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>
31 #include <rte_ether.h>
32 #include <rte_lcore.h>
34 #include <rte_memory.h>
35 #include <rte_memzone.h>
36 #include <rte_mempool.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_DFLT_NUM_DESCS XSK_RING_CONS__DEFAULT_NUM_DESCS
62 #define ETH_AF_XDP_DFLT_START_QUEUE_IDX 0
63 #define ETH_AF_XDP_DFLT_QUEUE_COUNT 1
65 #define ETH_AF_XDP_RX_BATCH_SIZE 32
66 #define ETH_AF_XDP_TX_BATCH_SIZE 32
69 struct xsk_umem_info {
70 struct xsk_ring_prod fq;
71 struct xsk_ring_cons cq;
72 struct xsk_umem *umem;
73 struct rte_ring *buf_ring;
74 const struct rte_memzone *mz;
75 struct rte_mempool *mb_pool;
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;
104 struct pkt_tx_queue {
105 struct xsk_ring_prod tx;
106 struct xsk_umem_info *umem;
108 struct tx_stats stats;
110 struct pkt_rx_queue *pair;
114 struct pmd_internals {
116 char if_name[IFNAMSIZ];
120 int combined_queue_cnt;
122 struct rte_ether_addr eth_addr;
124 struct pkt_rx_queue *rx_queues;
125 struct pkt_tx_queue *tx_queues;
128 #define ETH_AF_XDP_IFACE_ARG "iface"
129 #define ETH_AF_XDP_START_QUEUE_ARG "start_queue"
130 #define ETH_AF_XDP_QUEUE_COUNT_ARG "queue_count"
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,
139 static const struct rte_eth_link pmd_link = {
140 .link_speed = ETH_SPEED_NUM_10G,
141 .link_duplex = ETH_LINK_FULL_DUPLEX,
142 .link_status = ETH_LINK_DOWN,
143 .link_autoneg = ETH_LINK_AUTONEG
146 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
148 reserve_fill_queue_zc(struct xsk_umem_info *umem, uint16_t reserve_size,
149 struct rte_mbuf **bufs)
151 struct xsk_ring_prod *fq = &umem->fq;
155 if (unlikely(!xsk_ring_prod__reserve(fq, reserve_size, &idx))) {
156 for (i = 0; i < reserve_size; i++)
157 rte_pktmbuf_free(bufs[i]);
158 AF_XDP_LOG(DEBUG, "Failed to reserve enough fq descs.\n");
162 for (i = 0; i < reserve_size; i++) {
166 fq_addr = xsk_ring_prod__fill_addr(fq, idx++);
167 addr = (uint64_t)bufs[i] - (uint64_t)umem->buffer -
168 umem->mb_pool->header_size;
172 xsk_ring_prod__submit(fq, reserve_size);
178 reserve_fill_queue_cp(struct xsk_umem_info *umem, uint16_t reserve_size,
179 struct rte_mbuf **bufs __rte_unused)
181 struct xsk_ring_prod *fq = &umem->fq;
182 void *addrs[reserve_size];
186 if (rte_ring_dequeue_bulk(umem->buf_ring, addrs, reserve_size, NULL)
188 AF_XDP_LOG(DEBUG, "Failed to get enough buffers for fq.\n");
192 if (unlikely(!xsk_ring_prod__reserve(fq, reserve_size, &idx))) {
193 AF_XDP_LOG(DEBUG, "Failed to reserve enough fq descs.\n");
194 rte_ring_enqueue_bulk(umem->buf_ring, addrs,
199 for (i = 0; i < reserve_size; i++) {
202 fq_addr = xsk_ring_prod__fill_addr(fq, idx++);
203 *fq_addr = (uint64_t)addrs[i];
206 xsk_ring_prod__submit(fq, reserve_size);
213 reserve_fill_queue(struct xsk_umem_info *umem, uint16_t reserve_size,
214 struct rte_mbuf **bufs)
216 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
217 return reserve_fill_queue_zc(umem, reserve_size, bufs);
219 return reserve_fill_queue_cp(umem, reserve_size, bufs);
223 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
225 af_xdp_rx_zc(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
227 struct pkt_rx_queue *rxq = queue;
228 struct xsk_ring_cons *rx = &rxq->rx;
229 struct xsk_umem_info *umem = rxq->umem;
231 unsigned long rx_bytes = 0;
233 struct rte_mbuf *fq_bufs[ETH_AF_XDP_RX_BATCH_SIZE];
235 /* allocate bufs for fill queue replenishment after rx */
236 if (rte_pktmbuf_alloc_bulk(umem->mb_pool, fq_bufs, nb_pkts)) {
238 "Failed to get enough buffers for fq.\n");
242 rcvd = xsk_ring_cons__peek(rx, nb_pkts, &idx_rx);
245 #if defined(XDP_USE_NEED_WAKEUP)
246 if (xsk_ring_prod__needs_wakeup(&umem->fq))
247 (void)poll(rxq->fds, 1, 1000);
253 for (i = 0; i < rcvd; i++) {
254 const struct xdp_desc *desc;
259 desc = xsk_ring_cons__rx_desc(rx, idx_rx++);
263 offset = xsk_umem__extract_offset(addr);
264 addr = xsk_umem__extract_addr(addr);
266 bufs[i] = (struct rte_mbuf *)
267 xsk_umem__get_data(umem->buffer, addr +
268 umem->mb_pool->header_size);
269 bufs[i]->data_off = offset - sizeof(struct rte_mbuf) -
270 rte_pktmbuf_priv_size(umem->mb_pool) -
271 umem->mb_pool->header_size;
273 rte_pktmbuf_pkt_len(bufs[i]) = len;
274 rte_pktmbuf_data_len(bufs[i]) = len;
278 xsk_ring_cons__release(rx, rcvd);
280 (void)reserve_fill_queue(umem, rcvd, fq_bufs);
283 rxq->stats.rx_pkts += rcvd;
284 rxq->stats.rx_bytes += rx_bytes;
288 rte_mempool_put_bulk(umem->mb_pool, (void **)&fq_bufs[rcvd],
295 af_xdp_rx_cp(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
297 struct pkt_rx_queue *rxq = queue;
298 struct xsk_ring_cons *rx = &rxq->rx;
299 struct xsk_umem_info *umem = rxq->umem;
300 struct xsk_ring_prod *fq = &umem->fq;
302 unsigned long rx_bytes = 0;
304 uint32_t free_thresh = fq->size >> 1;
305 struct rte_mbuf *mbufs[ETH_AF_XDP_RX_BATCH_SIZE];
307 if (unlikely(rte_pktmbuf_alloc_bulk(rxq->mb_pool, mbufs, nb_pkts) != 0))
310 rcvd = xsk_ring_cons__peek(rx, nb_pkts, &idx_rx);
312 #if defined(XDP_USE_NEED_WAKEUP)
313 if (xsk_ring_prod__needs_wakeup(fq))
314 (void)poll(rxq->fds, 1, 1000);
320 if (xsk_prod_nb_free(fq, free_thresh) >= free_thresh)
321 (void)reserve_fill_queue(umem, ETH_AF_XDP_RX_BATCH_SIZE, NULL);
323 for (i = 0; i < rcvd; i++) {
324 const struct xdp_desc *desc;
329 desc = xsk_ring_cons__rx_desc(rx, idx_rx++);
332 pkt = xsk_umem__get_data(rxq->umem->mz->addr, addr);
334 rte_memcpy(rte_pktmbuf_mtod(mbufs[i], void *), pkt, len);
335 rte_ring_enqueue(umem->buf_ring, (void *)addr);
336 rte_pktmbuf_pkt_len(mbufs[i]) = len;
337 rte_pktmbuf_data_len(mbufs[i]) = len;
342 xsk_ring_cons__release(rx, rcvd);
345 rxq->stats.rx_pkts += rcvd;
346 rxq->stats.rx_bytes += rx_bytes;
350 rte_mempool_put_bulk(rxq->mb_pool, (void **)&mbufs[rcvd],
358 eth_af_xdp_rx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
360 nb_pkts = RTE_MIN(nb_pkts, ETH_AF_XDP_RX_BATCH_SIZE);
362 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
363 return af_xdp_rx_zc(queue, bufs, nb_pkts);
365 return af_xdp_rx_cp(queue, bufs, nb_pkts);
370 pull_umem_cq(struct xsk_umem_info *umem, int size)
372 struct xsk_ring_cons *cq = &umem->cq;
376 n = xsk_ring_cons__peek(cq, size, &idx_cq);
378 for (i = 0; i < n; i++) {
380 addr = *xsk_ring_cons__comp_addr(cq, idx_cq++);
381 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
382 addr = xsk_umem__extract_addr(addr);
383 rte_pktmbuf_free((struct rte_mbuf *)
384 xsk_umem__get_data(umem->buffer,
385 addr + umem->mb_pool->header_size));
387 rte_ring_enqueue(umem->buf_ring, (void *)addr);
391 xsk_ring_cons__release(cq, n);
395 kick_tx(struct pkt_tx_queue *txq)
397 struct xsk_umem_info *umem = txq->umem;
399 #if defined(XDP_USE_NEED_WAKEUP)
400 if (xsk_ring_prod__needs_wakeup(&txq->tx))
402 while (send(xsk_socket__fd(txq->pair->xsk), NULL,
403 0, MSG_DONTWAIT) < 0) {
404 /* some thing unexpected */
405 if (errno != EBUSY && errno != EAGAIN && errno != EINTR)
408 /* pull from completion queue to leave more space */
410 pull_umem_cq(umem, ETH_AF_XDP_TX_BATCH_SIZE);
412 #ifndef XDP_UMEM_UNALIGNED_CHUNK_FLAG
413 pull_umem_cq(umem, ETH_AF_XDP_TX_BATCH_SIZE);
417 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
419 af_xdp_tx_zc(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
421 struct pkt_tx_queue *txq = queue;
422 struct xsk_umem_info *umem = txq->umem;
423 struct rte_mbuf *mbuf;
424 unsigned long tx_bytes = 0;
428 struct xdp_desc *desc;
429 uint64_t addr, offset;
431 pull_umem_cq(umem, nb_pkts);
433 for (i = 0; i < nb_pkts; i++) {
436 if (mbuf->pool == umem->mb_pool) {
437 if (!xsk_ring_prod__reserve(&txq->tx, 1, &idx_tx)) {
441 desc = xsk_ring_prod__tx_desc(&txq->tx, idx_tx);
442 desc->len = mbuf->pkt_len;
443 addr = (uint64_t)mbuf - (uint64_t)umem->buffer -
444 umem->mb_pool->header_size;
445 offset = rte_pktmbuf_mtod(mbuf, uint64_t) -
447 umem->mb_pool->header_size;
448 offset = offset << XSK_UNALIGNED_BUF_OFFSET_SHIFT;
449 desc->addr = addr | offset;
452 struct rte_mbuf *local_mbuf =
453 rte_pktmbuf_alloc(umem->mb_pool);
456 if (local_mbuf == NULL)
459 if (!xsk_ring_prod__reserve(&txq->tx, 1, &idx_tx)) {
460 rte_pktmbuf_free(local_mbuf);
465 desc = xsk_ring_prod__tx_desc(&txq->tx, idx_tx);
466 desc->len = mbuf->pkt_len;
468 addr = (uint64_t)local_mbuf - (uint64_t)umem->buffer -
469 umem->mb_pool->header_size;
470 offset = rte_pktmbuf_mtod(local_mbuf, uint64_t) -
471 (uint64_t)local_mbuf +
472 umem->mb_pool->header_size;
473 pkt = xsk_umem__get_data(umem->buffer, addr + offset);
474 offset = offset << XSK_UNALIGNED_BUF_OFFSET_SHIFT;
475 desc->addr = addr | offset;
476 rte_memcpy(pkt, rte_pktmbuf_mtod(mbuf, void *),
478 rte_pktmbuf_free(mbuf);
482 tx_bytes += mbuf->pkt_len;
488 xsk_ring_prod__submit(&txq->tx, count);
490 txq->stats.tx_pkts += count;
491 txq->stats.tx_bytes += tx_bytes;
492 txq->stats.tx_dropped += nb_pkts - count;
498 af_xdp_tx_cp(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
500 struct pkt_tx_queue *txq = queue;
501 struct xsk_umem_info *umem = txq->umem;
502 struct rte_mbuf *mbuf;
503 void *addrs[ETH_AF_XDP_TX_BATCH_SIZE];
504 unsigned long tx_bytes = 0;
508 nb_pkts = RTE_MIN(nb_pkts, ETH_AF_XDP_TX_BATCH_SIZE);
510 pull_umem_cq(umem, nb_pkts);
512 nb_pkts = rte_ring_dequeue_bulk(umem->buf_ring, addrs,
517 if (xsk_ring_prod__reserve(&txq->tx, nb_pkts, &idx_tx) != nb_pkts) {
519 rte_ring_enqueue_bulk(umem->buf_ring, addrs, nb_pkts, NULL);
523 for (i = 0; i < nb_pkts; i++) {
524 struct xdp_desc *desc;
527 desc = xsk_ring_prod__tx_desc(&txq->tx, idx_tx + i);
529 desc->len = mbuf->pkt_len;
531 desc->addr = (uint64_t)addrs[i];
532 pkt = xsk_umem__get_data(umem->mz->addr,
534 rte_memcpy(pkt, rte_pktmbuf_mtod(mbuf, void *), desc->len);
535 tx_bytes += mbuf->pkt_len;
536 rte_pktmbuf_free(mbuf);
539 xsk_ring_prod__submit(&txq->tx, nb_pkts);
543 txq->stats.tx_pkts += nb_pkts;
544 txq->stats.tx_bytes += tx_bytes;
551 eth_af_xdp_tx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
553 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
554 return af_xdp_tx_zc(queue, bufs, nb_pkts);
556 return af_xdp_tx_cp(queue, bufs, nb_pkts);
561 eth_dev_start(struct rte_eth_dev *dev)
563 dev->data->dev_link.link_status = ETH_LINK_UP;
568 /* This function gets called when the current port gets stopped. */
570 eth_dev_stop(struct rte_eth_dev *dev)
572 dev->data->dev_link.link_status = ETH_LINK_DOWN;
576 eth_dev_configure(struct rte_eth_dev *dev)
578 /* rx/tx must be paired */
579 if (dev->data->nb_rx_queues != dev->data->nb_tx_queues)
586 eth_dev_info(struct rte_eth_dev *dev, struct rte_eth_dev_info *dev_info)
588 struct pmd_internals *internals = dev->data->dev_private;
590 dev_info->if_index = internals->if_index;
591 dev_info->max_mac_addrs = 1;
592 dev_info->max_rx_pktlen = ETH_FRAME_LEN;
593 dev_info->max_rx_queues = internals->queue_cnt;
594 dev_info->max_tx_queues = internals->queue_cnt;
596 dev_info->min_mtu = RTE_ETHER_MIN_MTU;
597 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
598 dev_info->max_mtu = getpagesize() -
599 sizeof(struct rte_mempool_objhdr) -
600 sizeof(struct rte_mbuf) -
601 RTE_PKTMBUF_HEADROOM - XDP_PACKET_HEADROOM;
603 dev_info->max_mtu = ETH_AF_XDP_FRAME_SIZE - XDP_PACKET_HEADROOM;
606 dev_info->default_rxportconf.nb_queues = 1;
607 dev_info->default_txportconf.nb_queues = 1;
608 dev_info->default_rxportconf.ring_size = ETH_AF_XDP_DFLT_NUM_DESCS;
609 dev_info->default_txportconf.ring_size = ETH_AF_XDP_DFLT_NUM_DESCS;
615 eth_stats_get(struct rte_eth_dev *dev, struct rte_eth_stats *stats)
617 struct pmd_internals *internals = dev->data->dev_private;
618 struct xdp_statistics xdp_stats;
619 struct pkt_rx_queue *rxq;
620 struct pkt_tx_queue *txq;
624 for (i = 0; i < dev->data->nb_rx_queues; i++) {
625 optlen = sizeof(struct xdp_statistics);
626 rxq = &internals->rx_queues[i];
628 stats->q_ipackets[i] = rxq->stats.rx_pkts;
629 stats->q_ibytes[i] = rxq->stats.rx_bytes;
631 stats->q_opackets[i] = txq->stats.tx_pkts;
632 stats->q_obytes[i] = txq->stats.tx_bytes;
634 stats->ipackets += stats->q_ipackets[i];
635 stats->ibytes += stats->q_ibytes[i];
636 stats->imissed += rxq->stats.rx_dropped;
637 stats->oerrors += txq->stats.tx_dropped;
638 ret = getsockopt(xsk_socket__fd(rxq->xsk), SOL_XDP,
639 XDP_STATISTICS, &xdp_stats, &optlen);
641 AF_XDP_LOG(ERR, "getsockopt() failed for XDP_STATISTICS.\n");
644 stats->imissed += xdp_stats.rx_dropped;
646 stats->opackets += stats->q_opackets[i];
647 stats->obytes += stats->q_obytes[i];
654 eth_stats_reset(struct rte_eth_dev *dev)
656 struct pmd_internals *internals = dev->data->dev_private;
659 for (i = 0; i < internals->queue_cnt; i++) {
660 memset(&internals->rx_queues[i].stats, 0,
661 sizeof(struct rx_stats));
662 memset(&internals->tx_queues[i].stats, 0,
663 sizeof(struct tx_stats));
670 remove_xdp_program(struct pmd_internals *internals)
672 uint32_t curr_prog_id = 0;
674 if (bpf_get_link_xdp_id(internals->if_index, &curr_prog_id,
675 XDP_FLAGS_UPDATE_IF_NOEXIST)) {
676 AF_XDP_LOG(ERR, "bpf_get_link_xdp_id failed\n");
679 bpf_set_link_xdp_fd(internals->if_index, -1,
680 XDP_FLAGS_UPDATE_IF_NOEXIST);
684 xdp_umem_destroy(struct xsk_umem_info *umem)
686 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
687 rte_mempool_free(umem->mb_pool);
688 umem->mb_pool = NULL;
690 rte_memzone_free(umem->mz);
693 rte_ring_free(umem->buf_ring);
694 umem->buf_ring = NULL;
702 eth_dev_close(struct rte_eth_dev *dev)
704 struct pmd_internals *internals = dev->data->dev_private;
705 struct pkt_rx_queue *rxq;
708 AF_XDP_LOG(INFO, "Closing AF_XDP ethdev on numa socket %u\n",
711 for (i = 0; i < internals->queue_cnt; i++) {
712 rxq = &internals->rx_queues[i];
713 if (rxq->umem == NULL)
715 xsk_socket__delete(rxq->xsk);
716 (void)xsk_umem__delete(rxq->umem->umem);
717 xdp_umem_destroy(rxq->umem);
719 /* free pkt_tx_queue */
725 * MAC is not allocated dynamically, setting it to NULL would prevent
726 * from releasing it in rte_eth_dev_release_port.
728 dev->data->mac_addrs = NULL;
730 remove_xdp_program(internals);
734 eth_queue_release(void *q __rte_unused)
739 eth_link_update(struct rte_eth_dev *dev __rte_unused,
740 int wait_to_complete __rte_unused)
745 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
746 static inline uint64_t get_base_addr(struct rte_mempool *mp)
748 struct rte_mempool_memhdr *memhdr;
750 memhdr = STAILQ_FIRST(&mp->mem_list);
751 return (uint64_t)memhdr->addr & ~(getpagesize() - 1);
755 xsk_umem_info *xdp_umem_configure(struct pmd_internals *internals __rte_unused,
756 struct pkt_rx_queue *rxq)
758 struct xsk_umem_info *umem;
760 struct xsk_umem_config usr_config = {
761 .fill_size = ETH_AF_XDP_DFLT_NUM_DESCS,
762 .comp_size = ETH_AF_XDP_DFLT_NUM_DESCS,
763 .flags = XDP_UMEM_UNALIGNED_CHUNK_FLAG};
764 void *base_addr = NULL;
765 struct rte_mempool *mb_pool = rxq->mb_pool;
767 usr_config.frame_size = rte_mempool_calc_obj_size(mb_pool->elt_size,
770 usr_config.frame_headroom = mb_pool->header_size +
771 sizeof(struct rte_mbuf) +
772 rte_pktmbuf_priv_size(mb_pool) +
773 RTE_PKTMBUF_HEADROOM;
775 umem = rte_zmalloc_socket("umem", sizeof(*umem), 0, rte_socket_id());
777 AF_XDP_LOG(ERR, "Failed to allocate umem info");
781 umem->mb_pool = mb_pool;
782 base_addr = (void *)get_base_addr(mb_pool);
784 ret = xsk_umem__create(&umem->umem, base_addr,
785 mb_pool->populated_size * usr_config.frame_size,
786 &umem->fq, &umem->cq,
790 AF_XDP_LOG(ERR, "Failed to create umem");
793 umem->buffer = base_addr;
797 xsk_umem_info *xdp_umem_configure(struct pmd_internals *internals,
798 struct pkt_rx_queue *rxq)
800 struct xsk_umem_info *umem;
801 const struct rte_memzone *mz;
802 struct xsk_umem_config usr_config = {
803 .fill_size = ETH_AF_XDP_DFLT_NUM_DESCS,
804 .comp_size = ETH_AF_XDP_DFLT_NUM_DESCS,
805 .frame_size = ETH_AF_XDP_FRAME_SIZE,
806 .frame_headroom = 0 };
807 char ring_name[RTE_RING_NAMESIZE];
808 char mz_name[RTE_MEMZONE_NAMESIZE];
812 umem = rte_zmalloc_socket("umem", sizeof(*umem), 0, rte_socket_id());
814 AF_XDP_LOG(ERR, "Failed to allocate umem info");
818 snprintf(ring_name, sizeof(ring_name), "af_xdp_ring_%s_%u",
819 internals->if_name, rxq->xsk_queue_idx);
820 umem->buf_ring = rte_ring_create(ring_name,
821 ETH_AF_XDP_NUM_BUFFERS,
823 RING_F_SP_ENQ | RING_F_SC_DEQ);
824 if (umem->buf_ring == NULL) {
825 AF_XDP_LOG(ERR, "Failed to create rte_ring\n");
829 for (i = 0; i < ETH_AF_XDP_NUM_BUFFERS; i++)
830 rte_ring_enqueue(umem->buf_ring,
831 (void *)(i * ETH_AF_XDP_FRAME_SIZE));
833 snprintf(mz_name, sizeof(mz_name), "af_xdp_umem_%s_%u",
834 internals->if_name, rxq->xsk_queue_idx);
835 mz = rte_memzone_reserve_aligned(mz_name,
836 ETH_AF_XDP_NUM_BUFFERS * ETH_AF_XDP_FRAME_SIZE,
837 rte_socket_id(), RTE_MEMZONE_IOVA_CONTIG,
840 AF_XDP_LOG(ERR, "Failed to reserve memzone for af_xdp umem.\n");
844 ret = xsk_umem__create(&umem->umem, mz->addr,
845 ETH_AF_XDP_NUM_BUFFERS * ETH_AF_XDP_FRAME_SIZE,
846 &umem->fq, &umem->cq,
850 AF_XDP_LOG(ERR, "Failed to create umem");
859 xdp_umem_destroy(umem);
864 xsk_configure(struct pmd_internals *internals, struct pkt_rx_queue *rxq,
867 struct xsk_socket_config cfg;
868 struct pkt_tx_queue *txq = rxq->pair;
870 int reserve_size = ETH_AF_XDP_DFLT_NUM_DESCS / 2;
871 struct rte_mbuf *fq_bufs[reserve_size];
873 rxq->umem = xdp_umem_configure(internals, rxq);
874 if (rxq->umem == NULL)
876 txq->umem = rxq->umem;
878 cfg.rx_size = ring_size;
879 cfg.tx_size = ring_size;
880 cfg.libbpf_flags = 0;
881 cfg.xdp_flags = XDP_FLAGS_UPDATE_IF_NOEXIST;
884 #if defined(XDP_USE_NEED_WAKEUP)
885 cfg.bind_flags |= XDP_USE_NEED_WAKEUP;
888 ret = xsk_socket__create(&rxq->xsk, internals->if_name,
889 rxq->xsk_queue_idx, rxq->umem->umem, &rxq->rx,
892 AF_XDP_LOG(ERR, "Failed to create xsk socket.\n");
896 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
897 if (rte_pktmbuf_alloc_bulk(rxq->umem->mb_pool, fq_bufs, reserve_size)) {
898 AF_XDP_LOG(DEBUG, "Failed to get enough buffers for fq.\n");
902 ret = reserve_fill_queue(rxq->umem, reserve_size, fq_bufs);
904 xsk_socket__delete(rxq->xsk);
905 AF_XDP_LOG(ERR, "Failed to reserve fill queue.\n");
912 xdp_umem_destroy(rxq->umem);
918 eth_rx_queue_setup(struct rte_eth_dev *dev,
919 uint16_t rx_queue_id,
921 unsigned int socket_id __rte_unused,
922 const struct rte_eth_rxconf *rx_conf __rte_unused,
923 struct rte_mempool *mb_pool)
925 struct pmd_internals *internals = dev->data->dev_private;
926 struct pkt_rx_queue *rxq;
929 rxq = &internals->rx_queues[rx_queue_id];
931 AF_XDP_LOG(INFO, "Set up rx queue, rx queue id: %d, xsk queue id: %d\n",
932 rx_queue_id, rxq->xsk_queue_idx);
934 #ifndef XDP_UMEM_UNALIGNED_CHUNK_FLAG
935 uint32_t buf_size, data_size;
937 /* Now get the space available for data in the mbuf */
938 buf_size = rte_pktmbuf_data_room_size(mb_pool) -
939 RTE_PKTMBUF_HEADROOM;
940 data_size = ETH_AF_XDP_FRAME_SIZE;
942 if (data_size > buf_size) {
943 AF_XDP_LOG(ERR, "%s: %d bytes will not fit in mbuf (%d bytes)\n",
944 dev->device->name, data_size, buf_size);
950 rxq->mb_pool = mb_pool;
952 if (xsk_configure(internals, rxq, nb_rx_desc)) {
953 AF_XDP_LOG(ERR, "Failed to configure xdp socket\n");
958 rxq->fds[0].fd = xsk_socket__fd(rxq->xsk);
959 rxq->fds[0].events = POLLIN;
961 dev->data->rx_queues[rx_queue_id] = rxq;
969 eth_tx_queue_setup(struct rte_eth_dev *dev,
970 uint16_t tx_queue_id,
971 uint16_t nb_tx_desc __rte_unused,
972 unsigned int socket_id __rte_unused,
973 const struct rte_eth_txconf *tx_conf __rte_unused)
975 struct pmd_internals *internals = dev->data->dev_private;
976 struct pkt_tx_queue *txq;
978 txq = &internals->tx_queues[tx_queue_id];
980 dev->data->tx_queues[tx_queue_id] = txq;
985 eth_dev_mtu_set(struct rte_eth_dev *dev, uint16_t mtu)
987 struct pmd_internals *internals = dev->data->dev_private;
988 struct ifreq ifr = { .ifr_mtu = mtu };
992 s = socket(PF_INET, SOCK_DGRAM, 0);
996 strlcpy(ifr.ifr_name, internals->if_name, IFNAMSIZ);
997 ret = ioctl(s, SIOCSIFMTU, &ifr);
1000 return (ret < 0) ? -errno : 0;
1004 eth_dev_change_flags(char *if_name, uint32_t flags, uint32_t mask)
1010 s = socket(PF_INET, SOCK_DGRAM, 0);
1014 strlcpy(ifr.ifr_name, if_name, IFNAMSIZ);
1015 if (ioctl(s, SIOCGIFFLAGS, &ifr) < 0) {
1019 ifr.ifr_flags &= mask;
1020 ifr.ifr_flags |= flags;
1021 if (ioctl(s, SIOCSIFFLAGS, &ifr) < 0) {
1031 eth_dev_promiscuous_enable(struct rte_eth_dev *dev)
1033 struct pmd_internals *internals = dev->data->dev_private;
1035 return eth_dev_change_flags(internals->if_name, IFF_PROMISC, ~0);
1039 eth_dev_promiscuous_disable(struct rte_eth_dev *dev)
1041 struct pmd_internals *internals = dev->data->dev_private;
1043 return eth_dev_change_flags(internals->if_name, 0, ~IFF_PROMISC);
1046 static const struct eth_dev_ops ops = {
1047 .dev_start = eth_dev_start,
1048 .dev_stop = eth_dev_stop,
1049 .dev_close = eth_dev_close,
1050 .dev_configure = eth_dev_configure,
1051 .dev_infos_get = eth_dev_info,
1052 .mtu_set = eth_dev_mtu_set,
1053 .promiscuous_enable = eth_dev_promiscuous_enable,
1054 .promiscuous_disable = eth_dev_promiscuous_disable,
1055 .rx_queue_setup = eth_rx_queue_setup,
1056 .tx_queue_setup = eth_tx_queue_setup,
1057 .rx_queue_release = eth_queue_release,
1058 .tx_queue_release = eth_queue_release,
1059 .link_update = eth_link_update,
1060 .stats_get = eth_stats_get,
1061 .stats_reset = eth_stats_reset,
1064 /** parse integer from integer argument */
1066 parse_integer_arg(const char *key __rte_unused,
1067 const char *value, void *extra_args)
1069 int *i = (int *)extra_args;
1072 *i = strtol(value, &end, 10);
1074 AF_XDP_LOG(ERR, "Argument has to be positive.\n");
1081 /** parse name argument */
1083 parse_name_arg(const char *key __rte_unused,
1084 const char *value, void *extra_args)
1086 char *name = extra_args;
1088 if (strnlen(value, IFNAMSIZ) > IFNAMSIZ - 1) {
1089 AF_XDP_LOG(ERR, "Invalid name %s, should be less than %u bytes.\n",
1094 strlcpy(name, value, IFNAMSIZ);
1100 xdp_get_channels_info(const char *if_name, int *max_queues,
1101 int *combined_queues)
1103 struct ethtool_channels channels;
1107 fd = socket(AF_INET, SOCK_DGRAM, 0);
1111 channels.cmd = ETHTOOL_GCHANNELS;
1112 ifr.ifr_data = (void *)&channels;
1113 strncpy(ifr.ifr_name, if_name, IFNAMSIZ);
1114 ret = ioctl(fd, SIOCETHTOOL, &ifr);
1116 if (errno == EOPNOTSUPP) {
1124 if (channels.max_combined == 0 || errno == EOPNOTSUPP) {
1125 /* If the device says it has no channels, then all traffic
1126 * is sent to a single stream, so max queues = 1.
1129 *combined_queues = 1;
1131 *max_queues = channels.max_combined;
1132 *combined_queues = channels.combined_count;
1141 parse_parameters(struct rte_kvargs *kvlist, char *if_name, int *start_queue,
1146 ret = rte_kvargs_process(kvlist, ETH_AF_XDP_IFACE_ARG,
1147 &parse_name_arg, if_name);
1151 ret = rte_kvargs_process(kvlist, ETH_AF_XDP_START_QUEUE_ARG,
1152 &parse_integer_arg, start_queue);
1156 ret = rte_kvargs_process(kvlist, ETH_AF_XDP_QUEUE_COUNT_ARG,
1157 &parse_integer_arg, queue_cnt);
1158 if (ret < 0 || *queue_cnt <= 0) {
1164 rte_kvargs_free(kvlist);
1169 get_iface_info(const char *if_name,
1170 struct rte_ether_addr *eth_addr,
1174 int sock = socket(AF_INET, SOCK_DGRAM, IPPROTO_IP);
1179 strlcpy(ifr.ifr_name, if_name, IFNAMSIZ);
1180 if (ioctl(sock, SIOCGIFINDEX, &ifr))
1183 *if_index = ifr.ifr_ifindex;
1185 if (ioctl(sock, SIOCGIFHWADDR, &ifr))
1188 rte_memcpy(eth_addr, ifr.ifr_hwaddr.sa_data, RTE_ETHER_ADDR_LEN);
1198 static struct rte_eth_dev *
1199 init_internals(struct rte_vdev_device *dev, const char *if_name,
1200 int start_queue_idx, int queue_cnt)
1202 const char *name = rte_vdev_device_name(dev);
1203 const unsigned int numa_node = dev->device.numa_node;
1204 struct pmd_internals *internals;
1205 struct rte_eth_dev *eth_dev;
1209 internals = rte_zmalloc_socket(name, sizeof(*internals), 0, numa_node);
1210 if (internals == NULL)
1213 internals->start_queue_idx = start_queue_idx;
1214 internals->queue_cnt = queue_cnt;
1215 strlcpy(internals->if_name, if_name, IFNAMSIZ);
1217 if (xdp_get_channels_info(if_name, &internals->max_queue_cnt,
1218 &internals->combined_queue_cnt)) {
1219 AF_XDP_LOG(ERR, "Failed to get channel info of interface: %s\n",
1221 goto err_free_internals;
1224 if (queue_cnt > internals->combined_queue_cnt) {
1225 AF_XDP_LOG(ERR, "Specified queue count %d is larger than combined queue count %d.\n",
1226 queue_cnt, internals->combined_queue_cnt);
1227 goto err_free_internals;
1230 internals->rx_queues = rte_zmalloc_socket(NULL,
1231 sizeof(struct pkt_rx_queue) * queue_cnt,
1233 if (internals->rx_queues == NULL) {
1234 AF_XDP_LOG(ERR, "Failed to allocate memory for rx queues.\n");
1235 goto err_free_internals;
1238 internals->tx_queues = rte_zmalloc_socket(NULL,
1239 sizeof(struct pkt_tx_queue) * queue_cnt,
1241 if (internals->tx_queues == NULL) {
1242 AF_XDP_LOG(ERR, "Failed to allocate memory for tx queues.\n");
1245 for (i = 0; i < queue_cnt; i++) {
1246 internals->tx_queues[i].pair = &internals->rx_queues[i];
1247 internals->rx_queues[i].pair = &internals->tx_queues[i];
1248 internals->rx_queues[i].xsk_queue_idx = start_queue_idx + i;
1249 internals->tx_queues[i].xsk_queue_idx = start_queue_idx + i;
1252 ret = get_iface_info(if_name, &internals->eth_addr,
1253 &internals->if_index);
1257 eth_dev = rte_eth_vdev_allocate(dev, 0);
1258 if (eth_dev == NULL)
1261 eth_dev->data->dev_private = internals;
1262 eth_dev->data->dev_link = pmd_link;
1263 eth_dev->data->mac_addrs = &internals->eth_addr;
1264 eth_dev->dev_ops = &ops;
1265 eth_dev->rx_pkt_burst = eth_af_xdp_rx;
1266 eth_dev->tx_pkt_burst = eth_af_xdp_tx;
1267 /* Let rte_eth_dev_close() release the port resources. */
1268 eth_dev->data->dev_flags |= RTE_ETH_DEV_CLOSE_REMOVE;
1270 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
1271 AF_XDP_LOG(INFO, "Zero copy between umem and mbuf enabled.\n");
1277 rte_free(internals->tx_queues);
1279 rte_free(internals->rx_queues);
1281 rte_free(internals);
1286 rte_pmd_af_xdp_probe(struct rte_vdev_device *dev)
1288 struct rte_kvargs *kvlist;
1289 char if_name[IFNAMSIZ] = {'\0'};
1290 int xsk_start_queue_idx = ETH_AF_XDP_DFLT_START_QUEUE_IDX;
1291 int xsk_queue_cnt = ETH_AF_XDP_DFLT_QUEUE_COUNT;
1292 struct rte_eth_dev *eth_dev = NULL;
1295 AF_XDP_LOG(INFO, "Initializing pmd_af_xdp for %s\n",
1296 rte_vdev_device_name(dev));
1298 name = rte_vdev_device_name(dev);
1299 if (rte_eal_process_type() == RTE_PROC_SECONDARY &&
1300 strlen(rte_vdev_device_args(dev)) == 0) {
1301 eth_dev = rte_eth_dev_attach_secondary(name);
1302 if (eth_dev == NULL) {
1303 AF_XDP_LOG(ERR, "Failed to probe %s\n", name);
1306 eth_dev->dev_ops = &ops;
1307 rte_eth_dev_probing_finish(eth_dev);
1311 kvlist = rte_kvargs_parse(rte_vdev_device_args(dev), valid_arguments);
1312 if (kvlist == NULL) {
1313 AF_XDP_LOG(ERR, "Invalid kvargs key\n");
1317 if (dev->device.numa_node == SOCKET_ID_ANY)
1318 dev->device.numa_node = rte_socket_id();
1320 if (parse_parameters(kvlist, if_name, &xsk_start_queue_idx,
1321 &xsk_queue_cnt) < 0) {
1322 AF_XDP_LOG(ERR, "Invalid kvargs value\n");
1326 if (strlen(if_name) == 0) {
1327 AF_XDP_LOG(ERR, "Network interface must be specified\n");
1331 eth_dev = init_internals(dev, if_name, xsk_start_queue_idx,
1333 if (eth_dev == NULL) {
1334 AF_XDP_LOG(ERR, "Failed to init internals\n");
1338 rte_eth_dev_probing_finish(eth_dev);
1344 rte_pmd_af_xdp_remove(struct rte_vdev_device *dev)
1346 struct rte_eth_dev *eth_dev = NULL;
1348 AF_XDP_LOG(INFO, "Removing AF_XDP ethdev on numa socket %u\n",
1354 /* find the ethdev entry */
1355 eth_dev = rte_eth_dev_allocated(rte_vdev_device_name(dev));
1356 if (eth_dev == NULL)
1359 eth_dev_close(eth_dev);
1360 rte_eth_dev_release_port(eth_dev);
1366 static struct rte_vdev_driver pmd_af_xdp_drv = {
1367 .probe = rte_pmd_af_xdp_probe,
1368 .remove = rte_pmd_af_xdp_remove,
1371 RTE_PMD_REGISTER_VDEV(net_af_xdp, pmd_af_xdp_drv);
1372 RTE_PMD_REGISTER_PARAM_STRING(net_af_xdp,
1374 "start_queue=<int> "
1375 "queue_count=<int> ");
1377 RTE_INIT(af_xdp_init_log)
1379 af_xdp_logtype = rte_log_register("pmd.net.af_xdp");
1380 if (af_xdp_logtype >= 0)
1381 rte_log_set_level(af_xdp_logtype, RTE_LOG_NOTICE);