xref: /openbmc/linux/net/xdp/xsk.c (revision b285d2ae)
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_drv.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 void xsk_set_rx_need_wakeup(struct xdp_umem *umem)
43 {
44 	if (umem->need_wakeup & XDP_WAKEUP_RX)
45 		return;
46 
47 	umem->fq->ring->flags |= XDP_RING_NEED_WAKEUP;
48 	umem->need_wakeup |= XDP_WAKEUP_RX;
49 }
50 EXPORT_SYMBOL(xsk_set_rx_need_wakeup);
51 
52 void xsk_set_tx_need_wakeup(struct xdp_umem *umem)
53 {
54 	struct xdp_sock *xs;
55 
56 	if (umem->need_wakeup & XDP_WAKEUP_TX)
57 		return;
58 
59 	rcu_read_lock();
60 	list_for_each_entry_rcu(xs, &umem->xsk_tx_list, list) {
61 		xs->tx->ring->flags |= XDP_RING_NEED_WAKEUP;
62 	}
63 	rcu_read_unlock();
64 
65 	umem->need_wakeup |= XDP_WAKEUP_TX;
66 }
67 EXPORT_SYMBOL(xsk_set_tx_need_wakeup);
68 
69 void xsk_clear_rx_need_wakeup(struct xdp_umem *umem)
70 {
71 	if (!(umem->need_wakeup & XDP_WAKEUP_RX))
72 		return;
73 
74 	umem->fq->ring->flags &= ~XDP_RING_NEED_WAKEUP;
75 	umem->need_wakeup &= ~XDP_WAKEUP_RX;
76 }
77 EXPORT_SYMBOL(xsk_clear_rx_need_wakeup);
78 
79 void xsk_clear_tx_need_wakeup(struct xdp_umem *umem)
80 {
81 	struct xdp_sock *xs;
82 
83 	if (!(umem->need_wakeup & XDP_WAKEUP_TX))
84 		return;
85 
86 	rcu_read_lock();
87 	list_for_each_entry_rcu(xs, &umem->xsk_tx_list, list) {
88 		xs->tx->ring->flags &= ~XDP_RING_NEED_WAKEUP;
89 	}
90 	rcu_read_unlock();
91 
92 	umem->need_wakeup &= ~XDP_WAKEUP_TX;
93 }
94 EXPORT_SYMBOL(xsk_clear_tx_need_wakeup);
95 
96 bool xsk_umem_uses_need_wakeup(struct xdp_umem *umem)
97 {
98 	return umem->flags & XDP_UMEM_USES_NEED_WAKEUP;
99 }
100 EXPORT_SYMBOL(xsk_umem_uses_need_wakeup);
101 
102 void xp_release(struct xdp_buff_xsk *xskb)
103 {
104 	xskb->pool->free_heads[xskb->pool->free_heads_cnt++] = xskb;
105 }
106 
107 static u64 xp_get_handle(struct xdp_buff_xsk *xskb)
108 {
109 	u64 offset = xskb->xdp.data - xskb->xdp.data_hard_start;
110 
111 	offset += xskb->pool->headroom;
112 	if (!xskb->pool->unaligned)
113 		return xskb->orig_addr + offset;
114 	return xskb->orig_addr + (offset << XSK_UNALIGNED_BUF_OFFSET_SHIFT);
115 }
116 
117 static int __xsk_rcv_zc(struct xdp_sock *xs, struct xdp_buff *xdp, u32 len)
118 {
119 	struct xdp_buff_xsk *xskb = container_of(xdp, struct xdp_buff_xsk, xdp);
120 	u64 addr;
121 	int err;
122 
123 	addr = xp_get_handle(xskb);
124 	err = xskq_prod_reserve_desc(xs->rx, addr, len);
125 	if (err) {
126 		xs->rx_queue_full++;
127 		return err;
128 	}
129 
130 	xp_release(xskb);
131 	return 0;
132 }
133 
134 static void xsk_copy_xdp(struct xdp_buff *to, struct xdp_buff *from, u32 len)
135 {
136 	void *from_buf, *to_buf;
137 	u32 metalen;
138 
139 	if (unlikely(xdp_data_meta_unsupported(from))) {
140 		from_buf = from->data;
141 		to_buf = to->data;
142 		metalen = 0;
143 	} else {
144 		from_buf = from->data_meta;
145 		metalen = from->data - from->data_meta;
146 		to_buf = to->data - metalen;
147 	}
148 
149 	memcpy(to_buf, from_buf, len + metalen);
150 }
151 
152 static int __xsk_rcv(struct xdp_sock *xs, struct xdp_buff *xdp, u32 len,
153 		     bool explicit_free)
154 {
155 	struct xdp_buff *xsk_xdp;
156 	int err;
157 
158 	if (len > xsk_umem_get_rx_frame_size(xs->umem)) {
159 		xs->rx_dropped++;
160 		return -ENOSPC;
161 	}
162 
163 	xsk_xdp = xsk_buff_alloc(xs->umem);
164 	if (!xsk_xdp) {
165 		xs->rx_dropped++;
166 		return -ENOSPC;
167 	}
168 
169 	xsk_copy_xdp(xsk_xdp, xdp, len);
170 	err = __xsk_rcv_zc(xs, xsk_xdp, len);
171 	if (err) {
172 		xsk_buff_free(xsk_xdp);
173 		return err;
174 	}
175 	if (explicit_free)
176 		xdp_return_buff(xdp);
177 	return 0;
178 }
179 
180 static bool xsk_is_bound(struct xdp_sock *xs)
181 {
182 	if (READ_ONCE(xs->state) == XSK_BOUND) {
183 		/* Matches smp_wmb() in bind(). */
184 		smp_rmb();
185 		return true;
186 	}
187 	return false;
188 }
189 
190 static int xsk_rcv(struct xdp_sock *xs, struct xdp_buff *xdp,
191 		   bool explicit_free)
192 {
193 	u32 len;
194 
195 	if (!xsk_is_bound(xs))
196 		return -EINVAL;
197 
198 	if (xs->dev != xdp->rxq->dev || xs->queue_id != xdp->rxq->queue_index)
199 		return -EINVAL;
200 
201 	len = xdp->data_end - xdp->data;
202 
203 	return xdp->rxq->mem.type == MEM_TYPE_XSK_BUFF_POOL ?
204 		__xsk_rcv_zc(xs, xdp, len) :
205 		__xsk_rcv(xs, xdp, len, explicit_free);
206 }
207 
208 static void xsk_flush(struct xdp_sock *xs)
209 {
210 	xskq_prod_submit(xs->rx);
211 	__xskq_cons_release(xs->umem->fq);
212 	sock_def_readable(&xs->sk);
213 }
214 
215 int xsk_generic_rcv(struct xdp_sock *xs, struct xdp_buff *xdp)
216 {
217 	int err;
218 
219 	spin_lock_bh(&xs->rx_lock);
220 	err = xsk_rcv(xs, xdp, false);
221 	xsk_flush(xs);
222 	spin_unlock_bh(&xs->rx_lock);
223 	return err;
224 }
225 
226 int __xsk_map_redirect(struct xdp_sock *xs, struct xdp_buff *xdp)
227 {
228 	struct list_head *flush_list = this_cpu_ptr(&xskmap_flush_list);
229 	int err;
230 
231 	err = xsk_rcv(xs, xdp, true);
232 	if (err)
233 		return err;
234 
235 	if (!xs->flush_node.prev)
236 		list_add(&xs->flush_node, flush_list);
237 
238 	return 0;
239 }
240 
241 void __xsk_map_flush(void)
242 {
243 	struct list_head *flush_list = this_cpu_ptr(&xskmap_flush_list);
244 	struct xdp_sock *xs, *tmp;
245 
246 	list_for_each_entry_safe(xs, tmp, flush_list, flush_node) {
247 		xsk_flush(xs);
248 		__list_del_clearprev(&xs->flush_node);
249 	}
250 }
251 
252 void xsk_umem_complete_tx(struct xdp_umem *umem, u32 nb_entries)
253 {
254 	xskq_prod_submit_n(umem->cq, nb_entries);
255 }
256 EXPORT_SYMBOL(xsk_umem_complete_tx);
257 
258 void xsk_umem_consume_tx_done(struct xdp_umem *umem)
259 {
260 	struct xdp_sock *xs;
261 
262 	rcu_read_lock();
263 	list_for_each_entry_rcu(xs, &umem->xsk_tx_list, list) {
264 		__xskq_cons_release(xs->tx);
265 		xs->sk.sk_write_space(&xs->sk);
266 	}
267 	rcu_read_unlock();
268 }
269 EXPORT_SYMBOL(xsk_umem_consume_tx_done);
270 
271 bool xsk_umem_consume_tx(struct xdp_umem *umem, struct xdp_desc *desc)
272 {
273 	struct xdp_sock *xs;
274 
275 	rcu_read_lock();
276 	list_for_each_entry_rcu(xs, &umem->xsk_tx_list, list) {
277 		if (!xskq_cons_peek_desc(xs->tx, desc, umem)) {
278 			xs->tx->queue_empty_descs++;
279 			continue;
280 		}
281 
282 		/* This is the backpressure mechanism for the Tx path.
283 		 * Reserve space in the completion queue and only proceed
284 		 * if there is space in it. This avoids having to implement
285 		 * any buffering in the Tx path.
286 		 */
287 		if (xskq_prod_reserve_addr(umem->cq, desc->addr))
288 			goto out;
289 
290 		xskq_cons_release(xs->tx);
291 		rcu_read_unlock();
292 		return true;
293 	}
294 
295 out:
296 	rcu_read_unlock();
297 	return false;
298 }
299 EXPORT_SYMBOL(xsk_umem_consume_tx);
300 
301 static int xsk_wakeup(struct xdp_sock *xs, u8 flags)
302 {
303 	struct net_device *dev = xs->dev;
304 	int err;
305 
306 	rcu_read_lock();
307 	err = dev->netdev_ops->ndo_xsk_wakeup(dev, xs->queue_id, flags);
308 	rcu_read_unlock();
309 
310 	return err;
311 }
312 
313 static int xsk_zc_xmit(struct xdp_sock *xs)
314 {
315 	return xsk_wakeup(xs, XDP_WAKEUP_TX);
316 }
317 
318 static void xsk_destruct_skb(struct sk_buff *skb)
319 {
320 	u64 addr = (u64)(long)skb_shinfo(skb)->destructor_arg;
321 	struct xdp_sock *xs = xdp_sk(skb->sk);
322 	unsigned long flags;
323 
324 	spin_lock_irqsave(&xs->tx_completion_lock, flags);
325 	xskq_prod_submit_addr(xs->umem->cq, addr);
326 	spin_unlock_irqrestore(&xs->tx_completion_lock, flags);
327 
328 	sock_wfree(skb);
329 }
330 
331 static int xsk_generic_xmit(struct sock *sk)
332 {
333 	struct xdp_sock *xs = xdp_sk(sk);
334 	u32 max_batch = TX_BATCH_SIZE;
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_cons_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 			goto out;
359 
360 		skb_put(skb, len);
361 		addr = desc.addr;
362 		buffer = xsk_buff_raw_get_data(xs->umem, addr);
363 		err = skb_store_bits(skb, 0, buffer, len);
364 		/* This is the backpressure mechanism for the Tx path.
365 		 * Reserve space in the completion queue and only proceed
366 		 * if there is space in it. This avoids having to implement
367 		 * any buffering in the Tx path.
368 		 */
369 		if (unlikely(err) || xskq_prod_reserve(xs->umem->cq)) {
370 			kfree_skb(skb);
371 			goto out;
372 		}
373 
374 		skb->dev = xs->dev;
375 		skb->priority = sk->sk_priority;
376 		skb->mark = sk->sk_mark;
377 		skb_shinfo(skb)->destructor_arg = (void *)(long)desc.addr;
378 		skb->destructor = xsk_destruct_skb;
379 
380 		err = dev_direct_xmit(skb, xs->queue_id);
381 		xskq_cons_release(xs->tx);
382 		/* Ignore NET_XMIT_CN as packet might have been sent */
383 		if (err == NET_XMIT_DROP || err == NETDEV_TX_BUSY) {
384 			/* SKB completed but not sent */
385 			err = -EBUSY;
386 			goto out;
387 		}
388 
389 		sent_frame = true;
390 	}
391 
392 	xs->tx->queue_empty_descs++;
393 
394 out:
395 	if (sent_frame)
396 		sk->sk_write_space(sk);
397 
398 	mutex_unlock(&xs->mutex);
399 	return err;
400 }
401 
402 static int __xsk_sendmsg(struct sock *sk)
403 {
404 	struct xdp_sock *xs = xdp_sk(sk);
405 
406 	if (unlikely(!(xs->dev->flags & IFF_UP)))
407 		return -ENETDOWN;
408 	if (unlikely(!xs->tx))
409 		return -ENOBUFS;
410 
411 	return xs->zc ? xsk_zc_xmit(xs) : xsk_generic_xmit(sk);
412 }
413 
414 static int xsk_sendmsg(struct socket *sock, struct msghdr *m, size_t total_len)
415 {
416 	bool need_wait = !(m->msg_flags & MSG_DONTWAIT);
417 	struct sock *sk = sock->sk;
418 	struct xdp_sock *xs = xdp_sk(sk);
419 
420 	if (unlikely(!xsk_is_bound(xs)))
421 		return -ENXIO;
422 	if (unlikely(need_wait))
423 		return -EOPNOTSUPP;
424 
425 	return __xsk_sendmsg(sk);
426 }
427 
428 static __poll_t xsk_poll(struct file *file, struct socket *sock,
429 			     struct poll_table_struct *wait)
430 {
431 	__poll_t mask = datagram_poll(file, sock, wait);
432 	struct sock *sk = sock->sk;
433 	struct xdp_sock *xs = xdp_sk(sk);
434 	struct xdp_umem *umem;
435 
436 	if (unlikely(!xsk_is_bound(xs)))
437 		return mask;
438 
439 	umem = xs->umem;
440 
441 	if (umem->need_wakeup) {
442 		if (xs->zc)
443 			xsk_wakeup(xs, umem->need_wakeup);
444 		else
445 			/* Poll needs to drive Tx also in copy mode */
446 			__xsk_sendmsg(sk);
447 	}
448 
449 	if (xs->rx && !xskq_prod_is_empty(xs->rx))
450 		mask |= EPOLLIN | EPOLLRDNORM;
451 	if (xs->tx && !xskq_cons_is_full(xs->tx))
452 		mask |= EPOLLOUT | EPOLLWRNORM;
453 
454 	return mask;
455 }
456 
457 static int xsk_init_queue(u32 entries, struct xsk_queue **queue,
458 			  bool umem_queue)
459 {
460 	struct xsk_queue *q;
461 
462 	if (entries == 0 || *queue || !is_power_of_2(entries))
463 		return -EINVAL;
464 
465 	q = xskq_create(entries, umem_queue);
466 	if (!q)
467 		return -ENOMEM;
468 
469 	/* Make sure queue is ready before it can be seen by others */
470 	smp_wmb();
471 	WRITE_ONCE(*queue, q);
472 	return 0;
473 }
474 
475 static void xsk_unbind_dev(struct xdp_sock *xs)
476 {
477 	struct net_device *dev = xs->dev;
478 
479 	if (xs->state != XSK_BOUND)
480 		return;
481 	WRITE_ONCE(xs->state, XSK_UNBOUND);
482 
483 	/* Wait for driver to stop using the xdp socket. */
484 	xdp_del_sk_umem(xs->umem, xs);
485 	xs->dev = NULL;
486 	synchronize_net();
487 	dev_put(dev);
488 }
489 
490 static struct xsk_map *xsk_get_map_list_entry(struct xdp_sock *xs,
491 					      struct xdp_sock ***map_entry)
492 {
493 	struct xsk_map *map = NULL;
494 	struct xsk_map_node *node;
495 
496 	*map_entry = NULL;
497 
498 	spin_lock_bh(&xs->map_list_lock);
499 	node = list_first_entry_or_null(&xs->map_list, struct xsk_map_node,
500 					node);
501 	if (node) {
502 		WARN_ON(xsk_map_inc(node->map));
503 		map = node->map;
504 		*map_entry = node->map_entry;
505 	}
506 	spin_unlock_bh(&xs->map_list_lock);
507 	return map;
508 }
509 
510 static void xsk_delete_from_maps(struct xdp_sock *xs)
511 {
512 	/* This function removes the current XDP socket from all the
513 	 * maps it resides in. We need to take extra care here, due to
514 	 * the two locks involved. Each map has a lock synchronizing
515 	 * updates to the entries, and each socket has a lock that
516 	 * synchronizes access to the list of maps (map_list). For
517 	 * deadlock avoidance the locks need to be taken in the order
518 	 * "map lock"->"socket map list lock". We start off by
519 	 * accessing the socket map list, and take a reference to the
520 	 * map to guarantee existence between the
521 	 * xsk_get_map_list_entry() and xsk_map_try_sock_delete()
522 	 * calls. Then we ask the map to remove the socket, which
523 	 * tries to remove the socket from the map. Note that there
524 	 * might be updates to the map between
525 	 * xsk_get_map_list_entry() and xsk_map_try_sock_delete().
526 	 */
527 	struct xdp_sock **map_entry = NULL;
528 	struct xsk_map *map;
529 
530 	while ((map = xsk_get_map_list_entry(xs, &map_entry))) {
531 		xsk_map_try_sock_delete(map, xs, map_entry);
532 		xsk_map_put(map);
533 	}
534 }
535 
536 static int xsk_release(struct socket *sock)
537 {
538 	struct sock *sk = sock->sk;
539 	struct xdp_sock *xs = xdp_sk(sk);
540 	struct net *net;
541 
542 	if (!sk)
543 		return 0;
544 
545 	net = sock_net(sk);
546 
547 	mutex_lock(&net->xdp.lock);
548 	sk_del_node_init_rcu(sk);
549 	mutex_unlock(&net->xdp.lock);
550 
551 	local_bh_disable();
552 	sock_prot_inuse_add(net, sk->sk_prot, -1);
553 	local_bh_enable();
554 
555 	xsk_delete_from_maps(xs);
556 	mutex_lock(&xs->mutex);
557 	xsk_unbind_dev(xs);
558 	mutex_unlock(&xs->mutex);
559 
560 	xskq_destroy(xs->rx);
561 	xskq_destroy(xs->tx);
562 
563 	sock_orphan(sk);
564 	sock->sk = NULL;
565 
566 	sk_refcnt_debug_release(sk);
567 	sock_put(sk);
568 
569 	return 0;
570 }
571 
572 static struct socket *xsk_lookup_xsk_from_fd(int fd)
573 {
574 	struct socket *sock;
575 	int err;
576 
577 	sock = sockfd_lookup(fd, &err);
578 	if (!sock)
579 		return ERR_PTR(-ENOTSOCK);
580 
581 	if (sock->sk->sk_family != PF_XDP) {
582 		sockfd_put(sock);
583 		return ERR_PTR(-ENOPROTOOPT);
584 	}
585 
586 	return sock;
587 }
588 
589 static int xsk_bind(struct socket *sock, struct sockaddr *addr, int addr_len)
590 {
591 	struct sockaddr_xdp *sxdp = (struct sockaddr_xdp *)addr;
592 	struct sock *sk = sock->sk;
593 	struct xdp_sock *xs = xdp_sk(sk);
594 	struct net_device *dev;
595 	u32 flags, qid;
596 	int err = 0;
597 
598 	if (addr_len < sizeof(struct sockaddr_xdp))
599 		return -EINVAL;
600 	if (sxdp->sxdp_family != AF_XDP)
601 		return -EINVAL;
602 
603 	flags = sxdp->sxdp_flags;
604 	if (flags & ~(XDP_SHARED_UMEM | XDP_COPY | XDP_ZEROCOPY |
605 		      XDP_USE_NEED_WAKEUP))
606 		return -EINVAL;
607 
608 	rtnl_lock();
609 	mutex_lock(&xs->mutex);
610 	if (xs->state != XSK_READY) {
611 		err = -EBUSY;
612 		goto out_release;
613 	}
614 
615 	dev = dev_get_by_index(sock_net(sk), sxdp->sxdp_ifindex);
616 	if (!dev) {
617 		err = -ENODEV;
618 		goto out_release;
619 	}
620 
621 	if (!xs->rx && !xs->tx) {
622 		err = -EINVAL;
623 		goto out_unlock;
624 	}
625 
626 	qid = sxdp->sxdp_queue_id;
627 
628 	if (flags & XDP_SHARED_UMEM) {
629 		struct xdp_sock *umem_xs;
630 		struct socket *sock;
631 
632 		if ((flags & XDP_COPY) || (flags & XDP_ZEROCOPY) ||
633 		    (flags & XDP_USE_NEED_WAKEUP)) {
634 			/* Cannot specify flags for shared sockets. */
635 			err = -EINVAL;
636 			goto out_unlock;
637 		}
638 
639 		if (xs->umem) {
640 			/* We have already our own. */
641 			err = -EINVAL;
642 			goto out_unlock;
643 		}
644 
645 		sock = xsk_lookup_xsk_from_fd(sxdp->sxdp_shared_umem_fd);
646 		if (IS_ERR(sock)) {
647 			err = PTR_ERR(sock);
648 			goto out_unlock;
649 		}
650 
651 		umem_xs = xdp_sk(sock->sk);
652 		if (!xsk_is_bound(umem_xs)) {
653 			err = -EBADF;
654 			sockfd_put(sock);
655 			goto out_unlock;
656 		}
657 		if (umem_xs->dev != dev || umem_xs->queue_id != qid) {
658 			err = -EINVAL;
659 			sockfd_put(sock);
660 			goto out_unlock;
661 		}
662 
663 		xdp_get_umem(umem_xs->umem);
664 		WRITE_ONCE(xs->umem, umem_xs->umem);
665 		sockfd_put(sock);
666 	} else if (!xs->umem || !xdp_umem_validate_queues(xs->umem)) {
667 		err = -EINVAL;
668 		goto out_unlock;
669 	} else {
670 		/* This xsk has its own umem. */
671 		err = xdp_umem_assign_dev(xs->umem, dev, qid, flags);
672 		if (err)
673 			goto out_unlock;
674 	}
675 
676 	xs->dev = dev;
677 	xs->zc = xs->umem->zc;
678 	xs->queue_id = qid;
679 	xdp_add_sk_umem(xs->umem, xs);
680 
681 out_unlock:
682 	if (err) {
683 		dev_put(dev);
684 	} else {
685 		/* Matches smp_rmb() in bind() for shared umem
686 		 * sockets, and xsk_is_bound().
687 		 */
688 		smp_wmb();
689 		WRITE_ONCE(xs->state, XSK_BOUND);
690 	}
691 out_release:
692 	mutex_unlock(&xs->mutex);
693 	rtnl_unlock();
694 	return err;
695 }
696 
697 struct xdp_umem_reg_v1 {
698 	__u64 addr; /* Start of packet data area */
699 	__u64 len; /* Length of packet data area */
700 	__u32 chunk_size;
701 	__u32 headroom;
702 };
703 
704 static int xsk_setsockopt(struct socket *sock, int level, int optname,
705 			  sockptr_t optval, unsigned int optlen)
706 {
707 	struct sock *sk = sock->sk;
708 	struct xdp_sock *xs = xdp_sk(sk);
709 	int err;
710 
711 	if (level != SOL_XDP)
712 		return -ENOPROTOOPT;
713 
714 	switch (optname) {
715 	case XDP_RX_RING:
716 	case XDP_TX_RING:
717 	{
718 		struct xsk_queue **q;
719 		int entries;
720 
721 		if (optlen < sizeof(entries))
722 			return -EINVAL;
723 		if (copy_from_sockptr(&entries, optval, sizeof(entries)))
724 			return -EFAULT;
725 
726 		mutex_lock(&xs->mutex);
727 		if (xs->state != XSK_READY) {
728 			mutex_unlock(&xs->mutex);
729 			return -EBUSY;
730 		}
731 		q = (optname == XDP_TX_RING) ? &xs->tx : &xs->rx;
732 		err = xsk_init_queue(entries, q, false);
733 		if (!err && optname == XDP_TX_RING)
734 			/* Tx needs to be explicitly woken up the first time */
735 			xs->tx->ring->flags |= XDP_RING_NEED_WAKEUP;
736 		mutex_unlock(&xs->mutex);
737 		return err;
738 	}
739 	case XDP_UMEM_REG:
740 	{
741 		size_t mr_size = sizeof(struct xdp_umem_reg);
742 		struct xdp_umem_reg mr = {};
743 		struct xdp_umem *umem;
744 
745 		if (optlen < sizeof(struct xdp_umem_reg_v1))
746 			return -EINVAL;
747 		else if (optlen < sizeof(mr))
748 			mr_size = sizeof(struct xdp_umem_reg_v1);
749 
750 		if (copy_from_sockptr(&mr, optval, mr_size))
751 			return -EFAULT;
752 
753 		mutex_lock(&xs->mutex);
754 		if (xs->state != XSK_READY || xs->umem) {
755 			mutex_unlock(&xs->mutex);
756 			return -EBUSY;
757 		}
758 
759 		umem = xdp_umem_create(&mr);
760 		if (IS_ERR(umem)) {
761 			mutex_unlock(&xs->mutex);
762 			return PTR_ERR(umem);
763 		}
764 
765 		/* Make sure umem is ready before it can be seen by others */
766 		smp_wmb();
767 		WRITE_ONCE(xs->umem, umem);
768 		mutex_unlock(&xs->mutex);
769 		return 0;
770 	}
771 	case XDP_UMEM_FILL_RING:
772 	case XDP_UMEM_COMPLETION_RING:
773 	{
774 		struct xsk_queue **q;
775 		int entries;
776 
777 		if (copy_from_sockptr(&entries, optval, sizeof(entries)))
778 			return -EFAULT;
779 
780 		mutex_lock(&xs->mutex);
781 		if (xs->state != XSK_READY) {
782 			mutex_unlock(&xs->mutex);
783 			return -EBUSY;
784 		}
785 		if (!xs->umem) {
786 			mutex_unlock(&xs->mutex);
787 			return -EINVAL;
788 		}
789 
790 		q = (optname == XDP_UMEM_FILL_RING) ? &xs->umem->fq :
791 			&xs->umem->cq;
792 		err = xsk_init_queue(entries, q, true);
793 		if (optname == XDP_UMEM_FILL_RING)
794 			xp_set_fq(xs->umem->pool, *q);
795 		mutex_unlock(&xs->mutex);
796 		return err;
797 	}
798 	default:
799 		break;
800 	}
801 
802 	return -ENOPROTOOPT;
803 }
804 
805 static void xsk_enter_rxtx_offsets(struct xdp_ring_offset_v1 *ring)
806 {
807 	ring->producer = offsetof(struct xdp_rxtx_ring, ptrs.producer);
808 	ring->consumer = offsetof(struct xdp_rxtx_ring, ptrs.consumer);
809 	ring->desc = offsetof(struct xdp_rxtx_ring, desc);
810 }
811 
812 static void xsk_enter_umem_offsets(struct xdp_ring_offset_v1 *ring)
813 {
814 	ring->producer = offsetof(struct xdp_umem_ring, ptrs.producer);
815 	ring->consumer = offsetof(struct xdp_umem_ring, ptrs.consumer);
816 	ring->desc = offsetof(struct xdp_umem_ring, desc);
817 }
818 
819 struct xdp_statistics_v1 {
820 	__u64 rx_dropped;
821 	__u64 rx_invalid_descs;
822 	__u64 tx_invalid_descs;
823 };
824 
825 static int xsk_getsockopt(struct socket *sock, int level, int optname,
826 			  char __user *optval, int __user *optlen)
827 {
828 	struct sock *sk = sock->sk;
829 	struct xdp_sock *xs = xdp_sk(sk);
830 	int len;
831 
832 	if (level != SOL_XDP)
833 		return -ENOPROTOOPT;
834 
835 	if (get_user(len, optlen))
836 		return -EFAULT;
837 	if (len < 0)
838 		return -EINVAL;
839 
840 	switch (optname) {
841 	case XDP_STATISTICS:
842 	{
843 		struct xdp_statistics stats = {};
844 		bool extra_stats = true;
845 		size_t stats_size;
846 
847 		if (len < sizeof(struct xdp_statistics_v1)) {
848 			return -EINVAL;
849 		} else if (len < sizeof(stats)) {
850 			extra_stats = false;
851 			stats_size = sizeof(struct xdp_statistics_v1);
852 		} else {
853 			stats_size = sizeof(stats);
854 		}
855 
856 		mutex_lock(&xs->mutex);
857 		stats.rx_dropped = xs->rx_dropped;
858 		if (extra_stats) {
859 			stats.rx_ring_full = xs->rx_queue_full;
860 			stats.rx_fill_ring_empty_descs =
861 				xs->umem ? xskq_nb_queue_empty_descs(xs->umem->fq) : 0;
862 			stats.tx_ring_empty_descs = xskq_nb_queue_empty_descs(xs->tx);
863 		} else {
864 			stats.rx_dropped += xs->rx_queue_full;
865 		}
866 		stats.rx_invalid_descs = xskq_nb_invalid_descs(xs->rx);
867 		stats.tx_invalid_descs = xskq_nb_invalid_descs(xs->tx);
868 		mutex_unlock(&xs->mutex);
869 
870 		if (copy_to_user(optval, &stats, stats_size))
871 			return -EFAULT;
872 		if (put_user(stats_size, optlen))
873 			return -EFAULT;
874 
875 		return 0;
876 	}
877 	case XDP_MMAP_OFFSETS:
878 	{
879 		struct xdp_mmap_offsets off;
880 		struct xdp_mmap_offsets_v1 off_v1;
881 		bool flags_supported = true;
882 		void *to_copy;
883 
884 		if (len < sizeof(off_v1))
885 			return -EINVAL;
886 		else if (len < sizeof(off))
887 			flags_supported = false;
888 
889 		if (flags_supported) {
890 			/* xdp_ring_offset is identical to xdp_ring_offset_v1
891 			 * except for the flags field added to the end.
892 			 */
893 			xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *)
894 					       &off.rx);
895 			xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *)
896 					       &off.tx);
897 			xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *)
898 					       &off.fr);
899 			xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *)
900 					       &off.cr);
901 			off.rx.flags = offsetof(struct xdp_rxtx_ring,
902 						ptrs.flags);
903 			off.tx.flags = offsetof(struct xdp_rxtx_ring,
904 						ptrs.flags);
905 			off.fr.flags = offsetof(struct xdp_umem_ring,
906 						ptrs.flags);
907 			off.cr.flags = offsetof(struct xdp_umem_ring,
908 						ptrs.flags);
909 
910 			len = sizeof(off);
911 			to_copy = &off;
912 		} else {
913 			xsk_enter_rxtx_offsets(&off_v1.rx);
914 			xsk_enter_rxtx_offsets(&off_v1.tx);
915 			xsk_enter_umem_offsets(&off_v1.fr);
916 			xsk_enter_umem_offsets(&off_v1.cr);
917 
918 			len = sizeof(off_v1);
919 			to_copy = &off_v1;
920 		}
921 
922 		if (copy_to_user(optval, to_copy, len))
923 			return -EFAULT;
924 		if (put_user(len, optlen))
925 			return -EFAULT;
926 
927 		return 0;
928 	}
929 	case XDP_OPTIONS:
930 	{
931 		struct xdp_options opts = {};
932 
933 		if (len < sizeof(opts))
934 			return -EINVAL;
935 
936 		mutex_lock(&xs->mutex);
937 		if (xs->zc)
938 			opts.flags |= XDP_OPTIONS_ZEROCOPY;
939 		mutex_unlock(&xs->mutex);
940 
941 		len = sizeof(opts);
942 		if (copy_to_user(optval, &opts, len))
943 			return -EFAULT;
944 		if (put_user(len, optlen))
945 			return -EFAULT;
946 
947 		return 0;
948 	}
949 	default:
950 		break;
951 	}
952 
953 	return -EOPNOTSUPP;
954 }
955 
956 static int xsk_mmap(struct file *file, struct socket *sock,
957 		    struct vm_area_struct *vma)
958 {
959 	loff_t offset = (loff_t)vma->vm_pgoff << PAGE_SHIFT;
960 	unsigned long size = vma->vm_end - vma->vm_start;
961 	struct xdp_sock *xs = xdp_sk(sock->sk);
962 	struct xsk_queue *q = NULL;
963 	struct xdp_umem *umem;
964 	unsigned long pfn;
965 	struct page *qpg;
966 
967 	if (READ_ONCE(xs->state) != XSK_READY)
968 		return -EBUSY;
969 
970 	if (offset == XDP_PGOFF_RX_RING) {
971 		q = READ_ONCE(xs->rx);
972 	} else if (offset == XDP_PGOFF_TX_RING) {
973 		q = READ_ONCE(xs->tx);
974 	} else {
975 		umem = READ_ONCE(xs->umem);
976 		if (!umem)
977 			return -EINVAL;
978 
979 		/* Matches the smp_wmb() in XDP_UMEM_REG */
980 		smp_rmb();
981 		if (offset == XDP_UMEM_PGOFF_FILL_RING)
982 			q = READ_ONCE(umem->fq);
983 		else if (offset == XDP_UMEM_PGOFF_COMPLETION_RING)
984 			q = READ_ONCE(umem->cq);
985 	}
986 
987 	if (!q)
988 		return -EINVAL;
989 
990 	/* Matches the smp_wmb() in xsk_init_queue */
991 	smp_rmb();
992 	qpg = virt_to_head_page(q->ring);
993 	if (size > page_size(qpg))
994 		return -EINVAL;
995 
996 	pfn = virt_to_phys(q->ring) >> PAGE_SHIFT;
997 	return remap_pfn_range(vma, vma->vm_start, pfn,
998 			       size, vma->vm_page_prot);
999 }
1000 
1001 static int xsk_notifier(struct notifier_block *this,
1002 			unsigned long msg, void *ptr)
1003 {
1004 	struct net_device *dev = netdev_notifier_info_to_dev(ptr);
1005 	struct net *net = dev_net(dev);
1006 	struct sock *sk;
1007 
1008 	switch (msg) {
1009 	case NETDEV_UNREGISTER:
1010 		mutex_lock(&net->xdp.lock);
1011 		sk_for_each(sk, &net->xdp.list) {
1012 			struct xdp_sock *xs = xdp_sk(sk);
1013 
1014 			mutex_lock(&xs->mutex);
1015 			if (xs->dev == dev) {
1016 				sk->sk_err = ENETDOWN;
1017 				if (!sock_flag(sk, SOCK_DEAD))
1018 					sk->sk_error_report(sk);
1019 
1020 				xsk_unbind_dev(xs);
1021 
1022 				/* Clear device references in umem. */
1023 				xdp_umem_clear_dev(xs->umem);
1024 			}
1025 			mutex_unlock(&xs->mutex);
1026 		}
1027 		mutex_unlock(&net->xdp.lock);
1028 		break;
1029 	}
1030 	return NOTIFY_DONE;
1031 }
1032 
1033 static struct proto xsk_proto = {
1034 	.name =		"XDP",
1035 	.owner =	THIS_MODULE,
1036 	.obj_size =	sizeof(struct xdp_sock),
1037 };
1038 
1039 static const struct proto_ops xsk_proto_ops = {
1040 	.family		= PF_XDP,
1041 	.owner		= THIS_MODULE,
1042 	.release	= xsk_release,
1043 	.bind		= xsk_bind,
1044 	.connect	= sock_no_connect,
1045 	.socketpair	= sock_no_socketpair,
1046 	.accept		= sock_no_accept,
1047 	.getname	= sock_no_getname,
1048 	.poll		= xsk_poll,
1049 	.ioctl		= sock_no_ioctl,
1050 	.listen		= sock_no_listen,
1051 	.shutdown	= sock_no_shutdown,
1052 	.setsockopt	= xsk_setsockopt,
1053 	.getsockopt	= xsk_getsockopt,
1054 	.sendmsg	= xsk_sendmsg,
1055 	.recvmsg	= sock_no_recvmsg,
1056 	.mmap		= xsk_mmap,
1057 	.sendpage	= sock_no_sendpage,
1058 };
1059 
1060 static void xsk_destruct(struct sock *sk)
1061 {
1062 	struct xdp_sock *xs = xdp_sk(sk);
1063 
1064 	if (!sock_flag(sk, SOCK_DEAD))
1065 		return;
1066 
1067 	xdp_put_umem(xs->umem);
1068 
1069 	sk_refcnt_debug_dec(sk);
1070 }
1071 
1072 static int xsk_create(struct net *net, struct socket *sock, int protocol,
1073 		      int kern)
1074 {
1075 	struct sock *sk;
1076 	struct xdp_sock *xs;
1077 
1078 	if (!ns_capable(net->user_ns, CAP_NET_RAW))
1079 		return -EPERM;
1080 	if (sock->type != SOCK_RAW)
1081 		return -ESOCKTNOSUPPORT;
1082 
1083 	if (protocol)
1084 		return -EPROTONOSUPPORT;
1085 
1086 	sock->state = SS_UNCONNECTED;
1087 
1088 	sk = sk_alloc(net, PF_XDP, GFP_KERNEL, &xsk_proto, kern);
1089 	if (!sk)
1090 		return -ENOBUFS;
1091 
1092 	sock->ops = &xsk_proto_ops;
1093 
1094 	sock_init_data(sock, sk);
1095 
1096 	sk->sk_family = PF_XDP;
1097 
1098 	sk->sk_destruct = xsk_destruct;
1099 	sk_refcnt_debug_inc(sk);
1100 
1101 	sock_set_flag(sk, SOCK_RCU_FREE);
1102 
1103 	xs = xdp_sk(sk);
1104 	xs->state = XSK_READY;
1105 	mutex_init(&xs->mutex);
1106 	spin_lock_init(&xs->rx_lock);
1107 	spin_lock_init(&xs->tx_completion_lock);
1108 
1109 	INIT_LIST_HEAD(&xs->map_list);
1110 	spin_lock_init(&xs->map_list_lock);
1111 
1112 	mutex_lock(&net->xdp.lock);
1113 	sk_add_node_rcu(sk, &net->xdp.list);
1114 	mutex_unlock(&net->xdp.lock);
1115 
1116 	local_bh_disable();
1117 	sock_prot_inuse_add(net, &xsk_proto, 1);
1118 	local_bh_enable();
1119 
1120 	return 0;
1121 }
1122 
1123 static const struct net_proto_family xsk_family_ops = {
1124 	.family = PF_XDP,
1125 	.create = xsk_create,
1126 	.owner	= THIS_MODULE,
1127 };
1128 
1129 static struct notifier_block xsk_netdev_notifier = {
1130 	.notifier_call	= xsk_notifier,
1131 };
1132 
1133 static int __net_init xsk_net_init(struct net *net)
1134 {
1135 	mutex_init(&net->xdp.lock);
1136 	INIT_HLIST_HEAD(&net->xdp.list);
1137 	return 0;
1138 }
1139 
1140 static void __net_exit xsk_net_exit(struct net *net)
1141 {
1142 	WARN_ON_ONCE(!hlist_empty(&net->xdp.list));
1143 }
1144 
1145 static struct pernet_operations xsk_net_ops = {
1146 	.init = xsk_net_init,
1147 	.exit = xsk_net_exit,
1148 };
1149 
1150 static int __init xsk_init(void)
1151 {
1152 	int err, cpu;
1153 
1154 	err = proto_register(&xsk_proto, 0 /* no slab */);
1155 	if (err)
1156 		goto out;
1157 
1158 	err = sock_register(&xsk_family_ops);
1159 	if (err)
1160 		goto out_proto;
1161 
1162 	err = register_pernet_subsys(&xsk_net_ops);
1163 	if (err)
1164 		goto out_sk;
1165 
1166 	err = register_netdevice_notifier(&xsk_netdev_notifier);
1167 	if (err)
1168 		goto out_pernet;
1169 
1170 	for_each_possible_cpu(cpu)
1171 		INIT_LIST_HEAD(&per_cpu(xskmap_flush_list, cpu));
1172 	return 0;
1173 
1174 out_pernet:
1175 	unregister_pernet_subsys(&xsk_net_ops);
1176 out_sk:
1177 	sock_unregister(PF_XDP);
1178 out_proto:
1179 	proto_unregister(&xsk_proto);
1180 out:
1181 	return err;
1182 }
1183 
1184 fs_initcall(xsk_init);
1185