xref: /linux/net/xdp/xsk.c (revision 79997eda0d31bc68203c95ecb978773ee6ce7a1f)
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 || likely(!frags))
171 		goto out;
172 
173 	xskb_list = &xskb->pool->xskb_list;
174 	list_for_each_entry_safe(pos, tmp, xskb_list, xskb_list_node) {
175 		if (list_is_singular(xskb_list))
176 			contd = 0;
177 		len = pos->xdp.data_end - pos->xdp.data;
178 		err = __xsk_rcv_zc(xs, pos, len, contd);
179 		if (err)
180 			return err;
181 		list_del(&pos->xskb_list_node);
182 	}
183 
184 out:
185 	return err;
186 }
187 
188 static void *xsk_copy_xdp_start(struct xdp_buff *from)
189 {
190 	if (unlikely(xdp_data_meta_unsupported(from)))
191 		return from->data;
192 	else
193 		return from->data_meta;
194 }
195 
196 static u32 xsk_copy_xdp(void *to, void **from, u32 to_len,
197 			u32 *from_len, skb_frag_t **frag, u32 rem)
198 {
199 	u32 copied = 0;
200 
201 	while (1) {
202 		u32 copy_len = min_t(u32, *from_len, to_len);
203 
204 		memcpy(to, *from, copy_len);
205 		copied += copy_len;
206 		if (rem == copied)
207 			return copied;
208 
209 		if (*from_len == copy_len) {
210 			*from = skb_frag_address(*frag);
211 			*from_len = skb_frag_size((*frag)++);
212 		} else {
213 			*from += copy_len;
214 			*from_len -= copy_len;
215 		}
216 		if (to_len == copy_len)
217 			return copied;
218 
219 		to_len -= copy_len;
220 		to += copy_len;
221 	}
222 }
223 
224 static int __xsk_rcv(struct xdp_sock *xs, struct xdp_buff *xdp, u32 len)
225 {
226 	u32 frame_size = xsk_pool_get_rx_frame_size(xs->pool);
227 	void *copy_from = xsk_copy_xdp_start(xdp), *copy_to;
228 	u32 from_len, meta_len, rem, num_desc;
229 	struct xdp_buff_xsk *xskb;
230 	struct xdp_buff *xsk_xdp;
231 	skb_frag_t *frag;
232 
233 	from_len = xdp->data_end - copy_from;
234 	meta_len = xdp->data - copy_from;
235 	rem = len + meta_len;
236 
237 	if (len <= frame_size && !xdp_buff_has_frags(xdp)) {
238 		int err;
239 
240 		xsk_xdp = xsk_buff_alloc(xs->pool);
241 		if (!xsk_xdp) {
242 			xs->rx_dropped++;
243 			return -ENOMEM;
244 		}
245 		memcpy(xsk_xdp->data - meta_len, copy_from, rem);
246 		xskb = container_of(xsk_xdp, struct xdp_buff_xsk, xdp);
247 		err = __xsk_rcv_zc(xs, xskb, len, 0);
248 		if (err) {
249 			xsk_buff_free(xsk_xdp);
250 			return err;
251 		}
252 
253 		return 0;
254 	}
255 
256 	num_desc = (len - 1) / frame_size + 1;
257 
258 	if (!xsk_buff_can_alloc(xs->pool, num_desc)) {
259 		xs->rx_dropped++;
260 		return -ENOMEM;
261 	}
262 	if (xskq_prod_nb_free(xs->rx, num_desc) < num_desc) {
263 		xs->rx_queue_full++;
264 		return -ENOBUFS;
265 	}
266 
267 	if (xdp_buff_has_frags(xdp)) {
268 		struct skb_shared_info *sinfo;
269 
270 		sinfo = xdp_get_shared_info_from_buff(xdp);
271 		frag =  &sinfo->frags[0];
272 	}
273 
274 	do {
275 		u32 to_len = frame_size + meta_len;
276 		u32 copied;
277 
278 		xsk_xdp = xsk_buff_alloc(xs->pool);
279 		copy_to = xsk_xdp->data - meta_len;
280 
281 		copied = xsk_copy_xdp(copy_to, &copy_from, to_len, &from_len, &frag, rem);
282 		rem -= copied;
283 
284 		xskb = container_of(xsk_xdp, struct xdp_buff_xsk, xdp);
285 		__xsk_rcv_zc(xs, xskb, copied - meta_len, rem ? XDP_PKT_CONTD : 0);
286 		meta_len = 0;
287 	} while (rem);
288 
289 	return 0;
290 }
291 
292 static bool xsk_tx_writeable(struct xdp_sock *xs)
293 {
294 	if (xskq_cons_present_entries(xs->tx) > xs->tx->nentries / 2)
295 		return false;
296 
297 	return true;
298 }
299 
300 static bool xsk_is_bound(struct xdp_sock *xs)
301 {
302 	if (READ_ONCE(xs->state) == XSK_BOUND) {
303 		/* Matches smp_wmb() in bind(). */
304 		smp_rmb();
305 		return true;
306 	}
307 	return false;
308 }
309 
310 static int xsk_rcv_check(struct xdp_sock *xs, struct xdp_buff *xdp, u32 len)
311 {
312 	if (!xsk_is_bound(xs))
313 		return -ENXIO;
314 
315 	if (xs->dev != xdp->rxq->dev || xs->queue_id != xdp->rxq->queue_index)
316 		return -EINVAL;
317 
318 	if (len > xsk_pool_get_rx_frame_size(xs->pool) && !xs->sg) {
319 		xs->rx_dropped++;
320 		return -ENOSPC;
321 	}
322 
323 	sk_mark_napi_id_once_xdp(&xs->sk, xdp);
324 	return 0;
325 }
326 
327 static void xsk_flush(struct xdp_sock *xs)
328 {
329 	xskq_prod_submit(xs->rx);
330 	__xskq_cons_release(xs->pool->fq);
331 	sock_def_readable(&xs->sk);
332 }
333 
334 int xsk_generic_rcv(struct xdp_sock *xs, struct xdp_buff *xdp)
335 {
336 	u32 len = xdp_get_buff_len(xdp);
337 	int err;
338 
339 	spin_lock_bh(&xs->rx_lock);
340 	err = xsk_rcv_check(xs, xdp, len);
341 	if (!err) {
342 		err = __xsk_rcv(xs, xdp, len);
343 		xsk_flush(xs);
344 	}
345 	spin_unlock_bh(&xs->rx_lock);
346 	return err;
347 }
348 
349 static int xsk_rcv(struct xdp_sock *xs, struct xdp_buff *xdp)
350 {
351 	u32 len = xdp_get_buff_len(xdp);
352 	int err;
353 
354 	err = xsk_rcv_check(xs, xdp, len);
355 	if (err)
356 		return err;
357 
358 	if (xdp->rxq->mem.type == MEM_TYPE_XSK_BUFF_POOL) {
359 		len = xdp->data_end - xdp->data;
360 		return xsk_rcv_zc(xs, xdp, len);
361 	}
362 
363 	err = __xsk_rcv(xs, xdp, len);
364 	if (!err)
365 		xdp_return_buff(xdp);
366 	return err;
367 }
368 
369 int __xsk_map_redirect(struct xdp_sock *xs, struct xdp_buff *xdp)
370 {
371 	struct list_head *flush_list = this_cpu_ptr(&xskmap_flush_list);
372 	int err;
373 
374 	err = xsk_rcv(xs, xdp);
375 	if (err)
376 		return err;
377 
378 	if (!xs->flush_node.prev)
379 		list_add(&xs->flush_node, flush_list);
380 
381 	return 0;
382 }
383 
384 void __xsk_map_flush(void)
385 {
386 	struct list_head *flush_list = this_cpu_ptr(&xskmap_flush_list);
387 	struct xdp_sock *xs, *tmp;
388 
389 	list_for_each_entry_safe(xs, tmp, flush_list, flush_node) {
390 		xsk_flush(xs);
391 		__list_del_clearprev(&xs->flush_node);
392 	}
393 }
394 
395 #ifdef CONFIG_DEBUG_NET
396 bool xsk_map_check_flush(void)
397 {
398 	if (list_empty(this_cpu_ptr(&xskmap_flush_list)))
399 		return false;
400 	__xsk_map_flush();
401 	return true;
402 }
403 #endif
404 
405 void xsk_tx_completed(struct xsk_buff_pool *pool, u32 nb_entries)
406 {
407 	xskq_prod_submit_n(pool->cq, nb_entries);
408 }
409 EXPORT_SYMBOL(xsk_tx_completed);
410 
411 void xsk_tx_release(struct xsk_buff_pool *pool)
412 {
413 	struct xdp_sock *xs;
414 
415 	rcu_read_lock();
416 	list_for_each_entry_rcu(xs, &pool->xsk_tx_list, tx_list) {
417 		__xskq_cons_release(xs->tx);
418 		if (xsk_tx_writeable(xs))
419 			xs->sk.sk_write_space(&xs->sk);
420 	}
421 	rcu_read_unlock();
422 }
423 EXPORT_SYMBOL(xsk_tx_release);
424 
425 bool xsk_tx_peek_desc(struct xsk_buff_pool *pool, struct xdp_desc *desc)
426 {
427 	bool budget_exhausted = false;
428 	struct xdp_sock *xs;
429 
430 	rcu_read_lock();
431 again:
432 	list_for_each_entry_rcu(xs, &pool->xsk_tx_list, tx_list) {
433 		if (xs->tx_budget_spent >= MAX_PER_SOCKET_BUDGET) {
434 			budget_exhausted = true;
435 			continue;
436 		}
437 
438 		if (!xskq_cons_peek_desc(xs->tx, desc, pool)) {
439 			if (xskq_has_descs(xs->tx))
440 				xskq_cons_release(xs->tx);
441 			continue;
442 		}
443 
444 		xs->tx_budget_spent++;
445 
446 		/* This is the backpressure mechanism for the Tx path.
447 		 * Reserve space in the completion queue and only proceed
448 		 * if there is space in it. This avoids having to implement
449 		 * any buffering in the Tx path.
450 		 */
451 		if (xskq_prod_reserve_addr(pool->cq, desc->addr))
452 			goto out;
453 
454 		xskq_cons_release(xs->tx);
455 		rcu_read_unlock();
456 		return true;
457 	}
458 
459 	if (budget_exhausted) {
460 		list_for_each_entry_rcu(xs, &pool->xsk_tx_list, tx_list)
461 			xs->tx_budget_spent = 0;
462 
463 		budget_exhausted = false;
464 		goto again;
465 	}
466 
467 out:
468 	rcu_read_unlock();
469 	return false;
470 }
471 EXPORT_SYMBOL(xsk_tx_peek_desc);
472 
473 static u32 xsk_tx_peek_release_fallback(struct xsk_buff_pool *pool, u32 max_entries)
474 {
475 	struct xdp_desc *descs = pool->tx_descs;
476 	u32 nb_pkts = 0;
477 
478 	while (nb_pkts < max_entries && xsk_tx_peek_desc(pool, &descs[nb_pkts]))
479 		nb_pkts++;
480 
481 	xsk_tx_release(pool);
482 	return nb_pkts;
483 }
484 
485 u32 xsk_tx_peek_release_desc_batch(struct xsk_buff_pool *pool, u32 nb_pkts)
486 {
487 	struct xdp_sock *xs;
488 
489 	rcu_read_lock();
490 	if (!list_is_singular(&pool->xsk_tx_list)) {
491 		/* Fallback to the non-batched version */
492 		rcu_read_unlock();
493 		return xsk_tx_peek_release_fallback(pool, nb_pkts);
494 	}
495 
496 	xs = list_first_or_null_rcu(&pool->xsk_tx_list, struct xdp_sock, tx_list);
497 	if (!xs) {
498 		nb_pkts = 0;
499 		goto out;
500 	}
501 
502 	nb_pkts = xskq_cons_nb_entries(xs->tx, nb_pkts);
503 
504 	/* This is the backpressure mechanism for the Tx path. Try to
505 	 * reserve space in the completion queue for all packets, but
506 	 * if there are fewer slots available, just process that many
507 	 * packets. This avoids having to implement any buffering in
508 	 * the Tx path.
509 	 */
510 	nb_pkts = xskq_prod_nb_free(pool->cq, nb_pkts);
511 	if (!nb_pkts)
512 		goto out;
513 
514 	nb_pkts = xskq_cons_read_desc_batch(xs->tx, pool, nb_pkts);
515 	if (!nb_pkts) {
516 		xs->tx->queue_empty_descs++;
517 		goto out;
518 	}
519 
520 	__xskq_cons_release(xs->tx);
521 	xskq_prod_write_addr_batch(pool->cq, pool->tx_descs, nb_pkts);
522 	xs->sk.sk_write_space(&xs->sk);
523 
524 out:
525 	rcu_read_unlock();
526 	return nb_pkts;
527 }
528 EXPORT_SYMBOL(xsk_tx_peek_release_desc_batch);
529 
530 static int xsk_wakeup(struct xdp_sock *xs, u8 flags)
531 {
532 	struct net_device *dev = xs->dev;
533 
534 	return dev->netdev_ops->ndo_xsk_wakeup(dev, xs->queue_id, flags);
535 }
536 
537 static int xsk_cq_reserve_addr_locked(struct xdp_sock *xs, u64 addr)
538 {
539 	unsigned long flags;
540 	int ret;
541 
542 	spin_lock_irqsave(&xs->pool->cq_lock, flags);
543 	ret = xskq_prod_reserve_addr(xs->pool->cq, addr);
544 	spin_unlock_irqrestore(&xs->pool->cq_lock, flags);
545 
546 	return ret;
547 }
548 
549 static void xsk_cq_submit_locked(struct xdp_sock *xs, u32 n)
550 {
551 	unsigned long flags;
552 
553 	spin_lock_irqsave(&xs->pool->cq_lock, flags);
554 	xskq_prod_submit_n(xs->pool->cq, n);
555 	spin_unlock_irqrestore(&xs->pool->cq_lock, flags);
556 }
557 
558 static void xsk_cq_cancel_locked(struct xdp_sock *xs, u32 n)
559 {
560 	unsigned long flags;
561 
562 	spin_lock_irqsave(&xs->pool->cq_lock, flags);
563 	xskq_prod_cancel_n(xs->pool->cq, n);
564 	spin_unlock_irqrestore(&xs->pool->cq_lock, flags);
565 }
566 
567 static u32 xsk_get_num_desc(struct sk_buff *skb)
568 {
569 	return skb ? (long)skb_shinfo(skb)->destructor_arg : 0;
570 }
571 
572 static void xsk_destruct_skb(struct sk_buff *skb)
573 {
574 	xsk_cq_submit_locked(xdp_sk(skb->sk), xsk_get_num_desc(skb));
575 	sock_wfree(skb);
576 }
577 
578 static void xsk_set_destructor_arg(struct sk_buff *skb)
579 {
580 	long num = xsk_get_num_desc(xdp_sk(skb->sk)->skb) + 1;
581 
582 	skb_shinfo(skb)->destructor_arg = (void *)num;
583 }
584 
585 static void xsk_consume_skb(struct sk_buff *skb)
586 {
587 	struct xdp_sock *xs = xdp_sk(skb->sk);
588 
589 	skb->destructor = sock_wfree;
590 	xsk_cq_cancel_locked(xs, xsk_get_num_desc(skb));
591 	/* Free skb without triggering the perf drop trace */
592 	consume_skb(skb);
593 	xs->skb = NULL;
594 }
595 
596 static void xsk_drop_skb(struct sk_buff *skb)
597 {
598 	xdp_sk(skb->sk)->tx->invalid_descs += xsk_get_num_desc(skb);
599 	xsk_consume_skb(skb);
600 }
601 
602 static struct sk_buff *xsk_build_skb_zerocopy(struct xdp_sock *xs,
603 					      struct xdp_desc *desc)
604 {
605 	struct xsk_buff_pool *pool = xs->pool;
606 	u32 hr, len, ts, offset, copy, copied;
607 	struct sk_buff *skb = xs->skb;
608 	struct page *page;
609 	void *buffer;
610 	int err, i;
611 	u64 addr;
612 
613 	if (!skb) {
614 		hr = max(NET_SKB_PAD, L1_CACHE_ALIGN(xs->dev->needed_headroom));
615 
616 		skb = sock_alloc_send_skb(&xs->sk, hr, 1, &err);
617 		if (unlikely(!skb))
618 			return ERR_PTR(err);
619 
620 		skb_reserve(skb, hr);
621 	}
622 
623 	addr = desc->addr;
624 	len = desc->len;
625 	ts = pool->unaligned ? len : pool->chunk_size;
626 
627 	buffer = xsk_buff_raw_get_data(pool, addr);
628 	offset = offset_in_page(buffer);
629 	addr = buffer - pool->addrs;
630 
631 	for (copied = 0, i = skb_shinfo(skb)->nr_frags; copied < len; i++) {
632 		if (unlikely(i >= MAX_SKB_FRAGS))
633 			return ERR_PTR(-EOVERFLOW);
634 
635 		page = pool->umem->pgs[addr >> PAGE_SHIFT];
636 		get_page(page);
637 
638 		copy = min_t(u32, PAGE_SIZE - offset, len - copied);
639 		skb_fill_page_desc(skb, i, page, offset, copy);
640 
641 		copied += copy;
642 		addr += copy;
643 		offset = 0;
644 	}
645 
646 	skb->len += len;
647 	skb->data_len += len;
648 	skb->truesize += ts;
649 
650 	refcount_add(ts, &xs->sk.sk_wmem_alloc);
651 
652 	return skb;
653 }
654 
655 static struct sk_buff *xsk_build_skb(struct xdp_sock *xs,
656 				     struct xdp_desc *desc)
657 {
658 	struct net_device *dev = xs->dev;
659 	struct sk_buff *skb = xs->skb;
660 	int err;
661 
662 	if (dev->priv_flags & IFF_TX_SKB_NO_LINEAR) {
663 		skb = xsk_build_skb_zerocopy(xs, desc);
664 		if (IS_ERR(skb)) {
665 			err = PTR_ERR(skb);
666 			goto free_err;
667 		}
668 	} else {
669 		u32 hr, tr, len;
670 		void *buffer;
671 
672 		buffer = xsk_buff_raw_get_data(xs->pool, desc->addr);
673 		len = desc->len;
674 
675 		if (!skb) {
676 			hr = max(NET_SKB_PAD, L1_CACHE_ALIGN(dev->needed_headroom));
677 			tr = dev->needed_tailroom;
678 			skb = sock_alloc_send_skb(&xs->sk, hr + len + tr, 1, &err);
679 			if (unlikely(!skb))
680 				goto free_err;
681 
682 			skb_reserve(skb, hr);
683 			skb_put(skb, len);
684 
685 			err = skb_store_bits(skb, 0, buffer, len);
686 			if (unlikely(err)) {
687 				kfree_skb(skb);
688 				goto free_err;
689 			}
690 		} else {
691 			int nr_frags = skb_shinfo(skb)->nr_frags;
692 			struct page *page;
693 			u8 *vaddr;
694 
695 			if (unlikely(nr_frags == (MAX_SKB_FRAGS - 1) && xp_mb_desc(desc))) {
696 				err = -EOVERFLOW;
697 				goto free_err;
698 			}
699 
700 			page = alloc_page(xs->sk.sk_allocation);
701 			if (unlikely(!page)) {
702 				err = -EAGAIN;
703 				goto free_err;
704 			}
705 
706 			vaddr = kmap_local_page(page);
707 			memcpy(vaddr, buffer, len);
708 			kunmap_local(vaddr);
709 
710 			skb_add_rx_frag(skb, nr_frags, page, 0, len, 0);
711 		}
712 	}
713 
714 	skb->dev = dev;
715 	skb->priority = READ_ONCE(xs->sk.sk_priority);
716 	skb->mark = READ_ONCE(xs->sk.sk_mark);
717 	skb->destructor = xsk_destruct_skb;
718 	xsk_set_destructor_arg(skb);
719 
720 	return skb;
721 
722 free_err:
723 	if (err == -EOVERFLOW) {
724 		/* Drop the packet */
725 		xsk_set_destructor_arg(xs->skb);
726 		xsk_drop_skb(xs->skb);
727 		xskq_cons_release(xs->tx);
728 	} else {
729 		/* Let application retry */
730 		xsk_cq_cancel_locked(xs, 1);
731 	}
732 
733 	return ERR_PTR(err);
734 }
735 
736 static int __xsk_generic_xmit(struct sock *sk)
737 {
738 	struct xdp_sock *xs = xdp_sk(sk);
739 	u32 max_batch = TX_BATCH_SIZE;
740 	bool sent_frame = false;
741 	struct xdp_desc desc;
742 	struct sk_buff *skb;
743 	int err = 0;
744 
745 	mutex_lock(&xs->mutex);
746 
747 	/* Since we dropped the RCU read lock, the socket state might have changed. */
748 	if (unlikely(!xsk_is_bound(xs))) {
749 		err = -ENXIO;
750 		goto out;
751 	}
752 
753 	if (xs->queue_id >= xs->dev->real_num_tx_queues)
754 		goto out;
755 
756 	while (xskq_cons_peek_desc(xs->tx, &desc, xs->pool)) {
757 		if (max_batch-- == 0) {
758 			err = -EAGAIN;
759 			goto out;
760 		}
761 
762 		/* This is the backpressure mechanism for the Tx path.
763 		 * Reserve space in the completion queue and only proceed
764 		 * if there is space in it. This avoids having to implement
765 		 * any buffering in the Tx path.
766 		 */
767 		if (xsk_cq_reserve_addr_locked(xs, desc.addr))
768 			goto out;
769 
770 		skb = xsk_build_skb(xs, &desc);
771 		if (IS_ERR(skb)) {
772 			err = PTR_ERR(skb);
773 			if (err != -EOVERFLOW)
774 				goto out;
775 			err = 0;
776 			continue;
777 		}
778 
779 		xskq_cons_release(xs->tx);
780 
781 		if (xp_mb_desc(&desc)) {
782 			xs->skb = skb;
783 			continue;
784 		}
785 
786 		err = __dev_direct_xmit(skb, xs->queue_id);
787 		if  (err == NETDEV_TX_BUSY) {
788 			/* Tell user-space to retry the send */
789 			xskq_cons_cancel_n(xs->tx, xsk_get_num_desc(skb));
790 			xsk_consume_skb(skb);
791 			err = -EAGAIN;
792 			goto out;
793 		}
794 
795 		/* Ignore NET_XMIT_CN as packet might have been sent */
796 		if (err == NET_XMIT_DROP) {
797 			/* SKB completed but not sent */
798 			err = -EBUSY;
799 			xs->skb = NULL;
800 			goto out;
801 		}
802 
803 		sent_frame = true;
804 		xs->skb = NULL;
805 	}
806 
807 	if (xskq_has_descs(xs->tx)) {
808 		if (xs->skb)
809 			xsk_drop_skb(xs->skb);
810 		xskq_cons_release(xs->tx);
811 	}
812 
813 out:
814 	if (sent_frame)
815 		if (xsk_tx_writeable(xs))
816 			sk->sk_write_space(sk);
817 
818 	mutex_unlock(&xs->mutex);
819 	return err;
820 }
821 
822 static int xsk_generic_xmit(struct sock *sk)
823 {
824 	int ret;
825 
826 	/* Drop the RCU lock since the SKB path might sleep. */
827 	rcu_read_unlock();
828 	ret = __xsk_generic_xmit(sk);
829 	/* Reaquire RCU lock before going into common code. */
830 	rcu_read_lock();
831 
832 	return ret;
833 }
834 
835 static bool xsk_no_wakeup(struct sock *sk)
836 {
837 #ifdef CONFIG_NET_RX_BUSY_POLL
838 	/* Prefer busy-polling, skip the wakeup. */
839 	return READ_ONCE(sk->sk_prefer_busy_poll) && READ_ONCE(sk->sk_ll_usec) &&
840 		READ_ONCE(sk->sk_napi_id) >= MIN_NAPI_ID;
841 #else
842 	return false;
843 #endif
844 }
845 
846 static int xsk_check_common(struct xdp_sock *xs)
847 {
848 	if (unlikely(!xsk_is_bound(xs)))
849 		return -ENXIO;
850 	if (unlikely(!(xs->dev->flags & IFF_UP)))
851 		return -ENETDOWN;
852 
853 	return 0;
854 }
855 
856 static int __xsk_sendmsg(struct socket *sock, struct msghdr *m, size_t total_len)
857 {
858 	bool need_wait = !(m->msg_flags & MSG_DONTWAIT);
859 	struct sock *sk = sock->sk;
860 	struct xdp_sock *xs = xdp_sk(sk);
861 	struct xsk_buff_pool *pool;
862 	int err;
863 
864 	err = xsk_check_common(xs);
865 	if (err)
866 		return err;
867 	if (unlikely(need_wait))
868 		return -EOPNOTSUPP;
869 	if (unlikely(!xs->tx))
870 		return -ENOBUFS;
871 
872 	if (sk_can_busy_loop(sk)) {
873 		if (xs->zc)
874 			__sk_mark_napi_id_once(sk, xsk_pool_get_napi_id(xs->pool));
875 		sk_busy_loop(sk, 1); /* only support non-blocking sockets */
876 	}
877 
878 	if (xs->zc && xsk_no_wakeup(sk))
879 		return 0;
880 
881 	pool = xs->pool;
882 	if (pool->cached_need_wakeup & XDP_WAKEUP_TX) {
883 		if (xs->zc)
884 			return xsk_wakeup(xs, XDP_WAKEUP_TX);
885 		return xsk_generic_xmit(sk);
886 	}
887 	return 0;
888 }
889 
890 static int xsk_sendmsg(struct socket *sock, struct msghdr *m, size_t total_len)
891 {
892 	int ret;
893 
894 	rcu_read_lock();
895 	ret = __xsk_sendmsg(sock, m, total_len);
896 	rcu_read_unlock();
897 
898 	return ret;
899 }
900 
901 static int __xsk_recvmsg(struct socket *sock, struct msghdr *m, size_t len, int flags)
902 {
903 	bool need_wait = !(flags & MSG_DONTWAIT);
904 	struct sock *sk = sock->sk;
905 	struct xdp_sock *xs = xdp_sk(sk);
906 	int err;
907 
908 	err = xsk_check_common(xs);
909 	if (err)
910 		return err;
911 	if (unlikely(!xs->rx))
912 		return -ENOBUFS;
913 	if (unlikely(need_wait))
914 		return -EOPNOTSUPP;
915 
916 	if (sk_can_busy_loop(sk))
917 		sk_busy_loop(sk, 1); /* only support non-blocking sockets */
918 
919 	if (xsk_no_wakeup(sk))
920 		return 0;
921 
922 	if (xs->pool->cached_need_wakeup & XDP_WAKEUP_RX && xs->zc)
923 		return xsk_wakeup(xs, XDP_WAKEUP_RX);
924 	return 0;
925 }
926 
927 static int xsk_recvmsg(struct socket *sock, struct msghdr *m, size_t len, int flags)
928 {
929 	int ret;
930 
931 	rcu_read_lock();
932 	ret = __xsk_recvmsg(sock, m, len, flags);
933 	rcu_read_unlock();
934 
935 	return ret;
936 }
937 
938 static __poll_t xsk_poll(struct file *file, struct socket *sock,
939 			     struct poll_table_struct *wait)
940 {
941 	__poll_t mask = 0;
942 	struct sock *sk = sock->sk;
943 	struct xdp_sock *xs = xdp_sk(sk);
944 	struct xsk_buff_pool *pool;
945 
946 	sock_poll_wait(file, sock, wait);
947 
948 	rcu_read_lock();
949 	if (xsk_check_common(xs))
950 		goto skip_tx;
951 
952 	pool = xs->pool;
953 
954 	if (pool->cached_need_wakeup) {
955 		if (xs->zc)
956 			xsk_wakeup(xs, pool->cached_need_wakeup);
957 		else if (xs->tx)
958 			/* Poll needs to drive Tx also in copy mode */
959 			xsk_generic_xmit(sk);
960 	}
961 
962 skip_tx:
963 	if (xs->rx && !xskq_prod_is_empty(xs->rx))
964 		mask |= EPOLLIN | EPOLLRDNORM;
965 	if (xs->tx && xsk_tx_writeable(xs))
966 		mask |= EPOLLOUT | EPOLLWRNORM;
967 
968 	rcu_read_unlock();
969 	return mask;
970 }
971 
972 static int xsk_init_queue(u32 entries, struct xsk_queue **queue,
973 			  bool umem_queue)
974 {
975 	struct xsk_queue *q;
976 
977 	if (entries == 0 || *queue || !is_power_of_2(entries))
978 		return -EINVAL;
979 
980 	q = xskq_create(entries, umem_queue);
981 	if (!q)
982 		return -ENOMEM;
983 
984 	/* Make sure queue is ready before it can be seen by others */
985 	smp_wmb();
986 	WRITE_ONCE(*queue, q);
987 	return 0;
988 }
989 
990 static void xsk_unbind_dev(struct xdp_sock *xs)
991 {
992 	struct net_device *dev = xs->dev;
993 
994 	if (xs->state != XSK_BOUND)
995 		return;
996 	WRITE_ONCE(xs->state, XSK_UNBOUND);
997 
998 	/* Wait for driver to stop using the xdp socket. */
999 	xp_del_xsk(xs->pool, xs);
1000 	synchronize_net();
1001 	dev_put(dev);
1002 }
1003 
1004 static struct xsk_map *xsk_get_map_list_entry(struct xdp_sock *xs,
1005 					      struct xdp_sock __rcu ***map_entry)
1006 {
1007 	struct xsk_map *map = NULL;
1008 	struct xsk_map_node *node;
1009 
1010 	*map_entry = NULL;
1011 
1012 	spin_lock_bh(&xs->map_list_lock);
1013 	node = list_first_entry_or_null(&xs->map_list, struct xsk_map_node,
1014 					node);
1015 	if (node) {
1016 		bpf_map_inc(&node->map->map);
1017 		map = node->map;
1018 		*map_entry = node->map_entry;
1019 	}
1020 	spin_unlock_bh(&xs->map_list_lock);
1021 	return map;
1022 }
1023 
1024 static void xsk_delete_from_maps(struct xdp_sock *xs)
1025 {
1026 	/* This function removes the current XDP socket from all the
1027 	 * maps it resides in. We need to take extra care here, due to
1028 	 * the two locks involved. Each map has a lock synchronizing
1029 	 * updates to the entries, and each socket has a lock that
1030 	 * synchronizes access to the list of maps (map_list). For
1031 	 * deadlock avoidance the locks need to be taken in the order
1032 	 * "map lock"->"socket map list lock". We start off by
1033 	 * accessing the socket map list, and take a reference to the
1034 	 * map to guarantee existence between the
1035 	 * xsk_get_map_list_entry() and xsk_map_try_sock_delete()
1036 	 * calls. Then we ask the map to remove the socket, which
1037 	 * tries to remove the socket from the map. Note that there
1038 	 * might be updates to the map between
1039 	 * xsk_get_map_list_entry() and xsk_map_try_sock_delete().
1040 	 */
1041 	struct xdp_sock __rcu **map_entry = NULL;
1042 	struct xsk_map *map;
1043 
1044 	while ((map = xsk_get_map_list_entry(xs, &map_entry))) {
1045 		xsk_map_try_sock_delete(map, xs, map_entry);
1046 		bpf_map_put(&map->map);
1047 	}
1048 }
1049 
1050 static int xsk_release(struct socket *sock)
1051 {
1052 	struct sock *sk = sock->sk;
1053 	struct xdp_sock *xs = xdp_sk(sk);
1054 	struct net *net;
1055 
1056 	if (!sk)
1057 		return 0;
1058 
1059 	net = sock_net(sk);
1060 
1061 	if (xs->skb)
1062 		xsk_drop_skb(xs->skb);
1063 
1064 	mutex_lock(&net->xdp.lock);
1065 	sk_del_node_init_rcu(sk);
1066 	mutex_unlock(&net->xdp.lock);
1067 
1068 	sock_prot_inuse_add(net, sk->sk_prot, -1);
1069 
1070 	xsk_delete_from_maps(xs);
1071 	mutex_lock(&xs->mutex);
1072 	xsk_unbind_dev(xs);
1073 	mutex_unlock(&xs->mutex);
1074 
1075 	xskq_destroy(xs->rx);
1076 	xskq_destroy(xs->tx);
1077 	xskq_destroy(xs->fq_tmp);
1078 	xskq_destroy(xs->cq_tmp);
1079 
1080 	sock_orphan(sk);
1081 	sock->sk = NULL;
1082 
1083 	sock_put(sk);
1084 
1085 	return 0;
1086 }
1087 
1088 static struct socket *xsk_lookup_xsk_from_fd(int fd)
1089 {
1090 	struct socket *sock;
1091 	int err;
1092 
1093 	sock = sockfd_lookup(fd, &err);
1094 	if (!sock)
1095 		return ERR_PTR(-ENOTSOCK);
1096 
1097 	if (sock->sk->sk_family != PF_XDP) {
1098 		sockfd_put(sock);
1099 		return ERR_PTR(-ENOPROTOOPT);
1100 	}
1101 
1102 	return sock;
1103 }
1104 
1105 static bool xsk_validate_queues(struct xdp_sock *xs)
1106 {
1107 	return xs->fq_tmp && xs->cq_tmp;
1108 }
1109 
1110 static int xsk_bind(struct socket *sock, struct sockaddr *addr, int addr_len)
1111 {
1112 	struct sockaddr_xdp *sxdp = (struct sockaddr_xdp *)addr;
1113 	struct sock *sk = sock->sk;
1114 	struct xdp_sock *xs = xdp_sk(sk);
1115 	struct net_device *dev;
1116 	int bound_dev_if;
1117 	u32 flags, qid;
1118 	int err = 0;
1119 
1120 	if (addr_len < sizeof(struct sockaddr_xdp))
1121 		return -EINVAL;
1122 	if (sxdp->sxdp_family != AF_XDP)
1123 		return -EINVAL;
1124 
1125 	flags = sxdp->sxdp_flags;
1126 	if (flags & ~(XDP_SHARED_UMEM | XDP_COPY | XDP_ZEROCOPY |
1127 		      XDP_USE_NEED_WAKEUP | XDP_USE_SG))
1128 		return -EINVAL;
1129 
1130 	bound_dev_if = READ_ONCE(sk->sk_bound_dev_if);
1131 	if (bound_dev_if && bound_dev_if != sxdp->sxdp_ifindex)
1132 		return -EINVAL;
1133 
1134 	rtnl_lock();
1135 	mutex_lock(&xs->mutex);
1136 	if (xs->state != XSK_READY) {
1137 		err = -EBUSY;
1138 		goto out_release;
1139 	}
1140 
1141 	dev = dev_get_by_index(sock_net(sk), sxdp->sxdp_ifindex);
1142 	if (!dev) {
1143 		err = -ENODEV;
1144 		goto out_release;
1145 	}
1146 
1147 	if (!xs->rx && !xs->tx) {
1148 		err = -EINVAL;
1149 		goto out_unlock;
1150 	}
1151 
1152 	qid = sxdp->sxdp_queue_id;
1153 
1154 	if (flags & XDP_SHARED_UMEM) {
1155 		struct xdp_sock *umem_xs;
1156 		struct socket *sock;
1157 
1158 		if ((flags & XDP_COPY) || (flags & XDP_ZEROCOPY) ||
1159 		    (flags & XDP_USE_NEED_WAKEUP) || (flags & XDP_USE_SG)) {
1160 			/* Cannot specify flags for shared sockets. */
1161 			err = -EINVAL;
1162 			goto out_unlock;
1163 		}
1164 
1165 		if (xs->umem) {
1166 			/* We have already our own. */
1167 			err = -EINVAL;
1168 			goto out_unlock;
1169 		}
1170 
1171 		sock = xsk_lookup_xsk_from_fd(sxdp->sxdp_shared_umem_fd);
1172 		if (IS_ERR(sock)) {
1173 			err = PTR_ERR(sock);
1174 			goto out_unlock;
1175 		}
1176 
1177 		umem_xs = xdp_sk(sock->sk);
1178 		if (!xsk_is_bound(umem_xs)) {
1179 			err = -EBADF;
1180 			sockfd_put(sock);
1181 			goto out_unlock;
1182 		}
1183 
1184 		if (umem_xs->queue_id != qid || umem_xs->dev != dev) {
1185 			/* Share the umem with another socket on another qid
1186 			 * and/or device.
1187 			 */
1188 			xs->pool = xp_create_and_assign_umem(xs,
1189 							     umem_xs->umem);
1190 			if (!xs->pool) {
1191 				err = -ENOMEM;
1192 				sockfd_put(sock);
1193 				goto out_unlock;
1194 			}
1195 
1196 			err = xp_assign_dev_shared(xs->pool, umem_xs, dev,
1197 						   qid);
1198 			if (err) {
1199 				xp_destroy(xs->pool);
1200 				xs->pool = NULL;
1201 				sockfd_put(sock);
1202 				goto out_unlock;
1203 			}
1204 		} else {
1205 			/* Share the buffer pool with the other socket. */
1206 			if (xs->fq_tmp || xs->cq_tmp) {
1207 				/* Do not allow setting your own fq or cq. */
1208 				err = -EINVAL;
1209 				sockfd_put(sock);
1210 				goto out_unlock;
1211 			}
1212 
1213 			xp_get_pool(umem_xs->pool);
1214 			xs->pool = umem_xs->pool;
1215 
1216 			/* If underlying shared umem was created without Tx
1217 			 * ring, allocate Tx descs array that Tx batching API
1218 			 * utilizes
1219 			 */
1220 			if (xs->tx && !xs->pool->tx_descs) {
1221 				err = xp_alloc_tx_descs(xs->pool, xs);
1222 				if (err) {
1223 					xp_put_pool(xs->pool);
1224 					xs->pool = NULL;
1225 					sockfd_put(sock);
1226 					goto out_unlock;
1227 				}
1228 			}
1229 		}
1230 
1231 		xdp_get_umem(umem_xs->umem);
1232 		WRITE_ONCE(xs->umem, umem_xs->umem);
1233 		sockfd_put(sock);
1234 	} else if (!xs->umem || !xsk_validate_queues(xs)) {
1235 		err = -EINVAL;
1236 		goto out_unlock;
1237 	} else {
1238 		/* This xsk has its own umem. */
1239 		xs->pool = xp_create_and_assign_umem(xs, xs->umem);
1240 		if (!xs->pool) {
1241 			err = -ENOMEM;
1242 			goto out_unlock;
1243 		}
1244 
1245 		err = xp_assign_dev(xs->pool, dev, qid, flags);
1246 		if (err) {
1247 			xp_destroy(xs->pool);
1248 			xs->pool = NULL;
1249 			goto out_unlock;
1250 		}
1251 	}
1252 
1253 	/* FQ and CQ are now owned by the buffer pool and cleaned up with it. */
1254 	xs->fq_tmp = NULL;
1255 	xs->cq_tmp = NULL;
1256 
1257 	xs->dev = dev;
1258 	xs->zc = xs->umem->zc;
1259 	xs->sg = !!(xs->umem->flags & XDP_UMEM_SG_FLAG);
1260 	xs->queue_id = qid;
1261 	xp_add_xsk(xs->pool, xs);
1262 
1263 out_unlock:
1264 	if (err) {
1265 		dev_put(dev);
1266 	} else {
1267 		/* Matches smp_rmb() in bind() for shared umem
1268 		 * sockets, and xsk_is_bound().
1269 		 */
1270 		smp_wmb();
1271 		WRITE_ONCE(xs->state, XSK_BOUND);
1272 	}
1273 out_release:
1274 	mutex_unlock(&xs->mutex);
1275 	rtnl_unlock();
1276 	return err;
1277 }
1278 
1279 struct xdp_umem_reg_v1 {
1280 	__u64 addr; /* Start of packet data area */
1281 	__u64 len; /* Length of packet data area */
1282 	__u32 chunk_size;
1283 	__u32 headroom;
1284 };
1285 
1286 static int xsk_setsockopt(struct socket *sock, int level, int optname,
1287 			  sockptr_t optval, unsigned int optlen)
1288 {
1289 	struct sock *sk = sock->sk;
1290 	struct xdp_sock *xs = xdp_sk(sk);
1291 	int err;
1292 
1293 	if (level != SOL_XDP)
1294 		return -ENOPROTOOPT;
1295 
1296 	switch (optname) {
1297 	case XDP_RX_RING:
1298 	case XDP_TX_RING:
1299 	{
1300 		struct xsk_queue **q;
1301 		int entries;
1302 
1303 		if (optlen < sizeof(entries))
1304 			return -EINVAL;
1305 		if (copy_from_sockptr(&entries, optval, sizeof(entries)))
1306 			return -EFAULT;
1307 
1308 		mutex_lock(&xs->mutex);
1309 		if (xs->state != XSK_READY) {
1310 			mutex_unlock(&xs->mutex);
1311 			return -EBUSY;
1312 		}
1313 		q = (optname == XDP_TX_RING) ? &xs->tx : &xs->rx;
1314 		err = xsk_init_queue(entries, q, false);
1315 		if (!err && optname == XDP_TX_RING)
1316 			/* Tx needs to be explicitly woken up the first time */
1317 			xs->tx->ring->flags |= XDP_RING_NEED_WAKEUP;
1318 		mutex_unlock(&xs->mutex);
1319 		return err;
1320 	}
1321 	case XDP_UMEM_REG:
1322 	{
1323 		size_t mr_size = sizeof(struct xdp_umem_reg);
1324 		struct xdp_umem_reg mr = {};
1325 		struct xdp_umem *umem;
1326 
1327 		if (optlen < sizeof(struct xdp_umem_reg_v1))
1328 			return -EINVAL;
1329 		else if (optlen < sizeof(mr))
1330 			mr_size = sizeof(struct xdp_umem_reg_v1);
1331 
1332 		if (copy_from_sockptr(&mr, optval, mr_size))
1333 			return -EFAULT;
1334 
1335 		mutex_lock(&xs->mutex);
1336 		if (xs->state != XSK_READY || xs->umem) {
1337 			mutex_unlock(&xs->mutex);
1338 			return -EBUSY;
1339 		}
1340 
1341 		umem = xdp_umem_create(&mr);
1342 		if (IS_ERR(umem)) {
1343 			mutex_unlock(&xs->mutex);
1344 			return PTR_ERR(umem);
1345 		}
1346 
1347 		/* Make sure umem is ready before it can be seen by others */
1348 		smp_wmb();
1349 		WRITE_ONCE(xs->umem, umem);
1350 		mutex_unlock(&xs->mutex);
1351 		return 0;
1352 	}
1353 	case XDP_UMEM_FILL_RING:
1354 	case XDP_UMEM_COMPLETION_RING:
1355 	{
1356 		struct xsk_queue **q;
1357 		int entries;
1358 
1359 		if (copy_from_sockptr(&entries, optval, sizeof(entries)))
1360 			return -EFAULT;
1361 
1362 		mutex_lock(&xs->mutex);
1363 		if (xs->state != XSK_READY) {
1364 			mutex_unlock(&xs->mutex);
1365 			return -EBUSY;
1366 		}
1367 
1368 		q = (optname == XDP_UMEM_FILL_RING) ? &xs->fq_tmp :
1369 			&xs->cq_tmp;
1370 		err = xsk_init_queue(entries, q, true);
1371 		mutex_unlock(&xs->mutex);
1372 		return err;
1373 	}
1374 	default:
1375 		break;
1376 	}
1377 
1378 	return -ENOPROTOOPT;
1379 }
1380 
1381 static void xsk_enter_rxtx_offsets(struct xdp_ring_offset_v1 *ring)
1382 {
1383 	ring->producer = offsetof(struct xdp_rxtx_ring, ptrs.producer);
1384 	ring->consumer = offsetof(struct xdp_rxtx_ring, ptrs.consumer);
1385 	ring->desc = offsetof(struct xdp_rxtx_ring, desc);
1386 }
1387 
1388 static void xsk_enter_umem_offsets(struct xdp_ring_offset_v1 *ring)
1389 {
1390 	ring->producer = offsetof(struct xdp_umem_ring, ptrs.producer);
1391 	ring->consumer = offsetof(struct xdp_umem_ring, ptrs.consumer);
1392 	ring->desc = offsetof(struct xdp_umem_ring, desc);
1393 }
1394 
1395 struct xdp_statistics_v1 {
1396 	__u64 rx_dropped;
1397 	__u64 rx_invalid_descs;
1398 	__u64 tx_invalid_descs;
1399 };
1400 
1401 static int xsk_getsockopt(struct socket *sock, int level, int optname,
1402 			  char __user *optval, int __user *optlen)
1403 {
1404 	struct sock *sk = sock->sk;
1405 	struct xdp_sock *xs = xdp_sk(sk);
1406 	int len;
1407 
1408 	if (level != SOL_XDP)
1409 		return -ENOPROTOOPT;
1410 
1411 	if (get_user(len, optlen))
1412 		return -EFAULT;
1413 	if (len < 0)
1414 		return -EINVAL;
1415 
1416 	switch (optname) {
1417 	case XDP_STATISTICS:
1418 	{
1419 		struct xdp_statistics stats = {};
1420 		bool extra_stats = true;
1421 		size_t stats_size;
1422 
1423 		if (len < sizeof(struct xdp_statistics_v1)) {
1424 			return -EINVAL;
1425 		} else if (len < sizeof(stats)) {
1426 			extra_stats = false;
1427 			stats_size = sizeof(struct xdp_statistics_v1);
1428 		} else {
1429 			stats_size = sizeof(stats);
1430 		}
1431 
1432 		mutex_lock(&xs->mutex);
1433 		stats.rx_dropped = xs->rx_dropped;
1434 		if (extra_stats) {
1435 			stats.rx_ring_full = xs->rx_queue_full;
1436 			stats.rx_fill_ring_empty_descs =
1437 				xs->pool ? xskq_nb_queue_empty_descs(xs->pool->fq) : 0;
1438 			stats.tx_ring_empty_descs = xskq_nb_queue_empty_descs(xs->tx);
1439 		} else {
1440 			stats.rx_dropped += xs->rx_queue_full;
1441 		}
1442 		stats.rx_invalid_descs = xskq_nb_invalid_descs(xs->rx);
1443 		stats.tx_invalid_descs = xskq_nb_invalid_descs(xs->tx);
1444 		mutex_unlock(&xs->mutex);
1445 
1446 		if (copy_to_user(optval, &stats, stats_size))
1447 			return -EFAULT;
1448 		if (put_user(stats_size, optlen))
1449 			return -EFAULT;
1450 
1451 		return 0;
1452 	}
1453 	case XDP_MMAP_OFFSETS:
1454 	{
1455 		struct xdp_mmap_offsets off;
1456 		struct xdp_mmap_offsets_v1 off_v1;
1457 		bool flags_supported = true;
1458 		void *to_copy;
1459 
1460 		if (len < sizeof(off_v1))
1461 			return -EINVAL;
1462 		else if (len < sizeof(off))
1463 			flags_supported = false;
1464 
1465 		if (flags_supported) {
1466 			/* xdp_ring_offset is identical to xdp_ring_offset_v1
1467 			 * except for the flags field added to the end.
1468 			 */
1469 			xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *)
1470 					       &off.rx);
1471 			xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *)
1472 					       &off.tx);
1473 			xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *)
1474 					       &off.fr);
1475 			xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *)
1476 					       &off.cr);
1477 			off.rx.flags = offsetof(struct xdp_rxtx_ring,
1478 						ptrs.flags);
1479 			off.tx.flags = offsetof(struct xdp_rxtx_ring,
1480 						ptrs.flags);
1481 			off.fr.flags = offsetof(struct xdp_umem_ring,
1482 						ptrs.flags);
1483 			off.cr.flags = offsetof(struct xdp_umem_ring,
1484 						ptrs.flags);
1485 
1486 			len = sizeof(off);
1487 			to_copy = &off;
1488 		} else {
1489 			xsk_enter_rxtx_offsets(&off_v1.rx);
1490 			xsk_enter_rxtx_offsets(&off_v1.tx);
1491 			xsk_enter_umem_offsets(&off_v1.fr);
1492 			xsk_enter_umem_offsets(&off_v1.cr);
1493 
1494 			len = sizeof(off_v1);
1495 			to_copy = &off_v1;
1496 		}
1497 
1498 		if (copy_to_user(optval, to_copy, len))
1499 			return -EFAULT;
1500 		if (put_user(len, optlen))
1501 			return -EFAULT;
1502 
1503 		return 0;
1504 	}
1505 	case XDP_OPTIONS:
1506 	{
1507 		struct xdp_options opts = {};
1508 
1509 		if (len < sizeof(opts))
1510 			return -EINVAL;
1511 
1512 		mutex_lock(&xs->mutex);
1513 		if (xs->zc)
1514 			opts.flags |= XDP_OPTIONS_ZEROCOPY;
1515 		mutex_unlock(&xs->mutex);
1516 
1517 		len = sizeof(opts);
1518 		if (copy_to_user(optval, &opts, len))
1519 			return -EFAULT;
1520 		if (put_user(len, optlen))
1521 			return -EFAULT;
1522 
1523 		return 0;
1524 	}
1525 	default:
1526 		break;
1527 	}
1528 
1529 	return -EOPNOTSUPP;
1530 }
1531 
1532 static int xsk_mmap(struct file *file, struct socket *sock,
1533 		    struct vm_area_struct *vma)
1534 {
1535 	loff_t offset = (loff_t)vma->vm_pgoff << PAGE_SHIFT;
1536 	unsigned long size = vma->vm_end - vma->vm_start;
1537 	struct xdp_sock *xs = xdp_sk(sock->sk);
1538 	int state = READ_ONCE(xs->state);
1539 	struct xsk_queue *q = NULL;
1540 
1541 	if (state != XSK_READY && state != XSK_BOUND)
1542 		return -EBUSY;
1543 
1544 	if (offset == XDP_PGOFF_RX_RING) {
1545 		q = READ_ONCE(xs->rx);
1546 	} else if (offset == XDP_PGOFF_TX_RING) {
1547 		q = READ_ONCE(xs->tx);
1548 	} else {
1549 		/* Matches the smp_wmb() in XDP_UMEM_REG */
1550 		smp_rmb();
1551 		if (offset == XDP_UMEM_PGOFF_FILL_RING)
1552 			q = state == XSK_READY ? READ_ONCE(xs->fq_tmp) :
1553 						 READ_ONCE(xs->pool->fq);
1554 		else if (offset == XDP_UMEM_PGOFF_COMPLETION_RING)
1555 			q = state == XSK_READY ? READ_ONCE(xs->cq_tmp) :
1556 						 READ_ONCE(xs->pool->cq);
1557 	}
1558 
1559 	if (!q)
1560 		return -EINVAL;
1561 
1562 	/* Matches the smp_wmb() in xsk_init_queue */
1563 	smp_rmb();
1564 	if (size > q->ring_vmalloc_size)
1565 		return -EINVAL;
1566 
1567 	return remap_vmalloc_range(vma, q->ring, 0);
1568 }
1569 
1570 static int xsk_notifier(struct notifier_block *this,
1571 			unsigned long msg, void *ptr)
1572 {
1573 	struct net_device *dev = netdev_notifier_info_to_dev(ptr);
1574 	struct net *net = dev_net(dev);
1575 	struct sock *sk;
1576 
1577 	switch (msg) {
1578 	case NETDEV_UNREGISTER:
1579 		mutex_lock(&net->xdp.lock);
1580 		sk_for_each(sk, &net->xdp.list) {
1581 			struct xdp_sock *xs = xdp_sk(sk);
1582 
1583 			mutex_lock(&xs->mutex);
1584 			if (xs->dev == dev) {
1585 				sk->sk_err = ENETDOWN;
1586 				if (!sock_flag(sk, SOCK_DEAD))
1587 					sk_error_report(sk);
1588 
1589 				xsk_unbind_dev(xs);
1590 
1591 				/* Clear device references. */
1592 				xp_clear_dev(xs->pool);
1593 			}
1594 			mutex_unlock(&xs->mutex);
1595 		}
1596 		mutex_unlock(&net->xdp.lock);
1597 		break;
1598 	}
1599 	return NOTIFY_DONE;
1600 }
1601 
1602 static struct proto xsk_proto = {
1603 	.name =		"XDP",
1604 	.owner =	THIS_MODULE,
1605 	.obj_size =	sizeof(struct xdp_sock),
1606 };
1607 
1608 static const struct proto_ops xsk_proto_ops = {
1609 	.family		= PF_XDP,
1610 	.owner		= THIS_MODULE,
1611 	.release	= xsk_release,
1612 	.bind		= xsk_bind,
1613 	.connect	= sock_no_connect,
1614 	.socketpair	= sock_no_socketpair,
1615 	.accept		= sock_no_accept,
1616 	.getname	= sock_no_getname,
1617 	.poll		= xsk_poll,
1618 	.ioctl		= sock_no_ioctl,
1619 	.listen		= sock_no_listen,
1620 	.shutdown	= sock_no_shutdown,
1621 	.setsockopt	= xsk_setsockopt,
1622 	.getsockopt	= xsk_getsockopt,
1623 	.sendmsg	= xsk_sendmsg,
1624 	.recvmsg	= xsk_recvmsg,
1625 	.mmap		= xsk_mmap,
1626 };
1627 
1628 static void xsk_destruct(struct sock *sk)
1629 {
1630 	struct xdp_sock *xs = xdp_sk(sk);
1631 
1632 	if (!sock_flag(sk, SOCK_DEAD))
1633 		return;
1634 
1635 	if (!xp_put_pool(xs->pool))
1636 		xdp_put_umem(xs->umem, !xs->pool);
1637 }
1638 
1639 static int xsk_create(struct net *net, struct socket *sock, int protocol,
1640 		      int kern)
1641 {
1642 	struct xdp_sock *xs;
1643 	struct sock *sk;
1644 
1645 	if (!ns_capable(net->user_ns, CAP_NET_RAW))
1646 		return -EPERM;
1647 	if (sock->type != SOCK_RAW)
1648 		return -ESOCKTNOSUPPORT;
1649 
1650 	if (protocol)
1651 		return -EPROTONOSUPPORT;
1652 
1653 	sock->state = SS_UNCONNECTED;
1654 
1655 	sk = sk_alloc(net, PF_XDP, GFP_KERNEL, &xsk_proto, kern);
1656 	if (!sk)
1657 		return -ENOBUFS;
1658 
1659 	sock->ops = &xsk_proto_ops;
1660 
1661 	sock_init_data(sock, sk);
1662 
1663 	sk->sk_family = PF_XDP;
1664 
1665 	sk->sk_destruct = xsk_destruct;
1666 
1667 	sock_set_flag(sk, SOCK_RCU_FREE);
1668 
1669 	xs = xdp_sk(sk);
1670 	xs->state = XSK_READY;
1671 	mutex_init(&xs->mutex);
1672 	spin_lock_init(&xs->rx_lock);
1673 
1674 	INIT_LIST_HEAD(&xs->map_list);
1675 	spin_lock_init(&xs->map_list_lock);
1676 
1677 	mutex_lock(&net->xdp.lock);
1678 	sk_add_node_rcu(sk, &net->xdp.list);
1679 	mutex_unlock(&net->xdp.lock);
1680 
1681 	sock_prot_inuse_add(net, &xsk_proto, 1);
1682 
1683 	return 0;
1684 }
1685 
1686 static const struct net_proto_family xsk_family_ops = {
1687 	.family = PF_XDP,
1688 	.create = xsk_create,
1689 	.owner	= THIS_MODULE,
1690 };
1691 
1692 static struct notifier_block xsk_netdev_notifier = {
1693 	.notifier_call	= xsk_notifier,
1694 };
1695 
1696 static int __net_init xsk_net_init(struct net *net)
1697 {
1698 	mutex_init(&net->xdp.lock);
1699 	INIT_HLIST_HEAD(&net->xdp.list);
1700 	return 0;
1701 }
1702 
1703 static void __net_exit xsk_net_exit(struct net *net)
1704 {
1705 	WARN_ON_ONCE(!hlist_empty(&net->xdp.list));
1706 }
1707 
1708 static struct pernet_operations xsk_net_ops = {
1709 	.init = xsk_net_init,
1710 	.exit = xsk_net_exit,
1711 };
1712 
1713 static int __init xsk_init(void)
1714 {
1715 	int err, cpu;
1716 
1717 	err = proto_register(&xsk_proto, 0 /* no slab */);
1718 	if (err)
1719 		goto out;
1720 
1721 	err = sock_register(&xsk_family_ops);
1722 	if (err)
1723 		goto out_proto;
1724 
1725 	err = register_pernet_subsys(&xsk_net_ops);
1726 	if (err)
1727 		goto out_sk;
1728 
1729 	err = register_netdevice_notifier(&xsk_netdev_notifier);
1730 	if (err)
1731 		goto out_pernet;
1732 
1733 	for_each_possible_cpu(cpu)
1734 		INIT_LIST_HEAD(&per_cpu(xskmap_flush_list, cpu));
1735 	return 0;
1736 
1737 out_pernet:
1738 	unregister_pernet_subsys(&xsk_net_ops);
1739 out_sk:
1740 	sock_unregister(PF_XDP);
1741 out_proto:
1742 	proto_unregister(&xsk_proto);
1743 out:
1744 	return err;
1745 }
1746 
1747 fs_initcall(xsk_init);
1748