xref: /openbmc/linux/net/xdp/xsk.c (revision 78beef62)
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 sock *sk)
309 {
310 	struct xdp_sock *xs = xdp_sk(sk);
311 	struct net_device *dev = xs->dev;
312 
313 	return dev->netdev_ops->ndo_xsk_wakeup(dev, xs->queue_id,
314 					       XDP_WAKEUP_TX);
315 }
316 
317 static void xsk_destruct_skb(struct sk_buff *skb)
318 {
319 	u64 addr = (u64)(long)skb_shinfo(skb)->destructor_arg;
320 	struct xdp_sock *xs = xdp_sk(skb->sk);
321 	unsigned long flags;
322 
323 	spin_lock_irqsave(&xs->tx_completion_lock, flags);
324 	WARN_ON_ONCE(xskq_produce_addr(xs->umem->cq, addr));
325 	spin_unlock_irqrestore(&xs->tx_completion_lock, flags);
326 
327 	sock_wfree(skb);
328 }
329 
330 static int xsk_generic_xmit(struct sock *sk, struct msghdr *m,
331 			    size_t total_len)
332 {
333 	u32 max_batch = TX_BATCH_SIZE;
334 	struct xdp_sock *xs = xdp_sk(sk);
335 	bool sent_frame = false;
336 	struct xdp_desc desc;
337 	struct sk_buff *skb;
338 	int err = 0;
339 
340 	mutex_lock(&xs->mutex);
341 
342 	if (xs->queue_id >= xs->dev->real_num_tx_queues)
343 		goto out;
344 
345 	while (xskq_peek_desc(xs->tx, &desc, xs->umem)) {
346 		char *buffer;
347 		u64 addr;
348 		u32 len;
349 
350 		if (max_batch-- == 0) {
351 			err = -EAGAIN;
352 			goto out;
353 		}
354 
355 		len = desc.len;
356 		skb = sock_alloc_send_skb(sk, len, 1, &err);
357 		if (unlikely(!skb)) {
358 			err = -EAGAIN;
359 			goto out;
360 		}
361 
362 		skb_put(skb, len);
363 		addr = desc.addr;
364 		buffer = xdp_umem_get_data(xs->umem, addr);
365 		err = skb_store_bits(skb, 0, buffer, len);
366 		if (unlikely(err) || xskq_reserve_addr(xs->umem->cq)) {
367 			kfree_skb(skb);
368 			goto out;
369 		}
370 
371 		skb->dev = xs->dev;
372 		skb->priority = sk->sk_priority;
373 		skb->mark = sk->sk_mark;
374 		skb_shinfo(skb)->destructor_arg = (void *)(long)desc.addr;
375 		skb->destructor = xsk_destruct_skb;
376 
377 		err = dev_direct_xmit(skb, xs->queue_id);
378 		xskq_discard_desc(xs->tx);
379 		/* Ignore NET_XMIT_CN as packet might have been sent */
380 		if (err == NET_XMIT_DROP || err == NETDEV_TX_BUSY) {
381 			/* SKB completed but not sent */
382 			err = -EBUSY;
383 			goto out;
384 		}
385 
386 		sent_frame = true;
387 	}
388 
389 out:
390 	if (sent_frame)
391 		sk->sk_write_space(sk);
392 
393 	mutex_unlock(&xs->mutex);
394 	return err;
395 }
396 
397 static int xsk_sendmsg(struct socket *sock, struct msghdr *m, size_t total_len)
398 {
399 	bool need_wait = !(m->msg_flags & MSG_DONTWAIT);
400 	struct sock *sk = sock->sk;
401 	struct xdp_sock *xs = xdp_sk(sk);
402 
403 	if (unlikely(!xsk_is_bound(xs)))
404 		return -ENXIO;
405 	if (unlikely(!(xs->dev->flags & IFF_UP)))
406 		return -ENETDOWN;
407 	if (unlikely(!xs->tx))
408 		return -ENOBUFS;
409 	if (need_wait)
410 		return -EOPNOTSUPP;
411 
412 	return (xs->zc) ? xsk_zc_xmit(sk) : xsk_generic_xmit(sk, m, total_len);
413 }
414 
415 static unsigned int xsk_poll(struct file *file, struct socket *sock,
416 			     struct poll_table_struct *wait)
417 {
418 	unsigned int mask = datagram_poll(file, sock, wait);
419 	struct xdp_sock *xs = xdp_sk(sock->sk);
420 	struct net_device *dev;
421 	struct xdp_umem *umem;
422 
423 	if (unlikely(!xsk_is_bound(xs)))
424 		return mask;
425 
426 	dev = xs->dev;
427 	umem = xs->umem;
428 
429 	if (umem->need_wakeup)
430 		dev->netdev_ops->ndo_xsk_wakeup(dev, xs->queue_id,
431 						umem->need_wakeup);
432 
433 	if (xs->rx && !xskq_empty_desc(xs->rx))
434 		mask |= POLLIN | POLLRDNORM;
435 	if (xs->tx && !xskq_full_desc(xs->tx))
436 		mask |= POLLOUT | POLLWRNORM;
437 
438 	return mask;
439 }
440 
441 static int xsk_init_queue(u32 entries, struct xsk_queue **queue,
442 			  bool umem_queue)
443 {
444 	struct xsk_queue *q;
445 
446 	if (entries == 0 || *queue || !is_power_of_2(entries))
447 		return -EINVAL;
448 
449 	q = xskq_create(entries, umem_queue);
450 	if (!q)
451 		return -ENOMEM;
452 
453 	/* Make sure queue is ready before it can be seen by others */
454 	smp_wmb();
455 	WRITE_ONCE(*queue, q);
456 	return 0;
457 }
458 
459 static void xsk_unbind_dev(struct xdp_sock *xs)
460 {
461 	struct net_device *dev = xs->dev;
462 
463 	if (xs->state != XSK_BOUND)
464 		return;
465 	WRITE_ONCE(xs->state, XSK_UNBOUND);
466 
467 	/* Wait for driver to stop using the xdp socket. */
468 	xdp_del_sk_umem(xs->umem, xs);
469 	xs->dev = NULL;
470 	synchronize_net();
471 	dev_put(dev);
472 }
473 
474 static struct xsk_map *xsk_get_map_list_entry(struct xdp_sock *xs,
475 					      struct xdp_sock ***map_entry)
476 {
477 	struct xsk_map *map = NULL;
478 	struct xsk_map_node *node;
479 
480 	*map_entry = NULL;
481 
482 	spin_lock_bh(&xs->map_list_lock);
483 	node = list_first_entry_or_null(&xs->map_list, struct xsk_map_node,
484 					node);
485 	if (node) {
486 		WARN_ON(xsk_map_inc(node->map));
487 		map = node->map;
488 		*map_entry = node->map_entry;
489 	}
490 	spin_unlock_bh(&xs->map_list_lock);
491 	return map;
492 }
493 
494 static void xsk_delete_from_maps(struct xdp_sock *xs)
495 {
496 	/* This function removes the current XDP socket from all the
497 	 * maps it resides in. We need to take extra care here, due to
498 	 * the two locks involved. Each map has a lock synchronizing
499 	 * updates to the entries, and each socket has a lock that
500 	 * synchronizes access to the list of maps (map_list). For
501 	 * deadlock avoidance the locks need to be taken in the order
502 	 * "map lock"->"socket map list lock". We start off by
503 	 * accessing the socket map list, and take a reference to the
504 	 * map to guarantee existence between the
505 	 * xsk_get_map_list_entry() and xsk_map_try_sock_delete()
506 	 * calls. Then we ask the map to remove the socket, which
507 	 * tries to remove the socket from the map. Note that there
508 	 * might be updates to the map between
509 	 * xsk_get_map_list_entry() and xsk_map_try_sock_delete().
510 	 */
511 	struct xdp_sock **map_entry = NULL;
512 	struct xsk_map *map;
513 
514 	while ((map = xsk_get_map_list_entry(xs, &map_entry))) {
515 		xsk_map_try_sock_delete(map, xs, map_entry);
516 		xsk_map_put(map);
517 	}
518 }
519 
520 static int xsk_release(struct socket *sock)
521 {
522 	struct sock *sk = sock->sk;
523 	struct xdp_sock *xs = xdp_sk(sk);
524 	struct net *net;
525 
526 	if (!sk)
527 		return 0;
528 
529 	net = sock_net(sk);
530 
531 	mutex_lock(&net->xdp.lock);
532 	sk_del_node_init_rcu(sk);
533 	mutex_unlock(&net->xdp.lock);
534 
535 	local_bh_disable();
536 	sock_prot_inuse_add(net, sk->sk_prot, -1);
537 	local_bh_enable();
538 
539 	xsk_delete_from_maps(xs);
540 	mutex_lock(&xs->mutex);
541 	xsk_unbind_dev(xs);
542 	mutex_unlock(&xs->mutex);
543 
544 	xskq_destroy(xs->rx);
545 	xskq_destroy(xs->tx);
546 
547 	sock_orphan(sk);
548 	sock->sk = NULL;
549 
550 	sk_refcnt_debug_release(sk);
551 	sock_put(sk);
552 
553 	return 0;
554 }
555 
556 static struct socket *xsk_lookup_xsk_from_fd(int fd)
557 {
558 	struct socket *sock;
559 	int err;
560 
561 	sock = sockfd_lookup(fd, &err);
562 	if (!sock)
563 		return ERR_PTR(-ENOTSOCK);
564 
565 	if (sock->sk->sk_family != PF_XDP) {
566 		sockfd_put(sock);
567 		return ERR_PTR(-ENOPROTOOPT);
568 	}
569 
570 	return sock;
571 }
572 
573 /* Check if umem pages are contiguous.
574  * If zero-copy mode, use the DMA address to do the page contiguity check
575  * For all other modes we use addr (kernel virtual address)
576  * Store the result in the low bits of addr.
577  */
578 static void xsk_check_page_contiguity(struct xdp_umem *umem, u32 flags)
579 {
580 	struct xdp_umem_page *pgs = umem->pages;
581 	int i, is_contig;
582 
583 	for (i = 0; i < umem->npgs - 1; i++) {
584 		is_contig = (flags & XDP_ZEROCOPY) ?
585 			(pgs[i].dma + PAGE_SIZE == pgs[i + 1].dma) :
586 			(pgs[i].addr + PAGE_SIZE == pgs[i + 1].addr);
587 		pgs[i].addr += is_contig << XSK_NEXT_PG_CONTIG_SHIFT;
588 	}
589 }
590 
591 static int xsk_bind(struct socket *sock, struct sockaddr *addr, int addr_len)
592 {
593 	struct sockaddr_xdp *sxdp = (struct sockaddr_xdp *)addr;
594 	struct sock *sk = sock->sk;
595 	struct xdp_sock *xs = xdp_sk(sk);
596 	struct net_device *dev;
597 	u32 flags, qid;
598 	int err = 0;
599 
600 	if (addr_len < sizeof(struct sockaddr_xdp))
601 		return -EINVAL;
602 	if (sxdp->sxdp_family != AF_XDP)
603 		return -EINVAL;
604 
605 	flags = sxdp->sxdp_flags;
606 	if (flags & ~(XDP_SHARED_UMEM | XDP_COPY | XDP_ZEROCOPY |
607 		      XDP_USE_NEED_WAKEUP))
608 		return -EINVAL;
609 
610 	rtnl_lock();
611 	mutex_lock(&xs->mutex);
612 	if (xs->state != XSK_READY) {
613 		err = -EBUSY;
614 		goto out_release;
615 	}
616 
617 	dev = dev_get_by_index(sock_net(sk), sxdp->sxdp_ifindex);
618 	if (!dev) {
619 		err = -ENODEV;
620 		goto out_release;
621 	}
622 
623 	if (!xs->rx && !xs->tx) {
624 		err = -EINVAL;
625 		goto out_unlock;
626 	}
627 
628 	qid = sxdp->sxdp_queue_id;
629 
630 	if (flags & XDP_SHARED_UMEM) {
631 		struct xdp_sock *umem_xs;
632 		struct socket *sock;
633 
634 		if ((flags & XDP_COPY) || (flags & XDP_ZEROCOPY) ||
635 		    (flags & XDP_USE_NEED_WAKEUP)) {
636 			/* Cannot specify flags for shared sockets. */
637 			err = -EINVAL;
638 			goto out_unlock;
639 		}
640 
641 		if (xs->umem) {
642 			/* We have already our own. */
643 			err = -EINVAL;
644 			goto out_unlock;
645 		}
646 
647 		sock = xsk_lookup_xsk_from_fd(sxdp->sxdp_shared_umem_fd);
648 		if (IS_ERR(sock)) {
649 			err = PTR_ERR(sock);
650 			goto out_unlock;
651 		}
652 
653 		umem_xs = xdp_sk(sock->sk);
654 		if (!xsk_is_bound(umem_xs)) {
655 			err = -EBADF;
656 			sockfd_put(sock);
657 			goto out_unlock;
658 		}
659 		if (umem_xs->dev != dev || umem_xs->queue_id != qid) {
660 			err = -EINVAL;
661 			sockfd_put(sock);
662 			goto out_unlock;
663 		}
664 
665 		xdp_get_umem(umem_xs->umem);
666 		WRITE_ONCE(xs->umem, umem_xs->umem);
667 		sockfd_put(sock);
668 	} else if (!xs->umem || !xdp_umem_validate_queues(xs->umem)) {
669 		err = -EINVAL;
670 		goto out_unlock;
671 	} else {
672 		/* This xsk has its own umem. */
673 		xskq_set_umem(xs->umem->fq, xs->umem->size,
674 			      xs->umem->chunk_mask);
675 		xskq_set_umem(xs->umem->cq, xs->umem->size,
676 			      xs->umem->chunk_mask);
677 
678 		err = xdp_umem_assign_dev(xs->umem, dev, qid, flags);
679 		if (err)
680 			goto out_unlock;
681 
682 		xsk_check_page_contiguity(xs->umem, flags);
683 	}
684 
685 	xs->dev = dev;
686 	xs->zc = xs->umem->zc;
687 	xs->queue_id = qid;
688 	xskq_set_umem(xs->rx, xs->umem->size, xs->umem->chunk_mask);
689 	xskq_set_umem(xs->tx, xs->umem->size, xs->umem->chunk_mask);
690 	xdp_add_sk_umem(xs->umem, xs);
691 
692 out_unlock:
693 	if (err) {
694 		dev_put(dev);
695 	} else {
696 		/* Matches smp_rmb() in bind() for shared umem
697 		 * sockets, and xsk_is_bound().
698 		 */
699 		smp_wmb();
700 		WRITE_ONCE(xs->state, XSK_BOUND);
701 	}
702 out_release:
703 	mutex_unlock(&xs->mutex);
704 	rtnl_unlock();
705 	return err;
706 }
707 
708 struct xdp_umem_reg_v1 {
709 	__u64 addr; /* Start of packet data area */
710 	__u64 len; /* Length of packet data area */
711 	__u32 chunk_size;
712 	__u32 headroom;
713 };
714 
715 static int xsk_setsockopt(struct socket *sock, int level, int optname,
716 			  char __user *optval, unsigned int optlen)
717 {
718 	struct sock *sk = sock->sk;
719 	struct xdp_sock *xs = xdp_sk(sk);
720 	int err;
721 
722 	if (level != SOL_XDP)
723 		return -ENOPROTOOPT;
724 
725 	switch (optname) {
726 	case XDP_RX_RING:
727 	case XDP_TX_RING:
728 	{
729 		struct xsk_queue **q;
730 		int entries;
731 
732 		if (optlen < sizeof(entries))
733 			return -EINVAL;
734 		if (copy_from_user(&entries, optval, sizeof(entries)))
735 			return -EFAULT;
736 
737 		mutex_lock(&xs->mutex);
738 		if (xs->state != XSK_READY) {
739 			mutex_unlock(&xs->mutex);
740 			return -EBUSY;
741 		}
742 		q = (optname == XDP_TX_RING) ? &xs->tx : &xs->rx;
743 		err = xsk_init_queue(entries, q, false);
744 		if (!err && optname == XDP_TX_RING)
745 			/* Tx needs to be explicitly woken up the first time */
746 			xs->tx->ring->flags |= XDP_RING_NEED_WAKEUP;
747 		mutex_unlock(&xs->mutex);
748 		return err;
749 	}
750 	case XDP_UMEM_REG:
751 	{
752 		size_t mr_size = sizeof(struct xdp_umem_reg);
753 		struct xdp_umem_reg mr = {};
754 		struct xdp_umem *umem;
755 
756 		if (optlen < sizeof(struct xdp_umem_reg_v1))
757 			return -EINVAL;
758 		else if (optlen < sizeof(mr))
759 			mr_size = sizeof(struct xdp_umem_reg_v1);
760 
761 		if (copy_from_user(&mr, optval, mr_size))
762 			return -EFAULT;
763 
764 		mutex_lock(&xs->mutex);
765 		if (xs->state != XSK_READY || xs->umem) {
766 			mutex_unlock(&xs->mutex);
767 			return -EBUSY;
768 		}
769 
770 		umem = xdp_umem_create(&mr);
771 		if (IS_ERR(umem)) {
772 			mutex_unlock(&xs->mutex);
773 			return PTR_ERR(umem);
774 		}
775 
776 		/* Make sure umem is ready before it can be seen by others */
777 		smp_wmb();
778 		WRITE_ONCE(xs->umem, umem);
779 		mutex_unlock(&xs->mutex);
780 		return 0;
781 	}
782 	case XDP_UMEM_FILL_RING:
783 	case XDP_UMEM_COMPLETION_RING:
784 	{
785 		struct xsk_queue **q;
786 		int entries;
787 
788 		if (copy_from_user(&entries, optval, sizeof(entries)))
789 			return -EFAULT;
790 
791 		mutex_lock(&xs->mutex);
792 		if (xs->state != XSK_READY) {
793 			mutex_unlock(&xs->mutex);
794 			return -EBUSY;
795 		}
796 		if (!xs->umem) {
797 			mutex_unlock(&xs->mutex);
798 			return -EINVAL;
799 		}
800 
801 		q = (optname == XDP_UMEM_FILL_RING) ? &xs->umem->fq :
802 			&xs->umem->cq;
803 		err = xsk_init_queue(entries, q, true);
804 		mutex_unlock(&xs->mutex);
805 		return err;
806 	}
807 	default:
808 		break;
809 	}
810 
811 	return -ENOPROTOOPT;
812 }
813 
814 static void xsk_enter_rxtx_offsets(struct xdp_ring_offset_v1 *ring)
815 {
816 	ring->producer = offsetof(struct xdp_rxtx_ring, ptrs.producer);
817 	ring->consumer = offsetof(struct xdp_rxtx_ring, ptrs.consumer);
818 	ring->desc = offsetof(struct xdp_rxtx_ring, desc);
819 }
820 
821 static void xsk_enter_umem_offsets(struct xdp_ring_offset_v1 *ring)
822 {
823 	ring->producer = offsetof(struct xdp_umem_ring, ptrs.producer);
824 	ring->consumer = offsetof(struct xdp_umem_ring, ptrs.consumer);
825 	ring->desc = offsetof(struct xdp_umem_ring, desc);
826 }
827 
828 static int xsk_getsockopt(struct socket *sock, int level, int optname,
829 			  char __user *optval, int __user *optlen)
830 {
831 	struct sock *sk = sock->sk;
832 	struct xdp_sock *xs = xdp_sk(sk);
833 	int len;
834 
835 	if (level != SOL_XDP)
836 		return -ENOPROTOOPT;
837 
838 	if (get_user(len, optlen))
839 		return -EFAULT;
840 	if (len < 0)
841 		return -EINVAL;
842 
843 	switch (optname) {
844 	case XDP_STATISTICS:
845 	{
846 		struct xdp_statistics stats;
847 
848 		if (len < sizeof(stats))
849 			return -EINVAL;
850 
851 		mutex_lock(&xs->mutex);
852 		stats.rx_dropped = xs->rx_dropped;
853 		stats.rx_invalid_descs = xskq_nb_invalid_descs(xs->rx);
854 		stats.tx_invalid_descs = xskq_nb_invalid_descs(xs->tx);
855 		mutex_unlock(&xs->mutex);
856 
857 		if (copy_to_user(optval, &stats, sizeof(stats)))
858 			return -EFAULT;
859 		if (put_user(sizeof(stats), optlen))
860 			return -EFAULT;
861 
862 		return 0;
863 	}
864 	case XDP_MMAP_OFFSETS:
865 	{
866 		struct xdp_mmap_offsets off;
867 		struct xdp_mmap_offsets_v1 off_v1;
868 		bool flags_supported = true;
869 		void *to_copy;
870 
871 		if (len < sizeof(off_v1))
872 			return -EINVAL;
873 		else if (len < sizeof(off))
874 			flags_supported = false;
875 
876 		if (flags_supported) {
877 			/* xdp_ring_offset is identical to xdp_ring_offset_v1
878 			 * except for the flags field added to the end.
879 			 */
880 			xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *)
881 					       &off.rx);
882 			xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *)
883 					       &off.tx);
884 			xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *)
885 					       &off.fr);
886 			xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *)
887 					       &off.cr);
888 			off.rx.flags = offsetof(struct xdp_rxtx_ring,
889 						ptrs.flags);
890 			off.tx.flags = offsetof(struct xdp_rxtx_ring,
891 						ptrs.flags);
892 			off.fr.flags = offsetof(struct xdp_umem_ring,
893 						ptrs.flags);
894 			off.cr.flags = offsetof(struct xdp_umem_ring,
895 						ptrs.flags);
896 
897 			len = sizeof(off);
898 			to_copy = &off;
899 		} else {
900 			xsk_enter_rxtx_offsets(&off_v1.rx);
901 			xsk_enter_rxtx_offsets(&off_v1.tx);
902 			xsk_enter_umem_offsets(&off_v1.fr);
903 			xsk_enter_umem_offsets(&off_v1.cr);
904 
905 			len = sizeof(off_v1);
906 			to_copy = &off_v1;
907 		}
908 
909 		if (copy_to_user(optval, to_copy, len))
910 			return -EFAULT;
911 		if (put_user(len, optlen))
912 			return -EFAULT;
913 
914 		return 0;
915 	}
916 	case XDP_OPTIONS:
917 	{
918 		struct xdp_options opts = {};
919 
920 		if (len < sizeof(opts))
921 			return -EINVAL;
922 
923 		mutex_lock(&xs->mutex);
924 		if (xs->zc)
925 			opts.flags |= XDP_OPTIONS_ZEROCOPY;
926 		mutex_unlock(&xs->mutex);
927 
928 		len = sizeof(opts);
929 		if (copy_to_user(optval, &opts, len))
930 			return -EFAULT;
931 		if (put_user(len, optlen))
932 			return -EFAULT;
933 
934 		return 0;
935 	}
936 	default:
937 		break;
938 	}
939 
940 	return -EOPNOTSUPP;
941 }
942 
943 static int xsk_mmap(struct file *file, struct socket *sock,
944 		    struct vm_area_struct *vma)
945 {
946 	loff_t offset = (loff_t)vma->vm_pgoff << PAGE_SHIFT;
947 	unsigned long size = vma->vm_end - vma->vm_start;
948 	struct xdp_sock *xs = xdp_sk(sock->sk);
949 	struct xsk_queue *q = NULL;
950 	struct xdp_umem *umem;
951 	unsigned long pfn;
952 	struct page *qpg;
953 
954 	if (READ_ONCE(xs->state) != XSK_READY)
955 		return -EBUSY;
956 
957 	if (offset == XDP_PGOFF_RX_RING) {
958 		q = READ_ONCE(xs->rx);
959 	} else if (offset == XDP_PGOFF_TX_RING) {
960 		q = READ_ONCE(xs->tx);
961 	} else {
962 		umem = READ_ONCE(xs->umem);
963 		if (!umem)
964 			return -EINVAL;
965 
966 		/* Matches the smp_wmb() in XDP_UMEM_REG */
967 		smp_rmb();
968 		if (offset == XDP_UMEM_PGOFF_FILL_RING)
969 			q = READ_ONCE(umem->fq);
970 		else if (offset == XDP_UMEM_PGOFF_COMPLETION_RING)
971 			q = READ_ONCE(umem->cq);
972 	}
973 
974 	if (!q)
975 		return -EINVAL;
976 
977 	/* Matches the smp_wmb() in xsk_init_queue */
978 	smp_rmb();
979 	qpg = virt_to_head_page(q->ring);
980 	if (size > (PAGE_SIZE << compound_order(qpg)))
981 		return -EINVAL;
982 
983 	pfn = virt_to_phys(q->ring) >> PAGE_SHIFT;
984 	return remap_pfn_range(vma, vma->vm_start, pfn,
985 			       size, vma->vm_page_prot);
986 }
987 
988 static int xsk_notifier(struct notifier_block *this,
989 			unsigned long msg, void *ptr)
990 {
991 	struct net_device *dev = netdev_notifier_info_to_dev(ptr);
992 	struct net *net = dev_net(dev);
993 	struct sock *sk;
994 
995 	switch (msg) {
996 	case NETDEV_UNREGISTER:
997 		mutex_lock(&net->xdp.lock);
998 		sk_for_each(sk, &net->xdp.list) {
999 			struct xdp_sock *xs = xdp_sk(sk);
1000 
1001 			mutex_lock(&xs->mutex);
1002 			if (xs->dev == dev) {
1003 				sk->sk_err = ENETDOWN;
1004 				if (!sock_flag(sk, SOCK_DEAD))
1005 					sk->sk_error_report(sk);
1006 
1007 				xsk_unbind_dev(xs);
1008 
1009 				/* Clear device references in umem. */
1010 				xdp_umem_clear_dev(xs->umem);
1011 			}
1012 			mutex_unlock(&xs->mutex);
1013 		}
1014 		mutex_unlock(&net->xdp.lock);
1015 		break;
1016 	}
1017 	return NOTIFY_DONE;
1018 }
1019 
1020 static struct proto xsk_proto = {
1021 	.name =		"XDP",
1022 	.owner =	THIS_MODULE,
1023 	.obj_size =	sizeof(struct xdp_sock),
1024 };
1025 
1026 static const struct proto_ops xsk_proto_ops = {
1027 	.family		= PF_XDP,
1028 	.owner		= THIS_MODULE,
1029 	.release	= xsk_release,
1030 	.bind		= xsk_bind,
1031 	.connect	= sock_no_connect,
1032 	.socketpair	= sock_no_socketpair,
1033 	.accept		= sock_no_accept,
1034 	.getname	= sock_no_getname,
1035 	.poll		= xsk_poll,
1036 	.ioctl		= sock_no_ioctl,
1037 	.listen		= sock_no_listen,
1038 	.shutdown	= sock_no_shutdown,
1039 	.setsockopt	= xsk_setsockopt,
1040 	.getsockopt	= xsk_getsockopt,
1041 	.sendmsg	= xsk_sendmsg,
1042 	.recvmsg	= sock_no_recvmsg,
1043 	.mmap		= xsk_mmap,
1044 	.sendpage	= sock_no_sendpage,
1045 };
1046 
1047 static void xsk_destruct(struct sock *sk)
1048 {
1049 	struct xdp_sock *xs = xdp_sk(sk);
1050 
1051 	if (!sock_flag(sk, SOCK_DEAD))
1052 		return;
1053 
1054 	xdp_put_umem(xs->umem);
1055 
1056 	sk_refcnt_debug_dec(sk);
1057 }
1058 
1059 static int xsk_create(struct net *net, struct socket *sock, int protocol,
1060 		      int kern)
1061 {
1062 	struct sock *sk;
1063 	struct xdp_sock *xs;
1064 
1065 	if (!ns_capable(net->user_ns, CAP_NET_RAW))
1066 		return -EPERM;
1067 	if (sock->type != SOCK_RAW)
1068 		return -ESOCKTNOSUPPORT;
1069 
1070 	if (protocol)
1071 		return -EPROTONOSUPPORT;
1072 
1073 	sock->state = SS_UNCONNECTED;
1074 
1075 	sk = sk_alloc(net, PF_XDP, GFP_KERNEL, &xsk_proto, kern);
1076 	if (!sk)
1077 		return -ENOBUFS;
1078 
1079 	sock->ops = &xsk_proto_ops;
1080 
1081 	sock_init_data(sock, sk);
1082 
1083 	sk->sk_family = PF_XDP;
1084 
1085 	sk->sk_destruct = xsk_destruct;
1086 	sk_refcnt_debug_inc(sk);
1087 
1088 	sock_set_flag(sk, SOCK_RCU_FREE);
1089 
1090 	xs = xdp_sk(sk);
1091 	xs->state = XSK_READY;
1092 	mutex_init(&xs->mutex);
1093 	spin_lock_init(&xs->rx_lock);
1094 	spin_lock_init(&xs->tx_completion_lock);
1095 
1096 	INIT_LIST_HEAD(&xs->map_list);
1097 	spin_lock_init(&xs->map_list_lock);
1098 
1099 	mutex_lock(&net->xdp.lock);
1100 	sk_add_node_rcu(sk, &net->xdp.list);
1101 	mutex_unlock(&net->xdp.lock);
1102 
1103 	local_bh_disable();
1104 	sock_prot_inuse_add(net, &xsk_proto, 1);
1105 	local_bh_enable();
1106 
1107 	return 0;
1108 }
1109 
1110 static const struct net_proto_family xsk_family_ops = {
1111 	.family = PF_XDP,
1112 	.create = xsk_create,
1113 	.owner	= THIS_MODULE,
1114 };
1115 
1116 static struct notifier_block xsk_netdev_notifier = {
1117 	.notifier_call	= xsk_notifier,
1118 };
1119 
1120 static int __net_init xsk_net_init(struct net *net)
1121 {
1122 	mutex_init(&net->xdp.lock);
1123 	INIT_HLIST_HEAD(&net->xdp.list);
1124 	return 0;
1125 }
1126 
1127 static void __net_exit xsk_net_exit(struct net *net)
1128 {
1129 	WARN_ON_ONCE(!hlist_empty(&net->xdp.list));
1130 }
1131 
1132 static struct pernet_operations xsk_net_ops = {
1133 	.init = xsk_net_init,
1134 	.exit = xsk_net_exit,
1135 };
1136 
1137 static int __init xsk_init(void)
1138 {
1139 	int err;
1140 
1141 	err = proto_register(&xsk_proto, 0 /* no slab */);
1142 	if (err)
1143 		goto out;
1144 
1145 	err = sock_register(&xsk_family_ops);
1146 	if (err)
1147 		goto out_proto;
1148 
1149 	err = register_pernet_subsys(&xsk_net_ops);
1150 	if (err)
1151 		goto out_sk;
1152 
1153 	err = register_netdevice_notifier(&xsk_netdev_notifier);
1154 	if (err)
1155 		goto out_pernet;
1156 
1157 	return 0;
1158 
1159 out_pernet:
1160 	unregister_pernet_subsys(&xsk_net_ops);
1161 out_sk:
1162 	sock_unregister(PF_XDP);
1163 out_proto:
1164 	proto_unregister(&xsk_proto);
1165 out:
1166 	return err;
1167 }
1168 
1169 fs_initcall(xsk_init);
1170