net/tls/tls_sw.c

   1 /*
   2  * Copyright (c) 2016-2017, Mellanox Technologies. All rights reserved.
   3  * Copyright (c) 2016-2017, Dave Watson <davejwatson@fb.com>. All rights reserved.
   4  * Copyright (c) 2016-2017, Lance Chao <lancerchao@fb.com>. All rights reserved.
   5  * Copyright (c) 2016, Fridolin Pokorny <fridolin.pokorny@gmail.com>. All rights reserved.
   6  * Copyright (c) 2016, Nikos Mavrogiannopoulos <nmav@gnutls.org>. All rights reserved.
   7  * Copyright (c) 2018, Covalent IO, Inc. http://covalent.io
   8  *
   9  * This software is available to you under a choice of one of two
  10  * licenses.  You may choose to be licensed under the terms of the GNU
  11  * General Public License (GPL) Version 2, available from the file
  12  * COPYING in the main directory of this source tree, or the
  13  * OpenIB.org BSD license below:
  14  *
  15  *     Redistribution and use in source and binary forms, with or
  16  *     without modification, are permitted provided that the following
  17  *     conditions are met:
  18  *
  19  *      - Redistributions of source code must retain the above
  20  *        copyright notice, this list of conditions and the following
  21  *        disclaimer.
  22  *
  23  *      - Redistributions in binary form must reproduce the above
  24  *        copyright notice, this list of conditions and the following
  25  *        disclaimer in the documentation and/or other materials
  26  *        provided with the distribution.
  27  *
  28  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  29  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  30  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  31  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  32  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  33  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  34  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  35  * SOFTWARE.
  36  */
  37
  38 #include <linux/bug.h>
  39 #include <linux/sched/signal.h>
  40 #include <linux/module.h>
  41 #include <linux/kernel.h>
  42 #include <linux/splice.h>
  43 #include <crypto/aead.h>
  44
  45 #include <net/strparser.h>
  46 #include <net/tls.h>
  47 #include <trace/events/sock.h>
  48
  49 #include "tls.h"
  50
  51 struct tls_decrypt_arg {
  52         struct_group(inargs,
  53         bool zc;
  54         bool async;
  55         u8 tail;
  56         );
  57
  58         struct sk_buff *skb;
  59 };
  60
  61 struct tls_decrypt_ctx {
  62         struct sock *sk;
  63         u8 iv[MAX_IV_SIZE];
  64         u8 aad[TLS_MAX_AAD_SIZE];
  65         u8 tail;
  66         struct scatterlist sg[];
  67 };
  68
  69 noinline void tls_err_abort(struct sock *sk, int err)
  70 {
  71         WARN_ON_ONCE(err >= 0);
  72         /* sk->sk_err should contain a positive error code. */
  73         WRITE_ONCE(sk->sk_err, -err);
  74         /* Paired with smp_rmb() in tcp_poll() */
  75         smp_wmb();
  76         sk_error_report(sk);
  77 }
  78
  79 static int __skb_nsg(struct sk_buff *skb, int offset, int len,
  80                      unsigned int recursion_level)
  81 {
  82         int start = skb_headlen(skb);
  83         int i, chunk = start - offset;
  84         struct sk_buff *frag_iter;
  85         int elt = 0;
  86
  87         if (unlikely(recursion_level >= 24))
  88                 return -EMSGSIZE;
  89
  90         if (chunk > 0) {
  91                 if (chunk > len)
  92                         chunk = len;
  93                 elt++;
  94                 len -= chunk;
  95                 if (len == 0)
  96                         return elt;
  97                 offset += chunk;
  98         }
  99
 100         for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) {
 101                 int end;
 102
 103                 WARN_ON(start > offset + len);
 104
 105                 end = start + skb_frag_size(&skb_shinfo(skb)->frags[i]);
 106                 chunk = end - offset;
 107                 if (chunk > 0) {
 108                         if (chunk > len)
 109                                 chunk = len;
 110                         elt++;
 111                         len -= chunk;
 112                         if (len == 0)
 113                                 return elt;
 114                         offset += chunk;
 115                 }
 116                 start = end;
 117         }
 118
 119         if (unlikely(skb_has_frag_list(skb))) {
 120                 skb_walk_frags(skb, frag_iter) {
 121                         int end, ret;
 122
 123                         WARN_ON(start > offset + len);
 124
 125                         end = start + frag_iter->len;
 126                         chunk = end - offset;
 127                         if (chunk > 0) {
 128                                 if (chunk > len)
 129                                         chunk = len;
 130                                 ret = __skb_nsg(frag_iter, offset - start, chunk,
 131                                                 recursion_level + 1);
 132                                 if (unlikely(ret < 0))
 133                                         return ret;
 134                                 elt += ret;
 135                                 len -= chunk;
 136                                 if (len == 0)
 137                                         return elt;
 138                                 offset += chunk;
 139                         }
 140                         start = end;
 141                 }
 142         }
 143         BUG_ON(len);
 144         return elt;
 145 }
 146
 147 /* Return the number of scatterlist elements required to completely map the
 148  * skb, or -EMSGSIZE if the recursion depth is exceeded.
 149  */
 150 static int skb_nsg(struct sk_buff *skb, int offset, int len)
 151 {
 152         return __skb_nsg(skb, offset, len, 0);
 153 }
 154
 155 static int tls_padding_length(struct tls_prot_info *prot, struct sk_buff *skb,
 156                               struct tls_decrypt_arg *darg)
 157 {
 158         struct strp_msg *rxm = strp_msg(skb);
 159         struct tls_msg *tlm = tls_msg(skb);
 160         int sub = 0;
 161
 162         /* Determine zero-padding length */
 163         if (prot->version == TLS_1_3_VERSION) {
 164                 int offset = rxm->full_len - TLS_TAG_SIZE - 1;
 165                 char content_type = darg->zc ? darg->tail : 0;
 166                 int err;
 167
 168                 while (content_type == 0) {
 169                         if (offset < prot->prepend_size)
 170                                 return -EBADMSG;
 171                         err = skb_copy_bits(skb, rxm->offset + offset,
 172                                             &content_type, 1);
 173                         if (err)
 174                                 return err;
 175                         if (content_type)
 176                                 break;
 177                         sub++;
 178                         offset--;
 179                 }
 180                 tlm->control = content_type;
 181         }
 182         return sub;
 183 }
 184
 185 static void tls_decrypt_done(void *data, int err)
 186 {
 187         struct aead_request *aead_req = data;
 188         struct crypto_aead *aead = crypto_aead_reqtfm(aead_req);
 189         struct scatterlist *sgout = aead_req->dst;
 190         struct scatterlist *sgin = aead_req->src;
 191         struct tls_sw_context_rx *ctx;
 192         struct tls_decrypt_ctx *dctx;
 193         struct tls_context *tls_ctx;
 194         struct scatterlist *sg;
 195         unsigned int pages;
 196         struct sock *sk;
 197         int aead_size;
 198
 199         aead_size = sizeof(*aead_req) + crypto_aead_reqsize(aead);
 200         aead_size = ALIGN(aead_size, __alignof__(*dctx));
 201         dctx = (void *)((u8 *)aead_req + aead_size);
 202
 203         sk = dctx->sk;
 204         tls_ctx = tls_get_ctx(sk);
 205         ctx = tls_sw_ctx_rx(tls_ctx);
 206
 207         /* Propagate if there was an err */
 208         if (err) {
 209                 if (err == -EBADMSG)
 210                         TLS_INC_STATS(sock_net(sk), LINUX_MIB_TLSDECRYPTERROR);
 211                 ctx->async_wait.err = err;
 212                 tls_err_abort(sk, err);
 213         }
 214
 215         /* Free the destination pages if skb was not decrypted inplace */
 216         if (sgout != sgin) {
 217                 /* Skip the first S/G entry as it points to AAD */
 218                 for_each_sg(sg_next(sgout), sg, UINT_MAX, pages) {
 219                         if (!sg)
 220                                 break;
 221                         put_page(sg_page(sg));
 222                 }
 223         }
 224
 225         kfree(aead_req);
 226
 227         spin_lock_bh(&ctx->decrypt_compl_lock);
 228         if (!atomic_dec_return(&ctx->decrypt_pending))
 229                 complete(&ctx->async_wait.completion);
 230         spin_unlock_bh(&ctx->decrypt_compl_lock);
 231 }
 232
 233 static int tls_do_decryption(struct sock *sk,
 234                              struct scatterlist *sgin,
 235                              struct scatterlist *sgout,
 236                              char *iv_recv,
 237                              size_t data_len,
 238                              struct aead_request *aead_req,
 239                              struct tls_decrypt_arg *darg)
 240 {
 241         struct tls_context *tls_ctx = tls_get_ctx(sk);
 242         struct tls_prot_info *prot = &tls_ctx->prot_info;
 243         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
 244         int ret;
 245
 246         aead_request_set_tfm(aead_req, ctx->aead_recv);
 247         aead_request_set_ad(aead_req, prot->aad_size);
 248         aead_request_set_crypt(aead_req, sgin, sgout,
 249                                data_len + prot->tag_size,
 250                                (u8 *)iv_recv);
 251
 252         if (darg->async) {
 253                 aead_request_set_callback(aead_req,
 254                                           CRYPTO_TFM_REQ_MAY_BACKLOG,
 255                                           tls_decrypt_done, aead_req);
 256                 atomic_inc(&ctx->decrypt_pending);
 257         } else {
 258                 aead_request_set_callback(aead_req,
 259                                           CRYPTO_TFM_REQ_MAY_BACKLOG,
 260                                           crypto_req_done, &ctx->async_wait);
 261         }
 262
 263         ret = crypto_aead_decrypt(aead_req);
 264         if (ret == -EINPROGRESS) {
 265                 if (darg->async)
 266                         return 0;
 267
 268                 ret = crypto_wait_req(ret, &ctx->async_wait);
 269         }
 270         darg->async = false;
 271
 272         return ret;
 273 }
 274
 275 static void tls_trim_both_msgs(struct sock *sk, int target_size)
 276 {
 277         struct tls_context *tls_ctx = tls_get_ctx(sk);
 278         struct tls_prot_info *prot = &tls_ctx->prot_info;
 279         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 280         struct tls_rec *rec = ctx->open_rec;
 281
 282         sk_msg_trim(sk, &rec->msg_plaintext, target_size);
 283         if (target_size > 0)
 284                 target_size += prot->overhead_size;
 285         sk_msg_trim(sk, &rec->msg_encrypted, target_size);
 286 }
 287
 288 static int tls_alloc_encrypted_msg(struct sock *sk, int len)
 289 {
 290         struct tls_context *tls_ctx = tls_get_ctx(sk);
 291         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 292         struct tls_rec *rec = ctx->open_rec;
 293         struct sk_msg *msg_en = &rec->msg_encrypted;
 294
 295         return sk_msg_alloc(sk, msg_en, len, 0);
 296 }
 297
 298 static int tls_clone_plaintext_msg(struct sock *sk, int required)
 299 {
 300         struct tls_context *tls_ctx = tls_get_ctx(sk);
 301         struct tls_prot_info *prot = &tls_ctx->prot_info;
 302         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 303         struct tls_rec *rec = ctx->open_rec;
 304         struct sk_msg *msg_pl = &rec->msg_plaintext;
 305         struct sk_msg *msg_en = &rec->msg_encrypted;
 306         int skip, len;
 307
 308         /* We add page references worth len bytes from encrypted sg
 309          * at the end of plaintext sg. It is guaranteed that msg_en
 310          * has enough required room (ensured by caller).
 311          */
 312         len = required - msg_pl->sg.size;
 313
 314         /* Skip initial bytes in msg_en's data to be able to use
 315          * same offset of both plain and encrypted data.
 316          */
 317         skip = prot->prepend_size + msg_pl->sg.size;
 318
 319         return sk_msg_clone(sk, msg_pl, msg_en, skip, len);
 320 }
 321
 322 static struct tls_rec *tls_get_rec(struct sock *sk)
 323 {
 324         struct tls_context *tls_ctx = tls_get_ctx(sk);
 325         struct tls_prot_info *prot = &tls_ctx->prot_info;
 326         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 327         struct sk_msg *msg_pl, *msg_en;
 328         struct tls_rec *rec;
 329         int mem_size;
 330
 331         mem_size = sizeof(struct tls_rec) + crypto_aead_reqsize(ctx->aead_send);
 332
 333         rec = kzalloc(mem_size, sk->sk_allocation);
 334         if (!rec)
 335                 return NULL;
 336
 337         msg_pl = &rec->msg_plaintext;
 338         msg_en = &rec->msg_encrypted;
 339
 340         sk_msg_init(msg_pl);
 341         sk_msg_init(msg_en);
 342
 343         sg_init_table(rec->sg_aead_in, 2);
 344         sg_set_buf(&rec->sg_aead_in[0], rec->aad_space, prot->aad_size);
 345         sg_unmark_end(&rec->sg_aead_in[1]);
 346
 347         sg_init_table(rec->sg_aead_out, 2);
 348         sg_set_buf(&rec->sg_aead_out[0], rec->aad_space, prot->aad_size);
 349         sg_unmark_end(&rec->sg_aead_out[1]);
 350
 351         rec->sk = sk;
 352
 353         return rec;
 354 }
 355
 356 static void tls_free_rec(struct sock *sk, struct tls_rec *rec)
 357 {
 358         sk_msg_free(sk, &rec->msg_encrypted);
 359         sk_msg_free(sk, &rec->msg_plaintext);
 360         kfree(rec);
 361 }
 362
 363 static void tls_free_open_rec(struct sock *sk)
 364 {
 365         struct tls_context *tls_ctx = tls_get_ctx(sk);
 366         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 367         struct tls_rec *rec = ctx->open_rec;
 368
 369         if (rec) {
 370                 tls_free_rec(sk, rec);
 371                 ctx->open_rec = NULL;
 372         }
 373 }
 374
 375 int tls_tx_records(struct sock *sk, int flags)
 376 {
 377         struct tls_context *tls_ctx = tls_get_ctx(sk);
 378         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 379         struct tls_rec *rec, *tmp;
 380         struct sk_msg *msg_en;
 381         int tx_flags, rc = 0;
 382
 383         if (tls_is_partially_sent_record(tls_ctx)) {
 384                 rec = list_first_entry(&ctx->tx_list,
 385                                        struct tls_rec, list);
 386
 387                 if (flags == -1)
 388                         tx_flags = rec->tx_flags;
 389                 else
 390                         tx_flags = flags;
 391
 392                 rc = tls_push_partial_record(sk, tls_ctx, tx_flags);
 393                 if (rc)
 394                         goto tx_err;
 395
 396                 /* Full record has been transmitted.
 397                  * Remove the head of tx_list
 398                  */
 399                 list_del(&rec->list);
 400                 sk_msg_free(sk, &rec->msg_plaintext);
 401                 kfree(rec);
 402         }
 403
 404         /* Tx all ready records */
 405         list_for_each_entry_safe(rec, tmp, &ctx->tx_list, list) {
 406                 if (READ_ONCE(rec->tx_ready)) {
 407                         if (flags == -1)
 408                                 tx_flags = rec->tx_flags;
 409                         else
 410                                 tx_flags = flags;
 411
 412                         msg_en = &rec->msg_encrypted;
 413                         rc = tls_push_sg(sk, tls_ctx,
 414                                          &msg_en->sg.data[msg_en->sg.curr],
 415                                          0, tx_flags);
 416                         if (rc)
 417                                 goto tx_err;
 418
 419                         list_del(&rec->list);
 420                         sk_msg_free(sk, &rec->msg_plaintext);
 421                         kfree(rec);
 422                 } else {
 423                         break;
 424                 }
 425         }
 426
 427 tx_err:
 428         if (rc < 0 && rc != -EAGAIN)
 429                 tls_err_abort(sk, -EBADMSG);
 430
 431         return rc;
 432 }
 433
 434 static void tls_encrypt_done(void *data, int err)
 435 {
 436         struct tls_sw_context_tx *ctx;
 437         struct tls_context *tls_ctx;
 438         struct tls_prot_info *prot;
 439         struct tls_rec *rec = data;
 440         struct scatterlist *sge;
 441         struct sk_msg *msg_en;
 442         bool ready = false;
 443         struct sock *sk;
 444         int pending;
 445
 446         msg_en = &rec->msg_encrypted;
 447
 448         sk = rec->sk;
 449         tls_ctx = tls_get_ctx(sk);
 450         prot = &tls_ctx->prot_info;
 451         ctx = tls_sw_ctx_tx(tls_ctx);
 452
 453         sge = sk_msg_elem(msg_en, msg_en->sg.curr);
 454         sge->offset -= prot->prepend_size;
 455         sge->length += prot->prepend_size;
 456
 457         /* Check if error is previously set on socket */
 458         if (err || sk->sk_err) {
 459                 rec = NULL;
 460
 461                 /* If err is already set on socket, return the same code */
 462                 if (sk->sk_err) {
 463                         ctx->async_wait.err = -sk->sk_err;
 464                 } else {
 465                         ctx->async_wait.err = err;
 466                         tls_err_abort(sk, err);
 467                 }
 468         }
 469
 470         if (rec) {
 471                 struct tls_rec *first_rec;
 472
 473                 /* Mark the record as ready for transmission */
 474                 smp_store_mb(rec->tx_ready, true);
 475
 476                 /* If received record is at head of tx_list, schedule tx */
 477                 first_rec = list_first_entry(&ctx->tx_list,
 478                                              struct tls_rec, list);
 479                 if (rec == first_rec)
 480                         ready = true;
 481         }
 482
 483         spin_lock_bh(&ctx->encrypt_compl_lock);
 484         pending = atomic_dec_return(&ctx->encrypt_pending);
 485
 486         if (!pending && ctx->async_notify)
 487                 complete(&ctx->async_wait.completion);
 488         spin_unlock_bh(&ctx->encrypt_compl_lock);
 489
 490         if (!ready)
 491                 return;
 492
 493         /* Schedule the transmission */
 494         if (!test_and_set_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask))
 495                 schedule_delayed_work(&ctx->tx_work.work, 1);
 496 }
 497
 498 static int tls_do_encryption(struct sock *sk,
 499                              struct tls_context *tls_ctx,
 500                              struct tls_sw_context_tx *ctx,
 501                              struct aead_request *aead_req,
 502                              size_t data_len, u32 start)
 503 {
 504         struct tls_prot_info *prot = &tls_ctx->prot_info;
 505         struct tls_rec *rec = ctx->open_rec;
 506         struct sk_msg *msg_en = &rec->msg_encrypted;
 507         struct scatterlist *sge = sk_msg_elem(msg_en, start);
 508         int rc, iv_offset = 0;
 509
 510         /* For CCM based ciphers, first byte of IV is a constant */
 511         switch (prot->cipher_type) {
 512         case TLS_CIPHER_AES_CCM_128:
 513                 rec->iv_data[0] = TLS_AES_CCM_IV_B0_BYTE;
 514                 iv_offset = 1;
 515                 break;
 516         case TLS_CIPHER_SM4_CCM:
 517                 rec->iv_data[0] = TLS_SM4_CCM_IV_B0_BYTE;
 518                 iv_offset = 1;
 519                 break;
 520         }
 521
 522         memcpy(&rec->iv_data[iv_offset], tls_ctx->tx.iv,
 523                prot->iv_size + prot->salt_size);
 524
 525         tls_xor_iv_with_seq(prot, rec->iv_data + iv_offset,
 526                             tls_ctx->tx.rec_seq);
 527
 528         sge->offset += prot->prepend_size;
 529         sge->length -= prot->prepend_size;
 530
 531         msg_en->sg.curr = start;
 532
 533         aead_request_set_tfm(aead_req, ctx->aead_send);
 534         aead_request_set_ad(aead_req, prot->aad_size);
 535         aead_request_set_crypt(aead_req, rec->sg_aead_in,
 536                                rec->sg_aead_out,
 537                                data_len, rec->iv_data);
 538
 539         aead_request_set_callback(aead_req, CRYPTO_TFM_REQ_MAY_BACKLOG,
 540                                   tls_encrypt_done, rec);
 541
 542         /* Add the record in tx_list */
 543         list_add_tail((struct list_head *)&rec->list, &ctx->tx_list);
 544         atomic_inc(&ctx->encrypt_pending);
 545
 546         rc = crypto_aead_encrypt(aead_req);
 547         if (!rc || rc != -EINPROGRESS) {
 548                 atomic_dec(&ctx->encrypt_pending);
 549                 sge->offset -= prot->prepend_size;
 550                 sge->length += prot->prepend_size;
 551         }
 552
 553         if (!rc) {
 554                 WRITE_ONCE(rec->tx_ready, true);
 555         } else if (rc != -EINPROGRESS) {
 556                 list_del(&rec->list);
 557                 return rc;
 558         }
 559
 560         /* Unhook the record from context if encryption is not failure */
 561         ctx->open_rec = NULL;
 562         tls_advance_record_sn(sk, prot, &tls_ctx->tx);
 563         return rc;
 564 }
 565
 566 static int tls_split_open_record(struct sock *sk, struct tls_rec *from,
 567                                  struct tls_rec **to, struct sk_msg *msg_opl,
 568                                  struct sk_msg *msg_oen, u32 split_point,
 569                                  u32 tx_overhead_size, u32 *orig_end)
 570 {
 571         u32 i, j, bytes = 0, apply = msg_opl->apply_bytes;
 572         struct scatterlist *sge, *osge, *nsge;
 573         u32 orig_size = msg_opl->sg.size;
 574         struct scatterlist tmp = { };
 575         struct sk_msg *msg_npl;
 576         struct tls_rec *new;
 577         int ret;
 578
 579         new = tls_get_rec(sk);
 580         if (!new)
 581                 return -ENOMEM;
 582         ret = sk_msg_alloc(sk, &new->msg_encrypted, msg_opl->sg.size +
 583                            tx_overhead_size, 0);
 584         if (ret < 0) {
 585                 tls_free_rec(sk, new);
 586                 return ret;
 587         }
 588
 589         *orig_end = msg_opl->sg.end;
 590         i = msg_opl->sg.start;
 591         sge = sk_msg_elem(msg_opl, i);
 592         while (apply && sge->length) {
 593                 if (sge->length > apply) {
 594                         u32 len = sge->length - apply;
 595
 596                         get_page(sg_page(sge));
 597                         sg_set_page(&tmp, sg_page(sge), len,
 598                                     sge->offset + apply);
 599                         sge->length = apply;
 600                         bytes += apply;
 601                         apply = 0;
 602                 } else {
 603                         apply -= sge->length;
 604                         bytes += sge->length;
 605                 }
 606
 607                 sk_msg_iter_var_next(i);
 608                 if (i == msg_opl->sg.end)
 609                         break;
 610                 sge = sk_msg_elem(msg_opl, i);
 611         }
 612
 613         msg_opl->sg.end = i;
 614         msg_opl->sg.curr = i;
 615         msg_opl->sg.copybreak = 0;
 616         msg_opl->apply_bytes = 0;
 617         msg_opl->sg.size = bytes;
 618
 619         msg_npl = &new->msg_plaintext;
 620         msg_npl->apply_bytes = apply;
 621         msg_npl->sg.size = orig_size - bytes;
 622
 623         j = msg_npl->sg.start;
 624         nsge = sk_msg_elem(msg_npl, j);
 625         if (tmp.length) {
 626                 memcpy(nsge, &tmp, sizeof(*nsge));
 627                 sk_msg_iter_var_next(j);
 628                 nsge = sk_msg_elem(msg_npl, j);
 629         }
 630
 631         osge = sk_msg_elem(msg_opl, i);
 632         while (osge->length) {
 633                 memcpy(nsge, osge, sizeof(*nsge));
 634                 sg_unmark_end(nsge);
 635                 sk_msg_iter_var_next(i);
 636                 sk_msg_iter_var_next(j);
 637                 if (i == *orig_end)
 638                         break;
 639                 osge = sk_msg_elem(msg_opl, i);
 640                 nsge = sk_msg_elem(msg_npl, j);
 641         }
 642
 643         msg_npl->sg.end = j;
 644         msg_npl->sg.curr = j;
 645         msg_npl->sg.copybreak = 0;
 646
 647         *to = new;
 648         return 0;
 649 }
 650
 651 static void tls_merge_open_record(struct sock *sk, struct tls_rec *to,
 652                                   struct tls_rec *from, u32 orig_end)
 653 {
 654         struct sk_msg *msg_npl = &from->msg_plaintext;
 655         struct sk_msg *msg_opl = &to->msg_plaintext;
 656         struct scatterlist *osge, *nsge;
 657         u32 i, j;
 658
 659         i = msg_opl->sg.end;
 660         sk_msg_iter_var_prev(i);
 661         j = msg_npl->sg.start;
 662
 663         osge = sk_msg_elem(msg_opl, i);
 664         nsge = sk_msg_elem(msg_npl, j);
 665
 666         if (sg_page(osge) == sg_page(nsge) &&
 667             osge->offset + osge->length == nsge->offset) {
 668                 osge->length += nsge->length;
 669                 put_page(sg_page(nsge));
 670         }
 671
 672         msg_opl->sg.end = orig_end;
 673         msg_opl->sg.curr = orig_end;
 674         msg_opl->sg.copybreak = 0;
 675         msg_opl->apply_bytes = msg_opl->sg.size + msg_npl->sg.size;
 676         msg_opl->sg.size += msg_npl->sg.size;
 677
 678         sk_msg_free(sk, &to->msg_encrypted);
 679         sk_msg_xfer_full(&to->msg_encrypted, &from->msg_encrypted);
 680
 681         kfree(from);
 682 }
 683
 684 static int tls_push_record(struct sock *sk, int flags,
 685                            unsigned char record_type)
 686 {
 687         struct tls_context *tls_ctx = tls_get_ctx(sk);
 688         struct tls_prot_info *prot = &tls_ctx->prot_info;
 689         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 690         struct tls_rec *rec = ctx->open_rec, *tmp = NULL;
 691         u32 i, split_point, orig_end;
 692         struct sk_msg *msg_pl, *msg_en;
 693         struct aead_request *req;
 694         bool split;
 695         int rc;
 696
 697         if (!rec)
 698                 return 0;
 699
 700         msg_pl = &rec->msg_plaintext;
 701         msg_en = &rec->msg_encrypted;
 702
 703         split_point = msg_pl->apply_bytes;
 704         split = split_point && split_point < msg_pl->sg.size;
 705         if (unlikely((!split &&
 706                       msg_pl->sg.size +
 707                       prot->overhead_size > msg_en->sg.size) ||
 708                      (split &&
 709                       split_point +
 710                       prot->overhead_size > msg_en->sg.size))) {
 711                 split = true;
 712                 split_point = msg_en->sg.size;
 713         }
 714         if (split) {
 715                 rc = tls_split_open_record(sk, rec, &tmp, msg_pl, msg_en,
 716                                            split_point, prot->overhead_size,
 717                                            &orig_end);
 718                 if (rc < 0)
 719                         return rc;
 720                 /* This can happen if above tls_split_open_record allocates
 721                  * a single large encryption buffer instead of two smaller
 722                  * ones. In this case adjust pointers and continue without
 723                  * split.
 724                  */
 725                 if (!msg_pl->sg.size) {
 726                         tls_merge_open_record(sk, rec, tmp, orig_end);
 727                         msg_pl = &rec->msg_plaintext;
 728                         msg_en = &rec->msg_encrypted;
 729                         split = false;
 730                 }
 731                 sk_msg_trim(sk, msg_en, msg_pl->sg.size +
 732                             prot->overhead_size);
 733         }
 734
 735         rec->tx_flags = flags;
 736         req = &rec->aead_req;
 737
 738         i = msg_pl->sg.end;
 739         sk_msg_iter_var_prev(i);
 740
 741         rec->content_type = record_type;
 742         if (prot->version == TLS_1_3_VERSION) {
 743                 /* Add content type to end of message.  No padding added */
 744                 sg_set_buf(&rec->sg_content_type, &rec->content_type, 1);
 745                 sg_mark_end(&rec->sg_content_type);
 746                 sg_chain(msg_pl->sg.data, msg_pl->sg.end + 1,
 747                          &rec->sg_content_type);
 748         } else {
 749                 sg_mark_end(sk_msg_elem(msg_pl, i));
 750         }
 751
 752         if (msg_pl->sg.end < msg_pl->sg.start) {
 753                 sg_chain(&msg_pl->sg.data[msg_pl->sg.start],
 754                          MAX_SKB_FRAGS - msg_pl->sg.start + 1,
 755                          msg_pl->sg.data);
 756         }
 757
 758         i = msg_pl->sg.start;
 759         sg_chain(rec->sg_aead_in, 2, &msg_pl->sg.data[i]);
 760
 761         i = msg_en->sg.end;
 762         sk_msg_iter_var_prev(i);
 763         sg_mark_end(sk_msg_elem(msg_en, i));
 764
 765         i = msg_en->sg.start;
 766         sg_chain(rec->sg_aead_out, 2, &msg_en->sg.data[i]);
 767
 768         tls_make_aad(rec->aad_space, msg_pl->sg.size + prot->tail_size,
 769                      tls_ctx->tx.rec_seq, record_type, prot);
 770
 771         tls_fill_prepend(tls_ctx,
 772                          page_address(sg_page(&msg_en->sg.data[i])) +
 773                          msg_en->sg.data[i].offset,
 774                          msg_pl->sg.size + prot->tail_size,
 775                          record_type);
 776
 777         tls_ctx->pending_open_record_frags = false;
 778
 779         rc = tls_do_encryption(sk, tls_ctx, ctx, req,
 780                                msg_pl->sg.size + prot->tail_size, i);
 781         if (rc < 0) {
 782                 if (rc != -EINPROGRESS) {
 783                         tls_err_abort(sk, -EBADMSG);
 784                         if (split) {
 785                                 tls_ctx->pending_open_record_frags = true;
 786                                 tls_merge_open_record(sk, rec, tmp, orig_end);
 787                         }
 788                 }
 789                 ctx->async_capable = 1;
 790                 return rc;
 791         } else if (split) {
 792                 msg_pl = &tmp->msg_plaintext;
 793                 msg_en = &tmp->msg_encrypted;
 794                 sk_msg_trim(sk, msg_en, msg_pl->sg.size + prot->overhead_size);
 795                 tls_ctx->pending_open_record_frags = true;
 796                 ctx->open_rec = tmp;
 797         }
 798
 799         return tls_tx_records(sk, flags);
 800 }
 801
 802 static int bpf_exec_tx_verdict(struct sk_msg *msg, struct sock *sk,
 803                                bool full_record, u8 record_type,
 804                                ssize_t *copied, int flags)
 805 {
 806         struct tls_context *tls_ctx = tls_get_ctx(sk);
 807         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 808         struct sk_msg msg_redir = { };
 809         struct sk_psock *psock;
 810         struct sock *sk_redir;
 811         struct tls_rec *rec;
 812         bool enospc, policy, redir_ingress;
 813         int err = 0, send;
 814         u32 delta = 0;
 815
 816         policy = !(flags & MSG_SENDPAGE_NOPOLICY);
 817         psock = sk_psock_get(sk);
 818         if (!psock || !policy) {
 819                 err = tls_push_record(sk, flags, record_type);
 820                 if (err && sk->sk_err == EBADMSG) {
 821                         *copied -= sk_msg_free(sk, msg);
 822                         tls_free_open_rec(sk);
 823                         err = -sk->sk_err;
 824                 }
 825                 if (psock)
 826                         sk_psock_put(sk, psock);
 827                 return err;
 828         }
 829 more_data:
 830         enospc = sk_msg_full(msg);
 831         if (psock->eval == __SK_NONE) {
 832                 delta = msg->sg.size;
 833                 psock->eval = sk_psock_msg_verdict(sk, psock, msg);
 834                 delta -= msg->sg.size;
 835         }
 836         if (msg->cork_bytes && msg->cork_bytes > msg->sg.size &&
 837             !enospc && !full_record) {
 838                 err = -ENOSPC;
 839                 goto out_err;
 840         }
 841         msg->cork_bytes = 0;
 842         send = msg->sg.size;
 843         if (msg->apply_bytes && msg->apply_bytes < send)
 844                 send = msg->apply_bytes;
 845
 846         switch (psock->eval) {
 847         case __SK_PASS:
 848                 err = tls_push_record(sk, flags, record_type);
 849                 if (err && sk->sk_err == EBADMSG) {
 850                         *copied -= sk_msg_free(sk, msg);
 851                         tls_free_open_rec(sk);
 852                         err = -sk->sk_err;
 853                         goto out_err;
 854                 }
 855                 break;
 856         case __SK_REDIRECT:
 857                 redir_ingress = psock->redir_ingress;
 858                 sk_redir = psock->sk_redir;
 859                 memcpy(&msg_redir, msg, sizeof(*msg));
 860                 if (msg->apply_bytes < send)
 861                         msg->apply_bytes = 0;
 862                 else
 863                         msg->apply_bytes -= send;
 864                 sk_msg_return_zero(sk, msg, send);
 865                 msg->sg.size -= send;
 866                 release_sock(sk);
 867                 err = tcp_bpf_sendmsg_redir(sk_redir, redir_ingress,
 868                                             &msg_redir, send, flags);
 869                 lock_sock(sk);
 870                 if (err < 0) {
 871                         *copied -= sk_msg_free_nocharge(sk, &msg_redir);
 872                         msg->sg.size = 0;
 873                 }
 874                 if (msg->sg.size == 0)
 875                         tls_free_open_rec(sk);
 876                 break;
 877         case __SK_DROP:
 878         default:
 879                 sk_msg_free_partial(sk, msg, send);
 880                 if (msg->apply_bytes < send)
 881                         msg->apply_bytes = 0;
 882                 else
 883                         msg->apply_bytes -= send;
 884                 if (msg->sg.size == 0)
 885                         tls_free_open_rec(sk);
 886                 *copied -= (send + delta);
 887                 err = -EACCES;
 888         }
 889
 890         if (likely(!err)) {
 891                 bool reset_eval = !ctx->open_rec;
 892
 893                 rec = ctx->open_rec;
 894                 if (rec) {
 895                         msg = &rec->msg_plaintext;
 896                         if (!msg->apply_bytes)
 897                                 reset_eval = true;
 898                 }
 899                 if (reset_eval) {
 900                         psock->eval = __SK_NONE;
 901                         if (psock->sk_redir) {
 902                                 sock_put(psock->sk_redir);
 903                                 psock->sk_redir = NULL;
 904                         }
 905                 }
 906                 if (rec)
 907                         goto more_data;
 908         }
 909  out_err:
 910         sk_psock_put(sk, psock);
 911         return err;
 912 }
 913
 914 static int tls_sw_push_pending_record(struct sock *sk, int flags)
 915 {
 916         struct tls_context *tls_ctx = tls_get_ctx(sk);
 917         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 918         struct tls_rec *rec = ctx->open_rec;
 919         struct sk_msg *msg_pl;
 920         size_t copied;
 921
 922         if (!rec)
 923                 return 0;
 924
 925         msg_pl = &rec->msg_plaintext;
 926         copied = msg_pl->sg.size;
 927         if (!copied)
 928                 return 0;
 929
 930         return bpf_exec_tx_verdict(msg_pl, sk, true, TLS_RECORD_TYPE_DATA,
 931                                    &copied, flags);
 932 }
 933
 934 static int tls_sw_sendmsg_splice(struct sock *sk, struct msghdr *msg,
 935                                  struct sk_msg *msg_pl, size_t try_to_copy,
 936                                  ssize_t *copied)
 937 {
 938         struct page *page = NULL, **pages = &page;
 939
 940         do {
 941                 ssize_t part;
 942                 size_t off;
 943
 944                 part = iov_iter_extract_pages(&msg->msg_iter, &pages,
 945                                               try_to_copy, 1, 0, &off);
 946                 if (part <= 0)
 947                         return part ?: -EIO;
 948
 949                 if (WARN_ON_ONCE(!sendpage_ok(page))) {
 950                         iov_iter_revert(&msg->msg_iter, part);
 951                         return -EIO;
 952                 }
 953
 954                 sk_msg_page_add(msg_pl, page, part, off);
 955                 sk_mem_charge(sk, part);
 956                 *copied += part;
 957                 try_to_copy -= part;
 958         } while (try_to_copy && !sk_msg_full(msg_pl));
 959
 960         return 0;
 961 }
 962
 963 static int tls_sw_sendmsg_locked(struct sock *sk, struct msghdr *msg,
 964                                  size_t size)
 965 {
 966         long timeo = sock_sndtimeo(sk, msg->msg_flags & MSG_DONTWAIT);
 967         struct tls_context *tls_ctx = tls_get_ctx(sk);
 968         struct tls_prot_info *prot = &tls_ctx->prot_info;
 969         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
 970         bool async_capable = ctx->async_capable;
 971         unsigned char record_type = TLS_RECORD_TYPE_DATA;
 972         bool is_kvec = iov_iter_is_kvec(&msg->msg_iter);
 973         bool eor = !(msg->msg_flags & MSG_MORE);
 974         size_t try_to_copy;
 975         ssize_t copied = 0;
 976         struct sk_msg *msg_pl, *msg_en;
 977         struct tls_rec *rec;
 978         int required_size;
 979         int num_async = 0;
 980         bool full_record;
 981         int record_room;
 982         int num_zc = 0;
 983         int orig_size;
 984         int ret = 0;
 985         int pending;
 986
 987         if (unlikely(msg->msg_controllen)) {
 988                 ret = tls_process_cmsg(sk, msg, &record_type);
 989                 if (ret) {
 990                         if (ret == -EINPROGRESS)
 991                                 num_async++;
 992                         else if (ret != -EAGAIN)
 993                                 goto send_end;
 994                 }
 995         }
 996
 997         while (msg_data_left(msg)) {
 998                 if (sk->sk_err) {
 999                         ret = -sk->sk_err;
1000                         goto send_end;
1001                 }
1002
1003                 if (ctx->open_rec)
1004                         rec = ctx->open_rec;
1005                 else
1006                         rec = ctx->open_rec = tls_get_rec(sk);
1007                 if (!rec) {
1008                         ret = -ENOMEM;
1009                         goto send_end;
1010                 }
1011
1012                 msg_pl = &rec->msg_plaintext;
1013                 msg_en = &rec->msg_encrypted;
1014
1015                 orig_size = msg_pl->sg.size;
1016                 full_record = false;
1017                 try_to_copy = msg_data_left(msg);
1018                 record_room = TLS_MAX_PAYLOAD_SIZE - msg_pl->sg.size;
1019                 if (try_to_copy >= record_room) {
1020                         try_to_copy = record_room;
1021                         full_record = true;
1022                 }
1023
1024                 required_size = msg_pl->sg.size + try_to_copy +
1025                                 prot->overhead_size;
1026
1027                 if (!sk_stream_memory_free(sk))
1028                         goto wait_for_sndbuf;
1029
1030 alloc_encrypted:
1031                 ret = tls_alloc_encrypted_msg(sk, required_size);
1032                 if (ret) {
1033                         if (ret != -ENOSPC)
1034                                 goto wait_for_memory;
1035
1036                         /* Adjust try_to_copy according to the amount that was
1037                          * actually allocated. The difference is due
1038                          * to max sg elements limit
1039                          */
1040                         try_to_copy -= required_size - msg_en->sg.size;
1041                         full_record = true;
1042                 }
1043
1044                 if (try_to_copy && (msg->msg_flags & MSG_SPLICE_PAGES)) {
1045                         ret = tls_sw_sendmsg_splice(sk, msg, msg_pl,
1046                                                     try_to_copy, &copied);
1047                         if (ret < 0)
1048                                 goto send_end;
1049                         tls_ctx->pending_open_record_frags = true;
1050                         if (full_record || eor || sk_msg_full(msg_pl))
1051                                 goto copied;
1052                         continue;
1053                 }
1054
1055                 if (!is_kvec && (full_record || eor) && !async_capable) {
1056                         u32 first = msg_pl->sg.end;
1057
1058                         ret = sk_msg_zerocopy_from_iter(sk, &msg->msg_iter,
1059                                                         msg_pl, try_to_copy);
1060                         if (ret)
1061                                 goto fallback_to_reg_send;
1062
1063                         num_zc++;
1064                         copied += try_to_copy;
1065
1066                         sk_msg_sg_copy_set(msg_pl, first);
1067                         ret = bpf_exec_tx_verdict(msg_pl, sk, full_record,
1068                                                   record_type, &copied,
1069                                                   msg->msg_flags);
1070                         if (ret) {
1071                                 if (ret == -EINPROGRESS)
1072                                         num_async++;
1073                                 else if (ret == -ENOMEM)
1074                                         goto wait_for_memory;
1075                                 else if (ctx->open_rec && ret == -ENOSPC)
1076                                         goto rollback_iter;
1077                                 else if (ret != -EAGAIN)
1078                                         goto send_end;
1079                         }
1080                         continue;
1081 rollback_iter:
1082                         copied -= try_to_copy;
1083                         sk_msg_sg_copy_clear(msg_pl, first);
1084                         iov_iter_revert(&msg->msg_iter,
1085                                         msg_pl->sg.size - orig_size);
1086 fallback_to_reg_send:
1087                         sk_msg_trim(sk, msg_pl, orig_size);
1088                 }
1089
1090                 required_size = msg_pl->sg.size + try_to_copy;
1091
1092                 ret = tls_clone_plaintext_msg(sk, required_size);
1093                 if (ret) {
1094                         if (ret != -ENOSPC)
1095                                 goto send_end;
1096
1097                         /* Adjust try_to_copy according to the amount that was
1098                          * actually allocated. The difference is due
1099                          * to max sg elements limit
1100                          */
1101                         try_to_copy -= required_size - msg_pl->sg.size;
1102                         full_record = true;
1103                         sk_msg_trim(sk, msg_en,
1104                                     msg_pl->sg.size + prot->overhead_size);
1105                 }
1106
1107                 if (try_to_copy) {
1108                         ret = sk_msg_memcopy_from_iter(sk, &msg->msg_iter,
1109                                                        msg_pl, try_to_copy);
1110                         if (ret < 0)
1111                                 goto trim_sgl;
1112                 }
1113
1114                 /* Open records defined only if successfully copied, otherwise
1115                  * we would trim the sg but not reset the open record frags.
1116                  */
1117                 tls_ctx->pending_open_record_frags = true;
1118                 copied += try_to_copy;
1119 copied:
1120                 if (full_record || eor) {
1121                         ret = bpf_exec_tx_verdict(msg_pl, sk, full_record,
1122                                                   record_type, &copied,
1123                                                   msg->msg_flags);
1124                         if (ret) {
1125                                 if (ret == -EINPROGRESS)
1126                                         num_async++;
1127                                 else if (ret == -ENOMEM)
1128                                         goto wait_for_memory;
1129                                 else if (ret != -EAGAIN) {
1130                                         if (ret == -ENOSPC)
1131                                                 ret = 0;
1132                                         goto send_end;
1133                                 }
1134                         }
1135                 }
1136
1137                 continue;
1138
1139 wait_for_sndbuf:
1140                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1141 wait_for_memory:
1142                 ret = sk_stream_wait_memory(sk, &timeo);
1143                 if (ret) {
1144 trim_sgl:
1145                         if (ctx->open_rec)
1146                                 tls_trim_both_msgs(sk, orig_size);
1147                         goto send_end;
1148                 }
1149
1150                 if (ctx->open_rec && msg_en->sg.size < required_size)
1151                         goto alloc_encrypted;
1152         }
1153
1154         if (!num_async) {
1155                 goto send_end;
1156         } else if (num_zc) {
1157                 /* Wait for pending encryptions to get completed */
1158                 spin_lock_bh(&ctx->encrypt_compl_lock);
1159                 ctx->async_notify = true;
1160
1161                 pending = atomic_read(&ctx->encrypt_pending);
1162                 spin_unlock_bh(&ctx->encrypt_compl_lock);
1163                 if (pending)
1164                         crypto_wait_req(-EINPROGRESS, &ctx->async_wait);
1165                 else
1166                         reinit_completion(&ctx->async_wait.completion);
1167
1168                 /* There can be no concurrent accesses, since we have no
1169                  * pending encrypt operations
1170                  */
1171                 WRITE_ONCE(ctx->async_notify, false);
1172
1173                 if (ctx->async_wait.err) {
1174                         ret = ctx->async_wait.err;
1175                         copied = 0;
1176                 }
1177         }
1178
1179         /* Transmit if any encryptions have completed */
1180         if (test_and_clear_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask)) {
1181                 cancel_delayed_work(&ctx->tx_work.work);
1182                 tls_tx_records(sk, msg->msg_flags);
1183         }
1184
1185 send_end:
1186         ret = sk_stream_error(sk, msg->msg_flags, ret);
1187         return copied > 0 ? copied : ret;
1188 }
1189
1190 int tls_sw_sendmsg(struct sock *sk, struct msghdr *msg, size_t size)
1191 {
1192         struct tls_context *tls_ctx = tls_get_ctx(sk);
1193         int ret;
1194
1195         if (msg->msg_flags & ~(MSG_MORE | MSG_DONTWAIT | MSG_NOSIGNAL |
1196                                MSG_CMSG_COMPAT | MSG_SPLICE_PAGES |
1197                                MSG_SENDPAGE_NOTLAST | MSG_SENDPAGE_NOPOLICY))
1198                 return -EOPNOTSUPP;
1199
1200         ret = mutex_lock_interruptible(&tls_ctx->tx_lock);
1201         if (ret)
1202                 return ret;
1203         lock_sock(sk);
1204         ret = tls_sw_sendmsg_locked(sk, msg, size);
1205         release_sock(sk);
1206         mutex_unlock(&tls_ctx->tx_lock);
1207         return ret;
1208 }
1209
1210 /*
1211  * Handle unexpected EOF during splice without SPLICE_F_MORE set.
1212  */
1213 void tls_sw_splice_eof(struct socket *sock)
1214 {
1215         struct sock *sk = sock->sk;
1216         struct tls_context *tls_ctx = tls_get_ctx(sk);
1217         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
1218         struct tls_rec *rec;
1219         struct sk_msg *msg_pl;
1220         ssize_t copied = 0;
1221         bool retrying = false;
1222         int ret = 0;
1223         int pending;
1224
1225         if (!ctx->open_rec)
1226                 return;
1227
1228         mutex_lock(&tls_ctx->tx_lock);
1229         lock_sock(sk);
1230
1231 retry:
1232         rec = ctx->open_rec;
1233         if (!rec)
1234                 goto unlock;
1235
1236         msg_pl = &rec->msg_plaintext;
1237
1238         /* Check the BPF advisor and perform transmission. */
1239         ret = bpf_exec_tx_verdict(msg_pl, sk, false, TLS_RECORD_TYPE_DATA,
1240                                   &copied, 0);
1241         switch (ret) {
1242         case 0:
1243         case -EAGAIN:
1244                 if (retrying)
1245                         goto unlock;
1246                 retrying = true;
1247                 goto retry;
1248         case -EINPROGRESS:
1249                 break;
1250         default:
1251                 goto unlock;
1252         }
1253
1254         /* Wait for pending encryptions to get completed */
1255         spin_lock_bh(&ctx->encrypt_compl_lock);
1256         ctx->async_notify = true;
1257
1258         pending = atomic_read(&ctx->encrypt_pending);
1259         spin_unlock_bh(&ctx->encrypt_compl_lock);
1260         if (pending)
1261                 crypto_wait_req(-EINPROGRESS, &ctx->async_wait);
1262         else
1263                 reinit_completion(&ctx->async_wait.completion);
1264
1265         /* There can be no concurrent accesses, since we have no pending
1266          * encrypt operations
1267          */
1268         WRITE_ONCE(ctx->async_notify, false);
1269
1270         if (ctx->async_wait.err)
1271                 goto unlock;
1272
1273         /* Transmit if any encryptions have completed */
1274         if (test_and_clear_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask)) {
1275                 cancel_delayed_work(&ctx->tx_work.work);
1276                 tls_tx_records(sk, 0);
1277         }
1278
1279 unlock:
1280         release_sock(sk);
1281         mutex_unlock(&tls_ctx->tx_lock);
1282 }
1283
1284 static int
1285 tls_rx_rec_wait(struct sock *sk, struct sk_psock *psock, bool nonblock,
1286                 bool released)
1287 {
1288         struct tls_context *tls_ctx = tls_get_ctx(sk);
1289         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1290         DEFINE_WAIT_FUNC(wait, woken_wake_function);
1291         long timeo;
1292
1293         timeo = sock_rcvtimeo(sk, nonblock);
1294
1295         while (!tls_strp_msg_ready(ctx)) {
1296                 if (!sk_psock_queue_empty(psock))
1297                         return 0;
1298
1299                 if (sk->sk_err)
1300                         return sock_error(sk);
1301
1302                 if (!skb_queue_empty(&sk->sk_receive_queue)) {
1303                         tls_strp_check_rcv(&ctx->strp);
1304                         if (tls_strp_msg_ready(ctx))
1305                                 break;
1306                 }
1307
1308                 if (sk->sk_shutdown & RCV_SHUTDOWN)
1309                         return 0;
1310
1311                 if (sock_flag(sk, SOCK_DONE))
1312                         return 0;
1313
1314                 if (!timeo)
1315                         return -EAGAIN;
1316
1317                 released = true;
1318                 add_wait_queue(sk_sleep(sk), &wait);
1319                 sk_set_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1320                 sk_wait_event(sk, &timeo,
1321                               tls_strp_msg_ready(ctx) ||
1322                               !sk_psock_queue_empty(psock),
1323                               &wait);
1324                 sk_clear_bit(SOCKWQ_ASYNC_WAITDATA, sk);
1325                 remove_wait_queue(sk_sleep(sk), &wait);
1326
1327                 /* Handle signals */
1328                 if (signal_pending(current))
1329                         return sock_intr_errno(timeo);
1330         }
1331
1332         tls_strp_msg_load(&ctx->strp, released);
1333
1334         return 1;
1335 }
1336
1337 static int tls_setup_from_iter(struct iov_iter *from,
1338                                int length, int *pages_used,
1339                                struct scatterlist *to,
1340                                int to_max_pages)
1341 {
1342         int rc = 0, i = 0, num_elem = *pages_used, maxpages;
1343         struct page *pages[MAX_SKB_FRAGS];
1344         unsigned int size = 0;
1345         ssize_t copied, use;
1346         size_t offset;
1347
1348         while (length > 0) {
1349                 i = 0;
1350                 maxpages = to_max_pages - num_elem;
1351                 if (maxpages == 0) {
1352                         rc = -EFAULT;
1353                         goto out;
1354                 }
1355                 copied = iov_iter_get_pages2(from, pages,
1356                                             length,
1357                                             maxpages, &offset);
1358                 if (copied <= 0) {
1359                         rc = -EFAULT;
1360                         goto out;
1361                 }
1362
1363                 length -= copied;
1364                 size += copied;
1365                 while (copied) {
1366                         use = min_t(int, copied, PAGE_SIZE - offset);
1367
1368                         sg_set_page(&to[num_elem],
1369                                     pages[i], use, offset);
1370                         sg_unmark_end(&to[num_elem]);
1371                         /* We do not uncharge memory from this API */
1372
1373                         offset = 0;
1374                         copied -= use;
1375
1376                         i++;
1377                         num_elem++;
1378                 }
1379         }
1380         /* Mark the end in the last sg entry if newly added */
1381         if (num_elem > *pages_used)
1382                 sg_mark_end(&to[num_elem - 1]);
1383 out:
1384         if (rc)
1385                 iov_iter_revert(from, size);
1386         *pages_used = num_elem;
1387
1388         return rc;
1389 }
1390
1391 static struct sk_buff *
1392 tls_alloc_clrtxt_skb(struct sock *sk, struct sk_buff *skb,
1393                      unsigned int full_len)
1394 {
1395         struct strp_msg *clr_rxm;
1396         struct sk_buff *clr_skb;
1397         int err;
1398
1399         clr_skb = alloc_skb_with_frags(0, full_len, TLS_PAGE_ORDER,
1400                                        &err, sk->sk_allocation);
1401         if (!clr_skb)
1402                 return NULL;
1403
1404         skb_copy_header(clr_skb, skb);
1405         clr_skb->len = full_len;
1406         clr_skb->data_len = full_len;
1407
1408         clr_rxm = strp_msg(clr_skb);
1409         clr_rxm->offset = 0;
1410
1411         return clr_skb;
1412 }
1413
1414 /* Decrypt handlers
1415  *
1416  * tls_decrypt_sw() and tls_decrypt_device() are decrypt handlers.
1417  * They must transform the darg in/out argument are as follows:
1418  *       |          Input            |         Output
1419  * -------------------------------------------------------------------
1420  *    zc | Zero-copy decrypt allowed | Zero-copy performed
1421  * async | Async decrypt allowed     | Async crypto used / in progress
1422  *   skb |            *              | Output skb
1423  *
1424  * If ZC decryption was performed darg.skb will point to the input skb.
1425  */
1426
1427 /* This function decrypts the input skb into either out_iov or in out_sg
1428  * or in skb buffers itself. The input parameter 'darg->zc' indicates if
1429  * zero-copy mode needs to be tried or not. With zero-copy mode, either
1430  * out_iov or out_sg must be non-NULL. In case both out_iov and out_sg are
1431  * NULL, then the decryption happens inside skb buffers itself, i.e.
1432  * zero-copy gets disabled and 'darg->zc' is updated.
1433  */
1434 static int tls_decrypt_sg(struct sock *sk, struct iov_iter *out_iov,
1435                           struct scatterlist *out_sg,
1436                           struct tls_decrypt_arg *darg)
1437 {
1438         struct tls_context *tls_ctx = tls_get_ctx(sk);
1439         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1440         struct tls_prot_info *prot = &tls_ctx->prot_info;
1441         int n_sgin, n_sgout, aead_size, err, pages = 0;
1442         struct sk_buff *skb = tls_strp_msg(ctx);
1443         const struct strp_msg *rxm = strp_msg(skb);
1444         const struct tls_msg *tlm = tls_msg(skb);
1445         struct aead_request *aead_req;
1446         struct scatterlist *sgin = NULL;
1447         struct scatterlist *sgout = NULL;
1448         const int data_len = rxm->full_len - prot->overhead_size;
1449         int tail_pages = !!prot->tail_size;
1450         struct tls_decrypt_ctx *dctx;
1451         struct sk_buff *clear_skb;
1452         int iv_offset = 0;
1453         u8 *mem;
1454
1455         n_sgin = skb_nsg(skb, rxm->offset + prot->prepend_size,
1456                          rxm->full_len - prot->prepend_size);
1457         if (n_sgin < 1)
1458                 return n_sgin ?: -EBADMSG;
1459
1460         if (darg->zc && (out_iov || out_sg)) {
1461                 clear_skb = NULL;
1462
1463                 if (out_iov)
1464                         n_sgout = 1 + tail_pages +
1465                                 iov_iter_npages_cap(out_iov, INT_MAX, data_len);
1466                 else
1467                         n_sgout = sg_nents(out_sg);
1468         } else {
1469                 darg->zc = false;
1470
1471                 clear_skb = tls_alloc_clrtxt_skb(sk, skb, rxm->full_len);
1472                 if (!clear_skb)
1473                         return -ENOMEM;
1474
1475                 n_sgout = 1 + skb_shinfo(clear_skb)->nr_frags;
1476         }
1477
1478         /* Increment to accommodate AAD */
1479         n_sgin = n_sgin + 1;
1480
1481         /* Allocate a single block of memory which contains
1482          *   aead_req || tls_decrypt_ctx.
1483          * Both structs are variable length.
1484          */
1485         aead_size = sizeof(*aead_req) + crypto_aead_reqsize(ctx->aead_recv);
1486         aead_size = ALIGN(aead_size, __alignof__(*dctx));
1487         mem = kmalloc(aead_size + struct_size(dctx, sg, n_sgin + n_sgout),
1488                       sk->sk_allocation);
1489         if (!mem) {
1490                 err = -ENOMEM;
1491                 goto exit_free_skb;
1492         }
1493
1494         /* Segment the allocated memory */
1495         aead_req = (struct aead_request *)mem;
1496         dctx = (struct tls_decrypt_ctx *)(mem + aead_size);
1497         dctx->sk = sk;
1498         sgin = &dctx->sg[0];
1499         sgout = &dctx->sg[n_sgin];
1500
1501         /* For CCM based ciphers, first byte of nonce+iv is a constant */
1502         switch (prot->cipher_type) {
1503         case TLS_CIPHER_AES_CCM_128:
1504                 dctx->iv[0] = TLS_AES_CCM_IV_B0_BYTE;
1505                 iv_offset = 1;
1506                 break;
1507         case TLS_CIPHER_SM4_CCM:
1508                 dctx->iv[0] = TLS_SM4_CCM_IV_B0_BYTE;
1509                 iv_offset = 1;
1510                 break;
1511         }
1512
1513         /* Prepare IV */
1514         if (prot->version == TLS_1_3_VERSION ||
1515             prot->cipher_type == TLS_CIPHER_CHACHA20_POLY1305) {
1516                 memcpy(&dctx->iv[iv_offset], tls_ctx->rx.iv,
1517                        prot->iv_size + prot->salt_size);
1518         } else {
1519                 err = skb_copy_bits(skb, rxm->offset + TLS_HEADER_SIZE,
1520                                     &dctx->iv[iv_offset] + prot->salt_size,
1521                                     prot->iv_size);
1522                 if (err < 0)
1523                         goto exit_free;
1524                 memcpy(&dctx->iv[iv_offset], tls_ctx->rx.iv, prot->salt_size);
1525         }
1526         tls_xor_iv_with_seq(prot, &dctx->iv[iv_offset], tls_ctx->rx.rec_seq);
1527
1528         /* Prepare AAD */
1529         tls_make_aad(dctx->aad, rxm->full_len - prot->overhead_size +
1530                      prot->tail_size,
1531                      tls_ctx->rx.rec_seq, tlm->control, prot);
1532
1533         /* Prepare sgin */
1534         sg_init_table(sgin, n_sgin);
1535         sg_set_buf(&sgin[0], dctx->aad, prot->aad_size);
1536         err = skb_to_sgvec(skb, &sgin[1],
1537                            rxm->offset + prot->prepend_size,
1538                            rxm->full_len - prot->prepend_size);
1539         if (err < 0)
1540                 goto exit_free;
1541
1542         if (clear_skb) {
1543                 sg_init_table(sgout, n_sgout);
1544                 sg_set_buf(&sgout[0], dctx->aad, prot->aad_size);
1545
1546                 err = skb_to_sgvec(clear_skb, &sgout[1], prot->prepend_size,
1547                                    data_len + prot->tail_size);
1548                 if (err < 0)
1549                         goto exit_free;
1550         } else if (out_iov) {
1551                 sg_init_table(sgout, n_sgout);
1552                 sg_set_buf(&sgout[0], dctx->aad, prot->aad_size);
1553
1554                 err = tls_setup_from_iter(out_iov, data_len, &pages, &sgout[1],
1555                                           (n_sgout - 1 - tail_pages));
1556                 if (err < 0)
1557                         goto exit_free_pages;
1558
1559                 if (prot->tail_size) {
1560                         sg_unmark_end(&sgout[pages]);
1561                         sg_set_buf(&sgout[pages + 1], &dctx->tail,
1562                                    prot->tail_size);
1563                         sg_mark_end(&sgout[pages + 1]);
1564                 }
1565         } else if (out_sg) {
1566                 memcpy(sgout, out_sg, n_sgout * sizeof(*sgout));
1567         }
1568
1569         /* Prepare and submit AEAD request */
1570         err = tls_do_decryption(sk, sgin, sgout, dctx->iv,
1571                                 data_len + prot->tail_size, aead_req, darg);
1572         if (err)
1573                 goto exit_free_pages;
1574
1575         darg->skb = clear_skb ?: tls_strp_msg(ctx);
1576         clear_skb = NULL;
1577
1578         if (unlikely(darg->async)) {
1579                 err = tls_strp_msg_hold(&ctx->strp, &ctx->async_hold);
1580                 if (err)
1581                         __skb_queue_tail(&ctx->async_hold, darg->skb);
1582                 return err;
1583         }
1584
1585         if (prot->tail_size)
1586                 darg->tail = dctx->tail;
1587
1588 exit_free_pages:
1589         /* Release the pages in case iov was mapped to pages */
1590         for (; pages > 0; pages--)
1591                 put_page(sg_page(&sgout[pages]));
1592 exit_free:
1593         kfree(mem);
1594 exit_free_skb:
1595         consume_skb(clear_skb);
1596         return err;
1597 }
1598
1599 static int
1600 tls_decrypt_sw(struct sock *sk, struct tls_context *tls_ctx,
1601                struct msghdr *msg, struct tls_decrypt_arg *darg)
1602 {
1603         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1604         struct tls_prot_info *prot = &tls_ctx->prot_info;
1605         struct strp_msg *rxm;
1606         int pad, err;
1607
1608         err = tls_decrypt_sg(sk, &msg->msg_iter, NULL, darg);
1609         if (err < 0) {
1610                 if (err == -EBADMSG)
1611                         TLS_INC_STATS(sock_net(sk), LINUX_MIB_TLSDECRYPTERROR);
1612                 return err;
1613         }
1614         /* keep going even for ->async, the code below is TLS 1.3 */
1615
1616         /* If opportunistic TLS 1.3 ZC failed retry without ZC */
1617         if (unlikely(darg->zc && prot->version == TLS_1_3_VERSION &&
1618                      darg->tail != TLS_RECORD_TYPE_DATA)) {
1619                 darg->zc = false;
1620                 if (!darg->tail)
1621                         TLS_INC_STATS(sock_net(sk), LINUX_MIB_TLSRXNOPADVIOL);
1622                 TLS_INC_STATS(sock_net(sk), LINUX_MIB_TLSDECRYPTRETRY);
1623                 return tls_decrypt_sw(sk, tls_ctx, msg, darg);
1624         }
1625
1626         pad = tls_padding_length(prot, darg->skb, darg);
1627         if (pad < 0) {
1628                 if (darg->skb != tls_strp_msg(ctx))
1629                         consume_skb(darg->skb);
1630                 return pad;
1631         }
1632
1633         rxm = strp_msg(darg->skb);
1634         rxm->full_len -= pad;
1635
1636         return 0;
1637 }
1638
1639 static int
1640 tls_decrypt_device(struct sock *sk, struct msghdr *msg,
1641                    struct tls_context *tls_ctx, struct tls_decrypt_arg *darg)
1642 {
1643         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1644         struct tls_prot_info *prot = &tls_ctx->prot_info;
1645         struct strp_msg *rxm;
1646         int pad, err;
1647
1648         if (tls_ctx->rx_conf != TLS_HW)
1649                 return 0;
1650
1651         err = tls_device_decrypted(sk, tls_ctx);
1652         if (err <= 0)
1653                 return err;
1654
1655         pad = tls_padding_length(prot, tls_strp_msg(ctx), darg);
1656         if (pad < 0)
1657                 return pad;
1658
1659         darg->async = false;
1660         darg->skb = tls_strp_msg(ctx);
1661         /* ->zc downgrade check, in case TLS 1.3 gets here */
1662         darg->zc &= !(prot->version == TLS_1_3_VERSION &&
1663                       tls_msg(darg->skb)->control != TLS_RECORD_TYPE_DATA);
1664
1665         rxm = strp_msg(darg->skb);
1666         rxm->full_len -= pad;
1667
1668         if (!darg->zc) {
1669                 /* Non-ZC case needs a real skb */
1670                 darg->skb = tls_strp_msg_detach(ctx);
1671                 if (!darg->skb)
1672                         return -ENOMEM;
1673         } else {
1674                 unsigned int off, len;
1675
1676                 /* In ZC case nobody cares about the output skb.
1677                  * Just copy the data here. Note the skb is not fully trimmed.
1678                  */
1679                 off = rxm->offset + prot->prepend_size;
1680                 len = rxm->full_len - prot->overhead_size;
1681
1682                 err = skb_copy_datagram_msg(darg->skb, off, msg, len);
1683                 if (err)
1684                         return err;
1685         }
1686         return 1;
1687 }
1688
1689 static int tls_rx_one_record(struct sock *sk, struct msghdr *msg,
1690                              struct tls_decrypt_arg *darg)
1691 {
1692         struct tls_context *tls_ctx = tls_get_ctx(sk);
1693         struct tls_prot_info *prot = &tls_ctx->prot_info;
1694         struct strp_msg *rxm;
1695         int err;
1696
1697         err = tls_decrypt_device(sk, msg, tls_ctx, darg);
1698         if (!err)
1699                 err = tls_decrypt_sw(sk, tls_ctx, msg, darg);
1700         if (err < 0)
1701                 return err;
1702
1703         rxm = strp_msg(darg->skb);
1704         rxm->offset += prot->prepend_size;
1705         rxm->full_len -= prot->overhead_size;
1706         tls_advance_record_sn(sk, prot, &tls_ctx->rx);
1707
1708         return 0;
1709 }
1710
1711 int decrypt_skb(struct sock *sk, struct scatterlist *sgout)
1712 {
1713         struct tls_decrypt_arg darg = { .zc = true, };
1714
1715         return tls_decrypt_sg(sk, NULL, sgout, &darg);
1716 }
1717
1718 static int tls_record_content_type(struct msghdr *msg, struct tls_msg *tlm,
1719                                    u8 *control)
1720 {
1721         int err;
1722
1723         if (!*control) {
1724                 *control = tlm->control;
1725                 if (!*control)
1726                         return -EBADMSG;
1727
1728                 err = put_cmsg(msg, SOL_TLS, TLS_GET_RECORD_TYPE,
1729                                sizeof(*control), control);
1730                 if (*control != TLS_RECORD_TYPE_DATA) {
1731                         if (err || msg->msg_flags & MSG_CTRUNC)
1732                                 return -EIO;
1733                 }
1734         } else if (*control != tlm->control) {
1735                 return 0;
1736         }
1737
1738         return 1;
1739 }
1740
1741 static void tls_rx_rec_done(struct tls_sw_context_rx *ctx)
1742 {
1743         tls_strp_msg_done(&ctx->strp);
1744 }
1745
1746 /* This function traverses the rx_list in tls receive context to copies the
1747  * decrypted records into the buffer provided by caller zero copy is not
1748  * true. Further, the records are removed from the rx_list if it is not a peek
1749  * case and the record has been consumed completely.
1750  */
1751 static int process_rx_list(struct tls_sw_context_rx *ctx,
1752                            struct msghdr *msg,
1753                            u8 *control,
1754                            size_t skip,
1755                            size_t len,
1756                            bool is_peek)
1757 {
1758         struct sk_buff *skb = skb_peek(&ctx->rx_list);
1759         struct tls_msg *tlm;
1760         ssize_t copied = 0;
1761         int err;
1762
1763         while (skip && skb) {
1764                 struct strp_msg *rxm = strp_msg(skb);
1765                 tlm = tls_msg(skb);
1766
1767                 err = tls_record_content_type(msg, tlm, control);
1768                 if (err <= 0)
1769                         goto out;
1770
1771                 if (skip < rxm->full_len)
1772                         break;
1773
1774                 skip = skip - rxm->full_len;
1775                 skb = skb_peek_next(skb, &ctx->rx_list);
1776         }
1777
1778         while (len && skb) {
1779                 struct sk_buff *next_skb;
1780                 struct strp_msg *rxm = strp_msg(skb);
1781                 int chunk = min_t(unsigned int, rxm->full_len - skip, len);
1782
1783                 tlm = tls_msg(skb);
1784
1785                 err = tls_record_content_type(msg, tlm, control);
1786                 if (err <= 0)
1787                         goto out;
1788
1789                 err = skb_copy_datagram_msg(skb, rxm->offset + skip,
1790                                             msg, chunk);
1791                 if (err < 0)
1792                         goto out;
1793
1794                 len = len - chunk;
1795                 copied = copied + chunk;
1796
1797                 /* Consume the data from record if it is non-peek case*/
1798                 if (!is_peek) {
1799                         rxm->offset = rxm->offset + chunk;
1800                         rxm->full_len = rxm->full_len - chunk;
1801
1802                         /* Return if there is unconsumed data in the record */
1803                         if (rxm->full_len - skip)
1804                                 break;
1805                 }
1806
1807                 /* The remaining skip-bytes must lie in 1st record in rx_list.
1808                  * So from the 2nd record, 'skip' should be 0.
1809                  */
1810                 skip = 0;
1811
1812                 if (msg)
1813                         msg->msg_flags |= MSG_EOR;
1814
1815                 next_skb = skb_peek_next(skb, &ctx->rx_list);
1816
1817                 if (!is_peek) {
1818                         __skb_unlink(skb, &ctx->rx_list);
1819                         consume_skb(skb);
1820                 }
1821
1822                 skb = next_skb;
1823         }
1824         err = 0;
1825
1826 out:
1827         return copied ? : err;
1828 }
1829
1830 static bool
1831 tls_read_flush_backlog(struct sock *sk, struct tls_prot_info *prot,
1832                        size_t len_left, size_t decrypted, ssize_t done,
1833                        size_t *flushed_at)
1834 {
1835         size_t max_rec;
1836
1837         if (len_left <= decrypted)
1838                 return false;
1839
1840         max_rec = prot->overhead_size - prot->tail_size + TLS_MAX_PAYLOAD_SIZE;
1841         if (done - *flushed_at < SZ_128K && tcp_inq(sk) > max_rec)
1842                 return false;
1843
1844         *flushed_at = done;
1845         return sk_flush_backlog(sk);
1846 }
1847
1848 static int tls_rx_reader_lock(struct sock *sk, struct tls_sw_context_rx *ctx,
1849                               bool nonblock)
1850 {
1851         long timeo;
1852         int err;
1853
1854         lock_sock(sk);
1855
1856         timeo = sock_rcvtimeo(sk, nonblock);
1857
1858         while (unlikely(ctx->reader_present)) {
1859                 DEFINE_WAIT_FUNC(wait, woken_wake_function);
1860
1861                 ctx->reader_contended = 1;
1862
1863                 add_wait_queue(&ctx->wq, &wait);
1864                 sk_wait_event(sk, &timeo,
1865                               !READ_ONCE(ctx->reader_present), &wait);
1866                 remove_wait_queue(&ctx->wq, &wait);
1867
1868                 if (timeo <= 0) {
1869                         err = -EAGAIN;
1870                         goto err_unlock;
1871                 }
1872                 if (signal_pending(current)) {
1873                         err = sock_intr_errno(timeo);
1874                         goto err_unlock;
1875                 }
1876         }
1877
1878         WRITE_ONCE(ctx->reader_present, 1);
1879
1880         return 0;
1881
1882 err_unlock:
1883         release_sock(sk);
1884         return err;
1885 }
1886
1887 static void tls_rx_reader_unlock(struct sock *sk, struct tls_sw_context_rx *ctx)
1888 {
1889         if (unlikely(ctx->reader_contended)) {
1890                 if (wq_has_sleeper(&ctx->wq))
1891                         wake_up(&ctx->wq);
1892                 else
1893                         ctx->reader_contended = 0;
1894
1895                 WARN_ON_ONCE(!ctx->reader_present);
1896         }
1897
1898         WRITE_ONCE(ctx->reader_present, 0);
1899         release_sock(sk);
1900 }
1901
1902 int tls_sw_recvmsg(struct sock *sk,
1903                    struct msghdr *msg,
1904                    size_t len,
1905                    int flags,
1906                    int *addr_len)
1907 {
1908         struct tls_context *tls_ctx = tls_get_ctx(sk);
1909         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
1910         struct tls_prot_info *prot = &tls_ctx->prot_info;
1911         ssize_t decrypted = 0, async_copy_bytes = 0;
1912         struct sk_psock *psock;
1913         unsigned char control = 0;
1914         size_t flushed_at = 0;
1915         struct strp_msg *rxm;
1916         struct tls_msg *tlm;
1917         ssize_t copied = 0;
1918         bool async = false;
1919         int target, err;
1920         bool is_kvec = iov_iter_is_kvec(&msg->msg_iter);
1921         bool is_peek = flags & MSG_PEEK;
1922         bool released = true;
1923         bool bpf_strp_enabled;
1924         bool zc_capable;
1925
1926         if (unlikely(flags & MSG_ERRQUEUE))
1927                 return sock_recv_errqueue(sk, msg, len, SOL_IP, IP_RECVERR);
1928
1929         psock = sk_psock_get(sk);
1930         err = tls_rx_reader_lock(sk, ctx, flags & MSG_DONTWAIT);
1931         if (err < 0)
1932                 return err;
1933         bpf_strp_enabled = sk_psock_strp_enabled(psock);
1934
1935         /* If crypto failed the connection is broken */
1936         err = ctx->async_wait.err;
1937         if (err)
1938                 goto end;
1939
1940         /* Process pending decrypted records. It must be non-zero-copy */
1941         err = process_rx_list(ctx, msg, &control, 0, len, is_peek);
1942         if (err < 0)
1943                 goto end;
1944
1945         copied = err;
1946         if (len <= copied)
1947                 goto end;
1948
1949         target = sock_rcvlowat(sk, flags & MSG_WAITALL, len);
1950         len = len - copied;
1951
1952         zc_capable = !bpf_strp_enabled && !is_kvec && !is_peek &&
1953                 ctx->zc_capable;
1954         decrypted = 0;
1955         while (len && (decrypted + copied < target || tls_strp_msg_ready(ctx))) {
1956                 struct tls_decrypt_arg darg;
1957                 int to_decrypt, chunk;
1958
1959                 err = tls_rx_rec_wait(sk, psock, flags & MSG_DONTWAIT,
1960                                       released);
1961                 if (err <= 0) {
1962                         if (psock) {
1963                                 chunk = sk_msg_recvmsg(sk, psock, msg, len,
1964                                                        flags);
1965                                 if (chunk > 0) {
1966                                         decrypted += chunk;
1967                                         len -= chunk;
1968                                         continue;
1969                                 }
1970                         }
1971                         goto recv_end;
1972                 }
1973
1974                 memset(&darg.inargs, 0, sizeof(darg.inargs));
1975
1976                 rxm = strp_msg(tls_strp_msg(ctx));
1977                 tlm = tls_msg(tls_strp_msg(ctx));
1978
1979                 to_decrypt = rxm->full_len - prot->overhead_size;
1980
1981                 if (zc_capable && to_decrypt <= len &&
1982                     tlm->control == TLS_RECORD_TYPE_DATA)
1983                         darg.zc = true;
1984
1985                 /* Do not use async mode if record is non-data */
1986                 if (tlm->control == TLS_RECORD_TYPE_DATA && !bpf_strp_enabled)
1987                         darg.async = ctx->async_capable;
1988                 else
1989                         darg.async = false;
1990
1991                 err = tls_rx_one_record(sk, msg, &darg);
1992                 if (err < 0) {
1993                         tls_err_abort(sk, -EBADMSG);
1994                         goto recv_end;
1995                 }
1996
1997                 async |= darg.async;
1998
1999                 /* If the type of records being processed is not known yet,
2000                  * set it to record type just dequeued. If it is already known,
2001                  * but does not match the record type just dequeued, go to end.
2002                  * We always get record type here since for tls1.2, record type
2003                  * is known just after record is dequeued from stream parser.
2004                  * For tls1.3, we disable async.
2005                  */
2006                 err = tls_record_content_type(msg, tls_msg(darg.skb), &control);
2007                 if (err <= 0) {
2008                         DEBUG_NET_WARN_ON_ONCE(darg.zc);
2009                         tls_rx_rec_done(ctx);
2010 put_on_rx_list_err:
2011                         __skb_queue_tail(&ctx->rx_list, darg.skb);
2012                         goto recv_end;
2013                 }
2014
2015                 /* periodically flush backlog, and feed strparser */
2016                 released = tls_read_flush_backlog(sk, prot, len, to_decrypt,
2017                                                   decrypted + copied,
2018                                                   &flushed_at);
2019
2020                 /* TLS 1.3 may have updated the length by more than overhead */
2021                 rxm = strp_msg(darg.skb);
2022                 chunk = rxm->full_len;
2023                 tls_rx_rec_done(ctx);
2024
2025                 if (!darg.zc) {
2026                         bool partially_consumed = chunk > len;
2027                         struct sk_buff *skb = darg.skb;
2028
2029                         DEBUG_NET_WARN_ON_ONCE(darg.skb == ctx->strp.anchor);
2030
2031                         if (async) {
2032                                 /* TLS 1.2-only, to_decrypt must be text len */
2033                                 chunk = min_t(int, to_decrypt, len);
2034                                 async_copy_bytes += chunk;
2035 put_on_rx_list:
2036                                 decrypted += chunk;
2037                                 len -= chunk;
2038                                 __skb_queue_tail(&ctx->rx_list, skb);
2039                                 continue;
2040                         }
2041
2042                         if (bpf_strp_enabled) {
2043                                 released = true;
2044                                 err = sk_psock_tls_strp_read(psock, skb);
2045                                 if (err != __SK_PASS) {
2046                                         rxm->offset = rxm->offset + rxm->full_len;
2047                                         rxm->full_len = 0;
2048                                         if (err == __SK_DROP)
2049                                                 consume_skb(skb);
2050                                         continue;
2051                                 }
2052                         }
2053
2054                         if (partially_consumed)
2055                                 chunk = len;
2056
2057                         err = skb_copy_datagram_msg(skb, rxm->offset,
2058                                                     msg, chunk);
2059                         if (err < 0)
2060                                 goto put_on_rx_list_err;
2061
2062                         if (is_peek)
2063                                 goto put_on_rx_list;
2064
2065                         if (partially_consumed) {
2066                                 rxm->offset += chunk;
2067                                 rxm->full_len -= chunk;
2068                                 goto put_on_rx_list;
2069                         }
2070
2071                         consume_skb(skb);
2072                 }
2073
2074                 decrypted += chunk;
2075                 len -= chunk;
2076
2077                 /* Return full control message to userspace before trying
2078                  * to parse another message type
2079                  */
2080                 msg->msg_flags |= MSG_EOR;
2081                 if (control != TLS_RECORD_TYPE_DATA)
2082                         break;
2083         }
2084
2085 recv_end:
2086         if (async) {
2087                 int ret, pending;
2088
2089                 /* Wait for all previously submitted records to be decrypted */
2090                 spin_lock_bh(&ctx->decrypt_compl_lock);
2091                 reinit_completion(&ctx->async_wait.completion);
2092                 pending = atomic_read(&ctx->decrypt_pending);
2093                 spin_unlock_bh(&ctx->decrypt_compl_lock);
2094                 ret = 0;
2095                 if (pending)
2096                         ret = crypto_wait_req(-EINPROGRESS, &ctx->async_wait);
2097                 __skb_queue_purge(&ctx->async_hold);
2098
2099                 if (ret) {
2100                         if (err >= 0 || err == -EINPROGRESS)
2101                                 err = ret;
2102                         decrypted = 0;
2103                         goto end;
2104                 }
2105
2106                 /* Drain records from the rx_list & copy if required */
2107                 if (is_peek || is_kvec)
2108                         err = process_rx_list(ctx, msg, &control, copied,
2109                                               decrypted, is_peek);
2110                 else
2111                         err = process_rx_list(ctx, msg, &control, 0,
2112                                               async_copy_bytes, is_peek);
2113                 decrypted += max(err, 0);
2114         }
2115
2116         copied += decrypted;
2117
2118 end:
2119         tls_rx_reader_unlock(sk, ctx);
2120         if (psock)
2121                 sk_psock_put(sk, psock);
2122         return copied ? : err;
2123 }
2124
2125 ssize_t tls_sw_splice_read(struct socket *sock,  loff_t *ppos,
2126                            struct pipe_inode_info *pipe,
2127                            size_t len, unsigned int flags)
2128 {
2129         struct tls_context *tls_ctx = tls_get_ctx(sock->sk);
2130         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2131         struct strp_msg *rxm = NULL;
2132         struct sock *sk = sock->sk;
2133         struct tls_msg *tlm;
2134         struct sk_buff *skb;
2135         ssize_t copied = 0;
2136         int chunk;
2137         int err;
2138
2139         err = tls_rx_reader_lock(sk, ctx, flags & SPLICE_F_NONBLOCK);
2140         if (err < 0)
2141                 return err;
2142
2143         if (!skb_queue_empty(&ctx->rx_list)) {
2144                 skb = __skb_dequeue(&ctx->rx_list);
2145         } else {
2146                 struct tls_decrypt_arg darg;
2147
2148                 err = tls_rx_rec_wait(sk, NULL, flags & SPLICE_F_NONBLOCK,
2149                                       true);
2150                 if (err <= 0)
2151                         goto splice_read_end;
2152
2153                 memset(&darg.inargs, 0, sizeof(darg.inargs));
2154
2155                 err = tls_rx_one_record(sk, NULL, &darg);
2156                 if (err < 0) {
2157                         tls_err_abort(sk, -EBADMSG);
2158                         goto splice_read_end;
2159                 }
2160
2161                 tls_rx_rec_done(ctx);
2162                 skb = darg.skb;
2163         }
2164
2165         rxm = strp_msg(skb);
2166         tlm = tls_msg(skb);
2167
2168         /* splice does not support reading control messages */
2169         if (tlm->control != TLS_RECORD_TYPE_DATA) {
2170                 err = -EINVAL;
2171                 goto splice_requeue;
2172         }
2173
2174         chunk = min_t(unsigned int, rxm->full_len, len);
2175         copied = skb_splice_bits(skb, sk, rxm->offset, pipe, chunk, flags);
2176         if (copied < 0)
2177                 goto splice_requeue;
2178
2179         if (chunk < rxm->full_len) {
2180                 rxm->offset += len;
2181                 rxm->full_len -= len;
2182                 goto splice_requeue;
2183         }
2184
2185         consume_skb(skb);
2186
2187 splice_read_end:
2188         tls_rx_reader_unlock(sk, ctx);
2189         return copied ? : err;
2190
2191 splice_requeue:
2192         __skb_queue_head(&ctx->rx_list, skb);
2193         goto splice_read_end;
2194 }
2195
2196 bool tls_sw_sock_is_readable(struct sock *sk)
2197 {
2198         struct tls_context *tls_ctx = tls_get_ctx(sk);
2199         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2200         bool ingress_empty = true;
2201         struct sk_psock *psock;
2202
2203         rcu_read_lock();
2204         psock = sk_psock(sk);
2205         if (psock)
2206                 ingress_empty = list_empty(&psock->ingress_msg);
2207         rcu_read_unlock();
2208
2209         return !ingress_empty || tls_strp_msg_ready(ctx) ||
2210                 !skb_queue_empty(&ctx->rx_list);
2211 }
2212
2213 int tls_rx_msg_size(struct tls_strparser *strp, struct sk_buff *skb)
2214 {
2215         struct tls_context *tls_ctx = tls_get_ctx(strp->sk);
2216         struct tls_prot_info *prot = &tls_ctx->prot_info;
2217         char header[TLS_HEADER_SIZE + MAX_IV_SIZE];
2218         size_t cipher_overhead;
2219         size_t data_len = 0;
2220         int ret;
2221
2222         /* Verify that we have a full TLS header, or wait for more data */
2223         if (strp->stm.offset + prot->prepend_size > skb->len)
2224                 return 0;
2225
2226         /* Sanity-check size of on-stack buffer. */
2227         if (WARN_ON(prot->prepend_size > sizeof(header))) {
2228                 ret = -EINVAL;
2229                 goto read_failure;
2230         }
2231
2232         /* Linearize header to local buffer */
2233         ret = skb_copy_bits(skb, strp->stm.offset, header, prot->prepend_size);
2234         if (ret < 0)
2235                 goto read_failure;
2236
2237         strp->mark = header[0];
2238
2239         data_len = ((header[4] & 0xFF) | (header[3] << 8));
2240
2241         cipher_overhead = prot->tag_size;
2242         if (prot->version != TLS_1_3_VERSION &&
2243             prot->cipher_type != TLS_CIPHER_CHACHA20_POLY1305)
2244                 cipher_overhead += prot->iv_size;
2245
2246         if (data_len > TLS_MAX_PAYLOAD_SIZE + cipher_overhead +
2247             prot->tail_size) {
2248                 ret = -EMSGSIZE;
2249                 goto read_failure;
2250         }
2251         if (data_len < cipher_overhead) {
2252                 ret = -EBADMSG;
2253                 goto read_failure;
2254         }
2255
2256         /* Note that both TLS1.3 and TLS1.2 use TLS_1_2 version here */
2257         if (header[1] != TLS_1_2_VERSION_MINOR ||
2258             header[2] != TLS_1_2_VERSION_MAJOR) {
2259                 ret = -EINVAL;
2260                 goto read_failure;
2261         }
2262
2263         tls_device_rx_resync_new_rec(strp->sk, data_len + TLS_HEADER_SIZE,
2264                                      TCP_SKB_CB(skb)->seq + strp->stm.offset);
2265         return data_len + TLS_HEADER_SIZE;
2266
2267 read_failure:
2268         tls_err_abort(strp->sk, ret);
2269
2270         return ret;
2271 }
2272
2273 void tls_rx_msg_ready(struct tls_strparser *strp)
2274 {
2275         struct tls_sw_context_rx *ctx;
2276
2277         ctx = container_of(strp, struct tls_sw_context_rx, strp);
2278         ctx->saved_data_ready(strp->sk);
2279 }
2280
2281 static void tls_data_ready(struct sock *sk)
2282 {
2283         struct tls_context *tls_ctx = tls_get_ctx(sk);
2284         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2285         struct sk_psock *psock;
2286         gfp_t alloc_save;
2287
2288         trace_sk_data_ready(sk);
2289
2290         alloc_save = sk->sk_allocation;
2291         sk->sk_allocation = GFP_ATOMIC;
2292         tls_strp_data_ready(&ctx->strp);
2293         sk->sk_allocation = alloc_save;
2294
2295         psock = sk_psock_get(sk);
2296         if (psock) {
2297                 if (!list_empty(&psock->ingress_msg))
2298                         ctx->saved_data_ready(sk);
2299                 sk_psock_put(sk, psock);
2300         }
2301 }
2302
2303 void tls_sw_cancel_work_tx(struct tls_context *tls_ctx)
2304 {
2305         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
2306
2307         set_bit(BIT_TX_CLOSING, &ctx->tx_bitmask);
2308         set_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask);
2309         cancel_delayed_work_sync(&ctx->tx_work.work);
2310 }
2311
2312 void tls_sw_release_resources_tx(struct sock *sk)
2313 {
2314         struct tls_context *tls_ctx = tls_get_ctx(sk);
2315         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
2316         struct tls_rec *rec, *tmp;
2317         int pending;
2318
2319         /* Wait for any pending async encryptions to complete */
2320         spin_lock_bh(&ctx->encrypt_compl_lock);
2321         ctx->async_notify = true;
2322         pending = atomic_read(&ctx->encrypt_pending);
2323         spin_unlock_bh(&ctx->encrypt_compl_lock);
2324
2325         if (pending)
2326                 crypto_wait_req(-EINPROGRESS, &ctx->async_wait);
2327
2328         tls_tx_records(sk, -1);
2329
2330         /* Free up un-sent records in tx_list. First, free
2331          * the partially sent record if any at head of tx_list.
2332          */
2333         if (tls_ctx->partially_sent_record) {
2334                 tls_free_partial_record(sk, tls_ctx);
2335                 rec = list_first_entry(&ctx->tx_list,
2336                                        struct tls_rec, list);
2337                 list_del(&rec->list);
2338                 sk_msg_free(sk, &rec->msg_plaintext);
2339                 kfree(rec);
2340         }
2341
2342         list_for_each_entry_safe(rec, tmp, &ctx->tx_list, list) {
2343                 list_del(&rec->list);
2344                 sk_msg_free(sk, &rec->msg_encrypted);
2345                 sk_msg_free(sk, &rec->msg_plaintext);
2346                 kfree(rec);
2347         }
2348
2349         crypto_free_aead(ctx->aead_send);
2350         tls_free_open_rec(sk);
2351 }
2352
2353 void tls_sw_free_ctx_tx(struct tls_context *tls_ctx)
2354 {
2355         struct tls_sw_context_tx *ctx = tls_sw_ctx_tx(tls_ctx);
2356
2357         kfree(ctx);
2358 }
2359
2360 void tls_sw_release_resources_rx(struct sock *sk)
2361 {
2362         struct tls_context *tls_ctx = tls_get_ctx(sk);
2363         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2364
2365         kfree(tls_ctx->rx.rec_seq);
2366         kfree(tls_ctx->rx.iv);
2367
2368         if (ctx->aead_recv) {
2369                 __skb_queue_purge(&ctx->rx_list);
2370                 crypto_free_aead(ctx->aead_recv);
2371                 tls_strp_stop(&ctx->strp);
2372                 /* If tls_sw_strparser_arm() was not called (cleanup paths)
2373                  * we still want to tls_strp_stop(), but sk->sk_data_ready was
2374                  * never swapped.
2375                  */
2376                 if (ctx->saved_data_ready) {
2377                         write_lock_bh(&sk->sk_callback_lock);
2378                         sk->sk_data_ready = ctx->saved_data_ready;
2379                         write_unlock_bh(&sk->sk_callback_lock);
2380                 }
2381         }
2382 }
2383
2384 void tls_sw_strparser_done(struct tls_context *tls_ctx)
2385 {
2386         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2387
2388         tls_strp_done(&ctx->strp);
2389 }
2390
2391 void tls_sw_free_ctx_rx(struct tls_context *tls_ctx)
2392 {
2393         struct tls_sw_context_rx *ctx = tls_sw_ctx_rx(tls_ctx);
2394
2395         kfree(ctx);
2396 }
2397
2398 void tls_sw_free_resources_rx(struct sock *sk)
2399 {
2400         struct tls_context *tls_ctx = tls_get_ctx(sk);
2401
2402         tls_sw_release_resources_rx(sk);
2403         tls_sw_free_ctx_rx(tls_ctx);
2404 }
2405
2406 /* The work handler to transmitt the encrypted records in tx_list */
2407 static void tx_work_handler(struct work_struct *work)
2408 {
2409         struct delayed_work *delayed_work = to_delayed_work(work);
2410         struct tx_work *tx_work = container_of(delayed_work,
2411                                                struct tx_work, work);
2412         struct sock *sk = tx_work->sk;
2413         struct tls_context *tls_ctx = tls_get_ctx(sk);
2414         struct tls_sw_context_tx *ctx;
2415
2416         if (unlikely(!tls_ctx))
2417                 return;
2418
2419         ctx = tls_sw_ctx_tx(tls_ctx);
2420         if (test_bit(BIT_TX_CLOSING, &ctx->tx_bitmask))
2421                 return;
2422
2423         if (!test_and_clear_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask))
2424                 return;
2425
2426         if (mutex_trylock(&tls_ctx->tx_lock)) {
2427                 lock_sock(sk);
2428                 tls_tx_records(sk, -1);
2429                 release_sock(sk);
2430                 mutex_unlock(&tls_ctx->tx_lock);
2431         } else if (!test_and_set_bit(BIT_TX_SCHEDULED, &ctx->tx_bitmask)) {
2432                 /* Someone is holding the tx_lock, they will likely run Tx
2433                  * and cancel the work on their way out of the lock section.
2434                  * Schedule a long delay just in case.
2435                  */
2436                 schedule_delayed_work(&ctx->tx_work.work, msecs_to_jiffies(10));
2437         }
2438 }
2439
2440 static bool tls_is_tx_ready(struct tls_sw_context_tx *ctx)
2441 {
2442         struct tls_rec *rec;
2443
2444         rec = list_first_entry_or_null(&ctx->tx_list, struct tls_rec, list);
2445         if (!rec)
2446                 return false;
2447
2448         return READ_ONCE(rec->tx_ready);
2449 }
2450
2451 void tls_sw_write_space(struct sock *sk, struct tls_context *ctx)
2452 {
2453         struct tls_sw_context_tx *tx_ctx = tls_sw_ctx_tx(ctx);
2454
2455         /* Schedule the transmission if tx list is ready */
2456         if (tls_is_tx_ready(tx_ctx) &&
2457             !test_and_set_bit(BIT_TX_SCHEDULED, &tx_ctx->tx_bitmask))
2458                 schedule_delayed_work(&tx_ctx->tx_work.work, 0);
2459 }
2460
2461 void tls_sw_strparser_arm(struct sock *sk, struct tls_context *tls_ctx)
2462 {
2463         struct tls_sw_context_rx *rx_ctx = tls_sw_ctx_rx(tls_ctx);
2464
2465         write_lock_bh(&sk->sk_callback_lock);
2466         rx_ctx->saved_data_ready = sk->sk_data_ready;
2467         sk->sk_data_ready = tls_data_ready;
2468         write_unlock_bh(&sk->sk_callback_lock);
2469 }
2470
2471 void tls_update_rx_zc_capable(struct tls_context *tls_ctx)
2472 {
2473         struct tls_sw_context_rx *rx_ctx = tls_sw_ctx_rx(tls_ctx);
2474
2475         rx_ctx->zc_capable = tls_ctx->rx_no_pad ||
2476                 tls_ctx->prot_info.version != TLS_1_3_VERSION;
2477 }
2478
2479 int tls_set_sw_offload(struct sock *sk, struct tls_context *ctx, int tx)
2480 {
2481         struct tls_context *tls_ctx = tls_get_ctx(sk);
2482         struct tls_prot_info *prot = &tls_ctx->prot_info;
2483         struct tls_crypto_info *crypto_info;
2484         struct tls_sw_context_tx *sw_ctx_tx = NULL;
2485         struct tls_sw_context_rx *sw_ctx_rx = NULL;
2486         struct cipher_context *cctx;
2487         struct crypto_aead **aead;
2488         u16 nonce_size, tag_size, iv_size, rec_seq_size, salt_size;
2489         struct crypto_tfm *tfm;
2490         char *iv, *rec_seq, *key, *salt, *cipher_name;
2491         size_t keysize;
2492         int rc = 0;
2493
2494         if (!ctx) {
2495                 rc = -EINVAL;
2496                 goto out;
2497         }
2498
2499         if (tx) {
2500                 if (!ctx->priv_ctx_tx) {
2501                         sw_ctx_tx = kzalloc(sizeof(*sw_ctx_tx), GFP_KERNEL);
2502                         if (!sw_ctx_tx) {
2503                                 rc = -ENOMEM;
2504                                 goto out;
2505                         }
2506                         ctx->priv_ctx_tx = sw_ctx_tx;
2507                 } else {
2508                         sw_ctx_tx =
2509                                 (struct tls_sw_context_tx *)ctx->priv_ctx_tx;
2510                 }
2511         } else {
2512                 if (!ctx->priv_ctx_rx) {
2513                         sw_ctx_rx = kzalloc(sizeof(*sw_ctx_rx), GFP_KERNEL);
2514                         if (!sw_ctx_rx) {
2515                                 rc = -ENOMEM;
2516                                 goto out;
2517                         }
2518                         ctx->priv_ctx_rx = sw_ctx_rx;
2519                 } else {
2520                         sw_ctx_rx =
2521                                 (struct tls_sw_context_rx *)ctx->priv_ctx_rx;
2522                 }
2523         }
2524
2525         if (tx) {
2526                 crypto_init_wait(&sw_ctx_tx->async_wait);
2527                 spin_lock_init(&sw_ctx_tx->encrypt_compl_lock);
2528                 crypto_info = &ctx->crypto_send.info;
2529                 cctx = &ctx->tx;
2530                 aead = &sw_ctx_tx->aead_send;
2531                 INIT_LIST_HEAD(&sw_ctx_tx->tx_list);
2532                 INIT_DELAYED_WORK(&sw_ctx_tx->tx_work.work, tx_work_handler);
2533                 sw_ctx_tx->tx_work.sk = sk;
2534         } else {
2535                 crypto_init_wait(&sw_ctx_rx->async_wait);
2536                 spin_lock_init(&sw_ctx_rx->decrypt_compl_lock);
2537                 init_waitqueue_head(&sw_ctx_rx->wq);
2538                 crypto_info = &ctx->crypto_recv.info;
2539                 cctx = &ctx->rx;
2540                 skb_queue_head_init(&sw_ctx_rx->rx_list);
2541                 skb_queue_head_init(&sw_ctx_rx->async_hold);
2542                 aead = &sw_ctx_rx->aead_recv;
2543         }
2544
2545         switch (crypto_info->cipher_type) {
2546         case TLS_CIPHER_AES_GCM_128: {
2547                 struct tls12_crypto_info_aes_gcm_128 *gcm_128_info;
2548
2549                 gcm_128_info = (void *)crypto_info;
2550                 nonce_size = TLS_CIPHER_AES_GCM_128_IV_SIZE;
2551                 tag_size = TLS_CIPHER_AES_GCM_128_TAG_SIZE;
2552                 iv_size = TLS_CIPHER_AES_GCM_128_IV_SIZE;
2553                 iv = gcm_128_info->iv;
2554                 rec_seq_size = TLS_CIPHER_AES_GCM_128_REC_SEQ_SIZE;
2555                 rec_seq = gcm_128_info->rec_seq;
2556                 keysize = TLS_CIPHER_AES_GCM_128_KEY_SIZE;
2557                 key = gcm_128_info->key;
2558                 salt = gcm_128_info->salt;
2559                 salt_size = TLS_CIPHER_AES_GCM_128_SALT_SIZE;
2560                 cipher_name = "gcm(aes)";
2561                 break;
2562         }
2563         case TLS_CIPHER_AES_GCM_256: {
2564                 struct tls12_crypto_info_aes_gcm_256 *gcm_256_info;
2565
2566                 gcm_256_info = (void *)crypto_info;
2567                 nonce_size = TLS_CIPHER_AES_GCM_256_IV_SIZE;
2568                 tag_size = TLS_CIPHER_AES_GCM_256_TAG_SIZE;
2569                 iv_size = TLS_CIPHER_AES_GCM_256_IV_SIZE;
2570                 iv = gcm_256_info->iv;
2571                 rec_seq_size = TLS_CIPHER_AES_GCM_256_REC_SEQ_SIZE;
2572                 rec_seq = gcm_256_info->rec_seq;
2573                 keysize = TLS_CIPHER_AES_GCM_256_KEY_SIZE;
2574                 key = gcm_256_info->key;
2575                 salt = gcm_256_info->salt;
2576                 salt_size = TLS_CIPHER_AES_GCM_256_SALT_SIZE;
2577                 cipher_name = "gcm(aes)";
2578                 break;
2579         }
2580         case TLS_CIPHER_AES_CCM_128: {
2581                 struct tls12_crypto_info_aes_ccm_128 *ccm_128_info;
2582
2583                 ccm_128_info = (void *)crypto_info;
2584                 nonce_size = TLS_CIPHER_AES_CCM_128_IV_SIZE;
2585                 tag_size = TLS_CIPHER_AES_CCM_128_TAG_SIZE;
2586                 iv_size = TLS_CIPHER_AES_CCM_128_IV_SIZE;
2587                 iv = ccm_128_info->iv;
2588                 rec_seq_size = TLS_CIPHER_AES_CCM_128_REC_SEQ_SIZE;
2589                 rec_seq = ccm_128_info->rec_seq;
2590                 keysize = TLS_CIPHER_AES_CCM_128_KEY_SIZE;
2591                 key = ccm_128_info->key;
2592                 salt = ccm_128_info->salt;
2593                 salt_size = TLS_CIPHER_AES_CCM_128_SALT_SIZE;
2594                 cipher_name = "ccm(aes)";
2595                 break;
2596         }
2597         case TLS_CIPHER_CHACHA20_POLY1305: {
2598                 struct tls12_crypto_info_chacha20_poly1305 *chacha20_poly1305_info;
2599
2600                 chacha20_poly1305_info = (void *)crypto_info;
2601                 nonce_size = 0;
2602                 tag_size = TLS_CIPHER_CHACHA20_POLY1305_TAG_SIZE;
2603                 iv_size = TLS_CIPHER_CHACHA20_POLY1305_IV_SIZE;
2604                 iv = chacha20_poly1305_info->iv;
2605                 rec_seq_size = TLS_CIPHER_CHACHA20_POLY1305_REC_SEQ_SIZE;
2606                 rec_seq = chacha20_poly1305_info->rec_seq;
2607                 keysize = TLS_CIPHER_CHACHA20_POLY1305_KEY_SIZE;
2608                 key = chacha20_poly1305_info->key;
2609                 salt = chacha20_poly1305_info->salt;
2610                 salt_size = TLS_CIPHER_CHACHA20_POLY1305_SALT_SIZE;
2611                 cipher_name = "rfc7539(chacha20,poly1305)";
2612                 break;
2613         }
2614         case TLS_CIPHER_SM4_GCM: {
2615                 struct tls12_crypto_info_sm4_gcm *sm4_gcm_info;
2616
2617                 sm4_gcm_info = (void *)crypto_info;
2618                 nonce_size = TLS_CIPHER_SM4_GCM_IV_SIZE;
2619                 tag_size = TLS_CIPHER_SM4_GCM_TAG_SIZE;
2620                 iv_size = TLS_CIPHER_SM4_GCM_IV_SIZE;
2621                 iv = sm4_gcm_info->iv;
2622                 rec_seq_size = TLS_CIPHER_SM4_GCM_REC_SEQ_SIZE;
2623                 rec_seq = sm4_gcm_info->rec_seq;
2624                 keysize = TLS_CIPHER_SM4_GCM_KEY_SIZE;
2625                 key = sm4_gcm_info->key;
2626                 salt = sm4_gcm_info->salt;
2627                 salt_size = TLS_CIPHER_SM4_GCM_SALT_SIZE;
2628                 cipher_name = "gcm(sm4)";
2629                 break;
2630         }
2631         case TLS_CIPHER_SM4_CCM: {
2632                 struct tls12_crypto_info_sm4_ccm *sm4_ccm_info;
2633
2634                 sm4_ccm_info = (void *)crypto_info;
2635                 nonce_size = TLS_CIPHER_SM4_CCM_IV_SIZE;
2636                 tag_size = TLS_CIPHER_SM4_CCM_TAG_SIZE;
2637                 iv_size = TLS_CIPHER_SM4_CCM_IV_SIZE;
2638                 iv = sm4_ccm_info->iv;
2639                 rec_seq_size = TLS_CIPHER_SM4_CCM_REC_SEQ_SIZE;
2640                 rec_seq = sm4_ccm_info->rec_seq;
2641                 keysize = TLS_CIPHER_SM4_CCM_KEY_SIZE;
2642                 key = sm4_ccm_info->key;
2643                 salt = sm4_ccm_info->salt;
2644                 salt_size = TLS_CIPHER_SM4_CCM_SALT_SIZE;
2645                 cipher_name = "ccm(sm4)";
2646                 break;
2647         }
2648         case TLS_CIPHER_ARIA_GCM_128: {
2649                 struct tls12_crypto_info_aria_gcm_128 *aria_gcm_128_info;
2650
2651                 aria_gcm_128_info = (void *)crypto_info;
2652                 nonce_size = TLS_CIPHER_ARIA_GCM_128_IV_SIZE;
2653                 tag_size = TLS_CIPHER_ARIA_GCM_128_TAG_SIZE;
2654                 iv_size = TLS_CIPHER_ARIA_GCM_128_IV_SIZE;
2655                 iv = aria_gcm_128_info->iv;
2656                 rec_seq_size = TLS_CIPHER_ARIA_GCM_128_REC_SEQ_SIZE;
2657                 rec_seq = aria_gcm_128_info->rec_seq;
2658                 keysize = TLS_CIPHER_ARIA_GCM_128_KEY_SIZE;
2659                 key = aria_gcm_128_info->key;
2660                 salt = aria_gcm_128_info->salt;
2661                 salt_size = TLS_CIPHER_ARIA_GCM_128_SALT_SIZE;
2662                 cipher_name = "gcm(aria)";
2663                 break;
2664         }
2665         case TLS_CIPHER_ARIA_GCM_256: {
2666                 struct tls12_crypto_info_aria_gcm_256 *gcm_256_info;
2667
2668                 gcm_256_info = (void *)crypto_info;
2669                 nonce_size = TLS_CIPHER_ARIA_GCM_256_IV_SIZE;
2670                 tag_size = TLS_CIPHER_ARIA_GCM_256_TAG_SIZE;
2671                 iv_size = TLS_CIPHER_ARIA_GCM_256_IV_SIZE;
2672                 iv = gcm_256_info->iv;
2673                 rec_seq_size = TLS_CIPHER_ARIA_GCM_256_REC_SEQ_SIZE;
2674                 rec_seq = gcm_256_info->rec_seq;
2675                 keysize = TLS_CIPHER_ARIA_GCM_256_KEY_SIZE;
2676                 key = gcm_256_info->key;
2677                 salt = gcm_256_info->salt;
2678                 salt_size = TLS_CIPHER_ARIA_GCM_256_SALT_SIZE;
2679                 cipher_name = "gcm(aria)";
2680                 break;
2681         }
2682         default:
2683                 rc = -EINVAL;
2684                 goto free_priv;
2685         }
2686
2687         if (crypto_info->version == TLS_1_3_VERSION) {
2688                 nonce_size = 0;
2689                 prot->aad_size = TLS_HEADER_SIZE;
2690                 prot->tail_size = 1;
2691         } else {
2692                 prot->aad_size = TLS_AAD_SPACE_SIZE;
2693                 prot->tail_size = 0;
2694         }
2695
2696         /* Sanity-check the sizes for stack allocations. */
2697         if (iv_size > MAX_IV_SIZE || nonce_size > MAX_IV_SIZE ||
2698             rec_seq_size > TLS_MAX_REC_SEQ_SIZE || tag_size != TLS_TAG_SIZE ||
2699             prot->aad_size > TLS_MAX_AAD_SIZE) {
2700                 rc = -EINVAL;
2701                 goto free_priv;
2702         }
2703
2704         prot->version = crypto_info->version;
2705         prot->cipher_type = crypto_info->cipher_type;
2706         prot->prepend_size = TLS_HEADER_SIZE + nonce_size;
2707         prot->tag_size = tag_size;
2708         prot->overhead_size = prot->prepend_size +
2709                               prot->tag_size + prot->tail_size;
2710         prot->iv_size = iv_size;
2711         prot->salt_size = salt_size;
2712         cctx->iv = kmalloc(iv_size + salt_size, GFP_KERNEL);
2713         if (!cctx->iv) {
2714                 rc = -ENOMEM;
2715                 goto free_priv;
2716         }
2717         /* Note: 128 & 256 bit salt are the same size */
2718         prot->rec_seq_size = rec_seq_size;
2719         memcpy(cctx->iv, salt, salt_size);
2720         memcpy(cctx->iv + salt_size, iv, iv_size);
2721         cctx->rec_seq = kmemdup(rec_seq, rec_seq_size, GFP_KERNEL);
2722         if (!cctx->rec_seq) {
2723                 rc = -ENOMEM;
2724                 goto free_iv;
2725         }
2726
2727         if (!*aead) {
2728                 *aead = crypto_alloc_aead(cipher_name, 0, 0);
2729                 if (IS_ERR(*aead)) {
2730                         rc = PTR_ERR(*aead);
2731                         *aead = NULL;
2732                         goto free_rec_seq;
2733                 }
2734         }
2735
2736         ctx->push_pending_record = tls_sw_push_pending_record;
2737
2738         rc = crypto_aead_setkey(*aead, key, keysize);
2739
2740         if (rc)
2741                 goto free_aead;
2742
2743         rc = crypto_aead_setauthsize(*aead, prot->tag_size);
2744         if (rc)
2745                 goto free_aead;
2746
2747         if (sw_ctx_rx) {
2748                 tfm = crypto_aead_tfm(sw_ctx_rx->aead_recv);
2749
2750                 tls_update_rx_zc_capable(ctx);
2751                 sw_ctx_rx->async_capable =
2752                         crypto_info->version != TLS_1_3_VERSION &&
2753                         !!(tfm->__crt_alg->cra_flags & CRYPTO_ALG_ASYNC);
2754
2755                 rc = tls_strp_init(&sw_ctx_rx->strp, sk);
2756                 if (rc)
2757                         goto free_aead;
2758         }
2759
2760         goto out;
2761
2762 free_aead:
2763         crypto_free_aead(*aead);
2764         *aead = NULL;
2765 free_rec_seq:
2766         kfree(cctx->rec_seq);
2767         cctx->rec_seq = NULL;
2768 free_iv:
2769         kfree(cctx->iv);
2770         cctx->iv = NULL;
2771 free_priv:
2772         if (tx) {
2773                 kfree(ctx->priv_ctx_tx);
2774                 ctx->priv_ctx_tx = NULL;
2775         } else {
2776                 kfree(ctx->priv_ctx_rx);
2777                 ctx->priv_ctx_rx = NULL;
2778         }
2779 out:
2780         return rc;
2781 }