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, redir_ingress;
 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 && err != -EINPROGRESS && 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 && err != -EINPROGRESS && 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                 redir_ingress = psock->redir_ingress;
 850                 sk_redir = psock->sk_redir;
 851                 memcpy(&msg_redir, msg, sizeof(*msg));
 852                 if (msg->apply_bytes < send)
 853                         msg->apply_bytes = 0;
 854                 else
 855                         msg->apply_bytes -= send;
 856                 sk_msg_return_zero(sk, msg, send);
 857                 msg->sg.size -= send;
 858                 release_sock(sk);
 859                 err = tcp_bpf_sendmsg_redir(sk_redir, redir_ingress,
 860                                             &msg_redir, send, flags);
 861                 lock_sock(sk);
 862                 if (err < 0) {
 863                         *copied -= sk_msg_free_nocharge(sk, &msg_redir);
 864                         msg->sg.size = 0;
 865                 }
 866                 if (msg->sg.size == 0)
 867                         tls_free_open_rec(sk);
 868                 break;
 869         case __SK_DROP:
 870         default:
 871                 sk_msg_free_partial(sk, msg, send);
 872                 if (msg->apply_bytes < send)
 873                         msg->apply_bytes = 0;
 874                 else
 875                         msg->apply_bytes -= send;
 876                 if (msg->sg.size == 0)
 877                         tls_free_open_rec(sk);
 878                 *copied -= (send + delta);
 879                 err = -EACCES;
 880         }
 881
 882         if (likely(!err)) {
 883                 bool reset_eval = !ctx->open_rec;
 884
 885                 rec = ctx->open_rec;
 886                 if (rec) {
 887                         msg = &rec->msg_plaintext;
 888                         if (!msg->apply_bytes)
 889                                 reset_eval = true;
 890                 }
 891                 if (reset_eval) {
 892                         psock->eval = __SK_NONE;
 893                         if (psock->sk_redir) {
 894                                 sock_put(psock->sk_redir);
 895                                 psock->sk_redir = NULL;
 896                         }
 897                 }
 898                 if (rec)
 899                         goto more_data;
 900         }
 901  out_err:
 902         sk_psock_put(sk, psock);
 903         return err;
 904 }
 905
 906 static int tls_sw_push_pending_record(struct sock *sk, int flags)
 907 {
 908         struct tls_context *tls_ctx = tls_get_ctx(sk);
 909         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 910         struct tls_rec *rec = ctx->open_rec;
 911         struct sk_msg *msg_pl;
 912         size_t copied;
 913
 914         if (!rec)
 915                 return 0;
 916
 917         msg_pl = &rec->msg_plaintext;
 918         copied = msg_pl->sg.size;
 919         if (!copied)
 920                 return 0;
 921
 922         return bpf_exec_tx_verdict(msg_pl, sk, true, TLS_RECORD_TYPE_DATA,
 923                                    &copied, flags);
 924 }
 925
 926 int tls_sw_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
 927 {
 928         long timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
 929         struct tls_context *tls_ctx = tls_get_ctx(sk);
 930         struct tls_prot_info *prot = &tls_ctx->prot_info;
 931         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 932         bool async_capable = ctx->async_capable;
 933         unsigned char record_type = TLS_RECORD_TYPE_DATA;
 934         bool is_kvec = iov_iter_is_kvec(&msg->msg_iter);
 935         bool eor = !(msg->msg_flags & MSG_MORE);
 936         size_t try_to_copy;
 937         ssize_t copied = 0;
 938         struct sk_msg *msg_pl, *msg_en;
 939         struct tls_rec *rec;
 940         int required_size;
 941         int num_async = 0;
 942         bool full_record;
 943         int record_room;
 944         int num_zc = 0;
 945         int orig_size;
 946         int ret = 0;
 947         int pending;
 948
 949         if (msg->msg_flags & ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL |
 950                                MSG_CMSG_COMPAT))
 951                 return -EOPNOTSUPP;
 952
 953         ret = mutex_lock_interruptible(&tls_ctx->tx_lock);
 954         if (ret)
 955                 return ret;
 956         lock_sock(sk);
 957
 958         if (unlikely(msg->msg_controllen)) {
 959                 ret = tls_proccess_cmsg(sk, msg, &record_type);
 960                 if (ret) {
 961                         if (ret == -EINPROGRESS)
 962                                 num_async++;
 963                         else if (ret != -EAGAIN)
 964                                 goto send_end;
 965                 }
 966         }
 967
 968         while (msg_data_left(msg)) {
 969                 if (sk->sk_err) {
 970                         ret = -sk->sk_err;
 971                         goto send_end;
 972                 }
 973
 974                 if (ctx->open_rec)
 975                         rec = ctx->open_rec;
 976                 else
 977                         rec = ctx->open_rec = tls_get_rec(sk);
 978                 if (!rec) {
 979                         ret = -ENOMEM;
 980                         goto send_end;
 981                 }
 982
 983                 msg_pl = &rec->msg_plaintext;
 984                 msg_en = &rec->msg_encrypted;
 985
 986                 orig_size = msg_pl->sg.size;
 987                 full_record = false;
 988                 try_to_copy = msg_data_left(msg);
 989                 record_room = TLS_MAX_PAYLOAD_SIZE - msg_pl->sg.size;
 990                 if (try_to_copy >= record_room) {
 991                         try_to_copy = record_room;
 992                         full_record = true;
 993                 }
 994
 995                 required_size = msg_pl->sg.size + try_to_copy +
 996                                 prot->overhead_size;
 997
 998                 if (!sk_stream_memory_free(sk))
 999                         goto wait_for_sndbuf;
1000
1001 alloc_encrypted:
1002                 ret = tls_alloc_encrypted_msg(sk, required_size);
1003                 if (ret) {
1004                         if (ret != -ENOSPC)
1005                                 goto wait_for_memory;
1006
1007                         /* Adjust try_to_copy according to the amount that was
1008                          * actually allocated. The difference is due
1009                          * to max sg elements limit
1010                          */
1011                         try_to_copy -= required_size - msg_en->sg.size;
1012                         full_record = true;
1013                 }
1014
1015                 if (!is_kvec && (full_record || eor) && !async_capable) {
1016                         u32 first = msg_pl->sg.end;
1017
1018                         ret = sk_msg_zerocopy_from_iter(sk, &msg->msg_iter,
1019                                                         msg_pl, try_to_copy);
1020                         if (ret)
1021                                 goto fallback_to_reg_send;
1022
1023                         num_zc++;
1024                         copied += try_to_copy;
1025
1026                         sk_msg_sg_copy_set(msg_pl, first);
1027                         ret = bpf_exec_tx_verdict(msg_pl, sk, full_record,
1028                                                   record_type, &copied,
1029                                                   msg->msg_flags);
1030                         if (ret) {
1031                                 if (ret == -EINPROGRESS)
1032                                         num_async++;
1033                                 else if (ret == -ENOMEM)
1034                                         goto wait_for_memory;
1035                                 else if (ctx->open_rec && ret == -ENOSPC)
1036                                         goto rollback_iter;
1037                                 else if (ret != -EAGAIN)
1038                                         goto send_end;
1039                         }
1040                         continue;
1041 rollback_iter:
1042                         copied -= try_to_copy;
1043                         sk_msg_sg_copy_clear(msg_pl, first);
1044                         iov_iter_revert(&msg->msg_iter,
1045                                         msg_pl->sg.size - orig_size);
1046 fallback_to_reg_send:
1047                         sk_msg_trim(sk, msg_pl, orig_size);
1048                 }
1049
1050                 required_size = msg_pl->sg.size + try_to_copy;
1051
1052                 ret = tls_clone_plaintext_msg(sk, required_size);
1053                 if (ret) {
1054                         if (ret != -ENOSPC)
1055                                 goto send_end;
1056
1057                         /* Adjust try_to_copy according to the amount that was
1058                          * actually allocated. The difference is due
1059                          * to max sg elements limit
1060                          */
1061                         try_to_copy -= required_size - msg_pl->sg.size;
1062                         full_record = true;
1063                         sk_msg_trim(sk, msg_en,
1064                                     msg_pl->sg.size + prot->overhead_size);
1065                 }
1066
1067                 if (try_to_copy) {
1068                         ret = sk_msg_memcopy_from_iter(sk, &msg->msg_iter,
1069                                                        msg_pl, try_to_copy);
1070                         if (ret < 0)
1071                                 goto trim_sgl;
1072                 }
1073
1074                 /* Open records defined only if successfully copied, otherwise
1075                  * we would trim the sg but not reset the open record frags.
1076                  */
1077                 tls_ctx->pending_open_record_frags = true;
1078                 copied += try_to_copy;
1079                 if (full_record || eor) {
1080                         ret = bpf_exec_tx_verdict(msg_pl, sk, full_record,
1081                                                   record_type, &copied,
1082                                                   msg->msg_flags);
1083                         if (ret) {
1084                                 if (ret == -EINPROGRESS)
1085                                         num_async++;
1086                                 else if (ret == -ENOMEM)
1087                                         goto wait_for_memory;
1088                                 else if (ret != -EAGAIN) {
1089                                         if (ret == -ENOSPC)
1090                                                 ret = 0;
1091                                         goto send_end;
1092                                 }
1093                         }
1094                 }
1095
1096                 continue;
1097
1098 wait_for_sndbuf:
1099                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1100 wait_for_memory:
1101                 ret = sk_stream_wait_memory(sk, &timeo);
1102                 if (ret) {
1103 trim_sgl:
1104                         if (ctx->open_rec)
1105                                 tls_trim_both_msgs(sk, orig_size);
1106                         goto send_end;
1107                 }
1108
1109                 if (ctx->open_rec && msg_en->sg.size < required_size)
1110                         goto alloc_encrypted;
1111         }
1112
1113         if (!num_async) {
1114                 goto send_end;
1115         } else if (num_zc) {
1116                 /* Wait for pending encryptions to get completed */
1117                 spin_lock_bh(&ctx->encrypt_compl_lock);
1118                 ctx->async_notify = true;
1119
1120                 pending = atomic_read(&ctx->encrypt_pending);
1121                 spin_unlock_bh(&ctx->encrypt_compl_lock);
1122                 if (pending)
1123                         crypto_wait_req(-EINPROGRESS, &ctx->async_wait);
1124                 else
1125                         reinit_completion(&ctx->async_wait.completion);
1126
1127                 /* There can be no concurrent accesses, since we have no
1128                  * pending encrypt operations
1129                  */
1130                 WRITE_ONCE(ctx->async_notify, false);
1131
1132                 if (ctx->async_wait.err) {
1133                         ret = ctx->async_wait.err;
1134                         copied = 0;
1135                 }
1136         }
1137
1138         /* Transmit if any encryptions have completed */
1139         if (test_and_clear_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask)) {
1140                 cancel_delayed_work(&ctx->tx_work.work);
1141                 tls_tx_records(sk, msg->msg_flags);
1142         }
1143
1144 send_end:
1145         ret = sk_stream_error(sk, msg->msg_flags, ret);
1146
1147         release_sock(sk);
1148         mutex_unlock(&tls_ctx->tx_lock);
1149         return copied > 0 ? copied : ret;
1150 }
1151
1152 static int tls_sw_do_sendpage(struct sock *sk, struct page *page,
1153                               int offset, size_t size, int flags)
1154 {
1155         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1156         struct tls_context *tls_ctx = tls_get_ctx(sk);
1157         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
1158         struct tls_prot_info *prot = &tls_ctx->prot_info;
1159         unsigned char record_type = TLS_RECORD_TYPE_DATA;
1160         struct sk_msg *msg_pl;
1161         struct tls_rec *rec;
1162         int num_async = 0;
1163         ssize_t copied = 0;
1164         bool full_record;
1165         int record_room;
1166         int ret = 0;
1167         bool eor;
1168
1169         eor = !(flags & MSG_SENDPAGE_NOTLAST);
1170         sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1171
1172         /* Call the sk_stream functions to manage the sndbuf mem. */
1173         while (size > 0) {
1174                 size_t copy, required_size;
1175
1176                 if (sk->sk_err) {
1177                         ret = -sk->sk_err;
1178                         goto sendpage_end;
1179                 }
1180
1181                 if (ctx->open_rec)
1182                         rec = ctx->open_rec;
1183                 else
1184                         rec = ctx->open_rec = tls_get_rec(sk);
1185                 if (!rec) {
1186                         ret = -ENOMEM;
1187                         goto sendpage_end;
1188                 }
1189
1190                 msg_pl = &rec->msg_plaintext;
1191
1192                 full_record = false;
1193                 record_room = TLS_MAX_PAYLOAD_SIZE - msg_pl->sg.size;
1194                 copy = size;
1195                 if (copy >= record_room) {
1196                         copy = record_room;
1197                         full_record = true;
1198                 }
1199
1200                 required_size = msg_pl->sg.size + copy + prot->overhead_size;
1201
1202                 if (!sk_stream_memory_free(sk))
1203                         goto wait_for_sndbuf;
1204 alloc_payload:
1205                 ret = tls_alloc_encrypted_msg(sk, required_size);
1206                 if (ret) {
1207                         if (ret != -ENOSPC)
1208                                 goto wait_for_memory;
1209
1210                         /* Adjust copy according to the amount that was
1211                          * actually allocated. The difference is due
1212                          * to max sg elements limit
1213                          */
1214                         copy -= required_size - msg_pl->sg.size;
1215                         full_record = true;
1216                 }
1217
1218                 sk_msg_page_add(msg_pl, page, copy, offset);
1219                 sk_mem_charge(sk, copy);
1220
1221                 offset += copy;
1222                 size -= copy;
1223                 copied += copy;
1224
1225                 tls_ctx->pending_open_record_frags = true;
1226                 if (full_record || eor || sk_msg_full(msg_pl)) {
1227                         ret = bpf_exec_tx_verdict(msg_pl, sk, full_record,
1228                                                   record_type, &copied, flags);
1229                         if (ret) {
1230                                 if (ret == -EINPROGRESS)
1231                                         num_async++;
1232                                 else if (ret == -ENOMEM)
1233                                         goto wait_for_memory;
1234                                 else if (ret != -EAGAIN) {
1235                                         if (ret == -ENOSPC)
1236                                                 ret = 0;
1237                                         goto sendpage_end;
1238                                 }
1239                         }
1240                 }
1241                 continue;
1242 wait_for_sndbuf:
1243                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1244 wait_for_memory:
1245                 ret = sk_stream_wait_memory(sk, &timeo);
1246                 if (ret) {
1247                         if (ctx->open_rec)
1248                                 tls_trim_both_msgs(sk, msg_pl->sg.size);
1249                         goto sendpage_end;
1250                 }
1251
1252                 if (ctx->open_rec)
1253                         goto alloc_payload;
1254         }
1255
1256         if (num_async) {
1257                 /* Transmit if any encryptions have completed */
1258                 if (test_and_clear_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask)) {
1259                         cancel_delayed_work(&ctx->tx_work.work);
1260                         tls_tx_records(sk, flags);
1261                 }
1262         }
1263 sendpage_end:
1264         ret = sk_stream_error(sk, flags, ret);
1265         return copied > 0 ? copied : ret;
1266 }
1267
1268 int tls_sw_sendpage_locked(struct sock *sk, struct page *page,
1269                            int offset, size_t size, int flags)
1270 {
1271         if (flags & ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL |
1272                       MSG_SENDPAGE_NOTLAST | MSG_SENDPAGE_NOPOLICY |
1273                       MSG_NO_SHARED_FRAGS))
1274                 return -EOPNOTSUPP;
1275
1276         return tls_sw_do_sendpage(sk, page, offset, size, flags);
1277 }
1278
1279 int tls_sw_sendpage(struct sock *sk, struct page *page,
1280                     int offset, size_t size, int flags)
1281 {
1282         struct tls_context *tls_ctx = tls_get_ctx(sk);
1283         int ret;
1284
1285         if (flags & ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL |
1286                       MSG_SENDPAGE_NOTLAST | MSG_SENDPAGE_NOPOLICY))
1287                 return -EOPNOTSUPP;
1288
1289         ret = mutex_lock_interruptible(&tls_ctx->tx_lock);
1290         if (ret)
1291                 return ret;
1292         lock_sock(sk);
1293         ret = tls_sw_do_sendpage(sk, page, offset, size, flags);
1294         release_sock(sk);
1295         mutex_unlock(&tls_ctx->tx_lock);
1296         return ret;
1297 }
1298
1299 static struct sk_buff *tls_wait_data(struct sock *sk, struct sk_psock *psock,
1300                                      bool nonblock, long timeo, int *err)
1301 {
1302         struct tls_context *tls_ctx = tls_get_ctx(sk);
1303         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1304         struct sk_buff *skb;
1305         DEFINE_WAIT_FUNC(wait, woken_wake_function);
1306
1307         while (!(skb = ctx->recv_pkt) && sk_psock_queue_empty(psock)) {
1308                 if (sk->sk_err) {
1309                         *err = sock_error(sk);
1310                         return NULL;
1311                 }
1312
1313                 if (!skb_queue_empty(&sk->sk_receive_queue)) {
1314                         __strp_unpause(&ctx->strp);
1315                         if (ctx->recv_pkt)
1316                                 return ctx->recv_pkt;
1317                 }
1318
1319                 if (sk->sk_shutdown & RCV_SHUTDOWN)
1320                         return NULL;
1321
1322                 if (sock_flag(sk, SOCK_DONE))
1323                         return NULL;
1324
1325                 if (nonblock || !timeo) {
1326                         *err = -EAGAIN;
1327                         return NULL;
1328                 }
1329
1330                 add_wait_queue(sk_sleep(sk), &wait);
1331                 sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1332                 sk_wait_event(sk, &timeo,
1333                               ctx->recv_pkt != skb ||
1334                               !sk_psock_queue_empty(psock),
1335                               &wait);
1336                 sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1337                 remove_wait_queue(sk_sleep(sk), &wait);
1338
1339                 /* Handle signals */
1340                 if (signal_pending(current)) {
1341                         *err = sock_intr_errno(timeo);
1342                         return NULL;
1343                 }
1344         }
1345
1346         return skb;
1347 }
1348
1349 static int tls_setup_from_iter(struct sock *sk, struct iov_iter *from,
1350                                int length, int *pages_used,
1351                                unsigned int *size_used,
1352                                struct scatterlist *to,
1353                                int to_max_pages)
1354 {
1355         int rc = 0, i = 0, num_elem = *pages_used, maxpages;
1356         struct page *pages[MAX_SKB_FRAGS];
1357         unsigned int size = *size_used;
1358         ssize_t copied, use;
1359         size_t offset;
1360
1361         while (length > 0) {
1362                 i = 0;
1363                 maxpages = to_max_pages - num_elem;
1364                 if (maxpages == 0) {
1365                         rc = -EFAULT;
1366                         goto out;
1367                 }
1368                 copied = iov_iter_get_pages(from, pages,
1369                                             length,
1370                                             maxpages, &offset);
1371                 if (copied <= 0) {
1372                         rc = -EFAULT;
1373                         goto out;
1374                 }
1375
1376                 iov_iter_advance(from, copied);
1377
1378                 length -= copied;
1379                 size += copied;
1380                 while (copied) {
1381                         use = min_t(int, copied, PAGE_SIZE - offset);
1382
1383                         sg_set_page(&to[num_elem],
1384                                     pages[i], use, offset);
1385                         sg_unmark_end(&to[num_elem]);
1386                         /* We do not uncharge memory from this API */
1387
1388                         offset = 0;
1389                         copied -= use;
1390
1391                         i++;
1392                         num_elem++;
1393                 }
1394         }
1395         /* Mark the end in the last sg entry if newly added */
1396         if (num_elem > *pages_used)
1397                 sg_mark_end(&to[num_elem - 1]);
1398 out:
1399         if (rc)
1400                 iov_iter_revert(from, size - *size_used);
1401         *size_used = size;
1402         *pages_used = num_elem;
1403
1404         return rc;
1405 }
1406
1407 /* This function decrypts the input skb into either out_iov or in out_sg
1408  * or in skb buffers itself. The input parameter 'zc' indicates if
1409  * zero-copy mode needs to be tried or not. With zero-copy mode, either
1410  * out_iov or out_sg must be non-NULL. In case both out_iov and out_sg are
1411  * NULL, then the decryption happens inside skb buffers itself, i.e.
1412  * zero-copy gets disabled and 'zc' is updated.
1413  */
1414
1415 static int decrypt_internal(struct sock *sk, struct sk_buff *skb,
1416                             struct iov_iter *out_iov,
1417                             struct scatterlist *out_sg,
1418                             int *chunk, bool *zc, bool async)
1419 {
1420         struct tls_context *tls_ctx = tls_get_ctx(sk);
1421         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1422         struct tls_prot_info *prot = &tls_ctx->prot_info;
1423         struct strp_msg *rxm = strp_msg(skb);
1424         int n_sgin, n_sgout, nsg, mem_size, aead_size, err, pages = 0;
1425         struct aead_request *aead_req;
1426         struct sk_buff *unused;
1427         u8 *aad, *iv, *mem = NULL;
1428         struct scatterlist *sgin = NULL;
1429         struct scatterlist *sgout = NULL;
1430         const int data_len = rxm->full_len - prot->overhead_size +
1431                              prot->tail_size;
1432         int iv_offset = 0;
1433
1434         if (*zc && (out_iov || out_sg)) {
1435                 if (out_iov)
1436                         n_sgout = iov_iter_npages(out_iov, INT_MAX) + 1;
1437                 else
1438                         n_sgout = sg_nents(out_sg);
1439                 n_sgin = skb_nsg(skb, rxm->offset + prot->prepend_size,
1440                                  rxm->full_len - prot->prepend_size);
1441         } else {
1442                 n_sgout = 0;
1443                 *zc = false;
1444                 n_sgin = skb_cow_data(skb, 0, &unused);
1445         }
1446
1447         if (n_sgin < 1)
1448                 return -EBADMSG;
1449
1450         /* Increment to accommodate AAD */
1451         n_sgin = n_sgin + 1;
1452
1453         nsg = n_sgin + n_sgout;
1454
1455         aead_size = sizeof(*aead_req) + crypto_aead_reqsize(ctx->aead_recv);
1456         mem_size = aead_size + (nsg * sizeof(struct scatterlist));
1457         mem_size = mem_size + prot->aad_size;
1458         mem_size = mem_size + crypto_aead_ivsize(ctx->aead_recv);
1459
1460         /* Allocate a single block of memory which contains
1461          * aead_req || sgin[] || sgout[] || aad || iv.
1462          * This order achieves correct alignment for aead_req, sgin, sgout.
1463          */
1464         mem = kmalloc(mem_size, sk->sk_allocation);
1465         if (!mem)
1466                 return -ENOMEM;
1467
1468         /* Segment the allocated memory */
1469         aead_req = (struct aead_request *)mem;
1470         sgin = (struct scatterlist *)(mem + aead_size);
1471         sgout = sgin + n_sgin;
1472         aad = (u8 *)(sgout + n_sgout);
1473         iv = aad + prot->aad_size;
1474
1475         /* For CCM based ciphers, first byte of nonce+iv is always '2' */
1476         if (prot->cipher_type == TLS_CIPHER_AES_CCM_128) {
1477                 iv[0] = 2;
1478                 iv_offset = 1;
1479         }
1480
1481         /* Prepare IV */
1482         err = skb_copy_bits(skb, rxm->offset + TLS_HEADER_SIZE,
1483                             iv + iv_offset + prot->salt_size,
1484                             prot->iv_size);
1485         if (err < 0) {
1486                 kfree(mem);
1487                 return err;
1488         }
1489         if (prot->version == TLS_1_3_VERSION ||
1490             prot->cipher_type == TLS_CIPHER_CHACHA20_POLY1305)
1491                 memcpy(iv + iv_offset, tls_ctx->rx.iv,
1492                        prot->iv_size + prot->salt_size);
1493         else
1494                 memcpy(iv + iv_offset, tls_ctx->rx.iv, prot->salt_size);
1495
1496         xor_iv_with_seq(prot, iv + iv_offset, tls_ctx->rx.rec_seq);
1497
1498         /* Prepare AAD */
1499         tls_make_aad(aad, rxm->full_len - prot->overhead_size +
1500                      prot->tail_size,
1501                      tls_ctx->rx.rec_seq, ctx->control, prot);
1502
1503         /* Prepare sgin */
1504         sg_init_table(sgin, n_sgin);
1505         sg_set_buf(&sgin[0], aad, prot->aad_size);
1506         err = skb_to_sgvec(skb, &sgin[1],
1507                            rxm->offset + prot->prepend_size,
1508                            rxm->full_len - prot->prepend_size);
1509         if (err < 0) {
1510                 kfree(mem);
1511                 return err;
1512         }
1513
1514         if (n_sgout) {
1515                 if (out_iov) {
1516                         sg_init_table(sgout, n_sgout);
1517                         sg_set_buf(&sgout[0], aad, prot->aad_size);
1518
1519                         *chunk = 0;
1520                         err = tls_setup_from_iter(sk, out_iov, data_len,
1521                                                   &pages, chunk, &sgout[1],
1522                                                   (n_sgout - 1));
1523                         if (err < 0)
1524                                 goto fallback_to_reg_recv;
1525                 } else if (out_sg) {
1526                         memcpy(sgout, out_sg, n_sgout * sizeof(*sgout));
1527                 } else {
1528                         goto fallback_to_reg_recv;
1529                 }
1530         } else {
1531 fallback_to_reg_recv:
1532                 sgout = sgin;
1533                 pages = 0;
1534                 *chunk = data_len;
1535                 *zc = false;
1536         }
1537
1538         /* Prepare and submit AEAD request */
1539         err = tls_do_decryption(sk, skb, sgin, sgout, iv,
1540                                 data_len, aead_req, async);
1541         if (err == -EINPROGRESS)
1542                 return err;
1543
1544         /* Release the pages in case iov was mapped to pages */
1545         for (; pages > 0; pages--)
1546                 put_page(sg_page(&sgout[pages]));
1547
1548         kfree(mem);
1549         return err;
1550 }
1551
1552 static int decrypt_skb_update(struct sock *sk, struct sk_buff *skb,
1553                               struct iov_iter *dest, int *chunk, bool *zc,
1554                               bool async)
1555 {
1556         struct tls_context *tls_ctx = tls_get_ctx(sk);
1557         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1558         struct tls_prot_info *prot = &tls_ctx->prot_info;
1559         struct strp_msg *rxm = strp_msg(skb);
1560         int pad, err = 0;
1561
1562         if (!ctx->decrypted) {
1563                 if (tls_ctx->rx_conf == TLS_HW) {
1564                         err = tls_device_decrypted(sk, tls_ctx, skb, rxm);
1565                         if (err < 0)
1566                                 return err;
1567                 }
1568
1569                 /* Still not decrypted after tls_device */
1570                 if (!ctx->decrypted) {
1571                         err = decrypt_internal(sk, skb, dest, NULL, chunk, zc,
1572                                                async);
1573                         if (err < 0) {
1574                                 if (err == -EINPROGRESS)
1575                                         tls_advance_record_sn(sk, prot,
1576                                                               &tls_ctx->rx);
1577                                 else if (err == -EBADMSG)
1578                                         TLS_INC_STATS(sock_net(sk),
1579                                                       LINUX_MIB_TLSDECRYPTERROR);
1580                                 return err;
1581                         }
1582                 } else {
1583                         *zc = false;
1584                 }
1585
1586                 pad = padding_length(ctx, prot, skb);
1587                 if (pad < 0)
1588                         return pad;
1589
1590                 rxm->full_len -= pad;
1591                 rxm->offset += prot->prepend_size;
1592                 rxm->full_len -= prot->overhead_size;
1593                 tls_advance_record_sn(sk, prot, &tls_ctx->rx);
1594                 ctx->decrypted = 1;
1595                 ctx->saved_data_ready(sk);
1596         } else {
1597                 *zc = false;
1598         }
1599
1600         return err;
1601 }
1602
1603 int decrypt_skb(struct sock *sk, struct sk_buff *skb,
1604                 struct scatterlist *sgout)
1605 {
1606         bool zc = true;
1607         int chunk;
1608
1609         return decrypt_internal(sk, skb, NULL, sgout, &chunk, &zc, false);
1610 }
1611
1612 static bool tls_sw_advance_skb(struct sock *sk, struct sk_buff *skb,
1613                                unsigned int len)
1614 {
1615         struct tls_context *tls_ctx = tls_get_ctx(sk);
1616         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1617
1618         if (skb) {
1619                 struct strp_msg *rxm = strp_msg(skb);
1620
1621                 if (len < rxm->full_len) {
1622                         rxm->offset += len;
1623                         rxm->full_len -= len;
1624                         return false;
1625                 }
1626                 consume_skb(skb);
1627         }
1628
1629         /* Finished with message */
1630         ctx->recv_pkt = NULL;
1631         __strp_unpause(&ctx->strp);
1632
1633         return true;
1634 }
1635
1636 /* This function traverses the rx_list in tls receive context to copies the
1637  * decrypted records into the buffer provided by caller zero copy is not
1638  * true. Further, the records are removed from the rx_list if it is not a peek
1639  * case and the record has been consumed completely.
1640  */
1641 static int process_rx_list(struct tls_sw_context_rx *ctx,
1642                            struct msghdr *msg,
1643                            u8 *control,
1644                            bool *cmsg,
1645                            size_t skip,
1646                            size_t len,
1647                            bool zc,
1648                            bool is_peek)
1649 {
1650         struct sk_buff *skb = skb_peek(&ctx->rx_list);
1651         u8 ctrl = *control;
1652         u8 msgc = *cmsg;
1653         struct tls_msg *tlm;
1654         ssize_t copied = 0;
1655
1656         /* Set the record type in 'control' if caller didn't pass it */
1657         if (!ctrl && skb) {
1658                 tlm = tls_msg(skb);
1659                 ctrl = tlm->control;
1660         }
1661
1662         while (skip && skb) {
1663                 struct strp_msg *rxm = strp_msg(skb);
1664                 tlm = tls_msg(skb);
1665
1666                 /* Cannot process a record of different type */
1667                 if (ctrl != tlm->control)
1668                         return 0;
1669
1670                 if (skip < rxm->full_len)
1671                         break;
1672
1673                 skip = skip - rxm->full_len;
1674                 skb = skb_peek_next(skb, &ctx->rx_list);
1675         }
1676
1677         while (len && skb) {
1678                 struct sk_buff *next_skb;
1679                 struct strp_msg *rxm = strp_msg(skb);
1680                 int chunk = min_t(unsigned int, rxm->full_len - skip, len);
1681
1682                 tlm = tls_msg(skb);
1683
1684                 /* Cannot process a record of different type */
1685                 if (ctrl != tlm->control)
1686                         return 0;
1687
1688                 /* Set record type if not already done. For a non-data record,
1689                  * do not proceed if record type could not be copied.
1690                  */
1691                 if (!msgc) {
1692                         int cerr = put_cmsg(msg, SOL_TLS, TLS_GET_RECORD_TYPE,
1693                                             sizeof(ctrl), &ctrl);
1694                         msgc = true;
1695                         if (ctrl != TLS_RECORD_TYPE_DATA) {
1696                                 if (cerr || msg->msg_flags & MSG_CTRUNC)
1697                                         return -EIO;
1698
1699                                 *cmsg = msgc;
1700                         }
1701                 }
1702
1703                 if (!zc || (rxm->full_len - skip) > len) {
1704                         int err = skb_copy_datagram_msg(skb, rxm->offset + skip,
1705                                                     msg, chunk);
1706                         if (err < 0)
1707                                 return err;
1708                 }
1709
1710                 len = len - chunk;
1711                 copied = copied + chunk;
1712
1713                 /* Consume the data from record if it is non-peek case*/
1714                 if (!is_peek) {
1715                         rxm->offset = rxm->offset + chunk;
1716                         rxm->full_len = rxm->full_len - chunk;
1717
1718                         /* Return if there is unconsumed data in the record */
1719                         if (rxm->full_len - skip)
1720                                 break;
1721                 }
1722
1723                 /* The remaining skip-bytes must lie in 1st record in rx_list.
1724                  * So from the 2nd record, 'skip' should be 0.
1725                  */
1726                 skip = 0;
1727
1728                 if (msg)
1729                         msg->msg_flags |= MSG_EOR;
1730
1731                 next_skb = skb_peek_next(skb, &ctx->rx_list);
1732
1733                 if (!is_peek) {
1734                         skb_unlink(skb, &ctx->rx_list);
1735                         consume_skb(skb);
1736                 }
1737
1738                 skb = next_skb;
1739         }
1740
1741         *control = ctrl;
1742         return copied;
1743 }
1744
1745 int tls_sw_recvmsg(struct sock *sk,
1746                    struct msghdr *msg,
1747                    size_t len,
1748                    int nonblock,
1749                    int flags,
1750                    int *addr_len)
1751 {
1752         struct tls_context *tls_ctx = tls_get_ctx(sk);
1753         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1754         struct tls_prot_info *prot = &tls_ctx->prot_info;
1755         struct sk_psock *psock;
1756         unsigned char control = 0;
1757         ssize_t decrypted = 0;
1758         struct strp_msg *rxm;
1759         struct tls_msg *tlm;
1760         struct sk_buff *skb;
1761         ssize_t copied = 0;
1762         bool cmsg = false;
1763         int target, err = 0;
1764         long timeo;
1765         bool is_kvec = iov_iter_is_kvec(&msg->msg_iter);
1766         bool is_peek = flags & MSG_PEEK;
1767         bool bpf_strp_enabled;
1768         int num_async = 0;
1769         int pending;
1770
1771         flags |= nonblock;
1772
1773         if (unlikely(flags & MSG_ERRQUEUE))
1774                 return sock_recv_errqueue(sk, msg, len, SOL_IP, IP_RECVERR);
1775
1776         psock = sk_psock_get(sk);
1777         lock_sock(sk);
1778         bpf_strp_enabled = sk_psock_strp_enabled(psock);
1779
1780         /* Process pending decrypted records. It must be non-zero-copy */
1781         err = process_rx_list(ctx, msg, &control, &cmsg, 0, len, false,
1782                               is_peek);
1783         if (err < 0) {
1784                 tls_err_abort(sk, err);
1785                 goto end;
1786         } else {
1787                 copied = err;
1788         }
1789
1790         if (len <= copied)
1791                 goto recv_end;
1792
1793         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1794         len = len - copied;
1795         timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
1796
1797         while (len && (decrypted + copied < target || ctx->recv_pkt)) {
1798                 bool retain_skb = false;
1799                 bool zc = false;
1800                 int to_decrypt;
1801                 int chunk = 0;
1802                 bool async_capable;
1803                 bool async = false;
1804
1805                 skb = tls_wait_data(sk, psock, flags & MSG_DONTWAIT, timeo, &err);
1806                 if (!skb) {
1807                         if (psock) {
1808                                 int ret = sk_msg_recvmsg(sk, psock, msg, len,
1809                                                          flags);
1810
1811                                 if (ret > 0) {
1812                                         decrypted += ret;
1813                                         len -= ret;
1814                                         continue;
1815                                 }
1816                         }
1817                         goto recv_end;
1818                 } else {
1819                         tlm = tls_msg(skb);
1820                         if (prot->version == TLS_1_3_VERSION)
1821                                 tlm->control = 0;
1822                         else
1823                                 tlm->control = ctx->control;
1824                 }
1825
1826                 rxm = strp_msg(skb);
1827
1828                 to_decrypt = rxm->full_len - prot->overhead_size;
1829
1830                 if (to_decrypt <= len && !is_kvec && !is_peek &&
1831                     ctx->control == TLS_RECORD_TYPE_DATA &&
1832                     prot->version != TLS_1_3_VERSION &&
1833                     !bpf_strp_enabled)
1834                         zc = true;
1835
1836                 /* Do not use async mode if record is non-data */
1837                 if (ctx->control == TLS_RECORD_TYPE_DATA && !bpf_strp_enabled)
1838                         async_capable = ctx->async_capable;
1839                 else
1840                         async_capable = false;
1841
1842                 err = decrypt_skb_update(sk, skb, &msg->msg_iter,
1843                                          &chunk, &zc, async_capable);
1844                 if (err < 0 && err != -EINPROGRESS) {
1845                         tls_err_abort(sk, -EBADMSG);
1846                         goto recv_end;
1847                 }
1848
1849                 if (err == -EINPROGRESS) {
1850                         async = true;
1851                         num_async++;
1852                 } else if (prot->version == TLS_1_3_VERSION) {
1853                         tlm->control = ctx->control;
1854                 }
1855
1856                 /* If the type of records being processed is not known yet,
1857                  * set it to record type just dequeued. If it is already known,
1858                  * but does not match the record type just dequeued, go to end.
1859                  * We always get record type here since for tls1.2, record type
1860                  * is known just after record is dequeued from stream parser.
1861                  * For tls1.3, we disable async.
1862                  */
1863
1864                 if (!control)
1865                         control = tlm->control;
1866                 else if (control != tlm->control)
1867                         goto recv_end;
1868
1869                 if (!cmsg) {
1870                         int cerr;
1871
1872                         cerr = put_cmsg(msg, SOL_TLS, TLS_GET_RECORD_TYPE,
1873                                         sizeof(control), &control);
1874                         cmsg = true;
1875                         if (control != TLS_RECORD_TYPE_DATA) {
1876                                 if (cerr || msg->msg_flags & MSG_CTRUNC) {
1877                                         err = -EIO;
1878                                         goto recv_end;
1879                                 }
1880                         }
1881                 }
1882
1883                 if (async)
1884                         goto pick_next_record;
1885
1886                 if (!zc) {
1887                         if (bpf_strp_enabled) {
1888                                 err = sk_psock_tls_strp_read(psock, skb);
1889                                 if (err != __SK_PASS) {
1890                                         rxm->offset = rxm->offset + rxm->full_len;
1891                                         rxm->full_len = 0;
1892                                         if (err == __SK_DROP)
1893                                                 consume_skb(skb);
1894                                         ctx->recv_pkt = NULL;
1895                                         __strp_unpause(&ctx->strp);
1896                                         continue;
1897                                 }
1898                         }
1899
1900                         if (rxm->full_len > len) {
1901                                 retain_skb = true;
1902                                 chunk = len;
1903                         } else {
1904                                 chunk = rxm->full_len;
1905                         }
1906
1907                         err = skb_copy_datagram_msg(skb, rxm->offset,
1908                                                     msg, chunk);
1909                         if (err < 0)
1910                                 goto recv_end;
1911
1912                         if (!is_peek) {
1913                                 rxm->offset = rxm->offset + chunk;
1914                                 rxm->full_len = rxm->full_len - chunk;
1915                         }
1916                 }
1917
1918 pick_next_record:
1919                 if (chunk > len)
1920                         chunk = len;
1921
1922                 decrypted += chunk;
1923                 len -= chunk;
1924
1925                 /* For async or peek case, queue the current skb */
1926                 if (async || is_peek || retain_skb) {
1927                         skb_queue_tail(&ctx->rx_list, skb);
1928                         skb = NULL;
1929                 }
1930
1931                 if (tls_sw_advance_skb(sk, skb, chunk)) {
1932                         /* Return full control message to
1933                          * userspace before trying to parse
1934                          * another message type
1935                          */
1936                         msg->msg_flags |= MSG_EOR;
1937                         if (control != TLS_RECORD_TYPE_DATA)
1938                                 goto recv_end;
1939                 } else {
1940                         break;
1941                 }
1942         }
1943
1944 recv_end:
1945         if (num_async) {
1946                 /* Wait for all previously submitted records to be decrypted */
1947                 spin_lock_bh(&ctx->decrypt_compl_lock);
1948                 ctx->async_notify = true;
1949                 pending = atomic_read(&ctx->decrypt_pending);
1950                 spin_unlock_bh(&ctx->decrypt_compl_lock);
1951                 if (pending) {
1952                         err = crypto_wait_req(-EINPROGRESS, &ctx->async_wait);
1953                         if (err) {
1954                                 /* one of async decrypt failed */
1955                                 tls_err_abort(sk, err);
1956                                 copied = 0;
1957                                 decrypted = 0;
1958                                 goto end;
1959                         }
1960                 } else {
1961                         reinit_completion(&ctx->async_wait.completion);
1962                 }
1963
1964                 /* There can be no concurrent accesses, since we have no
1965                  * pending decrypt operations
1966                  */
1967                 WRITE_ONCE(ctx->async_notify, false);
1968
1969                 /* Drain records from the rx_list & copy if required */
1970                 if (is_peek || is_kvec)
1971                         err = process_rx_list(ctx, msg, &control, &cmsg, copied,
1972                                               decrypted, false, is_peek);
1973                 else
1974                         err = process_rx_list(ctx, msg, &control, &cmsg, 0,
1975                                               decrypted, true, is_peek);
1976                 if (err < 0) {
1977                         tls_err_abort(sk, err);
1978                         copied = 0;
1979                         goto end;
1980                 }
1981         }
1982
1983         copied += decrypted;
1984
1985 end:
1986         release_sock(sk);
1987         if (psock)
1988                 sk_psock_put(sk, psock);
1989         return copied ? : err;
1990 }
1991
1992 ssize_t tls_sw_splice_read(struct socket *sock,  loff_t *ppos,
1993                            struct pipe_inode_info *pipe,
1994                            size_t len, unsigned int flags)
1995 {
1996         struct tls_context *tls_ctx = tls_get_ctx(sock->sk);
1997         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1998         struct strp_msg *rxm = NULL;
1999         struct sock *sk = sock->sk;
2000         struct sk_buff *skb;
2001         ssize_t copied = 0;
2002         bool from_queue;
2003         int err = 0;
2004         long timeo;
2005         int chunk;
2006         bool zc = false;
2007
2008         lock_sock(sk);
2009
2010         timeo = sock_rcvtimeo(sk, flags & SPLICE_F_NONBLOCK);
2011
2012         from_queue = !skb_queue_empty(&ctx->rx_list);
2013         if (from_queue) {
2014                 skb = __skb_dequeue(&ctx->rx_list);
2015         } else {
2016                 skb = tls_wait_data(sk, NULL, flags & SPLICE_F_NONBLOCK, timeo,
2017                                     &err);
2018                 if (!skb)
2019                         goto splice_read_end;
2020
2021                 err = decrypt_skb_update(sk, skb, NULL, &chunk, &zc, false);
2022                 if (err < 0) {
2023                         tls_err_abort(sk, -EBADMSG);
2024                         goto splice_read_end;
2025                 }
2026         }
2027
2028         /* splice does not support reading control messages */
2029         if (ctx->control != TLS_RECORD_TYPE_DATA) {
2030                 err = -EINVAL;
2031                 goto splice_read_end;
2032         }
2033
2034         rxm = strp_msg(skb);
2035
2036         chunk = min_t(unsigned int, rxm->full_len, len);
2037         copied = skb_splice_bits(skb, sk, rxm->offset, pipe, chunk, flags);
2038         if (copied < 0)
2039                 goto splice_read_end;
2040
2041         if (!from_queue) {
2042                 ctx->recv_pkt = NULL;
2043                 __strp_unpause(&ctx->strp);
2044         }
2045         if (chunk < rxm->full_len) {
2046                 __skb_queue_head(&ctx->rx_list, skb);
2047                 rxm->offset += len;
2048                 rxm->full_len -= len;
2049         } else {
2050                 consume_skb(skb);
2051         }
2052
2053 splice_read_end:
2054         release_sock(sk);
2055         return copied ? : err;
2056 }
2057
2058 bool tls_sw_sock_is_readable(struct sock *sk)
2059 {
2060         struct tls_context *tls_ctx = tls_get_ctx(sk);
2061         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2062         bool ingress_empty = true;
2063         struct sk_psock *psock;
2064
2065         rcu_read_lock();
2066         psock = sk_psock(sk);
2067         if (psock)
2068                 ingress_empty = list_empty(&psock->ingress_msg);
2069         rcu_read_unlock();
2070
2071         return !ingress_empty || ctx->recv_pkt ||
2072                 !skb_queue_empty(&ctx->rx_list);
2073 }
2074
2075 static int tls_read_size(struct strparser *strp, struct sk_buff *skb)
2076 {
2077         struct tls_context *tls_ctx = tls_get_ctx(strp->sk);
2078         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2079         struct tls_prot_info *prot = &tls_ctx->prot_info;
2080         char header[TLS_HEADER_SIZE + MAX_IV_SIZE];
2081         struct strp_msg *rxm = strp_msg(skb);
2082         size_t cipher_overhead;
2083         size_t data_len = 0;
2084         int ret;
2085
2086         /* Verify that we have a full TLS header, or wait for more data */
2087         if (rxm->offset + prot->prepend_size > skb->len)
2088                 return 0;
2089
2090         /* Sanity-check size of on-stack buffer. */
2091         if (WARN_ON(prot->prepend_size > sizeof(header))) {
2092                 ret = -EINVAL;
2093                 goto read_failure;
2094         }
2095
2096         /* Linearize header to local buffer */
2097         ret = skb_copy_bits(skb, rxm->offset, header, prot->prepend_size);
2098
2099         if (ret < 0)
2100                 goto read_failure;
2101
2102         ctx->control = header[0];
2103
2104         data_len = ((header[4] & 0xFF) | (header[3] << 8));
2105
2106         cipher_overhead = prot->tag_size;
2107         if (prot->version != TLS_1_3_VERSION &&
2108             prot->cipher_type != TLS_CIPHER_CHACHA20_POLY1305)
2109                 cipher_overhead += prot->iv_size;
2110
2111         if (data_len > TLS_MAX_PAYLOAD_SIZE + cipher_overhead +
2112             prot->tail_size) {
2113                 ret = -EMSGSIZE;
2114                 goto read_failure;
2115         }
2116         if (data_len < cipher_overhead) {
2117                 ret = -EBADMSG;
2118                 goto read_failure;
2119         }
2120
2121         /* Note that both TLS1.3 and TLS1.2 use TLS_1_2 version here */
2122         if (header[1] != TLS_1_2_VERSION_MINOR ||
2123             header[2] != TLS_1_2_VERSION_MAJOR) {
2124                 ret = -EINVAL;
2125                 goto read_failure;
2126         }
2127
2128         tls_device_rx_resync_new_rec(strp->sk, data_len + TLS_HEADER_SIZE,
2129                                      TCP_SKB_CB(skb)->seq + rxm->offset);
2130         return data_len + TLS_HEADER_SIZE;
2131
2132 read_failure:
2133         tls_err_abort(strp->sk, ret);
2134
2135         return ret;
2136 }
2137
2138 static void tls_queue(struct strparser *strp, struct sk_buff *skb)
2139 {
2140         struct tls_context *tls_ctx = tls_get_ctx(strp->sk);
2141         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2142
2143         ctx->decrypted = 0;
2144
2145         ctx->recv_pkt = skb;
2146         strp_pause(strp);
2147
2148         ctx->saved_data_ready(strp->sk);
2149 }
2150
2151 static void tls_data_ready(struct sock *sk)
2152 {
2153         struct tls_context *tls_ctx = tls_get_ctx(sk);
2154         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2155         struct sk_psock *psock;
2156
2157         strp_data_ready(&ctx->strp);
2158
2159         psock = sk_psock_get(sk);
2160         if (psock) {
2161                 if (!list_empty(&psock->ingress_msg))
2162                         ctx->saved_data_ready(sk);
2163                 sk_psock_put(sk, psock);
2164         }
2165 }
2166
2167 void tls_sw_cancel_work_tx(struct tls_context *tls_ctx)
2168 {
2169         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
2170
2171         set_bit(BIT_TX_CLOSING, &ctx->tx_bitmask);
2172         set_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask);
2173         cancel_delayed_work_sync(&ctx->tx_work.work);
2174 }
2175
2176 void tls_sw_release_resources_tx(struct sock *sk)
2177 {
2178         struct tls_context *tls_ctx = tls_get_ctx(sk);
2179         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
2180         struct tls_rec *rec, *tmp;
2181         int pending;
2182
2183         /* Wait for any pending async encryptions to complete */
2184         spin_lock_bh(&ctx->encrypt_compl_lock);
2185         ctx->async_notify = true;
2186         pending = atomic_read(&ctx->encrypt_pending);
2187         spin_unlock_bh(&ctx->encrypt_compl_lock);
2188
2189         if (pending)
2190                 crypto_wait_req(-EINPROGRESS, &ctx->async_wait);
2191
2192         tls_tx_records(sk, -1);
2193
2194         /* Free up un-sent records in tx_list. First, free
2195          * the partially sent record if any at head of tx_list.
2196          */
2197         if (tls_ctx->partially_sent_record) {
2198                 tls_free_partial_record(sk, tls_ctx);
2199                 rec = list_first_entry(&ctx->tx_list,
2200                                        struct tls_rec, list);
2201                 list_del(&rec->list);
2202                 sk_msg_free(sk, &rec->msg_plaintext);
2203                 kfree(rec);
2204         }
2205
2206         list_for_each_entry_safe(rec, tmp, &ctx->tx_list, list) {
2207                 list_del(&rec->list);
2208                 sk_msg_free(sk, &rec->msg_encrypted);
2209                 sk_msg_free(sk, &rec->msg_plaintext);
2210                 kfree(rec);
2211         }
2212
2213         crypto_free_aead(ctx->aead_send);
2214         tls_free_open_rec(sk);
2215 }
2216
2217 void tls_sw_free_ctx_tx(struct tls_context *tls_ctx)
2218 {
2219         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
2220
2221         kfree(ctx);
2222 }
2223
2224 void tls_sw_release_resources_rx(struct sock *sk)
2225 {
2226         struct tls_context *tls_ctx = tls_get_ctx(sk);
2227         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2228
2229         kfree(tls_ctx->rx.rec_seq);
2230         kfree(tls_ctx->rx.iv);
2231
2232         if (ctx->aead_recv) {
2233                 kfree_skb(ctx->recv_pkt);
2234                 ctx->recv_pkt = NULL;
2235                 skb_queue_purge(&ctx->rx_list);
2236                 crypto_free_aead(ctx->aead_recv);
2237                 strp_stop(&ctx->strp);
2238                 /* If tls_sw_strparser_arm() was not called (cleanup paths)
2239                  * we still want to strp_stop(), but sk->sk_data_ready was
2240                  * never swapped.
2241                  */
2242                 if (ctx->saved_data_ready) {
2243                         write_lock_bh(&sk->sk_callback_lock);
2244                         sk->sk_data_ready = ctx->saved_data_ready;
2245                         write_unlock_bh(&sk->sk_callback_lock);
2246                 }
2247         }
2248 }
2249
2250 void tls_sw_strparser_done(struct tls_context *tls_ctx)
2251 {
2252         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2253
2254         strp_done(&ctx->strp);
2255 }
2256
2257 void tls_sw_free_ctx_rx(struct tls_context *tls_ctx)
2258 {
2259         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2260
2261         kfree(ctx);
2262 }
2263
2264 void tls_sw_free_resources_rx(struct sock *sk)
2265 {
2266         struct tls_context *tls_ctx = tls_get_ctx(sk);
2267
2268         tls_sw_release_resources_rx(sk);
2269         tls_sw_free_ctx_rx(tls_ctx);
2270 }
2271
2272 /* The work handler to transmitt the encrypted records in tx_list */
2273 static void tx_work_handler(struct work_struct *work)
2274 {
2275         struct delayed_work *delayed_work = to_delayed_work(work);
2276         struct tx_work *tx_work = container_of(delayed_work,
2277                                                struct tx_work, work);
2278         struct sock *sk = tx_work->sk;
2279         struct tls_context *tls_ctx = tls_get_ctx(sk);
2280         struct tls_sw_context_tx *ctx;
2281
2282         if (unlikely(!tls_ctx))
2283                 return;
2284
2285         ctx = tls_sw_ctx_tx(tls_ctx);
2286         if (test_bit(BIT_TX_CLOSING, &ctx->tx_bitmask))
2287                 return;
2288
2289         if (!test_and_clear_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask))
2290                 return;
2291
2292         if (mutex_trylock(&tls_ctx->tx_lock)) {
2293                 lock_sock(sk);
2294                 tls_tx_records(sk, -1);
2295                 release_sock(sk);
2296                 mutex_unlock(&tls_ctx->tx_lock);
2297         } else if (!test_and_set_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask)) {
2298                 /* Someone is holding the tx_lock, they will likely run Tx
2299                  * and cancel the work on their way out of the lock section.
2300                  * Schedule a long delay just in case.
2301                  */
2302                 schedule_delayed_work(&ctx->tx_work.work, msecs_to_jiffies(10));
2303         }
2304 }
2305
2306 void tls_sw_write_space(struct sock *sk, struct tls_context *ctx)
2307 {
2308         struct tls_sw_context_tx *tx_ctx = tls_sw_ctx_tx(ctx);
2309
2310         /* Schedule the transmission if tx list is ready */
2311         if (is_tx_ready(tx_ctx) &&
2312             !test_and_set_bit(BIT_TX_SCHEDULED, &tx_ctx->tx_bitmask))
2313                 schedule_delayed_work(&tx_ctx->tx_work.work, 0);
2314 }
2315
2316 void tls_sw_strparser_arm(struct sock *sk, struct tls_context *tls_ctx)
2317 {
2318         struct tls_sw_context_rx *rx_ctx = tls_sw_ctx_rx(tls_ctx);
2319
2320         write_lock_bh(&sk->sk_callback_lock);
2321         rx_ctx->saved_data_ready = sk->sk_data_ready;
2322         sk->sk_data_ready = tls_data_ready;
2323         write_unlock_bh(&sk->sk_callback_lock);
2324
2325         strp_check_rcv(&rx_ctx->strp);
2326 }
2327
2328 int tls_set_sw_offload(struct sock *sk, struct tls_context *ctx, int tx)
2329 {
2330         struct tls_context *tls_ctx = tls_get_ctx(sk);
2331         struct tls_prot_info *prot = &tls_ctx->prot_info;
2332         struct tls_crypto_info *crypto_info;
2333         struct tls12_crypto_info_aes_gcm_128 *gcm_128_info;
2334         struct tls12_crypto_info_aes_gcm_256 *gcm_256_info;
2335         struct tls12_crypto_info_aes_ccm_128 *ccm_128_info;
2336         struct tls12_crypto_info_chacha20_poly1305 *chacha20_poly1305_info;
2337         struct tls_sw_context_tx *sw_ctx_tx = NULL;
2338         struct tls_sw_context_rx *sw_ctx_rx = NULL;
2339         struct cipher_context *cctx;
2340         struct crypto_aead **aead;
2341         struct strp_callbacks cb;
2342         u16 nonce_size, tag_size, iv_size, rec_seq_size, salt_size;
2343         struct crypto_tfm *tfm;
2344         char *iv, *rec_seq, *key, *salt, *cipher_name;
2345         size_t keysize;
2346         int rc = 0;
2347
2348         if (!ctx) {
2349                 rc = -EINVAL;
2350                 goto out;
2351         }
2352
2353         if (tx) {
2354                 if (!ctx->priv_ctx_tx) {
2355                         sw_ctx_tx = kzalloc(sizeof(*sw_ctx_tx), GFP_KERNEL);
2356                         if (!sw_ctx_tx) {
2357                                 rc = -ENOMEM;
2358                                 goto out;
2359                         }
2360                         ctx->priv_ctx_tx = sw_ctx_tx;
2361                 } else {
2362                         sw_ctx_tx =
2363                                 (struct tls_sw_context_tx *)ctx->priv_ctx_tx;
2364                 }
2365         } else {
2366                 if (!ctx->priv_ctx_rx) {
2367                         sw_ctx_rx = kzalloc(sizeof(*sw_ctx_rx), GFP_KERNEL);
2368                         if (!sw_ctx_rx) {
2369                                 rc = -ENOMEM;
2370                                 goto out;
2371                         }
2372                         ctx->priv_ctx_rx = sw_ctx_rx;
2373                 } else {
2374                         sw_ctx_rx =
2375                                 (struct tls_sw_context_rx *)ctx->priv_ctx_rx;
2376                 }
2377         }
2378
2379         if (tx) {
2380                 crypto_init_wait(&sw_ctx_tx->async_wait);
2381                 spin_lock_init(&sw_ctx_tx->encrypt_compl_lock);
2382                 crypto_info = &ctx->crypto_send.info;
2383                 cctx = &ctx->tx;
2384                 aead = &sw_ctx_tx->aead_send;
2385                 INIT_LIST_HEAD(&sw_ctx_tx->tx_list);
2386                 INIT_DELAYED_WORK(&sw_ctx_tx->tx_work.work, tx_work_handler);
2387                 sw_ctx_tx->tx_work.sk = sk;
2388         } else {
2389                 crypto_init_wait(&sw_ctx_rx->async_wait);
2390                 spin_lock_init(&sw_ctx_rx->decrypt_compl_lock);
2391                 crypto_info = &ctx->crypto_recv.info;
2392                 cctx = &ctx->rx;
2393                 skb_queue_head_init(&sw_ctx_rx->rx_list);
2394                 aead = &sw_ctx_rx->aead_recv;
2395         }
2396
2397         switch (crypto_info->cipher_type) {
2398         case TLS_CIPHER_AES_GCM_128: {
2399                 nonce_size = TLS_CIPHER_AES_GCM_128_IV_SIZE;
2400                 tag_size = TLS_CIPHER_AES_GCM_128_TAG_SIZE;
2401                 iv_size = TLS_CIPHER_AES_GCM_128_IV_SIZE;
2402                 iv = ((struct tls12_crypto_info_aes_gcm_128 *)crypto_info)->iv;
2403                 rec_seq_size = TLS_CIPHER_AES_GCM_128_REC_SEQ_SIZE;
2404                 rec_seq =
2405                  ((struct tls12_crypto_info_aes_gcm_128 *)crypto_info)->rec_seq;
2406                 gcm_128_info =
2407                         (struct tls12_crypto_info_aes_gcm_128 *)crypto_info;
2408                 keysize = TLS_CIPHER_AES_GCM_128_KEY_SIZE;
2409                 key = gcm_128_info->key;
2410                 salt = gcm_128_info->salt;
2411                 salt_size = TLS_CIPHER_AES_GCM_128_SALT_SIZE;
2412                 cipher_name = "gcm(aes)";
2413                 break;
2414         }
2415         case TLS_CIPHER_AES_GCM_256: {
2416                 nonce_size = TLS_CIPHER_AES_GCM_256_IV_SIZE;
2417                 tag_size = TLS_CIPHER_AES_GCM_256_TAG_SIZE;
2418                 iv_size = TLS_CIPHER_AES_GCM_256_IV_SIZE;
2419                 iv = ((struct tls12_crypto_info_aes_gcm_256 *)crypto_info)->iv;
2420                 rec_seq_size = TLS_CIPHER_AES_GCM_256_REC_SEQ_SIZE;
2421                 rec_seq =
2422                  ((struct tls12_crypto_info_aes_gcm_256 *)crypto_info)->rec_seq;
2423                 gcm_256_info =
2424                         (struct tls12_crypto_info_aes_gcm_256 *)crypto_info;
2425                 keysize = TLS_CIPHER_AES_GCM_256_KEY_SIZE;
2426                 key = gcm_256_info->key;
2427                 salt = gcm_256_info->salt;
2428                 salt_size = TLS_CIPHER_AES_GCM_256_SALT_SIZE;
2429                 cipher_name = "gcm(aes)";
2430                 break;
2431         }
2432         case TLS_CIPHER_AES_CCM_128: {
2433                 nonce_size = TLS_CIPHER_AES_CCM_128_IV_SIZE;
2434                 tag_size = TLS_CIPHER_AES_CCM_128_TAG_SIZE;
2435                 iv_size = TLS_CIPHER_AES_CCM_128_IV_SIZE;
2436                 iv = ((struct tls12_crypto_info_aes_ccm_128 *)crypto_info)->iv;
2437                 rec_seq_size = TLS_CIPHER_AES_CCM_128_REC_SEQ_SIZE;
2438                 rec_seq =
2439                 ((struct tls12_crypto_info_aes_ccm_128 *)crypto_info)->rec_seq;
2440                 ccm_128_info =
2441                 (struct tls12_crypto_info_aes_ccm_128 *)crypto_info;
2442                 keysize = TLS_CIPHER_AES_CCM_128_KEY_SIZE;
2443                 key = ccm_128_info->key;
2444                 salt = ccm_128_info->salt;
2445                 salt_size = TLS_CIPHER_AES_CCM_128_SALT_SIZE;
2446                 cipher_name = "ccm(aes)";
2447                 break;
2448         }
2449         case TLS_CIPHER_CHACHA20_POLY1305: {
2450                 chacha20_poly1305_info = (void *)crypto_info;
2451                 nonce_size = 0;
2452                 tag_size = TLS_CIPHER_CHACHA20_POLY1305_TAG_SIZE;
2453                 iv_size = TLS_CIPHER_CHACHA20_POLY1305_IV_SIZE;
2454                 iv = chacha20_poly1305_info->iv;
2455                 rec_seq_size = TLS_CIPHER_CHACHA20_POLY1305_REC_SEQ_SIZE;
2456                 rec_seq = chacha20_poly1305_info->rec_seq;
2457                 keysize = TLS_CIPHER_CHACHA20_POLY1305_KEY_SIZE;
2458                 key = chacha20_poly1305_info->key;
2459                 salt = chacha20_poly1305_info->salt;
2460                 salt_size = TLS_CIPHER_CHACHA20_POLY1305_SALT_SIZE;
2461                 cipher_name = "rfc7539(chacha20,poly1305)";
2462                 break;
2463         }
2464         default:
2465                 rc = -EINVAL;
2466                 goto free_priv;
2467         }
2468
2469         /* Sanity-check the sizes for stack allocations. */
2470         if (iv_size > MAX_IV_SIZE || nonce_size > MAX_IV_SIZE ||
2471             rec_seq_size > TLS_MAX_REC_SEQ_SIZE) {
2472                 rc = -EINVAL;
2473                 goto free_priv;
2474         }
2475
2476         if (crypto_info->version == TLS_1_3_VERSION) {
2477                 nonce_size = 0;
2478                 prot->aad_size = TLS_HEADER_SIZE;
2479                 prot->tail_size = 1;
2480         } else {
2481                 prot->aad_size = TLS_AAD_SPACE_SIZE;
2482                 prot->tail_size = 0;
2483         }
2484
2485         prot->version = crypto_info->version;
2486         prot->cipher_type = crypto_info->cipher_type;
2487         prot->prepend_size = TLS_HEADER_SIZE + nonce_size;
2488         prot->tag_size = tag_size;
2489         prot->overhead_size = prot->prepend_size +
2490                               prot->tag_size + prot->tail_size;
2491         prot->iv_size = iv_size;
2492         prot->salt_size = salt_size;
2493         cctx->iv = kmalloc(iv_size + salt_size, GFP_KERNEL);
2494         if (!cctx->iv) {
2495                 rc = -ENOMEM;
2496                 goto free_priv;
2497         }
2498         /* Note: 128 & 256 bit salt are the same size */
2499         prot->rec_seq_size = rec_seq_size;
2500         memcpy(cctx->iv, salt, salt_size);
2501         memcpy(cctx->iv + salt_size, iv, iv_size);
2502         cctx->rec_seq = kmemdup(rec_seq, rec_seq_size, GFP_KERNEL);
2503         if (!cctx->rec_seq) {
2504                 rc = -ENOMEM;
2505                 goto free_iv;
2506         }
2507
2508         if (!*aead) {
2509                 *aead = crypto_alloc_aead(cipher_name, 0, 0);
2510                 if (IS_ERR(*aead)) {
2511                         rc = PTR_ERR(*aead);
2512                         *aead = NULL;
2513                         goto free_rec_seq;
2514                 }
2515         }
2516
2517         ctx->push_pending_record = tls_sw_push_pending_record;
2518
2519         rc = crypto_aead_setkey(*aead, key, keysize);
2520
2521         if (rc)
2522                 goto free_aead;
2523
2524         rc = crypto_aead_setauthsize(*aead, prot->tag_size);
2525         if (rc)
2526                 goto free_aead;
2527
2528         if (sw_ctx_rx) {
2529                 tfm = crypto_aead_tfm(sw_ctx_rx->aead_recv);
2530
2531                 if (crypto_info->version == TLS_1_3_VERSION)
2532                         sw_ctx_rx->async_capable = 0;
2533                 else
2534                         sw_ctx_rx->async_capable =
2535                                 !!(tfm->__crt_alg->cra_flags &
2536                                    CRYPTO_ALG_ASYNC);
2537
2538                 /* Set up strparser */
2539                 memset(&cb, 0, sizeof(cb));
2540                 cb.rcv_msg = tls_queue;
2541                 cb.parse_msg = tls_read_size;
2542
2543                 strp_init(&sw_ctx_rx->strp, sk, &cb);
2544         }
2545
2546         goto out;
2547
2548 free_aead:
2549         crypto_free_aead(*aead);
2550         *aead = NULL;
2551 free_rec_seq:
2552         kfree(cctx->rec_seq);
2553         cctx->rec_seq = NULL;
2554 free_iv:
2555         kfree(cctx->iv);
2556         cctx->iv = NULL;
2557 free_priv:
2558         if (tx) {
2559                 kfree(ctx->priv_ctx_tx);
2560                 ctx->priv_ctx_tx = NULL;
2561         } else {
2562                 kfree(ctx->priv_ctx_rx);
2563                 ctx->priv_ctx_rx = NULL;
2564         }
2565 out:
2566         return rc;
2567 }