
1584         if (fd < 0)
1585                 return -1;
1586
1587         channels.cmd = ETHTOOL_GCHANNELS;
1588         ifr.ifr_data = (void *)&channels;
1589         strlcpy(ifr.ifr_name, if_name, IFNAMSIZ);
1590         ret = ioctl(fd, SIOCETHTOOL, &ifr);
1591         if (ret) {
1592                 if (errno == EOPNOTSUPP) {
1593                         ret = 0;
1594                 } else {
1595                         ret = -errno;
1596                         goto out;
1597                 }
1598         }
1599
1600         if (channels.max_combined == 0 || errno == EOPNOTSUPP) {
1601                 /* If the device says it has no channels, then all traffic
1602                  * is sent to a single stream, so max queues = 1.
1603                  */
1604                 *max_queues = 1;
1605                 *combined_queues = 1;
1606         } else {
1607                 *max_queues = channels.max_combined;
1608                 *combined_queues = channels.combined_count;
1609         }
1610
1611  out:
1612         close(fd);
1613         return ret;
1614 }
1615
1616 static int
1617 parse_parameters(struct rte_kvargs *kvlist, char *if_name, int *start_queue,
1618                         int *queue_cnt, int *shared_umem, char *prog_path,
1619                         int *busy_budget)
1620 {
1621         int ret;
1622
1623         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_IFACE_ARG,
1624                                  &parse_name_arg, if_name);
1625         if (ret < 0)
1626                 goto free_kvlist;
1627
1628         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_START_QUEUE_ARG,
1629                                  &parse_integer_arg, start_queue);
1630         if (ret < 0)
1631                 goto free_kvlist;
1632
1633         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_QUEUE_COUNT_ARG,
1634                                  &parse_integer_arg, queue_cnt);
1635         if (ret < 0 || *queue_cnt <= 0) {
1636                 ret = -EINVAL;
1637                 goto free_kvlist;
1638         }
1639
1640         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_SHARED_UMEM_ARG,
1641                                 &parse_integer_arg, shared_umem);
1642         if (ret < 0)
1643                 goto free_kvlist;
1644
1645         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_PROG_ARG,
1646                                  &parse_prog_arg, prog_path);
1647         if (ret < 0)
1648                 goto free_kvlist;
1649
1650         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_BUDGET_ARG,
1651                                 &parse_budget_arg, busy_budget);
1652         if (ret < 0)
1653                 goto free_kvlist;
1654
1655 free_kvlist:
1656         rte_kvargs_free(kvlist);
1657         return ret;
1658 }
1659
1660 static int
1661 get_iface_info(const char *if_name,
1662                struct rte_ether_addr *eth_addr,
1663                int *if_index)
1664 {
1665         struct ifreq ifr;
1666         int sock = socket(AF_INET, SOCK_DGRAM, IPPROTO_IP);
1667
1668         if (sock < 0)
1669                 return -1;
1670
1671         strlcpy(ifr.ifr_name, if_name, IFNAMSIZ);
1672         if (ioctl(sock, SIOCGIFINDEX, &ifr))
1673                 goto error;
1674
1675         *if_index = ifr.ifr_ifindex;
1676
1677         if (ioctl(sock, SIOCGIFHWADDR, &ifr))
1678                 goto error;
1679
1680         rte_memcpy(eth_addr, ifr.ifr_hwaddr.sa_data, RTE_ETHER_ADDR_LEN);
1681
1682         close(sock);
1683         return 0;
1684
1685 error:
1686         close(sock);
1687         return -1;
1688 }
1689
1690 static struct rte_eth_dev *
1691 init_internals(struct rte_vdev_device *dev, const char *if_name,
1692                 int start_queue_idx, int queue_cnt, int shared_umem,
1693                 const char *prog_path, int busy_budget)
1694 {
1695         const char *name = rte_vdev_device_name(dev);
1696         const unsigned int numa_node = dev->device.numa_node;
1697         struct pmd_internals *internals;
1698         struct rte_eth_dev *eth_dev;
1699         int ret;
1700         int i;
1701
1702         internals = rte_zmalloc_socket(name, sizeof(*internals), 0, numa_node);
1703         if (internals == NULL)
1704                 return NULL;
1705
1706         internals->start_queue_idx = start_queue_idx;
1707         internals->queue_cnt = queue_cnt;
1708         strlcpy(internals->if_name, if_name, IFNAMSIZ);
1709         strlcpy(internals->prog_path, prog_path, PATH_MAX);
1710         internals->custom_prog_configured = 0;
1711
1712 #ifndef ETH_AF_XDP_SHARED_UMEM
1713         if (shared_umem) {
1714                 AF_XDP_LOG(ERR, "Shared UMEM feature not available. "
1715                                 "Check kernel and libbpf version\n");
1716                 goto err_free_internals;
1717         }
1718 #endif
1719         internals->shared_umem = shared_umem;
1720
1721         if (xdp_get_channels_info(if_name, &internals->max_queue_cnt,
1722                                   &internals->combined_queue_cnt)) {
1723                 AF_XDP_LOG(ERR, "Failed to get channel info of interface: %s\n",
1724                                 if_name);
1725                 goto err_free_internals;
1726         }
1727
1728         if (queue_cnt > internals->combined_queue_cnt) {
1729                 AF_XDP_LOG(ERR, "Specified queue count %d is larger than combined queue count %d.\n",
1730                                 queue_cnt, internals->combined_queue_cnt);
1731                 goto err_free_internals;
1732         }
1733
1734         internals->rx_queues = rte_zmalloc_socket(NULL,
1735                                         sizeof(struct pkt_rx_queue) * queue_cnt,
1736                                         0, numa_node);
1737         if (internals->rx_queues == NULL) {
1738                 AF_XDP_LOG(ERR, "Failed to allocate memory for rx queues.\n");
1739                 goto err_free_internals;
1740         }
1741
1742         internals->tx_queues = rte_zmalloc_socket(NULL,
1743                                         sizeof(struct pkt_tx_queue) * queue_cnt,
1744                                         0, numa_node);
1745         if (internals->tx_queues == NULL) {
1746                 AF_XDP_LOG(ERR, "Failed to allocate memory for tx queues.\n");
1747                 goto err_free_rx;
1748         }
1749         for (i = 0; i < queue_cnt; i++) {
1750                 internals->tx_queues[i].pair = &internals->rx_queues[i];
1751                 internals->rx_queues[i].pair = &internals->tx_queues[i];
1752                 internals->rx_queues[i].xsk_queue_idx = start_queue_idx + i;
1753                 internals->tx_queues[i].xsk_queue_idx = start_queue_idx + i;
1754                 internals->rx_queues[i].busy_budget = busy_budget;
1755         }
1756
1757         ret = get_iface_info(if_name, &internals->eth_addr,
1758                              &internals->if_index);
1759         if (ret)
1760                 goto err_free_tx;
1761
1762         eth_dev = rte_eth_vdev_allocate(dev, 0);
1763         if (eth_dev == NULL)
1764                 goto err_free_tx;
1765
1766         eth_dev->data->dev_private = internals;
1767         eth_dev->data->dev_link = pmd_link;
1768         eth_dev->data->mac_addrs = &internals->eth_addr;
1769         eth_dev->data->dev_flags |= RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS;
1770         eth_dev->dev_ops = &ops;
1771         eth_dev->rx_pkt_burst = eth_af_xdp_rx;
1772         eth_dev->tx_pkt_burst = eth_af_xdp_tx;
1773
1774 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
1775         AF_XDP_LOG(INFO, "Zero copy between umem and mbuf enabled.\n");
1776 #endif
1777
1778         return eth_dev;
1779
1780 err_free_tx:
1781         rte_free(internals->tx_queues);
1782 err_free_rx:
1783         rte_free(internals->rx_queues);
1784 err_free_internals:
1785         rte_free(internals);
1786         return NULL;
1787 }
1788
1789 static int
1790 rte_pmd_af_xdp_probe(struct rte_vdev_device *dev)
1791 {
1792         struct rte_kvargs *kvlist;
1793         char if_name[IFNAMSIZ] = {'\0'};
1794         int xsk_start_queue_idx = ETH_AF_XDP_DFLT_START_QUEUE_IDX;
1795         int xsk_queue_cnt = ETH_AF_XDP_DFLT_QUEUE_COUNT;
1796         int shared_umem = 0;
1797         char prog_path[PATH_MAX] = {'\0'};
1798         int busy_budget = -1;
1799         struct rte_eth_dev *eth_dev = NULL;
1800         const char *name;
1801
1802         AF_XDP_LOG(INFO, "Initializing pmd_af_xdp for %s\n",
1803                 rte_vdev_device_name(dev));
1804
1805         name = rte_vdev_device_name(dev);
1806         if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1807                 AF_XDP_LOG(ERR, "Failed to probe %s. "
1808                                 "AF_XDP PMD does not support secondary processes.\n",
1809                                 name);
1810                 return -ENOTSUP;
1811         }
1812
1813         kvlist = rte_kvargs_parse(rte_vdev_device_args(dev), valid_arguments);
1814         if (kvlist == NULL) {
1815                 AF_XDP_LOG(ERR, "Invalid kvargs key\n");
1816                 return -EINVAL;
1817         }
1818
1819         if (dev->device.numa_node == SOCKET_ID_ANY)
1820                 dev->device.numa_node = rte_socket_id();
1821
1822         if (parse_parameters(kvlist, if_name, &xsk_start_queue_idx,
1823                              &xsk_queue_cnt, &shared_umem, prog_path,
1824                              &busy_budget) < 0) {
1825                 AF_XDP_LOG(ERR, "Invalid kvargs value\n");
1826                 return -EINVAL;
1827         }
1828
1829         if (strlen(if_name) == 0) {
1830                 AF_XDP_LOG(ERR, "Network interface must be specified\n");
1831                 return -EINVAL;
1832         }
1833
1834         busy_budget = busy_budget == -1 ? ETH_AF_XDP_DFLT_BUSY_BUDGET :
1835                                         busy_budget;
1836
1837         eth_dev = init_internals(dev, if_name, xsk_start_queue_idx,
1838                                         xsk_queue_cnt, shared_umem, prog_path,
1839                                         busy_budget);
1840         if (eth_dev == NULL) {
1841                 AF_XDP_LOG(ERR, "Failed to init internals\n");
1842                 return -1;
1843         }
1844
1845         rte_eth_dev_probing_finish(eth_dev);
1846
1847         return 0;
1848 }
1849
1850 static int
1851 rte_pmd_af_xdp_remove(struct rte_vdev_device *dev)
1852 {
1853         struct rte_eth_dev *eth_dev = NULL;
1854
1855         AF_XDP_LOG(INFO, "Removing AF_XDP ethdev on numa socket %u\n",
1856                 rte_socket_id());
1857
1858         if (dev == NULL)
1859                 return -1;
1860
1861         /* find the ethdev entry */
1862         eth_dev = rte_eth_dev_allocated(rte_vdev_device_name(dev));
1863         if (eth_dev == NULL)
1864                 return 0;
1865
1866         eth_dev_close(eth_dev);
1867         rte_eth_dev_release_port(eth_dev);
1868
1869
1870         return 0;
1871 }
1872
1873 static struct rte_vdev_driver pmd_af_xdp_drv = {
1874         .probe = rte_pmd_af_xdp_probe,
1875         .remove = rte_pmd_af_xdp_remove,
1876 };
1877
1878 RTE_PMD_REGISTER_VDEV(net_af_xdp, pmd_af_xdp_drv);
1879 RTE_PMD_REGISTER_PARAM_STRING(net_af_xdp,
1880                               "iface=<string> "
1881                               "start_queue=<int> "
1882                               "queue_count=<int> "
1883                               "shared_umem=<int> "
1884                               "xdp_prog=<string> "
1885                               "busy_budget=<int>");