1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2019 Intel Corporation.
9 #include <netinet/in.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"
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_config.h>
32 #include <rte_ether.h>
33 #include <rte_lcore.h>
35 #include <rte_memory.h>
36 #include <rte_memzone.h>
38 #include <rte_malloc.h>
53 static int af_xdp_logtype;
55 #define AF_XDP_LOG(level, fmt, args...) \
56 rte_log(RTE_LOG_ ## level, af_xdp_logtype, \
57 "%s(): " fmt, __func__, ##args)
59 #define ETH_AF_XDP_FRAME_SIZE 2048
60 #define ETH_AF_XDP_NUM_BUFFERS 4096
61 #define ETH_AF_XDP_DATA_HEADROOM 0
62 #define ETH_AF_XDP_DFLT_NUM_DESCS XSK_RING_CONS__DEFAULT_NUM_DESCS
63 #define ETH_AF_XDP_DFLT_START_QUEUE_IDX 0
64 #define ETH_AF_XDP_DFLT_QUEUE_COUNT 1
66 #define ETH_AF_XDP_RX_BATCH_SIZE 32
67 #define ETH_AF_XDP_TX_BATCH_SIZE 32
70 struct xsk_umem_info {
71 struct xsk_ring_prod fq;
72 struct xsk_ring_cons cq;
73 struct xsk_umem *umem;
74 struct rte_ring *buf_ring;
75 const struct rte_memzone *mz;
86 struct xsk_ring_cons rx;
87 struct xsk_umem_info *umem;
88 struct xsk_socket *xsk;
89 struct rte_mempool *mb_pool;
91 struct rx_stats stats;
93 struct pkt_tx_queue *pair;
103 struct pkt_tx_queue {
104 struct xsk_ring_prod tx;
106 struct tx_stats stats;
108 struct pkt_rx_queue *pair;
112 struct pmd_internals {
114 char if_name[IFNAMSIZ];
118 int combined_queue_cnt;
121 struct rte_ether_addr eth_addr;
123 struct pkt_rx_queue *rx_queues;
124 struct pkt_tx_queue *tx_queues;
127 #define ETH_AF_XDP_IFACE_ARG "iface"
128 #define ETH_AF_XDP_START_QUEUE_ARG "start_queue"
129 #define ETH_AF_XDP_QUEUE_COUNT_ARG "queue_count"
130 #define ETH_AF_XDP_PMD_ZC_ARG "pmd_zero_copy"
132 static const char * const valid_arguments[] = {
133 ETH_AF_XDP_IFACE_ARG,
134 ETH_AF_XDP_START_QUEUE_ARG,
135 ETH_AF_XDP_QUEUE_COUNT_ARG,
136 ETH_AF_XDP_PMD_ZC_ARG,
140 static const struct rte_eth_link pmd_link = {
141 .link_speed = ETH_SPEED_NUM_10G,
142 .link_duplex = ETH_LINK_FULL_DUPLEX,
143 .link_status = ETH_LINK_DOWN,
144 .link_autoneg = ETH_LINK_AUTONEG
148 reserve_fill_queue(struct xsk_umem_info *umem, uint16_t reserve_size)
150 struct xsk_ring_prod *fq = &umem->fq;
151 void *addrs[reserve_size];
155 if (rte_ring_dequeue_bulk(umem->buf_ring, addrs, reserve_size, NULL)
157 AF_XDP_LOG(DEBUG, "Failed to get enough buffers for fq.\n");
161 if (unlikely(!xsk_ring_prod__reserve(fq, reserve_size, &idx))) {
162 AF_XDP_LOG(DEBUG, "Failed to reserve enough fq descs.\n");
163 rte_ring_enqueue_bulk(umem->buf_ring, addrs,
168 for (i = 0; i < reserve_size; i++) {
171 fq_addr = xsk_ring_prod__fill_addr(fq, idx++);
172 *fq_addr = (uint64_t)addrs[i];
175 xsk_ring_prod__submit(fq, reserve_size);
181 umem_buf_release_to_fq(void *addr, void *opaque)
183 struct xsk_umem_info *umem = (struct xsk_umem_info *)opaque;
184 uint64_t umem_addr = (uint64_t)addr - umem->mz->addr_64;
186 rte_ring_enqueue(umem->buf_ring, (void *)umem_addr);
190 eth_af_xdp_rx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
192 struct pkt_rx_queue *rxq = queue;
193 struct xsk_ring_cons *rx = &rxq->rx;
194 struct xsk_umem_info *umem = rxq->umem;
195 struct xsk_ring_prod *fq = &umem->fq;
197 uint32_t free_thresh = fq->size >> 1;
198 int pmd_zc = umem->pmd_zc;
199 struct rte_mbuf *mbufs[ETH_AF_XDP_RX_BATCH_SIZE];
200 unsigned long dropped = 0;
201 unsigned long rx_bytes = 0;
204 nb_pkts = RTE_MIN(nb_pkts, ETH_AF_XDP_RX_BATCH_SIZE);
206 if (unlikely(rte_pktmbuf_alloc_bulk(rxq->mb_pool, mbufs, nb_pkts) != 0))
209 rcvd = xsk_ring_cons__peek(rx, nb_pkts, &idx_rx);
211 #if defined(XDP_USE_NEED_WAKEUP)
212 if (xsk_ring_prod__needs_wakeup(fq))
213 (void)poll(rxq->fds, 1, 1000);
219 if (xsk_prod_nb_free(fq, free_thresh) >= free_thresh)
220 (void)reserve_fill_queue(umem, ETH_AF_XDP_RX_BATCH_SIZE);
222 for (i = 0; i < rcvd; i++) {
223 const struct xdp_desc *desc;
227 uint16_t buf_len = ETH_AF_XDP_FRAME_SIZE;
228 struct rte_mbuf_ext_shared_info *shinfo;
230 desc = xsk_ring_cons__rx_desc(rx, idx_rx++);
233 pkt = xsk_umem__get_data(rxq->umem->mz->addr, addr);
236 shinfo = rte_pktmbuf_ext_shinfo_init_helper(pkt,
237 &buf_len, umem_buf_release_to_fq, umem);
239 rte_pktmbuf_attach_extbuf(mbufs[i], pkt, 0, buf_len,
242 rte_memcpy(rte_pktmbuf_mtod(mbufs[i], void *),
244 rte_ring_enqueue(umem->buf_ring, (void *)addr);
246 rte_pktmbuf_pkt_len(mbufs[i]) = len;
247 rte_pktmbuf_data_len(mbufs[i]) = len;
252 xsk_ring_cons__release(rx, rcvd);
255 rxq->stats.rx_pkts += (rcvd - dropped);
256 rxq->stats.rx_bytes += rx_bytes;
260 rte_mempool_put_bulk(rxq->mb_pool, (void **)&mbufs[rcvd],
267 pull_umem_cq(struct xsk_umem_info *umem, int size)
269 struct xsk_ring_cons *cq = &umem->cq;
273 n = xsk_ring_cons__peek(cq, size, &idx_cq);
275 for (i = 0; i < n; i++) {
277 addr = *xsk_ring_cons__comp_addr(cq, idx_cq++);
278 rte_ring_enqueue(umem->buf_ring, (void *)addr);
281 xsk_ring_cons__release(cq, n);
285 kick_tx(struct pkt_tx_queue *txq)
287 struct xsk_umem_info *umem = txq->pair->umem;
289 #if defined(XDP_USE_NEED_WAKEUP)
290 if (xsk_ring_prod__needs_wakeup(&txq->tx))
292 while (send(xsk_socket__fd(txq->pair->xsk), NULL,
293 0, MSG_DONTWAIT) < 0) {
294 /* some thing unexpected */
295 if (errno != EBUSY && errno != EAGAIN && errno != EINTR)
298 /* pull from completion queue to leave more space */
300 pull_umem_cq(umem, ETH_AF_XDP_TX_BATCH_SIZE);
302 pull_umem_cq(umem, ETH_AF_XDP_TX_BATCH_SIZE);
306 in_umem_range(struct xsk_umem_info *umem, uint64_t addr)
308 uint64_t mz_base_addr = umem->mz->addr_64;
310 return addr >= mz_base_addr && addr < mz_base_addr + umem->mz->len;
314 eth_af_xdp_tx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
316 struct pkt_tx_queue *txq = queue;
317 struct xsk_umem_info *umem = txq->pair->umem;
318 struct rte_mbuf *mbuf;
319 int pmd_zc = umem->pmd_zc;
320 void *addrs[ETH_AF_XDP_TX_BATCH_SIZE];
321 unsigned long tx_bytes = 0;
325 nb_pkts = RTE_MIN(nb_pkts, ETH_AF_XDP_TX_BATCH_SIZE);
327 pull_umem_cq(umem, nb_pkts);
329 nb_pkts = rte_ring_dequeue_bulk(umem->buf_ring, addrs,
334 if (xsk_ring_prod__reserve(&txq->tx, nb_pkts, &idx_tx) != nb_pkts) {
336 rte_ring_enqueue_bulk(umem->buf_ring, addrs, nb_pkts, NULL);
340 for (i = 0; i < nb_pkts; i++) {
341 struct xdp_desc *desc;
344 desc = xsk_ring_prod__tx_desc(&txq->tx, idx_tx + i);
346 desc->len = mbuf->pkt_len;
349 * We need to make sure the external mbuf address is within
350 * current port's umem memzone range
352 if (pmd_zc && RTE_MBUF_HAS_EXTBUF(mbuf) &&
353 in_umem_range(umem, (uint64_t)mbuf->buf_addr)) {
354 desc->addr = (uint64_t)mbuf->buf_addr -
356 mbuf->buf_addr = xsk_umem__get_data(umem->mz->addr,
359 desc->addr = (uint64_t)addrs[i];
360 pkt = xsk_umem__get_data(umem->mz->addr,
362 rte_memcpy(pkt, rte_pktmbuf_mtod(mbuf, void *),
365 tx_bytes += mbuf->pkt_len;
368 xsk_ring_prod__submit(&txq->tx, nb_pkts);
372 txq->stats.tx_pkts += nb_pkts;
373 txq->stats.tx_bytes += tx_bytes;
375 for (i = 0; i < nb_pkts; i++)
376 rte_pktmbuf_free(bufs[i]);
382 eth_dev_start(struct rte_eth_dev *dev)
384 dev->data->dev_link.link_status = ETH_LINK_UP;
389 /* This function gets called when the current port gets stopped. */
391 eth_dev_stop(struct rte_eth_dev *dev)
393 dev->data->dev_link.link_status = ETH_LINK_DOWN;
397 eth_dev_configure(struct rte_eth_dev *dev)
399 /* rx/tx must be paired */
400 if (dev->data->nb_rx_queues != dev->data->nb_tx_queues)
407 eth_dev_info(struct rte_eth_dev *dev, struct rte_eth_dev_info *dev_info)
409 struct pmd_internals *internals = dev->data->dev_private;
411 dev_info->if_index = internals->if_index;
412 dev_info->max_mac_addrs = 1;
413 dev_info->max_rx_pktlen = ETH_FRAME_LEN;
414 dev_info->max_rx_queues = internals->queue_cnt;
415 dev_info->max_tx_queues = internals->queue_cnt;
417 dev_info->min_mtu = RTE_ETHER_MIN_MTU;
418 dev_info->max_mtu = ETH_AF_XDP_FRAME_SIZE - ETH_AF_XDP_DATA_HEADROOM;
420 dev_info->default_rxportconf.nb_queues = 1;
421 dev_info->default_txportconf.nb_queues = 1;
422 dev_info->default_rxportconf.ring_size = ETH_AF_XDP_DFLT_NUM_DESCS;
423 dev_info->default_txportconf.ring_size = ETH_AF_XDP_DFLT_NUM_DESCS;
429 eth_stats_get(struct rte_eth_dev *dev, struct rte_eth_stats *stats)
431 struct pmd_internals *internals = dev->data->dev_private;
432 struct xdp_statistics xdp_stats;
433 struct pkt_rx_queue *rxq;
434 struct pkt_tx_queue *txq;
438 for (i = 0; i < dev->data->nb_rx_queues; i++) {
439 optlen = sizeof(struct xdp_statistics);
440 rxq = &internals->rx_queues[i];
442 stats->q_ipackets[i] = rxq->stats.rx_pkts;
443 stats->q_ibytes[i] = rxq->stats.rx_bytes;
445 stats->q_opackets[i] = txq->stats.tx_pkts;
446 stats->q_obytes[i] = txq->stats.tx_bytes;
448 stats->ipackets += stats->q_ipackets[i];
449 stats->ibytes += stats->q_ibytes[i];
450 stats->imissed += rxq->stats.rx_dropped;
451 ret = getsockopt(xsk_socket__fd(rxq->xsk), SOL_XDP,
452 XDP_STATISTICS, &xdp_stats, &optlen);
454 AF_XDP_LOG(ERR, "getsockopt() failed for XDP_STATISTICS.\n");
457 stats->imissed += xdp_stats.rx_dropped;
459 stats->opackets += stats->q_opackets[i];
460 stats->obytes += stats->q_obytes[i];
467 eth_stats_reset(struct rte_eth_dev *dev)
469 struct pmd_internals *internals = dev->data->dev_private;
472 for (i = 0; i < internals->queue_cnt; i++) {
473 memset(&internals->rx_queues[i].stats, 0,
474 sizeof(struct rx_stats));
475 memset(&internals->tx_queues[i].stats, 0,
476 sizeof(struct tx_stats));
483 remove_xdp_program(struct pmd_internals *internals)
485 uint32_t curr_prog_id = 0;
487 if (bpf_get_link_xdp_id(internals->if_index, &curr_prog_id,
488 XDP_FLAGS_UPDATE_IF_NOEXIST)) {
489 AF_XDP_LOG(ERR, "bpf_get_link_xdp_id failed\n");
492 bpf_set_link_xdp_fd(internals->if_index, -1,
493 XDP_FLAGS_UPDATE_IF_NOEXIST);
497 xdp_umem_destroy(struct xsk_umem_info *umem)
499 rte_memzone_free(umem->mz);
502 rte_ring_free(umem->buf_ring);
503 umem->buf_ring = NULL;
510 eth_dev_close(struct rte_eth_dev *dev)
512 struct pmd_internals *internals = dev->data->dev_private;
513 struct pkt_rx_queue *rxq;
516 AF_XDP_LOG(INFO, "Closing AF_XDP ethdev on numa socket %u\n",
519 for (i = 0; i < internals->queue_cnt; i++) {
520 rxq = &internals->rx_queues[i];
521 if (rxq->umem == NULL)
523 xsk_socket__delete(rxq->xsk);
524 (void)xsk_umem__delete(rxq->umem->umem);
525 xdp_umem_destroy(rxq->umem);
527 /* free pkt_tx_queue */
533 * MAC is not allocated dynamically, setting it to NULL would prevent
534 * from releasing it in rte_eth_dev_release_port.
536 dev->data->mac_addrs = NULL;
538 remove_xdp_program(internals);
542 eth_queue_release(void *q __rte_unused)
547 eth_link_update(struct rte_eth_dev *dev __rte_unused,
548 int wait_to_complete __rte_unused)
554 xsk_umem_info *xdp_umem_configure(struct pmd_internals *internals,
555 struct pkt_rx_queue *rxq)
557 struct xsk_umem_info *umem;
558 const struct rte_memzone *mz;
559 struct xsk_umem_config usr_config = {
560 .fill_size = ETH_AF_XDP_DFLT_NUM_DESCS,
561 .comp_size = ETH_AF_XDP_DFLT_NUM_DESCS,
562 .frame_size = ETH_AF_XDP_FRAME_SIZE,
563 .frame_headroom = ETH_AF_XDP_DATA_HEADROOM };
564 char ring_name[RTE_RING_NAMESIZE];
565 char mz_name[RTE_MEMZONE_NAMESIZE];
569 umem = rte_zmalloc_socket("umem", sizeof(*umem), 0, rte_socket_id());
571 AF_XDP_LOG(ERR, "Failed to allocate umem info");
575 snprintf(ring_name, sizeof(ring_name), "af_xdp_ring_%s_%u",
576 internals->if_name, rxq->xsk_queue_idx);
577 umem->buf_ring = rte_ring_create(ring_name,
578 ETH_AF_XDP_NUM_BUFFERS,
581 if (umem->buf_ring == NULL) {
582 AF_XDP_LOG(ERR, "Failed to create rte_ring\n");
586 for (i = 0; i < ETH_AF_XDP_NUM_BUFFERS; i++)
587 rte_ring_enqueue(umem->buf_ring,
588 (void *)(i * ETH_AF_XDP_FRAME_SIZE +
589 ETH_AF_XDP_DATA_HEADROOM));
591 snprintf(mz_name, sizeof(mz_name), "af_xdp_umem_%s_%u",
592 internals->if_name, rxq->xsk_queue_idx);
593 mz = rte_memzone_reserve_aligned(mz_name,
594 ETH_AF_XDP_NUM_BUFFERS * ETH_AF_XDP_FRAME_SIZE,
595 rte_socket_id(), RTE_MEMZONE_IOVA_CONTIG,
598 AF_XDP_LOG(ERR, "Failed to reserve memzone for af_xdp umem.\n");
602 ret = xsk_umem__create(&umem->umem, mz->addr,
603 ETH_AF_XDP_NUM_BUFFERS * ETH_AF_XDP_FRAME_SIZE,
604 &umem->fq, &umem->cq,
608 AF_XDP_LOG(ERR, "Failed to create umem");
616 xdp_umem_destroy(umem);
621 xsk_configure(struct pmd_internals *internals, struct pkt_rx_queue *rxq,
624 struct xsk_socket_config cfg;
625 struct pkt_tx_queue *txq = rxq->pair;
629 rxq->umem = xdp_umem_configure(internals, rxq);
630 if (rxq->umem == NULL)
633 cfg.rx_size = ring_size;
634 cfg.tx_size = ring_size;
635 cfg.libbpf_flags = 0;
636 cfg.xdp_flags = XDP_FLAGS_UPDATE_IF_NOEXIST;
639 #if defined(XDP_USE_NEED_WAKEUP)
640 cfg.bind_flags |= XDP_USE_NEED_WAKEUP;
643 ret = xsk_socket__create(&rxq->xsk, internals->if_name,
644 rxq->xsk_queue_idx, rxq->umem->umem, &rxq->rx,
647 AF_XDP_LOG(ERR, "Failed to create xsk socket.\n");
651 reserve_size = ETH_AF_XDP_DFLT_NUM_DESCS / 2;
652 ret = reserve_fill_queue(rxq->umem, reserve_size);
654 xsk_socket__delete(rxq->xsk);
655 AF_XDP_LOG(ERR, "Failed to reserve fill queue.\n");
662 xdp_umem_destroy(rxq->umem);
668 eth_rx_queue_setup(struct rte_eth_dev *dev,
669 uint16_t rx_queue_id,
671 unsigned int socket_id __rte_unused,
672 const struct rte_eth_rxconf *rx_conf __rte_unused,
673 struct rte_mempool *mb_pool)
675 struct pmd_internals *internals = dev->data->dev_private;
676 uint32_t buf_size, data_size;
677 struct pkt_rx_queue *rxq;
680 rxq = &internals->rx_queues[rx_queue_id];
682 AF_XDP_LOG(INFO, "Set up rx queue, rx queue id: %d, xsk queue id: %d\n",
683 rx_queue_id, rxq->xsk_queue_idx);
684 /* Now get the space available for data in the mbuf */
685 buf_size = rte_pktmbuf_data_room_size(mb_pool) -
686 RTE_PKTMBUF_HEADROOM;
687 data_size = ETH_AF_XDP_FRAME_SIZE - ETH_AF_XDP_DATA_HEADROOM;
689 if (data_size > buf_size) {
690 AF_XDP_LOG(ERR, "%s: %d bytes will not fit in mbuf (%d bytes)\n",
691 dev->device->name, data_size, buf_size);
696 rxq->mb_pool = mb_pool;
698 if (xsk_configure(internals, rxq, nb_rx_desc)) {
699 AF_XDP_LOG(ERR, "Failed to configure xdp socket\n");
704 rxq->fds[0].fd = xsk_socket__fd(rxq->xsk);
705 rxq->fds[0].events = POLLIN;
707 rxq->umem->pmd_zc = internals->pmd_zc;
709 dev->data->rx_queues[rx_queue_id] = rxq;
717 eth_tx_queue_setup(struct rte_eth_dev *dev,
718 uint16_t tx_queue_id,
719 uint16_t nb_tx_desc __rte_unused,
720 unsigned int socket_id __rte_unused,
721 const struct rte_eth_txconf *tx_conf __rte_unused)
723 struct pmd_internals *internals = dev->data->dev_private;
724 struct pkt_tx_queue *txq;
726 txq = &internals->tx_queues[tx_queue_id];
728 dev->data->tx_queues[tx_queue_id] = txq;
733 eth_dev_mtu_set(struct rte_eth_dev *dev, uint16_t mtu)
735 struct pmd_internals *internals = dev->data->dev_private;
736 struct ifreq ifr = { .ifr_mtu = mtu };
740 s = socket(PF_INET, SOCK_DGRAM, 0);
744 strlcpy(ifr.ifr_name, internals->if_name, IFNAMSIZ);
745 ret = ioctl(s, SIOCSIFMTU, &ifr);
748 return (ret < 0) ? -errno : 0;
752 eth_dev_change_flags(char *if_name, uint32_t flags, uint32_t mask)
758 s = socket(PF_INET, SOCK_DGRAM, 0);
762 strlcpy(ifr.ifr_name, if_name, IFNAMSIZ);
763 if (ioctl(s, SIOCGIFFLAGS, &ifr) < 0) {
767 ifr.ifr_flags &= mask;
768 ifr.ifr_flags |= flags;
769 if (ioctl(s, SIOCSIFFLAGS, &ifr) < 0) {
779 eth_dev_promiscuous_enable(struct rte_eth_dev *dev)
781 struct pmd_internals *internals = dev->data->dev_private;
783 return eth_dev_change_flags(internals->if_name, IFF_PROMISC, ~0);
787 eth_dev_promiscuous_disable(struct rte_eth_dev *dev)
789 struct pmd_internals *internals = dev->data->dev_private;
791 return eth_dev_change_flags(internals->if_name, 0, ~IFF_PROMISC);
794 static const struct eth_dev_ops ops = {
795 .dev_start = eth_dev_start,
796 .dev_stop = eth_dev_stop,
797 .dev_close = eth_dev_close,
798 .dev_configure = eth_dev_configure,
799 .dev_infos_get = eth_dev_info,
800 .mtu_set = eth_dev_mtu_set,
801 .promiscuous_enable = eth_dev_promiscuous_enable,
802 .promiscuous_disable = eth_dev_promiscuous_disable,
803 .rx_queue_setup = eth_rx_queue_setup,
804 .tx_queue_setup = eth_tx_queue_setup,
805 .rx_queue_release = eth_queue_release,
806 .tx_queue_release = eth_queue_release,
807 .link_update = eth_link_update,
808 .stats_get = eth_stats_get,
809 .stats_reset = eth_stats_reset,
812 /** parse integer from integer argument */
814 parse_integer_arg(const char *key __rte_unused,
815 const char *value, void *extra_args)
817 int *i = (int *)extra_args;
820 *i = strtol(value, &end, 10);
822 AF_XDP_LOG(ERR, "Argument has to be positive.\n");
829 /** parse name argument */
831 parse_name_arg(const char *key __rte_unused,
832 const char *value, void *extra_args)
834 char *name = extra_args;
836 if (strnlen(value, IFNAMSIZ) > IFNAMSIZ - 1) {
837 AF_XDP_LOG(ERR, "Invalid name %s, should be less than %u bytes.\n",
842 strlcpy(name, value, IFNAMSIZ);
848 xdp_get_channels_info(const char *if_name, int *max_queues,
849 int *combined_queues)
851 struct ethtool_channels channels;
855 fd = socket(AF_INET, SOCK_DGRAM, 0);
859 channels.cmd = ETHTOOL_GCHANNELS;
860 ifr.ifr_data = (void *)&channels;
861 strncpy(ifr.ifr_name, if_name, IFNAMSIZ);
862 ret = ioctl(fd, SIOCETHTOOL, &ifr);
864 if (errno == EOPNOTSUPP) {
872 if (channels.max_combined == 0 || errno == EOPNOTSUPP) {
873 /* If the device says it has no channels, then all traffic
874 * is sent to a single stream, so max queues = 1.
877 *combined_queues = 1;
879 *max_queues = channels.max_combined;
880 *combined_queues = channels.combined_count;
889 parse_parameters(struct rte_kvargs *kvlist, char *if_name, int *start_queue,
890 int *queue_cnt, int *pmd_zc)
894 ret = rte_kvargs_process(kvlist, ETH_AF_XDP_IFACE_ARG,
895 &parse_name_arg, if_name);
899 ret = rte_kvargs_process(kvlist, ETH_AF_XDP_START_QUEUE_ARG,
900 &parse_integer_arg, start_queue);
904 ret = rte_kvargs_process(kvlist, ETH_AF_XDP_QUEUE_COUNT_ARG,
905 &parse_integer_arg, queue_cnt);
906 if (ret < 0 || *queue_cnt <= 0) {
911 ret = rte_kvargs_process(kvlist, ETH_AF_XDP_PMD_ZC_ARG,
912 &parse_integer_arg, pmd_zc);
917 rte_kvargs_free(kvlist);
922 get_iface_info(const char *if_name,
923 struct rte_ether_addr *eth_addr,
927 int sock = socket(AF_INET, SOCK_DGRAM, IPPROTO_IP);
932 strlcpy(ifr.ifr_name, if_name, IFNAMSIZ);
933 if (ioctl(sock, SIOCGIFINDEX, &ifr))
936 *if_index = ifr.ifr_ifindex;
938 if (ioctl(sock, SIOCGIFHWADDR, &ifr))
941 rte_memcpy(eth_addr, ifr.ifr_hwaddr.sa_data, RTE_ETHER_ADDR_LEN);
951 static struct rte_eth_dev *
952 init_internals(struct rte_vdev_device *dev, const char *if_name,
953 int start_queue_idx, int queue_cnt, int pmd_zc)
955 const char *name = rte_vdev_device_name(dev);
956 const unsigned int numa_node = dev->device.numa_node;
957 struct pmd_internals *internals;
958 struct rte_eth_dev *eth_dev;
962 internals = rte_zmalloc_socket(name, sizeof(*internals), 0, numa_node);
963 if (internals == NULL)
966 internals->start_queue_idx = start_queue_idx;
967 internals->queue_cnt = queue_cnt;
968 internals->pmd_zc = pmd_zc;
969 strlcpy(internals->if_name, if_name, IFNAMSIZ);
971 if (xdp_get_channels_info(if_name, &internals->max_queue_cnt,
972 &internals->combined_queue_cnt)) {
973 AF_XDP_LOG(ERR, "Failed to get channel info of interface: %s\n",
975 goto err_free_internals;
978 if (queue_cnt > internals->combined_queue_cnt) {
979 AF_XDP_LOG(ERR, "Specified queue count %d is larger than combined queue count %d.\n",
980 queue_cnt, internals->combined_queue_cnt);
981 goto err_free_internals;
984 internals->rx_queues = rte_zmalloc_socket(NULL,
985 sizeof(struct pkt_rx_queue) * queue_cnt,
987 if (internals->rx_queues == NULL) {
988 AF_XDP_LOG(ERR, "Failed to allocate memory for rx queues.\n");
989 goto err_free_internals;
992 internals->tx_queues = rte_zmalloc_socket(NULL,
993 sizeof(struct pkt_tx_queue) * queue_cnt,
995 if (internals->tx_queues == NULL) {
996 AF_XDP_LOG(ERR, "Failed to allocate memory for tx queues.\n");
999 for (i = 0; i < queue_cnt; i++) {
1000 internals->tx_queues[i].pair = &internals->rx_queues[i];
1001 internals->rx_queues[i].pair = &internals->tx_queues[i];
1002 internals->rx_queues[i].xsk_queue_idx = start_queue_idx + i;
1003 internals->tx_queues[i].xsk_queue_idx = start_queue_idx + i;
1006 ret = get_iface_info(if_name, &internals->eth_addr,
1007 &internals->if_index);
1011 eth_dev = rte_eth_vdev_allocate(dev, 0);
1012 if (eth_dev == NULL)
1015 eth_dev->data->dev_private = internals;
1016 eth_dev->data->dev_link = pmd_link;
1017 eth_dev->data->mac_addrs = &internals->eth_addr;
1018 eth_dev->dev_ops = &ops;
1019 eth_dev->rx_pkt_burst = eth_af_xdp_rx;
1020 eth_dev->tx_pkt_burst = eth_af_xdp_tx;
1021 /* Let rte_eth_dev_close() release the port resources. */
1022 eth_dev->data->dev_flags |= RTE_ETH_DEV_CLOSE_REMOVE;
1024 if (internals->pmd_zc)
1025 AF_XDP_LOG(INFO, "Zero copy between umem and mbuf enabled.\n");
1030 rte_free(internals->tx_queues);
1032 rte_free(internals->rx_queues);
1034 rte_free(internals);
1039 rte_pmd_af_xdp_probe(struct rte_vdev_device *dev)
1041 struct rte_kvargs *kvlist;
1042 char if_name[IFNAMSIZ] = {'\0'};
1043 int xsk_start_queue_idx = ETH_AF_XDP_DFLT_START_QUEUE_IDX;
1044 int xsk_queue_cnt = ETH_AF_XDP_DFLT_QUEUE_COUNT;
1045 struct rte_eth_dev *eth_dev = NULL;
1049 AF_XDP_LOG(INFO, "Initializing pmd_af_xdp for %s\n",
1050 rte_vdev_device_name(dev));
1052 name = rte_vdev_device_name(dev);
1053 if (rte_eal_process_type() == RTE_PROC_SECONDARY &&
1054 strlen(rte_vdev_device_args(dev)) == 0) {
1055 eth_dev = rte_eth_dev_attach_secondary(name);
1056 if (eth_dev == NULL) {
1057 AF_XDP_LOG(ERR, "Failed to probe %s\n", name);
1060 eth_dev->dev_ops = &ops;
1061 rte_eth_dev_probing_finish(eth_dev);
1065 kvlist = rte_kvargs_parse(rte_vdev_device_args(dev), valid_arguments);
1066 if (kvlist == NULL) {
1067 AF_XDP_LOG(ERR, "Invalid kvargs key\n");
1071 if (dev->device.numa_node == SOCKET_ID_ANY)
1072 dev->device.numa_node = rte_socket_id();
1074 if (parse_parameters(kvlist, if_name, &xsk_start_queue_idx,
1075 &xsk_queue_cnt, &pmd_zc) < 0) {
1076 AF_XDP_LOG(ERR, "Invalid kvargs value\n");
1080 if (strlen(if_name) == 0) {
1081 AF_XDP_LOG(ERR, "Network interface must be specified\n");
1085 eth_dev = init_internals(dev, if_name, xsk_start_queue_idx,
1086 xsk_queue_cnt, pmd_zc);
1087 if (eth_dev == NULL) {
1088 AF_XDP_LOG(ERR, "Failed to init internals\n");
1092 rte_eth_dev_probing_finish(eth_dev);
1098 rte_pmd_af_xdp_remove(struct rte_vdev_device *dev)
1100 struct rte_eth_dev *eth_dev = NULL;
1102 AF_XDP_LOG(INFO, "Removing AF_XDP ethdev on numa socket %u\n",
1108 /* find the ethdev entry */
1109 eth_dev = rte_eth_dev_allocated(rte_vdev_device_name(dev));
1110 if (eth_dev == NULL)
1113 eth_dev_close(eth_dev);
1114 rte_eth_dev_release_port(eth_dev);
1120 static struct rte_vdev_driver pmd_af_xdp_drv = {
1121 .probe = rte_pmd_af_xdp_probe,
1122 .remove = rte_pmd_af_xdp_remove,
1125 RTE_PMD_REGISTER_VDEV(net_af_xdp, pmd_af_xdp_drv);
1126 RTE_PMD_REGISTER_PARAM_STRING(net_af_xdp,
1128 "start_queue=<int> "
1129 "queue_count=<int> "
1130 "pmd_zero_copy=<0|1>");
1132 RTE_INIT(af_xdp_init_log)
1134 af_xdp_logtype = rte_log_register("pmd.net.af_xdp");
1135 if (af_xdp_logtype >= 0)
1136 rte_log_set_level(af_xdp_logtype, RTE_LOG_NOTICE);