1 /* Copyright (c) 2018, Mellanox Technologies All rights reserved. 2 * 3 * This software is available to you under a choice of one of two 4 * licenses. You may choose to be licensed under the terms of the GNU 5 * General Public License (GPL) Version 2, available from the file 6 * COPYING in the main directory of this source tree, or the 7 * OpenIB.org BSD license below: 8 * 9 * Redistribution and use in source and binary forms, with or 10 * without modification, are permitted provided that the following 11 * conditions are met: 12 * 13 * - Redistributions of source code must retain the above 14 * copyright notice, this list of conditions and the following 15 * disclaimer. 16 * 17 * - Redistributions in binary form must reproduce the above 18 * copyright notice, this list of conditions and the following 19 * disclaimer in the documentation and/or other materials 20 * provided with the distribution. 21 * 22 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, 23 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF 24 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND 25 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS 26 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN 27 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN 28 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 29 * SOFTWARE. 30 */ 31 32 #include <crypto/aead.h> 33 #include <linux/highmem.h> 34 #include <linux/module.h> 35 #include <linux/netdevice.h> 36 #include <net/dst.h> 37 #include <net/inet_connection_sock.h> 38 #include <net/tcp.h> 39 #include <net/tls.h> 40 41 #include "tls.h" 42 #include "trace.h" 43 44 /* device_offload_lock is used to synchronize tls_dev_add 45 * against NETDEV_DOWN notifications. 46 */ 47 static DECLARE_RWSEM(device_offload_lock); 48 49 static struct workqueue_struct *destruct_wq __read_mostly; 50 51 static LIST_HEAD(tls_device_list); 52 static LIST_HEAD(tls_device_down_list); 53 static DEFINE_SPINLOCK(tls_device_lock); 54 55 static void tls_device_free_ctx(struct tls_context *ctx) 56 { 57 if (ctx->tx_conf == TLS_HW) { 58 kfree(tls_offload_ctx_tx(ctx)); 59 kfree(ctx->tx.rec_seq); 60 kfree(ctx->tx.iv); 61 } 62 63 if (ctx->rx_conf == TLS_HW) 64 kfree(tls_offload_ctx_rx(ctx)); 65 66 tls_ctx_free(NULL, ctx); 67 } 68 69 static void tls_device_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 88 static 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 */ 124 static struct net_device *get_netdev_for_sock(struct sock *sk) 125 { 126 struct dst_entry *dst = sk_dst_get(sk); 127 struct net_device *netdev = NULL; 128 129 if (likely(dst)) { 130 netdev = netdev_sk_get_lowest_dev(dst->dev, sk); 131 dev_hold(netdev); 132 } 133 134 dst_release(dst); 135 136 return netdev; 137 } 138 139 static void destroy_record(struct tls_record_info *record) 140 { 141 int i; 142 143 for (i = 0; i < record->num_frags; i++) 144 __skb_frag_unref(&record->frags[i], false); 145 kfree(record); 146 } 147 148 static void delete_all_records(struct tls_offload_context_tx *offload_ctx) 149 { 150 struct tls_record_info *info, *temp; 151 152 list_for_each_entry_safe(info, temp, &offload_ctx->records_list, list) { 153 list_del(&info->list); 154 destroy_record(info); 155 } 156 157 offload_ctx->retransmit_hint = NULL; 158 } 159 160 static void tls_icsk_clean_acked(struct sock *sk, u32 acked_seq) 161 { 162 struct tls_context *tls_ctx = tls_get_ctx(sk); 163 struct tls_record_info *info, *temp; 164 struct tls_offload_context_tx *ctx; 165 u64 deleted_records = 0; 166 unsigned long flags; 167 168 if (!tls_ctx) 169 return; 170 171 ctx = tls_offload_ctx_tx(tls_ctx); 172 173 spin_lock_irqsave(&ctx->lock, flags); 174 info = ctx->retransmit_hint; 175 if (info && !before(acked_seq, info->end_seq)) 176 ctx->retransmit_hint = NULL; 177 178 list_for_each_entry_safe(info, temp, &ctx->records_list, list) { 179 if (before(acked_seq, info->end_seq)) 180 break; 181 list_del(&info->list); 182 183 destroy_record(info); 184 deleted_records++; 185 } 186 187 ctx->unacked_record_sn += deleted_records; 188 spin_unlock_irqrestore(&ctx->lock, flags); 189 } 190 191 /* At this point, there should be no references on this 192 * socket and no in-flight SKBs associated with this 193 * socket, so it is safe to free all the resources. 194 */ 195 void tls_device_sk_destruct(struct sock *sk) 196 { 197 struct tls_context *tls_ctx = tls_get_ctx(sk); 198 struct tls_offload_context_tx *ctx = tls_offload_ctx_tx(tls_ctx); 199 200 tls_ctx->sk_destruct(sk); 201 202 if (tls_ctx->tx_conf == TLS_HW) { 203 if (ctx->open_record) 204 destroy_record(ctx->open_record); 205 delete_all_records(ctx); 206 crypto_free_aead(ctx->aead_send); 207 clean_acked_data_disable(inet_csk(sk)); 208 } 209 210 tls_device_queue_ctx_destruction(tls_ctx); 211 } 212 EXPORT_SYMBOL_GPL(tls_device_sk_destruct); 213 214 void tls_device_free_resources_tx(struct sock *sk) 215 { 216 struct tls_context *tls_ctx = tls_get_ctx(sk); 217 218 tls_free_partial_record(sk, tls_ctx); 219 } 220 221 void tls_offload_tx_resync_request(struct sock *sk, u32 got_seq, u32 exp_seq) 222 { 223 struct tls_context *tls_ctx = tls_get_ctx(sk); 224 225 trace_tls_device_tx_resync_req(sk, got_seq, exp_seq); 226 WARN_ON(test_and_set_bit(TLS_TX_SYNC_SCHED, &tls_ctx->flags)); 227 } 228 EXPORT_SYMBOL_GPL(tls_offload_tx_resync_request); 229 230 static void tls_device_resync_tx(struct sock *sk, struct tls_context *tls_ctx, 231 u32 seq) 232 { 233 struct net_device *netdev; 234 struct sk_buff *skb; 235 int err = 0; 236 u8 *rcd_sn; 237 238 skb = tcp_write_queue_tail(sk); 239 if (skb) 240 TCP_SKB_CB(skb)->eor = 1; 241 242 rcd_sn = tls_ctx->tx.rec_seq; 243 244 trace_tls_device_tx_resync_send(sk, seq, rcd_sn); 245 down_read(&device_offload_lock); 246 netdev = rcu_dereference_protected(tls_ctx->netdev, 247 lockdep_is_held(&device_offload_lock)); 248 if (netdev) 249 err = netdev->tlsdev_ops->tls_dev_resync(netdev, sk, seq, 250 rcd_sn, 251 TLS_OFFLOAD_CTX_DIR_TX); 252 up_read(&device_offload_lock); 253 if (err) 254 return; 255 256 clear_bit_unlock(TLS_TX_SYNC_SCHED, &tls_ctx->flags); 257 } 258 259 static void tls_append_frag(struct tls_record_info *record, 260 struct page_frag *pfrag, 261 int size) 262 { 263 skb_frag_t *frag; 264 265 frag = &record->frags[record->num_frags - 1]; 266 if (skb_frag_page(frag) == pfrag->page && 267 skb_frag_off(frag) + skb_frag_size(frag) == pfrag->offset) { 268 skb_frag_size_add(frag, size); 269 } else { 270 ++frag; 271 skb_frag_fill_page_desc(frag, pfrag->page, pfrag->offset, 272 size); 273 ++record->num_frags; 274 get_page(pfrag->page); 275 } 276 277 pfrag->offset += size; 278 record->len += size; 279 } 280 281 static int tls_push_record(struct sock *sk, 282 struct tls_context *ctx, 283 struct tls_offload_context_tx *offload_ctx, 284 struct tls_record_info *record, 285 int flags) 286 { 287 struct tls_prot_info *prot = &ctx->prot_info; 288 struct tcp_sock *tp = tcp_sk(sk); 289 skb_frag_t *frag; 290 int i; 291 292 record->end_seq = tp->write_seq + record->len; 293 list_add_tail_rcu(&record->list, &offload_ctx->records_list); 294 offload_ctx->open_record = NULL; 295 296 if (test_bit(TLS_TX_SYNC_SCHED, &ctx->flags)) 297 tls_device_resync_tx(sk, ctx, tp->write_seq); 298 299 tls_advance_record_sn(sk, prot, &ctx->tx); 300 301 for (i = 0; i < record->num_frags; i++) { 302 frag = &record->frags[i]; 303 sg_unmark_end(&offload_ctx->sg_tx_data[i]); 304 sg_set_page(&offload_ctx->sg_tx_data[i], skb_frag_page(frag), 305 skb_frag_size(frag), skb_frag_off(frag)); 306 sk_mem_charge(sk, skb_frag_size(frag)); 307 get_page(skb_frag_page(frag)); 308 } 309 sg_mark_end(&offload_ctx->sg_tx_data[record->num_frags - 1]); 310 311 /* all ready, send */ 312 return tls_push_sg(sk, ctx, offload_ctx->sg_tx_data, 0, flags); 313 } 314 315 static int tls_device_record_close(struct sock *sk, 316 struct tls_context *ctx, 317 struct tls_record_info *record, 318 struct page_frag *pfrag, 319 unsigned char record_type) 320 { 321 struct tls_prot_info *prot = &ctx->prot_info; 322 int ret; 323 324 /* append tag 325 * device will fill in the tag, we just need to append a placeholder 326 * use socket memory to improve coalescing (re-using a single buffer 327 * increases frag count) 328 * if we can't allocate memory now, steal some back from data 329 */ 330 if (likely(skb_page_frag_refill(prot->tag_size, pfrag, 331 sk->sk_allocation))) { 332 ret = 0; 333 tls_append_frag(record, pfrag, prot->tag_size); 334 } else { 335 ret = prot->tag_size; 336 if (record->len <= prot->overhead_size) 337 return -ENOMEM; 338 } 339 340 /* fill prepend */ 341 tls_fill_prepend(ctx, skb_frag_address(&record->frags[0]), 342 record->len - prot->overhead_size, 343 record_type); 344 return ret; 345 } 346 347 static int tls_create_new_record(struct tls_offload_context_tx *offload_ctx, 348 struct page_frag *pfrag, 349 size_t prepend_size) 350 { 351 struct tls_record_info *record; 352 skb_frag_t *frag; 353 354 record = kmalloc(sizeof(*record), GFP_KERNEL); 355 if (!record) 356 return -ENOMEM; 357 358 frag = &record->frags[0]; 359 skb_frag_fill_page_desc(frag, pfrag->page, pfrag->offset, 360 prepend_size); 361 362 get_page(pfrag->page); 363 pfrag->offset += prepend_size; 364 365 record->num_frags = 1; 366 record->len = prepend_size; 367 offload_ctx->open_record = record; 368 return 0; 369 } 370 371 static int tls_do_allocation(struct sock *sk, 372 struct tls_offload_context_tx *offload_ctx, 373 struct page_frag *pfrag, 374 size_t prepend_size) 375 { 376 int ret; 377 378 if (!offload_ctx->open_record) { 379 if (unlikely(!skb_page_frag_refill(prepend_size, pfrag, 380 sk->sk_allocation))) { 381 READ_ONCE(sk->sk_prot)->enter_memory_pressure(sk); 382 sk_stream_moderate_sndbuf(sk); 383 return -ENOMEM; 384 } 385 386 ret = tls_create_new_record(offload_ctx, pfrag, prepend_size); 387 if (ret) 388 return ret; 389 390 if (pfrag->size > pfrag->offset) 391 return 0; 392 } 393 394 if (!sk_page_frag_refill(sk, pfrag)) 395 return -ENOMEM; 396 397 return 0; 398 } 399 400 static int tls_device_copy_data(void *addr, size_t bytes, struct iov_iter *i) 401 { 402 size_t pre_copy, nocache; 403 404 pre_copy = ~((unsigned long)addr - 1) & (SMP_CACHE_BYTES - 1); 405 if (pre_copy) { 406 pre_copy = min(pre_copy, bytes); 407 if (copy_from_iter(addr, pre_copy, i) != pre_copy) 408 return -EFAULT; 409 bytes -= pre_copy; 410 addr += pre_copy; 411 } 412 413 nocache = round_down(bytes, SMP_CACHE_BYTES); 414 if (copy_from_iter_nocache(addr, nocache, i) != nocache) 415 return -EFAULT; 416 bytes -= nocache; 417 addr += nocache; 418 419 if (bytes && copy_from_iter(addr, bytes, i) != bytes) 420 return -EFAULT; 421 422 return 0; 423 } 424 425 union tls_iter_offset { 426 struct iov_iter *msg_iter; 427 int offset; 428 }; 429 430 static int tls_push_data(struct sock *sk, 431 union tls_iter_offset iter_offset, 432 size_t size, int flags, 433 unsigned char record_type, 434 struct page *zc_page) 435 { 436 struct tls_context *tls_ctx = tls_get_ctx(sk); 437 struct tls_prot_info *prot = &tls_ctx->prot_info; 438 struct tls_offload_context_tx *ctx = tls_offload_ctx_tx(tls_ctx); 439 struct tls_record_info *record; 440 int tls_push_record_flags; 441 struct page_frag *pfrag; 442 size_t orig_size = size; 443 u32 max_open_record_len; 444 bool more = false; 445 bool done = false; 446 int copy, rc = 0; 447 long timeo; 448 449 if (flags & 450 ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL | MSG_SENDPAGE_NOTLAST)) 451 return -EOPNOTSUPP; 452 453 if (unlikely(sk->sk_err)) 454 return -sk->sk_err; 455 456 flags |= MSG_SENDPAGE_DECRYPTED; 457 tls_push_record_flags = flags | MSG_SENDPAGE_NOTLAST; 458 459 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT); 460 if (tls_is_partially_sent_record(tls_ctx)) { 461 rc = tls_push_partial_record(sk, tls_ctx, flags); 462 if (rc < 0) 463 return rc; 464 } 465 466 pfrag = sk_page_frag(sk); 467 468 /* TLS_HEADER_SIZE is not counted as part of the TLS record, and 469 * we need to leave room for an authentication tag. 470 */ 471 max_open_record_len = TLS_MAX_PAYLOAD_SIZE + 472 prot->prepend_size; 473 do { 474 rc = tls_do_allocation(sk, ctx, pfrag, prot->prepend_size); 475 if (unlikely(rc)) { 476 rc = sk_stream_wait_memory(sk, &timeo); 477 if (!rc) 478 continue; 479 480 record = ctx->open_record; 481 if (!record) 482 break; 483 handle_error: 484 if (record_type != TLS_RECORD_TYPE_DATA) { 485 /* avoid sending partial 486 * record with type != 487 * application_data 488 */ 489 size = orig_size; 490 destroy_record(record); 491 ctx->open_record = NULL; 492 } else if (record->len > prot->prepend_size) { 493 goto last_record; 494 } 495 496 break; 497 } 498 499 record = ctx->open_record; 500 501 copy = min_t(size_t, size, max_open_record_len - record->len); 502 if (copy && zc_page) { 503 struct page_frag zc_pfrag; 504 505 zc_pfrag.page = zc_page; 506 zc_pfrag.offset = iter_offset.offset; 507 zc_pfrag.size = copy; 508 tls_append_frag(record, &zc_pfrag, copy); 509 510 iter_offset.offset += copy; 511 } else if (copy) { 512 copy = min_t(size_t, copy, pfrag->size - pfrag->offset); 513 514 rc = tls_device_copy_data(page_address(pfrag->page) + 515 pfrag->offset, copy, 516 iter_offset.msg_iter); 517 if (rc) 518 goto handle_error; 519 tls_append_frag(record, pfrag, copy); 520 } 521 522 size -= copy; 523 if (!size) { 524 last_record: 525 tls_push_record_flags = flags; 526 if (flags & (MSG_SENDPAGE_NOTLAST | MSG_MORE)) { 527 more = true; 528 break; 529 } 530 531 done = true; 532 } 533 534 if (done || record->len >= max_open_record_len || 535 (record->num_frags >= MAX_SKB_FRAGS - 1)) { 536 rc = tls_device_record_close(sk, tls_ctx, record, 537 pfrag, record_type); 538 if (rc) { 539 if (rc > 0) { 540 size += rc; 541 } else { 542 size = orig_size; 543 destroy_record(record); 544 ctx->open_record = NULL; 545 break; 546 } 547 } 548 549 rc = tls_push_record(sk, 550 tls_ctx, 551 ctx, 552 record, 553 tls_push_record_flags); 554 if (rc < 0) 555 break; 556 } 557 } while (!done); 558 559 tls_ctx->pending_open_record_frags = more; 560 561 if (orig_size - size > 0) 562 rc = orig_size - size; 563 564 return rc; 565 } 566 567 int tls_device_sendmsg(struct sock *sk, struct msghdr *msg, size_t size) 568 { 569 unsigned char record_type = TLS_RECORD_TYPE_DATA; 570 struct tls_context *tls_ctx = tls_get_ctx(sk); 571 union tls_iter_offset iter; 572 int rc; 573 574 mutex_lock(&tls_ctx->tx_lock); 575 lock_sock(sk); 576 577 if (unlikely(msg->msg_controllen)) { 578 rc = tls_process_cmsg(sk, msg, &record_type); 579 if (rc) 580 goto out; 581 } 582 583 iter.msg_iter = &msg->msg_iter; 584 rc = tls_push_data(sk, iter, size, msg->msg_flags, record_type, NULL); 585 586 out: 587 release_sock(sk); 588 mutex_unlock(&tls_ctx->tx_lock); 589 return rc; 590 } 591 592 int tls_device_sendpage(struct sock *sk, struct page *page, 593 int offset, size_t size, int flags) 594 { 595 struct tls_context *tls_ctx = tls_get_ctx(sk); 596 union tls_iter_offset iter_offset; 597 struct iov_iter msg_iter; 598 char *kaddr; 599 struct kvec iov; 600 int rc; 601 602 if (flags & MSG_SENDPAGE_NOTLAST) 603 flags |= MSG_MORE; 604 605 mutex_lock(&tls_ctx->tx_lock); 606 lock_sock(sk); 607 608 if (flags & MSG_OOB) { 609 rc = -EOPNOTSUPP; 610 goto out; 611 } 612 613 if (tls_ctx->zerocopy_sendfile) { 614 iter_offset.offset = offset; 615 rc = tls_push_data(sk, iter_offset, size, 616 flags, TLS_RECORD_TYPE_DATA, page); 617 goto out; 618 } 619 620 kaddr = kmap(page); 621 iov.iov_base = kaddr + offset; 622 iov.iov_len = size; 623 iov_iter_kvec(&msg_iter, ITER_SOURCE, &iov, 1, size); 624 iter_offset.msg_iter = &msg_iter; 625 rc = tls_push_data(sk, iter_offset, size, flags, TLS_RECORD_TYPE_DATA, 626 NULL); 627 kunmap(page); 628 629 out: 630 release_sock(sk); 631 mutex_unlock(&tls_ctx->tx_lock); 632 return rc; 633 } 634 635 struct tls_record_info *tls_get_record(struct tls_offload_context_tx *context, 636 u32 seq, u64 *p_record_sn) 637 { 638 u64 record_sn = context->hint_record_sn; 639 struct tls_record_info *info, *last; 640 641 info = context->retransmit_hint; 642 if (!info || 643 before(seq, info->end_seq - info->len)) { 644 /* if retransmit_hint is irrelevant start 645 * from the beginning of the list 646 */ 647 info = list_first_entry_or_null(&context->records_list, 648 struct tls_record_info, list); 649 if (!info) 650 return NULL; 651 /* send the start_marker record if seq number is before the 652 * tls offload start marker sequence number. This record is 653 * required to handle TCP packets which are before TLS offload 654 * started. 655 * And if it's not start marker, look if this seq number 656 * belongs to the list. 657 */ 658 if (likely(!tls_record_is_start_marker(info))) { 659 /* we have the first record, get the last record to see 660 * if this seq number belongs to the list. 661 */ 662 last = list_last_entry(&context->records_list, 663 struct tls_record_info, list); 664 665 if (!between(seq, tls_record_start_seq(info), 666 last->end_seq)) 667 return NULL; 668 } 669 record_sn = context->unacked_record_sn; 670 } 671 672 /* We just need the _rcu for the READ_ONCE() */ 673 rcu_read_lock(); 674 list_for_each_entry_from_rcu(info, &context->records_list, list) { 675 if (before(seq, info->end_seq)) { 676 if (!context->retransmit_hint || 677 after(info->end_seq, 678 context->retransmit_hint->end_seq)) { 679 context->hint_record_sn = record_sn; 680 context->retransmit_hint = info; 681 } 682 *p_record_sn = record_sn; 683 goto exit_rcu_unlock; 684 } 685 record_sn++; 686 } 687 info = NULL; 688 689 exit_rcu_unlock: 690 rcu_read_unlock(); 691 return info; 692 } 693 EXPORT_SYMBOL(tls_get_record); 694 695 static int tls_device_push_pending_record(struct sock *sk, int flags) 696 { 697 union tls_iter_offset iter; 698 struct iov_iter msg_iter; 699 700 iov_iter_kvec(&msg_iter, ITER_SOURCE, NULL, 0, 0); 701 iter.msg_iter = &msg_iter; 702 return tls_push_data(sk, iter, 0, flags, TLS_RECORD_TYPE_DATA, NULL); 703 } 704 705 void tls_device_write_space(struct sock *sk, struct tls_context *ctx) 706 { 707 if (tls_is_partially_sent_record(ctx)) { 708 gfp_t sk_allocation = sk->sk_allocation; 709 710 WARN_ON_ONCE(sk->sk_write_pending); 711 712 sk->sk_allocation = GFP_ATOMIC; 713 tls_push_partial_record(sk, ctx, 714 MSG_DONTWAIT | MSG_NOSIGNAL | 715 MSG_SENDPAGE_DECRYPTED); 716 sk->sk_allocation = sk_allocation; 717 } 718 } 719 720 static void tls_device_resync_rx(struct tls_context *tls_ctx, 721 struct sock *sk, u32 seq, u8 *rcd_sn) 722 { 723 struct tls_offload_context_rx *rx_ctx = tls_offload_ctx_rx(tls_ctx); 724 struct net_device *netdev; 725 726 trace_tls_device_rx_resync_send(sk, seq, rcd_sn, rx_ctx->resync_type); 727 rcu_read_lock(); 728 netdev = rcu_dereference(tls_ctx->netdev); 729 if (netdev) 730 netdev->tlsdev_ops->tls_dev_resync(netdev, sk, seq, rcd_sn, 731 TLS_OFFLOAD_CTX_DIR_RX); 732 rcu_read_unlock(); 733 TLS_INC_STATS(sock_net(sk), LINUX_MIB_TLSRXDEVICERESYNC); 734 } 735 736 static bool 737 tls_device_rx_resync_async(struct tls_offload_resync_async *resync_async, 738 s64 resync_req, u32 *seq, u16 *rcd_delta) 739 { 740 u32 is_async = resync_req & RESYNC_REQ_ASYNC; 741 u32 req_seq = resync_req >> 32; 742 u32 req_end = req_seq + ((resync_req >> 16) & 0xffff); 743 u16 i; 744 745 *rcd_delta = 0; 746 747 if (is_async) { 748 /* shouldn't get to wraparound: 749 * too long in async stage, something bad happened 750 */ 751 if (WARN_ON_ONCE(resync_async->rcd_delta == USHRT_MAX)) 752 return false; 753 754 /* asynchronous stage: log all headers seq such that 755 * req_seq <= seq <= end_seq, and wait for real resync request 756 */ 757 if (before(*seq, req_seq)) 758 return false; 759 if (!after(*seq, req_end) && 760 resync_async->loglen < TLS_DEVICE_RESYNC_ASYNC_LOGMAX) 761 resync_async->log[resync_async->loglen++] = *seq; 762 763 resync_async->rcd_delta++; 764 765 return false; 766 } 767 768 /* synchronous stage: check against the logged entries and 769 * proceed to check the next entries if no match was found 770 */ 771 for (i = 0; i < resync_async->loglen; i++) 772 if (req_seq == resync_async->log[i] && 773 atomic64_try_cmpxchg(&resync_async->req, &resync_req, 0)) { 774 *rcd_delta = resync_async->rcd_delta - i; 775 *seq = req_seq; 776 resync_async->loglen = 0; 777 resync_async->rcd_delta = 0; 778 return true; 779 } 780 781 resync_async->loglen = 0; 782 resync_async->rcd_delta = 0; 783 784 if (req_seq == *seq && 785 atomic64_try_cmpxchg(&resync_async->req, 786 &resync_req, 0)) 787 return true; 788 789 return false; 790 } 791 792 void tls_device_rx_resync_new_rec(struct sock *sk, u32 rcd_len, u32 seq) 793 { 794 struct tls_context *tls_ctx = tls_get_ctx(sk); 795 struct tls_offload_context_rx *rx_ctx; 796 u8 rcd_sn[TLS_MAX_REC_SEQ_SIZE]; 797 u32 sock_data, is_req_pending; 798 struct tls_prot_info *prot; 799 s64 resync_req; 800 u16 rcd_delta; 801 u32 req_seq; 802 803 if (tls_ctx->rx_conf != TLS_HW) 804 return; 805 if (unlikely(test_bit(TLS_RX_DEV_DEGRADED, &tls_ctx->flags))) 806 return; 807 808 prot = &tls_ctx->prot_info; 809 rx_ctx = tls_offload_ctx_rx(tls_ctx); 810 memcpy(rcd_sn, tls_ctx->rx.rec_seq, prot->rec_seq_size); 811 812 switch (rx_ctx->resync_type) { 813 case TLS_OFFLOAD_SYNC_TYPE_DRIVER_REQ: 814 resync_req = atomic64_read(&rx_ctx->resync_req); 815 req_seq = resync_req >> 32; 816 seq += TLS_HEADER_SIZE - 1; 817 is_req_pending = resync_req; 818 819 if (likely(!is_req_pending) || req_seq != seq || 820 !atomic64_try_cmpxchg(&rx_ctx->resync_req, &resync_req, 0)) 821 return; 822 break; 823 case TLS_OFFLOAD_SYNC_TYPE_CORE_NEXT_HINT: 824 if (likely(!rx_ctx->resync_nh_do_now)) 825 return; 826 827 /* head of next rec is already in, note that the sock_inq will 828 * include the currently parsed message when called from parser 829 */ 830 sock_data = tcp_inq(sk); 831 if (sock_data > rcd_len) { 832 trace_tls_device_rx_resync_nh_delay(sk, sock_data, 833 rcd_len); 834 return; 835 } 836 837 rx_ctx->resync_nh_do_now = 0; 838 seq += rcd_len; 839 tls_bigint_increment(rcd_sn, prot->rec_seq_size); 840 break; 841 case TLS_OFFLOAD_SYNC_TYPE_DRIVER_REQ_ASYNC: 842 resync_req = atomic64_read(&rx_ctx->resync_async->req); 843 is_req_pending = resync_req; 844 if (likely(!is_req_pending)) 845 return; 846 847 if (!tls_device_rx_resync_async(rx_ctx->resync_async, 848 resync_req, &seq, &rcd_delta)) 849 return; 850 tls_bigint_subtract(rcd_sn, rcd_delta); 851 break; 852 } 853 854 tls_device_resync_rx(tls_ctx, sk, seq, rcd_sn); 855 } 856 857 static void tls_device_core_ctrl_rx_resync(struct tls_context *tls_ctx, 858 struct tls_offload_context_rx *ctx, 859 struct sock *sk, struct sk_buff *skb) 860 { 861 struct strp_msg *rxm; 862 863 /* device will request resyncs by itself based on stream scan */ 864 if (ctx->resync_type != TLS_OFFLOAD_SYNC_TYPE_CORE_NEXT_HINT) 865 return; 866 /* already scheduled */ 867 if (ctx->resync_nh_do_now) 868 return; 869 /* seen decrypted fragments since last fully-failed record */ 870 if (ctx->resync_nh_reset) { 871 ctx->resync_nh_reset = 0; 872 ctx->resync_nh.decrypted_failed = 1; 873 ctx->resync_nh.decrypted_tgt = TLS_DEVICE_RESYNC_NH_START_IVAL; 874 return; 875 } 876 877 if (++ctx->resync_nh.decrypted_failed <= ctx->resync_nh.decrypted_tgt) 878 return; 879 880 /* doing resync, bump the next target in case it fails */ 881 if (ctx->resync_nh.decrypted_tgt < TLS_DEVICE_RESYNC_NH_MAX_IVAL) 882 ctx->resync_nh.decrypted_tgt *= 2; 883 else 884 ctx->resync_nh.decrypted_tgt += TLS_DEVICE_RESYNC_NH_MAX_IVAL; 885 886 rxm = strp_msg(skb); 887 888 /* head of next rec is already in, parser will sync for us */ 889 if (tcp_inq(sk) > rxm->full_len) { 890 trace_tls_device_rx_resync_nh_schedule(sk); 891 ctx->resync_nh_do_now = 1; 892 } else { 893 struct tls_prot_info *prot = &tls_ctx->prot_info; 894 u8 rcd_sn[TLS_MAX_REC_SEQ_SIZE]; 895 896 memcpy(rcd_sn, tls_ctx->rx.rec_seq, prot->rec_seq_size); 897 tls_bigint_increment(rcd_sn, prot->rec_seq_size); 898 899 tls_device_resync_rx(tls_ctx, sk, tcp_sk(sk)->copied_seq, 900 rcd_sn); 901 } 902 } 903 904 static int 905 tls_device_reencrypt(struct sock *sk, struct tls_context *tls_ctx) 906 { 907 struct tls_sw_context_rx *sw_ctx = tls_sw_ctx_rx(tls_ctx); 908 const struct tls_cipher_size_desc *cipher_sz; 909 int err, offset, copy, data_len, pos; 910 struct sk_buff *skb, *skb_iter; 911 struct scatterlist sg[1]; 912 struct strp_msg *rxm; 913 char *orig_buf, *buf; 914 915 switch (tls_ctx->crypto_recv.info.cipher_type) { 916 case TLS_CIPHER_AES_GCM_128: 917 case TLS_CIPHER_AES_GCM_256: 918 break; 919 default: 920 return -EINVAL; 921 } 922 cipher_sz = &tls_cipher_size_desc[tls_ctx->crypto_recv.info.cipher_type]; 923 924 rxm = strp_msg(tls_strp_msg(sw_ctx)); 925 orig_buf = kmalloc(rxm->full_len + TLS_HEADER_SIZE + cipher_sz->iv, 926 sk->sk_allocation); 927 if (!orig_buf) 928 return -ENOMEM; 929 buf = orig_buf; 930 931 err = tls_strp_msg_cow(sw_ctx); 932 if (unlikely(err)) 933 goto free_buf; 934 935 skb = tls_strp_msg(sw_ctx); 936 rxm = strp_msg(skb); 937 offset = rxm->offset; 938 939 sg_init_table(sg, 1); 940 sg_set_buf(&sg[0], buf, 941 rxm->full_len + TLS_HEADER_SIZE + cipher_sz->iv); 942 err = skb_copy_bits(skb, offset, buf, TLS_HEADER_SIZE + cipher_sz->iv); 943 if (err) 944 goto free_buf; 945 946 /* We are interested only in the decrypted data not the auth */ 947 err = decrypt_skb(sk, sg); 948 if (err != -EBADMSG) 949 goto free_buf; 950 else 951 err = 0; 952 953 data_len = rxm->full_len - cipher_sz->tag; 954 955 if (skb_pagelen(skb) > offset) { 956 copy = min_t(int, skb_pagelen(skb) - offset, data_len); 957 958 if (skb->decrypted) { 959 err = skb_store_bits(skb, offset, buf, copy); 960 if (err) 961 goto free_buf; 962 } 963 964 offset += copy; 965 buf += copy; 966 } 967 968 pos = skb_pagelen(skb); 969 skb_walk_frags(skb, skb_iter) { 970 int frag_pos; 971 972 /* Practically all frags must belong to msg if reencrypt 973 * is needed with current strparser and coalescing logic, 974 * but strparser may "get optimized", so let's be safe. 975 */ 976 if (pos + skb_iter->len <= offset) 977 goto done_with_frag; 978 if (pos >= data_len + rxm->offset) 979 break; 980 981 frag_pos = offset - pos; 982 copy = min_t(int, skb_iter->len - frag_pos, 983 data_len + rxm->offset - offset); 984 985 if (skb_iter->decrypted) { 986 err = skb_store_bits(skb_iter, frag_pos, buf, copy); 987 if (err) 988 goto free_buf; 989 } 990 991 offset += copy; 992 buf += copy; 993 done_with_frag: 994 pos += skb_iter->len; 995 } 996 997 free_buf: 998 kfree(orig_buf); 999 return err; 1000 } 1001 1002 int tls_device_decrypted(struct sock *sk, struct tls_context *tls_ctx) 1003 { 1004 struct tls_offload_context_rx *ctx = tls_offload_ctx_rx(tls_ctx); 1005 struct tls_sw_context_rx *sw_ctx = tls_sw_ctx_rx(tls_ctx); 1006 struct sk_buff *skb = tls_strp_msg(sw_ctx); 1007 struct strp_msg *rxm = strp_msg(skb); 1008 int is_decrypted, is_encrypted; 1009 1010 if (!tls_strp_msg_mixed_decrypted(sw_ctx)) { 1011 is_decrypted = skb->decrypted; 1012 is_encrypted = !is_decrypted; 1013 } else { 1014 is_decrypted = 0; 1015 is_encrypted = 0; 1016 } 1017 1018 trace_tls_device_decrypted(sk, tcp_sk(sk)->copied_seq - rxm->full_len, 1019 tls_ctx->rx.rec_seq, rxm->full_len, 1020 is_encrypted, is_decrypted); 1021 1022 if (unlikely(test_bit(TLS_RX_DEV_DEGRADED, &tls_ctx->flags))) { 1023 if (likely(is_encrypted || is_decrypted)) 1024 return is_decrypted; 1025 1026 /* After tls_device_down disables the offload, the next SKB will 1027 * likely have initial fragments decrypted, and final ones not 1028 * decrypted. We need to reencrypt that single SKB. 1029 */ 1030 return tls_device_reencrypt(sk, tls_ctx); 1031 } 1032 1033 /* Return immediately if the record is either entirely plaintext or 1034 * entirely ciphertext. Otherwise handle reencrypt partially decrypted 1035 * record. 1036 */ 1037 if (is_decrypted) { 1038 ctx->resync_nh_reset = 1; 1039 return is_decrypted; 1040 } 1041 if (is_encrypted) { 1042 tls_device_core_ctrl_rx_resync(tls_ctx, ctx, sk, skb); 1043 return 0; 1044 } 1045 1046 ctx->resync_nh_reset = 1; 1047 return tls_device_reencrypt(sk, tls_ctx); 1048 } 1049 1050 static void tls_device_attach(struct tls_context *ctx, struct sock *sk, 1051 struct net_device *netdev) 1052 { 1053 if (sk->sk_destruct != tls_device_sk_destruct) { 1054 refcount_set(&ctx->refcount, 1); 1055 dev_hold(netdev); 1056 RCU_INIT_POINTER(ctx->netdev, netdev); 1057 spin_lock_irq(&tls_device_lock); 1058 list_add_tail(&ctx->list, &tls_device_list); 1059 spin_unlock_irq(&tls_device_lock); 1060 1061 ctx->sk_destruct = sk->sk_destruct; 1062 smp_store_release(&sk->sk_destruct, tls_device_sk_destruct); 1063 } 1064 } 1065 1066 int tls_set_device_offload(struct sock *sk, struct tls_context *ctx) 1067 { 1068 struct tls_context *tls_ctx = tls_get_ctx(sk); 1069 struct tls_prot_info *prot = &tls_ctx->prot_info; 1070 const struct tls_cipher_size_desc *cipher_sz; 1071 struct tls_record_info *start_marker_record; 1072 struct tls_offload_context_tx *offload_ctx; 1073 struct tls_crypto_info *crypto_info; 1074 struct net_device *netdev; 1075 char *iv, *rec_seq; 1076 struct sk_buff *skb; 1077 __be64 rcd_sn; 1078 int rc; 1079 1080 if (!ctx) 1081 return -EINVAL; 1082 1083 if (ctx->priv_ctx_tx) 1084 return -EEXIST; 1085 1086 netdev = get_netdev_for_sock(sk); 1087 if (!netdev) { 1088 pr_err_ratelimited("%s: netdev not found\n", __func__); 1089 return -EINVAL; 1090 } 1091 1092 if (!(netdev->features & NETIF_F_HW_TLS_TX)) { 1093 rc = -EOPNOTSUPP; 1094 goto release_netdev; 1095 } 1096 1097 crypto_info = &ctx->crypto_send.info; 1098 if (crypto_info->version != TLS_1_2_VERSION) { 1099 rc = -EOPNOTSUPP; 1100 goto release_netdev; 1101 } 1102 1103 switch (crypto_info->cipher_type) { 1104 case TLS_CIPHER_AES_GCM_128: 1105 iv = ((struct tls12_crypto_info_aes_gcm_128 *)crypto_info)->iv; 1106 rec_seq = 1107 ((struct tls12_crypto_info_aes_gcm_128 *)crypto_info)->rec_seq; 1108 break; 1109 case TLS_CIPHER_AES_GCM_256: 1110 iv = ((struct tls12_crypto_info_aes_gcm_256 *)crypto_info)->iv; 1111 rec_seq = 1112 ((struct tls12_crypto_info_aes_gcm_256 *)crypto_info)->rec_seq; 1113 break; 1114 default: 1115 rc = -EINVAL; 1116 goto release_netdev; 1117 } 1118 cipher_sz = &tls_cipher_size_desc[crypto_info->cipher_type]; 1119 1120 /* Sanity-check the rec_seq_size for stack allocations */ 1121 if (cipher_sz->rec_seq > TLS_MAX_REC_SEQ_SIZE) { 1122 rc = -EINVAL; 1123 goto release_netdev; 1124 } 1125 1126 prot->version = crypto_info->version; 1127 prot->cipher_type = crypto_info->cipher_type; 1128 prot->prepend_size = TLS_HEADER_SIZE + cipher_sz->iv; 1129 prot->tag_size = cipher_sz->tag; 1130 prot->overhead_size = prot->prepend_size + prot->tag_size; 1131 prot->iv_size = cipher_sz->iv; 1132 prot->salt_size = cipher_sz->salt; 1133 ctx->tx.iv = kmalloc(cipher_sz->iv + cipher_sz->salt, GFP_KERNEL); 1134 if (!ctx->tx.iv) { 1135 rc = -ENOMEM; 1136 goto release_netdev; 1137 } 1138 1139 memcpy(ctx->tx.iv + cipher_sz->salt, iv, cipher_sz->iv); 1140 1141 prot->rec_seq_size = cipher_sz->rec_seq; 1142 ctx->tx.rec_seq = kmemdup(rec_seq, cipher_sz->rec_seq, GFP_KERNEL); 1143 if (!ctx->tx.rec_seq) { 1144 rc = -ENOMEM; 1145 goto free_iv; 1146 } 1147 1148 start_marker_record = kmalloc(sizeof(*start_marker_record), GFP_KERNEL); 1149 if (!start_marker_record) { 1150 rc = -ENOMEM; 1151 goto free_rec_seq; 1152 } 1153 1154 offload_ctx = kzalloc(TLS_OFFLOAD_CONTEXT_SIZE_TX, GFP_KERNEL); 1155 if (!offload_ctx) { 1156 rc = -ENOMEM; 1157 goto free_marker_record; 1158 } 1159 1160 rc = tls_sw_fallback_init(sk, offload_ctx, crypto_info); 1161 if (rc) 1162 goto free_offload_ctx; 1163 1164 /* start at rec_seq - 1 to account for the start marker record */ 1165 memcpy(&rcd_sn, ctx->tx.rec_seq, sizeof(rcd_sn)); 1166 offload_ctx->unacked_record_sn = be64_to_cpu(rcd_sn) - 1; 1167 1168 start_marker_record->end_seq = tcp_sk(sk)->write_seq; 1169 start_marker_record->len = 0; 1170 start_marker_record->num_frags = 0; 1171 1172 INIT_WORK(&offload_ctx->destruct_work, tls_device_tx_del_task); 1173 offload_ctx->ctx = ctx; 1174 1175 INIT_LIST_HEAD(&offload_ctx->records_list); 1176 list_add_tail(&start_marker_record->list, &offload_ctx->records_list); 1177 spin_lock_init(&offload_ctx->lock); 1178 sg_init_table(offload_ctx->sg_tx_data, 1179 ARRAY_SIZE(offload_ctx->sg_tx_data)); 1180 1181 clean_acked_data_enable(inet_csk(sk), &tls_icsk_clean_acked); 1182 ctx->push_pending_record = tls_device_push_pending_record; 1183 1184 /* TLS offload is greatly simplified if we don't send 1185 * SKBs where only part of the payload needs to be encrypted. 1186 * So mark the last skb in the write queue as end of record. 1187 */ 1188 skb = tcp_write_queue_tail(sk); 1189 if (skb) 1190 TCP_SKB_CB(skb)->eor = 1; 1191 1192 /* Avoid offloading if the device is down 1193 * We don't want to offload new flows after 1194 * the NETDEV_DOWN event 1195 * 1196 * device_offload_lock is taken in tls_devices's NETDEV_DOWN 1197 * handler thus protecting from the device going down before 1198 * ctx was added to tls_device_list. 1199 */ 1200 down_read(&device_offload_lock); 1201 if (!(netdev->flags & IFF_UP)) { 1202 rc = -EINVAL; 1203 goto release_lock; 1204 } 1205 1206 ctx->priv_ctx_tx = offload_ctx; 1207 rc = netdev->tlsdev_ops->tls_dev_add(netdev, sk, TLS_OFFLOAD_CTX_DIR_TX, 1208 &ctx->crypto_send.info, 1209 tcp_sk(sk)->write_seq); 1210 trace_tls_device_offload_set(sk, TLS_OFFLOAD_CTX_DIR_TX, 1211 tcp_sk(sk)->write_seq, rec_seq, rc); 1212 if (rc) 1213 goto release_lock; 1214 1215 tls_device_attach(ctx, sk, netdev); 1216 up_read(&device_offload_lock); 1217 1218 /* following this assignment tls_is_sk_tx_device_offloaded 1219 * will return true and the context might be accessed 1220 * by the netdev's xmit function. 1221 */ 1222 smp_store_release(&sk->sk_validate_xmit_skb, tls_validate_xmit_skb); 1223 dev_put(netdev); 1224 1225 return 0; 1226 1227 release_lock: 1228 up_read(&device_offload_lock); 1229 clean_acked_data_disable(inet_csk(sk)); 1230 crypto_free_aead(offload_ctx->aead_send); 1231 free_offload_ctx: 1232 kfree(offload_ctx); 1233 ctx->priv_ctx_tx = NULL; 1234 free_marker_record: 1235 kfree(start_marker_record); 1236 free_rec_seq: 1237 kfree(ctx->tx.rec_seq); 1238 free_iv: 1239 kfree(ctx->tx.iv); 1240 release_netdev: 1241 dev_put(netdev); 1242 return rc; 1243 } 1244 1245 int tls_set_device_offload_rx(struct sock *sk, struct tls_context *ctx) 1246 { 1247 struct tls12_crypto_info_aes_gcm_128 *info; 1248 struct tls_offload_context_rx *context; 1249 struct net_device *netdev; 1250 int rc = 0; 1251 1252 if (ctx->crypto_recv.info.version != TLS_1_2_VERSION) 1253 return -EOPNOTSUPP; 1254 1255 netdev = get_netdev_for_sock(sk); 1256 if (!netdev) { 1257 pr_err_ratelimited("%s: netdev not found\n", __func__); 1258 return -EINVAL; 1259 } 1260 1261 if (!(netdev->features & NETIF_F_HW_TLS_RX)) { 1262 rc = -EOPNOTSUPP; 1263 goto release_netdev; 1264 } 1265 1266 /* Avoid offloading if the device is down 1267 * We don't want to offload new flows after 1268 * the NETDEV_DOWN event 1269 * 1270 * device_offload_lock is taken in tls_devices's NETDEV_DOWN 1271 * handler thus protecting from the device going down before 1272 * ctx was added to tls_device_list. 1273 */ 1274 down_read(&device_offload_lock); 1275 if (!(netdev->flags & IFF_UP)) { 1276 rc = -EINVAL; 1277 goto release_lock; 1278 } 1279 1280 context = kzalloc(TLS_OFFLOAD_CONTEXT_SIZE_RX, GFP_KERNEL); 1281 if (!context) { 1282 rc = -ENOMEM; 1283 goto release_lock; 1284 } 1285 context->resync_nh_reset = 1; 1286 1287 ctx->priv_ctx_rx = context; 1288 rc = tls_set_sw_offload(sk, ctx, 0); 1289 if (rc) 1290 goto release_ctx; 1291 1292 rc = netdev->tlsdev_ops->tls_dev_add(netdev, sk, TLS_OFFLOAD_CTX_DIR_RX, 1293 &ctx->crypto_recv.info, 1294 tcp_sk(sk)->copied_seq); 1295 info = (void *)&ctx->crypto_recv.info; 1296 trace_tls_device_offload_set(sk, TLS_OFFLOAD_CTX_DIR_RX, 1297 tcp_sk(sk)->copied_seq, info->rec_seq, rc); 1298 if (rc) 1299 goto free_sw_resources; 1300 1301 tls_device_attach(ctx, sk, netdev); 1302 up_read(&device_offload_lock); 1303 1304 dev_put(netdev); 1305 1306 return 0; 1307 1308 free_sw_resources: 1309 up_read(&device_offload_lock); 1310 tls_sw_free_resources_rx(sk); 1311 down_read(&device_offload_lock); 1312 release_ctx: 1313 ctx->priv_ctx_rx = NULL; 1314 release_lock: 1315 up_read(&device_offload_lock); 1316 release_netdev: 1317 dev_put(netdev); 1318 return rc; 1319 } 1320 1321 void tls_device_offload_cleanup_rx(struct sock *sk) 1322 { 1323 struct tls_context *tls_ctx = tls_get_ctx(sk); 1324 struct net_device *netdev; 1325 1326 down_read(&device_offload_lock); 1327 netdev = rcu_dereference_protected(tls_ctx->netdev, 1328 lockdep_is_held(&device_offload_lock)); 1329 if (!netdev) 1330 goto out; 1331 1332 netdev->tlsdev_ops->tls_dev_del(netdev, tls_ctx, 1333 TLS_OFFLOAD_CTX_DIR_RX); 1334 1335 if (tls_ctx->tx_conf != TLS_HW) { 1336 dev_put(netdev); 1337 rcu_assign_pointer(tls_ctx->netdev, NULL); 1338 } else { 1339 set_bit(TLS_RX_DEV_CLOSED, &tls_ctx->flags); 1340 } 1341 out: 1342 up_read(&device_offload_lock); 1343 tls_sw_release_resources_rx(sk); 1344 } 1345 1346 static int tls_device_down(struct net_device *netdev) 1347 { 1348 struct tls_context *ctx, *tmp; 1349 unsigned long flags; 1350 LIST_HEAD(list); 1351 1352 /* Request a write lock to block new offload attempts */ 1353 down_write(&device_offload_lock); 1354 1355 spin_lock_irqsave(&tls_device_lock, flags); 1356 list_for_each_entry_safe(ctx, tmp, &tls_device_list, list) { 1357 struct net_device *ctx_netdev = 1358 rcu_dereference_protected(ctx->netdev, 1359 lockdep_is_held(&device_offload_lock)); 1360 1361 if (ctx_netdev != netdev || 1362 !refcount_inc_not_zero(&ctx->refcount)) 1363 continue; 1364 1365 list_move(&ctx->list, &list); 1366 } 1367 spin_unlock_irqrestore(&tls_device_lock, flags); 1368 1369 list_for_each_entry_safe(ctx, tmp, &list, list) { 1370 /* Stop offloaded TX and switch to the fallback. 1371 * tls_is_sk_tx_device_offloaded will return false. 1372 */ 1373 WRITE_ONCE(ctx->sk->sk_validate_xmit_skb, tls_validate_xmit_skb_sw); 1374 1375 /* Stop the RX and TX resync. 1376 * tls_dev_resync must not be called after tls_dev_del. 1377 */ 1378 rcu_assign_pointer(ctx->netdev, NULL); 1379 1380 /* Start skipping the RX resync logic completely. */ 1381 set_bit(TLS_RX_DEV_DEGRADED, &ctx->flags); 1382 1383 /* Sync with inflight packets. After this point: 1384 * TX: no non-encrypted packets will be passed to the driver. 1385 * RX: resync requests from the driver will be ignored. 1386 */ 1387 synchronize_net(); 1388 1389 /* Release the offload context on the driver side. */ 1390 if (ctx->tx_conf == TLS_HW) 1391 netdev->tlsdev_ops->tls_dev_del(netdev, ctx, 1392 TLS_OFFLOAD_CTX_DIR_TX); 1393 if (ctx->rx_conf == TLS_HW && 1394 !test_bit(TLS_RX_DEV_CLOSED, &ctx->flags)) 1395 netdev->tlsdev_ops->tls_dev_del(netdev, ctx, 1396 TLS_OFFLOAD_CTX_DIR_RX); 1397 1398 dev_put(netdev); 1399 1400 /* Move the context to a separate list for two reasons: 1401 * 1. When the context is deallocated, list_del is called. 1402 * 2. It's no longer an offloaded context, so we don't want to 1403 * run offload-specific code on this context. 1404 */ 1405 spin_lock_irqsave(&tls_device_lock, flags); 1406 list_move_tail(&ctx->list, &tls_device_down_list); 1407 spin_unlock_irqrestore(&tls_device_lock, flags); 1408 1409 /* Device contexts for RX and TX will be freed in on sk_destruct 1410 * by tls_device_free_ctx. rx_conf and tx_conf stay in TLS_HW. 1411 * Now release the ref taken above. 1412 */ 1413 if (refcount_dec_and_test(&ctx->refcount)) { 1414 /* sk_destruct ran after tls_device_down took a ref, and 1415 * it returned early. Complete the destruction here. 1416 */ 1417 list_del(&ctx->list); 1418 tls_device_free_ctx(ctx); 1419 } 1420 } 1421 1422 up_write(&device_offload_lock); 1423 1424 flush_workqueue(destruct_wq); 1425 1426 return NOTIFY_DONE; 1427 } 1428 1429 static int tls_dev_event(struct notifier_block *this, unsigned long event, 1430 void *ptr) 1431 { 1432 struct net_device *dev = netdev_notifier_info_to_dev(ptr); 1433 1434 if (!dev->tlsdev_ops && 1435 !(dev->features & (NETIF_F_HW_TLS_RX | NETIF_F_HW_TLS_TX))) 1436 return NOTIFY_DONE; 1437 1438 switch (event) { 1439 case NETDEV_REGISTER: 1440 case NETDEV_FEAT_CHANGE: 1441 if (netif_is_bond_master(dev)) 1442 return NOTIFY_DONE; 1443 if ((dev->features & NETIF_F_HW_TLS_RX) && 1444 !dev->tlsdev_ops->tls_dev_resync) 1445 return NOTIFY_BAD; 1446 1447 if (dev->tlsdev_ops && 1448 dev->tlsdev_ops->tls_dev_add && 1449 dev->tlsdev_ops->tls_dev_del) 1450 return NOTIFY_DONE; 1451 else 1452 return NOTIFY_BAD; 1453 case NETDEV_DOWN: 1454 return tls_device_down(dev); 1455 } 1456 return NOTIFY_DONE; 1457 } 1458 1459 static struct notifier_block tls_dev_notifier = { 1460 .notifier_call = tls_dev_event, 1461 }; 1462 1463 int __init tls_device_init(void) 1464 { 1465 int err; 1466 1467 destruct_wq = alloc_workqueue("ktls_device_destruct", 0, 0); 1468 if (!destruct_wq) 1469 return -ENOMEM; 1470 1471 err = register_netdevice_notifier(&tls_dev_notifier); 1472 if (err) 1473 destroy_workqueue(destruct_wq); 1474 1475 return err; 1476 } 1477 1478 void __exit tls_device_cleanup(void) 1479 { 1480 unregister_netdevice_notifier(&tls_dev_notifier); 1481 destroy_workqueue(destruct_wq); 1482 clean_acked_data_flush(); 1483 } 1484