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"
31 #define TX_BATCH_SIZE 16
33 static struct xdp_sock *xdp_sk(struct sock *sk)
35 return (struct xdp_sock *)sk;
38 bool xsk_is_setup_for_bpf_map(struct xdp_sock *xs)
40 return READ_ONCE(xs->rx) && READ_ONCE(xs->umem) &&
41 READ_ONCE(xs->umem->fq);
44 u64 *xsk_umem_peek_addr(struct xdp_umem *umem, u64 *addr)
46 return xskq_peek_addr(umem->fq, addr);
48 EXPORT_SYMBOL(xsk_umem_peek_addr);
50 void xsk_umem_discard_addr(struct xdp_umem *umem)
52 xskq_discard_addr(umem->fq);
54 EXPORT_SYMBOL(xsk_umem_discard_addr);
56 static int __xsk_rcv(struct xdp_sock *xs, struct xdp_buff *xdp, u32 len)
62 if (!xskq_peek_addr(xs->umem->fq, &addr) ||
63 len > xs->umem->chunk_size_nohr) {
68 addr += xs->umem->headroom;
70 buffer = xdp_umem_get_data(xs->umem, addr);
71 memcpy(buffer, xdp->data, len);
72 err = xskq_produce_batch_desc(xs->rx, addr, len);
74 xskq_discard_addr(xs->umem->fq);
83 static int __xsk_rcv_zc(struct xdp_sock *xs, struct xdp_buff *xdp, u32 len)
85 int err = xskq_produce_batch_desc(xs->rx, (u64)xdp->handle, len);
95 int xsk_rcv(struct xdp_sock *xs, struct xdp_buff *xdp)
99 if (xs->dev != xdp->rxq->dev || xs->queue_id != xdp->rxq->queue_index)
102 len = xdp->data_end - xdp->data;
104 return (xdp->rxq->mem.type == MEM_TYPE_ZERO_COPY) ?
105 __xsk_rcv_zc(xs, xdp, len) : __xsk_rcv(xs, xdp, len);
108 void xsk_flush(struct xdp_sock *xs)
110 xskq_produce_flush_desc(xs->rx);
111 xs->sk.sk_data_ready(&xs->sk);
114 int xsk_generic_rcv(struct xdp_sock *xs, struct xdp_buff *xdp)
116 u32 len = xdp->data_end - xdp->data;
121 if (xs->dev != xdp->rxq->dev || xs->queue_id != xdp->rxq->queue_index)
124 if (!xskq_peek_addr(xs->umem->fq, &addr) ||
125 len > xs->umem->chunk_size_nohr) {
130 addr += xs->umem->headroom;
132 buffer = xdp_umem_get_data(xs->umem, addr);
133 memcpy(buffer, xdp->data, len);
134 err = xskq_produce_batch_desc(xs->rx, addr, len);
136 xskq_discard_addr(xs->umem->fq);
145 void xsk_umem_complete_tx(struct xdp_umem *umem, u32 nb_entries)
147 xskq_produce_flush_addr_n(umem->cq, nb_entries);
149 EXPORT_SYMBOL(xsk_umem_complete_tx);
151 void xsk_umem_consume_tx_done(struct xdp_umem *umem)
156 list_for_each_entry_rcu(xs, &umem->xsk_list, list) {
157 xs->sk.sk_write_space(&xs->sk);
161 EXPORT_SYMBOL(xsk_umem_consume_tx_done);
163 bool xsk_umem_consume_tx(struct xdp_umem *umem, dma_addr_t *dma, u32 *len)
165 struct xdp_desc desc;
169 list_for_each_entry_rcu(xs, &umem->xsk_list, list) {
170 if (!xskq_peek_desc(xs->tx, &desc))
173 if (xskq_produce_addr_lazy(umem->cq, desc.addr))
176 *dma = xdp_umem_get_dma(umem, desc.addr);
179 xskq_discard_desc(xs->tx);
188 EXPORT_SYMBOL(xsk_umem_consume_tx);
190 static int xsk_zc_xmit(struct sock *sk)
192 struct xdp_sock *xs = xdp_sk(sk);
193 struct net_device *dev = xs->dev;
195 return dev->netdev_ops->ndo_xsk_async_xmit(dev, xs->queue_id);
198 static void xsk_destruct_skb(struct sk_buff *skb)
200 u64 addr = (u64)(long)skb_shinfo(skb)->destructor_arg;
201 struct xdp_sock *xs = xdp_sk(skb->sk);
204 spin_lock_irqsave(&xs->tx_completion_lock, flags);
205 WARN_ON_ONCE(xskq_produce_addr(xs->umem->cq, addr));
206 spin_unlock_irqrestore(&xs->tx_completion_lock, flags);
211 static int xsk_generic_xmit(struct sock *sk, struct msghdr *m,
214 u32 max_batch = TX_BATCH_SIZE;
215 struct xdp_sock *xs = xdp_sk(sk);
216 bool sent_frame = false;
217 struct xdp_desc desc;
221 mutex_lock(&xs->mutex);
223 while (xskq_peek_desc(xs->tx, &desc)) {
228 if (max_batch-- == 0) {
233 if (xskq_reserve_addr(xs->umem->cq))
236 if (xs->queue_id >= xs->dev->real_num_tx_queues)
240 skb = sock_alloc_send_skb(sk, len, 1, &err);
241 if (unlikely(!skb)) {
248 buffer = xdp_umem_get_data(xs->umem, addr);
249 err = skb_store_bits(skb, 0, buffer, len);
256 skb->priority = sk->sk_priority;
257 skb->mark = sk->sk_mark;
258 skb_shinfo(skb)->destructor_arg = (void *)(long)addr;
259 skb->destructor = xsk_destruct_skb;
261 err = dev_direct_xmit(skb, xs->queue_id);
262 xskq_discard_desc(xs->tx);
263 /* Ignore NET_XMIT_CN as packet might have been sent */
264 if (err == NET_XMIT_DROP || err == NETDEV_TX_BUSY) {
265 /* SKB completed but not sent */
275 sk->sk_write_space(sk);
277 mutex_unlock(&xs->mutex);
281 static int xsk_sendmsg(struct socket *sock, struct msghdr *m, size_t total_len)
283 bool need_wait = !(m->msg_flags & MSG_DONTWAIT);
284 struct sock *sk = sock->sk;
285 struct xdp_sock *xs = xdp_sk(sk);
287 if (unlikely(!xs->dev))
289 if (unlikely(!(xs->dev->flags & IFF_UP)))
291 if (unlikely(!xs->tx))
296 return (xs->zc) ? xsk_zc_xmit(sk) : xsk_generic_xmit(sk, m, total_len);
299 static unsigned int xsk_poll(struct file *file, struct socket *sock,
300 struct poll_table_struct *wait)
302 unsigned int mask = datagram_poll(file, sock, wait);
303 struct sock *sk = sock->sk;
304 struct xdp_sock *xs = xdp_sk(sk);
306 if (xs->rx && !xskq_empty_desc(xs->rx))
307 mask |= POLLIN | POLLRDNORM;
308 if (xs->tx && !xskq_full_desc(xs->tx))
309 mask |= POLLOUT | POLLWRNORM;
314 static int xsk_init_queue(u32 entries, struct xsk_queue **queue,
319 if (entries == 0 || *queue || !is_power_of_2(entries))
322 q = xskq_create(entries, umem_queue);
326 /* Make sure queue is ready before it can be seen by others */
332 static int xsk_release(struct socket *sock)
334 struct sock *sk = sock->sk;
335 struct xdp_sock *xs = xdp_sk(sk);
344 sock_prot_inuse_add(net, sk->sk_prot, -1);
348 /* Wait for driver to stop using the xdp socket. */
357 sk_refcnt_debug_release(sk);
363 static struct socket *xsk_lookup_xsk_from_fd(int fd)
368 sock = sockfd_lookup(fd, &err);
370 return ERR_PTR(-ENOTSOCK);
372 if (sock->sk->sk_family != PF_XDP) {
374 return ERR_PTR(-ENOPROTOOPT);
380 static int xsk_bind(struct socket *sock, struct sockaddr *addr, int addr_len)
382 struct sockaddr_xdp *sxdp = (struct sockaddr_xdp *)addr;
383 struct sock *sk = sock->sk;
384 struct xdp_sock *xs = xdp_sk(sk);
385 struct net_device *dev;
389 if (addr_len < sizeof(struct sockaddr_xdp))
391 if (sxdp->sxdp_family != AF_XDP)
394 mutex_lock(&xs->mutex);
400 dev = dev_get_by_index(sock_net(sk), sxdp->sxdp_ifindex);
406 if (!xs->rx && !xs->tx) {
411 qid = sxdp->sxdp_queue_id;
413 if ((xs->rx && qid >= dev->real_num_rx_queues) ||
414 (xs->tx && qid >= dev->real_num_tx_queues)) {
419 flags = sxdp->sxdp_flags;
421 if (flags & XDP_SHARED_UMEM) {
422 struct xdp_sock *umem_xs;
425 if ((flags & XDP_COPY) || (flags & XDP_ZEROCOPY)) {
426 /* Cannot specify flags for shared sockets. */
432 /* We have already our own. */
437 sock = xsk_lookup_xsk_from_fd(sxdp->sxdp_shared_umem_fd);
443 umem_xs = xdp_sk(sock->sk);
444 if (!umem_xs->umem) {
445 /* No umem to inherit. */
449 } else if (umem_xs->dev != dev || umem_xs->queue_id != qid) {
455 xdp_get_umem(umem_xs->umem);
456 xs->umem = umem_xs->umem;
458 } else if (!xs->umem || !xdp_umem_validate_queues(xs->umem)) {
462 /* This xsk has its own umem. */
463 xskq_set_umem(xs->umem->fq, &xs->umem->props);
464 xskq_set_umem(xs->umem->cq, &xs->umem->props);
466 err = xdp_umem_assign_dev(xs->umem, dev, qid, flags);
472 xs->zc = xs->umem->zc;
474 xskq_set_umem(xs->rx, &xs->umem->props);
475 xskq_set_umem(xs->tx, &xs->umem->props);
476 xdp_add_sk_umem(xs->umem, xs);
482 mutex_unlock(&xs->mutex);
486 static int xsk_setsockopt(struct socket *sock, int level, int optname,
487 char __user *optval, unsigned int optlen)
489 struct sock *sk = sock->sk;
490 struct xdp_sock *xs = xdp_sk(sk);
493 if (level != SOL_XDP)
500 struct xsk_queue **q;
503 if (optlen < sizeof(entries))
505 if (copy_from_user(&entries, optval, sizeof(entries)))
508 mutex_lock(&xs->mutex);
509 q = (optname == XDP_TX_RING) ? &xs->tx : &xs->rx;
510 err = xsk_init_queue(entries, q, false);
511 mutex_unlock(&xs->mutex);
516 struct xdp_umem_reg mr;
517 struct xdp_umem *umem;
519 if (copy_from_user(&mr, optval, sizeof(mr)))
522 mutex_lock(&xs->mutex);
524 mutex_unlock(&xs->mutex);
528 umem = xdp_umem_create(&mr);
530 mutex_unlock(&xs->mutex);
531 return PTR_ERR(umem);
534 /* Make sure umem is ready before it can be seen by others */
537 mutex_unlock(&xs->mutex);
540 case XDP_UMEM_FILL_RING:
541 case XDP_UMEM_COMPLETION_RING:
543 struct xsk_queue **q;
546 if (copy_from_user(&entries, optval, sizeof(entries)))
549 mutex_lock(&xs->mutex);
551 mutex_unlock(&xs->mutex);
555 q = (optname == XDP_UMEM_FILL_RING) ? &xs->umem->fq :
557 err = xsk_init_queue(entries, q, true);
558 mutex_unlock(&xs->mutex);
568 static int xsk_getsockopt(struct socket *sock, int level, int optname,
569 char __user *optval, int __user *optlen)
571 struct sock *sk = sock->sk;
572 struct xdp_sock *xs = xdp_sk(sk);
575 if (level != SOL_XDP)
578 if (get_user(len, optlen))
586 struct xdp_statistics stats;
588 if (len < sizeof(stats))
591 mutex_lock(&xs->mutex);
592 stats.rx_dropped = xs->rx_dropped;
593 stats.rx_invalid_descs = xskq_nb_invalid_descs(xs->rx);
594 stats.tx_invalid_descs = xskq_nb_invalid_descs(xs->tx);
595 mutex_unlock(&xs->mutex);
597 if (copy_to_user(optval, &stats, sizeof(stats)))
599 if (put_user(sizeof(stats), optlen))
604 case XDP_MMAP_OFFSETS:
606 struct xdp_mmap_offsets off;
608 if (len < sizeof(off))
611 off.rx.producer = offsetof(struct xdp_rxtx_ring, ptrs.producer);
612 off.rx.consumer = offsetof(struct xdp_rxtx_ring, ptrs.consumer);
613 off.rx.desc = offsetof(struct xdp_rxtx_ring, desc);
614 off.tx.producer = offsetof(struct xdp_rxtx_ring, ptrs.producer);
615 off.tx.consumer = offsetof(struct xdp_rxtx_ring, ptrs.consumer);
616 off.tx.desc = offsetof(struct xdp_rxtx_ring, desc);
618 off.fr.producer = offsetof(struct xdp_umem_ring, ptrs.producer);
619 off.fr.consumer = offsetof(struct xdp_umem_ring, ptrs.consumer);
620 off.fr.desc = offsetof(struct xdp_umem_ring, desc);
621 off.cr.producer = offsetof(struct xdp_umem_ring, ptrs.producer);
622 off.cr.consumer = offsetof(struct xdp_umem_ring, ptrs.consumer);
623 off.cr.desc = offsetof(struct xdp_umem_ring, desc);
626 if (copy_to_user(optval, &off, len))
628 if (put_user(len, optlen))
640 static int xsk_mmap(struct file *file, struct socket *sock,
641 struct vm_area_struct *vma)
643 loff_t offset = (loff_t)vma->vm_pgoff << PAGE_SHIFT;
644 unsigned long size = vma->vm_end - vma->vm_start;
645 struct xdp_sock *xs = xdp_sk(sock->sk);
646 struct xsk_queue *q = NULL;
647 struct xdp_umem *umem;
651 if (offset == XDP_PGOFF_RX_RING) {
652 q = READ_ONCE(xs->rx);
653 } else if (offset == XDP_PGOFF_TX_RING) {
654 q = READ_ONCE(xs->tx);
656 umem = READ_ONCE(xs->umem);
660 if (offset == XDP_UMEM_PGOFF_FILL_RING)
661 q = READ_ONCE(umem->fq);
662 else if (offset == XDP_UMEM_PGOFF_COMPLETION_RING)
663 q = READ_ONCE(umem->cq);
669 qpg = virt_to_head_page(q->ring);
670 if (size > (PAGE_SIZE << compound_order(qpg)))
673 pfn = virt_to_phys(q->ring) >> PAGE_SHIFT;
674 return remap_pfn_range(vma, vma->vm_start, pfn,
675 size, vma->vm_page_prot);
678 static struct proto xsk_proto = {
680 .owner = THIS_MODULE,
681 .obj_size = sizeof(struct xdp_sock),
684 static const struct proto_ops xsk_proto_ops = {
686 .owner = THIS_MODULE,
687 .release = xsk_release,
689 .connect = sock_no_connect,
690 .socketpair = sock_no_socketpair,
691 .accept = sock_no_accept,
692 .getname = sock_no_getname,
694 .ioctl = sock_no_ioctl,
695 .listen = sock_no_listen,
696 .shutdown = sock_no_shutdown,
697 .setsockopt = xsk_setsockopt,
698 .getsockopt = xsk_getsockopt,
699 .sendmsg = xsk_sendmsg,
700 .recvmsg = sock_no_recvmsg,
702 .sendpage = sock_no_sendpage,
705 static void xsk_destruct(struct sock *sk)
707 struct xdp_sock *xs = xdp_sk(sk);
709 if (!sock_flag(sk, SOCK_DEAD))
712 xskq_destroy(xs->rx);
713 xskq_destroy(xs->tx);
714 xdp_del_sk_umem(xs->umem, xs);
715 xdp_put_umem(xs->umem);
717 sk_refcnt_debug_dec(sk);
720 static int xsk_create(struct net *net, struct socket *sock, int protocol,
726 if (!ns_capable(net->user_ns, CAP_NET_RAW))
728 if (sock->type != SOCK_RAW)
729 return -ESOCKTNOSUPPORT;
732 return -EPROTONOSUPPORT;
734 sock->state = SS_UNCONNECTED;
736 sk = sk_alloc(net, PF_XDP, GFP_KERNEL, &xsk_proto, kern);
740 sock->ops = &xsk_proto_ops;
742 sock_init_data(sock, sk);
744 sk->sk_family = PF_XDP;
746 sk->sk_destruct = xsk_destruct;
747 sk_refcnt_debug_inc(sk);
750 mutex_init(&xs->mutex);
751 spin_lock_init(&xs->tx_completion_lock);
754 sock_prot_inuse_add(net, &xsk_proto, 1);
760 static const struct net_proto_family xsk_family_ops = {
762 .create = xsk_create,
763 .owner = THIS_MODULE,
766 static int __init xsk_init(void)
770 err = proto_register(&xsk_proto, 0 /* no slab */);
774 err = sock_register(&xsk_family_ops);
781 proto_unregister(&xsk_proto);
786 fs_initcall(xsk_init);