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 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 bool xsk_umem_has_addrs(struct xdp_umem *umem, u32 cnt)
44 return xskq_has_addrs(umem->fq, cnt);
46 EXPORT_SYMBOL(xsk_umem_has_addrs);
48 u64 *xsk_umem_peek_addr(struct xdp_umem *umem, u64 *addr)
50 return xskq_peek_addr(umem->fq, addr, umem);
52 EXPORT_SYMBOL(xsk_umem_peek_addr);
54 void xsk_umem_discard_addr(struct xdp_umem *umem)
56 xskq_discard_addr(umem->fq);
58 EXPORT_SYMBOL(xsk_umem_discard_addr);
60 void xsk_set_rx_need_wakeup(struct xdp_umem *umem)
62 if (umem->need_wakeup & XDP_WAKEUP_RX)
65 umem->fq->ring->flags |= XDP_RING_NEED_WAKEUP;
66 umem->need_wakeup |= XDP_WAKEUP_RX;
68 EXPORT_SYMBOL(xsk_set_rx_need_wakeup);
70 void xsk_set_tx_need_wakeup(struct xdp_umem *umem)
74 if (umem->need_wakeup & XDP_WAKEUP_TX)
78 list_for_each_entry_rcu(xs, &umem->xsk_list, list) {
79 xs->tx->ring->flags |= XDP_RING_NEED_WAKEUP;
83 umem->need_wakeup |= XDP_WAKEUP_TX;
85 EXPORT_SYMBOL(xsk_set_tx_need_wakeup);
87 void xsk_clear_rx_need_wakeup(struct xdp_umem *umem)
89 if (!(umem->need_wakeup & XDP_WAKEUP_RX))
92 umem->fq->ring->flags &= ~XDP_RING_NEED_WAKEUP;
93 umem->need_wakeup &= ~XDP_WAKEUP_RX;
95 EXPORT_SYMBOL(xsk_clear_rx_need_wakeup);
97 void xsk_clear_tx_need_wakeup(struct xdp_umem *umem)
101 if (!(umem->need_wakeup & XDP_WAKEUP_TX))
105 list_for_each_entry_rcu(xs, &umem->xsk_list, list) {
106 xs->tx->ring->flags &= ~XDP_RING_NEED_WAKEUP;
110 umem->need_wakeup &= ~XDP_WAKEUP_TX;
112 EXPORT_SYMBOL(xsk_clear_tx_need_wakeup);
114 bool xsk_umem_uses_need_wakeup(struct xdp_umem *umem)
116 return umem->flags & XDP_UMEM_USES_NEED_WAKEUP;
118 EXPORT_SYMBOL(xsk_umem_uses_need_wakeup);
120 /* If a buffer crosses a page boundary, we need to do 2 memcpy's, one for
121 * each page. This is only required in copy mode.
123 static void __xsk_rcv_memcpy(struct xdp_umem *umem, u64 addr, void *from_buf,
124 u32 len, u32 metalen)
126 void *to_buf = xdp_umem_get_data(umem, addr);
128 addr = xsk_umem_add_offset_to_addr(addr);
129 if (xskq_crosses_non_contig_pg(umem, addr, len + metalen)) {
130 void *next_pg_addr = umem->pages[(addr >> PAGE_SHIFT) + 1].addr;
131 u64 page_start = addr & ~(PAGE_SIZE - 1);
132 u64 first_len = PAGE_SIZE - (addr - page_start);
134 memcpy(to_buf, from_buf, first_len + metalen);
135 memcpy(next_pg_addr, from_buf + first_len, len - first_len);
140 memcpy(to_buf, from_buf, len + metalen);
143 static int __xsk_rcv(struct xdp_sock *xs, struct xdp_buff *xdp, u32 len)
145 u64 offset = xs->umem->headroom;
146 u64 addr, memcpy_addr;
151 if (!xskq_peek_addr(xs->umem->fq, &addr, xs->umem) ||
152 len > xs->umem->chunk_size_nohr - XDP_PACKET_HEADROOM) {
157 if (unlikely(xdp_data_meta_unsupported(xdp))) {
158 from_buf = xdp->data;
161 from_buf = xdp->data_meta;
162 metalen = xdp->data - xdp->data_meta;
165 memcpy_addr = xsk_umem_adjust_offset(xs->umem, addr, offset);
166 __xsk_rcv_memcpy(xs->umem, memcpy_addr, from_buf, len, metalen);
169 addr = xsk_umem_adjust_offset(xs->umem, addr, offset);
170 err = xskq_produce_batch_desc(xs->rx, addr, len);
172 xskq_discard_addr(xs->umem->fq);
173 xdp_return_buff(xdp);
181 static int __xsk_rcv_zc(struct xdp_sock *xs, struct xdp_buff *xdp, u32 len)
183 int err = xskq_produce_batch_desc(xs->rx, (u64)xdp->handle, len);
191 static bool xsk_is_bound(struct xdp_sock *xs)
193 if (READ_ONCE(xs->state) == XSK_BOUND) {
194 /* Matches smp_wmb() in bind(). */
201 static int xsk_rcv(struct xdp_sock *xs, struct xdp_buff *xdp)
205 if (!xsk_is_bound(xs))
208 if (xs->dev != xdp->rxq->dev || xs->queue_id != xdp->rxq->queue_index)
211 len = xdp->data_end - xdp->data;
213 return (xdp->rxq->mem.type == MEM_TYPE_ZERO_COPY) ?
214 __xsk_rcv_zc(xs, xdp, len) : __xsk_rcv(xs, xdp, len);
217 static void xsk_flush(struct xdp_sock *xs)
219 xskq_produce_flush_desc(xs->rx);
220 xs->sk.sk_data_ready(&xs->sk);
223 int xsk_generic_rcv(struct xdp_sock *xs, struct xdp_buff *xdp)
225 u32 metalen = xdp->data - xdp->data_meta;
226 u32 len = xdp->data_end - xdp->data;
227 u64 offset = xs->umem->headroom;
232 spin_lock_bh(&xs->rx_lock);
234 if (xs->dev != xdp->rxq->dev || xs->queue_id != xdp->rxq->queue_index) {
239 if (!xskq_peek_addr(xs->umem->fq, &addr, xs->umem) ||
240 len > xs->umem->chunk_size_nohr - XDP_PACKET_HEADROOM) {
245 addr = xsk_umem_adjust_offset(xs->umem, addr, offset);
246 buffer = xdp_umem_get_data(xs->umem, addr);
247 memcpy(buffer, xdp->data_meta, len + metalen);
249 addr = xsk_umem_adjust_offset(xs->umem, addr, metalen);
250 err = xskq_produce_batch_desc(xs->rx, addr, len);
254 xskq_discard_addr(xs->umem->fq);
255 xskq_produce_flush_desc(xs->rx);
257 spin_unlock_bh(&xs->rx_lock);
259 xs->sk.sk_data_ready(&xs->sk);
265 spin_unlock_bh(&xs->rx_lock);
269 int __xsk_map_redirect(struct xdp_sock *xs, struct xdp_buff *xdp)
271 struct list_head *flush_list = this_cpu_ptr(&xskmap_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(void)
286 struct list_head *flush_list = this_cpu_ptr(&xskmap_flush_list);
287 struct xdp_sock *xs, *tmp;
289 list_for_each_entry_safe(xs, tmp, flush_list, flush_node) {
291 __list_del_clearprev(&xs->flush_node);
295 void xsk_umem_complete_tx(struct xdp_umem *umem, u32 nb_entries)
297 xskq_produce_flush_addr_n(umem->cq, nb_entries);
299 EXPORT_SYMBOL(xsk_umem_complete_tx);
301 void xsk_umem_consume_tx_done(struct xdp_umem *umem)
306 list_for_each_entry_rcu(xs, &umem->xsk_list, list) {
307 xs->sk.sk_write_space(&xs->sk);
311 EXPORT_SYMBOL(xsk_umem_consume_tx_done);
313 bool xsk_umem_consume_tx(struct xdp_umem *umem, struct xdp_desc *desc)
318 list_for_each_entry_rcu(xs, &umem->xsk_list, list) {
319 if (!xskq_peek_desc(xs->tx, desc, umem))
322 if (xskq_produce_addr_lazy(umem->cq, desc->addr))
325 xskq_discard_desc(xs->tx);
334 EXPORT_SYMBOL(xsk_umem_consume_tx);
336 static int xsk_zc_xmit(struct xdp_sock *xs)
338 struct net_device *dev = xs->dev;
340 return dev->netdev_ops->ndo_xsk_wakeup(dev, xs->queue_id,
344 static void xsk_destruct_skb(struct sk_buff *skb)
346 u64 addr = (u64)(long)skb_shinfo(skb)->destructor_arg;
347 struct xdp_sock *xs = xdp_sk(skb->sk);
350 spin_lock_irqsave(&xs->tx_completion_lock, flags);
351 WARN_ON_ONCE(xskq_produce_addr(xs->umem->cq, addr));
352 spin_unlock_irqrestore(&xs->tx_completion_lock, flags);
357 static int xsk_generic_xmit(struct sock *sk)
359 struct xdp_sock *xs = xdp_sk(sk);
360 u32 max_batch = TX_BATCH_SIZE;
361 bool sent_frame = false;
362 struct xdp_desc desc;
366 mutex_lock(&xs->mutex);
368 if (xs->queue_id >= xs->dev->real_num_tx_queues)
371 while (xskq_peek_desc(xs->tx, &desc, xs->umem)) {
376 if (max_batch-- == 0) {
382 skb = sock_alloc_send_skb(sk, len, 1, &err);
383 if (unlikely(!skb)) {
390 buffer = xdp_umem_get_data(xs->umem, addr);
391 err = skb_store_bits(skb, 0, buffer, len);
392 if (unlikely(err) || xskq_reserve_addr(xs->umem->cq)) {
398 skb->priority = sk->sk_priority;
399 skb->mark = sk->sk_mark;
400 skb_shinfo(skb)->destructor_arg = (void *)(long)desc.addr;
401 skb->destructor = xsk_destruct_skb;
403 err = dev_direct_xmit(skb, xs->queue_id);
404 xskq_discard_desc(xs->tx);
405 /* Ignore NET_XMIT_CN as packet might have been sent */
406 if (err == NET_XMIT_DROP || err == NETDEV_TX_BUSY) {
407 /* SKB completed but not sent */
417 sk->sk_write_space(sk);
419 mutex_unlock(&xs->mutex);
423 static int __xsk_sendmsg(struct sock *sk)
425 struct xdp_sock *xs = xdp_sk(sk);
427 if (unlikely(!(xs->dev->flags & IFF_UP)))
429 if (unlikely(!xs->tx))
432 return xs->zc ? xsk_zc_xmit(xs) : xsk_generic_xmit(sk);
435 static int xsk_sendmsg(struct socket *sock, struct msghdr *m, size_t total_len)
437 bool need_wait = !(m->msg_flags & MSG_DONTWAIT);
438 struct sock *sk = sock->sk;
439 struct xdp_sock *xs = xdp_sk(sk);
441 if (unlikely(!xsk_is_bound(xs)))
443 if (unlikely(need_wait))
446 return __xsk_sendmsg(sk);
449 static __poll_t xsk_poll(struct file *file, struct socket *sock,
450 struct poll_table_struct *wait)
452 __poll_t mask = datagram_poll(file, sock, wait);
453 struct sock *sk = sock->sk;
454 struct xdp_sock *xs = xdp_sk(sk);
455 struct net_device *dev;
456 struct xdp_umem *umem;
458 if (unlikely(!xsk_is_bound(xs)))
464 if (umem->need_wakeup) {
465 if (dev->netdev_ops->ndo_xsk_wakeup)
466 dev->netdev_ops->ndo_xsk_wakeup(dev, xs->queue_id,
469 /* Poll needs to drive Tx also in copy mode */
473 if (xs->rx && !xskq_empty_desc(xs->rx))
474 mask |= EPOLLIN | EPOLLRDNORM;
475 if (xs->tx && !xskq_full_desc(xs->tx))
476 mask |= EPOLLOUT | EPOLLWRNORM;
481 static int xsk_init_queue(u32 entries, struct xsk_queue **queue,
486 if (entries == 0 || *queue || !is_power_of_2(entries))
489 q = xskq_create(entries, umem_queue);
493 /* Make sure queue is ready before it can be seen by others */
495 WRITE_ONCE(*queue, q);
499 static void xsk_unbind_dev(struct xdp_sock *xs)
501 struct net_device *dev = xs->dev;
503 if (xs->state != XSK_BOUND)
505 WRITE_ONCE(xs->state, XSK_UNBOUND);
507 /* Wait for driver to stop using the xdp socket. */
508 xdp_del_sk_umem(xs->umem, xs);
514 static struct xsk_map *xsk_get_map_list_entry(struct xdp_sock *xs,
515 struct xdp_sock ***map_entry)
517 struct xsk_map *map = NULL;
518 struct xsk_map_node *node;
522 spin_lock_bh(&xs->map_list_lock);
523 node = list_first_entry_or_null(&xs->map_list, struct xsk_map_node,
526 WARN_ON(xsk_map_inc(node->map));
528 *map_entry = node->map_entry;
530 spin_unlock_bh(&xs->map_list_lock);
534 static void xsk_delete_from_maps(struct xdp_sock *xs)
536 /* This function removes the current XDP socket from all the
537 * maps it resides in. We need to take extra care here, due to
538 * the two locks involved. Each map has a lock synchronizing
539 * updates to the entries, and each socket has a lock that
540 * synchronizes access to the list of maps (map_list). For
541 * deadlock avoidance the locks need to be taken in the order
542 * "map lock"->"socket map list lock". We start off by
543 * accessing the socket map list, and take a reference to the
544 * map to guarantee existence between the
545 * xsk_get_map_list_entry() and xsk_map_try_sock_delete()
546 * calls. Then we ask the map to remove the socket, which
547 * tries to remove the socket from the map. Note that there
548 * might be updates to the map between
549 * xsk_get_map_list_entry() and xsk_map_try_sock_delete().
551 struct xdp_sock **map_entry = NULL;
554 while ((map = xsk_get_map_list_entry(xs, &map_entry))) {
555 xsk_map_try_sock_delete(map, xs, map_entry);
560 static int xsk_release(struct socket *sock)
562 struct sock *sk = sock->sk;
563 struct xdp_sock *xs = xdp_sk(sk);
571 mutex_lock(&net->xdp.lock);
572 sk_del_node_init_rcu(sk);
573 mutex_unlock(&net->xdp.lock);
576 sock_prot_inuse_add(net, sk->sk_prot, -1);
579 xsk_delete_from_maps(xs);
580 mutex_lock(&xs->mutex);
582 mutex_unlock(&xs->mutex);
584 xskq_destroy(xs->rx);
585 xskq_destroy(xs->tx);
590 sk_refcnt_debug_release(sk);
596 static struct socket *xsk_lookup_xsk_from_fd(int fd)
601 sock = sockfd_lookup(fd, &err);
603 return ERR_PTR(-ENOTSOCK);
605 if (sock->sk->sk_family != PF_XDP) {
607 return ERR_PTR(-ENOPROTOOPT);
613 /* Check if umem pages are contiguous.
614 * If zero-copy mode, use the DMA address to do the page contiguity check
615 * For all other modes we use addr (kernel virtual address)
616 * Store the result in the low bits of addr.
618 static void xsk_check_page_contiguity(struct xdp_umem *umem, u32 flags)
620 struct xdp_umem_page *pgs = umem->pages;
623 for (i = 0; i < umem->npgs - 1; i++) {
624 is_contig = (flags & XDP_ZEROCOPY) ?
625 (pgs[i].dma + PAGE_SIZE == pgs[i + 1].dma) :
626 (pgs[i].addr + PAGE_SIZE == pgs[i + 1].addr);
627 pgs[i].addr += is_contig << XSK_NEXT_PG_CONTIG_SHIFT;
631 static int xsk_bind(struct socket *sock, struct sockaddr *addr, int addr_len)
633 struct sockaddr_xdp *sxdp = (struct sockaddr_xdp *)addr;
634 struct sock *sk = sock->sk;
635 struct xdp_sock *xs = xdp_sk(sk);
636 struct net_device *dev;
640 if (addr_len < sizeof(struct sockaddr_xdp))
642 if (sxdp->sxdp_family != AF_XDP)
645 flags = sxdp->sxdp_flags;
646 if (flags & ~(XDP_SHARED_UMEM | XDP_COPY | XDP_ZEROCOPY |
647 XDP_USE_NEED_WAKEUP))
651 mutex_lock(&xs->mutex);
652 if (xs->state != XSK_READY) {
657 dev = dev_get_by_index(sock_net(sk), sxdp->sxdp_ifindex);
663 if (!xs->rx && !xs->tx) {
668 qid = sxdp->sxdp_queue_id;
670 if (flags & XDP_SHARED_UMEM) {
671 struct xdp_sock *umem_xs;
674 if ((flags & XDP_COPY) || (flags & XDP_ZEROCOPY) ||
675 (flags & XDP_USE_NEED_WAKEUP)) {
676 /* Cannot specify flags for shared sockets. */
682 /* We have already our own. */
687 sock = xsk_lookup_xsk_from_fd(sxdp->sxdp_shared_umem_fd);
693 umem_xs = xdp_sk(sock->sk);
694 if (!xsk_is_bound(umem_xs)) {
699 if (umem_xs->dev != dev || umem_xs->queue_id != qid) {
705 xdp_get_umem(umem_xs->umem);
706 WRITE_ONCE(xs->umem, umem_xs->umem);
708 } else if (!xs->umem || !xdp_umem_validate_queues(xs->umem)) {
712 /* This xsk has its own umem. */
713 xskq_set_umem(xs->umem->fq, xs->umem->size,
714 xs->umem->chunk_mask);
715 xskq_set_umem(xs->umem->cq, xs->umem->size,
716 xs->umem->chunk_mask);
718 err = xdp_umem_assign_dev(xs->umem, dev, qid, flags);
722 xsk_check_page_contiguity(xs->umem, flags);
726 xs->zc = xs->umem->zc;
728 xskq_set_umem(xs->rx, xs->umem->size, xs->umem->chunk_mask);
729 xskq_set_umem(xs->tx, xs->umem->size, xs->umem->chunk_mask);
730 xdp_add_sk_umem(xs->umem, xs);
736 /* Matches smp_rmb() in bind() for shared umem
737 * sockets, and xsk_is_bound().
740 WRITE_ONCE(xs->state, XSK_BOUND);
743 mutex_unlock(&xs->mutex);
748 struct xdp_umem_reg_v1 {
749 __u64 addr; /* Start of packet data area */
750 __u64 len; /* Length of packet data area */
755 static int xsk_setsockopt(struct socket *sock, int level, int optname,
756 char __user *optval, unsigned int optlen)
758 struct sock *sk = sock->sk;
759 struct xdp_sock *xs = xdp_sk(sk);
762 if (level != SOL_XDP)
769 struct xsk_queue **q;
772 if (optlen < sizeof(entries))
774 if (copy_from_user(&entries, optval, sizeof(entries)))
777 mutex_lock(&xs->mutex);
778 if (xs->state != XSK_READY) {
779 mutex_unlock(&xs->mutex);
782 q = (optname == XDP_TX_RING) ? &xs->tx : &xs->rx;
783 err = xsk_init_queue(entries, q, false);
784 if (!err && optname == XDP_TX_RING)
785 /* Tx needs to be explicitly woken up the first time */
786 xs->tx->ring->flags |= XDP_RING_NEED_WAKEUP;
787 mutex_unlock(&xs->mutex);
792 size_t mr_size = sizeof(struct xdp_umem_reg);
793 struct xdp_umem_reg mr = {};
794 struct xdp_umem *umem;
796 if (optlen < sizeof(struct xdp_umem_reg_v1))
798 else if (optlen < sizeof(mr))
799 mr_size = sizeof(struct xdp_umem_reg_v1);
801 if (copy_from_user(&mr, optval, mr_size))
804 mutex_lock(&xs->mutex);
805 if (xs->state != XSK_READY || xs->umem) {
806 mutex_unlock(&xs->mutex);
810 umem = xdp_umem_create(&mr);
812 mutex_unlock(&xs->mutex);
813 return PTR_ERR(umem);
816 /* Make sure umem is ready before it can be seen by others */
818 WRITE_ONCE(xs->umem, umem);
819 mutex_unlock(&xs->mutex);
822 case XDP_UMEM_FILL_RING:
823 case XDP_UMEM_COMPLETION_RING:
825 struct xsk_queue **q;
828 if (copy_from_user(&entries, optval, sizeof(entries)))
831 mutex_lock(&xs->mutex);
832 if (xs->state != XSK_READY) {
833 mutex_unlock(&xs->mutex);
837 mutex_unlock(&xs->mutex);
841 q = (optname == XDP_UMEM_FILL_RING) ? &xs->umem->fq :
843 err = xsk_init_queue(entries, q, true);
844 mutex_unlock(&xs->mutex);
854 static void xsk_enter_rxtx_offsets(struct xdp_ring_offset_v1 *ring)
856 ring->producer = offsetof(struct xdp_rxtx_ring, ptrs.producer);
857 ring->consumer = offsetof(struct xdp_rxtx_ring, ptrs.consumer);
858 ring->desc = offsetof(struct xdp_rxtx_ring, desc);
861 static void xsk_enter_umem_offsets(struct xdp_ring_offset_v1 *ring)
863 ring->producer = offsetof(struct xdp_umem_ring, ptrs.producer);
864 ring->consumer = offsetof(struct xdp_umem_ring, ptrs.consumer);
865 ring->desc = offsetof(struct xdp_umem_ring, desc);
868 static int xsk_getsockopt(struct socket *sock, int level, int optname,
869 char __user *optval, int __user *optlen)
871 struct sock *sk = sock->sk;
872 struct xdp_sock *xs = xdp_sk(sk);
875 if (level != SOL_XDP)
878 if (get_user(len, optlen))
886 struct xdp_statistics stats;
888 if (len < sizeof(stats))
891 mutex_lock(&xs->mutex);
892 stats.rx_dropped = xs->rx_dropped;
893 stats.rx_invalid_descs = xskq_nb_invalid_descs(xs->rx);
894 stats.tx_invalid_descs = xskq_nb_invalid_descs(xs->tx);
895 mutex_unlock(&xs->mutex);
897 if (copy_to_user(optval, &stats, sizeof(stats)))
899 if (put_user(sizeof(stats), optlen))
904 case XDP_MMAP_OFFSETS:
906 struct xdp_mmap_offsets off;
907 struct xdp_mmap_offsets_v1 off_v1;
908 bool flags_supported = true;
911 if (len < sizeof(off_v1))
913 else if (len < sizeof(off))
914 flags_supported = false;
916 if (flags_supported) {
917 /* xdp_ring_offset is identical to xdp_ring_offset_v1
918 * except for the flags field added to the end.
920 xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *)
922 xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *)
924 xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *)
926 xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *)
928 off.rx.flags = offsetof(struct xdp_rxtx_ring,
930 off.tx.flags = offsetof(struct xdp_rxtx_ring,
932 off.fr.flags = offsetof(struct xdp_umem_ring,
934 off.cr.flags = offsetof(struct xdp_umem_ring,
940 xsk_enter_rxtx_offsets(&off_v1.rx);
941 xsk_enter_rxtx_offsets(&off_v1.tx);
942 xsk_enter_umem_offsets(&off_v1.fr);
943 xsk_enter_umem_offsets(&off_v1.cr);
945 len = sizeof(off_v1);
949 if (copy_to_user(optval, to_copy, len))
951 if (put_user(len, optlen))
958 struct xdp_options opts = {};
960 if (len < sizeof(opts))
963 mutex_lock(&xs->mutex);
965 opts.flags |= XDP_OPTIONS_ZEROCOPY;
966 mutex_unlock(&xs->mutex);
969 if (copy_to_user(optval, &opts, len))
971 if (put_user(len, optlen))
983 static int xsk_mmap(struct file *file, struct socket *sock,
984 struct vm_area_struct *vma)
986 loff_t offset = (loff_t)vma->vm_pgoff << PAGE_SHIFT;
987 unsigned long size = vma->vm_end - vma->vm_start;
988 struct xdp_sock *xs = xdp_sk(sock->sk);
989 struct xsk_queue *q = NULL;
990 struct xdp_umem *umem;
994 if (READ_ONCE(xs->state) != XSK_READY)
997 if (offset == XDP_PGOFF_RX_RING) {
998 q = READ_ONCE(xs->rx);
999 } else if (offset == XDP_PGOFF_TX_RING) {
1000 q = READ_ONCE(xs->tx);
1002 umem = READ_ONCE(xs->umem);
1006 /* Matches the smp_wmb() in XDP_UMEM_REG */
1008 if (offset == XDP_UMEM_PGOFF_FILL_RING)
1009 q = READ_ONCE(umem->fq);
1010 else if (offset == XDP_UMEM_PGOFF_COMPLETION_RING)
1011 q = READ_ONCE(umem->cq);
1017 /* Matches the smp_wmb() in xsk_init_queue */
1019 qpg = virt_to_head_page(q->ring);
1020 if (size > page_size(qpg))
1023 pfn = virt_to_phys(q->ring) >> PAGE_SHIFT;
1024 return remap_pfn_range(vma, vma->vm_start, pfn,
1025 size, vma->vm_page_prot);
1028 static int xsk_notifier(struct notifier_block *this,
1029 unsigned long msg, void *ptr)
1031 struct net_device *dev = netdev_notifier_info_to_dev(ptr);
1032 struct net *net = dev_net(dev);
1036 case NETDEV_UNREGISTER:
1037 mutex_lock(&net->xdp.lock);
1038 sk_for_each(sk, &net->xdp.list) {
1039 struct xdp_sock *xs = xdp_sk(sk);
1041 mutex_lock(&xs->mutex);
1042 if (xs->dev == dev) {
1043 sk->sk_err = ENETDOWN;
1044 if (!sock_flag(sk, SOCK_DEAD))
1045 sk->sk_error_report(sk);
1049 /* Clear device references in umem. */
1050 xdp_umem_clear_dev(xs->umem);
1052 mutex_unlock(&xs->mutex);
1054 mutex_unlock(&net->xdp.lock);
1060 static struct proto xsk_proto = {
1062 .owner = THIS_MODULE,
1063 .obj_size = sizeof(struct xdp_sock),
1066 static const struct proto_ops xsk_proto_ops = {
1068 .owner = THIS_MODULE,
1069 .release = xsk_release,
1071 .connect = sock_no_connect,
1072 .socketpair = sock_no_socketpair,
1073 .accept = sock_no_accept,
1074 .getname = sock_no_getname,
1076 .ioctl = sock_no_ioctl,
1077 .listen = sock_no_listen,
1078 .shutdown = sock_no_shutdown,
1079 .setsockopt = xsk_setsockopt,
1080 .getsockopt = xsk_getsockopt,
1081 .sendmsg = xsk_sendmsg,
1082 .recvmsg = sock_no_recvmsg,
1084 .sendpage = sock_no_sendpage,
1087 static void xsk_destruct(struct sock *sk)
1089 struct xdp_sock *xs = xdp_sk(sk);
1091 if (!sock_flag(sk, SOCK_DEAD))
1094 xdp_put_umem(xs->umem);
1096 sk_refcnt_debug_dec(sk);
1099 static int xsk_create(struct net *net, struct socket *sock, int protocol,
1103 struct xdp_sock *xs;
1105 if (!ns_capable(net->user_ns, CAP_NET_RAW))
1107 if (sock->type != SOCK_RAW)
1108 return -ESOCKTNOSUPPORT;
1111 return -EPROTONOSUPPORT;
1113 sock->state = SS_UNCONNECTED;
1115 sk = sk_alloc(net, PF_XDP, GFP_KERNEL, &xsk_proto, kern);
1119 sock->ops = &xsk_proto_ops;
1121 sock_init_data(sock, sk);
1123 sk->sk_family = PF_XDP;
1125 sk->sk_destruct = xsk_destruct;
1126 sk_refcnt_debug_inc(sk);
1128 sock_set_flag(sk, SOCK_RCU_FREE);
1131 xs->state = XSK_READY;
1132 mutex_init(&xs->mutex);
1133 spin_lock_init(&xs->rx_lock);
1134 spin_lock_init(&xs->tx_completion_lock);
1136 INIT_LIST_HEAD(&xs->map_list);
1137 spin_lock_init(&xs->map_list_lock);
1139 mutex_lock(&net->xdp.lock);
1140 sk_add_node_rcu(sk, &net->xdp.list);
1141 mutex_unlock(&net->xdp.lock);
1144 sock_prot_inuse_add(net, &xsk_proto, 1);
1150 static const struct net_proto_family xsk_family_ops = {
1152 .create = xsk_create,
1153 .owner = THIS_MODULE,
1156 static struct notifier_block xsk_netdev_notifier = {
1157 .notifier_call = xsk_notifier,
1160 static int __net_init xsk_net_init(struct net *net)
1162 mutex_init(&net->xdp.lock);
1163 INIT_HLIST_HEAD(&net->xdp.list);
1167 static void __net_exit xsk_net_exit(struct net *net)
1169 WARN_ON_ONCE(!hlist_empty(&net->xdp.list));
1172 static struct pernet_operations xsk_net_ops = {
1173 .init = xsk_net_init,
1174 .exit = xsk_net_exit,
1177 static int __init xsk_init(void)
1181 err = proto_register(&xsk_proto, 0 /* no slab */);
1185 err = sock_register(&xsk_family_ops);
1189 err = register_pernet_subsys(&xsk_net_ops);
1193 err = register_netdevice_notifier(&xsk_netdev_notifier);
1197 for_each_possible_cpu(cpu)
1198 INIT_LIST_HEAD(&per_cpu(xskmap_flush_list, cpu));
1202 unregister_pernet_subsys(&xsk_net_ops);
1204 sock_unregister(PF_XDP);
1206 proto_unregister(&xsk_proto);
1211 fs_initcall(xsk_init);