net/tls/tls_sw.c

   1 /*
   2  * Copyright (c) 2016-2017, Mellanox Technologies. All rights reserved.
   3  * Copyright (c) 2016-2017, Dave Watson <davejwatson@fb.com>. All rights reserved.
   4  * Copyright (c) 2016-2017, Lance Chao <lancerchao@fb.com>. All rights reserved.
   5  * Copyright (c) 2016, Fridolin Pokorny <fridolin.pokorny@gmail.com>. All rights reserved.
   6  * Copyright (c) 2016, Nikos Mavrogiannopoulos <nmav@gnutls.org>. All rights reserved.
   7  * Copyright (c) 2018, Covalent IO, Inc. http://covalent.io
   8  *
   9  * This software is available to you under a choice of one of two
  10  * licenses.  You may choose to be licensed under the terms of the GNU
  11  * General Public License (GPL) Version 2, available from the file
  12  * COPYING in the main directory of this source tree, or the
  13  * OpenIB.org BSD license below:
  14  *
  15  *     Redistribution and use in source and binary forms, with or
  16  *     without modification, are permitted provided that the following
  17  *     conditions are met:
  18  *
  19  *      - Redistributions of source code must retain the above
  20  *        copyright notice, this list of conditions and the following
  21  *        disclaimer.
  22  *
  23  *      - Redistributions in binary form must reproduce the above
  24  *        copyright notice, this list of conditions and the following
  25  *        disclaimer in the documentation and/or other materials
  26  *        provided with the distribution.
  27  *
  28  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  29  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  30  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  31  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  32  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  33  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  34  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  35  * SOFTWARE.
  36  */
  37
  38 #include <linux/bug.h>
  39 #include <linux/sched/signal.h>
  40 #include <linux/module.h>
  41 #include <linux/splice.h>
  42 #include <crypto/aead.h>
  43
  44 #include <net/strparser.h>
  45 #include <net/tls.h>
  46
  47 noinline void tls_err_abort(struct sock *sk, int err)
  48 {
  49         WARN_ON_ONCE(err >= 0);
  50         /* sk->sk_err should contain a positive error code. */
  51         sk->sk_err = -err;
  52         sk_error_report(sk);
  53 }
  54
  55 static int __skb_nsg(struct sk_buff *skb, int offset, int len,
  56                      unsigned int recursion_level)
  57 {
  58         int start = skb_headlen(skb);
  59         int i, chunk = start - offset;
  60         struct sk_buff *frag_iter;
  61         int elt = 0;
  62
  63         if (unlikely(recursion_level >= 24))
  64                 return -EMSGSIZE;
  65
  66         if (chunk > 0) {
  67                 if (chunk > len)
  68                         chunk = len;
  69                 elt++;
  70                 len -= chunk;
  71                 if (len == 0)
  72                         return elt;
  73                 offset += chunk;
  74         }
  75
  76         for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
  77                 int end;
  78
  79                 WARN_ON(start > offset + len);
  80
  81                 end = start + skb_frag_size(&skb_shinfo(skb)->frags[i]);
  82                 chunk = end - offset;
  83                 if (chunk > 0) {
  84                         if (chunk > len)
  85                                 chunk = len;
  86                         elt++;
  87                         len -= chunk;
  88                         if (len == 0)
  89                                 return elt;
  90                         offset += chunk;
  91                 }
  92                 start = end;
  93         }
  94
  95         if (unlikely(skb_has_frag_list(skb))) {
  96                 skb_walk_frags(skb, frag_iter) {
  97                         int end, ret;
  98
  99                         WARN_ON(start > offset + len);
 100
 101                         end = start + frag_iter->len;
 102                         chunk = end - offset;
 103                         if (chunk > 0) {
 104                                 if (chunk > len)
 105                                         chunk = len;
 106                                 ret = __skb_nsg(frag_iter, offset - start, chunk,
 107                                                 recursion_level + 1);
 108                                 if (unlikely(ret < 0))
 109                                         return ret;
 110                                 elt += ret;
 111                                 len -= chunk;
 112                                 if (len == 0)
 113                                         return elt;
 114                                 offset += chunk;
 115                         }
 116                         start = end;
 117                 }
 118         }
 119         BUG_ON(len);
 120         return elt;
 121 }
 122
 123 /* Return the number of scatterlist elements required to completely map the
 124  * skb, or -EMSGSIZE if the recursion depth is exceeded.
 125  */
 126 static int skb_nsg(struct sk_buff *skb, int offset, int len)
 127 {
 128         return __skb_nsg(skb, offset, len, 0);
 129 }
 130
 131 static int padding_length(struct tls_sw_context_rx *ctx,
 132                           struct tls_prot_info *prot, struct sk_buff *skb)
 133 {
 134         struct strp_msg *rxm = strp_msg(skb);
 135         int sub = 0;
 136
 137         /* Determine zero-padding length */
 138         if (prot->version == TLS_1_3_VERSION) {
 139                 char content_type = 0;
 140                 int err;
 141                 int back = 17;
 142
 143                 while (content_type == 0) {
 144                         if (back > rxm->full_len - prot->prepend_size)
 145                                 return -EBADMSG;
 146                         err = skb_copy_bits(skb,
 147                                             rxm->offset + rxm->full_len - back,
 148                                             &content_type, 1);
 149                         if (err)
 150                                 return err;
 151                         if (content_type)
 152                                 break;
 153                         sub++;
 154                         back++;
 155                 }
 156                 ctx->control = content_type;
 157         }
 158         return sub;
 159 }
 160
 161 static void tls_decrypt_done(struct crypto_async_request *req, int err)
 162 {
 163         struct aead_request *aead_req = (struct aead_request *)req;
 164         struct scatterlist *sgout = aead_req->dst;
 165         struct scatterlist *sgin = aead_req->src;
 166         struct tls_sw_context_rx *ctx;
 167         struct tls_context *tls_ctx;
 168         struct tls_prot_info *prot;
 169         struct scatterlist *sg;
 170         struct sk_buff *skb;
 171         unsigned int pages;
 172         int pending;
 173
 174         skb = (struct sk_buff *)req->data;
 175         tls_ctx = tls_get_ctx(skb->sk);
 176         ctx = tls_sw_ctx_rx(tls_ctx);
 177         prot = &tls_ctx->prot_info;
 178
 179         /* Propagate if there was an err */
 180         if (err) {
 181                 if (err == -EBADMSG)
 182                         TLS_INC_STATS(sock_net(skb->sk),
 183                                       LINUX_MIB_TLSDECRYPTERROR);
 184                 ctx->async_wait.err = err;
 185                 tls_err_abort(skb->sk, err);
 186         } else {
 187                 struct strp_msg *rxm = strp_msg(skb);
 188                 int pad;
 189
 190                 pad = padding_length(ctx, prot, skb);
 191                 if (pad < 0) {
 192                         ctx->async_wait.err = pad;
 193                         tls_err_abort(skb->sk, pad);
 194                 } else {
 195                         rxm->full_len -= pad;
 196                         rxm->offset += prot->prepend_size;
 197                         rxm->full_len -= prot->overhead_size;
 198                 }
 199         }
 200
 201         /* After using skb->sk to propagate sk through crypto async callback
 202          * we need to NULL it again.
 203          */
 204         skb->sk = NULL;
 205
 206
 207         /* Free the destination pages if skb was not decrypted inplace */
 208         if (sgout != sgin) {
 209                 /* Skip the first S/G entry as it points to AAD */
 210                 for_each_sg(sg_next(sgout), sg, UINT_MAX, pages) {
 211                         if (!sg)
 212                                 break;
 213                         put_page(sg_page(sg));
 214                 }
 215         }
 216
 217         kfree(aead_req);
 218
 219         spin_lock_bh(&ctx->decrypt_compl_lock);
 220         pending = atomic_dec_return(&ctx->decrypt_pending);
 221
 222         if (!pending && ctx->async_notify)
 223                 complete(&ctx->async_wait.completion);
 224         spin_unlock_bh(&ctx->decrypt_compl_lock);
 225 }
 226
 227 static int tls_do_decryption(struct sock *sk,
 228                              struct sk_buff *skb,
 229                              struct scatterlist *sgin,
 230                              struct scatterlist *sgout,
 231                              char *iv_recv,
 232                              size_t data_len,
 233                              struct aead_request *aead_req,
 234                              bool async)
 235 {
 236         struct tls_context *tls_ctx = tls_get_ctx(sk);
 237         struct tls_prot_info *prot = &tls_ctx->prot_info;
 238         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
 239         int ret;
 240
 241         aead_request_set_tfm(aead_req, ctx->aead_recv);
 242         aead_request_set_ad(aead_req, prot->aad_size);
 243         aead_request_set_crypt(aead_req, sgin, sgout,
 244                                data_len + prot->tag_size,
 245                                (u8 *)iv_recv);
 246
 247         if (async) {
 248                 /* Using skb->sk to push sk through to crypto async callback
 249                  * handler. This allows propagating errors up to the socket
 250                  * if needed. It _must_ be cleared in the async handler
 251                  * before consume_skb is called. We _know_ skb->sk is NULL
 252                  * because it is a clone from strparser.
 253                  */
 254                 skb->sk = sk;
 255                 aead_request_set_callback(aead_req,
 256                                           CRYPTO_TFM_REQ_MAY_BACKLOG,
 257                                           tls_decrypt_done, skb);
 258                 atomic_inc(&ctx->decrypt_pending);
 259         } else {
 260                 aead_request_set_callback(aead_req,
 261                                           CRYPTO_TFM_REQ_MAY_BACKLOG,
 262                                           crypto_req_done, &ctx->async_wait);
 263         }
 264
 265         ret = crypto_aead_decrypt(aead_req);
 266         if (ret == -EINPROGRESS) {
 267                 if (async)
 268                         return ret;
 269
 270                 ret = crypto_wait_req(ret, &ctx->async_wait);
 271         }
 272
 273         if (async)
 274                 atomic_dec(&ctx->decrypt_pending);
 275
 276         return ret;
 277 }
 278
 279 static void tls_trim_both_msgs(struct sock *sk, int target_size)
 280 {
 281         struct tls_context *tls_ctx = tls_get_ctx(sk);
 282         struct tls_prot_info *prot = &tls_ctx->prot_info;
 283         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 284         struct tls_rec *rec = ctx->open_rec;
 285
 286         sk_msg_trim(sk, &rec->msg_plaintext, target_size);
 287         if (target_size > 0)
 288                 target_size += prot->overhead_size;
 289         sk_msg_trim(sk, &rec->msg_encrypted, target_size);
 290 }
 291
 292 static int tls_alloc_encrypted_msg(struct sock *sk, int len)
 293 {
 294         struct tls_context *tls_ctx = tls_get_ctx(sk);
 295         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 296         struct tls_rec *rec = ctx->open_rec;
 297         struct sk_msg *msg_en = &rec->msg_encrypted;
 298
 299         return sk_msg_alloc(sk, msg_en, len, 0);
 300 }
 301
 302 static int tls_clone_plaintext_msg(struct sock *sk, int required)
 303 {
 304         struct tls_context *tls_ctx = tls_get_ctx(sk);
 305         struct tls_prot_info *prot = &tls_ctx->prot_info;
 306         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 307         struct tls_rec *rec = ctx->open_rec;
 308         struct sk_msg *msg_pl = &rec->msg_plaintext;
 309         struct sk_msg *msg_en = &rec->msg_encrypted;
 310         int skip, len;
 311
 312         /* We add page references worth len bytes from encrypted sg
 313          * at the end of plaintext sg. It is guaranteed that msg_en
 314          * has enough required room (ensured by caller).
 315          */
 316         len = required - msg_pl->sg.size;
 317
 318         /* Skip initial bytes in msg_en's data to be able to use
 319          * same offset of both plain and encrypted data.
 320          */
 321         skip = prot->prepend_size + msg_pl->sg.size;
 322
 323         return sk_msg_clone(sk, msg_pl, msg_en, skip, len);
 324 }
 325
 326 static struct tls_rec *tls_get_rec(struct sock *sk)
 327 {
 328         struct tls_context *tls_ctx = tls_get_ctx(sk);
 329         struct tls_prot_info *prot = &tls_ctx->prot_info;
 330         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 331         struct sk_msg *msg_pl, *msg_en;
 332         struct tls_rec *rec;
 333         int mem_size;
 334
 335         mem_size = sizeof(struct tls_rec) + crypto_aead_reqsize(ctx->aead_send);
 336
 337         rec = kzalloc(mem_size, sk->sk_allocation);
 338         if (!rec)
 339                 return NULL;
 340
 341         msg_pl = &rec->msg_plaintext;
 342         msg_en = &rec->msg_encrypted;
 343
 344         sk_msg_init(msg_pl);
 345         sk_msg_init(msg_en);
 346
 347         sg_init_table(rec->sg_aead_in, 2);
 348         sg_set_buf(&rec->sg_aead_in[0], rec->aad_space, prot->aad_size);
 349         sg_unmark_end(&rec->sg_aead_in[1]);
 350
 351         sg_init_table(rec->sg_aead_out, 2);
 352         sg_set_buf(&rec->sg_aead_out[0], rec->aad_space, prot->aad_size);
 353         sg_unmark_end(&rec->sg_aead_out[1]);
 354
 355         return rec;
 356 }
 357
 358 static void tls_free_rec(struct sock *sk, struct tls_rec *rec)
 359 {
 360         sk_msg_free(sk, &rec->msg_encrypted);
 361         sk_msg_free(sk, &rec->msg_plaintext);
 362         kfree(rec);
 363 }
 364
 365 static void tls_free_open_rec(struct sock *sk)
 366 {
 367         struct tls_context *tls_ctx = tls_get_ctx(sk);
 368         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 369         struct tls_rec *rec = ctx->open_rec;
 370
 371         if (rec) {
 372                 tls_free_rec(sk, rec);
 373                 ctx->open_rec = NULL;
 374         }
 375 }
 376
 377 int tls_tx_records(struct sock *sk, int flags)
 378 {
 379         struct tls_context *tls_ctx = tls_get_ctx(sk);
 380         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 381         struct tls_rec *rec, *tmp;
 382         struct sk_msg *msg_en;
 383         int tx_flags, rc = 0;
 384
 385         if (tls_is_partially_sent_record(tls_ctx)) {
 386                 rec = list_first_entry(&ctx->tx_list,
 387                                        struct tls_rec, list);
 388
 389                 if (flags == -1)
 390                         tx_flags = rec->tx_flags;
 391                 else
 392                         tx_flags = flags;
 393
 394                 rc = tls_push_partial_record(sk, tls_ctx, tx_flags);
 395                 if (rc)
 396                         goto tx_err;
 397
 398                 /* Full record has been transmitted.
 399                  * Remove the head of tx_list
 400                  */
 401                 list_del(&rec->list);
 402                 sk_msg_free(sk, &rec->msg_plaintext);
 403                 kfree(rec);
 404         }
 405
 406         /* Tx all ready records */
 407         list_for_each_entry_safe(rec, tmp, &ctx->tx_list, list) {
 408                 if (READ_ONCE(rec->tx_ready)) {
 409                         if (flags == -1)
 410                                 tx_flags = rec->tx_flags;
 411                         else
 412                                 tx_flags = flags;
 413
 414                         msg_en = &rec->msg_encrypted;
 415                         rc = tls_push_sg(sk, tls_ctx,
 416                                          &msg_en->sg.data[msg_en->sg.curr],
 417                                          0, tx_flags);
 418                         if (rc)
 419                                 goto tx_err;
 420
 421                         list_del(&rec->list);
 422                         sk_msg_free(sk, &rec->msg_plaintext);
 423                         kfree(rec);
 424                 } else {
 425                         break;
 426                 }
 427         }
 428
 429 tx_err:
 430         if (rc < 0 && rc != -EAGAIN)
 431                 tls_err_abort(sk, -EBADMSG);
 432
 433         return rc;
 434 }
 435
 436 static void tls_encrypt_done(struct crypto_async_request *req, int err)
 437 {
 438         struct aead_request *aead_req = (struct aead_request *)req;
 439         struct sock *sk = req->data;
 440         struct tls_context *tls_ctx = tls_get_ctx(sk);
 441         struct tls_prot_info *prot = &tls_ctx->prot_info;
 442         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 443         struct scatterlist *sge;
 444         struct sk_msg *msg_en;
 445         struct tls_rec *rec;
 446         bool ready = false;
 447         int pending;
 448
 449         rec = container_of(aead_req, struct tls_rec, aead_req);
 450         msg_en = &rec->msg_encrypted;
 451
 452         sge = sk_msg_elem(msg_en, msg_en->sg.curr);
 453         sge->offset -= prot->prepend_size;
 454         sge->length += prot->prepend_size;
 455
 456         /* Check if error is previously set on socket */
 457         if (err || sk->sk_err) {
 458                 rec = NULL;
 459
 460                 /* If err is already set on socket, return the same code */
 461                 if (sk->sk_err) {
 462                         ctx->async_wait.err = -sk->sk_err;
 463                 } else {
 464                         ctx->async_wait.err = err;
 465                         tls_err_abort(sk, err);
 466                 }
 467         }
 468
 469         if (rec) {
 470                 struct tls_rec *first_rec;
 471
 472                 /* Mark the record as ready for transmission */
 473                 smp_store_mb(rec->tx_ready, true);
 474
 475                 /* If received record is at head of tx_list, schedule tx */
 476                 first_rec = list_first_entry(&ctx->tx_list,
 477                                              struct tls_rec, list);
 478                 if (rec == first_rec)
 479                         ready = true;
 480         }
 481
 482         spin_lock_bh(&ctx->encrypt_compl_lock);
 483         pending = atomic_dec_return(&ctx->encrypt_pending);
 484
 485         if (!pending && ctx->async_notify)
 486                 complete(&ctx->async_wait.completion);
 487         spin_unlock_bh(&ctx->encrypt_compl_lock);
 488
 489         if (!ready)
 490                 return;
 491
 492         /* Schedule the transmission */
 493         if (!test_and_set_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask))
 494                 schedule_delayed_work(&ctx->tx_work.work, 1);
 495 }
 496
 497 static int tls_do_encryption(struct sock *sk,
 498                              struct tls_context *tls_ctx,
 499                              struct tls_sw_context_tx *ctx,
 500                              struct aead_request *aead_req,
 501                              size_t data_len, u32 start)
 502 {
 503         struct tls_prot_info *prot = &tls_ctx->prot_info;
 504         struct tls_rec *rec = ctx->open_rec;
 505         struct sk_msg *msg_en = &rec->msg_encrypted;
 506         struct scatterlist *sge = sk_msg_elem(msg_en, start);
 507         int rc, iv_offset = 0;
 508
 509         /* For CCM based ciphers, first byte of IV is a constant */
 510         if (prot->cipher_type == TLS_CIPHER_AES_CCM_128) {
 511                 rec->iv_data[0] = TLS_AES_CCM_IV_B0_BYTE;
 512                 iv_offset = 1;
 513         }
 514
 515         memcpy(&rec->iv_data[iv_offset], tls_ctx->tx.iv,
 516                prot->iv_size + prot->salt_size);
 517
 518         xor_iv_with_seq(prot, rec->iv_data + iv_offset, tls_ctx->tx.rec_seq);
 519
 520         sge->offset += prot->prepend_size;
 521         sge->length -= prot->prepend_size;
 522
 523         msg_en->sg.curr = start;
 524
 525         aead_request_set_tfm(aead_req, ctx->aead_send);
 526         aead_request_set_ad(aead_req, prot->aad_size);
 527         aead_request_set_crypt(aead_req, rec->sg_aead_in,
 528                                rec->sg_aead_out,
 529                                data_len, rec->iv_data);
 530
 531         aead_request_set_callback(aead_req, CRYPTO_TFM_REQ_MAY_BACKLOG,
 532                                   tls_encrypt_done, sk);
 533
 534         /* Add the record in tx_list */
 535         list_add_tail((struct list_head *)&rec->list, &ctx->tx_list);
 536         atomic_inc(&ctx->encrypt_pending);
 537
 538         rc = crypto_aead_encrypt(aead_req);
 539         if (!rc || rc != -EINPROGRESS) {
 540                 atomic_dec(&ctx->encrypt_pending);
 541                 sge->offset -= prot->prepend_size;
 542                 sge->length += prot->prepend_size;
 543         }
 544
 545         if (!rc) {
 546                 WRITE_ONCE(rec->tx_ready, true);
 547         } else if (rc != -EINPROGRESS) {
 548                 list_del(&rec->list);
 549                 return rc;
 550         }
 551
 552         /* Unhook the record from context if encryption is not failure */
 553         ctx->open_rec = NULL;
 554         tls_advance_record_sn(sk, prot, &tls_ctx->tx);
 555         return rc;
 556 }
 557
 558 static int tls_split_open_record(struct sock *sk, struct tls_rec *from,
 559                                  struct tls_rec **to, struct sk_msg *msg_opl,
 560                                  struct sk_msg *msg_oen, u32 split_point,
 561                                  u32 tx_overhead_size, u32 *orig_end)
 562 {
 563         u32 i, j, bytes = 0, apply = msg_opl->apply_bytes;
 564         struct scatterlist *sge, *osge, *nsge;
 565         u32 orig_size = msg_opl->sg.size;
 566         struct scatterlist tmp = { };
 567         struct sk_msg *msg_npl;
 568         struct tls_rec *new;
 569         int ret;
 570
 571         new = tls_get_rec(sk);
 572         if (!new)
 573                 return -ENOMEM;
 574         ret = sk_msg_alloc(sk, &new->msg_encrypted, msg_opl->sg.size +
 575                            tx_overhead_size, 0);
 576         if (ret < 0) {
 577                 tls_free_rec(sk, new);
 578                 return ret;
 579         }
 580
 581         *orig_end = msg_opl->sg.end;
 582         i = msg_opl->sg.start;
 583         sge = sk_msg_elem(msg_opl, i);
 584         while (apply && sge->length) {
 585                 if (sge->length > apply) {
 586                         u32 len = sge->length - apply;
 587
 588                         get_page(sg_page(sge));
 589                         sg_set_page(&tmp, sg_page(sge), len,
 590                                     sge->offset + apply);
 591                         sge->length = apply;
 592                         bytes += apply;
 593                         apply = 0;
 594                 } else {
 595                         apply -= sge->length;
 596                         bytes += sge->length;
 597                 }
 598
 599                 sk_msg_iter_var_next(i);
 600                 if (i == msg_opl->sg.end)
 601                         break;
 602                 sge = sk_msg_elem(msg_opl, i);
 603         }
 604
 605         msg_opl->sg.end = i;
 606         msg_opl->sg.curr = i;
 607         msg_opl->sg.copybreak = 0;
 608         msg_opl->apply_bytes = 0;
 609         msg_opl->sg.size = bytes;
 610
 611         msg_npl = &new->msg_plaintext;
 612         msg_npl->apply_bytes = apply;
 613         msg_npl->sg.size = orig_size - bytes;
 614
 615         j = msg_npl->sg.start;
 616         nsge = sk_msg_elem(msg_npl, j);
 617         if (tmp.length) {
 618                 memcpy(nsge, &tmp, sizeof(*nsge));
 619                 sk_msg_iter_var_next(j);
 620                 nsge = sk_msg_elem(msg_npl, j);
 621         }
 622
 623         osge = sk_msg_elem(msg_opl, i);
 624         while (osge->length) {
 625                 memcpy(nsge, osge, sizeof(*nsge));
 626                 sg_unmark_end(nsge);
 627                 sk_msg_iter_var_next(i);
 628                 sk_msg_iter_var_next(j);
 629                 if (i == *orig_end)
 630                         break;
 631                 osge = sk_msg_elem(msg_opl, i);
 632                 nsge = sk_msg_elem(msg_npl, j);
 633         }
 634
 635         msg_npl->sg.end = j;
 636         msg_npl->sg.curr = j;
 637         msg_npl->sg.copybreak = 0;
 638
 639         *to = new;
 640         return 0;
 641 }
 642
 643 static void tls_merge_open_record(struct sock *sk, struct tls_rec *to,
 644                                   struct tls_rec *from, u32 orig_end)
 645 {
 646         struct sk_msg *msg_npl = &from->msg_plaintext;
 647         struct sk_msg *msg_opl = &to->msg_plaintext;
 648         struct scatterlist *osge, *nsge;
 649         u32 i, j;
 650
 651         i = msg_opl->sg.end;
 652         sk_msg_iter_var_prev(i);
 653         j = msg_npl->sg.start;
 654
 655         osge = sk_msg_elem(msg_opl, i);
 656         nsge = sk_msg_elem(msg_npl, j);
 657
 658         if (sg_page(osge) == sg_page(nsge) &&
 659             osge->offset + osge->length == nsge->offset) {
 660                 osge->length += nsge->length;
 661                 put_page(sg_page(nsge));
 662         }
 663
 664         msg_opl->sg.end = orig_end;
 665         msg_opl->sg.curr = orig_end;
 666         msg_opl->sg.copybreak = 0;
 667         msg_opl->apply_bytes = msg_opl->sg.size + msg_npl->sg.size;
 668         msg_opl->sg.size += msg_npl->sg.size;
 669
 670         sk_msg_free(sk, &to->msg_encrypted);
 671         sk_msg_xfer_full(&to->msg_encrypted, &from->msg_encrypted);
 672
 673         kfree(from);
 674 }
 675
 676 static int tls_push_record(struct sock *sk, int flags,
 677                            unsigned char record_type)
 678 {
 679         struct tls_context *tls_ctx = tls_get_ctx(sk);
 680         struct tls_prot_info *prot = &tls_ctx->prot_info;
 681         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 682         struct tls_rec *rec = ctx->open_rec, *tmp = NULL;
 683         u32 i, split_point, orig_end;
 684         struct sk_msg *msg_pl, *msg_en;
 685         struct aead_request *req;
 686         bool split;
 687         int rc;
 688
 689         if (!rec)
 690                 return 0;
 691
 692         msg_pl = &rec->msg_plaintext;
 693         msg_en = &rec->msg_encrypted;
 694
 695         split_point = msg_pl->apply_bytes;
 696         split = split_point && split_point < msg_pl->sg.size;
 697         if (unlikely((!split &&
 698                       msg_pl->sg.size +
 699                       prot->overhead_size > msg_en->sg.size) ||
 700                      (split &&
 701                       split_point +
 702                       prot->overhead_size > msg_en->sg.size))) {
 703                 split = true;
 704                 split_point = msg_en->sg.size;
 705         }
 706         if (split) {
 707                 rc = tls_split_open_record(sk, rec, &tmp, msg_pl, msg_en,
 708                                            split_point, prot->overhead_size,
 709                                            &orig_end);
 710                 if (rc < 0)
 711                         return rc;
 712                 /* This can happen if above tls_split_open_record allocates
 713                  * a single large encryption buffer instead of two smaller
 714                  * ones. In this case adjust pointers and continue without
 715                  * split.
 716                  */
 717                 if (!msg_pl->sg.size) {
 718                         tls_merge_open_record(sk, rec, tmp, orig_end);
 719                         msg_pl = &rec->msg_plaintext;
 720                         msg_en = &rec->msg_encrypted;
 721                         split = false;
 722                 }
 723                 sk_msg_trim(sk, msg_en, msg_pl->sg.size +
 724                             prot->overhead_size);
 725         }
 726
 727         rec->tx_flags = flags;
 728         req = &rec->aead_req;
 729
 730         i = msg_pl->sg.end;
 731         sk_msg_iter_var_prev(i);
 732
 733         rec->content_type = record_type;
 734         if (prot->version == TLS_1_3_VERSION) {
 735                 /* Add content type to end of message.  No padding added */
 736                 sg_set_buf(&rec->sg_content_type, &rec->content_type, 1);
 737                 sg_mark_end(&rec->sg_content_type);
 738                 sg_chain(msg_pl->sg.data, msg_pl->sg.end + 1,
 739                          &rec->sg_content_type);
 740         } else {
 741                 sg_mark_end(sk_msg_elem(msg_pl, i));
 742         }
 743
 744         if (msg_pl->sg.end < msg_pl->sg.start) {
 745                 sg_chain(&msg_pl->sg.data[msg_pl->sg.start],
 746                          MAX_SKB_FRAGS - msg_pl->sg.start + 1,
 747                          msg_pl->sg.data);
 748         }
 749
 750         i = msg_pl->sg.start;
 751         sg_chain(rec->sg_aead_in, 2, &msg_pl->sg.data[i]);
 752
 753         i = msg_en->sg.end;
 754         sk_msg_iter_var_prev(i);
 755         sg_mark_end(sk_msg_elem(msg_en, i));
 756
 757         i = msg_en->sg.start;
 758         sg_chain(rec->sg_aead_out, 2, &msg_en->sg.data[i]);
 759
 760         tls_make_aad(rec->aad_space, msg_pl->sg.size + prot->tail_size,
 761                      tls_ctx->tx.rec_seq, record_type, prot);
 762
 763         tls_fill_prepend(tls_ctx,
 764                          page_address(sg_page(&msg_en->sg.data[i])) +
 765                          msg_en->sg.data[i].offset,
 766                          msg_pl->sg.size + prot->tail_size,
 767                          record_type);
 768
 769         tls_ctx->pending_open_record_frags = false;
 770
 771         rc = tls_do_encryption(sk, tls_ctx, ctx, req,
 772                                msg_pl->sg.size + prot->tail_size, i);
 773         if (rc < 0) {
 774                 if (rc != -EINPROGRESS) {
 775                         tls_err_abort(sk, -EBADMSG);
 776                         if (split) {
 777                                 tls_ctx->pending_open_record_frags = true;
 778                                 tls_merge_open_record(sk, rec, tmp, orig_end);
 779                         }
 780                 }
 781                 ctx->async_capable = 1;
 782                 return rc;
 783         } else if (split) {
 784                 msg_pl = &tmp->msg_plaintext;
 785                 msg_en = &tmp->msg_encrypted;
 786                 sk_msg_trim(sk, msg_en, msg_pl->sg.size + prot->overhead_size);
 787                 tls_ctx->pending_open_record_frags = true;
 788                 ctx->open_rec = tmp;
 789         }
 790
 791         return tls_tx_records(sk, flags);
 792 }
 793
 794 static int bpf_exec_tx_verdict(struct sk_msg *msg, struct sock *sk,
 795                                bool full_record, u8 record_type,
 796                                ssize_t *copied, int flags)
 797 {
 798         struct tls_context *tls_ctx = tls_get_ctx(sk);
 799         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 800         struct sk_msg msg_redir = { };
 801         struct sk_psock *psock;
 802         struct sock *sk_redir;
 803         struct tls_rec *rec;
 804         bool enospc, policy;
 805         int err = 0, send;
 806         u32 delta = 0;
 807
 808         policy = !(flags & MSG_SENDPAGE_NOPOLICY);
 809         psock = sk_psock_get(sk);
 810         if (!psock || !policy) {
 811                 err = tls_push_record(sk, flags, record_type);
 812                 if (err && sk->sk_err == EBADMSG) {
 813                         *copied -= sk_msg_free(sk, msg);
 814                         tls_free_open_rec(sk);
 815                         err = -sk->sk_err;
 816                 }
 817                 if (psock)
 818                         sk_psock_put(sk, psock);
 819                 return err;
 820         }
 821 more_data:
 822         enospc = sk_msg_full(msg);
 823         if (psock->eval == __SK_NONE) {
 824                 delta = msg->sg.size;
 825                 psock->eval = sk_psock_msg_verdict(sk, psock, msg);
 826                 delta -= msg->sg.size;
 827         }
 828         if (msg->cork_bytes && msg->cork_bytes > msg->sg.size &&
 829             !enospc && !full_record) {
 830                 err = -ENOSPC;
 831                 goto out_err;
 832         }
 833         msg->cork_bytes = 0;
 834         send = msg->sg.size;
 835         if (msg->apply_bytes && msg->apply_bytes < send)
 836                 send = msg->apply_bytes;
 837
 838         switch (psock->eval) {
 839         case __SK_PASS:
 840                 err = tls_push_record(sk, flags, record_type);
 841                 if (err && sk->sk_err == EBADMSG) {
 842                         *copied -= sk_msg_free(sk, msg);
 843                         tls_free_open_rec(sk);
 844                         err = -sk->sk_err;
 845                         goto out_err;
 846                 }
 847                 break;
 848         case __SK_REDIRECT:
 849                 sk_redir = psock->sk_redir;
 850                 memcpy(&msg_redir, msg, sizeof(*msg));
 851                 if (msg->apply_bytes < send)
 852                         msg->apply_bytes = 0;
 853                 else
 854                         msg->apply_bytes -= send;
 855                 sk_msg_return_zero(sk, msg, send);
 856                 msg->sg.size -= send;
 857                 release_sock(sk);
 858                 err = tcp_bpf_sendmsg_redir(sk_redir, &msg_redir, send, flags);
 859                 lock_sock(sk);
 860                 if (err < 0) {
 861                         *copied -= sk_msg_free_nocharge(sk, &msg_redir);
 862                         msg->sg.size = 0;
 863                 }
 864                 if (msg->sg.size == 0)
 865                         tls_free_open_rec(sk);
 866                 break;
 867         case __SK_DROP:
 868         default:
 869                 sk_msg_free_partial(sk, msg, send);
 870                 if (msg->apply_bytes < send)
 871                         msg->apply_bytes = 0;
 872                 else
 873                         msg->apply_bytes -= send;
 874                 if (msg->sg.size == 0)
 875                         tls_free_open_rec(sk);
 876                 *copied -= (send + delta);
 877                 err = -EACCES;
 878         }
 879
 880         if (likely(!err)) {
 881                 bool reset_eval = !ctx->open_rec;
 882
 883                 rec = ctx->open_rec;
 884                 if (rec) {
 885                         msg = &rec->msg_plaintext;
 886                         if (!msg->apply_bytes)
 887                                 reset_eval = true;
 888                 }
 889                 if (reset_eval) {
 890                         psock->eval = __SK_NONE;
 891                         if (psock->sk_redir) {
 892                                 sock_put(psock->sk_redir);
 893                                 psock->sk_redir = NULL;
 894                         }
 895                 }
 896                 if (rec)
 897                         goto more_data;
 898         }
 899  out_err:
 900         sk_psock_put(sk, psock);
 901         return err;
 902 }
 903
 904 static int tls_sw_push_pending_record(struct sock *sk, int flags)
 905 {
 906         struct tls_context *tls_ctx = tls_get_ctx(sk);
 907         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 908         struct tls_rec *rec = ctx->open_rec;
 909         struct sk_msg *msg_pl;
 910         size_t copied;
 911
 912         if (!rec)
 913                 return 0;
 914
 915         msg_pl = &rec->msg_plaintext;
 916         copied = msg_pl->sg.size;
 917         if (!copied)
 918                 return 0;
 919
 920         return bpf_exec_tx_verdict(msg_pl, sk, true, TLS_RECORD_TYPE_DATA,
 921                                    &copied, flags);
 922 }
 923
 924 int tls_sw_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
 925 {
 926         long timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
 927         struct tls_context *tls_ctx = tls_get_ctx(sk);
 928         struct tls_prot_info *prot = &tls_ctx->prot_info;
 929         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 930         bool async_capable = ctx->async_capable;
 931         unsigned char record_type = TLS_RECORD_TYPE_DATA;
 932         bool is_kvec = iov_iter_is_kvec(&msg->msg_iter);
 933         bool eor = !(msg->msg_flags & MSG_MORE);
 934         size_t try_to_copy;
 935         ssize_t copied = 0;
 936         struct sk_msg *msg_pl, *msg_en;
 937         struct tls_rec *rec;
 938         int required_size;
 939         int num_async = 0;
 940         bool full_record;
 941         int record_room;
 942         int num_zc = 0;
 943         int orig_size;
 944         int ret = 0;
 945         int pending;
 946
 947         if (msg->msg_flags & ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL |
 948                                MSG_CMSG_COMPAT))
 949                 return -EOPNOTSUPP;
 950
 951         mutex_lock(&tls_ctx->tx_lock);
 952         lock_sock(sk);
 953
 954         if (unlikely(msg->msg_controllen)) {
 955                 ret = tls_proccess_cmsg(sk, msg, &record_type);
 956                 if (ret) {
 957                         if (ret == -EINPROGRESS)
 958                                 num_async++;
 959                         else if (ret != -EAGAIN)
 960                                 goto send_end;
 961                 }
 962         }
 963
 964         while (msg_data_left(msg)) {
 965                 if (sk->sk_err) {
 966                         ret = -sk->sk_err;
 967                         goto send_end;
 968                 }
 969
 970                 if (ctx->open_rec)
 971                         rec = ctx->open_rec;
 972                 else
 973                         rec = ctx->open_rec = tls_get_rec(sk);
 974                 if (!rec) {
 975                         ret = -ENOMEM;
 976                         goto send_end;
 977                 }
 978
 979                 msg_pl = &rec->msg_plaintext;
 980                 msg_en = &rec->msg_encrypted;
 981
 982                 orig_size = msg_pl->sg.size;
 983                 full_record = false;
 984                 try_to_copy = msg_data_left(msg);
 985                 record_room = TLS_MAX_PAYLOAD_SIZE - msg_pl->sg.size;
 986                 if (try_to_copy >= record_room) {
 987                         try_to_copy = record_room;
 988                         full_record = true;
 989                 }
 990
 991                 required_size = msg_pl->sg.size + try_to_copy +
 992                                 prot->overhead_size;
 993
 994                 if (!sk_stream_memory_free(sk))
 995                         goto wait_for_sndbuf;
 996
 997 alloc_encrypted:
 998                 ret = tls_alloc_encrypted_msg(sk, required_size);
 999                 if (ret) {
1000                         if (ret != -ENOSPC)
1001                                 goto wait_for_memory;
1002
1003                         /* Adjust try_to_copy according to the amount that was
1004                          * actually allocated. The difference is due
1005                          * to max sg elements limit
1006                          */
1007                         try_to_copy -= required_size - msg_en->sg.size;
1008                         full_record = true;
1009                 }
1010
1011                 if (!is_kvec && (full_record || eor) && !async_capable) {
1012                         u32 first = msg_pl->sg.end;
1013
1014                         ret = sk_msg_zerocopy_from_iter(sk, &msg->msg_iter,
1015                                                         msg_pl, try_to_copy);
1016                         if (ret)
1017                                 goto fallback_to_reg_send;
1018
1019                         num_zc++;
1020                         copied += try_to_copy;
1021
1022                         sk_msg_sg_copy_set(msg_pl, first);
1023                         ret = bpf_exec_tx_verdict(msg_pl, sk, full_record,
1024                                                   record_type, &copied,
1025                                                   msg->msg_flags);
1026                         if (ret) {
1027                                 if (ret == -EINPROGRESS)
1028                                         num_async++;
1029                                 else if (ret == -ENOMEM)
1030                                         goto wait_for_memory;
1031                                 else if (ctx->open_rec && ret == -ENOSPC)
1032                                         goto rollback_iter;
1033                                 else if (ret != -EAGAIN)
1034                                         goto send_end;
1035                         }
1036                         continue;
1037 rollback_iter:
1038                         copied -= try_to_copy;
1039                         sk_msg_sg_copy_clear(msg_pl, first);
1040                         iov_iter_revert(&msg->msg_iter,
1041                                         msg_pl->sg.size - orig_size);
1042 fallback_to_reg_send:
1043                         sk_msg_trim(sk, msg_pl, orig_size);
1044                 }
1045
1046                 required_size = msg_pl->sg.size + try_to_copy;
1047
1048                 ret = tls_clone_plaintext_msg(sk, required_size);
1049                 if (ret) {
1050                         if (ret != -ENOSPC)
1051                                 goto send_end;
1052
1053                         /* Adjust try_to_copy according to the amount that was
1054                          * actually allocated. The difference is due
1055                          * to max sg elements limit
1056                          */
1057                         try_to_copy -= required_size - msg_pl->sg.size;
1058                         full_record = true;
1059                         sk_msg_trim(sk, msg_en,
1060                                     msg_pl->sg.size + prot->overhead_size);
1061                 }
1062
1063                 if (try_to_copy) {
1064                         ret = sk_msg_memcopy_from_iter(sk, &msg->msg_iter,
1065                                                        msg_pl, try_to_copy);
1066                         if (ret < 0)
1067                                 goto trim_sgl;
1068                 }
1069
1070                 /* Open records defined only if successfully copied, otherwise
1071                  * we would trim the sg but not reset the open record frags.
1072                  */
1073                 tls_ctx->pending_open_record_frags = true;
1074                 copied += try_to_copy;
1075                 if (full_record || eor) {
1076                         ret = bpf_exec_tx_verdict(msg_pl, sk, full_record,
1077                                                   record_type, &copied,
1078                                                   msg->msg_flags);
1079                         if (ret) {
1080                                 if (ret == -EINPROGRESS)
1081                                         num_async++;
1082                                 else if (ret == -ENOMEM)
1083                                         goto wait_for_memory;
1084                                 else if (ret != -EAGAIN) {
1085                                         if (ret == -ENOSPC)
1086                                                 ret = 0;
1087                                         goto send_end;
1088                                 }
1089                         }
1090                 }
1091
1092                 continue;
1093
1094 wait_for_sndbuf:
1095                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1096 wait_for_memory:
1097                 ret = sk_stream_wait_memory(sk, &timeo);
1098                 if (ret) {
1099 trim_sgl:
1100                         if (ctx->open_rec)
1101                                 tls_trim_both_msgs(sk, orig_size);
1102                         goto send_end;
1103                 }
1104
1105                 if (ctx->open_rec && msg_en->sg.size < required_size)
1106                         goto alloc_encrypted;
1107         }
1108
1109         if (!num_async) {
1110                 goto send_end;
1111         } else if (num_zc) {
1112                 /* Wait for pending encryptions to get completed */
1113                 spin_lock_bh(&ctx->encrypt_compl_lock);
1114                 ctx->async_notify = true;
1115
1116                 pending = atomic_read(&ctx->encrypt_pending);
1117                 spin_unlock_bh(&ctx->encrypt_compl_lock);
1118                 if (pending)
1119                         crypto_wait_req(-EINPROGRESS, &ctx->async_wait);
1120                 else
1121                         reinit_completion(&ctx->async_wait.completion);
1122
1123                 /* There can be no concurrent accesses, since we have no
1124                  * pending encrypt operations
1125                  */
1126                 WRITE_ONCE(ctx->async_notify, false);
1127
1128                 if (ctx->async_wait.err) {
1129                         ret = ctx->async_wait.err;
1130                         copied = 0;
1131                 }
1132         }
1133
1134         /* Transmit if any encryptions have completed */
1135         if (test_and_clear_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask)) {
1136                 cancel_delayed_work(&ctx->tx_work.work);
1137                 tls_tx_records(sk, msg->msg_flags);
1138         }
1139
1140 send_end:
1141         ret = sk_stream_error(sk, msg->msg_flags, ret);
1142
1143         release_sock(sk);
1144         mutex_unlock(&tls_ctx->tx_lock);
1145         return copied > 0 ? copied : ret;
1146 }
1147
1148 static int tls_sw_do_sendpage(struct sock *sk, struct page *page,
1149                               int offset, size_t size, int flags)
1150 {
1151         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1152         struct tls_context *tls_ctx = tls_get_ctx(sk);
1153         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
1154         struct tls_prot_info *prot = &tls_ctx->prot_info;
1155         unsigned char record_type = TLS_RECORD_TYPE_DATA;
1156         struct sk_msg *msg_pl;
1157         struct tls_rec *rec;
1158         int num_async = 0;
1159         ssize_t copied = 0;
1160         bool full_record;
1161         int record_room;
1162         int ret = 0;
1163         bool eor;
1164
1165         eor = !(flags & MSG_SENDPAGE_NOTLAST);
1166         sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1167
1168         /* Call the sk_stream functions to manage the sndbuf mem. */
1169         while (size > 0) {
1170                 size_t copy, required_size;
1171
1172                 if (sk->sk_err) {
1173                         ret = -sk->sk_err;
1174                         goto sendpage_end;
1175                 }
1176
1177                 if (ctx->open_rec)
1178                         rec = ctx->open_rec;
1179                 else
1180                         rec = ctx->open_rec = tls_get_rec(sk);
1181                 if (!rec) {
1182                         ret = -ENOMEM;
1183                         goto sendpage_end;
1184                 }
1185
1186                 msg_pl = &rec->msg_plaintext;
1187
1188                 full_record = false;
1189                 record_room = TLS_MAX_PAYLOAD_SIZE - msg_pl->sg.size;
1190                 copy = size;
1191                 if (copy >= record_room) {
1192                         copy = record_room;
1193                         full_record = true;
1194                 }
1195
1196                 required_size = msg_pl->sg.size + copy + prot->overhead_size;
1197
1198                 if (!sk_stream_memory_free(sk))
1199                         goto wait_for_sndbuf;
1200 alloc_payload:
1201                 ret = tls_alloc_encrypted_msg(sk, required_size);
1202                 if (ret) {
1203                         if (ret != -ENOSPC)
1204                                 goto wait_for_memory;
1205
1206                         /* Adjust copy according to the amount that was
1207                          * actually allocated. The difference is due
1208                          * to max sg elements limit
1209                          */
1210                         copy -= required_size - msg_pl->sg.size;
1211                         full_record = true;
1212                 }
1213
1214                 sk_msg_page_add(msg_pl, page, copy, offset);
1215                 sk_mem_charge(sk, copy);
1216
1217                 offset += copy;
1218                 size -= copy;
1219                 copied += copy;
1220
1221                 tls_ctx->pending_open_record_frags = true;
1222                 if (full_record || eor || sk_msg_full(msg_pl)) {
1223                         ret = bpf_exec_tx_verdict(msg_pl, sk, full_record,
1224                                                   record_type, &copied, flags);
1225                         if (ret) {
1226                                 if (ret == -EINPROGRESS)
1227                                         num_async++;
1228                                 else if (ret == -ENOMEM)
1229                                         goto wait_for_memory;
1230                                 else if (ret != -EAGAIN) {
1231                                         if (ret == -ENOSPC)
1232                                                 ret = 0;
1233                                         goto sendpage_end;
1234                                 }
1235                         }
1236                 }
1237                 continue;
1238 wait_for_sndbuf:
1239                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1240 wait_for_memory:
1241                 ret = sk_stream_wait_memory(sk, &timeo);
1242                 if (ret) {
1243                         if (ctx->open_rec)
1244                                 tls_trim_both_msgs(sk, msg_pl->sg.size);
1245                         goto sendpage_end;
1246                 }
1247
1248                 if (ctx->open_rec)
1249                         goto alloc_payload;
1250         }
1251
1252         if (num_async) {
1253                 /* Transmit if any encryptions have completed */
1254                 if (test_and_clear_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask)) {
1255                         cancel_delayed_work(&ctx->tx_work.work);
1256                         tls_tx_records(sk, flags);
1257                 }
1258         }
1259 sendpage_end:
1260         ret = sk_stream_error(sk, flags, ret);
1261         return copied > 0 ? copied : ret;
1262 }
1263
1264 int tls_sw_sendpage_locked(struct sock *sk, struct page *page,
1265                            int offset, size_t size, int flags)
1266 {
1267         if (flags & ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL |
1268                       MSG_SENDPAGE_NOTLAST | MSG_SENDPAGE_NOPOLICY |
1269                       MSG_NO_SHARED_FRAGS))
1270                 return -EOPNOTSUPP;
1271
1272         return tls_sw_do_sendpage(sk, page, offset, size, flags);
1273 }
1274
1275 int tls_sw_sendpage(struct sock *sk, struct page *page,
1276                     int offset, size_t size, int flags)
1277 {
1278         struct tls_context *tls_ctx = tls_get_ctx(sk);
1279         int ret;
1280
1281         if (flags & ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL |
1282                       MSG_SENDPAGE_NOTLAST | MSG_SENDPAGE_NOPOLICY))
1283                 return -EOPNOTSUPP;
1284
1285         mutex_lock(&tls_ctx->tx_lock);
1286         lock_sock(sk);
1287         ret = tls_sw_do_sendpage(sk, page, offset, size, flags);
1288         release_sock(sk);
1289         mutex_unlock(&tls_ctx->tx_lock);
1290         return ret;
1291 }
1292
1293 static struct sk_buff *tls_wait_data(struct sock *sk, struct sk_psock *psock,
1294                                      bool nonblock, long timeo, int *err)
1295 {
1296         struct tls_context *tls_ctx = tls_get_ctx(sk);
1297         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1298         struct sk_buff *skb;
1299         DEFINE_WAIT_FUNC(wait, woken_wake_function);
1300
1301         while (!(skb = ctx->recv_pkt) && sk_psock_queue_empty(psock)) {
1302                 if (sk->sk_err) {
1303                         *err = sock_error(sk);
1304                         return NULL;
1305                 }
1306
1307                 if (!skb_queue_empty(&sk->sk_receive_queue)) {
1308                         __strp_unpause(&ctx->strp);
1309                         if (ctx->recv_pkt)
1310                                 return ctx->recv_pkt;
1311                 }
1312
1313                 if (sk->sk_shutdown & RCV_SHUTDOWN)
1314                         return NULL;
1315
1316                 if (sock_flag(sk, SOCK_DONE))
1317                         return NULL;
1318
1319                 if (nonblock || !timeo) {
1320                         *err = -EAGAIN;
1321                         return NULL;
1322                 }
1323
1324                 add_wait_queue(sk_sleep(sk), &wait);
1325                 sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1326                 sk_wait_event(sk, &timeo,
1327                               ctx->recv_pkt != skb ||
1328                               !sk_psock_queue_empty(psock),
1329                               &wait);
1330                 sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1331                 remove_wait_queue(sk_sleep(sk), &wait);
1332
1333                 /* Handle signals */
1334                 if (signal_pending(current)) {
1335                         *err = sock_intr_errno(timeo);
1336                         return NULL;
1337                 }
1338         }
1339
1340         return skb;
1341 }
1342
1343 static int tls_setup_from_iter(struct sock *sk, struct iov_iter *from,
1344                                int length, int *pages_used,
1345                                unsigned int *size_used,
1346                                struct scatterlist *to,
1347                                int to_max_pages)
1348 {
1349         int rc = 0, i = 0, num_elem = *pages_used, maxpages;
1350         struct page *pages[MAX_SKB_FRAGS];
1351         unsigned int size = *size_used;
1352         ssize_t copied, use;
1353         size_t offset;
1354
1355         while (length > 0) {
1356                 i = 0;
1357                 maxpages = to_max_pages - num_elem;
1358                 if (maxpages == 0) {
1359                         rc = -EFAULT;
1360                         goto out;
1361                 }
1362                 copied = iov_iter_get_pages(from, pages,
1363                                             length,
1364                                             maxpages, &offset);
1365                 if (copied <= 0) {
1366                         rc = -EFAULT;
1367                         goto out;
1368                 }
1369
1370                 iov_iter_advance(from, copied);
1371
1372                 length -= copied;
1373                 size += copied;
1374                 while (copied) {
1375                         use = min_t(int, copied, PAGE_SIZE - offset);
1376
1377                         sg_set_page(&to[num_elem],
1378                                     pages[i], use, offset);
1379                         sg_unmark_end(&to[num_elem]);
1380                         /* We do not uncharge memory from this API */
1381
1382                         offset = 0;
1383                         copied -= use;
1384
1385                         i++;
1386                         num_elem++;
1387                 }
1388         }
1389         /* Mark the end in the last sg entry if newly added */
1390         if (num_elem > *pages_used)
1391                 sg_mark_end(&to[num_elem - 1]);
1392 out:
1393         if (rc)
1394                 iov_iter_revert(from, size - *size_used);
1395         *size_used = size;
1396         *pages_used = num_elem;
1397
1398         return rc;
1399 }
1400
1401 /* This function decrypts the input skb into either out_iov or in out_sg
1402  * or in skb buffers itself. The input parameter 'zc' indicates if
1403  * zero-copy mode needs to be tried or not. With zero-copy mode, either
1404  * out_iov or out_sg must be non-NULL. In case both out_iov and out_sg are
1405  * NULL, then the decryption happens inside skb buffers itself, i.e.
1406  * zero-copy gets disabled and 'zc' is updated.
1407  */
1408
1409 static int decrypt_internal(struct sock *sk, struct sk_buff *skb,
1410                             struct iov_iter *out_iov,
1411                             struct scatterlist *out_sg,
1412                             int *chunk, bool *zc, bool async)
1413 {
1414         struct tls_context *tls_ctx = tls_get_ctx(sk);
1415         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1416         struct tls_prot_info *prot = &tls_ctx->prot_info;
1417         struct strp_msg *rxm = strp_msg(skb);
1418         int n_sgin, n_sgout, nsg, mem_size, aead_size, err, pages = 0;
1419         struct aead_request *aead_req;
1420         struct sk_buff *unused;
1421         u8 *aad, *iv, *mem = NULL;
1422         struct scatterlist *sgin = NULL;
1423         struct scatterlist *sgout = NULL;
1424         const int data_len = rxm->full_len - prot->overhead_size +
1425                              prot->tail_size;
1426         int iv_offset = 0;
1427
1428         if (*zc && (out_iov || out_sg)) {
1429                 if (out_iov)
1430                         n_sgout = iov_iter_npages(out_iov, INT_MAX) + 1;
1431                 else
1432                         n_sgout = sg_nents(out_sg);
1433                 n_sgin = skb_nsg(skb, rxm->offset + prot->prepend_size,
1434                                  rxm->full_len - prot->prepend_size);
1435         } else {
1436                 n_sgout = 0;
1437                 *zc = false;
1438                 n_sgin = skb_cow_data(skb, 0, &unused);
1439         }
1440
1441         if (n_sgin < 1)
1442                 return -EBADMSG;
1443
1444         /* Increment to accommodate AAD */
1445         n_sgin = n_sgin + 1;
1446
1447         nsg = n_sgin + n_sgout;
1448
1449         aead_size = sizeof(*aead_req) + crypto_aead_reqsize(ctx->aead_recv);
1450         mem_size = aead_size + (nsg * sizeof(struct scatterlist));
1451         mem_size = mem_size + prot->aad_size;
1452         mem_size = mem_size + crypto_aead_ivsize(ctx->aead_recv);
1453
1454         /* Allocate a single block of memory which contains
1455          * aead_req || sgin[] || sgout[] || aad || iv.
1456          * This order achieves correct alignment for aead_req, sgin, sgout.
1457          */
1458         mem = kmalloc(mem_size, sk->sk_allocation);
1459         if (!mem)
1460                 return -ENOMEM;
1461
1462         /* Segment the allocated memory */
1463         aead_req = (struct aead_request *)mem;
1464         sgin = (struct scatterlist *)(mem + aead_size);
1465         sgout = sgin + n_sgin;
1466         aad = (u8 *)(sgout + n_sgout);
1467         iv = aad + prot->aad_size;
1468
1469         /* For CCM based ciphers, first byte of nonce+iv is always '2' */
1470         if (prot->cipher_type == TLS_CIPHER_AES_CCM_128) {
1471                 iv[0] = 2;
1472                 iv_offset = 1;
1473         }
1474
1475         /* Prepare IV */
1476         err = skb_copy_bits(skb, rxm->offset + TLS_HEADER_SIZE,
1477                             iv + iv_offset + prot->salt_size,
1478                             prot->iv_size);
1479         if (err < 0) {
1480                 kfree(mem);
1481                 return err;
1482         }
1483         if (prot->version == TLS_1_3_VERSION ||
1484             prot->cipher_type == TLS_CIPHER_CHACHA20_POLY1305)
1485                 memcpy(iv + iv_offset, tls_ctx->rx.iv,
1486                        prot->iv_size + prot->salt_size);
1487         else
1488                 memcpy(iv + iv_offset, tls_ctx->rx.iv, prot->salt_size);
1489
1490         xor_iv_with_seq(prot, iv + iv_offset, tls_ctx->rx.rec_seq);
1491
1492         /* Prepare AAD */
1493         tls_make_aad(aad, rxm->full_len - prot->overhead_size +
1494                      prot->tail_size,
1495                      tls_ctx->rx.rec_seq, ctx->control, prot);
1496
1497         /* Prepare sgin */
1498         sg_init_table(sgin, n_sgin);
1499         sg_set_buf(&sgin[0], aad, prot->aad_size);
1500         err = skb_to_sgvec(skb, &sgin[1],
1501                            rxm->offset + prot->prepend_size,
1502                            rxm->full_len - prot->prepend_size);
1503         if (err < 0) {
1504                 kfree(mem);
1505                 return err;
1506         }
1507
1508         if (n_sgout) {
1509                 if (out_iov) {
1510                         sg_init_table(sgout, n_sgout);
1511                         sg_set_buf(&sgout[0], aad, prot->aad_size);
1512
1513                         *chunk = 0;
1514                         err = tls_setup_from_iter(sk, out_iov, data_len,
1515                                                   &pages, chunk, &sgout[1],
1516                                                   (n_sgout - 1));
1517                         if (err < 0)
1518                                 goto fallback_to_reg_recv;
1519                 } else if (out_sg) {
1520                         memcpy(sgout, out_sg, n_sgout * sizeof(*sgout));
1521                 } else {
1522                         goto fallback_to_reg_recv;
1523                 }
1524         } else {
1525 fallback_to_reg_recv:
1526                 sgout = sgin;
1527                 pages = 0;
1528                 *chunk = data_len;
1529                 *zc = false;
1530         }
1531
1532         /* Prepare and submit AEAD request */
1533         err = tls_do_decryption(sk, skb, sgin, sgout, iv,
1534                                 data_len, aead_req, async);
1535         if (err == -EINPROGRESS)
1536                 return err;
1537
1538         /* Release the pages in case iov was mapped to pages */
1539         for (; pages > 0; pages--)
1540                 put_page(sg_page(&sgout[pages]));
1541
1542         kfree(mem);
1543         return err;
1544 }
1545
1546 static int decrypt_skb_update(struct sock *sk, struct sk_buff *skb,
1547                               struct iov_iter *dest, int *chunk, bool *zc,
1548                               bool async)
1549 {
1550         struct tls_context *tls_ctx = tls_get_ctx(sk);
1551         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1552         struct tls_prot_info *prot = &tls_ctx->prot_info;
1553         struct strp_msg *rxm = strp_msg(skb);
1554         int pad, err = 0;
1555
1556         if (!ctx->decrypted) {
1557                 if (tls_ctx->rx_conf == TLS_HW) {
1558                         err = tls_device_decrypted(sk, tls_ctx, skb, rxm);
1559                         if (err < 0)
1560                                 return err;
1561                 }
1562
1563                 /* Still not decrypted after tls_device */
1564                 if (!ctx->decrypted) {
1565                         err = decrypt_internal(sk, skb, dest, NULL, chunk, zc,
1566                                                async);
1567                         if (err < 0) {
1568                                 if (err == -EINPROGRESS)
1569                                         tls_advance_record_sn(sk, prot,
1570                                                               &tls_ctx->rx);
1571                                 else if (err == -EBADMSG)
1572                                         TLS_INC_STATS(sock_net(sk),
1573                                                       LINUX_MIB_TLSDECRYPTERROR);
1574                                 return err;
1575                         }
1576                 } else {
1577                         *zc = false;
1578                 }
1579
1580                 pad = padding_length(ctx, prot, skb);
1581                 if (pad < 0)
1582                         return pad;
1583
1584                 rxm->full_len -= pad;
1585                 rxm->offset += prot->prepend_size;
1586                 rxm->full_len -= prot->overhead_size;
1587                 tls_advance_record_sn(sk, prot, &tls_ctx->rx);
1588                 ctx->decrypted = 1;
1589                 ctx->saved_data_ready(sk);
1590         } else {
1591                 *zc = false;
1592         }
1593
1594         return err;
1595 }
1596
1597 int decrypt_skb(struct sock *sk, struct sk_buff *skb,
1598                 struct scatterlist *sgout)
1599 {
1600         bool zc = true;
1601         int chunk;
1602
1603         return decrypt_internal(sk, skb, NULL, sgout, &chunk, &zc, false);
1604 }
1605
1606 static bool tls_sw_advance_skb(struct sock *sk, struct sk_buff *skb,
1607                                unsigned int len)
1608 {
1609         struct tls_context *tls_ctx = tls_get_ctx(sk);
1610         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1611
1612         if (skb) {
1613                 struct strp_msg *rxm = strp_msg(skb);
1614
1615                 if (len < rxm->full_len) {
1616                         rxm->offset += len;
1617                         rxm->full_len -= len;
1618                         return false;
1619                 }
1620                 consume_skb(skb);
1621         }
1622
1623         /* Finished with message */
1624         ctx->recv_pkt = NULL;
1625         __strp_unpause(&ctx->strp);
1626
1627         return true;
1628 }
1629
1630 /* This function traverses the rx_list in tls receive context to copies the
1631  * decrypted records into the buffer provided by caller zero copy is not
1632  * true. Further, the records are removed from the rx_list if it is not a peek
1633  * case and the record has been consumed completely.
1634  */
1635 static int process_rx_list(struct tls_sw_context_rx *ctx,
1636                            struct msghdr *msg,
1637                            u8 *control,
1638                            bool *cmsg,
1639                            size_t skip,
1640                            size_t len,
1641                            bool zc,
1642                            bool is_peek)
1643 {
1644         struct sk_buff *skb = skb_peek(&ctx->rx_list);
1645         u8 ctrl = *control;
1646         u8 msgc = *cmsg;
1647         struct tls_msg *tlm;
1648         ssize_t copied = 0;
1649
1650         /* Set the record type in 'control' if caller didn't pass it */
1651         if (!ctrl && skb) {
1652                 tlm = tls_msg(skb);
1653                 ctrl = tlm->control;
1654         }
1655
1656         while (skip && skb) {
1657                 struct strp_msg *rxm = strp_msg(skb);
1658                 tlm = tls_msg(skb);
1659
1660                 /* Cannot process a record of different type */
1661                 if (ctrl != tlm->control)
1662                         return 0;
1663
1664                 if (skip < rxm->full_len)
1665                         break;
1666
1667                 skip = skip - rxm->full_len;
1668                 skb = skb_peek_next(skb, &ctx->rx_list);
1669         }
1670
1671         while (len && skb) {
1672                 struct sk_buff *next_skb;
1673                 struct strp_msg *rxm = strp_msg(skb);
1674                 int chunk = min_t(unsigned int, rxm->full_len - skip, len);
1675
1676                 tlm = tls_msg(skb);
1677
1678                 /* Cannot process a record of different type */
1679                 if (ctrl != tlm->control)
1680                         return 0;
1681
1682                 /* Set record type if not already done. For a non-data record,
1683                  * do not proceed if record type could not be copied.
1684                  */
1685                 if (!msgc) {
1686                         int cerr = put_cmsg(msg, SOL_TLS, TLS_GET_RECORD_TYPE,
1687                                             sizeof(ctrl), &ctrl);
1688                         msgc = true;
1689                         if (ctrl != TLS_RECORD_TYPE_DATA) {
1690                                 if (cerr || msg->msg_flags & MSG_CTRUNC)
1691                                         return -EIO;
1692
1693                                 *cmsg = msgc;
1694                         }
1695                 }
1696
1697                 if (!zc || (rxm->full_len - skip) > len) {
1698                         int err = skb_copy_datagram_msg(skb, rxm->offset + skip,
1699                                                     msg, chunk);
1700                         if (err < 0)
1701                                 return err;
1702                 }
1703
1704                 len = len - chunk;
1705                 copied = copied + chunk;
1706
1707                 /* Consume the data from record if it is non-peek case*/
1708                 if (!is_peek) {
1709                         rxm->offset = rxm->offset + chunk;
1710                         rxm->full_len = rxm->full_len - chunk;
1711
1712                         /* Return if there is unconsumed data in the record */
1713                         if (rxm->full_len - skip)
1714                                 break;
1715                 }
1716
1717                 /* The remaining skip-bytes must lie in 1st record in rx_list.
1718                  * So from the 2nd record, 'skip' should be 0.
1719                  */
1720                 skip = 0;
1721
1722                 if (msg)
1723                         msg->msg_flags |= MSG_EOR;
1724
1725                 next_skb = skb_peek_next(skb, &ctx->rx_list);
1726
1727                 if (!is_peek) {
1728                         skb_unlink(skb, &ctx->rx_list);
1729                         consume_skb(skb);
1730                 }
1731
1732                 skb = next_skb;
1733         }
1734
1735         *control = ctrl;
1736         return copied;
1737 }
1738
1739 int tls_sw_recvmsg(struct sock *sk,
1740                    struct msghdr *msg,
1741                    size_t len,
1742                    int nonblock,
1743                    int flags,
1744                    int *addr_len)
1745 {
1746         struct tls_context *tls_ctx = tls_get_ctx(sk);
1747         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1748         struct tls_prot_info *prot = &tls_ctx->prot_info;
1749         struct sk_psock *psock;
1750         unsigned char control = 0;
1751         ssize_t decrypted = 0;
1752         struct strp_msg *rxm;
1753         struct tls_msg *tlm;
1754         struct sk_buff *skb;
1755         ssize_t copied = 0;
1756         bool cmsg = false;
1757         int target, err = 0;
1758         long timeo;
1759         bool is_kvec = iov_iter_is_kvec(&msg->msg_iter);
1760         bool is_peek = flags & MSG_PEEK;
1761         bool bpf_strp_enabled;
1762         int num_async = 0;
1763         int pending;
1764
1765         flags |= nonblock;
1766
1767         if (unlikely(flags & MSG_ERRQUEUE))
1768                 return sock_recv_errqueue(sk, msg, len, SOL_IP, IP_RECVERR);
1769
1770         psock = sk_psock_get(sk);
1771         lock_sock(sk);
1772         bpf_strp_enabled = sk_psock_strp_enabled(psock);
1773
1774         /* Process pending decrypted records. It must be non-zero-copy */
1775         err = process_rx_list(ctx, msg, &control, &cmsg, 0, len, false,
1776                               is_peek);
1777         if (err < 0) {
1778                 tls_err_abort(sk, err);
1779                 goto end;
1780         } else {
1781                 copied = err;
1782         }
1783
1784         if (len <= copied)
1785                 goto recv_end;
1786
1787         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1788         len = len - copied;
1789         timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
1790
1791         while (len && (decrypted + copied < target || ctx->recv_pkt)) {
1792                 bool retain_skb = false;
1793                 bool zc = false;
1794                 int to_decrypt;
1795                 int chunk = 0;
1796                 bool async_capable;
1797                 bool async = false;
1798
1799                 skb = tls_wait_data(sk, psock, flags & MSG_DONTWAIT, timeo, &err);
1800                 if (!skb) {
1801                         if (psock) {
1802                                 int ret = sk_msg_recvmsg(sk, psock, msg, len,
1803                                                          flags);
1804
1805                                 if (ret > 0) {
1806                                         decrypted += ret;
1807                                         len -= ret;
1808                                         continue;
1809                                 }
1810                         }
1811                         goto recv_end;
1812                 } else {
1813                         tlm = tls_msg(skb);
1814                         if (prot->version == TLS_1_3_VERSION)
1815                                 tlm->control = 0;
1816                         else
1817                                 tlm->control = ctx->control;
1818                 }
1819
1820                 rxm = strp_msg(skb);
1821
1822                 to_decrypt = rxm->full_len - prot->overhead_size;
1823
1824                 if (to_decrypt <= len && !is_kvec && !is_peek &&
1825                     ctx->control == TLS_RECORD_TYPE_DATA &&
1826                     prot->version != TLS_1_3_VERSION &&
1827                     !bpf_strp_enabled)
1828                         zc = true;
1829
1830                 /* Do not use async mode if record is non-data */
1831                 if (ctx->control == TLS_RECORD_TYPE_DATA && !bpf_strp_enabled)
1832                         async_capable = ctx->async_capable;
1833                 else
1834                         async_capable = false;
1835
1836                 err = decrypt_skb_update(sk, skb, &msg->msg_iter,
1837                                          &chunk, &zc, async_capable);
1838                 if (err < 0 && err != -EINPROGRESS) {
1839                         tls_err_abort(sk, -EBADMSG);
1840                         goto recv_end;
1841                 }
1842
1843                 if (err == -EINPROGRESS) {
1844                         async = true;
1845                         num_async++;
1846                 } else if (prot->version == TLS_1_3_VERSION) {
1847                         tlm->control = ctx->control;
1848                 }
1849
1850                 /* If the type of records being processed is not known yet,
1851                  * set it to record type just dequeued. If it is already known,
1852                  * but does not match the record type just dequeued, go to end.
1853                  * We always get record type here since for tls1.2, record type
1854                  * is known just after record is dequeued from stream parser.
1855                  * For tls1.3, we disable async.
1856                  */
1857
1858                 if (!control)
1859                         control = tlm->control;
1860                 else if (control != tlm->control)
1861                         goto recv_end;
1862
1863                 if (!cmsg) {
1864                         int cerr;
1865
1866                         cerr = put_cmsg(msg, SOL_TLS, TLS_GET_RECORD_TYPE,
1867                                         sizeof(control), &control);
1868                         cmsg = true;
1869                         if (control != TLS_RECORD_TYPE_DATA) {
1870                                 if (cerr || msg->msg_flags & MSG_CTRUNC) {
1871                                         err = -EIO;
1872                                         goto recv_end;
1873                                 }
1874                         }
1875                 }
1876
1877                 if (async)
1878                         goto pick_next_record;
1879
1880                 if (!zc) {
1881                         if (bpf_strp_enabled) {
1882                                 err = sk_psock_tls_strp_read(psock, skb);
1883                                 if (err != __SK_PASS) {
1884                                         rxm->offset = rxm->offset + rxm->full_len;
1885                                         rxm->full_len = 0;
1886                                         if (err == __SK_DROP)
1887                                                 consume_skb(skb);
1888                                         ctx->recv_pkt = NULL;
1889                                         __strp_unpause(&ctx->strp);
1890                                         continue;
1891                                 }
1892                         }
1893
1894                         if (rxm->full_len > len) {
1895                                 retain_skb = true;
1896                                 chunk = len;
1897                         } else {
1898                                 chunk = rxm->full_len;
1899                         }
1900
1901                         err = skb_copy_datagram_msg(skb, rxm->offset,
1902                                                     msg, chunk);
1903                         if (err < 0)
1904                                 goto recv_end;
1905
1906                         if (!is_peek) {
1907                                 rxm->offset = rxm->offset + chunk;
1908                                 rxm->full_len = rxm->full_len - chunk;
1909                         }
1910                 }
1911
1912 pick_next_record:
1913                 if (chunk > len)
1914                         chunk = len;
1915
1916                 decrypted += chunk;
1917                 len -= chunk;
1918
1919                 /* For async or peek case, queue the current skb */
1920                 if (async || is_peek || retain_skb) {
1921                         skb_queue_tail(&ctx->rx_list, skb);
1922                         skb = NULL;
1923                 }
1924
1925                 if (tls_sw_advance_skb(sk, skb, chunk)) {
1926                         /* Return full control message to
1927                          * userspace before trying to parse
1928                          * another message type
1929                          */
1930                         msg->msg_flags |= MSG_EOR;
1931                         if (control != TLS_RECORD_TYPE_DATA)
1932                                 goto recv_end;
1933                 } else {
1934                         break;
1935                 }
1936         }
1937
1938 recv_end:
1939         if (num_async) {
1940                 /* Wait for all previously submitted records to be decrypted */
1941                 spin_lock_bh(&ctx->decrypt_compl_lock);
1942                 ctx->async_notify = true;
1943                 pending = atomic_read(&ctx->decrypt_pending);
1944                 spin_unlock_bh(&ctx->decrypt_compl_lock);
1945                 if (pending) {
1946                         err = crypto_wait_req(-EINPROGRESS, &ctx->async_wait);
1947                         if (err) {
1948                                 /* one of async decrypt failed */
1949                                 tls_err_abort(sk, err);
1950                                 copied = 0;
1951                                 decrypted = 0;
1952                                 goto end;
1953                         }
1954                 } else {
1955                         reinit_completion(&ctx->async_wait.completion);
1956                 }
1957
1958                 /* There can be no concurrent accesses, since we have no
1959                  * pending decrypt operations
1960                  */
1961                 WRITE_ONCE(ctx->async_notify, false);
1962
1963                 /* Drain records from the rx_list & copy if required */
1964                 if (is_peek || is_kvec)
1965                         err = process_rx_list(ctx, msg, &control, &cmsg, copied,
1966                                               decrypted, false, is_peek);
1967                 else
1968                         err = process_rx_list(ctx, msg, &control, &cmsg, 0,
1969                                               decrypted, true, is_peek);
1970                 if (err < 0) {
1971                         tls_err_abort(sk, err);
1972                         copied = 0;
1973                         goto end;
1974                 }
1975         }
1976
1977         copied += decrypted;
1978
1979 end:
1980         release_sock(sk);
1981         if (psock)
1982                 sk_psock_put(sk, psock);
1983         return copied ? : err;
1984 }
1985
1986 ssize_t tls_sw_splice_read(struct socket *sock,  loff_t *ppos,
1987                            struct pipe_inode_info *pipe,
1988                            size_t len, unsigned int flags)
1989 {
1990         struct tls_context *tls_ctx = tls_get_ctx(sock->sk);
1991         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1992         struct strp_msg *rxm = NULL;
1993         struct sock *sk = sock->sk;
1994         struct sk_buff *skb;
1995         ssize_t copied = 0;
1996         bool from_queue;
1997         int err = 0;
1998         long timeo;
1999         int chunk;
2000         bool zc = false;
2001
2002         lock_sock(sk);
2003
2004         timeo = sock_rcvtimeo(sk, flags & SPLICE_F_NONBLOCK);
2005
2006         from_queue = !skb_queue_empty(&ctx->rx_list);
2007         if (from_queue) {
2008                 skb = __skb_dequeue(&ctx->rx_list);
2009         } else {
2010                 skb = tls_wait_data(sk, NULL, flags & SPLICE_F_NONBLOCK, timeo,
2011                                     &err);
2012                 if (!skb)
2013                         goto splice_read_end;
2014
2015                 err = decrypt_skb_update(sk, skb, NULL, &chunk, &zc, false);
2016                 if (err < 0) {
2017                         tls_err_abort(sk, -EBADMSG);
2018                         goto splice_read_end;
2019                 }
2020         }
2021
2022         /* splice does not support reading control messages */
2023         if (ctx->control != TLS_RECORD_TYPE_DATA) {
2024                 err = -EINVAL;
2025                 goto splice_read_end;
2026         }
2027
2028         rxm = strp_msg(skb);
2029
2030         chunk = min_t(unsigned int, rxm->full_len, len);
2031         copied = skb_splice_bits(skb, sk, rxm->offset, pipe, chunk, flags);
2032         if (copied < 0)
2033                 goto splice_read_end;
2034
2035         if (!from_queue) {
2036                 ctx->recv_pkt = NULL;
2037                 __strp_unpause(&ctx->strp);
2038         }
2039         if (chunk < rxm->full_len) {
2040                 __skb_queue_head(&ctx->rx_list, skb);
2041                 rxm->offset += len;
2042                 rxm->full_len -= len;
2043         } else {
2044                 consume_skb(skb);
2045         }
2046
2047 splice_read_end:
2048         release_sock(sk);
2049         return copied ? : err;
2050 }
2051
2052 bool tls_sw_sock_is_readable(struct sock *sk)
2053 {
2054         struct tls_context *tls_ctx = tls_get_ctx(sk);
2055         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2056         bool ingress_empty = true;
2057         struct sk_psock *psock;
2058
2059         rcu_read_lock();
2060         psock = sk_psock(sk);
2061         if (psock)
2062                 ingress_empty = list_empty(&psock->ingress_msg);
2063         rcu_read_unlock();
2064
2065         return !ingress_empty || ctx->recv_pkt ||
2066                 !skb_queue_empty(&ctx->rx_list);
2067 }
2068
2069 static int tls_read_size(struct strparser *strp, struct sk_buff *skb)
2070 {
2071         struct tls_context *tls_ctx = tls_get_ctx(strp->sk);
2072         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2073         struct tls_prot_info *prot = &tls_ctx->prot_info;
2074         char header[TLS_HEADER_SIZE + MAX_IV_SIZE];
2075         struct strp_msg *rxm = strp_msg(skb);
2076         size_t cipher_overhead;
2077         size_t data_len = 0;
2078         int ret;
2079
2080         /* Verify that we have a full TLS header, or wait for more data */
2081         if (rxm->offset + prot->prepend_size > skb->len)
2082                 return 0;
2083
2084         /* Sanity-check size of on-stack buffer. */
2085         if (WARN_ON(prot->prepend_size > sizeof(header))) {
2086                 ret = -EINVAL;
2087                 goto read_failure;
2088         }
2089
2090         /* Linearize header to local buffer */
2091         ret = skb_copy_bits(skb, rxm->offset, header, prot->prepend_size);
2092
2093         if (ret < 0)
2094                 goto read_failure;
2095
2096         ctx->control = header[0];
2097
2098         data_len = ((header[4] & 0xFF) | (header[3] << 8));
2099
2100         cipher_overhead = prot->tag_size;
2101         if (prot->version != TLS_1_3_VERSION &&
2102             prot->cipher_type != TLS_CIPHER_CHACHA20_POLY1305)
2103                 cipher_overhead += prot->iv_size;
2104
2105         if (data_len > TLS_MAX_PAYLOAD_SIZE + cipher_overhead +
2106             prot->tail_size) {
2107                 ret = -EMSGSIZE;
2108                 goto read_failure;
2109         }
2110         if (data_len < cipher_overhead) {
2111                 ret = -EBADMSG;
2112                 goto read_failure;
2113         }
2114
2115         /* Note that both TLS1.3 and TLS1.2 use TLS_1_2 version here */
2116         if (header[1] != TLS_1_2_VERSION_MINOR ||
2117             header[2] != TLS_1_2_VERSION_MAJOR) {
2118                 ret = -EINVAL;
2119                 goto read_failure;
2120         }
2121
2122         tls_device_rx_resync_new_rec(strp->sk, data_len + TLS_HEADER_SIZE,
2123                                      TCP_SKB_CB(skb)->seq + rxm->offset);
2124         return data_len + TLS_HEADER_SIZE;
2125
2126 read_failure:
2127         tls_err_abort(strp->sk, ret);
2128
2129         return ret;
2130 }
2131
2132 static void tls_queue(struct strparser *strp, struct sk_buff *skb)
2133 {
2134         struct tls_context *tls_ctx = tls_get_ctx(strp->sk);
2135         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2136
2137         ctx->decrypted = 0;
2138
2139         ctx->recv_pkt = skb;
2140         strp_pause(strp);
2141
2142         ctx->saved_data_ready(strp->sk);
2143 }
2144
2145 static void tls_data_ready(struct sock *sk)
2146 {
2147         struct tls_context *tls_ctx = tls_get_ctx(sk);
2148         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2149         struct sk_psock *psock;
2150
2151         strp_data_ready(&ctx->strp);
2152
2153         psock = sk_psock_get(sk);
2154         if (psock) {
2155                 if (!list_empty(&psock->ingress_msg))
2156                         ctx->saved_data_ready(sk);
2157                 sk_psock_put(sk, psock);
2158         }
2159 }
2160
2161 void tls_sw_cancel_work_tx(struct tls_context *tls_ctx)
2162 {
2163         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
2164
2165         set_bit(BIT_TX_CLOSING, &ctx->tx_bitmask);
2166         set_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask);
2167         cancel_delayed_work_sync(&ctx->tx_work.work);
2168 }
2169
2170 void tls_sw_release_resources_tx(struct sock *sk)
2171 {
2172         struct tls_context *tls_ctx = tls_get_ctx(sk);
2173         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
2174         struct tls_rec *rec, *tmp;
2175         int pending;
2176
2177         /* Wait for any pending async encryptions to complete */
2178         spin_lock_bh(&ctx->encrypt_compl_lock);
2179         ctx->async_notify = true;
2180         pending = atomic_read(&ctx->encrypt_pending);
2181         spin_unlock_bh(&ctx->encrypt_compl_lock);
2182
2183         if (pending)
2184                 crypto_wait_req(-EINPROGRESS, &ctx->async_wait);
2185
2186         tls_tx_records(sk, -1);
2187
2188         /* Free up un-sent records in tx_list. First, free
2189          * the partially sent record if any at head of tx_list.
2190          */
2191         if (tls_ctx->partially_sent_record) {
2192                 tls_free_partial_record(sk, tls_ctx);
2193                 rec = list_first_entry(&ctx->tx_list,
2194                                        struct tls_rec, list);
2195                 list_del(&rec->list);
2196                 sk_msg_free(sk, &rec->msg_plaintext);
2197                 kfree(rec);
2198         }
2199
2200         list_for_each_entry_safe(rec, tmp, &ctx->tx_list, list) {
2201                 list_del(&rec->list);
2202                 sk_msg_free(sk, &rec->msg_encrypted);
2203                 sk_msg_free(sk, &rec->msg_plaintext);
2204                 kfree(rec);
2205         }
2206
2207         crypto_free_aead(ctx->aead_send);
2208         tls_free_open_rec(sk);
2209 }
2210
2211 void tls_sw_free_ctx_tx(struct tls_context *tls_ctx)
2212 {
2213         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
2214
2215         kfree(ctx);
2216 }
2217
2218 void tls_sw_release_resources_rx(struct sock *sk)
2219 {
2220         struct tls_context *tls_ctx = tls_get_ctx(sk);
2221         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2222
2223         kfree(tls_ctx->rx.rec_seq);
2224         kfree(tls_ctx->rx.iv);
2225
2226         if (ctx->aead_recv) {
2227                 kfree_skb(ctx->recv_pkt);
2228                 ctx->recv_pkt = NULL;
2229                 skb_queue_purge(&ctx->rx_list);
2230                 crypto_free_aead(ctx->aead_recv);
2231                 strp_stop(&ctx->strp);
2232                 /* If tls_sw_strparser_arm() was not called (cleanup paths)
2233                  * we still want to strp_stop(), but sk->sk_data_ready was
2234                  * never swapped.
2235                  */
2236                 if (ctx->saved_data_ready) {
2237                         write_lock_bh(&sk->sk_callback_lock);
2238                         sk->sk_data_ready = ctx->saved_data_ready;
2239                         write_unlock_bh(&sk->sk_callback_lock);
2240                 }
2241         }
2242 }
2243
2244 void tls_sw_strparser_done(struct tls_context *tls_ctx)
2245 {
2246         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2247
2248         strp_done(&ctx->strp);
2249 }
2250
2251 void tls_sw_free_ctx_rx(struct tls_context *tls_ctx)
2252 {
2253         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2254
2255         kfree(ctx);
2256 }
2257
2258 void tls_sw_free_resources_rx(struct sock *sk)
2259 {
2260         struct tls_context *tls_ctx = tls_get_ctx(sk);
2261
2262         tls_sw_release_resources_rx(sk);
2263         tls_sw_free_ctx_rx(tls_ctx);
2264 }
2265
2266 /* The work handler to transmitt the encrypted records in tx_list */
2267 static void tx_work_handler(struct work_struct *work)
2268 {
2269         struct delayed_work *delayed_work = to_delayed_work(work);
2270         struct tx_work *tx_work = container_of(delayed_work,
2271                                                struct tx_work, work);
2272         struct sock *sk = tx_work->sk;
2273         struct tls_context *tls_ctx = tls_get_ctx(sk);
2274         struct tls_sw_context_tx *ctx;
2275
2276         if (unlikely(!tls_ctx))
2277                 return;
2278
2279         ctx = tls_sw_ctx_tx(tls_ctx);
2280         if (test_bit(BIT_TX_CLOSING, &ctx->tx_bitmask))
2281                 return;
2282
2283         if (!test_and_clear_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask))
2284                 return;
2285         mutex_lock(&tls_ctx->tx_lock);
2286         lock_sock(sk);
2287         tls_tx_records(sk, -1);
2288         release_sock(sk);
2289         mutex_unlock(&tls_ctx->tx_lock);
2290 }
2291
2292 void tls_sw_write_space(struct sock *sk, struct tls_context *ctx)
2293 {
2294         struct tls_sw_context_tx *tx_ctx = tls_sw_ctx_tx(ctx);
2295
2296         /* Schedule the transmission if tx list is ready */
2297         if (is_tx_ready(tx_ctx) &&
2298             !test_and_set_bit(BIT_TX_SCHEDULED, &tx_ctx->tx_bitmask))
2299                 schedule_delayed_work(&tx_ctx->tx_work.work, 0);
2300 }
2301
2302 void tls_sw_strparser_arm(struct sock *sk, struct tls_context *tls_ctx)
2303 {
2304         struct tls_sw_context_rx *rx_ctx = tls_sw_ctx_rx(tls_ctx);
2305
2306         write_lock_bh(&sk->sk_callback_lock);
2307         rx_ctx->saved_data_ready = sk->sk_data_ready;
2308         sk->sk_data_ready = tls_data_ready;
2309         write_unlock_bh(&sk->sk_callback_lock);
2310
2311         strp_check_rcv(&rx_ctx->strp);
2312 }
2313
2314 int tls_set_sw_offload(struct sock *sk, struct tls_context *ctx, int tx)
2315 {
2316         struct tls_context *tls_ctx = tls_get_ctx(sk);
2317         struct tls_prot_info *prot = &tls_ctx->prot_info;
2318         struct tls_crypto_info *crypto_info;
2319         struct tls12_crypto_info_aes_gcm_128 *gcm_128_info;
2320         struct tls12_crypto_info_aes_gcm_256 *gcm_256_info;
2321         struct tls12_crypto_info_aes_ccm_128 *ccm_128_info;
2322         struct tls12_crypto_info_chacha20_poly1305 *chacha20_poly1305_info;
2323         struct tls_sw_context_tx *sw_ctx_tx = NULL;
2324         struct tls_sw_context_rx *sw_ctx_rx = NULL;
2325         struct cipher_context *cctx;
2326         struct crypto_aead **aead;
2327         struct strp_callbacks cb;
2328         u16 nonce_size, tag_size, iv_size, rec_seq_size, salt_size;
2329         struct crypto_tfm *tfm;
2330         char *iv, *rec_seq, *key, *salt, *cipher_name;
2331         size_t keysize;
2332         int rc = 0;
2333
2334         if (!ctx) {
2335                 rc = -EINVAL;
2336                 goto out;
2337         }
2338
2339         if (tx) {
2340                 if (!ctx->priv_ctx_tx) {
2341                         sw_ctx_tx = kzalloc(sizeof(*sw_ctx_tx), GFP_KERNEL);
2342                         if (!sw_ctx_tx) {
2343                                 rc = -ENOMEM;
2344                                 goto out;
2345                         }
2346                         ctx->priv_ctx_tx = sw_ctx_tx;
2347                 } else {
2348                         sw_ctx_tx =
2349                                 (struct tls_sw_context_tx *)ctx->priv_ctx_tx;
2350                 }
2351         } else {
2352                 if (!ctx->priv_ctx_rx) {
2353                         sw_ctx_rx = kzalloc(sizeof(*sw_ctx_rx), GFP_KERNEL);
2354                         if (!sw_ctx_rx) {
2355                                 rc = -ENOMEM;
2356                                 goto out;
2357                         }
2358                         ctx->priv_ctx_rx = sw_ctx_rx;
2359                 } else {
2360                         sw_ctx_rx =
2361                                 (struct tls_sw_context_rx *)ctx->priv_ctx_rx;
2362                 }
2363         }
2364
2365         if (tx) {
2366                 crypto_init_wait(&sw_ctx_tx->async_wait);
2367                 spin_lock_init(&sw_ctx_tx->encrypt_compl_lock);
2368                 crypto_info = &ctx->crypto_send.info;
2369                 cctx = &ctx->tx;
2370                 aead = &sw_ctx_tx->aead_send;
2371                 INIT_LIST_HEAD(&sw_ctx_tx->tx_list);
2372                 INIT_DELAYED_WORK(&sw_ctx_tx->tx_work.work, tx_work_handler);
2373                 sw_ctx_tx->tx_work.sk = sk;
2374         } else {
2375                 crypto_init_wait(&sw_ctx_rx->async_wait);
2376                 spin_lock_init(&sw_ctx_rx->decrypt_compl_lock);
2377                 crypto_info = &ctx->crypto_recv.info;
2378                 cctx = &ctx->rx;
2379                 skb_queue_head_init(&sw_ctx_rx->rx_list);
2380                 aead = &sw_ctx_rx->aead_recv;
2381         }
2382
2383         switch (crypto_info->cipher_type) {
2384         case TLS_CIPHER_AES_GCM_128: {
2385                 nonce_size = TLS_CIPHER_AES_GCM_128_IV_SIZE;
2386                 tag_size = TLS_CIPHER_AES_GCM_128_TAG_SIZE;
2387                 iv_size = TLS_CIPHER_AES_GCM_128_IV_SIZE;
2388                 iv = ((struct tls12_crypto_info_aes_gcm_128 *)crypto_info)->iv;
2389                 rec_seq_size = TLS_CIPHER_AES_GCM_128_REC_SEQ_SIZE;
2390                 rec_seq =
2391                  ((struct tls12_crypto_info_aes_gcm_128 *)crypto_info)->rec_seq;
2392                 gcm_128_info =
2393                         (struct tls12_crypto_info_aes_gcm_128 *)crypto_info;
2394                 keysize = TLS_CIPHER_AES_GCM_128_KEY_SIZE;
2395                 key = gcm_128_info->key;
2396                 salt = gcm_128_info->salt;
2397                 salt_size = TLS_CIPHER_AES_GCM_128_SALT_SIZE;
2398                 cipher_name = "gcm(aes)";
2399                 break;
2400         }
2401         case TLS_CIPHER_AES_GCM_256: {
2402                 nonce_size = TLS_CIPHER_AES_GCM_256_IV_SIZE;
2403                 tag_size = TLS_CIPHER_AES_GCM_256_TAG_SIZE;
2404                 iv_size = TLS_CIPHER_AES_GCM_256_IV_SIZE;
2405                 iv = ((struct tls12_crypto_info_aes_gcm_256 *)crypto_info)->iv;
2406                 rec_seq_size = TLS_CIPHER_AES_GCM_256_REC_SEQ_SIZE;
2407                 rec_seq =
2408                  ((struct tls12_crypto_info_aes_gcm_256 *)crypto_info)->rec_seq;
2409                 gcm_256_info =
2410                         (struct tls12_crypto_info_aes_gcm_256 *)crypto_info;
2411                 keysize = TLS_CIPHER_AES_GCM_256_KEY_SIZE;
2412                 key = gcm_256_info->key;
2413                 salt = gcm_256_info->salt;
2414                 salt_size = TLS_CIPHER_AES_GCM_256_SALT_SIZE;
2415                 cipher_name = "gcm(aes)";
2416                 break;
2417         }
2418         case TLS_CIPHER_AES_CCM_128: {
2419                 nonce_size = TLS_CIPHER_AES_CCM_128_IV_SIZE;
2420                 tag_size = TLS_CIPHER_AES_CCM_128_TAG_SIZE;
2421                 iv_size = TLS_CIPHER_AES_CCM_128_IV_SIZE;
2422                 iv = ((struct tls12_crypto_info_aes_ccm_128 *)crypto_info)->iv;
2423                 rec_seq_size = TLS_CIPHER_AES_CCM_128_REC_SEQ_SIZE;
2424                 rec_seq =
2425                 ((struct tls12_crypto_info_aes_ccm_128 *)crypto_info)->rec_seq;
2426                 ccm_128_info =
2427                 (struct tls12_crypto_info_aes_ccm_128 *)crypto_info;
2428                 keysize = TLS_CIPHER_AES_CCM_128_KEY_SIZE;
2429                 key = ccm_128_info->key;
2430                 salt = ccm_128_info->salt;
2431                 salt_size = TLS_CIPHER_AES_CCM_128_SALT_SIZE;
2432                 cipher_name = "ccm(aes)";
2433                 break;
2434         }
2435         case TLS_CIPHER_CHACHA20_POLY1305: {
2436                 chacha20_poly1305_info = (void *)crypto_info;
2437                 nonce_size = 0;
2438                 tag_size = TLS_CIPHER_CHACHA20_POLY1305_TAG_SIZE;
2439                 iv_size = TLS_CIPHER_CHACHA20_POLY1305_IV_SIZE;
2440                 iv = chacha20_poly1305_info->iv;
2441                 rec_seq_size = TLS_CIPHER_CHACHA20_POLY1305_REC_SEQ_SIZE;
2442                 rec_seq = chacha20_poly1305_info->rec_seq;
2443                 keysize = TLS_CIPHER_CHACHA20_POLY1305_KEY_SIZE;
2444                 key = chacha20_poly1305_info->key;
2445                 salt = chacha20_poly1305_info->salt;
2446                 salt_size = TLS_CIPHER_CHACHA20_POLY1305_SALT_SIZE;
2447                 cipher_name = "rfc7539(chacha20,poly1305)";
2448                 break;
2449         }
2450         default:
2451                 rc = -EINVAL;
2452                 goto free_priv;
2453         }
2454
2455         /* Sanity-check the sizes for stack allocations. */
2456         if (iv_size > MAX_IV_SIZE || nonce_size > MAX_IV_SIZE ||
2457             rec_seq_size > TLS_MAX_REC_SEQ_SIZE) {
2458                 rc = -EINVAL;
2459                 goto free_priv;
2460         }
2461
2462         if (crypto_info->version == TLS_1_3_VERSION) {
2463                 nonce_size = 0;
2464                 prot->aad_size = TLS_HEADER_SIZE;
2465                 prot->tail_size = 1;
2466         } else {
2467                 prot->aad_size = TLS_AAD_SPACE_SIZE;
2468                 prot->tail_size = 0;
2469         }
2470
2471         prot->version = crypto_info->version;
2472         prot->cipher_type = crypto_info->cipher_type;
2473         prot->prepend_size = TLS_HEADER_SIZE + nonce_size;
2474         prot->tag_size = tag_size;
2475         prot->overhead_size = prot->prepend_size +
2476                               prot->tag_size + prot->tail_size;
2477         prot->iv_size = iv_size;
2478         prot->salt_size = salt_size;
2479         cctx->iv = kmalloc(iv_size + salt_size, GFP_KERNEL);
2480         if (!cctx->iv) {
2481                 rc = -ENOMEM;
2482                 goto free_priv;
2483         }
2484         /* Note: 128 & 256 bit salt are the same size */
2485         prot->rec_seq_size = rec_seq_size;
2486         memcpy(cctx->iv, salt, salt_size);
2487         memcpy(cctx->iv + salt_size, iv, iv_size);
2488         cctx->rec_seq = kmemdup(rec_seq, rec_seq_size, GFP_KERNEL);
2489         if (!cctx->rec_seq) {
2490                 rc = -ENOMEM;
2491                 goto free_iv;
2492         }
2493
2494         if (!*aead) {
2495                 *aead = crypto_alloc_aead(cipher_name, 0, 0);
2496                 if (IS_ERR(*aead)) {
2497                         rc = PTR_ERR(*aead);
2498                         *aead = NULL;
2499                         goto free_rec_seq;
2500                 }
2501         }
2502
2503         ctx->push_pending_record = tls_sw_push_pending_record;
2504
2505         rc = crypto_aead_setkey(*aead, key, keysize);
2506
2507         if (rc)
2508                 goto free_aead;
2509
2510         rc = crypto_aead_setauthsize(*aead, prot->tag_size);
2511         if (rc)
2512                 goto free_aead;
2513
2514         if (sw_ctx_rx) {
2515                 tfm = crypto_aead_tfm(sw_ctx_rx->aead_recv);
2516
2517                 if (crypto_info->version == TLS_1_3_VERSION)
2518                         sw_ctx_rx->async_capable = 0;
2519                 else
2520                         sw_ctx_rx->async_capable =
2521                                 !!(tfm->__crt_alg->cra_flags &
2522                                    CRYPTO_ALG_ASYNC);
2523
2524                 /* Set up strparser */
2525                 memset(&cb, 0, sizeof(cb));
2526                 cb.rcv_msg = tls_queue;
2527                 cb.parse_msg = tls_read_size;
2528
2529                 strp_init(&sw_ctx_rx->strp, sk, &cb);
2530         }
2531
2532         goto out;
2533
2534 free_aead:
2535         crypto_free_aead(*aead);
2536         *aead = NULL;
2537 free_rec_seq:
2538         kfree(cctx->rec_seq);
2539         cctx->rec_seq = NULL;
2540 free_iv:
2541         kfree(cctx->iv);
2542         cctx->iv = NULL;
2543 free_priv:
2544         if (tx) {
2545                 kfree(ctx->priv_ctx_tx);
2546                 ctx->priv_ctx_tx = NULL;
2547         } else {
2548                 kfree(ctx->priv_ctx_rx);
2549                 ctx->priv_ctx_rx = NULL;
2550         }
2551 out:
2552         return rc;
2553 }