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 void tls_device_free_ctx(struct tls_context *ctx)
  56 {
  57         if (ctx->tx_conf == TLS_HW) {
  58                 kfree(tls_offload_ctx_tx(ctx));
  59                 kfree(ctx->tx.rec_seq);
  60                 kfree(ctx->tx.iv);
  61         }
  62
  63         if (ctx->rx_conf == TLS_HW)
  64                 kfree(tls_offload_ctx_rx(ctx));
  65
  66         tls_ctx_free(NULL, ctx);
  67 }
  68
  69 static void tls_device_tx_del_task(struct work_struct *work)
  70 {
  71         struct tls_offload_context_tx *offload_ctx =
  72                 container_of(work, struct tls_offload_context_tx, destruct_work);
  73         struct tls_context *ctx = offload_ctx->ctx;
  74         struct net_device *netdev;
  75
  76         /* Safe, because this is the destroy flow, refcount is 0, so
  77          * tls_device_down can't store this field in parallel.
  78          */
  79         netdev = rcu_dereference_protected(ctx->netdev,
  80                                            !refcount_read(&ctx->refcount));
  81
  82         netdev->tlsdev_ops->tls_dev_del(netdev, ctx, TLS_OFFLOAD_CTX_DIR_TX);
  83         dev_put(netdev);
  84         ctx->netdev = NULL;
  85         tls_device_free_ctx(ctx);
  86 }
  87
  88 static void tls_device_queue_ctx_destruction(struct tls_context *ctx)
  89 {
  90         struct net_device *netdev;
  91         unsigned long flags;
  92         bool async_cleanup;
  93
  94         spin_lock_irqsave(&tls_device_lock, flags);
  95         if (unlikely(!refcount_dec_and_test(&ctx->refcount))) {
  96                 spin_unlock_irqrestore(&tls_device_lock, flags);
  97                 return;
  98         }
  99
 100         list_del(&ctx->list); /* Remove from tls_device_list / tls_device_down_list */
 101
 102         /* Safe, because this is the destroy flow, refcount is 0, so
 103          * tls_device_down can't store this field in parallel.
 104          */
 105         netdev = rcu_dereference_protected(ctx->netdev,
 106                                            !refcount_read(&ctx->refcount));
 107
 108         async_cleanup = netdev && ctx->tx_conf == TLS_HW;
 109         if (async_cleanup) {
 110                 struct tls_offload_context_tx *offload_ctx = tls_offload_ctx_tx(ctx);
 111
 112                 /* queue_work inside the spinlock
 113                  * to make sure tls_device_down waits for that work.
 114                  */
 115                 queue_work(destruct_wq, &offload_ctx->destruct_work);
 116         }
 117         spin_unlock_irqrestore(&tls_device_lock, flags);
 118
 119         if (!async_cleanup)
 120                 tls_device_free_ctx(ctx);
 121 }
 122
 123 /* We assume that the socket is already connected */
 124 static struct net_device *get_netdev_for_sock(struct sock *sk)
 125 {
 126         struct dst_entry *dst = sk_dst_get(sk);
 127         struct net_device *netdev = NULL;
 128
 129         if (likely(dst)) {
 130                 netdev = netdev_sk_get_lowest_dev(dst->dev, sk);
 131                 dev_hold(netdev);
 132         }
 133
 134         dst_release(dst);
 135
 136         return netdev;
 137 }
 138
 139 static void destroy_record(struct tls_record_info *record)
 140 {
 141         int i;
 142
 143         for (i = 0; i < record->num_frags; i++)
 144                 __skb_frag_unref(&record->frags[i], false);
 145         kfree(record);
 146 }
 147
 148 static void delete_all_records(struct tls_offload_context_tx *offload_ctx)
 149 {
 150         struct tls_record_info *info, *temp;
 151
 152         list_for_each_entry_safe(info, temp, &offload_ctx->records_list, list) {
 153                 list_del(&info->list);
 154                 destroy_record(info);
 155         }
 156
 157         offload_ctx->retransmit_hint = NULL;
 158 }
 159
 160 static void tls_icsk_clean_acked(struct sock *sk, u32 acked_seq)
 161 {
 162         struct tls_context *tls_ctx = tls_get_ctx(sk);
 163         struct tls_record_info *info, *temp;
 164         struct tls_offload_context_tx *ctx;
 165         u64 deleted_records = 0;
 166         unsigned long flags;
 167
 168         if (!tls_ctx)
 169                 return;
 170
 171         ctx = tls_offload_ctx_tx(tls_ctx);
 172
 173         spin_lock_irqsave(&ctx->lock, flags);
 174         info = ctx->retransmit_hint;
 175         if (info && !before(acked_seq, info->end_seq))
 176                 ctx->retransmit_hint = NULL;
 177
 178         list_for_each_entry_safe(info, temp, &ctx->records_list, list) {
 179                 if (before(acked_seq, info->end_seq))
 180                         break;
 181                 list_del(&info->list);
 182
 183                 destroy_record(info);
 184                 deleted_records++;
 185         }
 186
 187         ctx->unacked_record_sn += deleted_records;
 188         spin_unlock_irqrestore(&ctx->lock, flags);
 189 }
 190
 191 /* At this point, there should be no references on this
 192  * socket and no in-flight SKBs associated with this
 193  * socket, so it is safe to free all the resources.
 194  */
 195 void tls_device_sk_destruct(struct sock *sk)
 196 {
 197         struct tls_context *tls_ctx = tls_get_ctx(sk);
 198         struct tls_offload_context_tx *ctx = tls_offload_ctx_tx(tls_ctx);
 199
 200         tls_ctx->sk_destruct(sk);
 201
 202         if (tls_ctx->tx_conf == TLS_HW) {
 203                 if (ctx->open_record)
 204                         destroy_record(ctx->open_record);
 205                 delete_all_records(ctx);
 206                 crypto_free_aead(ctx->aead_send);
 207                 clean_acked_data_disable(inet_csk(sk));
 208         }
 209
 210         tls_device_queue_ctx_destruction(tls_ctx);
 211 }
 212 EXPORT_SYMBOL_GPL(tls_device_sk_destruct);
 213
 214 void tls_device_free_resources_tx(struct sock *sk)
 215 {
 216         struct tls_context *tls_ctx = tls_get_ctx(sk);
 217
 218         tls_free_partial_record(sk, tls_ctx);
 219 }
 220
 221 void tls_offload_tx_resync_request(struct sock *sk, u32 got_seq, u32 exp_seq)
 222 {
 223         struct tls_context *tls_ctx = tls_get_ctx(sk);
 224
 225         trace_tls_device_tx_resync_req(sk, got_seq, exp_seq);
 226         WARN_ON(test_and_set_bit(TLS_TX_SYNC_SCHED, &tls_ctx->flags));
 227 }
 228 EXPORT_SYMBOL_GPL(tls_offload_tx_resync_request);
 229
 230 static void tls_device_resync_tx(struct sock *sk, struct tls_context *tls_ctx,
 231                                  u32 seq)
 232 {
 233         struct net_device *netdev;
 234         struct sk_buff *skb;
 235         int err = 0;
 236         u8 *rcd_sn;
 237
 238         skb = tcp_write_queue_tail(sk);
 239         if (skb)
 240                 TCP_SKB_CB(skb)->eor = 1;
 241
 242         rcd_sn = tls_ctx->tx.rec_seq;
 243
 244         trace_tls_device_tx_resync_send(sk, seq, rcd_sn);
 245         down_read(&device_offload_lock);
 246         netdev = rcu_dereference_protected(tls_ctx->netdev,
 247                                            lockdep_is_held(&device_offload_lock));
 248         if (netdev)
 249                 err = netdev->tlsdev_ops->tls_dev_resync(netdev, sk, seq,
 250                                                          rcd_sn,
 251                                                          TLS_OFFLOAD_CTX_DIR_TX);
 252         up_read(&device_offload_lock);
 253         if (err)
 254                 return;
 255
 256         clear_bit_unlock(TLS_TX_SYNC_SCHED, &tls_ctx->flags);
 257 }
 258
 259 static void tls_append_frag(struct tls_record_info *record,
 260                             struct page_frag *pfrag,
 261                             int size)
 262 {
 263         skb_frag_t *frag;
 264
 265         frag = &record->frags[record->num_frags - 1];
 266         if (skb_frag_page(frag) == pfrag->page &&
 267             skb_frag_off(frag) + skb_frag_size(frag) == pfrag->offset) {
 268                 skb_frag_size_add(frag, size);
 269         } else {
 270                 ++frag;
 271                 __skb_frag_set_page(frag, pfrag->page);
 272                 skb_frag_off_set(frag, pfrag->offset);
 273                 skb_frag_size_set(frag, size);
 274                 ++record->num_frags;
 275                 get_page(pfrag->page);
 276         }
 277
 278         pfrag->offset += size;
 279         record->len += size;
 280 }
 281
 282 static int tls_push_record(struct sock *sk,
 283                            struct tls_context *ctx,
 284                            struct tls_offload_context_tx *offload_ctx,
 285                            struct tls_record_info *record,
 286                            int flags)
 287 {
 288         struct tls_prot_info *prot = &ctx->prot_info;
 289         struct tcp_sock *tp = tcp_sk(sk);
 290         skb_frag_t *frag;
 291         int i;
 292
 293         record->end_seq = tp->write_seq + record->len;
 294         list_add_tail_rcu(&record->list, &offload_ctx->records_list);
 295         offload_ctx->open_record = NULL;
 296
 297         if (test_bit(TLS_TX_SYNC_SCHED, &ctx->flags))
 298                 tls_device_resync_tx(sk, ctx, tp->write_seq);
 299
 300         tls_advance_record_sn(sk, prot, &ctx->tx);
 301
 302         for (i = 0; i < record->num_frags; i++) {
 303                 frag = &record->frags[i];
 304                 sg_unmark_end(&offload_ctx->sg_tx_data[i]);
 305                 sg_set_page(&offload_ctx->sg_tx_data[i], skb_frag_page(frag),
 306                             skb_frag_size(frag), skb_frag_off(frag));
 307                 sk_mem_charge(sk, skb_frag_size(frag));
 308                 get_page(skb_frag_page(frag));
 309         }
 310         sg_mark_end(&offload_ctx->sg_tx_data[record->num_frags - 1]);
 311
 312         /* all ready, send */
 313         return tls_push_sg(sk, ctx, offload_ctx->sg_tx_data, 0, flags);
 314 }
 315
 316 static int tls_device_record_close(struct sock *sk,
 317                                    struct tls_context *ctx,
 318                                    struct tls_record_info *record,
 319                                    struct page_frag *pfrag,
 320                                    unsigned char record_type)
 321 {
 322         struct tls_prot_info *prot = &ctx->prot_info;
 323         int ret;
 324
 325         /* append tag
 326          * device will fill in the tag, we just need to append a placeholder
 327          * use socket memory to improve coalescing (re-using a single buffer
 328          * increases frag count)
 329          * if we can't allocate memory now, steal some back from data
 330          */
 331         if (likely(skb_page_frag_refill(prot->tag_size, pfrag,
 332                                         sk->sk_allocation))) {
 333                 ret = 0;
 334                 tls_append_frag(record, pfrag, prot->tag_size);
 335         } else {
 336                 ret = prot->tag_size;
 337                 if (record->len <= prot->overhead_size)
 338                         return -ENOMEM;
 339         }
 340
 341         /* fill prepend */
 342         tls_fill_prepend(ctx, skb_frag_address(&record->frags[0]),
 343                          record->len - prot->overhead_size,
 344                          record_type);
 345         return ret;
 346 }
 347
 348 static int tls_create_new_record(struct tls_offload_context_tx *offload_ctx,
 349                                  struct page_frag *pfrag,
 350                                  size_t prepend_size)
 351 {
 352         struct tls_record_info *record;
 353         skb_frag_t *frag;
 354
 355         record = kmalloc(sizeof(*record), GFP_KERNEL);
 356         if (!record)
 357                 return -ENOMEM;
 358
 359         frag = &record->frags[0];
 360         __skb_frag_set_page(frag, pfrag->page);
 361         skb_frag_off_set(frag, pfrag->offset);
 362         skb_frag_size_set(frag, prepend_size);
 363
 364         get_page(pfrag->page);
 365         pfrag->offset += prepend_size;
 366
 367         record->num_frags = 1;
 368         record->len = prepend_size;
 369         offload_ctx->open_record = record;
 370         return 0;
 371 }
 372
 373 static int tls_do_allocation(struct sock *sk,
 374                              struct tls_offload_context_tx *offload_ctx,
 375                              struct page_frag *pfrag,
 376                              size_t prepend_size)
 377 {
 378         int ret;
 379
 380         if (!offload_ctx->open_record) {
 381                 if (unlikely(!skb_page_frag_refill(prepend_size, pfrag,
 382                                                    sk->sk_allocation))) {
 383                         READ_ONCE(sk->sk_prot)->enter_memory_pressure(sk);
 384                         sk_stream_moderate_sndbuf(sk);
 385                         return -ENOMEM;
 386                 }
 387
 388                 ret = tls_create_new_record(offload_ctx, pfrag, prepend_size);
 389                 if (ret)
 390                         return ret;
 391
 392                 if (pfrag->size > pfrag->offset)
 393                         return 0;
 394         }
 395
 396         if (!sk_page_frag_refill(sk, pfrag))
 397                 return -ENOMEM;
 398
 399         return 0;
 400 }
 401
 402 static int tls_device_copy_data(void *addr, size_t bytes, struct iov_iter *i)
 403 {
 404         size_t pre_copy, nocache;
 405
 406         pre_copy = ~((unsigned long)addr - 1) & (SMP_CACHE_BYTES - 1);
 407         if (pre_copy) {
 408                 pre_copy = min(pre_copy, bytes);
 409                 if (copy_from_iter(addr, pre_copy, i) != pre_copy)
 410                         return -EFAULT;
 411                 bytes -= pre_copy;
 412                 addr += pre_copy;
 413         }
 414
 415         nocache = round_down(bytes, SMP_CACHE_BYTES);
 416         if (copy_from_iter_nocache(addr, nocache, i) != nocache)
 417                 return -EFAULT;
 418         bytes -= nocache;
 419         addr += nocache;
 420
 421         if (bytes && copy_from_iter(addr, bytes, i) != bytes)
 422                 return -EFAULT;
 423
 424         return 0;
 425 }
 426
 427 union tls_iter_offset {
 428         struct iov_iter *msg_iter;
 429         int offset;
 430 };
 431
 432 static int tls_push_data(struct sock *sk,
 433                          union tls_iter_offset iter_offset,
 434                          size_t size, int flags,
 435                          unsigned char record_type,
 436                          struct page *zc_page)
 437 {
 438         struct tls_context *tls_ctx = tls_get_ctx(sk);
 439         struct tls_prot_info *prot = &tls_ctx->prot_info;
 440         struct tls_offload_context_tx *ctx = tls_offload_ctx_tx(tls_ctx);
 441         struct tls_record_info *record;
 442         int tls_push_record_flags;
 443         struct page_frag *pfrag;
 444         size_t orig_size = size;
 445         u32 max_open_record_len;
 446         bool more = false;
 447         bool done = false;
 448         int copy, rc = 0;
 449         long timeo;
 450
 451         if (flags &
 452             ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL | MSG_SENDPAGE_NOTLAST))
 453                 return -EOPNOTSUPP;
 454
 455         if (unlikely(sk->sk_err))
 456                 return -sk->sk_err;
 457
 458         flags |= MSG_SENDPAGE_DECRYPTED;
 459         tls_push_record_flags = flags | MSG_SENDPAGE_NOTLAST;
 460
 461         timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
 462         if (tls_is_partially_sent_record(tls_ctx)) {
 463                 rc = tls_push_partial_record(sk, tls_ctx, flags);
 464                 if (rc < 0)
 465                         return rc;
 466         }
 467
 468         pfrag = sk_page_frag(sk);
 469
 470         /* TLS_HEADER_SIZE is not counted as part of the TLS record, and
 471          * we need to leave room for an authentication tag.
 472          */
 473         max_open_record_len = TLS_MAX_PAYLOAD_SIZE +
 474                               prot->prepend_size;
 475         do {
 476                 rc = tls_do_allocation(sk, ctx, pfrag, prot->prepend_size);
 477                 if (unlikely(rc)) {
 478                         rc = sk_stream_wait_memory(sk, &timeo);
 479                         if (!rc)
 480                                 continue;
 481
 482                         record = ctx->open_record;
 483                         if (!record)
 484                                 break;
 485 handle_error:
 486                         if (record_type != TLS_RECORD_TYPE_DATA) {
 487                                 /* avoid sending partial
 488                                  * record with type !=
 489                                  * application_data
 490                                  */
 491                                 size = orig_size;
 492                                 destroy_record(record);
 493                                 ctx->open_record = NULL;
 494                         } else if (record->len > prot->prepend_size) {
 495                                 goto last_record;
 496                         }
 497
 498                         break;
 499                 }
 500
 501                 record = ctx->open_record;
 502
 503                 copy = min_t(size_t, size, max_open_record_len - record->len);
 504                 if (copy && zc_page) {
 505                         struct page_frag zc_pfrag;
 506
 507                         zc_pfrag.page = zc_page;
 508                         zc_pfrag.offset = iter_offset.offset;
 509                         zc_pfrag.size = copy;
 510                         tls_append_frag(record, &zc_pfrag, copy);
 511                 } else if (copy) {
 512                         copy = min_t(size_t, copy, pfrag->size - pfrag->offset);
 513
 514                         rc = tls_device_copy_data(page_address(pfrag->page) +
 515                                                   pfrag->offset, copy,
 516                                                   iter_offset.msg_iter);
 517                         if (rc)
 518                                 goto handle_error;
 519                         tls_append_frag(record, pfrag, copy);
 520                 }
 521
 522                 size -= copy;
 523                 if (!size) {
 524 last_record:
 525                         tls_push_record_flags = flags;
 526                         if (flags & (MSG_SENDPAGE_NOTLAST | MSG_MORE)) {
 527                                 more = true;
 528                                 break;
 529                         }
 530
 531                         done = true;
 532                 }
 533
 534                 if (done || record->len >= max_open_record_len ||
 535                     (record->num_frags >= MAX_SKB_FRAGS - 1)) {
 536                         rc = tls_device_record_close(sk, tls_ctx, record,
 537                                                      pfrag, record_type);
 538                         if (rc) {
 539                                 if (rc > 0) {
 540                                         size += rc;
 541                                 } else {
 542                                         size = orig_size;
 543                                         destroy_record(record);
 544                                         ctx->open_record = NULL;
 545                                         break;
 546                                 }
 547                         }
 548
 549                         rc = tls_push_record(sk,
 550                                              tls_ctx,
 551                                              ctx,
 552                                              record,
 553                                              tls_push_record_flags);
 554                         if (rc < 0)
 555                                 break;
 556                 }
 557         } while (!done);
 558
 559         tls_ctx->pending_open_record_frags = more;
 560
 561         if (orig_size - size > 0)
 562                 rc = orig_size - size;
 563
 564         return rc;
 565 }
 566
 567 int tls_device_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
 568 {
 569         unsigned char record_type = TLS_RECORD_TYPE_DATA;
 570         struct tls_context *tls_ctx = tls_get_ctx(sk);
 571         union tls_iter_offset iter;
 572         int rc;
 573
 574         mutex_lock(&tls_ctx->tx_lock);
 575         lock_sock(sk);
 576
 577         if (unlikely(msg->msg_controllen)) {
 578                 rc = tls_process_cmsg(sk, msg, &record_type);
 579                 if (rc)
 580                         goto out;
 581         }
 582
 583         iter.msg_iter = &msg->msg_iter;
 584         rc = tls_push_data(sk, iter, size, msg->msg_flags, record_type, NULL);
 585
 586 out:
 587         release_sock(sk);
 588         mutex_unlock(&tls_ctx->tx_lock);
 589         return rc;
 590 }
 591
 592 int tls_device_sendpage(struct sock *sk, struct page *page,
 593                         int offset, size_t size, int flags)
 594 {
 595         struct tls_context *tls_ctx = tls_get_ctx(sk);
 596         union tls_iter_offset iter_offset;
 597         struct iov_iter msg_iter;
 598         char *kaddr;
 599         struct kvec iov;
 600         int rc;
 601
 602         if (flags & MSG_SENDPAGE_NOTLAST)
 603                 flags |= MSG_MORE;
 604
 605         mutex_lock(&tls_ctx->tx_lock);
 606         lock_sock(sk);
 607
 608         if (flags & MSG_OOB) {
 609                 rc = -EOPNOTSUPP;
 610                 goto out;
 611         }
 612
 613         if (tls_ctx->zerocopy_sendfile) {
 614                 iter_offset.offset = offset;
 615                 rc = tls_push_data(sk, iter_offset, size,
 616                                    flags, TLS_RECORD_TYPE_DATA, page);
 617                 goto out;
 618         }
 619
 620         kaddr = kmap(page);
 621         iov.iov_base = kaddr + offset;
 622         iov.iov_len = size;
 623         iov_iter_kvec(&msg_iter, ITER_SOURCE, &iov, 1, size);
 624         iter_offset.msg_iter = &msg_iter;
 625         rc = tls_push_data(sk, iter_offset, size, flags, TLS_RECORD_TYPE_DATA,
 626                            NULL);
 627         kunmap(page);
 628
 629 out:
 630         release_sock(sk);
 631         mutex_unlock(&tls_ctx->tx_lock);
 632         return rc;
 633 }
 634
 635 struct tls_record_info *tls_get_record(struct tls_offload_context_tx *context,
 636                                        u32 seq, u64 *p_record_sn)
 637 {
 638         u64 record_sn = context->hint_record_sn;
 639         struct tls_record_info *info, *last;
 640
 641         info = context->retransmit_hint;
 642         if (!info ||
 643             before(seq, info->end_seq - info->len)) {
 644                 /* if retransmit_hint is irrelevant start
 645                  * from the beginning of the list
 646                  */
 647                 info = list_first_entry_or_null(&context->records_list,
 648                                                 struct tls_record_info, list);
 649                 if (!info)
 650                         return NULL;
 651                 /* send the start_marker record if seq number is before the
 652                  * tls offload start marker sequence number. This record is
 653                  * required to handle TCP packets which are before TLS offload
 654                  * started.
 655                  *  And if it's not start marker, look if this seq number
 656                  * belongs to the list.
 657                  */
 658                 if (likely(!tls_record_is_start_marker(info))) {
 659                         /* we have the first record, get the last record to see
 660                          * if this seq number belongs to the list.
 661                          */
 662                         last = list_last_entry(&context->records_list,
 663                                                struct tls_record_info, list);
 664
 665                         if (!between(seq, tls_record_start_seq(info),
 666                                      last->end_seq))
 667                                 return NULL;
 668                 }
 669                 record_sn = context->unacked_record_sn;
 670         }
 671
 672         /* We just need the _rcu for the READ_ONCE() */
 673         rcu_read_lock();
 674         list_for_each_entry_from_rcu(info, &context->records_list, list) {
 675                 if (before(seq, info->end_seq)) {
 676                         if (!context->retransmit_hint ||
 677                             after(info->end_seq,
 678                                   context->retransmit_hint->end_seq)) {
 679                                 context->hint_record_sn = record_sn;
 680                                 context->retransmit_hint = info;
 681                         }
 682                         *p_record_sn = record_sn;
 683                         goto exit_rcu_unlock;
 684                 }
 685                 record_sn++;
 686         }
 687         info = NULL;
 688
 689 exit_rcu_unlock:
 690         rcu_read_unlock();
 691         return info;
 692 }
 693 EXPORT_SYMBOL(tls_get_record);
 694
 695 static int tls_device_push_pending_record(struct sock *sk, int flags)
 696 {
 697         union tls_iter_offset iter;
 698         struct iov_iter msg_iter;
 699
 700         iov_iter_kvec(&msg_iter, ITER_SOURCE, NULL, 0, 0);
 701         iter.msg_iter = &msg_iter;
 702         return tls_push_data(sk, iter, 0, flags, TLS_RECORD_TYPE_DATA, NULL);
 703 }
 704
 705 void tls_device_write_space(struct sock *sk, struct tls_context *ctx)
 706 {
 707         if (tls_is_partially_sent_record(ctx)) {
 708                 gfp_t sk_allocation = sk->sk_allocation;
 709
 710                 WARN_ON_ONCE(sk->sk_write_pending);
 711
 712                 sk->sk_allocation = GFP_ATOMIC;
 713                 tls_push_partial_record(sk, ctx,
 714                                         MSG_DONTWAIT | MSG_NOSIGNAL |
 715                                         MSG_SENDPAGE_DECRYPTED);
 716                 sk->sk_allocation = sk_allocation;
 717         }
 718 }
 719
 720 static void tls_device_resync_rx(struct tls_context *tls_ctx,
 721                                  struct sock *sk, u32 seq, u8 *rcd_sn)
 722 {
 723         struct tls_offload_context_rx *rx_ctx = tls_offload_ctx_rx(tls_ctx);
 724         struct net_device *netdev;
 725
 726         trace_tls_device_rx_resync_send(sk, seq, rcd_sn, rx_ctx->resync_type);
 727         rcu_read_lock();
 728         netdev = rcu_dereference(tls_ctx->netdev);
 729         if (netdev)
 730                 netdev->tlsdev_ops->tls_dev_resync(netdev, sk, seq, rcd_sn,
 731                                                    TLS_OFFLOAD_CTX_DIR_RX);
 732         rcu_read_unlock();
 733         TLS_INC_STATS(sock_net(sk), LINUX_MIB_TLSRXDEVICERESYNC);
 734 }
 735
 736 static bool
 737 tls_device_rx_resync_async(struct tls_offload_resync_async *resync_async,
 738                            s64 resync_req, u32 *seq, u16 *rcd_delta)
 739 {
 740         u32 is_async = resync_req & RESYNC_REQ_ASYNC;
 741         u32 req_seq = resync_req >> 32;
 742         u32 req_end = req_seq + ((resync_req >> 16) & 0xffff);
 743         u16 i;
 744
 745         *rcd_delta = 0;
 746
 747         if (is_async) {
 748                 /* shouldn't get to wraparound:
 749                  * too long in async stage, something bad happened
 750                  */
 751                 if (WARN_ON_ONCE(resync_async->rcd_delta == USHRT_MAX))
 752                         return false;
 753
 754                 /* asynchronous stage: log all headers seq such that
 755                  * req_seq <= seq <= end_seq, and wait for real resync request
 756                  */
 757                 if (before(*seq, req_seq))
 758                         return false;
 759                 if (!after(*seq, req_end) &&
 760                     resync_async->loglen < TLS_DEVICE_RESYNC_ASYNC_LOGMAX)
 761                         resync_async->log[resync_async->loglen++] = *seq;
 762
 763                 resync_async->rcd_delta++;
 764
 765                 return false;
 766         }
 767
 768         /* synchronous stage: check against the logged entries and
 769          * proceed to check the next entries if no match was found
 770          */
 771         for (i = 0; i < resync_async->loglen; i++)
 772                 if (req_seq == resync_async->log[i] &&
 773                     atomic64_try_cmpxchg(&resync_async->req, &resync_req, 0)) {
 774                         *rcd_delta = resync_async->rcd_delta - i;
 775                         *seq = req_seq;
 776                         resync_async->loglen = 0;
 777                         resync_async->rcd_delta = 0;
 778                         return true;
 779                 }
 780
 781         resync_async->loglen = 0;
 782         resync_async->rcd_delta = 0;
 783
 784         if (req_seq == *seq &&
 785             atomic64_try_cmpxchg(&resync_async->req,
 786                                  &resync_req, 0))
 787                 return true;
 788
 789         return false;
 790 }
 791
 792 void tls_device_rx_resync_new_rec(struct sock *sk, u32 rcd_len, u32 seq)
 793 {
 794         struct tls_context *tls_ctx = tls_get_ctx(sk);
 795         struct tls_offload_context_rx *rx_ctx;
 796         u8 rcd_sn[TLS_MAX_REC_SEQ_SIZE];
 797         u32 sock_data, is_req_pending;
 798         struct tls_prot_info *prot;
 799         s64 resync_req;
 800         u16 rcd_delta;
 801         u32 req_seq;
 802
 803         if (tls_ctx->rx_conf != TLS_HW)
 804                 return;
 805         if (unlikely(test_bit(TLS_RX_DEV_DEGRADED, &tls_ctx->flags)))
 806                 return;
 807
 808         prot = &tls_ctx->prot_info;
 809         rx_ctx = tls_offload_ctx_rx(tls_ctx);
 810         memcpy(rcd_sn, tls_ctx->rx.rec_seq, prot->rec_seq_size);
 811
 812         switch (rx_ctx->resync_type) {
 813         case TLS_OFFLOAD_SYNC_TYPE_DRIVER_REQ:
 814                 resync_req = atomic64_read(&rx_ctx->resync_req);
 815                 req_seq = resync_req >> 32;
 816                 seq += TLS_HEADER_SIZE - 1;
 817                 is_req_pending = resync_req;
 818
 819                 if (likely(!is_req_pending) || req_seq != seq ||
 820                     !atomic64_try_cmpxchg(&rx_ctx->resync_req, &resync_req, 0))
 821                         return;
 822                 break;
 823         case TLS_OFFLOAD_SYNC_TYPE_CORE_NEXT_HINT:
 824                 if (likely(!rx_ctx->resync_nh_do_now))
 825                         return;
 826
 827                 /* head of next rec is already in, note that the sock_inq will
 828                  * include the currently parsed message when called from parser
 829                  */
 830                 sock_data = tcp_inq(sk);
 831                 if (sock_data > rcd_len) {
 832                         trace_tls_device_rx_resync_nh_delay(sk, sock_data,
 833                                                             rcd_len);
 834                         return;
 835                 }
 836
 837                 rx_ctx->resync_nh_do_now = 0;
 838                 seq += rcd_len;
 839                 tls_bigint_increment(rcd_sn, prot->rec_seq_size);
 840                 break;
 841         case TLS_OFFLOAD_SYNC_TYPE_DRIVER_REQ_ASYNC:
 842                 resync_req = atomic64_read(&rx_ctx->resync_async->req);
 843                 is_req_pending = resync_req;
 844                 if (likely(!is_req_pending))
 845                         return;
 846
 847                 if (!tls_device_rx_resync_async(rx_ctx->resync_async,
 848                                                 resync_req, &seq, &rcd_delta))
 849                         return;
 850                 tls_bigint_subtract(rcd_sn, rcd_delta);
 851                 break;
 852         }
 853
 854         tls_device_resync_rx(tls_ctx, sk, seq, rcd_sn);
 855 }
 856
 857 static void tls_device_core_ctrl_rx_resync(struct tls_context *tls_ctx,
 858                                            struct tls_offload_context_rx *ctx,
 859                                            struct sock *sk, struct sk_buff *skb)
 860 {
 861         struct strp_msg *rxm;
 862
 863         /* device will request resyncs by itself based on stream scan */
 864         if (ctx->resync_type != TLS_OFFLOAD_SYNC_TYPE_CORE_NEXT_HINT)
 865                 return;
 866         /* already scheduled */
 867         if (ctx->resync_nh_do_now)
 868                 return;
 869         /* seen decrypted fragments since last fully-failed record */
 870         if (ctx->resync_nh_reset) {
 871                 ctx->resync_nh_reset = 0;
 872                 ctx->resync_nh.decrypted_failed = 1;
 873                 ctx->resync_nh.decrypted_tgt = TLS_DEVICE_RESYNC_NH_START_IVAL;
 874                 return;
 875         }
 876
 877         if (++ctx->resync_nh.decrypted_failed <= ctx->resync_nh.decrypted_tgt)
 878                 return;
 879
 880         /* doing resync, bump the next target in case it fails */
 881         if (ctx->resync_nh.decrypted_tgt < TLS_DEVICE_RESYNC_NH_MAX_IVAL)
 882                 ctx->resync_nh.decrypted_tgt *= 2;
 883         else
 884                 ctx->resync_nh.decrypted_tgt += TLS_DEVICE_RESYNC_NH_MAX_IVAL;
 885
 886         rxm = strp_msg(skb);
 887
 888         /* head of next rec is already in, parser will sync for us */
 889         if (tcp_inq(sk) > rxm->full_len) {
 890                 trace_tls_device_rx_resync_nh_schedule(sk);
 891                 ctx->resync_nh_do_now = 1;
 892         } else {
 893                 struct tls_prot_info *prot = &tls_ctx->prot_info;
 894                 u8 rcd_sn[TLS_MAX_REC_SEQ_SIZE];
 895
 896                 memcpy(rcd_sn, tls_ctx->rx.rec_seq, prot->rec_seq_size);
 897                 tls_bigint_increment(rcd_sn, prot->rec_seq_size);
 898
 899                 tls_device_resync_rx(tls_ctx, sk, tcp_sk(sk)->copied_seq,
 900                                      rcd_sn);
 901         }
 902 }
 903
 904 static int
 905 tls_device_reencrypt(struct sock *sk, struct tls_context *tls_ctx)
 906 {
 907         struct tls_sw_context_rx *sw_ctx = tls_sw_ctx_rx(tls_ctx);
 908         const struct tls_cipher_size_desc *cipher_sz;
 909         int err, offset, copy, data_len, pos;
 910         struct sk_buff *skb, *skb_iter;
 911         struct scatterlist sg[1];
 912         struct strp_msg *rxm;
 913         char *orig_buf, *buf;
 914
 915         switch (tls_ctx->crypto_recv.info.cipher_type) {
 916         case TLS_CIPHER_AES_GCM_128:
 917         case TLS_CIPHER_AES_GCM_256:
 918                 break;
 919         default:
 920                 return -EINVAL;
 921         }
 922         cipher_sz = &tls_cipher_size_desc[tls_ctx->crypto_recv.info.cipher_type];
 923
 924         rxm = strp_msg(tls_strp_msg(sw_ctx));
 925         orig_buf = kmalloc(rxm->full_len + TLS_HEADER_SIZE + cipher_sz->iv,
 926                            sk->sk_allocation);
 927         if (!orig_buf)
 928                 return -ENOMEM;
 929         buf = orig_buf;
 930
 931         err = tls_strp_msg_cow(sw_ctx);
 932         if (unlikely(err))
 933                 goto free_buf;
 934
 935         skb = tls_strp_msg(sw_ctx);
 936         rxm = strp_msg(skb);
 937         offset = rxm->offset;
 938
 939         sg_init_table(sg, 1);
 940         sg_set_buf(&sg[0], buf,
 941                    rxm->full_len + TLS_HEADER_SIZE + cipher_sz->iv);
 942         err = skb_copy_bits(skb, offset, buf, TLS_HEADER_SIZE + cipher_sz->iv);
 943         if (err)
 944                 goto free_buf;
 945
 946         /* We are interested only in the decrypted data not the auth */
 947         err = decrypt_skb(sk, sg);
 948         if (err != -EBADMSG)
 949                 goto free_buf;
 950         else
 951                 err = 0;
 952
 953         data_len = rxm->full_len - cipher_sz->tag;
 954
 955         if (skb_pagelen(skb) > offset) {
 956                 copy = min_t(int, skb_pagelen(skb) - offset, data_len);
 957
 958                 if (skb->decrypted) {
 959                         err = skb_store_bits(skb, offset, buf, copy);
 960                         if (err)
 961                                 goto free_buf;
 962                 }
 963
 964                 offset += copy;
 965                 buf += copy;
 966         }
 967
 968         pos = skb_pagelen(skb);
 969         skb_walk_frags(skb, skb_iter) {
 970                 int frag_pos;
 971
 972                 /* Practically all frags must belong to msg if reencrypt
 973                  * is needed with current strparser and coalescing logic,
 974                  * but strparser may "get optimized", so let's be safe.
 975                  */
 976                 if (pos + skb_iter->len <= offset)
 977                         goto done_with_frag;
 978                 if (pos >= data_len + rxm->offset)
 979                         break;
 980
 981                 frag_pos = offset - pos;
 982                 copy = min_t(int, skb_iter->len - frag_pos,
 983                              data_len + rxm->offset - offset);
 984
 985                 if (skb_iter->decrypted) {
 986                         err = skb_store_bits(skb_iter, frag_pos, buf, copy);
 987                         if (err)
 988                                 goto free_buf;
 989                 }
 990
 991                 offset += copy;
 992                 buf += copy;
 993 done_with_frag:
 994                 pos += skb_iter->len;
 995         }
 996
 997 free_buf:
 998         kfree(orig_buf);
 999         return err;
1000 }
1001
1002 int tls_device_decrypted(struct sock *sk, struct tls_context *tls_ctx)
1003 {
1004         struct tls_offload_context_rx *ctx = tls_offload_ctx_rx(tls_ctx);
1005         struct tls_sw_context_rx *sw_ctx = tls_sw_ctx_rx(tls_ctx);
1006         struct sk_buff *skb = tls_strp_msg(sw_ctx);
1007         struct strp_msg *rxm = strp_msg(skb);
1008         int is_decrypted = skb->decrypted;
1009         int is_encrypted = !is_decrypted;
1010         struct sk_buff *skb_iter;
1011         int left;
1012
1013         left = rxm->full_len - skb->len;
1014         /* Check if all the data is decrypted already */
1015         skb_iter = skb_shinfo(skb)->frag_list;
1016         while (skb_iter && left > 0) {
1017                 is_decrypted &= skb_iter->decrypted;
1018                 is_encrypted &= !skb_iter->decrypted;
1019
1020                 left -= skb_iter->len;
1021                 skb_iter = skb_iter->next;
1022         }
1023
1024         trace_tls_device_decrypted(sk, tcp_sk(sk)->copied_seq - rxm->full_len,
1025                                    tls_ctx->rx.rec_seq, rxm->full_len,
1026                                    is_encrypted, is_decrypted);
1027
1028         if (unlikely(test_bit(TLS_RX_DEV_DEGRADED, &tls_ctx->flags))) {
1029                 if (likely(is_encrypted || is_decrypted))
1030                         return is_decrypted;
1031
1032                 /* After tls_device_down disables the offload, the next SKB will
1033                  * likely have initial fragments decrypted, and final ones not
1034                  * decrypted. We need to reencrypt that single SKB.
1035                  */
1036                 return tls_device_reencrypt(sk, tls_ctx);
1037         }
1038
1039         /* Return immediately if the record is either entirely plaintext or
1040          * entirely ciphertext. Otherwise handle reencrypt partially decrypted
1041          * record.
1042          */
1043         if (is_decrypted) {
1044                 ctx->resync_nh_reset = 1;
1045                 return is_decrypted;
1046         }
1047         if (is_encrypted) {
1048                 tls_device_core_ctrl_rx_resync(tls_ctx, ctx, sk, skb);
1049                 return 0;
1050         }
1051
1052         ctx->resync_nh_reset = 1;
1053         return tls_device_reencrypt(sk, tls_ctx);
1054 }
1055
1056 static void tls_device_attach(struct tls_context *ctx, struct sock *sk,
1057                               struct net_device *netdev)
1058 {
1059         if (sk->sk_destruct != tls_device_sk_destruct) {
1060                 refcount_set(&ctx->refcount, 1);
1061                 dev_hold(netdev);
1062                 RCU_INIT_POINTER(ctx->netdev, netdev);
1063                 spin_lock_irq(&tls_device_lock);
1064                 list_add_tail(&ctx->list, &tls_device_list);
1065                 spin_unlock_irq(&tls_device_lock);
1066
1067                 ctx->sk_destruct = sk->sk_destruct;
1068                 smp_store_release(&sk->sk_destruct, tls_device_sk_destruct);
1069         }
1070 }
1071
1072 int tls_set_device_offload(struct sock *sk, struct tls_context *ctx)
1073 {
1074         struct tls_context *tls_ctx = tls_get_ctx(sk);
1075         struct tls_prot_info *prot = &tls_ctx->prot_info;
1076         const struct tls_cipher_size_desc *cipher_sz;
1077         struct tls_record_info *start_marker_record;
1078         struct tls_offload_context_tx *offload_ctx;
1079         struct tls_crypto_info *crypto_info;
1080         struct net_device *netdev;
1081         char *iv, *rec_seq;
1082         struct sk_buff *skb;
1083         __be64 rcd_sn;
1084         int rc;
1085
1086         if (!ctx)
1087                 return -EINVAL;
1088
1089         if (ctx->priv_ctx_tx)
1090                 return -EEXIST;
1091
1092         netdev = get_netdev_for_sock(sk);
1093         if (!netdev) {
1094                 pr_err_ratelimited("%s: netdev not found\n", __func__);
1095                 return -EINVAL;
1096         }
1097
1098         if (!(netdev->features & NETIF_F_HW_TLS_TX)) {
1099                 rc = -EOPNOTSUPP;
1100                 goto release_netdev;
1101         }
1102
1103         crypto_info = &ctx->crypto_send.info;
1104         if (crypto_info->version != TLS_1_2_VERSION) {
1105                 rc = -EOPNOTSUPP;
1106                 goto release_netdev;
1107         }
1108
1109         switch (crypto_info->cipher_type) {
1110         case TLS_CIPHER_AES_GCM_128:
1111                 iv = ((struct tls12_crypto_info_aes_gcm_128 *)crypto_info)->iv;
1112                 rec_seq =
1113                  ((struct tls12_crypto_info_aes_gcm_128 *)crypto_info)->rec_seq;
1114                 break;
1115         case TLS_CIPHER_AES_GCM_256:
1116                 iv = ((struct tls12_crypto_info_aes_gcm_256 *)crypto_info)->iv;
1117                 rec_seq =
1118                  ((struct tls12_crypto_info_aes_gcm_256 *)crypto_info)->rec_seq;
1119                 break;
1120         default:
1121                 rc = -EINVAL;
1122                 goto release_netdev;
1123         }
1124         cipher_sz = &tls_cipher_size_desc[crypto_info->cipher_type];
1125
1126         /* Sanity-check the rec_seq_size for stack allocations */
1127         if (cipher_sz->rec_seq > TLS_MAX_REC_SEQ_SIZE) {
1128                 rc = -EINVAL;
1129                 goto release_netdev;
1130         }
1131
1132         prot->version = crypto_info->version;
1133         prot->cipher_type = crypto_info->cipher_type;
1134         prot->prepend_size = TLS_HEADER_SIZE + cipher_sz->iv;
1135         prot->tag_size = cipher_sz->tag;
1136         prot->overhead_size = prot->prepend_size + prot->tag_size;
1137         prot->iv_size = cipher_sz->iv;
1138         prot->salt_size = cipher_sz->salt;
1139         ctx->tx.iv = kmalloc(cipher_sz->iv + cipher_sz->salt, GFP_KERNEL);
1140         if (!ctx->tx.iv) {
1141                 rc = -ENOMEM;
1142                 goto release_netdev;
1143         }
1144
1145         memcpy(ctx->tx.iv + cipher_sz->salt, iv, cipher_sz->iv);
1146
1147         prot->rec_seq_size = cipher_sz->rec_seq;
1148         ctx->tx.rec_seq = kmemdup(rec_seq, cipher_sz->rec_seq, GFP_KERNEL);
1149         if (!ctx->tx.rec_seq) {
1150                 rc = -ENOMEM;
1151                 goto free_iv;
1152         }
1153
1154         start_marker_record = kmalloc(sizeof(*start_marker_record), GFP_KERNEL);
1155         if (!start_marker_record) {
1156                 rc = -ENOMEM;
1157                 goto free_rec_seq;
1158         }
1159
1160         offload_ctx = kzalloc(TLS_OFFLOAD_CONTEXT_SIZE_TX, GFP_KERNEL);
1161         if (!offload_ctx) {
1162                 rc = -ENOMEM;
1163                 goto free_marker_record;
1164         }
1165
1166         rc = tls_sw_fallback_init(sk, offload_ctx, crypto_info);
1167         if (rc)
1168                 goto free_offload_ctx;
1169
1170         /* start at rec_seq - 1 to account for the start marker record */
1171         memcpy(&rcd_sn, ctx->tx.rec_seq, sizeof(rcd_sn));
1172         offload_ctx->unacked_record_sn = be64_to_cpu(rcd_sn) - 1;
1173
1174         start_marker_record->end_seq = tcp_sk(sk)->write_seq;
1175         start_marker_record->len = 0;
1176         start_marker_record->num_frags = 0;
1177
1178         INIT_WORK(&offload_ctx->destruct_work, tls_device_tx_del_task);
1179         offload_ctx->ctx = ctx;
1180
1181         INIT_LIST_HEAD(&offload_ctx->records_list);
1182         list_add_tail(&start_marker_record->list, &offload_ctx->records_list);
1183         spin_lock_init(&offload_ctx->lock);
1184         sg_init_table(offload_ctx->sg_tx_data,
1185                       ARRAY_SIZE(offload_ctx->sg_tx_data));
1186
1187         clean_acked_data_enable(inet_csk(sk), &tls_icsk_clean_acked);
1188         ctx->push_pending_record = tls_device_push_pending_record;
1189
1190         /* TLS offload is greatly simplified if we don't send
1191          * SKBs where only part of the payload needs to be encrypted.
1192          * So mark the last skb in the write queue as end of record.
1193          */
1194         skb = tcp_write_queue_tail(sk);
1195         if (skb)
1196                 TCP_SKB_CB(skb)->eor = 1;
1197
1198         /* Avoid offloading if the device is down
1199          * We don't want to offload new flows after
1200          * the NETDEV_DOWN event
1201          *
1202          * device_offload_lock is taken in tls_devices's NETDEV_DOWN
1203          * handler thus protecting from the device going down before
1204          * ctx was added to tls_device_list.
1205          */
1206         down_read(&device_offload_lock);
1207         if (!(netdev->flags & IFF_UP)) {
1208                 rc = -EINVAL;
1209                 goto release_lock;
1210         }
1211
1212         ctx->priv_ctx_tx = offload_ctx;
1213         rc = netdev->tlsdev_ops->tls_dev_add(netdev, sk, TLS_OFFLOAD_CTX_DIR_TX,
1214                                              &ctx->crypto_send.info,
1215                                              tcp_sk(sk)->write_seq);
1216         trace_tls_device_offload_set(sk, TLS_OFFLOAD_CTX_DIR_TX,
1217                                      tcp_sk(sk)->write_seq, rec_seq, rc);
1218         if (rc)
1219                 goto release_lock;
1220
1221         tls_device_attach(ctx, sk, netdev);
1222         up_read(&device_offload_lock);
1223
1224         /* following this assignment tls_is_sk_tx_device_offloaded
1225          * will return true and the context might be accessed
1226          * by the netdev's xmit function.
1227          */
1228         smp_store_release(&sk->sk_validate_xmit_skb, tls_validate_xmit_skb);
1229         dev_put(netdev);
1230
1231         return 0;
1232
1233 release_lock:
1234         up_read(&device_offload_lock);
1235         clean_acked_data_disable(inet_csk(sk));
1236         crypto_free_aead(offload_ctx->aead_send);
1237 free_offload_ctx:
1238         kfree(offload_ctx);
1239         ctx->priv_ctx_tx = NULL;
1240 free_marker_record:
1241         kfree(start_marker_record);
1242 free_rec_seq:
1243         kfree(ctx->tx.rec_seq);
1244 free_iv:
1245         kfree(ctx->tx.iv);
1246 release_netdev:
1247         dev_put(netdev);
1248         return rc;
1249 }
1250
1251 int tls_set_device_offload_rx(struct sock *sk, struct tls_context *ctx)
1252 {
1253         struct tls12_crypto_info_aes_gcm_128 *info;
1254         struct tls_offload_context_rx *context;
1255         struct net_device *netdev;
1256         int rc = 0;
1257
1258         if (ctx->crypto_recv.info.version != TLS_1_2_VERSION)
1259                 return -EOPNOTSUPP;
1260
1261         netdev = get_netdev_for_sock(sk);
1262         if (!netdev) {
1263                 pr_err_ratelimited("%s: netdev not found\n", __func__);
1264                 return -EINVAL;
1265         }
1266
1267         if (!(netdev->features & NETIF_F_HW_TLS_RX)) {
1268                 rc = -EOPNOTSUPP;
1269                 goto release_netdev;
1270         }
1271
1272         /* Avoid offloading if the device is down
1273          * We don't want to offload new flows after
1274          * the NETDEV_DOWN event
1275          *
1276          * device_offload_lock is taken in tls_devices's NETDEV_DOWN
1277          * handler thus protecting from the device going down before
1278          * ctx was added to tls_device_list.
1279          */
1280         down_read(&device_offload_lock);
1281         if (!(netdev->flags & IFF_UP)) {
1282                 rc = -EINVAL;
1283                 goto release_lock;
1284         }
1285
1286         context = kzalloc(TLS_OFFLOAD_CONTEXT_SIZE_RX, GFP_KERNEL);
1287         if (!context) {
1288                 rc = -ENOMEM;
1289                 goto release_lock;
1290         }
1291         context->resync_nh_reset = 1;
1292
1293         ctx->priv_ctx_rx = context;
1294         rc = tls_set_sw_offload(sk, ctx, 0);
1295         if (rc)
1296                 goto release_ctx;
1297
1298         rc = netdev->tlsdev_ops->tls_dev_add(netdev, sk, TLS_OFFLOAD_CTX_DIR_RX,
1299                                              &ctx->crypto_recv.info,
1300                                              tcp_sk(sk)->copied_seq);
1301         info = (void *)&ctx->crypto_recv.info;
1302         trace_tls_device_offload_set(sk, TLS_OFFLOAD_CTX_DIR_RX,
1303                                      tcp_sk(sk)->copied_seq, info->rec_seq, rc);
1304         if (rc)
1305                 goto free_sw_resources;
1306
1307         tls_device_attach(ctx, sk, netdev);
1308         up_read(&device_offload_lock);
1309
1310         dev_put(netdev);
1311
1312         return 0;
1313
1314 free_sw_resources:
1315         up_read(&device_offload_lock);
1316         tls_sw_free_resources_rx(sk);
1317         down_read(&device_offload_lock);
1318 release_ctx:
1319         ctx->priv_ctx_rx = NULL;
1320 release_lock:
1321         up_read(&device_offload_lock);
1322 release_netdev:
1323         dev_put(netdev);
1324         return rc;
1325 }
1326
1327 void tls_device_offload_cleanup_rx(struct sock *sk)
1328 {
1329         struct tls_context *tls_ctx = tls_get_ctx(sk);
1330         struct net_device *netdev;
1331
1332         down_read(&device_offload_lock);
1333         netdev = rcu_dereference_protected(tls_ctx->netdev,
1334                                            lockdep_is_held(&device_offload_lock));
1335         if (!netdev)
1336                 goto out;
1337
1338         netdev->tlsdev_ops->tls_dev_del(netdev, tls_ctx,
1339                                         TLS_OFFLOAD_CTX_DIR_RX);
1340
1341         if (tls_ctx->tx_conf != TLS_HW) {
1342                 dev_put(netdev);
1343                 rcu_assign_pointer(tls_ctx->netdev, NULL);
1344         } else {
1345                 set_bit(TLS_RX_DEV_CLOSED, &tls_ctx->flags);
1346         }
1347 out:
1348         up_read(&device_offload_lock);
1349         tls_sw_release_resources_rx(sk);
1350 }
1351
1352 static int tls_device_down(struct net_device *netdev)
1353 {
1354         struct tls_context *ctx, *tmp;
1355         unsigned long flags;
1356         LIST_HEAD(list);
1357
1358         /* Request a write lock to block new offload attempts */
1359         down_write(&device_offload_lock);
1360
1361         spin_lock_irqsave(&tls_device_lock, flags);
1362         list_for_each_entry_safe(ctx, tmp, &tls_device_list, list) {
1363                 struct net_device *ctx_netdev =
1364                         rcu_dereference_protected(ctx->netdev,
1365                                                   lockdep_is_held(&device_offload_lock));
1366
1367                 if (ctx_netdev != netdev ||
1368                     !refcount_inc_not_zero(&ctx->refcount))
1369                         continue;
1370
1371                 list_move(&ctx->list, &list);
1372         }
1373         spin_unlock_irqrestore(&tls_device_lock, flags);
1374
1375         list_for_each_entry_safe(ctx, tmp, &list, list) {
1376                 /* Stop offloaded TX and switch to the fallback.
1377                  * tls_is_sk_tx_device_offloaded will return false.
1378                  */
1379                 WRITE_ONCE(ctx->sk->sk_validate_xmit_skb, tls_validate_xmit_skb_sw);
1380
1381                 /* Stop the RX and TX resync.
1382                  * tls_dev_resync must not be called after tls_dev_del.
1383                  */
1384                 rcu_assign_pointer(ctx->netdev, NULL);
1385
1386                 /* Start skipping the RX resync logic completely. */
1387                 set_bit(TLS_RX_DEV_DEGRADED, &ctx->flags);
1388
1389                 /* Sync with inflight packets. After this point:
1390                  * TX: no non-encrypted packets will be passed to the driver.
1391                  * RX: resync requests from the driver will be ignored.
1392                  */
1393                 synchronize_net();
1394
1395                 /* Release the offload context on the driver side. */
1396                 if (ctx->tx_conf == TLS_HW)
1397                         netdev->tlsdev_ops->tls_dev_del(netdev, ctx,
1398                                                         TLS_OFFLOAD_CTX_DIR_TX);
1399                 if (ctx->rx_conf == TLS_HW &&
1400                     !test_bit(TLS_RX_DEV_CLOSED, &ctx->flags))
1401                         netdev->tlsdev_ops->tls_dev_del(netdev, ctx,
1402                                                         TLS_OFFLOAD_CTX_DIR_RX);
1403
1404                 dev_put(netdev);
1405
1406                 /* Move the context to a separate list for two reasons:
1407                  * 1. When the context is deallocated, list_del is called.
1408                  * 2. It's no longer an offloaded context, so we don't want to
1409                  *    run offload-specific code on this context.
1410                  */
1411                 spin_lock_irqsave(&tls_device_lock, flags);
1412                 list_move_tail(&ctx->list, &tls_device_down_list);
1413                 spin_unlock_irqrestore(&tls_device_lock, flags);
1414
1415                 /* Device contexts for RX and TX will be freed in on sk_destruct
1416                  * by tls_device_free_ctx. rx_conf and tx_conf stay in TLS_HW.
1417                  * Now release the ref taken above.
1418                  */
1419                 if (refcount_dec_and_test(&ctx->refcount)) {
1420                         /* sk_destruct ran after tls_device_down took a ref, and
1421                          * it returned early. Complete the destruction here.
1422                          */
1423                         list_del(&ctx->list);
1424                         tls_device_free_ctx(ctx);
1425                 }
1426         }
1427
1428         up_write(&device_offload_lock);
1429
1430         flush_workqueue(destruct_wq);
1431
1432         return NOTIFY_DONE;
1433 }
1434
1435 static int tls_dev_event(struct notifier_block *this, unsigned long event,
1436                          void *ptr)
1437 {
1438         struct net_device *dev = netdev_notifier_info_to_dev(ptr);
1439
1440         if (!dev->tlsdev_ops &&
1441             !(dev->features & (NETIF_F_HW_TLS_RX | NETIF_F_HW_TLS_TX)))
1442                 return NOTIFY_DONE;
1443
1444         switch (event) {
1445         case NETDEV_REGISTER:
1446         case NETDEV_FEAT_CHANGE:
1447                 if (netif_is_bond_master(dev))
1448                         return NOTIFY_DONE;
1449                 if ((dev->features & NETIF_F_HW_TLS_RX) &&
1450                     !dev->tlsdev_ops->tls_dev_resync)
1451                         return NOTIFY_BAD;
1452
1453                 if  (dev->tlsdev_ops &&
1454                      dev->tlsdev_ops->tls_dev_add &&
1455                      dev->tlsdev_ops->tls_dev_del)
1456                         return NOTIFY_DONE;
1457                 else
1458                         return NOTIFY_BAD;
1459         case NETDEV_DOWN:
1460                 return tls_device_down(dev);
1461         }
1462         return NOTIFY_DONE;
1463 }
1464
1465 static struct notifier_block tls_dev_notifier = {
1466         .notifier_call  = tls_dev_event,
1467 };
1468
1469 int __init tls_device_init(void)
1470 {
1471         int err;
1472
1473         destruct_wq = alloc_workqueue("ktls_device_destruct", 0, 0);
1474         if (!destruct_wq)
1475                 return -ENOMEM;
1476
1477         err = register_netdevice_notifier(&tls_dev_notifier);
1478         if (err)
1479                 destroy_workqueue(destruct_wq);
1480
1481         return err;
1482 }
1483
1484 void __exit tls_device_cleanup(void)
1485 {
1486         unregister_netdevice_notifier(&tls_dev_notifier);
1487         destroy_workqueue(destruct_wq);
1488         clean_acked_data_flush();
1489 }