xref: /linux/net/xdp/xsk.c (revision 17cfcb68af3bc7d5e8ae08779b1853310a2949f3)
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 bool xsk_is_setup_for_bpf_map(struct xdp_sock *xs)
35 {
36 	return READ_ONCE(xs->rx) &&  READ_ONCE(xs->umem) &&
37 		READ_ONCE(xs->umem->fq);
38 }
39 
40 bool xsk_umem_has_addrs(struct xdp_umem *umem, u32 cnt)
41 {
42 	return xskq_has_addrs(umem->fq, cnt);
43 }
44 EXPORT_SYMBOL(xsk_umem_has_addrs);
45 
46 u64 *xsk_umem_peek_addr(struct xdp_umem *umem, u64 *addr)
47 {
48 	return xskq_peek_addr(umem->fq, addr, umem);
49 }
50 EXPORT_SYMBOL(xsk_umem_peek_addr);
51 
52 void xsk_umem_discard_addr(struct xdp_umem *umem)
53 {
54 	xskq_discard_addr(umem->fq);
55 }
56 EXPORT_SYMBOL(xsk_umem_discard_addr);
57 
58 void xsk_set_rx_need_wakeup(struct xdp_umem *umem)
59 {
60 	if (umem->need_wakeup & XDP_WAKEUP_RX)
61 		return;
62 
63 	umem->fq->ring->flags |= XDP_RING_NEED_WAKEUP;
64 	umem->need_wakeup |= XDP_WAKEUP_RX;
65 }
66 EXPORT_SYMBOL(xsk_set_rx_need_wakeup);
67 
68 void xsk_set_tx_need_wakeup(struct xdp_umem *umem)
69 {
70 	struct xdp_sock *xs;
71 
72 	if (umem->need_wakeup & XDP_WAKEUP_TX)
73 		return;
74 
75 	rcu_read_lock();
76 	list_for_each_entry_rcu(xs, &umem->xsk_list, list) {
77 		xs->tx->ring->flags |= XDP_RING_NEED_WAKEUP;
78 	}
79 	rcu_read_unlock();
80 
81 	umem->need_wakeup |= XDP_WAKEUP_TX;
82 }
83 EXPORT_SYMBOL(xsk_set_tx_need_wakeup);
84 
85 void xsk_clear_rx_need_wakeup(struct xdp_umem *umem)
86 {
87 	if (!(umem->need_wakeup & XDP_WAKEUP_RX))
88 		return;
89 
90 	umem->fq->ring->flags &= ~XDP_RING_NEED_WAKEUP;
91 	umem->need_wakeup &= ~XDP_WAKEUP_RX;
92 }
93 EXPORT_SYMBOL(xsk_clear_rx_need_wakeup);
94 
95 void xsk_clear_tx_need_wakeup(struct xdp_umem *umem)
96 {
97 	struct xdp_sock *xs;
98 
99 	if (!(umem->need_wakeup & XDP_WAKEUP_TX))
100 		return;
101 
102 	rcu_read_lock();
103 	list_for_each_entry_rcu(xs, &umem->xsk_list, list) {
104 		xs->tx->ring->flags &= ~XDP_RING_NEED_WAKEUP;
105 	}
106 	rcu_read_unlock();
107 
108 	umem->need_wakeup &= ~XDP_WAKEUP_TX;
109 }
110 EXPORT_SYMBOL(xsk_clear_tx_need_wakeup);
111 
112 bool xsk_umem_uses_need_wakeup(struct xdp_umem *umem)
113 {
114 	return umem->flags & XDP_UMEM_USES_NEED_WAKEUP;
115 }
116 EXPORT_SYMBOL(xsk_umem_uses_need_wakeup);
117 
118 /* If a buffer crosses a page boundary, we need to do 2 memcpy's, one for
119  * each page. This is only required in copy mode.
120  */
121 static void __xsk_rcv_memcpy(struct xdp_umem *umem, u64 addr, void *from_buf,
122 			     u32 len, u32 metalen)
123 {
124 	void *to_buf = xdp_umem_get_data(umem, addr);
125 
126 	addr = xsk_umem_add_offset_to_addr(addr);
127 	if (xskq_crosses_non_contig_pg(umem, addr, len + metalen)) {
128 		void *next_pg_addr = umem->pages[(addr >> PAGE_SHIFT) + 1].addr;
129 		u64 page_start = addr & ~(PAGE_SIZE - 1);
130 		u64 first_len = PAGE_SIZE - (addr - page_start);
131 
132 		memcpy(to_buf, from_buf, first_len + metalen);
133 		memcpy(next_pg_addr, from_buf + first_len, len - first_len);
134 
135 		return;
136 	}
137 
138 	memcpy(to_buf, from_buf, len + metalen);
139 }
140 
141 static int __xsk_rcv(struct xdp_sock *xs, struct xdp_buff *xdp, u32 len)
142 {
143 	u64 offset = xs->umem->headroom;
144 	u64 addr, memcpy_addr;
145 	void *from_buf;
146 	u32 metalen;
147 	int err;
148 
149 	if (!xskq_peek_addr(xs->umem->fq, &addr, xs->umem) ||
150 	    len > xs->umem->chunk_size_nohr - XDP_PACKET_HEADROOM) {
151 		xs->rx_dropped++;
152 		return -ENOSPC;
153 	}
154 
155 	if (unlikely(xdp_data_meta_unsupported(xdp))) {
156 		from_buf = xdp->data;
157 		metalen = 0;
158 	} else {
159 		from_buf = xdp->data_meta;
160 		metalen = xdp->data - xdp->data_meta;
161 	}
162 
163 	memcpy_addr = xsk_umem_adjust_offset(xs->umem, addr, offset);
164 	__xsk_rcv_memcpy(xs->umem, memcpy_addr, from_buf, len, metalen);
165 
166 	offset += metalen;
167 	addr = xsk_umem_adjust_offset(xs->umem, addr, offset);
168 	err = xskq_produce_batch_desc(xs->rx, addr, len);
169 	if (!err) {
170 		xskq_discard_addr(xs->umem->fq);
171 		xdp_return_buff(xdp);
172 		return 0;
173 	}
174 
175 	xs->rx_dropped++;
176 	return err;
177 }
178 
179 static int __xsk_rcv_zc(struct xdp_sock *xs, struct xdp_buff *xdp, u32 len)
180 {
181 	int err = xskq_produce_batch_desc(xs->rx, (u64)xdp->handle, len);
182 
183 	if (err)
184 		xs->rx_dropped++;
185 
186 	return err;
187 }
188 
189 static bool xsk_is_bound(struct xdp_sock *xs)
190 {
191 	if (READ_ONCE(xs->state) == XSK_BOUND) {
192 		/* Matches smp_wmb() in bind(). */
193 		smp_rmb();
194 		return true;
195 	}
196 	return false;
197 }
198 
199 int xsk_rcv(struct xdp_sock *xs, struct xdp_buff *xdp)
200 {
201 	u32 len;
202 
203 	if (!xsk_is_bound(xs))
204 		return -EINVAL;
205 
206 	if (xs->dev != xdp->rxq->dev || xs->queue_id != xdp->rxq->queue_index)
207 		return -EINVAL;
208 
209 	len = xdp->data_end - xdp->data;
210 
211 	return (xdp->rxq->mem.type == MEM_TYPE_ZERO_COPY) ?
212 		__xsk_rcv_zc(xs, xdp, len) : __xsk_rcv(xs, xdp, len);
213 }
214 
215 void xsk_flush(struct xdp_sock *xs)
216 {
217 	xskq_produce_flush_desc(xs->rx);
218 	xs->sk.sk_data_ready(&xs->sk);
219 }
220 
221 int xsk_generic_rcv(struct xdp_sock *xs, struct xdp_buff *xdp)
222 {
223 	u32 metalen = xdp->data - xdp->data_meta;
224 	u32 len = xdp->data_end - xdp->data;
225 	u64 offset = xs->umem->headroom;
226 	void *buffer;
227 	u64 addr;
228 	int err;
229 
230 	spin_lock_bh(&xs->rx_lock);
231 
232 	if (xs->dev != xdp->rxq->dev || xs->queue_id != xdp->rxq->queue_index) {
233 		err = -EINVAL;
234 		goto out_unlock;
235 	}
236 
237 	if (!xskq_peek_addr(xs->umem->fq, &addr, xs->umem) ||
238 	    len > xs->umem->chunk_size_nohr - XDP_PACKET_HEADROOM) {
239 		err = -ENOSPC;
240 		goto out_drop;
241 	}
242 
243 	addr = xsk_umem_adjust_offset(xs->umem, addr, offset);
244 	buffer = xdp_umem_get_data(xs->umem, addr);
245 	memcpy(buffer, xdp->data_meta, len + metalen);
246 
247 	addr = xsk_umem_adjust_offset(xs->umem, addr, metalen);
248 	err = xskq_produce_batch_desc(xs->rx, addr, len);
249 	if (err)
250 		goto out_drop;
251 
252 	xskq_discard_addr(xs->umem->fq);
253 	xskq_produce_flush_desc(xs->rx);
254 
255 	spin_unlock_bh(&xs->rx_lock);
256 
257 	xs->sk.sk_data_ready(&xs->sk);
258 	return 0;
259 
260 out_drop:
261 	xs->rx_dropped++;
262 out_unlock:
263 	spin_unlock_bh(&xs->rx_lock);
264 	return err;
265 }
266 
267 void xsk_umem_complete_tx(struct xdp_umem *umem, u32 nb_entries)
268 {
269 	xskq_produce_flush_addr_n(umem->cq, nb_entries);
270 }
271 EXPORT_SYMBOL(xsk_umem_complete_tx);
272 
273 void xsk_umem_consume_tx_done(struct xdp_umem *umem)
274 {
275 	struct xdp_sock *xs;
276 
277 	rcu_read_lock();
278 	list_for_each_entry_rcu(xs, &umem->xsk_list, list) {
279 		xs->sk.sk_write_space(&xs->sk);
280 	}
281 	rcu_read_unlock();
282 }
283 EXPORT_SYMBOL(xsk_umem_consume_tx_done);
284 
285 bool xsk_umem_consume_tx(struct xdp_umem *umem, struct xdp_desc *desc)
286 {
287 	struct xdp_sock *xs;
288 
289 	rcu_read_lock();
290 	list_for_each_entry_rcu(xs, &umem->xsk_list, list) {
291 		if (!xskq_peek_desc(xs->tx, desc, umem))
292 			continue;
293 
294 		if (xskq_produce_addr_lazy(umem->cq, desc->addr))
295 			goto out;
296 
297 		xskq_discard_desc(xs->tx);
298 		rcu_read_unlock();
299 		return true;
300 	}
301 
302 out:
303 	rcu_read_unlock();
304 	return false;
305 }
306 EXPORT_SYMBOL(xsk_umem_consume_tx);
307 
308 static int xsk_zc_xmit(struct xdp_sock *xs)
309 {
310 	struct net_device *dev = xs->dev;
311 
312 	return dev->netdev_ops->ndo_xsk_wakeup(dev, xs->queue_id,
313 					       XDP_WAKEUP_TX);
314 }
315 
316 static void xsk_destruct_skb(struct sk_buff *skb)
317 {
318 	u64 addr = (u64)(long)skb_shinfo(skb)->destructor_arg;
319 	struct xdp_sock *xs = xdp_sk(skb->sk);
320 	unsigned long flags;
321 
322 	spin_lock_irqsave(&xs->tx_completion_lock, flags);
323 	WARN_ON_ONCE(xskq_produce_addr(xs->umem->cq, addr));
324 	spin_unlock_irqrestore(&xs->tx_completion_lock, flags);
325 
326 	sock_wfree(skb);
327 }
328 
329 static int xsk_generic_xmit(struct sock *sk)
330 {
331 	struct xdp_sock *xs = xdp_sk(sk);
332 	u32 max_batch = TX_BATCH_SIZE;
333 	bool sent_frame = false;
334 	struct xdp_desc desc;
335 	struct sk_buff *skb;
336 	int err = 0;
337 
338 	mutex_lock(&xs->mutex);
339 
340 	if (xs->queue_id >= xs->dev->real_num_tx_queues)
341 		goto out;
342 
343 	while (xskq_peek_desc(xs->tx, &desc, xs->umem)) {
344 		char *buffer;
345 		u64 addr;
346 		u32 len;
347 
348 		if (max_batch-- == 0) {
349 			err = -EAGAIN;
350 			goto out;
351 		}
352 
353 		len = desc.len;
354 		skb = sock_alloc_send_skb(sk, len, 1, &err);
355 		if (unlikely(!skb)) {
356 			err = -EAGAIN;
357 			goto out;
358 		}
359 
360 		skb_put(skb, len);
361 		addr = desc.addr;
362 		buffer = xdp_umem_get_data(xs->umem, addr);
363 		err = skb_store_bits(skb, 0, buffer, len);
364 		if (unlikely(err) || xskq_reserve_addr(xs->umem->cq)) {
365 			kfree_skb(skb);
366 			goto out;
367 		}
368 
369 		skb->dev = xs->dev;
370 		skb->priority = sk->sk_priority;
371 		skb->mark = sk->sk_mark;
372 		skb_shinfo(skb)->destructor_arg = (void *)(long)desc.addr;
373 		skb->destructor = xsk_destruct_skb;
374 
375 		err = dev_direct_xmit(skb, xs->queue_id);
376 		xskq_discard_desc(xs->tx);
377 		/* Ignore NET_XMIT_CN as packet might have been sent */
378 		if (err == NET_XMIT_DROP || err == NETDEV_TX_BUSY) {
379 			/* SKB completed but not sent */
380 			err = -EBUSY;
381 			goto out;
382 		}
383 
384 		sent_frame = true;
385 	}
386 
387 out:
388 	if (sent_frame)
389 		sk->sk_write_space(sk);
390 
391 	mutex_unlock(&xs->mutex);
392 	return err;
393 }
394 
395 static int __xsk_sendmsg(struct sock *sk)
396 {
397 	struct xdp_sock *xs = xdp_sk(sk);
398 
399 	if (unlikely(!(xs->dev->flags & IFF_UP)))
400 		return -ENETDOWN;
401 	if (unlikely(!xs->tx))
402 		return -ENOBUFS;
403 
404 	return xs->zc ? xsk_zc_xmit(xs) : xsk_generic_xmit(sk);
405 }
406 
407 static int xsk_sendmsg(struct socket *sock, struct msghdr *m, size_t total_len)
408 {
409 	bool need_wait = !(m->msg_flags & MSG_DONTWAIT);
410 	struct sock *sk = sock->sk;
411 	struct xdp_sock *xs = xdp_sk(sk);
412 
413 	if (unlikely(!xsk_is_bound(xs)))
414 		return -ENXIO;
415 	if (unlikely(need_wait))
416 		return -EOPNOTSUPP;
417 
418 	return __xsk_sendmsg(sk);
419 }
420 
421 static unsigned int xsk_poll(struct file *file, struct socket *sock,
422 			     struct poll_table_struct *wait)
423 {
424 	unsigned int mask = datagram_poll(file, sock, wait);
425 	struct sock *sk = sock->sk;
426 	struct xdp_sock *xs = xdp_sk(sk);
427 	struct net_device *dev;
428 	struct xdp_umem *umem;
429 
430 	if (unlikely(!xsk_is_bound(xs)))
431 		return mask;
432 
433 	dev = xs->dev;
434 	umem = xs->umem;
435 
436 	if (umem->need_wakeup) {
437 		if (dev->netdev_ops->ndo_xsk_wakeup)
438 			dev->netdev_ops->ndo_xsk_wakeup(dev, xs->queue_id,
439 							umem->need_wakeup);
440 		else
441 			/* Poll needs to drive Tx also in copy mode */
442 			__xsk_sendmsg(sk);
443 	}
444 
445 	if (xs->rx && !xskq_empty_desc(xs->rx))
446 		mask |= POLLIN | POLLRDNORM;
447 	if (xs->tx && !xskq_full_desc(xs->tx))
448 		mask |= POLLOUT | POLLWRNORM;
449 
450 	return mask;
451 }
452 
453 static int xsk_init_queue(u32 entries, struct xsk_queue **queue,
454 			  bool umem_queue)
455 {
456 	struct xsk_queue *q;
457 
458 	if (entries == 0 || *queue || !is_power_of_2(entries))
459 		return -EINVAL;
460 
461 	q = xskq_create(entries, umem_queue);
462 	if (!q)
463 		return -ENOMEM;
464 
465 	/* Make sure queue is ready before it can be seen by others */
466 	smp_wmb();
467 	WRITE_ONCE(*queue, q);
468 	return 0;
469 }
470 
471 static void xsk_unbind_dev(struct xdp_sock *xs)
472 {
473 	struct net_device *dev = xs->dev;
474 
475 	if (xs->state != XSK_BOUND)
476 		return;
477 	WRITE_ONCE(xs->state, XSK_UNBOUND);
478 
479 	/* Wait for driver to stop using the xdp socket. */
480 	xdp_del_sk_umem(xs->umem, xs);
481 	xs->dev = NULL;
482 	synchronize_net();
483 	dev_put(dev);
484 }
485 
486 static struct xsk_map *xsk_get_map_list_entry(struct xdp_sock *xs,
487 					      struct xdp_sock ***map_entry)
488 {
489 	struct xsk_map *map = NULL;
490 	struct xsk_map_node *node;
491 
492 	*map_entry = NULL;
493 
494 	spin_lock_bh(&xs->map_list_lock);
495 	node = list_first_entry_or_null(&xs->map_list, struct xsk_map_node,
496 					node);
497 	if (node) {
498 		WARN_ON(xsk_map_inc(node->map));
499 		map = node->map;
500 		*map_entry = node->map_entry;
501 	}
502 	spin_unlock_bh(&xs->map_list_lock);
503 	return map;
504 }
505 
506 static void xsk_delete_from_maps(struct xdp_sock *xs)
507 {
508 	/* This function removes the current XDP socket from all the
509 	 * maps it resides in. We need to take extra care here, due to
510 	 * the two locks involved. Each map has a lock synchronizing
511 	 * updates to the entries, and each socket has a lock that
512 	 * synchronizes access to the list of maps (map_list). For
513 	 * deadlock avoidance the locks need to be taken in the order
514 	 * "map lock"->"socket map list lock". We start off by
515 	 * accessing the socket map list, and take a reference to the
516 	 * map to guarantee existence between the
517 	 * xsk_get_map_list_entry() and xsk_map_try_sock_delete()
518 	 * calls. Then we ask the map to remove the socket, which
519 	 * tries to remove the socket from the map. Note that there
520 	 * might be updates to the map between
521 	 * xsk_get_map_list_entry() and xsk_map_try_sock_delete().
522 	 */
523 	struct xdp_sock **map_entry = NULL;
524 	struct xsk_map *map;
525 
526 	while ((map = xsk_get_map_list_entry(xs, &map_entry))) {
527 		xsk_map_try_sock_delete(map, xs, map_entry);
528 		xsk_map_put(map);
529 	}
530 }
531 
532 static int xsk_release(struct socket *sock)
533 {
534 	struct sock *sk = sock->sk;
535 	struct xdp_sock *xs = xdp_sk(sk);
536 	struct net *net;
537 
538 	if (!sk)
539 		return 0;
540 
541 	net = sock_net(sk);
542 
543 	mutex_lock(&net->xdp.lock);
544 	sk_del_node_init_rcu(sk);
545 	mutex_unlock(&net->xdp.lock);
546 
547 	local_bh_disable();
548 	sock_prot_inuse_add(net, sk->sk_prot, -1);
549 	local_bh_enable();
550 
551 	xsk_delete_from_maps(xs);
552 	mutex_lock(&xs->mutex);
553 	xsk_unbind_dev(xs);
554 	mutex_unlock(&xs->mutex);
555 
556 	xskq_destroy(xs->rx);
557 	xskq_destroy(xs->tx);
558 
559 	sock_orphan(sk);
560 	sock->sk = NULL;
561 
562 	sk_refcnt_debug_release(sk);
563 	sock_put(sk);
564 
565 	return 0;
566 }
567 
568 static struct socket *xsk_lookup_xsk_from_fd(int fd)
569 {
570 	struct socket *sock;
571 	int err;
572 
573 	sock = sockfd_lookup(fd, &err);
574 	if (!sock)
575 		return ERR_PTR(-ENOTSOCK);
576 
577 	if (sock->sk->sk_family != PF_XDP) {
578 		sockfd_put(sock);
579 		return ERR_PTR(-ENOPROTOOPT);
580 	}
581 
582 	return sock;
583 }
584 
585 /* Check if umem pages are contiguous.
586  * If zero-copy mode, use the DMA address to do the page contiguity check
587  * For all other modes we use addr (kernel virtual address)
588  * Store the result in the low bits of addr.
589  */
590 static void xsk_check_page_contiguity(struct xdp_umem *umem, u32 flags)
591 {
592 	struct xdp_umem_page *pgs = umem->pages;
593 	int i, is_contig;
594 
595 	for (i = 0; i < umem->npgs - 1; i++) {
596 		is_contig = (flags & XDP_ZEROCOPY) ?
597 			(pgs[i].dma + PAGE_SIZE == pgs[i + 1].dma) :
598 			(pgs[i].addr + PAGE_SIZE == pgs[i + 1].addr);
599 		pgs[i].addr += is_contig << XSK_NEXT_PG_CONTIG_SHIFT;
600 	}
601 }
602 
603 static int xsk_bind(struct socket *sock, struct sockaddr *addr, int addr_len)
604 {
605 	struct sockaddr_xdp *sxdp = (struct sockaddr_xdp *)addr;
606 	struct sock *sk = sock->sk;
607 	struct xdp_sock *xs = xdp_sk(sk);
608 	struct net_device *dev;
609 	u32 flags, qid;
610 	int err = 0;
611 
612 	if (addr_len < sizeof(struct sockaddr_xdp))
613 		return -EINVAL;
614 	if (sxdp->sxdp_family != AF_XDP)
615 		return -EINVAL;
616 
617 	flags = sxdp->sxdp_flags;
618 	if (flags & ~(XDP_SHARED_UMEM | XDP_COPY | XDP_ZEROCOPY |
619 		      XDP_USE_NEED_WAKEUP))
620 		return -EINVAL;
621 
622 	rtnl_lock();
623 	mutex_lock(&xs->mutex);
624 	if (xs->state != XSK_READY) {
625 		err = -EBUSY;
626 		goto out_release;
627 	}
628 
629 	dev = dev_get_by_index(sock_net(sk), sxdp->sxdp_ifindex);
630 	if (!dev) {
631 		err = -ENODEV;
632 		goto out_release;
633 	}
634 
635 	if (!xs->rx && !xs->tx) {
636 		err = -EINVAL;
637 		goto out_unlock;
638 	}
639 
640 	qid = sxdp->sxdp_queue_id;
641 
642 	if (flags & XDP_SHARED_UMEM) {
643 		struct xdp_sock *umem_xs;
644 		struct socket *sock;
645 
646 		if ((flags & XDP_COPY) || (flags & XDP_ZEROCOPY) ||
647 		    (flags & XDP_USE_NEED_WAKEUP)) {
648 			/* Cannot specify flags for shared sockets. */
649 			err = -EINVAL;
650 			goto out_unlock;
651 		}
652 
653 		if (xs->umem) {
654 			/* We have already our own. */
655 			err = -EINVAL;
656 			goto out_unlock;
657 		}
658 
659 		sock = xsk_lookup_xsk_from_fd(sxdp->sxdp_shared_umem_fd);
660 		if (IS_ERR(sock)) {
661 			err = PTR_ERR(sock);
662 			goto out_unlock;
663 		}
664 
665 		umem_xs = xdp_sk(sock->sk);
666 		if (!xsk_is_bound(umem_xs)) {
667 			err = -EBADF;
668 			sockfd_put(sock);
669 			goto out_unlock;
670 		}
671 		if (umem_xs->dev != dev || umem_xs->queue_id != qid) {
672 			err = -EINVAL;
673 			sockfd_put(sock);
674 			goto out_unlock;
675 		}
676 
677 		xdp_get_umem(umem_xs->umem);
678 		WRITE_ONCE(xs->umem, umem_xs->umem);
679 		sockfd_put(sock);
680 	} else if (!xs->umem || !xdp_umem_validate_queues(xs->umem)) {
681 		err = -EINVAL;
682 		goto out_unlock;
683 	} else {
684 		/* This xsk has its own umem. */
685 		xskq_set_umem(xs->umem->fq, xs->umem->size,
686 			      xs->umem->chunk_mask);
687 		xskq_set_umem(xs->umem->cq, xs->umem->size,
688 			      xs->umem->chunk_mask);
689 
690 		err = xdp_umem_assign_dev(xs->umem, dev, qid, flags);
691 		if (err)
692 			goto out_unlock;
693 
694 		xsk_check_page_contiguity(xs->umem, flags);
695 	}
696 
697 	xs->dev = dev;
698 	xs->zc = xs->umem->zc;
699 	xs->queue_id = qid;
700 	xskq_set_umem(xs->rx, xs->umem->size, xs->umem->chunk_mask);
701 	xskq_set_umem(xs->tx, xs->umem->size, xs->umem->chunk_mask);
702 	xdp_add_sk_umem(xs->umem, xs);
703 
704 out_unlock:
705 	if (err) {
706 		dev_put(dev);
707 	} else {
708 		/* Matches smp_rmb() in bind() for shared umem
709 		 * sockets, and xsk_is_bound().
710 		 */
711 		smp_wmb();
712 		WRITE_ONCE(xs->state, XSK_BOUND);
713 	}
714 out_release:
715 	mutex_unlock(&xs->mutex);
716 	rtnl_unlock();
717 	return err;
718 }
719 
720 struct xdp_umem_reg_v1 {
721 	__u64 addr; /* Start of packet data area */
722 	__u64 len; /* Length of packet data area */
723 	__u32 chunk_size;
724 	__u32 headroom;
725 };
726 
727 static int xsk_setsockopt(struct socket *sock, int level, int optname,
728 			  char __user *optval, unsigned int optlen)
729 {
730 	struct sock *sk = sock->sk;
731 	struct xdp_sock *xs = xdp_sk(sk);
732 	int err;
733 
734 	if (level != SOL_XDP)
735 		return -ENOPROTOOPT;
736 
737 	switch (optname) {
738 	case XDP_RX_RING:
739 	case XDP_TX_RING:
740 	{
741 		struct xsk_queue **q;
742 		int entries;
743 
744 		if (optlen < sizeof(entries))
745 			return -EINVAL;
746 		if (copy_from_user(&entries, optval, sizeof(entries)))
747 			return -EFAULT;
748 
749 		mutex_lock(&xs->mutex);
750 		if (xs->state != XSK_READY) {
751 			mutex_unlock(&xs->mutex);
752 			return -EBUSY;
753 		}
754 		q = (optname == XDP_TX_RING) ? &xs->tx : &xs->rx;
755 		err = xsk_init_queue(entries, q, false);
756 		if (!err && optname == XDP_TX_RING)
757 			/* Tx needs to be explicitly woken up the first time */
758 			xs->tx->ring->flags |= XDP_RING_NEED_WAKEUP;
759 		mutex_unlock(&xs->mutex);
760 		return err;
761 	}
762 	case XDP_UMEM_REG:
763 	{
764 		size_t mr_size = sizeof(struct xdp_umem_reg);
765 		struct xdp_umem_reg mr = {};
766 		struct xdp_umem *umem;
767 
768 		if (optlen < sizeof(struct xdp_umem_reg_v1))
769 			return -EINVAL;
770 		else if (optlen < sizeof(mr))
771 			mr_size = sizeof(struct xdp_umem_reg_v1);
772 
773 		if (copy_from_user(&mr, optval, mr_size))
774 			return -EFAULT;
775 
776 		mutex_lock(&xs->mutex);
777 		if (xs->state != XSK_READY || xs->umem) {
778 			mutex_unlock(&xs->mutex);
779 			return -EBUSY;
780 		}
781 
782 		umem = xdp_umem_create(&mr);
783 		if (IS_ERR(umem)) {
784 			mutex_unlock(&xs->mutex);
785 			return PTR_ERR(umem);
786 		}
787 
788 		/* Make sure umem is ready before it can be seen by others */
789 		smp_wmb();
790 		WRITE_ONCE(xs->umem, umem);
791 		mutex_unlock(&xs->mutex);
792 		return 0;
793 	}
794 	case XDP_UMEM_FILL_RING:
795 	case XDP_UMEM_COMPLETION_RING:
796 	{
797 		struct xsk_queue **q;
798 		int entries;
799 
800 		if (copy_from_user(&entries, optval, sizeof(entries)))
801 			return -EFAULT;
802 
803 		mutex_lock(&xs->mutex);
804 		if (xs->state != XSK_READY) {
805 			mutex_unlock(&xs->mutex);
806 			return -EBUSY;
807 		}
808 		if (!xs->umem) {
809 			mutex_unlock(&xs->mutex);
810 			return -EINVAL;
811 		}
812 
813 		q = (optname == XDP_UMEM_FILL_RING) ? &xs->umem->fq :
814 			&xs->umem->cq;
815 		err = xsk_init_queue(entries, q, true);
816 		mutex_unlock(&xs->mutex);
817 		return err;
818 	}
819 	default:
820 		break;
821 	}
822 
823 	return -ENOPROTOOPT;
824 }
825 
826 static void xsk_enter_rxtx_offsets(struct xdp_ring_offset_v1 *ring)
827 {
828 	ring->producer = offsetof(struct xdp_rxtx_ring, ptrs.producer);
829 	ring->consumer = offsetof(struct xdp_rxtx_ring, ptrs.consumer);
830 	ring->desc = offsetof(struct xdp_rxtx_ring, desc);
831 }
832 
833 static void xsk_enter_umem_offsets(struct xdp_ring_offset_v1 *ring)
834 {
835 	ring->producer = offsetof(struct xdp_umem_ring, ptrs.producer);
836 	ring->consumer = offsetof(struct xdp_umem_ring, ptrs.consumer);
837 	ring->desc = offsetof(struct xdp_umem_ring, desc);
838 }
839 
840 static int xsk_getsockopt(struct socket *sock, int level, int optname,
841 			  char __user *optval, int __user *optlen)
842 {
843 	struct sock *sk = sock->sk;
844 	struct xdp_sock *xs = xdp_sk(sk);
845 	int len;
846 
847 	if (level != SOL_XDP)
848 		return -ENOPROTOOPT;
849 
850 	if (get_user(len, optlen))
851 		return -EFAULT;
852 	if (len < 0)
853 		return -EINVAL;
854 
855 	switch (optname) {
856 	case XDP_STATISTICS:
857 	{
858 		struct xdp_statistics stats;
859 
860 		if (len < sizeof(stats))
861 			return -EINVAL;
862 
863 		mutex_lock(&xs->mutex);
864 		stats.rx_dropped = xs->rx_dropped;
865 		stats.rx_invalid_descs = xskq_nb_invalid_descs(xs->rx);
866 		stats.tx_invalid_descs = xskq_nb_invalid_descs(xs->tx);
867 		mutex_unlock(&xs->mutex);
868 
869 		if (copy_to_user(optval, &stats, sizeof(stats)))
870 			return -EFAULT;
871 		if (put_user(sizeof(stats), optlen))
872 			return -EFAULT;
873 
874 		return 0;
875 	}
876 	case XDP_MMAP_OFFSETS:
877 	{
878 		struct xdp_mmap_offsets off;
879 		struct xdp_mmap_offsets_v1 off_v1;
880 		bool flags_supported = true;
881 		void *to_copy;
882 
883 		if (len < sizeof(off_v1))
884 			return -EINVAL;
885 		else if (len < sizeof(off))
886 			flags_supported = false;
887 
888 		if (flags_supported) {
889 			/* xdp_ring_offset is identical to xdp_ring_offset_v1
890 			 * except for the flags field added to the end.
891 			 */
892 			xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *)
893 					       &off.rx);
894 			xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *)
895 					       &off.tx);
896 			xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *)
897 					       &off.fr);
898 			xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *)
899 					       &off.cr);
900 			off.rx.flags = offsetof(struct xdp_rxtx_ring,
901 						ptrs.flags);
902 			off.tx.flags = offsetof(struct xdp_rxtx_ring,
903 						ptrs.flags);
904 			off.fr.flags = offsetof(struct xdp_umem_ring,
905 						ptrs.flags);
906 			off.cr.flags = offsetof(struct xdp_umem_ring,
907 						ptrs.flags);
908 
909 			len = sizeof(off);
910 			to_copy = &off;
911 		} else {
912 			xsk_enter_rxtx_offsets(&off_v1.rx);
913 			xsk_enter_rxtx_offsets(&off_v1.tx);
914 			xsk_enter_umem_offsets(&off_v1.fr);
915 			xsk_enter_umem_offsets(&off_v1.cr);
916 
917 			len = sizeof(off_v1);
918 			to_copy = &off_v1;
919 		}
920 
921 		if (copy_to_user(optval, to_copy, len))
922 			return -EFAULT;
923 		if (put_user(len, optlen))
924 			return -EFAULT;
925 
926 		return 0;
927 	}
928 	case XDP_OPTIONS:
929 	{
930 		struct xdp_options opts = {};
931 
932 		if (len < sizeof(opts))
933 			return -EINVAL;
934 
935 		mutex_lock(&xs->mutex);
936 		if (xs->zc)
937 			opts.flags |= XDP_OPTIONS_ZEROCOPY;
938 		mutex_unlock(&xs->mutex);
939 
940 		len = sizeof(opts);
941 		if (copy_to_user(optval, &opts, len))
942 			return -EFAULT;
943 		if (put_user(len, optlen))
944 			return -EFAULT;
945 
946 		return 0;
947 	}
948 	default:
949 		break;
950 	}
951 
952 	return -EOPNOTSUPP;
953 }
954 
955 static int xsk_mmap(struct file *file, struct socket *sock,
956 		    struct vm_area_struct *vma)
957 {
958 	loff_t offset = (loff_t)vma->vm_pgoff << PAGE_SHIFT;
959 	unsigned long size = vma->vm_end - vma->vm_start;
960 	struct xdp_sock *xs = xdp_sk(sock->sk);
961 	struct xsk_queue *q = NULL;
962 	struct xdp_umem *umem;
963 	unsigned long pfn;
964 	struct page *qpg;
965 
966 	if (READ_ONCE(xs->state) != XSK_READY)
967 		return -EBUSY;
968 
969 	if (offset == XDP_PGOFF_RX_RING) {
970 		q = READ_ONCE(xs->rx);
971 	} else if (offset == XDP_PGOFF_TX_RING) {
972 		q = READ_ONCE(xs->tx);
973 	} else {
974 		umem = READ_ONCE(xs->umem);
975 		if (!umem)
976 			return -EINVAL;
977 
978 		/* Matches the smp_wmb() in XDP_UMEM_REG */
979 		smp_rmb();
980 		if (offset == XDP_UMEM_PGOFF_FILL_RING)
981 			q = READ_ONCE(umem->fq);
982 		else if (offset == XDP_UMEM_PGOFF_COMPLETION_RING)
983 			q = READ_ONCE(umem->cq);
984 	}
985 
986 	if (!q)
987 		return -EINVAL;
988 
989 	/* Matches the smp_wmb() in xsk_init_queue */
990 	smp_rmb();
991 	qpg = virt_to_head_page(q->ring);
992 	if (size > page_size(qpg))
993 		return -EINVAL;
994 
995 	pfn = virt_to_phys(q->ring) >> PAGE_SHIFT;
996 	return remap_pfn_range(vma, vma->vm_start, pfn,
997 			       size, vma->vm_page_prot);
998 }
999 
1000 static int xsk_notifier(struct notifier_block *this,
1001 			unsigned long msg, void *ptr)
1002 {
1003 	struct net_device *dev = netdev_notifier_info_to_dev(ptr);
1004 	struct net *net = dev_net(dev);
1005 	struct sock *sk;
1006 
1007 	switch (msg) {
1008 	case NETDEV_UNREGISTER:
1009 		mutex_lock(&net->xdp.lock);
1010 		sk_for_each(sk, &net->xdp.list) {
1011 			struct xdp_sock *xs = xdp_sk(sk);
1012 
1013 			mutex_lock(&xs->mutex);
1014 			if (xs->dev == dev) {
1015 				sk->sk_err = ENETDOWN;
1016 				if (!sock_flag(sk, SOCK_DEAD))
1017 					sk->sk_error_report(sk);
1018 
1019 				xsk_unbind_dev(xs);
1020 
1021 				/* Clear device references in umem. */
1022 				xdp_umem_clear_dev(xs->umem);
1023 			}
1024 			mutex_unlock(&xs->mutex);
1025 		}
1026 		mutex_unlock(&net->xdp.lock);
1027 		break;
1028 	}
1029 	return NOTIFY_DONE;
1030 }
1031 
1032 static struct proto xsk_proto = {
1033 	.name =		"XDP",
1034 	.owner =	THIS_MODULE,
1035 	.obj_size =	sizeof(struct xdp_sock),
1036 };
1037 
1038 static const struct proto_ops xsk_proto_ops = {
1039 	.family		= PF_XDP,
1040 	.owner		= THIS_MODULE,
1041 	.release	= xsk_release,
1042 	.bind		= xsk_bind,
1043 	.connect	= sock_no_connect,
1044 	.socketpair	= sock_no_socketpair,
1045 	.accept		= sock_no_accept,
1046 	.getname	= sock_no_getname,
1047 	.poll		= xsk_poll,
1048 	.ioctl		= sock_no_ioctl,
1049 	.listen		= sock_no_listen,
1050 	.shutdown	= sock_no_shutdown,
1051 	.setsockopt	= xsk_setsockopt,
1052 	.getsockopt	= xsk_getsockopt,
1053 	.sendmsg	= xsk_sendmsg,
1054 	.recvmsg	= sock_no_recvmsg,
1055 	.mmap		= xsk_mmap,
1056 	.sendpage	= sock_no_sendpage,
1057 };
1058 
1059 static void xsk_destruct(struct sock *sk)
1060 {
1061 	struct xdp_sock *xs = xdp_sk(sk);
1062 
1063 	if (!sock_flag(sk, SOCK_DEAD))
1064 		return;
1065 
1066 	xdp_put_umem(xs->umem);
1067 
1068 	sk_refcnt_debug_dec(sk);
1069 }
1070 
1071 static int xsk_create(struct net *net, struct socket *sock, int protocol,
1072 		      int kern)
1073 {
1074 	struct sock *sk;
1075 	struct xdp_sock *xs;
1076 
1077 	if (!ns_capable(net->user_ns, CAP_NET_RAW))
1078 		return -EPERM;
1079 	if (sock->type != SOCK_RAW)
1080 		return -ESOCKTNOSUPPORT;
1081 
1082 	if (protocol)
1083 		return -EPROTONOSUPPORT;
1084 
1085 	sock->state = SS_UNCONNECTED;
1086 
1087 	sk = sk_alloc(net, PF_XDP, GFP_KERNEL, &xsk_proto, kern);
1088 	if (!sk)
1089 		return -ENOBUFS;
1090 
1091 	sock->ops = &xsk_proto_ops;
1092 
1093 	sock_init_data(sock, sk);
1094 
1095 	sk->sk_family = PF_XDP;
1096 
1097 	sk->sk_destruct = xsk_destruct;
1098 	sk_refcnt_debug_inc(sk);
1099 
1100 	sock_set_flag(sk, SOCK_RCU_FREE);
1101 
1102 	xs = xdp_sk(sk);
1103 	xs->state = XSK_READY;
1104 	mutex_init(&xs->mutex);
1105 	spin_lock_init(&xs->rx_lock);
1106 	spin_lock_init(&xs->tx_completion_lock);
1107 
1108 	INIT_LIST_HEAD(&xs->map_list);
1109 	spin_lock_init(&xs->map_list_lock);
1110 
1111 	mutex_lock(&net->xdp.lock);
1112 	sk_add_node_rcu(sk, &net->xdp.list);
1113 	mutex_unlock(&net->xdp.lock);
1114 
1115 	local_bh_disable();
1116 	sock_prot_inuse_add(net, &xsk_proto, 1);
1117 	local_bh_enable();
1118 
1119 	return 0;
1120 }
1121 
1122 static const struct net_proto_family xsk_family_ops = {
1123 	.family = PF_XDP,
1124 	.create = xsk_create,
1125 	.owner	= THIS_MODULE,
1126 };
1127 
1128 static struct notifier_block xsk_netdev_notifier = {
1129 	.notifier_call	= xsk_notifier,
1130 };
1131 
1132 static int __net_init xsk_net_init(struct net *net)
1133 {
1134 	mutex_init(&net->xdp.lock);
1135 	INIT_HLIST_HEAD(&net->xdp.list);
1136 	return 0;
1137 }
1138 
1139 static void __net_exit xsk_net_exit(struct net *net)
1140 {
1141 	WARN_ON_ONCE(!hlist_empty(&net->xdp.list));
1142 }
1143 
1144 static struct pernet_operations xsk_net_ops = {
1145 	.init = xsk_net_init,
1146 	.exit = xsk_net_exit,
1147 };
1148 
1149 static int __init xsk_init(void)
1150 {
1151 	int err;
1152 
1153 	err = proto_register(&xsk_proto, 0 /* no slab */);
1154 	if (err)
1155 		goto out;
1156 
1157 	err = sock_register(&xsk_family_ops);
1158 	if (err)
1159 		goto out_proto;
1160 
1161 	err = register_pernet_subsys(&xsk_net_ops);
1162 	if (err)
1163 		goto out_sk;
1164 
1165 	err = register_netdevice_notifier(&xsk_netdev_notifier);
1166 	if (err)
1167 		goto out_pernet;
1168 
1169 	return 0;
1170 
1171 out_pernet:
1172 	unregister_pernet_subsys(&xsk_net_ops);
1173 out_sk:
1174 	sock_unregister(PF_XDP);
1175 out_proto:
1176 	proto_unregister(&xsk_proto);
1177 out:
1178 	return err;
1179 }
1180 
1181 fs_initcall(xsk_init);
1182