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         int more = flags & (MSG_SENDPAGE_NOTLAST | MSG_MORE);
 409         struct tls_record_info *record = ctx->open_record;
 410         int tls_push_record_flags;
 411         struct page_frag *pfrag;
 412         size_t orig_size = size;
 413         u32 max_open_record_len;
 414         int copy, rc = 0;
 415         bool done = false;
 416         long timeo;
 417
 418         if (flags &
 419             ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL | MSG_SENDPAGE_NOTLAST))
 420                 return -ENOTSUPP;
 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                 rc = tls_device_copy_data(page_address(pfrag->page) +
 474                                           pfrag->offset, copy, msg_iter);
 475                 if (rc)
 476                         goto handle_error;
 477                 tls_append_frag(record, pfrag, copy);
 478
 479                 size -= copy;
 480                 if (!size) {
 481 last_record:
 482                         tls_push_record_flags = flags;
 483                         if (more) {
 484                                 tls_ctx->pending_open_record_frags =
 485                                                 !!record->num_frags;
 486                                 break;
 487                         }
 488
 489                         done = true;
 490                 }
 491
 492                 if (done || record->len >= max_open_record_len ||
 493                     (record->num_frags >= MAX_SKB_FRAGS - 1)) {
 494                         rc = tls_device_record_close(sk, tls_ctx, record,
 495                                                      pfrag, record_type);
 496                         if (rc) {
 497                                 if (rc > 0) {
 498                                         size += rc;
 499                                 } else {
 500                                         size = orig_size;
 501                                         destroy_record(record);
 502                                         ctx->open_record = NULL;
 503                                         break;
 504                                 }
 505                         }
 506
 507                         rc = tls_push_record(sk,
 508                                              tls_ctx,
 509                                              ctx,
 510                                              record,
 511                                              tls_push_record_flags);
 512                         if (rc < 0)
 513                                 break;
 514                 }
 515         } while (!done);
 516
 517         if (orig_size - size > 0)
 518                 rc = orig_size - size;
 519
 520         return rc;
 521 }
 522
 523 int tls_device_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
 524 {
 525         unsigned char record_type = TLS_RECORD_TYPE_DATA;
 526         struct tls_context *tls_ctx = tls_get_ctx(sk);
 527         int rc;
 528
 529         mutex_lock(&tls_ctx->tx_lock);
 530         lock_sock(sk);
 531
 532         if (unlikely(msg->msg_controllen)) {
 533                 rc = tls_proccess_cmsg(sk, msg, &record_type);
 534                 if (rc)
 535                         goto out;
 536         }
 537
 538         rc = tls_push_data(sk, &msg->msg_iter, size,
 539                            msg->msg_flags, record_type);
 540
 541 out:
 542         release_sock(sk);
 543         mutex_unlock(&tls_ctx->tx_lock);
 544         return rc;
 545 }
 546
 547 int tls_device_sendpage(struct sock *sk, struct page *page,
 548                         int offset, size_t size, int flags)
 549 {
 550         struct tls_context *tls_ctx = tls_get_ctx(sk);
 551         struct iov_iter msg_iter;
 552         char *kaddr = kmap(page);
 553         struct kvec iov;
 554         int rc;
 555
 556         if (flags & MSG_SENDPAGE_NOTLAST)
 557                 flags |= MSG_MORE;
 558
 559         mutex_lock(&tls_ctx->tx_lock);
 560         lock_sock(sk);
 561
 562         if (flags & MSG_OOB) {
 563                 rc = -ENOTSUPP;
 564                 goto out;
 565         }
 566
 567         iov.iov_base = kaddr + offset;
 568         iov.iov_len = size;
 569         iov_iter_kvec(&msg_iter, WRITE, &iov, 1, size);
 570         rc = tls_push_data(sk, &msg_iter, size,
 571                            flags, TLS_RECORD_TYPE_DATA);
 572         kunmap(page);
 573
 574 out:
 575         release_sock(sk);
 576         mutex_unlock(&tls_ctx->tx_lock);
 577         return rc;
 578 }
 579
 580 struct tls_record_info *tls_get_record(struct tls_offload_context_tx *context,
 581                                        u32 seq, u64 *p_record_sn)
 582 {
 583         u64 record_sn = context->hint_record_sn;
 584         struct tls_record_info *info;
 585
 586         info = context->retransmit_hint;
 587         if (!info ||
 588             before(seq, info->end_seq - info->len)) {
 589                 /* if retransmit_hint is irrelevant start
 590                  * from the beggining of the list
 591                  */
 592                 info = list_first_entry_or_null(&context->records_list,
 593                                                 struct tls_record_info, list);
 594                 if (!info)
 595                         return NULL;
 596                 record_sn = context->unacked_record_sn;
 597         }
 598
 599         /* We just need the _rcu for the READ_ONCE() */
 600         rcu_read_lock();
 601         list_for_each_entry_from_rcu(info, &context->records_list, list) {
 602                 if (before(seq, info->end_seq)) {
 603                         if (!context->retransmit_hint ||
 604                             after(info->end_seq,
 605                                   context->retransmit_hint->end_seq)) {
 606                                 context->hint_record_sn = record_sn;
 607                                 context->retransmit_hint = info;
 608                         }
 609                         *p_record_sn = record_sn;
 610                         goto exit_rcu_unlock;
 611                 }
 612                 record_sn++;
 613         }
 614         info = NULL;
 615
 616 exit_rcu_unlock:
 617         rcu_read_unlock();
 618         return info;
 619 }
 620 EXPORT_SYMBOL(tls_get_record);
 621
 622 static int tls_device_push_pending_record(struct sock *sk, int flags)
 623 {
 624         struct iov_iter msg_iter;
 625
 626         iov_iter_kvec(&msg_iter, WRITE, NULL, 0, 0);
 627         return tls_push_data(sk, &msg_iter, 0, flags, TLS_RECORD_TYPE_DATA);
 628 }
 629
 630 void tls_device_write_space(struct sock *sk, struct tls_context *ctx)
 631 {
 632         if (tls_is_partially_sent_record(ctx)) {
 633                 gfp_t sk_allocation = sk->sk_allocation;
 634
 635                 WARN_ON_ONCE(sk->sk_write_pending);
 636
 637                 sk->sk_allocation = GFP_ATOMIC;
 638                 tls_push_partial_record(sk, ctx,
 639                                         MSG_DONTWAIT | MSG_NOSIGNAL |
 640                                         MSG_SENDPAGE_DECRYPTED);
 641                 sk->sk_allocation = sk_allocation;
 642         }
 643 }
 644
 645 static void tls_device_resync_rx(struct tls_context *tls_ctx,
 646                                  struct sock *sk, u32 seq, u8 *rcd_sn)
 647 {
 648         struct net_device *netdev;
 649
 650         if (WARN_ON(test_and_set_bit(TLS_RX_SYNC_RUNNING, &tls_ctx->flags)))
 651                 return;
 652         netdev = READ_ONCE(tls_ctx->netdev);
 653         if (netdev)
 654                 netdev->tlsdev_ops->tls_dev_resync(netdev, sk, seq, rcd_sn,
 655                                                    TLS_OFFLOAD_CTX_DIR_RX);
 656         clear_bit_unlock(TLS_RX_SYNC_RUNNING, &tls_ctx->flags);
 657 }
 658
 659 void tls_device_rx_resync_new_rec(struct sock *sk, u32 rcd_len, u32 seq)
 660 {
 661         struct tls_context *tls_ctx = tls_get_ctx(sk);
 662         struct tls_offload_context_rx *rx_ctx;
 663         u8 rcd_sn[TLS_MAX_REC_SEQ_SIZE];
 664         struct tls_prot_info *prot;
 665         u32 is_req_pending;
 666         s64 resync_req;
 667         u32 req_seq;
 668
 669         if (tls_ctx->rx_conf != TLS_HW)
 670                 return;
 671
 672         prot = &tls_ctx->prot_info;
 673         rx_ctx = tls_offload_ctx_rx(tls_ctx);
 674         memcpy(rcd_sn, tls_ctx->rx.rec_seq, prot->rec_seq_size);
 675
 676         switch (rx_ctx->resync_type) {
 677         case TLS_OFFLOAD_SYNC_TYPE_DRIVER_REQ:
 678                 resync_req = atomic64_read(&rx_ctx->resync_req);
 679                 req_seq = resync_req >> 32;
 680                 seq += TLS_HEADER_SIZE - 1;
 681                 is_req_pending = resync_req;
 682
 683                 if (likely(!is_req_pending) || req_seq != seq ||
 684                     !atomic64_try_cmpxchg(&rx_ctx->resync_req, &resync_req, 0))
 685                         return;
 686                 break;
 687         case TLS_OFFLOAD_SYNC_TYPE_CORE_NEXT_HINT:
 688                 if (likely(!rx_ctx->resync_nh_do_now))
 689                         return;
 690
 691                 /* head of next rec is already in, note that the sock_inq will
 692                  * include the currently parsed message when called from parser
 693                  */
 694                 if (tcp_inq(sk) > rcd_len)
 695                         return;
 696
 697                 rx_ctx->resync_nh_do_now = 0;
 698                 seq += rcd_len;
 699                 tls_bigint_increment(rcd_sn, prot->rec_seq_size);
 700                 break;
 701         }
 702
 703         tls_device_resync_rx(tls_ctx, sk, seq, rcd_sn);
 704 }
 705
 706 static void tls_device_core_ctrl_rx_resync(struct tls_context *tls_ctx,
 707                                            struct tls_offload_context_rx *ctx,
 708                                            struct sock *sk, struct sk_buff *skb)
 709 {
 710         struct strp_msg *rxm;
 711
 712         /* device will request resyncs by itself based on stream scan */
 713         if (ctx->resync_type != TLS_OFFLOAD_SYNC_TYPE_CORE_NEXT_HINT)
 714                 return;
 715         /* already scheduled */
 716         if (ctx->resync_nh_do_now)
 717                 return;
 718         /* seen decrypted fragments since last fully-failed record */
 719         if (ctx->resync_nh_reset) {
 720                 ctx->resync_nh_reset = 0;
 721                 ctx->resync_nh.decrypted_failed = 1;
 722                 ctx->resync_nh.decrypted_tgt = TLS_DEVICE_RESYNC_NH_START_IVAL;
 723                 return;
 724         }
 725
 726         if (++ctx->resync_nh.decrypted_failed <= ctx->resync_nh.decrypted_tgt)
 727                 return;
 728
 729         /* doing resync, bump the next target in case it fails */
 730         if (ctx->resync_nh.decrypted_tgt < TLS_DEVICE_RESYNC_NH_MAX_IVAL)
 731                 ctx->resync_nh.decrypted_tgt *= 2;
 732         else
 733                 ctx->resync_nh.decrypted_tgt += TLS_DEVICE_RESYNC_NH_MAX_IVAL;
 734
 735         rxm = strp_msg(skb);
 736
 737         /* head of next rec is already in, parser will sync for us */
 738         if (tcp_inq(sk) > rxm->full_len) {
 739                 ctx->resync_nh_do_now = 1;
 740         } else {
 741                 struct tls_prot_info *prot = &tls_ctx->prot_info;
 742                 u8 rcd_sn[TLS_MAX_REC_SEQ_SIZE];
 743
 744                 memcpy(rcd_sn, tls_ctx->rx.rec_seq, prot->rec_seq_size);
 745                 tls_bigint_increment(rcd_sn, prot->rec_seq_size);
 746
 747                 tls_device_resync_rx(tls_ctx, sk, tcp_sk(sk)->copied_seq,
 748                                      rcd_sn);
 749         }
 750 }
 751
 752 static int tls_device_reencrypt(struct sock *sk, struct sk_buff *skb)
 753 {
 754         struct strp_msg *rxm = strp_msg(skb);
 755         int err = 0, offset = rxm->offset, copy, nsg, data_len, pos;
 756         struct sk_buff *skb_iter, *unused;
 757         struct scatterlist sg[1];
 758         char *orig_buf, *buf;
 759
 760         orig_buf = kmalloc(rxm->full_len + TLS_HEADER_SIZE +
 761                            TLS_CIPHER_AES_GCM_128_IV_SIZE, sk->sk_allocation);
 762         if (!orig_buf)
 763                 return -ENOMEM;
 764         buf = orig_buf;
 765
 766         nsg = skb_cow_data(skb, 0, &unused);
 767         if (unlikely(nsg < 0)) {
 768                 err = nsg;
 769                 goto free_buf;
 770         }
 771
 772         sg_init_table(sg, 1);
 773         sg_set_buf(&sg[0], buf,
 774                    rxm->full_len + TLS_HEADER_SIZE +
 775                    TLS_CIPHER_AES_GCM_128_IV_SIZE);
 776         err = skb_copy_bits(skb, offset, buf,
 777                             TLS_HEADER_SIZE + TLS_CIPHER_AES_GCM_128_IV_SIZE);
 778         if (err)
 779                 goto free_buf;
 780
 781         /* We are interested only in the decrypted data not the auth */
 782         err = decrypt_skb(sk, skb, sg);
 783         if (err != -EBADMSG)
 784                 goto free_buf;
 785         else
 786                 err = 0;
 787
 788         data_len = rxm->full_len - TLS_CIPHER_AES_GCM_128_TAG_SIZE;
 789
 790         if (skb_pagelen(skb) > offset) {
 791                 copy = min_t(int, skb_pagelen(skb) - offset, data_len);
 792
 793                 if (skb->decrypted) {
 794                         err = skb_store_bits(skb, offset, buf, copy);
 795                         if (err)
 796                                 goto free_buf;
 797                 }
 798
 799                 offset += copy;
 800                 buf += copy;
 801         }
 802
 803         pos = skb_pagelen(skb);
 804         skb_walk_frags(skb, skb_iter) {
 805                 int frag_pos;
 806
 807                 /* Practically all frags must belong to msg if reencrypt
 808                  * is needed with current strparser and coalescing logic,
 809                  * but strparser may "get optimized", so let's be safe.
 810                  */
 811                 if (pos + skb_iter->len <= offset)
 812                         goto done_with_frag;
 813                 if (pos >= data_len + rxm->offset)
 814                         break;
 815
 816                 frag_pos = offset - pos;
 817                 copy = min_t(int, skb_iter->len - frag_pos,
 818                              data_len + rxm->offset - offset);
 819
 820                 if (skb_iter->decrypted) {
 821                         err = skb_store_bits(skb_iter, frag_pos, buf, copy);
 822                         if (err)
 823                                 goto free_buf;
 824                 }
 825
 826                 offset += copy;
 827                 buf += copy;
 828 done_with_frag:
 829                 pos += skb_iter->len;
 830         }
 831
 832 free_buf:
 833         kfree(orig_buf);
 834         return err;
 835 }
 836
 837 int tls_device_decrypted(struct sock *sk, struct sk_buff *skb)
 838 {
 839         struct tls_context *tls_ctx = tls_get_ctx(sk);
 840         struct tls_offload_context_rx *ctx = tls_offload_ctx_rx(tls_ctx);
 841         int is_decrypted = skb->decrypted;
 842         int is_encrypted = !is_decrypted;
 843         struct sk_buff *skb_iter;
 844
 845         /* Check if all the data is decrypted already */
 846         skb_walk_frags(skb, skb_iter) {
 847                 is_decrypted &= skb_iter->decrypted;
 848                 is_encrypted &= !skb_iter->decrypted;
 849         }
 850
 851         ctx->sw.decrypted |= is_decrypted;
 852
 853         /* Return immediately if the record is either entirely plaintext or
 854          * entirely ciphertext. Otherwise handle reencrypt partially decrypted
 855          * record.
 856          */
 857         if (is_decrypted) {
 858                 ctx->resync_nh_reset = 1;
 859                 return 0;
 860         }
 861         if (is_encrypted) {
 862                 tls_device_core_ctrl_rx_resync(tls_ctx, ctx, sk, skb);
 863                 return 0;
 864         }
 865
 866         ctx->resync_nh_reset = 1;
 867         return tls_device_reencrypt(sk, skb);
 868 }
 869
 870 static void tls_device_attach(struct tls_context *ctx, struct sock *sk,
 871                               struct net_device *netdev)
 872 {
 873         if (sk->sk_destruct != tls_device_sk_destruct) {
 874                 refcount_set(&ctx->refcount, 1);
 875                 dev_hold(netdev);
 876                 ctx->netdev = netdev;
 877                 spin_lock_irq(&tls_device_lock);
 878                 list_add_tail(&ctx->list, &tls_device_list);
 879                 spin_unlock_irq(&tls_device_lock);
 880
 881                 ctx->sk_destruct = sk->sk_destruct;
 882                 sk->sk_destruct = tls_device_sk_destruct;
 883         }
 884 }
 885
 886 int tls_set_device_offload(struct sock *sk, struct tls_context *ctx)
 887 {
 888         u16 nonce_size, tag_size, iv_size, rec_seq_size;
 889         struct tls_context *tls_ctx = tls_get_ctx(sk);
 890         struct tls_prot_info *prot = &tls_ctx->prot_info;
 891         struct tls_record_info *start_marker_record;
 892         struct tls_offload_context_tx *offload_ctx;
 893         struct tls_crypto_info *crypto_info;
 894         struct net_device *netdev;
 895         char *iv, *rec_seq;
 896         struct sk_buff *skb;
 897         __be64 rcd_sn;
 898         int rc;
 899
 900         if (!ctx)
 901                 return -EINVAL;
 902
 903         if (ctx->priv_ctx_tx)
 904                 return -EEXIST;
 905
 906         start_marker_record = kmalloc(sizeof(*start_marker_record), GFP_KERNEL);
 907         if (!start_marker_record)
 908                 return -ENOMEM;
 909
 910         offload_ctx = kzalloc(TLS_OFFLOAD_CONTEXT_SIZE_TX, GFP_KERNEL);
 911         if (!offload_ctx) {
 912                 rc = -ENOMEM;
 913                 goto free_marker_record;
 914         }
 915
 916         crypto_info = &ctx->crypto_send.info;
 917         if (crypto_info->version != TLS_1_2_VERSION) {
 918                 rc = -EOPNOTSUPP;
 919                 goto free_offload_ctx;
 920         }
 921
 922         switch (crypto_info->cipher_type) {
 923         case TLS_CIPHER_AES_GCM_128:
 924                 nonce_size = TLS_CIPHER_AES_GCM_128_IV_SIZE;
 925                 tag_size = TLS_CIPHER_AES_GCM_128_TAG_SIZE;
 926                 iv_size = TLS_CIPHER_AES_GCM_128_IV_SIZE;
 927                 iv = ((struct tls12_crypto_info_aes_gcm_128 *)crypto_info)->iv;
 928                 rec_seq_size = TLS_CIPHER_AES_GCM_128_REC_SEQ_SIZE;
 929                 rec_seq =
 930                  ((struct tls12_crypto_info_aes_gcm_128 *)crypto_info)->rec_seq;
 931                 break;
 932         default:
 933                 rc = -EINVAL;
 934                 goto free_offload_ctx;
 935         }
 936
 937         /* Sanity-check the rec_seq_size for stack allocations */
 938         if (rec_seq_size > TLS_MAX_REC_SEQ_SIZE) {
 939                 rc = -EINVAL;
 940                 goto free_offload_ctx;
 941         }
 942
 943         prot->version = crypto_info->version;
 944         prot->cipher_type = crypto_info->cipher_type;
 945         prot->prepend_size = TLS_HEADER_SIZE + nonce_size;
 946         prot->tag_size = tag_size;
 947         prot->overhead_size = prot->prepend_size + prot->tag_size;
 948         prot->iv_size = iv_size;
 949         ctx->tx.iv = kmalloc(iv_size + TLS_CIPHER_AES_GCM_128_SALT_SIZE,
 950                              GFP_KERNEL);
 951         if (!ctx->tx.iv) {
 952                 rc = -ENOMEM;
 953                 goto free_offload_ctx;
 954         }
 955
 956         memcpy(ctx->tx.iv + TLS_CIPHER_AES_GCM_128_SALT_SIZE, iv, iv_size);
 957
 958         prot->rec_seq_size = rec_seq_size;
 959         ctx->tx.rec_seq = kmemdup(rec_seq, rec_seq_size, GFP_KERNEL);
 960         if (!ctx->tx.rec_seq) {
 961                 rc = -ENOMEM;
 962                 goto free_iv;
 963         }
 964
 965         rc = tls_sw_fallback_init(sk, offload_ctx, crypto_info);
 966         if (rc)
 967                 goto free_rec_seq;
 968
 969         /* start at rec_seq - 1 to account for the start marker record */
 970         memcpy(&rcd_sn, ctx->tx.rec_seq, sizeof(rcd_sn));
 971         offload_ctx->unacked_record_sn = be64_to_cpu(rcd_sn) - 1;
 972
 973         start_marker_record->end_seq = tcp_sk(sk)->write_seq;
 974         start_marker_record->len = 0;
 975         start_marker_record->num_frags = 0;
 976
 977         INIT_LIST_HEAD(&offload_ctx->records_list);
 978         list_add_tail(&start_marker_record->list, &offload_ctx->records_list);
 979         spin_lock_init(&offload_ctx->lock);
 980         sg_init_table(offload_ctx->sg_tx_data,
 981                       ARRAY_SIZE(offload_ctx->sg_tx_data));
 982
 983         clean_acked_data_enable(inet_csk(sk), &tls_icsk_clean_acked);
 984         ctx->push_pending_record = tls_device_push_pending_record;
 985
 986         /* TLS offload is greatly simplified if we don't send
 987          * SKBs where only part of the payload needs to be encrypted.
 988          * So mark the last skb in the write queue as end of record.
 989          */
 990         skb = tcp_write_queue_tail(sk);
 991         if (skb)
 992                 TCP_SKB_CB(skb)->eor = 1;
 993
 994         netdev = get_netdev_for_sock(sk);
 995         if (!netdev) {
 996                 pr_err_ratelimited("%s: netdev not found\n", __func__);
 997                 rc = -EINVAL;
 998                 goto disable_cad;
 999         }
1000
1001         if (!(netdev->features & NETIF_F_HW_TLS_TX)) {
1002                 rc = -ENOTSUPP;
1003                 goto release_netdev;
1004         }
1005
1006         /* Avoid offloading if the device is down
1007          * We don't want to offload new flows after
1008          * the NETDEV_DOWN event
1009          *
1010          * device_offload_lock is taken in tls_devices's NETDEV_DOWN
1011          * handler thus protecting from the device going down before
1012          * ctx was added to tls_device_list.
1013          */
1014         down_read(&device_offload_lock);
1015         if (!(netdev->flags & IFF_UP)) {
1016                 rc = -EINVAL;
1017                 goto release_lock;
1018         }
1019
1020         ctx->priv_ctx_tx = offload_ctx;
1021         rc = netdev->tlsdev_ops->tls_dev_add(netdev, sk, TLS_OFFLOAD_CTX_DIR_TX,
1022                                              &ctx->crypto_send.info,
1023                                              tcp_sk(sk)->write_seq);
1024         if (rc)
1025                 goto release_lock;
1026
1027         tls_device_attach(ctx, sk, netdev);
1028         up_read(&device_offload_lock);
1029
1030         /* following this assignment tls_is_sk_tx_device_offloaded
1031          * will return true and the context might be accessed
1032          * by the netdev's xmit function.
1033          */
1034         smp_store_release(&sk->sk_validate_xmit_skb, tls_validate_xmit_skb);
1035         dev_put(netdev);
1036
1037         return 0;
1038
1039 release_lock:
1040         up_read(&device_offload_lock);
1041 release_netdev:
1042         dev_put(netdev);
1043 disable_cad:
1044         clean_acked_data_disable(inet_csk(sk));
1045         crypto_free_aead(offload_ctx->aead_send);
1046 free_rec_seq:
1047         kfree(ctx->tx.rec_seq);
1048 free_iv:
1049         kfree(ctx->tx.iv);
1050 free_offload_ctx:
1051         kfree(offload_ctx);
1052         ctx->priv_ctx_tx = NULL;
1053 free_marker_record:
1054         kfree(start_marker_record);
1055         return rc;
1056 }
1057
1058 int tls_set_device_offload_rx(struct sock *sk, struct tls_context *ctx)
1059 {
1060         struct tls_offload_context_rx *context;
1061         struct net_device *netdev;
1062         int rc = 0;
1063
1064         if (ctx->crypto_recv.info.version != TLS_1_2_VERSION)
1065                 return -EOPNOTSUPP;
1066
1067         netdev = get_netdev_for_sock(sk);
1068         if (!netdev) {
1069                 pr_err_ratelimited("%s: netdev not found\n", __func__);
1070                 return -EINVAL;
1071         }
1072
1073         if (!(netdev->features & NETIF_F_HW_TLS_RX)) {
1074                 rc = -ENOTSUPP;
1075                 goto release_netdev;
1076         }
1077
1078         /* Avoid offloading if the device is down
1079          * We don't want to offload new flows after
1080          * the NETDEV_DOWN event
1081          *
1082          * device_offload_lock is taken in tls_devices's NETDEV_DOWN
1083          * handler thus protecting from the device going down before
1084          * ctx was added to tls_device_list.
1085          */
1086         down_read(&device_offload_lock);
1087         if (!(netdev->flags & IFF_UP)) {
1088                 rc = -EINVAL;
1089                 goto release_lock;
1090         }
1091
1092         context = kzalloc(TLS_OFFLOAD_CONTEXT_SIZE_RX, GFP_KERNEL);
1093         if (!context) {
1094                 rc = -ENOMEM;
1095                 goto release_lock;
1096         }
1097         context->resync_nh_reset = 1;
1098
1099         ctx->priv_ctx_rx = context;
1100         rc = tls_set_sw_offload(sk, ctx, 0);
1101         if (rc)
1102                 goto release_ctx;
1103
1104         rc = netdev->tlsdev_ops->tls_dev_add(netdev, sk, TLS_OFFLOAD_CTX_DIR_RX,
1105                                              &ctx->crypto_recv.info,
1106                                              tcp_sk(sk)->copied_seq);
1107         if (rc)
1108                 goto free_sw_resources;
1109
1110         tls_device_attach(ctx, sk, netdev);
1111         up_read(&device_offload_lock);
1112
1113         dev_put(netdev);
1114
1115         return 0;
1116
1117 free_sw_resources:
1118         up_read(&device_offload_lock);
1119         tls_sw_free_resources_rx(sk);
1120         down_read(&device_offload_lock);
1121 release_ctx:
1122         ctx->priv_ctx_rx = NULL;
1123 release_lock:
1124         up_read(&device_offload_lock);
1125 release_netdev:
1126         dev_put(netdev);
1127         return rc;
1128 }
1129
1130 void tls_device_offload_cleanup_rx(struct sock *sk)
1131 {
1132         struct tls_context *tls_ctx = tls_get_ctx(sk);
1133         struct net_device *netdev;
1134
1135         down_read(&device_offload_lock);
1136         netdev = tls_ctx->netdev;
1137         if (!netdev)
1138                 goto out;
1139
1140         netdev->tlsdev_ops->tls_dev_del(netdev, tls_ctx,
1141                                         TLS_OFFLOAD_CTX_DIR_RX);
1142
1143         if (tls_ctx->tx_conf != TLS_HW) {
1144                 dev_put(netdev);
1145                 tls_ctx->netdev = NULL;
1146         }
1147 out:
1148         up_read(&device_offload_lock);
1149         tls_sw_release_resources_rx(sk);
1150 }
1151
1152 static int tls_device_down(struct net_device *netdev)
1153 {
1154         struct tls_context *ctx, *tmp;
1155         unsigned long flags;
1156         LIST_HEAD(list);
1157
1158         /* Request a write lock to block new offload attempts */
1159         down_write(&device_offload_lock);
1160
1161         spin_lock_irqsave(&tls_device_lock, flags);
1162         list_for_each_entry_safe(ctx, tmp, &tls_device_list, list) {
1163                 if (ctx->netdev != netdev ||
1164                     !refcount_inc_not_zero(&ctx->refcount))
1165                         continue;
1166
1167                 list_move(&ctx->list, &list);
1168         }
1169         spin_unlock_irqrestore(&tls_device_lock, flags);
1170
1171         list_for_each_entry_safe(ctx, tmp, &list, list) {
1172                 if (ctx->tx_conf == TLS_HW)
1173                         netdev->tlsdev_ops->tls_dev_del(netdev, ctx,
1174                                                         TLS_OFFLOAD_CTX_DIR_TX);
1175                 if (ctx->rx_conf == TLS_HW)
1176                         netdev->tlsdev_ops->tls_dev_del(netdev, ctx,
1177                                                         TLS_OFFLOAD_CTX_DIR_RX);
1178                 WRITE_ONCE(ctx->netdev, NULL);
1179                 smp_mb__before_atomic(); /* pairs with test_and_set_bit() */
1180                 while (test_bit(TLS_RX_SYNC_RUNNING, &ctx->flags))
1181                         usleep_range(10, 200);
1182                 dev_put(netdev);
1183                 list_del_init(&ctx->list);
1184
1185                 if (refcount_dec_and_test(&ctx->refcount))
1186                         tls_device_free_ctx(ctx);
1187         }
1188
1189         up_write(&device_offload_lock);
1190
1191         flush_work(&tls_device_gc_work);
1192
1193         return NOTIFY_DONE;
1194 }
1195
1196 static int tls_dev_event(struct notifier_block *this, unsigned long event,
1197                          void *ptr)
1198 {
1199         struct net_device *dev = netdev_notifier_info_to_dev(ptr);
1200
1201         if (!dev->tlsdev_ops &&
1202             !(dev->features & (NETIF_F_HW_TLS_RX | NETIF_F_HW_TLS_TX)))
1203                 return NOTIFY_DONE;
1204
1205         switch (event) {
1206         case NETDEV_REGISTER:
1207         case NETDEV_FEAT_CHANGE:
1208                 if ((dev->features & NETIF_F_HW_TLS_RX) &&
1209                     !dev->tlsdev_ops->tls_dev_resync)
1210                         return NOTIFY_BAD;
1211
1212                 if  (dev->tlsdev_ops &&
1213                      dev->tlsdev_ops->tls_dev_add &&
1214                      dev->tlsdev_ops->tls_dev_del)
1215                         return NOTIFY_DONE;
1216                 else
1217                         return NOTIFY_BAD;
1218         case NETDEV_DOWN:
1219                 return tls_device_down(dev);
1220         }
1221         return NOTIFY_DONE;
1222 }
1223
1224 static struct notifier_block tls_dev_notifier = {
1225         .notifier_call  = tls_dev_event,
1226 };
1227
1228 void __init tls_device_init(void)
1229 {
1230         register_netdevice_notifier(&tls_dev_notifier);
1231 }
1232
1233 void __exit tls_device_cleanup(void)
1234 {
1235         unregister_netdevice_notifier(&tls_dev_notifier);
1236         flush_work(&tls_device_gc_work);
1237         clean_acked_data_flush();
1238 }