xref: /linux/net/xdp/xsk.c (revision a97f9c8fcc6ed16f956346368fde3b1198163f6c)
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 <linux/vmalloc.h>
26 #include <net/xdp_sock_drv.h>
27 #include <net/busy_poll.h>
28 #include <net/netdev_rx_queue.h>
29 #include <net/xdp.h>
30 
31 #include "xsk_queue.h"
32 #include "xdp_umem.h"
33 #include "xsk.h"
34 
35 #define TX_BATCH_SIZE 32
36 #define MAX_PER_SOCKET_BUDGET (TX_BATCH_SIZE)
37 
38 static DEFINE_PER_CPU(struct list_head, xskmap_flush_list);
39 
40 void xsk_set_rx_need_wakeup(struct xsk_buff_pool *pool)
41 {
42 	if (pool->cached_need_wakeup & XDP_WAKEUP_RX)
43 		return;
44 
45 	pool->fq->ring->flags |= XDP_RING_NEED_WAKEUP;
46 	pool->cached_need_wakeup |= XDP_WAKEUP_RX;
47 }
48 EXPORT_SYMBOL(xsk_set_rx_need_wakeup);
49 
50 void xsk_set_tx_need_wakeup(struct xsk_buff_pool *pool)
51 {
52 	struct xdp_sock *xs;
53 
54 	if (pool->cached_need_wakeup & XDP_WAKEUP_TX)
55 		return;
56 
57 	rcu_read_lock();
58 	list_for_each_entry_rcu(xs, &pool->xsk_tx_list, tx_list) {
59 		xs->tx->ring->flags |= XDP_RING_NEED_WAKEUP;
60 	}
61 	rcu_read_unlock();
62 
63 	pool->cached_need_wakeup |= XDP_WAKEUP_TX;
64 }
65 EXPORT_SYMBOL(xsk_set_tx_need_wakeup);
66 
67 void xsk_clear_rx_need_wakeup(struct xsk_buff_pool *pool)
68 {
69 	if (!(pool->cached_need_wakeup & XDP_WAKEUP_RX))
70 		return;
71 
72 	pool->fq->ring->flags &= ~XDP_RING_NEED_WAKEUP;
73 	pool->cached_need_wakeup &= ~XDP_WAKEUP_RX;
74 }
75 EXPORT_SYMBOL(xsk_clear_rx_need_wakeup);
76 
77 void xsk_clear_tx_need_wakeup(struct xsk_buff_pool *pool)
78 {
79 	struct xdp_sock *xs;
80 
81 	if (!(pool->cached_need_wakeup & XDP_WAKEUP_TX))
82 		return;
83 
84 	rcu_read_lock();
85 	list_for_each_entry_rcu(xs, &pool->xsk_tx_list, tx_list) {
86 		xs->tx->ring->flags &= ~XDP_RING_NEED_WAKEUP;
87 	}
88 	rcu_read_unlock();
89 
90 	pool->cached_need_wakeup &= ~XDP_WAKEUP_TX;
91 }
92 EXPORT_SYMBOL(xsk_clear_tx_need_wakeup);
93 
94 bool xsk_uses_need_wakeup(struct xsk_buff_pool *pool)
95 {
96 	return pool->uses_need_wakeup;
97 }
98 EXPORT_SYMBOL(xsk_uses_need_wakeup);
99 
100 struct xsk_buff_pool *xsk_get_pool_from_qid(struct net_device *dev,
101 					    u16 queue_id)
102 {
103 	if (queue_id < dev->real_num_rx_queues)
104 		return dev->_rx[queue_id].pool;
105 	if (queue_id < dev->real_num_tx_queues)
106 		return dev->_tx[queue_id].pool;
107 
108 	return NULL;
109 }
110 EXPORT_SYMBOL(xsk_get_pool_from_qid);
111 
112 void xsk_clear_pool_at_qid(struct net_device *dev, u16 queue_id)
113 {
114 	if (queue_id < dev->num_rx_queues)
115 		dev->_rx[queue_id].pool = NULL;
116 	if (queue_id < dev->num_tx_queues)
117 		dev->_tx[queue_id].pool = NULL;
118 }
119 
120 /* The buffer pool is stored both in the _rx struct and the _tx struct as we do
121  * not know if the device has more tx queues than rx, or the opposite.
122  * This might also change during run time.
123  */
124 int xsk_reg_pool_at_qid(struct net_device *dev, struct xsk_buff_pool *pool,
125 			u16 queue_id)
126 {
127 	if (queue_id >= max_t(unsigned int,
128 			      dev->real_num_rx_queues,
129 			      dev->real_num_tx_queues))
130 		return -EINVAL;
131 
132 	if (queue_id < dev->real_num_rx_queues)
133 		dev->_rx[queue_id].pool = pool;
134 	if (queue_id < dev->real_num_tx_queues)
135 		dev->_tx[queue_id].pool = pool;
136 
137 	return 0;
138 }
139 
140 static int __xsk_rcv_zc(struct xdp_sock *xs, struct xdp_buff_xsk *xskb, u32 len,
141 			u32 flags)
142 {
143 	u64 addr;
144 	int err;
145 
146 	addr = xp_get_handle(xskb);
147 	err = xskq_prod_reserve_desc(xs->rx, addr, len, flags);
148 	if (err) {
149 		xs->rx_queue_full++;
150 		return err;
151 	}
152 
153 	xp_release(xskb);
154 	return 0;
155 }
156 
157 static int xsk_rcv_zc(struct xdp_sock *xs, struct xdp_buff *xdp, u32 len)
158 {
159 	struct xdp_buff_xsk *xskb = container_of(xdp, struct xdp_buff_xsk, xdp);
160 	u32 frags = xdp_buff_has_frags(xdp);
161 	struct xdp_buff_xsk *pos, *tmp;
162 	struct list_head *xskb_list;
163 	u32 contd = 0;
164 	int err;
165 
166 	if (frags)
167 		contd = XDP_PKT_CONTD;
168 
169 	err = __xsk_rcv_zc(xs, xskb, len, contd);
170 	if (err)
171 		goto err;
172 	if (likely(!frags))
173 		return 0;
174 
175 	xskb_list = &xskb->pool->xskb_list;
176 	list_for_each_entry_safe(pos, tmp, xskb_list, xskb_list_node) {
177 		if (list_is_singular(xskb_list))
178 			contd = 0;
179 		len = pos->xdp.data_end - pos->xdp.data;
180 		err = __xsk_rcv_zc(xs, pos, len, contd);
181 		if (err)
182 			goto err;
183 		list_del(&pos->xskb_list_node);
184 	}
185 
186 	return 0;
187 err:
188 	xsk_buff_free(xdp);
189 	return err;
190 }
191 
192 static void *xsk_copy_xdp_start(struct xdp_buff *from)
193 {
194 	if (unlikely(xdp_data_meta_unsupported(from)))
195 		return from->data;
196 	else
197 		return from->data_meta;
198 }
199 
200 static u32 xsk_copy_xdp(void *to, void **from, u32 to_len,
201 			u32 *from_len, skb_frag_t **frag, u32 rem)
202 {
203 	u32 copied = 0;
204 
205 	while (1) {
206 		u32 copy_len = min_t(u32, *from_len, to_len);
207 
208 		memcpy(to, *from, copy_len);
209 		copied += copy_len;
210 		if (rem == copied)
211 			return copied;
212 
213 		if (*from_len == copy_len) {
214 			*from = skb_frag_address(*frag);
215 			*from_len = skb_frag_size((*frag)++);
216 		} else {
217 			*from += copy_len;
218 			*from_len -= copy_len;
219 		}
220 		if (to_len == copy_len)
221 			return copied;
222 
223 		to_len -= copy_len;
224 		to += copy_len;
225 	}
226 }
227 
228 static int __xsk_rcv(struct xdp_sock *xs, struct xdp_buff *xdp, u32 len)
229 {
230 	u32 frame_size = xsk_pool_get_rx_frame_size(xs->pool);
231 	void *copy_from = xsk_copy_xdp_start(xdp), *copy_to;
232 	u32 from_len, meta_len, rem, num_desc;
233 	struct xdp_buff_xsk *xskb;
234 	struct xdp_buff *xsk_xdp;
235 	skb_frag_t *frag;
236 
237 	from_len = xdp->data_end - copy_from;
238 	meta_len = xdp->data - copy_from;
239 	rem = len + meta_len;
240 
241 	if (len <= frame_size && !xdp_buff_has_frags(xdp)) {
242 		int err;
243 
244 		xsk_xdp = xsk_buff_alloc(xs->pool);
245 		if (!xsk_xdp) {
246 			xs->rx_dropped++;
247 			return -ENOMEM;
248 		}
249 		memcpy(xsk_xdp->data - meta_len, copy_from, rem);
250 		xskb = container_of(xsk_xdp, struct xdp_buff_xsk, xdp);
251 		err = __xsk_rcv_zc(xs, xskb, len, 0);
252 		if (err) {
253 			xsk_buff_free(xsk_xdp);
254 			return err;
255 		}
256 
257 		return 0;
258 	}
259 
260 	num_desc = (len - 1) / frame_size + 1;
261 
262 	if (!xsk_buff_can_alloc(xs->pool, num_desc)) {
263 		xs->rx_dropped++;
264 		return -ENOMEM;
265 	}
266 	if (xskq_prod_nb_free(xs->rx, num_desc) < num_desc) {
267 		xs->rx_queue_full++;
268 		return -ENOBUFS;
269 	}
270 
271 	if (xdp_buff_has_frags(xdp)) {
272 		struct skb_shared_info *sinfo;
273 
274 		sinfo = xdp_get_shared_info_from_buff(xdp);
275 		frag =  &sinfo->frags[0];
276 	}
277 
278 	do {
279 		u32 to_len = frame_size + meta_len;
280 		u32 copied;
281 
282 		xsk_xdp = xsk_buff_alloc(xs->pool);
283 		copy_to = xsk_xdp->data - meta_len;
284 
285 		copied = xsk_copy_xdp(copy_to, &copy_from, to_len, &from_len, &frag, rem);
286 		rem -= copied;
287 
288 		xskb = container_of(xsk_xdp, struct xdp_buff_xsk, xdp);
289 		__xsk_rcv_zc(xs, xskb, copied - meta_len, rem ? XDP_PKT_CONTD : 0);
290 		meta_len = 0;
291 	} while (rem);
292 
293 	return 0;
294 }
295 
296 static bool xsk_tx_writeable(struct xdp_sock *xs)
297 {
298 	if (xskq_cons_present_entries(xs->tx) > xs->tx->nentries / 2)
299 		return false;
300 
301 	return true;
302 }
303 
304 static bool xsk_is_bound(struct xdp_sock *xs)
305 {
306 	if (READ_ONCE(xs->state) == XSK_BOUND) {
307 		/* Matches smp_wmb() in bind(). */
308 		smp_rmb();
309 		return true;
310 	}
311 	return false;
312 }
313 
314 static int xsk_rcv_check(struct xdp_sock *xs, struct xdp_buff *xdp, u32 len)
315 {
316 	if (!xsk_is_bound(xs))
317 		return -ENXIO;
318 
319 	if (xs->dev != xdp->rxq->dev || xs->queue_id != xdp->rxq->queue_index)
320 		return -EINVAL;
321 
322 	if (len > xsk_pool_get_rx_frame_size(xs->pool) && !xs->sg) {
323 		xs->rx_dropped++;
324 		return -ENOSPC;
325 	}
326 
327 	sk_mark_napi_id_once_xdp(&xs->sk, xdp);
328 	return 0;
329 }
330 
331 static void xsk_flush(struct xdp_sock *xs)
332 {
333 	xskq_prod_submit(xs->rx);
334 	__xskq_cons_release(xs->pool->fq);
335 	sock_def_readable(&xs->sk);
336 }
337 
338 int xsk_generic_rcv(struct xdp_sock *xs, struct xdp_buff *xdp)
339 {
340 	u32 len = xdp_get_buff_len(xdp);
341 	int err;
342 
343 	spin_lock_bh(&xs->rx_lock);
344 	err = xsk_rcv_check(xs, xdp, len);
345 	if (!err) {
346 		err = __xsk_rcv(xs, xdp, len);
347 		xsk_flush(xs);
348 	}
349 	spin_unlock_bh(&xs->rx_lock);
350 	return err;
351 }
352 
353 static int xsk_rcv(struct xdp_sock *xs, struct xdp_buff *xdp)
354 {
355 	u32 len = xdp_get_buff_len(xdp);
356 	int err;
357 
358 	err = xsk_rcv_check(xs, xdp, len);
359 	if (err)
360 		return err;
361 
362 	if (xdp->rxq->mem.type == MEM_TYPE_XSK_BUFF_POOL) {
363 		len = xdp->data_end - xdp->data;
364 		return xsk_rcv_zc(xs, xdp, len);
365 	}
366 
367 	err = __xsk_rcv(xs, xdp, len);
368 	if (!err)
369 		xdp_return_buff(xdp);
370 	return err;
371 }
372 
373 int __xsk_map_redirect(struct xdp_sock *xs, struct xdp_buff *xdp)
374 {
375 	struct list_head *flush_list = this_cpu_ptr(&xskmap_flush_list);
376 	int err;
377 
378 	err = xsk_rcv(xs, xdp);
379 	if (err)
380 		return err;
381 
382 	if (!xs->flush_node.prev)
383 		list_add(&xs->flush_node, flush_list);
384 
385 	return 0;
386 }
387 
388 void __xsk_map_flush(void)
389 {
390 	struct list_head *flush_list = this_cpu_ptr(&xskmap_flush_list);
391 	struct xdp_sock *xs, *tmp;
392 
393 	list_for_each_entry_safe(xs, tmp, flush_list, flush_node) {
394 		xsk_flush(xs);
395 		__list_del_clearprev(&xs->flush_node);
396 	}
397 }
398 
399 #ifdef CONFIG_DEBUG_NET
400 bool xsk_map_check_flush(void)
401 {
402 	if (list_empty(this_cpu_ptr(&xskmap_flush_list)))
403 		return false;
404 	__xsk_map_flush();
405 	return true;
406 }
407 #endif
408 
409 void xsk_tx_completed(struct xsk_buff_pool *pool, u32 nb_entries)
410 {
411 	xskq_prod_submit_n(pool->cq, nb_entries);
412 }
413 EXPORT_SYMBOL(xsk_tx_completed);
414 
415 void xsk_tx_release(struct xsk_buff_pool *pool)
416 {
417 	struct xdp_sock *xs;
418 
419 	rcu_read_lock();
420 	list_for_each_entry_rcu(xs, &pool->xsk_tx_list, tx_list) {
421 		__xskq_cons_release(xs->tx);
422 		if (xsk_tx_writeable(xs))
423 			xs->sk.sk_write_space(&xs->sk);
424 	}
425 	rcu_read_unlock();
426 }
427 EXPORT_SYMBOL(xsk_tx_release);
428 
429 bool xsk_tx_peek_desc(struct xsk_buff_pool *pool, struct xdp_desc *desc)
430 {
431 	bool budget_exhausted = false;
432 	struct xdp_sock *xs;
433 
434 	rcu_read_lock();
435 again:
436 	list_for_each_entry_rcu(xs, &pool->xsk_tx_list, tx_list) {
437 		if (xs->tx_budget_spent >= MAX_PER_SOCKET_BUDGET) {
438 			budget_exhausted = true;
439 			continue;
440 		}
441 
442 		if (!xskq_cons_peek_desc(xs->tx, desc, pool)) {
443 			if (xskq_has_descs(xs->tx))
444 				xskq_cons_release(xs->tx);
445 			continue;
446 		}
447 
448 		xs->tx_budget_spent++;
449 
450 		/* This is the backpressure mechanism for the Tx path.
451 		 * Reserve space in the completion queue and only proceed
452 		 * if there is space in it. This avoids having to implement
453 		 * any buffering in the Tx path.
454 		 */
455 		if (xskq_prod_reserve_addr(pool->cq, desc->addr))
456 			goto out;
457 
458 		xskq_cons_release(xs->tx);
459 		rcu_read_unlock();
460 		return true;
461 	}
462 
463 	if (budget_exhausted) {
464 		list_for_each_entry_rcu(xs, &pool->xsk_tx_list, tx_list)
465 			xs->tx_budget_spent = 0;
466 
467 		budget_exhausted = false;
468 		goto again;
469 	}
470 
471 out:
472 	rcu_read_unlock();
473 	return false;
474 }
475 EXPORT_SYMBOL(xsk_tx_peek_desc);
476 
477 static u32 xsk_tx_peek_release_fallback(struct xsk_buff_pool *pool, u32 max_entries)
478 {
479 	struct xdp_desc *descs = pool->tx_descs;
480 	u32 nb_pkts = 0;
481 
482 	while (nb_pkts < max_entries && xsk_tx_peek_desc(pool, &descs[nb_pkts]))
483 		nb_pkts++;
484 
485 	xsk_tx_release(pool);
486 	return nb_pkts;
487 }
488 
489 u32 xsk_tx_peek_release_desc_batch(struct xsk_buff_pool *pool, u32 nb_pkts)
490 {
491 	struct xdp_sock *xs;
492 
493 	rcu_read_lock();
494 	if (!list_is_singular(&pool->xsk_tx_list)) {
495 		/* Fallback to the non-batched version */
496 		rcu_read_unlock();
497 		return xsk_tx_peek_release_fallback(pool, nb_pkts);
498 	}
499 
500 	xs = list_first_or_null_rcu(&pool->xsk_tx_list, struct xdp_sock, tx_list);
501 	if (!xs) {
502 		nb_pkts = 0;
503 		goto out;
504 	}
505 
506 	nb_pkts = xskq_cons_nb_entries(xs->tx, nb_pkts);
507 
508 	/* This is the backpressure mechanism for the Tx path. Try to
509 	 * reserve space in the completion queue for all packets, but
510 	 * if there are fewer slots available, just process that many
511 	 * packets. This avoids having to implement any buffering in
512 	 * the Tx path.
513 	 */
514 	nb_pkts = xskq_prod_nb_free(pool->cq, nb_pkts);
515 	if (!nb_pkts)
516 		goto out;
517 
518 	nb_pkts = xskq_cons_read_desc_batch(xs->tx, pool, nb_pkts);
519 	if (!nb_pkts) {
520 		xs->tx->queue_empty_descs++;
521 		goto out;
522 	}
523 
524 	__xskq_cons_release(xs->tx);
525 	xskq_prod_write_addr_batch(pool->cq, pool->tx_descs, nb_pkts);
526 	xs->sk.sk_write_space(&xs->sk);
527 
528 out:
529 	rcu_read_unlock();
530 	return nb_pkts;
531 }
532 EXPORT_SYMBOL(xsk_tx_peek_release_desc_batch);
533 
534 static int xsk_wakeup(struct xdp_sock *xs, u8 flags)
535 {
536 	struct net_device *dev = xs->dev;
537 
538 	return dev->netdev_ops->ndo_xsk_wakeup(dev, xs->queue_id, flags);
539 }
540 
541 static int xsk_cq_reserve_addr_locked(struct xdp_sock *xs, u64 addr)
542 {
543 	unsigned long flags;
544 	int ret;
545 
546 	spin_lock_irqsave(&xs->pool->cq_lock, flags);
547 	ret = xskq_prod_reserve_addr(xs->pool->cq, addr);
548 	spin_unlock_irqrestore(&xs->pool->cq_lock, flags);
549 
550 	return ret;
551 }
552 
553 static void xsk_cq_submit_locked(struct xdp_sock *xs, u32 n)
554 {
555 	unsigned long flags;
556 
557 	spin_lock_irqsave(&xs->pool->cq_lock, flags);
558 	xskq_prod_submit_n(xs->pool->cq, n);
559 	spin_unlock_irqrestore(&xs->pool->cq_lock, flags);
560 }
561 
562 static void xsk_cq_cancel_locked(struct xdp_sock *xs, u32 n)
563 {
564 	unsigned long flags;
565 
566 	spin_lock_irqsave(&xs->pool->cq_lock, flags);
567 	xskq_prod_cancel_n(xs->pool->cq, n);
568 	spin_unlock_irqrestore(&xs->pool->cq_lock, flags);
569 }
570 
571 static u32 xsk_get_num_desc(struct sk_buff *skb)
572 {
573 	return skb ? (long)skb_shinfo(skb)->destructor_arg : 0;
574 }
575 
576 static void xsk_destruct_skb(struct sk_buff *skb)
577 {
578 	struct xsk_tx_metadata_compl *compl = &skb_shinfo(skb)->xsk_meta;
579 
580 	if (compl->tx_timestamp) {
581 		/* sw completion timestamp, not a real one */
582 		*compl->tx_timestamp = ktime_get_tai_fast_ns();
583 	}
584 
585 	xsk_cq_submit_locked(xdp_sk(skb->sk), xsk_get_num_desc(skb));
586 	sock_wfree(skb);
587 }
588 
589 static void xsk_set_destructor_arg(struct sk_buff *skb)
590 {
591 	long num = xsk_get_num_desc(xdp_sk(skb->sk)->skb) + 1;
592 
593 	skb_shinfo(skb)->destructor_arg = (void *)num;
594 }
595 
596 static void xsk_consume_skb(struct sk_buff *skb)
597 {
598 	struct xdp_sock *xs = xdp_sk(skb->sk);
599 
600 	skb->destructor = sock_wfree;
601 	xsk_cq_cancel_locked(xs, xsk_get_num_desc(skb));
602 	/* Free skb without triggering the perf drop trace */
603 	consume_skb(skb);
604 	xs->skb = NULL;
605 }
606 
607 static void xsk_drop_skb(struct sk_buff *skb)
608 {
609 	xdp_sk(skb->sk)->tx->invalid_descs += xsk_get_num_desc(skb);
610 	xsk_consume_skb(skb);
611 }
612 
613 static struct sk_buff *xsk_build_skb_zerocopy(struct xdp_sock *xs,
614 					      struct xdp_desc *desc)
615 {
616 	struct xsk_buff_pool *pool = xs->pool;
617 	u32 hr, len, ts, offset, copy, copied;
618 	struct sk_buff *skb = xs->skb;
619 	struct page *page;
620 	void *buffer;
621 	int err, i;
622 	u64 addr;
623 
624 	if (!skb) {
625 		hr = max(NET_SKB_PAD, L1_CACHE_ALIGN(xs->dev->needed_headroom));
626 
627 		skb = sock_alloc_send_skb(&xs->sk, hr, 1, &err);
628 		if (unlikely(!skb))
629 			return ERR_PTR(err);
630 
631 		skb_reserve(skb, hr);
632 	}
633 
634 	addr = desc->addr;
635 	len = desc->len;
636 	ts = pool->unaligned ? len : pool->chunk_size;
637 
638 	buffer = xsk_buff_raw_get_data(pool, addr);
639 	offset = offset_in_page(buffer);
640 	addr = buffer - pool->addrs;
641 
642 	for (copied = 0, i = skb_shinfo(skb)->nr_frags; copied < len; i++) {
643 		if (unlikely(i >= MAX_SKB_FRAGS))
644 			return ERR_PTR(-EOVERFLOW);
645 
646 		page = pool->umem->pgs[addr >> PAGE_SHIFT];
647 		get_page(page);
648 
649 		copy = min_t(u32, PAGE_SIZE - offset, len - copied);
650 		skb_fill_page_desc(skb, i, page, offset, copy);
651 
652 		copied += copy;
653 		addr += copy;
654 		offset = 0;
655 	}
656 
657 	skb->len += len;
658 	skb->data_len += len;
659 	skb->truesize += ts;
660 
661 	refcount_add(ts, &xs->sk.sk_wmem_alloc);
662 
663 	return skb;
664 }
665 
666 static struct sk_buff *xsk_build_skb(struct xdp_sock *xs,
667 				     struct xdp_desc *desc)
668 {
669 	struct xsk_tx_metadata *meta = NULL;
670 	struct net_device *dev = xs->dev;
671 	struct sk_buff *skb = xs->skb;
672 	bool first_frag = false;
673 	int err;
674 
675 	if (dev->priv_flags & IFF_TX_SKB_NO_LINEAR) {
676 		skb = xsk_build_skb_zerocopy(xs, desc);
677 		if (IS_ERR(skb)) {
678 			err = PTR_ERR(skb);
679 			goto free_err;
680 		}
681 	} else {
682 		u32 hr, tr, len;
683 		void *buffer;
684 
685 		buffer = xsk_buff_raw_get_data(xs->pool, desc->addr);
686 		len = desc->len;
687 
688 		if (!skb) {
689 			hr = max(NET_SKB_PAD, L1_CACHE_ALIGN(dev->needed_headroom));
690 			tr = dev->needed_tailroom;
691 			skb = sock_alloc_send_skb(&xs->sk, hr + len + tr, 1, &err);
692 			if (unlikely(!skb))
693 				goto free_err;
694 
695 			skb_reserve(skb, hr);
696 			skb_put(skb, len);
697 
698 			err = skb_store_bits(skb, 0, buffer, len);
699 			if (unlikely(err)) {
700 				kfree_skb(skb);
701 				goto free_err;
702 			}
703 
704 			first_frag = true;
705 		} else {
706 			int nr_frags = skb_shinfo(skb)->nr_frags;
707 			struct page *page;
708 			u8 *vaddr;
709 
710 			if (unlikely(nr_frags == (MAX_SKB_FRAGS - 1) && xp_mb_desc(desc))) {
711 				err = -EOVERFLOW;
712 				goto free_err;
713 			}
714 
715 			page = alloc_page(xs->sk.sk_allocation);
716 			if (unlikely(!page)) {
717 				err = -EAGAIN;
718 				goto free_err;
719 			}
720 
721 			vaddr = kmap_local_page(page);
722 			memcpy(vaddr, buffer, len);
723 			kunmap_local(vaddr);
724 
725 			skb_add_rx_frag(skb, nr_frags, page, 0, len, PAGE_SIZE);
726 			refcount_add(PAGE_SIZE, &xs->sk.sk_wmem_alloc);
727 		}
728 
729 		if (first_frag && desc->options & XDP_TX_METADATA) {
730 			if (unlikely(xs->pool->tx_metadata_len == 0)) {
731 				err = -EINVAL;
732 				goto free_err;
733 			}
734 
735 			meta = buffer - xs->pool->tx_metadata_len;
736 			if (unlikely(!xsk_buff_valid_tx_metadata(meta))) {
737 				err = -EINVAL;
738 				goto free_err;
739 			}
740 
741 			if (meta->flags & XDP_TXMD_FLAGS_CHECKSUM) {
742 				if (unlikely(meta->request.csum_start +
743 					     meta->request.csum_offset +
744 					     sizeof(__sum16) > len)) {
745 					err = -EINVAL;
746 					goto free_err;
747 				}
748 
749 				skb->csum_start = hr + meta->request.csum_start;
750 				skb->csum_offset = meta->request.csum_offset;
751 				skb->ip_summed = CHECKSUM_PARTIAL;
752 
753 				if (unlikely(xs->pool->tx_sw_csum)) {
754 					err = skb_checksum_help(skb);
755 					if (err)
756 						goto free_err;
757 				}
758 			}
759 		}
760 	}
761 
762 	skb->dev = dev;
763 	skb->priority = READ_ONCE(xs->sk.sk_priority);
764 	skb->mark = READ_ONCE(xs->sk.sk_mark);
765 	skb->destructor = xsk_destruct_skb;
766 	xsk_tx_metadata_to_compl(meta, &skb_shinfo(skb)->xsk_meta);
767 	xsk_set_destructor_arg(skb);
768 
769 	return skb;
770 
771 free_err:
772 	if (err == -EOVERFLOW) {
773 		/* Drop the packet */
774 		xsk_set_destructor_arg(xs->skb);
775 		xsk_drop_skb(xs->skb);
776 		xskq_cons_release(xs->tx);
777 	} else {
778 		/* Let application retry */
779 		xsk_cq_cancel_locked(xs, 1);
780 	}
781 
782 	return ERR_PTR(err);
783 }
784 
785 static int __xsk_generic_xmit(struct sock *sk)
786 {
787 	struct xdp_sock *xs = xdp_sk(sk);
788 	u32 max_batch = TX_BATCH_SIZE;
789 	bool sent_frame = false;
790 	struct xdp_desc desc;
791 	struct sk_buff *skb;
792 	int err = 0;
793 
794 	mutex_lock(&xs->mutex);
795 
796 	/* Since we dropped the RCU read lock, the socket state might have changed. */
797 	if (unlikely(!xsk_is_bound(xs))) {
798 		err = -ENXIO;
799 		goto out;
800 	}
801 
802 	if (xs->queue_id >= xs->dev->real_num_tx_queues)
803 		goto out;
804 
805 	while (xskq_cons_peek_desc(xs->tx, &desc, xs->pool)) {
806 		if (max_batch-- == 0) {
807 			err = -EAGAIN;
808 			goto out;
809 		}
810 
811 		/* This is the backpressure mechanism for the Tx path.
812 		 * Reserve space in the completion queue and only proceed
813 		 * if there is space in it. This avoids having to implement
814 		 * any buffering in the Tx path.
815 		 */
816 		if (xsk_cq_reserve_addr_locked(xs, desc.addr))
817 			goto out;
818 
819 		skb = xsk_build_skb(xs, &desc);
820 		if (IS_ERR(skb)) {
821 			err = PTR_ERR(skb);
822 			if (err != -EOVERFLOW)
823 				goto out;
824 			err = 0;
825 			continue;
826 		}
827 
828 		xskq_cons_release(xs->tx);
829 
830 		if (xp_mb_desc(&desc)) {
831 			xs->skb = skb;
832 			continue;
833 		}
834 
835 		err = __dev_direct_xmit(skb, xs->queue_id);
836 		if  (err == NETDEV_TX_BUSY) {
837 			/* Tell user-space to retry the send */
838 			xskq_cons_cancel_n(xs->tx, xsk_get_num_desc(skb));
839 			xsk_consume_skb(skb);
840 			err = -EAGAIN;
841 			goto out;
842 		}
843 
844 		/* Ignore NET_XMIT_CN as packet might have been sent */
845 		if (err == NET_XMIT_DROP) {
846 			/* SKB completed but not sent */
847 			err = -EBUSY;
848 			xs->skb = NULL;
849 			goto out;
850 		}
851 
852 		sent_frame = true;
853 		xs->skb = NULL;
854 	}
855 
856 	if (xskq_has_descs(xs->tx)) {
857 		if (xs->skb)
858 			xsk_drop_skb(xs->skb);
859 		xskq_cons_release(xs->tx);
860 	}
861 
862 out:
863 	if (sent_frame)
864 		if (xsk_tx_writeable(xs))
865 			sk->sk_write_space(sk);
866 
867 	mutex_unlock(&xs->mutex);
868 	return err;
869 }
870 
871 static int xsk_generic_xmit(struct sock *sk)
872 {
873 	int ret;
874 
875 	/* Drop the RCU lock since the SKB path might sleep. */
876 	rcu_read_unlock();
877 	ret = __xsk_generic_xmit(sk);
878 	/* Reaquire RCU lock before going into common code. */
879 	rcu_read_lock();
880 
881 	return ret;
882 }
883 
884 static bool xsk_no_wakeup(struct sock *sk)
885 {
886 #ifdef CONFIG_NET_RX_BUSY_POLL
887 	/* Prefer busy-polling, skip the wakeup. */
888 	return READ_ONCE(sk->sk_prefer_busy_poll) && READ_ONCE(sk->sk_ll_usec) &&
889 		READ_ONCE(sk->sk_napi_id) >= MIN_NAPI_ID;
890 #else
891 	return false;
892 #endif
893 }
894 
895 static int xsk_check_common(struct xdp_sock *xs)
896 {
897 	if (unlikely(!xsk_is_bound(xs)))
898 		return -ENXIO;
899 	if (unlikely(!(xs->dev->flags & IFF_UP)))
900 		return -ENETDOWN;
901 
902 	return 0;
903 }
904 
905 static int __xsk_sendmsg(struct socket *sock, struct msghdr *m, size_t total_len)
906 {
907 	bool need_wait = !(m->msg_flags & MSG_DONTWAIT);
908 	struct sock *sk = sock->sk;
909 	struct xdp_sock *xs = xdp_sk(sk);
910 	struct xsk_buff_pool *pool;
911 	int err;
912 
913 	err = xsk_check_common(xs);
914 	if (err)
915 		return err;
916 	if (unlikely(need_wait))
917 		return -EOPNOTSUPP;
918 	if (unlikely(!xs->tx))
919 		return -ENOBUFS;
920 
921 	if (sk_can_busy_loop(sk)) {
922 		if (xs->zc)
923 			__sk_mark_napi_id_once(sk, xsk_pool_get_napi_id(xs->pool));
924 		sk_busy_loop(sk, 1); /* only support non-blocking sockets */
925 	}
926 
927 	if (xs->zc && xsk_no_wakeup(sk))
928 		return 0;
929 
930 	pool = xs->pool;
931 	if (pool->cached_need_wakeup & XDP_WAKEUP_TX) {
932 		if (xs->zc)
933 			return xsk_wakeup(xs, XDP_WAKEUP_TX);
934 		return xsk_generic_xmit(sk);
935 	}
936 	return 0;
937 }
938 
939 static int xsk_sendmsg(struct socket *sock, struct msghdr *m, size_t total_len)
940 {
941 	int ret;
942 
943 	rcu_read_lock();
944 	ret = __xsk_sendmsg(sock, m, total_len);
945 	rcu_read_unlock();
946 
947 	return ret;
948 }
949 
950 static int __xsk_recvmsg(struct socket *sock, struct msghdr *m, size_t len, int flags)
951 {
952 	bool need_wait = !(flags & MSG_DONTWAIT);
953 	struct sock *sk = sock->sk;
954 	struct xdp_sock *xs = xdp_sk(sk);
955 	int err;
956 
957 	err = xsk_check_common(xs);
958 	if (err)
959 		return err;
960 	if (unlikely(!xs->rx))
961 		return -ENOBUFS;
962 	if (unlikely(need_wait))
963 		return -EOPNOTSUPP;
964 
965 	if (sk_can_busy_loop(sk))
966 		sk_busy_loop(sk, 1); /* only support non-blocking sockets */
967 
968 	if (xsk_no_wakeup(sk))
969 		return 0;
970 
971 	if (xs->pool->cached_need_wakeup & XDP_WAKEUP_RX && xs->zc)
972 		return xsk_wakeup(xs, XDP_WAKEUP_RX);
973 	return 0;
974 }
975 
976 static int xsk_recvmsg(struct socket *sock, struct msghdr *m, size_t len, int flags)
977 {
978 	int ret;
979 
980 	rcu_read_lock();
981 	ret = __xsk_recvmsg(sock, m, len, flags);
982 	rcu_read_unlock();
983 
984 	return ret;
985 }
986 
987 static __poll_t xsk_poll(struct file *file, struct socket *sock,
988 			     struct poll_table_struct *wait)
989 {
990 	__poll_t mask = 0;
991 	struct sock *sk = sock->sk;
992 	struct xdp_sock *xs = xdp_sk(sk);
993 	struct xsk_buff_pool *pool;
994 
995 	sock_poll_wait(file, sock, wait);
996 
997 	rcu_read_lock();
998 	if (xsk_check_common(xs))
999 		goto out;
1000 
1001 	pool = xs->pool;
1002 
1003 	if (pool->cached_need_wakeup) {
1004 		if (xs->zc)
1005 			xsk_wakeup(xs, pool->cached_need_wakeup);
1006 		else if (xs->tx)
1007 			/* Poll needs to drive Tx also in copy mode */
1008 			xsk_generic_xmit(sk);
1009 	}
1010 
1011 	if (xs->rx && !xskq_prod_is_empty(xs->rx))
1012 		mask |= EPOLLIN | EPOLLRDNORM;
1013 	if (xs->tx && xsk_tx_writeable(xs))
1014 		mask |= EPOLLOUT | EPOLLWRNORM;
1015 out:
1016 	rcu_read_unlock();
1017 	return mask;
1018 }
1019 
1020 static int xsk_init_queue(u32 entries, struct xsk_queue **queue,
1021 			  bool umem_queue)
1022 {
1023 	struct xsk_queue *q;
1024 
1025 	if (entries == 0 || *queue || !is_power_of_2(entries))
1026 		return -EINVAL;
1027 
1028 	q = xskq_create(entries, umem_queue);
1029 	if (!q)
1030 		return -ENOMEM;
1031 
1032 	/* Make sure queue is ready before it can be seen by others */
1033 	smp_wmb();
1034 	WRITE_ONCE(*queue, q);
1035 	return 0;
1036 }
1037 
1038 static void xsk_unbind_dev(struct xdp_sock *xs)
1039 {
1040 	struct net_device *dev = xs->dev;
1041 
1042 	if (xs->state != XSK_BOUND)
1043 		return;
1044 	WRITE_ONCE(xs->state, XSK_UNBOUND);
1045 
1046 	/* Wait for driver to stop using the xdp socket. */
1047 	xp_del_xsk(xs->pool, xs);
1048 	synchronize_net();
1049 	dev_put(dev);
1050 }
1051 
1052 static struct xsk_map *xsk_get_map_list_entry(struct xdp_sock *xs,
1053 					      struct xdp_sock __rcu ***map_entry)
1054 {
1055 	struct xsk_map *map = NULL;
1056 	struct xsk_map_node *node;
1057 
1058 	*map_entry = NULL;
1059 
1060 	spin_lock_bh(&xs->map_list_lock);
1061 	node = list_first_entry_or_null(&xs->map_list, struct xsk_map_node,
1062 					node);
1063 	if (node) {
1064 		bpf_map_inc(&node->map->map);
1065 		map = node->map;
1066 		*map_entry = node->map_entry;
1067 	}
1068 	spin_unlock_bh(&xs->map_list_lock);
1069 	return map;
1070 }
1071 
1072 static void xsk_delete_from_maps(struct xdp_sock *xs)
1073 {
1074 	/* This function removes the current XDP socket from all the
1075 	 * maps it resides in. We need to take extra care here, due to
1076 	 * the two locks involved. Each map has a lock synchronizing
1077 	 * updates to the entries, and each socket has a lock that
1078 	 * synchronizes access to the list of maps (map_list). For
1079 	 * deadlock avoidance the locks need to be taken in the order
1080 	 * "map lock"->"socket map list lock". We start off by
1081 	 * accessing the socket map list, and take a reference to the
1082 	 * map to guarantee existence between the
1083 	 * xsk_get_map_list_entry() and xsk_map_try_sock_delete()
1084 	 * calls. Then we ask the map to remove the socket, which
1085 	 * tries to remove the socket from the map. Note that there
1086 	 * might be updates to the map between
1087 	 * xsk_get_map_list_entry() and xsk_map_try_sock_delete().
1088 	 */
1089 	struct xdp_sock __rcu **map_entry = NULL;
1090 	struct xsk_map *map;
1091 
1092 	while ((map = xsk_get_map_list_entry(xs, &map_entry))) {
1093 		xsk_map_try_sock_delete(map, xs, map_entry);
1094 		bpf_map_put(&map->map);
1095 	}
1096 }
1097 
1098 static int xsk_release(struct socket *sock)
1099 {
1100 	struct sock *sk = sock->sk;
1101 	struct xdp_sock *xs = xdp_sk(sk);
1102 	struct net *net;
1103 
1104 	if (!sk)
1105 		return 0;
1106 
1107 	net = sock_net(sk);
1108 
1109 	if (xs->skb)
1110 		xsk_drop_skb(xs->skb);
1111 
1112 	mutex_lock(&net->xdp.lock);
1113 	sk_del_node_init_rcu(sk);
1114 	mutex_unlock(&net->xdp.lock);
1115 
1116 	sock_prot_inuse_add(net, sk->sk_prot, -1);
1117 
1118 	xsk_delete_from_maps(xs);
1119 	mutex_lock(&xs->mutex);
1120 	xsk_unbind_dev(xs);
1121 	mutex_unlock(&xs->mutex);
1122 
1123 	xskq_destroy(xs->rx);
1124 	xskq_destroy(xs->tx);
1125 	xskq_destroy(xs->fq_tmp);
1126 	xskq_destroy(xs->cq_tmp);
1127 
1128 	sock_orphan(sk);
1129 	sock->sk = NULL;
1130 
1131 	sock_put(sk);
1132 
1133 	return 0;
1134 }
1135 
1136 static struct socket *xsk_lookup_xsk_from_fd(int fd)
1137 {
1138 	struct socket *sock;
1139 	int err;
1140 
1141 	sock = sockfd_lookup(fd, &err);
1142 	if (!sock)
1143 		return ERR_PTR(-ENOTSOCK);
1144 
1145 	if (sock->sk->sk_family != PF_XDP) {
1146 		sockfd_put(sock);
1147 		return ERR_PTR(-ENOPROTOOPT);
1148 	}
1149 
1150 	return sock;
1151 }
1152 
1153 static bool xsk_validate_queues(struct xdp_sock *xs)
1154 {
1155 	return xs->fq_tmp && xs->cq_tmp;
1156 }
1157 
1158 static int xsk_bind(struct socket *sock, struct sockaddr *addr, int addr_len)
1159 {
1160 	struct sockaddr_xdp *sxdp = (struct sockaddr_xdp *)addr;
1161 	struct sock *sk = sock->sk;
1162 	struct xdp_sock *xs = xdp_sk(sk);
1163 	struct net_device *dev;
1164 	int bound_dev_if;
1165 	u32 flags, qid;
1166 	int err = 0;
1167 
1168 	if (addr_len < sizeof(struct sockaddr_xdp))
1169 		return -EINVAL;
1170 	if (sxdp->sxdp_family != AF_XDP)
1171 		return -EINVAL;
1172 
1173 	flags = sxdp->sxdp_flags;
1174 	if (flags & ~(XDP_SHARED_UMEM | XDP_COPY | XDP_ZEROCOPY |
1175 		      XDP_USE_NEED_WAKEUP | XDP_USE_SG))
1176 		return -EINVAL;
1177 
1178 	bound_dev_if = READ_ONCE(sk->sk_bound_dev_if);
1179 	if (bound_dev_if && bound_dev_if != sxdp->sxdp_ifindex)
1180 		return -EINVAL;
1181 
1182 	rtnl_lock();
1183 	mutex_lock(&xs->mutex);
1184 	if (xs->state != XSK_READY) {
1185 		err = -EBUSY;
1186 		goto out_release;
1187 	}
1188 
1189 	dev = dev_get_by_index(sock_net(sk), sxdp->sxdp_ifindex);
1190 	if (!dev) {
1191 		err = -ENODEV;
1192 		goto out_release;
1193 	}
1194 
1195 	if (!xs->rx && !xs->tx) {
1196 		err = -EINVAL;
1197 		goto out_unlock;
1198 	}
1199 
1200 	qid = sxdp->sxdp_queue_id;
1201 
1202 	if (flags & XDP_SHARED_UMEM) {
1203 		struct xdp_sock *umem_xs;
1204 		struct socket *sock;
1205 
1206 		if ((flags & XDP_COPY) || (flags & XDP_ZEROCOPY) ||
1207 		    (flags & XDP_USE_NEED_WAKEUP) || (flags & XDP_USE_SG)) {
1208 			/* Cannot specify flags for shared sockets. */
1209 			err = -EINVAL;
1210 			goto out_unlock;
1211 		}
1212 
1213 		if (xs->umem) {
1214 			/* We have already our own. */
1215 			err = -EINVAL;
1216 			goto out_unlock;
1217 		}
1218 
1219 		sock = xsk_lookup_xsk_from_fd(sxdp->sxdp_shared_umem_fd);
1220 		if (IS_ERR(sock)) {
1221 			err = PTR_ERR(sock);
1222 			goto out_unlock;
1223 		}
1224 
1225 		umem_xs = xdp_sk(sock->sk);
1226 		if (!xsk_is_bound(umem_xs)) {
1227 			err = -EBADF;
1228 			sockfd_put(sock);
1229 			goto out_unlock;
1230 		}
1231 
1232 		if (umem_xs->queue_id != qid || umem_xs->dev != dev) {
1233 			/* Share the umem with another socket on another qid
1234 			 * and/or device.
1235 			 */
1236 			xs->pool = xp_create_and_assign_umem(xs,
1237 							     umem_xs->umem);
1238 			if (!xs->pool) {
1239 				err = -ENOMEM;
1240 				sockfd_put(sock);
1241 				goto out_unlock;
1242 			}
1243 
1244 			err = xp_assign_dev_shared(xs->pool, umem_xs, dev,
1245 						   qid);
1246 			if (err) {
1247 				xp_destroy(xs->pool);
1248 				xs->pool = NULL;
1249 				sockfd_put(sock);
1250 				goto out_unlock;
1251 			}
1252 		} else {
1253 			/* Share the buffer pool with the other socket. */
1254 			if (xs->fq_tmp || xs->cq_tmp) {
1255 				/* Do not allow setting your own fq or cq. */
1256 				err = -EINVAL;
1257 				sockfd_put(sock);
1258 				goto out_unlock;
1259 			}
1260 
1261 			xp_get_pool(umem_xs->pool);
1262 			xs->pool = umem_xs->pool;
1263 
1264 			/* If underlying shared umem was created without Tx
1265 			 * ring, allocate Tx descs array that Tx batching API
1266 			 * utilizes
1267 			 */
1268 			if (xs->tx && !xs->pool->tx_descs) {
1269 				err = xp_alloc_tx_descs(xs->pool, xs);
1270 				if (err) {
1271 					xp_put_pool(xs->pool);
1272 					xs->pool = NULL;
1273 					sockfd_put(sock);
1274 					goto out_unlock;
1275 				}
1276 			}
1277 		}
1278 
1279 		xdp_get_umem(umem_xs->umem);
1280 		WRITE_ONCE(xs->umem, umem_xs->umem);
1281 		sockfd_put(sock);
1282 	} else if (!xs->umem || !xsk_validate_queues(xs)) {
1283 		err = -EINVAL;
1284 		goto out_unlock;
1285 	} else {
1286 		/* This xsk has its own umem. */
1287 		xs->pool = xp_create_and_assign_umem(xs, xs->umem);
1288 		if (!xs->pool) {
1289 			err = -ENOMEM;
1290 			goto out_unlock;
1291 		}
1292 
1293 		err = xp_assign_dev(xs->pool, dev, qid, flags);
1294 		if (err) {
1295 			xp_destroy(xs->pool);
1296 			xs->pool = NULL;
1297 			goto out_unlock;
1298 		}
1299 	}
1300 
1301 	/* FQ and CQ are now owned by the buffer pool and cleaned up with it. */
1302 	xs->fq_tmp = NULL;
1303 	xs->cq_tmp = NULL;
1304 
1305 	xs->dev = dev;
1306 	xs->zc = xs->umem->zc;
1307 	xs->sg = !!(xs->umem->flags & XDP_UMEM_SG_FLAG);
1308 	xs->queue_id = qid;
1309 	xp_add_xsk(xs->pool, xs);
1310 
1311 out_unlock:
1312 	if (err) {
1313 		dev_put(dev);
1314 	} else {
1315 		/* Matches smp_rmb() in bind() for shared umem
1316 		 * sockets, and xsk_is_bound().
1317 		 */
1318 		smp_wmb();
1319 		WRITE_ONCE(xs->state, XSK_BOUND);
1320 	}
1321 out_release:
1322 	mutex_unlock(&xs->mutex);
1323 	rtnl_unlock();
1324 	return err;
1325 }
1326 
1327 struct xdp_umem_reg_v1 {
1328 	__u64 addr; /* Start of packet data area */
1329 	__u64 len; /* Length of packet data area */
1330 	__u32 chunk_size;
1331 	__u32 headroom;
1332 };
1333 
1334 struct xdp_umem_reg_v2 {
1335 	__u64 addr; /* Start of packet data area */
1336 	__u64 len; /* Length of packet data area */
1337 	__u32 chunk_size;
1338 	__u32 headroom;
1339 	__u32 flags;
1340 };
1341 
1342 static int xsk_setsockopt(struct socket *sock, int level, int optname,
1343 			  sockptr_t optval, unsigned int optlen)
1344 {
1345 	struct sock *sk = sock->sk;
1346 	struct xdp_sock *xs = xdp_sk(sk);
1347 	int err;
1348 
1349 	if (level != SOL_XDP)
1350 		return -ENOPROTOOPT;
1351 
1352 	switch (optname) {
1353 	case XDP_RX_RING:
1354 	case XDP_TX_RING:
1355 	{
1356 		struct xsk_queue **q;
1357 		int entries;
1358 
1359 		if (optlen < sizeof(entries))
1360 			return -EINVAL;
1361 		if (copy_from_sockptr(&entries, optval, sizeof(entries)))
1362 			return -EFAULT;
1363 
1364 		mutex_lock(&xs->mutex);
1365 		if (xs->state != XSK_READY) {
1366 			mutex_unlock(&xs->mutex);
1367 			return -EBUSY;
1368 		}
1369 		q = (optname == XDP_TX_RING) ? &xs->tx : &xs->rx;
1370 		err = xsk_init_queue(entries, q, false);
1371 		if (!err && optname == XDP_TX_RING)
1372 			/* Tx needs to be explicitly woken up the first time */
1373 			xs->tx->ring->flags |= XDP_RING_NEED_WAKEUP;
1374 		mutex_unlock(&xs->mutex);
1375 		return err;
1376 	}
1377 	case XDP_UMEM_REG:
1378 	{
1379 		size_t mr_size = sizeof(struct xdp_umem_reg);
1380 		struct xdp_umem_reg mr = {};
1381 		struct xdp_umem *umem;
1382 
1383 		if (optlen < sizeof(struct xdp_umem_reg_v1))
1384 			return -EINVAL;
1385 		else if (optlen < sizeof(struct xdp_umem_reg_v2))
1386 			mr_size = sizeof(struct xdp_umem_reg_v1);
1387 		else if (optlen < sizeof(mr))
1388 			mr_size = sizeof(struct xdp_umem_reg_v2);
1389 
1390 		if (copy_from_sockptr(&mr, optval, mr_size))
1391 			return -EFAULT;
1392 
1393 		mutex_lock(&xs->mutex);
1394 		if (xs->state != XSK_READY || xs->umem) {
1395 			mutex_unlock(&xs->mutex);
1396 			return -EBUSY;
1397 		}
1398 
1399 		umem = xdp_umem_create(&mr);
1400 		if (IS_ERR(umem)) {
1401 			mutex_unlock(&xs->mutex);
1402 			return PTR_ERR(umem);
1403 		}
1404 
1405 		/* Make sure umem is ready before it can be seen by others */
1406 		smp_wmb();
1407 		WRITE_ONCE(xs->umem, umem);
1408 		mutex_unlock(&xs->mutex);
1409 		return 0;
1410 	}
1411 	case XDP_UMEM_FILL_RING:
1412 	case XDP_UMEM_COMPLETION_RING:
1413 	{
1414 		struct xsk_queue **q;
1415 		int entries;
1416 
1417 		if (copy_from_sockptr(&entries, optval, sizeof(entries)))
1418 			return -EFAULT;
1419 
1420 		mutex_lock(&xs->mutex);
1421 		if (xs->state != XSK_READY) {
1422 			mutex_unlock(&xs->mutex);
1423 			return -EBUSY;
1424 		}
1425 
1426 		q = (optname == XDP_UMEM_FILL_RING) ? &xs->fq_tmp :
1427 			&xs->cq_tmp;
1428 		err = xsk_init_queue(entries, q, true);
1429 		mutex_unlock(&xs->mutex);
1430 		return err;
1431 	}
1432 	default:
1433 		break;
1434 	}
1435 
1436 	return -ENOPROTOOPT;
1437 }
1438 
1439 static void xsk_enter_rxtx_offsets(struct xdp_ring_offset_v1 *ring)
1440 {
1441 	ring->producer = offsetof(struct xdp_rxtx_ring, ptrs.producer);
1442 	ring->consumer = offsetof(struct xdp_rxtx_ring, ptrs.consumer);
1443 	ring->desc = offsetof(struct xdp_rxtx_ring, desc);
1444 }
1445 
1446 static void xsk_enter_umem_offsets(struct xdp_ring_offset_v1 *ring)
1447 {
1448 	ring->producer = offsetof(struct xdp_umem_ring, ptrs.producer);
1449 	ring->consumer = offsetof(struct xdp_umem_ring, ptrs.consumer);
1450 	ring->desc = offsetof(struct xdp_umem_ring, desc);
1451 }
1452 
1453 struct xdp_statistics_v1 {
1454 	__u64 rx_dropped;
1455 	__u64 rx_invalid_descs;
1456 	__u64 tx_invalid_descs;
1457 };
1458 
1459 static int xsk_getsockopt(struct socket *sock, int level, int optname,
1460 			  char __user *optval, int __user *optlen)
1461 {
1462 	struct sock *sk = sock->sk;
1463 	struct xdp_sock *xs = xdp_sk(sk);
1464 	int len;
1465 
1466 	if (level != SOL_XDP)
1467 		return -ENOPROTOOPT;
1468 
1469 	if (get_user(len, optlen))
1470 		return -EFAULT;
1471 	if (len < 0)
1472 		return -EINVAL;
1473 
1474 	switch (optname) {
1475 	case XDP_STATISTICS:
1476 	{
1477 		struct xdp_statistics stats = {};
1478 		bool extra_stats = true;
1479 		size_t stats_size;
1480 
1481 		if (len < sizeof(struct xdp_statistics_v1)) {
1482 			return -EINVAL;
1483 		} else if (len < sizeof(stats)) {
1484 			extra_stats = false;
1485 			stats_size = sizeof(struct xdp_statistics_v1);
1486 		} else {
1487 			stats_size = sizeof(stats);
1488 		}
1489 
1490 		mutex_lock(&xs->mutex);
1491 		stats.rx_dropped = xs->rx_dropped;
1492 		if (extra_stats) {
1493 			stats.rx_ring_full = xs->rx_queue_full;
1494 			stats.rx_fill_ring_empty_descs =
1495 				xs->pool ? xskq_nb_queue_empty_descs(xs->pool->fq) : 0;
1496 			stats.tx_ring_empty_descs = xskq_nb_queue_empty_descs(xs->tx);
1497 		} else {
1498 			stats.rx_dropped += xs->rx_queue_full;
1499 		}
1500 		stats.rx_invalid_descs = xskq_nb_invalid_descs(xs->rx);
1501 		stats.tx_invalid_descs = xskq_nb_invalid_descs(xs->tx);
1502 		mutex_unlock(&xs->mutex);
1503 
1504 		if (copy_to_user(optval, &stats, stats_size))
1505 			return -EFAULT;
1506 		if (put_user(stats_size, optlen))
1507 			return -EFAULT;
1508 
1509 		return 0;
1510 	}
1511 	case XDP_MMAP_OFFSETS:
1512 	{
1513 		struct xdp_mmap_offsets off;
1514 		struct xdp_mmap_offsets_v1 off_v1;
1515 		bool flags_supported = true;
1516 		void *to_copy;
1517 
1518 		if (len < sizeof(off_v1))
1519 			return -EINVAL;
1520 		else if (len < sizeof(off))
1521 			flags_supported = false;
1522 
1523 		if (flags_supported) {
1524 			/* xdp_ring_offset is identical to xdp_ring_offset_v1
1525 			 * except for the flags field added to the end.
1526 			 */
1527 			xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *)
1528 					       &off.rx);
1529 			xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *)
1530 					       &off.tx);
1531 			xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *)
1532 					       &off.fr);
1533 			xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *)
1534 					       &off.cr);
1535 			off.rx.flags = offsetof(struct xdp_rxtx_ring,
1536 						ptrs.flags);
1537 			off.tx.flags = offsetof(struct xdp_rxtx_ring,
1538 						ptrs.flags);
1539 			off.fr.flags = offsetof(struct xdp_umem_ring,
1540 						ptrs.flags);
1541 			off.cr.flags = offsetof(struct xdp_umem_ring,
1542 						ptrs.flags);
1543 
1544 			len = sizeof(off);
1545 			to_copy = &off;
1546 		} else {
1547 			xsk_enter_rxtx_offsets(&off_v1.rx);
1548 			xsk_enter_rxtx_offsets(&off_v1.tx);
1549 			xsk_enter_umem_offsets(&off_v1.fr);
1550 			xsk_enter_umem_offsets(&off_v1.cr);
1551 
1552 			len = sizeof(off_v1);
1553 			to_copy = &off_v1;
1554 		}
1555 
1556 		if (copy_to_user(optval, to_copy, len))
1557 			return -EFAULT;
1558 		if (put_user(len, optlen))
1559 			return -EFAULT;
1560 
1561 		return 0;
1562 	}
1563 	case XDP_OPTIONS:
1564 	{
1565 		struct xdp_options opts = {};
1566 
1567 		if (len < sizeof(opts))
1568 			return -EINVAL;
1569 
1570 		mutex_lock(&xs->mutex);
1571 		if (xs->zc)
1572 			opts.flags |= XDP_OPTIONS_ZEROCOPY;
1573 		mutex_unlock(&xs->mutex);
1574 
1575 		len = sizeof(opts);
1576 		if (copy_to_user(optval, &opts, len))
1577 			return -EFAULT;
1578 		if (put_user(len, optlen))
1579 			return -EFAULT;
1580 
1581 		return 0;
1582 	}
1583 	default:
1584 		break;
1585 	}
1586 
1587 	return -EOPNOTSUPP;
1588 }
1589 
1590 static int xsk_mmap(struct file *file, struct socket *sock,
1591 		    struct vm_area_struct *vma)
1592 {
1593 	loff_t offset = (loff_t)vma->vm_pgoff << PAGE_SHIFT;
1594 	unsigned long size = vma->vm_end - vma->vm_start;
1595 	struct xdp_sock *xs = xdp_sk(sock->sk);
1596 	int state = READ_ONCE(xs->state);
1597 	struct xsk_queue *q = NULL;
1598 
1599 	if (state != XSK_READY && state != XSK_BOUND)
1600 		return -EBUSY;
1601 
1602 	if (offset == XDP_PGOFF_RX_RING) {
1603 		q = READ_ONCE(xs->rx);
1604 	} else if (offset == XDP_PGOFF_TX_RING) {
1605 		q = READ_ONCE(xs->tx);
1606 	} else {
1607 		/* Matches the smp_wmb() in XDP_UMEM_REG */
1608 		smp_rmb();
1609 		if (offset == XDP_UMEM_PGOFF_FILL_RING)
1610 			q = state == XSK_READY ? READ_ONCE(xs->fq_tmp) :
1611 						 READ_ONCE(xs->pool->fq);
1612 		else if (offset == XDP_UMEM_PGOFF_COMPLETION_RING)
1613 			q = state == XSK_READY ? READ_ONCE(xs->cq_tmp) :
1614 						 READ_ONCE(xs->pool->cq);
1615 	}
1616 
1617 	if (!q)
1618 		return -EINVAL;
1619 
1620 	/* Matches the smp_wmb() in xsk_init_queue */
1621 	smp_rmb();
1622 	if (size > q->ring_vmalloc_size)
1623 		return -EINVAL;
1624 
1625 	return remap_vmalloc_range(vma, q->ring, 0);
1626 }
1627 
1628 static int xsk_notifier(struct notifier_block *this,
1629 			unsigned long msg, void *ptr)
1630 {
1631 	struct net_device *dev = netdev_notifier_info_to_dev(ptr);
1632 	struct net *net = dev_net(dev);
1633 	struct sock *sk;
1634 
1635 	switch (msg) {
1636 	case NETDEV_UNREGISTER:
1637 		mutex_lock(&net->xdp.lock);
1638 		sk_for_each(sk, &net->xdp.list) {
1639 			struct xdp_sock *xs = xdp_sk(sk);
1640 
1641 			mutex_lock(&xs->mutex);
1642 			if (xs->dev == dev) {
1643 				sk->sk_err = ENETDOWN;
1644 				if (!sock_flag(sk, SOCK_DEAD))
1645 					sk_error_report(sk);
1646 
1647 				xsk_unbind_dev(xs);
1648 
1649 				/* Clear device references. */
1650 				xp_clear_dev(xs->pool);
1651 			}
1652 			mutex_unlock(&xs->mutex);
1653 		}
1654 		mutex_unlock(&net->xdp.lock);
1655 		break;
1656 	}
1657 	return NOTIFY_DONE;
1658 }
1659 
1660 static struct proto xsk_proto = {
1661 	.name =		"XDP",
1662 	.owner =	THIS_MODULE,
1663 	.obj_size =	sizeof(struct xdp_sock),
1664 };
1665 
1666 static const struct proto_ops xsk_proto_ops = {
1667 	.family		= PF_XDP,
1668 	.owner		= THIS_MODULE,
1669 	.release	= xsk_release,
1670 	.bind		= xsk_bind,
1671 	.connect	= sock_no_connect,
1672 	.socketpair	= sock_no_socketpair,
1673 	.accept		= sock_no_accept,
1674 	.getname	= sock_no_getname,
1675 	.poll		= xsk_poll,
1676 	.ioctl		= sock_no_ioctl,
1677 	.listen		= sock_no_listen,
1678 	.shutdown	= sock_no_shutdown,
1679 	.setsockopt	= xsk_setsockopt,
1680 	.getsockopt	= xsk_getsockopt,
1681 	.sendmsg	= xsk_sendmsg,
1682 	.recvmsg	= xsk_recvmsg,
1683 	.mmap		= xsk_mmap,
1684 };
1685 
1686 static void xsk_destruct(struct sock *sk)
1687 {
1688 	struct xdp_sock *xs = xdp_sk(sk);
1689 
1690 	if (!sock_flag(sk, SOCK_DEAD))
1691 		return;
1692 
1693 	if (!xp_put_pool(xs->pool))
1694 		xdp_put_umem(xs->umem, !xs->pool);
1695 }
1696 
1697 static int xsk_create(struct net *net, struct socket *sock, int protocol,
1698 		      int kern)
1699 {
1700 	struct xdp_sock *xs;
1701 	struct sock *sk;
1702 
1703 	if (!ns_capable(net->user_ns, CAP_NET_RAW))
1704 		return -EPERM;
1705 	if (sock->type != SOCK_RAW)
1706 		return -ESOCKTNOSUPPORT;
1707 
1708 	if (protocol)
1709 		return -EPROTONOSUPPORT;
1710 
1711 	sock->state = SS_UNCONNECTED;
1712 
1713 	sk = sk_alloc(net, PF_XDP, GFP_KERNEL, &xsk_proto, kern);
1714 	if (!sk)
1715 		return -ENOBUFS;
1716 
1717 	sock->ops = &xsk_proto_ops;
1718 
1719 	sock_init_data(sock, sk);
1720 
1721 	sk->sk_family = PF_XDP;
1722 
1723 	sk->sk_destruct = xsk_destruct;
1724 
1725 	sock_set_flag(sk, SOCK_RCU_FREE);
1726 
1727 	xs = xdp_sk(sk);
1728 	xs->state = XSK_READY;
1729 	mutex_init(&xs->mutex);
1730 	spin_lock_init(&xs->rx_lock);
1731 
1732 	INIT_LIST_HEAD(&xs->map_list);
1733 	spin_lock_init(&xs->map_list_lock);
1734 
1735 	mutex_lock(&net->xdp.lock);
1736 	sk_add_node_rcu(sk, &net->xdp.list);
1737 	mutex_unlock(&net->xdp.lock);
1738 
1739 	sock_prot_inuse_add(net, &xsk_proto, 1);
1740 
1741 	return 0;
1742 }
1743 
1744 static const struct net_proto_family xsk_family_ops = {
1745 	.family = PF_XDP,
1746 	.create = xsk_create,
1747 	.owner	= THIS_MODULE,
1748 };
1749 
1750 static struct notifier_block xsk_netdev_notifier = {
1751 	.notifier_call	= xsk_notifier,
1752 };
1753 
1754 static int __net_init xsk_net_init(struct net *net)
1755 {
1756 	mutex_init(&net->xdp.lock);
1757 	INIT_HLIST_HEAD(&net->xdp.list);
1758 	return 0;
1759 }
1760 
1761 static void __net_exit xsk_net_exit(struct net *net)
1762 {
1763 	WARN_ON_ONCE(!hlist_empty(&net->xdp.list));
1764 }
1765 
1766 static struct pernet_operations xsk_net_ops = {
1767 	.init = xsk_net_init,
1768 	.exit = xsk_net_exit,
1769 };
1770 
1771 static int __init xsk_init(void)
1772 {
1773 	int err, cpu;
1774 
1775 	err = proto_register(&xsk_proto, 0 /* no slab */);
1776 	if (err)
1777 		goto out;
1778 
1779 	err = sock_register(&xsk_family_ops);
1780 	if (err)
1781 		goto out_proto;
1782 
1783 	err = register_pernet_subsys(&xsk_net_ops);
1784 	if (err)
1785 		goto out_sk;
1786 
1787 	err = register_netdevice_notifier(&xsk_netdev_notifier);
1788 	if (err)
1789 		goto out_pernet;
1790 
1791 	for_each_possible_cpu(cpu)
1792 		INIT_LIST_HEAD(&per_cpu(xskmap_flush_list, cpu));
1793 	return 0;
1794 
1795 out_pernet:
1796 	unregister_pernet_subsys(&xsk_net_ops);
1797 out_sk:
1798 	sock_unregister(PF_XDP);
1799 out_proto:
1800 	proto_unregister(&xsk_proto);
1801 out:
1802 	return err;
1803 }
1804 
1805 fs_initcall(xsk_init);
1806