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