xref: /openbmc/linux/net/tls/tls_device.c (revision 86db9f28)
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 	tls_ctx_free(NULL, 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 i;
126 
127 	for (i = 0; i < record->num_frags; i++)
128 		__skb_frag_unref(&record->frags[i]);
129 	kfree(record);
130 }
131 
132 static void delete_all_records(struct tls_offload_context_tx *offload_ctx)
133 {
134 	struct tls_record_info *info, *temp;
135 
136 	list_for_each_entry_safe(info, temp, &offload_ctx->records_list, list) {
137 		list_del(&info->list);
138 		destroy_record(info);
139 	}
140 
141 	offload_ctx->retransmit_hint = NULL;
142 }
143 
144 static void tls_icsk_clean_acked(struct sock *sk, u32 acked_seq)
145 {
146 	struct tls_context *tls_ctx = tls_get_ctx(sk);
147 	struct tls_record_info *info, *temp;
148 	struct tls_offload_context_tx *ctx;
149 	u64 deleted_records = 0;
150 	unsigned long flags;
151 
152 	if (!tls_ctx)
153 		return;
154 
155 	ctx = tls_offload_ctx_tx(tls_ctx);
156 
157 	spin_lock_irqsave(&ctx->lock, flags);
158 	info = ctx->retransmit_hint;
159 	if (info && !before(acked_seq, info->end_seq))
160 		ctx->retransmit_hint = NULL;
161 
162 	list_for_each_entry_safe(info, temp, &ctx->records_list, list) {
163 		if (before(acked_seq, info->end_seq))
164 			break;
165 		list_del(&info->list);
166 
167 		destroy_record(info);
168 		deleted_records++;
169 	}
170 
171 	ctx->unacked_record_sn += deleted_records;
172 	spin_unlock_irqrestore(&ctx->lock, flags);
173 }
174 
175 /* At this point, there should be no references on this
176  * socket and no in-flight SKBs associated with this
177  * socket, so it is safe to free all the resources.
178  */
179 static void tls_device_sk_destruct(struct sock *sk)
180 {
181 	struct tls_context *tls_ctx = tls_get_ctx(sk);
182 	struct tls_offload_context_tx *ctx = tls_offload_ctx_tx(tls_ctx);
183 
184 	tls_ctx->sk_destruct(sk);
185 
186 	if (tls_ctx->tx_conf == TLS_HW) {
187 		if (ctx->open_record)
188 			destroy_record(ctx->open_record);
189 		delete_all_records(ctx);
190 		crypto_free_aead(ctx->aead_send);
191 		clean_acked_data_disable(inet_csk(sk));
192 	}
193 
194 	if (refcount_dec_and_test(&tls_ctx->refcount))
195 		tls_device_queue_ctx_destruction(tls_ctx);
196 }
197 
198 void tls_device_free_resources_tx(struct sock *sk)
199 {
200 	struct tls_context *tls_ctx = tls_get_ctx(sk);
201 
202 	tls_free_partial_record(sk, tls_ctx);
203 }
204 
205 static void tls_device_resync_tx(struct sock *sk, struct tls_context *tls_ctx,
206 				 u32 seq)
207 {
208 	struct net_device *netdev;
209 	struct sk_buff *skb;
210 	int err = 0;
211 	u8 *rcd_sn;
212 
213 	skb = tcp_write_queue_tail(sk);
214 	if (skb)
215 		TCP_SKB_CB(skb)->eor = 1;
216 
217 	rcd_sn = tls_ctx->tx.rec_seq;
218 
219 	down_read(&device_offload_lock);
220 	netdev = tls_ctx->netdev;
221 	if (netdev)
222 		err = netdev->tlsdev_ops->tls_dev_resync(netdev, sk, seq,
223 							 rcd_sn,
224 							 TLS_OFFLOAD_CTX_DIR_TX);
225 	up_read(&device_offload_lock);
226 	if (err)
227 		return;
228 
229 	clear_bit_unlock(TLS_TX_SYNC_SCHED, &tls_ctx->flags);
230 }
231 
232 static void tls_append_frag(struct tls_record_info *record,
233 			    struct page_frag *pfrag,
234 			    int size)
235 {
236 	skb_frag_t *frag;
237 
238 	frag = &record->frags[record->num_frags - 1];
239 	if (skb_frag_page(frag) == pfrag->page &&
240 	    skb_frag_off(frag) + skb_frag_size(frag) == pfrag->offset) {
241 		skb_frag_size_add(frag, size);
242 	} else {
243 		++frag;
244 		__skb_frag_set_page(frag, pfrag->page);
245 		skb_frag_off_set(frag, pfrag->offset);
246 		skb_frag_size_set(frag, size);
247 		++record->num_frags;
248 		get_page(pfrag->page);
249 	}
250 
251 	pfrag->offset += size;
252 	record->len += size;
253 }
254 
255 static int tls_push_record(struct sock *sk,
256 			   struct tls_context *ctx,
257 			   struct tls_offload_context_tx *offload_ctx,
258 			   struct tls_record_info *record,
259 			   int flags)
260 {
261 	struct tls_prot_info *prot = &ctx->prot_info;
262 	struct tcp_sock *tp = tcp_sk(sk);
263 	skb_frag_t *frag;
264 	int i;
265 
266 	record->end_seq = tp->write_seq + record->len;
267 	list_add_tail_rcu(&record->list, &offload_ctx->records_list);
268 	offload_ctx->open_record = NULL;
269 
270 	if (test_bit(TLS_TX_SYNC_SCHED, &ctx->flags))
271 		tls_device_resync_tx(sk, ctx, tp->write_seq);
272 
273 	tls_advance_record_sn(sk, prot, &ctx->tx);
274 
275 	for (i = 0; i < record->num_frags; i++) {
276 		frag = &record->frags[i];
277 		sg_unmark_end(&offload_ctx->sg_tx_data[i]);
278 		sg_set_page(&offload_ctx->sg_tx_data[i], skb_frag_page(frag),
279 			    skb_frag_size(frag), skb_frag_off(frag));
280 		sk_mem_charge(sk, skb_frag_size(frag));
281 		get_page(skb_frag_page(frag));
282 	}
283 	sg_mark_end(&offload_ctx->sg_tx_data[record->num_frags - 1]);
284 
285 	/* all ready, send */
286 	return tls_push_sg(sk, ctx, offload_ctx->sg_tx_data, 0, flags);
287 }
288 
289 static int tls_device_record_close(struct sock *sk,
290 				   struct tls_context *ctx,
291 				   struct tls_record_info *record,
292 				   struct page_frag *pfrag,
293 				   unsigned char record_type)
294 {
295 	struct tls_prot_info *prot = &ctx->prot_info;
296 	int ret;
297 
298 	/* append tag
299 	 * device will fill in the tag, we just need to append a placeholder
300 	 * use socket memory to improve coalescing (re-using a single buffer
301 	 * increases frag count)
302 	 * if we can't allocate memory now, steal some back from data
303 	 */
304 	if (likely(skb_page_frag_refill(prot->tag_size, pfrag,
305 					sk->sk_allocation))) {
306 		ret = 0;
307 		tls_append_frag(record, pfrag, prot->tag_size);
308 	} else {
309 		ret = prot->tag_size;
310 		if (record->len <= prot->overhead_size)
311 			return -ENOMEM;
312 	}
313 
314 	/* fill prepend */
315 	tls_fill_prepend(ctx, skb_frag_address(&record->frags[0]),
316 			 record->len - prot->overhead_size,
317 			 record_type, prot->version);
318 	return ret;
319 }
320 
321 static int tls_create_new_record(struct tls_offload_context_tx *offload_ctx,
322 				 struct page_frag *pfrag,
323 				 size_t prepend_size)
324 {
325 	struct tls_record_info *record;
326 	skb_frag_t *frag;
327 
328 	record = kmalloc(sizeof(*record), GFP_KERNEL);
329 	if (!record)
330 		return -ENOMEM;
331 
332 	frag = &record->frags[0];
333 	__skb_frag_set_page(frag, pfrag->page);
334 	skb_frag_off_set(frag, pfrag->offset);
335 	skb_frag_size_set(frag, prepend_size);
336 
337 	get_page(pfrag->page);
338 	pfrag->offset += prepend_size;
339 
340 	record->num_frags = 1;
341 	record->len = prepend_size;
342 	offload_ctx->open_record = record;
343 	return 0;
344 }
345 
346 static int tls_do_allocation(struct sock *sk,
347 			     struct tls_offload_context_tx *offload_ctx,
348 			     struct page_frag *pfrag,
349 			     size_t prepend_size)
350 {
351 	int ret;
352 
353 	if (!offload_ctx->open_record) {
354 		if (unlikely(!skb_page_frag_refill(prepend_size, pfrag,
355 						   sk->sk_allocation))) {
356 			sk->sk_prot->enter_memory_pressure(sk);
357 			sk_stream_moderate_sndbuf(sk);
358 			return -ENOMEM;
359 		}
360 
361 		ret = tls_create_new_record(offload_ctx, pfrag, prepend_size);
362 		if (ret)
363 			return ret;
364 
365 		if (pfrag->size > pfrag->offset)
366 			return 0;
367 	}
368 
369 	if (!sk_page_frag_refill(sk, pfrag))
370 		return -ENOMEM;
371 
372 	return 0;
373 }
374 
375 static int tls_device_copy_data(void *addr, size_t bytes, struct iov_iter *i)
376 {
377 	size_t pre_copy, nocache;
378 
379 	pre_copy = ~((unsigned long)addr - 1) & (SMP_CACHE_BYTES - 1);
380 	if (pre_copy) {
381 		pre_copy = min(pre_copy, bytes);
382 		if (copy_from_iter(addr, pre_copy, i) != pre_copy)
383 			return -EFAULT;
384 		bytes -= pre_copy;
385 		addr += pre_copy;
386 	}
387 
388 	nocache = round_down(bytes, SMP_CACHE_BYTES);
389 	if (copy_from_iter_nocache(addr, nocache, i) != nocache)
390 		return -EFAULT;
391 	bytes -= nocache;
392 	addr += nocache;
393 
394 	if (bytes && copy_from_iter(addr, bytes, i) != bytes)
395 		return -EFAULT;
396 
397 	return 0;
398 }
399 
400 static int tls_push_data(struct sock *sk,
401 			 struct iov_iter *msg_iter,
402 			 size_t size, int flags,
403 			 unsigned char record_type)
404 {
405 	struct tls_context *tls_ctx = tls_get_ctx(sk);
406 	struct tls_prot_info *prot = &tls_ctx->prot_info;
407 	struct tls_offload_context_tx *ctx = tls_offload_ctx_tx(tls_ctx);
408 	int more = flags & (MSG_SENDPAGE_NOTLAST | MSG_MORE);
409 	struct tls_record_info *record = ctx->open_record;
410 	int tls_push_record_flags;
411 	struct page_frag *pfrag;
412 	size_t orig_size = size;
413 	u32 max_open_record_len;
414 	int copy, rc = 0;
415 	bool done = false;
416 	long timeo;
417 
418 	if (flags &
419 	    ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL | MSG_SENDPAGE_NOTLAST))
420 		return -ENOTSUPP;
421 
422 	if (sk->sk_err)
423 		return -sk->sk_err;
424 
425 	flags |= MSG_SENDPAGE_DECRYPTED;
426 	tls_push_record_flags = flags | MSG_SENDPAGE_NOTLAST;
427 
428 	timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
429 	if (tls_is_partially_sent_record(tls_ctx)) {
430 		rc = tls_push_partial_record(sk, tls_ctx, flags);
431 		if (rc < 0)
432 			return rc;
433 	}
434 
435 	pfrag = sk_page_frag(sk);
436 
437 	/* TLS_HEADER_SIZE is not counted as part of the TLS record, and
438 	 * we need to leave room for an authentication tag.
439 	 */
440 	max_open_record_len = TLS_MAX_PAYLOAD_SIZE +
441 			      prot->prepend_size;
442 	do {
443 		rc = tls_do_allocation(sk, ctx, pfrag,
444 				       prot->prepend_size);
445 		if (rc) {
446 			rc = sk_stream_wait_memory(sk, &timeo);
447 			if (!rc)
448 				continue;
449 
450 			record = ctx->open_record;
451 			if (!record)
452 				break;
453 handle_error:
454 			if (record_type != TLS_RECORD_TYPE_DATA) {
455 				/* avoid sending partial
456 				 * record with type !=
457 				 * application_data
458 				 */
459 				size = orig_size;
460 				destroy_record(record);
461 				ctx->open_record = NULL;
462 			} else if (record->len > prot->prepend_size) {
463 				goto last_record;
464 			}
465 
466 			break;
467 		}
468 
469 		record = ctx->open_record;
470 		copy = min_t(size_t, size, (pfrag->size - pfrag->offset));
471 		copy = min_t(size_t, copy, (max_open_record_len - record->len));
472 
473 		rc = tls_device_copy_data(page_address(pfrag->page) +
474 					  pfrag->offset, copy, msg_iter);
475 		if (rc)
476 			goto handle_error;
477 		tls_append_frag(record, pfrag, copy);
478 
479 		size -= copy;
480 		if (!size) {
481 last_record:
482 			tls_push_record_flags = flags;
483 			if (more) {
484 				tls_ctx->pending_open_record_frags =
485 						!!record->num_frags;
486 				break;
487 			}
488 
489 			done = true;
490 		}
491 
492 		if (done || record->len >= max_open_record_len ||
493 		    (record->num_frags >= MAX_SKB_FRAGS - 1)) {
494 			rc = tls_device_record_close(sk, tls_ctx, record,
495 						     pfrag, record_type);
496 			if (rc) {
497 				if (rc > 0) {
498 					size += rc;
499 				} else {
500 					size = orig_size;
501 					destroy_record(record);
502 					ctx->open_record = NULL;
503 					break;
504 				}
505 			}
506 
507 			rc = tls_push_record(sk,
508 					     tls_ctx,
509 					     ctx,
510 					     record,
511 					     tls_push_record_flags);
512 			if (rc < 0)
513 				break;
514 		}
515 	} while (!done);
516 
517 	if (orig_size - size > 0)
518 		rc = orig_size - size;
519 
520 	return rc;
521 }
522 
523 int tls_device_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
524 {
525 	unsigned char record_type = TLS_RECORD_TYPE_DATA;
526 	int rc;
527 
528 	lock_sock(sk);
529 
530 	if (unlikely(msg->msg_controllen)) {
531 		rc = tls_proccess_cmsg(sk, msg, &record_type);
532 		if (rc)
533 			goto out;
534 	}
535 
536 	rc = tls_push_data(sk, &msg->msg_iter, size,
537 			   msg->msg_flags, record_type);
538 
539 out:
540 	release_sock(sk);
541 	return rc;
542 }
543 
544 int tls_device_sendpage(struct sock *sk, struct page *page,
545 			int offset, size_t size, int flags)
546 {
547 	struct iov_iter	msg_iter;
548 	char *kaddr = kmap(page);
549 	struct kvec iov;
550 	int rc;
551 
552 	if (flags & MSG_SENDPAGE_NOTLAST)
553 		flags |= MSG_MORE;
554 
555 	lock_sock(sk);
556 
557 	if (flags & MSG_OOB) {
558 		rc = -ENOTSUPP;
559 		goto out;
560 	}
561 
562 	iov.iov_base = kaddr + offset;
563 	iov.iov_len = size;
564 	iov_iter_kvec(&msg_iter, WRITE, &iov, 1, size);
565 	rc = tls_push_data(sk, &msg_iter, size,
566 			   flags, TLS_RECORD_TYPE_DATA);
567 	kunmap(page);
568 
569 out:
570 	release_sock(sk);
571 	return rc;
572 }
573 
574 struct tls_record_info *tls_get_record(struct tls_offload_context_tx *context,
575 				       u32 seq, u64 *p_record_sn)
576 {
577 	u64 record_sn = context->hint_record_sn;
578 	struct tls_record_info *info;
579 
580 	info = context->retransmit_hint;
581 	if (!info ||
582 	    before(seq, info->end_seq - info->len)) {
583 		/* if retransmit_hint is irrelevant start
584 		 * from the beggining of the list
585 		 */
586 		info = list_first_entry_or_null(&context->records_list,
587 						struct tls_record_info, list);
588 		if (!info)
589 			return NULL;
590 		record_sn = context->unacked_record_sn;
591 	}
592 
593 	/* We just need the _rcu for the READ_ONCE() */
594 	rcu_read_lock();
595 	list_for_each_entry_from_rcu(info, &context->records_list, list) {
596 		if (before(seq, info->end_seq)) {
597 			if (!context->retransmit_hint ||
598 			    after(info->end_seq,
599 				  context->retransmit_hint->end_seq)) {
600 				context->hint_record_sn = record_sn;
601 				context->retransmit_hint = info;
602 			}
603 			*p_record_sn = record_sn;
604 			goto exit_rcu_unlock;
605 		}
606 		record_sn++;
607 	}
608 	info = NULL;
609 
610 exit_rcu_unlock:
611 	rcu_read_unlock();
612 	return info;
613 }
614 EXPORT_SYMBOL(tls_get_record);
615 
616 static int tls_device_push_pending_record(struct sock *sk, int flags)
617 {
618 	struct iov_iter	msg_iter;
619 
620 	iov_iter_kvec(&msg_iter, WRITE, NULL, 0, 0);
621 	return tls_push_data(sk, &msg_iter, 0, flags, TLS_RECORD_TYPE_DATA);
622 }
623 
624 void tls_device_write_space(struct sock *sk, struct tls_context *ctx)
625 {
626 	if (!sk->sk_write_pending && tls_is_partially_sent_record(ctx)) {
627 		gfp_t sk_allocation = sk->sk_allocation;
628 
629 		sk->sk_allocation = GFP_ATOMIC;
630 		tls_push_partial_record(sk, ctx,
631 					MSG_DONTWAIT | MSG_NOSIGNAL |
632 					MSG_SENDPAGE_DECRYPTED);
633 		sk->sk_allocation = sk_allocation;
634 	}
635 }
636 
637 static void tls_device_resync_rx(struct tls_context *tls_ctx,
638 				 struct sock *sk, u32 seq, u8 *rcd_sn)
639 {
640 	struct net_device *netdev;
641 
642 	if (WARN_ON(test_and_set_bit(TLS_RX_SYNC_RUNNING, &tls_ctx->flags)))
643 		return;
644 	netdev = READ_ONCE(tls_ctx->netdev);
645 	if (netdev)
646 		netdev->tlsdev_ops->tls_dev_resync(netdev, sk, seq, rcd_sn,
647 						   TLS_OFFLOAD_CTX_DIR_RX);
648 	clear_bit_unlock(TLS_RX_SYNC_RUNNING, &tls_ctx->flags);
649 }
650 
651 void tls_device_rx_resync_new_rec(struct sock *sk, u32 rcd_len, u32 seq)
652 {
653 	struct tls_context *tls_ctx = tls_get_ctx(sk);
654 	struct tls_offload_context_rx *rx_ctx;
655 	u8 rcd_sn[TLS_MAX_REC_SEQ_SIZE];
656 	struct tls_prot_info *prot;
657 	u32 is_req_pending;
658 	s64 resync_req;
659 	u32 req_seq;
660 
661 	if (tls_ctx->rx_conf != TLS_HW)
662 		return;
663 
664 	prot = &tls_ctx->prot_info;
665 	rx_ctx = tls_offload_ctx_rx(tls_ctx);
666 	memcpy(rcd_sn, tls_ctx->rx.rec_seq, prot->rec_seq_size);
667 
668 	switch (rx_ctx->resync_type) {
669 	case TLS_OFFLOAD_SYNC_TYPE_DRIVER_REQ:
670 		resync_req = atomic64_read(&rx_ctx->resync_req);
671 		req_seq = resync_req >> 32;
672 		seq += TLS_HEADER_SIZE - 1;
673 		is_req_pending = resync_req;
674 
675 		if (likely(!is_req_pending) || req_seq != seq ||
676 		    !atomic64_try_cmpxchg(&rx_ctx->resync_req, &resync_req, 0))
677 			return;
678 		break;
679 	case TLS_OFFLOAD_SYNC_TYPE_CORE_NEXT_HINT:
680 		if (likely(!rx_ctx->resync_nh_do_now))
681 			return;
682 
683 		/* head of next rec is already in, note that the sock_inq will
684 		 * include the currently parsed message when called from parser
685 		 */
686 		if (tcp_inq(sk) > rcd_len)
687 			return;
688 
689 		rx_ctx->resync_nh_do_now = 0;
690 		seq += rcd_len;
691 		tls_bigint_increment(rcd_sn, prot->rec_seq_size);
692 		break;
693 	}
694 
695 	tls_device_resync_rx(tls_ctx, sk, seq, rcd_sn);
696 }
697 
698 static void tls_device_core_ctrl_rx_resync(struct tls_context *tls_ctx,
699 					   struct tls_offload_context_rx *ctx,
700 					   struct sock *sk, struct sk_buff *skb)
701 {
702 	struct strp_msg *rxm;
703 
704 	/* device will request resyncs by itself based on stream scan */
705 	if (ctx->resync_type != TLS_OFFLOAD_SYNC_TYPE_CORE_NEXT_HINT)
706 		return;
707 	/* already scheduled */
708 	if (ctx->resync_nh_do_now)
709 		return;
710 	/* seen decrypted fragments since last fully-failed record */
711 	if (ctx->resync_nh_reset) {
712 		ctx->resync_nh_reset = 0;
713 		ctx->resync_nh.decrypted_failed = 1;
714 		ctx->resync_nh.decrypted_tgt = TLS_DEVICE_RESYNC_NH_START_IVAL;
715 		return;
716 	}
717 
718 	if (++ctx->resync_nh.decrypted_failed <= ctx->resync_nh.decrypted_tgt)
719 		return;
720 
721 	/* doing resync, bump the next target in case it fails */
722 	if (ctx->resync_nh.decrypted_tgt < TLS_DEVICE_RESYNC_NH_MAX_IVAL)
723 		ctx->resync_nh.decrypted_tgt *= 2;
724 	else
725 		ctx->resync_nh.decrypted_tgt += TLS_DEVICE_RESYNC_NH_MAX_IVAL;
726 
727 	rxm = strp_msg(skb);
728 
729 	/* head of next rec is already in, parser will sync for us */
730 	if (tcp_inq(sk) > rxm->full_len) {
731 		ctx->resync_nh_do_now = 1;
732 	} else {
733 		struct tls_prot_info *prot = &tls_ctx->prot_info;
734 		u8 rcd_sn[TLS_MAX_REC_SEQ_SIZE];
735 
736 		memcpy(rcd_sn, tls_ctx->rx.rec_seq, prot->rec_seq_size);
737 		tls_bigint_increment(rcd_sn, prot->rec_seq_size);
738 
739 		tls_device_resync_rx(tls_ctx, sk, tcp_sk(sk)->copied_seq,
740 				     rcd_sn);
741 	}
742 }
743 
744 static int tls_device_reencrypt(struct sock *sk, struct sk_buff *skb)
745 {
746 	struct strp_msg *rxm = strp_msg(skb);
747 	int err = 0, offset = rxm->offset, copy, nsg, data_len, pos;
748 	struct sk_buff *skb_iter, *unused;
749 	struct scatterlist sg[1];
750 	char *orig_buf, *buf;
751 
752 	orig_buf = kmalloc(rxm->full_len + TLS_HEADER_SIZE +
753 			   TLS_CIPHER_AES_GCM_128_IV_SIZE, sk->sk_allocation);
754 	if (!orig_buf)
755 		return -ENOMEM;
756 	buf = orig_buf;
757 
758 	nsg = skb_cow_data(skb, 0, &unused);
759 	if (unlikely(nsg < 0)) {
760 		err = nsg;
761 		goto free_buf;
762 	}
763 
764 	sg_init_table(sg, 1);
765 	sg_set_buf(&sg[0], buf,
766 		   rxm->full_len + TLS_HEADER_SIZE +
767 		   TLS_CIPHER_AES_GCM_128_IV_SIZE);
768 	err = skb_copy_bits(skb, offset, buf,
769 			    TLS_HEADER_SIZE + TLS_CIPHER_AES_GCM_128_IV_SIZE);
770 	if (err)
771 		goto free_buf;
772 
773 	/* We are interested only in the decrypted data not the auth */
774 	err = decrypt_skb(sk, skb, sg);
775 	if (err != -EBADMSG)
776 		goto free_buf;
777 	else
778 		err = 0;
779 
780 	data_len = rxm->full_len - TLS_CIPHER_AES_GCM_128_TAG_SIZE;
781 
782 	if (skb_pagelen(skb) > offset) {
783 		copy = min_t(int, skb_pagelen(skb) - offset, data_len);
784 
785 		if (skb->decrypted) {
786 			err = skb_store_bits(skb, offset, buf, copy);
787 			if (err)
788 				goto free_buf;
789 		}
790 
791 		offset += copy;
792 		buf += copy;
793 	}
794 
795 	pos = skb_pagelen(skb);
796 	skb_walk_frags(skb, skb_iter) {
797 		int frag_pos;
798 
799 		/* Practically all frags must belong to msg if reencrypt
800 		 * is needed with current strparser and coalescing logic,
801 		 * but strparser may "get optimized", so let's be safe.
802 		 */
803 		if (pos + skb_iter->len <= offset)
804 			goto done_with_frag;
805 		if (pos >= data_len + rxm->offset)
806 			break;
807 
808 		frag_pos = offset - pos;
809 		copy = min_t(int, skb_iter->len - frag_pos,
810 			     data_len + rxm->offset - offset);
811 
812 		if (skb_iter->decrypted) {
813 			err = skb_store_bits(skb_iter, frag_pos, buf, copy);
814 			if (err)
815 				goto free_buf;
816 		}
817 
818 		offset += copy;
819 		buf += copy;
820 done_with_frag:
821 		pos += skb_iter->len;
822 	}
823 
824 free_buf:
825 	kfree(orig_buf);
826 	return err;
827 }
828 
829 int tls_device_decrypted(struct sock *sk, struct sk_buff *skb)
830 {
831 	struct tls_context *tls_ctx = tls_get_ctx(sk);
832 	struct tls_offload_context_rx *ctx = tls_offload_ctx_rx(tls_ctx);
833 	int is_decrypted = skb->decrypted;
834 	int is_encrypted = !is_decrypted;
835 	struct sk_buff *skb_iter;
836 
837 	/* Check if all the data is decrypted already */
838 	skb_walk_frags(skb, skb_iter) {
839 		is_decrypted &= skb_iter->decrypted;
840 		is_encrypted &= !skb_iter->decrypted;
841 	}
842 
843 	ctx->sw.decrypted |= is_decrypted;
844 
845 	/* Return immediately if the record is either entirely plaintext or
846 	 * entirely ciphertext. Otherwise handle reencrypt partially decrypted
847 	 * record.
848 	 */
849 	if (is_decrypted) {
850 		ctx->resync_nh_reset = 1;
851 		return 0;
852 	}
853 	if (is_encrypted) {
854 		tls_device_core_ctrl_rx_resync(tls_ctx, ctx, sk, skb);
855 		return 0;
856 	}
857 
858 	ctx->resync_nh_reset = 1;
859 	return tls_device_reencrypt(sk, skb);
860 }
861 
862 static void tls_device_attach(struct tls_context *ctx, struct sock *sk,
863 			      struct net_device *netdev)
864 {
865 	if (sk->sk_destruct != tls_device_sk_destruct) {
866 		refcount_set(&ctx->refcount, 1);
867 		dev_hold(netdev);
868 		ctx->netdev = netdev;
869 		spin_lock_irq(&tls_device_lock);
870 		list_add_tail(&ctx->list, &tls_device_list);
871 		spin_unlock_irq(&tls_device_lock);
872 
873 		ctx->sk_destruct = sk->sk_destruct;
874 		sk->sk_destruct = tls_device_sk_destruct;
875 	}
876 }
877 
878 int tls_set_device_offload(struct sock *sk, struct tls_context *ctx)
879 {
880 	u16 nonce_size, tag_size, iv_size, rec_seq_size;
881 	struct tls_context *tls_ctx = tls_get_ctx(sk);
882 	struct tls_prot_info *prot = &tls_ctx->prot_info;
883 	struct tls_record_info *start_marker_record;
884 	struct tls_offload_context_tx *offload_ctx;
885 	struct tls_crypto_info *crypto_info;
886 	struct net_device *netdev;
887 	char *iv, *rec_seq;
888 	struct sk_buff *skb;
889 	__be64 rcd_sn;
890 	int rc;
891 
892 	if (!ctx)
893 		return -EINVAL;
894 
895 	if (ctx->priv_ctx_tx)
896 		return -EEXIST;
897 
898 	start_marker_record = kmalloc(sizeof(*start_marker_record), GFP_KERNEL);
899 	if (!start_marker_record)
900 		return -ENOMEM;
901 
902 	offload_ctx = kzalloc(TLS_OFFLOAD_CONTEXT_SIZE_TX, GFP_KERNEL);
903 	if (!offload_ctx) {
904 		rc = -ENOMEM;
905 		goto free_marker_record;
906 	}
907 
908 	crypto_info = &ctx->crypto_send.info;
909 	if (crypto_info->version != TLS_1_2_VERSION) {
910 		rc = -EOPNOTSUPP;
911 		goto free_offload_ctx;
912 	}
913 
914 	switch (crypto_info->cipher_type) {
915 	case TLS_CIPHER_AES_GCM_128:
916 		nonce_size = TLS_CIPHER_AES_GCM_128_IV_SIZE;
917 		tag_size = TLS_CIPHER_AES_GCM_128_TAG_SIZE;
918 		iv_size = TLS_CIPHER_AES_GCM_128_IV_SIZE;
919 		iv = ((struct tls12_crypto_info_aes_gcm_128 *)crypto_info)->iv;
920 		rec_seq_size = TLS_CIPHER_AES_GCM_128_REC_SEQ_SIZE;
921 		rec_seq =
922 		 ((struct tls12_crypto_info_aes_gcm_128 *)crypto_info)->rec_seq;
923 		break;
924 	default:
925 		rc = -EINVAL;
926 		goto free_offload_ctx;
927 	}
928 
929 	/* Sanity-check the rec_seq_size for stack allocations */
930 	if (rec_seq_size > TLS_MAX_REC_SEQ_SIZE) {
931 		rc = -EINVAL;
932 		goto free_offload_ctx;
933 	}
934 
935 	prot->version = crypto_info->version;
936 	prot->cipher_type = crypto_info->cipher_type;
937 	prot->prepend_size = TLS_HEADER_SIZE + nonce_size;
938 	prot->tag_size = tag_size;
939 	prot->overhead_size = prot->prepend_size + prot->tag_size;
940 	prot->iv_size = iv_size;
941 	ctx->tx.iv = kmalloc(iv_size + TLS_CIPHER_AES_GCM_128_SALT_SIZE,
942 			     GFP_KERNEL);
943 	if (!ctx->tx.iv) {
944 		rc = -ENOMEM;
945 		goto free_offload_ctx;
946 	}
947 
948 	memcpy(ctx->tx.iv + TLS_CIPHER_AES_GCM_128_SALT_SIZE, iv, iv_size);
949 
950 	prot->rec_seq_size = rec_seq_size;
951 	ctx->tx.rec_seq = kmemdup(rec_seq, rec_seq_size, GFP_KERNEL);
952 	if (!ctx->tx.rec_seq) {
953 		rc = -ENOMEM;
954 		goto free_iv;
955 	}
956 
957 	rc = tls_sw_fallback_init(sk, offload_ctx, crypto_info);
958 	if (rc)
959 		goto free_rec_seq;
960 
961 	/* start at rec_seq - 1 to account for the start marker record */
962 	memcpy(&rcd_sn, ctx->tx.rec_seq, sizeof(rcd_sn));
963 	offload_ctx->unacked_record_sn = be64_to_cpu(rcd_sn) - 1;
964 
965 	start_marker_record->end_seq = tcp_sk(sk)->write_seq;
966 	start_marker_record->len = 0;
967 	start_marker_record->num_frags = 0;
968 
969 	INIT_LIST_HEAD(&offload_ctx->records_list);
970 	list_add_tail(&start_marker_record->list, &offload_ctx->records_list);
971 	spin_lock_init(&offload_ctx->lock);
972 	sg_init_table(offload_ctx->sg_tx_data,
973 		      ARRAY_SIZE(offload_ctx->sg_tx_data));
974 
975 	clean_acked_data_enable(inet_csk(sk), &tls_icsk_clean_acked);
976 	ctx->push_pending_record = tls_device_push_pending_record;
977 
978 	/* TLS offload is greatly simplified if we don't send
979 	 * SKBs where only part of the payload needs to be encrypted.
980 	 * So mark the last skb in the write queue as end of record.
981 	 */
982 	skb = tcp_write_queue_tail(sk);
983 	if (skb)
984 		TCP_SKB_CB(skb)->eor = 1;
985 
986 	netdev = get_netdev_for_sock(sk);
987 	if (!netdev) {
988 		pr_err_ratelimited("%s: netdev not found\n", __func__);
989 		rc = -EINVAL;
990 		goto disable_cad;
991 	}
992 
993 	if (!(netdev->features & NETIF_F_HW_TLS_TX)) {
994 		rc = -ENOTSUPP;
995 		goto release_netdev;
996 	}
997 
998 	/* Avoid offloading if the device is down
999 	 * We don't want to offload new flows after
1000 	 * the NETDEV_DOWN event
1001 	 *
1002 	 * device_offload_lock is taken in tls_devices's NETDEV_DOWN
1003 	 * handler thus protecting from the device going down before
1004 	 * ctx was added to tls_device_list.
1005 	 */
1006 	down_read(&device_offload_lock);
1007 	if (!(netdev->flags & IFF_UP)) {
1008 		rc = -EINVAL;
1009 		goto release_lock;
1010 	}
1011 
1012 	ctx->priv_ctx_tx = offload_ctx;
1013 	rc = netdev->tlsdev_ops->tls_dev_add(netdev, sk, TLS_OFFLOAD_CTX_DIR_TX,
1014 					     &ctx->crypto_send.info,
1015 					     tcp_sk(sk)->write_seq);
1016 	if (rc)
1017 		goto release_lock;
1018 
1019 	tls_device_attach(ctx, sk, netdev);
1020 	up_read(&device_offload_lock);
1021 
1022 	/* following this assignment tls_is_sk_tx_device_offloaded
1023 	 * will return true and the context might be accessed
1024 	 * by the netdev's xmit function.
1025 	 */
1026 	smp_store_release(&sk->sk_validate_xmit_skb, tls_validate_xmit_skb);
1027 	dev_put(netdev);
1028 
1029 	return 0;
1030 
1031 release_lock:
1032 	up_read(&device_offload_lock);
1033 release_netdev:
1034 	dev_put(netdev);
1035 disable_cad:
1036 	clean_acked_data_disable(inet_csk(sk));
1037 	crypto_free_aead(offload_ctx->aead_send);
1038 free_rec_seq:
1039 	kfree(ctx->tx.rec_seq);
1040 free_iv:
1041 	kfree(ctx->tx.iv);
1042 free_offload_ctx:
1043 	kfree(offload_ctx);
1044 	ctx->priv_ctx_tx = NULL;
1045 free_marker_record:
1046 	kfree(start_marker_record);
1047 	return rc;
1048 }
1049 
1050 int tls_set_device_offload_rx(struct sock *sk, struct tls_context *ctx)
1051 {
1052 	struct tls_offload_context_rx *context;
1053 	struct net_device *netdev;
1054 	int rc = 0;
1055 
1056 	if (ctx->crypto_recv.info.version != TLS_1_2_VERSION)
1057 		return -EOPNOTSUPP;
1058 
1059 	netdev = get_netdev_for_sock(sk);
1060 	if (!netdev) {
1061 		pr_err_ratelimited("%s: netdev not found\n", __func__);
1062 		return -EINVAL;
1063 	}
1064 
1065 	if (!(netdev->features & NETIF_F_HW_TLS_RX)) {
1066 		rc = -ENOTSUPP;
1067 		goto release_netdev;
1068 	}
1069 
1070 	/* Avoid offloading if the device is down
1071 	 * We don't want to offload new flows after
1072 	 * the NETDEV_DOWN event
1073 	 *
1074 	 * device_offload_lock is taken in tls_devices's NETDEV_DOWN
1075 	 * handler thus protecting from the device going down before
1076 	 * ctx was added to tls_device_list.
1077 	 */
1078 	down_read(&device_offload_lock);
1079 	if (!(netdev->flags & IFF_UP)) {
1080 		rc = -EINVAL;
1081 		goto release_lock;
1082 	}
1083 
1084 	context = kzalloc(TLS_OFFLOAD_CONTEXT_SIZE_RX, GFP_KERNEL);
1085 	if (!context) {
1086 		rc = -ENOMEM;
1087 		goto release_lock;
1088 	}
1089 	context->resync_nh_reset = 1;
1090 
1091 	ctx->priv_ctx_rx = context;
1092 	rc = tls_set_sw_offload(sk, ctx, 0);
1093 	if (rc)
1094 		goto release_ctx;
1095 
1096 	rc = netdev->tlsdev_ops->tls_dev_add(netdev, sk, TLS_OFFLOAD_CTX_DIR_RX,
1097 					     &ctx->crypto_recv.info,
1098 					     tcp_sk(sk)->copied_seq);
1099 	if (rc)
1100 		goto free_sw_resources;
1101 
1102 	tls_device_attach(ctx, sk, netdev);
1103 	up_read(&device_offload_lock);
1104 
1105 	dev_put(netdev);
1106 
1107 	return 0;
1108 
1109 free_sw_resources:
1110 	up_read(&device_offload_lock);
1111 	tls_sw_free_resources_rx(sk);
1112 	down_read(&device_offload_lock);
1113 release_ctx:
1114 	ctx->priv_ctx_rx = NULL;
1115 release_lock:
1116 	up_read(&device_offload_lock);
1117 release_netdev:
1118 	dev_put(netdev);
1119 	return rc;
1120 }
1121 
1122 void tls_device_offload_cleanup_rx(struct sock *sk)
1123 {
1124 	struct tls_context *tls_ctx = tls_get_ctx(sk);
1125 	struct net_device *netdev;
1126 
1127 	down_read(&device_offload_lock);
1128 	netdev = tls_ctx->netdev;
1129 	if (!netdev)
1130 		goto out;
1131 
1132 	netdev->tlsdev_ops->tls_dev_del(netdev, tls_ctx,
1133 					TLS_OFFLOAD_CTX_DIR_RX);
1134 
1135 	if (tls_ctx->tx_conf != TLS_HW) {
1136 		dev_put(netdev);
1137 		tls_ctx->netdev = NULL;
1138 	}
1139 out:
1140 	up_read(&device_offload_lock);
1141 	tls_sw_release_resources_rx(sk);
1142 }
1143 
1144 static int tls_device_down(struct net_device *netdev)
1145 {
1146 	struct tls_context *ctx, *tmp;
1147 	unsigned long flags;
1148 	LIST_HEAD(list);
1149 
1150 	/* Request a write lock to block new offload attempts */
1151 	down_write(&device_offload_lock);
1152 
1153 	spin_lock_irqsave(&tls_device_lock, flags);
1154 	list_for_each_entry_safe(ctx, tmp, &tls_device_list, list) {
1155 		if (ctx->netdev != netdev ||
1156 		    !refcount_inc_not_zero(&ctx->refcount))
1157 			continue;
1158 
1159 		list_move(&ctx->list, &list);
1160 	}
1161 	spin_unlock_irqrestore(&tls_device_lock, flags);
1162 
1163 	list_for_each_entry_safe(ctx, tmp, &list, list)	{
1164 		if (ctx->tx_conf == TLS_HW)
1165 			netdev->tlsdev_ops->tls_dev_del(netdev, ctx,
1166 							TLS_OFFLOAD_CTX_DIR_TX);
1167 		if (ctx->rx_conf == TLS_HW)
1168 			netdev->tlsdev_ops->tls_dev_del(netdev, ctx,
1169 							TLS_OFFLOAD_CTX_DIR_RX);
1170 		WRITE_ONCE(ctx->netdev, NULL);
1171 		smp_mb__before_atomic(); /* pairs with test_and_set_bit() */
1172 		while (test_bit(TLS_RX_SYNC_RUNNING, &ctx->flags))
1173 			usleep_range(10, 200);
1174 		dev_put(netdev);
1175 		list_del_init(&ctx->list);
1176 
1177 		if (refcount_dec_and_test(&ctx->refcount))
1178 			tls_device_free_ctx(ctx);
1179 	}
1180 
1181 	up_write(&device_offload_lock);
1182 
1183 	flush_work(&tls_device_gc_work);
1184 
1185 	return NOTIFY_DONE;
1186 }
1187 
1188 static int tls_dev_event(struct notifier_block *this, unsigned long event,
1189 			 void *ptr)
1190 {
1191 	struct net_device *dev = netdev_notifier_info_to_dev(ptr);
1192 
1193 	if (!dev->tlsdev_ops &&
1194 	    !(dev->features & (NETIF_F_HW_TLS_RX | NETIF_F_HW_TLS_TX)))
1195 		return NOTIFY_DONE;
1196 
1197 	switch (event) {
1198 	case NETDEV_REGISTER:
1199 	case NETDEV_FEAT_CHANGE:
1200 		if ((dev->features & NETIF_F_HW_TLS_RX) &&
1201 		    !dev->tlsdev_ops->tls_dev_resync)
1202 			return NOTIFY_BAD;
1203 
1204 		if  (dev->tlsdev_ops &&
1205 		     dev->tlsdev_ops->tls_dev_add &&
1206 		     dev->tlsdev_ops->tls_dev_del)
1207 			return NOTIFY_DONE;
1208 		else
1209 			return NOTIFY_BAD;
1210 	case NETDEV_DOWN:
1211 		return tls_device_down(dev);
1212 	}
1213 	return NOTIFY_DONE;
1214 }
1215 
1216 static struct notifier_block tls_dev_notifier = {
1217 	.notifier_call	= tls_dev_event,
1218 };
1219 
1220 void __init tls_device_init(void)
1221 {
1222 	register_netdevice_notifier(&tls_dev_notifier);
1223 }
1224 
1225 void __exit tls_device_cleanup(void)
1226 {
1227 	unregister_netdevice_notifier(&tls_dev_notifier);
1228 	flush_work(&tls_device_gc_work);
1229 	clean_acked_data_flush();
1230 }
1231