1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2019-2020 Intel Corporation.
8 #include <netinet/in.h>
10 #include <sys/socket.h>
11 #include <sys/ioctl.h>
12 #include <linux/if_ether.h>
13 #include <linux/if_xdp.h>
14 #include <linux/if_link.h>
15 #include <linux/ethtool.h>
16 #include <linux/sockios.h>
17 #include "af_xdp_deps.h"
20 #include <rte_ethdev.h>
21 #include <ethdev_driver.h>
22 #include <ethdev_vdev.h>
23 #include <rte_kvargs.h>
24 #include <rte_bus_vdev.h>
25 #include <rte_string_fns.h>
26 #include <rte_branch_prediction.h>
27 #include <rte_common.h>
30 #include <rte_ether.h>
31 #include <rte_lcore.h>
33 #include <rte_memory.h>
34 #include <rte_memzone.h>
35 #include <rte_mempool.h>
37 #include <rte_malloc.h>
39 #include <rte_spinlock.h>
43 #ifndef SO_PREFER_BUSY_POLL
44 #define SO_PREFER_BUSY_POLL 69
46 #ifndef SO_BUSY_POLL_BUDGET
47 #define SO_BUSY_POLL_BUDGET 70
63 RTE_LOG_REGISTER_DEFAULT(af_xdp_logtype, NOTICE);
65 #define AF_XDP_LOG(level, fmt, args...) \
66 rte_log(RTE_LOG_ ## level, af_xdp_logtype, \
67 "%s(): " fmt, __func__, ##args)
69 #define ETH_AF_XDP_FRAME_SIZE 2048
70 #define ETH_AF_XDP_NUM_BUFFERS 4096
71 #define ETH_AF_XDP_DFLT_NUM_DESCS XSK_RING_CONS__DEFAULT_NUM_DESCS
72 #define ETH_AF_XDP_DFLT_START_QUEUE_IDX 0
73 #define ETH_AF_XDP_DFLT_QUEUE_COUNT 1
74 #define ETH_AF_XDP_DFLT_BUSY_BUDGET 64
75 #define ETH_AF_XDP_DFLT_BUSY_TIMEOUT 20
77 #define ETH_AF_XDP_RX_BATCH_SIZE XSK_RING_CONS__DEFAULT_NUM_DESCS
78 #define ETH_AF_XDP_TX_BATCH_SIZE XSK_RING_CONS__DEFAULT_NUM_DESCS
81 struct xsk_umem_info {
82 struct xsk_umem *umem;
83 struct rte_ring *buf_ring;
84 const struct rte_memzone *mz;
85 struct rte_mempool *mb_pool;
98 struct xsk_ring_cons rx;
99 struct xsk_umem_info *umem;
100 struct xsk_socket *xsk;
101 struct rte_mempool *mb_pool;
103 struct rx_stats stats;
105 struct xsk_ring_prod fq;
106 struct xsk_ring_cons cq;
108 struct pkt_tx_queue *pair;
109 struct pollfd fds[1];
120 struct pkt_tx_queue {
121 struct xsk_ring_prod tx;
122 struct xsk_umem_info *umem;
124 struct tx_stats stats;
126 struct pkt_rx_queue *pair;
130 struct pmd_internals {
132 char if_name[IFNAMSIZ];
136 int combined_queue_cnt;
138 char prog_path[PATH_MAX];
139 bool custom_prog_configured;
141 struct rte_ether_addr eth_addr;
143 struct pkt_rx_queue *rx_queues;
144 struct pkt_tx_queue *tx_queues;
147 #define ETH_AF_XDP_IFACE_ARG "iface"
148 #define ETH_AF_XDP_START_QUEUE_ARG "start_queue"
149 #define ETH_AF_XDP_QUEUE_COUNT_ARG "queue_count"
150 #define ETH_AF_XDP_SHARED_UMEM_ARG "shared_umem"
151 #define ETH_AF_XDP_PROG_ARG "xdp_prog"
152 #define ETH_AF_XDP_BUDGET_ARG "busy_budget"
154 static const char * const valid_arguments[] = {
155 ETH_AF_XDP_IFACE_ARG,
156 ETH_AF_XDP_START_QUEUE_ARG,
157 ETH_AF_XDP_QUEUE_COUNT_ARG,
158 ETH_AF_XDP_SHARED_UMEM_ARG,
160 ETH_AF_XDP_BUDGET_ARG,
164 static const struct rte_eth_link pmd_link = {
165 .link_speed = ETH_SPEED_NUM_10G,
166 .link_duplex = ETH_LINK_FULL_DUPLEX,
167 .link_status = ETH_LINK_DOWN,
168 .link_autoneg = ETH_LINK_AUTONEG
171 /* List which tracks PMDs to facilitate sharing UMEMs across them. */
172 struct internal_list {
173 TAILQ_ENTRY(internal_list) next;
174 struct rte_eth_dev *eth_dev;
177 TAILQ_HEAD(internal_list_head, internal_list);
178 static struct internal_list_head internal_list =
179 TAILQ_HEAD_INITIALIZER(internal_list);
181 static pthread_mutex_t internal_list_lock = PTHREAD_MUTEX_INITIALIZER;
183 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
185 reserve_fill_queue_zc(struct xsk_umem_info *umem, uint16_t reserve_size,
186 struct rte_mbuf **bufs, struct xsk_ring_prod *fq)
191 if (unlikely(!xsk_ring_prod__reserve(fq, reserve_size, &idx))) {
192 for (i = 0; i < reserve_size; i++)
193 rte_pktmbuf_free(bufs[i]);
194 AF_XDP_LOG(DEBUG, "Failed to reserve enough fq descs.\n");
198 for (i = 0; i < reserve_size; i++) {
202 fq_addr = xsk_ring_prod__fill_addr(fq, idx++);
203 addr = (uint64_t)bufs[i] - (uint64_t)umem->buffer -
204 umem->mb_pool->header_size;
208 xsk_ring_prod__submit(fq, reserve_size);
214 reserve_fill_queue_cp(struct xsk_umem_info *umem, uint16_t reserve_size,
215 struct rte_mbuf **bufs __rte_unused,
216 struct xsk_ring_prod *fq)
218 void *addrs[reserve_size];
222 if (rte_ring_dequeue_bulk(umem->buf_ring, addrs, reserve_size, NULL)
224 AF_XDP_LOG(DEBUG, "Failed to get enough buffers for fq.\n");
228 if (unlikely(!xsk_ring_prod__reserve(fq, reserve_size, &idx))) {
229 AF_XDP_LOG(DEBUG, "Failed to reserve enough fq descs.\n");
230 rte_ring_enqueue_bulk(umem->buf_ring, addrs,
235 for (i = 0; i < reserve_size; i++) {
238 fq_addr = xsk_ring_prod__fill_addr(fq, idx++);
239 *fq_addr = (uint64_t)addrs[i];
242 xsk_ring_prod__submit(fq, reserve_size);
249 reserve_fill_queue(struct xsk_umem_info *umem, uint16_t reserve_size,
250 struct rte_mbuf **bufs, struct xsk_ring_prod *fq)
252 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
253 return reserve_fill_queue_zc(umem, reserve_size, bufs, fq);
255 return reserve_fill_queue_cp(umem, reserve_size, bufs, fq);
259 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
261 af_xdp_rx_zc(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
263 struct pkt_rx_queue *rxq = queue;
264 struct xsk_ring_cons *rx = &rxq->rx;
265 struct xsk_ring_prod *fq = &rxq->fq;
266 struct xsk_umem_info *umem = rxq->umem;
268 unsigned long rx_bytes = 0;
270 struct rte_mbuf *fq_bufs[ETH_AF_XDP_RX_BATCH_SIZE];
272 nb_pkts = xsk_ring_cons__peek(rx, nb_pkts, &idx_rx);
275 /* we can assume a kernel >= 5.11 is in use if busy polling is
276 * enabled and thus we can safely use the recvfrom() syscall
277 * which is only supported for AF_XDP sockets in kernels >=
280 if (rxq->busy_budget) {
281 (void)recvfrom(xsk_socket__fd(rxq->xsk), NULL, 0,
282 MSG_DONTWAIT, NULL, NULL);
283 } else if (xsk_ring_prod__needs_wakeup(fq)) {
284 (void)poll(&rxq->fds[0], 1, 1000);
290 /* allocate bufs for fill queue replenishment after rx */
291 if (rte_pktmbuf_alloc_bulk(umem->mb_pool, fq_bufs, nb_pkts)) {
293 "Failed to get enough buffers for fq.\n");
294 /* rollback cached_cons which is added by
295 * xsk_ring_cons__peek
297 rx->cached_cons -= nb_pkts;
301 for (i = 0; i < nb_pkts; i++) {
302 const struct xdp_desc *desc;
307 desc = xsk_ring_cons__rx_desc(rx, idx_rx++);
311 offset = xsk_umem__extract_offset(addr);
312 addr = xsk_umem__extract_addr(addr);
314 bufs[i] = (struct rte_mbuf *)
315 xsk_umem__get_data(umem->buffer, addr +
316 umem->mb_pool->header_size);
317 bufs[i]->data_off = offset - sizeof(struct rte_mbuf) -
318 rte_pktmbuf_priv_size(umem->mb_pool) -
319 umem->mb_pool->header_size;
321 rte_pktmbuf_pkt_len(bufs[i]) = len;
322 rte_pktmbuf_data_len(bufs[i]) = len;
326 xsk_ring_cons__release(rx, nb_pkts);
327 (void)reserve_fill_queue(umem, nb_pkts, fq_bufs, fq);
330 rxq->stats.rx_pkts += nb_pkts;
331 rxq->stats.rx_bytes += rx_bytes;
337 af_xdp_rx_cp(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
339 struct pkt_rx_queue *rxq = queue;
340 struct xsk_ring_cons *rx = &rxq->rx;
341 struct xsk_umem_info *umem = rxq->umem;
342 struct xsk_ring_prod *fq = &rxq->fq;
344 unsigned long rx_bytes = 0;
346 uint32_t free_thresh = fq->size >> 1;
347 struct rte_mbuf *mbufs[ETH_AF_XDP_RX_BATCH_SIZE];
349 if (xsk_prod_nb_free(fq, free_thresh) >= free_thresh)
350 (void)reserve_fill_queue(umem, nb_pkts, NULL, fq);
352 nb_pkts = xsk_ring_cons__peek(rx, nb_pkts, &idx_rx);
354 #if defined(XDP_USE_NEED_WAKEUP)
355 if (xsk_ring_prod__needs_wakeup(fq))
356 (void)poll(rxq->fds, 1, 1000);
361 if (unlikely(rte_pktmbuf_alloc_bulk(rxq->mb_pool, mbufs, nb_pkts))) {
362 /* rollback cached_cons which is added by
363 * xsk_ring_cons__peek
365 rx->cached_cons -= nb_pkts;
369 for (i = 0; i < nb_pkts; i++) {
370 const struct xdp_desc *desc;
375 desc = xsk_ring_cons__rx_desc(rx, idx_rx++);
378 pkt = xsk_umem__get_data(rxq->umem->mz->addr, addr);
380 rte_memcpy(rte_pktmbuf_mtod(mbufs[i], void *), pkt, len);
381 rte_ring_enqueue(umem->buf_ring, (void *)addr);
382 rte_pktmbuf_pkt_len(mbufs[i]) = len;
383 rte_pktmbuf_data_len(mbufs[i]) = len;
388 xsk_ring_cons__release(rx, nb_pkts);
391 rxq->stats.rx_pkts += nb_pkts;
392 rxq->stats.rx_bytes += rx_bytes;
399 af_xdp_rx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
401 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
402 return af_xdp_rx_zc(queue, bufs, nb_pkts);
404 return af_xdp_rx_cp(queue, bufs, nb_pkts);
409 eth_af_xdp_rx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
413 if (likely(nb_pkts <= ETH_AF_XDP_RX_BATCH_SIZE))
414 return af_xdp_rx(queue, bufs, nb_pkts);
416 /* Split larger batch into smaller batches of size
417 * ETH_AF_XDP_RX_BATCH_SIZE or less.
423 n = (uint16_t)RTE_MIN(nb_pkts, ETH_AF_XDP_RX_BATCH_SIZE);
424 ret = af_xdp_rx(queue, &bufs[nb_rx], n);
425 nb_rx = (uint16_t)(nb_rx + ret);
426 nb_pkts = (uint16_t)(nb_pkts - ret);
435 pull_umem_cq(struct xsk_umem_info *umem, int size, struct xsk_ring_cons *cq)
440 n = xsk_ring_cons__peek(cq, size, &idx_cq);
442 for (i = 0; i < n; i++) {
444 addr = *xsk_ring_cons__comp_addr(cq, idx_cq++);
445 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
446 addr = xsk_umem__extract_addr(addr);
447 rte_pktmbuf_free((struct rte_mbuf *)
448 xsk_umem__get_data(umem->buffer,
449 addr + umem->mb_pool->header_size));
451 rte_ring_enqueue(umem->buf_ring, (void *)addr);
455 xsk_ring_cons__release(cq, n);
459 kick_tx(struct pkt_tx_queue *txq, struct xsk_ring_cons *cq)
461 struct xsk_umem_info *umem = txq->umem;
463 pull_umem_cq(umem, XSK_RING_CONS__DEFAULT_NUM_DESCS, cq);
465 if (tx_syscall_needed(&txq->tx))
466 while (send(xsk_socket__fd(txq->pair->xsk), NULL,
467 0, MSG_DONTWAIT) < 0) {
468 /* some thing unexpected */
469 if (errno != EBUSY && errno != EAGAIN && errno != EINTR)
472 /* pull from completion queue to leave more space */
475 XSK_RING_CONS__DEFAULT_NUM_DESCS,
480 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
482 af_xdp_tx_zc(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
484 struct pkt_tx_queue *txq = queue;
485 struct xsk_umem_info *umem = txq->umem;
486 struct rte_mbuf *mbuf;
487 unsigned long tx_bytes = 0;
491 struct xdp_desc *desc;
492 uint64_t addr, offset;
493 struct xsk_ring_cons *cq = &txq->pair->cq;
494 uint32_t free_thresh = cq->size >> 1;
496 if (xsk_cons_nb_avail(cq, free_thresh) >= free_thresh)
497 pull_umem_cq(umem, XSK_RING_CONS__DEFAULT_NUM_DESCS, cq);
499 for (i = 0; i < nb_pkts; i++) {
502 if (mbuf->pool == umem->mb_pool) {
503 if (!xsk_ring_prod__reserve(&txq->tx, 1, &idx_tx)) {
505 if (!xsk_ring_prod__reserve(&txq->tx, 1,
509 desc = xsk_ring_prod__tx_desc(&txq->tx, idx_tx);
510 desc->len = mbuf->pkt_len;
511 addr = (uint64_t)mbuf - (uint64_t)umem->buffer -
512 umem->mb_pool->header_size;
513 offset = rte_pktmbuf_mtod(mbuf, uint64_t) -
515 umem->mb_pool->header_size;
516 offset = offset << XSK_UNALIGNED_BUF_OFFSET_SHIFT;
517 desc->addr = addr | offset;
520 struct rte_mbuf *local_mbuf =
521 rte_pktmbuf_alloc(umem->mb_pool);
524 if (local_mbuf == NULL)
527 if (!xsk_ring_prod__reserve(&txq->tx, 1, &idx_tx)) {
528 rte_pktmbuf_free(local_mbuf);
533 desc = xsk_ring_prod__tx_desc(&txq->tx, idx_tx);
534 desc->len = mbuf->pkt_len;
536 addr = (uint64_t)local_mbuf - (uint64_t)umem->buffer -
537 umem->mb_pool->header_size;
538 offset = rte_pktmbuf_mtod(local_mbuf, uint64_t) -
539 (uint64_t)local_mbuf +
540 umem->mb_pool->header_size;
541 pkt = xsk_umem__get_data(umem->buffer, addr + offset);
542 offset = offset << XSK_UNALIGNED_BUF_OFFSET_SHIFT;
543 desc->addr = addr | offset;
544 rte_memcpy(pkt, rte_pktmbuf_mtod(mbuf, void *),
546 rte_pktmbuf_free(mbuf);
550 tx_bytes += mbuf->pkt_len;
556 xsk_ring_prod__submit(&txq->tx, count);
558 txq->stats.tx_pkts += count;
559 txq->stats.tx_bytes += tx_bytes;
560 txq->stats.tx_dropped += nb_pkts - count;
566 af_xdp_tx_cp(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
568 struct pkt_tx_queue *txq = queue;
569 struct xsk_umem_info *umem = txq->umem;
570 struct rte_mbuf *mbuf;
571 void *addrs[ETH_AF_XDP_TX_BATCH_SIZE];
572 unsigned long tx_bytes = 0;
575 struct xsk_ring_cons *cq = &txq->pair->cq;
577 pull_umem_cq(umem, nb_pkts, cq);
579 nb_pkts = rte_ring_dequeue_bulk(umem->buf_ring, addrs,
584 if (xsk_ring_prod__reserve(&txq->tx, nb_pkts, &idx_tx) != nb_pkts) {
586 rte_ring_enqueue_bulk(umem->buf_ring, addrs, nb_pkts, NULL);
590 for (i = 0; i < nb_pkts; i++) {
591 struct xdp_desc *desc;
594 desc = xsk_ring_prod__tx_desc(&txq->tx, idx_tx + i);
596 desc->len = mbuf->pkt_len;
598 desc->addr = (uint64_t)addrs[i];
599 pkt = xsk_umem__get_data(umem->mz->addr,
601 rte_memcpy(pkt, rte_pktmbuf_mtod(mbuf, void *), desc->len);
602 tx_bytes += mbuf->pkt_len;
603 rte_pktmbuf_free(mbuf);
606 xsk_ring_prod__submit(&txq->tx, nb_pkts);
610 txq->stats.tx_pkts += nb_pkts;
611 txq->stats.tx_bytes += tx_bytes;
617 af_xdp_tx_cp_batch(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
621 if (likely(nb_pkts <= ETH_AF_XDP_TX_BATCH_SIZE))
622 return af_xdp_tx_cp(queue, bufs, nb_pkts);
628 /* Split larger batch into smaller batches of size
629 * ETH_AF_XDP_TX_BATCH_SIZE or less.
631 n = (uint16_t)RTE_MIN(nb_pkts, ETH_AF_XDP_TX_BATCH_SIZE);
632 ret = af_xdp_tx_cp(queue, &bufs[nb_tx], n);
633 nb_tx = (uint16_t)(nb_tx + ret);
634 nb_pkts = (uint16_t)(nb_pkts - ret);
644 eth_af_xdp_tx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
646 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
647 return af_xdp_tx_zc(queue, bufs, nb_pkts);
649 return af_xdp_tx_cp_batch(queue, bufs, nb_pkts);
654 eth_dev_start(struct rte_eth_dev *dev)
656 dev->data->dev_link.link_status = ETH_LINK_UP;
661 /* This function gets called when the current port gets stopped. */
663 eth_dev_stop(struct rte_eth_dev *dev)
665 dev->data->dev_link.link_status = ETH_LINK_DOWN;
669 /* Find ethdev in list */
670 static inline struct internal_list *
671 find_internal_resource(struct pmd_internals *port_int)
674 struct internal_list *list = NULL;
676 if (port_int == NULL)
679 pthread_mutex_lock(&internal_list_lock);
681 TAILQ_FOREACH(list, &internal_list, next) {
682 struct pmd_internals *list_int =
683 list->eth_dev->data->dev_private;
684 if (list_int == port_int) {
690 pthread_mutex_unlock(&internal_list_lock);
698 /* Check if the netdev,qid context already exists */
700 ctx_exists(struct pkt_rx_queue *rxq, const char *ifname,
701 struct pkt_rx_queue *list_rxq, const char *list_ifname)
705 if (rxq->xsk_queue_idx == list_rxq->xsk_queue_idx &&
706 !strncmp(ifname, list_ifname, IFNAMSIZ)) {
707 AF_XDP_LOG(ERR, "ctx %s,%i already exists, cannot share umem\n",
708 ifname, rxq->xsk_queue_idx);
715 /* Get a pointer to an existing UMEM which overlays the rxq's mb_pool */
717 get_shared_umem(struct pkt_rx_queue *rxq, const char *ifname,
718 struct xsk_umem_info **umem)
720 struct internal_list *list;
721 struct pmd_internals *internals;
723 struct rte_mempool *mb_pool = rxq->mb_pool;
728 pthread_mutex_lock(&internal_list_lock);
730 TAILQ_FOREACH(list, &internal_list, next) {
731 internals = list->eth_dev->data->dev_private;
732 for (i = 0; i < internals->queue_cnt; i++) {
733 struct pkt_rx_queue *list_rxq =
734 &internals->rx_queues[i];
737 if (mb_pool == internals->rx_queues[i].mb_pool) {
738 if (ctx_exists(rxq, ifname, list_rxq,
739 internals->if_name)) {
744 &internals->rx_queues[i].umem->refcnt,
746 *umem = internals->rx_queues[i].umem;
754 pthread_mutex_unlock(&internal_list_lock);
760 eth_dev_configure(struct rte_eth_dev *dev)
762 struct pmd_internals *internal = dev->data->dev_private;
764 /* rx/tx must be paired */
765 if (dev->data->nb_rx_queues != dev->data->nb_tx_queues)
768 if (internal->shared_umem) {
769 struct internal_list *list = NULL;
770 const char *name = dev->device->name;
772 /* Ensure PMD is not already inserted into the list */
773 list = find_internal_resource(internal);
777 list = rte_zmalloc_socket(name, sizeof(*list), 0,
778 dev->device->numa_node);
783 pthread_mutex_lock(&internal_list_lock);
784 TAILQ_INSERT_TAIL(&internal_list, list, next);
785 pthread_mutex_unlock(&internal_list_lock);
792 eth_dev_info(struct rte_eth_dev *dev, struct rte_eth_dev_info *dev_info)
794 struct pmd_internals *internals = dev->data->dev_private;
796 dev_info->if_index = internals->if_index;
797 dev_info->max_mac_addrs = 1;
798 dev_info->max_rx_pktlen = ETH_FRAME_LEN;
799 dev_info->max_rx_queues = internals->queue_cnt;
800 dev_info->max_tx_queues = internals->queue_cnt;
802 dev_info->min_mtu = RTE_ETHER_MIN_MTU;
803 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
804 dev_info->max_mtu = getpagesize() -
805 sizeof(struct rte_mempool_objhdr) -
806 sizeof(struct rte_mbuf) -
807 RTE_PKTMBUF_HEADROOM - XDP_PACKET_HEADROOM;
809 dev_info->max_mtu = ETH_AF_XDP_FRAME_SIZE - XDP_PACKET_HEADROOM;
812 dev_info->default_rxportconf.burst_size = ETH_AF_XDP_DFLT_BUSY_BUDGET;
813 dev_info->default_txportconf.burst_size = ETH_AF_XDP_DFLT_BUSY_BUDGET;
814 dev_info->default_rxportconf.nb_queues = 1;
815 dev_info->default_txportconf.nb_queues = 1;
816 dev_info->default_rxportconf.ring_size = ETH_AF_XDP_DFLT_NUM_DESCS;
817 dev_info->default_txportconf.ring_size = ETH_AF_XDP_DFLT_NUM_DESCS;
823 eth_stats_get(struct rte_eth_dev *dev, struct rte_eth_stats *stats)
825 struct pmd_internals *internals = dev->data->dev_private;
826 struct xdp_statistics xdp_stats;
827 struct pkt_rx_queue *rxq;
828 struct pkt_tx_queue *txq;
832 for (i = 0; i < dev->data->nb_rx_queues; i++) {
833 optlen = sizeof(struct xdp_statistics);
834 rxq = &internals->rx_queues[i];
836 stats->q_ipackets[i] = rxq->stats.rx_pkts;
837 stats->q_ibytes[i] = rxq->stats.rx_bytes;
839 stats->q_opackets[i] = txq->stats.tx_pkts;
840 stats->q_obytes[i] = txq->stats.tx_bytes;
842 stats->ipackets += stats->q_ipackets[i];
843 stats->ibytes += stats->q_ibytes[i];
844 stats->imissed += rxq->stats.rx_dropped;
845 stats->oerrors += txq->stats.tx_dropped;
846 ret = getsockopt(xsk_socket__fd(rxq->xsk), SOL_XDP,
847 XDP_STATISTICS, &xdp_stats, &optlen);
849 AF_XDP_LOG(ERR, "getsockopt() failed for XDP_STATISTICS.\n");
852 stats->imissed += xdp_stats.rx_dropped;
854 stats->opackets += stats->q_opackets[i];
855 stats->obytes += stats->q_obytes[i];
862 eth_stats_reset(struct rte_eth_dev *dev)
864 struct pmd_internals *internals = dev->data->dev_private;
867 for (i = 0; i < internals->queue_cnt; i++) {
868 memset(&internals->rx_queues[i].stats, 0,
869 sizeof(struct rx_stats));
870 memset(&internals->tx_queues[i].stats, 0,
871 sizeof(struct tx_stats));
878 remove_xdp_program(struct pmd_internals *internals)
880 uint32_t curr_prog_id = 0;
882 if (bpf_get_link_xdp_id(internals->if_index, &curr_prog_id,
883 XDP_FLAGS_UPDATE_IF_NOEXIST)) {
884 AF_XDP_LOG(ERR, "bpf_get_link_xdp_id failed\n");
887 bpf_set_link_xdp_fd(internals->if_index, -1,
888 XDP_FLAGS_UPDATE_IF_NOEXIST);
892 xdp_umem_destroy(struct xsk_umem_info *umem)
894 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
895 umem->mb_pool = NULL;
897 rte_memzone_free(umem->mz);
900 rte_ring_free(umem->buf_ring);
901 umem->buf_ring = NULL;
908 eth_dev_close(struct rte_eth_dev *dev)
910 struct pmd_internals *internals = dev->data->dev_private;
911 struct pkt_rx_queue *rxq;
914 if (rte_eal_process_type() != RTE_PROC_PRIMARY)
917 AF_XDP_LOG(INFO, "Closing AF_XDP ethdev on numa socket %u\n",
920 for (i = 0; i < internals->queue_cnt; i++) {
921 rxq = &internals->rx_queues[i];
922 if (rxq->umem == NULL)
924 xsk_socket__delete(rxq->xsk);
926 if (__atomic_sub_fetch(&rxq->umem->refcnt, 1, __ATOMIC_ACQUIRE)
928 (void)xsk_umem__delete(rxq->umem->umem);
929 xdp_umem_destroy(rxq->umem);
932 /* free pkt_tx_queue */
938 * MAC is not allocated dynamically, setting it to NULL would prevent
939 * from releasing it in rte_eth_dev_release_port.
941 dev->data->mac_addrs = NULL;
943 remove_xdp_program(internals);
945 if (internals->shared_umem) {
946 struct internal_list *list;
948 /* Remove ethdev from list used to track and share UMEMs */
949 list = find_internal_resource(internals);
951 pthread_mutex_lock(&internal_list_lock);
952 TAILQ_REMOVE(&internal_list, list, next);
953 pthread_mutex_unlock(&internal_list_lock);
962 eth_queue_release(void *q __rte_unused)
967 eth_link_update(struct rte_eth_dev *dev __rte_unused,
968 int wait_to_complete __rte_unused)
973 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
974 static inline uintptr_t get_base_addr(struct rte_mempool *mp, uint64_t *align)
976 struct rte_mempool_memhdr *memhdr;
977 uintptr_t memhdr_addr, aligned_addr;
979 memhdr = STAILQ_FIRST(&mp->mem_list);
980 memhdr_addr = (uintptr_t)memhdr->addr;
981 aligned_addr = memhdr_addr & ~(getpagesize() - 1);
982 *align = memhdr_addr - aligned_addr;
988 xsk_umem_info *xdp_umem_configure(struct pmd_internals *internals,
989 struct pkt_rx_queue *rxq)
991 struct xsk_umem_info *umem = NULL;
993 struct xsk_umem_config usr_config = {
994 .fill_size = ETH_AF_XDP_DFLT_NUM_DESCS * 2,
995 .comp_size = ETH_AF_XDP_DFLT_NUM_DESCS,
996 .flags = XDP_UMEM_UNALIGNED_CHUNK_FLAG};
997 void *base_addr = NULL;
998 struct rte_mempool *mb_pool = rxq->mb_pool;
999 uint64_t umem_size, align = 0;
1001 if (internals->shared_umem) {
1002 if (get_shared_umem(rxq, internals->if_name, &umem) < 0)
1006 __atomic_load_n(&umem->refcnt, __ATOMIC_ACQUIRE) <
1008 AF_XDP_LOG(INFO, "%s,qid%i sharing UMEM\n",
1009 internals->if_name, rxq->xsk_queue_idx);
1010 __atomic_fetch_add(&umem->refcnt, 1, __ATOMIC_ACQUIRE);
1015 usr_config.frame_size =
1016 rte_mempool_calc_obj_size(mb_pool->elt_size,
1017 mb_pool->flags, NULL);
1018 usr_config.frame_headroom = mb_pool->header_size +
1019 sizeof(struct rte_mbuf) +
1020 rte_pktmbuf_priv_size(mb_pool) +
1021 RTE_PKTMBUF_HEADROOM;
1023 umem = rte_zmalloc_socket("umem", sizeof(*umem), 0,
1026 AF_XDP_LOG(ERR, "Failed to allocate umem info");
1030 umem->mb_pool = mb_pool;
1031 base_addr = (void *)get_base_addr(mb_pool, &align);
1032 umem_size = (uint64_t)mb_pool->populated_size *
1033 (uint64_t)usr_config.frame_size +
1036 ret = xsk_umem__create(&umem->umem, base_addr, umem_size,
1037 &rxq->fq, &rxq->cq, &usr_config);
1039 AF_XDP_LOG(ERR, "Failed to create umem");
1042 umem->buffer = base_addr;
1044 if (internals->shared_umem) {
1045 umem->max_xsks = mb_pool->populated_size /
1046 ETH_AF_XDP_NUM_BUFFERS;
1047 AF_XDP_LOG(INFO, "Max xsks for UMEM %s: %u\n",
1048 mb_pool->name, umem->max_xsks);
1051 __atomic_store_n(&umem->refcnt, 1, __ATOMIC_RELEASE);
1056 xsk_umem_info *xdp_umem_configure(struct pmd_internals *internals,
1057 struct pkt_rx_queue *rxq)
1059 struct xsk_umem_info *umem;
1060 const struct rte_memzone *mz;
1061 struct xsk_umem_config usr_config = {
1062 .fill_size = ETH_AF_XDP_DFLT_NUM_DESCS,
1063 .comp_size = ETH_AF_XDP_DFLT_NUM_DESCS,
1064 .frame_size = ETH_AF_XDP_FRAME_SIZE,
1065 .frame_headroom = 0 };
1066 char ring_name[RTE_RING_NAMESIZE];
1067 char mz_name[RTE_MEMZONE_NAMESIZE];
1071 umem = rte_zmalloc_socket("umem", sizeof(*umem), 0, rte_socket_id());
1073 AF_XDP_LOG(ERR, "Failed to allocate umem info");
1077 snprintf(ring_name, sizeof(ring_name), "af_xdp_ring_%s_%u",
1078 internals->if_name, rxq->xsk_queue_idx);
1079 umem->buf_ring = rte_ring_create(ring_name,
1080 ETH_AF_XDP_NUM_BUFFERS,
1083 if (umem->buf_ring == NULL) {
1084 AF_XDP_LOG(ERR, "Failed to create rte_ring\n");
1088 for (i = 0; i < ETH_AF_XDP_NUM_BUFFERS; i++)
1089 rte_ring_enqueue(umem->buf_ring,
1090 (void *)(i * ETH_AF_XDP_FRAME_SIZE));
1092 snprintf(mz_name, sizeof(mz_name), "af_xdp_umem_%s_%u",
1093 internals->if_name, rxq->xsk_queue_idx);
1094 mz = rte_memzone_reserve_aligned(mz_name,
1095 ETH_AF_XDP_NUM_BUFFERS * ETH_AF_XDP_FRAME_SIZE,
1096 rte_socket_id(), RTE_MEMZONE_IOVA_CONTIG,
1099 AF_XDP_LOG(ERR, "Failed to reserve memzone for af_xdp umem.\n");
1103 ret = xsk_umem__create(&umem->umem, mz->addr,
1104 ETH_AF_XDP_NUM_BUFFERS * ETH_AF_XDP_FRAME_SIZE,
1109 AF_XDP_LOG(ERR, "Failed to create umem");
1118 xdp_umem_destroy(umem);
1123 load_custom_xdp_prog(const char *prog_path, int if_index)
1125 int ret, prog_fd = -1;
1126 struct bpf_object *obj;
1127 struct bpf_map *map;
1129 ret = bpf_prog_load(prog_path, BPF_PROG_TYPE_XDP, &obj, &prog_fd);
1131 AF_XDP_LOG(ERR, "Failed to load program %s\n", prog_path);
1136 * The loaded program must provision for a map of xsks, such that some
1137 * traffic can be redirected to userspace. When the xsk is created,
1138 * libbpf inserts it into the map.
1140 map = bpf_object__find_map_by_name(obj, "xsks_map");
1142 AF_XDP_LOG(ERR, "Failed to find xsks_map in %s\n", prog_path);
1146 /* Link the program with the given network device */
1147 ret = bpf_set_link_xdp_fd(if_index, prog_fd,
1148 XDP_FLAGS_UPDATE_IF_NOEXIST);
1150 AF_XDP_LOG(ERR, "Failed to set prog fd %d on interface\n",
1155 AF_XDP_LOG(INFO, "Successfully loaded XDP program %s with fd %d\n",
1156 prog_path, prog_fd);
1161 /* Detect support for busy polling through setsockopt(). */
1163 configure_preferred_busy_poll(struct pkt_rx_queue *rxq)
1166 int fd = xsk_socket__fd(rxq->xsk);
1169 ret = setsockopt(fd, SOL_SOCKET, SO_PREFER_BUSY_POLL,
1170 (void *)&sock_opt, sizeof(sock_opt));
1172 AF_XDP_LOG(DEBUG, "Failed to set SO_PREFER_BUSY_POLL\n");
1176 sock_opt = ETH_AF_XDP_DFLT_BUSY_TIMEOUT;
1177 ret = setsockopt(fd, SOL_SOCKET, SO_BUSY_POLL, (void *)&sock_opt,
1180 AF_XDP_LOG(DEBUG, "Failed to set SO_BUSY_POLL\n");
1184 sock_opt = rxq->busy_budget;
1185 ret = setsockopt(fd, SOL_SOCKET, SO_BUSY_POLL_BUDGET,
1186 (void *)&sock_opt, sizeof(sock_opt));
1188 AF_XDP_LOG(DEBUG, "Failed to set SO_BUSY_POLL_BUDGET\n");
1190 AF_XDP_LOG(INFO, "Busy polling budget set to: %u\n",
1195 /* setsockopt failure - attempt to restore xsk to default state and
1196 * proceed without busy polling support.
1199 ret = setsockopt(fd, SOL_SOCKET, SO_BUSY_POLL, (void *)&sock_opt,
1202 AF_XDP_LOG(ERR, "Failed to unset SO_BUSY_POLL\n");
1208 ret = setsockopt(fd, SOL_SOCKET, SO_PREFER_BUSY_POLL,
1209 (void *)&sock_opt, sizeof(sock_opt));
1211 AF_XDP_LOG(ERR, "Failed to unset SO_PREFER_BUSY_POLL\n");
1216 rxq->busy_budget = 0;
1221 xsk_configure(struct pmd_internals *internals, struct pkt_rx_queue *rxq,
1224 struct xsk_socket_config cfg;
1225 struct pkt_tx_queue *txq = rxq->pair;
1227 int reserve_size = ETH_AF_XDP_DFLT_NUM_DESCS;
1228 struct rte_mbuf *fq_bufs[reserve_size];
1230 rxq->umem = xdp_umem_configure(internals, rxq);
1231 if (rxq->umem == NULL)
1233 txq->umem = rxq->umem;
1235 cfg.rx_size = ring_size;
1236 cfg.tx_size = ring_size;
1237 cfg.libbpf_flags = 0;
1238 cfg.xdp_flags = XDP_FLAGS_UPDATE_IF_NOEXIST;
1241 #if defined(XDP_USE_NEED_WAKEUP)
1242 cfg.bind_flags |= XDP_USE_NEED_WAKEUP;
1245 if (strnlen(internals->prog_path, PATH_MAX) &&
1246 !internals->custom_prog_configured) {
1247 ret = load_custom_xdp_prog(internals->prog_path,
1248 internals->if_index);
1250 AF_XDP_LOG(ERR, "Failed to load custom XDP program %s\n",
1251 internals->prog_path);
1254 internals->custom_prog_configured = 1;
1257 if (internals->shared_umem)
1258 ret = create_shared_socket(&rxq->xsk, internals->if_name,
1259 rxq->xsk_queue_idx, rxq->umem->umem, &rxq->rx,
1260 &txq->tx, &rxq->fq, &rxq->cq, &cfg);
1262 ret = xsk_socket__create(&rxq->xsk, internals->if_name,
1263 rxq->xsk_queue_idx, rxq->umem->umem, &rxq->rx,
1267 AF_XDP_LOG(ERR, "Failed to create xsk socket.\n");
1271 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
1272 ret = rte_pktmbuf_alloc_bulk(rxq->umem->mb_pool, fq_bufs, reserve_size);
1274 AF_XDP_LOG(DEBUG, "Failed to get enough buffers for fq.\n");
1279 if (rxq->busy_budget) {
1280 ret = configure_preferred_busy_poll(rxq);
1282 AF_XDP_LOG(ERR, "Failed configure busy polling.\n");
1287 ret = reserve_fill_queue(rxq->umem, reserve_size, fq_bufs, &rxq->fq);
1289 xsk_socket__delete(rxq->xsk);
1290 AF_XDP_LOG(ERR, "Failed to reserve fill queue.\n");
1297 if (__atomic_sub_fetch(&rxq->umem->refcnt, 1, __ATOMIC_ACQUIRE) == 0)
1298 xdp_umem_destroy(rxq->umem);
1304 eth_rx_queue_setup(struct rte_eth_dev *dev,
1305 uint16_t rx_queue_id,
1306 uint16_t nb_rx_desc,
1307 unsigned int socket_id __rte_unused,
1308 const struct rte_eth_rxconf *rx_conf __rte_unused,
1309 struct rte_mempool *mb_pool)
1311 struct pmd_internals *internals = dev->data->dev_private;
1312 struct pkt_rx_queue *rxq;
1315 rxq = &internals->rx_queues[rx_queue_id];
1317 AF_XDP_LOG(INFO, "Set up rx queue, rx queue id: %d, xsk queue id: %d\n",
1318 rx_queue_id, rxq->xsk_queue_idx);
1320 #ifndef XDP_UMEM_UNALIGNED_CHUNK_FLAG
1321 uint32_t buf_size, data_size;
1323 /* Now get the space available for data in the mbuf */
1324 buf_size = rte_pktmbuf_data_room_size(mb_pool) -
1325 RTE_PKTMBUF_HEADROOM;
1326 data_size = ETH_AF_XDP_FRAME_SIZE;
1328 if (data_size > buf_size) {
1329 AF_XDP_LOG(ERR, "%s: %d bytes will not fit in mbuf (%d bytes)\n",
1330 dev->device->name, data_size, buf_size);
1336 rxq->mb_pool = mb_pool;
1338 if (xsk_configure(internals, rxq, nb_rx_desc)) {
1339 AF_XDP_LOG(ERR, "Failed to configure xdp socket\n");
1344 if (!rxq->busy_budget)
1345 AF_XDP_LOG(DEBUG, "Preferred busy polling not enabled\n");
1347 rxq->fds[0].fd = xsk_socket__fd(rxq->xsk);
1348 rxq->fds[0].events = POLLIN;
1350 dev->data->rx_queues[rx_queue_id] = rxq;
1358 eth_tx_queue_setup(struct rte_eth_dev *dev,
1359 uint16_t tx_queue_id,
1360 uint16_t nb_tx_desc __rte_unused,
1361 unsigned int socket_id __rte_unused,
1362 const struct rte_eth_txconf *tx_conf __rte_unused)
1364 struct pmd_internals *internals = dev->data->dev_private;
1365 struct pkt_tx_queue *txq;
1367 txq = &internals->tx_queues[tx_queue_id];
1369 dev->data->tx_queues[tx_queue_id] = txq;
1374 eth_dev_mtu_set(struct rte_eth_dev *dev, uint16_t mtu)
1376 struct pmd_internals *internals = dev->data->dev_private;
1377 struct ifreq ifr = { .ifr_mtu = mtu };
1381 s = socket(PF_INET, SOCK_DGRAM, 0);
1385 strlcpy(ifr.ifr_name, internals->if_name, IFNAMSIZ);
1386 ret = ioctl(s, SIOCSIFMTU, &ifr);
1389 return (ret < 0) ? -errno : 0;
1393 eth_dev_change_flags(char *if_name, uint32_t flags, uint32_t mask)
1399 s = socket(PF_INET, SOCK_DGRAM, 0);
1403 strlcpy(ifr.ifr_name, if_name, IFNAMSIZ);
1404 if (ioctl(s, SIOCGIFFLAGS, &ifr) < 0) {
1408 ifr.ifr_flags &= mask;
1409 ifr.ifr_flags |= flags;
1410 if (ioctl(s, SIOCSIFFLAGS, &ifr) < 0) {
1420 eth_dev_promiscuous_enable(struct rte_eth_dev *dev)
1422 struct pmd_internals *internals = dev->data->dev_private;
1424 return eth_dev_change_flags(internals->if_name, IFF_PROMISC, ~0);
1428 eth_dev_promiscuous_disable(struct rte_eth_dev *dev)
1430 struct pmd_internals *internals = dev->data->dev_private;
1432 return eth_dev_change_flags(internals->if_name, 0, ~IFF_PROMISC);
1435 static const struct eth_dev_ops ops = {
1436 .dev_start = eth_dev_start,
1437 .dev_stop = eth_dev_stop,
1438 .dev_close = eth_dev_close,
1439 .dev_configure = eth_dev_configure,
1440 .dev_infos_get = eth_dev_info,
1441 .mtu_set = eth_dev_mtu_set,
1442 .promiscuous_enable = eth_dev_promiscuous_enable,
1443 .promiscuous_disable = eth_dev_promiscuous_disable,
1444 .rx_queue_setup = eth_rx_queue_setup,
1445 .tx_queue_setup = eth_tx_queue_setup,
1446 .rx_queue_release = eth_queue_release,
1447 .tx_queue_release = eth_queue_release,
1448 .link_update = eth_link_update,
1449 .stats_get = eth_stats_get,
1450 .stats_reset = eth_stats_reset,
1453 /** parse busy_budget argument */
1455 parse_budget_arg(const char *key __rte_unused,
1456 const char *value, void *extra_args)
1458 int *i = (int *)extra_args;
1461 *i = strtol(value, &end, 10);
1462 if (*i < 0 || *i > UINT16_MAX) {
1463 AF_XDP_LOG(ERR, "Invalid busy_budget %i, must be >= 0 and <= %u\n",
1471 /** parse integer from integer argument */
1473 parse_integer_arg(const char *key __rte_unused,
1474 const char *value, void *extra_args)
1476 int *i = (int *)extra_args;
1479 *i = strtol(value, &end, 10);
1481 AF_XDP_LOG(ERR, "Argument has to be positive.\n");
1488 /** parse name argument */
1490 parse_name_arg(const char *key __rte_unused,
1491 const char *value, void *extra_args)
1493 char *name = extra_args;
1495 if (strnlen(value, IFNAMSIZ) > IFNAMSIZ - 1) {
1496 AF_XDP_LOG(ERR, "Invalid name %s, should be less than %u bytes.\n",
1501 strlcpy(name, value, IFNAMSIZ);
1506 /** parse xdp prog argument */
1508 parse_prog_arg(const char *key __rte_unused,
1509 const char *value, void *extra_args)
1511 char *path = extra_args;
1513 if (strnlen(value, PATH_MAX) == PATH_MAX) {
1514 AF_XDP_LOG(ERR, "Invalid path %s, should be less than %u bytes.\n",
1519 if (access(value, F_OK) != 0) {
1520 AF_XDP_LOG(ERR, "Error accessing %s: %s\n",
1521 value, strerror(errno));
1525 strlcpy(path, value, PATH_MAX);
1531 xdp_get_channels_info(const char *if_name, int *max_queues,
1532 int *combined_queues)
1534 struct ethtool_channels channels;
1538 fd = socket(AF_INET, SOCK_DGRAM, 0);
1542 channels.cmd = ETHTOOL_GCHANNELS;
1543 ifr.ifr_data = (void *)&channels;
1544 strlcpy(ifr.ifr_name, if_name, IFNAMSIZ);
1545 ret = ioctl(fd, SIOCETHTOOL, &ifr);
1547 if (errno == EOPNOTSUPP) {
1555 if (channels.max_combined == 0 || errno == EOPNOTSUPP) {
1556 /* If the device says it has no channels, then all traffic
1557 * is sent to a single stream, so max queues = 1.
1560 *combined_queues = 1;
1562 *max_queues = channels.max_combined;
1563 *combined_queues = channels.combined_count;
1572 parse_parameters(struct rte_kvargs *kvlist, char *if_name, int *start_queue,
1573 int *queue_cnt, int *shared_umem, char *prog_path,
1578 ret = rte_kvargs_process(kvlist, ETH_AF_XDP_IFACE_ARG,
1579 &parse_name_arg, if_name);
1583 ret = rte_kvargs_process(kvlist, ETH_AF_XDP_START_QUEUE_ARG,
1584 &parse_integer_arg, start_queue);
1588 ret = rte_kvargs_process(kvlist, ETH_AF_XDP_QUEUE_COUNT_ARG,
1589 &parse_integer_arg, queue_cnt);
1590 if (ret < 0 || *queue_cnt <= 0) {
1595 ret = rte_kvargs_process(kvlist, ETH_AF_XDP_SHARED_UMEM_ARG,
1596 &parse_integer_arg, shared_umem);
1600 ret = rte_kvargs_process(kvlist, ETH_AF_XDP_PROG_ARG,
1601 &parse_prog_arg, prog_path);
1605 ret = rte_kvargs_process(kvlist, ETH_AF_XDP_BUDGET_ARG,
1606 &parse_budget_arg, busy_budget);
1611 rte_kvargs_free(kvlist);
1616 get_iface_info(const char *if_name,
1617 struct rte_ether_addr *eth_addr,
1621 int sock = socket(AF_INET, SOCK_DGRAM, IPPROTO_IP);
1626 strlcpy(ifr.ifr_name, if_name, IFNAMSIZ);
1627 if (ioctl(sock, SIOCGIFINDEX, &ifr))
1630 *if_index = ifr.ifr_ifindex;
1632 if (ioctl(sock, SIOCGIFHWADDR, &ifr))
1635 rte_memcpy(eth_addr, ifr.ifr_hwaddr.sa_data, RTE_ETHER_ADDR_LEN);
1645 static struct rte_eth_dev *
1646 init_internals(struct rte_vdev_device *dev, const char *if_name,
1647 int start_queue_idx, int queue_cnt, int shared_umem,
1648 const char *prog_path, int busy_budget)
1650 const char *name = rte_vdev_device_name(dev);
1651 const unsigned int numa_node = dev->device.numa_node;
1652 struct pmd_internals *internals;
1653 struct rte_eth_dev *eth_dev;
1657 internals = rte_zmalloc_socket(name, sizeof(*internals), 0, numa_node);
1658 if (internals == NULL)
1661 internals->start_queue_idx = start_queue_idx;
1662 internals->queue_cnt = queue_cnt;
1663 strlcpy(internals->if_name, if_name, IFNAMSIZ);
1664 strlcpy(internals->prog_path, prog_path, PATH_MAX);
1665 internals->custom_prog_configured = 0;
1667 #ifndef ETH_AF_XDP_SHARED_UMEM
1669 AF_XDP_LOG(ERR, "Shared UMEM feature not available. "
1670 "Check kernel and libbpf version\n");
1671 goto err_free_internals;
1674 internals->shared_umem = shared_umem;
1676 if (xdp_get_channels_info(if_name, &internals->max_queue_cnt,
1677 &internals->combined_queue_cnt)) {
1678 AF_XDP_LOG(ERR, "Failed to get channel info of interface: %s\n",
1680 goto err_free_internals;
1683 if (queue_cnt > internals->combined_queue_cnt) {
1684 AF_XDP_LOG(ERR, "Specified queue count %d is larger than combined queue count %d.\n",
1685 queue_cnt, internals->combined_queue_cnt);
1686 goto err_free_internals;
1689 internals->rx_queues = rte_zmalloc_socket(NULL,
1690 sizeof(struct pkt_rx_queue) * queue_cnt,
1692 if (internals->rx_queues == NULL) {
1693 AF_XDP_LOG(ERR, "Failed to allocate memory for rx queues.\n");
1694 goto err_free_internals;
1697 internals->tx_queues = rte_zmalloc_socket(NULL,
1698 sizeof(struct pkt_tx_queue) * queue_cnt,
1700 if (internals->tx_queues == NULL) {
1701 AF_XDP_LOG(ERR, "Failed to allocate memory for tx queues.\n");
1704 for (i = 0; i < queue_cnt; i++) {
1705 internals->tx_queues[i].pair = &internals->rx_queues[i];
1706 internals->rx_queues[i].pair = &internals->tx_queues[i];
1707 internals->rx_queues[i].xsk_queue_idx = start_queue_idx + i;
1708 internals->tx_queues[i].xsk_queue_idx = start_queue_idx + i;
1709 internals->rx_queues[i].busy_budget = busy_budget;
1712 ret = get_iface_info(if_name, &internals->eth_addr,
1713 &internals->if_index);
1717 eth_dev = rte_eth_vdev_allocate(dev, 0);
1718 if (eth_dev == NULL)
1721 eth_dev->data->dev_private = internals;
1722 eth_dev->data->dev_link = pmd_link;
1723 eth_dev->data->mac_addrs = &internals->eth_addr;
1724 eth_dev->data->dev_flags |= RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS;
1725 eth_dev->dev_ops = &ops;
1726 eth_dev->rx_pkt_burst = eth_af_xdp_rx;
1727 eth_dev->tx_pkt_burst = eth_af_xdp_tx;
1729 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
1730 AF_XDP_LOG(INFO, "Zero copy between umem and mbuf enabled.\n");
1736 rte_free(internals->tx_queues);
1738 rte_free(internals->rx_queues);
1740 rte_free(internals);
1745 rte_pmd_af_xdp_probe(struct rte_vdev_device *dev)
1747 struct rte_kvargs *kvlist;
1748 char if_name[IFNAMSIZ] = {'\0'};
1749 int xsk_start_queue_idx = ETH_AF_XDP_DFLT_START_QUEUE_IDX;
1750 int xsk_queue_cnt = ETH_AF_XDP_DFLT_QUEUE_COUNT;
1751 int shared_umem = 0;
1752 char prog_path[PATH_MAX] = {'\0'};
1753 int busy_budget = -1;
1754 struct rte_eth_dev *eth_dev = NULL;
1757 AF_XDP_LOG(INFO, "Initializing pmd_af_xdp for %s\n",
1758 rte_vdev_device_name(dev));
1760 name = rte_vdev_device_name(dev);
1761 if (rte_eal_process_type() == RTE_PROC_SECONDARY &&
1762 strlen(rte_vdev_device_args(dev)) == 0) {
1763 eth_dev = rte_eth_dev_attach_secondary(name);
1764 if (eth_dev == NULL) {
1765 AF_XDP_LOG(ERR, "Failed to probe %s\n", name);
1768 eth_dev->dev_ops = &ops;
1769 rte_eth_dev_probing_finish(eth_dev);
1773 kvlist = rte_kvargs_parse(rte_vdev_device_args(dev), valid_arguments);
1774 if (kvlist == NULL) {
1775 AF_XDP_LOG(ERR, "Invalid kvargs key\n");
1779 if (dev->device.numa_node == SOCKET_ID_ANY)
1780 dev->device.numa_node = rte_socket_id();
1782 if (parse_parameters(kvlist, if_name, &xsk_start_queue_idx,
1783 &xsk_queue_cnt, &shared_umem, prog_path,
1784 &busy_budget) < 0) {
1785 AF_XDP_LOG(ERR, "Invalid kvargs value\n");
1789 if (strlen(if_name) == 0) {
1790 AF_XDP_LOG(ERR, "Network interface must be specified\n");
1794 busy_budget = busy_budget == -1 ? ETH_AF_XDP_DFLT_BUSY_BUDGET :
1797 eth_dev = init_internals(dev, if_name, xsk_start_queue_idx,
1798 xsk_queue_cnt, shared_umem, prog_path,
1800 if (eth_dev == NULL) {
1801 AF_XDP_LOG(ERR, "Failed to init internals\n");
1805 rte_eth_dev_probing_finish(eth_dev);
1811 rte_pmd_af_xdp_remove(struct rte_vdev_device *dev)
1813 struct rte_eth_dev *eth_dev = NULL;
1815 AF_XDP_LOG(INFO, "Removing AF_XDP ethdev on numa socket %u\n",
1821 /* find the ethdev entry */
1822 eth_dev = rte_eth_dev_allocated(rte_vdev_device_name(dev));
1823 if (eth_dev == NULL)
1826 eth_dev_close(eth_dev);
1827 rte_eth_dev_release_port(eth_dev);
1833 static struct rte_vdev_driver pmd_af_xdp_drv = {
1834 .probe = rte_pmd_af_xdp_probe,
1835 .remove = rte_pmd_af_xdp_remove,
1838 RTE_PMD_REGISTER_VDEV(net_af_xdp, pmd_af_xdp_drv);
1839 RTE_PMD_REGISTER_PARAM_STRING(net_af_xdp,
1841 "start_queue=<int> "
1842 "queue_count=<int> "
1843 "shared_umem=<int> "
1844 "xdp_prog=<string> "
1845 "busy_budget=<int>");