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 #include "tls.h"
  42 #include "trace.h"
  43
  44 /* device_offload_lock is used to synchronize tls_dev_add
  45  * against NETDEV_DOWN notifications.
  46  */
  47 static DECLARE_RWSEM(device_offload_lock);
  48
  49 static struct workqueue_struct *destruct_wq __read_mostly;
  50
  51 static LIST_HEAD(tls_device_list);
  52 static LIST_HEAD(tls_device_down_list);
  53 static DEFINE_SPINLOCK(tls_device_lock);
  54
  55 static struct page *dummy_page;
  56
  57 static void tls_device_free_ctx(struct tls_context *ctx)
  58 {
  59         if (ctx->tx_conf == TLS_HW) {
  60                 kfree(tls_offload_ctx_tx(ctx));
  61                 kfree(ctx->tx.iv);
  62         }
  63
  64         if (ctx->rx_conf == TLS_HW)
  65                 kfree(tls_offload_ctx_rx(ctx));
  66
  67         tls_ctx_free(NULL, ctx);
  68 }
  69
  70 static void tls_device_tx_del_task(struct work_struct *work)
  71 {
  72         struct tls_offload_context_tx *offload_ctx =
  73                 container_of(work, struct tls_offload_context_tx, destruct_work);
  74         struct tls_context *ctx = offload_ctx->ctx;
  75         struct net_device *netdev;
  76
  77         /* Safe, because this is the destroy flow, refcount is 0, so
  78          * tls_device_down can't store this field in parallel.
  79          */
  80         netdev = rcu_dereference_protected(ctx->netdev,
  81                                            !refcount_read(&ctx->refcount));
  82
  83         netdev->tlsdev_ops->tls_dev_del(netdev, ctx, TLS_OFFLOAD_CTX_DIR_TX);
  84         dev_put(netdev);
  85         ctx->netdev = NULL;
  86         tls_device_free_ctx(ctx);
  87 }
  88
  89 static void tls_device_queue_ctx_destruction(struct tls_context *ctx)
  90 {
  91         struct net_device *netdev;
  92         unsigned long flags;
  93         bool async_cleanup;
  94
  95         spin_lock_irqsave(&tls_device_lock, flags);
  96         if (unlikely(!refcount_dec_and_test(&ctx->refcount))) {
  97                 spin_unlock_irqrestore(&tls_device_lock, flags);
  98                 return;
  99         }
 100
 101         list_del(&ctx->list); /* Remove from tls_device_list / tls_device_down_list */
 102
 103         /* Safe, because this is the destroy flow, refcount is 0, so
 104          * tls_device_down can't store this field in parallel.
 105          */
 106         netdev = rcu_dereference_protected(ctx->netdev,
 107                                            !refcount_read(&ctx->refcount));
 108
 109         async_cleanup = netdev && ctx->tx_conf == TLS_HW;
 110         if (async_cleanup) {
 111                 struct tls_offload_context_tx *offload_ctx = tls_offload_ctx_tx(ctx);
 112
 113                 /* queue_work inside the spinlock
 114                  * to make sure tls_device_down waits for that work.
 115                  */
 116                 queue_work(destruct_wq, &offload_ctx->destruct_work);
 117         }
 118         spin_unlock_irqrestore(&tls_device_lock, flags);
 119
 120         if (!async_cleanup)
 121                 tls_device_free_ctx(ctx);
 122 }
 123
 124 /* We assume that the socket is already connected */
 125 static struct net_device *get_netdev_for_sock(struct sock *sk)
 126 {
 127         struct dst_entry *dst = sk_dst_get(sk);
 128         struct net_device *netdev = NULL;
 129
 130         if (likely(dst)) {
 131                 netdev = netdev_sk_get_lowest_dev(dst->dev, sk);
 132                 dev_hold(netdev);
 133         }
 134
 135         dst_release(dst);
 136
 137         return netdev;
 138 }
 139
 140 static void destroy_record(struct tls_record_info *record)
 141 {
 142         int i;
 143
 144         for (i = 0; i < record->num_frags; i++)
 145                 __skb_frag_unref(&record->frags[i], false);
 146         kfree(record);
 147 }
 148
 149 static void delete_all_records(struct tls_offload_context_tx *offload_ctx)
 150 {
 151         struct tls_record_info *info, *temp;
 152
 153         list_for_each_entry_safe(info, temp, &offload_ctx->records_list, list) {
 154                 list_del(&info->list);
 155                 destroy_record(info);
 156         }
 157
 158         offload_ctx->retransmit_hint = NULL;
 159 }
 160
 161 static void tls_icsk_clean_acked(struct sock *sk, u32 acked_seq)
 162 {
 163         struct tls_context *tls_ctx = tls_get_ctx(sk);
 164         struct tls_record_info *info, *temp;
 165         struct tls_offload_context_tx *ctx;
 166         u64 deleted_records = 0;
 167         unsigned long flags;
 168
 169         if (!tls_ctx)
 170                 return;
 171
 172         ctx = tls_offload_ctx_tx(tls_ctx);
 173
 174         spin_lock_irqsave(&ctx->lock, flags);
 175         info = ctx->retransmit_hint;
 176         if (info && !before(acked_seq, info->end_seq))
 177                 ctx->retransmit_hint = NULL;
 178
 179         list_for_each_entry_safe(info, temp, &ctx->records_list, list) {
 180                 if (before(acked_seq, info->end_seq))
 181                         break;
 182                 list_del(&info->list);
 183
 184                 destroy_record(info);
 185                 deleted_records++;
 186         }
 187
 188         ctx->unacked_record_sn += deleted_records;
 189         spin_unlock_irqrestore(&ctx->lock, flags);
 190 }
 191
 192 /* At this point, there should be no references on this
 193  * socket and no in-flight SKBs associated with this
 194  * socket, so it is safe to free all the resources.
 195  */
 196 void tls_device_sk_destruct(struct sock *sk)
 197 {
 198         struct tls_context *tls_ctx = tls_get_ctx(sk);
 199         struct tls_offload_context_tx *ctx = tls_offload_ctx_tx(tls_ctx);
 200
 201         tls_ctx->sk_destruct(sk);
 202
 203         if (tls_ctx->tx_conf == TLS_HW) {
 204                 if (ctx->open_record)
 205                         destroy_record(ctx->open_record);
 206                 delete_all_records(ctx);
 207                 crypto_free_aead(ctx->aead_send);
 208                 clean_acked_data_disable(inet_csk(sk));
 209         }
 210
 211         tls_device_queue_ctx_destruction(tls_ctx);
 212 }
 213 EXPORT_SYMBOL_GPL(tls_device_sk_destruct);
 214
 215 void tls_device_free_resources_tx(struct sock *sk)
 216 {
 217         struct tls_context *tls_ctx = tls_get_ctx(sk);
 218
 219         tls_free_partial_record(sk, tls_ctx);
 220 }
 221
 222 void tls_offload_tx_resync_request(struct sock *sk, u32 got_seq, u32 exp_seq)
 223 {
 224         struct tls_context *tls_ctx = tls_get_ctx(sk);
 225
 226         trace_tls_device_tx_resync_req(sk, got_seq, exp_seq);
 227         WARN_ON(test_and_set_bit(TLS_TX_SYNC_SCHED, &tls_ctx->flags));
 228 }
 229 EXPORT_SYMBOL_GPL(tls_offload_tx_resync_request);
 230
 231 static void tls_device_resync_tx(struct sock *sk, struct tls_context *tls_ctx,
 232                                  u32 seq)
 233 {
 234         struct net_device *netdev;
 235         struct sk_buff *skb;
 236         int err = 0;
 237         u8 *rcd_sn;
 238
 239         skb = tcp_write_queue_tail(sk);
 240         if (skb)
 241                 TCP_SKB_CB(skb)->eor = 1;
 242
 243         rcd_sn = tls_ctx->tx.rec_seq;
 244
 245         trace_tls_device_tx_resync_send(sk, seq, rcd_sn);
 246         down_read(&device_offload_lock);
 247         netdev = rcu_dereference_protected(tls_ctx->netdev,
 248                                            lockdep_is_held(&device_offload_lock));
 249         if (netdev)
 250                 err = netdev->tlsdev_ops->tls_dev_resync(netdev, sk, seq,
 251                                                          rcd_sn,
 252                                                          TLS_OFFLOAD_CTX_DIR_TX);
 253         up_read(&device_offload_lock);
 254         if (err)
 255                 return;
 256
 257         clear_bit_unlock(TLS_TX_SYNC_SCHED, &tls_ctx->flags);
 258 }
 259
 260 static void tls_append_frag(struct tls_record_info *record,
 261                             struct page_frag *pfrag,
 262                             int size)
 263 {
 264         skb_frag_t *frag;
 265
 266         frag = &record->frags[record->num_frags - 1];
 267         if (skb_frag_page(frag) == pfrag->page &&
 268             skb_frag_off(frag) + skb_frag_size(frag) == pfrag->offset) {
 269                 skb_frag_size_add(frag, size);
 270         } else {
 271                 ++frag;
 272                 skb_frag_fill_page_desc(frag, pfrag->page, pfrag->offset,
 273                                         size);
 274                 ++record->num_frags;
 275                 get_page(pfrag->page);
 276         }
 277
 278         pfrag->offset += size;
 279         record->len += size;
 280 }
 281
 282 static int tls_push_record(struct sock *sk,
 283                            struct tls_context *ctx,
 284                            struct tls_offload_context_tx *offload_ctx,
 285                            struct tls_record_info *record,
 286                            int flags)
 287 {
 288         struct tls_prot_info *prot = &ctx->prot_info;
 289         struct tcp_sock *tp = tcp_sk(sk);
 290         skb_frag_t *frag;
 291         int i;
 292
 293         record->end_seq = tp->write_seq + record->len;
 294         list_add_tail_rcu(&record->list, &offload_ctx->records_list);
 295         offload_ctx->open_record = NULL;
 296
 297         if (test_bit(TLS_TX_SYNC_SCHED, &ctx->flags))
 298                 tls_device_resync_tx(sk, ctx, tp->write_seq);
 299
 300         tls_advance_record_sn(sk, prot, &ctx->tx);
 301
 302         for (i = 0; i < record->num_frags; i++) {
 303                 frag = &record->frags[i];
 304                 sg_unmark_end(&offload_ctx->sg_tx_data[i]);
 305                 sg_set_page(&offload_ctx->sg_tx_data[i], skb_frag_page(frag),
 306                             skb_frag_size(frag), skb_frag_off(frag));
 307                 sk_mem_charge(sk, skb_frag_size(frag));
 308                 get_page(skb_frag_page(frag));
 309         }
 310         sg_mark_end(&offload_ctx->sg_tx_data[record->num_frags - 1]);
 311
 312         /* all ready, send */
 313         return tls_push_sg(sk, ctx, offload_ctx->sg_tx_data, 0, flags);
 314 }
 315
 316 static void tls_device_record_close(struct sock *sk,
 317                                     struct tls_context *ctx,
 318                                     struct tls_record_info *record,
 319                                     struct page_frag *pfrag,
 320                                     unsigned char record_type)
 321 {
 322         struct tls_prot_info *prot = &ctx->prot_info;
 323         struct page_frag dummy_tag_frag;
 324
 325         /* append tag
 326          * device will fill in the tag, we just need to append a placeholder
 327          * use socket memory to improve coalescing (re-using a single buffer
 328          * increases frag count)
 329          * if we can't allocate memory now use the dummy page
 330          */
 331         if (unlikely(pfrag->size - pfrag->offset < prot->tag_size) &&
 332             !skb_page_frag_refill(prot->tag_size, pfrag, sk->sk_allocation)) {
 333                 dummy_tag_frag.page = dummy_page;
 334                 dummy_tag_frag.offset = 0;
 335                 pfrag = &dummy_tag_frag;
 336         }
 337         tls_append_frag(record, pfrag, prot->tag_size);
 338
 339         /* fill prepend */
 340         tls_fill_prepend(ctx, skb_frag_address(&record->frags[0]),
 341                          record->len - prot->overhead_size,
 342                          record_type);
 343 }
 344
 345 static int tls_create_new_record(struct tls_offload_context_tx *offload_ctx,
 346                                  struct page_frag *pfrag,
 347                                  size_t prepend_size)
 348 {
 349         struct tls_record_info *record;
 350         skb_frag_t *frag;
 351
 352         record = kmalloc(sizeof(*record), GFP_KERNEL);
 353         if (!record)
 354                 return -ENOMEM;
 355
 356         frag = &record->frags[0];
 357         skb_frag_fill_page_desc(frag, pfrag->page, pfrag->offset,
 358                                 prepend_size);
 359
 360         get_page(pfrag->page);
 361         pfrag->offset += prepend_size;
 362
 363         record->num_frags = 1;
 364         record->len = prepend_size;
 365         offload_ctx->open_record = record;
 366         return 0;
 367 }
 368
 369 static int tls_do_allocation(struct sock *sk,
 370                              struct tls_offload_context_tx *offload_ctx,
 371                              struct page_frag *pfrag,
 372                              size_t prepend_size)
 373 {
 374         int ret;
 375
 376         if (!offload_ctx->open_record) {
 377                 if (unlikely(!skb_page_frag_refill(prepend_size, pfrag,
 378                                                    sk->sk_allocation))) {
 379                         READ_ONCE(sk->sk_prot)->enter_memory_pressure(sk);
 380                         sk_stream_moderate_sndbuf(sk);
 381                         return -ENOMEM;
 382                 }
 383
 384                 ret = tls_create_new_record(offload_ctx, pfrag, prepend_size);
 385                 if (ret)
 386                         return ret;
 387
 388                 if (pfrag->size > pfrag->offset)
 389                         return 0;
 390         }
 391
 392         if (!sk_page_frag_refill(sk, pfrag))
 393                 return -ENOMEM;
 394
 395         return 0;
 396 }
 397
 398 static int tls_device_copy_data(void *addr, size_t bytes, struct iov_iter *i)
 399 {
 400         size_t pre_copy, nocache;
 401
 402         pre_copy = ~((unsigned long)addr - 1) & (SMP_CACHE_BYTES - 1);
 403         if (pre_copy) {
 404                 pre_copy = min(pre_copy, bytes);
 405                 if (copy_from_iter(addr, pre_copy, i) != pre_copy)
 406                         return -EFAULT;
 407                 bytes -= pre_copy;
 408                 addr += pre_copy;
 409         }
 410
 411         nocache = round_down(bytes, SMP_CACHE_BYTES);
 412         if (copy_from_iter_nocache(addr, nocache, i) != nocache)
 413                 return -EFAULT;
 414         bytes -= nocache;
 415         addr += nocache;
 416
 417         if (bytes && copy_from_iter(addr, bytes, i) != bytes)
 418                 return -EFAULT;
 419
 420         return 0;
 421 }
 422
 423 static int tls_push_data(struct sock *sk,
 424                          struct iov_iter *iter,
 425                          size_t size, int flags,
 426                          unsigned char record_type)
 427 {
 428         struct tls_context *tls_ctx = tls_get_ctx(sk);
 429         struct tls_prot_info *prot = &tls_ctx->prot_info;
 430         struct tls_offload_context_tx *ctx = tls_offload_ctx_tx(tls_ctx);
 431         struct tls_record_info *record;
 432         int tls_push_record_flags;
 433         struct page_frag *pfrag;
 434         size_t orig_size = size;
 435         u32 max_open_record_len;
 436         bool more = false;
 437         bool done = false;
 438         int copy, rc = 0;
 439         long timeo;
 440
 441         if (flags &
 442             ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL |
 443               MSG_SPLICE_PAGES | MSG_EOR))
 444                 return -EOPNOTSUPP;
 445
 446         if ((flags & (MSG_MORE | MSG_EOR)) == (MSG_MORE | MSG_EOR))
 447                 return -EINVAL;
 448
 449         if (unlikely(sk->sk_err))
 450                 return -sk->sk_err;
 451
 452         flags |= MSG_SENDPAGE_DECRYPTED;
 453         tls_push_record_flags = flags | MSG_MORE;
 454
 455         timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
 456         if (tls_is_partially_sent_record(tls_ctx)) {
 457                 rc = tls_push_partial_record(sk, tls_ctx, flags);
 458                 if (rc < 0)
 459                         return rc;
 460         }
 461
 462         pfrag = sk_page_frag(sk);
 463
 464         /* TLS_HEADER_SIZE is not counted as part of the TLS record, and
 465          * we need to leave room for an authentication tag.
 466          */
 467         max_open_record_len = TLS_MAX_PAYLOAD_SIZE +
 468                               prot->prepend_size;
 469         do {
 470                 rc = tls_do_allocation(sk, ctx, pfrag, prot->prepend_size);
 471                 if (unlikely(rc)) {
 472                         rc = sk_stream_wait_memory(sk, &timeo);
 473                         if (!rc)
 474                                 continue;
 475
 476                         record = ctx->open_record;
 477                         if (!record)
 478                                 break;
 479 handle_error:
 480                         if (record_type != TLS_RECORD_TYPE_DATA) {
 481                                 /* avoid sending partial
 482                                  * record with type !=
 483                                  * application_data
 484                                  */
 485                                 size = orig_size;
 486                                 destroy_record(record);
 487                                 ctx->open_record = NULL;
 488                         } else if (record->len > prot->prepend_size) {
 489                                 goto last_record;
 490                         }
 491
 492                         break;
 493                 }
 494
 495                 record = ctx->open_record;
 496
 497                 copy = min_t(size_t, size, max_open_record_len - record->len);
 498                 if (copy && (flags & MSG_SPLICE_PAGES)) {
 499                         struct page_frag zc_pfrag;
 500                         struct page **pages = &zc_pfrag.page;
 501                         size_t off;
 502
 503                         rc = iov_iter_extract_pages(iter, &pages,
 504                                                     copy, 1, 0, &off);
 505                         if (rc <= 0) {
 506                                 if (rc == 0)
 507                                         rc = -EIO;
 508                                 goto handle_error;
 509                         }
 510                         copy = rc;
 511
 512                         if (WARN_ON_ONCE(!sendpage_ok(zc_pfrag.page))) {
 513                                 iov_iter_revert(iter, copy);
 514                                 rc = -EIO;
 515                                 goto handle_error;
 516                         }
 517
 518                         zc_pfrag.offset = off;
 519                         zc_pfrag.size = copy;
 520                         tls_append_frag(record, &zc_pfrag, copy);
 521                 } else if (copy) {
 522                         copy = min_t(size_t, copy, pfrag->size - pfrag->offset);
 523
 524                         rc = tls_device_copy_data(page_address(pfrag->page) +
 525                                                   pfrag->offset, copy,
 526                                                   iter);
 527                         if (rc)
 528                                 goto handle_error;
 529                         tls_append_frag(record, pfrag, copy);
 530                 }
 531
 532                 size -= copy;
 533                 if (!size) {
 534 last_record:
 535                         tls_push_record_flags = flags;
 536                         if (flags & MSG_MORE) {
 537                                 more = true;
 538                                 break;
 539                         }
 540
 541                         done = true;
 542                 }
 543
 544                 if (done || record->len >= max_open_record_len ||
 545                     (record->num_frags >= MAX_SKB_FRAGS - 1)) {
 546                         tls_device_record_close(sk, tls_ctx, record,
 547                                                 pfrag, record_type);
 548
 549                         rc = tls_push_record(sk,
 550                                              tls_ctx,
 551                                              ctx,
 552                                              record,
 553                                              tls_push_record_flags);
 554                         if (rc < 0)
 555                                 break;
 556                 }
 557         } while (!done);
 558
 559         tls_ctx->pending_open_record_frags = more;
 560
 561         if (orig_size - size > 0)
 562                 rc = orig_size - size;
 563
 564         return rc;
 565 }
 566
 567 int tls_device_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
 568 {
 569         unsigned char record_type = TLS_RECORD_TYPE_DATA;
 570         struct tls_context *tls_ctx = tls_get_ctx(sk);
 571         int rc;
 572
 573         if (!tls_ctx->zerocopy_sendfile)
 574                 msg->msg_flags &= ~MSG_SPLICE_PAGES;
 575
 576         mutex_lock(&tls_ctx->tx_lock);
 577         lock_sock(sk);
 578
 579         if (unlikely(msg->msg_controllen)) {
 580                 rc = tls_process_cmsg(sk, msg, &record_type);
 581                 if (rc)
 582                         goto out;
 583         }
 584
 585         rc = tls_push_data(sk, &msg->msg_iter, size, msg->msg_flags,
 586                            record_type);
 587
 588 out:
 589         release_sock(sk);
 590         mutex_unlock(&tls_ctx->tx_lock);
 591         return rc;
 592 }
 593
 594 void tls_device_splice_eof(struct socket *sock)
 595 {
 596         struct sock *sk = sock->sk;
 597         struct tls_context *tls_ctx = tls_get_ctx(sk);
 598         struct iov_iter iter = {};
 599
 600         if (!tls_is_partially_sent_record(tls_ctx))
 601                 return;
 602
 603         mutex_lock(&tls_ctx->tx_lock);
 604         lock_sock(sk);
 605
 606         if (tls_is_partially_sent_record(tls_ctx)) {
 607                 iov_iter_bvec(&iter, ITER_SOURCE, NULL, 0, 0);
 608                 tls_push_data(sk, &iter, 0, 0, TLS_RECORD_TYPE_DATA);
 609         }
 610
 611         release_sock(sk);
 612         mutex_unlock(&tls_ctx->tx_lock);
 613 }
 614
 615 struct tls_record_info *tls_get_record(struct tls_offload_context_tx *context,
 616                                        u32 seq, u64 *p_record_sn)
 617 {
 618         u64 record_sn = context->hint_record_sn;
 619         struct tls_record_info *info, *last;
 620
 621         info = context->retransmit_hint;
 622         if (!info ||
 623             before(seq, info->end_seq - info->len)) {
 624                 /* if retransmit_hint is irrelevant start
 625                  * from the beginning of the list
 626                  */
 627                 info = list_first_entry_or_null(&context->records_list,
 628                                                 struct tls_record_info, list);
 629                 if (!info)
 630                         return NULL;
 631                 /* send the start_marker record if seq number is before the
 632                  * tls offload start marker sequence number. This record is
 633                  * required to handle TCP packets which are before TLS offload
 634                  * started.
 635                  *  And if it's not start marker, look if this seq number
 636                  * belongs to the list.
 637                  */
 638                 if (likely(!tls_record_is_start_marker(info))) {
 639                         /* we have the first record, get the last record to see
 640                          * if this seq number belongs to the list.
 641                          */
 642                         last = list_last_entry(&context->records_list,
 643                                                struct tls_record_info, list);
 644
 645                         if (!between(seq, tls_record_start_seq(info),
 646                                      last->end_seq))
 647                                 return NULL;
 648                 }
 649                 record_sn = context->unacked_record_sn;
 650         }
 651
 652         /* We just need the _rcu for the READ_ONCE() */
 653         rcu_read_lock();
 654         list_for_each_entry_from_rcu(info, &context->records_list, list) {
 655                 if (before(seq, info->end_seq)) {
 656                         if (!context->retransmit_hint ||
 657                             after(info->end_seq,
 658                                   context->retransmit_hint->end_seq)) {
 659                                 context->hint_record_sn = record_sn;
 660                                 context->retransmit_hint = info;
 661                         }
 662                         *p_record_sn = record_sn;
 663                         goto exit_rcu_unlock;
 664                 }
 665                 record_sn++;
 666         }
 667         info = NULL;
 668
 669 exit_rcu_unlock:
 670         rcu_read_unlock();
 671         return info;
 672 }
 673 EXPORT_SYMBOL(tls_get_record);
 674
 675 static int tls_device_push_pending_record(struct sock *sk, int flags)
 676 {
 677         struct iov_iter iter;
 678
 679         iov_iter_kvec(&iter, ITER_SOURCE, NULL, 0, 0);
 680         return tls_push_data(sk, &iter, 0, flags, TLS_RECORD_TYPE_DATA);
 681 }
 682
 683 void tls_device_write_space(struct sock *sk, struct tls_context *ctx)
 684 {
 685         if (tls_is_partially_sent_record(ctx)) {
 686                 gfp_t sk_allocation = sk->sk_allocation;
 687
 688                 WARN_ON_ONCE(sk->sk_write_pending);
 689
 690                 sk->sk_allocation = GFP_ATOMIC;
 691                 tls_push_partial_record(sk, ctx,
 692                                         MSG_DONTWAIT | MSG_NOSIGNAL |
 693                                         MSG_SENDPAGE_DECRYPTED);
 694                 sk->sk_allocation = sk_allocation;
 695         }
 696 }
 697
 698 static void tls_device_resync_rx(struct tls_context *tls_ctx,
 699                                  struct sock *sk, u32 seq, u8 *rcd_sn)
 700 {
 701         struct tls_offload_context_rx *rx_ctx = tls_offload_ctx_rx(tls_ctx);
 702         struct net_device *netdev;
 703
 704         trace_tls_device_rx_resync_send(sk, seq, rcd_sn, rx_ctx->resync_type);
 705         rcu_read_lock();
 706         netdev = rcu_dereference(tls_ctx->netdev);
 707         if (netdev)
 708                 netdev->tlsdev_ops->tls_dev_resync(netdev, sk, seq, rcd_sn,
 709                                                    TLS_OFFLOAD_CTX_DIR_RX);
 710         rcu_read_unlock();
 711         TLS_INC_STATS(sock_net(sk), LINUX_MIB_TLSRXDEVICERESYNC);
 712 }
 713
 714 static bool
 715 tls_device_rx_resync_async(struct tls_offload_resync_async *resync_async,
 716                            s64 resync_req, u32 *seq, u16 *rcd_delta)
 717 {
 718         u32 is_async = resync_req & RESYNC_REQ_ASYNC;
 719         u32 req_seq = resync_req >> 32;
 720         u32 req_end = req_seq + ((resync_req >> 16) & 0xffff);
 721         u16 i;
 722
 723         *rcd_delta = 0;
 724
 725         if (is_async) {
 726                 /* shouldn't get to wraparound:
 727                  * too long in async stage, something bad happened
 728                  */
 729                 if (WARN_ON_ONCE(resync_async->rcd_delta == USHRT_MAX))
 730                         return false;
 731
 732                 /* asynchronous stage: log all headers seq such that
 733                  * req_seq <= seq <= end_seq, and wait for real resync request
 734                  */
 735                 if (before(*seq, req_seq))
 736                         return false;
 737                 if (!after(*seq, req_end) &&
 738                     resync_async->loglen < TLS_DEVICE_RESYNC_ASYNC_LOGMAX)
 739                         resync_async->log[resync_async->loglen++] = *seq;
 740
 741                 resync_async->rcd_delta++;
 742
 743                 return false;
 744         }
 745
 746         /* synchronous stage: check against the logged entries and
 747          * proceed to check the next entries if no match was found
 748          */
 749         for (i = 0; i < resync_async->loglen; i++)
 750                 if (req_seq == resync_async->log[i] &&
 751                     atomic64_try_cmpxchg(&resync_async->req, &resync_req, 0)) {
 752                         *rcd_delta = resync_async->rcd_delta - i;
 753                         *seq = req_seq;
 754                         resync_async->loglen = 0;
 755                         resync_async->rcd_delta = 0;
 756                         return true;
 757                 }
 758
 759         resync_async->loglen = 0;
 760         resync_async->rcd_delta = 0;
 761
 762         if (req_seq == *seq &&
 763             atomic64_try_cmpxchg(&resync_async->req,
 764                                  &resync_req, 0))
 765                 return true;
 766
 767         return false;
 768 }
 769
 770 void tls_device_rx_resync_new_rec(struct sock *sk, u32 rcd_len, u32 seq)
 771 {
 772         struct tls_context *tls_ctx = tls_get_ctx(sk);
 773         struct tls_offload_context_rx *rx_ctx;
 774         u8 rcd_sn[TLS_MAX_REC_SEQ_SIZE];
 775         u32 sock_data, is_req_pending;
 776         struct tls_prot_info *prot;
 777         s64 resync_req;
 778         u16 rcd_delta;
 779         u32 req_seq;
 780
 781         if (tls_ctx->rx_conf != TLS_HW)
 782                 return;
 783         if (unlikely(test_bit(TLS_RX_DEV_DEGRADED, &tls_ctx->flags)))
 784                 return;
 785
 786         prot = &tls_ctx->prot_info;
 787         rx_ctx = tls_offload_ctx_rx(tls_ctx);
 788         memcpy(rcd_sn, tls_ctx->rx.rec_seq, prot->rec_seq_size);
 789
 790         switch (rx_ctx->resync_type) {
 791         case TLS_OFFLOAD_SYNC_TYPE_DRIVER_REQ:
 792                 resync_req = atomic64_read(&rx_ctx->resync_req);
 793                 req_seq = resync_req >> 32;
 794                 seq += TLS_HEADER_SIZE - 1;
 795                 is_req_pending = resync_req;
 796
 797                 if (likely(!is_req_pending) || req_seq != seq ||
 798                     !atomic64_try_cmpxchg(&rx_ctx->resync_req, &resync_req, 0))
 799                         return;
 800                 break;
 801         case TLS_OFFLOAD_SYNC_TYPE_CORE_NEXT_HINT:
 802                 if (likely(!rx_ctx->resync_nh_do_now))
 803                         return;
 804
 805                 /* head of next rec is already in, note that the sock_inq will
 806                  * include the currently parsed message when called from parser
 807                  */
 808                 sock_data = tcp_inq(sk);
 809                 if (sock_data > rcd_len) {
 810                         trace_tls_device_rx_resync_nh_delay(sk, sock_data,
 811                                                             rcd_len);
 812                         return;
 813                 }
 814
 815                 rx_ctx->resync_nh_do_now = 0;
 816                 seq += rcd_len;
 817                 tls_bigint_increment(rcd_sn, prot->rec_seq_size);
 818                 break;
 819         case TLS_OFFLOAD_SYNC_TYPE_DRIVER_REQ_ASYNC:
 820                 resync_req = atomic64_read(&rx_ctx->resync_async->req);
 821                 is_req_pending = resync_req;
 822                 if (likely(!is_req_pending))
 823                         return;
 824
 825                 if (!tls_device_rx_resync_async(rx_ctx->resync_async,
 826                                                 resync_req, &seq, &rcd_delta))
 827                         return;
 828                 tls_bigint_subtract(rcd_sn, rcd_delta);
 829                 break;
 830         }
 831
 832         tls_device_resync_rx(tls_ctx, sk, seq, rcd_sn);
 833 }
 834
 835 static void tls_device_core_ctrl_rx_resync(struct tls_context *tls_ctx,
 836                                            struct tls_offload_context_rx *ctx,
 837                                            struct sock *sk, struct sk_buff *skb)
 838 {
 839         struct strp_msg *rxm;
 840
 841         /* device will request resyncs by itself based on stream scan */
 842         if (ctx->resync_type != TLS_OFFLOAD_SYNC_TYPE_CORE_NEXT_HINT)
 843                 return;
 844         /* already scheduled */
 845         if (ctx->resync_nh_do_now)
 846                 return;
 847         /* seen decrypted fragments since last fully-failed record */
 848         if (ctx->resync_nh_reset) {
 849                 ctx->resync_nh_reset = 0;
 850                 ctx->resync_nh.decrypted_failed = 1;
 851                 ctx->resync_nh.decrypted_tgt = TLS_DEVICE_RESYNC_NH_START_IVAL;
 852                 return;
 853         }
 854
 855         if (++ctx->resync_nh.decrypted_failed <= ctx->resync_nh.decrypted_tgt)
 856                 return;
 857
 858         /* doing resync, bump the next target in case it fails */
 859         if (ctx->resync_nh.decrypted_tgt < TLS_DEVICE_RESYNC_NH_MAX_IVAL)
 860                 ctx->resync_nh.decrypted_tgt *= 2;
 861         else
 862                 ctx->resync_nh.decrypted_tgt += TLS_DEVICE_RESYNC_NH_MAX_IVAL;
 863
 864         rxm = strp_msg(skb);
 865
 866         /* head of next rec is already in, parser will sync for us */
 867         if (tcp_inq(sk) > rxm->full_len) {
 868                 trace_tls_device_rx_resync_nh_schedule(sk);
 869                 ctx->resync_nh_do_now = 1;
 870         } else {
 871                 struct tls_prot_info *prot = &tls_ctx->prot_info;
 872                 u8 rcd_sn[TLS_MAX_REC_SEQ_SIZE];
 873
 874                 memcpy(rcd_sn, tls_ctx->rx.rec_seq, prot->rec_seq_size);
 875                 tls_bigint_increment(rcd_sn, prot->rec_seq_size);
 876
 877                 tls_device_resync_rx(tls_ctx, sk, tcp_sk(sk)->copied_seq,
 878                                      rcd_sn);
 879         }
 880 }
 881
 882 static int
 883 tls_device_reencrypt(struct sock *sk, struct tls_context *tls_ctx)
 884 {
 885         struct tls_sw_context_rx *sw_ctx = tls_sw_ctx_rx(tls_ctx);
 886         const struct tls_cipher_desc *cipher_desc;
 887         int err, offset, copy, data_len, pos;
 888         struct sk_buff *skb, *skb_iter;
 889         struct scatterlist sg[1];
 890         struct strp_msg *rxm;
 891         char *orig_buf, *buf;
 892
 893         cipher_desc = get_cipher_desc(tls_ctx->crypto_recv.info.cipher_type);
 894         DEBUG_NET_WARN_ON_ONCE(!cipher_desc || !cipher_desc->offloadable);
 895
 896         rxm = strp_msg(tls_strp_msg(sw_ctx));
 897         orig_buf = kmalloc(rxm->full_len + TLS_HEADER_SIZE + cipher_desc->iv,
 898                            sk->sk_allocation);
 899         if (!orig_buf)
 900                 return -ENOMEM;
 901         buf = orig_buf;
 902
 903         err = tls_strp_msg_cow(sw_ctx);
 904         if (unlikely(err))
 905                 goto free_buf;
 906
 907         skb = tls_strp_msg(sw_ctx);
 908         rxm = strp_msg(skb);
 909         offset = rxm->offset;
 910
 911         sg_init_table(sg, 1);
 912         sg_set_buf(&sg[0], buf,
 913                    rxm->full_len + TLS_HEADER_SIZE + cipher_desc->iv);
 914         err = skb_copy_bits(skb, offset, buf, TLS_HEADER_SIZE + cipher_desc->iv);
 915         if (err)
 916                 goto free_buf;
 917
 918         /* We are interested only in the decrypted data not the auth */
 919         err = decrypt_skb(sk, sg);
 920         if (err != -EBADMSG)
 921                 goto free_buf;
 922         else
 923                 err = 0;
 924
 925         data_len = rxm->full_len - cipher_desc->tag;
 926
 927         if (skb_pagelen(skb) > offset) {
 928                 copy = min_t(int, skb_pagelen(skb) - offset, data_len);
 929
 930                 if (skb->decrypted) {
 931                         err = skb_store_bits(skb, offset, buf, copy);
 932                         if (err)
 933                                 goto free_buf;
 934                 }
 935
 936                 offset += copy;
 937                 buf += copy;
 938         }
 939
 940         pos = skb_pagelen(skb);
 941         skb_walk_frags(skb, skb_iter) {
 942                 int frag_pos;
 943
 944                 /* Practically all frags must belong to msg if reencrypt
 945                  * is needed with current strparser and coalescing logic,
 946                  * but strparser may "get optimized", so let's be safe.
 947                  */
 948                 if (pos + skb_iter->len <= offset)
 949                         goto done_with_frag;
 950                 if (pos >= data_len + rxm->offset)
 951                         break;
 952
 953                 frag_pos = offset - pos;
 954                 copy = min_t(int, skb_iter->len - frag_pos,
 955                              data_len + rxm->offset - offset);
 956
 957                 if (skb_iter->decrypted) {
 958                         err = skb_store_bits(skb_iter, frag_pos, buf, copy);
 959                         if (err)
 960                                 goto free_buf;
 961                 }
 962
 963                 offset += copy;
 964                 buf += copy;
 965 done_with_frag:
 966                 pos += skb_iter->len;
 967         }
 968
 969 free_buf:
 970         kfree(orig_buf);
 971         return err;
 972 }
 973
 974 int tls_device_decrypted(struct sock *sk, struct tls_context *tls_ctx)
 975 {
 976         struct tls_offload_context_rx *ctx = tls_offload_ctx_rx(tls_ctx);
 977         struct tls_sw_context_rx *sw_ctx = tls_sw_ctx_rx(tls_ctx);
 978         struct sk_buff *skb = tls_strp_msg(sw_ctx);
 979         struct strp_msg *rxm = strp_msg(skb);
 980         int is_decrypted, is_encrypted;
 981
 982         if (!tls_strp_msg_mixed_decrypted(sw_ctx)) {
 983                 is_decrypted = skb->decrypted;
 984                 is_encrypted = !is_decrypted;
 985         } else {
 986                 is_decrypted = 0;
 987                 is_encrypted = 0;
 988         }
 989
 990         trace_tls_device_decrypted(sk, tcp_sk(sk)->copied_seq - rxm->full_len,
 991                                    tls_ctx->rx.rec_seq, rxm->full_len,
 992                                    is_encrypted, is_decrypted);
 993
 994         if (unlikely(test_bit(TLS_RX_DEV_DEGRADED, &tls_ctx->flags))) {
 995                 if (likely(is_encrypted || is_decrypted))
 996                         return is_decrypted;
 997
 998                 /* After tls_device_down disables the offload, the next SKB will
 999                  * likely have initial fragments decrypted, and final ones not
1000                  * decrypted. We need to reencrypt that single SKB.
1001                  */
1002                 return tls_device_reencrypt(sk, tls_ctx);
1003         }
1004
1005         /* Return immediately if the record is either entirely plaintext or
1006          * entirely ciphertext. Otherwise handle reencrypt partially decrypted
1007          * record.
1008          */
1009         if (is_decrypted) {
1010                 ctx->resync_nh_reset = 1;
1011                 return is_decrypted;
1012         }
1013         if (is_encrypted) {
1014                 tls_device_core_ctrl_rx_resync(tls_ctx, ctx, sk, skb);
1015                 return 0;
1016         }
1017
1018         ctx->resync_nh_reset = 1;
1019         return tls_device_reencrypt(sk, tls_ctx);
1020 }
1021
1022 static void tls_device_attach(struct tls_context *ctx, struct sock *sk,
1023                               struct net_device *netdev)
1024 {
1025         if (sk->sk_destruct != tls_device_sk_destruct) {
1026                 refcount_set(&ctx->refcount, 1);
1027                 dev_hold(netdev);
1028                 RCU_INIT_POINTER(ctx->netdev, netdev);
1029                 spin_lock_irq(&tls_device_lock);
1030                 list_add_tail(&ctx->list, &tls_device_list);
1031                 spin_unlock_irq(&tls_device_lock);
1032
1033                 ctx->sk_destruct = sk->sk_destruct;
1034                 smp_store_release(&sk->sk_destruct, tls_device_sk_destruct);
1035         }
1036 }
1037
1038 int tls_set_device_offload(struct sock *sk, struct tls_context *ctx)
1039 {
1040         struct tls_context *tls_ctx = tls_get_ctx(sk);
1041         struct tls_prot_info *prot = &tls_ctx->prot_info;
1042         const struct tls_cipher_desc *cipher_desc;
1043         struct tls_record_info *start_marker_record;
1044         struct tls_offload_context_tx *offload_ctx;
1045         struct tls_crypto_info *crypto_info;
1046         struct net_device *netdev;
1047         char *iv, *rec_seq;
1048         struct sk_buff *skb;
1049         __be64 rcd_sn;
1050         int rc;
1051
1052         if (!ctx)
1053                 return -EINVAL;
1054
1055         if (ctx->priv_ctx_tx)
1056                 return -EEXIST;
1057
1058         netdev = get_netdev_for_sock(sk);
1059         if (!netdev) {
1060                 pr_err_ratelimited("%s: netdev not found\n", __func__);
1061                 return -EINVAL;
1062         }
1063
1064         if (!(netdev->features & NETIF_F_HW_TLS_TX)) {
1065                 rc = -EOPNOTSUPP;
1066                 goto release_netdev;
1067         }
1068
1069         crypto_info = &ctx->crypto_send.info;
1070         if (crypto_info->version != TLS_1_2_VERSION) {
1071                 rc = -EOPNOTSUPP;
1072                 goto release_netdev;
1073         }
1074
1075         cipher_desc = get_cipher_desc(crypto_info->cipher_type);
1076         if (!cipher_desc || !cipher_desc->offloadable) {
1077                 rc = -EINVAL;
1078                 goto release_netdev;
1079         }
1080
1081         iv = crypto_info_iv(crypto_info, cipher_desc);
1082         rec_seq = crypto_info_rec_seq(crypto_info, cipher_desc);
1083
1084         prot->version = crypto_info->version;
1085         prot->cipher_type = crypto_info->cipher_type;
1086         prot->prepend_size = TLS_HEADER_SIZE + cipher_desc->iv;
1087         prot->tag_size = cipher_desc->tag;
1088         prot->overhead_size = prot->prepend_size + prot->tag_size;
1089         prot->iv_size = cipher_desc->iv;
1090         prot->salt_size = cipher_desc->salt;
1091         ctx->tx.iv = kmalloc(cipher_desc->iv + cipher_desc->salt, GFP_KERNEL);
1092         if (!ctx->tx.iv) {
1093                 rc = -ENOMEM;
1094                 goto release_netdev;
1095         }
1096
1097         memcpy(ctx->tx.iv + cipher_desc->salt, iv, cipher_desc->iv);
1098
1099         prot->rec_seq_size = cipher_desc->rec_seq;
1100         memcpy(ctx->tx.rec_seq, rec_seq, cipher_desc->rec_seq);
1101
1102         start_marker_record = kmalloc(sizeof(*start_marker_record), GFP_KERNEL);
1103         if (!start_marker_record) {
1104                 rc = -ENOMEM;
1105                 goto free_iv;
1106         }
1107
1108         offload_ctx = kzalloc(TLS_OFFLOAD_CONTEXT_SIZE_TX, GFP_KERNEL);
1109         if (!offload_ctx) {
1110                 rc = -ENOMEM;
1111                 goto free_marker_record;
1112         }
1113
1114         rc = tls_sw_fallback_init(sk, offload_ctx, crypto_info);
1115         if (rc)
1116                 goto free_offload_ctx;
1117
1118         /* start at rec_seq - 1 to account for the start marker record */
1119         memcpy(&rcd_sn, ctx->tx.rec_seq, sizeof(rcd_sn));
1120         offload_ctx->unacked_record_sn = be64_to_cpu(rcd_sn) - 1;
1121
1122         start_marker_record->end_seq = tcp_sk(sk)->write_seq;
1123         start_marker_record->len = 0;
1124         start_marker_record->num_frags = 0;
1125
1126         INIT_WORK(&offload_ctx->destruct_work, tls_device_tx_del_task);
1127         offload_ctx->ctx = ctx;
1128
1129         INIT_LIST_HEAD(&offload_ctx->records_list);
1130         list_add_tail(&start_marker_record->list, &offload_ctx->records_list);
1131         spin_lock_init(&offload_ctx->lock);
1132         sg_init_table(offload_ctx->sg_tx_data,
1133                       ARRAY_SIZE(offload_ctx->sg_tx_data));
1134
1135         clean_acked_data_enable(inet_csk(sk), &tls_icsk_clean_acked);
1136         ctx->push_pending_record = tls_device_push_pending_record;
1137
1138         /* TLS offload is greatly simplified if we don't send
1139          * SKBs where only part of the payload needs to be encrypted.
1140          * So mark the last skb in the write queue as end of record.
1141          */
1142         skb = tcp_write_queue_tail(sk);
1143         if (skb)
1144                 TCP_SKB_CB(skb)->eor = 1;
1145
1146         /* Avoid offloading if the device is down
1147          * We don't want to offload new flows after
1148          * the NETDEV_DOWN event
1149          *
1150          * device_offload_lock is taken in tls_devices's NETDEV_DOWN
1151          * handler thus protecting from the device going down before
1152          * ctx was added to tls_device_list.
1153          */
1154         down_read(&device_offload_lock);
1155         if (!(netdev->flags & IFF_UP)) {
1156                 rc = -EINVAL;
1157                 goto release_lock;
1158         }
1159
1160         ctx->priv_ctx_tx = offload_ctx;
1161         rc = netdev->tlsdev_ops->tls_dev_add(netdev, sk, TLS_OFFLOAD_CTX_DIR_TX,
1162                                              &ctx->crypto_send.info,
1163                                              tcp_sk(sk)->write_seq);
1164         trace_tls_device_offload_set(sk, TLS_OFFLOAD_CTX_DIR_TX,
1165                                      tcp_sk(sk)->write_seq, rec_seq, rc);
1166         if (rc)
1167                 goto release_lock;
1168
1169         tls_device_attach(ctx, sk, netdev);
1170         up_read(&device_offload_lock);
1171
1172         /* following this assignment tls_is_skb_tx_device_offloaded
1173          * will return true and the context might be accessed
1174          * by the netdev's xmit function.
1175          */
1176         smp_store_release(&sk->sk_validate_xmit_skb, tls_validate_xmit_skb);
1177         dev_put(netdev);
1178
1179         return 0;
1180
1181 release_lock:
1182         up_read(&device_offload_lock);
1183         clean_acked_data_disable(inet_csk(sk));
1184         crypto_free_aead(offload_ctx->aead_send);
1185 free_offload_ctx:
1186         kfree(offload_ctx);
1187         ctx->priv_ctx_tx = NULL;
1188 free_marker_record:
1189         kfree(start_marker_record);
1190 free_iv:
1191         kfree(ctx->tx.iv);
1192 release_netdev:
1193         dev_put(netdev);
1194         return rc;
1195 }
1196
1197 int tls_set_device_offload_rx(struct sock *sk, struct tls_context *ctx)
1198 {
1199         struct tls12_crypto_info_aes_gcm_128 *info;
1200         struct tls_offload_context_rx *context;
1201         struct net_device *netdev;
1202         int rc = 0;
1203
1204         if (ctx->crypto_recv.info.version != TLS_1_2_VERSION)
1205                 return -EOPNOTSUPP;
1206
1207         netdev = get_netdev_for_sock(sk);
1208         if (!netdev) {
1209                 pr_err_ratelimited("%s: netdev not found\n", __func__);
1210                 return -EINVAL;
1211         }
1212
1213         if (!(netdev->features & NETIF_F_HW_TLS_RX)) {
1214                 rc = -EOPNOTSUPP;
1215                 goto release_netdev;
1216         }
1217
1218         /* Avoid offloading if the device is down
1219          * We don't want to offload new flows after
1220          * the NETDEV_DOWN event
1221          *
1222          * device_offload_lock is taken in tls_devices's NETDEV_DOWN
1223          * handler thus protecting from the device going down before
1224          * ctx was added to tls_device_list.
1225          */
1226         down_read(&device_offload_lock);
1227         if (!(netdev->flags & IFF_UP)) {
1228                 rc = -EINVAL;
1229                 goto release_lock;
1230         }
1231
1232         context = kzalloc(TLS_OFFLOAD_CONTEXT_SIZE_RX, GFP_KERNEL);
1233         if (!context) {
1234                 rc = -ENOMEM;
1235                 goto release_lock;
1236         }
1237         context->resync_nh_reset = 1;
1238
1239         ctx->priv_ctx_rx = context;
1240         rc = tls_set_sw_offload(sk, ctx, 0);
1241         if (rc)
1242                 goto release_ctx;
1243
1244         rc = netdev->tlsdev_ops->tls_dev_add(netdev, sk, TLS_OFFLOAD_CTX_DIR_RX,
1245                                              &ctx->crypto_recv.info,
1246                                              tcp_sk(sk)->copied_seq);
1247         info = (void *)&ctx->crypto_recv.info;
1248         trace_tls_device_offload_set(sk, TLS_OFFLOAD_CTX_DIR_RX,
1249                                      tcp_sk(sk)->copied_seq, info->rec_seq, rc);
1250         if (rc)
1251                 goto free_sw_resources;
1252
1253         tls_device_attach(ctx, sk, netdev);
1254         up_read(&device_offload_lock);
1255
1256         dev_put(netdev);
1257
1258         return 0;
1259
1260 free_sw_resources:
1261         up_read(&device_offload_lock);
1262         tls_sw_free_resources_rx(sk);
1263         down_read(&device_offload_lock);
1264 release_ctx:
1265         ctx->priv_ctx_rx = NULL;
1266 release_lock:
1267         up_read(&device_offload_lock);
1268 release_netdev:
1269         dev_put(netdev);
1270         return rc;
1271 }
1272
1273 void tls_device_offload_cleanup_rx(struct sock *sk)
1274 {
1275         struct tls_context *tls_ctx = tls_get_ctx(sk);
1276         struct net_device *netdev;
1277
1278         down_read(&device_offload_lock);
1279         netdev = rcu_dereference_protected(tls_ctx->netdev,
1280                                            lockdep_is_held(&device_offload_lock));
1281         if (!netdev)
1282                 goto out;
1283
1284         netdev->tlsdev_ops->tls_dev_del(netdev, tls_ctx,
1285                                         TLS_OFFLOAD_CTX_DIR_RX);
1286
1287         if (tls_ctx->tx_conf != TLS_HW) {
1288                 dev_put(netdev);
1289                 rcu_assign_pointer(tls_ctx->netdev, NULL);
1290         } else {
1291                 set_bit(TLS_RX_DEV_CLOSED, &tls_ctx->flags);
1292         }
1293 out:
1294         up_read(&device_offload_lock);
1295         tls_sw_release_resources_rx(sk);
1296 }
1297
1298 static int tls_device_down(struct net_device *netdev)
1299 {
1300         struct tls_context *ctx, *tmp;
1301         unsigned long flags;
1302         LIST_HEAD(list);
1303
1304         /* Request a write lock to block new offload attempts */
1305         down_write(&device_offload_lock);
1306
1307         spin_lock_irqsave(&tls_device_lock, flags);
1308         list_for_each_entry_safe(ctx, tmp, &tls_device_list, list) {
1309                 struct net_device *ctx_netdev =
1310                         rcu_dereference_protected(ctx->netdev,
1311                                                   lockdep_is_held(&device_offload_lock));
1312
1313                 if (ctx_netdev != netdev ||
1314                     !refcount_inc_not_zero(&ctx->refcount))
1315                         continue;
1316
1317                 list_move(&ctx->list, &list);
1318         }
1319         spin_unlock_irqrestore(&tls_device_lock, flags);
1320
1321         list_for_each_entry_safe(ctx, tmp, &list, list) {
1322                 /* Stop offloaded TX and switch to the fallback.
1323                  * tls_is_skb_tx_device_offloaded will return false.
1324                  */
1325                 WRITE_ONCE(ctx->sk->sk_validate_xmit_skb, tls_validate_xmit_skb_sw);
1326
1327                 /* Stop the RX and TX resync.
1328                  * tls_dev_resync must not be called after tls_dev_del.
1329                  */
1330                 rcu_assign_pointer(ctx->netdev, NULL);
1331
1332                 /* Start skipping the RX resync logic completely. */
1333                 set_bit(TLS_RX_DEV_DEGRADED, &ctx->flags);
1334
1335                 /* Sync with inflight packets. After this point:
1336                  * TX: no non-encrypted packets will be passed to the driver.
1337                  * RX: resync requests from the driver will be ignored.
1338                  */
1339                 synchronize_net();
1340
1341                 /* Release the offload context on the driver side. */
1342                 if (ctx->tx_conf == TLS_HW)
1343                         netdev->tlsdev_ops->tls_dev_del(netdev, ctx,
1344                                                         TLS_OFFLOAD_CTX_DIR_TX);
1345                 if (ctx->rx_conf == TLS_HW &&
1346                     !test_bit(TLS_RX_DEV_CLOSED, &ctx->flags))
1347                         netdev->tlsdev_ops->tls_dev_del(netdev, ctx,
1348                                                         TLS_OFFLOAD_CTX_DIR_RX);
1349
1350                 dev_put(netdev);
1351
1352                 /* Move the context to a separate list for two reasons:
1353                  * 1. When the context is deallocated, list_del is called.
1354                  * 2. It's no longer an offloaded context, so we don't want to
1355                  *    run offload-specific code on this context.
1356                  */
1357                 spin_lock_irqsave(&tls_device_lock, flags);
1358                 list_move_tail(&ctx->list, &tls_device_down_list);
1359                 spin_unlock_irqrestore(&tls_device_lock, flags);
1360
1361                 /* Device contexts for RX and TX will be freed in on sk_destruct
1362                  * by tls_device_free_ctx. rx_conf and tx_conf stay in TLS_HW.
1363                  * Now release the ref taken above.
1364                  */
1365                 if (refcount_dec_and_test(&ctx->refcount)) {
1366                         /* sk_destruct ran after tls_device_down took a ref, and
1367                          * it returned early. Complete the destruction here.
1368                          */
1369                         list_del(&ctx->list);
1370                         tls_device_free_ctx(ctx);
1371                 }
1372         }
1373
1374         up_write(&device_offload_lock);
1375
1376         flush_workqueue(destruct_wq);
1377
1378         return NOTIFY_DONE;
1379 }
1380
1381 static int tls_dev_event(struct notifier_block *this, unsigned long event,
1382                          void *ptr)
1383 {
1384         struct net_device *dev = netdev_notifier_info_to_dev(ptr);
1385
1386         if (!dev->tlsdev_ops &&
1387             !(dev->features & (NETIF_F_HW_TLS_RX | NETIF_F_HW_TLS_TX)))
1388                 return NOTIFY_DONE;
1389
1390         switch (event) {
1391         case NETDEV_REGISTER:
1392         case NETDEV_FEAT_CHANGE:
1393                 if (netif_is_bond_master(dev))
1394                         return NOTIFY_DONE;
1395                 if ((dev->features & NETIF_F_HW_TLS_RX) &&
1396                     !dev->tlsdev_ops->tls_dev_resync)
1397                         return NOTIFY_BAD;
1398
1399                 if  (dev->tlsdev_ops &&
1400                      dev->tlsdev_ops->tls_dev_add &&
1401                      dev->tlsdev_ops->tls_dev_del)
1402                         return NOTIFY_DONE;
1403                 else
1404                         return NOTIFY_BAD;
1405         case NETDEV_DOWN:
1406                 return tls_device_down(dev);
1407         }
1408         return NOTIFY_DONE;
1409 }
1410
1411 static struct notifier_block tls_dev_notifier = {
1412         .notifier_call  = tls_dev_event,
1413 };
1414
1415 int __init tls_device_init(void)
1416 {
1417         int err;
1418
1419         dummy_page = alloc_page(GFP_KERNEL);
1420         if (!dummy_page)
1421                 return -ENOMEM;
1422
1423         destruct_wq = alloc_workqueue("ktls_device_destruct", 0, 0);
1424         if (!destruct_wq) {
1425                 err = -ENOMEM;
1426                 goto err_free_dummy;
1427         }
1428
1429         err = register_netdevice_notifier(&tls_dev_notifier);
1430         if (err)
1431                 goto err_destroy_wq;
1432
1433         return 0;
1434
1435 err_destroy_wq:
1436         destroy_workqueue(destruct_wq);
1437 err_free_dummy:
1438         put_page(dummy_page);
1439         return err;
1440 }
1441
1442 void __exit tls_device_cleanup(void)
1443 {
1444         unregister_netdevice_notifier(&tls_dev_notifier);
1445         destroy_workqueue(destruct_wq);
1446         clean_acked_data_flush();
1447         put_page(dummy_page);
1448 }