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 <linux/vmalloc.h>
26 #include <net/xdp_sock_drv.h>
27 #include <net/busy_poll.h>
28 #include <net/netdev_rx_queue.h>
31 #include "xsk_queue.h"
35 #define TX_BATCH_SIZE 32
37 static DEFINE_PER_CPU(struct list_head, xskmap_flush_list);
39 void xsk_set_rx_need_wakeup(struct xsk_buff_pool *pool)
41 if (pool->cached_need_wakeup & XDP_WAKEUP_RX)
44 pool->fq->ring->flags |= XDP_RING_NEED_WAKEUP;
45 pool->cached_need_wakeup |= XDP_WAKEUP_RX;
47 EXPORT_SYMBOL(xsk_set_rx_need_wakeup);
49 void xsk_set_tx_need_wakeup(struct xsk_buff_pool *pool)
53 if (pool->cached_need_wakeup & XDP_WAKEUP_TX)
57 list_for_each_entry_rcu(xs, &pool->xsk_tx_list, tx_list) {
58 xs->tx->ring->flags |= XDP_RING_NEED_WAKEUP;
62 pool->cached_need_wakeup |= XDP_WAKEUP_TX;
64 EXPORT_SYMBOL(xsk_set_tx_need_wakeup);
66 void xsk_clear_rx_need_wakeup(struct xsk_buff_pool *pool)
68 if (!(pool->cached_need_wakeup & XDP_WAKEUP_RX))
71 pool->fq->ring->flags &= ~XDP_RING_NEED_WAKEUP;
72 pool->cached_need_wakeup &= ~XDP_WAKEUP_RX;
74 EXPORT_SYMBOL(xsk_clear_rx_need_wakeup);
76 void xsk_clear_tx_need_wakeup(struct xsk_buff_pool *pool)
80 if (!(pool->cached_need_wakeup & XDP_WAKEUP_TX))
84 list_for_each_entry_rcu(xs, &pool->xsk_tx_list, tx_list) {
85 xs->tx->ring->flags &= ~XDP_RING_NEED_WAKEUP;
89 pool->cached_need_wakeup &= ~XDP_WAKEUP_TX;
91 EXPORT_SYMBOL(xsk_clear_tx_need_wakeup);
93 bool xsk_uses_need_wakeup(struct xsk_buff_pool *pool)
95 return pool->uses_need_wakeup;
97 EXPORT_SYMBOL(xsk_uses_need_wakeup);
99 struct xsk_buff_pool *xsk_get_pool_from_qid(struct net_device *dev,
102 if (queue_id < dev->real_num_rx_queues)
103 return dev->_rx[queue_id].pool;
104 if (queue_id < dev->real_num_tx_queues)
105 return dev->_tx[queue_id].pool;
109 EXPORT_SYMBOL(xsk_get_pool_from_qid);
111 void xsk_clear_pool_at_qid(struct net_device *dev, u16 queue_id)
113 if (queue_id < dev->num_rx_queues)
114 dev->_rx[queue_id].pool = NULL;
115 if (queue_id < dev->num_tx_queues)
116 dev->_tx[queue_id].pool = NULL;
119 /* The buffer pool is stored both in the _rx struct and the _tx struct as we do
120 * not know if the device has more tx queues than rx, or the opposite.
121 * This might also change during run time.
123 int xsk_reg_pool_at_qid(struct net_device *dev, struct xsk_buff_pool *pool,
126 if (queue_id >= max_t(unsigned int,
127 dev->real_num_rx_queues,
128 dev->real_num_tx_queues))
131 if (queue_id < dev->real_num_rx_queues)
132 dev->_rx[queue_id].pool = pool;
133 if (queue_id < dev->real_num_tx_queues)
134 dev->_tx[queue_id].pool = pool;
139 static int __xsk_rcv_zc(struct xdp_sock *xs, struct xdp_buff_xsk *xskb, u32 len,
145 addr = xp_get_handle(xskb);
146 err = xskq_prod_reserve_desc(xs->rx, addr, len, flags);
156 static int xsk_rcv_zc(struct xdp_sock *xs, struct xdp_buff *xdp, u32 len)
158 struct xdp_buff_xsk *xskb = container_of(xdp, struct xdp_buff_xsk, xdp);
159 u32 frags = xdp_buff_has_frags(xdp);
160 struct xdp_buff_xsk *pos, *tmp;
161 struct list_head *xskb_list;
166 contd = XDP_PKT_CONTD;
168 err = __xsk_rcv_zc(xs, xskb, len, contd);
169 if (err || likely(!frags))
172 xskb_list = &xskb->pool->xskb_list;
173 list_for_each_entry_safe(pos, tmp, xskb_list, xskb_list_node) {
174 if (list_is_singular(xskb_list))
176 len = pos->xdp.data_end - pos->xdp.data;
177 err = __xsk_rcv_zc(xs, pos, len, contd);
180 list_del(&pos->xskb_list_node);
187 static void *xsk_copy_xdp_start(struct xdp_buff *from)
189 if (unlikely(xdp_data_meta_unsupported(from)))
192 return from->data_meta;
195 static u32 xsk_copy_xdp(void *to, void **from, u32 to_len,
196 u32 *from_len, skb_frag_t **frag, u32 rem)
201 u32 copy_len = min_t(u32, *from_len, to_len);
203 memcpy(to, *from, copy_len);
208 if (*from_len == copy_len) {
209 *from = skb_frag_address(*frag);
210 *from_len = skb_frag_size((*frag)++);
213 *from_len -= copy_len;
215 if (to_len == copy_len)
223 static int __xsk_rcv(struct xdp_sock *xs, struct xdp_buff *xdp, u32 len)
225 u32 frame_size = xsk_pool_get_rx_frame_size(xs->pool);
226 void *copy_from = xsk_copy_xdp_start(xdp), *copy_to;
227 u32 from_len, meta_len, rem, num_desc;
228 struct xdp_buff_xsk *xskb;
229 struct xdp_buff *xsk_xdp;
232 from_len = xdp->data_end - copy_from;
233 meta_len = xdp->data - copy_from;
234 rem = len + meta_len;
236 if (len <= frame_size && !xdp_buff_has_frags(xdp)) {
239 xsk_xdp = xsk_buff_alloc(xs->pool);
244 memcpy(xsk_xdp->data - meta_len, copy_from, rem);
245 xskb = container_of(xsk_xdp, struct xdp_buff_xsk, xdp);
246 err = __xsk_rcv_zc(xs, xskb, len, 0);
248 xsk_buff_free(xsk_xdp);
255 num_desc = (len - 1) / frame_size + 1;
257 if (!xsk_buff_can_alloc(xs->pool, num_desc)) {
261 if (xskq_prod_nb_free(xs->rx, num_desc) < num_desc) {
266 if (xdp_buff_has_frags(xdp)) {
267 struct skb_shared_info *sinfo;
269 sinfo = xdp_get_shared_info_from_buff(xdp);
270 frag = &sinfo->frags[0];
274 u32 to_len = frame_size + meta_len;
277 xsk_xdp = xsk_buff_alloc(xs->pool);
278 copy_to = xsk_xdp->data - meta_len;
280 copied = xsk_copy_xdp(copy_to, ©_from, to_len, &from_len, &frag, rem);
283 xskb = container_of(xsk_xdp, struct xdp_buff_xsk, xdp);
284 __xsk_rcv_zc(xs, xskb, copied - meta_len, rem ? XDP_PKT_CONTD : 0);
291 static bool xsk_tx_writeable(struct xdp_sock *xs)
293 if (xskq_cons_present_entries(xs->tx) > xs->tx->nentries / 2)
299 static bool xsk_is_bound(struct xdp_sock *xs)
301 if (READ_ONCE(xs->state) == XSK_BOUND) {
302 /* Matches smp_wmb() in bind(). */
309 static int xsk_rcv_check(struct xdp_sock *xs, struct xdp_buff *xdp, u32 len)
311 if (!xsk_is_bound(xs))
314 if (xs->dev != xdp->rxq->dev || xs->queue_id != xdp->rxq->queue_index)
317 if (len > xsk_pool_get_rx_frame_size(xs->pool) && !xs->sg) {
322 sk_mark_napi_id_once_xdp(&xs->sk, xdp);
326 static void xsk_flush(struct xdp_sock *xs)
328 xskq_prod_submit(xs->rx);
329 __xskq_cons_release(xs->pool->fq);
330 sock_def_readable(&xs->sk);
333 int xsk_generic_rcv(struct xdp_sock *xs, struct xdp_buff *xdp)
335 u32 len = xdp_get_buff_len(xdp);
338 spin_lock_bh(&xs->rx_lock);
339 err = xsk_rcv_check(xs, xdp, len);
341 err = __xsk_rcv(xs, xdp, len);
344 spin_unlock_bh(&xs->rx_lock);
348 static int xsk_rcv(struct xdp_sock *xs, struct xdp_buff *xdp)
350 u32 len = xdp_get_buff_len(xdp);
353 err = xsk_rcv_check(xs, xdp, len);
357 if (xdp->rxq->mem.type == MEM_TYPE_XSK_BUFF_POOL) {
358 len = xdp->data_end - xdp->data;
359 return xsk_rcv_zc(xs, xdp, len);
362 err = __xsk_rcv(xs, xdp, len);
364 xdp_return_buff(xdp);
368 int __xsk_map_redirect(struct xdp_sock *xs, struct xdp_buff *xdp)
370 struct list_head *flush_list = this_cpu_ptr(&xskmap_flush_list);
373 err = xsk_rcv(xs, xdp);
377 if (!xs->flush_node.prev)
378 list_add(&xs->flush_node, flush_list);
383 void __xsk_map_flush(void)
385 struct list_head *flush_list = this_cpu_ptr(&xskmap_flush_list);
386 struct xdp_sock *xs, *tmp;
388 list_for_each_entry_safe(xs, tmp, flush_list, flush_node) {
390 __list_del_clearprev(&xs->flush_node);
394 void xsk_tx_completed(struct xsk_buff_pool *pool, u32 nb_entries)
396 xskq_prod_submit_n(pool->cq, nb_entries);
398 EXPORT_SYMBOL(xsk_tx_completed);
400 void xsk_tx_release(struct xsk_buff_pool *pool)
405 list_for_each_entry_rcu(xs, &pool->xsk_tx_list, tx_list) {
406 __xskq_cons_release(xs->tx);
407 if (xsk_tx_writeable(xs))
408 xs->sk.sk_write_space(&xs->sk);
412 EXPORT_SYMBOL(xsk_tx_release);
414 bool xsk_tx_peek_desc(struct xsk_buff_pool *pool, struct xdp_desc *desc)
419 list_for_each_entry_rcu(xs, &pool->xsk_tx_list, tx_list) {
420 if (!xskq_cons_peek_desc(xs->tx, desc, pool)) {
421 if (xskq_has_descs(xs->tx))
422 xskq_cons_release(xs->tx);
426 /* This is the backpressure mechanism for the Tx path.
427 * Reserve space in the completion queue and only proceed
428 * if there is space in it. This avoids having to implement
429 * any buffering in the Tx path.
431 if (xskq_prod_reserve_addr(pool->cq, desc->addr))
434 xskq_cons_release(xs->tx);
443 EXPORT_SYMBOL(xsk_tx_peek_desc);
445 static u32 xsk_tx_peek_release_fallback(struct xsk_buff_pool *pool, u32 max_entries)
447 struct xdp_desc *descs = pool->tx_descs;
450 while (nb_pkts < max_entries && xsk_tx_peek_desc(pool, &descs[nb_pkts]))
453 xsk_tx_release(pool);
457 u32 xsk_tx_peek_release_desc_batch(struct xsk_buff_pool *pool, u32 nb_pkts)
462 if (!list_is_singular(&pool->xsk_tx_list)) {
463 /* Fallback to the non-batched version */
465 return xsk_tx_peek_release_fallback(pool, nb_pkts);
468 xs = list_first_or_null_rcu(&pool->xsk_tx_list, struct xdp_sock, tx_list);
474 nb_pkts = xskq_cons_nb_entries(xs->tx, nb_pkts);
476 /* This is the backpressure mechanism for the Tx path. Try to
477 * reserve space in the completion queue for all packets, but
478 * if there are fewer slots available, just process that many
479 * packets. This avoids having to implement any buffering in
482 nb_pkts = xskq_prod_nb_free(pool->cq, nb_pkts);
486 nb_pkts = xskq_cons_read_desc_batch(xs->tx, pool, nb_pkts);
488 xs->tx->queue_empty_descs++;
492 __xskq_cons_release(xs->tx);
493 xskq_prod_write_addr_batch(pool->cq, pool->tx_descs, nb_pkts);
494 xs->sk.sk_write_space(&xs->sk);
500 EXPORT_SYMBOL(xsk_tx_peek_release_desc_batch);
502 static int xsk_wakeup(struct xdp_sock *xs, u8 flags)
504 struct net_device *dev = xs->dev;
506 return dev->netdev_ops->ndo_xsk_wakeup(dev, xs->queue_id, flags);
509 static int xsk_cq_reserve_addr_locked(struct xdp_sock *xs, u64 addr)
514 spin_lock_irqsave(&xs->pool->cq_lock, flags);
515 ret = xskq_prod_reserve_addr(xs->pool->cq, addr);
516 spin_unlock_irqrestore(&xs->pool->cq_lock, flags);
521 static void xsk_cq_submit_locked(struct xdp_sock *xs, u32 n)
525 spin_lock_irqsave(&xs->pool->cq_lock, flags);
526 xskq_prod_submit_n(xs->pool->cq, n);
527 spin_unlock_irqrestore(&xs->pool->cq_lock, flags);
530 static void xsk_cq_cancel_locked(struct xdp_sock *xs, u32 n)
534 spin_lock_irqsave(&xs->pool->cq_lock, flags);
535 xskq_prod_cancel_n(xs->pool->cq, n);
536 spin_unlock_irqrestore(&xs->pool->cq_lock, flags);
539 static u32 xsk_get_num_desc(struct sk_buff *skb)
541 return skb ? (long)skb_shinfo(skb)->destructor_arg : 0;
544 static void xsk_destruct_skb(struct sk_buff *skb)
546 xsk_cq_submit_locked(xdp_sk(skb->sk), xsk_get_num_desc(skb));
550 static void xsk_set_destructor_arg(struct sk_buff *skb)
552 long num = xsk_get_num_desc(xdp_sk(skb->sk)->skb) + 1;
554 skb_shinfo(skb)->destructor_arg = (void *)num;
557 static void xsk_consume_skb(struct sk_buff *skb)
559 struct xdp_sock *xs = xdp_sk(skb->sk);
561 skb->destructor = sock_wfree;
562 xsk_cq_cancel_locked(xs, xsk_get_num_desc(skb));
563 /* Free skb without triggering the perf drop trace */
568 static void xsk_drop_skb(struct sk_buff *skb)
570 xdp_sk(skb->sk)->tx->invalid_descs += xsk_get_num_desc(skb);
571 xsk_consume_skb(skb);
574 static struct sk_buff *xsk_build_skb_zerocopy(struct xdp_sock *xs,
575 struct xdp_desc *desc)
577 struct xsk_buff_pool *pool = xs->pool;
578 u32 hr, len, ts, offset, copy, copied;
579 struct sk_buff *skb = xs->skb;
586 hr = max(NET_SKB_PAD, L1_CACHE_ALIGN(xs->dev->needed_headroom));
588 skb = sock_alloc_send_skb(&xs->sk, hr, 1, &err);
592 skb_reserve(skb, hr);
597 ts = pool->unaligned ? len : pool->chunk_size;
599 buffer = xsk_buff_raw_get_data(pool, addr);
600 offset = offset_in_page(buffer);
601 addr = buffer - pool->addrs;
603 for (copied = 0, i = skb_shinfo(skb)->nr_frags; copied < len; i++) {
604 if (unlikely(i >= MAX_SKB_FRAGS))
605 return ERR_PTR(-EOVERFLOW);
607 page = pool->umem->pgs[addr >> PAGE_SHIFT];
610 copy = min_t(u32, PAGE_SIZE - offset, len - copied);
611 skb_fill_page_desc(skb, i, page, offset, copy);
619 skb->data_len += len;
622 refcount_add(ts, &xs->sk.sk_wmem_alloc);
627 static struct sk_buff *xsk_build_skb(struct xdp_sock *xs,
628 struct xdp_desc *desc)
630 struct net_device *dev = xs->dev;
631 struct sk_buff *skb = xs->skb;
634 if (dev->priv_flags & IFF_TX_SKB_NO_LINEAR) {
635 skb = xsk_build_skb_zerocopy(xs, desc);
644 buffer = xsk_buff_raw_get_data(xs->pool, desc->addr);
648 hr = max(NET_SKB_PAD, L1_CACHE_ALIGN(dev->needed_headroom));
649 tr = dev->needed_tailroom;
650 skb = sock_alloc_send_skb(&xs->sk, hr + len + tr, 1, &err);
654 skb_reserve(skb, hr);
657 err = skb_store_bits(skb, 0, buffer, len);
663 int nr_frags = skb_shinfo(skb)->nr_frags;
667 if (unlikely(nr_frags == (MAX_SKB_FRAGS - 1) && xp_mb_desc(desc))) {
672 page = alloc_page(xs->sk.sk_allocation);
673 if (unlikely(!page)) {
678 vaddr = kmap_local_page(page);
679 memcpy(vaddr, buffer, len);
682 skb_add_rx_frag(skb, nr_frags, page, 0, len, 0);
687 skb->priority = xs->sk.sk_priority;
688 skb->mark = READ_ONCE(xs->sk.sk_mark);
689 skb->destructor = xsk_destruct_skb;
690 xsk_set_destructor_arg(skb);
695 if (err == -EOVERFLOW) {
696 /* Drop the packet */
697 xsk_set_destructor_arg(xs->skb);
698 xsk_drop_skb(xs->skb);
699 xskq_cons_release(xs->tx);
701 /* Let application retry */
702 xsk_cq_cancel_locked(xs, 1);
708 static int __xsk_generic_xmit(struct sock *sk)
710 struct xdp_sock *xs = xdp_sk(sk);
711 u32 max_batch = TX_BATCH_SIZE;
712 bool sent_frame = false;
713 struct xdp_desc desc;
717 mutex_lock(&xs->mutex);
719 /* Since we dropped the RCU read lock, the socket state might have changed. */
720 if (unlikely(!xsk_is_bound(xs))) {
725 if (xs->queue_id >= xs->dev->real_num_tx_queues)
728 while (xskq_cons_peek_desc(xs->tx, &desc, xs->pool)) {
729 if (max_batch-- == 0) {
734 /* This is the backpressure mechanism for the Tx path.
735 * Reserve space in the completion queue and only proceed
736 * if there is space in it. This avoids having to implement
737 * any buffering in the Tx path.
739 if (xsk_cq_reserve_addr_locked(xs, desc.addr))
742 skb = xsk_build_skb(xs, &desc);
745 if (err != -EOVERFLOW)
751 xskq_cons_release(xs->tx);
753 if (xp_mb_desc(&desc)) {
758 err = __dev_direct_xmit(skb, xs->queue_id);
759 if (err == NETDEV_TX_BUSY) {
760 /* Tell user-space to retry the send */
761 xskq_cons_cancel_n(xs->tx, xsk_get_num_desc(skb));
762 xsk_consume_skb(skb);
767 /* Ignore NET_XMIT_CN as packet might have been sent */
768 if (err == NET_XMIT_DROP) {
769 /* SKB completed but not sent */
779 if (xskq_has_descs(xs->tx)) {
781 xsk_drop_skb(xs->skb);
782 xskq_cons_release(xs->tx);
787 if (xsk_tx_writeable(xs))
788 sk->sk_write_space(sk);
790 mutex_unlock(&xs->mutex);
794 static int xsk_generic_xmit(struct sock *sk)
798 /* Drop the RCU lock since the SKB path might sleep. */
800 ret = __xsk_generic_xmit(sk);
801 /* Reaquire RCU lock before going into common code. */
807 static bool xsk_no_wakeup(struct sock *sk)
809 #ifdef CONFIG_NET_RX_BUSY_POLL
810 /* Prefer busy-polling, skip the wakeup. */
811 return READ_ONCE(sk->sk_prefer_busy_poll) && READ_ONCE(sk->sk_ll_usec) &&
812 READ_ONCE(sk->sk_napi_id) >= MIN_NAPI_ID;
818 static int xsk_check_common(struct xdp_sock *xs)
820 if (unlikely(!xsk_is_bound(xs)))
822 if (unlikely(!(xs->dev->flags & IFF_UP)))
828 static int __xsk_sendmsg(struct socket *sock, struct msghdr *m, size_t total_len)
830 bool need_wait = !(m->msg_flags & MSG_DONTWAIT);
831 struct sock *sk = sock->sk;
832 struct xdp_sock *xs = xdp_sk(sk);
833 struct xsk_buff_pool *pool;
836 err = xsk_check_common(xs);
839 if (unlikely(need_wait))
841 if (unlikely(!xs->tx))
844 if (sk_can_busy_loop(sk)) {
846 __sk_mark_napi_id_once(sk, xsk_pool_get_napi_id(xs->pool));
847 sk_busy_loop(sk, 1); /* only support non-blocking sockets */
850 if (xs->zc && xsk_no_wakeup(sk))
854 if (pool->cached_need_wakeup & XDP_WAKEUP_TX) {
856 return xsk_wakeup(xs, XDP_WAKEUP_TX);
857 return xsk_generic_xmit(sk);
862 static int xsk_sendmsg(struct socket *sock, struct msghdr *m, size_t total_len)
867 ret = __xsk_sendmsg(sock, m, total_len);
873 static int __xsk_recvmsg(struct socket *sock, struct msghdr *m, size_t len, int flags)
875 bool need_wait = !(flags & MSG_DONTWAIT);
876 struct sock *sk = sock->sk;
877 struct xdp_sock *xs = xdp_sk(sk);
880 err = xsk_check_common(xs);
883 if (unlikely(!xs->rx))
885 if (unlikely(need_wait))
888 if (sk_can_busy_loop(sk))
889 sk_busy_loop(sk, 1); /* only support non-blocking sockets */
891 if (xsk_no_wakeup(sk))
894 if (xs->pool->cached_need_wakeup & XDP_WAKEUP_RX && xs->zc)
895 return xsk_wakeup(xs, XDP_WAKEUP_RX);
899 static int xsk_recvmsg(struct socket *sock, struct msghdr *m, size_t len, int flags)
904 ret = __xsk_recvmsg(sock, m, len, flags);
910 static __poll_t xsk_poll(struct file *file, struct socket *sock,
911 struct poll_table_struct *wait)
914 struct sock *sk = sock->sk;
915 struct xdp_sock *xs = xdp_sk(sk);
916 struct xsk_buff_pool *pool;
918 sock_poll_wait(file, sock, wait);
921 if (xsk_check_common(xs))
926 if (pool->cached_need_wakeup) {
928 xsk_wakeup(xs, pool->cached_need_wakeup);
930 /* Poll needs to drive Tx also in copy mode */
931 xsk_generic_xmit(sk);
935 if (xs->rx && !xskq_prod_is_empty(xs->rx))
936 mask |= EPOLLIN | EPOLLRDNORM;
937 if (xs->tx && xsk_tx_writeable(xs))
938 mask |= EPOLLOUT | EPOLLWRNORM;
944 static int xsk_init_queue(u32 entries, struct xsk_queue **queue,
949 if (entries == 0 || *queue || !is_power_of_2(entries))
952 q = xskq_create(entries, umem_queue);
956 /* Make sure queue is ready before it can be seen by others */
958 WRITE_ONCE(*queue, q);
962 static void xsk_unbind_dev(struct xdp_sock *xs)
964 struct net_device *dev = xs->dev;
966 if (xs->state != XSK_BOUND)
968 WRITE_ONCE(xs->state, XSK_UNBOUND);
970 /* Wait for driver to stop using the xdp socket. */
971 xp_del_xsk(xs->pool, xs);
976 static struct xsk_map *xsk_get_map_list_entry(struct xdp_sock *xs,
977 struct xdp_sock __rcu ***map_entry)
979 struct xsk_map *map = NULL;
980 struct xsk_map_node *node;
984 spin_lock_bh(&xs->map_list_lock);
985 node = list_first_entry_or_null(&xs->map_list, struct xsk_map_node,
988 bpf_map_inc(&node->map->map);
990 *map_entry = node->map_entry;
992 spin_unlock_bh(&xs->map_list_lock);
996 static void xsk_delete_from_maps(struct xdp_sock *xs)
998 /* This function removes the current XDP socket from all the
999 * maps it resides in. We need to take extra care here, due to
1000 * the two locks involved. Each map has a lock synchronizing
1001 * updates to the entries, and each socket has a lock that
1002 * synchronizes access to the list of maps (map_list). For
1003 * deadlock avoidance the locks need to be taken in the order
1004 * "map lock"->"socket map list lock". We start off by
1005 * accessing the socket map list, and take a reference to the
1006 * map to guarantee existence between the
1007 * xsk_get_map_list_entry() and xsk_map_try_sock_delete()
1008 * calls. Then we ask the map to remove the socket, which
1009 * tries to remove the socket from the map. Note that there
1010 * might be updates to the map between
1011 * xsk_get_map_list_entry() and xsk_map_try_sock_delete().
1013 struct xdp_sock __rcu **map_entry = NULL;
1014 struct xsk_map *map;
1016 while ((map = xsk_get_map_list_entry(xs, &map_entry))) {
1017 xsk_map_try_sock_delete(map, xs, map_entry);
1018 bpf_map_put(&map->map);
1022 static int xsk_release(struct socket *sock)
1024 struct sock *sk = sock->sk;
1025 struct xdp_sock *xs = xdp_sk(sk);
1034 xsk_drop_skb(xs->skb);
1036 mutex_lock(&net->xdp.lock);
1037 sk_del_node_init_rcu(sk);
1038 mutex_unlock(&net->xdp.lock);
1040 sock_prot_inuse_add(net, sk->sk_prot, -1);
1042 xsk_delete_from_maps(xs);
1043 mutex_lock(&xs->mutex);
1045 mutex_unlock(&xs->mutex);
1047 xskq_destroy(xs->rx);
1048 xskq_destroy(xs->tx);
1049 xskq_destroy(xs->fq_tmp);
1050 xskq_destroy(xs->cq_tmp);
1060 static struct socket *xsk_lookup_xsk_from_fd(int fd)
1062 struct socket *sock;
1065 sock = sockfd_lookup(fd, &err);
1067 return ERR_PTR(-ENOTSOCK);
1069 if (sock->sk->sk_family != PF_XDP) {
1071 return ERR_PTR(-ENOPROTOOPT);
1077 static bool xsk_validate_queues(struct xdp_sock *xs)
1079 return xs->fq_tmp && xs->cq_tmp;
1082 static int xsk_bind(struct socket *sock, struct sockaddr *addr, int addr_len)
1084 struct sockaddr_xdp *sxdp = (struct sockaddr_xdp *)addr;
1085 struct sock *sk = sock->sk;
1086 struct xdp_sock *xs = xdp_sk(sk);
1087 struct net_device *dev;
1092 if (addr_len < sizeof(struct sockaddr_xdp))
1094 if (sxdp->sxdp_family != AF_XDP)
1097 flags = sxdp->sxdp_flags;
1098 if (flags & ~(XDP_SHARED_UMEM | XDP_COPY | XDP_ZEROCOPY |
1099 XDP_USE_NEED_WAKEUP | XDP_USE_SG))
1102 bound_dev_if = READ_ONCE(sk->sk_bound_dev_if);
1103 if (bound_dev_if && bound_dev_if != sxdp->sxdp_ifindex)
1107 mutex_lock(&xs->mutex);
1108 if (xs->state != XSK_READY) {
1113 dev = dev_get_by_index(sock_net(sk), sxdp->sxdp_ifindex);
1119 if (!xs->rx && !xs->tx) {
1124 qid = sxdp->sxdp_queue_id;
1126 if (flags & XDP_SHARED_UMEM) {
1127 struct xdp_sock *umem_xs;
1128 struct socket *sock;
1130 if ((flags & XDP_COPY) || (flags & XDP_ZEROCOPY) ||
1131 (flags & XDP_USE_NEED_WAKEUP) || (flags & XDP_USE_SG)) {
1132 /* Cannot specify flags for shared sockets. */
1138 /* We have already our own. */
1143 sock = xsk_lookup_xsk_from_fd(sxdp->sxdp_shared_umem_fd);
1145 err = PTR_ERR(sock);
1149 umem_xs = xdp_sk(sock->sk);
1150 if (!xsk_is_bound(umem_xs)) {
1156 if (umem_xs->queue_id != qid || umem_xs->dev != dev) {
1157 /* Share the umem with another socket on another qid
1160 xs->pool = xp_create_and_assign_umem(xs,
1168 err = xp_assign_dev_shared(xs->pool, umem_xs, dev,
1171 xp_destroy(xs->pool);
1177 /* Share the buffer pool with the other socket. */
1178 if (xs->fq_tmp || xs->cq_tmp) {
1179 /* Do not allow setting your own fq or cq. */
1185 xp_get_pool(umem_xs->pool);
1186 xs->pool = umem_xs->pool;
1188 /* If underlying shared umem was created without Tx
1189 * ring, allocate Tx descs array that Tx batching API
1192 if (xs->tx && !xs->pool->tx_descs) {
1193 err = xp_alloc_tx_descs(xs->pool, xs);
1195 xp_put_pool(xs->pool);
1203 xdp_get_umem(umem_xs->umem);
1204 WRITE_ONCE(xs->umem, umem_xs->umem);
1206 } else if (!xs->umem || !xsk_validate_queues(xs)) {
1210 /* This xsk has its own umem. */
1211 xs->pool = xp_create_and_assign_umem(xs, xs->umem);
1217 err = xp_assign_dev(xs->pool, dev, qid, flags);
1219 xp_destroy(xs->pool);
1225 /* FQ and CQ are now owned by the buffer pool and cleaned up with it. */
1230 xs->zc = xs->umem->zc;
1231 xs->sg = !!(flags & XDP_USE_SG);
1233 xp_add_xsk(xs->pool, xs);
1239 /* Matches smp_rmb() in bind() for shared umem
1240 * sockets, and xsk_is_bound().
1243 WRITE_ONCE(xs->state, XSK_BOUND);
1246 mutex_unlock(&xs->mutex);
1251 struct xdp_umem_reg_v1 {
1252 __u64 addr; /* Start of packet data area */
1253 __u64 len; /* Length of packet data area */
1258 static int xsk_setsockopt(struct socket *sock, int level, int optname,
1259 sockptr_t optval, unsigned int optlen)
1261 struct sock *sk = sock->sk;
1262 struct xdp_sock *xs = xdp_sk(sk);
1265 if (level != SOL_XDP)
1266 return -ENOPROTOOPT;
1272 struct xsk_queue **q;
1275 if (optlen < sizeof(entries))
1277 if (copy_from_sockptr(&entries, optval, sizeof(entries)))
1280 mutex_lock(&xs->mutex);
1281 if (xs->state != XSK_READY) {
1282 mutex_unlock(&xs->mutex);
1285 q = (optname == XDP_TX_RING) ? &xs->tx : &xs->rx;
1286 err = xsk_init_queue(entries, q, false);
1287 if (!err && optname == XDP_TX_RING)
1288 /* Tx needs to be explicitly woken up the first time */
1289 xs->tx->ring->flags |= XDP_RING_NEED_WAKEUP;
1290 mutex_unlock(&xs->mutex);
1295 size_t mr_size = sizeof(struct xdp_umem_reg);
1296 struct xdp_umem_reg mr = {};
1297 struct xdp_umem *umem;
1299 if (optlen < sizeof(struct xdp_umem_reg_v1))
1301 else if (optlen < sizeof(mr))
1302 mr_size = sizeof(struct xdp_umem_reg_v1);
1304 if (copy_from_sockptr(&mr, optval, mr_size))
1307 mutex_lock(&xs->mutex);
1308 if (xs->state != XSK_READY || xs->umem) {
1309 mutex_unlock(&xs->mutex);
1313 umem = xdp_umem_create(&mr);
1315 mutex_unlock(&xs->mutex);
1316 return PTR_ERR(umem);
1319 /* Make sure umem is ready before it can be seen by others */
1321 WRITE_ONCE(xs->umem, umem);
1322 mutex_unlock(&xs->mutex);
1325 case XDP_UMEM_FILL_RING:
1326 case XDP_UMEM_COMPLETION_RING:
1328 struct xsk_queue **q;
1331 if (copy_from_sockptr(&entries, optval, sizeof(entries)))
1334 mutex_lock(&xs->mutex);
1335 if (xs->state != XSK_READY) {
1336 mutex_unlock(&xs->mutex);
1340 q = (optname == XDP_UMEM_FILL_RING) ? &xs->fq_tmp :
1342 err = xsk_init_queue(entries, q, true);
1343 mutex_unlock(&xs->mutex);
1350 return -ENOPROTOOPT;
1353 static void xsk_enter_rxtx_offsets(struct xdp_ring_offset_v1 *ring)
1355 ring->producer = offsetof(struct xdp_rxtx_ring, ptrs.producer);
1356 ring->consumer = offsetof(struct xdp_rxtx_ring, ptrs.consumer);
1357 ring->desc = offsetof(struct xdp_rxtx_ring, desc);
1360 static void xsk_enter_umem_offsets(struct xdp_ring_offset_v1 *ring)
1362 ring->producer = offsetof(struct xdp_umem_ring, ptrs.producer);
1363 ring->consumer = offsetof(struct xdp_umem_ring, ptrs.consumer);
1364 ring->desc = offsetof(struct xdp_umem_ring, desc);
1367 struct xdp_statistics_v1 {
1369 __u64 rx_invalid_descs;
1370 __u64 tx_invalid_descs;
1373 static int xsk_getsockopt(struct socket *sock, int level, int optname,
1374 char __user *optval, int __user *optlen)
1376 struct sock *sk = sock->sk;
1377 struct xdp_sock *xs = xdp_sk(sk);
1380 if (level != SOL_XDP)
1381 return -ENOPROTOOPT;
1383 if (get_user(len, optlen))
1389 case XDP_STATISTICS:
1391 struct xdp_statistics stats = {};
1392 bool extra_stats = true;
1395 if (len < sizeof(struct xdp_statistics_v1)) {
1397 } else if (len < sizeof(stats)) {
1398 extra_stats = false;
1399 stats_size = sizeof(struct xdp_statistics_v1);
1401 stats_size = sizeof(stats);
1404 mutex_lock(&xs->mutex);
1405 stats.rx_dropped = xs->rx_dropped;
1407 stats.rx_ring_full = xs->rx_queue_full;
1408 stats.rx_fill_ring_empty_descs =
1409 xs->pool ? xskq_nb_queue_empty_descs(xs->pool->fq) : 0;
1410 stats.tx_ring_empty_descs = xskq_nb_queue_empty_descs(xs->tx);
1412 stats.rx_dropped += xs->rx_queue_full;
1414 stats.rx_invalid_descs = xskq_nb_invalid_descs(xs->rx);
1415 stats.tx_invalid_descs = xskq_nb_invalid_descs(xs->tx);
1416 mutex_unlock(&xs->mutex);
1418 if (copy_to_user(optval, &stats, stats_size))
1420 if (put_user(stats_size, optlen))
1425 case XDP_MMAP_OFFSETS:
1427 struct xdp_mmap_offsets off;
1428 struct xdp_mmap_offsets_v1 off_v1;
1429 bool flags_supported = true;
1432 if (len < sizeof(off_v1))
1434 else if (len < sizeof(off))
1435 flags_supported = false;
1437 if (flags_supported) {
1438 /* xdp_ring_offset is identical to xdp_ring_offset_v1
1439 * except for the flags field added to the end.
1441 xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *)
1443 xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *)
1445 xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *)
1447 xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *)
1449 off.rx.flags = offsetof(struct xdp_rxtx_ring,
1451 off.tx.flags = offsetof(struct xdp_rxtx_ring,
1453 off.fr.flags = offsetof(struct xdp_umem_ring,
1455 off.cr.flags = offsetof(struct xdp_umem_ring,
1461 xsk_enter_rxtx_offsets(&off_v1.rx);
1462 xsk_enter_rxtx_offsets(&off_v1.tx);
1463 xsk_enter_umem_offsets(&off_v1.fr);
1464 xsk_enter_umem_offsets(&off_v1.cr);
1466 len = sizeof(off_v1);
1470 if (copy_to_user(optval, to_copy, len))
1472 if (put_user(len, optlen))
1479 struct xdp_options opts = {};
1481 if (len < sizeof(opts))
1484 mutex_lock(&xs->mutex);
1486 opts.flags |= XDP_OPTIONS_ZEROCOPY;
1487 mutex_unlock(&xs->mutex);
1490 if (copy_to_user(optval, &opts, len))
1492 if (put_user(len, optlen))
1504 static int xsk_mmap(struct file *file, struct socket *sock,
1505 struct vm_area_struct *vma)
1507 loff_t offset = (loff_t)vma->vm_pgoff << PAGE_SHIFT;
1508 unsigned long size = vma->vm_end - vma->vm_start;
1509 struct xdp_sock *xs = xdp_sk(sock->sk);
1510 int state = READ_ONCE(xs->state);
1511 struct xsk_queue *q = NULL;
1513 if (state != XSK_READY && state != XSK_BOUND)
1516 if (offset == XDP_PGOFF_RX_RING) {
1517 q = READ_ONCE(xs->rx);
1518 } else if (offset == XDP_PGOFF_TX_RING) {
1519 q = READ_ONCE(xs->tx);
1521 /* Matches the smp_wmb() in XDP_UMEM_REG */
1523 if (offset == XDP_UMEM_PGOFF_FILL_RING)
1524 q = state == XSK_READY ? READ_ONCE(xs->fq_tmp) :
1525 READ_ONCE(xs->pool->fq);
1526 else if (offset == XDP_UMEM_PGOFF_COMPLETION_RING)
1527 q = state == XSK_READY ? READ_ONCE(xs->cq_tmp) :
1528 READ_ONCE(xs->pool->cq);
1534 /* Matches the smp_wmb() in xsk_init_queue */
1536 if (size > q->ring_vmalloc_size)
1539 return remap_vmalloc_range(vma, q->ring, 0);
1542 static int xsk_notifier(struct notifier_block *this,
1543 unsigned long msg, void *ptr)
1545 struct net_device *dev = netdev_notifier_info_to_dev(ptr);
1546 struct net *net = dev_net(dev);
1550 case NETDEV_UNREGISTER:
1551 mutex_lock(&net->xdp.lock);
1552 sk_for_each(sk, &net->xdp.list) {
1553 struct xdp_sock *xs = xdp_sk(sk);
1555 mutex_lock(&xs->mutex);
1556 if (xs->dev == dev) {
1557 sk->sk_err = ENETDOWN;
1558 if (!sock_flag(sk, SOCK_DEAD))
1559 sk_error_report(sk);
1563 /* Clear device references. */
1564 xp_clear_dev(xs->pool);
1566 mutex_unlock(&xs->mutex);
1568 mutex_unlock(&net->xdp.lock);
1574 static struct proto xsk_proto = {
1576 .owner = THIS_MODULE,
1577 .obj_size = sizeof(struct xdp_sock),
1580 static const struct proto_ops xsk_proto_ops = {
1582 .owner = THIS_MODULE,
1583 .release = xsk_release,
1585 .connect = sock_no_connect,
1586 .socketpair = sock_no_socketpair,
1587 .accept = sock_no_accept,
1588 .getname = sock_no_getname,
1590 .ioctl = sock_no_ioctl,
1591 .listen = sock_no_listen,
1592 .shutdown = sock_no_shutdown,
1593 .setsockopt = xsk_setsockopt,
1594 .getsockopt = xsk_getsockopt,
1595 .sendmsg = xsk_sendmsg,
1596 .recvmsg = xsk_recvmsg,
1600 static void xsk_destruct(struct sock *sk)
1602 struct xdp_sock *xs = xdp_sk(sk);
1604 if (!sock_flag(sk, SOCK_DEAD))
1607 if (!xp_put_pool(xs->pool))
1608 xdp_put_umem(xs->umem, !xs->pool);
1611 static int xsk_create(struct net *net, struct socket *sock, int protocol,
1614 struct xdp_sock *xs;
1617 if (!ns_capable(net->user_ns, CAP_NET_RAW))
1619 if (sock->type != SOCK_RAW)
1620 return -ESOCKTNOSUPPORT;
1623 return -EPROTONOSUPPORT;
1625 sock->state = SS_UNCONNECTED;
1627 sk = sk_alloc(net, PF_XDP, GFP_KERNEL, &xsk_proto, kern);
1631 sock->ops = &xsk_proto_ops;
1633 sock_init_data(sock, sk);
1635 sk->sk_family = PF_XDP;
1637 sk->sk_destruct = xsk_destruct;
1639 sock_set_flag(sk, SOCK_RCU_FREE);
1642 xs->state = XSK_READY;
1643 mutex_init(&xs->mutex);
1644 spin_lock_init(&xs->rx_lock);
1646 INIT_LIST_HEAD(&xs->map_list);
1647 spin_lock_init(&xs->map_list_lock);
1649 mutex_lock(&net->xdp.lock);
1650 sk_add_node_rcu(sk, &net->xdp.list);
1651 mutex_unlock(&net->xdp.lock);
1653 sock_prot_inuse_add(net, &xsk_proto, 1);
1658 static const struct net_proto_family xsk_family_ops = {
1660 .create = xsk_create,
1661 .owner = THIS_MODULE,
1664 static struct notifier_block xsk_netdev_notifier = {
1665 .notifier_call = xsk_notifier,
1668 static int __net_init xsk_net_init(struct net *net)
1670 mutex_init(&net->xdp.lock);
1671 INIT_HLIST_HEAD(&net->xdp.list);
1675 static void __net_exit xsk_net_exit(struct net *net)
1677 WARN_ON_ONCE(!hlist_empty(&net->xdp.list));
1680 static struct pernet_operations xsk_net_ops = {
1681 .init = xsk_net_init,
1682 .exit = xsk_net_exit,
1685 static int __init xsk_init(void)
1689 err = proto_register(&xsk_proto, 0 /* no slab */);
1693 err = sock_register(&xsk_family_ops);
1697 err = register_pernet_subsys(&xsk_net_ops);
1701 err = register_netdevice_notifier(&xsk_netdev_notifier);
1705 for_each_possible_cpu(cpu)
1706 INIT_LIST_HEAD(&per_cpu(xskmap_flush_list, cpu));
1710 unregister_pernet_subsys(&xsk_net_ops);
1712 sock_unregister(PF_XDP);
1714 proto_unregister(&xsk_proto);
1719 fs_initcall(xsk_init);