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 <ethdev_driver.h>
  23 #include <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 i;
 259         struct rte_mbuf *fq_bufs[ETH_AF_XDP_RX_BATCH_SIZE];
 260
 261         nb_pkts = xsk_ring_cons__peek(rx, nb_pkts, &idx_rx);
 262
 263         if (nb_pkts == 0) {
 264 #if defined(XDP_USE_NEED_WAKEUP)
 265                 if (xsk_ring_prod__needs_wakeup(fq))
 266                         (void)poll(rxq->fds, 1, 1000);
 267 #endif
 268
 269                 return 0;
 270         }
 271
 272         /* allocate bufs for fill queue replenishment after rx */
 273         if (rte_pktmbuf_alloc_bulk(umem->mb_pool, fq_bufs, nb_pkts)) {
 274                 AF_XDP_LOG(DEBUG,
 275                         "Failed to get enough buffers for fq.\n");
 276                 /* rollback cached_cons which is added by
 277                  * xsk_ring_cons__peek
 278                  */
 279                 rx->cached_cons -= nb_pkts;
 280                 return 0;
 281         }
 282
 283         for (i = 0; i < nb_pkts; i++) {
 284                 const struct xdp_desc *desc;
 285                 uint64_t addr;
 286                 uint32_t len;
 287                 uint64_t offset;
 288
 289                 desc = xsk_ring_cons__rx_desc(rx, idx_rx++);
 290                 addr = desc->addr;
 291                 len = desc->len;
 292
 293                 offset = xsk_umem__extract_offset(addr);
 294                 addr = xsk_umem__extract_addr(addr);
 295
 296                 bufs[i] = (struct rte_mbuf *)
 297                                 xsk_umem__get_data(umem->buffer, addr +
 298                                         umem->mb_pool->header_size);
 299                 bufs[i]->data_off = offset - sizeof(struct rte_mbuf) -
 300                         rte_pktmbuf_priv_size(umem->mb_pool) -
 301                         umem->mb_pool->header_size;
 302
 303                 rte_pktmbuf_pkt_len(bufs[i]) = len;
 304                 rte_pktmbuf_data_len(bufs[i]) = len;
 305                 rx_bytes += len;
 306         }
 307
 308         xsk_ring_cons__release(rx, nb_pkts);
 309         (void)reserve_fill_queue(umem, nb_pkts, fq_bufs, fq);
 310
 311         /* statistics */
 312         rxq->stats.rx_pkts += nb_pkts;
 313         rxq->stats.rx_bytes += rx_bytes;
 314
 315         return nb_pkts;
 316 }
 317 #else
 318 static uint16_t
 319 af_xdp_rx_cp(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 320 {
 321         struct pkt_rx_queue *rxq = queue;
 322         struct xsk_ring_cons *rx = &rxq->rx;
 323         struct xsk_umem_info *umem = rxq->umem;
 324         struct xsk_ring_prod *fq = &rxq->fq;
 325         uint32_t idx_rx = 0;
 326         unsigned long rx_bytes = 0;
 327         int i;
 328         uint32_t free_thresh = fq->size >> 1;
 329         struct rte_mbuf *mbufs[ETH_AF_XDP_RX_BATCH_SIZE];
 330
 331         if (xsk_prod_nb_free(fq, free_thresh) >= free_thresh)
 332                 (void)reserve_fill_queue(umem, ETH_AF_XDP_RX_BATCH_SIZE,
 333                                          NULL, fq);
 334
 335         nb_pkts = xsk_ring_cons__peek(rx, nb_pkts, &idx_rx);
 336         if (nb_pkts == 0) {
 337 #if defined(XDP_USE_NEED_WAKEUP)
 338                 if (xsk_ring_prod__needs_wakeup(fq))
 339                         (void)poll(rxq->fds, 1, 1000);
 340 #endif
 341                 return 0;
 342         }
 343
 344         if (unlikely(rte_pktmbuf_alloc_bulk(rxq->mb_pool, mbufs, nb_pkts))) {
 345                 /* rollback cached_cons which is added by
 346                  * xsk_ring_cons__peek
 347                  */
 348                 rx->cached_cons -= nb_pkts;
 349                 return 0;
 350         }
 351
 352         for (i = 0; i < nb_pkts; i++) {
 353                 const struct xdp_desc *desc;
 354                 uint64_t addr;
 355                 uint32_t len;
 356                 void *pkt;
 357
 358                 desc = xsk_ring_cons__rx_desc(rx, idx_rx++);
 359                 addr = desc->addr;
 360                 len = desc->len;
 361                 pkt = xsk_umem__get_data(rxq->umem->mz->addr, addr);
 362
 363                 rte_memcpy(rte_pktmbuf_mtod(mbufs[i], void *), pkt, len);
 364                 rte_ring_enqueue(umem->buf_ring, (void *)addr);
 365                 rte_pktmbuf_pkt_len(mbufs[i]) = len;
 366                 rte_pktmbuf_data_len(mbufs[i]) = len;
 367                 rx_bytes += len;
 368                 bufs[i] = mbufs[i];
 369         }
 370
 371         xsk_ring_cons__release(rx, nb_pkts);
 372
 373         /* statistics */
 374         rxq->stats.rx_pkts += nb_pkts;
 375         rxq->stats.rx_bytes += rx_bytes;
 376
 377         return nb_pkts;
 378 }
 379 #endif
 380
 381 static uint16_t
 382 eth_af_xdp_rx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 383 {
 384         nb_pkts = RTE_MIN(nb_pkts, ETH_AF_XDP_RX_BATCH_SIZE);
 385
 386 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 387         return af_xdp_rx_zc(queue, bufs, nb_pkts);
 388 #else
 389         return af_xdp_rx_cp(queue, bufs, nb_pkts);
 390 #endif
 391 }
 392
 393 static void
 394 pull_umem_cq(struct xsk_umem_info *umem, int size, struct xsk_ring_cons *cq)
 395 {
 396         size_t i, n;
 397         uint32_t idx_cq = 0;
 398
 399         n = xsk_ring_cons__peek(cq, size, &idx_cq);
 400
 401         for (i = 0; i < n; i++) {
 402                 uint64_t addr;
 403                 addr = *xsk_ring_cons__comp_addr(cq, idx_cq++);
 404 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 405                 addr = xsk_umem__extract_addr(addr);
 406                 rte_pktmbuf_free((struct rte_mbuf *)
 407                                         xsk_umem__get_data(umem->buffer,
 408                                         addr + umem->mb_pool->header_size));
 409 #else
 410                 rte_ring_enqueue(umem->buf_ring, (void *)addr);
 411 #endif
 412         }
 413
 414         xsk_ring_cons__release(cq, n);
 415 }
 416
 417 static void
 418 kick_tx(struct pkt_tx_queue *txq, struct xsk_ring_cons *cq)
 419 {
 420         struct xsk_umem_info *umem = txq->umem;
 421
 422         pull_umem_cq(umem, XSK_RING_CONS__DEFAULT_NUM_DESCS, cq);
 423
 424 #if defined(XDP_USE_NEED_WAKEUP)
 425         if (xsk_ring_prod__needs_wakeup(&txq->tx))
 426 #endif
 427                 while (send(xsk_socket__fd(txq->pair->xsk), NULL,
 428                             0, MSG_DONTWAIT) < 0) {
 429                         /* some thing unexpected */
 430                         if (errno != EBUSY && errno != EAGAIN && errno != EINTR)
 431                                 break;
 432
 433                         /* pull from completion queue to leave more space */
 434                         if (errno == EAGAIN)
 435                                 pull_umem_cq(umem,
 436                                              XSK_RING_CONS__DEFAULT_NUM_DESCS,
 437                                              cq);
 438                 }
 439 }
 440
 441 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 442 static uint16_t
 443 af_xdp_tx_zc(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 444 {
 445         struct pkt_tx_queue *txq = queue;
 446         struct xsk_umem_info *umem = txq->umem;
 447         struct rte_mbuf *mbuf;
 448         unsigned long tx_bytes = 0;
 449         int i;
 450         uint32_t idx_tx;
 451         uint16_t count = 0;
 452         struct xdp_desc *desc;
 453         uint64_t addr, offset;
 454         struct xsk_ring_cons *cq = &txq->pair->cq;
 455         uint32_t free_thresh = cq->size >> 1;
 456
 457         if (xsk_cons_nb_avail(cq, free_thresh) >= free_thresh)
 458                 pull_umem_cq(umem, XSK_RING_CONS__DEFAULT_NUM_DESCS, cq);
 459
 460         for (i = 0; i < nb_pkts; i++) {
 461                 mbuf = bufs[i];
 462
 463                 if (mbuf->pool == umem->mb_pool) {
 464                         if (!xsk_ring_prod__reserve(&txq->tx, 1, &idx_tx)) {
 465                                 kick_tx(txq, cq);
 466                                 if (!xsk_ring_prod__reserve(&txq->tx, 1,
 467                                                             &idx_tx))
 468                                         goto out;
 469                         }
 470                         desc = xsk_ring_prod__tx_desc(&txq->tx, idx_tx);
 471                         desc->len = mbuf->pkt_len;
 472                         addr = (uint64_t)mbuf - (uint64_t)umem->buffer -
 473                                         umem->mb_pool->header_size;
 474                         offset = rte_pktmbuf_mtod(mbuf, uint64_t) -
 475                                         (uint64_t)mbuf +
 476                                         umem->mb_pool->header_size;
 477                         offset = offset << XSK_UNALIGNED_BUF_OFFSET_SHIFT;
 478                         desc->addr = addr | offset;
 479                         count++;
 480                 } else {
 481                         struct rte_mbuf *local_mbuf =
 482                                         rte_pktmbuf_alloc(umem->mb_pool);
 483                         void *pkt;
 484
 485                         if (local_mbuf == NULL)
 486                                 goto out;
 487
 488                         if (!xsk_ring_prod__reserve(&txq->tx, 1, &idx_tx)) {
 489                                 rte_pktmbuf_free(local_mbuf);
 490                                 kick_tx(txq, cq);
 491                                 goto out;
 492                         }
 493
 494                         desc = xsk_ring_prod__tx_desc(&txq->tx, idx_tx);
 495                         desc->len = mbuf->pkt_len;
 496
 497                         addr = (uint64_t)local_mbuf - (uint64_t)umem->buffer -
 498                                         umem->mb_pool->header_size;
 499                         offset = rte_pktmbuf_mtod(local_mbuf, uint64_t) -
 500                                         (uint64_t)local_mbuf +
 501                                         umem->mb_pool->header_size;
 502                         pkt = xsk_umem__get_data(umem->buffer, addr + offset);
 503                         offset = offset << XSK_UNALIGNED_BUF_OFFSET_SHIFT;
 504                         desc->addr = addr | offset;
 505                         rte_memcpy(pkt, rte_pktmbuf_mtod(mbuf, void *),
 506                                         desc->len);
 507                         rte_pktmbuf_free(mbuf);
 508                         count++;
 509                 }
 510
 511                 tx_bytes += mbuf->pkt_len;
 512         }
 513
 514         kick_tx(txq, cq);
 515
 516 out:
 517         xsk_ring_prod__submit(&txq->tx, count);
 518
 519         txq->stats.tx_pkts += count;
 520         txq->stats.tx_bytes += tx_bytes;
 521         txq->stats.tx_dropped += nb_pkts - count;
 522
 523         return count;
 524 }
 525 #else
 526 static uint16_t
 527 af_xdp_tx_cp(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 528 {
 529         struct pkt_tx_queue *txq = queue;
 530         struct xsk_umem_info *umem = txq->umem;
 531         struct rte_mbuf *mbuf;
 532         void *addrs[ETH_AF_XDP_TX_BATCH_SIZE];
 533         unsigned long tx_bytes = 0;
 534         int i;
 535         uint32_t idx_tx;
 536         struct xsk_ring_cons *cq = &txq->pair->cq;
 537
 538         nb_pkts = RTE_MIN(nb_pkts, ETH_AF_XDP_TX_BATCH_SIZE);
 539
 540         pull_umem_cq(umem, nb_pkts, cq);
 541
 542         nb_pkts = rte_ring_dequeue_bulk(umem->buf_ring, addrs,
 543                                         nb_pkts, NULL);
 544         if (nb_pkts == 0)
 545                 return 0;
 546
 547         if (xsk_ring_prod__reserve(&txq->tx, nb_pkts, &idx_tx) != nb_pkts) {
 548                 kick_tx(txq, cq);
 549                 rte_ring_enqueue_bulk(umem->buf_ring, addrs, nb_pkts, NULL);
 550                 return 0;
 551         }
 552
 553         for (i = 0; i < nb_pkts; i++) {
 554                 struct xdp_desc *desc;
 555                 void *pkt;
 556
 557                 desc = xsk_ring_prod__tx_desc(&txq->tx, idx_tx + i);
 558                 mbuf = bufs[i];
 559                 desc->len = mbuf->pkt_len;
 560
 561                 desc->addr = (uint64_t)addrs[i];
 562                 pkt = xsk_umem__get_data(umem->mz->addr,
 563                                          desc->addr);
 564                 rte_memcpy(pkt, rte_pktmbuf_mtod(mbuf, void *), desc->len);
 565                 tx_bytes += mbuf->pkt_len;
 566                 rte_pktmbuf_free(mbuf);
 567         }
 568
 569         xsk_ring_prod__submit(&txq->tx, nb_pkts);
 570
 571         kick_tx(txq, cq);
 572
 573         txq->stats.tx_pkts += nb_pkts;
 574         txq->stats.tx_bytes += tx_bytes;
 575
 576         return nb_pkts;
 577 }
 578 #endif
 579
 580 static uint16_t
 581 eth_af_xdp_tx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 582 {
 583 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 584         return af_xdp_tx_zc(queue, bufs, nb_pkts);
 585 #else
 586         return af_xdp_tx_cp(queue, bufs, nb_pkts);
 587 #endif
 588 }
 589
 590 static int
 591 eth_dev_start(struct rte_eth_dev *dev)
 592 {
 593         dev->data->dev_link.link_status = ETH_LINK_UP;
 594
 595         return 0;
 596 }
 597
 598 /* This function gets called when the current port gets stopped. */
 599 static int
 600 eth_dev_stop(struct rte_eth_dev *dev)
 601 {
 602         dev->data->dev_link.link_status = ETH_LINK_DOWN;
 603         return 0;
 604 }
 605
 606 /* Find ethdev in list */
 607 static inline struct internal_list *
 608 find_internal_resource(struct pmd_internals *port_int)
 609 {
 610         int found = 0;
 611         struct internal_list *list = NULL;
 612
 613         if (port_int == NULL)
 614                 return NULL;
 615
 616         pthread_mutex_lock(&internal_list_lock);
 617
 618         TAILQ_FOREACH(list, &internal_list, next) {
 619                 struct pmd_internals *list_int =
 620                                 list->eth_dev->data->dev_private;
 621                 if (list_int == port_int) {
 622                         found = 1;
 623                         break;
 624                 }
 625         }
 626
 627         pthread_mutex_unlock(&internal_list_lock);
 628
 629         if (!found)
 630                 return NULL;
 631
 632         return list;
 633 }
 634
 635 /* Check if the netdev,qid context already exists */
 636 static inline bool
 637 ctx_exists(struct pkt_rx_queue *rxq, const char *ifname,
 638                 struct pkt_rx_queue *list_rxq, const char *list_ifname)
 639 {
 640         bool exists = false;
 641
 642         if (rxq->xsk_queue_idx == list_rxq->xsk_queue_idx &&
 643                         !strncmp(ifname, list_ifname, IFNAMSIZ)) {
 644                 AF_XDP_LOG(ERR, "ctx %s,%i already exists, cannot share umem\n",
 645                                         ifname, rxq->xsk_queue_idx);
 646                 exists = true;
 647         }
 648
 649         return exists;
 650 }
 651
 652 /* Get a pointer to an existing UMEM which overlays the rxq's mb_pool */
 653 static inline int
 654 get_shared_umem(struct pkt_rx_queue *rxq, const char *ifname,
 655                         struct xsk_umem_info **umem)
 656 {
 657         struct internal_list *list;
 658         struct pmd_internals *internals;
 659         int i = 0, ret = 0;
 660         struct rte_mempool *mb_pool = rxq->mb_pool;
 661
 662         if (mb_pool == NULL)
 663                 return ret;
 664
 665         pthread_mutex_lock(&internal_list_lock);
 666
 667         TAILQ_FOREACH(list, &internal_list, next) {
 668                 internals = list->eth_dev->data->dev_private;
 669                 for (i = 0; i < internals->queue_cnt; i++) {
 670                         struct pkt_rx_queue *list_rxq =
 671                                                 &internals->rx_queues[i];
 672                         if (rxq == list_rxq)
 673                                 continue;
 674                         if (mb_pool == internals->rx_queues[i].mb_pool) {
 675                                 if (ctx_exists(rxq, ifname, list_rxq,
 676                                                 internals->if_name)) {
 677                                         ret = -1;
 678                                         goto out;
 679                                 }
 680                                 if (__atomic_load_n(
 681                                         &internals->rx_queues[i].umem->refcnt,
 682                                                         __ATOMIC_ACQUIRE)) {
 683                                         *umem = internals->rx_queues[i].umem;
 684                                         goto out;
 685                                 }
 686                         }
 687                 }
 688         }
 689
 690 out:
 691         pthread_mutex_unlock(&internal_list_lock);
 692
 693         return ret;
 694 }
 695
 696 static int
 697 eth_dev_configure(struct rte_eth_dev *dev)
 698 {
 699         struct pmd_internals *internal = dev->data->dev_private;
 700
 701         /* rx/tx must be paired */
 702         if (dev->data->nb_rx_queues != dev->data->nb_tx_queues)
 703                 return -EINVAL;
 704
 705         if (internal->shared_umem) {
 706                 struct internal_list *list = NULL;
 707                 const char *name = dev->device->name;
 708
 709                 /* Ensure PMD is not already inserted into the list */
 710                 list = find_internal_resource(internal);
 711                 if (list)
 712                         return 0;
 713
 714                 list = rte_zmalloc_socket(name, sizeof(*list), 0,
 715                                         dev->device->numa_node);
 716                 if (list == NULL)
 717                         return -1;
 718
 719                 list->eth_dev = dev;
 720                 pthread_mutex_lock(&internal_list_lock);
 721                 TAILQ_INSERT_TAIL(&internal_list, list, next);
 722                 pthread_mutex_unlock(&internal_list_lock);
 723         }
 724
 725         return 0;
 726 }
 727
 728 static int
 729 eth_dev_info(struct rte_eth_dev *dev, struct rte_eth_dev_info *dev_info)
 730 {
 731         struct pmd_internals *internals = dev->data->dev_private;
 732
 733         dev_info->if_index = internals->if_index;
 734         dev_info->max_mac_addrs = 1;
 735         dev_info->max_rx_pktlen = ETH_FRAME_LEN;
 736         dev_info->max_rx_queues = internals->queue_cnt;
 737         dev_info->max_tx_queues = internals->queue_cnt;
 738
 739         dev_info->min_mtu = RTE_ETHER_MIN_MTU;
 740 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 741         dev_info->max_mtu = getpagesize() -
 742                                 sizeof(struct rte_mempool_objhdr) -
 743                                 sizeof(struct rte_mbuf) -
 744                                 RTE_PKTMBUF_HEADROOM - XDP_PACKET_HEADROOM;
 745 #else
 746         dev_info->max_mtu = ETH_AF_XDP_FRAME_SIZE - XDP_PACKET_HEADROOM;
 747 #endif
 748
 749         dev_info->default_rxportconf.nb_queues = 1;
 750         dev_info->default_txportconf.nb_queues = 1;
 751         dev_info->default_rxportconf.ring_size = ETH_AF_XDP_DFLT_NUM_DESCS;
 752         dev_info->default_txportconf.ring_size = ETH_AF_XDP_DFLT_NUM_DESCS;
 753
 754         return 0;
 755 }
 756
 757 static int
 758 eth_stats_get(struct rte_eth_dev *dev, struct rte_eth_stats *stats)
 759 {
 760         struct pmd_internals *internals = dev->data->dev_private;
 761         struct xdp_statistics xdp_stats;
 762         struct pkt_rx_queue *rxq;
 763         struct pkt_tx_queue *txq;
 764         socklen_t optlen;
 765         int i, ret;
 766
 767         for (i = 0; i < dev->data->nb_rx_queues; i++) {
 768                 optlen = sizeof(struct xdp_statistics);
 769                 rxq = &internals->rx_queues[i];
 770                 txq = rxq->pair;
 771                 stats->q_ipackets[i] = rxq->stats.rx_pkts;
 772                 stats->q_ibytes[i] = rxq->stats.rx_bytes;
 773
 774                 stats->q_opackets[i] = txq->stats.tx_pkts;
 775                 stats->q_obytes[i] = txq->stats.tx_bytes;
 776
 777                 stats->ipackets += stats->q_ipackets[i];
 778                 stats->ibytes += stats->q_ibytes[i];
 779                 stats->imissed += rxq->stats.rx_dropped;
 780                 stats->oerrors += txq->stats.tx_dropped;
 781                 ret = getsockopt(xsk_socket__fd(rxq->xsk), SOL_XDP,
 782                                 XDP_STATISTICS, &xdp_stats, &optlen);
 783                 if (ret != 0) {
 784                         AF_XDP_LOG(ERR, "getsockopt() failed for XDP_STATISTICS.\n");
 785                         return -1;
 786                 }
 787                 stats->imissed += xdp_stats.rx_dropped;
 788
 789                 stats->opackets += stats->q_opackets[i];
 790                 stats->obytes += stats->q_obytes[i];
 791         }
 792
 793         return 0;
 794 }
 795
 796 static int
 797 eth_stats_reset(struct rte_eth_dev *dev)
 798 {
 799         struct pmd_internals *internals = dev->data->dev_private;
 800         int i;
 801
 802         for (i = 0; i < internals->queue_cnt; i++) {
 803                 memset(&internals->rx_queues[i].stats, 0,
 804                                         sizeof(struct rx_stats));
 805                 memset(&internals->tx_queues[i].stats, 0,
 806                                         sizeof(struct tx_stats));
 807         }
 808
 809         return 0;
 810 }
 811
 812 static void
 813 remove_xdp_program(struct pmd_internals *internals)
 814 {
 815         uint32_t curr_prog_id = 0;
 816
 817         if (bpf_get_link_xdp_id(internals->if_index, &curr_prog_id,
 818                                 XDP_FLAGS_UPDATE_IF_NOEXIST)) {
 819                 AF_XDP_LOG(ERR, "bpf_get_link_xdp_id failed\n");
 820                 return;
 821         }
 822         bpf_set_link_xdp_fd(internals->if_index, -1,
 823                         XDP_FLAGS_UPDATE_IF_NOEXIST);
 824 }
 825
 826 static void
 827 xdp_umem_destroy(struct xsk_umem_info *umem)
 828 {
 829 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 830         umem->mb_pool = NULL;
 831 #else
 832         rte_memzone_free(umem->mz);
 833         umem->mz = NULL;
 834
 835         rte_ring_free(umem->buf_ring);
 836         umem->buf_ring = NULL;
 837 #endif
 838
 839         rte_free(umem);
 840 }
 841
 842 static int
 843 eth_dev_close(struct rte_eth_dev *dev)
 844 {
 845         struct pmd_internals *internals = dev->data->dev_private;
 846         struct pkt_rx_queue *rxq;
 847         int i;
 848
 849         if (rte_eal_process_type() != RTE_PROC_PRIMARY)
 850                 return 0;
 851
 852         AF_XDP_LOG(INFO, "Closing AF_XDP ethdev on numa socket %u\n",
 853                 rte_socket_id());
 854
 855         for (i = 0; i < internals->queue_cnt; i++) {
 856                 rxq = &internals->rx_queues[i];
 857                 if (rxq->umem == NULL)
 858                         break;
 859                 xsk_socket__delete(rxq->xsk);
 860
 861                 if (__atomic_sub_fetch(&rxq->umem->refcnt, 1, __ATOMIC_ACQUIRE)
 862                                 == 0) {
 863                         (void)xsk_umem__delete(rxq->umem->umem);
 864                         xdp_umem_destroy(rxq->umem);
 865                 }
 866
 867                 /* free pkt_tx_queue */
 868                 rte_free(rxq->pair);
 869                 rte_free(rxq);
 870         }
 871
 872         /*
 873          * MAC is not allocated dynamically, setting it to NULL would prevent
 874          * from releasing it in rte_eth_dev_release_port.
 875          */
 876         dev->data->mac_addrs = NULL;
 877
 878         remove_xdp_program(internals);
 879
 880         if (internals->shared_umem) {
 881                 struct internal_list *list;
 882
 883                 /* Remove ethdev from list used to track and share UMEMs */
 884                 list = find_internal_resource(internals);
 885                 if (list) {
 886                         pthread_mutex_lock(&internal_list_lock);
 887                         TAILQ_REMOVE(&internal_list, list, next);
 888                         pthread_mutex_unlock(&internal_list_lock);
 889                         rte_free(list);
 890                 }
 891         }
 892
 893         return 0;
 894 }
 895
 896 static void
 897 eth_queue_release(void *q __rte_unused)
 898 {
 899 }
 900
 901 static int
 902 eth_link_update(struct rte_eth_dev *dev __rte_unused,
 903                 int wait_to_complete __rte_unused)
 904 {
 905         return 0;
 906 }
 907
 908 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 909 static inline uintptr_t get_base_addr(struct rte_mempool *mp, uint64_t *align)
 910 {
 911         struct rte_mempool_memhdr *memhdr;
 912         uintptr_t memhdr_addr, aligned_addr;
 913
 914         memhdr = STAILQ_FIRST(&mp->mem_list);
 915         memhdr_addr = (uintptr_t)memhdr->addr;
 916         aligned_addr = memhdr_addr & ~(getpagesize() - 1);
 917         *align = memhdr_addr - aligned_addr;
 918
 919         return aligned_addr;
 920 }
 921
 922 static struct
 923 xsk_umem_info *xdp_umem_configure(struct pmd_internals *internals,
 924                                   struct pkt_rx_queue *rxq)
 925 {
 926         struct xsk_umem_info *umem = NULL;
 927         int ret;
 928         struct xsk_umem_config usr_config = {
 929                 .fill_size = ETH_AF_XDP_DFLT_NUM_DESCS * 2,
 930                 .comp_size = ETH_AF_XDP_DFLT_NUM_DESCS,
 931                 .flags = XDP_UMEM_UNALIGNED_CHUNK_FLAG};
 932         void *base_addr = NULL;
 933         struct rte_mempool *mb_pool = rxq->mb_pool;
 934         uint64_t umem_size, align = 0;
 935
 936         if (internals->shared_umem) {
 937                 if (get_shared_umem(rxq, internals->if_name, &umem) < 0)
 938                         return NULL;
 939
 940                 if (umem != NULL &&
 941                         __atomic_load_n(&umem->refcnt, __ATOMIC_ACQUIRE) <
 942                                         umem->max_xsks) {
 943                         AF_XDP_LOG(INFO, "%s,qid%i sharing UMEM\n",
 944                                         internals->if_name, rxq->xsk_queue_idx);
 945                         __atomic_fetch_add(&umem->refcnt, 1, __ATOMIC_ACQUIRE);
 946                 }
 947         }
 948
 949         if (umem == NULL) {
 950                 usr_config.frame_size =
 951                         rte_mempool_calc_obj_size(mb_pool->elt_size,
 952                                                   mb_pool->flags, NULL);
 953                 usr_config.frame_headroom = mb_pool->header_size +
 954                                                 sizeof(struct rte_mbuf) +
 955                                                 rte_pktmbuf_priv_size(mb_pool) +
 956                                                 RTE_PKTMBUF_HEADROOM;
 957
 958                 umem = rte_zmalloc_socket("umem", sizeof(*umem), 0,
 959                                           rte_socket_id());
 960                 if (umem == NULL) {
 961                         AF_XDP_LOG(ERR, "Failed to allocate umem info");
 962                         return NULL;
 963                 }
 964
 965                 umem->mb_pool = mb_pool;
 966                 base_addr = (void *)get_base_addr(mb_pool, &align);
 967                 umem_size = (uint64_t)mb_pool->populated_size *
 968                                 (uint64_t)usr_config.frame_size +
 969                                 align;
 970
 971                 ret = xsk_umem__create(&umem->umem, base_addr, umem_size,
 972                                 &rxq->fq, &rxq->cq, &usr_config);
 973                 if (ret) {
 974                         AF_XDP_LOG(ERR, "Failed to create umem");
 975                         goto err;
 976                 }
 977                 umem->buffer = base_addr;
 978
 979                 if (internals->shared_umem) {
 980                         umem->max_xsks = mb_pool->populated_size /
 981                                                 ETH_AF_XDP_NUM_BUFFERS;
 982                         AF_XDP_LOG(INFO, "Max xsks for UMEM %s: %u\n",
 983                                                 mb_pool->name, umem->max_xsks);
 984                 }
 985
 986                 __atomic_store_n(&umem->refcnt, 1, __ATOMIC_RELEASE);
 987         }
 988
 989 #else
 990 static struct
 991 xsk_umem_info *xdp_umem_configure(struct pmd_internals *internals,
 992                                   struct pkt_rx_queue *rxq)
 993 {
 994         struct xsk_umem_info *umem;
 995         const struct rte_memzone *mz;
 996         struct xsk_umem_config usr_config = {
 997                 .fill_size = ETH_AF_XDP_DFLT_NUM_DESCS,
 998                 .comp_size = ETH_AF_XDP_DFLT_NUM_DESCS,
 999                 .frame_size = ETH_AF_XDP_FRAME_SIZE,
1000                 .frame_headroom = 0 };
1001         char ring_name[RTE_RING_NAMESIZE];
1002         char mz_name[RTE_MEMZONE_NAMESIZE];
1003         int ret;
1004         uint64_t i;
1005
1006         umem = rte_zmalloc_socket("umem", sizeof(*umem), 0, rte_socket_id());
1007         if (umem == NULL) {
1008                 AF_XDP_LOG(ERR, "Failed to allocate umem info");
1009                 return NULL;
1010         }
1011
1012         snprintf(ring_name, sizeof(ring_name), "af_xdp_ring_%s_%u",
1013                        internals->if_name, rxq->xsk_queue_idx);
1014         umem->buf_ring = rte_ring_create(ring_name,
1015                                          ETH_AF_XDP_NUM_BUFFERS,
1016                                          rte_socket_id(),
1017                                          0x0);
1018         if (umem->buf_ring == NULL) {
1019                 AF_XDP_LOG(ERR, "Failed to create rte_ring\n");
1020                 goto err;
1021         }
1022
1023         for (i = 0; i < ETH_AF_XDP_NUM_BUFFERS; i++)
1024                 rte_ring_enqueue(umem->buf_ring,
1025                                  (void *)(i * ETH_AF_XDP_FRAME_SIZE));
1026
1027         snprintf(mz_name, sizeof(mz_name), "af_xdp_umem_%s_%u",
1028                        internals->if_name, rxq->xsk_queue_idx);
1029         mz = rte_memzone_reserve_aligned(mz_name,
1030                         ETH_AF_XDP_NUM_BUFFERS * ETH_AF_XDP_FRAME_SIZE,
1031                         rte_socket_id(), RTE_MEMZONE_IOVA_CONTIG,
1032                         getpagesize());
1033         if (mz == NULL) {
1034                 AF_XDP_LOG(ERR, "Failed to reserve memzone for af_xdp umem.\n");
1035                 goto err;
1036         }
1037
1038         ret = xsk_umem__create(&umem->umem, mz->addr,
1039                                ETH_AF_XDP_NUM_BUFFERS * ETH_AF_XDP_FRAME_SIZE,
1040                                &rxq->fq, &rxq->cq,
1041                                &usr_config);
1042
1043         if (ret) {
1044                 AF_XDP_LOG(ERR, "Failed to create umem");
1045                 goto err;
1046         }
1047         umem->mz = mz;
1048
1049 #endif
1050         return umem;
1051
1052 err:
1053         xdp_umem_destroy(umem);
1054         return NULL;
1055 }
1056
1057 static int
1058 load_custom_xdp_prog(const char *prog_path, int if_index)
1059 {
1060         int ret, prog_fd = -1;
1061         struct bpf_object *obj;
1062         struct bpf_map *map;
1063
1064         ret = bpf_prog_load(prog_path, BPF_PROG_TYPE_XDP, &obj, &prog_fd);
1065         if (ret) {
1066                 AF_XDP_LOG(ERR, "Failed to load program %s\n", prog_path);
1067                 return ret;
1068         }
1069
1070         /*
1071          * The loaded program must provision for a map of xsks, such that some
1072          * traffic can be redirected to userspace. When the xsk is created,
1073          * libbpf inserts it into the map.
1074          */
1075         map = bpf_object__find_map_by_name(obj, "xsks_map");
1076         if (!map) {
1077                 AF_XDP_LOG(ERR, "Failed to find xsks_map in %s\n", prog_path);
1078                 return -1;
1079         }
1080
1081         /* Link the program with the given network device */
1082         ret = bpf_set_link_xdp_fd(if_index, prog_fd,
1083                                         XDP_FLAGS_UPDATE_IF_NOEXIST);
1084         if (ret) {
1085                 AF_XDP_LOG(ERR, "Failed to set prog fd %d on interface\n",
1086                                 prog_fd);
1087                 return -1;
1088         }
1089
1090         AF_XDP_LOG(INFO, "Successfully loaded XDP program %s with fd %d\n",
1091                                 prog_path, prog_fd);
1092
1093         return 0;
1094 }
1095
1096 static int
1097 xsk_configure(struct pmd_internals *internals, struct pkt_rx_queue *rxq,
1098               int ring_size)
1099 {
1100         struct xsk_socket_config cfg;
1101         struct pkt_tx_queue *txq = rxq->pair;
1102         int ret = 0;
1103         int reserve_size = ETH_AF_XDP_DFLT_NUM_DESCS;
1104         struct rte_mbuf *fq_bufs[reserve_size];
1105
1106         rxq->umem = xdp_umem_configure(internals, rxq);
1107         if (rxq->umem == NULL)
1108                 return -ENOMEM;
1109         txq->umem = rxq->umem;
1110
1111         cfg.rx_size = ring_size;
1112         cfg.tx_size = ring_size;
1113         cfg.libbpf_flags = 0;
1114         cfg.xdp_flags = XDP_FLAGS_UPDATE_IF_NOEXIST;
1115         cfg.bind_flags = 0;
1116
1117 #if defined(XDP_USE_NEED_WAKEUP)
1118         cfg.bind_flags |= XDP_USE_NEED_WAKEUP;
1119 #endif
1120
1121         if (strnlen(internals->prog_path, PATH_MAX) &&
1122                                 !internals->custom_prog_configured) {
1123                 ret = load_custom_xdp_prog(internals->prog_path,
1124                                            internals->if_index);
1125                 if (ret) {
1126                         AF_XDP_LOG(ERR, "Failed to load custom XDP program %s\n",
1127                                         internals->prog_path);
1128                         goto err;
1129                 }
1130                 internals->custom_prog_configured = 1;
1131         }
1132
1133         if (internals->shared_umem)
1134                 ret = create_shared_socket(&rxq->xsk, internals->if_name,
1135                                 rxq->xsk_queue_idx, rxq->umem->umem, &rxq->rx,
1136                                 &txq->tx, &rxq->fq, &rxq->cq, &cfg);
1137         else
1138                 ret = xsk_socket__create(&rxq->xsk, internals->if_name,
1139                                 rxq->xsk_queue_idx, rxq->umem->umem, &rxq->rx,
1140                                 &txq->tx, &cfg);
1141
1142         if (ret) {
1143                 AF_XDP_LOG(ERR, "Failed to create xsk socket.\n");
1144                 goto err;
1145         }
1146
1147 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
1148         ret = rte_pktmbuf_alloc_bulk(rxq->umem->mb_pool, fq_bufs, reserve_size);
1149         if (ret) {
1150                 AF_XDP_LOG(DEBUG, "Failed to get enough buffers for fq.\n");
1151                 goto err;
1152         }
1153 #endif
1154         ret = reserve_fill_queue(rxq->umem, reserve_size, fq_bufs, &rxq->fq);
1155         if (ret) {
1156                 xsk_socket__delete(rxq->xsk);
1157                 AF_XDP_LOG(ERR, "Failed to reserve fill queue.\n");
1158                 goto err;
1159         }
1160
1161         return 0;
1162
1163 err:
1164         if (__atomic_sub_fetch(&rxq->umem->refcnt, 1, __ATOMIC_ACQUIRE) == 0)
1165                 xdp_umem_destroy(rxq->umem);
1166
1167         return ret;
1168 }
1169
1170 static int
1171 eth_rx_queue_setup(struct rte_eth_dev *dev,
1172                    uint16_t rx_queue_id,
1173                    uint16_t nb_rx_desc,
1174                    unsigned int socket_id __rte_unused,
1175                    const struct rte_eth_rxconf *rx_conf __rte_unused,
1176                    struct rte_mempool *mb_pool)
1177 {
1178         struct pmd_internals *internals = dev->data->dev_private;
1179         struct pkt_rx_queue *rxq;
1180         int ret;
1181
1182         rxq = &internals->rx_queues[rx_queue_id];
1183
1184         AF_XDP_LOG(INFO, "Set up rx queue, rx queue id: %d, xsk queue id: %d\n",
1185                    rx_queue_id, rxq->xsk_queue_idx);
1186
1187 #ifndef XDP_UMEM_UNALIGNED_CHUNK_FLAG
1188         uint32_t buf_size, data_size;
1189
1190         /* Now get the space available for data in the mbuf */
1191         buf_size = rte_pktmbuf_data_room_size(mb_pool) -
1192                 RTE_PKTMBUF_HEADROOM;
1193         data_size = ETH_AF_XDP_FRAME_SIZE;
1194
1195         if (data_size > buf_size) {
1196                 AF_XDP_LOG(ERR, "%s: %d bytes will not fit in mbuf (%d bytes)\n",
1197                         dev->device->name, data_size, buf_size);
1198                 ret = -ENOMEM;
1199                 goto err;
1200         }
1201 #endif
1202
1203         rxq->mb_pool = mb_pool;
1204
1205         if (xsk_configure(internals, rxq, nb_rx_desc)) {
1206                 AF_XDP_LOG(ERR, "Failed to configure xdp socket\n");
1207                 ret = -EINVAL;
1208                 goto err;
1209         }
1210
1211         rxq->fds[0].fd = xsk_socket__fd(rxq->xsk);
1212         rxq->fds[0].events = POLLIN;
1213
1214         dev->data->rx_queues[rx_queue_id] = rxq;
1215         return 0;
1216
1217 err:
1218         return ret;
1219 }
1220
1221 static int
1222 eth_tx_queue_setup(struct rte_eth_dev *dev,
1223                    uint16_t tx_queue_id,
1224                    uint16_t nb_tx_desc __rte_unused,
1225                    unsigned int socket_id __rte_unused,
1226                    const struct rte_eth_txconf *tx_conf __rte_unused)
1227 {
1228         struct pmd_internals *internals = dev->data->dev_private;
1229         struct pkt_tx_queue *txq;
1230
1231         txq = &internals->tx_queues[tx_queue_id];
1232
1233         dev->data->tx_queues[tx_queue_id] = txq;
1234         return 0;
1235 }
1236
1237 static int
1238 eth_dev_mtu_set(struct rte_eth_dev *dev, uint16_t mtu)
1239 {
1240         struct pmd_internals *internals = dev->data->dev_private;
1241         struct ifreq ifr = { .ifr_mtu = mtu };
1242         int ret;
1243         int s;
1244
1245         s = socket(PF_INET, SOCK_DGRAM, 0);
1246         if (s < 0)
1247                 return -EINVAL;
1248
1249         strlcpy(ifr.ifr_name, internals->if_name, IFNAMSIZ);
1250         ret = ioctl(s, SIOCSIFMTU, &ifr);
1251         close(s);
1252
1253         return (ret < 0) ? -errno : 0;
1254 }
1255
1256 static int
1257 eth_dev_change_flags(char *if_name, uint32_t flags, uint32_t mask)
1258 {
1259         struct ifreq ifr;
1260         int ret = 0;
1261         int s;
1262
1263         s = socket(PF_INET, SOCK_DGRAM, 0);
1264         if (s < 0)
1265                 return -errno;
1266
1267         strlcpy(ifr.ifr_name, if_name, IFNAMSIZ);
1268         if (ioctl(s, SIOCGIFFLAGS, &ifr) < 0) {
1269                 ret = -errno;
1270                 goto out;
1271         }
1272         ifr.ifr_flags &= mask;
1273         ifr.ifr_flags |= flags;
1274         if (ioctl(s, SIOCSIFFLAGS, &ifr) < 0) {
1275                 ret = -errno;
1276                 goto out;
1277         }
1278 out:
1279         close(s);
1280         return ret;
1281 }
1282
1283 static int
1284 eth_dev_promiscuous_enable(struct rte_eth_dev *dev)
1285 {
1286         struct pmd_internals *internals = dev->data->dev_private;
1287
1288         return eth_dev_change_flags(internals->if_name, IFF_PROMISC, ~0);
1289 }
1290
1291 static int
1292 eth_dev_promiscuous_disable(struct rte_eth_dev *dev)
1293 {
1294         struct pmd_internals *internals = dev->data->dev_private;
1295
1296         return eth_dev_change_flags(internals->if_name, 0, ~IFF_PROMISC);
1297 }
1298
1299 static const struct eth_dev_ops ops = {
1300         .dev_start = eth_dev_start,
1301         .dev_stop = eth_dev_stop,
1302         .dev_close = eth_dev_close,
1303         .dev_configure = eth_dev_configure,
1304         .dev_infos_get = eth_dev_info,
1305         .mtu_set = eth_dev_mtu_set,
1306         .promiscuous_enable = eth_dev_promiscuous_enable,
1307         .promiscuous_disable = eth_dev_promiscuous_disable,
1308         .rx_queue_setup = eth_rx_queue_setup,
1309         .tx_queue_setup = eth_tx_queue_setup,
1310         .rx_queue_release = eth_queue_release,
1311         .tx_queue_release = eth_queue_release,
1312         .link_update = eth_link_update,
1313         .stats_get = eth_stats_get,
1314         .stats_reset = eth_stats_reset,
1315 };
1316
1317 /** parse integer from integer argument */
1318 static int
1319 parse_integer_arg(const char *key __rte_unused,
1320                   const char *value, void *extra_args)
1321 {
1322         int *i = (int *)extra_args;
1323         char *end;
1324
1325         *i = strtol(value, &end, 10);
1326         if (*i < 0) {
1327                 AF_XDP_LOG(ERR, "Argument has to be positive.\n");
1328                 return -EINVAL;
1329         }
1330
1331         return 0;
1332 }
1333
1334 /** parse name argument */
1335 static int
1336 parse_name_arg(const char *key __rte_unused,
1337                const char *value, void *extra_args)
1338 {
1339         char *name = extra_args;
1340
1341         if (strnlen(value, IFNAMSIZ) > IFNAMSIZ - 1) {
1342                 AF_XDP_LOG(ERR, "Invalid name %s, should be less than %u bytes.\n",
1343                            value, IFNAMSIZ);
1344                 return -EINVAL;
1345         }
1346
1347         strlcpy(name, value, IFNAMSIZ);
1348
1349         return 0;
1350 }
1351
1352 /** parse xdp prog argument */
1353 static int
1354 parse_prog_arg(const char *key __rte_unused,
1355                const char *value, void *extra_args)
1356 {
1357         char *path = extra_args;
1358
1359         if (strnlen(value, PATH_MAX) == PATH_MAX) {
1360                 AF_XDP_LOG(ERR, "Invalid path %s, should be less than %u bytes.\n",
1361                            value, PATH_MAX);
1362                 return -EINVAL;
1363         }
1364
1365         if (access(value, F_OK) != 0) {
1366                 AF_XDP_LOG(ERR, "Error accessing %s: %s\n",
1367                            value, strerror(errno));
1368                 return -EINVAL;
1369         }
1370
1371         strlcpy(path, value, PATH_MAX);
1372
1373         return 0;
1374 }
1375
1376 static int
1377 xdp_get_channels_info(const char *if_name, int *max_queues,
1378                                 int *combined_queues)
1379 {
1380         struct ethtool_channels channels;
1381         struct ifreq ifr;
1382         int fd, ret;
1383
1384         fd = socket(AF_INET, SOCK_DGRAM, 0);
1385         if (fd < 0)
1386                 return -1;
1387
1388         channels.cmd = ETHTOOL_GCHANNELS;
1389         ifr.ifr_data = (void *)&channels;
1390         strlcpy(ifr.ifr_name, if_name, IFNAMSIZ);
1391         ret = ioctl(fd, SIOCETHTOOL, &ifr);
1392         if (ret) {
1393                 if (errno == EOPNOTSUPP) {
1394                         ret = 0;
1395                 } else {
1396                         ret = -errno;
1397                         goto out;
1398                 }
1399         }
1400
1401         if (channels.max_combined == 0 || errno == EOPNOTSUPP) {
1402                 /* If the device says it has no channels, then all traffic
1403                  * is sent to a single stream, so max queues = 1.
1404                  */
1405                 *max_queues = 1;
1406                 *combined_queues = 1;
1407         } else {
1408                 *max_queues = channels.max_combined;
1409                 *combined_queues = channels.combined_count;
1410         }
1411
1412  out:
1413         close(fd);
1414         return ret;
1415 }
1416
1417 static int
1418 parse_parameters(struct rte_kvargs *kvlist, char *if_name, int *start_queue,
1419                         int *queue_cnt, int *shared_umem, char *prog_path)
1420 {
1421         int ret;
1422
1423         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_IFACE_ARG,
1424                                  &parse_name_arg, if_name);
1425         if (ret < 0)
1426                 goto free_kvlist;
1427
1428         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_START_QUEUE_ARG,
1429                                  &parse_integer_arg, start_queue);
1430         if (ret < 0)
1431                 goto free_kvlist;
1432
1433         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_QUEUE_COUNT_ARG,
1434                                  &parse_integer_arg, queue_cnt);
1435         if (ret < 0 || *queue_cnt <= 0) {
1436                 ret = -EINVAL;
1437                 goto free_kvlist;
1438         }
1439
1440         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_SHARED_UMEM_ARG,
1441                                 &parse_integer_arg, shared_umem);
1442         if (ret < 0)
1443                 goto free_kvlist;
1444
1445         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_PROG_ARG,
1446                                  &parse_prog_arg, prog_path);
1447         if (ret < 0)
1448                 goto free_kvlist;
1449
1450 free_kvlist:
1451         rte_kvargs_free(kvlist);
1452         return ret;
1453 }
1454
1455 static int
1456 get_iface_info(const char *if_name,
1457                struct rte_ether_addr *eth_addr,
1458                int *if_index)
1459 {
1460         struct ifreq ifr;
1461         int sock = socket(AF_INET, SOCK_DGRAM, IPPROTO_IP);
1462
1463         if (sock < 0)
1464                 return -1;
1465
1466         strlcpy(ifr.ifr_name, if_name, IFNAMSIZ);
1467         if (ioctl(sock, SIOCGIFINDEX, &ifr))
1468                 goto error;
1469
1470         *if_index = ifr.ifr_ifindex;
1471
1472         if (ioctl(sock, SIOCGIFHWADDR, &ifr))
1473                 goto error;
1474
1475         rte_memcpy(eth_addr, ifr.ifr_hwaddr.sa_data, RTE_ETHER_ADDR_LEN);
1476
1477         close(sock);
1478         return 0;
1479
1480 error:
1481         close(sock);
1482         return -1;
1483 }
1484
1485 static struct rte_eth_dev *
1486 init_internals(struct rte_vdev_device *dev, const char *if_name,
1487                 int start_queue_idx, int queue_cnt, int shared_umem,
1488                 const char *prog_path)
1489 {
1490         const char *name = rte_vdev_device_name(dev);
1491         const unsigned int numa_node = dev->device.numa_node;
1492         struct pmd_internals *internals;
1493         struct rte_eth_dev *eth_dev;
1494         int ret;
1495         int i;
1496
1497         internals = rte_zmalloc_socket(name, sizeof(*internals), 0, numa_node);
1498         if (internals == NULL)
1499                 return NULL;
1500
1501         internals->start_queue_idx = start_queue_idx;
1502         internals->queue_cnt = queue_cnt;
1503         strlcpy(internals->if_name, if_name, IFNAMSIZ);
1504         strlcpy(internals->prog_path, prog_path, PATH_MAX);
1505         internals->custom_prog_configured = 0;
1506
1507 #ifndef ETH_AF_XDP_SHARED_UMEM
1508         if (shared_umem) {
1509                 AF_XDP_LOG(ERR, "Shared UMEM feature not available. "
1510                                 "Check kernel and libbpf version\n");
1511                 goto err_free_internals;
1512         }
1513 #endif
1514         internals->shared_umem = shared_umem;
1515
1516         if (xdp_get_channels_info(if_name, &internals->max_queue_cnt,
1517                                   &internals->combined_queue_cnt)) {
1518                 AF_XDP_LOG(ERR, "Failed to get channel info of interface: %s\n",
1519                                 if_name);
1520                 goto err_free_internals;
1521         }
1522
1523         if (queue_cnt > internals->combined_queue_cnt) {
1524                 AF_XDP_LOG(ERR, "Specified queue count %d is larger than combined queue count %d.\n",
1525                                 queue_cnt, internals->combined_queue_cnt);
1526                 goto err_free_internals;
1527         }
1528
1529         internals->rx_queues = rte_zmalloc_socket(NULL,
1530                                         sizeof(struct pkt_rx_queue) * queue_cnt,
1531                                         0, numa_node);
1532         if (internals->rx_queues == NULL) {
1533                 AF_XDP_LOG(ERR, "Failed to allocate memory for rx queues.\n");
1534                 goto err_free_internals;
1535         }
1536
1537         internals->tx_queues = rte_zmalloc_socket(NULL,
1538                                         sizeof(struct pkt_tx_queue) * queue_cnt,
1539                                         0, numa_node);
1540         if (internals->tx_queues == NULL) {
1541                 AF_XDP_LOG(ERR, "Failed to allocate memory for tx queues.\n");
1542                 goto err_free_rx;
1543         }
1544         for (i = 0; i < queue_cnt; i++) {
1545                 internals->tx_queues[i].pair = &internals->rx_queues[i];
1546                 internals->rx_queues[i].pair = &internals->tx_queues[i];
1547                 internals->rx_queues[i].xsk_queue_idx = start_queue_idx + i;
1548                 internals->tx_queues[i].xsk_queue_idx = start_queue_idx + i;
1549         }
1550
1551         ret = get_iface_info(if_name, &internals->eth_addr,
1552                              &internals->if_index);
1553         if (ret)
1554                 goto err_free_tx;
1555
1556         eth_dev = rte_eth_vdev_allocate(dev, 0);
1557         if (eth_dev == NULL)
1558                 goto err_free_tx;
1559
1560         eth_dev->data->dev_private = internals;
1561         eth_dev->data->dev_link = pmd_link;
1562         eth_dev->data->mac_addrs = &internals->eth_addr;
1563         eth_dev->data->dev_flags |= RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS;
1564         eth_dev->dev_ops = &ops;
1565         eth_dev->rx_pkt_burst = eth_af_xdp_rx;
1566         eth_dev->tx_pkt_burst = eth_af_xdp_tx;
1567
1568 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
1569         AF_XDP_LOG(INFO, "Zero copy between umem and mbuf enabled.\n");
1570 #endif
1571
1572         return eth_dev;
1573
1574 err_free_tx:
1575         rte_free(internals->tx_queues);
1576 err_free_rx:
1577         rte_free(internals->rx_queues);
1578 err_free_internals:
1579         rte_free(internals);
1580         return NULL;
1581 }
1582
1583 static int
1584 rte_pmd_af_xdp_probe(struct rte_vdev_device *dev)
1585 {
1586         struct rte_kvargs *kvlist;
1587         char if_name[IFNAMSIZ] = {'\0'};
1588         int xsk_start_queue_idx = ETH_AF_XDP_DFLT_START_QUEUE_IDX;
1589         int xsk_queue_cnt = ETH_AF_XDP_DFLT_QUEUE_COUNT;
1590         int shared_umem = 0;
1591         char prog_path[PATH_MAX] = {'\0'};
1592         struct rte_eth_dev *eth_dev = NULL;
1593         const char *name;
1594
1595         AF_XDP_LOG(INFO, "Initializing pmd_af_xdp for %s\n",
1596                 rte_vdev_device_name(dev));
1597
1598         name = rte_vdev_device_name(dev);
1599         if (rte_eal_process_type() == RTE_PROC_SECONDARY &&
1600                 strlen(rte_vdev_device_args(dev)) == 0) {
1601                 eth_dev = rte_eth_dev_attach_secondary(name);
1602                 if (eth_dev == NULL) {
1603                         AF_XDP_LOG(ERR, "Failed to probe %s\n", name);
1604                         return -EINVAL;
1605                 }
1606                 eth_dev->dev_ops = &ops;
1607                 rte_eth_dev_probing_finish(eth_dev);
1608                 return 0;
1609         }
1610
1611         kvlist = rte_kvargs_parse(rte_vdev_device_args(dev), valid_arguments);
1612         if (kvlist == NULL) {
1613                 AF_XDP_LOG(ERR, "Invalid kvargs key\n");
1614                 return -EINVAL;
1615         }
1616
1617         if (dev->device.numa_node == SOCKET_ID_ANY)
1618                 dev->device.numa_node = rte_socket_id();
1619
1620         if (parse_parameters(kvlist, if_name, &xsk_start_queue_idx,
1621                              &xsk_queue_cnt, &shared_umem, prog_path) < 0) {
1622                 AF_XDP_LOG(ERR, "Invalid kvargs value\n");
1623                 return -EINVAL;
1624         }
1625
1626         if (strlen(if_name) == 0) {
1627                 AF_XDP_LOG(ERR, "Network interface must be specified\n");
1628                 return -EINVAL;
1629         }
1630
1631         eth_dev = init_internals(dev, if_name, xsk_start_queue_idx,
1632                                         xsk_queue_cnt, shared_umem, prog_path);
1633         if (eth_dev == NULL) {
1634                 AF_XDP_LOG(ERR, "Failed to init internals\n");
1635                 return -1;
1636         }
1637
1638         rte_eth_dev_probing_finish(eth_dev);
1639
1640         return 0;
1641 }
1642
1643 static int
1644 rte_pmd_af_xdp_remove(struct rte_vdev_device *dev)
1645 {
1646         struct rte_eth_dev *eth_dev = NULL;
1647
1648         AF_XDP_LOG(INFO, "Removing AF_XDP ethdev on numa socket %u\n",
1649                 rte_socket_id());
1650
1651         if (dev == NULL)
1652                 return -1;
1653
1654         /* find the ethdev entry */
1655         eth_dev = rte_eth_dev_allocated(rte_vdev_device_name(dev));
1656         if (eth_dev == NULL)
1657                 return 0;
1658
1659         eth_dev_close(eth_dev);
1660         rte_eth_dev_release_port(eth_dev);
1661
1662
1663         return 0;
1664 }
1665
1666 static struct rte_vdev_driver pmd_af_xdp_drv = {
1667         .probe = rte_pmd_af_xdp_probe,
1668         .remove = rte_pmd_af_xdp_remove,
1669 };
1670
1671 RTE_PMD_REGISTER_VDEV(net_af_xdp, pmd_af_xdp_drv);
1672 RTE_PMD_REGISTER_PARAM_STRING(net_af_xdp,
1673                               "iface=<string> "
1674                               "start_queue=<int> "
1675                               "queue_count=<int> "
1676                               "shared_umem=<int> "
1677                               "xdp_prog=<string> ");