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>
40 #include <rte_power_intrinsics.h>
44 #ifndef SO_PREFER_BUSY_POLL
45 #define SO_PREFER_BUSY_POLL 69
47 #ifndef SO_BUSY_POLL_BUDGET
48 #define SO_BUSY_POLL_BUDGET 70
64 RTE_LOG_REGISTER_DEFAULT(af_xdp_logtype, NOTICE);
66 #define AF_XDP_LOG(level, fmt, args...) \
67 rte_log(RTE_LOG_ ## level, af_xdp_logtype, \
68 "%s(): " fmt, __func__, ##args)
70 #define ETH_AF_XDP_FRAME_SIZE 2048
71 #define ETH_AF_XDP_NUM_BUFFERS 4096
72 #define ETH_AF_XDP_DFLT_NUM_DESCS XSK_RING_CONS__DEFAULT_NUM_DESCS
73 #define ETH_AF_XDP_DFLT_START_QUEUE_IDX 0
74 #define ETH_AF_XDP_DFLT_QUEUE_COUNT 1
75 #define ETH_AF_XDP_DFLT_BUSY_BUDGET 64
76 #define ETH_AF_XDP_DFLT_BUSY_TIMEOUT 20
78 #define ETH_AF_XDP_RX_BATCH_SIZE XSK_RING_CONS__DEFAULT_NUM_DESCS
79 #define ETH_AF_XDP_TX_BATCH_SIZE XSK_RING_CONS__DEFAULT_NUM_DESCS
81 #define ETH_AF_XDP_ETH_OVERHEAD (RTE_ETHER_HDR_LEN + RTE_ETHER_CRC_LEN)
83 struct xsk_umem_info {
84 struct xsk_umem *umem;
85 struct rte_ring *buf_ring;
86 const struct rte_memzone *mz;
87 struct rte_mempool *mb_pool;
100 struct xsk_ring_cons rx;
101 struct xsk_umem_info *umem;
102 struct xsk_socket *xsk;
103 struct rte_mempool *mb_pool;
105 struct rx_stats stats;
107 struct xsk_ring_prod fq;
108 struct xsk_ring_cons cq;
110 struct pkt_tx_queue *pair;
111 struct pollfd fds[1];
122 struct pkt_tx_queue {
123 struct xsk_ring_prod tx;
124 struct xsk_umem_info *umem;
126 struct tx_stats stats;
128 struct pkt_rx_queue *pair;
132 struct pmd_internals {
134 char if_name[IFNAMSIZ];
138 int combined_queue_cnt;
140 char prog_path[PATH_MAX];
141 bool custom_prog_configured;
143 struct rte_ether_addr eth_addr;
145 struct pkt_rx_queue *rx_queues;
146 struct pkt_tx_queue *tx_queues;
149 #define ETH_AF_XDP_IFACE_ARG "iface"
150 #define ETH_AF_XDP_START_QUEUE_ARG "start_queue"
151 #define ETH_AF_XDP_QUEUE_COUNT_ARG "queue_count"
152 #define ETH_AF_XDP_SHARED_UMEM_ARG "shared_umem"
153 #define ETH_AF_XDP_PROG_ARG "xdp_prog"
154 #define ETH_AF_XDP_BUDGET_ARG "busy_budget"
156 static const char * const valid_arguments[] = {
157 ETH_AF_XDP_IFACE_ARG,
158 ETH_AF_XDP_START_QUEUE_ARG,
159 ETH_AF_XDP_QUEUE_COUNT_ARG,
160 ETH_AF_XDP_SHARED_UMEM_ARG,
162 ETH_AF_XDP_BUDGET_ARG,
166 static const struct rte_eth_link pmd_link = {
167 .link_speed = RTE_ETH_SPEED_NUM_10G,
168 .link_duplex = RTE_ETH_LINK_FULL_DUPLEX,
169 .link_status = RTE_ETH_LINK_DOWN,
170 .link_autoneg = RTE_ETH_LINK_AUTONEG
173 /* List which tracks PMDs to facilitate sharing UMEMs across them. */
174 struct internal_list {
175 TAILQ_ENTRY(internal_list) next;
176 struct rte_eth_dev *eth_dev;
179 TAILQ_HEAD(internal_list_head, internal_list);
180 static struct internal_list_head internal_list =
181 TAILQ_HEAD_INITIALIZER(internal_list);
183 static pthread_mutex_t internal_list_lock = PTHREAD_MUTEX_INITIALIZER;
185 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
187 reserve_fill_queue_zc(struct xsk_umem_info *umem, uint16_t reserve_size,
188 struct rte_mbuf **bufs, struct xsk_ring_prod *fq)
193 if (unlikely(!xsk_ring_prod__reserve(fq, reserve_size, &idx))) {
194 for (i = 0; i < reserve_size; i++)
195 rte_pktmbuf_free(bufs[i]);
196 AF_XDP_LOG(DEBUG, "Failed to reserve enough fq descs.\n");
200 for (i = 0; i < reserve_size; i++) {
204 fq_addr = xsk_ring_prod__fill_addr(fq, idx++);
205 addr = (uint64_t)bufs[i] - (uint64_t)umem->buffer -
206 umem->mb_pool->header_size;
210 xsk_ring_prod__submit(fq, reserve_size);
216 reserve_fill_queue_cp(struct xsk_umem_info *umem, uint16_t reserve_size,
217 struct rte_mbuf **bufs __rte_unused,
218 struct xsk_ring_prod *fq)
220 void *addrs[reserve_size];
224 if (rte_ring_dequeue_bulk(umem->buf_ring, addrs, reserve_size, NULL)
226 AF_XDP_LOG(DEBUG, "Failed to get enough buffers for fq.\n");
230 if (unlikely(!xsk_ring_prod__reserve(fq, reserve_size, &idx))) {
231 AF_XDP_LOG(DEBUG, "Failed to reserve enough fq descs.\n");
232 rte_ring_enqueue_bulk(umem->buf_ring, addrs,
237 for (i = 0; i < reserve_size; i++) {
240 fq_addr = xsk_ring_prod__fill_addr(fq, idx++);
241 *fq_addr = (uint64_t)addrs[i];
244 xsk_ring_prod__submit(fq, reserve_size);
251 reserve_fill_queue(struct xsk_umem_info *umem, uint16_t reserve_size,
252 struct rte_mbuf **bufs, struct xsk_ring_prod *fq)
254 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
255 return reserve_fill_queue_zc(umem, reserve_size, bufs, fq);
257 return reserve_fill_queue_cp(umem, reserve_size, bufs, fq);
261 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
263 af_xdp_rx_zc(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
265 struct pkt_rx_queue *rxq = queue;
266 struct xsk_ring_cons *rx = &rxq->rx;
267 struct xsk_ring_prod *fq = &rxq->fq;
268 struct xsk_umem_info *umem = rxq->umem;
270 unsigned long rx_bytes = 0;
272 struct rte_mbuf *fq_bufs[ETH_AF_XDP_RX_BATCH_SIZE];
274 nb_pkts = xsk_ring_cons__peek(rx, nb_pkts, &idx_rx);
277 /* we can assume a kernel >= 5.11 is in use if busy polling is
278 * enabled and thus we can safely use the recvfrom() syscall
279 * which is only supported for AF_XDP sockets in kernels >=
282 if (rxq->busy_budget) {
283 (void)recvfrom(xsk_socket__fd(rxq->xsk), NULL, 0,
284 MSG_DONTWAIT, NULL, NULL);
285 } else if (xsk_ring_prod__needs_wakeup(fq)) {
286 (void)poll(&rxq->fds[0], 1, 1000);
292 /* allocate bufs for fill queue replenishment after rx */
293 if (rte_pktmbuf_alloc_bulk(umem->mb_pool, fq_bufs, nb_pkts)) {
295 "Failed to get enough buffers for fq.\n");
296 /* rollback cached_cons which is added by
297 * xsk_ring_cons__peek
299 rx->cached_cons -= nb_pkts;
303 for (i = 0; i < nb_pkts; i++) {
304 const struct xdp_desc *desc;
309 desc = xsk_ring_cons__rx_desc(rx, idx_rx++);
313 offset = xsk_umem__extract_offset(addr);
314 addr = xsk_umem__extract_addr(addr);
316 bufs[i] = (struct rte_mbuf *)
317 xsk_umem__get_data(umem->buffer, addr +
318 umem->mb_pool->header_size);
319 bufs[i]->data_off = offset - sizeof(struct rte_mbuf) -
320 rte_pktmbuf_priv_size(umem->mb_pool) -
321 umem->mb_pool->header_size;
323 rte_pktmbuf_pkt_len(bufs[i]) = len;
324 rte_pktmbuf_data_len(bufs[i]) = len;
328 xsk_ring_cons__release(rx, nb_pkts);
329 (void)reserve_fill_queue(umem, nb_pkts, fq_bufs, fq);
332 rxq->stats.rx_pkts += nb_pkts;
333 rxq->stats.rx_bytes += rx_bytes;
339 af_xdp_rx_cp(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
341 struct pkt_rx_queue *rxq = queue;
342 struct xsk_ring_cons *rx = &rxq->rx;
343 struct xsk_umem_info *umem = rxq->umem;
344 struct xsk_ring_prod *fq = &rxq->fq;
346 unsigned long rx_bytes = 0;
348 uint32_t free_thresh = fq->size >> 1;
349 struct rte_mbuf *mbufs[ETH_AF_XDP_RX_BATCH_SIZE];
351 if (xsk_prod_nb_free(fq, free_thresh) >= free_thresh)
352 (void)reserve_fill_queue(umem, nb_pkts, NULL, fq);
354 nb_pkts = xsk_ring_cons__peek(rx, nb_pkts, &idx_rx);
356 #if defined(XDP_USE_NEED_WAKEUP)
357 if (xsk_ring_prod__needs_wakeup(fq))
358 (void)poll(rxq->fds, 1, 1000);
363 if (unlikely(rte_pktmbuf_alloc_bulk(rxq->mb_pool, mbufs, nb_pkts))) {
364 /* rollback cached_cons which is added by
365 * xsk_ring_cons__peek
367 rx->cached_cons -= nb_pkts;
371 for (i = 0; i < nb_pkts; i++) {
372 const struct xdp_desc *desc;
377 desc = xsk_ring_cons__rx_desc(rx, idx_rx++);
380 pkt = xsk_umem__get_data(rxq->umem->mz->addr, addr);
382 rte_memcpy(rte_pktmbuf_mtod(mbufs[i], void *), pkt, len);
383 rte_ring_enqueue(umem->buf_ring, (void *)addr);
384 rte_pktmbuf_pkt_len(mbufs[i]) = len;
385 rte_pktmbuf_data_len(mbufs[i]) = len;
390 xsk_ring_cons__release(rx, nb_pkts);
393 rxq->stats.rx_pkts += nb_pkts;
394 rxq->stats.rx_bytes += rx_bytes;
401 af_xdp_rx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
403 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
404 return af_xdp_rx_zc(queue, bufs, nb_pkts);
406 return af_xdp_rx_cp(queue, bufs, nb_pkts);
411 eth_af_xdp_rx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
415 if (likely(nb_pkts <= ETH_AF_XDP_RX_BATCH_SIZE))
416 return af_xdp_rx(queue, bufs, nb_pkts);
418 /* Split larger batch into smaller batches of size
419 * ETH_AF_XDP_RX_BATCH_SIZE or less.
425 n = (uint16_t)RTE_MIN(nb_pkts, ETH_AF_XDP_RX_BATCH_SIZE);
426 ret = af_xdp_rx(queue, &bufs[nb_rx], n);
427 nb_rx = (uint16_t)(nb_rx + ret);
428 nb_pkts = (uint16_t)(nb_pkts - ret);
437 pull_umem_cq(struct xsk_umem_info *umem, int size, struct xsk_ring_cons *cq)
442 n = xsk_ring_cons__peek(cq, size, &idx_cq);
444 for (i = 0; i < n; i++) {
446 addr = *xsk_ring_cons__comp_addr(cq, idx_cq++);
447 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
448 addr = xsk_umem__extract_addr(addr);
449 rte_pktmbuf_free((struct rte_mbuf *)
450 xsk_umem__get_data(umem->buffer,
451 addr + umem->mb_pool->header_size));
453 rte_ring_enqueue(umem->buf_ring, (void *)addr);
457 xsk_ring_cons__release(cq, n);
461 kick_tx(struct pkt_tx_queue *txq, struct xsk_ring_cons *cq)
463 struct xsk_umem_info *umem = txq->umem;
465 pull_umem_cq(umem, XSK_RING_CONS__DEFAULT_NUM_DESCS, cq);
467 if (tx_syscall_needed(&txq->tx))
468 while (send(xsk_socket__fd(txq->pair->xsk), NULL,
469 0, MSG_DONTWAIT) < 0) {
470 /* some thing unexpected */
471 if (errno != EBUSY && errno != EAGAIN && errno != EINTR)
474 /* pull from completion queue to leave more space */
477 XSK_RING_CONS__DEFAULT_NUM_DESCS,
482 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
484 af_xdp_tx_zc(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
486 struct pkt_tx_queue *txq = queue;
487 struct xsk_umem_info *umem = txq->umem;
488 struct rte_mbuf *mbuf;
489 unsigned long tx_bytes = 0;
493 struct xdp_desc *desc;
494 uint64_t addr, offset;
495 struct xsk_ring_cons *cq = &txq->pair->cq;
496 uint32_t free_thresh = cq->size >> 1;
498 if (xsk_cons_nb_avail(cq, free_thresh) >= free_thresh)
499 pull_umem_cq(umem, XSK_RING_CONS__DEFAULT_NUM_DESCS, cq);
501 for (i = 0; i < nb_pkts; i++) {
504 if (mbuf->pool == umem->mb_pool) {
505 if (!xsk_ring_prod__reserve(&txq->tx, 1, &idx_tx)) {
507 if (!xsk_ring_prod__reserve(&txq->tx, 1,
511 desc = xsk_ring_prod__tx_desc(&txq->tx, idx_tx);
512 desc->len = mbuf->pkt_len;
513 addr = (uint64_t)mbuf - (uint64_t)umem->buffer -
514 umem->mb_pool->header_size;
515 offset = rte_pktmbuf_mtod(mbuf, uint64_t) -
517 umem->mb_pool->header_size;
518 offset = offset << XSK_UNALIGNED_BUF_OFFSET_SHIFT;
519 desc->addr = addr | offset;
522 struct rte_mbuf *local_mbuf =
523 rte_pktmbuf_alloc(umem->mb_pool);
526 if (local_mbuf == NULL)
529 if (!xsk_ring_prod__reserve(&txq->tx, 1, &idx_tx)) {
530 rte_pktmbuf_free(local_mbuf);
534 desc = xsk_ring_prod__tx_desc(&txq->tx, idx_tx);
535 desc->len = mbuf->pkt_len;
537 addr = (uint64_t)local_mbuf - (uint64_t)umem->buffer -
538 umem->mb_pool->header_size;
539 offset = rte_pktmbuf_mtod(local_mbuf, uint64_t) -
540 (uint64_t)local_mbuf +
541 umem->mb_pool->header_size;
542 pkt = xsk_umem__get_data(umem->buffer, addr + offset);
543 offset = offset << XSK_UNALIGNED_BUF_OFFSET_SHIFT;
544 desc->addr = addr | offset;
545 rte_memcpy(pkt, rte_pktmbuf_mtod(mbuf, void *),
547 rte_pktmbuf_free(mbuf);
551 tx_bytes += mbuf->pkt_len;
555 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 = RTE_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 = RTE_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);
791 #define CLB_VAL_IDX 0
793 eth_monitor_callback(const uint64_t value,
794 const uint64_t opaque[RTE_POWER_MONITOR_OPAQUE_SZ])
796 const uint64_t v = opaque[CLB_VAL_IDX];
797 const uint64_t m = (uint32_t)~0;
799 /* if the value has changed, abort entering power optimized state */
800 return (value & m) == v ? 0 : -1;
804 eth_get_monitor_addr(void *rx_queue, struct rte_power_monitor_cond *pmc)
806 struct pkt_rx_queue *rxq = rx_queue;
807 unsigned int *prod = rxq->rx.producer;
808 const uint32_t cur_val = rxq->rx.cached_prod; /* use cached value */
810 /* watch for changes in producer ring */
811 pmc->addr = (void *)prod;
813 /* store current value */
814 pmc->opaque[CLB_VAL_IDX] = cur_val;
815 pmc->fn = eth_monitor_callback;
817 /* AF_XDP producer ring index is 32-bit */
818 pmc->size = sizeof(uint32_t);
824 eth_dev_info(struct rte_eth_dev *dev, struct rte_eth_dev_info *dev_info)
826 struct pmd_internals *internals = dev->data->dev_private;
828 dev_info->if_index = internals->if_index;
829 dev_info->max_mac_addrs = 1;
830 dev_info->max_rx_queues = internals->queue_cnt;
831 dev_info->max_tx_queues = internals->queue_cnt;
833 dev_info->min_mtu = RTE_ETHER_MIN_MTU;
834 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
835 dev_info->max_rx_pktlen = getpagesize() -
836 sizeof(struct rte_mempool_objhdr) -
837 sizeof(struct rte_mbuf) -
838 RTE_PKTMBUF_HEADROOM - XDP_PACKET_HEADROOM;
840 dev_info->max_rx_pktlen = ETH_AF_XDP_FRAME_SIZE - XDP_PACKET_HEADROOM;
842 dev_info->max_mtu = dev_info->max_rx_pktlen - ETH_AF_XDP_ETH_OVERHEAD;
844 dev_info->default_rxportconf.burst_size = ETH_AF_XDP_DFLT_BUSY_BUDGET;
845 dev_info->default_txportconf.burst_size = ETH_AF_XDP_DFLT_BUSY_BUDGET;
846 dev_info->default_rxportconf.nb_queues = 1;
847 dev_info->default_txportconf.nb_queues = 1;
848 dev_info->default_rxportconf.ring_size = ETH_AF_XDP_DFLT_NUM_DESCS;
849 dev_info->default_txportconf.ring_size = ETH_AF_XDP_DFLT_NUM_DESCS;
855 eth_stats_get(struct rte_eth_dev *dev, struct rte_eth_stats *stats)
857 struct pmd_internals *internals = dev->data->dev_private;
858 struct xdp_statistics xdp_stats;
859 struct pkt_rx_queue *rxq;
860 struct pkt_tx_queue *txq;
864 for (i = 0; i < dev->data->nb_rx_queues; i++) {
865 optlen = sizeof(struct xdp_statistics);
866 rxq = &internals->rx_queues[i];
868 stats->q_ipackets[i] = rxq->stats.rx_pkts;
869 stats->q_ibytes[i] = rxq->stats.rx_bytes;
871 stats->q_opackets[i] = txq->stats.tx_pkts;
872 stats->q_obytes[i] = txq->stats.tx_bytes;
874 stats->ipackets += stats->q_ipackets[i];
875 stats->ibytes += stats->q_ibytes[i];
876 stats->imissed += rxq->stats.rx_dropped;
877 stats->oerrors += txq->stats.tx_dropped;
878 ret = getsockopt(xsk_socket__fd(rxq->xsk), SOL_XDP,
879 XDP_STATISTICS, &xdp_stats, &optlen);
881 AF_XDP_LOG(ERR, "getsockopt() failed for XDP_STATISTICS.\n");
884 stats->imissed += xdp_stats.rx_dropped;
886 stats->opackets += stats->q_opackets[i];
887 stats->obytes += stats->q_obytes[i];
894 eth_stats_reset(struct rte_eth_dev *dev)
896 struct pmd_internals *internals = dev->data->dev_private;
899 for (i = 0; i < internals->queue_cnt; i++) {
900 memset(&internals->rx_queues[i].stats, 0,
901 sizeof(struct rx_stats));
902 memset(&internals->tx_queues[i].stats, 0,
903 sizeof(struct tx_stats));
910 remove_xdp_program(struct pmd_internals *internals)
912 uint32_t curr_prog_id = 0;
914 if (bpf_get_link_xdp_id(internals->if_index, &curr_prog_id,
915 XDP_FLAGS_UPDATE_IF_NOEXIST)) {
916 AF_XDP_LOG(ERR, "bpf_get_link_xdp_id failed\n");
919 bpf_set_link_xdp_fd(internals->if_index, -1,
920 XDP_FLAGS_UPDATE_IF_NOEXIST);
924 xdp_umem_destroy(struct xsk_umem_info *umem)
926 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
927 umem->mb_pool = NULL;
929 rte_memzone_free(umem->mz);
932 rte_ring_free(umem->buf_ring);
933 umem->buf_ring = NULL;
940 eth_dev_close(struct rte_eth_dev *dev)
942 struct pmd_internals *internals = dev->data->dev_private;
943 struct pkt_rx_queue *rxq;
946 if (rte_eal_process_type() != RTE_PROC_PRIMARY)
949 AF_XDP_LOG(INFO, "Closing AF_XDP ethdev on numa socket %u\n",
952 for (i = 0; i < internals->queue_cnt; i++) {
953 rxq = &internals->rx_queues[i];
954 if (rxq->umem == NULL)
956 xsk_socket__delete(rxq->xsk);
958 if (__atomic_sub_fetch(&rxq->umem->refcnt, 1, __ATOMIC_ACQUIRE)
960 (void)xsk_umem__delete(rxq->umem->umem);
961 xdp_umem_destroy(rxq->umem);
964 /* free pkt_tx_queue */
970 * MAC is not allocated dynamically, setting it to NULL would prevent
971 * from releasing it in rte_eth_dev_release_port.
973 dev->data->mac_addrs = NULL;
975 remove_xdp_program(internals);
977 if (internals->shared_umem) {
978 struct internal_list *list;
980 /* Remove ethdev from list used to track and share UMEMs */
981 list = find_internal_resource(internals);
983 pthread_mutex_lock(&internal_list_lock);
984 TAILQ_REMOVE(&internal_list, list, next);
985 pthread_mutex_unlock(&internal_list_lock);
994 eth_link_update(struct rte_eth_dev *dev __rte_unused,
995 int wait_to_complete __rte_unused)
1000 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
1001 static inline uintptr_t get_base_addr(struct rte_mempool *mp, uint64_t *align)
1003 struct rte_mempool_memhdr *memhdr;
1004 uintptr_t memhdr_addr, aligned_addr;
1006 memhdr = STAILQ_FIRST(&mp->mem_list);
1007 memhdr_addr = (uintptr_t)memhdr->addr;
1008 aligned_addr = memhdr_addr & ~(getpagesize() - 1);
1009 *align = memhdr_addr - aligned_addr;
1011 return aligned_addr;
1015 xsk_umem_info *xdp_umem_configure(struct pmd_internals *internals,
1016 struct pkt_rx_queue *rxq)
1018 struct xsk_umem_info *umem = NULL;
1020 struct xsk_umem_config usr_config = {
1021 .fill_size = ETH_AF_XDP_DFLT_NUM_DESCS * 2,
1022 .comp_size = ETH_AF_XDP_DFLT_NUM_DESCS,
1023 .flags = XDP_UMEM_UNALIGNED_CHUNK_FLAG};
1024 void *base_addr = NULL;
1025 struct rte_mempool *mb_pool = rxq->mb_pool;
1026 uint64_t umem_size, align = 0;
1028 if (internals->shared_umem) {
1029 if (get_shared_umem(rxq, internals->if_name, &umem) < 0)
1033 __atomic_load_n(&umem->refcnt, __ATOMIC_ACQUIRE) <
1035 AF_XDP_LOG(INFO, "%s,qid%i sharing UMEM\n",
1036 internals->if_name, rxq->xsk_queue_idx);
1037 __atomic_fetch_add(&umem->refcnt, 1, __ATOMIC_ACQUIRE);
1042 usr_config.frame_size =
1043 rte_mempool_calc_obj_size(mb_pool->elt_size,
1044 mb_pool->flags, NULL);
1045 usr_config.frame_headroom = mb_pool->header_size +
1046 sizeof(struct rte_mbuf) +
1047 rte_pktmbuf_priv_size(mb_pool) +
1048 RTE_PKTMBUF_HEADROOM;
1050 umem = rte_zmalloc_socket("umem", sizeof(*umem), 0,
1053 AF_XDP_LOG(ERR, "Failed to allocate umem info");
1057 umem->mb_pool = mb_pool;
1058 base_addr = (void *)get_base_addr(mb_pool, &align);
1059 umem_size = (uint64_t)mb_pool->populated_size *
1060 (uint64_t)usr_config.frame_size +
1063 ret = xsk_umem__create(&umem->umem, base_addr, umem_size,
1064 &rxq->fq, &rxq->cq, &usr_config);
1066 AF_XDP_LOG(ERR, "Failed to create umem");
1069 umem->buffer = base_addr;
1071 if (internals->shared_umem) {
1072 umem->max_xsks = mb_pool->populated_size /
1073 ETH_AF_XDP_NUM_BUFFERS;
1074 AF_XDP_LOG(INFO, "Max xsks for UMEM %s: %u\n",
1075 mb_pool->name, umem->max_xsks);
1078 __atomic_store_n(&umem->refcnt, 1, __ATOMIC_RELEASE);
1083 xsk_umem_info *xdp_umem_configure(struct pmd_internals *internals,
1084 struct pkt_rx_queue *rxq)
1086 struct xsk_umem_info *umem;
1087 const struct rte_memzone *mz;
1088 struct xsk_umem_config usr_config = {
1089 .fill_size = ETH_AF_XDP_DFLT_NUM_DESCS,
1090 .comp_size = ETH_AF_XDP_DFLT_NUM_DESCS,
1091 .frame_size = ETH_AF_XDP_FRAME_SIZE,
1092 .frame_headroom = 0 };
1093 char ring_name[RTE_RING_NAMESIZE];
1094 char mz_name[RTE_MEMZONE_NAMESIZE];
1098 umem = rte_zmalloc_socket("umem", sizeof(*umem), 0, rte_socket_id());
1100 AF_XDP_LOG(ERR, "Failed to allocate umem info");
1104 snprintf(ring_name, sizeof(ring_name), "af_xdp_ring_%s_%u",
1105 internals->if_name, rxq->xsk_queue_idx);
1106 umem->buf_ring = rte_ring_create(ring_name,
1107 ETH_AF_XDP_NUM_BUFFERS,
1110 if (umem->buf_ring == NULL) {
1111 AF_XDP_LOG(ERR, "Failed to create rte_ring\n");
1115 for (i = 0; i < ETH_AF_XDP_NUM_BUFFERS; i++)
1116 rte_ring_enqueue(umem->buf_ring,
1117 (void *)(i * ETH_AF_XDP_FRAME_SIZE));
1119 snprintf(mz_name, sizeof(mz_name), "af_xdp_umem_%s_%u",
1120 internals->if_name, rxq->xsk_queue_idx);
1121 mz = rte_memzone_reserve_aligned(mz_name,
1122 ETH_AF_XDP_NUM_BUFFERS * ETH_AF_XDP_FRAME_SIZE,
1123 rte_socket_id(), RTE_MEMZONE_IOVA_CONTIG,
1126 AF_XDP_LOG(ERR, "Failed to reserve memzone for af_xdp umem.\n");
1130 ret = xsk_umem__create(&umem->umem, mz->addr,
1131 ETH_AF_XDP_NUM_BUFFERS * ETH_AF_XDP_FRAME_SIZE,
1136 AF_XDP_LOG(ERR, "Failed to create umem");
1145 xdp_umem_destroy(umem);
1150 load_custom_xdp_prog(const char *prog_path, int if_index)
1152 int ret, prog_fd = -1;
1153 struct bpf_object *obj;
1154 struct bpf_map *map;
1156 ret = bpf_prog_load(prog_path, BPF_PROG_TYPE_XDP, &obj, &prog_fd);
1158 AF_XDP_LOG(ERR, "Failed to load program %s\n", prog_path);
1163 * The loaded program must provision for a map of xsks, such that some
1164 * traffic can be redirected to userspace. When the xsk is created,
1165 * libbpf inserts it into the map.
1167 map = bpf_object__find_map_by_name(obj, "xsks_map");
1169 AF_XDP_LOG(ERR, "Failed to find xsks_map in %s\n", prog_path);
1173 /* Link the program with the given network device */
1174 ret = bpf_set_link_xdp_fd(if_index, prog_fd,
1175 XDP_FLAGS_UPDATE_IF_NOEXIST);
1177 AF_XDP_LOG(ERR, "Failed to set prog fd %d on interface\n",
1182 AF_XDP_LOG(INFO, "Successfully loaded XDP program %s with fd %d\n",
1183 prog_path, prog_fd);
1188 /* Detect support for busy polling through setsockopt(). */
1190 configure_preferred_busy_poll(struct pkt_rx_queue *rxq)
1193 int fd = xsk_socket__fd(rxq->xsk);
1196 ret = setsockopt(fd, SOL_SOCKET, SO_PREFER_BUSY_POLL,
1197 (void *)&sock_opt, sizeof(sock_opt));
1199 AF_XDP_LOG(DEBUG, "Failed to set SO_PREFER_BUSY_POLL\n");
1203 sock_opt = ETH_AF_XDP_DFLT_BUSY_TIMEOUT;
1204 ret = setsockopt(fd, SOL_SOCKET, SO_BUSY_POLL, (void *)&sock_opt,
1207 AF_XDP_LOG(DEBUG, "Failed to set SO_BUSY_POLL\n");
1211 sock_opt = rxq->busy_budget;
1212 ret = setsockopt(fd, SOL_SOCKET, SO_BUSY_POLL_BUDGET,
1213 (void *)&sock_opt, sizeof(sock_opt));
1215 AF_XDP_LOG(DEBUG, "Failed to set SO_BUSY_POLL_BUDGET\n");
1217 AF_XDP_LOG(INFO, "Busy polling budget set to: %u\n",
1222 /* setsockopt failure - attempt to restore xsk to default state and
1223 * proceed without busy polling support.
1226 ret = setsockopt(fd, SOL_SOCKET, SO_BUSY_POLL, (void *)&sock_opt,
1229 AF_XDP_LOG(ERR, "Failed to unset SO_BUSY_POLL\n");
1235 ret = setsockopt(fd, SOL_SOCKET, SO_PREFER_BUSY_POLL,
1236 (void *)&sock_opt, sizeof(sock_opt));
1238 AF_XDP_LOG(ERR, "Failed to unset SO_PREFER_BUSY_POLL\n");
1243 rxq->busy_budget = 0;
1248 xsk_configure(struct pmd_internals *internals, struct pkt_rx_queue *rxq,
1251 struct xsk_socket_config cfg;
1252 struct pkt_tx_queue *txq = rxq->pair;
1254 int reserve_size = ETH_AF_XDP_DFLT_NUM_DESCS;
1255 struct rte_mbuf *fq_bufs[reserve_size];
1257 rxq->umem = xdp_umem_configure(internals, rxq);
1258 if (rxq->umem == NULL)
1260 txq->umem = rxq->umem;
1262 cfg.rx_size = ring_size;
1263 cfg.tx_size = ring_size;
1264 cfg.libbpf_flags = 0;
1265 cfg.xdp_flags = XDP_FLAGS_UPDATE_IF_NOEXIST;
1268 #if defined(XDP_USE_NEED_WAKEUP)
1269 cfg.bind_flags |= XDP_USE_NEED_WAKEUP;
1272 if (strnlen(internals->prog_path, PATH_MAX) &&
1273 !internals->custom_prog_configured) {
1274 ret = load_custom_xdp_prog(internals->prog_path,
1275 internals->if_index);
1277 AF_XDP_LOG(ERR, "Failed to load custom XDP program %s\n",
1278 internals->prog_path);
1281 internals->custom_prog_configured = 1;
1284 if (internals->shared_umem)
1285 ret = create_shared_socket(&rxq->xsk, internals->if_name,
1286 rxq->xsk_queue_idx, rxq->umem->umem, &rxq->rx,
1287 &txq->tx, &rxq->fq, &rxq->cq, &cfg);
1289 ret = xsk_socket__create(&rxq->xsk, internals->if_name,
1290 rxq->xsk_queue_idx, rxq->umem->umem, &rxq->rx,
1294 AF_XDP_LOG(ERR, "Failed to create xsk socket.\n");
1298 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
1299 ret = rte_pktmbuf_alloc_bulk(rxq->umem->mb_pool, fq_bufs, reserve_size);
1301 AF_XDP_LOG(DEBUG, "Failed to get enough buffers for fq.\n");
1306 if (rxq->busy_budget) {
1307 ret = configure_preferred_busy_poll(rxq);
1309 AF_XDP_LOG(ERR, "Failed configure busy polling.\n");
1314 ret = reserve_fill_queue(rxq->umem, reserve_size, fq_bufs, &rxq->fq);
1316 xsk_socket__delete(rxq->xsk);
1317 AF_XDP_LOG(ERR, "Failed to reserve fill queue.\n");
1324 if (__atomic_sub_fetch(&rxq->umem->refcnt, 1, __ATOMIC_ACQUIRE) == 0)
1325 xdp_umem_destroy(rxq->umem);
1331 eth_rx_queue_setup(struct rte_eth_dev *dev,
1332 uint16_t rx_queue_id,
1333 uint16_t nb_rx_desc,
1334 unsigned int socket_id __rte_unused,
1335 const struct rte_eth_rxconf *rx_conf __rte_unused,
1336 struct rte_mempool *mb_pool)
1338 struct pmd_internals *internals = dev->data->dev_private;
1339 struct pkt_rx_queue *rxq;
1342 rxq = &internals->rx_queues[rx_queue_id];
1344 AF_XDP_LOG(INFO, "Set up rx queue, rx queue id: %d, xsk queue id: %d\n",
1345 rx_queue_id, rxq->xsk_queue_idx);
1347 #ifndef XDP_UMEM_UNALIGNED_CHUNK_FLAG
1348 uint32_t buf_size, data_size;
1350 /* Now get the space available for data in the mbuf */
1351 buf_size = rte_pktmbuf_data_room_size(mb_pool) -
1352 RTE_PKTMBUF_HEADROOM;
1353 data_size = ETH_AF_XDP_FRAME_SIZE;
1355 if (data_size > buf_size) {
1356 AF_XDP_LOG(ERR, "%s: %d bytes will not fit in mbuf (%d bytes)\n",
1357 dev->device->name, data_size, buf_size);
1363 rxq->mb_pool = mb_pool;
1365 if (xsk_configure(internals, rxq, nb_rx_desc)) {
1366 AF_XDP_LOG(ERR, "Failed to configure xdp socket\n");
1371 if (!rxq->busy_budget)
1372 AF_XDP_LOG(DEBUG, "Preferred busy polling not enabled\n");
1374 rxq->fds[0].fd = xsk_socket__fd(rxq->xsk);
1375 rxq->fds[0].events = POLLIN;
1377 dev->data->rx_queues[rx_queue_id] = rxq;
1385 eth_tx_queue_setup(struct rte_eth_dev *dev,
1386 uint16_t tx_queue_id,
1387 uint16_t nb_tx_desc __rte_unused,
1388 unsigned int socket_id __rte_unused,
1389 const struct rte_eth_txconf *tx_conf __rte_unused)
1391 struct pmd_internals *internals = dev->data->dev_private;
1392 struct pkt_tx_queue *txq;
1394 txq = &internals->tx_queues[tx_queue_id];
1396 dev->data->tx_queues[tx_queue_id] = txq;
1401 eth_dev_mtu_set(struct rte_eth_dev *dev, uint16_t mtu)
1403 struct pmd_internals *internals = dev->data->dev_private;
1404 struct ifreq ifr = { .ifr_mtu = mtu };
1408 s = socket(PF_INET, SOCK_DGRAM, 0);
1412 strlcpy(ifr.ifr_name, internals->if_name, IFNAMSIZ);
1413 ret = ioctl(s, SIOCSIFMTU, &ifr);
1416 return (ret < 0) ? -errno : 0;
1420 eth_dev_change_flags(char *if_name, uint32_t flags, uint32_t mask)
1426 s = socket(PF_INET, SOCK_DGRAM, 0);
1430 strlcpy(ifr.ifr_name, if_name, IFNAMSIZ);
1431 if (ioctl(s, SIOCGIFFLAGS, &ifr) < 0) {
1435 ifr.ifr_flags &= mask;
1436 ifr.ifr_flags |= flags;
1437 if (ioctl(s, SIOCSIFFLAGS, &ifr) < 0) {
1447 eth_dev_promiscuous_enable(struct rte_eth_dev *dev)
1449 struct pmd_internals *internals = dev->data->dev_private;
1451 return eth_dev_change_flags(internals->if_name, IFF_PROMISC, ~0);
1455 eth_dev_promiscuous_disable(struct rte_eth_dev *dev)
1457 struct pmd_internals *internals = dev->data->dev_private;
1459 return eth_dev_change_flags(internals->if_name, 0, ~IFF_PROMISC);
1462 static const struct eth_dev_ops ops = {
1463 .dev_start = eth_dev_start,
1464 .dev_stop = eth_dev_stop,
1465 .dev_close = eth_dev_close,
1466 .dev_configure = eth_dev_configure,
1467 .dev_infos_get = eth_dev_info,
1468 .mtu_set = eth_dev_mtu_set,
1469 .promiscuous_enable = eth_dev_promiscuous_enable,
1470 .promiscuous_disable = eth_dev_promiscuous_disable,
1471 .rx_queue_setup = eth_rx_queue_setup,
1472 .tx_queue_setup = eth_tx_queue_setup,
1473 .link_update = eth_link_update,
1474 .stats_get = eth_stats_get,
1475 .stats_reset = eth_stats_reset,
1476 .get_monitor_addr = eth_get_monitor_addr,
1479 /** parse busy_budget argument */
1481 parse_budget_arg(const char *key __rte_unused,
1482 const char *value, void *extra_args)
1484 int *i = (int *)extra_args;
1487 *i = strtol(value, &end, 10);
1488 if (*i < 0 || *i > UINT16_MAX) {
1489 AF_XDP_LOG(ERR, "Invalid busy_budget %i, must be >= 0 and <= %u\n",
1497 /** parse integer from integer argument */
1499 parse_integer_arg(const char *key __rte_unused,
1500 const char *value, void *extra_args)
1502 int *i = (int *)extra_args;
1505 *i = strtol(value, &end, 10);
1507 AF_XDP_LOG(ERR, "Argument has to be positive.\n");
1514 /** parse name argument */
1516 parse_name_arg(const char *key __rte_unused,
1517 const char *value, void *extra_args)
1519 char *name = extra_args;
1521 if (strnlen(value, IFNAMSIZ) > IFNAMSIZ - 1) {
1522 AF_XDP_LOG(ERR, "Invalid name %s, should be less than %u bytes.\n",
1527 strlcpy(name, value, IFNAMSIZ);
1532 /** parse xdp prog argument */
1534 parse_prog_arg(const char *key __rte_unused,
1535 const char *value, void *extra_args)
1537 char *path = extra_args;
1539 if (strnlen(value, PATH_MAX) == PATH_MAX) {
1540 AF_XDP_LOG(ERR, "Invalid path %s, should be less than %u bytes.\n",
1545 if (access(value, F_OK) != 0) {
1546 AF_XDP_LOG(ERR, "Error accessing %s: %s\n",
1547 value, strerror(errno));
1551 strlcpy(path, value, PATH_MAX);
1557 xdp_get_channels_info(const char *if_name, int *max_queues,
1558 int *combined_queues)
1560 struct ethtool_channels channels;
1564 fd = socket(AF_INET, SOCK_DGRAM, 0);
1568 channels.cmd = ETHTOOL_GCHANNELS;
1569 ifr.ifr_data = (void *)&channels;
1570 strlcpy(ifr.ifr_name, if_name, IFNAMSIZ);
1571 ret = ioctl(fd, SIOCETHTOOL, &ifr);
1573 if (errno == EOPNOTSUPP) {
1581 if (channels.max_combined == 0 || errno == EOPNOTSUPP) {
1582 /* If the device says it has no channels, then all traffic
1583 * is sent to a single stream, so max queues = 1.
1586 *combined_queues = 1;
1588 *max_queues = channels.max_combined;
1589 *combined_queues = channels.combined_count;
1598 parse_parameters(struct rte_kvargs *kvlist, char *if_name, int *start_queue,
1599 int *queue_cnt, int *shared_umem, char *prog_path,
1604 ret = rte_kvargs_process(kvlist, ETH_AF_XDP_IFACE_ARG,
1605 &parse_name_arg, if_name);
1609 ret = rte_kvargs_process(kvlist, ETH_AF_XDP_START_QUEUE_ARG,
1610 &parse_integer_arg, start_queue);
1614 ret = rte_kvargs_process(kvlist, ETH_AF_XDP_QUEUE_COUNT_ARG,
1615 &parse_integer_arg, queue_cnt);
1616 if (ret < 0 || *queue_cnt <= 0) {
1621 ret = rte_kvargs_process(kvlist, ETH_AF_XDP_SHARED_UMEM_ARG,
1622 &parse_integer_arg, shared_umem);
1626 ret = rte_kvargs_process(kvlist, ETH_AF_XDP_PROG_ARG,
1627 &parse_prog_arg, prog_path);
1631 ret = rte_kvargs_process(kvlist, ETH_AF_XDP_BUDGET_ARG,
1632 &parse_budget_arg, busy_budget);
1637 rte_kvargs_free(kvlist);
1642 get_iface_info(const char *if_name,
1643 struct rte_ether_addr *eth_addr,
1647 int sock = socket(AF_INET, SOCK_DGRAM, IPPROTO_IP);
1652 strlcpy(ifr.ifr_name, if_name, IFNAMSIZ);
1653 if (ioctl(sock, SIOCGIFINDEX, &ifr))
1656 *if_index = ifr.ifr_ifindex;
1658 if (ioctl(sock, SIOCGIFHWADDR, &ifr))
1661 rte_memcpy(eth_addr, ifr.ifr_hwaddr.sa_data, RTE_ETHER_ADDR_LEN);
1671 static struct rte_eth_dev *
1672 init_internals(struct rte_vdev_device *dev, const char *if_name,
1673 int start_queue_idx, int queue_cnt, int shared_umem,
1674 const char *prog_path, int busy_budget)
1676 const char *name = rte_vdev_device_name(dev);
1677 const unsigned int numa_node = dev->device.numa_node;
1678 struct pmd_internals *internals;
1679 struct rte_eth_dev *eth_dev;
1683 internals = rte_zmalloc_socket(name, sizeof(*internals), 0, numa_node);
1684 if (internals == NULL)
1687 internals->start_queue_idx = start_queue_idx;
1688 internals->queue_cnt = queue_cnt;
1689 strlcpy(internals->if_name, if_name, IFNAMSIZ);
1690 strlcpy(internals->prog_path, prog_path, PATH_MAX);
1691 internals->custom_prog_configured = 0;
1693 #ifndef ETH_AF_XDP_SHARED_UMEM
1695 AF_XDP_LOG(ERR, "Shared UMEM feature not available. "
1696 "Check kernel and libbpf version\n");
1697 goto err_free_internals;
1700 internals->shared_umem = shared_umem;
1702 if (xdp_get_channels_info(if_name, &internals->max_queue_cnt,
1703 &internals->combined_queue_cnt)) {
1704 AF_XDP_LOG(ERR, "Failed to get channel info of interface: %s\n",
1706 goto err_free_internals;
1709 if (queue_cnt > internals->combined_queue_cnt) {
1710 AF_XDP_LOG(ERR, "Specified queue count %d is larger than combined queue count %d.\n",
1711 queue_cnt, internals->combined_queue_cnt);
1712 goto err_free_internals;
1715 internals->rx_queues = rte_zmalloc_socket(NULL,
1716 sizeof(struct pkt_rx_queue) * queue_cnt,
1718 if (internals->rx_queues == NULL) {
1719 AF_XDP_LOG(ERR, "Failed to allocate memory for rx queues.\n");
1720 goto err_free_internals;
1723 internals->tx_queues = rte_zmalloc_socket(NULL,
1724 sizeof(struct pkt_tx_queue) * queue_cnt,
1726 if (internals->tx_queues == NULL) {
1727 AF_XDP_LOG(ERR, "Failed to allocate memory for tx queues.\n");
1730 for (i = 0; i < queue_cnt; i++) {
1731 internals->tx_queues[i].pair = &internals->rx_queues[i];
1732 internals->rx_queues[i].pair = &internals->tx_queues[i];
1733 internals->rx_queues[i].xsk_queue_idx = start_queue_idx + i;
1734 internals->tx_queues[i].xsk_queue_idx = start_queue_idx + i;
1735 internals->rx_queues[i].busy_budget = busy_budget;
1738 ret = get_iface_info(if_name, &internals->eth_addr,
1739 &internals->if_index);
1743 eth_dev = rte_eth_vdev_allocate(dev, 0);
1744 if (eth_dev == NULL)
1747 eth_dev->data->dev_private = internals;
1748 eth_dev->data->dev_link = pmd_link;
1749 eth_dev->data->mac_addrs = &internals->eth_addr;
1750 eth_dev->data->dev_flags |= RTE_ETH_DEV_AUTOFILL_QUEUE_XSTATS;
1751 eth_dev->dev_ops = &ops;
1752 eth_dev->rx_pkt_burst = eth_af_xdp_rx;
1753 eth_dev->tx_pkt_burst = eth_af_xdp_tx;
1755 #if defined(XDP_UMEM_UNALIGNED_CHUNK_FLAG)
1756 AF_XDP_LOG(INFO, "Zero copy between umem and mbuf enabled.\n");
1762 rte_free(internals->tx_queues);
1764 rte_free(internals->rx_queues);
1766 rte_free(internals);
1771 rte_pmd_af_xdp_probe(struct rte_vdev_device *dev)
1773 struct rte_kvargs *kvlist;
1774 char if_name[IFNAMSIZ] = {'\0'};
1775 int xsk_start_queue_idx = ETH_AF_XDP_DFLT_START_QUEUE_IDX;
1776 int xsk_queue_cnt = ETH_AF_XDP_DFLT_QUEUE_COUNT;
1777 int shared_umem = 0;
1778 char prog_path[PATH_MAX] = {'\0'};
1779 int busy_budget = -1;
1780 struct rte_eth_dev *eth_dev = NULL;
1783 AF_XDP_LOG(INFO, "Initializing pmd_af_xdp for %s\n",
1784 rte_vdev_device_name(dev));
1786 name = rte_vdev_device_name(dev);
1787 if (rte_eal_process_type() == RTE_PROC_SECONDARY) {
1788 AF_XDP_LOG(ERR, "Failed to probe %s. "
1789 "AF_XDP PMD does not support secondary processes.\n",
1794 kvlist = rte_kvargs_parse(rte_vdev_device_args(dev), valid_arguments);
1795 if (kvlist == NULL) {
1796 AF_XDP_LOG(ERR, "Invalid kvargs key\n");
1800 if (dev->device.numa_node == SOCKET_ID_ANY)
1801 dev->device.numa_node = rte_socket_id();
1803 if (parse_parameters(kvlist, if_name, &xsk_start_queue_idx,
1804 &xsk_queue_cnt, &shared_umem, prog_path,
1805 &busy_budget) < 0) {
1806 AF_XDP_LOG(ERR, "Invalid kvargs value\n");
1810 if (strlen(if_name) == 0) {
1811 AF_XDP_LOG(ERR, "Network interface must be specified\n");
1815 busy_budget = busy_budget == -1 ? ETH_AF_XDP_DFLT_BUSY_BUDGET :
1818 eth_dev = init_internals(dev, if_name, xsk_start_queue_idx,
1819 xsk_queue_cnt, shared_umem, prog_path,
1821 if (eth_dev == NULL) {
1822 AF_XDP_LOG(ERR, "Failed to init internals\n");
1826 rte_eth_dev_probing_finish(eth_dev);
1832 rte_pmd_af_xdp_remove(struct rte_vdev_device *dev)
1834 struct rte_eth_dev *eth_dev = NULL;
1836 AF_XDP_LOG(INFO, "Removing AF_XDP ethdev on numa socket %u\n",
1842 /* find the ethdev entry */
1843 eth_dev = rte_eth_dev_allocated(rte_vdev_device_name(dev));
1844 if (eth_dev == NULL)
1847 eth_dev_close(eth_dev);
1848 rte_eth_dev_release_port(eth_dev);
1854 static struct rte_vdev_driver pmd_af_xdp_drv = {
1855 .probe = rte_pmd_af_xdp_probe,
1856 .remove = rte_pmd_af_xdp_remove,
1859 RTE_PMD_REGISTER_VDEV(net_af_xdp, pmd_af_xdp_drv);
1860 RTE_PMD_REGISTER_PARAM_STRING(net_af_xdp,
1862 "start_queue=<int> "
1863 "queue_count=<int> "
1864 "shared_umem=<int> "
1865 "xdp_prog=<string> "
1866 "busy_budget=<int>");