xref: /linux/net/tls/tls_device.c (revision 987b741c52c7c6c68d46fbaeb95b8d1087f10b7f)
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 DEFINE_SPINLOCK(tls_device_lock);
54 
55 static void tls_device_free_ctx(struct tls_context *ctx)
56 {
57 	if (ctx->tx_conf == TLS_HW) {
58 		kfree(tls_offload_ctx_tx(ctx));
59 		kfree(ctx->tx.rec_seq);
60 		kfree(ctx->tx.iv);
61 	}
62 
63 	if (ctx->rx_conf == TLS_HW)
64 		kfree(tls_offload_ctx_rx(ctx));
65 
66 	tls_ctx_free(NULL, ctx);
67 }
68 
69 static void tls_device_gc_task(struct work_struct *work)
70 {
71 	struct tls_context *ctx, *tmp;
72 	unsigned long flags;
73 	LIST_HEAD(gc_list);
74 
75 	spin_lock_irqsave(&tls_device_lock, flags);
76 	list_splice_init(&tls_device_gc_list, &gc_list);
77 	spin_unlock_irqrestore(&tls_device_lock, flags);
78 
79 	list_for_each_entry_safe(ctx, tmp, &gc_list, list) {
80 		struct net_device *netdev = ctx->netdev;
81 
82 		if (netdev && ctx->tx_conf == TLS_HW) {
83 			netdev->tlsdev_ops->tls_dev_del(netdev, ctx,
84 							TLS_OFFLOAD_CTX_DIR_TX);
85 			dev_put(netdev);
86 			ctx->netdev = NULL;
87 		}
88 
89 		list_del(&ctx->list);
90 		tls_device_free_ctx(ctx);
91 	}
92 }
93 
94 static void tls_device_queue_ctx_destruction(struct tls_context *ctx)
95 {
96 	unsigned long flags;
97 
98 	spin_lock_irqsave(&tls_device_lock, flags);
99 	list_move_tail(&ctx->list, &tls_device_gc_list);
100 
101 	/* schedule_work inside the spinlock
102 	 * to make sure tls_device_down waits for that work.
103 	 */
104 	schedule_work(&tls_device_gc_work);
105 
106 	spin_unlock_irqrestore(&tls_device_lock, flags);
107 }
108 
109 /* We assume that the socket is already connected */
110 static struct net_device *get_netdev_for_sock(struct sock *sk)
111 {
112 	struct dst_entry *dst = sk_dst_get(sk);
113 	struct net_device *netdev = NULL;
114 
115 	if (likely(dst)) {
116 		netdev = netdev_sk_get_lowest_dev(dst->dev, sk);
117 		dev_hold(netdev);
118 	}
119 
120 	dst_release(dst);
121 
122 	return netdev;
123 }
124 
125 static void destroy_record(struct tls_record_info *record)
126 {
127 	int i;
128 
129 	for (i = 0; i < record->num_frags; i++)
130 		__skb_frag_unref(&record->frags[i]);
131 	kfree(record);
132 }
133 
134 static void delete_all_records(struct tls_offload_context_tx *offload_ctx)
135 {
136 	struct tls_record_info *info, *temp;
137 
138 	list_for_each_entry_safe(info, temp, &offload_ctx->records_list, list) {
139 		list_del(&info->list);
140 		destroy_record(info);
141 	}
142 
143 	offload_ctx->retransmit_hint = NULL;
144 }
145 
146 static void tls_icsk_clean_acked(struct sock *sk, u32 acked_seq)
147 {
148 	struct tls_context *tls_ctx = tls_get_ctx(sk);
149 	struct tls_record_info *info, *temp;
150 	struct tls_offload_context_tx *ctx;
151 	u64 deleted_records = 0;
152 	unsigned long flags;
153 
154 	if (!tls_ctx)
155 		return;
156 
157 	ctx = tls_offload_ctx_tx(tls_ctx);
158 
159 	spin_lock_irqsave(&ctx->lock, flags);
160 	info = ctx->retransmit_hint;
161 	if (info && !before(acked_seq, info->end_seq))
162 		ctx->retransmit_hint = NULL;
163 
164 	list_for_each_entry_safe(info, temp, &ctx->records_list, list) {
165 		if (before(acked_seq, info->end_seq))
166 			break;
167 		list_del(&info->list);
168 
169 		destroy_record(info);
170 		deleted_records++;
171 	}
172 
173 	ctx->unacked_record_sn += deleted_records;
174 	spin_unlock_irqrestore(&ctx->lock, flags);
175 }
176 
177 /* At this point, there should be no references on this
178  * socket and no in-flight SKBs associated with this
179  * socket, so it is safe to free all the resources.
180  */
181 void tls_device_sk_destruct(struct sock *sk)
182 {
183 	struct tls_context *tls_ctx = tls_get_ctx(sk);
184 	struct tls_offload_context_tx *ctx = tls_offload_ctx_tx(tls_ctx);
185 
186 	tls_ctx->sk_destruct(sk);
187 
188 	if (tls_ctx->tx_conf == TLS_HW) {
189 		if (ctx->open_record)
190 			destroy_record(ctx->open_record);
191 		delete_all_records(ctx);
192 		crypto_free_aead(ctx->aead_send);
193 		clean_acked_data_disable(inet_csk(sk));
194 	}
195 
196 	if (refcount_dec_and_test(&tls_ctx->refcount))
197 		tls_device_queue_ctx_destruction(tls_ctx);
198 }
199 EXPORT_SYMBOL_GPL(tls_device_sk_destruct);
200 
201 void tls_device_free_resources_tx(struct sock *sk)
202 {
203 	struct tls_context *tls_ctx = tls_get_ctx(sk);
204 
205 	tls_free_partial_record(sk, tls_ctx);
206 }
207 
208 void tls_offload_tx_resync_request(struct sock *sk, u32 got_seq, u32 exp_seq)
209 {
210 	struct tls_context *tls_ctx = tls_get_ctx(sk);
211 
212 	trace_tls_device_tx_resync_req(sk, got_seq, exp_seq);
213 	WARN_ON(test_and_set_bit(TLS_TX_SYNC_SCHED, &tls_ctx->flags));
214 }
215 EXPORT_SYMBOL_GPL(tls_offload_tx_resync_request);
216 
217 static void tls_device_resync_tx(struct sock *sk, struct tls_context *tls_ctx,
218 				 u32 seq)
219 {
220 	struct net_device *netdev;
221 	struct sk_buff *skb;
222 	int err = 0;
223 	u8 *rcd_sn;
224 
225 	skb = tcp_write_queue_tail(sk);
226 	if (skb)
227 		TCP_SKB_CB(skb)->eor = 1;
228 
229 	rcd_sn = tls_ctx->tx.rec_seq;
230 
231 	trace_tls_device_tx_resync_send(sk, seq, rcd_sn);
232 	down_read(&device_offload_lock);
233 	netdev = tls_ctx->netdev;
234 	if (netdev)
235 		err = netdev->tlsdev_ops->tls_dev_resync(netdev, sk, seq,
236 							 rcd_sn,
237 							 TLS_OFFLOAD_CTX_DIR_TX);
238 	up_read(&device_offload_lock);
239 	if (err)
240 		return;
241 
242 	clear_bit_unlock(TLS_TX_SYNC_SCHED, &tls_ctx->flags);
243 }
244 
245 static void tls_append_frag(struct tls_record_info *record,
246 			    struct page_frag *pfrag,
247 			    int size)
248 {
249 	skb_frag_t *frag;
250 
251 	frag = &record->frags[record->num_frags - 1];
252 	if (skb_frag_page(frag) == pfrag->page &&
253 	    skb_frag_off(frag) + skb_frag_size(frag) == pfrag->offset) {
254 		skb_frag_size_add(frag, size);
255 	} else {
256 		++frag;
257 		__skb_frag_set_page(frag, pfrag->page);
258 		skb_frag_off_set(frag, pfrag->offset);
259 		skb_frag_size_set(frag, size);
260 		++record->num_frags;
261 		get_page(pfrag->page);
262 	}
263 
264 	pfrag->offset += size;
265 	record->len += size;
266 }
267 
268 static int tls_push_record(struct sock *sk,
269 			   struct tls_context *ctx,
270 			   struct tls_offload_context_tx *offload_ctx,
271 			   struct tls_record_info *record,
272 			   int flags)
273 {
274 	struct tls_prot_info *prot = &ctx->prot_info;
275 	struct tcp_sock *tp = tcp_sk(sk);
276 	skb_frag_t *frag;
277 	int i;
278 
279 	record->end_seq = tp->write_seq + record->len;
280 	list_add_tail_rcu(&record->list, &offload_ctx->records_list);
281 	offload_ctx->open_record = NULL;
282 
283 	if (test_bit(TLS_TX_SYNC_SCHED, &ctx->flags))
284 		tls_device_resync_tx(sk, ctx, tp->write_seq);
285 
286 	tls_advance_record_sn(sk, prot, &ctx->tx);
287 
288 	for (i = 0; i < record->num_frags; i++) {
289 		frag = &record->frags[i];
290 		sg_unmark_end(&offload_ctx->sg_tx_data[i]);
291 		sg_set_page(&offload_ctx->sg_tx_data[i], skb_frag_page(frag),
292 			    skb_frag_size(frag), skb_frag_off(frag));
293 		sk_mem_charge(sk, skb_frag_size(frag));
294 		get_page(skb_frag_page(frag));
295 	}
296 	sg_mark_end(&offload_ctx->sg_tx_data[record->num_frags - 1]);
297 
298 	/* all ready, send */
299 	return tls_push_sg(sk, ctx, offload_ctx->sg_tx_data, 0, flags);
300 }
301 
302 static int tls_device_record_close(struct sock *sk,
303 				   struct tls_context *ctx,
304 				   struct tls_record_info *record,
305 				   struct page_frag *pfrag,
306 				   unsigned char record_type)
307 {
308 	struct tls_prot_info *prot = &ctx->prot_info;
309 	int ret;
310 
311 	/* append tag
312 	 * device will fill in the tag, we just need to append a placeholder
313 	 * use socket memory to improve coalescing (re-using a single buffer
314 	 * increases frag count)
315 	 * if we can't allocate memory now, steal some back from data
316 	 */
317 	if (likely(skb_page_frag_refill(prot->tag_size, pfrag,
318 					sk->sk_allocation))) {
319 		ret = 0;
320 		tls_append_frag(record, pfrag, prot->tag_size);
321 	} else {
322 		ret = prot->tag_size;
323 		if (record->len <= prot->overhead_size)
324 			return -ENOMEM;
325 	}
326 
327 	/* fill prepend */
328 	tls_fill_prepend(ctx, skb_frag_address(&record->frags[0]),
329 			 record->len - prot->overhead_size,
330 			 record_type);
331 	return ret;
332 }
333 
334 static int tls_create_new_record(struct tls_offload_context_tx *offload_ctx,
335 				 struct page_frag *pfrag,
336 				 size_t prepend_size)
337 {
338 	struct tls_record_info *record;
339 	skb_frag_t *frag;
340 
341 	record = kmalloc(sizeof(*record), GFP_KERNEL);
342 	if (!record)
343 		return -ENOMEM;
344 
345 	frag = &record->frags[0];
346 	__skb_frag_set_page(frag, pfrag->page);
347 	skb_frag_off_set(frag, pfrag->offset);
348 	skb_frag_size_set(frag, prepend_size);
349 
350 	get_page(pfrag->page);
351 	pfrag->offset += prepend_size;
352 
353 	record->num_frags = 1;
354 	record->len = prepend_size;
355 	offload_ctx->open_record = record;
356 	return 0;
357 }
358 
359 static int tls_do_allocation(struct sock *sk,
360 			     struct tls_offload_context_tx *offload_ctx,
361 			     struct page_frag *pfrag,
362 			     size_t prepend_size)
363 {
364 	int ret;
365 
366 	if (!offload_ctx->open_record) {
367 		if (unlikely(!skb_page_frag_refill(prepend_size, pfrag,
368 						   sk->sk_allocation))) {
369 			READ_ONCE(sk->sk_prot)->enter_memory_pressure(sk);
370 			sk_stream_moderate_sndbuf(sk);
371 			return -ENOMEM;
372 		}
373 
374 		ret = tls_create_new_record(offload_ctx, pfrag, prepend_size);
375 		if (ret)
376 			return ret;
377 
378 		if (pfrag->size > pfrag->offset)
379 			return 0;
380 	}
381 
382 	if (!sk_page_frag_refill(sk, pfrag))
383 		return -ENOMEM;
384 
385 	return 0;
386 }
387 
388 static int tls_device_copy_data(void *addr, size_t bytes, struct iov_iter *i)
389 {
390 	size_t pre_copy, nocache;
391 
392 	pre_copy = ~((unsigned long)addr - 1) & (SMP_CACHE_BYTES - 1);
393 	if (pre_copy) {
394 		pre_copy = min(pre_copy, bytes);
395 		if (copy_from_iter(addr, pre_copy, i) != pre_copy)
396 			return -EFAULT;
397 		bytes -= pre_copy;
398 		addr += pre_copy;
399 	}
400 
401 	nocache = round_down(bytes, SMP_CACHE_BYTES);
402 	if (copy_from_iter_nocache(addr, nocache, i) != nocache)
403 		return -EFAULT;
404 	bytes -= nocache;
405 	addr += nocache;
406 
407 	if (bytes && copy_from_iter(addr, bytes, i) != bytes)
408 		return -EFAULT;
409 
410 	return 0;
411 }
412 
413 static int tls_push_data(struct sock *sk,
414 			 struct iov_iter *msg_iter,
415 			 size_t size, int flags,
416 			 unsigned char record_type)
417 {
418 	struct tls_context *tls_ctx = tls_get_ctx(sk);
419 	struct tls_prot_info *prot = &tls_ctx->prot_info;
420 	struct tls_offload_context_tx *ctx = tls_offload_ctx_tx(tls_ctx);
421 	struct tls_record_info *record;
422 	int tls_push_record_flags;
423 	struct page_frag *pfrag;
424 	size_t orig_size = size;
425 	u32 max_open_record_len;
426 	bool more = false;
427 	bool done = false;
428 	int copy, rc = 0;
429 	long timeo;
430 
431 	if (flags &
432 	    ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL | MSG_SENDPAGE_NOTLAST))
433 		return -EOPNOTSUPP;
434 
435 	if (unlikely(sk->sk_err))
436 		return -sk->sk_err;
437 
438 	flags |= MSG_SENDPAGE_DECRYPTED;
439 	tls_push_record_flags = flags | MSG_SENDPAGE_NOTLAST;
440 
441 	timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
442 	if (tls_is_partially_sent_record(tls_ctx)) {
443 		rc = tls_push_partial_record(sk, tls_ctx, flags);
444 		if (rc < 0)
445 			return rc;
446 	}
447 
448 	pfrag = sk_page_frag(sk);
449 
450 	/* TLS_HEADER_SIZE is not counted as part of the TLS record, and
451 	 * we need to leave room for an authentication tag.
452 	 */
453 	max_open_record_len = TLS_MAX_PAYLOAD_SIZE +
454 			      prot->prepend_size;
455 	do {
456 		rc = tls_do_allocation(sk, ctx, pfrag, prot->prepend_size);
457 		if (unlikely(rc)) {
458 			rc = sk_stream_wait_memory(sk, &timeo);
459 			if (!rc)
460 				continue;
461 
462 			record = ctx->open_record;
463 			if (!record)
464 				break;
465 handle_error:
466 			if (record_type != TLS_RECORD_TYPE_DATA) {
467 				/* avoid sending partial
468 				 * record with type !=
469 				 * application_data
470 				 */
471 				size = orig_size;
472 				destroy_record(record);
473 				ctx->open_record = NULL;
474 			} else if (record->len > prot->prepend_size) {
475 				goto last_record;
476 			}
477 
478 			break;
479 		}
480 
481 		record = ctx->open_record;
482 		copy = min_t(size_t, size, (pfrag->size - pfrag->offset));
483 		copy = min_t(size_t, copy, (max_open_record_len - record->len));
484 
485 		rc = tls_device_copy_data(page_address(pfrag->page) +
486 					  pfrag->offset, copy, msg_iter);
487 		if (rc)
488 			goto handle_error;
489 		tls_append_frag(record, pfrag, copy);
490 
491 		size -= copy;
492 		if (!size) {
493 last_record:
494 			tls_push_record_flags = flags;
495 			if (flags & (MSG_SENDPAGE_NOTLAST | MSG_MORE)) {
496 				more = true;
497 				break;
498 			}
499 
500 			done = true;
501 		}
502 
503 		if (done || record->len >= max_open_record_len ||
504 		    (record->num_frags >= MAX_SKB_FRAGS - 1)) {
505 			rc = tls_device_record_close(sk, tls_ctx, record,
506 						     pfrag, record_type);
507 			if (rc) {
508 				if (rc > 0) {
509 					size += rc;
510 				} else {
511 					size = orig_size;
512 					destroy_record(record);
513 					ctx->open_record = NULL;
514 					break;
515 				}
516 			}
517 
518 			rc = tls_push_record(sk,
519 					     tls_ctx,
520 					     ctx,
521 					     record,
522 					     tls_push_record_flags);
523 			if (rc < 0)
524 				break;
525 		}
526 	} while (!done);
527 
528 	tls_ctx->pending_open_record_frags = more;
529 
530 	if (orig_size - size > 0)
531 		rc = orig_size - size;
532 
533 	return rc;
534 }
535 
536 int tls_device_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
537 {
538 	unsigned char record_type = TLS_RECORD_TYPE_DATA;
539 	struct tls_context *tls_ctx = tls_get_ctx(sk);
540 	int rc;
541 
542 	mutex_lock(&tls_ctx->tx_lock);
543 	lock_sock(sk);
544 
545 	if (unlikely(msg->msg_controllen)) {
546 		rc = tls_proccess_cmsg(sk, msg, &record_type);
547 		if (rc)
548 			goto out;
549 	}
550 
551 	rc = tls_push_data(sk, &msg->msg_iter, size,
552 			   msg->msg_flags, record_type);
553 
554 out:
555 	release_sock(sk);
556 	mutex_unlock(&tls_ctx->tx_lock);
557 	return rc;
558 }
559 
560 int tls_device_sendpage(struct sock *sk, struct page *page,
561 			int offset, size_t size, int flags)
562 {
563 	struct tls_context *tls_ctx = tls_get_ctx(sk);
564 	struct iov_iter	msg_iter;
565 	char *kaddr;
566 	struct kvec iov;
567 	int rc;
568 
569 	if (flags & MSG_SENDPAGE_NOTLAST)
570 		flags |= MSG_MORE;
571 
572 	mutex_lock(&tls_ctx->tx_lock);
573 	lock_sock(sk);
574 
575 	if (flags & MSG_OOB) {
576 		rc = -EOPNOTSUPP;
577 		goto out;
578 	}
579 
580 	kaddr = kmap(page);
581 	iov.iov_base = kaddr + offset;
582 	iov.iov_len = size;
583 	iov_iter_kvec(&msg_iter, WRITE, &iov, 1, size);
584 	rc = tls_push_data(sk, &msg_iter, size,
585 			   flags, TLS_RECORD_TYPE_DATA);
586 	kunmap(page);
587 
588 out:
589 	release_sock(sk);
590 	mutex_unlock(&tls_ctx->tx_lock);
591 	return rc;
592 }
593 
594 struct tls_record_info *tls_get_record(struct tls_offload_context_tx *context,
595 				       u32 seq, u64 *p_record_sn)
596 {
597 	u64 record_sn = context->hint_record_sn;
598 	struct tls_record_info *info, *last;
599 
600 	info = context->retransmit_hint;
601 	if (!info ||
602 	    before(seq, info->end_seq - info->len)) {
603 		/* if retransmit_hint is irrelevant start
604 		 * from the beginning of the list
605 		 */
606 		info = list_first_entry_or_null(&context->records_list,
607 						struct tls_record_info, list);
608 		if (!info)
609 			return NULL;
610 		/* send the start_marker record if seq number is before the
611 		 * tls offload start marker sequence number. This record is
612 		 * required to handle TCP packets which are before TLS offload
613 		 * started.
614 		 *  And if it's not start marker, look if this seq number
615 		 * belongs to the list.
616 		 */
617 		if (likely(!tls_record_is_start_marker(info))) {
618 			/* we have the first record, get the last record to see
619 			 * if this seq number belongs to the list.
620 			 */
621 			last = list_last_entry(&context->records_list,
622 					       struct tls_record_info, list);
623 
624 			if (!between(seq, tls_record_start_seq(info),
625 				     last->end_seq))
626 				return NULL;
627 		}
628 		record_sn = context->unacked_record_sn;
629 	}
630 
631 	/* We just need the _rcu for the READ_ONCE() */
632 	rcu_read_lock();
633 	list_for_each_entry_from_rcu(info, &context->records_list, list) {
634 		if (before(seq, info->end_seq)) {
635 			if (!context->retransmit_hint ||
636 			    after(info->end_seq,
637 				  context->retransmit_hint->end_seq)) {
638 				context->hint_record_sn = record_sn;
639 				context->retransmit_hint = info;
640 			}
641 			*p_record_sn = record_sn;
642 			goto exit_rcu_unlock;
643 		}
644 		record_sn++;
645 	}
646 	info = NULL;
647 
648 exit_rcu_unlock:
649 	rcu_read_unlock();
650 	return info;
651 }
652 EXPORT_SYMBOL(tls_get_record);
653 
654 static int tls_device_push_pending_record(struct sock *sk, int flags)
655 {
656 	struct iov_iter	msg_iter;
657 
658 	iov_iter_kvec(&msg_iter, WRITE, NULL, 0, 0);
659 	return tls_push_data(sk, &msg_iter, 0, flags, TLS_RECORD_TYPE_DATA);
660 }
661 
662 void tls_device_write_space(struct sock *sk, struct tls_context *ctx)
663 {
664 	if (tls_is_partially_sent_record(ctx)) {
665 		gfp_t sk_allocation = sk->sk_allocation;
666 
667 		WARN_ON_ONCE(sk->sk_write_pending);
668 
669 		sk->sk_allocation = GFP_ATOMIC;
670 		tls_push_partial_record(sk, ctx,
671 					MSG_DONTWAIT | MSG_NOSIGNAL |
672 					MSG_SENDPAGE_DECRYPTED);
673 		sk->sk_allocation = sk_allocation;
674 	}
675 }
676 
677 static void tls_device_resync_rx(struct tls_context *tls_ctx,
678 				 struct sock *sk, u32 seq, u8 *rcd_sn)
679 {
680 	struct tls_offload_context_rx *rx_ctx = tls_offload_ctx_rx(tls_ctx);
681 	struct net_device *netdev;
682 
683 	if (WARN_ON(test_and_set_bit(TLS_RX_SYNC_RUNNING, &tls_ctx->flags)))
684 		return;
685 
686 	trace_tls_device_rx_resync_send(sk, seq, rcd_sn, rx_ctx->resync_type);
687 	netdev = READ_ONCE(tls_ctx->netdev);
688 	if (netdev)
689 		netdev->tlsdev_ops->tls_dev_resync(netdev, sk, seq, rcd_sn,
690 						   TLS_OFFLOAD_CTX_DIR_RX);
691 	clear_bit_unlock(TLS_RX_SYNC_RUNNING, &tls_ctx->flags);
692 	TLS_INC_STATS(sock_net(sk), LINUX_MIB_TLSRXDEVICERESYNC);
693 }
694 
695 static bool
696 tls_device_rx_resync_async(struct tls_offload_resync_async *resync_async,
697 			   s64 resync_req, u32 *seq, u16 *rcd_delta)
698 {
699 	u32 is_async = resync_req & RESYNC_REQ_ASYNC;
700 	u32 req_seq = resync_req >> 32;
701 	u32 req_end = req_seq + ((resync_req >> 16) & 0xffff);
702 	u16 i;
703 
704 	*rcd_delta = 0;
705 
706 	if (is_async) {
707 		/* shouldn't get to wraparound:
708 		 * too long in async stage, something bad happened
709 		 */
710 		if (WARN_ON_ONCE(resync_async->rcd_delta == USHRT_MAX))
711 			return false;
712 
713 		/* asynchronous stage: log all headers seq such that
714 		 * req_seq <= seq <= end_seq, and wait for real resync request
715 		 */
716 		if (before(*seq, req_seq))
717 			return false;
718 		if (!after(*seq, req_end) &&
719 		    resync_async->loglen < TLS_DEVICE_RESYNC_ASYNC_LOGMAX)
720 			resync_async->log[resync_async->loglen++] = *seq;
721 
722 		resync_async->rcd_delta++;
723 
724 		return false;
725 	}
726 
727 	/* synchronous stage: check against the logged entries and
728 	 * proceed to check the next entries if no match was found
729 	 */
730 	for (i = 0; i < resync_async->loglen; i++)
731 		if (req_seq == resync_async->log[i] &&
732 		    atomic64_try_cmpxchg(&resync_async->req, &resync_req, 0)) {
733 			*rcd_delta = resync_async->rcd_delta - i;
734 			*seq = req_seq;
735 			resync_async->loglen = 0;
736 			resync_async->rcd_delta = 0;
737 			return true;
738 		}
739 
740 	resync_async->loglen = 0;
741 	resync_async->rcd_delta = 0;
742 
743 	if (req_seq == *seq &&
744 	    atomic64_try_cmpxchg(&resync_async->req,
745 				 &resync_req, 0))
746 		return true;
747 
748 	return false;
749 }
750 
751 void tls_device_rx_resync_new_rec(struct sock *sk, u32 rcd_len, u32 seq)
752 {
753 	struct tls_context *tls_ctx = tls_get_ctx(sk);
754 	struct tls_offload_context_rx *rx_ctx;
755 	u8 rcd_sn[TLS_MAX_REC_SEQ_SIZE];
756 	u32 sock_data, is_req_pending;
757 	struct tls_prot_info *prot;
758 	s64 resync_req;
759 	u16 rcd_delta;
760 	u32 req_seq;
761 
762 	if (tls_ctx->rx_conf != TLS_HW)
763 		return;
764 
765 	prot = &tls_ctx->prot_info;
766 	rx_ctx = tls_offload_ctx_rx(tls_ctx);
767 	memcpy(rcd_sn, tls_ctx->rx.rec_seq, prot->rec_seq_size);
768 
769 	switch (rx_ctx->resync_type) {
770 	case TLS_OFFLOAD_SYNC_TYPE_DRIVER_REQ:
771 		resync_req = atomic64_read(&rx_ctx->resync_req);
772 		req_seq = resync_req >> 32;
773 		seq += TLS_HEADER_SIZE - 1;
774 		is_req_pending = resync_req;
775 
776 		if (likely(!is_req_pending) || req_seq != seq ||
777 		    !atomic64_try_cmpxchg(&rx_ctx->resync_req, &resync_req, 0))
778 			return;
779 		break;
780 	case TLS_OFFLOAD_SYNC_TYPE_CORE_NEXT_HINT:
781 		if (likely(!rx_ctx->resync_nh_do_now))
782 			return;
783 
784 		/* head of next rec is already in, note that the sock_inq will
785 		 * include the currently parsed message when called from parser
786 		 */
787 		sock_data = tcp_inq(sk);
788 		if (sock_data > rcd_len) {
789 			trace_tls_device_rx_resync_nh_delay(sk, sock_data,
790 							    rcd_len);
791 			return;
792 		}
793 
794 		rx_ctx->resync_nh_do_now = 0;
795 		seq += rcd_len;
796 		tls_bigint_increment(rcd_sn, prot->rec_seq_size);
797 		break;
798 	case TLS_OFFLOAD_SYNC_TYPE_DRIVER_REQ_ASYNC:
799 		resync_req = atomic64_read(&rx_ctx->resync_async->req);
800 		is_req_pending = resync_req;
801 		if (likely(!is_req_pending))
802 			return;
803 
804 		if (!tls_device_rx_resync_async(rx_ctx->resync_async,
805 						resync_req, &seq, &rcd_delta))
806 			return;
807 		tls_bigint_subtract(rcd_sn, rcd_delta);
808 		break;
809 	}
810 
811 	tls_device_resync_rx(tls_ctx, sk, seq, rcd_sn);
812 }
813 
814 static void tls_device_core_ctrl_rx_resync(struct tls_context *tls_ctx,
815 					   struct tls_offload_context_rx *ctx,
816 					   struct sock *sk, struct sk_buff *skb)
817 {
818 	struct strp_msg *rxm;
819 
820 	/* device will request resyncs by itself based on stream scan */
821 	if (ctx->resync_type != TLS_OFFLOAD_SYNC_TYPE_CORE_NEXT_HINT)
822 		return;
823 	/* already scheduled */
824 	if (ctx->resync_nh_do_now)
825 		return;
826 	/* seen decrypted fragments since last fully-failed record */
827 	if (ctx->resync_nh_reset) {
828 		ctx->resync_nh_reset = 0;
829 		ctx->resync_nh.decrypted_failed = 1;
830 		ctx->resync_nh.decrypted_tgt = TLS_DEVICE_RESYNC_NH_START_IVAL;
831 		return;
832 	}
833 
834 	if (++ctx->resync_nh.decrypted_failed <= ctx->resync_nh.decrypted_tgt)
835 		return;
836 
837 	/* doing resync, bump the next target in case it fails */
838 	if (ctx->resync_nh.decrypted_tgt < TLS_DEVICE_RESYNC_NH_MAX_IVAL)
839 		ctx->resync_nh.decrypted_tgt *= 2;
840 	else
841 		ctx->resync_nh.decrypted_tgt += TLS_DEVICE_RESYNC_NH_MAX_IVAL;
842 
843 	rxm = strp_msg(skb);
844 
845 	/* head of next rec is already in, parser will sync for us */
846 	if (tcp_inq(sk) > rxm->full_len) {
847 		trace_tls_device_rx_resync_nh_schedule(sk);
848 		ctx->resync_nh_do_now = 1;
849 	} else {
850 		struct tls_prot_info *prot = &tls_ctx->prot_info;
851 		u8 rcd_sn[TLS_MAX_REC_SEQ_SIZE];
852 
853 		memcpy(rcd_sn, tls_ctx->rx.rec_seq, prot->rec_seq_size);
854 		tls_bigint_increment(rcd_sn, prot->rec_seq_size);
855 
856 		tls_device_resync_rx(tls_ctx, sk, tcp_sk(sk)->copied_seq,
857 				     rcd_sn);
858 	}
859 }
860 
861 static int tls_device_reencrypt(struct sock *sk, struct sk_buff *skb)
862 {
863 	struct strp_msg *rxm = strp_msg(skb);
864 	int err = 0, offset = rxm->offset, copy, nsg, data_len, pos;
865 	struct sk_buff *skb_iter, *unused;
866 	struct scatterlist sg[1];
867 	char *orig_buf, *buf;
868 
869 	orig_buf = kmalloc(rxm->full_len + TLS_HEADER_SIZE +
870 			   TLS_CIPHER_AES_GCM_128_IV_SIZE, sk->sk_allocation);
871 	if (!orig_buf)
872 		return -ENOMEM;
873 	buf = orig_buf;
874 
875 	nsg = skb_cow_data(skb, 0, &unused);
876 	if (unlikely(nsg < 0)) {
877 		err = nsg;
878 		goto free_buf;
879 	}
880 
881 	sg_init_table(sg, 1);
882 	sg_set_buf(&sg[0], buf,
883 		   rxm->full_len + TLS_HEADER_SIZE +
884 		   TLS_CIPHER_AES_GCM_128_IV_SIZE);
885 	err = skb_copy_bits(skb, offset, buf,
886 			    TLS_HEADER_SIZE + TLS_CIPHER_AES_GCM_128_IV_SIZE);
887 	if (err)
888 		goto free_buf;
889 
890 	/* We are interested only in the decrypted data not the auth */
891 	err = decrypt_skb(sk, skb, sg);
892 	if (err != -EBADMSG)
893 		goto free_buf;
894 	else
895 		err = 0;
896 
897 	data_len = rxm->full_len - TLS_CIPHER_AES_GCM_128_TAG_SIZE;
898 
899 	if (skb_pagelen(skb) > offset) {
900 		copy = min_t(int, skb_pagelen(skb) - offset, data_len);
901 
902 		if (skb->decrypted) {
903 			err = skb_store_bits(skb, offset, buf, copy);
904 			if (err)
905 				goto free_buf;
906 		}
907 
908 		offset += copy;
909 		buf += copy;
910 	}
911 
912 	pos = skb_pagelen(skb);
913 	skb_walk_frags(skb, skb_iter) {
914 		int frag_pos;
915 
916 		/* Practically all frags must belong to msg if reencrypt
917 		 * is needed with current strparser and coalescing logic,
918 		 * but strparser may "get optimized", so let's be safe.
919 		 */
920 		if (pos + skb_iter->len <= offset)
921 			goto done_with_frag;
922 		if (pos >= data_len + rxm->offset)
923 			break;
924 
925 		frag_pos = offset - pos;
926 		copy = min_t(int, skb_iter->len - frag_pos,
927 			     data_len + rxm->offset - offset);
928 
929 		if (skb_iter->decrypted) {
930 			err = skb_store_bits(skb_iter, frag_pos, buf, copy);
931 			if (err)
932 				goto free_buf;
933 		}
934 
935 		offset += copy;
936 		buf += copy;
937 done_with_frag:
938 		pos += skb_iter->len;
939 	}
940 
941 free_buf:
942 	kfree(orig_buf);
943 	return err;
944 }
945 
946 int tls_device_decrypted(struct sock *sk, struct tls_context *tls_ctx,
947 			 struct sk_buff *skb, struct strp_msg *rxm)
948 {
949 	struct tls_offload_context_rx *ctx = tls_offload_ctx_rx(tls_ctx);
950 	int is_decrypted = skb->decrypted;
951 	int is_encrypted = !is_decrypted;
952 	struct sk_buff *skb_iter;
953 
954 	/* Check if all the data is decrypted already */
955 	skb_walk_frags(skb, skb_iter) {
956 		is_decrypted &= skb_iter->decrypted;
957 		is_encrypted &= !skb_iter->decrypted;
958 	}
959 
960 	trace_tls_device_decrypted(sk, tcp_sk(sk)->copied_seq - rxm->full_len,
961 				   tls_ctx->rx.rec_seq, rxm->full_len,
962 				   is_encrypted, is_decrypted);
963 
964 	ctx->sw.decrypted |= is_decrypted;
965 
966 	/* Return immediately if the record is either entirely plaintext or
967 	 * entirely ciphertext. Otherwise handle reencrypt partially decrypted
968 	 * record.
969 	 */
970 	if (is_decrypted) {
971 		ctx->resync_nh_reset = 1;
972 		return 0;
973 	}
974 	if (is_encrypted) {
975 		tls_device_core_ctrl_rx_resync(tls_ctx, ctx, sk, skb);
976 		return 0;
977 	}
978 
979 	ctx->resync_nh_reset = 1;
980 	return tls_device_reencrypt(sk, skb);
981 }
982 
983 static void tls_device_attach(struct tls_context *ctx, struct sock *sk,
984 			      struct net_device *netdev)
985 {
986 	if (sk->sk_destruct != tls_device_sk_destruct) {
987 		refcount_set(&ctx->refcount, 1);
988 		dev_hold(netdev);
989 		ctx->netdev = netdev;
990 		spin_lock_irq(&tls_device_lock);
991 		list_add_tail(&ctx->list, &tls_device_list);
992 		spin_unlock_irq(&tls_device_lock);
993 
994 		ctx->sk_destruct = sk->sk_destruct;
995 		smp_store_release(&sk->sk_destruct, tls_device_sk_destruct);
996 	}
997 }
998 
999 int tls_set_device_offload(struct sock *sk, struct tls_context *ctx)
1000 {
1001 	u16 nonce_size, tag_size, iv_size, rec_seq_size, salt_size;
1002 	struct tls_context *tls_ctx = tls_get_ctx(sk);
1003 	struct tls_prot_info *prot = &tls_ctx->prot_info;
1004 	struct tls_record_info *start_marker_record;
1005 	struct tls_offload_context_tx *offload_ctx;
1006 	struct tls_crypto_info *crypto_info;
1007 	struct net_device *netdev;
1008 	char *iv, *rec_seq;
1009 	struct sk_buff *skb;
1010 	__be64 rcd_sn;
1011 	int rc;
1012 
1013 	if (!ctx)
1014 		return -EINVAL;
1015 
1016 	if (ctx->priv_ctx_tx)
1017 		return -EEXIST;
1018 
1019 	start_marker_record = kmalloc(sizeof(*start_marker_record), GFP_KERNEL);
1020 	if (!start_marker_record)
1021 		return -ENOMEM;
1022 
1023 	offload_ctx = kzalloc(TLS_OFFLOAD_CONTEXT_SIZE_TX, GFP_KERNEL);
1024 	if (!offload_ctx) {
1025 		rc = -ENOMEM;
1026 		goto free_marker_record;
1027 	}
1028 
1029 	crypto_info = &ctx->crypto_send.info;
1030 	if (crypto_info->version != TLS_1_2_VERSION) {
1031 		rc = -EOPNOTSUPP;
1032 		goto free_offload_ctx;
1033 	}
1034 
1035 	switch (crypto_info->cipher_type) {
1036 	case TLS_CIPHER_AES_GCM_128:
1037 		nonce_size = TLS_CIPHER_AES_GCM_128_IV_SIZE;
1038 		tag_size = TLS_CIPHER_AES_GCM_128_TAG_SIZE;
1039 		iv_size = TLS_CIPHER_AES_GCM_128_IV_SIZE;
1040 		iv = ((struct tls12_crypto_info_aes_gcm_128 *)crypto_info)->iv;
1041 		rec_seq_size = TLS_CIPHER_AES_GCM_128_REC_SEQ_SIZE;
1042 		salt_size = TLS_CIPHER_AES_GCM_128_SALT_SIZE;
1043 		rec_seq =
1044 		 ((struct tls12_crypto_info_aes_gcm_128 *)crypto_info)->rec_seq;
1045 		break;
1046 	default:
1047 		rc = -EINVAL;
1048 		goto free_offload_ctx;
1049 	}
1050 
1051 	/* Sanity-check the rec_seq_size for stack allocations */
1052 	if (rec_seq_size > TLS_MAX_REC_SEQ_SIZE) {
1053 		rc = -EINVAL;
1054 		goto free_offload_ctx;
1055 	}
1056 
1057 	prot->version = crypto_info->version;
1058 	prot->cipher_type = crypto_info->cipher_type;
1059 	prot->prepend_size = TLS_HEADER_SIZE + nonce_size;
1060 	prot->tag_size = tag_size;
1061 	prot->overhead_size = prot->prepend_size + prot->tag_size;
1062 	prot->iv_size = iv_size;
1063 	prot->salt_size = salt_size;
1064 	ctx->tx.iv = kmalloc(iv_size + TLS_CIPHER_AES_GCM_128_SALT_SIZE,
1065 			     GFP_KERNEL);
1066 	if (!ctx->tx.iv) {
1067 		rc = -ENOMEM;
1068 		goto free_offload_ctx;
1069 	}
1070 
1071 	memcpy(ctx->tx.iv + TLS_CIPHER_AES_GCM_128_SALT_SIZE, iv, iv_size);
1072 
1073 	prot->rec_seq_size = rec_seq_size;
1074 	ctx->tx.rec_seq = kmemdup(rec_seq, rec_seq_size, GFP_KERNEL);
1075 	if (!ctx->tx.rec_seq) {
1076 		rc = -ENOMEM;
1077 		goto free_iv;
1078 	}
1079 
1080 	rc = tls_sw_fallback_init(sk, offload_ctx, crypto_info);
1081 	if (rc)
1082 		goto free_rec_seq;
1083 
1084 	/* start at rec_seq - 1 to account for the start marker record */
1085 	memcpy(&rcd_sn, ctx->tx.rec_seq, sizeof(rcd_sn));
1086 	offload_ctx->unacked_record_sn = be64_to_cpu(rcd_sn) - 1;
1087 
1088 	start_marker_record->end_seq = tcp_sk(sk)->write_seq;
1089 	start_marker_record->len = 0;
1090 	start_marker_record->num_frags = 0;
1091 
1092 	INIT_LIST_HEAD(&offload_ctx->records_list);
1093 	list_add_tail(&start_marker_record->list, &offload_ctx->records_list);
1094 	spin_lock_init(&offload_ctx->lock);
1095 	sg_init_table(offload_ctx->sg_tx_data,
1096 		      ARRAY_SIZE(offload_ctx->sg_tx_data));
1097 
1098 	clean_acked_data_enable(inet_csk(sk), &tls_icsk_clean_acked);
1099 	ctx->push_pending_record = tls_device_push_pending_record;
1100 
1101 	/* TLS offload is greatly simplified if we don't send
1102 	 * SKBs where only part of the payload needs to be encrypted.
1103 	 * So mark the last skb in the write queue as end of record.
1104 	 */
1105 	skb = tcp_write_queue_tail(sk);
1106 	if (skb)
1107 		TCP_SKB_CB(skb)->eor = 1;
1108 
1109 	netdev = get_netdev_for_sock(sk);
1110 	if (!netdev) {
1111 		pr_err_ratelimited("%s: netdev not found\n", __func__);
1112 		rc = -EINVAL;
1113 		goto disable_cad;
1114 	}
1115 
1116 	if (!(netdev->features & NETIF_F_HW_TLS_TX)) {
1117 		rc = -EOPNOTSUPP;
1118 		goto release_netdev;
1119 	}
1120 
1121 	/* Avoid offloading if the device is down
1122 	 * We don't want to offload new flows after
1123 	 * the NETDEV_DOWN event
1124 	 *
1125 	 * device_offload_lock is taken in tls_devices's NETDEV_DOWN
1126 	 * handler thus protecting from the device going down before
1127 	 * ctx was added to tls_device_list.
1128 	 */
1129 	down_read(&device_offload_lock);
1130 	if (!(netdev->flags & IFF_UP)) {
1131 		rc = -EINVAL;
1132 		goto release_lock;
1133 	}
1134 
1135 	ctx->priv_ctx_tx = offload_ctx;
1136 	rc = netdev->tlsdev_ops->tls_dev_add(netdev, sk, TLS_OFFLOAD_CTX_DIR_TX,
1137 					     &ctx->crypto_send.info,
1138 					     tcp_sk(sk)->write_seq);
1139 	trace_tls_device_offload_set(sk, TLS_OFFLOAD_CTX_DIR_TX,
1140 				     tcp_sk(sk)->write_seq, rec_seq, rc);
1141 	if (rc)
1142 		goto release_lock;
1143 
1144 	tls_device_attach(ctx, sk, netdev);
1145 	up_read(&device_offload_lock);
1146 
1147 	/* following this assignment tls_is_sk_tx_device_offloaded
1148 	 * will return true and the context might be accessed
1149 	 * by the netdev's xmit function.
1150 	 */
1151 	smp_store_release(&sk->sk_validate_xmit_skb, tls_validate_xmit_skb);
1152 	dev_put(netdev);
1153 
1154 	return 0;
1155 
1156 release_lock:
1157 	up_read(&device_offload_lock);
1158 release_netdev:
1159 	dev_put(netdev);
1160 disable_cad:
1161 	clean_acked_data_disable(inet_csk(sk));
1162 	crypto_free_aead(offload_ctx->aead_send);
1163 free_rec_seq:
1164 	kfree(ctx->tx.rec_seq);
1165 free_iv:
1166 	kfree(ctx->tx.iv);
1167 free_offload_ctx:
1168 	kfree(offload_ctx);
1169 	ctx->priv_ctx_tx = NULL;
1170 free_marker_record:
1171 	kfree(start_marker_record);
1172 	return rc;
1173 }
1174 
1175 int tls_set_device_offload_rx(struct sock *sk, struct tls_context *ctx)
1176 {
1177 	struct tls12_crypto_info_aes_gcm_128 *info;
1178 	struct tls_offload_context_rx *context;
1179 	struct net_device *netdev;
1180 	int rc = 0;
1181 
1182 	if (ctx->crypto_recv.info.version != TLS_1_2_VERSION)
1183 		return -EOPNOTSUPP;
1184 
1185 	netdev = get_netdev_for_sock(sk);
1186 	if (!netdev) {
1187 		pr_err_ratelimited("%s: netdev not found\n", __func__);
1188 		return -EINVAL;
1189 	}
1190 
1191 	if (!(netdev->features & NETIF_F_HW_TLS_RX)) {
1192 		rc = -EOPNOTSUPP;
1193 		goto release_netdev;
1194 	}
1195 
1196 	/* Avoid offloading if the device is down
1197 	 * We don't want to offload new flows after
1198 	 * the NETDEV_DOWN event
1199 	 *
1200 	 * device_offload_lock is taken in tls_devices's NETDEV_DOWN
1201 	 * handler thus protecting from the device going down before
1202 	 * ctx was added to tls_device_list.
1203 	 */
1204 	down_read(&device_offload_lock);
1205 	if (!(netdev->flags & IFF_UP)) {
1206 		rc = -EINVAL;
1207 		goto release_lock;
1208 	}
1209 
1210 	context = kzalloc(TLS_OFFLOAD_CONTEXT_SIZE_RX, GFP_KERNEL);
1211 	if (!context) {
1212 		rc = -ENOMEM;
1213 		goto release_lock;
1214 	}
1215 	context->resync_nh_reset = 1;
1216 
1217 	ctx->priv_ctx_rx = context;
1218 	rc = tls_set_sw_offload(sk, ctx, 0);
1219 	if (rc)
1220 		goto release_ctx;
1221 
1222 	rc = netdev->tlsdev_ops->tls_dev_add(netdev, sk, TLS_OFFLOAD_CTX_DIR_RX,
1223 					     &ctx->crypto_recv.info,
1224 					     tcp_sk(sk)->copied_seq);
1225 	info = (void *)&ctx->crypto_recv.info;
1226 	trace_tls_device_offload_set(sk, TLS_OFFLOAD_CTX_DIR_RX,
1227 				     tcp_sk(sk)->copied_seq, info->rec_seq, rc);
1228 	if (rc)
1229 		goto free_sw_resources;
1230 
1231 	tls_device_attach(ctx, sk, netdev);
1232 	up_read(&device_offload_lock);
1233 
1234 	dev_put(netdev);
1235 
1236 	return 0;
1237 
1238 free_sw_resources:
1239 	up_read(&device_offload_lock);
1240 	tls_sw_free_resources_rx(sk);
1241 	down_read(&device_offload_lock);
1242 release_ctx:
1243 	ctx->priv_ctx_rx = NULL;
1244 release_lock:
1245 	up_read(&device_offload_lock);
1246 release_netdev:
1247 	dev_put(netdev);
1248 	return rc;
1249 }
1250 
1251 void tls_device_offload_cleanup_rx(struct sock *sk)
1252 {
1253 	struct tls_context *tls_ctx = tls_get_ctx(sk);
1254 	struct net_device *netdev;
1255 
1256 	down_read(&device_offload_lock);
1257 	netdev = tls_ctx->netdev;
1258 	if (!netdev)
1259 		goto out;
1260 
1261 	netdev->tlsdev_ops->tls_dev_del(netdev, tls_ctx,
1262 					TLS_OFFLOAD_CTX_DIR_RX);
1263 
1264 	if (tls_ctx->tx_conf != TLS_HW) {
1265 		dev_put(netdev);
1266 		tls_ctx->netdev = NULL;
1267 	} else {
1268 		set_bit(TLS_RX_DEV_CLOSED, &tls_ctx->flags);
1269 	}
1270 out:
1271 	up_read(&device_offload_lock);
1272 	tls_sw_release_resources_rx(sk);
1273 }
1274 
1275 static int tls_device_down(struct net_device *netdev)
1276 {
1277 	struct tls_context *ctx, *tmp;
1278 	unsigned long flags;
1279 	LIST_HEAD(list);
1280 
1281 	/* Request a write lock to block new offload attempts */
1282 	down_write(&device_offload_lock);
1283 
1284 	spin_lock_irqsave(&tls_device_lock, flags);
1285 	list_for_each_entry_safe(ctx, tmp, &tls_device_list, list) {
1286 		if (ctx->netdev != netdev ||
1287 		    !refcount_inc_not_zero(&ctx->refcount))
1288 			continue;
1289 
1290 		list_move(&ctx->list, &list);
1291 	}
1292 	spin_unlock_irqrestore(&tls_device_lock, flags);
1293 
1294 	list_for_each_entry_safe(ctx, tmp, &list, list)	{
1295 		if (ctx->tx_conf == TLS_HW)
1296 			netdev->tlsdev_ops->tls_dev_del(netdev, ctx,
1297 							TLS_OFFLOAD_CTX_DIR_TX);
1298 		if (ctx->rx_conf == TLS_HW &&
1299 		    !test_bit(TLS_RX_DEV_CLOSED, &ctx->flags))
1300 			netdev->tlsdev_ops->tls_dev_del(netdev, ctx,
1301 							TLS_OFFLOAD_CTX_DIR_RX);
1302 		WRITE_ONCE(ctx->netdev, NULL);
1303 		smp_mb__before_atomic(); /* pairs with test_and_set_bit() */
1304 		while (test_bit(TLS_RX_SYNC_RUNNING, &ctx->flags))
1305 			usleep_range(10, 200);
1306 		dev_put(netdev);
1307 		list_del_init(&ctx->list);
1308 
1309 		if (refcount_dec_and_test(&ctx->refcount))
1310 			tls_device_free_ctx(ctx);
1311 	}
1312 
1313 	up_write(&device_offload_lock);
1314 
1315 	flush_work(&tls_device_gc_work);
1316 
1317 	return NOTIFY_DONE;
1318 }
1319 
1320 static int tls_dev_event(struct notifier_block *this, unsigned long event,
1321 			 void *ptr)
1322 {
1323 	struct net_device *dev = netdev_notifier_info_to_dev(ptr);
1324 
1325 	if (!dev->tlsdev_ops &&
1326 	    !(dev->features & (NETIF_F_HW_TLS_RX | NETIF_F_HW_TLS_TX)))
1327 		return NOTIFY_DONE;
1328 
1329 	switch (event) {
1330 	case NETDEV_REGISTER:
1331 	case NETDEV_FEAT_CHANGE:
1332 		if (netif_is_bond_master(dev))
1333 			return NOTIFY_DONE;
1334 		if ((dev->features & NETIF_F_HW_TLS_RX) &&
1335 		    !dev->tlsdev_ops->tls_dev_resync)
1336 			return NOTIFY_BAD;
1337 
1338 		if  (dev->tlsdev_ops &&
1339 		     dev->tlsdev_ops->tls_dev_add &&
1340 		     dev->tlsdev_ops->tls_dev_del)
1341 			return NOTIFY_DONE;
1342 		else
1343 			return NOTIFY_BAD;
1344 	case NETDEV_DOWN:
1345 		return tls_device_down(dev);
1346 	}
1347 	return NOTIFY_DONE;
1348 }
1349 
1350 static struct notifier_block tls_dev_notifier = {
1351 	.notifier_call	= tls_dev_event,
1352 };
1353 
1354 void __init tls_device_init(void)
1355 {
1356 	register_netdevice_notifier(&tls_dev_notifier);
1357 }
1358 
1359 void __exit tls_device_cleanup(void)
1360 {
1361 	unregister_netdevice_notifier(&tls_dev_notifier);
1362 	flush_work(&tls_device_gc_work);
1363 	clean_acked_data_flush();
1364 }
1365