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