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