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 int
 603 eth_dev_stop(struct rte_eth_dev *dev)
 604 {
 605         dev->data->dev_link.link_status = ETH_LINK_DOWN;
 606         return 0;
 607 }
 608
 609 /* Find ethdev in list */
 610 static inline struct internal_list *
 611 find_internal_resource(struct pmd_internals *port_int)
 612 {
 613         int found = 0;
 614         struct internal_list *list = NULL;
 615
 616         if (port_int == NULL)
 617                 return NULL;
 618
 619         pthread_mutex_lock(&internal_list_lock);
 620
 621         TAILQ_FOREACH(list, &internal_list, next) {
 622                 struct pmd_internals *list_int =
 623                                 list->eth_dev->data->dev_private;
 624                 if (list_int == port_int) {
 625                         found = 1;
 626                         break;
 627                 }
 628         }
 629
 630         pthread_mutex_unlock(&internal_list_lock);
 631
 632         if (!found)
 633                 return NULL;
 634
 635         return list;
 636 }
 637
 638 /* Check if the netdev,qid context already exists */
 639 static inline bool
 640 ctx_exists(struct pkt_rx_queue *rxq, const char *ifname,
 641                 struct pkt_rx_queue *list_rxq, const char *list_ifname)
 642 {
 643         bool exists = false;
 644
 645         if (rxq->xsk_queue_idx == list_rxq->xsk_queue_idx &&
 646                         !strncmp(ifname, list_ifname, IFNAMSIZ)) {
 647                 AF_XDP_LOG(ERR, "ctx %s,%i already exists, cannot share umem\n",
 648                                         ifname, rxq->xsk_queue_idx);
 649                 exists = true;
 650         }
 651
 652         return exists;
 653 }
 654
 655 /* Get a pointer to an existing UMEM which overlays the rxq's mb_pool */
 656 static inline int
 657 get_shared_umem(struct pkt_rx_queue *rxq, const char *ifname,
 658                         struct xsk_umem_info **umem)
 659 {
 660         struct internal_list *list;
 661         struct pmd_internals *internals;
 662         int i = 0, ret = 0;
 663         struct rte_mempool *mb_pool = rxq->mb_pool;
 664
 665         if (mb_pool == NULL)
 666                 return ret;
 667
 668         pthread_mutex_lock(&internal_list_lock);
 669
 670         TAILQ_FOREACH(list, &internal_list, next) {
 671                 internals = list->eth_dev->data->dev_private;
 672                 for (i = 0; i < internals->queue_cnt; i++) {
 673                         struct pkt_rx_queue *list_rxq =
 674                                                 &internals->rx_queues[i];
 675                         if (rxq == list_rxq)
 676                                 continue;
 677                         if (mb_pool == internals->rx_queues[i].mb_pool) {
 678                                 if (ctx_exists(rxq, ifname, list_rxq,
 679                                                 internals->if_name)) {
 680                                         ret = -1;
 681                                         goto out;
 682                                 }
 683                                 if (__atomic_load_n(
 684                                         &internals->rx_queues[i].umem->refcnt,
 685                                                         __ATOMIC_ACQUIRE)) {
 686                                         *umem = internals->rx_queues[i].umem;
 687                                         goto out;
 688                                 }
 689                         }
 690                 }
 691         }
 692
 693 out:
 694         pthread_mutex_unlock(&internal_list_lock);
 695
 696         return ret;
 697 }
 698
 699 static int
 700 eth_dev_configure(struct rte_eth_dev *dev)
 701 {
 702         struct pmd_internals *internal = dev->data->dev_private;
 703
 704         /* rx/tx must be paired */
 705         if (dev->data->nb_rx_queues != dev->data->nb_tx_queues)
 706                 return -EINVAL;
 707
 708         if (internal->shared_umem) {
 709                 struct internal_list *list = NULL;
 710                 const char *name = dev->device->name;
 711
 712                 /* Ensure PMD is not already inserted into the list */
 713                 list = find_internal_resource(internal);
 714                 if (list)
 715                         return 0;
 716
 717                 list = rte_zmalloc_socket(name, sizeof(*list), 0,
 718                                         dev->device->numa_node);
 719                 if (list == NULL)
 720                         return -1;
 721
 722                 list->eth_dev = dev;
 723                 pthread_mutex_lock(&internal_list_lock);
 724                 TAILQ_INSERT_TAIL(&internal_list, list, next);
 725                 pthread_mutex_unlock(&internal_list_lock);
 726         }
 727
 728         return 0;
 729 }
 730
 731 static int
 732 eth_dev_info(struct rte_eth_dev *dev, struct rte_eth_dev_info *dev_info)
 733 {
 734         struct pmd_internals *internals = dev->data->dev_private;
 735
 736         dev_info->if_index = internals->if_index;
 737         dev_info->max_mac_addrs = 1;
 738         dev_info->max_rx_pktlen = ETH_FRAME_LEN;
 739         dev_info->max_rx_queues = internals->queue_cnt;
 740         dev_info->max_tx_queues = internals->queue_cnt;
 741
 742         dev_info->min_mtu = RTE_ETHER_MIN_MTU;
 743 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 744         dev_info->max_mtu = getpagesize() -
 745                                 sizeof(struct rte_mempool_objhdr) -
 746                                 sizeof(struct rte_mbuf) -
 747                                 RTE_PKTMBUF_HEADROOM - XDP_PACKET_HEADROOM;
 748 #else
 749         dev_info->max_mtu = ETH_AF_XDP_FRAME_SIZE - XDP_PACKET_HEADROOM;
 750 #endif
 751
 752         dev_info->default_rxportconf.nb_queues = 1;
 753         dev_info->default_txportconf.nb_queues = 1;
 754         dev_info->default_rxportconf.ring_size = ETH_AF_XDP_DFLT_NUM_DESCS;
 755         dev_info->default_txportconf.ring_size = ETH_AF_XDP_DFLT_NUM_DESCS;
 756
 757         return 0;
 758 }
 759
 760 static int
 761 eth_stats_get(struct rte_eth_dev *dev, struct rte_eth_stats *stats)
 762 {
 763         struct pmd_internals *internals = dev->data->dev_private;
 764         struct xdp_statistics xdp_stats;
 765         struct pkt_rx_queue *rxq;
 766         struct pkt_tx_queue *txq;
 767         socklen_t optlen;
 768         int i, ret;
 769
 770         for (i = 0; i < dev->data->nb_rx_queues; i++) {
 771                 optlen = sizeof(struct xdp_statistics);
 772                 rxq = &internals->rx_queues[i];
 773                 txq = rxq->pair;
 774                 stats->q_ipackets[i] = rxq->stats.rx_pkts;
 775                 stats->q_ibytes[i] = rxq->stats.rx_bytes;
 776
 777                 stats->q_opackets[i] = txq->stats.tx_pkts;
 778                 stats->q_obytes[i] = txq->stats.tx_bytes;
 779
 780                 stats->ipackets += stats->q_ipackets[i];
 781                 stats->ibytes += stats->q_ibytes[i];
 782                 stats->imissed += rxq->stats.rx_dropped;
 783                 stats->oerrors += txq->stats.tx_dropped;
 784                 ret = getsockopt(xsk_socket__fd(rxq->xsk), SOL_XDP,
 785                                 XDP_STATISTICS, &xdp_stats, &optlen);
 786                 if (ret != 0) {
 787                         AF_XDP_LOG(ERR, "getsockopt() failed for XDP_STATISTICS.\n");
 788                         return -1;
 789                 }
 790                 stats->imissed += xdp_stats.rx_dropped;
 791
 792                 stats->opackets += stats->q_opackets[i];
 793                 stats->obytes += stats->q_obytes[i];
 794         }
 795
 796         return 0;
 797 }
 798
 799 static int
 800 eth_stats_reset(struct rte_eth_dev *dev)
 801 {
 802         struct pmd_internals *internals = dev->data->dev_private;
 803         int i;
 804
 805         for (i = 0; i < internals->queue_cnt; i++) {
 806                 memset(&internals->rx_queues[i].stats, 0,
 807                                         sizeof(struct rx_stats));
 808                 memset(&internals->tx_queues[i].stats, 0,
 809                                         sizeof(struct tx_stats));
 810         }
 811
 812         return 0;
 813 }
 814
 815 static void
 816 remove_xdp_program(struct pmd_internals *internals)
 817 {
 818         uint32_t curr_prog_id = 0;
 819
 820         if (bpf_get_link_xdp_id(internals->if_index, &curr_prog_id,
 821                                 XDP_FLAGS_UPDATE_IF_NOEXIST)) {
 822                 AF_XDP_LOG(ERR, "bpf_get_link_xdp_id failed\n");
 823                 return;
 824         }
 825         bpf_set_link_xdp_fd(internals->if_index, -1,
 826                         XDP_FLAGS_UPDATE_IF_NOEXIST);
 827 }
 828
 829 static void
 830 xdp_umem_destroy(struct xsk_umem_info *umem)
 831 {
 832 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 833         umem->mb_pool = NULL;
 834 #else
 835         rte_memzone_free(umem->mz);
 836         umem->mz = NULL;
 837
 838         rte_ring_free(umem->buf_ring);
 839         umem->buf_ring = NULL;
 840 #endif
 841
 842         rte_free(umem);
 843         umem = NULL;
 844 }
 845
 846 static int
 847 eth_dev_close(struct rte_eth_dev *dev)
 848 {
 849         struct pmd_internals *internals = dev->data->dev_private;
 850         struct pkt_rx_queue *rxq;
 851         int i;
 852
 853         if (rte_eal_process_type() != RTE_PROC_PRIMARY)
 854                 return 0;
 855
 856         AF_XDP_LOG(INFO, "Closing AF_XDP ethdev on numa socket %u\n",
 857                 rte_socket_id());
 858
 859         for (i = 0; i < internals->queue_cnt; i++) {
 860                 rxq = &internals->rx_queues[i];
 861                 if (rxq->umem == NULL)
 862                         break;
 863                 xsk_socket__delete(rxq->xsk);
 864
 865                 if (__atomic_sub_fetch(&rxq->umem->refcnt, 1, __ATOMIC_ACQUIRE)
 866                                 == 0) {
 867                         (void)xsk_umem__delete(rxq->umem->umem);
 868                         xdp_umem_destroy(rxq->umem);
 869                 }
 870
 871                 /* free pkt_tx_queue */
 872                 rte_free(rxq->pair);
 873                 rte_free(rxq);
 874         }
 875
 876         /*
 877          * MAC is not allocated dynamically, setting it to NULL would prevent
 878          * from releasing it in rte_eth_dev_release_port.
 879          */
 880         dev->data->mac_addrs = NULL;
 881
 882         remove_xdp_program(internals);
 883
 884         if (internals->shared_umem) {
 885                 struct internal_list *list;
 886
 887                 /* Remove ethdev from list used to track and share UMEMs */
 888                 list = find_internal_resource(internals);
 889                 if (list) {
 890                         pthread_mutex_lock(&internal_list_lock);
 891                         TAILQ_REMOVE(&internal_list, list, next);
 892                         pthread_mutex_unlock(&internal_list_lock);
 893                         rte_free(list);
 894                 }
 895         }
 896
 897         return 0;
 898 }
 899
 900 static void
 901 eth_queue_release(void *q __rte_unused)
 902 {
 903 }
 904
 905 static int
 906 eth_link_update(struct rte_eth_dev *dev __rte_unused,
 907                 int wait_to_complete __rte_unused)
 908 {
 909         return 0;
 910 }
 911
 912 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
 913 static inline uint64_t get_base_addr(struct rte_mempool *mp, uint64_t *align)
 914 {
 915         struct rte_mempool_memhdr *memhdr;
 916         uint64_t memhdr_addr, aligned_addr;
 917
 918         memhdr = STAILQ_FIRST(&mp->mem_list);
 919         memhdr_addr = (uint64_t)memhdr->addr;
 920         aligned_addr = memhdr_addr & ~(getpagesize() - 1);
 921         *align = memhdr_addr - aligned_addr;
 922
 923         return aligned_addr;
 924 }
 925
 926 static struct
 927 xsk_umem_info *xdp_umem_configure(struct pmd_internals *internals,
 928                                   struct pkt_rx_queue *rxq)
 929 {
 930         struct xsk_umem_info *umem = NULL;
 931         int ret;
 932         struct xsk_umem_config usr_config = {
 933                 .fill_size = ETH_AF_XDP_DFLT_NUM_DESCS * 2,
 934                 .comp_size = ETH_AF_XDP_DFLT_NUM_DESCS,
 935                 .flags = XDP_UMEM_UNALIGNED_CHUNK_FLAG};
 936         void *base_addr = NULL;
 937         struct rte_mempool *mb_pool = rxq->mb_pool;
 938         uint64_t umem_size, align = 0;
 939
 940         if (internals->shared_umem) {
 941                 if (get_shared_umem(rxq, internals->if_name, &umem) < 0)
 942                         return NULL;
 943
 944                 if (umem != NULL &&
 945                         __atomic_load_n(&umem->refcnt, __ATOMIC_ACQUIRE) <
 946                                         umem->max_xsks) {
 947                         AF_XDP_LOG(INFO, "%s,qid%i sharing UMEM\n",
 948                                         internals->if_name, rxq->xsk_queue_idx);
 949                         __atomic_fetch_add(&umem->refcnt, 1, __ATOMIC_ACQUIRE);
 950                 }
 951         }
 952
 953         if (umem == NULL) {
 954                 usr_config.frame_size =
 955                         rte_mempool_calc_obj_size(mb_pool->elt_size,
 956                                                   mb_pool->flags, NULL);
 957                 usr_config.frame_headroom = mb_pool->header_size +
 958                                                 sizeof(struct rte_mbuf) +
 959                                                 rte_pktmbuf_priv_size(mb_pool) +
 960                                                 RTE_PKTMBUF_HEADROOM;
 961
 962                 umem = rte_zmalloc_socket("umem", sizeof(*umem), 0,
 963                                           rte_socket_id());
 964                 if (umem == NULL) {
 965                         AF_XDP_LOG(ERR, "Failed to allocate umem info");
 966                         return NULL;
 967                 }
 968
 969                 umem->mb_pool = mb_pool;
 970                 base_addr = (void *)get_base_addr(mb_pool, &align);
 971                 umem_size = mb_pool->populated_size * usr_config.frame_size +
 972                                 align;
 973
 974                 ret = xsk_umem__create(&umem->umem, base_addr, umem_size,
 975                                 &rxq->fq, &rxq->cq, &usr_config);
 976                 if (ret) {
 977                         AF_XDP_LOG(ERR, "Failed to create umem");
 978                         goto err;
 979                 }
 980                 umem->buffer = base_addr;
 981
 982                 if (internals->shared_umem) {
 983                         umem->max_xsks = mb_pool->populated_size /
 984                                                 ETH_AF_XDP_NUM_BUFFERS;
 985                         AF_XDP_LOG(INFO, "Max xsks for UMEM %s: %u\n",
 986                                                 mb_pool->name, umem->max_xsks);
 987                 }
 988
 989                 __atomic_store_n(&umem->refcnt, 1, __ATOMIC_RELEASE);
 990         }
 991
 992 #else
 993 static struct
 994 xsk_umem_info *xdp_umem_configure(struct pmd_internals *internals,
 995                                   struct pkt_rx_queue *rxq)
 996 {
 997         struct xsk_umem_info *umem;
 998         const struct rte_memzone *mz;
 999         struct xsk_umem_config usr_config = {
1000                 .fill_size = ETH_AF_XDP_DFLT_NUM_DESCS,
1001                 .comp_size = ETH_AF_XDP_DFLT_NUM_DESCS,
1002                 .frame_size = ETH_AF_XDP_FRAME_SIZE,
1003                 .frame_headroom = 0 };
1004         char ring_name[RTE_RING_NAMESIZE];
1005         char mz_name[RTE_MEMZONE_NAMESIZE];
1006         int ret;
1007         uint64_t i;
1008
1009         umem = rte_zmalloc_socket("umem", sizeof(*umem), 0, rte_socket_id());
1010         if (umem == NULL) {
1011                 AF_XDP_LOG(ERR, "Failed to allocate umem info");
1012                 return NULL;
1013         }
1014
1015         snprintf(ring_name, sizeof(ring_name), "af_xdp_ring_%s_%u",
1016                        internals->if_name, rxq->xsk_queue_idx);
1017         umem->buf_ring = rte_ring_create(ring_name,
1018                                          ETH_AF_XDP_NUM_BUFFERS,
1019                                          rte_socket_id(),
1020                                          0x0);
1021         if (umem->buf_ring == NULL) {
1022                 AF_XDP_LOG(ERR, "Failed to create rte_ring\n");
1023                 goto err;
1024         }
1025
1026         for (i = 0; i < ETH_AF_XDP_NUM_BUFFERS; i++)
1027                 rte_ring_enqueue(umem->buf_ring,
1028                                  (void *)(i * ETH_AF_XDP_FRAME_SIZE));
1029
1030         snprintf(mz_name, sizeof(mz_name), "af_xdp_umem_%s_%u",
1031                        internals->if_name, rxq->xsk_queue_idx);
1032         mz = rte_memzone_reserve_aligned(mz_name,
1033                         ETH_AF_XDP_NUM_BUFFERS * ETH_AF_XDP_FRAME_SIZE,
1034                         rte_socket_id(), RTE_MEMZONE_IOVA_CONTIG,
1035                         getpagesize());
1036         if (mz == NULL) {
1037                 AF_XDP_LOG(ERR, "Failed to reserve memzone for af_xdp umem.\n");
1038                 goto err;
1039         }
1040
1041         ret = xsk_umem__create(&umem->umem, mz->addr,
1042                                ETH_AF_XDP_NUM_BUFFERS * ETH_AF_XDP_FRAME_SIZE,
1043                                &rxq->fq, &rxq->cq,
1044                                &usr_config);
1045
1046         if (ret) {
1047                 AF_XDP_LOG(ERR, "Failed to create umem");
1048                 goto err;
1049         }
1050         umem->mz = mz;
1051
1052 #endif
1053         return umem;
1054
1055 err:
1056         xdp_umem_destroy(umem);
1057         return NULL;
1058 }
1059
1060 static int
1061 load_custom_xdp_prog(const char *prog_path, int if_index)
1062 {
1063         int ret, prog_fd = -1;
1064         struct bpf_object *obj;
1065         struct bpf_map *map;
1066
1067         ret = bpf_prog_load(prog_path, BPF_PROG_TYPE_XDP, &obj, &prog_fd);
1068         if (ret) {
1069                 AF_XDP_LOG(ERR, "Failed to load program %s\n", prog_path);
1070                 return ret;
1071         }
1072
1073         /*
1074          * The loaded program must provision for a map of xsks, such that some
1075          * traffic can be redirected to userspace. When the xsk is created,
1076          * libbpf inserts it into the map.
1077          */
1078         map = bpf_object__find_map_by_name(obj, "xsks_map");
1079         if (!map) {
1080                 AF_XDP_LOG(ERR, "Failed to find xsks_map in %s\n", prog_path);
1081                 return -1;
1082         }
1083
1084         /* Link the program with the given network device */
1085         ret = bpf_set_link_xdp_fd(if_index, prog_fd,
1086                                         XDP_FLAGS_UPDATE_IF_NOEXIST);
1087         if (ret) {
1088                 AF_XDP_LOG(ERR, "Failed to set prog fd %d on interface\n",
1089                                 prog_fd);
1090                 return -1;
1091         }
1092
1093         AF_XDP_LOG(INFO, "Successfully loaded XDP program %s with fd %d\n",
1094                                 prog_path, prog_fd);
1095
1096         return 0;
1097 }
1098
1099 static int
1100 xsk_configure(struct pmd_internals *internals, struct pkt_rx_queue *rxq,
1101               int ring_size)
1102 {
1103         struct xsk_socket_config cfg;
1104         struct pkt_tx_queue *txq = rxq->pair;
1105         int ret = 0;
1106         int reserve_size = ETH_AF_XDP_DFLT_NUM_DESCS;
1107         struct rte_mbuf *fq_bufs[reserve_size];
1108
1109         rxq->umem = xdp_umem_configure(internals, rxq);
1110         if (rxq->umem == NULL)
1111                 return -ENOMEM;
1112         txq->umem = rxq->umem;
1113
1114         cfg.rx_size = ring_size;
1115         cfg.tx_size = ring_size;
1116         cfg.libbpf_flags = 0;
1117         cfg.xdp_flags = XDP_FLAGS_UPDATE_IF_NOEXIST;
1118         cfg.bind_flags = 0;
1119
1120 #if defined(XDP_USE_NEED_WAKEUP)
1121         cfg.bind_flags |= XDP_USE_NEED_WAKEUP;
1122 #endif
1123
1124         if (strnlen(internals->prog_path, PATH_MAX) &&
1125                                 !internals->custom_prog_configured) {
1126                 ret = load_custom_xdp_prog(internals->prog_path,
1127                                            internals->if_index);
1128                 if (ret) {
1129                         AF_XDP_LOG(ERR, "Failed to load custom XDP program %s\n",
1130                                         internals->prog_path);
1131                         goto err;
1132                 }
1133                 internals->custom_prog_configured = 1;
1134         }
1135
1136         if (internals->shared_umem)
1137                 ret = create_shared_socket(&rxq->xsk, internals->if_name,
1138                                 rxq->xsk_queue_idx, rxq->umem->umem, &rxq->rx,
1139                                 &txq->tx, &rxq->fq, &rxq->cq, &cfg);
1140         else
1141                 ret = xsk_socket__create(&rxq->xsk, internals->if_name,
1142                                 rxq->xsk_queue_idx, rxq->umem->umem, &rxq->rx,
1143                                 &txq->tx, &cfg);
1144
1145         if (ret) {
1146                 AF_XDP_LOG(ERR, "Failed to create xsk socket.\n");
1147                 goto err;
1148         }
1149
1150 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
1151         if (rte_pktmbuf_alloc_bulk(rxq->umem->mb_pool, fq_bufs, reserve_size)) {
1152                 AF_XDP_LOG(DEBUG, "Failed to get enough buffers for fq.\n");
1153                 goto err;
1154         }
1155 #endif
1156         ret = reserve_fill_queue(rxq->umem, reserve_size, fq_bufs, &rxq->fq);
1157         if (ret) {
1158                 xsk_socket__delete(rxq->xsk);
1159                 AF_XDP_LOG(ERR, "Failed to reserve fill queue.\n");
1160                 goto err;
1161         }
1162
1163         return 0;
1164
1165 err:
1166         if (__atomic_sub_fetch(&rxq->umem->refcnt, 1, __ATOMIC_ACQUIRE) == 0)
1167                 xdp_umem_destroy(rxq->umem);
1168
1169         return ret;
1170 }
1171
1172 static int
1173 eth_rx_queue_setup(struct rte_eth_dev *dev,
1174                    uint16_t rx_queue_id,
1175                    uint16_t nb_rx_desc,
1176                    unsigned int socket_id __rte_unused,
1177                    const struct rte_eth_rxconf *rx_conf __rte_unused,
1178                    struct rte_mempool *mb_pool)
1179 {
1180         struct pmd_internals *internals = dev->data->dev_private;
1181         struct pkt_rx_queue *rxq;
1182         int ret;
1183
1184         rxq = &internals->rx_queues[rx_queue_id];
1185
1186         AF_XDP_LOG(INFO, "Set up rx queue, rx queue id: %d, xsk queue id: %d\n",
1187                    rx_queue_id, rxq->xsk_queue_idx);
1188
1189 #ifndef XDP_UMEM_UNALIGNED_CHUNK_FLAG
1190         uint32_t buf_size, data_size;
1191
1192         /* Now get the space available for data in the mbuf */
1193         buf_size = rte_pktmbuf_data_room_size(mb_pool) -
1194                 RTE_PKTMBUF_HEADROOM;
1195         data_size = ETH_AF_XDP_FRAME_SIZE;
1196
1197         if (data_size > buf_size) {
1198                 AF_XDP_LOG(ERR, "%s: %d bytes will not fit in mbuf (%d bytes)\n",
1199                         dev->device->name, data_size, buf_size);
1200                 ret = -ENOMEM;
1201                 goto err;
1202         }
1203 #endif
1204
1205         rxq->mb_pool = mb_pool;
1206
1207         if (xsk_configure(internals, rxq, nb_rx_desc)) {
1208                 AF_XDP_LOG(ERR, "Failed to configure xdp socket\n");
1209                 ret = -EINVAL;
1210                 goto err;
1211         }
1212
1213         rxq->fds[0].fd = xsk_socket__fd(rxq->xsk);
1214         rxq->fds[0].events = POLLIN;
1215
1216         dev->data->rx_queues[rx_queue_id] = rxq;
1217         return 0;
1218
1219 err:
1220         return ret;
1221 }
1222
1223 static int
1224 eth_tx_queue_setup(struct rte_eth_dev *dev,
1225                    uint16_t tx_queue_id,
1226                    uint16_t nb_tx_desc __rte_unused,
1227                    unsigned int socket_id __rte_unused,
1228                    const struct rte_eth_txconf *tx_conf __rte_unused)
1229 {
1230         struct pmd_internals *internals = dev->data->dev_private;
1231         struct pkt_tx_queue *txq;
1232
1233         txq = &internals->tx_queues[tx_queue_id];
1234
1235         dev->data->tx_queues[tx_queue_id] = txq;
1236         return 0;
1237 }
1238
1239 static int
1240 eth_dev_mtu_set(struct rte_eth_dev *dev, uint16_t mtu)
1241 {
1242         struct pmd_internals *internals = dev->data->dev_private;
1243         struct ifreq ifr = { .ifr_mtu = mtu };
1244         int ret;
1245         int s;
1246
1247         s = socket(PF_INET, SOCK_DGRAM, 0);
1248         if (s < 0)
1249                 return -EINVAL;
1250
1251         strlcpy(ifr.ifr_name, internals->if_name, IFNAMSIZ);
1252         ret = ioctl(s, SIOCSIFMTU, &ifr);
1253         close(s);
1254
1255         return (ret < 0) ? -errno : 0;
1256 }
1257
1258 static int
1259 eth_dev_change_flags(char *if_name, uint32_t flags, uint32_t mask)
1260 {
1261         struct ifreq ifr;
1262         int ret = 0;
1263         int s;
1264
1265         s = socket(PF_INET, SOCK_DGRAM, 0);
1266         if (s < 0)
1267                 return -errno;
1268
1269         strlcpy(ifr.ifr_name, if_name, IFNAMSIZ);
1270         if (ioctl(s, SIOCGIFFLAGS, &ifr) < 0) {
1271                 ret = -errno;
1272                 goto out;
1273         }
1274         ifr.ifr_flags &= mask;
1275         ifr.ifr_flags |= flags;
1276         if (ioctl(s, SIOCSIFFLAGS, &ifr) < 0) {
1277                 ret = -errno;
1278                 goto out;
1279         }
1280 out:
1281         close(s);
1282         return ret;
1283 }
1284
1285 static int
1286 eth_dev_promiscuous_enable(struct rte_eth_dev *dev)
1287 {
1288         struct pmd_internals *internals = dev->data->dev_private;
1289
1290         return eth_dev_change_flags(internals->if_name, IFF_PROMISC, ~0);
1291 }
1292
1293 static int
1294 eth_dev_promiscuous_disable(struct rte_eth_dev *dev)
1295 {
1296         struct pmd_internals *internals = dev->data->dev_private;
1297
1298         return eth_dev_change_flags(internals->if_name, 0, ~IFF_PROMISC);
1299 }
1300
1301 static const struct eth_dev_ops ops = {
1302         .dev_start = eth_dev_start,
1303         .dev_stop = eth_dev_stop,
1304         .dev_close = eth_dev_close,
1305         .dev_configure = eth_dev_configure,
1306         .dev_infos_get = eth_dev_info,
1307         .mtu_set = eth_dev_mtu_set,
1308         .promiscuous_enable = eth_dev_promiscuous_enable,
1309         .promiscuous_disable = eth_dev_promiscuous_disable,
1310         .rx_queue_setup = eth_rx_queue_setup,
1311         .tx_queue_setup = eth_tx_queue_setup,
1312         .rx_queue_release = eth_queue_release,
1313         .tx_queue_release = eth_queue_release,
1314         .link_update = eth_link_update,
1315         .stats_get = eth_stats_get,
1316         .stats_reset = eth_stats_reset,
1317 };
1318
1319 /** parse integer from integer argument */
1320 static int
1321 parse_integer_arg(const char *key __rte_unused,
1322                   const char *value, void *extra_args)
1323 {
1324         int *i = (int *)extra_args;
1325         char *end;
1326
1327         *i = strtol(value, &end, 10);
1328         if (*i < 0) {
1329                 AF_XDP_LOG(ERR, "Argument has to be positive.\n");
1330                 return -EINVAL;
1331         }
1332
1333         return 0;
1334 }
1335
1336 /** parse name argument */
1337 static int
1338 parse_name_arg(const char *key __rte_unused,
1339                const char *value, void *extra_args)
1340 {
1341         char *name = extra_args;
1342
1343         if (strnlen(value, IFNAMSIZ) > IFNAMSIZ - 1) {
1344                 AF_XDP_LOG(ERR, "Invalid name %s, should be less than %u bytes.\n",
1345                            value, IFNAMSIZ);
1346                 return -EINVAL;
1347         }
1348
1349         strlcpy(name, value, IFNAMSIZ);
1350
1351         return 0;
1352 }
1353
1354 /** parse xdp prog argument */
1355 static int
1356 parse_prog_arg(const char *key __rte_unused,
1357                const char *value, void *extra_args)
1358 {
1359         char *path = extra_args;
1360
1361         if (strnlen(value, PATH_MAX) == PATH_MAX) {
1362                 AF_XDP_LOG(ERR, "Invalid path %s, should be less than %u bytes.\n",
1363                            value, PATH_MAX);
1364                 return -EINVAL;
1365         }
1366
1367         if (access(value, F_OK) != 0) {
1368                 AF_XDP_LOG(ERR, "Error accessing %s: %s\n",
1369                            value, strerror(errno));
1370                 return -EINVAL;
1371         }
1372
1373         strlcpy(path, value, PATH_MAX);
1374
1375         return 0;
1376 }
1377
1378 static int
1379 xdp_get_channels_info(const char *if_name, int *max_queues,
1380                                 int *combined_queues)
1381 {
1382         struct ethtool_channels channels;
1383         struct ifreq ifr;
1384         int fd, ret;
1385
1386         fd = socket(AF_INET, SOCK_DGRAM, 0);
1387         if (fd < 0)
1388                 return -1;
1389
1390         channels.cmd = ETHTOOL_GCHANNELS;
1391         ifr.ifr_data = (void *)&channels;
1392         strlcpy(ifr.ifr_name, if_name, IFNAMSIZ);
1393         ret = ioctl(fd, SIOCETHTOOL, &ifr);
1394         if (ret) {
1395                 if (errno == EOPNOTSUPP) {
1396                         ret = 0;
1397                 } else {
1398                         ret = -errno;
1399                         goto out;
1400                 }
1401         }
1402
1403         if (channels.max_combined == 0 || errno == EOPNOTSUPP) {
1404                 /* If the device says it has no channels, then all traffic
1405                  * is sent to a single stream, so max queues = 1.
1406                  */
1407                 *max_queues = 1;
1408                 *combined_queues = 1;
1409         } else {
1410                 *max_queues = channels.max_combined;
1411                 *combined_queues = channels.combined_count;
1412         }
1413
1414  out:
1415         close(fd);
1416         return ret;
1417 }
1418
1419 static int
1420 parse_parameters(struct rte_kvargs *kvlist, char *if_name, int *start_queue,
1421                         int *queue_cnt, int *shared_umem, char *prog_path)
1422 {
1423         int ret;
1424
1425         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_IFACE_ARG,
1426                                  &parse_name_arg, if_name);
1427         if (ret < 0)
1428                 goto free_kvlist;
1429
1430         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_START_QUEUE_ARG,
1431                                  &parse_integer_arg, start_queue);
1432         if (ret < 0)
1433                 goto free_kvlist;
1434
1435         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_QUEUE_COUNT_ARG,
1436                                  &parse_integer_arg, queue_cnt);
1437         if (ret < 0 || *queue_cnt <= 0) {
1438                 ret = -EINVAL;
1439                 goto free_kvlist;
1440         }
1441
1442         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_SHARED_UMEM_ARG,
1443                                 &parse_integer_arg, shared_umem);
1444         if (ret < 0)
1445                 goto free_kvlist;
1446
1447         ret = rte_kvargs_process(kvlist, ETH_AF_XDP_PROG_ARG,
1448                                  &parse_prog_arg, prog_path);
1449         if (ret < 0)
1450                 goto free_kvlist;
1451
1452 free_kvlist:
1453         rte_kvargs_free(kvlist);
1454         return ret;
1455 }
1456
1457 static int
1458 get_iface_info(const char *if_name,
1459                struct rte_ether_addr *eth_addr,
1460                int *if_index)
1461 {
1462         struct ifreq ifr;
1463         int sock = socket(AF_INET, SOCK_DGRAM, IPPROTO_IP);
1464
1465         if (sock < 0)
1466                 return -1;
1467
1468         strlcpy(ifr.ifr_name, if_name, IFNAMSIZ);
1469         if (ioctl(sock, SIOCGIFINDEX, &ifr))
1470                 goto error;
1471
1472         *if_index = ifr.ifr_ifindex;
1473
1474         if (ioctl(sock, SIOCGIFHWADDR, &ifr))
1475                 goto error;
1476
1477         rte_memcpy(eth_addr, ifr.ifr_hwaddr.sa_data, RTE_ETHER_ADDR_LEN);
1478
1479         close(sock);
1480         return 0;
1481
1482 error:
1483         close(sock);
1484         return -1;
1485 }
1486
1487 static struct rte_eth_dev *
1488 init_internals(struct rte_vdev_device *dev, const char *if_name,
1489                 int start_queue_idx, int queue_cnt, int shared_umem,
1490                 const char *prog_path)
1491 {
1492         const char *name = rte_vdev_device_name(dev);
1493         const unsigned int numa_node = dev->device.numa_node;
1494         struct pmd_internals *internals;
1495         struct rte_eth_dev *eth_dev;
1496         int ret;
1497         int i;
1498
1499         internals = rte_zmalloc_socket(name, sizeof(*internals), 0, numa_node);
1500         if (internals == NULL)
1501                 return NULL;
1502
1503         internals->start_queue_idx = start_queue_idx;
1504         internals->queue_cnt = queue_cnt;
1505         strlcpy(internals->if_name, if_name, IFNAMSIZ);
1506         strlcpy(internals->prog_path, prog_path, PATH_MAX);
1507         internals->custom_prog_configured = 0;
1508
1509 #ifndef ETH_AF_XDP_SHARED_UMEM
1510         if (shared_umem) {
1511                 AF_XDP_LOG(ERR, "Shared UMEM feature not available. "
1512                                 "Check kernel and libbpf version\n");
1513                 goto err_free_internals;
1514         }
1515 #endif
1516         internals->shared_umem = shared_umem;
1517
1518         if (xdp_get_channels_info(if_name, &internals->max_queue_cnt,
1519                                   &internals->combined_queue_cnt)) {
1520                 AF_XDP_LOG(ERR, "Failed to get channel info of interface: %s\n",
1521                                 if_name);
1522                 goto err_free_internals;
1523         }
1524
1525         if (queue_cnt > internals->combined_queue_cnt) {
1526                 AF_XDP_LOG(ERR, "Specified queue count %d is larger than combined queue count %d.\n",
1527                                 queue_cnt, internals->combined_queue_cnt);
1528                 goto err_free_internals;
1529         }
1530
1531         internals->rx_queues = rte_zmalloc_socket(NULL,
1532                                         sizeof(struct pkt_rx_queue) * queue_cnt,
1533                                         0, numa_node);
1534         if (internals->rx_queues == NULL) {
1535                 AF_XDP_LOG(ERR, "Failed to allocate memory for rx queues.\n");
1536                 goto err_free_internals;
1537         }
1538
1539         internals->tx_queues = rte_zmalloc_socket(NULL,
1540                                         sizeof(struct pkt_tx_queue) * queue_cnt,
1541                                         0, numa_node);
1542         if (internals->tx_queues == NULL) {
1543                 AF_XDP_LOG(ERR, "Failed to allocate memory for tx queues.\n");
1544                 goto err_free_rx;
1545         }
1546         for (i = 0; i < queue_cnt; i++) {
1547                 internals->tx_queues[i].pair = &internals->rx_queues[i];
1548                 internals->rx_queues[i].pair = &internals->tx_queues[i];
1549                 internals->rx_queues[i].xsk_queue_idx = start_queue_idx + i;
1550                 internals->tx_queues[i].xsk_queue_idx = start_queue_idx + i;
1551         }
1552
1553         ret = get_iface_info(if_name, &internals->eth_addr,
1554                              &internals->if_index);
1555         if (ret)
1556                 goto err_free_tx;
1557
1558         eth_dev = rte_eth_vdev_allocate(dev, 0);
1559         if (eth_dev == NULL)
1560                 goto err_free_tx;
1561
1562         eth_dev->data->dev_private = internals;
1563         eth_dev->data->dev_link = pmd_link;
1564         eth_dev->data->mac_addrs = &internals->eth_addr;
1565         eth_dev->data->dev_flags |= RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS;
1566         eth_dev->dev_ops = &ops;
1567         eth_dev->rx_pkt_burst = eth_af_xdp_rx;
1568         eth_dev->tx_pkt_burst = eth_af_xdp_tx;
1569
1570 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
1571         AF_XDP_LOG(INFO, "Zero copy between umem and mbuf enabled.\n");
1572 #endif
1573
1574         return eth_dev;
1575
1576 err_free_tx:
1577         rte_free(internals->tx_queues);
1578 err_free_rx:
1579         rte_free(internals->rx_queues);
1580 err_free_internals:
1581         rte_free(internals);
1582         return NULL;
1583 }
1584
1585 static int
1586 rte_pmd_af_xdp_probe(struct rte_vdev_device *dev)
1587 {
1588         struct rte_kvargs *kvlist;
1589         char if_name[IFNAMSIZ] = {'\0'};
1590         int xsk_start_queue_idx = ETH_AF_XDP_DFLT_START_QUEUE_IDX;
1591         int xsk_queue_cnt = ETH_AF_XDP_DFLT_QUEUE_COUNT;
1592         int shared_umem = 0;
1593         char prog_path[PATH_MAX] = {'\0'};
1594         struct rte_eth_dev *eth_dev = NULL;
1595         const char *name;
1596
1597         AF_XDP_LOG(INFO, "Initializing pmd_af_xdp for %s\n",
1598                 rte_vdev_device_name(dev));
1599
1600         name = rte_vdev_device_name(dev);
1601         if (rte_eal_process_type() == RTE_PROC_SECONDARY &&
1602                 strlen(rte_vdev_device_args(dev)) == 0) {
1603                 eth_dev = rte_eth_dev_attach_secondary(name);
1604                 if (eth_dev == NULL) {
1605                         AF_XDP_LOG(ERR, "Failed to probe %s\n", name);
1606                         return -EINVAL;
1607                 }
1608                 eth_dev->dev_ops = &ops;
1609                 rte_eth_dev_probing_finish(eth_dev);
1610                 return 0;
1611         }
1612
1613         kvlist = rte_kvargs_parse(rte_vdev_device_args(dev), valid_arguments);
1614         if (kvlist == NULL) {
1615                 AF_XDP_LOG(ERR, "Invalid kvargs key\n");
1616                 return -EINVAL;
1617         }
1618
1619         if (dev->device.numa_node == SOCKET_ID_ANY)
1620                 dev->device.numa_node = rte_socket_id();
1621
1622         if (parse_parameters(kvlist, if_name, &xsk_start_queue_idx,
1623                              &xsk_queue_cnt, &shared_umem, prog_path) < 0) {
1624                 AF_XDP_LOG(ERR, "Invalid kvargs value\n");
1625                 return -EINVAL;
1626         }
1627
1628         if (strlen(if_name) == 0) {
1629                 AF_XDP_LOG(ERR, "Network interface must be specified\n");
1630                 return -EINVAL;
1631         }
1632
1633         eth_dev = init_internals(dev, if_name, xsk_start_queue_idx,
1634                                         xsk_queue_cnt, shared_umem, prog_path);
1635         if (eth_dev == NULL) {
1636                 AF_XDP_LOG(ERR, "Failed to init internals\n");
1637                 return -1;
1638         }
1639
1640         rte_eth_dev_probing_finish(eth_dev);
1641
1642         return 0;
1643 }
1644
1645 static int
1646 rte_pmd_af_xdp_remove(struct rte_vdev_device *dev)
1647 {
1648         struct rte_eth_dev *eth_dev = NULL;
1649
1650         AF_XDP_LOG(INFO, "Removing AF_XDP ethdev on numa socket %u\n",
1651                 rte_socket_id());
1652
1653         if (dev == NULL)
1654                 return -1;
1655
1656         /* find the ethdev entry */
1657         eth_dev = rte_eth_dev_allocated(rte_vdev_device_name(dev));
1658         if (eth_dev == NULL)
1659                 return 0;
1660
1661         eth_dev_close(eth_dev);
1662         rte_eth_dev_release_port(eth_dev);
1663
1664
1665         return 0;
1666 }
1667
1668 static struct rte_vdev_driver pmd_af_xdp_drv = {
1669         .probe = rte_pmd_af_xdp_probe,
1670         .remove = rte_pmd_af_xdp_remove,
1671 };
1672
1673 RTE_PMD_REGISTER_VDEV(net_af_xdp, pmd_af_xdp_drv);
1674 RTE_PMD_REGISTER_PARAM_STRING(net_af_xdp,
1675                               "iface=<string> "
1676                               "start_queue=<int> "
1677                               "queue_count=<int> "
1678                               "shared_umem=<int> "
1679                               "xdp_prog=<string> ");