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.
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)) {
278 xs->tx->queue_empty_descs++;
282 /* This is the backpressure mechanism for the Tx path.
283 * Reserve space in the completion queue and only proceed
284 * if there is space in it. This avoids having to implement
285 * any buffering in the Tx path.
287 if (xskq_prod_reserve_addr(umem->cq, desc->addr))
290 xskq_cons_release(xs->tx);
299 EXPORT_SYMBOL(xsk_umem_consume_tx);
301 static int xsk_wakeup(struct xdp_sock *xs, u8 flags)
303 struct net_device *dev = xs->dev;
307 err = dev->netdev_ops->ndo_xsk_wakeup(dev, xs->queue_id, flags);
313 static int xsk_zc_xmit(struct xdp_sock *xs)
315 return xsk_wakeup(xs, XDP_WAKEUP_TX);
318 static void xsk_destruct_skb(struct sk_buff *skb)
320 u64 addr = (u64)(long)skb_shinfo(skb)->destructor_arg;
321 struct xdp_sock *xs = xdp_sk(skb->sk);
324 spin_lock_irqsave(&xs->tx_completion_lock, flags);
325 xskq_prod_submit_addr(xs->umem->cq, addr);
326 spin_unlock_irqrestore(&xs->tx_completion_lock, flags);
331 static int xsk_generic_xmit(struct sock *sk)
333 struct xdp_sock *xs = xdp_sk(sk);
334 u32 max_batch = TX_BATCH_SIZE;
335 bool sent_frame = false;
336 struct xdp_desc desc;
340 mutex_lock(&xs->mutex);
342 if (xs->queue_id >= xs->dev->real_num_tx_queues)
345 while (xskq_cons_peek_desc(xs->tx, &desc, xs->umem)) {
350 if (max_batch-- == 0) {
356 skb = sock_alloc_send_skb(sk, len, 1, &err);
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 */
392 xs->tx->queue_empty_descs++;
396 sk->sk_write_space(sk);
398 mutex_unlock(&xs->mutex);
402 static int __xsk_sendmsg(struct sock *sk)
404 struct xdp_sock *xs = xdp_sk(sk);
406 if (unlikely(!(xs->dev->flags & IFF_UP)))
408 if (unlikely(!xs->tx))
411 return xs->zc ? xsk_zc_xmit(xs) : xsk_generic_xmit(sk);
414 static int xsk_sendmsg(struct socket *sock, struct msghdr *m, size_t total_len)
416 bool need_wait = !(m->msg_flags & MSG_DONTWAIT);
417 struct sock *sk = sock->sk;
418 struct xdp_sock *xs = xdp_sk(sk);
420 if (unlikely(!xsk_is_bound(xs)))
422 if (unlikely(need_wait))
425 return __xsk_sendmsg(sk);
428 static __poll_t xsk_poll(struct file *file, struct socket *sock,
429 struct poll_table_struct *wait)
431 __poll_t mask = datagram_poll(file, sock, wait);
432 struct sock *sk = sock->sk;
433 struct xdp_sock *xs = xdp_sk(sk);
434 struct xdp_umem *umem;
436 if (unlikely(!xsk_is_bound(xs)))
441 if (umem->need_wakeup) {
443 xsk_wakeup(xs, umem->need_wakeup);
445 /* Poll needs to drive Tx also in copy mode */
449 if (xs->rx && !xskq_prod_is_empty(xs->rx))
450 mask |= EPOLLIN | EPOLLRDNORM;
451 if (xs->tx && !xskq_cons_is_full(xs->tx))
452 mask |= EPOLLOUT | EPOLLWRNORM;
457 static int xsk_init_queue(u32 entries, struct xsk_queue **queue,
462 if (entries == 0 || *queue || !is_power_of_2(entries))
465 q = xskq_create(entries, umem_queue);
469 /* Make sure queue is ready before it can be seen by others */
471 WRITE_ONCE(*queue, q);
475 static void xsk_unbind_dev(struct xdp_sock *xs)
477 struct net_device *dev = xs->dev;
479 if (xs->state != XSK_BOUND)
481 WRITE_ONCE(xs->state, XSK_UNBOUND);
483 /* Wait for driver to stop using the xdp socket. */
484 xdp_del_sk_umem(xs->umem, xs);
490 static struct xsk_map *xsk_get_map_list_entry(struct xdp_sock *xs,
491 struct xdp_sock ***map_entry)
493 struct xsk_map *map = NULL;
494 struct xsk_map_node *node;
498 spin_lock_bh(&xs->map_list_lock);
499 node = list_first_entry_or_null(&xs->map_list, struct xsk_map_node,
502 WARN_ON(xsk_map_inc(node->map));
504 *map_entry = node->map_entry;
506 spin_unlock_bh(&xs->map_list_lock);
510 static void xsk_delete_from_maps(struct xdp_sock *xs)
512 /* This function removes the current XDP socket from all the
513 * maps it resides in. We need to take extra care here, due to
514 * the two locks involved. Each map has a lock synchronizing
515 * updates to the entries, and each socket has a lock that
516 * synchronizes access to the list of maps (map_list). For
517 * deadlock avoidance the locks need to be taken in the order
518 * "map lock"->"socket map list lock". We start off by
519 * accessing the socket map list, and take a reference to the
520 * map to guarantee existence between the
521 * xsk_get_map_list_entry() and xsk_map_try_sock_delete()
522 * calls. Then we ask the map to remove the socket, which
523 * tries to remove the socket from the map. Note that there
524 * might be updates to the map between
525 * xsk_get_map_list_entry() and xsk_map_try_sock_delete().
527 struct xdp_sock **map_entry = NULL;
530 while ((map = xsk_get_map_list_entry(xs, &map_entry))) {
531 xsk_map_try_sock_delete(map, xs, map_entry);
536 static int xsk_release(struct socket *sock)
538 struct sock *sk = sock->sk;
539 struct xdp_sock *xs = xdp_sk(sk);
547 mutex_lock(&net->xdp.lock);
548 sk_del_node_init_rcu(sk);
549 mutex_unlock(&net->xdp.lock);
552 sock_prot_inuse_add(net, sk->sk_prot, -1);
555 xsk_delete_from_maps(xs);
556 mutex_lock(&xs->mutex);
558 mutex_unlock(&xs->mutex);
560 xskq_destroy(xs->rx);
561 xskq_destroy(xs->tx);
566 sk_refcnt_debug_release(sk);
572 static struct socket *xsk_lookup_xsk_from_fd(int fd)
577 sock = sockfd_lookup(fd, &err);
579 return ERR_PTR(-ENOTSOCK);
581 if (sock->sk->sk_family != PF_XDP) {
583 return ERR_PTR(-ENOPROTOOPT);
589 static int xsk_bind(struct socket *sock, struct sockaddr *addr, int addr_len)
591 struct sockaddr_xdp *sxdp = (struct sockaddr_xdp *)addr;
592 struct sock *sk = sock->sk;
593 struct xdp_sock *xs = xdp_sk(sk);
594 struct net_device *dev;
598 if (addr_len < sizeof(struct sockaddr_xdp))
600 if (sxdp->sxdp_family != AF_XDP)
603 flags = sxdp->sxdp_flags;
604 if (flags & ~(XDP_SHARED_UMEM | XDP_COPY | XDP_ZEROCOPY |
605 XDP_USE_NEED_WAKEUP))
609 mutex_lock(&xs->mutex);
610 if (xs->state != XSK_READY) {
615 dev = dev_get_by_index(sock_net(sk), sxdp->sxdp_ifindex);
621 if (!xs->rx && !xs->tx) {
626 qid = sxdp->sxdp_queue_id;
628 if (flags & XDP_SHARED_UMEM) {
629 struct xdp_sock *umem_xs;
632 if ((flags & XDP_COPY) || (flags & XDP_ZEROCOPY) ||
633 (flags & XDP_USE_NEED_WAKEUP)) {
634 /* Cannot specify flags for shared sockets. */
640 /* We have already our own. */
645 sock = xsk_lookup_xsk_from_fd(sxdp->sxdp_shared_umem_fd);
651 umem_xs = xdp_sk(sock->sk);
652 if (!xsk_is_bound(umem_xs)) {
657 if (umem_xs->dev != dev || umem_xs->queue_id != qid) {
663 xdp_get_umem(umem_xs->umem);
664 WRITE_ONCE(xs->umem, umem_xs->umem);
666 } else if (!xs->umem || !xdp_umem_validate_queues(xs->umem)) {
670 /* This xsk has its own umem. */
671 err = xdp_umem_assign_dev(xs->umem, dev, qid, flags);
677 xs->zc = xs->umem->zc;
679 xdp_add_sk_umem(xs->umem, xs);
685 /* Matches smp_rmb() in bind() for shared umem
686 * sockets, and xsk_is_bound().
689 WRITE_ONCE(xs->state, XSK_BOUND);
692 mutex_unlock(&xs->mutex);
697 struct xdp_umem_reg_v1 {
698 __u64 addr; /* Start of packet data area */
699 __u64 len; /* Length of packet data area */
704 static int xsk_setsockopt(struct socket *sock, int level, int optname,
705 sockptr_t optval, unsigned int optlen)
707 struct sock *sk = sock->sk;
708 struct xdp_sock *xs = xdp_sk(sk);
711 if (level != SOL_XDP)
718 struct xsk_queue **q;
721 if (optlen < sizeof(entries))
723 if (copy_from_sockptr(&entries, optval, sizeof(entries)))
726 mutex_lock(&xs->mutex);
727 if (xs->state != XSK_READY) {
728 mutex_unlock(&xs->mutex);
731 q = (optname == XDP_TX_RING) ? &xs->tx : &xs->rx;
732 err = xsk_init_queue(entries, q, false);
733 if (!err && optname == XDP_TX_RING)
734 /* Tx needs to be explicitly woken up the first time */
735 xs->tx->ring->flags |= XDP_RING_NEED_WAKEUP;
736 mutex_unlock(&xs->mutex);
741 size_t mr_size = sizeof(struct xdp_umem_reg);
742 struct xdp_umem_reg mr = {};
743 struct xdp_umem *umem;
745 if (optlen < sizeof(struct xdp_umem_reg_v1))
747 else if (optlen < sizeof(mr))
748 mr_size = sizeof(struct xdp_umem_reg_v1);
750 if (copy_from_sockptr(&mr, optval, mr_size))
753 mutex_lock(&xs->mutex);
754 if (xs->state != XSK_READY || xs->umem) {
755 mutex_unlock(&xs->mutex);
759 umem = xdp_umem_create(&mr);
761 mutex_unlock(&xs->mutex);
762 return PTR_ERR(umem);
765 /* Make sure umem is ready before it can be seen by others */
767 WRITE_ONCE(xs->umem, umem);
768 mutex_unlock(&xs->mutex);
771 case XDP_UMEM_FILL_RING:
772 case XDP_UMEM_COMPLETION_RING:
774 struct xsk_queue **q;
777 if (copy_from_sockptr(&entries, optval, sizeof(entries)))
780 mutex_lock(&xs->mutex);
781 if (xs->state != XSK_READY) {
782 mutex_unlock(&xs->mutex);
786 mutex_unlock(&xs->mutex);
790 q = (optname == XDP_UMEM_FILL_RING) ? &xs->umem->fq :
792 err = xsk_init_queue(entries, q, true);
793 if (optname == XDP_UMEM_FILL_RING)
794 xp_set_fq(xs->umem->pool, *q);
795 mutex_unlock(&xs->mutex);
805 static void xsk_enter_rxtx_offsets(struct xdp_ring_offset_v1 *ring)
807 ring->producer = offsetof(struct xdp_rxtx_ring, ptrs.producer);
808 ring->consumer = offsetof(struct xdp_rxtx_ring, ptrs.consumer);
809 ring->desc = offsetof(struct xdp_rxtx_ring, desc);
812 static void xsk_enter_umem_offsets(struct xdp_ring_offset_v1 *ring)
814 ring->producer = offsetof(struct xdp_umem_ring, ptrs.producer);
815 ring->consumer = offsetof(struct xdp_umem_ring, ptrs.consumer);
816 ring->desc = offsetof(struct xdp_umem_ring, desc);
819 struct xdp_statistics_v1 {
821 __u64 rx_invalid_descs;
822 __u64 tx_invalid_descs;
825 static int xsk_getsockopt(struct socket *sock, int level, int optname,
826 char __user *optval, int __user *optlen)
828 struct sock *sk = sock->sk;
829 struct xdp_sock *xs = xdp_sk(sk);
832 if (level != SOL_XDP)
835 if (get_user(len, optlen))
843 struct xdp_statistics stats = {};
844 bool extra_stats = true;
847 if (len < sizeof(struct xdp_statistics_v1)) {
849 } else if (len < sizeof(stats)) {
851 stats_size = sizeof(struct xdp_statistics_v1);
853 stats_size = sizeof(stats);
856 mutex_lock(&xs->mutex);
857 stats.rx_dropped = xs->rx_dropped;
859 stats.rx_ring_full = xs->rx_queue_full;
860 stats.rx_fill_ring_empty_descs =
861 xs->umem ? xskq_nb_queue_empty_descs(xs->umem->fq) : 0;
862 stats.tx_ring_empty_descs = xskq_nb_queue_empty_descs(xs->tx);
864 stats.rx_dropped += xs->rx_queue_full;
866 stats.rx_invalid_descs = xskq_nb_invalid_descs(xs->rx);
867 stats.tx_invalid_descs = xskq_nb_invalid_descs(xs->tx);
868 mutex_unlock(&xs->mutex);
870 if (copy_to_user(optval, &stats, stats_size))
872 if (put_user(stats_size, optlen))
877 case XDP_MMAP_OFFSETS:
879 struct xdp_mmap_offsets off;
880 struct xdp_mmap_offsets_v1 off_v1;
881 bool flags_supported = true;
884 if (len < sizeof(off_v1))
886 else if (len < sizeof(off))
887 flags_supported = false;
889 if (flags_supported) {
890 /* xdp_ring_offset is identical to xdp_ring_offset_v1
891 * except for the flags field added to the end.
893 xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *)
895 xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *)
897 xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *)
899 xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *)
901 off.rx.flags = offsetof(struct xdp_rxtx_ring,
903 off.tx.flags = offsetof(struct xdp_rxtx_ring,
905 off.fr.flags = offsetof(struct xdp_umem_ring,
907 off.cr.flags = offsetof(struct xdp_umem_ring,
913 xsk_enter_rxtx_offsets(&off_v1.rx);
914 xsk_enter_rxtx_offsets(&off_v1.tx);
915 xsk_enter_umem_offsets(&off_v1.fr);
916 xsk_enter_umem_offsets(&off_v1.cr);
918 len = sizeof(off_v1);
922 if (copy_to_user(optval, to_copy, len))
924 if (put_user(len, optlen))
931 struct xdp_options opts = {};
933 if (len < sizeof(opts))
936 mutex_lock(&xs->mutex);
938 opts.flags |= XDP_OPTIONS_ZEROCOPY;
939 mutex_unlock(&xs->mutex);
942 if (copy_to_user(optval, &opts, len))
944 if (put_user(len, optlen))
956 static int xsk_mmap(struct file *file, struct socket *sock,
957 struct vm_area_struct *vma)
959 loff_t offset = (loff_t)vma->vm_pgoff << PAGE_SHIFT;
960 unsigned long size = vma->vm_end - vma->vm_start;
961 struct xdp_sock *xs = xdp_sk(sock->sk);
962 struct xsk_queue *q = NULL;
963 struct xdp_umem *umem;
967 if (READ_ONCE(xs->state) != XSK_READY)
970 if (offset == XDP_PGOFF_RX_RING) {
971 q = READ_ONCE(xs->rx);
972 } else if (offset == XDP_PGOFF_TX_RING) {
973 q = READ_ONCE(xs->tx);
975 umem = READ_ONCE(xs->umem);
979 /* Matches the smp_wmb() in XDP_UMEM_REG */
981 if (offset == XDP_UMEM_PGOFF_FILL_RING)
982 q = READ_ONCE(umem->fq);
983 else if (offset == XDP_UMEM_PGOFF_COMPLETION_RING)
984 q = READ_ONCE(umem->cq);
990 /* Matches the smp_wmb() in xsk_init_queue */
992 qpg = virt_to_head_page(q->ring);
993 if (size > page_size(qpg))
996 pfn = virt_to_phys(q->ring) >> PAGE_SHIFT;
997 return remap_pfn_range(vma, vma->vm_start, pfn,
998 size, vma->vm_page_prot);
1001 static int xsk_notifier(struct notifier_block *this,
1002 unsigned long msg, void *ptr)
1004 struct net_device *dev = netdev_notifier_info_to_dev(ptr);
1005 struct net *net = dev_net(dev);
1009 case NETDEV_UNREGISTER:
1010 mutex_lock(&net->xdp.lock);
1011 sk_for_each(sk, &net->xdp.list) {
1012 struct xdp_sock *xs = xdp_sk(sk);
1014 mutex_lock(&xs->mutex);
1015 if (xs->dev == dev) {
1016 sk->sk_err = ENETDOWN;
1017 if (!sock_flag(sk, SOCK_DEAD))
1018 sk->sk_error_report(sk);
1022 /* Clear device references in umem. */
1023 xdp_umem_clear_dev(xs->umem);
1025 mutex_unlock(&xs->mutex);
1027 mutex_unlock(&net->xdp.lock);
1033 static struct proto xsk_proto = {
1035 .owner = THIS_MODULE,
1036 .obj_size = sizeof(struct xdp_sock),
1039 static const struct proto_ops xsk_proto_ops = {
1041 .owner = THIS_MODULE,
1042 .release = xsk_release,
1044 .connect = sock_no_connect,
1045 .socketpair = sock_no_socketpair,
1046 .accept = sock_no_accept,
1047 .getname = sock_no_getname,
1049 .ioctl = sock_no_ioctl,
1050 .listen = sock_no_listen,
1051 .shutdown = sock_no_shutdown,
1052 .setsockopt = xsk_setsockopt,
1053 .getsockopt = xsk_getsockopt,
1054 .sendmsg = xsk_sendmsg,
1055 .recvmsg = sock_no_recvmsg,
1057 .sendpage = sock_no_sendpage,
1060 static void xsk_destruct(struct sock *sk)
1062 struct xdp_sock *xs = xdp_sk(sk);
1064 if (!sock_flag(sk, SOCK_DEAD))
1067 xdp_put_umem(xs->umem);
1069 sk_refcnt_debug_dec(sk);
1072 static int xsk_create(struct net *net, struct socket *sock, int protocol,
1076 struct xdp_sock *xs;
1078 if (!ns_capable(net->user_ns, CAP_NET_RAW))
1080 if (sock->type != SOCK_RAW)
1081 return -ESOCKTNOSUPPORT;
1084 return -EPROTONOSUPPORT;
1086 sock->state = SS_UNCONNECTED;
1088 sk = sk_alloc(net, PF_XDP, GFP_KERNEL, &xsk_proto, kern);
1092 sock->ops = &xsk_proto_ops;
1094 sock_init_data(sock, sk);
1096 sk->sk_family = PF_XDP;
1098 sk->sk_destruct = xsk_destruct;
1099 sk_refcnt_debug_inc(sk);
1101 sock_set_flag(sk, SOCK_RCU_FREE);
1104 xs->state = XSK_READY;
1105 mutex_init(&xs->mutex);
1106 spin_lock_init(&xs->rx_lock);
1107 spin_lock_init(&xs->tx_completion_lock);
1109 INIT_LIST_HEAD(&xs->map_list);
1110 spin_lock_init(&xs->map_list_lock);
1112 mutex_lock(&net->xdp.lock);
1113 sk_add_node_rcu(sk, &net->xdp.list);
1114 mutex_unlock(&net->xdp.lock);
1117 sock_prot_inuse_add(net, &xsk_proto, 1);
1123 static const struct net_proto_family xsk_family_ops = {
1125 .create = xsk_create,
1126 .owner = THIS_MODULE,
1129 static struct notifier_block xsk_netdev_notifier = {
1130 .notifier_call = xsk_notifier,
1133 static int __net_init xsk_net_init(struct net *net)
1135 mutex_init(&net->xdp.lock);
1136 INIT_HLIST_HEAD(&net->xdp.list);
1140 static void __net_exit xsk_net_exit(struct net *net)
1142 WARN_ON_ONCE(!hlist_empty(&net->xdp.list));
1145 static struct pernet_operations xsk_net_ops = {
1146 .init = xsk_net_init,
1147 .exit = xsk_net_exit,
1150 static int __init xsk_init(void)
1154 err = proto_register(&xsk_proto, 0 /* no slab */);
1158 err = sock_register(&xsk_family_ops);
1162 err = register_pernet_subsys(&xsk_net_ops);
1166 err = register_netdevice_notifier(&xsk_netdev_notifier);
1170 for_each_possible_cpu(cpu)
1171 INIT_LIST_HEAD(&per_cpu(xskmap_flush_list, cpu));
1175 unregister_pernet_subsys(&xsk_net_ops);
1177 sock_unregister(PF_XDP);
1179 proto_unregister(&xsk_proto);
1184 fs_initcall(xsk_init);