xref: /openbmc/linux/net/xdp/xsk.c (revision dc023681)
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/busy_poll.h>
27 #include <net/xdp.h>
28 
29 #include "xsk_queue.h"
30 #include "xdp_umem.h"
31 #include "xsk.h"
32 
33 #define TX_BATCH_SIZE 32
34 
35 static DEFINE_PER_CPU(struct list_head, xskmap_flush_list);
36 
37 void xsk_set_rx_need_wakeup(struct xsk_buff_pool *pool)
38 {
39 	if (pool->cached_need_wakeup & XDP_WAKEUP_RX)
40 		return;
41 
42 	pool->fq->ring->flags |= XDP_RING_NEED_WAKEUP;
43 	pool->cached_need_wakeup |= XDP_WAKEUP_RX;
44 }
45 EXPORT_SYMBOL(xsk_set_rx_need_wakeup);
46 
47 void xsk_set_tx_need_wakeup(struct xsk_buff_pool *pool)
48 {
49 	struct xdp_sock *xs;
50 
51 	if (pool->cached_need_wakeup & XDP_WAKEUP_TX)
52 		return;
53 
54 	rcu_read_lock();
55 	list_for_each_entry_rcu(xs, &pool->xsk_tx_list, tx_list) {
56 		xs->tx->ring->flags |= XDP_RING_NEED_WAKEUP;
57 	}
58 	rcu_read_unlock();
59 
60 	pool->cached_need_wakeup |= XDP_WAKEUP_TX;
61 }
62 EXPORT_SYMBOL(xsk_set_tx_need_wakeup);
63 
64 void xsk_clear_rx_need_wakeup(struct xsk_buff_pool *pool)
65 {
66 	if (!(pool->cached_need_wakeup & XDP_WAKEUP_RX))
67 		return;
68 
69 	pool->fq->ring->flags &= ~XDP_RING_NEED_WAKEUP;
70 	pool->cached_need_wakeup &= ~XDP_WAKEUP_RX;
71 }
72 EXPORT_SYMBOL(xsk_clear_rx_need_wakeup);
73 
74 void xsk_clear_tx_need_wakeup(struct xsk_buff_pool *pool)
75 {
76 	struct xdp_sock *xs;
77 
78 	if (!(pool->cached_need_wakeup & XDP_WAKEUP_TX))
79 		return;
80 
81 	rcu_read_lock();
82 	list_for_each_entry_rcu(xs, &pool->xsk_tx_list, tx_list) {
83 		xs->tx->ring->flags &= ~XDP_RING_NEED_WAKEUP;
84 	}
85 	rcu_read_unlock();
86 
87 	pool->cached_need_wakeup &= ~XDP_WAKEUP_TX;
88 }
89 EXPORT_SYMBOL(xsk_clear_tx_need_wakeup);
90 
91 bool xsk_uses_need_wakeup(struct xsk_buff_pool *pool)
92 {
93 	return pool->uses_need_wakeup;
94 }
95 EXPORT_SYMBOL(xsk_uses_need_wakeup);
96 
97 struct xsk_buff_pool *xsk_get_pool_from_qid(struct net_device *dev,
98 					    u16 queue_id)
99 {
100 	if (queue_id < dev->real_num_rx_queues)
101 		return dev->_rx[queue_id].pool;
102 	if (queue_id < dev->real_num_tx_queues)
103 		return dev->_tx[queue_id].pool;
104 
105 	return NULL;
106 }
107 EXPORT_SYMBOL(xsk_get_pool_from_qid);
108 
109 void xsk_clear_pool_at_qid(struct net_device *dev, u16 queue_id)
110 {
111 	if (queue_id < dev->num_rx_queues)
112 		dev->_rx[queue_id].pool = NULL;
113 	if (queue_id < dev->num_tx_queues)
114 		dev->_tx[queue_id].pool = NULL;
115 }
116 
117 /* The buffer pool is stored both in the _rx struct and the _tx struct as we do
118  * not know if the device has more tx queues than rx, or the opposite.
119  * This might also change during run time.
120  */
121 int xsk_reg_pool_at_qid(struct net_device *dev, struct xsk_buff_pool *pool,
122 			u16 queue_id)
123 {
124 	if (queue_id >= max_t(unsigned int,
125 			      dev->real_num_rx_queues,
126 			      dev->real_num_tx_queues))
127 		return -EINVAL;
128 
129 	if (queue_id < dev->real_num_rx_queues)
130 		dev->_rx[queue_id].pool = pool;
131 	if (queue_id < dev->real_num_tx_queues)
132 		dev->_tx[queue_id].pool = pool;
133 
134 	return 0;
135 }
136 
137 void xp_release(struct xdp_buff_xsk *xskb)
138 {
139 	xskb->pool->free_heads[xskb->pool->free_heads_cnt++] = xskb;
140 }
141 
142 static u64 xp_get_handle(struct xdp_buff_xsk *xskb)
143 {
144 	u64 offset = xskb->xdp.data - xskb->xdp.data_hard_start;
145 
146 	offset += xskb->pool->headroom;
147 	if (!xskb->pool->unaligned)
148 		return xskb->orig_addr + offset;
149 	return xskb->orig_addr + (offset << XSK_UNALIGNED_BUF_OFFSET_SHIFT);
150 }
151 
152 static int __xsk_rcv_zc(struct xdp_sock *xs, struct xdp_buff *xdp, u32 len)
153 {
154 	struct xdp_buff_xsk *xskb = container_of(xdp, struct xdp_buff_xsk, xdp);
155 	u64 addr;
156 	int err;
157 
158 	addr = xp_get_handle(xskb);
159 	err = xskq_prod_reserve_desc(xs->rx, addr, len);
160 	if (err) {
161 		xs->rx_queue_full++;
162 		return err;
163 	}
164 
165 	xp_release(xskb);
166 	return 0;
167 }
168 
169 static void xsk_copy_xdp(struct xdp_buff *to, struct xdp_buff *from, u32 len)
170 {
171 	void *from_buf, *to_buf;
172 	u32 metalen;
173 
174 	if (unlikely(xdp_data_meta_unsupported(from))) {
175 		from_buf = from->data;
176 		to_buf = to->data;
177 		metalen = 0;
178 	} else {
179 		from_buf = from->data_meta;
180 		metalen = from->data - from->data_meta;
181 		to_buf = to->data - metalen;
182 	}
183 
184 	memcpy(to_buf, from_buf, len + metalen);
185 }
186 
187 static int __xsk_rcv(struct xdp_sock *xs, struct xdp_buff *xdp)
188 {
189 	struct xdp_buff *xsk_xdp;
190 	int err;
191 	u32 len;
192 
193 	len = xdp->data_end - xdp->data;
194 	if (len > xsk_pool_get_rx_frame_size(xs->pool)) {
195 		xs->rx_dropped++;
196 		return -ENOSPC;
197 	}
198 
199 	xsk_xdp = xsk_buff_alloc(xs->pool);
200 	if (!xsk_xdp) {
201 		xs->rx_dropped++;
202 		return -ENOSPC;
203 	}
204 
205 	xsk_copy_xdp(xsk_xdp, xdp, len);
206 	err = __xsk_rcv_zc(xs, xsk_xdp, len);
207 	if (err) {
208 		xsk_buff_free(xsk_xdp);
209 		return err;
210 	}
211 	return 0;
212 }
213 
214 static bool xsk_tx_writeable(struct xdp_sock *xs)
215 {
216 	if (xskq_cons_present_entries(xs->tx) > xs->tx->nentries / 2)
217 		return false;
218 
219 	return true;
220 }
221 
222 static bool xsk_is_bound(struct xdp_sock *xs)
223 {
224 	if (READ_ONCE(xs->state) == XSK_BOUND) {
225 		/* Matches smp_wmb() in bind(). */
226 		smp_rmb();
227 		return true;
228 	}
229 	return false;
230 }
231 
232 static int xsk_rcv_check(struct xdp_sock *xs, struct xdp_buff *xdp)
233 {
234 	if (!xsk_is_bound(xs))
235 		return -EINVAL;
236 
237 	if (xs->dev != xdp->rxq->dev || xs->queue_id != xdp->rxq->queue_index)
238 		return -EINVAL;
239 
240 	sk_mark_napi_id_once_xdp(&xs->sk, xdp);
241 	return 0;
242 }
243 
244 static void xsk_flush(struct xdp_sock *xs)
245 {
246 	xskq_prod_submit(xs->rx);
247 	__xskq_cons_release(xs->pool->fq);
248 	sock_def_readable(&xs->sk);
249 }
250 
251 int xsk_generic_rcv(struct xdp_sock *xs, struct xdp_buff *xdp)
252 {
253 	int err;
254 
255 	spin_lock_bh(&xs->rx_lock);
256 	err = xsk_rcv_check(xs, xdp);
257 	if (!err) {
258 		err = __xsk_rcv(xs, xdp);
259 		xsk_flush(xs);
260 	}
261 	spin_unlock_bh(&xs->rx_lock);
262 	return err;
263 }
264 
265 static int xsk_rcv(struct xdp_sock *xs, struct xdp_buff *xdp)
266 {
267 	int err;
268 	u32 len;
269 
270 	err = xsk_rcv_check(xs, xdp);
271 	if (err)
272 		return err;
273 
274 	if (xdp->rxq->mem.type == MEM_TYPE_XSK_BUFF_POOL) {
275 		len = xdp->data_end - xdp->data;
276 		return __xsk_rcv_zc(xs, xdp, len);
277 	}
278 
279 	err = __xsk_rcv(xs, xdp);
280 	if (!err)
281 		xdp_return_buff(xdp);
282 	return err;
283 }
284 
285 int __xsk_map_redirect(struct xdp_sock *xs, struct xdp_buff *xdp)
286 {
287 	struct list_head *flush_list = this_cpu_ptr(&xskmap_flush_list);
288 	int err;
289 
290 	err = xsk_rcv(xs, xdp);
291 	if (err)
292 		return err;
293 
294 	if (!xs->flush_node.prev)
295 		list_add(&xs->flush_node, flush_list);
296 
297 	return 0;
298 }
299 
300 void __xsk_map_flush(void)
301 {
302 	struct list_head *flush_list = this_cpu_ptr(&xskmap_flush_list);
303 	struct xdp_sock *xs, *tmp;
304 
305 	list_for_each_entry_safe(xs, tmp, flush_list, flush_node) {
306 		xsk_flush(xs);
307 		__list_del_clearprev(&xs->flush_node);
308 	}
309 }
310 
311 void xsk_tx_completed(struct xsk_buff_pool *pool, u32 nb_entries)
312 {
313 	xskq_prod_submit_n(pool->cq, nb_entries);
314 }
315 EXPORT_SYMBOL(xsk_tx_completed);
316 
317 void xsk_tx_release(struct xsk_buff_pool *pool)
318 {
319 	struct xdp_sock *xs;
320 
321 	rcu_read_lock();
322 	list_for_each_entry_rcu(xs, &pool->xsk_tx_list, tx_list) {
323 		__xskq_cons_release(xs->tx);
324 		if (xsk_tx_writeable(xs))
325 			xs->sk.sk_write_space(&xs->sk);
326 	}
327 	rcu_read_unlock();
328 }
329 EXPORT_SYMBOL(xsk_tx_release);
330 
331 bool xsk_tx_peek_desc(struct xsk_buff_pool *pool, struct xdp_desc *desc)
332 {
333 	struct xdp_sock *xs;
334 
335 	rcu_read_lock();
336 	list_for_each_entry_rcu(xs, &pool->xsk_tx_list, tx_list) {
337 		if (!xskq_cons_peek_desc(xs->tx, desc, pool)) {
338 			xs->tx->queue_empty_descs++;
339 			continue;
340 		}
341 
342 		/* This is the backpressure mechanism for the Tx path.
343 		 * Reserve space in the completion queue and only proceed
344 		 * if there is space in it. This avoids having to implement
345 		 * any buffering in the Tx path.
346 		 */
347 		if (xskq_prod_reserve_addr(pool->cq, desc->addr))
348 			goto out;
349 
350 		xskq_cons_release(xs->tx);
351 		rcu_read_unlock();
352 		return true;
353 	}
354 
355 out:
356 	rcu_read_unlock();
357 	return false;
358 }
359 EXPORT_SYMBOL(xsk_tx_peek_desc);
360 
361 static u32 xsk_tx_peek_release_fallback(struct xsk_buff_pool *pool, struct xdp_desc *descs,
362 					u32 max_entries)
363 {
364 	u32 nb_pkts = 0;
365 
366 	while (nb_pkts < max_entries && xsk_tx_peek_desc(pool, &descs[nb_pkts]))
367 		nb_pkts++;
368 
369 	xsk_tx_release(pool);
370 	return nb_pkts;
371 }
372 
373 u32 xsk_tx_peek_release_desc_batch(struct xsk_buff_pool *pool, struct xdp_desc *descs,
374 				   u32 max_entries)
375 {
376 	struct xdp_sock *xs;
377 	u32 nb_pkts;
378 
379 	rcu_read_lock();
380 	if (!list_is_singular(&pool->xsk_tx_list)) {
381 		/* Fallback to the non-batched version */
382 		rcu_read_unlock();
383 		return xsk_tx_peek_release_fallback(pool, descs, max_entries);
384 	}
385 
386 	xs = list_first_or_null_rcu(&pool->xsk_tx_list, struct xdp_sock, tx_list);
387 	if (!xs) {
388 		nb_pkts = 0;
389 		goto out;
390 	}
391 
392 	nb_pkts = xskq_cons_peek_desc_batch(xs->tx, descs, pool, max_entries);
393 	if (!nb_pkts) {
394 		xs->tx->queue_empty_descs++;
395 		goto out;
396 	}
397 
398 	/* This is the backpressure mechanism for the Tx path. Try to
399 	 * reserve space in the completion queue for all packets, but
400 	 * if there are fewer slots available, just process that many
401 	 * packets. This avoids having to implement any buffering in
402 	 * the Tx path.
403 	 */
404 	nb_pkts = xskq_prod_reserve_addr_batch(pool->cq, descs, nb_pkts);
405 	if (!nb_pkts)
406 		goto out;
407 
408 	xskq_cons_release_n(xs->tx, nb_pkts);
409 	__xskq_cons_release(xs->tx);
410 	xs->sk.sk_write_space(&xs->sk);
411 
412 out:
413 	rcu_read_unlock();
414 	return nb_pkts;
415 }
416 EXPORT_SYMBOL(xsk_tx_peek_release_desc_batch);
417 
418 static int xsk_wakeup(struct xdp_sock *xs, u8 flags)
419 {
420 	struct net_device *dev = xs->dev;
421 	int err;
422 
423 	rcu_read_lock();
424 	err = dev->netdev_ops->ndo_xsk_wakeup(dev, xs->queue_id, flags);
425 	rcu_read_unlock();
426 
427 	return err;
428 }
429 
430 static int xsk_zc_xmit(struct xdp_sock *xs)
431 {
432 	return xsk_wakeup(xs, XDP_WAKEUP_TX);
433 }
434 
435 static void xsk_destruct_skb(struct sk_buff *skb)
436 {
437 	u64 addr = (u64)(long)skb_shinfo(skb)->destructor_arg;
438 	struct xdp_sock *xs = xdp_sk(skb->sk);
439 	unsigned long flags;
440 
441 	spin_lock_irqsave(&xs->pool->cq_lock, flags);
442 	xskq_prod_submit_addr(xs->pool->cq, addr);
443 	spin_unlock_irqrestore(&xs->pool->cq_lock, flags);
444 
445 	sock_wfree(skb);
446 }
447 
448 static struct sk_buff *xsk_build_skb_zerocopy(struct xdp_sock *xs,
449 					      struct xdp_desc *desc)
450 {
451 	struct xsk_buff_pool *pool = xs->pool;
452 	u32 hr, len, ts, offset, copy, copied;
453 	struct sk_buff *skb;
454 	struct page *page;
455 	void *buffer;
456 	int err, i;
457 	u64 addr;
458 
459 	hr = max(NET_SKB_PAD, L1_CACHE_ALIGN(xs->dev->needed_headroom));
460 
461 	skb = sock_alloc_send_skb(&xs->sk, hr, 1, &err);
462 	if (unlikely(!skb))
463 		return ERR_PTR(err);
464 
465 	skb_reserve(skb, hr);
466 
467 	addr = desc->addr;
468 	len = desc->len;
469 	ts = pool->unaligned ? len : pool->chunk_size;
470 
471 	buffer = xsk_buff_raw_get_data(pool, addr);
472 	offset = offset_in_page(buffer);
473 	addr = buffer - pool->addrs;
474 
475 	for (copied = 0, i = 0; copied < len; i++) {
476 		page = pool->umem->pgs[addr >> PAGE_SHIFT];
477 		get_page(page);
478 
479 		copy = min_t(u32, PAGE_SIZE - offset, len - copied);
480 		skb_fill_page_desc(skb, i, page, offset, copy);
481 
482 		copied += copy;
483 		addr += copy;
484 		offset = 0;
485 	}
486 
487 	skb->len += len;
488 	skb->data_len += len;
489 	skb->truesize += ts;
490 
491 	refcount_add(ts, &xs->sk.sk_wmem_alloc);
492 
493 	return skb;
494 }
495 
496 static struct sk_buff *xsk_build_skb(struct xdp_sock *xs,
497 				     struct xdp_desc *desc)
498 {
499 	struct net_device *dev = xs->dev;
500 	struct sk_buff *skb;
501 
502 	if (dev->priv_flags & IFF_TX_SKB_NO_LINEAR) {
503 		skb = xsk_build_skb_zerocopy(xs, desc);
504 		if (IS_ERR(skb))
505 			return skb;
506 	} else {
507 		u32 hr, tr, len;
508 		void *buffer;
509 		int err;
510 
511 		hr = max(NET_SKB_PAD, L1_CACHE_ALIGN(dev->needed_headroom));
512 		tr = dev->needed_tailroom;
513 		len = desc->len;
514 
515 		skb = sock_alloc_send_skb(&xs->sk, hr + len + tr, 1, &err);
516 		if (unlikely(!skb))
517 			return ERR_PTR(err);
518 
519 		skb_reserve(skb, hr);
520 		skb_put(skb, len);
521 
522 		buffer = xsk_buff_raw_get_data(xs->pool, desc->addr);
523 		err = skb_store_bits(skb, 0, buffer, len);
524 		if (unlikely(err)) {
525 			kfree_skb(skb);
526 			return ERR_PTR(err);
527 		}
528 	}
529 
530 	skb->dev = dev;
531 	skb->priority = xs->sk.sk_priority;
532 	skb->mark = xs->sk.sk_mark;
533 	skb_shinfo(skb)->destructor_arg = (void *)(long)desc->addr;
534 	skb->destructor = xsk_destruct_skb;
535 
536 	return skb;
537 }
538 
539 static int xsk_generic_xmit(struct sock *sk)
540 {
541 	struct xdp_sock *xs = xdp_sk(sk);
542 	u32 max_batch = TX_BATCH_SIZE;
543 	bool sent_frame = false;
544 	struct xdp_desc desc;
545 	struct sk_buff *skb;
546 	unsigned long flags;
547 	int err = 0;
548 
549 	mutex_lock(&xs->mutex);
550 
551 	if (xs->queue_id >= xs->dev->real_num_tx_queues)
552 		goto out;
553 
554 	while (xskq_cons_peek_desc(xs->tx, &desc, xs->pool)) {
555 		if (max_batch-- == 0) {
556 			err = -EAGAIN;
557 			goto out;
558 		}
559 
560 		skb = xsk_build_skb(xs, &desc);
561 		if (IS_ERR(skb)) {
562 			err = PTR_ERR(skb);
563 			goto out;
564 		}
565 
566 		/* This is the backpressure mechanism for the Tx path.
567 		 * Reserve space in the completion queue and only proceed
568 		 * if there is space in it. This avoids having to implement
569 		 * any buffering in the Tx path.
570 		 */
571 		spin_lock_irqsave(&xs->pool->cq_lock, flags);
572 		if (xskq_prod_reserve(xs->pool->cq)) {
573 			spin_unlock_irqrestore(&xs->pool->cq_lock, flags);
574 			kfree_skb(skb);
575 			goto out;
576 		}
577 		spin_unlock_irqrestore(&xs->pool->cq_lock, flags);
578 
579 		err = __dev_direct_xmit(skb, xs->queue_id);
580 		if  (err == NETDEV_TX_BUSY) {
581 			/* Tell user-space to retry the send */
582 			skb->destructor = sock_wfree;
583 			spin_lock_irqsave(&xs->pool->cq_lock, flags);
584 			xskq_prod_cancel(xs->pool->cq);
585 			spin_unlock_irqrestore(&xs->pool->cq_lock, flags);
586 			/* Free skb without triggering the perf drop trace */
587 			consume_skb(skb);
588 			err = -EAGAIN;
589 			goto out;
590 		}
591 
592 		xskq_cons_release(xs->tx);
593 		/* Ignore NET_XMIT_CN as packet might have been sent */
594 		if (err == NET_XMIT_DROP) {
595 			/* SKB completed but not sent */
596 			err = -EBUSY;
597 			goto out;
598 		}
599 
600 		sent_frame = true;
601 	}
602 
603 	xs->tx->queue_empty_descs++;
604 
605 out:
606 	if (sent_frame)
607 		if (xsk_tx_writeable(xs))
608 			sk->sk_write_space(sk);
609 
610 	mutex_unlock(&xs->mutex);
611 	return err;
612 }
613 
614 static int __xsk_sendmsg(struct sock *sk)
615 {
616 	struct xdp_sock *xs = xdp_sk(sk);
617 
618 	if (unlikely(!(xs->dev->flags & IFF_UP)))
619 		return -ENETDOWN;
620 	if (unlikely(!xs->tx))
621 		return -ENOBUFS;
622 
623 	return xs->zc ? xsk_zc_xmit(xs) : xsk_generic_xmit(sk);
624 }
625 
626 static bool xsk_no_wakeup(struct sock *sk)
627 {
628 #ifdef CONFIG_NET_RX_BUSY_POLL
629 	/* Prefer busy-polling, skip the wakeup. */
630 	return READ_ONCE(sk->sk_prefer_busy_poll) && READ_ONCE(sk->sk_ll_usec) &&
631 		READ_ONCE(sk->sk_napi_id) >= MIN_NAPI_ID;
632 #else
633 	return false;
634 #endif
635 }
636 
637 static int xsk_sendmsg(struct socket *sock, struct msghdr *m, size_t total_len)
638 {
639 	bool need_wait = !(m->msg_flags & MSG_DONTWAIT);
640 	struct sock *sk = sock->sk;
641 	struct xdp_sock *xs = xdp_sk(sk);
642 	struct xsk_buff_pool *pool;
643 
644 	if (unlikely(!xsk_is_bound(xs)))
645 		return -ENXIO;
646 	if (unlikely(need_wait))
647 		return -EOPNOTSUPP;
648 
649 	if (sk_can_busy_loop(sk))
650 		sk_busy_loop(sk, 1); /* only support non-blocking sockets */
651 
652 	if (xsk_no_wakeup(sk))
653 		return 0;
654 
655 	pool = xs->pool;
656 	if (pool->cached_need_wakeup & XDP_WAKEUP_TX)
657 		return __xsk_sendmsg(sk);
658 	return 0;
659 }
660 
661 static int xsk_recvmsg(struct socket *sock, struct msghdr *m, size_t len, int flags)
662 {
663 	bool need_wait = !(flags & MSG_DONTWAIT);
664 	struct sock *sk = sock->sk;
665 	struct xdp_sock *xs = xdp_sk(sk);
666 
667 	if (unlikely(!xsk_is_bound(xs)))
668 		return -ENXIO;
669 	if (unlikely(!(xs->dev->flags & IFF_UP)))
670 		return -ENETDOWN;
671 	if (unlikely(!xs->rx))
672 		return -ENOBUFS;
673 	if (unlikely(need_wait))
674 		return -EOPNOTSUPP;
675 
676 	if (sk_can_busy_loop(sk))
677 		sk_busy_loop(sk, 1); /* only support non-blocking sockets */
678 
679 	if (xsk_no_wakeup(sk))
680 		return 0;
681 
682 	if (xs->pool->cached_need_wakeup & XDP_WAKEUP_RX && xs->zc)
683 		return xsk_wakeup(xs, XDP_WAKEUP_RX);
684 	return 0;
685 }
686 
687 static __poll_t xsk_poll(struct file *file, struct socket *sock,
688 			     struct poll_table_struct *wait)
689 {
690 	__poll_t mask = 0;
691 	struct sock *sk = sock->sk;
692 	struct xdp_sock *xs = xdp_sk(sk);
693 	struct xsk_buff_pool *pool;
694 
695 	sock_poll_wait(file, sock, wait);
696 
697 	if (unlikely(!xsk_is_bound(xs)))
698 		return mask;
699 
700 	pool = xs->pool;
701 
702 	if (pool->cached_need_wakeup) {
703 		if (xs->zc)
704 			xsk_wakeup(xs, pool->cached_need_wakeup);
705 		else
706 			/* Poll needs to drive Tx also in copy mode */
707 			__xsk_sendmsg(sk);
708 	}
709 
710 	if (xs->rx && !xskq_prod_is_empty(xs->rx))
711 		mask |= EPOLLIN | EPOLLRDNORM;
712 	if (xs->tx && xsk_tx_writeable(xs))
713 		mask |= EPOLLOUT | EPOLLWRNORM;
714 
715 	return mask;
716 }
717 
718 static int xsk_init_queue(u32 entries, struct xsk_queue **queue,
719 			  bool umem_queue)
720 {
721 	struct xsk_queue *q;
722 
723 	if (entries == 0 || *queue || !is_power_of_2(entries))
724 		return -EINVAL;
725 
726 	q = xskq_create(entries, umem_queue);
727 	if (!q)
728 		return -ENOMEM;
729 
730 	/* Make sure queue is ready before it can be seen by others */
731 	smp_wmb();
732 	WRITE_ONCE(*queue, q);
733 	return 0;
734 }
735 
736 static void xsk_unbind_dev(struct xdp_sock *xs)
737 {
738 	struct net_device *dev = xs->dev;
739 
740 	if (xs->state != XSK_BOUND)
741 		return;
742 	WRITE_ONCE(xs->state, XSK_UNBOUND);
743 
744 	/* Wait for driver to stop using the xdp socket. */
745 	xp_del_xsk(xs->pool, xs);
746 	xs->dev = NULL;
747 	synchronize_net();
748 	dev_put(dev);
749 }
750 
751 static struct xsk_map *xsk_get_map_list_entry(struct xdp_sock *xs,
752 					      struct xdp_sock __rcu ***map_entry)
753 {
754 	struct xsk_map *map = NULL;
755 	struct xsk_map_node *node;
756 
757 	*map_entry = NULL;
758 
759 	spin_lock_bh(&xs->map_list_lock);
760 	node = list_first_entry_or_null(&xs->map_list, struct xsk_map_node,
761 					node);
762 	if (node) {
763 		bpf_map_inc(&node->map->map);
764 		map = node->map;
765 		*map_entry = node->map_entry;
766 	}
767 	spin_unlock_bh(&xs->map_list_lock);
768 	return map;
769 }
770 
771 static void xsk_delete_from_maps(struct xdp_sock *xs)
772 {
773 	/* This function removes the current XDP socket from all the
774 	 * maps it resides in. We need to take extra care here, due to
775 	 * the two locks involved. Each map has a lock synchronizing
776 	 * updates to the entries, and each socket has a lock that
777 	 * synchronizes access to the list of maps (map_list). For
778 	 * deadlock avoidance the locks need to be taken in the order
779 	 * "map lock"->"socket map list lock". We start off by
780 	 * accessing the socket map list, and take a reference to the
781 	 * map to guarantee existence between the
782 	 * xsk_get_map_list_entry() and xsk_map_try_sock_delete()
783 	 * calls. Then we ask the map to remove the socket, which
784 	 * tries to remove the socket from the map. Note that there
785 	 * might be updates to the map between
786 	 * xsk_get_map_list_entry() and xsk_map_try_sock_delete().
787 	 */
788 	struct xdp_sock __rcu **map_entry = NULL;
789 	struct xsk_map *map;
790 
791 	while ((map = xsk_get_map_list_entry(xs, &map_entry))) {
792 		xsk_map_try_sock_delete(map, xs, map_entry);
793 		bpf_map_put(&map->map);
794 	}
795 }
796 
797 static int xsk_release(struct socket *sock)
798 {
799 	struct sock *sk = sock->sk;
800 	struct xdp_sock *xs = xdp_sk(sk);
801 	struct net *net;
802 
803 	if (!sk)
804 		return 0;
805 
806 	net = sock_net(sk);
807 
808 	mutex_lock(&net->xdp.lock);
809 	sk_del_node_init_rcu(sk);
810 	mutex_unlock(&net->xdp.lock);
811 
812 	local_bh_disable();
813 	sock_prot_inuse_add(net, sk->sk_prot, -1);
814 	local_bh_enable();
815 
816 	xsk_delete_from_maps(xs);
817 	mutex_lock(&xs->mutex);
818 	xsk_unbind_dev(xs);
819 	mutex_unlock(&xs->mutex);
820 
821 	xskq_destroy(xs->rx);
822 	xskq_destroy(xs->tx);
823 	xskq_destroy(xs->fq_tmp);
824 	xskq_destroy(xs->cq_tmp);
825 
826 	sock_orphan(sk);
827 	sock->sk = NULL;
828 
829 	sk_refcnt_debug_release(sk);
830 	sock_put(sk);
831 
832 	return 0;
833 }
834 
835 static struct socket *xsk_lookup_xsk_from_fd(int fd)
836 {
837 	struct socket *sock;
838 	int err;
839 
840 	sock = sockfd_lookup(fd, &err);
841 	if (!sock)
842 		return ERR_PTR(-ENOTSOCK);
843 
844 	if (sock->sk->sk_family != PF_XDP) {
845 		sockfd_put(sock);
846 		return ERR_PTR(-ENOPROTOOPT);
847 	}
848 
849 	return sock;
850 }
851 
852 static bool xsk_validate_queues(struct xdp_sock *xs)
853 {
854 	return xs->fq_tmp && xs->cq_tmp;
855 }
856 
857 static int xsk_bind(struct socket *sock, struct sockaddr *addr, int addr_len)
858 {
859 	struct sockaddr_xdp *sxdp = (struct sockaddr_xdp *)addr;
860 	struct sock *sk = sock->sk;
861 	struct xdp_sock *xs = xdp_sk(sk);
862 	struct net_device *dev;
863 	u32 flags, qid;
864 	int err = 0;
865 
866 	if (addr_len < sizeof(struct sockaddr_xdp))
867 		return -EINVAL;
868 	if (sxdp->sxdp_family != AF_XDP)
869 		return -EINVAL;
870 
871 	flags = sxdp->sxdp_flags;
872 	if (flags & ~(XDP_SHARED_UMEM | XDP_COPY | XDP_ZEROCOPY |
873 		      XDP_USE_NEED_WAKEUP))
874 		return -EINVAL;
875 
876 	rtnl_lock();
877 	mutex_lock(&xs->mutex);
878 	if (xs->state != XSK_READY) {
879 		err = -EBUSY;
880 		goto out_release;
881 	}
882 
883 	dev = dev_get_by_index(sock_net(sk), sxdp->sxdp_ifindex);
884 	if (!dev) {
885 		err = -ENODEV;
886 		goto out_release;
887 	}
888 
889 	if (!xs->rx && !xs->tx) {
890 		err = -EINVAL;
891 		goto out_unlock;
892 	}
893 
894 	qid = sxdp->sxdp_queue_id;
895 
896 	if (flags & XDP_SHARED_UMEM) {
897 		struct xdp_sock *umem_xs;
898 		struct socket *sock;
899 
900 		if ((flags & XDP_COPY) || (flags & XDP_ZEROCOPY) ||
901 		    (flags & XDP_USE_NEED_WAKEUP)) {
902 			/* Cannot specify flags for shared sockets. */
903 			err = -EINVAL;
904 			goto out_unlock;
905 		}
906 
907 		if (xs->umem) {
908 			/* We have already our own. */
909 			err = -EINVAL;
910 			goto out_unlock;
911 		}
912 
913 		sock = xsk_lookup_xsk_from_fd(sxdp->sxdp_shared_umem_fd);
914 		if (IS_ERR(sock)) {
915 			err = PTR_ERR(sock);
916 			goto out_unlock;
917 		}
918 
919 		umem_xs = xdp_sk(sock->sk);
920 		if (!xsk_is_bound(umem_xs)) {
921 			err = -EBADF;
922 			sockfd_put(sock);
923 			goto out_unlock;
924 		}
925 
926 		if (umem_xs->queue_id != qid || umem_xs->dev != dev) {
927 			/* Share the umem with another socket on another qid
928 			 * and/or device.
929 			 */
930 			xs->pool = xp_create_and_assign_umem(xs,
931 							     umem_xs->umem);
932 			if (!xs->pool) {
933 				err = -ENOMEM;
934 				sockfd_put(sock);
935 				goto out_unlock;
936 			}
937 
938 			err = xp_assign_dev_shared(xs->pool, umem_xs->umem,
939 						   dev, qid);
940 			if (err) {
941 				xp_destroy(xs->pool);
942 				xs->pool = NULL;
943 				sockfd_put(sock);
944 				goto out_unlock;
945 			}
946 		} else {
947 			/* Share the buffer pool with the other socket. */
948 			if (xs->fq_tmp || xs->cq_tmp) {
949 				/* Do not allow setting your own fq or cq. */
950 				err = -EINVAL;
951 				sockfd_put(sock);
952 				goto out_unlock;
953 			}
954 
955 			xp_get_pool(umem_xs->pool);
956 			xs->pool = umem_xs->pool;
957 		}
958 
959 		xdp_get_umem(umem_xs->umem);
960 		WRITE_ONCE(xs->umem, umem_xs->umem);
961 		sockfd_put(sock);
962 	} else if (!xs->umem || !xsk_validate_queues(xs)) {
963 		err = -EINVAL;
964 		goto out_unlock;
965 	} else {
966 		/* This xsk has its own umem. */
967 		xs->pool = xp_create_and_assign_umem(xs, xs->umem);
968 		if (!xs->pool) {
969 			err = -ENOMEM;
970 			goto out_unlock;
971 		}
972 
973 		err = xp_assign_dev(xs->pool, dev, qid, flags);
974 		if (err) {
975 			xp_destroy(xs->pool);
976 			xs->pool = NULL;
977 			goto out_unlock;
978 		}
979 	}
980 
981 	/* FQ and CQ are now owned by the buffer pool and cleaned up with it. */
982 	xs->fq_tmp = NULL;
983 	xs->cq_tmp = NULL;
984 
985 	xs->dev = dev;
986 	xs->zc = xs->umem->zc;
987 	xs->queue_id = qid;
988 	xp_add_xsk(xs->pool, xs);
989 
990 out_unlock:
991 	if (err) {
992 		dev_put(dev);
993 	} else {
994 		/* Matches smp_rmb() in bind() for shared umem
995 		 * sockets, and xsk_is_bound().
996 		 */
997 		smp_wmb();
998 		WRITE_ONCE(xs->state, XSK_BOUND);
999 	}
1000 out_release:
1001 	mutex_unlock(&xs->mutex);
1002 	rtnl_unlock();
1003 	return err;
1004 }
1005 
1006 struct xdp_umem_reg_v1 {
1007 	__u64 addr; /* Start of packet data area */
1008 	__u64 len; /* Length of packet data area */
1009 	__u32 chunk_size;
1010 	__u32 headroom;
1011 };
1012 
1013 static int xsk_setsockopt(struct socket *sock, int level, int optname,
1014 			  sockptr_t optval, unsigned int optlen)
1015 {
1016 	struct sock *sk = sock->sk;
1017 	struct xdp_sock *xs = xdp_sk(sk);
1018 	int err;
1019 
1020 	if (level != SOL_XDP)
1021 		return -ENOPROTOOPT;
1022 
1023 	switch (optname) {
1024 	case XDP_RX_RING:
1025 	case XDP_TX_RING:
1026 	{
1027 		struct xsk_queue **q;
1028 		int entries;
1029 
1030 		if (optlen < sizeof(entries))
1031 			return -EINVAL;
1032 		if (copy_from_sockptr(&entries, optval, sizeof(entries)))
1033 			return -EFAULT;
1034 
1035 		mutex_lock(&xs->mutex);
1036 		if (xs->state != XSK_READY) {
1037 			mutex_unlock(&xs->mutex);
1038 			return -EBUSY;
1039 		}
1040 		q = (optname == XDP_TX_RING) ? &xs->tx : &xs->rx;
1041 		err = xsk_init_queue(entries, q, false);
1042 		if (!err && optname == XDP_TX_RING)
1043 			/* Tx needs to be explicitly woken up the first time */
1044 			xs->tx->ring->flags |= XDP_RING_NEED_WAKEUP;
1045 		mutex_unlock(&xs->mutex);
1046 		return err;
1047 	}
1048 	case XDP_UMEM_REG:
1049 	{
1050 		size_t mr_size = sizeof(struct xdp_umem_reg);
1051 		struct xdp_umem_reg mr = {};
1052 		struct xdp_umem *umem;
1053 
1054 		if (optlen < sizeof(struct xdp_umem_reg_v1))
1055 			return -EINVAL;
1056 		else if (optlen < sizeof(mr))
1057 			mr_size = sizeof(struct xdp_umem_reg_v1);
1058 
1059 		if (copy_from_sockptr(&mr, optval, mr_size))
1060 			return -EFAULT;
1061 
1062 		mutex_lock(&xs->mutex);
1063 		if (xs->state != XSK_READY || xs->umem) {
1064 			mutex_unlock(&xs->mutex);
1065 			return -EBUSY;
1066 		}
1067 
1068 		umem = xdp_umem_create(&mr);
1069 		if (IS_ERR(umem)) {
1070 			mutex_unlock(&xs->mutex);
1071 			return PTR_ERR(umem);
1072 		}
1073 
1074 		/* Make sure umem is ready before it can be seen by others */
1075 		smp_wmb();
1076 		WRITE_ONCE(xs->umem, umem);
1077 		mutex_unlock(&xs->mutex);
1078 		return 0;
1079 	}
1080 	case XDP_UMEM_FILL_RING:
1081 	case XDP_UMEM_COMPLETION_RING:
1082 	{
1083 		struct xsk_queue **q;
1084 		int entries;
1085 
1086 		if (copy_from_sockptr(&entries, optval, sizeof(entries)))
1087 			return -EFAULT;
1088 
1089 		mutex_lock(&xs->mutex);
1090 		if (xs->state != XSK_READY) {
1091 			mutex_unlock(&xs->mutex);
1092 			return -EBUSY;
1093 		}
1094 
1095 		q = (optname == XDP_UMEM_FILL_RING) ? &xs->fq_tmp :
1096 			&xs->cq_tmp;
1097 		err = xsk_init_queue(entries, q, true);
1098 		mutex_unlock(&xs->mutex);
1099 		return err;
1100 	}
1101 	default:
1102 		break;
1103 	}
1104 
1105 	return -ENOPROTOOPT;
1106 }
1107 
1108 static void xsk_enter_rxtx_offsets(struct xdp_ring_offset_v1 *ring)
1109 {
1110 	ring->producer = offsetof(struct xdp_rxtx_ring, ptrs.producer);
1111 	ring->consumer = offsetof(struct xdp_rxtx_ring, ptrs.consumer);
1112 	ring->desc = offsetof(struct xdp_rxtx_ring, desc);
1113 }
1114 
1115 static void xsk_enter_umem_offsets(struct xdp_ring_offset_v1 *ring)
1116 {
1117 	ring->producer = offsetof(struct xdp_umem_ring, ptrs.producer);
1118 	ring->consumer = offsetof(struct xdp_umem_ring, ptrs.consumer);
1119 	ring->desc = offsetof(struct xdp_umem_ring, desc);
1120 }
1121 
1122 struct xdp_statistics_v1 {
1123 	__u64 rx_dropped;
1124 	__u64 rx_invalid_descs;
1125 	__u64 tx_invalid_descs;
1126 };
1127 
1128 static int xsk_getsockopt(struct socket *sock, int level, int optname,
1129 			  char __user *optval, int __user *optlen)
1130 {
1131 	struct sock *sk = sock->sk;
1132 	struct xdp_sock *xs = xdp_sk(sk);
1133 	int len;
1134 
1135 	if (level != SOL_XDP)
1136 		return -ENOPROTOOPT;
1137 
1138 	if (get_user(len, optlen))
1139 		return -EFAULT;
1140 	if (len < 0)
1141 		return -EINVAL;
1142 
1143 	switch (optname) {
1144 	case XDP_STATISTICS:
1145 	{
1146 		struct xdp_statistics stats = {};
1147 		bool extra_stats = true;
1148 		size_t stats_size;
1149 
1150 		if (len < sizeof(struct xdp_statistics_v1)) {
1151 			return -EINVAL;
1152 		} else if (len < sizeof(stats)) {
1153 			extra_stats = false;
1154 			stats_size = sizeof(struct xdp_statistics_v1);
1155 		} else {
1156 			stats_size = sizeof(stats);
1157 		}
1158 
1159 		mutex_lock(&xs->mutex);
1160 		stats.rx_dropped = xs->rx_dropped;
1161 		if (extra_stats) {
1162 			stats.rx_ring_full = xs->rx_queue_full;
1163 			stats.rx_fill_ring_empty_descs =
1164 				xs->pool ? xskq_nb_queue_empty_descs(xs->pool->fq) : 0;
1165 			stats.tx_ring_empty_descs = xskq_nb_queue_empty_descs(xs->tx);
1166 		} else {
1167 			stats.rx_dropped += xs->rx_queue_full;
1168 		}
1169 		stats.rx_invalid_descs = xskq_nb_invalid_descs(xs->rx);
1170 		stats.tx_invalid_descs = xskq_nb_invalid_descs(xs->tx);
1171 		mutex_unlock(&xs->mutex);
1172 
1173 		if (copy_to_user(optval, &stats, stats_size))
1174 			return -EFAULT;
1175 		if (put_user(stats_size, optlen))
1176 			return -EFAULT;
1177 
1178 		return 0;
1179 	}
1180 	case XDP_MMAP_OFFSETS:
1181 	{
1182 		struct xdp_mmap_offsets off;
1183 		struct xdp_mmap_offsets_v1 off_v1;
1184 		bool flags_supported = true;
1185 		void *to_copy;
1186 
1187 		if (len < sizeof(off_v1))
1188 			return -EINVAL;
1189 		else if (len < sizeof(off))
1190 			flags_supported = false;
1191 
1192 		if (flags_supported) {
1193 			/* xdp_ring_offset is identical to xdp_ring_offset_v1
1194 			 * except for the flags field added to the end.
1195 			 */
1196 			xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *)
1197 					       &off.rx);
1198 			xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *)
1199 					       &off.tx);
1200 			xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *)
1201 					       &off.fr);
1202 			xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *)
1203 					       &off.cr);
1204 			off.rx.flags = offsetof(struct xdp_rxtx_ring,
1205 						ptrs.flags);
1206 			off.tx.flags = offsetof(struct xdp_rxtx_ring,
1207 						ptrs.flags);
1208 			off.fr.flags = offsetof(struct xdp_umem_ring,
1209 						ptrs.flags);
1210 			off.cr.flags = offsetof(struct xdp_umem_ring,
1211 						ptrs.flags);
1212 
1213 			len = sizeof(off);
1214 			to_copy = &off;
1215 		} else {
1216 			xsk_enter_rxtx_offsets(&off_v1.rx);
1217 			xsk_enter_rxtx_offsets(&off_v1.tx);
1218 			xsk_enter_umem_offsets(&off_v1.fr);
1219 			xsk_enter_umem_offsets(&off_v1.cr);
1220 
1221 			len = sizeof(off_v1);
1222 			to_copy = &off_v1;
1223 		}
1224 
1225 		if (copy_to_user(optval, to_copy, len))
1226 			return -EFAULT;
1227 		if (put_user(len, optlen))
1228 			return -EFAULT;
1229 
1230 		return 0;
1231 	}
1232 	case XDP_OPTIONS:
1233 	{
1234 		struct xdp_options opts = {};
1235 
1236 		if (len < sizeof(opts))
1237 			return -EINVAL;
1238 
1239 		mutex_lock(&xs->mutex);
1240 		if (xs->zc)
1241 			opts.flags |= XDP_OPTIONS_ZEROCOPY;
1242 		mutex_unlock(&xs->mutex);
1243 
1244 		len = sizeof(opts);
1245 		if (copy_to_user(optval, &opts, len))
1246 			return -EFAULT;
1247 		if (put_user(len, optlen))
1248 			return -EFAULT;
1249 
1250 		return 0;
1251 	}
1252 	default:
1253 		break;
1254 	}
1255 
1256 	return -EOPNOTSUPP;
1257 }
1258 
1259 static int xsk_mmap(struct file *file, struct socket *sock,
1260 		    struct vm_area_struct *vma)
1261 {
1262 	loff_t offset = (loff_t)vma->vm_pgoff << PAGE_SHIFT;
1263 	unsigned long size = vma->vm_end - vma->vm_start;
1264 	struct xdp_sock *xs = xdp_sk(sock->sk);
1265 	struct xsk_queue *q = NULL;
1266 	unsigned long pfn;
1267 	struct page *qpg;
1268 
1269 	if (READ_ONCE(xs->state) != XSK_READY)
1270 		return -EBUSY;
1271 
1272 	if (offset == XDP_PGOFF_RX_RING) {
1273 		q = READ_ONCE(xs->rx);
1274 	} else if (offset == XDP_PGOFF_TX_RING) {
1275 		q = READ_ONCE(xs->tx);
1276 	} else {
1277 		/* Matches the smp_wmb() in XDP_UMEM_REG */
1278 		smp_rmb();
1279 		if (offset == XDP_UMEM_PGOFF_FILL_RING)
1280 			q = READ_ONCE(xs->fq_tmp);
1281 		else if (offset == XDP_UMEM_PGOFF_COMPLETION_RING)
1282 			q = READ_ONCE(xs->cq_tmp);
1283 	}
1284 
1285 	if (!q)
1286 		return -EINVAL;
1287 
1288 	/* Matches the smp_wmb() in xsk_init_queue */
1289 	smp_rmb();
1290 	qpg = virt_to_head_page(q->ring);
1291 	if (size > page_size(qpg))
1292 		return -EINVAL;
1293 
1294 	pfn = virt_to_phys(q->ring) >> PAGE_SHIFT;
1295 	return remap_pfn_range(vma, vma->vm_start, pfn,
1296 			       size, vma->vm_page_prot);
1297 }
1298 
1299 static int xsk_notifier(struct notifier_block *this,
1300 			unsigned long msg, void *ptr)
1301 {
1302 	struct net_device *dev = netdev_notifier_info_to_dev(ptr);
1303 	struct net *net = dev_net(dev);
1304 	struct sock *sk;
1305 
1306 	switch (msg) {
1307 	case NETDEV_UNREGISTER:
1308 		mutex_lock(&net->xdp.lock);
1309 		sk_for_each(sk, &net->xdp.list) {
1310 			struct xdp_sock *xs = xdp_sk(sk);
1311 
1312 			mutex_lock(&xs->mutex);
1313 			if (xs->dev == dev) {
1314 				sk->sk_err = ENETDOWN;
1315 				if (!sock_flag(sk, SOCK_DEAD))
1316 					sk_error_report(sk);
1317 
1318 				xsk_unbind_dev(xs);
1319 
1320 				/* Clear device references. */
1321 				xp_clear_dev(xs->pool);
1322 			}
1323 			mutex_unlock(&xs->mutex);
1324 		}
1325 		mutex_unlock(&net->xdp.lock);
1326 		break;
1327 	}
1328 	return NOTIFY_DONE;
1329 }
1330 
1331 static struct proto xsk_proto = {
1332 	.name =		"XDP",
1333 	.owner =	THIS_MODULE,
1334 	.obj_size =	sizeof(struct xdp_sock),
1335 };
1336 
1337 static const struct proto_ops xsk_proto_ops = {
1338 	.family		= PF_XDP,
1339 	.owner		= THIS_MODULE,
1340 	.release	= xsk_release,
1341 	.bind		= xsk_bind,
1342 	.connect	= sock_no_connect,
1343 	.socketpair	= sock_no_socketpair,
1344 	.accept		= sock_no_accept,
1345 	.getname	= sock_no_getname,
1346 	.poll		= xsk_poll,
1347 	.ioctl		= sock_no_ioctl,
1348 	.listen		= sock_no_listen,
1349 	.shutdown	= sock_no_shutdown,
1350 	.setsockopt	= xsk_setsockopt,
1351 	.getsockopt	= xsk_getsockopt,
1352 	.sendmsg	= xsk_sendmsg,
1353 	.recvmsg	= xsk_recvmsg,
1354 	.mmap		= xsk_mmap,
1355 	.sendpage	= sock_no_sendpage,
1356 };
1357 
1358 static void xsk_destruct(struct sock *sk)
1359 {
1360 	struct xdp_sock *xs = xdp_sk(sk);
1361 
1362 	if (!sock_flag(sk, SOCK_DEAD))
1363 		return;
1364 
1365 	if (!xp_put_pool(xs->pool))
1366 		xdp_put_umem(xs->umem, !xs->pool);
1367 
1368 	sk_refcnt_debug_dec(sk);
1369 }
1370 
1371 static int xsk_create(struct net *net, struct socket *sock, int protocol,
1372 		      int kern)
1373 {
1374 	struct xdp_sock *xs;
1375 	struct sock *sk;
1376 
1377 	if (!ns_capable(net->user_ns, CAP_NET_RAW))
1378 		return -EPERM;
1379 	if (sock->type != SOCK_RAW)
1380 		return -ESOCKTNOSUPPORT;
1381 
1382 	if (protocol)
1383 		return -EPROTONOSUPPORT;
1384 
1385 	sock->state = SS_UNCONNECTED;
1386 
1387 	sk = sk_alloc(net, PF_XDP, GFP_KERNEL, &xsk_proto, kern);
1388 	if (!sk)
1389 		return -ENOBUFS;
1390 
1391 	sock->ops = &xsk_proto_ops;
1392 
1393 	sock_init_data(sock, sk);
1394 
1395 	sk->sk_family = PF_XDP;
1396 
1397 	sk->sk_destruct = xsk_destruct;
1398 	sk_refcnt_debug_inc(sk);
1399 
1400 	sock_set_flag(sk, SOCK_RCU_FREE);
1401 
1402 	xs = xdp_sk(sk);
1403 	xs->state = XSK_READY;
1404 	mutex_init(&xs->mutex);
1405 	spin_lock_init(&xs->rx_lock);
1406 
1407 	INIT_LIST_HEAD(&xs->map_list);
1408 	spin_lock_init(&xs->map_list_lock);
1409 
1410 	mutex_lock(&net->xdp.lock);
1411 	sk_add_node_rcu(sk, &net->xdp.list);
1412 	mutex_unlock(&net->xdp.lock);
1413 
1414 	local_bh_disable();
1415 	sock_prot_inuse_add(net, &xsk_proto, 1);
1416 	local_bh_enable();
1417 
1418 	return 0;
1419 }
1420 
1421 static const struct net_proto_family xsk_family_ops = {
1422 	.family = PF_XDP,
1423 	.create = xsk_create,
1424 	.owner	= THIS_MODULE,
1425 };
1426 
1427 static struct notifier_block xsk_netdev_notifier = {
1428 	.notifier_call	= xsk_notifier,
1429 };
1430 
1431 static int __net_init xsk_net_init(struct net *net)
1432 {
1433 	mutex_init(&net->xdp.lock);
1434 	INIT_HLIST_HEAD(&net->xdp.list);
1435 	return 0;
1436 }
1437 
1438 static void __net_exit xsk_net_exit(struct net *net)
1439 {
1440 	WARN_ON_ONCE(!hlist_empty(&net->xdp.list));
1441 }
1442 
1443 static struct pernet_operations xsk_net_ops = {
1444 	.init = xsk_net_init,
1445 	.exit = xsk_net_exit,
1446 };
1447 
1448 static int __init xsk_init(void)
1449 {
1450 	int err, cpu;
1451 
1452 	err = proto_register(&xsk_proto, 0 /* no slab */);
1453 	if (err)
1454 		goto out;
1455 
1456 	err = sock_register(&xsk_family_ops);
1457 	if (err)
1458 		goto out_proto;
1459 
1460 	err = register_pernet_subsys(&xsk_net_ops);
1461 	if (err)
1462 		goto out_sk;
1463 
1464 	err = register_netdevice_notifier(&xsk_netdev_notifier);
1465 	if (err)
1466 		goto out_pernet;
1467 
1468 	for_each_possible_cpu(cpu)
1469 		INIT_LIST_HEAD(&per_cpu(xskmap_flush_list, cpu));
1470 	return 0;
1471 
1472 out_pernet:
1473 	unregister_pernet_subsys(&xsk_net_ops);
1474 out_sk:
1475 	sock_unregister(PF_XDP);
1476 out_proto:
1477 	proto_unregister(&xsk_proto);
1478 out:
1479 	return err;
1480 }
1481 
1482 fs_initcall(xsk_init);
1483