xref: /linux/net/tls/tls_device.c (revision b7019ac550eb3916f34d79db583e9b7ea2524afa)
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 /* device_offload_lock is used to synchronize tls_dev_add
42  * against NETDEV_DOWN notifications.
43  */
44 static DECLARE_RWSEM(device_offload_lock);
45 
46 static void tls_device_gc_task(struct work_struct *work);
47 
48 static DECLARE_WORK(tls_device_gc_work, tls_device_gc_task);
49 static LIST_HEAD(tls_device_gc_list);
50 static LIST_HEAD(tls_device_list);
51 static DEFINE_SPINLOCK(tls_device_lock);
52 
53 static void tls_device_free_ctx(struct tls_context *ctx)
54 {
55 	if (ctx->tx_conf == TLS_HW) {
56 		kfree(tls_offload_ctx_tx(ctx));
57 		kfree(ctx->tx.rec_seq);
58 		kfree(ctx->tx.iv);
59 	}
60 
61 	if (ctx->rx_conf == TLS_HW)
62 		kfree(tls_offload_ctx_rx(ctx));
63 
64 	kfree(ctx);
65 }
66 
67 static void tls_device_gc_task(struct work_struct *work)
68 {
69 	struct tls_context *ctx, *tmp;
70 	unsigned long flags;
71 	LIST_HEAD(gc_list);
72 
73 	spin_lock_irqsave(&tls_device_lock, flags);
74 	list_splice_init(&tls_device_gc_list, &gc_list);
75 	spin_unlock_irqrestore(&tls_device_lock, flags);
76 
77 	list_for_each_entry_safe(ctx, tmp, &gc_list, list) {
78 		struct net_device *netdev = ctx->netdev;
79 
80 		if (netdev && ctx->tx_conf == TLS_HW) {
81 			netdev->tlsdev_ops->tls_dev_del(netdev, ctx,
82 							TLS_OFFLOAD_CTX_DIR_TX);
83 			dev_put(netdev);
84 			ctx->netdev = NULL;
85 		}
86 
87 		list_del(&ctx->list);
88 		tls_device_free_ctx(ctx);
89 	}
90 }
91 
92 static void tls_device_queue_ctx_destruction(struct tls_context *ctx)
93 {
94 	unsigned long flags;
95 
96 	spin_lock_irqsave(&tls_device_lock, flags);
97 	list_move_tail(&ctx->list, &tls_device_gc_list);
98 
99 	/* schedule_work inside the spinlock
100 	 * to make sure tls_device_down waits for that work.
101 	 */
102 	schedule_work(&tls_device_gc_work);
103 
104 	spin_unlock_irqrestore(&tls_device_lock, flags);
105 }
106 
107 /* We assume that the socket is already connected */
108 static struct net_device *get_netdev_for_sock(struct sock *sk)
109 {
110 	struct dst_entry *dst = sk_dst_get(sk);
111 	struct net_device *netdev = NULL;
112 
113 	if (likely(dst)) {
114 		netdev = dst->dev;
115 		dev_hold(netdev);
116 	}
117 
118 	dst_release(dst);
119 
120 	return netdev;
121 }
122 
123 static void destroy_record(struct tls_record_info *record)
124 {
125 	int nr_frags = record->num_frags;
126 	skb_frag_t *frag;
127 
128 	while (nr_frags-- > 0) {
129 		frag = &record->frags[nr_frags];
130 		__skb_frag_unref(frag);
131 	}
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 		list_del(&info->list);
165 		destroy_record(info);
166 		deleted_records++;
167 	}
168 
169 	list_for_each_entry_safe(info, temp, &ctx->records_list, list) {
170 		if (before(acked_seq, info->end_seq))
171 			break;
172 		list_del(&info->list);
173 
174 		destroy_record(info);
175 		deleted_records++;
176 	}
177 
178 	ctx->unacked_record_sn += deleted_records;
179 	spin_unlock_irqrestore(&ctx->lock, flags);
180 }
181 
182 /* At this point, there should be no references on this
183  * socket and no in-flight SKBs associated with this
184  * socket, so it is safe to free all the resources.
185  */
186 static void tls_device_sk_destruct(struct sock *sk)
187 {
188 	struct tls_context *tls_ctx = tls_get_ctx(sk);
189 	struct tls_offload_context_tx *ctx = tls_offload_ctx_tx(tls_ctx);
190 
191 	tls_ctx->sk_destruct(sk);
192 
193 	if (tls_ctx->tx_conf == TLS_HW) {
194 		if (ctx->open_record)
195 			destroy_record(ctx->open_record);
196 		delete_all_records(ctx);
197 		crypto_free_aead(ctx->aead_send);
198 		clean_acked_data_disable(inet_csk(sk));
199 	}
200 
201 	if (refcount_dec_and_test(&tls_ctx->refcount))
202 		tls_device_queue_ctx_destruction(tls_ctx);
203 }
204 
205 void tls_device_free_resources_tx(struct sock *sk)
206 {
207 	struct tls_context *tls_ctx = tls_get_ctx(sk);
208 
209 	tls_free_partial_record(sk, tls_ctx);
210 }
211 
212 static void tls_append_frag(struct tls_record_info *record,
213 			    struct page_frag *pfrag,
214 			    int size)
215 {
216 	skb_frag_t *frag;
217 
218 	frag = &record->frags[record->num_frags - 1];
219 	if (frag->page.p == pfrag->page &&
220 	    frag->page_offset + frag->size == pfrag->offset) {
221 		frag->size += size;
222 	} else {
223 		++frag;
224 		frag->page.p = pfrag->page;
225 		frag->page_offset = pfrag->offset;
226 		frag->size = size;
227 		++record->num_frags;
228 		get_page(pfrag->page);
229 	}
230 
231 	pfrag->offset += size;
232 	record->len += size;
233 }
234 
235 static int tls_push_record(struct sock *sk,
236 			   struct tls_context *ctx,
237 			   struct tls_offload_context_tx *offload_ctx,
238 			   struct tls_record_info *record,
239 			   struct page_frag *pfrag,
240 			   int flags,
241 			   unsigned char record_type)
242 {
243 	struct tls_prot_info *prot = &ctx->prot_info;
244 	struct tcp_sock *tp = tcp_sk(sk);
245 	struct page_frag dummy_tag_frag;
246 	skb_frag_t *frag;
247 	int i;
248 
249 	/* fill prepend */
250 	frag = &record->frags[0];
251 	tls_fill_prepend(ctx,
252 			 skb_frag_address(frag),
253 			 record->len - prot->prepend_size,
254 			 record_type,
255 			 ctx->crypto_send.info.version);
256 
257 	/* HW doesn't care about the data in the tag, because it fills it. */
258 	dummy_tag_frag.page = skb_frag_page(frag);
259 	dummy_tag_frag.offset = 0;
260 
261 	tls_append_frag(record, &dummy_tag_frag, prot->tag_size);
262 	record->end_seq = tp->write_seq + record->len;
263 	spin_lock_irq(&offload_ctx->lock);
264 	list_add_tail(&record->list, &offload_ctx->records_list);
265 	spin_unlock_irq(&offload_ctx->lock);
266 	offload_ctx->open_record = NULL;
267 	tls_advance_record_sn(sk, &ctx->tx, ctx->crypto_send.info.version);
268 
269 	for (i = 0; i < record->num_frags; i++) {
270 		frag = &record->frags[i];
271 		sg_unmark_end(&offload_ctx->sg_tx_data[i]);
272 		sg_set_page(&offload_ctx->sg_tx_data[i], skb_frag_page(frag),
273 			    frag->size, frag->page_offset);
274 		sk_mem_charge(sk, frag->size);
275 		get_page(skb_frag_page(frag));
276 	}
277 	sg_mark_end(&offload_ctx->sg_tx_data[record->num_frags - 1]);
278 
279 	/* all ready, send */
280 	return tls_push_sg(sk, ctx, offload_ctx->sg_tx_data, 0, flags);
281 }
282 
283 static int tls_create_new_record(struct tls_offload_context_tx *offload_ctx,
284 				 struct page_frag *pfrag,
285 				 size_t prepend_size)
286 {
287 	struct tls_record_info *record;
288 	skb_frag_t *frag;
289 
290 	record = kmalloc(sizeof(*record), GFP_KERNEL);
291 	if (!record)
292 		return -ENOMEM;
293 
294 	frag = &record->frags[0];
295 	__skb_frag_set_page(frag, pfrag->page);
296 	frag->page_offset = pfrag->offset;
297 	skb_frag_size_set(frag, prepend_size);
298 
299 	get_page(pfrag->page);
300 	pfrag->offset += prepend_size;
301 
302 	record->num_frags = 1;
303 	record->len = prepend_size;
304 	offload_ctx->open_record = record;
305 	return 0;
306 }
307 
308 static int tls_do_allocation(struct sock *sk,
309 			     struct tls_offload_context_tx *offload_ctx,
310 			     struct page_frag *pfrag,
311 			     size_t prepend_size)
312 {
313 	int ret;
314 
315 	if (!offload_ctx->open_record) {
316 		if (unlikely(!skb_page_frag_refill(prepend_size, pfrag,
317 						   sk->sk_allocation))) {
318 			sk->sk_prot->enter_memory_pressure(sk);
319 			sk_stream_moderate_sndbuf(sk);
320 			return -ENOMEM;
321 		}
322 
323 		ret = tls_create_new_record(offload_ctx, pfrag, prepend_size);
324 		if (ret)
325 			return ret;
326 
327 		if (pfrag->size > pfrag->offset)
328 			return 0;
329 	}
330 
331 	if (!sk_page_frag_refill(sk, pfrag))
332 		return -ENOMEM;
333 
334 	return 0;
335 }
336 
337 static int tls_push_data(struct sock *sk,
338 			 struct iov_iter *msg_iter,
339 			 size_t size, int flags,
340 			 unsigned char record_type)
341 {
342 	struct tls_context *tls_ctx = tls_get_ctx(sk);
343 	struct tls_prot_info *prot = &tls_ctx->prot_info;
344 	struct tls_offload_context_tx *ctx = tls_offload_ctx_tx(tls_ctx);
345 	int tls_push_record_flags = flags | MSG_SENDPAGE_NOTLAST;
346 	int more = flags & (MSG_SENDPAGE_NOTLAST | MSG_MORE);
347 	struct tls_record_info *record = ctx->open_record;
348 	struct page_frag *pfrag;
349 	size_t orig_size = size;
350 	u32 max_open_record_len;
351 	int copy, rc = 0;
352 	bool done = false;
353 	long timeo;
354 
355 	if (flags &
356 	    ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL | MSG_SENDPAGE_NOTLAST))
357 		return -ENOTSUPP;
358 
359 	if (sk->sk_err)
360 		return -sk->sk_err;
361 
362 	timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
363 	if (tls_is_partially_sent_record(tls_ctx)) {
364 		rc = tls_push_partial_record(sk, tls_ctx, flags);
365 		if (rc < 0)
366 			return rc;
367 	}
368 
369 	pfrag = sk_page_frag(sk);
370 
371 	/* TLS_HEADER_SIZE is not counted as part of the TLS record, and
372 	 * we need to leave room for an authentication tag.
373 	 */
374 	max_open_record_len = TLS_MAX_PAYLOAD_SIZE +
375 			      prot->prepend_size;
376 	do {
377 		rc = tls_do_allocation(sk, ctx, pfrag,
378 				       prot->prepend_size);
379 		if (rc) {
380 			rc = sk_stream_wait_memory(sk, &timeo);
381 			if (!rc)
382 				continue;
383 
384 			record = ctx->open_record;
385 			if (!record)
386 				break;
387 handle_error:
388 			if (record_type != TLS_RECORD_TYPE_DATA) {
389 				/* avoid sending partial
390 				 * record with type !=
391 				 * application_data
392 				 */
393 				size = orig_size;
394 				destroy_record(record);
395 				ctx->open_record = NULL;
396 			} else if (record->len > prot->prepend_size) {
397 				goto last_record;
398 			}
399 
400 			break;
401 		}
402 
403 		record = ctx->open_record;
404 		copy = min_t(size_t, size, (pfrag->size - pfrag->offset));
405 		copy = min_t(size_t, copy, (max_open_record_len - record->len));
406 
407 		if (copy_from_iter_nocache(page_address(pfrag->page) +
408 					       pfrag->offset,
409 					   copy, msg_iter) != copy) {
410 			rc = -EFAULT;
411 			goto handle_error;
412 		}
413 		tls_append_frag(record, pfrag, copy);
414 
415 		size -= copy;
416 		if (!size) {
417 last_record:
418 			tls_push_record_flags = flags;
419 			if (more) {
420 				tls_ctx->pending_open_record_frags =
421 						!!record->num_frags;
422 				break;
423 			}
424 
425 			done = true;
426 		}
427 
428 		if (done || record->len >= max_open_record_len ||
429 		    (record->num_frags >= MAX_SKB_FRAGS - 1)) {
430 			rc = tls_push_record(sk,
431 					     tls_ctx,
432 					     ctx,
433 					     record,
434 					     pfrag,
435 					     tls_push_record_flags,
436 					     record_type);
437 			if (rc < 0)
438 				break;
439 		}
440 	} while (!done);
441 
442 	if (orig_size - size > 0)
443 		rc = orig_size - size;
444 
445 	return rc;
446 }
447 
448 int tls_device_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
449 {
450 	unsigned char record_type = TLS_RECORD_TYPE_DATA;
451 	int rc;
452 
453 	lock_sock(sk);
454 
455 	if (unlikely(msg->msg_controllen)) {
456 		rc = tls_proccess_cmsg(sk, msg, &record_type);
457 		if (rc)
458 			goto out;
459 	}
460 
461 	rc = tls_push_data(sk, &msg->msg_iter, size,
462 			   msg->msg_flags, record_type);
463 
464 out:
465 	release_sock(sk);
466 	return rc;
467 }
468 
469 int tls_device_sendpage(struct sock *sk, struct page *page,
470 			int offset, size_t size, int flags)
471 {
472 	struct iov_iter	msg_iter;
473 	char *kaddr = kmap(page);
474 	struct kvec iov;
475 	int rc;
476 
477 	if (flags & MSG_SENDPAGE_NOTLAST)
478 		flags |= MSG_MORE;
479 
480 	lock_sock(sk);
481 
482 	if (flags & MSG_OOB) {
483 		rc = -ENOTSUPP;
484 		goto out;
485 	}
486 
487 	iov.iov_base = kaddr + offset;
488 	iov.iov_len = size;
489 	iov_iter_kvec(&msg_iter, WRITE, &iov, 1, size);
490 	rc = tls_push_data(sk, &msg_iter, size,
491 			   flags, TLS_RECORD_TYPE_DATA);
492 	kunmap(page);
493 
494 out:
495 	release_sock(sk);
496 	return rc;
497 }
498 
499 struct tls_record_info *tls_get_record(struct tls_offload_context_tx *context,
500 				       u32 seq, u64 *p_record_sn)
501 {
502 	u64 record_sn = context->hint_record_sn;
503 	struct tls_record_info *info;
504 
505 	info = context->retransmit_hint;
506 	if (!info ||
507 	    before(seq, info->end_seq - info->len)) {
508 		/* if retransmit_hint is irrelevant start
509 		 * from the beggining of the list
510 		 */
511 		info = list_first_entry(&context->records_list,
512 					struct tls_record_info, list);
513 		record_sn = context->unacked_record_sn;
514 	}
515 
516 	list_for_each_entry_from(info, &context->records_list, list) {
517 		if (before(seq, info->end_seq)) {
518 			if (!context->retransmit_hint ||
519 			    after(info->end_seq,
520 				  context->retransmit_hint->end_seq)) {
521 				context->hint_record_sn = record_sn;
522 				context->retransmit_hint = info;
523 			}
524 			*p_record_sn = record_sn;
525 			return info;
526 		}
527 		record_sn++;
528 	}
529 
530 	return NULL;
531 }
532 EXPORT_SYMBOL(tls_get_record);
533 
534 static int tls_device_push_pending_record(struct sock *sk, int flags)
535 {
536 	struct iov_iter	msg_iter;
537 
538 	iov_iter_kvec(&msg_iter, WRITE, NULL, 0, 0);
539 	return tls_push_data(sk, &msg_iter, 0, flags, TLS_RECORD_TYPE_DATA);
540 }
541 
542 void tls_device_write_space(struct sock *sk, struct tls_context *ctx)
543 {
544 	if (!sk->sk_write_pending && tls_is_partially_sent_record(ctx)) {
545 		gfp_t sk_allocation = sk->sk_allocation;
546 
547 		sk->sk_allocation = GFP_ATOMIC;
548 		tls_push_partial_record(sk, ctx, MSG_DONTWAIT | MSG_NOSIGNAL);
549 		sk->sk_allocation = sk_allocation;
550 	}
551 }
552 
553 static void tls_device_resync_rx(struct tls_context *tls_ctx,
554 				 struct sock *sk, u32 seq, u64 rcd_sn)
555 {
556 	struct net_device *netdev;
557 
558 	if (WARN_ON(test_and_set_bit(TLS_RX_SYNC_RUNNING, &tls_ctx->flags)))
559 		return;
560 	netdev = READ_ONCE(tls_ctx->netdev);
561 	if (netdev)
562 		netdev->tlsdev_ops->tls_dev_resync_rx(netdev, sk, seq, rcd_sn);
563 	clear_bit_unlock(TLS_RX_SYNC_RUNNING, &tls_ctx->flags);
564 }
565 
566 void handle_device_resync(struct sock *sk, u32 seq, u64 rcd_sn)
567 {
568 	struct tls_context *tls_ctx = tls_get_ctx(sk);
569 	struct tls_offload_context_rx *rx_ctx;
570 	u32 is_req_pending;
571 	s64 resync_req;
572 	u32 req_seq;
573 
574 	if (tls_ctx->rx_conf != TLS_HW)
575 		return;
576 
577 	rx_ctx = tls_offload_ctx_rx(tls_ctx);
578 	resync_req = atomic64_read(&rx_ctx->resync_req);
579 	req_seq = (resync_req >> 32) - ((u32)TLS_HEADER_SIZE - 1);
580 	is_req_pending = resync_req;
581 
582 	if (unlikely(is_req_pending) && req_seq == seq &&
583 	    atomic64_try_cmpxchg(&rx_ctx->resync_req, &resync_req, 0)) {
584 		seq += TLS_HEADER_SIZE - 1;
585 		tls_device_resync_rx(tls_ctx, sk, seq, rcd_sn);
586 	}
587 }
588 
589 static int tls_device_reencrypt(struct sock *sk, struct sk_buff *skb)
590 {
591 	struct strp_msg *rxm = strp_msg(skb);
592 	int err = 0, offset = rxm->offset, copy, nsg, data_len, pos;
593 	struct sk_buff *skb_iter, *unused;
594 	struct scatterlist sg[1];
595 	char *orig_buf, *buf;
596 
597 	orig_buf = kmalloc(rxm->full_len + TLS_HEADER_SIZE +
598 			   TLS_CIPHER_AES_GCM_128_IV_SIZE, sk->sk_allocation);
599 	if (!orig_buf)
600 		return -ENOMEM;
601 	buf = orig_buf;
602 
603 	nsg = skb_cow_data(skb, 0, &unused);
604 	if (unlikely(nsg < 0)) {
605 		err = nsg;
606 		goto free_buf;
607 	}
608 
609 	sg_init_table(sg, 1);
610 	sg_set_buf(&sg[0], buf,
611 		   rxm->full_len + TLS_HEADER_SIZE +
612 		   TLS_CIPHER_AES_GCM_128_IV_SIZE);
613 	skb_copy_bits(skb, offset, buf,
614 		      TLS_HEADER_SIZE + TLS_CIPHER_AES_GCM_128_IV_SIZE);
615 
616 	/* We are interested only in the decrypted data not the auth */
617 	err = decrypt_skb(sk, skb, sg);
618 	if (err != -EBADMSG)
619 		goto free_buf;
620 	else
621 		err = 0;
622 
623 	data_len = rxm->full_len - TLS_CIPHER_AES_GCM_128_TAG_SIZE;
624 
625 	if (skb_pagelen(skb) > offset) {
626 		copy = min_t(int, skb_pagelen(skb) - offset, data_len);
627 
628 		if (skb->decrypted)
629 			skb_store_bits(skb, offset, buf, copy);
630 
631 		offset += copy;
632 		buf += copy;
633 	}
634 
635 	pos = skb_pagelen(skb);
636 	skb_walk_frags(skb, skb_iter) {
637 		int frag_pos;
638 
639 		/* Practically all frags must belong to msg if reencrypt
640 		 * is needed with current strparser and coalescing logic,
641 		 * but strparser may "get optimized", so let's be safe.
642 		 */
643 		if (pos + skb_iter->len <= offset)
644 			goto done_with_frag;
645 		if (pos >= data_len + rxm->offset)
646 			break;
647 
648 		frag_pos = offset - pos;
649 		copy = min_t(int, skb_iter->len - frag_pos,
650 			     data_len + rxm->offset - offset);
651 
652 		if (skb_iter->decrypted)
653 			skb_store_bits(skb_iter, frag_pos, buf, copy);
654 
655 		offset += copy;
656 		buf += copy;
657 done_with_frag:
658 		pos += skb_iter->len;
659 	}
660 
661 free_buf:
662 	kfree(orig_buf);
663 	return err;
664 }
665 
666 int tls_device_decrypted(struct sock *sk, struct sk_buff *skb)
667 {
668 	struct tls_context *tls_ctx = tls_get_ctx(sk);
669 	struct tls_offload_context_rx *ctx = tls_offload_ctx_rx(tls_ctx);
670 	int is_decrypted = skb->decrypted;
671 	int is_encrypted = !is_decrypted;
672 	struct sk_buff *skb_iter;
673 
674 	/* Skip if it is already decrypted */
675 	if (ctx->sw.decrypted)
676 		return 0;
677 
678 	/* Check if all the data is decrypted already */
679 	skb_walk_frags(skb, skb_iter) {
680 		is_decrypted &= skb_iter->decrypted;
681 		is_encrypted &= !skb_iter->decrypted;
682 	}
683 
684 	ctx->sw.decrypted |= is_decrypted;
685 
686 	/* Return immedeatly if the record is either entirely plaintext or
687 	 * entirely ciphertext. Otherwise handle reencrypt partially decrypted
688 	 * record.
689 	 */
690 	return (is_encrypted || is_decrypted) ? 0 :
691 		tls_device_reencrypt(sk, skb);
692 }
693 
694 static void tls_device_attach(struct tls_context *ctx, struct sock *sk,
695 			      struct net_device *netdev)
696 {
697 	if (sk->sk_destruct != tls_device_sk_destruct) {
698 		refcount_set(&ctx->refcount, 1);
699 		dev_hold(netdev);
700 		ctx->netdev = netdev;
701 		spin_lock_irq(&tls_device_lock);
702 		list_add_tail(&ctx->list, &tls_device_list);
703 		spin_unlock_irq(&tls_device_lock);
704 
705 		ctx->sk_destruct = sk->sk_destruct;
706 		sk->sk_destruct = tls_device_sk_destruct;
707 	}
708 }
709 
710 int tls_set_device_offload(struct sock *sk, struct tls_context *ctx)
711 {
712 	u16 nonce_size, tag_size, iv_size, rec_seq_size;
713 	struct tls_context *tls_ctx = tls_get_ctx(sk);
714 	struct tls_prot_info *prot = &tls_ctx->prot_info;
715 	struct tls_record_info *start_marker_record;
716 	struct tls_offload_context_tx *offload_ctx;
717 	struct tls_crypto_info *crypto_info;
718 	struct net_device *netdev;
719 	char *iv, *rec_seq;
720 	struct sk_buff *skb;
721 	int rc = -EINVAL;
722 	__be64 rcd_sn;
723 
724 	if (!ctx)
725 		goto out;
726 
727 	if (ctx->priv_ctx_tx) {
728 		rc = -EEXIST;
729 		goto out;
730 	}
731 
732 	start_marker_record = kmalloc(sizeof(*start_marker_record), GFP_KERNEL);
733 	if (!start_marker_record) {
734 		rc = -ENOMEM;
735 		goto out;
736 	}
737 
738 	offload_ctx = kzalloc(TLS_OFFLOAD_CONTEXT_SIZE_TX, GFP_KERNEL);
739 	if (!offload_ctx) {
740 		rc = -ENOMEM;
741 		goto free_marker_record;
742 	}
743 
744 	crypto_info = &ctx->crypto_send.info;
745 	switch (crypto_info->cipher_type) {
746 	case TLS_CIPHER_AES_GCM_128:
747 		nonce_size = TLS_CIPHER_AES_GCM_128_IV_SIZE;
748 		tag_size = TLS_CIPHER_AES_GCM_128_TAG_SIZE;
749 		iv_size = TLS_CIPHER_AES_GCM_128_IV_SIZE;
750 		iv = ((struct tls12_crypto_info_aes_gcm_128 *)crypto_info)->iv;
751 		rec_seq_size = TLS_CIPHER_AES_GCM_128_REC_SEQ_SIZE;
752 		rec_seq =
753 		 ((struct tls12_crypto_info_aes_gcm_128 *)crypto_info)->rec_seq;
754 		break;
755 	default:
756 		rc = -EINVAL;
757 		goto free_offload_ctx;
758 	}
759 
760 	prot->prepend_size = TLS_HEADER_SIZE + nonce_size;
761 	prot->tag_size = tag_size;
762 	prot->overhead_size = prot->prepend_size + prot->tag_size;
763 	prot->iv_size = iv_size;
764 	ctx->tx.iv = kmalloc(iv_size + TLS_CIPHER_AES_GCM_128_SALT_SIZE,
765 			     GFP_KERNEL);
766 	if (!ctx->tx.iv) {
767 		rc = -ENOMEM;
768 		goto free_offload_ctx;
769 	}
770 
771 	memcpy(ctx->tx.iv + TLS_CIPHER_AES_GCM_128_SALT_SIZE, iv, iv_size);
772 
773 	prot->rec_seq_size = rec_seq_size;
774 	ctx->tx.rec_seq = kmemdup(rec_seq, rec_seq_size, GFP_KERNEL);
775 	if (!ctx->tx.rec_seq) {
776 		rc = -ENOMEM;
777 		goto free_iv;
778 	}
779 
780 	rc = tls_sw_fallback_init(sk, offload_ctx, crypto_info);
781 	if (rc)
782 		goto free_rec_seq;
783 
784 	/* start at rec_seq - 1 to account for the start marker record */
785 	memcpy(&rcd_sn, ctx->tx.rec_seq, sizeof(rcd_sn));
786 	offload_ctx->unacked_record_sn = be64_to_cpu(rcd_sn) - 1;
787 
788 	start_marker_record->end_seq = tcp_sk(sk)->write_seq;
789 	start_marker_record->len = 0;
790 	start_marker_record->num_frags = 0;
791 
792 	INIT_LIST_HEAD(&offload_ctx->records_list);
793 	list_add_tail(&start_marker_record->list, &offload_ctx->records_list);
794 	spin_lock_init(&offload_ctx->lock);
795 	sg_init_table(offload_ctx->sg_tx_data,
796 		      ARRAY_SIZE(offload_ctx->sg_tx_data));
797 
798 	clean_acked_data_enable(inet_csk(sk), &tls_icsk_clean_acked);
799 	ctx->push_pending_record = tls_device_push_pending_record;
800 
801 	/* TLS offload is greatly simplified if we don't send
802 	 * SKBs where only part of the payload needs to be encrypted.
803 	 * So mark the last skb in the write queue as end of record.
804 	 */
805 	skb = tcp_write_queue_tail(sk);
806 	if (skb)
807 		TCP_SKB_CB(skb)->eor = 1;
808 
809 	/* We support starting offload on multiple sockets
810 	 * concurrently, so we only need a read lock here.
811 	 * This lock must precede get_netdev_for_sock to prevent races between
812 	 * NETDEV_DOWN and setsockopt.
813 	 */
814 	down_read(&device_offload_lock);
815 	netdev = get_netdev_for_sock(sk);
816 	if (!netdev) {
817 		pr_err_ratelimited("%s: netdev not found\n", __func__);
818 		rc = -EINVAL;
819 		goto release_lock;
820 	}
821 
822 	if (!(netdev->features & NETIF_F_HW_TLS_TX)) {
823 		rc = -ENOTSUPP;
824 		goto release_netdev;
825 	}
826 
827 	/* Avoid offloading if the device is down
828 	 * We don't want to offload new flows after
829 	 * the NETDEV_DOWN event
830 	 */
831 	if (!(netdev->flags & IFF_UP)) {
832 		rc = -EINVAL;
833 		goto release_netdev;
834 	}
835 
836 	ctx->priv_ctx_tx = offload_ctx;
837 	rc = netdev->tlsdev_ops->tls_dev_add(netdev, sk, TLS_OFFLOAD_CTX_DIR_TX,
838 					     &ctx->crypto_send.info,
839 					     tcp_sk(sk)->write_seq);
840 	if (rc)
841 		goto release_netdev;
842 
843 	tls_device_attach(ctx, sk, netdev);
844 
845 	/* following this assignment tls_is_sk_tx_device_offloaded
846 	 * will return true and the context might be accessed
847 	 * by the netdev's xmit function.
848 	 */
849 	smp_store_release(&sk->sk_validate_xmit_skb, tls_validate_xmit_skb);
850 	dev_put(netdev);
851 	up_read(&device_offload_lock);
852 	goto out;
853 
854 release_netdev:
855 	dev_put(netdev);
856 release_lock:
857 	up_read(&device_offload_lock);
858 	clean_acked_data_disable(inet_csk(sk));
859 	crypto_free_aead(offload_ctx->aead_send);
860 free_rec_seq:
861 	kfree(ctx->tx.rec_seq);
862 free_iv:
863 	kfree(ctx->tx.iv);
864 free_offload_ctx:
865 	kfree(offload_ctx);
866 	ctx->priv_ctx_tx = NULL;
867 free_marker_record:
868 	kfree(start_marker_record);
869 out:
870 	return rc;
871 }
872 
873 int tls_set_device_offload_rx(struct sock *sk, struct tls_context *ctx)
874 {
875 	struct tls_offload_context_rx *context;
876 	struct net_device *netdev;
877 	int rc = 0;
878 
879 	/* We support starting offload on multiple sockets
880 	 * concurrently, so we only need a read lock here.
881 	 * This lock must precede get_netdev_for_sock to prevent races between
882 	 * NETDEV_DOWN and setsockopt.
883 	 */
884 	down_read(&device_offload_lock);
885 	netdev = get_netdev_for_sock(sk);
886 	if (!netdev) {
887 		pr_err_ratelimited("%s: netdev not found\n", __func__);
888 		rc = -EINVAL;
889 		goto release_lock;
890 	}
891 
892 	if (!(netdev->features & NETIF_F_HW_TLS_RX)) {
893 		rc = -ENOTSUPP;
894 		goto release_netdev;
895 	}
896 
897 	/* Avoid offloading if the device is down
898 	 * We don't want to offload new flows after
899 	 * the NETDEV_DOWN event
900 	 */
901 	if (!(netdev->flags & IFF_UP)) {
902 		rc = -EINVAL;
903 		goto release_netdev;
904 	}
905 
906 	context = kzalloc(TLS_OFFLOAD_CONTEXT_SIZE_RX, GFP_KERNEL);
907 	if (!context) {
908 		rc = -ENOMEM;
909 		goto release_netdev;
910 	}
911 
912 	ctx->priv_ctx_rx = context;
913 	rc = tls_set_sw_offload(sk, ctx, 0);
914 	if (rc)
915 		goto release_ctx;
916 
917 	rc = netdev->tlsdev_ops->tls_dev_add(netdev, sk, TLS_OFFLOAD_CTX_DIR_RX,
918 					     &ctx->crypto_recv.info,
919 					     tcp_sk(sk)->copied_seq);
920 	if (rc)
921 		goto free_sw_resources;
922 
923 	tls_device_attach(ctx, sk, netdev);
924 	goto release_netdev;
925 
926 free_sw_resources:
927 	up_read(&device_offload_lock);
928 	tls_sw_free_resources_rx(sk);
929 	down_read(&device_offload_lock);
930 release_ctx:
931 	ctx->priv_ctx_rx = NULL;
932 release_netdev:
933 	dev_put(netdev);
934 release_lock:
935 	up_read(&device_offload_lock);
936 	return rc;
937 }
938 
939 void tls_device_offload_cleanup_rx(struct sock *sk)
940 {
941 	struct tls_context *tls_ctx = tls_get_ctx(sk);
942 	struct net_device *netdev;
943 
944 	down_read(&device_offload_lock);
945 	netdev = tls_ctx->netdev;
946 	if (!netdev)
947 		goto out;
948 
949 	netdev->tlsdev_ops->tls_dev_del(netdev, tls_ctx,
950 					TLS_OFFLOAD_CTX_DIR_RX);
951 
952 	if (tls_ctx->tx_conf != TLS_HW) {
953 		dev_put(netdev);
954 		tls_ctx->netdev = NULL;
955 	}
956 out:
957 	up_read(&device_offload_lock);
958 	tls_sw_release_resources_rx(sk);
959 }
960 
961 static int tls_device_down(struct net_device *netdev)
962 {
963 	struct tls_context *ctx, *tmp;
964 	unsigned long flags;
965 	LIST_HEAD(list);
966 
967 	/* Request a write lock to block new offload attempts */
968 	down_write(&device_offload_lock);
969 
970 	spin_lock_irqsave(&tls_device_lock, flags);
971 	list_for_each_entry_safe(ctx, tmp, &tls_device_list, list) {
972 		if (ctx->netdev != netdev ||
973 		    !refcount_inc_not_zero(&ctx->refcount))
974 			continue;
975 
976 		list_move(&ctx->list, &list);
977 	}
978 	spin_unlock_irqrestore(&tls_device_lock, flags);
979 
980 	list_for_each_entry_safe(ctx, tmp, &list, list)	{
981 		if (ctx->tx_conf == TLS_HW)
982 			netdev->tlsdev_ops->tls_dev_del(netdev, ctx,
983 							TLS_OFFLOAD_CTX_DIR_TX);
984 		if (ctx->rx_conf == TLS_HW)
985 			netdev->tlsdev_ops->tls_dev_del(netdev, ctx,
986 							TLS_OFFLOAD_CTX_DIR_RX);
987 		WRITE_ONCE(ctx->netdev, NULL);
988 		smp_mb__before_atomic(); /* pairs with test_and_set_bit() */
989 		while (test_bit(TLS_RX_SYNC_RUNNING, &ctx->flags))
990 			usleep_range(10, 200);
991 		dev_put(netdev);
992 		list_del_init(&ctx->list);
993 
994 		if (refcount_dec_and_test(&ctx->refcount))
995 			tls_device_free_ctx(ctx);
996 	}
997 
998 	up_write(&device_offload_lock);
999 
1000 	flush_work(&tls_device_gc_work);
1001 
1002 	return NOTIFY_DONE;
1003 }
1004 
1005 static int tls_dev_event(struct notifier_block *this, unsigned long event,
1006 			 void *ptr)
1007 {
1008 	struct net_device *dev = netdev_notifier_info_to_dev(ptr);
1009 
1010 	if (!dev->tlsdev_ops &&
1011 	    !(dev->features & (NETIF_F_HW_TLS_RX | NETIF_F_HW_TLS_TX)))
1012 		return NOTIFY_DONE;
1013 
1014 	switch (event) {
1015 	case NETDEV_REGISTER:
1016 	case NETDEV_FEAT_CHANGE:
1017 		if ((dev->features & NETIF_F_HW_TLS_RX) &&
1018 		    !dev->tlsdev_ops->tls_dev_resync_rx)
1019 			return NOTIFY_BAD;
1020 
1021 		if  (dev->tlsdev_ops &&
1022 		     dev->tlsdev_ops->tls_dev_add &&
1023 		     dev->tlsdev_ops->tls_dev_del)
1024 			return NOTIFY_DONE;
1025 		else
1026 			return NOTIFY_BAD;
1027 	case NETDEV_DOWN:
1028 		return tls_device_down(dev);
1029 	}
1030 	return NOTIFY_DONE;
1031 }
1032 
1033 static struct notifier_block tls_dev_notifier = {
1034 	.notifier_call	= tls_dev_event,
1035 };
1036 
1037 void __init tls_device_init(void)
1038 {
1039 	register_netdevice_notifier(&tls_dev_notifier);
1040 }
1041 
1042 void __exit tls_device_cleanup(void)
1043 {
1044 	unregister_netdevice_notifier(&tls_dev_notifier);
1045 	flush_work(&tls_device_gc_work);
1046 	clean_acked_data_flush();
1047 }
1048