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