xref: /linux/net/xdp/xsk.c (revision 3be042cf46feeedf664152d063376b5c17026d1d)
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, 0);
726 		}
727 
728 		if (first_frag && desc->options & XDP_TX_METADATA) {
729 			if (unlikely(xs->pool->tx_metadata_len == 0)) {
730 				err = -EINVAL;
731 				goto free_err;
732 			}
733 
734 			meta = buffer - xs->pool->tx_metadata_len;
735 			if (unlikely(!xsk_buff_valid_tx_metadata(meta))) {
736 				err = -EINVAL;
737 				goto free_err;
738 			}
739 
740 			if (meta->flags & XDP_TXMD_FLAGS_CHECKSUM) {
741 				if (unlikely(meta->request.csum_start +
742 					     meta->request.csum_offset +
743 					     sizeof(__sum16) > len)) {
744 					err = -EINVAL;
745 					goto free_err;
746 				}
747 
748 				skb->csum_start = hr + meta->request.csum_start;
749 				skb->csum_offset = meta->request.csum_offset;
750 				skb->ip_summed = CHECKSUM_PARTIAL;
751 
752 				if (unlikely(xs->pool->tx_sw_csum)) {
753 					err = skb_checksum_help(skb);
754 					if (err)
755 						goto free_err;
756 				}
757 			}
758 		}
759 	}
760 
761 	skb->dev = dev;
762 	skb->priority = READ_ONCE(xs->sk.sk_priority);
763 	skb->mark = READ_ONCE(xs->sk.sk_mark);
764 	skb->destructor = xsk_destruct_skb;
765 	xsk_tx_metadata_to_compl(meta, &skb_shinfo(skb)->xsk_meta);
766 	xsk_set_destructor_arg(skb);
767 
768 	return skb;
769 
770 free_err:
771 	if (err == -EOVERFLOW) {
772 		/* Drop the packet */
773 		xsk_set_destructor_arg(xs->skb);
774 		xsk_drop_skb(xs->skb);
775 		xskq_cons_release(xs->tx);
776 	} else {
777 		/* Let application retry */
778 		xsk_cq_cancel_locked(xs, 1);
779 	}
780 
781 	return ERR_PTR(err);
782 }
783 
784 static int __xsk_generic_xmit(struct sock *sk)
785 {
786 	struct xdp_sock *xs = xdp_sk(sk);
787 	u32 max_batch = TX_BATCH_SIZE;
788 	bool sent_frame = false;
789 	struct xdp_desc desc;
790 	struct sk_buff *skb;
791 	int err = 0;
792 
793 	mutex_lock(&xs->mutex);
794 
795 	/* Since we dropped the RCU read lock, the socket state might have changed. */
796 	if (unlikely(!xsk_is_bound(xs))) {
797 		err = -ENXIO;
798 		goto out;
799 	}
800 
801 	if (xs->queue_id >= xs->dev->real_num_tx_queues)
802 		goto out;
803 
804 	while (xskq_cons_peek_desc(xs->tx, &desc, xs->pool)) {
805 		if (max_batch-- == 0) {
806 			err = -EAGAIN;
807 			goto out;
808 		}
809 
810 		/* This is the backpressure mechanism for the Tx path.
811 		 * Reserve space in the completion queue and only proceed
812 		 * if there is space in it. This avoids having to implement
813 		 * any buffering in the Tx path.
814 		 */
815 		if (xsk_cq_reserve_addr_locked(xs, desc.addr))
816 			goto out;
817 
818 		skb = xsk_build_skb(xs, &desc);
819 		if (IS_ERR(skb)) {
820 			err = PTR_ERR(skb);
821 			if (err != -EOVERFLOW)
822 				goto out;
823 			err = 0;
824 			continue;
825 		}
826 
827 		xskq_cons_release(xs->tx);
828 
829 		if (xp_mb_desc(&desc)) {
830 			xs->skb = skb;
831 			continue;
832 		}
833 
834 		err = __dev_direct_xmit(skb, xs->queue_id);
835 		if  (err == NETDEV_TX_BUSY) {
836 			/* Tell user-space to retry the send */
837 			xskq_cons_cancel_n(xs->tx, xsk_get_num_desc(skb));
838 			xsk_consume_skb(skb);
839 			err = -EAGAIN;
840 			goto out;
841 		}
842 
843 		/* Ignore NET_XMIT_CN as packet might have been sent */
844 		if (err == NET_XMIT_DROP) {
845 			/* SKB completed but not sent */
846 			err = -EBUSY;
847 			xs->skb = NULL;
848 			goto out;
849 		}
850 
851 		sent_frame = true;
852 		xs->skb = NULL;
853 	}
854 
855 	if (xskq_has_descs(xs->tx)) {
856 		if (xs->skb)
857 			xsk_drop_skb(xs->skb);
858 		xskq_cons_release(xs->tx);
859 	}
860 
861 out:
862 	if (sent_frame)
863 		if (xsk_tx_writeable(xs))
864 			sk->sk_write_space(sk);
865 
866 	mutex_unlock(&xs->mutex);
867 	return err;
868 }
869 
870 static int xsk_generic_xmit(struct sock *sk)
871 {
872 	int ret;
873 
874 	/* Drop the RCU lock since the SKB path might sleep. */
875 	rcu_read_unlock();
876 	ret = __xsk_generic_xmit(sk);
877 	/* Reaquire RCU lock before going into common code. */
878 	rcu_read_lock();
879 
880 	return ret;
881 }
882 
883 static bool xsk_no_wakeup(struct sock *sk)
884 {
885 #ifdef CONFIG_NET_RX_BUSY_POLL
886 	/* Prefer busy-polling, skip the wakeup. */
887 	return READ_ONCE(sk->sk_prefer_busy_poll) && READ_ONCE(sk->sk_ll_usec) &&
888 		READ_ONCE(sk->sk_napi_id) >= MIN_NAPI_ID;
889 #else
890 	return false;
891 #endif
892 }
893 
894 static int xsk_check_common(struct xdp_sock *xs)
895 {
896 	if (unlikely(!xsk_is_bound(xs)))
897 		return -ENXIO;
898 	if (unlikely(!(xs->dev->flags & IFF_UP)))
899 		return -ENETDOWN;
900 
901 	return 0;
902 }
903 
904 static int __xsk_sendmsg(struct socket *sock, struct msghdr *m, size_t total_len)
905 {
906 	bool need_wait = !(m->msg_flags & MSG_DONTWAIT);
907 	struct sock *sk = sock->sk;
908 	struct xdp_sock *xs = xdp_sk(sk);
909 	struct xsk_buff_pool *pool;
910 	int err;
911 
912 	err = xsk_check_common(xs);
913 	if (err)
914 		return err;
915 	if (unlikely(need_wait))
916 		return -EOPNOTSUPP;
917 	if (unlikely(!xs->tx))
918 		return -ENOBUFS;
919 
920 	if (sk_can_busy_loop(sk)) {
921 		if (xs->zc)
922 			__sk_mark_napi_id_once(sk, xsk_pool_get_napi_id(xs->pool));
923 		sk_busy_loop(sk, 1); /* only support non-blocking sockets */
924 	}
925 
926 	if (xs->zc && xsk_no_wakeup(sk))
927 		return 0;
928 
929 	pool = xs->pool;
930 	if (pool->cached_need_wakeup & XDP_WAKEUP_TX) {
931 		if (xs->zc)
932 			return xsk_wakeup(xs, XDP_WAKEUP_TX);
933 		return xsk_generic_xmit(sk);
934 	}
935 	return 0;
936 }
937 
938 static int xsk_sendmsg(struct socket *sock, struct msghdr *m, size_t total_len)
939 {
940 	int ret;
941 
942 	rcu_read_lock();
943 	ret = __xsk_sendmsg(sock, m, total_len);
944 	rcu_read_unlock();
945 
946 	return ret;
947 }
948 
949 static int __xsk_recvmsg(struct socket *sock, struct msghdr *m, size_t len, int flags)
950 {
951 	bool need_wait = !(flags & MSG_DONTWAIT);
952 	struct sock *sk = sock->sk;
953 	struct xdp_sock *xs = xdp_sk(sk);
954 	int err;
955 
956 	err = xsk_check_common(xs);
957 	if (err)
958 		return err;
959 	if (unlikely(!xs->rx))
960 		return -ENOBUFS;
961 	if (unlikely(need_wait))
962 		return -EOPNOTSUPP;
963 
964 	if (sk_can_busy_loop(sk))
965 		sk_busy_loop(sk, 1); /* only support non-blocking sockets */
966 
967 	if (xsk_no_wakeup(sk))
968 		return 0;
969 
970 	if (xs->pool->cached_need_wakeup & XDP_WAKEUP_RX && xs->zc)
971 		return xsk_wakeup(xs, XDP_WAKEUP_RX);
972 	return 0;
973 }
974 
975 static int xsk_recvmsg(struct socket *sock, struct msghdr *m, size_t len, int flags)
976 {
977 	int ret;
978 
979 	rcu_read_lock();
980 	ret = __xsk_recvmsg(sock, m, len, flags);
981 	rcu_read_unlock();
982 
983 	return ret;
984 }
985 
986 static __poll_t xsk_poll(struct file *file, struct socket *sock,
987 			     struct poll_table_struct *wait)
988 {
989 	__poll_t mask = 0;
990 	struct sock *sk = sock->sk;
991 	struct xdp_sock *xs = xdp_sk(sk);
992 	struct xsk_buff_pool *pool;
993 
994 	sock_poll_wait(file, sock, wait);
995 
996 	rcu_read_lock();
997 	if (xsk_check_common(xs))
998 		goto out;
999 
1000 	pool = xs->pool;
1001 
1002 	if (pool->cached_need_wakeup) {
1003 		if (xs->zc)
1004 			xsk_wakeup(xs, pool->cached_need_wakeup);
1005 		else if (xs->tx)
1006 			/* Poll needs to drive Tx also in copy mode */
1007 			xsk_generic_xmit(sk);
1008 	}
1009 
1010 	if (xs->rx && !xskq_prod_is_empty(xs->rx))
1011 		mask |= EPOLLIN | EPOLLRDNORM;
1012 	if (xs->tx && xsk_tx_writeable(xs))
1013 		mask |= EPOLLOUT | EPOLLWRNORM;
1014 out:
1015 	rcu_read_unlock();
1016 	return mask;
1017 }
1018 
1019 static int xsk_init_queue(u32 entries, struct xsk_queue **queue,
1020 			  bool umem_queue)
1021 {
1022 	struct xsk_queue *q;
1023 
1024 	if (entries == 0 || *queue || !is_power_of_2(entries))
1025 		return -EINVAL;
1026 
1027 	q = xskq_create(entries, umem_queue);
1028 	if (!q)
1029 		return -ENOMEM;
1030 
1031 	/* Make sure queue is ready before it can be seen by others */
1032 	smp_wmb();
1033 	WRITE_ONCE(*queue, q);
1034 	return 0;
1035 }
1036 
1037 static void xsk_unbind_dev(struct xdp_sock *xs)
1038 {
1039 	struct net_device *dev = xs->dev;
1040 
1041 	if (xs->state != XSK_BOUND)
1042 		return;
1043 	WRITE_ONCE(xs->state, XSK_UNBOUND);
1044 
1045 	/* Wait for driver to stop using the xdp socket. */
1046 	xp_del_xsk(xs->pool, xs);
1047 	synchronize_net();
1048 	dev_put(dev);
1049 }
1050 
1051 static struct xsk_map *xsk_get_map_list_entry(struct xdp_sock *xs,
1052 					      struct xdp_sock __rcu ***map_entry)
1053 {
1054 	struct xsk_map *map = NULL;
1055 	struct xsk_map_node *node;
1056 
1057 	*map_entry = NULL;
1058 
1059 	spin_lock_bh(&xs->map_list_lock);
1060 	node = list_first_entry_or_null(&xs->map_list, struct xsk_map_node,
1061 					node);
1062 	if (node) {
1063 		bpf_map_inc(&node->map->map);
1064 		map = node->map;
1065 		*map_entry = node->map_entry;
1066 	}
1067 	spin_unlock_bh(&xs->map_list_lock);
1068 	return map;
1069 }
1070 
1071 static void xsk_delete_from_maps(struct xdp_sock *xs)
1072 {
1073 	/* This function removes the current XDP socket from all the
1074 	 * maps it resides in. We need to take extra care here, due to
1075 	 * the two locks involved. Each map has a lock synchronizing
1076 	 * updates to the entries, and each socket has a lock that
1077 	 * synchronizes access to the list of maps (map_list). For
1078 	 * deadlock avoidance the locks need to be taken in the order
1079 	 * "map lock"->"socket map list lock". We start off by
1080 	 * accessing the socket map list, and take a reference to the
1081 	 * map to guarantee existence between the
1082 	 * xsk_get_map_list_entry() and xsk_map_try_sock_delete()
1083 	 * calls. Then we ask the map to remove the socket, which
1084 	 * tries to remove the socket from the map. Note that there
1085 	 * might be updates to the map between
1086 	 * xsk_get_map_list_entry() and xsk_map_try_sock_delete().
1087 	 */
1088 	struct xdp_sock __rcu **map_entry = NULL;
1089 	struct xsk_map *map;
1090 
1091 	while ((map = xsk_get_map_list_entry(xs, &map_entry))) {
1092 		xsk_map_try_sock_delete(map, xs, map_entry);
1093 		bpf_map_put(&map->map);
1094 	}
1095 }
1096 
1097 static int xsk_release(struct socket *sock)
1098 {
1099 	struct sock *sk = sock->sk;
1100 	struct xdp_sock *xs = xdp_sk(sk);
1101 	struct net *net;
1102 
1103 	if (!sk)
1104 		return 0;
1105 
1106 	net = sock_net(sk);
1107 
1108 	if (xs->skb)
1109 		xsk_drop_skb(xs->skb);
1110 
1111 	mutex_lock(&net->xdp.lock);
1112 	sk_del_node_init_rcu(sk);
1113 	mutex_unlock(&net->xdp.lock);
1114 
1115 	sock_prot_inuse_add(net, sk->sk_prot, -1);
1116 
1117 	xsk_delete_from_maps(xs);
1118 	mutex_lock(&xs->mutex);
1119 	xsk_unbind_dev(xs);
1120 	mutex_unlock(&xs->mutex);
1121 
1122 	xskq_destroy(xs->rx);
1123 	xskq_destroy(xs->tx);
1124 	xskq_destroy(xs->fq_tmp);
1125 	xskq_destroy(xs->cq_tmp);
1126 
1127 	sock_orphan(sk);
1128 	sock->sk = NULL;
1129 
1130 	sock_put(sk);
1131 
1132 	return 0;
1133 }
1134 
1135 static struct socket *xsk_lookup_xsk_from_fd(int fd)
1136 {
1137 	struct socket *sock;
1138 	int err;
1139 
1140 	sock = sockfd_lookup(fd, &err);
1141 	if (!sock)
1142 		return ERR_PTR(-ENOTSOCK);
1143 
1144 	if (sock->sk->sk_family != PF_XDP) {
1145 		sockfd_put(sock);
1146 		return ERR_PTR(-ENOPROTOOPT);
1147 	}
1148 
1149 	return sock;
1150 }
1151 
1152 static bool xsk_validate_queues(struct xdp_sock *xs)
1153 {
1154 	return xs->fq_tmp && xs->cq_tmp;
1155 }
1156 
1157 static int xsk_bind(struct socket *sock, struct sockaddr *addr, int addr_len)
1158 {
1159 	struct sockaddr_xdp *sxdp = (struct sockaddr_xdp *)addr;
1160 	struct sock *sk = sock->sk;
1161 	struct xdp_sock *xs = xdp_sk(sk);
1162 	struct net_device *dev;
1163 	int bound_dev_if;
1164 	u32 flags, qid;
1165 	int err = 0;
1166 
1167 	if (addr_len < sizeof(struct sockaddr_xdp))
1168 		return -EINVAL;
1169 	if (sxdp->sxdp_family != AF_XDP)
1170 		return -EINVAL;
1171 
1172 	flags = sxdp->sxdp_flags;
1173 	if (flags & ~(XDP_SHARED_UMEM | XDP_COPY | XDP_ZEROCOPY |
1174 		      XDP_USE_NEED_WAKEUP | XDP_USE_SG))
1175 		return -EINVAL;
1176 
1177 	bound_dev_if = READ_ONCE(sk->sk_bound_dev_if);
1178 	if (bound_dev_if && bound_dev_if != sxdp->sxdp_ifindex)
1179 		return -EINVAL;
1180 
1181 	rtnl_lock();
1182 	mutex_lock(&xs->mutex);
1183 	if (xs->state != XSK_READY) {
1184 		err = -EBUSY;
1185 		goto out_release;
1186 	}
1187 
1188 	dev = dev_get_by_index(sock_net(sk), sxdp->sxdp_ifindex);
1189 	if (!dev) {
1190 		err = -ENODEV;
1191 		goto out_release;
1192 	}
1193 
1194 	if (!xs->rx && !xs->tx) {
1195 		err = -EINVAL;
1196 		goto out_unlock;
1197 	}
1198 
1199 	qid = sxdp->sxdp_queue_id;
1200 
1201 	if (flags & XDP_SHARED_UMEM) {
1202 		struct xdp_sock *umem_xs;
1203 		struct socket *sock;
1204 
1205 		if ((flags & XDP_COPY) || (flags & XDP_ZEROCOPY) ||
1206 		    (flags & XDP_USE_NEED_WAKEUP) || (flags & XDP_USE_SG)) {
1207 			/* Cannot specify flags for shared sockets. */
1208 			err = -EINVAL;
1209 			goto out_unlock;
1210 		}
1211 
1212 		if (xs->umem) {
1213 			/* We have already our own. */
1214 			err = -EINVAL;
1215 			goto out_unlock;
1216 		}
1217 
1218 		sock = xsk_lookup_xsk_from_fd(sxdp->sxdp_shared_umem_fd);
1219 		if (IS_ERR(sock)) {
1220 			err = PTR_ERR(sock);
1221 			goto out_unlock;
1222 		}
1223 
1224 		umem_xs = xdp_sk(sock->sk);
1225 		if (!xsk_is_bound(umem_xs)) {
1226 			err = -EBADF;
1227 			sockfd_put(sock);
1228 			goto out_unlock;
1229 		}
1230 
1231 		if (umem_xs->queue_id != qid || umem_xs->dev != dev) {
1232 			/* Share the umem with another socket on another qid
1233 			 * and/or device.
1234 			 */
1235 			xs->pool = xp_create_and_assign_umem(xs,
1236 							     umem_xs->umem);
1237 			if (!xs->pool) {
1238 				err = -ENOMEM;
1239 				sockfd_put(sock);
1240 				goto out_unlock;
1241 			}
1242 
1243 			err = xp_assign_dev_shared(xs->pool, umem_xs, dev,
1244 						   qid);
1245 			if (err) {
1246 				xp_destroy(xs->pool);
1247 				xs->pool = NULL;
1248 				sockfd_put(sock);
1249 				goto out_unlock;
1250 			}
1251 		} else {
1252 			/* Share the buffer pool with the other socket. */
1253 			if (xs->fq_tmp || xs->cq_tmp) {
1254 				/* Do not allow setting your own fq or cq. */
1255 				err = -EINVAL;
1256 				sockfd_put(sock);
1257 				goto out_unlock;
1258 			}
1259 
1260 			xp_get_pool(umem_xs->pool);
1261 			xs->pool = umem_xs->pool;
1262 
1263 			/* If underlying shared umem was created without Tx
1264 			 * ring, allocate Tx descs array that Tx batching API
1265 			 * utilizes
1266 			 */
1267 			if (xs->tx && !xs->pool->tx_descs) {
1268 				err = xp_alloc_tx_descs(xs->pool, xs);
1269 				if (err) {
1270 					xp_put_pool(xs->pool);
1271 					xs->pool = NULL;
1272 					sockfd_put(sock);
1273 					goto out_unlock;
1274 				}
1275 			}
1276 		}
1277 
1278 		xdp_get_umem(umem_xs->umem);
1279 		WRITE_ONCE(xs->umem, umem_xs->umem);
1280 		sockfd_put(sock);
1281 	} else if (!xs->umem || !xsk_validate_queues(xs)) {
1282 		err = -EINVAL;
1283 		goto out_unlock;
1284 	} else {
1285 		/* This xsk has its own umem. */
1286 		xs->pool = xp_create_and_assign_umem(xs, xs->umem);
1287 		if (!xs->pool) {
1288 			err = -ENOMEM;
1289 			goto out_unlock;
1290 		}
1291 
1292 		err = xp_assign_dev(xs->pool, dev, qid, flags);
1293 		if (err) {
1294 			xp_destroy(xs->pool);
1295 			xs->pool = NULL;
1296 			goto out_unlock;
1297 		}
1298 	}
1299 
1300 	/* FQ and CQ are now owned by the buffer pool and cleaned up with it. */
1301 	xs->fq_tmp = NULL;
1302 	xs->cq_tmp = NULL;
1303 
1304 	xs->dev = dev;
1305 	xs->zc = xs->umem->zc;
1306 	xs->sg = !!(xs->umem->flags & XDP_UMEM_SG_FLAG);
1307 	xs->queue_id = qid;
1308 	xp_add_xsk(xs->pool, xs);
1309 
1310 out_unlock:
1311 	if (err) {
1312 		dev_put(dev);
1313 	} else {
1314 		/* Matches smp_rmb() in bind() for shared umem
1315 		 * sockets, and xsk_is_bound().
1316 		 */
1317 		smp_wmb();
1318 		WRITE_ONCE(xs->state, XSK_BOUND);
1319 	}
1320 out_release:
1321 	mutex_unlock(&xs->mutex);
1322 	rtnl_unlock();
1323 	return err;
1324 }
1325 
1326 struct xdp_umem_reg_v1 {
1327 	__u64 addr; /* Start of packet data area */
1328 	__u64 len; /* Length of packet data area */
1329 	__u32 chunk_size;
1330 	__u32 headroom;
1331 };
1332 
1333 struct xdp_umem_reg_v2 {
1334 	__u64 addr; /* Start of packet data area */
1335 	__u64 len; /* Length of packet data area */
1336 	__u32 chunk_size;
1337 	__u32 headroom;
1338 	__u32 flags;
1339 };
1340 
1341 static int xsk_setsockopt(struct socket *sock, int level, int optname,
1342 			  sockptr_t optval, unsigned int optlen)
1343 {
1344 	struct sock *sk = sock->sk;
1345 	struct xdp_sock *xs = xdp_sk(sk);
1346 	int err;
1347 
1348 	if (level != SOL_XDP)
1349 		return -ENOPROTOOPT;
1350 
1351 	switch (optname) {
1352 	case XDP_RX_RING:
1353 	case XDP_TX_RING:
1354 	{
1355 		struct xsk_queue **q;
1356 		int entries;
1357 
1358 		if (optlen < sizeof(entries))
1359 			return -EINVAL;
1360 		if (copy_from_sockptr(&entries, optval, sizeof(entries)))
1361 			return -EFAULT;
1362 
1363 		mutex_lock(&xs->mutex);
1364 		if (xs->state != XSK_READY) {
1365 			mutex_unlock(&xs->mutex);
1366 			return -EBUSY;
1367 		}
1368 		q = (optname == XDP_TX_RING) ? &xs->tx : &xs->rx;
1369 		err = xsk_init_queue(entries, q, false);
1370 		if (!err && optname == XDP_TX_RING)
1371 			/* Tx needs to be explicitly woken up the first time */
1372 			xs->tx->ring->flags |= XDP_RING_NEED_WAKEUP;
1373 		mutex_unlock(&xs->mutex);
1374 		return err;
1375 	}
1376 	case XDP_UMEM_REG:
1377 	{
1378 		size_t mr_size = sizeof(struct xdp_umem_reg);
1379 		struct xdp_umem_reg mr = {};
1380 		struct xdp_umem *umem;
1381 
1382 		if (optlen < sizeof(struct xdp_umem_reg_v1))
1383 			return -EINVAL;
1384 		else if (optlen < sizeof(struct xdp_umem_reg_v2))
1385 			mr_size = sizeof(struct xdp_umem_reg_v1);
1386 		else if (optlen < sizeof(mr))
1387 			mr_size = sizeof(struct xdp_umem_reg_v2);
1388 
1389 		if (copy_from_sockptr(&mr, optval, mr_size))
1390 			return -EFAULT;
1391 
1392 		mutex_lock(&xs->mutex);
1393 		if (xs->state != XSK_READY || xs->umem) {
1394 			mutex_unlock(&xs->mutex);
1395 			return -EBUSY;
1396 		}
1397 
1398 		umem = xdp_umem_create(&mr);
1399 		if (IS_ERR(umem)) {
1400 			mutex_unlock(&xs->mutex);
1401 			return PTR_ERR(umem);
1402 		}
1403 
1404 		/* Make sure umem is ready before it can be seen by others */
1405 		smp_wmb();
1406 		WRITE_ONCE(xs->umem, umem);
1407 		mutex_unlock(&xs->mutex);
1408 		return 0;
1409 	}
1410 	case XDP_UMEM_FILL_RING:
1411 	case XDP_UMEM_COMPLETION_RING:
1412 	{
1413 		struct xsk_queue **q;
1414 		int entries;
1415 
1416 		if (copy_from_sockptr(&entries, optval, sizeof(entries)))
1417 			return -EFAULT;
1418 
1419 		mutex_lock(&xs->mutex);
1420 		if (xs->state != XSK_READY) {
1421 			mutex_unlock(&xs->mutex);
1422 			return -EBUSY;
1423 		}
1424 
1425 		q = (optname == XDP_UMEM_FILL_RING) ? &xs->fq_tmp :
1426 			&xs->cq_tmp;
1427 		err = xsk_init_queue(entries, q, true);
1428 		mutex_unlock(&xs->mutex);
1429 		return err;
1430 	}
1431 	default:
1432 		break;
1433 	}
1434 
1435 	return -ENOPROTOOPT;
1436 }
1437 
1438 static void xsk_enter_rxtx_offsets(struct xdp_ring_offset_v1 *ring)
1439 {
1440 	ring->producer = offsetof(struct xdp_rxtx_ring, ptrs.producer);
1441 	ring->consumer = offsetof(struct xdp_rxtx_ring, ptrs.consumer);
1442 	ring->desc = offsetof(struct xdp_rxtx_ring, desc);
1443 }
1444 
1445 static void xsk_enter_umem_offsets(struct xdp_ring_offset_v1 *ring)
1446 {
1447 	ring->producer = offsetof(struct xdp_umem_ring, ptrs.producer);
1448 	ring->consumer = offsetof(struct xdp_umem_ring, ptrs.consumer);
1449 	ring->desc = offsetof(struct xdp_umem_ring, desc);
1450 }
1451 
1452 struct xdp_statistics_v1 {
1453 	__u64 rx_dropped;
1454 	__u64 rx_invalid_descs;
1455 	__u64 tx_invalid_descs;
1456 };
1457 
1458 static int xsk_getsockopt(struct socket *sock, int level, int optname,
1459 			  char __user *optval, int __user *optlen)
1460 {
1461 	struct sock *sk = sock->sk;
1462 	struct xdp_sock *xs = xdp_sk(sk);
1463 	int len;
1464 
1465 	if (level != SOL_XDP)
1466 		return -ENOPROTOOPT;
1467 
1468 	if (get_user(len, optlen))
1469 		return -EFAULT;
1470 	if (len < 0)
1471 		return -EINVAL;
1472 
1473 	switch (optname) {
1474 	case XDP_STATISTICS:
1475 	{
1476 		struct xdp_statistics stats = {};
1477 		bool extra_stats = true;
1478 		size_t stats_size;
1479 
1480 		if (len < sizeof(struct xdp_statistics_v1)) {
1481 			return -EINVAL;
1482 		} else if (len < sizeof(stats)) {
1483 			extra_stats = false;
1484 			stats_size = sizeof(struct xdp_statistics_v1);
1485 		} else {
1486 			stats_size = sizeof(stats);
1487 		}
1488 
1489 		mutex_lock(&xs->mutex);
1490 		stats.rx_dropped = xs->rx_dropped;
1491 		if (extra_stats) {
1492 			stats.rx_ring_full = xs->rx_queue_full;
1493 			stats.rx_fill_ring_empty_descs =
1494 				xs->pool ? xskq_nb_queue_empty_descs(xs->pool->fq) : 0;
1495 			stats.tx_ring_empty_descs = xskq_nb_queue_empty_descs(xs->tx);
1496 		} else {
1497 			stats.rx_dropped += xs->rx_queue_full;
1498 		}
1499 		stats.rx_invalid_descs = xskq_nb_invalid_descs(xs->rx);
1500 		stats.tx_invalid_descs = xskq_nb_invalid_descs(xs->tx);
1501 		mutex_unlock(&xs->mutex);
1502 
1503 		if (copy_to_user(optval, &stats, stats_size))
1504 			return -EFAULT;
1505 		if (put_user(stats_size, optlen))
1506 			return -EFAULT;
1507 
1508 		return 0;
1509 	}
1510 	case XDP_MMAP_OFFSETS:
1511 	{
1512 		struct xdp_mmap_offsets off;
1513 		struct xdp_mmap_offsets_v1 off_v1;
1514 		bool flags_supported = true;
1515 		void *to_copy;
1516 
1517 		if (len < sizeof(off_v1))
1518 			return -EINVAL;
1519 		else if (len < sizeof(off))
1520 			flags_supported = false;
1521 
1522 		if (flags_supported) {
1523 			/* xdp_ring_offset is identical to xdp_ring_offset_v1
1524 			 * except for the flags field added to the end.
1525 			 */
1526 			xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *)
1527 					       &off.rx);
1528 			xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *)
1529 					       &off.tx);
1530 			xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *)
1531 					       &off.fr);
1532 			xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *)
1533 					       &off.cr);
1534 			off.rx.flags = offsetof(struct xdp_rxtx_ring,
1535 						ptrs.flags);
1536 			off.tx.flags = offsetof(struct xdp_rxtx_ring,
1537 						ptrs.flags);
1538 			off.fr.flags = offsetof(struct xdp_umem_ring,
1539 						ptrs.flags);
1540 			off.cr.flags = offsetof(struct xdp_umem_ring,
1541 						ptrs.flags);
1542 
1543 			len = sizeof(off);
1544 			to_copy = &off;
1545 		} else {
1546 			xsk_enter_rxtx_offsets(&off_v1.rx);
1547 			xsk_enter_rxtx_offsets(&off_v1.tx);
1548 			xsk_enter_umem_offsets(&off_v1.fr);
1549 			xsk_enter_umem_offsets(&off_v1.cr);
1550 
1551 			len = sizeof(off_v1);
1552 			to_copy = &off_v1;
1553 		}
1554 
1555 		if (copy_to_user(optval, to_copy, len))
1556 			return -EFAULT;
1557 		if (put_user(len, optlen))
1558 			return -EFAULT;
1559 
1560 		return 0;
1561 	}
1562 	case XDP_OPTIONS:
1563 	{
1564 		struct xdp_options opts = {};
1565 
1566 		if (len < sizeof(opts))
1567 			return -EINVAL;
1568 
1569 		mutex_lock(&xs->mutex);
1570 		if (xs->zc)
1571 			opts.flags |= XDP_OPTIONS_ZEROCOPY;
1572 		mutex_unlock(&xs->mutex);
1573 
1574 		len = sizeof(opts);
1575 		if (copy_to_user(optval, &opts, len))
1576 			return -EFAULT;
1577 		if (put_user(len, optlen))
1578 			return -EFAULT;
1579 
1580 		return 0;
1581 	}
1582 	default:
1583 		break;
1584 	}
1585 
1586 	return -EOPNOTSUPP;
1587 }
1588 
1589 static int xsk_mmap(struct file *file, struct socket *sock,
1590 		    struct vm_area_struct *vma)
1591 {
1592 	loff_t offset = (loff_t)vma->vm_pgoff << PAGE_SHIFT;
1593 	unsigned long size = vma->vm_end - vma->vm_start;
1594 	struct xdp_sock *xs = xdp_sk(sock->sk);
1595 	int state = READ_ONCE(xs->state);
1596 	struct xsk_queue *q = NULL;
1597 
1598 	if (state != XSK_READY && state != XSK_BOUND)
1599 		return -EBUSY;
1600 
1601 	if (offset == XDP_PGOFF_RX_RING) {
1602 		q = READ_ONCE(xs->rx);
1603 	} else if (offset == XDP_PGOFF_TX_RING) {
1604 		q = READ_ONCE(xs->tx);
1605 	} else {
1606 		/* Matches the smp_wmb() in XDP_UMEM_REG */
1607 		smp_rmb();
1608 		if (offset == XDP_UMEM_PGOFF_FILL_RING)
1609 			q = state == XSK_READY ? READ_ONCE(xs->fq_tmp) :
1610 						 READ_ONCE(xs->pool->fq);
1611 		else if (offset == XDP_UMEM_PGOFF_COMPLETION_RING)
1612 			q = state == XSK_READY ? READ_ONCE(xs->cq_tmp) :
1613 						 READ_ONCE(xs->pool->cq);
1614 	}
1615 
1616 	if (!q)
1617 		return -EINVAL;
1618 
1619 	/* Matches the smp_wmb() in xsk_init_queue */
1620 	smp_rmb();
1621 	if (size > q->ring_vmalloc_size)
1622 		return -EINVAL;
1623 
1624 	return remap_vmalloc_range(vma, q->ring, 0);
1625 }
1626 
1627 static int xsk_notifier(struct notifier_block *this,
1628 			unsigned long msg, void *ptr)
1629 {
1630 	struct net_device *dev = netdev_notifier_info_to_dev(ptr);
1631 	struct net *net = dev_net(dev);
1632 	struct sock *sk;
1633 
1634 	switch (msg) {
1635 	case NETDEV_UNREGISTER:
1636 		mutex_lock(&net->xdp.lock);
1637 		sk_for_each(sk, &net->xdp.list) {
1638 			struct xdp_sock *xs = xdp_sk(sk);
1639 
1640 			mutex_lock(&xs->mutex);
1641 			if (xs->dev == dev) {
1642 				sk->sk_err = ENETDOWN;
1643 				if (!sock_flag(sk, SOCK_DEAD))
1644 					sk_error_report(sk);
1645 
1646 				xsk_unbind_dev(xs);
1647 
1648 				/* Clear device references. */
1649 				xp_clear_dev(xs->pool);
1650 			}
1651 			mutex_unlock(&xs->mutex);
1652 		}
1653 		mutex_unlock(&net->xdp.lock);
1654 		break;
1655 	}
1656 	return NOTIFY_DONE;
1657 }
1658 
1659 static struct proto xsk_proto = {
1660 	.name =		"XDP",
1661 	.owner =	THIS_MODULE,
1662 	.obj_size =	sizeof(struct xdp_sock),
1663 };
1664 
1665 static const struct proto_ops xsk_proto_ops = {
1666 	.family		= PF_XDP,
1667 	.owner		= THIS_MODULE,
1668 	.release	= xsk_release,
1669 	.bind		= xsk_bind,
1670 	.connect	= sock_no_connect,
1671 	.socketpair	= sock_no_socketpair,
1672 	.accept		= sock_no_accept,
1673 	.getname	= sock_no_getname,
1674 	.poll		= xsk_poll,
1675 	.ioctl		= sock_no_ioctl,
1676 	.listen		= sock_no_listen,
1677 	.shutdown	= sock_no_shutdown,
1678 	.setsockopt	= xsk_setsockopt,
1679 	.getsockopt	= xsk_getsockopt,
1680 	.sendmsg	= xsk_sendmsg,
1681 	.recvmsg	= xsk_recvmsg,
1682 	.mmap		= xsk_mmap,
1683 };
1684 
1685 static void xsk_destruct(struct sock *sk)
1686 {
1687 	struct xdp_sock *xs = xdp_sk(sk);
1688 
1689 	if (!sock_flag(sk, SOCK_DEAD))
1690 		return;
1691 
1692 	if (!xp_put_pool(xs->pool))
1693 		xdp_put_umem(xs->umem, !xs->pool);
1694 }
1695 
1696 static int xsk_create(struct net *net, struct socket *sock, int protocol,
1697 		      int kern)
1698 {
1699 	struct xdp_sock *xs;
1700 	struct sock *sk;
1701 
1702 	if (!ns_capable(net->user_ns, CAP_NET_RAW))
1703 		return -EPERM;
1704 	if (sock->type != SOCK_RAW)
1705 		return -ESOCKTNOSUPPORT;
1706 
1707 	if (protocol)
1708 		return -EPROTONOSUPPORT;
1709 
1710 	sock->state = SS_UNCONNECTED;
1711 
1712 	sk = sk_alloc(net, PF_XDP, GFP_KERNEL, &xsk_proto, kern);
1713 	if (!sk)
1714 		return -ENOBUFS;
1715 
1716 	sock->ops = &xsk_proto_ops;
1717 
1718 	sock_init_data(sock, sk);
1719 
1720 	sk->sk_family = PF_XDP;
1721 
1722 	sk->sk_destruct = xsk_destruct;
1723 
1724 	sock_set_flag(sk, SOCK_RCU_FREE);
1725 
1726 	xs = xdp_sk(sk);
1727 	xs->state = XSK_READY;
1728 	mutex_init(&xs->mutex);
1729 	spin_lock_init(&xs->rx_lock);
1730 
1731 	INIT_LIST_HEAD(&xs->map_list);
1732 	spin_lock_init(&xs->map_list_lock);
1733 
1734 	mutex_lock(&net->xdp.lock);
1735 	sk_add_node_rcu(sk, &net->xdp.list);
1736 	mutex_unlock(&net->xdp.lock);
1737 
1738 	sock_prot_inuse_add(net, &xsk_proto, 1);
1739 
1740 	return 0;
1741 }
1742 
1743 static const struct net_proto_family xsk_family_ops = {
1744 	.family = PF_XDP,
1745 	.create = xsk_create,
1746 	.owner	= THIS_MODULE,
1747 };
1748 
1749 static struct notifier_block xsk_netdev_notifier = {
1750 	.notifier_call	= xsk_notifier,
1751 };
1752 
1753 static int __net_init xsk_net_init(struct net *net)
1754 {
1755 	mutex_init(&net->xdp.lock);
1756 	INIT_HLIST_HEAD(&net->xdp.list);
1757 	return 0;
1758 }
1759 
1760 static void __net_exit xsk_net_exit(struct net *net)
1761 {
1762 	WARN_ON_ONCE(!hlist_empty(&net->xdp.list));
1763 }
1764 
1765 static struct pernet_operations xsk_net_ops = {
1766 	.init = xsk_net_init,
1767 	.exit = xsk_net_exit,
1768 };
1769 
1770 static int __init xsk_init(void)
1771 {
1772 	int err, cpu;
1773 
1774 	err = proto_register(&xsk_proto, 0 /* no slab */);
1775 	if (err)
1776 		goto out;
1777 
1778 	err = sock_register(&xsk_family_ops);
1779 	if (err)
1780 		goto out_proto;
1781 
1782 	err = register_pernet_subsys(&xsk_net_ops);
1783 	if (err)
1784 		goto out_sk;
1785 
1786 	err = register_netdevice_notifier(&xsk_netdev_notifier);
1787 	if (err)
1788 		goto out_pernet;
1789 
1790 	for_each_possible_cpu(cpu)
1791 		INIT_LIST_HEAD(&per_cpu(xskmap_flush_list, cpu));
1792 	return 0;
1793 
1794 out_pernet:
1795 	unregister_pernet_subsys(&xsk_net_ops);
1796 out_sk:
1797 	sock_unregister(PF_XDP);
1798 out_proto:
1799 	proto_unregister(&xsk_proto);
1800 out:
1801 	return err;
1802 }
1803 
1804 fs_initcall(xsk_init);
1805