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 void tls_device_gc_task(struct work_struct *work);
  50
  51 static DECLARE_WORK(tls_device_gc_work, tls_device_gc_task);
  52 static LIST_HEAD(tls_device_gc_list);
  53 static LIST_HEAD(tls_device_list);
  54 static LIST_HEAD(tls_device_down_list);
  55 static DEFINE_SPINLOCK(tls_device_lock);
  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.rec_seq);
  62                 kfree(ctx->tx.iv);
  63         }
  64
  65         if (ctx->rx_conf == TLS_HW)
  66                 kfree(tls_offload_ctx_rx(ctx));
  67
  68         tls_ctx_free(NULL, ctx);
  69 }
  70
  71 static void tls_device_gc_task(struct work_struct *work)
  72 {
  73         struct tls_context *ctx, *tmp;
  74         unsigned long flags;
  75         LIST_HEAD(gc_list);
  76
  77         spin_lock_irqsave(&tls_device_lock, flags);
  78         list_splice_init(&tls_device_gc_list, &gc_list);
  79         spin_unlock_irqrestore(&tls_device_lock, flags);
  80
  81         list_for_each_entry_safe(ctx, tmp, &gc_list, list) {
  82                 struct net_device *netdev = ctx->netdev;
  83
  84                 if (netdev && ctx->tx_conf == TLS_HW) {
  85                         netdev->tlsdev_ops->tls_dev_del(netdev, ctx,
  86                                                         TLS_OFFLOAD_CTX_DIR_TX);
  87                         dev_put(netdev);
  88                         ctx->netdev = NULL;
  89                 }
  90
  91                 list_del(&ctx->list);
  92                 tls_device_free_ctx(ctx);
  93         }
  94 }
  95
  96 static void tls_device_queue_ctx_destruction(struct tls_context *ctx)
  97 {
  98         unsigned long flags;
  99
 100         spin_lock_irqsave(&tls_device_lock, flags);
 101         list_move_tail(&ctx->list, &tls_device_gc_list);
 102
 103         /* schedule_work inside the spinlock
 104          * to make sure tls_device_down waits for that work.
 105          */
 106         schedule_work(&tls_device_gc_work);
 107
 108         spin_unlock_irqrestore(&tls_device_lock, flags);
 109 }
 110
 111 /* We assume that the socket is already connected */
 112 static struct net_device *get_netdev_for_sock(struct sock *sk)
 113 {
 114         struct dst_entry *dst = sk_dst_get(sk);
 115         struct net_device *netdev = NULL;
 116
 117         if (likely(dst)) {
 118                 netdev = netdev_sk_get_lowest_dev(dst->dev, sk);
 119                 dev_hold(netdev);
 120         }
 121
 122         dst_release(dst);
 123
 124         return netdev;
 125 }
 126
 127 static void destroy_record(struct tls_record_info *record)
 128 {
 129         int i;
 130
 131         for (i = 0; i < record->num_frags; i++)
 132                 __skb_frag_unref(&record->frags[i], false);
 133         kfree(record);
 134 }
 135
 136 static void delete_all_records(struct tls_offload_context_tx *offload_ctx)
 137 {
 138         struct tls_record_info *info, *temp;
 139
 140         list_for_each_entry_safe(info, temp, &offload_ctx->records_list, list) {
 141                 list_del(&info->list);
 142                 destroy_record(info);
 143         }
 144
 145         offload_ctx->retransmit_hint = NULL;
 146 }
 147
 148 static void tls_icsk_clean_acked(struct sock *sk, u32 acked_seq)
 149 {
 150         struct tls_context *tls_ctx = tls_get_ctx(sk);
 151         struct tls_record_info *info, *temp;
 152         struct tls_offload_context_tx *ctx;
 153         u64 deleted_records = 0;
 154         unsigned long flags;
 155
 156         if (!tls_ctx)
 157                 return;
 158
 159         ctx = tls_offload_ctx_tx(tls_ctx);
 160
 161         spin_lock_irqsave(&ctx->lock, flags);
 162         info = ctx->retransmit_hint;
 163         if (info && !before(acked_seq, info->end_seq))
 164                 ctx->retransmit_hint = NULL;
 165
 166         list_for_each_entry_safe(info, temp, &ctx->records_list, list) {
 167                 if (before(acked_seq, info->end_seq))
 168                         break;
 169                 list_del(&info->list);
 170
 171                 destroy_record(info);
 172                 deleted_records++;
 173         }
 174
 175         ctx->unacked_record_sn += deleted_records;
 176         spin_unlock_irqrestore(&ctx->lock, flags);
 177 }
 178
 179 /* At this point, there should be no references on this
 180  * socket and no in-flight SKBs associated with this
 181  * socket, so it is safe to free all the resources.
 182  */
 183 void tls_device_sk_destruct(struct sock *sk)
 184 {
 185         struct tls_context *tls_ctx = tls_get_ctx(sk);
 186         struct tls_offload_context_tx *ctx = tls_offload_ctx_tx(tls_ctx);
 187
 188         tls_ctx->sk_destruct(sk);
 189
 190         if (tls_ctx->tx_conf == TLS_HW) {
 191                 if (ctx->open_record)
 192                         destroy_record(ctx->open_record);
 193                 delete_all_records(ctx);
 194                 crypto_free_aead(ctx->aead_send);
 195                 clean_acked_data_disable(inet_csk(sk));
 196         }
 197
 198         if (refcount_dec_and_test(&tls_ctx->refcount))
 199                 tls_device_queue_ctx_destruction(tls_ctx);
 200 }
 201 EXPORT_SYMBOL_GPL(tls_device_sk_destruct);
 202
 203 void tls_device_free_resources_tx(struct sock *sk)
 204 {
 205         struct tls_context *tls_ctx = tls_get_ctx(sk);
 206
 207         tls_free_partial_record(sk, tls_ctx);
 208 }
 209
 210 void tls_offload_tx_resync_request(struct sock *sk, u32 got_seq, u32 exp_seq)
 211 {
 212         struct tls_context *tls_ctx = tls_get_ctx(sk);
 213
 214         trace_tls_device_tx_resync_req(sk, got_seq, exp_seq);
 215         WARN_ON(test_and_set_bit(TLS_TX_SYNC_SCHED, &tls_ctx->flags));
 216 }
 217 EXPORT_SYMBOL_GPL(tls_offload_tx_resync_request);
 218
 219 static void tls_device_resync_tx(struct sock *sk, struct tls_context *tls_ctx,
 220                                  u32 seq)
 221 {
 222         struct net_device *netdev;
 223         struct sk_buff *skb;
 224         int err = 0;
 225         u8 *rcd_sn;
 226
 227         skb = tcp_write_queue_tail(sk);
 228         if (skb)
 229                 TCP_SKB_CB(skb)->eor = 1;
 230
 231         rcd_sn = tls_ctx->tx.rec_seq;
 232
 233         trace_tls_device_tx_resync_send(sk, seq, rcd_sn);
 234         down_read(&device_offload_lock);
 235         netdev = tls_ctx->netdev;
 236         if (netdev)
 237                 err = netdev->tlsdev_ops->tls_dev_resync(netdev, sk, seq,
 238                                                          rcd_sn,
 239                                                          TLS_OFFLOAD_CTX_DIR_TX);
 240         up_read(&device_offload_lock);
 241         if (err)
 242                 return;
 243
 244         clear_bit_unlock(TLS_TX_SYNC_SCHED, &tls_ctx->flags);
 245 }
 246
 247 static void tls_append_frag(struct tls_record_info *record,
 248                             struct page_frag *pfrag,
 249                             int size)
 250 {
 251         skb_frag_t *frag;
 252
 253         frag = &record->frags[record->num_frags - 1];
 254         if (skb_frag_page(frag) == pfrag->page &&
 255             skb_frag_off(frag) + skb_frag_size(frag) == pfrag->offset) {
 256                 skb_frag_size_add(frag, size);
 257         } else {
 258                 ++frag;
 259                 __skb_frag_set_page(frag, pfrag->page);
 260                 skb_frag_off_set(frag, pfrag->offset);
 261                 skb_frag_size_set(frag, size);
 262                 ++record->num_frags;
 263                 get_page(pfrag->page);
 264         }
 265
 266         pfrag->offset += size;
 267         record->len += size;
 268 }
 269
 270 static int tls_push_record(struct sock *sk,
 271                            struct tls_context *ctx,
 272                            struct tls_offload_context_tx *offload_ctx,
 273                            struct tls_record_info *record,
 274                            int flags)
 275 {
 276         struct tls_prot_info *prot = &ctx->prot_info;
 277         struct tcp_sock *tp = tcp_sk(sk);
 278         skb_frag_t *frag;
 279         int i;
 280
 281         record->end_seq = tp->write_seq + record->len;
 282         list_add_tail_rcu(&record->list, &offload_ctx->records_list);
 283         offload_ctx->open_record = NULL;
 284
 285         if (test_bit(TLS_TX_SYNC_SCHED, &ctx->flags))
 286                 tls_device_resync_tx(sk, ctx, tp->write_seq);
 287
 288         tls_advance_record_sn(sk, prot, &ctx->tx);
 289
 290         for (i = 0; i < record->num_frags; i++) {
 291                 frag = &record->frags[i];
 292                 sg_unmark_end(&offload_ctx->sg_tx_data[i]);
 293                 sg_set_page(&offload_ctx->sg_tx_data[i], skb_frag_page(frag),
 294                             skb_frag_size(frag), skb_frag_off(frag));
 295                 sk_mem_charge(sk, skb_frag_size(frag));
 296                 get_page(skb_frag_page(frag));
 297         }
 298         sg_mark_end(&offload_ctx->sg_tx_data[record->num_frags - 1]);
 299
 300         /* all ready, send */
 301         return tls_push_sg(sk, ctx, offload_ctx->sg_tx_data, 0, flags);
 302 }
 303
 304 static int tls_device_record_close(struct sock *sk,
 305                                    struct tls_context *ctx,
 306                                    struct tls_record_info *record,
 307                                    struct page_frag *pfrag,
 308                                    unsigned char record_type)
 309 {
 310         struct tls_prot_info *prot = &ctx->prot_info;
 311         int ret;
 312
 313         /* append tag
 314          * device will fill in the tag, we just need to append a placeholder
 315          * use socket memory to improve coalescing (re-using a single buffer
 316          * increases frag count)
 317          * if we can't allocate memory now, steal some back from data
 318          */
 319         if (likely(skb_page_frag_refill(prot->tag_size, pfrag,
 320                                         sk->sk_allocation))) {
 321                 ret = 0;
 322                 tls_append_frag(record, pfrag, prot->tag_size);
 323         } else {
 324                 ret = prot->tag_size;
 325                 if (record->len <= prot->overhead_size)
 326                         return -ENOMEM;
 327         }
 328
 329         /* fill prepend */
 330         tls_fill_prepend(ctx, skb_frag_address(&record->frags[0]),
 331                          record->len - prot->overhead_size,
 332                          record_type);
 333         return ret;
 334 }
 335
 336 static int tls_create_new_record(struct tls_offload_context_tx *offload_ctx,
 337                                  struct page_frag *pfrag,
 338                                  size_t prepend_size)
 339 {
 340         struct tls_record_info *record;
 341         skb_frag_t *frag;
 342
 343         record = kmalloc(sizeof(*record), GFP_KERNEL);
 344         if (!record)
 345                 return -ENOMEM;
 346
 347         frag = &record->frags[0];
 348         __skb_frag_set_page(frag, pfrag->page);
 349         skb_frag_off_set(frag, pfrag->offset);
 350         skb_frag_size_set(frag, prepend_size);
 351
 352         get_page(pfrag->page);
 353         pfrag->offset += prepend_size;
 354
 355         record->num_frags = 1;
 356         record->len = prepend_size;
 357         offload_ctx->open_record = record;
 358         return 0;
 359 }
 360
 361 static int tls_do_allocation(struct sock *sk,
 362                              struct tls_offload_context_tx *offload_ctx,
 363                              struct page_frag *pfrag,
 364                              size_t prepend_size)
 365 {
 366         int ret;
 367
 368         if (!offload_ctx->open_record) {
 369                 if (unlikely(!skb_page_frag_refill(prepend_size, pfrag,
 370                                                    sk->sk_allocation))) {
 371                         READ_ONCE(sk->sk_prot)->enter_memory_pressure(sk);
 372                         sk_stream_moderate_sndbuf(sk);
 373                         return -ENOMEM;
 374                 }
 375
 376                 ret = tls_create_new_record(offload_ctx, pfrag, prepend_size);
 377                 if (ret)
 378                         return ret;
 379
 380                 if (pfrag->size > pfrag->offset)
 381                         return 0;
 382         }
 383
 384         if (!sk_page_frag_refill(sk, pfrag))
 385                 return -ENOMEM;
 386
 387         return 0;
 388 }
 389
 390 static int tls_device_copy_data(void *addr, size_t bytes, struct iov_iter *i)
 391 {
 392         size_t pre_copy, nocache;
 393
 394         pre_copy = ~((unsigned long)addr - 1) & (SMP_CACHE_BYTES - 1);
 395         if (pre_copy) {
 396                 pre_copy = min(pre_copy, bytes);
 397                 if (copy_from_iter(addr, pre_copy, i) != pre_copy)
 398                         return -EFAULT;
 399                 bytes -= pre_copy;
 400                 addr += pre_copy;
 401         }
 402
 403         nocache = round_down(bytes, SMP_CACHE_BYTES);
 404         if (copy_from_iter_nocache(addr, nocache, i) != nocache)
 405                 return -EFAULT;
 406         bytes -= nocache;
 407         addr += nocache;
 408
 409         if (bytes && copy_from_iter(addr, bytes, i) != bytes)
 410                 return -EFAULT;
 411
 412         return 0;
 413 }
 414
 415 union tls_iter_offset {
 416         struct iov_iter *msg_iter;
 417         int offset;
 418 };
 419
 420 static int tls_push_data(struct sock *sk,
 421                          union tls_iter_offset iter_offset,
 422                          size_t size, int flags,
 423                          unsigned char record_type,
 424                          struct page *zc_page)
 425 {
 426         struct tls_context *tls_ctx = tls_get_ctx(sk);
 427         struct tls_prot_info *prot = &tls_ctx->prot_info;
 428         struct tls_offload_context_tx *ctx = tls_offload_ctx_tx(tls_ctx);
 429         struct tls_record_info *record;
 430         int tls_push_record_flags;
 431         struct page_frag *pfrag;
 432         size_t orig_size = size;
 433         u32 max_open_record_len;
 434         bool more = false;
 435         bool done = false;
 436         int copy, rc = 0;
 437         long timeo;
 438
 439         if (flags &
 440             ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL | MSG_SENDPAGE_NOTLAST))
 441                 return -EOPNOTSUPP;
 442
 443         if (unlikely(sk->sk_err))
 444                 return -sk->sk_err;
 445
 446         flags |= MSG_SENDPAGE_DECRYPTED;
 447         tls_push_record_flags = flags | MSG_SENDPAGE_NOTLAST;
 448
 449         timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
 450         if (tls_is_partially_sent_record(tls_ctx)) {
 451                 rc = tls_push_partial_record(sk, tls_ctx, flags);
 452                 if (rc < 0)
 453                         return rc;
 454         }
 455
 456         pfrag = sk_page_frag(sk);
 457
 458         /* TLS_HEADER_SIZE is not counted as part of the TLS record, and
 459          * we need to leave room for an authentication tag.
 460          */
 461         max_open_record_len = TLS_MAX_PAYLOAD_SIZE +
 462                               prot->prepend_size;
 463         do {
 464                 rc = tls_do_allocation(sk, ctx, pfrag, prot->prepend_size);
 465                 if (unlikely(rc)) {
 466                         rc = sk_stream_wait_memory(sk, &timeo);
 467                         if (!rc)
 468                                 continue;
 469
 470                         record = ctx->open_record;
 471                         if (!record)
 472                                 break;
 473 handle_error:
 474                         if (record_type != TLS_RECORD_TYPE_DATA) {
 475                                 /* avoid sending partial
 476                                  * record with type !=
 477                                  * application_data
 478                                  */
 479                                 size = orig_size;
 480                                 destroy_record(record);
 481                                 ctx->open_record = NULL;
 482                         } else if (record->len > prot->prepend_size) {
 483                                 goto last_record;
 484                         }
 485
 486                         break;
 487                 }
 488
 489                 record = ctx->open_record;
 490
 491                 copy = min_t(size_t, size, max_open_record_len - record->len);
 492                 if (copy && zc_page) {
 493                         struct page_frag zc_pfrag;
 494
 495                         zc_pfrag.page = zc_page;
 496                         zc_pfrag.offset = iter_offset.offset;
 497                         zc_pfrag.size = copy;
 498                         tls_append_frag(record, &zc_pfrag, copy);
 499                 } else if (copy) {
 500                         copy = min_t(size_t, copy, pfrag->size - pfrag->offset);
 501
 502                         rc = tls_device_copy_data(page_address(pfrag->page) +
 503                                                   pfrag->offset, copy,
 504                                                   iter_offset.msg_iter);
 505                         if (rc)
 506                                 goto handle_error;
 507                         tls_append_frag(record, pfrag, copy);
 508                 }
 509
 510                 size -= copy;
 511                 if (!size) {
 512 last_record:
 513                         tls_push_record_flags = flags;
 514                         if (flags & (MSG_SENDPAGE_NOTLAST | MSG_MORE)) {
 515                                 more = true;
 516                                 break;
 517                         }
 518
 519                         done = true;
 520                 }
 521
 522                 if (done || record->len >= max_open_record_len ||
 523                     (record->num_frags >= MAX_SKB_FRAGS - 1)) {
 524                         rc = tls_device_record_close(sk, tls_ctx, record,
 525                                                      pfrag, record_type);
 526                         if (rc) {
 527                                 if (rc > 0) {
 528                                         size += rc;
 529                                 } else {
 530                                         size = orig_size;
 531                                         destroy_record(record);
 532                                         ctx->open_record = NULL;
 533                                         break;
 534                                 }
 535                         }
 536
 537                         rc = tls_push_record(sk,
 538                                              tls_ctx,
 539                                              ctx,
 540                                              record,
 541                                              tls_push_record_flags);
 542                         if (rc < 0)
 543                                 break;
 544                 }
 545         } while (!done);
 546
 547         tls_ctx->pending_open_record_frags = more;
 548
 549         if (orig_size - size > 0)
 550                 rc = orig_size - size;
 551
 552         return rc;
 553 }
 554
 555 int tls_device_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
 556 {
 557         unsigned char record_type = TLS_RECORD_TYPE_DATA;
 558         struct tls_context *tls_ctx = tls_get_ctx(sk);
 559         union tls_iter_offset iter;
 560         int rc;
 561
 562         mutex_lock(&tls_ctx->tx_lock);
 563         lock_sock(sk);
 564
 565         if (unlikely(msg->msg_controllen)) {
 566                 rc = tls_process_cmsg(sk, msg, &record_type);
 567                 if (rc)
 568                         goto out;
 569         }
 570
 571         iter.msg_iter = &msg->msg_iter;
 572         rc = tls_push_data(sk, iter, size, msg->msg_flags, record_type, NULL);
 573
 574 out:
 575         release_sock(sk);
 576         mutex_unlock(&tls_ctx->tx_lock);
 577         return rc;
 578 }
 579
 580 int tls_device_sendpage(struct sock *sk, struct page *page,
 581                         int offset, size_t size, int flags)
 582 {
 583         struct tls_context *tls_ctx = tls_get_ctx(sk);
 584         union tls_iter_offset iter_offset;
 585         struct iov_iter msg_iter;
 586         char *kaddr;
 587         struct kvec iov;
 588         int rc;
 589
 590         if (flags & MSG_SENDPAGE_NOTLAST)
 591                 flags |= MSG_MORE;
 592
 593         mutex_lock(&tls_ctx->tx_lock);
 594         lock_sock(sk);
 595
 596         if (flags & MSG_OOB) {
 597                 rc = -EOPNOTSUPP;
 598                 goto out;
 599         }
 600
 601         if (tls_ctx->zerocopy_sendfile) {
 602                 iter_offset.offset = offset;
 603                 rc = tls_push_data(sk, iter_offset, size,
 604                                    flags, TLS_RECORD_TYPE_DATA, page);
 605                 goto out;
 606         }
 607
 608         kaddr = kmap(page);
 609         iov.iov_base = kaddr + offset;
 610         iov.iov_len = size;
 611         iov_iter_kvec(&msg_iter, WRITE, &iov, 1, size);
 612         iter_offset.msg_iter = &msg_iter;
 613         rc = tls_push_data(sk, iter_offset, size, flags, TLS_RECORD_TYPE_DATA,
 614                            NULL);
 615         kunmap(page);
 616
 617 out:
 618         release_sock(sk);
 619         mutex_unlock(&tls_ctx->tx_lock);
 620         return rc;
 621 }
 622
 623 struct tls_record_info *tls_get_record(struct tls_offload_context_tx *context,
 624                                        u32 seq, u64 *p_record_sn)
 625 {
 626         u64 record_sn = context->hint_record_sn;
 627         struct tls_record_info *info, *last;
 628
 629         info = context->retransmit_hint;
 630         if (!info ||
 631             before(seq, info->end_seq - info->len)) {
 632                 /* if retransmit_hint is irrelevant start
 633                  * from the beginning of the list
 634                  */
 635                 info = list_first_entry_or_null(&context->records_list,
 636                                                 struct tls_record_info, list);
 637                 if (!info)
 638                         return NULL;
 639                 /* send the start_marker record if seq number is before the
 640                  * tls offload start marker sequence number. This record is
 641                  * required to handle TCP packets which are before TLS offload
 642                  * started.
 643                  *  And if it's not start marker, look if this seq number
 644                  * belongs to the list.
 645                  */
 646                 if (likely(!tls_record_is_start_marker(info))) {
 647                         /* we have the first record, get the last record to see
 648                          * if this seq number belongs to the list.
 649                          */
 650                         last = list_last_entry(&context->records_list,
 651                                                struct tls_record_info, list);
 652
 653                         if (!between(seq, tls_record_start_seq(info),
 654                                      last->end_seq))
 655                                 return NULL;
 656                 }
 657                 record_sn = context->unacked_record_sn;
 658         }
 659
 660         /* We just need the _rcu for the READ_ONCE() */
 661         rcu_read_lock();
 662         list_for_each_entry_from_rcu(info, &context->records_list, list) {
 663                 if (before(seq, info->end_seq)) {
 664                         if (!context->retransmit_hint ||
 665                             after(info->end_seq,
 666                                   context->retransmit_hint->end_seq)) {
 667                                 context->hint_record_sn = record_sn;
 668                                 context->retransmit_hint = info;
 669                         }
 670                         *p_record_sn = record_sn;
 671                         goto exit_rcu_unlock;
 672                 }
 673                 record_sn++;
 674         }
 675         info = NULL;
 676
 677 exit_rcu_unlock:
 678         rcu_read_unlock();
 679         return info;
 680 }
 681 EXPORT_SYMBOL(tls_get_record);
 682
 683 static int tls_device_push_pending_record(struct sock *sk, int flags)
 684 {
 685         union tls_iter_offset iter;
 686         struct iov_iter msg_iter;
 687
 688         iov_iter_kvec(&msg_iter, WRITE, NULL, 0, 0);
 689         iter.msg_iter = &msg_iter;
 690         return tls_push_data(sk, iter, 0, flags, TLS_RECORD_TYPE_DATA, NULL);
 691 }
 692
 693 void tls_device_write_space(struct sock *sk, struct tls_context *ctx)
 694 {
 695         if (tls_is_partially_sent_record(ctx)) {
 696                 gfp_t sk_allocation = sk->sk_allocation;
 697
 698                 WARN_ON_ONCE(sk->sk_write_pending);
 699
 700                 sk->sk_allocation = GFP_ATOMIC;
 701                 tls_push_partial_record(sk, ctx,
 702                                         MSG_DONTWAIT | MSG_NOSIGNAL |
 703                                         MSG_SENDPAGE_DECRYPTED);
 704                 sk->sk_allocation = sk_allocation;
 705         }
 706 }
 707
 708 static void tls_device_resync_rx(struct tls_context *tls_ctx,
 709                                  struct sock *sk, u32 seq, u8 *rcd_sn)
 710 {
 711         struct tls_offload_context_rx *rx_ctx = tls_offload_ctx_rx(tls_ctx);
 712         struct net_device *netdev;
 713
 714         trace_tls_device_rx_resync_send(sk, seq, rcd_sn, rx_ctx->resync_type);
 715         rcu_read_lock();
 716         netdev = READ_ONCE(tls_ctx->netdev);
 717         if (netdev)
 718                 netdev->tlsdev_ops->tls_dev_resync(netdev, sk, seq, rcd_sn,
 719                                                    TLS_OFFLOAD_CTX_DIR_RX);
 720         rcu_read_unlock();
 721         TLS_INC_STATS(sock_net(sk), LINUX_MIB_TLSRXDEVICERESYNC);
 722 }
 723
 724 static bool
 725 tls_device_rx_resync_async(struct tls_offload_resync_async *resync_async,
 726                            s64 resync_req, u32 *seq, u16 *rcd_delta)
 727 {
 728         u32 is_async = resync_req & RESYNC_REQ_ASYNC;
 729         u32 req_seq = resync_req >> 32;
 730         u32 req_end = req_seq + ((resync_req >> 16) & 0xffff);
 731         u16 i;
 732
 733         *rcd_delta = 0;
 734
 735         if (is_async) {
 736                 /* shouldn't get to wraparound:
 737                  * too long in async stage, something bad happened
 738                  */
 739                 if (WARN_ON_ONCE(resync_async->rcd_delta == USHRT_MAX))
 740                         return false;
 741
 742                 /* asynchronous stage: log all headers seq such that
 743                  * req_seq <= seq <= end_seq, and wait for real resync request
 744                  */
 745                 if (before(*seq, req_seq))
 746                         return false;
 747                 if (!after(*seq, req_end) &&
 748                     resync_async->loglen < TLS_DEVICE_RESYNC_ASYNC_LOGMAX)
 749                         resync_async->log[resync_async->loglen++] = *seq;
 750
 751                 resync_async->rcd_delta++;
 752
 753                 return false;
 754         }
 755
 756         /* synchronous stage: check against the logged entries and
 757          * proceed to check the next entries if no match was found
 758          */
 759         for (i = 0; i < resync_async->loglen; i++)
 760                 if (req_seq == resync_async->log[i] &&
 761                     atomic64_try_cmpxchg(&resync_async->req, &resync_req, 0)) {
 762                         *rcd_delta = resync_async->rcd_delta - i;
 763                         *seq = req_seq;
 764                         resync_async->loglen = 0;
 765                         resync_async->rcd_delta = 0;
 766                         return true;
 767                 }
 768
 769         resync_async->loglen = 0;
 770         resync_async->rcd_delta = 0;
 771
 772         if (req_seq == *seq &&
 773             atomic64_try_cmpxchg(&resync_async->req,
 774                                  &resync_req, 0))
 775                 return true;
 776
 777         return false;
 778 }
 779
 780 void tls_device_rx_resync_new_rec(struct sock *sk, u32 rcd_len, u32 seq)
 781 {
 782         struct tls_context *tls_ctx = tls_get_ctx(sk);
 783         struct tls_offload_context_rx *rx_ctx;
 784         u8 rcd_sn[TLS_MAX_REC_SEQ_SIZE];
 785         u32 sock_data, is_req_pending;
 786         struct tls_prot_info *prot;
 787         s64 resync_req;
 788         u16 rcd_delta;
 789         u32 req_seq;
 790
 791         if (tls_ctx->rx_conf != TLS_HW)
 792                 return;
 793         if (unlikely(test_bit(TLS_RX_DEV_DEGRADED, &tls_ctx->flags)))
 794                 return;
 795
 796         prot = &tls_ctx->prot_info;
 797         rx_ctx = tls_offload_ctx_rx(tls_ctx);
 798         memcpy(rcd_sn, tls_ctx->rx.rec_seq, prot->rec_seq_size);
 799
 800         switch (rx_ctx->resync_type) {
 801         case TLS_OFFLOAD_SYNC_TYPE_DRIVER_REQ:
 802                 resync_req = atomic64_read(&rx_ctx->resync_req);
 803                 req_seq = resync_req >> 32;
 804                 seq += TLS_HEADER_SIZE - 1;
 805                 is_req_pending = resync_req;
 806
 807                 if (likely(!is_req_pending) || req_seq != seq ||
 808                     !atomic64_try_cmpxchg(&rx_ctx->resync_req, &resync_req, 0))
 809                         return;
 810                 break;
 811         case TLS_OFFLOAD_SYNC_TYPE_CORE_NEXT_HINT:
 812                 if (likely(!rx_ctx->resync_nh_do_now))
 813                         return;
 814
 815                 /* head of next rec is already in, note that the sock_inq will
 816                  * include the currently parsed message when called from parser
 817                  */
 818                 sock_data = tcp_inq(sk);
 819                 if (sock_data > rcd_len) {
 820                         trace_tls_device_rx_resync_nh_delay(sk, sock_data,
 821                                                             rcd_len);
 822                         return;
 823                 }
 824
 825                 rx_ctx->resync_nh_do_now = 0;
 826                 seq += rcd_len;
 827                 tls_bigint_increment(rcd_sn, prot->rec_seq_size);
 828                 break;
 829         case TLS_OFFLOAD_SYNC_TYPE_DRIVER_REQ_ASYNC:
 830                 resync_req = atomic64_read(&rx_ctx->resync_async->req);
 831                 is_req_pending = resync_req;
 832                 if (likely(!is_req_pending))
 833                         return;
 834
 835                 if (!tls_device_rx_resync_async(rx_ctx->resync_async,
 836                                                 resync_req, &seq, &rcd_delta))
 837                         return;
 838                 tls_bigint_subtract(rcd_sn, rcd_delta);
 839                 break;
 840         }
 841
 842         tls_device_resync_rx(tls_ctx, sk, seq, rcd_sn);
 843 }
 844
 845 static void tls_device_core_ctrl_rx_resync(struct tls_context *tls_ctx,
 846                                            struct tls_offload_context_rx *ctx,
 847                                            struct sock *sk, struct sk_buff *skb)
 848 {
 849         struct strp_msg *rxm;
 850
 851         /* device will request resyncs by itself based on stream scan */
 852         if (ctx->resync_type != TLS_OFFLOAD_SYNC_TYPE_CORE_NEXT_HINT)
 853                 return;
 854         /* already scheduled */
 855         if (ctx->resync_nh_do_now)
 856                 return;
 857         /* seen decrypted fragments since last fully-failed record */
 858         if (ctx->resync_nh_reset) {
 859                 ctx->resync_nh_reset = 0;
 860                 ctx->resync_nh.decrypted_failed = 1;
 861                 ctx->resync_nh.decrypted_tgt = TLS_DEVICE_RESYNC_NH_START_IVAL;
 862                 return;
 863         }
 864
 865         if (++ctx->resync_nh.decrypted_failed <= ctx->resync_nh.decrypted_tgt)
 866                 return;
 867
 868         /* doing resync, bump the next target in case it fails */
 869         if (ctx->resync_nh.decrypted_tgt < TLS_DEVICE_RESYNC_NH_MAX_IVAL)
 870                 ctx->resync_nh.decrypted_tgt *= 2;
 871         else
 872                 ctx->resync_nh.decrypted_tgt += TLS_DEVICE_RESYNC_NH_MAX_IVAL;
 873
 874         rxm = strp_msg(skb);
 875
 876         /* head of next rec is already in, parser will sync for us */
 877         if (tcp_inq(sk) > rxm->full_len) {
 878                 trace_tls_device_rx_resync_nh_schedule(sk);
 879                 ctx->resync_nh_do_now = 1;
 880         } else {
 881                 struct tls_prot_info *prot = &tls_ctx->prot_info;
 882                 u8 rcd_sn[TLS_MAX_REC_SEQ_SIZE];
 883
 884                 memcpy(rcd_sn, tls_ctx->rx.rec_seq, prot->rec_seq_size);
 885                 tls_bigint_increment(rcd_sn, prot->rec_seq_size);
 886
 887                 tls_device_resync_rx(tls_ctx, sk, tcp_sk(sk)->copied_seq,
 888                                      rcd_sn);
 889         }
 890 }
 891
 892 static int tls_device_reencrypt(struct sock *sk, struct sk_buff *skb)
 893 {
 894         struct strp_msg *rxm = strp_msg(skb);
 895         int err = 0, offset = rxm->offset, copy, nsg, data_len, pos;
 896         struct sk_buff *skb_iter, *unused;
 897         struct scatterlist sg[1];
 898         char *orig_buf, *buf;
 899
 900         orig_buf = kmalloc(rxm->full_len + TLS_HEADER_SIZE +
 901                            TLS_CIPHER_AES_GCM_128_IV_SIZE, sk->sk_allocation);
 902         if (!orig_buf)
 903                 return -ENOMEM;
 904         buf = orig_buf;
 905
 906         nsg = skb_cow_data(skb, 0, &unused);
 907         if (unlikely(nsg < 0)) {
 908                 err = nsg;
 909                 goto free_buf;
 910         }
 911
 912         sg_init_table(sg, 1);
 913         sg_set_buf(&sg[0], buf,
 914                    rxm->full_len + TLS_HEADER_SIZE +
 915                    TLS_CIPHER_AES_GCM_128_IV_SIZE);
 916         err = skb_copy_bits(skb, offset, buf,
 917                             TLS_HEADER_SIZE + TLS_CIPHER_AES_GCM_128_IV_SIZE);
 918         if (err)
 919                 goto free_buf;
 920
 921         /* We are interested only in the decrypted data not the auth */
 922         err = decrypt_skb(sk, skb, sg);
 923         if (err != -EBADMSG)
 924                 goto free_buf;
 925         else
 926                 err = 0;
 927
 928         data_len = rxm->full_len - TLS_CIPHER_AES_GCM_128_TAG_SIZE;
 929
 930         if (skb_pagelen(skb) > offset) {
 931                 copy = min_t(int, skb_pagelen(skb) - offset, data_len);
 932
 933                 if (skb->decrypted) {
 934                         err = skb_store_bits(skb, offset, buf, copy);
 935                         if (err)
 936                                 goto free_buf;
 937                 }
 938
 939                 offset += copy;
 940                 buf += copy;
 941         }
 942
 943         pos = skb_pagelen(skb);
 944         skb_walk_frags(skb, skb_iter) {
 945                 int frag_pos;
 946
 947                 /* Practically all frags must belong to msg if reencrypt
 948                  * is needed with current strparser and coalescing logic,
 949                  * but strparser may "get optimized", so let's be safe.
 950                  */
 951                 if (pos + skb_iter->len <= offset)
 952                         goto done_with_frag;
 953                 if (pos >= data_len + rxm->offset)
 954                         break;
 955
 956                 frag_pos = offset - pos;
 957                 copy = min_t(int, skb_iter->len - frag_pos,
 958                              data_len + rxm->offset - offset);
 959
 960                 if (skb_iter->decrypted) {
 961                         err = skb_store_bits(skb_iter, frag_pos, buf, copy);
 962                         if (err)
 963                                 goto free_buf;
 964                 }
 965
 966                 offset += copy;
 967                 buf += copy;
 968 done_with_frag:
 969                 pos += skb_iter->len;
 970         }
 971
 972 free_buf:
 973         kfree(orig_buf);
 974         return err;
 975 }
 976
 977 int tls_device_decrypted(struct sock *sk, struct tls_context *tls_ctx,
 978                          struct sk_buff *skb, struct strp_msg *rxm)
 979 {
 980         struct tls_offload_context_rx *ctx = tls_offload_ctx_rx(tls_ctx);
 981         int is_decrypted = skb->decrypted;
 982         int is_encrypted = !is_decrypted;
 983         struct sk_buff *skb_iter;
 984
 985         /* Check if all the data is decrypted already */
 986         skb_walk_frags(skb, skb_iter) {
 987                 is_decrypted &= skb_iter->decrypted;
 988                 is_encrypted &= !skb_iter->decrypted;
 989         }
 990
 991         trace_tls_device_decrypted(sk, tcp_sk(sk)->copied_seq - rxm->full_len,
 992                                    tls_ctx->rx.rec_seq, rxm->full_len,
 993                                    is_encrypted, is_decrypted);
 994
 995         if (unlikely(test_bit(TLS_RX_DEV_DEGRADED, &tls_ctx->flags))) {
 996                 if (likely(is_encrypted || is_decrypted))
 997                         return is_decrypted;
 998
 999                 /* After tls_device_down disables the offload, the next SKB will
1000                  * likely have initial fragments decrypted, and final ones not
1001                  * decrypted. We need to reencrypt that single SKB.
1002                  */
1003                 return tls_device_reencrypt(sk, skb);
1004         }
1005
1006         /* Return immediately if the record is either entirely plaintext or
1007          * entirely ciphertext. Otherwise handle reencrypt partially decrypted
1008          * record.
1009          */
1010         if (is_decrypted) {
1011                 ctx->resync_nh_reset = 1;
1012                 return is_decrypted;
1013         }
1014         if (is_encrypted) {
1015                 tls_device_core_ctrl_rx_resync(tls_ctx, ctx, sk, skb);
1016                 return 0;
1017         }
1018
1019         ctx->resync_nh_reset = 1;
1020         return tls_device_reencrypt(sk, skb);
1021 }
1022
1023 static void tls_device_attach(struct tls_context *ctx, struct sock *sk,
1024                               struct net_device *netdev)
1025 {
1026         if (sk->sk_destruct != tls_device_sk_destruct) {
1027                 refcount_set(&ctx->refcount, 1);
1028                 dev_hold(netdev);
1029                 ctx->netdev = netdev;
1030                 spin_lock_irq(&tls_device_lock);
1031                 list_add_tail(&ctx->list, &tls_device_list);
1032                 spin_unlock_irq(&tls_device_lock);
1033
1034                 ctx->sk_destruct = sk->sk_destruct;
1035                 smp_store_release(&sk->sk_destruct, tls_device_sk_destruct);
1036         }
1037 }
1038
1039 int tls_set_device_offload(struct sock *sk, struct tls_context *ctx)
1040 {
1041         u16 nonce_size, tag_size, iv_size, rec_seq_size, salt_size;
1042         struct tls_context *tls_ctx = tls_get_ctx(sk);
1043         struct tls_prot_info *prot = &tls_ctx->prot_info;
1044         struct tls_record_info *start_marker_record;
1045         struct tls_offload_context_tx *offload_ctx;
1046         struct tls_crypto_info *crypto_info;
1047         struct net_device *netdev;
1048         char *iv, *rec_seq;
1049         struct sk_buff *skb;
1050         __be64 rcd_sn;
1051         int rc;
1052
1053         if (!ctx)
1054                 return -EINVAL;
1055
1056         if (ctx->priv_ctx_tx)
1057                 return -EEXIST;
1058
1059         netdev = get_netdev_for_sock(sk);
1060         if (!netdev) {
1061                 pr_err_ratelimited("%s: netdev not found\n", __func__);
1062                 return -EINVAL;
1063         }
1064
1065         if (!(netdev->features & NETIF_F_HW_TLS_TX)) {
1066                 rc = -EOPNOTSUPP;
1067                 goto release_netdev;
1068         }
1069
1070         crypto_info = &ctx->crypto_send.info;
1071         if (crypto_info->version != TLS_1_2_VERSION) {
1072                 rc = -EOPNOTSUPP;
1073                 goto release_netdev;
1074         }
1075
1076         switch (crypto_info->cipher_type) {
1077         case TLS_CIPHER_AES_GCM_128:
1078                 nonce_size = TLS_CIPHER_AES_GCM_128_IV_SIZE;
1079                 tag_size = TLS_CIPHER_AES_GCM_128_TAG_SIZE;
1080                 iv_size = TLS_CIPHER_AES_GCM_128_IV_SIZE;
1081                 iv = ((struct tls12_crypto_info_aes_gcm_128 *)crypto_info)->iv;
1082                 rec_seq_size = TLS_CIPHER_AES_GCM_128_REC_SEQ_SIZE;
1083                 salt_size = TLS_CIPHER_AES_GCM_128_SALT_SIZE;
1084                 rec_seq =
1085                  ((struct tls12_crypto_info_aes_gcm_128 *)crypto_info)->rec_seq;
1086                 break;
1087         default:
1088                 rc = -EINVAL;
1089                 goto release_netdev;
1090         }
1091
1092         /* Sanity-check the rec_seq_size for stack allocations */
1093         if (rec_seq_size > TLS_MAX_REC_SEQ_SIZE) {
1094                 rc = -EINVAL;
1095                 goto release_netdev;
1096         }
1097
1098         prot->version = crypto_info->version;
1099         prot->cipher_type = crypto_info->cipher_type;
1100         prot->prepend_size = TLS_HEADER_SIZE + nonce_size;
1101         prot->tag_size = tag_size;
1102         prot->overhead_size = prot->prepend_size + prot->tag_size;
1103         prot->iv_size = iv_size;
1104         prot->salt_size = salt_size;
1105         ctx->tx.iv = kmalloc(iv_size + TLS_CIPHER_AES_GCM_128_SALT_SIZE,
1106                              GFP_KERNEL);
1107         if (!ctx->tx.iv) {
1108                 rc = -ENOMEM;
1109                 goto release_netdev;
1110         }
1111
1112         memcpy(ctx->tx.iv + TLS_CIPHER_AES_GCM_128_SALT_SIZE, iv, iv_size);
1113
1114         prot->rec_seq_size = rec_seq_size;
1115         ctx->tx.rec_seq = kmemdup(rec_seq, rec_seq_size, GFP_KERNEL);
1116         if (!ctx->tx.rec_seq) {
1117                 rc = -ENOMEM;
1118                 goto free_iv;
1119         }
1120
1121         start_marker_record = kmalloc(sizeof(*start_marker_record), GFP_KERNEL);
1122         if (!start_marker_record) {
1123                 rc = -ENOMEM;
1124                 goto free_rec_seq;
1125         }
1126
1127         offload_ctx = kzalloc(TLS_OFFLOAD_CONTEXT_SIZE_TX, GFP_KERNEL);
1128         if (!offload_ctx) {
1129                 rc = -ENOMEM;
1130                 goto free_marker_record;
1131         }
1132
1133         rc = tls_sw_fallback_init(sk, offload_ctx, crypto_info);
1134         if (rc)
1135                 goto free_offload_ctx;
1136
1137         /* start at rec_seq - 1 to account for the start marker record */
1138         memcpy(&rcd_sn, ctx->tx.rec_seq, sizeof(rcd_sn));
1139         offload_ctx->unacked_record_sn = be64_to_cpu(rcd_sn) - 1;
1140
1141         start_marker_record->end_seq = tcp_sk(sk)->write_seq;
1142         start_marker_record->len = 0;
1143         start_marker_record->num_frags = 0;
1144
1145         INIT_LIST_HEAD(&offload_ctx->records_list);
1146         list_add_tail(&start_marker_record->list, &offload_ctx->records_list);
1147         spin_lock_init(&offload_ctx->lock);
1148         sg_init_table(offload_ctx->sg_tx_data,
1149                       ARRAY_SIZE(offload_ctx->sg_tx_data));
1150
1151         clean_acked_data_enable(inet_csk(sk), &tls_icsk_clean_acked);
1152         ctx->push_pending_record = tls_device_push_pending_record;
1153
1154         /* TLS offload is greatly simplified if we don't send
1155          * SKBs where only part of the payload needs to be encrypted.
1156          * So mark the last skb in the write queue as end of record.
1157          */
1158         skb = tcp_write_queue_tail(sk);
1159         if (skb)
1160                 TCP_SKB_CB(skb)->eor = 1;
1161
1162         /* Avoid offloading if the device is down
1163          * We don't want to offload new flows after
1164          * the NETDEV_DOWN event
1165          *
1166          * device_offload_lock is taken in tls_devices's NETDEV_DOWN
1167          * handler thus protecting from the device going down before
1168          * ctx was added to tls_device_list.
1169          */
1170         down_read(&device_offload_lock);
1171         if (!(netdev->flags & IFF_UP)) {
1172                 rc = -EINVAL;
1173                 goto release_lock;
1174         }
1175
1176         ctx->priv_ctx_tx = offload_ctx;
1177         rc = netdev->tlsdev_ops->tls_dev_add(netdev, sk, TLS_OFFLOAD_CTX_DIR_TX,
1178                                              &ctx->crypto_send.info,
1179                                              tcp_sk(sk)->write_seq);
1180         trace_tls_device_offload_set(sk, TLS_OFFLOAD_CTX_DIR_TX,
1181                                      tcp_sk(sk)->write_seq, rec_seq, rc);
1182         if (rc)
1183                 goto release_lock;
1184
1185         tls_device_attach(ctx, sk, netdev);
1186         up_read(&device_offload_lock);
1187
1188         /* following this assignment tls_is_sk_tx_device_offloaded
1189          * will return true and the context might be accessed
1190          * by the netdev's xmit function.
1191          */
1192         smp_store_release(&sk->sk_validate_xmit_skb, tls_validate_xmit_skb);
1193         dev_put(netdev);
1194
1195         return 0;
1196
1197 release_lock:
1198         up_read(&device_offload_lock);
1199         clean_acked_data_disable(inet_csk(sk));
1200         crypto_free_aead(offload_ctx->aead_send);
1201 free_offload_ctx:
1202         kfree(offload_ctx);
1203         ctx->priv_ctx_tx = NULL;
1204 free_marker_record:
1205         kfree(start_marker_record);
1206 free_rec_seq:
1207         kfree(ctx->tx.rec_seq);
1208 free_iv:
1209         kfree(ctx->tx.iv);
1210 release_netdev:
1211         dev_put(netdev);
1212         return rc;
1213 }
1214
1215 int tls_set_device_offload_rx(struct sock *sk, struct tls_context *ctx)
1216 {
1217         struct tls12_crypto_info_aes_gcm_128 *info;
1218         struct tls_offload_context_rx *context;
1219         struct net_device *netdev;
1220         int rc = 0;
1221
1222         if (ctx->crypto_recv.info.version != TLS_1_2_VERSION)
1223                 return -EOPNOTSUPP;
1224
1225         netdev = get_netdev_for_sock(sk);
1226         if (!netdev) {
1227                 pr_err_ratelimited("%s: netdev not found\n", __func__);
1228                 return -EINVAL;
1229         }
1230
1231         if (!(netdev->features & NETIF_F_HW_TLS_RX)) {
1232                 rc = -EOPNOTSUPP;
1233                 goto release_netdev;
1234         }
1235
1236         /* Avoid offloading if the device is down
1237          * We don't want to offload new flows after
1238          * the NETDEV_DOWN event
1239          *
1240          * device_offload_lock is taken in tls_devices's NETDEV_DOWN
1241          * handler thus protecting from the device going down before
1242          * ctx was added to tls_device_list.
1243          */
1244         down_read(&device_offload_lock);
1245         if (!(netdev->flags & IFF_UP)) {
1246                 rc = -EINVAL;
1247                 goto release_lock;
1248         }
1249
1250         context = kzalloc(TLS_OFFLOAD_CONTEXT_SIZE_RX, GFP_KERNEL);
1251         if (!context) {
1252                 rc = -ENOMEM;
1253                 goto release_lock;
1254         }
1255         context->resync_nh_reset = 1;
1256
1257         ctx->priv_ctx_rx = context;
1258         rc = tls_set_sw_offload(sk, ctx, 0);
1259         if (rc)
1260                 goto release_ctx;
1261
1262         rc = netdev->tlsdev_ops->tls_dev_add(netdev, sk, TLS_OFFLOAD_CTX_DIR_RX,
1263                                              &ctx->crypto_recv.info,
1264                                              tcp_sk(sk)->copied_seq);
1265         info = (void *)&ctx->crypto_recv.info;
1266         trace_tls_device_offload_set(sk, TLS_OFFLOAD_CTX_DIR_RX,
1267                                      tcp_sk(sk)->copied_seq, info->rec_seq, rc);
1268         if (rc)
1269                 goto free_sw_resources;
1270
1271         tls_device_attach(ctx, sk, netdev);
1272         up_read(&device_offload_lock);
1273
1274         dev_put(netdev);
1275
1276         return 0;
1277
1278 free_sw_resources:
1279         up_read(&device_offload_lock);
1280         tls_sw_free_resources_rx(sk);
1281         down_read(&device_offload_lock);
1282 release_ctx:
1283         ctx->priv_ctx_rx = NULL;
1284 release_lock:
1285         up_read(&device_offload_lock);
1286 release_netdev:
1287         dev_put(netdev);
1288         return rc;
1289 }
1290
1291 void tls_device_offload_cleanup_rx(struct sock *sk)
1292 {
1293         struct tls_context *tls_ctx = tls_get_ctx(sk);
1294         struct net_device *netdev;
1295
1296         down_read(&device_offload_lock);
1297         netdev = tls_ctx->netdev;
1298         if (!netdev)
1299                 goto out;
1300
1301         netdev->tlsdev_ops->tls_dev_del(netdev, tls_ctx,
1302                                         TLS_OFFLOAD_CTX_DIR_RX);
1303
1304         if (tls_ctx->tx_conf != TLS_HW) {
1305                 dev_put(netdev);
1306                 tls_ctx->netdev = NULL;
1307         } else {
1308                 set_bit(TLS_RX_DEV_CLOSED, &tls_ctx->flags);
1309         }
1310 out:
1311         up_read(&device_offload_lock);
1312         tls_sw_release_resources_rx(sk);
1313 }
1314
1315 static int tls_device_down(struct net_device *netdev)
1316 {
1317         struct tls_context *ctx, *tmp;
1318         unsigned long flags;
1319         LIST_HEAD(list);
1320
1321         /* Request a write lock to block new offload attempts */
1322         down_write(&device_offload_lock);
1323
1324         spin_lock_irqsave(&tls_device_lock, flags);
1325         list_for_each_entry_safe(ctx, tmp, &tls_device_list, list) {
1326                 if (ctx->netdev != netdev ||
1327                     !refcount_inc_not_zero(&ctx->refcount))
1328                         continue;
1329
1330                 list_move(&ctx->list, &list);
1331         }
1332         spin_unlock_irqrestore(&tls_device_lock, flags);
1333
1334         list_for_each_entry_safe(ctx, tmp, &list, list) {
1335                 /* Stop offloaded TX and switch to the fallback.
1336                  * tls_is_sk_tx_device_offloaded will return false.
1337                  */
1338                 WRITE_ONCE(ctx->sk->sk_validate_xmit_skb, tls_validate_xmit_skb_sw);
1339
1340                 /* Stop the RX and TX resync.
1341                  * tls_dev_resync must not be called after tls_dev_del.
1342                  */
1343                 WRITE_ONCE(ctx->netdev, NULL);
1344
1345                 /* Start skipping the RX resync logic completely. */
1346                 set_bit(TLS_RX_DEV_DEGRADED, &ctx->flags);
1347
1348                 /* Sync with inflight packets. After this point:
1349                  * TX: no non-encrypted packets will be passed to the driver.
1350                  * RX: resync requests from the driver will be ignored.
1351                  */
1352                 synchronize_net();
1353
1354                 /* Release the offload context on the driver side. */
1355                 if (ctx->tx_conf == TLS_HW)
1356                         netdev->tlsdev_ops->tls_dev_del(netdev, ctx,
1357                                                         TLS_OFFLOAD_CTX_DIR_TX);
1358                 if (ctx->rx_conf == TLS_HW &&
1359                     !test_bit(TLS_RX_DEV_CLOSED, &ctx->flags))
1360                         netdev->tlsdev_ops->tls_dev_del(netdev, ctx,
1361                                                         TLS_OFFLOAD_CTX_DIR_RX);
1362
1363                 dev_put(netdev);
1364
1365                 /* Move the context to a separate list for two reasons:
1366                  * 1. When the context is deallocated, list_del is called.
1367                  * 2. It's no longer an offloaded context, so we don't want to
1368                  *    run offload-specific code on this context.
1369                  */
1370                 spin_lock_irqsave(&tls_device_lock, flags);
1371                 list_move_tail(&ctx->list, &tls_device_down_list);
1372                 spin_unlock_irqrestore(&tls_device_lock, flags);
1373
1374                 /* Device contexts for RX and TX will be freed in on sk_destruct
1375                  * by tls_device_free_ctx. rx_conf and tx_conf stay in TLS_HW.
1376                  * Now release the ref taken above.
1377                  */
1378                 if (refcount_dec_and_test(&ctx->refcount))
1379                         tls_device_free_ctx(ctx);
1380         }
1381
1382         up_write(&device_offload_lock);
1383
1384         flush_work(&tls_device_gc_work);
1385
1386         return NOTIFY_DONE;
1387 }
1388
1389 static int tls_dev_event(struct notifier_block *this, unsigned long event,
1390                          void *ptr)
1391 {
1392         struct net_device *dev = netdev_notifier_info_to_dev(ptr);
1393
1394         if (!dev->tlsdev_ops &&
1395             !(dev->features & (NETIF_F_HW_TLS_RX | NETIF_F_HW_TLS_TX)))
1396                 return NOTIFY_DONE;
1397
1398         switch (event) {
1399         case NETDEV_REGISTER:
1400         case NETDEV_FEAT_CHANGE:
1401                 if (netif_is_bond_master(dev))
1402                         return NOTIFY_DONE;
1403                 if ((dev->features & NETIF_F_HW_TLS_RX) &&
1404                     !dev->tlsdev_ops->tls_dev_resync)
1405                         return NOTIFY_BAD;
1406
1407                 if  (dev->tlsdev_ops &&
1408                      dev->tlsdev_ops->tls_dev_add &&
1409                      dev->tlsdev_ops->tls_dev_del)
1410                         return NOTIFY_DONE;
1411                 else
1412                         return NOTIFY_BAD;
1413         case NETDEV_DOWN:
1414                 return tls_device_down(dev);
1415         }
1416         return NOTIFY_DONE;
1417 }
1418
1419 static struct notifier_block tls_dev_notifier = {
1420         .notifier_call  = tls_dev_event,
1421 };
1422
1423 void __init tls_device_init(void)
1424 {
1425         register_netdevice_notifier(&tls_dev_notifier);
1426 }
1427
1428 void __exit tls_device_cleanup(void)
1429 {
1430         unregister_netdevice_notifier(&tls_dev_notifier);
1431         flush_work(&tls_device_gc_work);
1432         clean_acked_data_flush();
1433 }