xref: /linux/net/tls/tls_device.c (revision bdd1a21b52557ea8f61d0a5dc2f77151b576eb70)
1 /* Copyright (c) 2018, Mellanox Technologies All rights reserved.
2  *
3  * This software is available to you under a choice of one of two
4  * licenses.  You may choose to be licensed under the terms of the GNU
5  * General Public License (GPL) Version 2, available from the file
6  * COPYING in the main directory of this source tree, or the
7  * OpenIB.org BSD license below:
8  *
9  *     Redistribution and use in source and binary forms, with or
10  *     without modification, are permitted provided that the following
11  *     conditions are met:
12  *
13  *      - Redistributions of source code must retain the above
14  *        copyright notice, this list of conditions and the following
15  *        disclaimer.
16  *
17  *      - Redistributions in binary form must reproduce the above
18  *        copyright notice, this list of conditions and the following
19  *        disclaimer in the documentation and/or other materials
20  *        provided with the distribution.
21  *
22  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
23  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
24  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
25  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
26  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
27  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
28  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
29  * SOFTWARE.
30  */
31 
32 #include <crypto/aead.h>
33 #include <linux/highmem.h>
34 #include <linux/module.h>
35 #include <linux/netdevice.h>
36 #include <net/dst.h>
37 #include <net/inet_connection_sock.h>
38 #include <net/tcp.h>
39 #include <net/tls.h>
40 
41 #include "trace.h"
42 
43 /* device_offload_lock is used to synchronize tls_dev_add
44  * against NETDEV_DOWN notifications.
45  */
46 static DECLARE_RWSEM(device_offload_lock);
47 
48 static void tls_device_gc_task(struct work_struct *work);
49 
50 static DECLARE_WORK(tls_device_gc_work, tls_device_gc_task);
51 static LIST_HEAD(tls_device_gc_list);
52 static LIST_HEAD(tls_device_list);
53 static LIST_HEAD(tls_device_down_list);
54 static DEFINE_SPINLOCK(tls_device_lock);
55 
56 static void tls_device_free_ctx(struct tls_context *ctx)
57 {
58 	if (ctx->tx_conf == TLS_HW) {
59 		kfree(tls_offload_ctx_tx(ctx));
60 		kfree(ctx->tx.rec_seq);
61 		kfree(ctx->tx.iv);
62 	}
63 
64 	if (ctx->rx_conf == TLS_HW)
65 		kfree(tls_offload_ctx_rx(ctx));
66 
67 	tls_ctx_free(NULL, ctx);
68 }
69 
70 static void tls_device_gc_task(struct work_struct *work)
71 {
72 	struct tls_context *ctx, *tmp;
73 	unsigned long flags;
74 	LIST_HEAD(gc_list);
75 
76 	spin_lock_irqsave(&tls_device_lock, flags);
77 	list_splice_init(&tls_device_gc_list, &gc_list);
78 	spin_unlock_irqrestore(&tls_device_lock, flags);
79 
80 	list_for_each_entry_safe(ctx, tmp, &gc_list, list) {
81 		struct net_device *netdev = ctx->netdev;
82 
83 		if (netdev && ctx->tx_conf == TLS_HW) {
84 			netdev->tlsdev_ops->tls_dev_del(netdev, ctx,
85 							TLS_OFFLOAD_CTX_DIR_TX);
86 			dev_put(netdev);
87 			ctx->netdev = NULL;
88 		}
89 
90 		list_del(&ctx->list);
91 		tls_device_free_ctx(ctx);
92 	}
93 }
94 
95 static void tls_device_queue_ctx_destruction(struct tls_context *ctx)
96 {
97 	unsigned long flags;
98 
99 	spin_lock_irqsave(&tls_device_lock, flags);
100 	list_move_tail(&ctx->list, &tls_device_gc_list);
101 
102 	/* schedule_work inside the spinlock
103 	 * to make sure tls_device_down waits for that work.
104 	 */
105 	schedule_work(&tls_device_gc_work);
106 
107 	spin_unlock_irqrestore(&tls_device_lock, flags);
108 }
109 
110 /* We assume that the socket is already connected */
111 static struct net_device *get_netdev_for_sock(struct sock *sk)
112 {
113 	struct dst_entry *dst = sk_dst_get(sk);
114 	struct net_device *netdev = NULL;
115 
116 	if (likely(dst)) {
117 		netdev = netdev_sk_get_lowest_dev(dst->dev, sk);
118 		dev_hold(netdev);
119 	}
120 
121 	dst_release(dst);
122 
123 	return netdev;
124 }
125 
126 static void destroy_record(struct tls_record_info *record)
127 {
128 	int i;
129 
130 	for (i = 0; i < record->num_frags; i++)
131 		__skb_frag_unref(&record->frags[i], false);
132 	kfree(record);
133 }
134 
135 static void delete_all_records(struct tls_offload_context_tx *offload_ctx)
136 {
137 	struct tls_record_info *info, *temp;
138 
139 	list_for_each_entry_safe(info, temp, &offload_ctx->records_list, list) {
140 		list_del(&info->list);
141 		destroy_record(info);
142 	}
143 
144 	offload_ctx->retransmit_hint = NULL;
145 }
146 
147 static void tls_icsk_clean_acked(struct sock *sk, u32 acked_seq)
148 {
149 	struct tls_context *tls_ctx = tls_get_ctx(sk);
150 	struct tls_record_info *info, *temp;
151 	struct tls_offload_context_tx *ctx;
152 	u64 deleted_records = 0;
153 	unsigned long flags;
154 
155 	if (!tls_ctx)
156 		return;
157 
158 	ctx = tls_offload_ctx_tx(tls_ctx);
159 
160 	spin_lock_irqsave(&ctx->lock, flags);
161 	info = ctx->retransmit_hint;
162 	if (info && !before(acked_seq, info->end_seq))
163 		ctx->retransmit_hint = NULL;
164 
165 	list_for_each_entry_safe(info, temp, &ctx->records_list, list) {
166 		if (before(acked_seq, info->end_seq))
167 			break;
168 		list_del(&info->list);
169 
170 		destroy_record(info);
171 		deleted_records++;
172 	}
173 
174 	ctx->unacked_record_sn += deleted_records;
175 	spin_unlock_irqrestore(&ctx->lock, flags);
176 }
177 
178 /* At this point, there should be no references on this
179  * socket and no in-flight SKBs associated with this
180  * socket, so it is safe to free all the resources.
181  */
182 void tls_device_sk_destruct(struct sock *sk)
183 {
184 	struct tls_context *tls_ctx = tls_get_ctx(sk);
185 	struct tls_offload_context_tx *ctx = tls_offload_ctx_tx(tls_ctx);
186 
187 	tls_ctx->sk_destruct(sk);
188 
189 	if (tls_ctx->tx_conf == TLS_HW) {
190 		if (ctx->open_record)
191 			destroy_record(ctx->open_record);
192 		delete_all_records(ctx);
193 		crypto_free_aead(ctx->aead_send);
194 		clean_acked_data_disable(inet_csk(sk));
195 	}
196 
197 	if (refcount_dec_and_test(&tls_ctx->refcount))
198 		tls_device_queue_ctx_destruction(tls_ctx);
199 }
200 EXPORT_SYMBOL_GPL(tls_device_sk_destruct);
201 
202 void tls_device_free_resources_tx(struct sock *sk)
203 {
204 	struct tls_context *tls_ctx = tls_get_ctx(sk);
205 
206 	tls_free_partial_record(sk, tls_ctx);
207 }
208 
209 void tls_offload_tx_resync_request(struct sock *sk, u32 got_seq, u32 exp_seq)
210 {
211 	struct tls_context *tls_ctx = tls_get_ctx(sk);
212 
213 	trace_tls_device_tx_resync_req(sk, got_seq, exp_seq);
214 	WARN_ON(test_and_set_bit(TLS_TX_SYNC_SCHED, &tls_ctx->flags));
215 }
216 EXPORT_SYMBOL_GPL(tls_offload_tx_resync_request);
217 
218 static void tls_device_resync_tx(struct sock *sk, struct tls_context *tls_ctx,
219 				 u32 seq)
220 {
221 	struct net_device *netdev;
222 	struct sk_buff *skb;
223 	int err = 0;
224 	u8 *rcd_sn;
225 
226 	skb = tcp_write_queue_tail(sk);
227 	if (skb)
228 		TCP_SKB_CB(skb)->eor = 1;
229 
230 	rcd_sn = tls_ctx->tx.rec_seq;
231 
232 	trace_tls_device_tx_resync_send(sk, seq, rcd_sn);
233 	down_read(&device_offload_lock);
234 	netdev = tls_ctx->netdev;
235 	if (netdev)
236 		err = netdev->tlsdev_ops->tls_dev_resync(netdev, sk, seq,
237 							 rcd_sn,
238 							 TLS_OFFLOAD_CTX_DIR_TX);
239 	up_read(&device_offload_lock);
240 	if (err)
241 		return;
242 
243 	clear_bit_unlock(TLS_TX_SYNC_SCHED, &tls_ctx->flags);
244 }
245 
246 static void tls_append_frag(struct tls_record_info *record,
247 			    struct page_frag *pfrag,
248 			    int size)
249 {
250 	skb_frag_t *frag;
251 
252 	frag = &record->frags[record->num_frags - 1];
253 	if (skb_frag_page(frag) == pfrag->page &&
254 	    skb_frag_off(frag) + skb_frag_size(frag) == pfrag->offset) {
255 		skb_frag_size_add(frag, size);
256 	} else {
257 		++frag;
258 		__skb_frag_set_page(frag, pfrag->page);
259 		skb_frag_off_set(frag, pfrag->offset);
260 		skb_frag_size_set(frag, size);
261 		++record->num_frags;
262 		get_page(pfrag->page);
263 	}
264 
265 	pfrag->offset += size;
266 	record->len += size;
267 }
268 
269 static int tls_push_record(struct sock *sk,
270 			   struct tls_context *ctx,
271 			   struct tls_offload_context_tx *offload_ctx,
272 			   struct tls_record_info *record,
273 			   int flags)
274 {
275 	struct tls_prot_info *prot = &ctx->prot_info;
276 	struct tcp_sock *tp = tcp_sk(sk);
277 	skb_frag_t *frag;
278 	int i;
279 
280 	record->end_seq = tp->write_seq + record->len;
281 	list_add_tail_rcu(&record->list, &offload_ctx->records_list);
282 	offload_ctx->open_record = NULL;
283 
284 	if (test_bit(TLS_TX_SYNC_SCHED, &ctx->flags))
285 		tls_device_resync_tx(sk, ctx, tp->write_seq);
286 
287 	tls_advance_record_sn(sk, prot, &ctx->tx);
288 
289 	for (i = 0; i < record->num_frags; i++) {
290 		frag = &record->frags[i];
291 		sg_unmark_end(&offload_ctx->sg_tx_data[i]);
292 		sg_set_page(&offload_ctx->sg_tx_data[i], skb_frag_page(frag),
293 			    skb_frag_size(frag), skb_frag_off(frag));
294 		sk_mem_charge(sk, skb_frag_size(frag));
295 		get_page(skb_frag_page(frag));
296 	}
297 	sg_mark_end(&offload_ctx->sg_tx_data[record->num_frags - 1]);
298 
299 	/* all ready, send */
300 	return tls_push_sg(sk, ctx, offload_ctx->sg_tx_data, 0, flags);
301 }
302 
303 static int tls_device_record_close(struct sock *sk,
304 				   struct tls_context *ctx,
305 				   struct tls_record_info *record,
306 				   struct page_frag *pfrag,
307 				   unsigned char record_type)
308 {
309 	struct tls_prot_info *prot = &ctx->prot_info;
310 	int ret;
311 
312 	/* append tag
313 	 * device will fill in the tag, we just need to append a placeholder
314 	 * use socket memory to improve coalescing (re-using a single buffer
315 	 * increases frag count)
316 	 * if we can't allocate memory now, steal some back from data
317 	 */
318 	if (likely(skb_page_frag_refill(prot->tag_size, pfrag,
319 					sk->sk_allocation))) {
320 		ret = 0;
321 		tls_append_frag(record, pfrag, prot->tag_size);
322 	} else {
323 		ret = prot->tag_size;
324 		if (record->len <= prot->overhead_size)
325 			return -ENOMEM;
326 	}
327 
328 	/* fill prepend */
329 	tls_fill_prepend(ctx, skb_frag_address(&record->frags[0]),
330 			 record->len - prot->overhead_size,
331 			 record_type);
332 	return ret;
333 }
334 
335 static int tls_create_new_record(struct tls_offload_context_tx *offload_ctx,
336 				 struct page_frag *pfrag,
337 				 size_t prepend_size)
338 {
339 	struct tls_record_info *record;
340 	skb_frag_t *frag;
341 
342 	record = kmalloc(sizeof(*record), GFP_KERNEL);
343 	if (!record)
344 		return -ENOMEM;
345 
346 	frag = &record->frags[0];
347 	__skb_frag_set_page(frag, pfrag->page);
348 	skb_frag_off_set(frag, pfrag->offset);
349 	skb_frag_size_set(frag, prepend_size);
350 
351 	get_page(pfrag->page);
352 	pfrag->offset += prepend_size;
353 
354 	record->num_frags = 1;
355 	record->len = prepend_size;
356 	offload_ctx->open_record = record;
357 	return 0;
358 }
359 
360 static int tls_do_allocation(struct sock *sk,
361 			     struct tls_offload_context_tx *offload_ctx,
362 			     struct page_frag *pfrag,
363 			     size_t prepend_size)
364 {
365 	int ret;
366 
367 	if (!offload_ctx->open_record) {
368 		if (unlikely(!skb_page_frag_refill(prepend_size, pfrag,
369 						   sk->sk_allocation))) {
370 			READ_ONCE(sk->sk_prot)->enter_memory_pressure(sk);
371 			sk_stream_moderate_sndbuf(sk);
372 			return -ENOMEM;
373 		}
374 
375 		ret = tls_create_new_record(offload_ctx, pfrag, prepend_size);
376 		if (ret)
377 			return ret;
378 
379 		if (pfrag->size > pfrag->offset)
380 			return 0;
381 	}
382 
383 	if (!sk_page_frag_refill(sk, pfrag))
384 		return -ENOMEM;
385 
386 	return 0;
387 }
388 
389 static int tls_device_copy_data(void *addr, size_t bytes, struct iov_iter *i)
390 {
391 	size_t pre_copy, nocache;
392 
393 	pre_copy = ~((unsigned long)addr - 1) & (SMP_CACHE_BYTES - 1);
394 	if (pre_copy) {
395 		pre_copy = min(pre_copy, bytes);
396 		if (copy_from_iter(addr, pre_copy, i) != pre_copy)
397 			return -EFAULT;
398 		bytes -= pre_copy;
399 		addr += pre_copy;
400 	}
401 
402 	nocache = round_down(bytes, SMP_CACHE_BYTES);
403 	if (copy_from_iter_nocache(addr, nocache, i) != nocache)
404 		return -EFAULT;
405 	bytes -= nocache;
406 	addr += nocache;
407 
408 	if (bytes && copy_from_iter(addr, bytes, i) != bytes)
409 		return -EFAULT;
410 
411 	return 0;
412 }
413 
414 static int tls_push_data(struct sock *sk,
415 			 struct iov_iter *msg_iter,
416 			 size_t size, int flags,
417 			 unsigned char record_type)
418 {
419 	struct tls_context *tls_ctx = tls_get_ctx(sk);
420 	struct tls_prot_info *prot = &tls_ctx->prot_info;
421 	struct tls_offload_context_tx *ctx = tls_offload_ctx_tx(tls_ctx);
422 	struct tls_record_info *record;
423 	int tls_push_record_flags;
424 	struct page_frag *pfrag;
425 	size_t orig_size = size;
426 	u32 max_open_record_len;
427 	bool more = false;
428 	bool done = false;
429 	int copy, rc = 0;
430 	long timeo;
431 
432 	if (flags &
433 	    ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL | MSG_SENDPAGE_NOTLAST))
434 		return -EOPNOTSUPP;
435 
436 	if (unlikely(sk->sk_err))
437 		return -sk->sk_err;
438 
439 	flags |= MSG_SENDPAGE_DECRYPTED;
440 	tls_push_record_flags = flags | MSG_SENDPAGE_NOTLAST;
441 
442 	timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
443 	if (tls_is_partially_sent_record(tls_ctx)) {
444 		rc = tls_push_partial_record(sk, tls_ctx, flags);
445 		if (rc < 0)
446 			return rc;
447 	}
448 
449 	pfrag = sk_page_frag(sk);
450 
451 	/* TLS_HEADER_SIZE is not counted as part of the TLS record, and
452 	 * we need to leave room for an authentication tag.
453 	 */
454 	max_open_record_len = TLS_MAX_PAYLOAD_SIZE +
455 			      prot->prepend_size;
456 	do {
457 		rc = tls_do_allocation(sk, ctx, pfrag, prot->prepend_size);
458 		if (unlikely(rc)) {
459 			rc = sk_stream_wait_memory(sk, &timeo);
460 			if (!rc)
461 				continue;
462 
463 			record = ctx->open_record;
464 			if (!record)
465 				break;
466 handle_error:
467 			if (record_type != TLS_RECORD_TYPE_DATA) {
468 				/* avoid sending partial
469 				 * record with type !=
470 				 * application_data
471 				 */
472 				size = orig_size;
473 				destroy_record(record);
474 				ctx->open_record = NULL;
475 			} else if (record->len > prot->prepend_size) {
476 				goto last_record;
477 			}
478 
479 			break;
480 		}
481 
482 		record = ctx->open_record;
483 		copy = min_t(size_t, size, (pfrag->size - pfrag->offset));
484 		copy = min_t(size_t, copy, (max_open_record_len - record->len));
485 
486 		rc = tls_device_copy_data(page_address(pfrag->page) +
487 					  pfrag->offset, copy, msg_iter);
488 		if (rc)
489 			goto handle_error;
490 		tls_append_frag(record, pfrag, copy);
491 
492 		size -= copy;
493 		if (!size) {
494 last_record:
495 			tls_push_record_flags = flags;
496 			if (flags & (MSG_SENDPAGE_NOTLAST | MSG_MORE)) {
497 				more = true;
498 				break;
499 			}
500 
501 			done = true;
502 		}
503 
504 		if (done || record->len >= max_open_record_len ||
505 		    (record->num_frags >= MAX_SKB_FRAGS - 1)) {
506 			rc = tls_device_record_close(sk, tls_ctx, record,
507 						     pfrag, record_type);
508 			if (rc) {
509 				if (rc > 0) {
510 					size += rc;
511 				} else {
512 					size = orig_size;
513 					destroy_record(record);
514 					ctx->open_record = NULL;
515 					break;
516 				}
517 			}
518 
519 			rc = tls_push_record(sk,
520 					     tls_ctx,
521 					     ctx,
522 					     record,
523 					     tls_push_record_flags);
524 			if (rc < 0)
525 				break;
526 		}
527 	} while (!done);
528 
529 	tls_ctx->pending_open_record_frags = more;
530 
531 	if (orig_size - size > 0)
532 		rc = orig_size - size;
533 
534 	return rc;
535 }
536 
537 int tls_device_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
538 {
539 	unsigned char record_type = TLS_RECORD_TYPE_DATA;
540 	struct tls_context *tls_ctx = tls_get_ctx(sk);
541 	int rc;
542 
543 	mutex_lock(&tls_ctx->tx_lock);
544 	lock_sock(sk);
545 
546 	if (unlikely(msg->msg_controllen)) {
547 		rc = tls_proccess_cmsg(sk, msg, &record_type);
548 		if (rc)
549 			goto out;
550 	}
551 
552 	rc = tls_push_data(sk, &msg->msg_iter, size,
553 			   msg->msg_flags, record_type);
554 
555 out:
556 	release_sock(sk);
557 	mutex_unlock(&tls_ctx->tx_lock);
558 	return rc;
559 }
560 
561 int tls_device_sendpage(struct sock *sk, struct page *page,
562 			int offset, size_t size, int flags)
563 {
564 	struct tls_context *tls_ctx = tls_get_ctx(sk);
565 	struct iov_iter	msg_iter;
566 	char *kaddr;
567 	struct kvec iov;
568 	int rc;
569 
570 	if (flags & MSG_SENDPAGE_NOTLAST)
571 		flags |= MSG_MORE;
572 
573 	mutex_lock(&tls_ctx->tx_lock);
574 	lock_sock(sk);
575 
576 	if (flags & MSG_OOB) {
577 		rc = -EOPNOTSUPP;
578 		goto out;
579 	}
580 
581 	kaddr = kmap(page);
582 	iov.iov_base = kaddr + offset;
583 	iov.iov_len = size;
584 	iov_iter_kvec(&msg_iter, WRITE, &iov, 1, size);
585 	rc = tls_push_data(sk, &msg_iter, size,
586 			   flags, TLS_RECORD_TYPE_DATA);
587 	kunmap(page);
588 
589 out:
590 	release_sock(sk);
591 	mutex_unlock(&tls_ctx->tx_lock);
592 	return rc;
593 }
594 
595 struct tls_record_info *tls_get_record(struct tls_offload_context_tx *context,
596 				       u32 seq, u64 *p_record_sn)
597 {
598 	u64 record_sn = context->hint_record_sn;
599 	struct tls_record_info *info, *last;
600 
601 	info = context->retransmit_hint;
602 	if (!info ||
603 	    before(seq, info->end_seq - info->len)) {
604 		/* if retransmit_hint is irrelevant start
605 		 * from the beginning of the list
606 		 */
607 		info = list_first_entry_or_null(&context->records_list,
608 						struct tls_record_info, list);
609 		if (!info)
610 			return NULL;
611 		/* send the start_marker record if seq number is before the
612 		 * tls offload start marker sequence number. This record is
613 		 * required to handle TCP packets which are before TLS offload
614 		 * started.
615 		 *  And if it's not start marker, look if this seq number
616 		 * belongs to the list.
617 		 */
618 		if (likely(!tls_record_is_start_marker(info))) {
619 			/* we have the first record, get the last record to see
620 			 * if this seq number belongs to the list.
621 			 */
622 			last = list_last_entry(&context->records_list,
623 					       struct tls_record_info, list);
624 
625 			if (!between(seq, tls_record_start_seq(info),
626 				     last->end_seq))
627 				return NULL;
628 		}
629 		record_sn = context->unacked_record_sn;
630 	}
631 
632 	/* We just need the _rcu for the READ_ONCE() */
633 	rcu_read_lock();
634 	list_for_each_entry_from_rcu(info, &context->records_list, list) {
635 		if (before(seq, info->end_seq)) {
636 			if (!context->retransmit_hint ||
637 			    after(info->end_seq,
638 				  context->retransmit_hint->end_seq)) {
639 				context->hint_record_sn = record_sn;
640 				context->retransmit_hint = info;
641 			}
642 			*p_record_sn = record_sn;
643 			goto exit_rcu_unlock;
644 		}
645 		record_sn++;
646 	}
647 	info = NULL;
648 
649 exit_rcu_unlock:
650 	rcu_read_unlock();
651 	return info;
652 }
653 EXPORT_SYMBOL(tls_get_record);
654 
655 static int tls_device_push_pending_record(struct sock *sk, int flags)
656 {
657 	struct iov_iter	msg_iter;
658 
659 	iov_iter_kvec(&msg_iter, WRITE, NULL, 0, 0);
660 	return tls_push_data(sk, &msg_iter, 0, flags, TLS_RECORD_TYPE_DATA);
661 }
662 
663 void tls_device_write_space(struct sock *sk, struct tls_context *ctx)
664 {
665 	if (tls_is_partially_sent_record(ctx)) {
666 		gfp_t sk_allocation = sk->sk_allocation;
667 
668 		WARN_ON_ONCE(sk->sk_write_pending);
669 
670 		sk->sk_allocation = GFP_ATOMIC;
671 		tls_push_partial_record(sk, ctx,
672 					MSG_DONTWAIT | MSG_NOSIGNAL |
673 					MSG_SENDPAGE_DECRYPTED);
674 		sk->sk_allocation = sk_allocation;
675 	}
676 }
677 
678 static void tls_device_resync_rx(struct tls_context *tls_ctx,
679 				 struct sock *sk, u32 seq, u8 *rcd_sn)
680 {
681 	struct tls_offload_context_rx *rx_ctx = tls_offload_ctx_rx(tls_ctx);
682 	struct net_device *netdev;
683 
684 	trace_tls_device_rx_resync_send(sk, seq, rcd_sn, rx_ctx->resync_type);
685 	rcu_read_lock();
686 	netdev = READ_ONCE(tls_ctx->netdev);
687 	if (netdev)
688 		netdev->tlsdev_ops->tls_dev_resync(netdev, sk, seq, rcd_sn,
689 						   TLS_OFFLOAD_CTX_DIR_RX);
690 	rcu_read_unlock();
691 	TLS_INC_STATS(sock_net(sk), LINUX_MIB_TLSRXDEVICERESYNC);
692 }
693 
694 static bool
695 tls_device_rx_resync_async(struct tls_offload_resync_async *resync_async,
696 			   s64 resync_req, u32 *seq, u16 *rcd_delta)
697 {
698 	u32 is_async = resync_req & RESYNC_REQ_ASYNC;
699 	u32 req_seq = resync_req >> 32;
700 	u32 req_end = req_seq + ((resync_req >> 16) & 0xffff);
701 	u16 i;
702 
703 	*rcd_delta = 0;
704 
705 	if (is_async) {
706 		/* shouldn't get to wraparound:
707 		 * too long in async stage, something bad happened
708 		 */
709 		if (WARN_ON_ONCE(resync_async->rcd_delta == USHRT_MAX))
710 			return false;
711 
712 		/* asynchronous stage: log all headers seq such that
713 		 * req_seq <= seq <= end_seq, and wait for real resync request
714 		 */
715 		if (before(*seq, req_seq))
716 			return false;
717 		if (!after(*seq, req_end) &&
718 		    resync_async->loglen < TLS_DEVICE_RESYNC_ASYNC_LOGMAX)
719 			resync_async->log[resync_async->loglen++] = *seq;
720 
721 		resync_async->rcd_delta++;
722 
723 		return false;
724 	}
725 
726 	/* synchronous stage: check against the logged entries and
727 	 * proceed to check the next entries if no match was found
728 	 */
729 	for (i = 0; i < resync_async->loglen; i++)
730 		if (req_seq == resync_async->log[i] &&
731 		    atomic64_try_cmpxchg(&resync_async->req, &resync_req, 0)) {
732 			*rcd_delta = resync_async->rcd_delta - i;
733 			*seq = req_seq;
734 			resync_async->loglen = 0;
735 			resync_async->rcd_delta = 0;
736 			return true;
737 		}
738 
739 	resync_async->loglen = 0;
740 	resync_async->rcd_delta = 0;
741 
742 	if (req_seq == *seq &&
743 	    atomic64_try_cmpxchg(&resync_async->req,
744 				 &resync_req, 0))
745 		return true;
746 
747 	return false;
748 }
749 
750 void tls_device_rx_resync_new_rec(struct sock *sk, u32 rcd_len, u32 seq)
751 {
752 	struct tls_context *tls_ctx = tls_get_ctx(sk);
753 	struct tls_offload_context_rx *rx_ctx;
754 	u8 rcd_sn[TLS_MAX_REC_SEQ_SIZE];
755 	u32 sock_data, is_req_pending;
756 	struct tls_prot_info *prot;
757 	s64 resync_req;
758 	u16 rcd_delta;
759 	u32 req_seq;
760 
761 	if (tls_ctx->rx_conf != TLS_HW)
762 		return;
763 	if (unlikely(test_bit(TLS_RX_DEV_DEGRADED, &tls_ctx->flags)))
764 		return;
765 
766 	prot = &tls_ctx->prot_info;
767 	rx_ctx = tls_offload_ctx_rx(tls_ctx);
768 	memcpy(rcd_sn, tls_ctx->rx.rec_seq, prot->rec_seq_size);
769 
770 	switch (rx_ctx->resync_type) {
771 	case TLS_OFFLOAD_SYNC_TYPE_DRIVER_REQ:
772 		resync_req = atomic64_read(&rx_ctx->resync_req);
773 		req_seq = resync_req >> 32;
774 		seq += TLS_HEADER_SIZE - 1;
775 		is_req_pending = resync_req;
776 
777 		if (likely(!is_req_pending) || req_seq != seq ||
778 		    !atomic64_try_cmpxchg(&rx_ctx->resync_req, &resync_req, 0))
779 			return;
780 		break;
781 	case TLS_OFFLOAD_SYNC_TYPE_CORE_NEXT_HINT:
782 		if (likely(!rx_ctx->resync_nh_do_now))
783 			return;
784 
785 		/* head of next rec is already in, note that the sock_inq will
786 		 * include the currently parsed message when called from parser
787 		 */
788 		sock_data = tcp_inq(sk);
789 		if (sock_data > rcd_len) {
790 			trace_tls_device_rx_resync_nh_delay(sk, sock_data,
791 							    rcd_len);
792 			return;
793 		}
794 
795 		rx_ctx->resync_nh_do_now = 0;
796 		seq += rcd_len;
797 		tls_bigint_increment(rcd_sn, prot->rec_seq_size);
798 		break;
799 	case TLS_OFFLOAD_SYNC_TYPE_DRIVER_REQ_ASYNC:
800 		resync_req = atomic64_read(&rx_ctx->resync_async->req);
801 		is_req_pending = resync_req;
802 		if (likely(!is_req_pending))
803 			return;
804 
805 		if (!tls_device_rx_resync_async(rx_ctx->resync_async,
806 						resync_req, &seq, &rcd_delta))
807 			return;
808 		tls_bigint_subtract(rcd_sn, rcd_delta);
809 		break;
810 	}
811 
812 	tls_device_resync_rx(tls_ctx, sk, seq, rcd_sn);
813 }
814 
815 static void tls_device_core_ctrl_rx_resync(struct tls_context *tls_ctx,
816 					   struct tls_offload_context_rx *ctx,
817 					   struct sock *sk, struct sk_buff *skb)
818 {
819 	struct strp_msg *rxm;
820 
821 	/* device will request resyncs by itself based on stream scan */
822 	if (ctx->resync_type != TLS_OFFLOAD_SYNC_TYPE_CORE_NEXT_HINT)
823 		return;
824 	/* already scheduled */
825 	if (ctx->resync_nh_do_now)
826 		return;
827 	/* seen decrypted fragments since last fully-failed record */
828 	if (ctx->resync_nh_reset) {
829 		ctx->resync_nh_reset = 0;
830 		ctx->resync_nh.decrypted_failed = 1;
831 		ctx->resync_nh.decrypted_tgt = TLS_DEVICE_RESYNC_NH_START_IVAL;
832 		return;
833 	}
834 
835 	if (++ctx->resync_nh.decrypted_failed <= ctx->resync_nh.decrypted_tgt)
836 		return;
837 
838 	/* doing resync, bump the next target in case it fails */
839 	if (ctx->resync_nh.decrypted_tgt < TLS_DEVICE_RESYNC_NH_MAX_IVAL)
840 		ctx->resync_nh.decrypted_tgt *= 2;
841 	else
842 		ctx->resync_nh.decrypted_tgt += TLS_DEVICE_RESYNC_NH_MAX_IVAL;
843 
844 	rxm = strp_msg(skb);
845 
846 	/* head of next rec is already in, parser will sync for us */
847 	if (tcp_inq(sk) > rxm->full_len) {
848 		trace_tls_device_rx_resync_nh_schedule(sk);
849 		ctx->resync_nh_do_now = 1;
850 	} else {
851 		struct tls_prot_info *prot = &tls_ctx->prot_info;
852 		u8 rcd_sn[TLS_MAX_REC_SEQ_SIZE];
853 
854 		memcpy(rcd_sn, tls_ctx->rx.rec_seq, prot->rec_seq_size);
855 		tls_bigint_increment(rcd_sn, prot->rec_seq_size);
856 
857 		tls_device_resync_rx(tls_ctx, sk, tcp_sk(sk)->copied_seq,
858 				     rcd_sn);
859 	}
860 }
861 
862 static int tls_device_reencrypt(struct sock *sk, struct sk_buff *skb)
863 {
864 	struct strp_msg *rxm = strp_msg(skb);
865 	int err = 0, offset = rxm->offset, copy, nsg, data_len, pos;
866 	struct sk_buff *skb_iter, *unused;
867 	struct scatterlist sg[1];
868 	char *orig_buf, *buf;
869 
870 	orig_buf = kmalloc(rxm->full_len + TLS_HEADER_SIZE +
871 			   TLS_CIPHER_AES_GCM_128_IV_SIZE, sk->sk_allocation);
872 	if (!orig_buf)
873 		return -ENOMEM;
874 	buf = orig_buf;
875 
876 	nsg = skb_cow_data(skb, 0, &unused);
877 	if (unlikely(nsg < 0)) {
878 		err = nsg;
879 		goto free_buf;
880 	}
881 
882 	sg_init_table(sg, 1);
883 	sg_set_buf(&sg[0], buf,
884 		   rxm->full_len + TLS_HEADER_SIZE +
885 		   TLS_CIPHER_AES_GCM_128_IV_SIZE);
886 	err = skb_copy_bits(skb, offset, buf,
887 			    TLS_HEADER_SIZE + TLS_CIPHER_AES_GCM_128_IV_SIZE);
888 	if (err)
889 		goto free_buf;
890 
891 	/* We are interested only in the decrypted data not the auth */
892 	err = decrypt_skb(sk, skb, sg);
893 	if (err != -EBADMSG)
894 		goto free_buf;
895 	else
896 		err = 0;
897 
898 	data_len = rxm->full_len - TLS_CIPHER_AES_GCM_128_TAG_SIZE;
899 
900 	if (skb_pagelen(skb) > offset) {
901 		copy = min_t(int, skb_pagelen(skb) - offset, data_len);
902 
903 		if (skb->decrypted) {
904 			err = skb_store_bits(skb, offset, buf, copy);
905 			if (err)
906 				goto free_buf;
907 		}
908 
909 		offset += copy;
910 		buf += copy;
911 	}
912 
913 	pos = skb_pagelen(skb);
914 	skb_walk_frags(skb, skb_iter) {
915 		int frag_pos;
916 
917 		/* Practically all frags must belong to msg if reencrypt
918 		 * is needed with current strparser and coalescing logic,
919 		 * but strparser may "get optimized", so let's be safe.
920 		 */
921 		if (pos + skb_iter->len <= offset)
922 			goto done_with_frag;
923 		if (pos >= data_len + rxm->offset)
924 			break;
925 
926 		frag_pos = offset - pos;
927 		copy = min_t(int, skb_iter->len - frag_pos,
928 			     data_len + rxm->offset - offset);
929 
930 		if (skb_iter->decrypted) {
931 			err = skb_store_bits(skb_iter, frag_pos, buf, copy);
932 			if (err)
933 				goto free_buf;
934 		}
935 
936 		offset += copy;
937 		buf += copy;
938 done_with_frag:
939 		pos += skb_iter->len;
940 	}
941 
942 free_buf:
943 	kfree(orig_buf);
944 	return err;
945 }
946 
947 int tls_device_decrypted(struct sock *sk, struct tls_context *tls_ctx,
948 			 struct sk_buff *skb, struct strp_msg *rxm)
949 {
950 	struct tls_offload_context_rx *ctx = tls_offload_ctx_rx(tls_ctx);
951 	int is_decrypted = skb->decrypted;
952 	int is_encrypted = !is_decrypted;
953 	struct sk_buff *skb_iter;
954 
955 	/* Check if all the data is decrypted already */
956 	skb_walk_frags(skb, skb_iter) {
957 		is_decrypted &= skb_iter->decrypted;
958 		is_encrypted &= !skb_iter->decrypted;
959 	}
960 
961 	trace_tls_device_decrypted(sk, tcp_sk(sk)->copied_seq - rxm->full_len,
962 				   tls_ctx->rx.rec_seq, rxm->full_len,
963 				   is_encrypted, is_decrypted);
964 
965 	ctx->sw.decrypted |= is_decrypted;
966 
967 	if (unlikely(test_bit(TLS_RX_DEV_DEGRADED, &tls_ctx->flags))) {
968 		if (likely(is_encrypted || is_decrypted))
969 			return 0;
970 
971 		/* After tls_device_down disables the offload, the next SKB will
972 		 * likely have initial fragments decrypted, and final ones not
973 		 * decrypted. We need to reencrypt that single SKB.
974 		 */
975 		return tls_device_reencrypt(sk, skb);
976 	}
977 
978 	/* Return immediately if the record is either entirely plaintext or
979 	 * entirely ciphertext. Otherwise handle reencrypt partially decrypted
980 	 * record.
981 	 */
982 	if (is_decrypted) {
983 		ctx->resync_nh_reset = 1;
984 		return 0;
985 	}
986 	if (is_encrypted) {
987 		tls_device_core_ctrl_rx_resync(tls_ctx, ctx, sk, skb);
988 		return 0;
989 	}
990 
991 	ctx->resync_nh_reset = 1;
992 	return tls_device_reencrypt(sk, skb);
993 }
994 
995 static void tls_device_attach(struct tls_context *ctx, struct sock *sk,
996 			      struct net_device *netdev)
997 {
998 	if (sk->sk_destruct != tls_device_sk_destruct) {
999 		refcount_set(&ctx->refcount, 1);
1000 		dev_hold(netdev);
1001 		ctx->netdev = netdev;
1002 		spin_lock_irq(&tls_device_lock);
1003 		list_add_tail(&ctx->list, &tls_device_list);
1004 		spin_unlock_irq(&tls_device_lock);
1005 
1006 		ctx->sk_destruct = sk->sk_destruct;
1007 		smp_store_release(&sk->sk_destruct, tls_device_sk_destruct);
1008 	}
1009 }
1010 
1011 int tls_set_device_offload(struct sock *sk, struct tls_context *ctx)
1012 {
1013 	u16 nonce_size, tag_size, iv_size, rec_seq_size, salt_size;
1014 	struct tls_context *tls_ctx = tls_get_ctx(sk);
1015 	struct tls_prot_info *prot = &tls_ctx->prot_info;
1016 	struct tls_record_info *start_marker_record;
1017 	struct tls_offload_context_tx *offload_ctx;
1018 	struct tls_crypto_info *crypto_info;
1019 	struct net_device *netdev;
1020 	char *iv, *rec_seq;
1021 	struct sk_buff *skb;
1022 	__be64 rcd_sn;
1023 	int rc;
1024 
1025 	if (!ctx)
1026 		return -EINVAL;
1027 
1028 	if (ctx->priv_ctx_tx)
1029 		return -EEXIST;
1030 
1031 	start_marker_record = kmalloc(sizeof(*start_marker_record), GFP_KERNEL);
1032 	if (!start_marker_record)
1033 		return -ENOMEM;
1034 
1035 	offload_ctx = kzalloc(TLS_OFFLOAD_CONTEXT_SIZE_TX, GFP_KERNEL);
1036 	if (!offload_ctx) {
1037 		rc = -ENOMEM;
1038 		goto free_marker_record;
1039 	}
1040 
1041 	crypto_info = &ctx->crypto_send.info;
1042 	if (crypto_info->version != TLS_1_2_VERSION) {
1043 		rc = -EOPNOTSUPP;
1044 		goto free_offload_ctx;
1045 	}
1046 
1047 	switch (crypto_info->cipher_type) {
1048 	case TLS_CIPHER_AES_GCM_128:
1049 		nonce_size = TLS_CIPHER_AES_GCM_128_IV_SIZE;
1050 		tag_size = TLS_CIPHER_AES_GCM_128_TAG_SIZE;
1051 		iv_size = TLS_CIPHER_AES_GCM_128_IV_SIZE;
1052 		iv = ((struct tls12_crypto_info_aes_gcm_128 *)crypto_info)->iv;
1053 		rec_seq_size = TLS_CIPHER_AES_GCM_128_REC_SEQ_SIZE;
1054 		salt_size = TLS_CIPHER_AES_GCM_128_SALT_SIZE;
1055 		rec_seq =
1056 		 ((struct tls12_crypto_info_aes_gcm_128 *)crypto_info)->rec_seq;
1057 		break;
1058 	default:
1059 		rc = -EINVAL;
1060 		goto free_offload_ctx;
1061 	}
1062 
1063 	/* Sanity-check the rec_seq_size for stack allocations */
1064 	if (rec_seq_size > TLS_MAX_REC_SEQ_SIZE) {
1065 		rc = -EINVAL;
1066 		goto free_offload_ctx;
1067 	}
1068 
1069 	prot->version = crypto_info->version;
1070 	prot->cipher_type = crypto_info->cipher_type;
1071 	prot->prepend_size = TLS_HEADER_SIZE + nonce_size;
1072 	prot->tag_size = tag_size;
1073 	prot->overhead_size = prot->prepend_size + prot->tag_size;
1074 	prot->iv_size = iv_size;
1075 	prot->salt_size = salt_size;
1076 	ctx->tx.iv = kmalloc(iv_size + TLS_CIPHER_AES_GCM_128_SALT_SIZE,
1077 			     GFP_KERNEL);
1078 	if (!ctx->tx.iv) {
1079 		rc = -ENOMEM;
1080 		goto free_offload_ctx;
1081 	}
1082 
1083 	memcpy(ctx->tx.iv + TLS_CIPHER_AES_GCM_128_SALT_SIZE, iv, iv_size);
1084 
1085 	prot->rec_seq_size = rec_seq_size;
1086 	ctx->tx.rec_seq = kmemdup(rec_seq, rec_seq_size, GFP_KERNEL);
1087 	if (!ctx->tx.rec_seq) {
1088 		rc = -ENOMEM;
1089 		goto free_iv;
1090 	}
1091 
1092 	rc = tls_sw_fallback_init(sk, offload_ctx, crypto_info);
1093 	if (rc)
1094 		goto free_rec_seq;
1095 
1096 	/* start at rec_seq - 1 to account for the start marker record */
1097 	memcpy(&rcd_sn, ctx->tx.rec_seq, sizeof(rcd_sn));
1098 	offload_ctx->unacked_record_sn = be64_to_cpu(rcd_sn) - 1;
1099 
1100 	start_marker_record->end_seq = tcp_sk(sk)->write_seq;
1101 	start_marker_record->len = 0;
1102 	start_marker_record->num_frags = 0;
1103 
1104 	INIT_LIST_HEAD(&offload_ctx->records_list);
1105 	list_add_tail(&start_marker_record->list, &offload_ctx->records_list);
1106 	spin_lock_init(&offload_ctx->lock);
1107 	sg_init_table(offload_ctx->sg_tx_data,
1108 		      ARRAY_SIZE(offload_ctx->sg_tx_data));
1109 
1110 	clean_acked_data_enable(inet_csk(sk), &tls_icsk_clean_acked);
1111 	ctx->push_pending_record = tls_device_push_pending_record;
1112 
1113 	/* TLS offload is greatly simplified if we don't send
1114 	 * SKBs where only part of the payload needs to be encrypted.
1115 	 * So mark the last skb in the write queue as end of record.
1116 	 */
1117 	skb = tcp_write_queue_tail(sk);
1118 	if (skb)
1119 		TCP_SKB_CB(skb)->eor = 1;
1120 
1121 	netdev = get_netdev_for_sock(sk);
1122 	if (!netdev) {
1123 		pr_err_ratelimited("%s: netdev not found\n", __func__);
1124 		rc = -EINVAL;
1125 		goto disable_cad;
1126 	}
1127 
1128 	if (!(netdev->features & NETIF_F_HW_TLS_TX)) {
1129 		rc = -EOPNOTSUPP;
1130 		goto release_netdev;
1131 	}
1132 
1133 	/* Avoid offloading if the device is down
1134 	 * We don't want to offload new flows after
1135 	 * the NETDEV_DOWN event
1136 	 *
1137 	 * device_offload_lock is taken in tls_devices's NETDEV_DOWN
1138 	 * handler thus protecting from the device going down before
1139 	 * ctx was added to tls_device_list.
1140 	 */
1141 	down_read(&device_offload_lock);
1142 	if (!(netdev->flags & IFF_UP)) {
1143 		rc = -EINVAL;
1144 		goto release_lock;
1145 	}
1146 
1147 	ctx->priv_ctx_tx = offload_ctx;
1148 	rc = netdev->tlsdev_ops->tls_dev_add(netdev, sk, TLS_OFFLOAD_CTX_DIR_TX,
1149 					     &ctx->crypto_send.info,
1150 					     tcp_sk(sk)->write_seq);
1151 	trace_tls_device_offload_set(sk, TLS_OFFLOAD_CTX_DIR_TX,
1152 				     tcp_sk(sk)->write_seq, rec_seq, rc);
1153 	if (rc)
1154 		goto release_lock;
1155 
1156 	tls_device_attach(ctx, sk, netdev);
1157 	up_read(&device_offload_lock);
1158 
1159 	/* following this assignment tls_is_sk_tx_device_offloaded
1160 	 * will return true and the context might be accessed
1161 	 * by the netdev's xmit function.
1162 	 */
1163 	smp_store_release(&sk->sk_validate_xmit_skb, tls_validate_xmit_skb);
1164 	dev_put(netdev);
1165 
1166 	return 0;
1167 
1168 release_lock:
1169 	up_read(&device_offload_lock);
1170 release_netdev:
1171 	dev_put(netdev);
1172 disable_cad:
1173 	clean_acked_data_disable(inet_csk(sk));
1174 	crypto_free_aead(offload_ctx->aead_send);
1175 free_rec_seq:
1176 	kfree(ctx->tx.rec_seq);
1177 free_iv:
1178 	kfree(ctx->tx.iv);
1179 free_offload_ctx:
1180 	kfree(offload_ctx);
1181 	ctx->priv_ctx_tx = NULL;
1182 free_marker_record:
1183 	kfree(start_marker_record);
1184 	return rc;
1185 }
1186 
1187 int tls_set_device_offload_rx(struct sock *sk, struct tls_context *ctx)
1188 {
1189 	struct tls12_crypto_info_aes_gcm_128 *info;
1190 	struct tls_offload_context_rx *context;
1191 	struct net_device *netdev;
1192 	int rc = 0;
1193 
1194 	if (ctx->crypto_recv.info.version != TLS_1_2_VERSION)
1195 		return -EOPNOTSUPP;
1196 
1197 	netdev = get_netdev_for_sock(sk);
1198 	if (!netdev) {
1199 		pr_err_ratelimited("%s: netdev not found\n", __func__);
1200 		return -EINVAL;
1201 	}
1202 
1203 	if (!(netdev->features & NETIF_F_HW_TLS_RX)) {
1204 		rc = -EOPNOTSUPP;
1205 		goto release_netdev;
1206 	}
1207 
1208 	/* Avoid offloading if the device is down
1209 	 * We don't want to offload new flows after
1210 	 * the NETDEV_DOWN event
1211 	 *
1212 	 * device_offload_lock is taken in tls_devices's NETDEV_DOWN
1213 	 * handler thus protecting from the device going down before
1214 	 * ctx was added to tls_device_list.
1215 	 */
1216 	down_read(&device_offload_lock);
1217 	if (!(netdev->flags & IFF_UP)) {
1218 		rc = -EINVAL;
1219 		goto release_lock;
1220 	}
1221 
1222 	context = kzalloc(TLS_OFFLOAD_CONTEXT_SIZE_RX, GFP_KERNEL);
1223 	if (!context) {
1224 		rc = -ENOMEM;
1225 		goto release_lock;
1226 	}
1227 	context->resync_nh_reset = 1;
1228 
1229 	ctx->priv_ctx_rx = context;
1230 	rc = tls_set_sw_offload(sk, ctx, 0);
1231 	if (rc)
1232 		goto release_ctx;
1233 
1234 	rc = netdev->tlsdev_ops->tls_dev_add(netdev, sk, TLS_OFFLOAD_CTX_DIR_RX,
1235 					     &ctx->crypto_recv.info,
1236 					     tcp_sk(sk)->copied_seq);
1237 	info = (void *)&ctx->crypto_recv.info;
1238 	trace_tls_device_offload_set(sk, TLS_OFFLOAD_CTX_DIR_RX,
1239 				     tcp_sk(sk)->copied_seq, info->rec_seq, rc);
1240 	if (rc)
1241 		goto free_sw_resources;
1242 
1243 	tls_device_attach(ctx, sk, netdev);
1244 	up_read(&device_offload_lock);
1245 
1246 	dev_put(netdev);
1247 
1248 	return 0;
1249 
1250 free_sw_resources:
1251 	up_read(&device_offload_lock);
1252 	tls_sw_free_resources_rx(sk);
1253 	down_read(&device_offload_lock);
1254 release_ctx:
1255 	ctx->priv_ctx_rx = NULL;
1256 release_lock:
1257 	up_read(&device_offload_lock);
1258 release_netdev:
1259 	dev_put(netdev);
1260 	return rc;
1261 }
1262 
1263 void tls_device_offload_cleanup_rx(struct sock *sk)
1264 {
1265 	struct tls_context *tls_ctx = tls_get_ctx(sk);
1266 	struct net_device *netdev;
1267 
1268 	down_read(&device_offload_lock);
1269 	netdev = tls_ctx->netdev;
1270 	if (!netdev)
1271 		goto out;
1272 
1273 	netdev->tlsdev_ops->tls_dev_del(netdev, tls_ctx,
1274 					TLS_OFFLOAD_CTX_DIR_RX);
1275 
1276 	if (tls_ctx->tx_conf != TLS_HW) {
1277 		dev_put(netdev);
1278 		tls_ctx->netdev = NULL;
1279 	} else {
1280 		set_bit(TLS_RX_DEV_CLOSED, &tls_ctx->flags);
1281 	}
1282 out:
1283 	up_read(&device_offload_lock);
1284 	tls_sw_release_resources_rx(sk);
1285 }
1286 
1287 static int tls_device_down(struct net_device *netdev)
1288 {
1289 	struct tls_context *ctx, *tmp;
1290 	unsigned long flags;
1291 	LIST_HEAD(list);
1292 
1293 	/* Request a write lock to block new offload attempts */
1294 	down_write(&device_offload_lock);
1295 
1296 	spin_lock_irqsave(&tls_device_lock, flags);
1297 	list_for_each_entry_safe(ctx, tmp, &tls_device_list, list) {
1298 		if (ctx->netdev != netdev ||
1299 		    !refcount_inc_not_zero(&ctx->refcount))
1300 			continue;
1301 
1302 		list_move(&ctx->list, &list);
1303 	}
1304 	spin_unlock_irqrestore(&tls_device_lock, flags);
1305 
1306 	list_for_each_entry_safe(ctx, tmp, &list, list)	{
1307 		/* Stop offloaded TX and switch to the fallback.
1308 		 * tls_is_sk_tx_device_offloaded will return false.
1309 		 */
1310 		WRITE_ONCE(ctx->sk->sk_validate_xmit_skb, tls_validate_xmit_skb_sw);
1311 
1312 		/* Stop the RX and TX resync.
1313 		 * tls_dev_resync must not be called after tls_dev_del.
1314 		 */
1315 		WRITE_ONCE(ctx->netdev, NULL);
1316 
1317 		/* Start skipping the RX resync logic completely. */
1318 		set_bit(TLS_RX_DEV_DEGRADED, &ctx->flags);
1319 
1320 		/* Sync with inflight packets. After this point:
1321 		 * TX: no non-encrypted packets will be passed to the driver.
1322 		 * RX: resync requests from the driver will be ignored.
1323 		 */
1324 		synchronize_net();
1325 
1326 		/* Release the offload context on the driver side. */
1327 		if (ctx->tx_conf == TLS_HW)
1328 			netdev->tlsdev_ops->tls_dev_del(netdev, ctx,
1329 							TLS_OFFLOAD_CTX_DIR_TX);
1330 		if (ctx->rx_conf == TLS_HW &&
1331 		    !test_bit(TLS_RX_DEV_CLOSED, &ctx->flags))
1332 			netdev->tlsdev_ops->tls_dev_del(netdev, ctx,
1333 							TLS_OFFLOAD_CTX_DIR_RX);
1334 
1335 		dev_put(netdev);
1336 
1337 		/* Move the context to a separate list for two reasons:
1338 		 * 1. When the context is deallocated, list_del is called.
1339 		 * 2. It's no longer an offloaded context, so we don't want to
1340 		 *    run offload-specific code on this context.
1341 		 */
1342 		spin_lock_irqsave(&tls_device_lock, flags);
1343 		list_move_tail(&ctx->list, &tls_device_down_list);
1344 		spin_unlock_irqrestore(&tls_device_lock, flags);
1345 
1346 		/* Device contexts for RX and TX will be freed in on sk_destruct
1347 		 * by tls_device_free_ctx. rx_conf and tx_conf stay in TLS_HW.
1348 		 */
1349 	}
1350 
1351 	up_write(&device_offload_lock);
1352 
1353 	flush_work(&tls_device_gc_work);
1354 
1355 	return NOTIFY_DONE;
1356 }
1357 
1358 static int tls_dev_event(struct notifier_block *this, unsigned long event,
1359 			 void *ptr)
1360 {
1361 	struct net_device *dev = netdev_notifier_info_to_dev(ptr);
1362 
1363 	if (!dev->tlsdev_ops &&
1364 	    !(dev->features & (NETIF_F_HW_TLS_RX | NETIF_F_HW_TLS_TX)))
1365 		return NOTIFY_DONE;
1366 
1367 	switch (event) {
1368 	case NETDEV_REGISTER:
1369 	case NETDEV_FEAT_CHANGE:
1370 		if (netif_is_bond_master(dev))
1371 			return NOTIFY_DONE;
1372 		if ((dev->features & NETIF_F_HW_TLS_RX) &&
1373 		    !dev->tlsdev_ops->tls_dev_resync)
1374 			return NOTIFY_BAD;
1375 
1376 		if  (dev->tlsdev_ops &&
1377 		     dev->tlsdev_ops->tls_dev_add &&
1378 		     dev->tlsdev_ops->tls_dev_del)
1379 			return NOTIFY_DONE;
1380 		else
1381 			return NOTIFY_BAD;
1382 	case NETDEV_DOWN:
1383 		return tls_device_down(dev);
1384 	}
1385 	return NOTIFY_DONE;
1386 }
1387 
1388 static struct notifier_block tls_dev_notifier = {
1389 	.notifier_call	= tls_dev_event,
1390 };
1391 
1392 void __init tls_device_init(void)
1393 {
1394 	register_netdevice_notifier(&tls_dev_notifier);
1395 }
1396 
1397 void __exit tls_device_cleanup(void)
1398 {
1399 	unregister_netdevice_notifier(&tls_dev_notifier);
1400 	flush_work(&tls_device_gc_work);
1401 	clean_acked_data_flush();
1402 }
1403