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;
619 if (addr_len < sizeof(struct sockaddr_xdp))
621 if (sxdp->sxdp_family != AF_XDP)
624 flags = sxdp->sxdp_flags;
625 if (flags & ~(XDP_SHARED_UMEM | XDP_COPY | XDP_ZEROCOPY |
626 XDP_USE_NEED_WAKEUP))
630 mutex_lock(&xs->mutex);
631 if (xs->state != XSK_READY) {
636 dev = dev_get_by_index(sock_net(sk), sxdp->sxdp_ifindex);
642 if (!xs->rx && !xs->tx) {
647 qid = sxdp->sxdp_queue_id;
649 if (flags & XDP_SHARED_UMEM) {
650 struct xdp_sock *umem_xs;
653 if ((flags & XDP_COPY) || (flags & XDP_ZEROCOPY) ||
654 (flags & XDP_USE_NEED_WAKEUP)) {
655 /* Cannot specify flags for shared sockets. */
661 /* We have already our own. */
666 sock = xsk_lookup_xsk_from_fd(sxdp->sxdp_shared_umem_fd);
672 umem_xs = xdp_sk(sock->sk);
673 if (!xsk_is_bound(umem_xs)) {
678 if (umem_xs->dev != dev || umem_xs->queue_id != qid) {
684 xdp_get_umem(umem_xs->umem);
685 WRITE_ONCE(xs->umem, umem_xs->umem);
687 } else if (!xs->umem || !xdp_umem_validate_queues(xs->umem)) {
691 /* This xsk has its own umem. */
692 xskq_set_umem(xs->umem->fq, xs->umem->size,
693 xs->umem->chunk_mask);
694 xskq_set_umem(xs->umem->cq, xs->umem->size,
695 xs->umem->chunk_mask);
697 err = xdp_umem_assign_dev(xs->umem, dev, qid, flags);
701 xsk_check_page_contiguity(xs->umem, flags);
705 xs->zc = xs->umem->zc;
707 xskq_set_umem(xs->rx, xs->umem->size, xs->umem->chunk_mask);
708 xskq_set_umem(xs->tx, xs->umem->size, xs->umem->chunk_mask);
709 xdp_add_sk_umem(xs->umem, xs);
715 /* Matches smp_rmb() in bind() for shared umem
716 * sockets, and xsk_is_bound().
719 WRITE_ONCE(xs->state, XSK_BOUND);
722 mutex_unlock(&xs->mutex);
727 struct xdp_umem_reg_v1 {
728 __u64 addr; /* Start of packet data area */
729 __u64 len; /* Length of packet data area */
734 static int xsk_setsockopt(struct socket *sock, int level, int optname,
735 char __user *optval, unsigned int optlen)
737 struct sock *sk = sock->sk;
738 struct xdp_sock *xs = xdp_sk(sk);
741 if (level != SOL_XDP)
748 struct xsk_queue **q;
751 if (optlen < sizeof(entries))
753 if (copy_from_user(&entries, optval, sizeof(entries)))
756 mutex_lock(&xs->mutex);
757 if (xs->state != XSK_READY) {
758 mutex_unlock(&xs->mutex);
761 q = (optname == XDP_TX_RING) ? &xs->tx : &xs->rx;
762 err = xsk_init_queue(entries, q, false);
763 if (!err && optname == XDP_TX_RING)
764 /* Tx needs to be explicitly woken up the first time */
765 xs->tx->ring->flags |= XDP_RING_NEED_WAKEUP;
766 mutex_unlock(&xs->mutex);
771 size_t mr_size = sizeof(struct xdp_umem_reg);
772 struct xdp_umem_reg mr = {};
773 struct xdp_umem *umem;
775 if (optlen < sizeof(struct xdp_umem_reg_v1))
777 else if (optlen < sizeof(mr))
778 mr_size = sizeof(struct xdp_umem_reg_v1);
780 if (copy_from_user(&mr, optval, mr_size))
783 mutex_lock(&xs->mutex);
784 if (xs->state != XSK_READY || xs->umem) {
785 mutex_unlock(&xs->mutex);
789 umem = xdp_umem_create(&mr);
791 mutex_unlock(&xs->mutex);
792 return PTR_ERR(umem);
795 /* Make sure umem is ready before it can be seen by others */
797 WRITE_ONCE(xs->umem, umem);
798 mutex_unlock(&xs->mutex);
801 case XDP_UMEM_FILL_RING:
802 case XDP_UMEM_COMPLETION_RING:
804 struct xsk_queue **q;
807 if (copy_from_user(&entries, optval, sizeof(entries)))
810 mutex_lock(&xs->mutex);
811 if (xs->state != XSK_READY) {
812 mutex_unlock(&xs->mutex);
816 mutex_unlock(&xs->mutex);
820 q = (optname == XDP_UMEM_FILL_RING) ? &xs->umem->fq :
822 err = xsk_init_queue(entries, q, true);
823 mutex_unlock(&xs->mutex);
833 static void xsk_enter_rxtx_offsets(struct xdp_ring_offset_v1 *ring)
835 ring->producer = offsetof(struct xdp_rxtx_ring, ptrs.producer);
836 ring->consumer = offsetof(struct xdp_rxtx_ring, ptrs.consumer);
837 ring->desc = offsetof(struct xdp_rxtx_ring, desc);
840 static void xsk_enter_umem_offsets(struct xdp_ring_offset_v1 *ring)
842 ring->producer = offsetof(struct xdp_umem_ring, ptrs.producer);
843 ring->consumer = offsetof(struct xdp_umem_ring, ptrs.consumer);
844 ring->desc = offsetof(struct xdp_umem_ring, desc);
847 static int xsk_getsockopt(struct socket *sock, int level, int optname,
848 char __user *optval, int __user *optlen)
850 struct sock *sk = sock->sk;
851 struct xdp_sock *xs = xdp_sk(sk);
854 if (level != SOL_XDP)
857 if (get_user(len, optlen))
865 struct xdp_statistics stats;
867 if (len < sizeof(stats))
870 mutex_lock(&xs->mutex);
871 stats.rx_dropped = xs->rx_dropped;
872 stats.rx_invalid_descs = xskq_nb_invalid_descs(xs->rx);
873 stats.tx_invalid_descs = xskq_nb_invalid_descs(xs->tx);
874 mutex_unlock(&xs->mutex);
876 if (copy_to_user(optval, &stats, sizeof(stats)))
878 if (put_user(sizeof(stats), optlen))
883 case XDP_MMAP_OFFSETS:
885 struct xdp_mmap_offsets off;
886 struct xdp_mmap_offsets_v1 off_v1;
887 bool flags_supported = true;
890 if (len < sizeof(off_v1))
892 else if (len < sizeof(off))
893 flags_supported = false;
895 if (flags_supported) {
896 /* xdp_ring_offset is identical to xdp_ring_offset_v1
897 * except for the flags field added to the end.
899 xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *)
901 xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *)
903 xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *)
905 xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *)
907 off.rx.flags = offsetof(struct xdp_rxtx_ring,
909 off.tx.flags = offsetof(struct xdp_rxtx_ring,
911 off.fr.flags = offsetof(struct xdp_umem_ring,
913 off.cr.flags = offsetof(struct xdp_umem_ring,
919 xsk_enter_rxtx_offsets(&off_v1.rx);
920 xsk_enter_rxtx_offsets(&off_v1.tx);
921 xsk_enter_umem_offsets(&off_v1.fr);
922 xsk_enter_umem_offsets(&off_v1.cr);
924 len = sizeof(off_v1);
928 if (copy_to_user(optval, to_copy, len))
930 if (put_user(len, optlen))
937 struct xdp_options opts = {};
939 if (len < sizeof(opts))
942 mutex_lock(&xs->mutex);
944 opts.flags |= XDP_OPTIONS_ZEROCOPY;
945 mutex_unlock(&xs->mutex);
948 if (copy_to_user(optval, &opts, len))
950 if (put_user(len, optlen))
962 static int xsk_mmap(struct file *file, struct socket *sock,
963 struct vm_area_struct *vma)
965 loff_t offset = (loff_t)vma->vm_pgoff << PAGE_SHIFT;
966 unsigned long size = vma->vm_end - vma->vm_start;
967 struct xdp_sock *xs = xdp_sk(sock->sk);
968 struct xsk_queue *q = NULL;
969 struct xdp_umem *umem;
973 if (READ_ONCE(xs->state) != XSK_READY)
976 if (offset == XDP_PGOFF_RX_RING) {
977 q = READ_ONCE(xs->rx);
978 } else if (offset == XDP_PGOFF_TX_RING) {
979 q = READ_ONCE(xs->tx);
981 umem = READ_ONCE(xs->umem);
985 /* Matches the smp_wmb() in XDP_UMEM_REG */
987 if (offset == XDP_UMEM_PGOFF_FILL_RING)
988 q = READ_ONCE(umem->fq);
989 else if (offset == XDP_UMEM_PGOFF_COMPLETION_RING)
990 q = READ_ONCE(umem->cq);
996 /* Matches the smp_wmb() in xsk_init_queue */
998 qpg = virt_to_head_page(q->ring);
999 if (size > page_size(qpg))
1002 pfn = virt_to_phys(q->ring) >> PAGE_SHIFT;
1003 return remap_pfn_range(vma, vma->vm_start, pfn,
1004 size, vma->vm_page_prot);
1007 static int xsk_notifier(struct notifier_block *this,
1008 unsigned long msg, void *ptr)
1010 struct net_device *dev = netdev_notifier_info_to_dev(ptr);
1011 struct net *net = dev_net(dev);
1015 case NETDEV_UNREGISTER:
1016 mutex_lock(&net->xdp.lock);
1017 sk_for_each(sk, &net->xdp.list) {
1018 struct xdp_sock *xs = xdp_sk(sk);
1020 mutex_lock(&xs->mutex);
1021 if (xs->dev == dev) {
1022 sk->sk_err = ENETDOWN;
1023 if (!sock_flag(sk, SOCK_DEAD))
1024 sk->sk_error_report(sk);
1028 /* Clear device references in umem. */
1029 xdp_umem_clear_dev(xs->umem);
1031 mutex_unlock(&xs->mutex);
1033 mutex_unlock(&net->xdp.lock);
1039 static struct proto xsk_proto = {
1041 .owner = THIS_MODULE,
1042 .obj_size = sizeof(struct xdp_sock),
1045 static const struct proto_ops xsk_proto_ops = {
1047 .owner = THIS_MODULE,
1048 .release = xsk_release,
1050 .connect = sock_no_connect,
1051 .socketpair = sock_no_socketpair,
1052 .accept = sock_no_accept,
1053 .getname = sock_no_getname,
1055 .ioctl = sock_no_ioctl,
1056 .listen = sock_no_listen,
1057 .shutdown = sock_no_shutdown,
1058 .setsockopt = xsk_setsockopt,
1059 .getsockopt = xsk_getsockopt,
1060 .sendmsg = xsk_sendmsg,
1061 .recvmsg = sock_no_recvmsg,
1063 .sendpage = sock_no_sendpage,
1066 static void xsk_destruct(struct sock *sk)
1068 struct xdp_sock *xs = xdp_sk(sk);
1070 if (!sock_flag(sk, SOCK_DEAD))
1073 xdp_put_umem(xs->umem);
1075 sk_refcnt_debug_dec(sk);
1078 static int xsk_create(struct net *net, struct socket *sock, int protocol,
1082 struct xdp_sock *xs;
1084 if (!ns_capable(net->user_ns, CAP_NET_RAW))
1086 if (sock->type != SOCK_RAW)
1087 return -ESOCKTNOSUPPORT;
1090 return -EPROTONOSUPPORT;
1092 sock->state = SS_UNCONNECTED;
1094 sk = sk_alloc(net, PF_XDP, GFP_KERNEL, &xsk_proto, kern);
1098 sock->ops = &xsk_proto_ops;
1100 sock_init_data(sock, sk);
1102 sk->sk_family = PF_XDP;
1104 sk->sk_destruct = xsk_destruct;
1105 sk_refcnt_debug_inc(sk);
1107 sock_set_flag(sk, SOCK_RCU_FREE);
1110 xs->state = XSK_READY;
1111 mutex_init(&xs->mutex);
1112 spin_lock_init(&xs->rx_lock);
1113 spin_lock_init(&xs->tx_completion_lock);
1115 INIT_LIST_HEAD(&xs->map_list);
1116 spin_lock_init(&xs->map_list_lock);
1118 mutex_lock(&net->xdp.lock);
1119 sk_add_node_rcu(sk, &net->xdp.list);
1120 mutex_unlock(&net->xdp.lock);
1123 sock_prot_inuse_add(net, &xsk_proto, 1);
1129 static const struct net_proto_family xsk_family_ops = {
1131 .create = xsk_create,
1132 .owner = THIS_MODULE,
1135 static struct notifier_block xsk_netdev_notifier = {
1136 .notifier_call = xsk_notifier,
1139 static int __net_init xsk_net_init(struct net *net)
1141 mutex_init(&net->xdp.lock);
1142 INIT_HLIST_HEAD(&net->xdp.list);
1146 static void __net_exit xsk_net_exit(struct net *net)
1148 WARN_ON_ONCE(!hlist_empty(&net->xdp.list));
1151 static struct pernet_operations xsk_net_ops = {
1152 .init = xsk_net_init,
1153 .exit = xsk_net_exit,
1156 static int __init xsk_init(void)
1160 err = proto_register(&xsk_proto, 0 /* no slab */);
1164 err = sock_register(&xsk_family_ops);
1168 err = register_pernet_subsys(&xsk_net_ops);
1172 err = register_netdevice_notifier(&xsk_netdev_notifier);
1179 unregister_pernet_subsys(&xsk_net_ops);
1181 sock_unregister(PF_XDP);
1183 proto_unregister(&xsk_proto);
1188 fs_initcall(xsk_init);