projects/linux·blob master · refs

1445 lines · 37.2 KiB

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#include <linux/skbuff_ref.h>
41
42#include "tls.h"
43#include "trace.h"
44
45/* device_offload_lock is used to synchronize tls_dev_add
46 * against NETDEV_DOWN notifications.
47 */
48static DECLARE_RWSEM(device_offload_lock);
49
50static struct workqueue_struct *destruct_wq __read_mostly;
51
52static LIST_HEAD(tls_device_list);
53static LIST_HEAD(tls_device_down_list);
54static DEFINE_SPINLOCK(tls_device_lock);
55
56static struct page *dummy_page;
57
58static void tls_device_free_ctx(struct tls_context *ctx)
59{
60 if (ctx->tx_conf == TLS_HW)
61 kfree(tls_offload_ctx_tx(ctx));
62
63 if (ctx->rx_conf == TLS_HW)
64 kfree(tls_offload_ctx_rx(ctx));
65
66 tls_ctx_free(NULL, ctx);
67}
68
69static void tls_device_tx_del_task(struct work_struct *work)
70{
71 struct tls_offload_context_tx *offload_ctx =
72 container_of(work, struct tls_offload_context_tx, destruct_work);
73 struct tls_context *ctx = offload_ctx->ctx;
74 struct net_device *netdev;
75
76 /* Safe, because this is the destroy flow, refcount is 0, so
77 * tls_device_down can't store this field in parallel.
78 */
79 netdev = rcu_dereference_protected(ctx->netdev,
80 !refcount_read(&ctx->refcount));
81
82 netdev->tlsdev_ops->tls_dev_del(netdev, ctx, TLS_OFFLOAD_CTX_DIR_TX);
83 dev_put(netdev);
84 ctx->netdev = NULL;
85 tls_device_free_ctx(ctx);
86}
87
88static void tls_device_queue_ctx_destruction(struct tls_context *ctx)
89{
90 struct net_device *netdev;
91 unsigned long flags;
92 bool async_cleanup;
93
94 spin_lock_irqsave(&tls_device_lock, flags);
95 if (unlikely(!refcount_dec_and_test(&ctx->refcount))) {
96 spin_unlock_irqrestore(&tls_device_lock, flags);
97 return;
98 }
99
100 list_del(&ctx->list); /* Remove from tls_device_list / tls_device_down_list */
101
102 /* Safe, because this is the destroy flow, refcount is 0, so
103 * tls_device_down can't store this field in parallel.
104 */
105 netdev = rcu_dereference_protected(ctx->netdev,
106 !refcount_read(&ctx->refcount));
107
108 async_cleanup = netdev && ctx->tx_conf == TLS_HW;
109 if (async_cleanup) {
110 struct tls_offload_context_tx *offload_ctx = tls_offload_ctx_tx(ctx);
111
112 /* queue_work inside the spinlock
113 * to make sure tls_device_down waits for that work.
114 */
115 queue_work(destruct_wq, &offload_ctx->destruct_work);
116 }
117 spin_unlock_irqrestore(&tls_device_lock, flags);
118
119 if (!async_cleanup)
120 tls_device_free_ctx(ctx);
121}
122
123/* We assume that the socket is already connected */
124static struct net_device *get_netdev_for_sock(struct sock *sk)
125{
126 struct net_device *dev, *lowest_dev = NULL;
127 struct dst_entry *dst;
128
129 rcu_read_lock();
130 dst = __sk_dst_get(sk);
131 dev = dst ? dst_dev_rcu(dst) : NULL;
132 if (likely(dev)) {
133 lowest_dev = netdev_sk_get_lowest_dev(dev, sk);
134 dev_hold(lowest_dev);
135 }
136 rcu_read_unlock();
137
138 return lowest_dev;
139}
140
141static void destroy_record(struct tls_record_info *record)
142{
143 int i;
144
145 for (i = 0; i < record->num_frags; i++)
146 __skb_frag_unref(&record->frags[i], false);
147 kfree(record);
148}
149
150static void delete_all_records(struct tls_offload_context_tx *offload_ctx)
151{
152 struct tls_record_info *info, *temp;
153
154 list_for_each_entry_safe(info, temp, &offload_ctx->records_list, list) {
155 list_del(&info->list);
156 destroy_record(info);
157 }
158
159 offload_ctx->retransmit_hint = NULL;
160}
161
162static void tls_tcp_clean_acked(struct sock *sk, u32 acked_seq)
163{
164 struct tls_context *tls_ctx = tls_get_ctx(sk);
165 struct tls_record_info *info, *temp;
166 struct tls_offload_context_tx *ctx;
167 u64 deleted_records = 0;
168 unsigned long flags;
169
170 if (!tls_ctx)
171 return;
172
173 ctx = tls_offload_ctx_tx(tls_ctx);
174
175 spin_lock_irqsave(&ctx->lock, flags);
176 info = ctx->retransmit_hint;
177 if (info && !before(acked_seq, info->end_seq))
178 ctx->retransmit_hint = NULL;
179
180 list_for_each_entry_safe(info, temp, &ctx->records_list, list) {
181 if (before(acked_seq, info->end_seq))
182 break;
183 list_del(&info->list);
184
185 destroy_record(info);
186 deleted_records++;
187 }
188
189 ctx->unacked_record_sn += deleted_records;
190 spin_unlock_irqrestore(&ctx->lock, flags);
191}
192
193/* At this point, there should be no references on this
194 * socket and no in-flight SKBs associated with this
195 * socket, so it is safe to free all the resources.
196 */
197void tls_device_sk_destruct(struct sock *sk)
198{
199 struct tls_context *tls_ctx = tls_get_ctx(sk);
200 struct tls_offload_context_tx *ctx = tls_offload_ctx_tx(tls_ctx);
201
202 tls_ctx->sk_destruct(sk);
203
204 if (tls_ctx->tx_conf == TLS_HW) {
205 if (ctx->open_record)
206 destroy_record(ctx->open_record);
207 delete_all_records(ctx);
208 crypto_free_aead(ctx->aead_send);
209 clean_acked_data_disable(tcp_sk(sk));
210 }
211
212 tls_device_queue_ctx_destruction(tls_ctx);
213}
214EXPORT_SYMBOL_GPL(tls_device_sk_destruct);
215
216void tls_device_free_resources_tx(struct sock *sk)
217{
218 struct tls_context *tls_ctx = tls_get_ctx(sk);
219
220 tls_free_partial_record(sk, tls_ctx);
221}
222
223void tls_offload_tx_resync_request(struct sock *sk, u32 got_seq, u32 exp_seq)
224{
225 struct tls_context *tls_ctx = tls_get_ctx(sk);
226
227 trace_tls_device_tx_resync_req(sk, got_seq, exp_seq);
228 WARN_ON(test_and_set_bit(TLS_TX_SYNC_SCHED, &tls_ctx->flags));
229}
230EXPORT_SYMBOL_GPL(tls_offload_tx_resync_request);
231
232static void tls_device_resync_tx(struct sock *sk, struct tls_context *tls_ctx,
233 u32 seq)
234{
235 struct net_device *netdev;
236 int err = 0;
237 u8 *rcd_sn;
238
239 tcp_write_collapse_fence(sk);
240 rcd_sn = tls_ctx->tx.rec_seq;
241
242 trace_tls_device_tx_resync_send(sk, seq, rcd_sn);
243 down_read(&device_offload_lock);
244 netdev = rcu_dereference_protected(tls_ctx->netdev,
245 lockdep_is_held(&device_offload_lock));
246 if (netdev)
247 err = netdev->tlsdev_ops->tls_dev_resync(netdev, sk, seq,
248 rcd_sn,
249 TLS_OFFLOAD_CTX_DIR_TX);
250 up_read(&device_offload_lock);
251 if (err)
252 return;
253
254 clear_bit_unlock(TLS_TX_SYNC_SCHED, &tls_ctx->flags);
255}
256
257static void tls_append_frag(struct tls_record_info *record,
258 struct page_frag *pfrag,
259 int size)
260{
261 skb_frag_t *frag;
262
263 frag = &record->frags[record->num_frags - 1];
264 if (skb_frag_page(frag) == pfrag->page &&
265 skb_frag_off(frag) + skb_frag_size(frag) == pfrag->offset) {
266 skb_frag_size_add(frag, size);
267 } else {
268 ++frag;
269 skb_frag_fill_page_desc(frag, pfrag->page, pfrag->offset,
270 size);
271 ++record->num_frags;
272 get_page(pfrag->page);
273 }
274
275 pfrag->offset += size;
276 record->len += size;
277}
278
279static int tls_push_record(struct sock *sk,
280 struct tls_context *ctx,
281 struct tls_offload_context_tx *offload_ctx,
282 struct tls_record_info *record,
283 int flags)
284{
285 struct tls_prot_info *prot = &ctx->prot_info;
286 struct tcp_sock *tp = tcp_sk(sk);
287 skb_frag_t *frag;
288 int i;
289
290 record->end_seq = tp->write_seq + record->len;
291 list_add_tail_rcu(&record->list, &offload_ctx->records_list);
292 offload_ctx->open_record = NULL;
293
294 if (test_bit(TLS_TX_SYNC_SCHED, &ctx->flags))
295 tls_device_resync_tx(sk, ctx, tp->write_seq);
296
297 tls_advance_record_sn(sk, prot, &ctx->tx);
298
299 for (i = 0; i < record->num_frags; i++) {
300 frag = &record->frags[i];
301 sg_unmark_end(&offload_ctx->sg_tx_data[i]);
302 sg_set_page(&offload_ctx->sg_tx_data[i], skb_frag_page(frag),
303 skb_frag_size(frag), skb_frag_off(frag));
304 sk_mem_charge(sk, skb_frag_size(frag));
305 get_page(skb_frag_page(frag));
306 }
307 sg_mark_end(&offload_ctx->sg_tx_data[record->num_frags - 1]);
308
309 /* all ready, send */
310 return tls_push_sg(sk, ctx, offload_ctx->sg_tx_data, 0, flags);
311}
312
313static void tls_device_record_close(struct sock *sk,
314 struct tls_context *ctx,
315 struct tls_record_info *record,
316 struct page_frag *pfrag,
317 unsigned char record_type)
318{
319 struct tls_prot_info *prot = &ctx->prot_info;
320 struct page_frag dummy_tag_frag;
321
322 /* append tag
323 * device will fill in the tag, we just need to append a placeholder
324 * use socket memory to improve coalescing (re-using a single buffer
325 * increases frag count)
326 * if we can't allocate memory now use the dummy page
327 */
328 if (unlikely(pfrag->size - pfrag->offset < prot->tag_size) &&
329 !skb_page_frag_refill(prot->tag_size, pfrag, sk->sk_allocation)) {
330 dummy_tag_frag.page = dummy_page;
331 dummy_tag_frag.offset = 0;
332 pfrag = &dummy_tag_frag;
333 }
334 tls_append_frag(record, pfrag, prot->tag_size);
335
336 /* fill prepend */
337 tls_fill_prepend(ctx, skb_frag_address(&record->frags[0]),
338 record->len - prot->overhead_size,
339 record_type);
340}
341
342static int tls_create_new_record(struct tls_offload_context_tx *offload_ctx,
343 struct page_frag *pfrag,
344 size_t prepend_size)
345{
346 struct tls_record_info *record;
347 skb_frag_t *frag;
348
349 record = kmalloc_obj(*record);
350 if (!record)
351 return -ENOMEM;
352
353 frag = &record->frags[0];
354 skb_frag_fill_page_desc(frag, pfrag->page, pfrag->offset,
355 prepend_size);
356
357 get_page(pfrag->page);
358 pfrag->offset += prepend_size;
359
360 record->num_frags = 1;
361 record->len = prepend_size;
362 offload_ctx->open_record = record;
363 return 0;
364}
365
366static int tls_do_allocation(struct sock *sk,
367 struct tls_offload_context_tx *offload_ctx,
368 struct page_frag *pfrag,
369 size_t prepend_size)
370{
371 int ret;
372
373 if (!offload_ctx->open_record) {
374 if (unlikely(!skb_page_frag_refill(prepend_size, pfrag,
375 sk->sk_allocation))) {
376 if (!sk->sk_bypass_prot_mem)
377 READ_ONCE(sk->sk_prot)->enter_memory_pressure(sk);
378 sk_stream_moderate_sndbuf(sk);
379 return -ENOMEM;
380 }
381
382 ret = tls_create_new_record(offload_ctx, pfrag, prepend_size);
383 if (ret)
384 return ret;
385
386 if (pfrag->size > pfrag->offset)
387 return 0;
388 }
389
390 if (!sk_page_frag_refill(sk, pfrag))
391 return -ENOMEM;
392
393 return 0;
394}
395
396static int tls_device_copy_data(void *addr, size_t bytes, struct iov_iter *i)
397{
398 size_t pre_copy, nocache;
399
400 pre_copy = ~((unsigned long)addr - 1) & (SMP_CACHE_BYTES - 1);
401 if (pre_copy) {
402 pre_copy = min(pre_copy, bytes);
403 if (copy_from_iter(addr, pre_copy, i) != pre_copy)
404 return -EFAULT;
405 bytes -= pre_copy;
406 addr += pre_copy;
407 }
408
409 nocache = round_down(bytes, SMP_CACHE_BYTES);
410 if (copy_from_iter_nocache(addr, nocache, i) != nocache)
411 return -EFAULT;
412 bytes -= nocache;
413 addr += nocache;
414
415 if (bytes && copy_from_iter(addr, bytes, i) != bytes)
416 return -EFAULT;
417
418 return 0;
419}
420
421static int tls_push_data(struct sock *sk,
422 struct iov_iter *iter,
423 size_t size, int flags,
424 unsigned char record_type)
425{
426 struct tls_context *tls_ctx = tls_get_ctx(sk);
427 struct tls_prot_info *prot = &tls_ctx->prot_info;
428 struct tls_offload_context_tx *ctx = tls_offload_ctx_tx(tls_ctx);
429 struct tls_record_info *record;
430 int tls_push_record_flags;
431 struct page_frag *pfrag;
432 size_t orig_size = size;
433 u32 max_open_record_len;
434 bool more = false;
435 bool done = false;
436 int copy, rc = 0;
437 long timeo;
438
439 if (flags &
440 ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL |
441 MSG_SPLICE_PAGES | MSG_EOR))
442 return -EOPNOTSUPP;
443
444 if ((flags & (MSG_MORE | MSG_EOR)) == (MSG_MORE | MSG_EOR))
445 return -EINVAL;
446
447 if (unlikely(sk->sk_err))
448 return -sk->sk_err;
449
450 flags |= MSG_SENDPAGE_DECRYPTED;
451 tls_push_record_flags = flags | MSG_MORE;
452
453 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
454 if (tls_is_partially_sent_record(tls_ctx)) {
455 rc = tls_push_partial_record(sk, tls_ctx, flags);
456 if (rc < 0)
457 return rc;
458 }
459
460 pfrag = sk_page_frag(sk);
461
462 /* TLS_HEADER_SIZE is not counted as part of the TLS record, and
463 * we need to leave room for an authentication tag.
464 */
465 max_open_record_len = tls_ctx->tx_max_payload_len +
466 prot->prepend_size;
467 do {
468 rc = tls_do_allocation(sk, ctx, pfrag, prot->prepend_size);
469 if (unlikely(rc)) {
470 rc = sk_stream_wait_memory(sk, &timeo);
471 if (!rc)
472 continue;
473
474 record = ctx->open_record;
475 if (!record)
476 break;
477handle_error:
478 if (record_type != TLS_RECORD_TYPE_DATA) {
479 /* avoid sending partial
480 * record with type !=
481 * application_data
482 */
483 size = orig_size;
484 destroy_record(record);
485 ctx->open_record = NULL;
486 } else if (record->len > prot->prepend_size) {
487 goto last_record;
488 }
489
490 break;
491 }
492
493 record = ctx->open_record;
494
495 copy = min_t(size_t, size, max_open_record_len - record->len);
496 if (copy && (flags & MSG_SPLICE_PAGES)) {
497 struct page_frag zc_pfrag;
498 struct page **pages = &zc_pfrag.page;
499 size_t off;
500
501 rc = iov_iter_extract_pages(iter, &pages,
502 copy, 1, 0, &off);
503 if (rc <= 0) {
504 if (rc == 0)
505 rc = -EIO;
506 goto handle_error;
507 }
508 copy = rc;
509
510 if (WARN_ON_ONCE(!sendpage_ok(zc_pfrag.page))) {
511 iov_iter_revert(iter, copy);
512 rc = -EIO;
513 goto handle_error;
514 }
515
516 zc_pfrag.offset = off;
517 zc_pfrag.size = copy;
518 tls_append_frag(record, &zc_pfrag, copy);
519 } else if (copy) {
520 copy = min_t(size_t, copy, pfrag->size - pfrag->offset);
521
522 rc = tls_device_copy_data(page_address(pfrag->page) +
523 pfrag->offset, copy,
524 iter);
525 if (rc)
526 goto handle_error;
527 tls_append_frag(record, pfrag, copy);
528 }
529
530 size -= copy;
531 if (!size) {
532last_record:
533 tls_push_record_flags = flags;
534 if ((flags & MSG_MORE) &&
535 record->num_frags < MAX_SKB_FRAGS - 1) {
536 more = true;
537 break;
538 }
539
540 done = true;
541 }
542
543 if (done || record->len >= max_open_record_len ||
544 (record->num_frags >= MAX_SKB_FRAGS - 1)) {
545 tls_device_record_close(sk, tls_ctx, record,
546 pfrag, record_type);
547
548 rc = tls_push_record(sk,
549 tls_ctx,
550 ctx,
551 record,
552 tls_push_record_flags);
553 if (rc < 0)
554 break;
555 }
556 } while (!done);
557
558 tls_ctx->pending_open_record_frags = more;
559
560 if (orig_size - size > 0)
561 rc = orig_size - size;
562
563 return rc;
564}
565
566int tls_device_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
567{
568 unsigned char record_type = TLS_RECORD_TYPE_DATA;
569 struct tls_context *tls_ctx = tls_get_ctx(sk);
570 int rc;
571
572 if (!tls_ctx->zerocopy_sendfile)
573 msg->msg_flags &= ~MSG_SPLICE_PAGES;
574
575 mutex_lock(&tls_ctx->tx_lock);
576 lock_sock(sk);
577
578 if (unlikely(msg->msg_controllen)) {
579 rc = tls_process_cmsg(sk, msg, &record_type);
580 if (rc)
581 goto out;
582 }
583
584 rc = tls_push_data(sk, &msg->msg_iter, size, msg->msg_flags,
585 record_type);
586
587out:
588 release_sock(sk);
589 mutex_unlock(&tls_ctx->tx_lock);
590 return rc;
591}
592
593void tls_device_splice_eof(struct socket *sock)
594{
595 struct sock *sk = sock->sk;
596 struct tls_context *tls_ctx = tls_get_ctx(sk);
597 struct iov_iter iter = {};
598
599 if (!tls_is_partially_sent_record(tls_ctx) &&
600 !tls_is_pending_open_record(tls_ctx))
601 return;
602
603 mutex_lock(&tls_ctx->tx_lock);
604 lock_sock(sk);
605
606 if (tls_is_partially_sent_record(tls_ctx) ||
607 tls_is_pending_open_record(tls_ctx)) {
608 iov_iter_bvec(&iter, ITER_SOURCE, NULL, 0, 0);
609 tls_push_data(sk, &iter, 0, 0, TLS_RECORD_TYPE_DATA);
610 }
611
612 release_sock(sk);
613 mutex_unlock(&tls_ctx->tx_lock);
614}
615
616struct tls_record_info *tls_get_record(struct tls_offload_context_tx *context,
617 u32 seq, u64 *p_record_sn)
618{
619 u64 record_sn = context->hint_record_sn;
620 struct tls_record_info *info, *last;
621
622 info = context->retransmit_hint;
623 if (!info ||
624 before(seq, info->end_seq - info->len)) {
625 /* if retransmit_hint is irrelevant start
626 * from the beginning of the list
627 */
628 info = list_first_entry_or_null(&context->records_list,
629 struct tls_record_info, list);
630 if (!info)
631 return NULL;
632 /* send the start_marker record if seq number is before the
633 * tls offload start marker sequence number. This record is
634 * required to handle TCP packets which are before TLS offload
635 * started.
636 * And if it's not start marker, look if this seq number
637 * belongs to the list.
638 */
639 if (likely(!tls_record_is_start_marker(info))) {
640 /* we have the first record, get the last record to see
641 * if this seq number belongs to the list.
642 */
643 last = list_last_entry(&context->records_list,
644 struct tls_record_info, list);
645
646 if (!between(seq, tls_record_start_seq(info),
647 last->end_seq))
648 return NULL;
649 }
650 record_sn = context->unacked_record_sn;
651 }
652
653 /* We just need the _rcu for the READ_ONCE() */
654 rcu_read_lock();
655 list_for_each_entry_from_rcu(info, &context->records_list, list) {
656 if (before(seq, info->end_seq)) {
657 if (!context->retransmit_hint ||
658 after(info->end_seq,
659 context->retransmit_hint->end_seq)) {
660 context->hint_record_sn = record_sn;
661 context->retransmit_hint = info;
662 }
663 *p_record_sn = record_sn;
664 goto exit_rcu_unlock;
665 }
666 record_sn++;
667 }
668 info = NULL;
669
670exit_rcu_unlock:
671 rcu_read_unlock();
672 return info;
673}
674EXPORT_SYMBOL(tls_get_record);
675
676static int tls_device_push_pending_record(struct sock *sk, int flags)
677{
678 struct iov_iter iter;
679
680 iov_iter_kvec(&iter, ITER_SOURCE, NULL, 0, 0);
681 return tls_push_data(sk, &iter, 0, flags, TLS_RECORD_TYPE_DATA);
682}
683
684void tls_device_write_space(struct sock *sk, struct tls_context *ctx)
685{
686 if (tls_is_partially_sent_record(ctx)) {
687 gfp_t sk_allocation = sk->sk_allocation;
688
689 WARN_ON_ONCE(sk->sk_write_pending);
690
691 sk->sk_allocation = GFP_ATOMIC;
692 tls_push_partial_record(sk, ctx,
693 MSG_DONTWAIT | MSG_NOSIGNAL |
694 MSG_SENDPAGE_DECRYPTED);
695 sk->sk_allocation = sk_allocation;
696 }
697}
698
699static void tls_device_resync_rx(struct tls_context *tls_ctx,
700 struct sock *sk, u32 seq, u8 *rcd_sn)
701{
702 struct tls_offload_context_rx *rx_ctx = tls_offload_ctx_rx(tls_ctx);
703 struct net_device *netdev;
704
705 trace_tls_device_rx_resync_send(sk, seq, rcd_sn, rx_ctx->resync_type);
706 rcu_read_lock();
707 netdev = rcu_dereference(tls_ctx->netdev);
708 if (netdev)
709 netdev->tlsdev_ops->tls_dev_resync(netdev, sk, seq, rcd_sn,
710 TLS_OFFLOAD_CTX_DIR_RX);
711 rcu_read_unlock();
712 TLS_INC_STATS(sock_net(sk), LINUX_MIB_TLSRXDEVICERESYNC);
713}
714
715static bool
716tls_device_rx_resync_async(struct tls_offload_resync_async *resync_async,
717 s64 resync_req, u32 *seq, u16 *rcd_delta)
718{
719 u32 is_async = resync_req & RESYNC_REQ_ASYNC;
720 u32 req_seq = resync_req >> 32;
721 u32 req_end = req_seq + ((resync_req >> 16) & 0xffff);
722 u16 i;
723
724 *rcd_delta = 0;
725
726 if (is_async) {
727 /* shouldn't get to wraparound:
728 * too long in async stage, something bad happened
729 */
730 if (WARN_ON_ONCE(resync_async->rcd_delta == USHRT_MAX)) {
731 tls_offload_rx_resync_async_request_cancel(resync_async);
732 return false;
733 }
734
735 /* asynchronous stage: log all headers seq such that
736 * req_seq <= seq <= end_seq, and wait for real resync request
737 */
738 if (before(*seq, req_seq))
739 return false;
740 if (!after(*seq, req_end) &&
741 resync_async->loglen < TLS_DEVICE_RESYNC_ASYNC_LOGMAX)
742 resync_async->log[resync_async->loglen++] = *seq;
743
744 resync_async->rcd_delta++;
745
746 return false;
747 }
748
749 /* synchronous stage: check against the logged entries and
750 * proceed to check the next entries if no match was found
751 */
752 for (i = 0; i < resync_async->loglen; i++)
753 if (req_seq == resync_async->log[i] &&
754 atomic64_try_cmpxchg(&resync_async->req, &resync_req, 0)) {
755 *rcd_delta = resync_async->rcd_delta - i;
756 *seq = req_seq;
757 resync_async->loglen = 0;
758 resync_async->rcd_delta = 0;
759 return true;
760 }
761
762 resync_async->loglen = 0;
763 resync_async->rcd_delta = 0;
764
765 if (req_seq == *seq &&
766 atomic64_try_cmpxchg(&resync_async->req,
767 &resync_req, 0))
768 return true;
769
770 return false;
771}
772
773void tls_device_rx_resync_new_rec(struct sock *sk, u32 rcd_len, u32 seq)
774{
775 struct tls_context *tls_ctx = tls_get_ctx(sk);
776 struct tls_offload_context_rx *rx_ctx;
777 u8 rcd_sn[TLS_MAX_REC_SEQ_SIZE];
778 u32 sock_data, is_req_pending;
779 struct tls_prot_info *prot;
780 s64 resync_req;
781 u16 rcd_delta;
782 u32 req_seq;
783
784 if (tls_ctx->rx_conf != TLS_HW)
785 return;
786 if (unlikely(test_bit(TLS_RX_DEV_DEGRADED, &tls_ctx->flags)))
787 return;
788
789 prot = &tls_ctx->prot_info;
790 rx_ctx = tls_offload_ctx_rx(tls_ctx);
791 memcpy(rcd_sn, tls_ctx->rx.rec_seq, prot->rec_seq_size);
792
793 switch (rx_ctx->resync_type) {
794 case TLS_OFFLOAD_SYNC_TYPE_DRIVER_REQ:
795 resync_req = atomic64_read(&rx_ctx->resync_req);
796 req_seq = resync_req >> 32;
797 seq += TLS_HEADER_SIZE - 1;
798 is_req_pending = resync_req;
799
800 if (likely(!is_req_pending) || req_seq != seq ||
801 !atomic64_try_cmpxchg(&rx_ctx->resync_req, &resync_req, 0))
802 return;
803 break;
804 case TLS_OFFLOAD_SYNC_TYPE_CORE_NEXT_HINT:
805 if (likely(!rx_ctx->resync_nh_do_now))
806 return;
807
808 /* head of next rec is already in, note that the sock_inq will
809 * include the currently parsed message when called from parser
810 */
811 sock_data = tcp_inq(sk);
812 if (sock_data > rcd_len) {
813 trace_tls_device_rx_resync_nh_delay(sk, sock_data,
814 rcd_len);
815 return;
816 }
817
818 rx_ctx->resync_nh_do_now = 0;
819 seq += rcd_len;
820 tls_bigint_increment(rcd_sn, prot->rec_seq_size);
821 break;
822 case TLS_OFFLOAD_SYNC_TYPE_DRIVER_REQ_ASYNC:
823 resync_req = atomic64_read(&rx_ctx->resync_async->req);
824 is_req_pending = resync_req;
825 if (likely(!is_req_pending))
826 return;
827
828 if (!tls_device_rx_resync_async(rx_ctx->resync_async,
829 resync_req, &seq, &rcd_delta))
830 return;
831 tls_bigint_subtract(rcd_sn, rcd_delta);
832 break;
833 }
834
835 tls_device_resync_rx(tls_ctx, sk, seq, rcd_sn);
836}
837
838static void tls_device_core_ctrl_rx_resync(struct tls_context *tls_ctx,
839 struct tls_offload_context_rx *ctx,
840 struct sock *sk, struct sk_buff *skb)
841{
842 struct strp_msg *rxm;
843
844 /* device will request resyncs by itself based on stream scan */
845 if (ctx->resync_type != TLS_OFFLOAD_SYNC_TYPE_CORE_NEXT_HINT)
846 return;
847 /* already scheduled */
848 if (ctx->resync_nh_do_now)
849 return;
850 /* seen decrypted fragments since last fully-failed record */
851 if (ctx->resync_nh_reset) {
852 ctx->resync_nh_reset = 0;
853 ctx->resync_nh.decrypted_failed = 1;
854 ctx->resync_nh.decrypted_tgt = TLS_DEVICE_RESYNC_NH_START_IVAL;
855 return;
856 }
857
858 if (++ctx->resync_nh.decrypted_failed <= ctx->resync_nh.decrypted_tgt)
859 return;
860
861 /* doing resync, bump the next target in case it fails */
862 if (ctx->resync_nh.decrypted_tgt < TLS_DEVICE_RESYNC_NH_MAX_IVAL)
863 ctx->resync_nh.decrypted_tgt *= 2;
864 else
865 ctx->resync_nh.decrypted_tgt += TLS_DEVICE_RESYNC_NH_MAX_IVAL;
866
867 rxm = strp_msg(skb);
868
869 /* head of next rec is already in, parser will sync for us */
870 if (tcp_inq(sk) > rxm->full_len) {
871 trace_tls_device_rx_resync_nh_schedule(sk);
872 ctx->resync_nh_do_now = 1;
873 } else {
874 struct tls_prot_info *prot = &tls_ctx->prot_info;
875 u8 rcd_sn[TLS_MAX_REC_SEQ_SIZE];
876
877 memcpy(rcd_sn, tls_ctx->rx.rec_seq, prot->rec_seq_size);
878 tls_bigint_increment(rcd_sn, prot->rec_seq_size);
879
880 tls_device_resync_rx(tls_ctx, sk, tcp_sk(sk)->copied_seq,
881 rcd_sn);
882 }
883}
884
885static int
886tls_device_reencrypt(struct sock *sk, struct tls_context *tls_ctx)
887{
888 struct tls_sw_context_rx *sw_ctx = tls_sw_ctx_rx(tls_ctx);
889 const struct tls_cipher_desc *cipher_desc;
890 int err, offset, copy, data_len, pos;
891 struct sk_buff *skb, *skb_iter;
892 struct scatterlist sg[1];
893 struct strp_msg *rxm;
894 char *orig_buf, *buf;
895
896 cipher_desc = get_cipher_desc(tls_ctx->crypto_recv.info.cipher_type);
897 DEBUG_NET_WARN_ON_ONCE(!cipher_desc || !cipher_desc->offloadable);
898
899 rxm = strp_msg(tls_strp_msg(sw_ctx));
900 orig_buf = kmalloc(rxm->full_len + TLS_HEADER_SIZE + cipher_desc->iv,
901 sk->sk_allocation);
902 if (!orig_buf)
903 return -ENOMEM;
904 buf = orig_buf;
905
906 err = tls_strp_msg_cow(sw_ctx);
907 if (unlikely(err))
908 goto free_buf;
909
910 skb = tls_strp_msg(sw_ctx);
911 rxm = strp_msg(skb);
912 offset = rxm->offset;
913
914 sg_init_table(sg, 1);
915 sg_set_buf(&sg[0], buf,
916 rxm->full_len + TLS_HEADER_SIZE + cipher_desc->iv);
917 err = skb_copy_bits(skb, offset, buf, TLS_HEADER_SIZE + cipher_desc->iv);
918 if (err)
919 goto free_buf;
920
921 /* We are interested only in the decrypted data not the auth */
922 err = decrypt_skb(sk, sg);
923 if (err != -EBADMSG)
924 goto free_buf;
925 else
926 err = 0;
927
928 data_len = rxm->full_len - cipher_desc->tag;
929
930 if (skb_pagelen(skb) > offset) {
931 copy = min_t(int, skb_pagelen(skb) - offset, data_len);
932
933 if (skb->decrypted) {
934 err = skb_store_bits(skb, offset, buf, copy);
935 if (err)
936 goto free_buf;
937 }
938
939 offset += copy;
940 buf += copy;
941 }
942
943 pos = skb_pagelen(skb);
944 skb_walk_frags(skb, skb_iter) {
945 int frag_pos;
946
947 /* Practically all frags must belong to msg if reencrypt
948 * is needed with current strparser and coalescing logic,
949 * but strparser may "get optimized", so let's be safe.
950 */
951 if (pos + skb_iter->len <= offset)
952 goto done_with_frag;
953 if (pos >= data_len + rxm->offset)
954 break;
955
956 frag_pos = offset - pos;
957 copy = min_t(int, skb_iter->len - frag_pos,
958 data_len + rxm->offset - offset);
959
960 if (skb_iter->decrypted) {
961 err = skb_store_bits(skb_iter, frag_pos, buf, copy);
962 if (err)
963 goto free_buf;
964 }
965
966 offset += copy;
967 buf += copy;
968done_with_frag:
969 pos += skb_iter->len;
970 }
971
972free_buf:
973 kfree(orig_buf);
974 return err;
975}
976
977int tls_device_decrypted(struct sock *sk, struct tls_context *tls_ctx)
978{
979 struct tls_offload_context_rx *ctx = tls_offload_ctx_rx(tls_ctx);
980 struct tls_sw_context_rx *sw_ctx = tls_sw_ctx_rx(tls_ctx);
981 struct sk_buff *skb = tls_strp_msg(sw_ctx);
982 struct strp_msg *rxm = strp_msg(skb);
983 int is_decrypted, is_encrypted;
984
985 if (!tls_strp_msg_mixed_decrypted(sw_ctx)) {
986 is_decrypted = skb->decrypted;
987 is_encrypted = !is_decrypted;
988 } else {
989 is_decrypted = 0;
990 is_encrypted = 0;
991 }
992
993 trace_tls_device_decrypted(sk, tcp_sk(sk)->copied_seq - rxm->full_len,
994 tls_ctx->rx.rec_seq, rxm->full_len,
995 is_encrypted, is_decrypted);
996
997 if (unlikely(test_bit(TLS_RX_DEV_DEGRADED, &tls_ctx->flags))) {
998 if (likely(is_encrypted || is_decrypted))
999 return is_decrypted;
1000
1001 /* After tls_device_down disables the offload, the next SKB will
1002 * likely have initial fragments decrypted, and final ones not
1003 * decrypted. We need to reencrypt that single SKB.
1004 */
1005 return tls_device_reencrypt(sk, tls_ctx);
1006 }
1007
1008 /* Return immediately if the record is either entirely plaintext or
1009 * entirely ciphertext. Otherwise handle reencrypt partially decrypted
1010 * record.
1011 */
1012 if (is_decrypted) {
1013 ctx->resync_nh_reset = 1;
1014 return is_decrypted;
1015 }
1016 if (is_encrypted) {
1017 tls_device_core_ctrl_rx_resync(tls_ctx, ctx, sk, skb);
1018 return 0;
1019 }
1020
1021 ctx->resync_nh_reset = 1;
1022 return tls_device_reencrypt(sk, tls_ctx);
1023}
1024
1025static void tls_device_attach(struct tls_context *ctx, struct sock *sk,
1026 struct net_device *netdev)
1027{
1028 if (sk->sk_destruct != tls_device_sk_destruct) {
1029 refcount_set(&ctx->refcount, 1);
1030 dev_hold(netdev);
1031 RCU_INIT_POINTER(ctx->netdev, netdev);
1032 spin_lock_irq(&tls_device_lock);
1033 list_add_tail(&ctx->list, &tls_device_list);
1034 spin_unlock_irq(&tls_device_lock);
1035
1036 ctx->sk_destruct = sk->sk_destruct;
1037 smp_store_release(&sk->sk_destruct, tls_device_sk_destruct);
1038 }
1039}
1040
1041static struct tls_offload_context_tx *alloc_offload_ctx_tx(struct tls_context *ctx)
1042{
1043 struct tls_offload_context_tx *offload_ctx;
1044 __be64 rcd_sn;
1045
1046 offload_ctx = kzalloc_obj(*offload_ctx);
1047 if (!offload_ctx)
1048 return NULL;
1049
1050 INIT_WORK(&offload_ctx->destruct_work, tls_device_tx_del_task);
1051 INIT_LIST_HEAD(&offload_ctx->records_list);
1052 spin_lock_init(&offload_ctx->lock);
1053 sg_init_table(offload_ctx->sg_tx_data,
1054 ARRAY_SIZE(offload_ctx->sg_tx_data));
1055
1056 /* start at rec_seq - 1 to account for the start marker record */
1057 memcpy(&rcd_sn, ctx->tx.rec_seq, sizeof(rcd_sn));
1058 offload_ctx->unacked_record_sn = be64_to_cpu(rcd_sn) - 1;
1059
1060 offload_ctx->ctx = ctx;
1061
1062 return offload_ctx;
1063}
1064
1065int tls_set_device_offload(struct sock *sk)
1066{
1067 struct tls_record_info *start_marker_record;
1068 struct tls_offload_context_tx *offload_ctx;
1069 const struct tls_cipher_desc *cipher_desc;
1070 struct tls_crypto_info *crypto_info;
1071 struct tls_prot_info *prot;
1072 struct net_device *netdev;
1073 struct tls_context *ctx;
1074 char *iv, *rec_seq;
1075 int rc;
1076
1077 ctx = tls_get_ctx(sk);
1078 prot = &ctx->prot_info;
1079
1080 if (ctx->priv_ctx_tx)
1081 return -EEXIST;
1082
1083 netdev = get_netdev_for_sock(sk);
1084 if (!netdev) {
1085 pr_err_ratelimited("%s: netdev not found\n", __func__);
1086 return -EINVAL;
1087 }
1088
1089 if (!(netdev->features & NETIF_F_HW_TLS_TX)) {
1090 rc = -EOPNOTSUPP;
1091 goto release_netdev;
1092 }
1093
1094 crypto_info = &ctx->crypto_send.info;
1095 if (crypto_info->version != TLS_1_2_VERSION) {
1096 rc = -EOPNOTSUPP;
1097 goto release_netdev;
1098 }
1099
1100 cipher_desc = get_cipher_desc(crypto_info->cipher_type);
1101 if (!cipher_desc || !cipher_desc->offloadable) {
1102 rc = -EINVAL;
1103 goto release_netdev;
1104 }
1105
1106 rc = init_prot_info(prot, crypto_info, cipher_desc);
1107 if (rc)
1108 goto release_netdev;
1109
1110 iv = crypto_info_iv(crypto_info, cipher_desc);
1111 rec_seq = crypto_info_rec_seq(crypto_info, cipher_desc);
1112
1113 memcpy(ctx->tx.iv + cipher_desc->salt, iv, cipher_desc->iv);
1114 memcpy(ctx->tx.rec_seq, rec_seq, cipher_desc->rec_seq);
1115
1116 start_marker_record = kmalloc_obj(*start_marker_record);
1117 if (!start_marker_record) {
1118 rc = -ENOMEM;
1119 goto release_netdev;
1120 }
1121
1122 offload_ctx = alloc_offload_ctx_tx(ctx);
1123 if (!offload_ctx) {
1124 rc = -ENOMEM;
1125 goto free_marker_record;
1126 }
1127
1128 rc = tls_sw_fallback_init(sk, offload_ctx, crypto_info);
1129 if (rc)
1130 goto free_offload_ctx;
1131
1132 start_marker_record->end_seq = tcp_sk(sk)->write_seq;
1133 start_marker_record->len = 0;
1134 start_marker_record->num_frags = 0;
1135 list_add_tail(&start_marker_record->list, &offload_ctx->records_list);
1136
1137 clean_acked_data_enable(tcp_sk(sk), &tls_tcp_clean_acked);
1138 ctx->push_pending_record = tls_device_push_pending_record;
1139
1140 /* TLS offload is greatly simplified if we don't send
1141 * SKBs where only part of the payload needs to be encrypted.
1142 * So mark the last skb in the write queue as end of record.
1143 */
1144 tcp_write_collapse_fence(sk);
1145
1146 /* Avoid offloading if the device is down
1147 * We don't want to offload new flows after
1148 * the NETDEV_DOWN event
1149 *
1150 * device_offload_lock is taken in tls_devices's NETDEV_DOWN
1151 * handler thus protecting from the device going down before
1152 * ctx was added to tls_device_list.
1153 */
1154 down_read(&device_offload_lock);
1155 if (!(netdev->flags & IFF_UP)) {
1156 rc = -EINVAL;
1157 goto release_lock;
1158 }
1159
1160 ctx->priv_ctx_tx = offload_ctx;
1161 rc = netdev->tlsdev_ops->tls_dev_add(netdev, sk, TLS_OFFLOAD_CTX_DIR_TX,
1162 &ctx->crypto_send.info,
1163 tcp_sk(sk)->write_seq);
1164 trace_tls_device_offload_set(sk, TLS_OFFLOAD_CTX_DIR_TX,
1165 tcp_sk(sk)->write_seq, rec_seq, rc);
1166 if (rc)
1167 goto release_lock;
1168
1169 tls_device_attach(ctx, sk, netdev);
1170 up_read(&device_offload_lock);
1171
1172 /* following this assignment tls_is_skb_tx_device_offloaded
1173 * will return true and the context might be accessed
1174 * by the netdev's xmit function.
1175 */
1176 smp_store_release(&sk->sk_validate_xmit_skb, tls_validate_xmit_skb);
1177 dev_put(netdev);
1178
1179 return 0;
1180
1181release_lock:
1182 up_read(&device_offload_lock);
1183 clean_acked_data_disable(tcp_sk(sk));
1184 crypto_free_aead(offload_ctx->aead_send);
1185free_offload_ctx:
1186 kfree(offload_ctx);
1187 ctx->priv_ctx_tx = NULL;
1188free_marker_record:
1189 kfree(start_marker_record);
1190release_netdev:
1191 dev_put(netdev);
1192 return rc;
1193}
1194
1195int tls_set_device_offload_rx(struct sock *sk, struct tls_context *ctx)
1196{
1197 struct tls12_crypto_info_aes_gcm_128 *info;
1198 struct tls_offload_context_rx *context;
1199 struct net_device *netdev;
1200 int rc = 0;
1201
1202 if (ctx->crypto_recv.info.version != TLS_1_2_VERSION)
1203 return -EOPNOTSUPP;
1204
1205 netdev = get_netdev_for_sock(sk);
1206 if (!netdev) {
1207 pr_err_ratelimited("%s: netdev not found\n", __func__);
1208 return -EINVAL;
1209 }
1210
1211 if (!(netdev->features & NETIF_F_HW_TLS_RX)) {
1212 rc = -EOPNOTSUPP;
1213 goto release_netdev;
1214 }
1215
1216 /* Avoid offloading if the device is down
1217 * We don't want to offload new flows after
1218 * the NETDEV_DOWN event
1219 *
1220 * device_offload_lock is taken in tls_devices's NETDEV_DOWN
1221 * handler thus protecting from the device going down before
1222 * ctx was added to tls_device_list.
1223 */
1224 down_read(&device_offload_lock);
1225 if (!(netdev->flags & IFF_UP)) {
1226 rc = -EINVAL;
1227 goto release_lock;
1228 }
1229
1230 context = kzalloc_obj(*context);
1231 if (!context) {
1232 rc = -ENOMEM;
1233 goto release_lock;
1234 }
1235 context->resync_nh_reset = 1;
1236
1237 ctx->priv_ctx_rx = context;
1238 rc = tls_set_sw_offload(sk, 0, NULL);
1239 if (rc)
1240 goto release_ctx;
1241
1242 rc = netdev->tlsdev_ops->tls_dev_add(netdev, sk, TLS_OFFLOAD_CTX_DIR_RX,
1243 &ctx->crypto_recv.info,
1244 tcp_sk(sk)->copied_seq);
1245 info = (void *)&ctx->crypto_recv.info;
1246 trace_tls_device_offload_set(sk, TLS_OFFLOAD_CTX_DIR_RX,
1247 tcp_sk(sk)->copied_seq, info->rec_seq, rc);
1248 if (rc)
1249 goto free_sw_resources;
1250
1251 tls_device_attach(ctx, sk, netdev);
1252 up_read(&device_offload_lock);
1253
1254 dev_put(netdev);
1255
1256 return 0;
1257
1258free_sw_resources:
1259 up_read(&device_offload_lock);
1260 tls_sw_free_resources_rx(sk);
1261 down_read(&device_offload_lock);
1262release_ctx:
1263 ctx->priv_ctx_rx = NULL;
1264release_lock:
1265 up_read(&device_offload_lock);
1266release_netdev:
1267 dev_put(netdev);
1268 return rc;
1269}
1270
1271void tls_device_offload_cleanup_rx(struct sock *sk)
1272{
1273 struct tls_context *tls_ctx = tls_get_ctx(sk);
1274 struct net_device *netdev;
1275
1276 down_read(&device_offload_lock);
1277 netdev = rcu_dereference_protected(tls_ctx->netdev,
1278 lockdep_is_held(&device_offload_lock));
1279 if (!netdev)
1280 goto out;
1281
1282 netdev->tlsdev_ops->tls_dev_del(netdev, tls_ctx,
1283 TLS_OFFLOAD_CTX_DIR_RX);
1284
1285 if (tls_ctx->tx_conf != TLS_HW) {
1286 dev_put(netdev);
1287 rcu_assign_pointer(tls_ctx->netdev, NULL);
1288 } else {
1289 set_bit(TLS_RX_DEV_CLOSED, &tls_ctx->flags);
1290 }
1291out:
1292 up_read(&device_offload_lock);
1293 tls_sw_release_resources_rx(sk);
1294}
1295
1296static int tls_device_down(struct net_device *netdev)
1297{
1298 struct tls_context *ctx, *tmp;
1299 unsigned long flags;
1300 LIST_HEAD(list);
1301
1302 /* Request a write lock to block new offload attempts */
1303 down_write(&device_offload_lock);
1304
1305 spin_lock_irqsave(&tls_device_lock, flags);
1306 list_for_each_entry_safe(ctx, tmp, &tls_device_list, list) {
1307 struct net_device *ctx_netdev =
1308 rcu_dereference_protected(ctx->netdev,
1309 lockdep_is_held(&device_offload_lock));
1310
1311 if (ctx_netdev != netdev ||
1312 !refcount_inc_not_zero(&ctx->refcount))
1313 continue;
1314
1315 list_move(&ctx->list, &list);
1316 }
1317 spin_unlock_irqrestore(&tls_device_lock, flags);
1318
1319 list_for_each_entry_safe(ctx, tmp, &list, list) {
1320 /* Stop offloaded TX and switch to the fallback.
1321 * tls_is_skb_tx_device_offloaded will return false.
1322 */
1323 WRITE_ONCE(ctx->sk->sk_validate_xmit_skb, tls_validate_xmit_skb_sw);
1324
1325 /* Stop the RX and TX resync.
1326 * tls_dev_resync must not be called after tls_dev_del.
1327 */
1328 rcu_assign_pointer(ctx->netdev, NULL);
1329
1330 /* Start skipping the RX resync logic completely. */
1331 set_bit(TLS_RX_DEV_DEGRADED, &ctx->flags);
1332
1333 /* Sync with inflight packets. After this point:
1334 * TX: no non-encrypted packets will be passed to the driver.
1335 * RX: resync requests from the driver will be ignored.
1336 */
1337 synchronize_net();
1338
1339 /* Release the offload context on the driver side. */
1340 if (ctx->tx_conf == TLS_HW)
1341 netdev->tlsdev_ops->tls_dev_del(netdev, ctx,
1342 TLS_OFFLOAD_CTX_DIR_TX);
1343 if (ctx->rx_conf == TLS_HW &&
1344 !test_bit(TLS_RX_DEV_CLOSED, &ctx->flags))
1345 netdev->tlsdev_ops->tls_dev_del(netdev, ctx,
1346 TLS_OFFLOAD_CTX_DIR_RX);
1347
1348 dev_put(netdev);
1349
1350 /* Move the context to a separate list for two reasons:
1351 * 1. When the context is deallocated, list_del is called.
1352 * 2. It's no longer an offloaded context, so we don't want to
1353 * run offload-specific code on this context.
1354 */
1355 spin_lock_irqsave(&tls_device_lock, flags);
1356 list_move_tail(&ctx->list, &tls_device_down_list);
1357 spin_unlock_irqrestore(&tls_device_lock, flags);
1358
1359 /* Device contexts for RX and TX will be freed in on sk_destruct
1360 * by tls_device_free_ctx. rx_conf and tx_conf stay in TLS_HW.
1361 * Now release the ref taken above.
1362 */
1363 if (refcount_dec_and_test(&ctx->refcount)) {
1364 /* sk_destruct ran after tls_device_down took a ref, and
1365 * it returned early. Complete the destruction here.
1366 */
1367 list_del(&ctx->list);
1368 tls_device_free_ctx(ctx);
1369 }
1370 }
1371
1372 up_write(&device_offload_lock);
1373
1374 flush_workqueue(destruct_wq);
1375
1376 return NOTIFY_DONE;
1377}
1378
1379static int tls_dev_event(struct notifier_block *this, unsigned long event,
1380 void *ptr)
1381{
1382 struct net_device *dev = netdev_notifier_info_to_dev(ptr);
1383
1384 if (!dev->tlsdev_ops &&
1385 !(dev->features & (NETIF_F_HW_TLS_RX | NETIF_F_HW_TLS_TX)))
1386 return NOTIFY_DONE;
1387
1388 switch (event) {
1389 case NETDEV_REGISTER:
1390 case NETDEV_FEAT_CHANGE:
1391 if (netif_is_bond_master(dev))
1392 return NOTIFY_DONE;
1393 if (!dev->tlsdev_ops ||
1394 !dev->tlsdev_ops->tls_dev_add ||
1395 !dev->tlsdev_ops->tls_dev_del)
1396 return NOTIFY_BAD;
1397 if ((dev->features & NETIF_F_HW_TLS_RX) &&
1398 !dev->tlsdev_ops->tls_dev_resync)
1399 return NOTIFY_BAD;
1400
1401 return NOTIFY_DONE;
1402 case NETDEV_DOWN:
1403 return tls_device_down(dev);
1404 }
1405 return NOTIFY_DONE;
1406}
1407
1408static struct notifier_block tls_dev_notifier = {
1409 .notifier_call = tls_dev_event,
1410};
1411
1412int __init tls_device_init(void)
1413{
1414 int err;
1415
1416 dummy_page = alloc_page(GFP_KERNEL);
1417 if (!dummy_page)
1418 return -ENOMEM;
1419
1420 destruct_wq = alloc_workqueue("ktls_device_destruct", WQ_PERCPU, 0);
1421 if (!destruct_wq) {
1422 err = -ENOMEM;
1423 goto err_free_dummy;
1424 }
1425
1426 err = register_netdevice_notifier(&tls_dev_notifier);
1427 if (err)
1428 goto err_destroy_wq;
1429
1430 return 0;
1431
1432err_destroy_wq:
1433 destroy_workqueue(destruct_wq);
1434err_free_dummy:
1435 put_page(dummy_page);
1436 return err;
1437}
1438
1439void __exit tls_device_cleanup(void)
1440{
1441 unregister_netdevice_notifier(&tls_dev_notifier);
1442 destroy_workqueue(destruct_wq);
1443 clean_acked_data_flush();
1444 put_page(dummy_page);
1445}