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 32
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
->num_rx_queues
)
112 dev
->_rx
[queue_id
].pool
= NULL
;
113 if (queue_id
< dev
->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
)
189 struct xdp_buff
*xsk_xdp
;
193 len
= xdp
->data_end
- xdp
->data
;
194 if (len
> xsk_pool_get_rx_frame_size(xs
->pool
)) {
199 xsk_xdp
= xsk_buff_alloc(xs
->pool
);
205 xsk_copy_xdp(xsk_xdp
, xdp
, len
);
206 err
= __xsk_rcv_zc(xs
, xsk_xdp
, len
);
208 xsk_buff_free(xsk_xdp
);
214 static bool xsk_tx_writeable(struct xdp_sock
*xs
)
216 if (xskq_cons_present_entries(xs
->tx
) > xs
->tx
->nentries
/ 2)
222 static bool xsk_is_bound(struct xdp_sock
*xs
)
224 if (READ_ONCE(xs
->state
) == XSK_BOUND
) {
225 /* Matches smp_wmb() in bind(). */
232 static int xsk_rcv_check(struct xdp_sock
*xs
, struct xdp_buff
*xdp
)
234 if (!xsk_is_bound(xs
))
237 if (xs
->dev
!= xdp
->rxq
->dev
|| xs
->queue_id
!= xdp
->rxq
->queue_index
)
240 sk_mark_napi_id_once_xdp(&xs
->sk
, xdp
);
244 static void xsk_flush(struct xdp_sock
*xs
)
246 xskq_prod_submit(xs
->rx
);
247 __xskq_cons_release(xs
->pool
->fq
);
248 sock_def_readable(&xs
->sk
);
251 int xsk_generic_rcv(struct xdp_sock
*xs
, struct xdp_buff
*xdp
)
255 spin_lock_bh(&xs
->rx_lock
);
256 err
= xsk_rcv_check(xs
, xdp
);
258 err
= __xsk_rcv(xs
, xdp
);
261 spin_unlock_bh(&xs
->rx_lock
);
265 static int xsk_rcv(struct xdp_sock
*xs
, struct xdp_buff
*xdp
)
270 err
= xsk_rcv_check(xs
, xdp
);
274 if (xdp
->rxq
->mem
.type
== MEM_TYPE_XSK_BUFF_POOL
) {
275 len
= xdp
->data_end
- xdp
->data
;
276 return __xsk_rcv_zc(xs
, xdp
, len
);
279 err
= __xsk_rcv(xs
, xdp
);
281 xdp_return_buff(xdp
);
285 int __xsk_map_redirect(struct xdp_sock
*xs
, struct xdp_buff
*xdp
)
287 struct list_head
*flush_list
= this_cpu_ptr(&xskmap_flush_list
);
290 err
= xsk_rcv(xs
, xdp
);
294 if (!xs
->flush_node
.prev
)
295 list_add(&xs
->flush_node
, flush_list
);
300 void __xsk_map_flush(void)
302 struct list_head
*flush_list
= this_cpu_ptr(&xskmap_flush_list
);
303 struct xdp_sock
*xs
, *tmp
;
305 list_for_each_entry_safe(xs
, tmp
, flush_list
, flush_node
) {
307 __list_del_clearprev(&xs
->flush_node
);
311 void xsk_tx_completed(struct xsk_buff_pool
*pool
, u32 nb_entries
)
313 xskq_prod_submit_n(pool
->cq
, nb_entries
);
315 EXPORT_SYMBOL(xsk_tx_completed
);
317 void xsk_tx_release(struct xsk_buff_pool
*pool
)
322 list_for_each_entry_rcu(xs
, &pool
->xsk_tx_list
, tx_list
) {
323 __xskq_cons_release(xs
->tx
);
324 if (xsk_tx_writeable(xs
))
325 xs
->sk
.sk_write_space(&xs
->sk
);
329 EXPORT_SYMBOL(xsk_tx_release
);
331 bool xsk_tx_peek_desc(struct xsk_buff_pool
*pool
, struct xdp_desc
*desc
)
336 list_for_each_entry_rcu(xs
, &pool
->xsk_tx_list
, tx_list
) {
337 if (!xskq_cons_peek_desc(xs
->tx
, desc
, pool
)) {
338 xs
->tx
->queue_empty_descs
++;
342 /* This is the backpressure mechanism for the Tx path.
343 * Reserve space in the completion queue and only proceed
344 * if there is space in it. This avoids having to implement
345 * any buffering in the Tx path.
347 if (xskq_prod_reserve_addr(pool
->cq
, desc
->addr
))
350 xskq_cons_release(xs
->tx
);
359 EXPORT_SYMBOL(xsk_tx_peek_desc
);
361 static u32
xsk_tx_peek_release_fallback(struct xsk_buff_pool
*pool
, struct xdp_desc
*descs
,
366 while (nb_pkts
< max_entries
&& xsk_tx_peek_desc(pool
, &descs
[nb_pkts
]))
369 xsk_tx_release(pool
);
373 u32
xsk_tx_peek_release_desc_batch(struct xsk_buff_pool
*pool
, struct xdp_desc
*descs
,
380 if (!list_is_singular(&pool
->xsk_tx_list
)) {
381 /* Fallback to the non-batched version */
383 return xsk_tx_peek_release_fallback(pool
, descs
, max_entries
);
386 xs
= list_first_or_null_rcu(&pool
->xsk_tx_list
, struct xdp_sock
, tx_list
);
392 nb_pkts
= xskq_cons_peek_desc_batch(xs
->tx
, descs
, pool
, max_entries
);
394 xs
->tx
->queue_empty_descs
++;
398 /* This is the backpressure mechanism for the Tx path. Try to
399 * reserve space in the completion queue for all packets, but
400 * if there are fewer slots available, just process that many
401 * packets. This avoids having to implement any buffering in
404 nb_pkts
= xskq_prod_reserve_addr_batch(pool
->cq
, descs
, nb_pkts
);
408 xskq_cons_release_n(xs
->tx
, nb_pkts
);
409 __xskq_cons_release(xs
->tx
);
410 xs
->sk
.sk_write_space(&xs
->sk
);
416 EXPORT_SYMBOL(xsk_tx_peek_release_desc_batch
);
418 static int xsk_wakeup(struct xdp_sock
*xs
, u8 flags
)
420 struct net_device
*dev
= xs
->dev
;
424 err
= dev
->netdev_ops
->ndo_xsk_wakeup(dev
, xs
->queue_id
, flags
);
430 static int xsk_zc_xmit(struct xdp_sock
*xs
)
432 return xsk_wakeup(xs
, XDP_WAKEUP_TX
);
435 static void xsk_destruct_skb(struct sk_buff
*skb
)
437 u64 addr
= (u64
)(long)skb_shinfo(skb
)->destructor_arg
;
438 struct xdp_sock
*xs
= xdp_sk(skb
->sk
);
441 spin_lock_irqsave(&xs
->pool
->cq_lock
, flags
);
442 xskq_prod_submit_addr(xs
->pool
->cq
, addr
);
443 spin_unlock_irqrestore(&xs
->pool
->cq_lock
, flags
);
448 static struct sk_buff
*xsk_build_skb_zerocopy(struct xdp_sock
*xs
,
449 struct xdp_desc
*desc
)
451 struct xsk_buff_pool
*pool
= xs
->pool
;
452 u32 hr
, len
, ts
, offset
, copy
, copied
;
459 hr
= max(NET_SKB_PAD
, L1_CACHE_ALIGN(xs
->dev
->needed_headroom
));
461 skb
= sock_alloc_send_skb(&xs
->sk
, hr
, 1, &err
);
465 skb_reserve(skb
, hr
);
469 ts
= pool
->unaligned
? len
: pool
->chunk_size
;
471 buffer
= xsk_buff_raw_get_data(pool
, addr
);
472 offset
= offset_in_page(buffer
);
473 addr
= buffer
- pool
->addrs
;
475 for (copied
= 0, i
= 0; copied
< len
; i
++) {
476 page
= pool
->umem
->pgs
[addr
>> PAGE_SHIFT
];
479 copy
= min_t(u32
, PAGE_SIZE
- offset
, len
- copied
);
480 skb_fill_page_desc(skb
, i
, page
, offset
, copy
);
488 skb
->data_len
+= len
;
491 refcount_add(ts
, &xs
->sk
.sk_wmem_alloc
);
496 static struct sk_buff
*xsk_build_skb(struct xdp_sock
*xs
,
497 struct xdp_desc
*desc
)
499 struct net_device
*dev
= xs
->dev
;
502 if (dev
->priv_flags
& IFF_TX_SKB_NO_LINEAR
) {
503 skb
= xsk_build_skb_zerocopy(xs
, desc
);
511 hr
= max(NET_SKB_PAD
, L1_CACHE_ALIGN(dev
->needed_headroom
));
512 tr
= dev
->needed_tailroom
;
515 skb
= sock_alloc_send_skb(&xs
->sk
, hr
+ len
+ tr
, 1, &err
);
519 skb_reserve(skb
, hr
);
522 buffer
= xsk_buff_raw_get_data(xs
->pool
, desc
->addr
);
523 err
= skb_store_bits(skb
, 0, buffer
, len
);
531 skb
->priority
= xs
->sk
.sk_priority
;
532 skb
->mark
= xs
->sk
.sk_mark
;
533 skb_shinfo(skb
)->destructor_arg
= (void *)(long)desc
->addr
;
534 skb
->destructor
= xsk_destruct_skb
;
539 static int xsk_generic_xmit(struct sock
*sk
)
541 struct xdp_sock
*xs
= xdp_sk(sk
);
542 u32 max_batch
= TX_BATCH_SIZE
;
543 bool sent_frame
= false;
544 struct xdp_desc desc
;
549 mutex_lock(&xs
->mutex
);
551 if (xs
->queue_id
>= xs
->dev
->real_num_tx_queues
)
554 while (xskq_cons_peek_desc(xs
->tx
, &desc
, xs
->pool
)) {
555 if (max_batch
-- == 0) {
560 skb
= xsk_build_skb(xs
, &desc
);
566 /* This is the backpressure mechanism for the Tx path.
567 * Reserve space in the completion queue and only proceed
568 * if there is space in it. This avoids having to implement
569 * any buffering in the Tx path.
571 spin_lock_irqsave(&xs
->pool
->cq_lock
, flags
);
572 if (xskq_prod_reserve(xs
->pool
->cq
)) {
573 spin_unlock_irqrestore(&xs
->pool
->cq_lock
, flags
);
577 spin_unlock_irqrestore(&xs
->pool
->cq_lock
, flags
);
579 err
= __dev_direct_xmit(skb
, xs
->queue_id
);
580 if (err
== NETDEV_TX_BUSY
) {
581 /* Tell user-space to retry the send */
582 skb
->destructor
= sock_wfree
;
583 spin_lock_irqsave(&xs
->pool
->cq_lock
, flags
);
584 xskq_prod_cancel(xs
->pool
->cq
);
585 spin_unlock_irqrestore(&xs
->pool
->cq_lock
, flags
);
586 /* Free skb without triggering the perf drop trace */
592 xskq_cons_release(xs
->tx
);
593 /* Ignore NET_XMIT_CN as packet might have been sent */
594 if (err
== NET_XMIT_DROP
) {
595 /* SKB completed but not sent */
603 xs
->tx
->queue_empty_descs
++;
607 if (xsk_tx_writeable(xs
))
608 sk
->sk_write_space(sk
);
610 mutex_unlock(&xs
->mutex
);
614 static int __xsk_sendmsg(struct sock
*sk
)
616 struct xdp_sock
*xs
= xdp_sk(sk
);
618 if (unlikely(!(xs
->dev
->flags
& IFF_UP
)))
620 if (unlikely(!xs
->tx
))
623 return xs
->zc
? xsk_zc_xmit(xs
) : xsk_generic_xmit(sk
);
626 static bool xsk_no_wakeup(struct sock
*sk
)
628 #ifdef CONFIG_NET_RX_BUSY_POLL
629 /* Prefer busy-polling, skip the wakeup. */
630 return READ_ONCE(sk
->sk_prefer_busy_poll
) && READ_ONCE(sk
->sk_ll_usec
) &&
631 READ_ONCE(sk
->sk_napi_id
) >= MIN_NAPI_ID
;
637 static int xsk_sendmsg(struct socket
*sock
, struct msghdr
*m
, size_t total_len
)
639 bool need_wait
= !(m
->msg_flags
& MSG_DONTWAIT
);
640 struct sock
*sk
= sock
->sk
;
641 struct xdp_sock
*xs
= xdp_sk(sk
);
642 struct xsk_buff_pool
*pool
;
644 if (unlikely(!xsk_is_bound(xs
)))
646 if (unlikely(need_wait
))
649 if (sk_can_busy_loop(sk
))
650 sk_busy_loop(sk
, 1); /* only support non-blocking sockets */
652 if (xsk_no_wakeup(sk
))
656 if (pool
->cached_need_wakeup
& XDP_WAKEUP_TX
)
657 return __xsk_sendmsg(sk
);
661 static int xsk_recvmsg(struct socket
*sock
, struct msghdr
*m
, size_t len
, int flags
)
663 bool need_wait
= !(flags
& MSG_DONTWAIT
);
664 struct sock
*sk
= sock
->sk
;
665 struct xdp_sock
*xs
= xdp_sk(sk
);
667 if (unlikely(!xsk_is_bound(xs
)))
669 if (unlikely(!(xs
->dev
->flags
& IFF_UP
)))
671 if (unlikely(!xs
->rx
))
673 if (unlikely(need_wait
))
676 if (sk_can_busy_loop(sk
))
677 sk_busy_loop(sk
, 1); /* only support non-blocking sockets */
679 if (xsk_no_wakeup(sk
))
682 if (xs
->pool
->cached_need_wakeup
& XDP_WAKEUP_RX
&& xs
->zc
)
683 return xsk_wakeup(xs
, XDP_WAKEUP_RX
);
687 static __poll_t
xsk_poll(struct file
*file
, struct socket
*sock
,
688 struct poll_table_struct
*wait
)
691 struct sock
*sk
= sock
->sk
;
692 struct xdp_sock
*xs
= xdp_sk(sk
);
693 struct xsk_buff_pool
*pool
;
695 sock_poll_wait(file
, sock
, wait
);
697 if (unlikely(!xsk_is_bound(xs
)))
702 if (pool
->cached_need_wakeup
) {
704 xsk_wakeup(xs
, pool
->cached_need_wakeup
);
706 /* Poll needs to drive Tx also in copy mode */
710 if (xs
->rx
&& !xskq_prod_is_empty(xs
->rx
))
711 mask
|= EPOLLIN
| EPOLLRDNORM
;
712 if (xs
->tx
&& xsk_tx_writeable(xs
))
713 mask
|= EPOLLOUT
| EPOLLWRNORM
;
718 static int xsk_init_queue(u32 entries
, struct xsk_queue
**queue
,
723 if (entries
== 0 || *queue
|| !is_power_of_2(entries
))
726 q
= xskq_create(entries
, umem_queue
);
730 /* Make sure queue is ready before it can be seen by others */
732 WRITE_ONCE(*queue
, q
);
736 static void xsk_unbind_dev(struct xdp_sock
*xs
)
738 struct net_device
*dev
= xs
->dev
;
740 if (xs
->state
!= XSK_BOUND
)
742 WRITE_ONCE(xs
->state
, XSK_UNBOUND
);
744 /* Wait for driver to stop using the xdp socket. */
745 xp_del_xsk(xs
->pool
, xs
);
751 static struct xsk_map
*xsk_get_map_list_entry(struct xdp_sock
*xs
,
752 struct xdp_sock __rcu
***map_entry
)
754 struct xsk_map
*map
= NULL
;
755 struct xsk_map_node
*node
;
759 spin_lock_bh(&xs
->map_list_lock
);
760 node
= list_first_entry_or_null(&xs
->map_list
, struct xsk_map_node
,
763 bpf_map_inc(&node
->map
->map
);
765 *map_entry
= node
->map_entry
;
767 spin_unlock_bh(&xs
->map_list_lock
);
771 static void xsk_delete_from_maps(struct xdp_sock
*xs
)
773 /* This function removes the current XDP socket from all the
774 * maps it resides in. We need to take extra care here, due to
775 * the two locks involved. Each map has a lock synchronizing
776 * updates to the entries, and each socket has a lock that
777 * synchronizes access to the list of maps (map_list). For
778 * deadlock avoidance the locks need to be taken in the order
779 * "map lock"->"socket map list lock". We start off by
780 * accessing the socket map list, and take a reference to the
781 * map to guarantee existence between the
782 * xsk_get_map_list_entry() and xsk_map_try_sock_delete()
783 * calls. Then we ask the map to remove the socket, which
784 * tries to remove the socket from the map. Note that there
785 * might be updates to the map between
786 * xsk_get_map_list_entry() and xsk_map_try_sock_delete().
788 struct xdp_sock __rcu
**map_entry
= NULL
;
791 while ((map
= xsk_get_map_list_entry(xs
, &map_entry
))) {
792 xsk_map_try_sock_delete(map
, xs
, map_entry
);
793 bpf_map_put(&map
->map
);
797 static int xsk_release(struct socket
*sock
)
799 struct sock
*sk
= sock
->sk
;
800 struct xdp_sock
*xs
= xdp_sk(sk
);
808 mutex_lock(&net
->xdp
.lock
);
809 sk_del_node_init_rcu(sk
);
810 mutex_unlock(&net
->xdp
.lock
);
813 sock_prot_inuse_add(net
, sk
->sk_prot
, -1);
816 xsk_delete_from_maps(xs
);
817 mutex_lock(&xs
->mutex
);
819 mutex_unlock(&xs
->mutex
);
821 xskq_destroy(xs
->rx
);
822 xskq_destroy(xs
->tx
);
823 xskq_destroy(xs
->fq_tmp
);
824 xskq_destroy(xs
->cq_tmp
);
829 sk_refcnt_debug_release(sk
);
835 static struct socket
*xsk_lookup_xsk_from_fd(int fd
)
840 sock
= sockfd_lookup(fd
, &err
);
842 return ERR_PTR(-ENOTSOCK
);
844 if (sock
->sk
->sk_family
!= PF_XDP
) {
846 return ERR_PTR(-ENOPROTOOPT
);
852 static bool xsk_validate_queues(struct xdp_sock
*xs
)
854 return xs
->fq_tmp
&& xs
->cq_tmp
;
857 static int xsk_bind(struct socket
*sock
, struct sockaddr
*addr
, int addr_len
)
859 struct sockaddr_xdp
*sxdp
= (struct sockaddr_xdp
*)addr
;
860 struct sock
*sk
= sock
->sk
;
861 struct xdp_sock
*xs
= xdp_sk(sk
);
862 struct net_device
*dev
;
866 if (addr_len
< sizeof(struct sockaddr_xdp
))
868 if (sxdp
->sxdp_family
!= AF_XDP
)
871 flags
= sxdp
->sxdp_flags
;
872 if (flags
& ~(XDP_SHARED_UMEM
| XDP_COPY
| XDP_ZEROCOPY
|
873 XDP_USE_NEED_WAKEUP
))
877 mutex_lock(&xs
->mutex
);
878 if (xs
->state
!= XSK_READY
) {
883 dev
= dev_get_by_index(sock_net(sk
), sxdp
->sxdp_ifindex
);
889 if (!xs
->rx
&& !xs
->tx
) {
894 qid
= sxdp
->sxdp_queue_id
;
896 if (flags
& XDP_SHARED_UMEM
) {
897 struct xdp_sock
*umem_xs
;
900 if ((flags
& XDP_COPY
) || (flags
& XDP_ZEROCOPY
) ||
901 (flags
& XDP_USE_NEED_WAKEUP
)) {
902 /* Cannot specify flags for shared sockets. */
908 /* We have already our own. */
913 sock
= xsk_lookup_xsk_from_fd(sxdp
->sxdp_shared_umem_fd
);
919 umem_xs
= xdp_sk(sock
->sk
);
920 if (!xsk_is_bound(umem_xs
)) {
926 if (umem_xs
->queue_id
!= qid
|| umem_xs
->dev
!= dev
) {
927 /* Share the umem with another socket on another qid
930 xs
->pool
= xp_create_and_assign_umem(xs
,
938 err
= xp_assign_dev_shared(xs
->pool
, umem_xs
->umem
,
941 xp_destroy(xs
->pool
);
947 /* Share the buffer pool with the other socket. */
948 if (xs
->fq_tmp
|| xs
->cq_tmp
) {
949 /* Do not allow setting your own fq or cq. */
955 xp_get_pool(umem_xs
->pool
);
956 xs
->pool
= umem_xs
->pool
;
959 xdp_get_umem(umem_xs
->umem
);
960 WRITE_ONCE(xs
->umem
, umem_xs
->umem
);
962 } else if (!xs
->umem
|| !xsk_validate_queues(xs
)) {
966 /* This xsk has its own umem. */
967 xs
->pool
= xp_create_and_assign_umem(xs
, xs
->umem
);
973 err
= xp_assign_dev(xs
->pool
, dev
, qid
, flags
);
975 xp_destroy(xs
->pool
);
981 /* FQ and CQ are now owned by the buffer pool and cleaned up with it. */
986 xs
->zc
= xs
->umem
->zc
;
988 xp_add_xsk(xs
->pool
, xs
);
994 /* Matches smp_rmb() in bind() for shared umem
995 * sockets, and xsk_is_bound().
998 WRITE_ONCE(xs
->state
, XSK_BOUND
);
1001 mutex_unlock(&xs
->mutex
);
1006 struct xdp_umem_reg_v1
{
1007 __u64 addr
; /* Start of packet data area */
1008 __u64 len
; /* Length of packet data area */
1013 static int xsk_setsockopt(struct socket
*sock
, int level
, int optname
,
1014 sockptr_t optval
, unsigned int optlen
)
1016 struct sock
*sk
= sock
->sk
;
1017 struct xdp_sock
*xs
= xdp_sk(sk
);
1020 if (level
!= SOL_XDP
)
1021 return -ENOPROTOOPT
;
1027 struct xsk_queue
**q
;
1030 if (optlen
< sizeof(entries
))
1032 if (copy_from_sockptr(&entries
, optval
, sizeof(entries
)))
1035 mutex_lock(&xs
->mutex
);
1036 if (xs
->state
!= XSK_READY
) {
1037 mutex_unlock(&xs
->mutex
);
1040 q
= (optname
== XDP_TX_RING
) ? &xs
->tx
: &xs
->rx
;
1041 err
= xsk_init_queue(entries
, q
, false);
1042 if (!err
&& optname
== XDP_TX_RING
)
1043 /* Tx needs to be explicitly woken up the first time */
1044 xs
->tx
->ring
->flags
|= XDP_RING_NEED_WAKEUP
;
1045 mutex_unlock(&xs
->mutex
);
1050 size_t mr_size
= sizeof(struct xdp_umem_reg
);
1051 struct xdp_umem_reg mr
= {};
1052 struct xdp_umem
*umem
;
1054 if (optlen
< sizeof(struct xdp_umem_reg_v1
))
1056 else if (optlen
< sizeof(mr
))
1057 mr_size
= sizeof(struct xdp_umem_reg_v1
);
1059 if (copy_from_sockptr(&mr
, optval
, mr_size
))
1062 mutex_lock(&xs
->mutex
);
1063 if (xs
->state
!= XSK_READY
|| xs
->umem
) {
1064 mutex_unlock(&xs
->mutex
);
1068 umem
= xdp_umem_create(&mr
);
1070 mutex_unlock(&xs
->mutex
);
1071 return PTR_ERR(umem
);
1074 /* Make sure umem is ready before it can be seen by others */
1076 WRITE_ONCE(xs
->umem
, umem
);
1077 mutex_unlock(&xs
->mutex
);
1080 case XDP_UMEM_FILL_RING
:
1081 case XDP_UMEM_COMPLETION_RING
:
1083 struct xsk_queue
**q
;
1086 if (copy_from_sockptr(&entries
, optval
, sizeof(entries
)))
1089 mutex_lock(&xs
->mutex
);
1090 if (xs
->state
!= XSK_READY
) {
1091 mutex_unlock(&xs
->mutex
);
1095 q
= (optname
== XDP_UMEM_FILL_RING
) ? &xs
->fq_tmp
:
1097 err
= xsk_init_queue(entries
, q
, true);
1098 mutex_unlock(&xs
->mutex
);
1105 return -ENOPROTOOPT
;
1108 static void xsk_enter_rxtx_offsets(struct xdp_ring_offset_v1
*ring
)
1110 ring
->producer
= offsetof(struct xdp_rxtx_ring
, ptrs
.producer
);
1111 ring
->consumer
= offsetof(struct xdp_rxtx_ring
, ptrs
.consumer
);
1112 ring
->desc
= offsetof(struct xdp_rxtx_ring
, desc
);
1115 static void xsk_enter_umem_offsets(struct xdp_ring_offset_v1
*ring
)
1117 ring
->producer
= offsetof(struct xdp_umem_ring
, ptrs
.producer
);
1118 ring
->consumer
= offsetof(struct xdp_umem_ring
, ptrs
.consumer
);
1119 ring
->desc
= offsetof(struct xdp_umem_ring
, desc
);
1122 struct xdp_statistics_v1
{
1124 __u64 rx_invalid_descs
;
1125 __u64 tx_invalid_descs
;
1128 static int xsk_getsockopt(struct socket
*sock
, int level
, int optname
,
1129 char __user
*optval
, int __user
*optlen
)
1131 struct sock
*sk
= sock
->sk
;
1132 struct xdp_sock
*xs
= xdp_sk(sk
);
1135 if (level
!= SOL_XDP
)
1136 return -ENOPROTOOPT
;
1138 if (get_user(len
, optlen
))
1144 case XDP_STATISTICS
:
1146 struct xdp_statistics stats
= {};
1147 bool extra_stats
= true;
1150 if (len
< sizeof(struct xdp_statistics_v1
)) {
1152 } else if (len
< sizeof(stats
)) {
1153 extra_stats
= false;
1154 stats_size
= sizeof(struct xdp_statistics_v1
);
1156 stats_size
= sizeof(stats
);
1159 mutex_lock(&xs
->mutex
);
1160 stats
.rx_dropped
= xs
->rx_dropped
;
1162 stats
.rx_ring_full
= xs
->rx_queue_full
;
1163 stats
.rx_fill_ring_empty_descs
=
1164 xs
->pool
? xskq_nb_queue_empty_descs(xs
->pool
->fq
) : 0;
1165 stats
.tx_ring_empty_descs
= xskq_nb_queue_empty_descs(xs
->tx
);
1167 stats
.rx_dropped
+= xs
->rx_queue_full
;
1169 stats
.rx_invalid_descs
= xskq_nb_invalid_descs(xs
->rx
);
1170 stats
.tx_invalid_descs
= xskq_nb_invalid_descs(xs
->tx
);
1171 mutex_unlock(&xs
->mutex
);
1173 if (copy_to_user(optval
, &stats
, stats_size
))
1175 if (put_user(stats_size
, optlen
))
1180 case XDP_MMAP_OFFSETS
:
1182 struct xdp_mmap_offsets off
;
1183 struct xdp_mmap_offsets_v1 off_v1
;
1184 bool flags_supported
= true;
1187 if (len
< sizeof(off_v1
))
1189 else if (len
< sizeof(off
))
1190 flags_supported
= false;
1192 if (flags_supported
) {
1193 /* xdp_ring_offset is identical to xdp_ring_offset_v1
1194 * except for the flags field added to the end.
1196 xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1
*)
1198 xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1
*)
1200 xsk_enter_umem_offsets((struct xdp_ring_offset_v1
*)
1202 xsk_enter_umem_offsets((struct xdp_ring_offset_v1
*)
1204 off
.rx
.flags
= offsetof(struct xdp_rxtx_ring
,
1206 off
.tx
.flags
= offsetof(struct xdp_rxtx_ring
,
1208 off
.fr
.flags
= offsetof(struct xdp_umem_ring
,
1210 off
.cr
.flags
= offsetof(struct xdp_umem_ring
,
1216 xsk_enter_rxtx_offsets(&off_v1
.rx
);
1217 xsk_enter_rxtx_offsets(&off_v1
.tx
);
1218 xsk_enter_umem_offsets(&off_v1
.fr
);
1219 xsk_enter_umem_offsets(&off_v1
.cr
);
1221 len
= sizeof(off_v1
);
1225 if (copy_to_user(optval
, to_copy
, len
))
1227 if (put_user(len
, optlen
))
1234 struct xdp_options opts
= {};
1236 if (len
< sizeof(opts
))
1239 mutex_lock(&xs
->mutex
);
1241 opts
.flags
|= XDP_OPTIONS_ZEROCOPY
;
1242 mutex_unlock(&xs
->mutex
);
1245 if (copy_to_user(optval
, &opts
, len
))
1247 if (put_user(len
, optlen
))
1259 static int xsk_mmap(struct file
*file
, struct socket
*sock
,
1260 struct vm_area_struct
*vma
)
1262 loff_t offset
= (loff_t
)vma
->vm_pgoff
<< PAGE_SHIFT
;
1263 unsigned long size
= vma
->vm_end
- vma
->vm_start
;
1264 struct xdp_sock
*xs
= xdp_sk(sock
->sk
);
1265 struct xsk_queue
*q
= NULL
;
1269 if (READ_ONCE(xs
->state
) != XSK_READY
)
1272 if (offset
== XDP_PGOFF_RX_RING
) {
1273 q
= READ_ONCE(xs
->rx
);
1274 } else if (offset
== XDP_PGOFF_TX_RING
) {
1275 q
= READ_ONCE(xs
->tx
);
1277 /* Matches the smp_wmb() in XDP_UMEM_REG */
1279 if (offset
== XDP_UMEM_PGOFF_FILL_RING
)
1280 q
= READ_ONCE(xs
->fq_tmp
);
1281 else if (offset
== XDP_UMEM_PGOFF_COMPLETION_RING
)
1282 q
= READ_ONCE(xs
->cq_tmp
);
1288 /* Matches the smp_wmb() in xsk_init_queue */
1290 qpg
= virt_to_head_page(q
->ring
);
1291 if (size
> page_size(qpg
))
1294 pfn
= virt_to_phys(q
->ring
) >> PAGE_SHIFT
;
1295 return remap_pfn_range(vma
, vma
->vm_start
, pfn
,
1296 size
, vma
->vm_page_prot
);
1299 static int xsk_notifier(struct notifier_block
*this,
1300 unsigned long msg
, void *ptr
)
1302 struct net_device
*dev
= netdev_notifier_info_to_dev(ptr
);
1303 struct net
*net
= dev_net(dev
);
1307 case NETDEV_UNREGISTER
:
1308 mutex_lock(&net
->xdp
.lock
);
1309 sk_for_each(sk
, &net
->xdp
.list
) {
1310 struct xdp_sock
*xs
= xdp_sk(sk
);
1312 mutex_lock(&xs
->mutex
);
1313 if (xs
->dev
== dev
) {
1314 sk
->sk_err
= ENETDOWN
;
1315 if (!sock_flag(sk
, SOCK_DEAD
))
1316 sk_error_report(sk
);
1320 /* Clear device references. */
1321 xp_clear_dev(xs
->pool
);
1323 mutex_unlock(&xs
->mutex
);
1325 mutex_unlock(&net
->xdp
.lock
);
1331 static struct proto xsk_proto
= {
1333 .owner
= THIS_MODULE
,
1334 .obj_size
= sizeof(struct xdp_sock
),
1337 static const struct proto_ops xsk_proto_ops
= {
1339 .owner
= THIS_MODULE
,
1340 .release
= xsk_release
,
1342 .connect
= sock_no_connect
,
1343 .socketpair
= sock_no_socketpair
,
1344 .accept
= sock_no_accept
,
1345 .getname
= sock_no_getname
,
1347 .ioctl
= sock_no_ioctl
,
1348 .listen
= sock_no_listen
,
1349 .shutdown
= sock_no_shutdown
,
1350 .setsockopt
= xsk_setsockopt
,
1351 .getsockopt
= xsk_getsockopt
,
1352 .sendmsg
= xsk_sendmsg
,
1353 .recvmsg
= xsk_recvmsg
,
1355 .sendpage
= sock_no_sendpage
,
1358 static void xsk_destruct(struct sock
*sk
)
1360 struct xdp_sock
*xs
= xdp_sk(sk
);
1362 if (!sock_flag(sk
, SOCK_DEAD
))
1365 if (!xp_put_pool(xs
->pool
))
1366 xdp_put_umem(xs
->umem
, !xs
->pool
);
1368 sk_refcnt_debug_dec(sk
);
1371 static int xsk_create(struct net
*net
, struct socket
*sock
, int protocol
,
1374 struct xdp_sock
*xs
;
1377 if (!ns_capable(net
->user_ns
, CAP_NET_RAW
))
1379 if (sock
->type
!= SOCK_RAW
)
1380 return -ESOCKTNOSUPPORT
;
1383 return -EPROTONOSUPPORT
;
1385 sock
->state
= SS_UNCONNECTED
;
1387 sk
= sk_alloc(net
, PF_XDP
, GFP_KERNEL
, &xsk_proto
, kern
);
1391 sock
->ops
= &xsk_proto_ops
;
1393 sock_init_data(sock
, sk
);
1395 sk
->sk_family
= PF_XDP
;
1397 sk
->sk_destruct
= xsk_destruct
;
1398 sk_refcnt_debug_inc(sk
);
1400 sock_set_flag(sk
, SOCK_RCU_FREE
);
1403 xs
->state
= XSK_READY
;
1404 mutex_init(&xs
->mutex
);
1405 spin_lock_init(&xs
->rx_lock
);
1407 INIT_LIST_HEAD(&xs
->map_list
);
1408 spin_lock_init(&xs
->map_list_lock
);
1410 mutex_lock(&net
->xdp
.lock
);
1411 sk_add_node_rcu(sk
, &net
->xdp
.list
);
1412 mutex_unlock(&net
->xdp
.lock
);
1415 sock_prot_inuse_add(net
, &xsk_proto
, 1);
1421 static const struct net_proto_family xsk_family_ops
= {
1423 .create
= xsk_create
,
1424 .owner
= THIS_MODULE
,
1427 static struct notifier_block xsk_netdev_notifier
= {
1428 .notifier_call
= xsk_notifier
,
1431 static int __net_init
xsk_net_init(struct net
*net
)
1433 mutex_init(&net
->xdp
.lock
);
1434 INIT_HLIST_HEAD(&net
->xdp
.list
);
1438 static void __net_exit
xsk_net_exit(struct net
*net
)
1440 WARN_ON_ONCE(!hlist_empty(&net
->xdp
.list
));
1443 static struct pernet_operations xsk_net_ops
= {
1444 .init
= xsk_net_init
,
1445 .exit
= xsk_net_exit
,
1448 static int __init
xsk_init(void)
1452 err
= proto_register(&xsk_proto
, 0 /* no slab */);
1456 err
= sock_register(&xsk_family_ops
);
1460 err
= register_pernet_subsys(&xsk_net_ops
);
1464 err
= register_netdevice_notifier(&xsk_netdev_notifier
);
1468 for_each_possible_cpu(cpu
)
1469 INIT_LIST_HEAD(&per_cpu(xskmap_flush_list
, cpu
));
1473 unregister_pernet_subsys(&xsk_net_ops
);
1475 sock_unregister(PF_XDP
);
1477 proto_unregister(&xsk_proto
);
1482 fs_initcall(xsk_init
);