1 // SPDX-License-Identifier: GPL-2.0
4 * AF_XDP sockets allows a channel between XDP programs and userspace
6 * Copyright(c) 2018 Intel Corporation.
8 * Author(s): Björn Töpel <bjorn.topel@intel.com>
9 * Magnus Karlsson <magnus.karlsson@intel.com>
12 #define pr_fmt(fmt) "AF_XDP: %s: " fmt, __func__
14 #include <linux/if_xdp.h>
15 #include <linux/init.h>
16 #include <linux/sched/mm.h>
17 #include <linux/sched/signal.h>
18 #include <linux/sched/task.h>
19 #include <linux/socket.h>
20 #include <linux/file.h>
21 #include <linux/uaccess.h>
22 #include <linux/net.h>
23 #include <linux/netdevice.h>
24 #include <linux/rculist.h>
25 #include <net/xdp_sock.h>
28 #include "xsk_queue.h"
32 #define TX_BATCH_SIZE 16
34 bool xsk_is_setup_for_bpf_map(struct xdp_sock *xs)
36 return READ_ONCE(xs->rx) && READ_ONCE(xs->umem) &&
37 READ_ONCE(xs->umem->fq);
40 bool xsk_umem_has_addrs(struct xdp_umem *umem, u32 cnt)
42 return xskq_has_addrs(umem->fq, cnt);
44 EXPORT_SYMBOL(xsk_umem_has_addrs);
46 u64 *xsk_umem_peek_addr(struct xdp_umem *umem, u64 *addr)
48 return xskq_peek_addr(umem->fq, addr, umem);
50 EXPORT_SYMBOL(xsk_umem_peek_addr);
52 void xsk_umem_discard_addr(struct xdp_umem *umem)
54 xskq_discard_addr(umem->fq);
56 EXPORT_SYMBOL(xsk_umem_discard_addr);
58 void xsk_set_rx_need_wakeup(struct xdp_umem *umem)
60 if (umem->need_wakeup & XDP_WAKEUP_RX)
63 umem->fq->ring->flags |= XDP_RING_NEED_WAKEUP;
64 umem->need_wakeup |= XDP_WAKEUP_RX;
66 EXPORT_SYMBOL(xsk_set_rx_need_wakeup);
68 void xsk_set_tx_need_wakeup(struct xdp_umem *umem)
72 if (umem->need_wakeup & XDP_WAKEUP_TX)
76 list_for_each_entry_rcu(xs, &umem->xsk_list, list) {
77 xs->tx->ring->flags |= XDP_RING_NEED_WAKEUP;
81 umem->need_wakeup |= XDP_WAKEUP_TX;
83 EXPORT_SYMBOL(xsk_set_tx_need_wakeup);
85 void xsk_clear_rx_need_wakeup(struct xdp_umem *umem)
87 if (!(umem->need_wakeup & XDP_WAKEUP_RX))
90 umem->fq->ring->flags &= ~XDP_RING_NEED_WAKEUP;
91 umem->need_wakeup &= ~XDP_WAKEUP_RX;
93 EXPORT_SYMBOL(xsk_clear_rx_need_wakeup);
95 void xsk_clear_tx_need_wakeup(struct xdp_umem *umem)
99 if (!(umem->need_wakeup & XDP_WAKEUP_TX))
103 list_for_each_entry_rcu(xs, &umem->xsk_list, list) {
104 xs->tx->ring->flags &= ~XDP_RING_NEED_WAKEUP;
108 umem->need_wakeup &= ~XDP_WAKEUP_TX;
110 EXPORT_SYMBOL(xsk_clear_tx_need_wakeup);
112 bool xsk_umem_uses_need_wakeup(struct xdp_umem *umem)
114 return umem->flags & XDP_UMEM_USES_NEED_WAKEUP;
116 EXPORT_SYMBOL(xsk_umem_uses_need_wakeup);
118 /* If a buffer crosses a page boundary, we need to do 2 memcpy's, one for
119 * each page. This is only required in copy mode.
121 static void __xsk_rcv_memcpy(struct xdp_umem *umem, u64 addr, void *from_buf,
122 u32 len, u32 metalen)
124 void *to_buf = xdp_umem_get_data(umem, addr);
126 addr = xsk_umem_add_offset_to_addr(addr);
127 if (xskq_crosses_non_contig_pg(umem, addr, len + metalen)) {
128 void *next_pg_addr = umem->pages[(addr >> PAGE_SHIFT) + 1].addr;
129 u64 page_start = addr & ~(PAGE_SIZE - 1);
130 u64 first_len = PAGE_SIZE - (addr - page_start);
132 memcpy(to_buf, from_buf, first_len + metalen);
133 memcpy(next_pg_addr, from_buf + first_len, len - first_len);
138 memcpy(to_buf, from_buf, len + metalen);
141 static int __xsk_rcv(struct xdp_sock *xs, struct xdp_buff *xdp, u32 len)
143 u64 offset = xs->umem->headroom;
144 u64 addr, memcpy_addr;
149 if (!xskq_peek_addr(xs->umem->fq, &addr, xs->umem) ||
150 len > xs->umem->chunk_size_nohr - XDP_PACKET_HEADROOM) {
155 if (unlikely(xdp_data_meta_unsupported(xdp))) {
156 from_buf = xdp->data;
159 from_buf = xdp->data_meta;
160 metalen = xdp->data - xdp->data_meta;
163 memcpy_addr = xsk_umem_adjust_offset(xs->umem, addr, offset);
164 __xsk_rcv_memcpy(xs->umem, memcpy_addr, from_buf, len, metalen);
167 addr = xsk_umem_adjust_offset(xs->umem, addr, offset);
168 err = xskq_produce_batch_desc(xs->rx, addr, len);
170 xskq_discard_addr(xs->umem->fq);
171 xdp_return_buff(xdp);
179 static int __xsk_rcv_zc(struct xdp_sock *xs, struct xdp_buff *xdp, u32 len)
181 int err = xskq_produce_batch_desc(xs->rx, (u64)xdp->handle, len);
189 static bool xsk_is_bound(struct xdp_sock *xs)
191 if (READ_ONCE(xs->state) == XSK_BOUND) {
192 /* Matches smp_wmb() in bind(). */
199 static int xsk_rcv(struct xdp_sock *xs, struct xdp_buff *xdp)
203 if (!xsk_is_bound(xs))
206 if (xs->dev != xdp->rxq->dev || xs->queue_id != xdp->rxq->queue_index)
209 len = xdp->data_end - xdp->data;
211 return (xdp->rxq->mem.type == MEM_TYPE_ZERO_COPY) ?
212 __xsk_rcv_zc(xs, xdp, len) : __xsk_rcv(xs, xdp, len);
215 static void xsk_flush(struct xdp_sock *xs)
217 xskq_produce_flush_desc(xs->rx);
218 xs->sk.sk_data_ready(&xs->sk);
221 int xsk_generic_rcv(struct xdp_sock *xs, struct xdp_buff *xdp)
223 u32 metalen = xdp->data - xdp->data_meta;
224 u32 len = xdp->data_end - xdp->data;
225 u64 offset = xs->umem->headroom;
230 spin_lock_bh(&xs->rx_lock);
232 if (xs->dev != xdp->rxq->dev || xs->queue_id != xdp->rxq->queue_index) {
237 if (!xskq_peek_addr(xs->umem->fq, &addr, xs->umem) ||
238 len > xs->umem->chunk_size_nohr - XDP_PACKET_HEADROOM) {
243 addr = xsk_umem_adjust_offset(xs->umem, addr, offset);
244 buffer = xdp_umem_get_data(xs->umem, addr);
245 memcpy(buffer, xdp->data_meta, len + metalen);
247 addr = xsk_umem_adjust_offset(xs->umem, addr, metalen);
248 err = xskq_produce_batch_desc(xs->rx, addr, len);
252 xskq_discard_addr(xs->umem->fq);
253 xskq_produce_flush_desc(xs->rx);
255 spin_unlock_bh(&xs->rx_lock);
257 xs->sk.sk_data_ready(&xs->sk);
263 spin_unlock_bh(&xs->rx_lock);
267 int __xsk_map_redirect(struct bpf_map *map, struct xdp_buff *xdp,
270 struct xsk_map *m = container_of(map, struct xsk_map, map);
271 struct list_head *flush_list = this_cpu_ptr(m->flush_list);
274 err = xsk_rcv(xs, xdp);
278 if (!xs->flush_node.prev)
279 list_add(&xs->flush_node, flush_list);
284 void __xsk_map_flush(struct bpf_map *map)
286 struct xsk_map *m = container_of(map, struct xsk_map, map);
287 struct list_head *flush_list = this_cpu_ptr(m->flush_list);
288 struct xdp_sock *xs, *tmp;
290 list_for_each_entry_safe(xs, tmp, flush_list, flush_node) {
292 __list_del_clearprev(&xs->flush_node);
296 void xsk_umem_complete_tx(struct xdp_umem *umem, u32 nb_entries)
298 xskq_produce_flush_addr_n(umem->cq, nb_entries);
300 EXPORT_SYMBOL(xsk_umem_complete_tx);
302 void xsk_umem_consume_tx_done(struct xdp_umem *umem)
307 list_for_each_entry_rcu(xs, &umem->xsk_list, list) {
308 xs->sk.sk_write_space(&xs->sk);
312 EXPORT_SYMBOL(xsk_umem_consume_tx_done);
314 bool xsk_umem_consume_tx(struct xdp_umem *umem, struct xdp_desc *desc)
319 list_for_each_entry_rcu(xs, &umem->xsk_list, list) {
320 if (!xskq_peek_desc(xs->tx, desc, umem))
323 if (xskq_produce_addr_lazy(umem->cq, desc->addr))
326 xskq_discard_desc(xs->tx);
335 EXPORT_SYMBOL(xsk_umem_consume_tx);
337 static int xsk_wakeup(struct xdp_sock *xs, u8 flags)
339 struct net_device *dev = xs->dev;
343 err = dev->netdev_ops->ndo_xsk_wakeup(dev, xs->queue_id, flags);
349 static int xsk_zc_xmit(struct xdp_sock *xs)
351 return xsk_wakeup(xs, XDP_WAKEUP_TX);
354 static void xsk_destruct_skb(struct sk_buff *skb)
356 u64 addr = (u64)(long)skb_shinfo(skb)->destructor_arg;
357 struct xdp_sock *xs = xdp_sk(skb->sk);
360 spin_lock_irqsave(&xs->tx_completion_lock, flags);
361 WARN_ON_ONCE(xskq_produce_addr(xs->umem->cq, addr));
362 spin_unlock_irqrestore(&xs->tx_completion_lock, flags);
367 static int xsk_generic_xmit(struct sock *sk)
369 struct xdp_sock *xs = xdp_sk(sk);
370 u32 max_batch = TX_BATCH_SIZE;
371 bool sent_frame = false;
372 struct xdp_desc desc;
376 mutex_lock(&xs->mutex);
378 if (xs->queue_id >= xs->dev->real_num_tx_queues)
381 while (xskq_peek_desc(xs->tx, &desc, xs->umem)) {
386 if (max_batch-- == 0) {
392 skb = sock_alloc_send_skb(sk, len, 1, &err);
393 if (unlikely(!skb)) {
400 buffer = xdp_umem_get_data(xs->umem, addr);
401 err = skb_store_bits(skb, 0, buffer, len);
402 if (unlikely(err) || xskq_reserve_addr(xs->umem->cq)) {
408 skb->priority = sk->sk_priority;
409 skb->mark = sk->sk_mark;
410 skb_shinfo(skb)->destructor_arg = (void *)(long)desc.addr;
411 skb->destructor = xsk_destruct_skb;
413 err = dev_direct_xmit(skb, xs->queue_id);
414 xskq_discard_desc(xs->tx);
415 /* Ignore NET_XMIT_CN as packet might have been sent */
416 if (err == NET_XMIT_DROP || err == NETDEV_TX_BUSY) {
417 /* SKB completed but not sent */
427 sk->sk_write_space(sk);
429 mutex_unlock(&xs->mutex);
433 static int __xsk_sendmsg(struct sock *sk)
435 struct xdp_sock *xs = xdp_sk(sk);
437 if (unlikely(!(xs->dev->flags & IFF_UP)))
439 if (unlikely(!xs->tx))
442 return xs->zc ? xsk_zc_xmit(xs) : xsk_generic_xmit(sk);
445 static int xsk_sendmsg(struct socket *sock, struct msghdr *m, size_t total_len)
447 bool need_wait = !(m->msg_flags & MSG_DONTWAIT);
448 struct sock *sk = sock->sk;
449 struct xdp_sock *xs = xdp_sk(sk);
451 if (unlikely(!xsk_is_bound(xs)))
453 if (unlikely(need_wait))
456 return __xsk_sendmsg(sk);
459 static __poll_t xsk_poll(struct file *file, struct socket *sock,
460 struct poll_table_struct *wait)
462 __poll_t mask = datagram_poll(file, sock, wait);
463 struct sock *sk = sock->sk;
464 struct xdp_sock *xs = xdp_sk(sk);
465 struct xdp_umem *umem;
467 if (unlikely(!xsk_is_bound(xs)))
472 if (umem->need_wakeup) {
474 xsk_wakeup(xs, umem->need_wakeup);
476 /* Poll needs to drive Tx also in copy mode */
480 if (xs->rx && !xskq_empty_desc(xs->rx))
481 mask |= EPOLLIN | EPOLLRDNORM;
482 if (xs->tx && !xskq_full_desc(xs->tx))
483 mask |= EPOLLOUT | EPOLLWRNORM;
488 static int xsk_init_queue(u32 entries, struct xsk_queue **queue,
493 if (entries == 0 || *queue || !is_power_of_2(entries))
496 q = xskq_create(entries, umem_queue);
500 /* Make sure queue is ready before it can be seen by others */
502 WRITE_ONCE(*queue, q);
506 static void xsk_unbind_dev(struct xdp_sock *xs)
508 struct net_device *dev = xs->dev;
510 if (xs->state != XSK_BOUND)
512 WRITE_ONCE(xs->state, XSK_UNBOUND);
514 /* Wait for driver to stop using the xdp socket. */
515 xdp_del_sk_umem(xs->umem, xs);
521 static struct xsk_map *xsk_get_map_list_entry(struct xdp_sock *xs,
522 struct xdp_sock ***map_entry)
524 struct xsk_map *map = NULL;
525 struct xsk_map_node *node;
529 spin_lock_bh(&xs->map_list_lock);
530 node = list_first_entry_or_null(&xs->map_list, struct xsk_map_node,
533 WARN_ON(xsk_map_inc(node->map));
535 *map_entry = node->map_entry;
537 spin_unlock_bh(&xs->map_list_lock);
541 static void xsk_delete_from_maps(struct xdp_sock *xs)
543 /* This function removes the current XDP socket from all the
544 * maps it resides in. We need to take extra care here, due to
545 * the two locks involved. Each map has a lock synchronizing
546 * updates to the entries, and each socket has a lock that
547 * synchronizes access to the list of maps (map_list). For
548 * deadlock avoidance the locks need to be taken in the order
549 * "map lock"->"socket map list lock". We start off by
550 * accessing the socket map list, and take a reference to the
551 * map to guarantee existence between the
552 * xsk_get_map_list_entry() and xsk_map_try_sock_delete()
553 * calls. Then we ask the map to remove the socket, which
554 * tries to remove the socket from the map. Note that there
555 * might be updates to the map between
556 * xsk_get_map_list_entry() and xsk_map_try_sock_delete().
558 struct xdp_sock **map_entry = NULL;
561 while ((map = xsk_get_map_list_entry(xs, &map_entry))) {
562 xsk_map_try_sock_delete(map, xs, map_entry);
567 static int xsk_release(struct socket *sock)
569 struct sock *sk = sock->sk;
570 struct xdp_sock *xs = xdp_sk(sk);
578 mutex_lock(&net->xdp.lock);
579 sk_del_node_init_rcu(sk);
580 mutex_unlock(&net->xdp.lock);
583 sock_prot_inuse_add(net, sk->sk_prot, -1);
586 xsk_delete_from_maps(xs);
587 mutex_lock(&xs->mutex);
589 mutex_unlock(&xs->mutex);
591 xskq_destroy(xs->rx);
592 xskq_destroy(xs->tx);
597 sk_refcnt_debug_release(sk);
603 static struct socket *xsk_lookup_xsk_from_fd(int fd)
608 sock = sockfd_lookup(fd, &err);
610 return ERR_PTR(-ENOTSOCK);
612 if (sock->sk->sk_family != PF_XDP) {
614 return ERR_PTR(-ENOPROTOOPT);
620 /* Check if umem pages are contiguous.
621 * If zero-copy mode, use the DMA address to do the page contiguity check
622 * For all other modes we use addr (kernel virtual address)
623 * Store the result in the low bits of addr.
625 static void xsk_check_page_contiguity(struct xdp_umem *umem, u32 flags)
627 struct xdp_umem_page *pgs = umem->pages;
630 for (i = 0; i < umem->npgs - 1; i++) {
631 is_contig = (flags & XDP_ZEROCOPY) ?
632 (pgs[i].dma + PAGE_SIZE == pgs[i + 1].dma) :
633 (pgs[i].addr + PAGE_SIZE == pgs[i + 1].addr);
634 pgs[i].addr += is_contig << XSK_NEXT_PG_CONTIG_SHIFT;
638 static int xsk_bind(struct socket *sock, struct sockaddr *addr, int addr_len)
640 struct sockaddr_xdp *sxdp = (struct sockaddr_xdp *)addr;
641 struct sock *sk = sock->sk;
642 struct xdp_sock *xs = xdp_sk(sk);
643 struct net_device *dev;
647 if (addr_len < sizeof(struct sockaddr_xdp))
649 if (sxdp->sxdp_family != AF_XDP)
652 flags = sxdp->sxdp_flags;
653 if (flags & ~(XDP_SHARED_UMEM | XDP_COPY | XDP_ZEROCOPY |
654 XDP_USE_NEED_WAKEUP))
658 mutex_lock(&xs->mutex);
659 if (xs->state != XSK_READY) {
664 dev = dev_get_by_index(sock_net(sk), sxdp->sxdp_ifindex);
670 if (!xs->rx && !xs->tx) {
675 qid = sxdp->sxdp_queue_id;
677 if (flags & XDP_SHARED_UMEM) {
678 struct xdp_sock *umem_xs;
681 if ((flags & XDP_COPY) || (flags & XDP_ZEROCOPY) ||
682 (flags & XDP_USE_NEED_WAKEUP)) {
683 /* Cannot specify flags for shared sockets. */
689 /* We have already our own. */
694 sock = xsk_lookup_xsk_from_fd(sxdp->sxdp_shared_umem_fd);
700 umem_xs = xdp_sk(sock->sk);
701 if (!xsk_is_bound(umem_xs)) {
706 if (umem_xs->dev != dev || umem_xs->queue_id != qid) {
712 xdp_get_umem(umem_xs->umem);
713 WRITE_ONCE(xs->umem, umem_xs->umem);
715 } else if (!xs->umem || !xdp_umem_validate_queues(xs->umem)) {
719 /* This xsk has its own umem. */
720 xskq_set_umem(xs->umem->fq, xs->umem->size,
721 xs->umem->chunk_mask);
722 xskq_set_umem(xs->umem->cq, xs->umem->size,
723 xs->umem->chunk_mask);
725 err = xdp_umem_assign_dev(xs->umem, dev, qid, flags);
729 xsk_check_page_contiguity(xs->umem, flags);
733 xs->zc = xs->umem->zc;
735 xskq_set_umem(xs->rx, xs->umem->size, xs->umem->chunk_mask);
736 xskq_set_umem(xs->tx, xs->umem->size, xs->umem->chunk_mask);
737 xdp_add_sk_umem(xs->umem, xs);
743 /* Matches smp_rmb() in bind() for shared umem
744 * sockets, and xsk_is_bound().
747 WRITE_ONCE(xs->state, XSK_BOUND);
750 mutex_unlock(&xs->mutex);
755 struct xdp_umem_reg_v1 {
756 __u64 addr; /* Start of packet data area */
757 __u64 len; /* Length of packet data area */
762 static int xsk_setsockopt(struct socket *sock, int level, int optname,
763 char __user *optval, unsigned int optlen)
765 struct sock *sk = sock->sk;
766 struct xdp_sock *xs = xdp_sk(sk);
769 if (level != SOL_XDP)
776 struct xsk_queue **q;
779 if (optlen < sizeof(entries))
781 if (copy_from_user(&entries, optval, sizeof(entries)))
784 mutex_lock(&xs->mutex);
785 if (xs->state != XSK_READY) {
786 mutex_unlock(&xs->mutex);
789 q = (optname == XDP_TX_RING) ? &xs->tx : &xs->rx;
790 err = xsk_init_queue(entries, q, false);
791 if (!err && optname == XDP_TX_RING)
792 /* Tx needs to be explicitly woken up the first time */
793 xs->tx->ring->flags |= XDP_RING_NEED_WAKEUP;
794 mutex_unlock(&xs->mutex);
799 size_t mr_size = sizeof(struct xdp_umem_reg);
800 struct xdp_umem_reg mr = {};
801 struct xdp_umem *umem;
803 if (optlen < sizeof(struct xdp_umem_reg_v1))
805 else if (optlen < sizeof(mr))
806 mr_size = sizeof(struct xdp_umem_reg_v1);
808 if (copy_from_user(&mr, optval, mr_size))
811 mutex_lock(&xs->mutex);
812 if (xs->state != XSK_READY || xs->umem) {
813 mutex_unlock(&xs->mutex);
817 umem = xdp_umem_create(&mr);
819 mutex_unlock(&xs->mutex);
820 return PTR_ERR(umem);
823 /* Make sure umem is ready before it can be seen by others */
825 WRITE_ONCE(xs->umem, umem);
826 mutex_unlock(&xs->mutex);
829 case XDP_UMEM_FILL_RING:
830 case XDP_UMEM_COMPLETION_RING:
832 struct xsk_queue **q;
835 if (copy_from_user(&entries, optval, sizeof(entries)))
838 mutex_lock(&xs->mutex);
839 if (xs->state != XSK_READY) {
840 mutex_unlock(&xs->mutex);
844 mutex_unlock(&xs->mutex);
848 q = (optname == XDP_UMEM_FILL_RING) ? &xs->umem->fq :
850 err = xsk_init_queue(entries, q, true);
851 mutex_unlock(&xs->mutex);
861 static void xsk_enter_rxtx_offsets(struct xdp_ring_offset_v1 *ring)
863 ring->producer = offsetof(struct xdp_rxtx_ring, ptrs.producer);
864 ring->consumer = offsetof(struct xdp_rxtx_ring, ptrs.consumer);
865 ring->desc = offsetof(struct xdp_rxtx_ring, desc);
868 static void xsk_enter_umem_offsets(struct xdp_ring_offset_v1 *ring)
870 ring->producer = offsetof(struct xdp_umem_ring, ptrs.producer);
871 ring->consumer = offsetof(struct xdp_umem_ring, ptrs.consumer);
872 ring->desc = offsetof(struct xdp_umem_ring, desc);
875 static int xsk_getsockopt(struct socket *sock, int level, int optname,
876 char __user *optval, int __user *optlen)
878 struct sock *sk = sock->sk;
879 struct xdp_sock *xs = xdp_sk(sk);
882 if (level != SOL_XDP)
885 if (get_user(len, optlen))
893 struct xdp_statistics stats;
895 if (len < sizeof(stats))
898 mutex_lock(&xs->mutex);
899 stats.rx_dropped = xs->rx_dropped;
900 stats.rx_invalid_descs = xskq_nb_invalid_descs(xs->rx);
901 stats.tx_invalid_descs = xskq_nb_invalid_descs(xs->tx);
902 mutex_unlock(&xs->mutex);
904 if (copy_to_user(optval, &stats, sizeof(stats)))
906 if (put_user(sizeof(stats), optlen))
911 case XDP_MMAP_OFFSETS:
913 struct xdp_mmap_offsets off;
914 struct xdp_mmap_offsets_v1 off_v1;
915 bool flags_supported = true;
918 if (len < sizeof(off_v1))
920 else if (len < sizeof(off))
921 flags_supported = false;
923 if (flags_supported) {
924 /* xdp_ring_offset is identical to xdp_ring_offset_v1
925 * except for the flags field added to the end.
927 xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *)
929 xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *)
931 xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *)
933 xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *)
935 off.rx.flags = offsetof(struct xdp_rxtx_ring,
937 off.tx.flags = offsetof(struct xdp_rxtx_ring,
939 off.fr.flags = offsetof(struct xdp_umem_ring,
941 off.cr.flags = offsetof(struct xdp_umem_ring,
947 xsk_enter_rxtx_offsets(&off_v1.rx);
948 xsk_enter_rxtx_offsets(&off_v1.tx);
949 xsk_enter_umem_offsets(&off_v1.fr);
950 xsk_enter_umem_offsets(&off_v1.cr);
952 len = sizeof(off_v1);
956 if (copy_to_user(optval, to_copy, len))
958 if (put_user(len, optlen))
965 struct xdp_options opts = {};
967 if (len < sizeof(opts))
970 mutex_lock(&xs->mutex);
972 opts.flags |= XDP_OPTIONS_ZEROCOPY;
973 mutex_unlock(&xs->mutex);
976 if (copy_to_user(optval, &opts, len))
978 if (put_user(len, optlen))
990 static int xsk_mmap(struct file *file, struct socket *sock,
991 struct vm_area_struct *vma)
993 loff_t offset = (loff_t)vma->vm_pgoff << PAGE_SHIFT;
994 unsigned long size = vma->vm_end - vma->vm_start;
995 struct xdp_sock *xs = xdp_sk(sock->sk);
996 struct xsk_queue *q = NULL;
997 struct xdp_umem *umem;
1001 if (READ_ONCE(xs->state) != XSK_READY)
1004 if (offset == XDP_PGOFF_RX_RING) {
1005 q = READ_ONCE(xs->rx);
1006 } else if (offset == XDP_PGOFF_TX_RING) {
1007 q = READ_ONCE(xs->tx);
1009 umem = READ_ONCE(xs->umem);
1013 /* Matches the smp_wmb() in XDP_UMEM_REG */
1015 if (offset == XDP_UMEM_PGOFF_FILL_RING)
1016 q = READ_ONCE(umem->fq);
1017 else if (offset == XDP_UMEM_PGOFF_COMPLETION_RING)
1018 q = READ_ONCE(umem->cq);
1024 /* Matches the smp_wmb() in xsk_init_queue */
1026 qpg = virt_to_head_page(q->ring);
1027 if (size > page_size(qpg))
1030 pfn = virt_to_phys(q->ring) >> PAGE_SHIFT;
1031 return remap_pfn_range(vma, vma->vm_start, pfn,
1032 size, vma->vm_page_prot);
1035 static int xsk_notifier(struct notifier_block *this,
1036 unsigned long msg, void *ptr)
1038 struct net_device *dev = netdev_notifier_info_to_dev(ptr);
1039 struct net *net = dev_net(dev);
1043 case NETDEV_UNREGISTER:
1044 mutex_lock(&net->xdp.lock);
1045 sk_for_each(sk, &net->xdp.list) {
1046 struct xdp_sock *xs = xdp_sk(sk);
1048 mutex_lock(&xs->mutex);
1049 if (xs->dev == dev) {
1050 sk->sk_err = ENETDOWN;
1051 if (!sock_flag(sk, SOCK_DEAD))
1052 sk->sk_error_report(sk);
1056 /* Clear device references in umem. */
1057 xdp_umem_clear_dev(xs->umem);
1059 mutex_unlock(&xs->mutex);
1061 mutex_unlock(&net->xdp.lock);
1067 static struct proto xsk_proto = {
1069 .owner = THIS_MODULE,
1070 .obj_size = sizeof(struct xdp_sock),
1073 static const struct proto_ops xsk_proto_ops = {
1075 .owner = THIS_MODULE,
1076 .release = xsk_release,
1078 .connect = sock_no_connect,
1079 .socketpair = sock_no_socketpair,
1080 .accept = sock_no_accept,
1081 .getname = sock_no_getname,
1083 .ioctl = sock_no_ioctl,
1084 .listen = sock_no_listen,
1085 .shutdown = sock_no_shutdown,
1086 .setsockopt = xsk_setsockopt,
1087 .getsockopt = xsk_getsockopt,
1088 .sendmsg = xsk_sendmsg,
1089 .recvmsg = sock_no_recvmsg,
1091 .sendpage = sock_no_sendpage,
1094 static void xsk_destruct(struct sock *sk)
1096 struct xdp_sock *xs = xdp_sk(sk);
1098 if (!sock_flag(sk, SOCK_DEAD))
1101 xdp_put_umem(xs->umem);
1103 sk_refcnt_debug_dec(sk);
1106 static int xsk_create(struct net *net, struct socket *sock, int protocol,
1110 struct xdp_sock *xs;
1112 if (!ns_capable(net->user_ns, CAP_NET_RAW))
1114 if (sock->type != SOCK_RAW)
1115 return -ESOCKTNOSUPPORT;
1118 return -EPROTONOSUPPORT;
1120 sock->state = SS_UNCONNECTED;
1122 sk = sk_alloc(net, PF_XDP, GFP_KERNEL, &xsk_proto, kern);
1126 sock->ops = &xsk_proto_ops;
1128 sock_init_data(sock, sk);
1130 sk->sk_family = PF_XDP;
1132 sk->sk_destruct = xsk_destruct;
1133 sk_refcnt_debug_inc(sk);
1135 sock_set_flag(sk, SOCK_RCU_FREE);
1138 xs->state = XSK_READY;
1139 mutex_init(&xs->mutex);
1140 spin_lock_init(&xs->rx_lock);
1141 spin_lock_init(&xs->tx_completion_lock);
1143 INIT_LIST_HEAD(&xs->map_list);
1144 spin_lock_init(&xs->map_list_lock);
1146 mutex_lock(&net->xdp.lock);
1147 sk_add_node_rcu(sk, &net->xdp.list);
1148 mutex_unlock(&net->xdp.lock);
1151 sock_prot_inuse_add(net, &xsk_proto, 1);
1157 static const struct net_proto_family xsk_family_ops = {
1159 .create = xsk_create,
1160 .owner = THIS_MODULE,
1163 static struct notifier_block xsk_netdev_notifier = {
1164 .notifier_call = xsk_notifier,
1167 static int __net_init xsk_net_init(struct net *net)
1169 mutex_init(&net->xdp.lock);
1170 INIT_HLIST_HEAD(&net->xdp.list);
1174 static void __net_exit xsk_net_exit(struct net *net)
1176 WARN_ON_ONCE(!hlist_empty(&net->xdp.list));
1179 static struct pernet_operations xsk_net_ops = {
1180 .init = xsk_net_init,
1181 .exit = xsk_net_exit,
1184 static int __init xsk_init(void)
1188 err = proto_register(&xsk_proto, 0 /* no slab */);
1192 err = sock_register(&xsk_family_ops);
1196 err = register_pernet_subsys(&xsk_net_ops);
1200 err = register_netdevice_notifier(&xsk_netdev_notifier);
1207 unregister_pernet_subsys(&xsk_net_ops);
1209 sock_unregister(PF_XDP);
1211 proto_unregister(&xsk_proto);
1216 fs_initcall(xsk_init);