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_drv.h>
28 #include "xsk_queue.h"
32 #define TX_BATCH_SIZE 16
34 static DEFINE_PER_CPU(struct list_head, xskmap_flush_list);
36 bool xsk_is_setup_for_bpf_map(struct xdp_sock *xs)
38 return READ_ONCE(xs->rx) && READ_ONCE(xs->umem) &&
39 READ_ONCE(xs->umem->fq);
42 void xsk_set_rx_need_wakeup(struct xdp_umem *umem)
44 if (umem->need_wakeup & XDP_WAKEUP_RX)
47 umem->fq->ring->flags |= XDP_RING_NEED_WAKEUP;
48 umem->need_wakeup |= XDP_WAKEUP_RX;
50 EXPORT_SYMBOL(xsk_set_rx_need_wakeup);
52 void xsk_set_tx_need_wakeup(struct xdp_umem *umem)
56 if (umem->need_wakeup & XDP_WAKEUP_TX)
60 list_for_each_entry_rcu(xs, &umem->xsk_tx_list, list) {
61 xs->tx->ring->flags |= XDP_RING_NEED_WAKEUP;
65 umem->need_wakeup |= XDP_WAKEUP_TX;
67 EXPORT_SYMBOL(xsk_set_tx_need_wakeup);
69 void xsk_clear_rx_need_wakeup(struct xdp_umem *umem)
71 if (!(umem->need_wakeup & XDP_WAKEUP_RX))
74 umem->fq->ring->flags &= ~XDP_RING_NEED_WAKEUP;
75 umem->need_wakeup &= ~XDP_WAKEUP_RX;
77 EXPORT_SYMBOL(xsk_clear_rx_need_wakeup);
79 void xsk_clear_tx_need_wakeup(struct xdp_umem *umem)
83 if (!(umem->need_wakeup & XDP_WAKEUP_TX))
87 list_for_each_entry_rcu(xs, &umem->xsk_tx_list, list) {
88 xs->tx->ring->flags &= ~XDP_RING_NEED_WAKEUP;
92 umem->need_wakeup &= ~XDP_WAKEUP_TX;
94 EXPORT_SYMBOL(xsk_clear_tx_need_wakeup);
96 bool xsk_umem_uses_need_wakeup(struct xdp_umem *umem)
98 return umem->flags & XDP_UMEM_USES_NEED_WAKEUP;
100 EXPORT_SYMBOL(xsk_umem_uses_need_wakeup);
102 void xp_release(struct xdp_buff_xsk *xskb)
104 xskb->pool->free_heads[xskb->pool->free_heads_cnt++] = xskb;
107 static u64 xp_get_handle(struct xdp_buff_xsk *xskb)
109 u64 offset = xskb->xdp.data - xskb->xdp.data_hard_start;
111 offset += xskb->pool->headroom;
112 if (!xskb->pool->unaligned)
113 return xskb->orig_addr + offset;
114 return xskb->orig_addr + (offset << XSK_UNALIGNED_BUF_OFFSET_SHIFT);
117 static int __xsk_rcv_zc(struct xdp_sock *xs, struct xdp_buff *xdp, u32 len)
119 struct xdp_buff_xsk *xskb = container_of(xdp, struct xdp_buff_xsk, xdp);
123 addr = xp_get_handle(xskb);
124 err = xskq_prod_reserve_desc(xs->rx, addr, len);
134 static void xsk_copy_xdp(struct xdp_buff *to, struct xdp_buff *from, u32 len)
136 void *from_buf, *to_buf;
139 if (unlikely(xdp_data_meta_unsupported(from))) {
140 from_buf = from->data;
144 from_buf = from->data_meta;
145 metalen = from->data - from->data_meta;
146 to_buf = to->data - metalen;
149 memcpy(to_buf, from_buf, len + metalen);
152 static int __xsk_rcv(struct xdp_sock *xs, struct xdp_buff *xdp, u32 len,
155 struct xdp_buff *xsk_xdp;
158 if (len > xsk_umem_get_rx_frame_size(xs->umem)) {
163 xsk_xdp = xsk_buff_alloc(xs->umem);
169 xsk_copy_xdp(xsk_xdp, xdp, len);
170 err = __xsk_rcv_zc(xs, xsk_xdp, len);
172 xsk_buff_free(xsk_xdp);
176 xdp_return_buff(xdp);
180 static bool xsk_is_bound(struct xdp_sock *xs)
182 if (READ_ONCE(xs->state) == XSK_BOUND) {
183 /* Matches smp_wmb() in bind(). */
190 static int xsk_rcv(struct xdp_sock *xs, struct xdp_buff *xdp,
195 if (!xsk_is_bound(xs))
198 if (xs->dev != xdp->rxq->dev || xs->queue_id != xdp->rxq->queue_index)
201 len = xdp->data_end - xdp->data;
203 return xdp->rxq->mem.type == MEM_TYPE_XSK_BUFF_POOL ?
204 __xsk_rcv_zc(xs, xdp, len) :
205 __xsk_rcv(xs, xdp, len, explicit_free);
208 static void xsk_flush(struct xdp_sock *xs)
210 xskq_prod_submit(xs->rx);
211 __xskq_cons_release(xs->umem->fq);
212 sock_def_readable(&xs->sk);
215 int xsk_generic_rcv(struct xdp_sock *xs, struct xdp_buff *xdp)
219 spin_lock_bh(&xs->rx_lock);
220 err = xsk_rcv(xs, xdp, false);
222 spin_unlock_bh(&xs->rx_lock);
226 int __xsk_map_redirect(struct xdp_sock *xs, struct xdp_buff *xdp)
228 struct list_head *flush_list = this_cpu_ptr(&xskmap_flush_list);
231 err = xsk_rcv(xs, xdp, true);
235 if (!xs->flush_node.prev)
236 list_add(&xs->flush_node, flush_list);
241 void __xsk_map_flush(void)
243 struct list_head *flush_list = this_cpu_ptr(&xskmap_flush_list);
244 struct xdp_sock *xs, *tmp;
246 list_for_each_entry_safe(xs, tmp, flush_list, flush_node) {
248 __list_del_clearprev(&xs->flush_node);
252 void xsk_umem_complete_tx(struct xdp_umem *umem, u32 nb_entries)
254 xskq_prod_submit_n(umem->cq, nb_entries);
256 EXPORT_SYMBOL(xsk_umem_complete_tx);
258 void xsk_umem_consume_tx_done(struct xdp_umem *umem)
263 list_for_each_entry_rcu(xs, &umem->xsk_tx_list, list) {
264 __xskq_cons_release(xs->tx);
265 xs->sk.sk_write_space(&xs->sk);
269 EXPORT_SYMBOL(xsk_umem_consume_tx_done);
271 bool xsk_umem_consume_tx(struct xdp_umem *umem, struct xdp_desc *desc)
276 list_for_each_entry_rcu(xs, &umem->xsk_tx_list, list) {
277 if (!xskq_cons_peek_desc(xs->tx, desc, umem))
280 /* This is the backpressure mechanism for the Tx path.
281 * Reserve space in the completion queue and only proceed
282 * if there is space in it. This avoids having to implement
283 * any buffering in the Tx path.
285 if (xskq_prod_reserve_addr(umem->cq, desc->addr))
288 xskq_cons_release(xs->tx);
297 EXPORT_SYMBOL(xsk_umem_consume_tx);
299 static int xsk_wakeup(struct xdp_sock *xs, u8 flags)
301 struct net_device *dev = xs->dev;
305 err = dev->netdev_ops->ndo_xsk_wakeup(dev, xs->queue_id, flags);
311 static int xsk_zc_xmit(struct xdp_sock *xs)
313 return xsk_wakeup(xs, XDP_WAKEUP_TX);
316 static void xsk_destruct_skb(struct sk_buff *skb)
318 u64 addr = (u64)(long)skb_shinfo(skb)->destructor_arg;
319 struct xdp_sock *xs = xdp_sk(skb->sk);
322 spin_lock_irqsave(&xs->tx_completion_lock, flags);
323 xskq_prod_submit_addr(xs->umem->cq, addr);
324 spin_unlock_irqrestore(&xs->tx_completion_lock, flags);
329 static int xsk_generic_xmit(struct sock *sk)
331 struct xdp_sock *xs = xdp_sk(sk);
332 u32 max_batch = TX_BATCH_SIZE;
333 bool sent_frame = false;
334 struct xdp_desc desc;
338 mutex_lock(&xs->mutex);
340 if (xs->queue_id >= xs->dev->real_num_tx_queues)
343 while (xskq_cons_peek_desc(xs->tx, &desc, xs->umem)) {
348 if (max_batch-- == 0) {
354 skb = sock_alloc_send_skb(sk, len, 1, &err);
355 if (unlikely(!skb)) {
362 buffer = xsk_buff_raw_get_data(xs->umem, addr);
363 err = skb_store_bits(skb, 0, buffer, len);
364 /* This is the backpressure mechanism for the Tx path.
365 * Reserve space in the completion queue and only proceed
366 * if there is space in it. This avoids having to implement
367 * any buffering in the Tx path.
369 if (unlikely(err) || xskq_prod_reserve(xs->umem->cq)) {
375 skb->priority = sk->sk_priority;
376 skb->mark = sk->sk_mark;
377 skb_shinfo(skb)->destructor_arg = (void *)(long)desc.addr;
378 skb->destructor = xsk_destruct_skb;
380 err = dev_direct_xmit(skb, xs->queue_id);
381 xskq_cons_release(xs->tx);
382 /* Ignore NET_XMIT_CN as packet might have been sent */
383 if (err == NET_XMIT_DROP || err == NETDEV_TX_BUSY) {
384 /* SKB completed but not sent */
394 sk->sk_write_space(sk);
396 mutex_unlock(&xs->mutex);
400 static int __xsk_sendmsg(struct sock *sk)
402 struct xdp_sock *xs = xdp_sk(sk);
404 if (unlikely(!(xs->dev->flags & IFF_UP)))
406 if (unlikely(!xs->tx))
409 return xs->zc ? xsk_zc_xmit(xs) : xsk_generic_xmit(sk);
412 static int xsk_sendmsg(struct socket *sock, struct msghdr *m, size_t total_len)
414 bool need_wait = !(m->msg_flags & MSG_DONTWAIT);
415 struct sock *sk = sock->sk;
416 struct xdp_sock *xs = xdp_sk(sk);
418 if (unlikely(!xsk_is_bound(xs)))
420 if (unlikely(need_wait))
423 return __xsk_sendmsg(sk);
426 static __poll_t xsk_poll(struct file *file, struct socket *sock,
427 struct poll_table_struct *wait)
429 __poll_t mask = datagram_poll(file, sock, wait);
430 struct sock *sk = sock->sk;
431 struct xdp_sock *xs = xdp_sk(sk);
432 struct xdp_umem *umem;
434 if (unlikely(!xsk_is_bound(xs)))
439 if (umem->need_wakeup) {
441 xsk_wakeup(xs, umem->need_wakeup);
443 /* Poll needs to drive Tx also in copy mode */
447 if (xs->rx && !xskq_prod_is_empty(xs->rx))
448 mask |= EPOLLIN | EPOLLRDNORM;
449 if (xs->tx && !xskq_cons_is_full(xs->tx))
450 mask |= EPOLLOUT | EPOLLWRNORM;
455 static int xsk_init_queue(u32 entries, struct xsk_queue **queue,
460 if (entries == 0 || *queue || !is_power_of_2(entries))
463 q = xskq_create(entries, umem_queue);
467 /* Make sure queue is ready before it can be seen by others */
469 WRITE_ONCE(*queue, q);
473 static void xsk_unbind_dev(struct xdp_sock *xs)
475 struct net_device *dev = xs->dev;
477 if (xs->state != XSK_BOUND)
479 WRITE_ONCE(xs->state, XSK_UNBOUND);
481 /* Wait for driver to stop using the xdp socket. */
482 xdp_del_sk_umem(xs->umem, xs);
488 static struct xsk_map *xsk_get_map_list_entry(struct xdp_sock *xs,
489 struct xdp_sock ***map_entry)
491 struct xsk_map *map = NULL;
492 struct xsk_map_node *node;
496 spin_lock_bh(&xs->map_list_lock);
497 node = list_first_entry_or_null(&xs->map_list, struct xsk_map_node,
500 WARN_ON(xsk_map_inc(node->map));
502 *map_entry = node->map_entry;
504 spin_unlock_bh(&xs->map_list_lock);
508 static void xsk_delete_from_maps(struct xdp_sock *xs)
510 /* This function removes the current XDP socket from all the
511 * maps it resides in. We need to take extra care here, due to
512 * the two locks involved. Each map has a lock synchronizing
513 * updates to the entries, and each socket has a lock that
514 * synchronizes access to the list of maps (map_list). For
515 * deadlock avoidance the locks need to be taken in the order
516 * "map lock"->"socket map list lock". We start off by
517 * accessing the socket map list, and take a reference to the
518 * map to guarantee existence between the
519 * xsk_get_map_list_entry() and xsk_map_try_sock_delete()
520 * calls. Then we ask the map to remove the socket, which
521 * tries to remove the socket from the map. Note that there
522 * might be updates to the map between
523 * xsk_get_map_list_entry() and xsk_map_try_sock_delete().
525 struct xdp_sock **map_entry = NULL;
528 while ((map = xsk_get_map_list_entry(xs, &map_entry))) {
529 xsk_map_try_sock_delete(map, xs, map_entry);
534 static int xsk_release(struct socket *sock)
536 struct sock *sk = sock->sk;
537 struct xdp_sock *xs = xdp_sk(sk);
545 mutex_lock(&net->xdp.lock);
546 sk_del_node_init_rcu(sk);
547 mutex_unlock(&net->xdp.lock);
550 sock_prot_inuse_add(net, sk->sk_prot, -1);
553 xsk_delete_from_maps(xs);
554 mutex_lock(&xs->mutex);
556 mutex_unlock(&xs->mutex);
558 xskq_destroy(xs->rx);
559 xskq_destroy(xs->tx);
564 sk_refcnt_debug_release(sk);
570 static struct socket *xsk_lookup_xsk_from_fd(int fd)
575 sock = sockfd_lookup(fd, &err);
577 return ERR_PTR(-ENOTSOCK);
579 if (sock->sk->sk_family != PF_XDP) {
581 return ERR_PTR(-ENOPROTOOPT);
587 static int xsk_bind(struct socket *sock, struct sockaddr *addr, int addr_len)
589 struct sockaddr_xdp *sxdp = (struct sockaddr_xdp *)addr;
590 struct sock *sk = sock->sk;
591 struct xdp_sock *xs = xdp_sk(sk);
592 struct net_device *dev;
596 if (addr_len < sizeof(struct sockaddr_xdp))
598 if (sxdp->sxdp_family != AF_XDP)
601 flags = sxdp->sxdp_flags;
602 if (flags & ~(XDP_SHARED_UMEM | XDP_COPY | XDP_ZEROCOPY |
603 XDP_USE_NEED_WAKEUP))
607 mutex_lock(&xs->mutex);
608 if (xs->state != XSK_READY) {
613 dev = dev_get_by_index(sock_net(sk), sxdp->sxdp_ifindex);
619 if (!xs->rx && !xs->tx) {
624 qid = sxdp->sxdp_queue_id;
626 if (flags & XDP_SHARED_UMEM) {
627 struct xdp_sock *umem_xs;
630 if ((flags & XDP_COPY) || (flags & XDP_ZEROCOPY) ||
631 (flags & XDP_USE_NEED_WAKEUP)) {
632 /* Cannot specify flags for shared sockets. */
638 /* We have already our own. */
643 sock = xsk_lookup_xsk_from_fd(sxdp->sxdp_shared_umem_fd);
649 umem_xs = xdp_sk(sock->sk);
650 if (!xsk_is_bound(umem_xs)) {
655 if (umem_xs->dev != dev || umem_xs->queue_id != qid) {
661 xdp_get_umem(umem_xs->umem);
662 WRITE_ONCE(xs->umem, umem_xs->umem);
664 } else if (!xs->umem || !xdp_umem_validate_queues(xs->umem)) {
668 /* This xsk has its own umem. */
669 err = xdp_umem_assign_dev(xs->umem, dev, qid, flags);
675 xs->zc = xs->umem->zc;
677 xdp_add_sk_umem(xs->umem, xs);
683 /* Matches smp_rmb() in bind() for shared umem
684 * sockets, and xsk_is_bound().
687 WRITE_ONCE(xs->state, XSK_BOUND);
690 mutex_unlock(&xs->mutex);
695 struct xdp_umem_reg_v1 {
696 __u64 addr; /* Start of packet data area */
697 __u64 len; /* Length of packet data area */
702 static int xsk_setsockopt(struct socket *sock, int level, int optname,
703 char __user *optval, unsigned int optlen)
705 struct sock *sk = sock->sk;
706 struct xdp_sock *xs = xdp_sk(sk);
709 if (level != SOL_XDP)
716 struct xsk_queue **q;
719 if (optlen < sizeof(entries))
721 if (copy_from_user(&entries, optval, sizeof(entries)))
724 mutex_lock(&xs->mutex);
725 if (xs->state != XSK_READY) {
726 mutex_unlock(&xs->mutex);
729 q = (optname == XDP_TX_RING) ? &xs->tx : &xs->rx;
730 err = xsk_init_queue(entries, q, false);
731 if (!err && optname == XDP_TX_RING)
732 /* Tx needs to be explicitly woken up the first time */
733 xs->tx->ring->flags |= XDP_RING_NEED_WAKEUP;
734 mutex_unlock(&xs->mutex);
739 size_t mr_size = sizeof(struct xdp_umem_reg);
740 struct xdp_umem_reg mr = {};
741 struct xdp_umem *umem;
743 if (optlen < sizeof(struct xdp_umem_reg_v1))
745 else if (optlen < sizeof(mr))
746 mr_size = sizeof(struct xdp_umem_reg_v1);
748 if (copy_from_user(&mr, optval, mr_size))
751 mutex_lock(&xs->mutex);
752 if (xs->state != XSK_READY || xs->umem) {
753 mutex_unlock(&xs->mutex);
757 umem = xdp_umem_create(&mr);
759 mutex_unlock(&xs->mutex);
760 return PTR_ERR(umem);
763 /* Make sure umem is ready before it can be seen by others */
765 WRITE_ONCE(xs->umem, umem);
766 mutex_unlock(&xs->mutex);
769 case XDP_UMEM_FILL_RING:
770 case XDP_UMEM_COMPLETION_RING:
772 struct xsk_queue **q;
775 if (copy_from_user(&entries, optval, sizeof(entries)))
778 mutex_lock(&xs->mutex);
779 if (xs->state != XSK_READY) {
780 mutex_unlock(&xs->mutex);
784 mutex_unlock(&xs->mutex);
788 q = (optname == XDP_UMEM_FILL_RING) ? &xs->umem->fq :
790 err = xsk_init_queue(entries, q, true);
791 if (optname == XDP_UMEM_FILL_RING)
792 xp_set_fq(xs->umem->pool, *q);
793 mutex_unlock(&xs->mutex);
803 static void xsk_enter_rxtx_offsets(struct xdp_ring_offset_v1 *ring)
805 ring->producer = offsetof(struct xdp_rxtx_ring, ptrs.producer);
806 ring->consumer = offsetof(struct xdp_rxtx_ring, ptrs.consumer);
807 ring->desc = offsetof(struct xdp_rxtx_ring, desc);
810 static void xsk_enter_umem_offsets(struct xdp_ring_offset_v1 *ring)
812 ring->producer = offsetof(struct xdp_umem_ring, ptrs.producer);
813 ring->consumer = offsetof(struct xdp_umem_ring, ptrs.consumer);
814 ring->desc = offsetof(struct xdp_umem_ring, desc);
817 static int xsk_getsockopt(struct socket *sock, int level, int optname,
818 char __user *optval, int __user *optlen)
820 struct sock *sk = sock->sk;
821 struct xdp_sock *xs = xdp_sk(sk);
824 if (level != SOL_XDP)
827 if (get_user(len, optlen))
835 struct xdp_statistics stats;
837 if (len < sizeof(stats))
840 mutex_lock(&xs->mutex);
841 stats.rx_dropped = xs->rx_dropped;
842 stats.rx_invalid_descs = xskq_nb_invalid_descs(xs->rx);
843 stats.tx_invalid_descs = xskq_nb_invalid_descs(xs->tx);
844 mutex_unlock(&xs->mutex);
846 if (copy_to_user(optval, &stats, sizeof(stats)))
848 if (put_user(sizeof(stats), optlen))
853 case XDP_MMAP_OFFSETS:
855 struct xdp_mmap_offsets off;
856 struct xdp_mmap_offsets_v1 off_v1;
857 bool flags_supported = true;
860 if (len < sizeof(off_v1))
862 else if (len < sizeof(off))
863 flags_supported = false;
865 if (flags_supported) {
866 /* xdp_ring_offset is identical to xdp_ring_offset_v1
867 * except for the flags field added to the end.
869 xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *)
871 xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *)
873 xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *)
875 xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *)
877 off.rx.flags = offsetof(struct xdp_rxtx_ring,
879 off.tx.flags = offsetof(struct xdp_rxtx_ring,
881 off.fr.flags = offsetof(struct xdp_umem_ring,
883 off.cr.flags = offsetof(struct xdp_umem_ring,
889 xsk_enter_rxtx_offsets(&off_v1.rx);
890 xsk_enter_rxtx_offsets(&off_v1.tx);
891 xsk_enter_umem_offsets(&off_v1.fr);
892 xsk_enter_umem_offsets(&off_v1.cr);
894 len = sizeof(off_v1);
898 if (copy_to_user(optval, to_copy, len))
900 if (put_user(len, optlen))
907 struct xdp_options opts = {};
909 if (len < sizeof(opts))
912 mutex_lock(&xs->mutex);
914 opts.flags |= XDP_OPTIONS_ZEROCOPY;
915 mutex_unlock(&xs->mutex);
918 if (copy_to_user(optval, &opts, len))
920 if (put_user(len, optlen))
932 static int xsk_mmap(struct file *file, struct socket *sock,
933 struct vm_area_struct *vma)
935 loff_t offset = (loff_t)vma->vm_pgoff << PAGE_SHIFT;
936 unsigned long size = vma->vm_end - vma->vm_start;
937 struct xdp_sock *xs = xdp_sk(sock->sk);
938 struct xsk_queue *q = NULL;
939 struct xdp_umem *umem;
943 if (READ_ONCE(xs->state) != XSK_READY)
946 if (offset == XDP_PGOFF_RX_RING) {
947 q = READ_ONCE(xs->rx);
948 } else if (offset == XDP_PGOFF_TX_RING) {
949 q = READ_ONCE(xs->tx);
951 umem = READ_ONCE(xs->umem);
955 /* Matches the smp_wmb() in XDP_UMEM_REG */
957 if (offset == XDP_UMEM_PGOFF_FILL_RING)
958 q = READ_ONCE(umem->fq);
959 else if (offset == XDP_UMEM_PGOFF_COMPLETION_RING)
960 q = READ_ONCE(umem->cq);
966 /* Matches the smp_wmb() in xsk_init_queue */
968 qpg = virt_to_head_page(q->ring);
969 if (size > page_size(qpg))
972 pfn = virt_to_phys(q->ring) >> PAGE_SHIFT;
973 return remap_pfn_range(vma, vma->vm_start, pfn,
974 size, vma->vm_page_prot);
977 static int xsk_notifier(struct notifier_block *this,
978 unsigned long msg, void *ptr)
980 struct net_device *dev = netdev_notifier_info_to_dev(ptr);
981 struct net *net = dev_net(dev);
985 case NETDEV_UNREGISTER:
986 mutex_lock(&net->xdp.lock);
987 sk_for_each(sk, &net->xdp.list) {
988 struct xdp_sock *xs = xdp_sk(sk);
990 mutex_lock(&xs->mutex);
991 if (xs->dev == dev) {
992 sk->sk_err = ENETDOWN;
993 if (!sock_flag(sk, SOCK_DEAD))
994 sk->sk_error_report(sk);
998 /* Clear device references in umem. */
999 xdp_umem_clear_dev(xs->umem);
1001 mutex_unlock(&xs->mutex);
1003 mutex_unlock(&net->xdp.lock);
1009 static struct proto xsk_proto = {
1011 .owner = THIS_MODULE,
1012 .obj_size = sizeof(struct xdp_sock),
1015 static const struct proto_ops xsk_proto_ops = {
1017 .owner = THIS_MODULE,
1018 .release = xsk_release,
1020 .connect = sock_no_connect,
1021 .socketpair = sock_no_socketpair,
1022 .accept = sock_no_accept,
1023 .getname = sock_no_getname,
1025 .ioctl = sock_no_ioctl,
1026 .listen = sock_no_listen,
1027 .shutdown = sock_no_shutdown,
1028 .setsockopt = xsk_setsockopt,
1029 .getsockopt = xsk_getsockopt,
1030 .sendmsg = xsk_sendmsg,
1031 .recvmsg = sock_no_recvmsg,
1033 .sendpage = sock_no_sendpage,
1036 static void xsk_destruct(struct sock *sk)
1038 struct xdp_sock *xs = xdp_sk(sk);
1040 if (!sock_flag(sk, SOCK_DEAD))
1043 xdp_put_umem(xs->umem);
1045 sk_refcnt_debug_dec(sk);
1048 static int xsk_create(struct net *net, struct socket *sock, int protocol,
1052 struct xdp_sock *xs;
1054 if (!ns_capable(net->user_ns, CAP_NET_RAW))
1056 if (sock->type != SOCK_RAW)
1057 return -ESOCKTNOSUPPORT;
1060 return -EPROTONOSUPPORT;
1062 sock->state = SS_UNCONNECTED;
1064 sk = sk_alloc(net, PF_XDP, GFP_KERNEL, &xsk_proto, kern);
1068 sock->ops = &xsk_proto_ops;
1070 sock_init_data(sock, sk);
1072 sk->sk_family = PF_XDP;
1074 sk->sk_destruct = xsk_destruct;
1075 sk_refcnt_debug_inc(sk);
1077 sock_set_flag(sk, SOCK_RCU_FREE);
1080 xs->state = XSK_READY;
1081 mutex_init(&xs->mutex);
1082 spin_lock_init(&xs->rx_lock);
1083 spin_lock_init(&xs->tx_completion_lock);
1085 INIT_LIST_HEAD(&xs->map_list);
1086 spin_lock_init(&xs->map_list_lock);
1088 mutex_lock(&net->xdp.lock);
1089 sk_add_node_rcu(sk, &net->xdp.list);
1090 mutex_unlock(&net->xdp.lock);
1093 sock_prot_inuse_add(net, &xsk_proto, 1);
1099 static const struct net_proto_family xsk_family_ops = {
1101 .create = xsk_create,
1102 .owner = THIS_MODULE,
1105 static struct notifier_block xsk_netdev_notifier = {
1106 .notifier_call = xsk_notifier,
1109 static int __net_init xsk_net_init(struct net *net)
1111 mutex_init(&net->xdp.lock);
1112 INIT_HLIST_HEAD(&net->xdp.list);
1116 static void __net_exit xsk_net_exit(struct net *net)
1118 WARN_ON_ONCE(!hlist_empty(&net->xdp.list));
1121 static struct pernet_operations xsk_net_ops = {
1122 .init = xsk_net_init,
1123 .exit = xsk_net_exit,
1126 static int __init xsk_init(void)
1130 err = proto_register(&xsk_proto, 0 /* no slab */);
1134 err = sock_register(&xsk_family_ops);
1138 err = register_pernet_subsys(&xsk_net_ops);
1142 err = register_netdevice_notifier(&xsk_netdev_notifier);
1146 for_each_possible_cpu(cpu)
1147 INIT_LIST_HEAD(&per_cpu(xskmap_flush_list, cpu));
1151 unregister_pernet_subsys(&xsk_net_ops);
1153 sock_unregister(PF_XDP);
1155 proto_unregister(&xsk_proto);
1160 fs_initcall(xsk_init);