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>
26 #include <net/busy_poll.h>
29 #include "xsk_queue.h"
33 #define TX_BATCH_SIZE 16
35 static DEFINE_PER_CPU(struct list_head, xskmap_flush_list);
37 void xsk_set_rx_need_wakeup(struct xsk_buff_pool *pool)
39 if (pool->cached_need_wakeup & XDP_WAKEUP_RX)
42 pool->fq->ring->flags |= XDP_RING_NEED_WAKEUP;
43 pool->cached_need_wakeup |= XDP_WAKEUP_RX;
45 EXPORT_SYMBOL(xsk_set_rx_need_wakeup);
47 void xsk_set_tx_need_wakeup(struct xsk_buff_pool *pool)
51 if (pool->cached_need_wakeup & XDP_WAKEUP_TX)
55 list_for_each_entry_rcu(xs, &pool->xsk_tx_list, tx_list) {
56 xs->tx->ring->flags |= XDP_RING_NEED_WAKEUP;
60 pool->cached_need_wakeup |= XDP_WAKEUP_TX;
62 EXPORT_SYMBOL(xsk_set_tx_need_wakeup);
64 void xsk_clear_rx_need_wakeup(struct xsk_buff_pool *pool)
66 if (!(pool->cached_need_wakeup & XDP_WAKEUP_RX))
69 pool->fq->ring->flags &= ~XDP_RING_NEED_WAKEUP;
70 pool->cached_need_wakeup &= ~XDP_WAKEUP_RX;
72 EXPORT_SYMBOL(xsk_clear_rx_need_wakeup);
74 void xsk_clear_tx_need_wakeup(struct xsk_buff_pool *pool)
78 if (!(pool->cached_need_wakeup & XDP_WAKEUP_TX))
82 list_for_each_entry_rcu(xs, &pool->xsk_tx_list, tx_list) {
83 xs->tx->ring->flags &= ~XDP_RING_NEED_WAKEUP;
87 pool->cached_need_wakeup &= ~XDP_WAKEUP_TX;
89 EXPORT_SYMBOL(xsk_clear_tx_need_wakeup);
91 bool xsk_uses_need_wakeup(struct xsk_buff_pool *pool)
93 return pool->uses_need_wakeup;
95 EXPORT_SYMBOL(xsk_uses_need_wakeup);
97 struct xsk_buff_pool *xsk_get_pool_from_qid(struct net_device *dev,
100 if (queue_id < dev->real_num_rx_queues)
101 return dev->_rx[queue_id].pool;
102 if (queue_id < dev->real_num_tx_queues)
103 return dev->_tx[queue_id].pool;
107 EXPORT_SYMBOL(xsk_get_pool_from_qid);
109 void xsk_clear_pool_at_qid(struct net_device *dev, u16 queue_id)
111 if (queue_id < dev->real_num_rx_queues)
112 dev->_rx[queue_id].pool = NULL;
113 if (queue_id < dev->real_num_tx_queues)
114 dev->_tx[queue_id].pool = NULL;
117 /* The buffer pool is stored both in the _rx struct and the _tx struct as we do
118 * not know if the device has more tx queues than rx, or the opposite.
119 * This might also change during run time.
121 int xsk_reg_pool_at_qid(struct net_device *dev, struct xsk_buff_pool *pool,
124 if (queue_id >= max_t(unsigned int,
125 dev->real_num_rx_queues,
126 dev->real_num_tx_queues))
129 if (queue_id < dev->real_num_rx_queues)
130 dev->_rx[queue_id].pool = pool;
131 if (queue_id < dev->real_num_tx_queues)
132 dev->_tx[queue_id].pool = pool;
137 void xp_release(struct xdp_buff_xsk *xskb)
139 xskb->pool->free_heads[xskb->pool->free_heads_cnt++] = xskb;
142 static u64 xp_get_handle(struct xdp_buff_xsk *xskb)
144 u64 offset = xskb->xdp.data - xskb->xdp.data_hard_start;
146 offset += xskb->pool->headroom;
147 if (!xskb->pool->unaligned)
148 return xskb->orig_addr + offset;
149 return xskb->orig_addr + (offset << XSK_UNALIGNED_BUF_OFFSET_SHIFT);
152 static int __xsk_rcv_zc(struct xdp_sock *xs, struct xdp_buff *xdp, u32 len)
154 struct xdp_buff_xsk *xskb = container_of(xdp, struct xdp_buff_xsk, xdp);
158 addr = xp_get_handle(xskb);
159 err = xskq_prod_reserve_desc(xs->rx, addr, len);
169 static void xsk_copy_xdp(struct xdp_buff *to, struct xdp_buff *from, u32 len)
171 void *from_buf, *to_buf;
174 if (unlikely(xdp_data_meta_unsupported(from))) {
175 from_buf = from->data;
179 from_buf = from->data_meta;
180 metalen = from->data - from->data_meta;
181 to_buf = to->data - metalen;
184 memcpy(to_buf, from_buf, len + metalen);
187 static int __xsk_rcv(struct xdp_sock *xs, struct xdp_buff *xdp, u32 len,
190 struct xdp_buff *xsk_xdp;
193 if (len > xsk_pool_get_rx_frame_size(xs->pool)) {
198 xsk_xdp = xsk_buff_alloc(xs->pool);
204 xsk_copy_xdp(xsk_xdp, xdp, len);
205 err = __xsk_rcv_zc(xs, xsk_xdp, len);
207 xsk_buff_free(xsk_xdp);
211 xdp_return_buff(xdp);
215 static bool xsk_tx_writeable(struct xdp_sock *xs)
217 if (xskq_cons_present_entries(xs->tx) > xs->tx->nentries / 2)
223 static bool xsk_is_bound(struct xdp_sock *xs)
225 if (READ_ONCE(xs->state) == XSK_BOUND) {
226 /* Matches smp_wmb() in bind(). */
233 static int xsk_rcv(struct xdp_sock *xs, struct xdp_buff *xdp,
238 if (!xsk_is_bound(xs))
241 if (xs->dev != xdp->rxq->dev || xs->queue_id != xdp->rxq->queue_index)
244 sk_mark_napi_id_once_xdp(&xs->sk, xdp);
245 len = xdp->data_end - xdp->data;
247 return xdp->rxq->mem.type == MEM_TYPE_XSK_BUFF_POOL ?
248 __xsk_rcv_zc(xs, xdp, len) :
249 __xsk_rcv(xs, xdp, len, explicit_free);
252 static void xsk_flush(struct xdp_sock *xs)
254 xskq_prod_submit(xs->rx);
255 __xskq_cons_release(xs->pool->fq);
256 sock_def_readable(&xs->sk);
259 int xsk_generic_rcv(struct xdp_sock *xs, struct xdp_buff *xdp)
263 spin_lock_bh(&xs->rx_lock);
264 err = xsk_rcv(xs, xdp, false);
266 spin_unlock_bh(&xs->rx_lock);
270 int __xsk_map_redirect(struct xdp_sock *xs, struct xdp_buff *xdp)
272 struct list_head *flush_list = this_cpu_ptr(&xskmap_flush_list);
275 err = xsk_rcv(xs, xdp, true);
279 if (!xs->flush_node.prev)
280 list_add(&xs->flush_node, flush_list);
285 void __xsk_map_flush(void)
287 struct list_head *flush_list = this_cpu_ptr(&xskmap_flush_list);
288 struct xdp_sock *xs, *tmp;
290 list_for_each_entry_safe(xs, tmp, flush_list, flush_node) {
292 __list_del_clearprev(&xs->flush_node);
296 void xsk_tx_completed(struct xsk_buff_pool *pool, u32 nb_entries)
298 xskq_prod_submit_n(pool->cq, nb_entries);
300 EXPORT_SYMBOL(xsk_tx_completed);
302 void xsk_tx_release(struct xsk_buff_pool *pool)
307 list_for_each_entry_rcu(xs, &pool->xsk_tx_list, tx_list) {
308 __xskq_cons_release(xs->tx);
309 if (xsk_tx_writeable(xs))
310 xs->sk.sk_write_space(&xs->sk);
314 EXPORT_SYMBOL(xsk_tx_release);
316 bool xsk_tx_peek_desc(struct xsk_buff_pool *pool, struct xdp_desc *desc)
321 list_for_each_entry_rcu(xs, &pool->xsk_tx_list, tx_list) {
322 if (!xskq_cons_peek_desc(xs->tx, desc, pool)) {
323 xs->tx->queue_empty_descs++;
327 /* This is the backpressure mechanism for the Tx path.
328 * Reserve space in the completion queue and only proceed
329 * if there is space in it. This avoids having to implement
330 * any buffering in the Tx path.
332 if (xskq_prod_reserve_addr(pool->cq, desc->addr))
335 xskq_cons_release(xs->tx);
344 EXPORT_SYMBOL(xsk_tx_peek_desc);
346 static u32 xsk_tx_peek_release_fallback(struct xsk_buff_pool *pool, struct xdp_desc *descs,
351 while (nb_pkts < max_entries && xsk_tx_peek_desc(pool, &descs[nb_pkts]))
354 xsk_tx_release(pool);
358 u32 xsk_tx_peek_release_desc_batch(struct xsk_buff_pool *pool, struct xdp_desc *descs,
365 if (!list_is_singular(&pool->xsk_tx_list)) {
366 /* Fallback to the non-batched version */
368 return xsk_tx_peek_release_fallback(pool, descs, max_entries);
371 xs = list_first_or_null_rcu(&pool->xsk_tx_list, struct xdp_sock, tx_list);
377 nb_pkts = xskq_cons_peek_desc_batch(xs->tx, descs, pool, max_entries);
379 xs->tx->queue_empty_descs++;
383 /* This is the backpressure mechanism for the Tx path. Try to
384 * reserve space in the completion queue for all packets, but
385 * if there are fewer slots available, just process that many
386 * packets. This avoids having to implement any buffering in
389 nb_pkts = xskq_prod_reserve_addr_batch(pool->cq, descs, nb_pkts);
393 xskq_cons_release_n(xs->tx, nb_pkts);
394 __xskq_cons_release(xs->tx);
395 xs->sk.sk_write_space(&xs->sk);
401 EXPORT_SYMBOL(xsk_tx_peek_release_desc_batch);
403 static int xsk_wakeup(struct xdp_sock *xs, u8 flags)
405 struct net_device *dev = xs->dev;
409 err = dev->netdev_ops->ndo_xsk_wakeup(dev, xs->queue_id, flags);
415 static int xsk_zc_xmit(struct xdp_sock *xs)
417 return xsk_wakeup(xs, XDP_WAKEUP_TX);
420 static void xsk_destruct_skb(struct sk_buff *skb)
422 u64 addr = (u64)(long)skb_shinfo(skb)->destructor_arg;
423 struct xdp_sock *xs = xdp_sk(skb->sk);
426 spin_lock_irqsave(&xs->tx_completion_lock, flags);
427 xskq_prod_submit_addr(xs->pool->cq, addr);
428 spin_unlock_irqrestore(&xs->tx_completion_lock, flags);
433 static int xsk_generic_xmit(struct sock *sk)
435 struct xdp_sock *xs = xdp_sk(sk);
436 u32 max_batch = TX_BATCH_SIZE;
437 bool sent_frame = false;
438 struct xdp_desc desc;
442 mutex_lock(&xs->mutex);
444 if (xs->queue_id >= xs->dev->real_num_tx_queues)
447 while (xskq_cons_peek_desc(xs->tx, &desc, xs->pool)) {
452 if (max_batch-- == 0) {
458 skb = sock_alloc_send_skb(sk, len, 1, &err);
464 buffer = xsk_buff_raw_get_data(xs->pool, addr);
465 err = skb_store_bits(skb, 0, buffer, len);
466 /* This is the backpressure mechanism for the Tx path.
467 * Reserve space in the completion queue and only proceed
468 * if there is space in it. This avoids having to implement
469 * any buffering in the Tx path.
471 if (unlikely(err) || xskq_prod_reserve(xs->pool->cq)) {
477 skb->priority = sk->sk_priority;
478 skb->mark = sk->sk_mark;
479 skb_shinfo(skb)->destructor_arg = (void *)(long)desc.addr;
480 skb->destructor = xsk_destruct_skb;
482 err = __dev_direct_xmit(skb, xs->queue_id);
483 if (err == NETDEV_TX_BUSY) {
484 /* Tell user-space to retry the send */
485 skb->destructor = sock_wfree;
486 /* Free skb without triggering the perf drop trace */
492 xskq_cons_release(xs->tx);
493 /* Ignore NET_XMIT_CN as packet might have been sent */
494 if (err == NET_XMIT_DROP) {
495 /* SKB completed but not sent */
503 xs->tx->queue_empty_descs++;
507 if (xsk_tx_writeable(xs))
508 sk->sk_write_space(sk);
510 mutex_unlock(&xs->mutex);
514 static int __xsk_sendmsg(struct sock *sk)
516 struct xdp_sock *xs = xdp_sk(sk);
518 if (unlikely(!(xs->dev->flags & IFF_UP)))
520 if (unlikely(!xs->tx))
523 return xs->zc ? xsk_zc_xmit(xs) : xsk_generic_xmit(sk);
526 static bool xsk_no_wakeup(struct sock *sk)
528 #ifdef CONFIG_NET_RX_BUSY_POLL
529 /* Prefer busy-polling, skip the wakeup. */
530 return READ_ONCE(sk->sk_prefer_busy_poll) && READ_ONCE(sk->sk_ll_usec) &&
531 READ_ONCE(sk->sk_napi_id) >= MIN_NAPI_ID;
537 static int xsk_sendmsg(struct socket *sock, struct msghdr *m, size_t total_len)
539 bool need_wait = !(m->msg_flags & MSG_DONTWAIT);
540 struct sock *sk = sock->sk;
541 struct xdp_sock *xs = xdp_sk(sk);
542 struct xsk_buff_pool *pool;
544 if (unlikely(!xsk_is_bound(xs)))
546 if (unlikely(need_wait))
549 if (sk_can_busy_loop(sk))
550 sk_busy_loop(sk, 1); /* only support non-blocking sockets */
552 if (xsk_no_wakeup(sk))
556 if (pool->cached_need_wakeup & XDP_WAKEUP_TX)
557 return __xsk_sendmsg(sk);
561 static int xsk_recvmsg(struct socket *sock, struct msghdr *m, size_t len, int flags)
563 bool need_wait = !(flags & MSG_DONTWAIT);
564 struct sock *sk = sock->sk;
565 struct xdp_sock *xs = xdp_sk(sk);
567 if (unlikely(!xsk_is_bound(xs)))
569 if (unlikely(!(xs->dev->flags & IFF_UP)))
571 if (unlikely(!xs->rx))
573 if (unlikely(need_wait))
576 if (sk_can_busy_loop(sk))
577 sk_busy_loop(sk, 1); /* only support non-blocking sockets */
579 if (xsk_no_wakeup(sk))
582 if (xs->pool->cached_need_wakeup & XDP_WAKEUP_RX && xs->zc)
583 return xsk_wakeup(xs, XDP_WAKEUP_RX);
587 static __poll_t xsk_poll(struct file *file, struct socket *sock,
588 struct poll_table_struct *wait)
591 struct sock *sk = sock->sk;
592 struct xdp_sock *xs = xdp_sk(sk);
593 struct xsk_buff_pool *pool;
595 sock_poll_wait(file, sock, wait);
597 if (unlikely(!xsk_is_bound(xs)))
602 if (pool->cached_need_wakeup) {
604 xsk_wakeup(xs, pool->cached_need_wakeup);
606 /* Poll needs to drive Tx also in copy mode */
610 if (xs->rx && !xskq_prod_is_empty(xs->rx))
611 mask |= EPOLLIN | EPOLLRDNORM;
612 if (xs->tx && xsk_tx_writeable(xs))
613 mask |= EPOLLOUT | EPOLLWRNORM;
618 static int xsk_init_queue(u32 entries, struct xsk_queue **queue,
623 if (entries == 0 || *queue || !is_power_of_2(entries))
626 q = xskq_create(entries, umem_queue);
630 /* Make sure queue is ready before it can be seen by others */
632 WRITE_ONCE(*queue, q);
636 static void xsk_unbind_dev(struct xdp_sock *xs)
638 struct net_device *dev = xs->dev;
640 if (xs->state != XSK_BOUND)
642 WRITE_ONCE(xs->state, XSK_UNBOUND);
644 /* Wait for driver to stop using the xdp socket. */
645 xp_del_xsk(xs->pool, xs);
651 static struct xsk_map *xsk_get_map_list_entry(struct xdp_sock *xs,
652 struct xdp_sock ***map_entry)
654 struct xsk_map *map = NULL;
655 struct xsk_map_node *node;
659 spin_lock_bh(&xs->map_list_lock);
660 node = list_first_entry_or_null(&xs->map_list, struct xsk_map_node,
663 bpf_map_inc(&node->map->map);
665 *map_entry = node->map_entry;
667 spin_unlock_bh(&xs->map_list_lock);
671 static void xsk_delete_from_maps(struct xdp_sock *xs)
673 /* This function removes the current XDP socket from all the
674 * maps it resides in. We need to take extra care here, due to
675 * the two locks involved. Each map has a lock synchronizing
676 * updates to the entries, and each socket has a lock that
677 * synchronizes access to the list of maps (map_list). For
678 * deadlock avoidance the locks need to be taken in the order
679 * "map lock"->"socket map list lock". We start off by
680 * accessing the socket map list, and take a reference to the
681 * map to guarantee existence between the
682 * xsk_get_map_list_entry() and xsk_map_try_sock_delete()
683 * calls. Then we ask the map to remove the socket, which
684 * tries to remove the socket from the map. Note that there
685 * might be updates to the map between
686 * xsk_get_map_list_entry() and xsk_map_try_sock_delete().
688 struct xdp_sock **map_entry = NULL;
691 while ((map = xsk_get_map_list_entry(xs, &map_entry))) {
692 xsk_map_try_sock_delete(map, xs, map_entry);
693 bpf_map_put(&map->map);
697 static int xsk_release(struct socket *sock)
699 struct sock *sk = sock->sk;
700 struct xdp_sock *xs = xdp_sk(sk);
708 mutex_lock(&net->xdp.lock);
709 sk_del_node_init_rcu(sk);
710 mutex_unlock(&net->xdp.lock);
713 sock_prot_inuse_add(net, sk->sk_prot, -1);
716 xsk_delete_from_maps(xs);
717 mutex_lock(&xs->mutex);
719 mutex_unlock(&xs->mutex);
721 xskq_destroy(xs->rx);
722 xskq_destroy(xs->tx);
723 xskq_destroy(xs->fq_tmp);
724 xskq_destroy(xs->cq_tmp);
729 sk_refcnt_debug_release(sk);
735 static struct socket *xsk_lookup_xsk_from_fd(int fd)
740 sock = sockfd_lookup(fd, &err);
742 return ERR_PTR(-ENOTSOCK);
744 if (sock->sk->sk_family != PF_XDP) {
746 return ERR_PTR(-ENOPROTOOPT);
752 static bool xsk_validate_queues(struct xdp_sock *xs)
754 return xs->fq_tmp && xs->cq_tmp;
757 static int xsk_bind(struct socket *sock, struct sockaddr *addr, int addr_len)
759 struct sockaddr_xdp *sxdp = (struct sockaddr_xdp *)addr;
760 struct sock *sk = sock->sk;
761 struct xdp_sock *xs = xdp_sk(sk);
762 struct net_device *dev;
766 if (addr_len < sizeof(struct sockaddr_xdp))
768 if (sxdp->sxdp_family != AF_XDP)
771 flags = sxdp->sxdp_flags;
772 if (flags & ~(XDP_SHARED_UMEM | XDP_COPY | XDP_ZEROCOPY |
773 XDP_USE_NEED_WAKEUP))
777 mutex_lock(&xs->mutex);
778 if (xs->state != XSK_READY) {
783 dev = dev_get_by_index(sock_net(sk), sxdp->sxdp_ifindex);
789 if (!xs->rx && !xs->tx) {
794 qid = sxdp->sxdp_queue_id;
796 if (flags & XDP_SHARED_UMEM) {
797 struct xdp_sock *umem_xs;
800 if ((flags & XDP_COPY) || (flags & XDP_ZEROCOPY) ||
801 (flags & XDP_USE_NEED_WAKEUP)) {
802 /* Cannot specify flags for shared sockets. */
808 /* We have already our own. */
813 sock = xsk_lookup_xsk_from_fd(sxdp->sxdp_shared_umem_fd);
819 umem_xs = xdp_sk(sock->sk);
820 if (!xsk_is_bound(umem_xs)) {
826 if (umem_xs->queue_id != qid || umem_xs->dev != dev) {
827 /* Share the umem with another socket on another qid
830 xs->pool = xp_create_and_assign_umem(xs,
838 err = xp_assign_dev_shared(xs->pool, umem_xs->umem,
841 xp_destroy(xs->pool);
847 /* Share the buffer pool with the other socket. */
848 if (xs->fq_tmp || xs->cq_tmp) {
849 /* Do not allow setting your own fq or cq. */
855 xp_get_pool(umem_xs->pool);
856 xs->pool = umem_xs->pool;
859 xdp_get_umem(umem_xs->umem);
860 WRITE_ONCE(xs->umem, umem_xs->umem);
862 } else if (!xs->umem || !xsk_validate_queues(xs)) {
866 /* This xsk has its own umem. */
867 xs->pool = xp_create_and_assign_umem(xs, xs->umem);
873 err = xp_assign_dev(xs->pool, dev, qid, flags);
875 xp_destroy(xs->pool);
882 xs->zc = xs->umem->zc;
884 xp_add_xsk(xs->pool, xs);
890 /* Matches smp_rmb() in bind() for shared umem
891 * sockets, and xsk_is_bound().
894 WRITE_ONCE(xs->state, XSK_BOUND);
897 mutex_unlock(&xs->mutex);
902 struct xdp_umem_reg_v1 {
903 __u64 addr; /* Start of packet data area */
904 __u64 len; /* Length of packet data area */
909 static int xsk_setsockopt(struct socket *sock, int level, int optname,
910 sockptr_t optval, unsigned int optlen)
912 struct sock *sk = sock->sk;
913 struct xdp_sock *xs = xdp_sk(sk);
916 if (level != SOL_XDP)
923 struct xsk_queue **q;
926 if (optlen < sizeof(entries))
928 if (copy_from_sockptr(&entries, optval, sizeof(entries)))
931 mutex_lock(&xs->mutex);
932 if (xs->state != XSK_READY) {
933 mutex_unlock(&xs->mutex);
936 q = (optname == XDP_TX_RING) ? &xs->tx : &xs->rx;
937 err = xsk_init_queue(entries, q, false);
938 if (!err && optname == XDP_TX_RING)
939 /* Tx needs to be explicitly woken up the first time */
940 xs->tx->ring->flags |= XDP_RING_NEED_WAKEUP;
941 mutex_unlock(&xs->mutex);
946 size_t mr_size = sizeof(struct xdp_umem_reg);
947 struct xdp_umem_reg mr = {};
948 struct xdp_umem *umem;
950 if (optlen < sizeof(struct xdp_umem_reg_v1))
952 else if (optlen < sizeof(mr))
953 mr_size = sizeof(struct xdp_umem_reg_v1);
955 if (copy_from_sockptr(&mr, optval, mr_size))
958 mutex_lock(&xs->mutex);
959 if (xs->state != XSK_READY || xs->umem) {
960 mutex_unlock(&xs->mutex);
964 umem = xdp_umem_create(&mr);
966 mutex_unlock(&xs->mutex);
967 return PTR_ERR(umem);
970 /* Make sure umem is ready before it can be seen by others */
972 WRITE_ONCE(xs->umem, umem);
973 mutex_unlock(&xs->mutex);
976 case XDP_UMEM_FILL_RING:
977 case XDP_UMEM_COMPLETION_RING:
979 struct xsk_queue **q;
982 if (copy_from_sockptr(&entries, optval, sizeof(entries)))
985 mutex_lock(&xs->mutex);
986 if (xs->state != XSK_READY) {
987 mutex_unlock(&xs->mutex);
991 q = (optname == XDP_UMEM_FILL_RING) ? &xs->fq_tmp :
993 err = xsk_init_queue(entries, q, true);
994 mutex_unlock(&xs->mutex);
1001 return -ENOPROTOOPT;
1004 static void xsk_enter_rxtx_offsets(struct xdp_ring_offset_v1 *ring)
1006 ring->producer = offsetof(struct xdp_rxtx_ring, ptrs.producer);
1007 ring->consumer = offsetof(struct xdp_rxtx_ring, ptrs.consumer);
1008 ring->desc = offsetof(struct xdp_rxtx_ring, desc);
1011 static void xsk_enter_umem_offsets(struct xdp_ring_offset_v1 *ring)
1013 ring->producer = offsetof(struct xdp_umem_ring, ptrs.producer);
1014 ring->consumer = offsetof(struct xdp_umem_ring, ptrs.consumer);
1015 ring->desc = offsetof(struct xdp_umem_ring, desc);
1018 struct xdp_statistics_v1 {
1020 __u64 rx_invalid_descs;
1021 __u64 tx_invalid_descs;
1024 static int xsk_getsockopt(struct socket *sock, int level, int optname,
1025 char __user *optval, int __user *optlen)
1027 struct sock *sk = sock->sk;
1028 struct xdp_sock *xs = xdp_sk(sk);
1031 if (level != SOL_XDP)
1032 return -ENOPROTOOPT;
1034 if (get_user(len, optlen))
1040 case XDP_STATISTICS:
1042 struct xdp_statistics stats = {};
1043 bool extra_stats = true;
1046 if (len < sizeof(struct xdp_statistics_v1)) {
1048 } else if (len < sizeof(stats)) {
1049 extra_stats = false;
1050 stats_size = sizeof(struct xdp_statistics_v1);
1052 stats_size = sizeof(stats);
1055 mutex_lock(&xs->mutex);
1056 stats.rx_dropped = xs->rx_dropped;
1058 stats.rx_ring_full = xs->rx_queue_full;
1059 stats.rx_fill_ring_empty_descs =
1060 xs->pool ? xskq_nb_queue_empty_descs(xs->pool->fq) : 0;
1061 stats.tx_ring_empty_descs = xskq_nb_queue_empty_descs(xs->tx);
1063 stats.rx_dropped += xs->rx_queue_full;
1065 stats.rx_invalid_descs = xskq_nb_invalid_descs(xs->rx);
1066 stats.tx_invalid_descs = xskq_nb_invalid_descs(xs->tx);
1067 mutex_unlock(&xs->mutex);
1069 if (copy_to_user(optval, &stats, stats_size))
1071 if (put_user(stats_size, optlen))
1076 case XDP_MMAP_OFFSETS:
1078 struct xdp_mmap_offsets off;
1079 struct xdp_mmap_offsets_v1 off_v1;
1080 bool flags_supported = true;
1083 if (len < sizeof(off_v1))
1085 else if (len < sizeof(off))
1086 flags_supported = false;
1088 if (flags_supported) {
1089 /* xdp_ring_offset is identical to xdp_ring_offset_v1
1090 * except for the flags field added to the end.
1092 xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *)
1094 xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *)
1096 xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *)
1098 xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *)
1100 off.rx.flags = offsetof(struct xdp_rxtx_ring,
1102 off.tx.flags = offsetof(struct xdp_rxtx_ring,
1104 off.fr.flags = offsetof(struct xdp_umem_ring,
1106 off.cr.flags = offsetof(struct xdp_umem_ring,
1112 xsk_enter_rxtx_offsets(&off_v1.rx);
1113 xsk_enter_rxtx_offsets(&off_v1.tx);
1114 xsk_enter_umem_offsets(&off_v1.fr);
1115 xsk_enter_umem_offsets(&off_v1.cr);
1117 len = sizeof(off_v1);
1121 if (copy_to_user(optval, to_copy, len))
1123 if (put_user(len, optlen))
1130 struct xdp_options opts = {};
1132 if (len < sizeof(opts))
1135 mutex_lock(&xs->mutex);
1137 opts.flags |= XDP_OPTIONS_ZEROCOPY;
1138 mutex_unlock(&xs->mutex);
1141 if (copy_to_user(optval, &opts, len))
1143 if (put_user(len, optlen))
1155 static int xsk_mmap(struct file *file, struct socket *sock,
1156 struct vm_area_struct *vma)
1158 loff_t offset = (loff_t)vma->vm_pgoff << PAGE_SHIFT;
1159 unsigned long size = vma->vm_end - vma->vm_start;
1160 struct xdp_sock *xs = xdp_sk(sock->sk);
1161 struct xsk_queue *q = NULL;
1165 if (READ_ONCE(xs->state) != XSK_READY)
1168 if (offset == XDP_PGOFF_RX_RING) {
1169 q = READ_ONCE(xs->rx);
1170 } else if (offset == XDP_PGOFF_TX_RING) {
1171 q = READ_ONCE(xs->tx);
1173 /* Matches the smp_wmb() in XDP_UMEM_REG */
1175 if (offset == XDP_UMEM_PGOFF_FILL_RING)
1176 q = READ_ONCE(xs->fq_tmp);
1177 else if (offset == XDP_UMEM_PGOFF_COMPLETION_RING)
1178 q = READ_ONCE(xs->cq_tmp);
1184 /* Matches the smp_wmb() in xsk_init_queue */
1186 qpg = virt_to_head_page(q->ring);
1187 if (size > page_size(qpg))
1190 pfn = virt_to_phys(q->ring) >> PAGE_SHIFT;
1191 return remap_pfn_range(vma, vma->vm_start, pfn,
1192 size, vma->vm_page_prot);
1195 static int xsk_notifier(struct notifier_block *this,
1196 unsigned long msg, void *ptr)
1198 struct net_device *dev = netdev_notifier_info_to_dev(ptr);
1199 struct net *net = dev_net(dev);
1203 case NETDEV_UNREGISTER:
1204 mutex_lock(&net->xdp.lock);
1205 sk_for_each(sk, &net->xdp.list) {
1206 struct xdp_sock *xs = xdp_sk(sk);
1208 mutex_lock(&xs->mutex);
1209 if (xs->dev == dev) {
1210 sk->sk_err = ENETDOWN;
1211 if (!sock_flag(sk, SOCK_DEAD))
1212 sk->sk_error_report(sk);
1216 /* Clear device references. */
1217 xp_clear_dev(xs->pool);
1219 mutex_unlock(&xs->mutex);
1221 mutex_unlock(&net->xdp.lock);
1227 static struct proto xsk_proto = {
1229 .owner = THIS_MODULE,
1230 .obj_size = sizeof(struct xdp_sock),
1233 static const struct proto_ops xsk_proto_ops = {
1235 .owner = THIS_MODULE,
1236 .release = xsk_release,
1238 .connect = sock_no_connect,
1239 .socketpair = sock_no_socketpair,
1240 .accept = sock_no_accept,
1241 .getname = sock_no_getname,
1243 .ioctl = sock_no_ioctl,
1244 .listen = sock_no_listen,
1245 .shutdown = sock_no_shutdown,
1246 .setsockopt = xsk_setsockopt,
1247 .getsockopt = xsk_getsockopt,
1248 .sendmsg = xsk_sendmsg,
1249 .recvmsg = xsk_recvmsg,
1251 .sendpage = sock_no_sendpage,
1254 static void xsk_destruct(struct sock *sk)
1256 struct xdp_sock *xs = xdp_sk(sk);
1258 if (!sock_flag(sk, SOCK_DEAD))
1261 if (!xp_put_pool(xs->pool))
1262 xdp_put_umem(xs->umem, !xs->pool);
1264 sk_refcnt_debug_dec(sk);
1267 static int xsk_create(struct net *net, struct socket *sock, int protocol,
1270 struct xdp_sock *xs;
1273 if (!ns_capable(net->user_ns, CAP_NET_RAW))
1275 if (sock->type != SOCK_RAW)
1276 return -ESOCKTNOSUPPORT;
1279 return -EPROTONOSUPPORT;
1281 sock->state = SS_UNCONNECTED;
1283 sk = sk_alloc(net, PF_XDP, GFP_KERNEL, &xsk_proto, kern);
1287 sock->ops = &xsk_proto_ops;
1289 sock_init_data(sock, sk);
1291 sk->sk_family = PF_XDP;
1293 sk->sk_destruct = xsk_destruct;
1294 sk_refcnt_debug_inc(sk);
1296 sock_set_flag(sk, SOCK_RCU_FREE);
1299 xs->state = XSK_READY;
1300 mutex_init(&xs->mutex);
1301 spin_lock_init(&xs->rx_lock);
1302 spin_lock_init(&xs->tx_completion_lock);
1304 INIT_LIST_HEAD(&xs->map_list);
1305 spin_lock_init(&xs->map_list_lock);
1307 mutex_lock(&net->xdp.lock);
1308 sk_add_node_rcu(sk, &net->xdp.list);
1309 mutex_unlock(&net->xdp.lock);
1312 sock_prot_inuse_add(net, &xsk_proto, 1);
1318 static const struct net_proto_family xsk_family_ops = {
1320 .create = xsk_create,
1321 .owner = THIS_MODULE,
1324 static struct notifier_block xsk_netdev_notifier = {
1325 .notifier_call = xsk_notifier,
1328 static int __net_init xsk_net_init(struct net *net)
1330 mutex_init(&net->xdp.lock);
1331 INIT_HLIST_HEAD(&net->xdp.list);
1335 static void __net_exit xsk_net_exit(struct net *net)
1337 WARN_ON_ONCE(!hlist_empty(&net->xdp.list));
1340 static struct pernet_operations xsk_net_ops = {
1341 .init = xsk_net_init,
1342 .exit = xsk_net_exit,
1345 static int __init xsk_init(void)
1349 err = proto_register(&xsk_proto, 0 /* no slab */);
1353 err = sock_register(&xsk_family_ops);
1357 err = register_pernet_subsys(&xsk_net_ops);
1361 err = register_netdevice_notifier(&xsk_netdev_notifier);
1365 for_each_possible_cpu(cpu)
1366 INIT_LIST_HEAD(&per_cpu(xskmap_flush_list, cpu));
1370 unregister_pernet_subsys(&xsk_net_ops);
1372 sock_unregister(PF_XDP);
1374 proto_unregister(&xsk_proto);
1379 fs_initcall(xsk_init);