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 "trace.h"
  42
  43 /* device_offload_lock is used to synchronize tls_dev_add
  44  * against NETDEV_DOWN notifications.
  45  */
  46 static DECLARE_RWSEM(device_offload_lock);
  47
  48 static void tls_device_gc_task(struct work_struct *work);
  49
  50 static DECLARE_WORK(tls_device_gc_work, tls_device_gc_task);
  51 static LIST_HEAD(tls_device_gc_list);
  52 static LIST_HEAD(tls_device_list);
  53 static LIST_HEAD(tls_device_down_list);
  54 static DEFINE_SPINLOCK(tls_device_lock);
  55
  56 static void tls_device_free_ctx(struct tls_context *ctx)
  57 {
  58         if (ctx->tx_conf == TLS_HW) {
  59                 kfree(tls_offload_ctx_tx(ctx));
  60                 kfree(ctx->tx.rec_seq);
  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_gc_task(struct work_struct *work)
  71 {
  72         struct tls_context *ctx, *tmp;
  73         unsigned long flags;
  74         LIST_HEAD(gc_list);
  75
  76         spin_lock_irqsave(&tls_device_lock, flags);
  77         list_splice_init(&tls_device_gc_list, &gc_list);
  78         spin_unlock_irqrestore(&tls_device_lock, flags);
  79
  80         list_for_each_entry_safe(ctx, tmp, &gc_list, list) {
  81                 struct net_device *netdev = ctx->netdev;
  82
  83                 if (netdev && ctx->tx_conf == TLS_HW) {
  84                         netdev->tlsdev_ops->tls_dev_del(netdev, ctx,
  85                                                         TLS_OFFLOAD_CTX_DIR_TX);
  86                         dev_put(netdev);
  87                         ctx->netdev = NULL;
  88                 }
  89
  90                 list_del(&ctx->list);
  91                 tls_device_free_ctx(ctx);
  92         }
  93 }
  94
  95 static void tls_device_queue_ctx_destruction(struct tls_context *ctx)
  96 {
  97         unsigned long flags;
  98
  99         spin_lock_irqsave(&tls_device_lock, flags);
 100         if (unlikely(!refcount_dec_and_test(&ctx->refcount)))
 101                 goto unlock;
 102
 103         list_move_tail(&ctx->list, &tls_device_gc_list);
 104
 105         /* schedule_work inside the spinlock
 106          * to make sure tls_device_down waits for that work.
 107          */
 108         schedule_work(&tls_device_gc_work);
 109 unlock:
 110         spin_unlock_irqrestore(&tls_device_lock, flags);
 111 }
 112
 113 /* We assume that the socket is already connected */
 114 static struct net_device *get_netdev_for_sock(struct sock *sk)
 115 {
 116         struct dst_entry *dst = sk_dst_get(sk);
 117         struct net_device *netdev = NULL;
 118
 119         if (likely(dst)) {
 120                 netdev = netdev_sk_get_lowest_dev(dst->dev, sk);
 121                 dev_hold(netdev);
 122         }
 123
 124         dst_release(dst);
 125
 126         return netdev;
 127 }
 128
 129 static void destroy_record(struct tls_record_info *record)
 130 {
 131         int i;
 132
 133         for (i = 0; i < record->num_frags; i++)
 134                 __skb_frag_unref(&record->frags[i], false);
 135         kfree(record);
 136 }
 137
 138 static void delete_all_records(struct tls_offload_context_tx *offload_ctx)
 139 {
 140         struct tls_record_info *info, *temp;
 141
 142         list_for_each_entry_safe(info, temp, &offload_ctx->records_list, list) {
 143                 list_del(&info->list);
 144                 destroy_record(info);
 145         }
 146
 147         offload_ctx->retransmit_hint = NULL;
 148 }
 149
 150 static void tls_icsk_clean_acked(struct sock *sk, u32 acked_seq)
 151 {
 152         struct tls_context *tls_ctx = tls_get_ctx(sk);
 153         struct tls_record_info *info, *temp;
 154         struct tls_offload_context_tx *ctx;
 155         u64 deleted_records = 0;
 156         unsigned long flags;
 157
 158         if (!tls_ctx)
 159                 return;
 160
 161         ctx = tls_offload_ctx_tx(tls_ctx);
 162
 163         spin_lock_irqsave(&ctx->lock, flags);
 164         info = ctx->retransmit_hint;
 165         if (info && !before(acked_seq, info->end_seq))
 166                 ctx->retransmit_hint = NULL;
 167
 168         list_for_each_entry_safe(info, temp, &ctx->records_list, list) {
 169                 if (before(acked_seq, info->end_seq))
 170                         break;
 171                 list_del(&info->list);
 172
 173                 destroy_record(info);
 174                 deleted_records++;
 175         }
 176
 177         ctx->unacked_record_sn += deleted_records;
 178         spin_unlock_irqrestore(&ctx->lock, flags);
 179 }
 180
 181 /* At this point, there should be no references on this
 182  * socket and no in-flight SKBs associated with this
 183  * socket, so it is safe to free all the resources.
 184  */
 185 void tls_device_sk_destruct(struct sock *sk)
 186 {
 187         struct tls_context *tls_ctx = tls_get_ctx(sk);
 188         struct tls_offload_context_tx *ctx = tls_offload_ctx_tx(tls_ctx);
 189
 190         tls_ctx->sk_destruct(sk);
 191
 192         if (tls_ctx->tx_conf == TLS_HW) {
 193                 if (ctx->open_record)
 194                         destroy_record(ctx->open_record);
 195                 delete_all_records(ctx);
 196                 crypto_free_aead(ctx->aead_send);
 197                 clean_acked_data_disable(inet_csk(sk));
 198         }
 199
 200         tls_device_queue_ctx_destruction(tls_ctx);
 201 }
 202 EXPORT_SYMBOL_GPL(tls_device_sk_destruct);
 203
 204 void tls_device_free_resources_tx(struct sock *sk)
 205 {
 206         struct tls_context *tls_ctx = tls_get_ctx(sk);
 207
 208         tls_free_partial_record(sk, tls_ctx);
 209 }
 210
 211 void tls_offload_tx_resync_request(struct sock *sk, u32 got_seq, u32 exp_seq)
 212 {
 213         struct tls_context *tls_ctx = tls_get_ctx(sk);
 214
 215         trace_tls_device_tx_resync_req(sk, got_seq, exp_seq);
 216         WARN_ON(test_and_set_bit(TLS_TX_SYNC_SCHED, &tls_ctx->flags));
 217 }
 218 EXPORT_SYMBOL_GPL(tls_offload_tx_resync_request);
 219
 220 static void tls_device_resync_tx(struct sock *sk, struct tls_context *tls_ctx,
 221                                  u32 seq)
 222 {
 223         struct net_device *netdev;
 224         struct sk_buff *skb;
 225         int err = 0;
 226         u8 *rcd_sn;
 227
 228         skb = tcp_write_queue_tail(sk);
 229         if (skb)
 230                 TCP_SKB_CB(skb)->eor = 1;
 231
 232         rcd_sn = tls_ctx->tx.rec_seq;
 233
 234         trace_tls_device_tx_resync_send(sk, seq, rcd_sn);
 235         down_read(&device_offload_lock);
 236         netdev = tls_ctx->netdev;
 237         if (netdev)
 238                 err = netdev->tlsdev_ops->tls_dev_resync(netdev, sk, seq,
 239                                                          rcd_sn,
 240                                                          TLS_OFFLOAD_CTX_DIR_TX);
 241         up_read(&device_offload_lock);
 242         if (err)
 243                 return;
 244
 245         clear_bit_unlock(TLS_TX_SYNC_SCHED, &tls_ctx->flags);
 246 }
 247
 248 static void tls_append_frag(struct tls_record_info *record,
 249                             struct page_frag *pfrag,
 250                             int size)
 251 {
 252         skb_frag_t *frag;
 253
 254         frag = &record->frags[record->num_frags - 1];
 255         if (skb_frag_page(frag) == pfrag->page &&
 256             skb_frag_off(frag) + skb_frag_size(frag) == pfrag->offset) {
 257                 skb_frag_size_add(frag, size);
 258         } else {
 259                 ++frag;
 260                 __skb_frag_set_page(frag, pfrag->page);
 261                 skb_frag_off_set(frag, pfrag->offset);
 262                 skb_frag_size_set(frag, size);
 263                 ++record->num_frags;
 264                 get_page(pfrag->page);
 265         }
 266
 267         pfrag->offset += size;
 268         record->len += size;
 269 }
 270
 271 static int tls_push_record(struct sock *sk,
 272                            struct tls_context *ctx,
 273                            struct tls_offload_context_tx *offload_ctx,
 274                            struct tls_record_info *record,
 275                            int flags)
 276 {
 277         struct tls_prot_info *prot = &ctx->prot_info;
 278         struct tcp_sock *tp = tcp_sk(sk);
 279         skb_frag_t *frag;
 280         int i;
 281
 282         record->end_seq = tp->write_seq + record->len;
 283         list_add_tail_rcu(&record->list, &offload_ctx->records_list);
 284         offload_ctx->open_record = NULL;
 285
 286         if (test_bit(TLS_TX_SYNC_SCHED, &ctx->flags))
 287                 tls_device_resync_tx(sk, ctx, tp->write_seq);
 288
 289         tls_advance_record_sn(sk, prot, &ctx->tx);
 290
 291         for (i = 0; i < record->num_frags; i++) {
 292                 frag = &record->frags[i];
 293                 sg_unmark_end(&offload_ctx->sg_tx_data[i]);
 294                 sg_set_page(&offload_ctx->sg_tx_data[i], skb_frag_page(frag),
 295                             skb_frag_size(frag), skb_frag_off(frag));
 296                 sk_mem_charge(sk, skb_frag_size(frag));
 297                 get_page(skb_frag_page(frag));
 298         }
 299         sg_mark_end(&offload_ctx->sg_tx_data[record->num_frags - 1]);
 300
 301         /* all ready, send */
 302         return tls_push_sg(sk, ctx, offload_ctx->sg_tx_data, 0, flags);
 303 }
 304
 305 static int tls_device_record_close(struct sock *sk,
 306                                    struct tls_context *ctx,
 307                                    struct tls_record_info *record,
 308                                    struct page_frag *pfrag,
 309                                    unsigned char record_type)
 310 {
 311         struct tls_prot_info *prot = &ctx->prot_info;
 312         int ret;
 313
 314         /* append tag
 315          * device will fill in the tag, we just need to append a placeholder
 316          * use socket memory to improve coalescing (re-using a single buffer
 317          * increases frag count)
 318          * if we can't allocate memory now, steal some back from data
 319          */
 320         if (likely(skb_page_frag_refill(prot->tag_size, pfrag,
 321                                         sk->sk_allocation))) {
 322                 ret = 0;
 323                 tls_append_frag(record, pfrag, prot->tag_size);
 324         } else {
 325                 ret = prot->tag_size;
 326                 if (record->len <= prot->overhead_size)
 327                         return -ENOMEM;
 328         }
 329
 330         /* fill prepend */
 331         tls_fill_prepend(ctx, skb_frag_address(&record->frags[0]),
 332                          record->len - prot->overhead_size,
 333                          record_type);
 334         return ret;
 335 }
 336
 337 static int tls_create_new_record(struct tls_offload_context_tx *offload_ctx,
 338                                  struct page_frag *pfrag,
 339                                  size_t prepend_size)
 340 {
 341         struct tls_record_info *record;
 342         skb_frag_t *frag;
 343
 344         record = kmalloc(sizeof(*record), GFP_KERNEL);
 345         if (!record)
 346                 return -ENOMEM;
 347
 348         frag = &record->frags[0];
 349         __skb_frag_set_page(frag, pfrag->page);
 350         skb_frag_off_set(frag, pfrag->offset);
 351         skb_frag_size_set(frag, prepend_size);
 352
 353         get_page(pfrag->page);
 354         pfrag->offset += prepend_size;
 355
 356         record->num_frags = 1;
 357         record->len = prepend_size;
 358         offload_ctx->open_record = record;
 359         return 0;
 360 }
 361
 362 static int tls_do_allocation(struct sock *sk,
 363                              struct tls_offload_context_tx *offload_ctx,
 364                              struct page_frag *pfrag,
 365                              size_t prepend_size)
 366 {
 367         int ret;
 368
 369         if (!offload_ctx->open_record) {
 370                 if (unlikely(!skb_page_frag_refill(prepend_size, pfrag,
 371                                                    sk->sk_allocation))) {
 372                         READ_ONCE(sk->sk_prot)->enter_memory_pressure(sk);
 373                         sk_stream_moderate_sndbuf(sk);
 374                         return -ENOMEM;
 375                 }
 376
 377                 ret = tls_create_new_record(offload_ctx, pfrag, prepend_size);
 378                 if (ret)
 379                         return ret;
 380
 381                 if (pfrag->size > pfrag->offset)
 382                         return 0;
 383         }
 384
 385         if (!sk_page_frag_refill(sk, pfrag))
 386                 return -ENOMEM;
 387
 388         return 0;
 389 }
 390
 391 static int tls_device_copy_data(void *addr, size_t bytes, struct iov_iter *i)
 392 {
 393         size_t pre_copy, nocache;
 394
 395         pre_copy = ~((unsigned long)addr - 1) & (SMP_CACHE_BYTES - 1);
 396         if (pre_copy) {
 397                 pre_copy = min(pre_copy, bytes);
 398                 if (copy_from_iter(addr, pre_copy, i) != pre_copy)
 399                         return -EFAULT;
 400                 bytes -= pre_copy;
 401                 addr += pre_copy;
 402         }
 403
 404         nocache = round_down(bytes, SMP_CACHE_BYTES);
 405         if (copy_from_iter_nocache(addr, nocache, i) != nocache)
 406                 return -EFAULT;
 407         bytes -= nocache;
 408         addr += nocache;
 409
 410         if (bytes && copy_from_iter(addr, bytes, i) != bytes)
 411                 return -EFAULT;
 412
 413         return 0;
 414 }
 415
 416 union tls_iter_offset {
 417         struct iov_iter *msg_iter;
 418         int offset;
 419 };
 420
 421 static int tls_push_data(struct sock *sk,
 422                          union tls_iter_offset iter_offset,
 423                          size_t size, int flags,
 424                          unsigned char record_type,
 425                          struct page *zc_page)
 426 {
 427         struct tls_context *tls_ctx = tls_get_ctx(sk);
 428         struct tls_prot_info *prot = &tls_ctx->prot_info;
 429         struct tls_offload_context_tx *ctx = tls_offload_ctx_tx(tls_ctx);
 430         struct tls_record_info *record;
 431         int tls_push_record_flags;
 432         struct page_frag *pfrag;
 433         size_t orig_size = size;
 434         u32 max_open_record_len;
 435         bool more = false;
 436         bool done = false;
 437         int copy, rc = 0;
 438         long timeo;
 439
 440         if (flags &
 441             ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL | MSG_SENDPAGE_NOTLAST))
 442                 return -EOPNOTSUPP;
 443
 444         if (unlikely(sk->sk_err))
 445                 return -sk->sk_err;
 446
 447         flags |= MSG_SENDPAGE_DECRYPTED;
 448         tls_push_record_flags = flags | MSG_SENDPAGE_NOTLAST;
 449
 450         timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
 451         if (tls_is_partially_sent_record(tls_ctx)) {
 452                 rc = tls_push_partial_record(sk, tls_ctx, flags);
 453                 if (rc < 0)
 454                         return rc;
 455         }
 456
 457         pfrag = sk_page_frag(sk);
 458
 459         /* TLS_HEADER_SIZE is not counted as part of the TLS record, and
 460          * we need to leave room for an authentication tag.
 461          */
 462         max_open_record_len = TLS_MAX_PAYLOAD_SIZE +
 463                               prot->prepend_size;
 464         do {
 465                 rc = tls_do_allocation(sk, ctx, pfrag, prot->prepend_size);
 466                 if (unlikely(rc)) {
 467                         rc = sk_stream_wait_memory(sk, &timeo);
 468                         if (!rc)
 469                                 continue;
 470
 471                         record = ctx->open_record;
 472                         if (!record)
 473                                 break;
 474 handle_error:
 475                         if (record_type != TLS_RECORD_TYPE_DATA) {
 476                                 /* avoid sending partial
 477                                  * record with type !=
 478                                  * application_data
 479                                  */
 480                                 size = orig_size;
 481                                 destroy_record(record);
 482                                 ctx->open_record = NULL;
 483                         } else if (record->len > prot->prepend_size) {
 484                                 goto last_record;
 485                         }
 486
 487                         break;
 488                 }
 489
 490                 record = ctx->open_record;
 491
 492                 copy = min_t(size_t, size, max_open_record_len - record->len);
 493                 if (copy && zc_page) {
 494                         struct page_frag zc_pfrag;
 495
 496                         zc_pfrag.page = zc_page;
 497                         zc_pfrag.offset = iter_offset.offset;
 498                         zc_pfrag.size = copy;
 499                         tls_append_frag(record, &zc_pfrag, copy);
 500                 } else if (copy) {
 501                         copy = min_t(size_t, copy, pfrag->size - pfrag->offset);
 502
 503                         rc = tls_device_copy_data(page_address(pfrag->page) +
 504                                                   pfrag->offset, copy,
 505                                                   iter_offset.msg_iter);
 506                         if (rc)
 507                                 goto handle_error;
 508                         tls_append_frag(record, pfrag, copy);
 509                 }
 510
 511                 size -= copy;
 512                 if (!size) {
 513 last_record:
 514                         tls_push_record_flags = flags;
 515                         if (flags & (MSG_SENDPAGE_NOTLAST | MSG_MORE)) {
 516                                 more = true;
 517                                 break;
 518                         }
 519
 520                         done = true;
 521                 }
 522
 523                 if (done || record->len >= max_open_record_len ||
 524                     (record->num_frags >= MAX_SKB_FRAGS - 1)) {
 525                         rc = tls_device_record_close(sk, tls_ctx, record,
 526                                                      pfrag, record_type);
 527                         if (rc) {
 528                                 if (rc > 0) {
 529                                         size += rc;
 530                                 } else {
 531                                         size = orig_size;
 532                                         destroy_record(record);
 533                                         ctx->open_record = NULL;
 534                                         break;
 535                                 }
 536                         }
 537
 538                         rc = tls_push_record(sk,
 539                                              tls_ctx,
 540                                              ctx,
 541                                              record,
 542                                              tls_push_record_flags);
 543                         if (rc < 0)
 544                                 break;
 545                 }
 546         } while (!done);
 547
 548         tls_ctx->pending_open_record_frags = more;
 549
 550         if (orig_size - size > 0)
 551                 rc = orig_size - size;
 552
 553         return rc;
 554 }
 555
 556 int tls_device_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
 557 {
 558         unsigned char record_type = TLS_RECORD_TYPE_DATA;
 559         struct tls_context *tls_ctx = tls_get_ctx(sk);
 560         union tls_iter_offset iter;
 561         int rc;
 562
 563         mutex_lock(&tls_ctx->tx_lock);
 564         lock_sock(sk);
 565
 566         if (unlikely(msg->msg_controllen)) {
 567                 rc = tls_proccess_cmsg(sk, msg, &record_type);
 568                 if (rc)
 569                         goto out;
 570         }
 571
 572         iter.msg_iter = &msg->msg_iter;
 573         rc = tls_push_data(sk, iter, size, msg->msg_flags, record_type, NULL);
 574
 575 out:
 576         release_sock(sk);
 577         mutex_unlock(&tls_ctx->tx_lock);
 578         return rc;
 579 }
 580
 581 int tls_device_sendpage(struct sock *sk, struct page *page,
 582                         int offset, size_t size, int flags)
 583 {
 584         struct tls_context *tls_ctx = tls_get_ctx(sk);
 585         union tls_iter_offset iter_offset;
 586         struct iov_iter msg_iter;
 587         char *kaddr;
 588         struct kvec iov;
 589         int rc;
 590
 591         if (flags & MSG_SENDPAGE_NOTLAST)
 592                 flags |= MSG_MORE;
 593
 594         mutex_lock(&tls_ctx->tx_lock);
 595         lock_sock(sk);
 596
 597         if (flags & MSG_OOB) {
 598                 rc = -EOPNOTSUPP;
 599                 goto out;
 600         }
 601
 602         if (tls_ctx->zerocopy_sendfile) {
 603                 iter_offset.offset = offset;
 604                 rc = tls_push_data(sk, iter_offset, size,
 605                                    flags, TLS_RECORD_TYPE_DATA, page);
 606                 goto out;
 607         }
 608
 609         kaddr = kmap(page);
 610         iov.iov_base = kaddr + offset;
 611         iov.iov_len = size;
 612         iov_iter_kvec(&msg_iter, WRITE, &iov, 1, size);
 613         iter_offset.msg_iter = &msg_iter;
 614         rc = tls_push_data(sk, iter_offset, size, flags, TLS_RECORD_TYPE_DATA,
 615                            NULL);
 616         kunmap(page);
 617
 618 out:
 619         release_sock(sk);
 620         mutex_unlock(&tls_ctx->tx_lock);
 621         return rc;
 622 }
 623
 624 struct tls_record_info *tls_get_record(struct tls_offload_context_tx *context,
 625                                        u32 seq, u64 *p_record_sn)
 626 {
 627         u64 record_sn = context->hint_record_sn;
 628         struct tls_record_info *info, *last;
 629
 630         info = context->retransmit_hint;
 631         if (!info ||
 632             before(seq, info->end_seq - info->len)) {
 633                 /* if retransmit_hint is irrelevant start
 634                  * from the beginning of the list
 635                  */
 636                 info = list_first_entry_or_null(&context->records_list,
 637                                                 struct tls_record_info, list);
 638                 if (!info)
 639                         return NULL;
 640                 /* send the start_marker record if seq number is before the
 641                  * tls offload start marker sequence number. This record is
 642                  * required to handle TCP packets which are before TLS offload
 643                  * started.
 644                  *  And if it's not start marker, look if this seq number
 645                  * belongs to the list.
 646                  */
 647                 if (likely(!tls_record_is_start_marker(info))) {
 648                         /* we have the first record, get the last record to see
 649                          * if this seq number belongs to the list.
 650                          */
 651                         last = list_last_entry(&context->records_list,
 652                                                struct tls_record_info, list);
 653
 654                         if (!between(seq, tls_record_start_seq(info),
 655                                      last->end_seq))
 656                                 return NULL;
 657                 }
 658                 record_sn = context->unacked_record_sn;
 659         }
 660
 661         /* We just need the _rcu for the READ_ONCE() */
 662         rcu_read_lock();
 663         list_for_each_entry_from_rcu(info, &context->records_list, list) {
 664                 if (before(seq, info->end_seq)) {
 665                         if (!context->retransmit_hint ||
 666                             after(info->end_seq,
 667                                   context->retransmit_hint->end_seq)) {
 668                                 context->hint_record_sn = record_sn;
 669                                 context->retransmit_hint = info;
 670                         }
 671                         *p_record_sn = record_sn;
 672                         goto exit_rcu_unlock;
 673                 }
 674                 record_sn++;
 675         }
 676         info = NULL;
 677
 678 exit_rcu_unlock:
 679         rcu_read_unlock();
 680         return info;
 681 }
 682 EXPORT_SYMBOL(tls_get_record);
 683
 684 static int tls_device_push_pending_record(struct sock *sk, int flags)
 685 {
 686         union tls_iter_offset iter;
 687         struct iov_iter msg_iter;
 688
 689         iov_iter_kvec(&msg_iter, WRITE, NULL, 0, 0);
 690         iter.msg_iter = &msg_iter;
 691         return tls_push_data(sk, iter, 0, flags, TLS_RECORD_TYPE_DATA, NULL);
 692 }
 693
 694 void tls_device_write_space(struct sock *sk, struct tls_context *ctx)
 695 {
 696         if (tls_is_partially_sent_record(ctx)) {
 697                 gfp_t sk_allocation = sk->sk_allocation;
 698
 699                 WARN_ON_ONCE(sk->sk_write_pending);
 700
 701                 sk->sk_allocation = GFP_ATOMIC;
 702                 tls_push_partial_record(sk, ctx,
 703                                         MSG_DONTWAIT | MSG_NOSIGNAL |
 704                                         MSG_SENDPAGE_DECRYPTED);
 705                 sk->sk_allocation = sk_allocation;
 706         }
 707 }
 708
 709 static void tls_device_resync_rx(struct tls_context *tls_ctx,
 710                                  struct sock *sk, u32 seq, u8 *rcd_sn)
 711 {
 712         struct tls_offload_context_rx *rx_ctx = tls_offload_ctx_rx(tls_ctx);
 713         struct net_device *netdev;
 714
 715         trace_tls_device_rx_resync_send(sk, seq, rcd_sn, rx_ctx->resync_type);
 716         rcu_read_lock();
 717         netdev = READ_ONCE(tls_ctx->netdev);
 718         if (netdev)
 719                 netdev->tlsdev_ops->tls_dev_resync(netdev, sk, seq, rcd_sn,
 720                                                    TLS_OFFLOAD_CTX_DIR_RX);
 721         rcu_read_unlock();
 722         TLS_INC_STATS(sock_net(sk), LINUX_MIB_TLSRXDEVICERESYNC);
 723 }
 724
 725 static bool
 726 tls_device_rx_resync_async(struct tls_offload_resync_async *resync_async,
 727                            s64 resync_req, u32 *seq, u16 *rcd_delta)
 728 {
 729         u32 is_async = resync_req & RESYNC_REQ_ASYNC;
 730         u32 req_seq = resync_req >> 32;
 731         u32 req_end = req_seq + ((resync_req >> 16) & 0xffff);
 732         u16 i;
 733
 734         *rcd_delta = 0;
 735
 736         if (is_async) {
 737                 /* shouldn't get to wraparound:
 738                  * too long in async stage, something bad happened
 739                  */
 740                 if (WARN_ON_ONCE(resync_async->rcd_delta == USHRT_MAX))
 741                         return false;
 742
 743                 /* asynchronous stage: log all headers seq such that
 744                  * req_seq <= seq <= end_seq, and wait for real resync request
 745                  */
 746                 if (before(*seq, req_seq))
 747                         return false;
 748                 if (!after(*seq, req_end) &&
 749                     resync_async->loglen < TLS_DEVICE_RESYNC_ASYNC_LOGMAX)
 750                         resync_async->log[resync_async->loglen++] = *seq;
 751
 752                 resync_async->rcd_delta++;
 753
 754                 return false;
 755         }
 756
 757         /* synchronous stage: check against the logged entries and
 758          * proceed to check the next entries if no match was found
 759          */
 760         for (i = 0; i < resync_async->loglen; i++)
 761                 if (req_seq == resync_async->log[i] &&
 762                     atomic64_try_cmpxchg(&resync_async->req, &resync_req, 0)) {
 763                         *rcd_delta = resync_async->rcd_delta - i;
 764                         *seq = req_seq;
 765                         resync_async->loglen = 0;
 766                         resync_async->rcd_delta = 0;
 767                         return true;
 768                 }
 769
 770         resync_async->loglen = 0;
 771         resync_async->rcd_delta = 0;
 772
 773         if (req_seq == *seq &&
 774             atomic64_try_cmpxchg(&resync_async->req,
 775                                  &resync_req, 0))
 776                 return true;
 777
 778         return false;
 779 }
 780
 781 void tls_device_rx_resync_new_rec(struct sock *sk, u32 rcd_len, u32 seq)
 782 {
 783         struct tls_context *tls_ctx = tls_get_ctx(sk);
 784         struct tls_offload_context_rx *rx_ctx;
 785         u8 rcd_sn[TLS_MAX_REC_SEQ_SIZE];
 786         u32 sock_data, is_req_pending;
 787         struct tls_prot_info *prot;
 788         s64 resync_req;
 789         u16 rcd_delta;
 790         u32 req_seq;
 791
 792         if (tls_ctx->rx_conf != TLS_HW)
 793                 return;
 794         if (unlikely(test_bit(TLS_RX_DEV_DEGRADED, &tls_ctx->flags)))
 795                 return;
 796
 797         prot = &tls_ctx->prot_info;
 798         rx_ctx = tls_offload_ctx_rx(tls_ctx);
 799         memcpy(rcd_sn, tls_ctx->rx.rec_seq, prot->rec_seq_size);
 800
 801         switch (rx_ctx->resync_type) {
 802         case TLS_OFFLOAD_SYNC_TYPE_DRIVER_REQ:
 803                 resync_req = atomic64_read(&rx_ctx->resync_req);
 804                 req_seq = resync_req >> 32;
 805                 seq += TLS_HEADER_SIZE - 1;
 806                 is_req_pending = resync_req;
 807
 808                 if (likely(!is_req_pending) || req_seq != seq ||
 809                     !atomic64_try_cmpxchg(&rx_ctx->resync_req, &resync_req, 0))
 810                         return;
 811                 break;
 812         case TLS_OFFLOAD_SYNC_TYPE_CORE_NEXT_HINT:
 813                 if (likely(!rx_ctx->resync_nh_do_now))
 814                         return;
 815
 816                 /* head of next rec is already in, note that the sock_inq will
 817                  * include the currently parsed message when called from parser
 818                  */
 819                 sock_data = tcp_inq(sk);
 820                 if (sock_data > rcd_len) {
 821                         trace_tls_device_rx_resync_nh_delay(sk, sock_data,
 822                                                             rcd_len);
 823                         return;
 824                 }
 825
 826                 rx_ctx->resync_nh_do_now = 0;
 827                 seq += rcd_len;
 828                 tls_bigint_increment(rcd_sn, prot->rec_seq_size);
 829                 break;
 830         case TLS_OFFLOAD_SYNC_TYPE_DRIVER_REQ_ASYNC:
 831                 resync_req = atomic64_read(&rx_ctx->resync_async->req);
 832                 is_req_pending = resync_req;
 833                 if (likely(!is_req_pending))
 834                         return;
 835
 836                 if (!tls_device_rx_resync_async(rx_ctx->resync_async,
 837                                                 resync_req, &seq, &rcd_delta))
 838                         return;
 839                 tls_bigint_subtract(rcd_sn, rcd_delta);
 840                 break;
 841         }
 842
 843         tls_device_resync_rx(tls_ctx, sk, seq, rcd_sn);
 844 }
 845
 846 static void tls_device_core_ctrl_rx_resync(struct tls_context *tls_ctx,
 847                                            struct tls_offload_context_rx *ctx,
 848                                            struct sock *sk, struct sk_buff *skb)
 849 {
 850         struct strp_msg *rxm;
 851
 852         /* device will request resyncs by itself based on stream scan */
 853         if (ctx->resync_type != TLS_OFFLOAD_SYNC_TYPE_CORE_NEXT_HINT)
 854                 return;
 855         /* already scheduled */
 856         if (ctx->resync_nh_do_now)
 857                 return;
 858         /* seen decrypted fragments since last fully-failed record */
 859         if (ctx->resync_nh_reset) {
 860                 ctx->resync_nh_reset = 0;
 861                 ctx->resync_nh.decrypted_failed = 1;
 862                 ctx->resync_nh.decrypted_tgt = TLS_DEVICE_RESYNC_NH_START_IVAL;
 863                 return;
 864         }
 865
 866         if (++ctx->resync_nh.decrypted_failed <= ctx->resync_nh.decrypted_tgt)
 867                 return;
 868
 869         /* doing resync, bump the next target in case it fails */
 870         if (ctx->resync_nh.decrypted_tgt < TLS_DEVICE_RESYNC_NH_MAX_IVAL)
 871                 ctx->resync_nh.decrypted_tgt *= 2;
 872         else
 873                 ctx->resync_nh.decrypted_tgt += TLS_DEVICE_RESYNC_NH_MAX_IVAL;
 874
 875         rxm = strp_msg(skb);
 876
 877         /* head of next rec is already in, parser will sync for us */
 878         if (tcp_inq(sk) > rxm->full_len) {
 879                 trace_tls_device_rx_resync_nh_schedule(sk);
 880                 ctx->resync_nh_do_now = 1;
 881         } else {
 882                 struct tls_prot_info *prot = &tls_ctx->prot_info;
 883                 u8 rcd_sn[TLS_MAX_REC_SEQ_SIZE];
 884
 885                 memcpy(rcd_sn, tls_ctx->rx.rec_seq, prot->rec_seq_size);
 886                 tls_bigint_increment(rcd_sn, prot->rec_seq_size);
 887
 888                 tls_device_resync_rx(tls_ctx, sk, tcp_sk(sk)->copied_seq,
 889                                      rcd_sn);
 890         }
 891 }
 892
 893 static int tls_device_reencrypt(struct sock *sk, struct sk_buff *skb)
 894 {
 895         struct strp_msg *rxm = strp_msg(skb);
 896         int err = 0, offset = rxm->offset, copy, nsg, data_len, pos;
 897         struct sk_buff *skb_iter, *unused;
 898         struct scatterlist sg[1];
 899         char *orig_buf, *buf;
 900
 901         orig_buf = kmalloc(rxm->full_len + TLS_HEADER_SIZE +
 902                            TLS_CIPHER_AES_GCM_128_IV_SIZE, sk->sk_allocation);
 903         if (!orig_buf)
 904                 return -ENOMEM;
 905         buf = orig_buf;
 906
 907         nsg = skb_cow_data(skb, 0, &unused);
 908         if (unlikely(nsg < 0)) {
 909                 err = nsg;
 910                 goto free_buf;
 911         }
 912
 913         sg_init_table(sg, 1);
 914         sg_set_buf(&sg[0], buf,
 915                    rxm->full_len + TLS_HEADER_SIZE +
 916                    TLS_CIPHER_AES_GCM_128_IV_SIZE);
 917         err = skb_copy_bits(skb, offset, buf,
 918                             TLS_HEADER_SIZE + TLS_CIPHER_AES_GCM_128_IV_SIZE);
 919         if (err)
 920                 goto free_buf;
 921
 922         /* We are interested only in the decrypted data not the auth */
 923         err = decrypt_skb(sk, skb, sg);
 924         if (err != -EBADMSG)
 925                 goto free_buf;
 926         else
 927                 err = 0;
 928
 929         data_len = rxm->full_len - TLS_CIPHER_AES_GCM_128_TAG_SIZE;
 930
 931         if (skb_pagelen(skb) > offset) {
 932                 copy = min_t(int, skb_pagelen(skb) - offset, data_len);
 933
 934                 if (skb->decrypted) {
 935                         err = skb_store_bits(skb, offset, buf, copy);
 936                         if (err)
 937                                 goto free_buf;
 938                 }
 939
 940                 offset += copy;
 941                 buf += copy;
 942         }
 943
 944         pos = skb_pagelen(skb);
 945         skb_walk_frags(skb, skb_iter) {
 946                 int frag_pos;
 947
 948                 /* Practically all frags must belong to msg if reencrypt
 949                  * is needed with current strparser and coalescing logic,
 950                  * but strparser may "get optimized", so let's be safe.
 951                  */
 952                 if (pos + skb_iter->len <= offset)
 953                         goto done_with_frag;
 954                 if (pos >= data_len + rxm->offset)
 955                         break;
 956
 957                 frag_pos = offset - pos;
 958                 copy = min_t(int, skb_iter->len - frag_pos,
 959                              data_len + rxm->offset - offset);
 960
 961                 if (skb_iter->decrypted) {
 962                         err = skb_store_bits(skb_iter, frag_pos, buf, copy);
 963                         if (err)
 964                                 goto free_buf;
 965                 }
 966
 967                 offset += copy;
 968                 buf += copy;
 969 done_with_frag:
 970                 pos += skb_iter->len;
 971         }
 972
 973 free_buf:
 974         kfree(orig_buf);
 975         return err;
 976 }
 977
 978 int tls_device_decrypted(struct sock *sk, struct tls_context *tls_ctx,
 979                          struct sk_buff *skb, struct strp_msg *rxm)
 980 {
 981         struct tls_offload_context_rx *ctx = tls_offload_ctx_rx(tls_ctx);
 982         int is_decrypted = skb->decrypted;
 983         int is_encrypted = !is_decrypted;
 984         struct sk_buff *skb_iter;
 985
 986         /* Check if all the data is decrypted already */
 987         skb_walk_frags(skb, skb_iter) {
 988                 is_decrypted &= skb_iter->decrypted;
 989                 is_encrypted &= !skb_iter->decrypted;
 990         }
 991
 992         trace_tls_device_decrypted(sk, tcp_sk(sk)->copied_seq - rxm->full_len,
 993                                    tls_ctx->rx.rec_seq, rxm->full_len,
 994                                    is_encrypted, is_decrypted);
 995
 996         if (unlikely(test_bit(TLS_RX_DEV_DEGRADED, &tls_ctx->flags))) {
 997                 if (likely(is_encrypted || is_decrypted))
 998                         return is_decrypted;
 999
1000                 /* After tls_device_down disables the offload, the next SKB will
1001                  * likely have initial fragments decrypted, and final ones not
1002                  * decrypted. We need to reencrypt that single SKB.
1003                  */
1004                 return tls_device_reencrypt(sk, skb);
1005         }
1006
1007         /* Return immediately if the record is either entirely plaintext or
1008          * entirely ciphertext. Otherwise handle reencrypt partially decrypted
1009          * record.
1010          */
1011         if (is_decrypted) {
1012                 ctx->resync_nh_reset = 1;
1013                 return is_decrypted;
1014         }
1015         if (is_encrypted) {
1016                 tls_device_core_ctrl_rx_resync(tls_ctx, ctx, sk, skb);
1017                 return 0;
1018         }
1019
1020         ctx->resync_nh_reset = 1;
1021         return tls_device_reencrypt(sk, skb);
1022 }
1023
1024 static void tls_device_attach(struct tls_context *ctx, struct sock *sk,
1025                               struct net_device *netdev)
1026 {
1027         if (sk->sk_destruct != tls_device_sk_destruct) {
1028                 refcount_set(&ctx->refcount, 1);
1029                 dev_hold(netdev);
1030                 ctx->netdev = netdev;
1031                 spin_lock_irq(&tls_device_lock);
1032                 list_add_tail(&ctx->list, &tls_device_list);
1033                 spin_unlock_irq(&tls_device_lock);
1034
1035                 ctx->sk_destruct = sk->sk_destruct;
1036                 smp_store_release(&sk->sk_destruct, tls_device_sk_destruct);
1037         }
1038 }
1039
1040 int tls_set_device_offload(struct sock *sk, struct tls_context *ctx)
1041 {
1042         u16 nonce_size, tag_size, iv_size, rec_seq_size, salt_size;
1043         struct tls_context *tls_ctx = tls_get_ctx(sk);
1044         struct tls_prot_info *prot = &tls_ctx->prot_info;
1045         struct tls_record_info *start_marker_record;
1046         struct tls_offload_context_tx *offload_ctx;
1047         struct tls_crypto_info *crypto_info;
1048         struct net_device *netdev;
1049         char *iv, *rec_seq;
1050         struct sk_buff *skb;
1051         __be64 rcd_sn;
1052         int rc;
1053
1054         if (!ctx)
1055                 return -EINVAL;
1056
1057         if (ctx->priv_ctx_tx)
1058                 return -EEXIST;
1059
1060         netdev = get_netdev_for_sock(sk);
1061         if (!netdev) {
1062                 pr_err_ratelimited("%s: netdev not found\n", __func__);
1063                 return -EINVAL;
1064         }
1065
1066         if (!(netdev->features & NETIF_F_HW_TLS_TX)) {
1067                 rc = -EOPNOTSUPP;
1068                 goto release_netdev;
1069         }
1070
1071         crypto_info = &ctx->crypto_send.info;
1072         if (crypto_info->version != TLS_1_2_VERSION) {
1073                 rc = -EOPNOTSUPP;
1074                 goto release_netdev;
1075         }
1076
1077         switch (crypto_info->cipher_type) {
1078         case TLS_CIPHER_AES_GCM_128:
1079                 nonce_size = TLS_CIPHER_AES_GCM_128_IV_SIZE;
1080                 tag_size = TLS_CIPHER_AES_GCM_128_TAG_SIZE;
1081                 iv_size = TLS_CIPHER_AES_GCM_128_IV_SIZE;
1082                 iv = ((struct tls12_crypto_info_aes_gcm_128 *)crypto_info)->iv;
1083                 rec_seq_size = TLS_CIPHER_AES_GCM_128_REC_SEQ_SIZE;
1084                 salt_size = TLS_CIPHER_AES_GCM_128_SALT_SIZE;
1085                 rec_seq =
1086                  ((struct tls12_crypto_info_aes_gcm_128 *)crypto_info)->rec_seq;
1087                 break;
1088         default:
1089                 rc = -EINVAL;
1090                 goto release_netdev;
1091         }
1092
1093         /* Sanity-check the rec_seq_size for stack allocations */
1094         if (rec_seq_size > TLS_MAX_REC_SEQ_SIZE) {
1095                 rc = -EINVAL;
1096                 goto release_netdev;
1097         }
1098
1099         prot->version = crypto_info->version;
1100         prot->cipher_type = crypto_info->cipher_type;
1101         prot->prepend_size = TLS_HEADER_SIZE + nonce_size;
1102         prot->tag_size = tag_size;
1103         prot->overhead_size = prot->prepend_size + prot->tag_size;
1104         prot->iv_size = iv_size;
1105         prot->salt_size = salt_size;
1106         ctx->tx.iv = kmalloc(iv_size + TLS_CIPHER_AES_GCM_128_SALT_SIZE,
1107                              GFP_KERNEL);
1108         if (!ctx->tx.iv) {
1109                 rc = -ENOMEM;
1110                 goto release_netdev;
1111         }
1112
1113         memcpy(ctx->tx.iv + TLS_CIPHER_AES_GCM_128_SALT_SIZE, iv, iv_size);
1114
1115         prot->rec_seq_size = rec_seq_size;
1116         ctx->tx.rec_seq = kmemdup(rec_seq, rec_seq_size, GFP_KERNEL);
1117         if (!ctx->tx.rec_seq) {
1118                 rc = -ENOMEM;
1119                 goto free_iv;
1120         }
1121
1122         start_marker_record = kmalloc(sizeof(*start_marker_record), GFP_KERNEL);
1123         if (!start_marker_record) {
1124                 rc = -ENOMEM;
1125                 goto free_rec_seq;
1126         }
1127
1128         offload_ctx = kzalloc(TLS_OFFLOAD_CONTEXT_SIZE_TX, GFP_KERNEL);
1129         if (!offload_ctx) {
1130                 rc = -ENOMEM;
1131                 goto free_marker_record;
1132         }
1133
1134         rc = tls_sw_fallback_init(sk, offload_ctx, crypto_info);
1135         if (rc)
1136                 goto free_offload_ctx;
1137
1138         /* start at rec_seq - 1 to account for the start marker record */
1139         memcpy(&rcd_sn, ctx->tx.rec_seq, sizeof(rcd_sn));
1140         offload_ctx->unacked_record_sn = be64_to_cpu(rcd_sn) - 1;
1141
1142         start_marker_record->end_seq = tcp_sk(sk)->write_seq;
1143         start_marker_record->len = 0;
1144         start_marker_record->num_frags = 0;
1145
1146         INIT_LIST_HEAD(&offload_ctx->records_list);
1147         list_add_tail(&start_marker_record->list, &offload_ctx->records_list);
1148         spin_lock_init(&offload_ctx->lock);
1149         sg_init_table(offload_ctx->sg_tx_data,
1150                       ARRAY_SIZE(offload_ctx->sg_tx_data));
1151
1152         clean_acked_data_enable(inet_csk(sk), &tls_icsk_clean_acked);
1153         ctx->push_pending_record = tls_device_push_pending_record;
1154
1155         /* TLS offload is greatly simplified if we don't send
1156          * SKBs where only part of the payload needs to be encrypted.
1157          * So mark the last skb in the write queue as end of record.
1158          */
1159         skb = tcp_write_queue_tail(sk);
1160         if (skb)
1161                 TCP_SKB_CB(skb)->eor = 1;
1162
1163         /* Avoid offloading if the device is down
1164          * We don't want to offload new flows after
1165          * the NETDEV_DOWN event
1166          *
1167          * device_offload_lock is taken in tls_devices's NETDEV_DOWN
1168          * handler thus protecting from the device going down before
1169          * ctx was added to tls_device_list.
1170          */
1171         down_read(&device_offload_lock);
1172         if (!(netdev->flags & IFF_UP)) {
1173                 rc = -EINVAL;
1174                 goto release_lock;
1175         }
1176
1177         ctx->priv_ctx_tx = offload_ctx;
1178         rc = netdev->tlsdev_ops->tls_dev_add(netdev, sk, TLS_OFFLOAD_CTX_DIR_TX,
1179                                              &ctx->crypto_send.info,
1180                                              tcp_sk(sk)->write_seq);
1181         trace_tls_device_offload_set(sk, TLS_OFFLOAD_CTX_DIR_TX,
1182                                      tcp_sk(sk)->write_seq, rec_seq, rc);
1183         if (rc)
1184                 goto release_lock;
1185
1186         tls_device_attach(ctx, sk, netdev);
1187         up_read(&device_offload_lock);
1188
1189         /* following this assignment tls_is_sk_tx_device_offloaded
1190          * will return true and the context might be accessed
1191          * by the netdev's xmit function.
1192          */
1193         smp_store_release(&sk->sk_validate_xmit_skb, tls_validate_xmit_skb);
1194         dev_put(netdev);
1195
1196         return 0;
1197
1198 release_lock:
1199         up_read(&device_offload_lock);
1200         clean_acked_data_disable(inet_csk(sk));
1201         crypto_free_aead(offload_ctx->aead_send);
1202 free_offload_ctx:
1203         kfree(offload_ctx);
1204         ctx->priv_ctx_tx = NULL;
1205 free_marker_record:
1206         kfree(start_marker_record);
1207 free_rec_seq:
1208         kfree(ctx->tx.rec_seq);
1209 free_iv:
1210         kfree(ctx->tx.iv);
1211 release_netdev:
1212         dev_put(netdev);
1213         return rc;
1214 }
1215
1216 int tls_set_device_offload_rx(struct sock *sk, struct tls_context *ctx)
1217 {
1218         struct tls12_crypto_info_aes_gcm_128 *info;
1219         struct tls_offload_context_rx *context;
1220         struct net_device *netdev;
1221         int rc = 0;
1222
1223         if (ctx->crypto_recv.info.version != TLS_1_2_VERSION)
1224                 return -EOPNOTSUPP;
1225
1226         netdev = get_netdev_for_sock(sk);
1227         if (!netdev) {
1228                 pr_err_ratelimited("%s: netdev not found\n", __func__);
1229                 return -EINVAL;
1230         }
1231
1232         if (!(netdev->features & NETIF_F_HW_TLS_RX)) {
1233                 rc = -EOPNOTSUPP;
1234                 goto release_netdev;
1235         }
1236
1237         /* Avoid offloading if the device is down
1238          * We don't want to offload new flows after
1239          * the NETDEV_DOWN event
1240          *
1241          * device_offload_lock is taken in tls_devices's NETDEV_DOWN
1242          * handler thus protecting from the device going down before
1243          * ctx was added to tls_device_list.
1244          */
1245         down_read(&device_offload_lock);
1246         if (!(netdev->flags & IFF_UP)) {
1247                 rc = -EINVAL;
1248                 goto release_lock;
1249         }
1250
1251         context = kzalloc(TLS_OFFLOAD_CONTEXT_SIZE_RX, GFP_KERNEL);
1252         if (!context) {
1253                 rc = -ENOMEM;
1254                 goto release_lock;
1255         }
1256         context->resync_nh_reset = 1;
1257
1258         ctx->priv_ctx_rx = context;
1259         rc = tls_set_sw_offload(sk, ctx, 0);
1260         if (rc)
1261                 goto release_ctx;
1262
1263         rc = netdev->tlsdev_ops->tls_dev_add(netdev, sk, TLS_OFFLOAD_CTX_DIR_RX,
1264                                              &ctx->crypto_recv.info,
1265                                              tcp_sk(sk)->copied_seq);
1266         info = (void *)&ctx->crypto_recv.info;
1267         trace_tls_device_offload_set(sk, TLS_OFFLOAD_CTX_DIR_RX,
1268                                      tcp_sk(sk)->copied_seq, info->rec_seq, rc);
1269         if (rc)
1270                 goto free_sw_resources;
1271
1272         tls_device_attach(ctx, sk, netdev);
1273         up_read(&device_offload_lock);
1274
1275         dev_put(netdev);
1276
1277         return 0;
1278
1279 free_sw_resources:
1280         up_read(&device_offload_lock);
1281         tls_sw_free_resources_rx(sk);
1282         down_read(&device_offload_lock);
1283 release_ctx:
1284         ctx->priv_ctx_rx = NULL;
1285 release_lock:
1286         up_read(&device_offload_lock);
1287 release_netdev:
1288         dev_put(netdev);
1289         return rc;
1290 }
1291
1292 void tls_device_offload_cleanup_rx(struct sock *sk)
1293 {
1294         struct tls_context *tls_ctx = tls_get_ctx(sk);
1295         struct net_device *netdev;
1296
1297         down_read(&device_offload_lock);
1298         netdev = tls_ctx->netdev;
1299         if (!netdev)
1300                 goto out;
1301
1302         netdev->tlsdev_ops->tls_dev_del(netdev, tls_ctx,
1303                                         TLS_OFFLOAD_CTX_DIR_RX);
1304
1305         if (tls_ctx->tx_conf != TLS_HW) {
1306                 dev_put(netdev);
1307                 tls_ctx->netdev = NULL;
1308         } else {
1309                 set_bit(TLS_RX_DEV_CLOSED, &tls_ctx->flags);
1310         }
1311 out:
1312         up_read(&device_offload_lock);
1313         tls_sw_release_resources_rx(sk);
1314 }
1315
1316 static int tls_device_down(struct net_device *netdev)
1317 {
1318         struct tls_context *ctx, *tmp;
1319         unsigned long flags;
1320         LIST_HEAD(list);
1321
1322         /* Request a write lock to block new offload attempts */
1323         down_write(&device_offload_lock);
1324
1325         spin_lock_irqsave(&tls_device_lock, flags);
1326         list_for_each_entry_safe(ctx, tmp, &tls_device_list, list) {
1327                 if (ctx->netdev != netdev ||
1328                     !refcount_inc_not_zero(&ctx->refcount))
1329                         continue;
1330
1331                 list_move(&ctx->list, &list);
1332         }
1333         spin_unlock_irqrestore(&tls_device_lock, flags);
1334
1335         list_for_each_entry_safe(ctx, tmp, &list, list) {
1336                 /* Stop offloaded TX and switch to the fallback.
1337                  * tls_is_sk_tx_device_offloaded will return false.
1338                  */
1339                 WRITE_ONCE(ctx->sk->sk_validate_xmit_skb, tls_validate_xmit_skb_sw);
1340
1341                 /* Stop the RX and TX resync.
1342                  * tls_dev_resync must not be called after tls_dev_del.
1343                  */
1344                 WRITE_ONCE(ctx->netdev, NULL);
1345
1346                 /* Start skipping the RX resync logic completely. */
1347                 set_bit(TLS_RX_DEV_DEGRADED, &ctx->flags);
1348
1349                 /* Sync with inflight packets. After this point:
1350                  * TX: no non-encrypted packets will be passed to the driver.
1351                  * RX: resync requests from the driver will be ignored.
1352                  */
1353                 synchronize_net();
1354
1355                 /* Release the offload context on the driver side. */
1356                 if (ctx->tx_conf == TLS_HW)
1357                         netdev->tlsdev_ops->tls_dev_del(netdev, ctx,
1358                                                         TLS_OFFLOAD_CTX_DIR_TX);
1359                 if (ctx->rx_conf == TLS_HW &&
1360                     !test_bit(TLS_RX_DEV_CLOSED, &ctx->flags))
1361                         netdev->tlsdev_ops->tls_dev_del(netdev, ctx,
1362                                                         TLS_OFFLOAD_CTX_DIR_RX);
1363
1364                 dev_put(netdev);
1365
1366                 /* Move the context to a separate list for two reasons:
1367                  * 1. When the context is deallocated, list_del is called.
1368                  * 2. It's no longer an offloaded context, so we don't want to
1369                  *    run offload-specific code on this context.
1370                  */
1371                 spin_lock_irqsave(&tls_device_lock, flags);
1372                 list_move_tail(&ctx->list, &tls_device_down_list);
1373                 spin_unlock_irqrestore(&tls_device_lock, flags);
1374
1375                 /* Device contexts for RX and TX will be freed in on sk_destruct
1376                  * by tls_device_free_ctx. rx_conf and tx_conf stay in TLS_HW.
1377                  * Now release the ref taken above.
1378                  */
1379                 if (refcount_dec_and_test(&ctx->refcount)) {
1380                         /* sk_destruct ran after tls_device_down took a ref, and
1381                          * it returned early. Complete the destruction here.
1382                          */
1383                         list_del(&ctx->list);
1384                         tls_device_free_ctx(ctx);
1385                 }
1386         }
1387
1388         up_write(&device_offload_lock);
1389
1390         flush_work(&tls_device_gc_work);
1391
1392         return NOTIFY_DONE;
1393 }
1394
1395 static int tls_dev_event(struct notifier_block *this, unsigned long event,
1396                          void *ptr)
1397 {
1398         struct net_device *dev = netdev_notifier_info_to_dev(ptr);
1399
1400         if (!dev->tlsdev_ops &&
1401             !(dev->features & (NETIF_F_HW_TLS_RX | NETIF_F_HW_TLS_TX)))
1402                 return NOTIFY_DONE;
1403
1404         switch (event) {
1405         case NETDEV_REGISTER:
1406         case NETDEV_FEAT_CHANGE:
1407                 if (netif_is_bond_master(dev))
1408                         return NOTIFY_DONE;
1409                 if ((dev->features & NETIF_F_HW_TLS_RX) &&
1410                     !dev->tlsdev_ops->tls_dev_resync)
1411                         return NOTIFY_BAD;
1412
1413                 if  (dev->tlsdev_ops &&
1414                      dev->tlsdev_ops->tls_dev_add &&
1415                      dev->tlsdev_ops->tls_dev_del)
1416                         return NOTIFY_DONE;
1417                 else
1418                         return NOTIFY_BAD;
1419         case NETDEV_DOWN:
1420                 return tls_device_down(dev);
1421         }
1422         return NOTIFY_DONE;
1423 }
1424
1425 static struct notifier_block tls_dev_notifier = {
1426         .notifier_call  = tls_dev_event,
1427 };
1428
1429 int __init tls_device_init(void)
1430 {
1431         return register_netdevice_notifier(&tls_dev_notifier);
1432 }
1433
1434 void __exit tls_device_cleanup(void)
1435 {
1436         unregister_netdevice_notifier(&tls_dev_notifier);
1437         flush_work(&tls_device_gc_work);
1438         clean_acked_data_flush();
1439 }