xref: /openbmc/linux/net/xdp/xsk.c (revision 165f2d28)
1 // SPDX-License-Identifier: GPL-2.0
2 /* XDP sockets
3  *
4  * AF_XDP sockets allows a channel between XDP programs and userspace
5  * applications.
6  * Copyright(c) 2018 Intel Corporation.
7  *
8  * Author(s): Björn Töpel <bjorn.topel@intel.com>
9  *	      Magnus Karlsson <magnus.karlsson@intel.com>
10  */
11 
12 #define pr_fmt(fmt) "AF_XDP: %s: " fmt, __func__
13 
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>
26 #include <net/xdp.h>
27 
28 #include "xsk_queue.h"
29 #include "xdp_umem.h"
30 #include "xsk.h"
31 
32 #define TX_BATCH_SIZE 16
33 
34 static DEFINE_PER_CPU(struct list_head, xskmap_flush_list);
35 
36 bool xsk_is_setup_for_bpf_map(struct xdp_sock *xs)
37 {
38 	return READ_ONCE(xs->rx) &&  READ_ONCE(xs->umem) &&
39 		READ_ONCE(xs->umem->fq);
40 }
41 
42 bool xsk_umem_has_addrs(struct xdp_umem *umem, u32 cnt)
43 {
44 	return xskq_cons_has_entries(umem->fq, cnt);
45 }
46 EXPORT_SYMBOL(xsk_umem_has_addrs);
47 
48 bool xsk_umem_peek_addr(struct xdp_umem *umem, u64 *addr)
49 {
50 	return xskq_cons_peek_addr(umem->fq, addr, umem);
51 }
52 EXPORT_SYMBOL(xsk_umem_peek_addr);
53 
54 void xsk_umem_release_addr(struct xdp_umem *umem)
55 {
56 	xskq_cons_release(umem->fq);
57 }
58 EXPORT_SYMBOL(xsk_umem_release_addr);
59 
60 void xsk_set_rx_need_wakeup(struct xdp_umem *umem)
61 {
62 	if (umem->need_wakeup & XDP_WAKEUP_RX)
63 		return;
64 
65 	umem->fq->ring->flags |= XDP_RING_NEED_WAKEUP;
66 	umem->need_wakeup |= XDP_WAKEUP_RX;
67 }
68 EXPORT_SYMBOL(xsk_set_rx_need_wakeup);
69 
70 void xsk_set_tx_need_wakeup(struct xdp_umem *umem)
71 {
72 	struct xdp_sock *xs;
73 
74 	if (umem->need_wakeup & XDP_WAKEUP_TX)
75 		return;
76 
77 	rcu_read_lock();
78 	list_for_each_entry_rcu(xs, &umem->xsk_list, list) {
79 		xs->tx->ring->flags |= XDP_RING_NEED_WAKEUP;
80 	}
81 	rcu_read_unlock();
82 
83 	umem->need_wakeup |= XDP_WAKEUP_TX;
84 }
85 EXPORT_SYMBOL(xsk_set_tx_need_wakeup);
86 
87 void xsk_clear_rx_need_wakeup(struct xdp_umem *umem)
88 {
89 	if (!(umem->need_wakeup & XDP_WAKEUP_RX))
90 		return;
91 
92 	umem->fq->ring->flags &= ~XDP_RING_NEED_WAKEUP;
93 	umem->need_wakeup &= ~XDP_WAKEUP_RX;
94 }
95 EXPORT_SYMBOL(xsk_clear_rx_need_wakeup);
96 
97 void xsk_clear_tx_need_wakeup(struct xdp_umem *umem)
98 {
99 	struct xdp_sock *xs;
100 
101 	if (!(umem->need_wakeup & XDP_WAKEUP_TX))
102 		return;
103 
104 	rcu_read_lock();
105 	list_for_each_entry_rcu(xs, &umem->xsk_list, list) {
106 		xs->tx->ring->flags &= ~XDP_RING_NEED_WAKEUP;
107 	}
108 	rcu_read_unlock();
109 
110 	umem->need_wakeup &= ~XDP_WAKEUP_TX;
111 }
112 EXPORT_SYMBOL(xsk_clear_tx_need_wakeup);
113 
114 bool xsk_umem_uses_need_wakeup(struct xdp_umem *umem)
115 {
116 	return umem->flags & XDP_UMEM_USES_NEED_WAKEUP;
117 }
118 EXPORT_SYMBOL(xsk_umem_uses_need_wakeup);
119 
120 /* If a buffer crosses a page boundary, we need to do 2 memcpy's, one for
121  * each page. This is only required in copy mode.
122  */
123 static void __xsk_rcv_memcpy(struct xdp_umem *umem, u64 addr, void *from_buf,
124 			     u32 len, u32 metalen)
125 {
126 	void *to_buf = xdp_umem_get_data(umem, addr);
127 
128 	addr = xsk_umem_add_offset_to_addr(addr);
129 	if (xskq_cons_crosses_non_contig_pg(umem, addr, len + metalen)) {
130 		void *next_pg_addr = umem->pages[(addr >> PAGE_SHIFT) + 1].addr;
131 		u64 page_start = addr & ~(PAGE_SIZE - 1);
132 		u64 first_len = PAGE_SIZE - (addr - page_start);
133 
134 		memcpy(to_buf, from_buf, first_len);
135 		memcpy(next_pg_addr, from_buf + first_len,
136 		       len + metalen - first_len);
137 
138 		return;
139 	}
140 
141 	memcpy(to_buf, from_buf, len + metalen);
142 }
143 
144 static int __xsk_rcv(struct xdp_sock *xs, struct xdp_buff *xdp, u32 len)
145 {
146 	u64 offset = xs->umem->headroom;
147 	u64 addr, memcpy_addr;
148 	void *from_buf;
149 	u32 metalen;
150 	int err;
151 
152 	if (!xskq_cons_peek_addr(xs->umem->fq, &addr, xs->umem) ||
153 	    len > xs->umem->chunk_size_nohr - XDP_PACKET_HEADROOM) {
154 		xs->rx_dropped++;
155 		return -ENOSPC;
156 	}
157 
158 	if (unlikely(xdp_data_meta_unsupported(xdp))) {
159 		from_buf = xdp->data;
160 		metalen = 0;
161 	} else {
162 		from_buf = xdp->data_meta;
163 		metalen = xdp->data - xdp->data_meta;
164 	}
165 
166 	memcpy_addr = xsk_umem_adjust_offset(xs->umem, addr, offset);
167 	__xsk_rcv_memcpy(xs->umem, memcpy_addr, from_buf, len, metalen);
168 
169 	offset += metalen;
170 	addr = xsk_umem_adjust_offset(xs->umem, addr, offset);
171 	err = xskq_prod_reserve_desc(xs->rx, addr, len);
172 	if (!err) {
173 		xskq_cons_release(xs->umem->fq);
174 		xdp_return_buff(xdp);
175 		return 0;
176 	}
177 
178 	xs->rx_dropped++;
179 	return err;
180 }
181 
182 static int __xsk_rcv_zc(struct xdp_sock *xs, struct xdp_buff *xdp, u32 len)
183 {
184 	int err = xskq_prod_reserve_desc(xs->rx, xdp->handle, len);
185 
186 	if (err)
187 		xs->rx_dropped++;
188 
189 	return err;
190 }
191 
192 static bool xsk_is_bound(struct xdp_sock *xs)
193 {
194 	if (READ_ONCE(xs->state) == XSK_BOUND) {
195 		/* Matches smp_wmb() in bind(). */
196 		smp_rmb();
197 		return true;
198 	}
199 	return false;
200 }
201 
202 static int xsk_rcv(struct xdp_sock *xs, struct xdp_buff *xdp)
203 {
204 	u32 len;
205 
206 	if (!xsk_is_bound(xs))
207 		return -EINVAL;
208 
209 	if (xs->dev != xdp->rxq->dev || xs->queue_id != xdp->rxq->queue_index)
210 		return -EINVAL;
211 
212 	len = xdp->data_end - xdp->data;
213 
214 	return (xdp->rxq->mem.type == MEM_TYPE_ZERO_COPY) ?
215 		__xsk_rcv_zc(xs, xdp, len) : __xsk_rcv(xs, xdp, len);
216 }
217 
218 static void xsk_flush(struct xdp_sock *xs)
219 {
220 	xskq_prod_submit(xs->rx);
221 	__xskq_cons_release(xs->umem->fq);
222 	sock_def_readable(&xs->sk);
223 }
224 
225 int xsk_generic_rcv(struct xdp_sock *xs, struct xdp_buff *xdp)
226 {
227 	u32 metalen = xdp->data - xdp->data_meta;
228 	u32 len = xdp->data_end - xdp->data;
229 	u64 offset = xs->umem->headroom;
230 	void *buffer;
231 	u64 addr;
232 	int err;
233 
234 	spin_lock_bh(&xs->rx_lock);
235 
236 	if (xs->dev != xdp->rxq->dev || xs->queue_id != xdp->rxq->queue_index) {
237 		err = -EINVAL;
238 		goto out_unlock;
239 	}
240 
241 	if (!xskq_cons_peek_addr(xs->umem->fq, &addr, xs->umem) ||
242 	    len > xs->umem->chunk_size_nohr - XDP_PACKET_HEADROOM) {
243 		err = -ENOSPC;
244 		goto out_drop;
245 	}
246 
247 	addr = xsk_umem_adjust_offset(xs->umem, addr, offset);
248 	buffer = xdp_umem_get_data(xs->umem, addr);
249 	memcpy(buffer, xdp->data_meta, len + metalen);
250 
251 	addr = xsk_umem_adjust_offset(xs->umem, addr, metalen);
252 	err = xskq_prod_reserve_desc(xs->rx, addr, len);
253 	if (err)
254 		goto out_drop;
255 
256 	xskq_cons_release(xs->umem->fq);
257 	xskq_prod_submit(xs->rx);
258 
259 	spin_unlock_bh(&xs->rx_lock);
260 
261 	xs->sk.sk_data_ready(&xs->sk);
262 	return 0;
263 
264 out_drop:
265 	xs->rx_dropped++;
266 out_unlock:
267 	spin_unlock_bh(&xs->rx_lock);
268 	return err;
269 }
270 
271 int __xsk_map_redirect(struct xdp_sock *xs, struct xdp_buff *xdp)
272 {
273 	struct list_head *flush_list = this_cpu_ptr(&xskmap_flush_list);
274 	int err;
275 
276 	err = xsk_rcv(xs, xdp);
277 	if (err)
278 		return err;
279 
280 	if (!xs->flush_node.prev)
281 		list_add(&xs->flush_node, flush_list);
282 
283 	return 0;
284 }
285 
286 void __xsk_map_flush(void)
287 {
288 	struct list_head *flush_list = this_cpu_ptr(&xskmap_flush_list);
289 	struct xdp_sock *xs, *tmp;
290 
291 	list_for_each_entry_safe(xs, tmp, flush_list, flush_node) {
292 		xsk_flush(xs);
293 		__list_del_clearprev(&xs->flush_node);
294 	}
295 }
296 
297 void xsk_umem_complete_tx(struct xdp_umem *umem, u32 nb_entries)
298 {
299 	xskq_prod_submit_n(umem->cq, nb_entries);
300 }
301 EXPORT_SYMBOL(xsk_umem_complete_tx);
302 
303 void xsk_umem_consume_tx_done(struct xdp_umem *umem)
304 {
305 	struct xdp_sock *xs;
306 
307 	rcu_read_lock();
308 	list_for_each_entry_rcu(xs, &umem->xsk_list, list) {
309 		__xskq_cons_release(xs->tx);
310 		xs->sk.sk_write_space(&xs->sk);
311 	}
312 	rcu_read_unlock();
313 }
314 EXPORT_SYMBOL(xsk_umem_consume_tx_done);
315 
316 bool xsk_umem_consume_tx(struct xdp_umem *umem, struct xdp_desc *desc)
317 {
318 	struct xdp_sock *xs;
319 
320 	rcu_read_lock();
321 	list_for_each_entry_rcu(xs, &umem->xsk_list, list) {
322 		if (!xskq_cons_peek_desc(xs->tx, desc, umem))
323 			continue;
324 
325 		/* This is the backpreassure mechanism for the Tx path.
326 		 * Reserve space in the completion queue and only proceed
327 		 * if there is space in it. This avoids having to implement
328 		 * any buffering in the Tx path.
329 		 */
330 		if (xskq_prod_reserve_addr(umem->cq, desc->addr))
331 			goto out;
332 
333 		xskq_cons_release(xs->tx);
334 		rcu_read_unlock();
335 		return true;
336 	}
337 
338 out:
339 	rcu_read_unlock();
340 	return false;
341 }
342 EXPORT_SYMBOL(xsk_umem_consume_tx);
343 
344 static int xsk_wakeup(struct xdp_sock *xs, u8 flags)
345 {
346 	struct net_device *dev = xs->dev;
347 	int err;
348 
349 	rcu_read_lock();
350 	err = dev->netdev_ops->ndo_xsk_wakeup(dev, xs->queue_id, flags);
351 	rcu_read_unlock();
352 
353 	return err;
354 }
355 
356 static int xsk_zc_xmit(struct xdp_sock *xs)
357 {
358 	return xsk_wakeup(xs, XDP_WAKEUP_TX);
359 }
360 
361 static void xsk_destruct_skb(struct sk_buff *skb)
362 {
363 	u64 addr = (u64)(long)skb_shinfo(skb)->destructor_arg;
364 	struct xdp_sock *xs = xdp_sk(skb->sk);
365 	unsigned long flags;
366 
367 	spin_lock_irqsave(&xs->tx_completion_lock, flags);
368 	xskq_prod_submit_addr(xs->umem->cq, addr);
369 	spin_unlock_irqrestore(&xs->tx_completion_lock, flags);
370 
371 	sock_wfree(skb);
372 }
373 
374 static int xsk_generic_xmit(struct sock *sk)
375 {
376 	struct xdp_sock *xs = xdp_sk(sk);
377 	u32 max_batch = TX_BATCH_SIZE;
378 	bool sent_frame = false;
379 	struct xdp_desc desc;
380 	struct sk_buff *skb;
381 	int err = 0;
382 
383 	mutex_lock(&xs->mutex);
384 
385 	if (xs->queue_id >= xs->dev->real_num_tx_queues)
386 		goto out;
387 
388 	while (xskq_cons_peek_desc(xs->tx, &desc, xs->umem)) {
389 		char *buffer;
390 		u64 addr;
391 		u32 len;
392 
393 		if (max_batch-- == 0) {
394 			err = -EAGAIN;
395 			goto out;
396 		}
397 
398 		len = desc.len;
399 		skb = sock_alloc_send_skb(sk, len, 1, &err);
400 		if (unlikely(!skb)) {
401 			err = -EAGAIN;
402 			goto out;
403 		}
404 
405 		skb_put(skb, len);
406 		addr = desc.addr;
407 		buffer = xdp_umem_get_data(xs->umem, addr);
408 		err = skb_store_bits(skb, 0, buffer, len);
409 		/* This is the backpreassure mechanism for the Tx path.
410 		 * Reserve space in the completion queue and only proceed
411 		 * if there is space in it. This avoids having to implement
412 		 * any buffering in the Tx path.
413 		 */
414 		if (unlikely(err) || xskq_prod_reserve(xs->umem->cq)) {
415 			kfree_skb(skb);
416 			goto out;
417 		}
418 
419 		skb->dev = xs->dev;
420 		skb->priority = sk->sk_priority;
421 		skb->mark = sk->sk_mark;
422 		skb_shinfo(skb)->destructor_arg = (void *)(long)desc.addr;
423 		skb->destructor = xsk_destruct_skb;
424 
425 		err = dev_direct_xmit(skb, xs->queue_id);
426 		xskq_cons_release(xs->tx);
427 		/* Ignore NET_XMIT_CN as packet might have been sent */
428 		if (err == NET_XMIT_DROP || err == NETDEV_TX_BUSY) {
429 			/* SKB completed but not sent */
430 			err = -EBUSY;
431 			goto out;
432 		}
433 
434 		sent_frame = true;
435 	}
436 
437 out:
438 	if (sent_frame)
439 		sk->sk_write_space(sk);
440 
441 	mutex_unlock(&xs->mutex);
442 	return err;
443 }
444 
445 static int __xsk_sendmsg(struct sock *sk)
446 {
447 	struct xdp_sock *xs = xdp_sk(sk);
448 
449 	if (unlikely(!(xs->dev->flags & IFF_UP)))
450 		return -ENETDOWN;
451 	if (unlikely(!xs->tx))
452 		return -ENOBUFS;
453 
454 	return xs->zc ? xsk_zc_xmit(xs) : xsk_generic_xmit(sk);
455 }
456 
457 static int xsk_sendmsg(struct socket *sock, struct msghdr *m, size_t total_len)
458 {
459 	bool need_wait = !(m->msg_flags & MSG_DONTWAIT);
460 	struct sock *sk = sock->sk;
461 	struct xdp_sock *xs = xdp_sk(sk);
462 
463 	if (unlikely(!xsk_is_bound(xs)))
464 		return -ENXIO;
465 	if (unlikely(need_wait))
466 		return -EOPNOTSUPP;
467 
468 	return __xsk_sendmsg(sk);
469 }
470 
471 static __poll_t xsk_poll(struct file *file, struct socket *sock,
472 			     struct poll_table_struct *wait)
473 {
474 	__poll_t mask = datagram_poll(file, sock, wait);
475 	struct sock *sk = sock->sk;
476 	struct xdp_sock *xs = xdp_sk(sk);
477 	struct xdp_umem *umem;
478 
479 	if (unlikely(!xsk_is_bound(xs)))
480 		return mask;
481 
482 	umem = xs->umem;
483 
484 	if (umem->need_wakeup) {
485 		if (xs->zc)
486 			xsk_wakeup(xs, umem->need_wakeup);
487 		else
488 			/* Poll needs to drive Tx also in copy mode */
489 			__xsk_sendmsg(sk);
490 	}
491 
492 	if (xs->rx && !xskq_prod_is_empty(xs->rx))
493 		mask |= EPOLLIN | EPOLLRDNORM;
494 	if (xs->tx && !xskq_cons_is_full(xs->tx))
495 		mask |= EPOLLOUT | EPOLLWRNORM;
496 
497 	return mask;
498 }
499 
500 static int xsk_init_queue(u32 entries, struct xsk_queue **queue,
501 			  bool umem_queue)
502 {
503 	struct xsk_queue *q;
504 
505 	if (entries == 0 || *queue || !is_power_of_2(entries))
506 		return -EINVAL;
507 
508 	q = xskq_create(entries, umem_queue);
509 	if (!q)
510 		return -ENOMEM;
511 
512 	/* Make sure queue is ready before it can be seen by others */
513 	smp_wmb();
514 	WRITE_ONCE(*queue, q);
515 	return 0;
516 }
517 
518 static void xsk_unbind_dev(struct xdp_sock *xs)
519 {
520 	struct net_device *dev = xs->dev;
521 
522 	if (xs->state != XSK_BOUND)
523 		return;
524 	WRITE_ONCE(xs->state, XSK_UNBOUND);
525 
526 	/* Wait for driver to stop using the xdp socket. */
527 	xdp_del_sk_umem(xs->umem, xs);
528 	xs->dev = NULL;
529 	synchronize_net();
530 	dev_put(dev);
531 }
532 
533 static struct xsk_map *xsk_get_map_list_entry(struct xdp_sock *xs,
534 					      struct xdp_sock ***map_entry)
535 {
536 	struct xsk_map *map = NULL;
537 	struct xsk_map_node *node;
538 
539 	*map_entry = NULL;
540 
541 	spin_lock_bh(&xs->map_list_lock);
542 	node = list_first_entry_or_null(&xs->map_list, struct xsk_map_node,
543 					node);
544 	if (node) {
545 		WARN_ON(xsk_map_inc(node->map));
546 		map = node->map;
547 		*map_entry = node->map_entry;
548 	}
549 	spin_unlock_bh(&xs->map_list_lock);
550 	return map;
551 }
552 
553 static void xsk_delete_from_maps(struct xdp_sock *xs)
554 {
555 	/* This function removes the current XDP socket from all the
556 	 * maps it resides in. We need to take extra care here, due to
557 	 * the two locks involved. Each map has a lock synchronizing
558 	 * updates to the entries, and each socket has a lock that
559 	 * synchronizes access to the list of maps (map_list). For
560 	 * deadlock avoidance the locks need to be taken in the order
561 	 * "map lock"->"socket map list lock". We start off by
562 	 * accessing the socket map list, and take a reference to the
563 	 * map to guarantee existence between the
564 	 * xsk_get_map_list_entry() and xsk_map_try_sock_delete()
565 	 * calls. Then we ask the map to remove the socket, which
566 	 * tries to remove the socket from the map. Note that there
567 	 * might be updates to the map between
568 	 * xsk_get_map_list_entry() and xsk_map_try_sock_delete().
569 	 */
570 	struct xdp_sock **map_entry = NULL;
571 	struct xsk_map *map;
572 
573 	while ((map = xsk_get_map_list_entry(xs, &map_entry))) {
574 		xsk_map_try_sock_delete(map, xs, map_entry);
575 		xsk_map_put(map);
576 	}
577 }
578 
579 static int xsk_release(struct socket *sock)
580 {
581 	struct sock *sk = sock->sk;
582 	struct xdp_sock *xs = xdp_sk(sk);
583 	struct net *net;
584 
585 	if (!sk)
586 		return 0;
587 
588 	net = sock_net(sk);
589 
590 	mutex_lock(&net->xdp.lock);
591 	sk_del_node_init_rcu(sk);
592 	mutex_unlock(&net->xdp.lock);
593 
594 	local_bh_disable();
595 	sock_prot_inuse_add(net, sk->sk_prot, -1);
596 	local_bh_enable();
597 
598 	xsk_delete_from_maps(xs);
599 	mutex_lock(&xs->mutex);
600 	xsk_unbind_dev(xs);
601 	mutex_unlock(&xs->mutex);
602 
603 	xskq_destroy(xs->rx);
604 	xskq_destroy(xs->tx);
605 
606 	sock_orphan(sk);
607 	sock->sk = NULL;
608 
609 	sk_refcnt_debug_release(sk);
610 	sock_put(sk);
611 
612 	return 0;
613 }
614 
615 static struct socket *xsk_lookup_xsk_from_fd(int fd)
616 {
617 	struct socket *sock;
618 	int err;
619 
620 	sock = sockfd_lookup(fd, &err);
621 	if (!sock)
622 		return ERR_PTR(-ENOTSOCK);
623 
624 	if (sock->sk->sk_family != PF_XDP) {
625 		sockfd_put(sock);
626 		return ERR_PTR(-ENOPROTOOPT);
627 	}
628 
629 	return sock;
630 }
631 
632 /* Check if umem pages are contiguous.
633  * If zero-copy mode, use the DMA address to do the page contiguity check
634  * For all other modes we use addr (kernel virtual address)
635  * Store the result in the low bits of addr.
636  */
637 static void xsk_check_page_contiguity(struct xdp_umem *umem, u32 flags)
638 {
639 	struct xdp_umem_page *pgs = umem->pages;
640 	int i, is_contig;
641 
642 	for (i = 0; i < umem->npgs - 1; i++) {
643 		is_contig = (flags & XDP_ZEROCOPY) ?
644 			(pgs[i].dma + PAGE_SIZE == pgs[i + 1].dma) :
645 			(pgs[i].addr + PAGE_SIZE == pgs[i + 1].addr);
646 		pgs[i].addr += is_contig << XSK_NEXT_PG_CONTIG_SHIFT;
647 	}
648 }
649 
650 static int xsk_bind(struct socket *sock, struct sockaddr *addr, int addr_len)
651 {
652 	struct sockaddr_xdp *sxdp = (struct sockaddr_xdp *)addr;
653 	struct sock *sk = sock->sk;
654 	struct xdp_sock *xs = xdp_sk(sk);
655 	struct net_device *dev;
656 	u32 flags, qid;
657 	int err = 0;
658 
659 	if (addr_len < sizeof(struct sockaddr_xdp))
660 		return -EINVAL;
661 	if (sxdp->sxdp_family != AF_XDP)
662 		return -EINVAL;
663 
664 	flags = sxdp->sxdp_flags;
665 	if (flags & ~(XDP_SHARED_UMEM | XDP_COPY | XDP_ZEROCOPY |
666 		      XDP_USE_NEED_WAKEUP))
667 		return -EINVAL;
668 
669 	rtnl_lock();
670 	mutex_lock(&xs->mutex);
671 	if (xs->state != XSK_READY) {
672 		err = -EBUSY;
673 		goto out_release;
674 	}
675 
676 	dev = dev_get_by_index(sock_net(sk), sxdp->sxdp_ifindex);
677 	if (!dev) {
678 		err = -ENODEV;
679 		goto out_release;
680 	}
681 
682 	if (!xs->rx && !xs->tx) {
683 		err = -EINVAL;
684 		goto out_unlock;
685 	}
686 
687 	qid = sxdp->sxdp_queue_id;
688 
689 	if (flags & XDP_SHARED_UMEM) {
690 		struct xdp_sock *umem_xs;
691 		struct socket *sock;
692 
693 		if ((flags & XDP_COPY) || (flags & XDP_ZEROCOPY) ||
694 		    (flags & XDP_USE_NEED_WAKEUP)) {
695 			/* Cannot specify flags for shared sockets. */
696 			err = -EINVAL;
697 			goto out_unlock;
698 		}
699 
700 		if (xs->umem) {
701 			/* We have already our own. */
702 			err = -EINVAL;
703 			goto out_unlock;
704 		}
705 
706 		sock = xsk_lookup_xsk_from_fd(sxdp->sxdp_shared_umem_fd);
707 		if (IS_ERR(sock)) {
708 			err = PTR_ERR(sock);
709 			goto out_unlock;
710 		}
711 
712 		umem_xs = xdp_sk(sock->sk);
713 		if (!xsk_is_bound(umem_xs)) {
714 			err = -EBADF;
715 			sockfd_put(sock);
716 			goto out_unlock;
717 		}
718 		if (umem_xs->dev != dev || umem_xs->queue_id != qid) {
719 			err = -EINVAL;
720 			sockfd_put(sock);
721 			goto out_unlock;
722 		}
723 
724 		xdp_get_umem(umem_xs->umem);
725 		WRITE_ONCE(xs->umem, umem_xs->umem);
726 		sockfd_put(sock);
727 	} else if (!xs->umem || !xdp_umem_validate_queues(xs->umem)) {
728 		err = -EINVAL;
729 		goto out_unlock;
730 	} else {
731 		/* This xsk has its own umem. */
732 		xskq_set_umem(xs->umem->fq, xs->umem->size,
733 			      xs->umem->chunk_mask);
734 		xskq_set_umem(xs->umem->cq, xs->umem->size,
735 			      xs->umem->chunk_mask);
736 
737 		err = xdp_umem_assign_dev(xs->umem, dev, qid, flags);
738 		if (err)
739 			goto out_unlock;
740 
741 		xsk_check_page_contiguity(xs->umem, flags);
742 	}
743 
744 	xs->dev = dev;
745 	xs->zc = xs->umem->zc;
746 	xs->queue_id = qid;
747 	xskq_set_umem(xs->rx, xs->umem->size, xs->umem->chunk_mask);
748 	xskq_set_umem(xs->tx, xs->umem->size, xs->umem->chunk_mask);
749 	xdp_add_sk_umem(xs->umem, xs);
750 
751 out_unlock:
752 	if (err) {
753 		dev_put(dev);
754 	} else {
755 		/* Matches smp_rmb() in bind() for shared umem
756 		 * sockets, and xsk_is_bound().
757 		 */
758 		smp_wmb();
759 		WRITE_ONCE(xs->state, XSK_BOUND);
760 	}
761 out_release:
762 	mutex_unlock(&xs->mutex);
763 	rtnl_unlock();
764 	return err;
765 }
766 
767 struct xdp_umem_reg_v1 {
768 	__u64 addr; /* Start of packet data area */
769 	__u64 len; /* Length of packet data area */
770 	__u32 chunk_size;
771 	__u32 headroom;
772 };
773 
774 static int xsk_setsockopt(struct socket *sock, int level, int optname,
775 			  char __user *optval, unsigned int optlen)
776 {
777 	struct sock *sk = sock->sk;
778 	struct xdp_sock *xs = xdp_sk(sk);
779 	int err;
780 
781 	if (level != SOL_XDP)
782 		return -ENOPROTOOPT;
783 
784 	switch (optname) {
785 	case XDP_RX_RING:
786 	case XDP_TX_RING:
787 	{
788 		struct xsk_queue **q;
789 		int entries;
790 
791 		if (optlen < sizeof(entries))
792 			return -EINVAL;
793 		if (copy_from_user(&entries, optval, sizeof(entries)))
794 			return -EFAULT;
795 
796 		mutex_lock(&xs->mutex);
797 		if (xs->state != XSK_READY) {
798 			mutex_unlock(&xs->mutex);
799 			return -EBUSY;
800 		}
801 		q = (optname == XDP_TX_RING) ? &xs->tx : &xs->rx;
802 		err = xsk_init_queue(entries, q, false);
803 		if (!err && optname == XDP_TX_RING)
804 			/* Tx needs to be explicitly woken up the first time */
805 			xs->tx->ring->flags |= XDP_RING_NEED_WAKEUP;
806 		mutex_unlock(&xs->mutex);
807 		return err;
808 	}
809 	case XDP_UMEM_REG:
810 	{
811 		size_t mr_size = sizeof(struct xdp_umem_reg);
812 		struct xdp_umem_reg mr = {};
813 		struct xdp_umem *umem;
814 
815 		if (optlen < sizeof(struct xdp_umem_reg_v1))
816 			return -EINVAL;
817 		else if (optlen < sizeof(mr))
818 			mr_size = sizeof(struct xdp_umem_reg_v1);
819 
820 		if (copy_from_user(&mr, optval, mr_size))
821 			return -EFAULT;
822 
823 		mutex_lock(&xs->mutex);
824 		if (xs->state != XSK_READY || xs->umem) {
825 			mutex_unlock(&xs->mutex);
826 			return -EBUSY;
827 		}
828 
829 		umem = xdp_umem_create(&mr);
830 		if (IS_ERR(umem)) {
831 			mutex_unlock(&xs->mutex);
832 			return PTR_ERR(umem);
833 		}
834 
835 		/* Make sure umem is ready before it can be seen by others */
836 		smp_wmb();
837 		WRITE_ONCE(xs->umem, umem);
838 		mutex_unlock(&xs->mutex);
839 		return 0;
840 	}
841 	case XDP_UMEM_FILL_RING:
842 	case XDP_UMEM_COMPLETION_RING:
843 	{
844 		struct xsk_queue **q;
845 		int entries;
846 
847 		if (copy_from_user(&entries, optval, sizeof(entries)))
848 			return -EFAULT;
849 
850 		mutex_lock(&xs->mutex);
851 		if (xs->state != XSK_READY) {
852 			mutex_unlock(&xs->mutex);
853 			return -EBUSY;
854 		}
855 		if (!xs->umem) {
856 			mutex_unlock(&xs->mutex);
857 			return -EINVAL;
858 		}
859 
860 		q = (optname == XDP_UMEM_FILL_RING) ? &xs->umem->fq :
861 			&xs->umem->cq;
862 		err = xsk_init_queue(entries, q, true);
863 		mutex_unlock(&xs->mutex);
864 		return err;
865 	}
866 	default:
867 		break;
868 	}
869 
870 	return -ENOPROTOOPT;
871 }
872 
873 static void xsk_enter_rxtx_offsets(struct xdp_ring_offset_v1 *ring)
874 {
875 	ring->producer = offsetof(struct xdp_rxtx_ring, ptrs.producer);
876 	ring->consumer = offsetof(struct xdp_rxtx_ring, ptrs.consumer);
877 	ring->desc = offsetof(struct xdp_rxtx_ring, desc);
878 }
879 
880 static void xsk_enter_umem_offsets(struct xdp_ring_offset_v1 *ring)
881 {
882 	ring->producer = offsetof(struct xdp_umem_ring, ptrs.producer);
883 	ring->consumer = offsetof(struct xdp_umem_ring, ptrs.consumer);
884 	ring->desc = offsetof(struct xdp_umem_ring, desc);
885 }
886 
887 static int xsk_getsockopt(struct socket *sock, int level, int optname,
888 			  char __user *optval, int __user *optlen)
889 {
890 	struct sock *sk = sock->sk;
891 	struct xdp_sock *xs = xdp_sk(sk);
892 	int len;
893 
894 	if (level != SOL_XDP)
895 		return -ENOPROTOOPT;
896 
897 	if (get_user(len, optlen))
898 		return -EFAULT;
899 	if (len < 0)
900 		return -EINVAL;
901 
902 	switch (optname) {
903 	case XDP_STATISTICS:
904 	{
905 		struct xdp_statistics stats;
906 
907 		if (len < sizeof(stats))
908 			return -EINVAL;
909 
910 		mutex_lock(&xs->mutex);
911 		stats.rx_dropped = xs->rx_dropped;
912 		stats.rx_invalid_descs = xskq_nb_invalid_descs(xs->rx);
913 		stats.tx_invalid_descs = xskq_nb_invalid_descs(xs->tx);
914 		mutex_unlock(&xs->mutex);
915 
916 		if (copy_to_user(optval, &stats, sizeof(stats)))
917 			return -EFAULT;
918 		if (put_user(sizeof(stats), optlen))
919 			return -EFAULT;
920 
921 		return 0;
922 	}
923 	case XDP_MMAP_OFFSETS:
924 	{
925 		struct xdp_mmap_offsets off;
926 		struct xdp_mmap_offsets_v1 off_v1;
927 		bool flags_supported = true;
928 		void *to_copy;
929 
930 		if (len < sizeof(off_v1))
931 			return -EINVAL;
932 		else if (len < sizeof(off))
933 			flags_supported = false;
934 
935 		if (flags_supported) {
936 			/* xdp_ring_offset is identical to xdp_ring_offset_v1
937 			 * except for the flags field added to the end.
938 			 */
939 			xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *)
940 					       &off.rx);
941 			xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *)
942 					       &off.tx);
943 			xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *)
944 					       &off.fr);
945 			xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *)
946 					       &off.cr);
947 			off.rx.flags = offsetof(struct xdp_rxtx_ring,
948 						ptrs.flags);
949 			off.tx.flags = offsetof(struct xdp_rxtx_ring,
950 						ptrs.flags);
951 			off.fr.flags = offsetof(struct xdp_umem_ring,
952 						ptrs.flags);
953 			off.cr.flags = offsetof(struct xdp_umem_ring,
954 						ptrs.flags);
955 
956 			len = sizeof(off);
957 			to_copy = &off;
958 		} else {
959 			xsk_enter_rxtx_offsets(&off_v1.rx);
960 			xsk_enter_rxtx_offsets(&off_v1.tx);
961 			xsk_enter_umem_offsets(&off_v1.fr);
962 			xsk_enter_umem_offsets(&off_v1.cr);
963 
964 			len = sizeof(off_v1);
965 			to_copy = &off_v1;
966 		}
967 
968 		if (copy_to_user(optval, to_copy, len))
969 			return -EFAULT;
970 		if (put_user(len, optlen))
971 			return -EFAULT;
972 
973 		return 0;
974 	}
975 	case XDP_OPTIONS:
976 	{
977 		struct xdp_options opts = {};
978 
979 		if (len < sizeof(opts))
980 			return -EINVAL;
981 
982 		mutex_lock(&xs->mutex);
983 		if (xs->zc)
984 			opts.flags |= XDP_OPTIONS_ZEROCOPY;
985 		mutex_unlock(&xs->mutex);
986 
987 		len = sizeof(opts);
988 		if (copy_to_user(optval, &opts, len))
989 			return -EFAULT;
990 		if (put_user(len, optlen))
991 			return -EFAULT;
992 
993 		return 0;
994 	}
995 	default:
996 		break;
997 	}
998 
999 	return -EOPNOTSUPP;
1000 }
1001 
1002 static int xsk_mmap(struct file *file, struct socket *sock,
1003 		    struct vm_area_struct *vma)
1004 {
1005 	loff_t offset = (loff_t)vma->vm_pgoff << PAGE_SHIFT;
1006 	unsigned long size = vma->vm_end - vma->vm_start;
1007 	struct xdp_sock *xs = xdp_sk(sock->sk);
1008 	struct xsk_queue *q = NULL;
1009 	struct xdp_umem *umem;
1010 	unsigned long pfn;
1011 	struct page *qpg;
1012 
1013 	if (READ_ONCE(xs->state) != XSK_READY)
1014 		return -EBUSY;
1015 
1016 	if (offset == XDP_PGOFF_RX_RING) {
1017 		q = READ_ONCE(xs->rx);
1018 	} else if (offset == XDP_PGOFF_TX_RING) {
1019 		q = READ_ONCE(xs->tx);
1020 	} else {
1021 		umem = READ_ONCE(xs->umem);
1022 		if (!umem)
1023 			return -EINVAL;
1024 
1025 		/* Matches the smp_wmb() in XDP_UMEM_REG */
1026 		smp_rmb();
1027 		if (offset == XDP_UMEM_PGOFF_FILL_RING)
1028 			q = READ_ONCE(umem->fq);
1029 		else if (offset == XDP_UMEM_PGOFF_COMPLETION_RING)
1030 			q = READ_ONCE(umem->cq);
1031 	}
1032 
1033 	if (!q)
1034 		return -EINVAL;
1035 
1036 	/* Matches the smp_wmb() in xsk_init_queue */
1037 	smp_rmb();
1038 	qpg = virt_to_head_page(q->ring);
1039 	if (size > page_size(qpg))
1040 		return -EINVAL;
1041 
1042 	pfn = virt_to_phys(q->ring) >> PAGE_SHIFT;
1043 	return remap_pfn_range(vma, vma->vm_start, pfn,
1044 			       size, vma->vm_page_prot);
1045 }
1046 
1047 static int xsk_notifier(struct notifier_block *this,
1048 			unsigned long msg, void *ptr)
1049 {
1050 	struct net_device *dev = netdev_notifier_info_to_dev(ptr);
1051 	struct net *net = dev_net(dev);
1052 	struct sock *sk;
1053 
1054 	switch (msg) {
1055 	case NETDEV_UNREGISTER:
1056 		mutex_lock(&net->xdp.lock);
1057 		sk_for_each(sk, &net->xdp.list) {
1058 			struct xdp_sock *xs = xdp_sk(sk);
1059 
1060 			mutex_lock(&xs->mutex);
1061 			if (xs->dev == dev) {
1062 				sk->sk_err = ENETDOWN;
1063 				if (!sock_flag(sk, SOCK_DEAD))
1064 					sk->sk_error_report(sk);
1065 
1066 				xsk_unbind_dev(xs);
1067 
1068 				/* Clear device references in umem. */
1069 				xdp_umem_clear_dev(xs->umem);
1070 			}
1071 			mutex_unlock(&xs->mutex);
1072 		}
1073 		mutex_unlock(&net->xdp.lock);
1074 		break;
1075 	}
1076 	return NOTIFY_DONE;
1077 }
1078 
1079 static struct proto xsk_proto = {
1080 	.name =		"XDP",
1081 	.owner =	THIS_MODULE,
1082 	.obj_size =	sizeof(struct xdp_sock),
1083 };
1084 
1085 static const struct proto_ops xsk_proto_ops = {
1086 	.family		= PF_XDP,
1087 	.owner		= THIS_MODULE,
1088 	.release	= xsk_release,
1089 	.bind		= xsk_bind,
1090 	.connect	= sock_no_connect,
1091 	.socketpair	= sock_no_socketpair,
1092 	.accept		= sock_no_accept,
1093 	.getname	= sock_no_getname,
1094 	.poll		= xsk_poll,
1095 	.ioctl		= sock_no_ioctl,
1096 	.listen		= sock_no_listen,
1097 	.shutdown	= sock_no_shutdown,
1098 	.setsockopt	= xsk_setsockopt,
1099 	.getsockopt	= xsk_getsockopt,
1100 	.sendmsg	= xsk_sendmsg,
1101 	.recvmsg	= sock_no_recvmsg,
1102 	.mmap		= xsk_mmap,
1103 	.sendpage	= sock_no_sendpage,
1104 };
1105 
1106 static void xsk_destruct(struct sock *sk)
1107 {
1108 	struct xdp_sock *xs = xdp_sk(sk);
1109 
1110 	if (!sock_flag(sk, SOCK_DEAD))
1111 		return;
1112 
1113 	xdp_put_umem(xs->umem);
1114 
1115 	sk_refcnt_debug_dec(sk);
1116 }
1117 
1118 static int xsk_create(struct net *net, struct socket *sock, int protocol,
1119 		      int kern)
1120 {
1121 	struct sock *sk;
1122 	struct xdp_sock *xs;
1123 
1124 	if (!ns_capable(net->user_ns, CAP_NET_RAW))
1125 		return -EPERM;
1126 	if (sock->type != SOCK_RAW)
1127 		return -ESOCKTNOSUPPORT;
1128 
1129 	if (protocol)
1130 		return -EPROTONOSUPPORT;
1131 
1132 	sock->state = SS_UNCONNECTED;
1133 
1134 	sk = sk_alloc(net, PF_XDP, GFP_KERNEL, &xsk_proto, kern);
1135 	if (!sk)
1136 		return -ENOBUFS;
1137 
1138 	sock->ops = &xsk_proto_ops;
1139 
1140 	sock_init_data(sock, sk);
1141 
1142 	sk->sk_family = PF_XDP;
1143 
1144 	sk->sk_destruct = xsk_destruct;
1145 	sk_refcnt_debug_inc(sk);
1146 
1147 	sock_set_flag(sk, SOCK_RCU_FREE);
1148 
1149 	xs = xdp_sk(sk);
1150 	xs->state = XSK_READY;
1151 	mutex_init(&xs->mutex);
1152 	spin_lock_init(&xs->rx_lock);
1153 	spin_lock_init(&xs->tx_completion_lock);
1154 
1155 	INIT_LIST_HEAD(&xs->map_list);
1156 	spin_lock_init(&xs->map_list_lock);
1157 
1158 	mutex_lock(&net->xdp.lock);
1159 	sk_add_node_rcu(sk, &net->xdp.list);
1160 	mutex_unlock(&net->xdp.lock);
1161 
1162 	local_bh_disable();
1163 	sock_prot_inuse_add(net, &xsk_proto, 1);
1164 	local_bh_enable();
1165 
1166 	return 0;
1167 }
1168 
1169 static const struct net_proto_family xsk_family_ops = {
1170 	.family = PF_XDP,
1171 	.create = xsk_create,
1172 	.owner	= THIS_MODULE,
1173 };
1174 
1175 static struct notifier_block xsk_netdev_notifier = {
1176 	.notifier_call	= xsk_notifier,
1177 };
1178 
1179 static int __net_init xsk_net_init(struct net *net)
1180 {
1181 	mutex_init(&net->xdp.lock);
1182 	INIT_HLIST_HEAD(&net->xdp.list);
1183 	return 0;
1184 }
1185 
1186 static void __net_exit xsk_net_exit(struct net *net)
1187 {
1188 	WARN_ON_ONCE(!hlist_empty(&net->xdp.list));
1189 }
1190 
1191 static struct pernet_operations xsk_net_ops = {
1192 	.init = xsk_net_init,
1193 	.exit = xsk_net_exit,
1194 };
1195 
1196 static int __init xsk_init(void)
1197 {
1198 	int err, cpu;
1199 
1200 	err = proto_register(&xsk_proto, 0 /* no slab */);
1201 	if (err)
1202 		goto out;
1203 
1204 	err = sock_register(&xsk_family_ops);
1205 	if (err)
1206 		goto out_proto;
1207 
1208 	err = register_pernet_subsys(&xsk_net_ops);
1209 	if (err)
1210 		goto out_sk;
1211 
1212 	err = register_netdevice_notifier(&xsk_netdev_notifier);
1213 	if (err)
1214 		goto out_pernet;
1215 
1216 	for_each_possible_cpu(cpu)
1217 		INIT_LIST_HEAD(&per_cpu(xskmap_flush_list, cpu));
1218 	return 0;
1219 
1220 out_pernet:
1221 	unregister_pernet_subsys(&xsk_net_ops);
1222 out_sk:
1223 	sock_unregister(PF_XDP);
1224 out_proto:
1225 	proto_unregister(&xsk_proto);
1226 out:
1227 	return err;
1228 }
1229 
1230 fs_initcall(xsk_init);
1231