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 RTE_LOG_REGISTER(af_xdp_logtype, pmd.net.af_xdp, NOTICE);
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 (xsk_prod_nb_free(fq, free_thresh) >= free_thresh)
308 (void)reserve_fill_queue(umem, ETH_AF_XDP_RX_BATCH_SIZE, NULL);
311 if (unlikely(rte_pktmbuf_alloc_bulk(rxq->mb_pool, mbufs, nb_pkts) != 0))
314 rcvd = xsk_ring_cons__peek(rx, nb_pkts, &idx_rx);
316 #if defined(XDP_USE_NEED_WAKEUP)
317 if (xsk_ring_prod__needs_wakeup(fq))
318 (void)poll(rxq->fds, 1, 1000);
324 for (i = 0; i < rcvd; i++) {
325 const struct xdp_desc *desc;
330 desc = xsk_ring_cons__rx_desc(rx, idx_rx++);
333 pkt = xsk_umem__get_data(rxq->umem->mz->addr, addr);
335 rte_memcpy(rte_pktmbuf_mtod(mbufs[i], void *), pkt, len);
336 rte_ring_enqueue(umem->buf_ring, (void *)addr);
337 rte_pktmbuf_pkt_len(mbufs[i]) = len;
338 rte_pktmbuf_data_len(mbufs[i]) = len;
343 xsk_ring_cons__release(rx, rcvd);
346 rxq->stats.rx_pkts += rcvd;
347 rxq->stats.rx_bytes += rx_bytes;
351 rte_mempool_put_bulk(rxq->mb_pool, (void **)&mbufs[rcvd],
359 eth_af_xdp_rx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
361 nb_pkts = RTE_MIN(nb_pkts, ETH_AF_XDP_RX_BATCH_SIZE);
363 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
364 return af_xdp_rx_zc(queue, bufs, nb_pkts);
366 return af_xdp_rx_cp(queue, bufs, nb_pkts);
371 pull_umem_cq(struct xsk_umem_info *umem, int size)
373 struct xsk_ring_cons *cq = &umem->cq;
377 n = xsk_ring_cons__peek(cq, size, &idx_cq);
379 for (i = 0; i < n; i++) {
381 addr = *xsk_ring_cons__comp_addr(cq, idx_cq++);
382 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
383 addr = xsk_umem__extract_addr(addr);
384 rte_pktmbuf_free((struct rte_mbuf *)
385 xsk_umem__get_data(umem->buffer,
386 addr + umem->mb_pool->header_size));
388 rte_ring_enqueue(umem->buf_ring, (void *)addr);
392 xsk_ring_cons__release(cq, n);
396 kick_tx(struct pkt_tx_queue *txq)
398 struct xsk_umem_info *umem = txq->umem;
400 pull_umem_cq(umem, XSK_RING_CONS__DEFAULT_NUM_DESCS);
402 #if defined(XDP_USE_NEED_WAKEUP)
403 if (xsk_ring_prod__needs_wakeup(&txq->tx))
405 while (send(xsk_socket__fd(txq->pair->xsk), NULL,
406 0, MSG_DONTWAIT) < 0) {
407 /* some thing unexpected */
408 if (errno != EBUSY && errno != EAGAIN && errno != EINTR)
411 /* pull from completion queue to leave more space */
414 XSK_RING_CONS__DEFAULT_NUM_DESCS);
418 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
420 af_xdp_tx_zc(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
422 struct pkt_tx_queue *txq = queue;
423 struct xsk_umem_info *umem = txq->umem;
424 struct rte_mbuf *mbuf;
425 unsigned long tx_bytes = 0;
429 struct xdp_desc *desc;
430 uint64_t addr, offset;
431 uint32_t free_thresh = umem->cq.size >> 1;
433 if (xsk_cons_nb_avail(&umem->cq, free_thresh) >= free_thresh)
434 pull_umem_cq(umem, XSK_RING_CONS__DEFAULT_NUM_DESCS);
436 for (i = 0; i < nb_pkts; i++) {
439 if (mbuf->pool == umem->mb_pool) {
440 if (!xsk_ring_prod__reserve(&txq->tx, 1, &idx_tx)) {
442 if (!xsk_ring_prod__reserve(&txq->tx, 1,
446 desc = xsk_ring_prod__tx_desc(&txq->tx, idx_tx);
447 desc->len = mbuf->pkt_len;
448 addr = (uint64_t)mbuf - (uint64_t)umem->buffer -
449 umem->mb_pool->header_size;
450 offset = rte_pktmbuf_mtod(mbuf, uint64_t) -
452 umem->mb_pool->header_size;
453 offset = offset << XSK_UNALIGNED_BUF_OFFSET_SHIFT;
454 desc->addr = addr | offset;
457 struct rte_mbuf *local_mbuf =
458 rte_pktmbuf_alloc(umem->mb_pool);
461 if (local_mbuf == NULL)
464 if (!xsk_ring_prod__reserve(&txq->tx, 1, &idx_tx)) {
465 rte_pktmbuf_free(local_mbuf);
470 desc = xsk_ring_prod__tx_desc(&txq->tx, idx_tx);
471 desc->len = mbuf->pkt_len;
473 addr = (uint64_t)local_mbuf - (uint64_t)umem->buffer -
474 umem->mb_pool->header_size;
475 offset = rte_pktmbuf_mtod(local_mbuf, uint64_t) -
476 (uint64_t)local_mbuf +
477 umem->mb_pool->header_size;
478 pkt = xsk_umem__get_data(umem->buffer, addr + offset);
479 offset = offset << XSK_UNALIGNED_BUF_OFFSET_SHIFT;
480 desc->addr = addr | offset;
481 rte_memcpy(pkt, rte_pktmbuf_mtod(mbuf, void *),
483 rte_pktmbuf_free(mbuf);
487 tx_bytes += mbuf->pkt_len;
493 xsk_ring_prod__submit(&txq->tx, count);
495 txq->stats.tx_pkts += count;
496 txq->stats.tx_bytes += tx_bytes;
497 txq->stats.tx_dropped += nb_pkts - count;
503 af_xdp_tx_cp(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
505 struct pkt_tx_queue *txq = queue;
506 struct xsk_umem_info *umem = txq->umem;
507 struct rte_mbuf *mbuf;
508 void *addrs[ETH_AF_XDP_TX_BATCH_SIZE];
509 unsigned long tx_bytes = 0;
513 nb_pkts = RTE_MIN(nb_pkts, ETH_AF_XDP_TX_BATCH_SIZE);
515 pull_umem_cq(umem, nb_pkts);
517 nb_pkts = rte_ring_dequeue_bulk(umem->buf_ring, addrs,
522 if (xsk_ring_prod__reserve(&txq->tx, nb_pkts, &idx_tx) != nb_pkts) {
524 rte_ring_enqueue_bulk(umem->buf_ring, addrs, nb_pkts, NULL);
528 for (i = 0; i < nb_pkts; i++) {
529 struct xdp_desc *desc;
532 desc = xsk_ring_prod__tx_desc(&txq->tx, idx_tx + i);
534 desc->len = mbuf->pkt_len;
536 desc->addr = (uint64_t)addrs[i];
537 pkt = xsk_umem__get_data(umem->mz->addr,
539 rte_memcpy(pkt, rte_pktmbuf_mtod(mbuf, void *), desc->len);
540 tx_bytes += mbuf->pkt_len;
541 rte_pktmbuf_free(mbuf);
544 xsk_ring_prod__submit(&txq->tx, nb_pkts);
548 txq->stats.tx_pkts += nb_pkts;
549 txq->stats.tx_bytes += tx_bytes;
556 eth_af_xdp_tx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
558 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
559 return af_xdp_tx_zc(queue, bufs, nb_pkts);
561 return af_xdp_tx_cp(queue, bufs, nb_pkts);
566 eth_dev_start(struct rte_eth_dev *dev)
568 dev->data->dev_link.link_status = ETH_LINK_UP;
573 /* This function gets called when the current port gets stopped. */
575 eth_dev_stop(struct rte_eth_dev *dev)
577 dev->data->dev_link.link_status = ETH_LINK_DOWN;
581 eth_dev_configure(struct rte_eth_dev *dev)
583 /* rx/tx must be paired */
584 if (dev->data->nb_rx_queues != dev->data->nb_tx_queues)
591 eth_dev_info(struct rte_eth_dev *dev, struct rte_eth_dev_info *dev_info)
593 struct pmd_internals *internals = dev->data->dev_private;
595 dev_info->if_index = internals->if_index;
596 dev_info->max_mac_addrs = 1;
597 dev_info->max_rx_pktlen = ETH_FRAME_LEN;
598 dev_info->max_rx_queues = internals->queue_cnt;
599 dev_info->max_tx_queues = internals->queue_cnt;
601 dev_info->min_mtu = RTE_ETHER_MIN_MTU;
602 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
603 dev_info->max_mtu = getpagesize() -
604 sizeof(struct rte_mempool_objhdr) -
605 sizeof(struct rte_mbuf) -
606 RTE_PKTMBUF_HEADROOM - XDP_PACKET_HEADROOM;
608 dev_info->max_mtu = ETH_AF_XDP_FRAME_SIZE - XDP_PACKET_HEADROOM;
611 dev_info->default_rxportconf.nb_queues = 1;
612 dev_info->default_txportconf.nb_queues = 1;
613 dev_info->default_rxportconf.ring_size = ETH_AF_XDP_DFLT_NUM_DESCS;
614 dev_info->default_txportconf.ring_size = ETH_AF_XDP_DFLT_NUM_DESCS;
620 eth_stats_get(struct rte_eth_dev *dev, struct rte_eth_stats *stats)
622 struct pmd_internals *internals = dev->data->dev_private;
623 struct xdp_statistics xdp_stats;
624 struct pkt_rx_queue *rxq;
625 struct pkt_tx_queue *txq;
629 for (i = 0; i < dev->data->nb_rx_queues; i++) {
630 optlen = sizeof(struct xdp_statistics);
631 rxq = &internals->rx_queues[i];
633 stats->q_ipackets[i] = rxq->stats.rx_pkts;
634 stats->q_ibytes[i] = rxq->stats.rx_bytes;
636 stats->q_opackets[i] = txq->stats.tx_pkts;
637 stats->q_obytes[i] = txq->stats.tx_bytes;
639 stats->ipackets += stats->q_ipackets[i];
640 stats->ibytes += stats->q_ibytes[i];
641 stats->imissed += rxq->stats.rx_dropped;
642 stats->oerrors += txq->stats.tx_dropped;
643 ret = getsockopt(xsk_socket__fd(rxq->xsk), SOL_XDP,
644 XDP_STATISTICS, &xdp_stats, &optlen);
646 AF_XDP_LOG(ERR, "getsockopt() failed for XDP_STATISTICS.\n");
649 stats->imissed += xdp_stats.rx_dropped;
651 stats->opackets += stats->q_opackets[i];
652 stats->obytes += stats->q_obytes[i];
659 eth_stats_reset(struct rte_eth_dev *dev)
661 struct pmd_internals *internals = dev->data->dev_private;
664 for (i = 0; i < internals->queue_cnt; i++) {
665 memset(&internals->rx_queues[i].stats, 0,
666 sizeof(struct rx_stats));
667 memset(&internals->tx_queues[i].stats, 0,
668 sizeof(struct tx_stats));
675 remove_xdp_program(struct pmd_internals *internals)
677 uint32_t curr_prog_id = 0;
679 if (bpf_get_link_xdp_id(internals->if_index, &curr_prog_id,
680 XDP_FLAGS_UPDATE_IF_NOEXIST)) {
681 AF_XDP_LOG(ERR, "bpf_get_link_xdp_id failed\n");
684 bpf_set_link_xdp_fd(internals->if_index, -1,
685 XDP_FLAGS_UPDATE_IF_NOEXIST);
689 xdp_umem_destroy(struct xsk_umem_info *umem)
691 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
692 umem->mb_pool = NULL;
694 rte_memzone_free(umem->mz);
697 rte_ring_free(umem->buf_ring);
698 umem->buf_ring = NULL;
706 eth_dev_close(struct rte_eth_dev *dev)
708 struct pmd_internals *internals = dev->data->dev_private;
709 struct pkt_rx_queue *rxq;
712 AF_XDP_LOG(INFO, "Closing AF_XDP ethdev on numa socket %u\n",
715 for (i = 0; i < internals->queue_cnt; i++) {
716 rxq = &internals->rx_queues[i];
717 if (rxq->umem == NULL)
719 xsk_socket__delete(rxq->xsk);
720 (void)xsk_umem__delete(rxq->umem->umem);
721 xdp_umem_destroy(rxq->umem);
723 /* free pkt_tx_queue */
729 * MAC is not allocated dynamically, setting it to NULL would prevent
730 * from releasing it in rte_eth_dev_release_port.
732 dev->data->mac_addrs = NULL;
734 remove_xdp_program(internals);
738 eth_queue_release(void *q __rte_unused)
743 eth_link_update(struct rte_eth_dev *dev __rte_unused,
744 int wait_to_complete __rte_unused)
749 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
750 static inline uint64_t get_base_addr(struct rte_mempool *mp, uint64_t *align)
752 struct rte_mempool_memhdr *memhdr;
753 uint64_t memhdr_addr, aligned_addr;
755 memhdr = STAILQ_FIRST(&mp->mem_list);
756 memhdr_addr = (uint64_t)memhdr->addr;
757 aligned_addr = memhdr_addr & ~(getpagesize() - 1);
758 *align = memhdr_addr - aligned_addr;
764 xsk_umem_info *xdp_umem_configure(struct pmd_internals *internals __rte_unused,
765 struct pkt_rx_queue *rxq)
767 struct xsk_umem_info *umem;
769 struct xsk_umem_config usr_config = {
770 .fill_size = ETH_AF_XDP_DFLT_NUM_DESCS * 2,
771 .comp_size = ETH_AF_XDP_DFLT_NUM_DESCS,
772 .flags = XDP_UMEM_UNALIGNED_CHUNK_FLAG};
773 void *base_addr = NULL;
774 struct rte_mempool *mb_pool = rxq->mb_pool;
775 uint64_t umem_size, align = 0;
777 usr_config.frame_size = rte_mempool_calc_obj_size(mb_pool->elt_size,
780 usr_config.frame_headroom = mb_pool->header_size +
781 sizeof(struct rte_mbuf) +
782 rte_pktmbuf_priv_size(mb_pool) +
783 RTE_PKTMBUF_HEADROOM;
785 umem = rte_zmalloc_socket("umem", sizeof(*umem), 0, rte_socket_id());
787 AF_XDP_LOG(ERR, "Failed to allocate umem info");
791 umem->mb_pool = mb_pool;
792 base_addr = (void *)get_base_addr(mb_pool, &align);
793 umem_size = mb_pool->populated_size * usr_config.frame_size + align;
795 ret = xsk_umem__create(&umem->umem, base_addr, umem_size,
796 &umem->fq, &umem->cq, &usr_config);
799 AF_XDP_LOG(ERR, "Failed to create umem");
802 umem->buffer = base_addr;
806 xsk_umem_info *xdp_umem_configure(struct pmd_internals *internals,
807 struct pkt_rx_queue *rxq)
809 struct xsk_umem_info *umem;
810 const struct rte_memzone *mz;
811 struct xsk_umem_config usr_config = {
812 .fill_size = ETH_AF_XDP_DFLT_NUM_DESCS,
813 .comp_size = ETH_AF_XDP_DFLT_NUM_DESCS,
814 .frame_size = ETH_AF_XDP_FRAME_SIZE,
815 .frame_headroom = 0 };
816 char ring_name[RTE_RING_NAMESIZE];
817 char mz_name[RTE_MEMZONE_NAMESIZE];
821 umem = rte_zmalloc_socket("umem", sizeof(*umem), 0, rte_socket_id());
823 AF_XDP_LOG(ERR, "Failed to allocate umem info");
827 snprintf(ring_name, sizeof(ring_name), "af_xdp_ring_%s_%u",
828 internals->if_name, rxq->xsk_queue_idx);
829 umem->buf_ring = rte_ring_create(ring_name,
830 ETH_AF_XDP_NUM_BUFFERS,
833 if (umem->buf_ring == NULL) {
834 AF_XDP_LOG(ERR, "Failed to create rte_ring\n");
838 for (i = 0; i < ETH_AF_XDP_NUM_BUFFERS; i++)
839 rte_ring_enqueue(umem->buf_ring,
840 (void *)(i * ETH_AF_XDP_FRAME_SIZE));
842 snprintf(mz_name, sizeof(mz_name), "af_xdp_umem_%s_%u",
843 internals->if_name, rxq->xsk_queue_idx);
844 mz = rte_memzone_reserve_aligned(mz_name,
845 ETH_AF_XDP_NUM_BUFFERS * ETH_AF_XDP_FRAME_SIZE,
846 rte_socket_id(), RTE_MEMZONE_IOVA_CONTIG,
849 AF_XDP_LOG(ERR, "Failed to reserve memzone for af_xdp umem.\n");
853 ret = xsk_umem__create(&umem->umem, mz->addr,
854 ETH_AF_XDP_NUM_BUFFERS * ETH_AF_XDP_FRAME_SIZE,
855 &umem->fq, &umem->cq,
859 AF_XDP_LOG(ERR, "Failed to create umem");
868 xdp_umem_destroy(umem);
873 xsk_configure(struct pmd_internals *internals, struct pkt_rx_queue *rxq,
876 struct xsk_socket_config cfg;
877 struct pkt_tx_queue *txq = rxq->pair;
879 int reserve_size = ETH_AF_XDP_DFLT_NUM_DESCS;
880 struct rte_mbuf *fq_bufs[reserve_size];
882 rxq->umem = xdp_umem_configure(internals, rxq);
883 if (rxq->umem == NULL)
885 txq->umem = rxq->umem;
887 cfg.rx_size = ring_size;
888 cfg.tx_size = ring_size;
889 cfg.libbpf_flags = 0;
890 cfg.xdp_flags = XDP_FLAGS_UPDATE_IF_NOEXIST;
893 #if defined(XDP_USE_NEED_WAKEUP)
894 cfg.bind_flags |= XDP_USE_NEED_WAKEUP;
897 ret = xsk_socket__create(&rxq->xsk, internals->if_name,
898 rxq->xsk_queue_idx, rxq->umem->umem, &rxq->rx,
901 AF_XDP_LOG(ERR, "Failed to create xsk socket.\n");
905 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
906 if (rte_pktmbuf_alloc_bulk(rxq->umem->mb_pool, fq_bufs, reserve_size)) {
907 AF_XDP_LOG(DEBUG, "Failed to get enough buffers for fq.\n");
911 ret = reserve_fill_queue(rxq->umem, reserve_size, fq_bufs);
913 xsk_socket__delete(rxq->xsk);
914 AF_XDP_LOG(ERR, "Failed to reserve fill queue.\n");
921 xdp_umem_destroy(rxq->umem);
927 eth_rx_queue_setup(struct rte_eth_dev *dev,
928 uint16_t rx_queue_id,
930 unsigned int socket_id __rte_unused,
931 const struct rte_eth_rxconf *rx_conf __rte_unused,
932 struct rte_mempool *mb_pool)
934 struct pmd_internals *internals = dev->data->dev_private;
935 struct pkt_rx_queue *rxq;
938 rxq = &internals->rx_queues[rx_queue_id];
940 AF_XDP_LOG(INFO, "Set up rx queue, rx queue id: %d, xsk queue id: %d\n",
941 rx_queue_id, rxq->xsk_queue_idx);
943 #ifndef XDP_UMEM_UNALIGNED_CHUNK_FLAG
944 uint32_t buf_size, data_size;
946 /* Now get the space available for data in the mbuf */
947 buf_size = rte_pktmbuf_data_room_size(mb_pool) -
948 RTE_PKTMBUF_HEADROOM;
949 data_size = ETH_AF_XDP_FRAME_SIZE;
951 if (data_size > buf_size) {
952 AF_XDP_LOG(ERR, "%s: %d bytes will not fit in mbuf (%d bytes)\n",
953 dev->device->name, data_size, buf_size);
959 rxq->mb_pool = mb_pool;
961 if (xsk_configure(internals, rxq, nb_rx_desc)) {
962 AF_XDP_LOG(ERR, "Failed to configure xdp socket\n");
967 rxq->fds[0].fd = xsk_socket__fd(rxq->xsk);
968 rxq->fds[0].events = POLLIN;
970 dev->data->rx_queues[rx_queue_id] = rxq;
978 eth_tx_queue_setup(struct rte_eth_dev *dev,
979 uint16_t tx_queue_id,
980 uint16_t nb_tx_desc __rte_unused,
981 unsigned int socket_id __rte_unused,
982 const struct rte_eth_txconf *tx_conf __rte_unused)
984 struct pmd_internals *internals = dev->data->dev_private;
985 struct pkt_tx_queue *txq;
987 txq = &internals->tx_queues[tx_queue_id];
989 dev->data->tx_queues[tx_queue_id] = txq;
994 eth_dev_mtu_set(struct rte_eth_dev *dev, uint16_t mtu)
996 struct pmd_internals *internals = dev->data->dev_private;
997 struct ifreq ifr = { .ifr_mtu = mtu };
1001 s = socket(PF_INET, SOCK_DGRAM, 0);
1005 strlcpy(ifr.ifr_name, internals->if_name, IFNAMSIZ);
1006 ret = ioctl(s, SIOCSIFMTU, &ifr);
1009 return (ret < 0) ? -errno : 0;
1013 eth_dev_change_flags(char *if_name, uint32_t flags, uint32_t mask)
1019 s = socket(PF_INET, SOCK_DGRAM, 0);
1023 strlcpy(ifr.ifr_name, if_name, IFNAMSIZ);
1024 if (ioctl(s, SIOCGIFFLAGS, &ifr) < 0) {
1028 ifr.ifr_flags &= mask;
1029 ifr.ifr_flags |= flags;
1030 if (ioctl(s, SIOCSIFFLAGS, &ifr) < 0) {
1040 eth_dev_promiscuous_enable(struct rte_eth_dev *dev)
1042 struct pmd_internals *internals = dev->data->dev_private;
1044 return eth_dev_change_flags(internals->if_name, IFF_PROMISC, ~0);
1048 eth_dev_promiscuous_disable(struct rte_eth_dev *dev)
1050 struct pmd_internals *internals = dev->data->dev_private;
1052 return eth_dev_change_flags(internals->if_name, 0, ~IFF_PROMISC);
1055 static const struct eth_dev_ops ops = {
1056 .dev_start = eth_dev_start,
1057 .dev_stop = eth_dev_stop,
1058 .dev_close = eth_dev_close,
1059 .dev_configure = eth_dev_configure,
1060 .dev_infos_get = eth_dev_info,
1061 .mtu_set = eth_dev_mtu_set,
1062 .promiscuous_enable = eth_dev_promiscuous_enable,
1063 .promiscuous_disable = eth_dev_promiscuous_disable,
1064 .rx_queue_setup = eth_rx_queue_setup,
1065 .tx_queue_setup = eth_tx_queue_setup,
1066 .rx_queue_release = eth_queue_release,
1067 .tx_queue_release = eth_queue_release,
1068 .link_update = eth_link_update,
1069 .stats_get = eth_stats_get,
1070 .stats_reset = eth_stats_reset,
1073 /** parse integer from integer argument */
1075 parse_integer_arg(const char *key __rte_unused,
1076 const char *value, void *extra_args)
1078 int *i = (int *)extra_args;
1081 *i = strtol(value, &end, 10);
1083 AF_XDP_LOG(ERR, "Argument has to be positive.\n");
1090 /** parse name argument */
1092 parse_name_arg(const char *key __rte_unused,
1093 const char *value, void *extra_args)
1095 char *name = extra_args;
1097 if (strnlen(value, IFNAMSIZ) > IFNAMSIZ - 1) {
1098 AF_XDP_LOG(ERR, "Invalid name %s, should be less than %u bytes.\n",
1103 strlcpy(name, value, IFNAMSIZ);
1109 xdp_get_channels_info(const char *if_name, int *max_queues,
1110 int *combined_queues)
1112 struct ethtool_channels channels;
1116 fd = socket(AF_INET, SOCK_DGRAM, 0);
1120 channels.cmd = ETHTOOL_GCHANNELS;
1121 ifr.ifr_data = (void *)&channels;
1122 strncpy(ifr.ifr_name, if_name, IFNAMSIZ);
1123 ret = ioctl(fd, SIOCETHTOOL, &ifr);
1125 if (errno == EOPNOTSUPP) {
1133 if (channels.max_combined == 0 || errno == EOPNOTSUPP) {
1134 /* If the device says it has no channels, then all traffic
1135 * is sent to a single stream, so max queues = 1.
1138 *combined_queues = 1;
1140 *max_queues = channels.max_combined;
1141 *combined_queues = channels.combined_count;
1150 parse_parameters(struct rte_kvargs *kvlist, char *if_name, int *start_queue,
1155 ret = rte_kvargs_process(kvlist, ETH_AF_XDP_IFACE_ARG,
1156 &parse_name_arg, if_name);
1160 ret = rte_kvargs_process(kvlist, ETH_AF_XDP_START_QUEUE_ARG,
1161 &parse_integer_arg, start_queue);
1165 ret = rte_kvargs_process(kvlist, ETH_AF_XDP_QUEUE_COUNT_ARG,
1166 &parse_integer_arg, queue_cnt);
1167 if (ret < 0 || *queue_cnt <= 0) {
1173 rte_kvargs_free(kvlist);
1178 get_iface_info(const char *if_name,
1179 struct rte_ether_addr *eth_addr,
1183 int sock = socket(AF_INET, SOCK_DGRAM, IPPROTO_IP);
1188 strlcpy(ifr.ifr_name, if_name, IFNAMSIZ);
1189 if (ioctl(sock, SIOCGIFINDEX, &ifr))
1192 *if_index = ifr.ifr_ifindex;
1194 if (ioctl(sock, SIOCGIFHWADDR, &ifr))
1197 rte_memcpy(eth_addr, ifr.ifr_hwaddr.sa_data, RTE_ETHER_ADDR_LEN);
1207 static struct rte_eth_dev *
1208 init_internals(struct rte_vdev_device *dev, const char *if_name,
1209 int start_queue_idx, int queue_cnt)
1211 const char *name = rte_vdev_device_name(dev);
1212 const unsigned int numa_node = dev->device.numa_node;
1213 struct pmd_internals *internals;
1214 struct rte_eth_dev *eth_dev;
1218 internals = rte_zmalloc_socket(name, sizeof(*internals), 0, numa_node);
1219 if (internals == NULL)
1222 internals->start_queue_idx = start_queue_idx;
1223 internals->queue_cnt = queue_cnt;
1224 strlcpy(internals->if_name, if_name, IFNAMSIZ);
1226 if (xdp_get_channels_info(if_name, &internals->max_queue_cnt,
1227 &internals->combined_queue_cnt)) {
1228 AF_XDP_LOG(ERR, "Failed to get channel info of interface: %s\n",
1230 goto err_free_internals;
1233 if (queue_cnt > internals->combined_queue_cnt) {
1234 AF_XDP_LOG(ERR, "Specified queue count %d is larger than combined queue count %d.\n",
1235 queue_cnt, internals->combined_queue_cnt);
1236 goto err_free_internals;
1239 internals->rx_queues = rte_zmalloc_socket(NULL,
1240 sizeof(struct pkt_rx_queue) * queue_cnt,
1242 if (internals->rx_queues == NULL) {
1243 AF_XDP_LOG(ERR, "Failed to allocate memory for rx queues.\n");
1244 goto err_free_internals;
1247 internals->tx_queues = rte_zmalloc_socket(NULL,
1248 sizeof(struct pkt_tx_queue) * queue_cnt,
1250 if (internals->tx_queues == NULL) {
1251 AF_XDP_LOG(ERR, "Failed to allocate memory for tx queues.\n");
1254 for (i = 0; i < queue_cnt; i++) {
1255 internals->tx_queues[i].pair = &internals->rx_queues[i];
1256 internals->rx_queues[i].pair = &internals->tx_queues[i];
1257 internals->rx_queues[i].xsk_queue_idx = start_queue_idx + i;
1258 internals->tx_queues[i].xsk_queue_idx = start_queue_idx + i;
1261 ret = get_iface_info(if_name, &internals->eth_addr,
1262 &internals->if_index);
1266 eth_dev = rte_eth_vdev_allocate(dev, 0);
1267 if (eth_dev == NULL)
1270 eth_dev->data->dev_private = internals;
1271 eth_dev->data->dev_link = pmd_link;
1272 eth_dev->data->mac_addrs = &internals->eth_addr;
1273 eth_dev->dev_ops = &ops;
1274 eth_dev->rx_pkt_burst = eth_af_xdp_rx;
1275 eth_dev->tx_pkt_burst = eth_af_xdp_tx;
1276 /* Let rte_eth_dev_close() release the port resources. */
1277 eth_dev->data->dev_flags |= RTE_ETH_DEV_CLOSE_REMOVE;
1279 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
1280 AF_XDP_LOG(INFO, "Zero copy between umem and mbuf enabled.\n");
1286 rte_free(internals->tx_queues);
1288 rte_free(internals->rx_queues);
1290 rte_free(internals);
1295 rte_pmd_af_xdp_probe(struct rte_vdev_device *dev)
1297 struct rte_kvargs *kvlist;
1298 char if_name[IFNAMSIZ] = {'\0'};
1299 int xsk_start_queue_idx = ETH_AF_XDP_DFLT_START_QUEUE_IDX;
1300 int xsk_queue_cnt = ETH_AF_XDP_DFLT_QUEUE_COUNT;
1301 struct rte_eth_dev *eth_dev = NULL;
1304 AF_XDP_LOG(INFO, "Initializing pmd_af_xdp for %s\n",
1305 rte_vdev_device_name(dev));
1307 name = rte_vdev_device_name(dev);
1308 if (rte_eal_process_type() == RTE_PROC_SECONDARY &&
1309 strlen(rte_vdev_device_args(dev)) == 0) {
1310 eth_dev = rte_eth_dev_attach_secondary(name);
1311 if (eth_dev == NULL) {
1312 AF_XDP_LOG(ERR, "Failed to probe %s\n", name);
1315 eth_dev->dev_ops = &ops;
1316 rte_eth_dev_probing_finish(eth_dev);
1320 kvlist = rte_kvargs_parse(rte_vdev_device_args(dev), valid_arguments);
1321 if (kvlist == NULL) {
1322 AF_XDP_LOG(ERR, "Invalid kvargs key\n");
1326 if (dev->device.numa_node == SOCKET_ID_ANY)
1327 dev->device.numa_node = rte_socket_id();
1329 if (parse_parameters(kvlist, if_name, &xsk_start_queue_idx,
1330 &xsk_queue_cnt) < 0) {
1331 AF_XDP_LOG(ERR, "Invalid kvargs value\n");
1335 if (strlen(if_name) == 0) {
1336 AF_XDP_LOG(ERR, "Network interface must be specified\n");
1340 eth_dev = init_internals(dev, if_name, xsk_start_queue_idx,
1342 if (eth_dev == NULL) {
1343 AF_XDP_LOG(ERR, "Failed to init internals\n");
1347 rte_eth_dev_probing_finish(eth_dev);
1353 rte_pmd_af_xdp_remove(struct rte_vdev_device *dev)
1355 struct rte_eth_dev *eth_dev = NULL;
1357 AF_XDP_LOG(INFO, "Removing AF_XDP ethdev on numa socket %u\n",
1363 /* find the ethdev entry */
1364 eth_dev = rte_eth_dev_allocated(rte_vdev_device_name(dev));
1365 if (eth_dev == NULL)
1368 eth_dev_close(eth_dev);
1369 rte_eth_dev_release_port(eth_dev);
1375 static struct rte_vdev_driver pmd_af_xdp_drv = {
1376 .probe = rte_pmd_af_xdp_probe,
1377 .remove = rte_pmd_af_xdp_remove,
1380 RTE_PMD_REGISTER_VDEV(net_af_xdp, pmd_af_xdp_drv);
1381 RTE_PMD_REGISTER_PARAM_STRING(net_af_xdp,
1383 "start_queue=<int> "
1384 "queue_count=<int> ");