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);
133 memcpy(next_pg_addr, from_buf + first_len,
134 len + metalen - first_len);
139 memcpy(to_buf, from_buf, len + metalen);
142 static int __xsk_rcv(struct xdp_sock *xs, struct xdp_buff *xdp, u32 len)
144 u64 offset = xs->umem->headroom;
145 u64 addr, memcpy_addr;
150 if (!xskq_peek_addr(xs->umem->fq, &addr, xs->umem) ||
151 len > xs->umem->chunk_size_nohr - XDP_PACKET_HEADROOM) {
156 if (unlikely(xdp_data_meta_unsupported(xdp))) {
157 from_buf = xdp->data;
160 from_buf = xdp->data_meta;
161 metalen = xdp->data - xdp->data_meta;
164 memcpy_addr = xsk_umem_adjust_offset(xs->umem, addr, offset);
165 __xsk_rcv_memcpy(xs->umem, memcpy_addr, from_buf, len, metalen);
168 addr = xsk_umem_adjust_offset(xs->umem, addr, offset);
169 err = xskq_produce_batch_desc(xs->rx, addr, len);
171 xskq_discard_addr(xs->umem->fq);
172 xdp_return_buff(xdp);
180 static int __xsk_rcv_zc(struct xdp_sock *xs, struct xdp_buff *xdp, u32 len)
182 int err = xskq_produce_batch_desc(xs->rx, (u64)xdp->handle, len);
190 static bool xsk_is_bound(struct xdp_sock *xs)
192 if (READ_ONCE(xs->state) == XSK_BOUND) {
193 /* Matches smp_wmb() in bind(). */
200 int xsk_rcv(struct xdp_sock *xs, struct xdp_buff *xdp)
204 if (!xsk_is_bound(xs))
207 if (xs->dev != xdp->rxq->dev || xs->queue_id != xdp->rxq->queue_index)
210 len = xdp->data_end - xdp->data;
212 return (xdp->rxq->mem.type == MEM_TYPE_ZERO_COPY) ?
213 __xsk_rcv_zc(xs, xdp, len) : __xsk_rcv(xs, xdp, len);
216 void xsk_flush(struct xdp_sock *xs)
218 xskq_produce_flush_desc(xs->rx);
219 xs->sk.sk_data_ready(&xs->sk);
222 int xsk_generic_rcv(struct xdp_sock *xs, struct xdp_buff *xdp)
224 u32 metalen = xdp->data - xdp->data_meta;
225 u32 len = xdp->data_end - xdp->data;
226 u64 offset = xs->umem->headroom;
231 spin_lock_bh(&xs->rx_lock);
233 if (xs->dev != xdp->rxq->dev || xs->queue_id != xdp->rxq->queue_index) {
238 if (!xskq_peek_addr(xs->umem->fq, &addr, xs->umem) ||
239 len > xs->umem->chunk_size_nohr - XDP_PACKET_HEADROOM) {
244 addr = xsk_umem_adjust_offset(xs->umem, addr, offset);
245 buffer = xdp_umem_get_data(xs->umem, addr);
246 memcpy(buffer, xdp->data_meta, len + metalen);
248 addr = xsk_umem_adjust_offset(xs->umem, addr, metalen);
249 err = xskq_produce_batch_desc(xs->rx, addr, len);
253 xskq_discard_addr(xs->umem->fq);
254 xskq_produce_flush_desc(xs->rx);
256 spin_unlock_bh(&xs->rx_lock);
258 xs->sk.sk_data_ready(&xs->sk);
264 spin_unlock_bh(&xs->rx_lock);
268 void xsk_umem_complete_tx(struct xdp_umem *umem, u32 nb_entries)
270 xskq_produce_flush_addr_n(umem->cq, nb_entries);
272 EXPORT_SYMBOL(xsk_umem_complete_tx);
274 void xsk_umem_consume_tx_done(struct xdp_umem *umem)
279 list_for_each_entry_rcu(xs, &umem->xsk_list, list) {
280 xs->sk.sk_write_space(&xs->sk);
284 EXPORT_SYMBOL(xsk_umem_consume_tx_done);
286 bool xsk_umem_consume_tx(struct xdp_umem *umem, struct xdp_desc *desc)
291 list_for_each_entry_rcu(xs, &umem->xsk_list, list) {
292 if (!xskq_peek_desc(xs->tx, desc, umem))
295 if (xskq_produce_addr_lazy(umem->cq, desc->addr))
298 xskq_discard_desc(xs->tx);
307 EXPORT_SYMBOL(xsk_umem_consume_tx);
309 static int xsk_wakeup(struct xdp_sock *xs, u8 flags)
311 struct net_device *dev = xs->dev;
315 err = dev->netdev_ops->ndo_xsk_wakeup(dev, xs->queue_id, flags);
321 static int xsk_zc_xmit(struct xdp_sock *xs)
323 return xsk_wakeup(xs, XDP_WAKEUP_TX);
326 static void xsk_destruct_skb(struct sk_buff *skb)
328 u64 addr = (u64)(long)skb_shinfo(skb)->destructor_arg;
329 struct xdp_sock *xs = xdp_sk(skb->sk);
332 spin_lock_irqsave(&xs->tx_completion_lock, flags);
333 WARN_ON_ONCE(xskq_produce_addr(xs->umem->cq, addr));
334 spin_unlock_irqrestore(&xs->tx_completion_lock, flags);
339 static int xsk_generic_xmit(struct sock *sk)
341 struct xdp_sock *xs = xdp_sk(sk);
342 u32 max_batch = TX_BATCH_SIZE;
343 bool sent_frame = false;
344 struct xdp_desc desc;
348 mutex_lock(&xs->mutex);
350 if (xs->queue_id >= xs->dev->real_num_tx_queues)
353 while (xskq_peek_desc(xs->tx, &desc, xs->umem)) {
358 if (max_batch-- == 0) {
364 skb = sock_alloc_send_skb(sk, len, 1, &err);
370 buffer = xdp_umem_get_data(xs->umem, addr);
371 err = skb_store_bits(skb, 0, buffer, len);
372 if (unlikely(err) || xskq_reserve_addr(xs->umem->cq)) {
378 skb->priority = sk->sk_priority;
379 skb->mark = sk->sk_mark;
380 skb_shinfo(skb)->destructor_arg = (void *)(long)desc.addr;
381 skb->destructor = xsk_destruct_skb;
383 err = dev_direct_xmit(skb, xs->queue_id);
384 xskq_discard_desc(xs->tx);
385 /* Ignore NET_XMIT_CN as packet might have been sent */
386 if (err == NET_XMIT_DROP || err == NETDEV_TX_BUSY) {
387 /* SKB completed but not sent */
397 sk->sk_write_space(sk);
399 mutex_unlock(&xs->mutex);
403 static int __xsk_sendmsg(struct sock *sk)
405 struct xdp_sock *xs = xdp_sk(sk);
407 if (unlikely(!(xs->dev->flags & IFF_UP)))
409 if (unlikely(!xs->tx))
412 return xs->zc ? xsk_zc_xmit(xs) : xsk_generic_xmit(sk);
415 static int xsk_sendmsg(struct socket *sock, struct msghdr *m, size_t total_len)
417 bool need_wait = !(m->msg_flags & MSG_DONTWAIT);
418 struct sock *sk = sock->sk;
419 struct xdp_sock *xs = xdp_sk(sk);
421 if (unlikely(!xsk_is_bound(xs)))
423 if (unlikely(need_wait))
426 return __xsk_sendmsg(sk);
429 static __poll_t xsk_poll(struct file *file, struct socket *sock,
430 struct poll_table_struct *wait)
433 struct sock *sk = sock->sk;
434 struct xdp_sock *xs = xdp_sk(sk);
435 struct xdp_umem *umem;
437 sock_poll_wait(file, sock, wait);
439 if (unlikely(!xsk_is_bound(xs)))
444 if (umem->need_wakeup) {
446 xsk_wakeup(xs, umem->need_wakeup);
448 /* Poll needs to drive Tx also in copy mode */
452 if (xs->rx && !xskq_empty_desc(xs->rx))
453 mask |= EPOLLIN | EPOLLRDNORM;
454 if (xs->tx && !xskq_full_desc(xs->tx))
455 mask |= EPOLLOUT | EPOLLWRNORM;
460 static int xsk_init_queue(u32 entries, struct xsk_queue **queue,
465 if (entries == 0 || *queue || !is_power_of_2(entries))
468 q = xskq_create(entries, umem_queue);
472 /* Make sure queue is ready before it can be seen by others */
474 WRITE_ONCE(*queue, q);
478 static void xsk_unbind_dev(struct xdp_sock *xs)
480 struct net_device *dev = xs->dev;
482 if (xs->state != XSK_BOUND)
484 WRITE_ONCE(xs->state, XSK_UNBOUND);
486 /* Wait for driver to stop using the xdp socket. */
487 xdp_del_sk_umem(xs->umem, xs);
493 static struct xsk_map *xsk_get_map_list_entry(struct xdp_sock *xs,
494 struct xdp_sock ***map_entry)
496 struct xsk_map *map = NULL;
497 struct xsk_map_node *node;
501 spin_lock_bh(&xs->map_list_lock);
502 node = list_first_entry_or_null(&xs->map_list, struct xsk_map_node,
505 WARN_ON(xsk_map_inc(node->map));
507 *map_entry = node->map_entry;
509 spin_unlock_bh(&xs->map_list_lock);
513 static void xsk_delete_from_maps(struct xdp_sock *xs)
515 /* This function removes the current XDP socket from all the
516 * maps it resides in. We need to take extra care here, due to
517 * the two locks involved. Each map has a lock synchronizing
518 * updates to the entries, and each socket has a lock that
519 * synchronizes access to the list of maps (map_list). For
520 * deadlock avoidance the locks need to be taken in the order
521 * "map lock"->"socket map list lock". We start off by
522 * accessing the socket map list, and take a reference to the
523 * map to guarantee existence between the
524 * xsk_get_map_list_entry() and xsk_map_try_sock_delete()
525 * calls. Then we ask the map to remove the socket, which
526 * tries to remove the socket from the map. Note that there
527 * might be updates to the map between
528 * xsk_get_map_list_entry() and xsk_map_try_sock_delete().
530 struct xdp_sock **map_entry = NULL;
533 while ((map = xsk_get_map_list_entry(xs, &map_entry))) {
534 xsk_map_try_sock_delete(map, xs, map_entry);
539 static int xsk_release(struct socket *sock)
541 struct sock *sk = sock->sk;
542 struct xdp_sock *xs = xdp_sk(sk);
550 mutex_lock(&net->xdp.lock);
551 sk_del_node_init_rcu(sk);
552 mutex_unlock(&net->xdp.lock);
555 sock_prot_inuse_add(net, sk->sk_prot, -1);
558 xsk_delete_from_maps(xs);
559 mutex_lock(&xs->mutex);
561 mutex_unlock(&xs->mutex);
563 xskq_destroy(xs->rx);
564 xskq_destroy(xs->tx);
569 sk_refcnt_debug_release(sk);
575 static struct socket *xsk_lookup_xsk_from_fd(int fd)
580 sock = sockfd_lookup(fd, &err);
582 return ERR_PTR(-ENOTSOCK);
584 if (sock->sk->sk_family != PF_XDP) {
586 return ERR_PTR(-ENOPROTOOPT);
592 /* Check if umem pages are contiguous.
593 * If zero-copy mode, use the DMA address to do the page contiguity check
594 * For all other modes we use addr (kernel virtual address)
595 * Store the result in the low bits of addr.
597 static void xsk_check_page_contiguity(struct xdp_umem *umem, u32 flags)
599 struct xdp_umem_page *pgs = umem->pages;
602 for (i = 0; i < umem->npgs - 1; i++) {
603 is_contig = (flags & XDP_ZEROCOPY) ?
604 (pgs[i].dma + PAGE_SIZE == pgs[i + 1].dma) :
605 (pgs[i].addr + PAGE_SIZE == pgs[i + 1].addr);
606 pgs[i].addr += is_contig << XSK_NEXT_PG_CONTIG_SHIFT;
610 static int xsk_bind(struct socket *sock, struct sockaddr *addr, int addr_len)
612 struct sockaddr_xdp *sxdp = (struct sockaddr_xdp *)addr;
613 struct sock *sk = sock->sk;
614 struct xdp_sock *xs = xdp_sk(sk);
615 struct net_device *dev;
620 if (addr_len < sizeof(struct sockaddr_xdp))
622 if (sxdp->sxdp_family != AF_XDP)
625 flags = sxdp->sxdp_flags;
626 if (flags & ~(XDP_SHARED_UMEM | XDP_COPY | XDP_ZEROCOPY |
627 XDP_USE_NEED_WAKEUP))
630 bound_dev_if = READ_ONCE(sk->sk_bound_dev_if);
631 if (bound_dev_if && bound_dev_if != sxdp->sxdp_ifindex)
635 mutex_lock(&xs->mutex);
636 if (xs->state != XSK_READY) {
641 dev = dev_get_by_index(sock_net(sk), sxdp->sxdp_ifindex);
647 if (!xs->rx && !xs->tx) {
652 qid = sxdp->sxdp_queue_id;
654 if (flags & XDP_SHARED_UMEM) {
655 struct xdp_sock *umem_xs;
658 if ((flags & XDP_COPY) || (flags & XDP_ZEROCOPY) ||
659 (flags & XDP_USE_NEED_WAKEUP)) {
660 /* Cannot specify flags for shared sockets. */
666 /* We have already our own. */
671 sock = xsk_lookup_xsk_from_fd(sxdp->sxdp_shared_umem_fd);
677 umem_xs = xdp_sk(sock->sk);
678 if (!xsk_is_bound(umem_xs)) {
683 if (umem_xs->dev != dev || umem_xs->queue_id != qid) {
689 xdp_get_umem(umem_xs->umem);
690 WRITE_ONCE(xs->umem, umem_xs->umem);
692 } else if (!xs->umem || !xdp_umem_validate_queues(xs->umem)) {
696 /* This xsk has its own umem. */
697 xskq_set_umem(xs->umem->fq, xs->umem->size,
698 xs->umem->chunk_mask);
699 xskq_set_umem(xs->umem->cq, xs->umem->size,
700 xs->umem->chunk_mask);
702 err = xdp_umem_assign_dev(xs->umem, dev, qid, flags);
706 xsk_check_page_contiguity(xs->umem, flags);
710 xs->zc = xs->umem->zc;
712 xskq_set_umem(xs->rx, xs->umem->size, xs->umem->chunk_mask);
713 xskq_set_umem(xs->tx, xs->umem->size, xs->umem->chunk_mask);
714 xdp_add_sk_umem(xs->umem, xs);
720 /* Matches smp_rmb() in bind() for shared umem
721 * sockets, and xsk_is_bound().
724 WRITE_ONCE(xs->state, XSK_BOUND);
727 mutex_unlock(&xs->mutex);
732 struct xdp_umem_reg_v1 {
733 __u64 addr; /* Start of packet data area */
734 __u64 len; /* Length of packet data area */
739 static int xsk_setsockopt(struct socket *sock, int level, int optname,
740 char __user *optval, unsigned int optlen)
742 struct sock *sk = sock->sk;
743 struct xdp_sock *xs = xdp_sk(sk);
746 if (level != SOL_XDP)
753 struct xsk_queue **q;
756 if (optlen < sizeof(entries))
758 if (copy_from_user(&entries, optval, sizeof(entries)))
761 mutex_lock(&xs->mutex);
762 if (xs->state != XSK_READY) {
763 mutex_unlock(&xs->mutex);
766 q = (optname == XDP_TX_RING) ? &xs->tx : &xs->rx;
767 err = xsk_init_queue(entries, q, false);
768 if (!err && optname == XDP_TX_RING)
769 /* Tx needs to be explicitly woken up the first time */
770 xs->tx->ring->flags |= XDP_RING_NEED_WAKEUP;
771 mutex_unlock(&xs->mutex);
776 size_t mr_size = sizeof(struct xdp_umem_reg);
777 struct xdp_umem_reg mr = {};
778 struct xdp_umem *umem;
780 if (optlen < sizeof(struct xdp_umem_reg_v1))
782 else if (optlen < sizeof(mr))
783 mr_size = sizeof(struct xdp_umem_reg_v1);
785 if (copy_from_user(&mr, optval, mr_size))
788 mutex_lock(&xs->mutex);
789 if (xs->state != XSK_READY || xs->umem) {
790 mutex_unlock(&xs->mutex);
794 umem = xdp_umem_create(&mr);
796 mutex_unlock(&xs->mutex);
797 return PTR_ERR(umem);
800 /* Make sure umem is ready before it can be seen by others */
802 WRITE_ONCE(xs->umem, umem);
803 mutex_unlock(&xs->mutex);
806 case XDP_UMEM_FILL_RING:
807 case XDP_UMEM_COMPLETION_RING:
809 struct xsk_queue **q;
812 if (copy_from_user(&entries, optval, sizeof(entries)))
815 mutex_lock(&xs->mutex);
816 if (xs->state != XSK_READY) {
817 mutex_unlock(&xs->mutex);
821 mutex_unlock(&xs->mutex);
825 q = (optname == XDP_UMEM_FILL_RING) ? &xs->umem->fq :
827 err = xsk_init_queue(entries, q, true);
828 mutex_unlock(&xs->mutex);
838 static void xsk_enter_rxtx_offsets(struct xdp_ring_offset_v1 *ring)
840 ring->producer = offsetof(struct xdp_rxtx_ring, ptrs.producer);
841 ring->consumer = offsetof(struct xdp_rxtx_ring, ptrs.consumer);
842 ring->desc = offsetof(struct xdp_rxtx_ring, desc);
845 static void xsk_enter_umem_offsets(struct xdp_ring_offset_v1 *ring)
847 ring->producer = offsetof(struct xdp_umem_ring, ptrs.producer);
848 ring->consumer = offsetof(struct xdp_umem_ring, ptrs.consumer);
849 ring->desc = offsetof(struct xdp_umem_ring, desc);
852 static int xsk_getsockopt(struct socket *sock, int level, int optname,
853 char __user *optval, int __user *optlen)
855 struct sock *sk = sock->sk;
856 struct xdp_sock *xs = xdp_sk(sk);
859 if (level != SOL_XDP)
862 if (get_user(len, optlen))
870 struct xdp_statistics stats;
872 if (len < sizeof(stats))
875 mutex_lock(&xs->mutex);
876 stats.rx_dropped = xs->rx_dropped;
877 stats.rx_invalid_descs = xskq_nb_invalid_descs(xs->rx);
878 stats.tx_invalid_descs = xskq_nb_invalid_descs(xs->tx);
879 mutex_unlock(&xs->mutex);
881 if (copy_to_user(optval, &stats, sizeof(stats)))
883 if (put_user(sizeof(stats), optlen))
888 case XDP_MMAP_OFFSETS:
890 struct xdp_mmap_offsets off;
891 struct xdp_mmap_offsets_v1 off_v1;
892 bool flags_supported = true;
895 if (len < sizeof(off_v1))
897 else if (len < sizeof(off))
898 flags_supported = false;
900 if (flags_supported) {
901 /* xdp_ring_offset is identical to xdp_ring_offset_v1
902 * except for the flags field added to the end.
904 xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *)
906 xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *)
908 xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *)
910 xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *)
912 off.rx.flags = offsetof(struct xdp_rxtx_ring,
914 off.tx.flags = offsetof(struct xdp_rxtx_ring,
916 off.fr.flags = offsetof(struct xdp_umem_ring,
918 off.cr.flags = offsetof(struct xdp_umem_ring,
924 xsk_enter_rxtx_offsets(&off_v1.rx);
925 xsk_enter_rxtx_offsets(&off_v1.tx);
926 xsk_enter_umem_offsets(&off_v1.fr);
927 xsk_enter_umem_offsets(&off_v1.cr);
929 len = sizeof(off_v1);
933 if (copy_to_user(optval, to_copy, len))
935 if (put_user(len, optlen))
942 struct xdp_options opts = {};
944 if (len < sizeof(opts))
947 mutex_lock(&xs->mutex);
949 opts.flags |= XDP_OPTIONS_ZEROCOPY;
950 mutex_unlock(&xs->mutex);
953 if (copy_to_user(optval, &opts, len))
955 if (put_user(len, optlen))
967 static int xsk_mmap(struct file *file, struct socket *sock,
968 struct vm_area_struct *vma)
970 loff_t offset = (loff_t)vma->vm_pgoff << PAGE_SHIFT;
971 unsigned long size = vma->vm_end - vma->vm_start;
972 struct xdp_sock *xs = xdp_sk(sock->sk);
973 struct xsk_queue *q = NULL;
974 struct xdp_umem *umem;
978 if (READ_ONCE(xs->state) != XSK_READY)
981 if (offset == XDP_PGOFF_RX_RING) {
982 q = READ_ONCE(xs->rx);
983 } else if (offset == XDP_PGOFF_TX_RING) {
984 q = READ_ONCE(xs->tx);
986 umem = READ_ONCE(xs->umem);
990 /* Matches the smp_wmb() in XDP_UMEM_REG */
992 if (offset == XDP_UMEM_PGOFF_FILL_RING)
993 q = READ_ONCE(umem->fq);
994 else if (offset == XDP_UMEM_PGOFF_COMPLETION_RING)
995 q = READ_ONCE(umem->cq);
1001 /* Matches the smp_wmb() in xsk_init_queue */
1003 qpg = virt_to_head_page(q->ring);
1004 if (size > page_size(qpg))
1007 pfn = virt_to_phys(q->ring) >> PAGE_SHIFT;
1008 return remap_pfn_range(vma, vma->vm_start, pfn,
1009 size, vma->vm_page_prot);
1012 static int xsk_notifier(struct notifier_block *this,
1013 unsigned long msg, void *ptr)
1015 struct net_device *dev = netdev_notifier_info_to_dev(ptr);
1016 struct net *net = dev_net(dev);
1020 case NETDEV_UNREGISTER:
1021 mutex_lock(&net->xdp.lock);
1022 sk_for_each(sk, &net->xdp.list) {
1023 struct xdp_sock *xs = xdp_sk(sk);
1025 mutex_lock(&xs->mutex);
1026 if (xs->dev == dev) {
1027 sk->sk_err = ENETDOWN;
1028 if (!sock_flag(sk, SOCK_DEAD))
1029 sk->sk_error_report(sk);
1033 /* Clear device references in umem. */
1034 xdp_umem_clear_dev(xs->umem);
1036 mutex_unlock(&xs->mutex);
1038 mutex_unlock(&net->xdp.lock);
1044 static struct proto xsk_proto = {
1046 .owner = THIS_MODULE,
1047 .obj_size = sizeof(struct xdp_sock),
1050 static const struct proto_ops xsk_proto_ops = {
1052 .owner = THIS_MODULE,
1053 .release = xsk_release,
1055 .connect = sock_no_connect,
1056 .socketpair = sock_no_socketpair,
1057 .accept = sock_no_accept,
1058 .getname = sock_no_getname,
1060 .ioctl = sock_no_ioctl,
1061 .listen = sock_no_listen,
1062 .shutdown = sock_no_shutdown,
1063 .setsockopt = xsk_setsockopt,
1064 .getsockopt = xsk_getsockopt,
1065 .sendmsg = xsk_sendmsg,
1066 .recvmsg = sock_no_recvmsg,
1068 .sendpage = sock_no_sendpage,
1071 static void xsk_destruct(struct sock *sk)
1073 struct xdp_sock *xs = xdp_sk(sk);
1075 if (!sock_flag(sk, SOCK_DEAD))
1078 xdp_put_umem(xs->umem);
1080 sk_refcnt_debug_dec(sk);
1083 static int xsk_create(struct net *net, struct socket *sock, int protocol,
1087 struct xdp_sock *xs;
1089 if (!ns_capable(net->user_ns, CAP_NET_RAW))
1091 if (sock->type != SOCK_RAW)
1092 return -ESOCKTNOSUPPORT;
1095 return -EPROTONOSUPPORT;
1097 sock->state = SS_UNCONNECTED;
1099 sk = sk_alloc(net, PF_XDP, GFP_KERNEL, &xsk_proto, kern);
1103 sock->ops = &xsk_proto_ops;
1105 sock_init_data(sock, sk);
1107 sk->sk_family = PF_XDP;
1109 sk->sk_destruct = xsk_destruct;
1110 sk_refcnt_debug_inc(sk);
1112 sock_set_flag(sk, SOCK_RCU_FREE);
1115 xs->state = XSK_READY;
1116 mutex_init(&xs->mutex);
1117 spin_lock_init(&xs->rx_lock);
1118 spin_lock_init(&xs->tx_completion_lock);
1120 INIT_LIST_HEAD(&xs->map_list);
1121 spin_lock_init(&xs->map_list_lock);
1123 mutex_lock(&net->xdp.lock);
1124 sk_add_node_rcu(sk, &net->xdp.list);
1125 mutex_unlock(&net->xdp.lock);
1128 sock_prot_inuse_add(net, &xsk_proto, 1);
1134 static const struct net_proto_family xsk_family_ops = {
1136 .create = xsk_create,
1137 .owner = THIS_MODULE,
1140 static struct notifier_block xsk_netdev_notifier = {
1141 .notifier_call = xsk_notifier,
1144 static int __net_init xsk_net_init(struct net *net)
1146 mutex_init(&net->xdp.lock);
1147 INIT_HLIST_HEAD(&net->xdp.list);
1151 static void __net_exit xsk_net_exit(struct net *net)
1153 WARN_ON_ONCE(!hlist_empty(&net->xdp.list));
1156 static struct pernet_operations xsk_net_ops = {
1157 .init = xsk_net_init,
1158 .exit = xsk_net_exit,
1161 static int __init xsk_init(void)
1165 err = proto_register(&xsk_proto, 0 /* no slab */);
1169 err = sock_register(&xsk_family_ops);
1173 err = register_pernet_subsys(&xsk_net_ops);
1177 err = register_netdevice_notifier(&xsk_netdev_notifier);
1184 unregister_pernet_subsys(&xsk_net_ops);
1186 sock_unregister(PF_XDP);
1188 proto_unregister(&xsk_proto);
1193 fs_initcall(xsk_init);