net/tls/tls_device.c

   1 /* Copyright (c) 2018, Mellanox Technologies All rights reserved.
   2  *
   3  * This software is available to you under a choice of one of two
   4  * licenses.  You may choose to be licensed under the terms of the GNU
   5  * General Public License (GPL) Version 2, available from the file
   6  * COPYING in the main directory of this source tree, or the
   7  * OpenIB.org BSD license below:
   8  *
   9  *     Redistribution and use in source and binary forms, with or
  10  *     without modification, are permitted provided that the following
  11  *     conditions are met:
  12  *
  13  *      - Redistributions of source code must retain the above
  14  *        copyright notice, this list of conditions and the following
  15  *        disclaimer.
  16  *
  17  *      - Redistributions in binary form must reproduce the above
  18  *        copyright notice, this list of conditions and the following
  19  *        disclaimer in the documentation and/or other materials
  20  *        provided with the distribution.
  21  *
  22  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  23  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  24  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  25  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  26  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  27  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  28  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  29  * SOFTWARE.
  30  */
  31
  32 #include <crypto/aead.h>
  33 #include <linux/highmem.h>
  34 #include <linux/module.h>
  35 #include <linux/netdevice.h>
  36 #include <net/dst.h>
  37 #include <net/inet_connection_sock.h>
  38 #include <net/tcp.h>
  39 #include <net/tls.h>
  40
  41 #include "tls.h"
  42 #include "trace.h"
  43
  44 /* device_offload_lock is used to synchronize tls_dev_add
  45  * against NETDEV_DOWN notifications.
  46  */
  47 static DECLARE_RWSEM(device_offload_lock);
  48
  49 static struct workqueue_struct *destruct_wq __read_mostly;
  50
  51 static LIST_HEAD(tls_device_list);
  52 static LIST_HEAD(tls_device_down_list);
  53 static DEFINE_SPINLOCK(tls_device_lock);
  54
  55 static struct page *dummy_page;
  56
  57 static void tls_device_free_ctx(struct tls_context *ctx)
  58 {
  59         if (ctx->tx_conf == TLS_HW) {
  60                 kfree(tls_offload_ctx_tx(ctx));
  61                 kfree(ctx->tx.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_tx_del_task(struct work_struct *work)
  72 {
  73         struct tls_offload_context_tx *offload_ctx =
  74                 container_of(work, struct tls_offload_context_tx, destruct_work);
  75         struct tls_context *ctx = offload_ctx->ctx;
  76         struct net_device *netdev;
  77
  78         /* Safe, because this is the destroy flow, refcount is 0, so
  79          * tls_device_down can't store this field in parallel.
  80          */
  81         netdev = rcu_dereference_protected(ctx->netdev,
  82                                            !refcount_read(&ctx->refcount));
  83
  84         netdev->tlsdev_ops->tls_dev_del(netdev, ctx, TLS_OFFLOAD_CTX_DIR_TX);
  85         dev_put(netdev);
  86         ctx->netdev = NULL;
  87         tls_device_free_ctx(ctx);
  88 }
  89
  90 static void tls_device_queue_ctx_destruction(struct tls_context *ctx)
  91 {
  92         struct net_device *netdev;
  93         unsigned long flags;
  94         bool async_cleanup;
  95
  96         spin_lock_irqsave(&tls_device_lock, flags);
  97         if (unlikely(!refcount_dec_and_test(&ctx->refcount))) {
  98                 spin_unlock_irqrestore(&tls_device_lock, flags);
  99                 return;
 100         }
 101
 102         list_del(&ctx->list); /* Remove from tls_device_list / tls_device_down_list */
 103
 104         /* Safe, because this is the destroy flow, refcount is 0, so
 105          * tls_device_down can't store this field in parallel.
 106          */
 107         netdev = rcu_dereference_protected(ctx->netdev,
 108                                            !refcount_read(&ctx->refcount));
 109
 110         async_cleanup = netdev && ctx->tx_conf == TLS_HW;
 111         if (async_cleanup) {
 112                 struct tls_offload_context_tx *offload_ctx = tls_offload_ctx_tx(ctx);
 113
 114                 /* queue_work inside the spinlock
 115                  * to make sure tls_device_down waits for that work.
 116                  */
 117                 queue_work(destruct_wq, &offload_ctx->destruct_work);
 118         }
 119         spin_unlock_irqrestore(&tls_device_lock, flags);
 120
 121         if (!async_cleanup)
 122                 tls_device_free_ctx(ctx);
 123 }
 124
 125 /* We assume that the socket is already connected */
 126 static struct net_device *get_netdev_for_sock(struct sock *sk)
 127 {
 128         struct dst_entry *dst = sk_dst_get(sk);
 129         struct net_device *netdev = NULL;
 130
 131         if (likely(dst)) {
 132                 netdev = netdev_sk_get_lowest_dev(dst->dev, sk);
 133                 dev_hold(netdev);
 134         }
 135
 136         dst_release(dst);
 137
 138         return netdev;
 139 }
 140
 141 static void destroy_record(struct tls_record_info *record)
 142 {
 143         int i;
 144
 145         for (i = 0; i < record->num_frags; i++)
 146                 __skb_frag_unref(&record->frags[i], false);
 147         kfree(record);
 148 }
 149
 150 static void delete_all_records(struct tls_offload_context_tx *offload_ctx)
 151 {
 152         struct tls_record_info *info, *temp;
 153
 154         list_for_each_entry_safe(info, temp, &offload_ctx->records_list, list) {
 155                 list_del(&info->list);
 156                 destroy_record(info);
 157         }
 158
 159         offload_ctx->retransmit_hint = NULL;
 160 }
 161
 162 static void tls_icsk_clean_acked(struct sock *sk, u32 acked_seq)
 163 {
 164         struct tls_context *tls_ctx = tls_get_ctx(sk);
 165         struct tls_record_info *info, *temp;
 166         struct tls_offload_context_tx *ctx;
 167         u64 deleted_records = 0;
 168         unsigned long flags;
 169
 170         if (!tls_ctx)
 171                 return;
 172
 173         ctx = tls_offload_ctx_tx(tls_ctx);
 174
 175         spin_lock_irqsave(&ctx->lock, flags);
 176         info = ctx->retransmit_hint;
 177         if (info && !before(acked_seq, info->end_seq))
 178                 ctx->retransmit_hint = NULL;
 179
 180         list_for_each_entry_safe(info, temp, &ctx->records_list, list) {
 181                 if (before(acked_seq, info->end_seq))
 182                         break;
 183                 list_del(&info->list);
 184
 185                 destroy_record(info);
 186                 deleted_records++;
 187         }
 188
 189         ctx->unacked_record_sn += deleted_records;
 190         spin_unlock_irqrestore(&ctx->lock, flags);
 191 }
 192
 193 /* At this point, there should be no references on this
 194  * socket and no in-flight SKBs associated with this
 195  * socket, so it is safe to free all the resources.
 196  */
 197 void tls_device_sk_destruct(struct sock *sk)
 198 {
 199         struct tls_context *tls_ctx = tls_get_ctx(sk);
 200         struct tls_offload_context_tx *ctx = tls_offload_ctx_tx(tls_ctx);
 201
 202         tls_ctx->sk_destruct(sk);
 203
 204         if (tls_ctx->tx_conf == TLS_HW) {
 205                 if (ctx->open_record)
 206                         destroy_record(ctx->open_record);
 207                 delete_all_records(ctx);
 208                 crypto_free_aead(ctx->aead_send);
 209                 clean_acked_data_disable(inet_csk(sk));
 210         }
 211
 212         tls_device_queue_ctx_destruction(tls_ctx);
 213 }
 214 EXPORT_SYMBOL_GPL(tls_device_sk_destruct);
 215
 216 void tls_device_free_resources_tx(struct sock *sk)
 217 {
 218         struct tls_context *tls_ctx = tls_get_ctx(sk);
 219
 220         tls_free_partial_record(sk, tls_ctx);
 221 }
 222
 223 void tls_offload_tx_resync_request(struct sock *sk, u32 got_seq, u32 exp_seq)
 224 {
 225         struct tls_context *tls_ctx = tls_get_ctx(sk);
 226
 227         trace_tls_device_tx_resync_req(sk, got_seq, exp_seq);
 228         WARN_ON(test_and_set_bit(TLS_TX_SYNC_SCHED, &tls_ctx->flags));
 229 }
 230 EXPORT_SYMBOL_GPL(tls_offload_tx_resync_request);
 231
 232 static void tls_device_resync_tx(struct sock *sk, struct tls_context *tls_ctx,
 233                                  u32 seq)
 234 {
 235         struct net_device *netdev;
 236         struct sk_buff *skb;
 237         int err = 0;
 238         u8 *rcd_sn;
 239
 240         skb = tcp_write_queue_tail(sk);
 241         if (skb)
 242                 TCP_SKB_CB(skb)->eor = 1;
 243
 244         rcd_sn = tls_ctx->tx.rec_seq;
 245
 246         trace_tls_device_tx_resync_send(sk, seq, rcd_sn);
 247         down_read(&device_offload_lock);
 248         netdev = rcu_dereference_protected(tls_ctx->netdev,
 249                                            lockdep_is_held(&device_offload_lock));
 250         if (netdev)
 251                 err = netdev->tlsdev_ops->tls_dev_resync(netdev, sk, seq,
 252                                                          rcd_sn,
 253                                                          TLS_OFFLOAD_CTX_DIR_TX);
 254         up_read(&device_offload_lock);
 255         if (err)
 256                 return;
 257
 258         clear_bit_unlock(TLS_TX_SYNC_SCHED, &tls_ctx->flags);
 259 }
 260
 261 static void tls_append_frag(struct tls_record_info *record,
 262                             struct page_frag *pfrag,
 263                             int size)
 264 {
 265         skb_frag_t *frag;
 266
 267         frag = &record->frags[record->num_frags - 1];
 268         if (skb_frag_page(frag) == pfrag->page &&
 269             skb_frag_off(frag) + skb_frag_size(frag) == pfrag->offset) {
 270                 skb_frag_size_add(frag, size);
 271         } else {
 272                 ++frag;
 273                 skb_frag_fill_page_desc(frag, pfrag->page, pfrag->offset,
 274                                         size);
 275                 ++record->num_frags;
 276                 get_page(pfrag->page);
 277         }
 278
 279         pfrag->offset += size;
 280         record->len += size;
 281 }
 282
 283 static int tls_push_record(struct sock *sk,
 284                            struct tls_context *ctx,
 285                            struct tls_offload_context_tx *offload_ctx,
 286                            struct tls_record_info *record,
 287                            int flags)
 288 {
 289         struct tls_prot_info *prot = &ctx->prot_info;
 290         struct tcp_sock *tp = tcp_sk(sk);
 291         skb_frag_t *frag;
 292         int i;
 293
 294         record->end_seq = tp->write_seq + record->len;
 295         list_add_tail_rcu(&record->list, &offload_ctx->records_list);
 296         offload_ctx->open_record = NULL;
 297
 298         if (test_bit(TLS_TX_SYNC_SCHED, &ctx->flags))
 299                 tls_device_resync_tx(sk, ctx, tp->write_seq);
 300
 301         tls_advance_record_sn(sk, prot, &ctx->tx);
 302
 303         for (i = 0; i < record->num_frags; i++) {
 304                 frag = &record->frags[i];
 305                 sg_unmark_end(&offload_ctx->sg_tx_data[i]);
 306                 sg_set_page(&offload_ctx->sg_tx_data[i], skb_frag_page(frag),
 307                             skb_frag_size(frag), skb_frag_off(frag));
 308                 sk_mem_charge(sk, skb_frag_size(frag));
 309                 get_page(skb_frag_page(frag));
 310         }
 311         sg_mark_end(&offload_ctx->sg_tx_data[record->num_frags - 1]);
 312
 313         /* all ready, send */
 314         return tls_push_sg(sk, ctx, offload_ctx->sg_tx_data, 0, flags);
 315 }
 316
 317 static void tls_device_record_close(struct sock *sk,
 318                                     struct tls_context *ctx,
 319                                     struct tls_record_info *record,
 320                                     struct page_frag *pfrag,
 321                                     unsigned char record_type)
 322 {
 323         struct tls_prot_info *prot = &ctx->prot_info;
 324         struct page_frag dummy_tag_frag;
 325
 326         /* append tag
 327          * device will fill in the tag, we just need to append a placeholder
 328          * use socket memory to improve coalescing (re-using a single buffer
 329          * increases frag count)
 330          * if we can't allocate memory now use the dummy page
 331          */
 332         if (unlikely(pfrag->size - pfrag->offset < prot->tag_size) &&
 333             !skb_page_frag_refill(prot->tag_size, pfrag, sk->sk_allocation)) {
 334                 dummy_tag_frag.page = dummy_page;
 335                 dummy_tag_frag.offset = 0;
 336                 pfrag = &dummy_tag_frag;
 337         }
 338         tls_append_frag(record, pfrag, prot->tag_size);
 339
 340         /* fill prepend */
 341         tls_fill_prepend(ctx, skb_frag_address(&record->frags[0]),
 342                          record->len - prot->overhead_size,
 343                          record_type);
 344 }
 345
 346 static int tls_create_new_record(struct tls_offload_context_tx *offload_ctx,
 347                                  struct page_frag *pfrag,
 348                                  size_t prepend_size)
 349 {
 350         struct tls_record_info *record;
 351         skb_frag_t *frag;
 352
 353         record = kmalloc(sizeof(*record), GFP_KERNEL);
 354         if (!record)
 355                 return -ENOMEM;
 356
 357         frag = &record->frags[0];
 358         skb_frag_fill_page_desc(frag, pfrag->page, pfrag->offset,
 359                                 prepend_size);
 360
 361         get_page(pfrag->page);
 362         pfrag->offset += prepend_size;
 363
 364         record->num_frags = 1;
 365         record->len = prepend_size;
 366         offload_ctx->open_record = record;
 367         return 0;
 368 }
 369
 370 static int tls_do_allocation(struct sock *sk,
 371                              struct tls_offload_context_tx *offload_ctx,
 372                              struct page_frag *pfrag,
 373                              size_t prepend_size)
 374 {
 375         int ret;
 376
 377         if (!offload_ctx->open_record) {
 378                 if (unlikely(!skb_page_frag_refill(prepend_size, pfrag,
 379                                                    sk->sk_allocation))) {
 380                         READ_ONCE(sk->sk_prot)->enter_memory_pressure(sk);
 381                         sk_stream_moderate_sndbuf(sk);
 382                         return -ENOMEM;
 383                 }
 384
 385                 ret = tls_create_new_record(offload_ctx, pfrag, prepend_size);
 386                 if (ret)
 387                         return ret;
 388
 389                 if (pfrag->size > pfrag->offset)
 390                         return 0;
 391         }
 392
 393         if (!sk_page_frag_refill(sk, pfrag))
 394                 return -ENOMEM;
 395
 396         return 0;
 397 }
 398
 399 static int tls_device_copy_data(void *addr, size_t bytes, struct iov_iter *i)
 400 {
 401         size_t pre_copy, nocache;
 402
 403         pre_copy = ~((unsigned long)addr - 1) & (SMP_CACHE_BYTES - 1);
 404         if (pre_copy) {
 405                 pre_copy = min(pre_copy, bytes);
 406                 if (copy_from_iter(addr, pre_copy, i) != pre_copy)
 407                         return -EFAULT;
 408                 bytes -= pre_copy;
 409                 addr += pre_copy;
 410         }
 411
 412         nocache = round_down(bytes, SMP_CACHE_BYTES);
 413         if (copy_from_iter_nocache(addr, nocache, i) != nocache)
 414                 return -EFAULT;
 415         bytes -= nocache;
 416         addr += nocache;
 417
 418         if (bytes && copy_from_iter(addr, bytes, i) != bytes)
 419                 return -EFAULT;
 420
 421         return 0;
 422 }
 423
 424 static int tls_push_data(struct sock *sk,
 425                          struct iov_iter *iter,
 426                          size_t size, int flags,
 427                          unsigned char record_type)
 428 {
 429         struct tls_context *tls_ctx = tls_get_ctx(sk);
 430         struct tls_prot_info *prot = &tls_ctx->prot_info;
 431         struct tls_offload_context_tx *ctx = tls_offload_ctx_tx(tls_ctx);
 432         struct tls_record_info *record;
 433         int tls_push_record_flags;
 434         struct page_frag *pfrag;
 435         size_t orig_size = size;
 436         u32 max_open_record_len;
 437         bool more = false;
 438         bool done = false;
 439         int copy, rc = 0;
 440         long timeo;
 441
 442         if (flags &
 443             ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL | MSG_SPLICE_PAGES))
 444                 return -EOPNOTSUPP;
 445
 446         if (unlikely(sk->sk_err))
 447                 return -sk->sk_err;
 448
 449         flags |= MSG_SENDPAGE_DECRYPTED;
 450         tls_push_record_flags = flags | MSG_MORE;
 451
 452         timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
 453         if (tls_is_partially_sent_record(tls_ctx)) {
 454                 rc = tls_push_partial_record(sk, tls_ctx, flags);
 455                 if (rc < 0)
 456                         return rc;
 457         }
 458
 459         pfrag = sk_page_frag(sk);
 460
 461         /* TLS_HEADER_SIZE is not counted as part of the TLS record, and
 462          * we need to leave room for an authentication tag.
 463          */
 464         max_open_record_len = TLS_MAX_PAYLOAD_SIZE +
 465                               prot->prepend_size;
 466         do {
 467                 rc = tls_do_allocation(sk, ctx, pfrag, prot->prepend_size);
 468                 if (unlikely(rc)) {
 469                         rc = sk_stream_wait_memory(sk, &timeo);
 470                         if (!rc)
 471                                 continue;
 472
 473                         record = ctx->open_record;
 474                         if (!record)
 475                                 break;
 476 handle_error:
 477                         if (record_type != TLS_RECORD_TYPE_DATA) {
 478                                 /* avoid sending partial
 479                                  * record with type !=
 480                                  * application_data
 481                                  */
 482                                 size = orig_size;
 483                                 destroy_record(record);
 484                                 ctx->open_record = NULL;
 485                         } else if (record->len > prot->prepend_size) {
 486                                 goto last_record;
 487                         }
 488
 489                         break;
 490                 }
 491
 492                 record = ctx->open_record;
 493
 494                 copy = min_t(size_t, size, max_open_record_len - record->len);
 495                 if (copy && (flags & MSG_SPLICE_PAGES)) {
 496                         struct page_frag zc_pfrag;
 497                         struct page **pages = &zc_pfrag.page;
 498                         size_t off;
 499
 500                         rc = iov_iter_extract_pages(iter, &pages,
 501                                                     copy, 1, 0, &off);
 502                         if (rc <= 0) {
 503                                 if (rc == 0)
 504                                         rc = -EIO;
 505                                 goto handle_error;
 506                         }
 507                         copy = rc;
 508
 509                         if (WARN_ON_ONCE(!sendpage_ok(zc_pfrag.page))) {
 510                                 iov_iter_revert(iter, copy);
 511                                 rc = -EIO;
 512                                 goto handle_error;
 513                         }
 514
 515                         zc_pfrag.offset = off;
 516                         zc_pfrag.size = copy;
 517                         tls_append_frag(record, &zc_pfrag, copy);
 518                 } else if (copy) {
 519                         copy = min_t(size_t, copy, pfrag->size - pfrag->offset);
 520
 521                         rc = tls_device_copy_data(page_address(pfrag->page) +
 522                                                   pfrag->offset, copy,
 523                                                   iter);
 524                         if (rc)
 525                                 goto handle_error;
 526                         tls_append_frag(record, pfrag, copy);
 527                 }
 528
 529                 size -= copy;
 530                 if (!size) {
 531 last_record:
 532                         tls_push_record_flags = flags;
 533                         if (flags & MSG_MORE) {
 534                                 more = true;
 535                                 break;
 536                         }
 537
 538                         done = true;
 539                 }
 540
 541                 if (done || record->len >= max_open_record_len ||
 542                     (record->num_frags >= MAX_SKB_FRAGS - 1)) {
 543                         tls_device_record_close(sk, tls_ctx, record,
 544                                                 pfrag, record_type);
 545
 546                         rc = tls_push_record(sk,
 547                                              tls_ctx,
 548                                              ctx,
 549                                              record,
 550                                              tls_push_record_flags);
 551                         if (rc < 0)
 552                                 break;
 553                 }
 554         } while (!done);
 555
 556         tls_ctx->pending_open_record_frags = more;
 557
 558         if (orig_size - size > 0)
 559                 rc = orig_size - size;
 560
 561         return rc;
 562 }
 563
 564 int tls_device_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
 565 {
 566         unsigned char record_type = TLS_RECORD_TYPE_DATA;
 567         struct tls_context *tls_ctx = tls_get_ctx(sk);
 568         int rc;
 569
 570         if (!tls_ctx->zerocopy_sendfile)
 571                 msg->msg_flags &= ~MSG_SPLICE_PAGES;
 572
 573         mutex_lock(&tls_ctx->tx_lock);
 574         lock_sock(sk);
 575
 576         if (unlikely(msg->msg_controllen)) {
 577                 rc = tls_process_cmsg(sk, msg, &record_type);
 578                 if (rc)
 579                         goto out;
 580         }
 581
 582         rc = tls_push_data(sk, &msg->msg_iter, size, msg->msg_flags,
 583                            record_type);
 584
 585 out:
 586         release_sock(sk);
 587         mutex_unlock(&tls_ctx->tx_lock);
 588         return rc;
 589 }
 590
 591 void tls_device_splice_eof(struct socket *sock)
 592 {
 593         struct sock *sk = sock->sk;
 594         struct tls_context *tls_ctx = tls_get_ctx(sk);
 595         struct iov_iter iter = {};
 596
 597         if (!tls_is_partially_sent_record(tls_ctx))
 598                 return;
 599
 600         mutex_lock(&tls_ctx->tx_lock);
 601         lock_sock(sk);
 602
 603         if (tls_is_partially_sent_record(tls_ctx)) {
 604                 iov_iter_bvec(&iter, ITER_SOURCE, NULL, 0, 0);
 605                 tls_push_data(sk, &iter, 0, 0, TLS_RECORD_TYPE_DATA);
 606         }
 607
 608         release_sock(sk);
 609         mutex_unlock(&tls_ctx->tx_lock);
 610 }
 611
 612 struct tls_record_info *tls_get_record(struct tls_offload_context_tx *context,
 613                                        u32 seq, u64 *p_record_sn)
 614 {
 615         u64 record_sn = context->hint_record_sn;
 616         struct tls_record_info *info, *last;
 617
 618         info = context->retransmit_hint;
 619         if (!info ||
 620             before(seq, info->end_seq - info->len)) {
 621                 /* if retransmit_hint is irrelevant start
 622                  * from the beginning of the list
 623                  */
 624                 info = list_first_entry_or_null(&context->records_list,
 625                                                 struct tls_record_info, list);
 626                 if (!info)
 627                         return NULL;
 628                 /* send the start_marker record if seq number is before the
 629                  * tls offload start marker sequence number. This record is
 630                  * required to handle TCP packets which are before TLS offload
 631                  * started.
 632                  *  And if it's not start marker, look if this seq number
 633                  * belongs to the list.
 634                  */
 635                 if (likely(!tls_record_is_start_marker(info))) {
 636                         /* we have the first record, get the last record to see
 637                          * if this seq number belongs to the list.
 638                          */
 639                         last = list_last_entry(&context->records_list,
 640                                                struct tls_record_info, list);
 641
 642                         if (!between(seq, tls_record_start_seq(info),
 643                                      last->end_seq))
 644                                 return NULL;
 645                 }
 646                 record_sn = context->unacked_record_sn;
 647         }
 648
 649         /* We just need the _rcu for the READ_ONCE() */
 650         rcu_read_lock();
 651         list_for_each_entry_from_rcu(info, &context->records_list, list) {
 652                 if (before(seq, info->end_seq)) {
 653                         if (!context->retransmit_hint ||
 654                             after(info->end_seq,
 655                                   context->retransmit_hint->end_seq)) {
 656                                 context->hint_record_sn = record_sn;
 657                                 context->retransmit_hint = info;
 658                         }
 659                         *p_record_sn = record_sn;
 660                         goto exit_rcu_unlock;
 661                 }
 662                 record_sn++;
 663         }
 664         info = NULL;
 665
 666 exit_rcu_unlock:
 667         rcu_read_unlock();
 668         return info;
 669 }
 670 EXPORT_SYMBOL(tls_get_record);
 671
 672 static int tls_device_push_pending_record(struct sock *sk, int flags)
 673 {
 674         struct iov_iter iter;
 675
 676         iov_iter_kvec(&iter, ITER_SOURCE, NULL, 0, 0);
 677         return tls_push_data(sk, &iter, 0, flags, TLS_RECORD_TYPE_DATA);
 678 }
 679
 680 void tls_device_write_space(struct sock *sk, struct tls_context *ctx)
 681 {
 682         if (tls_is_partially_sent_record(ctx)) {
 683                 gfp_t sk_allocation = sk->sk_allocation;
 684
 685                 WARN_ON_ONCE(sk->sk_write_pending);
 686
 687                 sk->sk_allocation = GFP_ATOMIC;
 688                 tls_push_partial_record(sk, ctx,
 689                                         MSG_DONTWAIT | MSG_NOSIGNAL |
 690                                         MSG_SENDPAGE_DECRYPTED);
 691                 sk->sk_allocation = sk_allocation;
 692         }
 693 }
 694
 695 static void tls_device_resync_rx(struct tls_context *tls_ctx,
 696                                  struct sock *sk, u32 seq, u8 *rcd_sn)
 697 {
 698         struct tls_offload_context_rx *rx_ctx = tls_offload_ctx_rx(tls_ctx);
 699         struct net_device *netdev;
 700
 701         trace_tls_device_rx_resync_send(sk, seq, rcd_sn, rx_ctx->resync_type);
 702         rcu_read_lock();
 703         netdev = rcu_dereference(tls_ctx->netdev);
 704         if (netdev)
 705                 netdev->tlsdev_ops->tls_dev_resync(netdev, sk, seq, rcd_sn,
 706                                                    TLS_OFFLOAD_CTX_DIR_RX);
 707         rcu_read_unlock();
 708         TLS_INC_STATS(sock_net(sk), LINUX_MIB_TLSRXDEVICERESYNC);
 709 }
 710
 711 static bool
 712 tls_device_rx_resync_async(struct tls_offload_resync_async *resync_async,
 713                            s64 resync_req, u32 *seq, u16 *rcd_delta)
 714 {
 715         u32 is_async = resync_req & RESYNC_REQ_ASYNC;
 716         u32 req_seq = resync_req >> 32;
 717         u32 req_end = req_seq + ((resync_req >> 16) & 0xffff);
 718         u16 i;
 719
 720         *rcd_delta = 0;
 721
 722         if (is_async) {
 723                 /* shouldn't get to wraparound:
 724                  * too long in async stage, something bad happened
 725                  */
 726                 if (WARN_ON_ONCE(resync_async->rcd_delta == USHRT_MAX))
 727                         return false;
 728
 729                 /* asynchronous stage: log all headers seq such that
 730                  * req_seq <= seq <= end_seq, and wait for real resync request
 731                  */
 732                 if (before(*seq, req_seq))
 733                         return false;
 734                 if (!after(*seq, req_end) &&
 735                     resync_async->loglen < TLS_DEVICE_RESYNC_ASYNC_LOGMAX)
 736                         resync_async->log[resync_async->loglen++] = *seq;
 737
 738                 resync_async->rcd_delta++;
 739
 740                 return false;
 741         }
 742
 743         /* synchronous stage: check against the logged entries and
 744          * proceed to check the next entries if no match was found
 745          */
 746         for (i = 0; i < resync_async->loglen; i++)
 747                 if (req_seq == resync_async->log[i] &&
 748                     atomic64_try_cmpxchg(&resync_async->req, &resync_req, 0)) {
 749                         *rcd_delta = resync_async->rcd_delta - i;
 750                         *seq = req_seq;
 751                         resync_async->loglen = 0;
 752                         resync_async->rcd_delta = 0;
 753                         return true;
 754                 }
 755
 756         resync_async->loglen = 0;
 757         resync_async->rcd_delta = 0;
 758
 759         if (req_seq == *seq &&
 760             atomic64_try_cmpxchg(&resync_async->req,
 761                                  &resync_req, 0))
 762                 return true;
 763
 764         return false;
 765 }
 766
 767 void tls_device_rx_resync_new_rec(struct sock *sk, u32 rcd_len, u32 seq)
 768 {
 769         struct tls_context *tls_ctx = tls_get_ctx(sk);
 770         struct tls_offload_context_rx *rx_ctx;
 771         u8 rcd_sn[TLS_MAX_REC_SEQ_SIZE];
 772         u32 sock_data, is_req_pending;
 773         struct tls_prot_info *prot;
 774         s64 resync_req;
 775         u16 rcd_delta;
 776         u32 req_seq;
 777
 778         if (tls_ctx->rx_conf != TLS_HW)
 779                 return;
 780         if (unlikely(test_bit(TLS_RX_DEV_DEGRADED, &tls_ctx->flags)))
 781                 return;
 782
 783         prot = &tls_ctx->prot_info;
 784         rx_ctx = tls_offload_ctx_rx(tls_ctx);
 785         memcpy(rcd_sn, tls_ctx->rx.rec_seq, prot->rec_seq_size);
 786
 787         switch (rx_ctx->resync_type) {
 788         case TLS_OFFLOAD_SYNC_TYPE_DRIVER_REQ:
 789                 resync_req = atomic64_read(&rx_ctx->resync_req);
 790                 req_seq = resync_req >> 32;
 791                 seq += TLS_HEADER_SIZE - 1;
 792                 is_req_pending = resync_req;
 793
 794                 if (likely(!is_req_pending) || req_seq != seq ||
 795                     !atomic64_try_cmpxchg(&rx_ctx->resync_req, &resync_req, 0))
 796                         return;
 797                 break;
 798         case TLS_OFFLOAD_SYNC_TYPE_CORE_NEXT_HINT:
 799                 if (likely(!rx_ctx->resync_nh_do_now))
 800                         return;
 801
 802                 /* head of next rec is already in, note that the sock_inq will
 803                  * include the currently parsed message when called from parser
 804                  */
 805                 sock_data = tcp_inq(sk);
 806                 if (sock_data > rcd_len) {
 807                         trace_tls_device_rx_resync_nh_delay(sk, sock_data,
 808                                                             rcd_len);
 809                         return;
 810                 }
 811
 812                 rx_ctx->resync_nh_do_now = 0;
 813                 seq += rcd_len;
 814                 tls_bigint_increment(rcd_sn, prot->rec_seq_size);
 815                 break;
 816         case TLS_OFFLOAD_SYNC_TYPE_DRIVER_REQ_ASYNC:
 817                 resync_req = atomic64_read(&rx_ctx->resync_async->req);
 818                 is_req_pending = resync_req;
 819                 if (likely(!is_req_pending))
 820                         return;
 821
 822                 if (!tls_device_rx_resync_async(rx_ctx->resync_async,
 823                                                 resync_req, &seq, &rcd_delta))
 824                         return;
 825                 tls_bigint_subtract(rcd_sn, rcd_delta);
 826                 break;
 827         }
 828
 829         tls_device_resync_rx(tls_ctx, sk, seq, rcd_sn);
 830 }
 831
 832 static void tls_device_core_ctrl_rx_resync(struct tls_context *tls_ctx,
 833                                            struct tls_offload_context_rx *ctx,
 834                                            struct sock *sk, struct sk_buff *skb)
 835 {
 836         struct strp_msg *rxm;
 837
 838         /* device will request resyncs by itself based on stream scan */
 839         if (ctx->resync_type != TLS_OFFLOAD_SYNC_TYPE_CORE_NEXT_HINT)
 840                 return;
 841         /* already scheduled */
 842         if (ctx->resync_nh_do_now)
 843                 return;
 844         /* seen decrypted fragments since last fully-failed record */
 845         if (ctx->resync_nh_reset) {
 846                 ctx->resync_nh_reset = 0;
 847                 ctx->resync_nh.decrypted_failed = 1;
 848                 ctx->resync_nh.decrypted_tgt = TLS_DEVICE_RESYNC_NH_START_IVAL;
 849                 return;
 850         }
 851
 852         if (++ctx->resync_nh.decrypted_failed <= ctx->resync_nh.decrypted_tgt)
 853                 return;
 854
 855         /* doing resync, bump the next target in case it fails */
 856         if (ctx->resync_nh.decrypted_tgt < TLS_DEVICE_RESYNC_NH_MAX_IVAL)
 857                 ctx->resync_nh.decrypted_tgt *= 2;
 858         else
 859                 ctx->resync_nh.decrypted_tgt += TLS_DEVICE_RESYNC_NH_MAX_IVAL;
 860
 861         rxm = strp_msg(skb);
 862
 863         /* head of next rec is already in, parser will sync for us */
 864         if (tcp_inq(sk) > rxm->full_len) {
 865                 trace_tls_device_rx_resync_nh_schedule(sk);
 866                 ctx->resync_nh_do_now = 1;
 867         } else {
 868                 struct tls_prot_info *prot = &tls_ctx->prot_info;
 869                 u8 rcd_sn[TLS_MAX_REC_SEQ_SIZE];
 870
 871                 memcpy(rcd_sn, tls_ctx->rx.rec_seq, prot->rec_seq_size);
 872                 tls_bigint_increment(rcd_sn, prot->rec_seq_size);
 873
 874                 tls_device_resync_rx(tls_ctx, sk, tcp_sk(sk)->copied_seq,
 875                                      rcd_sn);
 876         }
 877 }
 878
 879 static int
 880 tls_device_reencrypt(struct sock *sk, struct tls_context *tls_ctx)
 881 {
 882         struct tls_sw_context_rx *sw_ctx = tls_sw_ctx_rx(tls_ctx);
 883         const struct tls_cipher_size_desc *cipher_sz;
 884         int err, offset, copy, data_len, pos;
 885         struct sk_buff *skb, *skb_iter;
 886         struct scatterlist sg[1];
 887         struct strp_msg *rxm;
 888         char *orig_buf, *buf;
 889
 890         switch (tls_ctx->crypto_recv.info.cipher_type) {
 891         case TLS_CIPHER_AES_GCM_128:
 892         case TLS_CIPHER_AES_GCM_256:
 893                 break;
 894         default:
 895                 return -EINVAL;
 896         }
 897         cipher_sz = &tls_cipher_size_desc[tls_ctx->crypto_recv.info.cipher_type];
 898
 899         rxm = strp_msg(tls_strp_msg(sw_ctx));
 900         orig_buf = kmalloc(rxm->full_len + TLS_HEADER_SIZE + cipher_sz->iv,
 901                            sk->sk_allocation);
 902         if (!orig_buf)
 903                 return -ENOMEM;
 904         buf = orig_buf;
 905
 906         err = tls_strp_msg_cow(sw_ctx);
 907         if (unlikely(err))
 908                 goto free_buf;
 909
 910         skb = tls_strp_msg(sw_ctx);
 911         rxm = strp_msg(skb);
 912         offset = rxm->offset;
 913
 914         sg_init_table(sg, 1);
 915         sg_set_buf(&sg[0], buf,
 916                    rxm->full_len + TLS_HEADER_SIZE + cipher_sz->iv);
 917         err = skb_copy_bits(skb, offset, buf, TLS_HEADER_SIZE + cipher_sz->iv);
 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, sg);
 923         if (err != -EBADMSG)
 924                 goto free_buf;
 925         else
 926                 err = 0;
 927
 928         data_len = rxm->full_len - cipher_sz->tag;
 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 {
 979         struct tls_offload_context_rx *ctx = tls_offload_ctx_rx(tls_ctx);
 980         struct tls_sw_context_rx *sw_ctx = tls_sw_ctx_rx(tls_ctx);
 981         struct sk_buff *skb = tls_strp_msg(sw_ctx);
 982         struct strp_msg *rxm = strp_msg(skb);
 983         int is_decrypted, is_encrypted;
 984
 985         if (!tls_strp_msg_mixed_decrypted(sw_ctx)) {
 986                 is_decrypted = skb->decrypted;
 987                 is_encrypted = !is_decrypted;
 988         } else {
 989                 is_decrypted = 0;
 990                 is_encrypted = 0;
 991         }
 992
 993         trace_tls_device_decrypted(sk, tcp_sk(sk)->copied_seq - rxm->full_len,
 994                                    tls_ctx->rx.rec_seq, rxm->full_len,
 995                                    is_encrypted, is_decrypted);
 996
 997         if (unlikely(test_bit(TLS_RX_DEV_DEGRADED, &tls_ctx->flags))) {
 998                 if (likely(is_encrypted || is_decrypted))
 999                         return is_decrypted;
1000
1001                 /* After tls_device_down disables the offload, the next SKB will
1002                  * likely have initial fragments decrypted, and final ones not
1003                  * decrypted. We need to reencrypt that single SKB.
1004                  */
1005                 return tls_device_reencrypt(sk, tls_ctx);
1006         }
1007
1008         /* Return immediately if the record is either entirely plaintext or
1009          * entirely ciphertext. Otherwise handle reencrypt partially decrypted
1010          * record.
1011          */
1012         if (is_decrypted) {
1013                 ctx->resync_nh_reset = 1;
1014                 return is_decrypted;
1015         }
1016         if (is_encrypted) {
1017                 tls_device_core_ctrl_rx_resync(tls_ctx, ctx, sk, skb);
1018                 return 0;
1019         }
1020
1021         ctx->resync_nh_reset = 1;
1022         return tls_device_reencrypt(sk, tls_ctx);
1023 }
1024
1025 static void tls_device_attach(struct tls_context *ctx, struct sock *sk,
1026                               struct net_device *netdev)
1027 {
1028         if (sk->sk_destruct != tls_device_sk_destruct) {
1029                 refcount_set(&ctx->refcount, 1);
1030                 dev_hold(netdev);
1031                 RCU_INIT_POINTER(ctx->netdev, netdev);
1032                 spin_lock_irq(&tls_device_lock);
1033                 list_add_tail(&ctx->list, &tls_device_list);
1034                 spin_unlock_irq(&tls_device_lock);
1035
1036                 ctx->sk_destruct = sk->sk_destruct;
1037                 smp_store_release(&sk->sk_destruct, tls_device_sk_destruct);
1038         }
1039 }
1040
1041 int tls_set_device_offload(struct sock *sk, struct tls_context *ctx)
1042 {
1043         struct tls_context *tls_ctx = tls_get_ctx(sk);
1044         struct tls_prot_info *prot = &tls_ctx->prot_info;
1045         const struct tls_cipher_size_desc *cipher_sz;
1046         struct tls_record_info *start_marker_record;
1047         struct tls_offload_context_tx *offload_ctx;
1048         struct tls_crypto_info *crypto_info;
1049         struct net_device *netdev;
1050         char *iv, *rec_seq;
1051         struct sk_buff *skb;
1052         __be64 rcd_sn;
1053         int rc;
1054
1055         if (!ctx)
1056                 return -EINVAL;
1057
1058         if (ctx->priv_ctx_tx)
1059                 return -EEXIST;
1060
1061         netdev = get_netdev_for_sock(sk);
1062         if (!netdev) {
1063                 pr_err_ratelimited("%s: netdev not found\n", __func__);
1064                 return -EINVAL;
1065         }
1066
1067         if (!(netdev->features & NETIF_F_HW_TLS_TX)) {
1068                 rc = -EOPNOTSUPP;
1069                 goto release_netdev;
1070         }
1071
1072         crypto_info = &ctx->crypto_send.info;
1073         if (crypto_info->version != TLS_1_2_VERSION) {
1074                 rc = -EOPNOTSUPP;
1075                 goto release_netdev;
1076         }
1077
1078         switch (crypto_info->cipher_type) {
1079         case TLS_CIPHER_AES_GCM_128:
1080                 iv = ((struct tls12_crypto_info_aes_gcm_128 *)crypto_info)->iv;
1081                 rec_seq =
1082                  ((struct tls12_crypto_info_aes_gcm_128 *)crypto_info)->rec_seq;
1083                 break;
1084         case TLS_CIPHER_AES_GCM_256:
1085                 iv = ((struct tls12_crypto_info_aes_gcm_256 *)crypto_info)->iv;
1086                 rec_seq =
1087                  ((struct tls12_crypto_info_aes_gcm_256 *)crypto_info)->rec_seq;
1088                 break;
1089         default:
1090                 rc = -EINVAL;
1091                 goto release_netdev;
1092         }
1093         cipher_sz = &tls_cipher_size_desc[crypto_info->cipher_type];
1094
1095         /* Sanity-check the rec_seq_size for stack allocations */
1096         if (cipher_sz->rec_seq > TLS_MAX_REC_SEQ_SIZE) {
1097                 rc = -EINVAL;
1098                 goto release_netdev;
1099         }
1100
1101         prot->version = crypto_info->version;
1102         prot->cipher_type = crypto_info->cipher_type;
1103         prot->prepend_size = TLS_HEADER_SIZE + cipher_sz->iv;
1104         prot->tag_size = cipher_sz->tag;
1105         prot->overhead_size = prot->prepend_size + prot->tag_size;
1106         prot->iv_size = cipher_sz->iv;
1107         prot->salt_size = cipher_sz->salt;
1108         ctx->tx.iv = kmalloc(cipher_sz->iv + cipher_sz->salt, GFP_KERNEL);
1109         if (!ctx->tx.iv) {
1110                 rc = -ENOMEM;
1111                 goto release_netdev;
1112         }
1113
1114         memcpy(ctx->tx.iv + cipher_sz->salt, iv, cipher_sz->iv);
1115
1116         prot->rec_seq_size = cipher_sz->rec_seq;
1117         ctx->tx.rec_seq = kmemdup(rec_seq, cipher_sz->rec_seq, GFP_KERNEL);
1118         if (!ctx->tx.rec_seq) {
1119                 rc = -ENOMEM;
1120                 goto free_iv;
1121         }
1122
1123         start_marker_record = kmalloc(sizeof(*start_marker_record), GFP_KERNEL);
1124         if (!start_marker_record) {
1125                 rc = -ENOMEM;
1126                 goto free_rec_seq;
1127         }
1128
1129         offload_ctx = kzalloc(TLS_OFFLOAD_CONTEXT_SIZE_TX, GFP_KERNEL);
1130         if (!offload_ctx) {
1131                 rc = -ENOMEM;
1132                 goto free_marker_record;
1133         }
1134
1135         rc = tls_sw_fallback_init(sk, offload_ctx, crypto_info);
1136         if (rc)
1137                 goto free_offload_ctx;
1138
1139         /* start at rec_seq - 1 to account for the start marker record */
1140         memcpy(&rcd_sn, ctx->tx.rec_seq, sizeof(rcd_sn));
1141         offload_ctx->unacked_record_sn = be64_to_cpu(rcd_sn) - 1;
1142
1143         start_marker_record->end_seq = tcp_sk(sk)->write_seq;
1144         start_marker_record->len = 0;
1145         start_marker_record->num_frags = 0;
1146
1147         INIT_WORK(&offload_ctx->destruct_work, tls_device_tx_del_task);
1148         offload_ctx->ctx = ctx;
1149
1150         INIT_LIST_HEAD(&offload_ctx->records_list);
1151         list_add_tail(&start_marker_record->list, &offload_ctx->records_list);
1152         spin_lock_init(&offload_ctx->lock);
1153         sg_init_table(offload_ctx->sg_tx_data,
1154                       ARRAY_SIZE(offload_ctx->sg_tx_data));
1155
1156         clean_acked_data_enable(inet_csk(sk), &tls_icsk_clean_acked);
1157         ctx->push_pending_record = tls_device_push_pending_record;
1158
1159         /* TLS offload is greatly simplified if we don't send
1160          * SKBs where only part of the payload needs to be encrypted.
1161          * So mark the last skb in the write queue as end of record.
1162          */
1163         skb = tcp_write_queue_tail(sk);
1164         if (skb)
1165                 TCP_SKB_CB(skb)->eor = 1;
1166
1167         /* Avoid offloading if the device is down
1168          * We don't want to offload new flows after
1169          * the NETDEV_DOWN event
1170          *
1171          * device_offload_lock is taken in tls_devices's NETDEV_DOWN
1172          * handler thus protecting from the device going down before
1173          * ctx was added to tls_device_list.
1174          */
1175         down_read(&device_offload_lock);
1176         if (!(netdev->flags & IFF_UP)) {
1177                 rc = -EINVAL;
1178                 goto release_lock;
1179         }
1180
1181         ctx->priv_ctx_tx = offload_ctx;
1182         rc = netdev->tlsdev_ops->tls_dev_add(netdev, sk, TLS_OFFLOAD_CTX_DIR_TX,
1183                                              &ctx->crypto_send.info,
1184                                              tcp_sk(sk)->write_seq);
1185         trace_tls_device_offload_set(sk, TLS_OFFLOAD_CTX_DIR_TX,
1186                                      tcp_sk(sk)->write_seq, rec_seq, rc);
1187         if (rc)
1188                 goto release_lock;
1189
1190         tls_device_attach(ctx, sk, netdev);
1191         up_read(&device_offload_lock);
1192
1193         /* following this assignment tls_is_skb_tx_device_offloaded
1194          * will return true and the context might be accessed
1195          * by the netdev's xmit function.
1196          */
1197         smp_store_release(&sk->sk_validate_xmit_skb, tls_validate_xmit_skb);
1198         dev_put(netdev);
1199
1200         return 0;
1201
1202 release_lock:
1203         up_read(&device_offload_lock);
1204         clean_acked_data_disable(inet_csk(sk));
1205         crypto_free_aead(offload_ctx->aead_send);
1206 free_offload_ctx:
1207         kfree(offload_ctx);
1208         ctx->priv_ctx_tx = NULL;
1209 free_marker_record:
1210         kfree(start_marker_record);
1211 free_rec_seq:
1212         kfree(ctx->tx.rec_seq);
1213 free_iv:
1214         kfree(ctx->tx.iv);
1215 release_netdev:
1216         dev_put(netdev);
1217         return rc;
1218 }
1219
1220 int tls_set_device_offload_rx(struct sock *sk, struct tls_context *ctx)
1221 {
1222         struct tls12_crypto_info_aes_gcm_128 *info;
1223         struct tls_offload_context_rx *context;
1224         struct net_device *netdev;
1225         int rc = 0;
1226
1227         if (ctx->crypto_recv.info.version != TLS_1_2_VERSION)
1228                 return -EOPNOTSUPP;
1229
1230         netdev = get_netdev_for_sock(sk);
1231         if (!netdev) {
1232                 pr_err_ratelimited("%s: netdev not found\n", __func__);
1233                 return -EINVAL;
1234         }
1235
1236         if (!(netdev->features & NETIF_F_HW_TLS_RX)) {
1237                 rc = -EOPNOTSUPP;
1238                 goto release_netdev;
1239         }
1240
1241         /* Avoid offloading if the device is down
1242          * We don't want to offload new flows after
1243          * the NETDEV_DOWN event
1244          *
1245          * device_offload_lock is taken in tls_devices's NETDEV_DOWN
1246          * handler thus protecting from the device going down before
1247          * ctx was added to tls_device_list.
1248          */
1249         down_read(&device_offload_lock);
1250         if (!(netdev->flags & IFF_UP)) {
1251                 rc = -EINVAL;
1252                 goto release_lock;
1253         }
1254
1255         context = kzalloc(TLS_OFFLOAD_CONTEXT_SIZE_RX, GFP_KERNEL);
1256         if (!context) {
1257                 rc = -ENOMEM;
1258                 goto release_lock;
1259         }
1260         context->resync_nh_reset = 1;
1261
1262         ctx->priv_ctx_rx = context;
1263         rc = tls_set_sw_offload(sk, ctx, 0);
1264         if (rc)
1265                 goto release_ctx;
1266
1267         rc = netdev->tlsdev_ops->tls_dev_add(netdev, sk, TLS_OFFLOAD_CTX_DIR_RX,
1268                                              &ctx->crypto_recv.info,
1269                                              tcp_sk(sk)->copied_seq);
1270         info = (void *)&ctx->crypto_recv.info;
1271         trace_tls_device_offload_set(sk, TLS_OFFLOAD_CTX_DIR_RX,
1272                                      tcp_sk(sk)->copied_seq, info->rec_seq, rc);
1273         if (rc)
1274                 goto free_sw_resources;
1275
1276         tls_device_attach(ctx, sk, netdev);
1277         up_read(&device_offload_lock);
1278
1279         dev_put(netdev);
1280
1281         return 0;
1282
1283 free_sw_resources:
1284         up_read(&device_offload_lock);
1285         tls_sw_free_resources_rx(sk);
1286         down_read(&device_offload_lock);
1287 release_ctx:
1288         ctx->priv_ctx_rx = NULL;
1289 release_lock:
1290         up_read(&device_offload_lock);
1291 release_netdev:
1292         dev_put(netdev);
1293         return rc;
1294 }
1295
1296 void tls_device_offload_cleanup_rx(struct sock *sk)
1297 {
1298         struct tls_context *tls_ctx = tls_get_ctx(sk);
1299         struct net_device *netdev;
1300
1301         down_read(&device_offload_lock);
1302         netdev = rcu_dereference_protected(tls_ctx->netdev,
1303                                            lockdep_is_held(&device_offload_lock));
1304         if (!netdev)
1305                 goto out;
1306
1307         netdev->tlsdev_ops->tls_dev_del(netdev, tls_ctx,
1308                                         TLS_OFFLOAD_CTX_DIR_RX);
1309
1310         if (tls_ctx->tx_conf != TLS_HW) {
1311                 dev_put(netdev);
1312                 rcu_assign_pointer(tls_ctx->netdev, NULL);
1313         } else {
1314                 set_bit(TLS_RX_DEV_CLOSED, &tls_ctx->flags);
1315         }
1316 out:
1317         up_read(&device_offload_lock);
1318         tls_sw_release_resources_rx(sk);
1319 }
1320
1321 static int tls_device_down(struct net_device *netdev)
1322 {
1323         struct tls_context *ctx, *tmp;
1324         unsigned long flags;
1325         LIST_HEAD(list);
1326
1327         /* Request a write lock to block new offload attempts */
1328         down_write(&device_offload_lock);
1329
1330         spin_lock_irqsave(&tls_device_lock, flags);
1331         list_for_each_entry_safe(ctx, tmp, &tls_device_list, list) {
1332                 struct net_device *ctx_netdev =
1333                         rcu_dereference_protected(ctx->netdev,
1334                                                   lockdep_is_held(&device_offload_lock));
1335
1336                 if (ctx_netdev != netdev ||
1337                     !refcount_inc_not_zero(&ctx->refcount))
1338                         continue;
1339
1340                 list_move(&ctx->list, &list);
1341         }
1342         spin_unlock_irqrestore(&tls_device_lock, flags);
1343
1344         list_for_each_entry_safe(ctx, tmp, &list, list) {
1345                 /* Stop offloaded TX and switch to the fallback.
1346                  * tls_is_skb_tx_device_offloaded will return false.
1347                  */
1348                 WRITE_ONCE(ctx->sk->sk_validate_xmit_skb, tls_validate_xmit_skb_sw);
1349
1350                 /* Stop the RX and TX resync.
1351                  * tls_dev_resync must not be called after tls_dev_del.
1352                  */
1353                 rcu_assign_pointer(ctx->netdev, NULL);
1354
1355                 /* Start skipping the RX resync logic completely. */
1356                 set_bit(TLS_RX_DEV_DEGRADED, &ctx->flags);
1357
1358                 /* Sync with inflight packets. After this point:
1359                  * TX: no non-encrypted packets will be passed to the driver.
1360                  * RX: resync requests from the driver will be ignored.
1361                  */
1362                 synchronize_net();
1363
1364                 /* Release the offload context on the driver side. */
1365                 if (ctx->tx_conf == TLS_HW)
1366                         netdev->tlsdev_ops->tls_dev_del(netdev, ctx,
1367                                                         TLS_OFFLOAD_CTX_DIR_TX);
1368                 if (ctx->rx_conf == TLS_HW &&
1369                     !test_bit(TLS_RX_DEV_CLOSED, &ctx->flags))
1370                         netdev->tlsdev_ops->tls_dev_del(netdev, ctx,
1371                                                         TLS_OFFLOAD_CTX_DIR_RX);
1372
1373                 dev_put(netdev);
1374
1375                 /* Move the context to a separate list for two reasons:
1376                  * 1. When the context is deallocated, list_del is called.
1377                  * 2. It's no longer an offloaded context, so we don't want to
1378                  *    run offload-specific code on this context.
1379                  */
1380                 spin_lock_irqsave(&tls_device_lock, flags);
1381                 list_move_tail(&ctx->list, &tls_device_down_list);
1382                 spin_unlock_irqrestore(&tls_device_lock, flags);
1383
1384                 /* Device contexts for RX and TX will be freed in on sk_destruct
1385                  * by tls_device_free_ctx. rx_conf and tx_conf stay in TLS_HW.
1386                  * Now release the ref taken above.
1387                  */
1388                 if (refcount_dec_and_test(&ctx->refcount)) {
1389                         /* sk_destruct ran after tls_device_down took a ref, and
1390                          * it returned early. Complete the destruction here.
1391                          */
1392                         list_del(&ctx->list);
1393                         tls_device_free_ctx(ctx);
1394                 }
1395         }
1396
1397         up_write(&device_offload_lock);
1398
1399         flush_workqueue(destruct_wq);
1400
1401         return NOTIFY_DONE;
1402 }
1403
1404 static int tls_dev_event(struct notifier_block *this, unsigned long event,
1405                          void *ptr)
1406 {
1407         struct net_device *dev = netdev_notifier_info_to_dev(ptr);
1408
1409         if (!dev->tlsdev_ops &&
1410             !(dev->features & (NETIF_F_HW_TLS_RX | NETIF_F_HW_TLS_TX)))
1411                 return NOTIFY_DONE;
1412
1413         switch (event) {
1414         case NETDEV_REGISTER:
1415         case NETDEV_FEAT_CHANGE:
1416                 if (netif_is_bond_master(dev))
1417                         return NOTIFY_DONE;
1418                 if ((dev->features & NETIF_F_HW_TLS_RX) &&
1419                     !dev->tlsdev_ops->tls_dev_resync)
1420                         return NOTIFY_BAD;
1421
1422                 if  (dev->tlsdev_ops &&
1423                      dev->tlsdev_ops->tls_dev_add &&
1424                      dev->tlsdev_ops->tls_dev_del)
1425                         return NOTIFY_DONE;
1426                 else
1427                         return NOTIFY_BAD;
1428         case NETDEV_DOWN:
1429                 return tls_device_down(dev);
1430         }
1431         return NOTIFY_DONE;
1432 }
1433
1434 static struct notifier_block tls_dev_notifier = {
1435         .notifier_call  = tls_dev_event,
1436 };
1437
1438 int __init tls_device_init(void)
1439 {
1440         int err;
1441
1442         dummy_page = alloc_page(GFP_KERNEL);
1443         if (!dummy_page)
1444                 return -ENOMEM;
1445
1446         destruct_wq = alloc_workqueue("ktls_device_destruct", 0, 0);
1447         if (!destruct_wq) {
1448                 err = -ENOMEM;
1449                 goto err_free_dummy;
1450         }
1451
1452         err = register_netdevice_notifier(&tls_dev_notifier);
1453         if (err)
1454                 goto err_destroy_wq;
1455
1456         return 0;
1457
1458 err_destroy_wq:
1459         destroy_workqueue(destruct_wq);
1460 err_free_dummy:
1461         put_page(dummy_page);
1462         return err;
1463 }
1464
1465 void __exit tls_device_cleanup(void)
1466 {
1467         unregister_netdevice_notifier(&tls_dev_notifier);
1468         destroy_workqueue(destruct_wq);
1469         clean_acked_data_flush();
1470         put_page(dummy_page);
1471 }