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 "trace.h" 42 43 /* device_offload_lock is used to synchronize tls_dev_add 44 * against NETDEV_DOWN notifications. 45 */ 46 static DECLARE_RWSEM(device_offload_lock); 47 48 static void tls_device_gc_task(struct work_struct *work); 49 50 static DECLARE_WORK(tls_device_gc_work, tls_device_gc_task); 51 static LIST_HEAD(tls_device_gc_list); 52 static LIST_HEAD(tls_device_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_gc_task(struct work_struct *work) 70 { 71 struct tls_context *ctx, *tmp; 72 unsigned long flags; 73 LIST_HEAD(gc_list); 74 75 spin_lock_irqsave(&tls_device_lock, flags); 76 list_splice_init(&tls_device_gc_list, &gc_list); 77 spin_unlock_irqrestore(&tls_device_lock, flags); 78 79 list_for_each_entry_safe(ctx, tmp, &gc_list, list) { 80 struct net_device *netdev = ctx->netdev; 81 82 if (netdev && ctx->tx_conf == TLS_HW) { 83 netdev->tlsdev_ops->tls_dev_del(netdev, ctx, 84 TLS_OFFLOAD_CTX_DIR_TX); 85 dev_put(netdev); 86 ctx->netdev = NULL; 87 } 88 89 list_del(&ctx->list); 90 tls_device_free_ctx(ctx); 91 } 92 } 93 94 static void tls_device_queue_ctx_destruction(struct tls_context *ctx) 95 { 96 unsigned long flags; 97 98 spin_lock_irqsave(&tls_device_lock, flags); 99 list_move_tail(&ctx->list, &tls_device_gc_list); 100 101 /* schedule_work inside the spinlock 102 * to make sure tls_device_down waits for that work. 103 */ 104 schedule_work(&tls_device_gc_work); 105 106 spin_unlock_irqrestore(&tls_device_lock, flags); 107 } 108 109 /* We assume that the socket is already connected */ 110 static struct net_device *get_netdev_for_sock(struct sock *sk) 111 { 112 struct dst_entry *dst = sk_dst_get(sk); 113 struct net_device *netdev = NULL; 114 115 if (likely(dst)) { 116 netdev = dst->dev; 117 dev_hold(netdev); 118 } 119 120 dst_release(dst); 121 122 return netdev; 123 } 124 125 static void destroy_record(struct tls_record_info *record) 126 { 127 int i; 128 129 for (i = 0; i < record->num_frags; i++) 130 __skb_frag_unref(&record->frags[i]); 131 kfree(record); 132 } 133 134 static void delete_all_records(struct tls_offload_context_tx *offload_ctx) 135 { 136 struct tls_record_info *info, *temp; 137 138 list_for_each_entry_safe(info, temp, &offload_ctx->records_list, list) { 139 list_del(&info->list); 140 destroy_record(info); 141 } 142 143 offload_ctx->retransmit_hint = NULL; 144 } 145 146 static void tls_icsk_clean_acked(struct sock *sk, u32 acked_seq) 147 { 148 struct tls_context *tls_ctx = tls_get_ctx(sk); 149 struct tls_record_info *info, *temp; 150 struct tls_offload_context_tx *ctx; 151 u64 deleted_records = 0; 152 unsigned long flags; 153 154 if (!tls_ctx) 155 return; 156 157 ctx = tls_offload_ctx_tx(tls_ctx); 158 159 spin_lock_irqsave(&ctx->lock, flags); 160 info = ctx->retransmit_hint; 161 if (info && !before(acked_seq, info->end_seq)) 162 ctx->retransmit_hint = NULL; 163 164 list_for_each_entry_safe(info, temp, &ctx->records_list, list) { 165 if (before(acked_seq, info->end_seq)) 166 break; 167 list_del(&info->list); 168 169 destroy_record(info); 170 deleted_records++; 171 } 172 173 ctx->unacked_record_sn += deleted_records; 174 spin_unlock_irqrestore(&ctx->lock, flags); 175 } 176 177 /* At this point, there should be no references on this 178 * socket and no in-flight SKBs associated with this 179 * socket, so it is safe to free all the resources. 180 */ 181 void tls_device_sk_destruct(struct sock *sk) 182 { 183 struct tls_context *tls_ctx = tls_get_ctx(sk); 184 struct tls_offload_context_tx *ctx = tls_offload_ctx_tx(tls_ctx); 185 186 tls_ctx->sk_destruct(sk); 187 188 if (tls_ctx->tx_conf == TLS_HW) { 189 if (ctx->open_record) 190 destroy_record(ctx->open_record); 191 delete_all_records(ctx); 192 crypto_free_aead(ctx->aead_send); 193 clean_acked_data_disable(inet_csk(sk)); 194 } 195 196 if (refcount_dec_and_test(&tls_ctx->refcount)) 197 tls_device_queue_ctx_destruction(tls_ctx); 198 } 199 EXPORT_SYMBOL_GPL(tls_device_sk_destruct); 200 201 void tls_device_free_resources_tx(struct sock *sk) 202 { 203 struct tls_context *tls_ctx = tls_get_ctx(sk); 204 205 tls_free_partial_record(sk, tls_ctx); 206 } 207 208 void tls_offload_tx_resync_request(struct sock *sk, u32 got_seq, u32 exp_seq) 209 { 210 struct tls_context *tls_ctx = tls_get_ctx(sk); 211 212 trace_tls_device_tx_resync_req(sk, got_seq, exp_seq); 213 WARN_ON(test_and_set_bit(TLS_TX_SYNC_SCHED, &tls_ctx->flags)); 214 } 215 EXPORT_SYMBOL_GPL(tls_offload_tx_resync_request); 216 217 static void tls_device_resync_tx(struct sock *sk, struct tls_context *tls_ctx, 218 u32 seq) 219 { 220 struct net_device *netdev; 221 struct sk_buff *skb; 222 int err = 0; 223 u8 *rcd_sn; 224 225 skb = tcp_write_queue_tail(sk); 226 if (skb) 227 TCP_SKB_CB(skb)->eor = 1; 228 229 rcd_sn = tls_ctx->tx.rec_seq; 230 231 trace_tls_device_tx_resync_send(sk, seq, rcd_sn); 232 down_read(&device_offload_lock); 233 netdev = tls_ctx->netdev; 234 if (netdev) 235 err = netdev->tlsdev_ops->tls_dev_resync(netdev, sk, seq, 236 rcd_sn, 237 TLS_OFFLOAD_CTX_DIR_TX); 238 up_read(&device_offload_lock); 239 if (err) 240 return; 241 242 clear_bit_unlock(TLS_TX_SYNC_SCHED, &tls_ctx->flags); 243 } 244 245 static void tls_append_frag(struct tls_record_info *record, 246 struct page_frag *pfrag, 247 int size) 248 { 249 skb_frag_t *frag; 250 251 frag = &record->frags[record->num_frags - 1]; 252 if (skb_frag_page(frag) == pfrag->page && 253 skb_frag_off(frag) + skb_frag_size(frag) == pfrag->offset) { 254 skb_frag_size_add(frag, size); 255 } else { 256 ++frag; 257 __skb_frag_set_page(frag, pfrag->page); 258 skb_frag_off_set(frag, pfrag->offset); 259 skb_frag_size_set(frag, size); 260 ++record->num_frags; 261 get_page(pfrag->page); 262 } 263 264 pfrag->offset += size; 265 record->len += size; 266 } 267 268 static int tls_push_record(struct sock *sk, 269 struct tls_context *ctx, 270 struct tls_offload_context_tx *offload_ctx, 271 struct tls_record_info *record, 272 int flags) 273 { 274 struct tls_prot_info *prot = &ctx->prot_info; 275 struct tcp_sock *tp = tcp_sk(sk); 276 skb_frag_t *frag; 277 int i; 278 279 record->end_seq = tp->write_seq + record->len; 280 list_add_tail_rcu(&record->list, &offload_ctx->records_list); 281 offload_ctx->open_record = NULL; 282 283 if (test_bit(TLS_TX_SYNC_SCHED, &ctx->flags)) 284 tls_device_resync_tx(sk, ctx, tp->write_seq); 285 286 tls_advance_record_sn(sk, prot, &ctx->tx); 287 288 for (i = 0; i < record->num_frags; i++) { 289 frag = &record->frags[i]; 290 sg_unmark_end(&offload_ctx->sg_tx_data[i]); 291 sg_set_page(&offload_ctx->sg_tx_data[i], skb_frag_page(frag), 292 skb_frag_size(frag), skb_frag_off(frag)); 293 sk_mem_charge(sk, skb_frag_size(frag)); 294 get_page(skb_frag_page(frag)); 295 } 296 sg_mark_end(&offload_ctx->sg_tx_data[record->num_frags - 1]); 297 298 /* all ready, send */ 299 return tls_push_sg(sk, ctx, offload_ctx->sg_tx_data, 0, flags); 300 } 301 302 static int tls_device_record_close(struct sock *sk, 303 struct tls_context *ctx, 304 struct tls_record_info *record, 305 struct page_frag *pfrag, 306 unsigned char record_type) 307 { 308 struct tls_prot_info *prot = &ctx->prot_info; 309 int ret; 310 311 /* append tag 312 * device will fill in the tag, we just need to append a placeholder 313 * use socket memory to improve coalescing (re-using a single buffer 314 * increases frag count) 315 * if we can't allocate memory now, steal some back from data 316 */ 317 if (likely(skb_page_frag_refill(prot->tag_size, pfrag, 318 sk->sk_allocation))) { 319 ret = 0; 320 tls_append_frag(record, pfrag, prot->tag_size); 321 } else { 322 ret = prot->tag_size; 323 if (record->len <= prot->overhead_size) 324 return -ENOMEM; 325 } 326 327 /* fill prepend */ 328 tls_fill_prepend(ctx, skb_frag_address(&record->frags[0]), 329 record->len - prot->overhead_size, 330 record_type, prot->version); 331 return ret; 332 } 333 334 static int tls_create_new_record(struct tls_offload_context_tx *offload_ctx, 335 struct page_frag *pfrag, 336 size_t prepend_size) 337 { 338 struct tls_record_info *record; 339 skb_frag_t *frag; 340 341 record = kmalloc(sizeof(*record), GFP_KERNEL); 342 if (!record) 343 return -ENOMEM; 344 345 frag = &record->frags[0]; 346 __skb_frag_set_page(frag, pfrag->page); 347 skb_frag_off_set(frag, pfrag->offset); 348 skb_frag_size_set(frag, prepend_size); 349 350 get_page(pfrag->page); 351 pfrag->offset += prepend_size; 352 353 record->num_frags = 1; 354 record->len = prepend_size; 355 offload_ctx->open_record = record; 356 return 0; 357 } 358 359 static int tls_do_allocation(struct sock *sk, 360 struct tls_offload_context_tx *offload_ctx, 361 struct page_frag *pfrag, 362 size_t prepend_size) 363 { 364 int ret; 365 366 if (!offload_ctx->open_record) { 367 if (unlikely(!skb_page_frag_refill(prepend_size, pfrag, 368 sk->sk_allocation))) { 369 READ_ONCE(sk->sk_prot)->enter_memory_pressure(sk); 370 sk_stream_moderate_sndbuf(sk); 371 return -ENOMEM; 372 } 373 374 ret = tls_create_new_record(offload_ctx, pfrag, prepend_size); 375 if (ret) 376 return ret; 377 378 if (pfrag->size > pfrag->offset) 379 return 0; 380 } 381 382 if (!sk_page_frag_refill(sk, pfrag)) 383 return -ENOMEM; 384 385 return 0; 386 } 387 388 static int tls_device_copy_data(void *addr, size_t bytes, struct iov_iter *i) 389 { 390 size_t pre_copy, nocache; 391 392 pre_copy = ~((unsigned long)addr - 1) & (SMP_CACHE_BYTES - 1); 393 if (pre_copy) { 394 pre_copy = min(pre_copy, bytes); 395 if (copy_from_iter(addr, pre_copy, i) != pre_copy) 396 return -EFAULT; 397 bytes -= pre_copy; 398 addr += pre_copy; 399 } 400 401 nocache = round_down(bytes, SMP_CACHE_BYTES); 402 if (copy_from_iter_nocache(addr, nocache, i) != nocache) 403 return -EFAULT; 404 bytes -= nocache; 405 addr += nocache; 406 407 if (bytes && copy_from_iter(addr, bytes, i) != bytes) 408 return -EFAULT; 409 410 return 0; 411 } 412 413 static int tls_push_data(struct sock *sk, 414 struct iov_iter *msg_iter, 415 size_t size, int flags, 416 unsigned char record_type) 417 { 418 struct tls_context *tls_ctx = tls_get_ctx(sk); 419 struct tls_prot_info *prot = &tls_ctx->prot_info; 420 struct tls_offload_context_tx *ctx = tls_offload_ctx_tx(tls_ctx); 421 int more = flags & (MSG_SENDPAGE_NOTLAST | MSG_MORE); 422 struct tls_record_info *record = ctx->open_record; 423 int tls_push_record_flags; 424 struct page_frag *pfrag; 425 size_t orig_size = size; 426 u32 max_open_record_len; 427 int copy, rc = 0; 428 bool done = false; 429 long timeo; 430 431 if (flags & 432 ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL | MSG_SENDPAGE_NOTLAST)) 433 return -EOPNOTSUPP; 434 435 if (unlikely(sk->sk_err)) 436 return -sk->sk_err; 437 438 flags |= MSG_SENDPAGE_DECRYPTED; 439 tls_push_record_flags = flags | MSG_SENDPAGE_NOTLAST; 440 441 timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT); 442 if (tls_is_partially_sent_record(tls_ctx)) { 443 rc = tls_push_partial_record(sk, tls_ctx, flags); 444 if (rc < 0) 445 return rc; 446 } 447 448 pfrag = sk_page_frag(sk); 449 450 /* TLS_HEADER_SIZE is not counted as part of the TLS record, and 451 * we need to leave room for an authentication tag. 452 */ 453 max_open_record_len = TLS_MAX_PAYLOAD_SIZE + 454 prot->prepend_size; 455 do { 456 rc = tls_do_allocation(sk, ctx, pfrag, prot->prepend_size); 457 if (unlikely(rc)) { 458 rc = sk_stream_wait_memory(sk, &timeo); 459 if (!rc) 460 continue; 461 462 record = ctx->open_record; 463 if (!record) 464 break; 465 handle_error: 466 if (record_type != TLS_RECORD_TYPE_DATA) { 467 /* avoid sending partial 468 * record with type != 469 * application_data 470 */ 471 size = orig_size; 472 destroy_record(record); 473 ctx->open_record = NULL; 474 } else if (record->len > prot->prepend_size) { 475 goto last_record; 476 } 477 478 break; 479 } 480 481 record = ctx->open_record; 482 copy = min_t(size_t, size, (pfrag->size - pfrag->offset)); 483 copy = min_t(size_t, copy, (max_open_record_len - record->len)); 484 485 rc = tls_device_copy_data(page_address(pfrag->page) + 486 pfrag->offset, copy, msg_iter); 487 if (rc) 488 goto handle_error; 489 tls_append_frag(record, pfrag, copy); 490 491 size -= copy; 492 if (!size) { 493 last_record: 494 tls_push_record_flags = flags; 495 if (more) { 496 tls_ctx->pending_open_record_frags = 497 !!record->num_frags; 498 break; 499 } 500 501 done = true; 502 } 503 504 if (done || record->len >= max_open_record_len || 505 (record->num_frags >= MAX_SKB_FRAGS - 1)) { 506 rc = tls_device_record_close(sk, tls_ctx, record, 507 pfrag, record_type); 508 if (rc) { 509 if (rc > 0) { 510 size += rc; 511 } else { 512 size = orig_size; 513 destroy_record(record); 514 ctx->open_record = NULL; 515 break; 516 } 517 } 518 519 rc = tls_push_record(sk, 520 tls_ctx, 521 ctx, 522 record, 523 tls_push_record_flags); 524 if (rc < 0) 525 break; 526 } 527 } while (!done); 528 529 if (orig_size - size > 0) 530 rc = orig_size - size; 531 532 return rc; 533 } 534 535 int tls_device_sendmsg(struct sock *sk, struct msghdr *msg, size_t size) 536 { 537 unsigned char record_type = TLS_RECORD_TYPE_DATA; 538 struct tls_context *tls_ctx = tls_get_ctx(sk); 539 int rc; 540 541 mutex_lock(&tls_ctx->tx_lock); 542 lock_sock(sk); 543 544 if (unlikely(msg->msg_controllen)) { 545 rc = tls_proccess_cmsg(sk, msg, &record_type); 546 if (rc) 547 goto out; 548 } 549 550 rc = tls_push_data(sk, &msg->msg_iter, size, 551 msg->msg_flags, record_type); 552 553 out: 554 release_sock(sk); 555 mutex_unlock(&tls_ctx->tx_lock); 556 return rc; 557 } 558 559 int tls_device_sendpage(struct sock *sk, struct page *page, 560 int offset, size_t size, int flags) 561 { 562 struct tls_context *tls_ctx = tls_get_ctx(sk); 563 struct iov_iter msg_iter; 564 char *kaddr = kmap(page); 565 struct kvec iov; 566 int rc; 567 568 if (flags & MSG_SENDPAGE_NOTLAST) 569 flags |= MSG_MORE; 570 571 mutex_lock(&tls_ctx->tx_lock); 572 lock_sock(sk); 573 574 if (flags & MSG_OOB) { 575 rc = -EOPNOTSUPP; 576 goto out; 577 } 578 579 iov.iov_base = kaddr + offset; 580 iov.iov_len = size; 581 iov_iter_kvec(&msg_iter, WRITE, &iov, 1, size); 582 rc = tls_push_data(sk, &msg_iter, size, 583 flags, TLS_RECORD_TYPE_DATA); 584 kunmap(page); 585 586 out: 587 release_sock(sk); 588 mutex_unlock(&tls_ctx->tx_lock); 589 return rc; 590 } 591 592 struct tls_record_info *tls_get_record(struct tls_offload_context_tx *context, 593 u32 seq, u64 *p_record_sn) 594 { 595 u64 record_sn = context->hint_record_sn; 596 struct tls_record_info *info, *last; 597 598 info = context->retransmit_hint; 599 if (!info || 600 before(seq, info->end_seq - info->len)) { 601 /* if retransmit_hint is irrelevant start 602 * from the beggining of the list 603 */ 604 info = list_first_entry_or_null(&context->records_list, 605 struct tls_record_info, list); 606 if (!info) 607 return NULL; 608 /* send the start_marker record if seq number is before the 609 * tls offload start marker sequence number. This record is 610 * required to handle TCP packets which are before TLS offload 611 * started. 612 * And if it's not start marker, look if this seq number 613 * belongs to the list. 614 */ 615 if (likely(!tls_record_is_start_marker(info))) { 616 /* we have the first record, get the last record to see 617 * if this seq number belongs to the list. 618 */ 619 last = list_last_entry(&context->records_list, 620 struct tls_record_info, list); 621 622 if (!between(seq, tls_record_start_seq(info), 623 last->end_seq)) 624 return NULL; 625 } 626 record_sn = context->unacked_record_sn; 627 } 628 629 /* We just need the _rcu for the READ_ONCE() */ 630 rcu_read_lock(); 631 list_for_each_entry_from_rcu(info, &context->records_list, list) { 632 if (before(seq, info->end_seq)) { 633 if (!context->retransmit_hint || 634 after(info->end_seq, 635 context->retransmit_hint->end_seq)) { 636 context->hint_record_sn = record_sn; 637 context->retransmit_hint = info; 638 } 639 *p_record_sn = record_sn; 640 goto exit_rcu_unlock; 641 } 642 record_sn++; 643 } 644 info = NULL; 645 646 exit_rcu_unlock: 647 rcu_read_unlock(); 648 return info; 649 } 650 EXPORT_SYMBOL(tls_get_record); 651 652 static int tls_device_push_pending_record(struct sock *sk, int flags) 653 { 654 struct iov_iter msg_iter; 655 656 iov_iter_kvec(&msg_iter, WRITE, NULL, 0, 0); 657 return tls_push_data(sk, &msg_iter, 0, flags, TLS_RECORD_TYPE_DATA); 658 } 659 660 void tls_device_write_space(struct sock *sk, struct tls_context *ctx) 661 { 662 if (tls_is_partially_sent_record(ctx)) { 663 gfp_t sk_allocation = sk->sk_allocation; 664 665 WARN_ON_ONCE(sk->sk_write_pending); 666 667 sk->sk_allocation = GFP_ATOMIC; 668 tls_push_partial_record(sk, ctx, 669 MSG_DONTWAIT | MSG_NOSIGNAL | 670 MSG_SENDPAGE_DECRYPTED); 671 sk->sk_allocation = sk_allocation; 672 } 673 } 674 675 static void tls_device_resync_rx(struct tls_context *tls_ctx, 676 struct sock *sk, u32 seq, u8 *rcd_sn) 677 { 678 struct tls_offload_context_rx *rx_ctx = tls_offload_ctx_rx(tls_ctx); 679 struct net_device *netdev; 680 681 if (WARN_ON(test_and_set_bit(TLS_RX_SYNC_RUNNING, &tls_ctx->flags))) 682 return; 683 684 trace_tls_device_rx_resync_send(sk, seq, rcd_sn, rx_ctx->resync_type); 685 netdev = READ_ONCE(tls_ctx->netdev); 686 if (netdev) 687 netdev->tlsdev_ops->tls_dev_resync(netdev, sk, seq, rcd_sn, 688 TLS_OFFLOAD_CTX_DIR_RX); 689 clear_bit_unlock(TLS_RX_SYNC_RUNNING, &tls_ctx->flags); 690 TLS_INC_STATS(sock_net(sk), LINUX_MIB_TLSRXDEVICERESYNC); 691 } 692 693 void tls_device_rx_resync_new_rec(struct sock *sk, u32 rcd_len, u32 seq) 694 { 695 struct tls_context *tls_ctx = tls_get_ctx(sk); 696 struct tls_offload_context_rx *rx_ctx; 697 u8 rcd_sn[TLS_MAX_REC_SEQ_SIZE]; 698 u32 sock_data, is_req_pending; 699 struct tls_prot_info *prot; 700 s64 resync_req; 701 u32 req_seq; 702 703 if (tls_ctx->rx_conf != TLS_HW) 704 return; 705 706 prot = &tls_ctx->prot_info; 707 rx_ctx = tls_offload_ctx_rx(tls_ctx); 708 memcpy(rcd_sn, tls_ctx->rx.rec_seq, prot->rec_seq_size); 709 710 switch (rx_ctx->resync_type) { 711 case TLS_OFFLOAD_SYNC_TYPE_DRIVER_REQ: 712 resync_req = atomic64_read(&rx_ctx->resync_req); 713 req_seq = resync_req >> 32; 714 seq += TLS_HEADER_SIZE - 1; 715 is_req_pending = resync_req; 716 717 if (likely(!is_req_pending) || req_seq != seq || 718 !atomic64_try_cmpxchg(&rx_ctx->resync_req, &resync_req, 0)) 719 return; 720 break; 721 case TLS_OFFLOAD_SYNC_TYPE_CORE_NEXT_HINT: 722 if (likely(!rx_ctx->resync_nh_do_now)) 723 return; 724 725 /* head of next rec is already in, note that the sock_inq will 726 * include the currently parsed message when called from parser 727 */ 728 sock_data = tcp_inq(sk); 729 if (sock_data > rcd_len) { 730 trace_tls_device_rx_resync_nh_delay(sk, sock_data, 731 rcd_len); 732 return; 733 } 734 735 rx_ctx->resync_nh_do_now = 0; 736 seq += rcd_len; 737 tls_bigint_increment(rcd_sn, prot->rec_seq_size); 738 break; 739 } 740 741 tls_device_resync_rx(tls_ctx, sk, seq, rcd_sn); 742 } 743 744 static void tls_device_core_ctrl_rx_resync(struct tls_context *tls_ctx, 745 struct tls_offload_context_rx *ctx, 746 struct sock *sk, struct sk_buff *skb) 747 { 748 struct strp_msg *rxm; 749 750 /* device will request resyncs by itself based on stream scan */ 751 if (ctx->resync_type != TLS_OFFLOAD_SYNC_TYPE_CORE_NEXT_HINT) 752 return; 753 /* already scheduled */ 754 if (ctx->resync_nh_do_now) 755 return; 756 /* seen decrypted fragments since last fully-failed record */ 757 if (ctx->resync_nh_reset) { 758 ctx->resync_nh_reset = 0; 759 ctx->resync_nh.decrypted_failed = 1; 760 ctx->resync_nh.decrypted_tgt = TLS_DEVICE_RESYNC_NH_START_IVAL; 761 return; 762 } 763 764 if (++ctx->resync_nh.decrypted_failed <= ctx->resync_nh.decrypted_tgt) 765 return; 766 767 /* doing resync, bump the next target in case it fails */ 768 if (ctx->resync_nh.decrypted_tgt < TLS_DEVICE_RESYNC_NH_MAX_IVAL) 769 ctx->resync_nh.decrypted_tgt *= 2; 770 else 771 ctx->resync_nh.decrypted_tgt += TLS_DEVICE_RESYNC_NH_MAX_IVAL; 772 773 rxm = strp_msg(skb); 774 775 /* head of next rec is already in, parser will sync for us */ 776 if (tcp_inq(sk) > rxm->full_len) { 777 trace_tls_device_rx_resync_nh_schedule(sk); 778 ctx->resync_nh_do_now = 1; 779 } else { 780 struct tls_prot_info *prot = &tls_ctx->prot_info; 781 u8 rcd_sn[TLS_MAX_REC_SEQ_SIZE]; 782 783 memcpy(rcd_sn, tls_ctx->rx.rec_seq, prot->rec_seq_size); 784 tls_bigint_increment(rcd_sn, prot->rec_seq_size); 785 786 tls_device_resync_rx(tls_ctx, sk, tcp_sk(sk)->copied_seq, 787 rcd_sn); 788 } 789 } 790 791 static int tls_device_reencrypt(struct sock *sk, struct sk_buff *skb) 792 { 793 struct strp_msg *rxm = strp_msg(skb); 794 int err = 0, offset = rxm->offset, copy, nsg, data_len, pos; 795 struct sk_buff *skb_iter, *unused; 796 struct scatterlist sg[1]; 797 char *orig_buf, *buf; 798 799 orig_buf = kmalloc(rxm->full_len + TLS_HEADER_SIZE + 800 TLS_CIPHER_AES_GCM_128_IV_SIZE, sk->sk_allocation); 801 if (!orig_buf) 802 return -ENOMEM; 803 buf = orig_buf; 804 805 nsg = skb_cow_data(skb, 0, &unused); 806 if (unlikely(nsg < 0)) { 807 err = nsg; 808 goto free_buf; 809 } 810 811 sg_init_table(sg, 1); 812 sg_set_buf(&sg[0], buf, 813 rxm->full_len + TLS_HEADER_SIZE + 814 TLS_CIPHER_AES_GCM_128_IV_SIZE); 815 err = skb_copy_bits(skb, offset, buf, 816 TLS_HEADER_SIZE + TLS_CIPHER_AES_GCM_128_IV_SIZE); 817 if (err) 818 goto free_buf; 819 820 /* We are interested only in the decrypted data not the auth */ 821 err = decrypt_skb(sk, skb, sg); 822 if (err != -EBADMSG) 823 goto free_buf; 824 else 825 err = 0; 826 827 data_len = rxm->full_len - TLS_CIPHER_AES_GCM_128_TAG_SIZE; 828 829 if (skb_pagelen(skb) > offset) { 830 copy = min_t(int, skb_pagelen(skb) - offset, data_len); 831 832 if (skb->decrypted) { 833 err = skb_store_bits(skb, offset, buf, copy); 834 if (err) 835 goto free_buf; 836 } 837 838 offset += copy; 839 buf += copy; 840 } 841 842 pos = skb_pagelen(skb); 843 skb_walk_frags(skb, skb_iter) { 844 int frag_pos; 845 846 /* Practically all frags must belong to msg if reencrypt 847 * is needed with current strparser and coalescing logic, 848 * but strparser may "get optimized", so let's be safe. 849 */ 850 if (pos + skb_iter->len <= offset) 851 goto done_with_frag; 852 if (pos >= data_len + rxm->offset) 853 break; 854 855 frag_pos = offset - pos; 856 copy = min_t(int, skb_iter->len - frag_pos, 857 data_len + rxm->offset - offset); 858 859 if (skb_iter->decrypted) { 860 err = skb_store_bits(skb_iter, frag_pos, buf, copy); 861 if (err) 862 goto free_buf; 863 } 864 865 offset += copy; 866 buf += copy; 867 done_with_frag: 868 pos += skb_iter->len; 869 } 870 871 free_buf: 872 kfree(orig_buf); 873 return err; 874 } 875 876 int tls_device_decrypted(struct sock *sk, struct tls_context *tls_ctx, 877 struct sk_buff *skb, struct strp_msg *rxm) 878 { 879 struct tls_offload_context_rx *ctx = tls_offload_ctx_rx(tls_ctx); 880 int is_decrypted = skb->decrypted; 881 int is_encrypted = !is_decrypted; 882 struct sk_buff *skb_iter; 883 884 /* Check if all the data is decrypted already */ 885 skb_walk_frags(skb, skb_iter) { 886 is_decrypted &= skb_iter->decrypted; 887 is_encrypted &= !skb_iter->decrypted; 888 } 889 890 trace_tls_device_decrypted(sk, tcp_sk(sk)->copied_seq - rxm->full_len, 891 tls_ctx->rx.rec_seq, rxm->full_len, 892 is_encrypted, is_decrypted); 893 894 ctx->sw.decrypted |= is_decrypted; 895 896 /* Return immediately if the record is either entirely plaintext or 897 * entirely ciphertext. Otherwise handle reencrypt partially decrypted 898 * record. 899 */ 900 if (is_decrypted) { 901 ctx->resync_nh_reset = 1; 902 return 0; 903 } 904 if (is_encrypted) { 905 tls_device_core_ctrl_rx_resync(tls_ctx, ctx, sk, skb); 906 return 0; 907 } 908 909 ctx->resync_nh_reset = 1; 910 return tls_device_reencrypt(sk, skb); 911 } 912 913 static void tls_device_attach(struct tls_context *ctx, struct sock *sk, 914 struct net_device *netdev) 915 { 916 if (sk->sk_destruct != tls_device_sk_destruct) { 917 refcount_set(&ctx->refcount, 1); 918 dev_hold(netdev); 919 ctx->netdev = netdev; 920 spin_lock_irq(&tls_device_lock); 921 list_add_tail(&ctx->list, &tls_device_list); 922 spin_unlock_irq(&tls_device_lock); 923 924 ctx->sk_destruct = sk->sk_destruct; 925 smp_store_release(&sk->sk_destruct, tls_device_sk_destruct); 926 } 927 } 928 929 int tls_set_device_offload(struct sock *sk, struct tls_context *ctx) 930 { 931 u16 nonce_size, tag_size, iv_size, rec_seq_size; 932 struct tls_context *tls_ctx = tls_get_ctx(sk); 933 struct tls_prot_info *prot = &tls_ctx->prot_info; 934 struct tls_record_info *start_marker_record; 935 struct tls_offload_context_tx *offload_ctx; 936 struct tls_crypto_info *crypto_info; 937 struct net_device *netdev; 938 char *iv, *rec_seq; 939 struct sk_buff *skb; 940 __be64 rcd_sn; 941 int rc; 942 943 if (!ctx) 944 return -EINVAL; 945 946 if (ctx->priv_ctx_tx) 947 return -EEXIST; 948 949 start_marker_record = kmalloc(sizeof(*start_marker_record), GFP_KERNEL); 950 if (!start_marker_record) 951 return -ENOMEM; 952 953 offload_ctx = kzalloc(TLS_OFFLOAD_CONTEXT_SIZE_TX, GFP_KERNEL); 954 if (!offload_ctx) { 955 rc = -ENOMEM; 956 goto free_marker_record; 957 } 958 959 crypto_info = &ctx->crypto_send.info; 960 if (crypto_info->version != TLS_1_2_VERSION) { 961 rc = -EOPNOTSUPP; 962 goto free_offload_ctx; 963 } 964 965 switch (crypto_info->cipher_type) { 966 case TLS_CIPHER_AES_GCM_128: 967 nonce_size = TLS_CIPHER_AES_GCM_128_IV_SIZE; 968 tag_size = TLS_CIPHER_AES_GCM_128_TAG_SIZE; 969 iv_size = TLS_CIPHER_AES_GCM_128_IV_SIZE; 970 iv = ((struct tls12_crypto_info_aes_gcm_128 *)crypto_info)->iv; 971 rec_seq_size = TLS_CIPHER_AES_GCM_128_REC_SEQ_SIZE; 972 rec_seq = 973 ((struct tls12_crypto_info_aes_gcm_128 *)crypto_info)->rec_seq; 974 break; 975 default: 976 rc = -EINVAL; 977 goto free_offload_ctx; 978 } 979 980 /* Sanity-check the rec_seq_size for stack allocations */ 981 if (rec_seq_size > TLS_MAX_REC_SEQ_SIZE) { 982 rc = -EINVAL; 983 goto free_offload_ctx; 984 } 985 986 prot->version = crypto_info->version; 987 prot->cipher_type = crypto_info->cipher_type; 988 prot->prepend_size = TLS_HEADER_SIZE + nonce_size; 989 prot->tag_size = tag_size; 990 prot->overhead_size = prot->prepend_size + prot->tag_size; 991 prot->iv_size = iv_size; 992 ctx->tx.iv = kmalloc(iv_size + TLS_CIPHER_AES_GCM_128_SALT_SIZE, 993 GFP_KERNEL); 994 if (!ctx->tx.iv) { 995 rc = -ENOMEM; 996 goto free_offload_ctx; 997 } 998 999 memcpy(ctx->tx.iv + TLS_CIPHER_AES_GCM_128_SALT_SIZE, iv, iv_size); 1000 1001 prot->rec_seq_size = rec_seq_size; 1002 ctx->tx.rec_seq = kmemdup(rec_seq, rec_seq_size, GFP_KERNEL); 1003 if (!ctx->tx.rec_seq) { 1004 rc = -ENOMEM; 1005 goto free_iv; 1006 } 1007 1008 rc = tls_sw_fallback_init(sk, offload_ctx, crypto_info); 1009 if (rc) 1010 goto free_rec_seq; 1011 1012 /* start at rec_seq - 1 to account for the start marker record */ 1013 memcpy(&rcd_sn, ctx->tx.rec_seq, sizeof(rcd_sn)); 1014 offload_ctx->unacked_record_sn = be64_to_cpu(rcd_sn) - 1; 1015 1016 start_marker_record->end_seq = tcp_sk(sk)->write_seq; 1017 start_marker_record->len = 0; 1018 start_marker_record->num_frags = 0; 1019 1020 INIT_LIST_HEAD(&offload_ctx->records_list); 1021 list_add_tail(&start_marker_record->list, &offload_ctx->records_list); 1022 spin_lock_init(&offload_ctx->lock); 1023 sg_init_table(offload_ctx->sg_tx_data, 1024 ARRAY_SIZE(offload_ctx->sg_tx_data)); 1025 1026 clean_acked_data_enable(inet_csk(sk), &tls_icsk_clean_acked); 1027 ctx->push_pending_record = tls_device_push_pending_record; 1028 1029 /* TLS offload is greatly simplified if we don't send 1030 * SKBs where only part of the payload needs to be encrypted. 1031 * So mark the last skb in the write queue as end of record. 1032 */ 1033 skb = tcp_write_queue_tail(sk); 1034 if (skb) 1035 TCP_SKB_CB(skb)->eor = 1; 1036 1037 netdev = get_netdev_for_sock(sk); 1038 if (!netdev) { 1039 pr_err_ratelimited("%s: netdev not found\n", __func__); 1040 rc = -EINVAL; 1041 goto disable_cad; 1042 } 1043 1044 if (!(netdev->features & NETIF_F_HW_TLS_TX)) { 1045 rc = -EOPNOTSUPP; 1046 goto release_netdev; 1047 } 1048 1049 /* Avoid offloading if the device is down 1050 * We don't want to offload new flows after 1051 * the NETDEV_DOWN event 1052 * 1053 * device_offload_lock is taken in tls_devices's NETDEV_DOWN 1054 * handler thus protecting from the device going down before 1055 * ctx was added to tls_device_list. 1056 */ 1057 down_read(&device_offload_lock); 1058 if (!(netdev->flags & IFF_UP)) { 1059 rc = -EINVAL; 1060 goto release_lock; 1061 } 1062 1063 ctx->priv_ctx_tx = offload_ctx; 1064 rc = netdev->tlsdev_ops->tls_dev_add(netdev, sk, TLS_OFFLOAD_CTX_DIR_TX, 1065 &ctx->crypto_send.info, 1066 tcp_sk(sk)->write_seq); 1067 trace_tls_device_offload_set(sk, TLS_OFFLOAD_CTX_DIR_TX, 1068 tcp_sk(sk)->write_seq, rec_seq, rc); 1069 if (rc) 1070 goto release_lock; 1071 1072 tls_device_attach(ctx, sk, netdev); 1073 up_read(&device_offload_lock); 1074 1075 /* following this assignment tls_is_sk_tx_device_offloaded 1076 * will return true and the context might be accessed 1077 * by the netdev's xmit function. 1078 */ 1079 smp_store_release(&sk->sk_validate_xmit_skb, tls_validate_xmit_skb); 1080 dev_put(netdev); 1081 1082 return 0; 1083 1084 release_lock: 1085 up_read(&device_offload_lock); 1086 release_netdev: 1087 dev_put(netdev); 1088 disable_cad: 1089 clean_acked_data_disable(inet_csk(sk)); 1090 crypto_free_aead(offload_ctx->aead_send); 1091 free_rec_seq: 1092 kfree(ctx->tx.rec_seq); 1093 free_iv: 1094 kfree(ctx->tx.iv); 1095 free_offload_ctx: 1096 kfree(offload_ctx); 1097 ctx->priv_ctx_tx = NULL; 1098 free_marker_record: 1099 kfree(start_marker_record); 1100 return rc; 1101 } 1102 1103 int tls_set_device_offload_rx(struct sock *sk, struct tls_context *ctx) 1104 { 1105 struct tls12_crypto_info_aes_gcm_128 *info; 1106 struct tls_offload_context_rx *context; 1107 struct net_device *netdev; 1108 int rc = 0; 1109 1110 if (ctx->crypto_recv.info.version != TLS_1_2_VERSION) 1111 return -EOPNOTSUPP; 1112 1113 netdev = get_netdev_for_sock(sk); 1114 if (!netdev) { 1115 pr_err_ratelimited("%s: netdev not found\n", __func__); 1116 return -EINVAL; 1117 } 1118 1119 if (!(netdev->features & NETIF_F_HW_TLS_RX)) { 1120 rc = -EOPNOTSUPP; 1121 goto release_netdev; 1122 } 1123 1124 /* Avoid offloading if the device is down 1125 * We don't want to offload new flows after 1126 * the NETDEV_DOWN event 1127 * 1128 * device_offload_lock is taken in tls_devices's NETDEV_DOWN 1129 * handler thus protecting from the device going down before 1130 * ctx was added to tls_device_list. 1131 */ 1132 down_read(&device_offload_lock); 1133 if (!(netdev->flags & IFF_UP)) { 1134 rc = -EINVAL; 1135 goto release_lock; 1136 } 1137 1138 context = kzalloc(TLS_OFFLOAD_CONTEXT_SIZE_RX, GFP_KERNEL); 1139 if (!context) { 1140 rc = -ENOMEM; 1141 goto release_lock; 1142 } 1143 context->resync_nh_reset = 1; 1144 1145 ctx->priv_ctx_rx = context; 1146 rc = tls_set_sw_offload(sk, ctx, 0); 1147 if (rc) 1148 goto release_ctx; 1149 1150 rc = netdev->tlsdev_ops->tls_dev_add(netdev, sk, TLS_OFFLOAD_CTX_DIR_RX, 1151 &ctx->crypto_recv.info, 1152 tcp_sk(sk)->copied_seq); 1153 info = (void *)&ctx->crypto_recv.info; 1154 trace_tls_device_offload_set(sk, TLS_OFFLOAD_CTX_DIR_RX, 1155 tcp_sk(sk)->copied_seq, info->rec_seq, rc); 1156 if (rc) 1157 goto free_sw_resources; 1158 1159 tls_device_attach(ctx, sk, netdev); 1160 up_read(&device_offload_lock); 1161 1162 dev_put(netdev); 1163 1164 return 0; 1165 1166 free_sw_resources: 1167 up_read(&device_offload_lock); 1168 tls_sw_free_resources_rx(sk); 1169 down_read(&device_offload_lock); 1170 release_ctx: 1171 ctx->priv_ctx_rx = NULL; 1172 release_lock: 1173 up_read(&device_offload_lock); 1174 release_netdev: 1175 dev_put(netdev); 1176 return rc; 1177 } 1178 1179 void tls_device_offload_cleanup_rx(struct sock *sk) 1180 { 1181 struct tls_context *tls_ctx = tls_get_ctx(sk); 1182 struct net_device *netdev; 1183 1184 down_read(&device_offload_lock); 1185 netdev = tls_ctx->netdev; 1186 if (!netdev) 1187 goto out; 1188 1189 netdev->tlsdev_ops->tls_dev_del(netdev, tls_ctx, 1190 TLS_OFFLOAD_CTX_DIR_RX); 1191 1192 if (tls_ctx->tx_conf != TLS_HW) { 1193 dev_put(netdev); 1194 tls_ctx->netdev = NULL; 1195 } 1196 out: 1197 up_read(&device_offload_lock); 1198 tls_sw_release_resources_rx(sk); 1199 } 1200 1201 static int tls_device_down(struct net_device *netdev) 1202 { 1203 struct tls_context *ctx, *tmp; 1204 unsigned long flags; 1205 LIST_HEAD(list); 1206 1207 /* Request a write lock to block new offload attempts */ 1208 down_write(&device_offload_lock); 1209 1210 spin_lock_irqsave(&tls_device_lock, flags); 1211 list_for_each_entry_safe(ctx, tmp, &tls_device_list, list) { 1212 if (ctx->netdev != netdev || 1213 !refcount_inc_not_zero(&ctx->refcount)) 1214 continue; 1215 1216 list_move(&ctx->list, &list); 1217 } 1218 spin_unlock_irqrestore(&tls_device_lock, flags); 1219 1220 list_for_each_entry_safe(ctx, tmp, &list, list) { 1221 if (ctx->tx_conf == TLS_HW) 1222 netdev->tlsdev_ops->tls_dev_del(netdev, ctx, 1223 TLS_OFFLOAD_CTX_DIR_TX); 1224 if (ctx->rx_conf == TLS_HW) 1225 netdev->tlsdev_ops->tls_dev_del(netdev, ctx, 1226 TLS_OFFLOAD_CTX_DIR_RX); 1227 WRITE_ONCE(ctx->netdev, NULL); 1228 smp_mb__before_atomic(); /* pairs with test_and_set_bit() */ 1229 while (test_bit(TLS_RX_SYNC_RUNNING, &ctx->flags)) 1230 usleep_range(10, 200); 1231 dev_put(netdev); 1232 list_del_init(&ctx->list); 1233 1234 if (refcount_dec_and_test(&ctx->refcount)) 1235 tls_device_free_ctx(ctx); 1236 } 1237 1238 up_write(&device_offload_lock); 1239 1240 flush_work(&tls_device_gc_work); 1241 1242 return NOTIFY_DONE; 1243 } 1244 1245 static int tls_dev_event(struct notifier_block *this, unsigned long event, 1246 void *ptr) 1247 { 1248 struct net_device *dev = netdev_notifier_info_to_dev(ptr); 1249 1250 if (!dev->tlsdev_ops && 1251 !(dev->features & (NETIF_F_HW_TLS_RX | NETIF_F_HW_TLS_TX))) 1252 return NOTIFY_DONE; 1253 1254 switch (event) { 1255 case NETDEV_REGISTER: 1256 case NETDEV_FEAT_CHANGE: 1257 if ((dev->features & NETIF_F_HW_TLS_RX) && 1258 !dev->tlsdev_ops->tls_dev_resync) 1259 return NOTIFY_BAD; 1260 1261 if (dev->tlsdev_ops && 1262 dev->tlsdev_ops->tls_dev_add && 1263 dev->tlsdev_ops->tls_dev_del) 1264 return NOTIFY_DONE; 1265 else 1266 return NOTIFY_BAD; 1267 case NETDEV_DOWN: 1268 return tls_device_down(dev); 1269 } 1270 return NOTIFY_DONE; 1271 } 1272 1273 static struct notifier_block tls_dev_notifier = { 1274 .notifier_call = tls_dev_event, 1275 }; 1276 1277 void __init tls_device_init(void) 1278 { 1279 register_netdevice_notifier(&tls_dev_notifier); 1280 } 1281 1282 void __exit tls_device_cleanup(void) 1283 { 1284 unregister_netdevice_notifier(&tls_dev_notifier); 1285 flush_work(&tls_device_gc_work); 1286 clean_acked_data_flush(); 1287 } 1288