net/tls/tls_sw.c

   1 /*
   2  * Copyright (c) 2016-2017, Mellanox Technologies. All rights reserved.
   3  * Copyright (c) 2016-2017, Dave Watson <[email protected]>. All rights reserved.
   4  * Copyright (c) 2016-2017, Lance Chao <[email protected]>. All rights reserved.
   5  * Copyright (c) 2016, Fridolin Pokorny <[email protected]>. All rights reserved.
   6  * Copyright (c) 2016, Nikos Mavrogiannopoulos <[email protected]>. 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/sched/signal.h>
  39 #include <linux/module.h>
  40 #include <linux/splice.h>
  41 #include <crypto/aead.h>
  42
  43 #include <net/strparser.h>
  44 #include <net/tls.h>
  45
  46 static int __skb_nsg(struct sk_buff *skb, int offset, int len,
  47                      unsigned int recursion_level)
  48 {
  49         int start = skb_headlen(skb);
  50         int i, chunk = start - offset;
  51         struct sk_buff *frag_iter;
  52         int elt = 0;
  53
  54         if (unlikely(recursion_level >= 24))
  55                 return -EMSGSIZE;
  56
  57         if (chunk > 0) {
  58                 if (chunk > len)
  59                         chunk = len;
  60                 elt++;
  61                 len -= chunk;
  62                 if (len == 0)
  63                         return elt;
  64                 offset += chunk;
  65         }
  66
  67         for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
  68                 int end;
  69
  70                 WARN_ON(start > offset + len);
  71
  72                 end = start + skb_frag_size(&skb_shinfo(skb)->frags[i]);
  73                 chunk = end - offset;
  74                 if (chunk > 0) {
  75                         if (chunk > len)
  76                                 chunk = len;
  77                         elt++;
  78                         len -= chunk;
  79                         if (len == 0)
  80                                 return elt;
  81                         offset += chunk;
  82                 }
  83                 start = end;
  84         }
  85
  86         if (unlikely(skb_has_frag_list(skb))) {
  87                 skb_walk_frags(skb, frag_iter) {
  88                         int end, ret;
  89
  90                         WARN_ON(start > offset + len);
  91
  92                         end = start + frag_iter->len;
  93                         chunk = end - offset;
  94                         if (chunk > 0) {
  95                                 if (chunk > len)
  96                                         chunk = len;
  97                                 ret = __skb_nsg(frag_iter, offset - start, chunk,
  98                                                 recursion_level + 1);
  99                                 if (unlikely(ret < 0))
 100                                         return ret;
 101                                 elt += ret;
 102                                 len -= chunk;
 103                                 if (len == 0)
 104                                         return elt;
 105                                 offset += chunk;
 106                         }
 107                         start = end;
 108                 }
 109         }
 110         BUG_ON(len);
 111         return elt;
 112 }
 113
 114 /* Return the number of scatterlist elements required to completely map the
 115  * skb, or -EMSGSIZE if the recursion depth is exceeded.
 116  */
 117 static int skb_nsg(struct sk_buff *skb, int offset, int len)
 118 {
 119         return __skb_nsg(skb, offset, len, 0);
 120 }
 121
 122 static int padding_length(struct tls_sw_context_rx *ctx,
 123                           struct tls_prot_info *prot, struct sk_buff *skb)
 124 {
 125         struct strp_msg *rxm = strp_msg(skb);
 126         int sub = 0;
 127
 128         /* Determine zero-padding length */
 129         if (prot->version == TLS_1_3_VERSION) {
 130                 char content_type = 0;
 131                 int err;
 132                 int back = 17;
 133
 134                 while (content_type == 0) {
 135                         if (back > rxm->full_len - prot->prepend_size)
 136                                 return -EBADMSG;
 137                         err = skb_copy_bits(skb,
 138                                             rxm->offset + rxm->full_len - back,
 139                                             &content_type, 1);
 140                         if (err)
 141                                 return err;
 142                         if (content_type)
 143                                 break;
 144                         sub++;
 145                         back++;
 146                 }
 147                 ctx->control = content_type;
 148         }
 149         return sub;
 150 }
 151
 152 static void tls_decrypt_done(struct crypto_async_request *req, int err)
 153 {
 154         struct aead_request *aead_req = (struct aead_request *)req;
 155         struct scatterlist *sgout = aead_req->dst;
 156         struct scatterlist *sgin = aead_req->src;
 157         struct tls_sw_context_rx *ctx;
 158         struct tls_context *tls_ctx;
 159         struct tls_prot_info *prot;
 160         struct scatterlist *sg;
 161         struct sk_buff *skb;
 162         unsigned int pages;
 163         int pending;
 164
 165         skb = (struct sk_buff *)req->data;
 166         tls_ctx = tls_get_ctx(skb->sk);
 167         ctx = tls_sw_ctx_rx(tls_ctx);
 168         prot = &tls_ctx->prot_info;
 169
 170         /* Propagate if there was an err */
 171         if (err) {
 172                 if (err == -EBADMSG)
 173                         TLS_INC_STATS(sock_net(skb->sk),
 174                                       LINUX_MIB_TLSDECRYPTERROR);
 175                 ctx->async_wait.err = err;
 176                 tls_err_abort(skb->sk, err);
 177         } else {
 178                 struct strp_msg *rxm = strp_msg(skb);
 179                 int pad;
 180
 181                 pad = padding_length(ctx, prot, skb);
 182                 if (pad < 0) {
 183                         ctx->async_wait.err = pad;
 184                         tls_err_abort(skb->sk, pad);
 185                 } else {
 186                         rxm->full_len -= pad;
 187                         rxm->offset += prot->prepend_size;
 188                         rxm->full_len -= prot->overhead_size;
 189                 }
 190         }
 191
 192         /* After using skb->sk to propagate sk through crypto async callback
 193          * we need to NULL it again.
 194          */
 195         skb->sk = NULL;
 196
 197
 198         /* Free the destination pages if skb was not decrypted inplace */
 199         if (sgout != sgin) {
 200                 /* Skip the first S/G entry as it points to AAD */
 201                 for_each_sg(sg_next(sgout), sg, UINT_MAX, pages) {
 202                         if (!sg)
 203                                 break;
 204                         put_page(sg_page(sg));
 205                 }
 206         }
 207
 208         kfree(aead_req);
 209
 210         spin_lock_bh(&ctx->decrypt_compl_lock);
 211         pending = atomic_dec_return(&ctx->decrypt_pending);
 212
 213         if (!pending && ctx->async_notify)
 214                 complete(&ctx->async_wait.completion);
 215         spin_unlock_bh(&ctx->decrypt_compl_lock);
 216 }
 217
 218 static int tls_do_decryption(struct sock *sk,
 219                              struct sk_buff *skb,
 220                              struct scatterlist *sgin,
 221                              struct scatterlist *sgout,
 222                              char *iv_recv,
 223                              size_t data_len,
 224                              struct aead_request *aead_req,
 225                              bool async)
 226 {
 227         struct tls_context *tls_ctx = tls_get_ctx(sk);
 228         struct tls_prot_info *prot = &tls_ctx->prot_info;
 229         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
 230         int ret;
 231
 232         aead_request_set_tfm(aead_req, ctx->aead_recv);
 233         aead_request_set_ad(aead_req, prot->aad_size);
 234         aead_request_set_crypt(aead_req, sgin, sgout,
 235                                data_len + prot->tag_size,
 236                                (u8 *)iv_recv);
 237
 238         if (async) {
 239                 /* Using skb->sk to push sk through to crypto async callback
 240                  * handler. This allows propagating errors up to the socket
 241                  * if needed. It _must_ be cleared in the async handler
 242                  * before consume_skb is called. We _know_ skb->sk is NULL
 243                  * because it is a clone from strparser.
 244                  */
 245                 skb->sk = sk;
 246                 aead_request_set_callback(aead_req,
 247                                           CRYPTO_TFM_REQ_MAY_BACKLOG,
 248                                           tls_decrypt_done, skb);
 249                 atomic_inc(&ctx->decrypt_pending);
 250         } else {
 251                 aead_request_set_callback(aead_req,
 252                                           CRYPTO_TFM_REQ_MAY_BACKLOG,
 253                                           crypto_req_done, &ctx->async_wait);
 254         }
 255
 256         ret = crypto_aead_decrypt(aead_req);
 257         if (ret == -EINPROGRESS) {
 258                 if (async)
 259                         return ret;
 260
 261                 ret = crypto_wait_req(ret, &ctx->async_wait);
 262         }
 263
 264         if (async)
 265                 atomic_dec(&ctx->decrypt_pending);
 266
 267         return ret;
 268 }
 269
 270 static void tls_trim_both_msgs(struct sock *sk, int target_size)
 271 {
 272         struct tls_context *tls_ctx = tls_get_ctx(sk);
 273         struct tls_prot_info *prot = &tls_ctx->prot_info;
 274         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 275         struct tls_rec *rec = ctx->open_rec;
 276
 277         sk_msg_trim(sk, &rec->msg_plaintext, target_size);
 278         if (target_size > 0)
 279                 target_size += prot->overhead_size;
 280         sk_msg_trim(sk, &rec->msg_encrypted, target_size);
 281 }
 282
 283 static int tls_alloc_encrypted_msg(struct sock *sk, int len)
 284 {
 285         struct tls_context *tls_ctx = tls_get_ctx(sk);
 286         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 287         struct tls_rec *rec = ctx->open_rec;
 288         struct sk_msg *msg_en = &rec->msg_encrypted;
 289
 290         return sk_msg_alloc(sk, msg_en, len, 0);
 291 }
 292
 293 static int tls_clone_plaintext_msg(struct sock *sk, int required)
 294 {
 295         struct tls_context *tls_ctx = tls_get_ctx(sk);
 296         struct tls_prot_info *prot = &tls_ctx->prot_info;
 297         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 298         struct tls_rec *rec = ctx->open_rec;
 299         struct sk_msg *msg_pl = &rec->msg_plaintext;
 300         struct sk_msg *msg_en = &rec->msg_encrypted;
 301         int skip, len;
 302
 303         /* We add page references worth len bytes from encrypted sg
 304          * at the end of plaintext sg. It is guaranteed that msg_en
 305          * has enough required room (ensured by caller).
 306          */
 307         len = required - msg_pl->sg.size;
 308
 309         /* Skip initial bytes in msg_en's data to be able to use
 310          * same offset of both plain and encrypted data.
 311          */
 312         skip = prot->prepend_size + msg_pl->sg.size;
 313
 314         return sk_msg_clone(sk, msg_pl, msg_en, skip, len);
 315 }
 316
 317 static struct tls_rec *tls_get_rec(struct sock *sk)
 318 {
 319         struct tls_context *tls_ctx = tls_get_ctx(sk);
 320         struct tls_prot_info *prot = &tls_ctx->prot_info;
 321         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 322         struct sk_msg *msg_pl, *msg_en;
 323         struct tls_rec *rec;
 324         int mem_size;
 325
 326         mem_size = sizeof(struct tls_rec) + crypto_aead_reqsize(ctx->aead_send);
 327
 328         rec = kzalloc(mem_size, sk->sk_allocation);
 329         if (!rec)
 330                 return NULL;
 331
 332         msg_pl = &rec->msg_plaintext;
 333         msg_en = &rec->msg_encrypted;
 334
 335         sk_msg_init(msg_pl);
 336         sk_msg_init(msg_en);
 337
 338         sg_init_table(rec->sg_aead_in, 2);
 339         sg_set_buf(&rec->sg_aead_in[0], rec->aad_space, prot->aad_size);
 340         sg_unmark_end(&rec->sg_aead_in[1]);
 341
 342         sg_init_table(rec->sg_aead_out, 2);
 343         sg_set_buf(&rec->sg_aead_out[0], rec->aad_space, prot->aad_size);
 344         sg_unmark_end(&rec->sg_aead_out[1]);
 345
 346         return rec;
 347 }
 348
 349 static void tls_free_rec(struct sock *sk, struct tls_rec *rec)
 350 {
 351         sk_msg_free(sk, &rec->msg_encrypted);
 352         sk_msg_free(sk, &rec->msg_plaintext);
 353         kfree(rec);
 354 }
 355
 356 static void tls_free_open_rec(struct sock *sk)
 357 {
 358         struct tls_context *tls_ctx = tls_get_ctx(sk);
 359         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 360         struct tls_rec *rec = ctx->open_rec;
 361
 362         if (rec) {
 363                 tls_free_rec(sk, rec);
 364                 ctx->open_rec = NULL;
 365         }
 366 }
 367
 368 int tls_tx_records(struct sock *sk, int flags)
 369 {
 370         struct tls_context *tls_ctx = tls_get_ctx(sk);
 371         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 372         struct tls_rec *rec, *tmp;
 373         struct sk_msg *msg_en;
 374         int tx_flags, rc = 0;
 375
 376         if (tls_is_partially_sent_record(tls_ctx)) {
 377                 rec = list_first_entry(&ctx->tx_list,
 378                                        struct tls_rec, list);
 379
 380                 if (flags == -1)
 381                         tx_flags = rec->tx_flags;
 382                 else
 383                         tx_flags = flags;
 384
 385                 rc = tls_push_partial_record(sk, tls_ctx, tx_flags);
 386                 if (rc)
 387                         goto tx_err;
 388
 389                 /* Full record has been transmitted.
 390                  * Remove the head of tx_list
 391                  */
 392                 list_del(&rec->list);
 393                 sk_msg_free(sk, &rec->msg_plaintext);
 394                 kfree(rec);
 395         }
 396
 397         /* Tx all ready records */
 398         list_for_each_entry_safe(rec, tmp, &ctx->tx_list, list) {
 399                 if (READ_ONCE(rec->tx_ready)) {
 400                         if (flags == -1)
 401                                 tx_flags = rec->tx_flags;
 402                         else
 403                                 tx_flags = flags;
 404
 405                         msg_en = &rec->msg_encrypted;
 406                         rc = tls_push_sg(sk, tls_ctx,
 407                                          &msg_en->sg.data[msg_en->sg.curr],
 408                                          0, tx_flags);
 409                         if (rc)
 410                                 goto tx_err;
 411
 412                         list_del(&rec->list);
 413                         sk_msg_free(sk, &rec->msg_plaintext);
 414                         kfree(rec);
 415                 } else {
 416                         break;
 417                 }
 418         }
 419
 420 tx_err:
 421         if (rc < 0 && rc != -EAGAIN)
 422                 tls_err_abort(sk, EBADMSG);
 423
 424         return rc;
 425 }
 426
 427 static void tls_encrypt_done(struct crypto_async_request *req, int err)
 428 {
 429         struct aead_request *aead_req = (struct aead_request *)req;
 430         struct sock *sk = req->data;
 431         struct tls_context *tls_ctx = tls_get_ctx(sk);
 432         struct tls_prot_info *prot = &tls_ctx->prot_info;
 433         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 434         struct scatterlist *sge;
 435         struct sk_msg *msg_en;
 436         struct tls_rec *rec;
 437         bool ready = false;
 438         int pending;
 439
 440         rec = container_of(aead_req, struct tls_rec, aead_req);
 441         msg_en = &rec->msg_encrypted;
 442
 443         sge = sk_msg_elem(msg_en, msg_en->sg.curr);
 444         sge->offset -= prot->prepend_size;
 445         sge->length += prot->prepend_size;
 446
 447         /* Check if error is previously set on socket */
 448         if (err || sk->sk_err) {
 449                 rec = NULL;
 450
 451                 /* If err is already set on socket, return the same code */
 452                 if (sk->sk_err) {
 453                         ctx->async_wait.err = sk->sk_err;
 454                 } else {
 455                         ctx->async_wait.err = err;
 456                         tls_err_abort(sk, err);
 457                 }
 458         }
 459
 460         if (rec) {
 461                 struct tls_rec *first_rec;
 462
 463                 /* Mark the record as ready for transmission */
 464                 smp_store_mb(rec->tx_ready, true);
 465
 466                 /* If received record is at head of tx_list, schedule tx */
 467                 first_rec = list_first_entry(&ctx->tx_list,
 468                                              struct tls_rec, list);
 469                 if (rec == first_rec)
 470                         ready = true;
 471         }
 472
 473         spin_lock_bh(&ctx->encrypt_compl_lock);
 474         pending = atomic_dec_return(&ctx->encrypt_pending);
 475
 476         if (!pending && ctx->async_notify)
 477                 complete(&ctx->async_wait.completion);
 478         spin_unlock_bh(&ctx->encrypt_compl_lock);
 479
 480         if (!ready)
 481                 return;
 482
 483         /* Schedule the transmission */
 484         if (!test_and_set_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask))
 485                 schedule_delayed_work(&ctx->tx_work.work, 1);
 486 }
 487
 488 static int tls_do_encryption(struct sock *sk,
 489                              struct tls_context *tls_ctx,
 490                              struct tls_sw_context_tx *ctx,
 491                              struct aead_request *aead_req,
 492                              size_t data_len, u32 start)
 493 {
 494         struct tls_prot_info *prot = &tls_ctx->prot_info;
 495         struct tls_rec *rec = ctx->open_rec;
 496         struct sk_msg *msg_en = &rec->msg_encrypted;
 497         struct scatterlist *sge = sk_msg_elem(msg_en, start);
 498         int rc, iv_offset = 0;
 499
 500         /* For CCM based ciphers, first byte of IV is a constant */
 501         switch (prot->cipher_type) {
 502         case TLS_CIPHER_AES_CCM_128:
 503                 rec->iv_data[0] = TLS_AES_CCM_IV_B0_BYTE;
 504                 iv_offset = 1;
 505                 break;
 506         case TLS_CIPHER_SM4_CCM:
 507                 rec->iv_data[0] = TLS_SM4_CCM_IV_B0_BYTE;
 508                 iv_offset = 1;
 509                 break;
 510         }
 511
 512         memcpy(&rec->iv_data[iv_offset], tls_ctx->tx.iv,
 513                prot->iv_size + prot->salt_size);
 514
 515         xor_iv_with_seq(prot, rec->iv_data, tls_ctx->tx.rec_seq);
 516
 517         sge->offset += prot->prepend_size;
 518         sge->length -= prot->prepend_size;
 519
 520         msg_en->sg.curr = start;
 521
 522         aead_request_set_tfm(aead_req, ctx->aead_send);
 523         aead_request_set_ad(aead_req, prot->aad_size);
 524         aead_request_set_crypt(aead_req, rec->sg_aead_in,
 525                                rec->sg_aead_out,
 526                                data_len, rec->iv_data);
 527
 528         aead_request_set_callback(aead_req, CRYPTO_TFM_REQ_MAY_BACKLOG,
 529                                   tls_encrypt_done, sk);
 530
 531         /* Add the record in tx_list */
 532         list_add_tail((struct list_head *)&rec->list, &ctx->tx_list);
 533         atomic_inc(&ctx->encrypt_pending);
 534
 535         rc = crypto_aead_encrypt(aead_req);
 536         if (!rc || rc != -EINPROGRESS) {
 537                 atomic_dec(&ctx->encrypt_pending);
 538                 sge->offset -= prot->prepend_size;
 539                 sge->length += prot->prepend_size;
 540         }
 541
 542         if (!rc) {
 543                 WRITE_ONCE(rec->tx_ready, true);
 544         } else if (rc != -EINPROGRESS) {
 545                 list_del(&rec->list);
 546                 return rc;
 547         }
 548
 549         /* Unhook the record from context if encryption is not failure */
 550         ctx->open_rec = NULL;
 551         tls_advance_record_sn(sk, prot, &tls_ctx->tx);
 552         return rc;
 553 }
 554
 555 static int tls_split_open_record(struct sock *sk, struct tls_rec *from,
 556                                  struct tls_rec **to, struct sk_msg *msg_opl,
 557                                  struct sk_msg *msg_oen, u32 split_point,
 558                                  u32 tx_overhead_size, u32 *orig_end)
 559 {
 560         u32 i, j, bytes = 0, apply = msg_opl->apply_bytes;
 561         struct scatterlist *sge, *osge, *nsge;
 562         u32 orig_size = msg_opl->sg.size;
 563         struct scatterlist tmp = { };
 564         struct sk_msg *msg_npl;
 565         struct tls_rec *new;
 566         int ret;
 567
 568         new = tls_get_rec(sk);
 569         if (!new)
 570                 return -ENOMEM;
 571         ret = sk_msg_alloc(sk, &new->msg_encrypted, msg_opl->sg.size +
 572                            tx_overhead_size, 0);
 573         if (ret < 0) {
 574                 tls_free_rec(sk, new);
 575                 return ret;
 576         }
 577
 578         *orig_end = msg_opl->sg.end;
 579         i = msg_opl->sg.start;
 580         sge = sk_msg_elem(msg_opl, i);
 581         while (apply && sge->length) {
 582                 if (sge->length > apply) {
 583                         u32 len = sge->length - apply;
 584
 585                         get_page(sg_page(sge));
 586                         sg_set_page(&tmp, sg_page(sge), len,
 587                                     sge->offset + apply);
 588                         sge->length = apply;
 589                         bytes += apply;
 590                         apply = 0;
 591                 } else {
 592                         apply -= sge->length;
 593                         bytes += sge->length;
 594                 }
 595
 596                 sk_msg_iter_var_next(i);
 597                 if (i == msg_opl->sg.end)
 598                         break;
 599                 sge = sk_msg_elem(msg_opl, i);
 600         }
 601
 602         msg_opl->sg.end = i;
 603         msg_opl->sg.curr = i;
 604         msg_opl->sg.copybreak = 0;
 605         msg_opl->apply_bytes = 0;
 606         msg_opl->sg.size = bytes;
 607
 608         msg_npl = &new->msg_plaintext;
 609         msg_npl->apply_bytes = apply;
 610         msg_npl->sg.size = orig_size - bytes;
 611
 612         j = msg_npl->sg.start;
 613         nsge = sk_msg_elem(msg_npl, j);
 614         if (tmp.length) {
 615                 memcpy(nsge, &tmp, sizeof(*nsge));
 616                 sk_msg_iter_var_next(j);
 617                 nsge = sk_msg_elem(msg_npl, j);
 618         }
 619
 620         osge = sk_msg_elem(msg_opl, i);
 621         while (osge->length) {
 622                 memcpy(nsge, osge, sizeof(*nsge));
 623                 sg_unmark_end(nsge);
 624                 sk_msg_iter_var_next(i);
 625                 sk_msg_iter_var_next(j);
 626                 if (i == *orig_end)
 627                         break;
 628                 osge = sk_msg_elem(msg_opl, i);
 629                 nsge = sk_msg_elem(msg_npl, j);
 630         }
 631
 632         msg_npl->sg.end = j;
 633         msg_npl->sg.curr = j;
 634         msg_npl->sg.copybreak = 0;
 635
 636         *to = new;
 637         return 0;
 638 }
 639
 640 static void tls_merge_open_record(struct sock *sk, struct tls_rec *to,
 641                                   struct tls_rec *from, u32 orig_end)
 642 {
 643         struct sk_msg *msg_npl = &from->msg_plaintext;
 644         struct sk_msg *msg_opl = &to->msg_plaintext;
 645         struct scatterlist *osge, *nsge;
 646         u32 i, j;
 647
 648         i = msg_opl->sg.end;
 649         sk_msg_iter_var_prev(i);
 650         j = msg_npl->sg.start;
 651
 652         osge = sk_msg_elem(msg_opl, i);
 653         nsge = sk_msg_elem(msg_npl, j);
 654
 655         if (sg_page(osge) == sg_page(nsge) &&
 656             osge->offset + osge->length == nsge->offset) {
 657                 osge->length += nsge->length;
 658                 put_page(sg_page(nsge));
 659         }
 660
 661         msg_opl->sg.end = orig_end;
 662         msg_opl->sg.curr = orig_end;
 663         msg_opl->sg.copybreak = 0;
 664         msg_opl->apply_bytes = msg_opl->sg.size + msg_npl->sg.size;
 665         msg_opl->sg.size += msg_npl->sg.size;
 666
 667         sk_msg_free(sk, &to->msg_encrypted);
 668         sk_msg_xfer_full(&to->msg_encrypted, &from->msg_encrypted);
 669
 670         kfree(from);
 671 }
 672
 673 static int tls_push_record(struct sock *sk, int flags,
 674                            unsigned char record_type)
 675 {
 676         struct tls_context *tls_ctx = tls_get_ctx(sk);
 677         struct tls_prot_info *prot = &tls_ctx->prot_info;
 678         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 679         struct tls_rec *rec = ctx->open_rec, *tmp = NULL;
 680         u32 i, split_point, orig_end;
 681         struct sk_msg *msg_pl, *msg_en;
 682         struct aead_request *req;
 683         bool split;
 684         int rc;
 685
 686         if (!rec)
 687                 return 0;
 688
 689         msg_pl = &rec->msg_plaintext;
 690         msg_en = &rec->msg_encrypted;
 691
 692         split_point = msg_pl->apply_bytes;
 693         split = split_point && split_point < msg_pl->sg.size;
 694         if (unlikely((!split &&
 695                       msg_pl->sg.size +
 696                       prot->overhead_size > msg_en->sg.size) ||
 697                      (split &&
 698                       split_point +
 699                       prot->overhead_size > msg_en->sg.size))) {
 700                 split = true;
 701                 split_point = msg_en->sg.size;
 702         }
 703         if (split) {
 704                 rc = tls_split_open_record(sk, rec, &tmp, msg_pl, msg_en,
 705                                            split_point, prot->overhead_size,
 706                                            &orig_end);
 707                 if (rc < 0)
 708                         return rc;
 709                 /* This can happen if above tls_split_open_record allocates
 710                  * a single large encryption buffer instead of two smaller
 711                  * ones. In this case adjust pointers and continue without
 712                  * split.
 713                  */
 714                 if (!msg_pl->sg.size) {
 715                         tls_merge_open_record(sk, rec, tmp, orig_end);
 716                         msg_pl = &rec->msg_plaintext;
 717                         msg_en = &rec->msg_encrypted;
 718                         split = false;
 719                 }
 720                 sk_msg_trim(sk, msg_en, msg_pl->sg.size +
 721                             prot->overhead_size);
 722         }
 723
 724         rec->tx_flags = flags;
 725         req = &rec->aead_req;
 726
 727         i = msg_pl->sg.end;
 728         sk_msg_iter_var_prev(i);
 729
 730         rec->content_type = record_type;
 731         if (prot->version == TLS_1_3_VERSION) {
 732                 /* Add content type to end of message.  No padding added */
 733                 sg_set_buf(&rec->sg_content_type, &rec->content_type, 1);
 734                 sg_mark_end(&rec->sg_content_type);
 735                 sg_chain(msg_pl->sg.data, msg_pl->sg.end + 1,
 736                          &rec->sg_content_type);
 737         } else {
 738                 sg_mark_end(sk_msg_elem(msg_pl, i));
 739         }
 740
 741         if (msg_pl->sg.end < msg_pl->sg.start) {
 742                 sg_chain(&msg_pl->sg.data[msg_pl->sg.start],
 743                          MAX_SKB_FRAGS - msg_pl->sg.start + 1,
 744                          msg_pl->sg.data);
 745         }
 746
 747         i = msg_pl->sg.start;
 748         sg_chain(rec->sg_aead_in, 2, &msg_pl->sg.data[i]);
 749
 750         i = msg_en->sg.end;
 751         sk_msg_iter_var_prev(i);
 752         sg_mark_end(sk_msg_elem(msg_en, i));
 753
 754         i = msg_en->sg.start;
 755         sg_chain(rec->sg_aead_out, 2, &msg_en->sg.data[i]);
 756
 757         tls_make_aad(rec->aad_space, msg_pl->sg.size + prot->tail_size,
 758                      tls_ctx->tx.rec_seq, record_type, prot);
 759
 760         tls_fill_prepend(tls_ctx,
 761                          page_address(sg_page(&msg_en->sg.data[i])) +
 762                          msg_en->sg.data[i].offset,
 763                          msg_pl->sg.size + prot->tail_size,
 764                          record_type);
 765
 766         tls_ctx->pending_open_record_frags = false;
 767
 768         rc = tls_do_encryption(sk, tls_ctx, ctx, req,
 769                                msg_pl->sg.size + prot->tail_size, i);
 770         if (rc < 0) {
 771                 if (rc != -EINPROGRESS) {
 772                         tls_err_abort(sk, EBADMSG);
 773                         if (split) {
 774                                 tls_ctx->pending_open_record_frags = true;
 775                                 tls_merge_open_record(sk, rec, tmp, orig_end);
 776                         }
 777                 }
 778                 ctx->async_capable = 1;
 779                 return rc;
 780         } else if (split) {
 781                 msg_pl = &tmp->msg_plaintext;
 782                 msg_en = &tmp->msg_encrypted;
 783                 sk_msg_trim(sk, msg_en, msg_pl->sg.size + prot->overhead_size);
 784                 tls_ctx->pending_open_record_frags = true;
 785                 ctx->open_rec = tmp;
 786         }
 787
 788         return tls_tx_records(sk, flags);
 789 }
 790
 791 static int bpf_exec_tx_verdict(struct sk_msg *msg, struct sock *sk,
 792                                bool full_record, u8 record_type,
 793                                ssize_t *copied, int flags)
 794 {
 795         struct tls_context *tls_ctx = tls_get_ctx(sk);
 796         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 797         struct sk_msg msg_redir = { };
 798         struct sk_psock *psock;
 799         struct sock *sk_redir;
 800         struct tls_rec *rec;
 801         bool enospc, policy;
 802         int err = 0, send;
 803         u32 delta = 0;
 804
 805         policy = !(flags & MSG_SENDPAGE_NOPOLICY);
 806         psock = sk_psock_get(sk);
 807         if (!psock || !policy) {
 808                 err = tls_push_record(sk, flags, record_type);
 809                 if (err && sk->sk_err == EBADMSG) {
 810                         *copied -= sk_msg_free(sk, msg);
 811                         tls_free_open_rec(sk);
 812                         err = -sk->sk_err;
 813                 }
 814                 if (psock)
 815                         sk_psock_put(sk, psock);
 816                 return err;
 817         }
 818 more_data:
 819         enospc = sk_msg_full(msg);
 820         if (psock->eval == __SK_NONE) {
 821                 delta = msg->sg.size;
 822                 psock->eval = sk_psock_msg_verdict(sk, psock, msg);
 823                 delta -= msg->sg.size;
 824         }
 825         if (msg->cork_bytes && msg->cork_bytes > msg->sg.size &&
 826             !enospc && !full_record) {
 827                 err = -ENOSPC;
 828                 goto out_err;
 829         }
 830         msg->cork_bytes = 0;
 831         send = msg->sg.size;
 832         if (msg->apply_bytes && msg->apply_bytes < send)
 833                 send = msg->apply_bytes;
 834
 835         switch (psock->eval) {
 836         case __SK_PASS:
 837                 err = tls_push_record(sk, flags, record_type);
 838                 if (err && sk->sk_err == EBADMSG) {
 839                         *copied -= sk_msg_free(sk, msg);
 840                         tls_free_open_rec(sk);
 841                         err = -sk->sk_err;
 842                         goto out_err;
 843                 }
 844                 break;
 845         case __SK_REDIRECT:
 846                 sk_redir = psock->sk_redir;
 847                 memcpy(&msg_redir, msg, sizeof(*msg));
 848                 if (msg->apply_bytes < send)
 849                         msg->apply_bytes = 0;
 850                 else
 851                         msg->apply_bytes -= send;
 852                 sk_msg_return_zero(sk, msg, send);
 853                 msg->sg.size -= send;
 854                 release_sock(sk);
 855                 err = tcp_bpf_sendmsg_redir(sk_redir, &msg_redir, send, flags);
 856                 lock_sock(sk);
 857                 if (err < 0) {
 858                         *copied -= sk_msg_free_nocharge(sk, &msg_redir);
 859                         msg->sg.size = 0;
 860                 }
 861                 if (msg->sg.size == 0)
 862                         tls_free_open_rec(sk);
 863                 break;
 864         case __SK_DROP:
 865         default:
 866                 sk_msg_free_partial(sk, msg, send);
 867                 if (msg->apply_bytes < send)
 868                         msg->apply_bytes = 0;
 869                 else
 870                         msg->apply_bytes -= send;
 871                 if (msg->sg.size == 0)
 872                         tls_free_open_rec(sk);
 873                 *copied -= (send + delta);
 874                 err = -EACCES;
 875         }
 876
 877         if (likely(!err)) {
 878                 bool reset_eval = !ctx->open_rec;
 879
 880                 rec = ctx->open_rec;
 881                 if (rec) {
 882                         msg = &rec->msg_plaintext;
 883                         if (!msg->apply_bytes)
 884                                 reset_eval = true;
 885                 }
 886                 if (reset_eval) {
 887                         psock->eval = __SK_NONE;
 888                         if (psock->sk_redir) {
 889                                 sock_put(psock->sk_redir);
 890                                 psock->sk_redir = NULL;
 891                         }
 892                 }
 893                 if (rec)
 894                         goto more_data;
 895         }
 896  out_err:
 897         sk_psock_put(sk, psock);
 898         return err;
 899 }
 900
 901 static int tls_sw_push_pending_record(struct sock *sk, int flags)
 902 {
 903         struct tls_context *tls_ctx = tls_get_ctx(sk);
 904         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 905         struct tls_rec *rec = ctx->open_rec;
 906         struct sk_msg *msg_pl;
 907         size_t copied;
 908
 909         if (!rec)
 910                 return 0;
 911
 912         msg_pl = &rec->msg_plaintext;
 913         copied = msg_pl->sg.size;
 914         if (!copied)
 915                 return 0;
 916
 917         return bpf_exec_tx_verdict(msg_pl, sk, true, TLS_RECORD_TYPE_DATA,
 918                                    &copied, flags);
 919 }
 920
 921 int tls_sw_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
 922 {
 923         long timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
 924         struct tls_context *tls_ctx = tls_get_ctx(sk);
 925         struct tls_prot_info *prot = &tls_ctx->prot_info;
 926         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 927         bool async_capable = ctx->async_capable;
 928         unsigned char record_type = TLS_RECORD_TYPE_DATA;
 929         bool is_kvec = iov_iter_is_kvec(&msg->msg_iter);
 930         bool eor = !(msg->msg_flags & MSG_MORE);
 931         size_t try_to_copy;
 932         ssize_t copied = 0;
 933         struct sk_msg *msg_pl, *msg_en;
 934         struct tls_rec *rec;
 935         int required_size;
 936         int num_async = 0;
 937         bool full_record;
 938         int record_room;
 939         int num_zc = 0;
 940         int orig_size;
 941         int ret = 0;
 942         int pending;
 943
 944         if (msg->msg_flags & ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL |
 945                                MSG_CMSG_COMPAT))
 946                 return -EOPNOTSUPP;
 947
 948         mutex_lock(&tls_ctx->tx_lock);
 949         lock_sock(sk);
 950
 951         if (unlikely(msg->msg_controllen)) {
 952                 ret = tls_proccess_cmsg(sk, msg, &record_type);
 953                 if (ret) {
 954                         if (ret == -EINPROGRESS)
 955                                 num_async++;
 956                         else if (ret != -EAGAIN)
 957                                 goto send_end;
 958                 }
 959         }
 960
 961         while (msg_data_left(msg)) {
 962                 if (sk->sk_err) {
 963                         ret = -sk->sk_err;
 964                         goto send_end;
 965                 }
 966
 967                 if (ctx->open_rec)
 968                         rec = ctx->open_rec;
 969                 else
 970                         rec = ctx->open_rec = tls_get_rec(sk);
 971                 if (!rec) {
 972                         ret = -ENOMEM;
 973                         goto send_end;
 974                 }
 975
 976                 msg_pl = &rec->msg_plaintext;
 977                 msg_en = &rec->msg_encrypted;
 978
 979                 orig_size = msg_pl->sg.size;
 980                 full_record = false;
 981                 try_to_copy = msg_data_left(msg);
 982                 record_room = TLS_MAX_PAYLOAD_SIZE - msg_pl->sg.size;
 983                 if (try_to_copy >= record_room) {
 984                         try_to_copy = record_room;
 985                         full_record = true;
 986                 }
 987
 988                 required_size = msg_pl->sg.size + try_to_copy +
 989                                 prot->overhead_size;
 990
 991                 if (!sk_stream_memory_free(sk))
 992                         goto wait_for_sndbuf;
 993
 994 alloc_encrypted:
 995                 ret = tls_alloc_encrypted_msg(sk, required_size);
 996                 if (ret) {
 997                         if (ret != -ENOSPC)
 998                                 goto wait_for_memory;
 999
1000                         /* Adjust try_to_copy according to the amount that was
1001                          * actually allocated. The difference is due
1002                          * to max sg elements limit
1003                          */
1004                         try_to_copy -= required_size - msg_en->sg.size;
1005                         full_record = true;
1006                 }
1007
1008                 if (!is_kvec && (full_record || eor) && !async_capable) {
1009                         u32 first = msg_pl->sg.end;
1010
1011                         ret = sk_msg_zerocopy_from_iter(sk, &msg->msg_iter,
1012                                                         msg_pl, try_to_copy);
1013                         if (ret)
1014                                 goto fallback_to_reg_send;
1015
1016                         num_zc++;
1017                         copied += try_to_copy;
1018
1019                         sk_msg_sg_copy_set(msg_pl, first);
1020                         ret = bpf_exec_tx_verdict(msg_pl, sk, full_record,
1021                                                   record_type, &copied,
1022                                                   msg->msg_flags);
1023                         if (ret) {
1024                                 if (ret == -EINPROGRESS)
1025                                         num_async++;
1026                                 else if (ret == -ENOMEM)
1027                                         goto wait_for_memory;
1028                                 else if (ctx->open_rec && ret == -ENOSPC)
1029                                         goto rollback_iter;
1030                                 else if (ret != -EAGAIN)
1031                                         goto send_end;
1032                         }
1033                         continue;
1034 rollback_iter:
1035                         copied -= try_to_copy;
1036                         sk_msg_sg_copy_clear(msg_pl, first);
1037                         iov_iter_revert(&msg->msg_iter,
1038                                         msg_pl->sg.size - orig_size);
1039 fallback_to_reg_send:
1040                         sk_msg_trim(sk, msg_pl, orig_size);
1041                 }
1042
1043                 required_size = msg_pl->sg.size + try_to_copy;
1044
1045                 ret = tls_clone_plaintext_msg(sk, required_size);
1046                 if (ret) {
1047                         if (ret != -ENOSPC)
1048                                 goto send_end;
1049
1050                         /* Adjust try_to_copy according to the amount that was
1051                          * actually allocated. The difference is due
1052                          * to max sg elements limit
1053                          */
1054                         try_to_copy -= required_size - msg_pl->sg.size;
1055                         full_record = true;
1056                         sk_msg_trim(sk, msg_en,
1057                                     msg_pl->sg.size + prot->overhead_size);
1058                 }
1059
1060                 if (try_to_copy) {
1061                         ret = sk_msg_memcopy_from_iter(sk, &msg->msg_iter,
1062                                                        msg_pl, try_to_copy);
1063                         if (ret < 0)
1064                                 goto trim_sgl;
1065                 }
1066
1067                 /* Open records defined only if successfully copied, otherwise
1068                  * we would trim the sg but not reset the open record frags.
1069                  */
1070                 tls_ctx->pending_open_record_frags = true;
1071                 copied += try_to_copy;
1072                 if (full_record || eor) {
1073                         ret = bpf_exec_tx_verdict(msg_pl, sk, full_record,
1074                                                   record_type, &copied,
1075                                                   msg->msg_flags);
1076                         if (ret) {
1077                                 if (ret == -EINPROGRESS)
1078                                         num_async++;
1079                                 else if (ret == -ENOMEM)
1080                                         goto wait_for_memory;
1081                                 else if (ret != -EAGAIN) {
1082                                         if (ret == -ENOSPC)
1083                                                 ret = 0;
1084                                         goto send_end;
1085                                 }
1086                         }
1087                 }
1088
1089                 continue;
1090
1091 wait_for_sndbuf:
1092                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1093 wait_for_memory:
1094                 ret = sk_stream_wait_memory(sk, &timeo);
1095                 if (ret) {
1096 trim_sgl:
1097                         if (ctx->open_rec)
1098                                 tls_trim_both_msgs(sk, orig_size);
1099                         goto send_end;
1100                 }
1101
1102                 if (ctx->open_rec && msg_en->sg.size < required_size)
1103                         goto alloc_encrypted;
1104         }
1105
1106         if (!num_async) {
1107                 goto send_end;
1108         } else if (num_zc) {
1109                 /* Wait for pending encryptions to get completed */
1110                 spin_lock_bh(&ctx->encrypt_compl_lock);
1111                 ctx->async_notify = true;
1112
1113                 pending = atomic_read(&ctx->encrypt_pending);
1114                 spin_unlock_bh(&ctx->encrypt_compl_lock);
1115                 if (pending)
1116                         crypto_wait_req(-EINPROGRESS, &ctx->async_wait);
1117                 else
1118                         reinit_completion(&ctx->async_wait.completion);
1119
1120                 /* There can be no concurrent accesses, since we have no
1121                  * pending encrypt operations
1122                  */
1123                 WRITE_ONCE(ctx->async_notify, false);
1124
1125                 if (ctx->async_wait.err) {
1126                         ret = ctx->async_wait.err;
1127                         copied = 0;
1128                 }
1129         }
1130
1131         /* Transmit if any encryptions have completed */
1132         if (test_and_clear_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask)) {
1133                 cancel_delayed_work(&ctx->tx_work.work);
1134                 tls_tx_records(sk, msg->msg_flags);
1135         }
1136
1137 send_end:
1138         ret = sk_stream_error(sk, msg->msg_flags, ret);
1139
1140         release_sock(sk);
1141         mutex_unlock(&tls_ctx->tx_lock);
1142         return copied > 0 ? copied : ret;
1143 }
1144
1145 static int tls_sw_do_sendpage(struct sock *sk, struct page *page,
1146                               int offset, size_t size, int flags)
1147 {
1148         long timeo = sock_sndtimeo(sk, flags & MSG_DONTWAIT);
1149         struct tls_context *tls_ctx = tls_get_ctx(sk);
1150         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
1151         struct tls_prot_info *prot = &tls_ctx->prot_info;
1152         unsigned char record_type = TLS_RECORD_TYPE_DATA;
1153         struct sk_msg *msg_pl;
1154         struct tls_rec *rec;
1155         int num_async = 0;
1156         ssize_t copied = 0;
1157         bool full_record;
1158         int record_room;
1159         int ret = 0;
1160         bool eor;
1161
1162         eor = !(flags & MSG_SENDPAGE_NOTLAST);
1163         sk_clear_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1164
1165         /* Call the sk_stream functions to manage the sndbuf mem. */
1166         while (size > 0) {
1167                 size_t copy, required_size;
1168
1169                 if (sk->sk_err) {
1170                         ret = -sk->sk_err;
1171                         goto sendpage_end;
1172                 }
1173
1174                 if (ctx->open_rec)
1175                         rec = ctx->open_rec;
1176                 else
1177                         rec = ctx->open_rec = tls_get_rec(sk);
1178                 if (!rec) {
1179                         ret = -ENOMEM;
1180                         goto sendpage_end;
1181                 }
1182
1183                 msg_pl = &rec->msg_plaintext;
1184
1185                 full_record = false;
1186                 record_room = TLS_MAX_PAYLOAD_SIZE - msg_pl->sg.size;
1187                 copy = size;
1188                 if (copy >= record_room) {
1189                         copy = record_room;
1190                         full_record = true;
1191                 }
1192
1193                 required_size = msg_pl->sg.size + copy + prot->overhead_size;
1194
1195                 if (!sk_stream_memory_free(sk))
1196                         goto wait_for_sndbuf;
1197 alloc_payload:
1198                 ret = tls_alloc_encrypted_msg(sk, required_size);
1199                 if (ret) {
1200                         if (ret != -ENOSPC)
1201                                 goto wait_for_memory;
1202
1203                         /* Adjust copy according to the amount that was
1204                          * actually allocated. The difference is due
1205                          * to max sg elements limit
1206                          */
1207                         copy -= required_size - msg_pl->sg.size;
1208                         full_record = true;
1209                 }
1210
1211                 sk_msg_page_add(msg_pl, page, copy, offset);
1212                 sk_mem_charge(sk, copy);
1213
1214                 offset += copy;
1215                 size -= copy;
1216                 copied += copy;
1217
1218                 tls_ctx->pending_open_record_frags = true;
1219                 if (full_record || eor || sk_msg_full(msg_pl)) {
1220                         ret = bpf_exec_tx_verdict(msg_pl, sk, full_record,
1221                                                   record_type, &copied, flags);
1222                         if (ret) {
1223                                 if (ret == -EINPROGRESS)
1224                                         num_async++;
1225                                 else if (ret == -ENOMEM)
1226                                         goto wait_for_memory;
1227                                 else if (ret != -EAGAIN) {
1228                                         if (ret == -ENOSPC)
1229                                                 ret = 0;
1230                                         goto sendpage_end;
1231                                 }
1232                         }
1233                 }
1234                 continue;
1235 wait_for_sndbuf:
1236                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1237 wait_for_memory:
1238                 ret = sk_stream_wait_memory(sk, &timeo);
1239                 if (ret) {
1240                         if (ctx->open_rec)
1241                                 tls_trim_both_msgs(sk, msg_pl->sg.size);
1242                         goto sendpage_end;
1243                 }
1244
1245                 if (ctx->open_rec)
1246                         goto alloc_payload;
1247         }
1248
1249         if (num_async) {
1250                 /* Transmit if any encryptions have completed */
1251                 if (test_and_clear_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask)) {
1252                         cancel_delayed_work(&ctx->tx_work.work);
1253                         tls_tx_records(sk, flags);
1254                 }
1255         }
1256 sendpage_end:
1257         ret = sk_stream_error(sk, flags, ret);
1258         return copied > 0 ? copied : ret;
1259 }
1260
1261 int tls_sw_sendpage_locked(struct sock *sk, struct page *page,
1262                            int offset, size_t size, int flags)
1263 {
1264         if (flags & ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL |
1265                       MSG_SENDPAGE_NOTLAST | MSG_SENDPAGE_NOPOLICY |
1266                       MSG_NO_SHARED_FRAGS))
1267                 return -EOPNOTSUPP;
1268
1269         return tls_sw_do_sendpage(sk, page, offset, size, flags);
1270 }
1271
1272 int tls_sw_sendpage(struct sock *sk, struct page *page,
1273                     int offset, size_t size, int flags)
1274 {
1275         struct tls_context *tls_ctx = tls_get_ctx(sk);
1276         int ret;
1277
1278         if (flags & ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL |
1279                       MSG_SENDPAGE_NOTLAST | MSG_SENDPAGE_NOPOLICY))
1280                 return -EOPNOTSUPP;
1281
1282         mutex_lock(&tls_ctx->tx_lock);
1283         lock_sock(sk);
1284         ret = tls_sw_do_sendpage(sk, page, offset, size, flags);
1285         release_sock(sk);
1286         mutex_unlock(&tls_ctx->tx_lock);
1287         return ret;
1288 }
1289
1290 static struct sk_buff *tls_wait_data(struct sock *sk, struct sk_psock *psock,
1291                                      bool nonblock, long timeo, int *err)
1292 {
1293         struct tls_context *tls_ctx = tls_get_ctx(sk);
1294         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1295         struct sk_buff *skb;
1296         DEFINE_WAIT_FUNC(wait, woken_wake_function);
1297
1298         while (!(skb = ctx->recv_pkt) && sk_psock_queue_empty(psock)) {
1299                 if (sk->sk_err) {
1300                         *err = sock_error(sk);
1301                         return NULL;
1302                 }
1303
1304                 if (!skb_queue_empty(&sk->sk_receive_queue)) {
1305                         __strp_unpause(&ctx->strp);
1306                         if (ctx->recv_pkt)
1307                                 return ctx->recv_pkt;
1308                 }
1309
1310                 if (sk->sk_shutdown & RCV_SHUTDOWN)
1311                         return NULL;
1312
1313                 if (sock_flag(sk, SOCK_DONE))
1314                         return NULL;
1315
1316                 if (nonblock || !timeo) {
1317                         *err = -EAGAIN;
1318                         return NULL;
1319                 }
1320
1321                 add_wait_queue(sk_sleep(sk), &wait);
1322                 sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1323                 sk_wait_event(sk, &timeo,
1324                               ctx->recv_pkt != skb ||
1325                               !sk_psock_queue_empty(psock),
1326                               &wait);
1327                 sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1328                 remove_wait_queue(sk_sleep(sk), &wait);
1329
1330                 /* Handle signals */
1331                 if (signal_pending(current)) {
1332                         *err = sock_intr_errno(timeo);
1333                         return NULL;
1334                 }
1335         }
1336
1337         return skb;
1338 }
1339
1340 static int tls_setup_from_iter(struct sock *sk, struct iov_iter *from,
1341                                int length, int *pages_used,
1342                                unsigned int *size_used,
1343                                struct scatterlist *to,
1344                                int to_max_pages)
1345 {
1346         int rc = 0, i = 0, num_elem = *pages_used, maxpages;
1347         struct page *pages[MAX_SKB_FRAGS];
1348         unsigned int size = *size_used;
1349         ssize_t copied, use;
1350         size_t offset;
1351
1352         while (length > 0) {
1353                 i = 0;
1354                 maxpages = to_max_pages - num_elem;
1355                 if (maxpages == 0) {
1356                         rc = -EFAULT;
1357                         goto out;
1358                 }
1359                 copied = iov_iter_get_pages(from, pages,
1360                                             length,
1361                                             maxpages, &offset);
1362                 if (copied <= 0) {
1363                         rc = -EFAULT;
1364                         goto out;
1365                 }
1366
1367                 iov_iter_advance(from, copied);
1368
1369                 length -= copied;
1370                 size += copied;
1371                 while (copied) {
1372                         use = min_t(int, copied, PAGE_SIZE - offset);
1373
1374                         sg_set_page(&to[num_elem],
1375                                     pages[i], use, offset);
1376                         sg_unmark_end(&to[num_elem]);
1377                         /* We do not uncharge memory from this API */
1378
1379                         offset = 0;
1380                         copied -= use;
1381
1382                         i++;
1383                         num_elem++;
1384                 }
1385         }
1386         /* Mark the end in the last sg entry if newly added */
1387         if (num_elem > *pages_used)
1388                 sg_mark_end(&to[num_elem - 1]);
1389 out:
1390         if (rc)
1391                 iov_iter_revert(from, size - *size_used);
1392         *size_used = size;
1393         *pages_used = num_elem;
1394
1395         return rc;
1396 }
1397
1398 /* This function decrypts the input skb into either out_iov or in out_sg
1399  * or in skb buffers itself. The input parameter 'zc' indicates if
1400  * zero-copy mode needs to be tried or not. With zero-copy mode, either
1401  * out_iov or out_sg must be non-NULL. In case both out_iov and out_sg are
1402  * NULL, then the decryption happens inside skb buffers itself, i.e.
1403  * zero-copy gets disabled and 'zc' is updated.
1404  */
1405
1406 static int decrypt_internal(struct sock *sk, struct sk_buff *skb,
1407                             struct iov_iter *out_iov,
1408                             struct scatterlist *out_sg,
1409                             int *chunk, bool *zc, bool async)
1410 {
1411         struct tls_context *tls_ctx = tls_get_ctx(sk);
1412         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1413         struct tls_prot_info *prot = &tls_ctx->prot_info;
1414         struct strp_msg *rxm = strp_msg(skb);
1415         int n_sgin, n_sgout, nsg, mem_size, aead_size, err, pages = 0;
1416         struct aead_request *aead_req;
1417         struct sk_buff *unused;
1418         u8 *aad, *iv, *mem = NULL;
1419         struct scatterlist *sgin = NULL;
1420         struct scatterlist *sgout = NULL;
1421         const int data_len = rxm->full_len - prot->overhead_size +
1422                              prot->tail_size;
1423         int iv_offset = 0;
1424
1425         if (*zc && (out_iov || out_sg)) {
1426                 if (out_iov)
1427                         n_sgout = iov_iter_npages(out_iov, INT_MAX) + 1;
1428                 else
1429                         n_sgout = sg_nents(out_sg);
1430                 n_sgin = skb_nsg(skb, rxm->offset + prot->prepend_size,
1431                                  rxm->full_len - prot->prepend_size);
1432         } else {
1433                 n_sgout = 0;
1434                 *zc = false;
1435                 n_sgin = skb_cow_data(skb, 0, &unused);
1436         }
1437
1438         if (n_sgin < 1)
1439                 return -EBADMSG;
1440
1441         /* Increment to accommodate AAD */
1442         n_sgin = n_sgin + 1;
1443
1444         nsg = n_sgin + n_sgout;
1445
1446         aead_size = sizeof(*aead_req) + crypto_aead_reqsize(ctx->aead_recv);
1447         mem_size = aead_size + (nsg * sizeof(struct scatterlist));
1448         mem_size = mem_size + prot->aad_size;
1449         mem_size = mem_size + crypto_aead_ivsize(ctx->aead_recv);
1450
1451         /* Allocate a single block of memory which contains
1452          * aead_req || sgin[] || sgout[] || aad || iv.
1453          * This order achieves correct alignment for aead_req, sgin, sgout.
1454          */
1455         mem = kmalloc(mem_size, sk->sk_allocation);
1456         if (!mem)
1457                 return -ENOMEM;
1458
1459         /* Segment the allocated memory */
1460         aead_req = (struct aead_request *)mem;
1461         sgin = (struct scatterlist *)(mem + aead_size);
1462         sgout = sgin + n_sgin;
1463         aad = (u8 *)(sgout + n_sgout);
1464         iv = aad + prot->aad_size;
1465
1466         /* For CCM based ciphers, first byte of nonce+iv is a constant */
1467         switch (prot->cipher_type) {
1468         case TLS_CIPHER_AES_CCM_128:
1469                 iv[0] = TLS_AES_CCM_IV_B0_BYTE;
1470                 iv_offset = 1;
1471                 break;
1472         case TLS_CIPHER_SM4_CCM:
1473                 iv[0] = TLS_SM4_CCM_IV_B0_BYTE;
1474                 iv_offset = 1;
1475                 break;
1476         }
1477
1478         /* Prepare IV */
1479         err = skb_copy_bits(skb, rxm->offset + TLS_HEADER_SIZE,
1480                             iv + iv_offset + prot->salt_size,
1481                             prot->iv_size);
1482         if (err < 0) {
1483                 kfree(mem);
1484                 return err;
1485         }
1486         if (prot->version == TLS_1_3_VERSION ||
1487             prot->cipher_type == TLS_CIPHER_CHACHA20_POLY1305)
1488                 memcpy(iv + iv_offset, tls_ctx->rx.iv,
1489                        crypto_aead_ivsize(ctx->aead_recv));
1490         else
1491                 memcpy(iv + iv_offset, tls_ctx->rx.iv, prot->salt_size);
1492
1493         xor_iv_with_seq(prot, iv, tls_ctx->rx.rec_seq);
1494
1495         /* Prepare AAD */
1496         tls_make_aad(aad, rxm->full_len - prot->overhead_size +
1497                      prot->tail_size,
1498                      tls_ctx->rx.rec_seq, ctx->control, prot);
1499
1500         /* Prepare sgin */
1501         sg_init_table(sgin, n_sgin);
1502         sg_set_buf(&sgin[0], aad, prot->aad_size);
1503         err = skb_to_sgvec(skb, &sgin[1],
1504                            rxm->offset + prot->prepend_size,
1505                            rxm->full_len - prot->prepend_size);
1506         if (err < 0) {
1507                 kfree(mem);
1508                 return err;
1509         }
1510
1511         if (n_sgout) {
1512                 if (out_iov) {
1513                         sg_init_table(sgout, n_sgout);
1514                         sg_set_buf(&sgout[0], aad, prot->aad_size);
1515
1516                         *chunk = 0;
1517                         err = tls_setup_from_iter(sk, out_iov, data_len,
1518                                                   &pages, chunk, &sgout[1],
1519                                                   (n_sgout - 1));
1520                         if (err < 0)
1521                                 goto fallback_to_reg_recv;
1522                 } else if (out_sg) {
1523                         memcpy(sgout, out_sg, n_sgout * sizeof(*sgout));
1524                 } else {
1525                         goto fallback_to_reg_recv;
1526                 }
1527         } else {
1528 fallback_to_reg_recv:
1529                 sgout = sgin;
1530                 pages = 0;
1531                 *chunk = data_len;
1532                 *zc = false;
1533         }
1534
1535         /* Prepare and submit AEAD request */
1536         err = tls_do_decryption(sk, skb, sgin, sgout, iv,
1537                                 data_len, aead_req, async);
1538         if (err == -EINPROGRESS)
1539                 return err;
1540
1541         /* Release the pages in case iov was mapped to pages */
1542         for (; pages > 0; pages--)
1543                 put_page(sg_page(&sgout[pages]));
1544
1545         kfree(mem);
1546         return err;
1547 }
1548
1549 static int decrypt_skb_update(struct sock *sk, struct sk_buff *skb,
1550                               struct iov_iter *dest, int *chunk, bool *zc,
1551                               bool async)
1552 {
1553         struct tls_context *tls_ctx = tls_get_ctx(sk);
1554         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1555         struct tls_prot_info *prot = &tls_ctx->prot_info;
1556         struct strp_msg *rxm = strp_msg(skb);
1557         int pad, err = 0;
1558
1559         if (!ctx->decrypted) {
1560                 if (tls_ctx->rx_conf == TLS_HW) {
1561                         err = tls_device_decrypted(sk, tls_ctx, skb, rxm);
1562                         if (err < 0)
1563                                 return err;
1564                 }
1565
1566                 /* Still not decrypted after tls_device */
1567                 if (!ctx->decrypted) {
1568                         err = decrypt_internal(sk, skb, dest, NULL, chunk, zc,
1569                                                async);
1570                         if (err < 0) {
1571                                 if (err == -EINPROGRESS)
1572                                         tls_advance_record_sn(sk, prot,
1573                                                               &tls_ctx->rx);
1574                                 else if (err == -EBADMSG)
1575                                         TLS_INC_STATS(sock_net(sk),
1576                                                       LINUX_MIB_TLSDECRYPTERROR);
1577                                 return err;
1578                         }
1579                 } else {
1580                         *zc = false;
1581                 }
1582
1583                 pad = padding_length(ctx, prot, skb);
1584                 if (pad < 0)
1585                         return pad;
1586
1587                 rxm->full_len -= pad;
1588                 rxm->offset += prot->prepend_size;
1589                 rxm->full_len -= prot->overhead_size;
1590                 tls_advance_record_sn(sk, prot, &tls_ctx->rx);
1591                 ctx->decrypted = 1;
1592                 ctx->saved_data_ready(sk);
1593         } else {
1594                 *zc = false;
1595         }
1596
1597         return err;
1598 }
1599
1600 int decrypt_skb(struct sock *sk, struct sk_buff *skb,
1601                 struct scatterlist *sgout)
1602 {
1603         bool zc = true;
1604         int chunk;
1605
1606         return decrypt_internal(sk, skb, NULL, sgout, &chunk, &zc, false);
1607 }
1608
1609 static bool tls_sw_advance_skb(struct sock *sk, struct sk_buff *skb,
1610                                unsigned int len)
1611 {
1612         struct tls_context *tls_ctx = tls_get_ctx(sk);
1613         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1614
1615         if (skb) {
1616                 struct strp_msg *rxm = strp_msg(skb);
1617
1618                 if (len < rxm->full_len) {
1619                         rxm->offset += len;
1620                         rxm->full_len -= len;
1621                         return false;
1622                 }
1623                 consume_skb(skb);
1624         }
1625
1626         /* Finished with message */
1627         ctx->recv_pkt = NULL;
1628         __strp_unpause(&ctx->strp);
1629
1630         return true;
1631 }
1632
1633 /* This function traverses the rx_list in tls receive context to copies the
1634  * decrypted records into the buffer provided by caller zero copy is not
1635  * true. Further, the records are removed from the rx_list if it is not a peek
1636  * case and the record has been consumed completely.
1637  */
1638 static int process_rx_list(struct tls_sw_context_rx *ctx,
1639                            struct msghdr *msg,
1640                            u8 *control,
1641                            bool *cmsg,
1642                            size_t skip,
1643                            size_t len,
1644                            bool zc,
1645                            bool is_peek)
1646 {
1647         struct sk_buff *skb = skb_peek(&ctx->rx_list);
1648         u8 ctrl = *control;
1649         u8 msgc = *cmsg;
1650         struct tls_msg *tlm;
1651         ssize_t copied = 0;
1652
1653         /* Set the record type in 'control' if caller didn't pass it */
1654         if (!ctrl && skb) {
1655                 tlm = tls_msg(skb);
1656                 ctrl = tlm->control;
1657         }
1658
1659         while (skip && skb) {
1660                 struct strp_msg *rxm = strp_msg(skb);
1661                 tlm = tls_msg(skb);
1662
1663                 /* Cannot process a record of different type */
1664                 if (ctrl != tlm->control)
1665                         return 0;
1666
1667                 if (skip < rxm->full_len)
1668                         break;
1669
1670                 skip = skip - rxm->full_len;
1671                 skb = skb_peek_next(skb, &ctx->rx_list);
1672         }
1673
1674         while (len && skb) {
1675                 struct sk_buff *next_skb;
1676                 struct strp_msg *rxm = strp_msg(skb);
1677                 int chunk = min_t(unsigned int, rxm->full_len - skip, len);
1678
1679                 tlm = tls_msg(skb);
1680
1681                 /* Cannot process a record of different type */
1682                 if (ctrl != tlm->control)
1683                         return 0;
1684
1685                 /* Set record type if not already done. For a non-data record,
1686                  * do not proceed if record type could not be copied.
1687                  */
1688                 if (!msgc) {
1689                         int cerr = put_cmsg(msg, SOL_TLS, TLS_GET_RECORD_TYPE,
1690                                             sizeof(ctrl), &ctrl);
1691                         msgc = true;
1692                         if (ctrl != TLS_RECORD_TYPE_DATA) {
1693                                 if (cerr || msg->msg_flags & MSG_CTRUNC)
1694                                         return -EIO;
1695
1696                                 *cmsg = msgc;
1697                         }
1698                 }
1699
1700                 if (!zc || (rxm->full_len - skip) > len) {
1701                         int err = skb_copy_datagram_msg(skb, rxm->offset + skip,
1702                                                     msg, chunk);
1703                         if (err < 0)
1704                                 return err;
1705                 }
1706
1707                 len = len - chunk;
1708                 copied = copied + chunk;
1709
1710                 /* Consume the data from record if it is non-peek case*/
1711                 if (!is_peek) {
1712                         rxm->offset = rxm->offset + chunk;
1713                         rxm->full_len = rxm->full_len - chunk;
1714
1715                         /* Return if there is unconsumed data in the record */
1716                         if (rxm->full_len - skip)
1717                                 break;
1718                 }
1719
1720                 /* The remaining skip-bytes must lie in 1st record in rx_list.
1721                  * So from the 2nd record, 'skip' should be 0.
1722                  */
1723                 skip = 0;
1724
1725                 if (msg)
1726                         msg->msg_flags |= MSG_EOR;
1727
1728                 next_skb = skb_peek_next(skb, &ctx->rx_list);
1729
1730                 if (!is_peek) {
1731                         skb_unlink(skb, &ctx->rx_list);
1732                         consume_skb(skb);
1733                 }
1734
1735                 skb = next_skb;
1736         }
1737
1738         *control = ctrl;
1739         return copied;
1740 }
1741
1742 int tls_sw_recvmsg(struct sock *sk,
1743                    struct msghdr *msg,
1744                    size_t len,
1745                    int nonblock,
1746                    int flags,
1747                    int *addr_len)
1748 {
1749         struct tls_context *tls_ctx = tls_get_ctx(sk);
1750         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1751         struct tls_prot_info *prot = &tls_ctx->prot_info;
1752         struct sk_psock *psock;
1753         unsigned char control = 0;
1754         ssize_t decrypted = 0;
1755         struct strp_msg *rxm;
1756         struct tls_msg *tlm;
1757         struct sk_buff *skb;
1758         ssize_t copied = 0;
1759         bool cmsg = false;
1760         int target, err = 0;
1761         long timeo;
1762         bool is_kvec = iov_iter_is_kvec(&msg->msg_iter);
1763         bool is_peek = flags & MSG_PEEK;
1764         bool bpf_strp_enabled;
1765         int num_async = 0;
1766         int pending;
1767
1768         flags |= nonblock;
1769
1770         if (unlikely(flags & MSG_ERRQUEUE))
1771                 return sock_recv_errqueue(sk, msg, len, SOL_IP, IP_RECVERR);
1772
1773         psock = sk_psock_get(sk);
1774         lock_sock(sk);
1775         bpf_strp_enabled = sk_psock_strp_enabled(psock);
1776
1777         /* Process pending decrypted records. It must be non-zero-copy */
1778         err = process_rx_list(ctx, msg, &control, &cmsg, 0, len, false,
1779                               is_peek);
1780         if (err < 0) {
1781                 tls_err_abort(sk, err);
1782                 goto end;
1783         } else {
1784                 copied = err;
1785         }
1786
1787         if (len <= copied)
1788                 goto recv_end;
1789
1790         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1791         len = len - copied;
1792         timeo = sock_rcvtimeo(sk, flags & MSG_DONTWAIT);
1793
1794         while (len && (decrypted + copied < target || ctx->recv_pkt)) {
1795                 bool retain_skb = false;
1796                 bool zc = false;
1797                 int to_decrypt;
1798                 int chunk = 0;
1799                 bool async_capable;
1800                 bool async = false;
1801
1802                 skb = tls_wait_data(sk, psock, flags & MSG_DONTWAIT, timeo, &err);
1803                 if (!skb) {
1804                         if (psock) {
1805                                 int ret = sk_msg_recvmsg(sk, psock, msg, len,
1806                                                          flags);
1807
1808                                 if (ret > 0) {
1809                                         decrypted += ret;
1810                                         len -= ret;
1811                                         continue;
1812                                 }
1813                         }
1814                         goto recv_end;
1815                 } else {
1816                         tlm = tls_msg(skb);
1817                         if (prot->version == TLS_1_3_VERSION)
1818                                 tlm->control = 0;
1819                         else
1820                                 tlm->control = ctx->control;
1821                 }
1822
1823                 rxm = strp_msg(skb);
1824
1825                 to_decrypt = rxm->full_len - prot->overhead_size;
1826
1827                 if (to_decrypt <= len && !is_kvec && !is_peek &&
1828                     ctx->control == TLS_RECORD_TYPE_DATA &&
1829                     prot->version != TLS_1_3_VERSION &&
1830                     !bpf_strp_enabled)
1831                         zc = true;
1832
1833                 /* Do not use async mode if record is non-data */
1834                 if (ctx->control == TLS_RECORD_TYPE_DATA && !bpf_strp_enabled)
1835                         async_capable = ctx->async_capable;
1836                 else
1837                         async_capable = false;
1838
1839                 err = decrypt_skb_update(sk, skb, &msg->msg_iter,
1840                                          &chunk, &zc, async_capable);
1841                 if (err < 0 && err != -EINPROGRESS) {
1842                         tls_err_abort(sk, EBADMSG);
1843                         goto recv_end;
1844                 }
1845
1846                 if (err == -EINPROGRESS) {
1847                         async = true;
1848                         num_async++;
1849                 } else if (prot->version == TLS_1_3_VERSION) {
1850                         tlm->control = ctx->control;
1851                 }
1852
1853                 /* If the type of records being processed is not known yet,
1854                  * set it to record type just dequeued. If it is already known,
1855                  * but does not match the record type just dequeued, go to end.
1856                  * We always get record type here since for tls1.2, record type
1857                  * is known just after record is dequeued from stream parser.
1858                  * For tls1.3, we disable async.
1859                  */
1860
1861                 if (!control)
1862                         control = tlm->control;
1863                 else if (control != tlm->control)
1864                         goto recv_end;
1865
1866                 if (!cmsg) {
1867                         int cerr;
1868
1869                         cerr = put_cmsg(msg, SOL_TLS, TLS_GET_RECORD_TYPE,
1870                                         sizeof(control), &control);
1871                         cmsg = true;
1872                         if (control != TLS_RECORD_TYPE_DATA) {
1873                                 if (cerr || msg->msg_flags & MSG_CTRUNC) {
1874                                         err = -EIO;
1875                                         goto recv_end;
1876                                 }
1877                         }
1878                 }
1879
1880                 if (async)
1881                         goto pick_next_record;
1882
1883                 if (!zc) {
1884                         if (bpf_strp_enabled) {
1885                                 err = sk_psock_tls_strp_read(psock, skb);
1886                                 if (err != __SK_PASS) {
1887                                         rxm->offset = rxm->offset + rxm->full_len;
1888                                         rxm->full_len = 0;
1889                                         if (err == __SK_DROP)
1890                                                 consume_skb(skb);
1891                                         ctx->recv_pkt = NULL;
1892                                         __strp_unpause(&ctx->strp);
1893                                         continue;
1894                                 }
1895                         }
1896
1897                         if (rxm->full_len > len) {
1898                                 retain_skb = true;
1899                                 chunk = len;
1900                         } else {
1901                                 chunk = rxm->full_len;
1902                         }
1903
1904                         err = skb_copy_datagram_msg(skb, rxm->offset,
1905                                                     msg, chunk);
1906                         if (err < 0)
1907                                 goto recv_end;
1908
1909                         if (!is_peek) {
1910                                 rxm->offset = rxm->offset + chunk;
1911                                 rxm->full_len = rxm->full_len - chunk;
1912                         }
1913                 }
1914
1915 pick_next_record:
1916                 if (chunk > len)
1917                         chunk = len;
1918
1919                 decrypted += chunk;
1920                 len -= chunk;
1921
1922                 /* For async or peek case, queue the current skb */
1923                 if (async || is_peek || retain_skb) {
1924                         skb_queue_tail(&ctx->rx_list, skb);
1925                         skb = NULL;
1926                 }
1927
1928                 if (tls_sw_advance_skb(sk, skb, chunk)) {
1929                         /* Return full control message to
1930                          * userspace before trying to parse
1931                          * another message type
1932                          */
1933                         msg->msg_flags |= MSG_EOR;
1934                         if (control != TLS_RECORD_TYPE_DATA)
1935                                 goto recv_end;
1936                 } else {
1937                         break;
1938                 }
1939         }
1940
1941 recv_end:
1942         if (num_async) {
1943                 /* Wait for all previously submitted records to be decrypted */
1944                 spin_lock_bh(&ctx->decrypt_compl_lock);
1945                 ctx->async_notify = true;
1946                 pending = atomic_read(&ctx->decrypt_pending);
1947                 spin_unlock_bh(&ctx->decrypt_compl_lock);
1948                 if (pending) {
1949                         err = crypto_wait_req(-EINPROGRESS, &ctx->async_wait);
1950                         if (err) {
1951                                 /* one of async decrypt failed */
1952                                 tls_err_abort(sk, err);
1953                                 copied = 0;
1954                                 decrypted = 0;
1955                                 goto end;
1956                         }
1957                 } else {
1958                         reinit_completion(&ctx->async_wait.completion);
1959                 }
1960
1961                 /* There can be no concurrent accesses, since we have no
1962                  * pending decrypt operations
1963                  */
1964                 WRITE_ONCE(ctx->async_notify, false);
1965
1966                 /* Drain records from the rx_list & copy if required */
1967                 if (is_peek || is_kvec)
1968                         err = process_rx_list(ctx, msg, &control, &cmsg, copied,
1969                                               decrypted, false, is_peek);
1970                 else
1971                         err = process_rx_list(ctx, msg, &control, &cmsg, 0,
1972                                               decrypted, true, is_peek);
1973                 if (err < 0) {
1974                         tls_err_abort(sk, err);
1975                         copied = 0;
1976                         goto end;
1977                 }
1978         }
1979
1980         copied += decrypted;
1981
1982 end:
1983         release_sock(sk);
1984         if (psock)
1985                 sk_psock_put(sk, psock);
1986         return copied ? : err;
1987 }
1988
1989 ssize_t tls_sw_splice_read(struct socket *sock,  loff_t *ppos,
1990                            struct pipe_inode_info *pipe,
1991                            size_t len, unsigned int flags)
1992 {
1993         struct tls_context *tls_ctx = tls_get_ctx(sock->sk);
1994         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1995         struct strp_msg *rxm = NULL;
1996         struct sock *sk = sock->sk;
1997         struct sk_buff *skb;
1998         ssize_t copied = 0;
1999         int err = 0;
2000         long timeo;
2001         int chunk;
2002         bool zc = false;
2003
2004         lock_sock(sk);
2005
2006         timeo = sock_rcvtimeo(sk, flags & SPLICE_F_NONBLOCK);
2007
2008         skb = tls_wait_data(sk, NULL, flags & SPLICE_F_NONBLOCK, timeo, &err);
2009         if (!skb)
2010                 goto splice_read_end;
2011
2012         if (!ctx->decrypted) {
2013                 err = decrypt_skb_update(sk, skb, NULL, &chunk, &zc, false);
2014
2015                 /* splice does not support reading control messages */
2016                 if (ctx->control != TLS_RECORD_TYPE_DATA) {
2017                         err = -EINVAL;
2018                         goto splice_read_end;
2019                 }
2020
2021                 if (err < 0) {
2022                         tls_err_abort(sk, EBADMSG);
2023                         goto splice_read_end;
2024                 }
2025                 ctx->decrypted = 1;
2026         }
2027         rxm = strp_msg(skb);
2028
2029         chunk = min_t(unsigned int, rxm->full_len, len);
2030         copied = skb_splice_bits(skb, sk, rxm->offset, pipe, chunk, flags);
2031         if (copied < 0)
2032                 goto splice_read_end;
2033
2034         tls_sw_advance_skb(sk, skb, copied);
2035
2036 splice_read_end:
2037         release_sock(sk);
2038         return copied ? : err;
2039 }
2040
2041 bool tls_sw_stream_read(const struct sock *sk)
2042 {
2043         struct tls_context *tls_ctx = tls_get_ctx(sk);
2044         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2045         bool ingress_empty = true;
2046         struct sk_psock *psock;
2047
2048         rcu_read_lock();
2049         psock = sk_psock(sk);
2050         if (psock)
2051                 ingress_empty = list_empty(&psock->ingress_msg);
2052         rcu_read_unlock();
2053
2054         return !ingress_empty || ctx->recv_pkt ||
2055                 !skb_queue_empty(&ctx->rx_list);
2056 }
2057
2058 static int tls_read_size(struct strparser *strp, struct sk_buff *skb)
2059 {
2060         struct tls_context *tls_ctx = tls_get_ctx(strp->sk);
2061         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2062         struct tls_prot_info *prot = &tls_ctx->prot_info;
2063         char header[TLS_HEADER_SIZE + MAX_IV_SIZE];
2064         struct strp_msg *rxm = strp_msg(skb);
2065         size_t cipher_overhead;
2066         size_t data_len = 0;
2067         int ret;
2068
2069         /* Verify that we have a full TLS header, or wait for more data */
2070         if (rxm->offset + prot->prepend_size > skb->len)
2071                 return 0;
2072
2073         /* Sanity-check size of on-stack buffer. */
2074         if (WARN_ON(prot->prepend_size > sizeof(header))) {
2075                 ret = -EINVAL;
2076                 goto read_failure;
2077         }
2078
2079         /* Linearize header to local buffer */
2080         ret = skb_copy_bits(skb, rxm->offset, header, prot->prepend_size);
2081
2082         if (ret < 0)
2083                 goto read_failure;
2084
2085         ctx->control = header[0];
2086
2087         data_len = ((header[4] & 0xFF) | (header[3] << 8));
2088
2089         cipher_overhead = prot->tag_size;
2090         if (prot->version != TLS_1_3_VERSION &&
2091             prot->cipher_type != TLS_CIPHER_CHACHA20_POLY1305)
2092                 cipher_overhead += prot->iv_size;
2093
2094         if (data_len > TLS_MAX_PAYLOAD_SIZE + cipher_overhead +
2095             prot->tail_size) {
2096                 ret = -EMSGSIZE;
2097                 goto read_failure;
2098         }
2099         if (data_len < cipher_overhead) {
2100                 ret = -EBADMSG;
2101                 goto read_failure;
2102         }
2103
2104         /* Note that both TLS1.3 and TLS1.2 use TLS_1_2 version here */
2105         if (header[1] != TLS_1_2_VERSION_MINOR ||
2106             header[2] != TLS_1_2_VERSION_MAJOR) {
2107                 ret = -EINVAL;
2108                 goto read_failure;
2109         }
2110
2111         tls_device_rx_resync_new_rec(strp->sk, data_len + TLS_HEADER_SIZE,
2112                                      TCP_SKB_CB(skb)->seq + rxm->offset);
2113         return data_len + TLS_HEADER_SIZE;
2114
2115 read_failure:
2116         tls_err_abort(strp->sk, ret);
2117
2118         return ret;
2119 }
2120
2121 static void tls_queue(struct strparser *strp, struct sk_buff *skb)
2122 {
2123         struct tls_context *tls_ctx = tls_get_ctx(strp->sk);
2124         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2125
2126         ctx->decrypted = 0;
2127
2128         ctx->recv_pkt = skb;
2129         strp_pause(strp);
2130
2131         ctx->saved_data_ready(strp->sk);
2132 }
2133
2134 static void tls_data_ready(struct sock *sk)
2135 {
2136         struct tls_context *tls_ctx = tls_get_ctx(sk);
2137         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2138         struct sk_psock *psock;
2139
2140         strp_data_ready(&ctx->strp);
2141
2142         psock = sk_psock_get(sk);
2143         if (psock) {
2144                 if (!list_empty(&psock->ingress_msg))
2145                         ctx->saved_data_ready(sk);
2146                 sk_psock_put(sk, psock);
2147         }
2148 }
2149
2150 void tls_sw_cancel_work_tx(struct tls_context *tls_ctx)
2151 {
2152         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
2153
2154         set_bit(BIT_TX_CLOSING, &ctx->tx_bitmask);
2155         set_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask);
2156         cancel_delayed_work_sync(&ctx->tx_work.work);
2157 }
2158
2159 void tls_sw_release_resources_tx(struct sock *sk)
2160 {
2161         struct tls_context *tls_ctx = tls_get_ctx(sk);
2162         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
2163         struct tls_rec *rec, *tmp;
2164         int pending;
2165
2166         /* Wait for any pending async encryptions to complete */
2167         spin_lock_bh(&ctx->encrypt_compl_lock);
2168         ctx->async_notify = true;
2169         pending = atomic_read(&ctx->encrypt_pending);
2170         spin_unlock_bh(&ctx->encrypt_compl_lock);
2171
2172         if (pending)
2173                 crypto_wait_req(-EINPROGRESS, &ctx->async_wait);
2174
2175         tls_tx_records(sk, -1);
2176
2177         /* Free up un-sent records in tx_list. First, free
2178          * the partially sent record if any at head of tx_list.
2179          */
2180         if (tls_ctx->partially_sent_record) {
2181                 tls_free_partial_record(sk, tls_ctx);
2182                 rec = list_first_entry(&ctx->tx_list,
2183                                        struct tls_rec, list);
2184                 list_del(&rec->list);
2185                 sk_msg_free(sk, &rec->msg_plaintext);
2186                 kfree(rec);
2187         }
2188
2189         list_for_each_entry_safe(rec, tmp, &ctx->tx_list, list) {
2190                 list_del(&rec->list);
2191                 sk_msg_free(sk, &rec->msg_encrypted);
2192                 sk_msg_free(sk, &rec->msg_plaintext);
2193                 kfree(rec);
2194         }
2195
2196         crypto_free_aead(ctx->aead_send);
2197         tls_free_open_rec(sk);
2198 }
2199
2200 void tls_sw_free_ctx_tx(struct tls_context *tls_ctx)
2201 {
2202         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
2203
2204         kfree(ctx);
2205 }
2206
2207 void tls_sw_release_resources_rx(struct sock *sk)
2208 {
2209         struct tls_context *tls_ctx = tls_get_ctx(sk);
2210         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2211
2212         kfree(tls_ctx->rx.rec_seq);
2213         kfree(tls_ctx->rx.iv);
2214
2215         if (ctx->aead_recv) {
2216                 kfree_skb(ctx->recv_pkt);
2217                 ctx->recv_pkt = NULL;
2218                 skb_queue_purge(&ctx->rx_list);
2219                 crypto_free_aead(ctx->aead_recv);
2220                 strp_stop(&ctx->strp);
2221                 /* If tls_sw_strparser_arm() was not called (cleanup paths)
2222                  * we still want to strp_stop(), but sk->sk_data_ready was
2223                  * never swapped.
2224                  */
2225                 if (ctx->saved_data_ready) {
2226                         write_lock_bh(&sk->sk_callback_lock);
2227                         sk->sk_data_ready = ctx->saved_data_ready;
2228                         write_unlock_bh(&sk->sk_callback_lock);
2229                 }
2230         }
2231 }
2232
2233 void tls_sw_strparser_done(struct tls_context *tls_ctx)
2234 {
2235         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2236
2237         strp_done(&ctx->strp);
2238 }
2239
2240 void tls_sw_free_ctx_rx(struct tls_context *tls_ctx)
2241 {
2242         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2243
2244         kfree(ctx);
2245 }
2246
2247 void tls_sw_free_resources_rx(struct sock *sk)
2248 {
2249         struct tls_context *tls_ctx = tls_get_ctx(sk);
2250
2251         tls_sw_release_resources_rx(sk);
2252         tls_sw_free_ctx_rx(tls_ctx);
2253 }
2254
2255 /* The work handler to transmitt the encrypted records in tx_list */
2256 static void tx_work_handler(struct work_struct *work)
2257 {
2258         struct delayed_work *delayed_work = to_delayed_work(work);
2259         struct tx_work *tx_work = container_of(delayed_work,
2260                                                struct tx_work, work);
2261         struct sock *sk = tx_work->sk;
2262         struct tls_context *tls_ctx = tls_get_ctx(sk);
2263         struct tls_sw_context_tx *ctx;
2264
2265         if (unlikely(!tls_ctx))
2266                 return;
2267
2268         ctx = tls_sw_ctx_tx(tls_ctx);
2269         if (test_bit(BIT_TX_CLOSING, &ctx->tx_bitmask))
2270                 return;
2271
2272         if (!test_and_clear_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask))
2273                 return;
2274         mutex_lock(&tls_ctx->tx_lock);
2275         lock_sock(sk);
2276         tls_tx_records(sk, -1);
2277         release_sock(sk);
2278         mutex_unlock(&tls_ctx->tx_lock);
2279 }
2280
2281 void tls_sw_write_space(struct sock *sk, struct tls_context *ctx)
2282 {
2283         struct tls_sw_context_tx *tx_ctx = tls_sw_ctx_tx(ctx);
2284
2285         /* Schedule the transmission if tx list is ready */
2286         if (is_tx_ready(tx_ctx) &&
2287             !test_and_set_bit(BIT_TX_SCHEDULED, &tx_ctx->tx_bitmask))
2288                 schedule_delayed_work(&tx_ctx->tx_work.work, 0);
2289 }
2290
2291 void tls_sw_strparser_arm(struct sock *sk, struct tls_context *tls_ctx)
2292 {
2293         struct tls_sw_context_rx *rx_ctx = tls_sw_ctx_rx(tls_ctx);
2294
2295         write_lock_bh(&sk->sk_callback_lock);
2296         rx_ctx->saved_data_ready = sk->sk_data_ready;
2297         sk->sk_data_ready = tls_data_ready;
2298         write_unlock_bh(&sk->sk_callback_lock);
2299
2300         strp_check_rcv(&rx_ctx->strp);
2301 }
2302
2303 int tls_set_sw_offload(struct sock *sk, struct tls_context *ctx, int tx)
2304 {
2305         struct tls_context *tls_ctx = tls_get_ctx(sk);
2306         struct tls_prot_info *prot = &tls_ctx->prot_info;
2307         struct tls_crypto_info *crypto_info;
2308         struct tls12_crypto_info_aes_gcm_128 *gcm_128_info;
2309         struct tls12_crypto_info_aes_gcm_256 *gcm_256_info;
2310         struct tls12_crypto_info_aes_ccm_128 *ccm_128_info;
2311         struct tls12_crypto_info_chacha20_poly1305 *chacha20_poly1305_info;
2312         struct tls_sw_context_tx *sw_ctx_tx = NULL;
2313         struct tls_sw_context_rx *sw_ctx_rx = NULL;
2314         struct cipher_context *cctx;
2315         struct crypto_aead **aead;
2316         struct strp_callbacks cb;
2317         u16 nonce_size, tag_size, iv_size, rec_seq_size, salt_size;
2318         struct crypto_tfm *tfm;
2319         char *iv, *rec_seq, *key, *salt, *cipher_name;
2320         size_t keysize;
2321         int rc = 0;
2322
2323         if (!ctx) {
2324                 rc = -EINVAL;
2325                 goto out;
2326         }
2327
2328         if (tx) {
2329                 if (!ctx->priv_ctx_tx) {
2330                         sw_ctx_tx = kzalloc(sizeof(*sw_ctx_tx), GFP_KERNEL);
2331                         if (!sw_ctx_tx) {
2332                                 rc = -ENOMEM;
2333                                 goto out;
2334                         }
2335                         ctx->priv_ctx_tx = sw_ctx_tx;
2336                 } else {
2337                         sw_ctx_tx =
2338                                 (struct tls_sw_context_tx *)ctx->priv_ctx_tx;
2339                 }
2340         } else {
2341                 if (!ctx->priv_ctx_rx) {
2342                         sw_ctx_rx = kzalloc(sizeof(*sw_ctx_rx), GFP_KERNEL);
2343                         if (!sw_ctx_rx) {
2344                                 rc = -ENOMEM;
2345                                 goto out;
2346                         }
2347                         ctx->priv_ctx_rx = sw_ctx_rx;
2348                 } else {
2349                         sw_ctx_rx =
2350                                 (struct tls_sw_context_rx *)ctx->priv_ctx_rx;
2351                 }
2352         }
2353
2354         if (tx) {
2355                 crypto_init_wait(&sw_ctx_tx->async_wait);
2356                 spin_lock_init(&sw_ctx_tx->encrypt_compl_lock);
2357                 crypto_info = &ctx->crypto_send.info;
2358                 cctx = &ctx->tx;
2359                 aead = &sw_ctx_tx->aead_send;
2360                 INIT_LIST_HEAD(&sw_ctx_tx->tx_list);
2361                 INIT_DELAYED_WORK(&sw_ctx_tx->tx_work.work, tx_work_handler);
2362                 sw_ctx_tx->tx_work.sk = sk;
2363         } else {
2364                 crypto_init_wait(&sw_ctx_rx->async_wait);
2365                 spin_lock_init(&sw_ctx_rx->decrypt_compl_lock);
2366                 crypto_info = &ctx->crypto_recv.info;
2367                 cctx = &ctx->rx;
2368                 skb_queue_head_init(&sw_ctx_rx->rx_list);
2369                 aead = &sw_ctx_rx->aead_recv;
2370         }
2371
2372         switch (crypto_info->cipher_type) {
2373         case TLS_CIPHER_AES_GCM_128: {
2374                 nonce_size = TLS_CIPHER_AES_GCM_128_IV_SIZE;
2375                 tag_size = TLS_CIPHER_AES_GCM_128_TAG_SIZE;
2376                 iv_size = TLS_CIPHER_AES_GCM_128_IV_SIZE;
2377                 iv = ((struct tls12_crypto_info_aes_gcm_128 *)crypto_info)->iv;
2378                 rec_seq_size = TLS_CIPHER_AES_GCM_128_REC_SEQ_SIZE;
2379                 rec_seq =
2380                  ((struct tls12_crypto_info_aes_gcm_128 *)crypto_info)->rec_seq;
2381                 gcm_128_info =
2382                         (struct tls12_crypto_info_aes_gcm_128 *)crypto_info;
2383                 keysize = TLS_CIPHER_AES_GCM_128_KEY_SIZE;
2384                 key = gcm_128_info->key;
2385                 salt = gcm_128_info->salt;
2386                 salt_size = TLS_CIPHER_AES_GCM_128_SALT_SIZE;
2387                 cipher_name = "gcm(aes)";
2388                 break;
2389         }
2390         case TLS_CIPHER_AES_GCM_256: {
2391                 nonce_size = TLS_CIPHER_AES_GCM_256_IV_SIZE;
2392                 tag_size = TLS_CIPHER_AES_GCM_256_TAG_SIZE;
2393                 iv_size = TLS_CIPHER_AES_GCM_256_IV_SIZE;
2394                 iv = ((struct tls12_crypto_info_aes_gcm_256 *)crypto_info)->iv;
2395                 rec_seq_size = TLS_CIPHER_AES_GCM_256_REC_SEQ_SIZE;
2396                 rec_seq =
2397                  ((struct tls12_crypto_info_aes_gcm_256 *)crypto_info)->rec_seq;
2398                 gcm_256_info =
2399                         (struct tls12_crypto_info_aes_gcm_256 *)crypto_info;
2400                 keysize = TLS_CIPHER_AES_GCM_256_KEY_SIZE;
2401                 key = gcm_256_info->key;
2402                 salt = gcm_256_info->salt;
2403                 salt_size = TLS_CIPHER_AES_GCM_256_SALT_SIZE;
2404                 cipher_name = "gcm(aes)";
2405                 break;
2406         }
2407         case TLS_CIPHER_AES_CCM_128: {
2408                 nonce_size = TLS_CIPHER_AES_CCM_128_IV_SIZE;
2409                 tag_size = TLS_CIPHER_AES_CCM_128_TAG_SIZE;
2410                 iv_size = TLS_CIPHER_AES_CCM_128_IV_SIZE;
2411                 iv = ((struct tls12_crypto_info_aes_ccm_128 *)crypto_info)->iv;
2412                 rec_seq_size = TLS_CIPHER_AES_CCM_128_REC_SEQ_SIZE;
2413                 rec_seq =
2414                 ((struct tls12_crypto_info_aes_ccm_128 *)crypto_info)->rec_seq;
2415                 ccm_128_info =
2416                 (struct tls12_crypto_info_aes_ccm_128 *)crypto_info;
2417                 keysize = TLS_CIPHER_AES_CCM_128_KEY_SIZE;
2418                 key = ccm_128_info->key;
2419                 salt = ccm_128_info->salt;
2420                 salt_size = TLS_CIPHER_AES_CCM_128_SALT_SIZE;
2421                 cipher_name = "ccm(aes)";
2422                 break;
2423         }
2424         case TLS_CIPHER_CHACHA20_POLY1305: {
2425                 chacha20_poly1305_info = (void *)crypto_info;
2426                 nonce_size = 0;
2427                 tag_size = TLS_CIPHER_CHACHA20_POLY1305_TAG_SIZE;
2428                 iv_size = TLS_CIPHER_CHACHA20_POLY1305_IV_SIZE;
2429                 iv = chacha20_poly1305_info->iv;
2430                 rec_seq_size = TLS_CIPHER_CHACHA20_POLY1305_REC_SEQ_SIZE;
2431                 rec_seq = chacha20_poly1305_info->rec_seq;
2432                 keysize = TLS_CIPHER_CHACHA20_POLY1305_KEY_SIZE;
2433                 key = chacha20_poly1305_info->key;
2434                 salt = chacha20_poly1305_info->salt;
2435                 salt_size = TLS_CIPHER_CHACHA20_POLY1305_SALT_SIZE;
2436                 cipher_name = "rfc7539(chacha20,poly1305)";
2437                 break;
2438         }
2439         case TLS_CIPHER_SM4_GCM: {
2440                 struct tls12_crypto_info_sm4_gcm *sm4_gcm_info;
2441
2442                 sm4_gcm_info = (void *)crypto_info;
2443                 nonce_size = TLS_CIPHER_SM4_GCM_IV_SIZE;
2444                 tag_size = TLS_CIPHER_SM4_GCM_TAG_SIZE;
2445                 iv_size = TLS_CIPHER_SM4_GCM_IV_SIZE;
2446                 iv = sm4_gcm_info->iv;
2447                 rec_seq_size = TLS_CIPHER_SM4_GCM_REC_SEQ_SIZE;
2448                 rec_seq = sm4_gcm_info->rec_seq;
2449                 keysize = TLS_CIPHER_SM4_GCM_KEY_SIZE;
2450                 key = sm4_gcm_info->key;
2451                 salt = sm4_gcm_info->salt;
2452                 salt_size = TLS_CIPHER_SM4_GCM_SALT_SIZE;
2453                 cipher_name = "gcm(sm4)";
2454                 break;
2455         }
2456         case TLS_CIPHER_SM4_CCM: {
2457                 struct tls12_crypto_info_sm4_ccm *sm4_ccm_info;
2458
2459                 sm4_ccm_info = (void *)crypto_info;
2460                 nonce_size = TLS_CIPHER_SM4_CCM_IV_SIZE;
2461                 tag_size = TLS_CIPHER_SM4_CCM_TAG_SIZE;
2462                 iv_size = TLS_CIPHER_SM4_CCM_IV_SIZE;
2463                 iv = sm4_ccm_info->iv;
2464                 rec_seq_size = TLS_CIPHER_SM4_CCM_REC_SEQ_SIZE;
2465                 rec_seq = sm4_ccm_info->rec_seq;
2466                 keysize = TLS_CIPHER_SM4_CCM_KEY_SIZE;
2467                 key = sm4_ccm_info->key;
2468                 salt = sm4_ccm_info->salt;
2469                 salt_size = TLS_CIPHER_SM4_CCM_SALT_SIZE;
2470                 cipher_name = "ccm(sm4)";
2471                 break;
2472         }
2473         default:
2474                 rc = -EINVAL;
2475                 goto free_priv;
2476         }
2477
2478         /* Sanity-check the sizes for stack allocations. */
2479         if (iv_size > MAX_IV_SIZE || nonce_size > MAX_IV_SIZE ||
2480             rec_seq_size > TLS_MAX_REC_SEQ_SIZE) {
2481                 rc = -EINVAL;
2482                 goto free_priv;
2483         }
2484
2485         if (crypto_info->version == TLS_1_3_VERSION) {
2486                 nonce_size = 0;
2487                 prot->aad_size = TLS_HEADER_SIZE;
2488                 prot->tail_size = 1;
2489         } else {
2490                 prot->aad_size = TLS_AAD_SPACE_SIZE;
2491                 prot->tail_size = 0;
2492         }
2493
2494         prot->version = crypto_info->version;
2495         prot->cipher_type = crypto_info->cipher_type;
2496         prot->prepend_size = TLS_HEADER_SIZE + nonce_size;
2497         prot->tag_size = tag_size;
2498         prot->overhead_size = prot->prepend_size +
2499                               prot->tag_size + prot->tail_size;
2500         prot->iv_size = iv_size;
2501         prot->salt_size = salt_size;
2502         cctx->iv = kmalloc(iv_size + salt_size, GFP_KERNEL);
2503         if (!cctx->iv) {
2504                 rc = -ENOMEM;
2505                 goto free_priv;
2506         }
2507         /* Note: 128 & 256 bit salt are the same size */
2508         prot->rec_seq_size = rec_seq_size;
2509         memcpy(cctx->iv, salt, salt_size);
2510         memcpy(cctx->iv + salt_size, iv, iv_size);
2511         cctx->rec_seq = kmemdup(rec_seq, rec_seq_size, GFP_KERNEL);
2512         if (!cctx->rec_seq) {
2513                 rc = -ENOMEM;
2514                 goto free_iv;
2515         }
2516
2517         if (!*aead) {
2518                 *aead = crypto_alloc_aead(cipher_name, 0, 0);
2519                 if (IS_ERR(*aead)) {
2520                         rc = PTR_ERR(*aead);
2521                         *aead = NULL;
2522                         goto free_rec_seq;
2523                 }
2524         }
2525
2526         ctx->push_pending_record = tls_sw_push_pending_record;
2527
2528         rc = crypto_aead_setkey(*aead, key, keysize);
2529
2530         if (rc)
2531                 goto free_aead;
2532
2533         rc = crypto_aead_setauthsize(*aead, prot->tag_size);
2534         if (rc)
2535                 goto free_aead;
2536
2537         if (sw_ctx_rx) {
2538                 tfm = crypto_aead_tfm(sw_ctx_rx->aead_recv);
2539
2540                 if (crypto_info->version == TLS_1_3_VERSION)
2541                         sw_ctx_rx->async_capable = 0;
2542                 else
2543                         sw_ctx_rx->async_capable =
2544                                 !!(tfm->__crt_alg->cra_flags &
2545                                    CRYPTO_ALG_ASYNC);
2546
2547                 /* Set up strparser */
2548                 memset(&cb, 0, sizeof(cb));
2549                 cb.rcv_msg = tls_queue;
2550                 cb.parse_msg = tls_read_size;
2551
2552                 strp_init(&sw_ctx_rx->strp, sk, &cb);
2553         }
2554
2555         goto out;
2556
2557 free_aead:
2558         crypto_free_aead(*aead);
2559         *aead = NULL;
2560 free_rec_seq:
2561         kfree(cctx->rec_seq);
2562         cctx->rec_seq = NULL;
2563 free_iv:
2564         kfree(cctx->iv);
2565         cctx->iv = NULL;
2566 free_priv:
2567         if (tx) {
2568                 kfree(ctx->priv_ctx_tx);
2569                 ctx->priv_ctx_tx = NULL;
2570         } else {
2571                 kfree(ctx->priv_ctx_rx);
2572                 ctx->priv_ctx_rx = NULL;
2573         }
2574 out:
2575         return rc;
2576 }