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 #include "tls.h"
  48
  49 struct tls_decrypt_arg {
  50         bool zc;
  51         bool async;
  52         u8 tail;
  53 };
  54
  55 struct tls_decrypt_ctx {
  56         u8 iv[MAX_IV_SIZE];
  57         u8 aad[TLS_MAX_AAD_SIZE];
  58         u8 tail;
  59         struct scatterlist sg[];
  60 };
  61
  62 noinline void tls_err_abort(struct sock *sk, int err)
  63 {
  64         WARN_ON_ONCE(err >= 0);
  65         /* sk->sk_err should contain a positive error code. */
  66         sk->sk_err = -err;
  67         sk_error_report(sk);
  68 }
  69
  70 static int __skb_nsg(struct sk_buff *skb, int offset, int len,
  71                      unsigned int recursion_level)
  72 {
  73         int start = skb_headlen(skb);
  74         int i, chunk = start - offset;
  75         struct sk_buff *frag_iter;
  76         int elt = 0;
  77
  78         if (unlikely(recursion_level >= 24))
  79                 return -EMSGSIZE;
  80
  81         if (chunk > 0) {
  82                 if (chunk > len)
  83                         chunk = len;
  84                 elt++;
  85                 len -= chunk;
  86                 if (len == 0)
  87                         return elt;
  88                 offset += chunk;
  89         }
  90
  91         for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
  92                 int end;
  93
  94                 WARN_ON(start > offset + len);
  95
  96                 end = start + skb_frag_size(&skb_shinfo(skb)->frags[i]);
  97                 chunk = end - offset;
  98                 if (chunk > 0) {
  99                         if (chunk > len)
 100                                 chunk = len;
 101                         elt++;
 102                         len -= chunk;
 103                         if (len == 0)
 104                                 return elt;
 105                         offset += chunk;
 106                 }
 107                 start = end;
 108         }
 109
 110         if (unlikely(skb_has_frag_list(skb))) {
 111                 skb_walk_frags(skb, frag_iter) {
 112                         int end, ret;
 113
 114                         WARN_ON(start > offset + len);
 115
 116                         end = start + frag_iter->len;
 117                         chunk = end - offset;
 118                         if (chunk > 0) {
 119                                 if (chunk > len)
 120                                         chunk = len;
 121                                 ret = __skb_nsg(frag_iter, offset - start, chunk,
 122                                                 recursion_level + 1);
 123                                 if (unlikely(ret < 0))
 124                                         return ret;
 125                                 elt += ret;
 126                                 len -= chunk;
 127                                 if (len == 0)
 128                                         return elt;
 129                                 offset += chunk;
 130                         }
 131                         start = end;
 132                 }
 133         }
 134         BUG_ON(len);
 135         return elt;
 136 }
 137
 138 /* Return the number of scatterlist elements required to completely map the
 139  * skb, or -EMSGSIZE if the recursion depth is exceeded.
 140  */
 141 static int skb_nsg(struct sk_buff *skb, int offset, int len)
 142 {
 143         return __skb_nsg(skb, offset, len, 0);
 144 }
 145
 146 static int tls_padding_length(struct tls_prot_info *prot, struct sk_buff *skb,
 147                               struct tls_decrypt_arg *darg)
 148 {
 149         struct strp_msg *rxm = strp_msg(skb);
 150         struct tls_msg *tlm = tls_msg(skb);
 151         int sub = 0;
 152
 153         /* Determine zero-padding length */
 154         if (prot->version == TLS_1_3_VERSION) {
 155                 int offset = rxm->full_len - TLS_TAG_SIZE - 1;
 156                 char content_type = darg->zc ? darg->tail : 0;
 157                 int err;
 158
 159                 while (content_type == 0) {
 160                         if (offset < prot->prepend_size)
 161                                 return -EBADMSG;
 162                         err = skb_copy_bits(skb, rxm->offset + offset,
 163                                             &content_type, 1);
 164                         if (err)
 165                                 return err;
 166                         if (content_type)
 167                                 break;
 168                         sub++;
 169                         offset--;
 170                 }
 171                 tlm->control = content_type;
 172         }
 173         return sub;
 174 }
 175
 176 static void tls_decrypt_done(struct crypto_async_request *req, int err)
 177 {
 178         struct aead_request *aead_req = (struct aead_request *)req;
 179         struct scatterlist *sgout = aead_req->dst;
 180         struct scatterlist *sgin = aead_req->src;
 181         struct tls_sw_context_rx *ctx;
 182         struct tls_context *tls_ctx;
 183         struct tls_prot_info *prot;
 184         struct scatterlist *sg;
 185         struct sk_buff *skb;
 186         unsigned int pages;
 187
 188         skb = (struct sk_buff *)req->data;
 189         tls_ctx = tls_get_ctx(skb->sk);
 190         ctx = tls_sw_ctx_rx(tls_ctx);
 191         prot = &tls_ctx->prot_info;
 192
 193         /* Propagate if there was an err */
 194         if (err) {
 195                 if (err == -EBADMSG)
 196                         TLS_INC_STATS(sock_net(skb->sk),
 197                                       LINUX_MIB_TLSDECRYPTERROR);
 198                 ctx->async_wait.err = err;
 199                 tls_err_abort(skb->sk, err);
 200         } else {
 201                 struct strp_msg *rxm = strp_msg(skb);
 202
 203                 /* No TLS 1.3 support with async crypto */
 204                 WARN_ON(prot->tail_size);
 205
 206                 rxm->offset += prot->prepend_size;
 207                 rxm->full_len -= prot->overhead_size;
 208         }
 209
 210         /* After using skb->sk to propagate sk through crypto async callback
 211          * we need to NULL it again.
 212          */
 213         skb->sk = NULL;
 214
 215
 216         /* Free the destination pages if skb was not decrypted inplace */
 217         if (sgout != sgin) {
 218                 /* Skip the first S/G entry as it points to AAD */
 219                 for_each_sg(sg_next(sgout), sg, UINT_MAX, pages) {
 220                         if (!sg)
 221                                 break;
 222                         put_page(sg_page(sg));
 223                 }
 224         }
 225
 226         kfree(aead_req);
 227
 228         spin_lock_bh(&ctx->decrypt_compl_lock);
 229         if (!atomic_dec_return(&ctx->decrypt_pending))
 230                 complete(&ctx->async_wait.completion);
 231         spin_unlock_bh(&ctx->decrypt_compl_lock);
 232 }
 233
 234 static int tls_do_decryption(struct sock *sk,
 235                              struct sk_buff *skb,
 236                              struct scatterlist *sgin,
 237                              struct scatterlist *sgout,
 238                              char *iv_recv,
 239                              size_t data_len,
 240                              struct aead_request *aead_req,
 241                              struct tls_decrypt_arg *darg)
 242 {
 243         struct tls_context *tls_ctx = tls_get_ctx(sk);
 244         struct tls_prot_info *prot = &tls_ctx->prot_info;
 245         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
 246         int ret;
 247
 248         aead_request_set_tfm(aead_req, ctx->aead_recv);
 249         aead_request_set_ad(aead_req, prot->aad_size);
 250         aead_request_set_crypt(aead_req, sgin, sgout,
 251                                data_len + prot->tag_size,
 252                                (u8 *)iv_recv);
 253
 254         if (darg->async) {
 255                 /* Using skb->sk to push sk through to crypto async callback
 256                  * handler. This allows propagating errors up to the socket
 257                  * if needed. It _must_ be cleared in the async handler
 258                  * before consume_skb is called. We _know_ skb->sk is NULL
 259                  * because it is a clone from strparser.
 260                  */
 261                 skb->sk = sk;
 262                 aead_request_set_callback(aead_req,
 263                                           CRYPTO_TFM_REQ_MAY_BACKLOG,
 264                                           tls_decrypt_done, skb);
 265                 atomic_inc(&ctx->decrypt_pending);
 266         } else {
 267                 aead_request_set_callback(aead_req,
 268                                           CRYPTO_TFM_REQ_MAY_BACKLOG,
 269                                           crypto_req_done, &ctx->async_wait);
 270         }
 271
 272         ret = crypto_aead_decrypt(aead_req);
 273         if (ret == -EINPROGRESS) {
 274                 if (darg->async)
 275                         return 0;
 276
 277                 ret = crypto_wait_req(ret, &ctx->async_wait);
 278         }
 279         darg->async = false;
 280
 281         return ret;
 282 }
 283
 284 static void tls_trim_both_msgs(struct sock *sk, int target_size)
 285 {
 286         struct tls_context *tls_ctx = tls_get_ctx(sk);
 287         struct tls_prot_info *prot = &tls_ctx->prot_info;
 288         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 289         struct tls_rec *rec = ctx->open_rec;
 290
 291         sk_msg_trim(sk, &rec->msg_plaintext, target_size);
 292         if (target_size > 0)
 293                 target_size += prot->overhead_size;
 294         sk_msg_trim(sk, &rec->msg_encrypted, target_size);
 295 }
 296
 297 static int tls_alloc_encrypted_msg(struct sock *sk, int len)
 298 {
 299         struct tls_context *tls_ctx = tls_get_ctx(sk);
 300         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 301         struct tls_rec *rec = ctx->open_rec;
 302         struct sk_msg *msg_en = &rec->msg_encrypted;
 303
 304         return sk_msg_alloc(sk, msg_en, len, 0);
 305 }
 306
 307 static int tls_clone_plaintext_msg(struct sock *sk, int required)
 308 {
 309         struct tls_context *tls_ctx = tls_get_ctx(sk);
 310         struct tls_prot_info *prot = &tls_ctx->prot_info;
 311         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 312         struct tls_rec *rec = ctx->open_rec;
 313         struct sk_msg *msg_pl = &rec->msg_plaintext;
 314         struct sk_msg *msg_en = &rec->msg_encrypted;
 315         int skip, len;
 316
 317         /* We add page references worth len bytes from encrypted sg
 318          * at the end of plaintext sg. It is guaranteed that msg_en
 319          * has enough required room (ensured by caller).
 320          */
 321         len = required - msg_pl->sg.size;
 322
 323         /* Skip initial bytes in msg_en's data to be able to use
 324          * same offset of both plain and encrypted data.
 325          */
 326         skip = prot->prepend_size + msg_pl->sg.size;
 327
 328         return sk_msg_clone(sk, msg_pl, msg_en, skip, len);
 329 }
 330
 331 static struct tls_rec *tls_get_rec(struct sock *sk)
 332 {
 333         struct tls_context *tls_ctx = tls_get_ctx(sk);
 334         struct tls_prot_info *prot = &tls_ctx->prot_info;
 335         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 336         struct sk_msg *msg_pl, *msg_en;
 337         struct tls_rec *rec;
 338         int mem_size;
 339
 340         mem_size = sizeof(struct tls_rec) + crypto_aead_reqsize(ctx->aead_send);
 341
 342         rec = kzalloc(mem_size, sk->sk_allocation);
 343         if (!rec)
 344                 return NULL;
 345
 346         msg_pl = &rec->msg_plaintext;
 347         msg_en = &rec->msg_encrypted;
 348
 349         sk_msg_init(msg_pl);
 350         sk_msg_init(msg_en);
 351
 352         sg_init_table(rec->sg_aead_in, 2);
 353         sg_set_buf(&rec->sg_aead_in[0], rec->aad_space, prot->aad_size);
 354         sg_unmark_end(&rec->sg_aead_in[1]);
 355
 356         sg_init_table(rec->sg_aead_out, 2);
 357         sg_set_buf(&rec->sg_aead_out[0], rec->aad_space, prot->aad_size);
 358         sg_unmark_end(&rec->sg_aead_out[1]);
 359
 360         return rec;
 361 }
 362
 363 static void tls_free_rec(struct sock *sk, struct tls_rec *rec)
 364 {
 365         sk_msg_free(sk, &rec->msg_encrypted);
 366         sk_msg_free(sk, &rec->msg_plaintext);
 367         kfree(rec);
 368 }
 369
 370 static void tls_free_open_rec(struct sock *sk)
 371 {
 372         struct tls_context *tls_ctx = tls_get_ctx(sk);
 373         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 374         struct tls_rec *rec = ctx->open_rec;
 375
 376         if (rec) {
 377                 tls_free_rec(sk, rec);
 378                 ctx->open_rec = NULL;
 379         }
 380 }
 381
 382 int tls_tx_records(struct sock *sk, int flags)
 383 {
 384         struct tls_context *tls_ctx = tls_get_ctx(sk);
 385         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 386         struct tls_rec *rec, *tmp;
 387         struct sk_msg *msg_en;
 388         int tx_flags, rc = 0;
 389
 390         if (tls_is_partially_sent_record(tls_ctx)) {
 391                 rec = list_first_entry(&ctx->tx_list,
 392                                        struct tls_rec, list);
 393
 394                 if (flags == -1)
 395                         tx_flags = rec->tx_flags;
 396                 else
 397                         tx_flags = flags;
 398
 399                 rc = tls_push_partial_record(sk, tls_ctx, tx_flags);
 400                 if (rc)
 401                         goto tx_err;
 402
 403                 /* Full record has been transmitted.
 404                  * Remove the head of tx_list
 405                  */
 406                 list_del(&rec->list);
 407                 sk_msg_free(sk, &rec->msg_plaintext);
 408                 kfree(rec);
 409         }
 410
 411         /* Tx all ready records */
 412         list_for_each_entry_safe(rec, tmp, &ctx->tx_list, list) {
 413                 if (READ_ONCE(rec->tx_ready)) {
 414                         if (flags == -1)
 415                                 tx_flags = rec->tx_flags;
 416                         else
 417                                 tx_flags = flags;
 418
 419                         msg_en = &rec->msg_encrypted;
 420                         rc = tls_push_sg(sk, tls_ctx,
 421                                          &msg_en->sg.data[msg_en->sg.curr],
 422                                          0, tx_flags);
 423                         if (rc)
 424                                 goto tx_err;
 425
 426                         list_del(&rec->list);
 427                         sk_msg_free(sk, &rec->msg_plaintext);
 428                         kfree(rec);
 429                 } else {
 430                         break;
 431                 }
 432         }
 433
 434 tx_err:
 435         if (rc < 0 && rc != -EAGAIN)
 436                 tls_err_abort(sk, -EBADMSG);
 437
 438         return rc;
 439 }
 440
 441 static void tls_encrypt_done(struct crypto_async_request *req, int err)
 442 {
 443         struct aead_request *aead_req = (struct aead_request *)req;
 444         struct sock *sk = req->data;
 445         struct tls_context *tls_ctx = tls_get_ctx(sk);
 446         struct tls_prot_info *prot = &tls_ctx->prot_info;
 447         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 448         struct scatterlist *sge;
 449         struct sk_msg *msg_en;
 450         struct tls_rec *rec;
 451         bool ready = false;
 452         int pending;
 453
 454         rec = container_of(aead_req, struct tls_rec, aead_req);
 455         msg_en = &rec->msg_encrypted;
 456
 457         sge = sk_msg_elem(msg_en, msg_en->sg.curr);
 458         sge->offset -= prot->prepend_size;
 459         sge->length += prot->prepend_size;
 460
 461         /* Check if error is previously set on socket */
 462         if (err || sk->sk_err) {
 463                 rec = NULL;
 464
 465                 /* If err is already set on socket, return the same code */
 466                 if (sk->sk_err) {
 467                         ctx->async_wait.err = -sk->sk_err;
 468                 } else {
 469                         ctx->async_wait.err = err;
 470                         tls_err_abort(sk, err);
 471                 }
 472         }
 473
 474         if (rec) {
 475                 struct tls_rec *first_rec;
 476
 477                 /* Mark the record as ready for transmission */
 478                 smp_store_mb(rec->tx_ready, true);
 479
 480                 /* If received record is at head of tx_list, schedule tx */
 481                 first_rec = list_first_entry(&ctx->tx_list,
 482                                              struct tls_rec, list);
 483                 if (rec == first_rec)
 484                         ready = true;
 485         }
 486
 487         spin_lock_bh(&ctx->encrypt_compl_lock);
 488         pending = atomic_dec_return(&ctx->encrypt_pending);
 489
 490         if (!pending && ctx->async_notify)
 491                 complete(&ctx->async_wait.completion);
 492         spin_unlock_bh(&ctx->encrypt_compl_lock);
 493
 494         if (!ready)
 495                 return;
 496
 497         /* Schedule the transmission */
 498         if (!test_and_set_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask))
 499                 schedule_delayed_work(&ctx->tx_work.work, 1);
 500 }
 501
 502 static int tls_do_encryption(struct sock *sk,
 503                              struct tls_context *tls_ctx,
 504                              struct tls_sw_context_tx *ctx,
 505                              struct aead_request *aead_req,
 506                              size_t data_len, u32 start)
 507 {
 508         struct tls_prot_info *prot = &tls_ctx->prot_info;
 509         struct tls_rec *rec = ctx->open_rec;
 510         struct sk_msg *msg_en = &rec->msg_encrypted;
 511         struct scatterlist *sge = sk_msg_elem(msg_en, start);
 512         int rc, iv_offset = 0;
 513
 514         /* For CCM based ciphers, first byte of IV is a constant */
 515         switch (prot->cipher_type) {
 516         case TLS_CIPHER_AES_CCM_128:
 517                 rec->iv_data[0] = TLS_AES_CCM_IV_B0_BYTE;
 518                 iv_offset = 1;
 519                 break;
 520         case TLS_CIPHER_SM4_CCM:
 521                 rec->iv_data[0] = TLS_SM4_CCM_IV_B0_BYTE;
 522                 iv_offset = 1;
 523                 break;
 524         }
 525
 526         memcpy(&rec->iv_data[iv_offset], tls_ctx->tx.iv,
 527                prot->iv_size + prot->salt_size);
 528
 529         tls_xor_iv_with_seq(prot, rec->iv_data + iv_offset,
 530                             tls_ctx->tx.rec_seq);
 531
 532         sge->offset += prot->prepend_size;
 533         sge->length -= prot->prepend_size;
 534
 535         msg_en->sg.curr = start;
 536
 537         aead_request_set_tfm(aead_req, ctx->aead_send);
 538         aead_request_set_ad(aead_req, prot->aad_size);
 539         aead_request_set_crypt(aead_req, rec->sg_aead_in,
 540                                rec->sg_aead_out,
 541                                data_len, rec->iv_data);
 542
 543         aead_request_set_callback(aead_req, CRYPTO_TFM_REQ_MAY_BACKLOG,
 544                                   tls_encrypt_done, sk);
 545
 546         /* Add the record in tx_list */
 547         list_add_tail((struct list_head *)&rec->list, &ctx->tx_list);
 548         atomic_inc(&ctx->encrypt_pending);
 549
 550         rc = crypto_aead_encrypt(aead_req);
 551         if (!rc || rc != -EINPROGRESS) {
 552                 atomic_dec(&ctx->encrypt_pending);
 553                 sge->offset -= prot->prepend_size;
 554                 sge->length += prot->prepend_size;
 555         }
 556
 557         if (!rc) {
 558                 WRITE_ONCE(rec->tx_ready, true);
 559         } else if (rc != -EINPROGRESS) {
 560                 list_del(&rec->list);
 561                 return rc;
 562         }
 563
 564         /* Unhook the record from context if encryption is not failure */
 565         ctx->open_rec = NULL;
 566         tls_advance_record_sn(sk, prot, &tls_ctx->tx);
 567         return rc;
 568 }
 569
 570 static int tls_split_open_record(struct sock *sk, struct tls_rec *from,
 571                                  struct tls_rec **to, struct sk_msg *msg_opl,
 572                                  struct sk_msg *msg_oen, u32 split_point,
 573                                  u32 tx_overhead_size, u32 *orig_end)
 574 {
 575         u32 i, j, bytes = 0, apply = msg_opl->apply_bytes;
 576         struct scatterlist *sge, *osge, *nsge;
 577         u32 orig_size = msg_opl->sg.size;
 578         struct scatterlist tmp = { };
 579         struct sk_msg *msg_npl;
 580         struct tls_rec *new;
 581         int ret;
 582
 583         new = tls_get_rec(sk);
 584         if (!new)
 585                 return -ENOMEM;
 586         ret = sk_msg_alloc(sk, &new->msg_encrypted, msg_opl->sg.size +
 587                            tx_overhead_size, 0);
 588         if (ret < 0) {
 589                 tls_free_rec(sk, new);
 590                 return ret;
 591         }
 592
 593         *orig_end = msg_opl->sg.end;
 594         i = msg_opl->sg.start;
 595         sge = sk_msg_elem(msg_opl, i);
 596         while (apply && sge->length) {
 597                 if (sge->length > apply) {
 598                         u32 len = sge->length - apply;
 599
 600                         get_page(sg_page(sge));
 601                         sg_set_page(&tmp, sg_page(sge), len,
 602                                     sge->offset + apply);
 603                         sge->length = apply;
 604                         bytes += apply;
 605                         apply = 0;
 606                 } else {
 607                         apply -= sge->length;
 608                         bytes += sge->length;
 609                 }
 610
 611                 sk_msg_iter_var_next(i);
 612                 if (i == msg_opl->sg.end)
 613                         break;
 614                 sge = sk_msg_elem(msg_opl, i);
 615         }
 616
 617         msg_opl->sg.end = i;
 618         msg_opl->sg.curr = i;
 619         msg_opl->sg.copybreak = 0;
 620         msg_opl->apply_bytes = 0;
 621         msg_opl->sg.size = bytes;
 622
 623         msg_npl = &new->msg_plaintext;
 624         msg_npl->apply_bytes = apply;
 625         msg_npl->sg.size = orig_size - bytes;
 626
 627         j = msg_npl->sg.start;
 628         nsge = sk_msg_elem(msg_npl, j);
 629         if (tmp.length) {
 630                 memcpy(nsge, &tmp, sizeof(*nsge));
 631                 sk_msg_iter_var_next(j);
 632                 nsge = sk_msg_elem(msg_npl, j);
 633         }
 634
 635         osge = sk_msg_elem(msg_opl, i);
 636         while (osge->length) {
 637                 memcpy(nsge, osge, sizeof(*nsge));
 638                 sg_unmark_end(nsge);
 639                 sk_msg_iter_var_next(i);
 640                 sk_msg_iter_var_next(j);
 641                 if (i == *orig_end)
 642                         break;
 643                 osge = sk_msg_elem(msg_opl, i);
 644                 nsge = sk_msg_elem(msg_npl, j);
 645         }
 646
 647         msg_npl->sg.end = j;
 648         msg_npl->sg.curr = j;
 649         msg_npl->sg.copybreak = 0;
 650
 651         *to = new;
 652         return 0;
 653 }
 654
 655 static void tls_merge_open_record(struct sock *sk, struct tls_rec *to,
 656                                   struct tls_rec *from, u32 orig_end)
 657 {
 658         struct sk_msg *msg_npl = &from->msg_plaintext;
 659         struct sk_msg *msg_opl = &to->msg_plaintext;
 660         struct scatterlist *osge, *nsge;
 661         u32 i, j;
 662
 663         i = msg_opl->sg.end;
 664         sk_msg_iter_var_prev(i);
 665         j = msg_npl->sg.start;
 666
 667         osge = sk_msg_elem(msg_opl, i);
 668         nsge = sk_msg_elem(msg_npl, j);
 669
 670         if (sg_page(osge) == sg_page(nsge) &&
 671             osge->offset + osge->length == nsge->offset) {
 672                 osge->length += nsge->length;
 673                 put_page(sg_page(nsge));
 674         }
 675
 676         msg_opl->sg.end = orig_end;
 677         msg_opl->sg.curr = orig_end;
 678         msg_opl->sg.copybreak = 0;
 679         msg_opl->apply_bytes = msg_opl->sg.size + msg_npl->sg.size;
 680         msg_opl->sg.size += msg_npl->sg.size;
 681
 682         sk_msg_free(sk, &to->msg_encrypted);
 683         sk_msg_xfer_full(&to->msg_encrypted, &from->msg_encrypted);
 684
 685         kfree(from);
 686 }
 687
 688 static int tls_push_record(struct sock *sk, int flags,
 689                            unsigned char record_type)
 690 {
 691         struct tls_context *tls_ctx = tls_get_ctx(sk);
 692         struct tls_prot_info *prot = &tls_ctx->prot_info;
 693         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 694         struct tls_rec *rec = ctx->open_rec, *tmp = NULL;
 695         u32 i, split_point, orig_end;
 696         struct sk_msg *msg_pl, *msg_en;
 697         struct aead_request *req;
 698         bool split;
 699         int rc;
 700
 701         if (!rec)
 702                 return 0;
 703
 704         msg_pl = &rec->msg_plaintext;
 705         msg_en = &rec->msg_encrypted;
 706
 707         split_point = msg_pl->apply_bytes;
 708         split = split_point && split_point < msg_pl->sg.size;
 709         if (unlikely((!split &&
 710                       msg_pl->sg.size +
 711                       prot->overhead_size > msg_en->sg.size) ||
 712                      (split &&
 713                       split_point +
 714                       prot->overhead_size > msg_en->sg.size))) {
 715                 split = true;
 716                 split_point = msg_en->sg.size;
 717         }
 718         if (split) {
 719                 rc = tls_split_open_record(sk, rec, &tmp, msg_pl, msg_en,
 720                                            split_point, prot->overhead_size,
 721                                            &orig_end);
 722                 if (rc < 0)
 723                         return rc;
 724                 /* This can happen if above tls_split_open_record allocates
 725                  * a single large encryption buffer instead of two smaller
 726                  * ones. In this case adjust pointers and continue without
 727                  * split.
 728                  */
 729                 if (!msg_pl->sg.size) {
 730                         tls_merge_open_record(sk, rec, tmp, orig_end);
 731                         msg_pl = &rec->msg_plaintext;
 732                         msg_en = &rec->msg_encrypted;
 733                         split = false;
 734                 }
 735                 sk_msg_trim(sk, msg_en, msg_pl->sg.size +
 736                             prot->overhead_size);
 737         }
 738
 739         rec->tx_flags = flags;
 740         req = &rec->aead_req;
 741
 742         i = msg_pl->sg.end;
 743         sk_msg_iter_var_prev(i);
 744
 745         rec->content_type = record_type;
 746         if (prot->version == TLS_1_3_VERSION) {
 747                 /* Add content type to end of message.  No padding added */
 748                 sg_set_buf(&rec->sg_content_type, &rec->content_type, 1);
 749                 sg_mark_end(&rec->sg_content_type);
 750                 sg_chain(msg_pl->sg.data, msg_pl->sg.end + 1,
 751                          &rec->sg_content_type);
 752         } else {
 753                 sg_mark_end(sk_msg_elem(msg_pl, i));
 754         }
 755
 756         if (msg_pl->sg.end < msg_pl->sg.start) {
 757                 sg_chain(&msg_pl->sg.data[msg_pl->sg.start],
 758                          MAX_SKB_FRAGS - msg_pl->sg.start + 1,
 759                          msg_pl->sg.data);
 760         }
 761
 762         i = msg_pl->sg.start;
 763         sg_chain(rec->sg_aead_in, 2, &msg_pl->sg.data[i]);
 764
 765         i = msg_en->sg.end;
 766         sk_msg_iter_var_prev(i);
 767         sg_mark_end(sk_msg_elem(msg_en, i));
 768
 769         i = msg_en->sg.start;
 770         sg_chain(rec->sg_aead_out, 2, &msg_en->sg.data[i]);
 771
 772         tls_make_aad(rec->aad_space, msg_pl->sg.size + prot->tail_size,
 773                      tls_ctx->tx.rec_seq, record_type, prot);
 774
 775         tls_fill_prepend(tls_ctx,
 776                          page_address(sg_page(&msg_en->sg.data[i])) +
 777                          msg_en->sg.data[i].offset,
 778                          msg_pl->sg.size + prot->tail_size,
 779                          record_type);
 780
 781         tls_ctx->pending_open_record_frags = false;
 782
 783         rc = tls_do_encryption(sk, tls_ctx, ctx, req,
 784                                msg_pl->sg.size + prot->tail_size, i);
 785         if (rc < 0) {
 786                 if (rc != -EINPROGRESS) {
 787                         tls_err_abort(sk, -EBADMSG);
 788                         if (split) {
 789                                 tls_ctx->pending_open_record_frags = true;
 790                                 tls_merge_open_record(sk, rec, tmp, orig_end);
 791                         }
 792                 }
 793                 ctx->async_capable = 1;
 794                 return rc;
 795         } else if (split) {
 796                 msg_pl = &tmp->msg_plaintext;
 797                 msg_en = &tmp->msg_encrypted;
 798                 sk_msg_trim(sk, msg_en, msg_pl->sg.size + prot->overhead_size);
 799                 tls_ctx->pending_open_record_frags = true;
 800                 ctx->open_rec = tmp;
 801         }
 802
 803         return tls_tx_records(sk, flags);
 804 }
 805
 806 static int bpf_exec_tx_verdict(struct sk_msg *msg, struct sock *sk,
 807                                bool full_record, u8 record_type,
 808                                ssize_t *copied, int flags)
 809 {
 810         struct tls_context *tls_ctx = tls_get_ctx(sk);
 811         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 812         struct sk_msg msg_redir = { };
 813         struct sk_psock *psock;
 814         struct sock *sk_redir;
 815         struct tls_rec *rec;
 816         bool enospc, policy;
 817         int err = 0, send;
 818         u32 delta = 0;
 819
 820         policy = !(flags & MSG_SENDPAGE_NOPOLICY);
 821         psock = sk_psock_get(sk);
 822         if (!psock || !policy) {
 823                 err = tls_push_record(sk, flags, record_type);
 824                 if (err && sk->sk_err == EBADMSG) {
 825                         *copied -= sk_msg_free(sk, msg);
 826                         tls_free_open_rec(sk);
 827                         err = -sk->sk_err;
 828                 }
 829                 if (psock)
 830                         sk_psock_put(sk, psock);
 831                 return err;
 832         }
 833 more_data:
 834         enospc = sk_msg_full(msg);
 835         if (psock->eval == __SK_NONE) {
 836                 delta = msg->sg.size;
 837                 psock->eval = sk_psock_msg_verdict(sk, psock, msg);
 838                 delta -= msg->sg.size;
 839         }
 840         if (msg->cork_bytes && msg->cork_bytes > msg->sg.size &&
 841             !enospc && !full_record) {
 842                 err = -ENOSPC;
 843                 goto out_err;
 844         }
 845         msg->cork_bytes = 0;
 846         send = msg->sg.size;
 847         if (msg->apply_bytes && msg->apply_bytes < send)
 848                 send = msg->apply_bytes;
 849
 850         switch (psock->eval) {
 851         case __SK_PASS:
 852                 err = tls_push_record(sk, flags, record_type);
 853                 if (err && sk->sk_err == EBADMSG) {
 854                         *copied -= sk_msg_free(sk, msg);
 855                         tls_free_open_rec(sk);
 856                         err = -sk->sk_err;
 857                         goto out_err;
 858                 }
 859                 break;
 860         case __SK_REDIRECT:
 861                 sk_redir = psock->sk_redir;
 862                 memcpy(&msg_redir, msg, sizeof(*msg));
 863                 if (msg->apply_bytes < send)
 864                         msg->apply_bytes = 0;
 865                 else
 866                         msg->apply_bytes -= send;
 867                 sk_msg_return_zero(sk, msg, send);
 868                 msg->sg.size -= send;
 869                 release_sock(sk);
 870                 err = tcp_bpf_sendmsg_redir(sk_redir, &msg_redir, send, flags);
 871                 lock_sock(sk);
 872                 if (err < 0) {
 873                         *copied -= sk_msg_free_nocharge(sk, &msg_redir);
 874                         msg->sg.size = 0;
 875                 }
 876                 if (msg->sg.size == 0)
 877                         tls_free_open_rec(sk);
 878                 break;
 879         case __SK_DROP:
 880         default:
 881                 sk_msg_free_partial(sk, msg, send);
 882                 if (msg->apply_bytes < send)
 883                         msg->apply_bytes = 0;
 884                 else
 885                         msg->apply_bytes -= send;
 886                 if (msg->sg.size == 0)
 887                         tls_free_open_rec(sk);
 888                 *copied -= (send + delta);
 889                 err = -EACCES;
 890         }
 891
 892         if (likely(!err)) {
 893                 bool reset_eval = !ctx->open_rec;
 894
 895                 rec = ctx->open_rec;
 896                 if (rec) {
 897                         msg = &rec->msg_plaintext;
 898                         if (!msg->apply_bytes)
 899                                 reset_eval = true;
 900                 }
 901                 if (reset_eval) {
 902                         psock->eval = __SK_NONE;
 903                         if (psock->sk_redir) {
 904                                 sock_put(psock->sk_redir);
 905                                 psock->sk_redir = NULL;
 906                         }
 907                 }
 908                 if (rec)
 909                         goto more_data;
 910         }
 911  out_err:
 912         sk_psock_put(sk, psock);
 913         return err;
 914 }
 915
 916 static int tls_sw_push_pending_record(struct sock *sk, int flags)
 917 {
 918         struct tls_context *tls_ctx = tls_get_ctx(sk);
 919         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 920         struct tls_rec *rec = ctx->open_rec;
 921         struct sk_msg *msg_pl;
 922         size_t copied;
 923
 924         if (!rec)
 925                 return 0;
 926
 927         msg_pl = &rec->msg_plaintext;
 928         copied = msg_pl->sg.size;
 929         if (!copied)
 930                 return 0;
 931
 932         return bpf_exec_tx_verdict(msg_pl, sk, true, TLS_RECORD_TYPE_DATA,
 933                                    &copied, flags);
 934 }
 935
 936 int tls_sw_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
 937 {
 938         long timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
 939         struct tls_context *tls_ctx = tls_get_ctx(sk);
 940         struct tls_prot_info *prot = &tls_ctx->prot_info;
 941         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 942         bool async_capable = ctx->async_capable;
 943         unsigned char record_type = TLS_RECORD_TYPE_DATA;
 944         bool is_kvec = iov_iter_is_kvec(&msg->msg_iter);
 945         bool eor = !(msg->msg_flags & MSG_MORE);
 946         size_t try_to_copy;
 947         ssize_t copied = 0;
 948         struct sk_msg *msg_pl, *msg_en;
 949         struct tls_rec *rec;
 950         int required_size;
 951         int num_async = 0;
 952         bool full_record;
 953         int record_room;
 954         int num_zc = 0;
 955         int orig_size;
 956         int ret = 0;
 957         int pending;
 958
 959         if (msg->msg_flags & ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL |
 960                                MSG_CMSG_COMPAT))
 961                 return -EOPNOTSUPP;
 962
 963         mutex_lock(&tls_ctx->tx_lock);
 964         lock_sock(sk);
 965
 966         if (unlikely(msg->msg_controllen)) {
 967                 ret = tls_process_cmsg(sk, msg, &record_type);
 968                 if (ret) {
 969                         if (ret == -EINPROGRESS)
 970                                 num_async++;
 971                         else if (ret != -EAGAIN)
 972                                 goto send_end;
 973                 }
 974         }
 975
 976         while (msg_data_left(msg)) {
 977                 if (sk->sk_err) {
 978                         ret = -sk->sk_err;
 979                         goto send_end;
 980                 }
 981
 982                 if (ctx->open_rec)
 983                         rec = ctx->open_rec;
 984                 else
 985                         rec = ctx->open_rec = tls_get_rec(sk);
 986                 if (!rec) {
 987                         ret = -ENOMEM;
 988                         goto send_end;
 989                 }
 990
 991                 msg_pl = &rec->msg_plaintext;
 992                 msg_en = &rec->msg_encrypted;
 993
 994                 orig_size = msg_pl->sg.size;
 995                 full_record = false;
 996                 try_to_copy = msg_data_left(msg);
 997                 record_room = TLS_MAX_PAYLOAD_SIZE - msg_pl->sg.size;
 998                 if (try_to_copy >= record_room) {
 999                         try_to_copy = record_room;
1000                         full_record = true;
1001                 }
1002
1003                 required_size = msg_pl->sg.size + try_to_copy +
1004                                 prot->overhead_size;
1005
1006                 if (!sk_stream_memory_free(sk))
1007                         goto wait_for_sndbuf;
1008
1009 alloc_encrypted:
1010                 ret = tls_alloc_encrypted_msg(sk, required_size);
1011                 if (ret) {
1012                         if (ret != -ENOSPC)
1013                                 goto wait_for_memory;
1014
1015                         /* Adjust try_to_copy according to the amount that was
1016                          * actually allocated. The difference is due
1017                          * to max sg elements limit
1018                          */
1019                         try_to_copy -= required_size - msg_en->sg.size;
1020                         full_record = true;
1021                 }
1022
1023                 if (!is_kvec && (full_record || eor) && !async_capable) {
1024                         u32 first = msg_pl->sg.end;
1025
1026                         ret = sk_msg_zerocopy_from_iter(sk, &msg->msg_iter,
1027                                                         msg_pl, try_to_copy);
1028                         if (ret)
1029                                 goto fallback_to_reg_send;
1030
1031                         num_zc++;
1032                         copied += try_to_copy;
1033
1034                         sk_msg_sg_copy_set(msg_pl, first);
1035                         ret = bpf_exec_tx_verdict(msg_pl, sk, full_record,
1036                                                   record_type, &copied,
1037                                                   msg->msg_flags);
1038                         if (ret) {
1039                                 if (ret == -EINPROGRESS)
1040                                         num_async++;
1041                                 else if (ret == -ENOMEM)
1042                                         goto wait_for_memory;
1043                                 else if (ctx->open_rec && ret == -ENOSPC)
1044                                         goto rollback_iter;
1045                                 else if (ret != -EAGAIN)
1046                                         goto send_end;
1047                         }
1048                         continue;
1049 rollback_iter:
1050                         copied -= try_to_copy;
1051                         sk_msg_sg_copy_clear(msg_pl, first);
1052                         iov_iter_revert(&msg->msg_iter,
1053                                         msg_pl->sg.size - orig_size);
1054 fallback_to_reg_send:
1055                         sk_msg_trim(sk, msg_pl, orig_size);
1056                 }
1057
1058                 required_size = msg_pl->sg.size + try_to_copy;
1059
1060                 ret = tls_clone_plaintext_msg(sk, required_size);
1061                 if (ret) {
1062                         if (ret != -ENOSPC)
1063                                 goto send_end;
1064
1065                         /* Adjust try_to_copy according to the amount that was
1066                          * actually allocated. The difference is due
1067                          * to max sg elements limit
1068                          */
1069                         try_to_copy -= required_size - msg_pl->sg.size;
1070                         full_record = true;
1071                         sk_msg_trim(sk, msg_en,
1072                                     msg_pl->sg.size + prot->overhead_size);
1073                 }
1074
1075                 if (try_to_copy) {
1076                         ret = sk_msg_memcopy_from_iter(sk, &msg->msg_iter,
1077                                                        msg_pl, try_to_copy);
1078                         if (ret < 0)
1079                                 goto trim_sgl;
1080                 }
1081
1082                 /* Open records defined only if successfully copied, otherwise
1083                  * we would trim the sg but not reset the open record frags.
1084                  */
1085                 tls_ctx->pending_open_record_frags = true;
1086                 copied += try_to_copy;
1087                 if (full_record || eor) {
1088                         ret = bpf_exec_tx_verdict(msg_pl, sk, full_record,
1089                                                   record_type, &copied,
1090                                                   msg->msg_flags);
1091                         if (ret) {
1092                                 if (ret == -EINPROGRESS)
1093                                         num_async++;
1094                                 else if (ret == -ENOMEM)
1095                                         goto wait_for_memory;
1096                                 else if (ret != -EAGAIN) {
1097                                         if (ret == -ENOSPC)
1098                                                 ret = 0;
1099                                         goto send_end;
1100                                 }
1101                         }
1102                 }
1103
1104                 continue;
1105
1106 wait_for_sndbuf:
1107                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1108 wait_for_memory:
1109                 ret = sk_stream_wait_memory(sk, &timeo);
1110                 if (ret) {
1111 trim_sgl:
1112                         if (ctx->open_rec)
1113                                 tls_trim_both_msgs(sk, orig_size);
1114                         goto send_end;
1115                 }
1116
1117                 if (ctx->open_rec && msg_en->sg.size < required_size)
1118                         goto alloc_encrypted;
1119         }
1120
1121         if (!num_async) {
1122                 goto send_end;
1123         } else if (num_zc) {
1124                 /* Wait for pending encryptions to get completed */
1125                 spin_lock_bh(&ctx->encrypt_compl_lock);
1126                 ctx->async_notify = true;
1127
1128                 pending = atomic_read(&ctx->encrypt_pending);
1129                 spin_unlock_bh(&ctx->encrypt_compl_lock);
1130                 if (pending)
1131                         crypto_wait_req(-EINPROGRESS, &ctx->async_wait);
1132                 else
1133                         reinit_completion(&ctx->async_wait.completion);
1134
1135                 /* There can be no concurrent accesses, since we have no
1136                  * pending encrypt operations
1137                  */
1138                 WRITE_ONCE(ctx->async_notify, false);
1139
1140                 if (ctx->async_wait.err) {
1141                         ret = ctx->async_wait.err;
1142                         copied = 0;
1143                 }
1144         }
1145
1146         /* Transmit if any encryptions have completed */
1147         if (test_and_clear_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask)) {
1148                 cancel_delayed_work(&ctx->tx_work.work);
1149                 tls_tx_records(sk, msg->msg_flags);
1150         }
1151
1152 send_end:
1153         ret = sk_stream_error(sk, msg->msg_flags, ret);
1154
1155         release_sock(sk);
1156         mutex_unlock(&tls_ctx->tx_lock);
1157         return copied > 0 ? copied : ret;
1158 }
1159
1160 static int tls_sw_do_sendpage(struct sock *sk, struct page *page,
1161                               int offset, size_t size, int flags)
1162 {
1163         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1164         struct tls_context *tls_ctx = tls_get_ctx(sk);
1165         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
1166         struct tls_prot_info *prot = &tls_ctx->prot_info;
1167         unsigned char record_type = TLS_RECORD_TYPE_DATA;
1168         struct sk_msg *msg_pl;
1169         struct tls_rec *rec;
1170         int num_async = 0;
1171         ssize_t copied = 0;
1172         bool full_record;
1173         int record_room;
1174         int ret = 0;
1175         bool eor;
1176
1177         eor = !(flags & MSG_SENDPAGE_NOTLAST);
1178         sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1179
1180         /* Call the sk_stream functions to manage the sndbuf mem. */
1181         while (size > 0) {
1182                 size_t copy, required_size;
1183
1184                 if (sk->sk_err) {
1185                         ret = -sk->sk_err;
1186                         goto sendpage_end;
1187                 }
1188
1189                 if (ctx->open_rec)
1190                         rec = ctx->open_rec;
1191                 else
1192                         rec = ctx->open_rec = tls_get_rec(sk);
1193                 if (!rec) {
1194                         ret = -ENOMEM;
1195                         goto sendpage_end;
1196                 }
1197
1198                 msg_pl = &rec->msg_plaintext;
1199
1200                 full_record = false;
1201                 record_room = TLS_MAX_PAYLOAD_SIZE - msg_pl->sg.size;
1202                 copy = size;
1203                 if (copy >= record_room) {
1204                         copy = record_room;
1205                         full_record = true;
1206                 }
1207
1208                 required_size = msg_pl->sg.size + copy + prot->overhead_size;
1209
1210                 if (!sk_stream_memory_free(sk))
1211                         goto wait_for_sndbuf;
1212 alloc_payload:
1213                 ret = tls_alloc_encrypted_msg(sk, required_size);
1214                 if (ret) {
1215                         if (ret != -ENOSPC)
1216                                 goto wait_for_memory;
1217
1218                         /* Adjust copy according to the amount that was
1219                          * actually allocated. The difference is due
1220                          * to max sg elements limit
1221                          */
1222                         copy -= required_size - msg_pl->sg.size;
1223                         full_record = true;
1224                 }
1225
1226                 sk_msg_page_add(msg_pl, page, copy, offset);
1227                 sk_mem_charge(sk, copy);
1228
1229                 offset += copy;
1230                 size -= copy;
1231                 copied += copy;
1232
1233                 tls_ctx->pending_open_record_frags = true;
1234                 if (full_record || eor || sk_msg_full(msg_pl)) {
1235                         ret = bpf_exec_tx_verdict(msg_pl, sk, full_record,
1236                                                   record_type, &copied, flags);
1237                         if (ret) {
1238                                 if (ret == -EINPROGRESS)
1239                                         num_async++;
1240                                 else if (ret == -ENOMEM)
1241                                         goto wait_for_memory;
1242                                 else if (ret != -EAGAIN) {
1243                                         if (ret == -ENOSPC)
1244                                                 ret = 0;
1245                                         goto sendpage_end;
1246                                 }
1247                         }
1248                 }
1249                 continue;
1250 wait_for_sndbuf:
1251                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1252 wait_for_memory:
1253                 ret = sk_stream_wait_memory(sk, &timeo);
1254                 if (ret) {
1255                         if (ctx->open_rec)
1256                                 tls_trim_both_msgs(sk, msg_pl->sg.size);
1257                         goto sendpage_end;
1258                 }
1259
1260                 if (ctx->open_rec)
1261                         goto alloc_payload;
1262         }
1263
1264         if (num_async) {
1265                 /* Transmit if any encryptions have completed */
1266                 if (test_and_clear_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask)) {
1267                         cancel_delayed_work(&ctx->tx_work.work);
1268                         tls_tx_records(sk, flags);
1269                 }
1270         }
1271 sendpage_end:
1272         ret = sk_stream_error(sk, flags, ret);
1273         return copied > 0 ? copied : ret;
1274 }
1275
1276 int tls_sw_sendpage_locked(struct sock *sk, struct page *page,
1277                            int offset, size_t size, int flags)
1278 {
1279         if (flags & ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL |
1280                       MSG_SENDPAGE_NOTLAST | MSG_SENDPAGE_NOPOLICY |
1281                       MSG_NO_SHARED_FRAGS))
1282                 return -EOPNOTSUPP;
1283
1284         return tls_sw_do_sendpage(sk, page, offset, size, flags);
1285 }
1286
1287 int tls_sw_sendpage(struct sock *sk, struct page *page,
1288                     int offset, size_t size, int flags)
1289 {
1290         struct tls_context *tls_ctx = tls_get_ctx(sk);
1291         int ret;
1292
1293         if (flags & ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL |
1294                       MSG_SENDPAGE_NOTLAST | MSG_SENDPAGE_NOPOLICY))
1295                 return -EOPNOTSUPP;
1296
1297         mutex_lock(&tls_ctx->tx_lock);
1298         lock_sock(sk);
1299         ret = tls_sw_do_sendpage(sk, page, offset, size, flags);
1300         release_sock(sk);
1301         mutex_unlock(&tls_ctx->tx_lock);
1302         return ret;
1303 }
1304
1305 static int
1306 tls_rx_rec_wait(struct sock *sk, struct sk_psock *psock, bool nonblock,
1307                 long timeo)
1308 {
1309         struct tls_context *tls_ctx = tls_get_ctx(sk);
1310         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1311         DEFINE_WAIT_FUNC(wait, woken_wake_function);
1312
1313         while (!ctx->recv_pkt) {
1314                 if (!sk_psock_queue_empty(psock))
1315                         return 0;
1316
1317                 if (sk->sk_err)
1318                         return sock_error(sk);
1319
1320                 if (!skb_queue_empty(&sk->sk_receive_queue)) {
1321                         __strp_unpause(&ctx->strp);
1322                         if (ctx->recv_pkt)
1323                                 break;
1324                 }
1325
1326                 if (sk->sk_shutdown & RCV_SHUTDOWN)
1327                         return 0;
1328
1329                 if (sock_flag(sk, SOCK_DONE))
1330                         return 0;
1331
1332                 if (nonblock || !timeo)
1333                         return -EAGAIN;
1334
1335                 add_wait_queue(sk_sleep(sk), &wait);
1336                 sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1337                 sk_wait_event(sk, &timeo,
1338                               ctx->recv_pkt || !sk_psock_queue_empty(psock),
1339                               &wait);
1340                 sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1341                 remove_wait_queue(sk_sleep(sk), &wait);
1342
1343                 /* Handle signals */
1344                 if (signal_pending(current))
1345                         return sock_intr_errno(timeo);
1346         }
1347
1348         return 1;
1349 }
1350
1351 static int tls_setup_from_iter(struct iov_iter *from,
1352                                int length, int *pages_used,
1353                                struct scatterlist *to,
1354                                int to_max_pages)
1355 {
1356         int rc = 0, i = 0, num_elem = *pages_used, maxpages;
1357         struct page *pages[MAX_SKB_FRAGS];
1358         unsigned int size = 0;
1359         ssize_t copied, use;
1360         size_t offset;
1361
1362         while (length > 0) {
1363                 i = 0;
1364                 maxpages = to_max_pages - num_elem;
1365                 if (maxpages == 0) {
1366                         rc = -EFAULT;
1367                         goto out;
1368                 }
1369                 copied = iov_iter_get_pages(from, pages,
1370                                             length,
1371                                             maxpages, &offset);
1372                 if (copied <= 0) {
1373                         rc = -EFAULT;
1374                         goto out;
1375                 }
1376
1377                 iov_iter_advance(from, copied);
1378
1379                 length -= copied;
1380                 size += copied;
1381                 while (copied) {
1382                         use = min_t(int, copied, PAGE_SIZE - offset);
1383
1384                         sg_set_page(&to[num_elem],
1385                                     pages[i], use, offset);
1386                         sg_unmark_end(&to[num_elem]);
1387                         /* We do not uncharge memory from this API */
1388
1389                         offset = 0;
1390                         copied -= use;
1391
1392                         i++;
1393                         num_elem++;
1394                 }
1395         }
1396         /* Mark the end in the last sg entry if newly added */
1397         if (num_elem > *pages_used)
1398                 sg_mark_end(&to[num_elem - 1]);
1399 out:
1400         if (rc)
1401                 iov_iter_revert(from, 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                             struct tls_decrypt_arg *darg)
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         int n_sgin, n_sgout, aead_size, err, pages = 0;
1424         struct strp_msg *rxm = strp_msg(skb);
1425         struct tls_msg *tlm = tls_msg(skb);
1426         struct aead_request *aead_req;
1427         struct sk_buff *unused;
1428         struct scatterlist *sgin = NULL;
1429         struct scatterlist *sgout = NULL;
1430         const int data_len = rxm->full_len - prot->overhead_size;
1431         int tail_pages = !!prot->tail_size;
1432         struct tls_decrypt_ctx *dctx;
1433         int iv_offset = 0;
1434         u8 *mem;
1435
1436         if (darg->zc && (out_iov || out_sg)) {
1437                 if (out_iov)
1438                         n_sgout = 1 + tail_pages +
1439                                 iov_iter_npages_cap(out_iov, INT_MAX, data_len);
1440                 else
1441                         n_sgout = sg_nents(out_sg);
1442                 n_sgin = skb_nsg(skb, rxm->offset + prot->prepend_size,
1443                                  rxm->full_len - prot->prepend_size);
1444         } else {
1445                 n_sgout = 0;
1446                 darg->zc = false;
1447                 n_sgin = skb_cow_data(skb, 0, &unused);
1448         }
1449
1450         if (n_sgin < 1)
1451                 return -EBADMSG;
1452
1453         /* Increment to accommodate AAD */
1454         n_sgin = n_sgin + 1;
1455
1456         /* Allocate a single block of memory which contains
1457          *   aead_req || tls_decrypt_ctx.
1458          * Both structs are variable length.
1459          */
1460         aead_size = sizeof(*aead_req) + crypto_aead_reqsize(ctx->aead_recv);
1461         mem = kmalloc(aead_size + struct_size(dctx, sg, n_sgin + n_sgout),
1462                       sk->sk_allocation);
1463         if (!mem)
1464                 return -ENOMEM;
1465
1466         /* Segment the allocated memory */
1467         aead_req = (struct aead_request *)mem;
1468         dctx = (struct tls_decrypt_ctx *)(mem + aead_size);
1469         sgin = &dctx->sg[0];
1470         sgout = &dctx->sg[n_sgin];
1471
1472         /* For CCM based ciphers, first byte of nonce+iv is a constant */
1473         switch (prot->cipher_type) {
1474         case TLS_CIPHER_AES_CCM_128:
1475                 dctx->iv[0] = TLS_AES_CCM_IV_B0_BYTE;
1476                 iv_offset = 1;
1477                 break;
1478         case TLS_CIPHER_SM4_CCM:
1479                 dctx->iv[0] = TLS_SM4_CCM_IV_B0_BYTE;
1480                 iv_offset = 1;
1481                 break;
1482         }
1483
1484         /* Prepare IV */
1485         if (prot->version == TLS_1_3_VERSION ||
1486             prot->cipher_type == TLS_CIPHER_CHACHA20_POLY1305) {
1487                 memcpy(&dctx->iv[iv_offset], tls_ctx->rx.iv,
1488                        prot->iv_size + prot->salt_size);
1489         } else {
1490                 err = skb_copy_bits(skb, rxm->offset + TLS_HEADER_SIZE,
1491                                     &dctx->iv[iv_offset] + prot->salt_size,
1492                                     prot->iv_size);
1493                 if (err < 0)
1494                         goto exit_free;
1495                 memcpy(&dctx->iv[iv_offset], tls_ctx->rx.iv, prot->salt_size);
1496         }
1497         tls_xor_iv_with_seq(prot, &dctx->iv[iv_offset], tls_ctx->rx.rec_seq);
1498
1499         /* Prepare AAD */
1500         tls_make_aad(dctx->aad, rxm->full_len - prot->overhead_size +
1501                      prot->tail_size,
1502                      tls_ctx->rx.rec_seq, tlm->control, prot);
1503
1504         /* Prepare sgin */
1505         sg_init_table(sgin, n_sgin);
1506         sg_set_buf(&sgin[0], dctx->aad, prot->aad_size);
1507         err = skb_to_sgvec(skb, &sgin[1],
1508                            rxm->offset + prot->prepend_size,
1509                            rxm->full_len - prot->prepend_size);
1510         if (err < 0)
1511                 goto exit_free;
1512
1513         if (n_sgout) {
1514                 if (out_iov) {
1515                         sg_init_table(sgout, n_sgout);
1516                         sg_set_buf(&sgout[0], dctx->aad, prot->aad_size);
1517
1518                         err = tls_setup_from_iter(out_iov, data_len,
1519                                                   &pages, &sgout[1],
1520                                                   (n_sgout - 1 - tail_pages));
1521                         if (err < 0)
1522                                 goto fallback_to_reg_recv;
1523
1524                         if (prot->tail_size) {
1525                                 sg_unmark_end(&sgout[pages]);
1526                                 sg_set_buf(&sgout[pages + 1], &dctx->tail,
1527                                            prot->tail_size);
1528                                 sg_mark_end(&sgout[pages + 1]);
1529                         }
1530                 } else if (out_sg) {
1531                         memcpy(sgout, out_sg, n_sgout * sizeof(*sgout));
1532                 } else {
1533                         goto fallback_to_reg_recv;
1534                 }
1535         } else {
1536 fallback_to_reg_recv:
1537                 sgout = sgin;
1538                 pages = 0;
1539                 darg->zc = false;
1540         }
1541
1542         /* Prepare and submit AEAD request */
1543         err = tls_do_decryption(sk, skb, sgin, sgout, dctx->iv,
1544                                 data_len + prot->tail_size, aead_req, darg);
1545         if (darg->async)
1546                 return 0;
1547
1548         if (prot->tail_size)
1549                 darg->tail = dctx->tail;
1550
1551         /* Release the pages in case iov was mapped to pages */
1552         for (; pages > 0; pages--)
1553                 put_page(sg_page(&sgout[pages]));
1554 exit_free:
1555         kfree(mem);
1556         return err;
1557 }
1558
1559 static int decrypt_skb_update(struct sock *sk, struct sk_buff *skb,
1560                               struct iov_iter *dest,
1561                               struct tls_decrypt_arg *darg)
1562 {
1563         struct tls_context *tls_ctx = tls_get_ctx(sk);
1564         struct tls_prot_info *prot = &tls_ctx->prot_info;
1565         struct strp_msg *rxm = strp_msg(skb);
1566         struct tls_msg *tlm = tls_msg(skb);
1567         int pad, err;
1568
1569         if (tlm->decrypted) {
1570                 darg->zc = false;
1571                 darg->async = false;
1572                 return 0;
1573         }
1574
1575         if (tls_ctx->rx_conf == TLS_HW) {
1576                 err = tls_device_decrypted(sk, tls_ctx, skb, rxm);
1577                 if (err < 0)
1578                         return err;
1579                 if (err > 0) {
1580                         tlm->decrypted = 1;
1581                         darg->zc = false;
1582                         darg->async = false;
1583                         goto decrypt_done;
1584                 }
1585         }
1586
1587         err = decrypt_internal(sk, skb, dest, NULL, darg);
1588         if (err < 0) {
1589                 if (err == -EBADMSG)
1590                         TLS_INC_STATS(sock_net(sk), LINUX_MIB_TLSDECRYPTERROR);
1591                 return err;
1592         }
1593         if (darg->async)
1594                 goto decrypt_next;
1595         /* If opportunistic TLS 1.3 ZC failed retry without ZC */
1596         if (unlikely(darg->zc && prot->version == TLS_1_3_VERSION &&
1597                      darg->tail != TLS_RECORD_TYPE_DATA)) {
1598                 darg->zc = false;
1599                 if (!darg->tail)
1600                         TLS_INC_STATS(sock_net(sk), LINUX_MIB_TLSRXNOPADVIOL);
1601                 TLS_INC_STATS(sock_net(sk), LINUX_MIB_TLSDECRYPTRETRY);
1602                 return decrypt_skb_update(sk, skb, dest, darg);
1603         }
1604
1605 decrypt_done:
1606         pad = tls_padding_length(prot, skb, darg);
1607         if (pad < 0)
1608                 return pad;
1609
1610         rxm->full_len -= pad;
1611         rxm->offset += prot->prepend_size;
1612         rxm->full_len -= prot->overhead_size;
1613         tlm->decrypted = 1;
1614 decrypt_next:
1615         tls_advance_record_sn(sk, prot, &tls_ctx->rx);
1616
1617         return 0;
1618 }
1619
1620 int decrypt_skb(struct sock *sk, struct sk_buff *skb,
1621                 struct scatterlist *sgout)
1622 {
1623         struct tls_decrypt_arg darg = { .zc = true, };
1624
1625         return decrypt_internal(sk, skb, NULL, sgout, &darg);
1626 }
1627
1628 static int tls_record_content_type(struct msghdr *msg, struct tls_msg *tlm,
1629                                    u8 *control)
1630 {
1631         int err;
1632
1633         if (!*control) {
1634                 *control = tlm->control;
1635                 if (!*control)
1636                         return -EBADMSG;
1637
1638                 err = put_cmsg(msg, SOL_TLS, TLS_GET_RECORD_TYPE,
1639                                sizeof(*control), control);
1640                 if (*control != TLS_RECORD_TYPE_DATA) {
1641                         if (err || msg->msg_flags & MSG_CTRUNC)
1642                                 return -EIO;
1643                 }
1644         } else if (*control != tlm->control) {
1645                 return 0;
1646         }
1647
1648         return 1;
1649 }
1650
1651 /* This function traverses the rx_list in tls receive context to copies the
1652  * decrypted records into the buffer provided by caller zero copy is not
1653  * true. Further, the records are removed from the rx_list if it is not a peek
1654  * case and the record has been consumed completely.
1655  */
1656 static int process_rx_list(struct tls_sw_context_rx *ctx,
1657                            struct msghdr *msg,
1658                            u8 *control,
1659                            size_t skip,
1660                            size_t len,
1661                            bool zc,
1662                            bool is_peek)
1663 {
1664         struct sk_buff *skb = skb_peek(&ctx->rx_list);
1665         struct tls_msg *tlm;
1666         ssize_t copied = 0;
1667         int err;
1668
1669         while (skip && skb) {
1670                 struct strp_msg *rxm = strp_msg(skb);
1671                 tlm = tls_msg(skb);
1672
1673                 err = tls_record_content_type(msg, tlm, control);
1674                 if (err <= 0)
1675                         goto out;
1676
1677                 if (skip < rxm->full_len)
1678                         break;
1679
1680                 skip = skip - rxm->full_len;
1681                 skb = skb_peek_next(skb, &ctx->rx_list);
1682         }
1683
1684         while (len && skb) {
1685                 struct sk_buff *next_skb;
1686                 struct strp_msg *rxm = strp_msg(skb);
1687                 int chunk = min_t(unsigned int, rxm->full_len - skip, len);
1688
1689                 tlm = tls_msg(skb);
1690
1691                 err = tls_record_content_type(msg, tlm, control);
1692                 if (err <= 0)
1693                         goto out;
1694
1695                 if (!zc || (rxm->full_len - skip) > len) {
1696                         err = skb_copy_datagram_msg(skb, rxm->offset + skip,
1697                                                     msg, chunk);
1698                         if (err < 0)
1699                                 goto out;
1700                 }
1701
1702                 len = len - chunk;
1703                 copied = copied + chunk;
1704
1705                 /* Consume the data from record if it is non-peek case*/
1706                 if (!is_peek) {
1707                         rxm->offset = rxm->offset + chunk;
1708                         rxm->full_len = rxm->full_len - chunk;
1709
1710                         /* Return if there is unconsumed data in the record */
1711                         if (rxm->full_len - skip)
1712                                 break;
1713                 }
1714
1715                 /* The remaining skip-bytes must lie in 1st record in rx_list.
1716                  * So from the 2nd record, 'skip' should be 0.
1717                  */
1718                 skip = 0;
1719
1720                 if (msg)
1721                         msg->msg_flags |= MSG_EOR;
1722
1723                 next_skb = skb_peek_next(skb, &ctx->rx_list);
1724
1725                 if (!is_peek) {
1726                         __skb_unlink(skb, &ctx->rx_list);
1727                         consume_skb(skb);
1728                 }
1729
1730                 skb = next_skb;
1731         }
1732         err = 0;
1733
1734 out:
1735         return copied ? : err;
1736 }
1737
1738 static void
1739 tls_read_flush_backlog(struct sock *sk, struct tls_prot_info *prot,
1740                        size_t len_left, size_t decrypted, ssize_t done,
1741                        size_t *flushed_at)
1742 {
1743         size_t max_rec;
1744
1745         if (len_left <= decrypted)
1746                 return;
1747
1748         max_rec = prot->overhead_size - prot->tail_size + TLS_MAX_PAYLOAD_SIZE;
1749         if (done - *flushed_at < SZ_128K && tcp_inq(sk) > max_rec)
1750                 return;
1751
1752         *flushed_at = done;
1753         sk_flush_backlog(sk);
1754 }
1755
1756 static long tls_rx_reader_lock(struct sock *sk, struct tls_sw_context_rx *ctx,
1757                                bool nonblock)
1758 {
1759         long timeo;
1760
1761         lock_sock(sk);
1762
1763         timeo = sock_rcvtimeo(sk, nonblock);
1764
1765         while (unlikely(ctx->reader_present)) {
1766                 DEFINE_WAIT_FUNC(wait, woken_wake_function);
1767
1768                 ctx->reader_contended = 1;
1769
1770                 add_wait_queue(&ctx->wq, &wait);
1771                 sk_wait_event(sk, &timeo,
1772                               !READ_ONCE(ctx->reader_present), &wait);
1773                 remove_wait_queue(&ctx->wq, &wait);
1774
1775                 if (!timeo)
1776                         return -EAGAIN;
1777                 if (signal_pending(current))
1778                         return sock_intr_errno(timeo);
1779         }
1780
1781         WRITE_ONCE(ctx->reader_present, 1);
1782
1783         return timeo;
1784 }
1785
1786 static void tls_rx_reader_unlock(struct sock *sk, struct tls_sw_context_rx *ctx)
1787 {
1788         if (unlikely(ctx->reader_contended)) {
1789                 if (wq_has_sleeper(&ctx->wq))
1790                         wake_up(&ctx->wq);
1791                 else
1792                         ctx->reader_contended = 0;
1793
1794                 WARN_ON_ONCE(!ctx->reader_present);
1795         }
1796
1797         WRITE_ONCE(ctx->reader_present, 0);
1798         release_sock(sk);
1799 }
1800
1801 int tls_sw_recvmsg(struct sock *sk,
1802                    struct msghdr *msg,
1803                    size_t len,
1804                    int flags,
1805                    int *addr_len)
1806 {
1807         struct tls_context *tls_ctx = tls_get_ctx(sk);
1808         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1809         struct tls_prot_info *prot = &tls_ctx->prot_info;
1810         struct sk_psock *psock;
1811         unsigned char control = 0;
1812         ssize_t decrypted = 0;
1813         size_t flushed_at = 0;
1814         struct strp_msg *rxm;
1815         struct tls_msg *tlm;
1816         struct sk_buff *skb;
1817         ssize_t copied = 0;
1818         bool async = false;
1819         int target, err = 0;
1820         long timeo;
1821         bool is_kvec = iov_iter_is_kvec(&msg->msg_iter);
1822         bool is_peek = flags & MSG_PEEK;
1823         bool bpf_strp_enabled;
1824         bool zc_capable;
1825
1826         if (unlikely(flags & MSG_ERRQUEUE))
1827                 return sock_recv_errqueue(sk, msg, len, SOL_IP, IP_RECVERR);
1828
1829         psock = sk_psock_get(sk);
1830         timeo = tls_rx_reader_lock(sk, ctx, flags & MSG_DONTWAIT);
1831         if (timeo < 0)
1832                 return timeo;
1833         bpf_strp_enabled = sk_psock_strp_enabled(psock);
1834
1835         /* If crypto failed the connection is broken */
1836         err = ctx->async_wait.err;
1837         if (err)
1838                 goto end;
1839
1840         /* Process pending decrypted records. It must be non-zero-copy */
1841         err = process_rx_list(ctx, msg, &control, 0, len, false, is_peek);
1842         if (err < 0)
1843                 goto end;
1844
1845         copied = err;
1846         if (len <= copied)
1847                 goto end;
1848
1849         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1850         len = len - copied;
1851
1852         zc_capable = !bpf_strp_enabled && !is_kvec && !is_peek &&
1853                 ctx->zc_capable;
1854         decrypted = 0;
1855         while (len && (decrypted + copied < target || ctx->recv_pkt)) {
1856                 struct tls_decrypt_arg darg = {};
1857                 int to_decrypt, chunk;
1858
1859                 err = tls_rx_rec_wait(sk, psock, flags & MSG_DONTWAIT, timeo);
1860                 if (err <= 0) {
1861                         if (psock) {
1862                                 chunk = sk_msg_recvmsg(sk, psock, msg, len,
1863                                                        flags);
1864                                 if (chunk > 0)
1865                                         goto leave_on_list;
1866                         }
1867                         goto recv_end;
1868                 }
1869
1870                 skb = ctx->recv_pkt;
1871                 rxm = strp_msg(skb);
1872                 tlm = tls_msg(skb);
1873
1874                 to_decrypt = rxm->full_len - prot->overhead_size;
1875
1876                 if (zc_capable && to_decrypt <= len &&
1877                     tlm->control == TLS_RECORD_TYPE_DATA)
1878                         darg.zc = true;
1879
1880                 /* Do not use async mode if record is non-data */
1881                 if (tlm->control == TLS_RECORD_TYPE_DATA && !bpf_strp_enabled)
1882                         darg.async = ctx->async_capable;
1883                 else
1884                         darg.async = false;
1885
1886                 err = decrypt_skb_update(sk, skb, &msg->msg_iter, &darg);
1887                 if (err < 0) {
1888                         tls_err_abort(sk, -EBADMSG);
1889                         goto recv_end;
1890                 }
1891
1892                 async |= darg.async;
1893
1894                 /* If the type of records being processed is not known yet,
1895                  * set it to record type just dequeued. If it is already known,
1896                  * but does not match the record type just dequeued, go to end.
1897                  * We always get record type here since for tls1.2, record type
1898                  * is known just after record is dequeued from stream parser.
1899                  * For tls1.3, we disable async.
1900                  */
1901                 err = tls_record_content_type(msg, tlm, &control);
1902                 if (err <= 0)
1903                         goto recv_end;
1904
1905                 /* periodically flush backlog, and feed strparser */
1906                 tls_read_flush_backlog(sk, prot, len, to_decrypt,
1907                                        decrypted + copied, &flushed_at);
1908
1909                 ctx->recv_pkt = NULL;
1910                 __strp_unpause(&ctx->strp);
1911                 __skb_queue_tail(&ctx->rx_list, skb);
1912
1913                 if (async) {
1914                         /* TLS 1.2-only, to_decrypt must be text length */
1915                         chunk = min_t(int, to_decrypt, len);
1916 leave_on_list:
1917                         decrypted += chunk;
1918                         len -= chunk;
1919                         continue;
1920                 }
1921                 /* TLS 1.3 may have updated the length by more than overhead */
1922                 chunk = rxm->full_len;
1923
1924                 if (!darg.zc) {
1925                         bool partially_consumed = chunk > len;
1926
1927                         if (bpf_strp_enabled) {
1928                                 /* BPF may try to queue the skb */
1929                                 __skb_unlink(skb, &ctx->rx_list);
1930                                 err = sk_psock_tls_strp_read(psock, skb);
1931                                 if (err != __SK_PASS) {
1932                                         rxm->offset = rxm->offset + rxm->full_len;
1933                                         rxm->full_len = 0;
1934                                         if (err == __SK_DROP)
1935                                                 consume_skb(skb);
1936                                         continue;
1937                                 }
1938                                 __skb_queue_tail(&ctx->rx_list, skb);
1939                         }
1940
1941                         if (partially_consumed)
1942                                 chunk = len;
1943
1944                         err = skb_copy_datagram_msg(skb, rxm->offset,
1945                                                     msg, chunk);
1946                         if (err < 0)
1947                                 goto recv_end;
1948
1949                         if (is_peek)
1950                                 goto leave_on_list;
1951
1952                         if (partially_consumed) {
1953                                 rxm->offset += chunk;
1954                                 rxm->full_len -= chunk;
1955                                 goto leave_on_list;
1956                         }
1957                 }
1958
1959                 decrypted += chunk;
1960                 len -= chunk;
1961
1962                 __skb_unlink(skb, &ctx->rx_list);
1963                 consume_skb(skb);
1964
1965                 /* Return full control message to userspace before trying
1966                  * to parse another message type
1967                  */
1968                 msg->msg_flags |= MSG_EOR;
1969                 if (control != TLS_RECORD_TYPE_DATA)
1970                         break;
1971         }
1972
1973 recv_end:
1974         if (async) {
1975                 int ret, pending;
1976
1977                 /* Wait for all previously submitted records to be decrypted */
1978                 spin_lock_bh(&ctx->decrypt_compl_lock);
1979                 reinit_completion(&ctx->async_wait.completion);
1980                 pending = atomic_read(&ctx->decrypt_pending);
1981                 spin_unlock_bh(&ctx->decrypt_compl_lock);
1982                 if (pending) {
1983                         ret = crypto_wait_req(-EINPROGRESS, &ctx->async_wait);
1984                         if (ret) {
1985                                 if (err >= 0 || err == -EINPROGRESS)
1986                                         err = ret;
1987                                 decrypted = 0;
1988                                 goto end;
1989                         }
1990                 }
1991
1992                 /* Drain records from the rx_list & copy if required */
1993                 if (is_peek || is_kvec)
1994                         err = process_rx_list(ctx, msg, &control, copied,
1995                                               decrypted, false, is_peek);
1996                 else
1997                         err = process_rx_list(ctx, msg, &control, 0,
1998                                               decrypted, true, is_peek);
1999                 decrypted = max(err, 0);
2000         }
2001
2002         copied += decrypted;
2003
2004 end:
2005         tls_rx_reader_unlock(sk, ctx);
2006         if (psock)
2007                 sk_psock_put(sk, psock);
2008         return copied ? : err;
2009 }
2010
2011 ssize_t tls_sw_splice_read(struct socket *sock,  loff_t *ppos,
2012                            struct pipe_inode_info *pipe,
2013                            size_t len, unsigned int flags)
2014 {
2015         struct tls_context *tls_ctx = tls_get_ctx(sock->sk);
2016         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2017         struct strp_msg *rxm = NULL;
2018         struct sock *sk = sock->sk;
2019         struct tls_msg *tlm;
2020         struct sk_buff *skb;
2021         ssize_t copied = 0;
2022         bool from_queue;
2023         int err = 0;
2024         long timeo;
2025         int chunk;
2026
2027         timeo = tls_rx_reader_lock(sk, ctx, flags & SPLICE_F_NONBLOCK);
2028         if (timeo < 0)
2029                 return timeo;
2030
2031         from_queue = !skb_queue_empty(&ctx->rx_list);
2032         if (from_queue) {
2033                 skb = __skb_dequeue(&ctx->rx_list);
2034         } else {
2035                 struct tls_decrypt_arg darg = {};
2036
2037                 err = tls_rx_rec_wait(sk, NULL, flags & SPLICE_F_NONBLOCK,
2038                                       timeo);
2039                 if (err <= 0)
2040                         goto splice_read_end;
2041
2042                 skb = ctx->recv_pkt;
2043
2044                 err = decrypt_skb_update(sk, skb, NULL, &darg);
2045                 if (err < 0) {
2046                         tls_err_abort(sk, -EBADMSG);
2047                         goto splice_read_end;
2048                 }
2049         }
2050
2051         rxm = strp_msg(skb);
2052         tlm = tls_msg(skb);
2053
2054         /* splice does not support reading control messages */
2055         if (tlm->control != TLS_RECORD_TYPE_DATA) {
2056                 err = -EINVAL;
2057                 goto splice_read_end;
2058         }
2059
2060         chunk = min_t(unsigned int, rxm->full_len, len);
2061         copied = skb_splice_bits(skb, sk, rxm->offset, pipe, chunk, flags);
2062         if (copied < 0)
2063                 goto splice_read_end;
2064
2065         if (!from_queue) {
2066                 ctx->recv_pkt = NULL;
2067                 __strp_unpause(&ctx->strp);
2068         }
2069         if (chunk < rxm->full_len) {
2070                 __skb_queue_head(&ctx->rx_list, skb);
2071                 rxm->offset += len;
2072                 rxm->full_len -= len;
2073         } else {
2074                 consume_skb(skb);
2075         }
2076
2077 splice_read_end:
2078         tls_rx_reader_unlock(sk, ctx);
2079         return copied ? : err;
2080 }
2081
2082 bool tls_sw_sock_is_readable(struct sock *sk)
2083 {
2084         struct tls_context *tls_ctx = tls_get_ctx(sk);
2085         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2086         bool ingress_empty = true;
2087         struct sk_psock *psock;
2088
2089         rcu_read_lock();
2090         psock = sk_psock(sk);
2091         if (psock)
2092                 ingress_empty = list_empty(&psock->ingress_msg);
2093         rcu_read_unlock();
2094
2095         return !ingress_empty || ctx->recv_pkt ||
2096                 !skb_queue_empty(&ctx->rx_list);
2097 }
2098
2099 static int tls_read_size(struct strparser *strp, struct sk_buff *skb)
2100 {
2101         struct tls_context *tls_ctx = tls_get_ctx(strp->sk);
2102         struct tls_prot_info *prot = &tls_ctx->prot_info;
2103         char header[TLS_HEADER_SIZE + MAX_IV_SIZE];
2104         struct strp_msg *rxm = strp_msg(skb);
2105         struct tls_msg *tlm = tls_msg(skb);
2106         size_t cipher_overhead;
2107         size_t data_len = 0;
2108         int ret;
2109
2110         /* Verify that we have a full TLS header, or wait for more data */
2111         if (rxm->offset + prot->prepend_size > skb->len)
2112                 return 0;
2113
2114         /* Sanity-check size of on-stack buffer. */
2115         if (WARN_ON(prot->prepend_size > sizeof(header))) {
2116                 ret = -EINVAL;
2117                 goto read_failure;
2118         }
2119
2120         /* Linearize header to local buffer */
2121         ret = skb_copy_bits(skb, rxm->offset, header, prot->prepend_size);
2122         if (ret < 0)
2123                 goto read_failure;
2124
2125         tlm->decrypted = 0;
2126         tlm->control = header[0];
2127
2128         data_len = ((header[4] & 0xFF) | (header[3] << 8));
2129
2130         cipher_overhead = prot->tag_size;
2131         if (prot->version != TLS_1_3_VERSION &&
2132             prot->cipher_type != TLS_CIPHER_CHACHA20_POLY1305)
2133                 cipher_overhead += prot->iv_size;
2134
2135         if (data_len > TLS_MAX_PAYLOAD_SIZE + cipher_overhead +
2136             prot->tail_size) {
2137                 ret = -EMSGSIZE;
2138                 goto read_failure;
2139         }
2140         if (data_len < cipher_overhead) {
2141                 ret = -EBADMSG;
2142                 goto read_failure;
2143         }
2144
2145         /* Note that both TLS1.3 and TLS1.2 use TLS_1_2 version here */
2146         if (header[1] != TLS_1_2_VERSION_MINOR ||
2147             header[2] != TLS_1_2_VERSION_MAJOR) {
2148                 ret = -EINVAL;
2149                 goto read_failure;
2150         }
2151
2152         tls_device_rx_resync_new_rec(strp->sk, data_len + TLS_HEADER_SIZE,
2153                                      TCP_SKB_CB(skb)->seq + rxm->offset);
2154         return data_len + TLS_HEADER_SIZE;
2155
2156 read_failure:
2157         tls_err_abort(strp->sk, ret);
2158
2159         return ret;
2160 }
2161
2162 static void tls_queue(struct strparser *strp, struct sk_buff *skb)
2163 {
2164         struct tls_context *tls_ctx = tls_get_ctx(strp->sk);
2165         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2166
2167         ctx->recv_pkt = skb;
2168         strp_pause(strp);
2169
2170         ctx->saved_data_ready(strp->sk);
2171 }
2172
2173 static void tls_data_ready(struct sock *sk)
2174 {
2175         struct tls_context *tls_ctx = tls_get_ctx(sk);
2176         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2177         struct sk_psock *psock;
2178
2179         strp_data_ready(&ctx->strp);
2180
2181         psock = sk_psock_get(sk);
2182         if (psock) {
2183                 if (!list_empty(&psock->ingress_msg))
2184                         ctx->saved_data_ready(sk);
2185                 sk_psock_put(sk, psock);
2186         }
2187 }
2188
2189 void tls_sw_cancel_work_tx(struct tls_context *tls_ctx)
2190 {
2191         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
2192
2193         set_bit(BIT_TX_CLOSING, &ctx->tx_bitmask);
2194         set_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask);
2195         cancel_delayed_work_sync(&ctx->tx_work.work);
2196 }
2197
2198 void tls_sw_release_resources_tx(struct sock *sk)
2199 {
2200         struct tls_context *tls_ctx = tls_get_ctx(sk);
2201         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
2202         struct tls_rec *rec, *tmp;
2203         int pending;
2204
2205         /* Wait for any pending async encryptions to complete */
2206         spin_lock_bh(&ctx->encrypt_compl_lock);
2207         ctx->async_notify = true;
2208         pending = atomic_read(&ctx->encrypt_pending);
2209         spin_unlock_bh(&ctx->encrypt_compl_lock);
2210
2211         if (pending)
2212                 crypto_wait_req(-EINPROGRESS, &ctx->async_wait);
2213
2214         tls_tx_records(sk, -1);
2215
2216         /* Free up un-sent records in tx_list. First, free
2217          * the partially sent record if any at head of tx_list.
2218          */
2219         if (tls_ctx->partially_sent_record) {
2220                 tls_free_partial_record(sk, tls_ctx);
2221                 rec = list_first_entry(&ctx->tx_list,
2222                                        struct tls_rec, list);
2223                 list_del(&rec->list);
2224                 sk_msg_free(sk, &rec->msg_plaintext);
2225                 kfree(rec);
2226         }
2227
2228         list_for_each_entry_safe(rec, tmp, &ctx->tx_list, list) {
2229                 list_del(&rec->list);
2230                 sk_msg_free(sk, &rec->msg_encrypted);
2231                 sk_msg_free(sk, &rec->msg_plaintext);
2232                 kfree(rec);
2233         }
2234
2235         crypto_free_aead(ctx->aead_send);
2236         tls_free_open_rec(sk);
2237 }
2238
2239 void tls_sw_free_ctx_tx(struct tls_context *tls_ctx)
2240 {
2241         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
2242
2243         kfree(ctx);
2244 }
2245
2246 void tls_sw_release_resources_rx(struct sock *sk)
2247 {
2248         struct tls_context *tls_ctx = tls_get_ctx(sk);
2249         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2250
2251         kfree(tls_ctx->rx.rec_seq);
2252         kfree(tls_ctx->rx.iv);
2253
2254         if (ctx->aead_recv) {
2255                 kfree_skb(ctx->recv_pkt);
2256                 ctx->recv_pkt = NULL;
2257                 __skb_queue_purge(&ctx->rx_list);
2258                 crypto_free_aead(ctx->aead_recv);
2259                 strp_stop(&ctx->strp);
2260                 /* If tls_sw_strparser_arm() was not called (cleanup paths)
2261                  * we still want to strp_stop(), but sk->sk_data_ready was
2262                  * never swapped.
2263                  */
2264                 if (ctx->saved_data_ready) {
2265                         write_lock_bh(&sk->sk_callback_lock);
2266                         sk->sk_data_ready = ctx->saved_data_ready;
2267                         write_unlock_bh(&sk->sk_callback_lock);
2268                 }
2269         }
2270 }
2271
2272 void tls_sw_strparser_done(struct tls_context *tls_ctx)
2273 {
2274         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2275
2276         strp_done(&ctx->strp);
2277 }
2278
2279 void tls_sw_free_ctx_rx(struct tls_context *tls_ctx)
2280 {
2281         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2282
2283         kfree(ctx);
2284 }
2285
2286 void tls_sw_free_resources_rx(struct sock *sk)
2287 {
2288         struct tls_context *tls_ctx = tls_get_ctx(sk);
2289
2290         tls_sw_release_resources_rx(sk);
2291         tls_sw_free_ctx_rx(tls_ctx);
2292 }
2293
2294 /* The work handler to transmitt the encrypted records in tx_list */
2295 static void tx_work_handler(struct work_struct *work)
2296 {
2297         struct delayed_work *delayed_work = to_delayed_work(work);
2298         struct tx_work *tx_work = container_of(delayed_work,
2299                                                struct tx_work, work);
2300         struct sock *sk = tx_work->sk;
2301         struct tls_context *tls_ctx = tls_get_ctx(sk);
2302         struct tls_sw_context_tx *ctx;
2303
2304         if (unlikely(!tls_ctx))
2305                 return;
2306
2307         ctx = tls_sw_ctx_tx(tls_ctx);
2308         if (test_bit(BIT_TX_CLOSING, &ctx->tx_bitmask))
2309                 return;
2310
2311         if (!test_and_clear_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask))
2312                 return;
2313         mutex_lock(&tls_ctx->tx_lock);
2314         lock_sock(sk);
2315         tls_tx_records(sk, -1);
2316         release_sock(sk);
2317         mutex_unlock(&tls_ctx->tx_lock);
2318 }
2319
2320 static bool tls_is_tx_ready(struct tls_sw_context_tx *ctx)
2321 {
2322         struct tls_rec *rec;
2323
2324         rec = list_first_entry(&ctx->tx_list, struct tls_rec, list);
2325         if (!rec)
2326                 return false;
2327
2328         return READ_ONCE(rec->tx_ready);
2329 }
2330
2331 void tls_sw_write_space(struct sock *sk, struct tls_context *ctx)
2332 {
2333         struct tls_sw_context_tx *tx_ctx = tls_sw_ctx_tx(ctx);
2334
2335         /* Schedule the transmission if tx list is ready */
2336         if (tls_is_tx_ready(tx_ctx) &&
2337             !test_and_set_bit(BIT_TX_SCHEDULED, &tx_ctx->tx_bitmask))
2338                 schedule_delayed_work(&tx_ctx->tx_work.work, 0);
2339 }
2340
2341 void tls_sw_strparser_arm(struct sock *sk, struct tls_context *tls_ctx)
2342 {
2343         struct tls_sw_context_rx *rx_ctx = tls_sw_ctx_rx(tls_ctx);
2344
2345         write_lock_bh(&sk->sk_callback_lock);
2346         rx_ctx->saved_data_ready = sk->sk_data_ready;
2347         sk->sk_data_ready = tls_data_ready;
2348         write_unlock_bh(&sk->sk_callback_lock);
2349
2350         strp_check_rcv(&rx_ctx->strp);
2351 }
2352
2353 void tls_update_rx_zc_capable(struct tls_context *tls_ctx)
2354 {
2355         struct tls_sw_context_rx *rx_ctx = tls_sw_ctx_rx(tls_ctx);
2356
2357         rx_ctx->zc_capable = tls_ctx->rx_no_pad ||
2358                 tls_ctx->prot_info.version != TLS_1_3_VERSION;
2359 }
2360
2361 int tls_set_sw_offload(struct sock *sk, struct tls_context *ctx, int tx)
2362 {
2363         struct tls_context *tls_ctx = tls_get_ctx(sk);
2364         struct tls_prot_info *prot = &tls_ctx->prot_info;
2365         struct tls_crypto_info *crypto_info;
2366         struct tls_sw_context_tx *sw_ctx_tx = NULL;
2367         struct tls_sw_context_rx *sw_ctx_rx = NULL;
2368         struct cipher_context *cctx;
2369         struct crypto_aead **aead;
2370         struct strp_callbacks cb;
2371         u16 nonce_size, tag_size, iv_size, rec_seq_size, salt_size;
2372         struct crypto_tfm *tfm;
2373         char *iv, *rec_seq, *key, *salt, *cipher_name;
2374         size_t keysize;
2375         int rc = 0;
2376
2377         if (!ctx) {
2378                 rc = -EINVAL;
2379                 goto out;
2380         }
2381
2382         if (tx) {
2383                 if (!ctx->priv_ctx_tx) {
2384                         sw_ctx_tx = kzalloc(sizeof(*sw_ctx_tx), GFP_KERNEL);
2385                         if (!sw_ctx_tx) {
2386                                 rc = -ENOMEM;
2387                                 goto out;
2388                         }
2389                         ctx->priv_ctx_tx = sw_ctx_tx;
2390                 } else {
2391                         sw_ctx_tx =
2392                                 (struct tls_sw_context_tx *)ctx->priv_ctx_tx;
2393                 }
2394         } else {
2395                 if (!ctx->priv_ctx_rx) {
2396                         sw_ctx_rx = kzalloc(sizeof(*sw_ctx_rx), GFP_KERNEL);
2397                         if (!sw_ctx_rx) {
2398                                 rc = -ENOMEM;
2399                                 goto out;
2400                         }
2401                         ctx->priv_ctx_rx = sw_ctx_rx;
2402                 } else {
2403                         sw_ctx_rx =
2404                                 (struct tls_sw_context_rx *)ctx->priv_ctx_rx;
2405                 }
2406         }
2407
2408         if (tx) {
2409                 crypto_init_wait(&sw_ctx_tx->async_wait);
2410                 spin_lock_init(&sw_ctx_tx->encrypt_compl_lock);
2411                 crypto_info = &ctx->crypto_send.info;
2412                 cctx = &ctx->tx;
2413                 aead = &sw_ctx_tx->aead_send;
2414                 INIT_LIST_HEAD(&sw_ctx_tx->tx_list);
2415                 INIT_DELAYED_WORK(&sw_ctx_tx->tx_work.work, tx_work_handler);
2416                 sw_ctx_tx->tx_work.sk = sk;
2417         } else {
2418                 crypto_init_wait(&sw_ctx_rx->async_wait);
2419                 spin_lock_init(&sw_ctx_rx->decrypt_compl_lock);
2420                 init_waitqueue_head(&sw_ctx_rx->wq);
2421                 crypto_info = &ctx->crypto_recv.info;
2422                 cctx = &ctx->rx;
2423                 skb_queue_head_init(&sw_ctx_rx->rx_list);
2424                 aead = &sw_ctx_rx->aead_recv;
2425         }
2426
2427         switch (crypto_info->cipher_type) {
2428         case TLS_CIPHER_AES_GCM_128: {
2429                 struct tls12_crypto_info_aes_gcm_128 *gcm_128_info;
2430
2431                 gcm_128_info = (void *)crypto_info;
2432                 nonce_size = TLS_CIPHER_AES_GCM_128_IV_SIZE;
2433                 tag_size = TLS_CIPHER_AES_GCM_128_TAG_SIZE;
2434                 iv_size = TLS_CIPHER_AES_GCM_128_IV_SIZE;
2435                 iv = gcm_128_info->iv;
2436                 rec_seq_size = TLS_CIPHER_AES_GCM_128_REC_SEQ_SIZE;
2437                 rec_seq = gcm_128_info->rec_seq;
2438                 keysize = TLS_CIPHER_AES_GCM_128_KEY_SIZE;
2439                 key = gcm_128_info->key;
2440                 salt = gcm_128_info->salt;
2441                 salt_size = TLS_CIPHER_AES_GCM_128_SALT_SIZE;
2442                 cipher_name = "gcm(aes)";
2443                 break;
2444         }
2445         case TLS_CIPHER_AES_GCM_256: {
2446                 struct tls12_crypto_info_aes_gcm_256 *gcm_256_info;
2447
2448                 gcm_256_info = (void *)crypto_info;
2449                 nonce_size = TLS_CIPHER_AES_GCM_256_IV_SIZE;
2450                 tag_size = TLS_CIPHER_AES_GCM_256_TAG_SIZE;
2451                 iv_size = TLS_CIPHER_AES_GCM_256_IV_SIZE;
2452                 iv = gcm_256_info->iv;
2453                 rec_seq_size = TLS_CIPHER_AES_GCM_256_REC_SEQ_SIZE;
2454                 rec_seq = gcm_256_info->rec_seq;
2455                 keysize = TLS_CIPHER_AES_GCM_256_KEY_SIZE;
2456                 key = gcm_256_info->key;
2457                 salt = gcm_256_info->salt;
2458                 salt_size = TLS_CIPHER_AES_GCM_256_SALT_SIZE;
2459                 cipher_name = "gcm(aes)";
2460                 break;
2461         }
2462         case TLS_CIPHER_AES_CCM_128: {
2463                 struct tls12_crypto_info_aes_ccm_128 *ccm_128_info;
2464
2465                 ccm_128_info = (void *)crypto_info;
2466                 nonce_size = TLS_CIPHER_AES_CCM_128_IV_SIZE;
2467                 tag_size = TLS_CIPHER_AES_CCM_128_TAG_SIZE;
2468                 iv_size = TLS_CIPHER_AES_CCM_128_IV_SIZE;
2469                 iv = ccm_128_info->iv;
2470                 rec_seq_size = TLS_CIPHER_AES_CCM_128_REC_SEQ_SIZE;
2471                 rec_seq = ccm_128_info->rec_seq;
2472                 keysize = TLS_CIPHER_AES_CCM_128_KEY_SIZE;
2473                 key = ccm_128_info->key;
2474                 salt = ccm_128_info->salt;
2475                 salt_size = TLS_CIPHER_AES_CCM_128_SALT_SIZE;
2476                 cipher_name = "ccm(aes)";
2477                 break;
2478         }
2479         case TLS_CIPHER_CHACHA20_POLY1305: {
2480                 struct tls12_crypto_info_chacha20_poly1305 *chacha20_poly1305_info;
2481
2482                 chacha20_poly1305_info = (void *)crypto_info;
2483                 nonce_size = 0;
2484                 tag_size = TLS_CIPHER_CHACHA20_POLY1305_TAG_SIZE;
2485                 iv_size = TLS_CIPHER_CHACHA20_POLY1305_IV_SIZE;
2486                 iv = chacha20_poly1305_info->iv;
2487                 rec_seq_size = TLS_CIPHER_CHACHA20_POLY1305_REC_SEQ_SIZE;
2488                 rec_seq = chacha20_poly1305_info->rec_seq;
2489                 keysize = TLS_CIPHER_CHACHA20_POLY1305_KEY_SIZE;
2490                 key = chacha20_poly1305_info->key;
2491                 salt = chacha20_poly1305_info->salt;
2492                 salt_size = TLS_CIPHER_CHACHA20_POLY1305_SALT_SIZE;
2493                 cipher_name = "rfc7539(chacha20,poly1305)";
2494                 break;
2495         }
2496         case TLS_CIPHER_SM4_GCM: {
2497                 struct tls12_crypto_info_sm4_gcm *sm4_gcm_info;
2498
2499                 sm4_gcm_info = (void *)crypto_info;
2500                 nonce_size = TLS_CIPHER_SM4_GCM_IV_SIZE;
2501                 tag_size = TLS_CIPHER_SM4_GCM_TAG_SIZE;
2502                 iv_size = TLS_CIPHER_SM4_GCM_IV_SIZE;
2503                 iv = sm4_gcm_info->iv;
2504                 rec_seq_size = TLS_CIPHER_SM4_GCM_REC_SEQ_SIZE;
2505                 rec_seq = sm4_gcm_info->rec_seq;
2506                 keysize = TLS_CIPHER_SM4_GCM_KEY_SIZE;
2507                 key = sm4_gcm_info->key;
2508                 salt = sm4_gcm_info->salt;
2509                 salt_size = TLS_CIPHER_SM4_GCM_SALT_SIZE;
2510                 cipher_name = "gcm(sm4)";
2511                 break;
2512         }
2513         case TLS_CIPHER_SM4_CCM: {
2514                 struct tls12_crypto_info_sm4_ccm *sm4_ccm_info;
2515
2516                 sm4_ccm_info = (void *)crypto_info;
2517                 nonce_size = TLS_CIPHER_SM4_CCM_IV_SIZE;
2518                 tag_size = TLS_CIPHER_SM4_CCM_TAG_SIZE;
2519                 iv_size = TLS_CIPHER_SM4_CCM_IV_SIZE;
2520                 iv = sm4_ccm_info->iv;
2521                 rec_seq_size = TLS_CIPHER_SM4_CCM_REC_SEQ_SIZE;
2522                 rec_seq = sm4_ccm_info->rec_seq;
2523                 keysize = TLS_CIPHER_SM4_CCM_KEY_SIZE;
2524                 key = sm4_ccm_info->key;
2525                 salt = sm4_ccm_info->salt;
2526                 salt_size = TLS_CIPHER_SM4_CCM_SALT_SIZE;
2527                 cipher_name = "ccm(sm4)";
2528                 break;
2529         }
2530         default:
2531                 rc = -EINVAL;
2532                 goto free_priv;
2533         }
2534
2535         if (crypto_info->version == TLS_1_3_VERSION) {
2536                 nonce_size = 0;
2537                 prot->aad_size = TLS_HEADER_SIZE;
2538                 prot->tail_size = 1;
2539         } else {
2540                 prot->aad_size = TLS_AAD_SPACE_SIZE;
2541                 prot->tail_size = 0;
2542         }
2543
2544         /* Sanity-check the sizes for stack allocations. */
2545         if (iv_size > MAX_IV_SIZE || nonce_size > MAX_IV_SIZE ||
2546             rec_seq_size > TLS_MAX_REC_SEQ_SIZE || tag_size != TLS_TAG_SIZE ||
2547             prot->aad_size > TLS_MAX_AAD_SIZE) {
2548                 rc = -EINVAL;
2549                 goto free_priv;
2550         }
2551
2552         prot->version = crypto_info->version;
2553         prot->cipher_type = crypto_info->cipher_type;
2554         prot->prepend_size = TLS_HEADER_SIZE + nonce_size;
2555         prot->tag_size = tag_size;
2556         prot->overhead_size = prot->prepend_size +
2557                               prot->tag_size + prot->tail_size;
2558         prot->iv_size = iv_size;
2559         prot->salt_size = salt_size;
2560         cctx->iv = kmalloc(iv_size + salt_size, GFP_KERNEL);
2561         if (!cctx->iv) {
2562                 rc = -ENOMEM;
2563                 goto free_priv;
2564         }
2565         /* Note: 128 & 256 bit salt are the same size */
2566         prot->rec_seq_size = rec_seq_size;
2567         memcpy(cctx->iv, salt, salt_size);
2568         memcpy(cctx->iv + salt_size, iv, iv_size);
2569         cctx->rec_seq = kmemdup(rec_seq, rec_seq_size, GFP_KERNEL);
2570         if (!cctx->rec_seq) {
2571                 rc = -ENOMEM;
2572                 goto free_iv;
2573         }
2574
2575         if (!*aead) {
2576                 *aead = crypto_alloc_aead(cipher_name, 0, 0);
2577                 if (IS_ERR(*aead)) {
2578                         rc = PTR_ERR(*aead);
2579                         *aead = NULL;
2580                         goto free_rec_seq;
2581                 }
2582         }
2583
2584         ctx->push_pending_record = tls_sw_push_pending_record;
2585
2586         rc = crypto_aead_setkey(*aead, key, keysize);
2587
2588         if (rc)
2589                 goto free_aead;
2590
2591         rc = crypto_aead_setauthsize(*aead, prot->tag_size);
2592         if (rc)
2593                 goto free_aead;
2594
2595         if (sw_ctx_rx) {
2596                 tfm = crypto_aead_tfm(sw_ctx_rx->aead_recv);
2597
2598                 tls_update_rx_zc_capable(ctx);
2599                 sw_ctx_rx->async_capable =
2600                         crypto_info->version != TLS_1_3_VERSION &&
2601                         !!(tfm->__crt_alg->cra_flags & CRYPTO_ALG_ASYNC);
2602
2603                 /* Set up strparser */
2604                 memset(&cb, 0, sizeof(cb));
2605                 cb.rcv_msg = tls_queue;
2606                 cb.parse_msg = tls_read_size;
2607
2608                 strp_init(&sw_ctx_rx->strp, sk, &cb);
2609         }
2610
2611         goto out;
2612
2613 free_aead:
2614         crypto_free_aead(*aead);
2615         *aead = NULL;
2616 free_rec_seq:
2617         kfree(cctx->rec_seq);
2618         cctx->rec_seq = NULL;
2619 free_iv:
2620         kfree(cctx->iv);
2621         cctx->iv = NULL;
2622 free_priv:
2623         if (tx) {
2624                 kfree(ctx->priv_ctx_tx);
2625                 ctx->priv_ctx_tx = NULL;
2626         } else {
2627                 kfree(ctx->priv_ctx_rx);
2628                 ctx->priv_ctx_rx = NULL;
2629         }
2630 out:
2631         return rc;
2632 }