net/tls/tls_device.c

   1 /* Copyright (c) 2018, Mellanox Technologies All rights reserved.
   2  *
   3  * This software is available to you under a choice of one of two
   4  * licenses.  You may choose to be licensed under the terms of the GNU
   5  * General Public License (GPL) Version 2, available from the file
   6  * COPYING in the main directory of this source tree, or the
   7  * OpenIB.org BSD license below:
   8  *
   9  *     Redistribution and use in source and binary forms, with or
  10  *     without modification, are permitted provided that the following
  11  *     conditions are met:
  12  *
  13  *      - Redistributions of source code must retain the above
  14  *        copyright notice, this list of conditions and the following
  15  *        disclaimer.
  16  *
  17  *      - Redistributions in binary form must reproduce the above
  18  *        copyright notice, this list of conditions and the following
  19  *        disclaimer in the documentation and/or other materials
  20  *        provided with the distribution.
  21  *
  22  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  23  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  24  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  25  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  26  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  27  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  28  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  29  * SOFTWARE.
  30  */
  31
  32 #include <crypto/aead.h>
  33 #include <linux/highmem.h>
  34 #include <linux/module.h>
  35 #include <linux/netdevice.h>
  36 #include <net/dst.h>
  37 #include <net/inet_connection_sock.h>
  38 #include <net/tcp.h>
  39 #include <net/tls.h>
  40
  41 /* device_offload_lock is used to synchronize tls_dev_add
  42  * against NETDEV_DOWN notifications.
  43  */
  44 static DECLARE_RWSEM(device_offload_lock);
  45
  46 static void tls_device_gc_task(struct work_struct *work);
  47
  48 static DECLARE_WORK(tls_device_gc_work, tls_device_gc_task);
  49 static LIST_HEAD(tls_device_gc_list);
  50 static LIST_HEAD(tls_device_list);
  51 static DEFINE_SPINLOCK(tls_device_lock);
  52
  53 static void tls_device_free_ctx(struct tls_context *ctx)
  54 {
  55         if (ctx->tx_conf == TLS_HW) {
  56                 kfree(tls_offload_ctx_tx(ctx));
  57                 kfree(ctx->tx.rec_seq);
  58                 kfree(ctx->tx.iv);
  59         }
  60
  61         if (ctx->rx_conf == TLS_HW)
  62                 kfree(tls_offload_ctx_rx(ctx));
  63
  64         tls_ctx_free(ctx);
  65 }
  66
  67 static void tls_device_gc_task(struct work_struct *work)
  68 {
  69         struct tls_context *ctx, *tmp;
  70         unsigned long flags;
  71         LIST_HEAD(gc_list);
  72
  73         spin_lock_irqsave(&tls_device_lock, flags);
  74         list_splice_init(&tls_device_gc_list, &gc_list);
  75         spin_unlock_irqrestore(&tls_device_lock, flags);
  76
  77         list_for_each_entry_safe(ctx, tmp, &gc_list, list) {
  78                 struct net_device *netdev = ctx->netdev;
  79
  80                 if (netdev && ctx->tx_conf == TLS_HW) {
  81                         netdev->tlsdev_ops->tls_dev_del(netdev, ctx,
  82                                                         TLS_OFFLOAD_CTX_DIR_TX);
  83                         dev_put(netdev);
  84                         ctx->netdev = NULL;
  85                 }
  86
  87                 list_del(&ctx->list);
  88                 tls_device_free_ctx(ctx);
  89         }
  90 }
  91
  92 static void tls_device_queue_ctx_destruction(struct tls_context *ctx)
  93 {
  94         unsigned long flags;
  95
  96         spin_lock_irqsave(&tls_device_lock, flags);
  97         list_move_tail(&ctx->list, &tls_device_gc_list);
  98
  99         /* schedule_work inside the spinlock
 100          * to make sure tls_device_down waits for that work.
 101          */
 102         schedule_work(&tls_device_gc_work);
 103
 104         spin_unlock_irqrestore(&tls_device_lock, flags);
 105 }
 106
 107 /* We assume that the socket is already connected */
 108 static struct net_device *get_netdev_for_sock(struct sock *sk)
 109 {
 110         struct dst_entry *dst = sk_dst_get(sk);
 111         struct net_device *netdev = NULL;
 112
 113         if (likely(dst)) {
 114                 netdev = dst->dev;
 115                 dev_hold(netdev);
 116         }
 117
 118         dst_release(dst);
 119
 120         return netdev;
 121 }
 122
 123 static void destroy_record(struct tls_record_info *record)
 124 {
 125         int nr_frags = record->num_frags;
 126         skb_frag_t *frag;
 127
 128         while (nr_frags-- > 0) {
 129                 frag = &record->frags[nr_frags];
 130                 __skb_frag_unref(frag);
 131         }
 132         kfree(record);
 133 }
 134
 135 static void delete_all_records(struct tls_offload_context_tx *offload_ctx)
 136 {
 137         struct tls_record_info *info, *temp;
 138
 139         list_for_each_entry_safe(info, temp, &offload_ctx->records_list, list) {
 140                 list_del(&info->list);
 141                 destroy_record(info);
 142         }
 143
 144         offload_ctx->retransmit_hint = NULL;
 145 }
 146
 147 static void tls_icsk_clean_acked(struct sock *sk, u32 acked_seq)
 148 {
 149         struct tls_context *tls_ctx = tls_get_ctx(sk);
 150         struct tls_record_info *info, *temp;
 151         struct tls_offload_context_tx *ctx;
 152         u64 deleted_records = 0;
 153         unsigned long flags;
 154
 155         if (!tls_ctx)
 156                 return;
 157
 158         ctx = tls_offload_ctx_tx(tls_ctx);
 159
 160         spin_lock_irqsave(&ctx->lock, flags);
 161         info = ctx->retransmit_hint;
 162         if (info && !before(acked_seq, info->end_seq)) {
 163                 ctx->retransmit_hint = NULL;
 164                 list_del(&info->list);
 165                 destroy_record(info);
 166                 deleted_records++;
 167         }
 168
 169         list_for_each_entry_safe(info, temp, &ctx->records_list, list) {
 170                 if (before(acked_seq, info->end_seq))
 171                         break;
 172                 list_del(&info->list);
 173
 174                 destroy_record(info);
 175                 deleted_records++;
 176         }
 177
 178         ctx->unacked_record_sn += deleted_records;
 179         spin_unlock_irqrestore(&ctx->lock, flags);
 180 }
 181
 182 /* At this point, there should be no references on this
 183  * socket and no in-flight SKBs associated with this
 184  * socket, so it is safe to free all the resources.
 185  */
 186 static void tls_device_sk_destruct(struct sock *sk)
 187 {
 188         struct tls_context *tls_ctx = tls_get_ctx(sk);
 189         struct tls_offload_context_tx *ctx = tls_offload_ctx_tx(tls_ctx);
 190
 191         tls_ctx->sk_destruct(sk);
 192
 193         if (tls_ctx->tx_conf == TLS_HW) {
 194                 if (ctx->open_record)
 195                         destroy_record(ctx->open_record);
 196                 delete_all_records(ctx);
 197                 crypto_free_aead(ctx->aead_send);
 198                 clean_acked_data_disable(inet_csk(sk));
 199         }
 200
 201         if (refcount_dec_and_test(&tls_ctx->refcount))
 202                 tls_device_queue_ctx_destruction(tls_ctx);
 203 }
 204
 205 void tls_device_free_resources_tx(struct sock *sk)
 206 {
 207         struct tls_context *tls_ctx = tls_get_ctx(sk);
 208
 209         tls_free_partial_record(sk, tls_ctx);
 210 }
 211
 212 static void tls_device_resync_tx(struct sock *sk, struct tls_context *tls_ctx,
 213                                  u32 seq)
 214 {
 215         struct net_device *netdev;
 216         struct sk_buff *skb;
 217         int err = 0;
 218         u8 *rcd_sn;
 219
 220         skb = tcp_write_queue_tail(sk);
 221         if (skb)
 222                 TCP_SKB_CB(skb)->eor = 1;
 223
 224         rcd_sn = tls_ctx->tx.rec_seq;
 225
 226         down_read(&device_offload_lock);
 227         netdev = tls_ctx->netdev;
 228         if (netdev)
 229                 err = netdev->tlsdev_ops->tls_dev_resync(netdev, sk, seq,
 230                                                          rcd_sn,
 231                                                          TLS_OFFLOAD_CTX_DIR_TX);
 232         up_read(&device_offload_lock);
 233         if (err)
 234                 return;
 235
 236         clear_bit_unlock(TLS_TX_SYNC_SCHED, &tls_ctx->flags);
 237 }
 238
 239 static void tls_append_frag(struct tls_record_info *record,
 240                             struct page_frag *pfrag,
 241                             int size)
 242 {
 243         skb_frag_t *frag;
 244
 245         frag = &record->frags[record->num_frags - 1];
 246         if (frag->page.p == pfrag->page &&
 247             frag->page_offset + frag->size == pfrag->offset) {
 248                 frag->size += size;
 249         } else {
 250                 ++frag;
 251                 frag->page.p = pfrag->page;
 252                 frag->page_offset = pfrag->offset;
 253                 frag->size = size;
 254                 ++record->num_frags;
 255                 get_page(pfrag->page);
 256         }
 257
 258         pfrag->offset += size;
 259         record->len += size;
 260 }
 261
 262 static int tls_push_record(struct sock *sk,
 263                            struct tls_context *ctx,
 264                            struct tls_offload_context_tx *offload_ctx,
 265                            struct tls_record_info *record,
 266                            struct page_frag *pfrag,
 267                            int flags,
 268                            unsigned char record_type)
 269 {
 270         struct tls_prot_info *prot = &ctx->prot_info;
 271         struct tcp_sock *tp = tcp_sk(sk);
 272         struct page_frag dummy_tag_frag;
 273         skb_frag_t *frag;
 274         int i;
 275
 276         /* fill prepend */
 277         frag = &record->frags[0];
 278         tls_fill_prepend(ctx,
 279                          skb_frag_address(frag),
 280                          record->len - prot->prepend_size,
 281                          record_type,
 282                          prot->version);
 283
 284         /* HW doesn't care about the data in the tag, because it fills it. */
 285         dummy_tag_frag.page = skb_frag_page(frag);
 286         dummy_tag_frag.offset = 0;
 287
 288         tls_append_frag(record, &dummy_tag_frag, prot->tag_size);
 289         record->end_seq = tp->write_seq + record->len;
 290         spin_lock_irq(&offload_ctx->lock);
 291         list_add_tail(&record->list, &offload_ctx->records_list);
 292         spin_unlock_irq(&offload_ctx->lock);
 293         offload_ctx->open_record = NULL;
 294
 295         if (test_bit(TLS_TX_SYNC_SCHED, &ctx->flags))
 296                 tls_device_resync_tx(sk, ctx, tp->write_seq);
 297
 298         tls_advance_record_sn(sk, prot, &ctx->tx);
 299
 300         for (i = 0; i < record->num_frags; i++) {
 301                 frag = &record->frags[i];
 302                 sg_unmark_end(&offload_ctx->sg_tx_data[i]);
 303                 sg_set_page(&offload_ctx->sg_tx_data[i], skb_frag_page(frag),
 304                             frag->size, frag->page_offset);
 305                 sk_mem_charge(sk, frag->size);
 306                 get_page(skb_frag_page(frag));
 307         }
 308         sg_mark_end(&offload_ctx->sg_tx_data[record->num_frags - 1]);
 309
 310         /* all ready, send */
 311         return tls_push_sg(sk, ctx, offload_ctx->sg_tx_data, 0, flags);
 312 }
 313
 314 static int tls_create_new_record(struct tls_offload_context_tx *offload_ctx,
 315                                  struct page_frag *pfrag,
 316                                  size_t prepend_size)
 317 {
 318         struct tls_record_info *record;
 319         skb_frag_t *frag;
 320
 321         record = kmalloc(sizeof(*record), GFP_KERNEL);
 322         if (!record)
 323                 return -ENOMEM;
 324
 325         frag = &record->frags[0];
 326         __skb_frag_set_page(frag, pfrag->page);
 327         frag->page_offset = pfrag->offset;
 328         skb_frag_size_set(frag, prepend_size);
 329
 330         get_page(pfrag->page);
 331         pfrag->offset += prepend_size;
 332
 333         record->num_frags = 1;
 334         record->len = prepend_size;
 335         offload_ctx->open_record = record;
 336         return 0;
 337 }
 338
 339 static int tls_do_allocation(struct sock *sk,
 340                              struct tls_offload_context_tx *offload_ctx,
 341                              struct page_frag *pfrag,
 342                              size_t prepend_size)
 343 {
 344         int ret;
 345
 346         if (!offload_ctx->open_record) {
 347                 if (unlikely(!skb_page_frag_refill(prepend_size, pfrag,
 348                                                    sk->sk_allocation))) {
 349                         sk->sk_prot->enter_memory_pressure(sk);
 350                         sk_stream_moderate_sndbuf(sk);
 351                         return -ENOMEM;
 352                 }
 353
 354                 ret = tls_create_new_record(offload_ctx, pfrag, prepend_size);
 355                 if (ret)
 356                         return ret;
 357
 358                 if (pfrag->size > pfrag->offset)
 359                         return 0;
 360         }
 361
 362         if (!sk_page_frag_refill(sk, pfrag))
 363                 return -ENOMEM;
 364
 365         return 0;
 366 }
 367
 368 static int tls_push_data(struct sock *sk,
 369                          struct iov_iter *msg_iter,
 370                          size_t size, int flags,
 371                          unsigned char record_type)
 372 {
 373         struct tls_context *tls_ctx = tls_get_ctx(sk);
 374         struct tls_prot_info *prot = &tls_ctx->prot_info;
 375         struct tls_offload_context_tx *ctx = tls_offload_ctx_tx(tls_ctx);
 376         int more = flags & (MSG_SENDPAGE_NOTLAST | MSG_MORE);
 377         struct tls_record_info *record = ctx->open_record;
 378         int tls_push_record_flags;
 379         struct page_frag *pfrag;
 380         size_t orig_size = size;
 381         u32 max_open_record_len;
 382         int copy, rc = 0;
 383         bool done = false;
 384         long timeo;
 385
 386         if (flags &
 387             ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL | MSG_SENDPAGE_NOTLAST))
 388                 return -ENOTSUPP;
 389
 390         if (sk->sk_err)
 391                 return -sk->sk_err;
 392
 393         flags |= MSG_SENDPAGE_DECRYPTED;
 394         tls_push_record_flags = flags | MSG_SENDPAGE_NOTLAST;
 395
 396         timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
 397         if (tls_is_partially_sent_record(tls_ctx)) {
 398                 rc = tls_push_partial_record(sk, tls_ctx, flags);
 399                 if (rc < 0)
 400                         return rc;
 401         }
 402
 403         pfrag = sk_page_frag(sk);
 404
 405         /* TLS_HEADER_SIZE is not counted as part of the TLS record, and
 406          * we need to leave room for an authentication tag.
 407          */
 408         max_open_record_len = TLS_MAX_PAYLOAD_SIZE +
 409                               prot->prepend_size;
 410         do {
 411                 rc = tls_do_allocation(sk, ctx, pfrag,
 412                                        prot->prepend_size);
 413                 if (rc) {
 414                         rc = sk_stream_wait_memory(sk, &timeo);
 415                         if (!rc)
 416                                 continue;
 417
 418                         record = ctx->open_record;
 419                         if (!record)
 420                                 break;
 421 handle_error:
 422                         if (record_type != TLS_RECORD_TYPE_DATA) {
 423                                 /* avoid sending partial
 424                                  * record with type !=
 425                                  * application_data
 426                                  */
 427                                 size = orig_size;
 428                                 destroy_record(record);
 429                                 ctx->open_record = NULL;
 430                         } else if (record->len > prot->prepend_size) {
 431                                 goto last_record;
 432                         }
 433
 434                         break;
 435                 }
 436
 437                 record = ctx->open_record;
 438                 copy = min_t(size_t, size, (pfrag->size - pfrag->offset));
 439                 copy = min_t(size_t, copy, (max_open_record_len - record->len));
 440
 441                 if (copy_from_iter_nocache(page_address(pfrag->page) +
 442                                                pfrag->offset,
 443                                            copy, msg_iter) != copy) {
 444                         rc = -EFAULT;
 445                         goto handle_error;
 446                 }
 447                 tls_append_frag(record, pfrag, copy);
 448
 449                 size -= copy;
 450                 if (!size) {
 451 last_record:
 452                         tls_push_record_flags = flags;
 453                         if (more) {
 454                                 tls_ctx->pending_open_record_frags =
 455                                                 !!record->num_frags;
 456                                 break;
 457                         }
 458
 459                         done = true;
 460                 }
 461
 462                 if (done || record->len >= max_open_record_len ||
 463                     (record->num_frags >= MAX_SKB_FRAGS - 1)) {
 464                         rc = tls_push_record(sk,
 465                                              tls_ctx,
 466                                              ctx,
 467                                              record,
 468                                              pfrag,
 469                                              tls_push_record_flags,
 470                                              record_type);
 471                         if (rc < 0)
 472                                 break;
 473                 }
 474         } while (!done);
 475
 476         if (orig_size - size > 0)
 477                 rc = orig_size - size;
 478
 479         return rc;
 480 }
 481
 482 int tls_device_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
 483 {
 484         unsigned char record_type = TLS_RECORD_TYPE_DATA;
 485         int rc;
 486
 487         lock_sock(sk);
 488
 489         if (unlikely(msg->msg_controllen)) {
 490                 rc = tls_proccess_cmsg(sk, msg, &record_type);
 491                 if (rc)
 492                         goto out;
 493         }
 494
 495         rc = tls_push_data(sk, &msg->msg_iter, size,
 496                            msg->msg_flags, record_type);
 497
 498 out:
 499         release_sock(sk);
 500         return rc;
 501 }
 502
 503 int tls_device_sendpage(struct sock *sk, struct page *page,
 504                         int offset, size_t size, int flags)
 505 {
 506         struct iov_iter msg_iter;
 507         char *kaddr = kmap(page);
 508         struct kvec iov;
 509         int rc;
 510
 511         if (flags & MSG_SENDPAGE_NOTLAST)
 512                 flags |= MSG_MORE;
 513
 514         lock_sock(sk);
 515
 516         if (flags & MSG_OOB) {
 517                 rc = -ENOTSUPP;
 518                 goto out;
 519         }
 520
 521         iov.iov_base = kaddr + offset;
 522         iov.iov_len = size;
 523         iov_iter_kvec(&msg_iter, WRITE, &iov, 1, size);
 524         rc = tls_push_data(sk, &msg_iter, size,
 525                            flags, TLS_RECORD_TYPE_DATA);
 526         kunmap(page);
 527
 528 out:
 529         release_sock(sk);
 530         return rc;
 531 }
 532
 533 struct tls_record_info *tls_get_record(struct tls_offload_context_tx *context,
 534                                        u32 seq, u64 *p_record_sn)
 535 {
 536         u64 record_sn = context->hint_record_sn;
 537         struct tls_record_info *info;
 538
 539         info = context->retransmit_hint;
 540         if (!info ||
 541             before(seq, info->end_seq - info->len)) {
 542                 /* if retransmit_hint is irrelevant start
 543                  * from the beggining of the list
 544                  */
 545                 info = list_first_entry(&context->records_list,
 546                                         struct tls_record_info, list);
 547                 record_sn = context->unacked_record_sn;
 548         }
 549
 550         list_for_each_entry_from(info, &context->records_list, list) {
 551                 if (before(seq, info->end_seq)) {
 552                         if (!context->retransmit_hint ||
 553                             after(info->end_seq,
 554                                   context->retransmit_hint->end_seq)) {
 555                                 context->hint_record_sn = record_sn;
 556                                 context->retransmit_hint = info;
 557                         }
 558                         *p_record_sn = record_sn;
 559                         return info;
 560                 }
 561                 record_sn++;
 562         }
 563
 564         return NULL;
 565 }
 566 EXPORT_SYMBOL(tls_get_record);
 567
 568 static int tls_device_push_pending_record(struct sock *sk, int flags)
 569 {
 570         struct iov_iter msg_iter;
 571
 572         iov_iter_kvec(&msg_iter, WRITE, NULL, 0, 0);
 573         return tls_push_data(sk, &msg_iter, 0, flags, TLS_RECORD_TYPE_DATA);
 574 }
 575
 576 void tls_device_write_space(struct sock *sk, struct tls_context *ctx)
 577 {
 578         if (!sk->sk_write_pending && tls_is_partially_sent_record(ctx)) {
 579                 gfp_t sk_allocation = sk->sk_allocation;
 580
 581                 sk->sk_allocation = GFP_ATOMIC;
 582                 tls_push_partial_record(sk, ctx,
 583                                         MSG_DONTWAIT | MSG_NOSIGNAL |
 584                                         MSG_SENDPAGE_DECRYPTED);
 585                 sk->sk_allocation = sk_allocation;
 586         }
 587 }
 588
 589 static void tls_device_resync_rx(struct tls_context *tls_ctx,
 590                                  struct sock *sk, u32 seq, u8 *rcd_sn)
 591 {
 592         struct net_device *netdev;
 593
 594         if (WARN_ON(test_and_set_bit(TLS_RX_SYNC_RUNNING, &tls_ctx->flags)))
 595                 return;
 596         netdev = READ_ONCE(tls_ctx->netdev);
 597         if (netdev)
 598                 netdev->tlsdev_ops->tls_dev_resync(netdev, sk, seq, rcd_sn,
 599                                                    TLS_OFFLOAD_CTX_DIR_RX);
 600         clear_bit_unlock(TLS_RX_SYNC_RUNNING, &tls_ctx->flags);
 601 }
 602
 603 void tls_device_rx_resync_new_rec(struct sock *sk, u32 rcd_len, u32 seq)
 604 {
 605         struct tls_context *tls_ctx = tls_get_ctx(sk);
 606         struct tls_offload_context_rx *rx_ctx;
 607         u8 rcd_sn[TLS_MAX_REC_SEQ_SIZE];
 608         struct tls_prot_info *prot;
 609         u32 is_req_pending;
 610         s64 resync_req;
 611         u32 req_seq;
 612
 613         if (tls_ctx->rx_conf != TLS_HW)
 614                 return;
 615
 616         prot = &tls_ctx->prot_info;
 617         rx_ctx = tls_offload_ctx_rx(tls_ctx);
 618         memcpy(rcd_sn, tls_ctx->rx.rec_seq, prot->rec_seq_size);
 619
 620         switch (rx_ctx->resync_type) {
 621         case TLS_OFFLOAD_SYNC_TYPE_DRIVER_REQ:
 622                 resync_req = atomic64_read(&rx_ctx->resync_req);
 623                 req_seq = resync_req >> 32;
 624                 seq += TLS_HEADER_SIZE - 1;
 625                 is_req_pending = resync_req;
 626
 627                 if (likely(!is_req_pending) || req_seq != seq ||
 628                     !atomic64_try_cmpxchg(&rx_ctx->resync_req, &resync_req, 0))
 629                         return;
 630                 break;
 631         case TLS_OFFLOAD_SYNC_TYPE_CORE_NEXT_HINT:
 632                 if (likely(!rx_ctx->resync_nh_do_now))
 633                         return;
 634
 635                 /* head of next rec is already in, note that the sock_inq will
 636                  * include the currently parsed message when called from parser
 637                  */
 638                 if (tcp_inq(sk) > rcd_len)
 639                         return;
 640
 641                 rx_ctx->resync_nh_do_now = 0;
 642                 seq += rcd_len;
 643                 tls_bigint_increment(rcd_sn, prot->rec_seq_size);
 644                 break;
 645         }
 646
 647         tls_device_resync_rx(tls_ctx, sk, seq, rcd_sn);
 648 }
 649
 650 static void tls_device_core_ctrl_rx_resync(struct tls_context *tls_ctx,
 651                                            struct tls_offload_context_rx *ctx,
 652                                            struct sock *sk, struct sk_buff *skb)
 653 {
 654         struct strp_msg *rxm;
 655
 656         /* device will request resyncs by itself based on stream scan */
 657         if (ctx->resync_type != TLS_OFFLOAD_SYNC_TYPE_CORE_NEXT_HINT)
 658                 return;
 659         /* already scheduled */
 660         if (ctx->resync_nh_do_now)
 661                 return;
 662         /* seen decrypted fragments since last fully-failed record */
 663         if (ctx->resync_nh_reset) {
 664                 ctx->resync_nh_reset = 0;
 665                 ctx->resync_nh.decrypted_failed = 1;
 666                 ctx->resync_nh.decrypted_tgt = TLS_DEVICE_RESYNC_NH_START_IVAL;
 667                 return;
 668         }
 669
 670         if (++ctx->resync_nh.decrypted_failed <= ctx->resync_nh.decrypted_tgt)
 671                 return;
 672
 673         /* doing resync, bump the next target in case it fails */
 674         if (ctx->resync_nh.decrypted_tgt < TLS_DEVICE_RESYNC_NH_MAX_IVAL)
 675                 ctx->resync_nh.decrypted_tgt *= 2;
 676         else
 677                 ctx->resync_nh.decrypted_tgt += TLS_DEVICE_RESYNC_NH_MAX_IVAL;
 678
 679         rxm = strp_msg(skb);
 680
 681         /* head of next rec is already in, parser will sync for us */
 682         if (tcp_inq(sk) > rxm->full_len) {
 683                 ctx->resync_nh_do_now = 1;
 684         } else {
 685                 struct tls_prot_info *prot = &tls_ctx->prot_info;
 686                 u8 rcd_sn[TLS_MAX_REC_SEQ_SIZE];
 687
 688                 memcpy(rcd_sn, tls_ctx->rx.rec_seq, prot->rec_seq_size);
 689                 tls_bigint_increment(rcd_sn, prot->rec_seq_size);
 690
 691                 tls_device_resync_rx(tls_ctx, sk, tcp_sk(sk)->copied_seq,
 692                                      rcd_sn);
 693         }
 694 }
 695
 696 static int tls_device_reencrypt(struct sock *sk, struct sk_buff *skb)
 697 {
 698         struct strp_msg *rxm = strp_msg(skb);
 699         int err = 0, offset = rxm->offset, copy, nsg, data_len, pos;
 700         struct sk_buff *skb_iter, *unused;
 701         struct scatterlist sg[1];
 702         char *orig_buf, *buf;
 703
 704         orig_buf = kmalloc(rxm->full_len + TLS_HEADER_SIZE +
 705                            TLS_CIPHER_AES_GCM_128_IV_SIZE, sk->sk_allocation);
 706         if (!orig_buf)
 707                 return -ENOMEM;
 708         buf = orig_buf;
 709
 710         nsg = skb_cow_data(skb, 0, &unused);
 711         if (unlikely(nsg < 0)) {
 712                 err = nsg;
 713                 goto free_buf;
 714         }
 715
 716         sg_init_table(sg, 1);
 717         sg_set_buf(&sg[0], buf,
 718                    rxm->full_len + TLS_HEADER_SIZE +
 719                    TLS_CIPHER_AES_GCM_128_IV_SIZE);
 720         err = skb_copy_bits(skb, offset, buf,
 721                             TLS_HEADER_SIZE + TLS_CIPHER_AES_GCM_128_IV_SIZE);
 722         if (err)
 723                 goto free_buf;
 724
 725         /* We are interested only in the decrypted data not the auth */
 726         err = decrypt_skb(sk, skb, sg);
 727         if (err != -EBADMSG)
 728                 goto free_buf;
 729         else
 730                 err = 0;
 731
 732         data_len = rxm->full_len - TLS_CIPHER_AES_GCM_128_TAG_SIZE;
 733
 734         if (skb_pagelen(skb) > offset) {
 735                 copy = min_t(int, skb_pagelen(skb) - offset, data_len);
 736
 737                 if (skb->decrypted) {
 738                         err = skb_store_bits(skb, offset, buf, copy);
 739                         if (err)
 740                                 goto free_buf;
 741                 }
 742
 743                 offset += copy;
 744                 buf += copy;
 745         }
 746
 747         pos = skb_pagelen(skb);
 748         skb_walk_frags(skb, skb_iter) {
 749                 int frag_pos;
 750
 751                 /* Practically all frags must belong to msg if reencrypt
 752                  * is needed with current strparser and coalescing logic,
 753                  * but strparser may "get optimized", so let's be safe.
 754                  */
 755                 if (pos + skb_iter->len <= offset)
 756                         goto done_with_frag;
 757                 if (pos >= data_len + rxm->offset)
 758                         break;
 759
 760                 frag_pos = offset - pos;
 761                 copy = min_t(int, skb_iter->len - frag_pos,
 762                              data_len + rxm->offset - offset);
 763
 764                 if (skb_iter->decrypted) {
 765                         err = skb_store_bits(skb_iter, frag_pos, buf, copy);
 766                         if (err)
 767                                 goto free_buf;
 768                 }
 769
 770                 offset += copy;
 771                 buf += copy;
 772 done_with_frag:
 773                 pos += skb_iter->len;
 774         }
 775
 776 free_buf:
 777         kfree(orig_buf);
 778         return err;
 779 }
 780
 781 int tls_device_decrypted(struct sock *sk, struct sk_buff *skb)
 782 {
 783         struct tls_context *tls_ctx = tls_get_ctx(sk);
 784         struct tls_offload_context_rx *ctx = tls_offload_ctx_rx(tls_ctx);
 785         int is_decrypted = skb->decrypted;
 786         int is_encrypted = !is_decrypted;
 787         struct sk_buff *skb_iter;
 788
 789         /* Check if all the data is decrypted already */
 790         skb_walk_frags(skb, skb_iter) {
 791                 is_decrypted &= skb_iter->decrypted;
 792                 is_encrypted &= !skb_iter->decrypted;
 793         }
 794
 795         ctx->sw.decrypted |= is_decrypted;
 796
 797         /* Return immediately if the record is either entirely plaintext or
 798          * entirely ciphertext. Otherwise handle reencrypt partially decrypted
 799          * record.
 800          */
 801         if (is_decrypted) {
 802                 ctx->resync_nh_reset = 1;
 803                 return 0;
 804         }
 805         if (is_encrypted) {
 806                 tls_device_core_ctrl_rx_resync(tls_ctx, ctx, sk, skb);
 807                 return 0;
 808         }
 809
 810         ctx->resync_nh_reset = 1;
 811         return tls_device_reencrypt(sk, skb);
 812 }
 813
 814 static void tls_device_attach(struct tls_context *ctx, struct sock *sk,
 815                               struct net_device *netdev)
 816 {
 817         if (sk->sk_destruct != tls_device_sk_destruct) {
 818                 refcount_set(&ctx->refcount, 1);
 819                 dev_hold(netdev);
 820                 ctx->netdev = netdev;
 821                 spin_lock_irq(&tls_device_lock);
 822                 list_add_tail(&ctx->list, &tls_device_list);
 823                 spin_unlock_irq(&tls_device_lock);
 824
 825                 ctx->sk_destruct = sk->sk_destruct;
 826                 sk->sk_destruct = tls_device_sk_destruct;
 827         }
 828 }
 829
 830 int tls_set_device_offload(struct sock *sk, struct tls_context *ctx)
 831 {
 832         u16 nonce_size, tag_size, iv_size, rec_seq_size;
 833         struct tls_context *tls_ctx = tls_get_ctx(sk);
 834         struct tls_prot_info *prot = &tls_ctx->prot_info;
 835         struct tls_record_info *start_marker_record;
 836         struct tls_offload_context_tx *offload_ctx;
 837         struct tls_crypto_info *crypto_info;
 838         struct net_device *netdev;
 839         char *iv, *rec_seq;
 840         struct sk_buff *skb;
 841         int rc = -EINVAL;
 842         __be64 rcd_sn;
 843
 844         if (!ctx)
 845                 goto out;
 846
 847         if (ctx->priv_ctx_tx) {
 848                 rc = -EEXIST;
 849                 goto out;
 850         }
 851
 852         start_marker_record = kmalloc(sizeof(*start_marker_record), GFP_KERNEL);
 853         if (!start_marker_record) {
 854                 rc = -ENOMEM;
 855                 goto out;
 856         }
 857
 858         offload_ctx = kzalloc(TLS_OFFLOAD_CONTEXT_SIZE_TX, GFP_KERNEL);
 859         if (!offload_ctx) {
 860                 rc = -ENOMEM;
 861                 goto free_marker_record;
 862         }
 863
 864         crypto_info = &ctx->crypto_send.info;
 865         if (crypto_info->version != TLS_1_2_VERSION) {
 866                 rc = -EOPNOTSUPP;
 867                 goto free_offload_ctx;
 868         }
 869
 870         switch (crypto_info->cipher_type) {
 871         case TLS_CIPHER_AES_GCM_128:
 872                 nonce_size = TLS_CIPHER_AES_GCM_128_IV_SIZE;
 873                 tag_size = TLS_CIPHER_AES_GCM_128_TAG_SIZE;
 874                 iv_size = TLS_CIPHER_AES_GCM_128_IV_SIZE;
 875                 iv = ((struct tls12_crypto_info_aes_gcm_128 *)crypto_info)->iv;
 876                 rec_seq_size = TLS_CIPHER_AES_GCM_128_REC_SEQ_SIZE;
 877                 rec_seq =
 878                  ((struct tls12_crypto_info_aes_gcm_128 *)crypto_info)->rec_seq;
 879                 break;
 880         default:
 881                 rc = -EINVAL;
 882                 goto free_offload_ctx;
 883         }
 884
 885         /* Sanity-check the rec_seq_size for stack allocations */
 886         if (rec_seq_size > TLS_MAX_REC_SEQ_SIZE) {
 887                 rc = -EINVAL;
 888                 goto free_offload_ctx;
 889         }
 890
 891         prot->version = crypto_info->version;
 892         prot->cipher_type = crypto_info->cipher_type;
 893         prot->prepend_size = TLS_HEADER_SIZE + nonce_size;
 894         prot->tag_size = tag_size;
 895         prot->overhead_size = prot->prepend_size + prot->tag_size;
 896         prot->iv_size = iv_size;
 897         ctx->tx.iv = kmalloc(iv_size + TLS_CIPHER_AES_GCM_128_SALT_SIZE,
 898                              GFP_KERNEL);
 899         if (!ctx->tx.iv) {
 900                 rc = -ENOMEM;
 901                 goto free_offload_ctx;
 902         }
 903
 904         memcpy(ctx->tx.iv + TLS_CIPHER_AES_GCM_128_SALT_SIZE, iv, iv_size);
 905
 906         prot->rec_seq_size = rec_seq_size;
 907         ctx->tx.rec_seq = kmemdup(rec_seq, rec_seq_size, GFP_KERNEL);
 908         if (!ctx->tx.rec_seq) {
 909                 rc = -ENOMEM;
 910                 goto free_iv;
 911         }
 912
 913         rc = tls_sw_fallback_init(sk, offload_ctx, crypto_info);
 914         if (rc)
 915                 goto free_rec_seq;
 916
 917         /* start at rec_seq - 1 to account for the start marker record */
 918         memcpy(&rcd_sn, ctx->tx.rec_seq, sizeof(rcd_sn));
 919         offload_ctx->unacked_record_sn = be64_to_cpu(rcd_sn) - 1;
 920
 921         start_marker_record->end_seq = tcp_sk(sk)->write_seq;
 922         start_marker_record->len = 0;
 923         start_marker_record->num_frags = 0;
 924
 925         INIT_LIST_HEAD(&offload_ctx->records_list);
 926         list_add_tail(&start_marker_record->list, &offload_ctx->records_list);
 927         spin_lock_init(&offload_ctx->lock);
 928         sg_init_table(offload_ctx->sg_tx_data,
 929                       ARRAY_SIZE(offload_ctx->sg_tx_data));
 930
 931         clean_acked_data_enable(inet_csk(sk), &tls_icsk_clean_acked);
 932         ctx->push_pending_record = tls_device_push_pending_record;
 933
 934         /* TLS offload is greatly simplified if we don't send
 935          * SKBs where only part of the payload needs to be encrypted.
 936          * So mark the last skb in the write queue as end of record.
 937          */
 938         skb = tcp_write_queue_tail(sk);
 939         if (skb)
 940                 TCP_SKB_CB(skb)->eor = 1;
 941
 942         /* We support starting offload on multiple sockets
 943          * concurrently, so we only need a read lock here.
 944          * This lock must precede get_netdev_for_sock to prevent races between
 945          * NETDEV_DOWN and setsockopt.
 946          */
 947         down_read(&device_offload_lock);
 948         netdev = get_netdev_for_sock(sk);
 949         if (!netdev) {
 950                 pr_err_ratelimited("%s: netdev not found\n", __func__);
 951                 rc = -EINVAL;
 952                 goto release_lock;
 953         }
 954
 955         if (!(netdev->features & NETIF_F_HW_TLS_TX)) {
 956                 rc = -ENOTSUPP;
 957                 goto release_netdev;
 958         }
 959
 960         /* Avoid offloading if the device is down
 961          * We don't want to offload new flows after
 962          * the NETDEV_DOWN event
 963          */
 964         if (!(netdev->flags & IFF_UP)) {
 965                 rc = -EINVAL;
 966                 goto release_netdev;
 967         }
 968
 969         ctx->priv_ctx_tx = offload_ctx;
 970         rc = netdev->tlsdev_ops->tls_dev_add(netdev, sk, TLS_OFFLOAD_CTX_DIR_TX,
 971                                              &ctx->crypto_send.info,
 972                                              tcp_sk(sk)->write_seq);
 973         if (rc)
 974                 goto release_netdev;
 975
 976         tls_device_attach(ctx, sk, netdev);
 977
 978         /* following this assignment tls_is_sk_tx_device_offloaded
 979          * will return true and the context might be accessed
 980          * by the netdev's xmit function.
 981          */
 982         smp_store_release(&sk->sk_validate_xmit_skb, tls_validate_xmit_skb);
 983         dev_put(netdev);
 984         up_read(&device_offload_lock);
 985         goto out;
 986
 987 release_netdev:
 988         dev_put(netdev);
 989 release_lock:
 990         up_read(&device_offload_lock);
 991         clean_acked_data_disable(inet_csk(sk));
 992         crypto_free_aead(offload_ctx->aead_send);
 993 free_rec_seq:
 994         kfree(ctx->tx.rec_seq);
 995 free_iv:
 996         kfree(ctx->tx.iv);
 997 free_offload_ctx:
 998         kfree(offload_ctx);
 999         ctx->priv_ctx_tx = NULL;
1000 free_marker_record:
1001         kfree(start_marker_record);
1002 out:
1003         return rc;
1004 }
1005
1006 int tls_set_device_offload_rx(struct sock *sk, struct tls_context *ctx)
1007 {
1008         struct tls_offload_context_rx *context;
1009         struct net_device *netdev;
1010         int rc = 0;
1011
1012         if (ctx->crypto_recv.info.version != TLS_1_2_VERSION)
1013                 return -EOPNOTSUPP;
1014
1015         /* We support starting offload on multiple sockets
1016          * concurrently, so we only need a read lock here.
1017          * This lock must precede get_netdev_for_sock to prevent races between
1018          * NETDEV_DOWN and setsockopt.
1019          */
1020         down_read(&device_offload_lock);
1021         netdev = get_netdev_for_sock(sk);
1022         if (!netdev) {
1023                 pr_err_ratelimited("%s: netdev not found\n", __func__);
1024                 rc = -EINVAL;
1025                 goto release_lock;
1026         }
1027
1028         if (!(netdev->features & NETIF_F_HW_TLS_RX)) {
1029                 rc = -ENOTSUPP;
1030                 goto release_netdev;
1031         }
1032
1033         /* Avoid offloading if the device is down
1034          * We don't want to offload new flows after
1035          * the NETDEV_DOWN event
1036          */
1037         if (!(netdev->flags & IFF_UP)) {
1038                 rc = -EINVAL;
1039                 goto release_netdev;
1040         }
1041
1042         context = kzalloc(TLS_OFFLOAD_CONTEXT_SIZE_RX, GFP_KERNEL);
1043         if (!context) {
1044                 rc = -ENOMEM;
1045                 goto release_netdev;
1046         }
1047         context->resync_nh_reset = 1;
1048
1049         ctx->priv_ctx_rx = context;
1050         rc = tls_set_sw_offload(sk, ctx, 0);
1051         if (rc)
1052                 goto release_ctx;
1053
1054         rc = netdev->tlsdev_ops->tls_dev_add(netdev, sk, TLS_OFFLOAD_CTX_DIR_RX,
1055                                              &ctx->crypto_recv.info,
1056                                              tcp_sk(sk)->copied_seq);
1057         if (rc)
1058                 goto free_sw_resources;
1059
1060         tls_device_attach(ctx, sk, netdev);
1061         goto release_netdev;
1062
1063 free_sw_resources:
1064         up_read(&device_offload_lock);
1065         tls_sw_free_resources_rx(sk);
1066         down_read(&device_offload_lock);
1067 release_ctx:
1068         ctx->priv_ctx_rx = NULL;
1069 release_netdev:
1070         dev_put(netdev);
1071 release_lock:
1072         up_read(&device_offload_lock);
1073         return rc;
1074 }
1075
1076 void tls_device_offload_cleanup_rx(struct sock *sk)
1077 {
1078         struct tls_context *tls_ctx = tls_get_ctx(sk);
1079         struct net_device *netdev;
1080
1081         down_read(&device_offload_lock);
1082         netdev = tls_ctx->netdev;
1083         if (!netdev)
1084                 goto out;
1085
1086         netdev->tlsdev_ops->tls_dev_del(netdev, tls_ctx,
1087                                         TLS_OFFLOAD_CTX_DIR_RX);
1088
1089         if (tls_ctx->tx_conf != TLS_HW) {
1090                 dev_put(netdev);
1091                 tls_ctx->netdev = NULL;
1092         }
1093 out:
1094         up_read(&device_offload_lock);
1095         tls_sw_release_resources_rx(sk);
1096 }
1097
1098 static int tls_device_down(struct net_device *netdev)
1099 {
1100         struct tls_context *ctx, *tmp;
1101         unsigned long flags;
1102         LIST_HEAD(list);
1103
1104         /* Request a write lock to block new offload attempts */
1105         down_write(&device_offload_lock);
1106
1107         spin_lock_irqsave(&tls_device_lock, flags);
1108         list_for_each_entry_safe(ctx, tmp, &tls_device_list, list) {
1109                 if (ctx->netdev != netdev ||
1110                     !refcount_inc_not_zero(&ctx->refcount))
1111                         continue;
1112
1113                 list_move(&ctx->list, &list);
1114         }
1115         spin_unlock_irqrestore(&tls_device_lock, flags);
1116
1117         list_for_each_entry_safe(ctx, tmp, &list, list) {
1118                 if (ctx->tx_conf == TLS_HW)
1119                         netdev->tlsdev_ops->tls_dev_del(netdev, ctx,
1120                                                         TLS_OFFLOAD_CTX_DIR_TX);
1121                 if (ctx->rx_conf == TLS_HW)
1122                         netdev->tlsdev_ops->tls_dev_del(netdev, ctx,
1123                                                         TLS_OFFLOAD_CTX_DIR_RX);
1124                 WRITE_ONCE(ctx->netdev, NULL);
1125                 smp_mb__before_atomic(); /* pairs with test_and_set_bit() */
1126                 while (test_bit(TLS_RX_SYNC_RUNNING, &ctx->flags))
1127                         usleep_range(10, 200);
1128                 dev_put(netdev);
1129                 list_del_init(&ctx->list);
1130
1131                 if (refcount_dec_and_test(&ctx->refcount))
1132                         tls_device_free_ctx(ctx);
1133         }
1134
1135         up_write(&device_offload_lock);
1136
1137         flush_work(&tls_device_gc_work);
1138
1139         return NOTIFY_DONE;
1140 }
1141
1142 static int tls_dev_event(struct notifier_block *this, unsigned long event,
1143                          void *ptr)
1144 {
1145         struct net_device *dev = netdev_notifier_info_to_dev(ptr);
1146
1147         if (!dev->tlsdev_ops &&
1148             !(dev->features & (NETIF_F_HW_TLS_RX | NETIF_F_HW_TLS_TX)))
1149                 return NOTIFY_DONE;
1150
1151         switch (event) {
1152         case NETDEV_REGISTER:
1153         case NETDEV_FEAT_CHANGE:
1154                 if ((dev->features & NETIF_F_HW_TLS_RX) &&
1155                     !dev->tlsdev_ops->tls_dev_resync)
1156                         return NOTIFY_BAD;
1157
1158                 if  (dev->tlsdev_ops &&
1159                      dev->tlsdev_ops->tls_dev_add &&
1160                      dev->tlsdev_ops->tls_dev_del)
1161                         return NOTIFY_DONE;
1162                 else
1163                         return NOTIFY_BAD;
1164         case NETDEV_DOWN:
1165                 return tls_device_down(dev);
1166         }
1167         return NOTIFY_DONE;
1168 }
1169
1170 static struct notifier_block tls_dev_notifier = {
1171         .notifier_call  = tls_dev_event,
1172 };
1173
1174 void __init tls_device_init(void)
1175 {
1176         register_netdevice_notifier(&tls_dev_notifier);
1177 }
1178
1179 void __exit tls_device_cleanup(void)
1180 {
1181         unregister_netdevice_notifier(&tls_dev_notifier);
1182         flush_work(&tls_device_gc_work);
1183         clean_acked_data_flush();
1184 }