xref: /linux/net/tls/tls_device.c (revision 2dbc0838bcf24ca59cabc3130cf3b1d6809cdcd4)
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(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_device_resync_tx(struct sock *sk, struct tls_context *tls_ctx,
213 				 u32 seq)
214 {
215 	struct net_device *netdev;
216 	struct sk_buff *skb;
217 	int err = 0;
218 	u8 *rcd_sn;
219 
220 	skb = tcp_write_queue_tail(sk);
221 	if (skb)
222 		TCP_SKB_CB(skb)->eor = 1;
223 
224 	rcd_sn = tls_ctx->tx.rec_seq;
225 
226 	down_read(&device_offload_lock);
227 	netdev = tls_ctx->netdev;
228 	if (netdev)
229 		err = netdev->tlsdev_ops->tls_dev_resync(netdev, sk, seq,
230 							 rcd_sn,
231 							 TLS_OFFLOAD_CTX_DIR_TX);
232 	up_read(&device_offload_lock);
233 	if (err)
234 		return;
235 
236 	clear_bit_unlock(TLS_TX_SYNC_SCHED, &tls_ctx->flags);
237 }
238 
239 static void tls_append_frag(struct tls_record_info *record,
240 			    struct page_frag *pfrag,
241 			    int size)
242 {
243 	skb_frag_t *frag;
244 
245 	frag = &record->frags[record->num_frags - 1];
246 	if (frag->page.p == pfrag->page &&
247 	    frag->page_offset + frag->size == pfrag->offset) {
248 		frag->size += size;
249 	} else {
250 		++frag;
251 		frag->page.p = pfrag->page;
252 		frag->page_offset = pfrag->offset;
253 		frag->size = size;
254 		++record->num_frags;
255 		get_page(pfrag->page);
256 	}
257 
258 	pfrag->offset += size;
259 	record->len += size;
260 }
261 
262 static int tls_push_record(struct sock *sk,
263 			   struct tls_context *ctx,
264 			   struct tls_offload_context_tx *offload_ctx,
265 			   struct tls_record_info *record,
266 			   struct page_frag *pfrag,
267 			   int flags,
268 			   unsigned char record_type)
269 {
270 	struct tls_prot_info *prot = &ctx->prot_info;
271 	struct tcp_sock *tp = tcp_sk(sk);
272 	struct page_frag dummy_tag_frag;
273 	skb_frag_t *frag;
274 	int i;
275 
276 	/* fill prepend */
277 	frag = &record->frags[0];
278 	tls_fill_prepend(ctx,
279 			 skb_frag_address(frag),
280 			 record->len - prot->prepend_size,
281 			 record_type,
282 			 prot->version);
283 
284 	/* HW doesn't care about the data in the tag, because it fills it. */
285 	dummy_tag_frag.page = skb_frag_page(frag);
286 	dummy_tag_frag.offset = 0;
287 
288 	tls_append_frag(record, &dummy_tag_frag, prot->tag_size);
289 	record->end_seq = tp->write_seq + record->len;
290 	spin_lock_irq(&offload_ctx->lock);
291 	list_add_tail(&record->list, &offload_ctx->records_list);
292 	spin_unlock_irq(&offload_ctx->lock);
293 	offload_ctx->open_record = NULL;
294 
295 	if (test_bit(TLS_TX_SYNC_SCHED, &ctx->flags))
296 		tls_device_resync_tx(sk, ctx, tp->write_seq);
297 
298 	tls_advance_record_sn(sk, prot, &ctx->tx);
299 
300 	for (i = 0; i < record->num_frags; i++) {
301 		frag = &record->frags[i];
302 		sg_unmark_end(&offload_ctx->sg_tx_data[i]);
303 		sg_set_page(&offload_ctx->sg_tx_data[i], skb_frag_page(frag),
304 			    frag->size, frag->page_offset);
305 		sk_mem_charge(sk, frag->size);
306 		get_page(skb_frag_page(frag));
307 	}
308 	sg_mark_end(&offload_ctx->sg_tx_data[record->num_frags - 1]);
309 
310 	/* all ready, send */
311 	return tls_push_sg(sk, ctx, offload_ctx->sg_tx_data, 0, flags);
312 }
313 
314 static int tls_create_new_record(struct tls_offload_context_tx *offload_ctx,
315 				 struct page_frag *pfrag,
316 				 size_t prepend_size)
317 {
318 	struct tls_record_info *record;
319 	skb_frag_t *frag;
320 
321 	record = kmalloc(sizeof(*record), GFP_KERNEL);
322 	if (!record)
323 		return -ENOMEM;
324 
325 	frag = &record->frags[0];
326 	__skb_frag_set_page(frag, pfrag->page);
327 	frag->page_offset = pfrag->offset;
328 	skb_frag_size_set(frag, prepend_size);
329 
330 	get_page(pfrag->page);
331 	pfrag->offset += prepend_size;
332 
333 	record->num_frags = 1;
334 	record->len = prepend_size;
335 	offload_ctx->open_record = record;
336 	return 0;
337 }
338 
339 static int tls_do_allocation(struct sock *sk,
340 			     struct tls_offload_context_tx *offload_ctx,
341 			     struct page_frag *pfrag,
342 			     size_t prepend_size)
343 {
344 	int ret;
345 
346 	if (!offload_ctx->open_record) {
347 		if (unlikely(!skb_page_frag_refill(prepend_size, pfrag,
348 						   sk->sk_allocation))) {
349 			sk->sk_prot->enter_memory_pressure(sk);
350 			sk_stream_moderate_sndbuf(sk);
351 			return -ENOMEM;
352 		}
353 
354 		ret = tls_create_new_record(offload_ctx, pfrag, prepend_size);
355 		if (ret)
356 			return ret;
357 
358 		if (pfrag->size > pfrag->offset)
359 			return 0;
360 	}
361 
362 	if (!sk_page_frag_refill(sk, pfrag))
363 		return -ENOMEM;
364 
365 	return 0;
366 }
367 
368 static int tls_push_data(struct sock *sk,
369 			 struct iov_iter *msg_iter,
370 			 size_t size, int flags,
371 			 unsigned char record_type)
372 {
373 	struct tls_context *tls_ctx = tls_get_ctx(sk);
374 	struct tls_prot_info *prot = &tls_ctx->prot_info;
375 	struct tls_offload_context_tx *ctx = tls_offload_ctx_tx(tls_ctx);
376 	int tls_push_record_flags = flags | MSG_SENDPAGE_NOTLAST;
377 	int more = flags & (MSG_SENDPAGE_NOTLAST | MSG_MORE);
378 	struct tls_record_info *record = ctx->open_record;
379 	struct page_frag *pfrag;
380 	size_t orig_size = size;
381 	u32 max_open_record_len;
382 	int copy, rc = 0;
383 	bool done = false;
384 	long timeo;
385 
386 	if (flags &
387 	    ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL | MSG_SENDPAGE_NOTLAST))
388 		return -ENOTSUPP;
389 
390 	if (sk->sk_err)
391 		return -sk->sk_err;
392 
393 	timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
394 	if (tls_is_partially_sent_record(tls_ctx)) {
395 		rc = tls_push_partial_record(sk, tls_ctx, flags);
396 		if (rc < 0)
397 			return rc;
398 	}
399 
400 	pfrag = sk_page_frag(sk);
401 
402 	/* TLS_HEADER_SIZE is not counted as part of the TLS record, and
403 	 * we need to leave room for an authentication tag.
404 	 */
405 	max_open_record_len = TLS_MAX_PAYLOAD_SIZE +
406 			      prot->prepend_size;
407 	do {
408 		rc = tls_do_allocation(sk, ctx, pfrag,
409 				       prot->prepend_size);
410 		if (rc) {
411 			rc = sk_stream_wait_memory(sk, &timeo);
412 			if (!rc)
413 				continue;
414 
415 			record = ctx->open_record;
416 			if (!record)
417 				break;
418 handle_error:
419 			if (record_type != TLS_RECORD_TYPE_DATA) {
420 				/* avoid sending partial
421 				 * record with type !=
422 				 * application_data
423 				 */
424 				size = orig_size;
425 				destroy_record(record);
426 				ctx->open_record = NULL;
427 			} else if (record->len > prot->prepend_size) {
428 				goto last_record;
429 			}
430 
431 			break;
432 		}
433 
434 		record = ctx->open_record;
435 		copy = min_t(size_t, size, (pfrag->size - pfrag->offset));
436 		copy = min_t(size_t, copy, (max_open_record_len - record->len));
437 
438 		if (copy_from_iter_nocache(page_address(pfrag->page) +
439 					       pfrag->offset,
440 					   copy, msg_iter) != copy) {
441 			rc = -EFAULT;
442 			goto handle_error;
443 		}
444 		tls_append_frag(record, pfrag, copy);
445 
446 		size -= copy;
447 		if (!size) {
448 last_record:
449 			tls_push_record_flags = flags;
450 			if (more) {
451 				tls_ctx->pending_open_record_frags =
452 						!!record->num_frags;
453 				break;
454 			}
455 
456 			done = true;
457 		}
458 
459 		if (done || record->len >= max_open_record_len ||
460 		    (record->num_frags >= MAX_SKB_FRAGS - 1)) {
461 			rc = tls_push_record(sk,
462 					     tls_ctx,
463 					     ctx,
464 					     record,
465 					     pfrag,
466 					     tls_push_record_flags,
467 					     record_type);
468 			if (rc < 0)
469 				break;
470 		}
471 	} while (!done);
472 
473 	if (orig_size - size > 0)
474 		rc = orig_size - size;
475 
476 	return rc;
477 }
478 
479 int tls_device_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
480 {
481 	unsigned char record_type = TLS_RECORD_TYPE_DATA;
482 	int rc;
483 
484 	lock_sock(sk);
485 
486 	if (unlikely(msg->msg_controllen)) {
487 		rc = tls_proccess_cmsg(sk, msg, &record_type);
488 		if (rc)
489 			goto out;
490 	}
491 
492 	rc = tls_push_data(sk, &msg->msg_iter, size,
493 			   msg->msg_flags, record_type);
494 
495 out:
496 	release_sock(sk);
497 	return rc;
498 }
499 
500 int tls_device_sendpage(struct sock *sk, struct page *page,
501 			int offset, size_t size, int flags)
502 {
503 	struct iov_iter	msg_iter;
504 	char *kaddr = kmap(page);
505 	struct kvec iov;
506 	int rc;
507 
508 	if (flags & MSG_SENDPAGE_NOTLAST)
509 		flags |= MSG_MORE;
510 
511 	lock_sock(sk);
512 
513 	if (flags & MSG_OOB) {
514 		rc = -ENOTSUPP;
515 		goto out;
516 	}
517 
518 	iov.iov_base = kaddr + offset;
519 	iov.iov_len = size;
520 	iov_iter_kvec(&msg_iter, WRITE, &iov, 1, size);
521 	rc = tls_push_data(sk, &msg_iter, size,
522 			   flags, TLS_RECORD_TYPE_DATA);
523 	kunmap(page);
524 
525 out:
526 	release_sock(sk);
527 	return rc;
528 }
529 
530 struct tls_record_info *tls_get_record(struct tls_offload_context_tx *context,
531 				       u32 seq, u64 *p_record_sn)
532 {
533 	u64 record_sn = context->hint_record_sn;
534 	struct tls_record_info *info;
535 
536 	info = context->retransmit_hint;
537 	if (!info ||
538 	    before(seq, info->end_seq - info->len)) {
539 		/* if retransmit_hint is irrelevant start
540 		 * from the beggining of the list
541 		 */
542 		info = list_first_entry(&context->records_list,
543 					struct tls_record_info, list);
544 		record_sn = context->unacked_record_sn;
545 	}
546 
547 	list_for_each_entry_from(info, &context->records_list, list) {
548 		if (before(seq, info->end_seq)) {
549 			if (!context->retransmit_hint ||
550 			    after(info->end_seq,
551 				  context->retransmit_hint->end_seq)) {
552 				context->hint_record_sn = record_sn;
553 				context->retransmit_hint = info;
554 			}
555 			*p_record_sn = record_sn;
556 			return info;
557 		}
558 		record_sn++;
559 	}
560 
561 	return NULL;
562 }
563 EXPORT_SYMBOL(tls_get_record);
564 
565 static int tls_device_push_pending_record(struct sock *sk, int flags)
566 {
567 	struct iov_iter	msg_iter;
568 
569 	iov_iter_kvec(&msg_iter, WRITE, NULL, 0, 0);
570 	return tls_push_data(sk, &msg_iter, 0, flags, TLS_RECORD_TYPE_DATA);
571 }
572 
573 void tls_device_write_space(struct sock *sk, struct tls_context *ctx)
574 {
575 	if (!sk->sk_write_pending && tls_is_partially_sent_record(ctx)) {
576 		gfp_t sk_allocation = sk->sk_allocation;
577 
578 		sk->sk_allocation = GFP_ATOMIC;
579 		tls_push_partial_record(sk, ctx, MSG_DONTWAIT | MSG_NOSIGNAL);
580 		sk->sk_allocation = sk_allocation;
581 	}
582 }
583 
584 static void tls_device_resync_rx(struct tls_context *tls_ctx,
585 				 struct sock *sk, u32 seq, u8 *rcd_sn)
586 {
587 	struct net_device *netdev;
588 
589 	if (WARN_ON(test_and_set_bit(TLS_RX_SYNC_RUNNING, &tls_ctx->flags)))
590 		return;
591 	netdev = READ_ONCE(tls_ctx->netdev);
592 	if (netdev)
593 		netdev->tlsdev_ops->tls_dev_resync(netdev, sk, seq, rcd_sn,
594 						   TLS_OFFLOAD_CTX_DIR_RX);
595 	clear_bit_unlock(TLS_RX_SYNC_RUNNING, &tls_ctx->flags);
596 }
597 
598 void tls_device_rx_resync_new_rec(struct sock *sk, u32 rcd_len, u32 seq)
599 {
600 	struct tls_context *tls_ctx = tls_get_ctx(sk);
601 	struct tls_offload_context_rx *rx_ctx;
602 	u8 rcd_sn[TLS_MAX_REC_SEQ_SIZE];
603 	struct tls_prot_info *prot;
604 	u32 is_req_pending;
605 	s64 resync_req;
606 	u32 req_seq;
607 
608 	if (tls_ctx->rx_conf != TLS_HW)
609 		return;
610 
611 	prot = &tls_ctx->prot_info;
612 	rx_ctx = tls_offload_ctx_rx(tls_ctx);
613 	memcpy(rcd_sn, tls_ctx->rx.rec_seq, prot->rec_seq_size);
614 
615 	switch (rx_ctx->resync_type) {
616 	case TLS_OFFLOAD_SYNC_TYPE_DRIVER_REQ:
617 		resync_req = atomic64_read(&rx_ctx->resync_req);
618 		req_seq = resync_req >> 32;
619 		seq += TLS_HEADER_SIZE - 1;
620 		is_req_pending = resync_req;
621 
622 		if (likely(!is_req_pending) || req_seq != seq ||
623 		    !atomic64_try_cmpxchg(&rx_ctx->resync_req, &resync_req, 0))
624 			return;
625 		break;
626 	case TLS_OFFLOAD_SYNC_TYPE_CORE_NEXT_HINT:
627 		if (likely(!rx_ctx->resync_nh_do_now))
628 			return;
629 
630 		/* head of next rec is already in, note that the sock_inq will
631 		 * include the currently parsed message when called from parser
632 		 */
633 		if (tcp_inq(sk) > rcd_len)
634 			return;
635 
636 		rx_ctx->resync_nh_do_now = 0;
637 		seq += rcd_len;
638 		tls_bigint_increment(rcd_sn, prot->rec_seq_size);
639 		break;
640 	}
641 
642 	tls_device_resync_rx(tls_ctx, sk, seq, rcd_sn);
643 }
644 
645 static void tls_device_core_ctrl_rx_resync(struct tls_context *tls_ctx,
646 					   struct tls_offload_context_rx *ctx,
647 					   struct sock *sk, struct sk_buff *skb)
648 {
649 	struct strp_msg *rxm;
650 
651 	/* device will request resyncs by itself based on stream scan */
652 	if (ctx->resync_type != TLS_OFFLOAD_SYNC_TYPE_CORE_NEXT_HINT)
653 		return;
654 	/* already scheduled */
655 	if (ctx->resync_nh_do_now)
656 		return;
657 	/* seen decrypted fragments since last fully-failed record */
658 	if (ctx->resync_nh_reset) {
659 		ctx->resync_nh_reset = 0;
660 		ctx->resync_nh.decrypted_failed = 1;
661 		ctx->resync_nh.decrypted_tgt = TLS_DEVICE_RESYNC_NH_START_IVAL;
662 		return;
663 	}
664 
665 	if (++ctx->resync_nh.decrypted_failed <= ctx->resync_nh.decrypted_tgt)
666 		return;
667 
668 	/* doing resync, bump the next target in case it fails */
669 	if (ctx->resync_nh.decrypted_tgt < TLS_DEVICE_RESYNC_NH_MAX_IVAL)
670 		ctx->resync_nh.decrypted_tgt *= 2;
671 	else
672 		ctx->resync_nh.decrypted_tgt += TLS_DEVICE_RESYNC_NH_MAX_IVAL;
673 
674 	rxm = strp_msg(skb);
675 
676 	/* head of next rec is already in, parser will sync for us */
677 	if (tcp_inq(sk) > rxm->full_len) {
678 		ctx->resync_nh_do_now = 1;
679 	} else {
680 		struct tls_prot_info *prot = &tls_ctx->prot_info;
681 		u8 rcd_sn[TLS_MAX_REC_SEQ_SIZE];
682 
683 		memcpy(rcd_sn, tls_ctx->rx.rec_seq, prot->rec_seq_size);
684 		tls_bigint_increment(rcd_sn, prot->rec_seq_size);
685 
686 		tls_device_resync_rx(tls_ctx, sk, tcp_sk(sk)->copied_seq,
687 				     rcd_sn);
688 	}
689 }
690 
691 static int tls_device_reencrypt(struct sock *sk, struct sk_buff *skb)
692 {
693 	struct strp_msg *rxm = strp_msg(skb);
694 	int err = 0, offset = rxm->offset, copy, nsg, data_len, pos;
695 	struct sk_buff *skb_iter, *unused;
696 	struct scatterlist sg[1];
697 	char *orig_buf, *buf;
698 
699 	orig_buf = kmalloc(rxm->full_len + TLS_HEADER_SIZE +
700 			   TLS_CIPHER_AES_GCM_128_IV_SIZE, sk->sk_allocation);
701 	if (!orig_buf)
702 		return -ENOMEM;
703 	buf = orig_buf;
704 
705 	nsg = skb_cow_data(skb, 0, &unused);
706 	if (unlikely(nsg < 0)) {
707 		err = nsg;
708 		goto free_buf;
709 	}
710 
711 	sg_init_table(sg, 1);
712 	sg_set_buf(&sg[0], buf,
713 		   rxm->full_len + TLS_HEADER_SIZE +
714 		   TLS_CIPHER_AES_GCM_128_IV_SIZE);
715 	err = skb_copy_bits(skb, offset, buf,
716 			    TLS_HEADER_SIZE + TLS_CIPHER_AES_GCM_128_IV_SIZE);
717 	if (err)
718 		goto free_buf;
719 
720 	/* We are interested only in the decrypted data not the auth */
721 	err = decrypt_skb(sk, skb, sg);
722 	if (err != -EBADMSG)
723 		goto free_buf;
724 	else
725 		err = 0;
726 
727 	data_len = rxm->full_len - TLS_CIPHER_AES_GCM_128_TAG_SIZE;
728 
729 	if (skb_pagelen(skb) > offset) {
730 		copy = min_t(int, skb_pagelen(skb) - offset, data_len);
731 
732 		if (skb->decrypted) {
733 			err = skb_store_bits(skb, offset, buf, copy);
734 			if (err)
735 				goto free_buf;
736 		}
737 
738 		offset += copy;
739 		buf += copy;
740 	}
741 
742 	pos = skb_pagelen(skb);
743 	skb_walk_frags(skb, skb_iter) {
744 		int frag_pos;
745 
746 		/* Practically all frags must belong to msg if reencrypt
747 		 * is needed with current strparser and coalescing logic,
748 		 * but strparser may "get optimized", so let's be safe.
749 		 */
750 		if (pos + skb_iter->len <= offset)
751 			goto done_with_frag;
752 		if (pos >= data_len + rxm->offset)
753 			break;
754 
755 		frag_pos = offset - pos;
756 		copy = min_t(int, skb_iter->len - frag_pos,
757 			     data_len + rxm->offset - offset);
758 
759 		if (skb_iter->decrypted) {
760 			err = skb_store_bits(skb_iter, frag_pos, buf, copy);
761 			if (err)
762 				goto free_buf;
763 		}
764 
765 		offset += copy;
766 		buf += copy;
767 done_with_frag:
768 		pos += skb_iter->len;
769 	}
770 
771 free_buf:
772 	kfree(orig_buf);
773 	return err;
774 }
775 
776 int tls_device_decrypted(struct sock *sk, struct sk_buff *skb)
777 {
778 	struct tls_context *tls_ctx = tls_get_ctx(sk);
779 	struct tls_offload_context_rx *ctx = tls_offload_ctx_rx(tls_ctx);
780 	int is_decrypted = skb->decrypted;
781 	int is_encrypted = !is_decrypted;
782 	struct sk_buff *skb_iter;
783 
784 	/* Check if all the data is decrypted already */
785 	skb_walk_frags(skb, skb_iter) {
786 		is_decrypted &= skb_iter->decrypted;
787 		is_encrypted &= !skb_iter->decrypted;
788 	}
789 
790 	ctx->sw.decrypted |= is_decrypted;
791 
792 	/* Return immediately if the record is either entirely plaintext or
793 	 * entirely ciphertext. Otherwise handle reencrypt partially decrypted
794 	 * record.
795 	 */
796 	if (is_decrypted) {
797 		ctx->resync_nh_reset = 1;
798 		return 0;
799 	}
800 	if (is_encrypted) {
801 		tls_device_core_ctrl_rx_resync(tls_ctx, ctx, sk, skb);
802 		return 0;
803 	}
804 
805 	ctx->resync_nh_reset = 1;
806 	return tls_device_reencrypt(sk, skb);
807 }
808 
809 static void tls_device_attach(struct tls_context *ctx, struct sock *sk,
810 			      struct net_device *netdev)
811 {
812 	if (sk->sk_destruct != tls_device_sk_destruct) {
813 		refcount_set(&ctx->refcount, 1);
814 		dev_hold(netdev);
815 		ctx->netdev = netdev;
816 		spin_lock_irq(&tls_device_lock);
817 		list_add_tail(&ctx->list, &tls_device_list);
818 		spin_unlock_irq(&tls_device_lock);
819 
820 		ctx->sk_destruct = sk->sk_destruct;
821 		sk->sk_destruct = tls_device_sk_destruct;
822 	}
823 }
824 
825 int tls_set_device_offload(struct sock *sk, struct tls_context *ctx)
826 {
827 	u16 nonce_size, tag_size, iv_size, rec_seq_size;
828 	struct tls_context *tls_ctx = tls_get_ctx(sk);
829 	struct tls_prot_info *prot = &tls_ctx->prot_info;
830 	struct tls_record_info *start_marker_record;
831 	struct tls_offload_context_tx *offload_ctx;
832 	struct tls_crypto_info *crypto_info;
833 	struct net_device *netdev;
834 	char *iv, *rec_seq;
835 	struct sk_buff *skb;
836 	int rc = -EINVAL;
837 	__be64 rcd_sn;
838 
839 	if (!ctx)
840 		goto out;
841 
842 	if (ctx->priv_ctx_tx) {
843 		rc = -EEXIST;
844 		goto out;
845 	}
846 
847 	start_marker_record = kmalloc(sizeof(*start_marker_record), GFP_KERNEL);
848 	if (!start_marker_record) {
849 		rc = -ENOMEM;
850 		goto out;
851 	}
852 
853 	offload_ctx = kzalloc(TLS_OFFLOAD_CONTEXT_SIZE_TX, GFP_KERNEL);
854 	if (!offload_ctx) {
855 		rc = -ENOMEM;
856 		goto free_marker_record;
857 	}
858 
859 	crypto_info = &ctx->crypto_send.info;
860 	if (crypto_info->version != TLS_1_2_VERSION) {
861 		rc = -EOPNOTSUPP;
862 		goto free_offload_ctx;
863 	}
864 
865 	switch (crypto_info->cipher_type) {
866 	case TLS_CIPHER_AES_GCM_128:
867 		nonce_size = TLS_CIPHER_AES_GCM_128_IV_SIZE;
868 		tag_size = TLS_CIPHER_AES_GCM_128_TAG_SIZE;
869 		iv_size = TLS_CIPHER_AES_GCM_128_IV_SIZE;
870 		iv = ((struct tls12_crypto_info_aes_gcm_128 *)crypto_info)->iv;
871 		rec_seq_size = TLS_CIPHER_AES_GCM_128_REC_SEQ_SIZE;
872 		rec_seq =
873 		 ((struct tls12_crypto_info_aes_gcm_128 *)crypto_info)->rec_seq;
874 		break;
875 	default:
876 		rc = -EINVAL;
877 		goto free_offload_ctx;
878 	}
879 
880 	/* Sanity-check the rec_seq_size for stack allocations */
881 	if (rec_seq_size > TLS_MAX_REC_SEQ_SIZE) {
882 		rc = -EINVAL;
883 		goto free_offload_ctx;
884 	}
885 
886 	prot->version = crypto_info->version;
887 	prot->cipher_type = crypto_info->cipher_type;
888 	prot->prepend_size = TLS_HEADER_SIZE + nonce_size;
889 	prot->tag_size = tag_size;
890 	prot->overhead_size = prot->prepend_size + prot->tag_size;
891 	prot->iv_size = iv_size;
892 	ctx->tx.iv = kmalloc(iv_size + TLS_CIPHER_AES_GCM_128_SALT_SIZE,
893 			     GFP_KERNEL);
894 	if (!ctx->tx.iv) {
895 		rc = -ENOMEM;
896 		goto free_offload_ctx;
897 	}
898 
899 	memcpy(ctx->tx.iv + TLS_CIPHER_AES_GCM_128_SALT_SIZE, iv, iv_size);
900 
901 	prot->rec_seq_size = rec_seq_size;
902 	ctx->tx.rec_seq = kmemdup(rec_seq, rec_seq_size, GFP_KERNEL);
903 	if (!ctx->tx.rec_seq) {
904 		rc = -ENOMEM;
905 		goto free_iv;
906 	}
907 
908 	rc = tls_sw_fallback_init(sk, offload_ctx, crypto_info);
909 	if (rc)
910 		goto free_rec_seq;
911 
912 	/* start at rec_seq - 1 to account for the start marker record */
913 	memcpy(&rcd_sn, ctx->tx.rec_seq, sizeof(rcd_sn));
914 	offload_ctx->unacked_record_sn = be64_to_cpu(rcd_sn) - 1;
915 
916 	start_marker_record->end_seq = tcp_sk(sk)->write_seq;
917 	start_marker_record->len = 0;
918 	start_marker_record->num_frags = 0;
919 
920 	INIT_LIST_HEAD(&offload_ctx->records_list);
921 	list_add_tail(&start_marker_record->list, &offload_ctx->records_list);
922 	spin_lock_init(&offload_ctx->lock);
923 	sg_init_table(offload_ctx->sg_tx_data,
924 		      ARRAY_SIZE(offload_ctx->sg_tx_data));
925 
926 	clean_acked_data_enable(inet_csk(sk), &tls_icsk_clean_acked);
927 	ctx->push_pending_record = tls_device_push_pending_record;
928 
929 	/* TLS offload is greatly simplified if we don't send
930 	 * SKBs where only part of the payload needs to be encrypted.
931 	 * So mark the last skb in the write queue as end of record.
932 	 */
933 	skb = tcp_write_queue_tail(sk);
934 	if (skb)
935 		TCP_SKB_CB(skb)->eor = 1;
936 
937 	/* We support starting offload on multiple sockets
938 	 * concurrently, so we only need a read lock here.
939 	 * This lock must precede get_netdev_for_sock to prevent races between
940 	 * NETDEV_DOWN and setsockopt.
941 	 */
942 	down_read(&device_offload_lock);
943 	netdev = get_netdev_for_sock(sk);
944 	if (!netdev) {
945 		pr_err_ratelimited("%s: netdev not found\n", __func__);
946 		rc = -EINVAL;
947 		goto release_lock;
948 	}
949 
950 	if (!(netdev->features & NETIF_F_HW_TLS_TX)) {
951 		rc = -ENOTSUPP;
952 		goto release_netdev;
953 	}
954 
955 	/* Avoid offloading if the device is down
956 	 * We don't want to offload new flows after
957 	 * the NETDEV_DOWN event
958 	 */
959 	if (!(netdev->flags & IFF_UP)) {
960 		rc = -EINVAL;
961 		goto release_netdev;
962 	}
963 
964 	ctx->priv_ctx_tx = offload_ctx;
965 	rc = netdev->tlsdev_ops->tls_dev_add(netdev, sk, TLS_OFFLOAD_CTX_DIR_TX,
966 					     &ctx->crypto_send.info,
967 					     tcp_sk(sk)->write_seq);
968 	if (rc)
969 		goto release_netdev;
970 
971 	tls_device_attach(ctx, sk, netdev);
972 
973 	/* following this assignment tls_is_sk_tx_device_offloaded
974 	 * will return true and the context might be accessed
975 	 * by the netdev's xmit function.
976 	 */
977 	smp_store_release(&sk->sk_validate_xmit_skb, tls_validate_xmit_skb);
978 	dev_put(netdev);
979 	up_read(&device_offload_lock);
980 	goto out;
981 
982 release_netdev:
983 	dev_put(netdev);
984 release_lock:
985 	up_read(&device_offload_lock);
986 	clean_acked_data_disable(inet_csk(sk));
987 	crypto_free_aead(offload_ctx->aead_send);
988 free_rec_seq:
989 	kfree(ctx->tx.rec_seq);
990 free_iv:
991 	kfree(ctx->tx.iv);
992 free_offload_ctx:
993 	kfree(offload_ctx);
994 	ctx->priv_ctx_tx = NULL;
995 free_marker_record:
996 	kfree(start_marker_record);
997 out:
998 	return rc;
999 }
1000 
1001 int tls_set_device_offload_rx(struct sock *sk, struct tls_context *ctx)
1002 {
1003 	struct tls_offload_context_rx *context;
1004 	struct net_device *netdev;
1005 	int rc = 0;
1006 
1007 	if (ctx->crypto_recv.info.version != TLS_1_2_VERSION)
1008 		return -EOPNOTSUPP;
1009 
1010 	/* We support starting offload on multiple sockets
1011 	 * concurrently, so we only need a read lock here.
1012 	 * This lock must precede get_netdev_for_sock to prevent races between
1013 	 * NETDEV_DOWN and setsockopt.
1014 	 */
1015 	down_read(&device_offload_lock);
1016 	netdev = get_netdev_for_sock(sk);
1017 	if (!netdev) {
1018 		pr_err_ratelimited("%s: netdev not found\n", __func__);
1019 		rc = -EINVAL;
1020 		goto release_lock;
1021 	}
1022 
1023 	if (!(netdev->features & NETIF_F_HW_TLS_RX)) {
1024 		rc = -ENOTSUPP;
1025 		goto release_netdev;
1026 	}
1027 
1028 	/* Avoid offloading if the device is down
1029 	 * We don't want to offload new flows after
1030 	 * the NETDEV_DOWN event
1031 	 */
1032 	if (!(netdev->flags & IFF_UP)) {
1033 		rc = -EINVAL;
1034 		goto release_netdev;
1035 	}
1036 
1037 	context = kzalloc(TLS_OFFLOAD_CONTEXT_SIZE_RX, GFP_KERNEL);
1038 	if (!context) {
1039 		rc = -ENOMEM;
1040 		goto release_netdev;
1041 	}
1042 	context->resync_nh_reset = 1;
1043 
1044 	ctx->priv_ctx_rx = context;
1045 	rc = tls_set_sw_offload(sk, ctx, 0);
1046 	if (rc)
1047 		goto release_ctx;
1048 
1049 	rc = netdev->tlsdev_ops->tls_dev_add(netdev, sk, TLS_OFFLOAD_CTX_DIR_RX,
1050 					     &ctx->crypto_recv.info,
1051 					     tcp_sk(sk)->copied_seq);
1052 	if (rc)
1053 		goto free_sw_resources;
1054 
1055 	tls_device_attach(ctx, sk, netdev);
1056 	goto release_netdev;
1057 
1058 free_sw_resources:
1059 	up_read(&device_offload_lock);
1060 	tls_sw_free_resources_rx(sk);
1061 	down_read(&device_offload_lock);
1062 release_ctx:
1063 	ctx->priv_ctx_rx = NULL;
1064 release_netdev:
1065 	dev_put(netdev);
1066 release_lock:
1067 	up_read(&device_offload_lock);
1068 	return rc;
1069 }
1070 
1071 void tls_device_offload_cleanup_rx(struct sock *sk)
1072 {
1073 	struct tls_context *tls_ctx = tls_get_ctx(sk);
1074 	struct net_device *netdev;
1075 
1076 	down_read(&device_offload_lock);
1077 	netdev = tls_ctx->netdev;
1078 	if (!netdev)
1079 		goto out;
1080 
1081 	netdev->tlsdev_ops->tls_dev_del(netdev, tls_ctx,
1082 					TLS_OFFLOAD_CTX_DIR_RX);
1083 
1084 	if (tls_ctx->tx_conf != TLS_HW) {
1085 		dev_put(netdev);
1086 		tls_ctx->netdev = NULL;
1087 	}
1088 out:
1089 	up_read(&device_offload_lock);
1090 	tls_sw_release_resources_rx(sk);
1091 }
1092 
1093 static int tls_device_down(struct net_device *netdev)
1094 {
1095 	struct tls_context *ctx, *tmp;
1096 	unsigned long flags;
1097 	LIST_HEAD(list);
1098 
1099 	/* Request a write lock to block new offload attempts */
1100 	down_write(&device_offload_lock);
1101 
1102 	spin_lock_irqsave(&tls_device_lock, flags);
1103 	list_for_each_entry_safe(ctx, tmp, &tls_device_list, list) {
1104 		if (ctx->netdev != netdev ||
1105 		    !refcount_inc_not_zero(&ctx->refcount))
1106 			continue;
1107 
1108 		list_move(&ctx->list, &list);
1109 	}
1110 	spin_unlock_irqrestore(&tls_device_lock, flags);
1111 
1112 	list_for_each_entry_safe(ctx, tmp, &list, list)	{
1113 		if (ctx->tx_conf == TLS_HW)
1114 			netdev->tlsdev_ops->tls_dev_del(netdev, ctx,
1115 							TLS_OFFLOAD_CTX_DIR_TX);
1116 		if (ctx->rx_conf == TLS_HW)
1117 			netdev->tlsdev_ops->tls_dev_del(netdev, ctx,
1118 							TLS_OFFLOAD_CTX_DIR_RX);
1119 		WRITE_ONCE(ctx->netdev, NULL);
1120 		smp_mb__before_atomic(); /* pairs with test_and_set_bit() */
1121 		while (test_bit(TLS_RX_SYNC_RUNNING, &ctx->flags))
1122 			usleep_range(10, 200);
1123 		dev_put(netdev);
1124 		list_del_init(&ctx->list);
1125 
1126 		if (refcount_dec_and_test(&ctx->refcount))
1127 			tls_device_free_ctx(ctx);
1128 	}
1129 
1130 	up_write(&device_offload_lock);
1131 
1132 	flush_work(&tls_device_gc_work);
1133 
1134 	return NOTIFY_DONE;
1135 }
1136 
1137 static int tls_dev_event(struct notifier_block *this, unsigned long event,
1138 			 void *ptr)
1139 {
1140 	struct net_device *dev = netdev_notifier_info_to_dev(ptr);
1141 
1142 	if (!dev->tlsdev_ops &&
1143 	    !(dev->features & (NETIF_F_HW_TLS_RX | NETIF_F_HW_TLS_TX)))
1144 		return NOTIFY_DONE;
1145 
1146 	switch (event) {
1147 	case NETDEV_REGISTER:
1148 	case NETDEV_FEAT_CHANGE:
1149 		if ((dev->features & NETIF_F_HW_TLS_RX) &&
1150 		    !dev->tlsdev_ops->tls_dev_resync)
1151 			return NOTIFY_BAD;
1152 
1153 		if  (dev->tlsdev_ops &&
1154 		     dev->tlsdev_ops->tls_dev_add &&
1155 		     dev->tlsdev_ops->tls_dev_del)
1156 			return NOTIFY_DONE;
1157 		else
1158 			return NOTIFY_BAD;
1159 	case NETDEV_DOWN:
1160 		return tls_device_down(dev);
1161 	}
1162 	return NOTIFY_DONE;
1163 }
1164 
1165 static struct notifier_block tls_dev_notifier = {
1166 	.notifier_call	= tls_dev_event,
1167 };
1168 
1169 void __init tls_device_init(void)
1170 {
1171 	register_netdevice_notifier(&tls_dev_notifier);
1172 }
1173 
1174 void __exit tls_device_cleanup(void)
1175 {
1176 	unregister_netdevice_notifier(&tls_dev_notifier);
1177 	flush_work(&tls_device_gc_work);
1178 	clean_acked_data_flush();
1179 }
1180