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