net/tls/tls_device.c

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